Rev 6937 | Only display areas with differences | Regard whitespace | Details | Blame | Last modification | View Log | RSS feed
Rev 6937 | Rev 7144 | ||
---|---|---|---|
1 | /* |
1 | /* |
2 | * Copyright © 2008-2015 Intel Corporation |
2 | * Copyright © 2008-2015 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 |
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING |
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
21 | * IN THE SOFTWARE. |
21 | * IN THE SOFTWARE. |
22 | * |
22 | * |
23 | * Authors: |
23 | * Authors: |
24 | * Eric Anholt |
24 | * Eric Anholt |
25 | * |
25 | * |
26 | */ |
26 | */ |
27 | 27 | ||
28 | #include |
28 | #include |
29 | #include |
29 | #include |
30 | #include |
30 | #include |
31 | #include "i915_drv.h" |
31 | #include "i915_drv.h" |
32 | #include "i915_vgpu.h" |
32 | #include "i915_vgpu.h" |
33 | #include "i915_trace.h" |
33 | #include "i915_trace.h" |
34 | #include "intel_drv.h" |
34 | #include "intel_drv.h" |
35 | #include |
35 | #include |
36 | #include |
36 | #include |
37 | #include |
37 | #include |
38 | #include |
38 | #include |
39 | #include |
39 | #include |
40 | 40 | ||
41 | #define RQ_BUG_ON(expr) |
41 | #define RQ_BUG_ON(expr) |
42 | 42 | ||
43 | extern int x86_clflush_size; |
43 | extern int x86_clflush_size; |
44 | #define __copy_to_user_inatomic __copy_to_user |
- | |
45 | 44 | ||
46 | #define PROT_READ 0x1 /* page can be read */ |
45 | #define PROT_READ 0x1 /* page can be read */ |
47 | #define PROT_WRITE 0x2 /* page can be written */ |
46 | #define PROT_WRITE 0x2 /* page can be written */ |
48 | #define MAP_SHARED 0x01 /* Share changes */ |
47 | #define MAP_SHARED 0x01 /* Share changes */ |
49 | - | ||
50 | - | ||
51 | 48 | ||
52 | struct drm_i915_gem_object *get_fb_obj(); |
49 | struct drm_i915_gem_object *get_fb_obj(); |
53 | 50 | ||
54 | unsigned long vm_mmap(struct file *file, unsigned long addr, |
51 | unsigned long vm_mmap(struct file *file, unsigned long addr, |
55 | unsigned long len, unsigned long prot, |
52 | unsigned long len, unsigned long prot, |
56 | unsigned long flag, unsigned long offset); |
53 | unsigned long flag, unsigned long offset); |
57 | 54 | ||
58 | 55 | ||
59 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); |
56 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); |
60 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj); |
57 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj); |
61 | static void |
58 | static void |
62 | i915_gem_object_retire__write(struct drm_i915_gem_object *obj); |
59 | i915_gem_object_retire__write(struct drm_i915_gem_object *obj); |
63 | static void |
60 | static void |
64 | i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring); |
61 | i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring); |
65 | 62 | ||
66 | static bool cpu_cache_is_coherent(struct drm_device *dev, |
63 | static bool cpu_cache_is_coherent(struct drm_device *dev, |
67 | enum i915_cache_level level) |
64 | enum i915_cache_level level) |
68 | { |
65 | { |
69 | return HAS_LLC(dev) || level != I915_CACHE_NONE; |
66 | return HAS_LLC(dev) || level != I915_CACHE_NONE; |
70 | } |
67 | } |
71 | 68 | ||
72 | static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) |
69 | static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) |
73 | { |
70 | { |
74 | if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) |
71 | if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) |
75 | return true; |
72 | return true; |
76 | 73 | ||
77 | return obj->pin_display; |
74 | return obj->pin_display; |
78 | } |
75 | } |
79 | 76 | ||
80 | /* some bookkeeping */ |
77 | /* some bookkeeping */ |
81 | static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv, |
78 | static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv, |
82 | size_t size) |
79 | size_t size) |
83 | { |
80 | { |
84 | spin_lock(&dev_priv->mm.object_stat_lock); |
81 | spin_lock(&dev_priv->mm.object_stat_lock); |
85 | dev_priv->mm.object_count++; |
82 | dev_priv->mm.object_count++; |
86 | dev_priv->mm.object_memory += size; |
83 | dev_priv->mm.object_memory += size; |
87 | spin_unlock(&dev_priv->mm.object_stat_lock); |
84 | spin_unlock(&dev_priv->mm.object_stat_lock); |
88 | } |
85 | } |
89 | 86 | ||
90 | static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv, |
87 | static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv, |
91 | size_t size) |
88 | size_t size) |
92 | { |
89 | { |
93 | spin_lock(&dev_priv->mm.object_stat_lock); |
90 | spin_lock(&dev_priv->mm.object_stat_lock); |
94 | dev_priv->mm.object_count--; |
91 | dev_priv->mm.object_count--; |
95 | dev_priv->mm.object_memory -= size; |
92 | dev_priv->mm.object_memory -= size; |
96 | spin_unlock(&dev_priv->mm.object_stat_lock); |
93 | spin_unlock(&dev_priv->mm.object_stat_lock); |
97 | } |
94 | } |
98 | 95 | ||
99 | static int |
96 | static int |
100 | i915_gem_wait_for_error(struct i915_gpu_error *error) |
97 | i915_gem_wait_for_error(struct i915_gpu_error *error) |
101 | { |
98 | { |
102 | int ret; |
99 | int ret; |
103 | 100 | ||
104 | #define EXIT_COND (!i915_reset_in_progress(error)) |
101 | #define EXIT_COND (!i915_reset_in_progress(error)) |
105 | if (EXIT_COND) |
102 | if (EXIT_COND) |
106 | return 0; |
103 | return 0; |
107 | #if 0 |
104 | #if 0 |
108 | /* |
105 | /* |
109 | * Only wait 10 seconds for the gpu reset to complete to avoid hanging |
106 | * Only wait 10 seconds for the gpu reset to complete to avoid hanging |
110 | * userspace. If it takes that long something really bad is going on and |
107 | * userspace. If it takes that long something really bad is going on and |
111 | * we should simply try to bail out and fail as gracefully as possible. |
108 | * we should simply try to bail out and fail as gracefully as possible. |
112 | */ |
109 | */ |
113 | ret = wait_event_interruptible_timeout(error->reset_queue, |
110 | ret = wait_event_interruptible_timeout(error->reset_queue, |
114 | EXIT_COND, |
111 | EXIT_COND, |
115 | 10*HZ); |
112 | 10*HZ); |
116 | if (ret == 0) { |
113 | if (ret == 0) { |
117 | DRM_ERROR("Timed out waiting for the gpu reset to complete\n"); |
114 | DRM_ERROR("Timed out waiting for the gpu reset to complete\n"); |
118 | return -EIO; |
115 | return -EIO; |
119 | } else if (ret < 0) { |
116 | } else if (ret < 0) { |
120 | return ret; |
117 | return ret; |
121 | } |
118 | } |
122 | 119 | ||
123 | #endif |
120 | #endif |
124 | #undef EXIT_COND |
121 | #undef EXIT_COND |
125 | 122 | ||
126 | return 0; |
123 | return 0; |
127 | } |
124 | } |
128 | 125 | ||
129 | int i915_mutex_lock_interruptible(struct drm_device *dev) |
126 | int i915_mutex_lock_interruptible(struct drm_device *dev) |
130 | { |
127 | { |
131 | struct drm_i915_private *dev_priv = dev->dev_private; |
128 | struct drm_i915_private *dev_priv = dev->dev_private; |
132 | int ret; |
129 | int ret; |
133 | 130 | ||
134 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
131 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
135 | if (ret) |
132 | if (ret) |
136 | return ret; |
133 | return ret; |
137 | 134 | ||
138 | ret = mutex_lock_interruptible(&dev->struct_mutex); |
135 | ret = mutex_lock_interruptible(&dev->struct_mutex); |
139 | if (ret) |
136 | if (ret) |
140 | return ret; |
137 | return ret; |
141 | 138 | ||
142 | WARN_ON(i915_verify_lists(dev)); |
139 | WARN_ON(i915_verify_lists(dev)); |
143 | return 0; |
140 | return 0; |
144 | } |
141 | } |
145 | 142 | ||
146 | int |
143 | int |
147 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, |
144 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, |
148 | struct drm_file *file) |
145 | struct drm_file *file) |
149 | { |
146 | { |
150 | struct drm_i915_private *dev_priv = dev->dev_private; |
147 | struct drm_i915_private *dev_priv = dev->dev_private; |
151 | struct drm_i915_gem_get_aperture *args = data; |
148 | struct drm_i915_gem_get_aperture *args = data; |
152 | struct i915_gtt *ggtt = &dev_priv->gtt; |
149 | struct i915_gtt *ggtt = &dev_priv->gtt; |
153 | struct i915_vma *vma; |
150 | struct i915_vma *vma; |
154 | size_t pinned; |
151 | size_t pinned; |
155 | 152 | ||
156 | pinned = 0; |
153 | pinned = 0; |
157 | mutex_lock(&dev->struct_mutex); |
154 | mutex_lock(&dev->struct_mutex); |
158 | list_for_each_entry(vma, &ggtt->base.active_list, mm_list) |
155 | list_for_each_entry(vma, &ggtt->base.active_list, vm_link) |
159 | if (vma->pin_count) |
156 | if (vma->pin_count) |
160 | pinned += vma->node.size; |
157 | pinned += vma->node.size; |
161 | list_for_each_entry(vma, &ggtt->base.inactive_list, mm_list) |
158 | list_for_each_entry(vma, &ggtt->base.inactive_list, vm_link) |
162 | if (vma->pin_count) |
159 | if (vma->pin_count) |
163 | pinned += vma->node.size; |
160 | pinned += vma->node.size; |
164 | mutex_unlock(&dev->struct_mutex); |
161 | mutex_unlock(&dev->struct_mutex); |
165 | 162 | ||
166 | args->aper_size = dev_priv->gtt.base.total; |
163 | args->aper_size = dev_priv->gtt.base.total; |
167 | args->aper_available_size = args->aper_size - pinned; |
164 | args->aper_available_size = args->aper_size - pinned; |
168 | 165 | ||
169 | return 0; |
166 | return 0; |
170 | } |
167 | } |
171 | 168 | ||
172 | static int |
169 | static int |
173 | i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj) |
170 | i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj) |
174 | { |
171 | { |
175 | char *vaddr = obj->phys_handle->vaddr; |
172 | char *vaddr = obj->phys_handle->vaddr; |
176 | struct sg_table *st; |
173 | struct sg_table *st; |
177 | struct scatterlist *sg; |
174 | struct scatterlist *sg; |
178 | int i; |
175 | int i; |
179 | 176 | ||
180 | if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj))) |
177 | if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj))) |
181 | return -EINVAL; |
178 | return -EINVAL; |
182 | 179 | ||
183 | 180 | ||
184 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
181 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
185 | if (st == NULL) |
182 | if (st == NULL) |
186 | return -ENOMEM; |
183 | return -ENOMEM; |
187 | 184 | ||
188 | if (sg_alloc_table(st, 1, GFP_KERNEL)) { |
185 | if (sg_alloc_table(st, 1, GFP_KERNEL)) { |
189 | kfree(st); |
186 | kfree(st); |
190 | return -ENOMEM; |
187 | return -ENOMEM; |
191 | } |
188 | } |
192 | 189 | ||
193 | sg = st->sgl; |
190 | sg = st->sgl; |
194 | sg->offset = 0; |
191 | sg->offset = 0; |
195 | sg->length = obj->base.size; |
192 | sg->length = obj->base.size; |
196 | 193 | ||
197 | sg_dma_address(sg) = obj->phys_handle->busaddr; |
194 | sg_dma_address(sg) = obj->phys_handle->busaddr; |
198 | sg_dma_len(sg) = obj->base.size; |
195 | sg_dma_len(sg) = obj->base.size; |
199 | 196 | ||
200 | obj->pages = st; |
197 | obj->pages = st; |
201 | return 0; |
198 | return 0; |
202 | } |
199 | } |
203 | 200 | ||
204 | static void |
201 | static void |
205 | i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj) |
202 | i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj) |
206 | { |
203 | { |
207 | int ret; |
204 | int ret; |
208 | 205 | ||
209 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
206 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
210 | 207 | ||
211 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
208 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
212 | if (ret) { |
209 | if (ret) { |
213 | /* In the event of a disaster, abandon all caches and |
210 | /* In the event of a disaster, abandon all caches and |
214 | * hope for the best. |
211 | * hope for the best. |
215 | */ |
212 | */ |
216 | WARN_ON(ret != -EIO); |
213 | WARN_ON(ret != -EIO); |
217 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
214 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
218 | } |
215 | } |
219 | 216 | ||
220 | if (obj->madv == I915_MADV_DONTNEED) |
217 | if (obj->madv == I915_MADV_DONTNEED) |
221 | obj->dirty = 0; |
218 | obj->dirty = 0; |
222 | 219 | ||
223 | if (obj->dirty) { |
220 | if (obj->dirty) { |
224 | obj->dirty = 0; |
221 | obj->dirty = 0; |
225 | } |
222 | } |
226 | 223 | ||
227 | sg_free_table(obj->pages); |
224 | sg_free_table(obj->pages); |
228 | kfree(obj->pages); |
225 | kfree(obj->pages); |
229 | } |
226 | } |
230 | 227 | ||
231 | static void |
228 | static void |
232 | i915_gem_object_release_phys(struct drm_i915_gem_object *obj) |
229 | i915_gem_object_release_phys(struct drm_i915_gem_object *obj) |
233 | { |
230 | { |
234 | drm_pci_free(obj->base.dev, obj->phys_handle); |
231 | drm_pci_free(obj->base.dev, obj->phys_handle); |
235 | } |
232 | } |
236 | 233 | ||
237 | static const struct drm_i915_gem_object_ops i915_gem_phys_ops = { |
234 | static const struct drm_i915_gem_object_ops i915_gem_phys_ops = { |
238 | .get_pages = i915_gem_object_get_pages_phys, |
235 | .get_pages = i915_gem_object_get_pages_phys, |
239 | .put_pages = i915_gem_object_put_pages_phys, |
236 | .put_pages = i915_gem_object_put_pages_phys, |
240 | .release = i915_gem_object_release_phys, |
237 | .release = i915_gem_object_release_phys, |
241 | }; |
238 | }; |
242 | 239 | ||
243 | static int |
240 | static int |
244 | drop_pages(struct drm_i915_gem_object *obj) |
241 | drop_pages(struct drm_i915_gem_object *obj) |
245 | { |
242 | { |
246 | struct i915_vma *vma, *next; |
243 | struct i915_vma *vma, *next; |
247 | int ret; |
244 | int ret; |
248 | 245 | ||
249 | drm_gem_object_reference(&obj->base); |
246 | drm_gem_object_reference(&obj->base); |
250 | list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) |
247 | list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link) |
251 | if (i915_vma_unbind(vma)) |
248 | if (i915_vma_unbind(vma)) |
252 | break; |
249 | break; |
253 | 250 | ||
254 | ret = i915_gem_object_put_pages(obj); |
251 | ret = i915_gem_object_put_pages(obj); |
255 | drm_gem_object_unreference(&obj->base); |
252 | drm_gem_object_unreference(&obj->base); |
256 | 253 | ||
257 | return ret; |
254 | return ret; |
258 | } |
255 | } |
259 | 256 | ||
260 | int |
257 | int |
261 | i915_gem_object_attach_phys(struct drm_i915_gem_object *obj, |
258 | i915_gem_object_attach_phys(struct drm_i915_gem_object *obj, |
262 | int align) |
259 | int align) |
263 | { |
260 | { |
264 | drm_dma_handle_t *phys; |
261 | drm_dma_handle_t *phys; |
265 | int ret; |
262 | int ret; |
266 | 263 | ||
267 | if (obj->phys_handle) { |
264 | if (obj->phys_handle) { |
268 | if ((unsigned long)obj->phys_handle->vaddr & (align -1)) |
265 | if ((unsigned long)obj->phys_handle->vaddr & (align -1)) |
269 | return -EBUSY; |
266 | return -EBUSY; |
270 | 267 | ||
271 | return 0; |
268 | return 0; |
272 | } |
269 | } |
273 | 270 | ||
274 | if (obj->madv != I915_MADV_WILLNEED) |
271 | if (obj->madv != I915_MADV_WILLNEED) |
275 | return -EFAULT; |
272 | return -EFAULT; |
276 | 273 | ||
277 | if (obj->base.filp == NULL) |
274 | if (obj->base.filp == NULL) |
278 | return -EINVAL; |
275 | return -EINVAL; |
279 | 276 | ||
280 | ret = drop_pages(obj); |
277 | ret = drop_pages(obj); |
281 | if (ret) |
278 | if (ret) |
282 | return ret; |
279 | return ret; |
283 | 280 | ||
284 | /* create a new object */ |
281 | /* create a new object */ |
285 | phys = drm_pci_alloc(obj->base.dev, obj->base.size, align); |
282 | phys = drm_pci_alloc(obj->base.dev, obj->base.size, align); |
286 | if (!phys) |
283 | if (!phys) |
287 | return -ENOMEM; |
284 | return -ENOMEM; |
288 | 285 | ||
289 | obj->phys_handle = phys; |
286 | obj->phys_handle = phys; |
290 | obj->ops = &i915_gem_phys_ops; |
287 | obj->ops = &i915_gem_phys_ops; |
291 | 288 | ||
292 | return i915_gem_object_get_pages(obj); |
289 | return i915_gem_object_get_pages(obj); |
293 | } |
290 | } |
294 | void *i915_gem_object_alloc(struct drm_device *dev) |
291 | void *i915_gem_object_alloc(struct drm_device *dev) |
295 | { |
292 | { |
296 | struct drm_i915_private *dev_priv = dev->dev_private; |
293 | struct drm_i915_private *dev_priv = dev->dev_private; |
297 | return kzalloc(sizeof(struct drm_i915_gem_object), 0); |
294 | return kzalloc(sizeof(struct drm_i915_gem_object), 0); |
298 | } |
295 | } |
299 | 296 | ||
300 | void i915_gem_object_free(struct drm_i915_gem_object *obj) |
297 | void i915_gem_object_free(struct drm_i915_gem_object *obj) |
301 | { |
298 | { |
302 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
299 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
303 | kfree(obj); |
300 | kfree(obj); |
304 | } |
301 | } |
305 | 302 | ||
306 | static int |
303 | static int |
307 | i915_gem_create(struct drm_file *file, |
304 | i915_gem_create(struct drm_file *file, |
308 | struct drm_device *dev, |
305 | struct drm_device *dev, |
309 | uint64_t size, |
306 | uint64_t size, |
310 | uint32_t *handle_p) |
307 | uint32_t *handle_p) |
311 | { |
308 | { |
312 | struct drm_i915_gem_object *obj; |
309 | struct drm_i915_gem_object *obj; |
313 | int ret; |
310 | int ret; |
314 | u32 handle; |
311 | u32 handle; |
315 | 312 | ||
316 | size = roundup(size, PAGE_SIZE); |
313 | size = roundup(size, PAGE_SIZE); |
317 | if (size == 0) |
314 | if (size == 0) |
318 | return -EINVAL; |
315 | return -EINVAL; |
319 | 316 | ||
320 | /* Allocate the new object */ |
317 | /* Allocate the new object */ |
321 | obj = i915_gem_alloc_object(dev, size); |
318 | obj = i915_gem_alloc_object(dev, size); |
322 | if (obj == NULL) |
319 | if (obj == NULL) |
323 | return -ENOMEM; |
320 | return -ENOMEM; |
324 | 321 | ||
325 | ret = drm_gem_handle_create(file, &obj->base, &handle); |
322 | ret = drm_gem_handle_create(file, &obj->base, &handle); |
326 | /* drop reference from allocate - handle holds it now */ |
323 | /* drop reference from allocate - handle holds it now */ |
327 | drm_gem_object_unreference_unlocked(&obj->base); |
324 | drm_gem_object_unreference_unlocked(&obj->base); |
328 | if (ret) |
325 | if (ret) |
329 | return ret; |
326 | return ret; |
330 | 327 | ||
331 | *handle_p = handle; |
328 | *handle_p = handle; |
332 | return 0; |
329 | return 0; |
333 | } |
330 | } |
334 | 331 | ||
335 | int |
332 | int |
336 | i915_gem_dumb_create(struct drm_file *file, |
333 | i915_gem_dumb_create(struct drm_file *file, |
337 | struct drm_device *dev, |
334 | struct drm_device *dev, |
338 | struct drm_mode_create_dumb *args) |
335 | struct drm_mode_create_dumb *args) |
339 | { |
336 | { |
340 | /* have to work out size/pitch and return them */ |
337 | /* have to work out size/pitch and return them */ |
341 | args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64); |
338 | args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64); |
342 | args->size = args->pitch * args->height; |
339 | args->size = args->pitch * args->height; |
343 | return i915_gem_create(file, dev, |
340 | return i915_gem_create(file, dev, |
344 | args->size, &args->handle); |
341 | args->size, &args->handle); |
345 | } |
342 | } |
346 | 343 | ||
347 | /** |
344 | /** |
348 | * Creates a new mm object and returns a handle to it. |
345 | * Creates a new mm object and returns a handle to it. |
349 | */ |
346 | */ |
350 | int |
347 | int |
351 | i915_gem_create_ioctl(struct drm_device *dev, void *data, |
348 | i915_gem_create_ioctl(struct drm_device *dev, void *data, |
352 | struct drm_file *file) |
349 | struct drm_file *file) |
353 | { |
350 | { |
354 | struct drm_i915_gem_create *args = data; |
351 | struct drm_i915_gem_create *args = data; |
355 | 352 | ||
356 | return i915_gem_create(file, dev, |
353 | return i915_gem_create(file, dev, |
357 | args->size, &args->handle); |
354 | args->size, &args->handle); |
358 | } |
355 | } |
359 | 356 | ||
360 | static inline int |
357 | static inline int |
361 | __copy_to_user_swizzled(char __user *cpu_vaddr, |
358 | __copy_to_user_swizzled(char __user *cpu_vaddr, |
362 | const char *gpu_vaddr, int gpu_offset, |
359 | const char *gpu_vaddr, int gpu_offset, |
363 | int length) |
360 | int length) |
364 | { |
361 | { |
365 | int ret, cpu_offset = 0; |
362 | int ret, cpu_offset = 0; |
366 | 363 | ||
367 | while (length > 0) { |
364 | while (length > 0) { |
368 | int cacheline_end = ALIGN(gpu_offset + 1, 64); |
365 | int cacheline_end = ALIGN(gpu_offset + 1, 64); |
369 | int this_length = min(cacheline_end - gpu_offset, length); |
366 | int this_length = min(cacheline_end - gpu_offset, length); |
370 | int swizzled_gpu_offset = gpu_offset ^ 64; |
367 | int swizzled_gpu_offset = gpu_offset ^ 64; |
371 | 368 | ||
372 | ret = __copy_to_user(cpu_vaddr + cpu_offset, |
369 | ret = __copy_to_user(cpu_vaddr + cpu_offset, |
373 | gpu_vaddr + swizzled_gpu_offset, |
370 | gpu_vaddr + swizzled_gpu_offset, |
374 | this_length); |
371 | this_length); |
375 | if (ret) |
372 | if (ret) |
376 | return ret + length; |
373 | return ret + length; |
377 | 374 | ||
378 | cpu_offset += this_length; |
375 | cpu_offset += this_length; |
379 | gpu_offset += this_length; |
376 | gpu_offset += this_length; |
380 | length -= this_length; |
377 | length -= this_length; |
381 | } |
378 | } |
382 | 379 | ||
383 | return 0; |
380 | return 0; |
384 | } |
381 | } |
385 | 382 | ||
386 | static inline int |
383 | static inline int |
387 | __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, |
384 | __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, |
388 | const char __user *cpu_vaddr, |
385 | const char __user *cpu_vaddr, |
389 | int length) |
386 | int length) |
390 | { |
387 | { |
391 | int ret, cpu_offset = 0; |
388 | int ret, cpu_offset = 0; |
392 | 389 | ||
393 | while (length > 0) { |
390 | while (length > 0) { |
394 | int cacheline_end = ALIGN(gpu_offset + 1, 64); |
391 | int cacheline_end = ALIGN(gpu_offset + 1, 64); |
395 | int this_length = min(cacheline_end - gpu_offset, length); |
392 | int this_length = min(cacheline_end - gpu_offset, length); |
396 | int swizzled_gpu_offset = gpu_offset ^ 64; |
393 | int swizzled_gpu_offset = gpu_offset ^ 64; |
397 | 394 | ||
398 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, |
395 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, |
399 | cpu_vaddr + cpu_offset, |
396 | cpu_vaddr + cpu_offset, |
400 | this_length); |
397 | this_length); |
401 | if (ret) |
398 | if (ret) |
402 | return ret + length; |
399 | return ret + length; |
403 | 400 | ||
404 | cpu_offset += this_length; |
401 | cpu_offset += this_length; |
405 | gpu_offset += this_length; |
402 | gpu_offset += this_length; |
406 | length -= this_length; |
403 | length -= this_length; |
407 | } |
404 | } |
408 | 405 | ||
409 | return 0; |
406 | return 0; |
410 | } |
407 | } |
411 | 408 | ||
412 | /* |
409 | /* |
413 | * Pins the specified object's pages and synchronizes the object with |
410 | * Pins the specified object's pages and synchronizes the object with |
414 | * GPU accesses. Sets needs_clflush to non-zero if the caller should |
411 | * GPU accesses. Sets needs_clflush to non-zero if the caller should |
415 | * flush the object from the CPU cache. |
412 | * flush the object from the CPU cache. |
416 | */ |
413 | */ |
417 | int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj, |
414 | int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj, |
418 | int *needs_clflush) |
415 | int *needs_clflush) |
419 | { |
416 | { |
420 | int ret; |
417 | int ret; |
421 | 418 | ||
422 | *needs_clflush = 0; |
419 | *needs_clflush = 0; |
423 | 420 | ||
424 | if (!obj->base.filp) |
421 | if (!obj->base.filp) |
425 | return -EINVAL; |
422 | return -EINVAL; |
426 | 423 | ||
427 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) { |
424 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) { |
428 | /* If we're not in the cpu read domain, set ourself into the gtt |
425 | /* If we're not in the cpu read domain, set ourself into the gtt |
429 | * read domain and manually flush cachelines (if required). This |
426 | * read domain and manually flush cachelines (if required). This |
430 | * optimizes for the case when the gpu will dirty the data |
427 | * optimizes for the case when the gpu will dirty the data |
431 | * anyway again before the next pread happens. */ |
428 | * anyway again before the next pread happens. */ |
432 | *needs_clflush = !cpu_cache_is_coherent(obj->base.dev, |
429 | *needs_clflush = !cpu_cache_is_coherent(obj->base.dev, |
433 | obj->cache_level); |
430 | obj->cache_level); |
434 | ret = i915_gem_object_wait_rendering(obj, true); |
431 | ret = i915_gem_object_wait_rendering(obj, true); |
435 | if (ret) |
432 | if (ret) |
436 | return ret; |
433 | return ret; |
437 | } |
434 | } |
438 | 435 | ||
439 | ret = i915_gem_object_get_pages(obj); |
436 | ret = i915_gem_object_get_pages(obj); |
440 | if (ret) |
437 | if (ret) |
441 | return ret; |
438 | return ret; |
442 | 439 | ||
443 | i915_gem_object_pin_pages(obj); |
440 | i915_gem_object_pin_pages(obj); |
444 | 441 | ||
445 | return ret; |
442 | return ret; |
446 | } |
443 | } |
447 | 444 | ||
448 | /* Per-page copy function for the shmem pread fastpath. |
445 | /* Per-page copy function for the shmem pread fastpath. |
449 | * Flushes invalid cachelines before reading the target if |
446 | * Flushes invalid cachelines before reading the target if |
450 | * needs_clflush is set. */ |
447 | * needs_clflush is set. */ |
451 | static int |
448 | static int |
452 | shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length, |
449 | shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length, |
453 | char __user *user_data, |
450 | char __user *user_data, |
454 | bool page_do_bit17_swizzling, bool needs_clflush) |
451 | bool page_do_bit17_swizzling, bool needs_clflush) |
455 | { |
452 | { |
456 | char *vaddr; |
453 | char *vaddr; |
457 | int ret; |
454 | int ret; |
458 | 455 | ||
459 | if (unlikely(page_do_bit17_swizzling)) |
456 | if (unlikely(page_do_bit17_swizzling)) |
460 | return -EINVAL; |
457 | return -EINVAL; |
461 | 458 | ||
462 | vaddr = kmap_atomic(page); |
459 | vaddr = kmap_atomic(page); |
463 | if (needs_clflush) |
460 | if (needs_clflush) |
464 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
461 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
465 | page_length); |
462 | page_length); |
466 | ret = __copy_to_user_inatomic(user_data, |
463 | ret = __copy_to_user_inatomic(user_data, |
467 | vaddr + shmem_page_offset, |
464 | vaddr + shmem_page_offset, |
468 | page_length); |
465 | page_length); |
469 | kunmap_atomic(vaddr); |
466 | kunmap_atomic(vaddr); |
470 | 467 | ||
471 | return ret ? -EFAULT : 0; |
468 | return ret ? -EFAULT : 0; |
472 | } |
469 | } |
473 | 470 | ||
474 | static void |
471 | static void |
475 | shmem_clflush_swizzled_range(char *addr, unsigned long length, |
472 | shmem_clflush_swizzled_range(char *addr, unsigned long length, |
476 | bool swizzled) |
473 | bool swizzled) |
477 | { |
474 | { |
478 | if (unlikely(swizzled)) { |
475 | if (unlikely(swizzled)) { |
479 | unsigned long start = (unsigned long) addr; |
476 | unsigned long start = (unsigned long) addr; |
480 | unsigned long end = (unsigned long) addr + length; |
477 | unsigned long end = (unsigned long) addr + length; |
481 | 478 | ||
482 | /* For swizzling simply ensure that we always flush both |
479 | /* For swizzling simply ensure that we always flush both |
483 | * channels. Lame, but simple and it works. Swizzled |
480 | * channels. Lame, but simple and it works. Swizzled |
484 | * pwrite/pread is far from a hotpath - current userspace |
481 | * pwrite/pread is far from a hotpath - current userspace |
485 | * doesn't use it at all. */ |
482 | * doesn't use it at all. */ |
486 | start = round_down(start, 128); |
483 | start = round_down(start, 128); |
487 | end = round_up(end, 128); |
484 | end = round_up(end, 128); |
488 | 485 | ||
489 | drm_clflush_virt_range((void *)start, end - start); |
486 | drm_clflush_virt_range((void *)start, end - start); |
490 | } else { |
487 | } else { |
491 | drm_clflush_virt_range(addr, length); |
488 | drm_clflush_virt_range(addr, length); |
492 | } |
489 | } |
493 | 490 | ||
494 | } |
491 | } |
495 | 492 | ||
496 | /* Only difference to the fast-path function is that this can handle bit17 |
493 | /* Only difference to the fast-path function is that this can handle bit17 |
497 | * and uses non-atomic copy and kmap functions. */ |
494 | * and uses non-atomic copy and kmap functions. */ |
498 | static int |
495 | static int |
499 | shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length, |
496 | shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length, |
500 | char __user *user_data, |
497 | char __user *user_data, |
501 | bool page_do_bit17_swizzling, bool needs_clflush) |
498 | bool page_do_bit17_swizzling, bool needs_clflush) |
502 | { |
499 | { |
503 | char *vaddr; |
500 | char *vaddr; |
504 | int ret; |
501 | int ret; |
505 | 502 | ||
506 | vaddr = kmap(page); |
503 | vaddr = kmap(page); |
507 | if (needs_clflush) |
504 | if (needs_clflush) |
508 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
505 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
509 | page_length, |
506 | page_length, |
510 | page_do_bit17_swizzling); |
507 | page_do_bit17_swizzling); |
511 | 508 | ||
512 | if (page_do_bit17_swizzling) |
509 | if (page_do_bit17_swizzling) |
513 | ret = __copy_to_user_swizzled(user_data, |
510 | ret = __copy_to_user_swizzled(user_data, |
514 | vaddr, shmem_page_offset, |
511 | vaddr, shmem_page_offset, |
515 | page_length); |
512 | page_length); |
516 | else |
513 | else |
517 | ret = __copy_to_user(user_data, |
514 | ret = __copy_to_user(user_data, |
518 | vaddr + shmem_page_offset, |
515 | vaddr + shmem_page_offset, |
519 | page_length); |
516 | page_length); |
520 | kunmap(page); |
517 | kunmap(page); |
521 | 518 | ||
522 | return ret ? - EFAULT : 0; |
519 | return ret ? - EFAULT : 0; |
523 | } |
520 | } |
524 | 521 | ||
525 | static int |
522 | static int |
526 | i915_gem_shmem_pread(struct drm_device *dev, |
523 | i915_gem_shmem_pread(struct drm_device *dev, |
527 | struct drm_i915_gem_object *obj, |
524 | struct drm_i915_gem_object *obj, |
528 | struct drm_i915_gem_pread *args, |
525 | struct drm_i915_gem_pread *args, |
529 | struct drm_file *file) |
526 | struct drm_file *file) |
530 | { |
527 | { |
531 | char __user *user_data; |
528 | char __user *user_data; |
532 | ssize_t remain; |
529 | ssize_t remain; |
533 | loff_t offset; |
530 | loff_t offset; |
534 | int shmem_page_offset, page_length, ret = 0; |
531 | int shmem_page_offset, page_length, ret = 0; |
535 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
532 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
536 | int prefaulted = 0; |
533 | int prefaulted = 0; |
537 | int needs_clflush = 0; |
534 | int needs_clflush = 0; |
538 | struct sg_page_iter sg_iter; |
535 | struct sg_page_iter sg_iter; |
539 | 536 | ||
540 | user_data = to_user_ptr(args->data_ptr); |
537 | user_data = to_user_ptr(args->data_ptr); |
541 | remain = args->size; |
538 | remain = args->size; |
542 | 539 | ||
543 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
540 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
544 | 541 | ||
545 | ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush); |
542 | ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush); |
546 | if (ret) |
543 | if (ret) |
547 | return ret; |
544 | return ret; |
548 | 545 | ||
549 | offset = args->offset; |
546 | offset = args->offset; |
550 | 547 | ||
551 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
548 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
552 | offset >> PAGE_SHIFT) { |
549 | offset >> PAGE_SHIFT) { |
553 | struct page *page = sg_page_iter_page(&sg_iter); |
550 | struct page *page = sg_page_iter_page(&sg_iter); |
554 | 551 | ||
555 | if (remain <= 0) |
552 | if (remain <= 0) |
556 | break; |
553 | break; |
557 | 554 | ||
558 | /* Operation in this page |
555 | /* Operation in this page |
559 | * |
556 | * |
560 | * shmem_page_offset = offset within page in shmem file |
557 | * shmem_page_offset = offset within page in shmem file |
561 | * page_length = bytes to copy for this page |
558 | * page_length = bytes to copy for this page |
562 | */ |
559 | */ |
563 | shmem_page_offset = offset_in_page(offset); |
560 | shmem_page_offset = offset_in_page(offset); |
564 | page_length = remain; |
561 | page_length = remain; |
565 | if ((shmem_page_offset + page_length) > PAGE_SIZE) |
562 | if ((shmem_page_offset + page_length) > PAGE_SIZE) |
566 | page_length = PAGE_SIZE - shmem_page_offset; |
563 | page_length = PAGE_SIZE - shmem_page_offset; |
567 | 564 | ||
568 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
565 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
569 | (page_to_phys(page) & (1 << 17)) != 0; |
566 | (page_to_phys(page) & (1 << 17)) != 0; |
570 | 567 | ||
571 | ret = shmem_pread_fast(page, shmem_page_offset, page_length, |
568 | ret = shmem_pread_fast(page, shmem_page_offset, page_length, |
572 | user_data, page_do_bit17_swizzling, |
569 | user_data, page_do_bit17_swizzling, |
573 | needs_clflush); |
570 | needs_clflush); |
574 | if (ret == 0) |
571 | if (ret == 0) |
575 | goto next_page; |
572 | goto next_page; |
576 | 573 | ||
577 | mutex_unlock(&dev->struct_mutex); |
574 | mutex_unlock(&dev->struct_mutex); |
578 | 575 | ||
579 | ret = shmem_pread_slow(page, shmem_page_offset, page_length, |
576 | ret = shmem_pread_slow(page, shmem_page_offset, page_length, |
580 | user_data, page_do_bit17_swizzling, |
577 | user_data, page_do_bit17_swizzling, |
581 | needs_clflush); |
578 | needs_clflush); |
582 | 579 | ||
583 | mutex_lock(&dev->struct_mutex); |
580 | mutex_lock(&dev->struct_mutex); |
584 | 581 | ||
585 | if (ret) |
582 | if (ret) |
586 | goto out; |
583 | goto out; |
587 | 584 | ||
588 | next_page: |
585 | next_page: |
589 | remain -= page_length; |
586 | remain -= page_length; |
590 | user_data += page_length; |
587 | user_data += page_length; |
591 | offset += page_length; |
588 | offset += page_length; |
592 | } |
589 | } |
593 | 590 | ||
594 | out: |
591 | out: |
595 | i915_gem_object_unpin_pages(obj); |
592 | i915_gem_object_unpin_pages(obj); |
596 | 593 | ||
597 | return ret; |
594 | return ret; |
598 | } |
595 | } |
599 | 596 | ||
600 | /** |
597 | /** |
601 | * Reads data from the object referenced by handle. |
598 | * Reads data from the object referenced by handle. |
602 | * |
599 | * |
603 | * On error, the contents of *data are undefined. |
600 | * On error, the contents of *data are undefined. |
604 | */ |
601 | */ |
605 | int |
602 | int |
606 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, |
603 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, |
607 | struct drm_file *file) |
604 | struct drm_file *file) |
608 | { |
605 | { |
609 | struct drm_i915_gem_pread *args = data; |
606 | struct drm_i915_gem_pread *args = data; |
610 | struct drm_i915_gem_object *obj; |
607 | struct drm_i915_gem_object *obj; |
611 | int ret = 0; |
608 | int ret = 0; |
612 | 609 | ||
613 | if (args->size == 0) |
610 | if (args->size == 0) |
614 | return 0; |
611 | return 0; |
615 | 612 | ||
616 | ret = i915_mutex_lock_interruptible(dev); |
613 | ret = i915_mutex_lock_interruptible(dev); |
617 | if (ret) |
614 | if (ret) |
618 | return ret; |
615 | return ret; |
619 | 616 | ||
620 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
617 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
621 | if (&obj->base == NULL) { |
618 | if (&obj->base == NULL) { |
622 | ret = -ENOENT; |
619 | ret = -ENOENT; |
623 | goto unlock; |
620 | goto unlock; |
624 | } |
621 | } |
625 | 622 | ||
626 | /* Bounds check source. */ |
623 | /* Bounds check source. */ |
627 | if (args->offset > obj->base.size || |
624 | if (args->offset > obj->base.size || |
628 | args->size > obj->base.size - args->offset) { |
625 | args->size > obj->base.size - args->offset) { |
629 | ret = -EINVAL; |
626 | ret = -EINVAL; |
630 | goto out; |
627 | goto out; |
631 | } |
628 | } |
632 | 629 | ||
633 | /* prime objects have no backing filp to GEM pread/pwrite |
630 | /* prime objects have no backing filp to GEM pread/pwrite |
634 | * pages from. |
631 | * pages from. |
635 | */ |
632 | */ |
636 | if (!obj->base.filp) { |
633 | if (!obj->base.filp) { |
637 | ret = -EINVAL; |
634 | ret = -EINVAL; |
638 | goto out; |
635 | goto out; |
639 | } |
636 | } |
640 | 637 | ||
641 | trace_i915_gem_object_pread(obj, args->offset, args->size); |
638 | trace_i915_gem_object_pread(obj, args->offset, args->size); |
642 | 639 | ||
643 | ret = i915_gem_shmem_pread(dev, obj, args, file); |
640 | ret = i915_gem_shmem_pread(dev, obj, args, file); |
644 | 641 | ||
645 | out: |
642 | out: |
646 | drm_gem_object_unreference(&obj->base); |
643 | drm_gem_object_unreference(&obj->base); |
647 | unlock: |
644 | unlock: |
648 | mutex_unlock(&dev->struct_mutex); |
645 | mutex_unlock(&dev->struct_mutex); |
649 | return ret; |
646 | return ret; |
650 | } |
647 | } |
651 | 648 | ||
652 | /* This is the fast write path which cannot handle |
649 | /* This is the fast write path which cannot handle |
653 | * page faults in the source data |
650 | * page faults in the source data |
654 | */ |
651 | */ |
- | 652 | ||
- | 653 | static inline int |
|
- | 654 | fast_user_write(struct io_mapping *mapping, |
|
- | 655 | loff_t page_base, int page_offset, |
|
- | 656 | char __user *user_data, |
|
- | 657 | int length) |
|
- | 658 | { |
|
- | 659 | void __iomem *vaddr_atomic; |
|
- | 660 | void *vaddr; |
|
- | 661 | unsigned long unwritten; |
|
- | 662 | ||
- | 663 | vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base); |
|
- | 664 | /* We can use the cpu mem copy function because this is X86. */ |
|
- | 665 | vaddr = (void __force*)vaddr_atomic + page_offset; |
|
- | 666 | unwritten = __copy_from_user_inatomic_nocache(vaddr, |
|
- | 667 | user_data, length); |
|
- | 668 | io_mapping_unmap_atomic(vaddr_atomic); |
|
- | 669 | return unwritten; |
|
655 | 670 | } |
|
656 | 671 | ||
657 | /** |
672 | /** |
658 | * This is the fast pwrite path, where we copy the data directly from the |
673 | * This is the fast pwrite path, where we copy the data directly from the |
659 | * user into the GTT, uncached. |
674 | * user into the GTT, uncached. |
660 | */ |
675 | */ |
661 | static int |
676 | static int |
662 | i915_gem_gtt_pwrite_fast(struct drm_device *dev, |
677 | i915_gem_gtt_pwrite_fast(struct drm_device *dev, |
663 | struct drm_i915_gem_object *obj, |
678 | struct drm_i915_gem_object *obj, |
664 | struct drm_i915_gem_pwrite *args, |
679 | struct drm_i915_gem_pwrite *args, |
665 | struct drm_file *file) |
680 | struct drm_file *file) |
666 | { |
681 | { |
667 | struct drm_i915_private *dev_priv = dev->dev_private; |
682 | struct drm_i915_private *dev_priv = dev->dev_private; |
668 | ssize_t remain; |
683 | ssize_t remain; |
669 | loff_t offset, page_base; |
684 | loff_t offset, page_base; |
670 | char __user *user_data; |
685 | char __user *user_data; |
671 | int page_offset, page_length, ret; |
686 | int page_offset, page_length, ret; |
672 | 687 | ||
673 | ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK); |
688 | ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK); |
674 | if (ret) |
689 | if (ret) |
675 | goto out; |
690 | goto out; |
676 | 691 | ||
677 | ret = i915_gem_object_set_to_gtt_domain(obj, true); |
692 | ret = i915_gem_object_set_to_gtt_domain(obj, true); |
678 | if (ret) |
693 | if (ret) |
679 | goto out_unpin; |
694 | goto out_unpin; |
680 | 695 | ||
681 | ret = i915_gem_object_put_fence(obj); |
696 | ret = i915_gem_object_put_fence(obj); |
682 | if (ret) |
697 | if (ret) |
683 | goto out_unpin; |
698 | goto out_unpin; |
684 | 699 | ||
685 | user_data = to_user_ptr(args->data_ptr); |
700 | user_data = to_user_ptr(args->data_ptr); |
686 | remain = args->size; |
701 | remain = args->size; |
687 | 702 | ||
688 | offset = i915_gem_obj_ggtt_offset(obj) + args->offset; |
703 | offset = i915_gem_obj_ggtt_offset(obj) + args->offset; |
689 | 704 | ||
690 | intel_fb_obj_invalidate(obj, ORIGIN_GTT); |
705 | intel_fb_obj_invalidate(obj, ORIGIN_GTT); |
691 | 706 | ||
692 | while (remain > 0) { |
707 | while (remain > 0) { |
693 | /* Operation in this page |
708 | /* Operation in this page |
694 | * |
709 | * |
695 | * page_base = page offset within aperture |
710 | * page_base = page offset within aperture |
696 | * page_offset = offset within page |
711 | * page_offset = offset within page |
697 | * page_length = bytes to copy for this page |
712 | * page_length = bytes to copy for this page |
698 | */ |
713 | */ |
699 | page_base = offset & PAGE_MASK; |
714 | page_base = offset & PAGE_MASK; |
700 | page_offset = offset_in_page(offset); |
715 | page_offset = offset_in_page(offset); |
701 | page_length = remain; |
716 | page_length = remain; |
702 | if ((page_offset + remain) > PAGE_SIZE) |
717 | if ((page_offset + remain) > PAGE_SIZE) |
703 | page_length = PAGE_SIZE - page_offset; |
718 | page_length = PAGE_SIZE - page_offset; |
704 | 719 | ||
705 | MapPage(dev_priv->gtt.mappable, |
720 | /* If we get a fault while copying data, then (presumably) our |
- | 721 | * source page isn't available. Return the error and we'll |
|
706 | dev_priv->gtt.mappable_base+page_base, PG_WRITEC|PG_SW); |
722 | * retry in the slow path. |
- | 723 | */ |
|
707 | 724 | if (fast_user_write(dev_priv->gtt.mappable, page_base, |
|
- | 725 | page_offset, user_data, page_length)) { |
|
- | 726 | ret = -EFAULT; |
|
- | 727 | goto out_flush; |
|
708 | memcpy((char*)dev_priv->gtt.mappable+page_offset, user_data, page_length); |
728 | } |
709 | 729 | ||
710 | remain -= page_length; |
730 | remain -= page_length; |
711 | user_data += page_length; |
731 | user_data += page_length; |
712 | offset += page_length; |
732 | offset += page_length; |
713 | } |
733 | } |
714 | 734 | ||
715 | out_flush: |
735 | out_flush: |
716 | intel_fb_obj_flush(obj, false, ORIGIN_GTT); |
736 | intel_fb_obj_flush(obj, false, ORIGIN_GTT); |
717 | out_unpin: |
737 | out_unpin: |
718 | i915_gem_object_ggtt_unpin(obj); |
738 | i915_gem_object_ggtt_unpin(obj); |
719 | out: |
739 | out: |
720 | return ret; |
740 | return ret; |
721 | } |
741 | } |
722 | 742 | ||
723 | /* Per-page copy function for the shmem pwrite fastpath. |
743 | /* Per-page copy function for the shmem pwrite fastpath. |
724 | * Flushes invalid cachelines before writing to the target if |
744 | * Flushes invalid cachelines before writing to the target if |
725 | * needs_clflush_before is set and flushes out any written cachelines after |
745 | * needs_clflush_before is set and flushes out any written cachelines after |
726 | * writing if needs_clflush is set. */ |
746 | * writing if needs_clflush is set. */ |
727 | static int |
747 | static int |
728 | shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length, |
748 | shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length, |
729 | char __user *user_data, |
749 | char __user *user_data, |
730 | bool page_do_bit17_swizzling, |
750 | bool page_do_bit17_swizzling, |
731 | bool needs_clflush_before, |
751 | bool needs_clflush_before, |
732 | bool needs_clflush_after) |
752 | bool needs_clflush_after) |
733 | { |
753 | { |
734 | char *vaddr; |
754 | char *vaddr; |
735 | int ret; |
755 | int ret; |
736 | 756 | ||
737 | if (unlikely(page_do_bit17_swizzling)) |
757 | if (unlikely(page_do_bit17_swizzling)) |
738 | return -EINVAL; |
758 | return -EINVAL; |
739 | 759 | ||
740 | vaddr = kmap_atomic(page); |
760 | vaddr = kmap_atomic(page); |
741 | if (needs_clflush_before) |
761 | if (needs_clflush_before) |
742 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
762 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
743 | page_length); |
763 | page_length); |
744 | memcpy(vaddr + shmem_page_offset, |
764 | ret = __copy_from_user_inatomic(vaddr + shmem_page_offset, |
745 | user_data, |
- | |
746 | page_length); |
765 | user_data, page_length); |
747 | if (needs_clflush_after) |
766 | if (needs_clflush_after) |
748 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
767 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
749 | page_length); |
768 | page_length); |
750 | kunmap_atomic(vaddr); |
769 | kunmap_atomic(vaddr); |
751 | 770 | ||
752 | return ret ? -EFAULT : 0; |
771 | return ret ? -EFAULT : 0; |
753 | } |
772 | } |
754 | 773 | ||
755 | /* Only difference to the fast-path function is that this can handle bit17 |
774 | /* Only difference to the fast-path function is that this can handle bit17 |
756 | * and uses non-atomic copy and kmap functions. */ |
775 | * and uses non-atomic copy and kmap functions. */ |
757 | static int |
776 | static int |
758 | shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length, |
777 | shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length, |
759 | char __user *user_data, |
778 | char __user *user_data, |
760 | bool page_do_bit17_swizzling, |
779 | bool page_do_bit17_swizzling, |
761 | bool needs_clflush_before, |
780 | bool needs_clflush_before, |
762 | bool needs_clflush_after) |
781 | bool needs_clflush_after) |
763 | { |
782 | { |
764 | char *vaddr; |
783 | char *vaddr; |
765 | int ret; |
784 | int ret; |
766 | 785 | ||
767 | vaddr = kmap(page); |
786 | vaddr = kmap(page); |
768 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) |
787 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) |
769 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
788 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
770 | page_length, |
789 | page_length, |
771 | page_do_bit17_swizzling); |
790 | page_do_bit17_swizzling); |
772 | if (page_do_bit17_swizzling) |
791 | if (page_do_bit17_swizzling) |
773 | ret = __copy_from_user_swizzled(vaddr, shmem_page_offset, |
792 | ret = __copy_from_user_swizzled(vaddr, shmem_page_offset, |
774 | user_data, |
793 | user_data, |
775 | page_length); |
794 | page_length); |
776 | else |
795 | else |
777 | ret = __copy_from_user(vaddr + shmem_page_offset, |
796 | ret = __copy_from_user(vaddr + shmem_page_offset, |
778 | user_data, |
797 | user_data, |
779 | page_length); |
798 | page_length); |
780 | if (needs_clflush_after) |
799 | if (needs_clflush_after) |
781 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
800 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
782 | page_length, |
801 | page_length, |
783 | page_do_bit17_swizzling); |
802 | page_do_bit17_swizzling); |
784 | kunmap(page); |
803 | kunmap(page); |
785 | 804 | ||
786 | return ret ? -EFAULT : 0; |
805 | return ret ? -EFAULT : 0; |
787 | } |
806 | } |
788 | 807 | ||
789 | static int |
808 | static int |
790 | i915_gem_shmem_pwrite(struct drm_device *dev, |
809 | i915_gem_shmem_pwrite(struct drm_device *dev, |
791 | struct drm_i915_gem_object *obj, |
810 | struct drm_i915_gem_object *obj, |
792 | struct drm_i915_gem_pwrite *args, |
811 | struct drm_i915_gem_pwrite *args, |
793 | struct drm_file *file) |
812 | struct drm_file *file) |
794 | { |
813 | { |
795 | ssize_t remain; |
814 | ssize_t remain; |
796 | loff_t offset; |
815 | loff_t offset; |
797 | char __user *user_data; |
816 | char __user *user_data; |
798 | int shmem_page_offset, page_length, ret = 0; |
817 | int shmem_page_offset, page_length, ret = 0; |
799 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
818 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
800 | int hit_slowpath = 0; |
819 | int hit_slowpath = 0; |
801 | int needs_clflush_after = 0; |
820 | int needs_clflush_after = 0; |
802 | int needs_clflush_before = 0; |
821 | int needs_clflush_before = 0; |
803 | struct sg_page_iter sg_iter; |
822 | struct sg_page_iter sg_iter; |
804 | 823 | ||
805 | user_data = to_user_ptr(args->data_ptr); |
824 | user_data = to_user_ptr(args->data_ptr); |
806 | remain = args->size; |
825 | remain = args->size; |
807 | 826 | ||
808 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
827 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
809 | 828 | ||
810 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
829 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
811 | /* If we're not in the cpu write domain, set ourself into the gtt |
830 | /* If we're not in the cpu write domain, set ourself into the gtt |
812 | * write domain and manually flush cachelines (if required). This |
831 | * write domain and manually flush cachelines (if required). This |
813 | * optimizes for the case when the gpu will use the data |
832 | * optimizes for the case when the gpu will use the data |
814 | * right away and we therefore have to clflush anyway. */ |
833 | * right away and we therefore have to clflush anyway. */ |
815 | needs_clflush_after = cpu_write_needs_clflush(obj); |
834 | needs_clflush_after = cpu_write_needs_clflush(obj); |
816 | ret = i915_gem_object_wait_rendering(obj, false); |
835 | ret = i915_gem_object_wait_rendering(obj, false); |
817 | if (ret) |
836 | if (ret) |
818 | return ret; |
837 | return ret; |
819 | } |
838 | } |
820 | /* Same trick applies to invalidate partially written cachelines read |
839 | /* Same trick applies to invalidate partially written cachelines read |
821 | * before writing. */ |
840 | * before writing. */ |
822 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) |
841 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) |
823 | needs_clflush_before = |
842 | needs_clflush_before = |
824 | !cpu_cache_is_coherent(dev, obj->cache_level); |
843 | !cpu_cache_is_coherent(dev, obj->cache_level); |
825 | 844 | ||
826 | ret = i915_gem_object_get_pages(obj); |
845 | ret = i915_gem_object_get_pages(obj); |
827 | if (ret) |
846 | if (ret) |
828 | return ret; |
847 | return ret; |
829 | 848 | ||
830 | intel_fb_obj_invalidate(obj, ORIGIN_CPU); |
849 | intel_fb_obj_invalidate(obj, ORIGIN_CPU); |
831 | 850 | ||
832 | i915_gem_object_pin_pages(obj); |
851 | i915_gem_object_pin_pages(obj); |
833 | 852 | ||
834 | offset = args->offset; |
853 | offset = args->offset; |
835 | obj->dirty = 1; |
854 | obj->dirty = 1; |
836 | 855 | ||
837 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
856 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
838 | offset >> PAGE_SHIFT) { |
857 | offset >> PAGE_SHIFT) { |
839 | struct page *page = sg_page_iter_page(&sg_iter); |
858 | struct page *page = sg_page_iter_page(&sg_iter); |
840 | int partial_cacheline_write; |
859 | int partial_cacheline_write; |
841 | 860 | ||
842 | if (remain <= 0) |
861 | if (remain <= 0) |
843 | break; |
862 | break; |
844 | 863 | ||
845 | /* Operation in this page |
864 | /* Operation in this page |
846 | * |
865 | * |
847 | * shmem_page_offset = offset within page in shmem file |
866 | * shmem_page_offset = offset within page in shmem file |
848 | * page_length = bytes to copy for this page |
867 | * page_length = bytes to copy for this page |
849 | */ |
868 | */ |
850 | shmem_page_offset = offset_in_page(offset); |
869 | shmem_page_offset = offset_in_page(offset); |
851 | 870 | ||
852 | page_length = remain; |
871 | page_length = remain; |
853 | if ((shmem_page_offset + page_length) > PAGE_SIZE) |
872 | if ((shmem_page_offset + page_length) > PAGE_SIZE) |
854 | page_length = PAGE_SIZE - shmem_page_offset; |
873 | page_length = PAGE_SIZE - shmem_page_offset; |
855 | 874 | ||
856 | /* If we don't overwrite a cacheline completely we need to be |
875 | /* If we don't overwrite a cacheline completely we need to be |
857 | * careful to have up-to-date data by first clflushing. Don't |
876 | * careful to have up-to-date data by first clflushing. Don't |
858 | * overcomplicate things and flush the entire patch. */ |
877 | * overcomplicate things and flush the entire patch. */ |
859 | partial_cacheline_write = needs_clflush_before && |
878 | partial_cacheline_write = needs_clflush_before && |
860 | ((shmem_page_offset | page_length) |
879 | ((shmem_page_offset | page_length) |
861 | & (x86_clflush_size - 1)); |
880 | & (x86_clflush_size - 1)); |
862 | 881 | ||
863 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
882 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
864 | (page_to_phys(page) & (1 << 17)) != 0; |
883 | (page_to_phys(page) & (1 << 17)) != 0; |
865 | 884 | ||
866 | ret = shmem_pwrite_fast(page, shmem_page_offset, page_length, |
885 | ret = shmem_pwrite_fast(page, shmem_page_offset, page_length, |
867 | user_data, page_do_bit17_swizzling, |
886 | user_data, page_do_bit17_swizzling, |
868 | partial_cacheline_write, |
887 | partial_cacheline_write, |
869 | needs_clflush_after); |
888 | needs_clflush_after); |
870 | if (ret == 0) |
889 | if (ret == 0) |
871 | goto next_page; |
890 | goto next_page; |
872 | 891 | ||
873 | hit_slowpath = 1; |
892 | hit_slowpath = 1; |
874 | mutex_unlock(&dev->struct_mutex); |
893 | mutex_unlock(&dev->struct_mutex); |
875 | ret = shmem_pwrite_slow(page, shmem_page_offset, page_length, |
894 | ret = shmem_pwrite_slow(page, shmem_page_offset, page_length, |
876 | user_data, page_do_bit17_swizzling, |
895 | user_data, page_do_bit17_swizzling, |
877 | partial_cacheline_write, |
896 | partial_cacheline_write, |
878 | needs_clflush_after); |
897 | needs_clflush_after); |
879 | 898 | ||
880 | mutex_lock(&dev->struct_mutex); |
899 | mutex_lock(&dev->struct_mutex); |
881 | 900 | ||
882 | if (ret) |
901 | if (ret) |
883 | goto out; |
902 | goto out; |
884 | 903 | ||
885 | next_page: |
904 | next_page: |
886 | remain -= page_length; |
905 | remain -= page_length; |
887 | user_data += page_length; |
906 | user_data += page_length; |
888 | offset += page_length; |
907 | offset += page_length; |
889 | } |
908 | } |
890 | 909 | ||
891 | out: |
910 | out: |
892 | i915_gem_object_unpin_pages(obj); |
911 | i915_gem_object_unpin_pages(obj); |
893 | 912 | ||
894 | if (hit_slowpath) { |
913 | if (hit_slowpath) { |
895 | /* |
914 | /* |
896 | * Fixup: Flush cpu caches in case we didn't flush the dirty |
915 | * Fixup: Flush cpu caches in case we didn't flush the dirty |
897 | * cachelines in-line while writing and the object moved |
916 | * cachelines in-line while writing and the object moved |
898 | * out of the cpu write domain while we've dropped the lock. |
917 | * out of the cpu write domain while we've dropped the lock. |
899 | */ |
918 | */ |
900 | if (!needs_clflush_after && |
919 | if (!needs_clflush_after && |
901 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
920 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
902 | if (i915_gem_clflush_object(obj, obj->pin_display)) |
921 | if (i915_gem_clflush_object(obj, obj->pin_display)) |
903 | needs_clflush_after = true; |
922 | needs_clflush_after = true; |
904 | } |
923 | } |
905 | } |
924 | } |
906 | 925 | ||
907 | if (needs_clflush_after) |
926 | if (needs_clflush_after) |
908 | i915_gem_chipset_flush(dev); |
927 | i915_gem_chipset_flush(dev); |
909 | else |
928 | else |
910 | obj->cache_dirty = true; |
929 | obj->cache_dirty = true; |
911 | 930 | ||
912 | intel_fb_obj_flush(obj, false, ORIGIN_CPU); |
931 | intel_fb_obj_flush(obj, false, ORIGIN_CPU); |
913 | return ret; |
932 | return ret; |
914 | } |
933 | } |
915 | 934 | ||
916 | /** |
935 | /** |
917 | * Writes data to the object referenced by handle. |
936 | * Writes data to the object referenced by handle. |
918 | * |
937 | * |
919 | * On error, the contents of the buffer that were to be modified are undefined. |
938 | * On error, the contents of the buffer that were to be modified are undefined. |
920 | */ |
939 | */ |
921 | int |
940 | int |
922 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, |
941 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, |
923 | struct drm_file *file) |
942 | struct drm_file *file) |
924 | { |
943 | { |
925 | struct drm_i915_private *dev_priv = dev->dev_private; |
944 | struct drm_i915_private *dev_priv = dev->dev_private; |
926 | struct drm_i915_gem_pwrite *args = data; |
945 | struct drm_i915_gem_pwrite *args = data; |
927 | struct drm_i915_gem_object *obj; |
946 | struct drm_i915_gem_object *obj; |
928 | int ret; |
947 | int ret; |
929 | 948 | ||
930 | if (args->size == 0) |
949 | if (args->size == 0) |
931 | return 0; |
950 | return 0; |
932 | 951 | ||
933 | intel_runtime_pm_get(dev_priv); |
952 | intel_runtime_pm_get(dev_priv); |
934 | 953 | ||
935 | ret = i915_mutex_lock_interruptible(dev); |
954 | ret = i915_mutex_lock_interruptible(dev); |
936 | if (ret) |
955 | if (ret) |
937 | goto put_rpm; |
956 | goto put_rpm; |
938 | 957 | ||
939 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
958 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
940 | if (&obj->base == NULL) { |
959 | if (&obj->base == NULL) { |
941 | ret = -ENOENT; |
960 | ret = -ENOENT; |
942 | goto unlock; |
961 | goto unlock; |
943 | } |
962 | } |
944 | 963 | ||
945 | /* Bounds check destination. */ |
964 | /* Bounds check destination. */ |
946 | if (args->offset > obj->base.size || |
965 | if (args->offset > obj->base.size || |
947 | args->size > obj->base.size - args->offset) { |
966 | args->size > obj->base.size - args->offset) { |
948 | ret = -EINVAL; |
967 | ret = -EINVAL; |
949 | goto out; |
968 | goto out; |
950 | } |
969 | } |
951 | 970 | ||
952 | /* prime objects have no backing filp to GEM pread/pwrite |
971 | /* prime objects have no backing filp to GEM pread/pwrite |
953 | * pages from. |
972 | * pages from. |
954 | */ |
973 | */ |
955 | if (!obj->base.filp) { |
974 | if (!obj->base.filp) { |
956 | ret = -EINVAL; |
975 | ret = -EINVAL; |
957 | goto out; |
976 | goto out; |
958 | } |
977 | } |
959 | 978 | ||
960 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); |
979 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); |
961 | 980 | ||
962 | ret = -EFAULT; |
981 | ret = -EFAULT; |
963 | /* We can only do the GTT pwrite on untiled buffers, as otherwise |
982 | /* We can only do the GTT pwrite on untiled buffers, as otherwise |
964 | * it would end up going through the fenced access, and we'll get |
983 | * it would end up going through the fenced access, and we'll get |
965 | * different detiling behavior between reading and writing. |
984 | * different detiling behavior between reading and writing. |
966 | * pread/pwrite currently are reading and writing from the CPU |
985 | * pread/pwrite currently are reading and writing from the CPU |
967 | * perspective, requiring manual detiling by the client. |
986 | * perspective, requiring manual detiling by the client. |
968 | */ |
987 | */ |
969 | if (obj->tiling_mode == I915_TILING_NONE && |
988 | if (obj->tiling_mode == I915_TILING_NONE && |
970 | obj->base.write_domain != I915_GEM_DOMAIN_CPU && |
989 | obj->base.write_domain != I915_GEM_DOMAIN_CPU && |
971 | cpu_write_needs_clflush(obj)) { |
990 | cpu_write_needs_clflush(obj)) { |
972 | ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); |
991 | ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); |
973 | /* Note that the gtt paths might fail with non-page-backed user |
992 | /* Note that the gtt paths might fail with non-page-backed user |
974 | * pointers (e.g. gtt mappings when moving data between |
993 | * pointers (e.g. gtt mappings when moving data between |
975 | * textures). Fallback to the shmem path in that case. */ |
994 | * textures). Fallback to the shmem path in that case. */ |
976 | } |
995 | } |
977 | 996 | ||
978 | if (ret == -EFAULT || ret == -ENOSPC) { |
997 | if (ret == -EFAULT || ret == -ENOSPC) { |
979 | ret = i915_gem_shmem_pwrite(dev, obj, args, file); |
998 | ret = i915_gem_shmem_pwrite(dev, obj, args, file); |
980 | } |
999 | } |
981 | 1000 | ||
982 | out: |
1001 | out: |
983 | drm_gem_object_unreference(&obj->base); |
1002 | drm_gem_object_unreference(&obj->base); |
984 | unlock: |
1003 | unlock: |
985 | mutex_unlock(&dev->struct_mutex); |
1004 | mutex_unlock(&dev->struct_mutex); |
986 | put_rpm: |
1005 | put_rpm: |
987 | intel_runtime_pm_put(dev_priv); |
1006 | intel_runtime_pm_put(dev_priv); |
988 | 1007 | ||
989 | return ret; |
1008 | return ret; |
990 | } |
1009 | } |
991 | 1010 | ||
992 | int |
1011 | int |
993 | i915_gem_check_wedge(struct i915_gpu_error *error, |
1012 | i915_gem_check_wedge(struct i915_gpu_error *error, |
994 | bool interruptible) |
1013 | bool interruptible) |
995 | { |
1014 | { |
996 | if (i915_reset_in_progress(error)) { |
1015 | if (i915_reset_in_progress(error)) { |
997 | /* Non-interruptible callers can't handle -EAGAIN, hence return |
1016 | /* Non-interruptible callers can't handle -EAGAIN, hence return |
998 | * -EIO unconditionally for these. */ |
1017 | * -EIO unconditionally for these. */ |
999 | if (!interruptible) |
1018 | if (!interruptible) |
1000 | return -EIO; |
1019 | return -EIO; |
1001 | 1020 | ||
1002 | /* Recovery complete, but the reset failed ... */ |
1021 | /* Recovery complete, but the reset failed ... */ |
1003 | if (i915_terminally_wedged(error)) |
1022 | if (i915_terminally_wedged(error)) |
1004 | return -EIO; |
1023 | return -EIO; |
1005 | 1024 | ||
1006 | /* |
1025 | /* |
1007 | * Check if GPU Reset is in progress - we need intel_ring_begin |
1026 | * Check if GPU Reset is in progress - we need intel_ring_begin |
1008 | * to work properly to reinit the hw state while the gpu is |
1027 | * to work properly to reinit the hw state while the gpu is |
1009 | * still marked as reset-in-progress. Handle this with a flag. |
1028 | * still marked as reset-in-progress. Handle this with a flag. |
1010 | */ |
1029 | */ |
1011 | if (!error->reload_in_reset) |
1030 | if (!error->reload_in_reset) |
1012 | return -EAGAIN; |
1031 | return -EAGAIN; |
1013 | } |
1032 | } |
1014 | 1033 | ||
1015 | return 0; |
1034 | return 0; |
1016 | } |
1035 | } |
1017 | 1036 | ||
1018 | static void fake_irq(unsigned long data) |
1037 | static void fake_irq(unsigned long data) |
1019 | { |
1038 | { |
1020 | // wake_up_process((struct task_struct *)data); |
1039 | // wake_up_process((struct task_struct *)data); |
1021 | } |
1040 | } |
1022 | 1041 | ||
1023 | static bool missed_irq(struct drm_i915_private *dev_priv, |
1042 | static bool missed_irq(struct drm_i915_private *dev_priv, |
1024 | struct intel_engine_cs *ring) |
1043 | struct intel_engine_cs *ring) |
1025 | { |
1044 | { |
1026 | return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings); |
1045 | return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings); |
1027 | } |
1046 | } |
1028 | 1047 | ||
1029 | static unsigned long local_clock_us(unsigned *cpu) |
1048 | static unsigned long local_clock_us(unsigned *cpu) |
1030 | { |
1049 | { |
1031 | unsigned long t; |
1050 | unsigned long t; |
1032 | 1051 | ||
1033 | /* Cheaply and approximately convert from nanoseconds to microseconds. |
1052 | /* Cheaply and approximately convert from nanoseconds to microseconds. |
1034 | * The result and subsequent calculations are also defined in the same |
1053 | * The result and subsequent calculations are also defined in the same |
1035 | * approximate microseconds units. The principal source of timing |
1054 | * approximate microseconds units. The principal source of timing |
1036 | * error here is from the simple truncation. |
1055 | * error here is from the simple truncation. |
1037 | * |
1056 | * |
1038 | * Note that local_clock() is only defined wrt to the current CPU; |
1057 | * Note that local_clock() is only defined wrt to the current CPU; |
1039 | * the comparisons are no longer valid if we switch CPUs. Instead of |
1058 | * the comparisons are no longer valid if we switch CPUs. Instead of |
1040 | * blocking preemption for the entire busywait, we can detect the CPU |
1059 | * blocking preemption for the entire busywait, we can detect the CPU |
1041 | * switch and use that as indicator of system load and a reason to |
1060 | * switch and use that as indicator of system load and a reason to |
1042 | * stop busywaiting, see busywait_stop(). |
1061 | * stop busywaiting, see busywait_stop(). |
1043 | */ |
1062 | */ |
1044 | t = GetClockNs() >> 10; |
1063 | t = GetClockNs() >> 10; |
1045 | 1064 | ||
1046 | return t; |
1065 | return t; |
1047 | } |
1066 | } |
1048 | 1067 | ||
1049 | static bool busywait_stop(unsigned long timeout, unsigned cpu) |
1068 | static bool busywait_stop(unsigned long timeout, unsigned cpu) |
1050 | { |
1069 | { |
1051 | unsigned this_cpu = 0; |
1070 | unsigned this_cpu = 0; |
1052 | 1071 | ||
1053 | if (time_after(local_clock_us(&this_cpu), timeout)) |
1072 | if (time_after(local_clock_us(&this_cpu), timeout)) |
1054 | return true; |
1073 | return true; |
1055 | 1074 | ||
1056 | return this_cpu != cpu; |
1075 | return this_cpu != cpu; |
1057 | } |
1076 | } |
1058 | 1077 | ||
1059 | static int __i915_spin_request(struct drm_i915_gem_request *req, int state) |
1078 | static int __i915_spin_request(struct drm_i915_gem_request *req, int state) |
1060 | { |
1079 | { |
1061 | unsigned long timeout; |
1080 | unsigned long timeout; |
1062 | unsigned cpu; |
1081 | unsigned cpu; |
1063 | 1082 | ||
1064 | /* When waiting for high frequency requests, e.g. during synchronous |
1083 | /* When waiting for high frequency requests, e.g. during synchronous |
1065 | * rendering split between the CPU and GPU, the finite amount of time |
1084 | * rendering split between the CPU and GPU, the finite amount of time |
1066 | * required to set up the irq and wait upon it limits the response |
1085 | * required to set up the irq and wait upon it limits the response |
1067 | * rate. By busywaiting on the request completion for a short while we |
1086 | * rate. By busywaiting on the request completion for a short while we |
1068 | * can service the high frequency waits as quick as possible. However, |
1087 | * can service the high frequency waits as quick as possible. However, |
1069 | * if it is a slow request, we want to sleep as quickly as possible. |
1088 | * if it is a slow request, we want to sleep as quickly as possible. |
1070 | * The tradeoff between waiting and sleeping is roughly the time it |
1089 | * The tradeoff between waiting and sleeping is roughly the time it |
1071 | * takes to sleep on a request, on the order of a microsecond. |
1090 | * takes to sleep on a request, on the order of a microsecond. |
1072 | */ |
1091 | */ |
1073 | 1092 | ||
1074 | if (req->ring->irq_refcount) |
1093 | if (req->ring->irq_refcount) |
1075 | return -EBUSY; |
1094 | return -EBUSY; |
1076 | 1095 | ||
1077 | /* Only spin if we know the GPU is processing this request */ |
1096 | /* Only spin if we know the GPU is processing this request */ |
1078 | if (!i915_gem_request_started(req, true)) |
1097 | if (!i915_gem_request_started(req, true)) |
1079 | return -EAGAIN; |
1098 | return -EAGAIN; |
1080 | 1099 | ||
1081 | timeout = local_clock_us(&cpu) + 5; |
1100 | timeout = local_clock_us(&cpu) + 5; |
1082 | while (1 /*!need_resched()*/) { |
1101 | while (1 /*!need_resched()*/) { |
1083 | if (i915_gem_request_completed(req, true)) |
1102 | if (i915_gem_request_completed(req, true)) |
1084 | return 0; |
1103 | return 0; |
1085 | 1104 | ||
1086 | if (busywait_stop(timeout, cpu)) |
1105 | if (busywait_stop(timeout, cpu)) |
1087 | break; |
1106 | break; |
1088 | 1107 | ||
1089 | cpu_relax_lowlatency(); |
1108 | cpu_relax_lowlatency(); |
1090 | } |
1109 | } |
1091 | 1110 | ||
1092 | if (i915_gem_request_completed(req, false)) |
1111 | if (i915_gem_request_completed(req, false)) |
1093 | return 0; |
1112 | return 0; |
1094 | 1113 | ||
1095 | return -EAGAIN; |
1114 | return -EAGAIN; |
1096 | } |
1115 | } |
1097 | 1116 | ||
1098 | /** |
1117 | /** |
1099 | * __i915_wait_request - wait until execution of request has finished |
1118 | * __i915_wait_request - wait until execution of request has finished |
1100 | * @req: duh! |
1119 | * @req: duh! |
1101 | * @reset_counter: reset sequence associated with the given request |
1120 | * @reset_counter: reset sequence associated with the given request |
1102 | * @interruptible: do an interruptible wait (normally yes) |
1121 | * @interruptible: do an interruptible wait (normally yes) |
1103 | * @timeout: in - how long to wait (NULL forever); out - how much time remaining |
1122 | * @timeout: in - how long to wait (NULL forever); out - how much time remaining |
1104 | * |
1123 | * |
1105 | * Note: It is of utmost importance that the passed in seqno and reset_counter |
1124 | * Note: It is of utmost importance that the passed in seqno and reset_counter |
1106 | * values have been read by the caller in an smp safe manner. Where read-side |
1125 | * values have been read by the caller in an smp safe manner. Where read-side |
1107 | * locks are involved, it is sufficient to read the reset_counter before |
1126 | * locks are involved, it is sufficient to read the reset_counter before |
1108 | * unlocking the lock that protects the seqno. For lockless tricks, the |
1127 | * unlocking the lock that protects the seqno. For lockless tricks, the |
1109 | * reset_counter _must_ be read before, and an appropriate smp_rmb must be |
1128 | * reset_counter _must_ be read before, and an appropriate smp_rmb must be |
1110 | * inserted. |
1129 | * inserted. |
1111 | * |
1130 | * |
1112 | * Returns 0 if the request was found within the alloted time. Else returns the |
1131 | * Returns 0 if the request was found within the alloted time. Else returns the |
1113 | * errno with remaining time filled in timeout argument. |
1132 | * errno with remaining time filled in timeout argument. |
1114 | */ |
1133 | */ |
1115 | int __i915_wait_request(struct drm_i915_gem_request *req, |
1134 | int __i915_wait_request(struct drm_i915_gem_request *req, |
1116 | unsigned reset_counter, |
1135 | unsigned reset_counter, |
1117 | bool interruptible, |
1136 | bool interruptible, |
1118 | s64 *timeout, |
1137 | s64 *timeout, |
1119 | struct intel_rps_client *rps) |
1138 | struct intel_rps_client *rps) |
1120 | { |
1139 | { |
1121 | struct intel_engine_cs *ring = i915_gem_request_get_ring(req); |
1140 | struct intel_engine_cs *ring = i915_gem_request_get_ring(req); |
1122 | struct drm_device *dev = ring->dev; |
1141 | struct drm_device *dev = ring->dev; |
1123 | struct drm_i915_private *dev_priv = dev->dev_private; |
1142 | struct drm_i915_private *dev_priv = dev->dev_private; |
1124 | const bool irq_test_in_progress = |
1143 | const bool irq_test_in_progress = |
1125 | ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring); |
1144 | ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring); |
1126 | int state = interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE; |
1145 | int state = interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE; |
1127 | wait_queue_t wait; |
1146 | wait_queue_t wait; |
1128 | unsigned long timeout_expire; |
1147 | unsigned long timeout_expire; |
1129 | s64 before, now; |
1148 | s64 before = 0; /* Only to silence a compiler warning. */ |
1130 | int ret; |
1149 | int ret; |
1131 | 1150 | ||
1132 | WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled"); |
1151 | WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled"); |
1133 | 1152 | ||
1134 | if (list_empty(&req->list)) |
1153 | if (list_empty(&req->list)) |
1135 | return 0; |
1154 | return 0; |
1136 | 1155 | ||
1137 | if (i915_gem_request_completed(req, true)) |
1156 | if (i915_gem_request_completed(req, true)) |
1138 | return 0; |
1157 | return 0; |
1139 | 1158 | ||
1140 | timeout_expire = 0; |
1159 | timeout_expire = 0; |
1141 | if (timeout) { |
1160 | if (timeout) { |
1142 | if (WARN_ON(*timeout < 0)) |
1161 | if (WARN_ON(*timeout < 0)) |
1143 | return -EINVAL; |
1162 | return -EINVAL; |
1144 | 1163 | ||
1145 | if (*timeout == 0) |
1164 | if (*timeout == 0) |
1146 | return -ETIME; |
1165 | return -ETIME; |
1147 | 1166 | ||
1148 | timeout_expire = jiffies + nsecs_to_jiffies_timeout(*timeout); |
1167 | timeout_expire = jiffies + nsecs_to_jiffies_timeout(*timeout); |
- | 1168 | ||
- | 1169 | /* |
|
- | 1170 | * Record current time in case interrupted by signal, or wedged. |
|
- | 1171 | */ |
|
- | 1172 | before = ktime_get_raw_ns(); |
|
1149 | } |
1173 | } |
1150 | 1174 | ||
1151 | if (INTEL_INFO(dev_priv)->gen >= 6) |
1175 | if (INTEL_INFO(dev_priv)->gen >= 6) |
1152 | gen6_rps_boost(dev_priv, rps, req->emitted_jiffies); |
1176 | gen6_rps_boost(dev_priv, rps, req->emitted_jiffies); |
1153 | - | ||
1154 | /* Record current time in case interrupted by signal, or wedged */ |
1177 | |
1155 | trace_i915_gem_request_wait_begin(req); |
- | |
1156 | before = ktime_get_raw_ns(); |
1178 | trace_i915_gem_request_wait_begin(req); |
1157 | 1179 | ||
1158 | /* Optimistic spin for the next jiffie before touching IRQs */ |
1180 | /* Optimistic spin for the next jiffie before touching IRQs */ |
1159 | ret = __i915_spin_request(req, state); |
1181 | ret = __i915_spin_request(req, state); |
1160 | if (ret == 0) |
1182 | if (ret == 0) |
1161 | goto out; |
1183 | goto out; |
1162 | 1184 | ||
1163 | if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring))) { |
1185 | if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring))) { |
1164 | ret = -ENODEV; |
1186 | ret = -ENODEV; |
1165 | goto out; |
1187 | goto out; |
1166 | } |
1188 | } |
1167 | 1189 | ||
1168 | INIT_LIST_HEAD(&wait.task_list); |
1190 | INIT_LIST_HEAD(&wait.task_list); |
1169 | wait.evnt = CreateEvent(NULL, MANUAL_DESTROY); |
1191 | wait.evnt = CreateEvent(NULL, MANUAL_DESTROY); |
1170 | 1192 | ||
1171 | for (;;) { |
1193 | for (;;) { |
1172 | unsigned long flags; |
1194 | unsigned long flags; |
1173 | 1195 | ||
1174 | /* We need to check whether any gpu reset happened in between |
1196 | /* We need to check whether any gpu reset happened in between |
1175 | * the caller grabbing the seqno and now ... */ |
1197 | * the caller grabbing the seqno and now ... */ |
1176 | if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) { |
1198 | if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) { |
1177 | /* ... but upgrade the -EAGAIN to an -EIO if the gpu |
1199 | /* ... but upgrade the -EAGAIN to an -EIO if the gpu |
1178 | * is truely gone. */ |
1200 | * is truely gone. */ |
1179 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
1201 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
1180 | if (ret == 0) |
1202 | if (ret == 0) |
1181 | ret = -EAGAIN; |
1203 | ret = -EAGAIN; |
1182 | break; |
1204 | break; |
1183 | } |
1205 | } |
1184 | 1206 | ||
1185 | if (i915_gem_request_completed(req, false)) { |
1207 | if (i915_gem_request_completed(req, false)) { |
1186 | ret = 0; |
1208 | ret = 0; |
1187 | break; |
1209 | break; |
1188 | } |
1210 | } |
1189 | 1211 | ||
1190 | if (timeout && time_after_eq(jiffies, timeout_expire)) { |
1212 | if (timeout && time_after_eq(jiffies, timeout_expire)) { |
1191 | ret = -ETIME; |
1213 | ret = -ETIME; |
1192 | break; |
1214 | break; |
1193 | } |
1215 | } |
1194 | 1216 | ||
1195 | spin_lock_irqsave(&ring->irq_queue.lock, flags); |
1217 | spin_lock_irqsave(&ring->irq_queue.lock, flags); |
1196 | if (list_empty(&wait.task_list)) |
1218 | if (list_empty(&wait.task_list)) |
1197 | __add_wait_queue(&ring->irq_queue, &wait); |
1219 | __add_wait_queue(&ring->irq_queue, &wait); |
1198 | spin_unlock_irqrestore(&ring->irq_queue.lock, flags); |
1220 | spin_unlock_irqrestore(&ring->irq_queue.lock, flags); |
1199 | 1221 | ||
1200 | WaitEventTimeout(wait.evnt, 1); |
1222 | WaitEventTimeout(wait.evnt, 1); |
1201 | 1223 | ||
1202 | if (!list_empty(&wait.task_list)) { |
1224 | if (!list_empty(&wait.task_list)) { |
1203 | spin_lock_irqsave(&ring->irq_queue.lock, flags); |
1225 | spin_lock_irqsave(&ring->irq_queue.lock, flags); |
1204 | list_del_init(&wait.task_list); |
1226 | list_del_init(&wait.task_list); |
1205 | spin_unlock_irqrestore(&ring->irq_queue.lock, flags); |
1227 | spin_unlock_irqrestore(&ring->irq_queue.lock, flags); |
1206 | } |
1228 | } |
1207 | 1229 | ||
1208 | }; |
1230 | }; |
1209 | 1231 | ||
1210 | if (!irq_test_in_progress) |
1232 | if (!irq_test_in_progress) |
1211 | ring->irq_put(ring); |
1233 | ring->irq_put(ring); |
1212 | 1234 | ||
1213 | DestroyEvent(wait.evnt); |
1235 | DestroyEvent(wait.evnt); |
1214 | 1236 | ||
1215 | out: |
1237 | out: |
1216 | now = ktime_get_raw_ns(); |
- | |
1217 | trace_i915_gem_request_wait_end(req); |
1238 | trace_i915_gem_request_wait_end(req); |
1218 | 1239 | ||
1219 | if (timeout) { |
1240 | if (timeout) { |
1220 | s64 tres = *timeout - (now - before); |
1241 | s64 tres = *timeout - (ktime_get_raw_ns() - before); |
1221 | 1242 | ||
1222 | *timeout = tres < 0 ? 0 : tres; |
1243 | *timeout = tres < 0 ? 0 : tres; |
1223 | 1244 | ||
1224 | /* |
1245 | /* |
1225 | * Apparently ktime isn't accurate enough and occasionally has a |
1246 | * Apparently ktime isn't accurate enough and occasionally has a |
1226 | * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch |
1247 | * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch |
1227 | * things up to make the test happy. We allow up to 1 jiffy. |
1248 | * things up to make the test happy. We allow up to 1 jiffy. |
1228 | * |
1249 | * |
1229 | * This is a regrssion from the timespec->ktime conversion. |
1250 | * This is a regrssion from the timespec->ktime conversion. |
1230 | */ |
1251 | */ |
1231 | if (ret == -ETIME && *timeout < jiffies_to_usecs(1)*1000) |
1252 | if (ret == -ETIME && *timeout < jiffies_to_usecs(1)*1000) |
1232 | *timeout = 0; |
1253 | *timeout = 0; |
1233 | } |
1254 | } |
1234 | 1255 | ||
1235 | return ret; |
1256 | return ret; |
1236 | } |
1257 | } |
1237 | 1258 | ||
1238 | int i915_gem_request_add_to_client(struct drm_i915_gem_request *req, |
1259 | int i915_gem_request_add_to_client(struct drm_i915_gem_request *req, |
1239 | struct drm_file *file) |
1260 | struct drm_file *file) |
1240 | { |
1261 | { |
1241 | struct drm_i915_private *dev_private; |
1262 | struct drm_i915_private *dev_private; |
1242 | struct drm_i915_file_private *file_priv; |
1263 | struct drm_i915_file_private *file_priv; |
1243 | 1264 | ||
1244 | WARN_ON(!req || !file || req->file_priv); |
1265 | WARN_ON(!req || !file || req->file_priv); |
1245 | 1266 | ||
1246 | if (!req || !file) |
1267 | if (!req || !file) |
1247 | return -EINVAL; |
1268 | return -EINVAL; |
1248 | 1269 | ||
1249 | if (req->file_priv) |
1270 | if (req->file_priv) |
1250 | return -EINVAL; |
1271 | return -EINVAL; |
1251 | 1272 | ||
1252 | dev_private = req->ring->dev->dev_private; |
1273 | dev_private = req->ring->dev->dev_private; |
1253 | file_priv = file->driver_priv; |
1274 | file_priv = file->driver_priv; |
1254 | 1275 | ||
1255 | spin_lock(&file_priv->mm.lock); |
1276 | spin_lock(&file_priv->mm.lock); |
1256 | req->file_priv = file_priv; |
1277 | req->file_priv = file_priv; |
1257 | list_add_tail(&req->client_list, &file_priv->mm.request_list); |
1278 | list_add_tail(&req->client_list, &file_priv->mm.request_list); |
1258 | spin_unlock(&file_priv->mm.lock); |
1279 | spin_unlock(&file_priv->mm.lock); |
1259 | 1280 | ||
1260 | req->pid = (struct pid*)1; |
1281 | req->pid = (struct pid*)1; |
1261 | 1282 | ||
1262 | return 0; |
1283 | return 0; |
1263 | } |
1284 | } |
1264 | 1285 | ||
1265 | static inline void |
1286 | static inline void |
1266 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) |
1287 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) |
1267 | { |
1288 | { |
1268 | struct drm_i915_file_private *file_priv = request->file_priv; |
1289 | struct drm_i915_file_private *file_priv = request->file_priv; |
1269 | 1290 | ||
1270 | if (!file_priv) |
1291 | if (!file_priv) |
1271 | return; |
1292 | return; |
1272 | 1293 | ||
1273 | spin_lock(&file_priv->mm.lock); |
1294 | spin_lock(&file_priv->mm.lock); |
1274 | list_del(&request->client_list); |
1295 | list_del(&request->client_list); |
1275 | request->file_priv = NULL; |
1296 | request->file_priv = NULL; |
1276 | spin_unlock(&file_priv->mm.lock); |
1297 | spin_unlock(&file_priv->mm.lock); |
1277 | request->pid = NULL; |
1298 | request->pid = NULL; |
1278 | } |
1299 | } |
1279 | 1300 | ||
1280 | static void i915_gem_request_retire(struct drm_i915_gem_request *request) |
1301 | static void i915_gem_request_retire(struct drm_i915_gem_request *request) |
1281 | { |
1302 | { |
1282 | trace_i915_gem_request_retire(request); |
1303 | trace_i915_gem_request_retire(request); |
1283 | 1304 | ||
1284 | /* We know the GPU must have read the request to have |
1305 | /* We know the GPU must have read the request to have |
1285 | * sent us the seqno + interrupt, so use the position |
1306 | * sent us the seqno + interrupt, so use the position |
1286 | * of tail of the request to update the last known position |
1307 | * of tail of the request to update the last known position |
1287 | * of the GPU head. |
1308 | * of the GPU head. |
1288 | * |
1309 | * |
1289 | * Note this requires that we are always called in request |
1310 | * Note this requires that we are always called in request |
1290 | * completion order. |
1311 | * completion order. |
1291 | */ |
1312 | */ |
1292 | request->ringbuf->last_retired_head = request->postfix; |
1313 | request->ringbuf->last_retired_head = request->postfix; |
1293 | 1314 | ||
1294 | list_del_init(&request->list); |
1315 | list_del_init(&request->list); |
1295 | i915_gem_request_remove_from_client(request); |
1316 | i915_gem_request_remove_from_client(request); |
1296 | 1317 | ||
1297 | i915_gem_request_unreference(request); |
1318 | i915_gem_request_unreference(request); |
1298 | } |
1319 | } |
1299 | 1320 | ||
1300 | static void |
1321 | static void |
1301 | __i915_gem_request_retire__upto(struct drm_i915_gem_request *req) |
1322 | __i915_gem_request_retire__upto(struct drm_i915_gem_request *req) |
1302 | { |
1323 | { |
1303 | struct intel_engine_cs *engine = req->ring; |
1324 | struct intel_engine_cs *engine = req->ring; |
1304 | struct drm_i915_gem_request *tmp; |
1325 | struct drm_i915_gem_request *tmp; |
1305 | 1326 | ||
1306 | lockdep_assert_held(&engine->dev->struct_mutex); |
1327 | lockdep_assert_held(&engine->dev->struct_mutex); |
1307 | 1328 | ||
1308 | if (list_empty(&req->list)) |
1329 | if (list_empty(&req->list)) |
1309 | return; |
1330 | return; |
1310 | 1331 | ||
1311 | do { |
1332 | do { |
1312 | tmp = list_first_entry(&engine->request_list, |
1333 | tmp = list_first_entry(&engine->request_list, |
1313 | typeof(*tmp), list); |
1334 | typeof(*tmp), list); |
1314 | 1335 | ||
1315 | i915_gem_request_retire(tmp); |
1336 | i915_gem_request_retire(tmp); |
1316 | } while (tmp != req); |
1337 | } while (tmp != req); |
1317 | 1338 | ||
1318 | WARN_ON(i915_verify_lists(engine->dev)); |
1339 | WARN_ON(i915_verify_lists(engine->dev)); |
1319 | } |
1340 | } |
1320 | 1341 | ||
1321 | /** |
1342 | /** |
1322 | * Waits for a request to be signaled, and cleans up the |
1343 | * Waits for a request to be signaled, and cleans up the |
1323 | * request and object lists appropriately for that event. |
1344 | * request and object lists appropriately for that event. |
1324 | */ |
1345 | */ |
1325 | int |
1346 | int |
1326 | i915_wait_request(struct drm_i915_gem_request *req) |
1347 | i915_wait_request(struct drm_i915_gem_request *req) |
1327 | { |
1348 | { |
1328 | struct drm_device *dev; |
1349 | struct drm_device *dev; |
1329 | struct drm_i915_private *dev_priv; |
1350 | struct drm_i915_private *dev_priv; |
1330 | bool interruptible; |
1351 | bool interruptible; |
1331 | int ret; |
1352 | int ret; |
1332 | 1353 | ||
1333 | BUG_ON(req == NULL); |
1354 | BUG_ON(req == NULL); |
1334 | 1355 | ||
1335 | dev = req->ring->dev; |
1356 | dev = req->ring->dev; |
1336 | dev_priv = dev->dev_private; |
1357 | dev_priv = dev->dev_private; |
1337 | interruptible = dev_priv->mm.interruptible; |
1358 | interruptible = dev_priv->mm.interruptible; |
1338 | 1359 | ||
1339 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); |
1360 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); |
1340 | 1361 | ||
1341 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
1362 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
1342 | if (ret) |
1363 | if (ret) |
1343 | return ret; |
1364 | return ret; |
1344 | 1365 | ||
1345 | ret = __i915_wait_request(req, |
1366 | ret = __i915_wait_request(req, |
1346 | atomic_read(&dev_priv->gpu_error.reset_counter), |
1367 | atomic_read(&dev_priv->gpu_error.reset_counter), |
1347 | interruptible, NULL, NULL); |
1368 | interruptible, NULL, NULL); |
1348 | if (ret) |
1369 | if (ret) |
1349 | return ret; |
1370 | return ret; |
1350 | 1371 | ||
1351 | __i915_gem_request_retire__upto(req); |
1372 | __i915_gem_request_retire__upto(req); |
1352 | return 0; |
1373 | return 0; |
1353 | } |
1374 | } |
1354 | 1375 | ||
1355 | /** |
1376 | /** |
1356 | * Ensures that all rendering to the object has completed and the object is |
1377 | * Ensures that all rendering to the object has completed and the object is |
1357 | * safe to unbind from the GTT or access from the CPU. |
1378 | * safe to unbind from the GTT or access from the CPU. |
1358 | */ |
1379 | */ |
1359 | int |
1380 | int |
1360 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj, |
1381 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj, |
1361 | bool readonly) |
1382 | bool readonly) |
1362 | { |
1383 | { |
1363 | int ret, i; |
1384 | int ret, i; |
1364 | 1385 | ||
1365 | if (!obj->active) |
1386 | if (!obj->active) |
1366 | return 0; |
1387 | return 0; |
1367 | 1388 | ||
1368 | if (readonly) { |
1389 | if (readonly) { |
1369 | if (obj->last_write_req != NULL) { |
1390 | if (obj->last_write_req != NULL) { |
1370 | ret = i915_wait_request(obj->last_write_req); |
1391 | ret = i915_wait_request(obj->last_write_req); |
1371 | if (ret) |
1392 | if (ret) |
1372 | return ret; |
1393 | return ret; |
1373 | 1394 | ||
1374 | i = obj->last_write_req->ring->id; |
1395 | i = obj->last_write_req->ring->id; |
1375 | if (obj->last_read_req[i] == obj->last_write_req) |
1396 | if (obj->last_read_req[i] == obj->last_write_req) |
1376 | i915_gem_object_retire__read(obj, i); |
1397 | i915_gem_object_retire__read(obj, i); |
1377 | else |
1398 | else |
1378 | i915_gem_object_retire__write(obj); |
1399 | i915_gem_object_retire__write(obj); |
1379 | } |
1400 | } |
1380 | } else { |
1401 | } else { |
1381 | for (i = 0; i < I915_NUM_RINGS; i++) { |
1402 | for (i = 0; i < I915_NUM_RINGS; i++) { |
1382 | if (obj->last_read_req[i] == NULL) |
1403 | if (obj->last_read_req[i] == NULL) |
1383 | continue; |
1404 | continue; |
1384 | 1405 | ||
1385 | ret = i915_wait_request(obj->last_read_req[i]); |
1406 | ret = i915_wait_request(obj->last_read_req[i]); |
1386 | if (ret) |
1407 | if (ret) |
1387 | return ret; |
1408 | return ret; |
1388 | 1409 | ||
1389 | i915_gem_object_retire__read(obj, i); |
1410 | i915_gem_object_retire__read(obj, i); |
1390 | } |
1411 | } |
1391 | RQ_BUG_ON(obj->active); |
1412 | RQ_BUG_ON(obj->active); |
1392 | } |
1413 | } |
1393 | 1414 | ||
1394 | return 0; |
1415 | return 0; |
1395 | } |
1416 | } |
1396 | 1417 | ||
1397 | static void |
1418 | static void |
1398 | i915_gem_object_retire_request(struct drm_i915_gem_object *obj, |
1419 | i915_gem_object_retire_request(struct drm_i915_gem_object *obj, |
1399 | struct drm_i915_gem_request *req) |
1420 | struct drm_i915_gem_request *req) |
1400 | { |
1421 | { |
1401 | int ring = req->ring->id; |
1422 | int ring = req->ring->id; |
1402 | 1423 | ||
1403 | if (obj->last_read_req[ring] == req) |
1424 | if (obj->last_read_req[ring] == req) |
1404 | i915_gem_object_retire__read(obj, ring); |
1425 | i915_gem_object_retire__read(obj, ring); |
1405 | else if (obj->last_write_req == req) |
1426 | else if (obj->last_write_req == req) |
1406 | i915_gem_object_retire__write(obj); |
1427 | i915_gem_object_retire__write(obj); |
1407 | 1428 | ||
1408 | __i915_gem_request_retire__upto(req); |
1429 | __i915_gem_request_retire__upto(req); |
1409 | } |
1430 | } |
1410 | 1431 | ||
1411 | /* A nonblocking variant of the above wait. This is a highly dangerous routine |
1432 | /* A nonblocking variant of the above wait. This is a highly dangerous routine |
1412 | * as the object state may change during this call. |
1433 | * as the object state may change during this call. |
1413 | */ |
1434 | */ |
1414 | static __must_check int |
1435 | static __must_check int |
1415 | i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj, |
1436 | i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj, |
1416 | struct intel_rps_client *rps, |
1437 | struct intel_rps_client *rps, |
1417 | bool readonly) |
1438 | bool readonly) |
1418 | { |
1439 | { |
1419 | struct drm_device *dev = obj->base.dev; |
1440 | struct drm_device *dev = obj->base.dev; |
1420 | struct drm_i915_private *dev_priv = dev->dev_private; |
1441 | struct drm_i915_private *dev_priv = dev->dev_private; |
1421 | struct drm_i915_gem_request *requests[I915_NUM_RINGS]; |
1442 | struct drm_i915_gem_request *requests[I915_NUM_RINGS]; |
1422 | unsigned reset_counter; |
1443 | unsigned reset_counter; |
1423 | int ret, i, n = 0; |
1444 | int ret, i, n = 0; |
1424 | 1445 | ||
1425 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); |
1446 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); |
1426 | BUG_ON(!dev_priv->mm.interruptible); |
1447 | BUG_ON(!dev_priv->mm.interruptible); |
1427 | 1448 | ||
1428 | if (!obj->active) |
1449 | if (!obj->active) |
1429 | return 0; |
1450 | return 0; |
1430 | 1451 | ||
1431 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, true); |
1452 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, true); |
1432 | if (ret) |
1453 | if (ret) |
1433 | return ret; |
1454 | return ret; |
1434 | 1455 | ||
1435 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
1456 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
1436 | 1457 | ||
1437 | if (readonly) { |
1458 | if (readonly) { |
1438 | struct drm_i915_gem_request *req; |
1459 | struct drm_i915_gem_request *req; |
1439 | 1460 | ||
1440 | req = obj->last_write_req; |
1461 | req = obj->last_write_req; |
1441 | if (req == NULL) |
1462 | if (req == NULL) |
1442 | return 0; |
1463 | return 0; |
1443 | 1464 | ||
1444 | requests[n++] = i915_gem_request_reference(req); |
1465 | requests[n++] = i915_gem_request_reference(req); |
1445 | } else { |
1466 | } else { |
1446 | for (i = 0; i < I915_NUM_RINGS; i++) { |
1467 | for (i = 0; i < I915_NUM_RINGS; i++) { |
1447 | struct drm_i915_gem_request *req; |
1468 | struct drm_i915_gem_request *req; |
1448 | 1469 | ||
1449 | req = obj->last_read_req[i]; |
1470 | req = obj->last_read_req[i]; |
1450 | if (req == NULL) |
1471 | if (req == NULL) |
1451 | continue; |
1472 | continue; |
1452 | 1473 | ||
1453 | requests[n++] = i915_gem_request_reference(req); |
1474 | requests[n++] = i915_gem_request_reference(req); |
1454 | } |
1475 | } |
1455 | } |
1476 | } |
1456 | 1477 | ||
1457 | mutex_unlock(&dev->struct_mutex); |
1478 | mutex_unlock(&dev->struct_mutex); |
1458 | for (i = 0; ret == 0 && i < n; i++) |
1479 | for (i = 0; ret == 0 && i < n; i++) |
1459 | ret = __i915_wait_request(requests[i], reset_counter, true, |
1480 | ret = __i915_wait_request(requests[i], reset_counter, true, |
1460 | NULL, rps); |
1481 | NULL, rps); |
1461 | mutex_lock(&dev->struct_mutex); |
1482 | mutex_lock(&dev->struct_mutex); |
1462 | 1483 | ||
1463 | for (i = 0; i < n; i++) { |
1484 | for (i = 0; i < n; i++) { |
1464 | if (ret == 0) |
1485 | if (ret == 0) |
1465 | i915_gem_object_retire_request(obj, requests[i]); |
1486 | i915_gem_object_retire_request(obj, requests[i]); |
1466 | i915_gem_request_unreference(requests[i]); |
1487 | i915_gem_request_unreference(requests[i]); |
1467 | } |
1488 | } |
1468 | 1489 | ||
1469 | return ret; |
1490 | return ret; |
1470 | } |
1491 | } |
1471 | 1492 | ||
1472 | static struct intel_rps_client *to_rps_client(struct drm_file *file) |
1493 | static struct intel_rps_client *to_rps_client(struct drm_file *file) |
1473 | { |
1494 | { |
1474 | struct drm_i915_file_private *fpriv = file->driver_priv; |
1495 | struct drm_i915_file_private *fpriv = file->driver_priv; |
1475 | return &fpriv->rps; |
1496 | return &fpriv->rps; |
1476 | } |
1497 | } |
1477 | 1498 | ||
1478 | /** |
1499 | /** |
1479 | * Called when user space prepares to use an object with the CPU, either |
1500 | * Called when user space prepares to use an object with the CPU, either |
1480 | * through the mmap ioctl's mapping or a GTT mapping. |
1501 | * through the mmap ioctl's mapping or a GTT mapping. |
1481 | */ |
1502 | */ |
1482 | int |
1503 | int |
1483 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, |
1504 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, |
1484 | struct drm_file *file) |
1505 | struct drm_file *file) |
1485 | { |
1506 | { |
1486 | struct drm_i915_gem_set_domain *args = data; |
1507 | struct drm_i915_gem_set_domain *args = data; |
1487 | struct drm_i915_gem_object *obj; |
1508 | struct drm_i915_gem_object *obj; |
1488 | uint32_t read_domains = args->read_domains; |
1509 | uint32_t read_domains = args->read_domains; |
1489 | uint32_t write_domain = args->write_domain; |
1510 | uint32_t write_domain = args->write_domain; |
1490 | int ret; |
1511 | int ret; |
1491 | 1512 | ||
1492 | /* Only handle setting domains to types used by the CPU. */ |
1513 | /* Only handle setting domains to types used by the CPU. */ |
1493 | if (write_domain & I915_GEM_GPU_DOMAINS) |
1514 | if (write_domain & I915_GEM_GPU_DOMAINS) |
1494 | return -EINVAL; |
1515 | return -EINVAL; |
1495 | 1516 | ||
1496 | if (read_domains & I915_GEM_GPU_DOMAINS) |
1517 | if (read_domains & I915_GEM_GPU_DOMAINS) |
1497 | return -EINVAL; |
1518 | return -EINVAL; |
1498 | 1519 | ||
1499 | /* Having something in the write domain implies it's in the read |
1520 | /* Having something in the write domain implies it's in the read |
1500 | * domain, and only that read domain. Enforce that in the request. |
1521 | * domain, and only that read domain. Enforce that in the request. |
1501 | */ |
1522 | */ |
1502 | if (write_domain != 0 && read_domains != write_domain) |
1523 | if (write_domain != 0 && read_domains != write_domain) |
1503 | return -EINVAL; |
1524 | return -EINVAL; |
1504 | 1525 | ||
1505 | ret = i915_mutex_lock_interruptible(dev); |
1526 | ret = i915_mutex_lock_interruptible(dev); |
1506 | if (ret) |
1527 | if (ret) |
1507 | return ret; |
1528 | return ret; |
1508 | 1529 | ||
1509 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
1530 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
1510 | if (&obj->base == NULL) { |
1531 | if (&obj->base == NULL) { |
1511 | ret = -ENOENT; |
1532 | ret = -ENOENT; |
1512 | goto unlock; |
1533 | goto unlock; |
1513 | } |
1534 | } |
1514 | 1535 | ||
1515 | /* Try to flush the object off the GPU without holding the lock. |
1536 | /* Try to flush the object off the GPU without holding the lock. |
1516 | * We will repeat the flush holding the lock in the normal manner |
1537 | * We will repeat the flush holding the lock in the normal manner |
1517 | * to catch cases where we are gazumped. |
1538 | * to catch cases where we are gazumped. |
1518 | */ |
1539 | */ |
1519 | ret = i915_gem_object_wait_rendering__nonblocking(obj, |
1540 | ret = i915_gem_object_wait_rendering__nonblocking(obj, |
1520 | to_rps_client(file), |
1541 | to_rps_client(file), |
1521 | !write_domain); |
1542 | !write_domain); |
1522 | if (ret) |
1543 | if (ret) |
1523 | goto unref; |
1544 | goto unref; |
1524 | 1545 | ||
1525 | if (read_domains & I915_GEM_DOMAIN_GTT) |
1546 | if (read_domains & I915_GEM_DOMAIN_GTT) |
1526 | ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); |
1547 | ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); |
1527 | else |
1548 | else |
1528 | ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); |
1549 | ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); |
1529 | 1550 | ||
1530 | if (write_domain != 0) |
1551 | if (write_domain != 0) |
1531 | intel_fb_obj_invalidate(obj, |
1552 | intel_fb_obj_invalidate(obj, |
1532 | write_domain == I915_GEM_DOMAIN_GTT ? |
1553 | write_domain == I915_GEM_DOMAIN_GTT ? |
1533 | ORIGIN_GTT : ORIGIN_CPU); |
1554 | ORIGIN_GTT : ORIGIN_CPU); |
1534 | 1555 | ||
1535 | unref: |
1556 | unref: |
1536 | drm_gem_object_unreference(&obj->base); |
1557 | drm_gem_object_unreference(&obj->base); |
1537 | unlock: |
1558 | unlock: |
1538 | mutex_unlock(&dev->struct_mutex); |
1559 | mutex_unlock(&dev->struct_mutex); |
1539 | return ret; |
1560 | return ret; |
1540 | } |
1561 | } |
1541 | 1562 | ||
1542 | /** |
1563 | /** |
1543 | * Called when user space has done writes to this buffer |
1564 | * Called when user space has done writes to this buffer |
1544 | */ |
1565 | */ |
1545 | int |
1566 | int |
1546 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, |
1567 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, |
1547 | struct drm_file *file) |
1568 | struct drm_file *file) |
1548 | { |
1569 | { |
1549 | struct drm_i915_gem_sw_finish *args = data; |
1570 | struct drm_i915_gem_sw_finish *args = data; |
1550 | struct drm_i915_gem_object *obj; |
1571 | struct drm_i915_gem_object *obj; |
1551 | int ret = 0; |
1572 | int ret = 0; |
1552 | 1573 | ||
1553 | ret = i915_mutex_lock_interruptible(dev); |
1574 | ret = i915_mutex_lock_interruptible(dev); |
1554 | if (ret) |
1575 | if (ret) |
1555 | return ret; |
1576 | return ret; |
1556 | 1577 | ||
1557 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
1578 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
1558 | if (&obj->base == NULL) { |
1579 | if (&obj->base == NULL) { |
1559 | ret = -ENOENT; |
1580 | ret = -ENOENT; |
1560 | goto unlock; |
1581 | goto unlock; |
1561 | } |
1582 | } |
1562 | 1583 | ||
1563 | /* Pinned buffers may be scanout, so flush the cache */ |
1584 | /* Pinned buffers may be scanout, so flush the cache */ |
1564 | if (obj->pin_display) |
1585 | if (obj->pin_display) |
1565 | i915_gem_object_flush_cpu_write_domain(obj); |
1586 | i915_gem_object_flush_cpu_write_domain(obj); |
1566 | 1587 | ||
1567 | drm_gem_object_unreference(&obj->base); |
1588 | drm_gem_object_unreference(&obj->base); |
1568 | unlock: |
1589 | unlock: |
1569 | mutex_unlock(&dev->struct_mutex); |
1590 | mutex_unlock(&dev->struct_mutex); |
1570 | return ret; |
1591 | return ret; |
1571 | } |
1592 | } |
1572 | 1593 | ||
1573 | /** |
1594 | /** |
1574 | * Maps the contents of an object, returning the address it is mapped |
1595 | * Maps the contents of an object, returning the address it is mapped |
1575 | * into. |
1596 | * into. |
1576 | * |
1597 | * |
1577 | * While the mapping holds a reference on the contents of the object, it doesn't |
1598 | * While the mapping holds a reference on the contents of the object, it doesn't |
1578 | * imply a ref on the object itself. |
1599 | * imply a ref on the object itself. |
1579 | * |
1600 | * |
1580 | * IMPORTANT: |
1601 | * IMPORTANT: |
1581 | * |
1602 | * |
1582 | * DRM driver writers who look a this function as an example for how to do GEM |
1603 | * DRM driver writers who look a this function as an example for how to do GEM |
1583 | * mmap support, please don't implement mmap support like here. The modern way |
1604 | * mmap support, please don't implement mmap support like here. The modern way |
1584 | * to implement DRM mmap support is with an mmap offset ioctl (like |
1605 | * to implement DRM mmap support is with an mmap offset ioctl (like |
1585 | * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly. |
1606 | * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly. |
1586 | * That way debug tooling like valgrind will understand what's going on, hiding |
1607 | * That way debug tooling like valgrind will understand what's going on, hiding |
1587 | * the mmap call in a driver private ioctl will break that. The i915 driver only |
1608 | * the mmap call in a driver private ioctl will break that. The i915 driver only |
1588 | * does cpu mmaps this way because we didn't know better. |
1609 | * does cpu mmaps this way because we didn't know better. |
1589 | */ |
1610 | */ |
1590 | int |
1611 | int |
1591 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, |
1612 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, |
1592 | struct drm_file *file) |
1613 | struct drm_file *file) |
1593 | { |
1614 | { |
1594 | struct drm_i915_gem_mmap *args = data; |
1615 | struct drm_i915_gem_mmap *args = data; |
1595 | struct drm_gem_object *obj; |
1616 | struct drm_gem_object *obj; |
1596 | unsigned long addr; |
1617 | unsigned long addr; |
1597 | 1618 | ||
1598 | // if (args->flags & ~(I915_MMAP_WC)) |
1619 | // if (args->flags & ~(I915_MMAP_WC)) |
1599 | // return -EINVAL; |
1620 | // return -EINVAL; |
1600 | obj = drm_gem_object_lookup(dev, file, args->handle); |
1621 | obj = drm_gem_object_lookup(dev, file, args->handle); |
1601 | if (obj == NULL) |
1622 | if (obj == NULL) |
1602 | return -ENOENT; |
1623 | return -ENOENT; |
1603 | 1624 | ||
1604 | /* prime objects have no backing filp to GEM mmap |
1625 | /* prime objects have no backing filp to GEM mmap |
1605 | * pages from. |
1626 | * pages from. |
1606 | */ |
1627 | */ |
1607 | if (!obj->filp) { |
1628 | if (!obj->filp) { |
1608 | drm_gem_object_unreference_unlocked(obj); |
1629 | drm_gem_object_unreference_unlocked(obj); |
1609 | return -EINVAL; |
1630 | return -EINVAL; |
1610 | } |
1631 | } |
1611 | 1632 | ||
1612 | addr = vm_mmap(obj->filp, 0, args->size, |
1633 | addr = vm_mmap(obj->filp, 0, args->size, |
1613 | PROT_READ | PROT_WRITE, MAP_SHARED, |
1634 | PROT_READ | PROT_WRITE, MAP_SHARED, |
1614 | args->offset); |
1635 | args->offset); |
1615 | drm_gem_object_unreference_unlocked(obj); |
1636 | drm_gem_object_unreference_unlocked(obj); |
1616 | if (IS_ERR((void *)addr)) |
1637 | if (IS_ERR((void *)addr)) |
1617 | return addr; |
1638 | return addr; |
1618 | 1639 | ||
1619 | args->addr_ptr = (uint64_t) addr; |
1640 | args->addr_ptr = (uint64_t) addr; |
1620 | 1641 | ||
1621 | return 0; |
1642 | return 0; |
1622 | } |
1643 | } |
1623 | 1644 | ||
1624 | 1645 | ||
1625 | 1646 | ||
1626 | 1647 | ||
1627 | 1648 | ||
1628 | 1649 | ||
1629 | 1650 | ||
1630 | 1651 | ||
1631 | 1652 | ||
1632 | 1653 | ||
1633 | 1654 | ||
1634 | 1655 | ||
1635 | 1656 | ||
1636 | /** |
1657 | /** |
1637 | * i915_gem_release_mmap - remove physical page mappings |
1658 | * i915_gem_release_mmap - remove physical page mappings |
1638 | * @obj: obj in question |
1659 | * @obj: obj in question |
1639 | * |
1660 | * |
1640 | * Preserve the reservation of the mmapping with the DRM core code, but |
1661 | * Preserve the reservation of the mmapping with the DRM core code, but |
1641 | * relinquish ownership of the pages back to the system. |
1662 | * relinquish ownership of the pages back to the system. |
1642 | * |
1663 | * |
1643 | * It is vital that we remove the page mapping if we have mapped a tiled |
1664 | * It is vital that we remove the page mapping if we have mapped a tiled |
1644 | * object through the GTT and then lose the fence register due to |
1665 | * object through the GTT and then lose the fence register due to |
1645 | * resource pressure. Similarly if the object has been moved out of the |
1666 | * resource pressure. Similarly if the object has been moved out of the |
1646 | * aperture, than pages mapped into userspace must be revoked. Removing the |
1667 | * aperture, than pages mapped into userspace must be revoked. Removing the |
1647 | * mapping will then trigger a page fault on the next user access, allowing |
1668 | * mapping will then trigger a page fault on the next user access, allowing |
1648 | * fixup by i915_gem_fault(). |
1669 | * fixup by i915_gem_fault(). |
1649 | */ |
1670 | */ |
1650 | void |
1671 | void |
1651 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) |
1672 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) |
1652 | { |
1673 | { |
1653 | if (!obj->fault_mappable) |
1674 | if (!obj->fault_mappable) |
1654 | return; |
1675 | return; |
1655 | 1676 | ||
1656 | // drm_vma_node_unmap(&obj->base.vma_node, obj->base.dev->dev_mapping); |
1677 | // drm_vma_node_unmap(&obj->base.vma_node, obj->base.dev->dev_mapping); |
1657 | obj->fault_mappable = false; |
1678 | obj->fault_mappable = false; |
1658 | } |
1679 | } |
1659 | 1680 | ||
1660 | void |
1681 | void |
1661 | i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv) |
1682 | i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv) |
1662 | { |
1683 | { |
1663 | struct drm_i915_gem_object *obj; |
1684 | struct drm_i915_gem_object *obj; |
1664 | 1685 | ||
1665 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) |
1686 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) |
1666 | i915_gem_release_mmap(obj); |
1687 | i915_gem_release_mmap(obj); |
1667 | } |
1688 | } |
1668 | 1689 | ||
1669 | uint32_t |
1690 | uint32_t |
1670 | i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode) |
1691 | i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode) |
1671 | { |
1692 | { |
1672 | uint32_t gtt_size; |
1693 | uint32_t gtt_size; |
1673 | 1694 | ||
1674 | if (INTEL_INFO(dev)->gen >= 4 || |
1695 | if (INTEL_INFO(dev)->gen >= 4 || |
1675 | tiling_mode == I915_TILING_NONE) |
1696 | tiling_mode == I915_TILING_NONE) |
1676 | return size; |
1697 | return size; |
1677 | 1698 | ||
1678 | /* Previous chips need a power-of-two fence region when tiling */ |
1699 | /* Previous chips need a power-of-two fence region when tiling */ |
1679 | if (INTEL_INFO(dev)->gen == 3) |
1700 | if (INTEL_INFO(dev)->gen == 3) |
1680 | gtt_size = 1024*1024; |
1701 | gtt_size = 1024*1024; |
1681 | else |
1702 | else |
1682 | gtt_size = 512*1024; |
1703 | gtt_size = 512*1024; |
1683 | 1704 | ||
1684 | while (gtt_size < size) |
1705 | while (gtt_size < size) |
1685 | gtt_size <<= 1; |
1706 | gtt_size <<= 1; |
1686 | 1707 | ||
1687 | return gtt_size; |
1708 | return gtt_size; |
1688 | } |
1709 | } |
1689 | 1710 | ||
1690 | /** |
1711 | /** |
1691 | * i915_gem_get_gtt_alignment - return required GTT alignment for an object |
1712 | * i915_gem_get_gtt_alignment - return required GTT alignment for an object |
1692 | * @obj: object to check |
1713 | * @obj: object to check |
1693 | * |
1714 | * |
1694 | * Return the required GTT alignment for an object, taking into account |
1715 | * Return the required GTT alignment for an object, taking into account |
1695 | * potential fence register mapping. |
1716 | * potential fence register mapping. |
1696 | */ |
1717 | */ |
1697 | uint32_t |
1718 | uint32_t |
1698 | i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size, |
1719 | i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size, |
1699 | int tiling_mode, bool fenced) |
1720 | int tiling_mode, bool fenced) |
1700 | { |
1721 | { |
1701 | /* |
1722 | /* |
1702 | * Minimum alignment is 4k (GTT page size), but might be greater |
1723 | * Minimum alignment is 4k (GTT page size), but might be greater |
1703 | * if a fence register is needed for the object. |
1724 | * if a fence register is needed for the object. |
1704 | */ |
1725 | */ |
1705 | if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) || |
1726 | if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) || |
1706 | tiling_mode == I915_TILING_NONE) |
1727 | tiling_mode == I915_TILING_NONE) |
1707 | return 4096; |
1728 | return 4096; |
1708 | 1729 | ||
1709 | /* |
1730 | /* |
1710 | * Previous chips need to be aligned to the size of the smallest |
1731 | * Previous chips need to be aligned to the size of the smallest |
1711 | * fence register that can contain the object. |
1732 | * fence register that can contain the object. |
1712 | */ |
1733 | */ |
1713 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
1734 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
1714 | } |
1735 | } |
1715 | 1736 | ||
1716 | 1737 | ||
1717 | 1738 | ||
1718 | int |
1739 | int |
1719 | i915_gem_mmap_gtt(struct drm_file *file, |
1740 | i915_gem_mmap_gtt(struct drm_file *file, |
1720 | struct drm_device *dev, |
1741 | struct drm_device *dev, |
1721 | uint32_t handle, |
1742 | uint32_t handle, |
1722 | uint64_t *offset) |
1743 | uint64_t *offset) |
1723 | { |
1744 | { |
1724 | struct drm_i915_private *dev_priv = dev->dev_private; |
1745 | struct drm_i915_private *dev_priv = dev->dev_private; |
1725 | struct drm_i915_gem_object *obj; |
1746 | struct drm_i915_gem_object *obj; |
1726 | unsigned long pfn; |
1747 | unsigned long pfn; |
1727 | char *mem, *ptr; |
1748 | char *mem, *ptr; |
1728 | int ret; |
1749 | int ret; |
1729 | 1750 | ||
1730 | ret = i915_mutex_lock_interruptible(dev); |
1751 | ret = i915_mutex_lock_interruptible(dev); |
1731 | if (ret) |
1752 | if (ret) |
1732 | return ret; |
1753 | return ret; |
1733 | 1754 | ||
1734 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle)); |
1755 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle)); |
1735 | if (&obj->base == NULL) { |
1756 | if (&obj->base == NULL) { |
1736 | ret = -ENOENT; |
1757 | ret = -ENOENT; |
1737 | goto unlock; |
1758 | goto unlock; |
1738 | } |
1759 | } |
1739 | 1760 | ||
1740 | if (obj->madv != I915_MADV_WILLNEED) { |
1761 | if (obj->madv != I915_MADV_WILLNEED) { |
1741 | DRM_DEBUG("Attempting to mmap a purgeable buffer\n"); |
1762 | DRM_DEBUG("Attempting to mmap a purgeable buffer\n"); |
1742 | ret = -EFAULT; |
1763 | ret = -EFAULT; |
1743 | goto out; |
1764 | goto out; |
1744 | } |
1765 | } |
1745 | /* Now bind it into the GTT if needed */ |
1766 | /* Now bind it into the GTT if needed */ |
1746 | ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK); |
1767 | ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK); |
1747 | if (ret) |
1768 | if (ret) |
1748 | goto out; |
1769 | goto out; |
1749 | 1770 | ||
1750 | ret = i915_gem_object_set_to_gtt_domain(obj, 1); |
1771 | ret = i915_gem_object_set_to_gtt_domain(obj, 1); |
1751 | if (ret) |
1772 | if (ret) |
1752 | goto unpin; |
1773 | goto unpin; |
1753 | 1774 | ||
1754 | ret = i915_gem_object_get_fence(obj); |
1775 | ret = i915_gem_object_get_fence(obj); |
1755 | if (ret) |
1776 | if (ret) |
1756 | goto unpin; |
1777 | goto unpin; |
1757 | 1778 | ||
1758 | obj->fault_mappable = true; |
1779 | obj->fault_mappable = true; |
1759 | 1780 | ||
1760 | pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj); |
1781 | pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj); |
1761 | 1782 | ||
1762 | /* Finally, remap it using the new GTT offset */ |
1783 | /* Finally, remap it using the new GTT offset */ |
1763 | 1784 | ||
1764 | mem = UserAlloc(obj->base.size); |
1785 | mem = UserAlloc(obj->base.size); |
1765 | if(unlikely(mem == NULL)) |
1786 | if(unlikely(mem == NULL)) |
1766 | { |
1787 | { |
1767 | ret = -ENOMEM; |
1788 | ret = -ENOMEM; |
1768 | goto unpin; |
1789 | goto unpin; |
1769 | } |
1790 | } |
1770 | 1791 | ||
1771 | for(ptr = mem; ptr < mem + obj->base.size; ptr+= 4096, pfn+= 4096) |
1792 | for(ptr = mem; ptr < mem + obj->base.size; ptr+= 4096, pfn+= 4096) |
1772 | MapPage(ptr, pfn, PG_SHARED|PG_UW); |
1793 | MapPage(ptr, pfn, PG_SHARED|PG_UW); |
1773 | 1794 | ||
1774 | unpin: |
1795 | unpin: |
1775 | i915_gem_object_unpin_pages(obj); |
1796 | i915_gem_object_unpin_pages(obj); |
1776 | 1797 | ||
1777 | 1798 | ||
1778 | *offset = (uint32_t)mem; |
1799 | *offset = (uint32_t)mem; |
1779 | 1800 | ||
1780 | out: |
1801 | out: |
1781 | drm_gem_object_unreference(&obj->base); |
1802 | drm_gem_object_unreference(&obj->base); |
1782 | unlock: |
1803 | unlock: |
1783 | mutex_unlock(&dev->struct_mutex); |
1804 | mutex_unlock(&dev->struct_mutex); |
1784 | return ret; |
1805 | return ret; |
1785 | } |
1806 | } |
1786 | 1807 | ||
1787 | /** |
1808 | /** |
1788 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing |
1809 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing |
1789 | * @dev: DRM device |
1810 | * @dev: DRM device |
1790 | * @data: GTT mapping ioctl data |
1811 | * @data: GTT mapping ioctl data |
1791 | * @file: GEM object info |
1812 | * @file: GEM object info |
1792 | * |
1813 | * |
1793 | * Simply returns the fake offset to userspace so it can mmap it. |
1814 | * Simply returns the fake offset to userspace so it can mmap it. |
1794 | * The mmap call will end up in drm_gem_mmap(), which will set things |
1815 | * The mmap call will end up in drm_gem_mmap(), which will set things |
1795 | * up so we can get faults in the handler above. |
1816 | * up so we can get faults in the handler above. |
1796 | * |
1817 | * |
1797 | * The fault handler will take care of binding the object into the GTT |
1818 | * The fault handler will take care of binding the object into the GTT |
1798 | * (since it may have been evicted to make room for something), allocating |
1819 | * (since it may have been evicted to make room for something), allocating |
1799 | * a fence register, and mapping the appropriate aperture address into |
1820 | * a fence register, and mapping the appropriate aperture address into |
1800 | * userspace. |
1821 | * userspace. |
1801 | */ |
1822 | */ |
1802 | int |
1823 | int |
1803 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, |
1824 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, |
1804 | struct drm_file *file) |
1825 | struct drm_file *file) |
1805 | { |
1826 | { |
1806 | struct drm_i915_gem_mmap_gtt *args = data; |
1827 | struct drm_i915_gem_mmap_gtt *args = data; |
1807 | 1828 | ||
1808 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); |
1829 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); |
1809 | } |
1830 | } |
1810 | 1831 | ||
1811 | /* Immediately discard the backing storage */ |
1832 | /* Immediately discard the backing storage */ |
1812 | static void |
1833 | static void |
1813 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) |
1834 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) |
1814 | { |
1835 | { |
1815 | // i915_gem_object_free_mmap_offset(obj); |
1836 | // i915_gem_object_free_mmap_offset(obj); |
1816 | 1837 | ||
1817 | if (obj->base.filp == NULL) |
1838 | if (obj->base.filp == NULL) |
1818 | return; |
1839 | return; |
1819 | 1840 | ||
1820 | /* Our goal here is to return as much of the memory as |
1841 | /* Our goal here is to return as much of the memory as |
1821 | * is possible back to the system as we are called from OOM. |
1842 | * is possible back to the system as we are called from OOM. |
1822 | * To do this we must instruct the shmfs to drop all of its |
1843 | * To do this we must instruct the shmfs to drop all of its |
1823 | * backing pages, *now*. |
1844 | * backing pages, *now*. |
1824 | */ |
1845 | */ |
1825 | // shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); |
1846 | // shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); |
1826 | obj->madv = __I915_MADV_PURGED; |
1847 | obj->madv = __I915_MADV_PURGED; |
1827 | } |
1848 | } |
1828 | 1849 | ||
1829 | /* Try to discard unwanted pages */ |
1850 | /* Try to discard unwanted pages */ |
1830 | static void |
1851 | static void |
1831 | i915_gem_object_invalidate(struct drm_i915_gem_object *obj) |
1852 | i915_gem_object_invalidate(struct drm_i915_gem_object *obj) |
1832 | { |
1853 | { |
1833 | struct address_space *mapping; |
1854 | struct address_space *mapping; |
1834 | 1855 | ||
1835 | switch (obj->madv) { |
1856 | switch (obj->madv) { |
1836 | case I915_MADV_DONTNEED: |
1857 | case I915_MADV_DONTNEED: |
1837 | i915_gem_object_truncate(obj); |
1858 | i915_gem_object_truncate(obj); |
1838 | case __I915_MADV_PURGED: |
1859 | case __I915_MADV_PURGED: |
1839 | return; |
1860 | return; |
1840 | } |
1861 | } |
1841 | 1862 | ||
1842 | if (obj->base.filp == NULL) |
1863 | if (obj->base.filp == NULL) |
1843 | return; |
1864 | return; |
1844 | 1865 | ||
1845 | } |
1866 | } |
1846 | 1867 | ||
1847 | static void |
1868 | static void |
1848 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj) |
1869 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj) |
1849 | { |
1870 | { |
1850 | struct sg_page_iter sg_iter; |
1871 | struct sg_page_iter sg_iter; |
1851 | int ret; |
1872 | int ret; |
1852 | 1873 | ||
1853 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
1874 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
1854 | 1875 | ||
1855 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
1876 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
1856 | if (ret) { |
1877 | if (ret) { |
1857 | /* In the event of a disaster, abandon all caches and |
1878 | /* In the event of a disaster, abandon all caches and |
1858 | * hope for the best. |
1879 | * hope for the best. |
1859 | */ |
1880 | */ |
1860 | WARN_ON(ret != -EIO); |
1881 | WARN_ON(ret != -EIO); |
1861 | i915_gem_clflush_object(obj, true); |
1882 | i915_gem_clflush_object(obj, true); |
1862 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
1883 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
1863 | } |
1884 | } |
1864 | 1885 | ||
1865 | i915_gem_gtt_finish_object(obj); |
1886 | i915_gem_gtt_finish_object(obj); |
1866 | 1887 | ||
1867 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
1888 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
1868 | i915_gem_object_save_bit_17_swizzle(obj); |
1889 | i915_gem_object_save_bit_17_swizzle(obj); |
1869 | 1890 | ||
1870 | if (obj->madv == I915_MADV_DONTNEED) |
1891 | if (obj->madv == I915_MADV_DONTNEED) |
1871 | obj->dirty = 0; |
1892 | obj->dirty = 0; |
1872 | 1893 | ||
1873 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { |
1894 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { |
1874 | struct page *page = sg_page_iter_page(&sg_iter); |
1895 | struct page *page = sg_page_iter_page(&sg_iter); |
1875 | 1896 | ||
1876 | page_cache_release(page); |
1897 | page_cache_release(page); |
1877 | } |
1898 | } |
1878 | obj->dirty = 0; |
1899 | obj->dirty = 0; |
1879 | 1900 | ||
1880 | sg_free_table(obj->pages); |
1901 | sg_free_table(obj->pages); |
1881 | kfree(obj->pages); |
1902 | kfree(obj->pages); |
1882 | } |
1903 | } |
1883 | 1904 | ||
1884 | int |
1905 | int |
1885 | i915_gem_object_put_pages(struct drm_i915_gem_object *obj) |
1906 | i915_gem_object_put_pages(struct drm_i915_gem_object *obj) |
1886 | { |
1907 | { |
1887 | const struct drm_i915_gem_object_ops *ops = obj->ops; |
1908 | const struct drm_i915_gem_object_ops *ops = obj->ops; |
1888 | 1909 | ||
1889 | if (obj->pages == NULL) |
1910 | if (obj->pages == NULL) |
1890 | return 0; |
1911 | return 0; |
1891 | 1912 | ||
1892 | if (obj->pages_pin_count) |
1913 | if (obj->pages_pin_count) |
1893 | return -EBUSY; |
1914 | return -EBUSY; |
1894 | 1915 | ||
1895 | BUG_ON(i915_gem_obj_bound_any(obj)); |
1916 | BUG_ON(i915_gem_obj_bound_any(obj)); |
1896 | 1917 | ||
1897 | /* ->put_pages might need to allocate memory for the bit17 swizzle |
1918 | /* ->put_pages might need to allocate memory for the bit17 swizzle |
1898 | * array, hence protect them from being reaped by removing them from gtt |
1919 | * array, hence protect them from being reaped by removing them from gtt |
1899 | * lists early. */ |
1920 | * lists early. */ |
1900 | list_del(&obj->global_list); |
1921 | list_del(&obj->global_list); |
1901 | 1922 | ||
1902 | ops->put_pages(obj); |
1923 | ops->put_pages(obj); |
1903 | obj->pages = NULL; |
1924 | obj->pages = NULL; |
1904 | 1925 | ||
1905 | i915_gem_object_invalidate(obj); |
1926 | i915_gem_object_invalidate(obj); |
1906 | 1927 | ||
1907 | return 0; |
1928 | return 0; |
1908 | } |
1929 | } |
1909 | 1930 | ||
1910 | static int |
1931 | static int |
1911 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj) |
1932 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj) |
1912 | { |
1933 | { |
1913 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
1934 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
1914 | int page_count, i; |
1935 | int page_count, i; |
1915 | struct address_space *mapping; |
1936 | struct address_space *mapping; |
1916 | struct sg_table *st; |
1937 | struct sg_table *st; |
1917 | struct scatterlist *sg; |
1938 | struct scatterlist *sg; |
1918 | struct sg_page_iter sg_iter; |
1939 | struct sg_page_iter sg_iter; |
1919 | struct page *page; |
1940 | struct page *page; |
1920 | unsigned long last_pfn = 0; /* suppress gcc warning */ |
1941 | unsigned long last_pfn = 0; /* suppress gcc warning */ |
1921 | int ret; |
1942 | int ret; |
1922 | gfp_t gfp; |
1943 | gfp_t gfp; |
1923 | 1944 | ||
1924 | /* Assert that the object is not currently in any GPU domain. As it |
1945 | /* Assert that the object is not currently in any GPU domain. As it |
1925 | * wasn't in the GTT, there shouldn't be any way it could have been in |
1946 | * wasn't in the GTT, there shouldn't be any way it could have been in |
1926 | * a GPU cache |
1947 | * a GPU cache |
1927 | */ |
1948 | */ |
1928 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); |
1949 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); |
1929 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); |
1950 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); |
1930 | 1951 | ||
1931 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
1952 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
1932 | if (st == NULL) |
1953 | if (st == NULL) |
1933 | return -ENOMEM; |
1954 | return -ENOMEM; |
1934 | 1955 | ||
1935 | page_count = obj->base.size / PAGE_SIZE; |
1956 | page_count = obj->base.size / PAGE_SIZE; |
1936 | if (sg_alloc_table(st, page_count, GFP_KERNEL)) { |
1957 | if (sg_alloc_table(st, page_count, GFP_KERNEL)) { |
1937 | kfree(st); |
1958 | kfree(st); |
1938 | return -ENOMEM; |
1959 | return -ENOMEM; |
1939 | } |
1960 | } |
1940 | 1961 | ||
1941 | /* Get the list of pages out of our struct file. They'll be pinned |
1962 | /* Get the list of pages out of our struct file. They'll be pinned |
1942 | * at this point until we release them. |
1963 | * at this point until we release them. |
1943 | * |
1964 | * |
1944 | * Fail silently without starting the shrinker |
1965 | * Fail silently without starting the shrinker |
1945 | */ |
1966 | */ |
1946 | sg = st->sgl; |
1967 | sg = st->sgl; |
1947 | st->nents = 0; |
1968 | st->nents = 0; |
1948 | for (i = 0; i < page_count; i++) { |
1969 | for (i = 0; i < page_count; i++) { |
1949 | page = shmem_read_mapping_page_gfp(obj->base.filp, i, gfp); |
1970 | page = shmem_read_mapping_page_gfp(obj->base.filp, i, gfp); |
1950 | if (IS_ERR(page)) { |
1971 | if (IS_ERR(page)) { |
1951 | dbgprintf("%s invalid page %p\n", __FUNCTION__, page); |
1972 | dbgprintf("%s invalid page %p\n", __FUNCTION__, page); |
1952 | goto err_pages; |
1973 | goto err_pages; |
1953 | } |
1974 | } |
1954 | #ifdef CONFIG_SWIOTLB |
1975 | #ifdef CONFIG_SWIOTLB |
1955 | if (swiotlb_nr_tbl()) { |
1976 | if (swiotlb_nr_tbl()) { |
1956 | st->nents++; |
1977 | st->nents++; |
1957 | sg_set_page(sg, page, PAGE_SIZE, 0); |
1978 | sg_set_page(sg, page, PAGE_SIZE, 0); |
1958 | sg = sg_next(sg); |
1979 | sg = sg_next(sg); |
1959 | continue; |
1980 | continue; |
1960 | } |
1981 | } |
1961 | #endif |
1982 | #endif |
1962 | if (!i || page_to_pfn(page) != last_pfn + 1) { |
1983 | if (!i || page_to_pfn(page) != last_pfn + 1) { |
1963 | if (i) |
1984 | if (i) |
1964 | sg = sg_next(sg); |
1985 | sg = sg_next(sg); |
1965 | st->nents++; |
1986 | st->nents++; |
1966 | sg_set_page(sg, page, PAGE_SIZE, 0); |
1987 | sg_set_page(sg, page, PAGE_SIZE, 0); |
1967 | } else { |
1988 | } else { |
1968 | sg->length += PAGE_SIZE; |
1989 | sg->length += PAGE_SIZE; |
1969 | } |
1990 | } |
1970 | last_pfn = page_to_pfn(page); |
1991 | last_pfn = page_to_pfn(page); |
1971 | 1992 | ||
1972 | /* Check that the i965g/gm workaround works. */ |
1993 | /* Check that the i965g/gm workaround works. */ |
1973 | WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL)); |
1994 | WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL)); |
1974 | } |
1995 | } |
1975 | #ifdef CONFIG_SWIOTLB |
1996 | #ifdef CONFIG_SWIOTLB |
1976 | if (!swiotlb_nr_tbl()) |
1997 | if (!swiotlb_nr_tbl()) |
1977 | #endif |
1998 | #endif |
1978 | sg_mark_end(sg); |
1999 | sg_mark_end(sg); |
1979 | obj->pages = st; |
2000 | obj->pages = st; |
1980 | 2001 | ||
1981 | ret = i915_gem_gtt_prepare_object(obj); |
2002 | ret = i915_gem_gtt_prepare_object(obj); |
1982 | if (ret) |
2003 | if (ret) |
1983 | goto err_pages; |
2004 | goto err_pages; |
1984 | 2005 | ||
1985 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
2006 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
1986 | i915_gem_object_do_bit_17_swizzle(obj); |
2007 | i915_gem_object_do_bit_17_swizzle(obj); |
1987 | 2008 | ||
1988 | if (obj->tiling_mode != I915_TILING_NONE && |
2009 | if (obj->tiling_mode != I915_TILING_NONE && |
1989 | dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) |
2010 | dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) |
1990 | i915_gem_object_pin_pages(obj); |
2011 | i915_gem_object_pin_pages(obj); |
1991 | 2012 | ||
1992 | return 0; |
2013 | return 0; |
1993 | 2014 | ||
1994 | err_pages: |
2015 | err_pages: |
1995 | sg_mark_end(sg); |
2016 | sg_mark_end(sg); |
1996 | for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) |
2017 | for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) |
1997 | page_cache_release(sg_page_iter_page(&sg_iter)); |
2018 | page_cache_release(sg_page_iter_page(&sg_iter)); |
1998 | sg_free_table(st); |
2019 | sg_free_table(st); |
1999 | kfree(st); |
2020 | kfree(st); |
2000 | 2021 | ||
2001 | return PTR_ERR(page); |
2022 | return PTR_ERR(page); |
2002 | } |
2023 | } |
2003 | 2024 | ||
2004 | /* Ensure that the associated pages are gathered from the backing storage |
2025 | /* Ensure that the associated pages are gathered from the backing storage |
2005 | * and pinned into our object. i915_gem_object_get_pages() may be called |
2026 | * and pinned into our object. i915_gem_object_get_pages() may be called |
2006 | * multiple times before they are released by a single call to |
2027 | * multiple times before they are released by a single call to |
2007 | * i915_gem_object_put_pages() - once the pages are no longer referenced |
2028 | * i915_gem_object_put_pages() - once the pages are no longer referenced |
2008 | * either as a result of memory pressure (reaping pages under the shrinker) |
2029 | * either as a result of memory pressure (reaping pages under the shrinker) |
2009 | * or as the object is itself released. |
2030 | * or as the object is itself released. |
2010 | */ |
2031 | */ |
2011 | int |
2032 | int |
2012 | i915_gem_object_get_pages(struct drm_i915_gem_object *obj) |
2033 | i915_gem_object_get_pages(struct drm_i915_gem_object *obj) |
2013 | { |
2034 | { |
2014 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
2035 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
2015 | const struct drm_i915_gem_object_ops *ops = obj->ops; |
2036 | const struct drm_i915_gem_object_ops *ops = obj->ops; |
2016 | int ret; |
2037 | int ret; |
2017 | 2038 | ||
2018 | if (obj->pages) |
2039 | if (obj->pages) |
2019 | return 0; |
2040 | return 0; |
2020 | 2041 | ||
2021 | if (obj->madv != I915_MADV_WILLNEED) { |
2042 | if (obj->madv != I915_MADV_WILLNEED) { |
2022 | DRM_DEBUG("Attempting to obtain a purgeable object\n"); |
2043 | DRM_DEBUG("Attempting to obtain a purgeable object\n"); |
2023 | return -EFAULT; |
2044 | return -EFAULT; |
2024 | } |
2045 | } |
2025 | 2046 | ||
2026 | BUG_ON(obj->pages_pin_count); |
2047 | BUG_ON(obj->pages_pin_count); |
2027 | 2048 | ||
2028 | ret = ops->get_pages(obj); |
2049 | ret = ops->get_pages(obj); |
2029 | if (ret) |
2050 | if (ret) |
2030 | return ret; |
2051 | return ret; |
2031 | 2052 | ||
2032 | list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
2053 | list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
2033 | 2054 | ||
2034 | obj->get_page.sg = obj->pages->sgl; |
2055 | obj->get_page.sg = obj->pages->sgl; |
2035 | obj->get_page.last = 0; |
2056 | obj->get_page.last = 0; |
2036 | 2057 | ||
2037 | return 0; |
2058 | return 0; |
2038 | } |
2059 | } |
2039 | 2060 | ||
2040 | void i915_vma_move_to_active(struct i915_vma *vma, |
2061 | void i915_vma_move_to_active(struct i915_vma *vma, |
2041 | struct drm_i915_gem_request *req) |
2062 | struct drm_i915_gem_request *req) |
2042 | { |
2063 | { |
2043 | struct drm_i915_gem_object *obj = vma->obj; |
2064 | struct drm_i915_gem_object *obj = vma->obj; |
2044 | struct intel_engine_cs *ring; |
2065 | struct intel_engine_cs *ring; |
2045 | 2066 | ||
2046 | ring = i915_gem_request_get_ring(req); |
2067 | ring = i915_gem_request_get_ring(req); |
2047 | 2068 | ||
2048 | /* Add a reference if we're newly entering the active list. */ |
2069 | /* Add a reference if we're newly entering the active list. */ |
2049 | if (obj->active == 0) |
2070 | if (obj->active == 0) |
2050 | drm_gem_object_reference(&obj->base); |
2071 | drm_gem_object_reference(&obj->base); |
2051 | obj->active |= intel_ring_flag(ring); |
2072 | obj->active |= intel_ring_flag(ring); |
2052 | 2073 | ||
2053 | list_move_tail(&obj->ring_list[ring->id], &ring->active_list); |
2074 | list_move_tail(&obj->ring_list[ring->id], &ring->active_list); |
2054 | i915_gem_request_assign(&obj->last_read_req[ring->id], req); |
2075 | i915_gem_request_assign(&obj->last_read_req[ring->id], req); |
2055 | 2076 | ||
2056 | list_move_tail(&vma->mm_list, &vma->vm->active_list); |
2077 | list_move_tail(&vma->vm_link, &vma->vm->active_list); |
2057 | } |
2078 | } |
2058 | 2079 | ||
2059 | static void |
2080 | static void |
2060 | i915_gem_object_retire__write(struct drm_i915_gem_object *obj) |
2081 | i915_gem_object_retire__write(struct drm_i915_gem_object *obj) |
2061 | { |
2082 | { |
2062 | RQ_BUG_ON(obj->last_write_req == NULL); |
2083 | RQ_BUG_ON(obj->last_write_req == NULL); |
2063 | RQ_BUG_ON(!(obj->active & intel_ring_flag(obj->last_write_req->ring))); |
2084 | RQ_BUG_ON(!(obj->active & intel_ring_flag(obj->last_write_req->ring))); |
2064 | 2085 | ||
2065 | i915_gem_request_assign(&obj->last_write_req, NULL); |
2086 | i915_gem_request_assign(&obj->last_write_req, NULL); |
2066 | intel_fb_obj_flush(obj, true, ORIGIN_CS); |
2087 | intel_fb_obj_flush(obj, true, ORIGIN_CS); |
2067 | } |
2088 | } |
2068 | 2089 | ||
2069 | static void |
2090 | static void |
2070 | i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring) |
2091 | i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring) |
2071 | { |
2092 | { |
2072 | struct i915_vma *vma; |
2093 | struct i915_vma *vma; |
2073 | 2094 | ||
2074 | RQ_BUG_ON(obj->last_read_req[ring] == NULL); |
2095 | RQ_BUG_ON(obj->last_read_req[ring] == NULL); |
2075 | RQ_BUG_ON(!(obj->active & (1 << ring))); |
2096 | RQ_BUG_ON(!(obj->active & (1 << ring))); |
2076 | 2097 | ||
2077 | list_del_init(&obj->ring_list[ring]); |
2098 | list_del_init(&obj->ring_list[ring]); |
2078 | i915_gem_request_assign(&obj->last_read_req[ring], NULL); |
2099 | i915_gem_request_assign(&obj->last_read_req[ring], NULL); |
2079 | 2100 | ||
2080 | if (obj->last_write_req && obj->last_write_req->ring->id == ring) |
2101 | if (obj->last_write_req && obj->last_write_req->ring->id == ring) |
2081 | i915_gem_object_retire__write(obj); |
2102 | i915_gem_object_retire__write(obj); |
2082 | 2103 | ||
2083 | obj->active &= ~(1 << ring); |
2104 | obj->active &= ~(1 << ring); |
2084 | if (obj->active) |
2105 | if (obj->active) |
2085 | return; |
2106 | return; |
2086 | 2107 | ||
2087 | /* Bump our place on the bound list to keep it roughly in LRU order |
2108 | /* Bump our place on the bound list to keep it roughly in LRU order |
2088 | * so that we don't steal from recently used but inactive objects |
2109 | * so that we don't steal from recently used but inactive objects |
2089 | * (unless we are forced to ofc!) |
2110 | * (unless we are forced to ofc!) |
2090 | */ |
2111 | */ |
2091 | list_move_tail(&obj->global_list, |
2112 | list_move_tail(&obj->global_list, |
2092 | &to_i915(obj->base.dev)->mm.bound_list); |
2113 | &to_i915(obj->base.dev)->mm.bound_list); |
2093 | 2114 | ||
2094 | list_for_each_entry(vma, &obj->vma_list, vma_link) { |
2115 | list_for_each_entry(vma, &obj->vma_list, obj_link) { |
2095 | if (!list_empty(&vma->mm_list)) |
2116 | if (!list_empty(&vma->vm_link)) |
2096 | list_move_tail(&vma->mm_list, &vma->vm->inactive_list); |
2117 | list_move_tail(&vma->vm_link, &vma->vm->inactive_list); |
2097 | } |
2118 | } |
2098 | 2119 | ||
2099 | i915_gem_request_assign(&obj->last_fenced_req, NULL); |
2120 | i915_gem_request_assign(&obj->last_fenced_req, NULL); |
2100 | drm_gem_object_unreference(&obj->base); |
2121 | drm_gem_object_unreference(&obj->base); |
2101 | } |
2122 | } |
2102 | 2123 | ||
2103 | static int |
2124 | static int |
2104 | i915_gem_init_seqno(struct drm_device *dev, u32 seqno) |
2125 | i915_gem_init_seqno(struct drm_device *dev, u32 seqno) |
2105 | { |
2126 | { |
2106 | struct drm_i915_private *dev_priv = dev->dev_private; |
2127 | struct drm_i915_private *dev_priv = dev->dev_private; |
2107 | struct intel_engine_cs *ring; |
2128 | struct intel_engine_cs *ring; |
2108 | int ret, i, j; |
2129 | int ret, i, j; |
2109 | 2130 | ||
2110 | /* Carefully retire all requests without writing to the rings */ |
2131 | /* Carefully retire all requests without writing to the rings */ |
2111 | for_each_ring(ring, dev_priv, i) { |
2132 | for_each_ring(ring, dev_priv, i) { |
2112 | ret = intel_ring_idle(ring); |
2133 | ret = intel_ring_idle(ring); |
2113 | if (ret) |
2134 | if (ret) |
2114 | return ret; |
2135 | return ret; |
2115 | } |
2136 | } |
2116 | i915_gem_retire_requests(dev); |
2137 | i915_gem_retire_requests(dev); |
2117 | 2138 | ||
2118 | /* Finally reset hw state */ |
2139 | /* Finally reset hw state */ |
2119 | for_each_ring(ring, dev_priv, i) { |
2140 | for_each_ring(ring, dev_priv, i) { |
2120 | intel_ring_init_seqno(ring, seqno); |
2141 | intel_ring_init_seqno(ring, seqno); |
2121 | 2142 | ||
2122 | for (j = 0; j < ARRAY_SIZE(ring->semaphore.sync_seqno); j++) |
2143 | for (j = 0; j < ARRAY_SIZE(ring->semaphore.sync_seqno); j++) |
2123 | ring->semaphore.sync_seqno[j] = 0; |
2144 | ring->semaphore.sync_seqno[j] = 0; |
2124 | } |
2145 | } |
2125 | 2146 | ||
2126 | return 0; |
2147 | return 0; |
2127 | } |
2148 | } |
2128 | 2149 | ||
2129 | int i915_gem_set_seqno(struct drm_device *dev, u32 seqno) |
2150 | int i915_gem_set_seqno(struct drm_device *dev, u32 seqno) |
2130 | { |
2151 | { |
2131 | struct drm_i915_private *dev_priv = dev->dev_private; |
2152 | struct drm_i915_private *dev_priv = dev->dev_private; |
2132 | int ret; |
2153 | int ret; |
2133 | 2154 | ||
2134 | if (seqno == 0) |
2155 | if (seqno == 0) |
2135 | return -EINVAL; |
2156 | return -EINVAL; |
2136 | 2157 | ||
2137 | /* HWS page needs to be set less than what we |
2158 | /* HWS page needs to be set less than what we |
2138 | * will inject to ring |
2159 | * will inject to ring |
2139 | */ |
2160 | */ |
2140 | ret = i915_gem_init_seqno(dev, seqno - 1); |
2161 | ret = i915_gem_init_seqno(dev, seqno - 1); |
2141 | if (ret) |
2162 | if (ret) |
2142 | return ret; |
2163 | return ret; |
2143 | 2164 | ||
2144 | /* Carefully set the last_seqno value so that wrap |
2165 | /* Carefully set the last_seqno value so that wrap |
2145 | * detection still works |
2166 | * detection still works |
2146 | */ |
2167 | */ |
2147 | dev_priv->next_seqno = seqno; |
2168 | dev_priv->next_seqno = seqno; |
2148 | dev_priv->last_seqno = seqno - 1; |
2169 | dev_priv->last_seqno = seqno - 1; |
2149 | if (dev_priv->last_seqno == 0) |
2170 | if (dev_priv->last_seqno == 0) |
2150 | dev_priv->last_seqno--; |
2171 | dev_priv->last_seqno--; |
2151 | 2172 | ||
2152 | return 0; |
2173 | return 0; |
2153 | } |
2174 | } |
2154 | 2175 | ||
2155 | int |
2176 | int |
2156 | i915_gem_get_seqno(struct drm_device *dev, u32 *seqno) |
2177 | i915_gem_get_seqno(struct drm_device *dev, u32 *seqno) |
2157 | { |
2178 | { |
2158 | struct drm_i915_private *dev_priv = dev->dev_private; |
2179 | struct drm_i915_private *dev_priv = dev->dev_private; |
2159 | 2180 | ||
2160 | /* reserve 0 for non-seqno */ |
2181 | /* reserve 0 for non-seqno */ |
2161 | if (dev_priv->next_seqno == 0) { |
2182 | if (dev_priv->next_seqno == 0) { |
2162 | int ret = i915_gem_init_seqno(dev, 0); |
2183 | int ret = i915_gem_init_seqno(dev, 0); |
2163 | if (ret) |
2184 | if (ret) |
2164 | return ret; |
2185 | return ret; |
2165 | 2186 | ||
2166 | dev_priv->next_seqno = 1; |
2187 | dev_priv->next_seqno = 1; |
2167 | } |
2188 | } |
2168 | 2189 | ||
2169 | *seqno = dev_priv->last_seqno = dev_priv->next_seqno++; |
2190 | *seqno = dev_priv->last_seqno = dev_priv->next_seqno++; |
2170 | return 0; |
2191 | return 0; |
2171 | } |
2192 | } |
2172 | 2193 | ||
2173 | /* |
2194 | /* |
2174 | * NB: This function is not allowed to fail. Doing so would mean the the |
2195 | * NB: This function is not allowed to fail. Doing so would mean the the |
2175 | * request is not being tracked for completion but the work itself is |
2196 | * request is not being tracked for completion but the work itself is |
2176 | * going to happen on the hardware. This would be a Bad Thing(tm). |
2197 | * going to happen on the hardware. This would be a Bad Thing(tm). |
2177 | */ |
2198 | */ |
2178 | void __i915_add_request(struct drm_i915_gem_request *request, |
2199 | void __i915_add_request(struct drm_i915_gem_request *request, |
2179 | struct drm_i915_gem_object *obj, |
2200 | struct drm_i915_gem_object *obj, |
2180 | bool flush_caches) |
2201 | bool flush_caches) |
2181 | { |
2202 | { |
2182 | struct intel_engine_cs *ring; |
2203 | struct intel_engine_cs *ring; |
2183 | struct drm_i915_private *dev_priv; |
2204 | struct drm_i915_private *dev_priv; |
2184 | struct intel_ringbuffer *ringbuf; |
2205 | struct intel_ringbuffer *ringbuf; |
2185 | u32 request_start; |
2206 | u32 request_start; |
2186 | int ret; |
2207 | int ret; |
2187 | 2208 | ||
2188 | if (WARN_ON(request == NULL)) |
2209 | if (WARN_ON(request == NULL)) |
2189 | return; |
2210 | return; |
2190 | 2211 | ||
2191 | ring = request->ring; |
2212 | ring = request->ring; |
2192 | dev_priv = ring->dev->dev_private; |
2213 | dev_priv = ring->dev->dev_private; |
2193 | ringbuf = request->ringbuf; |
2214 | ringbuf = request->ringbuf; |
2194 | 2215 | ||
2195 | /* |
2216 | /* |
2196 | * To ensure that this call will not fail, space for its emissions |
2217 | * To ensure that this call will not fail, space for its emissions |
2197 | * should already have been reserved in the ring buffer. Let the ring |
2218 | * should already have been reserved in the ring buffer. Let the ring |
2198 | * know that it is time to use that space up. |
2219 | * know that it is time to use that space up. |
2199 | */ |
2220 | */ |
2200 | intel_ring_reserved_space_use(ringbuf); |
2221 | intel_ring_reserved_space_use(ringbuf); |
2201 | 2222 | ||
2202 | request_start = intel_ring_get_tail(ringbuf); |
2223 | request_start = intel_ring_get_tail(ringbuf); |
2203 | /* |
2224 | /* |
2204 | * Emit any outstanding flushes - execbuf can fail to emit the flush |
2225 | * Emit any outstanding flushes - execbuf can fail to emit the flush |
2205 | * after having emitted the batchbuffer command. Hence we need to fix |
2226 | * after having emitted the batchbuffer command. Hence we need to fix |
2206 | * things up similar to emitting the lazy request. The difference here |
2227 | * things up similar to emitting the lazy request. The difference here |
2207 | * is that the flush _must_ happen before the next request, no matter |
2228 | * is that the flush _must_ happen before the next request, no matter |
2208 | * what. |
2229 | * what. |
2209 | */ |
2230 | */ |
2210 | if (flush_caches) { |
2231 | if (flush_caches) { |
2211 | if (i915.enable_execlists) |
2232 | if (i915.enable_execlists) |
2212 | ret = logical_ring_flush_all_caches(request); |
2233 | ret = logical_ring_flush_all_caches(request); |
2213 | else |
2234 | else |
2214 | ret = intel_ring_flush_all_caches(request); |
2235 | ret = intel_ring_flush_all_caches(request); |
2215 | /* Not allowed to fail! */ |
2236 | /* Not allowed to fail! */ |
2216 | WARN(ret, "*_ring_flush_all_caches failed: %d!\n", ret); |
2237 | WARN(ret, "*_ring_flush_all_caches failed: %d!\n", ret); |
2217 | } |
2238 | } |
2218 | 2239 | ||
2219 | /* Record the position of the start of the request so that |
2240 | /* Record the position of the start of the request so that |
2220 | * should we detect the updated seqno part-way through the |
2241 | * should we detect the updated seqno part-way through the |
2221 | * GPU processing the request, we never over-estimate the |
2242 | * GPU processing the request, we never over-estimate the |
2222 | * position of the head. |
2243 | * position of the head. |
2223 | */ |
2244 | */ |
2224 | request->postfix = intel_ring_get_tail(ringbuf); |
2245 | request->postfix = intel_ring_get_tail(ringbuf); |
2225 | 2246 | ||
2226 | if (i915.enable_execlists) |
2247 | if (i915.enable_execlists) |
2227 | ret = ring->emit_request(request); |
2248 | ret = ring->emit_request(request); |
2228 | else { |
2249 | else { |
2229 | ret = ring->add_request(request); |
2250 | ret = ring->add_request(request); |
2230 | 2251 | ||
2231 | request->tail = intel_ring_get_tail(ringbuf); |
2252 | request->tail = intel_ring_get_tail(ringbuf); |
2232 | } |
2253 | } |
2233 | /* Not allowed to fail! */ |
2254 | /* Not allowed to fail! */ |
2234 | WARN(ret, "emit|add_request failed: %d!\n", ret); |
2255 | WARN(ret, "emit|add_request failed: %d!\n", ret); |
2235 | 2256 | ||
2236 | request->head = request_start; |
2257 | request->head = request_start; |
2237 | 2258 | ||
2238 | /* Whilst this request exists, batch_obj will be on the |
2259 | /* Whilst this request exists, batch_obj will be on the |
2239 | * active_list, and so will hold the active reference. Only when this |
2260 | * active_list, and so will hold the active reference. Only when this |
2240 | * request is retired will the the batch_obj be moved onto the |
2261 | * request is retired will the the batch_obj be moved onto the |
2241 | * inactive_list and lose its active reference. Hence we do not need |
2262 | * inactive_list and lose its active reference. Hence we do not need |
2242 | * to explicitly hold another reference here. |
2263 | * to explicitly hold another reference here. |
2243 | */ |
2264 | */ |
2244 | request->batch_obj = obj; |
2265 | request->batch_obj = obj; |
2245 | 2266 | ||
2246 | request->emitted_jiffies = jiffies; |
2267 | request->emitted_jiffies = jiffies; |
2247 | request->previous_seqno = ring->last_submitted_seqno; |
2268 | request->previous_seqno = ring->last_submitted_seqno; |
2248 | ring->last_submitted_seqno = request->seqno; |
2269 | ring->last_submitted_seqno = request->seqno; |
2249 | list_add_tail(&request->list, &ring->request_list); |
2270 | list_add_tail(&request->list, &ring->request_list); |
2250 | 2271 | ||
2251 | trace_i915_gem_request_add(request); |
2272 | trace_i915_gem_request_add(request); |
2252 | 2273 | ||
2253 | // i915_queue_hangcheck(ring->dev); |
2274 | i915_queue_hangcheck(ring->dev); |
2254 | 2275 | ||
2255 | queue_delayed_work(dev_priv->wq, |
2276 | queue_delayed_work(dev_priv->wq, |
2256 | &dev_priv->mm.retire_work, |
2277 | &dev_priv->mm.retire_work, |
2257 | round_jiffies_up_relative(HZ)); |
2278 | round_jiffies_up_relative(HZ)); |
2258 | intel_mark_busy(dev_priv->dev); |
2279 | intel_mark_busy(dev_priv->dev); |
2259 | 2280 | ||
2260 | /* Sanity check that the reserved size was large enough. */ |
2281 | /* Sanity check that the reserved size was large enough. */ |
2261 | intel_ring_reserved_space_end(ringbuf); |
2282 | intel_ring_reserved_space_end(ringbuf); |
2262 | } |
2283 | } |
2263 | 2284 | ||
2264 | static bool i915_context_is_banned(struct drm_i915_private *dev_priv, |
2285 | static bool i915_context_is_banned(struct drm_i915_private *dev_priv, |
2265 | const struct intel_context *ctx) |
2286 | const struct intel_context *ctx) |
2266 | { |
2287 | { |
2267 | unsigned long elapsed; |
2288 | unsigned long elapsed; |
2268 | 2289 | ||
2269 | elapsed = GetTimerTicks()/100 - ctx->hang_stats.guilty_ts; |
2290 | elapsed = GetTimerTicks()/100 - ctx->hang_stats.guilty_ts; |
2270 | 2291 | ||
2271 | if (ctx->hang_stats.banned) |
2292 | if (ctx->hang_stats.banned) |
2272 | return true; |
2293 | return true; |
2273 | 2294 | ||
2274 | if (ctx->hang_stats.ban_period_seconds && |
2295 | if (ctx->hang_stats.ban_period_seconds && |
2275 | elapsed <= ctx->hang_stats.ban_period_seconds) { |
2296 | elapsed <= ctx->hang_stats.ban_period_seconds) { |
2276 | if (!i915_gem_context_is_default(ctx)) { |
2297 | if (!i915_gem_context_is_default(ctx)) { |
2277 | DRM_DEBUG("context hanging too fast, banning!\n"); |
2298 | DRM_DEBUG("context hanging too fast, banning!\n"); |
2278 | return true; |
2299 | return true; |
2279 | } else if (i915_stop_ring_allow_ban(dev_priv)) { |
2300 | } else if (i915_stop_ring_allow_ban(dev_priv)) { |
2280 | if (i915_stop_ring_allow_warn(dev_priv)) |
2301 | if (i915_stop_ring_allow_warn(dev_priv)) |
2281 | DRM_ERROR("gpu hanging too fast, banning!\n"); |
2302 | DRM_ERROR("gpu hanging too fast, banning!\n"); |
2282 | return true; |
2303 | return true; |
2283 | } |
2304 | } |
2284 | } |
2305 | } |
2285 | 2306 | ||
2286 | return false; |
2307 | return false; |
2287 | } |
2308 | } |
2288 | 2309 | ||
2289 | static void i915_set_reset_status(struct drm_i915_private *dev_priv, |
2310 | static void i915_set_reset_status(struct drm_i915_private *dev_priv, |
2290 | struct intel_context *ctx, |
2311 | struct intel_context *ctx, |
2291 | const bool guilty) |
2312 | const bool guilty) |
2292 | { |
2313 | { |
2293 | struct i915_ctx_hang_stats *hs; |
2314 | struct i915_ctx_hang_stats *hs; |
2294 | 2315 | ||
2295 | if (WARN_ON(!ctx)) |
2316 | if (WARN_ON(!ctx)) |
2296 | return; |
2317 | return; |
2297 | 2318 | ||
2298 | hs = &ctx->hang_stats; |
2319 | hs = &ctx->hang_stats; |
2299 | 2320 | ||
2300 | if (guilty) { |
2321 | if (guilty) { |
2301 | hs->banned = i915_context_is_banned(dev_priv, ctx); |
2322 | hs->banned = i915_context_is_banned(dev_priv, ctx); |
2302 | hs->batch_active++; |
2323 | hs->batch_active++; |
2303 | hs->guilty_ts = GetTimerTicks()/100; |
2324 | hs->guilty_ts = GetTimerTicks()/100; |
2304 | } else { |
2325 | } else { |
2305 | hs->batch_pending++; |
2326 | hs->batch_pending++; |
2306 | } |
2327 | } |
2307 | } |
2328 | } |
2308 | 2329 | ||
2309 | void i915_gem_request_free(struct kref *req_ref) |
2330 | void i915_gem_request_free(struct kref *req_ref) |
2310 | { |
2331 | { |
2311 | struct drm_i915_gem_request *req = container_of(req_ref, |
2332 | struct drm_i915_gem_request *req = container_of(req_ref, |
2312 | typeof(*req), ref); |
2333 | typeof(*req), ref); |
2313 | struct intel_context *ctx = req->ctx; |
2334 | struct intel_context *ctx = req->ctx; |
2314 | 2335 | ||
2315 | if (req->file_priv) |
2336 | if (req->file_priv) |
2316 | i915_gem_request_remove_from_client(req); |
2337 | i915_gem_request_remove_from_client(req); |
2317 | 2338 | ||
2318 | if (ctx) { |
2339 | if (ctx) { |
2319 | if (i915.enable_execlists) { |
- | |
2320 | if (ctx != req->ring->default_context) |
2340 | if (i915.enable_execlists && ctx != req->i915->kernel_context) |
2321 | intel_lr_context_unpin(req); |
2341 | intel_lr_context_unpin(ctx, req->ring); |
2322 | } |
- | |
2323 | 2342 | ||
2324 | i915_gem_context_unreference(ctx); |
2343 | i915_gem_context_unreference(ctx); |
2325 | } |
2344 | } |
2326 | 2345 | ||
2327 | kfree(req); |
2346 | kfree(req); |
2328 | } |
2347 | } |
- | 2348 | ||
2329 | 2349 | static inline int |
|
2330 | int i915_gem_request_alloc(struct intel_engine_cs *ring, |
2350 | __i915_gem_request_alloc(struct intel_engine_cs *ring, |
2331 | struct intel_context *ctx, |
2351 | struct intel_context *ctx, |
2332 | struct drm_i915_gem_request **req_out) |
2352 | struct drm_i915_gem_request **req_out) |
2333 | { |
2353 | { |
2334 | struct drm_i915_private *dev_priv = to_i915(ring->dev); |
2354 | struct drm_i915_private *dev_priv = to_i915(ring->dev); |
2335 | struct drm_i915_gem_request *req; |
2355 | struct drm_i915_gem_request *req; |
2336 | int ret; |
2356 | int ret; |
2337 | 2357 | ||
2338 | if (!req_out) |
2358 | if (!req_out) |
2339 | return -EINVAL; |
2359 | return -EINVAL; |
2340 | 2360 | ||
2341 | *req_out = NULL; |
2361 | *req_out = NULL; |
2342 | 2362 | ||
2343 | // req = kmem_cache_zalloc(dev_priv->requests, GFP_KERNEL); |
2363 | // req = kmem_cache_zalloc(dev_priv->requests, GFP_KERNEL); |
2344 | req = kzalloc(sizeof(*req),0); |
2364 | req = kzalloc(sizeof(*req),0); |
2345 | if (req == NULL) |
2365 | if (req == NULL) |
2346 | return -ENOMEM; |
2366 | return -ENOMEM; |
2347 | 2367 | ||
2348 | ret = i915_gem_get_seqno(ring->dev, &req->seqno); |
2368 | ret = i915_gem_get_seqno(ring->dev, &req->seqno); |
2349 | if (ret) |
2369 | if (ret) |
2350 | goto err; |
2370 | goto err; |
2351 | 2371 | ||
2352 | kref_init(&req->ref); |
2372 | kref_init(&req->ref); |
2353 | req->i915 = dev_priv; |
2373 | req->i915 = dev_priv; |
2354 | req->ring = ring; |
2374 | req->ring = ring; |
2355 | req->ctx = ctx; |
2375 | req->ctx = ctx; |
2356 | i915_gem_context_reference(req->ctx); |
2376 | i915_gem_context_reference(req->ctx); |
2357 | 2377 | ||
2358 | if (i915.enable_execlists) |
2378 | if (i915.enable_execlists) |
2359 | ret = intel_logical_ring_alloc_request_extras(req); |
2379 | ret = intel_logical_ring_alloc_request_extras(req); |
2360 | else |
2380 | else |
2361 | ret = intel_ring_alloc_request_extras(req); |
2381 | ret = intel_ring_alloc_request_extras(req); |
2362 | if (ret) { |
2382 | if (ret) { |
2363 | i915_gem_context_unreference(req->ctx); |
2383 | i915_gem_context_unreference(req->ctx); |
2364 | goto err; |
2384 | goto err; |
2365 | } |
2385 | } |
2366 | 2386 | ||
2367 | /* |
2387 | /* |
2368 | * Reserve space in the ring buffer for all the commands required to |
2388 | * Reserve space in the ring buffer for all the commands required to |
2369 | * eventually emit this request. This is to guarantee that the |
2389 | * eventually emit this request. This is to guarantee that the |
2370 | * i915_add_request() call can't fail. Note that the reserve may need |
2390 | * i915_add_request() call can't fail. Note that the reserve may need |
2371 | * to be redone if the request is not actually submitted straight |
2391 | * to be redone if the request is not actually submitted straight |
2372 | * away, e.g. because a GPU scheduler has deferred it. |
2392 | * away, e.g. because a GPU scheduler has deferred it. |
2373 | */ |
2393 | */ |
2374 | if (i915.enable_execlists) |
2394 | if (i915.enable_execlists) |
2375 | ret = intel_logical_ring_reserve_space(req); |
2395 | ret = intel_logical_ring_reserve_space(req); |
2376 | else |
2396 | else |
2377 | ret = intel_ring_reserve_space(req); |
2397 | ret = intel_ring_reserve_space(req); |
2378 | if (ret) { |
2398 | if (ret) { |
2379 | /* |
2399 | /* |
2380 | * At this point, the request is fully allocated even if not |
2400 | * At this point, the request is fully allocated even if not |
2381 | * fully prepared. Thus it can be cleaned up using the proper |
2401 | * fully prepared. Thus it can be cleaned up using the proper |
2382 | * free code. |
2402 | * free code. |
2383 | */ |
2403 | */ |
2384 | i915_gem_request_cancel(req); |
2404 | i915_gem_request_cancel(req); |
2385 | return ret; |
2405 | return ret; |
2386 | } |
2406 | } |
2387 | 2407 | ||
2388 | *req_out = req; |
2408 | *req_out = req; |
2389 | return 0; |
2409 | return 0; |
2390 | 2410 | ||
2391 | err: |
2411 | err: |
2392 | kfree(req); |
2412 | kfree(req); |
2393 | return ret; |
2413 | return ret; |
2394 | } |
2414 | } |
- | 2415 | ||
- | 2416 | /** |
|
- | 2417 | * i915_gem_request_alloc - allocate a request structure |
|
- | 2418 | * |
|
- | 2419 | * @engine: engine that we wish to issue the request on. |
|
- | 2420 | * @ctx: context that the request will be associated with. |
|
- | 2421 | * This can be NULL if the request is not directly related to |
|
- | 2422 | * any specific user context, in which case this function will |
|
- | 2423 | * choose an appropriate context to use. |
|
- | 2424 | * |
|
- | 2425 | * Returns a pointer to the allocated request if successful, |
|
- | 2426 | * or an error code if not. |
|
- | 2427 | */ |
|
- | 2428 | struct drm_i915_gem_request * |
|
- | 2429 | i915_gem_request_alloc(struct intel_engine_cs *engine, |
|
- | 2430 | struct intel_context *ctx) |
|
- | 2431 | { |
|
- | 2432 | struct drm_i915_gem_request *req; |
|
- | 2433 | int err; |
|
- | 2434 | ||
- | 2435 | if (ctx == NULL) |
|
- | 2436 | ctx = to_i915(engine->dev)->kernel_context; |
|
- | 2437 | err = __i915_gem_request_alloc(engine, ctx, &req); |
|
- | 2438 | return err ? ERR_PTR(err) : req; |
|
- | 2439 | } |
|
2395 | 2440 | ||
2396 | void i915_gem_request_cancel(struct drm_i915_gem_request *req) |
2441 | void i915_gem_request_cancel(struct drm_i915_gem_request *req) |
2397 | { |
2442 | { |
2398 | intel_ring_reserved_space_cancel(req->ringbuf); |
2443 | intel_ring_reserved_space_cancel(req->ringbuf); |
2399 | 2444 | ||
2400 | i915_gem_request_unreference(req); |
2445 | i915_gem_request_unreference(req); |
2401 | } |
2446 | } |
2402 | 2447 | ||
2403 | struct drm_i915_gem_request * |
2448 | struct drm_i915_gem_request * |
2404 | i915_gem_find_active_request(struct intel_engine_cs *ring) |
2449 | i915_gem_find_active_request(struct intel_engine_cs *ring) |
2405 | { |
2450 | { |
2406 | struct drm_i915_gem_request *request; |
2451 | struct drm_i915_gem_request *request; |
2407 | 2452 | ||
2408 | list_for_each_entry(request, &ring->request_list, list) { |
2453 | list_for_each_entry(request, &ring->request_list, list) { |
2409 | if (i915_gem_request_completed(request, false)) |
2454 | if (i915_gem_request_completed(request, false)) |
2410 | continue; |
2455 | continue; |
2411 | 2456 | ||
2412 | return request; |
2457 | return request; |
2413 | } |
2458 | } |
2414 | 2459 | ||
2415 | return NULL; |
2460 | return NULL; |
2416 | } |
2461 | } |
2417 | 2462 | ||
2418 | static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv, |
2463 | static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv, |
2419 | struct intel_engine_cs *ring) |
2464 | struct intel_engine_cs *ring) |
2420 | { |
2465 | { |
2421 | struct drm_i915_gem_request *request; |
2466 | struct drm_i915_gem_request *request; |
2422 | bool ring_hung; |
2467 | bool ring_hung; |
2423 | 2468 | ||
2424 | request = i915_gem_find_active_request(ring); |
2469 | request = i915_gem_find_active_request(ring); |
2425 | 2470 | ||
2426 | if (request == NULL) |
2471 | if (request == NULL) |
2427 | return; |
2472 | return; |
2428 | 2473 | ||
2429 | ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG; |
2474 | ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG; |
2430 | 2475 | ||
2431 | i915_set_reset_status(dev_priv, request->ctx, ring_hung); |
2476 | i915_set_reset_status(dev_priv, request->ctx, ring_hung); |
2432 | 2477 | ||
2433 | list_for_each_entry_continue(request, &ring->request_list, list) |
2478 | list_for_each_entry_continue(request, &ring->request_list, list) |
2434 | i915_set_reset_status(dev_priv, request->ctx, false); |
2479 | i915_set_reset_status(dev_priv, request->ctx, false); |
2435 | } |
2480 | } |
2436 | 2481 | ||
2437 | static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv, |
2482 | static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv, |
2438 | struct intel_engine_cs *ring) |
2483 | struct intel_engine_cs *ring) |
2439 | { |
2484 | { |
2440 | struct intel_ringbuffer *buffer; |
2485 | struct intel_ringbuffer *buffer; |
2441 | 2486 | ||
2442 | while (!list_empty(&ring->active_list)) { |
2487 | while (!list_empty(&ring->active_list)) { |
2443 | struct drm_i915_gem_object *obj; |
2488 | struct drm_i915_gem_object *obj; |
2444 | 2489 | ||
2445 | obj = list_first_entry(&ring->active_list, |
2490 | obj = list_first_entry(&ring->active_list, |
2446 | struct drm_i915_gem_object, |
2491 | struct drm_i915_gem_object, |
2447 | ring_list[ring->id]); |
2492 | ring_list[ring->id]); |
2448 | 2493 | ||
2449 | i915_gem_object_retire__read(obj, ring->id); |
2494 | i915_gem_object_retire__read(obj, ring->id); |
2450 | } |
2495 | } |
2451 | 2496 | ||
2452 | /* |
2497 | /* |
2453 | * Clear the execlists queue up before freeing the requests, as those |
2498 | * Clear the execlists queue up before freeing the requests, as those |
2454 | * are the ones that keep the context and ringbuffer backing objects |
2499 | * are the ones that keep the context and ringbuffer backing objects |
2455 | * pinned in place. |
2500 | * pinned in place. |
2456 | */ |
2501 | */ |
2457 | 2502 | ||
2458 | if (i915.enable_execlists) { |
2503 | if (i915.enable_execlists) { |
2459 | spin_lock_irq(&ring->execlist_lock); |
2504 | spin_lock_irq(&ring->execlist_lock); |
2460 | 2505 | ||
2461 | /* list_splice_tail_init checks for empty lists */ |
2506 | /* list_splice_tail_init checks for empty lists */ |
2462 | list_splice_tail_init(&ring->execlist_queue, |
2507 | list_splice_tail_init(&ring->execlist_queue, |
2463 | &ring->execlist_retired_req_list); |
2508 | &ring->execlist_retired_req_list); |
2464 | 2509 | ||
2465 | spin_unlock_irq(&ring->execlist_lock); |
2510 | spin_unlock_irq(&ring->execlist_lock); |
2466 | intel_execlists_retire_requests(ring); |
2511 | intel_execlists_retire_requests(ring); |
2467 | } |
2512 | } |
2468 | 2513 | ||
2469 | /* |
2514 | /* |
2470 | * We must free the requests after all the corresponding objects have |
2515 | * We must free the requests after all the corresponding objects have |
2471 | * been moved off active lists. Which is the same order as the normal |
2516 | * been moved off active lists. Which is the same order as the normal |
2472 | * retire_requests function does. This is important if object hold |
2517 | * retire_requests function does. This is important if object hold |
2473 | * implicit references on things like e.g. ppgtt address spaces through |
2518 | * implicit references on things like e.g. ppgtt address spaces through |
2474 | * the request. |
2519 | * the request. |
2475 | */ |
2520 | */ |
2476 | while (!list_empty(&ring->request_list)) { |
2521 | while (!list_empty(&ring->request_list)) { |
2477 | struct drm_i915_gem_request *request; |
2522 | struct drm_i915_gem_request *request; |
2478 | 2523 | ||
2479 | request = list_first_entry(&ring->request_list, |
2524 | request = list_first_entry(&ring->request_list, |
2480 | struct drm_i915_gem_request, |
2525 | struct drm_i915_gem_request, |
2481 | list); |
2526 | list); |
2482 | 2527 | ||
2483 | i915_gem_request_retire(request); |
2528 | i915_gem_request_retire(request); |
2484 | } |
2529 | } |
2485 | 2530 | ||
2486 | /* Having flushed all requests from all queues, we know that all |
2531 | /* Having flushed all requests from all queues, we know that all |
2487 | * ringbuffers must now be empty. However, since we do not reclaim |
2532 | * ringbuffers must now be empty. However, since we do not reclaim |
2488 | * all space when retiring the request (to prevent HEADs colliding |
2533 | * all space when retiring the request (to prevent HEADs colliding |
2489 | * with rapid ringbuffer wraparound) the amount of available space |
2534 | * with rapid ringbuffer wraparound) the amount of available space |
2490 | * upon reset is less than when we start. Do one more pass over |
2535 | * upon reset is less than when we start. Do one more pass over |
2491 | * all the ringbuffers to reset last_retired_head. |
2536 | * all the ringbuffers to reset last_retired_head. |
2492 | */ |
2537 | */ |
2493 | list_for_each_entry(buffer, &ring->buffers, link) { |
2538 | list_for_each_entry(buffer, &ring->buffers, link) { |
2494 | buffer->last_retired_head = buffer->tail; |
2539 | buffer->last_retired_head = buffer->tail; |
2495 | intel_ring_update_space(buffer); |
2540 | intel_ring_update_space(buffer); |
2496 | } |
2541 | } |
2497 | } |
2542 | } |
2498 | 2543 | ||
2499 | void i915_gem_reset(struct drm_device *dev) |
2544 | void i915_gem_reset(struct drm_device *dev) |
2500 | { |
2545 | { |
2501 | struct drm_i915_private *dev_priv = dev->dev_private; |
2546 | struct drm_i915_private *dev_priv = dev->dev_private; |
2502 | struct intel_engine_cs *ring; |
2547 | struct intel_engine_cs *ring; |
2503 | int i; |
2548 | int i; |
2504 | 2549 | ||
2505 | /* |
2550 | /* |
2506 | * Before we free the objects from the requests, we need to inspect |
2551 | * Before we free the objects from the requests, we need to inspect |
2507 | * them for finding the guilty party. As the requests only borrow |
2552 | * them for finding the guilty party. As the requests only borrow |
2508 | * their reference to the objects, the inspection must be done first. |
2553 | * their reference to the objects, the inspection must be done first. |
2509 | */ |
2554 | */ |
2510 | for_each_ring(ring, dev_priv, i) |
2555 | for_each_ring(ring, dev_priv, i) |
2511 | i915_gem_reset_ring_status(dev_priv, ring); |
2556 | i915_gem_reset_ring_status(dev_priv, ring); |
2512 | 2557 | ||
2513 | for_each_ring(ring, dev_priv, i) |
2558 | for_each_ring(ring, dev_priv, i) |
2514 | i915_gem_reset_ring_cleanup(dev_priv, ring); |
2559 | i915_gem_reset_ring_cleanup(dev_priv, ring); |
2515 | 2560 | ||
2516 | i915_gem_context_reset(dev); |
2561 | i915_gem_context_reset(dev); |
2517 | 2562 | ||
2518 | i915_gem_restore_fences(dev); |
2563 | i915_gem_restore_fences(dev); |
2519 | 2564 | ||
2520 | WARN_ON(i915_verify_lists(dev)); |
2565 | WARN_ON(i915_verify_lists(dev)); |
2521 | } |
2566 | } |
2522 | 2567 | ||
2523 | /** |
2568 | /** |
2524 | * This function clears the request list as sequence numbers are passed. |
2569 | * This function clears the request list as sequence numbers are passed. |
2525 | */ |
2570 | */ |
2526 | void |
2571 | void |
2527 | i915_gem_retire_requests_ring(struct intel_engine_cs *ring) |
2572 | i915_gem_retire_requests_ring(struct intel_engine_cs *ring) |
2528 | { |
2573 | { |
2529 | WARN_ON(i915_verify_lists(ring->dev)); |
2574 | WARN_ON(i915_verify_lists(ring->dev)); |
2530 | 2575 | ||
2531 | /* Retire requests first as we use it above for the early return. |
2576 | /* Retire requests first as we use it above for the early return. |
2532 | * If we retire requests last, we may use a later seqno and so clear |
2577 | * If we retire requests last, we may use a later seqno and so clear |
2533 | * the requests lists without clearing the active list, leading to |
2578 | * the requests lists without clearing the active list, leading to |
2534 | * confusion. |
2579 | * confusion. |
2535 | */ |
2580 | */ |
2536 | while (!list_empty(&ring->request_list)) { |
2581 | while (!list_empty(&ring->request_list)) { |
2537 | struct drm_i915_gem_request *request; |
2582 | struct drm_i915_gem_request *request; |
2538 | 2583 | ||
2539 | request = list_first_entry(&ring->request_list, |
2584 | request = list_first_entry(&ring->request_list, |
2540 | struct drm_i915_gem_request, |
2585 | struct drm_i915_gem_request, |
2541 | list); |
2586 | list); |
2542 | 2587 | ||
2543 | if (!i915_gem_request_completed(request, true)) |
2588 | if (!i915_gem_request_completed(request, true)) |
2544 | break; |
2589 | break; |
2545 | 2590 | ||
2546 | i915_gem_request_retire(request); |
2591 | i915_gem_request_retire(request); |
2547 | } |
2592 | } |
2548 | 2593 | ||
2549 | /* Move any buffers on the active list that are no longer referenced |
2594 | /* Move any buffers on the active list that are no longer referenced |
2550 | * by the ringbuffer to the flushing/inactive lists as appropriate, |
2595 | * by the ringbuffer to the flushing/inactive lists as appropriate, |
2551 | * before we free the context associated with the requests. |
2596 | * before we free the context associated with the requests. |
2552 | */ |
2597 | */ |
2553 | while (!list_empty(&ring->active_list)) { |
2598 | while (!list_empty(&ring->active_list)) { |
2554 | struct drm_i915_gem_object *obj; |
2599 | struct drm_i915_gem_object *obj; |
2555 | 2600 | ||
2556 | obj = list_first_entry(&ring->active_list, |
2601 | obj = list_first_entry(&ring->active_list, |
2557 | struct drm_i915_gem_object, |
2602 | struct drm_i915_gem_object, |
2558 | ring_list[ring->id]); |
2603 | ring_list[ring->id]); |
2559 | 2604 | ||
2560 | if (!list_empty(&obj->last_read_req[ring->id]->list)) |
2605 | if (!list_empty(&obj->last_read_req[ring->id]->list)) |
2561 | break; |
2606 | break; |
2562 | 2607 | ||
2563 | i915_gem_object_retire__read(obj, ring->id); |
2608 | i915_gem_object_retire__read(obj, ring->id); |
2564 | } |
2609 | } |
2565 | 2610 | ||
2566 | if (unlikely(ring->trace_irq_req && |
2611 | if (unlikely(ring->trace_irq_req && |
2567 | i915_gem_request_completed(ring->trace_irq_req, true))) { |
2612 | i915_gem_request_completed(ring->trace_irq_req, true))) { |
2568 | ring->irq_put(ring); |
2613 | ring->irq_put(ring); |
2569 | i915_gem_request_assign(&ring->trace_irq_req, NULL); |
2614 | i915_gem_request_assign(&ring->trace_irq_req, NULL); |
2570 | } |
2615 | } |
2571 | 2616 | ||
2572 | WARN_ON(i915_verify_lists(ring->dev)); |
2617 | WARN_ON(i915_verify_lists(ring->dev)); |
2573 | } |
2618 | } |
2574 | 2619 | ||
2575 | bool |
2620 | bool |
2576 | i915_gem_retire_requests(struct drm_device *dev) |
2621 | i915_gem_retire_requests(struct drm_device *dev) |
2577 | { |
2622 | { |
2578 | struct drm_i915_private *dev_priv = dev->dev_private; |
2623 | struct drm_i915_private *dev_priv = dev->dev_private; |
2579 | struct intel_engine_cs *ring; |
2624 | struct intel_engine_cs *ring; |
2580 | bool idle = true; |
2625 | bool idle = true; |
2581 | int i; |
2626 | int i; |
2582 | 2627 | ||
2583 | for_each_ring(ring, dev_priv, i) { |
2628 | for_each_ring(ring, dev_priv, i) { |
2584 | i915_gem_retire_requests_ring(ring); |
2629 | i915_gem_retire_requests_ring(ring); |
2585 | idle &= list_empty(&ring->request_list); |
2630 | idle &= list_empty(&ring->request_list); |
2586 | if (i915.enable_execlists) { |
2631 | if (i915.enable_execlists) { |
2587 | unsigned long flags; |
- | |
2588 | - | ||
2589 | spin_lock_irqsave(&ring->execlist_lock, flags); |
2632 | spin_lock_irq(&ring->execlist_lock); |
2590 | idle &= list_empty(&ring->execlist_queue); |
2633 | idle &= list_empty(&ring->execlist_queue); |
2591 | spin_unlock_irqrestore(&ring->execlist_lock, flags); |
2634 | spin_unlock_irq(&ring->execlist_lock); |
2592 | 2635 | ||
2593 | intel_execlists_retire_requests(ring); |
2636 | intel_execlists_retire_requests(ring); |
2594 | } |
2637 | } |
2595 | } |
2638 | } |
2596 | 2639 | ||
2597 | // if (idle) |
2640 | // if (idle) |
2598 | // mod_delayed_work(dev_priv->wq, |
2641 | // mod_delayed_work(dev_priv->wq, |
2599 | // &dev_priv->mm.idle_work, |
2642 | // &dev_priv->mm.idle_work, |
2600 | // msecs_to_jiffies(100)); |
2643 | // msecs_to_jiffies(100)); |
2601 | 2644 | ||
2602 | return idle; |
2645 | return idle; |
2603 | } |
2646 | } |
2604 | 2647 | ||
2605 | static void |
2648 | static void |
2606 | i915_gem_retire_work_handler(struct work_struct *work) |
2649 | i915_gem_retire_work_handler(struct work_struct *work) |
2607 | { |
2650 | { |
2608 | struct drm_i915_private *dev_priv = |
2651 | struct drm_i915_private *dev_priv = |
2609 | container_of(work, typeof(*dev_priv), mm.retire_work.work); |
2652 | container_of(work, typeof(*dev_priv), mm.retire_work.work); |
2610 | struct drm_device *dev = dev_priv->dev; |
2653 | struct drm_device *dev = dev_priv->dev; |
2611 | bool idle; |
2654 | bool idle; |
2612 | 2655 | ||
2613 | /* Come back later if the device is busy... */ |
2656 | /* Come back later if the device is busy... */ |
2614 | idle = false; |
2657 | idle = false; |
2615 | if (mutex_trylock(&dev->struct_mutex)) { |
2658 | if (mutex_trylock(&dev->struct_mutex)) { |
2616 | idle = i915_gem_retire_requests(dev); |
2659 | idle = i915_gem_retire_requests(dev); |
2617 | mutex_unlock(&dev->struct_mutex); |
2660 | mutex_unlock(&dev->struct_mutex); |
2618 | } |
2661 | } |
2619 | if (!idle) |
2662 | if (!idle) |
2620 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2663 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2621 | round_jiffies_up_relative(HZ)); |
2664 | round_jiffies_up_relative(HZ)); |
2622 | } |
2665 | } |
2623 | 2666 | ||
2624 | static void |
2667 | static void |
2625 | i915_gem_idle_work_handler(struct work_struct *work) |
2668 | i915_gem_idle_work_handler(struct work_struct *work) |
2626 | { |
2669 | { |
2627 | struct drm_i915_private *dev_priv = |
2670 | struct drm_i915_private *dev_priv = |
2628 | container_of(work, typeof(*dev_priv), mm.idle_work.work); |
2671 | container_of(work, typeof(*dev_priv), mm.idle_work.work); |
2629 | struct drm_device *dev = dev_priv->dev; |
2672 | struct drm_device *dev = dev_priv->dev; |
2630 | struct intel_engine_cs *ring; |
2673 | struct intel_engine_cs *ring; |
2631 | int i; |
2674 | int i; |
2632 | 2675 | ||
2633 | for_each_ring(ring, dev_priv, i) |
2676 | for_each_ring(ring, dev_priv, i) |
2634 | if (!list_empty(&ring->request_list)) |
2677 | if (!list_empty(&ring->request_list)) |
2635 | return; |
2678 | return; |
2636 | 2679 | ||
2637 | /* we probably should sync with hangcheck here, using cancel_work_sync. |
2680 | /* we probably should sync with hangcheck here, using cancel_work_sync. |
2638 | * Also locking seems to be fubar here, ring->request_list is protected |
2681 | * Also locking seems to be fubar here, ring->request_list is protected |
2639 | * by dev->struct_mutex. */ |
2682 | * by dev->struct_mutex. */ |
2640 | 2683 | ||
2641 | intel_mark_idle(dev); |
2684 | intel_mark_idle(dev); |
2642 | 2685 | ||
2643 | if (mutex_trylock(&dev->struct_mutex)) { |
2686 | if (mutex_trylock(&dev->struct_mutex)) { |
2644 | struct intel_engine_cs *ring; |
2687 | struct intel_engine_cs *ring; |
2645 | int i; |
2688 | int i; |
2646 | 2689 | ||
2647 | for_each_ring(ring, dev_priv, i) |
2690 | for_each_ring(ring, dev_priv, i) |
2648 | i915_gem_batch_pool_fini(&ring->batch_pool); |
2691 | i915_gem_batch_pool_fini(&ring->batch_pool); |
2649 | 2692 | ||
2650 | mutex_unlock(&dev->struct_mutex); |
2693 | mutex_unlock(&dev->struct_mutex); |
2651 | } |
2694 | } |
2652 | } |
2695 | } |
2653 | 2696 | ||
2654 | /** |
2697 | /** |
2655 | * Ensures that an object will eventually get non-busy by flushing any required |
2698 | * Ensures that an object will eventually get non-busy by flushing any required |
2656 | * write domains, emitting any outstanding lazy request and retiring and |
2699 | * write domains, emitting any outstanding lazy request and retiring and |
2657 | * completed requests. |
2700 | * completed requests. |
2658 | */ |
2701 | */ |
2659 | static int |
2702 | static int |
2660 | i915_gem_object_flush_active(struct drm_i915_gem_object *obj) |
2703 | i915_gem_object_flush_active(struct drm_i915_gem_object *obj) |
2661 | { |
2704 | { |
2662 | int i; |
2705 | int i; |
2663 | 2706 | ||
2664 | if (!obj->active) |
2707 | if (!obj->active) |
2665 | return 0; |
2708 | return 0; |
2666 | 2709 | ||
2667 | for (i = 0; i < I915_NUM_RINGS; i++) { |
2710 | for (i = 0; i < I915_NUM_RINGS; i++) { |
2668 | struct drm_i915_gem_request *req; |
2711 | struct drm_i915_gem_request *req; |
2669 | 2712 | ||
2670 | req = obj->last_read_req[i]; |
2713 | req = obj->last_read_req[i]; |
2671 | if (req == NULL) |
2714 | if (req == NULL) |
2672 | continue; |
2715 | continue; |
2673 | 2716 | ||
2674 | if (list_empty(&req->list)) |
2717 | if (list_empty(&req->list)) |
2675 | goto retire; |
2718 | goto retire; |
2676 | 2719 | ||
2677 | if (i915_gem_request_completed(req, true)) { |
2720 | if (i915_gem_request_completed(req, true)) { |
2678 | __i915_gem_request_retire__upto(req); |
2721 | __i915_gem_request_retire__upto(req); |
2679 | retire: |
2722 | retire: |
2680 | i915_gem_object_retire__read(obj, i); |
2723 | i915_gem_object_retire__read(obj, i); |
2681 | } |
2724 | } |
2682 | } |
2725 | } |
2683 | 2726 | ||
2684 | return 0; |
2727 | return 0; |
2685 | } |
2728 | } |
2686 | 2729 | ||
2687 | /** |
2730 | /** |
2688 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT |
2731 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT |
2689 | * @DRM_IOCTL_ARGS: standard ioctl arguments |
2732 | * @DRM_IOCTL_ARGS: standard ioctl arguments |
2690 | * |
2733 | * |
2691 | * Returns 0 if successful, else an error is returned with the remaining time in |
2734 | * Returns 0 if successful, else an error is returned with the remaining time in |
2692 | * the timeout parameter. |
2735 | * the timeout parameter. |
2693 | * -ETIME: object is still busy after timeout |
2736 | * -ETIME: object is still busy after timeout |
2694 | * -ERESTARTSYS: signal interrupted the wait |
2737 | * -ERESTARTSYS: signal interrupted the wait |
2695 | * -ENONENT: object doesn't exist |
2738 | * -ENONENT: object doesn't exist |
2696 | * Also possible, but rare: |
2739 | * Also possible, but rare: |
2697 | * -EAGAIN: GPU wedged |
2740 | * -EAGAIN: GPU wedged |
2698 | * -ENOMEM: damn |
2741 | * -ENOMEM: damn |
2699 | * -ENODEV: Internal IRQ fail |
2742 | * -ENODEV: Internal IRQ fail |
2700 | * -E?: The add request failed |
2743 | * -E?: The add request failed |
2701 | * |
2744 | * |
2702 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any |
2745 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any |
2703 | * non-zero timeout parameter the wait ioctl will wait for the given number of |
2746 | * non-zero timeout parameter the wait ioctl will wait for the given number of |
2704 | * nanoseconds on an object becoming unbusy. Since the wait itself does so |
2747 | * nanoseconds on an object becoming unbusy. Since the wait itself does so |
2705 | * without holding struct_mutex the object may become re-busied before this |
2748 | * without holding struct_mutex the object may become re-busied before this |
2706 | * function completes. A similar but shorter * race condition exists in the busy |
2749 | * function completes. A similar but shorter * race condition exists in the busy |
2707 | * ioctl |
2750 | * ioctl |
2708 | */ |
2751 | */ |
2709 | int |
2752 | int |
2710 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) |
2753 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) |
2711 | { |
2754 | { |
2712 | struct drm_i915_private *dev_priv = dev->dev_private; |
2755 | struct drm_i915_private *dev_priv = dev->dev_private; |
2713 | struct drm_i915_gem_wait *args = data; |
2756 | struct drm_i915_gem_wait *args = data; |
2714 | struct drm_i915_gem_object *obj; |
2757 | struct drm_i915_gem_object *obj; |
2715 | struct drm_i915_gem_request *req[I915_NUM_RINGS]; |
2758 | struct drm_i915_gem_request *req[I915_NUM_RINGS]; |
2716 | unsigned reset_counter; |
2759 | unsigned reset_counter; |
2717 | int i, n = 0; |
2760 | int i, n = 0; |
2718 | int ret; |
2761 | int ret; |
2719 | 2762 | ||
2720 | if (args->flags != 0) |
2763 | if (args->flags != 0) |
2721 | return -EINVAL; |
2764 | return -EINVAL; |
2722 | 2765 | ||
2723 | ret = i915_mutex_lock_interruptible(dev); |
2766 | ret = i915_mutex_lock_interruptible(dev); |
2724 | if (ret) |
2767 | if (ret) |
2725 | return ret; |
2768 | return ret; |
2726 | 2769 | ||
2727 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle)); |
2770 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle)); |
2728 | if (&obj->base == NULL) { |
2771 | if (&obj->base == NULL) { |
2729 | mutex_unlock(&dev->struct_mutex); |
2772 | mutex_unlock(&dev->struct_mutex); |
2730 | return -ENOENT; |
2773 | return -ENOENT; |
2731 | } |
2774 | } |
2732 | 2775 | ||
2733 | /* Need to make sure the object gets inactive eventually. */ |
2776 | /* Need to make sure the object gets inactive eventually. */ |
2734 | ret = i915_gem_object_flush_active(obj); |
2777 | ret = i915_gem_object_flush_active(obj); |
2735 | if (ret) |
2778 | if (ret) |
2736 | goto out; |
2779 | goto out; |
2737 | 2780 | ||
2738 | if (!obj->active) |
2781 | if (!obj->active) |
2739 | goto out; |
2782 | goto out; |
2740 | 2783 | ||
2741 | /* Do this after OLR check to make sure we make forward progress polling |
2784 | /* Do this after OLR check to make sure we make forward progress polling |
2742 | * on this IOCTL with a timeout == 0 (like busy ioctl) |
2785 | * on this IOCTL with a timeout == 0 (like busy ioctl) |
2743 | */ |
2786 | */ |
2744 | if (args->timeout_ns == 0) { |
2787 | if (args->timeout_ns == 0) { |
2745 | ret = -ETIME; |
2788 | ret = -ETIME; |
2746 | goto out; |
2789 | goto out; |
2747 | } |
2790 | } |
2748 | 2791 | ||
2749 | drm_gem_object_unreference(&obj->base); |
2792 | drm_gem_object_unreference(&obj->base); |
2750 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
2793 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
2751 | 2794 | ||
2752 | for (i = 0; i < I915_NUM_RINGS; i++) { |
2795 | for (i = 0; i < I915_NUM_RINGS; i++) { |
2753 | if (obj->last_read_req[i] == NULL) |
2796 | if (obj->last_read_req[i] == NULL) |
2754 | continue; |
2797 | continue; |
2755 | 2798 | ||
2756 | req[n++] = i915_gem_request_reference(obj->last_read_req[i]); |
2799 | req[n++] = i915_gem_request_reference(obj->last_read_req[i]); |
2757 | } |
2800 | } |
2758 | 2801 | ||
2759 | mutex_unlock(&dev->struct_mutex); |
2802 | mutex_unlock(&dev->struct_mutex); |
2760 | 2803 | ||
2761 | for (i = 0; i < n; i++) { |
2804 | for (i = 0; i < n; i++) { |
2762 | if (ret == 0) |
2805 | if (ret == 0) |
2763 | ret = __i915_wait_request(req[i], reset_counter, true, |
2806 | ret = __i915_wait_request(req[i], reset_counter, true, |
2764 | args->timeout_ns > 0 ? &args->timeout_ns : NULL, |
2807 | args->timeout_ns > 0 ? &args->timeout_ns : NULL, |
2765 | to_rps_client(file)); |
2808 | to_rps_client(file)); |
2766 | i915_gem_request_unreference__unlocked(req[i]); |
2809 | i915_gem_request_unreference__unlocked(req[i]); |
2767 | } |
2810 | } |
2768 | return ret; |
2811 | return ret; |
2769 | 2812 | ||
2770 | out: |
2813 | out: |
2771 | drm_gem_object_unreference(&obj->base); |
2814 | drm_gem_object_unreference(&obj->base); |
2772 | mutex_unlock(&dev->struct_mutex); |
2815 | mutex_unlock(&dev->struct_mutex); |
2773 | return ret; |
2816 | return ret; |
2774 | } |
2817 | } |
2775 | 2818 | ||
2776 | static int |
2819 | static int |
2777 | __i915_gem_object_sync(struct drm_i915_gem_object *obj, |
2820 | __i915_gem_object_sync(struct drm_i915_gem_object *obj, |
2778 | struct intel_engine_cs *to, |
2821 | struct intel_engine_cs *to, |
2779 | struct drm_i915_gem_request *from_req, |
2822 | struct drm_i915_gem_request *from_req, |
2780 | struct drm_i915_gem_request **to_req) |
2823 | struct drm_i915_gem_request **to_req) |
2781 | { |
2824 | { |
2782 | struct intel_engine_cs *from; |
2825 | struct intel_engine_cs *from; |
2783 | int ret; |
2826 | int ret; |
2784 | 2827 | ||
2785 | from = i915_gem_request_get_ring(from_req); |
2828 | from = i915_gem_request_get_ring(from_req); |
2786 | if (to == from) |
2829 | if (to == from) |
2787 | return 0; |
2830 | return 0; |
2788 | 2831 | ||
2789 | if (i915_gem_request_completed(from_req, true)) |
2832 | if (i915_gem_request_completed(from_req, true)) |
2790 | return 0; |
2833 | return 0; |
2791 | 2834 | ||
2792 | if (!i915_semaphore_is_enabled(obj->base.dev)) { |
2835 | if (!i915_semaphore_is_enabled(obj->base.dev)) { |
2793 | struct drm_i915_private *i915 = to_i915(obj->base.dev); |
2836 | struct drm_i915_private *i915 = to_i915(obj->base.dev); |
2794 | ret = __i915_wait_request(from_req, |
2837 | ret = __i915_wait_request(from_req, |
2795 | atomic_read(&i915->gpu_error.reset_counter), |
2838 | atomic_read(&i915->gpu_error.reset_counter), |
2796 | i915->mm.interruptible, |
2839 | i915->mm.interruptible, |
2797 | NULL, |
2840 | NULL, |
2798 | &i915->rps.semaphores); |
2841 | &i915->rps.semaphores); |
2799 | if (ret) |
2842 | if (ret) |
2800 | return ret; |
2843 | return ret; |
2801 | 2844 | ||
2802 | i915_gem_object_retire_request(obj, from_req); |
2845 | i915_gem_object_retire_request(obj, from_req); |
2803 | } else { |
2846 | } else { |
2804 | int idx = intel_ring_sync_index(from, to); |
2847 | int idx = intel_ring_sync_index(from, to); |
2805 | u32 seqno = i915_gem_request_get_seqno(from_req); |
2848 | u32 seqno = i915_gem_request_get_seqno(from_req); |
2806 | 2849 | ||
2807 | WARN_ON(!to_req); |
2850 | WARN_ON(!to_req); |
2808 | 2851 | ||
2809 | if (seqno <= from->semaphore.sync_seqno[idx]) |
2852 | if (seqno <= from->semaphore.sync_seqno[idx]) |
2810 | return 0; |
2853 | return 0; |
2811 | 2854 | ||
2812 | if (*to_req == NULL) { |
2855 | if (*to_req == NULL) { |
- | 2856 | struct drm_i915_gem_request *req; |
|
- | 2857 | ||
2813 | ret = i915_gem_request_alloc(to, to->default_context, to_req); |
2858 | req = i915_gem_request_alloc(to, NULL); |
2814 | if (ret) |
2859 | if (IS_ERR(req)) |
2815 | return ret; |
2860 | return PTR_ERR(req); |
- | 2861 | ||
- | 2862 | *to_req = req; |
|
2816 | } |
2863 | } |
2817 | 2864 | ||
2818 | trace_i915_gem_ring_sync_to(*to_req, from, from_req); |
2865 | trace_i915_gem_ring_sync_to(*to_req, from, from_req); |
2819 | ret = to->semaphore.sync_to(*to_req, from, seqno); |
2866 | ret = to->semaphore.sync_to(*to_req, from, seqno); |
2820 | if (ret) |
2867 | if (ret) |
2821 | return ret; |
2868 | return ret; |
2822 | 2869 | ||
2823 | /* We use last_read_req because sync_to() |
2870 | /* We use last_read_req because sync_to() |
2824 | * might have just caused seqno wrap under |
2871 | * might have just caused seqno wrap under |
2825 | * the radar. |
2872 | * the radar. |
2826 | */ |
2873 | */ |
2827 | from->semaphore.sync_seqno[idx] = |
2874 | from->semaphore.sync_seqno[idx] = |
2828 | i915_gem_request_get_seqno(obj->last_read_req[from->id]); |
2875 | i915_gem_request_get_seqno(obj->last_read_req[from->id]); |
2829 | } |
2876 | } |
2830 | 2877 | ||
2831 | return 0; |
2878 | return 0; |
2832 | } |
2879 | } |
2833 | 2880 | ||
2834 | /** |
2881 | /** |
2835 | * i915_gem_object_sync - sync an object to a ring. |
2882 | * i915_gem_object_sync - sync an object to a ring. |
2836 | * |
2883 | * |
2837 | * @obj: object which may be in use on another ring. |
2884 | * @obj: object which may be in use on another ring. |
2838 | * @to: ring we wish to use the object on. May be NULL. |
2885 | * @to: ring we wish to use the object on. May be NULL. |
2839 | * @to_req: request we wish to use the object for. See below. |
2886 | * @to_req: request we wish to use the object for. See below. |
2840 | * This will be allocated and returned if a request is |
2887 | * This will be allocated and returned if a request is |
2841 | * required but not passed in. |
2888 | * required but not passed in. |
2842 | * |
2889 | * |
2843 | * This code is meant to abstract object synchronization with the GPU. |
2890 | * This code is meant to abstract object synchronization with the GPU. |
2844 | * Calling with NULL implies synchronizing the object with the CPU |
2891 | * Calling with NULL implies synchronizing the object with the CPU |
2845 | * rather than a particular GPU ring. Conceptually we serialise writes |
2892 | * rather than a particular GPU ring. Conceptually we serialise writes |
2846 | * between engines inside the GPU. We only allow one engine to write |
2893 | * between engines inside the GPU. We only allow one engine to write |
2847 | * into a buffer at any time, but multiple readers. To ensure each has |
2894 | * into a buffer at any time, but multiple readers. To ensure each has |
2848 | * a coherent view of memory, we must: |
2895 | * a coherent view of memory, we must: |
2849 | * |
2896 | * |
2850 | * - If there is an outstanding write request to the object, the new |
2897 | * - If there is an outstanding write request to the object, the new |
2851 | * request must wait for it to complete (either CPU or in hw, requests |
2898 | * request must wait for it to complete (either CPU or in hw, requests |
2852 | * on the same ring will be naturally ordered). |
2899 | * on the same ring will be naturally ordered). |
2853 | * |
2900 | * |
2854 | * - If we are a write request (pending_write_domain is set), the new |
2901 | * - If we are a write request (pending_write_domain is set), the new |
2855 | * request must wait for outstanding read requests to complete. |
2902 | * request must wait for outstanding read requests to complete. |
2856 | * |
2903 | * |
2857 | * For CPU synchronisation (NULL to) no request is required. For syncing with |
2904 | * For CPU synchronisation (NULL to) no request is required. For syncing with |
2858 | * rings to_req must be non-NULL. However, a request does not have to be |
2905 | * rings to_req must be non-NULL. However, a request does not have to be |
2859 | * pre-allocated. If *to_req is NULL and sync commands will be emitted then a |
2906 | * pre-allocated. If *to_req is NULL and sync commands will be emitted then a |
2860 | * request will be allocated automatically and returned through *to_req. Note |
2907 | * request will be allocated automatically and returned through *to_req. Note |
2861 | * that it is not guaranteed that commands will be emitted (because the system |
2908 | * that it is not guaranteed that commands will be emitted (because the system |
2862 | * might already be idle). Hence there is no need to create a request that |
2909 | * might already be idle). Hence there is no need to create a request that |
2863 | * might never have any work submitted. Note further that if a request is |
2910 | * might never have any work submitted. Note further that if a request is |
2864 | * returned in *to_req, it is the responsibility of the caller to submit |
2911 | * returned in *to_req, it is the responsibility of the caller to submit |
2865 | * that request (after potentially adding more work to it). |
2912 | * that request (after potentially adding more work to it). |
2866 | * |
2913 | * |
2867 | * Returns 0 if successful, else propagates up the lower layer error. |
2914 | * Returns 0 if successful, else propagates up the lower layer error. |
2868 | */ |
2915 | */ |
2869 | int |
2916 | int |
2870 | i915_gem_object_sync(struct drm_i915_gem_object *obj, |
2917 | i915_gem_object_sync(struct drm_i915_gem_object *obj, |
2871 | struct intel_engine_cs *to, |
2918 | struct intel_engine_cs *to, |
2872 | struct drm_i915_gem_request **to_req) |
2919 | struct drm_i915_gem_request **to_req) |
2873 | { |
2920 | { |
2874 | const bool readonly = obj->base.pending_write_domain == 0; |
2921 | const bool readonly = obj->base.pending_write_domain == 0; |
2875 | struct drm_i915_gem_request *req[I915_NUM_RINGS]; |
2922 | struct drm_i915_gem_request *req[I915_NUM_RINGS]; |
2876 | int ret, i, n; |
2923 | int ret, i, n; |
2877 | 2924 | ||
2878 | if (!obj->active) |
2925 | if (!obj->active) |
2879 | return 0; |
2926 | return 0; |
2880 | 2927 | ||
2881 | if (to == NULL) |
2928 | if (to == NULL) |
2882 | return i915_gem_object_wait_rendering(obj, readonly); |
2929 | return i915_gem_object_wait_rendering(obj, readonly); |
2883 | 2930 | ||
2884 | n = 0; |
2931 | n = 0; |
2885 | if (readonly) { |
2932 | if (readonly) { |
2886 | if (obj->last_write_req) |
2933 | if (obj->last_write_req) |
2887 | req[n++] = obj->last_write_req; |
2934 | req[n++] = obj->last_write_req; |
2888 | } else { |
2935 | } else { |
2889 | for (i = 0; i < I915_NUM_RINGS; i++) |
2936 | for (i = 0; i < I915_NUM_RINGS; i++) |
2890 | if (obj->last_read_req[i]) |
2937 | if (obj->last_read_req[i]) |
2891 | req[n++] = obj->last_read_req[i]; |
2938 | req[n++] = obj->last_read_req[i]; |
2892 | } |
2939 | } |
2893 | for (i = 0; i < n; i++) { |
2940 | for (i = 0; i < n; i++) { |
2894 | ret = __i915_gem_object_sync(obj, to, req[i], to_req); |
2941 | ret = __i915_gem_object_sync(obj, to, req[i], to_req); |
2895 | if (ret) |
2942 | if (ret) |
2896 | return ret; |
2943 | return ret; |
2897 | } |
2944 | } |
2898 | 2945 | ||
2899 | return 0; |
2946 | return 0; |
2900 | } |
2947 | } |
2901 | 2948 | ||
2902 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) |
2949 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) |
2903 | { |
2950 | { |
2904 | u32 old_write_domain, old_read_domains; |
2951 | u32 old_write_domain, old_read_domains; |
2905 | 2952 | ||
2906 | /* Force a pagefault for domain tracking on next user access */ |
2953 | /* Force a pagefault for domain tracking on next user access */ |
2907 | i915_gem_release_mmap(obj); |
2954 | i915_gem_release_mmap(obj); |
2908 | 2955 | ||
2909 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
2956 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
2910 | return; |
2957 | return; |
2911 | 2958 | ||
2912 | /* Wait for any direct GTT access to complete */ |
2959 | /* Wait for any direct GTT access to complete */ |
2913 | mb(); |
2960 | mb(); |
2914 | 2961 | ||
2915 | old_read_domains = obj->base.read_domains; |
2962 | old_read_domains = obj->base.read_domains; |
2916 | old_write_domain = obj->base.write_domain; |
2963 | old_write_domain = obj->base.write_domain; |
2917 | 2964 | ||
2918 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; |
2965 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; |
2919 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; |
2966 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; |
2920 | 2967 | ||
2921 | trace_i915_gem_object_change_domain(obj, |
2968 | trace_i915_gem_object_change_domain(obj, |
2922 | old_read_domains, |
2969 | old_read_domains, |
2923 | old_write_domain); |
2970 | old_write_domain); |
2924 | } |
2971 | } |
2925 | 2972 | ||
2926 | static int __i915_vma_unbind(struct i915_vma *vma, bool wait) |
2973 | static int __i915_vma_unbind(struct i915_vma *vma, bool wait) |
2927 | { |
2974 | { |
2928 | struct drm_i915_gem_object *obj = vma->obj; |
2975 | struct drm_i915_gem_object *obj = vma->obj; |
2929 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
2976 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
2930 | int ret; |
2977 | int ret; |
2931 | 2978 | ||
2932 | if (list_empty(&vma->vma_link)) |
2979 | if (list_empty(&vma->obj_link)) |
2933 | return 0; |
2980 | return 0; |
2934 | 2981 | ||
2935 | if (!drm_mm_node_allocated(&vma->node)) { |
2982 | if (!drm_mm_node_allocated(&vma->node)) { |
2936 | i915_gem_vma_destroy(vma); |
2983 | i915_gem_vma_destroy(vma); |
2937 | return 0; |
2984 | return 0; |
2938 | } |
2985 | } |
2939 | 2986 | ||
2940 | if (vma->pin_count) |
2987 | if (vma->pin_count) |
2941 | return -EBUSY; |
2988 | return -EBUSY; |
2942 | 2989 | ||
2943 | BUG_ON(obj->pages == NULL); |
2990 | BUG_ON(obj->pages == NULL); |
2944 | 2991 | ||
2945 | if (wait) { |
2992 | if (wait) { |
2946 | ret = i915_gem_object_wait_rendering(obj, false); |
2993 | ret = i915_gem_object_wait_rendering(obj, false); |
2947 | if (ret) |
2994 | if (ret) |
2948 | return ret; |
2995 | return ret; |
2949 | } |
2996 | } |
2950 | - | ||
2951 | if (i915_is_ggtt(vma->vm) && |
2997 | |
2952 | vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) { |
2998 | if (vma->is_ggtt && vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) { |
2953 | i915_gem_object_finish_gtt(obj); |
2999 | i915_gem_object_finish_gtt(obj); |
2954 | 3000 | ||
2955 | /* release the fence reg _after_ flushing */ |
3001 | /* release the fence reg _after_ flushing */ |
2956 | ret = i915_gem_object_put_fence(obj); |
3002 | ret = i915_gem_object_put_fence(obj); |
2957 | if (ret) |
3003 | if (ret) |
2958 | return ret; |
3004 | return ret; |
2959 | } |
3005 | } |
2960 | 3006 | ||
2961 | trace_i915_vma_unbind(vma); |
3007 | trace_i915_vma_unbind(vma); |
2962 | 3008 | ||
2963 | vma->vm->unbind_vma(vma); |
3009 | vma->vm->unbind_vma(vma); |
2964 | vma->bound = 0; |
3010 | vma->bound = 0; |
2965 | 3011 | ||
2966 | list_del_init(&vma->mm_list); |
3012 | list_del_init(&vma->vm_link); |
2967 | if (i915_is_ggtt(vma->vm)) { |
3013 | if (vma->is_ggtt) { |
2968 | if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) { |
3014 | if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) { |
2969 | obj->map_and_fenceable = false; |
3015 | obj->map_and_fenceable = false; |
2970 | } else if (vma->ggtt_view.pages) { |
3016 | } else if (vma->ggtt_view.pages) { |
2971 | sg_free_table(vma->ggtt_view.pages); |
3017 | sg_free_table(vma->ggtt_view.pages); |
2972 | kfree(vma->ggtt_view.pages); |
3018 | kfree(vma->ggtt_view.pages); |
2973 | } |
3019 | } |
2974 | vma->ggtt_view.pages = NULL; |
3020 | vma->ggtt_view.pages = NULL; |
2975 | } |
3021 | } |
2976 | 3022 | ||
2977 | drm_mm_remove_node(&vma->node); |
3023 | drm_mm_remove_node(&vma->node); |
2978 | i915_gem_vma_destroy(vma); |
3024 | i915_gem_vma_destroy(vma); |
2979 | 3025 | ||
2980 | /* Since the unbound list is global, only move to that list if |
3026 | /* Since the unbound list is global, only move to that list if |
2981 | * no more VMAs exist. */ |
3027 | * no more VMAs exist. */ |
2982 | if (list_empty(&obj->vma_list)) |
3028 | if (list_empty(&obj->vma_list)) |
2983 | list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
3029 | list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
2984 | 3030 | ||
2985 | /* And finally now the object is completely decoupled from this vma, |
3031 | /* And finally now the object is completely decoupled from this vma, |
2986 | * we can drop its hold on the backing storage and allow it to be |
3032 | * we can drop its hold on the backing storage and allow it to be |
2987 | * reaped by the shrinker. |
3033 | * reaped by the shrinker. |
2988 | */ |
3034 | */ |
2989 | i915_gem_object_unpin_pages(obj); |
3035 | i915_gem_object_unpin_pages(obj); |
2990 | 3036 | ||
2991 | return 0; |
3037 | return 0; |
2992 | } |
3038 | } |
2993 | 3039 | ||
2994 | int i915_vma_unbind(struct i915_vma *vma) |
3040 | int i915_vma_unbind(struct i915_vma *vma) |
2995 | { |
3041 | { |
2996 | return __i915_vma_unbind(vma, true); |
3042 | return __i915_vma_unbind(vma, true); |
2997 | } |
3043 | } |
2998 | 3044 | ||
2999 | int __i915_vma_unbind_no_wait(struct i915_vma *vma) |
3045 | int __i915_vma_unbind_no_wait(struct i915_vma *vma) |
3000 | { |
3046 | { |
3001 | return __i915_vma_unbind(vma, false); |
3047 | return __i915_vma_unbind(vma, false); |
3002 | } |
3048 | } |
3003 | 3049 | ||
3004 | int i915_gpu_idle(struct drm_device *dev) |
3050 | int i915_gpu_idle(struct drm_device *dev) |
3005 | { |
3051 | { |
3006 | struct drm_i915_private *dev_priv = dev->dev_private; |
3052 | struct drm_i915_private *dev_priv = dev->dev_private; |
3007 | struct intel_engine_cs *ring; |
3053 | struct intel_engine_cs *ring; |
3008 | int ret, i; |
3054 | int ret, i; |
3009 | 3055 | ||
3010 | /* Flush everything onto the inactive list. */ |
3056 | /* Flush everything onto the inactive list. */ |
3011 | for_each_ring(ring, dev_priv, i) { |
3057 | for_each_ring(ring, dev_priv, i) { |
3012 | if (!i915.enable_execlists) { |
3058 | if (!i915.enable_execlists) { |
3013 | struct drm_i915_gem_request *req; |
3059 | struct drm_i915_gem_request *req; |
3014 | 3060 | ||
3015 | ret = i915_gem_request_alloc(ring, ring->default_context, &req); |
3061 | req = i915_gem_request_alloc(ring, NULL); |
3016 | if (ret) |
3062 | if (IS_ERR(req)) |
3017 | return ret; |
3063 | return PTR_ERR(req); |
3018 | 3064 | ||
3019 | ret = i915_switch_context(req); |
3065 | ret = i915_switch_context(req); |
3020 | if (ret) { |
3066 | if (ret) { |
3021 | i915_gem_request_cancel(req); |
3067 | i915_gem_request_cancel(req); |
3022 | return ret; |
3068 | return ret; |
3023 | } |
3069 | } |
3024 | 3070 | ||
3025 | i915_add_request_no_flush(req); |
3071 | i915_add_request_no_flush(req); |
3026 | } |
3072 | } |
3027 | 3073 | ||
3028 | ret = intel_ring_idle(ring); |
3074 | ret = intel_ring_idle(ring); |
3029 | if (ret) |
3075 | if (ret) |
3030 | return ret; |
3076 | return ret; |
3031 | } |
3077 | } |
3032 | 3078 | ||
3033 | WARN_ON(i915_verify_lists(dev)); |
3079 | WARN_ON(i915_verify_lists(dev)); |
3034 | return 0; |
3080 | return 0; |
3035 | } |
3081 | } |
3036 | 3082 | ||
3037 | static bool i915_gem_valid_gtt_space(struct i915_vma *vma, |
3083 | static bool i915_gem_valid_gtt_space(struct i915_vma *vma, |
3038 | unsigned long cache_level) |
3084 | unsigned long cache_level) |
3039 | { |
3085 | { |
3040 | struct drm_mm_node *gtt_space = &vma->node; |
3086 | struct drm_mm_node *gtt_space = &vma->node; |
3041 | struct drm_mm_node *other; |
3087 | struct drm_mm_node *other; |
3042 | 3088 | ||
3043 | /* |
3089 | /* |
3044 | * On some machines we have to be careful when putting differing types |
3090 | * On some machines we have to be careful when putting differing types |
3045 | * of snoopable memory together to avoid the prefetcher crossing memory |
3091 | * of snoopable memory together to avoid the prefetcher crossing memory |
3046 | * domains and dying. During vm initialisation, we decide whether or not |
3092 | * domains and dying. During vm initialisation, we decide whether or not |
3047 | * these constraints apply and set the drm_mm.color_adjust |
3093 | * these constraints apply and set the drm_mm.color_adjust |
3048 | * appropriately. |
3094 | * appropriately. |
3049 | */ |
3095 | */ |
3050 | if (vma->vm->mm.color_adjust == NULL) |
3096 | if (vma->vm->mm.color_adjust == NULL) |
3051 | return true; |
3097 | return true; |
3052 | 3098 | ||
3053 | if (!drm_mm_node_allocated(gtt_space)) |
3099 | if (!drm_mm_node_allocated(gtt_space)) |
3054 | return true; |
3100 | return true; |
3055 | 3101 | ||
3056 | if (list_empty(>t_space->node_list)) |
3102 | if (list_empty(>t_space->node_list)) |
3057 | return true; |
3103 | return true; |
3058 | 3104 | ||
3059 | other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list); |
3105 | other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list); |
3060 | if (other->allocated && !other->hole_follows && other->color != cache_level) |
3106 | if (other->allocated && !other->hole_follows && other->color != cache_level) |
3061 | return false; |
3107 | return false; |
3062 | 3108 | ||
3063 | other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list); |
3109 | other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list); |
3064 | if (other->allocated && !gtt_space->hole_follows && other->color != cache_level) |
3110 | if (other->allocated && !gtt_space->hole_follows && other->color != cache_level) |
3065 | return false; |
3111 | return false; |
3066 | 3112 | ||
3067 | return true; |
3113 | return true; |
3068 | } |
3114 | } |
3069 | 3115 | ||
3070 | /** |
3116 | /** |
3071 | * Finds free space in the GTT aperture and binds the object or a view of it |
3117 | * Finds free space in the GTT aperture and binds the object or a view of it |
3072 | * there. |
3118 | * there. |
3073 | */ |
3119 | */ |
3074 | static struct i915_vma * |
3120 | static struct i915_vma * |
3075 | i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, |
3121 | i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, |
3076 | struct i915_address_space *vm, |
3122 | struct i915_address_space *vm, |
3077 | const struct i915_ggtt_view *ggtt_view, |
3123 | const struct i915_ggtt_view *ggtt_view, |
3078 | unsigned alignment, |
3124 | unsigned alignment, |
3079 | uint64_t flags) |
3125 | uint64_t flags) |
3080 | { |
3126 | { |
3081 | struct drm_device *dev = obj->base.dev; |
3127 | struct drm_device *dev = obj->base.dev; |
3082 | struct drm_i915_private *dev_priv = dev->dev_private; |
3128 | struct drm_i915_private *dev_priv = dev->dev_private; |
3083 | u32 fence_alignment, unfenced_alignment; |
3129 | u32 fence_alignment, unfenced_alignment; |
3084 | u32 search_flag, alloc_flag; |
3130 | u32 search_flag, alloc_flag; |
3085 | u64 start, end; |
3131 | u64 start, end; |
3086 | u64 size, fence_size; |
3132 | u64 size, fence_size; |
3087 | struct i915_vma *vma; |
3133 | struct i915_vma *vma; |
3088 | int ret; |
3134 | int ret; |
3089 | 3135 | ||
3090 | if (i915_is_ggtt(vm)) { |
3136 | if (i915_is_ggtt(vm)) { |
3091 | u32 view_size; |
3137 | u32 view_size; |
3092 | 3138 | ||
3093 | if (WARN_ON(!ggtt_view)) |
3139 | if (WARN_ON(!ggtt_view)) |
3094 | return ERR_PTR(-EINVAL); |
3140 | return ERR_PTR(-EINVAL); |
3095 | 3141 | ||
3096 | view_size = i915_ggtt_view_size(obj, ggtt_view); |
3142 | view_size = i915_ggtt_view_size(obj, ggtt_view); |
3097 | 3143 | ||
3098 | fence_size = i915_gem_get_gtt_size(dev, |
3144 | fence_size = i915_gem_get_gtt_size(dev, |
3099 | view_size, |
3145 | view_size, |
3100 | obj->tiling_mode); |
3146 | obj->tiling_mode); |
3101 | fence_alignment = i915_gem_get_gtt_alignment(dev, |
3147 | fence_alignment = i915_gem_get_gtt_alignment(dev, |
3102 | view_size, |
3148 | view_size, |
3103 | obj->tiling_mode, |
3149 | obj->tiling_mode, |
3104 | true); |
3150 | true); |
3105 | unfenced_alignment = i915_gem_get_gtt_alignment(dev, |
3151 | unfenced_alignment = i915_gem_get_gtt_alignment(dev, |
3106 | view_size, |
3152 | view_size, |
3107 | obj->tiling_mode, |
3153 | obj->tiling_mode, |
3108 | false); |
3154 | false); |
3109 | size = flags & PIN_MAPPABLE ? fence_size : view_size; |
3155 | size = flags & PIN_MAPPABLE ? fence_size : view_size; |
3110 | } else { |
3156 | } else { |
3111 | fence_size = i915_gem_get_gtt_size(dev, |
3157 | fence_size = i915_gem_get_gtt_size(dev, |
3112 | obj->base.size, |
3158 | obj->base.size, |
3113 | obj->tiling_mode); |
3159 | obj->tiling_mode); |
3114 | fence_alignment = i915_gem_get_gtt_alignment(dev, |
3160 | fence_alignment = i915_gem_get_gtt_alignment(dev, |
3115 | obj->base.size, |
3161 | obj->base.size, |
3116 | obj->tiling_mode, |
3162 | obj->tiling_mode, |
3117 | true); |
3163 | true); |
3118 | unfenced_alignment = |
3164 | unfenced_alignment = |
3119 | i915_gem_get_gtt_alignment(dev, |
3165 | i915_gem_get_gtt_alignment(dev, |
3120 | obj->base.size, |
3166 | obj->base.size, |
3121 | obj->tiling_mode, |
3167 | obj->tiling_mode, |
3122 | false); |
3168 | false); |
3123 | size = flags & PIN_MAPPABLE ? fence_size : obj->base.size; |
3169 | size = flags & PIN_MAPPABLE ? fence_size : obj->base.size; |
3124 | } |
3170 | } |
3125 | 3171 | ||
3126 | start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0; |
3172 | start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0; |
3127 | end = vm->total; |
3173 | end = vm->total; |
3128 | if (flags & PIN_MAPPABLE) |
3174 | if (flags & PIN_MAPPABLE) |
3129 | end = min_t(u64, end, dev_priv->gtt.mappable_end); |
3175 | end = min_t(u64, end, dev_priv->gtt.mappable_end); |
3130 | if (flags & PIN_ZONE_4G) |
3176 | if (flags & PIN_ZONE_4G) |
3131 | end = min_t(u64, end, (1ULL << 32) - PAGE_SIZE); |
3177 | end = min_t(u64, end, (1ULL << 32) - PAGE_SIZE); |
3132 | 3178 | ||
3133 | if (alignment == 0) |
3179 | if (alignment == 0) |
3134 | alignment = flags & PIN_MAPPABLE ? fence_alignment : |
3180 | alignment = flags & PIN_MAPPABLE ? fence_alignment : |
3135 | unfenced_alignment; |
3181 | unfenced_alignment; |
3136 | if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) { |
3182 | if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) { |
3137 | DRM_DEBUG("Invalid object (view type=%u) alignment requested %u\n", |
3183 | DRM_DEBUG("Invalid object (view type=%u) alignment requested %u\n", |
3138 | ggtt_view ? ggtt_view->type : 0, |
3184 | ggtt_view ? ggtt_view->type : 0, |
3139 | alignment); |
3185 | alignment); |
3140 | return ERR_PTR(-EINVAL); |
3186 | return ERR_PTR(-EINVAL); |
3141 | } |
3187 | } |
3142 | 3188 | ||
3143 | /* If binding the object/GGTT view requires more space than the entire |
3189 | /* If binding the object/GGTT view requires more space than the entire |
3144 | * aperture has, reject it early before evicting everything in a vain |
3190 | * aperture has, reject it early before evicting everything in a vain |
3145 | * attempt to find space. |
3191 | * attempt to find space. |
3146 | */ |
3192 | */ |
3147 | if (size > end) { |
3193 | if (size > end) { |
3148 | DRM_DEBUG("Attempting to bind an object (view type=%u) larger than the aperture: size=%llu > %s aperture=%llu\n", |
3194 | DRM_DEBUG("Attempting to bind an object (view type=%u) larger than the aperture: size=%llu > %s aperture=%llu\n", |
3149 | ggtt_view ? ggtt_view->type : 0, |
3195 | ggtt_view ? ggtt_view->type : 0, |
3150 | size, |
3196 | size, |
3151 | flags & PIN_MAPPABLE ? "mappable" : "total", |
3197 | flags & PIN_MAPPABLE ? "mappable" : "total", |
3152 | end); |
3198 | end); |
3153 | return ERR_PTR(-E2BIG); |
3199 | return ERR_PTR(-E2BIG); |
3154 | } |
3200 | } |
3155 | 3201 | ||
3156 | ret = i915_gem_object_get_pages(obj); |
3202 | ret = i915_gem_object_get_pages(obj); |
3157 | if (ret) |
3203 | if (ret) |
3158 | return ERR_PTR(ret); |
3204 | return ERR_PTR(ret); |
3159 | 3205 | ||
3160 | i915_gem_object_pin_pages(obj); |
3206 | i915_gem_object_pin_pages(obj); |
3161 | 3207 | ||
3162 | vma = ggtt_view ? i915_gem_obj_lookup_or_create_ggtt_vma(obj, ggtt_view) : |
3208 | vma = ggtt_view ? i915_gem_obj_lookup_or_create_ggtt_vma(obj, ggtt_view) : |
3163 | i915_gem_obj_lookup_or_create_vma(obj, vm); |
3209 | i915_gem_obj_lookup_or_create_vma(obj, vm); |
3164 | 3210 | ||
3165 | if (IS_ERR(vma)) |
3211 | if (IS_ERR(vma)) |
3166 | goto err_unpin; |
3212 | goto err_unpin; |
3167 | 3213 | ||
3168 | if (flags & PIN_OFFSET_FIXED) { |
3214 | if (flags & PIN_OFFSET_FIXED) { |
3169 | uint64_t offset = flags & PIN_OFFSET_MASK; |
3215 | uint64_t offset = flags & PIN_OFFSET_MASK; |
3170 | 3216 | ||
3171 | if (offset & (alignment - 1) || offset + size > end) { |
3217 | if (offset & (alignment - 1) || offset + size > end) { |
3172 | ret = -EINVAL; |
3218 | ret = -EINVAL; |
3173 | goto err_free_vma; |
3219 | goto err_free_vma; |
3174 | } |
3220 | } |
3175 | vma->node.start = offset; |
3221 | vma->node.start = offset; |
3176 | vma->node.size = size; |
3222 | vma->node.size = size; |
3177 | vma->node.color = obj->cache_level; |
3223 | vma->node.color = obj->cache_level; |
3178 | ret = drm_mm_reserve_node(&vm->mm, &vma->node); |
3224 | ret = drm_mm_reserve_node(&vm->mm, &vma->node); |
3179 | if (ret) |
3225 | if (ret) |
3180 | goto err_free_vma; |
3226 | goto err_free_vma; |
3181 | } else { |
3227 | } else { |
3182 | if (flags & PIN_HIGH) { |
3228 | if (flags & PIN_HIGH) { |
3183 | search_flag = DRM_MM_SEARCH_BELOW; |
3229 | search_flag = DRM_MM_SEARCH_BELOW; |
3184 | alloc_flag = DRM_MM_CREATE_TOP; |
3230 | alloc_flag = DRM_MM_CREATE_TOP; |
3185 | } else { |
3231 | } else { |
3186 | search_flag = DRM_MM_SEARCH_DEFAULT; |
3232 | search_flag = DRM_MM_SEARCH_DEFAULT; |
3187 | alloc_flag = DRM_MM_CREATE_DEFAULT; |
3233 | alloc_flag = DRM_MM_CREATE_DEFAULT; |
3188 | } |
3234 | } |
3189 | 3235 | ||
3190 | search_free: |
3236 | search_free: |
3191 | ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node, |
3237 | ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node, |
3192 | size, alignment, |
3238 | size, alignment, |
3193 | obj->cache_level, |
3239 | obj->cache_level, |
3194 | start, end, |
3240 | start, end, |
3195 | search_flag, |
3241 | search_flag, |
3196 | alloc_flag); |
3242 | alloc_flag); |
3197 | if (ret) { |
3243 | if (ret) { |
3198 | 3244 | ||
3199 | goto err_free_vma; |
3245 | goto err_free_vma; |
3200 | } |
3246 | } |
3201 | } |
3247 | } |
3202 | if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) { |
3248 | if (WARN_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level))) { |
3203 | ret = -EINVAL; |
3249 | ret = -EINVAL; |
3204 | goto err_remove_node; |
3250 | goto err_remove_node; |
3205 | } |
3251 | } |
3206 | 3252 | ||
3207 | trace_i915_vma_bind(vma, flags); |
3253 | trace_i915_vma_bind(vma, flags); |
3208 | ret = i915_vma_bind(vma, obj->cache_level, flags); |
3254 | ret = i915_vma_bind(vma, obj->cache_level, flags); |
3209 | if (ret) |
3255 | if (ret) |
3210 | goto err_remove_node; |
3256 | goto err_remove_node; |
3211 | 3257 | ||
3212 | list_move_tail(&obj->global_list, &dev_priv->mm.bound_list); |
3258 | list_move_tail(&obj->global_list, &dev_priv->mm.bound_list); |
3213 | list_add_tail(&vma->mm_list, &vm->inactive_list); |
3259 | list_add_tail(&vma->vm_link, &vm->inactive_list); |
3214 | 3260 | ||
3215 | return vma; |
3261 | return vma; |
3216 | 3262 | ||
3217 | err_remove_node: |
3263 | err_remove_node: |
3218 | drm_mm_remove_node(&vma->node); |
3264 | drm_mm_remove_node(&vma->node); |
3219 | err_free_vma: |
3265 | err_free_vma: |
3220 | i915_gem_vma_destroy(vma); |
3266 | i915_gem_vma_destroy(vma); |
3221 | vma = ERR_PTR(ret); |
3267 | vma = ERR_PTR(ret); |
3222 | err_unpin: |
3268 | err_unpin: |
3223 | i915_gem_object_unpin_pages(obj); |
3269 | i915_gem_object_unpin_pages(obj); |
3224 | return vma; |
3270 | return vma; |
3225 | } |
3271 | } |
3226 | 3272 | ||
3227 | bool |
3273 | bool |
3228 | i915_gem_clflush_object(struct drm_i915_gem_object *obj, |
3274 | i915_gem_clflush_object(struct drm_i915_gem_object *obj, |
3229 | bool force) |
3275 | bool force) |
3230 | { |
3276 | { |
3231 | /* If we don't have a page list set up, then we're not pinned |
3277 | /* If we don't have a page list set up, then we're not pinned |
3232 | * to GPU, and we can ignore the cache flush because it'll happen |
3278 | * to GPU, and we can ignore the cache flush because it'll happen |
3233 | * again at bind time. |
3279 | * again at bind time. |
3234 | */ |
3280 | */ |
3235 | if (obj->pages == NULL) |
3281 | if (obj->pages == NULL) |
3236 | return false; |
3282 | return false; |
3237 | 3283 | ||
3238 | /* |
3284 | /* |
3239 | * Stolen memory is always coherent with the GPU as it is explicitly |
3285 | * Stolen memory is always coherent with the GPU as it is explicitly |
3240 | * marked as wc by the system, or the system is cache-coherent. |
3286 | * marked as wc by the system, or the system is cache-coherent. |
3241 | */ |
3287 | */ |
3242 | if (obj->stolen || obj->phys_handle) |
3288 | if (obj->stolen || obj->phys_handle) |
3243 | return false; |
3289 | return false; |
3244 | 3290 | ||
3245 | /* If the GPU is snooping the contents of the CPU cache, |
3291 | /* If the GPU is snooping the contents of the CPU cache, |
3246 | * we do not need to manually clear the CPU cache lines. However, |
3292 | * we do not need to manually clear the CPU cache lines. However, |
3247 | * the caches are only snooped when the render cache is |
3293 | * the caches are only snooped when the render cache is |
3248 | * flushed/invalidated. As we always have to emit invalidations |
3294 | * flushed/invalidated. As we always have to emit invalidations |
3249 | * and flushes when moving into and out of the RENDER domain, correct |
3295 | * and flushes when moving into and out of the RENDER domain, correct |
3250 | * snooping behaviour occurs naturally as the result of our domain |
3296 | * snooping behaviour occurs naturally as the result of our domain |
3251 | * tracking. |
3297 | * tracking. |
3252 | */ |
3298 | */ |
3253 | if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) { |
3299 | if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) { |
3254 | obj->cache_dirty = true; |
3300 | obj->cache_dirty = true; |
3255 | return false; |
3301 | return false; |
3256 | } |
3302 | } |
3257 | 3303 | ||
3258 | trace_i915_gem_object_clflush(obj); |
3304 | trace_i915_gem_object_clflush(obj); |
3259 | drm_clflush_sg(obj->pages); |
3305 | drm_clflush_sg(obj->pages); |
3260 | obj->cache_dirty = false; |
3306 | obj->cache_dirty = false; |
3261 | 3307 | ||
3262 | return true; |
3308 | return true; |
3263 | } |
3309 | } |
3264 | 3310 | ||
3265 | /** Flushes the GTT write domain for the object if it's dirty. */ |
3311 | /** Flushes the GTT write domain for the object if it's dirty. */ |
3266 | static void |
3312 | static void |
3267 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
3313 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
3268 | { |
3314 | { |
3269 | uint32_t old_write_domain; |
3315 | uint32_t old_write_domain; |
3270 | 3316 | ||
3271 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
3317 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
3272 | return; |
3318 | return; |
3273 | 3319 | ||
3274 | /* No actual flushing is required for the GTT write domain. Writes |
3320 | /* No actual flushing is required for the GTT write domain. Writes |
3275 | * to it immediately go to main memory as far as we know, so there's |
3321 | * to it immediately go to main memory as far as we know, so there's |
3276 | * no chipset flush. It also doesn't land in render cache. |
3322 | * no chipset flush. It also doesn't land in render cache. |
3277 | * |
3323 | * |
3278 | * However, we do have to enforce the order so that all writes through |
3324 | * However, we do have to enforce the order so that all writes through |
3279 | * the GTT land before any writes to the device, such as updates to |
3325 | * the GTT land before any writes to the device, such as updates to |
3280 | * the GATT itself. |
3326 | * the GATT itself. |
3281 | */ |
3327 | */ |
3282 | wmb(); |
3328 | wmb(); |
3283 | 3329 | ||
3284 | old_write_domain = obj->base.write_domain; |
3330 | old_write_domain = obj->base.write_domain; |
3285 | obj->base.write_domain = 0; |
3331 | obj->base.write_domain = 0; |
3286 | 3332 | ||
3287 | intel_fb_obj_flush(obj, false, ORIGIN_GTT); |
3333 | intel_fb_obj_flush(obj, false, ORIGIN_GTT); |
3288 | 3334 | ||
3289 | trace_i915_gem_object_change_domain(obj, |
3335 | trace_i915_gem_object_change_domain(obj, |
3290 | obj->base.read_domains, |
3336 | obj->base.read_domains, |
3291 | old_write_domain); |
3337 | old_write_domain); |
3292 | } |
3338 | } |
3293 | 3339 | ||
3294 | /** Flushes the CPU write domain for the object if it's dirty. */ |
3340 | /** Flushes the CPU write domain for the object if it's dirty. */ |
3295 | static void |
3341 | static void |
3296 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj) |
3342 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj) |
3297 | { |
3343 | { |
3298 | uint32_t old_write_domain; |
3344 | uint32_t old_write_domain; |
3299 | 3345 | ||
3300 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
3346 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
3301 | return; |
3347 | return; |
3302 | 3348 | ||
3303 | if (i915_gem_clflush_object(obj, obj->pin_display)) |
3349 | if (i915_gem_clflush_object(obj, obj->pin_display)) |
3304 | i915_gem_chipset_flush(obj->base.dev); |
3350 | i915_gem_chipset_flush(obj->base.dev); |
3305 | 3351 | ||
3306 | old_write_domain = obj->base.write_domain; |
3352 | old_write_domain = obj->base.write_domain; |
3307 | obj->base.write_domain = 0; |
3353 | obj->base.write_domain = 0; |
3308 | 3354 | ||
3309 | intel_fb_obj_flush(obj, false, ORIGIN_CPU); |
3355 | intel_fb_obj_flush(obj, false, ORIGIN_CPU); |
3310 | 3356 | ||
3311 | trace_i915_gem_object_change_domain(obj, |
3357 | trace_i915_gem_object_change_domain(obj, |
3312 | obj->base.read_domains, |
3358 | obj->base.read_domains, |
3313 | old_write_domain); |
3359 | old_write_domain); |
3314 | } |
3360 | } |
3315 | 3361 | ||
3316 | /** |
3362 | /** |
3317 | * Moves a single object to the GTT read, and possibly write domain. |
3363 | * Moves a single object to the GTT read, and possibly write domain. |
3318 | * |
3364 | * |
3319 | * This function returns when the move is complete, including waiting on |
3365 | * This function returns when the move is complete, including waiting on |
3320 | * flushes to occur. |
3366 | * flushes to occur. |
3321 | */ |
3367 | */ |
3322 | int |
3368 | int |
3323 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
3369 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
3324 | { |
3370 | { |
3325 | uint32_t old_write_domain, old_read_domains; |
3371 | uint32_t old_write_domain, old_read_domains; |
3326 | struct i915_vma *vma; |
3372 | struct i915_vma *vma; |
3327 | int ret; |
3373 | int ret; |
3328 | 3374 | ||
3329 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
3375 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
3330 | return 0; |
3376 | return 0; |
3331 | 3377 | ||
3332 | ret = i915_gem_object_wait_rendering(obj, !write); |
3378 | ret = i915_gem_object_wait_rendering(obj, !write); |
3333 | if (ret) |
3379 | if (ret) |
3334 | return ret; |
3380 | return ret; |
3335 | 3381 | ||
3336 | /* Flush and acquire obj->pages so that we are coherent through |
3382 | /* Flush and acquire obj->pages so that we are coherent through |
3337 | * direct access in memory with previous cached writes through |
3383 | * direct access in memory with previous cached writes through |
3338 | * shmemfs and that our cache domain tracking remains valid. |
3384 | * shmemfs and that our cache domain tracking remains valid. |
3339 | * For example, if the obj->filp was moved to swap without us |
3385 | * For example, if the obj->filp was moved to swap without us |
3340 | * being notified and releasing the pages, we would mistakenly |
3386 | * being notified and releasing the pages, we would mistakenly |
3341 | * continue to assume that the obj remained out of the CPU cached |
3387 | * continue to assume that the obj remained out of the CPU cached |
3342 | * domain. |
3388 | * domain. |
3343 | */ |
3389 | */ |
3344 | ret = i915_gem_object_get_pages(obj); |
3390 | ret = i915_gem_object_get_pages(obj); |
3345 | if (ret) |
3391 | if (ret) |
3346 | return ret; |
3392 | return ret; |
3347 | 3393 | ||
3348 | i915_gem_object_flush_cpu_write_domain(obj); |
3394 | i915_gem_object_flush_cpu_write_domain(obj); |
3349 | 3395 | ||
3350 | /* Serialise direct access to this object with the barriers for |
3396 | /* Serialise direct access to this object with the barriers for |
3351 | * coherent writes from the GPU, by effectively invalidating the |
3397 | * coherent writes from the GPU, by effectively invalidating the |
3352 | * GTT domain upon first access. |
3398 | * GTT domain upon first access. |
3353 | */ |
3399 | */ |
3354 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
3400 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
3355 | mb(); |
3401 | mb(); |
3356 | 3402 | ||
3357 | old_write_domain = obj->base.write_domain; |
3403 | old_write_domain = obj->base.write_domain; |
3358 | old_read_domains = obj->base.read_domains; |
3404 | old_read_domains = obj->base.read_domains; |
3359 | 3405 | ||
3360 | /* It should now be out of any other write domains, and we can update |
3406 | /* It should now be out of any other write domains, and we can update |
3361 | * the domain values for our changes. |
3407 | * the domain values for our changes. |
3362 | */ |
3408 | */ |
3363 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
3409 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
3364 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
3410 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
3365 | if (write) { |
3411 | if (write) { |
3366 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
3412 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
3367 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; |
3413 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; |
3368 | obj->dirty = 1; |
3414 | obj->dirty = 1; |
3369 | } |
3415 | } |
3370 | 3416 | ||
3371 | trace_i915_gem_object_change_domain(obj, |
3417 | trace_i915_gem_object_change_domain(obj, |
3372 | old_read_domains, |
3418 | old_read_domains, |
3373 | old_write_domain); |
3419 | old_write_domain); |
3374 | 3420 | ||
3375 | /* And bump the LRU for this access */ |
3421 | /* And bump the LRU for this access */ |
3376 | vma = i915_gem_obj_to_ggtt(obj); |
3422 | vma = i915_gem_obj_to_ggtt(obj); |
3377 | if (vma && drm_mm_node_allocated(&vma->node) && !obj->active) |
3423 | if (vma && drm_mm_node_allocated(&vma->node) && !obj->active) |
3378 | list_move_tail(&vma->mm_list, |
3424 | list_move_tail(&vma->vm_link, |
3379 | &to_i915(obj->base.dev)->gtt.base.inactive_list); |
3425 | &to_i915(obj->base.dev)->gtt.base.inactive_list); |
3380 | 3426 | ||
3381 | return 0; |
3427 | return 0; |
3382 | } |
3428 | } |
3383 | 3429 | ||
3384 | /** |
3430 | /** |
3385 | * Changes the cache-level of an object across all VMA. |
3431 | * Changes the cache-level of an object across all VMA. |
3386 | * |
3432 | * |
3387 | * After this function returns, the object will be in the new cache-level |
3433 | * After this function returns, the object will be in the new cache-level |
3388 | * across all GTT and the contents of the backing storage will be coherent, |
3434 | * across all GTT and the contents of the backing storage will be coherent, |
3389 | * with respect to the new cache-level. In order to keep the backing storage |
3435 | * with respect to the new cache-level. In order to keep the backing storage |
3390 | * coherent for all users, we only allow a single cache level to be set |
3436 | * coherent for all users, we only allow a single cache level to be set |
3391 | * globally on the object and prevent it from being changed whilst the |
3437 | * globally on the object and prevent it from being changed whilst the |
3392 | * hardware is reading from the object. That is if the object is currently |
3438 | * hardware is reading from the object. That is if the object is currently |
3393 | * on the scanout it will be set to uncached (or equivalent display |
3439 | * on the scanout it will be set to uncached (or equivalent display |
3394 | * cache coherency) and all non-MOCS GPU access will also be uncached so |
3440 | * cache coherency) and all non-MOCS GPU access will also be uncached so |
3395 | * that all direct access to the scanout remains coherent. |
3441 | * that all direct access to the scanout remains coherent. |
3396 | */ |
3442 | */ |
3397 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
3443 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
3398 | enum i915_cache_level cache_level) |
3444 | enum i915_cache_level cache_level) |
3399 | { |
3445 | { |
3400 | struct drm_device *dev = obj->base.dev; |
3446 | struct drm_device *dev = obj->base.dev; |
3401 | struct i915_vma *vma, *next; |
3447 | struct i915_vma *vma, *next; |
3402 | bool bound = false; |
3448 | bool bound = false; |
3403 | int ret = 0; |
3449 | int ret = 0; |
3404 | 3450 | ||
3405 | if (obj->cache_level == cache_level) |
3451 | if (obj->cache_level == cache_level) |
3406 | goto out; |
3452 | goto out; |
3407 | 3453 | ||
3408 | /* Inspect the list of currently bound VMA and unbind any that would |
3454 | /* Inspect the list of currently bound VMA and unbind any that would |
3409 | * be invalid given the new cache-level. This is principally to |
3455 | * be invalid given the new cache-level. This is principally to |
3410 | * catch the issue of the CS prefetch crossing page boundaries and |
3456 | * catch the issue of the CS prefetch crossing page boundaries and |
3411 | * reading an invalid PTE on older architectures. |
3457 | * reading an invalid PTE on older architectures. |
3412 | */ |
3458 | */ |
3413 | list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) { |
3459 | list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link) { |
3414 | if (!drm_mm_node_allocated(&vma->node)) |
3460 | if (!drm_mm_node_allocated(&vma->node)) |
3415 | continue; |
3461 | continue; |
3416 | 3462 | ||
3417 | if (vma->pin_count) { |
3463 | if (vma->pin_count) { |
3418 | DRM_DEBUG("can not change the cache level of pinned objects\n"); |
3464 | DRM_DEBUG("can not change the cache level of pinned objects\n"); |
3419 | return -EBUSY; |
3465 | return -EBUSY; |
3420 | } |
3466 | } |
3421 | 3467 | ||
3422 | if (!i915_gem_valid_gtt_space(vma, cache_level)) { |
3468 | if (!i915_gem_valid_gtt_space(vma, cache_level)) { |
3423 | ret = i915_vma_unbind(vma); |
3469 | ret = i915_vma_unbind(vma); |
3424 | if (ret) |
3470 | if (ret) |
3425 | return ret; |
3471 | return ret; |
3426 | } else |
3472 | } else |
3427 | bound = true; |
3473 | bound = true; |
3428 | } |
3474 | } |
3429 | 3475 | ||
3430 | /* We can reuse the existing drm_mm nodes but need to change the |
3476 | /* We can reuse the existing drm_mm nodes but need to change the |
3431 | * cache-level on the PTE. We could simply unbind them all and |
3477 | * cache-level on the PTE. We could simply unbind them all and |
3432 | * rebind with the correct cache-level on next use. However since |
3478 | * rebind with the correct cache-level on next use. However since |
3433 | * we already have a valid slot, dma mapping, pages etc, we may as |
3479 | * we already have a valid slot, dma mapping, pages etc, we may as |
3434 | * rewrite the PTE in the belief that doing so tramples upon less |
3480 | * rewrite the PTE in the belief that doing so tramples upon less |
3435 | * state and so involves less work. |
3481 | * state and so involves less work. |
3436 | */ |
3482 | */ |
3437 | if (bound) { |
3483 | if (bound) { |
3438 | /* Before we change the PTE, the GPU must not be accessing it. |
3484 | /* Before we change the PTE, the GPU must not be accessing it. |
3439 | * If we wait upon the object, we know that all the bound |
3485 | * If we wait upon the object, we know that all the bound |
3440 | * VMA are no longer active. |
3486 | * VMA are no longer active. |
3441 | */ |
3487 | */ |
3442 | ret = i915_gem_object_wait_rendering(obj, false); |
3488 | ret = i915_gem_object_wait_rendering(obj, false); |
3443 | if (ret) |
3489 | if (ret) |
3444 | return ret; |
3490 | return ret; |
3445 | 3491 | ||
3446 | if (!HAS_LLC(dev) && cache_level != I915_CACHE_NONE) { |
3492 | if (!HAS_LLC(dev) && cache_level != I915_CACHE_NONE) { |
3447 | /* Access to snoopable pages through the GTT is |
3493 | /* Access to snoopable pages through the GTT is |
3448 | * incoherent and on some machines causes a hard |
3494 | * incoherent and on some machines causes a hard |
3449 | * lockup. Relinquish the CPU mmaping to force |
3495 | * lockup. Relinquish the CPU mmaping to force |
3450 | * userspace to refault in the pages and we can |
3496 | * userspace to refault in the pages and we can |
3451 | * then double check if the GTT mapping is still |
3497 | * then double check if the GTT mapping is still |
3452 | * valid for that pointer access. |
3498 | * valid for that pointer access. |
3453 | */ |
3499 | */ |
3454 | i915_gem_release_mmap(obj); |
3500 | i915_gem_release_mmap(obj); |
3455 | 3501 | ||
3456 | /* As we no longer need a fence for GTT access, |
3502 | /* As we no longer need a fence for GTT access, |
3457 | * we can relinquish it now (and so prevent having |
3503 | * we can relinquish it now (and so prevent having |
3458 | * to steal a fence from someone else on the next |
3504 | * to steal a fence from someone else on the next |
3459 | * fence request). Note GPU activity would have |
3505 | * fence request). Note GPU activity would have |
3460 | * dropped the fence as all snoopable access is |
3506 | * dropped the fence as all snoopable access is |
3461 | * supposed to be linear. |
3507 | * supposed to be linear. |
3462 | */ |
3508 | */ |
3463 | ret = i915_gem_object_put_fence(obj); |
3509 | ret = i915_gem_object_put_fence(obj); |
3464 | if (ret) |
3510 | if (ret) |
3465 | return ret; |
3511 | return ret; |
3466 | } else { |
3512 | } else { |
3467 | /* We either have incoherent backing store and |
3513 | /* We either have incoherent backing store and |
3468 | * so no GTT access or the architecture is fully |
3514 | * so no GTT access or the architecture is fully |
3469 | * coherent. In such cases, existing GTT mmaps |
3515 | * coherent. In such cases, existing GTT mmaps |
3470 | * ignore the cache bit in the PTE and we can |
3516 | * ignore the cache bit in the PTE and we can |
3471 | * rewrite it without confusing the GPU or having |
3517 | * rewrite it without confusing the GPU or having |
3472 | * to force userspace to fault back in its mmaps. |
3518 | * to force userspace to fault back in its mmaps. |
3473 | */ |
3519 | */ |
3474 | } |
3520 | } |
3475 | 3521 | ||
3476 | list_for_each_entry(vma, &obj->vma_list, vma_link) { |
3522 | list_for_each_entry(vma, &obj->vma_list, obj_link) { |
3477 | if (!drm_mm_node_allocated(&vma->node)) |
3523 | if (!drm_mm_node_allocated(&vma->node)) |
3478 | continue; |
3524 | continue; |
3479 | 3525 | ||
3480 | ret = i915_vma_bind(vma, cache_level, PIN_UPDATE); |
3526 | ret = i915_vma_bind(vma, cache_level, PIN_UPDATE); |
3481 | if (ret) |
3527 | if (ret) |
3482 | return ret; |
3528 | return ret; |
3483 | } |
3529 | } |
3484 | } |
3530 | } |
3485 | 3531 | ||
3486 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
3532 | list_for_each_entry(vma, &obj->vma_list, obj_link) |
3487 | vma->node.color = cache_level; |
3533 | vma->node.color = cache_level; |
3488 | obj->cache_level = cache_level; |
3534 | obj->cache_level = cache_level; |
3489 | 3535 | ||
3490 | out: |
3536 | out: |
3491 | /* Flush the dirty CPU caches to the backing storage so that the |
3537 | /* Flush the dirty CPU caches to the backing storage so that the |
3492 | * object is now coherent at its new cache level (with respect |
3538 | * object is now coherent at its new cache level (with respect |
3493 | * to the access domain). |
3539 | * to the access domain). |
3494 | */ |
3540 | */ |
3495 | if (obj->cache_dirty && |
3541 | if (obj->cache_dirty && |
3496 | obj->base.write_domain != I915_GEM_DOMAIN_CPU && |
3542 | obj->base.write_domain != I915_GEM_DOMAIN_CPU && |
3497 | cpu_write_needs_clflush(obj)) { |
3543 | cpu_write_needs_clflush(obj)) { |
3498 | if (i915_gem_clflush_object(obj, true)) |
3544 | if (i915_gem_clflush_object(obj, true)) |
3499 | i915_gem_chipset_flush(obj->base.dev); |
3545 | i915_gem_chipset_flush(obj->base.dev); |
3500 | } |
3546 | } |
3501 | 3547 | ||
3502 | return 0; |
3548 | return 0; |
3503 | } |
3549 | } |
3504 | 3550 | ||
3505 | int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, |
3551 | int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, |
3506 | struct drm_file *file) |
3552 | struct drm_file *file) |
3507 | { |
3553 | { |
3508 | struct drm_i915_gem_caching *args = data; |
3554 | struct drm_i915_gem_caching *args = data; |
3509 | struct drm_i915_gem_object *obj; |
3555 | struct drm_i915_gem_object *obj; |
3510 | 3556 | ||
3511 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3557 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3512 | if (&obj->base == NULL) |
3558 | if (&obj->base == NULL) |
3513 | return -ENOENT; |
3559 | return -ENOENT; |
3514 | 3560 | ||
3515 | switch (obj->cache_level) { |
3561 | switch (obj->cache_level) { |
3516 | case I915_CACHE_LLC: |
3562 | case I915_CACHE_LLC: |
3517 | case I915_CACHE_L3_LLC: |
3563 | case I915_CACHE_L3_LLC: |
3518 | args->caching = I915_CACHING_CACHED; |
3564 | args->caching = I915_CACHING_CACHED; |
3519 | break; |
3565 | break; |
3520 | 3566 | ||
3521 | case I915_CACHE_WT: |
3567 | case I915_CACHE_WT: |
3522 | args->caching = I915_CACHING_DISPLAY; |
3568 | args->caching = I915_CACHING_DISPLAY; |
3523 | break; |
3569 | break; |
3524 | 3570 | ||
3525 | default: |
3571 | default: |
3526 | args->caching = I915_CACHING_NONE; |
3572 | args->caching = I915_CACHING_NONE; |
3527 | break; |
3573 | break; |
3528 | } |
3574 | } |
3529 | 3575 | ||
3530 | drm_gem_object_unreference_unlocked(&obj->base); |
3576 | drm_gem_object_unreference_unlocked(&obj->base); |
3531 | return 0; |
3577 | return 0; |
3532 | } |
3578 | } |
3533 | 3579 | ||
3534 | int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data, |
3580 | int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data, |
3535 | struct drm_file *file) |
3581 | struct drm_file *file) |
3536 | { |
3582 | { |
3537 | struct drm_i915_private *dev_priv = dev->dev_private; |
3583 | struct drm_i915_private *dev_priv = dev->dev_private; |
3538 | struct drm_i915_gem_caching *args = data; |
3584 | struct drm_i915_gem_caching *args = data; |
3539 | struct drm_i915_gem_object *obj; |
3585 | struct drm_i915_gem_object *obj; |
3540 | enum i915_cache_level level; |
3586 | enum i915_cache_level level; |
3541 | int ret; |
3587 | int ret; |
3542 | 3588 | ||
3543 | switch (args->caching) { |
3589 | switch (args->caching) { |
3544 | case I915_CACHING_NONE: |
3590 | case I915_CACHING_NONE: |
3545 | level = I915_CACHE_NONE; |
3591 | level = I915_CACHE_NONE; |
3546 | break; |
3592 | break; |
3547 | case I915_CACHING_CACHED: |
3593 | case I915_CACHING_CACHED: |
3548 | /* |
3594 | /* |
3549 | * Due to a HW issue on BXT A stepping, GPU stores via a |
3595 | * Due to a HW issue on BXT A stepping, GPU stores via a |
3550 | * snooped mapping may leave stale data in a corresponding CPU |
3596 | * snooped mapping may leave stale data in a corresponding CPU |
3551 | * cacheline, whereas normally such cachelines would get |
3597 | * cacheline, whereas normally such cachelines would get |
3552 | * invalidated. |
3598 | * invalidated. |
3553 | */ |
3599 | */ |
3554 | if (IS_BXT_REVID(dev, 0, BXT_REVID_A1)) |
3600 | if (IS_BXT_REVID(dev, 0, BXT_REVID_A1)) |
3555 | return -ENODEV; |
3601 | return -ENODEV; |
3556 | 3602 | ||
3557 | level = I915_CACHE_LLC; |
3603 | level = I915_CACHE_LLC; |
3558 | break; |
3604 | break; |
3559 | case I915_CACHING_DISPLAY: |
3605 | case I915_CACHING_DISPLAY: |
3560 | level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE; |
3606 | level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE; |
3561 | break; |
3607 | break; |
3562 | default: |
3608 | default: |
3563 | return -EINVAL; |
3609 | return -EINVAL; |
3564 | } |
3610 | } |
3565 | 3611 | ||
3566 | intel_runtime_pm_get(dev_priv); |
3612 | intel_runtime_pm_get(dev_priv); |
3567 | 3613 | ||
3568 | ret = i915_mutex_lock_interruptible(dev); |
3614 | ret = i915_mutex_lock_interruptible(dev); |
3569 | if (ret) |
3615 | if (ret) |
3570 | goto rpm_put; |
3616 | goto rpm_put; |
3571 | 3617 | ||
3572 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3618 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3573 | if (&obj->base == NULL) { |
3619 | if (&obj->base == NULL) { |
3574 | ret = -ENOENT; |
3620 | ret = -ENOENT; |
3575 | goto unlock; |
3621 | goto unlock; |
3576 | } |
3622 | } |
3577 | 3623 | ||
3578 | ret = i915_gem_object_set_cache_level(obj, level); |
3624 | ret = i915_gem_object_set_cache_level(obj, level); |
3579 | 3625 | ||
3580 | drm_gem_object_unreference(&obj->base); |
3626 | drm_gem_object_unreference(&obj->base); |
3581 | unlock: |
3627 | unlock: |
3582 | mutex_unlock(&dev->struct_mutex); |
3628 | mutex_unlock(&dev->struct_mutex); |
3583 | rpm_put: |
3629 | rpm_put: |
3584 | intel_runtime_pm_put(dev_priv); |
3630 | intel_runtime_pm_put(dev_priv); |
3585 | 3631 | ||
3586 | return ret; |
3632 | return ret; |
3587 | } |
3633 | } |
3588 | 3634 | ||
3589 | /* |
3635 | /* |
3590 | * Prepare buffer for display plane (scanout, cursors, etc). |
3636 | * Prepare buffer for display plane (scanout, cursors, etc). |
3591 | * Can be called from an uninterruptible phase (modesetting) and allows |
3637 | * Can be called from an uninterruptible phase (modesetting) and allows |
3592 | * any flushes to be pipelined (for pageflips). |
3638 | * any flushes to be pipelined (for pageflips). |
3593 | */ |
3639 | */ |
3594 | int |
3640 | int |
3595 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
3641 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
3596 | u32 alignment, |
3642 | u32 alignment, |
3597 | const struct i915_ggtt_view *view) |
3643 | const struct i915_ggtt_view *view) |
3598 | { |
3644 | { |
3599 | u32 old_read_domains, old_write_domain; |
3645 | u32 old_read_domains, old_write_domain; |
3600 | int ret; |
3646 | int ret; |
3601 | 3647 | ||
3602 | /* Mark the pin_display early so that we account for the |
3648 | /* Mark the pin_display early so that we account for the |
3603 | * display coherency whilst setting up the cache domains. |
3649 | * display coherency whilst setting up the cache domains. |
3604 | */ |
3650 | */ |
3605 | obj->pin_display++; |
3651 | obj->pin_display++; |
3606 | 3652 | ||
3607 | /* The display engine is not coherent with the LLC cache on gen6. As |
3653 | /* The display engine is not coherent with the LLC cache on gen6. As |
3608 | * a result, we make sure that the pinning that is about to occur is |
3654 | * a result, we make sure that the pinning that is about to occur is |
3609 | * done with uncached PTEs. This is lowest common denominator for all |
3655 | * done with uncached PTEs. This is lowest common denominator for all |
3610 | * chipsets. |
3656 | * chipsets. |
3611 | * |
3657 | * |
3612 | * However for gen6+, we could do better by using the GFDT bit instead |
3658 | * However for gen6+, we could do better by using the GFDT bit instead |
3613 | * of uncaching, which would allow us to flush all the LLC-cached data |
3659 | * of uncaching, which would allow us to flush all the LLC-cached data |
3614 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. |
3660 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. |
3615 | */ |
3661 | */ |
3616 | ret = i915_gem_object_set_cache_level(obj, |
3662 | ret = i915_gem_object_set_cache_level(obj, |
3617 | HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE); |
3663 | HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE); |
3618 | if (ret) |
3664 | if (ret) |
3619 | goto err_unpin_display; |
3665 | goto err_unpin_display; |
3620 | 3666 | ||
3621 | /* As the user may map the buffer once pinned in the display plane |
3667 | /* As the user may map the buffer once pinned in the display plane |
3622 | * (e.g. libkms for the bootup splash), we have to ensure that we |
3668 | * (e.g. libkms for the bootup splash), we have to ensure that we |
3623 | * always use map_and_fenceable for all scanout buffers. |
3669 | * always use map_and_fenceable for all scanout buffers. |
3624 | */ |
3670 | */ |
3625 | ret = i915_gem_object_ggtt_pin(obj, view, alignment, |
3671 | ret = i915_gem_object_ggtt_pin(obj, view, alignment, |
3626 | view->type == I915_GGTT_VIEW_NORMAL ? |
3672 | view->type == I915_GGTT_VIEW_NORMAL ? |
3627 | PIN_MAPPABLE : 0); |
3673 | PIN_MAPPABLE : 0); |
3628 | if (ret) |
3674 | if (ret) |
3629 | goto err_unpin_display; |
3675 | goto err_unpin_display; |
3630 | 3676 | ||
3631 | i915_gem_object_flush_cpu_write_domain(obj); |
3677 | i915_gem_object_flush_cpu_write_domain(obj); |
3632 | 3678 | ||
3633 | old_write_domain = obj->base.write_domain; |
3679 | old_write_domain = obj->base.write_domain; |
3634 | old_read_domains = obj->base.read_domains; |
3680 | old_read_domains = obj->base.read_domains; |
3635 | 3681 | ||
3636 | /* It should now be out of any other write domains, and we can update |
3682 | /* It should now be out of any other write domains, and we can update |
3637 | * the domain values for our changes. |
3683 | * the domain values for our changes. |
3638 | */ |
3684 | */ |
3639 | obj->base.write_domain = 0; |
3685 | obj->base.write_domain = 0; |
3640 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
3686 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
3641 | 3687 | ||
3642 | trace_i915_gem_object_change_domain(obj, |
3688 | trace_i915_gem_object_change_domain(obj, |
3643 | old_read_domains, |
3689 | old_read_domains, |
3644 | old_write_domain); |
3690 | old_write_domain); |
3645 | 3691 | ||
3646 | return 0; |
3692 | return 0; |
3647 | 3693 | ||
3648 | err_unpin_display: |
3694 | err_unpin_display: |
3649 | obj->pin_display--; |
3695 | obj->pin_display--; |
3650 | return ret; |
3696 | return ret; |
3651 | } |
3697 | } |
3652 | 3698 | ||
3653 | void |
3699 | void |
3654 | i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj, |
3700 | i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj, |
3655 | const struct i915_ggtt_view *view) |
3701 | const struct i915_ggtt_view *view) |
3656 | { |
3702 | { |
3657 | if (WARN_ON(obj->pin_display == 0)) |
3703 | if (WARN_ON(obj->pin_display == 0)) |
3658 | return; |
3704 | return; |
3659 | 3705 | ||
3660 | i915_gem_object_ggtt_unpin_view(obj, view); |
3706 | i915_gem_object_ggtt_unpin_view(obj, view); |
3661 | 3707 | ||
3662 | obj->pin_display--; |
3708 | obj->pin_display--; |
3663 | } |
3709 | } |
3664 | 3710 | ||
3665 | /** |
3711 | /** |
3666 | * Moves a single object to the CPU read, and possibly write domain. |
3712 | * Moves a single object to the CPU read, and possibly write domain. |
3667 | * |
3713 | * |
3668 | * This function returns when the move is complete, including waiting on |
3714 | * This function returns when the move is complete, including waiting on |
3669 | * flushes to occur. |
3715 | * flushes to occur. |
3670 | */ |
3716 | */ |
3671 | int |
3717 | int |
3672 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
3718 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
3673 | { |
3719 | { |
3674 | uint32_t old_write_domain, old_read_domains; |
3720 | uint32_t old_write_domain, old_read_domains; |
3675 | int ret; |
3721 | int ret; |
3676 | 3722 | ||
3677 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3723 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3678 | return 0; |
3724 | return 0; |
3679 | 3725 | ||
3680 | ret = i915_gem_object_wait_rendering(obj, !write); |
3726 | ret = i915_gem_object_wait_rendering(obj, !write); |
3681 | if (ret) |
3727 | if (ret) |
3682 | return ret; |
3728 | return ret; |
3683 | 3729 | ||
3684 | i915_gem_object_flush_gtt_write_domain(obj); |
3730 | i915_gem_object_flush_gtt_write_domain(obj); |
3685 | 3731 | ||
3686 | old_write_domain = obj->base.write_domain; |
3732 | old_write_domain = obj->base.write_domain; |
3687 | old_read_domains = obj->base.read_domains; |
3733 | old_read_domains = obj->base.read_domains; |
3688 | 3734 | ||
3689 | /* Flush the CPU cache if it's still invalid. */ |
3735 | /* Flush the CPU cache if it's still invalid. */ |
3690 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
3736 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
3691 | i915_gem_clflush_object(obj, false); |
3737 | i915_gem_clflush_object(obj, false); |
3692 | 3738 | ||
3693 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
3739 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
3694 | } |
3740 | } |
3695 | 3741 | ||
3696 | /* It should now be out of any other write domains, and we can update |
3742 | /* It should now be out of any other write domains, and we can update |
3697 | * the domain values for our changes. |
3743 | * the domain values for our changes. |
3698 | */ |
3744 | */ |
3699 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
3745 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
3700 | 3746 | ||
3701 | /* If we're writing through the CPU, then the GPU read domains will |
3747 | /* If we're writing through the CPU, then the GPU read domains will |
3702 | * need to be invalidated at next use. |
3748 | * need to be invalidated at next use. |
3703 | */ |
3749 | */ |
3704 | if (write) { |
3750 | if (write) { |
3705 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3751 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3706 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
3752 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
3707 | } |
3753 | } |
3708 | 3754 | ||
3709 | trace_i915_gem_object_change_domain(obj, |
3755 | trace_i915_gem_object_change_domain(obj, |
3710 | old_read_domains, |
3756 | old_read_domains, |
3711 | old_write_domain); |
3757 | old_write_domain); |
3712 | 3758 | ||
3713 | return 0; |
3759 | return 0; |
3714 | } |
3760 | } |
3715 | 3761 | ||
3716 | /* Throttle our rendering by waiting until the ring has completed our requests |
3762 | /* Throttle our rendering by waiting until the ring has completed our requests |
3717 | * emitted over 20 msec ago. |
3763 | * emitted over 20 msec ago. |
3718 | * |
3764 | * |
3719 | * Note that if we were to use the current jiffies each time around the loop, |
3765 | * Note that if we were to use the current jiffies each time around the loop, |
3720 | * we wouldn't escape the function with any frames outstanding if the time to |
3766 | * we wouldn't escape the function with any frames outstanding if the time to |
3721 | * render a frame was over 20ms. |
3767 | * render a frame was over 20ms. |
3722 | * |
3768 | * |
3723 | * This should get us reasonable parallelism between CPU and GPU but also |
3769 | * This should get us reasonable parallelism between CPU and GPU but also |
3724 | * relatively low latency when blocking on a particular request to finish. |
3770 | * relatively low latency when blocking on a particular request to finish. |
3725 | */ |
3771 | */ |
3726 | static int |
3772 | static int |
3727 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
3773 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
3728 | { |
3774 | { |
3729 | struct drm_i915_private *dev_priv = dev->dev_private; |
3775 | struct drm_i915_private *dev_priv = dev->dev_private; |
3730 | struct drm_i915_file_private *file_priv = file->driver_priv; |
3776 | struct drm_i915_file_private *file_priv = file->driver_priv; |
3731 | unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES; |
3777 | unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES; |
3732 | struct drm_i915_gem_request *request, *target = NULL; |
3778 | struct drm_i915_gem_request *request, *target = NULL; |
3733 | unsigned reset_counter; |
3779 | unsigned reset_counter; |
3734 | int ret; |
3780 | int ret; |
3735 | 3781 | ||
3736 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
3782 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
3737 | if (ret) |
3783 | if (ret) |
3738 | return ret; |
3784 | return ret; |
3739 | 3785 | ||
3740 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, false); |
3786 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, false); |
3741 | if (ret) |
3787 | if (ret) |
3742 | return ret; |
3788 | return ret; |
3743 | 3789 | ||
3744 | spin_lock(&file_priv->mm.lock); |
3790 | spin_lock(&file_priv->mm.lock); |
3745 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { |
3791 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { |
3746 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
3792 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
3747 | break; |
3793 | break; |
3748 | 3794 | ||
3749 | /* |
3795 | /* |
3750 | * Note that the request might not have been submitted yet. |
3796 | * Note that the request might not have been submitted yet. |
3751 | * In which case emitted_jiffies will be zero. |
3797 | * In which case emitted_jiffies will be zero. |
3752 | */ |
3798 | */ |
3753 | if (!request->emitted_jiffies) |
3799 | if (!request->emitted_jiffies) |
3754 | continue; |
3800 | continue; |
3755 | 3801 | ||
3756 | target = request; |
3802 | target = request; |
3757 | } |
3803 | } |
3758 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
3804 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
3759 | if (target) |
3805 | if (target) |
3760 | i915_gem_request_reference(target); |
3806 | i915_gem_request_reference(target); |
3761 | spin_unlock(&file_priv->mm.lock); |
3807 | spin_unlock(&file_priv->mm.lock); |
3762 | 3808 | ||
3763 | if (target == NULL) |
3809 | if (target == NULL) |
3764 | return 0; |
3810 | return 0; |
3765 | 3811 | ||
3766 | ret = __i915_wait_request(target, reset_counter, true, NULL, NULL); |
3812 | ret = __i915_wait_request(target, reset_counter, true, NULL, NULL); |
3767 | if (ret == 0) |
3813 | if (ret == 0) |
3768 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); |
3814 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); |
3769 | 3815 | ||
3770 | i915_gem_request_unreference__unlocked(target); |
3816 | i915_gem_request_unreference__unlocked(target); |
3771 | 3817 | ||
3772 | return ret; |
3818 | return ret; |
3773 | } |
3819 | } |
3774 | 3820 | ||
3775 | static bool |
3821 | static bool |
3776 | i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags) |
3822 | i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags) |
3777 | { |
3823 | { |
3778 | struct drm_i915_gem_object *obj = vma->obj; |
3824 | struct drm_i915_gem_object *obj = vma->obj; |
3779 | 3825 | ||
3780 | if (alignment && |
3826 | if (alignment && |
3781 | vma->node.start & (alignment - 1)) |
3827 | vma->node.start & (alignment - 1)) |
3782 | return true; |
3828 | return true; |
3783 | 3829 | ||
3784 | if (flags & PIN_MAPPABLE && !obj->map_and_fenceable) |
3830 | if (flags & PIN_MAPPABLE && !obj->map_and_fenceable) |
3785 | return true; |
3831 | return true; |
3786 | 3832 | ||
3787 | if (flags & PIN_OFFSET_BIAS && |
3833 | if (flags & PIN_OFFSET_BIAS && |
3788 | vma->node.start < (flags & PIN_OFFSET_MASK)) |
3834 | vma->node.start < (flags & PIN_OFFSET_MASK)) |
3789 | return true; |
3835 | return true; |
3790 | 3836 | ||
3791 | if (flags & PIN_OFFSET_FIXED && |
3837 | if (flags & PIN_OFFSET_FIXED && |
3792 | vma->node.start != (flags & PIN_OFFSET_MASK)) |
3838 | vma->node.start != (flags & PIN_OFFSET_MASK)) |
3793 | return true; |
3839 | return true; |
3794 | 3840 | ||
3795 | return false; |
3841 | return false; |
3796 | } |
3842 | } |
3797 | 3843 | ||
3798 | void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) |
3844 | void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) |
3799 | { |
3845 | { |
3800 | struct drm_i915_gem_object *obj = vma->obj; |
3846 | struct drm_i915_gem_object *obj = vma->obj; |
3801 | bool mappable, fenceable; |
3847 | bool mappable, fenceable; |
3802 | u32 fence_size, fence_alignment; |
3848 | u32 fence_size, fence_alignment; |
3803 | 3849 | ||
3804 | fence_size = i915_gem_get_gtt_size(obj->base.dev, |
3850 | fence_size = i915_gem_get_gtt_size(obj->base.dev, |
3805 | obj->base.size, |
3851 | obj->base.size, |
3806 | obj->tiling_mode); |
3852 | obj->tiling_mode); |
3807 | fence_alignment = i915_gem_get_gtt_alignment(obj->base.dev, |
3853 | fence_alignment = i915_gem_get_gtt_alignment(obj->base.dev, |
3808 | obj->base.size, |
3854 | obj->base.size, |
3809 | obj->tiling_mode, |
3855 | obj->tiling_mode, |
3810 | true); |
3856 | true); |
3811 | 3857 | ||
3812 | fenceable = (vma->node.size == fence_size && |
3858 | fenceable = (vma->node.size == fence_size && |
3813 | (vma->node.start & (fence_alignment - 1)) == 0); |
3859 | (vma->node.start & (fence_alignment - 1)) == 0); |
3814 | 3860 | ||
3815 | mappable = (vma->node.start + fence_size <= |
3861 | mappable = (vma->node.start + fence_size <= |
3816 | to_i915(obj->base.dev)->gtt.mappable_end); |
3862 | to_i915(obj->base.dev)->gtt.mappable_end); |
3817 | 3863 | ||
3818 | obj->map_and_fenceable = mappable && fenceable; |
3864 | obj->map_and_fenceable = mappable && fenceable; |
3819 | } |
3865 | } |
3820 | 3866 | ||
3821 | static int |
3867 | static int |
3822 | i915_gem_object_do_pin(struct drm_i915_gem_object *obj, |
3868 | i915_gem_object_do_pin(struct drm_i915_gem_object *obj, |
3823 | struct i915_address_space *vm, |
3869 | struct i915_address_space *vm, |
3824 | const struct i915_ggtt_view *ggtt_view, |
3870 | const struct i915_ggtt_view *ggtt_view, |
3825 | uint32_t alignment, |
3871 | uint32_t alignment, |
3826 | uint64_t flags) |
3872 | uint64_t flags) |
3827 | { |
3873 | { |
3828 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
3874 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
3829 | struct i915_vma *vma; |
3875 | struct i915_vma *vma; |
3830 | unsigned bound; |
3876 | unsigned bound; |
3831 | int ret; |
3877 | int ret; |
3832 | 3878 | ||
3833 | if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base)) |
3879 | if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base)) |
3834 | return -ENODEV; |
3880 | return -ENODEV; |
3835 | 3881 | ||
3836 | if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm))) |
3882 | if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm))) |
3837 | return -EINVAL; |
3883 | return -EINVAL; |
3838 | 3884 | ||
3839 | if (WARN_ON((flags & (PIN_MAPPABLE | PIN_GLOBAL)) == PIN_MAPPABLE)) |
3885 | if (WARN_ON((flags & (PIN_MAPPABLE | PIN_GLOBAL)) == PIN_MAPPABLE)) |
3840 | return -EINVAL; |
3886 | return -EINVAL; |
3841 | 3887 | ||
3842 | if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view)) |
3888 | if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view)) |
3843 | return -EINVAL; |
3889 | return -EINVAL; |
3844 | 3890 | ||
3845 | vma = ggtt_view ? i915_gem_obj_to_ggtt_view(obj, ggtt_view) : |
3891 | vma = ggtt_view ? i915_gem_obj_to_ggtt_view(obj, ggtt_view) : |
3846 | i915_gem_obj_to_vma(obj, vm); |
3892 | i915_gem_obj_to_vma(obj, vm); |
3847 | 3893 | ||
3848 | if (IS_ERR(vma)) |
3894 | if (IS_ERR(vma)) |
3849 | return PTR_ERR(vma); |
3895 | return PTR_ERR(vma); |
3850 | 3896 | ||
3851 | if (vma) { |
3897 | if (vma) { |
3852 | if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT)) |
3898 | if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT)) |
3853 | return -EBUSY; |
3899 | return -EBUSY; |
3854 | 3900 | ||
3855 | if (i915_vma_misplaced(vma, alignment, flags)) { |
3901 | if (i915_vma_misplaced(vma, alignment, flags)) { |
3856 | WARN(vma->pin_count, |
3902 | WARN(vma->pin_count, |
3857 | "bo is already pinned in %s with incorrect alignment:" |
3903 | "bo is already pinned in %s with incorrect alignment:" |
3858 | " offset=%08x %08x, req.alignment=%x, req.map_and_fenceable=%d," |
3904 | " offset=%08x %08x, req.alignment=%x, req.map_and_fenceable=%d," |
3859 | " obj->map_and_fenceable=%d\n", |
3905 | " obj->map_and_fenceable=%d\n", |
3860 | ggtt_view ? "ggtt" : "ppgtt", |
3906 | ggtt_view ? "ggtt" : "ppgtt", |
3861 | upper_32_bits(vma->node.start), |
3907 | upper_32_bits(vma->node.start), |
3862 | lower_32_bits(vma->node.start), |
3908 | lower_32_bits(vma->node.start), |
3863 | alignment, |
3909 | alignment, |
3864 | !!(flags & PIN_MAPPABLE), |
3910 | !!(flags & PIN_MAPPABLE), |
3865 | obj->map_and_fenceable); |
3911 | obj->map_and_fenceable); |
3866 | ret = i915_vma_unbind(vma); |
3912 | ret = i915_vma_unbind(vma); |
3867 | if (ret) |
3913 | if (ret) |
3868 | return ret; |
3914 | return ret; |
3869 | 3915 | ||
3870 | vma = NULL; |
3916 | vma = NULL; |
3871 | } |
3917 | } |
3872 | } |
3918 | } |
3873 | 3919 | ||
3874 | bound = vma ? vma->bound : 0; |
3920 | bound = vma ? vma->bound : 0; |
3875 | if (vma == NULL || !drm_mm_node_allocated(&vma->node)) { |
3921 | if (vma == NULL || !drm_mm_node_allocated(&vma->node)) { |
3876 | vma = i915_gem_object_bind_to_vm(obj, vm, ggtt_view, alignment, |
3922 | vma = i915_gem_object_bind_to_vm(obj, vm, ggtt_view, alignment, |
3877 | flags); |
3923 | flags); |
3878 | if (IS_ERR(vma)) |
3924 | if (IS_ERR(vma)) |
3879 | return PTR_ERR(vma); |
3925 | return PTR_ERR(vma); |
3880 | } else { |
3926 | } else { |
3881 | ret = i915_vma_bind(vma, obj->cache_level, flags); |
3927 | ret = i915_vma_bind(vma, obj->cache_level, flags); |
3882 | if (ret) |
3928 | if (ret) |
3883 | return ret; |
3929 | return ret; |
3884 | } |
3930 | } |
3885 | 3931 | ||
3886 | if (ggtt_view && ggtt_view->type == I915_GGTT_VIEW_NORMAL && |
3932 | if (ggtt_view && ggtt_view->type == I915_GGTT_VIEW_NORMAL && |
3887 | (bound ^ vma->bound) & GLOBAL_BIND) { |
3933 | (bound ^ vma->bound) & GLOBAL_BIND) { |
3888 | __i915_vma_set_map_and_fenceable(vma); |
3934 | __i915_vma_set_map_and_fenceable(vma); |
3889 | WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable); |
3935 | WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable); |
3890 | } |
3936 | } |
3891 | 3937 | ||
3892 | vma->pin_count++; |
3938 | vma->pin_count++; |
3893 | return 0; |
3939 | return 0; |
3894 | } |
3940 | } |
3895 | 3941 | ||
3896 | int |
3942 | int |
3897 | i915_gem_object_pin(struct drm_i915_gem_object *obj, |
3943 | i915_gem_object_pin(struct drm_i915_gem_object *obj, |
3898 | struct i915_address_space *vm, |
3944 | struct i915_address_space *vm, |
3899 | uint32_t alignment, |
3945 | uint32_t alignment, |
3900 | uint64_t flags) |
3946 | uint64_t flags) |
3901 | { |
3947 | { |
3902 | return i915_gem_object_do_pin(obj, vm, |
3948 | return i915_gem_object_do_pin(obj, vm, |
3903 | i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL, |
3949 | i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL, |
3904 | alignment, flags); |
3950 | alignment, flags); |
3905 | } |
3951 | } |
3906 | 3952 | ||
3907 | int |
3953 | int |
3908 | i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj, |
3954 | i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj, |
3909 | const struct i915_ggtt_view *view, |
3955 | const struct i915_ggtt_view *view, |
3910 | uint32_t alignment, |
3956 | uint32_t alignment, |
3911 | uint64_t flags) |
3957 | uint64_t flags) |
3912 | { |
3958 | { |
3913 | if (WARN_ONCE(!view, "no view specified")) |
3959 | if (WARN_ONCE(!view, "no view specified")) |
3914 | return -EINVAL; |
3960 | return -EINVAL; |
3915 | 3961 | ||
3916 | return i915_gem_object_do_pin(obj, i915_obj_to_ggtt(obj), view, |
3962 | return i915_gem_object_do_pin(obj, i915_obj_to_ggtt(obj), view, |
3917 | alignment, flags | PIN_GLOBAL); |
3963 | alignment, flags | PIN_GLOBAL); |
3918 | } |
3964 | } |
3919 | 3965 | ||
3920 | void |
3966 | void |
3921 | i915_gem_object_ggtt_unpin_view(struct drm_i915_gem_object *obj, |
3967 | i915_gem_object_ggtt_unpin_view(struct drm_i915_gem_object *obj, |
3922 | const struct i915_ggtt_view *view) |
3968 | const struct i915_ggtt_view *view) |
3923 | { |
3969 | { |
3924 | struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view); |
3970 | struct i915_vma *vma = i915_gem_obj_to_ggtt_view(obj, view); |
3925 | 3971 | ||
3926 | BUG_ON(!vma); |
3972 | BUG_ON(!vma); |
3927 | WARN_ON(vma->pin_count == 0); |
3973 | WARN_ON(vma->pin_count == 0); |
3928 | WARN_ON(!i915_gem_obj_ggtt_bound_view(obj, view)); |
3974 | WARN_ON(!i915_gem_obj_ggtt_bound_view(obj, view)); |
3929 | 3975 | ||
3930 | --vma->pin_count; |
3976 | --vma->pin_count; |
3931 | } |
3977 | } |
3932 | 3978 | ||
3933 | int |
3979 | int |
3934 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, |
3980 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, |
3935 | struct drm_file *file) |
3981 | struct drm_file *file) |
3936 | { |
3982 | { |
3937 | struct drm_i915_gem_busy *args = data; |
3983 | struct drm_i915_gem_busy *args = data; |
3938 | struct drm_i915_gem_object *obj; |
3984 | struct drm_i915_gem_object *obj; |
3939 | int ret; |
3985 | int ret; |
3940 | 3986 | ||
3941 | ret = i915_mutex_lock_interruptible(dev); |
3987 | ret = i915_mutex_lock_interruptible(dev); |
3942 | if (ret) |
3988 | if (ret) |
3943 | return ret; |
3989 | return ret; |
3944 | 3990 | ||
3945 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3991 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3946 | if (&obj->base == NULL) { |
3992 | if (&obj->base == NULL) { |
3947 | ret = -ENOENT; |
3993 | ret = -ENOENT; |
3948 | goto unlock; |
3994 | goto unlock; |
3949 | } |
3995 | } |
3950 | 3996 | ||
3951 | /* Count all active objects as busy, even if they are currently not used |
3997 | /* Count all active objects as busy, even if they are currently not used |
3952 | * by the gpu. Users of this interface expect objects to eventually |
3998 | * by the gpu. Users of this interface expect objects to eventually |
3953 | * become non-busy without any further actions, therefore emit any |
3999 | * become non-busy without any further actions, therefore emit any |
3954 | * necessary flushes here. |
4000 | * necessary flushes here. |
3955 | */ |
4001 | */ |
3956 | ret = i915_gem_object_flush_active(obj); |
4002 | ret = i915_gem_object_flush_active(obj); |
3957 | if (ret) |
4003 | if (ret) |
3958 | goto unref; |
4004 | goto unref; |
- | 4005 | ||
- | 4006 | args->busy = 0; |
|
- | 4007 | if (obj->active) { |
|
- | 4008 | int i; |
|
3959 | 4009 | ||
- | 4010 | for (i = 0; i < I915_NUM_RINGS; i++) { |
|
- | 4011 | struct drm_i915_gem_request *req; |
|
3960 | BUILD_BUG_ON(I915_NUM_RINGS > 16); |
4012 | |
- | 4013 | req = obj->last_read_req[i]; |
|
- | 4014 | if (req) |
|
- | 4015 | args->busy |= 1 << (16 + req->ring->exec_id); |
|
3961 | args->busy = obj->active << 16; |
4016 | } |
3962 | if (obj->last_write_req) |
4017 | if (obj->last_write_req) |
- | 4018 | args->busy |= obj->last_write_req->ring->exec_id; |
|
3963 | args->busy |= obj->last_write_req->ring->id; |
4019 | } |
3964 | 4020 | ||
3965 | unref: |
4021 | unref: |
3966 | drm_gem_object_unreference(&obj->base); |
4022 | drm_gem_object_unreference(&obj->base); |
3967 | unlock: |
4023 | unlock: |
3968 | mutex_unlock(&dev->struct_mutex); |
4024 | mutex_unlock(&dev->struct_mutex); |
3969 | return ret; |
4025 | return ret; |
3970 | } |
4026 | } |
3971 | 4027 | ||
3972 | int |
4028 | int |
3973 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, |
4029 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, |
3974 | struct drm_file *file_priv) |
4030 | struct drm_file *file_priv) |
3975 | { |
4031 | { |
3976 | return i915_gem_ring_throttle(dev, file_priv); |
4032 | return i915_gem_ring_throttle(dev, file_priv); |
3977 | } |
4033 | } |
3978 | 4034 | ||
3979 | #if 0 |
4035 | #if 0 |
3980 | 4036 | ||
3981 | int |
4037 | int |
3982 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, |
4038 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, |
3983 | struct drm_file *file_priv) |
4039 | struct drm_file *file_priv) |
3984 | { |
4040 | { |
3985 | struct drm_i915_private *dev_priv = dev->dev_private; |
4041 | struct drm_i915_private *dev_priv = dev->dev_private; |
3986 | struct drm_i915_gem_madvise *args = data; |
4042 | struct drm_i915_gem_madvise *args = data; |
3987 | struct drm_i915_gem_object *obj; |
4043 | struct drm_i915_gem_object *obj; |
3988 | int ret; |
4044 | int ret; |
3989 | 4045 | ||
3990 | switch (args->madv) { |
4046 | switch (args->madv) { |
3991 | case I915_MADV_DONTNEED: |
4047 | case I915_MADV_DONTNEED: |
3992 | case I915_MADV_WILLNEED: |
4048 | case I915_MADV_WILLNEED: |
3993 | break; |
4049 | break; |
3994 | default: |
4050 | default: |
3995 | return -EINVAL; |
4051 | return -EINVAL; |
3996 | } |
4052 | } |
3997 | 4053 | ||
3998 | ret = i915_mutex_lock_interruptible(dev); |
4054 | ret = i915_mutex_lock_interruptible(dev); |
3999 | if (ret) |
4055 | if (ret) |
4000 | return ret; |
4056 | return ret; |
4001 | 4057 | ||
4002 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); |
4058 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); |
4003 | if (&obj->base == NULL) { |
4059 | if (&obj->base == NULL) { |
4004 | ret = -ENOENT; |
4060 | ret = -ENOENT; |
4005 | goto unlock; |
4061 | goto unlock; |
4006 | } |
4062 | } |
4007 | 4063 | ||
4008 | if (i915_gem_obj_is_pinned(obj)) { |
4064 | if (i915_gem_obj_is_pinned(obj)) { |
4009 | ret = -EINVAL; |
4065 | ret = -EINVAL; |
4010 | goto out; |
4066 | goto out; |
4011 | } |
4067 | } |
4012 | 4068 | ||
4013 | if (obj->pages && |
4069 | if (obj->pages && |
4014 | obj->tiling_mode != I915_TILING_NONE && |
4070 | obj->tiling_mode != I915_TILING_NONE && |
4015 | dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) { |
4071 | dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) { |
4016 | if (obj->madv == I915_MADV_WILLNEED) |
4072 | if (obj->madv == I915_MADV_WILLNEED) |
4017 | i915_gem_object_unpin_pages(obj); |
4073 | i915_gem_object_unpin_pages(obj); |
4018 | if (args->madv == I915_MADV_WILLNEED) |
4074 | if (args->madv == I915_MADV_WILLNEED) |
4019 | i915_gem_object_pin_pages(obj); |
4075 | i915_gem_object_pin_pages(obj); |
4020 | } |
4076 | } |
4021 | 4077 | ||
4022 | if (obj->madv != __I915_MADV_PURGED) |
4078 | if (obj->madv != __I915_MADV_PURGED) |
4023 | obj->madv = args->madv; |
4079 | obj->madv = args->madv; |
4024 | 4080 | ||
4025 | /* if the object is no longer attached, discard its backing storage */ |
4081 | /* if the object is no longer attached, discard its backing storage */ |
4026 | if (obj->madv == I915_MADV_DONTNEED && obj->pages == NULL) |
4082 | if (obj->madv == I915_MADV_DONTNEED && obj->pages == NULL) |
4027 | i915_gem_object_truncate(obj); |
4083 | i915_gem_object_truncate(obj); |
4028 | 4084 | ||
4029 | args->retained = obj->madv != __I915_MADV_PURGED; |
4085 | args->retained = obj->madv != __I915_MADV_PURGED; |
4030 | 4086 | ||
4031 | out: |
4087 | out: |
4032 | drm_gem_object_unreference(&obj->base); |
4088 | drm_gem_object_unreference(&obj->base); |
4033 | unlock: |
4089 | unlock: |
4034 | mutex_unlock(&dev->struct_mutex); |
4090 | mutex_unlock(&dev->struct_mutex); |
4035 | return ret; |
4091 | return ret; |
4036 | } |
4092 | } |
4037 | #endif |
4093 | #endif |
4038 | 4094 | ||
4039 | void i915_gem_object_init(struct drm_i915_gem_object *obj, |
4095 | void i915_gem_object_init(struct drm_i915_gem_object *obj, |
4040 | const struct drm_i915_gem_object_ops *ops) |
4096 | const struct drm_i915_gem_object_ops *ops) |
4041 | { |
4097 | { |
4042 | int i; |
4098 | int i; |
4043 | 4099 | ||
4044 | INIT_LIST_HEAD(&obj->global_list); |
4100 | INIT_LIST_HEAD(&obj->global_list); |
4045 | for (i = 0; i < I915_NUM_RINGS; i++) |
4101 | for (i = 0; i < I915_NUM_RINGS; i++) |
4046 | INIT_LIST_HEAD(&obj->ring_list[i]); |
4102 | INIT_LIST_HEAD(&obj->ring_list[i]); |
4047 | INIT_LIST_HEAD(&obj->obj_exec_link); |
4103 | INIT_LIST_HEAD(&obj->obj_exec_link); |
4048 | INIT_LIST_HEAD(&obj->vma_list); |
4104 | INIT_LIST_HEAD(&obj->vma_list); |
4049 | INIT_LIST_HEAD(&obj->batch_pool_link); |
4105 | INIT_LIST_HEAD(&obj->batch_pool_link); |
4050 | 4106 | ||
4051 | obj->ops = ops; |
4107 | obj->ops = ops; |
4052 | 4108 | ||
4053 | obj->fence_reg = I915_FENCE_REG_NONE; |
4109 | obj->fence_reg = I915_FENCE_REG_NONE; |
4054 | obj->madv = I915_MADV_WILLNEED; |
4110 | obj->madv = I915_MADV_WILLNEED; |
4055 | 4111 | ||
4056 | i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size); |
4112 | i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size); |
4057 | } |
4113 | } |
4058 | 4114 | ||
4059 | static const struct drm_i915_gem_object_ops i915_gem_object_ops = { |
4115 | static const struct drm_i915_gem_object_ops i915_gem_object_ops = { |
4060 | .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE, |
4116 | .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE, |
4061 | .get_pages = i915_gem_object_get_pages_gtt, |
4117 | .get_pages = i915_gem_object_get_pages_gtt, |
4062 | .put_pages = i915_gem_object_put_pages_gtt, |
4118 | .put_pages = i915_gem_object_put_pages_gtt, |
4063 | }; |
4119 | }; |
4064 | 4120 | ||
4065 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, |
4121 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, |
4066 | size_t size) |
4122 | size_t size) |
4067 | { |
4123 | { |
4068 | struct drm_i915_gem_object *obj; |
4124 | struct drm_i915_gem_object *obj; |
4069 | struct address_space *mapping; |
4125 | struct address_space *mapping; |
4070 | gfp_t mask; |
4126 | gfp_t mask; |
4071 | 4127 | ||
4072 | obj = i915_gem_object_alloc(dev); |
4128 | obj = i915_gem_object_alloc(dev); |
4073 | if (obj == NULL) |
4129 | if (obj == NULL) |
4074 | return NULL; |
4130 | return NULL; |
4075 | 4131 | ||
4076 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { |
4132 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { |
4077 | i915_gem_object_free(obj); |
4133 | i915_gem_object_free(obj); |
4078 | return NULL; |
4134 | return NULL; |
4079 | } |
4135 | } |
4080 | 4136 | ||
4081 | 4137 | ||
4082 | i915_gem_object_init(obj, &i915_gem_object_ops); |
4138 | i915_gem_object_init(obj, &i915_gem_object_ops); |
4083 | 4139 | ||
4084 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
4140 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
4085 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
4141 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
4086 | 4142 | ||
4087 | if (HAS_LLC(dev)) { |
4143 | if (HAS_LLC(dev)) { |
4088 | /* On some devices, we can have the GPU use the LLC (the CPU |
4144 | /* On some devices, we can have the GPU use the LLC (the CPU |
4089 | * cache) for about a 10% performance improvement |
4145 | * cache) for about a 10% performance improvement |
4090 | * compared to uncached. Graphics requests other than |
4146 | * compared to uncached. Graphics requests other than |
4091 | * display scanout are coherent with the CPU in |
4147 | * display scanout are coherent with the CPU in |
4092 | * accessing this cache. This means in this mode we |
4148 | * accessing this cache. This means in this mode we |
4093 | * don't need to clflush on the CPU side, and on the |
4149 | * don't need to clflush on the CPU side, and on the |
4094 | * GPU side we only need to flush internal caches to |
4150 | * GPU side we only need to flush internal caches to |
4095 | * get data visible to the CPU. |
4151 | * get data visible to the CPU. |
4096 | * |
4152 | * |
4097 | * However, we maintain the display planes as UC, and so |
4153 | * However, we maintain the display planes as UC, and so |
4098 | * need to rebind when first used as such. |
4154 | * need to rebind when first used as such. |
4099 | */ |
4155 | */ |
4100 | obj->cache_level = I915_CACHE_LLC; |
4156 | obj->cache_level = I915_CACHE_LLC; |
4101 | } else |
4157 | } else |
4102 | obj->cache_level = I915_CACHE_NONE; |
4158 | obj->cache_level = I915_CACHE_NONE; |
4103 | 4159 | ||
4104 | trace_i915_gem_object_create(obj); |
4160 | trace_i915_gem_object_create(obj); |
4105 | 4161 | ||
4106 | return obj; |
4162 | return obj; |
4107 | } |
4163 | } |
4108 | 4164 | ||
4109 | static bool discard_backing_storage(struct drm_i915_gem_object *obj) |
4165 | static bool discard_backing_storage(struct drm_i915_gem_object *obj) |
4110 | { |
4166 | { |
4111 | /* If we are the last user of the backing storage (be it shmemfs |
4167 | /* If we are the last user of the backing storage (be it shmemfs |
4112 | * pages or stolen etc), we know that the pages are going to be |
4168 | * pages or stolen etc), we know that the pages are going to be |
4113 | * immediately released. In this case, we can then skip copying |
4169 | * immediately released. In this case, we can then skip copying |
4114 | * back the contents from the GPU. |
4170 | * back the contents from the GPU. |
4115 | */ |
4171 | */ |
4116 | 4172 | ||
4117 | if (obj->madv != I915_MADV_WILLNEED) |
4173 | if (obj->madv != I915_MADV_WILLNEED) |
4118 | return false; |
4174 | return false; |
4119 | 4175 | ||
4120 | if (obj->base.filp == NULL) |
4176 | if (obj->base.filp == NULL) |
4121 | return true; |
4177 | return true; |
4122 | 4178 | ||
4123 | // printf("filp %p\n", obj->base.filp); |
4179 | // printf("filp %p\n", obj->base.filp); |
4124 | shmem_file_delete(obj->base.filp); |
4180 | shmem_file_delete(obj->base.filp); |
4125 | return true; |
4181 | return true; |
4126 | } |
4182 | } |
4127 | 4183 | ||
4128 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
4184 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
4129 | { |
4185 | { |
4130 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); |
4186 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); |
4131 | struct drm_device *dev = obj->base.dev; |
4187 | struct drm_device *dev = obj->base.dev; |
4132 | struct drm_i915_private *dev_priv = dev->dev_private; |
4188 | struct drm_i915_private *dev_priv = dev->dev_private; |
4133 | struct i915_vma *vma, *next; |
4189 | struct i915_vma *vma, *next; |
4134 | 4190 | ||
4135 | intel_runtime_pm_get(dev_priv); |
4191 | intel_runtime_pm_get(dev_priv); |
4136 | 4192 | ||
4137 | trace_i915_gem_object_destroy(obj); |
4193 | trace_i915_gem_object_destroy(obj); |
4138 | 4194 | ||
4139 | list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) { |
4195 | list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link) { |
4140 | int ret; |
4196 | int ret; |
4141 | 4197 | ||
4142 | vma->pin_count = 0; |
4198 | vma->pin_count = 0; |
4143 | ret = i915_vma_unbind(vma); |
4199 | ret = i915_vma_unbind(vma); |
4144 | if (WARN_ON(ret == -ERESTARTSYS)) { |
4200 | if (WARN_ON(ret == -ERESTARTSYS)) { |
4145 | bool was_interruptible; |
4201 | bool was_interruptible; |
4146 | 4202 | ||
4147 | was_interruptible = dev_priv->mm.interruptible; |
4203 | was_interruptible = dev_priv->mm.interruptible; |
4148 | dev_priv->mm.interruptible = false; |
4204 | dev_priv->mm.interruptible = false; |
4149 | 4205 | ||
4150 | WARN_ON(i915_vma_unbind(vma)); |
4206 | WARN_ON(i915_vma_unbind(vma)); |
4151 | 4207 | ||
4152 | dev_priv->mm.interruptible = was_interruptible; |
4208 | dev_priv->mm.interruptible = was_interruptible; |
4153 | } |
4209 | } |
4154 | } |
4210 | } |
4155 | 4211 | ||
4156 | /* Stolen objects don't hold a ref, but do hold pin count. Fix that up |
4212 | /* Stolen objects don't hold a ref, but do hold pin count. Fix that up |
4157 | * before progressing. */ |
4213 | * before progressing. */ |
4158 | if (obj->stolen) |
4214 | if (obj->stolen) |
4159 | i915_gem_object_unpin_pages(obj); |
4215 | i915_gem_object_unpin_pages(obj); |
4160 | 4216 | ||
4161 | WARN_ON(obj->frontbuffer_bits); |
4217 | WARN_ON(obj->frontbuffer_bits); |
4162 | 4218 | ||
4163 | if (obj->pages && obj->madv == I915_MADV_WILLNEED && |
4219 | if (obj->pages && obj->madv == I915_MADV_WILLNEED && |
4164 | dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES && |
4220 | dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES && |
4165 | obj->tiling_mode != I915_TILING_NONE) |
4221 | obj->tiling_mode != I915_TILING_NONE) |
4166 | i915_gem_object_unpin_pages(obj); |
4222 | i915_gem_object_unpin_pages(obj); |
4167 | 4223 | ||
4168 | if (WARN_ON(obj->pages_pin_count)) |
4224 | if (WARN_ON(obj->pages_pin_count)) |
4169 | obj->pages_pin_count = 0; |
4225 | obj->pages_pin_count = 0; |
4170 | if (discard_backing_storage(obj)) |
4226 | if (discard_backing_storage(obj)) |
4171 | obj->madv = I915_MADV_DONTNEED; |
4227 | obj->madv = I915_MADV_DONTNEED; |
4172 | i915_gem_object_put_pages(obj); |
4228 | i915_gem_object_put_pages(obj); |
4173 | // i915_gem_object_free_mmap_offset(obj); |
4229 | // i915_gem_object_free_mmap_offset(obj); |
4174 | 4230 | ||
4175 | BUG_ON(obj->pages); |
4231 | BUG_ON(obj->pages); |
4176 | 4232 | ||
4177 | if (obj->ops->release) |
4233 | if (obj->ops->release) |
4178 | obj->ops->release(obj); |
4234 | obj->ops->release(obj); |
4179 | 4235 | ||
4180 | drm_gem_object_release(&obj->base); |
4236 | drm_gem_object_release(&obj->base); |
4181 | i915_gem_info_remove_obj(dev_priv, obj->base.size); |
4237 | i915_gem_info_remove_obj(dev_priv, obj->base.size); |
4182 | 4238 | ||
4183 | kfree(obj->bit_17); |
4239 | kfree(obj->bit_17); |
4184 | i915_gem_object_free(obj); |
4240 | i915_gem_object_free(obj); |
4185 | 4241 | ||
4186 | intel_runtime_pm_put(dev_priv); |
4242 | intel_runtime_pm_put(dev_priv); |
4187 | } |
4243 | } |
4188 | 4244 | ||
4189 | struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj, |
4245 | struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj, |
4190 | struct i915_address_space *vm) |
4246 | struct i915_address_space *vm) |
4191 | { |
4247 | { |
4192 | struct i915_vma *vma; |
4248 | struct i915_vma *vma; |
4193 | list_for_each_entry(vma, &obj->vma_list, vma_link) { |
4249 | list_for_each_entry(vma, &obj->vma_list, obj_link) { |
4194 | if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL && |
4250 | if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL && |
4195 | vma->vm == vm) |
4251 | vma->vm == vm) |
4196 | return vma; |
4252 | return vma; |
4197 | } |
4253 | } |
4198 | return NULL; |
4254 | return NULL; |
4199 | } |
4255 | } |
4200 | 4256 | ||
4201 | struct i915_vma *i915_gem_obj_to_ggtt_view(struct drm_i915_gem_object *obj, |
4257 | struct i915_vma *i915_gem_obj_to_ggtt_view(struct drm_i915_gem_object *obj, |
4202 | const struct i915_ggtt_view *view) |
4258 | const struct i915_ggtt_view *view) |
4203 | { |
4259 | { |
4204 | struct i915_address_space *ggtt = i915_obj_to_ggtt(obj); |
4260 | struct i915_address_space *ggtt = i915_obj_to_ggtt(obj); |
4205 | struct i915_vma *vma; |
4261 | struct i915_vma *vma; |
4206 | 4262 | ||
4207 | if (WARN_ONCE(!view, "no view specified")) |
4263 | if (WARN_ONCE(!view, "no view specified")) |
4208 | return ERR_PTR(-EINVAL); |
4264 | return ERR_PTR(-EINVAL); |
4209 | 4265 | ||
4210 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
4266 | list_for_each_entry(vma, &obj->vma_list, obj_link) |
4211 | if (vma->vm == ggtt && |
4267 | if (vma->vm == ggtt && |
4212 | i915_ggtt_view_equal(&vma->ggtt_view, view)) |
4268 | i915_ggtt_view_equal(&vma->ggtt_view, view)) |
4213 | return vma; |
4269 | return vma; |
4214 | return NULL; |
4270 | return NULL; |
4215 | } |
4271 | } |
4216 | 4272 | ||
4217 | void i915_gem_vma_destroy(struct i915_vma *vma) |
4273 | void i915_gem_vma_destroy(struct i915_vma *vma) |
4218 | { |
4274 | { |
4219 | struct i915_address_space *vm = NULL; |
- | |
4220 | WARN_ON(vma->node.allocated); |
4275 | WARN_ON(vma->node.allocated); |
4221 | 4276 | ||
4222 | /* Keep the vma as a placeholder in the execbuffer reservation lists */ |
4277 | /* Keep the vma as a placeholder in the execbuffer reservation lists */ |
4223 | if (!list_empty(&vma->exec_list)) |
4278 | if (!list_empty(&vma->exec_list)) |
4224 | return; |
4279 | return; |
4225 | - | ||
4226 | vm = vma->vm; |
- | |
4227 | 4280 | ||
4228 | if (!i915_is_ggtt(vm)) |
4281 | if (!vma->is_ggtt) |
4229 | i915_ppgtt_put(i915_vm_to_ppgtt(vm)); |
4282 | i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm)); |
4230 | 4283 | ||
4231 | list_del(&vma->vma_link); |
4284 | list_del(&vma->obj_link); |
4232 | 4285 | ||
4233 | kfree(vma); |
4286 | kfree(vma); |
4234 | } |
4287 | } |
4235 | 4288 | ||
4236 | static void |
4289 | static void |
4237 | i915_gem_stop_ringbuffers(struct drm_device *dev) |
4290 | i915_gem_stop_ringbuffers(struct drm_device *dev) |
4238 | { |
4291 | { |
4239 | struct drm_i915_private *dev_priv = dev->dev_private; |
4292 | struct drm_i915_private *dev_priv = dev->dev_private; |
4240 | struct intel_engine_cs *ring; |
4293 | struct intel_engine_cs *ring; |
4241 | int i; |
4294 | int i; |
4242 | 4295 | ||
4243 | for_each_ring(ring, dev_priv, i) |
4296 | for_each_ring(ring, dev_priv, i) |
4244 | dev_priv->gt.stop_ring(ring); |
4297 | dev_priv->gt.stop_ring(ring); |
4245 | } |
4298 | } |
4246 | 4299 | ||
4247 | #if 0 |
4300 | #if 0 |
4248 | int |
4301 | int |
4249 | i915_gem_suspend(struct drm_device *dev) |
4302 | i915_gem_suspend(struct drm_device *dev) |
4250 | { |
4303 | { |
4251 | struct drm_i915_private *dev_priv = dev->dev_private; |
4304 | struct drm_i915_private *dev_priv = dev->dev_private; |
4252 | int ret = 0; |
4305 | int ret = 0; |
4253 | 4306 | ||
4254 | mutex_lock(&dev->struct_mutex); |
4307 | mutex_lock(&dev->struct_mutex); |
4255 | ret = i915_gpu_idle(dev); |
4308 | ret = i915_gpu_idle(dev); |
4256 | if (ret) |
4309 | if (ret) |
4257 | goto err; |
4310 | goto err; |
4258 | 4311 | ||
4259 | i915_gem_retire_requests(dev); |
4312 | i915_gem_retire_requests(dev); |
4260 | 4313 | ||
4261 | i915_gem_stop_ringbuffers(dev); |
4314 | i915_gem_stop_ringbuffers(dev); |
4262 | mutex_unlock(&dev->struct_mutex); |
4315 | mutex_unlock(&dev->struct_mutex); |
4263 | 4316 | ||
4264 | cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work); |
4317 | cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work); |
4265 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); |
4318 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); |
4266 | flush_delayed_work(&dev_priv->mm.idle_work); |
4319 | flush_delayed_work(&dev_priv->mm.idle_work); |
4267 | 4320 | ||
4268 | /* Assert that we sucessfully flushed all the work and |
4321 | /* Assert that we sucessfully flushed all the work and |
4269 | * reset the GPU back to its idle, low power state. |
4322 | * reset the GPU back to its idle, low power state. |
4270 | */ |
4323 | */ |
4271 | WARN_ON(dev_priv->mm.busy); |
4324 | WARN_ON(dev_priv->mm.busy); |
4272 | 4325 | ||
4273 | return 0; |
4326 | return 0; |
4274 | 4327 | ||
4275 | err: |
4328 | err: |
4276 | mutex_unlock(&dev->struct_mutex); |
4329 | mutex_unlock(&dev->struct_mutex); |
4277 | return ret; |
4330 | return ret; |
4278 | } |
4331 | } |
4279 | #endif |
4332 | #endif |
4280 | 4333 | ||
4281 | int i915_gem_l3_remap(struct drm_i915_gem_request *req, int slice) |
4334 | int i915_gem_l3_remap(struct drm_i915_gem_request *req, int slice) |
4282 | { |
4335 | { |
4283 | struct intel_engine_cs *ring = req->ring; |
4336 | struct intel_engine_cs *ring = req->ring; |
4284 | struct drm_device *dev = ring->dev; |
4337 | struct drm_device *dev = ring->dev; |
4285 | struct drm_i915_private *dev_priv = dev->dev_private; |
4338 | struct drm_i915_private *dev_priv = dev->dev_private; |
4286 | u32 *remap_info = dev_priv->l3_parity.remap_info[slice]; |
4339 | u32 *remap_info = dev_priv->l3_parity.remap_info[slice]; |
4287 | int i, ret; |
4340 | int i, ret; |
4288 | 4341 | ||
4289 | if (!HAS_L3_DPF(dev) || !remap_info) |
4342 | if (!HAS_L3_DPF(dev) || !remap_info) |
4290 | return 0; |
4343 | return 0; |
4291 | 4344 | ||
4292 | ret = intel_ring_begin(req, GEN7_L3LOG_SIZE / 4 * 3); |
4345 | ret = intel_ring_begin(req, GEN7_L3LOG_SIZE / 4 * 3); |
4293 | if (ret) |
4346 | if (ret) |
4294 | return ret; |
4347 | return ret; |
4295 | 4348 | ||
4296 | /* |
4349 | /* |
4297 | * Note: We do not worry about the concurrent register cacheline hang |
4350 | * Note: We do not worry about the concurrent register cacheline hang |
4298 | * here because no other code should access these registers other than |
4351 | * here because no other code should access these registers other than |
4299 | * at initialization time. |
4352 | * at initialization time. |
4300 | */ |
4353 | */ |
4301 | for (i = 0; i < GEN7_L3LOG_SIZE / 4; i++) { |
4354 | for (i = 0; i < GEN7_L3LOG_SIZE / 4; i++) { |
4302 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1)); |
4355 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1)); |
4303 | intel_ring_emit_reg(ring, GEN7_L3LOG(slice, i)); |
4356 | intel_ring_emit_reg(ring, GEN7_L3LOG(slice, i)); |
4304 | intel_ring_emit(ring, remap_info[i]); |
4357 | intel_ring_emit(ring, remap_info[i]); |
4305 | } |
4358 | } |
4306 | 4359 | ||
4307 | intel_ring_advance(ring); |
4360 | intel_ring_advance(ring); |
4308 | 4361 | ||
4309 | return ret; |
4362 | return ret; |
4310 | } |
4363 | } |
4311 | 4364 | ||
4312 | void i915_gem_init_swizzling(struct drm_device *dev) |
4365 | void i915_gem_init_swizzling(struct drm_device *dev) |
4313 | { |
4366 | { |
4314 | struct drm_i915_private *dev_priv = dev->dev_private; |
4367 | struct drm_i915_private *dev_priv = dev->dev_private; |
4315 | 4368 | ||
4316 | if (INTEL_INFO(dev)->gen < 5 || |
4369 | if (INTEL_INFO(dev)->gen < 5 || |
4317 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
4370 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
4318 | return; |
4371 | return; |
4319 | 4372 | ||
4320 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | |
4373 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | |
4321 | DISP_TILE_SURFACE_SWIZZLING); |
4374 | DISP_TILE_SURFACE_SWIZZLING); |
4322 | 4375 | ||
4323 | if (IS_GEN5(dev)) |
4376 | if (IS_GEN5(dev)) |
4324 | return; |
4377 | return; |
4325 | 4378 | ||
4326 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
4379 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
4327 | if (IS_GEN6(dev)) |
4380 | if (IS_GEN6(dev)) |
4328 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); |
4381 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); |
4329 | else if (IS_GEN7(dev)) |
4382 | else if (IS_GEN7(dev)) |
4330 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); |
4383 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); |
4331 | else if (IS_GEN8(dev)) |
4384 | else if (IS_GEN8(dev)) |
4332 | I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW)); |
4385 | I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW)); |
4333 | else |
4386 | else |
4334 | BUG(); |
4387 | BUG(); |
4335 | } |
4388 | } |
4336 | 4389 | ||
4337 | static void init_unused_ring(struct drm_device *dev, u32 base) |
4390 | static void init_unused_ring(struct drm_device *dev, u32 base) |
4338 | { |
4391 | { |
4339 | struct drm_i915_private *dev_priv = dev->dev_private; |
4392 | struct drm_i915_private *dev_priv = dev->dev_private; |
4340 | 4393 | ||
4341 | I915_WRITE(RING_CTL(base), 0); |
4394 | I915_WRITE(RING_CTL(base), 0); |
4342 | I915_WRITE(RING_HEAD(base), 0); |
4395 | I915_WRITE(RING_HEAD(base), 0); |
4343 | I915_WRITE(RING_TAIL(base), 0); |
4396 | I915_WRITE(RING_TAIL(base), 0); |
4344 | I915_WRITE(RING_START(base), 0); |
4397 | I915_WRITE(RING_START(base), 0); |
4345 | } |
4398 | } |
4346 | 4399 | ||
4347 | static void init_unused_rings(struct drm_device *dev) |
4400 | static void init_unused_rings(struct drm_device *dev) |
4348 | { |
4401 | { |
4349 | if (IS_I830(dev)) { |
4402 | if (IS_I830(dev)) { |
4350 | init_unused_ring(dev, PRB1_BASE); |
4403 | init_unused_ring(dev, PRB1_BASE); |
4351 | init_unused_ring(dev, SRB0_BASE); |
4404 | init_unused_ring(dev, SRB0_BASE); |
4352 | init_unused_ring(dev, SRB1_BASE); |
4405 | init_unused_ring(dev, SRB1_BASE); |
4353 | init_unused_ring(dev, SRB2_BASE); |
4406 | init_unused_ring(dev, SRB2_BASE); |
4354 | init_unused_ring(dev, SRB3_BASE); |
4407 | init_unused_ring(dev, SRB3_BASE); |
4355 | } else if (IS_GEN2(dev)) { |
4408 | } else if (IS_GEN2(dev)) { |
4356 | init_unused_ring(dev, SRB0_BASE); |
4409 | init_unused_ring(dev, SRB0_BASE); |
4357 | init_unused_ring(dev, SRB1_BASE); |
4410 | init_unused_ring(dev, SRB1_BASE); |
4358 | } else if (IS_GEN3(dev)) { |
4411 | } else if (IS_GEN3(dev)) { |
4359 | init_unused_ring(dev, PRB1_BASE); |
4412 | init_unused_ring(dev, PRB1_BASE); |
4360 | init_unused_ring(dev, PRB2_BASE); |
4413 | init_unused_ring(dev, PRB2_BASE); |
4361 | } |
4414 | } |
4362 | } |
4415 | } |
4363 | 4416 | ||
4364 | int i915_gem_init_rings(struct drm_device *dev) |
4417 | int i915_gem_init_rings(struct drm_device *dev) |
4365 | { |
4418 | { |
4366 | struct drm_i915_private *dev_priv = dev->dev_private; |
4419 | struct drm_i915_private *dev_priv = dev->dev_private; |
4367 | int ret; |
4420 | int ret; |
4368 | 4421 | ||
4369 | ret = intel_init_render_ring_buffer(dev); |
4422 | ret = intel_init_render_ring_buffer(dev); |
4370 | if (ret) |
4423 | if (ret) |
4371 | return ret; |
4424 | return ret; |
4372 | 4425 | ||
4373 | if (HAS_BSD(dev)) { |
4426 | if (HAS_BSD(dev)) { |
4374 | ret = intel_init_bsd_ring_buffer(dev); |
4427 | ret = intel_init_bsd_ring_buffer(dev); |
4375 | if (ret) |
4428 | if (ret) |
4376 | goto cleanup_render_ring; |
4429 | goto cleanup_render_ring; |
4377 | } |
4430 | } |
4378 | 4431 | ||
4379 | if (HAS_BLT(dev)) { |
4432 | if (HAS_BLT(dev)) { |
4380 | ret = intel_init_blt_ring_buffer(dev); |
4433 | ret = intel_init_blt_ring_buffer(dev); |
4381 | if (ret) |
4434 | if (ret) |
4382 | goto cleanup_bsd_ring; |
4435 | goto cleanup_bsd_ring; |
4383 | } |
4436 | } |
4384 | 4437 | ||
4385 | if (HAS_VEBOX(dev)) { |
4438 | if (HAS_VEBOX(dev)) { |
4386 | ret = intel_init_vebox_ring_buffer(dev); |
4439 | ret = intel_init_vebox_ring_buffer(dev); |
4387 | if (ret) |
4440 | if (ret) |
4388 | goto cleanup_blt_ring; |
4441 | goto cleanup_blt_ring; |
4389 | } |
4442 | } |
4390 | 4443 | ||
4391 | if (HAS_BSD2(dev)) { |
4444 | if (HAS_BSD2(dev)) { |
4392 | ret = intel_init_bsd2_ring_buffer(dev); |
4445 | ret = intel_init_bsd2_ring_buffer(dev); |
4393 | if (ret) |
4446 | if (ret) |
4394 | goto cleanup_vebox_ring; |
4447 | goto cleanup_vebox_ring; |
4395 | } |
4448 | } |
4396 | 4449 | ||
4397 | return 0; |
4450 | return 0; |
4398 | 4451 | ||
4399 | cleanup_vebox_ring: |
4452 | cleanup_vebox_ring: |
4400 | intel_cleanup_ring_buffer(&dev_priv->ring[VECS]); |
4453 | intel_cleanup_ring_buffer(&dev_priv->ring[VECS]); |
4401 | cleanup_blt_ring: |
4454 | cleanup_blt_ring: |
4402 | intel_cleanup_ring_buffer(&dev_priv->ring[BCS]); |
4455 | intel_cleanup_ring_buffer(&dev_priv->ring[BCS]); |
4403 | cleanup_bsd_ring: |
4456 | cleanup_bsd_ring: |
4404 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); |
4457 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); |
4405 | cleanup_render_ring: |
4458 | cleanup_render_ring: |
4406 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); |
4459 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); |
4407 | 4460 | ||
4408 | return ret; |
4461 | return ret; |
4409 | } |
4462 | } |
4410 | 4463 | ||
4411 | int |
4464 | int |
4412 | i915_gem_init_hw(struct drm_device *dev) |
4465 | i915_gem_init_hw(struct drm_device *dev) |
4413 | { |
4466 | { |
4414 | struct drm_i915_private *dev_priv = dev->dev_private; |
4467 | struct drm_i915_private *dev_priv = dev->dev_private; |
4415 | struct intel_engine_cs *ring; |
4468 | struct intel_engine_cs *ring; |
4416 | int ret, i, j; |
4469 | int ret, i, j; |
4417 | 4470 | ||
4418 | if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt()) |
4471 | if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt()) |
4419 | return -EIO; |
4472 | return -EIO; |
4420 | 4473 | ||
4421 | /* Double layer security blanket, see i915_gem_init() */ |
4474 | /* Double layer security blanket, see i915_gem_init() */ |
4422 | intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL); |
4475 | intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL); |
4423 | 4476 | ||
4424 | if (dev_priv->ellc_size) |
4477 | if (dev_priv->ellc_size) |
4425 | I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf)); |
4478 | I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf)); |
4426 | 4479 | ||
4427 | if (IS_HASWELL(dev)) |
4480 | if (IS_HASWELL(dev)) |
4428 | I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ? |
4481 | I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ? |
4429 | LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED); |
4482 | LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED); |
4430 | 4483 | ||
4431 | if (HAS_PCH_NOP(dev)) { |
4484 | if (HAS_PCH_NOP(dev)) { |
4432 | if (IS_IVYBRIDGE(dev)) { |
4485 | if (IS_IVYBRIDGE(dev)) { |
4433 | u32 temp = I915_READ(GEN7_MSG_CTL); |
4486 | u32 temp = I915_READ(GEN7_MSG_CTL); |
4434 | temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK); |
4487 | temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK); |
4435 | I915_WRITE(GEN7_MSG_CTL, temp); |
4488 | I915_WRITE(GEN7_MSG_CTL, temp); |
4436 | } else if (INTEL_INFO(dev)->gen >= 7) { |
4489 | } else if (INTEL_INFO(dev)->gen >= 7) { |
4437 | u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT); |
4490 | u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT); |
4438 | temp &= ~RESET_PCH_HANDSHAKE_ENABLE; |
4491 | temp &= ~RESET_PCH_HANDSHAKE_ENABLE; |
4439 | I915_WRITE(HSW_NDE_RSTWRN_OPT, temp); |
4492 | I915_WRITE(HSW_NDE_RSTWRN_OPT, temp); |
4440 | } |
4493 | } |
4441 | } |
4494 | } |
4442 | 4495 | ||
4443 | i915_gem_init_swizzling(dev); |
4496 | i915_gem_init_swizzling(dev); |
4444 | 4497 | ||
4445 | /* |
4498 | /* |
4446 | * At least 830 can leave some of the unused rings |
4499 | * At least 830 can leave some of the unused rings |
4447 | * "active" (ie. head != tail) after resume which |
4500 | * "active" (ie. head != tail) after resume which |
4448 | * will prevent c3 entry. Makes sure all unused rings |
4501 | * will prevent c3 entry. Makes sure all unused rings |
4449 | * are totally idle. |
4502 | * are totally idle. |
4450 | */ |
4503 | */ |
4451 | init_unused_rings(dev); |
4504 | init_unused_rings(dev); |
4452 | 4505 | ||
4453 | BUG_ON(!dev_priv->ring[RCS].default_context); |
4506 | BUG_ON(!dev_priv->kernel_context); |
4454 | 4507 | ||
4455 | ret = i915_ppgtt_init_hw(dev); |
4508 | ret = i915_ppgtt_init_hw(dev); |
4456 | if (ret) { |
4509 | if (ret) { |
4457 | DRM_ERROR("PPGTT enable HW failed %d\n", ret); |
4510 | DRM_ERROR("PPGTT enable HW failed %d\n", ret); |
4458 | goto out; |
4511 | goto out; |
4459 | } |
4512 | } |
4460 | 4513 | ||
4461 | /* Need to do basic initialisation of all rings first: */ |
4514 | /* Need to do basic initialisation of all rings first: */ |
4462 | for_each_ring(ring, dev_priv, i) { |
4515 | for_each_ring(ring, dev_priv, i) { |
4463 | ret = ring->init_hw(ring); |
4516 | ret = ring->init_hw(ring); |
4464 | if (ret) |
4517 | if (ret) |
4465 | goto out; |
4518 | goto out; |
4466 | } |
4519 | } |
4467 | 4520 | ||
4468 | /* We can't enable contexts until all firmware is loaded */ |
4521 | /* We can't enable contexts until all firmware is loaded */ |
4469 | if (HAS_GUC_UCODE(dev)) { |
4522 | if (HAS_GUC_UCODE(dev)) { |
4470 | ret = intel_guc_ucode_load(dev); |
4523 | ret = intel_guc_ucode_load(dev); |
4471 | if (ret) { |
4524 | if (ret) { |
4472 | DRM_ERROR("Failed to initialize GuC, error %d\n", ret); |
4525 | DRM_ERROR("Failed to initialize GuC, error %d\n", ret); |
4473 | ret = -EIO; |
4526 | ret = -EIO; |
4474 | goto out; |
4527 | goto out; |
4475 | } |
4528 | } |
4476 | } |
4529 | } |
4477 | 4530 | ||
4478 | /* |
4531 | /* |
4479 | * Increment the next seqno by 0x100 so we have a visible break |
4532 | * Increment the next seqno by 0x100 so we have a visible break |
4480 | * on re-initialisation |
4533 | * on re-initialisation |
4481 | */ |
4534 | */ |
4482 | ret = i915_gem_set_seqno(dev, dev_priv->next_seqno+0x100); |
4535 | ret = i915_gem_set_seqno(dev, dev_priv->next_seqno+0x100); |
4483 | if (ret) |
4536 | if (ret) |
4484 | goto out; |
4537 | goto out; |
4485 | 4538 | ||
4486 | /* Now it is safe to go back round and do everything else: */ |
4539 | /* Now it is safe to go back round and do everything else: */ |
4487 | for_each_ring(ring, dev_priv, i) { |
4540 | for_each_ring(ring, dev_priv, i) { |
4488 | struct drm_i915_gem_request *req; |
4541 | struct drm_i915_gem_request *req; |
4489 | - | ||
4490 | WARN_ON(!ring->default_context); |
- | |
4491 | 4542 | ||
4492 | ret = i915_gem_request_alloc(ring, ring->default_context, &req); |
4543 | req = i915_gem_request_alloc(ring, NULL); |
- | 4544 | if (IS_ERR(req)) { |
|
4493 | if (ret) { |
4545 | ret = PTR_ERR(req); |
4494 | i915_gem_cleanup_ringbuffer(dev); |
4546 | i915_gem_cleanup_ringbuffer(dev); |
4495 | goto out; |
4547 | goto out; |
4496 | } |
4548 | } |
4497 | 4549 | ||
4498 | if (ring->id == RCS) { |
4550 | if (ring->id == RCS) { |
4499 | for (j = 0; j < NUM_L3_SLICES(dev); j++) |
4551 | for (j = 0; j < NUM_L3_SLICES(dev); j++) |
4500 | i915_gem_l3_remap(req, j); |
4552 | i915_gem_l3_remap(req, j); |
4501 | } |
4553 | } |
4502 | 4554 | ||
4503 | ret = i915_ppgtt_init_ring(req); |
4555 | ret = i915_ppgtt_init_ring(req); |
4504 | if (ret && ret != -EIO) { |
4556 | if (ret && ret != -EIO) { |
4505 | DRM_ERROR("PPGTT enable ring #%d failed %d\n", i, ret); |
4557 | DRM_ERROR("PPGTT enable ring #%d failed %d\n", i, ret); |
4506 | i915_gem_request_cancel(req); |
4558 | i915_gem_request_cancel(req); |
4507 | i915_gem_cleanup_ringbuffer(dev); |
4559 | i915_gem_cleanup_ringbuffer(dev); |
4508 | goto out; |
4560 | goto out; |
4509 | } |
4561 | } |
4510 | 4562 | ||
4511 | ret = i915_gem_context_enable(req); |
4563 | ret = i915_gem_context_enable(req); |
4512 | if (ret && ret != -EIO) { |
4564 | if (ret && ret != -EIO) { |
4513 | DRM_ERROR("Context enable ring #%d failed %d\n", i, ret); |
4565 | DRM_ERROR("Context enable ring #%d failed %d\n", i, ret); |
4514 | i915_gem_request_cancel(req); |
4566 | i915_gem_request_cancel(req); |
4515 | i915_gem_cleanup_ringbuffer(dev); |
4567 | i915_gem_cleanup_ringbuffer(dev); |
4516 | goto out; |
4568 | goto out; |
4517 | } |
4569 | } |
4518 | 4570 | ||
4519 | i915_add_request_no_flush(req); |
4571 | i915_add_request_no_flush(req); |
4520 | } |
4572 | } |
4521 | 4573 | ||
4522 | out: |
4574 | out: |
4523 | intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL); |
4575 | intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL); |
4524 | return ret; |
4576 | return ret; |
4525 | } |
4577 | } |
4526 | 4578 | ||
4527 | int i915_gem_init(struct drm_device *dev) |
4579 | int i915_gem_init(struct drm_device *dev) |
4528 | { |
4580 | { |
4529 | struct drm_i915_private *dev_priv = dev->dev_private; |
4581 | struct drm_i915_private *dev_priv = dev->dev_private; |
4530 | int ret; |
4582 | int ret; |
4531 | 4583 | ||
4532 | i915.enable_execlists = intel_sanitize_enable_execlists(dev, |
4584 | i915.enable_execlists = intel_sanitize_enable_execlists(dev, |
4533 | i915.enable_execlists); |
4585 | i915.enable_execlists); |
4534 | 4586 | ||
4535 | mutex_lock(&dev->struct_mutex); |
4587 | mutex_lock(&dev->struct_mutex); |
4536 | 4588 | ||
4537 | if (!i915.enable_execlists) { |
4589 | if (!i915.enable_execlists) { |
4538 | dev_priv->gt.execbuf_submit = i915_gem_ringbuffer_submission; |
4590 | dev_priv->gt.execbuf_submit = i915_gem_ringbuffer_submission; |
4539 | dev_priv->gt.init_rings = i915_gem_init_rings; |
4591 | dev_priv->gt.init_rings = i915_gem_init_rings; |
4540 | dev_priv->gt.cleanup_ring = intel_cleanup_ring_buffer; |
4592 | dev_priv->gt.cleanup_ring = intel_cleanup_ring_buffer; |
4541 | dev_priv->gt.stop_ring = intel_stop_ring_buffer; |
4593 | dev_priv->gt.stop_ring = intel_stop_ring_buffer; |
4542 | } else { |
4594 | } else { |
4543 | dev_priv->gt.execbuf_submit = intel_execlists_submission; |
4595 | dev_priv->gt.execbuf_submit = intel_execlists_submission; |
4544 | dev_priv->gt.init_rings = intel_logical_rings_init; |
4596 | dev_priv->gt.init_rings = intel_logical_rings_init; |
4545 | dev_priv->gt.cleanup_ring = intel_logical_ring_cleanup; |
4597 | dev_priv->gt.cleanup_ring = intel_logical_ring_cleanup; |
4546 | dev_priv->gt.stop_ring = intel_logical_ring_stop; |
4598 | dev_priv->gt.stop_ring = intel_logical_ring_stop; |
4547 | } |
4599 | } |
4548 | 4600 | ||
4549 | /* This is just a security blanket to placate dragons. |
4601 | /* This is just a security blanket to placate dragons. |
4550 | * On some systems, we very sporadically observe that the first TLBs |
4602 | * On some systems, we very sporadically observe that the first TLBs |
4551 | * used by the CS may be stale, despite us poking the TLB reset. If |
4603 | * used by the CS may be stale, despite us poking the TLB reset. If |
4552 | * we hold the forcewake during initialisation these problems |
4604 | * we hold the forcewake during initialisation these problems |
4553 | * just magically go away. |
4605 | * just magically go away. |
4554 | */ |
4606 | */ |
4555 | intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL); |
4607 | intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL); |
4556 | 4608 | ||
4557 | // ret = i915_gem_init_userptr(dev); |
4609 | // ret = i915_gem_init_userptr(dev); |
4558 | // if (ret) |
4610 | // if (ret) |
4559 | // goto out_unlock; |
4611 | // goto out_unlock; |
4560 | 4612 | ||
4561 | i915_gem_init_global_gtt(dev); |
4613 | i915_gem_init_global_gtt(dev); |
4562 | 4614 | ||
4563 | ret = i915_gem_context_init(dev); |
4615 | ret = i915_gem_context_init(dev); |
4564 | if (ret) |
4616 | if (ret) |
4565 | goto out_unlock; |
4617 | goto out_unlock; |
4566 | 4618 | ||
4567 | ret = dev_priv->gt.init_rings(dev); |
4619 | ret = dev_priv->gt.init_rings(dev); |
4568 | if (ret) |
4620 | if (ret) |
4569 | goto out_unlock; |
4621 | goto out_unlock; |
4570 | 4622 | ||
4571 | ret = i915_gem_init_hw(dev); |
4623 | ret = i915_gem_init_hw(dev); |
4572 | if (ret == -EIO) { |
4624 | if (ret == -EIO) { |
4573 | /* Allow ring initialisation to fail by marking the GPU as |
4625 | /* Allow ring initialisation to fail by marking the GPU as |
4574 | * wedged. But we only want to do this where the GPU is angry, |
4626 | * wedged. But we only want to do this where the GPU is angry, |
4575 | * for all other failure, such as an allocation failure, bail. |
4627 | * for all other failure, such as an allocation failure, bail. |
4576 | */ |
4628 | */ |
4577 | DRM_ERROR("Failed to initialize GPU, declaring it wedged\n"); |
4629 | DRM_ERROR("Failed to initialize GPU, declaring it wedged\n"); |
4578 | atomic_or(I915_WEDGED, &dev_priv->gpu_error.reset_counter); |
4630 | atomic_or(I915_WEDGED, &dev_priv->gpu_error.reset_counter); |
4579 | ret = 0; |
4631 | ret = 0; |
4580 | } |
4632 | } |
4581 | 4633 | ||
4582 | out_unlock: |
4634 | out_unlock: |
4583 | intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL); |
4635 | intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL); |
4584 | mutex_unlock(&dev->struct_mutex); |
4636 | mutex_unlock(&dev->struct_mutex); |
4585 | 4637 | ||
4586 | return ret; |
4638 | return ret; |
4587 | } |
4639 | } |
4588 | 4640 | ||
4589 | void |
4641 | void |
4590 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) |
4642 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) |
4591 | { |
4643 | { |
4592 | struct drm_i915_private *dev_priv = dev->dev_private; |
4644 | struct drm_i915_private *dev_priv = dev->dev_private; |
4593 | struct intel_engine_cs *ring; |
4645 | struct intel_engine_cs *ring; |
4594 | int i; |
4646 | int i; |
4595 | 4647 | ||
4596 | for_each_ring(ring, dev_priv, i) |
4648 | for_each_ring(ring, dev_priv, i) |
4597 | dev_priv->gt.cleanup_ring(ring); |
4649 | dev_priv->gt.cleanup_ring(ring); |
- | 4650 | ||
- | 4651 | if (i915.enable_execlists) |
|
- | 4652 | /* |
|
- | 4653 | * Neither the BIOS, ourselves or any other kernel |
|
- | 4654 | * expects the system to be in execlists mode on startup, |
|
- | 4655 | * so we need to reset the GPU back to legacy mode. |
|
- | 4656 | */ |
|
- | 4657 | intel_gpu_reset(dev); |
|
4598 | } |
4658 | } |
4599 | 4659 | ||
4600 | static void |
4660 | static void |
4601 | init_ring_lists(struct intel_engine_cs *ring) |
4661 | init_ring_lists(struct intel_engine_cs *ring) |
4602 | { |
4662 | { |
4603 | INIT_LIST_HEAD(&ring->active_list); |
4663 | INIT_LIST_HEAD(&ring->active_list); |
4604 | INIT_LIST_HEAD(&ring->request_list); |
4664 | INIT_LIST_HEAD(&ring->request_list); |
4605 | } |
4665 | } |
4606 | 4666 | ||
4607 | void |
4667 | void |
4608 | i915_gem_load(struct drm_device *dev) |
4668 | i915_gem_load_init(struct drm_device *dev) |
4609 | { |
4669 | { |
4610 | struct drm_i915_private *dev_priv = dev->dev_private; |
4670 | struct drm_i915_private *dev_priv = dev->dev_private; |
4611 | int i; |
4671 | int i; |
4612 | 4672 | ||
4613 | INIT_LIST_HEAD(&dev_priv->vm_list); |
4673 | INIT_LIST_HEAD(&dev_priv->vm_list); |
4614 | INIT_LIST_HEAD(&dev_priv->context_list); |
4674 | INIT_LIST_HEAD(&dev_priv->context_list); |
4615 | INIT_LIST_HEAD(&dev_priv->mm.unbound_list); |
4675 | INIT_LIST_HEAD(&dev_priv->mm.unbound_list); |
4616 | INIT_LIST_HEAD(&dev_priv->mm.bound_list); |
4676 | INIT_LIST_HEAD(&dev_priv->mm.bound_list); |
4617 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
4677 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
4618 | for (i = 0; i < I915_NUM_RINGS; i++) |
4678 | for (i = 0; i < I915_NUM_RINGS; i++) |
4619 | init_ring_lists(&dev_priv->ring[i]); |
4679 | init_ring_lists(&dev_priv->ring[i]); |
4620 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) |
4680 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) |
4621 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); |
4681 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); |
4622 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, |
4682 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, |
4623 | i915_gem_retire_work_handler); |
4683 | i915_gem_retire_work_handler); |
4624 | INIT_DELAYED_WORK(&dev_priv->mm.idle_work, |
4684 | INIT_DELAYED_WORK(&dev_priv->mm.idle_work, |
4625 | i915_gem_idle_work_handler); |
4685 | i915_gem_idle_work_handler); |
4626 | init_waitqueue_head(&dev_priv->gpu_error.reset_queue); |
4686 | init_waitqueue_head(&dev_priv->gpu_error.reset_queue); |
4627 | 4687 | ||
4628 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; |
4688 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; |
4629 | 4689 | ||
4630 | if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev)) |
4690 | if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev)) |
4631 | dev_priv->num_fence_regs = 32; |
4691 | dev_priv->num_fence_regs = 32; |
4632 | else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) |
4692 | else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) |
4633 | dev_priv->num_fence_regs = 16; |
4693 | dev_priv->num_fence_regs = 16; |
4634 | else |
4694 | else |
4635 | dev_priv->num_fence_regs = 8; |
4695 | dev_priv->num_fence_regs = 8; |
4636 | 4696 | ||
4637 | if (intel_vgpu_active(dev)) |
4697 | if (intel_vgpu_active(dev)) |
4638 | dev_priv->num_fence_regs = |
4698 | dev_priv->num_fence_regs = |
4639 | I915_READ(vgtif_reg(avail_rs.fence_num)); |
4699 | I915_READ(vgtif_reg(avail_rs.fence_num)); |
4640 | 4700 | ||
4641 | /* |
4701 | /* |
4642 | * Set initial sequence number for requests. |
4702 | * Set initial sequence number for requests. |
4643 | * Using this number allows the wraparound to happen early, |
4703 | * Using this number allows the wraparound to happen early, |
4644 | * catching any obvious problems. |
4704 | * catching any obvious problems. |
4645 | */ |
4705 | */ |
4646 | dev_priv->next_seqno = ((u32)~0 - 0x1100); |
4706 | dev_priv->next_seqno = ((u32)~0 - 0x1100); |
4647 | dev_priv->last_seqno = ((u32)~0 - 0x1101); |
4707 | dev_priv->last_seqno = ((u32)~0 - 0x1101); |
4648 | 4708 | ||
4649 | /* Initialize fence registers to zero */ |
4709 | /* Initialize fence registers to zero */ |
4650 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
4710 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
4651 | i915_gem_restore_fences(dev); |
4711 | i915_gem_restore_fences(dev); |
4652 | 4712 | ||
4653 | i915_gem_detect_bit_6_swizzle(dev); |
4713 | i915_gem_detect_bit_6_swizzle(dev); |
- | 4714 | init_waitqueue_head(&dev_priv->pending_flip_queue); |
|
4654 | 4715 | ||
4655 | dev_priv->mm.interruptible = true; |
4716 | dev_priv->mm.interruptible = true; |
4656 | 4717 | ||
4657 | mutex_init(&dev_priv->fb_tracking.lock); |
4718 | mutex_init(&dev_priv->fb_tracking.lock); |
4658 | } |
4719 | } |
4659 | 4720 | ||
4660 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
4721 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
4661 | { |
4722 | { |
4662 | struct drm_i915_file_private *file_priv = file->driver_priv; |
4723 | struct drm_i915_file_private *file_priv = file->driver_priv; |
4663 | 4724 | ||
4664 | /* Clean up our request list when the client is going away, so that |
4725 | /* Clean up our request list when the client is going away, so that |
4665 | * later retire_requests won't dereference our soon-to-be-gone |
4726 | * later retire_requests won't dereference our soon-to-be-gone |
4666 | * file_priv. |
4727 | * file_priv. |
4667 | */ |
4728 | */ |
4668 | spin_lock(&file_priv->mm.lock); |
4729 | spin_lock(&file_priv->mm.lock); |
4669 | while (!list_empty(&file_priv->mm.request_list)) { |
4730 | while (!list_empty(&file_priv->mm.request_list)) { |
4670 | struct drm_i915_gem_request *request; |
4731 | struct drm_i915_gem_request *request; |
4671 | 4732 | ||
4672 | request = list_first_entry(&file_priv->mm.request_list, |
4733 | request = list_first_entry(&file_priv->mm.request_list, |
4673 | struct drm_i915_gem_request, |
4734 | struct drm_i915_gem_request, |
4674 | client_list); |
4735 | client_list); |
4675 | list_del(&request->client_list); |
4736 | list_del(&request->client_list); |
4676 | request->file_priv = NULL; |
4737 | request->file_priv = NULL; |
4677 | } |
4738 | } |
4678 | spin_unlock(&file_priv->mm.lock); |
4739 | spin_unlock(&file_priv->mm.lock); |
4679 | 4740 | ||
4680 | if (!list_empty(&file_priv->rps.link)) { |
4741 | if (!list_empty(&file_priv->rps.link)) { |
4681 | spin_lock(&to_i915(dev)->rps.client_lock); |
4742 | spin_lock(&to_i915(dev)->rps.client_lock); |
4682 | list_del(&file_priv->rps.link); |
4743 | list_del(&file_priv->rps.link); |
4683 | spin_unlock(&to_i915(dev)->rps.client_lock); |
4744 | spin_unlock(&to_i915(dev)->rps.client_lock); |
4684 | } |
4745 | } |
4685 | } |
4746 | } |
4686 | 4747 | ||
4687 | int i915_gem_open(struct drm_device *dev, struct drm_file *file) |
4748 | int i915_gem_open(struct drm_device *dev, struct drm_file *file) |
4688 | { |
4749 | { |
4689 | struct drm_i915_file_private *file_priv; |
4750 | struct drm_i915_file_private *file_priv; |
4690 | int ret; |
4751 | int ret; |
4691 | 4752 | ||
4692 | DRM_DEBUG_DRIVER("\n"); |
4753 | DRM_DEBUG_DRIVER("\n"); |
4693 | 4754 | ||
4694 | file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL); |
4755 | file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL); |
4695 | if (!file_priv) |
4756 | if (!file_priv) |
4696 | return -ENOMEM; |
4757 | return -ENOMEM; |
4697 | 4758 | ||
4698 | file->driver_priv = file_priv; |
4759 | file->driver_priv = file_priv; |
4699 | file_priv->dev_priv = dev->dev_private; |
4760 | file_priv->dev_priv = dev->dev_private; |
4700 | file_priv->file = file; |
4761 | file_priv->file = file; |
4701 | INIT_LIST_HEAD(&file_priv->rps.link); |
4762 | INIT_LIST_HEAD(&file_priv->rps.link); |
4702 | 4763 | ||
4703 | spin_lock_init(&file_priv->mm.lock); |
4764 | spin_lock_init(&file_priv->mm.lock); |
4704 | INIT_LIST_HEAD(&file_priv->mm.request_list); |
4765 | INIT_LIST_HEAD(&file_priv->mm.request_list); |
- | 4766 | ||
- | 4767 | file_priv->bsd_ring = -1; |
|
4705 | 4768 | ||
4706 | ret = i915_gem_context_open(dev, file); |
4769 | ret = i915_gem_context_open(dev, file); |
4707 | if (ret) |
4770 | if (ret) |
4708 | kfree(file_priv); |
4771 | kfree(file_priv); |
4709 | 4772 | ||
4710 | return ret; |
4773 | return ret; |
4711 | } |
4774 | } |
4712 | 4775 | ||
4713 | /** |
4776 | /** |
4714 | * i915_gem_track_fb - update frontbuffer tracking |
4777 | * i915_gem_track_fb - update frontbuffer tracking |
4715 | * @old: current GEM buffer for the frontbuffer slots |
4778 | * @old: current GEM buffer for the frontbuffer slots |
4716 | * @new: new GEM buffer for the frontbuffer slots |
4779 | * @new: new GEM buffer for the frontbuffer slots |
4717 | * @frontbuffer_bits: bitmask of frontbuffer slots |
4780 | * @frontbuffer_bits: bitmask of frontbuffer slots |
4718 | * |
4781 | * |
4719 | * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them |
4782 | * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them |
4720 | * from @old and setting them in @new. Both @old and @new can be NULL. |
4783 | * from @old and setting them in @new. Both @old and @new can be NULL. |
4721 | */ |
4784 | */ |
4722 | void i915_gem_track_fb(struct drm_i915_gem_object *old, |
4785 | void i915_gem_track_fb(struct drm_i915_gem_object *old, |
4723 | struct drm_i915_gem_object *new, |
4786 | struct drm_i915_gem_object *new, |
4724 | unsigned frontbuffer_bits) |
4787 | unsigned frontbuffer_bits) |
4725 | { |
4788 | { |
4726 | if (old) { |
4789 | if (old) { |
4727 | WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex)); |
4790 | WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex)); |
4728 | WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits)); |
4791 | WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits)); |
4729 | old->frontbuffer_bits &= ~frontbuffer_bits; |
4792 | old->frontbuffer_bits &= ~frontbuffer_bits; |
4730 | } |
4793 | } |
4731 | 4794 | ||
4732 | if (new) { |
4795 | if (new) { |
4733 | WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex)); |
4796 | WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex)); |
4734 | WARN_ON(new->frontbuffer_bits & frontbuffer_bits); |
4797 | WARN_ON(new->frontbuffer_bits & frontbuffer_bits); |
4735 | new->frontbuffer_bits |= frontbuffer_bits; |
4798 | new->frontbuffer_bits |= frontbuffer_bits; |
4736 | } |
4799 | } |
4737 | } |
4800 | } |
4738 | 4801 | ||
4739 | /* All the new VM stuff */ |
4802 | /* All the new VM stuff */ |
4740 | u64 i915_gem_obj_offset(struct drm_i915_gem_object *o, |
4803 | u64 i915_gem_obj_offset(struct drm_i915_gem_object *o, |
4741 | struct i915_address_space *vm) |
4804 | struct i915_address_space *vm) |
4742 | { |
4805 | { |
4743 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; |
4806 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; |
4744 | struct i915_vma *vma; |
4807 | struct i915_vma *vma; |
4745 | 4808 | ||
4746 | WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base); |
4809 | WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base); |
4747 | 4810 | ||
4748 | list_for_each_entry(vma, &o->vma_list, vma_link) { |
4811 | list_for_each_entry(vma, &o->vma_list, obj_link) { |
4749 | if (i915_is_ggtt(vma->vm) && |
4812 | if (vma->is_ggtt && |
4750 | vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL) |
4813 | vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL) |
4751 | continue; |
4814 | continue; |
4752 | if (vma->vm == vm) |
4815 | if (vma->vm == vm) |
4753 | return vma->node.start; |
4816 | return vma->node.start; |
4754 | } |
4817 | } |
4755 | 4818 | ||
4756 | WARN(1, "%s vma for this object not found.\n", |
4819 | WARN(1, "%s vma for this object not found.\n", |
4757 | i915_is_ggtt(vm) ? "global" : "ppgtt"); |
4820 | i915_is_ggtt(vm) ? "global" : "ppgtt"); |
4758 | return -1; |
4821 | return -1; |
4759 | } |
4822 | } |
4760 | 4823 | ||
4761 | u64 i915_gem_obj_ggtt_offset_view(struct drm_i915_gem_object *o, |
4824 | u64 i915_gem_obj_ggtt_offset_view(struct drm_i915_gem_object *o, |
4762 | const struct i915_ggtt_view *view) |
4825 | const struct i915_ggtt_view *view) |
4763 | { |
4826 | { |
4764 | struct i915_address_space *ggtt = i915_obj_to_ggtt(o); |
4827 | struct i915_address_space *ggtt = i915_obj_to_ggtt(o); |
4765 | struct i915_vma *vma; |
4828 | struct i915_vma *vma; |
4766 | 4829 | ||
4767 | list_for_each_entry(vma, &o->vma_list, vma_link) |
4830 | list_for_each_entry(vma, &o->vma_list, obj_link) |
4768 | if (vma->vm == ggtt && |
4831 | if (vma->vm == ggtt && |
4769 | i915_ggtt_view_equal(&vma->ggtt_view, view)) |
4832 | i915_ggtt_view_equal(&vma->ggtt_view, view)) |
4770 | return vma->node.start; |
4833 | return vma->node.start; |
4771 | 4834 | ||
4772 | WARN(1, "global vma for this object not found. (view=%u)\n", view->type); |
4835 | WARN(1, "global vma for this object not found. (view=%u)\n", view->type); |
4773 | return -1; |
4836 | return -1; |
4774 | } |
4837 | } |
4775 | 4838 | ||
4776 | bool i915_gem_obj_bound(struct drm_i915_gem_object *o, |
4839 | bool i915_gem_obj_bound(struct drm_i915_gem_object *o, |
4777 | struct i915_address_space *vm) |
4840 | struct i915_address_space *vm) |
4778 | { |
4841 | { |
4779 | struct i915_vma *vma; |
4842 | struct i915_vma *vma; |
4780 | 4843 | ||
4781 | list_for_each_entry(vma, &o->vma_list, vma_link) { |
4844 | list_for_each_entry(vma, &o->vma_list, obj_link) { |
4782 | if (i915_is_ggtt(vma->vm) && |
4845 | if (vma->is_ggtt && |
4783 | vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL) |
4846 | vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL) |
4784 | continue; |
4847 | continue; |
4785 | if (vma->vm == vm && drm_mm_node_allocated(&vma->node)) |
4848 | if (vma->vm == vm && drm_mm_node_allocated(&vma->node)) |
4786 | return true; |
4849 | return true; |
4787 | } |
4850 | } |
4788 | 4851 | ||
4789 | return false; |
4852 | return false; |
4790 | } |
4853 | } |
4791 | 4854 | ||
4792 | bool i915_gem_obj_ggtt_bound_view(struct drm_i915_gem_object *o, |
4855 | bool i915_gem_obj_ggtt_bound_view(struct drm_i915_gem_object *o, |
4793 | const struct i915_ggtt_view *view) |
4856 | const struct i915_ggtt_view *view) |
4794 | { |
4857 | { |
4795 | struct i915_address_space *ggtt = i915_obj_to_ggtt(o); |
4858 | struct i915_address_space *ggtt = i915_obj_to_ggtt(o); |
4796 | struct i915_vma *vma; |
4859 | struct i915_vma *vma; |
4797 | 4860 | ||
4798 | list_for_each_entry(vma, &o->vma_list, vma_link) |
4861 | list_for_each_entry(vma, &o->vma_list, obj_link) |
4799 | if (vma->vm == ggtt && |
4862 | if (vma->vm == ggtt && |
4800 | i915_ggtt_view_equal(&vma->ggtt_view, view) && |
4863 | i915_ggtt_view_equal(&vma->ggtt_view, view) && |
4801 | drm_mm_node_allocated(&vma->node)) |
4864 | drm_mm_node_allocated(&vma->node)) |
4802 | return true; |
4865 | return true; |
4803 | 4866 | ||
4804 | return false; |
4867 | return false; |
4805 | } |
4868 | } |
4806 | 4869 | ||
4807 | bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o) |
4870 | bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o) |
4808 | { |
4871 | { |
4809 | struct i915_vma *vma; |
4872 | struct i915_vma *vma; |
4810 | 4873 | ||
4811 | list_for_each_entry(vma, &o->vma_list, vma_link) |
4874 | list_for_each_entry(vma, &o->vma_list, obj_link) |
4812 | if (drm_mm_node_allocated(&vma->node)) |
4875 | if (drm_mm_node_allocated(&vma->node)) |
4813 | return true; |
4876 | return true; |
4814 | 4877 | ||
4815 | return false; |
4878 | return false; |
4816 | } |
4879 | } |
4817 | 4880 | ||
4818 | unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o, |
4881 | unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o, |
4819 | struct i915_address_space *vm) |
4882 | struct i915_address_space *vm) |
4820 | { |
4883 | { |
4821 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; |
4884 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; |
4822 | struct i915_vma *vma; |
4885 | struct i915_vma *vma; |
4823 | 4886 | ||
4824 | WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base); |
4887 | WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base); |
4825 | 4888 | ||
4826 | BUG_ON(list_empty(&o->vma_list)); |
4889 | BUG_ON(list_empty(&o->vma_list)); |
4827 | 4890 | ||
4828 | list_for_each_entry(vma, &o->vma_list, vma_link) { |
4891 | list_for_each_entry(vma, &o->vma_list, obj_link) { |
4829 | if (i915_is_ggtt(vma->vm) && |
4892 | if (vma->is_ggtt && |
4830 | vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL) |
4893 | vma->ggtt_view.type != I915_GGTT_VIEW_NORMAL) |
4831 | continue; |
4894 | continue; |
4832 | if (vma->vm == vm) |
4895 | if (vma->vm == vm) |
4833 | return vma->node.size; |
4896 | return vma->node.size; |
4834 | } |
4897 | } |
4835 | return 0; |
4898 | return 0; |
4836 | } |
4899 | } |
4837 | 4900 | ||
4838 | bool i915_gem_obj_is_pinned(struct drm_i915_gem_object *obj) |
4901 | bool i915_gem_obj_is_pinned(struct drm_i915_gem_object *obj) |
4839 | { |
4902 | { |
4840 | struct i915_vma *vma; |
4903 | struct i915_vma *vma; |
4841 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
4904 | list_for_each_entry(vma, &obj->vma_list, obj_link) |
4842 | if (vma->pin_count > 0) |
4905 | if (vma->pin_count > 0) |
4843 | return true; |
4906 | return true; |
4844 | 4907 | ||
4845 | return false; |
4908 | return false; |
4846 | } |
4909 | } |
4847 | 4910 | ||
4848 | /* Like i915_gem_object_get_page(), but mark the returned page dirty */ |
4911 | /* Like i915_gem_object_get_page(), but mark the returned page dirty */ |
4849 | struct page * |
4912 | struct page * |
4850 | i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, int n) |
4913 | i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, int n) |
4851 | { |
4914 | { |
4852 | struct page *page; |
4915 | struct page *page; |
4853 | 4916 | ||
4854 | /* Only default objects have per-page dirty tracking */ |
4917 | /* Only default objects have per-page dirty tracking */ |
4855 | if (WARN_ON((obj->ops->flags & I915_GEM_OBJECT_HAS_STRUCT_PAGE) == 0)) |
4918 | if (WARN_ON((obj->ops->flags & I915_GEM_OBJECT_HAS_STRUCT_PAGE) == 0)) |
4856 | return NULL; |
4919 | return NULL; |
4857 | 4920 | ||
4858 | page = i915_gem_object_get_page(obj, n); |
4921 | page = i915_gem_object_get_page(obj, n); |
4859 | set_page_dirty(page); |
4922 | set_page_dirty(page); |
4860 | return page; |
4923 | return page; |
4861 | } |
4924 | } |
4862 | 4925 | ||
4863 | /* Allocate a new GEM object and fill it with the supplied data */ |
4926 | /* Allocate a new GEM object and fill it with the supplied data */ |
4864 | struct drm_i915_gem_object * |
4927 | struct drm_i915_gem_object * |
4865 | i915_gem_object_create_from_data(struct drm_device *dev, |
4928 | i915_gem_object_create_from_data(struct drm_device *dev, |
4866 | const void *data, size_t size) |
4929 | const void *data, size_t size) |
4867 | { |
4930 | { |
4868 | struct drm_i915_gem_object *obj; |
4931 | struct drm_i915_gem_object *obj; |
4869 | struct sg_table *sg; |
4932 | struct sg_table *sg; |
4870 | size_t bytes; |
4933 | size_t bytes; |
4871 | int ret; |
4934 | int ret; |
4872 | 4935 | ||
4873 | obj = i915_gem_alloc_object(dev, round_up(size, PAGE_SIZE)); |
4936 | obj = i915_gem_alloc_object(dev, round_up(size, PAGE_SIZE)); |
4874 | if (IS_ERR_OR_NULL(obj)) |
4937 | if (IS_ERR_OR_NULL(obj)) |
4875 | return obj; |
4938 | return obj; |
4876 | 4939 | ||
4877 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
4940 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
4878 | if (ret) |
4941 | if (ret) |
4879 | goto fail; |
4942 | goto fail; |
4880 | 4943 | ||
4881 | ret = i915_gem_object_get_pages(obj); |
4944 | ret = i915_gem_object_get_pages(obj); |
4882 | if (ret) |
4945 | if (ret) |
4883 | goto fail; |
4946 | goto fail; |
4884 | 4947 | ||
4885 | i915_gem_object_pin_pages(obj); |
4948 | i915_gem_object_pin_pages(obj); |
4886 | sg = obj->pages; |
4949 | sg = obj->pages; |
4887 | bytes = sg_copy_from_buffer(sg->sgl, sg->nents, (void *)data, size); |
4950 | bytes = sg_copy_from_buffer(sg->sgl, sg->nents, (void *)data, size); |
4888 | obj->dirty = 1; /* Backing store is now out of date */ |
4951 | obj->dirty = 1; /* Backing store is now out of date */ |
4889 | i915_gem_object_unpin_pages(obj); |
4952 | i915_gem_object_unpin_pages(obj); |
4890 | 4953 | ||
4891 | if (WARN_ON(bytes != size)) { |
4954 | if (WARN_ON(bytes != size)) { |
4892 | DRM_ERROR("Incomplete copy, wrote %zu of %zu", bytes, size); |
4955 | DRM_ERROR("Incomplete copy, wrote %zu of %zu", bytes, size); |
4893 | ret = -EFAULT; |
4956 | ret = -EFAULT; |
4894 | goto fail; |
4957 | goto fail; |
4895 | } |
4958 | } |
4896 | 4959 | ||
4897 | return obj; |
4960 | return obj; |
4898 | 4961 | ||
4899 | fail: |
4962 | fail: |
4900 | drm_gem_object_unreference(&obj->base); |
4963 | drm_gem_object_unreference(&obj->base); |
4901 | return ERR_PTR(ret); |
4964 | return ERR_PTR(ret); |
4902 | }>>>>>>>><>= |
4965 | }>>>>>>>><>>= |
4903 | >>><>>>=>>>>=>>><>><>>>=><=>>>>>>>->->>>><>=>><>=>> |
4966 | >>><>>>=>>>>=>>><>><>>>=><=>>>>>>>->->>>><>=>><>=>> |