]>
Commit | Line | Data |
---|---|---|
673a394b EA |
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 <eric@anholt.net> | |
25 | * | |
26 | */ | |
27 | ||
760285e7 | 28 | #include <drm/drmP.h> |
0de23977 | 29 | #include <drm/drm_vma_manager.h> |
760285e7 | 30 | #include <drm/i915_drm.h> |
673a394b | 31 | #include "i915_drv.h" |
1c5d22f7 | 32 | #include "i915_trace.h" |
652c393a | 33 | #include "intel_drv.h" |
2cfcd32a | 34 | #include <linux/oom.h> |
5949eac4 | 35 | #include <linux/shmem_fs.h> |
5a0e3ad6 | 36 | #include <linux/slab.h> |
673a394b | 37 | #include <linux/swap.h> |
79e53945 | 38 | #include <linux/pci.h> |
1286ff73 | 39 | #include <linux/dma-buf.h> |
673a394b | 40 | |
05394f39 | 41 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); |
2c22569b CW |
42 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj, |
43 | bool force); | |
07fe0b12 | 44 | static __must_check int |
23f54483 BW |
45 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj, |
46 | bool readonly); | |
c8725f3d CW |
47 | static void |
48 | i915_gem_object_retire(struct drm_i915_gem_object *obj); | |
49 | ||
61050808 CW |
50 | static void i915_gem_write_fence(struct drm_device *dev, int reg, |
51 | struct drm_i915_gem_object *obj); | |
52 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, | |
53 | struct drm_i915_fence_reg *fence, | |
54 | bool enable); | |
55 | ||
ceabbba5 | 56 | static unsigned long i915_gem_shrinker_count(struct shrinker *shrinker, |
7dc19d5a | 57 | struct shrink_control *sc); |
ceabbba5 | 58 | static unsigned long i915_gem_shrinker_scan(struct shrinker *shrinker, |
7dc19d5a | 59 | struct shrink_control *sc); |
2cfcd32a CW |
60 | static int i915_gem_shrinker_oom(struct notifier_block *nb, |
61 | unsigned long event, | |
62 | void *ptr); | |
d9973b43 CW |
63 | static unsigned long i915_gem_purge(struct drm_i915_private *dev_priv, long target); |
64 | static unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv); | |
31169714 | 65 | |
c76ce038 CW |
66 | static bool cpu_cache_is_coherent(struct drm_device *dev, |
67 | enum i915_cache_level level) | |
68 | { | |
69 | return HAS_LLC(dev) || level != I915_CACHE_NONE; | |
70 | } | |
71 | ||
2c22569b CW |
72 | static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) |
73 | { | |
74 | if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) | |
75 | return true; | |
76 | ||
77 | return obj->pin_display; | |
78 | } | |
79 | ||
61050808 CW |
80 | static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj) |
81 | { | |
82 | if (obj->tiling_mode) | |
83 | i915_gem_release_mmap(obj); | |
84 | ||
85 | /* As we do not have an associated fence register, we will force | |
86 | * a tiling change if we ever need to acquire one. | |
87 | */ | |
5d82e3e6 | 88 | obj->fence_dirty = false; |
61050808 CW |
89 | obj->fence_reg = I915_FENCE_REG_NONE; |
90 | } | |
91 | ||
73aa808f CW |
92 | /* some bookkeeping */ |
93 | static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv, | |
94 | size_t size) | |
95 | { | |
c20e8355 | 96 | spin_lock(&dev_priv->mm.object_stat_lock); |
73aa808f CW |
97 | dev_priv->mm.object_count++; |
98 | dev_priv->mm.object_memory += size; | |
c20e8355 | 99 | spin_unlock(&dev_priv->mm.object_stat_lock); |
73aa808f CW |
100 | } |
101 | ||
102 | static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv, | |
103 | size_t size) | |
104 | { | |
c20e8355 | 105 | spin_lock(&dev_priv->mm.object_stat_lock); |
73aa808f CW |
106 | dev_priv->mm.object_count--; |
107 | dev_priv->mm.object_memory -= size; | |
c20e8355 | 108 | spin_unlock(&dev_priv->mm.object_stat_lock); |
73aa808f CW |
109 | } |
110 | ||
21dd3734 | 111 | static int |
33196ded | 112 | i915_gem_wait_for_error(struct i915_gpu_error *error) |
30dbf0c0 | 113 | { |
30dbf0c0 CW |
114 | int ret; |
115 | ||
7abb690a DV |
116 | #define EXIT_COND (!i915_reset_in_progress(error) || \ |
117 | i915_terminally_wedged(error)) | |
1f83fee0 | 118 | if (EXIT_COND) |
30dbf0c0 CW |
119 | return 0; |
120 | ||
0a6759c6 DV |
121 | /* |
122 | * Only wait 10 seconds for the gpu reset to complete to avoid hanging | |
123 | * userspace. If it takes that long something really bad is going on and | |
124 | * we should simply try to bail out and fail as gracefully as possible. | |
125 | */ | |
1f83fee0 DV |
126 | ret = wait_event_interruptible_timeout(error->reset_queue, |
127 | EXIT_COND, | |
128 | 10*HZ); | |
0a6759c6 DV |
129 | if (ret == 0) { |
130 | DRM_ERROR("Timed out waiting for the gpu reset to complete\n"); | |
131 | return -EIO; | |
132 | } else if (ret < 0) { | |
30dbf0c0 | 133 | return ret; |
0a6759c6 | 134 | } |
1f83fee0 | 135 | #undef EXIT_COND |
30dbf0c0 | 136 | |
21dd3734 | 137 | return 0; |
30dbf0c0 CW |
138 | } |
139 | ||
54cf91dc | 140 | int i915_mutex_lock_interruptible(struct drm_device *dev) |
76c1dec1 | 141 | { |
33196ded | 142 | struct drm_i915_private *dev_priv = dev->dev_private; |
76c1dec1 CW |
143 | int ret; |
144 | ||
33196ded | 145 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
76c1dec1 CW |
146 | if (ret) |
147 | return ret; | |
148 | ||
149 | ret = mutex_lock_interruptible(&dev->struct_mutex); | |
150 | if (ret) | |
151 | return ret; | |
152 | ||
23bc5982 | 153 | WARN_ON(i915_verify_lists(dev)); |
76c1dec1 CW |
154 | return 0; |
155 | } | |
30dbf0c0 | 156 | |
7d1c4804 | 157 | static inline bool |
05394f39 | 158 | i915_gem_object_is_inactive(struct drm_i915_gem_object *obj) |
7d1c4804 | 159 | { |
9843877d | 160 | return i915_gem_obj_bound_any(obj) && !obj->active; |
7d1c4804 CW |
161 | } |
162 | ||
79e53945 JB |
163 | int |
164 | i915_gem_init_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 165 | struct drm_file *file) |
79e53945 | 166 | { |
93d18799 | 167 | struct drm_i915_private *dev_priv = dev->dev_private; |
79e53945 | 168 | struct drm_i915_gem_init *args = data; |
2021746e | 169 | |
7bb6fb8d DV |
170 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
171 | return -ENODEV; | |
172 | ||
2021746e CW |
173 | if (args->gtt_start >= args->gtt_end || |
174 | (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1)) | |
175 | return -EINVAL; | |
79e53945 | 176 | |
f534bc0b DV |
177 | /* GEM with user mode setting was never supported on ilk and later. */ |
178 | if (INTEL_INFO(dev)->gen >= 5) | |
179 | return -ENODEV; | |
180 | ||
79e53945 | 181 | mutex_lock(&dev->struct_mutex); |
d7e5008f BW |
182 | i915_gem_setup_global_gtt(dev, args->gtt_start, args->gtt_end, |
183 | args->gtt_end); | |
93d18799 | 184 | dev_priv->gtt.mappable_end = args->gtt_end; |
673a394b EA |
185 | mutex_unlock(&dev->struct_mutex); |
186 | ||
2021746e | 187 | return 0; |
673a394b EA |
188 | } |
189 | ||
5a125c3c EA |
190 | int |
191 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 192 | struct drm_file *file) |
5a125c3c | 193 | { |
73aa808f | 194 | struct drm_i915_private *dev_priv = dev->dev_private; |
5a125c3c | 195 | struct drm_i915_gem_get_aperture *args = data; |
6299f992 CW |
196 | struct drm_i915_gem_object *obj; |
197 | size_t pinned; | |
5a125c3c | 198 | |
6299f992 | 199 | pinned = 0; |
73aa808f | 200 | mutex_lock(&dev->struct_mutex); |
35c20a60 | 201 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) |
d7f46fc4 | 202 | if (i915_gem_obj_is_pinned(obj)) |
f343c5f6 | 203 | pinned += i915_gem_obj_ggtt_size(obj); |
73aa808f | 204 | mutex_unlock(&dev->struct_mutex); |
5a125c3c | 205 | |
853ba5d2 | 206 | args->aper_size = dev_priv->gtt.base.total; |
0206e353 | 207 | args->aper_available_size = args->aper_size - pinned; |
6299f992 | 208 | |
5a125c3c EA |
209 | return 0; |
210 | } | |
211 | ||
00731155 CW |
212 | static void i915_gem_object_detach_phys(struct drm_i915_gem_object *obj) |
213 | { | |
214 | drm_dma_handle_t *phys = obj->phys_handle; | |
215 | ||
216 | if (!phys) | |
217 | return; | |
218 | ||
219 | if (obj->madv == I915_MADV_WILLNEED) { | |
220 | struct address_space *mapping = file_inode(obj->base.filp)->i_mapping; | |
221 | char *vaddr = phys->vaddr; | |
222 | int i; | |
223 | ||
224 | for (i = 0; i < obj->base.size / PAGE_SIZE; i++) { | |
225 | struct page *page = shmem_read_mapping_page(mapping, i); | |
226 | if (!IS_ERR(page)) { | |
227 | char *dst = kmap_atomic(page); | |
228 | memcpy(dst, vaddr, PAGE_SIZE); | |
229 | drm_clflush_virt_range(dst, PAGE_SIZE); | |
230 | kunmap_atomic(dst); | |
231 | ||
232 | set_page_dirty(page); | |
233 | mark_page_accessed(page); | |
234 | page_cache_release(page); | |
235 | } | |
236 | vaddr += PAGE_SIZE; | |
237 | } | |
238 | i915_gem_chipset_flush(obj->base.dev); | |
239 | } | |
240 | ||
241 | #ifdef CONFIG_X86 | |
242 | set_memory_wb((unsigned long)phys->vaddr, phys->size / PAGE_SIZE); | |
243 | #endif | |
244 | drm_pci_free(obj->base.dev, phys); | |
245 | obj->phys_handle = NULL; | |
246 | } | |
247 | ||
248 | int | |
249 | i915_gem_object_attach_phys(struct drm_i915_gem_object *obj, | |
250 | int align) | |
251 | { | |
252 | drm_dma_handle_t *phys; | |
253 | struct address_space *mapping; | |
254 | char *vaddr; | |
255 | int i; | |
256 | ||
257 | if (obj->phys_handle) { | |
258 | if ((unsigned long)obj->phys_handle->vaddr & (align -1)) | |
259 | return -EBUSY; | |
260 | ||
261 | return 0; | |
262 | } | |
263 | ||
264 | if (obj->madv != I915_MADV_WILLNEED) | |
265 | return -EFAULT; | |
266 | ||
267 | if (obj->base.filp == NULL) | |
268 | return -EINVAL; | |
269 | ||
270 | /* create a new object */ | |
271 | phys = drm_pci_alloc(obj->base.dev, obj->base.size, align); | |
272 | if (!phys) | |
273 | return -ENOMEM; | |
274 | ||
275 | vaddr = phys->vaddr; | |
276 | #ifdef CONFIG_X86 | |
277 | set_memory_wc((unsigned long)vaddr, phys->size / PAGE_SIZE); | |
278 | #endif | |
279 | mapping = file_inode(obj->base.filp)->i_mapping; | |
280 | for (i = 0; i < obj->base.size / PAGE_SIZE; i++) { | |
281 | struct page *page; | |
282 | char *src; | |
283 | ||
284 | page = shmem_read_mapping_page(mapping, i); | |
285 | if (IS_ERR(page)) { | |
286 | #ifdef CONFIG_X86 | |
287 | set_memory_wb((unsigned long)phys->vaddr, phys->size / PAGE_SIZE); | |
288 | #endif | |
289 | drm_pci_free(obj->base.dev, phys); | |
290 | return PTR_ERR(page); | |
291 | } | |
292 | ||
293 | src = kmap_atomic(page); | |
294 | memcpy(vaddr, src, PAGE_SIZE); | |
295 | kunmap_atomic(src); | |
296 | ||
297 | mark_page_accessed(page); | |
298 | page_cache_release(page); | |
299 | ||
300 | vaddr += PAGE_SIZE; | |
301 | } | |
302 | ||
303 | obj->phys_handle = phys; | |
304 | return 0; | |
305 | } | |
306 | ||
307 | static int | |
308 | i915_gem_phys_pwrite(struct drm_i915_gem_object *obj, | |
309 | struct drm_i915_gem_pwrite *args, | |
310 | struct drm_file *file_priv) | |
311 | { | |
312 | struct drm_device *dev = obj->base.dev; | |
313 | void *vaddr = obj->phys_handle->vaddr + args->offset; | |
314 | char __user *user_data = to_user_ptr(args->data_ptr); | |
315 | ||
316 | if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) { | |
317 | unsigned long unwritten; | |
318 | ||
319 | /* The physical object once assigned is fixed for the lifetime | |
320 | * of the obj, so we can safely drop the lock and continue | |
321 | * to access vaddr. | |
322 | */ | |
323 | mutex_unlock(&dev->struct_mutex); | |
324 | unwritten = copy_from_user(vaddr, user_data, args->size); | |
325 | mutex_lock(&dev->struct_mutex); | |
326 | if (unwritten) | |
327 | return -EFAULT; | |
328 | } | |
329 | ||
330 | i915_gem_chipset_flush(dev); | |
331 | return 0; | |
332 | } | |
333 | ||
42dcedd4 CW |
334 | void *i915_gem_object_alloc(struct drm_device *dev) |
335 | { | |
336 | struct drm_i915_private *dev_priv = dev->dev_private; | |
fac15c10 | 337 | return kmem_cache_zalloc(dev_priv->slab, GFP_KERNEL); |
42dcedd4 CW |
338 | } |
339 | ||
340 | void i915_gem_object_free(struct drm_i915_gem_object *obj) | |
341 | { | |
342 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
343 | kmem_cache_free(dev_priv->slab, obj); | |
344 | } | |
345 | ||
ff72145b DA |
346 | static int |
347 | i915_gem_create(struct drm_file *file, | |
348 | struct drm_device *dev, | |
349 | uint64_t size, | |
350 | uint32_t *handle_p) | |
673a394b | 351 | { |
05394f39 | 352 | struct drm_i915_gem_object *obj; |
a1a2d1d3 PP |
353 | int ret; |
354 | u32 handle; | |
673a394b | 355 | |
ff72145b | 356 | size = roundup(size, PAGE_SIZE); |
8ffc0246 CW |
357 | if (size == 0) |
358 | return -EINVAL; | |
673a394b EA |
359 | |
360 | /* Allocate the new object */ | |
ff72145b | 361 | obj = i915_gem_alloc_object(dev, size); |
673a394b EA |
362 | if (obj == NULL) |
363 | return -ENOMEM; | |
364 | ||
05394f39 | 365 | ret = drm_gem_handle_create(file, &obj->base, &handle); |
202f2fef | 366 | /* drop reference from allocate - handle holds it now */ |
d861e338 DV |
367 | drm_gem_object_unreference_unlocked(&obj->base); |
368 | if (ret) | |
369 | return ret; | |
202f2fef | 370 | |
ff72145b | 371 | *handle_p = handle; |
673a394b EA |
372 | return 0; |
373 | } | |
374 | ||
ff72145b DA |
375 | int |
376 | i915_gem_dumb_create(struct drm_file *file, | |
377 | struct drm_device *dev, | |
378 | struct drm_mode_create_dumb *args) | |
379 | { | |
380 | /* have to work out size/pitch and return them */ | |
de45eaf7 | 381 | args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64); |
ff72145b DA |
382 | args->size = args->pitch * args->height; |
383 | return i915_gem_create(file, dev, | |
384 | args->size, &args->handle); | |
385 | } | |
386 | ||
ff72145b DA |
387 | /** |
388 | * Creates a new mm object and returns a handle to it. | |
389 | */ | |
390 | int | |
391 | i915_gem_create_ioctl(struct drm_device *dev, void *data, | |
392 | struct drm_file *file) | |
393 | { | |
394 | struct drm_i915_gem_create *args = data; | |
63ed2cb2 | 395 | |
ff72145b DA |
396 | return i915_gem_create(file, dev, |
397 | args->size, &args->handle); | |
398 | } | |
399 | ||
8461d226 DV |
400 | static inline int |
401 | __copy_to_user_swizzled(char __user *cpu_vaddr, | |
402 | const char *gpu_vaddr, int gpu_offset, | |
403 | int length) | |
404 | { | |
405 | int ret, cpu_offset = 0; | |
406 | ||
407 | while (length > 0) { | |
408 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
409 | int this_length = min(cacheline_end - gpu_offset, length); | |
410 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
411 | ||
412 | ret = __copy_to_user(cpu_vaddr + cpu_offset, | |
413 | gpu_vaddr + swizzled_gpu_offset, | |
414 | this_length); | |
415 | if (ret) | |
416 | return ret + length; | |
417 | ||
418 | cpu_offset += this_length; | |
419 | gpu_offset += this_length; | |
420 | length -= this_length; | |
421 | } | |
422 | ||
423 | return 0; | |
424 | } | |
425 | ||
8c59967c | 426 | static inline int |
4f0c7cfb BW |
427 | __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, |
428 | const char __user *cpu_vaddr, | |
8c59967c DV |
429 | int length) |
430 | { | |
431 | int ret, cpu_offset = 0; | |
432 | ||
433 | while (length > 0) { | |
434 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
435 | int this_length = min(cacheline_end - gpu_offset, length); | |
436 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
437 | ||
438 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, | |
439 | cpu_vaddr + cpu_offset, | |
440 | this_length); | |
441 | if (ret) | |
442 | return ret + length; | |
443 | ||
444 | cpu_offset += this_length; | |
445 | gpu_offset += this_length; | |
446 | length -= this_length; | |
447 | } | |
448 | ||
449 | return 0; | |
450 | } | |
451 | ||
4c914c0c BV |
452 | /* |
453 | * Pins the specified object's pages and synchronizes the object with | |
454 | * GPU accesses. Sets needs_clflush to non-zero if the caller should | |
455 | * flush the object from the CPU cache. | |
456 | */ | |
457 | int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj, | |
458 | int *needs_clflush) | |
459 | { | |
460 | int ret; | |
461 | ||
462 | *needs_clflush = 0; | |
463 | ||
464 | if (!obj->base.filp) | |
465 | return -EINVAL; | |
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. */ | |
472 | *needs_clflush = !cpu_cache_is_coherent(obj->base.dev, | |
473 | obj->cache_level); | |
474 | ret = i915_gem_object_wait_rendering(obj, true); | |
475 | if (ret) | |
476 | return ret; | |
c8725f3d CW |
477 | |
478 | i915_gem_object_retire(obj); | |
4c914c0c BV |
479 | } |
480 | ||
481 | ret = i915_gem_object_get_pages(obj); | |
482 | if (ret) | |
483 | return ret; | |
484 | ||
485 | i915_gem_object_pin_pages(obj); | |
486 | ||
487 | return ret; | |
488 | } | |
489 | ||
d174bd64 DV |
490 | /* Per-page copy function for the shmem pread fastpath. |
491 | * Flushes invalid cachelines before reading the target if | |
492 | * needs_clflush is set. */ | |
eb01459f | 493 | static int |
d174bd64 DV |
494 | shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length, |
495 | char __user *user_data, | |
496 | bool page_do_bit17_swizzling, bool needs_clflush) | |
497 | { | |
498 | char *vaddr; | |
499 | int ret; | |
500 | ||
e7e58eb5 | 501 | if (unlikely(page_do_bit17_swizzling)) |
d174bd64 DV |
502 | return -EINVAL; |
503 | ||
504 | vaddr = kmap_atomic(page); | |
505 | if (needs_clflush) | |
506 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
507 | page_length); | |
508 | ret = __copy_to_user_inatomic(user_data, | |
509 | vaddr + shmem_page_offset, | |
510 | page_length); | |
511 | kunmap_atomic(vaddr); | |
512 | ||
f60d7f0c | 513 | return ret ? -EFAULT : 0; |
d174bd64 DV |
514 | } |
515 | ||
23c18c71 DV |
516 | static void |
517 | shmem_clflush_swizzled_range(char *addr, unsigned long length, | |
518 | bool swizzled) | |
519 | { | |
e7e58eb5 | 520 | if (unlikely(swizzled)) { |
23c18c71 DV |
521 | unsigned long start = (unsigned long) addr; |
522 | unsigned long end = (unsigned long) addr + length; | |
523 | ||
524 | /* For swizzling simply ensure that we always flush both | |
525 | * channels. Lame, but simple and it works. Swizzled | |
526 | * pwrite/pread is far from a hotpath - current userspace | |
527 | * doesn't use it at all. */ | |
528 | start = round_down(start, 128); | |
529 | end = round_up(end, 128); | |
530 | ||
531 | drm_clflush_virt_range((void *)start, end - start); | |
532 | } else { | |
533 | drm_clflush_virt_range(addr, length); | |
534 | } | |
535 | ||
536 | } | |
537 | ||
d174bd64 DV |
538 | /* Only difference to the fast-path function is that this can handle bit17 |
539 | * and uses non-atomic copy and kmap functions. */ | |
540 | static int | |
541 | shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length, | |
542 | char __user *user_data, | |
543 | bool page_do_bit17_swizzling, bool needs_clflush) | |
544 | { | |
545 | char *vaddr; | |
546 | int ret; | |
547 | ||
548 | vaddr = kmap(page); | |
549 | if (needs_clflush) | |
23c18c71 DV |
550 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
551 | page_length, | |
552 | page_do_bit17_swizzling); | |
d174bd64 DV |
553 | |
554 | if (page_do_bit17_swizzling) | |
555 | ret = __copy_to_user_swizzled(user_data, | |
556 | vaddr, shmem_page_offset, | |
557 | page_length); | |
558 | else | |
559 | ret = __copy_to_user(user_data, | |
560 | vaddr + shmem_page_offset, | |
561 | page_length); | |
562 | kunmap(page); | |
563 | ||
f60d7f0c | 564 | return ret ? - EFAULT : 0; |
d174bd64 DV |
565 | } |
566 | ||
eb01459f | 567 | static int |
dbf7bff0 DV |
568 | i915_gem_shmem_pread(struct drm_device *dev, |
569 | struct drm_i915_gem_object *obj, | |
570 | struct drm_i915_gem_pread *args, | |
571 | struct drm_file *file) | |
eb01459f | 572 | { |
8461d226 | 573 | char __user *user_data; |
eb01459f | 574 | ssize_t remain; |
8461d226 | 575 | loff_t offset; |
eb2c0c81 | 576 | int shmem_page_offset, page_length, ret = 0; |
8461d226 | 577 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
96d79b52 | 578 | int prefaulted = 0; |
8489731c | 579 | int needs_clflush = 0; |
67d5a50c | 580 | struct sg_page_iter sg_iter; |
eb01459f | 581 | |
2bb4629a | 582 | user_data = to_user_ptr(args->data_ptr); |
eb01459f EA |
583 | remain = args->size; |
584 | ||
8461d226 | 585 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
eb01459f | 586 | |
4c914c0c | 587 | ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush); |
f60d7f0c CW |
588 | if (ret) |
589 | return ret; | |
590 | ||
8461d226 | 591 | offset = args->offset; |
eb01459f | 592 | |
67d5a50c ID |
593 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
594 | offset >> PAGE_SHIFT) { | |
2db76d7c | 595 | struct page *page = sg_page_iter_page(&sg_iter); |
9da3da66 CW |
596 | |
597 | if (remain <= 0) | |
598 | break; | |
599 | ||
eb01459f EA |
600 | /* Operation in this page |
601 | * | |
eb01459f | 602 | * shmem_page_offset = offset within page in shmem file |
eb01459f EA |
603 | * page_length = bytes to copy for this page |
604 | */ | |
c8cbbb8b | 605 | shmem_page_offset = offset_in_page(offset); |
eb01459f EA |
606 | page_length = remain; |
607 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
608 | page_length = PAGE_SIZE - shmem_page_offset; | |
eb01459f | 609 | |
8461d226 DV |
610 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
611 | (page_to_phys(page) & (1 << 17)) != 0; | |
612 | ||
d174bd64 DV |
613 | ret = shmem_pread_fast(page, shmem_page_offset, page_length, |
614 | user_data, page_do_bit17_swizzling, | |
615 | needs_clflush); | |
616 | if (ret == 0) | |
617 | goto next_page; | |
dbf7bff0 | 618 | |
dbf7bff0 DV |
619 | mutex_unlock(&dev->struct_mutex); |
620 | ||
d330a953 | 621 | if (likely(!i915.prefault_disable) && !prefaulted) { |
f56f821f | 622 | ret = fault_in_multipages_writeable(user_data, remain); |
96d79b52 DV |
623 | /* Userspace is tricking us, but we've already clobbered |
624 | * its pages with the prefault and promised to write the | |
625 | * data up to the first fault. Hence ignore any errors | |
626 | * and just continue. */ | |
627 | (void)ret; | |
628 | prefaulted = 1; | |
629 | } | |
eb01459f | 630 | |
d174bd64 DV |
631 | ret = shmem_pread_slow(page, shmem_page_offset, page_length, |
632 | user_data, page_do_bit17_swizzling, | |
633 | needs_clflush); | |
eb01459f | 634 | |
dbf7bff0 | 635 | mutex_lock(&dev->struct_mutex); |
f60d7f0c | 636 | |
f60d7f0c | 637 | if (ret) |
8461d226 | 638 | goto out; |
8461d226 | 639 | |
17793c9a | 640 | next_page: |
eb01459f | 641 | remain -= page_length; |
8461d226 | 642 | user_data += page_length; |
eb01459f EA |
643 | offset += page_length; |
644 | } | |
645 | ||
4f27b75d | 646 | out: |
f60d7f0c CW |
647 | i915_gem_object_unpin_pages(obj); |
648 | ||
eb01459f EA |
649 | return ret; |
650 | } | |
651 | ||
673a394b EA |
652 | /** |
653 | * Reads data from the object referenced by handle. | |
654 | * | |
655 | * On error, the contents of *data are undefined. | |
656 | */ | |
657 | int | |
658 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 659 | struct drm_file *file) |
673a394b EA |
660 | { |
661 | struct drm_i915_gem_pread *args = data; | |
05394f39 | 662 | struct drm_i915_gem_object *obj; |
35b62a89 | 663 | int ret = 0; |
673a394b | 664 | |
51311d0a CW |
665 | if (args->size == 0) |
666 | return 0; | |
667 | ||
668 | if (!access_ok(VERIFY_WRITE, | |
2bb4629a | 669 | to_user_ptr(args->data_ptr), |
51311d0a CW |
670 | args->size)) |
671 | return -EFAULT; | |
672 | ||
4f27b75d | 673 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 674 | if (ret) |
4f27b75d | 675 | return ret; |
673a394b | 676 | |
05394f39 | 677 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 678 | if (&obj->base == NULL) { |
1d7cfea1 CW |
679 | ret = -ENOENT; |
680 | goto unlock; | |
4f27b75d | 681 | } |
673a394b | 682 | |
7dcd2499 | 683 | /* Bounds check source. */ |
05394f39 CW |
684 | if (args->offset > obj->base.size || |
685 | args->size > obj->base.size - args->offset) { | |
ce9d419d | 686 | ret = -EINVAL; |
35b62a89 | 687 | goto out; |
ce9d419d CW |
688 | } |
689 | ||
1286ff73 DV |
690 | /* prime objects have no backing filp to GEM pread/pwrite |
691 | * pages from. | |
692 | */ | |
693 | if (!obj->base.filp) { | |
694 | ret = -EINVAL; | |
695 | goto out; | |
696 | } | |
697 | ||
db53a302 CW |
698 | trace_i915_gem_object_pread(obj, args->offset, args->size); |
699 | ||
dbf7bff0 | 700 | ret = i915_gem_shmem_pread(dev, obj, args, file); |
673a394b | 701 | |
35b62a89 | 702 | out: |
05394f39 | 703 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 704 | unlock: |
4f27b75d | 705 | mutex_unlock(&dev->struct_mutex); |
eb01459f | 706 | return ret; |
673a394b EA |
707 | } |
708 | ||
0839ccb8 KP |
709 | /* This is the fast write path which cannot handle |
710 | * page faults in the source data | |
9b7530cc | 711 | */ |
0839ccb8 KP |
712 | |
713 | static inline int | |
714 | fast_user_write(struct io_mapping *mapping, | |
715 | loff_t page_base, int page_offset, | |
716 | char __user *user_data, | |
717 | int length) | |
9b7530cc | 718 | { |
4f0c7cfb BW |
719 | void __iomem *vaddr_atomic; |
720 | void *vaddr; | |
0839ccb8 | 721 | unsigned long unwritten; |
9b7530cc | 722 | |
3e4d3af5 | 723 | vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base); |
4f0c7cfb BW |
724 | /* We can use the cpu mem copy function because this is X86. */ |
725 | vaddr = (void __force*)vaddr_atomic + page_offset; | |
726 | unwritten = __copy_from_user_inatomic_nocache(vaddr, | |
0839ccb8 | 727 | user_data, length); |
3e4d3af5 | 728 | io_mapping_unmap_atomic(vaddr_atomic); |
fbd5a26d | 729 | return unwritten; |
0839ccb8 KP |
730 | } |
731 | ||
3de09aa3 EA |
732 | /** |
733 | * This is the fast pwrite path, where we copy the data directly from the | |
734 | * user into the GTT, uncached. | |
735 | */ | |
673a394b | 736 | static int |
05394f39 CW |
737 | i915_gem_gtt_pwrite_fast(struct drm_device *dev, |
738 | struct drm_i915_gem_object *obj, | |
3de09aa3 | 739 | struct drm_i915_gem_pwrite *args, |
05394f39 | 740 | struct drm_file *file) |
673a394b | 741 | { |
3e31c6c0 | 742 | struct drm_i915_private *dev_priv = dev->dev_private; |
673a394b | 743 | ssize_t remain; |
0839ccb8 | 744 | loff_t offset, page_base; |
673a394b | 745 | char __user *user_data; |
935aaa69 DV |
746 | int page_offset, page_length, ret; |
747 | ||
1ec9e26d | 748 | ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK); |
935aaa69 DV |
749 | if (ret) |
750 | goto out; | |
751 | ||
752 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
753 | if (ret) | |
754 | goto out_unpin; | |
755 | ||
756 | ret = i915_gem_object_put_fence(obj); | |
757 | if (ret) | |
758 | goto out_unpin; | |
673a394b | 759 | |
2bb4629a | 760 | user_data = to_user_ptr(args->data_ptr); |
673a394b | 761 | remain = args->size; |
673a394b | 762 | |
f343c5f6 | 763 | offset = i915_gem_obj_ggtt_offset(obj) + args->offset; |
673a394b EA |
764 | |
765 | while (remain > 0) { | |
766 | /* Operation in this page | |
767 | * | |
0839ccb8 KP |
768 | * page_base = page offset within aperture |
769 | * page_offset = offset within page | |
770 | * page_length = bytes to copy for this page | |
673a394b | 771 | */ |
c8cbbb8b CW |
772 | page_base = offset & PAGE_MASK; |
773 | page_offset = offset_in_page(offset); | |
0839ccb8 KP |
774 | page_length = remain; |
775 | if ((page_offset + remain) > PAGE_SIZE) | |
776 | page_length = PAGE_SIZE - page_offset; | |
777 | ||
0839ccb8 | 778 | /* If we get a fault while copying data, then (presumably) our |
3de09aa3 EA |
779 | * source page isn't available. Return the error and we'll |
780 | * retry in the slow path. | |
0839ccb8 | 781 | */ |
5d4545ae | 782 | if (fast_user_write(dev_priv->gtt.mappable, page_base, |
935aaa69 DV |
783 | page_offset, user_data, page_length)) { |
784 | ret = -EFAULT; | |
785 | goto out_unpin; | |
786 | } | |
673a394b | 787 | |
0839ccb8 KP |
788 | remain -= page_length; |
789 | user_data += page_length; | |
790 | offset += page_length; | |
673a394b | 791 | } |
673a394b | 792 | |
935aaa69 | 793 | out_unpin: |
d7f46fc4 | 794 | i915_gem_object_ggtt_unpin(obj); |
935aaa69 | 795 | out: |
3de09aa3 | 796 | return ret; |
673a394b EA |
797 | } |
798 | ||
d174bd64 DV |
799 | /* Per-page copy function for the shmem pwrite fastpath. |
800 | * Flushes invalid cachelines before writing to the target if | |
801 | * needs_clflush_before is set and flushes out any written cachelines after | |
802 | * writing if needs_clflush is set. */ | |
3043c60c | 803 | static int |
d174bd64 DV |
804 | shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length, |
805 | char __user *user_data, | |
806 | bool page_do_bit17_swizzling, | |
807 | bool needs_clflush_before, | |
808 | bool needs_clflush_after) | |
673a394b | 809 | { |
d174bd64 | 810 | char *vaddr; |
673a394b | 811 | int ret; |
3de09aa3 | 812 | |
e7e58eb5 | 813 | if (unlikely(page_do_bit17_swizzling)) |
d174bd64 | 814 | return -EINVAL; |
3de09aa3 | 815 | |
d174bd64 DV |
816 | vaddr = kmap_atomic(page); |
817 | if (needs_clflush_before) | |
818 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
819 | page_length); | |
c2831a94 CW |
820 | ret = __copy_from_user_inatomic(vaddr + shmem_page_offset, |
821 | user_data, page_length); | |
d174bd64 DV |
822 | if (needs_clflush_after) |
823 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
824 | page_length); | |
825 | kunmap_atomic(vaddr); | |
3de09aa3 | 826 | |
755d2218 | 827 | return ret ? -EFAULT : 0; |
3de09aa3 EA |
828 | } |
829 | ||
d174bd64 DV |
830 | /* Only difference to the fast-path function is that this can handle bit17 |
831 | * and uses non-atomic copy and kmap functions. */ | |
3043c60c | 832 | static int |
d174bd64 DV |
833 | shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length, |
834 | char __user *user_data, | |
835 | bool page_do_bit17_swizzling, | |
836 | bool needs_clflush_before, | |
837 | bool needs_clflush_after) | |
673a394b | 838 | { |
d174bd64 DV |
839 | char *vaddr; |
840 | int ret; | |
e5281ccd | 841 | |
d174bd64 | 842 | vaddr = kmap(page); |
e7e58eb5 | 843 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) |
23c18c71 DV |
844 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
845 | page_length, | |
846 | page_do_bit17_swizzling); | |
d174bd64 DV |
847 | if (page_do_bit17_swizzling) |
848 | ret = __copy_from_user_swizzled(vaddr, shmem_page_offset, | |
e5281ccd CW |
849 | user_data, |
850 | page_length); | |
d174bd64 DV |
851 | else |
852 | ret = __copy_from_user(vaddr + shmem_page_offset, | |
853 | user_data, | |
854 | page_length); | |
855 | if (needs_clflush_after) | |
23c18c71 DV |
856 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
857 | page_length, | |
858 | page_do_bit17_swizzling); | |
d174bd64 | 859 | kunmap(page); |
40123c1f | 860 | |
755d2218 | 861 | return ret ? -EFAULT : 0; |
40123c1f EA |
862 | } |
863 | ||
40123c1f | 864 | static int |
e244a443 DV |
865 | i915_gem_shmem_pwrite(struct drm_device *dev, |
866 | struct drm_i915_gem_object *obj, | |
867 | struct drm_i915_gem_pwrite *args, | |
868 | struct drm_file *file) | |
40123c1f | 869 | { |
40123c1f | 870 | ssize_t remain; |
8c59967c DV |
871 | loff_t offset; |
872 | char __user *user_data; | |
eb2c0c81 | 873 | int shmem_page_offset, page_length, ret = 0; |
8c59967c | 874 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
e244a443 | 875 | int hit_slowpath = 0; |
58642885 DV |
876 | int needs_clflush_after = 0; |
877 | int needs_clflush_before = 0; | |
67d5a50c | 878 | struct sg_page_iter sg_iter; |
40123c1f | 879 | |
2bb4629a | 880 | user_data = to_user_ptr(args->data_ptr); |
40123c1f EA |
881 | remain = args->size; |
882 | ||
8c59967c | 883 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
40123c1f | 884 | |
58642885 DV |
885 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
886 | /* If we're not in the cpu write domain, set ourself into the gtt | |
887 | * write domain and manually flush cachelines (if required). This | |
888 | * optimizes for the case when the gpu will use the data | |
889 | * right away and we therefore have to clflush anyway. */ | |
2c22569b | 890 | needs_clflush_after = cpu_write_needs_clflush(obj); |
23f54483 BW |
891 | ret = i915_gem_object_wait_rendering(obj, false); |
892 | if (ret) | |
893 | return ret; | |
c8725f3d CW |
894 | |
895 | i915_gem_object_retire(obj); | |
58642885 | 896 | } |
c76ce038 CW |
897 | /* Same trick applies to invalidate partially written cachelines read |
898 | * before writing. */ | |
899 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) | |
900 | needs_clflush_before = | |
901 | !cpu_cache_is_coherent(dev, obj->cache_level); | |
58642885 | 902 | |
755d2218 CW |
903 | ret = i915_gem_object_get_pages(obj); |
904 | if (ret) | |
905 | return ret; | |
906 | ||
907 | i915_gem_object_pin_pages(obj); | |
908 | ||
673a394b | 909 | offset = args->offset; |
05394f39 | 910 | obj->dirty = 1; |
673a394b | 911 | |
67d5a50c ID |
912 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
913 | offset >> PAGE_SHIFT) { | |
2db76d7c | 914 | struct page *page = sg_page_iter_page(&sg_iter); |
58642885 | 915 | int partial_cacheline_write; |
e5281ccd | 916 | |
9da3da66 CW |
917 | if (remain <= 0) |
918 | break; | |
919 | ||
40123c1f EA |
920 | /* Operation in this page |
921 | * | |
40123c1f | 922 | * shmem_page_offset = offset within page in shmem file |
40123c1f EA |
923 | * page_length = bytes to copy for this page |
924 | */ | |
c8cbbb8b | 925 | shmem_page_offset = offset_in_page(offset); |
40123c1f EA |
926 | |
927 | page_length = remain; | |
928 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
929 | page_length = PAGE_SIZE - shmem_page_offset; | |
40123c1f | 930 | |
58642885 DV |
931 | /* If we don't overwrite a cacheline completely we need to be |
932 | * careful to have up-to-date data by first clflushing. Don't | |
933 | * overcomplicate things and flush the entire patch. */ | |
934 | partial_cacheline_write = needs_clflush_before && | |
935 | ((shmem_page_offset | page_length) | |
936 | & (boot_cpu_data.x86_clflush_size - 1)); | |
937 | ||
8c59967c DV |
938 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
939 | (page_to_phys(page) & (1 << 17)) != 0; | |
940 | ||
d174bd64 DV |
941 | ret = shmem_pwrite_fast(page, shmem_page_offset, page_length, |
942 | user_data, page_do_bit17_swizzling, | |
943 | partial_cacheline_write, | |
944 | needs_clflush_after); | |
945 | if (ret == 0) | |
946 | goto next_page; | |
e244a443 DV |
947 | |
948 | hit_slowpath = 1; | |
e244a443 | 949 | mutex_unlock(&dev->struct_mutex); |
d174bd64 DV |
950 | ret = shmem_pwrite_slow(page, shmem_page_offset, page_length, |
951 | user_data, page_do_bit17_swizzling, | |
952 | partial_cacheline_write, | |
953 | needs_clflush_after); | |
40123c1f | 954 | |
e244a443 | 955 | mutex_lock(&dev->struct_mutex); |
755d2218 | 956 | |
755d2218 | 957 | if (ret) |
8c59967c | 958 | goto out; |
8c59967c | 959 | |
17793c9a | 960 | next_page: |
40123c1f | 961 | remain -= page_length; |
8c59967c | 962 | user_data += page_length; |
40123c1f | 963 | offset += page_length; |
673a394b EA |
964 | } |
965 | ||
fbd5a26d | 966 | out: |
755d2218 CW |
967 | i915_gem_object_unpin_pages(obj); |
968 | ||
e244a443 | 969 | if (hit_slowpath) { |
8dcf015e DV |
970 | /* |
971 | * Fixup: Flush cpu caches in case we didn't flush the dirty | |
972 | * cachelines in-line while writing and the object moved | |
973 | * out of the cpu write domain while we've dropped the lock. | |
974 | */ | |
975 | if (!needs_clflush_after && | |
976 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { | |
000433b6 CW |
977 | if (i915_gem_clflush_object(obj, obj->pin_display)) |
978 | i915_gem_chipset_flush(dev); | |
e244a443 | 979 | } |
8c59967c | 980 | } |
673a394b | 981 | |
58642885 | 982 | if (needs_clflush_after) |
e76e9aeb | 983 | i915_gem_chipset_flush(dev); |
58642885 | 984 | |
40123c1f | 985 | return ret; |
673a394b EA |
986 | } |
987 | ||
988 | /** | |
989 | * Writes data to the object referenced by handle. | |
990 | * | |
991 | * On error, the contents of the buffer that were to be modified are undefined. | |
992 | */ | |
993 | int | |
994 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, | |
fbd5a26d | 995 | struct drm_file *file) |
673a394b EA |
996 | { |
997 | struct drm_i915_gem_pwrite *args = data; | |
05394f39 | 998 | struct drm_i915_gem_object *obj; |
51311d0a CW |
999 | int ret; |
1000 | ||
1001 | if (args->size == 0) | |
1002 | return 0; | |
1003 | ||
1004 | if (!access_ok(VERIFY_READ, | |
2bb4629a | 1005 | to_user_ptr(args->data_ptr), |
51311d0a CW |
1006 | args->size)) |
1007 | return -EFAULT; | |
1008 | ||
d330a953 | 1009 | if (likely(!i915.prefault_disable)) { |
0b74b508 XZ |
1010 | ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr), |
1011 | args->size); | |
1012 | if (ret) | |
1013 | return -EFAULT; | |
1014 | } | |
673a394b | 1015 | |
fbd5a26d | 1016 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1017 | if (ret) |
fbd5a26d | 1018 | return ret; |
1d7cfea1 | 1019 | |
05394f39 | 1020 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 1021 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1022 | ret = -ENOENT; |
1023 | goto unlock; | |
fbd5a26d | 1024 | } |
673a394b | 1025 | |
7dcd2499 | 1026 | /* Bounds check destination. */ |
05394f39 CW |
1027 | if (args->offset > obj->base.size || |
1028 | args->size > obj->base.size - args->offset) { | |
ce9d419d | 1029 | ret = -EINVAL; |
35b62a89 | 1030 | goto out; |
ce9d419d CW |
1031 | } |
1032 | ||
1286ff73 DV |
1033 | /* prime objects have no backing filp to GEM pread/pwrite |
1034 | * pages from. | |
1035 | */ | |
1036 | if (!obj->base.filp) { | |
1037 | ret = -EINVAL; | |
1038 | goto out; | |
1039 | } | |
1040 | ||
db53a302 CW |
1041 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); |
1042 | ||
935aaa69 | 1043 | ret = -EFAULT; |
673a394b EA |
1044 | /* We can only do the GTT pwrite on untiled buffers, as otherwise |
1045 | * it would end up going through the fenced access, and we'll get | |
1046 | * different detiling behavior between reading and writing. | |
1047 | * pread/pwrite currently are reading and writing from the CPU | |
1048 | * perspective, requiring manual detiling by the client. | |
1049 | */ | |
00731155 CW |
1050 | if (obj->phys_handle) { |
1051 | ret = i915_gem_phys_pwrite(obj, args, file); | |
5c0480f2 DV |
1052 | goto out; |
1053 | } | |
1054 | ||
2c22569b CW |
1055 | if (obj->tiling_mode == I915_TILING_NONE && |
1056 | obj->base.write_domain != I915_GEM_DOMAIN_CPU && | |
1057 | cpu_write_needs_clflush(obj)) { | |
fbd5a26d | 1058 | ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); |
935aaa69 DV |
1059 | /* Note that the gtt paths might fail with non-page-backed user |
1060 | * pointers (e.g. gtt mappings when moving data between | |
1061 | * textures). Fallback to the shmem path in that case. */ | |
fbd5a26d | 1062 | } |
673a394b | 1063 | |
86a1ee26 | 1064 | if (ret == -EFAULT || ret == -ENOSPC) |
935aaa69 | 1065 | ret = i915_gem_shmem_pwrite(dev, obj, args, file); |
5c0480f2 | 1066 | |
35b62a89 | 1067 | out: |
05394f39 | 1068 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1069 | unlock: |
fbd5a26d | 1070 | mutex_unlock(&dev->struct_mutex); |
673a394b EA |
1071 | return ret; |
1072 | } | |
1073 | ||
b361237b | 1074 | int |
33196ded | 1075 | i915_gem_check_wedge(struct i915_gpu_error *error, |
b361237b CW |
1076 | bool interruptible) |
1077 | { | |
1f83fee0 | 1078 | if (i915_reset_in_progress(error)) { |
b361237b CW |
1079 | /* Non-interruptible callers can't handle -EAGAIN, hence return |
1080 | * -EIO unconditionally for these. */ | |
1081 | if (!interruptible) | |
1082 | return -EIO; | |
1083 | ||
1f83fee0 DV |
1084 | /* Recovery complete, but the reset failed ... */ |
1085 | if (i915_terminally_wedged(error)) | |
b361237b CW |
1086 | return -EIO; |
1087 | ||
6689c167 MA |
1088 | /* |
1089 | * Check if GPU Reset is in progress - we need intel_ring_begin | |
1090 | * to work properly to reinit the hw state while the gpu is | |
1091 | * still marked as reset-in-progress. Handle this with a flag. | |
1092 | */ | |
1093 | if (!error->reload_in_reset) | |
1094 | return -EAGAIN; | |
b361237b CW |
1095 | } |
1096 | ||
1097 | return 0; | |
1098 | } | |
1099 | ||
1100 | /* | |
1101 | * Compare seqno against outstanding lazy request. Emit a request if they are | |
1102 | * equal. | |
1103 | */ | |
84c33a64 | 1104 | int |
a4872ba6 | 1105 | i915_gem_check_olr(struct intel_engine_cs *ring, u32 seqno) |
b361237b CW |
1106 | { |
1107 | int ret; | |
1108 | ||
1109 | BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex)); | |
1110 | ||
1111 | ret = 0; | |
1823521d | 1112 | if (seqno == ring->outstanding_lazy_seqno) |
0025c077 | 1113 | ret = i915_add_request(ring, NULL); |
b361237b CW |
1114 | |
1115 | return ret; | |
1116 | } | |
1117 | ||
094f9a54 CW |
1118 | static void fake_irq(unsigned long data) |
1119 | { | |
1120 | wake_up_process((struct task_struct *)data); | |
1121 | } | |
1122 | ||
1123 | static bool missed_irq(struct drm_i915_private *dev_priv, | |
a4872ba6 | 1124 | struct intel_engine_cs *ring) |
094f9a54 CW |
1125 | { |
1126 | return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings); | |
1127 | } | |
1128 | ||
b29c19b6 CW |
1129 | static bool can_wait_boost(struct drm_i915_file_private *file_priv) |
1130 | { | |
1131 | if (file_priv == NULL) | |
1132 | return true; | |
1133 | ||
1134 | return !atomic_xchg(&file_priv->rps_wait_boost, true); | |
1135 | } | |
1136 | ||
b361237b CW |
1137 | /** |
1138 | * __wait_seqno - wait until execution of seqno has finished | |
1139 | * @ring: the ring expected to report seqno | |
1140 | * @seqno: duh! | |
f69061be | 1141 | * @reset_counter: reset sequence associated with the given seqno |
b361237b CW |
1142 | * @interruptible: do an interruptible wait (normally yes) |
1143 | * @timeout: in - how long to wait (NULL forever); out - how much time remaining | |
1144 | * | |
f69061be DV |
1145 | * Note: It is of utmost importance that the passed in seqno and reset_counter |
1146 | * values have been read by the caller in an smp safe manner. Where read-side | |
1147 | * locks are involved, it is sufficient to read the reset_counter before | |
1148 | * unlocking the lock that protects the seqno. For lockless tricks, the | |
1149 | * reset_counter _must_ be read before, and an appropriate smp_rmb must be | |
1150 | * inserted. | |
1151 | * | |
b361237b CW |
1152 | * Returns 0 if the seqno was found within the alloted time. Else returns the |
1153 | * errno with remaining time filled in timeout argument. | |
1154 | */ | |
a4872ba6 | 1155 | static int __wait_seqno(struct intel_engine_cs *ring, u32 seqno, |
f69061be | 1156 | unsigned reset_counter, |
b29c19b6 | 1157 | bool interruptible, |
5ed0bdf2 | 1158 | s64 *timeout, |
b29c19b6 | 1159 | struct drm_i915_file_private *file_priv) |
b361237b | 1160 | { |
3d13ef2e | 1161 | struct drm_device *dev = ring->dev; |
3e31c6c0 | 1162 | struct drm_i915_private *dev_priv = dev->dev_private; |
168c3f21 MK |
1163 | const bool irq_test_in_progress = |
1164 | ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring); | |
094f9a54 | 1165 | DEFINE_WAIT(wait); |
47e9766d | 1166 | unsigned long timeout_expire; |
5ed0bdf2 | 1167 | s64 before, now; |
b361237b CW |
1168 | int ret; |
1169 | ||
9df7575f | 1170 | WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled"); |
c67a470b | 1171 | |
b361237b CW |
1172 | if (i915_seqno_passed(ring->get_seqno(ring, true), seqno)) |
1173 | return 0; | |
1174 | ||
5ed0bdf2 | 1175 | timeout_expire = timeout ? jiffies + nsecs_to_jiffies((u64)*timeout) : 0; |
b361237b | 1176 | |
ec5cc0f9 | 1177 | if (INTEL_INFO(dev)->gen >= 6 && ring->id == RCS && can_wait_boost(file_priv)) { |
b29c19b6 CW |
1178 | gen6_rps_boost(dev_priv); |
1179 | if (file_priv) | |
1180 | mod_delayed_work(dev_priv->wq, | |
1181 | &file_priv->mm.idle_work, | |
1182 | msecs_to_jiffies(100)); | |
1183 | } | |
1184 | ||
168c3f21 | 1185 | if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring))) |
b361237b CW |
1186 | return -ENODEV; |
1187 | ||
094f9a54 CW |
1188 | /* Record current time in case interrupted by signal, or wedged */ |
1189 | trace_i915_gem_request_wait_begin(ring, seqno); | |
5ed0bdf2 | 1190 | before = ktime_get_raw_ns(); |
094f9a54 CW |
1191 | for (;;) { |
1192 | struct timer_list timer; | |
b361237b | 1193 | |
094f9a54 CW |
1194 | prepare_to_wait(&ring->irq_queue, &wait, |
1195 | interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE); | |
b361237b | 1196 | |
f69061be DV |
1197 | /* We need to check whether any gpu reset happened in between |
1198 | * the caller grabbing the seqno and now ... */ | |
094f9a54 CW |
1199 | if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) { |
1200 | /* ... but upgrade the -EAGAIN to an -EIO if the gpu | |
1201 | * is truely gone. */ | |
1202 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); | |
1203 | if (ret == 0) | |
1204 | ret = -EAGAIN; | |
1205 | break; | |
1206 | } | |
f69061be | 1207 | |
094f9a54 CW |
1208 | if (i915_seqno_passed(ring->get_seqno(ring, false), seqno)) { |
1209 | ret = 0; | |
1210 | break; | |
1211 | } | |
b361237b | 1212 | |
094f9a54 CW |
1213 | if (interruptible && signal_pending(current)) { |
1214 | ret = -ERESTARTSYS; | |
1215 | break; | |
1216 | } | |
1217 | ||
47e9766d | 1218 | if (timeout && time_after_eq(jiffies, timeout_expire)) { |
094f9a54 CW |
1219 | ret = -ETIME; |
1220 | break; | |
1221 | } | |
1222 | ||
1223 | timer.function = NULL; | |
1224 | if (timeout || missed_irq(dev_priv, ring)) { | |
47e9766d MK |
1225 | unsigned long expire; |
1226 | ||
094f9a54 | 1227 | setup_timer_on_stack(&timer, fake_irq, (unsigned long)current); |
47e9766d | 1228 | expire = missed_irq(dev_priv, ring) ? jiffies + 1 : timeout_expire; |
094f9a54 CW |
1229 | mod_timer(&timer, expire); |
1230 | } | |
1231 | ||
5035c275 | 1232 | io_schedule(); |
094f9a54 | 1233 | |
094f9a54 CW |
1234 | if (timer.function) { |
1235 | del_singleshot_timer_sync(&timer); | |
1236 | destroy_timer_on_stack(&timer); | |
1237 | } | |
1238 | } | |
5ed0bdf2 | 1239 | now = ktime_get_raw_ns(); |
094f9a54 | 1240 | trace_i915_gem_request_wait_end(ring, seqno); |
b361237b | 1241 | |
168c3f21 MK |
1242 | if (!irq_test_in_progress) |
1243 | ring->irq_put(ring); | |
094f9a54 CW |
1244 | |
1245 | finish_wait(&ring->irq_queue, &wait); | |
b361237b CW |
1246 | |
1247 | if (timeout) { | |
5ed0bdf2 TG |
1248 | s64 tres = *timeout - (now - before); |
1249 | ||
1250 | *timeout = tres < 0 ? 0 : tres; | |
b361237b CW |
1251 | } |
1252 | ||
094f9a54 | 1253 | return ret; |
b361237b CW |
1254 | } |
1255 | ||
1256 | /** | |
1257 | * Waits for a sequence number to be signaled, and cleans up the | |
1258 | * request and object lists appropriately for that event. | |
1259 | */ | |
1260 | int | |
a4872ba6 | 1261 | i915_wait_seqno(struct intel_engine_cs *ring, uint32_t seqno) |
b361237b CW |
1262 | { |
1263 | struct drm_device *dev = ring->dev; | |
1264 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1265 | bool interruptible = dev_priv->mm.interruptible; | |
1266 | int ret; | |
1267 | ||
1268 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); | |
1269 | BUG_ON(seqno == 0); | |
1270 | ||
33196ded | 1271 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
b361237b CW |
1272 | if (ret) |
1273 | return ret; | |
1274 | ||
1275 | ret = i915_gem_check_olr(ring, seqno); | |
1276 | if (ret) | |
1277 | return ret; | |
1278 | ||
f69061be DV |
1279 | return __wait_seqno(ring, seqno, |
1280 | atomic_read(&dev_priv->gpu_error.reset_counter), | |
b29c19b6 | 1281 | interruptible, NULL, NULL); |
b361237b CW |
1282 | } |
1283 | ||
d26e3af8 CW |
1284 | static int |
1285 | i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj, | |
a4872ba6 | 1286 | struct intel_engine_cs *ring) |
d26e3af8 | 1287 | { |
c8725f3d CW |
1288 | if (!obj->active) |
1289 | return 0; | |
d26e3af8 CW |
1290 | |
1291 | /* Manually manage the write flush as we may have not yet | |
1292 | * retired the buffer. | |
1293 | * | |
1294 | * Note that the last_write_seqno is always the earlier of | |
1295 | * the two (read/write) seqno, so if we haved successfully waited, | |
1296 | * we know we have passed the last write. | |
1297 | */ | |
1298 | obj->last_write_seqno = 0; | |
d26e3af8 CW |
1299 | |
1300 | return 0; | |
1301 | } | |
1302 | ||
b361237b CW |
1303 | /** |
1304 | * Ensures that all rendering to the object has completed and the object is | |
1305 | * safe to unbind from the GTT or access from the CPU. | |
1306 | */ | |
1307 | static __must_check int | |
1308 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj, | |
1309 | bool readonly) | |
1310 | { | |
a4872ba6 | 1311 | struct intel_engine_cs *ring = obj->ring; |
b361237b CW |
1312 | u32 seqno; |
1313 | int ret; | |
1314 | ||
1315 | seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno; | |
1316 | if (seqno == 0) | |
1317 | return 0; | |
1318 | ||
1319 | ret = i915_wait_seqno(ring, seqno); | |
1320 | if (ret) | |
1321 | return ret; | |
1322 | ||
d26e3af8 | 1323 | return i915_gem_object_wait_rendering__tail(obj, ring); |
b361237b CW |
1324 | } |
1325 | ||
3236f57a CW |
1326 | /* A nonblocking variant of the above wait. This is a highly dangerous routine |
1327 | * as the object state may change during this call. | |
1328 | */ | |
1329 | static __must_check int | |
1330 | i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj, | |
6e4930f6 | 1331 | struct drm_i915_file_private *file_priv, |
3236f57a CW |
1332 | bool readonly) |
1333 | { | |
1334 | struct drm_device *dev = obj->base.dev; | |
1335 | struct drm_i915_private *dev_priv = dev->dev_private; | |
a4872ba6 | 1336 | struct intel_engine_cs *ring = obj->ring; |
f69061be | 1337 | unsigned reset_counter; |
3236f57a CW |
1338 | u32 seqno; |
1339 | int ret; | |
1340 | ||
1341 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); | |
1342 | BUG_ON(!dev_priv->mm.interruptible); | |
1343 | ||
1344 | seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno; | |
1345 | if (seqno == 0) | |
1346 | return 0; | |
1347 | ||
33196ded | 1348 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, true); |
3236f57a CW |
1349 | if (ret) |
1350 | return ret; | |
1351 | ||
1352 | ret = i915_gem_check_olr(ring, seqno); | |
1353 | if (ret) | |
1354 | return ret; | |
1355 | ||
f69061be | 1356 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
3236f57a | 1357 | mutex_unlock(&dev->struct_mutex); |
6e4930f6 | 1358 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL, file_priv); |
3236f57a | 1359 | mutex_lock(&dev->struct_mutex); |
d26e3af8 CW |
1360 | if (ret) |
1361 | return ret; | |
3236f57a | 1362 | |
d26e3af8 | 1363 | return i915_gem_object_wait_rendering__tail(obj, ring); |
3236f57a CW |
1364 | } |
1365 | ||
673a394b | 1366 | /** |
2ef7eeaa EA |
1367 | * Called when user space prepares to use an object with the CPU, either |
1368 | * through the mmap ioctl's mapping or a GTT mapping. | |
673a394b EA |
1369 | */ |
1370 | int | |
1371 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1372 | struct drm_file *file) |
673a394b EA |
1373 | { |
1374 | struct drm_i915_gem_set_domain *args = data; | |
05394f39 | 1375 | struct drm_i915_gem_object *obj; |
2ef7eeaa EA |
1376 | uint32_t read_domains = args->read_domains; |
1377 | uint32_t write_domain = args->write_domain; | |
673a394b EA |
1378 | int ret; |
1379 | ||
2ef7eeaa | 1380 | /* Only handle setting domains to types used by the CPU. */ |
21d509e3 | 1381 | if (write_domain & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
1382 | return -EINVAL; |
1383 | ||
21d509e3 | 1384 | if (read_domains & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
1385 | return -EINVAL; |
1386 | ||
1387 | /* Having something in the write domain implies it's in the read | |
1388 | * domain, and only that read domain. Enforce that in the request. | |
1389 | */ | |
1390 | if (write_domain != 0 && read_domains != write_domain) | |
1391 | return -EINVAL; | |
1392 | ||
76c1dec1 | 1393 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1394 | if (ret) |
76c1dec1 | 1395 | return ret; |
1d7cfea1 | 1396 | |
05394f39 | 1397 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 1398 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1399 | ret = -ENOENT; |
1400 | goto unlock; | |
76c1dec1 | 1401 | } |
673a394b | 1402 | |
3236f57a CW |
1403 | /* Try to flush the object off the GPU without holding the lock. |
1404 | * We will repeat the flush holding the lock in the normal manner | |
1405 | * to catch cases where we are gazumped. | |
1406 | */ | |
6e4930f6 CW |
1407 | ret = i915_gem_object_wait_rendering__nonblocking(obj, |
1408 | file->driver_priv, | |
1409 | !write_domain); | |
3236f57a CW |
1410 | if (ret) |
1411 | goto unref; | |
1412 | ||
2ef7eeaa EA |
1413 | if (read_domains & I915_GEM_DOMAIN_GTT) { |
1414 | ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); | |
02354392 EA |
1415 | |
1416 | /* Silently promote "you're not bound, there was nothing to do" | |
1417 | * to success, since the client was just asking us to | |
1418 | * make sure everything was done. | |
1419 | */ | |
1420 | if (ret == -EINVAL) | |
1421 | ret = 0; | |
2ef7eeaa | 1422 | } else { |
e47c68e9 | 1423 | ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); |
2ef7eeaa EA |
1424 | } |
1425 | ||
3236f57a | 1426 | unref: |
05394f39 | 1427 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1428 | unlock: |
673a394b EA |
1429 | mutex_unlock(&dev->struct_mutex); |
1430 | return ret; | |
1431 | } | |
1432 | ||
1433 | /** | |
1434 | * Called when user space has done writes to this buffer | |
1435 | */ | |
1436 | int | |
1437 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1438 | struct drm_file *file) |
673a394b EA |
1439 | { |
1440 | struct drm_i915_gem_sw_finish *args = data; | |
05394f39 | 1441 | struct drm_i915_gem_object *obj; |
673a394b EA |
1442 | int ret = 0; |
1443 | ||
76c1dec1 | 1444 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1445 | if (ret) |
76c1dec1 | 1446 | return ret; |
1d7cfea1 | 1447 | |
05394f39 | 1448 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 1449 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1450 | ret = -ENOENT; |
1451 | goto unlock; | |
673a394b EA |
1452 | } |
1453 | ||
673a394b | 1454 | /* Pinned buffers may be scanout, so flush the cache */ |
2c22569b CW |
1455 | if (obj->pin_display) |
1456 | i915_gem_object_flush_cpu_write_domain(obj, true); | |
e47c68e9 | 1457 | |
05394f39 | 1458 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1459 | unlock: |
673a394b EA |
1460 | mutex_unlock(&dev->struct_mutex); |
1461 | return ret; | |
1462 | } | |
1463 | ||
1464 | /** | |
1465 | * Maps the contents of an object, returning the address it is mapped | |
1466 | * into. | |
1467 | * | |
1468 | * While the mapping holds a reference on the contents of the object, it doesn't | |
1469 | * imply a ref on the object itself. | |
1470 | */ | |
1471 | int | |
1472 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1473 | struct drm_file *file) |
673a394b EA |
1474 | { |
1475 | struct drm_i915_gem_mmap *args = data; | |
1476 | struct drm_gem_object *obj; | |
673a394b EA |
1477 | unsigned long addr; |
1478 | ||
05394f39 | 1479 | obj = drm_gem_object_lookup(dev, file, args->handle); |
673a394b | 1480 | if (obj == NULL) |
bf79cb91 | 1481 | return -ENOENT; |
673a394b | 1482 | |
1286ff73 DV |
1483 | /* prime objects have no backing filp to GEM mmap |
1484 | * pages from. | |
1485 | */ | |
1486 | if (!obj->filp) { | |
1487 | drm_gem_object_unreference_unlocked(obj); | |
1488 | return -EINVAL; | |
1489 | } | |
1490 | ||
6be5ceb0 | 1491 | addr = vm_mmap(obj->filp, 0, args->size, |
673a394b EA |
1492 | PROT_READ | PROT_WRITE, MAP_SHARED, |
1493 | args->offset); | |
bc9025bd | 1494 | drm_gem_object_unreference_unlocked(obj); |
673a394b EA |
1495 | if (IS_ERR((void *)addr)) |
1496 | return addr; | |
1497 | ||
1498 | args->addr_ptr = (uint64_t) addr; | |
1499 | ||
1500 | return 0; | |
1501 | } | |
1502 | ||
de151cf6 JB |
1503 | /** |
1504 | * i915_gem_fault - fault a page into the GTT | |
1505 | * vma: VMA in question | |
1506 | * vmf: fault info | |
1507 | * | |
1508 | * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped | |
1509 | * from userspace. The fault handler takes care of binding the object to | |
1510 | * the GTT (if needed), allocating and programming a fence register (again, | |
1511 | * only if needed based on whether the old reg is still valid or the object | |
1512 | * is tiled) and inserting a new PTE into the faulting process. | |
1513 | * | |
1514 | * Note that the faulting process may involve evicting existing objects | |
1515 | * from the GTT and/or fence registers to make room. So performance may | |
1516 | * suffer if the GTT working set is large or there are few fence registers | |
1517 | * left. | |
1518 | */ | |
1519 | int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
1520 | { | |
05394f39 CW |
1521 | struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data); |
1522 | struct drm_device *dev = obj->base.dev; | |
3e31c6c0 | 1523 | struct drm_i915_private *dev_priv = dev->dev_private; |
de151cf6 JB |
1524 | pgoff_t page_offset; |
1525 | unsigned long pfn; | |
1526 | int ret = 0; | |
0f973f27 | 1527 | bool write = !!(vmf->flags & FAULT_FLAG_WRITE); |
de151cf6 | 1528 | |
f65c9168 PZ |
1529 | intel_runtime_pm_get(dev_priv); |
1530 | ||
de151cf6 JB |
1531 | /* We don't use vmf->pgoff since that has the fake offset */ |
1532 | page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >> | |
1533 | PAGE_SHIFT; | |
1534 | ||
d9bc7e9f CW |
1535 | ret = i915_mutex_lock_interruptible(dev); |
1536 | if (ret) | |
1537 | goto out; | |
a00b10c3 | 1538 | |
db53a302 CW |
1539 | trace_i915_gem_object_fault(obj, page_offset, true, write); |
1540 | ||
6e4930f6 CW |
1541 | /* Try to flush the object off the GPU first without holding the lock. |
1542 | * Upon reacquiring the lock, we will perform our sanity checks and then | |
1543 | * repeat the flush holding the lock in the normal manner to catch cases | |
1544 | * where we are gazumped. | |
1545 | */ | |
1546 | ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write); | |
1547 | if (ret) | |
1548 | goto unlock; | |
1549 | ||
eb119bd6 CW |
1550 | /* Access to snoopable pages through the GTT is incoherent. */ |
1551 | if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) { | |
ddeff6ee | 1552 | ret = -EFAULT; |
eb119bd6 CW |
1553 | goto unlock; |
1554 | } | |
1555 | ||
d9bc7e9f | 1556 | /* Now bind it into the GTT if needed */ |
1ec9e26d | 1557 | ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE); |
c9839303 CW |
1558 | if (ret) |
1559 | goto unlock; | |
4a684a41 | 1560 | |
c9839303 CW |
1561 | ret = i915_gem_object_set_to_gtt_domain(obj, write); |
1562 | if (ret) | |
1563 | goto unpin; | |
74898d7e | 1564 | |
06d98131 | 1565 | ret = i915_gem_object_get_fence(obj); |
d9e86c0e | 1566 | if (ret) |
c9839303 | 1567 | goto unpin; |
7d1c4804 | 1568 | |
b90b91d8 | 1569 | /* Finally, remap it using the new GTT offset */ |
f343c5f6 BW |
1570 | pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj); |
1571 | pfn >>= PAGE_SHIFT; | |
de151cf6 | 1572 | |
b90b91d8 | 1573 | if (!obj->fault_mappable) { |
beff0d0f VS |
1574 | unsigned long size = min_t(unsigned long, |
1575 | vma->vm_end - vma->vm_start, | |
1576 | obj->base.size); | |
b90b91d8 CW |
1577 | int i; |
1578 | ||
beff0d0f | 1579 | for (i = 0; i < size >> PAGE_SHIFT; i++) { |
b90b91d8 CW |
1580 | ret = vm_insert_pfn(vma, |
1581 | (unsigned long)vma->vm_start + i * PAGE_SIZE, | |
1582 | pfn + i); | |
1583 | if (ret) | |
1584 | break; | |
1585 | } | |
1586 | ||
1587 | obj->fault_mappable = true; | |
1588 | } else | |
1589 | ret = vm_insert_pfn(vma, | |
1590 | (unsigned long)vmf->virtual_address, | |
1591 | pfn + page_offset); | |
c9839303 | 1592 | unpin: |
d7f46fc4 | 1593 | i915_gem_object_ggtt_unpin(obj); |
c715089f | 1594 | unlock: |
de151cf6 | 1595 | mutex_unlock(&dev->struct_mutex); |
d9bc7e9f | 1596 | out: |
de151cf6 | 1597 | switch (ret) { |
d9bc7e9f | 1598 | case -EIO: |
2232f031 DV |
1599 | /* |
1600 | * We eat errors when the gpu is terminally wedged to avoid | |
1601 | * userspace unduly crashing (gl has no provisions for mmaps to | |
1602 | * fail). But any other -EIO isn't ours (e.g. swap in failure) | |
1603 | * and so needs to be reported. | |
1604 | */ | |
1605 | if (!i915_terminally_wedged(&dev_priv->gpu_error)) { | |
f65c9168 PZ |
1606 | ret = VM_FAULT_SIGBUS; |
1607 | break; | |
1608 | } | |
045e769a | 1609 | case -EAGAIN: |
571c608d DV |
1610 | /* |
1611 | * EAGAIN means the gpu is hung and we'll wait for the error | |
1612 | * handler to reset everything when re-faulting in | |
1613 | * i915_mutex_lock_interruptible. | |
d9bc7e9f | 1614 | */ |
c715089f CW |
1615 | case 0: |
1616 | case -ERESTARTSYS: | |
bed636ab | 1617 | case -EINTR: |
e79e0fe3 DR |
1618 | case -EBUSY: |
1619 | /* | |
1620 | * EBUSY is ok: this just means that another thread | |
1621 | * already did the job. | |
1622 | */ | |
f65c9168 PZ |
1623 | ret = VM_FAULT_NOPAGE; |
1624 | break; | |
de151cf6 | 1625 | case -ENOMEM: |
f65c9168 PZ |
1626 | ret = VM_FAULT_OOM; |
1627 | break; | |
a7c2e1aa | 1628 | case -ENOSPC: |
45d67817 | 1629 | case -EFAULT: |
f65c9168 PZ |
1630 | ret = VM_FAULT_SIGBUS; |
1631 | break; | |
de151cf6 | 1632 | default: |
a7c2e1aa | 1633 | WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret); |
f65c9168 PZ |
1634 | ret = VM_FAULT_SIGBUS; |
1635 | break; | |
de151cf6 | 1636 | } |
f65c9168 PZ |
1637 | |
1638 | intel_runtime_pm_put(dev_priv); | |
1639 | return ret; | |
de151cf6 JB |
1640 | } |
1641 | ||
901782b2 CW |
1642 | /** |
1643 | * i915_gem_release_mmap - remove physical page mappings | |
1644 | * @obj: obj in question | |
1645 | * | |
af901ca1 | 1646 | * Preserve the reservation of the mmapping with the DRM core code, but |
901782b2 CW |
1647 | * relinquish ownership of the pages back to the system. |
1648 | * | |
1649 | * It is vital that we remove the page mapping if we have mapped a tiled | |
1650 | * object through the GTT and then lose the fence register due to | |
1651 | * resource pressure. Similarly if the object has been moved out of the | |
1652 | * aperture, than pages mapped into userspace must be revoked. Removing the | |
1653 | * mapping will then trigger a page fault on the next user access, allowing | |
1654 | * fixup by i915_gem_fault(). | |
1655 | */ | |
d05ca301 | 1656 | void |
05394f39 | 1657 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) |
901782b2 | 1658 | { |
6299f992 CW |
1659 | if (!obj->fault_mappable) |
1660 | return; | |
901782b2 | 1661 | |
6796cb16 DH |
1662 | drm_vma_node_unmap(&obj->base.vma_node, |
1663 | obj->base.dev->anon_inode->i_mapping); | |
6299f992 | 1664 | obj->fault_mappable = false; |
901782b2 CW |
1665 | } |
1666 | ||
eedd10f4 CW |
1667 | void |
1668 | i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv) | |
1669 | { | |
1670 | struct drm_i915_gem_object *obj; | |
1671 | ||
1672 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) | |
1673 | i915_gem_release_mmap(obj); | |
1674 | } | |
1675 | ||
0fa87796 | 1676 | uint32_t |
e28f8711 | 1677 | i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode) |
92b88aeb | 1678 | { |
e28f8711 | 1679 | uint32_t gtt_size; |
92b88aeb CW |
1680 | |
1681 | if (INTEL_INFO(dev)->gen >= 4 || | |
e28f8711 CW |
1682 | tiling_mode == I915_TILING_NONE) |
1683 | return size; | |
92b88aeb CW |
1684 | |
1685 | /* Previous chips need a power-of-two fence region when tiling */ | |
1686 | if (INTEL_INFO(dev)->gen == 3) | |
e28f8711 | 1687 | gtt_size = 1024*1024; |
92b88aeb | 1688 | else |
e28f8711 | 1689 | gtt_size = 512*1024; |
92b88aeb | 1690 | |
e28f8711 CW |
1691 | while (gtt_size < size) |
1692 | gtt_size <<= 1; | |
92b88aeb | 1693 | |
e28f8711 | 1694 | return gtt_size; |
92b88aeb CW |
1695 | } |
1696 | ||
de151cf6 JB |
1697 | /** |
1698 | * i915_gem_get_gtt_alignment - return required GTT alignment for an object | |
1699 | * @obj: object to check | |
1700 | * | |
1701 | * Return the required GTT alignment for an object, taking into account | |
5e783301 | 1702 | * potential fence register mapping. |
de151cf6 | 1703 | */ |
d865110c ID |
1704 | uint32_t |
1705 | i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size, | |
1706 | int tiling_mode, bool fenced) | |
de151cf6 | 1707 | { |
de151cf6 JB |
1708 | /* |
1709 | * Minimum alignment is 4k (GTT page size), but might be greater | |
1710 | * if a fence register is needed for the object. | |
1711 | */ | |
d865110c | 1712 | if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) || |
e28f8711 | 1713 | tiling_mode == I915_TILING_NONE) |
de151cf6 JB |
1714 | return 4096; |
1715 | ||
a00b10c3 CW |
1716 | /* |
1717 | * Previous chips need to be aligned to the size of the smallest | |
1718 | * fence register that can contain the object. | |
1719 | */ | |
e28f8711 | 1720 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
a00b10c3 CW |
1721 | } |
1722 | ||
d8cb5086 CW |
1723 | static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj) |
1724 | { | |
1725 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
1726 | int ret; | |
1727 | ||
0de23977 | 1728 | if (drm_vma_node_has_offset(&obj->base.vma_node)) |
d8cb5086 CW |
1729 | return 0; |
1730 | ||
da494d7c DV |
1731 | dev_priv->mm.shrinker_no_lock_stealing = true; |
1732 | ||
d8cb5086 CW |
1733 | ret = drm_gem_create_mmap_offset(&obj->base); |
1734 | if (ret != -ENOSPC) | |
da494d7c | 1735 | goto out; |
d8cb5086 CW |
1736 | |
1737 | /* Badly fragmented mmap space? The only way we can recover | |
1738 | * space is by destroying unwanted objects. We can't randomly release | |
1739 | * mmap_offsets as userspace expects them to be persistent for the | |
1740 | * lifetime of the objects. The closest we can is to release the | |
1741 | * offsets on purgeable objects by truncating it and marking it purged, | |
1742 | * which prevents userspace from ever using that object again. | |
1743 | */ | |
1744 | i915_gem_purge(dev_priv, obj->base.size >> PAGE_SHIFT); | |
1745 | ret = drm_gem_create_mmap_offset(&obj->base); | |
1746 | if (ret != -ENOSPC) | |
da494d7c | 1747 | goto out; |
d8cb5086 CW |
1748 | |
1749 | i915_gem_shrink_all(dev_priv); | |
da494d7c DV |
1750 | ret = drm_gem_create_mmap_offset(&obj->base); |
1751 | out: | |
1752 | dev_priv->mm.shrinker_no_lock_stealing = false; | |
1753 | ||
1754 | return ret; | |
d8cb5086 CW |
1755 | } |
1756 | ||
1757 | static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj) | |
1758 | { | |
d8cb5086 CW |
1759 | drm_gem_free_mmap_offset(&obj->base); |
1760 | } | |
1761 | ||
de151cf6 | 1762 | int |
ff72145b DA |
1763 | i915_gem_mmap_gtt(struct drm_file *file, |
1764 | struct drm_device *dev, | |
1765 | uint32_t handle, | |
1766 | uint64_t *offset) | |
de151cf6 | 1767 | { |
da761a6e | 1768 | struct drm_i915_private *dev_priv = dev->dev_private; |
05394f39 | 1769 | struct drm_i915_gem_object *obj; |
de151cf6 JB |
1770 | int ret; |
1771 | ||
76c1dec1 | 1772 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1773 | if (ret) |
76c1dec1 | 1774 | return ret; |
de151cf6 | 1775 | |
ff72145b | 1776 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle)); |
c8725226 | 1777 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1778 | ret = -ENOENT; |
1779 | goto unlock; | |
1780 | } | |
de151cf6 | 1781 | |
5d4545ae | 1782 | if (obj->base.size > dev_priv->gtt.mappable_end) { |
da761a6e | 1783 | ret = -E2BIG; |
ff56b0bc | 1784 | goto out; |
da761a6e CW |
1785 | } |
1786 | ||
05394f39 | 1787 | if (obj->madv != I915_MADV_WILLNEED) { |
bd9b6a4e | 1788 | DRM_DEBUG("Attempting to mmap a purgeable buffer\n"); |
8c99e57d | 1789 | ret = -EFAULT; |
1d7cfea1 | 1790 | goto out; |
ab18282d CW |
1791 | } |
1792 | ||
d8cb5086 CW |
1793 | ret = i915_gem_object_create_mmap_offset(obj); |
1794 | if (ret) | |
1795 | goto out; | |
de151cf6 | 1796 | |
0de23977 | 1797 | *offset = drm_vma_node_offset_addr(&obj->base.vma_node); |
de151cf6 | 1798 | |
1d7cfea1 | 1799 | out: |
05394f39 | 1800 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1801 | unlock: |
de151cf6 | 1802 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 1803 | return ret; |
de151cf6 JB |
1804 | } |
1805 | ||
ff72145b DA |
1806 | /** |
1807 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing | |
1808 | * @dev: DRM device | |
1809 | * @data: GTT mapping ioctl data | |
1810 | * @file: GEM object info | |
1811 | * | |
1812 | * Simply returns the fake offset to userspace so it can mmap it. | |
1813 | * The mmap call will end up in drm_gem_mmap(), which will set things | |
1814 | * up so we can get faults in the handler above. | |
1815 | * | |
1816 | * The fault handler will take care of binding the object into the GTT | |
1817 | * (since it may have been evicted to make room for something), allocating | |
1818 | * a fence register, and mapping the appropriate aperture address into | |
1819 | * userspace. | |
1820 | */ | |
1821 | int | |
1822 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, | |
1823 | struct drm_file *file) | |
1824 | { | |
1825 | struct drm_i915_gem_mmap_gtt *args = data; | |
1826 | ||
ff72145b DA |
1827 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); |
1828 | } | |
1829 | ||
5537252b CW |
1830 | static inline int |
1831 | i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj) | |
1832 | { | |
1833 | return obj->madv == I915_MADV_DONTNEED; | |
1834 | } | |
1835 | ||
225067ee DV |
1836 | /* Immediately discard the backing storage */ |
1837 | static void | |
1838 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) | |
e5281ccd | 1839 | { |
4d6294bf | 1840 | i915_gem_object_free_mmap_offset(obj); |
1286ff73 | 1841 | |
4d6294bf CW |
1842 | if (obj->base.filp == NULL) |
1843 | return; | |
e5281ccd | 1844 | |
225067ee DV |
1845 | /* Our goal here is to return as much of the memory as |
1846 | * is possible back to the system as we are called from OOM. | |
1847 | * To do this we must instruct the shmfs to drop all of its | |
1848 | * backing pages, *now*. | |
1849 | */ | |
5537252b | 1850 | shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); |
225067ee DV |
1851 | obj->madv = __I915_MADV_PURGED; |
1852 | } | |
e5281ccd | 1853 | |
5537252b CW |
1854 | /* Try to discard unwanted pages */ |
1855 | static void | |
1856 | i915_gem_object_invalidate(struct drm_i915_gem_object *obj) | |
225067ee | 1857 | { |
5537252b CW |
1858 | struct address_space *mapping; |
1859 | ||
1860 | switch (obj->madv) { | |
1861 | case I915_MADV_DONTNEED: | |
1862 | i915_gem_object_truncate(obj); | |
1863 | case __I915_MADV_PURGED: | |
1864 | return; | |
1865 | } | |
1866 | ||
1867 | if (obj->base.filp == NULL) | |
1868 | return; | |
1869 | ||
1870 | mapping = file_inode(obj->base.filp)->i_mapping, | |
1871 | invalidate_mapping_pages(mapping, 0, (loff_t)-1); | |
e5281ccd CW |
1872 | } |
1873 | ||
5cdf5881 | 1874 | static void |
05394f39 | 1875 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj) |
673a394b | 1876 | { |
90797e6d ID |
1877 | struct sg_page_iter sg_iter; |
1878 | int ret; | |
1286ff73 | 1879 | |
05394f39 | 1880 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
673a394b | 1881 | |
6c085a72 CW |
1882 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
1883 | if (ret) { | |
1884 | /* In the event of a disaster, abandon all caches and | |
1885 | * hope for the best. | |
1886 | */ | |
1887 | WARN_ON(ret != -EIO); | |
2c22569b | 1888 | i915_gem_clflush_object(obj, true); |
6c085a72 CW |
1889 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
1890 | } | |
1891 | ||
6dacfd2f | 1892 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
280b713b EA |
1893 | i915_gem_object_save_bit_17_swizzle(obj); |
1894 | ||
05394f39 CW |
1895 | if (obj->madv == I915_MADV_DONTNEED) |
1896 | obj->dirty = 0; | |
3ef94daa | 1897 | |
90797e6d | 1898 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { |
2db76d7c | 1899 | struct page *page = sg_page_iter_page(&sg_iter); |
9da3da66 | 1900 | |
05394f39 | 1901 | if (obj->dirty) |
9da3da66 | 1902 | set_page_dirty(page); |
3ef94daa | 1903 | |
05394f39 | 1904 | if (obj->madv == I915_MADV_WILLNEED) |
9da3da66 | 1905 | mark_page_accessed(page); |
3ef94daa | 1906 | |
9da3da66 | 1907 | page_cache_release(page); |
3ef94daa | 1908 | } |
05394f39 | 1909 | obj->dirty = 0; |
673a394b | 1910 | |
9da3da66 CW |
1911 | sg_free_table(obj->pages); |
1912 | kfree(obj->pages); | |
37e680a1 | 1913 | } |
6c085a72 | 1914 | |
dd624afd | 1915 | int |
37e680a1 CW |
1916 | i915_gem_object_put_pages(struct drm_i915_gem_object *obj) |
1917 | { | |
1918 | const struct drm_i915_gem_object_ops *ops = obj->ops; | |
1919 | ||
2f745ad3 | 1920 | if (obj->pages == NULL) |
37e680a1 CW |
1921 | return 0; |
1922 | ||
a5570178 CW |
1923 | if (obj->pages_pin_count) |
1924 | return -EBUSY; | |
1925 | ||
9843877d | 1926 | BUG_ON(i915_gem_obj_bound_any(obj)); |
3e123027 | 1927 | |
a2165e31 CW |
1928 | /* ->put_pages might need to allocate memory for the bit17 swizzle |
1929 | * array, hence protect them from being reaped by removing them from gtt | |
1930 | * lists early. */ | |
35c20a60 | 1931 | list_del(&obj->global_list); |
a2165e31 | 1932 | |
37e680a1 | 1933 | ops->put_pages(obj); |
05394f39 | 1934 | obj->pages = NULL; |
37e680a1 | 1935 | |
5537252b | 1936 | i915_gem_object_invalidate(obj); |
6c085a72 CW |
1937 | |
1938 | return 0; | |
1939 | } | |
1940 | ||
d9973b43 | 1941 | static unsigned long |
93927ca5 DV |
1942 | __i915_gem_shrink(struct drm_i915_private *dev_priv, long target, |
1943 | bool purgeable_only) | |
6c085a72 | 1944 | { |
c8725f3d CW |
1945 | struct list_head still_in_list; |
1946 | struct drm_i915_gem_object *obj; | |
d9973b43 | 1947 | unsigned long count = 0; |
6c085a72 | 1948 | |
57094f82 | 1949 | /* |
c8725f3d | 1950 | * As we may completely rewrite the (un)bound list whilst unbinding |
57094f82 CW |
1951 | * (due to retiring requests) we have to strictly process only |
1952 | * one element of the list at the time, and recheck the list | |
1953 | * on every iteration. | |
c8725f3d CW |
1954 | * |
1955 | * In particular, we must hold a reference whilst removing the | |
1956 | * object as we may end up waiting for and/or retiring the objects. | |
1957 | * This might release the final reference (held by the active list) | |
1958 | * and result in the object being freed from under us. This is | |
1959 | * similar to the precautions the eviction code must take whilst | |
1960 | * removing objects. | |
1961 | * | |
1962 | * Also note that although these lists do not hold a reference to | |
1963 | * the object we can safely grab one here: The final object | |
1964 | * unreferencing and the bound_list are both protected by the | |
1965 | * dev->struct_mutex and so we won't ever be able to observe an | |
1966 | * object on the bound_list with a reference count equals 0. | |
57094f82 | 1967 | */ |
c8725f3d CW |
1968 | INIT_LIST_HEAD(&still_in_list); |
1969 | while (count < target && !list_empty(&dev_priv->mm.unbound_list)) { | |
1970 | obj = list_first_entry(&dev_priv->mm.unbound_list, | |
1971 | typeof(*obj), global_list); | |
1972 | list_move_tail(&obj->global_list, &still_in_list); | |
1973 | ||
1974 | if (!i915_gem_object_is_purgeable(obj) && purgeable_only) | |
1975 | continue; | |
1976 | ||
1977 | drm_gem_object_reference(&obj->base); | |
1978 | ||
1979 | if (i915_gem_object_put_pages(obj) == 0) | |
1980 | count += obj->base.size >> PAGE_SHIFT; | |
1981 | ||
1982 | drm_gem_object_unreference(&obj->base); | |
1983 | } | |
1984 | list_splice(&still_in_list, &dev_priv->mm.unbound_list); | |
1985 | ||
1986 | INIT_LIST_HEAD(&still_in_list); | |
57094f82 | 1987 | while (count < target && !list_empty(&dev_priv->mm.bound_list)) { |
07fe0b12 | 1988 | struct i915_vma *vma, *v; |
80dcfdbd | 1989 | |
57094f82 CW |
1990 | obj = list_first_entry(&dev_priv->mm.bound_list, |
1991 | typeof(*obj), global_list); | |
c8725f3d | 1992 | list_move_tail(&obj->global_list, &still_in_list); |
57094f82 | 1993 | |
80dcfdbd BW |
1994 | if (!i915_gem_object_is_purgeable(obj) && purgeable_only) |
1995 | continue; | |
1996 | ||
57094f82 CW |
1997 | drm_gem_object_reference(&obj->base); |
1998 | ||
07fe0b12 BW |
1999 | list_for_each_entry_safe(vma, v, &obj->vma_list, vma_link) |
2000 | if (i915_vma_unbind(vma)) | |
2001 | break; | |
80dcfdbd | 2002 | |
57094f82 | 2003 | if (i915_gem_object_put_pages(obj) == 0) |
6c085a72 | 2004 | count += obj->base.size >> PAGE_SHIFT; |
57094f82 CW |
2005 | |
2006 | drm_gem_object_unreference(&obj->base); | |
6c085a72 | 2007 | } |
c8725f3d | 2008 | list_splice(&still_in_list, &dev_priv->mm.bound_list); |
6c085a72 CW |
2009 | |
2010 | return count; | |
2011 | } | |
2012 | ||
d9973b43 | 2013 | static unsigned long |
93927ca5 DV |
2014 | i915_gem_purge(struct drm_i915_private *dev_priv, long target) |
2015 | { | |
2016 | return __i915_gem_shrink(dev_priv, target, true); | |
2017 | } | |
2018 | ||
d9973b43 | 2019 | static unsigned long |
6c085a72 CW |
2020 | i915_gem_shrink_all(struct drm_i915_private *dev_priv) |
2021 | { | |
6c085a72 | 2022 | i915_gem_evict_everything(dev_priv->dev); |
c8725f3d | 2023 | return __i915_gem_shrink(dev_priv, LONG_MAX, false); |
225067ee DV |
2024 | } |
2025 | ||
37e680a1 | 2026 | static int |
6c085a72 | 2027 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj) |
e5281ccd | 2028 | { |
6c085a72 | 2029 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
e5281ccd CW |
2030 | int page_count, i; |
2031 | struct address_space *mapping; | |
9da3da66 CW |
2032 | struct sg_table *st; |
2033 | struct scatterlist *sg; | |
90797e6d | 2034 | struct sg_page_iter sg_iter; |
e5281ccd | 2035 | struct page *page; |
90797e6d | 2036 | unsigned long last_pfn = 0; /* suppress gcc warning */ |
6c085a72 | 2037 | gfp_t gfp; |
e5281ccd | 2038 | |
6c085a72 CW |
2039 | /* Assert that the object is not currently in any GPU domain. As it |
2040 | * wasn't in the GTT, there shouldn't be any way it could have been in | |
2041 | * a GPU cache | |
2042 | */ | |
2043 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); | |
2044 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); | |
2045 | ||
9da3da66 CW |
2046 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
2047 | if (st == NULL) | |
2048 | return -ENOMEM; | |
2049 | ||
05394f39 | 2050 | page_count = obj->base.size / PAGE_SIZE; |
9da3da66 | 2051 | if (sg_alloc_table(st, page_count, GFP_KERNEL)) { |
9da3da66 | 2052 | kfree(st); |
e5281ccd | 2053 | return -ENOMEM; |
9da3da66 | 2054 | } |
e5281ccd | 2055 | |
9da3da66 CW |
2056 | /* Get the list of pages out of our struct file. They'll be pinned |
2057 | * at this point until we release them. | |
2058 | * | |
2059 | * Fail silently without starting the shrinker | |
2060 | */ | |
496ad9aa | 2061 | mapping = file_inode(obj->base.filp)->i_mapping; |
6c085a72 | 2062 | gfp = mapping_gfp_mask(mapping); |
caf49191 | 2063 | gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD; |
6c085a72 | 2064 | gfp &= ~(__GFP_IO | __GFP_WAIT); |
90797e6d ID |
2065 | sg = st->sgl; |
2066 | st->nents = 0; | |
2067 | for (i = 0; i < page_count; i++) { | |
6c085a72 CW |
2068 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); |
2069 | if (IS_ERR(page)) { | |
2070 | i915_gem_purge(dev_priv, page_count); | |
2071 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); | |
2072 | } | |
2073 | if (IS_ERR(page)) { | |
2074 | /* We've tried hard to allocate the memory by reaping | |
2075 | * our own buffer, now let the real VM do its job and | |
2076 | * go down in flames if truly OOM. | |
2077 | */ | |
6c085a72 | 2078 | i915_gem_shrink_all(dev_priv); |
f461d1be | 2079 | page = shmem_read_mapping_page(mapping, i); |
6c085a72 CW |
2080 | if (IS_ERR(page)) |
2081 | goto err_pages; | |
6c085a72 | 2082 | } |
426729dc KRW |
2083 | #ifdef CONFIG_SWIOTLB |
2084 | if (swiotlb_nr_tbl()) { | |
2085 | st->nents++; | |
2086 | sg_set_page(sg, page, PAGE_SIZE, 0); | |
2087 | sg = sg_next(sg); | |
2088 | continue; | |
2089 | } | |
2090 | #endif | |
90797e6d ID |
2091 | if (!i || page_to_pfn(page) != last_pfn + 1) { |
2092 | if (i) | |
2093 | sg = sg_next(sg); | |
2094 | st->nents++; | |
2095 | sg_set_page(sg, page, PAGE_SIZE, 0); | |
2096 | } else { | |
2097 | sg->length += PAGE_SIZE; | |
2098 | } | |
2099 | last_pfn = page_to_pfn(page); | |
3bbbe706 DV |
2100 | |
2101 | /* Check that the i965g/gm workaround works. */ | |
2102 | WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL)); | |
e5281ccd | 2103 | } |
426729dc KRW |
2104 | #ifdef CONFIG_SWIOTLB |
2105 | if (!swiotlb_nr_tbl()) | |
2106 | #endif | |
2107 | sg_mark_end(sg); | |
74ce6b6c CW |
2108 | obj->pages = st; |
2109 | ||
6dacfd2f | 2110 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
e5281ccd CW |
2111 | i915_gem_object_do_bit_17_swizzle(obj); |
2112 | ||
2113 | return 0; | |
2114 | ||
2115 | err_pages: | |
90797e6d ID |
2116 | sg_mark_end(sg); |
2117 | for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) | |
2db76d7c | 2118 | page_cache_release(sg_page_iter_page(&sg_iter)); |
9da3da66 CW |
2119 | sg_free_table(st); |
2120 | kfree(st); | |
0820baf3 CW |
2121 | |
2122 | /* shmemfs first checks if there is enough memory to allocate the page | |
2123 | * and reports ENOSPC should there be insufficient, along with the usual | |
2124 | * ENOMEM for a genuine allocation failure. | |
2125 | * | |
2126 | * We use ENOSPC in our driver to mean that we have run out of aperture | |
2127 | * space and so want to translate the error from shmemfs back to our | |
2128 | * usual understanding of ENOMEM. | |
2129 | */ | |
2130 | if (PTR_ERR(page) == -ENOSPC) | |
2131 | return -ENOMEM; | |
2132 | else | |
2133 | return PTR_ERR(page); | |
673a394b EA |
2134 | } |
2135 | ||
37e680a1 CW |
2136 | /* Ensure that the associated pages are gathered from the backing storage |
2137 | * and pinned into our object. i915_gem_object_get_pages() may be called | |
2138 | * multiple times before they are released by a single call to | |
2139 | * i915_gem_object_put_pages() - once the pages are no longer referenced | |
2140 | * either as a result of memory pressure (reaping pages under the shrinker) | |
2141 | * or as the object is itself released. | |
2142 | */ | |
2143 | int | |
2144 | i915_gem_object_get_pages(struct drm_i915_gem_object *obj) | |
2145 | { | |
2146 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
2147 | const struct drm_i915_gem_object_ops *ops = obj->ops; | |
2148 | int ret; | |
2149 | ||
2f745ad3 | 2150 | if (obj->pages) |
37e680a1 CW |
2151 | return 0; |
2152 | ||
43e28f09 | 2153 | if (obj->madv != I915_MADV_WILLNEED) { |
bd9b6a4e | 2154 | DRM_DEBUG("Attempting to obtain a purgeable object\n"); |
8c99e57d | 2155 | return -EFAULT; |
43e28f09 CW |
2156 | } |
2157 | ||
a5570178 CW |
2158 | BUG_ON(obj->pages_pin_count); |
2159 | ||
37e680a1 CW |
2160 | ret = ops->get_pages(obj); |
2161 | if (ret) | |
2162 | return ret; | |
2163 | ||
35c20a60 | 2164 | list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
37e680a1 | 2165 | return 0; |
673a394b EA |
2166 | } |
2167 | ||
e2d05a8b | 2168 | static void |
05394f39 | 2169 | i915_gem_object_move_to_active(struct drm_i915_gem_object *obj, |
a4872ba6 | 2170 | struct intel_engine_cs *ring) |
673a394b | 2171 | { |
9d773091 | 2172 | u32 seqno = intel_ring_get_seqno(ring); |
617dbe27 | 2173 | |
852835f3 | 2174 | BUG_ON(ring == NULL); |
02978ff5 CW |
2175 | if (obj->ring != ring && obj->last_write_seqno) { |
2176 | /* Keep the seqno relative to the current ring */ | |
2177 | obj->last_write_seqno = seqno; | |
2178 | } | |
05394f39 | 2179 | obj->ring = ring; |
673a394b EA |
2180 | |
2181 | /* Add a reference if we're newly entering the active list. */ | |
05394f39 CW |
2182 | if (!obj->active) { |
2183 | drm_gem_object_reference(&obj->base); | |
2184 | obj->active = 1; | |
673a394b | 2185 | } |
e35a41de | 2186 | |
05394f39 | 2187 | list_move_tail(&obj->ring_list, &ring->active_list); |
caea7476 | 2188 | |
0201f1ec | 2189 | obj->last_read_seqno = seqno; |
caea7476 CW |
2190 | } |
2191 | ||
e2d05a8b | 2192 | void i915_vma_move_to_active(struct i915_vma *vma, |
a4872ba6 | 2193 | struct intel_engine_cs *ring) |
e2d05a8b BW |
2194 | { |
2195 | list_move_tail(&vma->mm_list, &vma->vm->active_list); | |
2196 | return i915_gem_object_move_to_active(vma->obj, ring); | |
2197 | } | |
2198 | ||
caea7476 | 2199 | static void |
caea7476 | 2200 | i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj) |
ce44b0ea | 2201 | { |
ca191b13 | 2202 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
feb822cf BW |
2203 | struct i915_address_space *vm; |
2204 | struct i915_vma *vma; | |
ce44b0ea | 2205 | |
65ce3027 | 2206 | BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS); |
05394f39 | 2207 | BUG_ON(!obj->active); |
caea7476 | 2208 | |
feb822cf BW |
2209 | list_for_each_entry(vm, &dev_priv->vm_list, global_link) { |
2210 | vma = i915_gem_obj_to_vma(obj, vm); | |
2211 | if (vma && !list_empty(&vma->mm_list)) | |
2212 | list_move_tail(&vma->mm_list, &vm->inactive_list); | |
2213 | } | |
caea7476 | 2214 | |
f99d7069 DV |
2215 | intel_fb_obj_flush(obj, true); |
2216 | ||
65ce3027 | 2217 | list_del_init(&obj->ring_list); |
caea7476 CW |
2218 | obj->ring = NULL; |
2219 | ||
65ce3027 CW |
2220 | obj->last_read_seqno = 0; |
2221 | obj->last_write_seqno = 0; | |
2222 | obj->base.write_domain = 0; | |
2223 | ||
2224 | obj->last_fenced_seqno = 0; | |
caea7476 CW |
2225 | |
2226 | obj->active = 0; | |
2227 | drm_gem_object_unreference(&obj->base); | |
2228 | ||
2229 | WARN_ON(i915_verify_lists(dev)); | |
ce44b0ea | 2230 | } |
673a394b | 2231 | |
c8725f3d CW |
2232 | static void |
2233 | i915_gem_object_retire(struct drm_i915_gem_object *obj) | |
2234 | { | |
a4872ba6 | 2235 | struct intel_engine_cs *ring = obj->ring; |
c8725f3d CW |
2236 | |
2237 | if (ring == NULL) | |
2238 | return; | |
2239 | ||
2240 | if (i915_seqno_passed(ring->get_seqno(ring, true), | |
2241 | obj->last_read_seqno)) | |
2242 | i915_gem_object_move_to_inactive(obj); | |
2243 | } | |
2244 | ||
9d773091 | 2245 | static int |
fca26bb4 | 2246 | i915_gem_init_seqno(struct drm_device *dev, u32 seqno) |
53d227f2 | 2247 | { |
9d773091 | 2248 | struct drm_i915_private *dev_priv = dev->dev_private; |
a4872ba6 | 2249 | struct intel_engine_cs *ring; |
9d773091 | 2250 | int ret, i, j; |
53d227f2 | 2251 | |
107f27a5 | 2252 | /* Carefully retire all requests without writing to the rings */ |
9d773091 | 2253 | for_each_ring(ring, dev_priv, i) { |
107f27a5 CW |
2254 | ret = intel_ring_idle(ring); |
2255 | if (ret) | |
2256 | return ret; | |
9d773091 | 2257 | } |
9d773091 | 2258 | i915_gem_retire_requests(dev); |
107f27a5 CW |
2259 | |
2260 | /* Finally reset hw state */ | |
9d773091 | 2261 | for_each_ring(ring, dev_priv, i) { |
fca26bb4 | 2262 | intel_ring_init_seqno(ring, seqno); |
498d2ac1 | 2263 | |
ebc348b2 BW |
2264 | for (j = 0; j < ARRAY_SIZE(ring->semaphore.sync_seqno); j++) |
2265 | ring->semaphore.sync_seqno[j] = 0; | |
9d773091 | 2266 | } |
53d227f2 | 2267 | |
9d773091 | 2268 | return 0; |
53d227f2 DV |
2269 | } |
2270 | ||
fca26bb4 MK |
2271 | int i915_gem_set_seqno(struct drm_device *dev, u32 seqno) |
2272 | { | |
2273 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2274 | int ret; | |
2275 | ||
2276 | if (seqno == 0) | |
2277 | return -EINVAL; | |
2278 | ||
2279 | /* HWS page needs to be set less than what we | |
2280 | * will inject to ring | |
2281 | */ | |
2282 | ret = i915_gem_init_seqno(dev, seqno - 1); | |
2283 | if (ret) | |
2284 | return ret; | |
2285 | ||
2286 | /* Carefully set the last_seqno value so that wrap | |
2287 | * detection still works | |
2288 | */ | |
2289 | dev_priv->next_seqno = seqno; | |
2290 | dev_priv->last_seqno = seqno - 1; | |
2291 | if (dev_priv->last_seqno == 0) | |
2292 | dev_priv->last_seqno--; | |
2293 | ||
2294 | return 0; | |
2295 | } | |
2296 | ||
9d773091 CW |
2297 | int |
2298 | i915_gem_get_seqno(struct drm_device *dev, u32 *seqno) | |
53d227f2 | 2299 | { |
9d773091 CW |
2300 | struct drm_i915_private *dev_priv = dev->dev_private; |
2301 | ||
2302 | /* reserve 0 for non-seqno */ | |
2303 | if (dev_priv->next_seqno == 0) { | |
fca26bb4 | 2304 | int ret = i915_gem_init_seqno(dev, 0); |
9d773091 CW |
2305 | if (ret) |
2306 | return ret; | |
53d227f2 | 2307 | |
9d773091 CW |
2308 | dev_priv->next_seqno = 1; |
2309 | } | |
53d227f2 | 2310 | |
f72b3435 | 2311 | *seqno = dev_priv->last_seqno = dev_priv->next_seqno++; |
9d773091 | 2312 | return 0; |
53d227f2 DV |
2313 | } |
2314 | ||
a4872ba6 | 2315 | int __i915_add_request(struct intel_engine_cs *ring, |
0025c077 | 2316 | struct drm_file *file, |
7d736f4f | 2317 | struct drm_i915_gem_object *obj, |
0025c077 | 2318 | u32 *out_seqno) |
673a394b | 2319 | { |
3e31c6c0 | 2320 | struct drm_i915_private *dev_priv = ring->dev->dev_private; |
acb868d3 | 2321 | struct drm_i915_gem_request *request; |
48e29f55 | 2322 | struct intel_ringbuffer *ringbuf; |
7d736f4f | 2323 | u32 request_ring_position, request_start; |
3cce469c CW |
2324 | int ret; |
2325 | ||
48e29f55 OM |
2326 | request = ring->preallocated_lazy_request; |
2327 | if (WARN_ON(request == NULL)) | |
2328 | return -ENOMEM; | |
2329 | ||
2330 | if (i915.enable_execlists) { | |
2331 | struct intel_context *ctx = request->ctx; | |
2332 | ringbuf = ctx->engine[ring->id].ringbuf; | |
2333 | } else | |
2334 | ringbuf = ring->buffer; | |
2335 | ||
2336 | request_start = intel_ring_get_tail(ringbuf); | |
cc889e0f DV |
2337 | /* |
2338 | * Emit any outstanding flushes - execbuf can fail to emit the flush | |
2339 | * after having emitted the batchbuffer command. Hence we need to fix | |
2340 | * things up similar to emitting the lazy request. The difference here | |
2341 | * is that the flush _must_ happen before the next request, no matter | |
2342 | * what. | |
2343 | */ | |
48e29f55 OM |
2344 | if (i915.enable_execlists) { |
2345 | ret = logical_ring_flush_all_caches(ringbuf); | |
2346 | if (ret) | |
2347 | return ret; | |
2348 | } else { | |
2349 | ret = intel_ring_flush_all_caches(ring); | |
2350 | if (ret) | |
2351 | return ret; | |
2352 | } | |
cc889e0f | 2353 | |
a71d8d94 CW |
2354 | /* Record the position of the start of the request so that |
2355 | * should we detect the updated seqno part-way through the | |
2356 | * GPU processing the request, we never over-estimate the | |
2357 | * position of the head. | |
2358 | */ | |
48e29f55 | 2359 | request_ring_position = intel_ring_get_tail(ringbuf); |
a71d8d94 | 2360 | |
48e29f55 OM |
2361 | if (i915.enable_execlists) { |
2362 | ret = ring->emit_request(ringbuf); | |
2363 | if (ret) | |
2364 | return ret; | |
2365 | } else { | |
2366 | ret = ring->add_request(ring); | |
2367 | if (ret) | |
2368 | return ret; | |
2369 | } | |
673a394b | 2370 | |
9d773091 | 2371 | request->seqno = intel_ring_get_seqno(ring); |
852835f3 | 2372 | request->ring = ring; |
7d736f4f | 2373 | request->head = request_start; |
a71d8d94 | 2374 | request->tail = request_ring_position; |
7d736f4f MK |
2375 | |
2376 | /* Whilst this request exists, batch_obj will be on the | |
2377 | * active_list, and so will hold the active reference. Only when this | |
2378 | * request is retired will the the batch_obj be moved onto the | |
2379 | * inactive_list and lose its active reference. Hence we do not need | |
2380 | * to explicitly hold another reference here. | |
2381 | */ | |
9a7e0c2a | 2382 | request->batch_obj = obj; |
0e50e96b | 2383 | |
48e29f55 OM |
2384 | if (!i915.enable_execlists) { |
2385 | /* Hold a reference to the current context so that we can inspect | |
2386 | * it later in case a hangcheck error event fires. | |
2387 | */ | |
2388 | request->ctx = ring->last_context; | |
2389 | if (request->ctx) | |
2390 | i915_gem_context_reference(request->ctx); | |
2391 | } | |
0e50e96b | 2392 | |
673a394b | 2393 | request->emitted_jiffies = jiffies; |
852835f3 | 2394 | list_add_tail(&request->list, &ring->request_list); |
3bb73aba | 2395 | request->file_priv = NULL; |
852835f3 | 2396 | |
db53a302 CW |
2397 | if (file) { |
2398 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
2399 | ||
1c25595f | 2400 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 2401 | request->file_priv = file_priv; |
b962442e | 2402 | list_add_tail(&request->client_list, |
f787a5f5 | 2403 | &file_priv->mm.request_list); |
1c25595f | 2404 | spin_unlock(&file_priv->mm.lock); |
b962442e | 2405 | } |
673a394b | 2406 | |
9d773091 | 2407 | trace_i915_gem_request_add(ring, request->seqno); |
1823521d | 2408 | ring->outstanding_lazy_seqno = 0; |
3c0e234c | 2409 | ring->preallocated_lazy_request = NULL; |
db53a302 | 2410 | |
db1b76ca | 2411 | if (!dev_priv->ums.mm_suspended) { |
10cd45b6 MK |
2412 | i915_queue_hangcheck(ring->dev); |
2413 | ||
f62a0076 CW |
2414 | cancel_delayed_work_sync(&dev_priv->mm.idle_work); |
2415 | queue_delayed_work(dev_priv->wq, | |
2416 | &dev_priv->mm.retire_work, | |
2417 | round_jiffies_up_relative(HZ)); | |
2418 | intel_mark_busy(dev_priv->dev); | |
f65d9421 | 2419 | } |
cc889e0f | 2420 | |
acb868d3 | 2421 | if (out_seqno) |
9d773091 | 2422 | *out_seqno = request->seqno; |
3cce469c | 2423 | return 0; |
673a394b EA |
2424 | } |
2425 | ||
f787a5f5 CW |
2426 | static inline void |
2427 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) | |
673a394b | 2428 | { |
1c25595f | 2429 | struct drm_i915_file_private *file_priv = request->file_priv; |
673a394b | 2430 | |
1c25595f CW |
2431 | if (!file_priv) |
2432 | return; | |
1c5d22f7 | 2433 | |
1c25595f | 2434 | spin_lock(&file_priv->mm.lock); |
b29c19b6 CW |
2435 | list_del(&request->client_list); |
2436 | request->file_priv = NULL; | |
1c25595f | 2437 | spin_unlock(&file_priv->mm.lock); |
673a394b | 2438 | } |
673a394b | 2439 | |
939fd762 | 2440 | static bool i915_context_is_banned(struct drm_i915_private *dev_priv, |
273497e5 | 2441 | const struct intel_context *ctx) |
be62acb4 | 2442 | { |
44e2c070 | 2443 | unsigned long elapsed; |
be62acb4 | 2444 | |
44e2c070 MK |
2445 | elapsed = get_seconds() - ctx->hang_stats.guilty_ts; |
2446 | ||
2447 | if (ctx->hang_stats.banned) | |
be62acb4 MK |
2448 | return true; |
2449 | ||
2450 | if (elapsed <= DRM_I915_CTX_BAN_PERIOD) { | |
ccc7bed0 | 2451 | if (!i915_gem_context_is_default(ctx)) { |
3fac8978 | 2452 | DRM_DEBUG("context hanging too fast, banning!\n"); |
ccc7bed0 | 2453 | return true; |
88b4aa87 MK |
2454 | } else if (i915_stop_ring_allow_ban(dev_priv)) { |
2455 | if (i915_stop_ring_allow_warn(dev_priv)) | |
2456 | DRM_ERROR("gpu hanging too fast, banning!\n"); | |
ccc7bed0 | 2457 | return true; |
3fac8978 | 2458 | } |
be62acb4 MK |
2459 | } |
2460 | ||
2461 | return false; | |
2462 | } | |
2463 | ||
939fd762 | 2464 | static void i915_set_reset_status(struct drm_i915_private *dev_priv, |
273497e5 | 2465 | struct intel_context *ctx, |
b6b0fac0 | 2466 | const bool guilty) |
aa60c664 | 2467 | { |
44e2c070 MK |
2468 | struct i915_ctx_hang_stats *hs; |
2469 | ||
2470 | if (WARN_ON(!ctx)) | |
2471 | return; | |
aa60c664 | 2472 | |
44e2c070 MK |
2473 | hs = &ctx->hang_stats; |
2474 | ||
2475 | if (guilty) { | |
939fd762 | 2476 | hs->banned = i915_context_is_banned(dev_priv, ctx); |
44e2c070 MK |
2477 | hs->batch_active++; |
2478 | hs->guilty_ts = get_seconds(); | |
2479 | } else { | |
2480 | hs->batch_pending++; | |
aa60c664 MK |
2481 | } |
2482 | } | |
2483 | ||
0e50e96b MK |
2484 | static void i915_gem_free_request(struct drm_i915_gem_request *request) |
2485 | { | |
2486 | list_del(&request->list); | |
2487 | i915_gem_request_remove_from_client(request); | |
2488 | ||
2489 | if (request->ctx) | |
2490 | i915_gem_context_unreference(request->ctx); | |
2491 | ||
2492 | kfree(request); | |
2493 | } | |
2494 | ||
8d9fc7fd | 2495 | struct drm_i915_gem_request * |
a4872ba6 | 2496 | i915_gem_find_active_request(struct intel_engine_cs *ring) |
9375e446 | 2497 | { |
4db080f9 | 2498 | struct drm_i915_gem_request *request; |
8d9fc7fd CW |
2499 | u32 completed_seqno; |
2500 | ||
2501 | completed_seqno = ring->get_seqno(ring, false); | |
4db080f9 CW |
2502 | |
2503 | list_for_each_entry(request, &ring->request_list, list) { | |
2504 | if (i915_seqno_passed(completed_seqno, request->seqno)) | |
2505 | continue; | |
aa60c664 | 2506 | |
b6b0fac0 | 2507 | return request; |
4db080f9 | 2508 | } |
b6b0fac0 MK |
2509 | |
2510 | return NULL; | |
2511 | } | |
2512 | ||
2513 | static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv, | |
a4872ba6 | 2514 | struct intel_engine_cs *ring) |
b6b0fac0 MK |
2515 | { |
2516 | struct drm_i915_gem_request *request; | |
2517 | bool ring_hung; | |
2518 | ||
8d9fc7fd | 2519 | request = i915_gem_find_active_request(ring); |
b6b0fac0 MK |
2520 | |
2521 | if (request == NULL) | |
2522 | return; | |
2523 | ||
2524 | ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG; | |
2525 | ||
939fd762 | 2526 | i915_set_reset_status(dev_priv, request->ctx, ring_hung); |
b6b0fac0 MK |
2527 | |
2528 | list_for_each_entry_continue(request, &ring->request_list, list) | |
939fd762 | 2529 | i915_set_reset_status(dev_priv, request->ctx, false); |
4db080f9 | 2530 | } |
aa60c664 | 2531 | |
4db080f9 | 2532 | static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv, |
a4872ba6 | 2533 | struct intel_engine_cs *ring) |
4db080f9 | 2534 | { |
dfaae392 | 2535 | while (!list_empty(&ring->active_list)) { |
05394f39 | 2536 | struct drm_i915_gem_object *obj; |
9375e446 | 2537 | |
05394f39 CW |
2538 | obj = list_first_entry(&ring->active_list, |
2539 | struct drm_i915_gem_object, | |
2540 | ring_list); | |
9375e446 | 2541 | |
05394f39 | 2542 | i915_gem_object_move_to_inactive(obj); |
673a394b | 2543 | } |
1d62beea BW |
2544 | |
2545 | /* | |
2546 | * We must free the requests after all the corresponding objects have | |
2547 | * been moved off active lists. Which is the same order as the normal | |
2548 | * retire_requests function does. This is important if object hold | |
2549 | * implicit references on things like e.g. ppgtt address spaces through | |
2550 | * the request. | |
2551 | */ | |
2552 | while (!list_empty(&ring->request_list)) { | |
2553 | struct drm_i915_gem_request *request; | |
2554 | ||
2555 | request = list_first_entry(&ring->request_list, | |
2556 | struct drm_i915_gem_request, | |
2557 | list); | |
2558 | ||
2559 | i915_gem_free_request(request); | |
2560 | } | |
e3efda49 | 2561 | |
cc9130be OM |
2562 | while (!list_empty(&ring->execlist_queue)) { |
2563 | struct intel_ctx_submit_request *submit_req; | |
2564 | ||
2565 | submit_req = list_first_entry(&ring->execlist_queue, | |
2566 | struct intel_ctx_submit_request, | |
2567 | execlist_link); | |
2568 | list_del(&submit_req->execlist_link); | |
2569 | intel_runtime_pm_put(dev_priv); | |
2570 | i915_gem_context_unreference(submit_req->ctx); | |
2571 | kfree(submit_req); | |
2572 | } | |
2573 | ||
e3efda49 CW |
2574 | /* These may not have been flush before the reset, do so now */ |
2575 | kfree(ring->preallocated_lazy_request); | |
2576 | ring->preallocated_lazy_request = NULL; | |
2577 | ring->outstanding_lazy_seqno = 0; | |
673a394b EA |
2578 | } |
2579 | ||
19b2dbde | 2580 | void i915_gem_restore_fences(struct drm_device *dev) |
312817a3 CW |
2581 | { |
2582 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2583 | int i; | |
2584 | ||
4b9de737 | 2585 | for (i = 0; i < dev_priv->num_fence_regs; i++) { |
312817a3 | 2586 | struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i]; |
7d2cb39c | 2587 | |
94a335db DV |
2588 | /* |
2589 | * Commit delayed tiling changes if we have an object still | |
2590 | * attached to the fence, otherwise just clear the fence. | |
2591 | */ | |
2592 | if (reg->obj) { | |
2593 | i915_gem_object_update_fence(reg->obj, reg, | |
2594 | reg->obj->tiling_mode); | |
2595 | } else { | |
2596 | i915_gem_write_fence(dev, i, NULL); | |
2597 | } | |
312817a3 CW |
2598 | } |
2599 | } | |
2600 | ||
069efc1d | 2601 | void i915_gem_reset(struct drm_device *dev) |
673a394b | 2602 | { |
77f01230 | 2603 | struct drm_i915_private *dev_priv = dev->dev_private; |
a4872ba6 | 2604 | struct intel_engine_cs *ring; |
1ec14ad3 | 2605 | int i; |
673a394b | 2606 | |
4db080f9 CW |
2607 | /* |
2608 | * Before we free the objects from the requests, we need to inspect | |
2609 | * them for finding the guilty party. As the requests only borrow | |
2610 | * their reference to the objects, the inspection must be done first. | |
2611 | */ | |
2612 | for_each_ring(ring, dev_priv, i) | |
2613 | i915_gem_reset_ring_status(dev_priv, ring); | |
2614 | ||
b4519513 | 2615 | for_each_ring(ring, dev_priv, i) |
4db080f9 | 2616 | i915_gem_reset_ring_cleanup(dev_priv, ring); |
dfaae392 | 2617 | |
acce9ffa BW |
2618 | i915_gem_context_reset(dev); |
2619 | ||
19b2dbde | 2620 | i915_gem_restore_fences(dev); |
673a394b EA |
2621 | } |
2622 | ||
2623 | /** | |
2624 | * This function clears the request list as sequence numbers are passed. | |
2625 | */ | |
1cf0ba14 | 2626 | void |
a4872ba6 | 2627 | i915_gem_retire_requests_ring(struct intel_engine_cs *ring) |
673a394b | 2628 | { |
673a394b EA |
2629 | uint32_t seqno; |
2630 | ||
db53a302 | 2631 | if (list_empty(&ring->request_list)) |
6c0594a3 KW |
2632 | return; |
2633 | ||
db53a302 | 2634 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b | 2635 | |
b2eadbc8 | 2636 | seqno = ring->get_seqno(ring, true); |
1ec14ad3 | 2637 | |
e9103038 CW |
2638 | /* Move any buffers on the active list that are no longer referenced |
2639 | * by the ringbuffer to the flushing/inactive lists as appropriate, | |
2640 | * before we free the context associated with the requests. | |
2641 | */ | |
2642 | while (!list_empty(&ring->active_list)) { | |
2643 | struct drm_i915_gem_object *obj; | |
2644 | ||
2645 | obj = list_first_entry(&ring->active_list, | |
2646 | struct drm_i915_gem_object, | |
2647 | ring_list); | |
2648 | ||
2649 | if (!i915_seqno_passed(seqno, obj->last_read_seqno)) | |
2650 | break; | |
2651 | ||
2652 | i915_gem_object_move_to_inactive(obj); | |
2653 | } | |
2654 | ||
2655 | ||
852835f3 | 2656 | while (!list_empty(&ring->request_list)) { |
673a394b | 2657 | struct drm_i915_gem_request *request; |
48e29f55 | 2658 | struct intel_ringbuffer *ringbuf; |
673a394b | 2659 | |
852835f3 | 2660 | request = list_first_entry(&ring->request_list, |
673a394b EA |
2661 | struct drm_i915_gem_request, |
2662 | list); | |
673a394b | 2663 | |
dfaae392 | 2664 | if (!i915_seqno_passed(seqno, request->seqno)) |
b84d5f0c CW |
2665 | break; |
2666 | ||
db53a302 | 2667 | trace_i915_gem_request_retire(ring, request->seqno); |
48e29f55 OM |
2668 | |
2669 | /* This is one of the few common intersection points | |
2670 | * between legacy ringbuffer submission and execlists: | |
2671 | * we need to tell them apart in order to find the correct | |
2672 | * ringbuffer to which the request belongs to. | |
2673 | */ | |
2674 | if (i915.enable_execlists) { | |
2675 | struct intel_context *ctx = request->ctx; | |
2676 | ringbuf = ctx->engine[ring->id].ringbuf; | |
2677 | } else | |
2678 | ringbuf = ring->buffer; | |
2679 | ||
a71d8d94 CW |
2680 | /* We know the GPU must have read the request to have |
2681 | * sent us the seqno + interrupt, so use the position | |
2682 | * of tail of the request to update the last known position | |
2683 | * of the GPU head. | |
2684 | */ | |
48e29f55 | 2685 | ringbuf->last_retired_head = request->tail; |
b84d5f0c | 2686 | |
0e50e96b | 2687 | i915_gem_free_request(request); |
b84d5f0c | 2688 | } |
673a394b | 2689 | |
db53a302 CW |
2690 | if (unlikely(ring->trace_irq_seqno && |
2691 | i915_seqno_passed(seqno, ring->trace_irq_seqno))) { | |
1ec14ad3 | 2692 | ring->irq_put(ring); |
db53a302 | 2693 | ring->trace_irq_seqno = 0; |
9d34e5db | 2694 | } |
23bc5982 | 2695 | |
db53a302 | 2696 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b EA |
2697 | } |
2698 | ||
b29c19b6 | 2699 | bool |
b09a1fec CW |
2700 | i915_gem_retire_requests(struct drm_device *dev) |
2701 | { | |
3e31c6c0 | 2702 | struct drm_i915_private *dev_priv = dev->dev_private; |
a4872ba6 | 2703 | struct intel_engine_cs *ring; |
b29c19b6 | 2704 | bool idle = true; |
1ec14ad3 | 2705 | int i; |
b09a1fec | 2706 | |
b29c19b6 | 2707 | for_each_ring(ring, dev_priv, i) { |
b4519513 | 2708 | i915_gem_retire_requests_ring(ring); |
b29c19b6 CW |
2709 | idle &= list_empty(&ring->request_list); |
2710 | } | |
2711 | ||
2712 | if (idle) | |
2713 | mod_delayed_work(dev_priv->wq, | |
2714 | &dev_priv->mm.idle_work, | |
2715 | msecs_to_jiffies(100)); | |
2716 | ||
2717 | return idle; | |
b09a1fec CW |
2718 | } |
2719 | ||
75ef9da2 | 2720 | static void |
673a394b EA |
2721 | i915_gem_retire_work_handler(struct work_struct *work) |
2722 | { | |
b29c19b6 CW |
2723 | struct drm_i915_private *dev_priv = |
2724 | container_of(work, typeof(*dev_priv), mm.retire_work.work); | |
2725 | struct drm_device *dev = dev_priv->dev; | |
0a58705b | 2726 | bool idle; |
673a394b | 2727 | |
891b48cf | 2728 | /* Come back later if the device is busy... */ |
b29c19b6 CW |
2729 | idle = false; |
2730 | if (mutex_trylock(&dev->struct_mutex)) { | |
2731 | idle = i915_gem_retire_requests(dev); | |
2732 | mutex_unlock(&dev->struct_mutex); | |
673a394b | 2733 | } |
b29c19b6 | 2734 | if (!idle) |
bcb45086 CW |
2735 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2736 | round_jiffies_up_relative(HZ)); | |
b29c19b6 | 2737 | } |
0a58705b | 2738 | |
b29c19b6 CW |
2739 | static void |
2740 | i915_gem_idle_work_handler(struct work_struct *work) | |
2741 | { | |
2742 | struct drm_i915_private *dev_priv = | |
2743 | container_of(work, typeof(*dev_priv), mm.idle_work.work); | |
2744 | ||
2745 | intel_mark_idle(dev_priv->dev); | |
673a394b EA |
2746 | } |
2747 | ||
30dfebf3 DV |
2748 | /** |
2749 | * Ensures that an object will eventually get non-busy by flushing any required | |
2750 | * write domains, emitting any outstanding lazy request and retiring and | |
2751 | * completed requests. | |
2752 | */ | |
2753 | static int | |
2754 | i915_gem_object_flush_active(struct drm_i915_gem_object *obj) | |
2755 | { | |
2756 | int ret; | |
2757 | ||
2758 | if (obj->active) { | |
0201f1ec | 2759 | ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno); |
30dfebf3 DV |
2760 | if (ret) |
2761 | return ret; | |
2762 | ||
30dfebf3 DV |
2763 | i915_gem_retire_requests_ring(obj->ring); |
2764 | } | |
2765 | ||
2766 | return 0; | |
2767 | } | |
2768 | ||
23ba4fd0 BW |
2769 | /** |
2770 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT | |
2771 | * @DRM_IOCTL_ARGS: standard ioctl arguments | |
2772 | * | |
2773 | * Returns 0 if successful, else an error is returned with the remaining time in | |
2774 | * the timeout parameter. | |
2775 | * -ETIME: object is still busy after timeout | |
2776 | * -ERESTARTSYS: signal interrupted the wait | |
2777 | * -ENONENT: object doesn't exist | |
2778 | * Also possible, but rare: | |
2779 | * -EAGAIN: GPU wedged | |
2780 | * -ENOMEM: damn | |
2781 | * -ENODEV: Internal IRQ fail | |
2782 | * -E?: The add request failed | |
2783 | * | |
2784 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any | |
2785 | * non-zero timeout parameter the wait ioctl will wait for the given number of | |
2786 | * nanoseconds on an object becoming unbusy. Since the wait itself does so | |
2787 | * without holding struct_mutex the object may become re-busied before this | |
2788 | * function completes. A similar but shorter * race condition exists in the busy | |
2789 | * ioctl | |
2790 | */ | |
2791 | int | |
2792 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
2793 | { | |
3e31c6c0 | 2794 | struct drm_i915_private *dev_priv = dev->dev_private; |
23ba4fd0 BW |
2795 | struct drm_i915_gem_wait *args = data; |
2796 | struct drm_i915_gem_object *obj; | |
a4872ba6 | 2797 | struct intel_engine_cs *ring = NULL; |
f69061be | 2798 | unsigned reset_counter; |
23ba4fd0 BW |
2799 | u32 seqno = 0; |
2800 | int ret = 0; | |
2801 | ||
23ba4fd0 BW |
2802 | ret = i915_mutex_lock_interruptible(dev); |
2803 | if (ret) | |
2804 | return ret; | |
2805 | ||
2806 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle)); | |
2807 | if (&obj->base == NULL) { | |
2808 | mutex_unlock(&dev->struct_mutex); | |
2809 | return -ENOENT; | |
2810 | } | |
2811 | ||
30dfebf3 DV |
2812 | /* Need to make sure the object gets inactive eventually. */ |
2813 | ret = i915_gem_object_flush_active(obj); | |
23ba4fd0 BW |
2814 | if (ret) |
2815 | goto out; | |
2816 | ||
2817 | if (obj->active) { | |
0201f1ec | 2818 | seqno = obj->last_read_seqno; |
23ba4fd0 BW |
2819 | ring = obj->ring; |
2820 | } | |
2821 | ||
2822 | if (seqno == 0) | |
2823 | goto out; | |
2824 | ||
23ba4fd0 | 2825 | /* Do this after OLR check to make sure we make forward progress polling |
5ed0bdf2 | 2826 | * on this IOCTL with a timeout <=0 (like busy ioctl) |
23ba4fd0 | 2827 | */ |
5ed0bdf2 | 2828 | if (args->timeout_ns <= 0) { |
23ba4fd0 BW |
2829 | ret = -ETIME; |
2830 | goto out; | |
2831 | } | |
2832 | ||
2833 | drm_gem_object_unreference(&obj->base); | |
f69061be | 2834 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
23ba4fd0 BW |
2835 | mutex_unlock(&dev->struct_mutex); |
2836 | ||
5ed0bdf2 TG |
2837 | return __wait_seqno(ring, seqno, reset_counter, true, &args->timeout_ns, |
2838 | file->driver_priv); | |
23ba4fd0 BW |
2839 | |
2840 | out: | |
2841 | drm_gem_object_unreference(&obj->base); | |
2842 | mutex_unlock(&dev->struct_mutex); | |
2843 | return ret; | |
2844 | } | |
2845 | ||
5816d648 BW |
2846 | /** |
2847 | * i915_gem_object_sync - sync an object to a ring. | |
2848 | * | |
2849 | * @obj: object which may be in use on another ring. | |
2850 | * @to: ring we wish to use the object on. May be NULL. | |
2851 | * | |
2852 | * This code is meant to abstract object synchronization with the GPU. | |
2853 | * Calling with NULL implies synchronizing the object with the CPU | |
2854 | * rather than a particular GPU ring. | |
2855 | * | |
2856 | * Returns 0 if successful, else propagates up the lower layer error. | |
2857 | */ | |
2911a35b BW |
2858 | int |
2859 | i915_gem_object_sync(struct drm_i915_gem_object *obj, | |
a4872ba6 | 2860 | struct intel_engine_cs *to) |
2911a35b | 2861 | { |
a4872ba6 | 2862 | struct intel_engine_cs *from = obj->ring; |
2911a35b BW |
2863 | u32 seqno; |
2864 | int ret, idx; | |
2865 | ||
2866 | if (from == NULL || to == from) | |
2867 | return 0; | |
2868 | ||
5816d648 | 2869 | if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev)) |
0201f1ec | 2870 | return i915_gem_object_wait_rendering(obj, false); |
2911a35b BW |
2871 | |
2872 | idx = intel_ring_sync_index(from, to); | |
2873 | ||
0201f1ec | 2874 | seqno = obj->last_read_seqno; |
ddd4dbc6 RV |
2875 | /* Optimization: Avoid semaphore sync when we are sure we already |
2876 | * waited for an object with higher seqno */ | |
ebc348b2 | 2877 | if (seqno <= from->semaphore.sync_seqno[idx]) |
2911a35b BW |
2878 | return 0; |
2879 | ||
b4aca010 BW |
2880 | ret = i915_gem_check_olr(obj->ring, seqno); |
2881 | if (ret) | |
2882 | return ret; | |
2911a35b | 2883 | |
b52b89da | 2884 | trace_i915_gem_ring_sync_to(from, to, seqno); |
ebc348b2 | 2885 | ret = to->semaphore.sync_to(to, from, seqno); |
e3a5a225 | 2886 | if (!ret) |
7b01e260 MK |
2887 | /* We use last_read_seqno because sync_to() |
2888 | * might have just caused seqno wrap under | |
2889 | * the radar. | |
2890 | */ | |
ebc348b2 | 2891 | from->semaphore.sync_seqno[idx] = obj->last_read_seqno; |
2911a35b | 2892 | |
e3a5a225 | 2893 | return ret; |
2911a35b BW |
2894 | } |
2895 | ||
b5ffc9bc CW |
2896 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) |
2897 | { | |
2898 | u32 old_write_domain, old_read_domains; | |
2899 | ||
b5ffc9bc CW |
2900 | /* Force a pagefault for domain tracking on next user access */ |
2901 | i915_gem_release_mmap(obj); | |
2902 | ||
b97c3d9c KP |
2903 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
2904 | return; | |
2905 | ||
97c809fd CW |
2906 | /* Wait for any direct GTT access to complete */ |
2907 | mb(); | |
2908 | ||
b5ffc9bc CW |
2909 | old_read_domains = obj->base.read_domains; |
2910 | old_write_domain = obj->base.write_domain; | |
2911 | ||
2912 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; | |
2913 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; | |
2914 | ||
2915 | trace_i915_gem_object_change_domain(obj, | |
2916 | old_read_domains, | |
2917 | old_write_domain); | |
2918 | } | |
2919 | ||
07fe0b12 | 2920 | int i915_vma_unbind(struct i915_vma *vma) |
673a394b | 2921 | { |
07fe0b12 | 2922 | struct drm_i915_gem_object *obj = vma->obj; |
3e31c6c0 | 2923 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
43e28f09 | 2924 | int ret; |
673a394b | 2925 | |
07fe0b12 | 2926 | if (list_empty(&vma->vma_link)) |
673a394b EA |
2927 | return 0; |
2928 | ||
0ff501cb DV |
2929 | if (!drm_mm_node_allocated(&vma->node)) { |
2930 | i915_gem_vma_destroy(vma); | |
0ff501cb DV |
2931 | return 0; |
2932 | } | |
433544bd | 2933 | |
d7f46fc4 | 2934 | if (vma->pin_count) |
31d8d651 | 2935 | return -EBUSY; |
673a394b | 2936 | |
c4670ad0 CW |
2937 | BUG_ON(obj->pages == NULL); |
2938 | ||
a8198eea | 2939 | ret = i915_gem_object_finish_gpu(obj); |
1488fc08 | 2940 | if (ret) |
a8198eea CW |
2941 | return ret; |
2942 | /* Continue on if we fail due to EIO, the GPU is hung so we | |
2943 | * should be safe and we need to cleanup or else we might | |
2944 | * cause memory corruption through use-after-free. | |
2945 | */ | |
2946 | ||
8b1bc9b4 DV |
2947 | if (i915_is_ggtt(vma->vm)) { |
2948 | i915_gem_object_finish_gtt(obj); | |
5323fd04 | 2949 | |
8b1bc9b4 DV |
2950 | /* release the fence reg _after_ flushing */ |
2951 | ret = i915_gem_object_put_fence(obj); | |
2952 | if (ret) | |
2953 | return ret; | |
2954 | } | |
96b47b65 | 2955 | |
07fe0b12 | 2956 | trace_i915_vma_unbind(vma); |
db53a302 | 2957 | |
6f65e29a BW |
2958 | vma->unbind_vma(vma); |
2959 | ||
64bf9303 | 2960 | list_del_init(&vma->mm_list); |
5cacaac7 | 2961 | if (i915_is_ggtt(vma->vm)) |
e6a84468 | 2962 | obj->map_and_fenceable = false; |
673a394b | 2963 | |
2f633156 BW |
2964 | drm_mm_remove_node(&vma->node); |
2965 | i915_gem_vma_destroy(vma); | |
2966 | ||
2967 | /* Since the unbound list is global, only move to that list if | |
b93dab6e | 2968 | * no more VMAs exist. */ |
9490edb5 AR |
2969 | if (list_empty(&obj->vma_list)) { |
2970 | i915_gem_gtt_finish_object(obj); | |
2f633156 | 2971 | list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
9490edb5 | 2972 | } |
673a394b | 2973 | |
70903c3b CW |
2974 | /* And finally now the object is completely decoupled from this vma, |
2975 | * we can drop its hold on the backing storage and allow it to be | |
2976 | * reaped by the shrinker. | |
2977 | */ | |
2978 | i915_gem_object_unpin_pages(obj); | |
2979 | ||
88241785 | 2980 | return 0; |
54cf91dc CW |
2981 | } |
2982 | ||
b2da9fe5 | 2983 | int i915_gpu_idle(struct drm_device *dev) |
4df2faf4 | 2984 | { |
3e31c6c0 | 2985 | struct drm_i915_private *dev_priv = dev->dev_private; |
a4872ba6 | 2986 | struct intel_engine_cs *ring; |
1ec14ad3 | 2987 | int ret, i; |
4df2faf4 | 2988 | |
4df2faf4 | 2989 | /* Flush everything onto the inactive list. */ |
b4519513 | 2990 | for_each_ring(ring, dev_priv, i) { |
ecdb5fd8 TD |
2991 | if (!i915.enable_execlists) { |
2992 | ret = i915_switch_context(ring, ring->default_context); | |
2993 | if (ret) | |
2994 | return ret; | |
2995 | } | |
b6c7488d | 2996 | |
3e960501 | 2997 | ret = intel_ring_idle(ring); |
1ec14ad3 CW |
2998 | if (ret) |
2999 | return ret; | |
3000 | } | |
4df2faf4 | 3001 | |
8a1a49f9 | 3002 | return 0; |
4df2faf4 DV |
3003 | } |
3004 | ||
9ce079e4 CW |
3005 | static void i965_write_fence_reg(struct drm_device *dev, int reg, |
3006 | struct drm_i915_gem_object *obj) | |
de151cf6 | 3007 | { |
3e31c6c0 | 3008 | struct drm_i915_private *dev_priv = dev->dev_private; |
56c844e5 ID |
3009 | int fence_reg; |
3010 | int fence_pitch_shift; | |
de151cf6 | 3011 | |
56c844e5 ID |
3012 | if (INTEL_INFO(dev)->gen >= 6) { |
3013 | fence_reg = FENCE_REG_SANDYBRIDGE_0; | |
3014 | fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT; | |
3015 | } else { | |
3016 | fence_reg = FENCE_REG_965_0; | |
3017 | fence_pitch_shift = I965_FENCE_PITCH_SHIFT; | |
3018 | } | |
3019 | ||
d18b9619 CW |
3020 | fence_reg += reg * 8; |
3021 | ||
3022 | /* To w/a incoherency with non-atomic 64-bit register updates, | |
3023 | * we split the 64-bit update into two 32-bit writes. In order | |
3024 | * for a partial fence not to be evaluated between writes, we | |
3025 | * precede the update with write to turn off the fence register, | |
3026 | * and only enable the fence as the last step. | |
3027 | * | |
3028 | * For extra levels of paranoia, we make sure each step lands | |
3029 | * before applying the next step. | |
3030 | */ | |
3031 | I915_WRITE(fence_reg, 0); | |
3032 | POSTING_READ(fence_reg); | |
3033 | ||
9ce079e4 | 3034 | if (obj) { |
f343c5f6 | 3035 | u32 size = i915_gem_obj_ggtt_size(obj); |
d18b9619 | 3036 | uint64_t val; |
de151cf6 | 3037 | |
f343c5f6 | 3038 | val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) & |
9ce079e4 | 3039 | 0xfffff000) << 32; |
f343c5f6 | 3040 | val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000; |
56c844e5 | 3041 | val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift; |
9ce079e4 CW |
3042 | if (obj->tiling_mode == I915_TILING_Y) |
3043 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; | |
3044 | val |= I965_FENCE_REG_VALID; | |
c6642782 | 3045 | |
d18b9619 CW |
3046 | I915_WRITE(fence_reg + 4, val >> 32); |
3047 | POSTING_READ(fence_reg + 4); | |
3048 | ||
3049 | I915_WRITE(fence_reg + 0, val); | |
3050 | POSTING_READ(fence_reg); | |
3051 | } else { | |
3052 | I915_WRITE(fence_reg + 4, 0); | |
3053 | POSTING_READ(fence_reg + 4); | |
3054 | } | |
de151cf6 JB |
3055 | } |
3056 | ||
9ce079e4 CW |
3057 | static void i915_write_fence_reg(struct drm_device *dev, int reg, |
3058 | struct drm_i915_gem_object *obj) | |
de151cf6 | 3059 | { |
3e31c6c0 | 3060 | struct drm_i915_private *dev_priv = dev->dev_private; |
9ce079e4 | 3061 | u32 val; |
de151cf6 | 3062 | |
9ce079e4 | 3063 | if (obj) { |
f343c5f6 | 3064 | u32 size = i915_gem_obj_ggtt_size(obj); |
9ce079e4 CW |
3065 | int pitch_val; |
3066 | int tile_width; | |
c6642782 | 3067 | |
f343c5f6 | 3068 | WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) || |
9ce079e4 | 3069 | (size & -size) != size || |
f343c5f6 BW |
3070 | (i915_gem_obj_ggtt_offset(obj) & (size - 1)), |
3071 | "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n", | |
3072 | i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size); | |
c6642782 | 3073 | |
9ce079e4 CW |
3074 | if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)) |
3075 | tile_width = 128; | |
3076 | else | |
3077 | tile_width = 512; | |
3078 | ||
3079 | /* Note: pitch better be a power of two tile widths */ | |
3080 | pitch_val = obj->stride / tile_width; | |
3081 | pitch_val = ffs(pitch_val) - 1; | |
3082 | ||
f343c5f6 | 3083 | val = i915_gem_obj_ggtt_offset(obj); |
9ce079e4 CW |
3084 | if (obj->tiling_mode == I915_TILING_Y) |
3085 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
3086 | val |= I915_FENCE_SIZE_BITS(size); | |
3087 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
3088 | val |= I830_FENCE_REG_VALID; | |
3089 | } else | |
3090 | val = 0; | |
3091 | ||
3092 | if (reg < 8) | |
3093 | reg = FENCE_REG_830_0 + reg * 4; | |
3094 | else | |
3095 | reg = FENCE_REG_945_8 + (reg - 8) * 4; | |
3096 | ||
3097 | I915_WRITE(reg, val); | |
3098 | POSTING_READ(reg); | |
de151cf6 JB |
3099 | } |
3100 | ||
9ce079e4 CW |
3101 | static void i830_write_fence_reg(struct drm_device *dev, int reg, |
3102 | struct drm_i915_gem_object *obj) | |
de151cf6 | 3103 | { |
3e31c6c0 | 3104 | struct drm_i915_private *dev_priv = dev->dev_private; |
de151cf6 | 3105 | uint32_t val; |
de151cf6 | 3106 | |
9ce079e4 | 3107 | if (obj) { |
f343c5f6 | 3108 | u32 size = i915_gem_obj_ggtt_size(obj); |
9ce079e4 | 3109 | uint32_t pitch_val; |
de151cf6 | 3110 | |
f343c5f6 | 3111 | WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) || |
9ce079e4 | 3112 | (size & -size) != size || |
f343c5f6 BW |
3113 | (i915_gem_obj_ggtt_offset(obj) & (size - 1)), |
3114 | "object 0x%08lx not 512K or pot-size 0x%08x aligned\n", | |
3115 | i915_gem_obj_ggtt_offset(obj), size); | |
e76a16de | 3116 | |
9ce079e4 CW |
3117 | pitch_val = obj->stride / 128; |
3118 | pitch_val = ffs(pitch_val) - 1; | |
de151cf6 | 3119 | |
f343c5f6 | 3120 | val = i915_gem_obj_ggtt_offset(obj); |
9ce079e4 CW |
3121 | if (obj->tiling_mode == I915_TILING_Y) |
3122 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
3123 | val |= I830_FENCE_SIZE_BITS(size); | |
3124 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
3125 | val |= I830_FENCE_REG_VALID; | |
3126 | } else | |
3127 | val = 0; | |
c6642782 | 3128 | |
9ce079e4 CW |
3129 | I915_WRITE(FENCE_REG_830_0 + reg * 4, val); |
3130 | POSTING_READ(FENCE_REG_830_0 + reg * 4); | |
3131 | } | |
3132 | ||
d0a57789 CW |
3133 | inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj) |
3134 | { | |
3135 | return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT; | |
3136 | } | |
3137 | ||
9ce079e4 CW |
3138 | static void i915_gem_write_fence(struct drm_device *dev, int reg, |
3139 | struct drm_i915_gem_object *obj) | |
3140 | { | |
d0a57789 CW |
3141 | struct drm_i915_private *dev_priv = dev->dev_private; |
3142 | ||
3143 | /* Ensure that all CPU reads are completed before installing a fence | |
3144 | * and all writes before removing the fence. | |
3145 | */ | |
3146 | if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj)) | |
3147 | mb(); | |
3148 | ||
94a335db DV |
3149 | WARN(obj && (!obj->stride || !obj->tiling_mode), |
3150 | "bogus fence setup with stride: 0x%x, tiling mode: %i\n", | |
3151 | obj->stride, obj->tiling_mode); | |
3152 | ||
9ce079e4 | 3153 | switch (INTEL_INFO(dev)->gen) { |
5ab31333 | 3154 | case 8: |
9ce079e4 | 3155 | case 7: |
56c844e5 | 3156 | case 6: |
9ce079e4 CW |
3157 | case 5: |
3158 | case 4: i965_write_fence_reg(dev, reg, obj); break; | |
3159 | case 3: i915_write_fence_reg(dev, reg, obj); break; | |
3160 | case 2: i830_write_fence_reg(dev, reg, obj); break; | |
7dbf9d6e | 3161 | default: BUG(); |
9ce079e4 | 3162 | } |
d0a57789 CW |
3163 | |
3164 | /* And similarly be paranoid that no direct access to this region | |
3165 | * is reordered to before the fence is installed. | |
3166 | */ | |
3167 | if (i915_gem_object_needs_mb(obj)) | |
3168 | mb(); | |
de151cf6 JB |
3169 | } |
3170 | ||
61050808 CW |
3171 | static inline int fence_number(struct drm_i915_private *dev_priv, |
3172 | struct drm_i915_fence_reg *fence) | |
3173 | { | |
3174 | return fence - dev_priv->fence_regs; | |
3175 | } | |
3176 | ||
3177 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, | |
3178 | struct drm_i915_fence_reg *fence, | |
3179 | bool enable) | |
3180 | { | |
2dc8aae0 | 3181 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
46a0b638 CW |
3182 | int reg = fence_number(dev_priv, fence); |
3183 | ||
3184 | i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL); | |
61050808 CW |
3185 | |
3186 | if (enable) { | |
46a0b638 | 3187 | obj->fence_reg = reg; |
61050808 CW |
3188 | fence->obj = obj; |
3189 | list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list); | |
3190 | } else { | |
3191 | obj->fence_reg = I915_FENCE_REG_NONE; | |
3192 | fence->obj = NULL; | |
3193 | list_del_init(&fence->lru_list); | |
3194 | } | |
94a335db | 3195 | obj->fence_dirty = false; |
61050808 CW |
3196 | } |
3197 | ||
d9e86c0e | 3198 | static int |
d0a57789 | 3199 | i915_gem_object_wait_fence(struct drm_i915_gem_object *obj) |
d9e86c0e | 3200 | { |
1c293ea3 | 3201 | if (obj->last_fenced_seqno) { |
86d5bc37 | 3202 | int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno); |
18991845 CW |
3203 | if (ret) |
3204 | return ret; | |
d9e86c0e CW |
3205 | |
3206 | obj->last_fenced_seqno = 0; | |
d9e86c0e CW |
3207 | } |
3208 | ||
3209 | return 0; | |
3210 | } | |
3211 | ||
3212 | int | |
3213 | i915_gem_object_put_fence(struct drm_i915_gem_object *obj) | |
3214 | { | |
61050808 | 3215 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
f9c513e9 | 3216 | struct drm_i915_fence_reg *fence; |
d9e86c0e CW |
3217 | int ret; |
3218 | ||
d0a57789 | 3219 | ret = i915_gem_object_wait_fence(obj); |
d9e86c0e CW |
3220 | if (ret) |
3221 | return ret; | |
3222 | ||
61050808 CW |
3223 | if (obj->fence_reg == I915_FENCE_REG_NONE) |
3224 | return 0; | |
d9e86c0e | 3225 | |
f9c513e9 CW |
3226 | fence = &dev_priv->fence_regs[obj->fence_reg]; |
3227 | ||
aff10b30 DV |
3228 | if (WARN_ON(fence->pin_count)) |
3229 | return -EBUSY; | |
3230 | ||
61050808 | 3231 | i915_gem_object_fence_lost(obj); |
f9c513e9 | 3232 | i915_gem_object_update_fence(obj, fence, false); |
d9e86c0e CW |
3233 | |
3234 | return 0; | |
3235 | } | |
3236 | ||
3237 | static struct drm_i915_fence_reg * | |
a360bb1a | 3238 | i915_find_fence_reg(struct drm_device *dev) |
ae3db24a | 3239 | { |
ae3db24a | 3240 | struct drm_i915_private *dev_priv = dev->dev_private; |
8fe301ad | 3241 | struct drm_i915_fence_reg *reg, *avail; |
d9e86c0e | 3242 | int i; |
ae3db24a DV |
3243 | |
3244 | /* First try to find a free reg */ | |
d9e86c0e | 3245 | avail = NULL; |
ae3db24a DV |
3246 | for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) { |
3247 | reg = &dev_priv->fence_regs[i]; | |
3248 | if (!reg->obj) | |
d9e86c0e | 3249 | return reg; |
ae3db24a | 3250 | |
1690e1eb | 3251 | if (!reg->pin_count) |
d9e86c0e | 3252 | avail = reg; |
ae3db24a DV |
3253 | } |
3254 | ||
d9e86c0e | 3255 | if (avail == NULL) |
5dce5b93 | 3256 | goto deadlock; |
ae3db24a DV |
3257 | |
3258 | /* None available, try to steal one or wait for a user to finish */ | |
d9e86c0e | 3259 | list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) { |
1690e1eb | 3260 | if (reg->pin_count) |
ae3db24a DV |
3261 | continue; |
3262 | ||
8fe301ad | 3263 | return reg; |
ae3db24a DV |
3264 | } |
3265 | ||
5dce5b93 CW |
3266 | deadlock: |
3267 | /* Wait for completion of pending flips which consume fences */ | |
3268 | if (intel_has_pending_fb_unpin(dev)) | |
3269 | return ERR_PTR(-EAGAIN); | |
3270 | ||
3271 | return ERR_PTR(-EDEADLK); | |
ae3db24a DV |
3272 | } |
3273 | ||
de151cf6 | 3274 | /** |
9a5a53b3 | 3275 | * i915_gem_object_get_fence - set up fencing for an object |
de151cf6 JB |
3276 | * @obj: object to map through a fence reg |
3277 | * | |
3278 | * When mapping objects through the GTT, userspace wants to be able to write | |
3279 | * to them without having to worry about swizzling if the object is tiled. | |
de151cf6 JB |
3280 | * This function walks the fence regs looking for a free one for @obj, |
3281 | * stealing one if it can't find any. | |
3282 | * | |
3283 | * It then sets up the reg based on the object's properties: address, pitch | |
3284 | * and tiling format. | |
9a5a53b3 CW |
3285 | * |
3286 | * For an untiled surface, this removes any existing fence. | |
de151cf6 | 3287 | */ |
8c4b8c3f | 3288 | int |
06d98131 | 3289 | i915_gem_object_get_fence(struct drm_i915_gem_object *obj) |
de151cf6 | 3290 | { |
05394f39 | 3291 | struct drm_device *dev = obj->base.dev; |
79e53945 | 3292 | struct drm_i915_private *dev_priv = dev->dev_private; |
14415745 | 3293 | bool enable = obj->tiling_mode != I915_TILING_NONE; |
d9e86c0e | 3294 | struct drm_i915_fence_reg *reg; |
ae3db24a | 3295 | int ret; |
de151cf6 | 3296 | |
14415745 CW |
3297 | /* Have we updated the tiling parameters upon the object and so |
3298 | * will need to serialise the write to the associated fence register? | |
3299 | */ | |
5d82e3e6 | 3300 | if (obj->fence_dirty) { |
d0a57789 | 3301 | ret = i915_gem_object_wait_fence(obj); |
14415745 CW |
3302 | if (ret) |
3303 | return ret; | |
3304 | } | |
9a5a53b3 | 3305 | |
d9e86c0e | 3306 | /* Just update our place in the LRU if our fence is getting reused. */ |
05394f39 CW |
3307 | if (obj->fence_reg != I915_FENCE_REG_NONE) { |
3308 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
5d82e3e6 | 3309 | if (!obj->fence_dirty) { |
14415745 CW |
3310 | list_move_tail(®->lru_list, |
3311 | &dev_priv->mm.fence_list); | |
3312 | return 0; | |
3313 | } | |
3314 | } else if (enable) { | |
e6a84468 CW |
3315 | if (WARN_ON(!obj->map_and_fenceable)) |
3316 | return -EINVAL; | |
3317 | ||
14415745 | 3318 | reg = i915_find_fence_reg(dev); |
5dce5b93 CW |
3319 | if (IS_ERR(reg)) |
3320 | return PTR_ERR(reg); | |
d9e86c0e | 3321 | |
14415745 CW |
3322 | if (reg->obj) { |
3323 | struct drm_i915_gem_object *old = reg->obj; | |
3324 | ||
d0a57789 | 3325 | ret = i915_gem_object_wait_fence(old); |
29c5a587 CW |
3326 | if (ret) |
3327 | return ret; | |
3328 | ||
14415745 | 3329 | i915_gem_object_fence_lost(old); |
29c5a587 | 3330 | } |
14415745 | 3331 | } else |
a09ba7fa | 3332 | return 0; |
a09ba7fa | 3333 | |
14415745 | 3334 | i915_gem_object_update_fence(obj, reg, enable); |
14415745 | 3335 | |
9ce079e4 | 3336 | return 0; |
de151cf6 JB |
3337 | } |
3338 | ||
42d6ab48 CW |
3339 | static bool i915_gem_valid_gtt_space(struct drm_device *dev, |
3340 | struct drm_mm_node *gtt_space, | |
3341 | unsigned long cache_level) | |
3342 | { | |
3343 | struct drm_mm_node *other; | |
3344 | ||
3345 | /* On non-LLC machines we have to be careful when putting differing | |
3346 | * types of snoopable memory together to avoid the prefetcher | |
4239ca77 | 3347 | * crossing memory domains and dying. |
42d6ab48 CW |
3348 | */ |
3349 | if (HAS_LLC(dev)) | |
3350 | return true; | |
3351 | ||
c6cfb325 | 3352 | if (!drm_mm_node_allocated(gtt_space)) |
42d6ab48 CW |
3353 | return true; |
3354 | ||
3355 | if (list_empty(>t_space->node_list)) | |
3356 | return true; | |
3357 | ||
3358 | other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list); | |
3359 | if (other->allocated && !other->hole_follows && other->color != cache_level) | |
3360 | return false; | |
3361 | ||
3362 | other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list); | |
3363 | if (other->allocated && !gtt_space->hole_follows && other->color != cache_level) | |
3364 | return false; | |
3365 | ||
3366 | return true; | |
3367 | } | |
3368 | ||
3369 | static void i915_gem_verify_gtt(struct drm_device *dev) | |
3370 | { | |
3371 | #if WATCH_GTT | |
3372 | struct drm_i915_private *dev_priv = dev->dev_private; | |
3373 | struct drm_i915_gem_object *obj; | |
3374 | int err = 0; | |
3375 | ||
35c20a60 | 3376 | list_for_each_entry(obj, &dev_priv->mm.gtt_list, global_list) { |
42d6ab48 CW |
3377 | if (obj->gtt_space == NULL) { |
3378 | printk(KERN_ERR "object found on GTT list with no space reserved\n"); | |
3379 | err++; | |
3380 | continue; | |
3381 | } | |
3382 | ||
3383 | if (obj->cache_level != obj->gtt_space->color) { | |
3384 | printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n", | |
f343c5f6 BW |
3385 | i915_gem_obj_ggtt_offset(obj), |
3386 | i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj), | |
42d6ab48 CW |
3387 | obj->cache_level, |
3388 | obj->gtt_space->color); | |
3389 | err++; | |
3390 | continue; | |
3391 | } | |
3392 | ||
3393 | if (!i915_gem_valid_gtt_space(dev, | |
3394 | obj->gtt_space, | |
3395 | obj->cache_level)) { | |
3396 | printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n", | |
f343c5f6 BW |
3397 | i915_gem_obj_ggtt_offset(obj), |
3398 | i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj), | |
42d6ab48 CW |
3399 | obj->cache_level); |
3400 | err++; | |
3401 | continue; | |
3402 | } | |
3403 | } | |
3404 | ||
3405 | WARN_ON(err); | |
3406 | #endif | |
3407 | } | |
3408 | ||
673a394b EA |
3409 | /** |
3410 | * Finds free space in the GTT aperture and binds the object there. | |
3411 | */ | |
262de145 | 3412 | static struct i915_vma * |
07fe0b12 BW |
3413 | i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, |
3414 | struct i915_address_space *vm, | |
3415 | unsigned alignment, | |
d23db88c | 3416 | uint64_t flags) |
673a394b | 3417 | { |
05394f39 | 3418 | struct drm_device *dev = obj->base.dev; |
3e31c6c0 | 3419 | struct drm_i915_private *dev_priv = dev->dev_private; |
5e783301 | 3420 | u32 size, fence_size, fence_alignment, unfenced_alignment; |
d23db88c CW |
3421 | unsigned long start = |
3422 | flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0; | |
3423 | unsigned long end = | |
1ec9e26d | 3424 | flags & PIN_MAPPABLE ? dev_priv->gtt.mappable_end : vm->total; |
2f633156 | 3425 | struct i915_vma *vma; |
07f73f69 | 3426 | int ret; |
673a394b | 3427 | |
e28f8711 CW |
3428 | fence_size = i915_gem_get_gtt_size(dev, |
3429 | obj->base.size, | |
3430 | obj->tiling_mode); | |
3431 | fence_alignment = i915_gem_get_gtt_alignment(dev, | |
3432 | obj->base.size, | |
d865110c | 3433 | obj->tiling_mode, true); |
e28f8711 | 3434 | unfenced_alignment = |
d865110c | 3435 | i915_gem_get_gtt_alignment(dev, |
1ec9e26d DV |
3436 | obj->base.size, |
3437 | obj->tiling_mode, false); | |
a00b10c3 | 3438 | |
673a394b | 3439 | if (alignment == 0) |
1ec9e26d | 3440 | alignment = flags & PIN_MAPPABLE ? fence_alignment : |
5e783301 | 3441 | unfenced_alignment; |
1ec9e26d | 3442 | if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) { |
bd9b6a4e | 3443 | DRM_DEBUG("Invalid object alignment requested %u\n", alignment); |
262de145 | 3444 | return ERR_PTR(-EINVAL); |
673a394b EA |
3445 | } |
3446 | ||
1ec9e26d | 3447 | size = flags & PIN_MAPPABLE ? fence_size : obj->base.size; |
a00b10c3 | 3448 | |
654fc607 CW |
3449 | /* If the object is bigger than the entire aperture, reject it early |
3450 | * before evicting everything in a vain attempt to find space. | |
3451 | */ | |
d23db88c CW |
3452 | if (obj->base.size > end) { |
3453 | DRM_DEBUG("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%lu\n", | |
a36689cb | 3454 | obj->base.size, |
1ec9e26d | 3455 | flags & PIN_MAPPABLE ? "mappable" : "total", |
d23db88c | 3456 | end); |
262de145 | 3457 | return ERR_PTR(-E2BIG); |
654fc607 CW |
3458 | } |
3459 | ||
37e680a1 | 3460 | ret = i915_gem_object_get_pages(obj); |
6c085a72 | 3461 | if (ret) |
262de145 | 3462 | return ERR_PTR(ret); |
6c085a72 | 3463 | |
fbdda6fb CW |
3464 | i915_gem_object_pin_pages(obj); |
3465 | ||
accfef2e | 3466 | vma = i915_gem_obj_lookup_or_create_vma(obj, vm); |
262de145 | 3467 | if (IS_ERR(vma)) |
bc6bc15b | 3468 | goto err_unpin; |
2f633156 | 3469 | |
0a9ae0d7 | 3470 | search_free: |
07fe0b12 | 3471 | ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node, |
0a9ae0d7 | 3472 | size, alignment, |
d23db88c CW |
3473 | obj->cache_level, |
3474 | start, end, | |
62347f9e LK |
3475 | DRM_MM_SEARCH_DEFAULT, |
3476 | DRM_MM_CREATE_DEFAULT); | |
dc9dd7a2 | 3477 | if (ret) { |
f6cd1f15 | 3478 | ret = i915_gem_evict_something(dev, vm, size, alignment, |
d23db88c CW |
3479 | obj->cache_level, |
3480 | start, end, | |
3481 | flags); | |
dc9dd7a2 CW |
3482 | if (ret == 0) |
3483 | goto search_free; | |
9731129c | 3484 | |
bc6bc15b | 3485 | goto err_free_vma; |
673a394b | 3486 | } |
2f633156 | 3487 | if (WARN_ON(!i915_gem_valid_gtt_space(dev, &vma->node, |
c6cfb325 | 3488 | obj->cache_level))) { |
2f633156 | 3489 | ret = -EINVAL; |
bc6bc15b | 3490 | goto err_remove_node; |
673a394b EA |
3491 | } |
3492 | ||
74163907 | 3493 | ret = i915_gem_gtt_prepare_object(obj); |
2f633156 | 3494 | if (ret) |
bc6bc15b | 3495 | goto err_remove_node; |
673a394b | 3496 | |
35c20a60 | 3497 | list_move_tail(&obj->global_list, &dev_priv->mm.bound_list); |
ca191b13 | 3498 | list_add_tail(&vma->mm_list, &vm->inactive_list); |
bf1a1092 | 3499 | |
4bd561b3 BW |
3500 | if (i915_is_ggtt(vm)) { |
3501 | bool mappable, fenceable; | |
a00b10c3 | 3502 | |
49987099 DV |
3503 | fenceable = (vma->node.size == fence_size && |
3504 | (vma->node.start & (fence_alignment - 1)) == 0); | |
4bd561b3 | 3505 | |
49987099 DV |
3506 | mappable = (vma->node.start + obj->base.size <= |
3507 | dev_priv->gtt.mappable_end); | |
a00b10c3 | 3508 | |
5cacaac7 | 3509 | obj->map_and_fenceable = mappable && fenceable; |
4bd561b3 | 3510 | } |
75e9e915 | 3511 | |
1ec9e26d | 3512 | WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable); |
75e9e915 | 3513 | |
1ec9e26d | 3514 | trace_i915_vma_bind(vma, flags); |
8ea99c92 DV |
3515 | vma->bind_vma(vma, obj->cache_level, |
3516 | flags & (PIN_MAPPABLE | PIN_GLOBAL) ? GLOBAL_BIND : 0); | |
3517 | ||
42d6ab48 | 3518 | i915_gem_verify_gtt(dev); |
262de145 | 3519 | return vma; |
2f633156 | 3520 | |
bc6bc15b | 3521 | err_remove_node: |
6286ef9b | 3522 | drm_mm_remove_node(&vma->node); |
bc6bc15b | 3523 | err_free_vma: |
2f633156 | 3524 | i915_gem_vma_destroy(vma); |
262de145 | 3525 | vma = ERR_PTR(ret); |
bc6bc15b | 3526 | err_unpin: |
2f633156 | 3527 | i915_gem_object_unpin_pages(obj); |
262de145 | 3528 | return vma; |
673a394b EA |
3529 | } |
3530 | ||
000433b6 | 3531 | bool |
2c22569b CW |
3532 | i915_gem_clflush_object(struct drm_i915_gem_object *obj, |
3533 | bool force) | |
673a394b | 3534 | { |
673a394b EA |
3535 | /* If we don't have a page list set up, then we're not pinned |
3536 | * to GPU, and we can ignore the cache flush because it'll happen | |
3537 | * again at bind time. | |
3538 | */ | |
05394f39 | 3539 | if (obj->pages == NULL) |
000433b6 | 3540 | return false; |
673a394b | 3541 | |
769ce464 ID |
3542 | /* |
3543 | * Stolen memory is always coherent with the GPU as it is explicitly | |
3544 | * marked as wc by the system, or the system is cache-coherent. | |
3545 | */ | |
3546 | if (obj->stolen) | |
000433b6 | 3547 | return false; |
769ce464 | 3548 | |
9c23f7fc CW |
3549 | /* If the GPU is snooping the contents of the CPU cache, |
3550 | * we do not need to manually clear the CPU cache lines. However, | |
3551 | * the caches are only snooped when the render cache is | |
3552 | * flushed/invalidated. As we always have to emit invalidations | |
3553 | * and flushes when moving into and out of the RENDER domain, correct | |
3554 | * snooping behaviour occurs naturally as the result of our domain | |
3555 | * tracking. | |
3556 | */ | |
2c22569b | 3557 | if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) |
000433b6 | 3558 | return false; |
9c23f7fc | 3559 | |
1c5d22f7 | 3560 | trace_i915_gem_object_clflush(obj); |
9da3da66 | 3561 | drm_clflush_sg(obj->pages); |
000433b6 CW |
3562 | |
3563 | return true; | |
e47c68e9 EA |
3564 | } |
3565 | ||
3566 | /** Flushes the GTT write domain for the object if it's dirty. */ | |
3567 | static void | |
05394f39 | 3568 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 3569 | { |
1c5d22f7 CW |
3570 | uint32_t old_write_domain; |
3571 | ||
05394f39 | 3572 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
e47c68e9 EA |
3573 | return; |
3574 | ||
63256ec5 | 3575 | /* No actual flushing is required for the GTT write domain. Writes |
e47c68e9 EA |
3576 | * to it immediately go to main memory as far as we know, so there's |
3577 | * no chipset flush. It also doesn't land in render cache. | |
63256ec5 CW |
3578 | * |
3579 | * However, we do have to enforce the order so that all writes through | |
3580 | * the GTT land before any writes to the device, such as updates to | |
3581 | * the GATT itself. | |
e47c68e9 | 3582 | */ |
63256ec5 CW |
3583 | wmb(); |
3584 | ||
05394f39 CW |
3585 | old_write_domain = obj->base.write_domain; |
3586 | obj->base.write_domain = 0; | |
1c5d22f7 | 3587 | |
f99d7069 DV |
3588 | intel_fb_obj_flush(obj, false); |
3589 | ||
1c5d22f7 | 3590 | trace_i915_gem_object_change_domain(obj, |
05394f39 | 3591 | obj->base.read_domains, |
1c5d22f7 | 3592 | old_write_domain); |
e47c68e9 EA |
3593 | } |
3594 | ||
3595 | /** Flushes the CPU write domain for the object if it's dirty. */ | |
3596 | static void | |
2c22569b CW |
3597 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj, |
3598 | bool force) | |
e47c68e9 | 3599 | { |
1c5d22f7 | 3600 | uint32_t old_write_domain; |
e47c68e9 | 3601 | |
05394f39 | 3602 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
e47c68e9 EA |
3603 | return; |
3604 | ||
000433b6 CW |
3605 | if (i915_gem_clflush_object(obj, force)) |
3606 | i915_gem_chipset_flush(obj->base.dev); | |
3607 | ||
05394f39 CW |
3608 | old_write_domain = obj->base.write_domain; |
3609 | obj->base.write_domain = 0; | |
1c5d22f7 | 3610 | |
f99d7069 DV |
3611 | intel_fb_obj_flush(obj, false); |
3612 | ||
1c5d22f7 | 3613 | trace_i915_gem_object_change_domain(obj, |
05394f39 | 3614 | obj->base.read_domains, |
1c5d22f7 | 3615 | old_write_domain); |
e47c68e9 EA |
3616 | } |
3617 | ||
2ef7eeaa EA |
3618 | /** |
3619 | * Moves a single object to the GTT read, and possibly write domain. | |
3620 | * | |
3621 | * This function returns when the move is complete, including waiting on | |
3622 | * flushes to occur. | |
3623 | */ | |
79e53945 | 3624 | int |
2021746e | 3625 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
2ef7eeaa | 3626 | { |
3e31c6c0 | 3627 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
dc8cd1e7 | 3628 | struct i915_vma *vma = i915_gem_obj_to_ggtt(obj); |
1c5d22f7 | 3629 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 | 3630 | int ret; |
2ef7eeaa | 3631 | |
02354392 | 3632 | /* Not valid to be called on unbound objects. */ |
dc8cd1e7 | 3633 | if (vma == NULL) |
02354392 EA |
3634 | return -EINVAL; |
3635 | ||
8d7e3de1 CW |
3636 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
3637 | return 0; | |
3638 | ||
0201f1ec | 3639 | ret = i915_gem_object_wait_rendering(obj, !write); |
88241785 CW |
3640 | if (ret) |
3641 | return ret; | |
3642 | ||
c8725f3d | 3643 | i915_gem_object_retire(obj); |
2c22569b | 3644 | i915_gem_object_flush_cpu_write_domain(obj, false); |
1c5d22f7 | 3645 | |
d0a57789 CW |
3646 | /* Serialise direct access to this object with the barriers for |
3647 | * coherent writes from the GPU, by effectively invalidating the | |
3648 | * GTT domain upon first access. | |
3649 | */ | |
3650 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) | |
3651 | mb(); | |
3652 | ||
05394f39 CW |
3653 | old_write_domain = obj->base.write_domain; |
3654 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 3655 | |
e47c68e9 EA |
3656 | /* It should now be out of any other write domains, and we can update |
3657 | * the domain values for our changes. | |
3658 | */ | |
05394f39 CW |
3659 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
3660 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; | |
e47c68e9 | 3661 | if (write) { |
05394f39 CW |
3662 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
3663 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; | |
3664 | obj->dirty = 1; | |
2ef7eeaa EA |
3665 | } |
3666 | ||
f99d7069 DV |
3667 | if (write) |
3668 | intel_fb_obj_invalidate(obj, NULL); | |
3669 | ||
1c5d22f7 CW |
3670 | trace_i915_gem_object_change_domain(obj, |
3671 | old_read_domains, | |
3672 | old_write_domain); | |
3673 | ||
8325a09d | 3674 | /* And bump the LRU for this access */ |
dc8cd1e7 CW |
3675 | if (i915_gem_object_is_inactive(obj)) |
3676 | list_move_tail(&vma->mm_list, | |
3677 | &dev_priv->gtt.base.inactive_list); | |
8325a09d | 3678 | |
e47c68e9 EA |
3679 | return 0; |
3680 | } | |
3681 | ||
e4ffd173 CW |
3682 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
3683 | enum i915_cache_level cache_level) | |
3684 | { | |
7bddb01f | 3685 | struct drm_device *dev = obj->base.dev; |
df6f783a | 3686 | struct i915_vma *vma, *next; |
e4ffd173 CW |
3687 | int ret; |
3688 | ||
3689 | if (obj->cache_level == cache_level) | |
3690 | return 0; | |
3691 | ||
d7f46fc4 | 3692 | if (i915_gem_obj_is_pinned(obj)) { |
e4ffd173 CW |
3693 | DRM_DEBUG("can not change the cache level of pinned objects\n"); |
3694 | return -EBUSY; | |
3695 | } | |
3696 | ||
df6f783a | 3697 | list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) { |
3089c6f2 | 3698 | if (!i915_gem_valid_gtt_space(dev, &vma->node, cache_level)) { |
07fe0b12 | 3699 | ret = i915_vma_unbind(vma); |
3089c6f2 BW |
3700 | if (ret) |
3701 | return ret; | |
3089c6f2 | 3702 | } |
42d6ab48 CW |
3703 | } |
3704 | ||
3089c6f2 | 3705 | if (i915_gem_obj_bound_any(obj)) { |
e4ffd173 CW |
3706 | ret = i915_gem_object_finish_gpu(obj); |
3707 | if (ret) | |
3708 | return ret; | |
3709 | ||
3710 | i915_gem_object_finish_gtt(obj); | |
3711 | ||
3712 | /* Before SandyBridge, you could not use tiling or fence | |
3713 | * registers with snooped memory, so relinquish any fences | |
3714 | * currently pointing to our region in the aperture. | |
3715 | */ | |
42d6ab48 | 3716 | if (INTEL_INFO(dev)->gen < 6) { |
e4ffd173 CW |
3717 | ret = i915_gem_object_put_fence(obj); |
3718 | if (ret) | |
3719 | return ret; | |
3720 | } | |
3721 | ||
6f65e29a | 3722 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
8ea99c92 DV |
3723 | if (drm_mm_node_allocated(&vma->node)) |
3724 | vma->bind_vma(vma, cache_level, | |
3725 | obj->has_global_gtt_mapping ? GLOBAL_BIND : 0); | |
e4ffd173 CW |
3726 | } |
3727 | ||
2c22569b CW |
3728 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
3729 | vma->node.color = cache_level; | |
3730 | obj->cache_level = cache_level; | |
3731 | ||
3732 | if (cpu_write_needs_clflush(obj)) { | |
e4ffd173 CW |
3733 | u32 old_read_domains, old_write_domain; |
3734 | ||
3735 | /* If we're coming from LLC cached, then we haven't | |
3736 | * actually been tracking whether the data is in the | |
3737 | * CPU cache or not, since we only allow one bit set | |
3738 | * in obj->write_domain and have been skipping the clflushes. | |
3739 | * Just set it to the CPU cache for now. | |
3740 | */ | |
c8725f3d | 3741 | i915_gem_object_retire(obj); |
e4ffd173 | 3742 | WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU); |
e4ffd173 CW |
3743 | |
3744 | old_read_domains = obj->base.read_domains; | |
3745 | old_write_domain = obj->base.write_domain; | |
3746 | ||
3747 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
3748 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
3749 | ||
3750 | trace_i915_gem_object_change_domain(obj, | |
3751 | old_read_domains, | |
3752 | old_write_domain); | |
3753 | } | |
3754 | ||
42d6ab48 | 3755 | i915_gem_verify_gtt(dev); |
e4ffd173 CW |
3756 | return 0; |
3757 | } | |
3758 | ||
199adf40 BW |
3759 | int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, |
3760 | struct drm_file *file) | |
e6994aee | 3761 | { |
199adf40 | 3762 | struct drm_i915_gem_caching *args = data; |
e6994aee CW |
3763 | struct drm_i915_gem_object *obj; |
3764 | int ret; | |
3765 | ||
3766 | ret = i915_mutex_lock_interruptible(dev); | |
3767 | if (ret) | |
3768 | return ret; | |
3769 | ||
3770 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
3771 | if (&obj->base == NULL) { | |
3772 | ret = -ENOENT; | |
3773 | goto unlock; | |
3774 | } | |
3775 | ||
651d794f CW |
3776 | switch (obj->cache_level) { |
3777 | case I915_CACHE_LLC: | |
3778 | case I915_CACHE_L3_LLC: | |
3779 | args->caching = I915_CACHING_CACHED; | |
3780 | break; | |
3781 | ||
4257d3ba CW |
3782 | case I915_CACHE_WT: |
3783 | args->caching = I915_CACHING_DISPLAY; | |
3784 | break; | |
3785 | ||
651d794f CW |
3786 | default: |
3787 | args->caching = I915_CACHING_NONE; | |
3788 | break; | |
3789 | } | |
e6994aee CW |
3790 | |
3791 | drm_gem_object_unreference(&obj->base); | |
3792 | unlock: | |
3793 | mutex_unlock(&dev->struct_mutex); | |
3794 | return ret; | |
3795 | } | |
3796 | ||
199adf40 BW |
3797 | int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data, |
3798 | struct drm_file *file) | |
e6994aee | 3799 | { |
199adf40 | 3800 | struct drm_i915_gem_caching *args = data; |
e6994aee CW |
3801 | struct drm_i915_gem_object *obj; |
3802 | enum i915_cache_level level; | |
3803 | int ret; | |
3804 | ||
199adf40 BW |
3805 | switch (args->caching) { |
3806 | case I915_CACHING_NONE: | |
e6994aee CW |
3807 | level = I915_CACHE_NONE; |
3808 | break; | |
199adf40 | 3809 | case I915_CACHING_CACHED: |
e6994aee CW |
3810 | level = I915_CACHE_LLC; |
3811 | break; | |
4257d3ba CW |
3812 | case I915_CACHING_DISPLAY: |
3813 | level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE; | |
3814 | break; | |
e6994aee CW |
3815 | default: |
3816 | return -EINVAL; | |
3817 | } | |
3818 | ||
3bc2913e BW |
3819 | ret = i915_mutex_lock_interruptible(dev); |
3820 | if (ret) | |
3821 | return ret; | |
3822 | ||
e6994aee CW |
3823 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3824 | if (&obj->base == NULL) { | |
3825 | ret = -ENOENT; | |
3826 | goto unlock; | |
3827 | } | |
3828 | ||
3829 | ret = i915_gem_object_set_cache_level(obj, level); | |
3830 | ||
3831 | drm_gem_object_unreference(&obj->base); | |
3832 | unlock: | |
3833 | mutex_unlock(&dev->struct_mutex); | |
3834 | return ret; | |
3835 | } | |
3836 | ||
cc98b413 CW |
3837 | static bool is_pin_display(struct drm_i915_gem_object *obj) |
3838 | { | |
19656430 OM |
3839 | struct i915_vma *vma; |
3840 | ||
19656430 OM |
3841 | vma = i915_gem_obj_to_ggtt(obj); |
3842 | if (!vma) | |
3843 | return false; | |
3844 | ||
cc98b413 CW |
3845 | /* There are 3 sources that pin objects: |
3846 | * 1. The display engine (scanouts, sprites, cursors); | |
3847 | * 2. Reservations for execbuffer; | |
3848 | * 3. The user. | |
3849 | * | |
3850 | * We can ignore reservations as we hold the struct_mutex and | |
3851 | * are only called outside of the reservation path. The user | |
3852 | * can only increment pin_count once, and so if after | |
3853 | * subtracting the potential reference by the user, any pin_count | |
3854 | * remains, it must be due to another use by the display engine. | |
3855 | */ | |
19656430 | 3856 | return vma->pin_count - !!obj->user_pin_count; |
cc98b413 CW |
3857 | } |
3858 | ||
b9241ea3 | 3859 | /* |
2da3b9b9 CW |
3860 | * Prepare buffer for display plane (scanout, cursors, etc). |
3861 | * Can be called from an uninterruptible phase (modesetting) and allows | |
3862 | * any flushes to be pipelined (for pageflips). | |
b9241ea3 ZW |
3863 | */ |
3864 | int | |
2da3b9b9 CW |
3865 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
3866 | u32 alignment, | |
a4872ba6 | 3867 | struct intel_engine_cs *pipelined) |
b9241ea3 | 3868 | { |
2da3b9b9 | 3869 | u32 old_read_domains, old_write_domain; |
19656430 | 3870 | bool was_pin_display; |
b9241ea3 ZW |
3871 | int ret; |
3872 | ||
0be73284 | 3873 | if (pipelined != obj->ring) { |
2911a35b BW |
3874 | ret = i915_gem_object_sync(obj, pipelined); |
3875 | if (ret) | |
b9241ea3 ZW |
3876 | return ret; |
3877 | } | |
3878 | ||
cc98b413 CW |
3879 | /* Mark the pin_display early so that we account for the |
3880 | * display coherency whilst setting up the cache domains. | |
3881 | */ | |
19656430 | 3882 | was_pin_display = obj->pin_display; |
cc98b413 CW |
3883 | obj->pin_display = true; |
3884 | ||
a7ef0640 EA |
3885 | /* The display engine is not coherent with the LLC cache on gen6. As |
3886 | * a result, we make sure that the pinning that is about to occur is | |
3887 | * done with uncached PTEs. This is lowest common denominator for all | |
3888 | * chipsets. | |
3889 | * | |
3890 | * However for gen6+, we could do better by using the GFDT bit instead | |
3891 | * of uncaching, which would allow us to flush all the LLC-cached data | |
3892 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. | |
3893 | */ | |
651d794f CW |
3894 | ret = i915_gem_object_set_cache_level(obj, |
3895 | HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE); | |
a7ef0640 | 3896 | if (ret) |
cc98b413 | 3897 | goto err_unpin_display; |
a7ef0640 | 3898 | |
2da3b9b9 CW |
3899 | /* As the user may map the buffer once pinned in the display plane |
3900 | * (e.g. libkms for the bootup splash), we have to ensure that we | |
3901 | * always use map_and_fenceable for all scanout buffers. | |
3902 | */ | |
1ec9e26d | 3903 | ret = i915_gem_obj_ggtt_pin(obj, alignment, PIN_MAPPABLE); |
2da3b9b9 | 3904 | if (ret) |
cc98b413 | 3905 | goto err_unpin_display; |
2da3b9b9 | 3906 | |
2c22569b | 3907 | i915_gem_object_flush_cpu_write_domain(obj, true); |
b118c1e3 | 3908 | |
2da3b9b9 | 3909 | old_write_domain = obj->base.write_domain; |
05394f39 | 3910 | old_read_domains = obj->base.read_domains; |
2da3b9b9 CW |
3911 | |
3912 | /* It should now be out of any other write domains, and we can update | |
3913 | * the domain values for our changes. | |
3914 | */ | |
e5f1d962 | 3915 | obj->base.write_domain = 0; |
05394f39 | 3916 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
b9241ea3 ZW |
3917 | |
3918 | trace_i915_gem_object_change_domain(obj, | |
3919 | old_read_domains, | |
2da3b9b9 | 3920 | old_write_domain); |
b9241ea3 ZW |
3921 | |
3922 | return 0; | |
cc98b413 CW |
3923 | |
3924 | err_unpin_display: | |
19656430 OM |
3925 | WARN_ON(was_pin_display != is_pin_display(obj)); |
3926 | obj->pin_display = was_pin_display; | |
cc98b413 CW |
3927 | return ret; |
3928 | } | |
3929 | ||
3930 | void | |
3931 | i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj) | |
3932 | { | |
d7f46fc4 | 3933 | i915_gem_object_ggtt_unpin(obj); |
cc98b413 | 3934 | obj->pin_display = is_pin_display(obj); |
b9241ea3 ZW |
3935 | } |
3936 | ||
85345517 | 3937 | int |
a8198eea | 3938 | i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj) |
85345517 | 3939 | { |
88241785 CW |
3940 | int ret; |
3941 | ||
a8198eea | 3942 | if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0) |
85345517 CW |
3943 | return 0; |
3944 | ||
0201f1ec | 3945 | ret = i915_gem_object_wait_rendering(obj, false); |
c501ae7f CW |
3946 | if (ret) |
3947 | return ret; | |
3948 | ||
a8198eea CW |
3949 | /* Ensure that we invalidate the GPU's caches and TLBs. */ |
3950 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; | |
c501ae7f | 3951 | return 0; |
85345517 CW |
3952 | } |
3953 | ||
e47c68e9 EA |
3954 | /** |
3955 | * Moves a single object to the CPU read, and possibly write domain. | |
3956 | * | |
3957 | * This function returns when the move is complete, including waiting on | |
3958 | * flushes to occur. | |
3959 | */ | |
dabdfe02 | 3960 | int |
919926ae | 3961 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
e47c68e9 | 3962 | { |
1c5d22f7 | 3963 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 EA |
3964 | int ret; |
3965 | ||
8d7e3de1 CW |
3966 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3967 | return 0; | |
3968 | ||
0201f1ec | 3969 | ret = i915_gem_object_wait_rendering(obj, !write); |
88241785 CW |
3970 | if (ret) |
3971 | return ret; | |
3972 | ||
c8725f3d | 3973 | i915_gem_object_retire(obj); |
e47c68e9 | 3974 | i915_gem_object_flush_gtt_write_domain(obj); |
2ef7eeaa | 3975 | |
05394f39 CW |
3976 | old_write_domain = obj->base.write_domain; |
3977 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 3978 | |
e47c68e9 | 3979 | /* Flush the CPU cache if it's still invalid. */ |
05394f39 | 3980 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
2c22569b | 3981 | i915_gem_clflush_object(obj, false); |
2ef7eeaa | 3982 | |
05394f39 | 3983 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
2ef7eeaa EA |
3984 | } |
3985 | ||
3986 | /* It should now be out of any other write domains, and we can update | |
3987 | * the domain values for our changes. | |
3988 | */ | |
05394f39 | 3989 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
e47c68e9 EA |
3990 | |
3991 | /* If we're writing through the CPU, then the GPU read domains will | |
3992 | * need to be invalidated at next use. | |
3993 | */ | |
3994 | if (write) { | |
05394f39 CW |
3995 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3996 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
e47c68e9 | 3997 | } |
2ef7eeaa | 3998 | |
f99d7069 DV |
3999 | if (write) |
4000 | intel_fb_obj_invalidate(obj, NULL); | |
4001 | ||
1c5d22f7 CW |
4002 | trace_i915_gem_object_change_domain(obj, |
4003 | old_read_domains, | |
4004 | old_write_domain); | |
4005 | ||
2ef7eeaa EA |
4006 | return 0; |
4007 | } | |
4008 | ||
673a394b EA |
4009 | /* Throttle our rendering by waiting until the ring has completed our requests |
4010 | * emitted over 20 msec ago. | |
4011 | * | |
b962442e EA |
4012 | * Note that if we were to use the current jiffies each time around the loop, |
4013 | * we wouldn't escape the function with any frames outstanding if the time to | |
4014 | * render a frame was over 20ms. | |
4015 | * | |
673a394b EA |
4016 | * This should get us reasonable parallelism between CPU and GPU but also |
4017 | * relatively low latency when blocking on a particular request to finish. | |
4018 | */ | |
40a5f0de | 4019 | static int |
f787a5f5 | 4020 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
40a5f0de | 4021 | { |
f787a5f5 CW |
4022 | struct drm_i915_private *dev_priv = dev->dev_private; |
4023 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
b962442e | 4024 | unsigned long recent_enough = jiffies - msecs_to_jiffies(20); |
f787a5f5 | 4025 | struct drm_i915_gem_request *request; |
a4872ba6 | 4026 | struct intel_engine_cs *ring = NULL; |
f69061be | 4027 | unsigned reset_counter; |
f787a5f5 CW |
4028 | u32 seqno = 0; |
4029 | int ret; | |
93533c29 | 4030 | |
308887aa DV |
4031 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
4032 | if (ret) | |
4033 | return ret; | |
4034 | ||
4035 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, false); | |
4036 | if (ret) | |
4037 | return ret; | |
e110e8d6 | 4038 | |
1c25595f | 4039 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 4040 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { |
b962442e EA |
4041 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
4042 | break; | |
40a5f0de | 4043 | |
f787a5f5 CW |
4044 | ring = request->ring; |
4045 | seqno = request->seqno; | |
b962442e | 4046 | } |
f69061be | 4047 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
1c25595f | 4048 | spin_unlock(&file_priv->mm.lock); |
40a5f0de | 4049 | |
f787a5f5 CW |
4050 | if (seqno == 0) |
4051 | return 0; | |
2bc43b5c | 4052 | |
b29c19b6 | 4053 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL, NULL); |
f787a5f5 CW |
4054 | if (ret == 0) |
4055 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); | |
40a5f0de EA |
4056 | |
4057 | return ret; | |
4058 | } | |
4059 | ||
d23db88c CW |
4060 | static bool |
4061 | i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags) | |
4062 | { | |
4063 | struct drm_i915_gem_object *obj = vma->obj; | |
4064 | ||
4065 | if (alignment && | |
4066 | vma->node.start & (alignment - 1)) | |
4067 | return true; | |
4068 | ||
4069 | if (flags & PIN_MAPPABLE && !obj->map_and_fenceable) | |
4070 | return true; | |
4071 | ||
4072 | if (flags & PIN_OFFSET_BIAS && | |
4073 | vma->node.start < (flags & PIN_OFFSET_MASK)) | |
4074 | return true; | |
4075 | ||
4076 | return false; | |
4077 | } | |
4078 | ||
673a394b | 4079 | int |
05394f39 | 4080 | i915_gem_object_pin(struct drm_i915_gem_object *obj, |
c37e2204 | 4081 | struct i915_address_space *vm, |
05394f39 | 4082 | uint32_t alignment, |
d23db88c | 4083 | uint64_t flags) |
673a394b | 4084 | { |
6e7186af | 4085 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
07fe0b12 | 4086 | struct i915_vma *vma; |
673a394b EA |
4087 | int ret; |
4088 | ||
6e7186af BW |
4089 | if (WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base)) |
4090 | return -ENODEV; | |
4091 | ||
bf3d149b | 4092 | if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm))) |
1ec9e26d | 4093 | return -EINVAL; |
07fe0b12 BW |
4094 | |
4095 | vma = i915_gem_obj_to_vma(obj, vm); | |
07fe0b12 | 4096 | if (vma) { |
d7f46fc4 BW |
4097 | if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT)) |
4098 | return -EBUSY; | |
4099 | ||
d23db88c | 4100 | if (i915_vma_misplaced(vma, alignment, flags)) { |
d7f46fc4 | 4101 | WARN(vma->pin_count, |
ae7d49d8 | 4102 | "bo is already pinned with incorrect alignment:" |
f343c5f6 | 4103 | " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d," |
75e9e915 | 4104 | " obj->map_and_fenceable=%d\n", |
07fe0b12 | 4105 | i915_gem_obj_offset(obj, vm), alignment, |
d23db88c | 4106 | !!(flags & PIN_MAPPABLE), |
05394f39 | 4107 | obj->map_and_fenceable); |
07fe0b12 | 4108 | ret = i915_vma_unbind(vma); |
ac0c6b5a CW |
4109 | if (ret) |
4110 | return ret; | |
8ea99c92 DV |
4111 | |
4112 | vma = NULL; | |
ac0c6b5a CW |
4113 | } |
4114 | } | |
4115 | ||
8ea99c92 | 4116 | if (vma == NULL || !drm_mm_node_allocated(&vma->node)) { |
262de145 DV |
4117 | vma = i915_gem_object_bind_to_vm(obj, vm, alignment, flags); |
4118 | if (IS_ERR(vma)) | |
4119 | return PTR_ERR(vma); | |
22c344e9 | 4120 | } |
76446cac | 4121 | |
8ea99c92 DV |
4122 | if (flags & PIN_GLOBAL && !obj->has_global_gtt_mapping) |
4123 | vma->bind_vma(vma, obj->cache_level, GLOBAL_BIND); | |
74898d7e | 4124 | |
8ea99c92 | 4125 | vma->pin_count++; |
1ec9e26d DV |
4126 | if (flags & PIN_MAPPABLE) |
4127 | obj->pin_mappable |= true; | |
673a394b EA |
4128 | |
4129 | return 0; | |
4130 | } | |
4131 | ||
4132 | void | |
d7f46fc4 | 4133 | i915_gem_object_ggtt_unpin(struct drm_i915_gem_object *obj) |
673a394b | 4134 | { |
d7f46fc4 | 4135 | struct i915_vma *vma = i915_gem_obj_to_ggtt(obj); |
673a394b | 4136 | |
d7f46fc4 BW |
4137 | BUG_ON(!vma); |
4138 | BUG_ON(vma->pin_count == 0); | |
4139 | BUG_ON(!i915_gem_obj_ggtt_bound(obj)); | |
4140 | ||
4141 | if (--vma->pin_count == 0) | |
6299f992 | 4142 | obj->pin_mappable = false; |
673a394b EA |
4143 | } |
4144 | ||
d8ffa60b DV |
4145 | bool |
4146 | i915_gem_object_pin_fence(struct drm_i915_gem_object *obj) | |
4147 | { | |
4148 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
4149 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
4150 | struct i915_vma *ggtt_vma = i915_gem_obj_to_ggtt(obj); | |
4151 | ||
4152 | WARN_ON(!ggtt_vma || | |
4153 | dev_priv->fence_regs[obj->fence_reg].pin_count > | |
4154 | ggtt_vma->pin_count); | |
4155 | dev_priv->fence_regs[obj->fence_reg].pin_count++; | |
4156 | return true; | |
4157 | } else | |
4158 | return false; | |
4159 | } | |
4160 | ||
4161 | void | |
4162 | i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj) | |
4163 | { | |
4164 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
4165 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
4166 | WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count <= 0); | |
4167 | dev_priv->fence_regs[obj->fence_reg].pin_count--; | |
4168 | } | |
4169 | } | |
4170 | ||
673a394b EA |
4171 | int |
4172 | i915_gem_pin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 4173 | struct drm_file *file) |
673a394b EA |
4174 | { |
4175 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 4176 | struct drm_i915_gem_object *obj; |
673a394b EA |
4177 | int ret; |
4178 | ||
02f6bccc DV |
4179 | if (INTEL_INFO(dev)->gen >= 6) |
4180 | return -ENODEV; | |
4181 | ||
1d7cfea1 CW |
4182 | ret = i915_mutex_lock_interruptible(dev); |
4183 | if (ret) | |
4184 | return ret; | |
673a394b | 4185 | |
05394f39 | 4186 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 4187 | if (&obj->base == NULL) { |
1d7cfea1 CW |
4188 | ret = -ENOENT; |
4189 | goto unlock; | |
673a394b | 4190 | } |
673a394b | 4191 | |
05394f39 | 4192 | if (obj->madv != I915_MADV_WILLNEED) { |
bd9b6a4e | 4193 | DRM_DEBUG("Attempting to pin a purgeable buffer\n"); |
8c99e57d | 4194 | ret = -EFAULT; |
1d7cfea1 | 4195 | goto out; |
3ef94daa CW |
4196 | } |
4197 | ||
05394f39 | 4198 | if (obj->pin_filp != NULL && obj->pin_filp != file) { |
bd9b6a4e | 4199 | DRM_DEBUG("Already pinned in i915_gem_pin_ioctl(): %d\n", |
79e53945 | 4200 | args->handle); |
1d7cfea1 CW |
4201 | ret = -EINVAL; |
4202 | goto out; | |
79e53945 JB |
4203 | } |
4204 | ||
aa5f8021 DV |
4205 | if (obj->user_pin_count == ULONG_MAX) { |
4206 | ret = -EBUSY; | |
4207 | goto out; | |
4208 | } | |
4209 | ||
93be8788 | 4210 | if (obj->user_pin_count == 0) { |
1ec9e26d | 4211 | ret = i915_gem_obj_ggtt_pin(obj, args->alignment, PIN_MAPPABLE); |
1d7cfea1 CW |
4212 | if (ret) |
4213 | goto out; | |
673a394b EA |
4214 | } |
4215 | ||
93be8788 CW |
4216 | obj->user_pin_count++; |
4217 | obj->pin_filp = file; | |
4218 | ||
f343c5f6 | 4219 | args->offset = i915_gem_obj_ggtt_offset(obj); |
1d7cfea1 | 4220 | out: |
05394f39 | 4221 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 4222 | unlock: |
673a394b | 4223 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 4224 | return ret; |
673a394b EA |
4225 | } |
4226 | ||
4227 | int | |
4228 | i915_gem_unpin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 4229 | struct drm_file *file) |
673a394b EA |
4230 | { |
4231 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 4232 | struct drm_i915_gem_object *obj; |
76c1dec1 | 4233 | int ret; |
673a394b | 4234 | |
1d7cfea1 CW |
4235 | ret = i915_mutex_lock_interruptible(dev); |
4236 | if (ret) | |
4237 | return ret; | |
673a394b | 4238 | |
05394f39 | 4239 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 4240 | if (&obj->base == NULL) { |
1d7cfea1 CW |
4241 | ret = -ENOENT; |
4242 | goto unlock; | |
673a394b | 4243 | } |
76c1dec1 | 4244 | |
05394f39 | 4245 | if (obj->pin_filp != file) { |
bd9b6a4e | 4246 | DRM_DEBUG("Not pinned by caller in i915_gem_pin_ioctl(): %d\n", |
79e53945 | 4247 | args->handle); |
1d7cfea1 CW |
4248 | ret = -EINVAL; |
4249 | goto out; | |
79e53945 | 4250 | } |
05394f39 CW |
4251 | obj->user_pin_count--; |
4252 | if (obj->user_pin_count == 0) { | |
4253 | obj->pin_filp = NULL; | |
d7f46fc4 | 4254 | i915_gem_object_ggtt_unpin(obj); |
79e53945 | 4255 | } |
673a394b | 4256 | |
1d7cfea1 | 4257 | out: |
05394f39 | 4258 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 4259 | unlock: |
673a394b | 4260 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 4261 | return ret; |
673a394b EA |
4262 | } |
4263 | ||
4264 | int | |
4265 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 4266 | struct drm_file *file) |
673a394b EA |
4267 | { |
4268 | struct drm_i915_gem_busy *args = data; | |
05394f39 | 4269 | struct drm_i915_gem_object *obj; |
30dbf0c0 CW |
4270 | int ret; |
4271 | ||
76c1dec1 | 4272 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 4273 | if (ret) |
76c1dec1 | 4274 | return ret; |
673a394b | 4275 | |
05394f39 | 4276 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 4277 | if (&obj->base == NULL) { |
1d7cfea1 CW |
4278 | ret = -ENOENT; |
4279 | goto unlock; | |
673a394b | 4280 | } |
d1b851fc | 4281 | |
0be555b6 CW |
4282 | /* Count all active objects as busy, even if they are currently not used |
4283 | * by the gpu. Users of this interface expect objects to eventually | |
4284 | * become non-busy without any further actions, therefore emit any | |
4285 | * necessary flushes here. | |
c4de0a5d | 4286 | */ |
30dfebf3 | 4287 | ret = i915_gem_object_flush_active(obj); |
0be555b6 | 4288 | |
30dfebf3 | 4289 | args->busy = obj->active; |
e9808edd CW |
4290 | if (obj->ring) { |
4291 | BUILD_BUG_ON(I915_NUM_RINGS > 16); | |
4292 | args->busy |= intel_ring_flag(obj->ring) << 16; | |
4293 | } | |
673a394b | 4294 | |
05394f39 | 4295 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 4296 | unlock: |
673a394b | 4297 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 4298 | return ret; |
673a394b EA |
4299 | } |
4300 | ||
4301 | int | |
4302 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, | |
4303 | struct drm_file *file_priv) | |
4304 | { | |
0206e353 | 4305 | return i915_gem_ring_throttle(dev, file_priv); |
673a394b EA |
4306 | } |
4307 | ||
3ef94daa CW |
4308 | int |
4309 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, | |
4310 | struct drm_file *file_priv) | |
4311 | { | |
4312 | struct drm_i915_gem_madvise *args = data; | |
05394f39 | 4313 | struct drm_i915_gem_object *obj; |
76c1dec1 | 4314 | int ret; |
3ef94daa CW |
4315 | |
4316 | switch (args->madv) { | |
4317 | case I915_MADV_DONTNEED: | |
4318 | case I915_MADV_WILLNEED: | |
4319 | break; | |
4320 | default: | |
4321 | return -EINVAL; | |
4322 | } | |
4323 | ||
1d7cfea1 CW |
4324 | ret = i915_mutex_lock_interruptible(dev); |
4325 | if (ret) | |
4326 | return ret; | |
4327 | ||
05394f39 | 4328 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); |
c8725226 | 4329 | if (&obj->base == NULL) { |
1d7cfea1 CW |
4330 | ret = -ENOENT; |
4331 | goto unlock; | |
3ef94daa | 4332 | } |
3ef94daa | 4333 | |
d7f46fc4 | 4334 | if (i915_gem_obj_is_pinned(obj)) { |
1d7cfea1 CW |
4335 | ret = -EINVAL; |
4336 | goto out; | |
3ef94daa CW |
4337 | } |
4338 | ||
05394f39 CW |
4339 | if (obj->madv != __I915_MADV_PURGED) |
4340 | obj->madv = args->madv; | |
3ef94daa | 4341 | |
6c085a72 CW |
4342 | /* if the object is no longer attached, discard its backing storage */ |
4343 | if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL) | |
2d7ef395 CW |
4344 | i915_gem_object_truncate(obj); |
4345 | ||
05394f39 | 4346 | args->retained = obj->madv != __I915_MADV_PURGED; |
bb6baf76 | 4347 | |
1d7cfea1 | 4348 | out: |
05394f39 | 4349 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 4350 | unlock: |
3ef94daa | 4351 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 4352 | return ret; |
3ef94daa CW |
4353 | } |
4354 | ||
37e680a1 CW |
4355 | void i915_gem_object_init(struct drm_i915_gem_object *obj, |
4356 | const struct drm_i915_gem_object_ops *ops) | |
0327d6ba | 4357 | { |
35c20a60 | 4358 | INIT_LIST_HEAD(&obj->global_list); |
0327d6ba | 4359 | INIT_LIST_HEAD(&obj->ring_list); |
b25cb2f8 | 4360 | INIT_LIST_HEAD(&obj->obj_exec_link); |
2f633156 | 4361 | INIT_LIST_HEAD(&obj->vma_list); |
0327d6ba | 4362 | |
37e680a1 CW |
4363 | obj->ops = ops; |
4364 | ||
0327d6ba CW |
4365 | obj->fence_reg = I915_FENCE_REG_NONE; |
4366 | obj->madv = I915_MADV_WILLNEED; | |
0327d6ba CW |
4367 | |
4368 | i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size); | |
4369 | } | |
4370 | ||
37e680a1 CW |
4371 | static const struct drm_i915_gem_object_ops i915_gem_object_ops = { |
4372 | .get_pages = i915_gem_object_get_pages_gtt, | |
4373 | .put_pages = i915_gem_object_put_pages_gtt, | |
4374 | }; | |
4375 | ||
05394f39 CW |
4376 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, |
4377 | size_t size) | |
ac52bc56 | 4378 | { |
c397b908 | 4379 | struct drm_i915_gem_object *obj; |
5949eac4 | 4380 | struct address_space *mapping; |
1a240d4d | 4381 | gfp_t mask; |
ac52bc56 | 4382 | |
42dcedd4 | 4383 | obj = i915_gem_object_alloc(dev); |
c397b908 DV |
4384 | if (obj == NULL) |
4385 | return NULL; | |
673a394b | 4386 | |
c397b908 | 4387 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { |
42dcedd4 | 4388 | i915_gem_object_free(obj); |
c397b908 DV |
4389 | return NULL; |
4390 | } | |
673a394b | 4391 | |
bed1ea95 CW |
4392 | mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; |
4393 | if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) { | |
4394 | /* 965gm cannot relocate objects above 4GiB. */ | |
4395 | mask &= ~__GFP_HIGHMEM; | |
4396 | mask |= __GFP_DMA32; | |
4397 | } | |
4398 | ||
496ad9aa | 4399 | mapping = file_inode(obj->base.filp)->i_mapping; |
bed1ea95 | 4400 | mapping_set_gfp_mask(mapping, mask); |
5949eac4 | 4401 | |
37e680a1 | 4402 | i915_gem_object_init(obj, &i915_gem_object_ops); |
73aa808f | 4403 | |
c397b908 DV |
4404 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
4405 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
673a394b | 4406 | |
3d29b842 ED |
4407 | if (HAS_LLC(dev)) { |
4408 | /* On some devices, we can have the GPU use the LLC (the CPU | |
a1871112 EA |
4409 | * cache) for about a 10% performance improvement |
4410 | * compared to uncached. Graphics requests other than | |
4411 | * display scanout are coherent with the CPU in | |
4412 | * accessing this cache. This means in this mode we | |
4413 | * don't need to clflush on the CPU side, and on the | |
4414 | * GPU side we only need to flush internal caches to | |
4415 | * get data visible to the CPU. | |
4416 | * | |
4417 | * However, we maintain the display planes as UC, and so | |
4418 | * need to rebind when first used as such. | |
4419 | */ | |
4420 | obj->cache_level = I915_CACHE_LLC; | |
4421 | } else | |
4422 | obj->cache_level = I915_CACHE_NONE; | |
4423 | ||
d861e338 DV |
4424 | trace_i915_gem_object_create(obj); |
4425 | ||
05394f39 | 4426 | return obj; |
c397b908 DV |
4427 | } |
4428 | ||
340fbd8c CW |
4429 | static bool discard_backing_storage(struct drm_i915_gem_object *obj) |
4430 | { | |
4431 | /* If we are the last user of the backing storage (be it shmemfs | |
4432 | * pages or stolen etc), we know that the pages are going to be | |
4433 | * immediately released. In this case, we can then skip copying | |
4434 | * back the contents from the GPU. | |
4435 | */ | |
4436 | ||
4437 | if (obj->madv != I915_MADV_WILLNEED) | |
4438 | return false; | |
4439 | ||
4440 | if (obj->base.filp == NULL) | |
4441 | return true; | |
4442 | ||
4443 | /* At first glance, this looks racy, but then again so would be | |
4444 | * userspace racing mmap against close. However, the first external | |
4445 | * reference to the filp can only be obtained through the | |
4446 | * i915_gem_mmap_ioctl() which safeguards us against the user | |
4447 | * acquiring such a reference whilst we are in the middle of | |
4448 | * freeing the object. | |
4449 | */ | |
4450 | return atomic_long_read(&obj->base.filp->f_count) == 1; | |
4451 | } | |
4452 | ||
1488fc08 | 4453 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
673a394b | 4454 | { |
1488fc08 | 4455 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); |
05394f39 | 4456 | struct drm_device *dev = obj->base.dev; |
3e31c6c0 | 4457 | struct drm_i915_private *dev_priv = dev->dev_private; |
07fe0b12 | 4458 | struct i915_vma *vma, *next; |
673a394b | 4459 | |
f65c9168 PZ |
4460 | intel_runtime_pm_get(dev_priv); |
4461 | ||
26e12f89 CW |
4462 | trace_i915_gem_object_destroy(obj); |
4463 | ||
07fe0b12 | 4464 | list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) { |
d7f46fc4 BW |
4465 | int ret; |
4466 | ||
4467 | vma->pin_count = 0; | |
4468 | ret = i915_vma_unbind(vma); | |
07fe0b12 BW |
4469 | if (WARN_ON(ret == -ERESTARTSYS)) { |
4470 | bool was_interruptible; | |
1488fc08 | 4471 | |
07fe0b12 BW |
4472 | was_interruptible = dev_priv->mm.interruptible; |
4473 | dev_priv->mm.interruptible = false; | |
1488fc08 | 4474 | |
07fe0b12 | 4475 | WARN_ON(i915_vma_unbind(vma)); |
1488fc08 | 4476 | |
07fe0b12 BW |
4477 | dev_priv->mm.interruptible = was_interruptible; |
4478 | } | |
1488fc08 CW |
4479 | } |
4480 | ||
00731155 CW |
4481 | i915_gem_object_detach_phys(obj); |
4482 | ||
1d64ae71 BW |
4483 | /* Stolen objects don't hold a ref, but do hold pin count. Fix that up |
4484 | * before progressing. */ | |
4485 | if (obj->stolen) | |
4486 | i915_gem_object_unpin_pages(obj); | |
4487 | ||
a071fa00 DV |
4488 | WARN_ON(obj->frontbuffer_bits); |
4489 | ||
401c29f6 BW |
4490 | if (WARN_ON(obj->pages_pin_count)) |
4491 | obj->pages_pin_count = 0; | |
340fbd8c | 4492 | if (discard_backing_storage(obj)) |
5537252b | 4493 | obj->madv = I915_MADV_DONTNEED; |
37e680a1 | 4494 | i915_gem_object_put_pages(obj); |
d8cb5086 | 4495 | i915_gem_object_free_mmap_offset(obj); |
de151cf6 | 4496 | |
9da3da66 CW |
4497 | BUG_ON(obj->pages); |
4498 | ||
2f745ad3 CW |
4499 | if (obj->base.import_attach) |
4500 | drm_prime_gem_destroy(&obj->base, NULL); | |
de151cf6 | 4501 | |
5cc9ed4b CW |
4502 | if (obj->ops->release) |
4503 | obj->ops->release(obj); | |
4504 | ||
05394f39 CW |
4505 | drm_gem_object_release(&obj->base); |
4506 | i915_gem_info_remove_obj(dev_priv, obj->base.size); | |
c397b908 | 4507 | |
05394f39 | 4508 | kfree(obj->bit_17); |
42dcedd4 | 4509 | i915_gem_object_free(obj); |
f65c9168 PZ |
4510 | |
4511 | intel_runtime_pm_put(dev_priv); | |
673a394b EA |
4512 | } |
4513 | ||
e656a6cb | 4514 | struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj, |
2f633156 | 4515 | struct i915_address_space *vm) |
e656a6cb DV |
4516 | { |
4517 | struct i915_vma *vma; | |
4518 | list_for_each_entry(vma, &obj->vma_list, vma_link) | |
4519 | if (vma->vm == vm) | |
4520 | return vma; | |
4521 | ||
4522 | return NULL; | |
4523 | } | |
4524 | ||
2f633156 BW |
4525 | void i915_gem_vma_destroy(struct i915_vma *vma) |
4526 | { | |
b9d06dd9 | 4527 | struct i915_address_space *vm = NULL; |
2f633156 | 4528 | WARN_ON(vma->node.allocated); |
aaa05667 CW |
4529 | |
4530 | /* Keep the vma as a placeholder in the execbuffer reservation lists */ | |
4531 | if (!list_empty(&vma->exec_list)) | |
4532 | return; | |
4533 | ||
b9d06dd9 | 4534 | vm = vma->vm; |
b9d06dd9 | 4535 | |
841cd773 DV |
4536 | if (!i915_is_ggtt(vm)) |
4537 | i915_ppgtt_put(i915_vm_to_ppgtt(vm)); | |
b9d06dd9 | 4538 | |
8b9c2b94 | 4539 | list_del(&vma->vma_link); |
b93dab6e | 4540 | |
2f633156 BW |
4541 | kfree(vma); |
4542 | } | |
4543 | ||
e3efda49 CW |
4544 | static void |
4545 | i915_gem_stop_ringbuffers(struct drm_device *dev) | |
4546 | { | |
4547 | struct drm_i915_private *dev_priv = dev->dev_private; | |
a4872ba6 | 4548 | struct intel_engine_cs *ring; |
e3efda49 CW |
4549 | int i; |
4550 | ||
4551 | for_each_ring(ring, dev_priv, i) | |
a83014d3 | 4552 | dev_priv->gt.stop_ring(ring); |
e3efda49 CW |
4553 | } |
4554 | ||
29105ccc | 4555 | int |
45c5f202 | 4556 | i915_gem_suspend(struct drm_device *dev) |
29105ccc | 4557 | { |
3e31c6c0 | 4558 | struct drm_i915_private *dev_priv = dev->dev_private; |
45c5f202 | 4559 | int ret = 0; |
28dfe52a | 4560 | |
45c5f202 | 4561 | mutex_lock(&dev->struct_mutex); |
f7403347 | 4562 | if (dev_priv->ums.mm_suspended) |
45c5f202 | 4563 | goto err; |
28dfe52a | 4564 | |
b2da9fe5 | 4565 | ret = i915_gpu_idle(dev); |
f7403347 | 4566 | if (ret) |
45c5f202 | 4567 | goto err; |
f7403347 | 4568 | |
b2da9fe5 | 4569 | i915_gem_retire_requests(dev); |
673a394b | 4570 | |
29105ccc | 4571 | /* Under UMS, be paranoid and evict. */ |
a39d7efc | 4572 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
6c085a72 | 4573 | i915_gem_evict_everything(dev); |
29105ccc | 4574 | |
29105ccc | 4575 | i915_kernel_lost_context(dev); |
e3efda49 | 4576 | i915_gem_stop_ringbuffers(dev); |
29105ccc | 4577 | |
45c5f202 CW |
4578 | /* Hack! Don't let anybody do execbuf while we don't control the chip. |
4579 | * We need to replace this with a semaphore, or something. | |
4580 | * And not confound ums.mm_suspended! | |
4581 | */ | |
4582 | dev_priv->ums.mm_suspended = !drm_core_check_feature(dev, | |
4583 | DRIVER_MODESET); | |
4584 | mutex_unlock(&dev->struct_mutex); | |
4585 | ||
4586 | del_timer_sync(&dev_priv->gpu_error.hangcheck_timer); | |
29105ccc | 4587 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); |
274fa1c1 | 4588 | flush_delayed_work(&dev_priv->mm.idle_work); |
29105ccc | 4589 | |
673a394b | 4590 | return 0; |
45c5f202 CW |
4591 | |
4592 | err: | |
4593 | mutex_unlock(&dev->struct_mutex); | |
4594 | return ret; | |
673a394b EA |
4595 | } |
4596 | ||
a4872ba6 | 4597 | int i915_gem_l3_remap(struct intel_engine_cs *ring, int slice) |
b9524a1e | 4598 | { |
c3787e2e | 4599 | struct drm_device *dev = ring->dev; |
3e31c6c0 | 4600 | struct drm_i915_private *dev_priv = dev->dev_private; |
35a85ac6 BW |
4601 | u32 reg_base = GEN7_L3LOG_BASE + (slice * 0x200); |
4602 | u32 *remap_info = dev_priv->l3_parity.remap_info[slice]; | |
c3787e2e | 4603 | int i, ret; |
b9524a1e | 4604 | |
040d2baa | 4605 | if (!HAS_L3_DPF(dev) || !remap_info) |
c3787e2e | 4606 | return 0; |
b9524a1e | 4607 | |
c3787e2e BW |
4608 | ret = intel_ring_begin(ring, GEN7_L3LOG_SIZE / 4 * 3); |
4609 | if (ret) | |
4610 | return ret; | |
b9524a1e | 4611 | |
c3787e2e BW |
4612 | /* |
4613 | * Note: We do not worry about the concurrent register cacheline hang | |
4614 | * here because no other code should access these registers other than | |
4615 | * at initialization time. | |
4616 | */ | |
b9524a1e | 4617 | for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) { |
c3787e2e BW |
4618 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1)); |
4619 | intel_ring_emit(ring, reg_base + i); | |
4620 | intel_ring_emit(ring, remap_info[i/4]); | |
b9524a1e BW |
4621 | } |
4622 | ||
c3787e2e | 4623 | intel_ring_advance(ring); |
b9524a1e | 4624 | |
c3787e2e | 4625 | return ret; |
b9524a1e BW |
4626 | } |
4627 | ||
f691e2f4 DV |
4628 | void i915_gem_init_swizzling(struct drm_device *dev) |
4629 | { | |
3e31c6c0 | 4630 | struct drm_i915_private *dev_priv = dev->dev_private; |
f691e2f4 | 4631 | |
11782b02 | 4632 | if (INTEL_INFO(dev)->gen < 5 || |
f691e2f4 DV |
4633 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
4634 | return; | |
4635 | ||
4636 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | | |
4637 | DISP_TILE_SURFACE_SWIZZLING); | |
4638 | ||
11782b02 DV |
4639 | if (IS_GEN5(dev)) |
4640 | return; | |
4641 | ||
f691e2f4 DV |
4642 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
4643 | if (IS_GEN6(dev)) | |
6b26c86d | 4644 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); |
8782e26c | 4645 | else if (IS_GEN7(dev)) |
6b26c86d | 4646 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); |
31a5336e BW |
4647 | else if (IS_GEN8(dev)) |
4648 | I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW)); | |
8782e26c BW |
4649 | else |
4650 | BUG(); | |
f691e2f4 | 4651 | } |
e21af88d | 4652 | |
67b1b571 CW |
4653 | static bool |
4654 | intel_enable_blt(struct drm_device *dev) | |
4655 | { | |
4656 | if (!HAS_BLT(dev)) | |
4657 | return false; | |
4658 | ||
4659 | /* The blitter was dysfunctional on early prototypes */ | |
4660 | if (IS_GEN6(dev) && dev->pdev->revision < 8) { | |
4661 | DRM_INFO("BLT not supported on this pre-production hardware;" | |
4662 | " graphics performance will be degraded.\n"); | |
4663 | return false; | |
4664 | } | |
4665 | ||
4666 | return true; | |
4667 | } | |
4668 | ||
81e7f200 VS |
4669 | static void init_unused_ring(struct drm_device *dev, u32 base) |
4670 | { | |
4671 | struct drm_i915_private *dev_priv = dev->dev_private; | |
4672 | ||
4673 | I915_WRITE(RING_CTL(base), 0); | |
4674 | I915_WRITE(RING_HEAD(base), 0); | |
4675 | I915_WRITE(RING_TAIL(base), 0); | |
4676 | I915_WRITE(RING_START(base), 0); | |
4677 | } | |
4678 | ||
4679 | static void init_unused_rings(struct drm_device *dev) | |
4680 | { | |
4681 | if (IS_I830(dev)) { | |
4682 | init_unused_ring(dev, PRB1_BASE); | |
4683 | init_unused_ring(dev, SRB0_BASE); | |
4684 | init_unused_ring(dev, SRB1_BASE); | |
4685 | init_unused_ring(dev, SRB2_BASE); | |
4686 | init_unused_ring(dev, SRB3_BASE); | |
4687 | } else if (IS_GEN2(dev)) { | |
4688 | init_unused_ring(dev, SRB0_BASE); | |
4689 | init_unused_ring(dev, SRB1_BASE); | |
4690 | } else if (IS_GEN3(dev)) { | |
4691 | init_unused_ring(dev, PRB1_BASE); | |
4692 | init_unused_ring(dev, PRB2_BASE); | |
4693 | } | |
4694 | } | |
4695 | ||
a83014d3 | 4696 | int i915_gem_init_rings(struct drm_device *dev) |
8187a2b7 | 4697 | { |
4fc7c971 | 4698 | struct drm_i915_private *dev_priv = dev->dev_private; |
8187a2b7 | 4699 | int ret; |
68f95ba9 | 4700 | |
81e7f200 VS |
4701 | /* |
4702 | * At least 830 can leave some of the unused rings | |
4703 | * "active" (ie. head != tail) after resume which | |
4704 | * will prevent c3 entry. Makes sure all unused rings | |
4705 | * are totally idle. | |
4706 | */ | |
4707 | init_unused_rings(dev); | |
4708 | ||
5c1143bb | 4709 | ret = intel_init_render_ring_buffer(dev); |
68f95ba9 | 4710 | if (ret) |
b6913e4b | 4711 | return ret; |
68f95ba9 CW |
4712 | |
4713 | if (HAS_BSD(dev)) { | |
5c1143bb | 4714 | ret = intel_init_bsd_ring_buffer(dev); |
68f95ba9 CW |
4715 | if (ret) |
4716 | goto cleanup_render_ring; | |
d1b851fc | 4717 | } |
68f95ba9 | 4718 | |
67b1b571 | 4719 | if (intel_enable_blt(dev)) { |
549f7365 CW |
4720 | ret = intel_init_blt_ring_buffer(dev); |
4721 | if (ret) | |
4722 | goto cleanup_bsd_ring; | |
4723 | } | |
4724 | ||
9a8a2213 BW |
4725 | if (HAS_VEBOX(dev)) { |
4726 | ret = intel_init_vebox_ring_buffer(dev); | |
4727 | if (ret) | |
4728 | goto cleanup_blt_ring; | |
4729 | } | |
4730 | ||
845f74a7 ZY |
4731 | if (HAS_BSD2(dev)) { |
4732 | ret = intel_init_bsd2_ring_buffer(dev); | |
4733 | if (ret) | |
4734 | goto cleanup_vebox_ring; | |
4735 | } | |
9a8a2213 | 4736 | |
99433931 | 4737 | ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000)); |
4fc7c971 | 4738 | if (ret) |
845f74a7 | 4739 | goto cleanup_bsd2_ring; |
4fc7c971 BW |
4740 | |
4741 | return 0; | |
4742 | ||
845f74a7 ZY |
4743 | cleanup_bsd2_ring: |
4744 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS2]); | |
9a8a2213 BW |
4745 | cleanup_vebox_ring: |
4746 | intel_cleanup_ring_buffer(&dev_priv->ring[VECS]); | |
4fc7c971 BW |
4747 | cleanup_blt_ring: |
4748 | intel_cleanup_ring_buffer(&dev_priv->ring[BCS]); | |
4749 | cleanup_bsd_ring: | |
4750 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); | |
4751 | cleanup_render_ring: | |
4752 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); | |
4753 | ||
4754 | return ret; | |
4755 | } | |
4756 | ||
4757 | int | |
4758 | i915_gem_init_hw(struct drm_device *dev) | |
4759 | { | |
3e31c6c0 | 4760 | struct drm_i915_private *dev_priv = dev->dev_private; |
35a85ac6 | 4761 | int ret, i; |
4fc7c971 BW |
4762 | |
4763 | if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt()) | |
4764 | return -EIO; | |
4765 | ||
59124506 | 4766 | if (dev_priv->ellc_size) |
05e21cc4 | 4767 | I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf)); |
4fc7c971 | 4768 | |
0bf21347 VS |
4769 | if (IS_HASWELL(dev)) |
4770 | I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ? | |
4771 | LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED); | |
9435373e | 4772 | |
88a2b2a3 | 4773 | if (HAS_PCH_NOP(dev)) { |
6ba844b0 DV |
4774 | if (IS_IVYBRIDGE(dev)) { |
4775 | u32 temp = I915_READ(GEN7_MSG_CTL); | |
4776 | temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK); | |
4777 | I915_WRITE(GEN7_MSG_CTL, temp); | |
4778 | } else if (INTEL_INFO(dev)->gen >= 7) { | |
4779 | u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT); | |
4780 | temp &= ~RESET_PCH_HANDSHAKE_ENABLE; | |
4781 | I915_WRITE(HSW_NDE_RSTWRN_OPT, temp); | |
4782 | } | |
88a2b2a3 BW |
4783 | } |
4784 | ||
4fc7c971 BW |
4785 | i915_gem_init_swizzling(dev); |
4786 | ||
a83014d3 | 4787 | ret = dev_priv->gt.init_rings(dev); |
99433931 MK |
4788 | if (ret) |
4789 | return ret; | |
4790 | ||
c3787e2e BW |
4791 | for (i = 0; i < NUM_L3_SLICES(dev); i++) |
4792 | i915_gem_l3_remap(&dev_priv->ring[RCS], i); | |
4793 | ||
254f965c | 4794 | /* |
2fa48d8d BW |
4795 | * XXX: Contexts should only be initialized once. Doing a switch to the |
4796 | * default context switch however is something we'd like to do after | |
4797 | * reset or thaw (the latter may not actually be necessary for HW, but | |
4798 | * goes with our code better). Context switching requires rings (for | |
4799 | * the do_switch), but before enabling PPGTT. So don't move this. | |
254f965c | 4800 | */ |
2fa48d8d | 4801 | ret = i915_gem_context_enable(dev_priv); |
60990320 | 4802 | if (ret && ret != -EIO) { |
2fa48d8d | 4803 | DRM_ERROR("Context enable failed %d\n", ret); |
60990320 | 4804 | i915_gem_cleanup_ringbuffer(dev); |
82460d97 DV |
4805 | |
4806 | return ret; | |
4807 | } | |
4808 | ||
4809 | ret = i915_ppgtt_init_hw(dev); | |
4810 | if (ret && ret != -EIO) { | |
4811 | DRM_ERROR("PPGTT enable failed %d\n", ret); | |
4812 | i915_gem_cleanup_ringbuffer(dev); | |
b7c36d25 | 4813 | } |
e21af88d | 4814 | |
2fa48d8d | 4815 | return ret; |
8187a2b7 ZN |
4816 | } |
4817 | ||
1070a42b CW |
4818 | int i915_gem_init(struct drm_device *dev) |
4819 | { | |
4820 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1070a42b CW |
4821 | int ret; |
4822 | ||
127f1003 OM |
4823 | i915.enable_execlists = intel_sanitize_enable_execlists(dev, |
4824 | i915.enable_execlists); | |
4825 | ||
1070a42b | 4826 | mutex_lock(&dev->struct_mutex); |
d62b4892 JB |
4827 | |
4828 | if (IS_VALLEYVIEW(dev)) { | |
4829 | /* VLVA0 (potential hack), BIOS isn't actually waking us */ | |
981a5aea ID |
4830 | I915_WRITE(VLV_GTLC_WAKE_CTRL, VLV_GTLC_ALLOWWAKEREQ); |
4831 | if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) & | |
4832 | VLV_GTLC_ALLOWWAKEACK), 10)) | |
d62b4892 JB |
4833 | DRM_DEBUG_DRIVER("allow wake ack timed out\n"); |
4834 | } | |
4835 | ||
a83014d3 OM |
4836 | if (!i915.enable_execlists) { |
4837 | dev_priv->gt.do_execbuf = i915_gem_ringbuffer_submission; | |
4838 | dev_priv->gt.init_rings = i915_gem_init_rings; | |
4839 | dev_priv->gt.cleanup_ring = intel_cleanup_ring_buffer; | |
4840 | dev_priv->gt.stop_ring = intel_stop_ring_buffer; | |
454afebd OM |
4841 | } else { |
4842 | dev_priv->gt.do_execbuf = intel_execlists_submission; | |
4843 | dev_priv->gt.init_rings = intel_logical_rings_init; | |
4844 | dev_priv->gt.cleanup_ring = intel_logical_ring_cleanup; | |
4845 | dev_priv->gt.stop_ring = intel_logical_ring_stop; | |
a83014d3 OM |
4846 | } |
4847 | ||
6c5566a8 DV |
4848 | ret = i915_gem_init_userptr(dev); |
4849 | if (ret) { | |
4850 | mutex_unlock(&dev->struct_mutex); | |
4851 | return ret; | |
4852 | } | |
4853 | ||
d7e5008f | 4854 | i915_gem_init_global_gtt(dev); |
d62b4892 | 4855 | |
2fa48d8d | 4856 | ret = i915_gem_context_init(dev); |
e3848694 MK |
4857 | if (ret) { |
4858 | mutex_unlock(&dev->struct_mutex); | |
2fa48d8d | 4859 | return ret; |
e3848694 | 4860 | } |
2fa48d8d | 4861 | |
1070a42b | 4862 | ret = i915_gem_init_hw(dev); |
60990320 CW |
4863 | if (ret == -EIO) { |
4864 | /* Allow ring initialisation to fail by marking the GPU as | |
4865 | * wedged. But we only want to do this where the GPU is angry, | |
4866 | * for all other failure, such as an allocation failure, bail. | |
4867 | */ | |
4868 | DRM_ERROR("Failed to initialize GPU, declaring it wedged\n"); | |
4869 | atomic_set_mask(I915_WEDGED, &dev_priv->gpu_error.reset_counter); | |
4870 | ret = 0; | |
1070a42b | 4871 | } |
60990320 | 4872 | mutex_unlock(&dev->struct_mutex); |
1070a42b | 4873 | |
53ca26ca DV |
4874 | /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */ |
4875 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) | |
4876 | dev_priv->dri1.allow_batchbuffer = 1; | |
60990320 | 4877 | return ret; |
1070a42b CW |
4878 | } |
4879 | ||
8187a2b7 ZN |
4880 | void |
4881 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) | |
4882 | { | |
3e31c6c0 | 4883 | struct drm_i915_private *dev_priv = dev->dev_private; |
a4872ba6 | 4884 | struct intel_engine_cs *ring; |
1ec14ad3 | 4885 | int i; |
8187a2b7 | 4886 | |
b4519513 | 4887 | for_each_ring(ring, dev_priv, i) |
a83014d3 | 4888 | dev_priv->gt.cleanup_ring(ring); |
8187a2b7 ZN |
4889 | } |
4890 | ||
673a394b EA |
4891 | int |
4892 | i915_gem_entervt_ioctl(struct drm_device *dev, void *data, | |
4893 | struct drm_file *file_priv) | |
4894 | { | |
db1b76ca | 4895 | struct drm_i915_private *dev_priv = dev->dev_private; |
b4519513 | 4896 | int ret; |
673a394b | 4897 | |
79e53945 JB |
4898 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4899 | return 0; | |
4900 | ||
1f83fee0 | 4901 | if (i915_reset_in_progress(&dev_priv->gpu_error)) { |
673a394b | 4902 | DRM_ERROR("Reenabling wedged hardware, good luck\n"); |
1f83fee0 | 4903 | atomic_set(&dev_priv->gpu_error.reset_counter, 0); |
673a394b EA |
4904 | } |
4905 | ||
673a394b | 4906 | mutex_lock(&dev->struct_mutex); |
db1b76ca | 4907 | dev_priv->ums.mm_suspended = 0; |
9bb2d6f9 | 4908 | |
f691e2f4 | 4909 | ret = i915_gem_init_hw(dev); |
d816f6ac WF |
4910 | if (ret != 0) { |
4911 | mutex_unlock(&dev->struct_mutex); | |
9bb2d6f9 | 4912 | return ret; |
d816f6ac | 4913 | } |
9bb2d6f9 | 4914 | |
5cef07e1 | 4915 | BUG_ON(!list_empty(&dev_priv->gtt.base.active_list)); |
dbb19d30 | 4916 | |
bb0f1b5c | 4917 | ret = drm_irq_install(dev, dev->pdev->irq); |
5f35308b CW |
4918 | if (ret) |
4919 | goto cleanup_ringbuffer; | |
e090c53b | 4920 | mutex_unlock(&dev->struct_mutex); |
dbb19d30 | 4921 | |
673a394b | 4922 | return 0; |
5f35308b CW |
4923 | |
4924 | cleanup_ringbuffer: | |
5f35308b | 4925 | i915_gem_cleanup_ringbuffer(dev); |
db1b76ca | 4926 | dev_priv->ums.mm_suspended = 1; |
5f35308b CW |
4927 | mutex_unlock(&dev->struct_mutex); |
4928 | ||
4929 | return ret; | |
673a394b EA |
4930 | } |
4931 | ||
4932 | int | |
4933 | i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, | |
4934 | struct drm_file *file_priv) | |
4935 | { | |
79e53945 JB |
4936 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4937 | return 0; | |
4938 | ||
e090c53b | 4939 | mutex_lock(&dev->struct_mutex); |
dbb19d30 | 4940 | drm_irq_uninstall(dev); |
e090c53b | 4941 | mutex_unlock(&dev->struct_mutex); |
db1b76ca | 4942 | |
45c5f202 | 4943 | return i915_gem_suspend(dev); |
673a394b EA |
4944 | } |
4945 | ||
4946 | void | |
4947 | i915_gem_lastclose(struct drm_device *dev) | |
4948 | { | |
4949 | int ret; | |
673a394b | 4950 | |
e806b495 EA |
4951 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4952 | return; | |
4953 | ||
45c5f202 | 4954 | ret = i915_gem_suspend(dev); |
6dbe2772 KP |
4955 | if (ret) |
4956 | DRM_ERROR("failed to idle hardware: %d\n", ret); | |
673a394b EA |
4957 | } |
4958 | ||
64193406 | 4959 | static void |
a4872ba6 | 4960 | init_ring_lists(struct intel_engine_cs *ring) |
64193406 CW |
4961 | { |
4962 | INIT_LIST_HEAD(&ring->active_list); | |
4963 | INIT_LIST_HEAD(&ring->request_list); | |
64193406 CW |
4964 | } |
4965 | ||
7e0d96bc BW |
4966 | void i915_init_vm(struct drm_i915_private *dev_priv, |
4967 | struct i915_address_space *vm) | |
fc8c067e | 4968 | { |
7e0d96bc BW |
4969 | if (!i915_is_ggtt(vm)) |
4970 | drm_mm_init(&vm->mm, vm->start, vm->total); | |
fc8c067e BW |
4971 | vm->dev = dev_priv->dev; |
4972 | INIT_LIST_HEAD(&vm->active_list); | |
4973 | INIT_LIST_HEAD(&vm->inactive_list); | |
4974 | INIT_LIST_HEAD(&vm->global_link); | |
f72d21ed | 4975 | list_add_tail(&vm->global_link, &dev_priv->vm_list); |
fc8c067e BW |
4976 | } |
4977 | ||
673a394b EA |
4978 | void |
4979 | i915_gem_load(struct drm_device *dev) | |
4980 | { | |
3e31c6c0 | 4981 | struct drm_i915_private *dev_priv = dev->dev_private; |
42dcedd4 CW |
4982 | int i; |
4983 | ||
4984 | dev_priv->slab = | |
4985 | kmem_cache_create("i915_gem_object", | |
4986 | sizeof(struct drm_i915_gem_object), 0, | |
4987 | SLAB_HWCACHE_ALIGN, | |
4988 | NULL); | |
673a394b | 4989 | |
fc8c067e BW |
4990 | INIT_LIST_HEAD(&dev_priv->vm_list); |
4991 | i915_init_vm(dev_priv, &dev_priv->gtt.base); | |
4992 | ||
a33afea5 | 4993 | INIT_LIST_HEAD(&dev_priv->context_list); |
6c085a72 CW |
4994 | INIT_LIST_HEAD(&dev_priv->mm.unbound_list); |
4995 | INIT_LIST_HEAD(&dev_priv->mm.bound_list); | |
a09ba7fa | 4996 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
1ec14ad3 CW |
4997 | for (i = 0; i < I915_NUM_RINGS; i++) |
4998 | init_ring_lists(&dev_priv->ring[i]); | |
4b9de737 | 4999 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) |
007cc8ac | 5000 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); |
673a394b EA |
5001 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, |
5002 | i915_gem_retire_work_handler); | |
b29c19b6 CW |
5003 | INIT_DELAYED_WORK(&dev_priv->mm.idle_work, |
5004 | i915_gem_idle_work_handler); | |
1f83fee0 | 5005 | init_waitqueue_head(&dev_priv->gpu_error.reset_queue); |
31169714 | 5006 | |
94400120 | 5007 | /* On GEN3 we really need to make sure the ARB C3 LP bit is set */ |
dbb42748 | 5008 | if (!drm_core_check_feature(dev, DRIVER_MODESET) && IS_GEN3(dev)) { |
50743298 DV |
5009 | I915_WRITE(MI_ARB_STATE, |
5010 | _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE)); | |
94400120 DA |
5011 | } |
5012 | ||
72bfa19c CW |
5013 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; |
5014 | ||
de151cf6 | 5015 | /* Old X drivers will take 0-2 for front, back, depth buffers */ |
b397c836 EA |
5016 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
5017 | dev_priv->fence_reg_start = 3; | |
de151cf6 | 5018 | |
42b5aeab VS |
5019 | if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev)) |
5020 | dev_priv->num_fence_regs = 32; | |
5021 | else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) | |
de151cf6 JB |
5022 | dev_priv->num_fence_regs = 16; |
5023 | else | |
5024 | dev_priv->num_fence_regs = 8; | |
5025 | ||
b5aa8a0f | 5026 | /* Initialize fence registers to zero */ |
19b2dbde CW |
5027 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
5028 | i915_gem_restore_fences(dev); | |
10ed13e4 | 5029 | |
673a394b | 5030 | i915_gem_detect_bit_6_swizzle(dev); |
6b95a207 | 5031 | init_waitqueue_head(&dev_priv->pending_flip_queue); |
17250b71 | 5032 | |
ce453d81 CW |
5033 | dev_priv->mm.interruptible = true; |
5034 | ||
ceabbba5 CW |
5035 | dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan; |
5036 | dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count; | |
5037 | dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS; | |
5038 | register_shrinker(&dev_priv->mm.shrinker); | |
2cfcd32a CW |
5039 | |
5040 | dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom; | |
5041 | register_oom_notifier(&dev_priv->mm.oom_notifier); | |
f99d7069 DV |
5042 | |
5043 | mutex_init(&dev_priv->fb_tracking.lock); | |
673a394b | 5044 | } |
71acb5eb | 5045 | |
f787a5f5 | 5046 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
b962442e | 5047 | { |
f787a5f5 | 5048 | struct drm_i915_file_private *file_priv = file->driver_priv; |
b962442e | 5049 | |
b29c19b6 CW |
5050 | cancel_delayed_work_sync(&file_priv->mm.idle_work); |
5051 | ||
b962442e EA |
5052 | /* Clean up our request list when the client is going away, so that |
5053 | * later retire_requests won't dereference our soon-to-be-gone | |
5054 | * file_priv. | |
5055 | */ | |
1c25595f | 5056 | spin_lock(&file_priv->mm.lock); |
f787a5f5 CW |
5057 | while (!list_empty(&file_priv->mm.request_list)) { |
5058 | struct drm_i915_gem_request *request; | |
5059 | ||
5060 | request = list_first_entry(&file_priv->mm.request_list, | |
5061 | struct drm_i915_gem_request, | |
5062 | client_list); | |
5063 | list_del(&request->client_list); | |
5064 | request->file_priv = NULL; | |
5065 | } | |
1c25595f | 5066 | spin_unlock(&file_priv->mm.lock); |
b962442e | 5067 | } |
31169714 | 5068 | |
b29c19b6 CW |
5069 | static void |
5070 | i915_gem_file_idle_work_handler(struct work_struct *work) | |
5071 | { | |
5072 | struct drm_i915_file_private *file_priv = | |
5073 | container_of(work, typeof(*file_priv), mm.idle_work.work); | |
5074 | ||
5075 | atomic_set(&file_priv->rps_wait_boost, false); | |
5076 | } | |
5077 | ||
5078 | int i915_gem_open(struct drm_device *dev, struct drm_file *file) | |
5079 | { | |
5080 | struct drm_i915_file_private *file_priv; | |
e422b888 | 5081 | int ret; |
b29c19b6 CW |
5082 | |
5083 | DRM_DEBUG_DRIVER("\n"); | |
5084 | ||
5085 | file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL); | |
5086 | if (!file_priv) | |
5087 | return -ENOMEM; | |
5088 | ||
5089 | file->driver_priv = file_priv; | |
5090 | file_priv->dev_priv = dev->dev_private; | |
ab0e7ff9 | 5091 | file_priv->file = file; |
b29c19b6 CW |
5092 | |
5093 | spin_lock_init(&file_priv->mm.lock); | |
5094 | INIT_LIST_HEAD(&file_priv->mm.request_list); | |
5095 | INIT_DELAYED_WORK(&file_priv->mm.idle_work, | |
5096 | i915_gem_file_idle_work_handler); | |
5097 | ||
e422b888 BW |
5098 | ret = i915_gem_context_open(dev, file); |
5099 | if (ret) | |
5100 | kfree(file_priv); | |
b29c19b6 | 5101 | |
e422b888 | 5102 | return ret; |
b29c19b6 CW |
5103 | } |
5104 | ||
a071fa00 DV |
5105 | void i915_gem_track_fb(struct drm_i915_gem_object *old, |
5106 | struct drm_i915_gem_object *new, | |
5107 | unsigned frontbuffer_bits) | |
5108 | { | |
5109 | if (old) { | |
5110 | WARN_ON(!mutex_is_locked(&old->base.dev->struct_mutex)); | |
5111 | WARN_ON(!(old->frontbuffer_bits & frontbuffer_bits)); | |
5112 | old->frontbuffer_bits &= ~frontbuffer_bits; | |
5113 | } | |
5114 | ||
5115 | if (new) { | |
5116 | WARN_ON(!mutex_is_locked(&new->base.dev->struct_mutex)); | |
5117 | WARN_ON(new->frontbuffer_bits & frontbuffer_bits); | |
5118 | new->frontbuffer_bits |= frontbuffer_bits; | |
5119 | } | |
5120 | } | |
5121 | ||
5774506f CW |
5122 | static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task) |
5123 | { | |
5124 | if (!mutex_is_locked(mutex)) | |
5125 | return false; | |
5126 | ||
5127 | #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES) | |
5128 | return mutex->owner == task; | |
5129 | #else | |
5130 | /* Since UP may be pre-empted, we cannot assume that we own the lock */ | |
5131 | return false; | |
5132 | #endif | |
5133 | } | |
5134 | ||
b453c4db CW |
5135 | static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock) |
5136 | { | |
5137 | if (!mutex_trylock(&dev->struct_mutex)) { | |
5138 | if (!mutex_is_locked_by(&dev->struct_mutex, current)) | |
5139 | return false; | |
5140 | ||
5141 | if (to_i915(dev)->mm.shrinker_no_lock_stealing) | |
5142 | return false; | |
5143 | ||
5144 | *unlock = false; | |
5145 | } else | |
5146 | *unlock = true; | |
5147 | ||
5148 | return true; | |
5149 | } | |
5150 | ||
ceabbba5 CW |
5151 | static int num_vma_bound(struct drm_i915_gem_object *obj) |
5152 | { | |
5153 | struct i915_vma *vma; | |
5154 | int count = 0; | |
5155 | ||
5156 | list_for_each_entry(vma, &obj->vma_list, vma_link) | |
5157 | if (drm_mm_node_allocated(&vma->node)) | |
5158 | count++; | |
5159 | ||
5160 | return count; | |
5161 | } | |
5162 | ||
7dc19d5a | 5163 | static unsigned long |
ceabbba5 | 5164 | i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc) |
31169714 | 5165 | { |
17250b71 | 5166 | struct drm_i915_private *dev_priv = |
ceabbba5 | 5167 | container_of(shrinker, struct drm_i915_private, mm.shrinker); |
17250b71 | 5168 | struct drm_device *dev = dev_priv->dev; |
6c085a72 | 5169 | struct drm_i915_gem_object *obj; |
7dc19d5a | 5170 | unsigned long count; |
b453c4db | 5171 | bool unlock; |
17250b71 | 5172 | |
b453c4db CW |
5173 | if (!i915_gem_shrinker_lock(dev, &unlock)) |
5174 | return 0; | |
31169714 | 5175 | |
7dc19d5a | 5176 | count = 0; |
35c20a60 | 5177 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) |
a5570178 | 5178 | if (obj->pages_pin_count == 0) |
7dc19d5a | 5179 | count += obj->base.size >> PAGE_SHIFT; |
fcb4a578 BW |
5180 | |
5181 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { | |
ceabbba5 CW |
5182 | if (!i915_gem_obj_is_pinned(obj) && |
5183 | obj->pages_pin_count == num_vma_bound(obj)) | |
7dc19d5a | 5184 | count += obj->base.size >> PAGE_SHIFT; |
fcb4a578 | 5185 | } |
17250b71 | 5186 | |
5774506f CW |
5187 | if (unlock) |
5188 | mutex_unlock(&dev->struct_mutex); | |
d9973b43 | 5189 | |
7dc19d5a | 5190 | return count; |
31169714 | 5191 | } |
a70a3148 BW |
5192 | |
5193 | /* All the new VM stuff */ | |
5194 | unsigned long i915_gem_obj_offset(struct drm_i915_gem_object *o, | |
5195 | struct i915_address_space *vm) | |
5196 | { | |
5197 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; | |
5198 | struct i915_vma *vma; | |
5199 | ||
896ab1a5 | 5200 | WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base); |
a70a3148 | 5201 | |
a70a3148 BW |
5202 | list_for_each_entry(vma, &o->vma_list, vma_link) { |
5203 | if (vma->vm == vm) | |
5204 | return vma->node.start; | |
5205 | ||
5206 | } | |
f25748ea DV |
5207 | WARN(1, "%s vma for this object not found.\n", |
5208 | i915_is_ggtt(vm) ? "global" : "ppgtt"); | |
a70a3148 BW |
5209 | return -1; |
5210 | } | |
5211 | ||
5212 | bool i915_gem_obj_bound(struct drm_i915_gem_object *o, | |
5213 | struct i915_address_space *vm) | |
5214 | { | |
5215 | struct i915_vma *vma; | |
5216 | ||
5217 | list_for_each_entry(vma, &o->vma_list, vma_link) | |
8b9c2b94 | 5218 | if (vma->vm == vm && drm_mm_node_allocated(&vma->node)) |
a70a3148 BW |
5219 | return true; |
5220 | ||
5221 | return false; | |
5222 | } | |
5223 | ||
5224 | bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o) | |
5225 | { | |
5a1d5eb0 | 5226 | struct i915_vma *vma; |
a70a3148 | 5227 | |
5a1d5eb0 CW |
5228 | list_for_each_entry(vma, &o->vma_list, vma_link) |
5229 | if (drm_mm_node_allocated(&vma->node)) | |
a70a3148 BW |
5230 | return true; |
5231 | ||
5232 | return false; | |
5233 | } | |
5234 | ||
5235 | unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o, | |
5236 | struct i915_address_space *vm) | |
5237 | { | |
5238 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; | |
5239 | struct i915_vma *vma; | |
5240 | ||
896ab1a5 | 5241 | WARN_ON(vm == &dev_priv->mm.aliasing_ppgtt->base); |
a70a3148 BW |
5242 | |
5243 | BUG_ON(list_empty(&o->vma_list)); | |
5244 | ||
5245 | list_for_each_entry(vma, &o->vma_list, vma_link) | |
5246 | if (vma->vm == vm) | |
5247 | return vma->node.size; | |
5248 | ||
5249 | return 0; | |
5250 | } | |
5251 | ||
7dc19d5a | 5252 | static unsigned long |
ceabbba5 | 5253 | i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc) |
7dc19d5a DC |
5254 | { |
5255 | struct drm_i915_private *dev_priv = | |
ceabbba5 | 5256 | container_of(shrinker, struct drm_i915_private, mm.shrinker); |
7dc19d5a | 5257 | struct drm_device *dev = dev_priv->dev; |
7dc19d5a | 5258 | unsigned long freed; |
b453c4db | 5259 | bool unlock; |
7dc19d5a | 5260 | |
b453c4db CW |
5261 | if (!i915_gem_shrinker_lock(dev, &unlock)) |
5262 | return SHRINK_STOP; | |
7dc19d5a | 5263 | |
d9973b43 CW |
5264 | freed = i915_gem_purge(dev_priv, sc->nr_to_scan); |
5265 | if (freed < sc->nr_to_scan) | |
5266 | freed += __i915_gem_shrink(dev_priv, | |
5267 | sc->nr_to_scan - freed, | |
5268 | false); | |
7dc19d5a DC |
5269 | if (unlock) |
5270 | mutex_unlock(&dev->struct_mutex); | |
d9973b43 | 5271 | |
7dc19d5a DC |
5272 | return freed; |
5273 | } | |
5c2abbea | 5274 | |
2cfcd32a CW |
5275 | static int |
5276 | i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr) | |
5277 | { | |
5278 | struct drm_i915_private *dev_priv = | |
5279 | container_of(nb, struct drm_i915_private, mm.oom_notifier); | |
5280 | struct drm_device *dev = dev_priv->dev; | |
5281 | struct drm_i915_gem_object *obj; | |
5282 | unsigned long timeout = msecs_to_jiffies(5000) + 1; | |
5283 | unsigned long pinned, bound, unbound, freed; | |
5284 | bool was_interruptible; | |
5285 | bool unlock; | |
5286 | ||
a1db2fa7 | 5287 | while (!i915_gem_shrinker_lock(dev, &unlock) && --timeout) { |
2cfcd32a | 5288 | schedule_timeout_killable(1); |
a1db2fa7 CW |
5289 | if (fatal_signal_pending(current)) |
5290 | return NOTIFY_DONE; | |
5291 | } | |
2cfcd32a CW |
5292 | if (timeout == 0) { |
5293 | pr_err("Unable to purge GPU memory due lock contention.\n"); | |
5294 | return NOTIFY_DONE; | |
5295 | } | |
5296 | ||
5297 | was_interruptible = dev_priv->mm.interruptible; | |
5298 | dev_priv->mm.interruptible = false; | |
5299 | ||
5300 | freed = i915_gem_shrink_all(dev_priv); | |
5301 | ||
5302 | dev_priv->mm.interruptible = was_interruptible; | |
5303 | ||
5304 | /* Because we may be allocating inside our own driver, we cannot | |
5305 | * assert that there are no objects with pinned pages that are not | |
5306 | * being pointed to by hardware. | |
5307 | */ | |
5308 | unbound = bound = pinned = 0; | |
5309 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) { | |
5310 | if (!obj->base.filp) /* not backed by a freeable object */ | |
5311 | continue; | |
5312 | ||
5313 | if (obj->pages_pin_count) | |
5314 | pinned += obj->base.size; | |
5315 | else | |
5316 | unbound += obj->base.size; | |
5317 | } | |
5318 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { | |
5319 | if (!obj->base.filp) | |
5320 | continue; | |
5321 | ||
5322 | if (obj->pages_pin_count) | |
5323 | pinned += obj->base.size; | |
5324 | else | |
5325 | bound += obj->base.size; | |
5326 | } | |
5327 | ||
5328 | if (unlock) | |
5329 | mutex_unlock(&dev->struct_mutex); | |
5330 | ||
5331 | pr_info("Purging GPU memory, %lu bytes freed, %lu bytes still pinned.\n", | |
5332 | freed, pinned); | |
5333 | if (unbound || bound) | |
5334 | pr_err("%lu and %lu bytes still available in the " | |
5335 | "bound and unbound GPU page lists.\n", | |
5336 | bound, unbound); | |
5337 | ||
5338 | *(unsigned long *)ptr += freed; | |
5339 | return NOTIFY_DONE; | |
5340 | } | |
5341 | ||
5c2abbea BW |
5342 | struct i915_vma *i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj) |
5343 | { | |
5344 | struct i915_vma *vma; | |
5345 | ||
5c2abbea | 5346 | vma = list_first_entry(&obj->vma_list, typeof(*vma), vma_link); |
5dc383b0 | 5347 | if (vma->vm != i915_obj_to_ggtt(obj)) |
5c2abbea BW |
5348 | return NULL; |
5349 | ||
5350 | return vma; | |
5351 | } |