Rev 4251 | Go to most recent revision | Only display areas with differences | Regard whitespace | Details | Blame | Last modification | View Log | RSS feed
Rev 4251 | Rev 4281 | ||
---|---|---|---|
1 | /* |
1 | /* |
2 | * Copyright (c) 2011 Intel Corporation |
2 | * Copyright (c) 2011 Intel Corporation |
3 | * |
3 | * |
4 | * Permission is hereby granted, free of charge, to any person obtaining a |
4 | * Permission is hereby granted, free of charge, to any person obtaining a |
5 | * copy of this software and associated documentation files (the "Software"), |
5 | * copy of this software and associated documentation files (the "Software"), |
6 | * to deal in the Software without restriction, including without limitation |
6 | * to deal in the Software without restriction, including without limitation |
7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
8 | * and/or sell copies of the Software, and to permit persons to whom the |
8 | * and/or sell copies of the Software, and to permit persons to whom the |
9 | * Software is furnished to do so, subject to the following conditions: |
9 | * Software is furnished to do so, subject to the following conditions: |
10 | * |
10 | * |
11 | * The above copyright notice and this permission notice (including the next |
11 | * The above copyright notice and this permission notice (including the next |
12 | * paragraph) shall be included in all copies or substantial portions of the |
12 | * paragraph) shall be included in all copies or substantial portions of the |
13 | * Software. |
13 | * Software. |
14 | * |
14 | * |
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
20 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
20 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
21 | * SOFTWARE. |
21 | * SOFTWARE. |
22 | * |
22 | * |
23 | * Authors: |
23 | * Authors: |
24 | * Chris Wilson |
24 | * Chris Wilson |
25 | * |
25 | * |
26 | */ |
26 | */ |
27 | 27 | ||
28 | #ifdef HAVE_CONFIG_H |
28 | #ifdef HAVE_CONFIG_H |
29 | #include "config.h" |
29 | #include "config.h" |
30 | #endif |
30 | #endif |
31 | 31 | ||
32 | #include "sna.h" |
32 | #include "sna.h" |
33 | #include "sna_reg.h" |
33 | #include "sna_reg.h" |
34 | 34 | ||
35 | #include |
35 | #include |
36 | #include |
36 | #include |
37 | #include |
37 | #include |
38 | 38 | ||
39 | #ifdef HAVE_VALGRIND |
39 | #ifdef HAVE_VALGRIND |
40 | #include |
40 | #include |
41 | #include |
41 | #include |
42 | #endif |
42 | #endif |
43 | 43 | ||
44 | #ifdef HAVE_STRUCT_SYSINFO_TOTALRAM |
44 | #ifdef HAVE_STRUCT_SYSINFO_TOTALRAM |
45 | #include |
45 | #include |
46 | #endif |
46 | #endif |
47 | 47 | ||
48 | #include "sna_cpuid.h" |
48 | #include "sna_cpuid.h" |
49 | 49 | ||
50 | static struct kgem_bo * |
50 | static struct kgem_bo * |
51 | search_linear_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags); |
51 | search_linear_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags); |
52 | 52 | ||
53 | static struct kgem_bo * |
53 | static struct kgem_bo * |
54 | search_snoop_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags); |
54 | search_snoop_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags); |
55 | 55 | ||
56 | #define DBG_NO_HW 0 |
56 | #define DBG_NO_HW 0 |
57 | #define DBG_NO_TILING 0 |
57 | #define DBG_NO_TILING 0 |
58 | #define DBG_NO_CACHE 0 |
58 | #define DBG_NO_CACHE 0 |
59 | #define DBG_NO_CACHE_LEVEL 0 |
59 | #define DBG_NO_CACHE_LEVEL 0 |
60 | #define DBG_NO_CPU 0 |
60 | #define DBG_NO_CPU 0 |
61 | #define DBG_NO_CREATE2 1 |
61 | #define DBG_NO_CREATE2 1 |
62 | #define DBG_NO_USERPTR 0 |
62 | #define DBG_NO_USERPTR 0 |
63 | #define DBG_NO_UNSYNCHRONIZED_USERPTR 0 |
63 | #define DBG_NO_UNSYNCHRONIZED_USERPTR 0 |
64 | #define DBG_NO_LLC 0 |
64 | #define DBG_NO_LLC 0 |
65 | #define DBG_NO_SEMAPHORES 0 |
65 | #define DBG_NO_SEMAPHORES 0 |
66 | #define DBG_NO_MADV 1 |
66 | #define DBG_NO_MADV 1 |
67 | #define DBG_NO_UPLOAD_CACHE 0 |
67 | #define DBG_NO_UPLOAD_CACHE 0 |
68 | #define DBG_NO_UPLOAD_ACTIVE 0 |
68 | #define DBG_NO_UPLOAD_ACTIVE 0 |
69 | #define DBG_NO_MAP_UPLOAD 0 |
69 | #define DBG_NO_MAP_UPLOAD 0 |
70 | #define DBG_NO_RELAXED_FENCING 0 |
70 | #define DBG_NO_RELAXED_FENCING 0 |
71 | #define DBG_NO_SECURE_BATCHES 0 |
71 | #define DBG_NO_SECURE_BATCHES 0 |
72 | #define DBG_NO_PINNED_BATCHES 0 |
72 | #define DBG_NO_PINNED_BATCHES 0 |
73 | #define DBG_NO_FAST_RELOC 0 |
73 | #define DBG_NO_FAST_RELOC 0 |
74 | #define DBG_NO_HANDLE_LUT 1 |
74 | #define DBG_NO_HANDLE_LUT 1 |
75 | #define DBG_NO_WT 0 |
75 | #define DBG_NO_WT 0 |
76 | #define DBG_DUMP 0 |
76 | #define DBG_DUMP 0 |
77 | 77 | ||
78 | #define FORCE_MMAP_SYNC 0 /* ((1 << DOMAIN_CPU) | (1 << DOMAIN_GTT)) */ |
78 | #define FORCE_MMAP_SYNC 0 /* ((1 << DOMAIN_CPU) | (1 << DOMAIN_GTT)) */ |
79 | 79 | ||
80 | #ifndef DEBUG_SYNC |
80 | #ifndef DEBUG_SYNC |
81 | #define DEBUG_SYNC 0 |
81 | #define DEBUG_SYNC 0 |
82 | #endif |
82 | #endif |
83 | 83 | ||
84 | 84 | ||
85 | #if 0 |
85 | #if 0 |
86 | #define ASSERT_IDLE(kgem__, handle__) assert(!__kgem_busy(kgem__, handle__)) |
86 | #define ASSERT_IDLE(kgem__, handle__) assert(!__kgem_busy(kgem__, handle__)) |
87 | #define ASSERT_MAYBE_IDLE(kgem__, handle__, expect__) assert(!(expect__) || !__kgem_busy(kgem__, handle__)) |
87 | #define ASSERT_MAYBE_IDLE(kgem__, handle__, expect__) assert(!(expect__) || !__kgem_busy(kgem__, handle__)) |
88 | #else |
88 | #else |
89 | #define ASSERT_IDLE(kgem__, handle__) |
89 | #define ASSERT_IDLE(kgem__, handle__) |
90 | #define ASSERT_MAYBE_IDLE(kgem__, handle__, expect__) |
90 | #define ASSERT_MAYBE_IDLE(kgem__, handle__, expect__) |
91 | #endif |
91 | #endif |
92 | 92 | ||
93 | /* Worst case seems to be 965gm where we cannot write within a cacheline that |
93 | /* Worst case seems to be 965gm where we cannot write within a cacheline that |
94 | * is being simultaneously being read by the GPU, or within the sampler |
94 | * is being simultaneously being read by the GPU, or within the sampler |
95 | * prefetch. In general, the chipsets seem to have a requirement that sampler |
95 | * prefetch. In general, the chipsets seem to have a requirement that sampler |
96 | * offsets be aligned to a cacheline (64 bytes). |
96 | * offsets be aligned to a cacheline (64 bytes). |
97 | */ |
97 | */ |
98 | #define UPLOAD_ALIGNMENT 128 |
98 | #define UPLOAD_ALIGNMENT 128 |
99 | 99 | ||
100 | #define PAGE_ALIGN(x) ALIGN(x, PAGE_SIZE) |
100 | #define PAGE_ALIGN(x) ALIGN(x, PAGE_SIZE) |
101 | #define NUM_PAGES(x) (((x) + PAGE_SIZE-1) / PAGE_SIZE) |
101 | #define NUM_PAGES(x) (((x) + PAGE_SIZE-1) / PAGE_SIZE) |
102 | 102 | ||
103 | #define MAX_GTT_VMA_CACHE 512 |
103 | #define MAX_GTT_VMA_CACHE 512 |
104 | #define MAX_CPU_VMA_CACHE INT16_MAX |
104 | #define MAX_CPU_VMA_CACHE INT16_MAX |
105 | #define MAP_PRESERVE_TIME 10 |
105 | #define MAP_PRESERVE_TIME 10 |
106 | 106 | ||
107 | #define MAKE_CPU_MAP(ptr) ((void*)((uintptr_t)(ptr) | 1)) |
107 | #define MAKE_CPU_MAP(ptr) ((void*)((uintptr_t)(ptr) | 1)) |
108 | #define MAKE_USER_MAP(ptr) ((void*)((uintptr_t)(ptr) | 3)) |
108 | #define MAKE_USER_MAP(ptr) ((void*)((uintptr_t)(ptr) | 3)) |
109 | #define IS_USER_MAP(ptr) ((uintptr_t)(ptr) & 2) |
109 | #define IS_USER_MAP(ptr) ((uintptr_t)(ptr) & 2) |
110 | #define __MAP_TYPE(ptr) ((uintptr_t)(ptr) & 3) |
110 | #define __MAP_TYPE(ptr) ((uintptr_t)(ptr) & 3) |
111 | 111 | ||
112 | #define MAKE_REQUEST(rq, ring) ((struct kgem_request *)((uintptr_t)(rq) | (ring))) |
112 | #define MAKE_REQUEST(rq, ring) ((struct kgem_request *)((uintptr_t)(rq) | (ring))) |
113 | 113 | ||
114 | #define LOCAL_I915_PARAM_HAS_BLT 11 |
114 | #define LOCAL_I915_PARAM_HAS_BLT 11 |
115 | #define LOCAL_I915_PARAM_HAS_RELAXED_FENCING 12 |
115 | #define LOCAL_I915_PARAM_HAS_RELAXED_FENCING 12 |
116 | #define LOCAL_I915_PARAM_HAS_RELAXED_DELTA 15 |
116 | #define LOCAL_I915_PARAM_HAS_RELAXED_DELTA 15 |
117 | #define LOCAL_I915_PARAM_HAS_SEMAPHORES 20 |
117 | #define LOCAL_I915_PARAM_HAS_SEMAPHORES 20 |
118 | #define LOCAL_I915_PARAM_HAS_SECURE_BATCHES 23 |
118 | #define LOCAL_I915_PARAM_HAS_SECURE_BATCHES 23 |
119 | #define LOCAL_I915_PARAM_HAS_PINNED_BATCHES 24 |
119 | #define LOCAL_I915_PARAM_HAS_PINNED_BATCHES 24 |
120 | #define LOCAL_I915_PARAM_HAS_NO_RELOC 25 |
120 | #define LOCAL_I915_PARAM_HAS_NO_RELOC 25 |
121 | #define LOCAL_I915_PARAM_HAS_HANDLE_LUT 26 |
121 | #define LOCAL_I915_PARAM_HAS_HANDLE_LUT 26 |
122 | #define LOCAL_I915_PARAM_HAS_WT 27 |
122 | #define LOCAL_I915_PARAM_HAS_WT 27 |
123 | 123 | ||
124 | #define LOCAL_I915_EXEC_IS_PINNED (1<<10) |
124 | #define LOCAL_I915_EXEC_IS_PINNED (1<<10) |
125 | #define LOCAL_I915_EXEC_NO_RELOC (1<<11) |
125 | #define LOCAL_I915_EXEC_NO_RELOC (1<<11) |
126 | #define LOCAL_I915_EXEC_HANDLE_LUT (1<<12) |
126 | #define LOCAL_I915_EXEC_HANDLE_LUT (1<<12) |
127 | struct local_i915_gem_userptr { |
127 | struct local_i915_gem_userptr { |
128 | uint64_t user_ptr; |
128 | uint64_t user_ptr; |
129 | uint64_t user_size; |
129 | uint64_t user_size; |
130 | uint32_t flags; |
130 | uint32_t flags; |
131 | #define I915_USERPTR_READ_ONLY (1<<0) |
131 | #define I915_USERPTR_READ_ONLY (1<<0) |
132 | #define I915_USERPTR_UNSYNCHRONIZED (1<<31) |
132 | #define I915_USERPTR_UNSYNCHRONIZED (1<<31) |
133 | uint32_t handle; |
133 | uint32_t handle; |
134 | }; |
134 | }; |
135 | 135 | ||
136 | #define UNCACHED 0 |
136 | #define UNCACHED 0 |
137 | #define SNOOPED 1 |
137 | #define SNOOPED 1 |
138 | #define DISPLAY 2 |
138 | #define DISPLAY 2 |
139 | 139 | ||
140 | struct local_i915_gem_caching { |
140 | struct local_i915_gem_caching { |
141 | uint32_t handle; |
141 | uint32_t handle; |
142 | uint32_t caching; |
142 | uint32_t caching; |
143 | }; |
143 | }; |
144 | 144 | ||
145 | #define LOCAL_IOCTL_I915_GEM_SET_CACHING SRV_I915_GEM_SET_CACHING |
145 | #define LOCAL_IOCTL_I915_GEM_SET_CACHING SRV_I915_GEM_SET_CACHING |
146 | 146 | ||
147 | struct local_fbinfo { |
147 | struct local_fbinfo { |
148 | int width; |
148 | int width; |
149 | int height; |
149 | int height; |
150 | int pitch; |
150 | int pitch; |
151 | int tiling; |
151 | int tiling; |
152 | }; |
152 | }; |
153 | 153 | ||
154 | struct kgem_buffer { |
154 | struct kgem_buffer { |
155 | struct kgem_bo base; |
155 | struct kgem_bo base; |
156 | void *mem; |
156 | void *mem; |
157 | uint32_t used; |
157 | uint32_t used; |
158 | uint32_t need_io : 1; |
158 | uint32_t need_io : 1; |
159 | uint32_t write : 2; |
159 | uint32_t write : 2; |
160 | uint32_t mmapped : 1; |
160 | uint32_t mmapped : 1; |
161 | }; |
161 | }; |
162 | 162 | ||
163 | static struct kgem_bo *__kgem_freed_bo; |
163 | static struct kgem_bo *__kgem_freed_bo; |
164 | static struct kgem_request *__kgem_freed_request; |
164 | static struct kgem_request *__kgem_freed_request; |
165 | static struct drm_i915_gem_exec_object2 _kgem_dummy_exec; |
165 | static struct drm_i915_gem_exec_object2 _kgem_dummy_exec; |
166 | 166 | ||
167 | static inline int bytes(struct kgem_bo *bo) |
167 | static inline int bytes(struct kgem_bo *bo) |
168 | { |
168 | { |
169 | return __kgem_bo_size(bo); |
169 | return __kgem_bo_size(bo); |
170 | } |
170 | } |
171 | 171 | ||
172 | #define bucket(B) (B)->size.pages.bucket |
172 | #define bucket(B) (B)->size.pages.bucket |
173 | #define num_pages(B) (B)->size.pages.count |
173 | #define num_pages(B) (B)->size.pages.count |
174 | 174 | ||
175 | #ifdef DEBUG_MEMORY |
175 | #ifdef DEBUG_MEMORY |
176 | static void debug_alloc(struct kgem *kgem, size_t size) |
176 | static void debug_alloc(struct kgem *kgem, size_t size) |
177 | { |
177 | { |
178 | kgem->debug_memory.bo_allocs++; |
178 | kgem->debug_memory.bo_allocs++; |
179 | kgem->debug_memory.bo_bytes += size; |
179 | kgem->debug_memory.bo_bytes += size; |
180 | } |
180 | } |
181 | static void debug_alloc__bo(struct kgem *kgem, struct kgem_bo *bo) |
181 | static void debug_alloc__bo(struct kgem *kgem, struct kgem_bo *bo) |
182 | { |
182 | { |
183 | debug_alloc(kgem, bytes(bo)); |
183 | debug_alloc(kgem, bytes(bo)); |
184 | } |
184 | } |
185 | #else |
185 | #else |
186 | #define debug_alloc(k, b) |
186 | #define debug_alloc(k, b) |
187 | #define debug_alloc__bo(k, b) |
187 | #define debug_alloc__bo(k, b) |
188 | #endif |
188 | #endif |
189 | 189 | ||
190 | #ifndef NDEBUG |
190 | #ifndef NDEBUG |
191 | static void assert_tiling(struct kgem *kgem, struct kgem_bo *bo) |
191 | static void assert_tiling(struct kgem *kgem, struct kgem_bo *bo) |
192 | { |
192 | { |
193 | struct drm_i915_gem_get_tiling tiling; |
193 | struct drm_i915_gem_get_tiling tiling; |
194 | 194 | ||
195 | assert(bo); |
195 | assert(bo); |
196 | 196 | ||
197 | VG_CLEAR(tiling); |
197 | VG_CLEAR(tiling); |
198 | tiling.handle = bo->handle; |
198 | tiling.handle = bo->handle; |
199 | tiling.tiling_mode = -1; |
199 | tiling.tiling_mode = -1; |
200 | (void)drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_GET_TILING, &tiling); |
200 | (void)drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_GET_TILING, &tiling); |
201 | assert(tiling.tiling_mode == bo->tiling); |
201 | assert(tiling.tiling_mode == bo->tiling); |
202 | } |
202 | } |
203 | #else |
203 | #else |
204 | #define assert_tiling(kgem, bo) |
204 | #define assert_tiling(kgem, bo) |
205 | #endif |
205 | #endif |
206 | 206 | ||
207 | static void kgem_sna_reset(struct kgem *kgem) |
207 | static void kgem_sna_reset(struct kgem *kgem) |
208 | { |
208 | { |
209 | struct sna *sna = container_of(kgem, struct sna, kgem); |
209 | struct sna *sna = container_of(kgem, struct sna, kgem); |
210 | 210 | ||
211 | sna->render.reset(sna); |
211 | sna->render.reset(sna); |
212 | sna->blt_state.fill_bo = 0; |
212 | sna->blt_state.fill_bo = 0; |
213 | } |
213 | } |
214 | 214 | ||
215 | static void kgem_sna_flush(struct kgem *kgem) |
215 | static void kgem_sna_flush(struct kgem *kgem) |
216 | { |
216 | { |
217 | struct sna *sna = container_of(kgem, struct sna, kgem); |
217 | struct sna *sna = container_of(kgem, struct sna, kgem); |
218 | 218 | ||
219 | sna->render.flush(sna); |
219 | sna->render.flush(sna); |
220 | 220 | ||
221 | // if (sna->render.solid_cache.dirty) |
221 | // if (sna->render.solid_cache.dirty) |
222 | // sna_render_flush_solid(sna); |
222 | // sna_render_flush_solid(sna); |
223 | } |
223 | } |
224 | 224 | ||
225 | static bool gem_set_tiling(int fd, uint32_t handle, int tiling, int stride) |
225 | static bool gem_set_tiling(int fd, uint32_t handle, int tiling, int stride) |
226 | { |
226 | { |
227 | struct drm_i915_gem_set_tiling set_tiling; |
227 | struct drm_i915_gem_set_tiling set_tiling; |
228 | int ret; |
228 | int ret; |
229 | 229 | ||
230 | if (DBG_NO_TILING) |
230 | if (DBG_NO_TILING) |
231 | return false; |
231 | return false; |
232 | 232 | ||
233 | VG_CLEAR(set_tiling); |
233 | VG_CLEAR(set_tiling); |
234 | do { |
234 | do { |
235 | set_tiling.handle = handle; |
235 | set_tiling.handle = handle; |
236 | set_tiling.tiling_mode = tiling; |
236 | set_tiling.tiling_mode = tiling; |
237 | set_tiling.stride = stride; |
237 | set_tiling.stride = stride; |
238 | 238 | ||
239 | ret = drmIoctl(fd, DRM_IOCTL_I915_GEM_SET_TILING, &set_tiling); |
239 | ret = drmIoctl(fd, DRM_IOCTL_I915_GEM_SET_TILING, &set_tiling); |
240 | } while (ret != 0); |
240 | } while (ret != 0); |
241 | return ret == 0; |
241 | return ret == 0; |
242 | } |
242 | } |
243 | 243 | ||
244 | static bool gem_set_caching(int fd, uint32_t handle, int caching) |
244 | static bool gem_set_caching(int fd, uint32_t handle, int caching) |
245 | { |
245 | { |
246 | struct local_i915_gem_caching arg; |
246 | struct local_i915_gem_caching arg; |
247 | 247 | ||
248 | VG_CLEAR(arg); |
248 | VG_CLEAR(arg); |
249 | arg.handle = handle; |
249 | arg.handle = handle; |
250 | arg.caching = caching; |
250 | arg.caching = caching; |
251 | return drmIoctl(fd, LOCAL_IOCTL_I915_GEM_SET_CACHING, &arg) == 0; |
251 | return drmIoctl(fd, LOCAL_IOCTL_I915_GEM_SET_CACHING, &arg) == 0; |
252 | } |
252 | } |
253 | 253 | ||
254 | 254 | ||
255 | 255 | ||
256 | 256 | ||
257 | 257 | ||
258 | static bool __kgem_throttle_retire(struct kgem *kgem, unsigned flags) |
258 | static bool __kgem_throttle_retire(struct kgem *kgem, unsigned flags) |
259 | { |
259 | { |
260 | if (flags & CREATE_NO_RETIRE) { |
260 | if (flags & CREATE_NO_RETIRE) { |
261 | DBG(("%s: not retiring per-request\n", __FUNCTION__)); |
261 | DBG(("%s: not retiring per-request\n", __FUNCTION__)); |
262 | return false; |
262 | return false; |
263 | } |
263 | } |
264 | 264 | ||
265 | if (!kgem->need_retire) { |
265 | if (!kgem->need_retire) { |
266 | DBG(("%s: nothing to retire\n", __FUNCTION__)); |
266 | DBG(("%s: nothing to retire\n", __FUNCTION__)); |
267 | return false; |
267 | return false; |
268 | } |
268 | } |
269 | 269 | ||
270 | if (kgem_retire(kgem)) |
270 | if (kgem_retire(kgem)) |
271 | return true; |
271 | return true; |
272 | 272 | ||
273 | if (flags & CREATE_NO_THROTTLE || !kgem->need_throttle) { |
273 | if (flags & CREATE_NO_THROTTLE || !kgem->need_throttle) { |
274 | DBG(("%s: not throttling\n", __FUNCTION__)); |
274 | DBG(("%s: not throttling\n", __FUNCTION__)); |
275 | return false; |
275 | return false; |
276 | } |
276 | } |
277 | 277 | ||
278 | kgem_throttle(kgem); |
278 | kgem_throttle(kgem); |
279 | return kgem_retire(kgem); |
279 | return kgem_retire(kgem); |
280 | } |
280 | } |
281 | 281 | ||
282 | static void *__kgem_bo_map__gtt(struct kgem *kgem, struct kgem_bo *bo) |
282 | static void *__kgem_bo_map__gtt(struct kgem *kgem, struct kgem_bo *bo) |
283 | { |
283 | { |
284 | struct drm_i915_gem_mmap_gtt mmap_arg; |
284 | struct drm_i915_gem_mmap_gtt mmap_arg; |
285 | void *ptr; |
285 | void *ptr; |
286 | 286 | ||
287 | DBG(("%s(handle=%d, size=%d)\n", __FUNCTION__, |
287 | DBG(("%s(handle=%d, size=%d)\n", __FUNCTION__, |
288 | bo->handle, bytes(bo))); |
288 | bo->handle, bytes(bo))); |
289 | assert(bo->proxy == NULL); |
289 | assert(bo->proxy == NULL); |
290 | assert(!bo->snoop); |
290 | assert(!bo->snoop); |
291 | assert(kgem_bo_can_map(kgem, bo)); |
291 | assert(kgem_bo_can_map(kgem, bo)); |
292 | 292 | ||
293 | retry_gtt: |
293 | retry_gtt: |
294 | VG_CLEAR(mmap_arg); |
294 | VG_CLEAR(mmap_arg); |
295 | mmap_arg.handle = bo->handle; |
295 | mmap_arg.handle = bo->handle; |
296 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MMAP_GTT, &mmap_arg)) { |
296 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MMAP_GTT, &mmap_arg)) { |
297 | 297 | ||
298 | (void)__kgem_throttle_retire(kgem, 0); |
298 | (void)__kgem_throttle_retire(kgem, 0); |
299 | if (kgem_expire_cache(kgem)) |
299 | if (kgem_expire_cache(kgem)) |
300 | goto retry_gtt; |
300 | goto retry_gtt; |
301 | 301 | ||
302 | if (kgem->need_expire) { |
302 | if (kgem->need_expire) { |
303 | kgem_cleanup_cache(kgem); |
303 | kgem_cleanup_cache(kgem); |
304 | goto retry_gtt; |
304 | goto retry_gtt; |
305 | } |
305 | } |
306 | 306 | ||
307 | printf("%s: failed to retrieve GTT offset for handle=%d\n", |
307 | printf("%s: failed to retrieve GTT offset for handle=%d\n", |
308 | __FUNCTION__, bo->handle); |
308 | __FUNCTION__, bo->handle); |
309 | return NULL; |
309 | return NULL; |
310 | } |
310 | } |
311 | 311 | ||
312 | retry_mmap: |
312 | retry_mmap: |
313 | ptr = (void*)(int)mmap_arg.offset; |
313 | ptr = (void*)(int)mmap_arg.offset; |
314 | if (ptr == NULL) { |
314 | if (ptr == NULL) { |
315 | ErrorF("%s: failed to mmap handle=%d, %d bytes, into GTT domain\n", |
315 | ErrorF("%s: failed to mmap handle=%d, %d bytes, into GTT domain\n", |
316 | __FUNCTION__, bo->handle, bytes(bo)); |
316 | __FUNCTION__, bo->handle, bytes(bo)); |
317 | ptr = NULL; |
317 | ptr = NULL; |
318 | } |
318 | } |
319 | 319 | ||
320 | return ptr; |
320 | return ptr; |
321 | } |
321 | } |
322 | 322 | ||
323 | static int __gem_write(int fd, uint32_t handle, |
323 | static int __gem_write(int fd, uint32_t handle, |
324 | int offset, int length, |
324 | int offset, int length, |
325 | const void *src) |
325 | const void *src) |
326 | { |
326 | { |
327 | struct drm_i915_gem_pwrite pwrite; |
327 | struct drm_i915_gem_pwrite pwrite; |
328 | 328 | ||
329 | DBG(("%s(handle=%d, offset=%d, len=%d)\n", __FUNCTION__, |
329 | DBG(("%s(handle=%d, offset=%d, len=%d)\n", __FUNCTION__, |
330 | handle, offset, length)); |
330 | handle, offset, length)); |
331 | 331 | ||
332 | VG_CLEAR(pwrite); |
332 | VG_CLEAR(pwrite); |
333 | pwrite.handle = handle; |
333 | pwrite.handle = handle; |
334 | pwrite.offset = offset; |
334 | pwrite.offset = offset; |
335 | pwrite.size = length; |
335 | pwrite.size = length; |
336 | pwrite.data_ptr = (uintptr_t)src; |
336 | pwrite.data_ptr = (uintptr_t)src; |
337 | return drmIoctl(fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite); |
337 | return drmIoctl(fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite); |
338 | } |
338 | } |
339 | 339 | ||
340 | static int gem_write(int fd, uint32_t handle, |
340 | static int gem_write(int fd, uint32_t handle, |
341 | int offset, int length, |
341 | int offset, int length, |
342 | const void *src) |
342 | const void *src) |
343 | { |
343 | { |
344 | struct drm_i915_gem_pwrite pwrite; |
344 | struct drm_i915_gem_pwrite pwrite; |
345 | 345 | ||
346 | DBG(("%s(handle=%d, offset=%d, len=%d)\n", __FUNCTION__, |
346 | DBG(("%s(handle=%d, offset=%d, len=%d)\n", __FUNCTION__, |
347 | handle, offset, length)); |
347 | handle, offset, length)); |
348 | 348 | ||
349 | VG_CLEAR(pwrite); |
349 | VG_CLEAR(pwrite); |
350 | pwrite.handle = handle; |
350 | pwrite.handle = handle; |
351 | /* align the transfer to cachelines; fortuitously this is safe! */ |
351 | /* align the transfer to cachelines; fortuitously this is safe! */ |
352 | if ((offset | length) & 63) { |
352 | if ((offset | length) & 63) { |
353 | pwrite.offset = offset & ~63; |
353 | pwrite.offset = offset & ~63; |
354 | pwrite.size = ALIGN(offset+length, 64) - pwrite.offset; |
354 | pwrite.size = ALIGN(offset+length, 64) - pwrite.offset; |
355 | pwrite.data_ptr = (uintptr_t)src + pwrite.offset - offset; |
355 | pwrite.data_ptr = (uintptr_t)src + pwrite.offset - offset; |
356 | } else { |
356 | } else { |
357 | pwrite.offset = offset; |
357 | pwrite.offset = offset; |
358 | pwrite.size = length; |
358 | pwrite.size = length; |
359 | pwrite.data_ptr = (uintptr_t)src; |
359 | pwrite.data_ptr = (uintptr_t)src; |
360 | } |
360 | } |
361 | return drmIoctl(fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite); |
361 | return drmIoctl(fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite); |
362 | } |
362 | } |
363 | 363 | ||
364 | 364 | ||
365 | bool __kgem_busy(struct kgem *kgem, int handle) |
365 | bool __kgem_busy(struct kgem *kgem, int handle) |
366 | { |
366 | { |
367 | struct drm_i915_gem_busy busy; |
367 | struct drm_i915_gem_busy busy; |
368 | 368 | ||
369 | VG_CLEAR(busy); |
369 | VG_CLEAR(busy); |
370 | busy.handle = handle; |
370 | busy.handle = handle; |
371 | busy.busy = !kgem->wedged; |
371 | busy.busy = !kgem->wedged; |
372 | (void)drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_BUSY, &busy); |
372 | (void)drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_BUSY, &busy); |
373 | DBG(("%s: handle=%d, busy=%d, wedged=%d\n", |
373 | DBG(("%s: handle=%d, busy=%d, wedged=%d\n", |
374 | __FUNCTION__, handle, busy.busy, kgem->wedged)); |
374 | __FUNCTION__, handle, busy.busy, kgem->wedged)); |
375 | 375 | ||
376 | return busy.busy; |
376 | return busy.busy; |
377 | } |
377 | } |
378 | 378 | ||
379 | static void kgem_bo_retire(struct kgem *kgem, struct kgem_bo *bo) |
379 | static void kgem_bo_retire(struct kgem *kgem, struct kgem_bo *bo) |
380 | { |
380 | { |
381 | DBG(("%s: retiring bo handle=%d (needed flush? %d), rq? %d [busy?=%d]\n", |
381 | DBG(("%s: retiring bo handle=%d (needed flush? %d), rq? %d [busy?=%d]\n", |
382 | __FUNCTION__, bo->handle, bo->needs_flush, bo->rq != NULL, |
382 | __FUNCTION__, bo->handle, bo->needs_flush, bo->rq != NULL, |
383 | __kgem_busy(kgem, bo->handle))); |
383 | __kgem_busy(kgem, bo->handle))); |
384 | assert(bo->exec == NULL); |
384 | assert(bo->exec == NULL); |
385 | assert(list_is_empty(&bo->vma)); |
385 | assert(list_is_empty(&bo->vma)); |
386 | 386 | ||
387 | if (bo->rq) { |
387 | if (bo->rq) { |
388 | if (!__kgem_busy(kgem, bo->handle)) { |
388 | if (!__kgem_busy(kgem, bo->handle)) { |
389 | __kgem_bo_clear_busy(bo); |
389 | __kgem_bo_clear_busy(bo); |
390 | kgem_retire(kgem); |
390 | kgem_retire(kgem); |
391 | } |
391 | } |
392 | } else { |
392 | } else { |
393 | assert(!bo->needs_flush); |
393 | assert(!bo->needs_flush); |
394 | ASSERT_IDLE(kgem, bo->handle); |
394 | ASSERT_IDLE(kgem, bo->handle); |
395 | } |
395 | } |
396 | } |
396 | } |
397 | 397 | ||
398 | bool kgem_bo_write(struct kgem *kgem, struct kgem_bo *bo, |
398 | bool kgem_bo_write(struct kgem *kgem, struct kgem_bo *bo, |
399 | const void *data, int length) |
399 | const void *data, int length) |
400 | { |
400 | { |
401 | assert(bo->refcnt); |
401 | assert(bo->refcnt); |
402 | assert(!bo->purged); |
402 | assert(!bo->purged); |
403 | assert(bo->proxy == NULL); |
403 | assert(bo->proxy == NULL); |
404 | ASSERT_IDLE(kgem, bo->handle); |
404 | ASSERT_IDLE(kgem, bo->handle); |
405 | 405 | ||
406 | assert(length <= bytes(bo)); |
406 | assert(length <= bytes(bo)); |
407 | if (gem_write(kgem->fd, bo->handle, 0, length, data)) |
407 | if (gem_write(kgem->fd, bo->handle, 0, length, data)) |
408 | return false; |
408 | return false; |
409 | 409 | ||
410 | DBG(("%s: flush=%d, domain=%d\n", __FUNCTION__, bo->flush, bo->domain)); |
410 | DBG(("%s: flush=%d, domain=%d\n", __FUNCTION__, bo->flush, bo->domain)); |
411 | if (bo->exec == NULL) { |
411 | if (bo->exec == NULL) { |
412 | kgem_bo_retire(kgem, bo); |
412 | kgem_bo_retire(kgem, bo); |
413 | bo->domain = DOMAIN_NONE; |
413 | bo->domain = DOMAIN_NONE; |
414 | } |
414 | } |
415 | bo->gtt_dirty = true; |
415 | bo->gtt_dirty = true; |
416 | return true; |
416 | return true; |
417 | } |
417 | } |
418 | 418 | ||
419 | static uint32_t gem_create(int fd, int num_pages) |
419 | static uint32_t gem_create(int fd, int num_pages) |
420 | { |
420 | { |
421 | struct drm_i915_gem_create create; |
421 | struct drm_i915_gem_create create; |
422 | 422 | ||
423 | VG_CLEAR(create); |
423 | VG_CLEAR(create); |
424 | create.handle = 0; |
424 | create.handle = 0; |
425 | create.size = PAGE_SIZE * num_pages; |
425 | create.size = PAGE_SIZE * num_pages; |
426 | (void)drmIoctl(fd, DRM_IOCTL_I915_GEM_CREATE, &create); |
426 | (void)drmIoctl(fd, DRM_IOCTL_I915_GEM_CREATE, &create); |
427 | 427 | ||
428 | return create.handle; |
428 | return create.handle; |
429 | } |
429 | } |
430 | 430 | ||
431 | static bool |
431 | static bool |
432 | kgem_bo_set_purgeable(struct kgem *kgem, struct kgem_bo *bo) |
432 | kgem_bo_set_purgeable(struct kgem *kgem, struct kgem_bo *bo) |
433 | { |
433 | { |
434 | #if DBG_NO_MADV |
434 | #if DBG_NO_MADV |
435 | return true; |
435 | return true; |
436 | #else |
436 | #else |
437 | struct drm_i915_gem_madvise madv; |
437 | struct drm_i915_gem_madvise madv; |
438 | 438 | ||
439 | assert(bo->exec == NULL); |
439 | assert(bo->exec == NULL); |
440 | assert(!bo->purged); |
440 | assert(!bo->purged); |
441 | 441 | ||
442 | VG_CLEAR(madv); |
442 | VG_CLEAR(madv); |
443 | madv.handle = bo->handle; |
443 | madv.handle = bo->handle; |
444 | madv.madv = I915_MADV_DONTNEED; |
444 | madv.madv = I915_MADV_DONTNEED; |
445 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0) { |
445 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0) { |
446 | bo->purged = 1; |
446 | bo->purged = 1; |
447 | kgem->need_purge |= !madv.retained && bo->domain == DOMAIN_GPU; |
447 | kgem->need_purge |= !madv.retained && bo->domain == DOMAIN_GPU; |
448 | return madv.retained; |
448 | return madv.retained; |
449 | } |
449 | } |
450 | 450 | ||
451 | return true; |
451 | return true; |
452 | #endif |
452 | #endif |
453 | } |
453 | } |
454 | 454 | ||
455 | static bool |
455 | static bool |
456 | kgem_bo_is_retained(struct kgem *kgem, struct kgem_bo *bo) |
456 | kgem_bo_is_retained(struct kgem *kgem, struct kgem_bo *bo) |
457 | { |
457 | { |
458 | #if DBG_NO_MADV |
458 | #if DBG_NO_MADV |
459 | return true; |
459 | return true; |
460 | #else |
460 | #else |
461 | struct drm_i915_gem_madvise madv; |
461 | struct drm_i915_gem_madvise madv; |
462 | 462 | ||
463 | if (!bo->purged) |
463 | if (!bo->purged) |
464 | return true; |
464 | return true; |
465 | 465 | ||
466 | VG_CLEAR(madv); |
466 | VG_CLEAR(madv); |
467 | madv.handle = bo->handle; |
467 | madv.handle = bo->handle; |
468 | madv.madv = I915_MADV_DONTNEED; |
468 | madv.madv = I915_MADV_DONTNEED; |
469 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0) |
469 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0) |
470 | return madv.retained; |
470 | return madv.retained; |
471 | 471 | ||
472 | return false; |
472 | return false; |
473 | #endif |
473 | #endif |
474 | } |
474 | } |
475 | 475 | ||
476 | static bool |
476 | static bool |
477 | kgem_bo_clear_purgeable(struct kgem *kgem, struct kgem_bo *bo) |
477 | kgem_bo_clear_purgeable(struct kgem *kgem, struct kgem_bo *bo) |
478 | { |
478 | { |
479 | #if DBG_NO_MADV |
479 | #if DBG_NO_MADV |
480 | return true; |
480 | return true; |
481 | #else |
481 | #else |
482 | struct drm_i915_gem_madvise madv; |
482 | struct drm_i915_gem_madvise madv; |
483 | 483 | ||
484 | assert(bo->purged); |
484 | assert(bo->purged); |
485 | 485 | ||
486 | VG_CLEAR(madv); |
486 | VG_CLEAR(madv); |
487 | madv.handle = bo->handle; |
487 | madv.handle = bo->handle; |
488 | madv.madv = I915_MADV_WILLNEED; |
488 | madv.madv = I915_MADV_WILLNEED; |
489 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0) { |
489 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0) { |
490 | bo->purged = !madv.retained; |
490 | bo->purged = !madv.retained; |
491 | kgem->need_purge |= !madv.retained && bo->domain == DOMAIN_GPU; |
491 | kgem->need_purge |= !madv.retained && bo->domain == DOMAIN_GPU; |
492 | return madv.retained; |
492 | return madv.retained; |
493 | } |
493 | } |
494 | 494 | ||
495 | return false; |
495 | return false; |
496 | #endif |
496 | #endif |
497 | } |
497 | } |
498 | 498 | ||
499 | static void gem_close(int fd, uint32_t handle) |
499 | static void gem_close(int fd, uint32_t handle) |
500 | { |
500 | { |
501 | struct drm_gem_close close; |
501 | struct drm_gem_close close; |
502 | 502 | ||
503 | VG_CLEAR(close); |
503 | VG_CLEAR(close); |
504 | close.handle = handle; |
504 | close.handle = handle; |
505 | (void)drmIoctl(fd, DRM_IOCTL_GEM_CLOSE, &close); |
505 | (void)drmIoctl(fd, DRM_IOCTL_GEM_CLOSE, &close); |
506 | } |
506 | } |
507 | 507 | ||
508 | constant inline static unsigned long __fls(unsigned long word) |
508 | constant inline static unsigned long __fls(unsigned long word) |
509 | { |
509 | { |
510 | #if defined(__GNUC__) && (defined(__i386__) || defined(__x86__) || defined(__x86_64__)) |
510 | #if defined(__GNUC__) && (defined(__i386__) || defined(__x86__) || defined(__x86_64__)) |
511 | asm("bsr %1,%0" |
511 | asm("bsr %1,%0" |
512 | : "=r" (word) |
512 | : "=r" (word) |
513 | : "rm" (word)); |
513 | : "rm" (word)); |
514 | return word; |
514 | return word; |
515 | #else |
515 | #else |
516 | unsigned int v = 0; |
516 | unsigned int v = 0; |
517 | 517 | ||
518 | while (word >>= 1) |
518 | while (word >>= 1) |
519 | v++; |
519 | v++; |
520 | 520 | ||
521 | return v; |
521 | return v; |
522 | #endif |
522 | #endif |
523 | } |
523 | } |
524 | 524 | ||
525 | constant inline static int cache_bucket(int num_pages) |
525 | constant inline static int cache_bucket(int num_pages) |
526 | { |
526 | { |
527 | return __fls(num_pages); |
527 | return __fls(num_pages); |
528 | } |
528 | } |
529 | 529 | ||
530 | static struct kgem_bo *__kgem_bo_init(struct kgem_bo *bo, |
530 | static struct kgem_bo *__kgem_bo_init(struct kgem_bo *bo, |
531 | int handle, int num_pages) |
531 | int handle, int num_pages) |
532 | { |
532 | { |
533 | assert(num_pages); |
533 | assert(num_pages); |
534 | memset(bo, 0, sizeof(*bo)); |
534 | memset(bo, 0, sizeof(*bo)); |
535 | 535 | ||
536 | bo->refcnt = 1; |
536 | bo->refcnt = 1; |
537 | bo->handle = handle; |
537 | bo->handle = handle; |
538 | bo->target_handle = -1; |
538 | bo->target_handle = -1; |
539 | num_pages(bo) = num_pages; |
539 | num_pages(bo) = num_pages; |
540 | bucket(bo) = cache_bucket(num_pages); |
540 | bucket(bo) = cache_bucket(num_pages); |
541 | bo->reusable = true; |
541 | bo->reusable = true; |
542 | bo->domain = DOMAIN_CPU; |
542 | bo->domain = DOMAIN_CPU; |
543 | list_init(&bo->request); |
543 | list_init(&bo->request); |
544 | list_init(&bo->list); |
544 | list_init(&bo->list); |
545 | list_init(&bo->vma); |
545 | list_init(&bo->vma); |
546 | 546 | ||
547 | return bo; |
547 | return bo; |
548 | } |
548 | } |
549 | 549 | ||
550 | static struct kgem_bo *__kgem_bo_alloc(int handle, int num_pages) |
550 | static struct kgem_bo *__kgem_bo_alloc(int handle, int num_pages) |
551 | { |
551 | { |
552 | struct kgem_bo *bo; |
552 | struct kgem_bo *bo; |
553 | 553 | ||
554 | if (__kgem_freed_bo) { |
554 | if (__kgem_freed_bo) { |
555 | bo = __kgem_freed_bo; |
555 | bo = __kgem_freed_bo; |
556 | __kgem_freed_bo = *(struct kgem_bo **)bo; |
556 | __kgem_freed_bo = *(struct kgem_bo **)bo; |
557 | } else { |
557 | } else { |
558 | bo = malloc(sizeof(*bo)); |
558 | bo = malloc(sizeof(*bo)); |
559 | if (bo == NULL) |
559 | if (bo == NULL) |
560 | return NULL; |
560 | return NULL; |
561 | } |
561 | } |
562 | 562 | ||
563 | return __kgem_bo_init(bo, handle, num_pages); |
563 | return __kgem_bo_init(bo, handle, num_pages); |
564 | } |
564 | } |
565 | 565 | ||
566 | static struct kgem_request *__kgem_request_alloc(struct kgem *kgem) |
566 | static struct kgem_request *__kgem_request_alloc(struct kgem *kgem) |
567 | { |
567 | { |
568 | struct kgem_request *rq; |
568 | struct kgem_request *rq; |
569 | 569 | ||
570 | rq = __kgem_freed_request; |
570 | rq = __kgem_freed_request; |
571 | if (rq) { |
571 | if (rq) { |
572 | __kgem_freed_request = *(struct kgem_request **)rq; |
572 | __kgem_freed_request = *(struct kgem_request **)rq; |
573 | } else { |
573 | } else { |
574 | rq = malloc(sizeof(*rq)); |
574 | rq = malloc(sizeof(*rq)); |
575 | if (rq == NULL) |
575 | if (rq == NULL) |
576 | rq = &kgem->static_request; |
576 | rq = &kgem->static_request; |
577 | } |
577 | } |
578 | 578 | ||
579 | list_init(&rq->buffers); |
579 | list_init(&rq->buffers); |
580 | rq->bo = NULL; |
580 | rq->bo = NULL; |
581 | rq->ring = 0; |
581 | rq->ring = 0; |
582 | 582 | ||
583 | return rq; |
583 | return rq; |
584 | } |
584 | } |
585 | 585 | ||
586 | static void __kgem_request_free(struct kgem_request *rq) |
586 | static void __kgem_request_free(struct kgem_request *rq) |
587 | { |
587 | { |
588 | _list_del(&rq->list); |
588 | _list_del(&rq->list); |
589 | *(struct kgem_request **)rq = __kgem_freed_request; |
589 | *(struct kgem_request **)rq = __kgem_freed_request; |
590 | __kgem_freed_request = rq; |
590 | __kgem_freed_request = rq; |
591 | } |
591 | } |
592 | 592 | ||
593 | static struct list *inactive(struct kgem *kgem, int num_pages) |
593 | static struct list *inactive(struct kgem *kgem, int num_pages) |
594 | { |
594 | { |
595 | assert(num_pages < MAX_CACHE_SIZE / PAGE_SIZE); |
595 | assert(num_pages < MAX_CACHE_SIZE / PAGE_SIZE); |
596 | assert(cache_bucket(num_pages) < NUM_CACHE_BUCKETS); |
596 | assert(cache_bucket(num_pages) < NUM_CACHE_BUCKETS); |
597 | return &kgem->inactive[cache_bucket(num_pages)]; |
597 | return &kgem->inactive[cache_bucket(num_pages)]; |
598 | } |
598 | } |
599 | 599 | ||
600 | static struct list *active(struct kgem *kgem, int num_pages, int tiling) |
600 | static struct list *active(struct kgem *kgem, int num_pages, int tiling) |
601 | { |
601 | { |
602 | assert(num_pages < MAX_CACHE_SIZE / PAGE_SIZE); |
602 | assert(num_pages < MAX_CACHE_SIZE / PAGE_SIZE); |
603 | assert(cache_bucket(num_pages) < NUM_CACHE_BUCKETS); |
603 | assert(cache_bucket(num_pages) < NUM_CACHE_BUCKETS); |
604 | return &kgem->active[cache_bucket(num_pages)][tiling]; |
604 | return &kgem->active[cache_bucket(num_pages)][tiling]; |
605 | } |
605 | } |
606 | 606 | ||
607 | static size_t |
607 | static size_t |
608 | agp_aperture_size(struct pci_device *dev, unsigned gen) |
608 | agp_aperture_size(struct pci_device *dev, unsigned gen) |
609 | { |
609 | { |
610 | /* XXX assume that only future chipsets are unknown and follow |
610 | /* XXX assume that only future chipsets are unknown and follow |
611 | * the post gen2 PCI layout. |
611 | * the post gen2 PCI layout. |
612 | */ |
612 | */ |
613 | return 0; |
613 | return 0; |
614 | } |
614 | } |
615 | 615 | ||
616 | static size_t |
616 | static size_t |
617 | total_ram_size(void) |
617 | total_ram_size(void) |
618 | { |
618 | { |
619 | uint32_t data[9]; |
619 | uint32_t data[9]; |
620 | size_t size = 0; |
620 | size_t size = 0; |
621 | 621 | ||
622 | asm volatile("int $0x40" |
622 | asm volatile("int $0x40" |
623 | : "=a" (size) |
623 | : "=a" (size) |
624 | : "a" (18),"b"(20), "c" (data) |
624 | : "a" (18),"b"(20), "c" (data) |
625 | : "memory"); |
625 | : "memory"); |
626 | 626 | ||
627 | return size != -1 ? size : 0; |
627 | return size != -1 ? size : 0; |
628 | } |
628 | } |
629 | 629 | ||
630 | static unsigned |
630 | static unsigned |
631 | cpu_cache_size__cpuid4(void) |
631 | cpu_cache_size__cpuid4(void) |
632 | { |
632 | { |
633 | /* Deterministic Cache Parmaeters (Function 04h)": |
633 | /* Deterministic Cache Parmaeters (Function 04h)": |
634 | * When EAX is initialized to a value of 4, the CPUID instruction |
634 | * When EAX is initialized to a value of 4, the CPUID instruction |
635 | * returns deterministic cache information in the EAX, EBX, ECX |
635 | * returns deterministic cache information in the EAX, EBX, ECX |
636 | * and EDX registers. This function requires ECX be initialized |
636 | * and EDX registers. This function requires ECX be initialized |
637 | * with an index which indicates which cache to return information |
637 | * with an index which indicates which cache to return information |
638 | * about. The OS is expected to call this function (CPUID.4) with |
638 | * about. The OS is expected to call this function (CPUID.4) with |
639 | * ECX = 0, 1, 2, until EAX[4:0] == 0, indicating no more caches. |
639 | * ECX = 0, 1, 2, until EAX[4:0] == 0, indicating no more caches. |
640 | * The order in which the caches are returned is not specified |
640 | * The order in which the caches are returned is not specified |
641 | * and may change at Intel's discretion. |
641 | * and may change at Intel's discretion. |
642 | * |
642 | * |
643 | * Calculating the Cache Size in bytes: |
643 | * Calculating the Cache Size in bytes: |
644 | * = (Ways +1) * (Partitions +1) * (Line Size +1) * (Sets +1) |
644 | * = (Ways +1) * (Partitions +1) * (Line Size +1) * (Sets +1) |
645 | */ |
645 | */ |
646 | 646 | ||
647 | unsigned int eax, ebx, ecx, edx; |
647 | unsigned int eax, ebx, ecx, edx; |
648 | unsigned int llc_size = 0; |
648 | unsigned int llc_size = 0; |
649 | int cnt = 0; |
649 | int cnt = 0; |
650 | 650 | ||
651 | if (__get_cpuid_max(BASIC_CPUID, NULL) < 4) |
651 | if (__get_cpuid_max(BASIC_CPUID, NULL) < 4) |
652 | return 0; |
652 | return 0; |
653 | 653 | ||
654 | do { |
654 | do { |
655 | unsigned associativity, line_partitions, line_size, sets; |
655 | unsigned associativity, line_partitions, line_size, sets; |
656 | 656 | ||
657 | __cpuid_count(4, cnt++, eax, ebx, ecx, edx); |
657 | __cpuid_count(4, cnt++, eax, ebx, ecx, edx); |
658 | 658 | ||
659 | if ((eax & 0x1f) == 0) |
659 | if ((eax & 0x1f) == 0) |
660 | break; |
660 | break; |
661 | 661 | ||
662 | associativity = ((ebx >> 22) & 0x3ff) + 1; |
662 | associativity = ((ebx >> 22) & 0x3ff) + 1; |
663 | line_partitions = ((ebx >> 12) & 0x3ff) + 1; |
663 | line_partitions = ((ebx >> 12) & 0x3ff) + 1; |
664 | line_size = (ebx & 0xfff) + 1; |
664 | line_size = (ebx & 0xfff) + 1; |
665 | sets = ecx + 1; |
665 | sets = ecx + 1; |
666 | 666 | ||
667 | llc_size = associativity * line_partitions * line_size * sets; |
667 | llc_size = associativity * line_partitions * line_size * sets; |
668 | } while (1); |
668 | } while (1); |
669 | 669 | ||
670 | return llc_size; |
670 | return llc_size; |
671 | } |
671 | } |
672 | 672 | ||
673 | static int gem_param(struct kgem *kgem, int name) |
673 | static int gem_param(struct kgem *kgem, int name) |
674 | { |
674 | { |
675 | drm_i915_getparam_t gp; |
675 | drm_i915_getparam_t gp; |
676 | int v = -1; /* No param uses the sign bit, reserve it for errors */ |
676 | int v = -1; /* No param uses the sign bit, reserve it for errors */ |
677 | 677 | ||
678 | VG_CLEAR(gp); |
678 | VG_CLEAR(gp); |
679 | gp.param = name; |
679 | gp.param = name; |
680 | gp.value = &v; |
680 | gp.value = &v; |
681 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GETPARAM, &gp)) |
681 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GETPARAM, &gp)) |
682 | return -1; |
682 | return -1; |
683 | 683 | ||
684 | VG(VALGRIND_MAKE_MEM_DEFINED(&v, sizeof(v))); |
684 | VG(VALGRIND_MAKE_MEM_DEFINED(&v, sizeof(v))); |
685 | return v; |
685 | return v; |
686 | } |
686 | } |
687 | 687 | ||
688 | static bool test_has_execbuffer2(struct kgem *kgem) |
688 | static bool test_has_execbuffer2(struct kgem *kgem) |
689 | { |
689 | { |
690 | return 1; |
690 | return 1; |
691 | } |
691 | } |
692 | 692 | ||
693 | static bool test_has_no_reloc(struct kgem *kgem) |
693 | static bool test_has_no_reloc(struct kgem *kgem) |
694 | { |
694 | { |
695 | if (DBG_NO_FAST_RELOC) |
695 | if (DBG_NO_FAST_RELOC) |
696 | return false; |
696 | return false; |
697 | 697 | ||
698 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_NO_RELOC) > 0; |
698 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_NO_RELOC) > 0; |
699 | } |
699 | } |
700 | 700 | ||
701 | static bool test_has_handle_lut(struct kgem *kgem) |
701 | static bool test_has_handle_lut(struct kgem *kgem) |
702 | { |
702 | { |
703 | if (DBG_NO_HANDLE_LUT) |
703 | if (DBG_NO_HANDLE_LUT) |
704 | return false; |
704 | return false; |
705 | 705 | ||
706 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_HANDLE_LUT) > 0; |
706 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_HANDLE_LUT) > 0; |
707 | } |
707 | } |
708 | 708 | ||
709 | static bool test_has_wt(struct kgem *kgem) |
709 | static bool test_has_wt(struct kgem *kgem) |
710 | { |
710 | { |
711 | if (DBG_NO_WT) |
711 | if (DBG_NO_WT) |
712 | return false; |
712 | return false; |
713 | 713 | ||
714 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_WT) > 0; |
714 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_WT) > 0; |
715 | } |
715 | } |
716 | 716 | ||
717 | static bool test_has_semaphores_enabled(struct kgem *kgem) |
717 | static bool test_has_semaphores_enabled(struct kgem *kgem) |
718 | { |
718 | { |
719 | bool detected = false; |
719 | bool detected = false; |
720 | int ret; |
720 | int ret; |
721 | 721 | ||
722 | if (DBG_NO_SEMAPHORES) |
722 | if (DBG_NO_SEMAPHORES) |
723 | return false; |
723 | return false; |
724 | 724 | ||
725 | ret = gem_param(kgem, LOCAL_I915_PARAM_HAS_SEMAPHORES); |
725 | ret = gem_param(kgem, LOCAL_I915_PARAM_HAS_SEMAPHORES); |
726 | if (ret != -1) |
726 | if (ret != -1) |
727 | return ret > 0; |
727 | return ret > 0; |
728 | 728 | ||
729 | return detected; |
729 | return detected; |
730 | } |
730 | } |
731 | 731 | ||
732 | static bool __kgem_throttle(struct kgem *kgem) |
732 | static bool __kgem_throttle(struct kgem *kgem) |
733 | { |
733 | { |
734 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_THROTTLE, NULL) == 0) |
734 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_THROTTLE, NULL) == 0) |
735 | return false; |
735 | return false; |
736 | 736 | ||
737 | return errno == EIO; |
737 | return errno == EIO; |
738 | } |
738 | } |
739 | 739 | ||
740 | static bool is_hw_supported(struct kgem *kgem, |
740 | static bool is_hw_supported(struct kgem *kgem, |
741 | struct pci_device *dev) |
741 | struct pci_device *dev) |
742 | { |
742 | { |
743 | if (DBG_NO_HW) |
743 | if (DBG_NO_HW) |
744 | return false; |
744 | return false; |
745 | 745 | ||
746 | if (!test_has_execbuffer2(kgem)) |
746 | if (!test_has_execbuffer2(kgem)) |
747 | return false; |
747 | return false; |
748 | 748 | ||
749 | if (kgem->gen == (unsigned)-1) /* unknown chipset, assume future gen */ |
749 | if (kgem->gen == (unsigned)-1) /* unknown chipset, assume future gen */ |
750 | return kgem->has_blt; |
750 | return kgem->has_blt; |
751 | 751 | ||
752 | /* Although pre-855gm the GMCH is fubar, it works mostly. So |
752 | /* Although pre-855gm the GMCH is fubar, it works mostly. So |
753 | * let the user decide through "NoAccel" whether or not to risk |
753 | * let the user decide through "NoAccel" whether or not to risk |
754 | * hw acceleration. |
754 | * hw acceleration. |
755 | */ |
755 | */ |
756 | 756 | ||
757 | if (kgem->gen == 060 && dev->revision < 8) { |
757 | if (kgem->gen == 060 && dev->revision < 8) { |
758 | /* pre-production SNB with dysfunctional BLT */ |
758 | /* pre-production SNB with dysfunctional BLT */ |
759 | return false; |
759 | return false; |
760 | } |
760 | } |
761 | 761 | ||
762 | if (kgem->gen >= 060) /* Only if the kernel supports the BLT ring */ |
762 | if (kgem->gen >= 060) /* Only if the kernel supports the BLT ring */ |
763 | return kgem->has_blt; |
763 | return kgem->has_blt; |
764 | 764 | ||
765 | return true; |
765 | return true; |
766 | } |
766 | } |
767 | 767 | ||
768 | static bool test_has_relaxed_fencing(struct kgem *kgem) |
768 | static bool test_has_relaxed_fencing(struct kgem *kgem) |
769 | { |
769 | { |
770 | if (kgem->gen < 040) { |
770 | if (kgem->gen < 040) { |
771 | if (DBG_NO_RELAXED_FENCING) |
771 | if (DBG_NO_RELAXED_FENCING) |
772 | return false; |
772 | return false; |
773 | 773 | ||
774 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_RELAXED_FENCING) > 0; |
774 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_RELAXED_FENCING) > 0; |
775 | } else |
775 | } else |
776 | return true; |
776 | return true; |
777 | } |
777 | } |
778 | 778 | ||
779 | static bool test_has_llc(struct kgem *kgem) |
779 | static bool test_has_llc(struct kgem *kgem) |
780 | { |
780 | { |
781 | int has_llc = -1; |
781 | int has_llc = -1; |
782 | 782 | ||
783 | if (DBG_NO_LLC) |
783 | if (DBG_NO_LLC) |
784 | return false; |
784 | return false; |
785 | 785 | ||
786 | #if defined(I915_PARAM_HAS_LLC) /* Expected in libdrm-2.4.31 */ |
786 | #if defined(I915_PARAM_HAS_LLC) /* Expected in libdrm-2.4.31 */ |
787 | has_llc = gem_param(kgem, I915_PARAM_HAS_LLC); |
787 | has_llc = gem_param(kgem, I915_PARAM_HAS_LLC); |
788 | #endif |
788 | #endif |
789 | if (has_llc == -1) { |
789 | if (has_llc == -1) { |
790 | DBG(("%s: no kernel/drm support for HAS_LLC, assuming support for LLC based on GPU generation\n", __FUNCTION__)); |
790 | DBG(("%s: no kernel/drm support for HAS_LLC, assuming support for LLC based on GPU generation\n", __FUNCTION__)); |
791 | has_llc = kgem->gen >= 060; |
791 | has_llc = kgem->gen >= 060; |
792 | } |
792 | } |
793 | 793 | ||
794 | return has_llc; |
794 | return has_llc; |
795 | } |
795 | } |
796 | 796 | ||
797 | static bool test_has_caching(struct kgem *kgem) |
797 | static bool test_has_caching(struct kgem *kgem) |
798 | { |
798 | { |
799 | uint32_t handle; |
799 | uint32_t handle; |
800 | bool ret; |
800 | bool ret; |
801 | 801 | ||
802 | if (DBG_NO_CACHE_LEVEL) |
802 | if (DBG_NO_CACHE_LEVEL) |
803 | return false; |
803 | return false; |
804 | 804 | ||
805 | /* Incoherent blt and sampler hangs the GPU */ |
805 | /* Incoherent blt and sampler hangs the GPU */ |
806 | if (kgem->gen == 040) |
806 | if (kgem->gen == 040) |
807 | return false; |
807 | return false; |
808 | 808 | ||
809 | handle = gem_create(kgem->fd, 1); |
809 | handle = gem_create(kgem->fd, 1); |
810 | if (handle == 0) |
810 | if (handle == 0) |
811 | return false; |
811 | return false; |
812 | 812 | ||
813 | ret = gem_set_caching(kgem->fd, handle, UNCACHED); |
813 | ret = gem_set_caching(kgem->fd, handle, UNCACHED); |
814 | gem_close(kgem->fd, handle); |
814 | gem_close(kgem->fd, handle); |
815 | return ret; |
815 | return ret; |
816 | } |
816 | } |
817 | 817 | ||
818 | static bool test_has_userptr(struct kgem *kgem) |
818 | static bool test_has_userptr(struct kgem *kgem) |
819 | { |
819 | { |
820 | #if defined(USE_USERPTR) |
820 | #if defined(USE_USERPTR) |
821 | uint32_t handle; |
821 | uint32_t handle; |
822 | void *ptr; |
822 | void *ptr; |
823 | 823 | ||
824 | if (DBG_NO_USERPTR) |
824 | if (DBG_NO_USERPTR) |
825 | return false; |
825 | return false; |
826 | 826 | ||
827 | /* Incoherent blt and sampler hangs the GPU */ |
827 | /* Incoherent blt and sampler hangs the GPU */ |
828 | if (kgem->gen == 040) |
828 | if (kgem->gen == 040) |
829 | return false; |
829 | return false; |
830 | 830 | ||
831 | if (posix_memalign(&ptr, PAGE_SIZE, PAGE_SIZE)) |
831 | if (posix_memalign(&ptr, PAGE_SIZE, PAGE_SIZE)) |
832 | return false; |
832 | return false; |
833 | 833 | ||
834 | handle = gem_userptr(kgem->fd, ptr, PAGE_SIZE, false); |
834 | handle = gem_userptr(kgem->fd, ptr, PAGE_SIZE, false); |
835 | gem_close(kgem->fd, handle); |
835 | gem_close(kgem->fd, handle); |
836 | free(ptr); |
836 | free(ptr); |
837 | 837 | ||
838 | return handle != 0; |
838 | return handle != 0; |
839 | #else |
839 | #else |
840 | return false; |
840 | return false; |
841 | #endif |
841 | #endif |
842 | } |
842 | } |
843 | 843 | ||
844 | static bool test_has_create2(struct kgem *kgem) |
844 | static bool test_has_create2(struct kgem *kgem) |
845 | { |
845 | { |
846 | #if defined(USE_CREATE2) |
846 | #if defined(USE_CREATE2) |
847 | struct local_i915_gem_create2 args; |
847 | struct local_i915_gem_create2 args; |
848 | 848 | ||
849 | if (DBG_NO_CREATE2) |
849 | if (DBG_NO_CREATE2) |
850 | return false; |
850 | return false; |
851 | 851 | ||
852 | memset(&args, 0, sizeof(args)); |
852 | memset(&args, 0, sizeof(args)); |
853 | args.size = PAGE_SIZE; |
853 | args.size = PAGE_SIZE; |
854 | args.caching = DISPLAY; |
854 | args.caching = DISPLAY; |
855 | if (drmIoctl(kgem->fd, LOCAL_IOCTL_I915_GEM_CREATE2, &args) == 0) |
855 | if (drmIoctl(kgem->fd, LOCAL_IOCTL_I915_GEM_CREATE2, &args) == 0) |
856 | gem_close(kgem->fd, args.handle); |
856 | gem_close(kgem->fd, args.handle); |
857 | 857 | ||
858 | return args.handle != 0; |
858 | return args.handle != 0; |
859 | #else |
859 | #else |
860 | return false; |
860 | return false; |
861 | #endif |
861 | #endif |
862 | } |
862 | } |
863 | 863 | ||
864 | static bool test_has_secure_batches(struct kgem *kgem) |
864 | static bool test_has_secure_batches(struct kgem *kgem) |
865 | { |
865 | { |
866 | if (DBG_NO_SECURE_BATCHES) |
866 | if (DBG_NO_SECURE_BATCHES) |
867 | return false; |
867 | return false; |
868 | 868 | ||
869 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_SECURE_BATCHES) > 0; |
869 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_SECURE_BATCHES) > 0; |
870 | } |
870 | } |
871 | 871 | ||
872 | static bool test_has_pinned_batches(struct kgem *kgem) |
872 | static bool test_has_pinned_batches(struct kgem *kgem) |
873 | { |
873 | { |
874 | if (DBG_NO_PINNED_BATCHES) |
874 | if (DBG_NO_PINNED_BATCHES) |
875 | return false; |
875 | return false; |
876 | 876 | ||
877 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_PINNED_BATCHES) > 0; |
877 | return gem_param(kgem, LOCAL_I915_PARAM_HAS_PINNED_BATCHES) > 0; |
878 | } |
878 | } |
879 | 879 | ||
880 | 880 | ||
881 | static bool kgem_init_pinned_batches(struct kgem *kgem) |
881 | static bool kgem_init_pinned_batches(struct kgem *kgem) |
882 | { |
882 | { |
883 | int count[2] = { 2, 1 }; |
883 | int count[2] = { 2, 1 }; |
884 | int size[2] = { 1, 2 }; |
884 | int size[2] = { 1, 2 }; |
885 | int n, i; |
885 | int n, i; |
886 | 886 | ||
887 | if (kgem->wedged) |
887 | if (kgem->wedged) |
888 | return true; |
888 | return true; |
889 | 889 | ||
890 | for (n = 0; n < ARRAY_SIZE(count); n++) { |
890 | for (n = 0; n < ARRAY_SIZE(count); n++) { |
891 | for (i = 0; i < count[n]; i++) { |
891 | for (i = 0; i < count[n]; i++) { |
892 | struct drm_i915_gem_pin pin; |
892 | struct drm_i915_gem_pin pin; |
893 | struct kgem_bo *bo; |
893 | struct kgem_bo *bo; |
894 | 894 | ||
895 | VG_CLEAR(pin); |
895 | VG_CLEAR(pin); |
896 | 896 | ||
897 | pin.handle = gem_create(kgem->fd, size[n]); |
897 | pin.handle = gem_create(kgem->fd, size[n]); |
898 | if (pin.handle == 0) |
898 | if (pin.handle == 0) |
899 | goto err; |
899 | goto err; |
900 | 900 | ||
901 | DBG(("%s: new handle=%d, num_pages=%d\n", |
901 | DBG(("%s: new handle=%d, num_pages=%d\n", |
902 | __FUNCTION__, pin.handle, size[n])); |
902 | __FUNCTION__, pin.handle, size[n])); |
903 | 903 | ||
904 | bo = __kgem_bo_alloc(pin.handle, size[n]); |
904 | bo = __kgem_bo_alloc(pin.handle, size[n]); |
905 | if (bo == NULL) { |
905 | if (bo == NULL) { |
906 | gem_close(kgem->fd, pin.handle); |
906 | gem_close(kgem->fd, pin.handle); |
907 | goto err; |
907 | goto err; |
908 | } |
908 | } |
909 | 909 | ||
910 | pin.alignment = 0; |
910 | pin.alignment = 0; |
911 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_PIN, &pin)) { |
911 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_PIN, &pin)) { |
912 | gem_close(kgem->fd, pin.handle); |
912 | gem_close(kgem->fd, pin.handle); |
913 | goto err; |
913 | goto err; |
914 | } |
914 | } |
915 | bo->presumed_offset = pin.offset; |
915 | bo->presumed_offset = pin.offset; |
916 | debug_alloc__bo(kgem, bo); |
916 | debug_alloc__bo(kgem, bo); |
917 | list_add(&bo->list, &kgem->pinned_batches[n]); |
917 | list_add(&bo->list, &kgem->pinned_batches[n]); |
918 | } |
918 | } |
919 | } |
919 | } |
920 | 920 | ||
921 | return true; |
921 | return true; |
922 | 922 | ||
923 | err: |
923 | err: |
924 | for (n = 0; n < ARRAY_SIZE(kgem->pinned_batches); n++) { |
924 | for (n = 0; n < ARRAY_SIZE(kgem->pinned_batches); n++) { |
925 | while (!list_is_empty(&kgem->pinned_batches[n])) { |
925 | while (!list_is_empty(&kgem->pinned_batches[n])) { |
926 | kgem_bo_destroy(kgem, |
926 | kgem_bo_destroy(kgem, |
927 | list_first_entry(&kgem->pinned_batches[n], |
927 | list_first_entry(&kgem->pinned_batches[n], |
928 | struct kgem_bo, list)); |
928 | struct kgem_bo, list)); |
929 | } |
929 | } |
930 | } |
930 | } |
931 | 931 | ||
932 | /* For simplicity populate the lists with a single unpinned bo */ |
932 | /* For simplicity populate the lists with a single unpinned bo */ |
933 | for (n = 0; n < ARRAY_SIZE(count); n++) { |
933 | for (n = 0; n < ARRAY_SIZE(count); n++) { |
934 | struct kgem_bo *bo; |
934 | struct kgem_bo *bo; |
935 | uint32_t handle; |
935 | uint32_t handle; |
936 | 936 | ||
937 | handle = gem_create(kgem->fd, size[n]); |
937 | handle = gem_create(kgem->fd, size[n]); |
938 | if (handle == 0) |
938 | if (handle == 0) |
939 | break; |
939 | break; |
940 | 940 | ||
941 | bo = __kgem_bo_alloc(handle, size[n]); |
941 | bo = __kgem_bo_alloc(handle, size[n]); |
942 | if (bo == NULL) { |
942 | if (bo == NULL) { |
943 | gem_close(kgem->fd, handle); |
943 | gem_close(kgem->fd, handle); |
944 | break; |
944 | break; |
945 | } |
945 | } |
946 | 946 | ||
947 | debug_alloc__bo(kgem, bo); |
947 | debug_alloc__bo(kgem, bo); |
948 | list_add(&bo->list, &kgem->pinned_batches[n]); |
948 | list_add(&bo->list, &kgem->pinned_batches[n]); |
949 | } |
949 | } |
950 | return false; |
950 | return false; |
951 | } |
951 | } |
952 | 952 | ||
953 | void kgem_init(struct kgem *kgem, int fd, struct pci_device *dev, unsigned gen) |
953 | void kgem_init(struct kgem *kgem, int fd, struct pci_device *dev, unsigned gen) |
954 | { |
954 | { |
955 | struct drm_i915_gem_get_aperture aperture; |
955 | struct drm_i915_gem_get_aperture aperture; |
956 | size_t totalram; |
956 | size_t totalram; |
957 | unsigned half_gpu_max; |
957 | unsigned half_gpu_max; |
958 | unsigned int i, j; |
958 | unsigned int i, j; |
959 | 959 | ||
960 | DBG(("%s: fd=%d, gen=%d\n", __FUNCTION__, fd, gen)); |
960 | DBG(("%s: fd=%d, gen=%d\n", __FUNCTION__, fd, gen)); |
961 | 961 | ||
962 | memset(kgem, 0, sizeof(*kgem)); |
962 | memset(kgem, 0, sizeof(*kgem)); |
963 | 963 | ||
964 | kgem->fd = fd; |
964 | kgem->fd = fd; |
965 | kgem->gen = gen; |
965 | kgem->gen = gen; |
966 | 966 | ||
967 | list_init(&kgem->requests[0]); |
967 | list_init(&kgem->requests[0]); |
968 | list_init(&kgem->requests[1]); |
968 | list_init(&kgem->requests[1]); |
969 | list_init(&kgem->batch_buffers); |
969 | list_init(&kgem->batch_buffers); |
970 | list_init(&kgem->active_buffers); |
970 | list_init(&kgem->active_buffers); |
971 | list_init(&kgem->flushing); |
971 | list_init(&kgem->flushing); |
972 | list_init(&kgem->large); |
972 | list_init(&kgem->large); |
973 | list_init(&kgem->large_inactive); |
973 | list_init(&kgem->large_inactive); |
974 | list_init(&kgem->snoop); |
974 | list_init(&kgem->snoop); |
975 | list_init(&kgem->scanout); |
975 | list_init(&kgem->scanout); |
976 | for (i = 0; i < ARRAY_SIZE(kgem->pinned_batches); i++) |
976 | for (i = 0; i < ARRAY_SIZE(kgem->pinned_batches); i++) |
977 | list_init(&kgem->pinned_batches[i]); |
977 | list_init(&kgem->pinned_batches[i]); |
978 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) |
978 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) |
979 | list_init(&kgem->inactive[i]); |
979 | list_init(&kgem->inactive[i]); |
980 | for (i = 0; i < ARRAY_SIZE(kgem->active); i++) { |
980 | for (i = 0; i < ARRAY_SIZE(kgem->active); i++) { |
981 | for (j = 0; j < ARRAY_SIZE(kgem->active[i]); j++) |
981 | for (j = 0; j < ARRAY_SIZE(kgem->active[i]); j++) |
982 | list_init(&kgem->active[i][j]); |
982 | list_init(&kgem->active[i][j]); |
983 | } |
983 | } |
984 | for (i = 0; i < ARRAY_SIZE(kgem->vma); i++) { |
984 | for (i = 0; i < ARRAY_SIZE(kgem->vma); i++) { |
985 | for (j = 0; j < ARRAY_SIZE(kgem->vma[i].inactive); j++) |
985 | for (j = 0; j < ARRAY_SIZE(kgem->vma[i].inactive); j++) |
986 | list_init(&kgem->vma[i].inactive[j]); |
986 | list_init(&kgem->vma[i].inactive[j]); |
987 | } |
987 | } |
988 | kgem->vma[MAP_GTT].count = -MAX_GTT_VMA_CACHE; |
988 | kgem->vma[MAP_GTT].count = -MAX_GTT_VMA_CACHE; |
989 | kgem->vma[MAP_CPU].count = -MAX_CPU_VMA_CACHE; |
989 | kgem->vma[MAP_CPU].count = -MAX_CPU_VMA_CACHE; |
990 | 990 | ||
991 | kgem->has_blt = gem_param(kgem, LOCAL_I915_PARAM_HAS_BLT) > 0; |
991 | kgem->has_blt = gem_param(kgem, LOCAL_I915_PARAM_HAS_BLT) > 0; |
992 | DBG(("%s: has BLT ring? %d\n", __FUNCTION__, |
992 | DBG(("%s: has BLT ring? %d\n", __FUNCTION__, |
993 | kgem->has_blt)); |
993 | kgem->has_blt)); |
994 | 994 | ||
995 | kgem->has_relaxed_delta = |
995 | kgem->has_relaxed_delta = |
996 | gem_param(kgem, LOCAL_I915_PARAM_HAS_RELAXED_DELTA) > 0; |
996 | gem_param(kgem, LOCAL_I915_PARAM_HAS_RELAXED_DELTA) > 0; |
997 | DBG(("%s: has relaxed delta? %d\n", __FUNCTION__, |
997 | DBG(("%s: has relaxed delta? %d\n", __FUNCTION__, |
998 | kgem->has_relaxed_delta)); |
998 | kgem->has_relaxed_delta)); |
999 | 999 | ||
1000 | kgem->has_relaxed_fencing = test_has_relaxed_fencing(kgem); |
1000 | kgem->has_relaxed_fencing = test_has_relaxed_fencing(kgem); |
1001 | DBG(("%s: has relaxed fencing? %d\n", __FUNCTION__, |
1001 | DBG(("%s: has relaxed fencing? %d\n", __FUNCTION__, |
1002 | kgem->has_relaxed_fencing)); |
1002 | kgem->has_relaxed_fencing)); |
1003 | 1003 | ||
1004 | kgem->has_llc = test_has_llc(kgem); |
1004 | kgem->has_llc = test_has_llc(kgem); |
1005 | DBG(("%s: has shared last-level-cache? %d\n", __FUNCTION__, |
1005 | DBG(("%s: has shared last-level-cache? %d\n", __FUNCTION__, |
1006 | kgem->has_llc)); |
1006 | kgem->has_llc)); |
1007 | 1007 | ||
1008 | kgem->has_wt = test_has_wt(kgem); |
1008 | kgem->has_wt = test_has_wt(kgem); |
1009 | DBG(("%s: has write-through caching for scanouts? %d\n", __FUNCTION__, |
1009 | DBG(("%s: has write-through caching for scanouts? %d\n", __FUNCTION__, |
1010 | kgem->has_wt)); |
1010 | kgem->has_wt)); |
1011 | 1011 | ||
1012 | kgem->has_caching = test_has_caching(kgem); |
1012 | kgem->has_caching = test_has_caching(kgem); |
1013 | DBG(("%s: has set-cache-level? %d\n", __FUNCTION__, |
1013 | DBG(("%s: has set-cache-level? %d\n", __FUNCTION__, |
1014 | kgem->has_caching)); |
1014 | kgem->has_caching)); |
1015 | 1015 | ||
1016 | kgem->has_userptr = test_has_userptr(kgem); |
1016 | kgem->has_userptr = test_has_userptr(kgem); |
1017 | DBG(("%s: has userptr? %d\n", __FUNCTION__, |
1017 | DBG(("%s: has userptr? %d\n", __FUNCTION__, |
1018 | kgem->has_userptr)); |
1018 | kgem->has_userptr)); |
1019 | 1019 | ||
1020 | kgem->has_create2 = test_has_create2(kgem); |
1020 | kgem->has_create2 = test_has_create2(kgem); |
1021 | kgem->has_create2 = 0; |
1021 | kgem->has_create2 = 0; |
1022 | DBG(("%s: has create2? %d\n", __FUNCTION__, |
1022 | DBG(("%s: has create2? %d\n", __FUNCTION__, |
1023 | kgem->has_create2)); |
1023 | kgem->has_create2)); |
1024 | 1024 | ||
1025 | kgem->has_no_reloc = test_has_no_reloc(kgem); |
1025 | kgem->has_no_reloc = test_has_no_reloc(kgem); |
1026 | DBG(("%s: has no-reloc? %d\n", __FUNCTION__, |
1026 | DBG(("%s: has no-reloc? %d\n", __FUNCTION__, |
1027 | kgem->has_no_reloc)); |
1027 | kgem->has_no_reloc)); |
1028 | 1028 | ||
1029 | kgem->has_handle_lut = test_has_handle_lut(kgem); |
1029 | kgem->has_handle_lut = test_has_handle_lut(kgem); |
1030 | kgem->has_handle_lut = 0; |
1030 | kgem->has_handle_lut = 0; |
1031 | DBG(("%s: has handle-lut? %d\n", __FUNCTION__, |
1031 | DBG(("%s: has handle-lut? %d\n", __FUNCTION__, |
1032 | kgem->has_handle_lut)); |
1032 | kgem->has_handle_lut)); |
1033 | 1033 | ||
1034 | kgem->has_semaphores = false; |
1034 | kgem->has_semaphores = false; |
1035 | if (kgem->has_blt && test_has_semaphores_enabled(kgem)) |
1035 | if (kgem->has_blt && test_has_semaphores_enabled(kgem)) |
1036 | kgem->has_semaphores = true; |
1036 | kgem->has_semaphores = true; |
1037 | DBG(("%s: semaphores enabled? %d\n", __FUNCTION__, |
1037 | DBG(("%s: semaphores enabled? %d\n", __FUNCTION__, |
1038 | kgem->has_semaphores)); |
1038 | kgem->has_semaphores)); |
1039 | 1039 | ||
1040 | kgem->can_blt_cpu = gen >= 030; |
1040 | kgem->can_blt_cpu = gen >= 030; |
1041 | DBG(("%s: can blt to cpu? %d\n", __FUNCTION__, |
1041 | DBG(("%s: can blt to cpu? %d\n", __FUNCTION__, |
1042 | kgem->can_blt_cpu)); |
1042 | kgem->can_blt_cpu)); |
1043 | 1043 | ||
1044 | kgem->has_secure_batches = test_has_secure_batches(kgem); |
1044 | kgem->has_secure_batches = test_has_secure_batches(kgem); |
1045 | DBG(("%s: can use privileged batchbuffers? %d\n", __FUNCTION__, |
1045 | DBG(("%s: can use privileged batchbuffers? %d\n", __FUNCTION__, |
1046 | kgem->has_secure_batches)); |
1046 | kgem->has_secure_batches)); |
1047 | 1047 | ||
1048 | kgem->has_pinned_batches = test_has_pinned_batches(kgem); |
1048 | kgem->has_pinned_batches = test_has_pinned_batches(kgem); |
1049 | DBG(("%s: can use pinned batchbuffers (to avoid CS w/a)? %d\n", __FUNCTION__, |
1049 | DBG(("%s: can use pinned batchbuffers (to avoid CS w/a)? %d\n", __FUNCTION__, |
1050 | kgem->has_pinned_batches)); |
1050 | kgem->has_pinned_batches)); |
1051 | 1051 | ||
1052 | if (!is_hw_supported(kgem, dev)) { |
1052 | if (!is_hw_supported(kgem, dev)) { |
1053 | printf("Detected unsupported/dysfunctional hardware, disabling acceleration.\n"); |
1053 | printf("Detected unsupported/dysfunctional hardware, disabling acceleration.\n"); |
1054 | kgem->wedged = 1; |
1054 | kgem->wedged = 1; |
1055 | } else if (__kgem_throttle(kgem)) { |
1055 | } else if (__kgem_throttle(kgem)) { |
1056 | printf("Detected a hung GPU, disabling acceleration.\n"); |
1056 | printf("Detected a hung GPU, disabling acceleration.\n"); |
1057 | kgem->wedged = 1; |
1057 | kgem->wedged = 1; |
1058 | } |
1058 | } |
1059 | 1059 | ||
1060 | kgem->batch_size = ARRAY_SIZE(kgem->batch); |
1060 | kgem->batch_size = ARRAY_SIZE(kgem->batch); |
1061 | if (gen == 020 && !kgem->has_pinned_batches) |
1061 | if (gen == 020 && !kgem->has_pinned_batches) |
1062 | /* Limited to what we can pin */ |
1062 | /* Limited to what we can pin */ |
1063 | kgem->batch_size = 4*1024; |
1063 | kgem->batch_size = 4*1024; |
1064 | if (gen == 022) |
1064 | if (gen == 022) |
1065 | /* 865g cannot handle a batch spanning multiple pages */ |
1065 | /* 865g cannot handle a batch spanning multiple pages */ |
1066 | kgem->batch_size = PAGE_SIZE / sizeof(uint32_t); |
1066 | kgem->batch_size = PAGE_SIZE / sizeof(uint32_t); |
1067 | if ((gen >> 3) == 7) |
1067 | if ((gen >> 3) == 7) |
1068 | kgem->batch_size = 16*1024; |
1068 | kgem->batch_size = 16*1024; |
1069 | if (!kgem->has_relaxed_delta && kgem->batch_size > 4*1024) |
1069 | if (!kgem->has_relaxed_delta && kgem->batch_size > 4*1024) |
1070 | kgem->batch_size = 4*1024; |
1070 | kgem->batch_size = 4*1024; |
1071 | 1071 | ||
1072 | if (!kgem_init_pinned_batches(kgem) && gen == 020) { |
1072 | if (!kgem_init_pinned_batches(kgem) && gen == 020) { |
1073 | printf("Unable to reserve memory for GPU, disabling acceleration.\n"); |
1073 | printf("Unable to reserve memory for GPU, disabling acceleration.\n"); |
1074 | kgem->wedged = 1; |
1074 | kgem->wedged = 1; |
1075 | } |
1075 | } |
1076 | 1076 | ||
1077 | DBG(("%s: maximum batch size? %d\n", __FUNCTION__, |
1077 | DBG(("%s: maximum batch size? %d\n", __FUNCTION__, |
1078 | kgem->batch_size)); |
1078 | kgem->batch_size)); |
1079 | 1079 | ||
1080 | kgem->min_alignment = 4; |
1080 | kgem->min_alignment = 4; |
1081 | if (gen < 040) |
1081 | if (gen < 040) |
1082 | kgem->min_alignment = 64; |
1082 | kgem->min_alignment = 64; |
1083 | 1083 | ||
1084 | kgem->half_cpu_cache_pages = cpu_cache_size() >> 13; |
1084 | kgem->half_cpu_cache_pages = cpu_cache_size() >> 13; |
1085 | DBG(("%s: last-level cache size: %d bytes, threshold in pages: %d\n", |
1085 | DBG(("%s: last-level cache size: %d bytes, threshold in pages: %d\n", |
1086 | __FUNCTION__, cpu_cache_size(), kgem->half_cpu_cache_pages)); |
1086 | __FUNCTION__, cpu_cache_size(), kgem->half_cpu_cache_pages)); |
1087 | 1087 | ||
1088 | kgem->next_request = __kgem_request_alloc(kgem); |
1088 | kgem->next_request = __kgem_request_alloc(kgem); |
1089 | 1089 | ||
1090 | DBG(("%s: cpu bo enabled %d: llc? %d, set-cache-level? %d, userptr? %d\n", __FUNCTION__, |
1090 | DBG(("%s: cpu bo enabled %d: llc? %d, set-cache-level? %d, userptr? %d\n", __FUNCTION__, |
1091 | !DBG_NO_CPU && (kgem->has_llc | kgem->has_userptr | kgem->has_caching), |
1091 | !DBG_NO_CPU && (kgem->has_llc | kgem->has_userptr | kgem->has_caching), |
1092 | kgem->has_llc, kgem->has_caching, kgem->has_userptr)); |
1092 | kgem->has_llc, kgem->has_caching, kgem->has_userptr)); |
1093 | 1093 | ||
1094 | VG_CLEAR(aperture); |
1094 | VG_CLEAR(aperture); |
1095 | aperture.aper_size = 0; |
1095 | aperture.aper_size = 0; |
1096 | (void)drmIoctl(fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture); |
1096 | (void)drmIoctl(fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture); |
1097 | if (aperture.aper_size == 0) |
1097 | if (aperture.aper_size == 0) |
1098 | aperture.aper_size = 64*1024*1024; |
1098 | aperture.aper_size = 64*1024*1024; |
1099 | 1099 | ||
1100 | DBG(("%s: aperture size %lld, available now %lld\n", |
1100 | DBG(("%s: aperture size %lld, available now %lld\n", |
1101 | __FUNCTION__, |
1101 | __FUNCTION__, |
1102 | (long long)aperture.aper_size, |
1102 | (long long)aperture.aper_size, |
1103 | (long long)aperture.aper_available_size)); |
1103 | (long long)aperture.aper_available_size)); |
1104 | 1104 | ||
1105 | kgem->aperture_total = aperture.aper_size; |
1105 | kgem->aperture_total = aperture.aper_size; |
1106 | kgem->aperture_high = aperture.aper_size * 3/4; |
1106 | kgem->aperture_high = aperture.aper_size * 3/4; |
1107 | kgem->aperture_low = aperture.aper_size * 1/3; |
1107 | kgem->aperture_low = aperture.aper_size * 1/3; |
1108 | if (gen < 033) { |
1108 | if (gen < 033) { |
1109 | /* Severe alignment penalties */ |
1109 | /* Severe alignment penalties */ |
1110 | kgem->aperture_high /= 2; |
1110 | kgem->aperture_high /= 2; |
1111 | kgem->aperture_low /= 2; |
1111 | kgem->aperture_low /= 2; |
1112 | } |
1112 | } |
1113 | DBG(("%s: aperture low=%d [%d], high=%d [%d]\n", __FUNCTION__, |
1113 | DBG(("%s: aperture low=%d [%d], high=%d [%d]\n", __FUNCTION__, |
1114 | kgem->aperture_low, kgem->aperture_low / (1024*1024), |
1114 | kgem->aperture_low, kgem->aperture_low / (1024*1024), |
1115 | kgem->aperture_high, kgem->aperture_high / (1024*1024))); |
1115 | kgem->aperture_high, kgem->aperture_high / (1024*1024))); |
1116 | 1116 | ||
1117 | kgem->aperture_mappable = agp_aperture_size(dev, gen); |
1117 | kgem->aperture_mappable = agp_aperture_size(dev, gen); |
1118 | if (kgem->aperture_mappable == 0 || |
1118 | if (kgem->aperture_mappable == 0 || |
1119 | kgem->aperture_mappable > aperture.aper_size) |
1119 | kgem->aperture_mappable > aperture.aper_size) |
1120 | kgem->aperture_mappable = aperture.aper_size; |
1120 | kgem->aperture_mappable = aperture.aper_size; |
1121 | DBG(("%s: aperture mappable=%d [%d MiB]\n", __FUNCTION__, |
1121 | DBG(("%s: aperture mappable=%d [%d MiB]\n", __FUNCTION__, |
1122 | kgem->aperture_mappable, kgem->aperture_mappable / (1024*1024))); |
1122 | kgem->aperture_mappable, kgem->aperture_mappable / (1024*1024))); |
1123 | 1123 | ||
1124 | kgem->buffer_size = 64 * 1024; |
1124 | kgem->buffer_size = 64 * 1024; |
1125 | while (kgem->buffer_size < kgem->aperture_mappable >> 10) |
1125 | while (kgem->buffer_size < kgem->aperture_mappable >> 10) |
1126 | kgem->buffer_size *= 2; |
1126 | kgem->buffer_size *= 2; |
1127 | if (kgem->buffer_size >> 12 > kgem->half_cpu_cache_pages) |
1127 | if (kgem->buffer_size >> 12 > kgem->half_cpu_cache_pages) |
1128 | kgem->buffer_size = kgem->half_cpu_cache_pages << 12; |
1128 | kgem->buffer_size = kgem->half_cpu_cache_pages << 12; |
1129 | kgem->buffer_size = 1 << __fls(kgem->buffer_size); |
1129 | kgem->buffer_size = 1 << __fls(kgem->buffer_size); |
1130 | DBG(("%s: buffer size=%d [%d KiB]\n", __FUNCTION__, |
1130 | DBG(("%s: buffer size=%d [%d KiB]\n", __FUNCTION__, |
1131 | kgem->buffer_size, kgem->buffer_size / 1024)); |
1131 | kgem->buffer_size, kgem->buffer_size / 1024)); |
1132 | assert(kgem->buffer_size); |
1132 | assert(kgem->buffer_size); |
1133 | 1133 | ||
1134 | kgem->max_object_size = 3 * (kgem->aperture_high >> 12) << 10; |
1134 | kgem->max_object_size = 3 * (kgem->aperture_high >> 12) << 10; |
1135 | kgem->max_gpu_size = kgem->max_object_size; |
1135 | kgem->max_gpu_size = kgem->max_object_size; |
1136 | if (!kgem->has_llc && kgem->max_gpu_size > MAX_CACHE_SIZE) |
1136 | if (!kgem->has_llc && kgem->max_gpu_size > MAX_CACHE_SIZE) |
1137 | kgem->max_gpu_size = MAX_CACHE_SIZE; |
1137 | kgem->max_gpu_size = MAX_CACHE_SIZE; |
1138 | 1138 | ||
1139 | totalram = total_ram_size(); |
1139 | totalram = total_ram_size(); |
1140 | if (totalram == 0) { |
1140 | if (totalram == 0) { |
1141 | DBG(("%s: total ram size unknown, assuming maximum of total aperture\n", |
1141 | DBG(("%s: total ram size unknown, assuming maximum of total aperture\n", |
1142 | __FUNCTION__)); |
1142 | __FUNCTION__)); |
1143 | totalram = kgem->aperture_total; |
1143 | totalram = kgem->aperture_total; |
1144 | } |
1144 | } |
1145 | DBG(("%s: total ram=%ld\n", __FUNCTION__, (long)totalram)); |
1145 | DBG(("%s: total ram=%ld\n", __FUNCTION__, (long)totalram)); |
1146 | if (kgem->max_object_size > totalram / 2) |
1146 | if (kgem->max_object_size > totalram / 2) |
1147 | kgem->max_object_size = totalram / 2; |
1147 | kgem->max_object_size = totalram / 2; |
1148 | if (kgem->max_gpu_size > totalram / 4) |
1148 | if (kgem->max_gpu_size > totalram / 4) |
1149 | kgem->max_gpu_size = totalram / 4; |
1149 | kgem->max_gpu_size = totalram / 4; |
1150 | 1150 | ||
1151 | kgem->max_cpu_size = kgem->max_object_size; |
1151 | kgem->max_cpu_size = kgem->max_object_size; |
1152 | 1152 | ||
1153 | half_gpu_max = kgem->max_gpu_size / 2; |
1153 | half_gpu_max = kgem->max_gpu_size / 2; |
1154 | kgem->max_copy_tile_size = (MAX_CACHE_SIZE + 1)/2; |
1154 | kgem->max_copy_tile_size = (MAX_CACHE_SIZE + 1)/2; |
1155 | if (kgem->max_copy_tile_size > half_gpu_max) |
1155 | if (kgem->max_copy_tile_size > half_gpu_max) |
1156 | kgem->max_copy_tile_size = half_gpu_max; |
1156 | kgem->max_copy_tile_size = half_gpu_max; |
1157 | 1157 | ||
1158 | if (kgem->has_llc) |
1158 | if (kgem->has_llc) |
1159 | kgem->max_upload_tile_size = kgem->max_copy_tile_size; |
1159 | kgem->max_upload_tile_size = kgem->max_copy_tile_size; |
1160 | else |
1160 | else |
1161 | kgem->max_upload_tile_size = kgem->aperture_mappable / 4; |
1161 | kgem->max_upload_tile_size = kgem->aperture_mappable / 4; |
1162 | if (kgem->max_upload_tile_size > half_gpu_max) |
1162 | if (kgem->max_upload_tile_size > half_gpu_max) |
1163 | kgem->max_upload_tile_size = half_gpu_max; |
1163 | kgem->max_upload_tile_size = half_gpu_max; |
1164 | if (kgem->max_upload_tile_size > kgem->aperture_high/2) |
1164 | if (kgem->max_upload_tile_size > kgem->aperture_high/2) |
1165 | kgem->max_upload_tile_size = kgem->aperture_high/2; |
1165 | kgem->max_upload_tile_size = kgem->aperture_high/2; |
1166 | if (kgem->max_upload_tile_size > kgem->aperture_low) |
1166 | if (kgem->max_upload_tile_size > kgem->aperture_low) |
1167 | kgem->max_upload_tile_size = kgem->aperture_low; |
1167 | kgem->max_upload_tile_size = kgem->aperture_low; |
1168 | if (kgem->max_upload_tile_size < 16*PAGE_SIZE) |
1168 | if (kgem->max_upload_tile_size < 16*PAGE_SIZE) |
1169 | kgem->max_upload_tile_size = 16*PAGE_SIZE; |
1169 | kgem->max_upload_tile_size = 16*PAGE_SIZE; |
1170 | 1170 | ||
1171 | kgem->large_object_size = MAX_CACHE_SIZE; |
1171 | kgem->large_object_size = MAX_CACHE_SIZE; |
1172 | if (kgem->large_object_size > half_gpu_max) |
1172 | if (kgem->large_object_size > half_gpu_max) |
1173 | kgem->large_object_size = half_gpu_max; |
1173 | kgem->large_object_size = half_gpu_max; |
1174 | if (kgem->max_copy_tile_size > kgem->aperture_high/2) |
1174 | if (kgem->max_copy_tile_size > kgem->aperture_high/2) |
1175 | kgem->max_copy_tile_size = kgem->aperture_high/2; |
1175 | kgem->max_copy_tile_size = kgem->aperture_high/2; |
1176 | if (kgem->max_copy_tile_size > kgem->aperture_low) |
1176 | if (kgem->max_copy_tile_size > kgem->aperture_low) |
1177 | kgem->max_copy_tile_size = kgem->aperture_low; |
1177 | kgem->max_copy_tile_size = kgem->aperture_low; |
1178 | if (kgem->max_copy_tile_size < 16*PAGE_SIZE) |
1178 | if (kgem->max_copy_tile_size < 16*PAGE_SIZE) |
1179 | kgem->max_copy_tile_size = 16*PAGE_SIZE; |
1179 | kgem->max_copy_tile_size = 16*PAGE_SIZE; |
1180 | 1180 | ||
1181 | if (kgem->has_llc | kgem->has_caching | kgem->has_userptr) { |
1181 | if (kgem->has_llc | kgem->has_caching | kgem->has_userptr) { |
1182 | if (kgem->large_object_size > kgem->max_cpu_size) |
1182 | if (kgem->large_object_size > kgem->max_cpu_size) |
1183 | kgem->large_object_size = kgem->max_cpu_size; |
1183 | kgem->large_object_size = kgem->max_cpu_size; |
1184 | } else |
1184 | } else |
1185 | kgem->max_cpu_size = 0; |
1185 | kgem->max_cpu_size = 0; |
1186 | if (DBG_NO_CPU) |
1186 | if (DBG_NO_CPU) |
1187 | kgem->max_cpu_size = 0; |
1187 | kgem->max_cpu_size = 0; |
1188 | 1188 | ||
1189 | DBG(("%s: maximum object size=%d\n", |
1189 | DBG(("%s: maximum object size=%d\n", |
1190 | __FUNCTION__, kgem->max_object_size)); |
1190 | __FUNCTION__, kgem->max_object_size)); |
1191 | DBG(("%s: large object thresold=%d\n", |
1191 | DBG(("%s: large object thresold=%d\n", |
1192 | __FUNCTION__, kgem->large_object_size)); |
1192 | __FUNCTION__, kgem->large_object_size)); |
1193 | DBG(("%s: max object sizes (gpu=%d, cpu=%d, tile upload=%d, copy=%d)\n", |
1193 | DBG(("%s: max object sizes (gpu=%d, cpu=%d, tile upload=%d, copy=%d)\n", |
1194 | __FUNCTION__, |
1194 | __FUNCTION__, |
1195 | kgem->max_gpu_size, kgem->max_cpu_size, |
1195 | kgem->max_gpu_size, kgem->max_cpu_size, |
1196 | kgem->max_upload_tile_size, kgem->max_copy_tile_size)); |
1196 | kgem->max_upload_tile_size, kgem->max_copy_tile_size)); |
1197 | 1197 | ||
1198 | /* Convert the aperture thresholds to pages */ |
1198 | /* Convert the aperture thresholds to pages */ |
1199 | kgem->aperture_low /= PAGE_SIZE; |
1199 | kgem->aperture_low /= PAGE_SIZE; |
1200 | kgem->aperture_high /= PAGE_SIZE; |
1200 | kgem->aperture_high /= PAGE_SIZE; |
1201 | 1201 | ||
1202 | kgem->fence_max = gem_param(kgem, I915_PARAM_NUM_FENCES_AVAIL) - 2; |
1202 | kgem->fence_max = gem_param(kgem, I915_PARAM_NUM_FENCES_AVAIL) - 2; |
1203 | if ((int)kgem->fence_max < 0) |
1203 | if ((int)kgem->fence_max < 0) |
1204 | kgem->fence_max = 5; /* minimum safe value for all hw */ |
1204 | kgem->fence_max = 5; /* minimum safe value for all hw */ |
1205 | DBG(("%s: max fences=%d\n", __FUNCTION__, kgem->fence_max)); |
1205 | DBG(("%s: max fences=%d\n", __FUNCTION__, kgem->fence_max)); |
1206 | 1206 | ||
1207 | kgem->batch_flags_base = 0; |
1207 | kgem->batch_flags_base = 0; |
1208 | if (kgem->has_no_reloc) |
1208 | if (kgem->has_no_reloc) |
1209 | kgem->batch_flags_base |= LOCAL_I915_EXEC_NO_RELOC; |
1209 | kgem->batch_flags_base |= LOCAL_I915_EXEC_NO_RELOC; |
1210 | if (kgem->has_handle_lut) |
1210 | if (kgem->has_handle_lut) |
1211 | kgem->batch_flags_base |= LOCAL_I915_EXEC_HANDLE_LUT; |
1211 | kgem->batch_flags_base |= LOCAL_I915_EXEC_HANDLE_LUT; |
1212 | if (kgem->has_pinned_batches) |
1212 | if (kgem->has_pinned_batches) |
1213 | kgem->batch_flags_base |= LOCAL_I915_EXEC_IS_PINNED; |
1213 | kgem->batch_flags_base |= LOCAL_I915_EXEC_IS_PINNED; |
1214 | } |
1214 | } |
1215 | 1215 | ||
1216 | /* XXX hopefully a good approximation */ |
1216 | /* XXX hopefully a good approximation */ |
1217 | uint32_t kgem_get_unique_id(struct kgem *kgem) |
1217 | uint32_t kgem_get_unique_id(struct kgem *kgem) |
1218 | { |
1218 | { |
1219 | uint32_t id; |
1219 | uint32_t id; |
1220 | id = ++kgem->unique_id; |
1220 | id = ++kgem->unique_id; |
1221 | if (id == 0) |
1221 | if (id == 0) |
1222 | id = ++kgem->unique_id; |
1222 | id = ++kgem->unique_id; |
1223 | return id; |
1223 | return id; |
1224 | } |
1224 | } |
1225 | 1225 | ||
1226 | inline static uint32_t kgem_pitch_alignment(struct kgem *kgem, unsigned flags) |
1226 | inline static uint32_t kgem_pitch_alignment(struct kgem *kgem, unsigned flags) |
1227 | { |
1227 | { |
1228 | if (flags & CREATE_PRIME) |
1228 | if (flags & CREATE_PRIME) |
1229 | return 256; |
1229 | return 256; |
1230 | if (flags & CREATE_SCANOUT) |
1230 | if (flags & CREATE_SCANOUT) |
1231 | return 64; |
1231 | return 64; |
1232 | return kgem->min_alignment; |
1232 | return kgem->min_alignment; |
1233 | } |
1233 | } |
1234 | 1234 | ||
1235 | void kgem_get_tile_size(struct kgem *kgem, int tiling, |
1235 | void kgem_get_tile_size(struct kgem *kgem, int tiling, |
1236 | int *tile_width, int *tile_height, int *tile_size) |
1236 | int *tile_width, int *tile_height, int *tile_size) |
1237 | { |
1237 | { |
1238 | if (kgem->gen <= 030) { |
1238 | if (kgem->gen <= 030) { |
1239 | if (tiling) { |
1239 | if (tiling) { |
1240 | if (kgem->gen < 030) { |
1240 | if (kgem->gen < 030) { |
1241 | *tile_width = 128; |
1241 | *tile_width = 128; |
1242 | *tile_height = 16; |
1242 | *tile_height = 16; |
1243 | *tile_size = 2048; |
1243 | *tile_size = 2048; |
1244 | } else { |
1244 | } else { |
1245 | *tile_width = 512; |
1245 | *tile_width = 512; |
1246 | *tile_height = 8; |
1246 | *tile_height = 8; |
1247 | *tile_size = 4096; |
1247 | *tile_size = 4096; |
1248 | } |
1248 | } |
1249 | } else { |
1249 | } else { |
1250 | *tile_width = 1; |
1250 | *tile_width = 1; |
1251 | *tile_height = 1; |
1251 | *tile_height = 1; |
1252 | *tile_size = 1; |
1252 | *tile_size = 1; |
1253 | } |
1253 | } |
1254 | } else switch (tiling) { |
1254 | } else switch (tiling) { |
1255 | default: |
1255 | default: |
1256 | case I915_TILING_NONE: |
1256 | case I915_TILING_NONE: |
1257 | *tile_width = 1; |
1257 | *tile_width = 1; |
1258 | *tile_height = 1; |
1258 | *tile_height = 1; |
1259 | *tile_size = 1; |
1259 | *tile_size = 1; |
1260 | break; |
1260 | break; |
1261 | case I915_TILING_X: |
1261 | case I915_TILING_X: |
1262 | *tile_width = 512; |
1262 | *tile_width = 512; |
1263 | *tile_height = 8; |
1263 | *tile_height = 8; |
1264 | *tile_size = 4096; |
1264 | *tile_size = 4096; |
1265 | break; |
1265 | break; |
1266 | case I915_TILING_Y: |
1266 | case I915_TILING_Y: |
1267 | *tile_width = 128; |
1267 | *tile_width = 128; |
1268 | *tile_height = 32; |
1268 | *tile_height = 32; |
1269 | *tile_size = 4096; |
1269 | *tile_size = 4096; |
1270 | break; |
1270 | break; |
1271 | } |
1271 | } |
1272 | } |
1272 | } |
1273 | 1273 | ||
1274 | uint32_t kgem_surface_size(struct kgem *kgem, |
1274 | uint32_t kgem_surface_size(struct kgem *kgem, |
1275 | bool relaxed_fencing, |
1275 | bool relaxed_fencing, |
1276 | unsigned flags, |
1276 | unsigned flags, |
1277 | uint32_t width, |
1277 | uint32_t width, |
1278 | uint32_t height, |
1278 | uint32_t height, |
1279 | uint32_t bpp, |
1279 | uint32_t bpp, |
1280 | uint32_t tiling, |
1280 | uint32_t tiling, |
1281 | uint32_t *pitch) |
1281 | uint32_t *pitch) |
1282 | { |
1282 | { |
1283 | uint32_t tile_width, tile_height; |
1283 | uint32_t tile_width, tile_height; |
1284 | uint32_t size; |
1284 | uint32_t size; |
1285 | 1285 | ||
1286 | assert(width <= MAXSHORT); |
1286 | assert(width <= MAXSHORT); |
1287 | assert(height <= MAXSHORT); |
1287 | assert(height <= MAXSHORT); |
1288 | assert(bpp >= 8); |
1288 | assert(bpp >= 8); |
1289 | 1289 | ||
1290 | if (kgem->gen <= 030) { |
1290 | if (kgem->gen <= 030) { |
1291 | if (tiling) { |
1291 | if (tiling) { |
1292 | if (kgem->gen < 030) { |
1292 | if (kgem->gen < 030) { |
1293 | tile_width = 128; |
1293 | tile_width = 128; |
1294 | tile_height = 32; |
1294 | tile_height = 32; |
1295 | } else { |
1295 | } else { |
1296 | tile_width = 512; |
1296 | tile_width = 512; |
1297 | tile_height = 16; |
1297 | tile_height = 16; |
1298 | } |
1298 | } |
1299 | } else { |
1299 | } else { |
1300 | tile_width = 2 * bpp >> 3; |
1300 | tile_width = 2 * bpp >> 3; |
1301 | tile_width = ALIGN(tile_width, |
1301 | tile_width = ALIGN(tile_width, |
1302 | kgem_pitch_alignment(kgem, flags)); |
1302 | kgem_pitch_alignment(kgem, flags)); |
1303 | tile_height = 2; |
1303 | tile_height = 2; |
1304 | } |
1304 | } |
1305 | } else switch (tiling) { |
1305 | } else switch (tiling) { |
1306 | default: |
1306 | default: |
1307 | case I915_TILING_NONE: |
1307 | case I915_TILING_NONE: |
1308 | tile_width = 2 * bpp >> 3; |
1308 | tile_width = 2 * bpp >> 3; |
1309 | tile_width = ALIGN(tile_width, |
1309 | tile_width = ALIGN(tile_width, |
1310 | kgem_pitch_alignment(kgem, flags)); |
1310 | kgem_pitch_alignment(kgem, flags)); |
1311 | tile_height = 2; |
1311 | tile_height = 2; |
1312 | break; |
1312 | break; |
1313 | 1313 | ||
1314 | /* XXX align to an even tile row */ |
1314 | /* XXX align to an even tile row */ |
1315 | case I915_TILING_X: |
1315 | case I915_TILING_X: |
1316 | tile_width = 512; |
1316 | tile_width = 512; |
1317 | tile_height = 16; |
1317 | tile_height = 16; |
1318 | break; |
1318 | break; |
1319 | case I915_TILING_Y: |
1319 | case I915_TILING_Y: |
1320 | tile_width = 128; |
1320 | tile_width = 128; |
1321 | tile_height = 64; |
1321 | tile_height = 64; |
1322 | break; |
1322 | break; |
1323 | } |
1323 | } |
1324 | 1324 | ||
1325 | *pitch = ALIGN(width * bpp / 8, tile_width); |
1325 | *pitch = ALIGN(width * bpp / 8, tile_width); |
1326 | height = ALIGN(height, tile_height); |
1326 | height = ALIGN(height, tile_height); |
1327 | if (kgem->gen >= 040) |
1327 | if (kgem->gen >= 040) |
1328 | return PAGE_ALIGN(*pitch * height); |
1328 | return PAGE_ALIGN(*pitch * height); |
1329 | 1329 | ||
1330 | /* If it is too wide for the blitter, don't even bother. */ |
1330 | /* If it is too wide for the blitter, don't even bother. */ |
1331 | if (tiling != I915_TILING_NONE) { |
1331 | if (tiling != I915_TILING_NONE) { |
1332 | if (*pitch > 8192) |
1332 | if (*pitch > 8192) |
1333 | return 0; |
1333 | return 0; |
1334 | 1334 | ||
1335 | for (size = tile_width; size < *pitch; size <<= 1) |
1335 | for (size = tile_width; size < *pitch; size <<= 1) |
1336 | ; |
1336 | ; |
1337 | *pitch = size; |
1337 | *pitch = size; |
1338 | } else { |
1338 | } else { |
1339 | if (*pitch >= 32768) |
1339 | if (*pitch >= 32768) |
1340 | return 0; |
1340 | return 0; |
1341 | } |
1341 | } |
1342 | 1342 | ||
1343 | size = *pitch * height; |
1343 | size = *pitch * height; |
1344 | if (relaxed_fencing || tiling == I915_TILING_NONE) |
1344 | if (relaxed_fencing || tiling == I915_TILING_NONE) |
1345 | return PAGE_ALIGN(size); |
1345 | return PAGE_ALIGN(size); |
1346 | 1346 | ||
1347 | /* We need to allocate a pot fence region for a tiled buffer. */ |
1347 | /* We need to allocate a pot fence region for a tiled buffer. */ |
1348 | if (kgem->gen < 030) |
1348 | if (kgem->gen < 030) |
1349 | tile_width = 512 * 1024; |
1349 | tile_width = 512 * 1024; |
1350 | else |
1350 | else |
1351 | tile_width = 1024 * 1024; |
1351 | tile_width = 1024 * 1024; |
1352 | while (tile_width < size) |
1352 | while (tile_width < size) |
1353 | tile_width *= 2; |
1353 | tile_width *= 2; |
1354 | return tile_width; |
1354 | return tile_width; |
1355 | } |
1355 | } |
1356 | 1356 | ||
1357 | static uint32_t kgem_aligned_height(struct kgem *kgem, |
1357 | static uint32_t kgem_aligned_height(struct kgem *kgem, |
1358 | uint32_t height, uint32_t tiling) |
1358 | uint32_t height, uint32_t tiling) |
1359 | { |
1359 | { |
1360 | uint32_t tile_height; |
1360 | uint32_t tile_height; |
1361 | 1361 | ||
1362 | if (kgem->gen <= 030) { |
1362 | if (kgem->gen <= 030) { |
1363 | tile_height = tiling ? kgem->gen < 030 ? 32 : 16 : 1; |
1363 | tile_height = tiling ? kgem->gen < 030 ? 32 : 16 : 1; |
1364 | } else switch (tiling) { |
1364 | } else switch (tiling) { |
1365 | /* XXX align to an even tile row */ |
1365 | /* XXX align to an even tile row */ |
1366 | default: |
1366 | default: |
1367 | case I915_TILING_NONE: |
1367 | case I915_TILING_NONE: |
1368 | tile_height = 1; |
1368 | tile_height = 1; |
1369 | break; |
1369 | break; |
1370 | case I915_TILING_X: |
1370 | case I915_TILING_X: |
1371 | tile_height = 16; |
1371 | tile_height = 16; |
1372 | break; |
1372 | break; |
1373 | case I915_TILING_Y: |
1373 | case I915_TILING_Y: |
1374 | tile_height = 64; |
1374 | tile_height = 64; |
1375 | break; |
1375 | break; |
1376 | } |
1376 | } |
1377 | 1377 | ||
1378 | return ALIGN(height, tile_height); |
1378 | return ALIGN(height, tile_height); |
1379 | } |
1379 | } |
1380 | 1380 | ||
1381 | static struct drm_i915_gem_exec_object2 * |
1381 | static struct drm_i915_gem_exec_object2 * |
1382 | kgem_add_handle(struct kgem *kgem, struct kgem_bo *bo) |
1382 | kgem_add_handle(struct kgem *kgem, struct kgem_bo *bo) |
1383 | { |
1383 | { |
1384 | struct drm_i915_gem_exec_object2 *exec; |
1384 | struct drm_i915_gem_exec_object2 *exec; |
1385 | 1385 | ||
1386 | DBG(("%s: handle=%d, index=%d\n", |
1386 | DBG(("%s: handle=%d, index=%d\n", |
1387 | __FUNCTION__, bo->handle, kgem->nexec)); |
1387 | __FUNCTION__, bo->handle, kgem->nexec)); |
1388 | 1388 | ||
1389 | assert(kgem->nexec < ARRAY_SIZE(kgem->exec)); |
1389 | assert(kgem->nexec < ARRAY_SIZE(kgem->exec)); |
1390 | bo->target_handle = kgem->has_handle_lut ? kgem->nexec : bo->handle; |
1390 | bo->target_handle = kgem->has_handle_lut ? kgem->nexec : bo->handle; |
1391 | exec = memset(&kgem->exec[kgem->nexec++], 0, sizeof(*exec)); |
1391 | exec = memset(&kgem->exec[kgem->nexec++], 0, sizeof(*exec)); |
1392 | exec->handle = bo->handle; |
1392 | exec->handle = bo->handle; |
1393 | exec->offset = bo->presumed_offset; |
1393 | exec->offset = bo->presumed_offset; |
1394 | 1394 | ||
1395 | kgem->aperture += num_pages(bo); |
1395 | kgem->aperture += num_pages(bo); |
1396 | 1396 | ||
1397 | return exec; |
1397 | return exec; |
1398 | } |
1398 | } |
1399 | 1399 | ||
1400 | static void kgem_add_bo(struct kgem *kgem, struct kgem_bo *bo) |
1400 | static void kgem_add_bo(struct kgem *kgem, struct kgem_bo *bo) |
1401 | { |
1401 | { |
1402 | bo->exec = kgem_add_handle(kgem, bo); |
1402 | bo->exec = kgem_add_handle(kgem, bo); |
1403 | bo->rq = MAKE_REQUEST(kgem->next_request, kgem->ring); |
1403 | bo->rq = MAKE_REQUEST(kgem->next_request, kgem->ring); |
1404 | 1404 | ||
1405 | list_move_tail(&bo->request, &kgem->next_request->buffers); |
1405 | list_move_tail(&bo->request, &kgem->next_request->buffers); |
1406 | 1406 | ||
1407 | /* XXX is it worth working around gcc here? */ |
1407 | /* XXX is it worth working around gcc here? */ |
1408 | kgem->flush |= bo->flush; |
1408 | kgem->flush |= bo->flush; |
1409 | } |
1409 | } |
1410 | 1410 | ||
1411 | static uint32_t kgem_end_batch(struct kgem *kgem) |
1411 | static uint32_t kgem_end_batch(struct kgem *kgem) |
1412 | { |
1412 | { |
1413 | kgem->batch[kgem->nbatch++] = MI_BATCH_BUFFER_END; |
1413 | kgem->batch[kgem->nbatch++] = MI_BATCH_BUFFER_END; |
1414 | if (kgem->nbatch & 1) |
1414 | if (kgem->nbatch & 1) |
1415 | kgem->batch[kgem->nbatch++] = MI_NOOP; |
1415 | kgem->batch[kgem->nbatch++] = MI_NOOP; |
1416 | 1416 | ||
1417 | return kgem->nbatch; |
1417 | return kgem->nbatch; |
1418 | } |
1418 | } |
1419 | 1419 | ||
1420 | static void kgem_fixup_self_relocs(struct kgem *kgem, struct kgem_bo *bo) |
1420 | static void kgem_fixup_self_relocs(struct kgem *kgem, struct kgem_bo *bo) |
1421 | { |
1421 | { |
1422 | int n; |
1422 | int n; |
1423 | 1423 | ||
1424 | assert(kgem->nreloc__self <= 256); |
1424 | assert(kgem->nreloc__self <= 256); |
1425 | if (kgem->nreloc__self == 0) |
1425 | if (kgem->nreloc__self == 0) |
1426 | return; |
1426 | return; |
1427 | 1427 | ||
1428 | for (n = 0; n < kgem->nreloc__self; n++) { |
1428 | for (n = 0; n < kgem->nreloc__self; n++) { |
1429 | int i = kgem->reloc__self[n]; |
1429 | int i = kgem->reloc__self[n]; |
1430 | assert(kgem->reloc[i].target_handle == ~0U); |
1430 | assert(kgem->reloc[i].target_handle == ~0U); |
1431 | kgem->reloc[i].target_handle = bo->target_handle; |
1431 | kgem->reloc[i].target_handle = bo->target_handle; |
1432 | kgem->reloc[i].presumed_offset = bo->presumed_offset; |
1432 | kgem->reloc[i].presumed_offset = bo->presumed_offset; |
1433 | kgem->batch[kgem->reloc[i].offset/sizeof(kgem->batch[0])] = |
1433 | kgem->batch[kgem->reloc[i].offset/sizeof(kgem->batch[0])] = |
1434 | kgem->reloc[i].delta + bo->presumed_offset; |
1434 | kgem->reloc[i].delta + bo->presumed_offset; |
1435 | } |
1435 | } |
1436 | 1436 | ||
1437 | if (n == 256) { |
1437 | if (n == 256) { |
1438 | for (n = kgem->reloc__self[255]; n < kgem->nreloc; n++) { |
1438 | for (n = kgem->reloc__self[255]; n < kgem->nreloc; n++) { |
1439 | if (kgem->reloc[n].target_handle == ~0U) { |
1439 | if (kgem->reloc[n].target_handle == ~0U) { |
1440 | kgem->reloc[n].target_handle = bo->target_handle; |
1440 | kgem->reloc[n].target_handle = bo->target_handle; |
1441 | kgem->reloc[n].presumed_offset = bo->presumed_offset; |
1441 | kgem->reloc[n].presumed_offset = bo->presumed_offset; |
1442 | kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0])] = |
1442 | kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0])] = |
1443 | kgem->reloc[n].delta + bo->presumed_offset; |
1443 | kgem->reloc[n].delta + bo->presumed_offset; |
1444 | } |
1444 | } |
1445 | } |
1445 | } |
1446 | 1446 | ||
1447 | } |
1447 | } |
1448 | 1448 | ||
1449 | } |
1449 | } |
1450 | 1450 | ||
1451 | static void kgem_bo_binding_free(struct kgem *kgem, struct kgem_bo *bo) |
1451 | static void kgem_bo_binding_free(struct kgem *kgem, struct kgem_bo *bo) |
1452 | { |
1452 | { |
1453 | struct kgem_bo_binding *b; |
1453 | struct kgem_bo_binding *b; |
1454 | 1454 | ||
1455 | b = bo->binding.next; |
1455 | b = bo->binding.next; |
1456 | while (b) { |
1456 | while (b) { |
1457 | struct kgem_bo_binding *next = b->next; |
1457 | struct kgem_bo_binding *next = b->next; |
1458 | free (b); |
1458 | free (b); |
1459 | b = next; |
1459 | b = next; |
1460 | } |
1460 | } |
1461 | } |
1461 | } |
1462 | 1462 | ||
1463 | static void kgem_bo_release_map(struct kgem *kgem, struct kgem_bo *bo) |
1463 | static void kgem_bo_release_map(struct kgem *kgem, struct kgem_bo *bo) |
1464 | { |
1464 | { |
1465 | int type = IS_CPU_MAP(bo->map); |
1465 | int type = IS_CPU_MAP(bo->map); |
1466 | 1466 | ||
1467 | assert(!IS_USER_MAP(bo->map)); |
1467 | assert(!IS_USER_MAP(bo->map)); |
1468 | 1468 | ||
1469 | DBG(("%s: releasing %s vma for handle=%d, count=%d\n", |
1469 | DBG(("%s: releasing %s vma for handle=%d, count=%d\n", |
1470 | __FUNCTION__, type ? "CPU" : "GTT", |
1470 | __FUNCTION__, type ? "CPU" : "GTT", |
1471 | bo->handle, kgem->vma[type].count)); |
1471 | bo->handle, kgem->vma[type].count)); |
1472 | 1472 | ||
1473 | VG(if (type) VALGRIND_MAKE_MEM_NOACCESS(MAP(bo->map), bytes(bo))); |
1473 | VG(if (type) VALGRIND_MAKE_MEM_NOACCESS(MAP(bo->map), bytes(bo))); |
1474 | user_free(MAP(bo->map)); |
1474 | user_free(MAP(bo->map)); |
1475 | bo->map = NULL; |
1475 | bo->map = NULL; |
1476 | 1476 | ||
1477 | if (!list_is_empty(&bo->vma)) { |
1477 | if (!list_is_empty(&bo->vma)) { |
1478 | list_del(&bo->vma); |
1478 | list_del(&bo->vma); |
1479 | kgem->vma[type].count--; |
1479 | kgem->vma[type].count--; |
1480 | } |
1480 | } |
1481 | } |
1481 | } |
1482 | 1482 | ||
1483 | static void kgem_bo_free(struct kgem *kgem, struct kgem_bo *bo) |
1483 | static void kgem_bo_free(struct kgem *kgem, struct kgem_bo *bo) |
1484 | { |
1484 | { |
1485 | DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle)); |
1485 | DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle)); |
1486 | assert(bo->refcnt == 0); |
1486 | assert(bo->refcnt == 0); |
1487 | assert(bo->proxy == NULL); |
1487 | assert(bo->proxy == NULL); |
1488 | assert(bo->exec == NULL); |
1488 | assert(bo->exec == NULL); |
1489 | assert(!bo->snoop || bo->rq == NULL); |
1489 | assert(!bo->snoop || bo->rq == NULL); |
1490 | 1490 | ||
1491 | #ifdef DEBUG_MEMORY |
1491 | #ifdef DEBUG_MEMORY |
1492 | kgem->debug_memory.bo_allocs--; |
1492 | kgem->debug_memory.bo_allocs--; |
1493 | kgem->debug_memory.bo_bytes -= bytes(bo); |
1493 | kgem->debug_memory.bo_bytes -= bytes(bo); |
1494 | #endif |
1494 | #endif |
1495 | 1495 | ||
1496 | kgem_bo_binding_free(kgem, bo); |
1496 | kgem_bo_binding_free(kgem, bo); |
1497 | 1497 | ||
1498 | if (IS_USER_MAP(bo->map)) { |
1498 | if (IS_USER_MAP(bo->map)) { |
1499 | assert(bo->rq == NULL); |
1499 | assert(bo->rq == NULL); |
1500 | assert(!__kgem_busy(kgem, bo->handle)); |
1500 | assert(!__kgem_busy(kgem, bo->handle)); |
1501 | assert(MAP(bo->map) != bo || bo->io || bo->flush); |
1501 | assert(MAP(bo->map) != bo || bo->io || bo->flush); |
1502 | if (!(bo->io || bo->flush)) { |
1502 | if (!(bo->io || bo->flush)) { |
1503 | DBG(("%s: freeing snooped base\n", __FUNCTION__)); |
1503 | DBG(("%s: freeing snooped base\n", __FUNCTION__)); |
1504 | assert(bo != MAP(bo->map)); |
1504 | assert(bo != MAP(bo->map)); |
1505 | free(MAP(bo->map)); |
1505 | free(MAP(bo->map)); |
1506 | } |
1506 | } |
1507 | bo->map = NULL; |
1507 | bo->map = NULL; |
1508 | } |
1508 | } |
1509 | if (bo->map) |
1509 | if (bo->map) |
1510 | kgem_bo_release_map(kgem, bo); |
1510 | kgem_bo_release_map(kgem, bo); |
1511 | assert(list_is_empty(&bo->vma)); |
1511 | assert(list_is_empty(&bo->vma)); |
1512 | assert(bo->map == NULL); |
1512 | assert(bo->map == NULL); |
1513 | 1513 | ||
1514 | _list_del(&bo->list); |
1514 | _list_del(&bo->list); |
1515 | _list_del(&bo->request); |
1515 | _list_del(&bo->request); |
1516 | gem_close(kgem->fd, bo->handle); |
1516 | gem_close(kgem->fd, bo->handle); |
1517 | 1517 | ||
1518 | if (!bo->io) { |
1518 | if (!bo->io) { |
1519 | *(struct kgem_bo **)bo = __kgem_freed_bo; |
1519 | *(struct kgem_bo **)bo = __kgem_freed_bo; |
1520 | __kgem_freed_bo = bo; |
1520 | __kgem_freed_bo = bo; |
1521 | } else |
1521 | } else |
1522 | free(bo); |
1522 | free(bo); |
1523 | } |
1523 | } |
1524 | 1524 | ||
1525 | inline static void kgem_bo_move_to_inactive(struct kgem *kgem, |
1525 | inline static void kgem_bo_move_to_inactive(struct kgem *kgem, |
1526 | struct kgem_bo *bo) |
1526 | struct kgem_bo *bo) |
1527 | { |
1527 | { |
1528 | DBG(("%s: moving handle=%d to inactive\n", __FUNCTION__, bo->handle)); |
1528 | DBG(("%s: moving handle=%d to inactive\n", __FUNCTION__, bo->handle)); |
1529 | 1529 | ||
1530 | assert(bo->refcnt == 0); |
1530 | assert(bo->refcnt == 0); |
1531 | assert(bo->reusable); |
1531 | assert(bo->reusable); |
1532 | assert(bo->rq == NULL); |
1532 | assert(bo->rq == NULL); |
1533 | assert(bo->exec == NULL); |
1533 | assert(bo->exec == NULL); |
1534 | assert(bo->domain != DOMAIN_GPU); |
1534 | assert(bo->domain != DOMAIN_GPU); |
1535 | assert(!bo->proxy); |
1535 | assert(!bo->proxy); |
1536 | assert(!bo->io); |
1536 | assert(!bo->io); |
1537 | assert(!bo->scanout); |
1537 | assert(!bo->scanout); |
1538 | assert(!bo->snoop); |
1538 | assert(!bo->snoop); |
1539 | assert(!bo->flush); |
1539 | assert(!bo->flush); |
1540 | assert(!bo->needs_flush); |
1540 | assert(!bo->needs_flush); |
1541 | assert(list_is_empty(&bo->vma)); |
1541 | assert(list_is_empty(&bo->vma)); |
1542 | assert_tiling(kgem, bo); |
1542 | assert_tiling(kgem, bo); |
1543 | ASSERT_IDLE(kgem, bo->handle); |
1543 | ASSERT_IDLE(kgem, bo->handle); |
1544 | 1544 | ||
1545 | kgem->need_expire = true; |
1545 | kgem->need_expire = true; |
1546 | 1546 | ||
1547 | if (bucket(bo) >= NUM_CACHE_BUCKETS) { |
1547 | if (bucket(bo) >= NUM_CACHE_BUCKETS) { |
1548 | list_move(&bo->list, &kgem->large_inactive); |
1548 | list_move(&bo->list, &kgem->large_inactive); |
1549 | return; |
1549 | return; |
1550 | } |
1550 | } |
1551 | 1551 | ||
1552 | assert(bo->flush == false); |
1552 | assert(bo->flush == false); |
1553 | list_move(&bo->list, &kgem->inactive[bucket(bo)]); |
1553 | list_move(&bo->list, &kgem->inactive[bucket(bo)]); |
1554 | if (bo->map) { |
1554 | if (bo->map) { |
1555 | int type = IS_CPU_MAP(bo->map); |
1555 | int type = IS_CPU_MAP(bo->map); |
1556 | if (bucket(bo) >= NUM_CACHE_BUCKETS || |
1556 | if (bucket(bo) >= NUM_CACHE_BUCKETS || |
1557 | (!type && !__kgem_bo_is_mappable(kgem, bo))) { |
1557 | (!type && !__kgem_bo_is_mappable(kgem, bo))) { |
1558 | // munmap(MAP(bo->map), bytes(bo)); |
1558 | // munmap(MAP(bo->map), bytes(bo)); |
1559 | bo->map = NULL; |
1559 | bo->map = NULL; |
1560 | } |
1560 | } |
1561 | if (bo->map) { |
1561 | if (bo->map) { |
1562 | list_add(&bo->vma, &kgem->vma[type].inactive[bucket(bo)]); |
1562 | list_add(&bo->vma, &kgem->vma[type].inactive[bucket(bo)]); |
1563 | kgem->vma[type].count++; |
1563 | kgem->vma[type].count++; |
1564 | } |
1564 | } |
1565 | } |
1565 | } |
1566 | } |
1566 | } |
1567 | 1567 | ||
1568 | static struct kgem_bo *kgem_bo_replace_io(struct kgem_bo *bo) |
1568 | static struct kgem_bo *kgem_bo_replace_io(struct kgem_bo *bo) |
1569 | { |
1569 | { |
1570 | struct kgem_bo *base; |
1570 | struct kgem_bo *base; |
1571 | 1571 | ||
1572 | if (!bo->io) |
1572 | if (!bo->io) |
1573 | return bo; |
1573 | return bo; |
1574 | 1574 | ||
1575 | assert(!bo->snoop); |
1575 | assert(!bo->snoop); |
1576 | base = malloc(sizeof(*base)); |
1576 | base = malloc(sizeof(*base)); |
1577 | if (base) { |
1577 | if (base) { |
1578 | DBG(("%s: transferring io handle=%d to bo\n", |
1578 | DBG(("%s: transferring io handle=%d to bo\n", |
1579 | __FUNCTION__, bo->handle)); |
1579 | __FUNCTION__, bo->handle)); |
1580 | /* transfer the handle to a minimum bo */ |
1580 | /* transfer the handle to a minimum bo */ |
1581 | memcpy(base, bo, sizeof(*base)); |
1581 | memcpy(base, bo, sizeof(*base)); |
1582 | base->io = false; |
1582 | base->io = false; |
1583 | list_init(&base->list); |
1583 | list_init(&base->list); |
1584 | list_replace(&bo->request, &base->request); |
1584 | list_replace(&bo->request, &base->request); |
1585 | list_replace(&bo->vma, &base->vma); |
1585 | list_replace(&bo->vma, &base->vma); |
1586 | free(bo); |
1586 | free(bo); |
1587 | bo = base; |
1587 | bo = base; |
1588 | } else |
1588 | } else |
1589 | bo->reusable = false; |
1589 | bo->reusable = false; |
1590 | 1590 | ||
1591 | return bo; |
1591 | return bo; |
1592 | } |
1592 | } |
1593 | 1593 | ||
1594 | inline static void kgem_bo_remove_from_inactive(struct kgem *kgem, |
1594 | inline static void kgem_bo_remove_from_inactive(struct kgem *kgem, |
1595 | struct kgem_bo *bo) |
1595 | struct kgem_bo *bo) |
1596 | { |
1596 | { |
1597 | DBG(("%s: removing handle=%d from inactive\n", __FUNCTION__, bo->handle)); |
1597 | DBG(("%s: removing handle=%d from inactive\n", __FUNCTION__, bo->handle)); |
1598 | 1598 | ||
1599 | list_del(&bo->list); |
1599 | list_del(&bo->list); |
1600 | assert(bo->rq == NULL); |
1600 | assert(bo->rq == NULL); |
1601 | assert(bo->exec == NULL); |
1601 | assert(bo->exec == NULL); |
1602 | if (bo->map) { |
1602 | if (bo->map) { |
1603 | assert(!list_is_empty(&bo->vma)); |
1603 | assert(!list_is_empty(&bo->vma)); |
1604 | list_del(&bo->vma); |
1604 | list_del(&bo->vma); |
1605 | kgem->vma[IS_CPU_MAP(bo->map)].count--; |
1605 | kgem->vma[IS_CPU_MAP(bo->map)].count--; |
1606 | } |
1606 | } |
1607 | } |
1607 | } |
1608 | 1608 | ||
1609 | inline static void kgem_bo_remove_from_active(struct kgem *kgem, |
1609 | inline static void kgem_bo_remove_from_active(struct kgem *kgem, |
1610 | struct kgem_bo *bo) |
1610 | struct kgem_bo *bo) |
1611 | { |
1611 | { |
1612 | DBG(("%s: removing handle=%d from active\n", __FUNCTION__, bo->handle)); |
1612 | DBG(("%s: removing handle=%d from active\n", __FUNCTION__, bo->handle)); |
1613 | 1613 | ||
1614 | list_del(&bo->list); |
1614 | list_del(&bo->list); |
1615 | assert(bo->rq != NULL); |
1615 | assert(bo->rq != NULL); |
1616 | if (bo->rq == (void *)kgem) |
1616 | if (bo->rq == (void *)kgem) |
1617 | list_del(&bo->request); |
1617 | list_del(&bo->request); |
1618 | assert(list_is_empty(&bo->vma)); |
1618 | assert(list_is_empty(&bo->vma)); |
1619 | } |
1619 | } |
1620 | 1620 | ||
1621 | static void _kgem_bo_delete_buffer(struct kgem *kgem, struct kgem_bo *bo) |
1621 | static void _kgem_bo_delete_buffer(struct kgem *kgem, struct kgem_bo *bo) |
1622 | { |
1622 | { |
1623 | struct kgem_buffer *io = (struct kgem_buffer *)bo->proxy; |
1623 | struct kgem_buffer *io = (struct kgem_buffer *)bo->proxy; |
1624 | 1624 | ||
1625 | DBG(("%s: size=%d, offset=%d, parent used=%d\n", |
1625 | DBG(("%s: size=%d, offset=%d, parent used=%d\n", |
1626 | __FUNCTION__, bo->size.bytes, bo->delta, io->used)); |
1626 | __FUNCTION__, bo->size.bytes, bo->delta, io->used)); |
1627 | 1627 | ||
1628 | if (ALIGN(bo->delta + bo->size.bytes, UPLOAD_ALIGNMENT) == io->used) |
1628 | if (ALIGN(bo->delta + bo->size.bytes, UPLOAD_ALIGNMENT) == io->used) |
1629 | io->used = bo->delta; |
1629 | io->used = bo->delta; |
1630 | } |
1630 | } |
1631 | 1631 | ||
1632 | static void kgem_bo_move_to_scanout(struct kgem *kgem, struct kgem_bo *bo) |
1632 | static void kgem_bo_move_to_scanout(struct kgem *kgem, struct kgem_bo *bo) |
1633 | { |
1633 | { |
1634 | assert(bo->refcnt == 0); |
1634 | assert(bo->refcnt == 0); |
1635 | assert(bo->scanout); |
1635 | assert(bo->scanout); |
1636 | assert(bo->delta); |
1636 | assert(bo->delta); |
1637 | assert(!bo->flush); |
1637 | assert(!bo->flush); |
1638 | assert(!bo->snoop); |
1638 | assert(!bo->snoop); |
1639 | assert(!bo->io); |
1639 | assert(!bo->io); |
1640 | 1640 | ||
1641 | if (bo->purged) { |
1641 | if (bo->purged) { |
1642 | DBG(("%s: discarding purged scanout - external name?\n", |
1642 | DBG(("%s: discarding purged scanout - external name?\n", |
1643 | __FUNCTION__)); |
1643 | __FUNCTION__)); |
1644 | kgem_bo_free(kgem, bo); |
1644 | kgem_bo_free(kgem, bo); |
1645 | return; |
1645 | return; |
1646 | } |
1646 | } |
1647 | 1647 | ||
1648 | DBG(("%s: moving %d [fb %d] to scanout cache, active? %d\n", |
1648 | DBG(("%s: moving %d [fb %d] to scanout cache, active? %d\n", |
1649 | __FUNCTION__, bo->handle, bo->delta, bo->rq != NULL)); |
1649 | __FUNCTION__, bo->handle, bo->delta, bo->rq != NULL)); |
1650 | if (bo->rq) |
1650 | if (bo->rq) |
1651 | list_move_tail(&bo->list, &kgem->scanout); |
1651 | list_move_tail(&bo->list, &kgem->scanout); |
1652 | else |
1652 | else |
1653 | list_move(&bo->list, &kgem->scanout); |
1653 | list_move(&bo->list, &kgem->scanout); |
1654 | } |
1654 | } |
1655 | 1655 | ||
1656 | static void kgem_bo_move_to_snoop(struct kgem *kgem, struct kgem_bo *bo) |
1656 | static void kgem_bo_move_to_snoop(struct kgem *kgem, struct kgem_bo *bo) |
1657 | { |
1657 | { |
1658 | assert(bo->reusable); |
1658 | assert(bo->reusable); |
1659 | assert(!bo->flush); |
1659 | assert(!bo->flush); |
1660 | assert(!bo->needs_flush); |
1660 | assert(!bo->needs_flush); |
1661 | assert(bo->refcnt == 0); |
1661 | assert(bo->refcnt == 0); |
1662 | assert(bo->exec == NULL); |
1662 | assert(bo->exec == NULL); |
1663 | 1663 | ||
1664 | if (num_pages(bo) > kgem->max_cpu_size >> 13) { |
1664 | if (num_pages(bo) > kgem->max_cpu_size >> 13) { |
1665 | DBG(("%s handle=%d discarding large CPU buffer (%d >%d pages)\n", |
1665 | DBG(("%s handle=%d discarding large CPU buffer (%d >%d pages)\n", |
1666 | __FUNCTION__, bo->handle, num_pages(bo), kgem->max_cpu_size >> 13)); |
1666 | __FUNCTION__, bo->handle, num_pages(bo), kgem->max_cpu_size >> 13)); |
1667 | kgem_bo_free(kgem, bo); |
1667 | kgem_bo_free(kgem, bo); |
1668 | return; |
1668 | return; |
1669 | } |
1669 | } |
1670 | 1670 | ||
1671 | assert(bo->tiling == I915_TILING_NONE); |
1671 | assert(bo->tiling == I915_TILING_NONE); |
1672 | assert(bo->rq == NULL); |
1672 | assert(bo->rq == NULL); |
1673 | 1673 | ||
1674 | DBG(("%s: moving %d to snoop cachee\n", __FUNCTION__, bo->handle)); |
1674 | DBG(("%s: moving %d to snoop cachee\n", __FUNCTION__, bo->handle)); |
1675 | list_add(&bo->list, &kgem->snoop); |
1675 | list_add(&bo->list, &kgem->snoop); |
1676 | } |
1676 | } |
1677 | 1677 | ||
1678 | static struct kgem_bo * |
1678 | static struct kgem_bo * |
1679 | search_snoop_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags) |
1679 | search_snoop_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags) |
1680 | { |
1680 | { |
1681 | struct kgem_bo *bo, *first = NULL; |
1681 | struct kgem_bo *bo, *first = NULL; |
1682 | 1682 | ||
1683 | DBG(("%s: num_pages=%d, flags=%x\n", __FUNCTION__, num_pages, flags)); |
1683 | DBG(("%s: num_pages=%d, flags=%x\n", __FUNCTION__, num_pages, flags)); |
1684 | 1684 | ||
1685 | if ((kgem->has_caching | kgem->has_userptr) == 0) |
1685 | if ((kgem->has_caching | kgem->has_userptr) == 0) |
1686 | return NULL; |
1686 | return NULL; |
1687 | 1687 | ||
1688 | if (list_is_empty(&kgem->snoop)) { |
1688 | if (list_is_empty(&kgem->snoop)) { |
1689 | DBG(("%s: inactive and cache empty\n", __FUNCTION__)); |
1689 | DBG(("%s: inactive and cache empty\n", __FUNCTION__)); |
1690 | if (!__kgem_throttle_retire(kgem, flags)) { |
1690 | if (!__kgem_throttle_retire(kgem, flags)) { |
1691 | DBG(("%s: nothing retired\n", __FUNCTION__)); |
1691 | DBG(("%s: nothing retired\n", __FUNCTION__)); |
1692 | return NULL; |
1692 | return NULL; |
1693 | } |
1693 | } |
1694 | } |
1694 | } |
1695 | 1695 | ||
1696 | list_for_each_entry(bo, &kgem->snoop, list) { |
1696 | list_for_each_entry(bo, &kgem->snoop, list) { |
1697 | assert(bo->refcnt == 0); |
1697 | assert(bo->refcnt == 0); |
1698 | assert(bo->snoop); |
1698 | assert(bo->snoop); |
1699 | assert(!bo->scanout); |
1699 | assert(!bo->scanout); |
1700 | assert(!bo->purged); |
1700 | assert(!bo->purged); |
1701 | assert(bo->proxy == NULL); |
1701 | assert(bo->proxy == NULL); |
1702 | assert(bo->tiling == I915_TILING_NONE); |
1702 | assert(bo->tiling == I915_TILING_NONE); |
1703 | assert(bo->rq == NULL); |
1703 | assert(bo->rq == NULL); |
1704 | assert(bo->exec == NULL); |
1704 | assert(bo->exec == NULL); |
1705 | 1705 | ||
1706 | if (num_pages > num_pages(bo)) |
1706 | if (num_pages > num_pages(bo)) |
1707 | continue; |
1707 | continue; |
1708 | 1708 | ||
1709 | if (num_pages(bo) > 2*num_pages) { |
1709 | if (num_pages(bo) > 2*num_pages) { |
1710 | if (first == NULL) |
1710 | if (first == NULL) |
1711 | first = bo; |
1711 | first = bo; |
1712 | continue; |
1712 | continue; |
1713 | } |
1713 | } |
1714 | 1714 | ||
1715 | list_del(&bo->list); |
1715 | list_del(&bo->list); |
1716 | bo->pitch = 0; |
1716 | bo->pitch = 0; |
1717 | bo->delta = 0; |
1717 | bo->delta = 0; |
1718 | 1718 | ||
1719 | DBG((" %s: found handle=%d (num_pages=%d) in snoop cache\n", |
1719 | DBG((" %s: found handle=%d (num_pages=%d) in snoop cache\n", |
1720 | __FUNCTION__, bo->handle, num_pages(bo))); |
1720 | __FUNCTION__, bo->handle, num_pages(bo))); |
1721 | return bo; |
1721 | return bo; |
1722 | } |
1722 | } |
1723 | 1723 | ||
1724 | if (first) { |
1724 | if (first) { |
1725 | list_del(&first->list); |
1725 | list_del(&first->list); |
1726 | first->pitch = 0; |
1726 | first->pitch = 0; |
1727 | first->delta = 0; |
1727 | first->delta = 0; |
1728 | 1728 | ||
1729 | DBG((" %s: found handle=%d (num_pages=%d) in snoop cache\n", |
1729 | DBG((" %s: found handle=%d (num_pages=%d) in snoop cache\n", |
1730 | __FUNCTION__, first->handle, num_pages(first))); |
1730 | __FUNCTION__, first->handle, num_pages(first))); |
1731 | return first; |
1731 | return first; |
1732 | } |
1732 | } |
1733 | 1733 | ||
1734 | return NULL; |
1734 | return NULL; |
1735 | } |
1735 | } |
1736 | 1736 | ||
1737 | void kgem_bo_undo(struct kgem *kgem, struct kgem_bo *bo) |
1737 | void kgem_bo_undo(struct kgem *kgem, struct kgem_bo *bo) |
1738 | { |
1738 | { |
1739 | if (kgem->nexec != 1 || bo->exec == NULL) |
1739 | if (kgem->nexec != 1 || bo->exec == NULL) |
1740 | return; |
1740 | return; |
1741 | 1741 | ||
1742 | DBG(("%s: only handle in batch, discarding last operations for handle=%d\n", |
1742 | DBG(("%s: only handle in batch, discarding last operations for handle=%d\n", |
1743 | __FUNCTION__, bo->handle)); |
1743 | __FUNCTION__, bo->handle)); |
1744 | 1744 | ||
1745 | assert(bo->exec == &kgem->exec[0]); |
1745 | assert(bo->exec == &kgem->exec[0]); |
1746 | assert(kgem->exec[0].handle == bo->handle); |
1746 | assert(kgem->exec[0].handle == bo->handle); |
1747 | assert(RQ(bo->rq) == kgem->next_request); |
1747 | assert(RQ(bo->rq) == kgem->next_request); |
1748 | 1748 | ||
1749 | bo->refcnt++; |
1749 | bo->refcnt++; |
1750 | kgem_reset(kgem); |
1750 | kgem_reset(kgem); |
1751 | bo->refcnt--; |
1751 | bo->refcnt--; |
1752 | } |
1752 | } |
1753 | 1753 | ||
1754 | static void __kgem_bo_destroy(struct kgem *kgem, struct kgem_bo *bo) |
1754 | static void __kgem_bo_destroy(struct kgem *kgem, struct kgem_bo *bo) |
1755 | { |
1755 | { |
1756 | DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle)); |
1756 | DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle)); |
1757 | 1757 | ||
1758 | assert(list_is_empty(&bo->list)); |
1758 | assert(list_is_empty(&bo->list)); |
1759 | assert(bo->refcnt == 0); |
1759 | assert(bo->refcnt == 0); |
1760 | assert(!bo->purged || !bo->reusable); |
1760 | assert(!bo->purged || !bo->reusable); |
1761 | assert(bo->proxy == NULL); |
1761 | assert(bo->proxy == NULL); |
1762 | assert_tiling(kgem, bo); |
1762 | assert_tiling(kgem, bo); |
1763 | 1763 | ||
1764 | bo->binding.offset = 0; |
1764 | bo->binding.offset = 0; |
1765 | 1765 | ||
1766 | if (DBG_NO_CACHE) |
1766 | if (DBG_NO_CACHE) |
1767 | goto destroy; |
1767 | goto destroy; |
1768 | 1768 | ||
1769 | if (bo->snoop && !bo->flush) { |
1769 | if (bo->snoop && !bo->flush) { |
1770 | DBG(("%s: handle=%d is snooped\n", __FUNCTION__, bo->handle)); |
1770 | DBG(("%s: handle=%d is snooped\n", __FUNCTION__, bo->handle)); |
1771 | assert(bo->reusable); |
1771 | assert(bo->reusable); |
1772 | assert(list_is_empty(&bo->list)); |
1772 | assert(list_is_empty(&bo->list)); |
1773 | if (bo->exec == NULL && bo->rq && !__kgem_busy(kgem, bo->handle)) |
1773 | if (bo->exec == NULL && bo->rq && !__kgem_busy(kgem, bo->handle)) |
1774 | __kgem_bo_clear_busy(bo); |
1774 | __kgem_bo_clear_busy(bo); |
1775 | if (bo->rq == NULL) |
1775 | if (bo->rq == NULL) |
1776 | kgem_bo_move_to_snoop(kgem, bo); |
1776 | kgem_bo_move_to_snoop(kgem, bo); |
1777 | return; |
1777 | return; |
1778 | } |
1778 | } |
1779 | if (!IS_USER_MAP(bo->map)) |
1779 | if (!IS_USER_MAP(bo->map)) |
1780 | bo->flush = false; |
1780 | bo->flush = false; |
1781 | 1781 | ||
1782 | if (bo->scanout) { |
1782 | if (bo->scanout) { |
1783 | kgem_bo_move_to_scanout(kgem, bo); |
1783 | kgem_bo_move_to_scanout(kgem, bo); |
1784 | return; |
1784 | return; |
1785 | } |
1785 | } |
1786 | 1786 | ||
1787 | if (bo->io) |
1787 | if (bo->io) |
1788 | bo = kgem_bo_replace_io(bo); |
1788 | bo = kgem_bo_replace_io(bo); |
1789 | if (!bo->reusable) { |
1789 | if (!bo->reusable) { |
1790 | DBG(("%s: handle=%d, not reusable\n", |
1790 | DBG(("%s: handle=%d, not reusable\n", |
1791 | __FUNCTION__, bo->handle)); |
1791 | __FUNCTION__, bo->handle)); |
1792 | goto destroy; |
1792 | goto destroy; |
1793 | } |
1793 | } |
1794 | 1794 | ||
1795 | if (!kgem->has_llc && IS_CPU_MAP(bo->map) && bo->domain != DOMAIN_CPU) |
1795 | if (!kgem->has_llc && IS_CPU_MAP(bo->map) && bo->domain != DOMAIN_CPU) |
1796 | kgem_bo_release_map(kgem, bo); |
1796 | kgem_bo_release_map(kgem, bo); |
1797 | 1797 | ||
1798 | assert(list_is_empty(&bo->vma)); |
1798 | assert(list_is_empty(&bo->vma)); |
1799 | assert(list_is_empty(&bo->list)); |
1799 | assert(list_is_empty(&bo->list)); |
1800 | assert(bo->flush == false); |
1800 | assert(bo->flush == false); |
1801 | assert(bo->snoop == false); |
1801 | assert(bo->snoop == false); |
1802 | assert(bo->io == false); |
1802 | assert(bo->io == false); |
1803 | assert(bo->scanout == false); |
1803 | assert(bo->scanout == false); |
1804 | 1804 | ||
1805 | kgem_bo_undo(kgem, bo); |
1805 | kgem_bo_undo(kgem, bo); |
1806 | assert(bo->refcnt == 0); |
1806 | assert(bo->refcnt == 0); |
1807 | 1807 | ||
1808 | if (bo->rq && bo->exec == NULL && !__kgem_busy(kgem, bo->handle)) |
1808 | if (bo->rq && bo->exec == NULL && !__kgem_busy(kgem, bo->handle)) |
1809 | __kgem_bo_clear_busy(bo); |
1809 | __kgem_bo_clear_busy(bo); |
1810 | 1810 | ||
1811 | if (bo->rq) { |
1811 | if (bo->rq) { |
1812 | struct list *cache; |
1812 | struct list *cache; |
1813 | 1813 | ||
1814 | DBG(("%s: handle=%d -> active\n", __FUNCTION__, bo->handle)); |
1814 | DBG(("%s: handle=%d -> active\n", __FUNCTION__, bo->handle)); |
1815 | if (bucket(bo) < NUM_CACHE_BUCKETS) |
1815 | if (bucket(bo) < NUM_CACHE_BUCKETS) |
1816 | cache = &kgem->active[bucket(bo)][bo->tiling]; |
1816 | cache = &kgem->active[bucket(bo)][bo->tiling]; |
1817 | else |
1817 | else |
1818 | cache = &kgem->large; |
1818 | cache = &kgem->large; |
1819 | list_add(&bo->list, cache); |
1819 | list_add(&bo->list, cache); |
1820 | return; |
1820 | return; |
1821 | } |
1821 | } |
1822 | 1822 | ||
1823 | assert(bo->exec == NULL); |
1823 | assert(bo->exec == NULL); |
1824 | assert(list_is_empty(&bo->request)); |
1824 | assert(list_is_empty(&bo->request)); |
1825 | 1825 | ||
1826 | if (!IS_CPU_MAP(bo->map)) { |
1826 | if (!IS_CPU_MAP(bo->map)) { |
1827 | if (!kgem_bo_set_purgeable(kgem, bo)) |
1827 | if (!kgem_bo_set_purgeable(kgem, bo)) |
1828 | goto destroy; |
1828 | goto destroy; |
1829 | 1829 | ||
1830 | if (!kgem->has_llc && bo->domain == DOMAIN_CPU) |
1830 | if (!kgem->has_llc && bo->domain == DOMAIN_CPU) |
1831 | goto destroy; |
1831 | goto destroy; |
1832 | 1832 | ||
1833 | DBG(("%s: handle=%d, purged\n", |
1833 | DBG(("%s: handle=%d, purged\n", |
1834 | __FUNCTION__, bo->handle)); |
1834 | __FUNCTION__, bo->handle)); |
1835 | } |
1835 | } |
1836 | 1836 | ||
1837 | kgem_bo_move_to_inactive(kgem, bo); |
1837 | kgem_bo_move_to_inactive(kgem, bo); |
1838 | return; |
1838 | return; |
1839 | 1839 | ||
1840 | destroy: |
1840 | destroy: |
1841 | if (!bo->exec) |
1841 | if (!bo->exec) |
1842 | kgem_bo_free(kgem, bo); |
1842 | kgem_bo_free(kgem, bo); |
1843 | } |
1843 | } |
1844 | 1844 | ||
1845 | static void kgem_bo_unref(struct kgem *kgem, struct kgem_bo *bo) |
1845 | static void kgem_bo_unref(struct kgem *kgem, struct kgem_bo *bo) |
1846 | { |
1846 | { |
1847 | assert(bo->refcnt); |
1847 | assert(bo->refcnt); |
1848 | if (--bo->refcnt == 0) |
1848 | if (--bo->refcnt == 0) |
1849 | __kgem_bo_destroy(kgem, bo); |
1849 | __kgem_bo_destroy(kgem, bo); |
1850 | } |
1850 | } |
1851 | 1851 | ||
1852 | static void kgem_buffer_release(struct kgem *kgem, struct kgem_buffer *bo) |
1852 | static void kgem_buffer_release(struct kgem *kgem, struct kgem_buffer *bo) |
1853 | { |
1853 | { |
1854 | while (!list_is_empty(&bo->base.vma)) { |
1854 | while (!list_is_empty(&bo->base.vma)) { |
1855 | struct kgem_bo *cached; |
1855 | struct kgem_bo *cached; |
1856 | 1856 | ||
1857 | cached = list_first_entry(&bo->base.vma, struct kgem_bo, vma); |
1857 | cached = list_first_entry(&bo->base.vma, struct kgem_bo, vma); |
1858 | assert(cached->proxy == &bo->base); |
1858 | assert(cached->proxy == &bo->base); |
1859 | list_del(&cached->vma); |
1859 | list_del(&cached->vma); |
1860 | 1860 | ||
1861 | assert(*(struct kgem_bo **)cached->map == cached); |
1861 | assert(*(struct kgem_bo **)cached->map == cached); |
1862 | *(struct kgem_bo **)cached->map = NULL; |
1862 | *(struct kgem_bo **)cached->map = NULL; |
1863 | cached->map = NULL; |
1863 | cached->map = NULL; |
1864 | 1864 | ||
1865 | kgem_bo_destroy(kgem, cached); |
1865 | kgem_bo_destroy(kgem, cached); |
1866 | } |
1866 | } |
1867 | } |
1867 | } |
1868 | 1868 | ||
1869 | static bool kgem_retire__buffers(struct kgem *kgem) |
1869 | static bool kgem_retire__buffers(struct kgem *kgem) |
1870 | { |
1870 | { |
1871 | bool retired = false; |
1871 | bool retired = false; |
1872 | 1872 | ||
1873 | while (!list_is_empty(&kgem->active_buffers)) { |
1873 | while (!list_is_empty(&kgem->active_buffers)) { |
1874 | struct kgem_buffer *bo = |
1874 | struct kgem_buffer *bo = |
1875 | list_last_entry(&kgem->active_buffers, |
1875 | list_last_entry(&kgem->active_buffers, |
1876 | struct kgem_buffer, |
1876 | struct kgem_buffer, |
1877 | base.list); |
1877 | base.list); |
1878 | 1878 | ||
1879 | if (bo->base.rq) |
1879 | if (bo->base.rq) |
1880 | break; |
1880 | break; |
1881 | 1881 | ||
1882 | DBG(("%s: releasing upload cache for handle=%d? %d\n", |
1882 | DBG(("%s: releasing upload cache for handle=%d? %d\n", |
1883 | __FUNCTION__, bo->base.handle, !list_is_empty(&bo->base.vma))); |
1883 | __FUNCTION__, bo->base.handle, !list_is_empty(&bo->base.vma))); |
1884 | list_del(&bo->base.list); |
1884 | list_del(&bo->base.list); |
1885 | kgem_buffer_release(kgem, bo); |
1885 | kgem_buffer_release(kgem, bo); |
1886 | kgem_bo_unref(kgem, &bo->base); |
1886 | kgem_bo_unref(kgem, &bo->base); |
1887 | retired = true; |
1887 | retired = true; |
1888 | } |
1888 | } |
1889 | 1889 | ||
1890 | return retired; |
1890 | return retired; |
1891 | } |
1891 | } |
1892 | 1892 | ||
1893 | static bool kgem_retire__flushing(struct kgem *kgem) |
1893 | static bool kgem_retire__flushing(struct kgem *kgem) |
1894 | { |
1894 | { |
1895 | struct kgem_bo *bo, *next; |
1895 | struct kgem_bo *bo, *next; |
1896 | bool retired = false; |
1896 | bool retired = false; |
1897 | 1897 | ||
1898 | list_for_each_entry_safe(bo, next, &kgem->flushing, request) { |
1898 | list_for_each_entry_safe(bo, next, &kgem->flushing, request) { |
1899 | assert(bo->rq == (void *)kgem); |
1899 | assert(bo->rq == (void *)kgem); |
1900 | assert(bo->exec == NULL); |
1900 | assert(bo->exec == NULL); |
1901 | 1901 | ||
1902 | if (__kgem_busy(kgem, bo->handle)) |
1902 | if (__kgem_busy(kgem, bo->handle)) |
1903 | break; |
1903 | break; |
1904 | 1904 | ||
1905 | __kgem_bo_clear_busy(bo); |
1905 | __kgem_bo_clear_busy(bo); |
1906 | 1906 | ||
1907 | if (bo->refcnt) |
1907 | if (bo->refcnt) |
1908 | continue; |
1908 | continue; |
1909 | 1909 | ||
1910 | if (bo->snoop) { |
1910 | if (bo->snoop) { |
1911 | kgem_bo_move_to_snoop(kgem, bo); |
1911 | kgem_bo_move_to_snoop(kgem, bo); |
1912 | } else if (bo->scanout) { |
1912 | } else if (bo->scanout) { |
1913 | kgem_bo_move_to_scanout(kgem, bo); |
1913 | kgem_bo_move_to_scanout(kgem, bo); |
1914 | } else if ((bo = kgem_bo_replace_io(bo))->reusable && |
1914 | } else if ((bo = kgem_bo_replace_io(bo))->reusable && |
1915 | kgem_bo_set_purgeable(kgem, bo)) { |
1915 | kgem_bo_set_purgeable(kgem, bo)) { |
1916 | kgem_bo_move_to_inactive(kgem, bo); |
1916 | kgem_bo_move_to_inactive(kgem, bo); |
1917 | retired = true; |
1917 | retired = true; |
1918 | } else |
1918 | } else |
1919 | kgem_bo_free(kgem, bo); |
1919 | kgem_bo_free(kgem, bo); |
1920 | } |
1920 | } |
1921 | #if HAS_DEBUG_FULL |
1921 | #if HAS_DEBUG_FULL |
1922 | { |
1922 | { |
1923 | int count = 0; |
1923 | int count = 0; |
1924 | list_for_each_entry(bo, &kgem->flushing, request) |
1924 | list_for_each_entry(bo, &kgem->flushing, request) |
1925 | count++; |
1925 | count++; |
1926 | ErrorF("%s: %d bo on flushing list\n", __FUNCTION__, count); |
1926 | ErrorF("%s: %d bo on flushing list\n", __FUNCTION__, count); |
1927 | } |
1927 | } |
1928 | #endif |
1928 | #endif |
1929 | 1929 | ||
1930 | kgem->need_retire |= !list_is_empty(&kgem->flushing); |
1930 | kgem->need_retire |= !list_is_empty(&kgem->flushing); |
1931 | 1931 | ||
1932 | return retired; |
1932 | return retired; |
1933 | } |
1933 | } |
1934 | 1934 | ||
1935 | 1935 | ||
1936 | static bool __kgem_retire_rq(struct kgem *kgem, struct kgem_request *rq) |
1936 | static bool __kgem_retire_rq(struct kgem *kgem, struct kgem_request *rq) |
1937 | { |
1937 | { |
1938 | bool retired = false; |
1938 | bool retired = false; |
1939 | 1939 | ||
1940 | DBG(("%s: request %d complete\n", |
1940 | DBG(("%s: request %d complete\n", |
1941 | __FUNCTION__, rq->bo->handle)); |
1941 | __FUNCTION__, rq->bo->handle)); |
1942 | 1942 | ||
1943 | while (!list_is_empty(&rq->buffers)) { |
1943 | while (!list_is_empty(&rq->buffers)) { |
1944 | struct kgem_bo *bo; |
1944 | struct kgem_bo *bo; |
1945 | 1945 | ||
1946 | bo = list_first_entry(&rq->buffers, |
1946 | bo = list_first_entry(&rq->buffers, |
1947 | struct kgem_bo, |
1947 | struct kgem_bo, |
1948 | request); |
1948 | request); |
1949 | 1949 | ||
1950 | assert(RQ(bo->rq) == rq); |
1950 | assert(RQ(bo->rq) == rq); |
1951 | assert(bo->exec == NULL); |
1951 | assert(bo->exec == NULL); |
1952 | assert(bo->domain == DOMAIN_GPU || bo->domain == DOMAIN_NONE); |
1952 | assert(bo->domain == DOMAIN_GPU || bo->domain == DOMAIN_NONE); |
1953 | 1953 | ||
1954 | list_del(&bo->request); |
1954 | list_del(&bo->request); |
1955 | 1955 | ||
1956 | if (bo->needs_flush) |
1956 | if (bo->needs_flush) |
1957 | bo->needs_flush = __kgem_busy(kgem, bo->handle); |
1957 | bo->needs_flush = __kgem_busy(kgem, bo->handle); |
1958 | if (bo->needs_flush) { |
1958 | if (bo->needs_flush) { |
1959 | DBG(("%s: moving %d to flushing\n", |
1959 | DBG(("%s: moving %d to flushing\n", |
1960 | __FUNCTION__, bo->handle)); |
1960 | __FUNCTION__, bo->handle)); |
1961 | list_add(&bo->request, &kgem->flushing); |
1961 | list_add(&bo->request, &kgem->flushing); |
1962 | bo->rq = (void *)kgem; |
1962 | bo->rq = (void *)kgem; |
1963 | continue; |
1963 | continue; |
1964 | } |
1964 | } |
1965 | 1965 | ||
1966 | bo->domain = DOMAIN_NONE; |
1966 | bo->domain = DOMAIN_NONE; |
1967 | bo->rq = NULL; |
1967 | bo->rq = NULL; |
1968 | if (bo->refcnt) |
1968 | if (bo->refcnt) |
1969 | continue; |
1969 | continue; |
1970 | 1970 | ||
1971 | if (bo->snoop) { |
1971 | if (bo->snoop) { |
1972 | kgem_bo_move_to_snoop(kgem, bo); |
1972 | kgem_bo_move_to_snoop(kgem, bo); |
1973 | } else if (bo->scanout) { |
1973 | } else if (bo->scanout) { |
1974 | kgem_bo_move_to_scanout(kgem, bo); |
1974 | kgem_bo_move_to_scanout(kgem, bo); |
1975 | } else if ((bo = kgem_bo_replace_io(bo))->reusable && |
1975 | } else if ((bo = kgem_bo_replace_io(bo))->reusable && |
1976 | kgem_bo_set_purgeable(kgem, bo)) { |
1976 | kgem_bo_set_purgeable(kgem, bo)) { |
1977 | kgem_bo_move_to_inactive(kgem, bo); |
1977 | kgem_bo_move_to_inactive(kgem, bo); |
1978 | retired = true; |
1978 | retired = true; |
1979 | } else { |
1979 | } else { |
1980 | DBG(("%s: closing %d\n", |
1980 | DBG(("%s: closing %d\n", |
1981 | __FUNCTION__, bo->handle)); |
1981 | __FUNCTION__, bo->handle)); |
1982 | kgem_bo_free(kgem, bo); |
1982 | kgem_bo_free(kgem, bo); |
1983 | } |
1983 | } |
1984 | } |
1984 | } |
1985 | 1985 | ||
1986 | assert(rq->bo->rq == NULL); |
1986 | assert(rq->bo->rq == NULL); |
1987 | assert(list_is_empty(&rq->bo->request)); |
1987 | assert(list_is_empty(&rq->bo->request)); |
1988 | 1988 | ||
1989 | if (--rq->bo->refcnt == 0) { |
1989 | if (--rq->bo->refcnt == 0) { |
1990 | if (kgem_bo_set_purgeable(kgem, rq->bo)) { |
1990 | if (kgem_bo_set_purgeable(kgem, rq->bo)) { |
1991 | kgem_bo_move_to_inactive(kgem, rq->bo); |
1991 | kgem_bo_move_to_inactive(kgem, rq->bo); |
1992 | retired = true; |
1992 | retired = true; |
1993 | } else { |
1993 | } else { |
1994 | DBG(("%s: closing %d\n", |
1994 | DBG(("%s: closing %d\n", |
1995 | __FUNCTION__, rq->bo->handle)); |
1995 | __FUNCTION__, rq->bo->handle)); |
1996 | kgem_bo_free(kgem, rq->bo); |
1996 | kgem_bo_free(kgem, rq->bo); |
1997 | } |
1997 | } |
1998 | } |
1998 | } |
1999 | 1999 | ||
2000 | __kgem_request_free(rq); |
2000 | __kgem_request_free(rq); |
2001 | return retired; |
2001 | return retired; |
2002 | } |
2002 | } |
2003 | 2003 | ||
2004 | static bool kgem_retire__requests_ring(struct kgem *kgem, int ring) |
2004 | static bool kgem_retire__requests_ring(struct kgem *kgem, int ring) |
2005 | { |
2005 | { |
2006 | bool retired = false; |
2006 | bool retired = false; |
2007 | 2007 | ||
2008 | while (!list_is_empty(&kgem->requests[ring])) { |
2008 | while (!list_is_empty(&kgem->requests[ring])) { |
2009 | struct kgem_request *rq; |
2009 | struct kgem_request *rq; |
2010 | 2010 | ||
2011 | rq = list_first_entry(&kgem->requests[ring], |
2011 | rq = list_first_entry(&kgem->requests[ring], |
2012 | struct kgem_request, |
2012 | struct kgem_request, |
2013 | list); |
2013 | list); |
2014 | if (__kgem_busy(kgem, rq->bo->handle)) |
2014 | if (__kgem_busy(kgem, rq->bo->handle)) |
2015 | break; |
2015 | break; |
2016 | 2016 | ||
2017 | retired |= __kgem_retire_rq(kgem, rq); |
2017 | retired |= __kgem_retire_rq(kgem, rq); |
2018 | } |
2018 | } |
2019 | 2019 | ||
2020 | #if HAS_DEBUG_FULL |
2020 | #if HAS_DEBUG_FULL |
2021 | { |
2021 | { |
2022 | struct kgem_bo *bo; |
2022 | struct kgem_bo *bo; |
2023 | int count = 0; |
2023 | int count = 0; |
2024 | 2024 | ||
2025 | list_for_each_entry(bo, &kgem->requests[ring], request) |
2025 | list_for_each_entry(bo, &kgem->requests[ring], request) |
2026 | count++; |
2026 | count++; |
2027 | 2027 | ||
2028 | bo = NULL; |
2028 | bo = NULL; |
2029 | if (!list_is_empty(&kgem->requests[ring])) |
2029 | if (!list_is_empty(&kgem->requests[ring])) |
2030 | bo = list_first_entry(&kgem->requests[ring], |
2030 | bo = list_first_entry(&kgem->requests[ring], |
2031 | struct kgem_request, |
2031 | struct kgem_request, |
2032 | list)->bo; |
2032 | list)->bo; |
2033 | 2033 | ||
2034 | ErrorF("%s: ring=%d, %d outstanding requests, oldest=%d\n", |
2034 | ErrorF("%s: ring=%d, %d outstanding requests, oldest=%d\n", |
2035 | __FUNCTION__, ring, count, bo ? bo->handle : 0); |
2035 | __FUNCTION__, ring, count, bo ? bo->handle : 0); |
2036 | } |
2036 | } |
2037 | #endif |
2037 | #endif |
2038 | 2038 | ||
2039 | return retired; |
2039 | return retired; |
2040 | } |
2040 | } |
2041 | 2041 | ||
2042 | static bool kgem_retire__requests(struct kgem *kgem) |
2042 | static bool kgem_retire__requests(struct kgem *kgem) |
2043 | { |
2043 | { |
2044 | bool retired = false; |
2044 | bool retired = false; |
2045 | int n; |
2045 | int n; |
2046 | 2046 | ||
2047 | for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) { |
2047 | for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) { |
2048 | retired |= kgem_retire__requests_ring(kgem, n); |
2048 | retired |= kgem_retire__requests_ring(kgem, n); |
2049 | kgem->need_retire |= !list_is_empty(&kgem->requests[n]); |
2049 | kgem->need_retire |= !list_is_empty(&kgem->requests[n]); |
2050 | } |
2050 | } |
2051 | 2051 | ||
2052 | return retired; |
2052 | return retired; |
2053 | } |
2053 | } |
2054 | 2054 | ||
2055 | bool kgem_retire(struct kgem *kgem) |
2055 | bool kgem_retire(struct kgem *kgem) |
2056 | { |
2056 | { |
2057 | bool retired = false; |
2057 | bool retired = false; |
2058 | 2058 | ||
2059 | DBG(("%s\n", __FUNCTION__)); |
2059 | DBG(("%s\n", __FUNCTION__)); |
2060 | 2060 | ||
2061 | kgem->need_retire = false; |
2061 | kgem->need_retire = false; |
2062 | 2062 | ||
2063 | retired |= kgem_retire__flushing(kgem); |
2063 | retired |= kgem_retire__flushing(kgem); |
2064 | retired |= kgem_retire__requests(kgem); |
2064 | retired |= kgem_retire__requests(kgem); |
2065 | retired |= kgem_retire__buffers(kgem); |
2065 | retired |= kgem_retire__buffers(kgem); |
2066 | 2066 | ||
2067 | DBG(("%s -- retired=%d, need_retire=%d\n", |
2067 | DBG(("%s -- retired=%d, need_retire=%d\n", |
2068 | __FUNCTION__, retired, kgem->need_retire)); |
2068 | __FUNCTION__, retired, kgem->need_retire)); |
2069 | 2069 | ||
2070 | kgem->retire(kgem); |
2070 | kgem->retire(kgem); |
2071 | 2071 | ||
2072 | return retired; |
2072 | return retired; |
2073 | } |
2073 | } |
2074 | 2074 | ||
2075 | bool __kgem_ring_is_idle(struct kgem *kgem, int ring) |
2075 | bool __kgem_ring_is_idle(struct kgem *kgem, int ring) |
2076 | { |
2076 | { |
2077 | struct kgem_request *rq; |
2077 | struct kgem_request *rq; |
2078 | 2078 | ||
2079 | assert(!list_is_empty(&kgem->requests[ring])); |
2079 | assert(!list_is_empty(&kgem->requests[ring])); |
2080 | 2080 | ||
2081 | rq = list_last_entry(&kgem->requests[ring], |
2081 | rq = list_last_entry(&kgem->requests[ring], |
2082 | struct kgem_request, list); |
2082 | struct kgem_request, list); |
2083 | if (__kgem_busy(kgem, rq->bo->handle)) { |
2083 | if (__kgem_busy(kgem, rq->bo->handle)) { |
2084 | DBG(("%s: last requests handle=%d still busy\n", |
2084 | DBG(("%s: last requests handle=%d still busy\n", |
2085 | __FUNCTION__, rq->bo->handle)); |
2085 | __FUNCTION__, rq->bo->handle)); |
2086 | return false; |
2086 | return false; |
2087 | } |
2087 | } |
2088 | 2088 | ||
2089 | DBG(("%s: ring=%d idle (handle=%d)\n", |
2089 | DBG(("%s: ring=%d idle (handle=%d)\n", |
2090 | __FUNCTION__, ring, rq->bo->handle)); |
2090 | __FUNCTION__, ring, rq->bo->handle)); |
2091 | 2091 | ||
2092 | kgem_retire__requests_ring(kgem, ring); |
2092 | kgem_retire__requests_ring(kgem, ring); |
2093 | assert(list_is_empty(&kgem->requests[ring])); |
2093 | assert(list_is_empty(&kgem->requests[ring])); |
2094 | return true; |
2094 | return true; |
2095 | } |
2095 | } |
2096 | 2096 | ||
2097 | static void kgem_commit(struct kgem *kgem) |
2097 | static void kgem_commit(struct kgem *kgem) |
2098 | { |
2098 | { |
2099 | struct kgem_request *rq = kgem->next_request; |
2099 | struct kgem_request *rq = kgem->next_request; |
2100 | struct kgem_bo *bo, *next; |
2100 | struct kgem_bo *bo, *next; |
2101 | 2101 | ||
2102 | list_for_each_entry_safe(bo, next, &rq->buffers, request) { |
2102 | list_for_each_entry_safe(bo, next, &rq->buffers, request) { |
2103 | assert(next->request.prev == &bo->request); |
2103 | assert(next->request.prev == &bo->request); |
2104 | 2104 | ||
2105 | DBG(("%s: release handle=%d (proxy? %d), dirty? %d flush? %d, snoop? %d -> offset=%x\n", |
2105 | DBG(("%s: release handle=%d (proxy? %d), dirty? %d flush? %d, snoop? %d -> offset=%x\n", |
2106 | __FUNCTION__, bo->handle, bo->proxy != NULL, |
2106 | __FUNCTION__, bo->handle, bo->proxy != NULL, |
2107 | bo->gpu_dirty, bo->needs_flush, bo->snoop, |
2107 | bo->gpu_dirty, bo->needs_flush, bo->snoop, |
2108 | (unsigned)bo->exec->offset)); |
2108 | (unsigned)bo->exec->offset)); |
2109 | 2109 | ||
2110 | assert(bo->exec); |
2110 | assert(bo->exec); |
2111 | assert(bo->proxy == NULL || bo->exec == &_kgem_dummy_exec); |
2111 | assert(bo->proxy == NULL || bo->exec == &_kgem_dummy_exec); |
2112 | assert(RQ(bo->rq) == rq || (RQ(bo->proxy->rq) == rq)); |
2112 | assert(RQ(bo->rq) == rq || (RQ(bo->proxy->rq) == rq)); |
2113 | 2113 | ||
2114 | bo->presumed_offset = bo->exec->offset; |
2114 | bo->presumed_offset = bo->exec->offset; |
2115 | bo->exec = NULL; |
2115 | bo->exec = NULL; |
2116 | bo->target_handle = -1; |
2116 | bo->target_handle = -1; |
2117 | 2117 | ||
2118 | if (!bo->refcnt && !bo->reusable) { |
2118 | if (!bo->refcnt && !bo->reusable) { |
2119 | assert(!bo->snoop); |
2119 | assert(!bo->snoop); |
2120 | kgem_bo_free(kgem, bo); |
2120 | kgem_bo_free(kgem, bo); |
2121 | continue; |
2121 | continue; |
2122 | } |
2122 | } |
2123 | 2123 | ||
2124 | bo->binding.offset = 0; |
2124 | bo->binding.offset = 0; |
2125 | bo->domain = DOMAIN_GPU; |
2125 | bo->domain = DOMAIN_GPU; |
2126 | bo->gpu_dirty = false; |
2126 | bo->gpu_dirty = false; |
2127 | 2127 | ||
2128 | if (bo->proxy) { |
2128 | if (bo->proxy) { |
2129 | /* proxies are not used for domain tracking */ |
2129 | /* proxies are not used for domain tracking */ |
2130 | bo->exec = NULL; |
2130 | bo->exec = NULL; |
2131 | __kgem_bo_clear_busy(bo); |
2131 | __kgem_bo_clear_busy(bo); |
2132 | } |
2132 | } |
2133 | 2133 | ||
2134 | kgem->scanout_busy |= bo->scanout; |
2134 | kgem->scanout_busy |= bo->scanout; |
2135 | } |
2135 | } |
2136 | 2136 | ||
2137 | if (rq == &kgem->static_request) { |
2137 | if (rq == &kgem->static_request) { |
2138 | struct drm_i915_gem_set_domain set_domain; |
2138 | struct drm_i915_gem_set_domain set_domain; |
2139 | 2139 | ||
2140 | DBG(("%s: syncing due to allocation failure\n", __FUNCTION__)); |
2140 | DBG(("%s: syncing due to allocation failure\n", __FUNCTION__)); |
2141 | 2141 | ||
2142 | VG_CLEAR(set_domain); |
2142 | VG_CLEAR(set_domain); |
2143 | set_domain.handle = rq->bo->handle; |
2143 | set_domain.handle = rq->bo->handle; |
2144 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
2144 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
2145 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
2145 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
2146 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) { |
2146 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) { |
2147 | DBG(("%s: sync: GPU hang detected\n", __FUNCTION__)); |
2147 | DBG(("%s: sync: GPU hang detected\n", __FUNCTION__)); |
2148 | kgem_throttle(kgem); |
2148 | kgem_throttle(kgem); |
2149 | } |
2149 | } |
2150 | 2150 | ||
2151 | kgem_retire(kgem); |
2151 | kgem_retire(kgem); |
2152 | assert(list_is_empty(&rq->buffers)); |
2152 | assert(list_is_empty(&rq->buffers)); |
2153 | 2153 | ||
2154 | assert(rq->bo->map == NULL); |
2154 | assert(rq->bo->map == NULL); |
2155 | gem_close(kgem->fd, rq->bo->handle); |
2155 | gem_close(kgem->fd, rq->bo->handle); |
2156 | kgem_cleanup_cache(kgem); |
2156 | kgem_cleanup_cache(kgem); |
2157 | } else { |
2157 | } else { |
2158 | list_add_tail(&rq->list, &kgem->requests[rq->ring]); |
2158 | list_add_tail(&rq->list, &kgem->requests[rq->ring]); |
2159 | kgem->need_throttle = kgem->need_retire = 1; |
2159 | kgem->need_throttle = kgem->need_retire = 1; |
2160 | } |
2160 | } |
2161 | 2161 | ||
2162 | kgem->next_request = NULL; |
2162 | kgem->next_request = NULL; |
2163 | } |
2163 | } |
2164 | 2164 | ||
2165 | static void kgem_close_list(struct kgem *kgem, struct list *head) |
2165 | static void kgem_close_list(struct kgem *kgem, struct list *head) |
2166 | { |
2166 | { |
2167 | while (!list_is_empty(head)) |
2167 | while (!list_is_empty(head)) |
2168 | kgem_bo_free(kgem, list_first_entry(head, struct kgem_bo, list)); |
2168 | kgem_bo_free(kgem, list_first_entry(head, struct kgem_bo, list)); |
2169 | } |
2169 | } |
2170 | 2170 | ||
2171 | static void kgem_close_inactive(struct kgem *kgem) |
2171 | static void kgem_close_inactive(struct kgem *kgem) |
2172 | { |
2172 | { |
2173 | unsigned int i; |
2173 | unsigned int i; |
2174 | 2174 | ||
2175 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) |
2175 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) |
2176 | kgem_close_list(kgem, &kgem->inactive[i]); |
2176 | kgem_close_list(kgem, &kgem->inactive[i]); |
2177 | } |
2177 | } |
2178 | 2178 | ||
2179 | static void kgem_finish_buffers(struct kgem *kgem) |
2179 | static void kgem_finish_buffers(struct kgem *kgem) |
2180 | { |
2180 | { |
2181 | struct kgem_buffer *bo, *next; |
2181 | struct kgem_buffer *bo, *next; |
2182 | 2182 | ||
2183 | list_for_each_entry_safe(bo, next, &kgem->batch_buffers, base.list) { |
2183 | list_for_each_entry_safe(bo, next, &kgem->batch_buffers, base.list) { |
2184 | DBG(("%s: buffer handle=%d, used=%d, exec?=%d, write=%d, mmapped=%s\n", |
2184 | DBG(("%s: buffer handle=%d, used=%d, exec?=%d, write=%d, mmapped=%s\n", |
2185 | __FUNCTION__, bo->base.handle, bo->used, bo->base.exec!=NULL, |
2185 | __FUNCTION__, bo->base.handle, bo->used, bo->base.exec!=NULL, |
2186 | bo->write, bo->mmapped ? IS_CPU_MAP(bo->base.map) ? "cpu" : "gtt" : "no")); |
2186 | bo->write, bo->mmapped ? IS_CPU_MAP(bo->base.map) ? "cpu" : "gtt" : "no")); |
2187 | 2187 | ||
2188 | assert(next->base.list.prev == &bo->base.list); |
2188 | assert(next->base.list.prev == &bo->base.list); |
2189 | assert(bo->base.io); |
2189 | assert(bo->base.io); |
2190 | assert(bo->base.refcnt >= 1); |
2190 | assert(bo->base.refcnt >= 1); |
2191 | 2191 | ||
2192 | if (!bo->base.exec) { |
2192 | if (!bo->base.exec) { |
2193 | DBG(("%s: skipping unattached handle=%d, used=%d\n", |
2193 | DBG(("%s: skipping unattached handle=%d, used=%d\n", |
2194 | __FUNCTION__, bo->base.handle, bo->used)); |
2194 | __FUNCTION__, bo->base.handle, bo->used)); |
2195 | continue; |
2195 | continue; |
2196 | } |
2196 | } |
2197 | 2197 | ||
2198 | if (!bo->write) { |
2198 | if (!bo->write) { |
2199 | assert(bo->base.exec || bo->base.refcnt > 1); |
2199 | assert(bo->base.exec || bo->base.refcnt > 1); |
2200 | goto decouple; |
2200 | goto decouple; |
2201 | } |
2201 | } |
2202 | 2202 | ||
2203 | if (bo->mmapped) { |
2203 | if (bo->mmapped) { |
2204 | int used; |
2204 | int used; |
2205 | 2205 | ||
2206 | assert(!bo->need_io); |
2206 | assert(!bo->need_io); |
2207 | 2207 | ||
2208 | used = ALIGN(bo->used, PAGE_SIZE); |
2208 | used = ALIGN(bo->used, PAGE_SIZE); |
2209 | if (!DBG_NO_UPLOAD_ACTIVE && |
2209 | if (!DBG_NO_UPLOAD_ACTIVE && |
2210 | used + PAGE_SIZE <= bytes(&bo->base) && |
2210 | used + PAGE_SIZE <= bytes(&bo->base) && |
2211 | (kgem->has_llc || !IS_CPU_MAP(bo->base.map) || bo->base.snoop)) { |
2211 | (kgem->has_llc || !IS_CPU_MAP(bo->base.map) || bo->base.snoop)) { |
2212 | DBG(("%s: retaining upload buffer (%d/%d)\n", |
2212 | DBG(("%s: retaining upload buffer (%d/%d)\n", |
2213 | __FUNCTION__, bo->used, bytes(&bo->base))); |
2213 | __FUNCTION__, bo->used, bytes(&bo->base))); |
2214 | bo->used = used; |
2214 | bo->used = used; |
2215 | list_move(&bo->base.list, |
2215 | list_move(&bo->base.list, |
2216 | &kgem->active_buffers); |
2216 | &kgem->active_buffers); |
2217 | continue; |
2217 | continue; |
2218 | } |
2218 | } |
2219 | DBG(("%s: discarding mmapped buffer, used=%d, map type=%d\n", |
2219 | DBG(("%s: discarding mmapped buffer, used=%d, map type=%d\n", |
2220 | __FUNCTION__, bo->used, (int)__MAP_TYPE(bo->base.map))); |
2220 | __FUNCTION__, bo->used, (int)__MAP_TYPE(bo->base.map))); |
2221 | goto decouple; |
2221 | goto decouple; |
2222 | } |
2222 | } |
2223 | 2223 | ||
2224 | if (!bo->used) { |
2224 | if (!bo->used) { |
2225 | /* Unless we replace the handle in the execbuffer, |
2225 | /* Unless we replace the handle in the execbuffer, |
2226 | * then this bo will become active. So decouple it |
2226 | * then this bo will become active. So decouple it |
2227 | * from the buffer list and track it in the normal |
2227 | * from the buffer list and track it in the normal |
2228 | * manner. |
2228 | * manner. |
2229 | */ |
2229 | */ |
2230 | goto decouple; |
2230 | goto decouple; |
2231 | } |
2231 | } |
2232 | 2232 | ||
2233 | assert(bo->need_io); |
2233 | assert(bo->need_io); |
2234 | assert(bo->base.rq == MAKE_REQUEST(kgem->next_request, kgem->ring)); |
2234 | assert(bo->base.rq == MAKE_REQUEST(kgem->next_request, kgem->ring)); |
2235 | assert(bo->base.domain != DOMAIN_GPU); |
2235 | assert(bo->base.domain != DOMAIN_GPU); |
2236 | 2236 | ||
2237 | if (bo->base.refcnt == 1 && |
2237 | if (bo->base.refcnt == 1 && |
2238 | bo->base.size.pages.count > 1 && |
2238 | bo->base.size.pages.count > 1 && |
2239 | bo->used < bytes(&bo->base) / 2) { |
2239 | bo->used < bytes(&bo->base) / 2) { |
2240 | struct kgem_bo *shrink; |
2240 | struct kgem_bo *shrink; |
2241 | unsigned alloc = NUM_PAGES(bo->used); |
2241 | unsigned alloc = NUM_PAGES(bo->used); |
2242 | 2242 | ||
2243 | shrink = search_snoop_cache(kgem, alloc, |
2243 | shrink = search_snoop_cache(kgem, alloc, |
2244 | CREATE_INACTIVE | CREATE_NO_RETIRE); |
2244 | CREATE_INACTIVE | CREATE_NO_RETIRE); |
2245 | if (shrink) { |
2245 | if (shrink) { |
2246 | void *map; |
2246 | void *map; |
2247 | int n; |
2247 | int n; |
2248 | 2248 | ||
2249 | DBG(("%s: used=%d, shrinking %d to %d, handle %d to %d\n", |
2249 | DBG(("%s: used=%d, shrinking %d to %d, handle %d to %d\n", |
2250 | __FUNCTION__, |
2250 | __FUNCTION__, |
2251 | bo->used, bytes(&bo->base), bytes(shrink), |
2251 | bo->used, bytes(&bo->base), bytes(shrink), |
2252 | bo->base.handle, shrink->handle)); |
2252 | bo->base.handle, shrink->handle)); |
2253 | 2253 | ||
2254 | assert(bo->used <= bytes(shrink)); |
2254 | assert(bo->used <= bytes(shrink)); |
2255 | map = kgem_bo_map__cpu(kgem, shrink); |
2255 | map = kgem_bo_map__cpu(kgem, shrink); |
2256 | if (map) { |
2256 | if (map) { |
2257 | kgem_bo_sync__cpu(kgem, shrink); |
2257 | kgem_bo_sync__cpu(kgem, shrink); |
2258 | memcpy(map, bo->mem, bo->used); |
2258 | memcpy(map, bo->mem, bo->used); |
2259 | 2259 | ||
2260 | shrink->target_handle = |
2260 | shrink->target_handle = |
2261 | kgem->has_handle_lut ? bo->base.target_handle : shrink->handle; |
2261 | kgem->has_handle_lut ? bo->base.target_handle : shrink->handle; |
2262 | for (n = 0; n < kgem->nreloc; n++) { |
2262 | for (n = 0; n < kgem->nreloc; n++) { |
2263 | if (kgem->reloc[n].target_handle == bo->base.target_handle) { |
2263 | if (kgem->reloc[n].target_handle == bo->base.target_handle) { |
2264 | kgem->reloc[n].target_handle = shrink->target_handle; |
2264 | kgem->reloc[n].target_handle = shrink->target_handle; |
2265 | kgem->reloc[n].presumed_offset = shrink->presumed_offset; |
2265 | kgem->reloc[n].presumed_offset = shrink->presumed_offset; |
2266 | kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0])] = |
2266 | kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0])] = |
2267 | kgem->reloc[n].delta + shrink->presumed_offset; |
2267 | kgem->reloc[n].delta + shrink->presumed_offset; |
2268 | } |
2268 | } |
2269 | } |
2269 | } |
2270 | 2270 | ||
2271 | bo->base.exec->handle = shrink->handle; |
2271 | bo->base.exec->handle = shrink->handle; |
2272 | bo->base.exec->offset = shrink->presumed_offset; |
2272 | bo->base.exec->offset = shrink->presumed_offset; |
2273 | shrink->exec = bo->base.exec; |
2273 | shrink->exec = bo->base.exec; |
2274 | shrink->rq = bo->base.rq; |
2274 | shrink->rq = bo->base.rq; |
2275 | list_replace(&bo->base.request, |
2275 | list_replace(&bo->base.request, |
2276 | &shrink->request); |
2276 | &shrink->request); |
2277 | list_init(&bo->base.request); |
2277 | list_init(&bo->base.request); |
2278 | shrink->needs_flush = bo->base.gpu_dirty; |
2278 | shrink->needs_flush = bo->base.gpu_dirty; |
2279 | 2279 | ||
2280 | bo->base.exec = NULL; |
2280 | bo->base.exec = NULL; |
2281 | bo->base.rq = NULL; |
2281 | bo->base.rq = NULL; |
2282 | bo->base.gpu_dirty = false; |
2282 | bo->base.gpu_dirty = false; |
2283 | bo->base.needs_flush = false; |
2283 | bo->base.needs_flush = false; |
2284 | bo->used = 0; |
2284 | bo->used = 0; |
2285 | 2285 | ||
2286 | goto decouple; |
2286 | goto decouple; |
2287 | } |
2287 | } |
2288 | 2288 | ||
2289 | __kgem_bo_destroy(kgem, shrink); |
2289 | __kgem_bo_destroy(kgem, shrink); |
2290 | } |
2290 | } |
2291 | 2291 | ||
2292 | shrink = search_linear_cache(kgem, alloc, |
2292 | shrink = search_linear_cache(kgem, alloc, |
2293 | CREATE_INACTIVE | CREATE_NO_RETIRE); |
2293 | CREATE_INACTIVE | CREATE_NO_RETIRE); |
2294 | if (shrink) { |
2294 | if (shrink) { |
2295 | int n; |
2295 | int n; |
2296 | 2296 | ||
2297 | DBG(("%s: used=%d, shrinking %d to %d, handle %d to %d\n", |
2297 | DBG(("%s: used=%d, shrinking %d to %d, handle %d to %d\n", |
2298 | __FUNCTION__, |
2298 | __FUNCTION__, |
2299 | bo->used, bytes(&bo->base), bytes(shrink), |
2299 | bo->used, bytes(&bo->base), bytes(shrink), |
2300 | bo->base.handle, shrink->handle)); |
2300 | bo->base.handle, shrink->handle)); |
2301 | 2301 | ||
2302 | assert(bo->used <= bytes(shrink)); |
2302 | assert(bo->used <= bytes(shrink)); |
2303 | if (gem_write(kgem->fd, shrink->handle, |
2303 | if (gem_write(kgem->fd, shrink->handle, |
2304 | 0, bo->used, bo->mem) == 0) { |
2304 | 0, bo->used, bo->mem) == 0) { |
2305 | shrink->target_handle = |
2305 | shrink->target_handle = |
2306 | kgem->has_handle_lut ? bo->base.target_handle : shrink->handle; |
2306 | kgem->has_handle_lut ? bo->base.target_handle : shrink->handle; |
2307 | for (n = 0; n < kgem->nreloc; n++) { |
2307 | for (n = 0; n < kgem->nreloc; n++) { |
2308 | if (kgem->reloc[n].target_handle == bo->base.target_handle) { |
2308 | if (kgem->reloc[n].target_handle == bo->base.target_handle) { |
2309 | kgem->reloc[n].target_handle = shrink->target_handle; |
2309 | kgem->reloc[n].target_handle = shrink->target_handle; |
2310 | kgem->reloc[n].presumed_offset = shrink->presumed_offset; |
2310 | kgem->reloc[n].presumed_offset = shrink->presumed_offset; |
2311 | kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0])] = |
2311 | kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0])] = |
2312 | kgem->reloc[n].delta + shrink->presumed_offset; |
2312 | kgem->reloc[n].delta + shrink->presumed_offset; |
2313 | } |
2313 | } |
2314 | } |
2314 | } |
2315 | 2315 | ||
2316 | bo->base.exec->handle = shrink->handle; |
2316 | bo->base.exec->handle = shrink->handle; |
2317 | bo->base.exec->offset = shrink->presumed_offset; |
2317 | bo->base.exec->offset = shrink->presumed_offset; |
2318 | shrink->exec = bo->base.exec; |
2318 | shrink->exec = bo->base.exec; |
2319 | shrink->rq = bo->base.rq; |
2319 | shrink->rq = bo->base.rq; |
2320 | list_replace(&bo->base.request, |
2320 | list_replace(&bo->base.request, |
2321 | &shrink->request); |
2321 | &shrink->request); |
2322 | list_init(&bo->base.request); |
2322 | list_init(&bo->base.request); |
2323 | shrink->needs_flush = bo->base.gpu_dirty; |
2323 | shrink->needs_flush = bo->base.gpu_dirty; |
2324 | 2324 | ||
2325 | bo->base.exec = NULL; |
2325 | bo->base.exec = NULL; |
2326 | bo->base.rq = NULL; |
2326 | bo->base.rq = NULL; |
2327 | bo->base.gpu_dirty = false; |
2327 | bo->base.gpu_dirty = false; |
2328 | bo->base.needs_flush = false; |
2328 | bo->base.needs_flush = false; |
2329 | bo->used = 0; |
2329 | bo->used = 0; |
2330 | 2330 | ||
2331 | goto decouple; |
2331 | goto decouple; |
2332 | } |
2332 | } |
2333 | 2333 | ||
2334 | __kgem_bo_destroy(kgem, shrink); |
2334 | __kgem_bo_destroy(kgem, shrink); |
2335 | } |
2335 | } |
2336 | } |
2336 | } |
2337 | 2337 | ||
2338 | DBG(("%s: handle=%d, uploading %d/%d\n", |
2338 | DBG(("%s: handle=%d, uploading %d/%d\n", |
2339 | __FUNCTION__, bo->base.handle, bo->used, bytes(&bo->base))); |
2339 | __FUNCTION__, bo->base.handle, bo->used, bytes(&bo->base))); |
2340 | ASSERT_IDLE(kgem, bo->base.handle); |
2340 | ASSERT_IDLE(kgem, bo->base.handle); |
2341 | assert(bo->used <= bytes(&bo->base)); |
2341 | assert(bo->used <= bytes(&bo->base)); |
2342 | gem_write(kgem->fd, bo->base.handle, |
2342 | gem_write(kgem->fd, bo->base.handle, |
2343 | 0, bo->used, bo->mem); |
2343 | 0, bo->used, bo->mem); |
2344 | bo->need_io = 0; |
2344 | bo->need_io = 0; |
2345 | 2345 | ||
2346 | decouple: |
2346 | decouple: |
2347 | DBG(("%s: releasing handle=%d\n", |
2347 | DBG(("%s: releasing handle=%d\n", |
2348 | __FUNCTION__, bo->base.handle)); |
2348 | __FUNCTION__, bo->base.handle)); |
2349 | list_del(&bo->base.list); |
2349 | list_del(&bo->base.list); |
2350 | kgem_bo_unref(kgem, &bo->base); |
2350 | kgem_bo_unref(kgem, &bo->base); |
2351 | } |
2351 | } |
2352 | } |
2352 | } |
2353 | 2353 | ||
2354 | static void kgem_cleanup(struct kgem *kgem) |
2354 | static void kgem_cleanup(struct kgem *kgem) |
2355 | { |
2355 | { |
2356 | int n; |
2356 | int n; |
2357 | 2357 | ||
2358 | for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) { |
2358 | for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) { |
2359 | while (!list_is_empty(&kgem->requests[n])) { |
2359 | while (!list_is_empty(&kgem->requests[n])) { |
2360 | struct kgem_request *rq; |
2360 | struct kgem_request *rq; |
2361 | 2361 | ||
2362 | rq = list_first_entry(&kgem->requests[n], |
2362 | rq = list_first_entry(&kgem->requests[n], |
2363 | struct kgem_request, |
2363 | struct kgem_request, |
2364 | list); |
2364 | list); |
2365 | while (!list_is_empty(&rq->buffers)) { |
2365 | while (!list_is_empty(&rq->buffers)) { |
2366 | struct kgem_bo *bo; |
2366 | struct kgem_bo *bo; |
2367 | 2367 | ||
2368 | bo = list_first_entry(&rq->buffers, |
2368 | bo = list_first_entry(&rq->buffers, |
2369 | struct kgem_bo, |
2369 | struct kgem_bo, |
2370 | request); |
2370 | request); |
2371 | 2371 | ||
2372 | bo->exec = NULL; |
2372 | bo->exec = NULL; |
2373 | bo->gpu_dirty = false; |
2373 | bo->gpu_dirty = false; |
2374 | __kgem_bo_clear_busy(bo); |
2374 | __kgem_bo_clear_busy(bo); |
2375 | if (bo->refcnt == 0) |
2375 | if (bo->refcnt == 0) |
2376 | kgem_bo_free(kgem, bo); |
2376 | kgem_bo_free(kgem, bo); |
2377 | } |
2377 | } |
2378 | 2378 | ||
2379 | __kgem_request_free(rq); |
2379 | __kgem_request_free(rq); |
2380 | } |
2380 | } |
2381 | } |
2381 | } |
2382 | 2382 | ||
2383 | kgem_close_inactive(kgem); |
2383 | kgem_close_inactive(kgem); |
2384 | } |
2384 | } |
2385 | 2385 | ||
2386 | static int kgem_batch_write(struct kgem *kgem, uint32_t handle, uint32_t size) |
2386 | static int kgem_batch_write(struct kgem *kgem, uint32_t handle, uint32_t size) |
2387 | { |
2387 | { |
2388 | int ret; |
2388 | int ret; |
2389 | 2389 | ||
2390 | ASSERT_IDLE(kgem, handle); |
2390 | ASSERT_IDLE(kgem, handle); |
2391 | 2391 | ||
2392 | /* If there is no surface data, just upload the batch */ |
2392 | /* If there is no surface data, just upload the batch */ |
2393 | if (kgem->surface == kgem->batch_size) |
2393 | if (kgem->surface == kgem->batch_size) |
2394 | return gem_write(kgem->fd, handle, |
2394 | return gem_write(kgem->fd, handle, |
2395 | 0, sizeof(uint32_t)*kgem->nbatch, |
2395 | 0, sizeof(uint32_t)*kgem->nbatch, |
2396 | kgem->batch); |
2396 | kgem->batch); |
2397 | 2397 | ||
2398 | /* Are the batch pages conjoint with the surface pages? */ |
2398 | /* Are the batch pages conjoint with the surface pages? */ |
2399 | if (kgem->surface < kgem->nbatch + PAGE_SIZE/sizeof(uint32_t)) { |
2399 | if (kgem->surface < kgem->nbatch + PAGE_SIZE/sizeof(uint32_t)) { |
2400 | assert(size == PAGE_ALIGN(kgem->batch_size*sizeof(uint32_t))); |
2400 | assert(size == PAGE_ALIGN(kgem->batch_size*sizeof(uint32_t))); |
2401 | return gem_write(kgem->fd, handle, |
2401 | return gem_write(kgem->fd, handle, |
2402 | 0, kgem->batch_size*sizeof(uint32_t), |
2402 | 0, kgem->batch_size*sizeof(uint32_t), |
2403 | kgem->batch); |
2403 | kgem->batch); |
2404 | } |
2404 | } |
2405 | 2405 | ||
2406 | /* Disjoint surface/batch, upload separately */ |
2406 | /* Disjoint surface/batch, upload separately */ |
2407 | ret = gem_write(kgem->fd, handle, |
2407 | ret = gem_write(kgem->fd, handle, |
2408 | 0, sizeof(uint32_t)*kgem->nbatch, |
2408 | 0, sizeof(uint32_t)*kgem->nbatch, |
2409 | kgem->batch); |
2409 | kgem->batch); |
2410 | if (ret) |
2410 | if (ret) |
2411 | return ret; |
2411 | return ret; |
2412 | 2412 | ||
2413 | ret = PAGE_ALIGN(sizeof(uint32_t) * kgem->batch_size); |
2413 | ret = PAGE_ALIGN(sizeof(uint32_t) * kgem->batch_size); |
2414 | ret -= sizeof(uint32_t) * kgem->surface; |
2414 | ret -= sizeof(uint32_t) * kgem->surface; |
2415 | assert(size-ret >= kgem->nbatch*sizeof(uint32_t)); |
2415 | assert(size-ret >= kgem->nbatch*sizeof(uint32_t)); |
2416 | return __gem_write(kgem->fd, handle, |
2416 | return __gem_write(kgem->fd, handle, |
2417 | size - ret, (kgem->batch_size - kgem->surface)*sizeof(uint32_t), |
2417 | size - ret, (kgem->batch_size - kgem->surface)*sizeof(uint32_t), |
2418 | kgem->batch + kgem->surface); |
2418 | kgem->batch + kgem->surface); |
2419 | } |
2419 | } |
2420 | 2420 | ||
2421 | void kgem_reset(struct kgem *kgem) |
2421 | void kgem_reset(struct kgem *kgem) |
2422 | { |
2422 | { |
2423 | if (kgem->next_request) { |
2423 | if (kgem->next_request) { |
2424 | struct kgem_request *rq = kgem->next_request; |
2424 | struct kgem_request *rq = kgem->next_request; |
2425 | 2425 | ||
2426 | while (!list_is_empty(&rq->buffers)) { |
2426 | while (!list_is_empty(&rq->buffers)) { |
2427 | struct kgem_bo *bo = |
2427 | struct kgem_bo *bo = |
2428 | list_first_entry(&rq->buffers, |
2428 | list_first_entry(&rq->buffers, |
2429 | struct kgem_bo, |
2429 | struct kgem_bo, |
2430 | request); |
2430 | request); |
2431 | list_del(&bo->request); |
2431 | list_del(&bo->request); |
2432 | 2432 | ||
2433 | assert(RQ(bo->rq) == rq); |
2433 | assert(RQ(bo->rq) == rq); |
2434 | 2434 | ||
2435 | bo->binding.offset = 0; |
2435 | bo->binding.offset = 0; |
2436 | bo->exec = NULL; |
2436 | bo->exec = NULL; |
2437 | bo->target_handle = -1; |
2437 | bo->target_handle = -1; |
2438 | bo->gpu_dirty = false; |
2438 | bo->gpu_dirty = false; |
2439 | 2439 | ||
2440 | if (bo->needs_flush && __kgem_busy(kgem, bo->handle)) { |
2440 | if (bo->needs_flush && __kgem_busy(kgem, bo->handle)) { |
2441 | assert(bo->domain == DOMAIN_GPU || bo->domain == DOMAIN_NONE); |
2441 | assert(bo->domain == DOMAIN_GPU || bo->domain == DOMAIN_NONE); |
2442 | list_add(&bo->request, &kgem->flushing); |
2442 | list_add(&bo->request, &kgem->flushing); |
2443 | bo->rq = (void *)kgem; |
2443 | bo->rq = (void *)kgem; |
2444 | } else |
2444 | } else |
2445 | __kgem_bo_clear_busy(bo); |
2445 | __kgem_bo_clear_busy(bo); |
2446 | 2446 | ||
2447 | if (bo->refcnt || bo->rq) |
2447 | if (bo->refcnt || bo->rq) |
2448 | continue; |
2448 | continue; |
2449 | 2449 | ||
2450 | if (bo->snoop) { |
2450 | if (bo->snoop) { |
2451 | kgem_bo_move_to_snoop(kgem, bo); |
2451 | kgem_bo_move_to_snoop(kgem, bo); |
2452 | } else if (bo->scanout) { |
2452 | } else if (bo->scanout) { |
2453 | kgem_bo_move_to_scanout(kgem, bo); |
2453 | kgem_bo_move_to_scanout(kgem, bo); |
2454 | } else if ((bo = kgem_bo_replace_io(bo))->reusable && |
2454 | } else if ((bo = kgem_bo_replace_io(bo))->reusable && |
2455 | kgem_bo_set_purgeable(kgem, bo)) { |
2455 | kgem_bo_set_purgeable(kgem, bo)) { |
2456 | kgem_bo_move_to_inactive(kgem, bo); |
2456 | kgem_bo_move_to_inactive(kgem, bo); |
2457 | } else { |
2457 | } else { |
2458 | DBG(("%s: closing %d\n", |
2458 | DBG(("%s: closing %d\n", |
2459 | __FUNCTION__, bo->handle)); |
2459 | __FUNCTION__, bo->handle)); |
2460 | kgem_bo_free(kgem, bo); |
2460 | kgem_bo_free(kgem, bo); |
2461 | } |
2461 | } |
2462 | } |
2462 | } |
2463 | 2463 | ||
2464 | if (rq != &kgem->static_request) { |
2464 | if (rq != &kgem->static_request) { |
2465 | list_init(&rq->list); |
2465 | list_init(&rq->list); |
2466 | __kgem_request_free(rq); |
2466 | __kgem_request_free(rq); |
2467 | } |
2467 | } |
2468 | } |
2468 | } |
2469 | 2469 | ||
2470 | kgem->nfence = 0; |
2470 | kgem->nfence = 0; |
2471 | kgem->nexec = 0; |
2471 | kgem->nexec = 0; |
2472 | kgem->nreloc = 0; |
2472 | kgem->nreloc = 0; |
2473 | kgem->nreloc__self = 0; |
2473 | kgem->nreloc__self = 0; |
2474 | kgem->aperture = 0; |
2474 | kgem->aperture = 0; |
2475 | kgem->aperture_fenced = 0; |
2475 | kgem->aperture_fenced = 0; |
2476 | kgem->nbatch = 0; |
2476 | kgem->nbatch = 0; |
2477 | kgem->surface = kgem->batch_size; |
2477 | kgem->surface = kgem->batch_size; |
2478 | kgem->mode = KGEM_NONE; |
2478 | kgem->mode = KGEM_NONE; |
2479 | kgem->flush = 0; |
2479 | kgem->flush = 0; |
2480 | kgem->batch_flags = kgem->batch_flags_base; |
2480 | kgem->batch_flags = kgem->batch_flags_base; |
2481 | 2481 | ||
2482 | kgem->next_request = __kgem_request_alloc(kgem); |
2482 | kgem->next_request = __kgem_request_alloc(kgem); |
2483 | 2483 | ||
2484 | kgem_sna_reset(kgem); |
2484 | kgem_sna_reset(kgem); |
2485 | } |
2485 | } |
2486 | 2486 | ||
2487 | static int compact_batch_surface(struct kgem *kgem) |
2487 | static int compact_batch_surface(struct kgem *kgem) |
2488 | { |
2488 | { |
2489 | int size, shrink, n; |
2489 | int size, shrink, n; |
2490 | 2490 | ||
2491 | if (!kgem->has_relaxed_delta) |
2491 | if (!kgem->has_relaxed_delta) |
2492 | return kgem->batch_size; |
2492 | return kgem->batch_size; |
2493 | 2493 | ||
2494 | /* See if we can pack the contents into one or two pages */ |
2494 | /* See if we can pack the contents into one or two pages */ |
2495 | n = ALIGN(kgem->batch_size, 1024); |
2495 | n = ALIGN(kgem->batch_size, 1024); |
2496 | size = n - kgem->surface + kgem->nbatch; |
2496 | size = n - kgem->surface + kgem->nbatch; |
2497 | size = ALIGN(size, 1024); |
2497 | size = ALIGN(size, 1024); |
2498 | 2498 | ||
2499 | shrink = n - size; |
2499 | shrink = n - size; |
2500 | if (shrink) { |
2500 | if (shrink) { |
2501 | DBG(("shrinking from %d to %d\n", kgem->batch_size, size)); |
2501 | DBG(("shrinking from %d to %d\n", kgem->batch_size, size)); |
2502 | 2502 | ||
2503 | shrink *= sizeof(uint32_t); |
2503 | shrink *= sizeof(uint32_t); |
2504 | for (n = 0; n < kgem->nreloc; n++) { |
2504 | for (n = 0; n < kgem->nreloc; n++) { |
2505 | if (kgem->reloc[n].read_domains == I915_GEM_DOMAIN_INSTRUCTION && |
2505 | if (kgem->reloc[n].read_domains == I915_GEM_DOMAIN_INSTRUCTION && |
2506 | kgem->reloc[n].target_handle == ~0U) |
2506 | kgem->reloc[n].target_handle == ~0U) |
2507 | kgem->reloc[n].delta -= shrink; |
2507 | kgem->reloc[n].delta -= shrink; |
2508 | 2508 | ||
2509 | if (kgem->reloc[n].offset >= sizeof(uint32_t)*kgem->nbatch) |
2509 | if (kgem->reloc[n].offset >= sizeof(uint32_t)*kgem->nbatch) |
2510 | kgem->reloc[n].offset -= shrink; |
2510 | kgem->reloc[n].offset -= shrink; |
2511 | } |
2511 | } |
2512 | } |
2512 | } |
2513 | 2513 | ||
2514 | return size * sizeof(uint32_t); |
2514 | return size * sizeof(uint32_t); |
2515 | } |
2515 | } |
2516 | 2516 | ||
2517 | static struct kgem_bo * |
2517 | static struct kgem_bo * |
2518 | kgem_create_batch(struct kgem *kgem, int size) |
2518 | kgem_create_batch(struct kgem *kgem, int size) |
2519 | { |
2519 | { |
2520 | struct drm_i915_gem_set_domain set_domain; |
2520 | struct drm_i915_gem_set_domain set_domain; |
2521 | struct kgem_bo *bo; |
2521 | struct kgem_bo *bo; |
2522 | 2522 | ||
2523 | if (size <= 4096) { |
2523 | if (size <= 4096) { |
2524 | bo = list_first_entry(&kgem->pinned_batches[0], |
2524 | bo = list_first_entry(&kgem->pinned_batches[0], |
2525 | struct kgem_bo, |
2525 | struct kgem_bo, |
2526 | list); |
2526 | list); |
2527 | if (!bo->rq) { |
2527 | if (!bo->rq) { |
2528 | out_4096: |
2528 | out_4096: |
2529 | list_move_tail(&bo->list, &kgem->pinned_batches[0]); |
2529 | list_move_tail(&bo->list, &kgem->pinned_batches[0]); |
2530 | return kgem_bo_reference(bo); |
2530 | return kgem_bo_reference(bo); |
2531 | } |
2531 | } |
2532 | 2532 | ||
2533 | if (!__kgem_busy(kgem, bo->handle)) { |
2533 | if (!__kgem_busy(kgem, bo->handle)) { |
2534 | assert(RQ(bo->rq)->bo == bo); |
2534 | assert(RQ(bo->rq)->bo == bo); |
2535 | __kgem_retire_rq(kgem, RQ(bo->rq)); |
2535 | __kgem_retire_rq(kgem, RQ(bo->rq)); |
2536 | goto out_4096; |
2536 | goto out_4096; |
2537 | } |
2537 | } |
2538 | } |
2538 | } |
2539 | 2539 | ||
2540 | if (size <= 16384) { |
2540 | if (size <= 16384) { |
2541 | bo = list_first_entry(&kgem->pinned_batches[1], |
2541 | bo = list_first_entry(&kgem->pinned_batches[1], |
2542 | struct kgem_bo, |
2542 | struct kgem_bo, |
2543 | list); |
2543 | list); |
2544 | if (!bo->rq) { |
2544 | if (!bo->rq) { |
2545 | out_16384: |
2545 | out_16384: |
2546 | list_move_tail(&bo->list, &kgem->pinned_batches[1]); |
2546 | list_move_tail(&bo->list, &kgem->pinned_batches[1]); |
2547 | return kgem_bo_reference(bo); |
2547 | return kgem_bo_reference(bo); |
2548 | } |
2548 | } |
2549 | 2549 | ||
2550 | if (!__kgem_busy(kgem, bo->handle)) { |
2550 | if (!__kgem_busy(kgem, bo->handle)) { |
2551 | assert(RQ(bo->rq)->bo == bo); |
2551 | assert(RQ(bo->rq)->bo == bo); |
2552 | __kgem_retire_rq(kgem, RQ(bo->rq)); |
2552 | __kgem_retire_rq(kgem, RQ(bo->rq)); |
2553 | goto out_16384; |
2553 | goto out_16384; |
2554 | } |
2554 | } |
2555 | } |
2555 | } |
2556 | 2556 | ||
2557 | if (kgem->gen == 020 && !kgem->has_pinned_batches) { |
2557 | if (kgem->gen == 020 && !kgem->has_pinned_batches) { |
2558 | assert(size <= 16384); |
2558 | assert(size <= 16384); |
2559 | 2559 | ||
2560 | bo = list_first_entry(&kgem->pinned_batches[size > 4096], |
2560 | bo = list_first_entry(&kgem->pinned_batches[size > 4096], |
2561 | struct kgem_bo, |
2561 | struct kgem_bo, |
2562 | list); |
2562 | list); |
2563 | list_move_tail(&bo->list, &kgem->pinned_batches[size > 4096]); |
2563 | list_move_tail(&bo->list, &kgem->pinned_batches[size > 4096]); |
2564 | 2564 | ||
2565 | DBG(("%s: syncing due to busy batches\n", __FUNCTION__)); |
2565 | DBG(("%s: syncing due to busy batches\n", __FUNCTION__)); |
2566 | 2566 | ||
2567 | VG_CLEAR(set_domain); |
2567 | VG_CLEAR(set_domain); |
2568 | set_domain.handle = bo->handle; |
2568 | set_domain.handle = bo->handle; |
2569 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
2569 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
2570 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
2570 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
2571 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) { |
2571 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) { |
2572 | DBG(("%s: sync: GPU hang detected\n", __FUNCTION__)); |
2572 | DBG(("%s: sync: GPU hang detected\n", __FUNCTION__)); |
2573 | kgem_throttle(kgem); |
2573 | kgem_throttle(kgem); |
2574 | return NULL; |
2574 | return NULL; |
2575 | } |
2575 | } |
2576 | 2576 | ||
2577 | kgem_retire(kgem); |
2577 | kgem_retire(kgem); |
2578 | assert(bo->rq == NULL); |
2578 | assert(bo->rq == NULL); |
2579 | return kgem_bo_reference(bo); |
2579 | return kgem_bo_reference(bo); |
2580 | } |
2580 | } |
2581 | 2581 | ||
2582 | return kgem_create_linear(kgem, size, CREATE_NO_THROTTLE); |
2582 | return kgem_create_linear(kgem, size, CREATE_NO_THROTTLE); |
2583 | } |
2583 | } |
2584 | 2584 | ||
2585 | void _kgem_submit(struct kgem *kgem) |
2585 | void _kgem_submit(struct kgem *kgem) |
2586 | { |
2586 | { |
2587 | struct kgem_request *rq; |
2587 | struct kgem_request *rq; |
2588 | uint32_t batch_end; |
2588 | uint32_t batch_end; |
2589 | int size; |
2589 | int size; |
2590 | 2590 | ||
2591 | assert(!DBG_NO_HW); |
2591 | assert(!DBG_NO_HW); |
2592 | assert(!kgem->wedged); |
2592 | assert(!kgem->wedged); |
2593 | 2593 | ||
2594 | assert(kgem->nbatch); |
2594 | assert(kgem->nbatch); |
2595 | assert(kgem->nbatch <= KGEM_BATCH_SIZE(kgem)); |
2595 | assert(kgem->nbatch <= KGEM_BATCH_SIZE(kgem)); |
2596 | assert(kgem->nbatch <= kgem->surface); |
2596 | assert(kgem->nbatch <= kgem->surface); |
2597 | 2597 | ||
2598 | batch_end = kgem_end_batch(kgem); |
2598 | batch_end = kgem_end_batch(kgem); |
2599 | kgem_sna_flush(kgem); |
2599 | kgem_sna_flush(kgem); |
2600 | 2600 | ||
2601 | DBG(("batch[%d/%d, flags=%x]: %d %d %d %d, nreloc=%d, nexec=%d, nfence=%d, aperture=%d\n", |
2601 | DBG(("batch[%d/%d, flags=%x]: %d %d %d %d, nreloc=%d, nexec=%d, nfence=%d, aperture=%d\n", |
2602 | kgem->mode, kgem->ring, kgem->batch_flags, |
2602 | kgem->mode, kgem->ring, kgem->batch_flags, |
2603 | batch_end, kgem->nbatch, kgem->surface, kgem->batch_size, |
2603 | batch_end, kgem->nbatch, kgem->surface, kgem->batch_size, |
2604 | kgem->nreloc, kgem->nexec, kgem->nfence, kgem->aperture)); |
2604 | kgem->nreloc, kgem->nexec, kgem->nfence, kgem->aperture)); |
2605 | 2605 | ||
2606 | assert(kgem->nbatch <= kgem->batch_size); |
2606 | assert(kgem->nbatch <= kgem->batch_size); |
2607 | assert(kgem->nbatch <= kgem->surface); |
2607 | assert(kgem->nbatch <= kgem->surface); |
2608 | assert(kgem->nreloc <= ARRAY_SIZE(kgem->reloc)); |
2608 | assert(kgem->nreloc <= ARRAY_SIZE(kgem->reloc)); |
2609 | assert(kgem->nexec < ARRAY_SIZE(kgem->exec)); |
2609 | assert(kgem->nexec < ARRAY_SIZE(kgem->exec)); |
2610 | assert(kgem->nfence <= kgem->fence_max); |
2610 | assert(kgem->nfence <= kgem->fence_max); |
2611 | 2611 | ||
2612 | kgem_finish_buffers(kgem); |
2612 | kgem_finish_buffers(kgem); |
2613 | 2613 | ||
2614 | #if SHOW_BATCH |
2614 | #if SHOW_BATCH |
2615 | __kgem_batch_debug(kgem, batch_end); |
2615 | __kgem_batch_debug(kgem, batch_end); |
2616 | #endif |
2616 | #endif |
2617 | 2617 | ||
2618 | rq = kgem->next_request; |
2618 | rq = kgem->next_request; |
2619 | if (kgem->surface != kgem->batch_size) |
2619 | if (kgem->surface != kgem->batch_size) |
2620 | size = compact_batch_surface(kgem); |
2620 | size = compact_batch_surface(kgem); |
2621 | else |
2621 | else |
2622 | size = kgem->nbatch * sizeof(kgem->batch[0]); |
2622 | size = kgem->nbatch * sizeof(kgem->batch[0]); |
2623 | rq->bo = kgem_create_batch(kgem, size); |
2623 | rq->bo = kgem_create_batch(kgem, size); |
2624 | if (rq->bo) { |
2624 | if (rq->bo) { |
2625 | uint32_t handle = rq->bo->handle; |
2625 | uint32_t handle = rq->bo->handle; |
2626 | int i; |
2626 | int i; |
2627 | 2627 | ||
2628 | assert(!rq->bo->needs_flush); |
2628 | assert(!rq->bo->needs_flush); |
2629 | 2629 | ||
2630 | i = kgem->nexec++; |
2630 | i = kgem->nexec++; |
2631 | kgem->exec[i].handle = handle; |
2631 | kgem->exec[i].handle = handle; |
2632 | kgem->exec[i].relocation_count = kgem->nreloc; |
2632 | kgem->exec[i].relocation_count = kgem->nreloc; |
2633 | kgem->exec[i].relocs_ptr = (uintptr_t)kgem->reloc; |
2633 | kgem->exec[i].relocs_ptr = (uintptr_t)kgem->reloc; |
2634 | kgem->exec[i].alignment = 0; |
2634 | kgem->exec[i].alignment = 0; |
2635 | kgem->exec[i].offset = rq->bo->presumed_offset; |
2635 | kgem->exec[i].offset = rq->bo->presumed_offset; |
2636 | kgem->exec[i].flags = 0; |
2636 | kgem->exec[i].flags = 0; |
2637 | kgem->exec[i].rsvd1 = 0; |
2637 | kgem->exec[i].rsvd1 = 0; |
2638 | kgem->exec[i].rsvd2 = 0; |
2638 | kgem->exec[i].rsvd2 = 0; |
2639 | 2639 | ||
2640 | rq->bo->target_handle = kgem->has_handle_lut ? i : handle; |
2640 | rq->bo->target_handle = kgem->has_handle_lut ? i : handle; |
2641 | rq->bo->exec = &kgem->exec[i]; |
2641 | rq->bo->exec = &kgem->exec[i]; |
2642 | rq->bo->rq = MAKE_REQUEST(rq, kgem->ring); /* useful sanity check */ |
2642 | rq->bo->rq = MAKE_REQUEST(rq, kgem->ring); /* useful sanity check */ |
2643 | list_add(&rq->bo->request, &rq->buffers); |
2643 | list_add(&rq->bo->request, &rq->buffers); |
2644 | rq->ring = kgem->ring == KGEM_BLT; |
2644 | rq->ring = kgem->ring == KGEM_BLT; |
2645 | 2645 | ||
2646 | kgem_fixup_self_relocs(kgem, rq->bo); |
2646 | kgem_fixup_self_relocs(kgem, rq->bo); |
2647 | 2647 | ||
2648 | if (kgem_batch_write(kgem, handle, size) == 0) { |
2648 | if (kgem_batch_write(kgem, handle, size) == 0) { |
2649 | struct drm_i915_gem_execbuffer2 execbuf; |
2649 | struct drm_i915_gem_execbuffer2 execbuf; |
2650 | int ret, retry = 3; |
2650 | int ret, retry = 3; |
2651 | 2651 | ||
2652 | memset(&execbuf, 0, sizeof(execbuf)); |
2652 | memset(&execbuf, 0, sizeof(execbuf)); |
2653 | execbuf.buffers_ptr = (uintptr_t)kgem->exec; |
2653 | execbuf.buffers_ptr = (uintptr_t)kgem->exec; |
2654 | execbuf.buffer_count = kgem->nexec; |
2654 | execbuf.buffer_count = kgem->nexec; |
2655 | execbuf.batch_len = batch_end*sizeof(uint32_t); |
2655 | execbuf.batch_len = batch_end*sizeof(uint32_t); |
2656 | execbuf.flags = kgem->ring | kgem->batch_flags; |
2656 | execbuf.flags = kgem->ring | kgem->batch_flags; |
2657 | 2657 | ||
2658 | 2658 | ||
2659 | ret = drmIoctl(kgem->fd, |
2659 | ret = drmIoctl(kgem->fd, |
2660 | DRM_IOCTL_I915_GEM_EXECBUFFER2, |
2660 | DRM_IOCTL_I915_GEM_EXECBUFFER2, |
2661 | &execbuf); |
2661 | &execbuf); |
2662 | while (ret == -1 && errno == EBUSY && retry--) { |
2662 | while (ret == -1 && errno == EBUSY && retry--) { |
2663 | __kgem_throttle(kgem); |
2663 | __kgem_throttle(kgem); |
2664 | ret = drmIoctl(kgem->fd, |
2664 | ret = drmIoctl(kgem->fd, |
2665 | DRM_IOCTL_I915_GEM_EXECBUFFER2, |
2665 | DRM_IOCTL_I915_GEM_EXECBUFFER2, |
2666 | &execbuf); |
2666 | &execbuf); |
2667 | } |
2667 | } |
2668 | if (DEBUG_SYNC && ret == 0) { |
2668 | if (DEBUG_SYNC && ret == 0) { |
2669 | struct drm_i915_gem_set_domain set_domain; |
2669 | struct drm_i915_gem_set_domain set_domain; |
2670 | 2670 | ||
2671 | VG_CLEAR(set_domain); |
2671 | VG_CLEAR(set_domain); |
2672 | set_domain.handle = handle; |
2672 | set_domain.handle = handle; |
2673 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
2673 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
2674 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
2674 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
2675 | 2675 | ||
2676 | ret = drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain); |
2676 | ret = drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain); |
2677 | } |
2677 | } |
2678 | if (ret == -1) { |
2678 | if (ret == -1) { |
2679 | // DBG(("%s: GPU hang detected [%d]\n", |
2679 | // DBG(("%s: GPU hang detected [%d]\n", |
2680 | // __FUNCTION__, errno)); |
2680 | // __FUNCTION__, errno)); |
2681 | kgem_throttle(kgem); |
2681 | kgem_throttle(kgem); |
2682 | kgem->wedged = true; |
2682 | kgem->wedged = true; |
2683 | 2683 | ||
2684 | #if 0 |
2684 | #if 0 |
2685 | ret = errno; |
2685 | ret = errno; |
2686 | ErrorF("batch[%d/%d]: %d %d %d, nreloc=%d, nexec=%d, nfence=%d, aperture=%d: errno=%d\n", |
2686 | ErrorF("batch[%d/%d]: %d %d %d, nreloc=%d, nexec=%d, nfence=%d, aperture=%d: errno=%d\n", |
2687 | kgem->mode, kgem->ring, batch_end, kgem->nbatch, kgem->surface, |
2687 | kgem->mode, kgem->ring, batch_end, kgem->nbatch, kgem->surface, |
2688 | kgem->nreloc, kgem->nexec, kgem->nfence, kgem->aperture, errno); |
2688 | kgem->nreloc, kgem->nexec, kgem->nfence, kgem->aperture, errno); |
2689 | 2689 | ||
2690 | for (i = 0; i < kgem->nexec; i++) { |
2690 | for (i = 0; i < kgem->nexec; i++) { |
2691 | struct kgem_bo *bo, *found = NULL; |
2691 | struct kgem_bo *bo, *found = NULL; |
2692 | 2692 | ||
2693 | list_for_each_entry(bo, &kgem->next_request->buffers, request) { |
2693 | list_for_each_entry(bo, &kgem->next_request->buffers, request) { |
2694 | if (bo->handle == kgem->exec[i].handle) { |
2694 | if (bo->handle == kgem->exec[i].handle) { |
2695 | found = bo; |
2695 | found = bo; |
2696 | break; |
2696 | break; |
2697 | } |
2697 | } |
2698 | } |
2698 | } |
2699 | ErrorF("exec[%d] = handle:%d, presumed offset: %x, size: %d, tiling %d, fenced %d, snooped %d, deleted %d\n", |
2699 | ErrorF("exec[%d] = handle:%d, presumed offset: %x, size: %d, tiling %d, fenced %d, snooped %d, deleted %d\n", |
2700 | i, |
2700 | i, |
2701 | kgem->exec[i].handle, |
2701 | kgem->exec[i].handle, |
2702 | (int)kgem->exec[i].offset, |
2702 | (int)kgem->exec[i].offset, |
2703 | found ? kgem_bo_size(found) : -1, |
2703 | found ? kgem_bo_size(found) : -1, |
2704 | found ? found->tiling : -1, |
2704 | found ? found->tiling : -1, |
2705 | (int)(kgem->exec[i].flags & EXEC_OBJECT_NEEDS_FENCE), |
2705 | (int)(kgem->exec[i].flags & EXEC_OBJECT_NEEDS_FENCE), |
2706 | found ? found->snoop : -1, |
2706 | found ? found->snoop : -1, |
2707 | found ? found->purged : -1); |
2707 | found ? found->purged : -1); |
2708 | } |
2708 | } |
2709 | for (i = 0; i < kgem->nreloc; i++) { |
2709 | for (i = 0; i < kgem->nreloc; i++) { |
2710 | ErrorF("reloc[%d] = pos:%d, target:%d, delta:%d, read:%x, write:%x, offset:%x\n", |
2710 | ErrorF("reloc[%d] = pos:%d, target:%d, delta:%d, read:%x, write:%x, offset:%x\n", |
2711 | i, |
2711 | i, |
2712 | (int)kgem->reloc[i].offset, |
2712 | (int)kgem->reloc[i].offset, |
2713 | kgem->reloc[i].target_handle, |
2713 | kgem->reloc[i].target_handle, |
2714 | kgem->reloc[i].delta, |
2714 | kgem->reloc[i].delta, |
2715 | kgem->reloc[i].read_domains, |
2715 | kgem->reloc[i].read_domains, |
2716 | kgem->reloc[i].write_domain, |
2716 | kgem->reloc[i].write_domain, |
2717 | (int)kgem->reloc[i].presumed_offset); |
2717 | (int)kgem->reloc[i].presumed_offset); |
2718 | } |
2718 | } |
2719 | 2719 | ||
2720 | if (DEBUG_SYNC) { |
2720 | if (DEBUG_SYNC) { |
2721 | int fd = open("/tmp/batchbuffer", O_WRONLY | O_CREAT | O_APPEND, 0666); |
2721 | int fd = open("/tmp/batchbuffer", O_WRONLY | O_CREAT | O_APPEND, 0666); |
2722 | if (fd != -1) { |
2722 | if (fd != -1) { |
2723 | write(fd, kgem->batch, batch_end*sizeof(uint32_t)); |
2723 | write(fd, kgem->batch, batch_end*sizeof(uint32_t)); |
2724 | close(fd); |
2724 | close(fd); |
2725 | } |
2725 | } |
2726 | 2726 | ||
2727 | FatalError("SNA: failed to submit batchbuffer, errno=%d\n", ret); |
2727 | FatalError("SNA: failed to submit batchbuffer, errno=%d\n", ret); |
2728 | } |
2728 | } |
2729 | #endif |
2729 | #endif |
2730 | } |
2730 | } |
2731 | } |
2731 | } |
2732 | 2732 | ||
2733 | kgem_commit(kgem); |
2733 | kgem_commit(kgem); |
2734 | } |
2734 | } |
2735 | if (kgem->wedged) |
2735 | if (kgem->wedged) |
2736 | kgem_cleanup(kgem); |
2736 | kgem_cleanup(kgem); |
2737 | 2737 | ||
2738 | kgem_reset(kgem); |
2738 | kgem_reset(kgem); |
2739 | 2739 | ||
2740 | assert(kgem->next_request != NULL); |
2740 | assert(kgem->next_request != NULL); |
2741 | } |
2741 | } |
2742 | 2742 | ||
2743 | void kgem_throttle(struct kgem *kgem) |
2743 | void kgem_throttle(struct kgem *kgem) |
2744 | { |
2744 | { |
2745 | kgem->need_throttle = 0; |
2745 | kgem->need_throttle = 0; |
2746 | if (kgem->wedged) |
2746 | if (kgem->wedged) |
2747 | return; |
2747 | return; |
2748 | 2748 | ||
2749 | kgem->wedged = __kgem_throttle(kgem); |
2749 | kgem->wedged = __kgem_throttle(kgem); |
2750 | if (kgem->wedged) { |
2750 | if (kgem->wedged) { |
2751 | printf("Detected a hung GPU, disabling acceleration.\n"); |
2751 | printf("Detected a hung GPU, disabling acceleration.\n"); |
2752 | printf("When reporting this, please include i915_error_state from debugfs and the full dmesg.\n"); |
2752 | printf("When reporting this, please include i915_error_state from debugfs and the full dmesg.\n"); |
2753 | } |
2753 | } |
2754 | } |
2754 | } |
2755 | 2755 | ||
2756 | void kgem_purge_cache(struct kgem *kgem) |
2756 | void kgem_purge_cache(struct kgem *kgem) |
2757 | { |
2757 | { |
2758 | struct kgem_bo *bo, *next; |
2758 | struct kgem_bo *bo, *next; |
2759 | int i; |
2759 | int i; |
2760 | 2760 | ||
2761 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) { |
2761 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) { |
2762 | list_for_each_entry_safe(bo, next, &kgem->inactive[i], list) { |
2762 | list_for_each_entry_safe(bo, next, &kgem->inactive[i], list) { |
2763 | if (!kgem_bo_is_retained(kgem, bo)) { |
2763 | if (!kgem_bo_is_retained(kgem, bo)) { |
2764 | DBG(("%s: purging %d\n", |
2764 | DBG(("%s: purging %d\n", |
2765 | __FUNCTION__, bo->handle)); |
2765 | __FUNCTION__, bo->handle)); |
2766 | kgem_bo_free(kgem, bo); |
2766 | kgem_bo_free(kgem, bo); |
2767 | } |
2767 | } |
2768 | } |
2768 | } |
2769 | } |
2769 | } |
2770 | 2770 | ||
2771 | kgem->need_purge = false; |
2771 | kgem->need_purge = false; |
2772 | } |
2772 | } |
2773 | 2773 | ||
2774 | 2774 | ||
2775 | void kgem_clean_large_cache(struct kgem *kgem) |
2775 | void kgem_clean_large_cache(struct kgem *kgem) |
2776 | { |
2776 | { |
2777 | while (!list_is_empty(&kgem->large_inactive)) { |
2777 | while (!list_is_empty(&kgem->large_inactive)) { |
2778 | kgem_bo_free(kgem, |
2778 | kgem_bo_free(kgem, |
2779 | list_first_entry(&kgem->large_inactive, |
2779 | list_first_entry(&kgem->large_inactive, |
2780 | struct kgem_bo, list)); |
2780 | struct kgem_bo, list)); |
2781 | 2781 | ||
2782 | } |
2782 | } |
2783 | } |
2783 | } |
2784 | 2784 | ||
2785 | bool kgem_expire_cache(struct kgem *kgem) |
2785 | bool kgem_expire_cache(struct kgem *kgem) |
2786 | { |
2786 | { |
2787 | time_t now, expire; |
2787 | time_t now, expire; |
2788 | struct kgem_bo *bo; |
2788 | struct kgem_bo *bo; |
2789 | unsigned int size = 0, count = 0; |
2789 | unsigned int size = 0, count = 0; |
2790 | bool idle; |
2790 | bool idle; |
2791 | unsigned int i; |
2791 | unsigned int i; |
2792 | 2792 | ||
2793 | time(&now); |
2793 | time(&now); |
2794 | 2794 | ||
2795 | while (__kgem_freed_bo) { |
2795 | while (__kgem_freed_bo) { |
2796 | bo = __kgem_freed_bo; |
2796 | bo = __kgem_freed_bo; |
2797 | __kgem_freed_bo = *(struct kgem_bo **)bo; |
2797 | __kgem_freed_bo = *(struct kgem_bo **)bo; |
2798 | free(bo); |
2798 | free(bo); |
2799 | } |
2799 | } |
2800 | 2800 | ||
2801 | while (__kgem_freed_request) { |
2801 | while (__kgem_freed_request) { |
2802 | struct kgem_request *rq = __kgem_freed_request; |
2802 | struct kgem_request *rq = __kgem_freed_request; |
2803 | __kgem_freed_request = *(struct kgem_request **)rq; |
2803 | __kgem_freed_request = *(struct kgem_request **)rq; |
2804 | free(rq); |
2804 | free(rq); |
2805 | } |
2805 | } |
2806 | 2806 | ||
2807 | kgem_clean_large_cache(kgem); |
2807 | kgem_clean_large_cache(kgem); |
2808 | 2808 | ||
2809 | expire = 0; |
2809 | expire = 0; |
2810 | list_for_each_entry(bo, &kgem->snoop, list) { |
2810 | list_for_each_entry(bo, &kgem->snoop, list) { |
2811 | if (bo->delta) { |
2811 | if (bo->delta) { |
2812 | expire = now - MAX_INACTIVE_TIME/2; |
2812 | expire = now - MAX_INACTIVE_TIME/2; |
2813 | break; |
2813 | break; |
2814 | } |
2814 | } |
2815 | 2815 | ||
2816 | bo->delta = now; |
2816 | bo->delta = now; |
2817 | } |
2817 | } |
2818 | if (expire) { |
2818 | if (expire) { |
2819 | while (!list_is_empty(&kgem->snoop)) { |
2819 | while (!list_is_empty(&kgem->snoop)) { |
2820 | bo = list_last_entry(&kgem->snoop, struct kgem_bo, list); |
2820 | bo = list_last_entry(&kgem->snoop, struct kgem_bo, list); |
2821 | 2821 | ||
2822 | if (bo->delta > expire) |
2822 | if (bo->delta > expire) |
2823 | break; |
2823 | break; |
2824 | 2824 | ||
2825 | kgem_bo_free(kgem, bo); |
2825 | kgem_bo_free(kgem, bo); |
2826 | } |
2826 | } |
2827 | } |
2827 | } |
2828 | #ifdef DEBUG_MEMORY |
2828 | #ifdef DEBUG_MEMORY |
2829 | { |
2829 | { |
2830 | long snoop_size = 0; |
2830 | long snoop_size = 0; |
2831 | int snoop_count = 0; |
2831 | int snoop_count = 0; |
2832 | list_for_each_entry(bo, &kgem->snoop, list) |
2832 | list_for_each_entry(bo, &kgem->snoop, list) |
2833 | snoop_count++, snoop_size += bytes(bo); |
2833 | snoop_count++, snoop_size += bytes(bo); |
2834 | ErrorF("%s: still allocated %d bo, %ld bytes, in snoop cache\n", |
2834 | ErrorF("%s: still allocated %d bo, %ld bytes, in snoop cache\n", |
2835 | __FUNCTION__, snoop_count, snoop_size); |
2835 | __FUNCTION__, snoop_count, snoop_size); |
2836 | } |
2836 | } |
2837 | #endif |
2837 | #endif |
2838 | 2838 | ||
2839 | kgem_retire(kgem); |
2839 | kgem_retire(kgem); |
2840 | if (kgem->wedged) |
2840 | if (kgem->wedged) |
2841 | kgem_cleanup(kgem); |
2841 | kgem_cleanup(kgem); |
2842 | 2842 | ||
2843 | kgem->expire(kgem); |
2843 | kgem->expire(kgem); |
2844 | 2844 | ||
2845 | if (kgem->need_purge) |
2845 | if (kgem->need_purge) |
2846 | kgem_purge_cache(kgem); |
2846 | kgem_purge_cache(kgem); |
2847 | 2847 | ||
2848 | expire = 0; |
2848 | expire = 0; |
2849 | 2849 | ||
2850 | idle = !kgem->need_retire; |
2850 | idle = !kgem->need_retire; |
2851 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) { |
2851 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) { |
2852 | idle &= list_is_empty(&kgem->inactive[i]); |
2852 | idle &= list_is_empty(&kgem->inactive[i]); |
2853 | list_for_each_entry(bo, &kgem->inactive[i], list) { |
2853 | list_for_each_entry(bo, &kgem->inactive[i], list) { |
2854 | if (bo->delta) { |
2854 | if (bo->delta) { |
2855 | expire = now - MAX_INACTIVE_TIME; |
2855 | expire = now - MAX_INACTIVE_TIME; |
2856 | break; |
2856 | break; |
2857 | } |
2857 | } |
2858 | 2858 | ||
2859 | bo->delta = now; |
2859 | bo->delta = now; |
2860 | } |
2860 | } |
2861 | } |
2861 | } |
2862 | if (idle) { |
2862 | if (idle) { |
2863 | DBG(("%s: idle\n", __FUNCTION__)); |
2863 | DBG(("%s: idle\n", __FUNCTION__)); |
2864 | kgem->need_expire = false; |
2864 | kgem->need_expire = false; |
2865 | return false; |
2865 | return false; |
2866 | } |
2866 | } |
2867 | if (expire == 0) |
2867 | if (expire == 0) |
2868 | return true; |
2868 | return true; |
2869 | 2869 | ||
2870 | idle = !kgem->need_retire; |
2870 | idle = !kgem->need_retire; |
2871 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) { |
2871 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) { |
2872 | struct list preserve; |
2872 | struct list preserve; |
2873 | 2873 | ||
2874 | list_init(&preserve); |
2874 | list_init(&preserve); |
2875 | while (!list_is_empty(&kgem->inactive[i])) { |
2875 | while (!list_is_empty(&kgem->inactive[i])) { |
2876 | bo = list_last_entry(&kgem->inactive[i], |
2876 | bo = list_last_entry(&kgem->inactive[i], |
2877 | struct kgem_bo, list); |
2877 | struct kgem_bo, list); |
2878 | 2878 | ||
2879 | if (bo->delta > expire) { |
2879 | if (bo->delta > expire) { |
2880 | idle = false; |
2880 | idle = false; |
2881 | break; |
2881 | break; |
2882 | } |
2882 | } |
2883 | 2883 | ||
2884 | if (bo->map && bo->delta + MAP_PRESERVE_TIME > expire) { |
2884 | if (bo->map && bo->delta + MAP_PRESERVE_TIME > expire) { |
2885 | idle = false; |
2885 | idle = false; |
2886 | list_move_tail(&bo->list, &preserve); |
2886 | list_move_tail(&bo->list, &preserve); |
2887 | } else { |
2887 | } else { |
2888 | count++; |
2888 | count++; |
2889 | size += bytes(bo); |
2889 | size += bytes(bo); |
2890 | kgem_bo_free(kgem, bo); |
2890 | kgem_bo_free(kgem, bo); |
2891 | DBG(("%s: expiring %d\n", |
2891 | DBG(("%s: expiring %d\n", |
2892 | __FUNCTION__, bo->handle)); |
2892 | __FUNCTION__, bo->handle)); |
2893 | } |
2893 | } |
2894 | } |
2894 | } |
2895 | if (!list_is_empty(&preserve)) { |
2895 | if (!list_is_empty(&preserve)) { |
2896 | preserve.prev->next = kgem->inactive[i].next; |
2896 | preserve.prev->next = kgem->inactive[i].next; |
2897 | kgem->inactive[i].next->prev = preserve.prev; |
2897 | kgem->inactive[i].next->prev = preserve.prev; |
2898 | kgem->inactive[i].next = preserve.next; |
2898 | kgem->inactive[i].next = preserve.next; |
2899 | preserve.next->prev = &kgem->inactive[i]; |
2899 | preserve.next->prev = &kgem->inactive[i]; |
2900 | } |
2900 | } |
2901 | } |
2901 | } |
2902 | 2902 | ||
2903 | #ifdef DEBUG_MEMORY |
2903 | #ifdef DEBUG_MEMORY |
2904 | { |
2904 | { |
2905 | long inactive_size = 0; |
2905 | long inactive_size = 0; |
2906 | int inactive_count = 0; |
2906 | int inactive_count = 0; |
2907 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) |
2907 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) |
2908 | list_for_each_entry(bo, &kgem->inactive[i], list) |
2908 | list_for_each_entry(bo, &kgem->inactive[i], list) |
2909 | inactive_count++, inactive_size += bytes(bo); |
2909 | inactive_count++, inactive_size += bytes(bo); |
2910 | ErrorF("%s: still allocated %d bo, %ld bytes, in inactive cache\n", |
2910 | ErrorF("%s: still allocated %d bo, %ld bytes, in inactive cache\n", |
2911 | __FUNCTION__, inactive_count, inactive_size); |
2911 | __FUNCTION__, inactive_count, inactive_size); |
2912 | } |
2912 | } |
2913 | #endif |
2913 | #endif |
2914 | 2914 | ||
2915 | DBG(("%s: expired %d objects, %d bytes, idle? %d\n", |
2915 | DBG(("%s: expired %d objects, %d bytes, idle? %d\n", |
2916 | __FUNCTION__, count, size, idle)); |
2916 | __FUNCTION__, count, size, idle)); |
2917 | 2917 | ||
2918 | kgem->need_expire = !idle; |
2918 | kgem->need_expire = !idle; |
2919 | return !idle; |
2919 | return !idle; |
2920 | (void)count; |
2920 | (void)count; |
2921 | (void)size; |
2921 | (void)size; |
2922 | } |
2922 | } |
2923 | 2923 | ||
2924 | void kgem_cleanup_cache(struct kgem *kgem) |
2924 | void kgem_cleanup_cache(struct kgem *kgem) |
2925 | { |
2925 | { |
2926 | unsigned int i; |
2926 | unsigned int i; |
2927 | int n; |
2927 | int n; |
2928 | 2928 | ||
2929 | /* sync to the most recent request */ |
2929 | /* sync to the most recent request */ |
2930 | for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) { |
2930 | for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) { |
2931 | if (!list_is_empty(&kgem->requests[n])) { |
2931 | if (!list_is_empty(&kgem->requests[n])) { |
2932 | struct kgem_request *rq; |
2932 | struct kgem_request *rq; |
2933 | struct drm_i915_gem_set_domain set_domain; |
2933 | struct drm_i915_gem_set_domain set_domain; |
2934 | 2934 | ||
2935 | rq = list_first_entry(&kgem->requests[n], |
2935 | rq = list_first_entry(&kgem->requests[n], |
2936 | struct kgem_request, |
2936 | struct kgem_request, |
2937 | list); |
2937 | list); |
2938 | 2938 | ||
2939 | DBG(("%s: sync on cleanup\n", __FUNCTION__)); |
2939 | DBG(("%s: sync on cleanup\n", __FUNCTION__)); |
2940 | 2940 | ||
2941 | VG_CLEAR(set_domain); |
2941 | VG_CLEAR(set_domain); |
2942 | set_domain.handle = rq->bo->handle; |
2942 | set_domain.handle = rq->bo->handle; |
2943 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
2943 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
2944 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
2944 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
2945 | (void)drmIoctl(kgem->fd, |
2945 | (void)drmIoctl(kgem->fd, |
2946 | DRM_IOCTL_I915_GEM_SET_DOMAIN, |
2946 | DRM_IOCTL_I915_GEM_SET_DOMAIN, |
2947 | &set_domain); |
2947 | &set_domain); |
2948 | } |
2948 | } |
2949 | } |
2949 | } |
2950 | 2950 | ||
2951 | kgem_retire(kgem); |
2951 | kgem_retire(kgem); |
2952 | kgem_cleanup(kgem); |
2952 | kgem_cleanup(kgem); |
2953 | 2953 | ||
2954 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) { |
2954 | for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) { |
2955 | while (!list_is_empty(&kgem->inactive[i])) |
2955 | while (!list_is_empty(&kgem->inactive[i])) |
2956 | kgem_bo_free(kgem, |
2956 | kgem_bo_free(kgem, |
2957 | list_last_entry(&kgem->inactive[i], |
2957 | list_last_entry(&kgem->inactive[i], |
2958 | struct kgem_bo, list)); |
2958 | struct kgem_bo, list)); |
2959 | } |
2959 | } |
2960 | 2960 | ||
2961 | kgem_clean_large_cache(kgem); |
2961 | kgem_clean_large_cache(kgem); |
2962 | 2962 | ||
2963 | while (!list_is_empty(&kgem->snoop)) |
2963 | while (!list_is_empty(&kgem->snoop)) |
2964 | kgem_bo_free(kgem, |
2964 | kgem_bo_free(kgem, |
2965 | list_last_entry(&kgem->snoop, |
2965 | list_last_entry(&kgem->snoop, |
2966 | struct kgem_bo, list)); |
2966 | struct kgem_bo, list)); |
2967 | 2967 | ||
2968 | while (__kgem_freed_bo) { |
2968 | while (__kgem_freed_bo) { |
2969 | struct kgem_bo *bo = __kgem_freed_bo; |
2969 | struct kgem_bo *bo = __kgem_freed_bo; |
2970 | __kgem_freed_bo = *(struct kgem_bo **)bo; |
2970 | __kgem_freed_bo = *(struct kgem_bo **)bo; |
2971 | free(bo); |
2971 | free(bo); |
2972 | } |
2972 | } |
2973 | 2973 | ||
2974 | kgem->need_purge = false; |
2974 | kgem->need_purge = false; |
2975 | kgem->need_expire = false; |
2975 | kgem->need_expire = false; |
2976 | } |
2976 | } |
2977 | 2977 | ||
2978 | static struct kgem_bo * |
2978 | static struct kgem_bo * |
2979 | search_linear_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags) |
2979 | search_linear_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags) |
2980 | { |
2980 | { |
2981 | struct kgem_bo *bo, *first = NULL; |
2981 | struct kgem_bo *bo, *first = NULL; |
2982 | bool use_active = (flags & CREATE_INACTIVE) == 0; |
2982 | bool use_active = (flags & CREATE_INACTIVE) == 0; |
2983 | struct list *cache; |
2983 | struct list *cache; |
2984 | 2984 | ||
2985 | DBG(("%s: num_pages=%d, flags=%x, use_active? %d, use_large=%d [max=%d]\n", |
2985 | DBG(("%s: num_pages=%d, flags=%x, use_active? %d, use_large=%d [max=%d]\n", |
2986 | __FUNCTION__, num_pages, flags, use_active, |
2986 | __FUNCTION__, num_pages, flags, use_active, |
2987 | num_pages >= MAX_CACHE_SIZE / PAGE_SIZE, |
2987 | num_pages >= MAX_CACHE_SIZE / PAGE_SIZE, |
2988 | MAX_CACHE_SIZE / PAGE_SIZE)); |
2988 | MAX_CACHE_SIZE / PAGE_SIZE)); |
2989 | 2989 | ||
2990 | assert(num_pages); |
2990 | assert(num_pages); |
2991 | 2991 | ||
2992 | if (num_pages >= MAX_CACHE_SIZE / PAGE_SIZE) { |
2992 | if (num_pages >= MAX_CACHE_SIZE / PAGE_SIZE) { |
2993 | DBG(("%s: searching large buffers\n", __FUNCTION__)); |
2993 | DBG(("%s: searching large buffers\n", __FUNCTION__)); |
2994 | retry_large: |
2994 | retry_large: |
2995 | cache = use_active ? &kgem->large : &kgem->large_inactive; |
2995 | cache = use_active ? &kgem->large : &kgem->large_inactive; |
2996 | list_for_each_entry_safe(bo, first, cache, list) { |
2996 | list_for_each_entry_safe(bo, first, cache, list) { |
2997 | assert(bo->refcnt == 0); |
2997 | assert(bo->refcnt == 0); |
2998 | assert(bo->reusable); |
2998 | assert(bo->reusable); |
2999 | assert(!bo->scanout); |
2999 | assert(!bo->scanout); |
3000 | 3000 | ||
3001 | if (num_pages > num_pages(bo)) |
3001 | if (num_pages > num_pages(bo)) |
3002 | goto discard; |
3002 | goto discard; |
3003 | 3003 | ||
3004 | if (bo->tiling != I915_TILING_NONE) { |
3004 | if (bo->tiling != I915_TILING_NONE) { |
3005 | if (use_active) |
3005 | if (use_active) |
3006 | goto discard; |
3006 | goto discard; |
3007 | 3007 | ||
3008 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3008 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3009 | I915_TILING_NONE, 0)) |
3009 | I915_TILING_NONE, 0)) |
3010 | goto discard; |
3010 | goto discard; |
3011 | 3011 | ||
3012 | bo->tiling = I915_TILING_NONE; |
3012 | bo->tiling = I915_TILING_NONE; |
3013 | bo->pitch = 0; |
3013 | bo->pitch = 0; |
3014 | } |
3014 | } |
3015 | 3015 | ||
3016 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) |
3016 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) |
3017 | goto discard; |
3017 | goto discard; |
3018 | 3018 | ||
3019 | list_del(&bo->list); |
3019 | list_del(&bo->list); |
3020 | if (bo->rq == (void *)kgem) |
3020 | if (bo->rq == (void *)kgem) |
3021 | list_del(&bo->request); |
3021 | list_del(&bo->request); |
3022 | 3022 | ||
3023 | bo->delta = 0; |
3023 | bo->delta = 0; |
3024 | assert_tiling(kgem, bo); |
3024 | assert_tiling(kgem, bo); |
3025 | return bo; |
3025 | return bo; |
3026 | 3026 | ||
3027 | discard: |
3027 | discard: |
3028 | if (!use_active) |
3028 | if (!use_active) |
3029 | kgem_bo_free(kgem, bo); |
3029 | kgem_bo_free(kgem, bo); |
3030 | } |
3030 | } |
3031 | 3031 | ||
3032 | if (use_active) { |
3032 | if (use_active) { |
3033 | use_active = false; |
3033 | use_active = false; |
3034 | goto retry_large; |
3034 | goto retry_large; |
3035 | } |
3035 | } |
3036 | 3036 | ||
3037 | if (__kgem_throttle_retire(kgem, flags)) |
3037 | if (__kgem_throttle_retire(kgem, flags)) |
3038 | goto retry_large; |
3038 | goto retry_large; |
3039 | 3039 | ||
3040 | return NULL; |
3040 | return NULL; |
3041 | } |
3041 | } |
3042 | 3042 | ||
3043 | if (!use_active && list_is_empty(inactive(kgem, num_pages))) { |
3043 | if (!use_active && list_is_empty(inactive(kgem, num_pages))) { |
3044 | DBG(("%s: inactive and cache bucket empty\n", |
3044 | DBG(("%s: inactive and cache bucket empty\n", |
3045 | __FUNCTION__)); |
3045 | __FUNCTION__)); |
3046 | 3046 | ||
3047 | if (flags & CREATE_NO_RETIRE) { |
3047 | if (flags & CREATE_NO_RETIRE) { |
3048 | DBG(("%s: can not retire\n", __FUNCTION__)); |
3048 | DBG(("%s: can not retire\n", __FUNCTION__)); |
3049 | return NULL; |
3049 | return NULL; |
3050 | } |
3050 | } |
3051 | 3051 | ||
3052 | if (list_is_empty(active(kgem, num_pages, I915_TILING_NONE))) { |
3052 | if (list_is_empty(active(kgem, num_pages, I915_TILING_NONE))) { |
3053 | DBG(("%s: active cache bucket empty\n", __FUNCTION__)); |
3053 | DBG(("%s: active cache bucket empty\n", __FUNCTION__)); |
3054 | return NULL; |
3054 | return NULL; |
3055 | } |
3055 | } |
3056 | 3056 | ||
3057 | if (!__kgem_throttle_retire(kgem, flags)) { |
3057 | if (!__kgem_throttle_retire(kgem, flags)) { |
3058 | DBG(("%s: nothing retired\n", __FUNCTION__)); |
3058 | DBG(("%s: nothing retired\n", __FUNCTION__)); |
3059 | return NULL; |
3059 | return NULL; |
3060 | } |
3060 | } |
3061 | 3061 | ||
3062 | if (list_is_empty(inactive(kgem, num_pages))) { |
3062 | if (list_is_empty(inactive(kgem, num_pages))) { |
3063 | DBG(("%s: active cache bucket still empty after retire\n", |
3063 | DBG(("%s: active cache bucket still empty after retire\n", |
3064 | __FUNCTION__)); |
3064 | __FUNCTION__)); |
3065 | return NULL; |
3065 | return NULL; |
3066 | } |
3066 | } |
3067 | } |
3067 | } |
3068 | 3068 | ||
3069 | if (!use_active && flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) { |
3069 | if (!use_active && flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) { |
3070 | int for_cpu = !!(flags & CREATE_CPU_MAP); |
3070 | int for_cpu = !!(flags & CREATE_CPU_MAP); |
3071 | DBG(("%s: searching for inactive %s map\n", |
3071 | DBG(("%s: searching for inactive %s map\n", |
3072 | __FUNCTION__, for_cpu ? "cpu" : "gtt")); |
3072 | __FUNCTION__, for_cpu ? "cpu" : "gtt")); |
3073 | cache = &kgem->vma[for_cpu].inactive[cache_bucket(num_pages)]; |
3073 | cache = &kgem->vma[for_cpu].inactive[cache_bucket(num_pages)]; |
3074 | list_for_each_entry(bo, cache, vma) { |
3074 | list_for_each_entry(bo, cache, vma) { |
3075 | assert(IS_CPU_MAP(bo->map) == for_cpu); |
3075 | assert(IS_CPU_MAP(bo->map) == for_cpu); |
3076 | assert(bucket(bo) == cache_bucket(num_pages)); |
3076 | assert(bucket(bo) == cache_bucket(num_pages)); |
3077 | assert(bo->proxy == NULL); |
3077 | assert(bo->proxy == NULL); |
3078 | assert(bo->rq == NULL); |
3078 | assert(bo->rq == NULL); |
3079 | assert(bo->exec == NULL); |
3079 | assert(bo->exec == NULL); |
3080 | assert(!bo->scanout); |
3080 | assert(!bo->scanout); |
3081 | 3081 | ||
3082 | if (num_pages > num_pages(bo)) { |
3082 | if (num_pages > num_pages(bo)) { |
3083 | DBG(("inactive too small: %d < %d\n", |
3083 | DBG(("inactive too small: %d < %d\n", |
3084 | num_pages(bo), num_pages)); |
3084 | num_pages(bo), num_pages)); |
3085 | continue; |
3085 | continue; |
3086 | } |
3086 | } |
3087 | 3087 | ||
3088 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3088 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3089 | kgem_bo_free(kgem, bo); |
3089 | kgem_bo_free(kgem, bo); |
3090 | break; |
3090 | break; |
3091 | } |
3091 | } |
3092 | 3092 | ||
3093 | if (I915_TILING_NONE != bo->tiling && |
3093 | if (I915_TILING_NONE != bo->tiling && |
3094 | !gem_set_tiling(kgem->fd, bo->handle, |
3094 | !gem_set_tiling(kgem->fd, bo->handle, |
3095 | I915_TILING_NONE, 0)) |
3095 | I915_TILING_NONE, 0)) |
3096 | continue; |
3096 | continue; |
3097 | 3097 | ||
3098 | kgem_bo_remove_from_inactive(kgem, bo); |
3098 | kgem_bo_remove_from_inactive(kgem, bo); |
3099 | 3099 | ||
3100 | bo->tiling = I915_TILING_NONE; |
3100 | bo->tiling = I915_TILING_NONE; |
3101 | bo->pitch = 0; |
3101 | bo->pitch = 0; |
3102 | bo->delta = 0; |
3102 | bo->delta = 0; |
3103 | DBG((" %s: found handle=%d (num_pages=%d) in linear vma cache\n", |
3103 | DBG((" %s: found handle=%d (num_pages=%d) in linear vma cache\n", |
3104 | __FUNCTION__, bo->handle, num_pages(bo))); |
3104 | __FUNCTION__, bo->handle, num_pages(bo))); |
3105 | assert(use_active || bo->domain != DOMAIN_GPU); |
3105 | assert(use_active || bo->domain != DOMAIN_GPU); |
3106 | assert(!bo->needs_flush); |
3106 | assert(!bo->needs_flush); |
3107 | assert_tiling(kgem, bo); |
3107 | assert_tiling(kgem, bo); |
3108 | ASSERT_MAYBE_IDLE(kgem, bo->handle, !use_active); |
3108 | ASSERT_MAYBE_IDLE(kgem, bo->handle, !use_active); |
3109 | return bo; |
3109 | return bo; |
3110 | } |
3110 | } |
3111 | 3111 | ||
3112 | if (flags & CREATE_EXACT) |
3112 | if (flags & CREATE_EXACT) |
3113 | return NULL; |
3113 | return NULL; |
3114 | 3114 | ||
3115 | if (flags & CREATE_CPU_MAP && !kgem->has_llc) |
3115 | if (flags & CREATE_CPU_MAP && !kgem->has_llc) |
3116 | return NULL; |
3116 | return NULL; |
3117 | } |
3117 | } |
3118 | 3118 | ||
3119 | cache = use_active ? active(kgem, num_pages, I915_TILING_NONE) : inactive(kgem, num_pages); |
3119 | cache = use_active ? active(kgem, num_pages, I915_TILING_NONE) : inactive(kgem, num_pages); |
3120 | list_for_each_entry(bo, cache, list) { |
3120 | list_for_each_entry(bo, cache, list) { |
3121 | assert(bo->refcnt == 0); |
3121 | assert(bo->refcnt == 0); |
3122 | assert(bo->reusable); |
3122 | assert(bo->reusable); |
3123 | assert(!!bo->rq == !!use_active); |
3123 | assert(!!bo->rq == !!use_active); |
3124 | assert(bo->proxy == NULL); |
3124 | assert(bo->proxy == NULL); |
3125 | assert(!bo->scanout); |
3125 | assert(!bo->scanout); |
3126 | 3126 | ||
3127 | if (num_pages > num_pages(bo)) |
3127 | if (num_pages > num_pages(bo)) |
3128 | continue; |
3128 | continue; |
3129 | 3129 | ||
3130 | if (use_active && |
3130 | if (use_active && |
3131 | kgem->gen <= 040 && |
3131 | kgem->gen <= 040 && |
3132 | bo->tiling != I915_TILING_NONE) |
3132 | bo->tiling != I915_TILING_NONE) |
3133 | continue; |
3133 | continue; |
3134 | 3134 | ||
3135 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3135 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3136 | kgem_bo_free(kgem, bo); |
3136 | kgem_bo_free(kgem, bo); |
3137 | break; |
3137 | break; |
3138 | } |
3138 | } |
3139 | 3139 | ||
3140 | if (I915_TILING_NONE != bo->tiling) { |
3140 | if (I915_TILING_NONE != bo->tiling) { |
3141 | if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) |
3141 | if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) |
3142 | continue; |
3142 | continue; |
3143 | 3143 | ||
3144 | if (first) |
3144 | if (first) |
3145 | continue; |
3145 | continue; |
3146 | 3146 | ||
3147 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3147 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3148 | I915_TILING_NONE, 0)) |
3148 | I915_TILING_NONE, 0)) |
3149 | continue; |
3149 | continue; |
3150 | 3150 | ||
3151 | bo->tiling = I915_TILING_NONE; |
3151 | bo->tiling = I915_TILING_NONE; |
3152 | bo->pitch = 0; |
3152 | bo->pitch = 0; |
3153 | } |
3153 | } |
3154 | 3154 | ||
3155 | if (bo->map) { |
3155 | if (bo->map) { |
3156 | if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) { |
3156 | if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) { |
3157 | int for_cpu = !!(flags & CREATE_CPU_MAP); |
3157 | int for_cpu = !!(flags & CREATE_CPU_MAP); |
3158 | if (IS_CPU_MAP(bo->map) != for_cpu) { |
3158 | if (IS_CPU_MAP(bo->map) != for_cpu) { |
3159 | if (first != NULL) |
3159 | if (first != NULL) |
3160 | break; |
3160 | break; |
3161 | 3161 | ||
3162 | first = bo; |
3162 | first = bo; |
3163 | continue; |
3163 | continue; |
3164 | } |
3164 | } |
3165 | } else { |
3165 | } else { |
3166 | if (first != NULL) |
3166 | if (first != NULL) |
3167 | break; |
3167 | break; |
3168 | 3168 | ||
3169 | first = bo; |
3169 | first = bo; |
3170 | continue; |
3170 | continue; |
3171 | } |
3171 | } |
3172 | } else { |
3172 | } else { |
3173 | if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) { |
3173 | if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) { |
3174 | if (first != NULL) |
3174 | if (first != NULL) |
3175 | break; |
3175 | break; |
3176 | 3176 | ||
3177 | first = bo; |
3177 | first = bo; |
3178 | continue; |
3178 | continue; |
3179 | } |
3179 | } |
3180 | } |
3180 | } |
3181 | 3181 | ||
3182 | if (use_active) |
3182 | if (use_active) |
3183 | kgem_bo_remove_from_active(kgem, bo); |
3183 | kgem_bo_remove_from_active(kgem, bo); |
3184 | else |
3184 | else |
3185 | kgem_bo_remove_from_inactive(kgem, bo); |
3185 | kgem_bo_remove_from_inactive(kgem, bo); |
3186 | 3186 | ||
3187 | assert(bo->tiling == I915_TILING_NONE); |
3187 | assert(bo->tiling == I915_TILING_NONE); |
3188 | bo->pitch = 0; |
3188 | bo->pitch = 0; |
3189 | bo->delta = 0; |
3189 | bo->delta = 0; |
3190 | DBG((" %s: found handle=%d (num_pages=%d) in linear %s cache\n", |
3190 | DBG((" %s: found handle=%d (num_pages=%d) in linear %s cache\n", |
3191 | __FUNCTION__, bo->handle, num_pages(bo), |
3191 | __FUNCTION__, bo->handle, num_pages(bo), |
3192 | use_active ? "active" : "inactive")); |
3192 | use_active ? "active" : "inactive")); |
3193 | assert(list_is_empty(&bo->list)); |
3193 | assert(list_is_empty(&bo->list)); |
3194 | assert(use_active || bo->domain != DOMAIN_GPU); |
3194 | assert(use_active || bo->domain != DOMAIN_GPU); |
3195 | assert(!bo->needs_flush || use_active); |
3195 | assert(!bo->needs_flush || use_active); |
3196 | assert_tiling(kgem, bo); |
3196 | assert_tiling(kgem, bo); |
3197 | ASSERT_MAYBE_IDLE(kgem, bo->handle, !use_active); |
3197 | ASSERT_MAYBE_IDLE(kgem, bo->handle, !use_active); |
3198 | return bo; |
3198 | return bo; |
3199 | } |
3199 | } |
3200 | 3200 | ||
3201 | if (first) { |
3201 | if (first) { |
3202 | assert(first->tiling == I915_TILING_NONE); |
3202 | assert(first->tiling == I915_TILING_NONE); |
3203 | 3203 | ||
3204 | if (use_active) |
3204 | if (use_active) |
3205 | kgem_bo_remove_from_active(kgem, first); |
3205 | kgem_bo_remove_from_active(kgem, first); |
3206 | else |
3206 | else |
3207 | kgem_bo_remove_from_inactive(kgem, first); |
3207 | kgem_bo_remove_from_inactive(kgem, first); |
3208 | 3208 | ||
3209 | first->pitch = 0; |
3209 | first->pitch = 0; |
3210 | first->delta = 0; |
3210 | first->delta = 0; |
3211 | DBG((" %s: found handle=%d (near-miss) (num_pages=%d) in linear %s cache\n", |
3211 | DBG((" %s: found handle=%d (near-miss) (num_pages=%d) in linear %s cache\n", |
3212 | __FUNCTION__, first->handle, num_pages(first), |
3212 | __FUNCTION__, first->handle, num_pages(first), |
3213 | use_active ? "active" : "inactive")); |
3213 | use_active ? "active" : "inactive")); |
3214 | assert(list_is_empty(&first->list)); |
3214 | assert(list_is_empty(&first->list)); |
3215 | assert(use_active || first->domain != DOMAIN_GPU); |
3215 | assert(use_active || first->domain != DOMAIN_GPU); |
3216 | assert(!first->needs_flush || use_active); |
3216 | assert(!first->needs_flush || use_active); |
3217 | ASSERT_MAYBE_IDLE(kgem, first->handle, !use_active); |
3217 | ASSERT_MAYBE_IDLE(kgem, first->handle, !use_active); |
3218 | return first; |
3218 | return first; |
3219 | } |
3219 | } |
3220 | 3220 | ||
3221 | return NULL; |
3221 | return NULL; |
3222 | } |
3222 | } |
3223 | 3223 | ||
3224 | 3224 | ||
3225 | struct kgem_bo *kgem_create_linear(struct kgem *kgem, int size, unsigned flags) |
3225 | struct kgem_bo *kgem_create_linear(struct kgem *kgem, int size, unsigned flags) |
3226 | { |
3226 | { |
3227 | struct kgem_bo *bo; |
3227 | struct kgem_bo *bo; |
3228 | uint32_t handle; |
3228 | uint32_t handle; |
3229 | 3229 | ||
3230 | DBG(("%s(%d)\n", __FUNCTION__, size)); |
3230 | DBG(("%s(%d)\n", __FUNCTION__, size)); |
3231 | assert(size); |
3231 | assert(size); |
3232 | 3232 | ||
3233 | if (flags & CREATE_GTT_MAP && kgem->has_llc) { |
3233 | if (flags & CREATE_GTT_MAP && kgem->has_llc) { |
3234 | flags &= ~CREATE_GTT_MAP; |
3234 | flags &= ~CREATE_GTT_MAP; |
3235 | flags |= CREATE_CPU_MAP; |
3235 | flags |= CREATE_CPU_MAP; |
3236 | } |
3236 | } |
3237 | 3237 | ||
3238 | size = NUM_PAGES(size); |
3238 | size = NUM_PAGES(size); |
3239 | bo = search_linear_cache(kgem, size, CREATE_INACTIVE | flags); |
3239 | bo = search_linear_cache(kgem, size, CREATE_INACTIVE | flags); |
3240 | if (bo) { |
3240 | if (bo) { |
3241 | assert(bo->domain != DOMAIN_GPU); |
3241 | assert(bo->domain != DOMAIN_GPU); |
3242 | ASSERT_IDLE(kgem, bo->handle); |
3242 | ASSERT_IDLE(kgem, bo->handle); |
3243 | bo->refcnt = 1; |
3243 | bo->refcnt = 1; |
3244 | return bo; |
3244 | return bo; |
3245 | } |
3245 | } |
3246 | 3246 | ||
3247 | if (flags & CREATE_CACHED) |
3247 | if (flags & CREATE_CACHED) |
3248 | return NULL; |
3248 | return NULL; |
3249 | 3249 | ||
3250 | handle = gem_create(kgem->fd, size); |
3250 | handle = gem_create(kgem->fd, size); |
3251 | if (handle == 0) |
3251 | if (handle == 0) |
3252 | return NULL; |
3252 | return NULL; |
3253 | 3253 | ||
3254 | DBG(("%s: new handle=%d, num_pages=%d\n", __FUNCTION__, handle, size)); |
3254 | DBG(("%s: new handle=%d, num_pages=%d\n", __FUNCTION__, handle, size)); |
3255 | bo = __kgem_bo_alloc(handle, size); |
3255 | bo = __kgem_bo_alloc(handle, size); |
3256 | if (bo == NULL) { |
3256 | if (bo == NULL) { |
3257 | gem_close(kgem->fd, handle); |
3257 | gem_close(kgem->fd, handle); |
3258 | return NULL; |
3258 | return NULL; |
3259 | } |
3259 | } |
3260 | 3260 | ||
3261 | debug_alloc__bo(kgem, bo); |
3261 | debug_alloc__bo(kgem, bo); |
3262 | return bo; |
3262 | return bo; |
3263 | } |
3263 | } |
3264 | 3264 | ||
3265 | inline int kgem_bo_fenced_size(struct kgem *kgem, struct kgem_bo *bo) |
3265 | inline int kgem_bo_fenced_size(struct kgem *kgem, struct kgem_bo *bo) |
3266 | { |
3266 | { |
3267 | unsigned int size; |
3267 | unsigned int size; |
3268 | 3268 | ||
3269 | assert(bo->tiling); |
3269 | assert(bo->tiling); |
3270 | assert_tiling(kgem, bo); |
3270 | assert_tiling(kgem, bo); |
3271 | assert(kgem->gen < 040); |
3271 | assert(kgem->gen < 040); |
3272 | 3272 | ||
3273 | if (kgem->gen < 030) |
3273 | if (kgem->gen < 030) |
3274 | size = 512 * 1024; |
3274 | size = 512 * 1024; |
3275 | else |
3275 | else |
3276 | size = 1024 * 1024; |
3276 | size = 1024 * 1024; |
3277 | while (size < bytes(bo)) |
3277 | while (size < bytes(bo)) |
3278 | size *= 2; |
3278 | size *= 2; |
3279 | 3279 | ||
3280 | return size; |
3280 | return size; |
3281 | } |
3281 | } |
3282 | 3282 | ||
3283 | struct kgem_bo *kgem_create_2d(struct kgem *kgem, |
3283 | struct kgem_bo *kgem_create_2d(struct kgem *kgem, |
3284 | int width, |
3284 | int width, |
3285 | int height, |
3285 | int height, |
3286 | int bpp, |
3286 | int bpp, |
3287 | int tiling, |
3287 | int tiling, |
3288 | uint32_t flags) |
3288 | uint32_t flags) |
3289 | { |
3289 | { |
3290 | struct list *cache; |
3290 | struct list *cache; |
3291 | struct kgem_bo *bo; |
3291 | struct kgem_bo *bo; |
3292 | uint32_t pitch, tiled_height, size; |
3292 | uint32_t pitch, tiled_height, size; |
3293 | uint32_t handle; |
3293 | uint32_t handle; |
3294 | int i, bucket, retry; |
3294 | int i, bucket, retry; |
3295 | bool exact = flags & (CREATE_EXACT | CREATE_SCANOUT); |
3295 | bool exact = flags & (CREATE_EXACT | CREATE_SCANOUT); |
3296 | 3296 | ||
3297 | if (tiling < 0) |
3297 | if (tiling < 0) |
3298 | exact = true, tiling = -tiling; |
3298 | exact = true, tiling = -tiling; |
3299 | 3299 | ||
3300 | 3300 | ||
3301 | DBG(("%s(%dx%d, bpp=%d, tiling=%d, exact=%d, inactive=%d, cpu-mapping=%d, gtt-mapping=%d, scanout?=%d, prime?=%d, temp?=%d)\n", __FUNCTION__, |
3301 | DBG(("%s(%dx%d, bpp=%d, tiling=%d, exact=%d, inactive=%d, cpu-mapping=%d, gtt-mapping=%d, scanout?=%d, prime?=%d, temp?=%d)\n", __FUNCTION__, |
3302 | width, height, bpp, tiling, exact, |
3302 | width, height, bpp, tiling, exact, |
3303 | !!(flags & CREATE_INACTIVE), |
3303 | !!(flags & CREATE_INACTIVE), |
3304 | !!(flags & CREATE_CPU_MAP), |
3304 | !!(flags & CREATE_CPU_MAP), |
3305 | !!(flags & CREATE_GTT_MAP), |
3305 | !!(flags & CREATE_GTT_MAP), |
3306 | !!(flags & CREATE_SCANOUT), |
3306 | !!(flags & CREATE_SCANOUT), |
3307 | !!(flags & CREATE_PRIME), |
3307 | !!(flags & CREATE_PRIME), |
3308 | !!(flags & CREATE_TEMPORARY))); |
3308 | !!(flags & CREATE_TEMPORARY))); |
3309 | 3309 | ||
3310 | size = kgem_surface_size(kgem, kgem->has_relaxed_fencing, flags, |
3310 | size = kgem_surface_size(kgem, kgem->has_relaxed_fencing, flags, |
3311 | width, height, bpp, tiling, &pitch); |
3311 | width, height, bpp, tiling, &pitch); |
3312 | assert(size && size <= kgem->max_object_size); |
3312 | assert(size && size <= kgem->max_object_size); |
3313 | size /= PAGE_SIZE; |
3313 | size /= PAGE_SIZE; |
3314 | bucket = cache_bucket(size); |
3314 | bucket = cache_bucket(size); |
3315 | 3315 | ||
3316 | if (flags & CREATE_SCANOUT) { |
3316 | if (flags & CREATE_SCANOUT) { |
3317 | struct kgem_bo *last = NULL; |
3317 | struct kgem_bo *last = NULL; |
3318 | 3318 | ||
3319 | list_for_each_entry_reverse(bo, &kgem->scanout, list) { |
3319 | list_for_each_entry_reverse(bo, &kgem->scanout, list) { |
3320 | assert(bo->scanout); |
3320 | assert(bo->scanout); |
3321 | assert(bo->delta); |
3321 | assert(bo->delta); |
3322 | assert(!bo->flush); |
3322 | assert(!bo->flush); |
3323 | assert_tiling(kgem, bo); |
3323 | assert_tiling(kgem, bo); |
3324 | 3324 | ||
3325 | if (size > num_pages(bo) || num_pages(bo) > 2*size) |
3325 | if (size > num_pages(bo) || num_pages(bo) > 2*size) |
3326 | continue; |
3326 | continue; |
3327 | 3327 | ||
3328 | if (bo->tiling != tiling || |
3328 | if (bo->tiling != tiling || |
3329 | (tiling != I915_TILING_NONE && bo->pitch != pitch)) { |
3329 | (tiling != I915_TILING_NONE && bo->pitch != pitch)) { |
3330 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3330 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3331 | tiling, pitch)) |
3331 | tiling, pitch)) |
3332 | continue; |
3332 | continue; |
3333 | 3333 | ||
3334 | bo->tiling = tiling; |
3334 | bo->tiling = tiling; |
3335 | bo->pitch = pitch; |
3335 | bo->pitch = pitch; |
3336 | } |
3336 | } |
3337 | 3337 | ||
3338 | if (flags & CREATE_INACTIVE && bo->rq) { |
3338 | if (flags & CREATE_INACTIVE && bo->rq) { |
3339 | last = bo; |
3339 | last = bo; |
3340 | continue; |
3340 | continue; |
3341 | } |
3341 | } |
3342 | 3342 | ||
3343 | list_del(&bo->list); |
3343 | list_del(&bo->list); |
3344 | 3344 | ||
3345 | bo->unique_id = kgem_get_unique_id(kgem); |
3345 | bo->unique_id = kgem_get_unique_id(kgem); |
3346 | DBG((" 1:from scanout: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3346 | DBG((" 1:from scanout: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3347 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3347 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3348 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3348 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3349 | assert_tiling(kgem, bo); |
3349 | assert_tiling(kgem, bo); |
3350 | bo->refcnt = 1; |
3350 | bo->refcnt = 1; |
3351 | return bo; |
3351 | return bo; |
3352 | } |
3352 | } |
3353 | 3353 | ||
3354 | if (last) { |
3354 | if (last) { |
3355 | list_del(&last->list); |
3355 | list_del(&last->list); |
3356 | 3356 | ||
3357 | last->unique_id = kgem_get_unique_id(kgem); |
3357 | last->unique_id = kgem_get_unique_id(kgem); |
3358 | DBG((" 1:from scanout: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3358 | DBG((" 1:from scanout: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3359 | last->pitch, last->tiling, last->handle, last->unique_id)); |
3359 | last->pitch, last->tiling, last->handle, last->unique_id)); |
3360 | assert(last->pitch*kgem_aligned_height(kgem, height, last->tiling) <= kgem_bo_size(last)); |
3360 | assert(last->pitch*kgem_aligned_height(kgem, height, last->tiling) <= kgem_bo_size(last)); |
3361 | assert_tiling(kgem, last); |
3361 | assert_tiling(kgem, last); |
3362 | last->refcnt = 1; |
3362 | last->refcnt = 1; |
3363 | return last; |
3363 | return last; |
3364 | } |
3364 | } |
3365 | 3365 | ||
3366 | bo = NULL; //__kgem_bo_create_as_display(kgem, size, tiling, pitch); |
3366 | bo = NULL; //__kgem_bo_create_as_display(kgem, size, tiling, pitch); |
3367 | if (bo) |
3367 | if (bo) |
3368 | return bo; |
3368 | return bo; |
3369 | } |
3369 | } |
3370 | 3370 | ||
3371 | if (bucket >= NUM_CACHE_BUCKETS) { |
3371 | if (bucket >= NUM_CACHE_BUCKETS) { |
3372 | DBG(("%s: large bo num pages=%d, bucket=%d\n", |
3372 | DBG(("%s: large bo num pages=%d, bucket=%d\n", |
3373 | __FUNCTION__, size, bucket)); |
3373 | __FUNCTION__, size, bucket)); |
3374 | 3374 | ||
3375 | if (flags & CREATE_INACTIVE) |
3375 | if (flags & CREATE_INACTIVE) |
3376 | goto large_inactive; |
3376 | goto large_inactive; |
3377 | 3377 | ||
3378 | tiled_height = kgem_aligned_height(kgem, height, tiling); |
3378 | tiled_height = kgem_aligned_height(kgem, height, tiling); |
3379 | 3379 | ||
3380 | list_for_each_entry(bo, &kgem->large, list) { |
3380 | list_for_each_entry(bo, &kgem->large, list) { |
3381 | assert(!bo->purged); |
3381 | assert(!bo->purged); |
3382 | assert(!bo->scanout); |
3382 | assert(!bo->scanout); |
3383 | assert(bo->refcnt == 0); |
3383 | assert(bo->refcnt == 0); |
3384 | assert(bo->reusable); |
3384 | assert(bo->reusable); |
3385 | assert_tiling(kgem, bo); |
3385 | assert_tiling(kgem, bo); |
3386 | 3386 | ||
3387 | if (kgem->gen < 040) { |
3387 | if (kgem->gen < 040) { |
3388 | if (bo->pitch < pitch) { |
3388 | if (bo->pitch < pitch) { |
3389 | DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n", |
3389 | DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n", |
3390 | bo->tiling, tiling, |
3390 | bo->tiling, tiling, |
3391 | bo->pitch, pitch)); |
3391 | bo->pitch, pitch)); |
3392 | continue; |
3392 | continue; |
3393 | } |
3393 | } |
3394 | 3394 | ||
3395 | if (bo->pitch * tiled_height > bytes(bo)) |
3395 | if (bo->pitch * tiled_height > bytes(bo)) |
3396 | continue; |
3396 | continue; |
3397 | } else { |
3397 | } else { |
3398 | if (num_pages(bo) < size) |
3398 | if (num_pages(bo) < size) |
3399 | continue; |
3399 | continue; |
3400 | 3400 | ||
3401 | if (bo->pitch != pitch || bo->tiling != tiling) { |
3401 | if (bo->pitch != pitch || bo->tiling != tiling) { |
3402 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3402 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3403 | tiling, pitch)) |
3403 | tiling, pitch)) |
3404 | continue; |
3404 | continue; |
3405 | 3405 | ||
3406 | bo->pitch = pitch; |
3406 | bo->pitch = pitch; |
3407 | bo->tiling = tiling; |
3407 | bo->tiling = tiling; |
3408 | } |
3408 | } |
3409 | } |
3409 | } |
3410 | 3410 | ||
3411 | kgem_bo_remove_from_active(kgem, bo); |
3411 | kgem_bo_remove_from_active(kgem, bo); |
3412 | 3412 | ||
3413 | bo->unique_id = kgem_get_unique_id(kgem); |
3413 | bo->unique_id = kgem_get_unique_id(kgem); |
3414 | bo->delta = 0; |
3414 | bo->delta = 0; |
3415 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3415 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3416 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3416 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3417 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3417 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3418 | assert_tiling(kgem, bo); |
3418 | assert_tiling(kgem, bo); |
3419 | bo->refcnt = 1; |
3419 | bo->refcnt = 1; |
3420 | bo->flush = true; |
3420 | bo->flush = true; |
3421 | return bo; |
3421 | return bo; |
3422 | } |
3422 | } |
3423 | 3423 | ||
3424 | large_inactive: |
3424 | large_inactive: |
3425 | __kgem_throttle_retire(kgem, flags); |
3425 | __kgem_throttle_retire(kgem, flags); |
3426 | list_for_each_entry(bo, &kgem->large_inactive, list) { |
3426 | list_for_each_entry(bo, &kgem->large_inactive, list) { |
3427 | assert(bo->refcnt == 0); |
3427 | assert(bo->refcnt == 0); |
3428 | assert(bo->reusable); |
3428 | assert(bo->reusable); |
3429 | assert(!bo->scanout); |
3429 | assert(!bo->scanout); |
3430 | assert_tiling(kgem, bo); |
3430 | assert_tiling(kgem, bo); |
3431 | 3431 | ||
3432 | if (size > num_pages(bo)) |
3432 | if (size > num_pages(bo)) |
3433 | continue; |
3433 | continue; |
3434 | 3434 | ||
3435 | if (bo->tiling != tiling || |
3435 | if (bo->tiling != tiling || |
3436 | (tiling != I915_TILING_NONE && bo->pitch != pitch)) { |
3436 | (tiling != I915_TILING_NONE && bo->pitch != pitch)) { |
3437 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3437 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3438 | tiling, pitch)) |
3438 | tiling, pitch)) |
3439 | continue; |
3439 | continue; |
3440 | 3440 | ||
3441 | bo->tiling = tiling; |
3441 | bo->tiling = tiling; |
3442 | bo->pitch = pitch; |
3442 | bo->pitch = pitch; |
3443 | } |
3443 | } |
3444 | 3444 | ||
3445 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3445 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3446 | kgem_bo_free(kgem, bo); |
3446 | kgem_bo_free(kgem, bo); |
3447 | break; |
3447 | break; |
3448 | } |
3448 | } |
3449 | 3449 | ||
3450 | list_del(&bo->list); |
3450 | list_del(&bo->list); |
3451 | 3451 | ||
3452 | assert(bo->domain != DOMAIN_GPU); |
3452 | assert(bo->domain != DOMAIN_GPU); |
3453 | bo->unique_id = kgem_get_unique_id(kgem); |
3453 | bo->unique_id = kgem_get_unique_id(kgem); |
3454 | bo->pitch = pitch; |
3454 | bo->pitch = pitch; |
3455 | bo->delta = 0; |
3455 | bo->delta = 0; |
3456 | DBG((" 1:from large inactive: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3456 | DBG((" 1:from large inactive: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3457 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3457 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3458 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3458 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3459 | assert_tiling(kgem, bo); |
3459 | assert_tiling(kgem, bo); |
3460 | bo->refcnt = 1; |
3460 | bo->refcnt = 1; |
3461 | return bo; |
3461 | return bo; |
3462 | } |
3462 | } |
3463 | 3463 | ||
3464 | goto create; |
3464 | goto create; |
3465 | } |
3465 | } |
3466 | 3466 | ||
3467 | if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) { |
3467 | if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) { |
3468 | int for_cpu = !!(flags & CREATE_CPU_MAP); |
3468 | int for_cpu = !!(flags & CREATE_CPU_MAP); |
3469 | if (kgem->has_llc && tiling == I915_TILING_NONE) |
3469 | if (kgem->has_llc && tiling == I915_TILING_NONE) |
3470 | for_cpu = 1; |
3470 | for_cpu = 1; |
3471 | /* We presume that we will need to upload to this bo, |
3471 | /* We presume that we will need to upload to this bo, |
3472 | * and so would prefer to have an active VMA. |
3472 | * and so would prefer to have an active VMA. |
3473 | */ |
3473 | */ |
3474 | cache = &kgem->vma[for_cpu].inactive[bucket]; |
3474 | cache = &kgem->vma[for_cpu].inactive[bucket]; |
3475 | do { |
3475 | do { |
3476 | list_for_each_entry(bo, cache, vma) { |
3476 | list_for_each_entry(bo, cache, vma) { |
3477 | assert(bucket(bo) == bucket); |
3477 | assert(bucket(bo) == bucket); |
3478 | assert(bo->refcnt == 0); |
3478 | assert(bo->refcnt == 0); |
3479 | assert(!bo->scanout); |
3479 | assert(!bo->scanout); |
3480 | assert(bo->map); |
3480 | assert(bo->map); |
3481 | assert(IS_CPU_MAP(bo->map) == for_cpu); |
3481 | assert(IS_CPU_MAP(bo->map) == for_cpu); |
3482 | assert(bo->rq == NULL); |
3482 | assert(bo->rq == NULL); |
3483 | assert(list_is_empty(&bo->request)); |
3483 | assert(list_is_empty(&bo->request)); |
3484 | assert(bo->flush == false); |
3484 | assert(bo->flush == false); |
3485 | assert_tiling(kgem, bo); |
3485 | assert_tiling(kgem, bo); |
3486 | 3486 | ||
3487 | if (size > num_pages(bo)) { |
3487 | if (size > num_pages(bo)) { |
3488 | DBG(("inactive too small: %d < %d\n", |
3488 | DBG(("inactive too small: %d < %d\n", |
3489 | num_pages(bo), size)); |
3489 | num_pages(bo), size)); |
3490 | continue; |
3490 | continue; |
3491 | } |
3491 | } |
3492 | 3492 | ||
3493 | if (bo->tiling != tiling || |
3493 | if (bo->tiling != tiling || |
3494 | (tiling != I915_TILING_NONE && bo->pitch != pitch)) { |
3494 | (tiling != I915_TILING_NONE && bo->pitch != pitch)) { |
3495 | DBG(("inactive vma with wrong tiling: %d < %d\n", |
3495 | DBG(("inactive vma with wrong tiling: %d < %d\n", |
3496 | bo->tiling, tiling)); |
3496 | bo->tiling, tiling)); |
3497 | continue; |
3497 | continue; |
3498 | } |
3498 | } |
3499 | 3499 | ||
3500 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3500 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3501 | kgem_bo_free(kgem, bo); |
3501 | kgem_bo_free(kgem, bo); |
3502 | break; |
3502 | break; |
3503 | } |
3503 | } |
3504 | 3504 | ||
3505 | assert(bo->tiling == tiling); |
3505 | assert(bo->tiling == tiling); |
3506 | bo->pitch = pitch; |
3506 | bo->pitch = pitch; |
3507 | bo->delta = 0; |
3507 | bo->delta = 0; |
3508 | bo->unique_id = kgem_get_unique_id(kgem); |
3508 | bo->unique_id = kgem_get_unique_id(kgem); |
3509 | bo->domain = DOMAIN_NONE; |
3509 | bo->domain = DOMAIN_NONE; |
3510 | 3510 | ||
3511 | kgem_bo_remove_from_inactive(kgem, bo); |
3511 | kgem_bo_remove_from_inactive(kgem, bo); |
3512 | 3512 | ||
3513 | DBG((" from inactive vma: pitch=%d, tiling=%d: handle=%d, id=%d\n", |
3513 | DBG((" from inactive vma: pitch=%d, tiling=%d: handle=%d, id=%d\n", |
3514 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3514 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3515 | assert(bo->reusable); |
3515 | assert(bo->reusable); |
3516 | assert(bo->domain != DOMAIN_GPU); |
3516 | assert(bo->domain != DOMAIN_GPU); |
3517 | ASSERT_IDLE(kgem, bo->handle); |
3517 | ASSERT_IDLE(kgem, bo->handle); |
3518 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3518 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3519 | assert_tiling(kgem, bo); |
3519 | assert_tiling(kgem, bo); |
3520 | bo->refcnt = 1; |
3520 | bo->refcnt = 1; |
3521 | return bo; |
3521 | return bo; |
3522 | } |
3522 | } |
3523 | } while (!list_is_empty(cache) && |
3523 | } while (!list_is_empty(cache) && |
3524 | __kgem_throttle_retire(kgem, flags)); |
3524 | __kgem_throttle_retire(kgem, flags)); |
3525 | 3525 | ||
3526 | if (flags & CREATE_CPU_MAP && !kgem->has_llc) { |
3526 | if (flags & CREATE_CPU_MAP && !kgem->has_llc) { |
3527 | if (list_is_empty(&kgem->active[bucket][tiling]) && |
3527 | if (list_is_empty(&kgem->active[bucket][tiling]) && |
3528 | list_is_empty(&kgem->inactive[bucket])) |
3528 | list_is_empty(&kgem->inactive[bucket])) |
3529 | flags &= ~CREATE_CACHED; |
3529 | flags &= ~CREATE_CACHED; |
3530 | 3530 | ||
3531 | goto create; |
3531 | goto create; |
3532 | } |
3532 | } |
3533 | } |
3533 | } |
3534 | 3534 | ||
3535 | if (flags & CREATE_INACTIVE) |
3535 | if (flags & CREATE_INACTIVE) |
3536 | goto skip_active_search; |
3536 | goto skip_active_search; |
3537 | 3537 | ||
3538 | /* Best active match */ |
3538 | /* Best active match */ |
3539 | retry = NUM_CACHE_BUCKETS - bucket; |
3539 | retry = NUM_CACHE_BUCKETS - bucket; |
3540 | if (retry > 3 && (flags & CREATE_TEMPORARY) == 0) |
3540 | if (retry > 3 && (flags & CREATE_TEMPORARY) == 0) |
3541 | retry = 3; |
3541 | retry = 3; |
3542 | search_again: |
3542 | search_again: |
3543 | assert(bucket < NUM_CACHE_BUCKETS); |
3543 | assert(bucket < NUM_CACHE_BUCKETS); |
3544 | cache = &kgem->active[bucket][tiling]; |
3544 | cache = &kgem->active[bucket][tiling]; |
3545 | if (tiling) { |
3545 | if (tiling) { |
3546 | tiled_height = kgem_aligned_height(kgem, height, tiling); |
3546 | tiled_height = kgem_aligned_height(kgem, height, tiling); |
3547 | list_for_each_entry(bo, cache, list) { |
3547 | list_for_each_entry(bo, cache, list) { |
3548 | assert(!bo->purged); |
3548 | assert(!bo->purged); |
3549 | assert(bo->refcnt == 0); |
3549 | assert(bo->refcnt == 0); |
3550 | assert(bucket(bo) == bucket); |
3550 | assert(bucket(bo) == bucket); |
3551 | assert(bo->reusable); |
3551 | assert(bo->reusable); |
3552 | assert(bo->tiling == tiling); |
3552 | assert(bo->tiling == tiling); |
3553 | assert(bo->flush == false); |
3553 | assert(bo->flush == false); |
3554 | assert(!bo->scanout); |
3554 | assert(!bo->scanout); |
3555 | assert_tiling(kgem, bo); |
3555 | assert_tiling(kgem, bo); |
3556 | 3556 | ||
3557 | if (kgem->gen < 040) { |
3557 | if (kgem->gen < 040) { |
3558 | if (bo->pitch < pitch) { |
3558 | if (bo->pitch < pitch) { |
3559 | DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n", |
3559 | DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n", |
3560 | bo->tiling, tiling, |
3560 | bo->tiling, tiling, |
3561 | bo->pitch, pitch)); |
3561 | bo->pitch, pitch)); |
3562 | continue; |
3562 | continue; |
3563 | } |
3563 | } |
3564 | 3564 | ||
3565 | if (bo->pitch * tiled_height > bytes(bo)) |
3565 | if (bo->pitch * tiled_height > bytes(bo)) |
3566 | continue; |
3566 | continue; |
3567 | } else { |
3567 | } else { |
3568 | if (num_pages(bo) < size) |
3568 | if (num_pages(bo) < size) |
3569 | continue; |
3569 | continue; |
3570 | 3570 | ||
3571 | if (bo->pitch != pitch) { |
3571 | if (bo->pitch != pitch) { |
3572 | if (!gem_set_tiling(kgem->fd, |
3572 | if (!gem_set_tiling(kgem->fd, |
3573 | bo->handle, |
3573 | bo->handle, |
3574 | tiling, pitch)) |
3574 | tiling, pitch)) |
3575 | continue; |
3575 | continue; |
3576 | 3576 | ||
3577 | bo->pitch = pitch; |
3577 | bo->pitch = pitch; |
3578 | } |
3578 | } |
3579 | } |
3579 | } |
3580 | 3580 | ||
3581 | kgem_bo_remove_from_active(kgem, bo); |
3581 | kgem_bo_remove_from_active(kgem, bo); |
3582 | 3582 | ||
3583 | bo->unique_id = kgem_get_unique_id(kgem); |
3583 | bo->unique_id = kgem_get_unique_id(kgem); |
3584 | bo->delta = 0; |
3584 | bo->delta = 0; |
3585 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3585 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3586 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3586 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3587 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3587 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3588 | assert_tiling(kgem, bo); |
3588 | assert_tiling(kgem, bo); |
3589 | bo->refcnt = 1; |
3589 | bo->refcnt = 1; |
3590 | return bo; |
3590 | return bo; |
3591 | } |
3591 | } |
3592 | } else { |
3592 | } else { |
3593 | list_for_each_entry(bo, cache, list) { |
3593 | list_for_each_entry(bo, cache, list) { |
3594 | assert(bucket(bo) == bucket); |
3594 | assert(bucket(bo) == bucket); |
3595 | assert(!bo->purged); |
3595 | assert(!bo->purged); |
3596 | assert(bo->refcnt == 0); |
3596 | assert(bo->refcnt == 0); |
3597 | assert(bo->reusable); |
3597 | assert(bo->reusable); |
3598 | assert(!bo->scanout); |
3598 | assert(!bo->scanout); |
3599 | assert(bo->tiling == tiling); |
3599 | assert(bo->tiling == tiling); |
3600 | assert(bo->flush == false); |
3600 | assert(bo->flush == false); |
3601 | assert_tiling(kgem, bo); |
3601 | assert_tiling(kgem, bo); |
3602 | 3602 | ||
3603 | if (num_pages(bo) < size) |
3603 | if (num_pages(bo) < size) |
3604 | continue; |
3604 | continue; |
3605 | 3605 | ||
3606 | kgem_bo_remove_from_active(kgem, bo); |
3606 | kgem_bo_remove_from_active(kgem, bo); |
3607 | 3607 | ||
3608 | bo->pitch = pitch; |
3608 | bo->pitch = pitch; |
3609 | bo->unique_id = kgem_get_unique_id(kgem); |
3609 | bo->unique_id = kgem_get_unique_id(kgem); |
3610 | bo->delta = 0; |
3610 | bo->delta = 0; |
3611 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3611 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3612 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3612 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3613 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3613 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3614 | assert_tiling(kgem, bo); |
3614 | assert_tiling(kgem, bo); |
3615 | bo->refcnt = 1; |
3615 | bo->refcnt = 1; |
3616 | return bo; |
3616 | return bo; |
3617 | } |
3617 | } |
3618 | } |
3618 | } |
3619 | 3619 | ||
3620 | if (--retry && exact) { |
3620 | if (--retry && exact) { |
3621 | if (kgem->gen >= 040) { |
3621 | if (kgem->gen >= 040) { |
3622 | for (i = I915_TILING_NONE; i <= I915_TILING_Y; i++) { |
3622 | for (i = I915_TILING_NONE; i <= I915_TILING_Y; i++) { |
3623 | if (i == tiling) |
3623 | if (i == tiling) |
3624 | continue; |
3624 | continue; |
3625 | 3625 | ||
3626 | cache = &kgem->active[bucket][i]; |
3626 | cache = &kgem->active[bucket][i]; |
3627 | list_for_each_entry(bo, cache, list) { |
3627 | list_for_each_entry(bo, cache, list) { |
3628 | assert(!bo->purged); |
3628 | assert(!bo->purged); |
3629 | assert(bo->refcnt == 0); |
3629 | assert(bo->refcnt == 0); |
3630 | assert(bo->reusable); |
3630 | assert(bo->reusable); |
3631 | assert(!bo->scanout); |
3631 | assert(!bo->scanout); |
3632 | assert(bo->flush == false); |
3632 | assert(bo->flush == false); |
3633 | assert_tiling(kgem, bo); |
3633 | assert_tiling(kgem, bo); |
3634 | 3634 | ||
3635 | if (num_pages(bo) < size) |
3635 | if (num_pages(bo) < size) |
3636 | continue; |
3636 | continue; |
3637 | 3637 | ||
3638 | if (!gem_set_tiling(kgem->fd, |
3638 | if (!gem_set_tiling(kgem->fd, |
3639 | bo->handle, |
3639 | bo->handle, |
3640 | tiling, pitch)) |
3640 | tiling, pitch)) |
3641 | continue; |
3641 | continue; |
3642 | 3642 | ||
3643 | kgem_bo_remove_from_active(kgem, bo); |
3643 | kgem_bo_remove_from_active(kgem, bo); |
3644 | 3644 | ||
3645 | bo->unique_id = kgem_get_unique_id(kgem); |
3645 | bo->unique_id = kgem_get_unique_id(kgem); |
3646 | bo->pitch = pitch; |
3646 | bo->pitch = pitch; |
3647 | bo->tiling = tiling; |
3647 | bo->tiling = tiling; |
3648 | bo->delta = 0; |
3648 | bo->delta = 0; |
3649 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3649 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3650 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3650 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3651 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3651 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3652 | assert_tiling(kgem, bo); |
3652 | assert_tiling(kgem, bo); |
3653 | bo->refcnt = 1; |
3653 | bo->refcnt = 1; |
3654 | return bo; |
3654 | return bo; |
3655 | } |
3655 | } |
3656 | } |
3656 | } |
3657 | } |
3657 | } |
3658 | 3658 | ||
3659 | bucket++; |
3659 | bucket++; |
3660 | goto search_again; |
3660 | goto search_again; |
3661 | } |
3661 | } |
3662 | 3662 | ||
3663 | if (!exact) { /* allow an active near-miss? */ |
3663 | if (!exact) { /* allow an active near-miss? */ |
3664 | i = tiling; |
3664 | i = tiling; |
3665 | while (--i >= 0) { |
3665 | while (--i >= 0) { |
3666 | tiled_height = kgem_surface_size(kgem, kgem->has_relaxed_fencing, flags, |
3666 | tiled_height = kgem_surface_size(kgem, kgem->has_relaxed_fencing, flags, |
3667 | width, height, bpp, tiling, &pitch); |
3667 | width, height, bpp, tiling, &pitch); |
3668 | cache = active(kgem, tiled_height / PAGE_SIZE, i); |
3668 | cache = active(kgem, tiled_height / PAGE_SIZE, i); |
3669 | tiled_height = kgem_aligned_height(kgem, height, i); |
3669 | tiled_height = kgem_aligned_height(kgem, height, i); |
3670 | list_for_each_entry(bo, cache, list) { |
3670 | list_for_each_entry(bo, cache, list) { |
3671 | assert(!bo->purged); |
3671 | assert(!bo->purged); |
3672 | assert(bo->refcnt == 0); |
3672 | assert(bo->refcnt == 0); |
3673 | assert(bo->reusable); |
3673 | assert(bo->reusable); |
3674 | assert(!bo->scanout); |
3674 | assert(!bo->scanout); |
3675 | assert(bo->flush == false); |
3675 | assert(bo->flush == false); |
3676 | assert_tiling(kgem, bo); |
3676 | assert_tiling(kgem, bo); |
3677 | 3677 | ||
3678 | if (bo->tiling) { |
3678 | if (bo->tiling) { |
3679 | if (bo->pitch < pitch) { |
3679 | if (bo->pitch < pitch) { |
3680 | DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n", |
3680 | DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n", |
3681 | bo->tiling, tiling, |
3681 | bo->tiling, tiling, |
3682 | bo->pitch, pitch)); |
3682 | bo->pitch, pitch)); |
3683 | continue; |
3683 | continue; |
3684 | } |
3684 | } |
3685 | } else |
3685 | } else |
3686 | bo->pitch = pitch; |
3686 | bo->pitch = pitch; |
3687 | 3687 | ||
3688 | if (bo->pitch * tiled_height > bytes(bo)) |
3688 | if (bo->pitch * tiled_height > bytes(bo)) |
3689 | continue; |
3689 | continue; |
3690 | 3690 | ||
3691 | kgem_bo_remove_from_active(kgem, bo); |
3691 | kgem_bo_remove_from_active(kgem, bo); |
3692 | 3692 | ||
3693 | bo->unique_id = kgem_get_unique_id(kgem); |
3693 | bo->unique_id = kgem_get_unique_id(kgem); |
3694 | bo->delta = 0; |
3694 | bo->delta = 0; |
3695 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3695 | DBG((" 1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n", |
3696 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3696 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3697 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3697 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3698 | assert_tiling(kgem, bo); |
3698 | assert_tiling(kgem, bo); |
3699 | bo->refcnt = 1; |
3699 | bo->refcnt = 1; |
3700 | return bo; |
3700 | return bo; |
3701 | } |
3701 | } |
3702 | } |
3702 | } |
3703 | } |
3703 | } |
3704 | 3704 | ||
3705 | skip_active_search: |
3705 | skip_active_search: |
3706 | bucket = cache_bucket(size); |
3706 | bucket = cache_bucket(size); |
3707 | retry = NUM_CACHE_BUCKETS - bucket; |
3707 | retry = NUM_CACHE_BUCKETS - bucket; |
3708 | if (retry > 3) |
3708 | if (retry > 3) |
3709 | retry = 3; |
3709 | retry = 3; |
3710 | search_inactive: |
3710 | search_inactive: |
3711 | /* Now just look for a close match and prefer any currently active */ |
3711 | /* Now just look for a close match and prefer any currently active */ |
3712 | assert(bucket < NUM_CACHE_BUCKETS); |
3712 | assert(bucket < NUM_CACHE_BUCKETS); |
3713 | cache = &kgem->inactive[bucket]; |
3713 | cache = &kgem->inactive[bucket]; |
3714 | list_for_each_entry(bo, cache, list) { |
3714 | list_for_each_entry(bo, cache, list) { |
3715 | assert(bucket(bo) == bucket); |
3715 | assert(bucket(bo) == bucket); |
3716 | assert(bo->reusable); |
3716 | assert(bo->reusable); |
3717 | assert(!bo->scanout); |
3717 | assert(!bo->scanout); |
3718 | assert(bo->flush == false); |
3718 | assert(bo->flush == false); |
3719 | assert_tiling(kgem, bo); |
3719 | assert_tiling(kgem, bo); |
3720 | 3720 | ||
3721 | if (size > num_pages(bo)) { |
3721 | if (size > num_pages(bo)) { |
3722 | DBG(("inactive too small: %d < %d\n", |
3722 | DBG(("inactive too small: %d < %d\n", |
3723 | num_pages(bo), size)); |
3723 | num_pages(bo), size)); |
3724 | continue; |
3724 | continue; |
3725 | } |
3725 | } |
3726 | 3726 | ||
3727 | if (bo->tiling != tiling || |
3727 | if (bo->tiling != tiling || |
3728 | (tiling != I915_TILING_NONE && bo->pitch != pitch)) { |
3728 | (tiling != I915_TILING_NONE && bo->pitch != pitch)) { |
3729 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3729 | if (!gem_set_tiling(kgem->fd, bo->handle, |
3730 | tiling, pitch)) |
3730 | tiling, pitch)) |
3731 | continue; |
3731 | continue; |
3732 | 3732 | ||
3733 | if (bo->map) |
3733 | if (bo->map) |
3734 | kgem_bo_release_map(kgem, bo); |
3734 | kgem_bo_release_map(kgem, bo); |
3735 | } |
3735 | } |
3736 | 3736 | ||
3737 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3737 | if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) { |
3738 | kgem_bo_free(kgem, bo); |
3738 | kgem_bo_free(kgem, bo); |
3739 | break; |
3739 | break; |
3740 | } |
3740 | } |
3741 | 3741 | ||
3742 | kgem_bo_remove_from_inactive(kgem, bo); |
3742 | kgem_bo_remove_from_inactive(kgem, bo); |
3743 | 3743 | ||
3744 | bo->pitch = pitch; |
3744 | bo->pitch = pitch; |
3745 | bo->tiling = tiling; |
3745 | bo->tiling = tiling; |
3746 | 3746 | ||
3747 | bo->delta = 0; |
3747 | bo->delta = 0; |
3748 | bo->unique_id = kgem_get_unique_id(kgem); |
3748 | bo->unique_id = kgem_get_unique_id(kgem); |
3749 | assert(bo->pitch); |
3749 | assert(bo->pitch); |
3750 | DBG((" from inactive: pitch=%d, tiling=%d: handle=%d, id=%d\n", |
3750 | DBG((" from inactive: pitch=%d, tiling=%d: handle=%d, id=%d\n", |
3751 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3751 | bo->pitch, bo->tiling, bo->handle, bo->unique_id)); |
3752 | assert(bo->refcnt == 0); |
3752 | assert(bo->refcnt == 0); |
3753 | assert(bo->reusable); |
3753 | assert(bo->reusable); |
3754 | assert((flags & CREATE_INACTIVE) == 0 || bo->domain != DOMAIN_GPU); |
3754 | assert((flags & CREATE_INACTIVE) == 0 || bo->domain != DOMAIN_GPU); |
3755 | ASSERT_MAYBE_IDLE(kgem, bo->handle, flags & CREATE_INACTIVE); |
3755 | ASSERT_MAYBE_IDLE(kgem, bo->handle, flags & CREATE_INACTIVE); |
3756 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3756 | assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo)); |
3757 | assert_tiling(kgem, bo); |
3757 | assert_tiling(kgem, bo); |
3758 | bo->refcnt = 1; |
3758 | bo->refcnt = 1; |
3759 | return bo; |
3759 | return bo; |
3760 | } |
3760 | } |
3761 | 3761 | ||
3762 | if (flags & CREATE_INACTIVE && |
3762 | if (flags & CREATE_INACTIVE && |
3763 | !list_is_empty(&kgem->active[bucket][tiling]) && |
3763 | !list_is_empty(&kgem->active[bucket][tiling]) && |
3764 | __kgem_throttle_retire(kgem, flags)) { |
3764 | __kgem_throttle_retire(kgem, flags)) { |
3765 | flags &= ~CREATE_INACTIVE; |
3765 | flags &= ~CREATE_INACTIVE; |
3766 | goto search_inactive; |
3766 | goto search_inactive; |
3767 | } |
3767 | } |
3768 | 3768 | ||
3769 | if (--retry) { |
3769 | if (--retry) { |
3770 | bucket++; |
3770 | bucket++; |
3771 | flags &= ~CREATE_INACTIVE; |
3771 | flags &= ~CREATE_INACTIVE; |
3772 | goto search_inactive; |
3772 | goto search_inactive; |
3773 | } |
3773 | } |
3774 | 3774 | ||
3775 | create: |
3775 | create: |
3776 | if (flags & CREATE_CACHED) |
3776 | if (flags & CREATE_CACHED) |
3777 | return NULL; |
3777 | return NULL; |
3778 | 3778 | ||
3779 | if (bucket >= NUM_CACHE_BUCKETS) |
3779 | if (bucket >= NUM_CACHE_BUCKETS) |
3780 | size = ALIGN(size, 1024); |
3780 | size = ALIGN(size, 1024); |
3781 | handle = gem_create(kgem->fd, size); |
3781 | handle = gem_create(kgem->fd, size); |
3782 | if (handle == 0) |
3782 | if (handle == 0) |
3783 | return NULL; |
3783 | return NULL; |
3784 | 3784 | ||
3785 | bo = __kgem_bo_alloc(handle, size); |
3785 | bo = __kgem_bo_alloc(handle, size); |
3786 | if (!bo) { |
3786 | if (!bo) { |
3787 | gem_close(kgem->fd, handle); |
3787 | gem_close(kgem->fd, handle); |
3788 | return NULL; |
3788 | return NULL; |
3789 | } |
3789 | } |
3790 | 3790 | ||
3791 | if (bucket >= NUM_CACHE_BUCKETS) { |
3791 | if (bucket >= NUM_CACHE_BUCKETS) { |
3792 | DBG(("%s: marking large bo for automatic flushing\n", |
3792 | DBG(("%s: marking large bo for automatic flushing\n", |
3793 | __FUNCTION__)); |
3793 | __FUNCTION__)); |
3794 | bo->flush = true; |
3794 | bo->flush = true; |
3795 | } |
3795 | } |
3796 | 3796 | ||
3797 | bo->unique_id = kgem_get_unique_id(kgem); |
3797 | bo->unique_id = kgem_get_unique_id(kgem); |
3798 | if (tiling == I915_TILING_NONE || |
3798 | if (tiling == I915_TILING_NONE || |
3799 | gem_set_tiling(kgem->fd, handle, tiling, pitch)) { |
3799 | gem_set_tiling(kgem->fd, handle, tiling, pitch)) { |
3800 | bo->tiling = tiling; |
3800 | bo->tiling = tiling; |
3801 | bo->pitch = pitch; |
3801 | bo->pitch = pitch; |
3802 | } else { |
3802 | } else { |
3803 | if (flags & CREATE_EXACT) { |
3803 | if (flags & CREATE_EXACT) { |
3804 | if (bo->pitch != pitch || bo->tiling != tiling) { |
3804 | if (bo->pitch != pitch || bo->tiling != tiling) { |
3805 | kgem_bo_free(kgem, bo); |
3805 | kgem_bo_free(kgem, bo); |
3806 | return NULL; |
3806 | return NULL; |
3807 | } |
3807 | } |
3808 | } |
3808 | } |
3809 | } |
3809 | } |
3810 | 3810 | ||
3811 | assert(bytes(bo) >= bo->pitch * kgem_aligned_height(kgem, height, bo->tiling)); |
3811 | assert(bytes(bo) >= bo->pitch * kgem_aligned_height(kgem, height, bo->tiling)); |
3812 | assert_tiling(kgem, bo); |
3812 | assert_tiling(kgem, bo); |
3813 | 3813 | ||
3814 | debug_alloc__bo(kgem, bo); |
3814 | debug_alloc__bo(kgem, bo); |
3815 | 3815 | ||
3816 | DBG((" new pitch=%d, tiling=%d, handle=%d, id=%d, num_pages=%d [%d], bucket=%d\n", |
3816 | DBG((" new pitch=%d, tiling=%d, handle=%d, id=%d, num_pages=%d [%d], bucket=%d\n", |
3817 | bo->pitch, bo->tiling, bo->handle, bo->unique_id, |
3817 | bo->pitch, bo->tiling, bo->handle, bo->unique_id, |
3818 | size, num_pages(bo), bucket(bo))); |
3818 | size, num_pages(bo), bucket(bo))); |
3819 | return bo; |
3819 | return bo; |
3820 | } |
3820 | } |
3821 | 3821 | ||
3822 | #if 0 |
3822 | #if 0 |
3823 | struct kgem_bo *kgem_create_cpu_2d(struct kgem *kgem, |
3823 | struct kgem_bo *kgem_create_cpu_2d(struct kgem *kgem, |
3824 | int width, |
3824 | int width, |
3825 | int height, |
3825 | int height, |
3826 | int bpp, |
3826 | int bpp, |
3827 | uint32_t flags) |
3827 | uint32_t flags) |
3828 | { |
3828 | { |
3829 | struct kgem_bo *bo; |
3829 | struct kgem_bo *bo; |
3830 | int stride, size; |
3830 | int stride, size; |
3831 | 3831 | ||
3832 | if (DBG_NO_CPU) |
3832 | if (DBG_NO_CPU) |
3833 | return NULL; |
3833 | return NULL; |
3834 | 3834 | ||
3835 | DBG(("%s(%dx%d, bpp=%d)\n", __FUNCTION__, width, height, bpp)); |
3835 | DBG(("%s(%dx%d, bpp=%d)\n", __FUNCTION__, width, height, bpp)); |
3836 | 3836 | ||
3837 | if (kgem->has_llc) { |
3837 | if (kgem->has_llc) { |
3838 | bo = kgem_create_2d(kgem, width, height, bpp, |
3838 | bo = kgem_create_2d(kgem, width, height, bpp, |
3839 | I915_TILING_NONE, flags); |
3839 | I915_TILING_NONE, flags); |
3840 | if (bo == NULL) |
3840 | if (bo == NULL) |
3841 | return bo; |
3841 | return bo; |
3842 | 3842 | ||
3843 | assert(bo->tiling == I915_TILING_NONE); |
3843 | assert(bo->tiling == I915_TILING_NONE); |
3844 | assert_tiling(kgem, bo); |
3844 | assert_tiling(kgem, bo); |
3845 | 3845 | ||
3846 | if (kgem_bo_map__cpu(kgem, bo) == NULL) { |
3846 | if (kgem_bo_map__cpu(kgem, bo) == NULL) { |
3847 | kgem_bo_destroy(kgem, bo); |
3847 | kgem_bo_destroy(kgem, bo); |
3848 | return NULL; |
3848 | return NULL; |
3849 | } |
3849 | } |
3850 | 3850 | ||
3851 | return bo; |
3851 | return bo; |
3852 | } |
3852 | } |
3853 | 3853 | ||
3854 | assert(width > 0 && height > 0); |
3854 | assert(width > 0 && height > 0); |
3855 | stride = ALIGN(width, 2) * bpp >> 3; |
3855 | stride = ALIGN(width, 2) * bpp >> 3; |
3856 | stride = ALIGN(stride, 4); |
3856 | stride = ALIGN(stride, 4); |
3857 | size = stride * ALIGN(height, 2); |
3857 | size = stride * ALIGN(height, 2); |
3858 | assert(size >= PAGE_SIZE); |
3858 | assert(size >= PAGE_SIZE); |
3859 | 3859 | ||
3860 | DBG(("%s: %dx%d, %d bpp, stride=%d\n", |
3860 | DBG(("%s: %dx%d, %d bpp, stride=%d\n", |
3861 | __FUNCTION__, width, height, bpp, stride)); |
3861 | __FUNCTION__, width, height, bpp, stride)); |
3862 | 3862 | ||
3863 | bo = search_snoop_cache(kgem, NUM_PAGES(size), 0); |
3863 | bo = search_snoop_cache(kgem, NUM_PAGES(size), 0); |
3864 | if (bo) { |
3864 | if (bo) { |
3865 | assert(bo->tiling == I915_TILING_NONE); |
3865 | assert(bo->tiling == I915_TILING_NONE); |
3866 | assert_tiling(kgem, bo); |
3866 | assert_tiling(kgem, bo); |
3867 | assert(bo->snoop); |
3867 | assert(bo->snoop); |
3868 | bo->refcnt = 1; |
3868 | bo->refcnt = 1; |
3869 | bo->pitch = stride; |
3869 | bo->pitch = stride; |
3870 | bo->unique_id = kgem_get_unique_id(kgem); |
3870 | bo->unique_id = kgem_get_unique_id(kgem); |
3871 | return bo; |
3871 | return bo; |
3872 | } |
3872 | } |
3873 | 3873 | ||
3874 | if (kgem->has_caching) { |
3874 | if (kgem->has_caching) { |
3875 | bo = kgem_create_linear(kgem, size, flags); |
3875 | bo = kgem_create_linear(kgem, size, flags); |
3876 | if (bo == NULL) |
3876 | if (bo == NULL) |
3877 | return NULL; |
3877 | return NULL; |
3878 | 3878 | ||
3879 | assert(bo->tiling == I915_TILING_NONE); |
3879 | assert(bo->tiling == I915_TILING_NONE); |
3880 | assert_tiling(kgem, bo); |
3880 | assert_tiling(kgem, bo); |
3881 | 3881 | ||
3882 | if (!gem_set_caching(kgem->fd, bo->handle, SNOOPED)) { |
3882 | if (!gem_set_caching(kgem->fd, bo->handle, SNOOPED)) { |
3883 | kgem_bo_destroy(kgem, bo); |
3883 | kgem_bo_destroy(kgem, bo); |
3884 | return NULL; |
3884 | return NULL; |
3885 | } |
3885 | } |
3886 | bo->snoop = true; |
3886 | bo->snoop = true; |
3887 | 3887 | ||
3888 | if (kgem_bo_map__cpu(kgem, bo) == NULL) { |
3888 | if (kgem_bo_map__cpu(kgem, bo) == NULL) { |
3889 | kgem_bo_destroy(kgem, bo); |
3889 | kgem_bo_destroy(kgem, bo); |
3890 | return NULL; |
3890 | return NULL; |
3891 | } |
3891 | } |
3892 | 3892 | ||
3893 | bo->pitch = stride; |
3893 | bo->pitch = stride; |
3894 | bo->unique_id = kgem_get_unique_id(kgem); |
3894 | bo->unique_id = kgem_get_unique_id(kgem); |
3895 | return bo; |
3895 | return bo; |
3896 | } |
3896 | } |
3897 | 3897 | ||
3898 | if (kgem->has_userptr) { |
3898 | if (kgem->has_userptr) { |
3899 | void *ptr; |
3899 | void *ptr; |
3900 | 3900 | ||
3901 | /* XXX */ |
3901 | /* XXX */ |
3902 | //if (posix_memalign(&ptr, 64, ALIGN(size, 64))) |
3902 | //if (posix_memalign(&ptr, 64, ALIGN(size, 64))) |
3903 | if (posix_memalign(&ptr, PAGE_SIZE, ALIGN(size, PAGE_SIZE))) |
3903 | if (posix_memalign(&ptr, PAGE_SIZE, ALIGN(size, PAGE_SIZE))) |
3904 | return NULL; |
3904 | return NULL; |
3905 | 3905 | ||
3906 | bo = kgem_create_map(kgem, ptr, size, false); |
3906 | bo = kgem_create_map(kgem, ptr, size, false); |
3907 | if (bo == NULL) { |
3907 | if (bo == NULL) { |
3908 | free(ptr); |
3908 | free(ptr); |
3909 | return NULL; |
3909 | return NULL; |
3910 | } |
3910 | } |
3911 | 3911 | ||
3912 | bo->pitch = stride; |
3912 | bo->pitch = stride; |
3913 | bo->unique_id = kgem_get_unique_id(kgem); |
3913 | bo->unique_id = kgem_get_unique_id(kgem); |
3914 | return bo; |
3914 | return bo; |
3915 | } |
3915 | } |
3916 | 3916 | ||
3917 | return NULL; |
3917 | return NULL; |
3918 | } |
3918 | } |
3919 | #endif |
3919 | #endif |
3920 | 3920 | ||
3921 | void _kgem_bo_destroy(struct kgem *kgem, struct kgem_bo *bo) |
3921 | void _kgem_bo_destroy(struct kgem *kgem, struct kgem_bo *bo) |
3922 | { |
3922 | { |
3923 | DBG(("%s: handle=%d, proxy? %d\n", |
3923 | DBG(("%s: handle=%d, proxy? %d\n", |
3924 | __FUNCTION__, bo->handle, bo->proxy != NULL)); |
3924 | __FUNCTION__, bo->handle, bo->proxy != NULL)); |
3925 | 3925 | ||
3926 | if (bo->proxy) { |
3926 | if (bo->proxy) { |
3927 | _list_del(&bo->vma); |
3927 | _list_del(&bo->vma); |
3928 | _list_del(&bo->request); |
3928 | _list_del(&bo->request); |
3929 | if (bo->io && bo->exec == NULL) |
3929 | if (bo->io && bo->exec == NULL) |
3930 | _kgem_bo_delete_buffer(kgem, bo); |
3930 | _kgem_bo_delete_buffer(kgem, bo); |
3931 | kgem_bo_unref(kgem, bo->proxy); |
3931 | kgem_bo_unref(kgem, bo->proxy); |
3932 | kgem_bo_binding_free(kgem, bo); |
3932 | kgem_bo_binding_free(kgem, bo); |
3933 | free(bo); |
3933 | free(bo); |
3934 | return; |
3934 | return; |
3935 | } |
3935 | } |
3936 | 3936 | ||
3937 | __kgem_bo_destroy(kgem, bo); |
3937 | __kgem_bo_destroy(kgem, bo); |
3938 | } |
3938 | } |
3939 | 3939 | ||
3940 | static void __kgem_flush(struct kgem *kgem, struct kgem_bo *bo) |
3940 | static void __kgem_flush(struct kgem *kgem, struct kgem_bo *bo) |
3941 | { |
3941 | { |
3942 | assert(bo->rq); |
3942 | assert(bo->rq); |
3943 | assert(bo->exec == NULL); |
3943 | assert(bo->exec == NULL); |
3944 | assert(bo->needs_flush); |
3944 | assert(bo->needs_flush); |
3945 | 3945 | ||
3946 | /* The kernel will emit a flush *and* update its own flushing lists. */ |
3946 | /* The kernel will emit a flush *and* update its own flushing lists. */ |
3947 | if (!__kgem_busy(kgem, bo->handle)) |
3947 | if (!__kgem_busy(kgem, bo->handle)) |
3948 | __kgem_bo_clear_busy(bo); |
3948 | __kgem_bo_clear_busy(bo); |
3949 | 3949 | ||
3950 | DBG(("%s: handle=%d, busy?=%d\n", |
3950 | DBG(("%s: handle=%d, busy?=%d\n", |
3951 | __FUNCTION__, bo->handle, bo->rq != NULL)); |
3951 | __FUNCTION__, bo->handle, bo->rq != NULL)); |
3952 | } |
3952 | } |
3953 | 3953 | ||
3954 | void kgem_scanout_flush(struct kgem *kgem, struct kgem_bo *bo) |
3954 | void kgem_scanout_flush(struct kgem *kgem, struct kgem_bo *bo) |
3955 | { |
3955 | { |
3956 | kgem_bo_submit(kgem, bo); |
3956 | kgem_bo_submit(kgem, bo); |
3957 | if (!bo->needs_flush) |
3957 | if (!bo->needs_flush) |
3958 | return; |
3958 | return; |
3959 | 3959 | ||
3960 | /* If the kernel fails to emit the flush, then it will be forced when |
3960 | /* If the kernel fails to emit the flush, then it will be forced when |
3961 | * we assume direct access. And as the usual failure is EIO, we do |
3961 | * we assume direct access. And as the usual failure is EIO, we do |
3962 | * not actually care. |
3962 | * not actually care. |
3963 | */ |
3963 | */ |
3964 | assert(bo->exec == NULL); |
3964 | assert(bo->exec == NULL); |
3965 | if (bo->rq) |
3965 | if (bo->rq) |
3966 | __kgem_flush(kgem, bo); |
3966 | __kgem_flush(kgem, bo); |
3967 | 3967 | ||
3968 | /* Whatever actually happens, we can regard the GTT write domain |
3968 | /* Whatever actually happens, we can regard the GTT write domain |
3969 | * as being flushed. |
3969 | * as being flushed. |
3970 | */ |
3970 | */ |
3971 | bo->gtt_dirty = false; |
3971 | bo->gtt_dirty = false; |
3972 | bo->needs_flush = false; |
3972 | bo->needs_flush = false; |
3973 | bo->domain = DOMAIN_NONE; |
3973 | bo->domain = DOMAIN_NONE; |
3974 | } |
3974 | } |
3975 | 3975 | ||
3976 | inline static bool needs_semaphore(struct kgem *kgem, struct kgem_bo *bo) |
3976 | inline static bool needs_semaphore(struct kgem *kgem, struct kgem_bo *bo) |
3977 | { |
3977 | { |
3978 | return kgem->nreloc && bo->rq && RQ_RING(bo->rq) != kgem->ring; |
3978 | return kgem->nreloc && bo->rq && RQ_RING(bo->rq) != kgem->ring; |
3979 | } |
3979 | } |
3980 | 3980 | ||
3981 | bool kgem_check_bo(struct kgem *kgem, ...) |
3981 | bool kgem_check_bo(struct kgem *kgem, ...) |
3982 | { |
3982 | { |
3983 | va_list ap; |
3983 | va_list ap; |
3984 | struct kgem_bo *bo; |
3984 | struct kgem_bo *bo; |
3985 | int num_exec = 0; |
3985 | int num_exec = 0; |
3986 | int num_pages = 0; |
3986 | int num_pages = 0; |
3987 | bool flush = false; |
3987 | bool flush = false; |
3988 | 3988 | ||
3989 | va_start(ap, kgem); |
3989 | va_start(ap, kgem); |
3990 | while ((bo = va_arg(ap, struct kgem_bo *))) { |
3990 | while ((bo = va_arg(ap, struct kgem_bo *))) { |
3991 | while (bo->proxy) |
3991 | while (bo->proxy) |
3992 | bo = bo->proxy; |
3992 | bo = bo->proxy; |
3993 | if (bo->exec) |
3993 | if (bo->exec) |
3994 | continue; |
3994 | continue; |
3995 | 3995 | ||
3996 | if (needs_semaphore(kgem, bo)) |
3996 | if (needs_semaphore(kgem, bo)) |
3997 | return false; |
3997 | return false; |
3998 | 3998 | ||
3999 | num_pages += num_pages(bo); |
3999 | num_pages += num_pages(bo); |
4000 | num_exec++; |
4000 | num_exec++; |
4001 | 4001 | ||
4002 | flush |= bo->flush; |
4002 | flush |= bo->flush; |
4003 | } |
4003 | } |
4004 | va_end(ap); |
4004 | va_end(ap); |
4005 | 4005 | ||
4006 | DBG(("%s: num_pages=+%d, num_exec=+%d\n", |
4006 | DBG(("%s: num_pages=+%d, num_exec=+%d\n", |
4007 | __FUNCTION__, num_pages, num_exec)); |
4007 | __FUNCTION__, num_pages, num_exec)); |
4008 | 4008 | ||
4009 | if (!num_pages) |
4009 | if (!num_pages) |
4010 | return true; |
4010 | return true; |
4011 | 4011 | ||
4012 | if (kgem_flush(kgem, flush)) |
4012 | if (kgem_flush(kgem, flush)) |
4013 | return false; |
4013 | return false; |
4014 | 4014 | ||
4015 | if (kgem->aperture > kgem->aperture_low && |
4015 | if (kgem->aperture > kgem->aperture_low && |
4016 | kgem_ring_is_idle(kgem, kgem->ring)) { |
4016 | kgem_ring_is_idle(kgem, kgem->ring)) { |
4017 | DBG(("%s: current aperture usage (%d) is greater than low water mark (%d)\n", |
4017 | DBG(("%s: current aperture usage (%d) is greater than low water mark (%d)\n", |
4018 | __FUNCTION__, kgem->aperture, kgem->aperture_low)); |
4018 | __FUNCTION__, kgem->aperture, kgem->aperture_low)); |
4019 | return false; |
4019 | return false; |
4020 | } |
4020 | } |
4021 | 4021 | ||
4022 | if (num_pages + kgem->aperture > kgem->aperture_high) { |
4022 | if (num_pages + kgem->aperture > kgem->aperture_high) { |
4023 | DBG(("%s: final aperture usage (%d) is greater than high water mark (%d)\n", |
4023 | DBG(("%s: final aperture usage (%d) is greater than high water mark (%d)\n", |
4024 | __FUNCTION__, num_pages + kgem->aperture, kgem->aperture_high)); |
4024 | __FUNCTION__, num_pages + kgem->aperture, kgem->aperture_high)); |
4025 | return false; |
4025 | return false; |
4026 | } |
4026 | } |
4027 | 4027 | ||
4028 | if (kgem->nexec + num_exec >= KGEM_EXEC_SIZE(kgem)) { |
4028 | if (kgem->nexec + num_exec >= KGEM_EXEC_SIZE(kgem)) { |
4029 | DBG(("%s: out of exec slots (%d + %d / %d)\n", __FUNCTION__, |
4029 | DBG(("%s: out of exec slots (%d + %d / %d)\n", __FUNCTION__, |
4030 | kgem->nexec, num_exec, KGEM_EXEC_SIZE(kgem))); |
4030 | kgem->nexec, num_exec, KGEM_EXEC_SIZE(kgem))); |
4031 | return false; |
4031 | return false; |
4032 | } |
4032 | } |
4033 | 4033 | ||
4034 | return true; |
4034 | return true; |
4035 | } |
4035 | } |
4036 | 4036 | ||
4037 | 4037 | ||
4038 | 4038 | ||
4039 | 4039 | ||
4040 | 4040 | ||
4041 | 4041 | ||
4042 | 4042 | ||
4043 | 4043 | ||
4044 | 4044 | ||
4045 | 4045 | ||
4046 | 4046 | ||
4047 | 4047 | ||
4048 | 4048 | ||
4049 | 4049 | ||
4050 | 4050 | ||
4051 | 4051 | ||
4052 | 4052 | ||
4053 | 4053 | ||
4054 | 4054 | ||
4055 | 4055 | ||
4056 | 4056 | ||
4057 | 4057 | ||
4058 | 4058 | ||
4059 | 4059 | ||
4060 | 4060 | ||
4061 | 4061 | ||
4062 | 4062 | ||
4063 | 4063 | ||
4064 | 4064 | ||
4065 | 4065 | ||
4066 | uint32_t kgem_add_reloc(struct kgem *kgem, |
4066 | uint32_t kgem_add_reloc(struct kgem *kgem, |
4067 | uint32_t pos, |
4067 | uint32_t pos, |
4068 | struct kgem_bo *bo, |
4068 | struct kgem_bo *bo, |
4069 | uint32_t read_write_domain, |
4069 | uint32_t read_write_domain, |
4070 | uint32_t delta) |
4070 | uint32_t delta) |
4071 | { |
4071 | { |
4072 | int index; |
4072 | int index; |
4073 | 4073 | ||
4074 | DBG(("%s: handle=%d, pos=%d, delta=%d, domains=%08x\n", |
4074 | DBG(("%s: handle=%d, pos=%d, delta=%d, domains=%08x\n", |
4075 | __FUNCTION__, bo ? bo->handle : 0, pos, delta, read_write_domain)); |
4075 | __FUNCTION__, bo ? bo->handle : 0, pos, delta, read_write_domain)); |
4076 | 4076 | ||
4077 | assert((read_write_domain & 0x7fff) == 0 || bo != NULL); |
4077 | assert((read_write_domain & 0x7fff) == 0 || bo != NULL); |
4078 | 4078 | ||
4079 | if( bo != NULL && bo->handle == -2) |
4079 | if( bo != NULL && bo->handle == -2) |
4080 | { |
4080 | { |
4081 | if (bo->exec == NULL) |
4081 | if (bo->exec == NULL) |
4082 | kgem_add_bo(kgem, bo); |
4082 | kgem_add_bo(kgem, bo); |
4083 | 4083 | ||
4084 | if (read_write_domain & 0x7fff && !bo->gpu_dirty) { |
4084 | if (read_write_domain & 0x7fff && !bo->gpu_dirty) { |
4085 | __kgem_bo_mark_dirty(bo); |
4085 | __kgem_bo_mark_dirty(bo); |
4086 | } |
4086 | } |
4087 | return 0; |
4087 | return 0; |
4088 | }; |
4088 | }; |
4089 | 4089 | ||
4090 | index = kgem->nreloc++; |
4090 | index = kgem->nreloc++; |
4091 | assert(index < ARRAY_SIZE(kgem->reloc)); |
4091 | assert(index < ARRAY_SIZE(kgem->reloc)); |
4092 | kgem->reloc[index].offset = pos * sizeof(kgem->batch[0]); |
4092 | kgem->reloc[index].offset = pos * sizeof(kgem->batch[0]); |
4093 | if (bo) { |
4093 | if (bo) { |
4094 | assert(bo->refcnt); |
4094 | assert(bo->refcnt); |
4095 | while (bo->proxy) { |
4095 | while (bo->proxy) { |
4096 | DBG(("%s: adding proxy [delta=%d] for handle=%d\n", |
4096 | DBG(("%s: adding proxy [delta=%d] for handle=%d\n", |
4097 | __FUNCTION__, bo->delta, bo->handle)); |
4097 | __FUNCTION__, bo->delta, bo->handle)); |
4098 | delta += bo->delta; |
4098 | delta += bo->delta; |
4099 | assert(bo->handle == bo->proxy->handle); |
4099 | assert(bo->handle == bo->proxy->handle); |
4100 | /* need to release the cache upon batch submit */ |
4100 | /* need to release the cache upon batch submit */ |
4101 | if (bo->exec == NULL) { |
4101 | if (bo->exec == NULL) { |
4102 | list_move_tail(&bo->request, |
4102 | list_move_tail(&bo->request, |
4103 | &kgem->next_request->buffers); |
4103 | &kgem->next_request->buffers); |
4104 | bo->rq = MAKE_REQUEST(kgem->next_request, |
4104 | bo->rq = MAKE_REQUEST(kgem->next_request, |
4105 | kgem->ring); |
4105 | kgem->ring); |
4106 | bo->exec = &_kgem_dummy_exec; |
4106 | bo->exec = &_kgem_dummy_exec; |
4107 | } |
4107 | } |
4108 | 4108 | ||
4109 | if (read_write_domain & 0x7fff && !bo->gpu_dirty) |
4109 | if (read_write_domain & 0x7fff && !bo->gpu_dirty) |
4110 | __kgem_bo_mark_dirty(bo); |
4110 | __kgem_bo_mark_dirty(bo); |
4111 | 4111 | ||
4112 | bo = bo->proxy; |
4112 | bo = bo->proxy; |
4113 | assert(bo->refcnt); |
4113 | assert(bo->refcnt); |
4114 | } |
4114 | } |
4115 | assert(bo->refcnt); |
4115 | assert(bo->refcnt); |
4116 | 4116 | ||
4117 | if (bo->exec == NULL) |
4117 | if (bo->exec == NULL) |
4118 | kgem_add_bo(kgem, bo); |
4118 | kgem_add_bo(kgem, bo); |
4119 | assert(bo->rq == MAKE_REQUEST(kgem->next_request, kgem->ring)); |
4119 | assert(bo->rq == MAKE_REQUEST(kgem->next_request, kgem->ring)); |
4120 | assert(RQ_RING(bo->rq) == kgem->ring); |
4120 | assert(RQ_RING(bo->rq) == kgem->ring); |
4121 | 4121 | ||
4122 | if (kgem->gen < 040 && read_write_domain & KGEM_RELOC_FENCED) { |
4122 | if (kgem->gen < 040 && read_write_domain & KGEM_RELOC_FENCED) { |
4123 | if (bo->tiling && |
4123 | if (bo->tiling && |
4124 | (bo->exec->flags & EXEC_OBJECT_NEEDS_FENCE) == 0) { |
4124 | (bo->exec->flags & EXEC_OBJECT_NEEDS_FENCE) == 0) { |
4125 | assert(kgem->nfence < kgem->fence_max); |
4125 | assert(kgem->nfence < kgem->fence_max); |
4126 | kgem->aperture_fenced += |
4126 | kgem->aperture_fenced += |
4127 | kgem_bo_fenced_size(kgem, bo); |
4127 | kgem_bo_fenced_size(kgem, bo); |
4128 | kgem->nfence++; |
4128 | kgem->nfence++; |
4129 | } |
4129 | } |
4130 | bo->exec->flags |= EXEC_OBJECT_NEEDS_FENCE; |
4130 | bo->exec->flags |= EXEC_OBJECT_NEEDS_FENCE; |
4131 | } |
4131 | } |
4132 | 4132 | ||
4133 | kgem->reloc[index].delta = delta; |
4133 | kgem->reloc[index].delta = delta; |
4134 | kgem->reloc[index].target_handle = bo->target_handle; |
4134 | kgem->reloc[index].target_handle = bo->target_handle; |
4135 | kgem->reloc[index].presumed_offset = bo->presumed_offset; |
4135 | kgem->reloc[index].presumed_offset = bo->presumed_offset; |
4136 | 4136 | ||
4137 | if (read_write_domain & 0x7fff && !bo->gpu_dirty) { |
4137 | if (read_write_domain & 0x7fff && !bo->gpu_dirty) { |
4138 | assert(!bo->snoop || kgem->can_blt_cpu); |
4138 | assert(!bo->snoop || kgem->can_blt_cpu); |
4139 | __kgem_bo_mark_dirty(bo); |
4139 | __kgem_bo_mark_dirty(bo); |
4140 | } |
4140 | } |
4141 | 4141 | ||
4142 | delta += bo->presumed_offset; |
4142 | delta += bo->presumed_offset; |
4143 | } else { |
4143 | } else { |
4144 | kgem->reloc[index].delta = delta; |
4144 | kgem->reloc[index].delta = delta; |
4145 | kgem->reloc[index].target_handle = ~0U; |
4145 | kgem->reloc[index].target_handle = ~0U; |
4146 | kgem->reloc[index].presumed_offset = 0; |
4146 | kgem->reloc[index].presumed_offset = 0; |
4147 | if (kgem->nreloc__self < 256) |
4147 | if (kgem->nreloc__self < 256) |
4148 | kgem->reloc__self[kgem->nreloc__self++] = index; |
4148 | kgem->reloc__self[kgem->nreloc__self++] = index; |
4149 | } |
4149 | } |
4150 | kgem->reloc[index].read_domains = read_write_domain >> 16; |
4150 | kgem->reloc[index].read_domains = read_write_domain >> 16; |
4151 | kgem->reloc[index].write_domain = read_write_domain & 0x7fff; |
4151 | kgem->reloc[index].write_domain = read_write_domain & 0x7fff; |
4152 | 4152 | ||
4153 | return delta; |
4153 | return delta; |
4154 | } |
4154 | } |
4155 | 4155 | ||
4156 | static void kgem_trim_vma_cache(struct kgem *kgem, int type, int bucket) |
4156 | static void kgem_trim_vma_cache(struct kgem *kgem, int type, int bucket) |
4157 | { |
4157 | { |
4158 | int i, j; |
4158 | int i, j; |
4159 | 4159 | ||
4160 | DBG(("%s: type=%d, count=%d (bucket: %d)\n", |
4160 | DBG(("%s: type=%d, count=%d (bucket: %d)\n", |
4161 | __FUNCTION__, type, kgem->vma[type].count, bucket)); |
4161 | __FUNCTION__, type, kgem->vma[type].count, bucket)); |
4162 | if (kgem->vma[type].count <= 0) |
4162 | if (kgem->vma[type].count <= 0) |
4163 | return; |
4163 | return; |
4164 | 4164 | ||
4165 | if (kgem->need_purge) |
4165 | if (kgem->need_purge) |
4166 | kgem_purge_cache(kgem); |
4166 | kgem_purge_cache(kgem); |
4167 | 4167 | ||
4168 | /* vma are limited on a per-process basis to around 64k. |
4168 | /* vma are limited on a per-process basis to around 64k. |
4169 | * This includes all malloc arenas as well as other file |
4169 | * This includes all malloc arenas as well as other file |
4170 | * mappings. In order to be fair and not hog the cache, |
4170 | * mappings. In order to be fair and not hog the cache, |
4171 | * and more importantly not to exhaust that limit and to |
4171 | * and more importantly not to exhaust that limit and to |
4172 | * start failing mappings, we keep our own number of open |
4172 | * start failing mappings, we keep our own number of open |
4173 | * vma to within a conservative value. |
4173 | * vma to within a conservative value. |
4174 | */ |
4174 | */ |
4175 | i = 0; |
4175 | i = 0; |
4176 | while (kgem->vma[type].count > 0) { |
4176 | while (kgem->vma[type].count > 0) { |
4177 | struct kgem_bo *bo = NULL; |
4177 | struct kgem_bo *bo = NULL; |
4178 | 4178 | ||
4179 | for (j = 0; |
4179 | for (j = 0; |
4180 | bo == NULL && j < ARRAY_SIZE(kgem->vma[type].inactive); |
4180 | bo == NULL && j < ARRAY_SIZE(kgem->vma[type].inactive); |
4181 | j++) { |
4181 | j++) { |
4182 | struct list *head = &kgem->vma[type].inactive[i++%ARRAY_SIZE(kgem->vma[type].inactive)]; |
4182 | struct list *head = &kgem->vma[type].inactive[i++%ARRAY_SIZE(kgem->vma[type].inactive)]; |
4183 | if (!list_is_empty(head)) |
4183 | if (!list_is_empty(head)) |
4184 | bo = list_last_entry(head, struct kgem_bo, vma); |
4184 | bo = list_last_entry(head, struct kgem_bo, vma); |
4185 | } |
4185 | } |
4186 | if (bo == NULL) |
4186 | if (bo == NULL) |
4187 | break; |
4187 | break; |
4188 | 4188 | ||
4189 | DBG(("%s: discarding inactive %s vma cache for %d\n", |
4189 | DBG(("%s: discarding inactive %s vma cache for %d\n", |
4190 | __FUNCTION__, |
4190 | __FUNCTION__, |
4191 | IS_CPU_MAP(bo->map) ? "CPU" : "GTT", bo->handle)); |
4191 | IS_CPU_MAP(bo->map) ? "CPU" : "GTT", bo->handle)); |
4192 | assert(IS_CPU_MAP(bo->map) == type); |
4192 | assert(IS_CPU_MAP(bo->map) == type); |
4193 | assert(bo->map); |
4193 | assert(bo->map); |
4194 | assert(bo->rq == NULL); |
4194 | assert(bo->rq == NULL); |
4195 | 4195 | ||
4196 | VG(if (type) VALGRIND_MAKE_MEM_NOACCESS(MAP(bo->map), bytes(bo))); |
4196 | VG(if (type) VALGRIND_MAKE_MEM_NOACCESS(MAP(bo->map), bytes(bo))); |
4197 | // munmap(MAP(bo->map), bytes(bo)); |
4197 | // munmap(MAP(bo->map), bytes(bo)); |
4198 | bo->map = NULL; |
4198 | bo->map = NULL; |
4199 | list_del(&bo->vma); |
4199 | list_del(&bo->vma); |
4200 | kgem->vma[type].count--; |
4200 | kgem->vma[type].count--; |
4201 | 4201 | ||
4202 | if (!bo->purged && !kgem_bo_set_purgeable(kgem, bo)) { |
4202 | if (!bo->purged && !kgem_bo_set_purgeable(kgem, bo)) { |
4203 | DBG(("%s: freeing unpurgeable old mapping\n", |
4203 | DBG(("%s: freeing unpurgeable old mapping\n", |
4204 | __FUNCTION__)); |
4204 | __FUNCTION__)); |
4205 | kgem_bo_free(kgem, bo); |
4205 | kgem_bo_free(kgem, bo); |
4206 | } |
4206 | } |
4207 | } |
4207 | } |
4208 | } |
4208 | } |
4209 | 4209 | ||
4210 | void *kgem_bo_map__async(struct kgem *kgem, struct kgem_bo *bo) |
4210 | void *kgem_bo_map__async(struct kgem *kgem, struct kgem_bo *bo) |
4211 | { |
4211 | { |
4212 | void *ptr; |
4212 | void *ptr; |
4213 | 4213 | ||
4214 | DBG(("%s: handle=%d, offset=%d, tiling=%d, map=%p, domain=%d\n", __FUNCTION__, |
4214 | DBG(("%s: handle=%d, offset=%d, tiling=%d, map=%p, domain=%d\n", __FUNCTION__, |
4215 | bo->handle, bo->presumed_offset, bo->tiling, bo->map, bo->domain)); |
4215 | bo->handle, bo->presumed_offset, bo->tiling, bo->map, bo->domain)); |
4216 | 4216 | ||
4217 | assert(bo->proxy == NULL); |
4217 | assert(bo->proxy == NULL); |
4218 | assert(list_is_empty(&bo->list)); |
4218 | assert(list_is_empty(&bo->list)); |
4219 | assert(!IS_USER_MAP(bo->map)); |
4219 | assert(!IS_USER_MAP(bo->map)); |
4220 | assert_tiling(kgem, bo); |
4220 | assert_tiling(kgem, bo); |
4221 | 4221 | ||
4222 | if (bo->tiling == I915_TILING_NONE && !bo->scanout && kgem->has_llc) { |
4222 | if (bo->tiling == I915_TILING_NONE && !bo->scanout && kgem->has_llc) { |
4223 | DBG(("%s: converting request for GTT map into CPU map\n", |
4223 | DBG(("%s: converting request for GTT map into CPU map\n", |
4224 | __FUNCTION__)); |
4224 | __FUNCTION__)); |
4225 | return kgem_bo_map__cpu(kgem, bo); |
4225 | return kgem_bo_map__cpu(kgem, bo); |
4226 | } |
4226 | } |
4227 | 4227 | ||
4228 | if (IS_CPU_MAP(bo->map)) |
4228 | if (IS_CPU_MAP(bo->map)) |
4229 | kgem_bo_release_map(kgem, bo); |
4229 | kgem_bo_release_map(kgem, bo); |
4230 | 4230 | ||
4231 | ptr = bo->map; |
4231 | ptr = bo->map; |
4232 | if (ptr == NULL) { |
4232 | if (ptr == NULL) { |
4233 | assert(kgem_bo_size(bo) <= kgem->aperture_mappable / 2); |
4233 | assert(kgem_bo_size(bo) <= kgem->aperture_mappable / 2); |
4234 | 4234 | ||
4235 | kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo)); |
4235 | kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo)); |
4236 | 4236 | ||
4237 | ptr = __kgem_bo_map__gtt(kgem, bo); |
4237 | ptr = __kgem_bo_map__gtt(kgem, bo); |
4238 | if (ptr == NULL) |
4238 | if (ptr == NULL) |
4239 | return NULL; |
4239 | return NULL; |
4240 | 4240 | ||
4241 | /* Cache this mapping to avoid the overhead of an |
4241 | /* Cache this mapping to avoid the overhead of an |
4242 | * excruciatingly slow GTT pagefault. This is more an |
4242 | * excruciatingly slow GTT pagefault. This is more an |
4243 | * issue with compositing managers which need to frequently |
4243 | * issue with compositing managers which need to frequently |
4244 | * flush CPU damage to their GPU bo. |
4244 | * flush CPU damage to their GPU bo. |
4245 | */ |
4245 | */ |
4246 | bo->map = ptr; |
4246 | bo->map = ptr; |
4247 | DBG(("%s: caching GTT vma for %d\n", __FUNCTION__, bo->handle)); |
4247 | DBG(("%s: caching GTT vma for %d\n", __FUNCTION__, bo->handle)); |
4248 | } |
4248 | } |
4249 | 4249 | ||
4250 | return ptr; |
4250 | return ptr; |
4251 | } |
4251 | } |
4252 | 4252 | ||
4253 | void *kgem_bo_map(struct kgem *kgem, struct kgem_bo *bo) |
4253 | void *kgem_bo_map(struct kgem *kgem, struct kgem_bo *bo) |
4254 | { |
4254 | { |
4255 | void *ptr; |
4255 | void *ptr; |
4256 | 4256 | ||
4257 | DBG(("%s: handle=%d, offset=%d, tiling=%d, map=%p, domain=%d\n", __FUNCTION__, |
4257 | DBG(("%s: handle=%d, offset=%d, tiling=%d, map=%p, domain=%d\n", __FUNCTION__, |
4258 | bo->handle, bo->presumed_offset, bo->tiling, bo->map, bo->domain)); |
4258 | bo->handle, bo->presumed_offset, bo->tiling, bo->map, bo->domain)); |
4259 | 4259 | ||
4260 | assert(bo->proxy == NULL); |
4260 | assert(bo->proxy == NULL); |
4261 | assert(list_is_empty(&bo->list)); |
4261 | assert(list_is_empty(&bo->list)); |
4262 | assert(!IS_USER_MAP(bo->map)); |
4262 | assert(!IS_USER_MAP(bo->map)); |
4263 | assert(bo->exec == NULL); |
4263 | assert(bo->exec == NULL); |
4264 | assert_tiling(kgem, bo); |
4264 | assert_tiling(kgem, bo); |
4265 | 4265 | ||
4266 | if (bo->tiling == I915_TILING_NONE && !bo->scanout && |
4266 | if (bo->tiling == I915_TILING_NONE && !bo->scanout && |
4267 | (kgem->has_llc || bo->domain == DOMAIN_CPU)) { |
4267 | (kgem->has_llc || bo->domain == DOMAIN_CPU)) { |
4268 | DBG(("%s: converting request for GTT map into CPU map\n", |
4268 | DBG(("%s: converting request for GTT map into CPU map\n", |
4269 | __FUNCTION__)); |
4269 | __FUNCTION__)); |
4270 | ptr = kgem_bo_map__cpu(kgem, bo); |
4270 | ptr = kgem_bo_map__cpu(kgem, bo); |
4271 | if (ptr) |
4271 | if (ptr) |
4272 | kgem_bo_sync__cpu(kgem, bo); |
4272 | kgem_bo_sync__cpu(kgem, bo); |
4273 | return ptr; |
4273 | return ptr; |
4274 | } |
4274 | } |
4275 | 4275 | ||
4276 | if (IS_CPU_MAP(bo->map)) |
4276 | if (IS_CPU_MAP(bo->map)) |
4277 | kgem_bo_release_map(kgem, bo); |
4277 | kgem_bo_release_map(kgem, bo); |
4278 | 4278 | ||
4279 | ptr = bo->map; |
4279 | ptr = bo->map; |
4280 | if (ptr == NULL) { |
4280 | if (ptr == NULL) { |
4281 | assert(kgem_bo_size(bo) <= kgem->aperture_mappable / 2); |
4281 | assert(kgem_bo_size(bo) <= kgem->aperture_mappable / 2); |
4282 | assert(kgem->gen != 021 || bo->tiling != I915_TILING_Y); |
4282 | assert(kgem->gen != 021 || bo->tiling != I915_TILING_Y); |
4283 | 4283 | ||
4284 | kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo)); |
4284 | kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo)); |
4285 | 4285 | ||
4286 | ptr = __kgem_bo_map__gtt(kgem, bo); |
4286 | ptr = __kgem_bo_map__gtt(kgem, bo); |
4287 | if (ptr == NULL) |
4287 | if (ptr == NULL) |
4288 | return NULL; |
4288 | return NULL; |
4289 | 4289 | ||
4290 | /* Cache this mapping to avoid the overhead of an |
4290 | /* Cache this mapping to avoid the overhead of an |
4291 | * excruciatingly slow GTT pagefault. This is more an |
4291 | * excruciatingly slow GTT pagefault. This is more an |
4292 | * issue with compositing managers which need to frequently |
4292 | * issue with compositing managers which need to frequently |
4293 | * flush CPU damage to their GPU bo. |
4293 | * flush CPU damage to their GPU bo. |
4294 | */ |
4294 | */ |
4295 | bo->map = ptr; |
4295 | bo->map = ptr; |
4296 | DBG(("%s: caching GTT vma for %d\n", __FUNCTION__, bo->handle)); |
4296 | DBG(("%s: caching GTT vma for %d\n", __FUNCTION__, bo->handle)); |
4297 | } |
4297 | } |
4298 | 4298 | ||
4299 | if (bo->domain != DOMAIN_GTT || FORCE_MMAP_SYNC & (1 << DOMAIN_GTT)) { |
4299 | if (bo->domain != DOMAIN_GTT || FORCE_MMAP_SYNC & (1 << DOMAIN_GTT)) { |
4300 | struct drm_i915_gem_set_domain set_domain; |
4300 | struct drm_i915_gem_set_domain set_domain; |
4301 | 4301 | ||
4302 | DBG(("%s: sync: needs_flush? %d, domain? %d, busy? %d\n", __FUNCTION__, |
4302 | DBG(("%s: sync: needs_flush? %d, domain? %d, busy? %d\n", __FUNCTION__, |
4303 | bo->needs_flush, bo->domain, __kgem_busy(kgem, bo->handle))); |
4303 | bo->needs_flush, bo->domain, __kgem_busy(kgem, bo->handle))); |
4304 | 4304 | ||
4305 | /* XXX use PROT_READ to avoid the write flush? */ |
4305 | /* XXX use PROT_READ to avoid the write flush? */ |
4306 | 4306 | ||
4307 | VG_CLEAR(set_domain); |
4307 | VG_CLEAR(set_domain); |
4308 | set_domain.handle = bo->handle; |
4308 | set_domain.handle = bo->handle; |
4309 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
4309 | set_domain.read_domains = I915_GEM_DOMAIN_GTT; |
4310 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
4310 | set_domain.write_domain = I915_GEM_DOMAIN_GTT; |
4311 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain) == 0) { |
4311 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain) == 0) { |
4312 | kgem_bo_retire(kgem, bo); |
4312 | kgem_bo_retire(kgem, bo); |
4313 | bo->domain = DOMAIN_GTT; |
4313 | bo->domain = DOMAIN_GTT; |
4314 | bo->gtt_dirty = true; |
4314 | bo->gtt_dirty = true; |
4315 | } |
4315 | } |
4316 | } |
4316 | } |
4317 | 4317 | ||
4318 | return ptr; |
4318 | return ptr; |
4319 | } |
4319 | } |
4320 | 4320 | ||
4321 | void *kgem_bo_map__gtt(struct kgem *kgem, struct kgem_bo *bo) |
4321 | void *kgem_bo_map__gtt(struct kgem *kgem, struct kgem_bo *bo) |
4322 | { |
4322 | { |
4323 | void *ptr; |
4323 | void *ptr; |
4324 | 4324 | ||
4325 | DBG(("%s: handle=%d, offset=%d, tiling=%d, map=%p, domain=%d\n", __FUNCTION__, |
4325 | DBG(("%s: handle=%d, offset=%d, tiling=%d, map=%p, domain=%d\n", __FUNCTION__, |
4326 | bo->handle, bo->presumed_offset, bo->tiling, bo->map, bo->domain)); |
4326 | bo->handle, bo->presumed_offset, bo->tiling, bo->map, bo->domain)); |
4327 | 4327 | ||
4328 | assert(bo->exec == NULL); |
4328 | assert(bo->exec == NULL); |
4329 | assert(list_is_empty(&bo->list)); |
4329 | assert(list_is_empty(&bo->list)); |
4330 | assert(!IS_USER_MAP(bo->map)); |
4330 | assert(!IS_USER_MAP(bo->map)); |
4331 | assert_tiling(kgem, bo); |
4331 | assert_tiling(kgem, bo); |
4332 | 4332 | ||
4333 | if (IS_CPU_MAP(bo->map)) |
4333 | if (IS_CPU_MAP(bo->map)) |
4334 | kgem_bo_release_map(kgem, bo); |
4334 | kgem_bo_release_map(kgem, bo); |
4335 | 4335 | ||
4336 | ptr = bo->map; |
4336 | ptr = bo->map; |
4337 | if (ptr == NULL) { |
4337 | if (ptr == NULL) { |
4338 | assert(bytes(bo) <= kgem->aperture_mappable / 4); |
4338 | assert(bytes(bo) <= kgem->aperture_mappable / 4); |
4339 | 4339 | ||
4340 | kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo)); |
4340 | kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo)); |
4341 | 4341 | ||
4342 | ptr = __kgem_bo_map__gtt(kgem, bo); |
4342 | ptr = __kgem_bo_map__gtt(kgem, bo); |
4343 | if (ptr == NULL) |
4343 | if (ptr == NULL) |
4344 | return NULL; |
4344 | return NULL; |
4345 | 4345 | ||
4346 | /* Cache this mapping to avoid the overhead of an |
4346 | /* Cache this mapping to avoid the overhead of an |
4347 | * excruciatingly slow GTT pagefault. This is more an |
4347 | * excruciatingly slow GTT pagefault. This is more an |
4348 | * issue with compositing managers which need to frequently |
4348 | * issue with compositing managers which need to frequently |
4349 | * flush CPU damage to their GPU bo. |
4349 | * flush CPU damage to their GPU bo. |
4350 | */ |
4350 | */ |
4351 | bo->map = ptr; |
4351 | bo->map = ptr; |
4352 | DBG(("%s: caching GTT vma for %d\n", __FUNCTION__, bo->handle)); |
4352 | DBG(("%s: caching GTT vma for %d\n", __FUNCTION__, bo->handle)); |
4353 | } |
4353 | } |
4354 | 4354 | ||
4355 | return ptr; |
4355 | return ptr; |
4356 | } |
4356 | } |
4357 | 4357 | ||
4358 | void *kgem_bo_map__debug(struct kgem *kgem, struct kgem_bo *bo) |
4358 | void *kgem_bo_map__debug(struct kgem *kgem, struct kgem_bo *bo) |
4359 | { |
4359 | { |
4360 | if (bo->map) |
4360 | if (bo->map) |
4361 | return MAP(bo->map); |
4361 | return MAP(bo->map); |
4362 | 4362 | ||
4363 | kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo)); |
4363 | kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo)); |
4364 | return bo->map = __kgem_bo_map__gtt(kgem, bo); |
4364 | return bo->map = __kgem_bo_map__gtt(kgem, bo); |
4365 | } |
4365 | } |
4366 | 4366 | ||
4367 | void *kgem_bo_map__cpu(struct kgem *kgem, struct kgem_bo *bo) |
4367 | void *kgem_bo_map__cpu(struct kgem *kgem, struct kgem_bo *bo) |
4368 | { |
4368 | { |
4369 | struct drm_i915_gem_mmap mmap_arg; |
4369 | struct drm_i915_gem_mmap mmap_arg; |
4370 | 4370 | ||
4371 | DBG(("%s(handle=%d, size=%d, mapped? %d)\n", |
4371 | DBG(("%s(handle=%d, size=%d, mapped? %d)\n", |
4372 | __FUNCTION__, bo->handle, bytes(bo), (int)__MAP_TYPE(bo->map))); |
4372 | __FUNCTION__, bo->handle, bytes(bo), (int)__MAP_TYPE(bo->map))); |
4373 | assert(!bo->purged); |
4373 | assert(!bo->purged); |
4374 | assert(list_is_empty(&bo->list)); |
4374 | assert(list_is_empty(&bo->list)); |
4375 | assert(bo->proxy == NULL); |
4375 | assert(bo->proxy == NULL); |
4376 | 4376 | ||
4377 | if (IS_CPU_MAP(bo->map)) |
4377 | if (IS_CPU_MAP(bo->map)) |
4378 | return MAP(bo->map); |
4378 | return MAP(bo->map); |
4379 | 4379 | ||
4380 | if (bo->map) |
4380 | if (bo->map) |
4381 | kgem_bo_release_map(kgem, bo); |
4381 | kgem_bo_release_map(kgem, bo); |
4382 | 4382 | ||
4383 | kgem_trim_vma_cache(kgem, MAP_CPU, bucket(bo)); |
4383 | kgem_trim_vma_cache(kgem, MAP_CPU, bucket(bo)); |
4384 | 4384 | ||
4385 | retry: |
4385 | retry: |
4386 | VG_CLEAR(mmap_arg); |
4386 | VG_CLEAR(mmap_arg); |
4387 | mmap_arg.handle = bo->handle; |
4387 | mmap_arg.handle = bo->handle; |
4388 | mmap_arg.offset = 0; |
4388 | mmap_arg.offset = 0; |
4389 | mmap_arg.size = bytes(bo); |
4389 | mmap_arg.size = bytes(bo); |
4390 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg)) { |
4390 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg)) { |
4391 | 4391 | ||
4392 | if (__kgem_throttle_retire(kgem, 0)) |
4392 | if (__kgem_throttle_retire(kgem, 0)) |
4393 | goto retry; |
4393 | goto retry; |
4394 | 4394 | ||
4395 | if (kgem->need_expire) { |
4395 | if (kgem->need_expire) { |
4396 | kgem_cleanup_cache(kgem); |
4396 | kgem_cleanup_cache(kgem); |
4397 | goto retry; |
4397 | goto retry; |
4398 | } |
4398 | } |
4399 | 4399 | ||
4400 | ErrorF("%s: failed to mmap handle=%d, %d bytes, into CPU domain\n", |
4400 | ErrorF("%s: failed to mmap handle=%d, %d bytes, into CPU domain\n", |
4401 | __FUNCTION__, bo->handle, bytes(bo)); |
4401 | __FUNCTION__, bo->handle, bytes(bo)); |
4402 | return NULL; |
4402 | return NULL; |
4403 | } |
4403 | } |
4404 | 4404 | ||
4405 | VG(VALGRIND_MAKE_MEM_DEFINED(mmap_arg.addr_ptr, bytes(bo))); |
4405 | VG(VALGRIND_MAKE_MEM_DEFINED(mmap_arg.addr_ptr, bytes(bo))); |
4406 | 4406 | ||
4407 | DBG(("%s: caching CPU vma for %d\n", __FUNCTION__, bo->handle)); |
4407 | DBG(("%s: caching CPU vma for %d\n", __FUNCTION__, bo->handle)); |
4408 | bo->map = MAKE_CPU_MAP(mmap_arg.addr_ptr); |
4408 | bo->map = MAKE_CPU_MAP(mmap_arg.addr_ptr); |
4409 | return (void *)(uintptr_t)mmap_arg.addr_ptr; |
4409 | return (void *)(uintptr_t)mmap_arg.addr_ptr; |
4410 | } |
4410 | } |
4411 | 4411 | ||
4412 | void *__kgem_bo_map__cpu(struct kgem *kgem, struct kgem_bo *bo) |
4412 | void *__kgem_bo_map__cpu(struct kgem *kgem, struct kgem_bo *bo) |
4413 | { |
4413 | { |
4414 | struct drm_i915_gem_mmap mmap_arg; |
4414 | struct drm_i915_gem_mmap mmap_arg; |
4415 | 4415 | ||
4416 | DBG(("%s(handle=%d, size=%d, mapped? %d)\n", |
4416 | DBG(("%s(handle=%d, size=%d, mapped? %d)\n", |
4417 | __FUNCTION__, bo->handle, bytes(bo), (int)__MAP_TYPE(bo->map))); |
4417 | __FUNCTION__, bo->handle, bytes(bo), (int)__MAP_TYPE(bo->map))); |
4418 | assert(bo->refcnt); |
4418 | assert(bo->refcnt); |
4419 | assert(!bo->purged); |
4419 | assert(!bo->purged); |
4420 | assert(list_is_empty(&bo->list)); |
4420 | assert(list_is_empty(&bo->list)); |
4421 | assert(bo->proxy == NULL); |
4421 | assert(bo->proxy == NULL); |
4422 | 4422 | ||
4423 | if (IS_CPU_MAP(bo->map)) |
4423 | if (IS_CPU_MAP(bo->map)) |
4424 | return MAP(bo->map); |
4424 | return MAP(bo->map); |
4425 | 4425 | ||
4426 | retry: |
4426 | retry: |
4427 | VG_CLEAR(mmap_arg); |
4427 | VG_CLEAR(mmap_arg); |
4428 | mmap_arg.handle = bo->handle; |
4428 | mmap_arg.handle = bo->handle; |
4429 | mmap_arg.offset = 0; |
4429 | mmap_arg.offset = 0; |
4430 | mmap_arg.size = bytes(bo); |
4430 | mmap_arg.size = bytes(bo); |
4431 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg)) { |
4431 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg)) { |
4432 | int err = errno; |
4432 | int err = errno; |
4433 | 4433 | ||
4434 | assert(err != EINVAL); |
4434 | assert(err != EINVAL); |
4435 | 4435 | ||
4436 | if (__kgem_throttle_retire(kgem, 0)) |
4436 | if (__kgem_throttle_retire(kgem, 0)) |
4437 | goto retry; |
4437 | goto retry; |
4438 | 4438 | ||
4439 | if (kgem->need_expire) { |
4439 | if (kgem->need_expire) { |
4440 | kgem_cleanup_cache(kgem); |
4440 | kgem_cleanup_cache(kgem); |
4441 | goto retry; |
4441 | goto retry; |
4442 | } |
4442 | } |
4443 | 4443 | ||
4444 | ErrorF("%s: failed to mmap handle=%d, %d bytes, into CPU domain: %d\n", |
4444 | ErrorF("%s: failed to mmap handle=%d, %d bytes, into CPU domain: %d\n", |
4445 | __FUNCTION__, bo->handle, bytes(bo), err); |
4445 | __FUNCTION__, bo->handle, bytes(bo), err); |
4446 | return NULL; |
4446 | return NULL; |
4447 | } |
4447 | } |
4448 | 4448 | ||
4449 | VG(VALGRIND_MAKE_MEM_DEFINED(mmap_arg.addr_ptr, bytes(bo))); |
4449 | VG(VALGRIND_MAKE_MEM_DEFINED(mmap_arg.addr_ptr, bytes(bo))); |
4450 | if (bo->map && bo->domain == DOMAIN_CPU) { |
4450 | if (bo->map && bo->domain == DOMAIN_CPU) { |
4451 | DBG(("%s: discarding GTT vma for %d\n", __FUNCTION__, bo->handle)); |
4451 | DBG(("%s: discarding GTT vma for %d\n", __FUNCTION__, bo->handle)); |
4452 | kgem_bo_release_map(kgem, bo); |
4452 | kgem_bo_release_map(kgem, bo); |
4453 | } |
4453 | } |
4454 | if (bo->map == NULL) { |
4454 | if (bo->map == NULL) { |
4455 | DBG(("%s: caching CPU vma for %d\n", __FUNCTION__, bo->handle)); |
4455 | DBG(("%s: caching CPU vma for %d\n", __FUNCTION__, bo->handle)); |
4456 | bo->map = MAKE_CPU_MAP(mmap_arg.addr_ptr); |
4456 | bo->map = MAKE_CPU_MAP(mmap_arg.addr_ptr); |
4457 | } |
4457 | } |
4458 | return (void *)(uintptr_t)mmap_arg.addr_ptr; |
4458 | return (void *)(uintptr_t)mmap_arg.addr_ptr; |
4459 | } |
4459 | } |
4460 | void kgem_bo_sync__cpu(struct kgem *kgem, struct kgem_bo *bo) |
4460 | void kgem_bo_sync__cpu(struct kgem *kgem, struct kgem_bo *bo) |
4461 | { |
4461 | { |
4462 | DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle)); |
4462 | DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle)); |
4463 | assert(!bo->scanout); |
4463 | assert(!bo->scanout); |
4464 | kgem_bo_submit(kgem, bo); |
4464 | kgem_bo_submit(kgem, bo); |
4465 | 4465 | ||
4466 | /* SHM pixmaps use proxies for subpage offsets */ |
4466 | /* SHM pixmaps use proxies for subpage offsets */ |
4467 | assert(!bo->purged); |
4467 | assert(!bo->purged); |
4468 | while (bo->proxy) |
4468 | while (bo->proxy) |
4469 | bo = bo->proxy; |
4469 | bo = bo->proxy; |
4470 | assert(!bo->purged); |
4470 | assert(!bo->purged); |
4471 | 4471 | ||
4472 | if (bo->domain != DOMAIN_CPU || FORCE_MMAP_SYNC & (1 << DOMAIN_CPU)) { |
4472 | if (bo->domain != DOMAIN_CPU || FORCE_MMAP_SYNC & (1 << DOMAIN_CPU)) { |
4473 | struct drm_i915_gem_set_domain set_domain; |
4473 | struct drm_i915_gem_set_domain set_domain; |
4474 | 4474 | ||
4475 | DBG(("%s: SYNC: handle=%d, needs_flush? %d, domain? %d, busy? %d\n", |
4475 | DBG(("%s: SYNC: handle=%d, needs_flush? %d, domain? %d, busy? %d\n", |
4476 | __FUNCTION__, bo->handle, |
4476 | __FUNCTION__, bo->handle, |
4477 | bo->needs_flush, bo->domain, |
4477 | bo->needs_flush, bo->domain, |
4478 | __kgem_busy(kgem, bo->handle))); |
4478 | __kgem_busy(kgem, bo->handle))); |
4479 | 4479 | ||
4480 | VG_CLEAR(set_domain); |
4480 | VG_CLEAR(set_domain); |
4481 | set_domain.handle = bo->handle; |
4481 | set_domain.handle = bo->handle; |
4482 | set_domain.read_domains = I915_GEM_DOMAIN_CPU; |
4482 | set_domain.read_domains = I915_GEM_DOMAIN_CPU; |
4483 | set_domain.write_domain = I915_GEM_DOMAIN_CPU; |
4483 | set_domain.write_domain = I915_GEM_DOMAIN_CPU; |
4484 | 4484 | ||
4485 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain) == 0) { |
4485 | if (drmIoctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain) == 0) { |
4486 | kgem_bo_retire(kgem, bo); |
4486 | kgem_bo_retire(kgem, bo); |
4487 | bo->domain = DOMAIN_CPU; |
4487 | bo->domain = DOMAIN_CPU; |
4488 | } |
4488 | } |
4489 | } |
4489 | } |
4490 | } |
4490 | } |
4491 | 4491 | ||
4492 | void kgem_clear_dirty(struct kgem *kgem) |
4492 | void kgem_clear_dirty(struct kgem *kgem) |
4493 | { |
4493 | { |
4494 | struct list * const buffers = &kgem->next_request->buffers; |
4494 | struct list * const buffers = &kgem->next_request->buffers; |
4495 | struct kgem_bo *bo; |
4495 | struct kgem_bo *bo; |
4496 | 4496 | ||
4497 | list_for_each_entry(bo, buffers, request) { |
4497 | list_for_each_entry(bo, buffers, request) { |
4498 | if (!bo->gpu_dirty) |
4498 | if (!bo->gpu_dirty) |
4499 | break; |
4499 | break; |
4500 | 4500 | ||
4501 | bo->gpu_dirty = false; |
4501 | bo->gpu_dirty = false; |
4502 | } |
4502 | } |
4503 | } |
4503 | } |
4504 | 4504 | ||
4505 | struct kgem_bo *kgem_create_proxy(struct kgem *kgem, |
4505 | struct kgem_bo *kgem_create_proxy(struct kgem *kgem, |
4506 | struct kgem_bo *target, |
4506 | struct kgem_bo *target, |
4507 | int offset, int length) |
4507 | int offset, int length) |
4508 | { |
4508 | { |
4509 | struct kgem_bo *bo; |
4509 | struct kgem_bo *bo; |
4510 | 4510 | ||
4511 | DBG(("%s: target handle=%d [proxy? %d], offset=%d, length=%d, io=%d\n", |
4511 | DBG(("%s: target handle=%d [proxy? %d], offset=%d, length=%d, io=%d\n", |
4512 | __FUNCTION__, target->handle, target->proxy ? target->proxy->delta : -1, |
4512 | __FUNCTION__, target->handle, target->proxy ? target->proxy->delta : -1, |
4513 | offset, length, target->io)); |
4513 | offset, length, target->io)); |
4514 | 4514 | ||
4515 | bo = __kgem_bo_alloc(target->handle, length); |
4515 | bo = __kgem_bo_alloc(target->handle, length); |
4516 | if (bo == NULL) |
4516 | if (bo == NULL) |
4517 | return NULL; |
4517 | return NULL; |
4518 | 4518 | ||
4519 | bo->unique_id = kgem_get_unique_id(kgem); |
4519 | bo->unique_id = kgem_get_unique_id(kgem); |
4520 | bo->reusable = false; |
4520 | bo->reusable = false; |
4521 | bo->size.bytes = length; |
4521 | bo->size.bytes = length; |
4522 | 4522 | ||
4523 | bo->io = target->io && target->proxy == NULL; |
4523 | bo->io = target->io && target->proxy == NULL; |
4524 | bo->gpu_dirty = target->gpu_dirty; |
4524 | bo->gpu_dirty = target->gpu_dirty; |
4525 | bo->tiling = target->tiling; |
4525 | bo->tiling = target->tiling; |
4526 | bo->pitch = target->pitch; |
4526 | bo->pitch = target->pitch; |
4527 | bo->flush = target->flush; |
4527 | bo->flush = target->flush; |
4528 | bo->snoop = target->snoop; |
4528 | bo->snoop = target->snoop; |
4529 | 4529 | ||
4530 | assert(!bo->scanout); |
4530 | assert(!bo->scanout); |
4531 | bo->proxy = kgem_bo_reference(target); |
4531 | bo->proxy = kgem_bo_reference(target); |
4532 | bo->delta = offset; |
4532 | bo->delta = offset; |
4533 | 4533 | ||
4534 | if (target->exec) { |
4534 | if (target->exec) { |
4535 | list_move_tail(&bo->request, &kgem->next_request->buffers); |
4535 | list_move_tail(&bo->request, &kgem->next_request->buffers); |
4536 | bo->exec = &_kgem_dummy_exec; |
4536 | bo->exec = &_kgem_dummy_exec; |
4537 | } |
4537 | } |
4538 | bo->rq = target->rq; |
4538 | bo->rq = target->rq; |
4539 | 4539 | ||
4540 | return bo; |
4540 | return bo; |
4541 | } |
4541 | } |
4542 | 4542 | ||
4543 | #if 0 |
4543 | #if 0 |
4544 | static struct kgem_buffer * |
4544 | static struct kgem_buffer * |
4545 | buffer_alloc(void) |
4545 | buffer_alloc(void) |
4546 | { |
4546 | { |
4547 | struct kgem_buffer *bo; |
4547 | struct kgem_buffer *bo; |
4548 | 4548 | ||
4549 | bo = malloc(sizeof(*bo)); |
4549 | bo = malloc(sizeof(*bo)); |
4550 | if (bo == NULL) |
4550 | if (bo == NULL) |
4551 | return NULL; |
4551 | return NULL; |
4552 | 4552 | ||
4553 | bo->mem = NULL; |
4553 | bo->mem = NULL; |
4554 | bo->need_io = false; |
4554 | bo->need_io = false; |
4555 | bo->mmapped = true; |
4555 | bo->mmapped = true; |
4556 | 4556 | ||
4557 | return bo; |
4557 | return bo; |
4558 | } |
4558 | } |
4559 | 4559 | ||
4560 | static struct kgem_buffer * |
4560 | static struct kgem_buffer * |
4561 | buffer_alloc_with_data(int num_pages) |
4561 | buffer_alloc_with_data(int num_pages) |
4562 | { |
4562 | { |
4563 | struct kgem_buffer *bo; |
4563 | struct kgem_buffer *bo; |
4564 | 4564 | ||
4565 | bo = malloc(sizeof(*bo) + 2*UPLOAD_ALIGNMENT + num_pages * PAGE_SIZE); |
4565 | bo = malloc(sizeof(*bo) + 2*UPLOAD_ALIGNMENT + num_pages * PAGE_SIZE); |
4566 | if (bo == NULL) |
4566 | if (bo == NULL) |
4567 | return NULL; |
4567 | return NULL; |
4568 | 4568 | ||
4569 | bo->mem = (void *)ALIGN((uintptr_t)bo + sizeof(*bo), UPLOAD_ALIGNMENT); |
4569 | bo->mem = (void *)ALIGN((uintptr_t)bo + sizeof(*bo), UPLOAD_ALIGNMENT); |
4570 | bo->mmapped = false; |
4570 | bo->mmapped = false; |
4571 | return bo; |
4571 | return bo; |
4572 | } |
4572 | } |
4573 | 4573 | ||
4574 | static inline bool |
4574 | static inline bool |
4575 | use_snoopable_buffer(struct kgem *kgem, uint32_t flags) |
4575 | use_snoopable_buffer(struct kgem *kgem, uint32_t flags) |
4576 | { |
4576 | { |
4577 | if ((flags & KGEM_BUFFER_WRITE) == 0) |
4577 | if ((flags & KGEM_BUFFER_WRITE) == 0) |
4578 | return kgem->gen >= 030; |
4578 | return kgem->gen >= 030; |
4579 | 4579 | ||
4580 | return true; |
4580 | return true; |
4581 | } |
4581 | } |
4582 | 4582 | ||
4583 | static void |
4583 | static void |
4584 | init_buffer_from_bo(struct kgem_buffer *bo, struct kgem_bo *old) |
4584 | init_buffer_from_bo(struct kgem_buffer *bo, struct kgem_bo *old) |
4585 | { |
4585 | { |
4586 | DBG(("%s: reusing handle=%d for buffer\n", |
4586 | DBG(("%s: reusing handle=%d for buffer\n", |
4587 | __FUNCTION__, old->handle)); |
4587 | __FUNCTION__, old->handle)); |
4588 | 4588 | ||
4589 | assert(old->proxy == NULL); |
4589 | assert(old->proxy == NULL); |
4590 | 4590 | ||
4591 | memcpy(&bo->base, old, sizeof(*old)); |
4591 | memcpy(&bo->base, old, sizeof(*old)); |
4592 | if (old->rq) |
4592 | if (old->rq) |
4593 | list_replace(&old->request, &bo->base.request); |
4593 | list_replace(&old->request, &bo->base.request); |
4594 | else |
4594 | else |
4595 | list_init(&bo->base.request); |
4595 | list_init(&bo->base.request); |
4596 | list_replace(&old->vma, &bo->base.vma); |
4596 | list_replace(&old->vma, &bo->base.vma); |
4597 | list_init(&bo->base.list); |
4597 | list_init(&bo->base.list); |
4598 | free(old); |
4598 | free(old); |
4599 | 4599 | ||
4600 | assert(bo->base.tiling == I915_TILING_NONE); |
4600 | assert(bo->base.tiling == I915_TILING_NONE); |
4601 | 4601 | ||
4602 | bo->base.refcnt = 1; |
4602 | bo->base.refcnt = 1; |
4603 | } |
4603 | } |
4604 | 4604 | ||
4605 | static struct kgem_buffer * |
4605 | static struct kgem_buffer * |
4606 | search_snoopable_buffer(struct kgem *kgem, unsigned alloc) |
4606 | search_snoopable_buffer(struct kgem *kgem, unsigned alloc) |
4607 | { |
4607 | { |
4608 | struct kgem_buffer *bo; |
4608 | struct kgem_buffer *bo; |
4609 | struct kgem_bo *old; |
4609 | struct kgem_bo *old; |
4610 | 4610 | ||
4611 | old = search_snoop_cache(kgem, alloc, 0); |
4611 | old = search_snoop_cache(kgem, alloc, 0); |
4612 | if (old) { |
4612 | if (old) { |
4613 | if (!old->io) { |
4613 | if (!old->io) { |
4614 | bo = buffer_alloc(); |
4614 | bo = buffer_alloc(); |
4615 | if (bo == NULL) |
4615 | if (bo == NULL) |
4616 | return NULL; |
4616 | return NULL; |
4617 | 4617 | ||
4618 | init_buffer_from_bo(bo, old); |
4618 | init_buffer_from_bo(bo, old); |
4619 | } else { |
4619 | } else { |
4620 | bo = (struct kgem_buffer *)old; |
4620 | bo = (struct kgem_buffer *)old; |
4621 | bo->base.refcnt = 1; |
4621 | bo->base.refcnt = 1; |
4622 | } |
4622 | } |
4623 | 4623 | ||
4624 | DBG(("%s: created CPU handle=%d for buffer, size %d\n", |
4624 | DBG(("%s: created CPU handle=%d for buffer, size %d\n", |
4625 | __FUNCTION__, bo->base.handle, num_pages(&bo->base))); |
4625 | __FUNCTION__, bo->base.handle, num_pages(&bo->base))); |
4626 | 4626 | ||
4627 | assert(bo->base.snoop); |
4627 | assert(bo->base.snoop); |
4628 | assert(bo->base.tiling == I915_TILING_NONE); |
4628 | assert(bo->base.tiling == I915_TILING_NONE); |
4629 | assert(num_pages(&bo->base) >= alloc); |
4629 | assert(num_pages(&bo->base) >= alloc); |
4630 | assert(bo->mmapped == true); |
4630 | assert(bo->mmapped == true); |
4631 | assert(bo->need_io == false); |
4631 | assert(bo->need_io == false); |
4632 | 4632 | ||
4633 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
4633 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
4634 | if (bo->mem == NULL) { |
4634 | if (bo->mem == NULL) { |
4635 | bo->base.refcnt = 0; |
4635 | bo->base.refcnt = 0; |
4636 | kgem_bo_free(kgem, &bo->base); |
4636 | kgem_bo_free(kgem, &bo->base); |
4637 | bo = NULL; |
4637 | bo = NULL; |
4638 | } |
4638 | } |
4639 | 4639 | ||
4640 | return bo; |
4640 | return bo; |
4641 | } |
4641 | } |
4642 | 4642 | ||
4643 | return NULL; |
4643 | return NULL; |
4644 | } |
4644 | } |
4645 | 4645 | ||
4646 | static struct kgem_buffer * |
4646 | static struct kgem_buffer * |
4647 | create_snoopable_buffer(struct kgem *kgem, unsigned alloc) |
4647 | create_snoopable_buffer(struct kgem *kgem, unsigned alloc) |
4648 | { |
4648 | { |
4649 | struct kgem_buffer *bo; |
4649 | struct kgem_buffer *bo; |
4650 | uint32_t handle; |
4650 | uint32_t handle; |
4651 | 4651 | ||
4652 | if (kgem->has_llc) { |
4652 | if (kgem->has_llc) { |
4653 | struct kgem_bo *old; |
4653 | struct kgem_bo *old; |
4654 | 4654 | ||
4655 | bo = buffer_alloc(); |
4655 | bo = buffer_alloc(); |
4656 | if (bo == NULL) |
4656 | if (bo == NULL) |
4657 | return NULL; |
4657 | return NULL; |
4658 | 4658 | ||
4659 | old = search_linear_cache(kgem, alloc, |
4659 | old = search_linear_cache(kgem, alloc, |
4660 | CREATE_INACTIVE | CREATE_CPU_MAP | CREATE_EXACT); |
4660 | CREATE_INACTIVE | CREATE_CPU_MAP | CREATE_EXACT); |
4661 | if (old) { |
4661 | if (old) { |
4662 | init_buffer_from_bo(bo, old); |
4662 | init_buffer_from_bo(bo, old); |
4663 | } else { |
4663 | } else { |
4664 | handle = gem_create(kgem->fd, alloc); |
4664 | handle = gem_create(kgem->fd, alloc); |
4665 | if (handle == 0) { |
4665 | if (handle == 0) { |
4666 | free(bo); |
4666 | free(bo); |
4667 | return NULL; |
4667 | return NULL; |
4668 | } |
4668 | } |
4669 | 4669 | ||
4670 | debug_alloc(kgem, alloc); |
4670 | debug_alloc(kgem, alloc); |
4671 | __kgem_bo_init(&bo->base, handle, alloc); |
4671 | __kgem_bo_init(&bo->base, handle, alloc); |
4672 | DBG(("%s: created CPU (LLC) handle=%d for buffer, size %d\n", |
4672 | DBG(("%s: created CPU (LLC) handle=%d for buffer, size %d\n", |
4673 | __FUNCTION__, bo->base.handle, alloc)); |
4673 | __FUNCTION__, bo->base.handle, alloc)); |
4674 | } |
4674 | } |
4675 | 4675 | ||
4676 | assert(bo->base.refcnt == 1); |
4676 | assert(bo->base.refcnt == 1); |
4677 | assert(bo->mmapped == true); |
4677 | assert(bo->mmapped == true); |
4678 | assert(bo->need_io == false); |
4678 | assert(bo->need_io == false); |
4679 | 4679 | ||
4680 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
4680 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
4681 | if (bo->mem != NULL) |
4681 | if (bo->mem != NULL) |
4682 | return bo; |
4682 | return bo; |
4683 | 4683 | ||
4684 | bo->base.refcnt = 0; /* for valgrind */ |
4684 | bo->base.refcnt = 0; /* for valgrind */ |
4685 | kgem_bo_free(kgem, &bo->base); |
4685 | kgem_bo_free(kgem, &bo->base); |
4686 | } |
4686 | } |
4687 | 4687 | ||
4688 | if (kgem->has_caching) { |
4688 | if (kgem->has_caching) { |
4689 | struct kgem_bo *old; |
4689 | struct kgem_bo *old; |
4690 | 4690 | ||
4691 | bo = buffer_alloc(); |
4691 | bo = buffer_alloc(); |
4692 | if (bo == NULL) |
4692 | if (bo == NULL) |
4693 | return NULL; |
4693 | return NULL; |
4694 | 4694 | ||
4695 | old = search_linear_cache(kgem, alloc, |
4695 | old = search_linear_cache(kgem, alloc, |
4696 | CREATE_INACTIVE | CREATE_CPU_MAP | CREATE_EXACT); |
4696 | CREATE_INACTIVE | CREATE_CPU_MAP | CREATE_EXACT); |
4697 | if (old) { |
4697 | if (old) { |
4698 | init_buffer_from_bo(bo, old); |
4698 | init_buffer_from_bo(bo, old); |
4699 | } else { |
4699 | } else { |
4700 | handle = gem_create(kgem->fd, alloc); |
4700 | handle = gem_create(kgem->fd, alloc); |
4701 | if (handle == 0) { |
4701 | if (handle == 0) { |
4702 | free(bo); |
4702 | free(bo); |
4703 | return NULL; |
4703 | return NULL; |
4704 | } |
4704 | } |
4705 | 4705 | ||
4706 | debug_alloc(kgem, alloc); |
4706 | debug_alloc(kgem, alloc); |
4707 | __kgem_bo_init(&bo->base, handle, alloc); |
4707 | __kgem_bo_init(&bo->base, handle, alloc); |
4708 | DBG(("%s: created CPU handle=%d for buffer, size %d\n", |
4708 | DBG(("%s: created CPU handle=%d for buffer, size %d\n", |
4709 | __FUNCTION__, bo->base.handle, alloc)); |
4709 | __FUNCTION__, bo->base.handle, alloc)); |
4710 | } |
4710 | } |
4711 | 4711 | ||
4712 | assert(bo->base.refcnt == 1); |
4712 | assert(bo->base.refcnt == 1); |
4713 | assert(bo->mmapped == true); |
4713 | assert(bo->mmapped == true); |
4714 | assert(bo->need_io == false); |
4714 | assert(bo->need_io == false); |
4715 | 4715 | ||
4716 | if (!gem_set_caching(kgem->fd, bo->base.handle, SNOOPED)) |
4716 | if (!gem_set_caching(kgem->fd, bo->base.handle, SNOOPED)) |
4717 | goto free_caching; |
4717 | goto free_caching; |
4718 | 4718 | ||
4719 | bo->base.snoop = true; |
4719 | bo->base.snoop = true; |
4720 | 4720 | ||
4721 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
4721 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
4722 | if (bo->mem == NULL) |
4722 | if (bo->mem == NULL) |
4723 | goto free_caching; |
4723 | goto free_caching; |
4724 | 4724 | ||
4725 | return bo; |
4725 | return bo; |
4726 | 4726 | ||
4727 | free_caching: |
4727 | free_caching: |
4728 | bo->base.refcnt = 0; /* for valgrind */ |
4728 | bo->base.refcnt = 0; /* for valgrind */ |
4729 | kgem_bo_free(kgem, &bo->base); |
4729 | kgem_bo_free(kgem, &bo->base); |
4730 | } |
4730 | } |
4731 | 4731 | ||
4732 | if (kgem->has_userptr) { |
4732 | if (kgem->has_userptr) { |
4733 | bo = buffer_alloc(); |
4733 | bo = buffer_alloc(); |
4734 | if (bo == NULL) |
4734 | if (bo == NULL) |
4735 | return NULL; |
4735 | return NULL; |
4736 | 4736 | ||
4737 | //if (posix_memalign(&ptr, 64, ALIGN(size, 64))) |
4737 | //if (posix_memalign(&ptr, 64, ALIGN(size, 64))) |
4738 | if (posix_memalign(&bo->mem, PAGE_SIZE, alloc * PAGE_SIZE)) { |
4738 | if (posix_memalign(&bo->mem, PAGE_SIZE, alloc * PAGE_SIZE)) { |
4739 | free(bo); |
4739 | free(bo); |
4740 | return NULL; |
4740 | return NULL; |
4741 | } |
4741 | } |
4742 | 4742 | ||
4743 | handle = gem_userptr(kgem->fd, bo->mem, alloc * PAGE_SIZE, false); |
4743 | handle = gem_userptr(kgem->fd, bo->mem, alloc * PAGE_SIZE, false); |
4744 | if (handle == 0) { |
4744 | if (handle == 0) { |
4745 | free(bo->mem); |
4745 | free(bo->mem); |
4746 | free(bo); |
4746 | free(bo); |
4747 | return NULL; |
4747 | return NULL; |
4748 | } |
4748 | } |
4749 | 4749 | ||
4750 | debug_alloc(kgem, alloc); |
4750 | debug_alloc(kgem, alloc); |
4751 | __kgem_bo_init(&bo->base, handle, alloc); |
4751 | __kgem_bo_init(&bo->base, handle, alloc); |
4752 | DBG(("%s: created snoop handle=%d for buffer\n", |
4752 | DBG(("%s: created snoop handle=%d for buffer\n", |
4753 | __FUNCTION__, bo->base.handle)); |
4753 | __FUNCTION__, bo->base.handle)); |
4754 | 4754 | ||
4755 | assert(bo->mmapped == true); |
4755 | assert(bo->mmapped == true); |
4756 | assert(bo->need_io == false); |
4756 | assert(bo->need_io == false); |
4757 | 4757 | ||
4758 | bo->base.refcnt = 1; |
4758 | bo->base.refcnt = 1; |
4759 | bo->base.snoop = true; |
4759 | bo->base.snoop = true; |
4760 | bo->base.map = MAKE_USER_MAP(bo->mem); |
4760 | bo->base.map = MAKE_USER_MAP(bo->mem); |
4761 | 4761 | ||
4762 | return bo; |
4762 | return bo; |
4763 | } |
4763 | } |
4764 | 4764 | ||
4765 | return NULL; |
4765 | return NULL; |
4766 | } |
4766 | } |
4767 | 4767 | ||
4768 | struct kgem_bo *kgem_create_buffer(struct kgem *kgem, |
4768 | struct kgem_bo *kgem_create_buffer(struct kgem *kgem, |
4769 | uint32_t size, uint32_t flags, |
4769 | uint32_t size, uint32_t flags, |
4770 | void **ret) |
4770 | void **ret) |
4771 | { |
4771 | { |
4772 | struct kgem_buffer *bo; |
4772 | struct kgem_buffer *bo; |
4773 | unsigned offset, alloc; |
4773 | unsigned offset, alloc; |
4774 | struct kgem_bo *old; |
4774 | struct kgem_bo *old; |
4775 | 4775 | ||
4776 | DBG(("%s: size=%d, flags=%x [write?=%d, inplace?=%d, last?=%d]\n", |
4776 | DBG(("%s: size=%d, flags=%x [write?=%d, inplace?=%d, last?=%d]\n", |
4777 | __FUNCTION__, size, flags, |
4777 | __FUNCTION__, size, flags, |
4778 | !!(flags & KGEM_BUFFER_WRITE), |
4778 | !!(flags & KGEM_BUFFER_WRITE), |
4779 | !!(flags & KGEM_BUFFER_INPLACE), |
4779 | !!(flags & KGEM_BUFFER_INPLACE), |
4780 | !!(flags & KGEM_BUFFER_LAST))); |
4780 | !!(flags & KGEM_BUFFER_LAST))); |
4781 | assert(size); |
4781 | assert(size); |
4782 | /* we should never be asked to create anything TOO large */ |
4782 | /* we should never be asked to create anything TOO large */ |
4783 | assert(size <= kgem->max_object_size); |
4783 | assert(size <= kgem->max_object_size); |
4784 | 4784 | ||
4785 | #if !DBG_NO_UPLOAD_CACHE |
4785 | #if !DBG_NO_UPLOAD_CACHE |
4786 | list_for_each_entry(bo, &kgem->batch_buffers, base.list) { |
4786 | list_for_each_entry(bo, &kgem->batch_buffers, base.list) { |
4787 | assert(bo->base.io); |
4787 | assert(bo->base.io); |
4788 | assert(bo->base.refcnt >= 1); |
4788 | assert(bo->base.refcnt >= 1); |
4789 | 4789 | ||
4790 | /* We can reuse any write buffer which we can fit */ |
4790 | /* We can reuse any write buffer which we can fit */ |
4791 | if (flags == KGEM_BUFFER_LAST && |
4791 | if (flags == KGEM_BUFFER_LAST && |
4792 | bo->write == KGEM_BUFFER_WRITE && |
4792 | bo->write == KGEM_BUFFER_WRITE && |
4793 | bo->base.refcnt == 1 && !bo->mmapped && |
4793 | bo->base.refcnt == 1 && !bo->mmapped && |
4794 | size <= bytes(&bo->base)) { |
4794 | size <= bytes(&bo->base)) { |
4795 | DBG(("%s: reusing write buffer for read of %d bytes? used=%d, total=%d\n", |
4795 | DBG(("%s: reusing write buffer for read of %d bytes? used=%d, total=%d\n", |
4796 | __FUNCTION__, size, bo->used, bytes(&bo->base))); |
4796 | __FUNCTION__, size, bo->used, bytes(&bo->base))); |
4797 | gem_write(kgem->fd, bo->base.handle, |
4797 | gem_write(kgem->fd, bo->base.handle, |
4798 | 0, bo->used, bo->mem); |
4798 | 0, bo->used, bo->mem); |
4799 | kgem_buffer_release(kgem, bo); |
4799 | kgem_buffer_release(kgem, bo); |
4800 | bo->need_io = 0; |
4800 | bo->need_io = 0; |
4801 | bo->write = 0; |
4801 | bo->write = 0; |
4802 | offset = 0; |
4802 | offset = 0; |
4803 | bo->used = size; |
4803 | bo->used = size; |
4804 | goto done; |
4804 | goto done; |
4805 | } |
4805 | } |
4806 | 4806 | ||
4807 | if (flags & KGEM_BUFFER_WRITE) { |
4807 | if (flags & KGEM_BUFFER_WRITE) { |
4808 | if ((bo->write & KGEM_BUFFER_WRITE) == 0 || |
4808 | if ((bo->write & KGEM_BUFFER_WRITE) == 0 || |
4809 | (((bo->write & ~flags) & KGEM_BUFFER_INPLACE) && |
4809 | (((bo->write & ~flags) & KGEM_BUFFER_INPLACE) && |
4810 | !bo->base.snoop)) { |
4810 | !bo->base.snoop)) { |
4811 | DBG(("%s: skip write %x buffer, need %x\n", |
4811 | DBG(("%s: skip write %x buffer, need %x\n", |
4812 | __FUNCTION__, bo->write, flags)); |
4812 | __FUNCTION__, bo->write, flags)); |
4813 | continue; |
4813 | continue; |
4814 | } |
4814 | } |
4815 | assert(bo->mmapped || bo->need_io); |
4815 | assert(bo->mmapped || bo->need_io); |
4816 | } else { |
4816 | } else { |
4817 | if (bo->write & KGEM_BUFFER_WRITE) { |
4817 | if (bo->write & KGEM_BUFFER_WRITE) { |
4818 | DBG(("%s: skip write %x buffer, need %x\n", |
4818 | DBG(("%s: skip write %x buffer, need %x\n", |
4819 | __FUNCTION__, bo->write, flags)); |
4819 | __FUNCTION__, bo->write, flags)); |
4820 | continue; |
4820 | continue; |
4821 | } |
4821 | } |
4822 | } |
4822 | } |
4823 | 4823 | ||
4824 | if (bo->used + size <= bytes(&bo->base)) { |
4824 | if (bo->used + size <= bytes(&bo->base)) { |
4825 | DBG(("%s: reusing buffer? used=%d + size=%d, total=%d\n", |
4825 | DBG(("%s: reusing buffer? used=%d + size=%d, total=%d\n", |
4826 | __FUNCTION__, bo->used, size, bytes(&bo->base))); |
4826 | __FUNCTION__, bo->used, size, bytes(&bo->base))); |
4827 | offset = bo->used; |
4827 | offset = bo->used; |
4828 | bo->used += size; |
4828 | bo->used += size; |
4829 | goto done; |
4829 | goto done; |
4830 | } |
4830 | } |
4831 | } |
4831 | } |
4832 | 4832 | ||
4833 | if (flags & KGEM_BUFFER_WRITE) { |
4833 | if (flags & KGEM_BUFFER_WRITE) { |
4834 | list_for_each_entry(bo, &kgem->active_buffers, base.list) { |
4834 | list_for_each_entry(bo, &kgem->active_buffers, base.list) { |
4835 | assert(bo->base.io); |
4835 | assert(bo->base.io); |
4836 | assert(bo->base.refcnt >= 1); |
4836 | assert(bo->base.refcnt >= 1); |
4837 | assert(bo->mmapped); |
4837 | assert(bo->mmapped); |
4838 | assert(!IS_CPU_MAP(bo->base.map) || kgem->has_llc || bo->base.snoop); |
4838 | assert(!IS_CPU_MAP(bo->base.map) || kgem->has_llc || bo->base.snoop); |
4839 | 4839 | ||
4840 | if (!kgem->has_llc && (bo->write & ~flags) & KGEM_BUFFER_INPLACE) { |
4840 | if (!kgem->has_llc && (bo->write & ~flags) & KGEM_BUFFER_INPLACE) { |
4841 | DBG(("%s: skip write %x buffer, need %x\n", |
4841 | DBG(("%s: skip write %x buffer, need %x\n", |
4842 | __FUNCTION__, bo->write, flags)); |
4842 | __FUNCTION__, bo->write, flags)); |
4843 | continue; |
4843 | continue; |
4844 | } |
4844 | } |
4845 | 4845 | ||
4846 | if (bo->used + size <= bytes(&bo->base)) { |
4846 | if (bo->used + size <= bytes(&bo->base)) { |
4847 | DBG(("%s: reusing buffer? used=%d + size=%d, total=%d\n", |
4847 | DBG(("%s: reusing buffer? used=%d + size=%d, total=%d\n", |
4848 | __FUNCTION__, bo->used, size, bytes(&bo->base))); |
4848 | __FUNCTION__, bo->used, size, bytes(&bo->base))); |
4849 | offset = bo->used; |
4849 | offset = bo->used; |
4850 | bo->used += size; |
4850 | bo->used += size; |
4851 | list_move(&bo->base.list, &kgem->batch_buffers); |
4851 | list_move(&bo->base.list, &kgem->batch_buffers); |
4852 | goto done; |
4852 | goto done; |
4853 | } |
4853 | } |
4854 | } |
4854 | } |
4855 | } |
4855 | } |
4856 | #endif |
4856 | #endif |
4857 | 4857 | ||
4858 | #if !DBG_NO_MAP_UPLOAD |
4858 | #if !DBG_NO_MAP_UPLOAD |
4859 | /* Be a little more generous and hope to hold fewer mmappings */ |
4859 | /* Be a little more generous and hope to hold fewer mmappings */ |
4860 | alloc = ALIGN(2*size, kgem->buffer_size); |
4860 | alloc = ALIGN(2*size, kgem->buffer_size); |
4861 | if (alloc > MAX_CACHE_SIZE) |
4861 | if (alloc > MAX_CACHE_SIZE) |
4862 | alloc = ALIGN(size, kgem->buffer_size); |
4862 | alloc = ALIGN(size, kgem->buffer_size); |
4863 | if (alloc > MAX_CACHE_SIZE) |
4863 | if (alloc > MAX_CACHE_SIZE) |
4864 | alloc = PAGE_ALIGN(size); |
4864 | alloc = PAGE_ALIGN(size); |
4865 | assert(alloc); |
4865 | assert(alloc); |
4866 | 4866 | ||
4867 | if (alloc > kgem->aperture_mappable / 4) |
4867 | if (alloc > kgem->aperture_mappable / 4) |
4868 | flags &= ~KGEM_BUFFER_INPLACE; |
4868 | flags &= ~KGEM_BUFFER_INPLACE; |
4869 | alloc /= PAGE_SIZE; |
4869 | alloc /= PAGE_SIZE; |
4870 | 4870 | ||
4871 | if (kgem->has_llc && |
4871 | if (kgem->has_llc && |
4872 | (flags & KGEM_BUFFER_WRITE_INPLACE) != KGEM_BUFFER_WRITE_INPLACE) { |
4872 | (flags & KGEM_BUFFER_WRITE_INPLACE) != KGEM_BUFFER_WRITE_INPLACE) { |
4873 | bo = buffer_alloc(); |
4873 | bo = buffer_alloc(); |
4874 | if (bo == NULL) |
4874 | if (bo == NULL) |
4875 | goto skip_llc; |
4875 | goto skip_llc; |
4876 | 4876 | ||
4877 | old = NULL; |
4877 | old = NULL; |
4878 | if ((flags & KGEM_BUFFER_WRITE) == 0) |
4878 | if ((flags & KGEM_BUFFER_WRITE) == 0) |
4879 | old = search_linear_cache(kgem, alloc, CREATE_CPU_MAP); |
4879 | old = search_linear_cache(kgem, alloc, CREATE_CPU_MAP); |
4880 | if (old == NULL) |
4880 | if (old == NULL) |
4881 | old = search_linear_cache(kgem, alloc, CREATE_INACTIVE | CREATE_CPU_MAP); |
4881 | old = search_linear_cache(kgem, alloc, CREATE_INACTIVE | CREATE_CPU_MAP); |
4882 | if (old == NULL) |
4882 | if (old == NULL) |
4883 | old = search_linear_cache(kgem, NUM_PAGES(size), CREATE_INACTIVE | CREATE_CPU_MAP); |
4883 | old = search_linear_cache(kgem, NUM_PAGES(size), CREATE_INACTIVE | CREATE_CPU_MAP); |
4884 | if (old) { |
4884 | if (old) { |
4885 | DBG(("%s: found LLC handle=%d for buffer\n", |
4885 | DBG(("%s: found LLC handle=%d for buffer\n", |
4886 | __FUNCTION__, old->handle)); |
4886 | __FUNCTION__, old->handle)); |
4887 | 4887 | ||
4888 | init_buffer_from_bo(bo, old); |
4888 | init_buffer_from_bo(bo, old); |
4889 | } else { |
4889 | } else { |
4890 | uint32_t handle = gem_create(kgem->fd, alloc); |
4890 | uint32_t handle = gem_create(kgem->fd, alloc); |
4891 | if (handle == 0) { |
4891 | if (handle == 0) { |
4892 | free(bo); |
4892 | free(bo); |
4893 | goto skip_llc; |
4893 | goto skip_llc; |
4894 | } |
4894 | } |
4895 | __kgem_bo_init(&bo->base, handle, alloc); |
4895 | __kgem_bo_init(&bo->base, handle, alloc); |
4896 | DBG(("%s: created LLC handle=%d for buffer\n", |
4896 | DBG(("%s: created LLC handle=%d for buffer\n", |
4897 | __FUNCTION__, bo->base.handle)); |
4897 | __FUNCTION__, bo->base.handle)); |
4898 | 4898 | ||
4899 | debug_alloc(kgem, alloc); |
4899 | debug_alloc(kgem, alloc); |
4900 | } |
4900 | } |
4901 | 4901 | ||
4902 | assert(bo->mmapped); |
4902 | assert(bo->mmapped); |
4903 | assert(!bo->need_io); |
4903 | assert(!bo->need_io); |
4904 | 4904 | ||
4905 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
4905 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
4906 | if (bo->mem) { |
4906 | if (bo->mem) { |
4907 | if (flags & KGEM_BUFFER_WRITE) |
4907 | if (flags & KGEM_BUFFER_WRITE) |
4908 | kgem_bo_sync__cpu(kgem, &bo->base); |
4908 | kgem_bo_sync__cpu(kgem, &bo->base); |
4909 | flags &= ~KGEM_BUFFER_INPLACE; |
4909 | flags &= ~KGEM_BUFFER_INPLACE; |
4910 | goto init; |
4910 | goto init; |
4911 | } else { |
4911 | } else { |
4912 | bo->base.refcnt = 0; /* for valgrind */ |
4912 | bo->base.refcnt = 0; /* for valgrind */ |
4913 | kgem_bo_free(kgem, &bo->base); |
4913 | kgem_bo_free(kgem, &bo->base); |
4914 | } |
4914 | } |
4915 | } |
4915 | } |
4916 | skip_llc: |
4916 | skip_llc: |
4917 | 4917 | ||
4918 | if ((flags & KGEM_BUFFER_WRITE_INPLACE) == KGEM_BUFFER_WRITE_INPLACE) { |
4918 | if ((flags & KGEM_BUFFER_WRITE_INPLACE) == KGEM_BUFFER_WRITE_INPLACE) { |
4919 | /* The issue with using a GTT upload buffer is that we may |
4919 | /* The issue with using a GTT upload buffer is that we may |
4920 | * cause eviction-stalls in order to free up some GTT space. |
4920 | * cause eviction-stalls in order to free up some GTT space. |
4921 | * An is-mappable? ioctl could help us detect when we are |
4921 | * An is-mappable? ioctl could help us detect when we are |
4922 | * about to block, or some per-page magic in the kernel. |
4922 | * about to block, or some per-page magic in the kernel. |
4923 | * |
4923 | * |
4924 | * XXX This is especially noticeable on memory constrained |
4924 | * XXX This is especially noticeable on memory constrained |
4925 | * devices like gen2 or with relatively slow gpu like i3. |
4925 | * devices like gen2 or with relatively slow gpu like i3. |
4926 | */ |
4926 | */ |
4927 | DBG(("%s: searching for an inactive GTT map for upload\n", |
4927 | DBG(("%s: searching for an inactive GTT map for upload\n", |
4928 | __FUNCTION__)); |
4928 | __FUNCTION__)); |
4929 | old = search_linear_cache(kgem, alloc, |
4929 | old = search_linear_cache(kgem, alloc, |
4930 | CREATE_EXACT | CREATE_INACTIVE | CREATE_GTT_MAP); |
4930 | CREATE_EXACT | CREATE_INACTIVE | CREATE_GTT_MAP); |
4931 | #if HAVE_I915_GEM_BUFFER_INFO |
4931 | #if HAVE_I915_GEM_BUFFER_INFO |
4932 | if (old) { |
4932 | if (old) { |
4933 | struct drm_i915_gem_buffer_info info; |
4933 | struct drm_i915_gem_buffer_info info; |
4934 | 4934 | ||
4935 | /* An example of such a non-blocking ioctl might work */ |
4935 | /* An example of such a non-blocking ioctl might work */ |
4936 | 4936 | ||
4937 | VG_CLEAR(info); |
4937 | VG_CLEAR(info); |
4938 | info.handle = handle; |
4938 | info.handle = handle; |
4939 | if (drmIoctl(kgem->fd, |
4939 | if (drmIoctl(kgem->fd, |
4940 | DRM_IOCTL_I915_GEM_BUFFER_INFO, |
4940 | DRM_IOCTL_I915_GEM_BUFFER_INFO, |
4941 | &fino) == 0) { |
4941 | &fino) == 0) { |
4942 | old->presumed_offset = info.addr; |
4942 | old->presumed_offset = info.addr; |
4943 | if ((info.flags & I915_GEM_MAPPABLE) == 0) { |
4943 | if ((info.flags & I915_GEM_MAPPABLE) == 0) { |
4944 | kgem_bo_move_to_inactive(kgem, old); |
4944 | kgem_bo_move_to_inactive(kgem, old); |
4945 | old = NULL; |
4945 | old = NULL; |
4946 | } |
4946 | } |
4947 | } |
4947 | } |
4948 | } |
4948 | } |
4949 | #endif |
4949 | #endif |
4950 | if (old == NULL) |
4950 | if (old == NULL) |
4951 | old = search_linear_cache(kgem, NUM_PAGES(size), |
4951 | old = search_linear_cache(kgem, NUM_PAGES(size), |
4952 | CREATE_EXACT | CREATE_INACTIVE | CREATE_GTT_MAP); |
4952 | CREATE_EXACT | CREATE_INACTIVE | CREATE_GTT_MAP); |
4953 | if (old == NULL) { |
4953 | if (old == NULL) { |
4954 | old = search_linear_cache(kgem, alloc, CREATE_INACTIVE); |
4954 | old = search_linear_cache(kgem, alloc, CREATE_INACTIVE); |
4955 | if (old && !__kgem_bo_is_mappable(kgem, old)) { |
4955 | if (old && !__kgem_bo_is_mappable(kgem, old)) { |
4956 | _kgem_bo_destroy(kgem, old); |
4956 | _kgem_bo_destroy(kgem, old); |
4957 | old = NULL; |
4957 | old = NULL; |
4958 | } |
4958 | } |
4959 | } |
4959 | } |
4960 | if (old) { |
4960 | if (old) { |
4961 | DBG(("%s: reusing handle=%d for buffer\n", |
4961 | DBG(("%s: reusing handle=%d for buffer\n", |
4962 | __FUNCTION__, old->handle)); |
4962 | __FUNCTION__, old->handle)); |
4963 | assert(__kgem_bo_is_mappable(kgem, old)); |
4963 | assert(__kgem_bo_is_mappable(kgem, old)); |
4964 | assert(!old->snoop); |
4964 | assert(!old->snoop); |
4965 | assert(old->rq == NULL); |
4965 | assert(old->rq == NULL); |
4966 | 4966 | ||
4967 | bo = buffer_alloc(); |
4967 | bo = buffer_alloc(); |
4968 | if (bo == NULL) |
4968 | if (bo == NULL) |
4969 | return NULL; |
4969 | return NULL; |
4970 | 4970 | ||
4971 | init_buffer_from_bo(bo, old); |
4971 | init_buffer_from_bo(bo, old); |
4972 | assert(num_pages(&bo->base) >= NUM_PAGES(size)); |
4972 | assert(num_pages(&bo->base) >= NUM_PAGES(size)); |
4973 | 4973 | ||
4974 | assert(bo->mmapped); |
4974 | assert(bo->mmapped); |
4975 | assert(bo->base.refcnt == 1); |
4975 | assert(bo->base.refcnt == 1); |
4976 | 4976 | ||
4977 | bo->mem = kgem_bo_map(kgem, &bo->base); |
4977 | bo->mem = kgem_bo_map(kgem, &bo->base); |
4978 | if (bo->mem) { |
4978 | if (bo->mem) { |
4979 | if (IS_CPU_MAP(bo->base.map)) |
4979 | if (IS_CPU_MAP(bo->base.map)) |
4980 | flags &= ~KGEM_BUFFER_INPLACE; |
4980 | flags &= ~KGEM_BUFFER_INPLACE; |
4981 | goto init; |
4981 | goto init; |
4982 | } else { |
4982 | } else { |
4983 | bo->base.refcnt = 0; |
4983 | bo->base.refcnt = 0; |
4984 | kgem_bo_free(kgem, &bo->base); |
4984 | kgem_bo_free(kgem, &bo->base); |
4985 | } |
4985 | } |
4986 | } |
4986 | } |
4987 | } |
4987 | } |
4988 | #else |
4988 | #else |
4989 | flags &= ~KGEM_BUFFER_INPLACE; |
4989 | flags &= ~KGEM_BUFFER_INPLACE; |
4990 | #endif |
4990 | #endif |
4991 | /* Be more parsimonious with pwrite/pread/cacheable buffers */ |
4991 | /* Be more parsimonious with pwrite/pread/cacheable buffers */ |
4992 | if ((flags & KGEM_BUFFER_INPLACE) == 0) |
4992 | if ((flags & KGEM_BUFFER_INPLACE) == 0) |
4993 | alloc = NUM_PAGES(size); |
4993 | alloc = NUM_PAGES(size); |
4994 | 4994 | ||
4995 | if (use_snoopable_buffer(kgem, flags)) { |
4995 | if (use_snoopable_buffer(kgem, flags)) { |
4996 | bo = search_snoopable_buffer(kgem, alloc); |
4996 | bo = search_snoopable_buffer(kgem, alloc); |
4997 | if (bo) { |
4997 | if (bo) { |
4998 | if (flags & KGEM_BUFFER_WRITE) |
4998 | if (flags & KGEM_BUFFER_WRITE) |
4999 | kgem_bo_sync__cpu(kgem, &bo->base); |
4999 | kgem_bo_sync__cpu(kgem, &bo->base); |
5000 | flags &= ~KGEM_BUFFER_INPLACE; |
5000 | flags &= ~KGEM_BUFFER_INPLACE; |
5001 | goto init; |
5001 | goto init; |
5002 | } |
5002 | } |
5003 | 5003 | ||
5004 | if ((flags & KGEM_BUFFER_INPLACE) == 0) { |
5004 | if ((flags & KGEM_BUFFER_INPLACE) == 0) { |
5005 | bo = create_snoopable_buffer(kgem, alloc); |
5005 | bo = create_snoopable_buffer(kgem, alloc); |
5006 | if (bo) |
5006 | if (bo) |
5007 | goto init; |
5007 | goto init; |
5008 | } |
5008 | } |
5009 | } |
5009 | } |
5010 | 5010 | ||
5011 | flags &= ~KGEM_BUFFER_INPLACE; |
5011 | flags &= ~KGEM_BUFFER_INPLACE; |
5012 | 5012 | ||
5013 | old = NULL; |
5013 | old = NULL; |
5014 | if ((flags & KGEM_BUFFER_WRITE) == 0) |
5014 | if ((flags & KGEM_BUFFER_WRITE) == 0) |
5015 | old = search_linear_cache(kgem, alloc, 0); |
5015 | old = search_linear_cache(kgem, alloc, 0); |
5016 | if (old == NULL) |
5016 | if (old == NULL) |
5017 | old = search_linear_cache(kgem, alloc, CREATE_INACTIVE); |
5017 | old = search_linear_cache(kgem, alloc, CREATE_INACTIVE); |
5018 | if (old) { |
5018 | if (old) { |
5019 | DBG(("%s: reusing ordinary handle %d for io\n", |
5019 | DBG(("%s: reusing ordinary handle %d for io\n", |
5020 | __FUNCTION__, old->handle)); |
5020 | __FUNCTION__, old->handle)); |
5021 | bo = buffer_alloc_with_data(num_pages(old)); |
5021 | bo = buffer_alloc_with_data(num_pages(old)); |
5022 | if (bo == NULL) |
5022 | if (bo == NULL) |
5023 | return NULL; |
5023 | return NULL; |
5024 | 5024 | ||
5025 | init_buffer_from_bo(bo, old); |
5025 | init_buffer_from_bo(bo, old); |
5026 | bo->need_io = flags & KGEM_BUFFER_WRITE; |
5026 | bo->need_io = flags & KGEM_BUFFER_WRITE; |
5027 | } else { |
5027 | } else { |
5028 | unsigned hint; |
5028 | unsigned hint; |
5029 | 5029 | ||
5030 | if (use_snoopable_buffer(kgem, flags)) { |
5030 | if (use_snoopable_buffer(kgem, flags)) { |
5031 | bo = create_snoopable_buffer(kgem, alloc); |
5031 | bo = create_snoopable_buffer(kgem, alloc); |
5032 | if (bo) |
5032 | if (bo) |
5033 | goto init; |
5033 | goto init; |
5034 | } |
5034 | } |
5035 | 5035 | ||
5036 | bo = buffer_alloc(); |
5036 | bo = buffer_alloc(); |
5037 | if (bo == NULL) |
5037 | if (bo == NULL) |
5038 | return NULL; |
5038 | return NULL; |
5039 | 5039 | ||
5040 | hint = CREATE_INACTIVE; |
5040 | hint = CREATE_INACTIVE; |
5041 | if (flags & KGEM_BUFFER_WRITE) |
5041 | if (flags & KGEM_BUFFER_WRITE) |
5042 | hint |= CREATE_CPU_MAP; |
5042 | hint |= CREATE_CPU_MAP; |
5043 | old = search_linear_cache(kgem, alloc, hint); |
5043 | old = search_linear_cache(kgem, alloc, hint); |
5044 | if (old) { |
5044 | if (old) { |
5045 | DBG(("%s: reusing handle=%d for buffer\n", |
5045 | DBG(("%s: reusing handle=%d for buffer\n", |
5046 | __FUNCTION__, old->handle)); |
5046 | __FUNCTION__, old->handle)); |
5047 | 5047 | ||
5048 | init_buffer_from_bo(bo, old); |
5048 | init_buffer_from_bo(bo, old); |
5049 | } else { |
5049 | } else { |
5050 | uint32_t handle = gem_create(kgem->fd, alloc); |
5050 | uint32_t handle = gem_create(kgem->fd, alloc); |
5051 | if (handle == 0) { |
5051 | if (handle == 0) { |
5052 | free(bo); |
5052 | free(bo); |
5053 | return NULL; |
5053 | return NULL; |
5054 | } |
5054 | } |
5055 | 5055 | ||
5056 | DBG(("%s: created handle=%d for buffer\n", |
5056 | DBG(("%s: created handle=%d for buffer\n", |
5057 | __FUNCTION__, handle)); |
5057 | __FUNCTION__, handle)); |
5058 | 5058 | ||
5059 | __kgem_bo_init(&bo->base, handle, alloc); |
5059 | __kgem_bo_init(&bo->base, handle, alloc); |
5060 | debug_alloc(kgem, alloc * PAGE_SIZE); |
5060 | debug_alloc(kgem, alloc * PAGE_SIZE); |
5061 | } |
5061 | } |
5062 | 5062 | ||
5063 | assert(bo->mmapped); |
5063 | assert(bo->mmapped); |
5064 | assert(!bo->need_io); |
5064 | assert(!bo->need_io); |
5065 | assert(bo->base.refcnt == 1); |
5065 | assert(bo->base.refcnt == 1); |
5066 | 5066 | ||
5067 | if (flags & KGEM_BUFFER_WRITE) { |
5067 | if (flags & KGEM_BUFFER_WRITE) { |
5068 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
5068 | bo->mem = kgem_bo_map__cpu(kgem, &bo->base); |
5069 | if (bo->mem != NULL) { |
5069 | if (bo->mem != NULL) { |
5070 | kgem_bo_sync__cpu(kgem, &bo->base); |
5070 | kgem_bo_sync__cpu(kgem, &bo->base); |
5071 | goto init; |
5071 | goto init; |
5072 | } |
5072 | } |
5073 | } |
5073 | } |
5074 | 5074 | ||
5075 | DBG(("%s: failing back to new pwrite buffer\n", __FUNCTION__)); |
5075 | DBG(("%s: failing back to new pwrite buffer\n", __FUNCTION__)); |
5076 | old = &bo->base; |
5076 | old = &bo->base; |
5077 | bo = buffer_alloc_with_data(num_pages(old)); |
5077 | bo = buffer_alloc_with_data(num_pages(old)); |
5078 | if (bo == NULL) { |
5078 | if (bo == NULL) { |
5079 | old->refcnt= 0; |
5079 | old->refcnt= 0; |
5080 | kgem_bo_free(kgem, old); |
5080 | kgem_bo_free(kgem, old); |
5081 | return NULL; |
5081 | return NULL; |
5082 | } |
5082 | } |
5083 | 5083 | ||
5084 | init_buffer_from_bo(bo, old); |
5084 | init_buffer_from_bo(bo, old); |
5085 | 5085 | ||
5086 | assert(bo->mem); |
5086 | assert(bo->mem); |
5087 | assert(!bo->mmapped); |
5087 | assert(!bo->mmapped); |
5088 | assert(bo->base.refcnt == 1); |
5088 | assert(bo->base.refcnt == 1); |
5089 | 5089 | ||
5090 | bo->need_io = flags & KGEM_BUFFER_WRITE; |
5090 | bo->need_io = flags & KGEM_BUFFER_WRITE; |
5091 | } |
5091 | } |
5092 | init: |
5092 | init: |
5093 | bo->base.io = true; |
5093 | bo->base.io = true; |
5094 | assert(bo->base.refcnt == 1); |
5094 | assert(bo->base.refcnt == 1); |
5095 | assert(num_pages(&bo->base) >= NUM_PAGES(size)); |
5095 | assert(num_pages(&bo->base) >= NUM_PAGES(size)); |
5096 | assert(!bo->need_io || !bo->base.needs_flush); |
5096 | assert(!bo->need_io || !bo->base.needs_flush); |
5097 | assert(!bo->need_io || bo->base.domain != DOMAIN_GPU); |
5097 | assert(!bo->need_io || bo->base.domain != DOMAIN_GPU); |
5098 | assert(bo->mem); |
5098 | assert(bo->mem); |
5099 | assert(!bo->mmapped || bo->base.map != NULL); |
5099 | assert(!bo->mmapped || bo->base.map != NULL); |
5100 | 5100 | ||
5101 | bo->used = size; |
5101 | bo->used = size; |
5102 | bo->write = flags & KGEM_BUFFER_WRITE_INPLACE; |
5102 | bo->write = flags & KGEM_BUFFER_WRITE_INPLACE; |
5103 | offset = 0; |
5103 | offset = 0; |
5104 | 5104 | ||
5105 | assert(list_is_empty(&bo->base.list)); |
5105 | assert(list_is_empty(&bo->base.list)); |
5106 | list_add(&bo->base.list, &kgem->batch_buffers); |
5106 | list_add(&bo->base.list, &kgem->batch_buffers); |
5107 | 5107 | ||
5108 | DBG(("%s(pages=%d [%d]) new handle=%d, used=%d, write=%d\n", |
5108 | DBG(("%s(pages=%d [%d]) new handle=%d, used=%d, write=%d\n", |
5109 | __FUNCTION__, num_pages(&bo->base), alloc, bo->base.handle, bo->used, bo->write)); |
5109 | __FUNCTION__, num_pages(&bo->base), alloc, bo->base.handle, bo->used, bo->write)); |
5110 | 5110 | ||
5111 | done: |
5111 | done: |
5112 | bo->used = ALIGN(bo->used, UPLOAD_ALIGNMENT); |
5112 | bo->used = ALIGN(bo->used, UPLOAD_ALIGNMENT); |
5113 | assert(bo->mem); |
5113 | assert(bo->mem); |
5114 | *ret = (char *)bo->mem + offset; |
5114 | *ret = (char *)bo->mem + offset; |
5115 | return kgem_create_proxy(kgem, &bo->base, offset, size); |
5115 | return kgem_create_proxy(kgem, &bo->base, offset, size); |
5116 | } |
5116 | } |
5117 | 5117 | ||
5118 | bool kgem_buffer_is_inplace(struct kgem_bo *_bo) |
5118 | bool kgem_buffer_is_inplace(struct kgem_bo *_bo) |
5119 | { |
5119 | { |
5120 | struct kgem_buffer *bo = (struct kgem_buffer *)_bo->proxy; |
5120 | struct kgem_buffer *bo = (struct kgem_buffer *)_bo->proxy; |
5121 | return bo->write & KGEM_BUFFER_WRITE_INPLACE; |
5121 | return bo->write & KGEM_BUFFER_WRITE_INPLACE; |
5122 | } |
5122 | } |
5123 | 5123 | ||
5124 | struct kgem_bo *kgem_create_buffer_2d(struct kgem *kgem, |
5124 | struct kgem_bo *kgem_create_buffer_2d(struct kgem *kgem, |
5125 | int width, int height, int bpp, |
5125 | int width, int height, int bpp, |
5126 | uint32_t flags, |
5126 | uint32_t flags, |
5127 | void **ret) |
5127 | void **ret) |
5128 | { |
5128 | { |
5129 | struct kgem_bo *bo; |
5129 | struct kgem_bo *bo; |
5130 | int stride; |
5130 | int stride; |
5131 | 5131 | ||
5132 | assert(width > 0 && height > 0); |
5132 | assert(width > 0 && height > 0); |
5133 | assert(ret != NULL); |
5133 | assert(ret != NULL); |
5134 | stride = ALIGN(width, 2) * bpp >> 3; |
5134 | stride = ALIGN(width, 2) * bpp >> 3; |
5135 | stride = ALIGN(stride, 4); |
5135 | stride = ALIGN(stride, 4); |
5136 | 5136 | ||
5137 | DBG(("%s: %dx%d, %d bpp, stride=%d\n", |
5137 | DBG(("%s: %dx%d, %d bpp, stride=%d\n", |
5138 | __FUNCTION__, width, height, bpp, stride)); |
5138 | __FUNCTION__, width, height, bpp, stride)); |
5139 | 5139 | ||
5140 | bo = kgem_create_buffer(kgem, stride * ALIGN(height, 2), flags, ret); |
5140 | bo = kgem_create_buffer(kgem, stride * ALIGN(height, 2), flags, ret); |
5141 | if (bo == NULL) { |
5141 | if (bo == NULL) { |
5142 | DBG(("%s: allocation failure for upload buffer\n", |
5142 | DBG(("%s: allocation failure for upload buffer\n", |
5143 | __FUNCTION__)); |
5143 | __FUNCTION__)); |
5144 | return NULL; |
5144 | return NULL; |
5145 | } |
5145 | } |
5146 | assert(*ret != NULL); |
5146 | assert(*ret != NULL); |
5147 | assert(bo->proxy != NULL); |
5147 | assert(bo->proxy != NULL); |
5148 | 5148 | ||
5149 | if (height & 1) { |
5149 | if (height & 1) { |
5150 | struct kgem_buffer *io = (struct kgem_buffer *)bo->proxy; |
5150 | struct kgem_buffer *io = (struct kgem_buffer *)bo->proxy; |
5151 | int min; |
5151 | int min; |
5152 | 5152 | ||
5153 | assert(io->used); |
5153 | assert(io->used); |
5154 | 5154 | ||
5155 | /* Having padded this surface to ensure that accesses to |
5155 | /* Having padded this surface to ensure that accesses to |
5156 | * the last pair of rows is valid, remove the padding so |
5156 | * the last pair of rows is valid, remove the padding so |
5157 | * that it can be allocated to other pixmaps. |
5157 | * that it can be allocated to other pixmaps. |
5158 | */ |
5158 | */ |
5159 | min = bo->delta + height * stride; |
5159 | min = bo->delta + height * stride; |
5160 | min = ALIGN(min, UPLOAD_ALIGNMENT); |
5160 | min = ALIGN(min, UPLOAD_ALIGNMENT); |
5161 | if (io->used != min) { |
5161 | if (io->used != min) { |
5162 | DBG(("%s: trimming buffer from %d to %d\n", |
5162 | DBG(("%s: trimming buffer from %d to %d\n", |
5163 | __FUNCTION__, io->used, min)); |
5163 | __FUNCTION__, io->used, min)); |
5164 | io->used = min; |
5164 | io->used = min; |
5165 | } |
5165 | } |
5166 | bo->size.bytes -= stride; |
5166 | bo->size.bytes -= stride; |
5167 | } |
5167 | } |
5168 | 5168 | ||
5169 | bo->map = MAKE_CPU_MAP(*ret); |
5169 | bo->map = MAKE_CPU_MAP(*ret); |
5170 | bo->pitch = stride; |
5170 | bo->pitch = stride; |
5171 | bo->unique_id = kgem_get_unique_id(kgem); |
5171 | bo->unique_id = kgem_get_unique_id(kgem); |
5172 | return bo; |
5172 | return bo; |
5173 | } |
5173 | } |
5174 | 5174 | ||
5175 | struct kgem_bo *kgem_upload_source_image(struct kgem *kgem, |
5175 | struct kgem_bo *kgem_upload_source_image(struct kgem *kgem, |
5176 | const void *data, |
5176 | const void *data, |
5177 | const BoxRec *box, |
5177 | const BoxRec *box, |
5178 | int stride, int bpp) |
5178 | int stride, int bpp) |
5179 | { |
5179 | { |
5180 | int width = box->x2 - box->x1; |
5180 | int width = box->x2 - box->x1; |
5181 | int height = box->y2 - box->y1; |
5181 | int height = box->y2 - box->y1; |
5182 | struct kgem_bo *bo; |
5182 | struct kgem_bo *bo; |
5183 | void *dst; |
5183 | void *dst; |
5184 | 5184 | ||
5185 | if (!kgem_can_create_2d(kgem, width, height, bpp)) |
5185 | if (!kgem_can_create_2d(kgem, width, height, bpp)) |
5186 | return NULL; |
5186 | return NULL; |
5187 | 5187 | ||
5188 | DBG(("%s : (%d, %d), (%d, %d), stride=%d, bpp=%d\n", |
5188 | DBG(("%s : (%d, %d), (%d, %d), stride=%d, bpp=%d\n", |
5189 | __FUNCTION__, box->x1, box->y1, box->x2, box->y2, stride, bpp)); |
5189 | __FUNCTION__, box->x1, box->y1, box->x2, box->y2, stride, bpp)); |
5190 | 5190 | ||
5191 | assert(data); |
5191 | assert(data); |
5192 | assert(width > 0); |
5192 | assert(width > 0); |
5193 | assert(height > 0); |
5193 | assert(height > 0); |
5194 | assert(stride); |
5194 | assert(stride); |
5195 | assert(bpp); |
5195 | assert(bpp); |
5196 | 5196 | ||
5197 | bo = kgem_create_buffer_2d(kgem, |
5197 | bo = kgem_create_buffer_2d(kgem, |
5198 | width, height, bpp, |
5198 | width, height, bpp, |
5199 | KGEM_BUFFER_WRITE_INPLACE, &dst); |
5199 | KGEM_BUFFER_WRITE_INPLACE, &dst); |
5200 | if (bo) |
5200 | if (bo) |
5201 | memcpy_blt(data, dst, bpp, |
5201 | memcpy_blt(data, dst, bpp, |
5202 | stride, bo->pitch, |
5202 | stride, bo->pitch, |
5203 | box->x1, box->y1, |
5203 | box->x1, box->y1, |
5204 | 0, 0, |
5204 | 0, 0, |
5205 | width, height); |
5205 | width, height); |
5206 | 5206 | ||
5207 | return bo; |
5207 | return bo; |
5208 | } |
5208 | } |
5209 | 5209 | ||
5210 | void kgem_proxy_bo_attach(struct kgem_bo *bo, |
5210 | void kgem_proxy_bo_attach(struct kgem_bo *bo, |
5211 | struct kgem_bo **ptr) |
5211 | struct kgem_bo **ptr) |
5212 | { |
5212 | { |
5213 | DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle)); |
5213 | DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle)); |
5214 | assert(bo->map == NULL || IS_CPU_MAP(bo->map)); |
5214 | assert(bo->map == NULL || IS_CPU_MAP(bo->map)); |
5215 | assert(bo->proxy); |
5215 | assert(bo->proxy); |
5216 | list_add(&bo->vma, &bo->proxy->vma); |
5216 | list_add(&bo->vma, &bo->proxy->vma); |
5217 | bo->map = ptr; |
5217 | bo->map = ptr; |
5218 | *ptr = kgem_bo_reference(bo); |
5218 | *ptr = kgem_bo_reference(bo); |
5219 | } |
5219 | } |
5220 | 5220 | ||
5221 | void kgem_buffer_read_sync(struct kgem *kgem, struct kgem_bo *_bo) |
5221 | void kgem_buffer_read_sync(struct kgem *kgem, struct kgem_bo *_bo) |
5222 | { |
5222 | { |
5223 | struct kgem_buffer *bo; |
5223 | struct kgem_buffer *bo; |
5224 | uint32_t offset = _bo->delta, length = _bo->size.bytes; |
5224 | uint32_t offset = _bo->delta, length = _bo->size.bytes; |
5225 | 5225 | ||
5226 | /* We expect the caller to have already submitted the batch */ |
5226 | /* We expect the caller to have already submitted the batch */ |
5227 | assert(_bo->io); |
5227 | assert(_bo->io); |
5228 | assert(_bo->exec == NULL); |
5228 | assert(_bo->exec == NULL); |
5229 | assert(_bo->rq == NULL); |
5229 | assert(_bo->rq == NULL); |
5230 | assert(_bo->proxy); |
5230 | assert(_bo->proxy); |
5231 | 5231 | ||
5232 | _bo = _bo->proxy; |
5232 | _bo = _bo->proxy; |
5233 | assert(_bo->proxy == NULL); |
5233 | assert(_bo->proxy == NULL); |
5234 | assert(_bo->exec == NULL); |
5234 | assert(_bo->exec == NULL); |
5235 | 5235 | ||
5236 | bo = (struct kgem_buffer *)_bo; |
5236 | bo = (struct kgem_buffer *)_bo; |
5237 | 5237 | ||
5238 | DBG(("%s(offset=%d, length=%d, snooped=%d)\n", __FUNCTION__, |
5238 | DBG(("%s(offset=%d, length=%d, snooped=%d)\n", __FUNCTION__, |
5239 | offset, length, bo->base.snoop)); |
5239 | offset, length, bo->base.snoop)); |
5240 | 5240 | ||
5241 | if (bo->mmapped) { |
5241 | if (bo->mmapped) { |
5242 | struct drm_i915_gem_set_domain set_domain; |
5242 | struct drm_i915_gem_set_domain set_domain; |
5243 | 5243 | ||
5244 | DBG(("%s: sync: needs_flush? %d, domain? %d, busy? %d\n", |
5244 | DBG(("%s: sync: needs_flush? %d, domain? %d, busy? %d\n", |
5245 | __FUNCTION__, |
5245 | __FUNCTION__, |
5246 | bo->base.needs_flush, |
5246 | bo->base.needs_flush, |
5247 | bo->base.domain, |
5247 | bo->base.domain, |
5248 | __kgem_busy(kgem, bo->base.handle))); |
5248 | __kgem_busy(kgem, bo->base.handle))); |
5249 | 5249 | ||
5250 | assert(!IS_CPU_MAP(bo->base.map) || bo->base.snoop || kgem->has_llc); |
5250 | assert(!IS_CPU_MAP(bo->base.map) || bo->base.snoop || kgem->has_llc); |
5251 | 5251 | ||
5252 | VG_CLEAR(set_domain); |
5252 | VG_CLEAR(set_domain); |
5253 | set_domain.handle = bo->base.handle; |
5253 | set_domain.handle = bo->base.handle; |
5254 | set_domain.write_domain = 0; |
5254 | set_domain.write_domain = 0; |
5255 | set_domain.read_domains = |
5255 | set_domain.read_domains = |
5256 | IS_CPU_MAP(bo->base.map) ? I915_GEM_DOMAIN_CPU : I915_GEM_DOMAIN_GTT; |
5256 | IS_CPU_MAP(bo->base.map) ? I915_GEM_DOMAIN_CPU : I915_GEM_DOMAIN_GTT; |
5257 | 5257 | ||
5258 | if (drmIoctl(kgem->fd, |
5258 | if (drmIoctl(kgem->fd, |
5259 | DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) |
5259 | DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) |
5260 | return; |
5260 | return; |
5261 | } else { |
5261 | } else { |
5262 | if (gem_read(kgem->fd, |
5262 | if (gem_read(kgem->fd, |
5263 | bo->base.handle, (char *)bo->mem+offset, |
5263 | bo->base.handle, (char *)bo->mem+offset, |
5264 | offset, length)) |
5264 | offset, length)) |
5265 | return; |
5265 | return; |
5266 | } |
5266 | } |
5267 | kgem_bo_retire(kgem, &bo->base); |
5267 | kgem_bo_retire(kgem, &bo->base); |
5268 | bo->base.domain = DOMAIN_NONE; |
5268 | bo->base.domain = DOMAIN_NONE; |
5269 | } |
5269 | } |
5270 | #endif |
5270 | #endif |
5271 | 5271 | ||
5272 | uint32_t kgem_bo_get_binding(struct kgem_bo *bo, uint32_t format) |
5272 | uint32_t kgem_bo_get_binding(struct kgem_bo *bo, uint32_t format) |
5273 | { |
5273 | { |
5274 | struct kgem_bo_binding *b; |
5274 | struct kgem_bo_binding *b; |
5275 | 5275 | ||
5276 | for (b = &bo->binding; b && b->offset; b = b->next) |
5276 | for (b = &bo->binding; b && b->offset; b = b->next) |
5277 | if (format == b->format) |
5277 | if (format == b->format) |
5278 | return b->offset; |
5278 | return b->offset; |
5279 | 5279 | ||
5280 | return 0; |
5280 | return 0; |
5281 | } |
5281 | } |
5282 | 5282 | ||
5283 | void kgem_bo_set_binding(struct kgem_bo *bo, uint32_t format, uint16_t offset) |
5283 | void kgem_bo_set_binding(struct kgem_bo *bo, uint32_t format, uint16_t offset) |
5284 | { |
5284 | { |
5285 | struct kgem_bo_binding *b; |
5285 | struct kgem_bo_binding *b; |
5286 | 5286 | ||
5287 | for (b = &bo->binding; b; b = b->next) { |
5287 | for (b = &bo->binding; b; b = b->next) { |
5288 | if (b->offset) |
5288 | if (b->offset) |
5289 | continue; |
5289 | continue; |
5290 | 5290 | ||
5291 | b->offset = offset; |
5291 | b->offset = offset; |
5292 | b->format = format; |
5292 | b->format = format; |
5293 | 5293 | ||
5294 | if (b->next) |
5294 | if (b->next) |
5295 | b->next->offset = 0; |
5295 | b->next->offset = 0; |
5296 | 5296 | ||
5297 | return; |
5297 | return; |
5298 | } |
5298 | } |
5299 | 5299 | ||
5300 | b = malloc(sizeof(*b)); |
5300 | b = malloc(sizeof(*b)); |
5301 | if (b) { |
5301 | if (b) { |
5302 | b->next = bo->binding.next; |
5302 | b->next = bo->binding.next; |
5303 | b->format = format; |
5303 | b->format = format; |
5304 | b->offset = offset; |
5304 | b->offset = offset; |
5305 | bo->binding.next = b; |
5305 | bo->binding.next = b; |
5306 | } |
5306 | } |
5307 | } |
5307 | } |
5308 | 5308 | ||
5309 | int kgem_init_fb(struct kgem *kgem, struct sna_fb *fb) |
5309 | int kgem_init_fb(struct kgem *kgem, struct sna_fb *fb) |
5310 | { |
5310 | { |
5311 | struct kgem_bo *bo; |
5311 | struct kgem_bo *bo; |
5312 | size_t size; |
5312 | size_t size; |
5313 | int ret; |
5313 | int ret; |
5314 | 5314 | ||
5315 | ret = drmIoctl(kgem->fd, SRV_FBINFO, fb); |
5315 | ret = drmIoctl(kgem->fd, SRV_FBINFO, fb); |
5316 | if( ret != 0 ) |
5316 | if( ret != 0 ) |
5317 | return 0; |
5317 | return 0; |
5318 | 5318 | ||
5319 | size = fb->pitch * fb->height / PAGE_SIZE; |
5319 | size = fb->pitch * fb->height / PAGE_SIZE; |
5320 | 5320 | ||
5321 | bo = __kgem_bo_alloc(-2, size); |
5321 | bo = __kgem_bo_alloc(-2, size); |
5322 | if (!bo) { |
5322 | if (!bo) { |
5323 | return 0; |
5323 | return 0; |
5324 | } |
5324 | } |
5325 | 5325 | ||
5326 | bo->domain = DOMAIN_GTT; |
5326 | bo->domain = DOMAIN_GTT; |
5327 | bo->unique_id = kgem_get_unique_id(kgem); |
5327 | bo->unique_id = kgem_get_unique_id(kgem); |
5328 | bo->pitch = fb->pitch; |
5328 | bo->pitch = fb->pitch; |
5329 | bo->tiling = I915_TILING_NONE; |
5329 | bo->tiling = I915_TILING_X; |
5330 | bo->scanout = 1; |
5330 | bo->scanout = 1; |
5331 | fb->fb_bo = bo; |
5331 | fb->fb_bo = bo; |
5332 | 5332 | ||
5333 | // printf("fb width %d height %d pitch %d bo %p\n", |
5333 | // printf("fb width %d height %d pitch %d bo %p\n", |
5334 | // fb->width, fb->height, fb->pitch, fb->fb_bo); |
5334 | // fb->width, fb->height, fb->pitch, fb->fb_bo); |
5335 | 5335 | ||
5336 | return 1; |
5336 | return 1; |
5337 | }; |
5337 | }; |
5338 | 5338 | ||
5339 | 5339 | ||
5340 | int kgem_update_fb(struct kgem *kgem, struct sna_fb *fb) |
5340 | int kgem_update_fb(struct kgem *kgem, struct sna_fb *fb) |
5341 | { |
5341 | { |
5342 | struct kgem_bo *bo; |
5342 | struct kgem_bo *bo; |
5343 | size_t size; |
5343 | size_t size; |
5344 | int ret; |
5344 | int ret; |
5345 | 5345 | ||
5346 | bo = fb->fb_bo; |
5346 | bo = fb->fb_bo; |
5347 | 5347 | ||
5348 | ret = drmIoctl(kgem->fd, SRV_FBINFO, fb); |
5348 | ret = drmIoctl(kgem->fd, SRV_FBINFO, fb); |
5349 | if( ret != 0 ) |
5349 | if( ret != 0 ) |
5350 | return 0; |
5350 | return 0; |
5351 | 5351 | ||
5352 | fb->fb_bo = bo; |
5352 | fb->fb_bo = bo; |
5353 | 5353 | ||
5354 | size = fb->pitch * fb->height / PAGE_SIZE; |
5354 | size = fb->pitch * fb->height / PAGE_SIZE; |
5355 | 5355 | ||
5356 | if((size != bo->size.pages.count) || |
5356 | if((size != bo->size.pages.count) || |
5357 | (fb->pitch != bo->pitch)) |
5357 | (fb->pitch != bo->pitch)) |
5358 | { |
5358 | { |
5359 | bo->size.pages.count = size; |
5359 | bo->size.pages.count = size; |
5360 | bo->pitch = fb->pitch; |
5360 | bo->pitch = fb->pitch; |
5361 | 5361 | ||
5362 | printf("fb width %d height %d pitch %d bo %p\n", |
5362 | printf("fb width %d height %d pitch %d bo %p\n", |
5363 | fb->width, fb->height, fb->pitch, fb->fb_bo); |
5363 | fb->width, fb->height, fb->pitch, fb->fb_bo); |
5364 | 5364 | ||
5365 | return 1; |
5365 | return 1; |
5366 | } |
5366 | } |
5367 | 5367 | ||
5368 | return 0; |
5368 | return 0; |
5369 | }; |
5369 | }; |
5370 | 5370 | ||
5371 | void sna_bo_destroy(struct kgem *kgem, struct kgem_bo *bo) |
5371 | void sna_bo_destroy(struct kgem *kgem, struct kgem_bo *bo) |
5372 | { |
5372 | { |
5373 | kgem_bo_destroy(kgem, bo); |
5373 | kgem_bo_destroy(kgem, bo); |
5374 | kgem_bo_free(kgem, bo); |
5374 | kgem_bo_free(kgem, bo); |
5375 | } |
5375 | } |
5376 | 5376 | ||
5377 | 5377 | ||
5378 | void kgem_close_batches(struct kgem *kgem) |
5378 | void kgem_close_batches(struct kgem *kgem) |
5379 | { |
5379 | { |
5380 | int n; |
5380 | int n; |
5381 | 5381 | ||
5382 | for (n = 0; n < ARRAY_SIZE(kgem->pinned_batches); n++) { |
5382 | for (n = 0; n < ARRAY_SIZE(kgem->pinned_batches); n++) { |
5383 | while (!list_is_empty(&kgem->pinned_batches[n])) { |
5383 | while (!list_is_empty(&kgem->pinned_batches[n])) { |
5384 | kgem_bo_destroy(kgem, |
5384 | kgem_bo_destroy(kgem, |
5385 | list_first_entry(&kgem->pinned_batches[n], |
5385 | list_first_entry(&kgem->pinned_batches[n], |
5386 | struct kgem_bo, list)); |
5386 | struct kgem_bo, list)); |
5387 | } |
5387 | } |
5388 | } |
5388 | } |
5389 | }; |
5389 | }; |
5390 | 5390 | ||
5391 | struct kgem_bo *kgem_bo_from_handle(struct kgem *kgem, int handle, |
5391 | struct kgem_bo *kgem_bo_from_handle(struct kgem *kgem, int handle, |
5392 | int pitch, int height) |
5392 | int pitch, int height) |
5393 | { |
5393 | { |
5394 | struct kgem_bo *bo; |
5394 | struct kgem_bo *bo; |
5395 | int size; |
5395 | int size; |
5396 | 5396 | ||
5397 | size = pitch * height / PAGE_SIZE; |
5397 | size = pitch * height / PAGE_SIZE; |
5398 | 5398 | ||
5399 | bo = __kgem_bo_alloc(handle, size); |
5399 | bo = __kgem_bo_alloc(handle, size); |
5400 | if(bo == NULL) |
5400 | if(bo == NULL) |
5401 | return NULL; |
5401 | return NULL; |
5402 | 5402 | ||
5403 | bo->domain = DOMAIN_GTT; |
5403 | bo->domain = DOMAIN_GTT; |
5404 | bo->unique_id = kgem_get_unique_id(kgem); |
5404 | bo->unique_id = kgem_get_unique_id(kgem); |
5405 | bo->pitch = pitch; |
5405 | bo->pitch = pitch; |
5406 | bo->tiling = I915_TILING_X; |
5406 | bo->tiling = I915_TILING_X; |
5407 | bo->scanout = 0; |
5407 | bo->scanout = 0; |
5408 | 5408 | ||
5409 | return bo; |
5409 | return bo; |
5410 | }>=>=>=>=>><>=>><>=>=>>=>>>>>=>>>=>>=>>=>=>>=>>>>>=>>>=>=>>>>=>=>=>>>>>=>>>>>>>>>>=>>=>=>=>=>=>=>=>=>>>>=>>=>>=>>=>>>>>>=>>>=>>>=><=>>>=>=>=>>=>>>>><>><>><>>>>>>>>>>>>>>>>>>>>>=>31) |
5410 | }>=>=>=>=>><>=>><>=>=>>=>>>>>=>>>=>>=>>=>=>>=>>>>>=>>>=>=>>>>=>=>=>>>>>=>>>>>>>>>>=>>=>=>=>=>=>=>=>=>>>>=>>=>>=>>=>>>>>>=>>>=>>>=><=>>>=>=>=>>=>>>>><>><>><>>>>>>>>>>>>>>>>>>>>>=>31) |
5411 | ><31) |
5411 | ><31) |
5412 | >0) |
5412 | >0) |
5413 | #define><0) |
5413 | #define><0) |
5414 | #define>12) |
5414 | #define>12) |
5415 | struct><12) |
5415 | struct><12) |
5416 | struct>11) |
5416 | struct>11) |
5417 | #define><11) |
5417 | #define><11) |
5418 | #define>10) |
5418 | #define>10) |
5419 | #define><10) |
5419 | #define><10) |
5420 | #define>><>><> |
5420 | #define>><>><> |