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" |
5949eac4 | 34 | #include <linux/shmem_fs.h> |
5a0e3ad6 | 35 | #include <linux/slab.h> |
673a394b | 36 | #include <linux/swap.h> |
79e53945 | 37 | #include <linux/pci.h> |
1286ff73 | 38 | #include <linux/dma-buf.h> |
673a394b | 39 | |
05394f39 | 40 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); |
2c22569b CW |
41 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj, |
42 | bool force); | |
07fe0b12 BW |
43 | static __must_check int |
44 | i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, | |
45 | struct i915_address_space *vm, | |
46 | unsigned alignment, | |
47 | bool map_and_fenceable, | |
48 | bool nonblocking); | |
05394f39 CW |
49 | static int i915_gem_phys_pwrite(struct drm_device *dev, |
50 | struct drm_i915_gem_object *obj, | |
71acb5eb | 51 | struct drm_i915_gem_pwrite *args, |
05394f39 | 52 | struct drm_file *file); |
673a394b | 53 | |
61050808 CW |
54 | static void i915_gem_write_fence(struct drm_device *dev, int reg, |
55 | struct drm_i915_gem_object *obj); | |
56 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, | |
57 | struct drm_i915_fence_reg *fence, | |
58 | bool enable); | |
59 | ||
17250b71 | 60 | static int i915_gem_inactive_shrink(struct shrinker *shrinker, |
1495f230 | 61 | struct shrink_control *sc); |
6c085a72 CW |
62 | static long i915_gem_purge(struct drm_i915_private *dev_priv, long target); |
63 | static void i915_gem_shrink_all(struct drm_i915_private *dev_priv); | |
8c59967c | 64 | static void i915_gem_object_truncate(struct drm_i915_gem_object *obj); |
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) |
1b50247a | 202 | if (obj->pin_count) |
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 | ||
42dcedd4 CW |
212 | void *i915_gem_object_alloc(struct drm_device *dev) |
213 | { | |
214 | struct drm_i915_private *dev_priv = dev->dev_private; | |
fac15c10 | 215 | return kmem_cache_zalloc(dev_priv->slab, GFP_KERNEL); |
42dcedd4 CW |
216 | } |
217 | ||
218 | void i915_gem_object_free(struct drm_i915_gem_object *obj) | |
219 | { | |
220 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
221 | kmem_cache_free(dev_priv->slab, obj); | |
222 | } | |
223 | ||
ff72145b DA |
224 | static int |
225 | i915_gem_create(struct drm_file *file, | |
226 | struct drm_device *dev, | |
227 | uint64_t size, | |
228 | uint32_t *handle_p) | |
673a394b | 229 | { |
05394f39 | 230 | struct drm_i915_gem_object *obj; |
a1a2d1d3 PP |
231 | int ret; |
232 | u32 handle; | |
673a394b | 233 | |
ff72145b | 234 | size = roundup(size, PAGE_SIZE); |
8ffc0246 CW |
235 | if (size == 0) |
236 | return -EINVAL; | |
673a394b EA |
237 | |
238 | /* Allocate the new object */ | |
ff72145b | 239 | obj = i915_gem_alloc_object(dev, size); |
673a394b EA |
240 | if (obj == NULL) |
241 | return -ENOMEM; | |
242 | ||
05394f39 | 243 | ret = drm_gem_handle_create(file, &obj->base, &handle); |
202f2fef | 244 | /* drop reference from allocate - handle holds it now */ |
d861e338 DV |
245 | drm_gem_object_unreference_unlocked(&obj->base); |
246 | if (ret) | |
247 | return ret; | |
202f2fef | 248 | |
ff72145b | 249 | *handle_p = handle; |
673a394b EA |
250 | return 0; |
251 | } | |
252 | ||
ff72145b DA |
253 | int |
254 | i915_gem_dumb_create(struct drm_file *file, | |
255 | struct drm_device *dev, | |
256 | struct drm_mode_create_dumb *args) | |
257 | { | |
258 | /* have to work out size/pitch and return them */ | |
ed0291fd | 259 | args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64); |
ff72145b DA |
260 | args->size = args->pitch * args->height; |
261 | return i915_gem_create(file, dev, | |
262 | args->size, &args->handle); | |
263 | } | |
264 | ||
ff72145b DA |
265 | /** |
266 | * Creates a new mm object and returns a handle to it. | |
267 | */ | |
268 | int | |
269 | i915_gem_create_ioctl(struct drm_device *dev, void *data, | |
270 | struct drm_file *file) | |
271 | { | |
272 | struct drm_i915_gem_create *args = data; | |
63ed2cb2 | 273 | |
ff72145b DA |
274 | return i915_gem_create(file, dev, |
275 | args->size, &args->handle); | |
276 | } | |
277 | ||
8461d226 DV |
278 | static inline int |
279 | __copy_to_user_swizzled(char __user *cpu_vaddr, | |
280 | const char *gpu_vaddr, int gpu_offset, | |
281 | int length) | |
282 | { | |
283 | int ret, cpu_offset = 0; | |
284 | ||
285 | while (length > 0) { | |
286 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
287 | int this_length = min(cacheline_end - gpu_offset, length); | |
288 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
289 | ||
290 | ret = __copy_to_user(cpu_vaddr + cpu_offset, | |
291 | gpu_vaddr + swizzled_gpu_offset, | |
292 | this_length); | |
293 | if (ret) | |
294 | return ret + length; | |
295 | ||
296 | cpu_offset += this_length; | |
297 | gpu_offset += this_length; | |
298 | length -= this_length; | |
299 | } | |
300 | ||
301 | return 0; | |
302 | } | |
303 | ||
8c59967c | 304 | static inline int |
4f0c7cfb BW |
305 | __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, |
306 | const char __user *cpu_vaddr, | |
8c59967c DV |
307 | int length) |
308 | { | |
309 | int ret, cpu_offset = 0; | |
310 | ||
311 | while (length > 0) { | |
312 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
313 | int this_length = min(cacheline_end - gpu_offset, length); | |
314 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
315 | ||
316 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, | |
317 | cpu_vaddr + cpu_offset, | |
318 | this_length); | |
319 | if (ret) | |
320 | return ret + length; | |
321 | ||
322 | cpu_offset += this_length; | |
323 | gpu_offset += this_length; | |
324 | length -= this_length; | |
325 | } | |
326 | ||
327 | return 0; | |
328 | } | |
329 | ||
d174bd64 DV |
330 | /* Per-page copy function for the shmem pread fastpath. |
331 | * Flushes invalid cachelines before reading the target if | |
332 | * needs_clflush is set. */ | |
eb01459f | 333 | static int |
d174bd64 DV |
334 | shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length, |
335 | char __user *user_data, | |
336 | bool page_do_bit17_swizzling, bool needs_clflush) | |
337 | { | |
338 | char *vaddr; | |
339 | int ret; | |
340 | ||
e7e58eb5 | 341 | if (unlikely(page_do_bit17_swizzling)) |
d174bd64 DV |
342 | return -EINVAL; |
343 | ||
344 | vaddr = kmap_atomic(page); | |
345 | if (needs_clflush) | |
346 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
347 | page_length); | |
348 | ret = __copy_to_user_inatomic(user_data, | |
349 | vaddr + shmem_page_offset, | |
350 | page_length); | |
351 | kunmap_atomic(vaddr); | |
352 | ||
f60d7f0c | 353 | return ret ? -EFAULT : 0; |
d174bd64 DV |
354 | } |
355 | ||
23c18c71 DV |
356 | static void |
357 | shmem_clflush_swizzled_range(char *addr, unsigned long length, | |
358 | bool swizzled) | |
359 | { | |
e7e58eb5 | 360 | if (unlikely(swizzled)) { |
23c18c71 DV |
361 | unsigned long start = (unsigned long) addr; |
362 | unsigned long end = (unsigned long) addr + length; | |
363 | ||
364 | /* For swizzling simply ensure that we always flush both | |
365 | * channels. Lame, but simple and it works. Swizzled | |
366 | * pwrite/pread is far from a hotpath - current userspace | |
367 | * doesn't use it at all. */ | |
368 | start = round_down(start, 128); | |
369 | end = round_up(end, 128); | |
370 | ||
371 | drm_clflush_virt_range((void *)start, end - start); | |
372 | } else { | |
373 | drm_clflush_virt_range(addr, length); | |
374 | } | |
375 | ||
376 | } | |
377 | ||
d174bd64 DV |
378 | /* Only difference to the fast-path function is that this can handle bit17 |
379 | * and uses non-atomic copy and kmap functions. */ | |
380 | static int | |
381 | shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length, | |
382 | char __user *user_data, | |
383 | bool page_do_bit17_swizzling, bool needs_clflush) | |
384 | { | |
385 | char *vaddr; | |
386 | int ret; | |
387 | ||
388 | vaddr = kmap(page); | |
389 | if (needs_clflush) | |
23c18c71 DV |
390 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
391 | page_length, | |
392 | page_do_bit17_swizzling); | |
d174bd64 DV |
393 | |
394 | if (page_do_bit17_swizzling) | |
395 | ret = __copy_to_user_swizzled(user_data, | |
396 | vaddr, shmem_page_offset, | |
397 | page_length); | |
398 | else | |
399 | ret = __copy_to_user(user_data, | |
400 | vaddr + shmem_page_offset, | |
401 | page_length); | |
402 | kunmap(page); | |
403 | ||
f60d7f0c | 404 | return ret ? - EFAULT : 0; |
d174bd64 DV |
405 | } |
406 | ||
eb01459f | 407 | static int |
dbf7bff0 DV |
408 | i915_gem_shmem_pread(struct drm_device *dev, |
409 | struct drm_i915_gem_object *obj, | |
410 | struct drm_i915_gem_pread *args, | |
411 | struct drm_file *file) | |
eb01459f | 412 | { |
8461d226 | 413 | char __user *user_data; |
eb01459f | 414 | ssize_t remain; |
8461d226 | 415 | loff_t offset; |
eb2c0c81 | 416 | int shmem_page_offset, page_length, ret = 0; |
8461d226 | 417 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
96d79b52 | 418 | int prefaulted = 0; |
8489731c | 419 | int needs_clflush = 0; |
67d5a50c | 420 | struct sg_page_iter sg_iter; |
eb01459f | 421 | |
2bb4629a | 422 | user_data = to_user_ptr(args->data_ptr); |
eb01459f EA |
423 | remain = args->size; |
424 | ||
8461d226 | 425 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
eb01459f | 426 | |
8489731c DV |
427 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) { |
428 | /* If we're not in the cpu read domain, set ourself into the gtt | |
429 | * read domain and manually flush cachelines (if required). This | |
430 | * optimizes for the case when the gpu will dirty the data | |
431 | * anyway again before the next pread happens. */ | |
c76ce038 | 432 | needs_clflush = !cpu_cache_is_coherent(dev, obj->cache_level); |
9843877d | 433 | if (i915_gem_obj_bound_any(obj)) { |
6c085a72 CW |
434 | ret = i915_gem_object_set_to_gtt_domain(obj, false); |
435 | if (ret) | |
436 | return ret; | |
437 | } | |
8489731c | 438 | } |
eb01459f | 439 | |
f60d7f0c CW |
440 | ret = i915_gem_object_get_pages(obj); |
441 | if (ret) | |
442 | return ret; | |
443 | ||
444 | i915_gem_object_pin_pages(obj); | |
445 | ||
8461d226 | 446 | offset = args->offset; |
eb01459f | 447 | |
67d5a50c ID |
448 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
449 | offset >> PAGE_SHIFT) { | |
2db76d7c | 450 | struct page *page = sg_page_iter_page(&sg_iter); |
9da3da66 CW |
451 | |
452 | if (remain <= 0) | |
453 | break; | |
454 | ||
eb01459f EA |
455 | /* Operation in this page |
456 | * | |
eb01459f | 457 | * shmem_page_offset = offset within page in shmem file |
eb01459f EA |
458 | * page_length = bytes to copy for this page |
459 | */ | |
c8cbbb8b | 460 | shmem_page_offset = offset_in_page(offset); |
eb01459f EA |
461 | page_length = remain; |
462 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
463 | page_length = PAGE_SIZE - shmem_page_offset; | |
eb01459f | 464 | |
8461d226 DV |
465 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
466 | (page_to_phys(page) & (1 << 17)) != 0; | |
467 | ||
d174bd64 DV |
468 | ret = shmem_pread_fast(page, shmem_page_offset, page_length, |
469 | user_data, page_do_bit17_swizzling, | |
470 | needs_clflush); | |
471 | if (ret == 0) | |
472 | goto next_page; | |
dbf7bff0 | 473 | |
dbf7bff0 DV |
474 | mutex_unlock(&dev->struct_mutex); |
475 | ||
0b74b508 | 476 | if (likely(!i915_prefault_disable) && !prefaulted) { |
f56f821f | 477 | ret = fault_in_multipages_writeable(user_data, remain); |
96d79b52 DV |
478 | /* Userspace is tricking us, but we've already clobbered |
479 | * its pages with the prefault and promised to write the | |
480 | * data up to the first fault. Hence ignore any errors | |
481 | * and just continue. */ | |
482 | (void)ret; | |
483 | prefaulted = 1; | |
484 | } | |
eb01459f | 485 | |
d174bd64 DV |
486 | ret = shmem_pread_slow(page, shmem_page_offset, page_length, |
487 | user_data, page_do_bit17_swizzling, | |
488 | needs_clflush); | |
eb01459f | 489 | |
dbf7bff0 | 490 | mutex_lock(&dev->struct_mutex); |
f60d7f0c | 491 | |
dbf7bff0 | 492 | next_page: |
e5281ccd | 493 | mark_page_accessed(page); |
e5281ccd | 494 | |
f60d7f0c | 495 | if (ret) |
8461d226 | 496 | goto out; |
8461d226 | 497 | |
eb01459f | 498 | remain -= page_length; |
8461d226 | 499 | user_data += page_length; |
eb01459f EA |
500 | offset += page_length; |
501 | } | |
502 | ||
4f27b75d | 503 | out: |
f60d7f0c CW |
504 | i915_gem_object_unpin_pages(obj); |
505 | ||
eb01459f EA |
506 | return ret; |
507 | } | |
508 | ||
673a394b EA |
509 | /** |
510 | * Reads data from the object referenced by handle. | |
511 | * | |
512 | * On error, the contents of *data are undefined. | |
513 | */ | |
514 | int | |
515 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 516 | struct drm_file *file) |
673a394b EA |
517 | { |
518 | struct drm_i915_gem_pread *args = data; | |
05394f39 | 519 | struct drm_i915_gem_object *obj; |
35b62a89 | 520 | int ret = 0; |
673a394b | 521 | |
51311d0a CW |
522 | if (args->size == 0) |
523 | return 0; | |
524 | ||
525 | if (!access_ok(VERIFY_WRITE, | |
2bb4629a | 526 | to_user_ptr(args->data_ptr), |
51311d0a CW |
527 | args->size)) |
528 | return -EFAULT; | |
529 | ||
4f27b75d | 530 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 531 | if (ret) |
4f27b75d | 532 | return ret; |
673a394b | 533 | |
05394f39 | 534 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 535 | if (&obj->base == NULL) { |
1d7cfea1 CW |
536 | ret = -ENOENT; |
537 | goto unlock; | |
4f27b75d | 538 | } |
673a394b | 539 | |
7dcd2499 | 540 | /* Bounds check source. */ |
05394f39 CW |
541 | if (args->offset > obj->base.size || |
542 | args->size > obj->base.size - args->offset) { | |
ce9d419d | 543 | ret = -EINVAL; |
35b62a89 | 544 | goto out; |
ce9d419d CW |
545 | } |
546 | ||
1286ff73 DV |
547 | /* prime objects have no backing filp to GEM pread/pwrite |
548 | * pages from. | |
549 | */ | |
550 | if (!obj->base.filp) { | |
551 | ret = -EINVAL; | |
552 | goto out; | |
553 | } | |
554 | ||
db53a302 CW |
555 | trace_i915_gem_object_pread(obj, args->offset, args->size); |
556 | ||
dbf7bff0 | 557 | ret = i915_gem_shmem_pread(dev, obj, args, file); |
673a394b | 558 | |
35b62a89 | 559 | out: |
05394f39 | 560 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 561 | unlock: |
4f27b75d | 562 | mutex_unlock(&dev->struct_mutex); |
eb01459f | 563 | return ret; |
673a394b EA |
564 | } |
565 | ||
0839ccb8 KP |
566 | /* This is the fast write path which cannot handle |
567 | * page faults in the source data | |
9b7530cc | 568 | */ |
0839ccb8 KP |
569 | |
570 | static inline int | |
571 | fast_user_write(struct io_mapping *mapping, | |
572 | loff_t page_base, int page_offset, | |
573 | char __user *user_data, | |
574 | int length) | |
9b7530cc | 575 | { |
4f0c7cfb BW |
576 | void __iomem *vaddr_atomic; |
577 | void *vaddr; | |
0839ccb8 | 578 | unsigned long unwritten; |
9b7530cc | 579 | |
3e4d3af5 | 580 | vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base); |
4f0c7cfb BW |
581 | /* We can use the cpu mem copy function because this is X86. */ |
582 | vaddr = (void __force*)vaddr_atomic + page_offset; | |
583 | unwritten = __copy_from_user_inatomic_nocache(vaddr, | |
0839ccb8 | 584 | user_data, length); |
3e4d3af5 | 585 | io_mapping_unmap_atomic(vaddr_atomic); |
fbd5a26d | 586 | return unwritten; |
0839ccb8 KP |
587 | } |
588 | ||
3de09aa3 EA |
589 | /** |
590 | * This is the fast pwrite path, where we copy the data directly from the | |
591 | * user into the GTT, uncached. | |
592 | */ | |
673a394b | 593 | static int |
05394f39 CW |
594 | i915_gem_gtt_pwrite_fast(struct drm_device *dev, |
595 | struct drm_i915_gem_object *obj, | |
3de09aa3 | 596 | struct drm_i915_gem_pwrite *args, |
05394f39 | 597 | struct drm_file *file) |
673a394b | 598 | { |
0839ccb8 | 599 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b | 600 | ssize_t remain; |
0839ccb8 | 601 | loff_t offset, page_base; |
673a394b | 602 | char __user *user_data; |
935aaa69 DV |
603 | int page_offset, page_length, ret; |
604 | ||
c37e2204 | 605 | ret = i915_gem_obj_ggtt_pin(obj, 0, true, true); |
935aaa69 DV |
606 | if (ret) |
607 | goto out; | |
608 | ||
609 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
610 | if (ret) | |
611 | goto out_unpin; | |
612 | ||
613 | ret = i915_gem_object_put_fence(obj); | |
614 | if (ret) | |
615 | goto out_unpin; | |
673a394b | 616 | |
2bb4629a | 617 | user_data = to_user_ptr(args->data_ptr); |
673a394b | 618 | remain = args->size; |
673a394b | 619 | |
f343c5f6 | 620 | offset = i915_gem_obj_ggtt_offset(obj) + args->offset; |
673a394b EA |
621 | |
622 | while (remain > 0) { | |
623 | /* Operation in this page | |
624 | * | |
0839ccb8 KP |
625 | * page_base = page offset within aperture |
626 | * page_offset = offset within page | |
627 | * page_length = bytes to copy for this page | |
673a394b | 628 | */ |
c8cbbb8b CW |
629 | page_base = offset & PAGE_MASK; |
630 | page_offset = offset_in_page(offset); | |
0839ccb8 KP |
631 | page_length = remain; |
632 | if ((page_offset + remain) > PAGE_SIZE) | |
633 | page_length = PAGE_SIZE - page_offset; | |
634 | ||
0839ccb8 | 635 | /* If we get a fault while copying data, then (presumably) our |
3de09aa3 EA |
636 | * source page isn't available. Return the error and we'll |
637 | * retry in the slow path. | |
0839ccb8 | 638 | */ |
5d4545ae | 639 | if (fast_user_write(dev_priv->gtt.mappable, page_base, |
935aaa69 DV |
640 | page_offset, user_data, page_length)) { |
641 | ret = -EFAULT; | |
642 | goto out_unpin; | |
643 | } | |
673a394b | 644 | |
0839ccb8 KP |
645 | remain -= page_length; |
646 | user_data += page_length; | |
647 | offset += page_length; | |
673a394b | 648 | } |
673a394b | 649 | |
935aaa69 DV |
650 | out_unpin: |
651 | i915_gem_object_unpin(obj); | |
652 | out: | |
3de09aa3 | 653 | return ret; |
673a394b EA |
654 | } |
655 | ||
d174bd64 DV |
656 | /* Per-page copy function for the shmem pwrite fastpath. |
657 | * Flushes invalid cachelines before writing to the target if | |
658 | * needs_clflush_before is set and flushes out any written cachelines after | |
659 | * writing if needs_clflush is set. */ | |
3043c60c | 660 | static int |
d174bd64 DV |
661 | shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length, |
662 | char __user *user_data, | |
663 | bool page_do_bit17_swizzling, | |
664 | bool needs_clflush_before, | |
665 | bool needs_clflush_after) | |
673a394b | 666 | { |
d174bd64 | 667 | char *vaddr; |
673a394b | 668 | int ret; |
3de09aa3 | 669 | |
e7e58eb5 | 670 | if (unlikely(page_do_bit17_swizzling)) |
d174bd64 | 671 | return -EINVAL; |
3de09aa3 | 672 | |
d174bd64 DV |
673 | vaddr = kmap_atomic(page); |
674 | if (needs_clflush_before) | |
675 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
676 | page_length); | |
677 | ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset, | |
678 | user_data, | |
679 | page_length); | |
680 | if (needs_clflush_after) | |
681 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
682 | page_length); | |
683 | kunmap_atomic(vaddr); | |
3de09aa3 | 684 | |
755d2218 | 685 | return ret ? -EFAULT : 0; |
3de09aa3 EA |
686 | } |
687 | ||
d174bd64 DV |
688 | /* Only difference to the fast-path function is that this can handle bit17 |
689 | * and uses non-atomic copy and kmap functions. */ | |
3043c60c | 690 | static int |
d174bd64 DV |
691 | shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length, |
692 | char __user *user_data, | |
693 | bool page_do_bit17_swizzling, | |
694 | bool needs_clflush_before, | |
695 | bool needs_clflush_after) | |
673a394b | 696 | { |
d174bd64 DV |
697 | char *vaddr; |
698 | int ret; | |
e5281ccd | 699 | |
d174bd64 | 700 | vaddr = kmap(page); |
e7e58eb5 | 701 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) |
23c18c71 DV |
702 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
703 | page_length, | |
704 | page_do_bit17_swizzling); | |
d174bd64 DV |
705 | if (page_do_bit17_swizzling) |
706 | ret = __copy_from_user_swizzled(vaddr, shmem_page_offset, | |
e5281ccd CW |
707 | user_data, |
708 | page_length); | |
d174bd64 DV |
709 | else |
710 | ret = __copy_from_user(vaddr + shmem_page_offset, | |
711 | user_data, | |
712 | page_length); | |
713 | if (needs_clflush_after) | |
23c18c71 DV |
714 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
715 | page_length, | |
716 | page_do_bit17_swizzling); | |
d174bd64 | 717 | kunmap(page); |
40123c1f | 718 | |
755d2218 | 719 | return ret ? -EFAULT : 0; |
40123c1f EA |
720 | } |
721 | ||
40123c1f | 722 | static int |
e244a443 DV |
723 | i915_gem_shmem_pwrite(struct drm_device *dev, |
724 | struct drm_i915_gem_object *obj, | |
725 | struct drm_i915_gem_pwrite *args, | |
726 | struct drm_file *file) | |
40123c1f | 727 | { |
40123c1f | 728 | ssize_t remain; |
8c59967c DV |
729 | loff_t offset; |
730 | char __user *user_data; | |
eb2c0c81 | 731 | int shmem_page_offset, page_length, ret = 0; |
8c59967c | 732 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
e244a443 | 733 | int hit_slowpath = 0; |
58642885 DV |
734 | int needs_clflush_after = 0; |
735 | int needs_clflush_before = 0; | |
67d5a50c | 736 | struct sg_page_iter sg_iter; |
40123c1f | 737 | |
2bb4629a | 738 | user_data = to_user_ptr(args->data_ptr); |
40123c1f EA |
739 | remain = args->size; |
740 | ||
8c59967c | 741 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
40123c1f | 742 | |
58642885 DV |
743 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
744 | /* If we're not in the cpu write domain, set ourself into the gtt | |
745 | * write domain and manually flush cachelines (if required). This | |
746 | * optimizes for the case when the gpu will use the data | |
747 | * right away and we therefore have to clflush anyway. */ | |
2c22569b | 748 | needs_clflush_after = cpu_write_needs_clflush(obj); |
9843877d | 749 | if (i915_gem_obj_bound_any(obj)) { |
6c085a72 CW |
750 | ret = i915_gem_object_set_to_gtt_domain(obj, true); |
751 | if (ret) | |
752 | return ret; | |
753 | } | |
58642885 | 754 | } |
c76ce038 CW |
755 | /* Same trick applies to invalidate partially written cachelines read |
756 | * before writing. */ | |
757 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) | |
758 | needs_clflush_before = | |
759 | !cpu_cache_is_coherent(dev, obj->cache_level); | |
58642885 | 760 | |
755d2218 CW |
761 | ret = i915_gem_object_get_pages(obj); |
762 | if (ret) | |
763 | return ret; | |
764 | ||
765 | i915_gem_object_pin_pages(obj); | |
766 | ||
673a394b | 767 | offset = args->offset; |
05394f39 | 768 | obj->dirty = 1; |
673a394b | 769 | |
67d5a50c ID |
770 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, |
771 | offset >> PAGE_SHIFT) { | |
2db76d7c | 772 | struct page *page = sg_page_iter_page(&sg_iter); |
58642885 | 773 | int partial_cacheline_write; |
e5281ccd | 774 | |
9da3da66 CW |
775 | if (remain <= 0) |
776 | break; | |
777 | ||
40123c1f EA |
778 | /* Operation in this page |
779 | * | |
40123c1f | 780 | * shmem_page_offset = offset within page in shmem file |
40123c1f EA |
781 | * page_length = bytes to copy for this page |
782 | */ | |
c8cbbb8b | 783 | shmem_page_offset = offset_in_page(offset); |
40123c1f EA |
784 | |
785 | page_length = remain; | |
786 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
787 | page_length = PAGE_SIZE - shmem_page_offset; | |
40123c1f | 788 | |
58642885 DV |
789 | /* If we don't overwrite a cacheline completely we need to be |
790 | * careful to have up-to-date data by first clflushing. Don't | |
791 | * overcomplicate things and flush the entire patch. */ | |
792 | partial_cacheline_write = needs_clflush_before && | |
793 | ((shmem_page_offset | page_length) | |
794 | & (boot_cpu_data.x86_clflush_size - 1)); | |
795 | ||
8c59967c DV |
796 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
797 | (page_to_phys(page) & (1 << 17)) != 0; | |
798 | ||
d174bd64 DV |
799 | ret = shmem_pwrite_fast(page, shmem_page_offset, page_length, |
800 | user_data, page_do_bit17_swizzling, | |
801 | partial_cacheline_write, | |
802 | needs_clflush_after); | |
803 | if (ret == 0) | |
804 | goto next_page; | |
e244a443 DV |
805 | |
806 | hit_slowpath = 1; | |
e244a443 | 807 | mutex_unlock(&dev->struct_mutex); |
d174bd64 DV |
808 | ret = shmem_pwrite_slow(page, shmem_page_offset, page_length, |
809 | user_data, page_do_bit17_swizzling, | |
810 | partial_cacheline_write, | |
811 | needs_clflush_after); | |
40123c1f | 812 | |
e244a443 | 813 | mutex_lock(&dev->struct_mutex); |
755d2218 | 814 | |
e244a443 | 815 | next_page: |
e5281ccd CW |
816 | set_page_dirty(page); |
817 | mark_page_accessed(page); | |
e5281ccd | 818 | |
755d2218 | 819 | if (ret) |
8c59967c | 820 | goto out; |
8c59967c | 821 | |
40123c1f | 822 | remain -= page_length; |
8c59967c | 823 | user_data += page_length; |
40123c1f | 824 | offset += page_length; |
673a394b EA |
825 | } |
826 | ||
fbd5a26d | 827 | out: |
755d2218 CW |
828 | i915_gem_object_unpin_pages(obj); |
829 | ||
e244a443 | 830 | if (hit_slowpath) { |
8dcf015e DV |
831 | /* |
832 | * Fixup: Flush cpu caches in case we didn't flush the dirty | |
833 | * cachelines in-line while writing and the object moved | |
834 | * out of the cpu write domain while we've dropped the lock. | |
835 | */ | |
836 | if (!needs_clflush_after && | |
837 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { | |
000433b6 CW |
838 | if (i915_gem_clflush_object(obj, obj->pin_display)) |
839 | i915_gem_chipset_flush(dev); | |
e244a443 | 840 | } |
8c59967c | 841 | } |
673a394b | 842 | |
58642885 | 843 | if (needs_clflush_after) |
e76e9aeb | 844 | i915_gem_chipset_flush(dev); |
58642885 | 845 | |
40123c1f | 846 | return ret; |
673a394b EA |
847 | } |
848 | ||
849 | /** | |
850 | * Writes data to the object referenced by handle. | |
851 | * | |
852 | * On error, the contents of the buffer that were to be modified are undefined. | |
853 | */ | |
854 | int | |
855 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, | |
fbd5a26d | 856 | struct drm_file *file) |
673a394b EA |
857 | { |
858 | struct drm_i915_gem_pwrite *args = data; | |
05394f39 | 859 | struct drm_i915_gem_object *obj; |
51311d0a CW |
860 | int ret; |
861 | ||
862 | if (args->size == 0) | |
863 | return 0; | |
864 | ||
865 | if (!access_ok(VERIFY_READ, | |
2bb4629a | 866 | to_user_ptr(args->data_ptr), |
51311d0a CW |
867 | args->size)) |
868 | return -EFAULT; | |
869 | ||
0b74b508 XZ |
870 | if (likely(!i915_prefault_disable)) { |
871 | ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr), | |
872 | args->size); | |
873 | if (ret) | |
874 | return -EFAULT; | |
875 | } | |
673a394b | 876 | |
fbd5a26d | 877 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 878 | if (ret) |
fbd5a26d | 879 | return ret; |
1d7cfea1 | 880 | |
05394f39 | 881 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 882 | if (&obj->base == NULL) { |
1d7cfea1 CW |
883 | ret = -ENOENT; |
884 | goto unlock; | |
fbd5a26d | 885 | } |
673a394b | 886 | |
7dcd2499 | 887 | /* Bounds check destination. */ |
05394f39 CW |
888 | if (args->offset > obj->base.size || |
889 | args->size > obj->base.size - args->offset) { | |
ce9d419d | 890 | ret = -EINVAL; |
35b62a89 | 891 | goto out; |
ce9d419d CW |
892 | } |
893 | ||
1286ff73 DV |
894 | /* prime objects have no backing filp to GEM pread/pwrite |
895 | * pages from. | |
896 | */ | |
897 | if (!obj->base.filp) { | |
898 | ret = -EINVAL; | |
899 | goto out; | |
900 | } | |
901 | ||
db53a302 CW |
902 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); |
903 | ||
935aaa69 | 904 | ret = -EFAULT; |
673a394b EA |
905 | /* We can only do the GTT pwrite on untiled buffers, as otherwise |
906 | * it would end up going through the fenced access, and we'll get | |
907 | * different detiling behavior between reading and writing. | |
908 | * pread/pwrite currently are reading and writing from the CPU | |
909 | * perspective, requiring manual detiling by the client. | |
910 | */ | |
5c0480f2 | 911 | if (obj->phys_obj) { |
fbd5a26d | 912 | ret = i915_gem_phys_pwrite(dev, obj, args, file); |
5c0480f2 DV |
913 | goto out; |
914 | } | |
915 | ||
2c22569b CW |
916 | if (obj->tiling_mode == I915_TILING_NONE && |
917 | obj->base.write_domain != I915_GEM_DOMAIN_CPU && | |
918 | cpu_write_needs_clflush(obj)) { | |
fbd5a26d | 919 | ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); |
935aaa69 DV |
920 | /* Note that the gtt paths might fail with non-page-backed user |
921 | * pointers (e.g. gtt mappings when moving data between | |
922 | * textures). Fallback to the shmem path in that case. */ | |
fbd5a26d | 923 | } |
673a394b | 924 | |
86a1ee26 | 925 | if (ret == -EFAULT || ret == -ENOSPC) |
935aaa69 | 926 | ret = i915_gem_shmem_pwrite(dev, obj, args, file); |
5c0480f2 | 927 | |
35b62a89 | 928 | out: |
05394f39 | 929 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 930 | unlock: |
fbd5a26d | 931 | mutex_unlock(&dev->struct_mutex); |
673a394b EA |
932 | return ret; |
933 | } | |
934 | ||
b361237b | 935 | int |
33196ded | 936 | i915_gem_check_wedge(struct i915_gpu_error *error, |
b361237b CW |
937 | bool interruptible) |
938 | { | |
1f83fee0 | 939 | if (i915_reset_in_progress(error)) { |
b361237b CW |
940 | /* Non-interruptible callers can't handle -EAGAIN, hence return |
941 | * -EIO unconditionally for these. */ | |
942 | if (!interruptible) | |
943 | return -EIO; | |
944 | ||
1f83fee0 DV |
945 | /* Recovery complete, but the reset failed ... */ |
946 | if (i915_terminally_wedged(error)) | |
b361237b CW |
947 | return -EIO; |
948 | ||
949 | return -EAGAIN; | |
950 | } | |
951 | ||
952 | return 0; | |
953 | } | |
954 | ||
955 | /* | |
956 | * Compare seqno against outstanding lazy request. Emit a request if they are | |
957 | * equal. | |
958 | */ | |
959 | static int | |
960 | i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno) | |
961 | { | |
962 | int ret; | |
963 | ||
964 | BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex)); | |
965 | ||
966 | ret = 0; | |
1823521d | 967 | if (seqno == ring->outstanding_lazy_seqno) |
0025c077 | 968 | ret = i915_add_request(ring, NULL); |
b361237b CW |
969 | |
970 | return ret; | |
971 | } | |
972 | ||
973 | /** | |
974 | * __wait_seqno - wait until execution of seqno has finished | |
975 | * @ring: the ring expected to report seqno | |
976 | * @seqno: duh! | |
f69061be | 977 | * @reset_counter: reset sequence associated with the given seqno |
b361237b CW |
978 | * @interruptible: do an interruptible wait (normally yes) |
979 | * @timeout: in - how long to wait (NULL forever); out - how much time remaining | |
980 | * | |
f69061be DV |
981 | * Note: It is of utmost importance that the passed in seqno and reset_counter |
982 | * values have been read by the caller in an smp safe manner. Where read-side | |
983 | * locks are involved, it is sufficient to read the reset_counter before | |
984 | * unlocking the lock that protects the seqno. For lockless tricks, the | |
985 | * reset_counter _must_ be read before, and an appropriate smp_rmb must be | |
986 | * inserted. | |
987 | * | |
b361237b CW |
988 | * Returns 0 if the seqno was found within the alloted time. Else returns the |
989 | * errno with remaining time filled in timeout argument. | |
990 | */ | |
991 | static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno, | |
f69061be | 992 | unsigned reset_counter, |
b361237b CW |
993 | bool interruptible, struct timespec *timeout) |
994 | { | |
995 | drm_i915_private_t *dev_priv = ring->dev->dev_private; | |
996 | struct timespec before, now, wait_time={1,0}; | |
997 | unsigned long timeout_jiffies; | |
998 | long end; | |
999 | bool wait_forever = true; | |
1000 | int ret; | |
1001 | ||
c67a470b PZ |
1002 | WARN(dev_priv->pc8.irqs_disabled, "IRQs disabled\n"); |
1003 | ||
b361237b CW |
1004 | if (i915_seqno_passed(ring->get_seqno(ring, true), seqno)) |
1005 | return 0; | |
1006 | ||
1007 | trace_i915_gem_request_wait_begin(ring, seqno); | |
1008 | ||
1009 | if (timeout != NULL) { | |
1010 | wait_time = *timeout; | |
1011 | wait_forever = false; | |
1012 | } | |
1013 | ||
e054cc39 | 1014 | timeout_jiffies = timespec_to_jiffies_timeout(&wait_time); |
b361237b CW |
1015 | |
1016 | if (WARN_ON(!ring->irq_get(ring))) | |
1017 | return -ENODEV; | |
1018 | ||
1019 | /* Record current time in case interrupted by signal, or wedged * */ | |
1020 | getrawmonotonic(&before); | |
1021 | ||
1022 | #define EXIT_COND \ | |
1023 | (i915_seqno_passed(ring->get_seqno(ring, false), seqno) || \ | |
f69061be DV |
1024 | i915_reset_in_progress(&dev_priv->gpu_error) || \ |
1025 | reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) | |
b361237b CW |
1026 | do { |
1027 | if (interruptible) | |
1028 | end = wait_event_interruptible_timeout(ring->irq_queue, | |
1029 | EXIT_COND, | |
1030 | timeout_jiffies); | |
1031 | else | |
1032 | end = wait_event_timeout(ring->irq_queue, EXIT_COND, | |
1033 | timeout_jiffies); | |
1034 | ||
f69061be DV |
1035 | /* We need to check whether any gpu reset happened in between |
1036 | * the caller grabbing the seqno and now ... */ | |
1037 | if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) | |
1038 | end = -EAGAIN; | |
1039 | ||
1040 | /* ... but upgrade the -EGAIN to an -EIO if the gpu is truely | |
1041 | * gone. */ | |
33196ded | 1042 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
b361237b CW |
1043 | if (ret) |
1044 | end = ret; | |
1045 | } while (end == 0 && wait_forever); | |
1046 | ||
1047 | getrawmonotonic(&now); | |
1048 | ||
1049 | ring->irq_put(ring); | |
1050 | trace_i915_gem_request_wait_end(ring, seqno); | |
1051 | #undef EXIT_COND | |
1052 | ||
1053 | if (timeout) { | |
1054 | struct timespec sleep_time = timespec_sub(now, before); | |
1055 | *timeout = timespec_sub(*timeout, sleep_time); | |
4f42f4ef CW |
1056 | if (!timespec_valid(timeout)) /* i.e. negative time remains */ |
1057 | set_normalized_timespec(timeout, 0, 0); | |
b361237b CW |
1058 | } |
1059 | ||
1060 | switch (end) { | |
1061 | case -EIO: | |
1062 | case -EAGAIN: /* Wedged */ | |
1063 | case -ERESTARTSYS: /* Signal */ | |
1064 | return (int)end; | |
1065 | case 0: /* Timeout */ | |
b361237b CW |
1066 | return -ETIME; |
1067 | default: /* Completed */ | |
1068 | WARN_ON(end < 0); /* We're not aware of other errors */ | |
1069 | return 0; | |
1070 | } | |
1071 | } | |
1072 | ||
1073 | /** | |
1074 | * Waits for a sequence number to be signaled, and cleans up the | |
1075 | * request and object lists appropriately for that event. | |
1076 | */ | |
1077 | int | |
1078 | i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno) | |
1079 | { | |
1080 | struct drm_device *dev = ring->dev; | |
1081 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1082 | bool interruptible = dev_priv->mm.interruptible; | |
1083 | int ret; | |
1084 | ||
1085 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); | |
1086 | BUG_ON(seqno == 0); | |
1087 | ||
33196ded | 1088 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); |
b361237b CW |
1089 | if (ret) |
1090 | return ret; | |
1091 | ||
1092 | ret = i915_gem_check_olr(ring, seqno); | |
1093 | if (ret) | |
1094 | return ret; | |
1095 | ||
f69061be DV |
1096 | return __wait_seqno(ring, seqno, |
1097 | atomic_read(&dev_priv->gpu_error.reset_counter), | |
1098 | interruptible, NULL); | |
b361237b CW |
1099 | } |
1100 | ||
d26e3af8 CW |
1101 | static int |
1102 | i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj, | |
1103 | struct intel_ring_buffer *ring) | |
1104 | { | |
1105 | i915_gem_retire_requests_ring(ring); | |
1106 | ||
1107 | /* Manually manage the write flush as we may have not yet | |
1108 | * retired the buffer. | |
1109 | * | |
1110 | * Note that the last_write_seqno is always the earlier of | |
1111 | * the two (read/write) seqno, so if we haved successfully waited, | |
1112 | * we know we have passed the last write. | |
1113 | */ | |
1114 | obj->last_write_seqno = 0; | |
1115 | obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS; | |
1116 | ||
1117 | return 0; | |
1118 | } | |
1119 | ||
b361237b CW |
1120 | /** |
1121 | * Ensures that all rendering to the object has completed and the object is | |
1122 | * safe to unbind from the GTT or access from the CPU. | |
1123 | */ | |
1124 | static __must_check int | |
1125 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj, | |
1126 | bool readonly) | |
1127 | { | |
1128 | struct intel_ring_buffer *ring = obj->ring; | |
1129 | u32 seqno; | |
1130 | int ret; | |
1131 | ||
1132 | seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno; | |
1133 | if (seqno == 0) | |
1134 | return 0; | |
1135 | ||
1136 | ret = i915_wait_seqno(ring, seqno); | |
1137 | if (ret) | |
1138 | return ret; | |
1139 | ||
d26e3af8 | 1140 | return i915_gem_object_wait_rendering__tail(obj, ring); |
b361237b CW |
1141 | } |
1142 | ||
3236f57a CW |
1143 | /* A nonblocking variant of the above wait. This is a highly dangerous routine |
1144 | * as the object state may change during this call. | |
1145 | */ | |
1146 | static __must_check int | |
1147 | i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj, | |
1148 | bool readonly) | |
1149 | { | |
1150 | struct drm_device *dev = obj->base.dev; | |
1151 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1152 | struct intel_ring_buffer *ring = obj->ring; | |
f69061be | 1153 | unsigned reset_counter; |
3236f57a CW |
1154 | u32 seqno; |
1155 | int ret; | |
1156 | ||
1157 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); | |
1158 | BUG_ON(!dev_priv->mm.interruptible); | |
1159 | ||
1160 | seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno; | |
1161 | if (seqno == 0) | |
1162 | return 0; | |
1163 | ||
33196ded | 1164 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, true); |
3236f57a CW |
1165 | if (ret) |
1166 | return ret; | |
1167 | ||
1168 | ret = i915_gem_check_olr(ring, seqno); | |
1169 | if (ret) | |
1170 | return ret; | |
1171 | ||
f69061be | 1172 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
3236f57a | 1173 | mutex_unlock(&dev->struct_mutex); |
f69061be | 1174 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL); |
3236f57a | 1175 | mutex_lock(&dev->struct_mutex); |
d26e3af8 CW |
1176 | if (ret) |
1177 | return ret; | |
3236f57a | 1178 | |
d26e3af8 | 1179 | return i915_gem_object_wait_rendering__tail(obj, ring); |
3236f57a CW |
1180 | } |
1181 | ||
673a394b | 1182 | /** |
2ef7eeaa EA |
1183 | * Called when user space prepares to use an object with the CPU, either |
1184 | * through the mmap ioctl's mapping or a GTT mapping. | |
673a394b EA |
1185 | */ |
1186 | int | |
1187 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1188 | struct drm_file *file) |
673a394b EA |
1189 | { |
1190 | struct drm_i915_gem_set_domain *args = data; | |
05394f39 | 1191 | struct drm_i915_gem_object *obj; |
2ef7eeaa EA |
1192 | uint32_t read_domains = args->read_domains; |
1193 | uint32_t write_domain = args->write_domain; | |
673a394b EA |
1194 | int ret; |
1195 | ||
2ef7eeaa | 1196 | /* Only handle setting domains to types used by the CPU. */ |
21d509e3 | 1197 | if (write_domain & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
1198 | return -EINVAL; |
1199 | ||
21d509e3 | 1200 | if (read_domains & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
1201 | return -EINVAL; |
1202 | ||
1203 | /* Having something in the write domain implies it's in the read | |
1204 | * domain, and only that read domain. Enforce that in the request. | |
1205 | */ | |
1206 | if (write_domain != 0 && read_domains != write_domain) | |
1207 | return -EINVAL; | |
1208 | ||
76c1dec1 | 1209 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1210 | if (ret) |
76c1dec1 | 1211 | return ret; |
1d7cfea1 | 1212 | |
05394f39 | 1213 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 1214 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1215 | ret = -ENOENT; |
1216 | goto unlock; | |
76c1dec1 | 1217 | } |
673a394b | 1218 | |
3236f57a CW |
1219 | /* Try to flush the object off the GPU without holding the lock. |
1220 | * We will repeat the flush holding the lock in the normal manner | |
1221 | * to catch cases where we are gazumped. | |
1222 | */ | |
1223 | ret = i915_gem_object_wait_rendering__nonblocking(obj, !write_domain); | |
1224 | if (ret) | |
1225 | goto unref; | |
1226 | ||
2ef7eeaa EA |
1227 | if (read_domains & I915_GEM_DOMAIN_GTT) { |
1228 | ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); | |
02354392 EA |
1229 | |
1230 | /* Silently promote "you're not bound, there was nothing to do" | |
1231 | * to success, since the client was just asking us to | |
1232 | * make sure everything was done. | |
1233 | */ | |
1234 | if (ret == -EINVAL) | |
1235 | ret = 0; | |
2ef7eeaa | 1236 | } else { |
e47c68e9 | 1237 | ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); |
2ef7eeaa EA |
1238 | } |
1239 | ||
3236f57a | 1240 | unref: |
05394f39 | 1241 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1242 | unlock: |
673a394b EA |
1243 | mutex_unlock(&dev->struct_mutex); |
1244 | return ret; | |
1245 | } | |
1246 | ||
1247 | /** | |
1248 | * Called when user space has done writes to this buffer | |
1249 | */ | |
1250 | int | |
1251 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1252 | struct drm_file *file) |
673a394b EA |
1253 | { |
1254 | struct drm_i915_gem_sw_finish *args = data; | |
05394f39 | 1255 | struct drm_i915_gem_object *obj; |
673a394b EA |
1256 | int ret = 0; |
1257 | ||
76c1dec1 | 1258 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1259 | if (ret) |
76c1dec1 | 1260 | return ret; |
1d7cfea1 | 1261 | |
05394f39 | 1262 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 1263 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1264 | ret = -ENOENT; |
1265 | goto unlock; | |
673a394b EA |
1266 | } |
1267 | ||
673a394b | 1268 | /* Pinned buffers may be scanout, so flush the cache */ |
2c22569b CW |
1269 | if (obj->pin_display) |
1270 | i915_gem_object_flush_cpu_write_domain(obj, true); | |
e47c68e9 | 1271 | |
05394f39 | 1272 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1273 | unlock: |
673a394b EA |
1274 | mutex_unlock(&dev->struct_mutex); |
1275 | return ret; | |
1276 | } | |
1277 | ||
1278 | /** | |
1279 | * Maps the contents of an object, returning the address it is mapped | |
1280 | * into. | |
1281 | * | |
1282 | * While the mapping holds a reference on the contents of the object, it doesn't | |
1283 | * imply a ref on the object itself. | |
1284 | */ | |
1285 | int | |
1286 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1287 | struct drm_file *file) |
673a394b EA |
1288 | { |
1289 | struct drm_i915_gem_mmap *args = data; | |
1290 | struct drm_gem_object *obj; | |
673a394b EA |
1291 | unsigned long addr; |
1292 | ||
05394f39 | 1293 | obj = drm_gem_object_lookup(dev, file, args->handle); |
673a394b | 1294 | if (obj == NULL) |
bf79cb91 | 1295 | return -ENOENT; |
673a394b | 1296 | |
1286ff73 DV |
1297 | /* prime objects have no backing filp to GEM mmap |
1298 | * pages from. | |
1299 | */ | |
1300 | if (!obj->filp) { | |
1301 | drm_gem_object_unreference_unlocked(obj); | |
1302 | return -EINVAL; | |
1303 | } | |
1304 | ||
6be5ceb0 | 1305 | addr = vm_mmap(obj->filp, 0, args->size, |
673a394b EA |
1306 | PROT_READ | PROT_WRITE, MAP_SHARED, |
1307 | args->offset); | |
bc9025bd | 1308 | drm_gem_object_unreference_unlocked(obj); |
673a394b EA |
1309 | if (IS_ERR((void *)addr)) |
1310 | return addr; | |
1311 | ||
1312 | args->addr_ptr = (uint64_t) addr; | |
1313 | ||
1314 | return 0; | |
1315 | } | |
1316 | ||
de151cf6 JB |
1317 | /** |
1318 | * i915_gem_fault - fault a page into the GTT | |
1319 | * vma: VMA in question | |
1320 | * vmf: fault info | |
1321 | * | |
1322 | * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped | |
1323 | * from userspace. The fault handler takes care of binding the object to | |
1324 | * the GTT (if needed), allocating and programming a fence register (again, | |
1325 | * only if needed based on whether the old reg is still valid or the object | |
1326 | * is tiled) and inserting a new PTE into the faulting process. | |
1327 | * | |
1328 | * Note that the faulting process may involve evicting existing objects | |
1329 | * from the GTT and/or fence registers to make room. So performance may | |
1330 | * suffer if the GTT working set is large or there are few fence registers | |
1331 | * left. | |
1332 | */ | |
1333 | int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
1334 | { | |
05394f39 CW |
1335 | struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data); |
1336 | struct drm_device *dev = obj->base.dev; | |
7d1c4804 | 1337 | drm_i915_private_t *dev_priv = dev->dev_private; |
de151cf6 JB |
1338 | pgoff_t page_offset; |
1339 | unsigned long pfn; | |
1340 | int ret = 0; | |
0f973f27 | 1341 | bool write = !!(vmf->flags & FAULT_FLAG_WRITE); |
de151cf6 JB |
1342 | |
1343 | /* We don't use vmf->pgoff since that has the fake offset */ | |
1344 | page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >> | |
1345 | PAGE_SHIFT; | |
1346 | ||
d9bc7e9f CW |
1347 | ret = i915_mutex_lock_interruptible(dev); |
1348 | if (ret) | |
1349 | goto out; | |
a00b10c3 | 1350 | |
db53a302 CW |
1351 | trace_i915_gem_object_fault(obj, page_offset, true, write); |
1352 | ||
eb119bd6 CW |
1353 | /* Access to snoopable pages through the GTT is incoherent. */ |
1354 | if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) { | |
1355 | ret = -EINVAL; | |
1356 | goto unlock; | |
1357 | } | |
1358 | ||
d9bc7e9f | 1359 | /* Now bind it into the GTT if needed */ |
c37e2204 | 1360 | ret = i915_gem_obj_ggtt_pin(obj, 0, true, false); |
c9839303 CW |
1361 | if (ret) |
1362 | goto unlock; | |
4a684a41 | 1363 | |
c9839303 CW |
1364 | ret = i915_gem_object_set_to_gtt_domain(obj, write); |
1365 | if (ret) | |
1366 | goto unpin; | |
74898d7e | 1367 | |
06d98131 | 1368 | ret = i915_gem_object_get_fence(obj); |
d9e86c0e | 1369 | if (ret) |
c9839303 | 1370 | goto unpin; |
7d1c4804 | 1371 | |
6299f992 CW |
1372 | obj->fault_mappable = true; |
1373 | ||
f343c5f6 BW |
1374 | pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj); |
1375 | pfn >>= PAGE_SHIFT; | |
1376 | pfn += page_offset; | |
de151cf6 JB |
1377 | |
1378 | /* Finally, remap it using the new GTT offset */ | |
1379 | ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn); | |
c9839303 CW |
1380 | unpin: |
1381 | i915_gem_object_unpin(obj); | |
c715089f | 1382 | unlock: |
de151cf6 | 1383 | mutex_unlock(&dev->struct_mutex); |
d9bc7e9f | 1384 | out: |
de151cf6 | 1385 | switch (ret) { |
d9bc7e9f | 1386 | case -EIO: |
a9340cca DV |
1387 | /* If this -EIO is due to a gpu hang, give the reset code a |
1388 | * chance to clean up the mess. Otherwise return the proper | |
1389 | * SIGBUS. */ | |
1f83fee0 | 1390 | if (i915_terminally_wedged(&dev_priv->gpu_error)) |
a9340cca | 1391 | return VM_FAULT_SIGBUS; |
045e769a | 1392 | case -EAGAIN: |
d9bc7e9f CW |
1393 | /* Give the error handler a chance to run and move the |
1394 | * objects off the GPU active list. Next time we service the | |
1395 | * fault, we should be able to transition the page into the | |
1396 | * GTT without touching the GPU (and so avoid further | |
1397 | * EIO/EGAIN). If the GPU is wedged, then there is no issue | |
1398 | * with coherency, just lost writes. | |
1399 | */ | |
045e769a | 1400 | set_need_resched(); |
c715089f CW |
1401 | case 0: |
1402 | case -ERESTARTSYS: | |
bed636ab | 1403 | case -EINTR: |
e79e0fe3 DR |
1404 | case -EBUSY: |
1405 | /* | |
1406 | * EBUSY is ok: this just means that another thread | |
1407 | * already did the job. | |
1408 | */ | |
c715089f | 1409 | return VM_FAULT_NOPAGE; |
de151cf6 | 1410 | case -ENOMEM: |
de151cf6 | 1411 | return VM_FAULT_OOM; |
a7c2e1aa DV |
1412 | case -ENOSPC: |
1413 | return VM_FAULT_SIGBUS; | |
de151cf6 | 1414 | default: |
a7c2e1aa | 1415 | WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret); |
c715089f | 1416 | return VM_FAULT_SIGBUS; |
de151cf6 JB |
1417 | } |
1418 | } | |
1419 | ||
901782b2 CW |
1420 | /** |
1421 | * i915_gem_release_mmap - remove physical page mappings | |
1422 | * @obj: obj in question | |
1423 | * | |
af901ca1 | 1424 | * Preserve the reservation of the mmapping with the DRM core code, but |
901782b2 CW |
1425 | * relinquish ownership of the pages back to the system. |
1426 | * | |
1427 | * It is vital that we remove the page mapping if we have mapped a tiled | |
1428 | * object through the GTT and then lose the fence register due to | |
1429 | * resource pressure. Similarly if the object has been moved out of the | |
1430 | * aperture, than pages mapped into userspace must be revoked. Removing the | |
1431 | * mapping will then trigger a page fault on the next user access, allowing | |
1432 | * fixup by i915_gem_fault(). | |
1433 | */ | |
d05ca301 | 1434 | void |
05394f39 | 1435 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) |
901782b2 | 1436 | { |
6299f992 CW |
1437 | if (!obj->fault_mappable) |
1438 | return; | |
901782b2 | 1439 | |
51335df9 | 1440 | drm_vma_node_unmap(&obj->base.vma_node, obj->base.dev->dev_mapping); |
6299f992 | 1441 | obj->fault_mappable = false; |
901782b2 CW |
1442 | } |
1443 | ||
0fa87796 | 1444 | uint32_t |
e28f8711 | 1445 | i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode) |
92b88aeb | 1446 | { |
e28f8711 | 1447 | uint32_t gtt_size; |
92b88aeb CW |
1448 | |
1449 | if (INTEL_INFO(dev)->gen >= 4 || | |
e28f8711 CW |
1450 | tiling_mode == I915_TILING_NONE) |
1451 | return size; | |
92b88aeb CW |
1452 | |
1453 | /* Previous chips need a power-of-two fence region when tiling */ | |
1454 | if (INTEL_INFO(dev)->gen == 3) | |
e28f8711 | 1455 | gtt_size = 1024*1024; |
92b88aeb | 1456 | else |
e28f8711 | 1457 | gtt_size = 512*1024; |
92b88aeb | 1458 | |
e28f8711 CW |
1459 | while (gtt_size < size) |
1460 | gtt_size <<= 1; | |
92b88aeb | 1461 | |
e28f8711 | 1462 | return gtt_size; |
92b88aeb CW |
1463 | } |
1464 | ||
de151cf6 JB |
1465 | /** |
1466 | * i915_gem_get_gtt_alignment - return required GTT alignment for an object | |
1467 | * @obj: object to check | |
1468 | * | |
1469 | * Return the required GTT alignment for an object, taking into account | |
5e783301 | 1470 | * potential fence register mapping. |
de151cf6 | 1471 | */ |
d865110c ID |
1472 | uint32_t |
1473 | i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size, | |
1474 | int tiling_mode, bool fenced) | |
de151cf6 | 1475 | { |
de151cf6 JB |
1476 | /* |
1477 | * Minimum alignment is 4k (GTT page size), but might be greater | |
1478 | * if a fence register is needed for the object. | |
1479 | */ | |
d865110c | 1480 | if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) || |
e28f8711 | 1481 | tiling_mode == I915_TILING_NONE) |
de151cf6 JB |
1482 | return 4096; |
1483 | ||
a00b10c3 CW |
1484 | /* |
1485 | * Previous chips need to be aligned to the size of the smallest | |
1486 | * fence register that can contain the object. | |
1487 | */ | |
e28f8711 | 1488 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
a00b10c3 CW |
1489 | } |
1490 | ||
d8cb5086 CW |
1491 | static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj) |
1492 | { | |
1493 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
1494 | int ret; | |
1495 | ||
0de23977 | 1496 | if (drm_vma_node_has_offset(&obj->base.vma_node)) |
d8cb5086 CW |
1497 | return 0; |
1498 | ||
da494d7c DV |
1499 | dev_priv->mm.shrinker_no_lock_stealing = true; |
1500 | ||
d8cb5086 CW |
1501 | ret = drm_gem_create_mmap_offset(&obj->base); |
1502 | if (ret != -ENOSPC) | |
da494d7c | 1503 | goto out; |
d8cb5086 CW |
1504 | |
1505 | /* Badly fragmented mmap space? The only way we can recover | |
1506 | * space is by destroying unwanted objects. We can't randomly release | |
1507 | * mmap_offsets as userspace expects them to be persistent for the | |
1508 | * lifetime of the objects. The closest we can is to release the | |
1509 | * offsets on purgeable objects by truncating it and marking it purged, | |
1510 | * which prevents userspace from ever using that object again. | |
1511 | */ | |
1512 | i915_gem_purge(dev_priv, obj->base.size >> PAGE_SHIFT); | |
1513 | ret = drm_gem_create_mmap_offset(&obj->base); | |
1514 | if (ret != -ENOSPC) | |
da494d7c | 1515 | goto out; |
d8cb5086 CW |
1516 | |
1517 | i915_gem_shrink_all(dev_priv); | |
da494d7c DV |
1518 | ret = drm_gem_create_mmap_offset(&obj->base); |
1519 | out: | |
1520 | dev_priv->mm.shrinker_no_lock_stealing = false; | |
1521 | ||
1522 | return ret; | |
d8cb5086 CW |
1523 | } |
1524 | ||
1525 | static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj) | |
1526 | { | |
d8cb5086 CW |
1527 | drm_gem_free_mmap_offset(&obj->base); |
1528 | } | |
1529 | ||
de151cf6 | 1530 | int |
ff72145b DA |
1531 | i915_gem_mmap_gtt(struct drm_file *file, |
1532 | struct drm_device *dev, | |
1533 | uint32_t handle, | |
1534 | uint64_t *offset) | |
de151cf6 | 1535 | { |
da761a6e | 1536 | struct drm_i915_private *dev_priv = dev->dev_private; |
05394f39 | 1537 | struct drm_i915_gem_object *obj; |
de151cf6 JB |
1538 | int ret; |
1539 | ||
76c1dec1 | 1540 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1541 | if (ret) |
76c1dec1 | 1542 | return ret; |
de151cf6 | 1543 | |
ff72145b | 1544 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle)); |
c8725226 | 1545 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1546 | ret = -ENOENT; |
1547 | goto unlock; | |
1548 | } | |
de151cf6 | 1549 | |
5d4545ae | 1550 | if (obj->base.size > dev_priv->gtt.mappable_end) { |
da761a6e | 1551 | ret = -E2BIG; |
ff56b0bc | 1552 | goto out; |
da761a6e CW |
1553 | } |
1554 | ||
05394f39 | 1555 | if (obj->madv != I915_MADV_WILLNEED) { |
ab18282d | 1556 | DRM_ERROR("Attempting to mmap a purgeable buffer\n"); |
1d7cfea1 CW |
1557 | ret = -EINVAL; |
1558 | goto out; | |
ab18282d CW |
1559 | } |
1560 | ||
d8cb5086 CW |
1561 | ret = i915_gem_object_create_mmap_offset(obj); |
1562 | if (ret) | |
1563 | goto out; | |
de151cf6 | 1564 | |
0de23977 | 1565 | *offset = drm_vma_node_offset_addr(&obj->base.vma_node); |
de151cf6 | 1566 | |
1d7cfea1 | 1567 | out: |
05394f39 | 1568 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1569 | unlock: |
de151cf6 | 1570 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 1571 | return ret; |
de151cf6 JB |
1572 | } |
1573 | ||
ff72145b DA |
1574 | /** |
1575 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing | |
1576 | * @dev: DRM device | |
1577 | * @data: GTT mapping ioctl data | |
1578 | * @file: GEM object info | |
1579 | * | |
1580 | * Simply returns the fake offset to userspace so it can mmap it. | |
1581 | * The mmap call will end up in drm_gem_mmap(), which will set things | |
1582 | * up so we can get faults in the handler above. | |
1583 | * | |
1584 | * The fault handler will take care of binding the object into the GTT | |
1585 | * (since it may have been evicted to make room for something), allocating | |
1586 | * a fence register, and mapping the appropriate aperture address into | |
1587 | * userspace. | |
1588 | */ | |
1589 | int | |
1590 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, | |
1591 | struct drm_file *file) | |
1592 | { | |
1593 | struct drm_i915_gem_mmap_gtt *args = data; | |
1594 | ||
ff72145b DA |
1595 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); |
1596 | } | |
1597 | ||
225067ee DV |
1598 | /* Immediately discard the backing storage */ |
1599 | static void | |
1600 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) | |
e5281ccd | 1601 | { |
e5281ccd | 1602 | struct inode *inode; |
e5281ccd | 1603 | |
4d6294bf | 1604 | i915_gem_object_free_mmap_offset(obj); |
1286ff73 | 1605 | |
4d6294bf CW |
1606 | if (obj->base.filp == NULL) |
1607 | return; | |
e5281ccd | 1608 | |
225067ee DV |
1609 | /* Our goal here is to return as much of the memory as |
1610 | * is possible back to the system as we are called from OOM. | |
1611 | * To do this we must instruct the shmfs to drop all of its | |
1612 | * backing pages, *now*. | |
1613 | */ | |
496ad9aa | 1614 | inode = file_inode(obj->base.filp); |
225067ee | 1615 | shmem_truncate_range(inode, 0, (loff_t)-1); |
e5281ccd | 1616 | |
225067ee DV |
1617 | obj->madv = __I915_MADV_PURGED; |
1618 | } | |
e5281ccd | 1619 | |
225067ee DV |
1620 | static inline int |
1621 | i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj) | |
1622 | { | |
1623 | return obj->madv == I915_MADV_DONTNEED; | |
e5281ccd CW |
1624 | } |
1625 | ||
5cdf5881 | 1626 | static void |
05394f39 | 1627 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj) |
673a394b | 1628 | { |
90797e6d ID |
1629 | struct sg_page_iter sg_iter; |
1630 | int ret; | |
1286ff73 | 1631 | |
05394f39 | 1632 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
673a394b | 1633 | |
6c085a72 CW |
1634 | ret = i915_gem_object_set_to_cpu_domain(obj, true); |
1635 | if (ret) { | |
1636 | /* In the event of a disaster, abandon all caches and | |
1637 | * hope for the best. | |
1638 | */ | |
1639 | WARN_ON(ret != -EIO); | |
2c22569b | 1640 | i915_gem_clflush_object(obj, true); |
6c085a72 CW |
1641 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
1642 | } | |
1643 | ||
6dacfd2f | 1644 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
280b713b EA |
1645 | i915_gem_object_save_bit_17_swizzle(obj); |
1646 | ||
05394f39 CW |
1647 | if (obj->madv == I915_MADV_DONTNEED) |
1648 | obj->dirty = 0; | |
3ef94daa | 1649 | |
90797e6d | 1650 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { |
2db76d7c | 1651 | struct page *page = sg_page_iter_page(&sg_iter); |
9da3da66 | 1652 | |
05394f39 | 1653 | if (obj->dirty) |
9da3da66 | 1654 | set_page_dirty(page); |
3ef94daa | 1655 | |
05394f39 | 1656 | if (obj->madv == I915_MADV_WILLNEED) |
9da3da66 | 1657 | mark_page_accessed(page); |
3ef94daa | 1658 | |
9da3da66 | 1659 | page_cache_release(page); |
3ef94daa | 1660 | } |
05394f39 | 1661 | obj->dirty = 0; |
673a394b | 1662 | |
9da3da66 CW |
1663 | sg_free_table(obj->pages); |
1664 | kfree(obj->pages); | |
37e680a1 | 1665 | } |
6c085a72 | 1666 | |
dd624afd | 1667 | int |
37e680a1 CW |
1668 | i915_gem_object_put_pages(struct drm_i915_gem_object *obj) |
1669 | { | |
1670 | const struct drm_i915_gem_object_ops *ops = obj->ops; | |
1671 | ||
2f745ad3 | 1672 | if (obj->pages == NULL) |
37e680a1 CW |
1673 | return 0; |
1674 | ||
a5570178 CW |
1675 | if (obj->pages_pin_count) |
1676 | return -EBUSY; | |
1677 | ||
9843877d | 1678 | BUG_ON(i915_gem_obj_bound_any(obj)); |
3e123027 | 1679 | |
a2165e31 CW |
1680 | /* ->put_pages might need to allocate memory for the bit17 swizzle |
1681 | * array, hence protect them from being reaped by removing them from gtt | |
1682 | * lists early. */ | |
35c20a60 | 1683 | list_del(&obj->global_list); |
a2165e31 | 1684 | |
37e680a1 | 1685 | ops->put_pages(obj); |
05394f39 | 1686 | obj->pages = NULL; |
37e680a1 | 1687 | |
6c085a72 CW |
1688 | if (i915_gem_object_is_purgeable(obj)) |
1689 | i915_gem_object_truncate(obj); | |
1690 | ||
1691 | return 0; | |
1692 | } | |
1693 | ||
1694 | static long | |
93927ca5 DV |
1695 | __i915_gem_shrink(struct drm_i915_private *dev_priv, long target, |
1696 | bool purgeable_only) | |
6c085a72 CW |
1697 | { |
1698 | struct drm_i915_gem_object *obj, *next; | |
1699 | long count = 0; | |
1700 | ||
1701 | list_for_each_entry_safe(obj, next, | |
1702 | &dev_priv->mm.unbound_list, | |
35c20a60 | 1703 | global_list) { |
93927ca5 | 1704 | if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) && |
37e680a1 | 1705 | i915_gem_object_put_pages(obj) == 0) { |
6c085a72 CW |
1706 | count += obj->base.size >> PAGE_SHIFT; |
1707 | if (count >= target) | |
1708 | return count; | |
1709 | } | |
1710 | } | |
1711 | ||
07fe0b12 BW |
1712 | list_for_each_entry_safe(obj, next, &dev_priv->mm.bound_list, |
1713 | global_list) { | |
1714 | struct i915_vma *vma, *v; | |
80dcfdbd BW |
1715 | |
1716 | if (!i915_gem_object_is_purgeable(obj) && purgeable_only) | |
1717 | continue; | |
1718 | ||
07fe0b12 BW |
1719 | list_for_each_entry_safe(vma, v, &obj->vma_list, vma_link) |
1720 | if (i915_vma_unbind(vma)) | |
1721 | break; | |
80dcfdbd BW |
1722 | |
1723 | if (!i915_gem_object_put_pages(obj)) { | |
6c085a72 CW |
1724 | count += obj->base.size >> PAGE_SHIFT; |
1725 | if (count >= target) | |
1726 | return count; | |
1727 | } | |
1728 | } | |
1729 | ||
1730 | return count; | |
1731 | } | |
1732 | ||
93927ca5 DV |
1733 | static long |
1734 | i915_gem_purge(struct drm_i915_private *dev_priv, long target) | |
1735 | { | |
1736 | return __i915_gem_shrink(dev_priv, target, true); | |
1737 | } | |
1738 | ||
6c085a72 CW |
1739 | static void |
1740 | i915_gem_shrink_all(struct drm_i915_private *dev_priv) | |
1741 | { | |
1742 | struct drm_i915_gem_object *obj, *next; | |
1743 | ||
1744 | i915_gem_evict_everything(dev_priv->dev); | |
1745 | ||
35c20a60 BW |
1746 | list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list, |
1747 | global_list) | |
37e680a1 | 1748 | i915_gem_object_put_pages(obj); |
225067ee DV |
1749 | } |
1750 | ||
37e680a1 | 1751 | static int |
6c085a72 | 1752 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj) |
e5281ccd | 1753 | { |
6c085a72 | 1754 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
e5281ccd CW |
1755 | int page_count, i; |
1756 | struct address_space *mapping; | |
9da3da66 CW |
1757 | struct sg_table *st; |
1758 | struct scatterlist *sg; | |
90797e6d | 1759 | struct sg_page_iter sg_iter; |
e5281ccd | 1760 | struct page *page; |
90797e6d | 1761 | unsigned long last_pfn = 0; /* suppress gcc warning */ |
6c085a72 | 1762 | gfp_t gfp; |
e5281ccd | 1763 | |
6c085a72 CW |
1764 | /* Assert that the object is not currently in any GPU domain. As it |
1765 | * wasn't in the GTT, there shouldn't be any way it could have been in | |
1766 | * a GPU cache | |
1767 | */ | |
1768 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); | |
1769 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); | |
1770 | ||
9da3da66 CW |
1771 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
1772 | if (st == NULL) | |
1773 | return -ENOMEM; | |
1774 | ||
05394f39 | 1775 | page_count = obj->base.size / PAGE_SIZE; |
9da3da66 | 1776 | if (sg_alloc_table(st, page_count, GFP_KERNEL)) { |
9da3da66 | 1777 | kfree(st); |
e5281ccd | 1778 | return -ENOMEM; |
9da3da66 | 1779 | } |
e5281ccd | 1780 | |
9da3da66 CW |
1781 | /* Get the list of pages out of our struct file. They'll be pinned |
1782 | * at this point until we release them. | |
1783 | * | |
1784 | * Fail silently without starting the shrinker | |
1785 | */ | |
496ad9aa | 1786 | mapping = file_inode(obj->base.filp)->i_mapping; |
6c085a72 | 1787 | gfp = mapping_gfp_mask(mapping); |
caf49191 | 1788 | gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD; |
6c085a72 | 1789 | gfp &= ~(__GFP_IO | __GFP_WAIT); |
90797e6d ID |
1790 | sg = st->sgl; |
1791 | st->nents = 0; | |
1792 | for (i = 0; i < page_count; i++) { | |
6c085a72 CW |
1793 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); |
1794 | if (IS_ERR(page)) { | |
1795 | i915_gem_purge(dev_priv, page_count); | |
1796 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); | |
1797 | } | |
1798 | if (IS_ERR(page)) { | |
1799 | /* We've tried hard to allocate the memory by reaping | |
1800 | * our own buffer, now let the real VM do its job and | |
1801 | * go down in flames if truly OOM. | |
1802 | */ | |
caf49191 | 1803 | gfp &= ~(__GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD); |
6c085a72 CW |
1804 | gfp |= __GFP_IO | __GFP_WAIT; |
1805 | ||
1806 | i915_gem_shrink_all(dev_priv); | |
1807 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); | |
1808 | if (IS_ERR(page)) | |
1809 | goto err_pages; | |
1810 | ||
caf49191 | 1811 | gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD; |
6c085a72 CW |
1812 | gfp &= ~(__GFP_IO | __GFP_WAIT); |
1813 | } | |
426729dc KRW |
1814 | #ifdef CONFIG_SWIOTLB |
1815 | if (swiotlb_nr_tbl()) { | |
1816 | st->nents++; | |
1817 | sg_set_page(sg, page, PAGE_SIZE, 0); | |
1818 | sg = sg_next(sg); | |
1819 | continue; | |
1820 | } | |
1821 | #endif | |
90797e6d ID |
1822 | if (!i || page_to_pfn(page) != last_pfn + 1) { |
1823 | if (i) | |
1824 | sg = sg_next(sg); | |
1825 | st->nents++; | |
1826 | sg_set_page(sg, page, PAGE_SIZE, 0); | |
1827 | } else { | |
1828 | sg->length += PAGE_SIZE; | |
1829 | } | |
1830 | last_pfn = page_to_pfn(page); | |
e5281ccd | 1831 | } |
426729dc KRW |
1832 | #ifdef CONFIG_SWIOTLB |
1833 | if (!swiotlb_nr_tbl()) | |
1834 | #endif | |
1835 | sg_mark_end(sg); | |
74ce6b6c CW |
1836 | obj->pages = st; |
1837 | ||
6dacfd2f | 1838 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
e5281ccd CW |
1839 | i915_gem_object_do_bit_17_swizzle(obj); |
1840 | ||
1841 | return 0; | |
1842 | ||
1843 | err_pages: | |
90797e6d ID |
1844 | sg_mark_end(sg); |
1845 | for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) | |
2db76d7c | 1846 | page_cache_release(sg_page_iter_page(&sg_iter)); |
9da3da66 CW |
1847 | sg_free_table(st); |
1848 | kfree(st); | |
e5281ccd | 1849 | return PTR_ERR(page); |
673a394b EA |
1850 | } |
1851 | ||
37e680a1 CW |
1852 | /* Ensure that the associated pages are gathered from the backing storage |
1853 | * and pinned into our object. i915_gem_object_get_pages() may be called | |
1854 | * multiple times before they are released by a single call to | |
1855 | * i915_gem_object_put_pages() - once the pages are no longer referenced | |
1856 | * either as a result of memory pressure (reaping pages under the shrinker) | |
1857 | * or as the object is itself released. | |
1858 | */ | |
1859 | int | |
1860 | i915_gem_object_get_pages(struct drm_i915_gem_object *obj) | |
1861 | { | |
1862 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
1863 | const struct drm_i915_gem_object_ops *ops = obj->ops; | |
1864 | int ret; | |
1865 | ||
2f745ad3 | 1866 | if (obj->pages) |
37e680a1 CW |
1867 | return 0; |
1868 | ||
43e28f09 CW |
1869 | if (obj->madv != I915_MADV_WILLNEED) { |
1870 | DRM_ERROR("Attempting to obtain a purgeable object\n"); | |
1871 | return -EINVAL; | |
1872 | } | |
1873 | ||
a5570178 CW |
1874 | BUG_ON(obj->pages_pin_count); |
1875 | ||
37e680a1 CW |
1876 | ret = ops->get_pages(obj); |
1877 | if (ret) | |
1878 | return ret; | |
1879 | ||
35c20a60 | 1880 | list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
37e680a1 | 1881 | return 0; |
673a394b EA |
1882 | } |
1883 | ||
54cf91dc | 1884 | void |
05394f39 | 1885 | i915_gem_object_move_to_active(struct drm_i915_gem_object *obj, |
9d773091 | 1886 | struct intel_ring_buffer *ring) |
673a394b | 1887 | { |
05394f39 | 1888 | struct drm_device *dev = obj->base.dev; |
69dc4987 | 1889 | struct drm_i915_private *dev_priv = dev->dev_private; |
9d773091 | 1890 | u32 seqno = intel_ring_get_seqno(ring); |
617dbe27 | 1891 | |
852835f3 | 1892 | BUG_ON(ring == NULL); |
02978ff5 CW |
1893 | if (obj->ring != ring && obj->last_write_seqno) { |
1894 | /* Keep the seqno relative to the current ring */ | |
1895 | obj->last_write_seqno = seqno; | |
1896 | } | |
05394f39 | 1897 | obj->ring = ring; |
673a394b EA |
1898 | |
1899 | /* Add a reference if we're newly entering the active list. */ | |
05394f39 CW |
1900 | if (!obj->active) { |
1901 | drm_gem_object_reference(&obj->base); | |
1902 | obj->active = 1; | |
673a394b | 1903 | } |
e35a41de | 1904 | |
05394f39 | 1905 | list_move_tail(&obj->ring_list, &ring->active_list); |
caea7476 | 1906 | |
0201f1ec | 1907 | obj->last_read_seqno = seqno; |
caea7476 | 1908 | |
7dd49065 | 1909 | if (obj->fenced_gpu_access) { |
caea7476 | 1910 | obj->last_fenced_seqno = seqno; |
caea7476 | 1911 | |
7dd49065 CW |
1912 | /* Bump MRU to take account of the delayed flush */ |
1913 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
1914 | struct drm_i915_fence_reg *reg; | |
1915 | ||
1916 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
1917 | list_move_tail(®->lru_list, | |
1918 | &dev_priv->mm.fence_list); | |
1919 | } | |
caea7476 CW |
1920 | } |
1921 | } | |
1922 | ||
1923 | static void | |
caea7476 | 1924 | i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj) |
ce44b0ea | 1925 | { |
ca191b13 BW |
1926 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
1927 | struct i915_address_space *ggtt_vm = &dev_priv->gtt.base; | |
1928 | struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm); | |
ce44b0ea | 1929 | |
65ce3027 | 1930 | BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS); |
05394f39 | 1931 | BUG_ON(!obj->active); |
caea7476 | 1932 | |
ca191b13 | 1933 | list_move_tail(&vma->mm_list, &ggtt_vm->inactive_list); |
caea7476 | 1934 | |
65ce3027 | 1935 | list_del_init(&obj->ring_list); |
caea7476 CW |
1936 | obj->ring = NULL; |
1937 | ||
65ce3027 CW |
1938 | obj->last_read_seqno = 0; |
1939 | obj->last_write_seqno = 0; | |
1940 | obj->base.write_domain = 0; | |
1941 | ||
1942 | obj->last_fenced_seqno = 0; | |
caea7476 | 1943 | obj->fenced_gpu_access = false; |
caea7476 CW |
1944 | |
1945 | obj->active = 0; | |
1946 | drm_gem_object_unreference(&obj->base); | |
1947 | ||
1948 | WARN_ON(i915_verify_lists(dev)); | |
ce44b0ea | 1949 | } |
673a394b | 1950 | |
9d773091 | 1951 | static int |
fca26bb4 | 1952 | i915_gem_init_seqno(struct drm_device *dev, u32 seqno) |
53d227f2 | 1953 | { |
9d773091 CW |
1954 | struct drm_i915_private *dev_priv = dev->dev_private; |
1955 | struct intel_ring_buffer *ring; | |
1956 | int ret, i, j; | |
53d227f2 | 1957 | |
107f27a5 | 1958 | /* Carefully retire all requests without writing to the rings */ |
9d773091 | 1959 | for_each_ring(ring, dev_priv, i) { |
107f27a5 CW |
1960 | ret = intel_ring_idle(ring); |
1961 | if (ret) | |
1962 | return ret; | |
9d773091 | 1963 | } |
9d773091 | 1964 | i915_gem_retire_requests(dev); |
107f27a5 CW |
1965 | |
1966 | /* Finally reset hw state */ | |
9d773091 | 1967 | for_each_ring(ring, dev_priv, i) { |
fca26bb4 | 1968 | intel_ring_init_seqno(ring, seqno); |
498d2ac1 | 1969 | |
9d773091 CW |
1970 | for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++) |
1971 | ring->sync_seqno[j] = 0; | |
1972 | } | |
53d227f2 | 1973 | |
9d773091 | 1974 | return 0; |
53d227f2 DV |
1975 | } |
1976 | ||
fca26bb4 MK |
1977 | int i915_gem_set_seqno(struct drm_device *dev, u32 seqno) |
1978 | { | |
1979 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1980 | int ret; | |
1981 | ||
1982 | if (seqno == 0) | |
1983 | return -EINVAL; | |
1984 | ||
1985 | /* HWS page needs to be set less than what we | |
1986 | * will inject to ring | |
1987 | */ | |
1988 | ret = i915_gem_init_seqno(dev, seqno - 1); | |
1989 | if (ret) | |
1990 | return ret; | |
1991 | ||
1992 | /* Carefully set the last_seqno value so that wrap | |
1993 | * detection still works | |
1994 | */ | |
1995 | dev_priv->next_seqno = seqno; | |
1996 | dev_priv->last_seqno = seqno - 1; | |
1997 | if (dev_priv->last_seqno == 0) | |
1998 | dev_priv->last_seqno--; | |
1999 | ||
2000 | return 0; | |
2001 | } | |
2002 | ||
9d773091 CW |
2003 | int |
2004 | i915_gem_get_seqno(struct drm_device *dev, u32 *seqno) | |
53d227f2 | 2005 | { |
9d773091 CW |
2006 | struct drm_i915_private *dev_priv = dev->dev_private; |
2007 | ||
2008 | /* reserve 0 for non-seqno */ | |
2009 | if (dev_priv->next_seqno == 0) { | |
fca26bb4 | 2010 | int ret = i915_gem_init_seqno(dev, 0); |
9d773091 CW |
2011 | if (ret) |
2012 | return ret; | |
53d227f2 | 2013 | |
9d773091 CW |
2014 | dev_priv->next_seqno = 1; |
2015 | } | |
53d227f2 | 2016 | |
f72b3435 | 2017 | *seqno = dev_priv->last_seqno = dev_priv->next_seqno++; |
9d773091 | 2018 | return 0; |
53d227f2 DV |
2019 | } |
2020 | ||
0025c077 MK |
2021 | int __i915_add_request(struct intel_ring_buffer *ring, |
2022 | struct drm_file *file, | |
7d736f4f | 2023 | struct drm_i915_gem_object *obj, |
0025c077 | 2024 | u32 *out_seqno) |
673a394b | 2025 | { |
db53a302 | 2026 | drm_i915_private_t *dev_priv = ring->dev->dev_private; |
acb868d3 | 2027 | struct drm_i915_gem_request *request; |
7d736f4f | 2028 | u32 request_ring_position, request_start; |
673a394b | 2029 | int was_empty; |
3cce469c CW |
2030 | int ret; |
2031 | ||
7d736f4f | 2032 | request_start = intel_ring_get_tail(ring); |
cc889e0f DV |
2033 | /* |
2034 | * Emit any outstanding flushes - execbuf can fail to emit the flush | |
2035 | * after having emitted the batchbuffer command. Hence we need to fix | |
2036 | * things up similar to emitting the lazy request. The difference here | |
2037 | * is that the flush _must_ happen before the next request, no matter | |
2038 | * what. | |
2039 | */ | |
a7b9761d CW |
2040 | ret = intel_ring_flush_all_caches(ring); |
2041 | if (ret) | |
2042 | return ret; | |
cc889e0f | 2043 | |
acb868d3 CW |
2044 | request = kmalloc(sizeof(*request), GFP_KERNEL); |
2045 | if (request == NULL) | |
2046 | return -ENOMEM; | |
cc889e0f | 2047 | |
a71d8d94 CW |
2048 | /* Record the position of the start of the request so that |
2049 | * should we detect the updated seqno part-way through the | |
2050 | * GPU processing the request, we never over-estimate the | |
2051 | * position of the head. | |
2052 | */ | |
2053 | request_ring_position = intel_ring_get_tail(ring); | |
2054 | ||
9d773091 | 2055 | ret = ring->add_request(ring); |
3bb73aba CW |
2056 | if (ret) { |
2057 | kfree(request); | |
2058 | return ret; | |
2059 | } | |
673a394b | 2060 | |
9d773091 | 2061 | request->seqno = intel_ring_get_seqno(ring); |
852835f3 | 2062 | request->ring = ring; |
7d736f4f | 2063 | request->head = request_start; |
a71d8d94 | 2064 | request->tail = request_ring_position; |
7d736f4f MK |
2065 | |
2066 | /* Whilst this request exists, batch_obj will be on the | |
2067 | * active_list, and so will hold the active reference. Only when this | |
2068 | * request is retired will the the batch_obj be moved onto the | |
2069 | * inactive_list and lose its active reference. Hence we do not need | |
2070 | * to explicitly hold another reference here. | |
2071 | */ | |
9a7e0c2a | 2072 | request->batch_obj = obj; |
0e50e96b | 2073 | |
9a7e0c2a CW |
2074 | /* Hold a reference to the current context so that we can inspect |
2075 | * it later in case a hangcheck error event fires. | |
2076 | */ | |
2077 | request->ctx = ring->last_context; | |
0e50e96b MK |
2078 | if (request->ctx) |
2079 | i915_gem_context_reference(request->ctx); | |
2080 | ||
673a394b | 2081 | request->emitted_jiffies = jiffies; |
852835f3 ZN |
2082 | was_empty = list_empty(&ring->request_list); |
2083 | list_add_tail(&request->list, &ring->request_list); | |
3bb73aba | 2084 | request->file_priv = NULL; |
852835f3 | 2085 | |
db53a302 CW |
2086 | if (file) { |
2087 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
2088 | ||
1c25595f | 2089 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 2090 | request->file_priv = file_priv; |
b962442e | 2091 | list_add_tail(&request->client_list, |
f787a5f5 | 2092 | &file_priv->mm.request_list); |
1c25595f | 2093 | spin_unlock(&file_priv->mm.lock); |
b962442e | 2094 | } |
673a394b | 2095 | |
9d773091 | 2096 | trace_i915_gem_request_add(ring, request->seqno); |
1823521d | 2097 | ring->outstanding_lazy_seqno = 0; |
db53a302 | 2098 | |
db1b76ca | 2099 | if (!dev_priv->ums.mm_suspended) { |
10cd45b6 MK |
2100 | i915_queue_hangcheck(ring->dev); |
2101 | ||
f047e395 | 2102 | if (was_empty) { |
b3b079db | 2103 | queue_delayed_work(dev_priv->wq, |
bcb45086 CW |
2104 | &dev_priv->mm.retire_work, |
2105 | round_jiffies_up_relative(HZ)); | |
f047e395 CW |
2106 | intel_mark_busy(dev_priv->dev); |
2107 | } | |
f65d9421 | 2108 | } |
cc889e0f | 2109 | |
acb868d3 | 2110 | if (out_seqno) |
9d773091 | 2111 | *out_seqno = request->seqno; |
3cce469c | 2112 | return 0; |
673a394b EA |
2113 | } |
2114 | ||
f787a5f5 CW |
2115 | static inline void |
2116 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) | |
673a394b | 2117 | { |
1c25595f | 2118 | struct drm_i915_file_private *file_priv = request->file_priv; |
673a394b | 2119 | |
1c25595f CW |
2120 | if (!file_priv) |
2121 | return; | |
1c5d22f7 | 2122 | |
1c25595f | 2123 | spin_lock(&file_priv->mm.lock); |
09bfa517 HRK |
2124 | if (request->file_priv) { |
2125 | list_del(&request->client_list); | |
2126 | request->file_priv = NULL; | |
2127 | } | |
1c25595f | 2128 | spin_unlock(&file_priv->mm.lock); |
673a394b | 2129 | } |
673a394b | 2130 | |
d1ccbb5d BW |
2131 | static bool i915_head_inside_object(u32 acthd, struct drm_i915_gem_object *obj, |
2132 | struct i915_address_space *vm) | |
aa60c664 | 2133 | { |
d1ccbb5d BW |
2134 | if (acthd >= i915_gem_obj_offset(obj, vm) && |
2135 | acthd < i915_gem_obj_offset(obj, vm) + obj->base.size) | |
aa60c664 MK |
2136 | return true; |
2137 | ||
2138 | return false; | |
2139 | } | |
2140 | ||
2141 | static bool i915_head_inside_request(const u32 acthd_unmasked, | |
2142 | const u32 request_start, | |
2143 | const u32 request_end) | |
2144 | { | |
2145 | const u32 acthd = acthd_unmasked & HEAD_ADDR; | |
2146 | ||
2147 | if (request_start < request_end) { | |
2148 | if (acthd >= request_start && acthd < request_end) | |
2149 | return true; | |
2150 | } else if (request_start > request_end) { | |
2151 | if (acthd >= request_start || acthd < request_end) | |
2152 | return true; | |
2153 | } | |
2154 | ||
2155 | return false; | |
2156 | } | |
2157 | ||
d1ccbb5d BW |
2158 | static struct i915_address_space * |
2159 | request_to_vm(struct drm_i915_gem_request *request) | |
2160 | { | |
2161 | struct drm_i915_private *dev_priv = request->ring->dev->dev_private; | |
2162 | struct i915_address_space *vm; | |
2163 | ||
2164 | vm = &dev_priv->gtt.base; | |
2165 | ||
2166 | return vm; | |
2167 | } | |
2168 | ||
aa60c664 MK |
2169 | static bool i915_request_guilty(struct drm_i915_gem_request *request, |
2170 | const u32 acthd, bool *inside) | |
2171 | { | |
2172 | /* There is a possibility that unmasked head address | |
2173 | * pointing inside the ring, matches the batch_obj address range. | |
2174 | * However this is extremely unlikely. | |
2175 | */ | |
aa60c664 | 2176 | if (request->batch_obj) { |
d1ccbb5d BW |
2177 | if (i915_head_inside_object(acthd, request->batch_obj, |
2178 | request_to_vm(request))) { | |
aa60c664 MK |
2179 | *inside = true; |
2180 | return true; | |
2181 | } | |
2182 | } | |
2183 | ||
2184 | if (i915_head_inside_request(acthd, request->head, request->tail)) { | |
2185 | *inside = false; | |
2186 | return true; | |
2187 | } | |
2188 | ||
2189 | return false; | |
2190 | } | |
2191 | ||
2192 | static void i915_set_reset_status(struct intel_ring_buffer *ring, | |
2193 | struct drm_i915_gem_request *request, | |
2194 | u32 acthd) | |
2195 | { | |
2196 | struct i915_ctx_hang_stats *hs = NULL; | |
2197 | bool inside, guilty; | |
d1ccbb5d | 2198 | unsigned long offset = 0; |
aa60c664 MK |
2199 | |
2200 | /* Innocent until proven guilty */ | |
2201 | guilty = false; | |
2202 | ||
d1ccbb5d BW |
2203 | if (request->batch_obj) |
2204 | offset = i915_gem_obj_offset(request->batch_obj, | |
2205 | request_to_vm(request)); | |
2206 | ||
f2f4d82f | 2207 | if (ring->hangcheck.action != HANGCHECK_WAIT && |
aa60c664 | 2208 | i915_request_guilty(request, acthd, &inside)) { |
f343c5f6 | 2209 | DRM_ERROR("%s hung %s bo (0x%lx ctx %d) at 0x%x\n", |
aa60c664 MK |
2210 | ring->name, |
2211 | inside ? "inside" : "flushing", | |
d1ccbb5d | 2212 | offset, |
aa60c664 MK |
2213 | request->ctx ? request->ctx->id : 0, |
2214 | acthd); | |
2215 | ||
2216 | guilty = true; | |
2217 | } | |
2218 | ||
2219 | /* If contexts are disabled or this is the default context, use | |
2220 | * file_priv->reset_state | |
2221 | */ | |
2222 | if (request->ctx && request->ctx->id != DEFAULT_CONTEXT_ID) | |
2223 | hs = &request->ctx->hang_stats; | |
2224 | else if (request->file_priv) | |
2225 | hs = &request->file_priv->hang_stats; | |
2226 | ||
2227 | if (hs) { | |
2228 | if (guilty) | |
2229 | hs->batch_active++; | |
2230 | else | |
2231 | hs->batch_pending++; | |
2232 | } | |
2233 | } | |
2234 | ||
0e50e96b MK |
2235 | static void i915_gem_free_request(struct drm_i915_gem_request *request) |
2236 | { | |
2237 | list_del(&request->list); | |
2238 | i915_gem_request_remove_from_client(request); | |
2239 | ||
2240 | if (request->ctx) | |
2241 | i915_gem_context_unreference(request->ctx); | |
2242 | ||
2243 | kfree(request); | |
2244 | } | |
2245 | ||
dfaae392 CW |
2246 | static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv, |
2247 | struct intel_ring_buffer *ring) | |
9375e446 | 2248 | { |
aa60c664 MK |
2249 | u32 completed_seqno; |
2250 | u32 acthd; | |
2251 | ||
2252 | acthd = intel_ring_get_active_head(ring); | |
2253 | completed_seqno = ring->get_seqno(ring, false); | |
2254 | ||
dfaae392 CW |
2255 | while (!list_empty(&ring->request_list)) { |
2256 | struct drm_i915_gem_request *request; | |
673a394b | 2257 | |
dfaae392 CW |
2258 | request = list_first_entry(&ring->request_list, |
2259 | struct drm_i915_gem_request, | |
2260 | list); | |
de151cf6 | 2261 | |
aa60c664 MK |
2262 | if (request->seqno > completed_seqno) |
2263 | i915_set_reset_status(ring, request, acthd); | |
2264 | ||
0e50e96b | 2265 | i915_gem_free_request(request); |
dfaae392 | 2266 | } |
673a394b | 2267 | |
dfaae392 | 2268 | while (!list_empty(&ring->active_list)) { |
05394f39 | 2269 | struct drm_i915_gem_object *obj; |
9375e446 | 2270 | |
05394f39 CW |
2271 | obj = list_first_entry(&ring->active_list, |
2272 | struct drm_i915_gem_object, | |
2273 | ring_list); | |
9375e446 | 2274 | |
05394f39 | 2275 | i915_gem_object_move_to_inactive(obj); |
673a394b EA |
2276 | } |
2277 | } | |
2278 | ||
19b2dbde | 2279 | void i915_gem_restore_fences(struct drm_device *dev) |
312817a3 CW |
2280 | { |
2281 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2282 | int i; | |
2283 | ||
4b9de737 | 2284 | for (i = 0; i < dev_priv->num_fence_regs; i++) { |
312817a3 | 2285 | struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i]; |
7d2cb39c | 2286 | |
94a335db DV |
2287 | /* |
2288 | * Commit delayed tiling changes if we have an object still | |
2289 | * attached to the fence, otherwise just clear the fence. | |
2290 | */ | |
2291 | if (reg->obj) { | |
2292 | i915_gem_object_update_fence(reg->obj, reg, | |
2293 | reg->obj->tiling_mode); | |
2294 | } else { | |
2295 | i915_gem_write_fence(dev, i, NULL); | |
2296 | } | |
312817a3 CW |
2297 | } |
2298 | } | |
2299 | ||
069efc1d | 2300 | void i915_gem_reset(struct drm_device *dev) |
673a394b | 2301 | { |
77f01230 | 2302 | struct drm_i915_private *dev_priv = dev->dev_private; |
b4519513 | 2303 | struct intel_ring_buffer *ring; |
1ec14ad3 | 2304 | int i; |
673a394b | 2305 | |
b4519513 CW |
2306 | for_each_ring(ring, dev_priv, i) |
2307 | i915_gem_reset_ring_lists(dev_priv, ring); | |
dfaae392 | 2308 | |
19b2dbde | 2309 | i915_gem_restore_fences(dev); |
673a394b EA |
2310 | } |
2311 | ||
2312 | /** | |
2313 | * This function clears the request list as sequence numbers are passed. | |
2314 | */ | |
a71d8d94 | 2315 | void |
db53a302 | 2316 | i915_gem_retire_requests_ring(struct intel_ring_buffer *ring) |
673a394b | 2317 | { |
673a394b EA |
2318 | uint32_t seqno; |
2319 | ||
db53a302 | 2320 | if (list_empty(&ring->request_list)) |
6c0594a3 KW |
2321 | return; |
2322 | ||
db53a302 | 2323 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b | 2324 | |
b2eadbc8 | 2325 | seqno = ring->get_seqno(ring, true); |
1ec14ad3 | 2326 | |
852835f3 | 2327 | while (!list_empty(&ring->request_list)) { |
673a394b | 2328 | struct drm_i915_gem_request *request; |
673a394b | 2329 | |
852835f3 | 2330 | request = list_first_entry(&ring->request_list, |
673a394b EA |
2331 | struct drm_i915_gem_request, |
2332 | list); | |
673a394b | 2333 | |
dfaae392 | 2334 | if (!i915_seqno_passed(seqno, request->seqno)) |
b84d5f0c CW |
2335 | break; |
2336 | ||
db53a302 | 2337 | trace_i915_gem_request_retire(ring, request->seqno); |
a71d8d94 CW |
2338 | /* We know the GPU must have read the request to have |
2339 | * sent us the seqno + interrupt, so use the position | |
2340 | * of tail of the request to update the last known position | |
2341 | * of the GPU head. | |
2342 | */ | |
2343 | ring->last_retired_head = request->tail; | |
b84d5f0c | 2344 | |
0e50e96b | 2345 | i915_gem_free_request(request); |
b84d5f0c | 2346 | } |
673a394b | 2347 | |
b84d5f0c CW |
2348 | /* Move any buffers on the active list that are no longer referenced |
2349 | * by the ringbuffer to the flushing/inactive lists as appropriate. | |
2350 | */ | |
2351 | while (!list_empty(&ring->active_list)) { | |
05394f39 | 2352 | struct drm_i915_gem_object *obj; |
b84d5f0c | 2353 | |
0206e353 | 2354 | obj = list_first_entry(&ring->active_list, |
05394f39 CW |
2355 | struct drm_i915_gem_object, |
2356 | ring_list); | |
673a394b | 2357 | |
0201f1ec | 2358 | if (!i915_seqno_passed(seqno, obj->last_read_seqno)) |
673a394b | 2359 | break; |
b84d5f0c | 2360 | |
65ce3027 | 2361 | i915_gem_object_move_to_inactive(obj); |
673a394b | 2362 | } |
9d34e5db | 2363 | |
db53a302 CW |
2364 | if (unlikely(ring->trace_irq_seqno && |
2365 | i915_seqno_passed(seqno, ring->trace_irq_seqno))) { | |
1ec14ad3 | 2366 | ring->irq_put(ring); |
db53a302 | 2367 | ring->trace_irq_seqno = 0; |
9d34e5db | 2368 | } |
23bc5982 | 2369 | |
db53a302 | 2370 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b EA |
2371 | } |
2372 | ||
b09a1fec CW |
2373 | void |
2374 | i915_gem_retire_requests(struct drm_device *dev) | |
2375 | { | |
2376 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 2377 | struct intel_ring_buffer *ring; |
1ec14ad3 | 2378 | int i; |
b09a1fec | 2379 | |
b4519513 CW |
2380 | for_each_ring(ring, dev_priv, i) |
2381 | i915_gem_retire_requests_ring(ring); | |
b09a1fec CW |
2382 | } |
2383 | ||
75ef9da2 | 2384 | static void |
673a394b EA |
2385 | i915_gem_retire_work_handler(struct work_struct *work) |
2386 | { | |
2387 | drm_i915_private_t *dev_priv; | |
2388 | struct drm_device *dev; | |
b4519513 | 2389 | struct intel_ring_buffer *ring; |
0a58705b CW |
2390 | bool idle; |
2391 | int i; | |
673a394b EA |
2392 | |
2393 | dev_priv = container_of(work, drm_i915_private_t, | |
2394 | mm.retire_work.work); | |
2395 | dev = dev_priv->dev; | |
2396 | ||
891b48cf CW |
2397 | /* Come back later if the device is busy... */ |
2398 | if (!mutex_trylock(&dev->struct_mutex)) { | |
bcb45086 CW |
2399 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2400 | round_jiffies_up_relative(HZ)); | |
891b48cf CW |
2401 | return; |
2402 | } | |
673a394b | 2403 | |
b09a1fec | 2404 | i915_gem_retire_requests(dev); |
673a394b | 2405 | |
0a58705b CW |
2406 | /* Send a periodic flush down the ring so we don't hold onto GEM |
2407 | * objects indefinitely. | |
673a394b | 2408 | */ |
0a58705b | 2409 | idle = true; |
b4519513 | 2410 | for_each_ring(ring, dev_priv, i) { |
3bb73aba | 2411 | if (ring->gpu_caches_dirty) |
0025c077 | 2412 | i915_add_request(ring, NULL); |
0a58705b CW |
2413 | |
2414 | idle &= list_empty(&ring->request_list); | |
673a394b EA |
2415 | } |
2416 | ||
db1b76ca | 2417 | if (!dev_priv->ums.mm_suspended && !idle) |
bcb45086 CW |
2418 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2419 | round_jiffies_up_relative(HZ)); | |
f047e395 CW |
2420 | if (idle) |
2421 | intel_mark_idle(dev); | |
0a58705b | 2422 | |
673a394b | 2423 | mutex_unlock(&dev->struct_mutex); |
673a394b EA |
2424 | } |
2425 | ||
30dfebf3 DV |
2426 | /** |
2427 | * Ensures that an object will eventually get non-busy by flushing any required | |
2428 | * write domains, emitting any outstanding lazy request and retiring and | |
2429 | * completed requests. | |
2430 | */ | |
2431 | static int | |
2432 | i915_gem_object_flush_active(struct drm_i915_gem_object *obj) | |
2433 | { | |
2434 | int ret; | |
2435 | ||
2436 | if (obj->active) { | |
0201f1ec | 2437 | ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno); |
30dfebf3 DV |
2438 | if (ret) |
2439 | return ret; | |
2440 | ||
30dfebf3 DV |
2441 | i915_gem_retire_requests_ring(obj->ring); |
2442 | } | |
2443 | ||
2444 | return 0; | |
2445 | } | |
2446 | ||
23ba4fd0 BW |
2447 | /** |
2448 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT | |
2449 | * @DRM_IOCTL_ARGS: standard ioctl arguments | |
2450 | * | |
2451 | * Returns 0 if successful, else an error is returned with the remaining time in | |
2452 | * the timeout parameter. | |
2453 | * -ETIME: object is still busy after timeout | |
2454 | * -ERESTARTSYS: signal interrupted the wait | |
2455 | * -ENONENT: object doesn't exist | |
2456 | * Also possible, but rare: | |
2457 | * -EAGAIN: GPU wedged | |
2458 | * -ENOMEM: damn | |
2459 | * -ENODEV: Internal IRQ fail | |
2460 | * -E?: The add request failed | |
2461 | * | |
2462 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any | |
2463 | * non-zero timeout parameter the wait ioctl will wait for the given number of | |
2464 | * nanoseconds on an object becoming unbusy. Since the wait itself does so | |
2465 | * without holding struct_mutex the object may become re-busied before this | |
2466 | * function completes. A similar but shorter * race condition exists in the busy | |
2467 | * ioctl | |
2468 | */ | |
2469 | int | |
2470 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
2471 | { | |
f69061be | 2472 | drm_i915_private_t *dev_priv = dev->dev_private; |
23ba4fd0 BW |
2473 | struct drm_i915_gem_wait *args = data; |
2474 | struct drm_i915_gem_object *obj; | |
2475 | struct intel_ring_buffer *ring = NULL; | |
eac1f14f | 2476 | struct timespec timeout_stack, *timeout = NULL; |
f69061be | 2477 | unsigned reset_counter; |
23ba4fd0 BW |
2478 | u32 seqno = 0; |
2479 | int ret = 0; | |
2480 | ||
eac1f14f BW |
2481 | if (args->timeout_ns >= 0) { |
2482 | timeout_stack = ns_to_timespec(args->timeout_ns); | |
2483 | timeout = &timeout_stack; | |
2484 | } | |
23ba4fd0 BW |
2485 | |
2486 | ret = i915_mutex_lock_interruptible(dev); | |
2487 | if (ret) | |
2488 | return ret; | |
2489 | ||
2490 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle)); | |
2491 | if (&obj->base == NULL) { | |
2492 | mutex_unlock(&dev->struct_mutex); | |
2493 | return -ENOENT; | |
2494 | } | |
2495 | ||
30dfebf3 DV |
2496 | /* Need to make sure the object gets inactive eventually. */ |
2497 | ret = i915_gem_object_flush_active(obj); | |
23ba4fd0 BW |
2498 | if (ret) |
2499 | goto out; | |
2500 | ||
2501 | if (obj->active) { | |
0201f1ec | 2502 | seqno = obj->last_read_seqno; |
23ba4fd0 BW |
2503 | ring = obj->ring; |
2504 | } | |
2505 | ||
2506 | if (seqno == 0) | |
2507 | goto out; | |
2508 | ||
23ba4fd0 BW |
2509 | /* Do this after OLR check to make sure we make forward progress polling |
2510 | * on this IOCTL with a 0 timeout (like busy ioctl) | |
2511 | */ | |
2512 | if (!args->timeout_ns) { | |
2513 | ret = -ETIME; | |
2514 | goto out; | |
2515 | } | |
2516 | ||
2517 | drm_gem_object_unreference(&obj->base); | |
f69061be | 2518 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
23ba4fd0 BW |
2519 | mutex_unlock(&dev->struct_mutex); |
2520 | ||
f69061be | 2521 | ret = __wait_seqno(ring, seqno, reset_counter, true, timeout); |
4f42f4ef | 2522 | if (timeout) |
eac1f14f | 2523 | args->timeout_ns = timespec_to_ns(timeout); |
23ba4fd0 BW |
2524 | return ret; |
2525 | ||
2526 | out: | |
2527 | drm_gem_object_unreference(&obj->base); | |
2528 | mutex_unlock(&dev->struct_mutex); | |
2529 | return ret; | |
2530 | } | |
2531 | ||
5816d648 BW |
2532 | /** |
2533 | * i915_gem_object_sync - sync an object to a ring. | |
2534 | * | |
2535 | * @obj: object which may be in use on another ring. | |
2536 | * @to: ring we wish to use the object on. May be NULL. | |
2537 | * | |
2538 | * This code is meant to abstract object synchronization with the GPU. | |
2539 | * Calling with NULL implies synchronizing the object with the CPU | |
2540 | * rather than a particular GPU ring. | |
2541 | * | |
2542 | * Returns 0 if successful, else propagates up the lower layer error. | |
2543 | */ | |
2911a35b BW |
2544 | int |
2545 | i915_gem_object_sync(struct drm_i915_gem_object *obj, | |
2546 | struct intel_ring_buffer *to) | |
2547 | { | |
2548 | struct intel_ring_buffer *from = obj->ring; | |
2549 | u32 seqno; | |
2550 | int ret, idx; | |
2551 | ||
2552 | if (from == NULL || to == from) | |
2553 | return 0; | |
2554 | ||
5816d648 | 2555 | if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev)) |
0201f1ec | 2556 | return i915_gem_object_wait_rendering(obj, false); |
2911a35b BW |
2557 | |
2558 | idx = intel_ring_sync_index(from, to); | |
2559 | ||
0201f1ec | 2560 | seqno = obj->last_read_seqno; |
2911a35b BW |
2561 | if (seqno <= from->sync_seqno[idx]) |
2562 | return 0; | |
2563 | ||
b4aca010 BW |
2564 | ret = i915_gem_check_olr(obj->ring, seqno); |
2565 | if (ret) | |
2566 | return ret; | |
2911a35b | 2567 | |
1500f7ea | 2568 | ret = to->sync_to(to, from, seqno); |
e3a5a225 | 2569 | if (!ret) |
7b01e260 MK |
2570 | /* We use last_read_seqno because sync_to() |
2571 | * might have just caused seqno wrap under | |
2572 | * the radar. | |
2573 | */ | |
2574 | from->sync_seqno[idx] = obj->last_read_seqno; | |
2911a35b | 2575 | |
e3a5a225 | 2576 | return ret; |
2911a35b BW |
2577 | } |
2578 | ||
b5ffc9bc CW |
2579 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) |
2580 | { | |
2581 | u32 old_write_domain, old_read_domains; | |
2582 | ||
b5ffc9bc CW |
2583 | /* Force a pagefault for domain tracking on next user access */ |
2584 | i915_gem_release_mmap(obj); | |
2585 | ||
b97c3d9c KP |
2586 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
2587 | return; | |
2588 | ||
97c809fd CW |
2589 | /* Wait for any direct GTT access to complete */ |
2590 | mb(); | |
2591 | ||
b5ffc9bc CW |
2592 | old_read_domains = obj->base.read_domains; |
2593 | old_write_domain = obj->base.write_domain; | |
2594 | ||
2595 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; | |
2596 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; | |
2597 | ||
2598 | trace_i915_gem_object_change_domain(obj, | |
2599 | old_read_domains, | |
2600 | old_write_domain); | |
2601 | } | |
2602 | ||
07fe0b12 | 2603 | int i915_vma_unbind(struct i915_vma *vma) |
673a394b | 2604 | { |
07fe0b12 | 2605 | struct drm_i915_gem_object *obj = vma->obj; |
7bddb01f | 2606 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
43e28f09 | 2607 | int ret; |
673a394b | 2608 | |
b93dab6e DV |
2609 | /* For now we only ever use 1 vma per object */ |
2610 | WARN_ON(!list_is_singular(&obj->vma_list)); | |
2611 | ||
07fe0b12 | 2612 | if (list_empty(&vma->vma_link)) |
673a394b EA |
2613 | return 0; |
2614 | ||
0ff501cb DV |
2615 | if (!drm_mm_node_allocated(&vma->node)) { |
2616 | i915_gem_vma_destroy(vma); | |
2617 | ||
2618 | return 0; | |
2619 | } | |
433544bd | 2620 | |
31d8d651 CW |
2621 | if (obj->pin_count) |
2622 | return -EBUSY; | |
673a394b | 2623 | |
c4670ad0 CW |
2624 | BUG_ON(obj->pages == NULL); |
2625 | ||
a8198eea | 2626 | ret = i915_gem_object_finish_gpu(obj); |
1488fc08 | 2627 | if (ret) |
a8198eea CW |
2628 | return ret; |
2629 | /* Continue on if we fail due to EIO, the GPU is hung so we | |
2630 | * should be safe and we need to cleanup or else we might | |
2631 | * cause memory corruption through use-after-free. | |
2632 | */ | |
2633 | ||
b5ffc9bc | 2634 | i915_gem_object_finish_gtt(obj); |
5323fd04 | 2635 | |
96b47b65 | 2636 | /* release the fence reg _after_ flushing */ |
d9e86c0e | 2637 | ret = i915_gem_object_put_fence(obj); |
1488fc08 | 2638 | if (ret) |
d9e86c0e | 2639 | return ret; |
96b47b65 | 2640 | |
07fe0b12 | 2641 | trace_i915_vma_unbind(vma); |
db53a302 | 2642 | |
74898d7e DV |
2643 | if (obj->has_global_gtt_mapping) |
2644 | i915_gem_gtt_unbind_object(obj); | |
7bddb01f DV |
2645 | if (obj->has_aliasing_ppgtt_mapping) { |
2646 | i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj); | |
2647 | obj->has_aliasing_ppgtt_mapping = 0; | |
2648 | } | |
74163907 | 2649 | i915_gem_gtt_finish_object(obj); |
401c29f6 | 2650 | i915_gem_object_unpin_pages(obj); |
7bddb01f | 2651 | |
ca191b13 | 2652 | list_del(&vma->mm_list); |
75e9e915 | 2653 | /* Avoid an unnecessary call to unbind on rebind. */ |
5cacaac7 BW |
2654 | if (i915_is_ggtt(vma->vm)) |
2655 | obj->map_and_fenceable = true; | |
673a394b | 2656 | |
2f633156 | 2657 | drm_mm_remove_node(&vma->node); |
433544bd | 2658 | |
2f633156 BW |
2659 | i915_gem_vma_destroy(vma); |
2660 | ||
2661 | /* Since the unbound list is global, only move to that list if | |
b93dab6e | 2662 | * no more VMAs exist. */ |
2f633156 BW |
2663 | if (list_empty(&obj->vma_list)) |
2664 | list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list); | |
673a394b | 2665 | |
88241785 | 2666 | return 0; |
54cf91dc CW |
2667 | } |
2668 | ||
07fe0b12 BW |
2669 | /** |
2670 | * Unbinds an object from the global GTT aperture. | |
2671 | */ | |
2672 | int | |
2673 | i915_gem_object_ggtt_unbind(struct drm_i915_gem_object *obj) | |
2674 | { | |
2675 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
2676 | struct i915_address_space *ggtt = &dev_priv->gtt.base; | |
2677 | ||
58e73e15 | 2678 | if (!i915_gem_obj_ggtt_bound(obj)) |
07fe0b12 BW |
2679 | return 0; |
2680 | ||
2681 | if (obj->pin_count) | |
2682 | return -EBUSY; | |
2683 | ||
2684 | BUG_ON(obj->pages == NULL); | |
2685 | ||
2686 | return i915_vma_unbind(i915_gem_obj_to_vma(obj, ggtt)); | |
2687 | } | |
2688 | ||
b2da9fe5 | 2689 | int i915_gpu_idle(struct drm_device *dev) |
4df2faf4 DV |
2690 | { |
2691 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 2692 | struct intel_ring_buffer *ring; |
1ec14ad3 | 2693 | int ret, i; |
4df2faf4 | 2694 | |
4df2faf4 | 2695 | /* Flush everything onto the inactive list. */ |
b4519513 | 2696 | for_each_ring(ring, dev_priv, i) { |
b6c7488d BW |
2697 | ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID); |
2698 | if (ret) | |
2699 | return ret; | |
2700 | ||
3e960501 | 2701 | ret = intel_ring_idle(ring); |
1ec14ad3 CW |
2702 | if (ret) |
2703 | return ret; | |
2704 | } | |
4df2faf4 | 2705 | |
8a1a49f9 | 2706 | return 0; |
4df2faf4 DV |
2707 | } |
2708 | ||
9ce079e4 CW |
2709 | static void i965_write_fence_reg(struct drm_device *dev, int reg, |
2710 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2711 | { |
de151cf6 | 2712 | drm_i915_private_t *dev_priv = dev->dev_private; |
56c844e5 ID |
2713 | int fence_reg; |
2714 | int fence_pitch_shift; | |
de151cf6 | 2715 | |
56c844e5 ID |
2716 | if (INTEL_INFO(dev)->gen >= 6) { |
2717 | fence_reg = FENCE_REG_SANDYBRIDGE_0; | |
2718 | fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT; | |
2719 | } else { | |
2720 | fence_reg = FENCE_REG_965_0; | |
2721 | fence_pitch_shift = I965_FENCE_PITCH_SHIFT; | |
2722 | } | |
2723 | ||
d18b9619 CW |
2724 | fence_reg += reg * 8; |
2725 | ||
2726 | /* To w/a incoherency with non-atomic 64-bit register updates, | |
2727 | * we split the 64-bit update into two 32-bit writes. In order | |
2728 | * for a partial fence not to be evaluated between writes, we | |
2729 | * precede the update with write to turn off the fence register, | |
2730 | * and only enable the fence as the last step. | |
2731 | * | |
2732 | * For extra levels of paranoia, we make sure each step lands | |
2733 | * before applying the next step. | |
2734 | */ | |
2735 | I915_WRITE(fence_reg, 0); | |
2736 | POSTING_READ(fence_reg); | |
2737 | ||
9ce079e4 | 2738 | if (obj) { |
f343c5f6 | 2739 | u32 size = i915_gem_obj_ggtt_size(obj); |
d18b9619 | 2740 | uint64_t val; |
de151cf6 | 2741 | |
f343c5f6 | 2742 | val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) & |
9ce079e4 | 2743 | 0xfffff000) << 32; |
f343c5f6 | 2744 | val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000; |
56c844e5 | 2745 | val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift; |
9ce079e4 CW |
2746 | if (obj->tiling_mode == I915_TILING_Y) |
2747 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; | |
2748 | val |= I965_FENCE_REG_VALID; | |
c6642782 | 2749 | |
d18b9619 CW |
2750 | I915_WRITE(fence_reg + 4, val >> 32); |
2751 | POSTING_READ(fence_reg + 4); | |
2752 | ||
2753 | I915_WRITE(fence_reg + 0, val); | |
2754 | POSTING_READ(fence_reg); | |
2755 | } else { | |
2756 | I915_WRITE(fence_reg + 4, 0); | |
2757 | POSTING_READ(fence_reg + 4); | |
2758 | } | |
de151cf6 JB |
2759 | } |
2760 | ||
9ce079e4 CW |
2761 | static void i915_write_fence_reg(struct drm_device *dev, int reg, |
2762 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2763 | { |
de151cf6 | 2764 | drm_i915_private_t *dev_priv = dev->dev_private; |
9ce079e4 | 2765 | u32 val; |
de151cf6 | 2766 | |
9ce079e4 | 2767 | if (obj) { |
f343c5f6 | 2768 | u32 size = i915_gem_obj_ggtt_size(obj); |
9ce079e4 CW |
2769 | int pitch_val; |
2770 | int tile_width; | |
c6642782 | 2771 | |
f343c5f6 | 2772 | WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) || |
9ce079e4 | 2773 | (size & -size) != size || |
f343c5f6 BW |
2774 | (i915_gem_obj_ggtt_offset(obj) & (size - 1)), |
2775 | "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n", | |
2776 | i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size); | |
c6642782 | 2777 | |
9ce079e4 CW |
2778 | if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)) |
2779 | tile_width = 128; | |
2780 | else | |
2781 | tile_width = 512; | |
2782 | ||
2783 | /* Note: pitch better be a power of two tile widths */ | |
2784 | pitch_val = obj->stride / tile_width; | |
2785 | pitch_val = ffs(pitch_val) - 1; | |
2786 | ||
f343c5f6 | 2787 | val = i915_gem_obj_ggtt_offset(obj); |
9ce079e4 CW |
2788 | if (obj->tiling_mode == I915_TILING_Y) |
2789 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
2790 | val |= I915_FENCE_SIZE_BITS(size); | |
2791 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
2792 | val |= I830_FENCE_REG_VALID; | |
2793 | } else | |
2794 | val = 0; | |
2795 | ||
2796 | if (reg < 8) | |
2797 | reg = FENCE_REG_830_0 + reg * 4; | |
2798 | else | |
2799 | reg = FENCE_REG_945_8 + (reg - 8) * 4; | |
2800 | ||
2801 | I915_WRITE(reg, val); | |
2802 | POSTING_READ(reg); | |
de151cf6 JB |
2803 | } |
2804 | ||
9ce079e4 CW |
2805 | static void i830_write_fence_reg(struct drm_device *dev, int reg, |
2806 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2807 | { |
de151cf6 | 2808 | drm_i915_private_t *dev_priv = dev->dev_private; |
de151cf6 | 2809 | uint32_t val; |
de151cf6 | 2810 | |
9ce079e4 | 2811 | if (obj) { |
f343c5f6 | 2812 | u32 size = i915_gem_obj_ggtt_size(obj); |
9ce079e4 | 2813 | uint32_t pitch_val; |
de151cf6 | 2814 | |
f343c5f6 | 2815 | WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) || |
9ce079e4 | 2816 | (size & -size) != size || |
f343c5f6 BW |
2817 | (i915_gem_obj_ggtt_offset(obj) & (size - 1)), |
2818 | "object 0x%08lx not 512K or pot-size 0x%08x aligned\n", | |
2819 | i915_gem_obj_ggtt_offset(obj), size); | |
e76a16de | 2820 | |
9ce079e4 CW |
2821 | pitch_val = obj->stride / 128; |
2822 | pitch_val = ffs(pitch_val) - 1; | |
de151cf6 | 2823 | |
f343c5f6 | 2824 | val = i915_gem_obj_ggtt_offset(obj); |
9ce079e4 CW |
2825 | if (obj->tiling_mode == I915_TILING_Y) |
2826 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
2827 | val |= I830_FENCE_SIZE_BITS(size); | |
2828 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
2829 | val |= I830_FENCE_REG_VALID; | |
2830 | } else | |
2831 | val = 0; | |
c6642782 | 2832 | |
9ce079e4 CW |
2833 | I915_WRITE(FENCE_REG_830_0 + reg * 4, val); |
2834 | POSTING_READ(FENCE_REG_830_0 + reg * 4); | |
2835 | } | |
2836 | ||
d0a57789 CW |
2837 | inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj) |
2838 | { | |
2839 | return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT; | |
2840 | } | |
2841 | ||
9ce079e4 CW |
2842 | static void i915_gem_write_fence(struct drm_device *dev, int reg, |
2843 | struct drm_i915_gem_object *obj) | |
2844 | { | |
d0a57789 CW |
2845 | struct drm_i915_private *dev_priv = dev->dev_private; |
2846 | ||
2847 | /* Ensure that all CPU reads are completed before installing a fence | |
2848 | * and all writes before removing the fence. | |
2849 | */ | |
2850 | if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj)) | |
2851 | mb(); | |
2852 | ||
94a335db DV |
2853 | WARN(obj && (!obj->stride || !obj->tiling_mode), |
2854 | "bogus fence setup with stride: 0x%x, tiling mode: %i\n", | |
2855 | obj->stride, obj->tiling_mode); | |
2856 | ||
9ce079e4 CW |
2857 | switch (INTEL_INFO(dev)->gen) { |
2858 | case 7: | |
56c844e5 | 2859 | case 6: |
9ce079e4 CW |
2860 | case 5: |
2861 | case 4: i965_write_fence_reg(dev, reg, obj); break; | |
2862 | case 3: i915_write_fence_reg(dev, reg, obj); break; | |
2863 | case 2: i830_write_fence_reg(dev, reg, obj); break; | |
7dbf9d6e | 2864 | default: BUG(); |
9ce079e4 | 2865 | } |
d0a57789 CW |
2866 | |
2867 | /* And similarly be paranoid that no direct access to this region | |
2868 | * is reordered to before the fence is installed. | |
2869 | */ | |
2870 | if (i915_gem_object_needs_mb(obj)) | |
2871 | mb(); | |
de151cf6 JB |
2872 | } |
2873 | ||
61050808 CW |
2874 | static inline int fence_number(struct drm_i915_private *dev_priv, |
2875 | struct drm_i915_fence_reg *fence) | |
2876 | { | |
2877 | return fence - dev_priv->fence_regs; | |
2878 | } | |
2879 | ||
2880 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, | |
2881 | struct drm_i915_fence_reg *fence, | |
2882 | bool enable) | |
2883 | { | |
2dc8aae0 | 2884 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
46a0b638 CW |
2885 | int reg = fence_number(dev_priv, fence); |
2886 | ||
2887 | i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL); | |
61050808 CW |
2888 | |
2889 | if (enable) { | |
46a0b638 | 2890 | obj->fence_reg = reg; |
61050808 CW |
2891 | fence->obj = obj; |
2892 | list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list); | |
2893 | } else { | |
2894 | obj->fence_reg = I915_FENCE_REG_NONE; | |
2895 | fence->obj = NULL; | |
2896 | list_del_init(&fence->lru_list); | |
2897 | } | |
94a335db | 2898 | obj->fence_dirty = false; |
61050808 CW |
2899 | } |
2900 | ||
d9e86c0e | 2901 | static int |
d0a57789 | 2902 | i915_gem_object_wait_fence(struct drm_i915_gem_object *obj) |
d9e86c0e | 2903 | { |
1c293ea3 | 2904 | if (obj->last_fenced_seqno) { |
86d5bc37 | 2905 | int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno); |
18991845 CW |
2906 | if (ret) |
2907 | return ret; | |
d9e86c0e CW |
2908 | |
2909 | obj->last_fenced_seqno = 0; | |
d9e86c0e CW |
2910 | } |
2911 | ||
86d5bc37 | 2912 | obj->fenced_gpu_access = false; |
d9e86c0e CW |
2913 | return 0; |
2914 | } | |
2915 | ||
2916 | int | |
2917 | i915_gem_object_put_fence(struct drm_i915_gem_object *obj) | |
2918 | { | |
61050808 | 2919 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
f9c513e9 | 2920 | struct drm_i915_fence_reg *fence; |
d9e86c0e CW |
2921 | int ret; |
2922 | ||
d0a57789 | 2923 | ret = i915_gem_object_wait_fence(obj); |
d9e86c0e CW |
2924 | if (ret) |
2925 | return ret; | |
2926 | ||
61050808 CW |
2927 | if (obj->fence_reg == I915_FENCE_REG_NONE) |
2928 | return 0; | |
d9e86c0e | 2929 | |
f9c513e9 CW |
2930 | fence = &dev_priv->fence_regs[obj->fence_reg]; |
2931 | ||
61050808 | 2932 | i915_gem_object_fence_lost(obj); |
f9c513e9 | 2933 | i915_gem_object_update_fence(obj, fence, false); |
d9e86c0e CW |
2934 | |
2935 | return 0; | |
2936 | } | |
2937 | ||
2938 | static struct drm_i915_fence_reg * | |
a360bb1a | 2939 | i915_find_fence_reg(struct drm_device *dev) |
ae3db24a | 2940 | { |
ae3db24a | 2941 | struct drm_i915_private *dev_priv = dev->dev_private; |
8fe301ad | 2942 | struct drm_i915_fence_reg *reg, *avail; |
d9e86c0e | 2943 | int i; |
ae3db24a DV |
2944 | |
2945 | /* First try to find a free reg */ | |
d9e86c0e | 2946 | avail = NULL; |
ae3db24a DV |
2947 | for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) { |
2948 | reg = &dev_priv->fence_regs[i]; | |
2949 | if (!reg->obj) | |
d9e86c0e | 2950 | return reg; |
ae3db24a | 2951 | |
1690e1eb | 2952 | if (!reg->pin_count) |
d9e86c0e | 2953 | avail = reg; |
ae3db24a DV |
2954 | } |
2955 | ||
d9e86c0e CW |
2956 | if (avail == NULL) |
2957 | return NULL; | |
ae3db24a DV |
2958 | |
2959 | /* None available, try to steal one or wait for a user to finish */ | |
d9e86c0e | 2960 | list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) { |
1690e1eb | 2961 | if (reg->pin_count) |
ae3db24a DV |
2962 | continue; |
2963 | ||
8fe301ad | 2964 | return reg; |
ae3db24a DV |
2965 | } |
2966 | ||
8fe301ad | 2967 | return NULL; |
ae3db24a DV |
2968 | } |
2969 | ||
de151cf6 | 2970 | /** |
9a5a53b3 | 2971 | * i915_gem_object_get_fence - set up fencing for an object |
de151cf6 JB |
2972 | * @obj: object to map through a fence reg |
2973 | * | |
2974 | * When mapping objects through the GTT, userspace wants to be able to write | |
2975 | * to them without having to worry about swizzling if the object is tiled. | |
de151cf6 JB |
2976 | * This function walks the fence regs looking for a free one for @obj, |
2977 | * stealing one if it can't find any. | |
2978 | * | |
2979 | * It then sets up the reg based on the object's properties: address, pitch | |
2980 | * and tiling format. | |
9a5a53b3 CW |
2981 | * |
2982 | * For an untiled surface, this removes any existing fence. | |
de151cf6 | 2983 | */ |
8c4b8c3f | 2984 | int |
06d98131 | 2985 | i915_gem_object_get_fence(struct drm_i915_gem_object *obj) |
de151cf6 | 2986 | { |
05394f39 | 2987 | struct drm_device *dev = obj->base.dev; |
79e53945 | 2988 | struct drm_i915_private *dev_priv = dev->dev_private; |
14415745 | 2989 | bool enable = obj->tiling_mode != I915_TILING_NONE; |
d9e86c0e | 2990 | struct drm_i915_fence_reg *reg; |
ae3db24a | 2991 | int ret; |
de151cf6 | 2992 | |
14415745 CW |
2993 | /* Have we updated the tiling parameters upon the object and so |
2994 | * will need to serialise the write to the associated fence register? | |
2995 | */ | |
5d82e3e6 | 2996 | if (obj->fence_dirty) { |
d0a57789 | 2997 | ret = i915_gem_object_wait_fence(obj); |
14415745 CW |
2998 | if (ret) |
2999 | return ret; | |
3000 | } | |
9a5a53b3 | 3001 | |
d9e86c0e | 3002 | /* Just update our place in the LRU if our fence is getting reused. */ |
05394f39 CW |
3003 | if (obj->fence_reg != I915_FENCE_REG_NONE) { |
3004 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
5d82e3e6 | 3005 | if (!obj->fence_dirty) { |
14415745 CW |
3006 | list_move_tail(®->lru_list, |
3007 | &dev_priv->mm.fence_list); | |
3008 | return 0; | |
3009 | } | |
3010 | } else if (enable) { | |
3011 | reg = i915_find_fence_reg(dev); | |
3012 | if (reg == NULL) | |
3013 | return -EDEADLK; | |
d9e86c0e | 3014 | |
14415745 CW |
3015 | if (reg->obj) { |
3016 | struct drm_i915_gem_object *old = reg->obj; | |
3017 | ||
d0a57789 | 3018 | ret = i915_gem_object_wait_fence(old); |
29c5a587 CW |
3019 | if (ret) |
3020 | return ret; | |
3021 | ||
14415745 | 3022 | i915_gem_object_fence_lost(old); |
29c5a587 | 3023 | } |
14415745 | 3024 | } else |
a09ba7fa | 3025 | return 0; |
a09ba7fa | 3026 | |
14415745 | 3027 | i915_gem_object_update_fence(obj, reg, enable); |
14415745 | 3028 | |
9ce079e4 | 3029 | return 0; |
de151cf6 JB |
3030 | } |
3031 | ||
42d6ab48 CW |
3032 | static bool i915_gem_valid_gtt_space(struct drm_device *dev, |
3033 | struct drm_mm_node *gtt_space, | |
3034 | unsigned long cache_level) | |
3035 | { | |
3036 | struct drm_mm_node *other; | |
3037 | ||
3038 | /* On non-LLC machines we have to be careful when putting differing | |
3039 | * types of snoopable memory together to avoid the prefetcher | |
4239ca77 | 3040 | * crossing memory domains and dying. |
42d6ab48 CW |
3041 | */ |
3042 | if (HAS_LLC(dev)) | |
3043 | return true; | |
3044 | ||
c6cfb325 | 3045 | if (!drm_mm_node_allocated(gtt_space)) |
42d6ab48 CW |
3046 | return true; |
3047 | ||
3048 | if (list_empty(>t_space->node_list)) | |
3049 | return true; | |
3050 | ||
3051 | other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list); | |
3052 | if (other->allocated && !other->hole_follows && other->color != cache_level) | |
3053 | return false; | |
3054 | ||
3055 | other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list); | |
3056 | if (other->allocated && !gtt_space->hole_follows && other->color != cache_level) | |
3057 | return false; | |
3058 | ||
3059 | return true; | |
3060 | } | |
3061 | ||
3062 | static void i915_gem_verify_gtt(struct drm_device *dev) | |
3063 | { | |
3064 | #if WATCH_GTT | |
3065 | struct drm_i915_private *dev_priv = dev->dev_private; | |
3066 | struct drm_i915_gem_object *obj; | |
3067 | int err = 0; | |
3068 | ||
35c20a60 | 3069 | list_for_each_entry(obj, &dev_priv->mm.gtt_list, global_list) { |
42d6ab48 CW |
3070 | if (obj->gtt_space == NULL) { |
3071 | printk(KERN_ERR "object found on GTT list with no space reserved\n"); | |
3072 | err++; | |
3073 | continue; | |
3074 | } | |
3075 | ||
3076 | if (obj->cache_level != obj->gtt_space->color) { | |
3077 | printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n", | |
f343c5f6 BW |
3078 | i915_gem_obj_ggtt_offset(obj), |
3079 | i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj), | |
42d6ab48 CW |
3080 | obj->cache_level, |
3081 | obj->gtt_space->color); | |
3082 | err++; | |
3083 | continue; | |
3084 | } | |
3085 | ||
3086 | if (!i915_gem_valid_gtt_space(dev, | |
3087 | obj->gtt_space, | |
3088 | obj->cache_level)) { | |
3089 | printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n", | |
f343c5f6 BW |
3090 | i915_gem_obj_ggtt_offset(obj), |
3091 | i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj), | |
42d6ab48 CW |
3092 | obj->cache_level); |
3093 | err++; | |
3094 | continue; | |
3095 | } | |
3096 | } | |
3097 | ||
3098 | WARN_ON(err); | |
3099 | #endif | |
3100 | } | |
3101 | ||
673a394b EA |
3102 | /** |
3103 | * Finds free space in the GTT aperture and binds the object there. | |
3104 | */ | |
3105 | static int | |
07fe0b12 BW |
3106 | i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, |
3107 | struct i915_address_space *vm, | |
3108 | unsigned alignment, | |
3109 | bool map_and_fenceable, | |
3110 | bool nonblocking) | |
673a394b | 3111 | { |
05394f39 | 3112 | struct drm_device *dev = obj->base.dev; |
673a394b | 3113 | drm_i915_private_t *dev_priv = dev->dev_private; |
5e783301 | 3114 | u32 size, fence_size, fence_alignment, unfenced_alignment; |
07fe0b12 BW |
3115 | size_t gtt_max = |
3116 | map_and_fenceable ? dev_priv->gtt.mappable_end : vm->total; | |
2f633156 | 3117 | struct i915_vma *vma; |
07f73f69 | 3118 | int ret; |
673a394b | 3119 | |
e28f8711 CW |
3120 | fence_size = i915_gem_get_gtt_size(dev, |
3121 | obj->base.size, | |
3122 | obj->tiling_mode); | |
3123 | fence_alignment = i915_gem_get_gtt_alignment(dev, | |
3124 | obj->base.size, | |
d865110c | 3125 | obj->tiling_mode, true); |
e28f8711 | 3126 | unfenced_alignment = |
d865110c | 3127 | i915_gem_get_gtt_alignment(dev, |
e28f8711 | 3128 | obj->base.size, |
d865110c | 3129 | obj->tiling_mode, false); |
a00b10c3 | 3130 | |
673a394b | 3131 | if (alignment == 0) |
5e783301 DV |
3132 | alignment = map_and_fenceable ? fence_alignment : |
3133 | unfenced_alignment; | |
75e9e915 | 3134 | if (map_and_fenceable && alignment & (fence_alignment - 1)) { |
673a394b EA |
3135 | DRM_ERROR("Invalid object alignment requested %u\n", alignment); |
3136 | return -EINVAL; | |
3137 | } | |
3138 | ||
05394f39 | 3139 | size = map_and_fenceable ? fence_size : obj->base.size; |
a00b10c3 | 3140 | |
654fc607 CW |
3141 | /* If the object is bigger than the entire aperture, reject it early |
3142 | * before evicting everything in a vain attempt to find space. | |
3143 | */ | |
0a9ae0d7 | 3144 | if (obj->base.size > gtt_max) { |
3765f304 | 3145 | DRM_ERROR("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%zu\n", |
a36689cb CW |
3146 | obj->base.size, |
3147 | map_and_fenceable ? "mappable" : "total", | |
0a9ae0d7 | 3148 | gtt_max); |
654fc607 CW |
3149 | return -E2BIG; |
3150 | } | |
3151 | ||
37e680a1 | 3152 | ret = i915_gem_object_get_pages(obj); |
6c085a72 CW |
3153 | if (ret) |
3154 | return ret; | |
3155 | ||
fbdda6fb CW |
3156 | i915_gem_object_pin_pages(obj); |
3157 | ||
07fe0b12 | 3158 | BUG_ON(!i915_is_ggtt(vm)); |
07fe0b12 | 3159 | |
accfef2e | 3160 | vma = i915_gem_obj_lookup_or_create_vma(obj, vm); |
db473b36 | 3161 | if (IS_ERR(vma)) { |
bc6bc15b DV |
3162 | ret = PTR_ERR(vma); |
3163 | goto err_unpin; | |
2f633156 BW |
3164 | } |
3165 | ||
accfef2e BW |
3166 | /* For now we only ever use 1 vma per object */ |
3167 | WARN_ON(!list_is_singular(&obj->vma_list)); | |
3168 | ||
0a9ae0d7 | 3169 | search_free: |
07fe0b12 | 3170 | ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node, |
0a9ae0d7 | 3171 | size, alignment, |
31e5d7c6 DH |
3172 | obj->cache_level, 0, gtt_max, |
3173 | DRM_MM_SEARCH_DEFAULT); | |
dc9dd7a2 | 3174 | if (ret) { |
f6cd1f15 | 3175 | ret = i915_gem_evict_something(dev, vm, size, alignment, |
42d6ab48 | 3176 | obj->cache_level, |
86a1ee26 CW |
3177 | map_and_fenceable, |
3178 | nonblocking); | |
dc9dd7a2 CW |
3179 | if (ret == 0) |
3180 | goto search_free; | |
9731129c | 3181 | |
bc6bc15b | 3182 | goto err_free_vma; |
673a394b | 3183 | } |
2f633156 | 3184 | if (WARN_ON(!i915_gem_valid_gtt_space(dev, &vma->node, |
c6cfb325 | 3185 | obj->cache_level))) { |
2f633156 | 3186 | ret = -EINVAL; |
bc6bc15b | 3187 | goto err_remove_node; |
673a394b EA |
3188 | } |
3189 | ||
74163907 | 3190 | ret = i915_gem_gtt_prepare_object(obj); |
2f633156 | 3191 | if (ret) |
bc6bc15b | 3192 | goto err_remove_node; |
673a394b | 3193 | |
35c20a60 | 3194 | list_move_tail(&obj->global_list, &dev_priv->mm.bound_list); |
ca191b13 | 3195 | list_add_tail(&vma->mm_list, &vm->inactive_list); |
bf1a1092 | 3196 | |
4bd561b3 BW |
3197 | if (i915_is_ggtt(vm)) { |
3198 | bool mappable, fenceable; | |
a00b10c3 | 3199 | |
49987099 DV |
3200 | fenceable = (vma->node.size == fence_size && |
3201 | (vma->node.start & (fence_alignment - 1)) == 0); | |
4bd561b3 | 3202 | |
49987099 DV |
3203 | mappable = (vma->node.start + obj->base.size <= |
3204 | dev_priv->gtt.mappable_end); | |
a00b10c3 | 3205 | |
5cacaac7 | 3206 | obj->map_and_fenceable = mappable && fenceable; |
4bd561b3 | 3207 | } |
75e9e915 | 3208 | |
7ace7ef2 | 3209 | WARN_ON(map_and_fenceable && !obj->map_and_fenceable); |
75e9e915 | 3210 | |
07fe0b12 | 3211 | trace_i915_vma_bind(vma, map_and_fenceable); |
42d6ab48 | 3212 | i915_gem_verify_gtt(dev); |
673a394b | 3213 | return 0; |
2f633156 | 3214 | |
bc6bc15b | 3215 | err_remove_node: |
6286ef9b | 3216 | drm_mm_remove_node(&vma->node); |
bc6bc15b | 3217 | err_free_vma: |
2f633156 | 3218 | i915_gem_vma_destroy(vma); |
bc6bc15b | 3219 | err_unpin: |
2f633156 | 3220 | i915_gem_object_unpin_pages(obj); |
2f633156 | 3221 | return ret; |
673a394b EA |
3222 | } |
3223 | ||
000433b6 | 3224 | bool |
2c22569b CW |
3225 | i915_gem_clflush_object(struct drm_i915_gem_object *obj, |
3226 | bool force) | |
673a394b | 3227 | { |
673a394b EA |
3228 | /* If we don't have a page list set up, then we're not pinned |
3229 | * to GPU, and we can ignore the cache flush because it'll happen | |
3230 | * again at bind time. | |
3231 | */ | |
05394f39 | 3232 | if (obj->pages == NULL) |
000433b6 | 3233 | return false; |
673a394b | 3234 | |
769ce464 ID |
3235 | /* |
3236 | * Stolen memory is always coherent with the GPU as it is explicitly | |
3237 | * marked as wc by the system, or the system is cache-coherent. | |
3238 | */ | |
3239 | if (obj->stolen) | |
000433b6 | 3240 | return false; |
769ce464 | 3241 | |
9c23f7fc CW |
3242 | /* If the GPU is snooping the contents of the CPU cache, |
3243 | * we do not need to manually clear the CPU cache lines. However, | |
3244 | * the caches are only snooped when the render cache is | |
3245 | * flushed/invalidated. As we always have to emit invalidations | |
3246 | * and flushes when moving into and out of the RENDER domain, correct | |
3247 | * snooping behaviour occurs naturally as the result of our domain | |
3248 | * tracking. | |
3249 | */ | |
2c22569b | 3250 | if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) |
000433b6 | 3251 | return false; |
9c23f7fc | 3252 | |
1c5d22f7 | 3253 | trace_i915_gem_object_clflush(obj); |
9da3da66 | 3254 | drm_clflush_sg(obj->pages); |
000433b6 CW |
3255 | |
3256 | return true; | |
e47c68e9 EA |
3257 | } |
3258 | ||
3259 | /** Flushes the GTT write domain for the object if it's dirty. */ | |
3260 | static void | |
05394f39 | 3261 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 3262 | { |
1c5d22f7 CW |
3263 | uint32_t old_write_domain; |
3264 | ||
05394f39 | 3265 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
e47c68e9 EA |
3266 | return; |
3267 | ||
63256ec5 | 3268 | /* No actual flushing is required for the GTT write domain. Writes |
e47c68e9 EA |
3269 | * to it immediately go to main memory as far as we know, so there's |
3270 | * no chipset flush. It also doesn't land in render cache. | |
63256ec5 CW |
3271 | * |
3272 | * However, we do have to enforce the order so that all writes through | |
3273 | * the GTT land before any writes to the device, such as updates to | |
3274 | * the GATT itself. | |
e47c68e9 | 3275 | */ |
63256ec5 CW |
3276 | wmb(); |
3277 | ||
05394f39 CW |
3278 | old_write_domain = obj->base.write_domain; |
3279 | obj->base.write_domain = 0; | |
1c5d22f7 CW |
3280 | |
3281 | trace_i915_gem_object_change_domain(obj, | |
05394f39 | 3282 | obj->base.read_domains, |
1c5d22f7 | 3283 | old_write_domain); |
e47c68e9 EA |
3284 | } |
3285 | ||
3286 | /** Flushes the CPU write domain for the object if it's dirty. */ | |
3287 | static void | |
2c22569b CW |
3288 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj, |
3289 | bool force) | |
e47c68e9 | 3290 | { |
1c5d22f7 | 3291 | uint32_t old_write_domain; |
e47c68e9 | 3292 | |
05394f39 | 3293 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
e47c68e9 EA |
3294 | return; |
3295 | ||
000433b6 CW |
3296 | if (i915_gem_clflush_object(obj, force)) |
3297 | i915_gem_chipset_flush(obj->base.dev); | |
3298 | ||
05394f39 CW |
3299 | old_write_domain = obj->base.write_domain; |
3300 | obj->base.write_domain = 0; | |
1c5d22f7 CW |
3301 | |
3302 | trace_i915_gem_object_change_domain(obj, | |
05394f39 | 3303 | obj->base.read_domains, |
1c5d22f7 | 3304 | old_write_domain); |
e47c68e9 EA |
3305 | } |
3306 | ||
2ef7eeaa EA |
3307 | /** |
3308 | * Moves a single object to the GTT read, and possibly write domain. | |
3309 | * | |
3310 | * This function returns when the move is complete, including waiting on | |
3311 | * flushes to occur. | |
3312 | */ | |
79e53945 | 3313 | int |
2021746e | 3314 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
2ef7eeaa | 3315 | { |
8325a09d | 3316 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
1c5d22f7 | 3317 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 | 3318 | int ret; |
2ef7eeaa | 3319 | |
02354392 | 3320 | /* Not valid to be called on unbound objects. */ |
9843877d | 3321 | if (!i915_gem_obj_bound_any(obj)) |
02354392 EA |
3322 | return -EINVAL; |
3323 | ||
8d7e3de1 CW |
3324 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
3325 | return 0; | |
3326 | ||
0201f1ec | 3327 | ret = i915_gem_object_wait_rendering(obj, !write); |
88241785 CW |
3328 | if (ret) |
3329 | return ret; | |
3330 | ||
2c22569b | 3331 | i915_gem_object_flush_cpu_write_domain(obj, false); |
1c5d22f7 | 3332 | |
d0a57789 CW |
3333 | /* Serialise direct access to this object with the barriers for |
3334 | * coherent writes from the GPU, by effectively invalidating the | |
3335 | * GTT domain upon first access. | |
3336 | */ | |
3337 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) | |
3338 | mb(); | |
3339 | ||
05394f39 CW |
3340 | old_write_domain = obj->base.write_domain; |
3341 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 3342 | |
e47c68e9 EA |
3343 | /* It should now be out of any other write domains, and we can update |
3344 | * the domain values for our changes. | |
3345 | */ | |
05394f39 CW |
3346 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
3347 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; | |
e47c68e9 | 3348 | if (write) { |
05394f39 CW |
3349 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
3350 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; | |
3351 | obj->dirty = 1; | |
2ef7eeaa EA |
3352 | } |
3353 | ||
1c5d22f7 CW |
3354 | trace_i915_gem_object_change_domain(obj, |
3355 | old_read_domains, | |
3356 | old_write_domain); | |
3357 | ||
8325a09d | 3358 | /* And bump the LRU for this access */ |
ca191b13 BW |
3359 | if (i915_gem_object_is_inactive(obj)) { |
3360 | struct i915_vma *vma = i915_gem_obj_to_vma(obj, | |
3361 | &dev_priv->gtt.base); | |
3362 | if (vma) | |
3363 | list_move_tail(&vma->mm_list, | |
3364 | &dev_priv->gtt.base.inactive_list); | |
3365 | ||
3366 | } | |
8325a09d | 3367 | |
e47c68e9 EA |
3368 | return 0; |
3369 | } | |
3370 | ||
e4ffd173 CW |
3371 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
3372 | enum i915_cache_level cache_level) | |
3373 | { | |
7bddb01f DV |
3374 | struct drm_device *dev = obj->base.dev; |
3375 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3089c6f2 | 3376 | struct i915_vma *vma; |
e4ffd173 CW |
3377 | int ret; |
3378 | ||
3379 | if (obj->cache_level == cache_level) | |
3380 | return 0; | |
3381 | ||
3382 | if (obj->pin_count) { | |
3383 | DRM_DEBUG("can not change the cache level of pinned objects\n"); | |
3384 | return -EBUSY; | |
3385 | } | |
3386 | ||
3089c6f2 BW |
3387 | list_for_each_entry(vma, &obj->vma_list, vma_link) { |
3388 | if (!i915_gem_valid_gtt_space(dev, &vma->node, cache_level)) { | |
07fe0b12 | 3389 | ret = i915_vma_unbind(vma); |
3089c6f2 BW |
3390 | if (ret) |
3391 | return ret; | |
3392 | ||
3393 | break; | |
3394 | } | |
42d6ab48 CW |
3395 | } |
3396 | ||
3089c6f2 | 3397 | if (i915_gem_obj_bound_any(obj)) { |
e4ffd173 CW |
3398 | ret = i915_gem_object_finish_gpu(obj); |
3399 | if (ret) | |
3400 | return ret; | |
3401 | ||
3402 | i915_gem_object_finish_gtt(obj); | |
3403 | ||
3404 | /* Before SandyBridge, you could not use tiling or fence | |
3405 | * registers with snooped memory, so relinquish any fences | |
3406 | * currently pointing to our region in the aperture. | |
3407 | */ | |
42d6ab48 | 3408 | if (INTEL_INFO(dev)->gen < 6) { |
e4ffd173 CW |
3409 | ret = i915_gem_object_put_fence(obj); |
3410 | if (ret) | |
3411 | return ret; | |
3412 | } | |
3413 | ||
74898d7e DV |
3414 | if (obj->has_global_gtt_mapping) |
3415 | i915_gem_gtt_bind_object(obj, cache_level); | |
7bddb01f DV |
3416 | if (obj->has_aliasing_ppgtt_mapping) |
3417 | i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt, | |
3418 | obj, cache_level); | |
e4ffd173 CW |
3419 | } |
3420 | ||
2c22569b CW |
3421 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
3422 | vma->node.color = cache_level; | |
3423 | obj->cache_level = cache_level; | |
3424 | ||
3425 | if (cpu_write_needs_clflush(obj)) { | |
e4ffd173 CW |
3426 | u32 old_read_domains, old_write_domain; |
3427 | ||
3428 | /* If we're coming from LLC cached, then we haven't | |
3429 | * actually been tracking whether the data is in the | |
3430 | * CPU cache or not, since we only allow one bit set | |
3431 | * in obj->write_domain and have been skipping the clflushes. | |
3432 | * Just set it to the CPU cache for now. | |
3433 | */ | |
3434 | WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU); | |
e4ffd173 CW |
3435 | |
3436 | old_read_domains = obj->base.read_domains; | |
3437 | old_write_domain = obj->base.write_domain; | |
3438 | ||
3439 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
3440 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
3441 | ||
3442 | trace_i915_gem_object_change_domain(obj, | |
3443 | old_read_domains, | |
3444 | old_write_domain); | |
3445 | } | |
3446 | ||
42d6ab48 | 3447 | i915_gem_verify_gtt(dev); |
e4ffd173 CW |
3448 | return 0; |
3449 | } | |
3450 | ||
199adf40 BW |
3451 | int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, |
3452 | struct drm_file *file) | |
e6994aee | 3453 | { |
199adf40 | 3454 | struct drm_i915_gem_caching *args = data; |
e6994aee CW |
3455 | struct drm_i915_gem_object *obj; |
3456 | int ret; | |
3457 | ||
3458 | ret = i915_mutex_lock_interruptible(dev); | |
3459 | if (ret) | |
3460 | return ret; | |
3461 | ||
3462 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
3463 | if (&obj->base == NULL) { | |
3464 | ret = -ENOENT; | |
3465 | goto unlock; | |
3466 | } | |
3467 | ||
651d794f CW |
3468 | switch (obj->cache_level) { |
3469 | case I915_CACHE_LLC: | |
3470 | case I915_CACHE_L3_LLC: | |
3471 | args->caching = I915_CACHING_CACHED; | |
3472 | break; | |
3473 | ||
4257d3ba CW |
3474 | case I915_CACHE_WT: |
3475 | args->caching = I915_CACHING_DISPLAY; | |
3476 | break; | |
3477 | ||
651d794f CW |
3478 | default: |
3479 | args->caching = I915_CACHING_NONE; | |
3480 | break; | |
3481 | } | |
e6994aee CW |
3482 | |
3483 | drm_gem_object_unreference(&obj->base); | |
3484 | unlock: | |
3485 | mutex_unlock(&dev->struct_mutex); | |
3486 | return ret; | |
3487 | } | |
3488 | ||
199adf40 BW |
3489 | int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data, |
3490 | struct drm_file *file) | |
e6994aee | 3491 | { |
199adf40 | 3492 | struct drm_i915_gem_caching *args = data; |
e6994aee CW |
3493 | struct drm_i915_gem_object *obj; |
3494 | enum i915_cache_level level; | |
3495 | int ret; | |
3496 | ||
199adf40 BW |
3497 | switch (args->caching) { |
3498 | case I915_CACHING_NONE: | |
e6994aee CW |
3499 | level = I915_CACHE_NONE; |
3500 | break; | |
199adf40 | 3501 | case I915_CACHING_CACHED: |
e6994aee CW |
3502 | level = I915_CACHE_LLC; |
3503 | break; | |
4257d3ba CW |
3504 | case I915_CACHING_DISPLAY: |
3505 | level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE; | |
3506 | break; | |
e6994aee CW |
3507 | default: |
3508 | return -EINVAL; | |
3509 | } | |
3510 | ||
3bc2913e BW |
3511 | ret = i915_mutex_lock_interruptible(dev); |
3512 | if (ret) | |
3513 | return ret; | |
3514 | ||
e6994aee CW |
3515 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3516 | if (&obj->base == NULL) { | |
3517 | ret = -ENOENT; | |
3518 | goto unlock; | |
3519 | } | |
3520 | ||
3521 | ret = i915_gem_object_set_cache_level(obj, level); | |
3522 | ||
3523 | drm_gem_object_unreference(&obj->base); | |
3524 | unlock: | |
3525 | mutex_unlock(&dev->struct_mutex); | |
3526 | return ret; | |
3527 | } | |
3528 | ||
cc98b413 CW |
3529 | static bool is_pin_display(struct drm_i915_gem_object *obj) |
3530 | { | |
3531 | /* There are 3 sources that pin objects: | |
3532 | * 1. The display engine (scanouts, sprites, cursors); | |
3533 | * 2. Reservations for execbuffer; | |
3534 | * 3. The user. | |
3535 | * | |
3536 | * We can ignore reservations as we hold the struct_mutex and | |
3537 | * are only called outside of the reservation path. The user | |
3538 | * can only increment pin_count once, and so if after | |
3539 | * subtracting the potential reference by the user, any pin_count | |
3540 | * remains, it must be due to another use by the display engine. | |
3541 | */ | |
3542 | return obj->pin_count - !!obj->user_pin_count; | |
3543 | } | |
3544 | ||
b9241ea3 | 3545 | /* |
2da3b9b9 CW |
3546 | * Prepare buffer for display plane (scanout, cursors, etc). |
3547 | * Can be called from an uninterruptible phase (modesetting) and allows | |
3548 | * any flushes to be pipelined (for pageflips). | |
b9241ea3 ZW |
3549 | */ |
3550 | int | |
2da3b9b9 CW |
3551 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
3552 | u32 alignment, | |
919926ae | 3553 | struct intel_ring_buffer *pipelined) |
b9241ea3 | 3554 | { |
2da3b9b9 | 3555 | u32 old_read_domains, old_write_domain; |
b9241ea3 ZW |
3556 | int ret; |
3557 | ||
0be73284 | 3558 | if (pipelined != obj->ring) { |
2911a35b BW |
3559 | ret = i915_gem_object_sync(obj, pipelined); |
3560 | if (ret) | |
b9241ea3 ZW |
3561 | return ret; |
3562 | } | |
3563 | ||
cc98b413 CW |
3564 | /* Mark the pin_display early so that we account for the |
3565 | * display coherency whilst setting up the cache domains. | |
3566 | */ | |
3567 | obj->pin_display = true; | |
3568 | ||
a7ef0640 EA |
3569 | /* The display engine is not coherent with the LLC cache on gen6. As |
3570 | * a result, we make sure that the pinning that is about to occur is | |
3571 | * done with uncached PTEs. This is lowest common denominator for all | |
3572 | * chipsets. | |
3573 | * | |
3574 | * However for gen6+, we could do better by using the GFDT bit instead | |
3575 | * of uncaching, which would allow us to flush all the LLC-cached data | |
3576 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. | |
3577 | */ | |
651d794f CW |
3578 | ret = i915_gem_object_set_cache_level(obj, |
3579 | HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE); | |
a7ef0640 | 3580 | if (ret) |
cc98b413 | 3581 | goto err_unpin_display; |
a7ef0640 | 3582 | |
2da3b9b9 CW |
3583 | /* As the user may map the buffer once pinned in the display plane |
3584 | * (e.g. libkms for the bootup splash), we have to ensure that we | |
3585 | * always use map_and_fenceable for all scanout buffers. | |
3586 | */ | |
c37e2204 | 3587 | ret = i915_gem_obj_ggtt_pin(obj, alignment, true, false); |
2da3b9b9 | 3588 | if (ret) |
cc98b413 | 3589 | goto err_unpin_display; |
2da3b9b9 | 3590 | |
2c22569b | 3591 | i915_gem_object_flush_cpu_write_domain(obj, true); |
b118c1e3 | 3592 | |
2da3b9b9 | 3593 | old_write_domain = obj->base.write_domain; |
05394f39 | 3594 | old_read_domains = obj->base.read_domains; |
2da3b9b9 CW |
3595 | |
3596 | /* It should now be out of any other write domains, and we can update | |
3597 | * the domain values for our changes. | |
3598 | */ | |
e5f1d962 | 3599 | obj->base.write_domain = 0; |
05394f39 | 3600 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
b9241ea3 ZW |
3601 | |
3602 | trace_i915_gem_object_change_domain(obj, | |
3603 | old_read_domains, | |
2da3b9b9 | 3604 | old_write_domain); |
b9241ea3 ZW |
3605 | |
3606 | return 0; | |
cc98b413 CW |
3607 | |
3608 | err_unpin_display: | |
3609 | obj->pin_display = is_pin_display(obj); | |
3610 | return ret; | |
3611 | } | |
3612 | ||
3613 | void | |
3614 | i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj) | |
3615 | { | |
3616 | i915_gem_object_unpin(obj); | |
3617 | obj->pin_display = is_pin_display(obj); | |
b9241ea3 ZW |
3618 | } |
3619 | ||
85345517 | 3620 | int |
a8198eea | 3621 | i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj) |
85345517 | 3622 | { |
88241785 CW |
3623 | int ret; |
3624 | ||
a8198eea | 3625 | if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0) |
85345517 CW |
3626 | return 0; |
3627 | ||
0201f1ec | 3628 | ret = i915_gem_object_wait_rendering(obj, false); |
c501ae7f CW |
3629 | if (ret) |
3630 | return ret; | |
3631 | ||
a8198eea CW |
3632 | /* Ensure that we invalidate the GPU's caches and TLBs. */ |
3633 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; | |
c501ae7f | 3634 | return 0; |
85345517 CW |
3635 | } |
3636 | ||
e47c68e9 EA |
3637 | /** |
3638 | * Moves a single object to the CPU read, and possibly write domain. | |
3639 | * | |
3640 | * This function returns when the move is complete, including waiting on | |
3641 | * flushes to occur. | |
3642 | */ | |
dabdfe02 | 3643 | int |
919926ae | 3644 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
e47c68e9 | 3645 | { |
1c5d22f7 | 3646 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 EA |
3647 | int ret; |
3648 | ||
8d7e3de1 CW |
3649 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3650 | return 0; | |
3651 | ||
0201f1ec | 3652 | ret = i915_gem_object_wait_rendering(obj, !write); |
88241785 CW |
3653 | if (ret) |
3654 | return ret; | |
3655 | ||
e47c68e9 | 3656 | i915_gem_object_flush_gtt_write_domain(obj); |
2ef7eeaa | 3657 | |
05394f39 CW |
3658 | old_write_domain = obj->base.write_domain; |
3659 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 3660 | |
e47c68e9 | 3661 | /* Flush the CPU cache if it's still invalid. */ |
05394f39 | 3662 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
2c22569b | 3663 | i915_gem_clflush_object(obj, false); |
2ef7eeaa | 3664 | |
05394f39 | 3665 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
2ef7eeaa EA |
3666 | } |
3667 | ||
3668 | /* It should now be out of any other write domains, and we can update | |
3669 | * the domain values for our changes. | |
3670 | */ | |
05394f39 | 3671 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
e47c68e9 EA |
3672 | |
3673 | /* If we're writing through the CPU, then the GPU read domains will | |
3674 | * need to be invalidated at next use. | |
3675 | */ | |
3676 | if (write) { | |
05394f39 CW |
3677 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3678 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
e47c68e9 | 3679 | } |
2ef7eeaa | 3680 | |
1c5d22f7 CW |
3681 | trace_i915_gem_object_change_domain(obj, |
3682 | old_read_domains, | |
3683 | old_write_domain); | |
3684 | ||
2ef7eeaa EA |
3685 | return 0; |
3686 | } | |
3687 | ||
673a394b EA |
3688 | /* Throttle our rendering by waiting until the ring has completed our requests |
3689 | * emitted over 20 msec ago. | |
3690 | * | |
b962442e EA |
3691 | * Note that if we were to use the current jiffies each time around the loop, |
3692 | * we wouldn't escape the function with any frames outstanding if the time to | |
3693 | * render a frame was over 20ms. | |
3694 | * | |
673a394b EA |
3695 | * This should get us reasonable parallelism between CPU and GPU but also |
3696 | * relatively low latency when blocking on a particular request to finish. | |
3697 | */ | |
40a5f0de | 3698 | static int |
f787a5f5 | 3699 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
40a5f0de | 3700 | { |
f787a5f5 CW |
3701 | struct drm_i915_private *dev_priv = dev->dev_private; |
3702 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
b962442e | 3703 | unsigned long recent_enough = jiffies - msecs_to_jiffies(20); |
f787a5f5 CW |
3704 | struct drm_i915_gem_request *request; |
3705 | struct intel_ring_buffer *ring = NULL; | |
f69061be | 3706 | unsigned reset_counter; |
f787a5f5 CW |
3707 | u32 seqno = 0; |
3708 | int ret; | |
93533c29 | 3709 | |
308887aa DV |
3710 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
3711 | if (ret) | |
3712 | return ret; | |
3713 | ||
3714 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, false); | |
3715 | if (ret) | |
3716 | return ret; | |
e110e8d6 | 3717 | |
1c25595f | 3718 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 3719 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { |
b962442e EA |
3720 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
3721 | break; | |
40a5f0de | 3722 | |
f787a5f5 CW |
3723 | ring = request->ring; |
3724 | seqno = request->seqno; | |
b962442e | 3725 | } |
f69061be | 3726 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
1c25595f | 3727 | spin_unlock(&file_priv->mm.lock); |
40a5f0de | 3728 | |
f787a5f5 CW |
3729 | if (seqno == 0) |
3730 | return 0; | |
2bc43b5c | 3731 | |
f69061be | 3732 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL); |
f787a5f5 CW |
3733 | if (ret == 0) |
3734 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); | |
40a5f0de EA |
3735 | |
3736 | return ret; | |
3737 | } | |
3738 | ||
673a394b | 3739 | int |
05394f39 | 3740 | i915_gem_object_pin(struct drm_i915_gem_object *obj, |
c37e2204 | 3741 | struct i915_address_space *vm, |
05394f39 | 3742 | uint32_t alignment, |
86a1ee26 CW |
3743 | bool map_and_fenceable, |
3744 | bool nonblocking) | |
673a394b | 3745 | { |
07fe0b12 | 3746 | struct i915_vma *vma; |
673a394b EA |
3747 | int ret; |
3748 | ||
7e81a42e CW |
3749 | if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT)) |
3750 | return -EBUSY; | |
ac0c6b5a | 3751 | |
07fe0b12 BW |
3752 | WARN_ON(map_and_fenceable && !i915_is_ggtt(vm)); |
3753 | ||
3754 | vma = i915_gem_obj_to_vma(obj, vm); | |
3755 | ||
3756 | if (vma) { | |
3757 | if ((alignment && | |
3758 | vma->node.start & (alignment - 1)) || | |
05394f39 CW |
3759 | (map_and_fenceable && !obj->map_and_fenceable)) { |
3760 | WARN(obj->pin_count, | |
ae7d49d8 | 3761 | "bo is already pinned with incorrect alignment:" |
f343c5f6 | 3762 | " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d," |
75e9e915 | 3763 | " obj->map_and_fenceable=%d\n", |
07fe0b12 | 3764 | i915_gem_obj_offset(obj, vm), alignment, |
75e9e915 | 3765 | map_and_fenceable, |
05394f39 | 3766 | obj->map_and_fenceable); |
07fe0b12 | 3767 | ret = i915_vma_unbind(vma); |
ac0c6b5a CW |
3768 | if (ret) |
3769 | return ret; | |
3770 | } | |
3771 | } | |
3772 | ||
07fe0b12 | 3773 | if (!i915_gem_obj_bound(obj, vm)) { |
8742267a CW |
3774 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
3775 | ||
07fe0b12 BW |
3776 | ret = i915_gem_object_bind_to_vm(obj, vm, alignment, |
3777 | map_and_fenceable, | |
3778 | nonblocking); | |
9731129c | 3779 | if (ret) |
673a394b | 3780 | return ret; |
8742267a CW |
3781 | |
3782 | if (!dev_priv->mm.aliasing_ppgtt) | |
3783 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
22c344e9 | 3784 | } |
76446cac | 3785 | |
74898d7e DV |
3786 | if (!obj->has_global_gtt_mapping && map_and_fenceable) |
3787 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
3788 | ||
1b50247a | 3789 | obj->pin_count++; |
6299f992 | 3790 | obj->pin_mappable |= map_and_fenceable; |
673a394b EA |
3791 | |
3792 | return 0; | |
3793 | } | |
3794 | ||
3795 | void | |
05394f39 | 3796 | i915_gem_object_unpin(struct drm_i915_gem_object *obj) |
673a394b | 3797 | { |
05394f39 | 3798 | BUG_ON(obj->pin_count == 0); |
9843877d | 3799 | BUG_ON(!i915_gem_obj_bound_any(obj)); |
673a394b | 3800 | |
1b50247a | 3801 | if (--obj->pin_count == 0) |
6299f992 | 3802 | obj->pin_mappable = false; |
673a394b EA |
3803 | } |
3804 | ||
3805 | int | |
3806 | i915_gem_pin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3807 | struct drm_file *file) |
673a394b EA |
3808 | { |
3809 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 3810 | struct drm_i915_gem_object *obj; |
673a394b EA |
3811 | int ret; |
3812 | ||
1d7cfea1 CW |
3813 | ret = i915_mutex_lock_interruptible(dev); |
3814 | if (ret) | |
3815 | return ret; | |
673a394b | 3816 | |
05394f39 | 3817 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3818 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3819 | ret = -ENOENT; |
3820 | goto unlock; | |
673a394b | 3821 | } |
673a394b | 3822 | |
05394f39 | 3823 | if (obj->madv != I915_MADV_WILLNEED) { |
bb6baf76 | 3824 | DRM_ERROR("Attempting to pin a purgeable buffer\n"); |
1d7cfea1 CW |
3825 | ret = -EINVAL; |
3826 | goto out; | |
3ef94daa CW |
3827 | } |
3828 | ||
05394f39 | 3829 | if (obj->pin_filp != NULL && obj->pin_filp != file) { |
79e53945 JB |
3830 | DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n", |
3831 | args->handle); | |
1d7cfea1 CW |
3832 | ret = -EINVAL; |
3833 | goto out; | |
79e53945 JB |
3834 | } |
3835 | ||
93be8788 | 3836 | if (obj->user_pin_count == 0) { |
c37e2204 | 3837 | ret = i915_gem_obj_ggtt_pin(obj, args->alignment, true, false); |
1d7cfea1 CW |
3838 | if (ret) |
3839 | goto out; | |
673a394b EA |
3840 | } |
3841 | ||
93be8788 CW |
3842 | obj->user_pin_count++; |
3843 | obj->pin_filp = file; | |
3844 | ||
f343c5f6 | 3845 | args->offset = i915_gem_obj_ggtt_offset(obj); |
1d7cfea1 | 3846 | out: |
05394f39 | 3847 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3848 | unlock: |
673a394b | 3849 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3850 | return ret; |
673a394b EA |
3851 | } |
3852 | ||
3853 | int | |
3854 | i915_gem_unpin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3855 | struct drm_file *file) |
673a394b EA |
3856 | { |
3857 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 3858 | struct drm_i915_gem_object *obj; |
76c1dec1 | 3859 | int ret; |
673a394b | 3860 | |
1d7cfea1 CW |
3861 | ret = i915_mutex_lock_interruptible(dev); |
3862 | if (ret) | |
3863 | return ret; | |
673a394b | 3864 | |
05394f39 | 3865 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3866 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3867 | ret = -ENOENT; |
3868 | goto unlock; | |
673a394b | 3869 | } |
76c1dec1 | 3870 | |
05394f39 | 3871 | if (obj->pin_filp != file) { |
79e53945 JB |
3872 | DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n", |
3873 | args->handle); | |
1d7cfea1 CW |
3874 | ret = -EINVAL; |
3875 | goto out; | |
79e53945 | 3876 | } |
05394f39 CW |
3877 | obj->user_pin_count--; |
3878 | if (obj->user_pin_count == 0) { | |
3879 | obj->pin_filp = NULL; | |
79e53945 JB |
3880 | i915_gem_object_unpin(obj); |
3881 | } | |
673a394b | 3882 | |
1d7cfea1 | 3883 | out: |
05394f39 | 3884 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3885 | unlock: |
673a394b | 3886 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3887 | return ret; |
673a394b EA |
3888 | } |
3889 | ||
3890 | int | |
3891 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3892 | struct drm_file *file) |
673a394b EA |
3893 | { |
3894 | struct drm_i915_gem_busy *args = data; | |
05394f39 | 3895 | struct drm_i915_gem_object *obj; |
30dbf0c0 CW |
3896 | int ret; |
3897 | ||
76c1dec1 | 3898 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 3899 | if (ret) |
76c1dec1 | 3900 | return ret; |
673a394b | 3901 | |
05394f39 | 3902 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3903 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3904 | ret = -ENOENT; |
3905 | goto unlock; | |
673a394b | 3906 | } |
d1b851fc | 3907 | |
0be555b6 CW |
3908 | /* Count all active objects as busy, even if they are currently not used |
3909 | * by the gpu. Users of this interface expect objects to eventually | |
3910 | * become non-busy without any further actions, therefore emit any | |
3911 | * necessary flushes here. | |
c4de0a5d | 3912 | */ |
30dfebf3 | 3913 | ret = i915_gem_object_flush_active(obj); |
0be555b6 | 3914 | |
30dfebf3 | 3915 | args->busy = obj->active; |
e9808edd CW |
3916 | if (obj->ring) { |
3917 | BUILD_BUG_ON(I915_NUM_RINGS > 16); | |
3918 | args->busy |= intel_ring_flag(obj->ring) << 16; | |
3919 | } | |
673a394b | 3920 | |
05394f39 | 3921 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3922 | unlock: |
673a394b | 3923 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3924 | return ret; |
673a394b EA |
3925 | } |
3926 | ||
3927 | int | |
3928 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, | |
3929 | struct drm_file *file_priv) | |
3930 | { | |
0206e353 | 3931 | return i915_gem_ring_throttle(dev, file_priv); |
673a394b EA |
3932 | } |
3933 | ||
3ef94daa CW |
3934 | int |
3935 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, | |
3936 | struct drm_file *file_priv) | |
3937 | { | |
3938 | struct drm_i915_gem_madvise *args = data; | |
05394f39 | 3939 | struct drm_i915_gem_object *obj; |
76c1dec1 | 3940 | int ret; |
3ef94daa CW |
3941 | |
3942 | switch (args->madv) { | |
3943 | case I915_MADV_DONTNEED: | |
3944 | case I915_MADV_WILLNEED: | |
3945 | break; | |
3946 | default: | |
3947 | return -EINVAL; | |
3948 | } | |
3949 | ||
1d7cfea1 CW |
3950 | ret = i915_mutex_lock_interruptible(dev); |
3951 | if (ret) | |
3952 | return ret; | |
3953 | ||
05394f39 | 3954 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); |
c8725226 | 3955 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3956 | ret = -ENOENT; |
3957 | goto unlock; | |
3ef94daa | 3958 | } |
3ef94daa | 3959 | |
05394f39 | 3960 | if (obj->pin_count) { |
1d7cfea1 CW |
3961 | ret = -EINVAL; |
3962 | goto out; | |
3ef94daa CW |
3963 | } |
3964 | ||
05394f39 CW |
3965 | if (obj->madv != __I915_MADV_PURGED) |
3966 | obj->madv = args->madv; | |
3ef94daa | 3967 | |
6c085a72 CW |
3968 | /* if the object is no longer attached, discard its backing storage */ |
3969 | if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL) | |
2d7ef395 CW |
3970 | i915_gem_object_truncate(obj); |
3971 | ||
05394f39 | 3972 | args->retained = obj->madv != __I915_MADV_PURGED; |
bb6baf76 | 3973 | |
1d7cfea1 | 3974 | out: |
05394f39 | 3975 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3976 | unlock: |
3ef94daa | 3977 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3978 | return ret; |
3ef94daa CW |
3979 | } |
3980 | ||
37e680a1 CW |
3981 | void i915_gem_object_init(struct drm_i915_gem_object *obj, |
3982 | const struct drm_i915_gem_object_ops *ops) | |
0327d6ba | 3983 | { |
35c20a60 | 3984 | INIT_LIST_HEAD(&obj->global_list); |
0327d6ba | 3985 | INIT_LIST_HEAD(&obj->ring_list); |
b25cb2f8 | 3986 | INIT_LIST_HEAD(&obj->obj_exec_link); |
2f633156 | 3987 | INIT_LIST_HEAD(&obj->vma_list); |
0327d6ba | 3988 | |
37e680a1 CW |
3989 | obj->ops = ops; |
3990 | ||
0327d6ba CW |
3991 | obj->fence_reg = I915_FENCE_REG_NONE; |
3992 | obj->madv = I915_MADV_WILLNEED; | |
3993 | /* Avoid an unnecessary call to unbind on the first bind. */ | |
3994 | obj->map_and_fenceable = true; | |
3995 | ||
3996 | i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size); | |
3997 | } | |
3998 | ||
37e680a1 CW |
3999 | static const struct drm_i915_gem_object_ops i915_gem_object_ops = { |
4000 | .get_pages = i915_gem_object_get_pages_gtt, | |
4001 | .put_pages = i915_gem_object_put_pages_gtt, | |
4002 | }; | |
4003 | ||
05394f39 CW |
4004 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, |
4005 | size_t size) | |
ac52bc56 | 4006 | { |
c397b908 | 4007 | struct drm_i915_gem_object *obj; |
5949eac4 | 4008 | struct address_space *mapping; |
1a240d4d | 4009 | gfp_t mask; |
ac52bc56 | 4010 | |
42dcedd4 | 4011 | obj = i915_gem_object_alloc(dev); |
c397b908 DV |
4012 | if (obj == NULL) |
4013 | return NULL; | |
673a394b | 4014 | |
c397b908 | 4015 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { |
42dcedd4 | 4016 | i915_gem_object_free(obj); |
c397b908 DV |
4017 | return NULL; |
4018 | } | |
673a394b | 4019 | |
bed1ea95 CW |
4020 | mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; |
4021 | if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) { | |
4022 | /* 965gm cannot relocate objects above 4GiB. */ | |
4023 | mask &= ~__GFP_HIGHMEM; | |
4024 | mask |= __GFP_DMA32; | |
4025 | } | |
4026 | ||
496ad9aa | 4027 | mapping = file_inode(obj->base.filp)->i_mapping; |
bed1ea95 | 4028 | mapping_set_gfp_mask(mapping, mask); |
5949eac4 | 4029 | |
37e680a1 | 4030 | i915_gem_object_init(obj, &i915_gem_object_ops); |
73aa808f | 4031 | |
c397b908 DV |
4032 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
4033 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
673a394b | 4034 | |
3d29b842 ED |
4035 | if (HAS_LLC(dev)) { |
4036 | /* On some devices, we can have the GPU use the LLC (the CPU | |
a1871112 EA |
4037 | * cache) for about a 10% performance improvement |
4038 | * compared to uncached. Graphics requests other than | |
4039 | * display scanout are coherent with the CPU in | |
4040 | * accessing this cache. This means in this mode we | |
4041 | * don't need to clflush on the CPU side, and on the | |
4042 | * GPU side we only need to flush internal caches to | |
4043 | * get data visible to the CPU. | |
4044 | * | |
4045 | * However, we maintain the display planes as UC, and so | |
4046 | * need to rebind when first used as such. | |
4047 | */ | |
4048 | obj->cache_level = I915_CACHE_LLC; | |
4049 | } else | |
4050 | obj->cache_level = I915_CACHE_NONE; | |
4051 | ||
d861e338 DV |
4052 | trace_i915_gem_object_create(obj); |
4053 | ||
05394f39 | 4054 | return obj; |
c397b908 DV |
4055 | } |
4056 | ||
4057 | int i915_gem_init_object(struct drm_gem_object *obj) | |
4058 | { | |
4059 | BUG(); | |
de151cf6 | 4060 | |
673a394b EA |
4061 | return 0; |
4062 | } | |
4063 | ||
1488fc08 | 4064 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
673a394b | 4065 | { |
1488fc08 | 4066 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); |
05394f39 | 4067 | struct drm_device *dev = obj->base.dev; |
be72615b | 4068 | drm_i915_private_t *dev_priv = dev->dev_private; |
07fe0b12 | 4069 | struct i915_vma *vma, *next; |
673a394b | 4070 | |
26e12f89 CW |
4071 | trace_i915_gem_object_destroy(obj); |
4072 | ||
1488fc08 CW |
4073 | if (obj->phys_obj) |
4074 | i915_gem_detach_phys_object(dev, obj); | |
4075 | ||
4076 | obj->pin_count = 0; | |
07fe0b12 BW |
4077 | /* NB: 0 or 1 elements */ |
4078 | WARN_ON(!list_empty(&obj->vma_list) && | |
4079 | !list_is_singular(&obj->vma_list)); | |
4080 | list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) { | |
4081 | int ret = i915_vma_unbind(vma); | |
4082 | if (WARN_ON(ret == -ERESTARTSYS)) { | |
4083 | bool was_interruptible; | |
1488fc08 | 4084 | |
07fe0b12 BW |
4085 | was_interruptible = dev_priv->mm.interruptible; |
4086 | dev_priv->mm.interruptible = false; | |
1488fc08 | 4087 | |
07fe0b12 | 4088 | WARN_ON(i915_vma_unbind(vma)); |
1488fc08 | 4089 | |
07fe0b12 BW |
4090 | dev_priv->mm.interruptible = was_interruptible; |
4091 | } | |
1488fc08 CW |
4092 | } |
4093 | ||
1d64ae71 BW |
4094 | /* Stolen objects don't hold a ref, but do hold pin count. Fix that up |
4095 | * before progressing. */ | |
4096 | if (obj->stolen) | |
4097 | i915_gem_object_unpin_pages(obj); | |
4098 | ||
401c29f6 BW |
4099 | if (WARN_ON(obj->pages_pin_count)) |
4100 | obj->pages_pin_count = 0; | |
37e680a1 | 4101 | i915_gem_object_put_pages(obj); |
d8cb5086 | 4102 | i915_gem_object_free_mmap_offset(obj); |
0104fdbb | 4103 | i915_gem_object_release_stolen(obj); |
de151cf6 | 4104 | |
9da3da66 CW |
4105 | BUG_ON(obj->pages); |
4106 | ||
2f745ad3 CW |
4107 | if (obj->base.import_attach) |
4108 | drm_prime_gem_destroy(&obj->base, NULL); | |
de151cf6 | 4109 | |
05394f39 CW |
4110 | drm_gem_object_release(&obj->base); |
4111 | i915_gem_info_remove_obj(dev_priv, obj->base.size); | |
c397b908 | 4112 | |
05394f39 | 4113 | kfree(obj->bit_17); |
42dcedd4 | 4114 | i915_gem_object_free(obj); |
673a394b EA |
4115 | } |
4116 | ||
e656a6cb | 4117 | struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj, |
2f633156 | 4118 | struct i915_address_space *vm) |
e656a6cb DV |
4119 | { |
4120 | struct i915_vma *vma; | |
4121 | list_for_each_entry(vma, &obj->vma_list, vma_link) | |
4122 | if (vma->vm == vm) | |
4123 | return vma; | |
4124 | ||
4125 | return NULL; | |
4126 | } | |
4127 | ||
4128 | static struct i915_vma *__i915_gem_vma_create(struct drm_i915_gem_object *obj, | |
4129 | struct i915_address_space *vm) | |
2f633156 BW |
4130 | { |
4131 | struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL); | |
4132 | if (vma == NULL) | |
4133 | return ERR_PTR(-ENOMEM); | |
4134 | ||
4135 | INIT_LIST_HEAD(&vma->vma_link); | |
ca191b13 | 4136 | INIT_LIST_HEAD(&vma->mm_list); |
82a55ad1 | 4137 | INIT_LIST_HEAD(&vma->exec_list); |
2f633156 BW |
4138 | vma->vm = vm; |
4139 | vma->obj = obj; | |
4140 | ||
8b9c2b94 BW |
4141 | /* Keep GGTT vmas first to make debug easier */ |
4142 | if (i915_is_ggtt(vm)) | |
4143 | list_add(&vma->vma_link, &obj->vma_list); | |
4144 | else | |
4145 | list_add_tail(&vma->vma_link, &obj->vma_list); | |
4146 | ||
2f633156 BW |
4147 | return vma; |
4148 | } | |
4149 | ||
e656a6cb DV |
4150 | struct i915_vma * |
4151 | i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj, | |
4152 | struct i915_address_space *vm) | |
4153 | { | |
4154 | struct i915_vma *vma; | |
4155 | ||
4156 | vma = i915_gem_obj_to_vma(obj, vm); | |
4157 | if (!vma) | |
4158 | vma = __i915_gem_vma_create(obj, vm); | |
4159 | ||
4160 | return vma; | |
4161 | } | |
4162 | ||
2f633156 BW |
4163 | void i915_gem_vma_destroy(struct i915_vma *vma) |
4164 | { | |
4165 | WARN_ON(vma->node.allocated); | |
aaa05667 CW |
4166 | |
4167 | /* Keep the vma as a placeholder in the execbuffer reservation lists */ | |
4168 | if (!list_empty(&vma->exec_list)) | |
4169 | return; | |
4170 | ||
b93dab6e DV |
4171 | list_del(&vma->vma_link); |
4172 | ||
2f633156 BW |
4173 | kfree(vma); |
4174 | } | |
4175 | ||
29105ccc CW |
4176 | int |
4177 | i915_gem_idle(struct drm_device *dev) | |
4178 | { | |
4179 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4180 | int ret; | |
28dfe52a | 4181 | |
db1b76ca | 4182 | if (dev_priv->ums.mm_suspended) { |
29105ccc CW |
4183 | mutex_unlock(&dev->struct_mutex); |
4184 | return 0; | |
28dfe52a EA |
4185 | } |
4186 | ||
b2da9fe5 | 4187 | ret = i915_gpu_idle(dev); |
6dbe2772 KP |
4188 | if (ret) { |
4189 | mutex_unlock(&dev->struct_mutex); | |
673a394b | 4190 | return ret; |
6dbe2772 | 4191 | } |
b2da9fe5 | 4192 | i915_gem_retire_requests(dev); |
673a394b | 4193 | |
29105ccc | 4194 | /* Under UMS, be paranoid and evict. */ |
a39d7efc | 4195 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
6c085a72 | 4196 | i915_gem_evict_everything(dev); |
29105ccc | 4197 | |
99584db3 | 4198 | del_timer_sync(&dev_priv->gpu_error.hangcheck_timer); |
29105ccc CW |
4199 | |
4200 | i915_kernel_lost_context(dev); | |
6dbe2772 | 4201 | i915_gem_cleanup_ringbuffer(dev); |
29105ccc | 4202 | |
29105ccc CW |
4203 | /* Cancel the retire work handler, which should be idle now. */ |
4204 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); | |
4205 | ||
673a394b EA |
4206 | return 0; |
4207 | } | |
4208 | ||
b9524a1e BW |
4209 | void i915_gem_l3_remap(struct drm_device *dev) |
4210 | { | |
4211 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4212 | u32 misccpctl; | |
4213 | int i; | |
4214 | ||
eb32e458 | 4215 | if (!HAS_L3_GPU_CACHE(dev)) |
b9524a1e BW |
4216 | return; |
4217 | ||
a4da4fa4 | 4218 | if (!dev_priv->l3_parity.remap_info) |
b9524a1e BW |
4219 | return; |
4220 | ||
4221 | misccpctl = I915_READ(GEN7_MISCCPCTL); | |
4222 | I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE); | |
4223 | POSTING_READ(GEN7_MISCCPCTL); | |
4224 | ||
4225 | for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) { | |
4226 | u32 remap = I915_READ(GEN7_L3LOG_BASE + i); | |
a4da4fa4 | 4227 | if (remap && remap != dev_priv->l3_parity.remap_info[i/4]) |
b9524a1e BW |
4228 | DRM_DEBUG("0x%x was already programmed to %x\n", |
4229 | GEN7_L3LOG_BASE + i, remap); | |
a4da4fa4 | 4230 | if (remap && !dev_priv->l3_parity.remap_info[i/4]) |
b9524a1e | 4231 | DRM_DEBUG_DRIVER("Clearing remapped register\n"); |
a4da4fa4 | 4232 | I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]); |
b9524a1e BW |
4233 | } |
4234 | ||
4235 | /* Make sure all the writes land before disabling dop clock gating */ | |
4236 | POSTING_READ(GEN7_L3LOG_BASE); | |
4237 | ||
4238 | I915_WRITE(GEN7_MISCCPCTL, misccpctl); | |
4239 | } | |
4240 | ||
f691e2f4 DV |
4241 | void i915_gem_init_swizzling(struct drm_device *dev) |
4242 | { | |
4243 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4244 | ||
11782b02 | 4245 | if (INTEL_INFO(dev)->gen < 5 || |
f691e2f4 DV |
4246 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
4247 | return; | |
4248 | ||
4249 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | | |
4250 | DISP_TILE_SURFACE_SWIZZLING); | |
4251 | ||
11782b02 DV |
4252 | if (IS_GEN5(dev)) |
4253 | return; | |
4254 | ||
f691e2f4 DV |
4255 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
4256 | if (IS_GEN6(dev)) | |
6b26c86d | 4257 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); |
8782e26c | 4258 | else if (IS_GEN7(dev)) |
6b26c86d | 4259 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); |
8782e26c BW |
4260 | else |
4261 | BUG(); | |
f691e2f4 | 4262 | } |
e21af88d | 4263 | |
67b1b571 CW |
4264 | static bool |
4265 | intel_enable_blt(struct drm_device *dev) | |
4266 | { | |
4267 | if (!HAS_BLT(dev)) | |
4268 | return false; | |
4269 | ||
4270 | /* The blitter was dysfunctional on early prototypes */ | |
4271 | if (IS_GEN6(dev) && dev->pdev->revision < 8) { | |
4272 | DRM_INFO("BLT not supported on this pre-production hardware;" | |
4273 | " graphics performance will be degraded.\n"); | |
4274 | return false; | |
4275 | } | |
4276 | ||
4277 | return true; | |
4278 | } | |
4279 | ||
4fc7c971 | 4280 | static int i915_gem_init_rings(struct drm_device *dev) |
8187a2b7 | 4281 | { |
4fc7c971 | 4282 | struct drm_i915_private *dev_priv = dev->dev_private; |
8187a2b7 | 4283 | int ret; |
68f95ba9 | 4284 | |
5c1143bb | 4285 | ret = intel_init_render_ring_buffer(dev); |
68f95ba9 | 4286 | if (ret) |
b6913e4b | 4287 | return ret; |
68f95ba9 CW |
4288 | |
4289 | if (HAS_BSD(dev)) { | |
5c1143bb | 4290 | ret = intel_init_bsd_ring_buffer(dev); |
68f95ba9 CW |
4291 | if (ret) |
4292 | goto cleanup_render_ring; | |
d1b851fc | 4293 | } |
68f95ba9 | 4294 | |
67b1b571 | 4295 | if (intel_enable_blt(dev)) { |
549f7365 CW |
4296 | ret = intel_init_blt_ring_buffer(dev); |
4297 | if (ret) | |
4298 | goto cleanup_bsd_ring; | |
4299 | } | |
4300 | ||
9a8a2213 BW |
4301 | if (HAS_VEBOX(dev)) { |
4302 | ret = intel_init_vebox_ring_buffer(dev); | |
4303 | if (ret) | |
4304 | goto cleanup_blt_ring; | |
4305 | } | |
4306 | ||
4307 | ||
99433931 | 4308 | ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000)); |
4fc7c971 | 4309 | if (ret) |
9a8a2213 | 4310 | goto cleanup_vebox_ring; |
4fc7c971 BW |
4311 | |
4312 | return 0; | |
4313 | ||
9a8a2213 BW |
4314 | cleanup_vebox_ring: |
4315 | intel_cleanup_ring_buffer(&dev_priv->ring[VECS]); | |
4fc7c971 BW |
4316 | cleanup_blt_ring: |
4317 | intel_cleanup_ring_buffer(&dev_priv->ring[BCS]); | |
4318 | cleanup_bsd_ring: | |
4319 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); | |
4320 | cleanup_render_ring: | |
4321 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); | |
4322 | ||
4323 | return ret; | |
4324 | } | |
4325 | ||
4326 | int | |
4327 | i915_gem_init_hw(struct drm_device *dev) | |
4328 | { | |
4329 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4330 | int ret; | |
4331 | ||
4332 | if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt()) | |
4333 | return -EIO; | |
4334 | ||
59124506 | 4335 | if (dev_priv->ellc_size) |
05e21cc4 | 4336 | I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf)); |
4fc7c971 | 4337 | |
9435373e RV |
4338 | if (IS_HSW_GT3(dev)) |
4339 | I915_WRITE(MI_PREDICATE_RESULT_2, LOWER_SLICE_ENABLED); | |
4340 | else | |
4341 | I915_WRITE(MI_PREDICATE_RESULT_2, LOWER_SLICE_DISABLED); | |
4342 | ||
88a2b2a3 BW |
4343 | if (HAS_PCH_NOP(dev)) { |
4344 | u32 temp = I915_READ(GEN7_MSG_CTL); | |
4345 | temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK); | |
4346 | I915_WRITE(GEN7_MSG_CTL, temp); | |
4347 | } | |
4348 | ||
4fc7c971 BW |
4349 | i915_gem_l3_remap(dev); |
4350 | ||
4351 | i915_gem_init_swizzling(dev); | |
4352 | ||
4353 | ret = i915_gem_init_rings(dev); | |
99433931 MK |
4354 | if (ret) |
4355 | return ret; | |
4356 | ||
254f965c BW |
4357 | /* |
4358 | * XXX: There was some w/a described somewhere suggesting loading | |
4359 | * contexts before PPGTT. | |
4360 | */ | |
4361 | i915_gem_context_init(dev); | |
b7c36d25 BW |
4362 | if (dev_priv->mm.aliasing_ppgtt) { |
4363 | ret = dev_priv->mm.aliasing_ppgtt->enable(dev); | |
4364 | if (ret) { | |
4365 | i915_gem_cleanup_aliasing_ppgtt(dev); | |
4366 | DRM_INFO("PPGTT enable failed. This is not fatal, but unexpected\n"); | |
4367 | } | |
4368 | } | |
e21af88d | 4369 | |
68f95ba9 | 4370 | return 0; |
8187a2b7 ZN |
4371 | } |
4372 | ||
1070a42b CW |
4373 | int i915_gem_init(struct drm_device *dev) |
4374 | { | |
4375 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1070a42b CW |
4376 | int ret; |
4377 | ||
1070a42b | 4378 | mutex_lock(&dev->struct_mutex); |
d62b4892 JB |
4379 | |
4380 | if (IS_VALLEYVIEW(dev)) { | |
4381 | /* VLVA0 (potential hack), BIOS isn't actually waking us */ | |
4382 | I915_WRITE(VLV_GTLC_WAKE_CTRL, 1); | |
4383 | if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) & 1) == 1, 10)) | |
4384 | DRM_DEBUG_DRIVER("allow wake ack timed out\n"); | |
4385 | } | |
4386 | ||
d7e5008f | 4387 | i915_gem_init_global_gtt(dev); |
d62b4892 | 4388 | |
1070a42b CW |
4389 | ret = i915_gem_init_hw(dev); |
4390 | mutex_unlock(&dev->struct_mutex); | |
4391 | if (ret) { | |
4392 | i915_gem_cleanup_aliasing_ppgtt(dev); | |
4393 | return ret; | |
4394 | } | |
4395 | ||
53ca26ca DV |
4396 | /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */ |
4397 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) | |
4398 | dev_priv->dri1.allow_batchbuffer = 1; | |
1070a42b CW |
4399 | return 0; |
4400 | } | |
4401 | ||
8187a2b7 ZN |
4402 | void |
4403 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) | |
4404 | { | |
4405 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 4406 | struct intel_ring_buffer *ring; |
1ec14ad3 | 4407 | int i; |
8187a2b7 | 4408 | |
b4519513 CW |
4409 | for_each_ring(ring, dev_priv, i) |
4410 | intel_cleanup_ring_buffer(ring); | |
8187a2b7 ZN |
4411 | } |
4412 | ||
673a394b EA |
4413 | int |
4414 | i915_gem_entervt_ioctl(struct drm_device *dev, void *data, | |
4415 | struct drm_file *file_priv) | |
4416 | { | |
db1b76ca | 4417 | struct drm_i915_private *dev_priv = dev->dev_private; |
b4519513 | 4418 | int ret; |
673a394b | 4419 | |
79e53945 JB |
4420 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4421 | return 0; | |
4422 | ||
1f83fee0 | 4423 | if (i915_reset_in_progress(&dev_priv->gpu_error)) { |
673a394b | 4424 | DRM_ERROR("Reenabling wedged hardware, good luck\n"); |
1f83fee0 | 4425 | atomic_set(&dev_priv->gpu_error.reset_counter, 0); |
673a394b EA |
4426 | } |
4427 | ||
673a394b | 4428 | mutex_lock(&dev->struct_mutex); |
db1b76ca | 4429 | dev_priv->ums.mm_suspended = 0; |
9bb2d6f9 | 4430 | |
f691e2f4 | 4431 | ret = i915_gem_init_hw(dev); |
d816f6ac WF |
4432 | if (ret != 0) { |
4433 | mutex_unlock(&dev->struct_mutex); | |
9bb2d6f9 | 4434 | return ret; |
d816f6ac | 4435 | } |
9bb2d6f9 | 4436 | |
5cef07e1 | 4437 | BUG_ON(!list_empty(&dev_priv->gtt.base.active_list)); |
673a394b | 4438 | mutex_unlock(&dev->struct_mutex); |
dbb19d30 | 4439 | |
5f35308b CW |
4440 | ret = drm_irq_install(dev); |
4441 | if (ret) | |
4442 | goto cleanup_ringbuffer; | |
dbb19d30 | 4443 | |
673a394b | 4444 | return 0; |
5f35308b CW |
4445 | |
4446 | cleanup_ringbuffer: | |
4447 | mutex_lock(&dev->struct_mutex); | |
4448 | i915_gem_cleanup_ringbuffer(dev); | |
db1b76ca | 4449 | dev_priv->ums.mm_suspended = 1; |
5f35308b CW |
4450 | mutex_unlock(&dev->struct_mutex); |
4451 | ||
4452 | return ret; | |
673a394b EA |
4453 | } |
4454 | ||
4455 | int | |
4456 | i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, | |
4457 | struct drm_file *file_priv) | |
4458 | { | |
db1b76ca DV |
4459 | struct drm_i915_private *dev_priv = dev->dev_private; |
4460 | int ret; | |
4461 | ||
79e53945 JB |
4462 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4463 | return 0; | |
4464 | ||
dbb19d30 | 4465 | drm_irq_uninstall(dev); |
db1b76ca DV |
4466 | |
4467 | mutex_lock(&dev->struct_mutex); | |
4468 | ret = i915_gem_idle(dev); | |
4469 | ||
4470 | /* Hack! Don't let anybody do execbuf while we don't control the chip. | |
4471 | * We need to replace this with a semaphore, or something. | |
4472 | * And not confound ums.mm_suspended! | |
4473 | */ | |
4474 | if (ret != 0) | |
4475 | dev_priv->ums.mm_suspended = 1; | |
4476 | mutex_unlock(&dev->struct_mutex); | |
4477 | ||
4478 | return ret; | |
673a394b EA |
4479 | } |
4480 | ||
4481 | void | |
4482 | i915_gem_lastclose(struct drm_device *dev) | |
4483 | { | |
4484 | int ret; | |
673a394b | 4485 | |
e806b495 EA |
4486 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4487 | return; | |
4488 | ||
db1b76ca | 4489 | mutex_lock(&dev->struct_mutex); |
6dbe2772 KP |
4490 | ret = i915_gem_idle(dev); |
4491 | if (ret) | |
4492 | DRM_ERROR("failed to idle hardware: %d\n", ret); | |
db1b76ca | 4493 | mutex_unlock(&dev->struct_mutex); |
673a394b EA |
4494 | } |
4495 | ||
64193406 CW |
4496 | static void |
4497 | init_ring_lists(struct intel_ring_buffer *ring) | |
4498 | { | |
4499 | INIT_LIST_HEAD(&ring->active_list); | |
4500 | INIT_LIST_HEAD(&ring->request_list); | |
64193406 CW |
4501 | } |
4502 | ||
fc8c067e BW |
4503 | static void i915_init_vm(struct drm_i915_private *dev_priv, |
4504 | struct i915_address_space *vm) | |
4505 | { | |
4506 | vm->dev = dev_priv->dev; | |
4507 | INIT_LIST_HEAD(&vm->active_list); | |
4508 | INIT_LIST_HEAD(&vm->inactive_list); | |
4509 | INIT_LIST_HEAD(&vm->global_link); | |
4510 | list_add(&vm->global_link, &dev_priv->vm_list); | |
4511 | } | |
4512 | ||
673a394b EA |
4513 | void |
4514 | i915_gem_load(struct drm_device *dev) | |
4515 | { | |
4516 | drm_i915_private_t *dev_priv = dev->dev_private; | |
42dcedd4 CW |
4517 | int i; |
4518 | ||
4519 | dev_priv->slab = | |
4520 | kmem_cache_create("i915_gem_object", | |
4521 | sizeof(struct drm_i915_gem_object), 0, | |
4522 | SLAB_HWCACHE_ALIGN, | |
4523 | NULL); | |
673a394b | 4524 | |
fc8c067e BW |
4525 | INIT_LIST_HEAD(&dev_priv->vm_list); |
4526 | i915_init_vm(dev_priv, &dev_priv->gtt.base); | |
4527 | ||
6c085a72 CW |
4528 | INIT_LIST_HEAD(&dev_priv->mm.unbound_list); |
4529 | INIT_LIST_HEAD(&dev_priv->mm.bound_list); | |
a09ba7fa | 4530 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
1ec14ad3 CW |
4531 | for (i = 0; i < I915_NUM_RINGS; i++) |
4532 | init_ring_lists(&dev_priv->ring[i]); | |
4b9de737 | 4533 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) |
007cc8ac | 4534 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); |
673a394b EA |
4535 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, |
4536 | i915_gem_retire_work_handler); | |
1f83fee0 | 4537 | init_waitqueue_head(&dev_priv->gpu_error.reset_queue); |
31169714 | 4538 | |
94400120 DA |
4539 | /* On GEN3 we really need to make sure the ARB C3 LP bit is set */ |
4540 | if (IS_GEN3(dev)) { | |
50743298 DV |
4541 | I915_WRITE(MI_ARB_STATE, |
4542 | _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE)); | |
94400120 DA |
4543 | } |
4544 | ||
72bfa19c CW |
4545 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; |
4546 | ||
de151cf6 | 4547 | /* Old X drivers will take 0-2 for front, back, depth buffers */ |
b397c836 EA |
4548 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
4549 | dev_priv->fence_reg_start = 3; | |
de151cf6 | 4550 | |
42b5aeab VS |
4551 | if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev)) |
4552 | dev_priv->num_fence_regs = 32; | |
4553 | else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) | |
de151cf6 JB |
4554 | dev_priv->num_fence_regs = 16; |
4555 | else | |
4556 | dev_priv->num_fence_regs = 8; | |
4557 | ||
b5aa8a0f | 4558 | /* Initialize fence registers to zero */ |
19b2dbde CW |
4559 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
4560 | i915_gem_restore_fences(dev); | |
10ed13e4 | 4561 | |
673a394b | 4562 | i915_gem_detect_bit_6_swizzle(dev); |
6b95a207 | 4563 | init_waitqueue_head(&dev_priv->pending_flip_queue); |
17250b71 | 4564 | |
ce453d81 CW |
4565 | dev_priv->mm.interruptible = true; |
4566 | ||
17250b71 CW |
4567 | dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink; |
4568 | dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS; | |
4569 | register_shrinker(&dev_priv->mm.inactive_shrinker); | |
673a394b | 4570 | } |
71acb5eb DA |
4571 | |
4572 | /* | |
4573 | * Create a physically contiguous memory object for this object | |
4574 | * e.g. for cursor + overlay regs | |
4575 | */ | |
995b6762 CW |
4576 | static int i915_gem_init_phys_object(struct drm_device *dev, |
4577 | int id, int size, int align) | |
71acb5eb DA |
4578 | { |
4579 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4580 | struct drm_i915_gem_phys_object *phys_obj; | |
4581 | int ret; | |
4582 | ||
4583 | if (dev_priv->mm.phys_objs[id - 1] || !size) | |
4584 | return 0; | |
4585 | ||
9a298b2a | 4586 | phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL); |
71acb5eb DA |
4587 | if (!phys_obj) |
4588 | return -ENOMEM; | |
4589 | ||
4590 | phys_obj->id = id; | |
4591 | ||
6eeefaf3 | 4592 | phys_obj->handle = drm_pci_alloc(dev, size, align); |
71acb5eb DA |
4593 | if (!phys_obj->handle) { |
4594 | ret = -ENOMEM; | |
4595 | goto kfree_obj; | |
4596 | } | |
4597 | #ifdef CONFIG_X86 | |
4598 | set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
4599 | #endif | |
4600 | ||
4601 | dev_priv->mm.phys_objs[id - 1] = phys_obj; | |
4602 | ||
4603 | return 0; | |
4604 | kfree_obj: | |
9a298b2a | 4605 | kfree(phys_obj); |
71acb5eb DA |
4606 | return ret; |
4607 | } | |
4608 | ||
995b6762 | 4609 | static void i915_gem_free_phys_object(struct drm_device *dev, int id) |
71acb5eb DA |
4610 | { |
4611 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4612 | struct drm_i915_gem_phys_object *phys_obj; | |
4613 | ||
4614 | if (!dev_priv->mm.phys_objs[id - 1]) | |
4615 | return; | |
4616 | ||
4617 | phys_obj = dev_priv->mm.phys_objs[id - 1]; | |
4618 | if (phys_obj->cur_obj) { | |
4619 | i915_gem_detach_phys_object(dev, phys_obj->cur_obj); | |
4620 | } | |
4621 | ||
4622 | #ifdef CONFIG_X86 | |
4623 | set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
4624 | #endif | |
4625 | drm_pci_free(dev, phys_obj->handle); | |
4626 | kfree(phys_obj); | |
4627 | dev_priv->mm.phys_objs[id - 1] = NULL; | |
4628 | } | |
4629 | ||
4630 | void i915_gem_free_all_phys_object(struct drm_device *dev) | |
4631 | { | |
4632 | int i; | |
4633 | ||
260883c8 | 4634 | for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++) |
71acb5eb DA |
4635 | i915_gem_free_phys_object(dev, i); |
4636 | } | |
4637 | ||
4638 | void i915_gem_detach_phys_object(struct drm_device *dev, | |
05394f39 | 4639 | struct drm_i915_gem_object *obj) |
71acb5eb | 4640 | { |
496ad9aa | 4641 | struct address_space *mapping = file_inode(obj->base.filp)->i_mapping; |
e5281ccd | 4642 | char *vaddr; |
71acb5eb | 4643 | int i; |
71acb5eb DA |
4644 | int page_count; |
4645 | ||
05394f39 | 4646 | if (!obj->phys_obj) |
71acb5eb | 4647 | return; |
05394f39 | 4648 | vaddr = obj->phys_obj->handle->vaddr; |
71acb5eb | 4649 | |
05394f39 | 4650 | page_count = obj->base.size / PAGE_SIZE; |
71acb5eb | 4651 | for (i = 0; i < page_count; i++) { |
5949eac4 | 4652 | struct page *page = shmem_read_mapping_page(mapping, i); |
e5281ccd CW |
4653 | if (!IS_ERR(page)) { |
4654 | char *dst = kmap_atomic(page); | |
4655 | memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE); | |
4656 | kunmap_atomic(dst); | |
4657 | ||
4658 | drm_clflush_pages(&page, 1); | |
4659 | ||
4660 | set_page_dirty(page); | |
4661 | mark_page_accessed(page); | |
4662 | page_cache_release(page); | |
4663 | } | |
71acb5eb | 4664 | } |
e76e9aeb | 4665 | i915_gem_chipset_flush(dev); |
d78b47b9 | 4666 | |
05394f39 CW |
4667 | obj->phys_obj->cur_obj = NULL; |
4668 | obj->phys_obj = NULL; | |
71acb5eb DA |
4669 | } |
4670 | ||
4671 | int | |
4672 | i915_gem_attach_phys_object(struct drm_device *dev, | |
05394f39 | 4673 | struct drm_i915_gem_object *obj, |
6eeefaf3 CW |
4674 | int id, |
4675 | int align) | |
71acb5eb | 4676 | { |
496ad9aa | 4677 | struct address_space *mapping = file_inode(obj->base.filp)->i_mapping; |
71acb5eb | 4678 | drm_i915_private_t *dev_priv = dev->dev_private; |
71acb5eb DA |
4679 | int ret = 0; |
4680 | int page_count; | |
4681 | int i; | |
4682 | ||
4683 | if (id > I915_MAX_PHYS_OBJECT) | |
4684 | return -EINVAL; | |
4685 | ||
05394f39 CW |
4686 | if (obj->phys_obj) { |
4687 | if (obj->phys_obj->id == id) | |
71acb5eb DA |
4688 | return 0; |
4689 | i915_gem_detach_phys_object(dev, obj); | |
4690 | } | |
4691 | ||
71acb5eb DA |
4692 | /* create a new object */ |
4693 | if (!dev_priv->mm.phys_objs[id - 1]) { | |
4694 | ret = i915_gem_init_phys_object(dev, id, | |
05394f39 | 4695 | obj->base.size, align); |
71acb5eb | 4696 | if (ret) { |
05394f39 CW |
4697 | DRM_ERROR("failed to init phys object %d size: %zu\n", |
4698 | id, obj->base.size); | |
e5281ccd | 4699 | return ret; |
71acb5eb DA |
4700 | } |
4701 | } | |
4702 | ||
4703 | /* bind to the object */ | |
05394f39 CW |
4704 | obj->phys_obj = dev_priv->mm.phys_objs[id - 1]; |
4705 | obj->phys_obj->cur_obj = obj; | |
71acb5eb | 4706 | |
05394f39 | 4707 | page_count = obj->base.size / PAGE_SIZE; |
71acb5eb DA |
4708 | |
4709 | for (i = 0; i < page_count; i++) { | |
e5281ccd CW |
4710 | struct page *page; |
4711 | char *dst, *src; | |
4712 | ||
5949eac4 | 4713 | page = shmem_read_mapping_page(mapping, i); |
e5281ccd CW |
4714 | if (IS_ERR(page)) |
4715 | return PTR_ERR(page); | |
71acb5eb | 4716 | |
ff75b9bc | 4717 | src = kmap_atomic(page); |
05394f39 | 4718 | dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE); |
71acb5eb | 4719 | memcpy(dst, src, PAGE_SIZE); |
3e4d3af5 | 4720 | kunmap_atomic(src); |
71acb5eb | 4721 | |
e5281ccd CW |
4722 | mark_page_accessed(page); |
4723 | page_cache_release(page); | |
4724 | } | |
d78b47b9 | 4725 | |
71acb5eb | 4726 | return 0; |
71acb5eb DA |
4727 | } |
4728 | ||
4729 | static int | |
05394f39 CW |
4730 | i915_gem_phys_pwrite(struct drm_device *dev, |
4731 | struct drm_i915_gem_object *obj, | |
71acb5eb DA |
4732 | struct drm_i915_gem_pwrite *args, |
4733 | struct drm_file *file_priv) | |
4734 | { | |
05394f39 | 4735 | void *vaddr = obj->phys_obj->handle->vaddr + args->offset; |
2bb4629a | 4736 | char __user *user_data = to_user_ptr(args->data_ptr); |
71acb5eb | 4737 | |
b47b30cc CW |
4738 | if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) { |
4739 | unsigned long unwritten; | |
4740 | ||
4741 | /* The physical object once assigned is fixed for the lifetime | |
4742 | * of the obj, so we can safely drop the lock and continue | |
4743 | * to access vaddr. | |
4744 | */ | |
4745 | mutex_unlock(&dev->struct_mutex); | |
4746 | unwritten = copy_from_user(vaddr, user_data, args->size); | |
4747 | mutex_lock(&dev->struct_mutex); | |
4748 | if (unwritten) | |
4749 | return -EFAULT; | |
4750 | } | |
71acb5eb | 4751 | |
e76e9aeb | 4752 | i915_gem_chipset_flush(dev); |
71acb5eb DA |
4753 | return 0; |
4754 | } | |
b962442e | 4755 | |
f787a5f5 | 4756 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
b962442e | 4757 | { |
f787a5f5 | 4758 | struct drm_i915_file_private *file_priv = file->driver_priv; |
b962442e EA |
4759 | |
4760 | /* Clean up our request list when the client is going away, so that | |
4761 | * later retire_requests won't dereference our soon-to-be-gone | |
4762 | * file_priv. | |
4763 | */ | |
1c25595f | 4764 | spin_lock(&file_priv->mm.lock); |
f787a5f5 CW |
4765 | while (!list_empty(&file_priv->mm.request_list)) { |
4766 | struct drm_i915_gem_request *request; | |
4767 | ||
4768 | request = list_first_entry(&file_priv->mm.request_list, | |
4769 | struct drm_i915_gem_request, | |
4770 | client_list); | |
4771 | list_del(&request->client_list); | |
4772 | request->file_priv = NULL; | |
4773 | } | |
1c25595f | 4774 | spin_unlock(&file_priv->mm.lock); |
b962442e | 4775 | } |
31169714 | 4776 | |
5774506f CW |
4777 | static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task) |
4778 | { | |
4779 | if (!mutex_is_locked(mutex)) | |
4780 | return false; | |
4781 | ||
4782 | #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES) | |
4783 | return mutex->owner == task; | |
4784 | #else | |
4785 | /* Since UP may be pre-empted, we cannot assume that we own the lock */ | |
4786 | return false; | |
4787 | #endif | |
4788 | } | |
4789 | ||
31169714 | 4790 | static int |
1495f230 | 4791 | i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc) |
31169714 | 4792 | { |
17250b71 CW |
4793 | struct drm_i915_private *dev_priv = |
4794 | container_of(shrinker, | |
4795 | struct drm_i915_private, | |
4796 | mm.inactive_shrinker); | |
4797 | struct drm_device *dev = dev_priv->dev; | |
6c085a72 | 4798 | struct drm_i915_gem_object *obj; |
1495f230 | 4799 | int nr_to_scan = sc->nr_to_scan; |
5774506f | 4800 | bool unlock = true; |
17250b71 CW |
4801 | int cnt; |
4802 | ||
5774506f CW |
4803 | if (!mutex_trylock(&dev->struct_mutex)) { |
4804 | if (!mutex_is_locked_by(&dev->struct_mutex, current)) | |
4805 | return 0; | |
4806 | ||
677feac2 DV |
4807 | if (dev_priv->mm.shrinker_no_lock_stealing) |
4808 | return 0; | |
4809 | ||
5774506f CW |
4810 | unlock = false; |
4811 | } | |
31169714 | 4812 | |
6c085a72 CW |
4813 | if (nr_to_scan) { |
4814 | nr_to_scan -= i915_gem_purge(dev_priv, nr_to_scan); | |
93927ca5 DV |
4815 | if (nr_to_scan > 0) |
4816 | nr_to_scan -= __i915_gem_shrink(dev_priv, nr_to_scan, | |
4817 | false); | |
6c085a72 CW |
4818 | if (nr_to_scan > 0) |
4819 | i915_gem_shrink_all(dev_priv); | |
31169714 CW |
4820 | } |
4821 | ||
17250b71 | 4822 | cnt = 0; |
35c20a60 | 4823 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) |
a5570178 CW |
4824 | if (obj->pages_pin_count == 0) |
4825 | cnt += obj->base.size >> PAGE_SHIFT; | |
fcb4a578 BW |
4826 | |
4827 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { | |
4828 | if (obj->active) | |
4829 | continue; | |
4830 | ||
a5570178 | 4831 | if (obj->pin_count == 0 && obj->pages_pin_count == 0) |
6c085a72 | 4832 | cnt += obj->base.size >> PAGE_SHIFT; |
fcb4a578 | 4833 | } |
17250b71 | 4834 | |
5774506f CW |
4835 | if (unlock) |
4836 | mutex_unlock(&dev->struct_mutex); | |
6c085a72 | 4837 | return cnt; |
31169714 | 4838 | } |
a70a3148 BW |
4839 | |
4840 | /* All the new VM stuff */ | |
4841 | unsigned long i915_gem_obj_offset(struct drm_i915_gem_object *o, | |
4842 | struct i915_address_space *vm) | |
4843 | { | |
4844 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; | |
4845 | struct i915_vma *vma; | |
4846 | ||
4847 | if (vm == &dev_priv->mm.aliasing_ppgtt->base) | |
4848 | vm = &dev_priv->gtt.base; | |
4849 | ||
4850 | BUG_ON(list_empty(&o->vma_list)); | |
4851 | list_for_each_entry(vma, &o->vma_list, vma_link) { | |
4852 | if (vma->vm == vm) | |
4853 | return vma->node.start; | |
4854 | ||
4855 | } | |
4856 | return -1; | |
4857 | } | |
4858 | ||
4859 | bool i915_gem_obj_bound(struct drm_i915_gem_object *o, | |
4860 | struct i915_address_space *vm) | |
4861 | { | |
4862 | struct i915_vma *vma; | |
4863 | ||
4864 | list_for_each_entry(vma, &o->vma_list, vma_link) | |
8b9c2b94 | 4865 | if (vma->vm == vm && drm_mm_node_allocated(&vma->node)) |
a70a3148 BW |
4866 | return true; |
4867 | ||
4868 | return false; | |
4869 | } | |
4870 | ||
4871 | bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o) | |
4872 | { | |
4873 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; | |
4874 | struct i915_address_space *vm; | |
4875 | ||
4876 | list_for_each_entry(vm, &dev_priv->vm_list, global_link) | |
4877 | if (i915_gem_obj_bound(o, vm)) | |
4878 | return true; | |
4879 | ||
4880 | return false; | |
4881 | } | |
4882 | ||
4883 | unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o, | |
4884 | struct i915_address_space *vm) | |
4885 | { | |
4886 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; | |
4887 | struct i915_vma *vma; | |
4888 | ||
4889 | if (vm == &dev_priv->mm.aliasing_ppgtt->base) | |
4890 | vm = &dev_priv->gtt.base; | |
4891 | ||
4892 | BUG_ON(list_empty(&o->vma_list)); | |
4893 | ||
4894 | list_for_each_entry(vma, &o->vma_list, vma_link) | |
4895 | if (vma->vm == vm) | |
4896 | return vma->node.size; | |
4897 | ||
4898 | return 0; | |
4899 | } |