Rev 4246 | Rev 4293 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
2326 | Serge | 1 | /* |
2 | * Copyright © 2008 Intel Corporation |
||
3 | * |
||
4 | * Permission is hereby granted, free of charge, to any person obtaining a |
||
5 | * copy of this software and associated documentation files (the "Software"), |
||
6 | * to deal in the Software without restriction, including without limitation |
||
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 |
||
9 | * Software is furnished to do so, subject to the following conditions: |
||
10 | * |
||
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 |
||
13 | * Software. |
||
14 | * |
||
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, |
||
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 |
||
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 |
||
21 | * IN THE SOFTWARE. |
||
22 | * |
||
23 | * Authors: |
||
24 | * Eric Anholt |
||
25 | * |
||
26 | */ |
||
27 | |||
3031 | serge | 28 | #include |
4280 | Serge | 29 | #include |
3031 | serge | 30 | #include |
2326 | Serge | 31 | #include "i915_drv.h" |
2351 | Serge | 32 | #include "i915_trace.h" |
2326 | Serge | 33 | #include "intel_drv.h" |
3260 | Serge | 34 | #include |
2330 | Serge | 35 | #include |
2326 | Serge | 36 | //#include |
3746 | Serge | 37 | #include |
2326 | Serge | 38 | #include |
39 | |||
2344 | Serge | 40 | extern int x86_clflush_size; |
2332 | Serge | 41 | |
3263 | Serge | 42 | #define PROT_READ 0x1 /* page can be read */ |
43 | #define PROT_WRITE 0x2 /* page can be written */ |
||
44 | #define MAP_SHARED 0x01 /* Share changes */ |
||
45 | |||
2344 | Serge | 46 | #undef mb |
47 | #undef rmb |
||
48 | #undef wmb |
||
49 | #define mb() asm volatile("mfence") |
||
50 | #define rmb() asm volatile ("lfence") |
||
51 | #define wmb() asm volatile ("sfence") |
||
52 | |||
3266 | Serge | 53 | struct drm_i915_gem_object *get_fb_obj(); |
54 | |||
3263 | Serge | 55 | unsigned long vm_mmap(struct file *file, unsigned long addr, |
56 | unsigned long len, unsigned long prot, |
||
57 | unsigned long flag, unsigned long offset); |
||
58 | |||
2344 | Serge | 59 | static inline void clflush(volatile void *__p) |
60 | { |
||
61 | asm volatile("clflush %0" : "+m" (*(volatile char*)__p)); |
||
62 | } |
||
63 | |||
2332 | Serge | 64 | #define MAX_ERRNO 4095 |
65 | |||
66 | #define IS_ERR_VALUE(x) unlikely((x) >= (unsigned long)-MAX_ERRNO) |
||
67 | |||
68 | |||
2326 | Serge | 69 | #define I915_EXEC_CONSTANTS_MASK (3<<6) |
70 | #define I915_EXEC_CONSTANTS_REL_GENERAL (0<<6) /* default */ |
||
71 | #define I915_EXEC_CONSTANTS_ABSOLUTE (1<<6) |
||
72 | #define I915_EXEC_CONSTANTS_REL_SURFACE (2<<6) /* gen4/5 only */ |
||
73 | |||
2332 | Serge | 74 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); |
4104 | Serge | 75 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj, |
76 | bool force); |
||
77 | static __must_check int |
||
78 | i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, |
||
79 | struct i915_address_space *vm, |
||
2332 | Serge | 80 | unsigned alignment, |
3031 | serge | 81 | bool map_and_fenceable, |
82 | bool nonblocking); |
||
2332 | Serge | 83 | static int i915_gem_phys_pwrite(struct drm_device *dev, |
84 | struct drm_i915_gem_object *obj, |
||
85 | struct drm_i915_gem_pwrite *args, |
||
86 | struct drm_file *file); |
||
2326 | Serge | 87 | |
3031 | serge | 88 | static void i915_gem_write_fence(struct drm_device *dev, int reg, |
89 | struct drm_i915_gem_object *obj); |
||
90 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, |
||
91 | struct drm_i915_fence_reg *fence, |
||
92 | bool enable); |
||
2332 | Serge | 93 | |
3031 | serge | 94 | static long i915_gem_purge(struct drm_i915_private *dev_priv, long target); |
4104 | Serge | 95 | static long i915_gem_shrink_all(struct drm_i915_private *dev_priv); |
3031 | serge | 96 | static void i915_gem_object_truncate(struct drm_i915_gem_object *obj); |
97 | |||
4104 | Serge | 98 | static bool cpu_cache_is_coherent(struct drm_device *dev, |
99 | enum i915_cache_level level) |
||
100 | { |
||
101 | return HAS_LLC(dev) || level != I915_CACHE_NONE; |
||
102 | } |
||
103 | |||
104 | static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) |
||
105 | { |
||
106 | if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) |
||
107 | return true; |
||
108 | |||
109 | return obj->pin_display; |
||
110 | } |
||
111 | |||
3031 | serge | 112 | static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj) |
113 | { |
||
114 | if (obj->tiling_mode) |
||
115 | i915_gem_release_mmap(obj); |
||
116 | |||
117 | /* As we do not have an associated fence register, we will force |
||
118 | * a tiling change if we ever need to acquire one. |
||
119 | */ |
||
120 | obj->fence_dirty = false; |
||
121 | obj->fence_reg = I915_FENCE_REG_NONE; |
||
122 | } |
||
123 | |||
2332 | Serge | 124 | /* some bookkeeping */ |
125 | static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv, |
||
126 | size_t size) |
||
127 | { |
||
4104 | Serge | 128 | spin_lock(&dev_priv->mm.object_stat_lock); |
2332 | Serge | 129 | dev_priv->mm.object_count++; |
130 | dev_priv->mm.object_memory += size; |
||
4104 | Serge | 131 | spin_unlock(&dev_priv->mm.object_stat_lock); |
2332 | Serge | 132 | } |
133 | |||
134 | static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv, |
||
135 | size_t size) |
||
136 | { |
||
4104 | Serge | 137 | spin_lock(&dev_priv->mm.object_stat_lock); |
2332 | Serge | 138 | dev_priv->mm.object_count--; |
139 | dev_priv->mm.object_memory -= size; |
||
4104 | Serge | 140 | spin_unlock(&dev_priv->mm.object_stat_lock); |
2332 | Serge | 141 | } |
142 | |||
143 | static int |
||
3480 | Serge | 144 | i915_gem_wait_for_error(struct i915_gpu_error *error) |
2332 | Serge | 145 | { |
146 | int ret; |
||
147 | |||
3480 | Serge | 148 | #define EXIT_COND (!i915_reset_in_progress(error)) |
149 | if (EXIT_COND) |
||
2332 | Serge | 150 | return 0; |
3255 | Serge | 151 | #if 0 |
3031 | serge | 152 | /* |
153 | * Only wait 10 seconds for the gpu reset to complete to avoid hanging |
||
154 | * userspace. If it takes that long something really bad is going on and |
||
155 | * we should simply try to bail out and fail as gracefully as possible. |
||
156 | */ |
||
3480 | Serge | 157 | ret = wait_event_interruptible_timeout(error->reset_queue, |
158 | EXIT_COND, |
||
159 | 10*HZ); |
||
3031 | serge | 160 | if (ret == 0) { |
161 | DRM_ERROR("Timed out waiting for the gpu reset to complete\n"); |
||
162 | return -EIO; |
||
163 | } else if (ret < 0) { |
||
2332 | Serge | 164 | return ret; |
3031 | serge | 165 | } |
2332 | Serge | 166 | |
3255 | Serge | 167 | #endif |
3480 | Serge | 168 | #undef EXIT_COND |
3255 | Serge | 169 | |
2332 | Serge | 170 | return 0; |
171 | } |
||
172 | |||
173 | int i915_mutex_lock_interruptible(struct drm_device *dev) |
||
174 | { |
||
3480 | Serge | 175 | struct drm_i915_private *dev_priv = dev->dev_private; |
2332 | Serge | 176 | int ret; |
177 | |||
3480 | Serge | 178 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
2332 | Serge | 179 | if (ret) |
180 | return ret; |
||
181 | |||
3480 | Serge | 182 | ret = mutex_lock_interruptible(&dev->struct_mutex); |
183 | if (ret) |
||
184 | return ret; |
||
2332 | Serge | 185 | |
186 | WARN_ON(i915_verify_lists(dev)); |
||
187 | return 0; |
||
188 | } |
||
189 | |||
190 | static inline bool |
||
191 | i915_gem_object_is_inactive(struct drm_i915_gem_object *obj) |
||
192 | { |
||
4104 | Serge | 193 | return i915_gem_obj_bound_any(obj) && !obj->active; |
2332 | Serge | 194 | } |
195 | |||
196 | |||
197 | #if 0 |
||
198 | |||
199 | int |
||
200 | i915_gem_init_ioctl(struct drm_device *dev, void *data, |
||
201 | struct drm_file *file) |
||
202 | { |
||
3480 | Serge | 203 | struct drm_i915_private *dev_priv = dev->dev_private; |
2332 | Serge | 204 | struct drm_i915_gem_init *args = data; |
205 | |||
3031 | serge | 206 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
207 | return -ENODEV; |
||
208 | |||
2332 | Serge | 209 | if (args->gtt_start >= args->gtt_end || |
210 | (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1)) |
||
211 | return -EINVAL; |
||
212 | |||
3031 | serge | 213 | /* GEM with user mode setting was never supported on ilk and later. */ |
214 | if (INTEL_INFO(dev)->gen >= 5) |
||
215 | return -ENODEV; |
||
216 | |||
2332 | Serge | 217 | mutex_lock(&dev->struct_mutex); |
3480 | Serge | 218 | i915_gem_setup_global_gtt(dev, args->gtt_start, args->gtt_end, |
219 | args->gtt_end); |
||
220 | dev_priv->gtt.mappable_end = args->gtt_end; |
||
2332 | Serge | 221 | mutex_unlock(&dev->struct_mutex); |
222 | |||
223 | return 0; |
||
224 | } |
||
2351 | Serge | 225 | #endif |
2332 | Serge | 226 | |
227 | int |
||
228 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, |
||
229 | struct drm_file *file) |
||
230 | { |
||
231 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
232 | struct drm_i915_gem_get_aperture *args = data; |
||
233 | struct drm_i915_gem_object *obj; |
||
234 | size_t pinned; |
||
235 | |||
236 | pinned = 0; |
||
237 | mutex_lock(&dev->struct_mutex); |
||
4104 | Serge | 238 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) |
3031 | serge | 239 | if (obj->pin_count) |
4104 | Serge | 240 | pinned += i915_gem_obj_ggtt_size(obj); |
2332 | Serge | 241 | mutex_unlock(&dev->struct_mutex); |
242 | |||
4104 | Serge | 243 | args->aper_size = dev_priv->gtt.base.total; |
2342 | Serge | 244 | args->aper_available_size = args->aper_size - pinned; |
2332 | Serge | 245 | |
246 | return 0; |
||
247 | } |
||
248 | |||
3480 | Serge | 249 | void *i915_gem_object_alloc(struct drm_device *dev) |
250 | { |
||
251 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
252 | return kmalloc(sizeof(struct drm_i915_gem_object), 0); |
||
253 | } |
||
254 | |||
255 | void i915_gem_object_free(struct drm_i915_gem_object *obj) |
||
256 | { |
||
257 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
||
258 | kfree(obj); |
||
259 | } |
||
260 | |||
3031 | serge | 261 | static int |
262 | i915_gem_create(struct drm_file *file, |
||
2332 | Serge | 263 | struct drm_device *dev, |
264 | uint64_t size, |
||
265 | uint32_t *handle_p) |
||
266 | { |
||
267 | struct drm_i915_gem_object *obj; |
||
268 | int ret; |
||
269 | u32 handle; |
||
270 | |||
271 | size = roundup(size, PAGE_SIZE); |
||
2342 | Serge | 272 | if (size == 0) |
273 | return -EINVAL; |
||
2332 | Serge | 274 | |
275 | /* Allocate the new object */ |
||
276 | obj = i915_gem_alloc_object(dev, size); |
||
277 | if (obj == NULL) |
||
278 | return -ENOMEM; |
||
279 | |||
280 | ret = drm_gem_handle_create(file, &obj->base, &handle); |
||
4104 | Serge | 281 | /* drop reference from allocate - handle holds it now */ |
282 | drm_gem_object_unreference_unlocked(&obj->base); |
||
283 | if (ret) |
||
2332 | Serge | 284 | return ret; |
285 | |||
286 | *handle_p = handle; |
||
287 | return 0; |
||
288 | } |
||
289 | |||
290 | int |
||
291 | i915_gem_dumb_create(struct drm_file *file, |
||
292 | struct drm_device *dev, |
||
293 | struct drm_mode_create_dumb *args) |
||
294 | { |
||
295 | /* have to work out size/pitch and return them */ |
||
296 | args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64); |
||
297 | args->size = args->pitch * args->height; |
||
298 | return i915_gem_create(file, dev, |
||
299 | args->size, &args->handle); |
||
300 | } |
||
301 | |||
2326 | Serge | 302 | /** |
2332 | Serge | 303 | * Creates a new mm object and returns a handle to it. |
304 | */ |
||
305 | int |
||
306 | i915_gem_create_ioctl(struct drm_device *dev, void *data, |
||
307 | struct drm_file *file) |
||
308 | { |
||
309 | struct drm_i915_gem_create *args = data; |
||
3031 | serge | 310 | |
2332 | Serge | 311 | return i915_gem_create(file, dev, |
312 | args->size, &args->handle); |
||
313 | } |
||
314 | |||
315 | |||
3260 | Serge | 316 | #if 0 |
2332 | Serge | 317 | |
3031 | serge | 318 | static inline int |
319 | __copy_to_user_swizzled(char __user *cpu_vaddr, |
||
320 | const char *gpu_vaddr, int gpu_offset, |
||
2332 | Serge | 321 | int length) |
322 | { |
||
3031 | serge | 323 | int ret, cpu_offset = 0; |
2332 | Serge | 324 | |
3031 | serge | 325 | while (length > 0) { |
326 | int cacheline_end = ALIGN(gpu_offset + 1, 64); |
||
327 | int this_length = min(cacheline_end - gpu_offset, length); |
||
328 | int swizzled_gpu_offset = gpu_offset ^ 64; |
||
2332 | Serge | 329 | |
3031 | serge | 330 | ret = __copy_to_user(cpu_vaddr + cpu_offset, |
331 | gpu_vaddr + swizzled_gpu_offset, |
||
332 | this_length); |
||
333 | if (ret) |
||
334 | return ret + length; |
||
2332 | Serge | 335 | |
3031 | serge | 336 | cpu_offset += this_length; |
337 | gpu_offset += this_length; |
||
338 | length -= this_length; |
||
339 | } |
||
340 | |||
341 | return 0; |
||
2332 | Serge | 342 | } |
343 | |||
3031 | serge | 344 | static inline int |
345 | __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, |
||
346 | const char __user *cpu_vaddr, |
||
347 | int length) |
||
2332 | Serge | 348 | { |
3031 | serge | 349 | int ret, cpu_offset = 0; |
2332 | Serge | 350 | |
351 | while (length > 0) { |
||
352 | int cacheline_end = ALIGN(gpu_offset + 1, 64); |
||
353 | int this_length = min(cacheline_end - gpu_offset, length); |
||
354 | int swizzled_gpu_offset = gpu_offset ^ 64; |
||
355 | |||
3031 | serge | 356 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, |
2332 | Serge | 357 | cpu_vaddr + cpu_offset, |
358 | this_length); |
||
3031 | serge | 359 | if (ret) |
360 | return ret + length; |
||
361 | |||
2332 | Serge | 362 | cpu_offset += this_length; |
363 | gpu_offset += this_length; |
||
364 | length -= this_length; |
||
365 | } |
||
366 | |||
3031 | serge | 367 | return 0; |
2332 | Serge | 368 | } |
369 | |||
3031 | serge | 370 | /* Per-page copy function for the shmem pread fastpath. |
371 | * Flushes invalid cachelines before reading the target if |
||
372 | * needs_clflush is set. */ |
||
2332 | Serge | 373 | static int |
3031 | serge | 374 | shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length, |
375 | char __user *user_data, |
||
376 | bool page_do_bit17_swizzling, bool needs_clflush) |
||
377 | { |
||
378 | char *vaddr; |
||
379 | int ret; |
||
380 | |||
381 | if (unlikely(page_do_bit17_swizzling)) |
||
382 | return -EINVAL; |
||
383 | |||
384 | vaddr = kmap_atomic(page); |
||
385 | if (needs_clflush) |
||
386 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
||
387 | page_length); |
||
388 | ret = __copy_to_user_inatomic(user_data, |
||
389 | vaddr + shmem_page_offset, |
||
390 | page_length); |
||
391 | kunmap_atomic(vaddr); |
||
392 | |||
393 | return ret ? -EFAULT : 0; |
||
394 | } |
||
395 | |||
396 | static void |
||
397 | shmem_clflush_swizzled_range(char *addr, unsigned long length, |
||
398 | bool swizzled) |
||
399 | { |
||
400 | if (unlikely(swizzled)) { |
||
401 | unsigned long start = (unsigned long) addr; |
||
402 | unsigned long end = (unsigned long) addr + length; |
||
403 | |||
404 | /* For swizzling simply ensure that we always flush both |
||
405 | * channels. Lame, but simple and it works. Swizzled |
||
406 | * pwrite/pread is far from a hotpath - current userspace |
||
407 | * doesn't use it at all. */ |
||
408 | start = round_down(start, 128); |
||
409 | end = round_up(end, 128); |
||
410 | |||
411 | drm_clflush_virt_range((void *)start, end - start); |
||
412 | } else { |
||
413 | drm_clflush_virt_range(addr, length); |
||
414 | } |
||
415 | |||
416 | } |
||
417 | |||
418 | /* Only difference to the fast-path function is that this can handle bit17 |
||
419 | * and uses non-atomic copy and kmap functions. */ |
||
420 | static int |
||
421 | shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length, |
||
422 | char __user *user_data, |
||
423 | bool page_do_bit17_swizzling, bool needs_clflush) |
||
424 | { |
||
425 | char *vaddr; |
||
426 | int ret; |
||
427 | |||
428 | vaddr = kmap(page); |
||
429 | if (needs_clflush) |
||
430 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
||
431 | page_length, |
||
432 | page_do_bit17_swizzling); |
||
433 | |||
434 | if (page_do_bit17_swizzling) |
||
435 | ret = __copy_to_user_swizzled(user_data, |
||
436 | vaddr, shmem_page_offset, |
||
437 | page_length); |
||
438 | else |
||
439 | ret = __copy_to_user(user_data, |
||
440 | vaddr + shmem_page_offset, |
||
441 | page_length); |
||
442 | kunmap(page); |
||
443 | |||
444 | return ret ? - EFAULT : 0; |
||
445 | } |
||
446 | |||
447 | static int |
||
448 | i915_gem_shmem_pread(struct drm_device *dev, |
||
2332 | Serge | 449 | struct drm_i915_gem_object *obj, |
450 | struct drm_i915_gem_pread *args, |
||
451 | struct drm_file *file) |
||
452 | { |
||
3031 | serge | 453 | char __user *user_data; |
2332 | Serge | 454 | ssize_t remain; |
455 | loff_t offset; |
||
3031 | serge | 456 | int shmem_page_offset, page_length, ret = 0; |
457 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
||
458 | int prefaulted = 0; |
||
459 | int needs_clflush = 0; |
||
3746 | Serge | 460 | struct sg_page_iter sg_iter; |
2332 | Serge | 461 | |
3746 | Serge | 462 | user_data = to_user_ptr(args->data_ptr); |
2332 | Serge | 463 | remain = args->size; |
464 | |||
3031 | serge | 465 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
466 | |||
467 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) { |
||
468 | /* If we're not in the cpu read domain, set ourself into the gtt |
||
469 | * read domain and manually flush cachelines (if required). This |
||
470 | * optimizes for the case when the gpu will dirty the data |
||
471 | * anyway again before the next pread happens. */ |
||
4104 | Serge | 472 | needs_clflush = !cpu_cache_is_coherent(dev, obj->cache_level); |
473 | if (i915_gem_obj_bound_any(obj)) { |
||
3031 | serge | 474 | ret = i915_gem_object_set_to_gtt_domain(obj, false); |
475 | if (ret) |
||
476 | return ret; |
||
477 | } |
||
478 | } |
||
479 | |||
480 | ret = i915_gem_object_get_pages(obj); |
||
481 | if (ret) |
||
482 | return ret; |
||
483 | |||
484 | i915_gem_object_pin_pages(obj); |
||
485 | |||
2332 | Serge | 486 | offset = args->offset; |
487 | |||
3746 | Serge | 488 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
489 | offset >> PAGE_SHIFT) { |
||
490 | struct page *page = sg_page_iter_page(&sg_iter); |
||
2332 | Serge | 491 | |
3031 | serge | 492 | if (remain <= 0) |
493 | break; |
||
494 | |||
2332 | Serge | 495 | /* Operation in this page |
496 | * |
||
3031 | serge | 497 | * shmem_page_offset = offset within page in shmem file |
2332 | Serge | 498 | * page_length = bytes to copy for this page |
499 | */ |
||
3031 | serge | 500 | shmem_page_offset = offset_in_page(offset); |
2332 | Serge | 501 | page_length = remain; |
3031 | serge | 502 | if ((shmem_page_offset + page_length) > PAGE_SIZE) |
503 | page_length = PAGE_SIZE - shmem_page_offset; |
||
2332 | Serge | 504 | |
3031 | serge | 505 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
506 | (page_to_phys(page) & (1 << 17)) != 0; |
||
2332 | Serge | 507 | |
3031 | serge | 508 | ret = shmem_pread_fast(page, shmem_page_offset, page_length, |
509 | user_data, page_do_bit17_swizzling, |
||
510 | needs_clflush); |
||
511 | if (ret == 0) |
||
512 | goto next_page; |
||
2332 | Serge | 513 | |
3031 | serge | 514 | mutex_unlock(&dev->struct_mutex); |
515 | |||
4104 | Serge | 516 | if (likely(!i915_prefault_disable) && !prefaulted) { |
3031 | serge | 517 | ret = fault_in_multipages_writeable(user_data, remain); |
518 | /* Userspace is tricking us, but we've already clobbered |
||
519 | * its pages with the prefault and promised to write the |
||
520 | * data up to the first fault. Hence ignore any errors |
||
521 | * and just continue. */ |
||
522 | (void)ret; |
||
523 | prefaulted = 1; |
||
524 | } |
||
525 | |||
526 | ret = shmem_pread_slow(page, shmem_page_offset, page_length, |
||
527 | user_data, page_do_bit17_swizzling, |
||
528 | needs_clflush); |
||
529 | |||
530 | mutex_lock(&dev->struct_mutex); |
||
531 | |||
532 | next_page: |
||
2332 | Serge | 533 | mark_page_accessed(page); |
3031 | serge | 534 | |
2332 | Serge | 535 | if (ret) |
3031 | serge | 536 | goto out; |
2332 | Serge | 537 | |
538 | remain -= page_length; |
||
539 | user_data += page_length; |
||
540 | offset += page_length; |
||
541 | } |
||
542 | |||
3031 | serge | 543 | out: |
544 | i915_gem_object_unpin_pages(obj); |
||
545 | |||
546 | return ret; |
||
2332 | Serge | 547 | } |
548 | |||
549 | /** |
||
3031 | serge | 550 | * Reads data from the object referenced by handle. |
551 | * |
||
552 | * On error, the contents of *data are undefined. |
||
2332 | Serge | 553 | */ |
3031 | serge | 554 | int |
555 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, |
||
556 | struct drm_file *file) |
||
557 | { |
||
558 | struct drm_i915_gem_pread *args = data; |
||
559 | struct drm_i915_gem_object *obj; |
||
560 | int ret = 0; |
||
561 | |||
562 | if (args->size == 0) |
||
563 | return 0; |
||
564 | |||
565 | if (!access_ok(VERIFY_WRITE, |
||
3746 | Serge | 566 | to_user_ptr(args->data_ptr), |
3031 | serge | 567 | args->size)) |
568 | return -EFAULT; |
||
569 | |||
570 | ret = i915_mutex_lock_interruptible(dev); |
||
571 | if (ret) |
||
572 | return ret; |
||
573 | |||
574 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
||
575 | if (&obj->base == NULL) { |
||
576 | ret = -ENOENT; |
||
577 | goto unlock; |
||
578 | } |
||
579 | |||
580 | /* Bounds check source. */ |
||
581 | if (args->offset > obj->base.size || |
||
582 | args->size > obj->base.size - args->offset) { |
||
583 | ret = -EINVAL; |
||
584 | goto out; |
||
585 | } |
||
586 | |||
587 | /* prime objects have no backing filp to GEM pread/pwrite |
||
588 | * pages from. |
||
589 | */ |
||
590 | if (!obj->base.filp) { |
||
591 | ret = -EINVAL; |
||
592 | goto out; |
||
593 | } |
||
594 | |||
595 | trace_i915_gem_object_pread(obj, args->offset, args->size); |
||
596 | |||
597 | ret = i915_gem_shmem_pread(dev, obj, args, file); |
||
598 | |||
599 | out: |
||
600 | drm_gem_object_unreference(&obj->base); |
||
601 | unlock: |
||
602 | mutex_unlock(&dev->struct_mutex); |
||
603 | return ret; |
||
604 | } |
||
605 | |||
606 | /* This is the fast write path which cannot handle |
||
607 | * page faults in the source data |
||
608 | */ |
||
609 | |||
610 | static inline int |
||
611 | fast_user_write(struct io_mapping *mapping, |
||
612 | loff_t page_base, int page_offset, |
||
613 | char __user *user_data, |
||
614 | int length) |
||
615 | { |
||
616 | void __iomem *vaddr_atomic; |
||
617 | void *vaddr; |
||
618 | unsigned long unwritten; |
||
619 | |||
620 | vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base); |
||
621 | /* We can use the cpu mem copy function because this is X86. */ |
||
622 | vaddr = (void __force*)vaddr_atomic + page_offset; |
||
623 | unwritten = __copy_from_user_inatomic_nocache(vaddr, |
||
624 | user_data, length); |
||
625 | io_mapping_unmap_atomic(vaddr_atomic); |
||
626 | return unwritten; |
||
627 | } |
||
3260 | Serge | 628 | #endif |
3031 | serge | 629 | |
3260 | Serge | 630 | #define offset_in_page(p) ((unsigned long)(p) & ~PAGE_MASK) |
3031 | serge | 631 | /** |
632 | * This is the fast pwrite path, where we copy the data directly from the |
||
633 | * user into the GTT, uncached. |
||
634 | */ |
||
2332 | Serge | 635 | static int |
3031 | serge | 636 | i915_gem_gtt_pwrite_fast(struct drm_device *dev, |
637 | struct drm_i915_gem_object *obj, |
||
638 | struct drm_i915_gem_pwrite *args, |
||
639 | struct drm_file *file) |
||
2332 | Serge | 640 | { |
3031 | serge | 641 | drm_i915_private_t *dev_priv = dev->dev_private; |
2332 | Serge | 642 | ssize_t remain; |
3031 | serge | 643 | loff_t offset, page_base; |
644 | char __user *user_data; |
||
645 | int page_offset, page_length, ret; |
||
3260 | Serge | 646 | char *vaddr; |
2332 | Serge | 647 | |
4104 | Serge | 648 | ret = i915_gem_obj_ggtt_pin(obj, 0, true, true); |
3031 | serge | 649 | if (ret) |
650 | goto out; |
||
651 | |||
652 | ret = i915_gem_object_set_to_gtt_domain(obj, true); |
||
653 | if (ret) |
||
654 | goto out_unpin; |
||
655 | |||
656 | ret = i915_gem_object_put_fence(obj); |
||
657 | if (ret) |
||
658 | goto out_unpin; |
||
659 | |||
3260 | Serge | 660 | vaddr = AllocKernelSpace(4096); |
661 | if(vaddr == NULL) |
||
662 | { |
||
663 | ret = -ENOSPC; |
||
664 | goto out_unpin; |
||
665 | }; |
||
666 | |||
3031 | serge | 667 | user_data = (char __user *) (uintptr_t) args->data_ptr; |
2332 | Serge | 668 | remain = args->size; |
669 | |||
4104 | Serge | 670 | offset = i915_gem_obj_ggtt_offset(obj) + args->offset; |
2332 | Serge | 671 | |
3031 | serge | 672 | while (remain > 0) { |
673 | /* Operation in this page |
||
674 | * |
||
675 | * page_base = page offset within aperture |
||
676 | * page_offset = offset within page |
||
677 | * page_length = bytes to copy for this page |
||
678 | */ |
||
679 | page_base = offset & PAGE_MASK; |
||
680 | page_offset = offset_in_page(offset); |
||
681 | page_length = remain; |
||
682 | if ((page_offset + remain) > PAGE_SIZE) |
||
683 | page_length = PAGE_SIZE - page_offset; |
||
2332 | Serge | 684 | |
3260 | Serge | 685 | MapPage(vaddr, page_base, PG_SW|PG_NOCACHE); |
3031 | serge | 686 | |
3260 | Serge | 687 | memcpy(vaddr+page_offset, user_data, page_length); |
688 | |||
3031 | serge | 689 | remain -= page_length; |
690 | user_data += page_length; |
||
691 | offset += page_length; |
||
2332 | Serge | 692 | } |
693 | |||
3260 | Serge | 694 | FreeKernelSpace(vaddr); |
695 | |||
3031 | serge | 696 | out_unpin: |
697 | i915_gem_object_unpin(obj); |
||
698 | out: |
||
699 | return ret; |
||
700 | } |
||
701 | |||
702 | /* Per-page copy function for the shmem pwrite fastpath. |
||
703 | * Flushes invalid cachelines before writing to the target if |
||
704 | * needs_clflush_before is set and flushes out any written cachelines after |
||
705 | * writing if needs_clflush is set. */ |
||
706 | static int |
||
707 | shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length, |
||
708 | char __user *user_data, |
||
709 | bool page_do_bit17_swizzling, |
||
710 | bool needs_clflush_before, |
||
711 | bool needs_clflush_after) |
||
712 | { |
||
713 | char *vaddr; |
||
3260 | Serge | 714 | int ret = 0; |
3031 | serge | 715 | |
716 | if (unlikely(page_do_bit17_swizzling)) |
||
717 | return -EINVAL; |
||
718 | |||
3260 | Serge | 719 | vaddr = (char *)MapIoMem((addr_t)page, 4096, PG_SW); |
3031 | serge | 720 | if (needs_clflush_before) |
721 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
||
722 | page_length); |
||
3260 | Serge | 723 | memcpy(vaddr + shmem_page_offset, |
3031 | serge | 724 | user_data, |
725 | page_length); |
||
726 | if (needs_clflush_after) |
||
727 | drm_clflush_virt_range(vaddr + shmem_page_offset, |
||
728 | page_length); |
||
3260 | Serge | 729 | FreeKernelSpace(vaddr); |
3031 | serge | 730 | |
731 | return ret ? -EFAULT : 0; |
||
732 | } |
||
3260 | Serge | 733 | #if 0 |
3031 | serge | 734 | |
735 | /* Only difference to the fast-path function is that this can handle bit17 |
||
736 | * and uses non-atomic copy and kmap functions. */ |
||
737 | static int |
||
738 | shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length, |
||
739 | char __user *user_data, |
||
740 | bool page_do_bit17_swizzling, |
||
741 | bool needs_clflush_before, |
||
742 | bool needs_clflush_after) |
||
743 | { |
||
744 | char *vaddr; |
||
745 | int ret; |
||
746 | |||
747 | vaddr = kmap(page); |
||
748 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) |
||
749 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
||
750 | page_length, |
||
751 | page_do_bit17_swizzling); |
||
752 | if (page_do_bit17_swizzling) |
||
753 | ret = __copy_from_user_swizzled(vaddr, shmem_page_offset, |
||
754 | user_data, |
||
755 | page_length); |
||
756 | else |
||
757 | ret = __copy_from_user(vaddr + shmem_page_offset, |
||
758 | user_data, |
||
759 | page_length); |
||
760 | if (needs_clflush_after) |
||
761 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
||
762 | page_length, |
||
763 | page_do_bit17_swizzling); |
||
764 | kunmap(page); |
||
765 | |||
766 | return ret ? -EFAULT : 0; |
||
767 | } |
||
3260 | Serge | 768 | #endif |
3031 | serge | 769 | |
3260 | Serge | 770 | |
3031 | serge | 771 | static int |
772 | i915_gem_shmem_pwrite(struct drm_device *dev, |
||
773 | struct drm_i915_gem_object *obj, |
||
774 | struct drm_i915_gem_pwrite *args, |
||
775 | struct drm_file *file) |
||
776 | { |
||
777 | ssize_t remain; |
||
778 | loff_t offset; |
||
779 | char __user *user_data; |
||
780 | int shmem_page_offset, page_length, ret = 0; |
||
781 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
||
782 | int hit_slowpath = 0; |
||
783 | int needs_clflush_after = 0; |
||
784 | int needs_clflush_before = 0; |
||
3746 | Serge | 785 | struct sg_page_iter sg_iter; |
3031 | serge | 786 | |
3746 | Serge | 787 | user_data = to_user_ptr(args->data_ptr); |
3031 | serge | 788 | remain = args->size; |
789 | |||
790 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
||
791 | |||
792 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
||
793 | /* If we're not in the cpu write domain, set ourself into the gtt |
||
794 | * write domain and manually flush cachelines (if required). This |
||
795 | * optimizes for the case when the gpu will use the data |
||
796 | * right away and we therefore have to clflush anyway. */ |
||
4104 | Serge | 797 | needs_clflush_after = cpu_write_needs_clflush(obj); |
798 | if (i915_gem_obj_bound_any(obj)) { |
||
3031 | serge | 799 | ret = i915_gem_object_set_to_gtt_domain(obj, true); |
800 | if (ret) |
||
801 | return ret; |
||
802 | } |
||
803 | } |
||
4104 | Serge | 804 | /* Same trick applies to invalidate partially written cachelines read |
805 | * before writing. */ |
||
806 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) |
||
807 | needs_clflush_before = |
||
808 | !cpu_cache_is_coherent(dev, obj->cache_level); |
||
3031 | serge | 809 | |
810 | ret = i915_gem_object_get_pages(obj); |
||
2332 | Serge | 811 | if (ret) |
3031 | serge | 812 | return ret; |
2332 | Serge | 813 | |
3031 | serge | 814 | i915_gem_object_pin_pages(obj); |
2332 | Serge | 815 | |
816 | offset = args->offset; |
||
3031 | serge | 817 | obj->dirty = 1; |
2332 | Serge | 818 | |
3746 | Serge | 819 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
820 | offset >> PAGE_SHIFT) { |
||
821 | struct page *page = sg_page_iter_page(&sg_iter); |
||
3031 | serge | 822 | int partial_cacheline_write; |
2332 | Serge | 823 | |
3031 | serge | 824 | if (remain <= 0) |
825 | break; |
||
826 | |||
2332 | Serge | 827 | /* Operation in this page |
828 | * |
||
829 | * shmem_page_offset = offset within page in shmem file |
||
830 | * page_length = bytes to copy for this page |
||
831 | */ |
||
832 | shmem_page_offset = offset_in_page(offset); |
||
833 | |||
834 | page_length = remain; |
||
835 | if ((shmem_page_offset + page_length) > PAGE_SIZE) |
||
836 | page_length = PAGE_SIZE - shmem_page_offset; |
||
837 | |||
3031 | serge | 838 | /* If we don't overwrite a cacheline completely we need to be |
839 | * careful to have up-to-date data by first clflushing. Don't |
||
840 | * overcomplicate things and flush the entire patch. */ |
||
841 | partial_cacheline_write = needs_clflush_before && |
||
842 | ((shmem_page_offset | page_length) |
||
3260 | Serge | 843 | & (x86_clflush_size - 1)); |
2332 | Serge | 844 | |
3031 | serge | 845 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
846 | (page_to_phys(page) & (1 << 17)) != 0; |
||
2332 | Serge | 847 | |
3031 | serge | 848 | ret = shmem_pwrite_fast(page, shmem_page_offset, page_length, |
849 | user_data, page_do_bit17_swizzling, |
||
850 | partial_cacheline_write, |
||
851 | needs_clflush_after); |
||
852 | if (ret == 0) |
||
853 | goto next_page; |
||
854 | |||
855 | hit_slowpath = 1; |
||
856 | mutex_unlock(&dev->struct_mutex); |
||
3260 | Serge | 857 | dbgprintf("%s need shmem_pwrite_slow\n",__FUNCTION__); |
3031 | serge | 858 | |
3260 | Serge | 859 | // ret = shmem_pwrite_slow(page, shmem_page_offset, page_length, |
860 | // user_data, page_do_bit17_swizzling, |
||
861 | // partial_cacheline_write, |
||
862 | // needs_clflush_after); |
||
863 | |||
3031 | serge | 864 | mutex_lock(&dev->struct_mutex); |
865 | |||
866 | next_page: |
||
2332 | Serge | 867 | |
3031 | serge | 868 | if (ret) |
869 | goto out; |
||
870 | |||
2332 | Serge | 871 | remain -= page_length; |
3031 | serge | 872 | user_data += page_length; |
2332 | Serge | 873 | offset += page_length; |
874 | } |
||
875 | |||
876 | out: |
||
3031 | serge | 877 | i915_gem_object_unpin_pages(obj); |
878 | |||
879 | if (hit_slowpath) { |
||
3480 | Serge | 880 | /* |
881 | * Fixup: Flush cpu caches in case we didn't flush the dirty |
||
882 | * cachelines in-line while writing and the object moved |
||
883 | * out of the cpu write domain while we've dropped the lock. |
||
884 | */ |
||
885 | if (!needs_clflush_after && |
||
886 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
||
4104 | Serge | 887 | if (i915_gem_clflush_object(obj, obj->pin_display)) |
3243 | Serge | 888 | i915_gem_chipset_flush(dev); |
3031 | serge | 889 | } |
2332 | Serge | 890 | } |
891 | |||
3031 | serge | 892 | if (needs_clflush_after) |
3243 | Serge | 893 | i915_gem_chipset_flush(dev); |
3031 | serge | 894 | |
2332 | Serge | 895 | return ret; |
896 | } |
||
3031 | serge | 897 | |
898 | /** |
||
899 | * Writes data to the object referenced by handle. |
||
900 | * |
||
901 | * On error, the contents of the buffer that were to be modified are undefined. |
||
902 | */ |
||
903 | int |
||
904 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, |
||
905 | struct drm_file *file) |
||
906 | { |
||
907 | struct drm_i915_gem_pwrite *args = data; |
||
908 | struct drm_i915_gem_object *obj; |
||
909 | int ret; |
||
910 | |||
4104 | Serge | 911 | if (args->size == 0) |
912 | return 0; |
||
913 | |||
3480 | Serge | 914 | if(args->handle == -2) |
915 | { |
||
916 | printf("%s handle %d\n", __FUNCTION__, args->handle); |
||
917 | return 0; |
||
918 | } |
||
919 | |||
3031 | serge | 920 | ret = i915_mutex_lock_interruptible(dev); |
921 | if (ret) |
||
922 | return ret; |
||
923 | |||
924 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
||
925 | if (&obj->base == NULL) { |
||
926 | ret = -ENOENT; |
||
927 | goto unlock; |
||
928 | } |
||
929 | |||
930 | /* Bounds check destination. */ |
||
931 | if (args->offset > obj->base.size || |
||
932 | args->size > obj->base.size - args->offset) { |
||
933 | ret = -EINVAL; |
||
934 | goto out; |
||
935 | } |
||
936 | |||
937 | /* prime objects have no backing filp to GEM pread/pwrite |
||
938 | * pages from. |
||
939 | */ |
||
940 | if (!obj->base.filp) { |
||
941 | ret = -EINVAL; |
||
942 | goto out; |
||
943 | } |
||
944 | |||
945 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); |
||
946 | |||
947 | ret = -EFAULT; |
||
948 | /* We can only do the GTT pwrite on untiled buffers, as otherwise |
||
949 | * it would end up going through the fenced access, and we'll get |
||
950 | * different detiling behavior between reading and writing. |
||
951 | * pread/pwrite currently are reading and writing from the CPU |
||
952 | * perspective, requiring manual detiling by the client. |
||
953 | */ |
||
3260 | Serge | 954 | // if (obj->phys_obj) { |
955 | // ret = i915_gem_phys_pwrite(dev, obj, args, file); |
||
956 | // goto out; |
||
957 | // } |
||
3031 | serge | 958 | |
4104 | Serge | 959 | if (obj->tiling_mode == I915_TILING_NONE && |
960 | obj->base.write_domain != I915_GEM_DOMAIN_CPU && |
||
961 | cpu_write_needs_clflush(obj)) { |
||
3031 | serge | 962 | ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); |
963 | /* Note that the gtt paths might fail with non-page-backed user |
||
964 | * pointers (e.g. gtt mappings when moving data between |
||
965 | * textures). Fallback to the shmem path in that case. */ |
||
966 | } |
||
967 | |||
968 | if (ret == -EFAULT || ret == -ENOSPC) |
||
3260 | Serge | 969 | ret = i915_gem_shmem_pwrite(dev, obj, args, file); |
3031 | serge | 970 | |
971 | out: |
||
972 | drm_gem_object_unreference(&obj->base); |
||
973 | unlock: |
||
974 | mutex_unlock(&dev->struct_mutex); |
||
975 | return ret; |
||
976 | } |
||
977 | |||
978 | int |
||
3480 | Serge | 979 | i915_gem_check_wedge(struct i915_gpu_error *error, |
3031 | serge | 980 | bool interruptible) |
981 | { |
||
3480 | Serge | 982 | if (i915_reset_in_progress(error)) { |
3031 | serge | 983 | /* Non-interruptible callers can't handle -EAGAIN, hence return |
984 | * -EIO unconditionally for these. */ |
||
985 | if (!interruptible) |
||
986 | return -EIO; |
||
2332 | Serge | 987 | |
3480 | Serge | 988 | /* Recovery complete, but the reset failed ... */ |
989 | if (i915_terminally_wedged(error)) |
||
3031 | serge | 990 | return -EIO; |
2332 | Serge | 991 | |
3031 | serge | 992 | return -EAGAIN; |
993 | } |
||
2332 | Serge | 994 | |
3031 | serge | 995 | return 0; |
996 | } |
||
2332 | Serge | 997 | |
3031 | serge | 998 | /* |
999 | * Compare seqno against outstanding lazy request. Emit a request if they are |
||
1000 | * equal. |
||
1001 | */ |
||
1002 | static int |
||
1003 | i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno) |
||
1004 | { |
||
1005 | int ret; |
||
2332 | Serge | 1006 | |
3031 | serge | 1007 | BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex)); |
2332 | Serge | 1008 | |
3031 | serge | 1009 | ret = 0; |
1010 | if (seqno == ring->outstanding_lazy_request) |
||
4104 | Serge | 1011 | ret = i915_add_request(ring, NULL); |
2332 | Serge | 1012 | |
3031 | serge | 1013 | return ret; |
1014 | } |
||
2332 | Serge | 1015 | |
3031 | serge | 1016 | /** |
1017 | * __wait_seqno - wait until execution of seqno has finished |
||
1018 | * @ring: the ring expected to report seqno |
||
1019 | * @seqno: duh! |
||
3480 | Serge | 1020 | * @reset_counter: reset sequence associated with the given seqno |
3031 | serge | 1021 | * @interruptible: do an interruptible wait (normally yes) |
1022 | * @timeout: in - how long to wait (NULL forever); out - how much time remaining |
||
1023 | * |
||
3480 | Serge | 1024 | * Note: It is of utmost importance that the passed in seqno and reset_counter |
1025 | * values have been read by the caller in an smp safe manner. Where read-side |
||
1026 | * locks are involved, it is sufficient to read the reset_counter before |
||
1027 | * unlocking the lock that protects the seqno. For lockless tricks, the |
||
1028 | * reset_counter _must_ be read before, and an appropriate smp_rmb must be |
||
1029 | * inserted. |
||
1030 | * |
||
3031 | serge | 1031 | * Returns 0 if the seqno was found within the alloted time. Else returns the |
1032 | * errno with remaining time filled in timeout argument. |
||
1033 | */ |
||
1034 | static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno, |
||
3480 | Serge | 1035 | unsigned reset_counter, |
3031 | serge | 1036 | bool interruptible, struct timespec *timeout) |
1037 | { |
||
1038 | drm_i915_private_t *dev_priv = ring->dev->dev_private; |
||
1039 | struct timespec before, now, wait_time={1,0}; |
||
1040 | unsigned long timeout_jiffies; |
||
1041 | long end; |
||
1042 | bool wait_forever = true; |
||
1043 | int ret; |
||
2332 | Serge | 1044 | |
4104 | Serge | 1045 | WARN(dev_priv->pc8.irqs_disabled, "IRQs disabled\n"); |
1046 | |||
3031 | serge | 1047 | if (i915_seqno_passed(ring->get_seqno(ring, true), seqno)) |
1048 | return 0; |
||
2332 | Serge | 1049 | |
3031 | serge | 1050 | trace_i915_gem_request_wait_begin(ring, seqno); |
2332 | Serge | 1051 | |
3031 | serge | 1052 | if (timeout != NULL) { |
1053 | wait_time = *timeout; |
||
1054 | wait_forever = false; |
||
1055 | } |
||
2332 | Serge | 1056 | |
4104 | Serge | 1057 | timeout_jiffies = timespec_to_jiffies_timeout(&wait_time); |
2332 | Serge | 1058 | |
3031 | serge | 1059 | if (WARN_ON(!ring->irq_get(ring))) |
1060 | return -ENODEV; |
||
2332 | Serge | 1061 | |
3031 | serge | 1062 | /* Record current time in case interrupted by signal, or wedged * */ |
1063 | getrawmonotonic(&before); |
||
2332 | Serge | 1064 | |
3031 | serge | 1065 | #define EXIT_COND \ |
1066 | (i915_seqno_passed(ring->get_seqno(ring, false), seqno) || \ |
||
3480 | Serge | 1067 | i915_reset_in_progress(&dev_priv->gpu_error) || \ |
1068 | reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) |
||
3031 | serge | 1069 | do { |
3266 | Serge | 1070 | if (interruptible) |
1071 | end = wait_event_interruptible_timeout(ring->irq_queue, |
||
1072 | EXIT_COND, |
||
1073 | timeout_jiffies); |
||
1074 | else |
||
3031 | serge | 1075 | end = wait_event_timeout(ring->irq_queue, EXIT_COND, |
1076 | timeout_jiffies); |
||
2332 | Serge | 1077 | |
3480 | Serge | 1078 | /* We need to check whether any gpu reset happened in between |
1079 | * the caller grabbing the seqno and now ... */ |
||
1080 | if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) |
||
1081 | end = -EAGAIN; |
||
1082 | |||
1083 | /* ... but upgrade the -EGAIN to an -EIO if the gpu is truely |
||
1084 | * gone. */ |
||
1085 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
||
3031 | serge | 1086 | if (ret) |
1087 | end = ret; |
||
1088 | } while (end == 0 && wait_forever); |
||
2332 | Serge | 1089 | |
3031 | serge | 1090 | getrawmonotonic(&now); |
2332 | Serge | 1091 | |
3031 | serge | 1092 | ring->irq_put(ring); |
1093 | trace_i915_gem_request_wait_end(ring, seqno); |
||
1094 | #undef EXIT_COND |
||
2332 | Serge | 1095 | |
3031 | serge | 1096 | if (timeout) { |
4104 | Serge | 1097 | // struct timespec sleep_time = timespec_sub(now, before); |
1098 | // *timeout = timespec_sub(*timeout, sleep_time); |
||
3031 | serge | 1099 | } |
2332 | Serge | 1100 | |
3031 | serge | 1101 | switch (end) { |
1102 | case -EIO: |
||
1103 | case -EAGAIN: /* Wedged */ |
||
1104 | case -ERESTARTSYS: /* Signal */ |
||
1105 | return (int)end; |
||
1106 | case 0: /* Timeout */ |
||
1107 | return -ETIME; |
||
1108 | default: /* Completed */ |
||
1109 | WARN_ON(end < 0); /* We're not aware of other errors */ |
||
1110 | return 0; |
||
1111 | } |
||
1112 | } |
||
2332 | Serge | 1113 | |
3031 | serge | 1114 | /** |
1115 | * Waits for a sequence number to be signaled, and cleans up the |
||
1116 | * request and object lists appropriately for that event. |
||
1117 | */ |
||
1118 | int |
||
1119 | i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno) |
||
1120 | { |
||
1121 | struct drm_device *dev = ring->dev; |
||
1122 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
1123 | bool interruptible = dev_priv->mm.interruptible; |
||
1124 | int ret; |
||
2332 | Serge | 1125 | |
3031 | serge | 1126 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); |
1127 | BUG_ON(seqno == 0); |
||
2332 | Serge | 1128 | |
3480 | Serge | 1129 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
3031 | serge | 1130 | if (ret) |
1131 | return ret; |
||
2332 | Serge | 1132 | |
3031 | serge | 1133 | ret = i915_gem_check_olr(ring, seqno); |
1134 | if (ret) |
||
1135 | return ret; |
||
2332 | Serge | 1136 | |
3480 | Serge | 1137 | return __wait_seqno(ring, seqno, |
1138 | atomic_read(&dev_priv->gpu_error.reset_counter), |
||
1139 | interruptible, NULL); |
||
3031 | serge | 1140 | } |
2332 | Serge | 1141 | |
4104 | Serge | 1142 | static int |
1143 | i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj, |
||
1144 | struct intel_ring_buffer *ring) |
||
1145 | { |
||
1146 | i915_gem_retire_requests_ring(ring); |
||
1147 | |||
1148 | /* Manually manage the write flush as we may have not yet |
||
1149 | * retired the buffer. |
||
1150 | * |
||
1151 | * Note that the last_write_seqno is always the earlier of |
||
1152 | * the two (read/write) seqno, so if we haved successfully waited, |
||
1153 | * we know we have passed the last write. |
||
1154 | */ |
||
1155 | obj->last_write_seqno = 0; |
||
1156 | obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS; |
||
1157 | |||
1158 | return 0; |
||
1159 | } |
||
1160 | |||
3031 | serge | 1161 | /** |
1162 | * Ensures that all rendering to the object has completed and the object is |
||
1163 | * safe to unbind from the GTT or access from the CPU. |
||
1164 | */ |
||
1165 | static __must_check int |
||
1166 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj, |
||
1167 | bool readonly) |
||
1168 | { |
||
1169 | struct intel_ring_buffer *ring = obj->ring; |
||
1170 | u32 seqno; |
||
1171 | int ret; |
||
2332 | Serge | 1172 | |
3031 | serge | 1173 | seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno; |
1174 | if (seqno == 0) |
||
1175 | return 0; |
||
2332 | Serge | 1176 | |
3031 | serge | 1177 | ret = i915_wait_seqno(ring, seqno); |
4104 | Serge | 1178 | if (ret) |
1179 | return ret; |
||
2332 | Serge | 1180 | |
4104 | Serge | 1181 | return i915_gem_object_wait_rendering__tail(obj, ring); |
3031 | serge | 1182 | } |
2332 | Serge | 1183 | |
3260 | Serge | 1184 | /* A nonblocking variant of the above wait. This is a highly dangerous routine |
1185 | * as the object state may change during this call. |
||
1186 | */ |
||
1187 | static __must_check int |
||
1188 | i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj, |
||
1189 | bool readonly) |
||
1190 | { |
||
1191 | struct drm_device *dev = obj->base.dev; |
||
1192 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
1193 | struct intel_ring_buffer *ring = obj->ring; |
||
3480 | Serge | 1194 | unsigned reset_counter; |
3260 | Serge | 1195 | u32 seqno; |
1196 | int ret; |
||
2332 | Serge | 1197 | |
3260 | Serge | 1198 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); |
1199 | BUG_ON(!dev_priv->mm.interruptible); |
||
2332 | Serge | 1200 | |
3260 | Serge | 1201 | seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno; |
1202 | if (seqno == 0) |
||
1203 | return 0; |
||
2332 | Serge | 1204 | |
3480 | Serge | 1205 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, true); |
3260 | Serge | 1206 | if (ret) |
1207 | return ret; |
||
2332 | Serge | 1208 | |
3260 | Serge | 1209 | ret = i915_gem_check_olr(ring, seqno); |
1210 | if (ret) |
||
1211 | return ret; |
||
2332 | Serge | 1212 | |
3480 | Serge | 1213 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
3260 | Serge | 1214 | mutex_unlock(&dev->struct_mutex); |
3480 | Serge | 1215 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL); |
3260 | Serge | 1216 | mutex_lock(&dev->struct_mutex); |
4104 | Serge | 1217 | if (ret) |
1218 | return ret; |
||
2332 | Serge | 1219 | |
4104 | Serge | 1220 | return i915_gem_object_wait_rendering__tail(obj, ring); |
3260 | Serge | 1221 | } |
2332 | Serge | 1222 | |
3260 | Serge | 1223 | /** |
1224 | * Called when user space prepares to use an object with the CPU, either |
||
1225 | * through the mmap ioctl's mapping or a GTT mapping. |
||
1226 | */ |
||
1227 | int |
||
1228 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, |
||
1229 | struct drm_file *file) |
||
1230 | { |
||
1231 | struct drm_i915_gem_set_domain *args = data; |
||
1232 | struct drm_i915_gem_object *obj; |
||
1233 | uint32_t read_domains = args->read_domains; |
||
1234 | uint32_t write_domain = args->write_domain; |
||
1235 | int ret; |
||
2332 | Serge | 1236 | |
3480 | Serge | 1237 | |
1238 | if(args->handle == -2) |
||
1239 | { |
||
1240 | printf("%s handle %d\n", __FUNCTION__, args->handle); |
||
1241 | return 0; |
||
1242 | } |
||
1243 | |||
3260 | Serge | 1244 | /* Only handle setting domains to types used by the CPU. */ |
1245 | if (write_domain & I915_GEM_GPU_DOMAINS) |
||
1246 | return -EINVAL; |
||
2332 | Serge | 1247 | |
3260 | Serge | 1248 | if (read_domains & I915_GEM_GPU_DOMAINS) |
1249 | return -EINVAL; |
||
2332 | Serge | 1250 | |
3260 | Serge | 1251 | /* Having something in the write domain implies it's in the read |
1252 | * domain, and only that read domain. Enforce that in the request. |
||
1253 | */ |
||
1254 | if (write_domain != 0 && read_domains != write_domain) |
||
1255 | return -EINVAL; |
||
2332 | Serge | 1256 | |
3260 | Serge | 1257 | ret = i915_mutex_lock_interruptible(dev); |
1258 | if (ret) |
||
1259 | return ret; |
||
2332 | Serge | 1260 | |
3260 | Serge | 1261 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
1262 | if (&obj->base == NULL) { |
||
1263 | ret = -ENOENT; |
||
1264 | goto unlock; |
||
1265 | } |
||
2332 | Serge | 1266 | |
3260 | Serge | 1267 | /* Try to flush the object off the GPU without holding the lock. |
1268 | * We will repeat the flush holding the lock in the normal manner |
||
1269 | * to catch cases where we are gazumped. |
||
1270 | */ |
||
1271 | ret = i915_gem_object_wait_rendering__nonblocking(obj, !write_domain); |
||
1272 | if (ret) |
||
1273 | goto unref; |
||
2332 | Serge | 1274 | |
3260 | Serge | 1275 | if (read_domains & I915_GEM_DOMAIN_GTT) { |
1276 | ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); |
||
2332 | Serge | 1277 | |
3260 | Serge | 1278 | /* Silently promote "you're not bound, there was nothing to do" |
1279 | * to success, since the client was just asking us to |
||
1280 | * make sure everything was done. |
||
1281 | */ |
||
1282 | if (ret == -EINVAL) |
||
1283 | ret = 0; |
||
1284 | } else { |
||
1285 | ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); |
||
1286 | } |
||
2332 | Serge | 1287 | |
3260 | Serge | 1288 | unref: |
1289 | drm_gem_object_unreference(&obj->base); |
||
1290 | unlock: |
||
1291 | mutex_unlock(&dev->struct_mutex); |
||
1292 | return ret; |
||
1293 | } |
||
2332 | Serge | 1294 | |
1295 | |||
1296 | |||
1297 | |||
1298 | |||
1299 | |||
3260 | Serge | 1300 | /** |
1301 | * Maps the contents of an object, returning the address it is mapped |
||
1302 | * into. |
||
1303 | * |
||
1304 | * While the mapping holds a reference on the contents of the object, it doesn't |
||
1305 | * imply a ref on the object itself. |
||
1306 | */ |
||
1307 | int |
||
1308 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, |
||
1309 | struct drm_file *file) |
||
1310 | { |
||
1311 | struct drm_i915_gem_mmap *args = data; |
||
1312 | struct drm_gem_object *obj; |
||
1313 | unsigned long addr = 0; |
||
2332 | Serge | 1314 | |
3480 | Serge | 1315 | if(args->handle == -2) |
1316 | { |
||
1317 | printf("%s handle %d\n", __FUNCTION__, args->handle); |
||
1318 | return 0; |
||
1319 | } |
||
1320 | |||
3260 | Serge | 1321 | obj = drm_gem_object_lookup(dev, file, args->handle); |
1322 | if (obj == NULL) |
||
1323 | return -ENOENT; |
||
4104 | Serge | 1324 | |
3260 | Serge | 1325 | /* prime objects have no backing filp to GEM mmap |
1326 | * pages from. |
||
1327 | */ |
||
1328 | if (!obj->filp) { |
||
1329 | drm_gem_object_unreference_unlocked(obj); |
||
1330 | return -EINVAL; |
||
1331 | } |
||
2332 | Serge | 1332 | |
3263 | Serge | 1333 | addr = vm_mmap(obj->filp, 0, args->size, |
1334 | PROT_READ | PROT_WRITE, MAP_SHARED, |
||
1335 | args->offset); |
||
3260 | Serge | 1336 | drm_gem_object_unreference_unlocked(obj); |
3263 | Serge | 1337 | if (IS_ERR((void *)addr)) |
1338 | return addr; |
||
2332 | Serge | 1339 | |
3260 | Serge | 1340 | args->addr_ptr = (uint64_t) addr; |
2332 | Serge | 1341 | |
3263 | Serge | 1342 | return 0; |
3260 | Serge | 1343 | } |
2332 | Serge | 1344 | |
1345 | |||
1346 | |||
1347 | |||
1348 | |||
1349 | |||
1350 | |||
1351 | |||
3031 | serge | 1352 | |
1353 | |||
1354 | |||
1355 | |||
1356 | |||
1357 | /** |
||
1358 | * i915_gem_release_mmap - remove physical page mappings |
||
1359 | * @obj: obj in question |
||
1360 | * |
||
1361 | * Preserve the reservation of the mmapping with the DRM core code, but |
||
1362 | * relinquish ownership of the pages back to the system. |
||
1363 | * |
||
1364 | * It is vital that we remove the page mapping if we have mapped a tiled |
||
1365 | * object through the GTT and then lose the fence register due to |
||
1366 | * resource pressure. Similarly if the object has been moved out of the |
||
1367 | * aperture, than pages mapped into userspace must be revoked. Removing the |
||
1368 | * mapping will then trigger a page fault on the next user access, allowing |
||
1369 | * fixup by i915_gem_fault(). |
||
1370 | */ |
||
1371 | void |
||
1372 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) |
||
1373 | { |
||
1374 | if (!obj->fault_mappable) |
||
1375 | return; |
||
1376 | |||
4104 | Serge | 1377 | // drm_vma_node_unmap(&obj->base.vma_node, obj->base.dev->dev_mapping); |
3031 | serge | 1378 | obj->fault_mappable = false; |
1379 | } |
||
1380 | |||
3480 | Serge | 1381 | uint32_t |
2332 | Serge | 1382 | i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode) |
1383 | { |
||
1384 | uint32_t gtt_size; |
||
1385 | |||
1386 | if (INTEL_INFO(dev)->gen >= 4 || |
||
1387 | tiling_mode == I915_TILING_NONE) |
||
1388 | return size; |
||
1389 | |||
1390 | /* Previous chips need a power-of-two fence region when tiling */ |
||
1391 | if (INTEL_INFO(dev)->gen == 3) |
||
1392 | gtt_size = 1024*1024; |
||
1393 | else |
||
1394 | gtt_size = 512*1024; |
||
1395 | |||
1396 | while (gtt_size < size) |
||
1397 | gtt_size <<= 1; |
||
1398 | |||
1399 | return gtt_size; |
||
1400 | } |
||
1401 | |||
1402 | /** |
||
1403 | * i915_gem_get_gtt_alignment - return required GTT alignment for an object |
||
1404 | * @obj: object to check |
||
1405 | * |
||
1406 | * Return the required GTT alignment for an object, taking into account |
||
1407 | * potential fence register mapping. |
||
1408 | */ |
||
3480 | Serge | 1409 | uint32_t |
1410 | i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size, |
||
1411 | int tiling_mode, bool fenced) |
||
2332 | Serge | 1412 | { |
1413 | /* |
||
1414 | * Minimum alignment is 4k (GTT page size), but might be greater |
||
1415 | * if a fence register is needed for the object. |
||
1416 | */ |
||
3480 | Serge | 1417 | if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) || |
2332 | Serge | 1418 | tiling_mode == I915_TILING_NONE) |
1419 | return 4096; |
||
1420 | |||
1421 | /* |
||
1422 | * Previous chips need to be aligned to the size of the smallest |
||
1423 | * fence register that can contain the object. |
||
1424 | */ |
||
1425 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
||
1426 | } |
||
1427 | |||
1428 | /** |
||
1429 | * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an |
||
1430 | * unfenced object |
||
1431 | * @dev: the device |
||
1432 | * @size: size of the object |
||
1433 | * @tiling_mode: tiling mode of the object |
||
1434 | * |
||
1435 | * Return the required GTT alignment for an object, only taking into account |
||
1436 | * unfenced tiled surface requirements. |
||
1437 | */ |
||
1438 | uint32_t |
||
1439 | i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev, |
||
1440 | uint32_t size, |
||
1441 | int tiling_mode) |
||
1442 | { |
||
1443 | /* |
||
1444 | * Minimum alignment is 4k (GTT page size) for sane hw. |
||
1445 | */ |
||
1446 | if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) || |
||
1447 | tiling_mode == I915_TILING_NONE) |
||
1448 | return 4096; |
||
1449 | |||
1450 | /* Previous hardware however needs to be aligned to a power-of-two |
||
1451 | * tile height. The simplest method for determining this is to reuse |
||
1452 | * the power-of-tile object size. |
||
1453 | */ |
||
1454 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
||
1455 | } |
||
1456 | |||
3480 | Serge | 1457 | int |
1458 | i915_gem_mmap_gtt(struct drm_file *file, |
||
1459 | struct drm_device *dev, |
||
1460 | uint32_t handle, |
||
1461 | uint64_t *offset) |
||
1462 | { |
||
1463 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
1464 | struct drm_i915_gem_object *obj; |
||
1465 | unsigned long pfn; |
||
1466 | char *mem, *ptr; |
||
1467 | int ret; |
||
1468 | |||
1469 | ret = i915_mutex_lock_interruptible(dev); |
||
1470 | if (ret) |
||
1471 | return ret; |
||
1472 | |||
1473 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle)); |
||
1474 | if (&obj->base == NULL) { |
||
1475 | ret = -ENOENT; |
||
1476 | goto unlock; |
||
1477 | } |
||
1478 | |||
1479 | if (obj->base.size > dev_priv->gtt.mappable_end) { |
||
1480 | ret = -E2BIG; |
||
1481 | goto out; |
||
1482 | } |
||
1483 | |||
1484 | if (obj->madv != I915_MADV_WILLNEED) { |
||
1485 | DRM_ERROR("Attempting to mmap a purgeable buffer\n"); |
||
1486 | ret = -EINVAL; |
||
1487 | goto out; |
||
1488 | } |
||
1489 | /* Now bind it into the GTT if needed */ |
||
4104 | Serge | 1490 | ret = i915_gem_obj_ggtt_pin(obj, 0, true, false); |
3480 | Serge | 1491 | if (ret) |
1492 | goto out; |
||
1493 | |||
1494 | ret = i915_gem_object_set_to_gtt_domain(obj, 1); |
||
1495 | if (ret) |
||
1496 | goto unpin; |
||
1497 | |||
1498 | ret = i915_gem_object_get_fence(obj); |
||
1499 | if (ret) |
||
1500 | goto unpin; |
||
1501 | |||
1502 | obj->fault_mappable = true; |
||
1503 | |||
4104 | Serge | 1504 | pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj); |
3480 | Serge | 1505 | |
1506 | /* Finally, remap it using the new GTT offset */ |
||
1507 | |||
1508 | mem = UserAlloc(obj->base.size); |
||
1509 | if(unlikely(mem == NULL)) |
||
1510 | { |
||
1511 | ret = -ENOMEM; |
||
1512 | goto unpin; |
||
1513 | } |
||
1514 | |||
1515 | for(ptr = mem; ptr < mem + obj->base.size; ptr+= 4096, pfn+= 4096) |
||
1516 | MapPage(ptr, pfn, PG_SHARED|PG_UW); |
||
1517 | |||
1518 | unpin: |
||
1519 | i915_gem_object_unpin(obj); |
||
1520 | |||
1521 | |||
4104 | Serge | 1522 | *offset = mem; |
3480 | Serge | 1523 | |
1524 | out: |
||
1525 | drm_gem_object_unreference(&obj->base); |
||
1526 | unlock: |
||
1527 | mutex_unlock(&dev->struct_mutex); |
||
1528 | return ret; |
||
1529 | } |
||
1530 | |||
1531 | /** |
||
1532 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing |
||
1533 | * @dev: DRM device |
||
1534 | * @data: GTT mapping ioctl data |
||
1535 | * @file: GEM object info |
||
1536 | * |
||
1537 | * Simply returns the fake offset to userspace so it can mmap it. |
||
1538 | * The mmap call will end up in drm_gem_mmap(), which will set things |
||
1539 | * up so we can get faults in the handler above. |
||
1540 | * |
||
1541 | * The fault handler will take care of binding the object into the GTT |
||
1542 | * (since it may have been evicted to make room for something), allocating |
||
1543 | * a fence register, and mapping the appropriate aperture address into |
||
1544 | * userspace. |
||
1545 | */ |
||
1546 | int |
||
1547 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, |
||
1548 | struct drm_file *file) |
||
1549 | { |
||
1550 | struct drm_i915_gem_mmap_gtt *args = data; |
||
1551 | |||
1552 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); |
||
1553 | } |
||
1554 | |||
3031 | serge | 1555 | /* Immediately discard the backing storage */ |
1556 | static void |
||
1557 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) |
||
1558 | { |
||
1559 | // struct inode *inode; |
||
2332 | Serge | 1560 | |
3031 | serge | 1561 | // i915_gem_object_free_mmap_offset(obj); |
2332 | Serge | 1562 | |
3263 | Serge | 1563 | if (obj->base.filp == NULL) |
1564 | return; |
||
2332 | Serge | 1565 | |
3031 | serge | 1566 | /* Our goal here is to return as much of the memory as |
1567 | * is possible back to the system as we are called from OOM. |
||
1568 | * To do this we must instruct the shmfs to drop all of its |
||
1569 | * backing pages, *now*. |
||
1570 | */ |
||
1571 | // inode = obj->base.filp->f_path.dentry->d_inode; |
||
1572 | // shmem_truncate_range(inode, 0, (loff_t)-1); |
||
2332 | Serge | 1573 | |
3031 | serge | 1574 | obj->madv = __I915_MADV_PURGED; |
1575 | } |
||
2332 | Serge | 1576 | |
3031 | serge | 1577 | static inline int |
1578 | i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj) |
||
1579 | { |
||
1580 | return obj->madv == I915_MADV_DONTNEED; |
||
1581 | } |
||
2332 | Serge | 1582 | |
3031 | serge | 1583 | static void |
1584 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj) |
||
1585 | { |
||
3746 | Serge | 1586 | struct sg_page_iter sg_iter; |
1587 | int ret; |
||
2332 | Serge | 1588 | |
3031 | serge | 1589 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
2332 | Serge | 1590 | |
3031 | serge | 1591 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
1592 | if (ret) { |
||
1593 | /* In the event of a disaster, abandon all caches and |
||
1594 | * hope for the best. |
||
1595 | */ |
||
1596 | WARN_ON(ret != -EIO); |
||
4104 | Serge | 1597 | i915_gem_clflush_object(obj, true); |
3031 | serge | 1598 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
1599 | } |
||
2332 | Serge | 1600 | |
3031 | serge | 1601 | if (obj->madv == I915_MADV_DONTNEED) |
1602 | obj->dirty = 0; |
||
2332 | Serge | 1603 | |
3746 | Serge | 1604 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { |
1605 | struct page *page = sg_page_iter_page(&sg_iter); |
||
2332 | Serge | 1606 | |
3290 | Serge | 1607 | page_cache_release(page); |
3243 | Serge | 1608 | } |
1609 | //DRM_DEBUG_KMS("%s release %d pages\n", __FUNCTION__, page_count); |
||
3290 | Serge | 1610 | |
4104 | Serge | 1611 | obj->dirty = 0; |
3243 | Serge | 1612 | |
1613 | sg_free_table(obj->pages); |
||
1614 | kfree(obj->pages); |
||
3031 | serge | 1615 | } |
2332 | Serge | 1616 | |
3480 | Serge | 1617 | int |
3031 | serge | 1618 | i915_gem_object_put_pages(struct drm_i915_gem_object *obj) |
1619 | { |
||
1620 | const struct drm_i915_gem_object_ops *ops = obj->ops; |
||
2332 | Serge | 1621 | |
3243 | Serge | 1622 | if (obj->pages == NULL) |
3031 | serge | 1623 | return 0; |
2332 | Serge | 1624 | |
3031 | serge | 1625 | if (obj->pages_pin_count) |
1626 | return -EBUSY; |
||
1627 | |||
4104 | Serge | 1628 | BUG_ON(i915_gem_obj_bound_any(obj)); |
1629 | |||
3243 | Serge | 1630 | /* ->put_pages might need to allocate memory for the bit17 swizzle |
1631 | * array, hence protect them from being reaped by removing them from gtt |
||
1632 | * lists early. */ |
||
4104 | Serge | 1633 | list_del(&obj->global_list); |
3243 | Serge | 1634 | |
3031 | serge | 1635 | ops->put_pages(obj); |
3243 | Serge | 1636 | obj->pages = NULL; |
3031 | serge | 1637 | |
1638 | if (i915_gem_object_is_purgeable(obj)) |
||
1639 | i915_gem_object_truncate(obj); |
||
1640 | |||
1641 | return 0; |
||
1642 | } |
||
1643 | |||
1644 | |||
1645 | |||
1646 | |||
1647 | |||
1648 | |||
1649 | |||
1650 | |||
2332 | Serge | 1651 | static int |
3031 | serge | 1652 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj) |
2332 | Serge | 1653 | { |
3260 | Serge | 1654 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
3243 | Serge | 1655 | int page_count, i; |
4104 | Serge | 1656 | struct sg_table *st; |
3243 | Serge | 1657 | struct scatterlist *sg; |
3746 | Serge | 1658 | struct sg_page_iter sg_iter; |
3243 | Serge | 1659 | struct page *page; |
3746 | Serge | 1660 | unsigned long last_pfn = 0; /* suppress gcc warning */ |
3243 | Serge | 1661 | gfp_t gfp; |
2332 | Serge | 1662 | |
3243 | Serge | 1663 | /* Assert that the object is not currently in any GPU domain. As it |
1664 | * wasn't in the GTT, there shouldn't be any way it could have been in |
||
1665 | * a GPU cache |
||
2332 | Serge | 1666 | */ |
3243 | Serge | 1667 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); |
1668 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); |
||
1669 | |||
1670 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
||
1671 | if (st == NULL) |
||
1672 | return -ENOMEM; |
||
1673 | |||
2332 | Serge | 1674 | page_count = obj->base.size / PAGE_SIZE; |
3243 | Serge | 1675 | if (sg_alloc_table(st, page_count, GFP_KERNEL)) { |
1676 | kfree(st); |
||
3746 | Serge | 1677 | FAIL(); |
2332 | Serge | 1678 | return -ENOMEM; |
3243 | Serge | 1679 | } |
2332 | Serge | 1680 | |
3243 | Serge | 1681 | /* Get the list of pages out of our struct file. They'll be pinned |
1682 | * at this point until we release them. |
||
1683 | * |
||
1684 | * Fail silently without starting the shrinker |
||
1685 | */ |
||
3746 | Serge | 1686 | sg = st->sgl; |
1687 | st->nents = 0; |
||
1688 | for (i = 0; i < page_count; i++) { |
||
4104 | Serge | 1689 | page = shmem_read_mapping_page_gfp(obj->base.filp, i, gfp); |
3260 | Serge | 1690 | if (IS_ERR(page)) { |
1691 | dbgprintf("%s invalid page %p\n", __FUNCTION__, page); |
||
2332 | Serge | 1692 | goto err_pages; |
1693 | |||
3260 | Serge | 1694 | } |
3746 | Serge | 1695 | |
1696 | if (!i || page_to_pfn(page) != last_pfn + 1) { |
||
1697 | if (i) |
||
1698 | sg = sg_next(sg); |
||
1699 | st->nents++; |
||
3243 | Serge | 1700 | sg_set_page(sg, page, PAGE_SIZE, 0); |
3746 | Serge | 1701 | } else { |
1702 | sg->length += PAGE_SIZE; |
||
1703 | } |
||
1704 | last_pfn = page_to_pfn(page); |
||
3243 | Serge | 1705 | } |
3031 | serge | 1706 | |
3746 | Serge | 1707 | sg_mark_end(sg); |
3243 | Serge | 1708 | obj->pages = st; |
3031 | serge | 1709 | |
2332 | Serge | 1710 | return 0; |
1711 | |||
1712 | err_pages: |
||
3746 | Serge | 1713 | sg_mark_end(sg); |
1714 | for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) |
||
1715 | page_cache_release(sg_page_iter_page(&sg_iter)); |
||
3243 | Serge | 1716 | sg_free_table(st); |
1717 | kfree(st); |
||
3746 | Serge | 1718 | FAIL(); |
3243 | Serge | 1719 | return PTR_ERR(page); |
2332 | Serge | 1720 | } |
1721 | |||
3031 | serge | 1722 | /* Ensure that the associated pages are gathered from the backing storage |
1723 | * and pinned into our object. i915_gem_object_get_pages() may be called |
||
1724 | * multiple times before they are released by a single call to |
||
1725 | * i915_gem_object_put_pages() - once the pages are no longer referenced |
||
1726 | * either as a result of memory pressure (reaping pages under the shrinker) |
||
1727 | * or as the object is itself released. |
||
1728 | */ |
||
1729 | int |
||
1730 | i915_gem_object_get_pages(struct drm_i915_gem_object *obj) |
||
2332 | Serge | 1731 | { |
3031 | serge | 1732 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
1733 | const struct drm_i915_gem_object_ops *ops = obj->ops; |
||
1734 | int ret; |
||
2332 | Serge | 1735 | |
3243 | Serge | 1736 | if (obj->pages) |
3031 | serge | 1737 | return 0; |
2332 | Serge | 1738 | |
3031 | serge | 1739 | BUG_ON(obj->pages_pin_count); |
2332 | Serge | 1740 | |
3031 | serge | 1741 | ret = ops->get_pages(obj); |
1742 | if (ret) |
||
1743 | return ret; |
||
2344 | Serge | 1744 | |
4104 | Serge | 1745 | list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
3243 | Serge | 1746 | return 0; |
2332 | Serge | 1747 | } |
1748 | |||
1749 | void |
||
1750 | i915_gem_object_move_to_active(struct drm_i915_gem_object *obj, |
||
3243 | Serge | 1751 | struct intel_ring_buffer *ring) |
2332 | Serge | 1752 | { |
1753 | struct drm_device *dev = obj->base.dev; |
||
1754 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
3243 | Serge | 1755 | u32 seqno = intel_ring_get_seqno(ring); |
2332 | Serge | 1756 | |
1757 | BUG_ON(ring == NULL); |
||
4104 | Serge | 1758 | if (obj->ring != ring && obj->last_write_seqno) { |
1759 | /* Keep the seqno relative to the current ring */ |
||
1760 | obj->last_write_seqno = seqno; |
||
1761 | } |
||
2332 | Serge | 1762 | obj->ring = ring; |
1763 | |||
1764 | /* Add a reference if we're newly entering the active list. */ |
||
1765 | if (!obj->active) { |
||
2344 | Serge | 1766 | drm_gem_object_reference(&obj->base); |
2332 | Serge | 1767 | obj->active = 1; |
1768 | } |
||
1769 | |||
1770 | list_move_tail(&obj->ring_list, &ring->active_list); |
||
1771 | |||
3031 | serge | 1772 | obj->last_read_seqno = seqno; |
1773 | |||
2332 | Serge | 1774 | if (obj->fenced_gpu_access) { |
3031 | serge | 1775 | obj->last_fenced_seqno = seqno; |
1776 | |||
1777 | /* Bump MRU to take account of the delayed flush */ |
||
1778 | if (obj->fence_reg != I915_FENCE_REG_NONE) { |
||
2332 | Serge | 1779 | struct drm_i915_fence_reg *reg; |
1780 | |||
1781 | reg = &dev_priv->fence_regs[obj->fence_reg]; |
||
3031 | serge | 1782 | list_move_tail(®->lru_list, |
1783 | &dev_priv->mm.fence_list); |
||
1784 | } |
||
2332 | Serge | 1785 | } |
1786 | } |
||
1787 | |||
2344 | Serge | 1788 | static void |
3031 | serge | 1789 | i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj) |
2344 | Serge | 1790 | { |
4104 | Serge | 1791 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
1792 | struct i915_address_space *ggtt_vm = &dev_priv->gtt.base; |
||
1793 | struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm); |
||
2332 | Serge | 1794 | |
3031 | serge | 1795 | BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS); |
2344 | Serge | 1796 | BUG_ON(!obj->active); |
2332 | Serge | 1797 | |
4104 | Serge | 1798 | list_move_tail(&vma->mm_list, &ggtt_vm->inactive_list); |
2344 | Serge | 1799 | |
3031 | serge | 1800 | list_del_init(&obj->ring_list); |
2352 | Serge | 1801 | obj->ring = NULL; |
2344 | Serge | 1802 | |
3031 | serge | 1803 | obj->last_read_seqno = 0; |
1804 | obj->last_write_seqno = 0; |
||
1805 | obj->base.write_domain = 0; |
||
1806 | |||
1807 | obj->last_fenced_seqno = 0; |
||
2352 | Serge | 1808 | obj->fenced_gpu_access = false; |
2344 | Serge | 1809 | |
2352 | Serge | 1810 | obj->active = 0; |
1811 | drm_gem_object_unreference(&obj->base); |
||
1812 | |||
1813 | WARN_ON(i915_verify_lists(dev)); |
||
1814 | } |
||
1815 | |||
3243 | Serge | 1816 | static int |
3480 | Serge | 1817 | i915_gem_init_seqno(struct drm_device *dev, u32 seqno) |
2344 | Serge | 1818 | { |
3243 | Serge | 1819 | struct drm_i915_private *dev_priv = dev->dev_private; |
1820 | struct intel_ring_buffer *ring; |
||
1821 | int ret, i, j; |
||
2344 | Serge | 1822 | |
3480 | Serge | 1823 | /* Carefully retire all requests without writing to the rings */ |
3243 | Serge | 1824 | for_each_ring(ring, dev_priv, i) { |
3480 | Serge | 1825 | ret = intel_ring_idle(ring); |
3243 | Serge | 1826 | if (ret) |
1827 | return ret; |
||
3480 | Serge | 1828 | } |
1829 | i915_gem_retire_requests(dev); |
||
3243 | Serge | 1830 | |
3480 | Serge | 1831 | /* Finally reset hw state */ |
3243 | Serge | 1832 | for_each_ring(ring, dev_priv, i) { |
3480 | Serge | 1833 | intel_ring_init_seqno(ring, seqno); |
1834 | |||
3243 | Serge | 1835 | for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++) |
1836 | ring->sync_seqno[j] = 0; |
||
1837 | } |
||
1838 | |||
1839 | return 0; |
||
2344 | Serge | 1840 | } |
1841 | |||
3480 | Serge | 1842 | int i915_gem_set_seqno(struct drm_device *dev, u32 seqno) |
1843 | { |
||
1844 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
1845 | int ret; |
||
1846 | |||
1847 | if (seqno == 0) |
||
1848 | return -EINVAL; |
||
1849 | |||
1850 | /* HWS page needs to be set less than what we |
||
1851 | * will inject to ring |
||
1852 | */ |
||
1853 | ret = i915_gem_init_seqno(dev, seqno - 1); |
||
1854 | if (ret) |
||
1855 | return ret; |
||
1856 | |||
1857 | /* Carefully set the last_seqno value so that wrap |
||
1858 | * detection still works |
||
1859 | */ |
||
1860 | dev_priv->next_seqno = seqno; |
||
1861 | dev_priv->last_seqno = seqno - 1; |
||
1862 | if (dev_priv->last_seqno == 0) |
||
1863 | dev_priv->last_seqno--; |
||
1864 | |||
1865 | return 0; |
||
1866 | } |
||
1867 | |||
3243 | Serge | 1868 | int |
1869 | i915_gem_get_seqno(struct drm_device *dev, u32 *seqno) |
||
2344 | Serge | 1870 | { |
3243 | Serge | 1871 | struct drm_i915_private *dev_priv = dev->dev_private; |
2344 | Serge | 1872 | |
3243 | Serge | 1873 | /* reserve 0 for non-seqno */ |
1874 | if (dev_priv->next_seqno == 0) { |
||
3480 | Serge | 1875 | int ret = i915_gem_init_seqno(dev, 0); |
3243 | Serge | 1876 | if (ret) |
1877 | return ret; |
||
1878 | |||
1879 | dev_priv->next_seqno = 1; |
||
1880 | } |
||
1881 | |||
3480 | Serge | 1882 | *seqno = dev_priv->last_seqno = dev_priv->next_seqno++; |
3243 | Serge | 1883 | return 0; |
2332 | Serge | 1884 | } |
1885 | |||
4104 | Serge | 1886 | int __i915_add_request(struct intel_ring_buffer *ring, |
2352 | Serge | 1887 | struct drm_file *file, |
4104 | Serge | 1888 | struct drm_i915_gem_object *obj, |
3031 | serge | 1889 | u32 *out_seqno) |
2352 | Serge | 1890 | { |
1891 | drm_i915_private_t *dev_priv = ring->dev->dev_private; |
||
3031 | serge | 1892 | struct drm_i915_gem_request *request; |
4104 | Serge | 1893 | u32 request_ring_position, request_start; |
2352 | Serge | 1894 | int was_empty; |
1895 | int ret; |
||
2332 | Serge | 1896 | |
4104 | Serge | 1897 | request_start = intel_ring_get_tail(ring); |
3031 | serge | 1898 | /* |
1899 | * Emit any outstanding flushes - execbuf can fail to emit the flush |
||
1900 | * after having emitted the batchbuffer command. Hence we need to fix |
||
1901 | * things up similar to emitting the lazy request. The difference here |
||
1902 | * is that the flush _must_ happen before the next request, no matter |
||
1903 | * what. |
||
1904 | */ |
||
4104 | Serge | 1905 | ret = intel_ring_flush_all_caches(ring); |
1906 | if (ret) |
||
1907 | return ret; |
||
2332 | Serge | 1908 | |
3031 | serge | 1909 | request = kmalloc(sizeof(*request), GFP_KERNEL); |
1910 | if (request == NULL) |
||
1911 | return -ENOMEM; |
||
1912 | |||
1913 | |||
1914 | /* Record the position of the start of the request so that |
||
1915 | * should we detect the updated seqno part-way through the |
||
4104 | Serge | 1916 | * GPU processing the request, we never over-estimate the |
3031 | serge | 1917 | * position of the head. |
1918 | */ |
||
4104 | Serge | 1919 | request_ring_position = intel_ring_get_tail(ring); |
3031 | serge | 1920 | |
3243 | Serge | 1921 | ret = ring->add_request(ring); |
3031 | serge | 1922 | if (ret) { |
1923 | kfree(request); |
||
4104 | Serge | 1924 | return ret; |
3031 | serge | 1925 | } |
2332 | Serge | 1926 | |
3243 | Serge | 1927 | request->seqno = intel_ring_get_seqno(ring); |
2352 | Serge | 1928 | request->ring = ring; |
4104 | Serge | 1929 | request->head = request_start; |
3031 | serge | 1930 | request->tail = request_ring_position; |
4104 | Serge | 1931 | request->ctx = ring->last_context; |
1932 | request->batch_obj = obj; |
||
1933 | |||
1934 | /* Whilst this request exists, batch_obj will be on the |
||
1935 | * active_list, and so will hold the active reference. Only when this |
||
1936 | * request is retired will the the batch_obj be moved onto the |
||
1937 | * inactive_list and lose its active reference. Hence we do not need |
||
1938 | * to explicitly hold another reference here. |
||
1939 | */ |
||
1940 | |||
1941 | if (request->ctx) |
||
1942 | i915_gem_context_reference(request->ctx); |
||
1943 | |||
3031 | serge | 1944 | request->emitted_jiffies = GetTimerTicks(); |
2352 | Serge | 1945 | was_empty = list_empty(&ring->request_list); |
1946 | list_add_tail(&request->list, &ring->request_list); |
||
3031 | serge | 1947 | request->file_priv = NULL; |
2332 | Serge | 1948 | |
3263 | Serge | 1949 | if (file) { |
1950 | struct drm_i915_file_private *file_priv = file->driver_priv; |
||
2332 | Serge | 1951 | |
3263 | Serge | 1952 | spin_lock(&file_priv->mm.lock); |
1953 | request->file_priv = file_priv; |
||
1954 | list_add_tail(&request->client_list, |
||
1955 | &file_priv->mm.request_list); |
||
1956 | spin_unlock(&file_priv->mm.lock); |
||
1957 | } |
||
1958 | |||
1959 | trace_i915_gem_request_add(ring, request->seqno); |
||
3031 | serge | 1960 | ring->outstanding_lazy_request = 0; |
2332 | Serge | 1961 | |
4104 | Serge | 1962 | if (!dev_priv->ums.mm_suspended) { |
1963 | // i915_queue_hangcheck(ring->dev); |
||
1964 | |||
1965 | if (was_empty) { |
||
2360 | Serge | 1966 | queue_delayed_work(dev_priv->wq, |
3482 | Serge | 1967 | &dev_priv->mm.retire_work, |
1968 | round_jiffies_up_relative(HZ)); |
||
4104 | Serge | 1969 | intel_mark_busy(dev_priv->dev); |
1970 | } |
||
1971 | } |
||
3031 | serge | 1972 | |
1973 | if (out_seqno) |
||
3243 | Serge | 1974 | *out_seqno = request->seqno; |
2352 | Serge | 1975 | return 0; |
1976 | } |
||
2332 | Serge | 1977 | |
3263 | Serge | 1978 | static inline void |
1979 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) |
||
1980 | { |
||
1981 | struct drm_i915_file_private *file_priv = request->file_priv; |
||
2332 | Serge | 1982 | |
3263 | Serge | 1983 | if (!file_priv) |
1984 | return; |
||
2332 | Serge | 1985 | |
3263 | Serge | 1986 | spin_lock(&file_priv->mm.lock); |
1987 | if (request->file_priv) { |
||
1988 | list_del(&request->client_list); |
||
1989 | request->file_priv = NULL; |
||
1990 | } |
||
1991 | spin_unlock(&file_priv->mm.lock); |
||
1992 | } |
||
2332 | Serge | 1993 | |
4104 | Serge | 1994 | static bool i915_head_inside_object(u32 acthd, struct drm_i915_gem_object *obj, |
1995 | struct i915_address_space *vm) |
||
1996 | { |
||
1997 | if (acthd >= i915_gem_obj_offset(obj, vm) && |
||
1998 | acthd < i915_gem_obj_offset(obj, vm) + obj->base.size) |
||
1999 | return true; |
||
2000 | |||
2001 | return false; |
||
2002 | } |
||
2003 | |||
2004 | static bool i915_head_inside_request(const u32 acthd_unmasked, |
||
2005 | const u32 request_start, |
||
2006 | const u32 request_end) |
||
2007 | { |
||
2008 | const u32 acthd = acthd_unmasked & HEAD_ADDR; |
||
2009 | |||
2010 | if (request_start < request_end) { |
||
2011 | if (acthd >= request_start && acthd < request_end) |
||
2012 | return true; |
||
2013 | } else if (request_start > request_end) { |
||
2014 | if (acthd >= request_start || acthd < request_end) |
||
2015 | return true; |
||
2016 | } |
||
2017 | |||
2018 | return false; |
||
2019 | } |
||
2020 | |||
2021 | static struct i915_address_space * |
||
2022 | request_to_vm(struct drm_i915_gem_request *request) |
||
2023 | { |
||
2024 | struct drm_i915_private *dev_priv = request->ring->dev->dev_private; |
||
2025 | struct i915_address_space *vm; |
||
2026 | |||
2027 | vm = &dev_priv->gtt.base; |
||
2028 | |||
2029 | return vm; |
||
2030 | } |
||
2031 | |||
2032 | static bool i915_request_guilty(struct drm_i915_gem_request *request, |
||
2033 | const u32 acthd, bool *inside) |
||
2034 | { |
||
2035 | /* There is a possibility that unmasked head address |
||
2036 | * pointing inside the ring, matches the batch_obj address range. |
||
2037 | * However this is extremely unlikely. |
||
2038 | */ |
||
2039 | if (request->batch_obj) { |
||
2040 | if (i915_head_inside_object(acthd, request->batch_obj, |
||
2041 | request_to_vm(request))) { |
||
2042 | *inside = true; |
||
2043 | return true; |
||
2044 | } |
||
2045 | } |
||
2046 | |||
2047 | if (i915_head_inside_request(acthd, request->head, request->tail)) { |
||
2048 | *inside = false; |
||
2049 | return true; |
||
2050 | } |
||
2051 | |||
2052 | return false; |
||
2053 | } |
||
2054 | |||
2055 | static void i915_set_reset_status(struct intel_ring_buffer *ring, |
||
2056 | struct drm_i915_gem_request *request, |
||
2057 | u32 acthd) |
||
2058 | { |
||
2059 | struct i915_ctx_hang_stats *hs = NULL; |
||
2060 | bool inside, guilty; |
||
2061 | unsigned long offset = 0; |
||
2062 | |||
2063 | /* Innocent until proven guilty */ |
||
2064 | guilty = false; |
||
2065 | |||
2066 | if (request->batch_obj) |
||
2067 | offset = i915_gem_obj_offset(request->batch_obj, |
||
2068 | request_to_vm(request)); |
||
2069 | |||
2070 | if (ring->hangcheck.action != HANGCHECK_WAIT && |
||
2071 | i915_request_guilty(request, acthd, &inside)) { |
||
2072 | DRM_ERROR("%s hung %s bo (0x%lx ctx %d) at 0x%x\n", |
||
2073 | ring->name, |
||
2074 | inside ? "inside" : "flushing", |
||
2075 | offset, |
||
2076 | request->ctx ? request->ctx->id : 0, |
||
2077 | acthd); |
||
2078 | |||
2079 | guilty = true; |
||
2080 | } |
||
2081 | |||
2082 | /* If contexts are disabled or this is the default context, use |
||
2083 | * file_priv->reset_state |
||
2084 | */ |
||
2085 | if (request->ctx && request->ctx->id != DEFAULT_CONTEXT_ID) |
||
2086 | hs = &request->ctx->hang_stats; |
||
2087 | else if (request->file_priv) |
||
2088 | hs = &request->file_priv->hang_stats; |
||
2089 | |||
2090 | if (hs) { |
||
2091 | if (guilty) |
||
2092 | hs->batch_active++; |
||
2093 | else |
||
2094 | hs->batch_pending++; |
||
2095 | } |
||
2096 | } |
||
2097 | |||
2098 | static void i915_gem_free_request(struct drm_i915_gem_request *request) |
||
2099 | { |
||
2100 | list_del(&request->list); |
||
2101 | i915_gem_request_remove_from_client(request); |
||
2102 | |||
2103 | if (request->ctx) |
||
2104 | i915_gem_context_unreference(request->ctx); |
||
2105 | |||
2106 | kfree(request); |
||
2107 | } |
||
2108 | |||
3031 | serge | 2109 | static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv, |
2110 | struct intel_ring_buffer *ring) |
||
2111 | { |
||
4104 | Serge | 2112 | u32 completed_seqno; |
2113 | u32 acthd; |
||
2114 | |||
2115 | acthd = intel_ring_get_active_head(ring); |
||
2116 | completed_seqno = ring->get_seqno(ring, false); |
||
2117 | |||
3031 | serge | 2118 | while (!list_empty(&ring->request_list)) { |
2119 | struct drm_i915_gem_request *request; |
||
2332 | Serge | 2120 | |
3031 | serge | 2121 | request = list_first_entry(&ring->request_list, |
2122 | struct drm_i915_gem_request, |
||
2123 | list); |
||
2332 | Serge | 2124 | |
4104 | Serge | 2125 | if (request->seqno > completed_seqno) |
2126 | i915_set_reset_status(ring, request, acthd); |
||
2127 | |||
2128 | i915_gem_free_request(request); |
||
3031 | serge | 2129 | } |
2332 | Serge | 2130 | |
3031 | serge | 2131 | while (!list_empty(&ring->active_list)) { |
2132 | struct drm_i915_gem_object *obj; |
||
2332 | Serge | 2133 | |
3031 | serge | 2134 | obj = list_first_entry(&ring->active_list, |
2135 | struct drm_i915_gem_object, |
||
2136 | ring_list); |
||
2332 | Serge | 2137 | |
3031 | serge | 2138 | i915_gem_object_move_to_inactive(obj); |
2139 | } |
||
2140 | } |
||
2332 | Serge | 2141 | |
3746 | Serge | 2142 | void i915_gem_restore_fences(struct drm_device *dev) |
3031 | serge | 2143 | { |
2144 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
2145 | int i; |
||
2332 | Serge | 2146 | |
3031 | serge | 2147 | for (i = 0; i < dev_priv->num_fence_regs; i++) { |
2148 | struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i]; |
||
4104 | Serge | 2149 | |
2150 | /* |
||
2151 | * Commit delayed tiling changes if we have an object still |
||
2152 | * attached to the fence, otherwise just clear the fence. |
||
2153 | */ |
||
2154 | if (reg->obj) { |
||
2155 | i915_gem_object_update_fence(reg->obj, reg, |
||
2156 | reg->obj->tiling_mode); |
||
2157 | } else { |
||
2158 | i915_gem_write_fence(dev, i, NULL); |
||
2159 | } |
||
3031 | serge | 2160 | } |
2161 | } |
||
2360 | Serge | 2162 | |
3031 | serge | 2163 | void i915_gem_reset(struct drm_device *dev) |
2164 | { |
||
2165 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
2166 | struct intel_ring_buffer *ring; |
||
2167 | int i; |
||
2360 | Serge | 2168 | |
3031 | serge | 2169 | for_each_ring(ring, dev_priv, i) |
2170 | i915_gem_reset_ring_lists(dev_priv, ring); |
||
2360 | Serge | 2171 | |
3746 | Serge | 2172 | i915_gem_restore_fences(dev); |
3031 | serge | 2173 | } |
2360 | Serge | 2174 | |
2352 | Serge | 2175 | /** |
2176 | * This function clears the request list as sequence numbers are passed. |
||
2177 | */ |
||
3031 | serge | 2178 | void |
2352 | Serge | 2179 | i915_gem_retire_requests_ring(struct intel_ring_buffer *ring) |
2180 | { |
||
2181 | uint32_t seqno; |
||
2332 | Serge | 2182 | |
2352 | Serge | 2183 | if (list_empty(&ring->request_list)) |
2184 | return; |
||
2332 | Serge | 2185 | |
2352 | Serge | 2186 | WARN_ON(i915_verify_lists(ring->dev)); |
2332 | Serge | 2187 | |
3031 | serge | 2188 | seqno = ring->get_seqno(ring, true); |
2332 | Serge | 2189 | |
2352 | Serge | 2190 | while (!list_empty(&ring->request_list)) { |
2191 | struct drm_i915_gem_request *request; |
||
2332 | Serge | 2192 | |
2352 | Serge | 2193 | request = list_first_entry(&ring->request_list, |
2194 | struct drm_i915_gem_request, |
||
2195 | list); |
||
2332 | Serge | 2196 | |
2352 | Serge | 2197 | if (!i915_seqno_passed(seqno, request->seqno)) |
2198 | break; |
||
2332 | Serge | 2199 | |
2352 | Serge | 2200 | trace_i915_gem_request_retire(ring, request->seqno); |
3031 | serge | 2201 | /* We know the GPU must have read the request to have |
2202 | * sent us the seqno + interrupt, so use the position |
||
2203 | * of tail of the request to update the last known position |
||
2204 | * of the GPU head. |
||
2205 | */ |
||
2206 | ring->last_retired_head = request->tail; |
||
2332 | Serge | 2207 | |
4104 | Serge | 2208 | i915_gem_free_request(request); |
2352 | Serge | 2209 | } |
2332 | Serge | 2210 | |
2352 | Serge | 2211 | /* Move any buffers on the active list that are no longer referenced |
2212 | * by the ringbuffer to the flushing/inactive lists as appropriate. |
||
2213 | */ |
||
2214 | while (!list_empty(&ring->active_list)) { |
||
2215 | struct drm_i915_gem_object *obj; |
||
2332 | Serge | 2216 | |
2352 | Serge | 2217 | obj = list_first_entry(&ring->active_list, |
2218 | struct drm_i915_gem_object, |
||
2219 | ring_list); |
||
2332 | Serge | 2220 | |
3031 | serge | 2221 | if (!i915_seqno_passed(seqno, obj->last_read_seqno)) |
2352 | Serge | 2222 | break; |
2332 | Serge | 2223 | |
2352 | Serge | 2224 | i915_gem_object_move_to_inactive(obj); |
2225 | } |
||
2332 | Serge | 2226 | |
2352 | Serge | 2227 | if (unlikely(ring->trace_irq_seqno && |
2228 | i915_seqno_passed(seqno, ring->trace_irq_seqno))) { |
||
2229 | ring->irq_put(ring); |
||
2230 | ring->trace_irq_seqno = 0; |
||
2231 | } |
||
2332 | Serge | 2232 | |
2352 | Serge | 2233 | WARN_ON(i915_verify_lists(ring->dev)); |
2234 | } |
||
2332 | Serge | 2235 | |
2352 | Serge | 2236 | void |
2237 | i915_gem_retire_requests(struct drm_device *dev) |
||
2238 | { |
||
2239 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
3031 | serge | 2240 | struct intel_ring_buffer *ring; |
2352 | Serge | 2241 | int i; |
2332 | Serge | 2242 | |
3031 | serge | 2243 | for_each_ring(ring, dev_priv, i) |
2244 | i915_gem_retire_requests_ring(ring); |
||
2352 | Serge | 2245 | } |
2246 | |||
2360 | Serge | 2247 | static void |
2248 | i915_gem_retire_work_handler(struct work_struct *work) |
||
2249 | { |
||
2250 | drm_i915_private_t *dev_priv; |
||
2251 | struct drm_device *dev; |
||
3031 | serge | 2252 | struct intel_ring_buffer *ring; |
2360 | Serge | 2253 | bool idle; |
2254 | int i; |
||
2352 | Serge | 2255 | |
2360 | Serge | 2256 | dev_priv = container_of(work, drm_i915_private_t, |
2257 | mm.retire_work.work); |
||
2258 | dev = dev_priv->dev; |
||
2352 | Serge | 2259 | |
2360 | Serge | 2260 | /* Come back later if the device is busy... */ |
2261 | if (!mutex_trylock(&dev->struct_mutex)) { |
||
3482 | Serge | 2262 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2263 | round_jiffies_up_relative(HZ)); |
||
3243 | Serge | 2264 | return; |
2360 | Serge | 2265 | } |
2352 | Serge | 2266 | |
2360 | Serge | 2267 | i915_gem_retire_requests(dev); |
2352 | Serge | 2268 | |
2360 | Serge | 2269 | /* Send a periodic flush down the ring so we don't hold onto GEM |
2270 | * objects indefinitely. |
||
2271 | */ |
||
2272 | idle = true; |
||
3031 | serge | 2273 | for_each_ring(ring, dev_priv, i) { |
2274 | if (ring->gpu_caches_dirty) |
||
4104 | Serge | 2275 | i915_add_request(ring, NULL); |
2352 | Serge | 2276 | |
2360 | Serge | 2277 | idle &= list_empty(&ring->request_list); |
2278 | } |
||
2352 | Serge | 2279 | |
4104 | Serge | 2280 | if (!dev_priv->ums.mm_suspended && !idle) |
3482 | Serge | 2281 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2282 | round_jiffies_up_relative(HZ)); |
||
3031 | serge | 2283 | if (idle) |
2284 | intel_mark_idle(dev); |
||
2360 | Serge | 2285 | |
2286 | mutex_unlock(&dev->struct_mutex); |
||
2287 | } |
||
2288 | |||
2344 | Serge | 2289 | /** |
3031 | serge | 2290 | * Ensures that an object will eventually get non-busy by flushing any required |
2291 | * write domains, emitting any outstanding lazy request and retiring and |
||
2292 | * completed requests. |
||
2352 | Serge | 2293 | */ |
3031 | serge | 2294 | static int |
2295 | i915_gem_object_flush_active(struct drm_i915_gem_object *obj) |
||
2352 | Serge | 2296 | { |
3031 | serge | 2297 | int ret; |
2352 | Serge | 2298 | |
3031 | serge | 2299 | if (obj->active) { |
2300 | ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno); |
||
2301 | if (ret) |
||
2302 | return ret; |
||
2352 | Serge | 2303 | |
3031 | serge | 2304 | i915_gem_retire_requests_ring(obj->ring); |
2305 | } |
||
2352 | Serge | 2306 | |
3031 | serge | 2307 | return 0; |
2308 | } |
||
2352 | Serge | 2309 | |
3243 | Serge | 2310 | /** |
2311 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT |
||
2312 | * @DRM_IOCTL_ARGS: standard ioctl arguments |
||
2313 | * |
||
2314 | * Returns 0 if successful, else an error is returned with the remaining time in |
||
2315 | * the timeout parameter. |
||
2316 | * -ETIME: object is still busy after timeout |
||
2317 | * -ERESTARTSYS: signal interrupted the wait |
||
2318 | * -ENONENT: object doesn't exist |
||
2319 | * Also possible, but rare: |
||
2320 | * -EAGAIN: GPU wedged |
||
2321 | * -ENOMEM: damn |
||
2322 | * -ENODEV: Internal IRQ fail |
||
2323 | * -E?: The add request failed |
||
2324 | * |
||
2325 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any |
||
2326 | * non-zero timeout parameter the wait ioctl will wait for the given number of |
||
2327 | * nanoseconds on an object becoming unbusy. Since the wait itself does so |
||
2328 | * without holding struct_mutex the object may become re-busied before this |
||
2329 | * function completes. A similar but shorter * race condition exists in the busy |
||
2330 | * ioctl |
||
2331 | */ |
||
4246 | Serge | 2332 | int |
2333 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) |
||
2334 | { |
||
2335 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
2336 | struct drm_i915_gem_wait *args = data; |
||
2337 | struct drm_i915_gem_object *obj; |
||
2338 | struct intel_ring_buffer *ring = NULL; |
||
2339 | struct timespec timeout_stack, *timeout = NULL; |
||
2340 | unsigned reset_counter; |
||
2341 | u32 seqno = 0; |
||
2342 | int ret = 0; |
||
2352 | Serge | 2343 | |
4246 | Serge | 2344 | if (args->timeout_ns >= 0) { |
2345 | timeout_stack = ns_to_timespec(args->timeout_ns); |
||
2346 | timeout = &timeout_stack; |
||
2347 | } |
||
2352 | Serge | 2348 | |
4246 | Serge | 2349 | ret = i915_mutex_lock_interruptible(dev); |
2350 | if (ret) |
||
2351 | return ret; |
||
2352 | Serge | 2352 | |
4246 | Serge | 2353 | if(args->bo_handle == -2) |
2354 | { |
||
2355 | obj = get_fb_obj(); |
||
2356 | drm_gem_object_reference(&obj->base); |
||
2357 | } |
||
2358 | else |
||
2359 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle)); |
||
2360 | if (&obj->base == NULL) { |
||
2361 | mutex_unlock(&dev->struct_mutex); |
||
2362 | return -ENOENT; |
||
2363 | } |
||
2352 | Serge | 2364 | |
4246 | Serge | 2365 | /* Need to make sure the object gets inactive eventually. */ |
2366 | ret = i915_gem_object_flush_active(obj); |
||
2367 | if (ret) |
||
2368 | goto out; |
||
2352 | Serge | 2369 | |
4246 | Serge | 2370 | if (obj->active) { |
2371 | seqno = obj->last_read_seqno; |
||
2372 | ring = obj->ring; |
||
2373 | } |
||
2352 | Serge | 2374 | |
4246 | Serge | 2375 | if (seqno == 0) |
2376 | goto out; |
||
2352 | Serge | 2377 | |
4246 | Serge | 2378 | /* Do this after OLR check to make sure we make forward progress polling |
2379 | * on this IOCTL with a 0 timeout (like busy ioctl) |
||
2380 | */ |
||
2381 | if (!args->timeout_ns) { |
||
2382 | ret = -ETIME; |
||
2383 | goto out; |
||
2384 | } |
||
2352 | Serge | 2385 | |
4246 | Serge | 2386 | drm_gem_object_unreference(&obj->base); |
2387 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
||
2388 | mutex_unlock(&dev->struct_mutex); |
||
2352 | Serge | 2389 | |
4246 | Serge | 2390 | ret = __wait_seqno(ring, seqno, reset_counter, true, timeout); |
2391 | if (timeout) |
||
2392 | args->timeout_ns = timespec_to_ns(timeout); |
||
2393 | return ret; |
||
3243 | Serge | 2394 | |
4246 | Serge | 2395 | out: |
2396 | drm_gem_object_unreference(&obj->base); |
||
2397 | mutex_unlock(&dev->struct_mutex); |
||
2398 | return ret; |
||
2399 | } |
||
3243 | Serge | 2400 | |
2352 | Serge | 2401 | /** |
3031 | serge | 2402 | * i915_gem_object_sync - sync an object to a ring. |
2403 | * |
||
2404 | * @obj: object which may be in use on another ring. |
||
2405 | * @to: ring we wish to use the object on. May be NULL. |
||
2406 | * |
||
2407 | * This code is meant to abstract object synchronization with the GPU. |
||
2408 | * Calling with NULL implies synchronizing the object with the CPU |
||
2409 | * rather than a particular GPU ring. |
||
2410 | * |
||
2411 | * Returns 0 if successful, else propagates up the lower layer error. |
||
2344 | Serge | 2412 | */ |
2413 | int |
||
3031 | serge | 2414 | i915_gem_object_sync(struct drm_i915_gem_object *obj, |
2415 | struct intel_ring_buffer *to) |
||
2344 | Serge | 2416 | { |
3031 | serge | 2417 | struct intel_ring_buffer *from = obj->ring; |
2418 | u32 seqno; |
||
2419 | int ret, idx; |
||
2332 | Serge | 2420 | |
3031 | serge | 2421 | if (from == NULL || to == from) |
2422 | return 0; |
||
2332 | Serge | 2423 | |
3031 | serge | 2424 | if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev)) |
2425 | return i915_gem_object_wait_rendering(obj, false); |
||
2332 | Serge | 2426 | |
3031 | serge | 2427 | idx = intel_ring_sync_index(from, to); |
2428 | |||
2429 | seqno = obj->last_read_seqno; |
||
2430 | if (seqno <= from->sync_seqno[idx]) |
||
2431 | return 0; |
||
2432 | |||
2433 | ret = i915_gem_check_olr(obj->ring, seqno); |
||
2434 | if (ret) |
||
2435 | return ret; |
||
2436 | |||
2437 | ret = to->sync_to(to, from, seqno); |
||
2438 | if (!ret) |
||
3243 | Serge | 2439 | /* We use last_read_seqno because sync_to() |
2440 | * might have just caused seqno wrap under |
||
2441 | * the radar. |
||
2442 | */ |
||
2443 | from->sync_seqno[idx] = obj->last_read_seqno; |
||
3031 | serge | 2444 | |
2445 | return ret; |
||
2344 | Serge | 2446 | } |
2332 | Serge | 2447 | |
2344 | Serge | 2448 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) |
2449 | { |
||
2450 | u32 old_write_domain, old_read_domains; |
||
2332 | Serge | 2451 | |
2344 | Serge | 2452 | /* Force a pagefault for domain tracking on next user access */ |
2453 | // i915_gem_release_mmap(obj); |
||
2332 | Serge | 2454 | |
2344 | Serge | 2455 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
2456 | return; |
||
2332 | Serge | 2457 | |
3480 | Serge | 2458 | /* Wait for any direct GTT access to complete */ |
2459 | mb(); |
||
2460 | |||
2344 | Serge | 2461 | old_read_domains = obj->base.read_domains; |
2462 | old_write_domain = obj->base.write_domain; |
||
2351 | Serge | 2463 | |
2344 | Serge | 2464 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; |
2465 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; |
||
2332 | Serge | 2466 | |
2351 | Serge | 2467 | trace_i915_gem_object_change_domain(obj, |
2468 | old_read_domains, |
||
2469 | old_write_domain); |
||
2344 | Serge | 2470 | } |
2332 | Serge | 2471 | |
4104 | Serge | 2472 | int i915_vma_unbind(struct i915_vma *vma) |
2344 | Serge | 2473 | { |
4104 | Serge | 2474 | struct drm_i915_gem_object *obj = vma->obj; |
3031 | serge | 2475 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
3480 | Serge | 2476 | int ret; |
2332 | Serge | 2477 | |
3263 | Serge | 2478 | if(obj == get_fb_obj()) |
2479 | return 0; |
||
2480 | |||
4104 | Serge | 2481 | if (list_empty(&vma->vma_link)) |
2344 | Serge | 2482 | return 0; |
2332 | Serge | 2483 | |
4104 | Serge | 2484 | if (!drm_mm_node_allocated(&vma->node)) |
2485 | goto destroy; |
||
2486 | |||
3031 | serge | 2487 | if (obj->pin_count) |
2488 | return -EBUSY; |
||
2332 | Serge | 2489 | |
3243 | Serge | 2490 | BUG_ON(obj->pages == NULL); |
3031 | serge | 2491 | |
2344 | Serge | 2492 | ret = i915_gem_object_finish_gpu(obj); |
3031 | serge | 2493 | if (ret) |
2344 | Serge | 2494 | return ret; |
2495 | /* Continue on if we fail due to EIO, the GPU is hung so we |
||
2496 | * should be safe and we need to cleanup or else we might |
||
2497 | * cause memory corruption through use-after-free. |
||
2498 | */ |
||
2332 | Serge | 2499 | |
2344 | Serge | 2500 | i915_gem_object_finish_gtt(obj); |
2332 | Serge | 2501 | |
2344 | Serge | 2502 | /* release the fence reg _after_ flushing */ |
2503 | ret = i915_gem_object_put_fence(obj); |
||
3031 | serge | 2504 | if (ret) |
2344 | Serge | 2505 | return ret; |
2332 | Serge | 2506 | |
4104 | Serge | 2507 | trace_i915_vma_unbind(vma); |
2332 | Serge | 2508 | |
3031 | serge | 2509 | if (obj->has_global_gtt_mapping) |
3243 | Serge | 2510 | i915_gem_gtt_unbind_object(obj); |
3031 | serge | 2511 | if (obj->has_aliasing_ppgtt_mapping) { |
2512 | i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj); |
||
2513 | obj->has_aliasing_ppgtt_mapping = 0; |
||
2514 | } |
||
2515 | i915_gem_gtt_finish_object(obj); |
||
4104 | Serge | 2516 | i915_gem_object_unpin_pages(obj); |
2332 | Serge | 2517 | |
4104 | Serge | 2518 | list_del(&vma->mm_list); |
2344 | Serge | 2519 | /* Avoid an unnecessary call to unbind on rebind. */ |
4104 | Serge | 2520 | if (i915_is_ggtt(vma->vm)) |
2344 | Serge | 2521 | obj->map_and_fenceable = true; |
2332 | Serge | 2522 | |
4104 | Serge | 2523 | drm_mm_remove_node(&vma->node); |
2332 | Serge | 2524 | |
4104 | Serge | 2525 | destroy: |
2526 | i915_gem_vma_destroy(vma); |
||
2527 | |||
2528 | /* Since the unbound list is global, only move to that list if |
||
2529 | * no more VMAs exist. |
||
2530 | * NB: Until we have real VMAs there will only ever be one */ |
||
2531 | WARN_ON(!list_empty(&obj->vma_list)); |
||
2532 | if (list_empty(&obj->vma_list)) |
||
2533 | list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
||
2534 | |||
2344 | Serge | 2535 | return 0; |
2536 | } |
||
2332 | Serge | 2537 | |
4104 | Serge | 2538 | /** |
2539 | * Unbinds an object from the global GTT aperture. |
||
2540 | */ |
||
2541 | int |
||
2542 | i915_gem_object_ggtt_unbind(struct drm_i915_gem_object *obj) |
||
2543 | { |
||
2544 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
||
2545 | struct i915_address_space *ggtt = &dev_priv->gtt.base; |
||
2546 | |||
2547 | if (!i915_gem_obj_ggtt_bound(obj)) |
||
2548 | return 0; |
||
2549 | |||
2550 | if (obj->pin_count) |
||
2551 | return -EBUSY; |
||
2552 | |||
2553 | BUG_ON(obj->pages == NULL); |
||
2554 | |||
2555 | return i915_vma_unbind(i915_gem_obj_to_vma(obj, ggtt)); |
||
2556 | } |
||
2557 | |||
3031 | serge | 2558 | int i915_gpu_idle(struct drm_device *dev) |
2344 | Serge | 2559 | { |
2560 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
3031 | serge | 2561 | struct intel_ring_buffer *ring; |
2344 | Serge | 2562 | int ret, i; |
2332 | Serge | 2563 | |
2344 | Serge | 2564 | /* Flush everything onto the inactive list. */ |
3031 | serge | 2565 | for_each_ring(ring, dev_priv, i) { |
2566 | ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID); |
||
2344 | Serge | 2567 | if (ret) |
2568 | return ret; |
||
3031 | serge | 2569 | |
3243 | Serge | 2570 | ret = intel_ring_idle(ring); |
3031 | serge | 2571 | if (ret) |
2572 | return ret; |
||
2344 | Serge | 2573 | } |
2332 | Serge | 2574 | |
2344 | Serge | 2575 | return 0; |
2576 | } |
||
2332 | Serge | 2577 | |
3480 | Serge | 2578 | static void i965_write_fence_reg(struct drm_device *dev, int reg, |
3031 | serge | 2579 | struct drm_i915_gem_object *obj) |
2580 | { |
||
2581 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
3480 | Serge | 2582 | int fence_reg; |
2583 | int fence_pitch_shift; |
||
2332 | Serge | 2584 | |
3480 | Serge | 2585 | if (INTEL_INFO(dev)->gen >= 6) { |
2586 | fence_reg = FENCE_REG_SANDYBRIDGE_0; |
||
2587 | fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT; |
||
2588 | } else { |
||
2589 | fence_reg = FENCE_REG_965_0; |
||
2590 | fence_pitch_shift = I965_FENCE_PITCH_SHIFT; |
||
2591 | } |
||
2332 | Serge | 2592 | |
4104 | Serge | 2593 | fence_reg += reg * 8; |
2594 | |||
2595 | /* To w/a incoherency with non-atomic 64-bit register updates, |
||
2596 | * we split the 64-bit update into two 32-bit writes. In order |
||
2597 | * for a partial fence not to be evaluated between writes, we |
||
2598 | * precede the update with write to turn off the fence register, |
||
2599 | * and only enable the fence as the last step. |
||
2600 | * |
||
2601 | * For extra levels of paranoia, we make sure each step lands |
||
2602 | * before applying the next step. |
||
2603 | */ |
||
2604 | I915_WRITE(fence_reg, 0); |
||
2605 | POSTING_READ(fence_reg); |
||
2606 | |||
3031 | serge | 2607 | if (obj) { |
4104 | Serge | 2608 | u32 size = i915_gem_obj_ggtt_size(obj); |
2609 | uint64_t val; |
||
2332 | Serge | 2610 | |
4104 | Serge | 2611 | val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) & |
3031 | serge | 2612 | 0xfffff000) << 32; |
4104 | Serge | 2613 | val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000; |
3480 | Serge | 2614 | val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift; |
3031 | serge | 2615 | if (obj->tiling_mode == I915_TILING_Y) |
2616 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; |
||
2617 | val |= I965_FENCE_REG_VALID; |
||
2332 | Serge | 2618 | |
4104 | Serge | 2619 | I915_WRITE(fence_reg + 4, val >> 32); |
2620 | POSTING_READ(fence_reg + 4); |
||
2621 | |||
2622 | I915_WRITE(fence_reg + 0, val); |
||
4280 | Serge | 2623 | |
2624 | dbgprintf("%s val %x%x\n",__FUNCTION__, (int)(val >> 32), (int)val); |
||
2625 | |||
3480 | Serge | 2626 | POSTING_READ(fence_reg); |
4104 | Serge | 2627 | } else { |
2628 | I915_WRITE(fence_reg + 4, 0); |
||
2629 | POSTING_READ(fence_reg + 4); |
||
2630 | } |
||
3031 | serge | 2631 | } |
2332 | Serge | 2632 | |
3031 | serge | 2633 | static void i915_write_fence_reg(struct drm_device *dev, int reg, |
2634 | struct drm_i915_gem_object *obj) |
||
2635 | { |
||
2636 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
2637 | u32 val; |
||
2332 | Serge | 2638 | |
3031 | serge | 2639 | if (obj) { |
4104 | Serge | 2640 | u32 size = i915_gem_obj_ggtt_size(obj); |
3031 | serge | 2641 | int pitch_val; |
2642 | int tile_width; |
||
2332 | Serge | 2643 | |
4104 | Serge | 2644 | WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) || |
3031 | serge | 2645 | (size & -size) != size || |
4104 | Serge | 2646 | (i915_gem_obj_ggtt_offset(obj) & (size - 1)), |
2647 | "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n", |
||
2648 | i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size); |
||
2332 | Serge | 2649 | |
3031 | serge | 2650 | if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)) |
2651 | tile_width = 128; |
||
2652 | else |
||
2653 | tile_width = 512; |
||
2332 | Serge | 2654 | |
3031 | serge | 2655 | /* Note: pitch better be a power of two tile widths */ |
2656 | pitch_val = obj->stride / tile_width; |
||
2657 | pitch_val = ffs(pitch_val) - 1; |
||
2332 | Serge | 2658 | |
4104 | Serge | 2659 | val = i915_gem_obj_ggtt_offset(obj); |
3031 | serge | 2660 | if (obj->tiling_mode == I915_TILING_Y) |
2661 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; |
||
2662 | val |= I915_FENCE_SIZE_BITS(size); |
||
2663 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; |
||
2664 | val |= I830_FENCE_REG_VALID; |
||
2665 | } else |
||
2666 | val = 0; |
||
2332 | Serge | 2667 | |
3031 | serge | 2668 | if (reg < 8) |
2669 | reg = FENCE_REG_830_0 + reg * 4; |
||
2670 | else |
||
2671 | reg = FENCE_REG_945_8 + (reg - 8) * 4; |
||
2332 | Serge | 2672 | |
3031 | serge | 2673 | I915_WRITE(reg, val); |
2674 | POSTING_READ(reg); |
||
2675 | } |
||
2332 | Serge | 2676 | |
3031 | serge | 2677 | static void i830_write_fence_reg(struct drm_device *dev, int reg, |
2678 | struct drm_i915_gem_object *obj) |
||
2679 | { |
||
2680 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
2681 | uint32_t val; |
||
2344 | Serge | 2682 | |
3031 | serge | 2683 | if (obj) { |
4104 | Serge | 2684 | u32 size = i915_gem_obj_ggtt_size(obj); |
3031 | serge | 2685 | uint32_t pitch_val; |
2344 | Serge | 2686 | |
4104 | Serge | 2687 | WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) || |
3031 | serge | 2688 | (size & -size) != size || |
4104 | Serge | 2689 | (i915_gem_obj_ggtt_offset(obj) & (size - 1)), |
2690 | "object 0x%08lx not 512K or pot-size 0x%08x aligned\n", |
||
2691 | i915_gem_obj_ggtt_offset(obj), size); |
||
2344 | Serge | 2692 | |
3031 | serge | 2693 | pitch_val = obj->stride / 128; |
2694 | pitch_val = ffs(pitch_val) - 1; |
||
2344 | Serge | 2695 | |
4104 | Serge | 2696 | val = i915_gem_obj_ggtt_offset(obj); |
3031 | serge | 2697 | if (obj->tiling_mode == I915_TILING_Y) |
2698 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; |
||
2699 | val |= I830_FENCE_SIZE_BITS(size); |
||
2700 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; |
||
2701 | val |= I830_FENCE_REG_VALID; |
||
2702 | } else |
||
2703 | val = 0; |
||
2704 | |||
2705 | I915_WRITE(FENCE_REG_830_0 + reg * 4, val); |
||
2706 | POSTING_READ(FENCE_REG_830_0 + reg * 4); |
||
2707 | } |
||
2708 | |||
3480 | Serge | 2709 | inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj) |
2710 | { |
||
2711 | return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT; |
||
2712 | } |
||
2713 | |||
3031 | serge | 2714 | static void i915_gem_write_fence(struct drm_device *dev, int reg, |
2715 | struct drm_i915_gem_object *obj) |
||
2332 | Serge | 2716 | { |
3480 | Serge | 2717 | struct drm_i915_private *dev_priv = dev->dev_private; |
2718 | |||
2719 | /* Ensure that all CPU reads are completed before installing a fence |
||
2720 | * and all writes before removing the fence. |
||
2721 | */ |
||
2722 | if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj)) |
||
2723 | mb(); |
||
2724 | |||
4104 | Serge | 2725 | WARN(obj && (!obj->stride || !obj->tiling_mode), |
2726 | "bogus fence setup with stride: 0x%x, tiling mode: %i\n", |
||
2727 | obj->stride, obj->tiling_mode); |
||
2728 | |||
3031 | serge | 2729 | switch (INTEL_INFO(dev)->gen) { |
2730 | case 7: |
||
3480 | Serge | 2731 | case 6: |
3031 | serge | 2732 | case 5: |
2733 | case 4: i965_write_fence_reg(dev, reg, obj); break; |
||
2734 | case 3: i915_write_fence_reg(dev, reg, obj); break; |
||
2735 | case 2: i830_write_fence_reg(dev, reg, obj); break; |
||
3480 | Serge | 2736 | default: BUG(); |
3031 | serge | 2737 | } |
3480 | Serge | 2738 | |
2739 | /* And similarly be paranoid that no direct access to this region |
||
2740 | * is reordered to before the fence is installed. |
||
2741 | */ |
||
2742 | if (i915_gem_object_needs_mb(obj)) |
||
2743 | mb(); |
||
2344 | Serge | 2744 | } |
2745 | |||
3031 | serge | 2746 | static inline int fence_number(struct drm_i915_private *dev_priv, |
2747 | struct drm_i915_fence_reg *fence) |
||
2344 | Serge | 2748 | { |
3031 | serge | 2749 | return fence - dev_priv->fence_regs; |
2750 | } |
||
2332 | Serge | 2751 | |
3031 | serge | 2752 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, |
2753 | struct drm_i915_fence_reg *fence, |
||
2754 | bool enable) |
||
2755 | { |
||
4104 | Serge | 2756 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
2757 | int reg = fence_number(dev_priv, fence); |
||
2332 | Serge | 2758 | |
4104 | Serge | 2759 | i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL); |
3031 | serge | 2760 | |
2761 | if (enable) { |
||
4104 | Serge | 2762 | obj->fence_reg = reg; |
3031 | serge | 2763 | fence->obj = obj; |
2764 | list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list); |
||
2765 | } else { |
||
2766 | obj->fence_reg = I915_FENCE_REG_NONE; |
||
2767 | fence->obj = NULL; |
||
2768 | list_del_init(&fence->lru_list); |
||
2344 | Serge | 2769 | } |
4104 | Serge | 2770 | obj->fence_dirty = false; |
3031 | serge | 2771 | } |
2344 | Serge | 2772 | |
3031 | serge | 2773 | static int |
3480 | Serge | 2774 | i915_gem_object_wait_fence(struct drm_i915_gem_object *obj) |
3031 | serge | 2775 | { |
2776 | if (obj->last_fenced_seqno) { |
||
2777 | int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno); |
||
2352 | Serge | 2778 | if (ret) |
2779 | return ret; |
||
2344 | Serge | 2780 | |
2781 | obj->last_fenced_seqno = 0; |
||
2782 | } |
||
2783 | |||
3031 | serge | 2784 | obj->fenced_gpu_access = false; |
2332 | Serge | 2785 | return 0; |
2786 | } |
||
2787 | |||
2788 | int |
||
2344 | Serge | 2789 | i915_gem_object_put_fence(struct drm_i915_gem_object *obj) |
2332 | Serge | 2790 | { |
3031 | serge | 2791 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
3746 | Serge | 2792 | struct drm_i915_fence_reg *fence; |
2332 | Serge | 2793 | int ret; |
2794 | |||
3480 | Serge | 2795 | ret = i915_gem_object_wait_fence(obj); |
2332 | Serge | 2796 | if (ret) |
2797 | return ret; |
||
2798 | |||
3031 | serge | 2799 | if (obj->fence_reg == I915_FENCE_REG_NONE) |
2800 | return 0; |
||
2332 | Serge | 2801 | |
3746 | Serge | 2802 | fence = &dev_priv->fence_regs[obj->fence_reg]; |
2803 | |||
3031 | serge | 2804 | i915_gem_object_fence_lost(obj); |
3746 | Serge | 2805 | i915_gem_object_update_fence(obj, fence, false); |
2344 | Serge | 2806 | |
2332 | Serge | 2807 | return 0; |
2808 | } |
||
2809 | |||
3031 | serge | 2810 | static struct drm_i915_fence_reg * |
2811 | i915_find_fence_reg(struct drm_device *dev) |
||
2812 | { |
||
2813 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
2814 | struct drm_i915_fence_reg *reg, *avail; |
||
2815 | int i; |
||
2332 | Serge | 2816 | |
3031 | serge | 2817 | /* First try to find a free reg */ |
2818 | avail = NULL; |
||
2819 | for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) { |
||
2820 | reg = &dev_priv->fence_regs[i]; |
||
2821 | if (!reg->obj) |
||
2822 | return reg; |
||
2332 | Serge | 2823 | |
3031 | serge | 2824 | if (!reg->pin_count) |
2825 | avail = reg; |
||
2826 | } |
||
2332 | Serge | 2827 | |
3031 | serge | 2828 | if (avail == NULL) |
2829 | return NULL; |
||
2332 | Serge | 2830 | |
3031 | serge | 2831 | /* None available, try to steal one or wait for a user to finish */ |
2832 | list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) { |
||
2833 | if (reg->pin_count) |
||
2834 | continue; |
||
2332 | Serge | 2835 | |
3031 | serge | 2836 | return reg; |
2837 | } |
||
2332 | Serge | 2838 | |
3031 | serge | 2839 | return NULL; |
2840 | } |
||
2332 | Serge | 2841 | |
3031 | serge | 2842 | /** |
2843 | * i915_gem_object_get_fence - set up fencing for an object |
||
2844 | * @obj: object to map through a fence reg |
||
2845 | * |
||
2846 | * When mapping objects through the GTT, userspace wants to be able to write |
||
2847 | * to them without having to worry about swizzling if the object is tiled. |
||
2848 | * This function walks the fence regs looking for a free one for @obj, |
||
2849 | * stealing one if it can't find any. |
||
2850 | * |
||
2851 | * It then sets up the reg based on the object's properties: address, pitch |
||
2852 | * and tiling format. |
||
2853 | * |
||
2854 | * For an untiled surface, this removes any existing fence. |
||
2855 | */ |
||
2856 | int |
||
2857 | i915_gem_object_get_fence(struct drm_i915_gem_object *obj) |
||
2858 | { |
||
2859 | struct drm_device *dev = obj->base.dev; |
||
2860 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
2861 | bool enable = obj->tiling_mode != I915_TILING_NONE; |
||
2862 | struct drm_i915_fence_reg *reg; |
||
2863 | int ret; |
||
2332 | Serge | 2864 | |
3031 | serge | 2865 | /* Have we updated the tiling parameters upon the object and so |
2866 | * will need to serialise the write to the associated fence register? |
||
2867 | */ |
||
2868 | if (obj->fence_dirty) { |
||
3480 | Serge | 2869 | ret = i915_gem_object_wait_fence(obj); |
3031 | serge | 2870 | if (ret) |
2871 | return ret; |
||
2872 | } |
||
2332 | Serge | 2873 | |
3031 | serge | 2874 | /* Just update our place in the LRU if our fence is getting reused. */ |
2875 | if (obj->fence_reg != I915_FENCE_REG_NONE) { |
||
2876 | reg = &dev_priv->fence_regs[obj->fence_reg]; |
||
2877 | if (!obj->fence_dirty) { |
||
2878 | list_move_tail(®->lru_list, |
||
2879 | &dev_priv->mm.fence_list); |
||
2880 | return 0; |
||
2881 | } |
||
2882 | } else if (enable) { |
||
2883 | reg = i915_find_fence_reg(dev); |
||
2884 | if (reg == NULL) |
||
2885 | return -EDEADLK; |
||
2332 | Serge | 2886 | |
3031 | serge | 2887 | if (reg->obj) { |
2888 | struct drm_i915_gem_object *old = reg->obj; |
||
2332 | Serge | 2889 | |
3480 | Serge | 2890 | ret = i915_gem_object_wait_fence(old); |
3031 | serge | 2891 | if (ret) |
2892 | return ret; |
||
2332 | Serge | 2893 | |
3031 | serge | 2894 | i915_gem_object_fence_lost(old); |
2895 | } |
||
2896 | } else |
||
2897 | return 0; |
||
2332 | Serge | 2898 | |
3031 | serge | 2899 | i915_gem_object_update_fence(obj, reg, enable); |
2332 | Serge | 2900 | |
3031 | serge | 2901 | return 0; |
2902 | } |
||
2332 | Serge | 2903 | |
3031 | serge | 2904 | static bool i915_gem_valid_gtt_space(struct drm_device *dev, |
2905 | struct drm_mm_node *gtt_space, |
||
2906 | unsigned long cache_level) |
||
2907 | { |
||
2908 | struct drm_mm_node *other; |
||
2332 | Serge | 2909 | |
3031 | serge | 2910 | /* On non-LLC machines we have to be careful when putting differing |
2911 | * types of snoopable memory together to avoid the prefetcher |
||
3480 | Serge | 2912 | * crossing memory domains and dying. |
3031 | serge | 2913 | */ |
2914 | if (HAS_LLC(dev)) |
||
2915 | return true; |
||
2332 | Serge | 2916 | |
4104 | Serge | 2917 | if (!drm_mm_node_allocated(gtt_space)) |
3031 | serge | 2918 | return true; |
2332 | Serge | 2919 | |
3031 | serge | 2920 | if (list_empty(>t_space->node_list)) |
2921 | return true; |
||
2332 | Serge | 2922 | |
3031 | serge | 2923 | other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list); |
2924 | if (other->allocated && !other->hole_follows && other->color != cache_level) |
||
2925 | return false; |
||
2344 | Serge | 2926 | |
3031 | serge | 2927 | other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list); |
2928 | if (other->allocated && !gtt_space->hole_follows && other->color != cache_level) |
||
2929 | return false; |
||
2344 | Serge | 2930 | |
3031 | serge | 2931 | return true; |
2932 | } |
||
2344 | Serge | 2933 | |
3031 | serge | 2934 | static void i915_gem_verify_gtt(struct drm_device *dev) |
2935 | { |
||
2936 | #if WATCH_GTT |
||
2937 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
2938 | struct drm_i915_gem_object *obj; |
||
2939 | int err = 0; |
||
2344 | Serge | 2940 | |
4104 | Serge | 2941 | list_for_each_entry(obj, &dev_priv->mm.gtt_list, global_list) { |
3031 | serge | 2942 | if (obj->gtt_space == NULL) { |
2943 | printk(KERN_ERR "object found on GTT list with no space reserved\n"); |
||
2944 | err++; |
||
2945 | continue; |
||
2946 | } |
||
2344 | Serge | 2947 | |
3031 | serge | 2948 | if (obj->cache_level != obj->gtt_space->color) { |
2949 | printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n", |
||
4104 | Serge | 2950 | i915_gem_obj_ggtt_offset(obj), |
2951 | i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj), |
||
3031 | serge | 2952 | obj->cache_level, |
2953 | obj->gtt_space->color); |
||
2954 | err++; |
||
2955 | continue; |
||
2956 | } |
||
2344 | Serge | 2957 | |
3031 | serge | 2958 | if (!i915_gem_valid_gtt_space(dev, |
2959 | obj->gtt_space, |
||
2960 | obj->cache_level)) { |
||
2961 | printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n", |
||
4104 | Serge | 2962 | i915_gem_obj_ggtt_offset(obj), |
2963 | i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj), |
||
3031 | serge | 2964 | obj->cache_level); |
2965 | err++; |
||
2966 | continue; |
||
2967 | } |
||
2968 | } |
||
2344 | Serge | 2969 | |
3031 | serge | 2970 | WARN_ON(err); |
2971 | #endif |
||
2326 | Serge | 2972 | } |
2973 | |||
2332 | Serge | 2974 | /** |
2975 | * Finds free space in the GTT aperture and binds the object there. |
||
2976 | */ |
||
2977 | static int |
||
4104 | Serge | 2978 | i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, |
2979 | struct i915_address_space *vm, |
||
2332 | Serge | 2980 | unsigned alignment, |
3031 | serge | 2981 | bool map_and_fenceable, |
2982 | bool nonblocking) |
||
2332 | Serge | 2983 | { |
2984 | struct drm_device *dev = obj->base.dev; |
||
2985 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
2986 | u32 size, fence_size, fence_alignment, unfenced_alignment; |
||
4104 | Serge | 2987 | size_t gtt_max = |
2988 | map_and_fenceable ? dev_priv->gtt.mappable_end : vm->total; |
||
2989 | struct i915_vma *vma; |
||
2332 | Serge | 2990 | int ret; |
2326 | Serge | 2991 | |
2332 | Serge | 2992 | fence_size = i915_gem_get_gtt_size(dev, |
2993 | obj->base.size, |
||
2994 | obj->tiling_mode); |
||
2995 | fence_alignment = i915_gem_get_gtt_alignment(dev, |
||
2996 | obj->base.size, |
||
3480 | Serge | 2997 | obj->tiling_mode, true); |
2332 | Serge | 2998 | unfenced_alignment = |
3480 | Serge | 2999 | i915_gem_get_gtt_alignment(dev, |
2332 | Serge | 3000 | obj->base.size, |
3480 | Serge | 3001 | obj->tiling_mode, false); |
2332 | Serge | 3002 | |
3003 | if (alignment == 0) |
||
3004 | alignment = map_and_fenceable ? fence_alignment : |
||
3005 | unfenced_alignment; |
||
3006 | if (map_and_fenceable && alignment & (fence_alignment - 1)) { |
||
3007 | DRM_ERROR("Invalid object alignment requested %u\n", alignment); |
||
3008 | return -EINVAL; |
||
3009 | } |
||
3010 | |||
3011 | size = map_and_fenceable ? fence_size : obj->base.size; |
||
3012 | |||
3013 | /* If the object is bigger than the entire aperture, reject it early |
||
3014 | * before evicting everything in a vain attempt to find space. |
||
3015 | */ |
||
4104 | Serge | 3016 | if (obj->base.size > gtt_max) { |
3017 | DRM_ERROR("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%zu\n", |
||
3018 | obj->base.size, |
||
3019 | map_and_fenceable ? "mappable" : "total", |
||
3020 | gtt_max); |
||
2332 | Serge | 3021 | return -E2BIG; |
3022 | } |
||
3023 | |||
3031 | serge | 3024 | ret = i915_gem_object_get_pages(obj); |
3025 | if (ret) |
||
3026 | return ret; |
||
3027 | |||
3243 | Serge | 3028 | i915_gem_object_pin_pages(obj); |
3029 | |||
4104 | Serge | 3030 | BUG_ON(!i915_is_ggtt(vm)); |
3031 | |||
3032 | vma = i915_gem_obj_lookup_or_create_vma(obj, vm); |
||
3033 | if (IS_ERR(vma)) { |
||
3034 | ret = PTR_ERR(vma); |
||
3035 | goto err_unpin; |
||
3243 | Serge | 3036 | } |
3037 | |||
4104 | Serge | 3038 | /* For now we only ever use 1 vma per object */ |
3039 | WARN_ON(!list_is_singular(&obj->vma_list)); |
||
3040 | |||
3041 | search_free: |
||
3042 | ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node, |
||
3043 | size, alignment, |
||
3044 | obj->cache_level, 0, gtt_max, |
||
3045 | DRM_MM_SEARCH_DEFAULT); |
||
3243 | Serge | 3046 | if (ret) { |
2332 | Serge | 3047 | |
4104 | Serge | 3048 | goto err_free_vma; |
2332 | Serge | 3049 | } |
4104 | Serge | 3050 | if (WARN_ON(!i915_gem_valid_gtt_space(dev, &vma->node, |
3051 | obj->cache_level))) { |
||
3052 | ret = -EINVAL; |
||
3053 | goto err_remove_node; |
||
3031 | serge | 3054 | } |
2332 | Serge | 3055 | |
3031 | serge | 3056 | ret = i915_gem_gtt_prepare_object(obj); |
4104 | Serge | 3057 | if (ret) |
3058 | goto err_remove_node; |
||
2332 | Serge | 3059 | |
4104 | Serge | 3060 | list_move_tail(&obj->global_list, &dev_priv->mm.bound_list); |
3061 | list_add_tail(&vma->mm_list, &vm->inactive_list); |
||
2332 | Serge | 3062 | |
4104 | Serge | 3063 | if (i915_is_ggtt(vm)) { |
3064 | bool mappable, fenceable; |
||
2332 | Serge | 3065 | |
4104 | Serge | 3066 | fenceable = (vma->node.size == fence_size && |
3067 | (vma->node.start & (fence_alignment - 1)) == 0); |
||
2332 | Serge | 3068 | |
4104 | Serge | 3069 | mappable = (vma->node.start + obj->base.size <= |
3070 | dev_priv->gtt.mappable_end); |
||
2332 | Serge | 3071 | |
3072 | obj->map_and_fenceable = mappable && fenceable; |
||
4104 | Serge | 3073 | } |
2332 | Serge | 3074 | |
4104 | Serge | 3075 | WARN_ON(map_and_fenceable && !obj->map_and_fenceable); |
3076 | |||
3077 | trace_i915_vma_bind(vma, map_and_fenceable); |
||
3031 | serge | 3078 | i915_gem_verify_gtt(dev); |
2332 | Serge | 3079 | return 0; |
4104 | Serge | 3080 | |
3081 | err_remove_node: |
||
3082 | drm_mm_remove_node(&vma->node); |
||
3083 | err_free_vma: |
||
3084 | i915_gem_vma_destroy(vma); |
||
3085 | err_unpin: |
||
3086 | i915_gem_object_unpin_pages(obj); |
||
3087 | return ret; |
||
2332 | Serge | 3088 | } |
3089 | |||
4104 | Serge | 3090 | bool |
3091 | i915_gem_clflush_object(struct drm_i915_gem_object *obj, |
||
3092 | bool force) |
||
2332 | Serge | 3093 | { |
3094 | /* If we don't have a page list set up, then we're not pinned |
||
3095 | * to GPU, and we can ignore the cache flush because it'll happen |
||
3096 | * again at bind time. |
||
3097 | */ |
||
3243 | Serge | 3098 | if (obj->pages == NULL) |
4104 | Serge | 3099 | return false; |
2332 | Serge | 3100 | |
3480 | Serge | 3101 | /* |
3102 | * Stolen memory is always coherent with the GPU as it is explicitly |
||
3103 | * marked as wc by the system, or the system is cache-coherent. |
||
3104 | */ |
||
3105 | if (obj->stolen) |
||
4104 | Serge | 3106 | return false; |
3480 | Serge | 3107 | |
2332 | Serge | 3108 | /* If the GPU is snooping the contents of the CPU cache, |
3109 | * we do not need to manually clear the CPU cache lines. However, |
||
3110 | * the caches are only snooped when the render cache is |
||
3111 | * flushed/invalidated. As we always have to emit invalidations |
||
3112 | * and flushes when moving into and out of the RENDER domain, correct |
||
3113 | * snooping behaviour occurs naturally as the result of our domain |
||
3114 | * tracking. |
||
3115 | */ |
||
4104 | Serge | 3116 | if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) |
3117 | return false; |
||
3243 | Serge | 3118 | #if 0 |
2344 | Serge | 3119 | if(obj->mapped != NULL) |
3120 | { |
||
3121 | uint8_t *page_virtual; |
||
3122 | unsigned int i; |
||
2332 | Serge | 3123 | |
2344 | Serge | 3124 | page_virtual = obj->mapped; |
3125 | asm volatile("mfence"); |
||
3126 | for (i = 0; i < obj->base.size; i += x86_clflush_size) |
||
3127 | clflush(page_virtual + i); |
||
3128 | asm volatile("mfence"); |
||
3129 | } |
||
3130 | else |
||
3131 | { |
||
3132 | uint8_t *page_virtual; |
||
3133 | unsigned int i; |
||
3134 | page_virtual = AllocKernelSpace(obj->base.size); |
||
3135 | if(page_virtual != NULL) |
||
3136 | { |
||
3031 | serge | 3137 | dma_addr_t *src, *dst; |
2344 | Serge | 3138 | u32 count; |
3139 | |||
3140 | #define page_tabs 0xFDC00000 /* really dirty hack */ |
||
3141 | |||
3031 | serge | 3142 | src = obj->pages.page; |
3143 | dst = &((dma_addr_t*)page_tabs)[(u32_t)page_virtual >> 12]; |
||
2344 | Serge | 3144 | count = obj->base.size/4096; |
3145 | |||
3146 | while(count--) |
||
3147 | { |
||
3148 | *dst++ = (0xFFFFF000 & *src++) | 0x001 ; |
||
3149 | }; |
||
3150 | |||
3151 | asm volatile("mfence"); |
||
3152 | for (i = 0; i < obj->base.size; i += x86_clflush_size) |
||
3153 | clflush(page_virtual + i); |
||
3154 | asm volatile("mfence"); |
||
3155 | FreeKernelSpace(page_virtual); |
||
3156 | } |
||
3157 | else |
||
3158 | { |
||
3159 | asm volatile ( |
||
3160 | "mfence \n" |
||
3161 | "wbinvd \n" /* this is really ugly */ |
||
3162 | "mfence"); |
||
3163 | } |
||
3164 | } |
||
3243 | Serge | 3165 | #endif |
3166 | |||
4104 | Serge | 3167 | return true; |
2332 | Serge | 3168 | } |
3169 | |||
2344 | Serge | 3170 | /** Flushes the GTT write domain for the object if it's dirty. */ |
3171 | static void |
||
3172 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
||
3173 | { |
||
3174 | uint32_t old_write_domain; |
||
2332 | Serge | 3175 | |
2344 | Serge | 3176 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
3177 | return; |
||
2332 | Serge | 3178 | |
2344 | Serge | 3179 | /* No actual flushing is required for the GTT write domain. Writes |
3180 | * to it immediately go to main memory as far as we know, so there's |
||
3181 | * no chipset flush. It also doesn't land in render cache. |
||
3182 | * |
||
3183 | * However, we do have to enforce the order so that all writes through |
||
3184 | * the GTT land before any writes to the device, such as updates to |
||
3185 | * the GATT itself. |
||
3186 | */ |
||
3187 | wmb(); |
||
2332 | Serge | 3188 | |
2344 | Serge | 3189 | old_write_domain = obj->base.write_domain; |
3190 | obj->base.write_domain = 0; |
||
2332 | Serge | 3191 | |
2351 | Serge | 3192 | trace_i915_gem_object_change_domain(obj, |
3193 | obj->base.read_domains, |
||
3194 | old_write_domain); |
||
2344 | Serge | 3195 | } |
2332 | Serge | 3196 | |
3197 | /** Flushes the CPU write domain for the object if it's dirty. */ |
||
2326 | Serge | 3198 | static void |
4104 | Serge | 3199 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj, |
3200 | bool force) |
||
2332 | Serge | 3201 | { |
3202 | uint32_t old_write_domain; |
||
3203 | |||
3204 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
||
3205 | return; |
||
3206 | |||
4104 | Serge | 3207 | if (i915_gem_clflush_object(obj, force)) |
3243 | Serge | 3208 | i915_gem_chipset_flush(obj->base.dev); |
4104 | Serge | 3209 | |
2332 | Serge | 3210 | old_write_domain = obj->base.write_domain; |
3211 | obj->base.write_domain = 0; |
||
3212 | |||
2351 | Serge | 3213 | trace_i915_gem_object_change_domain(obj, |
3214 | obj->base.read_domains, |
||
3215 | old_write_domain); |
||
2332 | Serge | 3216 | } |
3217 | |||
3218 | /** |
||
3219 | * Moves a single object to the GTT read, and possibly write domain. |
||
3220 | * |
||
3221 | * This function returns when the move is complete, including waiting on |
||
3222 | * flushes to occur. |
||
3223 | */ |
||
3224 | int |
||
3225 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
||
3226 | { |
||
3031 | serge | 3227 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
2332 | Serge | 3228 | uint32_t old_write_domain, old_read_domains; |
3229 | int ret; |
||
3230 | |||
3231 | /* Not valid to be called on unbound objects. */ |
||
4104 | Serge | 3232 | if (!i915_gem_obj_bound_any(obj)) |
2332 | Serge | 3233 | return -EINVAL; |
3234 | |||
3235 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
||
3236 | return 0; |
||
3237 | |||
3031 | serge | 3238 | ret = i915_gem_object_wait_rendering(obj, !write); |
2332 | Serge | 3239 | if (ret) |
3240 | return ret; |
||
3241 | |||
4104 | Serge | 3242 | i915_gem_object_flush_cpu_write_domain(obj, false); |
2332 | Serge | 3243 | |
3480 | Serge | 3244 | /* Serialise direct access to this object with the barriers for |
3245 | * coherent writes from the GPU, by effectively invalidating the |
||
3246 | * GTT domain upon first access. |
||
3247 | */ |
||
3248 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
||
3249 | mb(); |
||
3250 | |||
2332 | Serge | 3251 | old_write_domain = obj->base.write_domain; |
3252 | old_read_domains = obj->base.read_domains; |
||
3253 | |||
3254 | /* It should now be out of any other write domains, and we can update |
||
3255 | * the domain values for our changes. |
||
3256 | */ |
||
3257 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
||
3258 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
||
3259 | if (write) { |
||
3260 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
||
3261 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; |
||
3262 | obj->dirty = 1; |
||
3263 | } |
||
3264 | |||
2351 | Serge | 3265 | trace_i915_gem_object_change_domain(obj, |
3266 | old_read_domains, |
||
3267 | old_write_domain); |
||
3268 | |||
3031 | serge | 3269 | /* And bump the LRU for this access */ |
4104 | Serge | 3270 | if (i915_gem_object_is_inactive(obj)) { |
3271 | struct i915_vma *vma = i915_gem_obj_to_vma(obj, |
||
3272 | &dev_priv->gtt.base); |
||
3273 | if (vma) |
||
3274 | list_move_tail(&vma->mm_list, |
||
3275 | &dev_priv->gtt.base.inactive_list); |
||
3031 | serge | 3276 | |
4104 | Serge | 3277 | } |
3278 | |||
2332 | Serge | 3279 | return 0; |
3280 | } |
||
3281 | |||
2335 | Serge | 3282 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
3283 | enum i915_cache_level cache_level) |
||
3284 | { |
||
3031 | serge | 3285 | struct drm_device *dev = obj->base.dev; |
3286 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
4104 | Serge | 3287 | struct i915_vma *vma; |
2335 | Serge | 3288 | int ret; |
2332 | Serge | 3289 | |
2335 | Serge | 3290 | if (obj->cache_level == cache_level) |
3291 | return 0; |
||
2332 | Serge | 3292 | |
2335 | Serge | 3293 | if (obj->pin_count) { |
3294 | DRM_DEBUG("can not change the cache level of pinned objects\n"); |
||
3295 | return -EBUSY; |
||
3296 | } |
||
2332 | Serge | 3297 | |
4104 | Serge | 3298 | list_for_each_entry(vma, &obj->vma_list, vma_link) { |
3299 | if (!i915_gem_valid_gtt_space(dev, &vma->node, cache_level)) { |
||
3300 | ret = i915_vma_unbind(vma); |
||
3031 | serge | 3301 | if (ret) |
3302 | return ret; |
||
4104 | Serge | 3303 | |
3304 | break; |
||
3305 | } |
||
3031 | serge | 3306 | } |
3307 | |||
4104 | Serge | 3308 | if (i915_gem_obj_bound_any(obj)) { |
2335 | Serge | 3309 | ret = i915_gem_object_finish_gpu(obj); |
3310 | if (ret) |
||
3311 | return ret; |
||
2332 | Serge | 3312 | |
2335 | Serge | 3313 | i915_gem_object_finish_gtt(obj); |
2332 | Serge | 3314 | |
2335 | Serge | 3315 | /* Before SandyBridge, you could not use tiling or fence |
3316 | * registers with snooped memory, so relinquish any fences |
||
3317 | * currently pointing to our region in the aperture. |
||
3318 | */ |
||
3031 | serge | 3319 | if (INTEL_INFO(dev)->gen < 6) { |
2335 | Serge | 3320 | ret = i915_gem_object_put_fence(obj); |
3321 | if (ret) |
||
3322 | return ret; |
||
3323 | } |
||
2332 | Serge | 3324 | |
3031 | serge | 3325 | if (obj->has_global_gtt_mapping) |
3326 | i915_gem_gtt_bind_object(obj, cache_level); |
||
3327 | if (obj->has_aliasing_ppgtt_mapping) |
||
3328 | i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt, |
||
3329 | obj, cache_level); |
||
2335 | Serge | 3330 | } |
2332 | Serge | 3331 | |
4104 | Serge | 3332 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
3333 | vma->node.color = cache_level; |
||
3334 | obj->cache_level = cache_level; |
||
3335 | |||
3336 | if (cpu_write_needs_clflush(obj)) { |
||
2335 | Serge | 3337 | u32 old_read_domains, old_write_domain; |
2332 | Serge | 3338 | |
2335 | Serge | 3339 | /* If we're coming from LLC cached, then we haven't |
3340 | * actually been tracking whether the data is in the |
||
3341 | * CPU cache or not, since we only allow one bit set |
||
3342 | * in obj->write_domain and have been skipping the clflushes. |
||
3343 | * Just set it to the CPU cache for now. |
||
3344 | */ |
||
3345 | WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU); |
||
2332 | Serge | 3346 | |
2335 | Serge | 3347 | old_read_domains = obj->base.read_domains; |
3348 | old_write_domain = obj->base.write_domain; |
||
2332 | Serge | 3349 | |
2335 | Serge | 3350 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3351 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
||
2332 | Serge | 3352 | |
2351 | Serge | 3353 | trace_i915_gem_object_change_domain(obj, |
3354 | old_read_domains, |
||
3355 | old_write_domain); |
||
2344 | Serge | 3356 | } |
2332 | Serge | 3357 | |
3031 | serge | 3358 | i915_gem_verify_gtt(dev); |
2335 | Serge | 3359 | return 0; |
3360 | } |
||
2332 | Serge | 3361 | |
3260 | Serge | 3362 | int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, |
3363 | struct drm_file *file) |
||
3364 | { |
||
3365 | struct drm_i915_gem_caching *args = data; |
||
3366 | struct drm_i915_gem_object *obj; |
||
3367 | int ret; |
||
3368 | |||
3480 | Serge | 3369 | if(args->handle == -2) |
3370 | { |
||
3371 | printf("%s handle %d\n", __FUNCTION__, args->handle); |
||
3372 | return 0; |
||
3373 | } |
||
3374 | |||
3260 | Serge | 3375 | ret = i915_mutex_lock_interruptible(dev); |
3376 | if (ret) |
||
3377 | return ret; |
||
3378 | |||
3379 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
||
3380 | if (&obj->base == NULL) { |
||
3381 | ret = -ENOENT; |
||
3382 | goto unlock; |
||
3383 | } |
||
3384 | |||
4104 | Serge | 3385 | switch (obj->cache_level) { |
3386 | case I915_CACHE_LLC: |
||
3387 | case I915_CACHE_L3_LLC: |
||
3388 | args->caching = I915_CACHING_CACHED; |
||
3389 | break; |
||
3260 | Serge | 3390 | |
4104 | Serge | 3391 | case I915_CACHE_WT: |
3392 | args->caching = I915_CACHING_DISPLAY; |
||
3393 | break; |
||
3394 | |||
3395 | default: |
||
3396 | args->caching = I915_CACHING_NONE; |
||
3397 | break; |
||
3398 | } |
||
3399 | |||
3260 | Serge | 3400 | drm_gem_object_unreference(&obj->base); |
3401 | unlock: |
||
3402 | mutex_unlock(&dev->struct_mutex); |
||
3403 | return ret; |
||
3404 | } |
||
3405 | |||
3406 | int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data, |
||
3407 | struct drm_file *file) |
||
3408 | { |
||
3409 | struct drm_i915_gem_caching *args = data; |
||
3410 | struct drm_i915_gem_object *obj; |
||
3411 | enum i915_cache_level level; |
||
3412 | int ret; |
||
3413 | |||
3480 | Serge | 3414 | if(args->handle == -2) |
3415 | { |
||
3416 | printf("%s handle %d\n", __FUNCTION__, args->handle); |
||
3417 | return 0; |
||
3418 | } |
||
3419 | |||
3260 | Serge | 3420 | switch (args->caching) { |
3421 | case I915_CACHING_NONE: |
||
3422 | level = I915_CACHE_NONE; |
||
3423 | break; |
||
3424 | case I915_CACHING_CACHED: |
||
3425 | level = I915_CACHE_LLC; |
||
3426 | break; |
||
4104 | Serge | 3427 | case I915_CACHING_DISPLAY: |
3428 | level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE; |
||
3429 | break; |
||
3260 | Serge | 3430 | default: |
3431 | return -EINVAL; |
||
3432 | } |
||
3433 | |||
3434 | ret = i915_mutex_lock_interruptible(dev); |
||
3435 | if (ret) |
||
3436 | return ret; |
||
3437 | |||
3438 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
||
3439 | if (&obj->base == NULL) { |
||
3440 | ret = -ENOENT; |
||
3441 | goto unlock; |
||
3442 | } |
||
3443 | |||
3444 | ret = i915_gem_object_set_cache_level(obj, level); |
||
3445 | |||
3446 | drm_gem_object_unreference(&obj->base); |
||
3447 | unlock: |
||
3448 | mutex_unlock(&dev->struct_mutex); |
||
3449 | return ret; |
||
3450 | } |
||
3451 | |||
4104 | Serge | 3452 | static bool is_pin_display(struct drm_i915_gem_object *obj) |
3453 | { |
||
3454 | /* There are 3 sources that pin objects: |
||
3455 | * 1. The display engine (scanouts, sprites, cursors); |
||
3456 | * 2. Reservations for execbuffer; |
||
3457 | * 3. The user. |
||
3458 | * |
||
3459 | * We can ignore reservations as we hold the struct_mutex and |
||
3460 | * are only called outside of the reservation path. The user |
||
3461 | * can only increment pin_count once, and so if after |
||
3462 | * subtracting the potential reference by the user, any pin_count |
||
3463 | * remains, it must be due to another use by the display engine. |
||
3464 | */ |
||
3465 | return obj->pin_count - !!obj->user_pin_count; |
||
3466 | } |
||
3467 | |||
2335 | Serge | 3468 | /* |
3469 | * Prepare buffer for display plane (scanout, cursors, etc). |
||
3470 | * Can be called from an uninterruptible phase (modesetting) and allows |
||
3471 | * any flushes to be pipelined (for pageflips). |
||
3472 | */ |
||
3473 | int |
||
3474 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
||
3475 | u32 alignment, |
||
3476 | struct intel_ring_buffer *pipelined) |
||
3477 | { |
||
3478 | u32 old_read_domains, old_write_domain; |
||
3479 | int ret; |
||
2332 | Serge | 3480 | |
3031 | serge | 3481 | if (pipelined != obj->ring) { |
3482 | ret = i915_gem_object_sync(obj, pipelined); |
||
2335 | Serge | 3483 | if (ret) |
3484 | return ret; |
||
3485 | } |
||
2332 | Serge | 3486 | |
4104 | Serge | 3487 | /* Mark the pin_display early so that we account for the |
3488 | * display coherency whilst setting up the cache domains. |
||
3489 | */ |
||
3490 | obj->pin_display = true; |
||
3491 | |||
2335 | Serge | 3492 | /* The display engine is not coherent with the LLC cache on gen6. As |
3493 | * a result, we make sure that the pinning that is about to occur is |
||
3494 | * done with uncached PTEs. This is lowest common denominator for all |
||
3495 | * chipsets. |
||
3496 | * |
||
3497 | * However for gen6+, we could do better by using the GFDT bit instead |
||
3498 | * of uncaching, which would allow us to flush all the LLC-cached data |
||
3499 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. |
||
3500 | */ |
||
4104 | Serge | 3501 | ret = i915_gem_object_set_cache_level(obj, |
3502 | HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE); |
||
2360 | Serge | 3503 | if (ret) |
4104 | Serge | 3504 | goto err_unpin_display; |
2332 | Serge | 3505 | |
2335 | Serge | 3506 | /* As the user may map the buffer once pinned in the display plane |
3507 | * (e.g. libkms for the bootup splash), we have to ensure that we |
||
3508 | * always use map_and_fenceable for all scanout buffers. |
||
3509 | */ |
||
4104 | Serge | 3510 | ret = i915_gem_obj_ggtt_pin(obj, alignment, true, false); |
2335 | Serge | 3511 | if (ret) |
4104 | Serge | 3512 | goto err_unpin_display; |
2332 | Serge | 3513 | |
4104 | Serge | 3514 | i915_gem_object_flush_cpu_write_domain(obj, true); |
2332 | Serge | 3515 | |
2335 | Serge | 3516 | old_write_domain = obj->base.write_domain; |
3517 | old_read_domains = obj->base.read_domains; |
||
2332 | Serge | 3518 | |
2335 | Serge | 3519 | /* It should now be out of any other write domains, and we can update |
3520 | * the domain values for our changes. |
||
3521 | */ |
||
3031 | serge | 3522 | obj->base.write_domain = 0; |
2335 | Serge | 3523 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
2332 | Serge | 3524 | |
2351 | Serge | 3525 | trace_i915_gem_object_change_domain(obj, |
3526 | old_read_domains, |
||
3527 | old_write_domain); |
||
2332 | Serge | 3528 | |
2335 | Serge | 3529 | return 0; |
4104 | Serge | 3530 | |
3531 | err_unpin_display: |
||
3532 | obj->pin_display = is_pin_display(obj); |
||
3533 | return ret; |
||
2335 | Serge | 3534 | } |
2332 | Serge | 3535 | |
4104 | Serge | 3536 | void |
3537 | i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj) |
||
3538 | { |
||
3539 | i915_gem_object_unpin(obj); |
||
3540 | obj->pin_display = is_pin_display(obj); |
||
3541 | } |
||
3542 | |||
2344 | Serge | 3543 | int |
3544 | i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj) |
||
3545 | { |
||
3546 | int ret; |
||
2332 | Serge | 3547 | |
2344 | Serge | 3548 | if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0) |
3549 | return 0; |
||
2332 | Serge | 3550 | |
3031 | serge | 3551 | ret = i915_gem_object_wait_rendering(obj, false); |
3243 | Serge | 3552 | if (ret) |
3553 | return ret; |
||
2332 | Serge | 3554 | |
2344 | Serge | 3555 | /* Ensure that we invalidate the GPU's caches and TLBs. */ |
3556 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; |
||
3031 | serge | 3557 | return 0; |
2344 | Serge | 3558 | } |
2332 | Serge | 3559 | |
2344 | Serge | 3560 | /** |
3561 | * Moves a single object to the CPU read, and possibly write domain. |
||
3562 | * |
||
3563 | * This function returns when the move is complete, including waiting on |
||
3564 | * flushes to occur. |
||
3565 | */ |
||
3031 | serge | 3566 | int |
2344 | Serge | 3567 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
3568 | { |
||
3569 | uint32_t old_write_domain, old_read_domains; |
||
3570 | int ret; |
||
2332 | Serge | 3571 | |
2344 | Serge | 3572 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3573 | return 0; |
||
2332 | Serge | 3574 | |
3031 | serge | 3575 | ret = i915_gem_object_wait_rendering(obj, !write); |
2344 | Serge | 3576 | if (ret) |
3577 | return ret; |
||
2332 | Serge | 3578 | |
2344 | Serge | 3579 | i915_gem_object_flush_gtt_write_domain(obj); |
2332 | Serge | 3580 | |
2344 | Serge | 3581 | old_write_domain = obj->base.write_domain; |
3582 | old_read_domains = obj->base.read_domains; |
||
2332 | Serge | 3583 | |
2344 | Serge | 3584 | /* Flush the CPU cache if it's still invalid. */ |
3585 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
||
4104 | Serge | 3586 | i915_gem_clflush_object(obj, false); |
2332 | Serge | 3587 | |
2344 | Serge | 3588 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
3589 | } |
||
2332 | Serge | 3590 | |
2344 | Serge | 3591 | /* It should now be out of any other write domains, and we can update |
3592 | * the domain values for our changes. |
||
3593 | */ |
||
3594 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
||
2332 | Serge | 3595 | |
2344 | Serge | 3596 | /* If we're writing through the CPU, then the GPU read domains will |
3597 | * need to be invalidated at next use. |
||
3598 | */ |
||
3599 | if (write) { |
||
3600 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
||
3601 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
||
3602 | } |
||
2332 | Serge | 3603 | |
2351 | Serge | 3604 | trace_i915_gem_object_change_domain(obj, |
3605 | old_read_domains, |
||
3606 | old_write_domain); |
||
2332 | Serge | 3607 | |
2344 | Serge | 3608 | return 0; |
3609 | } |
||
2332 | Serge | 3610 | |
3031 | serge | 3611 | /* Throttle our rendering by waiting until the ring has completed our requests |
3612 | * emitted over 20 msec ago. |
||
2344 | Serge | 3613 | * |
3031 | serge | 3614 | * Note that if we were to use the current jiffies each time around the loop, |
3615 | * we wouldn't escape the function with any frames outstanding if the time to |
||
3616 | * render a frame was over 20ms. |
||
3617 | * |
||
3618 | * This should get us reasonable parallelism between CPU and GPU but also |
||
3619 | * relatively low latency when blocking on a particular request to finish. |
||
2344 | Serge | 3620 | */ |
3031 | serge | 3621 | static int |
3622 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
||
2344 | Serge | 3623 | { |
3031 | serge | 3624 | struct drm_i915_private *dev_priv = dev->dev_private; |
3625 | struct drm_i915_file_private *file_priv = file->driver_priv; |
||
3263 | Serge | 3626 | unsigned long recent_enough = GetTimerTicks() - msecs_to_jiffies(20); |
3031 | serge | 3627 | struct drm_i915_gem_request *request; |
3628 | struct intel_ring_buffer *ring = NULL; |
||
3480 | Serge | 3629 | unsigned reset_counter; |
3031 | serge | 3630 | u32 seqno = 0; |
3631 | int ret; |
||
2332 | Serge | 3632 | |
3480 | Serge | 3633 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
3634 | if (ret) |
||
3635 | return ret; |
||
2332 | Serge | 3636 | |
3480 | Serge | 3637 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, false); |
3638 | if (ret) |
||
3639 | return ret; |
||
3640 | |||
3031 | serge | 3641 | spin_lock(&file_priv->mm.lock); |
3642 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { |
||
3643 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
||
3644 | break; |
||
2332 | Serge | 3645 | |
3031 | serge | 3646 | ring = request->ring; |
3647 | seqno = request->seqno; |
||
3648 | } |
||
3480 | Serge | 3649 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
3031 | serge | 3650 | spin_unlock(&file_priv->mm.lock); |
2332 | Serge | 3651 | |
3031 | serge | 3652 | if (seqno == 0) |
3653 | return 0; |
||
2332 | Serge | 3654 | |
3480 | Serge | 3655 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL); |
3031 | serge | 3656 | if (ret == 0) |
3657 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); |
||
2332 | Serge | 3658 | |
3031 | serge | 3659 | return ret; |
2352 | Serge | 3660 | } |
2332 | Serge | 3661 | |
3662 | int |
||
3663 | i915_gem_object_pin(struct drm_i915_gem_object *obj, |
||
4104 | Serge | 3664 | struct i915_address_space *vm, |
2332 | Serge | 3665 | uint32_t alignment, |
3031 | serge | 3666 | bool map_and_fenceable, |
3667 | bool nonblocking) |
||
2332 | Serge | 3668 | { |
4104 | Serge | 3669 | struct i915_vma *vma; |
2332 | Serge | 3670 | int ret; |
3671 | |||
3031 | serge | 3672 | if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT)) |
3673 | return -EBUSY; |
||
2332 | Serge | 3674 | |
4104 | Serge | 3675 | WARN_ON(map_and_fenceable && !i915_is_ggtt(vm)); |
3676 | |||
3677 | vma = i915_gem_obj_to_vma(obj, vm); |
||
3678 | |||
3679 | if (vma) { |
||
3680 | if ((alignment && |
||
3681 | vma->node.start & (alignment - 1)) || |
||
2332 | Serge | 3682 | (map_and_fenceable && !obj->map_and_fenceable)) { |
3683 | WARN(obj->pin_count, |
||
3684 | "bo is already pinned with incorrect alignment:" |
||
4104 | Serge | 3685 | " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d," |
2332 | Serge | 3686 | " obj->map_and_fenceable=%d\n", |
4104 | Serge | 3687 | i915_gem_obj_offset(obj, vm), alignment, |
2332 | Serge | 3688 | map_and_fenceable, |
3689 | obj->map_and_fenceable); |
||
4104 | Serge | 3690 | ret = i915_vma_unbind(vma); |
2332 | Serge | 3691 | if (ret) |
3692 | return ret; |
||
3693 | } |
||
3694 | } |
||
3695 | |||
4104 | Serge | 3696 | if (!i915_gem_obj_bound(obj, vm)) { |
3243 | Serge | 3697 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
3698 | |||
4104 | Serge | 3699 | ret = i915_gem_object_bind_to_vm(obj, vm, alignment, |
3031 | serge | 3700 | map_and_fenceable, |
3701 | nonblocking); |
||
2332 | Serge | 3702 | if (ret) |
3703 | return ret; |
||
3243 | Serge | 3704 | |
3705 | if (!dev_priv->mm.aliasing_ppgtt) |
||
3706 | i915_gem_gtt_bind_object(obj, obj->cache_level); |
||
2332 | Serge | 3707 | } |
3708 | |||
3031 | serge | 3709 | if (!obj->has_global_gtt_mapping && map_and_fenceable) |
3710 | i915_gem_gtt_bind_object(obj, obj->cache_level); |
||
3711 | |||
3712 | obj->pin_count++; |
||
2332 | Serge | 3713 | obj->pin_mappable |= map_and_fenceable; |
3714 | |||
3715 | return 0; |
||
3716 | } |
||
3717 | |||
2344 | Serge | 3718 | void |
3719 | i915_gem_object_unpin(struct drm_i915_gem_object *obj) |
||
3720 | { |
||
3721 | BUG_ON(obj->pin_count == 0); |
||
4104 | Serge | 3722 | BUG_ON(!i915_gem_obj_bound_any(obj)); |
2332 | Serge | 3723 | |
3031 | serge | 3724 | if (--obj->pin_count == 0) |
2344 | Serge | 3725 | obj->pin_mappable = false; |
3726 | } |
||
2332 | Serge | 3727 | |
3031 | serge | 3728 | int |
3729 | i915_gem_pin_ioctl(struct drm_device *dev, void *data, |
||
3730 | struct drm_file *file) |
||
3731 | { |
||
3732 | struct drm_i915_gem_pin *args = data; |
||
3733 | struct drm_i915_gem_object *obj; |
||
3734 | int ret; |
||
2332 | Serge | 3735 | |
3480 | Serge | 3736 | if(args->handle == -2) |
3737 | { |
||
3738 | printf("%s handle %d\n", __FUNCTION__, args->handle); |
||
3739 | return 0; |
||
3740 | } |
||
3741 | |||
3031 | serge | 3742 | ret = i915_mutex_lock_interruptible(dev); |
3743 | if (ret) |
||
3744 | return ret; |
||
2332 | Serge | 3745 | |
3031 | serge | 3746 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3747 | if (&obj->base == NULL) { |
||
3748 | ret = -ENOENT; |
||
3749 | goto unlock; |
||
3750 | } |
||
2332 | Serge | 3751 | |
3031 | serge | 3752 | if (obj->madv != I915_MADV_WILLNEED) { |
3753 | DRM_ERROR("Attempting to pin a purgeable buffer\n"); |
||
3754 | ret = -EINVAL; |
||
3755 | goto out; |
||
3756 | } |
||
2332 | Serge | 3757 | |
3031 | serge | 3758 | if (obj->pin_filp != NULL && obj->pin_filp != file) { |
3759 | DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n", |
||
3760 | args->handle); |
||
3761 | ret = -EINVAL; |
||
3762 | goto out; |
||
3763 | } |
||
2332 | Serge | 3764 | |
3243 | Serge | 3765 | if (obj->user_pin_count == 0) { |
4104 | Serge | 3766 | ret = i915_gem_obj_ggtt_pin(obj, args->alignment, true, false); |
3031 | serge | 3767 | if (ret) |
3768 | goto out; |
||
3769 | } |
||
2332 | Serge | 3770 | |
3243 | Serge | 3771 | obj->user_pin_count++; |
3772 | obj->pin_filp = file; |
||
3773 | |||
4104 | Serge | 3774 | args->offset = i915_gem_obj_ggtt_offset(obj); |
3031 | serge | 3775 | out: |
3776 | drm_gem_object_unreference(&obj->base); |
||
3777 | unlock: |
||
3778 | mutex_unlock(&dev->struct_mutex); |
||
3779 | return ret; |
||
3780 | } |
||
2332 | Serge | 3781 | |
3031 | serge | 3782 | int |
3783 | i915_gem_unpin_ioctl(struct drm_device *dev, void *data, |
||
3784 | struct drm_file *file) |
||
3785 | { |
||
3786 | struct drm_i915_gem_pin *args = data; |
||
3787 | struct drm_i915_gem_object *obj; |
||
3788 | int ret; |
||
2332 | Serge | 3789 | |
3031 | serge | 3790 | ret = i915_mutex_lock_interruptible(dev); |
3791 | if (ret) |
||
3792 | return ret; |
||
2332 | Serge | 3793 | |
4246 | Serge | 3794 | if(args->handle == -2) |
3795 | { |
||
3796 | obj = get_fb_obj(); |
||
3797 | drm_gem_object_reference(&obj->base); |
||
3798 | } |
||
3799 | else |
||
3031 | serge | 3800 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3801 | if (&obj->base == NULL) { |
||
3802 | ret = -ENOENT; |
||
3803 | goto unlock; |
||
3804 | } |
||
2332 | Serge | 3805 | |
3031 | serge | 3806 | if (obj->pin_filp != file) { |
3807 | DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n", |
||
3808 | args->handle); |
||
3809 | ret = -EINVAL; |
||
3810 | goto out; |
||
3811 | } |
||
3812 | obj->user_pin_count--; |
||
3813 | if (obj->user_pin_count == 0) { |
||
3814 | obj->pin_filp = NULL; |
||
3815 | i915_gem_object_unpin(obj); |
||
3816 | } |
||
2332 | Serge | 3817 | |
3031 | serge | 3818 | out: |
3819 | drm_gem_object_unreference(&obj->base); |
||
3820 | unlock: |
||
3821 | mutex_unlock(&dev->struct_mutex); |
||
3822 | return ret; |
||
3823 | } |
||
2332 | Serge | 3824 | |
3031 | serge | 3825 | int |
3826 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, |
||
3827 | struct drm_file *file) |
||
3828 | { |
||
3829 | struct drm_i915_gem_busy *args = data; |
||
3830 | struct drm_i915_gem_object *obj; |
||
3831 | int ret; |
||
2332 | Serge | 3832 | |
3031 | serge | 3833 | ret = i915_mutex_lock_interruptible(dev); |
3834 | if (ret) |
||
3835 | return ret; |
||
2332 | Serge | 3836 | |
3480 | Serge | 3837 | if(args->handle == -2) |
3838 | { |
||
3839 | obj = get_fb_obj(); |
||
3840 | drm_gem_object_reference(&obj->base); |
||
3841 | } |
||
3842 | else |
||
4104 | Serge | 3843 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3031 | serge | 3844 | if (&obj->base == NULL) { |
3845 | ret = -ENOENT; |
||
3846 | goto unlock; |
||
3847 | } |
||
2332 | Serge | 3848 | |
3031 | serge | 3849 | /* Count all active objects as busy, even if they are currently not used |
3850 | * by the gpu. Users of this interface expect objects to eventually |
||
3851 | * become non-busy without any further actions, therefore emit any |
||
3852 | * necessary flushes here. |
||
3853 | */ |
||
3854 | ret = i915_gem_object_flush_active(obj); |
||
2332 | Serge | 3855 | |
3031 | serge | 3856 | args->busy = obj->active; |
3857 | if (obj->ring) { |
||
3858 | BUILD_BUG_ON(I915_NUM_RINGS > 16); |
||
3859 | args->busy |= intel_ring_flag(obj->ring) << 16; |
||
3860 | } |
||
2332 | Serge | 3861 | |
3031 | serge | 3862 | drm_gem_object_unreference(&obj->base); |
3863 | unlock: |
||
3864 | mutex_unlock(&dev->struct_mutex); |
||
3865 | return ret; |
||
3866 | } |
||
2332 | Serge | 3867 | |
3031 | serge | 3868 | int |
3869 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, |
||
3870 | struct drm_file *file_priv) |
||
3871 | { |
||
3872 | return i915_gem_ring_throttle(dev, file_priv); |
||
3873 | } |
||
2332 | Serge | 3874 | |
3263 | Serge | 3875 | #if 0 |
3876 | |||
3031 | serge | 3877 | int |
3878 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, |
||
3879 | struct drm_file *file_priv) |
||
3880 | { |
||
3881 | struct drm_i915_gem_madvise *args = data; |
||
3882 | struct drm_i915_gem_object *obj; |
||
3883 | int ret; |
||
2332 | Serge | 3884 | |
3031 | serge | 3885 | switch (args->madv) { |
3886 | case I915_MADV_DONTNEED: |
||
3887 | case I915_MADV_WILLNEED: |
||
3888 | break; |
||
3889 | default: |
||
3890 | return -EINVAL; |
||
3891 | } |
||
2332 | Serge | 3892 | |
3031 | serge | 3893 | ret = i915_mutex_lock_interruptible(dev); |
3894 | if (ret) |
||
3895 | return ret; |
||
2332 | Serge | 3896 | |
3031 | serge | 3897 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); |
3898 | if (&obj->base == NULL) { |
||
3899 | ret = -ENOENT; |
||
3900 | goto unlock; |
||
3901 | } |
||
2332 | Serge | 3902 | |
3031 | serge | 3903 | if (obj->pin_count) { |
3904 | ret = -EINVAL; |
||
3905 | goto out; |
||
3906 | } |
||
2332 | Serge | 3907 | |
3031 | serge | 3908 | if (obj->madv != __I915_MADV_PURGED) |
3909 | obj->madv = args->madv; |
||
2332 | Serge | 3910 | |
3031 | serge | 3911 | /* if the object is no longer attached, discard its backing storage */ |
3912 | if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL) |
||
3913 | i915_gem_object_truncate(obj); |
||
2332 | Serge | 3914 | |
3031 | serge | 3915 | args->retained = obj->madv != __I915_MADV_PURGED; |
2332 | Serge | 3916 | |
3031 | serge | 3917 | out: |
3918 | drm_gem_object_unreference(&obj->base); |
||
3919 | unlock: |
||
3920 | mutex_unlock(&dev->struct_mutex); |
||
3921 | return ret; |
||
3922 | } |
||
3923 | #endif |
||
2332 | Serge | 3924 | |
3031 | serge | 3925 | void i915_gem_object_init(struct drm_i915_gem_object *obj, |
3926 | const struct drm_i915_gem_object_ops *ops) |
||
3927 | { |
||
4104 | Serge | 3928 | INIT_LIST_HEAD(&obj->global_list); |
3031 | serge | 3929 | INIT_LIST_HEAD(&obj->ring_list); |
3930 | INIT_LIST_HEAD(&obj->exec_list); |
||
4104 | Serge | 3931 | INIT_LIST_HEAD(&obj->obj_exec_link); |
3932 | INIT_LIST_HEAD(&obj->vma_list); |
||
2332 | Serge | 3933 | |
3031 | serge | 3934 | obj->ops = ops; |
3935 | |||
3936 | obj->fence_reg = I915_FENCE_REG_NONE; |
||
3937 | obj->madv = I915_MADV_WILLNEED; |
||
3938 | /* Avoid an unnecessary call to unbind on the first bind. */ |
||
3939 | obj->map_and_fenceable = true; |
||
3940 | |||
3941 | i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size); |
||
3942 | } |
||
3943 | |||
3944 | static const struct drm_i915_gem_object_ops i915_gem_object_ops = { |
||
3945 | .get_pages = i915_gem_object_get_pages_gtt, |
||
3946 | .put_pages = i915_gem_object_put_pages_gtt, |
||
3947 | }; |
||
3948 | |||
2332 | Serge | 3949 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, |
3950 | size_t size) |
||
3951 | { |
||
3952 | struct drm_i915_gem_object *obj; |
||
3031 | serge | 3953 | struct address_space *mapping; |
3480 | Serge | 3954 | gfp_t mask; |
2340 | Serge | 3955 | |
3746 | Serge | 3956 | obj = i915_gem_object_alloc(dev); |
2332 | Serge | 3957 | if (obj == NULL) |
3958 | return NULL; |
||
3959 | |||
3960 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { |
||
4104 | Serge | 3961 | i915_gem_object_free(obj); |
2332 | Serge | 3962 | return NULL; |
3963 | } |
||
3964 | |||
3965 | |||
3031 | serge | 3966 | i915_gem_object_init(obj, &i915_gem_object_ops); |
2332 | Serge | 3967 | |
3968 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
||
3969 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
||
3970 | |||
3031 | serge | 3971 | if (HAS_LLC(dev)) { |
3972 | /* On some devices, we can have the GPU use the LLC (the CPU |
||
2332 | Serge | 3973 | * cache) for about a 10% performance improvement |
3974 | * compared to uncached. Graphics requests other than |
||
3975 | * display scanout are coherent with the CPU in |
||
3976 | * accessing this cache. This means in this mode we |
||
3977 | * don't need to clflush on the CPU side, and on the |
||
3978 | * GPU side we only need to flush internal caches to |
||
3979 | * get data visible to the CPU. |
||
3980 | * |
||
3981 | * However, we maintain the display planes as UC, and so |
||
3982 | * need to rebind when first used as such. |
||
3983 | */ |
||
3984 | obj->cache_level = I915_CACHE_LLC; |
||
3985 | } else |
||
3986 | obj->cache_level = I915_CACHE_NONE; |
||
3987 | |||
3988 | return obj; |
||
3989 | } |
||
3990 | |||
2344 | Serge | 3991 | int i915_gem_init_object(struct drm_gem_object *obj) |
3992 | { |
||
3993 | BUG(); |
||
2332 | Serge | 3994 | |
2344 | Serge | 3995 | return 0; |
3996 | } |
||
2332 | Serge | 3997 | |
3031 | serge | 3998 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
2344 | Serge | 3999 | { |
3031 | serge | 4000 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); |
2344 | Serge | 4001 | struct drm_device *dev = obj->base.dev; |
4002 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
4104 | Serge | 4003 | struct i915_vma *vma, *next; |
2332 | Serge | 4004 | |
3031 | serge | 4005 | trace_i915_gem_object_destroy(obj); |
4006 | |||
4007 | |||
4008 | obj->pin_count = 0; |
||
4104 | Serge | 4009 | /* NB: 0 or 1 elements */ |
4010 | WARN_ON(!list_empty(&obj->vma_list) && |
||
4011 | !list_is_singular(&obj->vma_list)); |
||
4012 | list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) { |
||
4013 | int ret = i915_vma_unbind(vma); |
||
4014 | if (WARN_ON(ret == -ERESTARTSYS)) { |
||
3031 | serge | 4015 | bool was_interruptible; |
4016 | |||
4017 | was_interruptible = dev_priv->mm.interruptible; |
||
4018 | dev_priv->mm.interruptible = false; |
||
4019 | |||
4104 | Serge | 4020 | WARN_ON(i915_vma_unbind(vma)); |
3031 | serge | 4021 | |
4022 | dev_priv->mm.interruptible = was_interruptible; |
||
2344 | Serge | 4023 | } |
4104 | Serge | 4024 | } |
2332 | Serge | 4025 | |
4104 | Serge | 4026 | /* Stolen objects don't hold a ref, but do hold pin count. Fix that up |
4027 | * before progressing. */ |
||
4028 | if (obj->stolen) |
||
4029 | i915_gem_object_unpin_pages(obj); |
||
4030 | |||
4031 | if (WARN_ON(obj->pages_pin_count)) |
||
3031 | serge | 4032 | obj->pages_pin_count = 0; |
4033 | i915_gem_object_put_pages(obj); |
||
4034 | // i915_gem_object_free_mmap_offset(obj); |
||
4104 | Serge | 4035 | i915_gem_object_release_stolen(obj); |
2332 | Serge | 4036 | |
3243 | Serge | 4037 | BUG_ON(obj->pages); |
2332 | Serge | 4038 | |
3031 | serge | 4039 | |
3290 | Serge | 4040 | if(obj->base.filp != NULL) |
4041 | { |
||
3298 | Serge | 4042 | // printf("filp %p\n", obj->base.filp); |
3290 | Serge | 4043 | shmem_file_delete(obj->base.filp); |
4044 | } |
||
4045 | |||
2344 | Serge | 4046 | drm_gem_object_release(&obj->base); |
4047 | i915_gem_info_remove_obj(dev_priv, obj->base.size); |
||
2332 | Serge | 4048 | |
2344 | Serge | 4049 | kfree(obj->bit_17); |
4104 | Serge | 4050 | i915_gem_object_free(obj); |
2344 | Serge | 4051 | } |
2332 | Serge | 4052 | |
4104 | Serge | 4053 | struct i915_vma *i915_gem_vma_create(struct drm_i915_gem_object *obj, |
4054 | struct i915_address_space *vm) |
||
4055 | { |
||
4056 | struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL); |
||
4057 | if (vma == NULL) |
||
4058 | return ERR_PTR(-ENOMEM); |
||
4059 | |||
4060 | INIT_LIST_HEAD(&vma->vma_link); |
||
4061 | INIT_LIST_HEAD(&vma->mm_list); |
||
4062 | INIT_LIST_HEAD(&vma->exec_list); |
||
4063 | vma->vm = vm; |
||
4064 | vma->obj = obj; |
||
4065 | |||
4066 | /* Keep GGTT vmas first to make debug easier */ |
||
4067 | if (i915_is_ggtt(vm)) |
||
4068 | list_add(&vma->vma_link, &obj->vma_list); |
||
4069 | else |
||
4070 | list_add_tail(&vma->vma_link, &obj->vma_list); |
||
4071 | |||
4072 | return vma; |
||
4073 | } |
||
4074 | |||
4075 | void i915_gem_vma_destroy(struct i915_vma *vma) |
||
4076 | { |
||
4077 | WARN_ON(vma->node.allocated); |
||
4078 | list_del(&vma->vma_link); |
||
4079 | kfree(vma); |
||
4080 | } |
||
4081 | |||
3031 | serge | 4082 | #if 0 |
4083 | int |
||
4084 | i915_gem_idle(struct drm_device *dev) |
||
2344 | Serge | 4085 | { |
3031 | serge | 4086 | drm_i915_private_t *dev_priv = dev->dev_private; |
4087 | int ret; |
||
2332 | Serge | 4088 | |
4104 | Serge | 4089 | if (dev_priv->ums.mm_suspended) { |
3031 | serge | 4090 | mutex_unlock(&dev->struct_mutex); |
4091 | return 0; |
||
4092 | } |
||
2332 | Serge | 4093 | |
3031 | serge | 4094 | ret = i915_gpu_idle(dev); |
4095 | if (ret) { |
||
4096 | mutex_unlock(&dev->struct_mutex); |
||
4097 | return ret; |
||
4098 | } |
||
4099 | i915_gem_retire_requests(dev); |
||
4100 | |||
3480 | Serge | 4101 | /* Under UMS, be paranoid and evict. */ |
4102 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
||
4103 | i915_gem_evict_everything(dev); |
||
4104 | |||
4105 | del_timer_sync(&dev_priv->gpu_error.hangcheck_timer); |
||
3031 | serge | 4106 | |
4107 | i915_kernel_lost_context(dev); |
||
4108 | i915_gem_cleanup_ringbuffer(dev); |
||
4109 | |||
4110 | /* Cancel the retire work handler, which should be idle now. */ |
||
3263 | Serge | 4111 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); |
3031 | serge | 4112 | |
4113 | return 0; |
||
2344 | Serge | 4114 | } |
3031 | serge | 4115 | #endif |
2332 | Serge | 4116 | |
3031 | serge | 4117 | void i915_gem_l3_remap(struct drm_device *dev) |
4118 | { |
||
4119 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
4120 | u32 misccpctl; |
||
4121 | int i; |
||
2332 | Serge | 4122 | |
3480 | Serge | 4123 | if (!HAS_L3_GPU_CACHE(dev)) |
3031 | serge | 4124 | return; |
2332 | Serge | 4125 | |
3243 | Serge | 4126 | if (!dev_priv->l3_parity.remap_info) |
3031 | serge | 4127 | return; |
2332 | Serge | 4128 | |
3031 | serge | 4129 | misccpctl = I915_READ(GEN7_MISCCPCTL); |
4130 | I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE); |
||
4131 | POSTING_READ(GEN7_MISCCPCTL); |
||
2332 | Serge | 4132 | |
3031 | serge | 4133 | for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) { |
4134 | u32 remap = I915_READ(GEN7_L3LOG_BASE + i); |
||
3243 | Serge | 4135 | if (remap && remap != dev_priv->l3_parity.remap_info[i/4]) |
3031 | serge | 4136 | DRM_DEBUG("0x%x was already programmed to %x\n", |
4137 | GEN7_L3LOG_BASE + i, remap); |
||
3243 | Serge | 4138 | if (remap && !dev_priv->l3_parity.remap_info[i/4]) |
3031 | serge | 4139 | DRM_DEBUG_DRIVER("Clearing remapped register\n"); |
3243 | Serge | 4140 | I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]); |
3031 | serge | 4141 | } |
2332 | Serge | 4142 | |
3031 | serge | 4143 | /* Make sure all the writes land before disabling dop clock gating */ |
4144 | POSTING_READ(GEN7_L3LOG_BASE); |
||
2332 | Serge | 4145 | |
3031 | serge | 4146 | I915_WRITE(GEN7_MISCCPCTL, misccpctl); |
4147 | } |
||
2332 | Serge | 4148 | |
3031 | serge | 4149 | void i915_gem_init_swizzling(struct drm_device *dev) |
4150 | { |
||
4151 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
2332 | Serge | 4152 | |
3031 | serge | 4153 | if (INTEL_INFO(dev)->gen < 5 || |
4154 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
||
4155 | return; |
||
2332 | Serge | 4156 | |
3031 | serge | 4157 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | |
4158 | DISP_TILE_SURFACE_SWIZZLING); |
||
2332 | Serge | 4159 | |
3031 | serge | 4160 | if (IS_GEN5(dev)) |
4161 | return; |
||
2344 | Serge | 4162 | |
3031 | serge | 4163 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
4164 | if (IS_GEN6(dev)) |
||
4165 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); |
||
3480 | Serge | 4166 | else if (IS_GEN7(dev)) |
4167 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); |
||
3031 | serge | 4168 | else |
3480 | Serge | 4169 | BUG(); |
3031 | serge | 4170 | } |
4171 | |||
4172 | static bool |
||
4173 | intel_enable_blt(struct drm_device *dev) |
||
4174 | { |
||
4175 | if (!HAS_BLT(dev)) |
||
4176 | return false; |
||
4177 | |||
4178 | /* The blitter was dysfunctional on early prototypes */ |
||
4179 | if (IS_GEN6(dev) && dev->pdev->revision < 8) { |
||
4180 | DRM_INFO("BLT not supported on this pre-production hardware;" |
||
4181 | " graphics performance will be degraded.\n"); |
||
4182 | return false; |
||
4183 | } |
||
4184 | |||
4185 | return true; |
||
4186 | } |
||
4187 | |||
3480 | Serge | 4188 | static int i915_gem_init_rings(struct drm_device *dev) |
2332 | Serge | 4189 | { |
3480 | Serge | 4190 | struct drm_i915_private *dev_priv = dev->dev_private; |
2332 | Serge | 4191 | int ret; |
2351 | Serge | 4192 | |
2332 | Serge | 4193 | ret = intel_init_render_ring_buffer(dev); |
4194 | if (ret) |
||
4195 | return ret; |
||
4196 | |||
4197 | if (HAS_BSD(dev)) { |
||
4198 | ret = intel_init_bsd_ring_buffer(dev); |
||
4199 | if (ret) |
||
4200 | goto cleanup_render_ring; |
||
4201 | } |
||
4202 | |||
3031 | serge | 4203 | if (intel_enable_blt(dev)) { |
2332 | Serge | 4204 | ret = intel_init_blt_ring_buffer(dev); |
4205 | if (ret) |
||
4206 | goto cleanup_bsd_ring; |
||
4207 | } |
||
4208 | |||
4104 | Serge | 4209 | if (HAS_VEBOX(dev)) { |
4210 | ret = intel_init_vebox_ring_buffer(dev); |
||
4211 | if (ret) |
||
4212 | goto cleanup_blt_ring; |
||
4213 | } |
||
4214 | |||
4215 | |||
3480 | Serge | 4216 | ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000)); |
4217 | if (ret) |
||
4104 | Serge | 4218 | goto cleanup_vebox_ring; |
2351 | Serge | 4219 | |
2332 | Serge | 4220 | return 0; |
4221 | |||
4104 | Serge | 4222 | cleanup_vebox_ring: |
4223 | intel_cleanup_ring_buffer(&dev_priv->ring[VECS]); |
||
3480 | Serge | 4224 | cleanup_blt_ring: |
4225 | intel_cleanup_ring_buffer(&dev_priv->ring[BCS]); |
||
2332 | Serge | 4226 | cleanup_bsd_ring: |
4227 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); |
||
4228 | cleanup_render_ring: |
||
4229 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); |
||
3480 | Serge | 4230 | |
2332 | Serge | 4231 | return ret; |
4232 | } |
||
4233 | |||
3480 | Serge | 4234 | int |
4235 | i915_gem_init_hw(struct drm_device *dev) |
||
3031 | serge | 4236 | { |
3480 | Serge | 4237 | drm_i915_private_t *dev_priv = dev->dev_private; |
4238 | int ret; |
||
3031 | serge | 4239 | |
3480 | Serge | 4240 | if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt()) |
4241 | return -EIO; |
||
3031 | serge | 4242 | |
4104 | Serge | 4243 | if (dev_priv->ellc_size) |
4244 | I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf)); |
||
3480 | Serge | 4245 | |
3746 | Serge | 4246 | if (HAS_PCH_NOP(dev)) { |
4247 | u32 temp = I915_READ(GEN7_MSG_CTL); |
||
4248 | temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK); |
||
4249 | I915_WRITE(GEN7_MSG_CTL, temp); |
||
4250 | } |
||
4251 | |||
3480 | Serge | 4252 | i915_gem_l3_remap(dev); |
4253 | |||
4254 | i915_gem_init_swizzling(dev); |
||
4255 | |||
4256 | ret = i915_gem_init_rings(dev); |
||
4257 | if (ret) |
||
4258 | return ret; |
||
4259 | |||
4260 | /* |
||
4261 | * XXX: There was some w/a described somewhere suggesting loading |
||
4262 | * contexts before PPGTT. |
||
4263 | */ |
||
4264 | i915_gem_context_init(dev); |
||
3746 | Serge | 4265 | if (dev_priv->mm.aliasing_ppgtt) { |
4266 | ret = dev_priv->mm.aliasing_ppgtt->enable(dev); |
||
4267 | if (ret) { |
||
4268 | i915_gem_cleanup_aliasing_ppgtt(dev); |
||
4269 | DRM_INFO("PPGTT enable failed. This is not fatal, but unexpected\n"); |
||
4270 | } |
||
4271 | } |
||
3480 | Serge | 4272 | |
4273 | return 0; |
||
3031 | serge | 4274 | } |
4275 | |||
4276 | int i915_gem_init(struct drm_device *dev) |
||
4277 | { |
||
4278 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
4279 | int ret; |
||
4280 | |||
4281 | mutex_lock(&dev->struct_mutex); |
||
3746 | Serge | 4282 | |
4283 | if (IS_VALLEYVIEW(dev)) { |
||
4284 | /* VLVA0 (potential hack), BIOS isn't actually waking us */ |
||
4285 | I915_WRITE(VLV_GTLC_WAKE_CTRL, 1); |
||
4286 | if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) & 1) == 1, 10)) |
||
4287 | DRM_DEBUG_DRIVER("allow wake ack timed out\n"); |
||
4288 | } |
||
4289 | |||
3480 | Serge | 4290 | i915_gem_init_global_gtt(dev); |
3746 | Serge | 4291 | |
3031 | serge | 4292 | ret = i915_gem_init_hw(dev); |
4293 | mutex_unlock(&dev->struct_mutex); |
||
4294 | if (ret) { |
||
4295 | i915_gem_cleanup_aliasing_ppgtt(dev); |
||
4296 | return ret; |
||
4297 | } |
||
4298 | |||
3746 | Serge | 4299 | |
3031 | serge | 4300 | return 0; |
4301 | } |
||
4302 | |||
2332 | Serge | 4303 | void |
4304 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) |
||
4305 | { |
||
4306 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
3031 | serge | 4307 | struct intel_ring_buffer *ring; |
2332 | Serge | 4308 | int i; |
4309 | |||
3031 | serge | 4310 | for_each_ring(ring, dev_priv, i) |
4311 | intel_cleanup_ring_buffer(ring); |
||
2332 | Serge | 4312 | } |
4313 | |||
3031 | serge | 4314 | #if 0 |
4315 | |||
2332 | Serge | 4316 | int |
4317 | i915_gem_entervt_ioctl(struct drm_device *dev, void *data, |
||
4318 | struct drm_file *file_priv) |
||
4319 | { |
||
4104 | Serge | 4320 | struct drm_i915_private *dev_priv = dev->dev_private; |
3031 | serge | 4321 | int ret; |
2332 | Serge | 4322 | |
4323 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
||
4324 | return 0; |
||
4325 | |||
3480 | Serge | 4326 | if (i915_reset_in_progress(&dev_priv->gpu_error)) { |
2332 | Serge | 4327 | DRM_ERROR("Reenabling wedged hardware, good luck\n"); |
3480 | Serge | 4328 | atomic_set(&dev_priv->gpu_error.reset_counter, 0); |
2332 | Serge | 4329 | } |
4330 | |||
4331 | mutex_lock(&dev->struct_mutex); |
||
4104 | Serge | 4332 | dev_priv->ums.mm_suspended = 0; |
2332 | Serge | 4333 | |
3031 | serge | 4334 | ret = i915_gem_init_hw(dev); |
2332 | Serge | 4335 | if (ret != 0) { |
4336 | mutex_unlock(&dev->struct_mutex); |
||
4337 | return ret; |
||
4338 | } |
||
4339 | |||
4104 | Serge | 4340 | BUG_ON(!list_empty(&dev_priv->gtt.base.active_list)); |
2332 | Serge | 4341 | mutex_unlock(&dev->struct_mutex); |
4342 | |||
4343 | ret = drm_irq_install(dev); |
||
4344 | if (ret) |
||
4345 | goto cleanup_ringbuffer; |
||
4346 | |||
4347 | return 0; |
||
4348 | |||
4349 | cleanup_ringbuffer: |
||
4350 | mutex_lock(&dev->struct_mutex); |
||
4351 | i915_gem_cleanup_ringbuffer(dev); |
||
4104 | Serge | 4352 | dev_priv->ums.mm_suspended = 1; |
2332 | Serge | 4353 | mutex_unlock(&dev->struct_mutex); |
4354 | |||
4355 | return ret; |
||
4356 | } |
||
4357 | |||
4358 | int |
||
4359 | i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, |
||
4360 | struct drm_file *file_priv) |
||
4361 | { |
||
4104 | Serge | 4362 | struct drm_i915_private *dev_priv = dev->dev_private; |
4363 | int ret; |
||
4364 | |||
2332 | Serge | 4365 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4366 | return 0; |
||
4367 | |||
4368 | drm_irq_uninstall(dev); |
||
4104 | Serge | 4369 | |
4370 | mutex_lock(&dev->struct_mutex); |
||
4371 | ret = i915_gem_idle(dev); |
||
4372 | |||
4373 | /* Hack! Don't let anybody do execbuf while we don't control the chip. |
||
4374 | * We need to replace this with a semaphore, or something. |
||
4375 | * And not confound ums.mm_suspended! |
||
4376 | */ |
||
4377 | if (ret != 0) |
||
4378 | dev_priv->ums.mm_suspended = 1; |
||
4379 | mutex_unlock(&dev->struct_mutex); |
||
4380 | |||
4381 | return ret; |
||
2332 | Serge | 4382 | } |
4383 | |||
4384 | void |
||
4385 | i915_gem_lastclose(struct drm_device *dev) |
||
4386 | { |
||
4387 | int ret; |
||
4388 | |||
4389 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
||
4390 | return; |
||
4391 | |||
4104 | Serge | 4392 | mutex_lock(&dev->struct_mutex); |
2332 | Serge | 4393 | ret = i915_gem_idle(dev); |
4394 | if (ret) |
||
4395 | DRM_ERROR("failed to idle hardware: %d\n", ret); |
||
4104 | Serge | 4396 | mutex_unlock(&dev->struct_mutex); |
2332 | Serge | 4397 | } |
4398 | #endif |
||
4399 | |||
4400 | static void |
||
2326 | Serge | 4401 | init_ring_lists(struct intel_ring_buffer *ring) |
4402 | { |
||
4403 | INIT_LIST_HEAD(&ring->active_list); |
||
4404 | INIT_LIST_HEAD(&ring->request_list); |
||
4405 | } |
||
4406 | |||
4104 | Serge | 4407 | static void i915_init_vm(struct drm_i915_private *dev_priv, |
4408 | struct i915_address_space *vm) |
||
4409 | { |
||
4410 | vm->dev = dev_priv->dev; |
||
4411 | INIT_LIST_HEAD(&vm->active_list); |
||
4412 | INIT_LIST_HEAD(&vm->inactive_list); |
||
4413 | INIT_LIST_HEAD(&vm->global_link); |
||
4414 | list_add(&vm->global_link, &dev_priv->vm_list); |
||
4415 | } |
||
4416 | |||
2326 | Serge | 4417 | void |
4418 | i915_gem_load(struct drm_device *dev) |
||
4419 | { |
||
3480 | Serge | 4420 | drm_i915_private_t *dev_priv = dev->dev_private; |
2326 | Serge | 4421 | int i; |
4422 | |||
4104 | Serge | 4423 | INIT_LIST_HEAD(&dev_priv->vm_list); |
4424 | i915_init_vm(dev_priv, &dev_priv->gtt.base); |
||
4425 | |||
3031 | serge | 4426 | INIT_LIST_HEAD(&dev_priv->mm.unbound_list); |
4427 | INIT_LIST_HEAD(&dev_priv->mm.bound_list); |
||
2326 | Serge | 4428 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
4429 | for (i = 0; i < I915_NUM_RINGS; i++) |
||
4430 | init_ring_lists(&dev_priv->ring[i]); |
||
2342 | Serge | 4431 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) |
2326 | Serge | 4432 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); |
2360 | Serge | 4433 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, |
4434 | i915_gem_retire_work_handler); |
||
3480 | Serge | 4435 | init_waitqueue_head(&dev_priv->gpu_error.reset_queue); |
2326 | Serge | 4436 | |
4437 | /* On GEN3 we really need to make sure the ARB C3 LP bit is set */ |
||
4438 | if (IS_GEN3(dev)) { |
||
3031 | serge | 4439 | I915_WRITE(MI_ARB_STATE, |
4440 | _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE)); |
||
2326 | Serge | 4441 | } |
4442 | |||
4443 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; |
||
4444 | |||
3746 | Serge | 4445 | if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev)) |
4446 | dev_priv->num_fence_regs = 32; |
||
4447 | else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) |
||
2326 | Serge | 4448 | dev_priv->num_fence_regs = 16; |
4449 | else |
||
4450 | dev_priv->num_fence_regs = 8; |
||
4451 | |||
4452 | /* Initialize fence registers to zero */ |
||
3746 | Serge | 4453 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
4454 | i915_gem_restore_fences(dev); |
||
2326 | Serge | 4455 | |
4456 | i915_gem_detect_bit_6_swizzle(dev); |
||
4457 | |||
4458 | dev_priv->mm.interruptible = true; |
||
4459 | |||
4460 | } |
||
4461 | |||
4104 | Serge | 4462 | #if 0 |
4463 | /* |
||
4464 | * Create a physically contiguous memory object for this object |
||
4465 | * e.g. for cursor + overlay regs |
||
4466 | */ |
||
4467 | static int i915_gem_init_phys_object(struct drm_device *dev, |
||
4468 | int id, int size, int align) |
||
4469 | { |
||
4470 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
4471 | struct drm_i915_gem_phys_object *phys_obj; |
||
4472 | int ret; |
||
2326 | Serge | 4473 | |
4104 | Serge | 4474 | if (dev_priv->mm.phys_objs[id - 1] || !size) |
4475 | return 0; |
||
4476 | |||
4477 | phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL); |
||
4478 | if (!phys_obj) |
||
4479 | return -ENOMEM; |
||
4480 | |||
4481 | phys_obj->id = id; |
||
4482 | |||
4483 | phys_obj->handle = drm_pci_alloc(dev, size, align); |
||
4484 | if (!phys_obj->handle) { |
||
4485 | ret = -ENOMEM; |
||
4486 | goto kfree_obj; |
||
4487 | } |
||
4488 | #ifdef CONFIG_X86 |
||
4489 | set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); |
||
4490 | #endif |
||
4491 | |||
4492 | dev_priv->mm.phys_objs[id - 1] = phys_obj; |
||
4493 | |||
4494 | return 0; |
||
4495 | kfree_obj: |
||
4496 | kfree(phys_obj); |
||
4497 | return ret; |
||
4498 | } |
||
4499 | |||
4500 | static void i915_gem_free_phys_object(struct drm_device *dev, int id) |
||
4501 | { |
||
4502 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
4503 | struct drm_i915_gem_phys_object *phys_obj; |
||
4504 | |||
4505 | if (!dev_priv->mm.phys_objs[id - 1]) |
||
4506 | return; |
||
4507 | |||
4508 | phys_obj = dev_priv->mm.phys_objs[id - 1]; |
||
4509 | if (phys_obj->cur_obj) { |
||
4510 | i915_gem_detach_phys_object(dev, phys_obj->cur_obj); |
||
4511 | } |
||
4512 | |||
4513 | #ifdef CONFIG_X86 |
||
4514 | set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); |
||
4515 | #endif |
||
4516 | drm_pci_free(dev, phys_obj->handle); |
||
4517 | kfree(phys_obj); |
||
4518 | dev_priv->mm.phys_objs[id - 1] = NULL; |
||
4519 | } |
||
4520 | |||
4521 | void i915_gem_free_all_phys_object(struct drm_device *dev) |
||
4522 | { |
||
4523 | int i; |
||
4524 | |||
4525 | for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++) |
||
4526 | i915_gem_free_phys_object(dev, i); |
||
4527 | } |
||
4528 | |||
4529 | void i915_gem_detach_phys_object(struct drm_device *dev, |
||
4530 | struct drm_i915_gem_object *obj) |
||
4531 | { |
||
4532 | struct address_space *mapping = file_inode(obj->base.filp)->i_mapping; |
||
4533 | char *vaddr; |
||
4534 | int i; |
||
4535 | int page_count; |
||
4536 | |||
4537 | if (!obj->phys_obj) |
||
4538 | return; |
||
4539 | vaddr = obj->phys_obj->handle->vaddr; |
||
4540 | |||
4541 | page_count = obj->base.size / PAGE_SIZE; |
||
4542 | for (i = 0; i < page_count; i++) { |
||
4543 | struct page *page = shmem_read_mapping_page(mapping, i); |
||
4544 | if (!IS_ERR(page)) { |
||
4545 | char *dst = kmap_atomic(page); |
||
4546 | memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE); |
||
4547 | kunmap_atomic(dst); |
||
4548 | |||
4549 | drm_clflush_pages(&page, 1); |
||
4550 | |||
4551 | set_page_dirty(page); |
||
4552 | mark_page_accessed(page); |
||
4553 | page_cache_release(page); |
||
4554 | } |
||
4555 | } |
||
4556 | i915_gem_chipset_flush(dev); |
||
4557 | |||
4558 | obj->phys_obj->cur_obj = NULL; |
||
4559 | obj->phys_obj = NULL; |
||
4560 | } |
||
4561 | |||
4562 | int |
||
4563 | i915_gem_attach_phys_object(struct drm_device *dev, |
||
4564 | struct drm_i915_gem_object *obj, |
||
4565 | int id, |
||
4566 | int align) |
||
4567 | { |
||
4568 | struct address_space *mapping = file_inode(obj->base.filp)->i_mapping; |
||
4569 | drm_i915_private_t *dev_priv = dev->dev_private; |
||
4570 | int ret = 0; |
||
4571 | int page_count; |
||
4572 | int i; |
||
4573 | |||
4574 | if (id > I915_MAX_PHYS_OBJECT) |
||
4575 | return -EINVAL; |
||
4576 | |||
4577 | if (obj->phys_obj) { |
||
4578 | if (obj->phys_obj->id == id) |
||
4579 | return 0; |
||
4580 | i915_gem_detach_phys_object(dev, obj); |
||
4581 | } |
||
4582 | |||
4583 | /* create a new object */ |
||
4584 | if (!dev_priv->mm.phys_objs[id - 1]) { |
||
4585 | ret = i915_gem_init_phys_object(dev, id, |
||
4586 | obj->base.size, align); |
||
4587 | if (ret) { |
||
4588 | DRM_ERROR("failed to init phys object %d size: %zu\n", |
||
4589 | id, obj->base.size); |
||
4590 | return ret; |
||
4591 | } |
||
4592 | } |
||
4593 | |||
4594 | /* bind to the object */ |
||
4595 | obj->phys_obj = dev_priv->mm.phys_objs[id - 1]; |
||
4596 | obj->phys_obj->cur_obj = obj; |
||
4597 | |||
4598 | page_count = obj->base.size / PAGE_SIZE; |
||
4599 | |||
4600 | for (i = 0; i < page_count; i++) { |
||
4601 | struct page *page; |
||
4602 | char *dst, *src; |
||
4603 | |||
4604 | page = shmem_read_mapping_page(mapping, i); |
||
4605 | if (IS_ERR(page)) |
||
4606 | return PTR_ERR(page); |
||
4607 | |||
4608 | src = kmap_atomic(page); |
||
4609 | dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE); |
||
4610 | memcpy(dst, src, PAGE_SIZE); |
||
4611 | kunmap_atomic(src); |
||
4612 | |||
4613 | mark_page_accessed(page); |
||
4614 | page_cache_release(page); |
||
4615 | } |
||
4616 | |||
4617 | return 0; |
||
4618 | } |
||
4619 | |||
4620 | static int |
||
4621 | i915_gem_phys_pwrite(struct drm_device *dev, |
||
4622 | struct drm_i915_gem_object *obj, |
||
4623 | struct drm_i915_gem_pwrite *args, |
||
4624 | struct drm_file *file_priv) |
||
4625 | { |
||
4626 | void *vaddr = obj->phys_obj->handle->vaddr + args->offset; |
||
4627 | char __user *user_data = to_user_ptr(args->data_ptr); |
||
4628 | |||
4629 | if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) { |
||
4630 | unsigned long unwritten; |
||
4631 | |||
4632 | /* The physical object once assigned is fixed for the lifetime |
||
4633 | * of the obj, so we can safely drop the lock and continue |
||
4634 | * to access vaddr. |
||
4635 | */ |
||
4636 | mutex_unlock(&dev->struct_mutex); |
||
4637 | unwritten = copy_from_user(vaddr, user_data, args->size); |
||
4638 | mutex_lock(&dev->struct_mutex); |
||
4639 | if (unwritten) |
||
4640 | return -EFAULT; |
||
4641 | } |
||
4642 | |||
4643 | i915_gem_chipset_flush(dev); |
||
4644 | return 0; |
||
4645 | } |
||
4646 | |||
4647 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
||
4648 | { |
||
4649 | struct drm_i915_file_private *file_priv = file->driver_priv; |
||
4650 | |||
4651 | /* Clean up our request list when the client is going away, so that |
||
4652 | * later retire_requests won't dereference our soon-to-be-gone |
||
4653 | * file_priv. |
||
4654 | */ |
||
4655 | spin_lock(&file_priv->mm.lock); |
||
4656 | while (!list_empty(&file_priv->mm.request_list)) { |
||
4657 | struct drm_i915_gem_request *request; |
||
4658 | |||
4659 | request = list_first_entry(&file_priv->mm.request_list, |
||
4660 | struct drm_i915_gem_request, |
||
4661 | client_list); |
||
4662 | list_del(&request->client_list); |
||
4663 | request->file_priv = NULL; |
||
4664 | } |
||
4665 | spin_unlock(&file_priv->mm.lock); |
||
4666 | } |
||
4667 | #endif |
||
4668 | |||
4669 | static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task) |
||
4670 | { |
||
4671 | if (!mutex_is_locked(mutex)) |
||
4672 | return false; |
||
4673 | |||
4674 | #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES) |
||
4675 | return mutex->owner == task; |
||
4676 | #else |
||
4677 | /* Since UP may be pre-empted, we cannot assume that we own the lock */ |
||
4678 | return false; |
||
4679 | #endif |
||
4680 | } |
||
4681 | |||
4682 | /* All the new VM stuff */ |
||
4683 | unsigned long i915_gem_obj_offset(struct drm_i915_gem_object *o, |
||
4684 | struct i915_address_space *vm) |
||
4685 | { |
||
4686 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; |
||
4687 | struct i915_vma *vma; |
||
4688 | |||
4689 | if (vm == &dev_priv->mm.aliasing_ppgtt->base) |
||
4690 | vm = &dev_priv->gtt.base; |
||
4691 | |||
4692 | BUG_ON(list_empty(&o->vma_list)); |
||
4693 | list_for_each_entry(vma, &o->vma_list, vma_link) { |
||
4694 | if (vma->vm == vm) |
||
4695 | return vma->node.start; |
||
4696 | |||
4697 | } |
||
4698 | return 0; //-1; |
||
4699 | } |
||
4700 | |||
4701 | bool i915_gem_obj_bound(struct drm_i915_gem_object *o, |
||
4702 | struct i915_address_space *vm) |
||
4703 | { |
||
4704 | struct i915_vma *vma; |
||
4705 | |||
4706 | list_for_each_entry(vma, &o->vma_list, vma_link) |
||
4707 | if (vma->vm == vm && drm_mm_node_allocated(&vma->node)) |
||
4708 | return true; |
||
4709 | |||
4710 | return false; |
||
4711 | } |
||
4712 | |||
4713 | bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o) |
||
4714 | { |
||
4715 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; |
||
4716 | struct i915_address_space *vm; |
||
4717 | |||
4718 | list_for_each_entry(vm, &dev_priv->vm_list, global_link) |
||
4719 | if (i915_gem_obj_bound(o, vm)) |
||
4720 | return true; |
||
4721 | |||
4722 | return false; |
||
4723 | } |
||
4724 | |||
4725 | unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o, |
||
4726 | struct i915_address_space *vm) |
||
4727 | { |
||
4728 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; |
||
4729 | struct i915_vma *vma; |
||
4730 | |||
4731 | if (vm == &dev_priv->mm.aliasing_ppgtt->base) |
||
4732 | vm = &dev_priv->gtt.base; |
||
4733 | |||
4734 | BUG_ON(list_empty(&o->vma_list)); |
||
4735 | |||
4736 | list_for_each_entry(vma, &o->vma_list, vma_link) |
||
4737 | if (vma->vm == vm) |
||
4738 | return vma->node.size; |
||
4739 | |||
4740 | return 0; |
||
4741 | } |
||
4742 | struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj, |
||
4743 | struct i915_address_space *vm) |
||
4744 | { |
||
4745 | struct i915_vma *vma; |
||
4746 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
||
4747 | if (vma->vm == vm) |
||
4748 | return vma; |
||
4749 | |||
4750 | return NULL; |
||
4751 | } |
||
4752 | |||
4753 | struct i915_vma * |
||
4754 | i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj, |
||
4755 | struct i915_address_space *vm) |
||
4756 | { |
||
4757 | struct i915_vma *vma; |
||
4758 | |||
4759 | vma = i915_gem_obj_to_vma(obj, vm); |
||
4760 | if (!vma) |
||
4761 | vma = i915_gem_vma_create(obj, vm); |
||
4762 | |||
4763 | return vma; |
||
4764 | }>>=>>>>>>>><>>>>= |