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; | |
967 | if (seqno == ring->outstanding_lazy_request) | |
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 | 1697 | { |
57094f82 | 1698 | struct list_head still_bound_list; |
6c085a72 CW |
1699 | struct drm_i915_gem_object *obj, *next; |
1700 | long count = 0; | |
1701 | ||
1702 | list_for_each_entry_safe(obj, next, | |
1703 | &dev_priv->mm.unbound_list, | |
35c20a60 | 1704 | global_list) { |
93927ca5 | 1705 | if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) && |
37e680a1 | 1706 | i915_gem_object_put_pages(obj) == 0) { |
6c085a72 CW |
1707 | count += obj->base.size >> PAGE_SHIFT; |
1708 | if (count >= target) | |
1709 | return count; | |
1710 | } | |
1711 | } | |
1712 | ||
57094f82 CW |
1713 | /* |
1714 | * As we may completely rewrite the bound list whilst unbinding | |
1715 | * (due to retiring requests) we have to strictly process only | |
1716 | * one element of the list at the time, and recheck the list | |
1717 | * on every iteration. | |
1718 | */ | |
1719 | INIT_LIST_HEAD(&still_bound_list); | |
1720 | while (count < target && !list_empty(&dev_priv->mm.bound_list)) { | |
07fe0b12 | 1721 | struct i915_vma *vma, *v; |
80dcfdbd | 1722 | |
57094f82 CW |
1723 | obj = list_first_entry(&dev_priv->mm.bound_list, |
1724 | typeof(*obj), global_list); | |
1725 | list_move_tail(&obj->global_list, &still_bound_list); | |
1726 | ||
80dcfdbd BW |
1727 | if (!i915_gem_object_is_purgeable(obj) && purgeable_only) |
1728 | continue; | |
1729 | ||
57094f82 CW |
1730 | /* |
1731 | * Hold a reference whilst we unbind this object, as we may | |
1732 | * end up waiting for and retiring requests. This might | |
1733 | * release the final reference (held by the active list) | |
1734 | * and result in the object being freed from under us. | |
1735 | * in this object being freed. | |
1736 | * | |
1737 | * Note 1: Shrinking the bound list is special since only active | |
1738 | * (and hence bound objects) can contain such limbo objects, so | |
1739 | * we don't need special tricks for shrinking the unbound list. | |
1740 | * The only other place where we have to be careful with active | |
1741 | * objects suddenly disappearing due to retiring requests is the | |
1742 | * eviction code. | |
1743 | * | |
1744 | * Note 2: Even though the bound list doesn't hold a reference | |
1745 | * to the object we can safely grab one here: The final object | |
1746 | * unreferencing and the bound_list are both protected by the | |
1747 | * dev->struct_mutex and so we won't ever be able to observe an | |
1748 | * object on the bound_list with a reference count equals 0. | |
1749 | */ | |
1750 | drm_gem_object_reference(&obj->base); | |
1751 | ||
07fe0b12 BW |
1752 | list_for_each_entry_safe(vma, v, &obj->vma_list, vma_link) |
1753 | if (i915_vma_unbind(vma)) | |
1754 | break; | |
80dcfdbd | 1755 | |
57094f82 | 1756 | if (i915_gem_object_put_pages(obj) == 0) |
6c085a72 | 1757 | count += obj->base.size >> PAGE_SHIFT; |
57094f82 CW |
1758 | |
1759 | drm_gem_object_unreference(&obj->base); | |
6c085a72 | 1760 | } |
57094f82 | 1761 | list_splice(&still_bound_list, &dev_priv->mm.bound_list); |
6c085a72 CW |
1762 | |
1763 | return count; | |
1764 | } | |
1765 | ||
93927ca5 DV |
1766 | static long |
1767 | i915_gem_purge(struct drm_i915_private *dev_priv, long target) | |
1768 | { | |
1769 | return __i915_gem_shrink(dev_priv, target, true); | |
1770 | } | |
1771 | ||
6c085a72 CW |
1772 | static void |
1773 | i915_gem_shrink_all(struct drm_i915_private *dev_priv) | |
1774 | { | |
1775 | struct drm_i915_gem_object *obj, *next; | |
1776 | ||
1777 | i915_gem_evict_everything(dev_priv->dev); | |
1778 | ||
35c20a60 BW |
1779 | list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list, |
1780 | global_list) | |
37e680a1 | 1781 | i915_gem_object_put_pages(obj); |
225067ee DV |
1782 | } |
1783 | ||
37e680a1 | 1784 | static int |
6c085a72 | 1785 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj) |
e5281ccd | 1786 | { |
6c085a72 | 1787 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
e5281ccd CW |
1788 | int page_count, i; |
1789 | struct address_space *mapping; | |
9da3da66 CW |
1790 | struct sg_table *st; |
1791 | struct scatterlist *sg; | |
90797e6d | 1792 | struct sg_page_iter sg_iter; |
e5281ccd | 1793 | struct page *page; |
90797e6d | 1794 | unsigned long last_pfn = 0; /* suppress gcc warning */ |
6c085a72 | 1795 | gfp_t gfp; |
e5281ccd | 1796 | |
6c085a72 CW |
1797 | /* Assert that the object is not currently in any GPU domain. As it |
1798 | * wasn't in the GTT, there shouldn't be any way it could have been in | |
1799 | * a GPU cache | |
1800 | */ | |
1801 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); | |
1802 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); | |
1803 | ||
9da3da66 CW |
1804 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
1805 | if (st == NULL) | |
1806 | return -ENOMEM; | |
1807 | ||
05394f39 | 1808 | page_count = obj->base.size / PAGE_SIZE; |
9da3da66 | 1809 | if (sg_alloc_table(st, page_count, GFP_KERNEL)) { |
9da3da66 | 1810 | kfree(st); |
e5281ccd | 1811 | return -ENOMEM; |
9da3da66 | 1812 | } |
e5281ccd | 1813 | |
9da3da66 CW |
1814 | /* Get the list of pages out of our struct file. They'll be pinned |
1815 | * at this point until we release them. | |
1816 | * | |
1817 | * Fail silently without starting the shrinker | |
1818 | */ | |
496ad9aa | 1819 | mapping = file_inode(obj->base.filp)->i_mapping; |
6c085a72 | 1820 | gfp = mapping_gfp_mask(mapping); |
caf49191 | 1821 | gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD; |
6c085a72 | 1822 | gfp &= ~(__GFP_IO | __GFP_WAIT); |
90797e6d ID |
1823 | sg = st->sgl; |
1824 | st->nents = 0; | |
1825 | for (i = 0; i < page_count; i++) { | |
6c085a72 CW |
1826 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); |
1827 | if (IS_ERR(page)) { | |
1828 | i915_gem_purge(dev_priv, page_count); | |
1829 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); | |
1830 | } | |
1831 | if (IS_ERR(page)) { | |
1832 | /* We've tried hard to allocate the memory by reaping | |
1833 | * our own buffer, now let the real VM do its job and | |
1834 | * go down in flames if truly OOM. | |
1835 | */ | |
caf49191 | 1836 | gfp &= ~(__GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD); |
6c085a72 CW |
1837 | gfp |= __GFP_IO | __GFP_WAIT; |
1838 | ||
1839 | i915_gem_shrink_all(dev_priv); | |
1840 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); | |
1841 | if (IS_ERR(page)) | |
1842 | goto err_pages; | |
1843 | ||
caf49191 | 1844 | gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD; |
6c085a72 CW |
1845 | gfp &= ~(__GFP_IO | __GFP_WAIT); |
1846 | } | |
426729dc KRW |
1847 | #ifdef CONFIG_SWIOTLB |
1848 | if (swiotlb_nr_tbl()) { | |
1849 | st->nents++; | |
1850 | sg_set_page(sg, page, PAGE_SIZE, 0); | |
1851 | sg = sg_next(sg); | |
1852 | continue; | |
1853 | } | |
1854 | #endif | |
90797e6d ID |
1855 | if (!i || page_to_pfn(page) != last_pfn + 1) { |
1856 | if (i) | |
1857 | sg = sg_next(sg); | |
1858 | st->nents++; | |
1859 | sg_set_page(sg, page, PAGE_SIZE, 0); | |
1860 | } else { | |
1861 | sg->length += PAGE_SIZE; | |
1862 | } | |
1863 | last_pfn = page_to_pfn(page); | |
e5281ccd | 1864 | } |
426729dc KRW |
1865 | #ifdef CONFIG_SWIOTLB |
1866 | if (!swiotlb_nr_tbl()) | |
1867 | #endif | |
1868 | sg_mark_end(sg); | |
74ce6b6c CW |
1869 | obj->pages = st; |
1870 | ||
6dacfd2f | 1871 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
e5281ccd CW |
1872 | i915_gem_object_do_bit_17_swizzle(obj); |
1873 | ||
1874 | return 0; | |
1875 | ||
1876 | err_pages: | |
90797e6d ID |
1877 | sg_mark_end(sg); |
1878 | for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) | |
2db76d7c | 1879 | page_cache_release(sg_page_iter_page(&sg_iter)); |
9da3da66 CW |
1880 | sg_free_table(st); |
1881 | kfree(st); | |
e5281ccd | 1882 | return PTR_ERR(page); |
673a394b EA |
1883 | } |
1884 | ||
37e680a1 CW |
1885 | /* Ensure that the associated pages are gathered from the backing storage |
1886 | * and pinned into our object. i915_gem_object_get_pages() may be called | |
1887 | * multiple times before they are released by a single call to | |
1888 | * i915_gem_object_put_pages() - once the pages are no longer referenced | |
1889 | * either as a result of memory pressure (reaping pages under the shrinker) | |
1890 | * or as the object is itself released. | |
1891 | */ | |
1892 | int | |
1893 | i915_gem_object_get_pages(struct drm_i915_gem_object *obj) | |
1894 | { | |
1895 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
1896 | const struct drm_i915_gem_object_ops *ops = obj->ops; | |
1897 | int ret; | |
1898 | ||
2f745ad3 | 1899 | if (obj->pages) |
37e680a1 CW |
1900 | return 0; |
1901 | ||
43e28f09 CW |
1902 | if (obj->madv != I915_MADV_WILLNEED) { |
1903 | DRM_ERROR("Attempting to obtain a purgeable object\n"); | |
1904 | return -EINVAL; | |
1905 | } | |
1906 | ||
a5570178 CW |
1907 | BUG_ON(obj->pages_pin_count); |
1908 | ||
37e680a1 CW |
1909 | ret = ops->get_pages(obj); |
1910 | if (ret) | |
1911 | return ret; | |
1912 | ||
35c20a60 | 1913 | list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list); |
37e680a1 | 1914 | return 0; |
673a394b EA |
1915 | } |
1916 | ||
54cf91dc | 1917 | void |
05394f39 | 1918 | i915_gem_object_move_to_active(struct drm_i915_gem_object *obj, |
9d773091 | 1919 | struct intel_ring_buffer *ring) |
673a394b | 1920 | { |
05394f39 | 1921 | struct drm_device *dev = obj->base.dev; |
69dc4987 | 1922 | struct drm_i915_private *dev_priv = dev->dev_private; |
9d773091 | 1923 | u32 seqno = intel_ring_get_seqno(ring); |
617dbe27 | 1924 | |
852835f3 | 1925 | BUG_ON(ring == NULL); |
02978ff5 CW |
1926 | if (obj->ring != ring && obj->last_write_seqno) { |
1927 | /* Keep the seqno relative to the current ring */ | |
1928 | obj->last_write_seqno = seqno; | |
1929 | } | |
05394f39 | 1930 | obj->ring = ring; |
673a394b EA |
1931 | |
1932 | /* Add a reference if we're newly entering the active list. */ | |
05394f39 CW |
1933 | if (!obj->active) { |
1934 | drm_gem_object_reference(&obj->base); | |
1935 | obj->active = 1; | |
673a394b | 1936 | } |
e35a41de | 1937 | |
05394f39 | 1938 | list_move_tail(&obj->ring_list, &ring->active_list); |
caea7476 | 1939 | |
0201f1ec | 1940 | obj->last_read_seqno = seqno; |
caea7476 | 1941 | |
7dd49065 | 1942 | if (obj->fenced_gpu_access) { |
caea7476 | 1943 | obj->last_fenced_seqno = seqno; |
caea7476 | 1944 | |
7dd49065 CW |
1945 | /* Bump MRU to take account of the delayed flush */ |
1946 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
1947 | struct drm_i915_fence_reg *reg; | |
1948 | ||
1949 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
1950 | list_move_tail(®->lru_list, | |
1951 | &dev_priv->mm.fence_list); | |
1952 | } | |
caea7476 CW |
1953 | } |
1954 | } | |
1955 | ||
1956 | static void | |
caea7476 | 1957 | i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj) |
ce44b0ea | 1958 | { |
ca191b13 BW |
1959 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
1960 | struct i915_address_space *ggtt_vm = &dev_priv->gtt.base; | |
1961 | struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm); | |
ce44b0ea | 1962 | |
65ce3027 | 1963 | BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS); |
05394f39 | 1964 | BUG_ON(!obj->active); |
caea7476 | 1965 | |
ca191b13 | 1966 | list_move_tail(&vma->mm_list, &ggtt_vm->inactive_list); |
caea7476 | 1967 | |
65ce3027 | 1968 | list_del_init(&obj->ring_list); |
caea7476 CW |
1969 | obj->ring = NULL; |
1970 | ||
65ce3027 CW |
1971 | obj->last_read_seqno = 0; |
1972 | obj->last_write_seqno = 0; | |
1973 | obj->base.write_domain = 0; | |
1974 | ||
1975 | obj->last_fenced_seqno = 0; | |
caea7476 | 1976 | obj->fenced_gpu_access = false; |
caea7476 CW |
1977 | |
1978 | obj->active = 0; | |
1979 | drm_gem_object_unreference(&obj->base); | |
1980 | ||
1981 | WARN_ON(i915_verify_lists(dev)); | |
ce44b0ea | 1982 | } |
673a394b | 1983 | |
9d773091 | 1984 | static int |
fca26bb4 | 1985 | i915_gem_init_seqno(struct drm_device *dev, u32 seqno) |
53d227f2 | 1986 | { |
9d773091 CW |
1987 | struct drm_i915_private *dev_priv = dev->dev_private; |
1988 | struct intel_ring_buffer *ring; | |
1989 | int ret, i, j; | |
53d227f2 | 1990 | |
107f27a5 | 1991 | /* Carefully retire all requests without writing to the rings */ |
9d773091 | 1992 | for_each_ring(ring, dev_priv, i) { |
107f27a5 CW |
1993 | ret = intel_ring_idle(ring); |
1994 | if (ret) | |
1995 | return ret; | |
9d773091 | 1996 | } |
9d773091 | 1997 | i915_gem_retire_requests(dev); |
107f27a5 CW |
1998 | |
1999 | /* Finally reset hw state */ | |
9d773091 | 2000 | for_each_ring(ring, dev_priv, i) { |
fca26bb4 | 2001 | intel_ring_init_seqno(ring, seqno); |
498d2ac1 | 2002 | |
9d773091 CW |
2003 | for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++) |
2004 | ring->sync_seqno[j] = 0; | |
2005 | } | |
53d227f2 | 2006 | |
9d773091 | 2007 | return 0; |
53d227f2 DV |
2008 | } |
2009 | ||
fca26bb4 MK |
2010 | int i915_gem_set_seqno(struct drm_device *dev, u32 seqno) |
2011 | { | |
2012 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2013 | int ret; | |
2014 | ||
2015 | if (seqno == 0) | |
2016 | return -EINVAL; | |
2017 | ||
2018 | /* HWS page needs to be set less than what we | |
2019 | * will inject to ring | |
2020 | */ | |
2021 | ret = i915_gem_init_seqno(dev, seqno - 1); | |
2022 | if (ret) | |
2023 | return ret; | |
2024 | ||
2025 | /* Carefully set the last_seqno value so that wrap | |
2026 | * detection still works | |
2027 | */ | |
2028 | dev_priv->next_seqno = seqno; | |
2029 | dev_priv->last_seqno = seqno - 1; | |
2030 | if (dev_priv->last_seqno == 0) | |
2031 | dev_priv->last_seqno--; | |
2032 | ||
2033 | return 0; | |
2034 | } | |
2035 | ||
9d773091 CW |
2036 | int |
2037 | i915_gem_get_seqno(struct drm_device *dev, u32 *seqno) | |
53d227f2 | 2038 | { |
9d773091 CW |
2039 | struct drm_i915_private *dev_priv = dev->dev_private; |
2040 | ||
2041 | /* reserve 0 for non-seqno */ | |
2042 | if (dev_priv->next_seqno == 0) { | |
fca26bb4 | 2043 | int ret = i915_gem_init_seqno(dev, 0); |
9d773091 CW |
2044 | if (ret) |
2045 | return ret; | |
53d227f2 | 2046 | |
9d773091 CW |
2047 | dev_priv->next_seqno = 1; |
2048 | } | |
53d227f2 | 2049 | |
f72b3435 | 2050 | *seqno = dev_priv->last_seqno = dev_priv->next_seqno++; |
9d773091 | 2051 | return 0; |
53d227f2 DV |
2052 | } |
2053 | ||
0025c077 MK |
2054 | int __i915_add_request(struct intel_ring_buffer *ring, |
2055 | struct drm_file *file, | |
7d736f4f | 2056 | struct drm_i915_gem_object *obj, |
0025c077 | 2057 | u32 *out_seqno) |
673a394b | 2058 | { |
db53a302 | 2059 | drm_i915_private_t *dev_priv = ring->dev->dev_private; |
acb868d3 | 2060 | struct drm_i915_gem_request *request; |
7d736f4f | 2061 | u32 request_ring_position, request_start; |
673a394b | 2062 | int was_empty; |
3cce469c CW |
2063 | int ret; |
2064 | ||
7d736f4f | 2065 | request_start = intel_ring_get_tail(ring); |
cc889e0f DV |
2066 | /* |
2067 | * Emit any outstanding flushes - execbuf can fail to emit the flush | |
2068 | * after having emitted the batchbuffer command. Hence we need to fix | |
2069 | * things up similar to emitting the lazy request. The difference here | |
2070 | * is that the flush _must_ happen before the next request, no matter | |
2071 | * what. | |
2072 | */ | |
a7b9761d CW |
2073 | ret = intel_ring_flush_all_caches(ring); |
2074 | if (ret) | |
2075 | return ret; | |
cc889e0f | 2076 | |
acb868d3 CW |
2077 | request = kmalloc(sizeof(*request), GFP_KERNEL); |
2078 | if (request == NULL) | |
2079 | return -ENOMEM; | |
cc889e0f | 2080 | |
673a394b | 2081 | |
a71d8d94 CW |
2082 | /* Record the position of the start of the request so that |
2083 | * should we detect the updated seqno part-way through the | |
2084 | * GPU processing the request, we never over-estimate the | |
2085 | * position of the head. | |
2086 | */ | |
2087 | request_ring_position = intel_ring_get_tail(ring); | |
2088 | ||
9d773091 | 2089 | ret = ring->add_request(ring); |
3bb73aba CW |
2090 | if (ret) { |
2091 | kfree(request); | |
2092 | return ret; | |
2093 | } | |
673a394b | 2094 | |
9d773091 | 2095 | request->seqno = intel_ring_get_seqno(ring); |
852835f3 | 2096 | request->ring = ring; |
7d736f4f | 2097 | request->head = request_start; |
a71d8d94 | 2098 | request->tail = request_ring_position; |
0e50e96b | 2099 | request->ctx = ring->last_context; |
7d736f4f MK |
2100 | request->batch_obj = obj; |
2101 | ||
2102 | /* Whilst this request exists, batch_obj will be on the | |
2103 | * active_list, and so will hold the active reference. Only when this | |
2104 | * request is retired will the the batch_obj be moved onto the | |
2105 | * inactive_list and lose its active reference. Hence we do not need | |
2106 | * to explicitly hold another reference here. | |
2107 | */ | |
0e50e96b MK |
2108 | |
2109 | if (request->ctx) | |
2110 | i915_gem_context_reference(request->ctx); | |
2111 | ||
673a394b | 2112 | request->emitted_jiffies = jiffies; |
852835f3 ZN |
2113 | was_empty = list_empty(&ring->request_list); |
2114 | list_add_tail(&request->list, &ring->request_list); | |
3bb73aba | 2115 | request->file_priv = NULL; |
852835f3 | 2116 | |
db53a302 CW |
2117 | if (file) { |
2118 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
2119 | ||
1c25595f | 2120 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 2121 | request->file_priv = file_priv; |
b962442e | 2122 | list_add_tail(&request->client_list, |
f787a5f5 | 2123 | &file_priv->mm.request_list); |
1c25595f | 2124 | spin_unlock(&file_priv->mm.lock); |
b962442e | 2125 | } |
673a394b | 2126 | |
9d773091 | 2127 | trace_i915_gem_request_add(ring, request->seqno); |
5391d0cf | 2128 | ring->outstanding_lazy_request = 0; |
db53a302 | 2129 | |
db1b76ca | 2130 | if (!dev_priv->ums.mm_suspended) { |
10cd45b6 MK |
2131 | i915_queue_hangcheck(ring->dev); |
2132 | ||
f047e395 | 2133 | if (was_empty) { |
b3b079db | 2134 | queue_delayed_work(dev_priv->wq, |
bcb45086 CW |
2135 | &dev_priv->mm.retire_work, |
2136 | round_jiffies_up_relative(HZ)); | |
f047e395 CW |
2137 | intel_mark_busy(dev_priv->dev); |
2138 | } | |
f65d9421 | 2139 | } |
cc889e0f | 2140 | |
acb868d3 | 2141 | if (out_seqno) |
9d773091 | 2142 | *out_seqno = request->seqno; |
3cce469c | 2143 | return 0; |
673a394b EA |
2144 | } |
2145 | ||
f787a5f5 CW |
2146 | static inline void |
2147 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) | |
673a394b | 2148 | { |
1c25595f | 2149 | struct drm_i915_file_private *file_priv = request->file_priv; |
673a394b | 2150 | |
1c25595f CW |
2151 | if (!file_priv) |
2152 | return; | |
1c5d22f7 | 2153 | |
1c25595f | 2154 | spin_lock(&file_priv->mm.lock); |
09bfa517 HRK |
2155 | if (request->file_priv) { |
2156 | list_del(&request->client_list); | |
2157 | request->file_priv = NULL; | |
2158 | } | |
1c25595f | 2159 | spin_unlock(&file_priv->mm.lock); |
673a394b | 2160 | } |
673a394b | 2161 | |
d1ccbb5d BW |
2162 | static bool i915_head_inside_object(u32 acthd, struct drm_i915_gem_object *obj, |
2163 | struct i915_address_space *vm) | |
aa60c664 | 2164 | { |
d1ccbb5d BW |
2165 | if (acthd >= i915_gem_obj_offset(obj, vm) && |
2166 | acthd < i915_gem_obj_offset(obj, vm) + obj->base.size) | |
aa60c664 MK |
2167 | return true; |
2168 | ||
2169 | return false; | |
2170 | } | |
2171 | ||
2172 | static bool i915_head_inside_request(const u32 acthd_unmasked, | |
2173 | const u32 request_start, | |
2174 | const u32 request_end) | |
2175 | { | |
2176 | const u32 acthd = acthd_unmasked & HEAD_ADDR; | |
2177 | ||
2178 | if (request_start < request_end) { | |
2179 | if (acthd >= request_start && acthd < request_end) | |
2180 | return true; | |
2181 | } else if (request_start > request_end) { | |
2182 | if (acthd >= request_start || acthd < request_end) | |
2183 | return true; | |
2184 | } | |
2185 | ||
2186 | return false; | |
2187 | } | |
2188 | ||
d1ccbb5d BW |
2189 | static struct i915_address_space * |
2190 | request_to_vm(struct drm_i915_gem_request *request) | |
2191 | { | |
2192 | struct drm_i915_private *dev_priv = request->ring->dev->dev_private; | |
2193 | struct i915_address_space *vm; | |
2194 | ||
2195 | vm = &dev_priv->gtt.base; | |
2196 | ||
2197 | return vm; | |
2198 | } | |
2199 | ||
aa60c664 MK |
2200 | static bool i915_request_guilty(struct drm_i915_gem_request *request, |
2201 | const u32 acthd, bool *inside) | |
2202 | { | |
2203 | /* There is a possibility that unmasked head address | |
2204 | * pointing inside the ring, matches the batch_obj address range. | |
2205 | * However this is extremely unlikely. | |
2206 | */ | |
aa60c664 | 2207 | if (request->batch_obj) { |
d1ccbb5d BW |
2208 | if (i915_head_inside_object(acthd, request->batch_obj, |
2209 | request_to_vm(request))) { | |
aa60c664 MK |
2210 | *inside = true; |
2211 | return true; | |
2212 | } | |
2213 | } | |
2214 | ||
2215 | if (i915_head_inside_request(acthd, request->head, request->tail)) { | |
2216 | *inside = false; | |
2217 | return true; | |
2218 | } | |
2219 | ||
2220 | return false; | |
2221 | } | |
2222 | ||
2223 | static void i915_set_reset_status(struct intel_ring_buffer *ring, | |
2224 | struct drm_i915_gem_request *request, | |
2225 | u32 acthd) | |
2226 | { | |
2227 | struct i915_ctx_hang_stats *hs = NULL; | |
2228 | bool inside, guilty; | |
d1ccbb5d | 2229 | unsigned long offset = 0; |
aa60c664 MK |
2230 | |
2231 | /* Innocent until proven guilty */ | |
2232 | guilty = false; | |
2233 | ||
d1ccbb5d BW |
2234 | if (request->batch_obj) |
2235 | offset = i915_gem_obj_offset(request->batch_obj, | |
2236 | request_to_vm(request)); | |
2237 | ||
f2f4d82f | 2238 | if (ring->hangcheck.action != HANGCHECK_WAIT && |
aa60c664 | 2239 | i915_request_guilty(request, acthd, &inside)) { |
f343c5f6 | 2240 | DRM_ERROR("%s hung %s bo (0x%lx ctx %d) at 0x%x\n", |
aa60c664 MK |
2241 | ring->name, |
2242 | inside ? "inside" : "flushing", | |
d1ccbb5d | 2243 | offset, |
aa60c664 MK |
2244 | request->ctx ? request->ctx->id : 0, |
2245 | acthd); | |
2246 | ||
2247 | guilty = true; | |
2248 | } | |
2249 | ||
2250 | /* If contexts are disabled or this is the default context, use | |
2251 | * file_priv->reset_state | |
2252 | */ | |
2253 | if (request->ctx && request->ctx->id != DEFAULT_CONTEXT_ID) | |
2254 | hs = &request->ctx->hang_stats; | |
2255 | else if (request->file_priv) | |
2256 | hs = &request->file_priv->hang_stats; | |
2257 | ||
2258 | if (hs) { | |
2259 | if (guilty) | |
2260 | hs->batch_active++; | |
2261 | else | |
2262 | hs->batch_pending++; | |
2263 | } | |
2264 | } | |
2265 | ||
0e50e96b MK |
2266 | static void i915_gem_free_request(struct drm_i915_gem_request *request) |
2267 | { | |
2268 | list_del(&request->list); | |
2269 | i915_gem_request_remove_from_client(request); | |
2270 | ||
2271 | if (request->ctx) | |
2272 | i915_gem_context_unreference(request->ctx); | |
2273 | ||
2274 | kfree(request); | |
2275 | } | |
2276 | ||
dfaae392 CW |
2277 | static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv, |
2278 | struct intel_ring_buffer *ring) | |
9375e446 | 2279 | { |
aa60c664 MK |
2280 | u32 completed_seqno; |
2281 | u32 acthd; | |
2282 | ||
2283 | acthd = intel_ring_get_active_head(ring); | |
2284 | completed_seqno = ring->get_seqno(ring, false); | |
2285 | ||
dfaae392 CW |
2286 | while (!list_empty(&ring->request_list)) { |
2287 | struct drm_i915_gem_request *request; | |
673a394b | 2288 | |
dfaae392 CW |
2289 | request = list_first_entry(&ring->request_list, |
2290 | struct drm_i915_gem_request, | |
2291 | list); | |
de151cf6 | 2292 | |
aa60c664 MK |
2293 | if (request->seqno > completed_seqno) |
2294 | i915_set_reset_status(ring, request, acthd); | |
2295 | ||
0e50e96b | 2296 | i915_gem_free_request(request); |
dfaae392 | 2297 | } |
673a394b | 2298 | |
dfaae392 | 2299 | while (!list_empty(&ring->active_list)) { |
05394f39 | 2300 | struct drm_i915_gem_object *obj; |
9375e446 | 2301 | |
05394f39 CW |
2302 | obj = list_first_entry(&ring->active_list, |
2303 | struct drm_i915_gem_object, | |
2304 | ring_list); | |
9375e446 | 2305 | |
05394f39 | 2306 | i915_gem_object_move_to_inactive(obj); |
673a394b EA |
2307 | } |
2308 | } | |
2309 | ||
19b2dbde | 2310 | void i915_gem_restore_fences(struct drm_device *dev) |
312817a3 CW |
2311 | { |
2312 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2313 | int i; | |
2314 | ||
4b9de737 | 2315 | for (i = 0; i < dev_priv->num_fence_regs; i++) { |
312817a3 | 2316 | struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i]; |
7d2cb39c | 2317 | |
94a335db DV |
2318 | /* |
2319 | * Commit delayed tiling changes if we have an object still | |
2320 | * attached to the fence, otherwise just clear the fence. | |
2321 | */ | |
2322 | if (reg->obj) { | |
2323 | i915_gem_object_update_fence(reg->obj, reg, | |
2324 | reg->obj->tiling_mode); | |
2325 | } else { | |
2326 | i915_gem_write_fence(dev, i, NULL); | |
2327 | } | |
312817a3 CW |
2328 | } |
2329 | } | |
2330 | ||
069efc1d | 2331 | void i915_gem_reset(struct drm_device *dev) |
673a394b | 2332 | { |
77f01230 | 2333 | struct drm_i915_private *dev_priv = dev->dev_private; |
b4519513 | 2334 | struct intel_ring_buffer *ring; |
1ec14ad3 | 2335 | int i; |
673a394b | 2336 | |
b4519513 CW |
2337 | for_each_ring(ring, dev_priv, i) |
2338 | i915_gem_reset_ring_lists(dev_priv, ring); | |
dfaae392 | 2339 | |
19b2dbde | 2340 | i915_gem_restore_fences(dev); |
673a394b EA |
2341 | } |
2342 | ||
2343 | /** | |
2344 | * This function clears the request list as sequence numbers are passed. | |
2345 | */ | |
a71d8d94 | 2346 | void |
db53a302 | 2347 | i915_gem_retire_requests_ring(struct intel_ring_buffer *ring) |
673a394b | 2348 | { |
673a394b EA |
2349 | uint32_t seqno; |
2350 | ||
db53a302 | 2351 | if (list_empty(&ring->request_list)) |
6c0594a3 KW |
2352 | return; |
2353 | ||
db53a302 | 2354 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b | 2355 | |
b2eadbc8 | 2356 | seqno = ring->get_seqno(ring, true); |
1ec14ad3 | 2357 | |
852835f3 | 2358 | while (!list_empty(&ring->request_list)) { |
673a394b | 2359 | struct drm_i915_gem_request *request; |
673a394b | 2360 | |
852835f3 | 2361 | request = list_first_entry(&ring->request_list, |
673a394b EA |
2362 | struct drm_i915_gem_request, |
2363 | list); | |
673a394b | 2364 | |
dfaae392 | 2365 | if (!i915_seqno_passed(seqno, request->seqno)) |
b84d5f0c CW |
2366 | break; |
2367 | ||
db53a302 | 2368 | trace_i915_gem_request_retire(ring, request->seqno); |
a71d8d94 CW |
2369 | /* We know the GPU must have read the request to have |
2370 | * sent us the seqno + interrupt, so use the position | |
2371 | * of tail of the request to update the last known position | |
2372 | * of the GPU head. | |
2373 | */ | |
2374 | ring->last_retired_head = request->tail; | |
b84d5f0c | 2375 | |
0e50e96b | 2376 | i915_gem_free_request(request); |
b84d5f0c | 2377 | } |
673a394b | 2378 | |
b84d5f0c CW |
2379 | /* Move any buffers on the active list that are no longer referenced |
2380 | * by the ringbuffer to the flushing/inactive lists as appropriate. | |
2381 | */ | |
2382 | while (!list_empty(&ring->active_list)) { | |
05394f39 | 2383 | struct drm_i915_gem_object *obj; |
b84d5f0c | 2384 | |
0206e353 | 2385 | obj = list_first_entry(&ring->active_list, |
05394f39 CW |
2386 | struct drm_i915_gem_object, |
2387 | ring_list); | |
673a394b | 2388 | |
0201f1ec | 2389 | if (!i915_seqno_passed(seqno, obj->last_read_seqno)) |
673a394b | 2390 | break; |
b84d5f0c | 2391 | |
65ce3027 | 2392 | i915_gem_object_move_to_inactive(obj); |
673a394b | 2393 | } |
9d34e5db | 2394 | |
db53a302 CW |
2395 | if (unlikely(ring->trace_irq_seqno && |
2396 | i915_seqno_passed(seqno, ring->trace_irq_seqno))) { | |
1ec14ad3 | 2397 | ring->irq_put(ring); |
db53a302 | 2398 | ring->trace_irq_seqno = 0; |
9d34e5db | 2399 | } |
23bc5982 | 2400 | |
db53a302 | 2401 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b EA |
2402 | } |
2403 | ||
b09a1fec CW |
2404 | void |
2405 | i915_gem_retire_requests(struct drm_device *dev) | |
2406 | { | |
2407 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 2408 | struct intel_ring_buffer *ring; |
1ec14ad3 | 2409 | int i; |
b09a1fec | 2410 | |
b4519513 CW |
2411 | for_each_ring(ring, dev_priv, i) |
2412 | i915_gem_retire_requests_ring(ring); | |
b09a1fec CW |
2413 | } |
2414 | ||
75ef9da2 | 2415 | static void |
673a394b EA |
2416 | i915_gem_retire_work_handler(struct work_struct *work) |
2417 | { | |
2418 | drm_i915_private_t *dev_priv; | |
2419 | struct drm_device *dev; | |
b4519513 | 2420 | struct intel_ring_buffer *ring; |
0a58705b CW |
2421 | bool idle; |
2422 | int i; | |
673a394b EA |
2423 | |
2424 | dev_priv = container_of(work, drm_i915_private_t, | |
2425 | mm.retire_work.work); | |
2426 | dev = dev_priv->dev; | |
2427 | ||
891b48cf CW |
2428 | /* Come back later if the device is busy... */ |
2429 | if (!mutex_trylock(&dev->struct_mutex)) { | |
bcb45086 CW |
2430 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2431 | round_jiffies_up_relative(HZ)); | |
891b48cf CW |
2432 | return; |
2433 | } | |
673a394b | 2434 | |
b09a1fec | 2435 | i915_gem_retire_requests(dev); |
673a394b | 2436 | |
0a58705b CW |
2437 | /* Send a periodic flush down the ring so we don't hold onto GEM |
2438 | * objects indefinitely. | |
673a394b | 2439 | */ |
0a58705b | 2440 | idle = true; |
b4519513 | 2441 | for_each_ring(ring, dev_priv, i) { |
3bb73aba | 2442 | if (ring->gpu_caches_dirty) |
0025c077 | 2443 | i915_add_request(ring, NULL); |
0a58705b CW |
2444 | |
2445 | idle &= list_empty(&ring->request_list); | |
673a394b EA |
2446 | } |
2447 | ||
db1b76ca | 2448 | if (!dev_priv->ums.mm_suspended && !idle) |
bcb45086 CW |
2449 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, |
2450 | round_jiffies_up_relative(HZ)); | |
f047e395 CW |
2451 | if (idle) |
2452 | intel_mark_idle(dev); | |
0a58705b | 2453 | |
673a394b | 2454 | mutex_unlock(&dev->struct_mutex); |
673a394b EA |
2455 | } |
2456 | ||
30dfebf3 DV |
2457 | /** |
2458 | * Ensures that an object will eventually get non-busy by flushing any required | |
2459 | * write domains, emitting any outstanding lazy request and retiring and | |
2460 | * completed requests. | |
2461 | */ | |
2462 | static int | |
2463 | i915_gem_object_flush_active(struct drm_i915_gem_object *obj) | |
2464 | { | |
2465 | int ret; | |
2466 | ||
2467 | if (obj->active) { | |
0201f1ec | 2468 | ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno); |
30dfebf3 DV |
2469 | if (ret) |
2470 | return ret; | |
2471 | ||
30dfebf3 DV |
2472 | i915_gem_retire_requests_ring(obj->ring); |
2473 | } | |
2474 | ||
2475 | return 0; | |
2476 | } | |
2477 | ||
23ba4fd0 BW |
2478 | /** |
2479 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT | |
2480 | * @DRM_IOCTL_ARGS: standard ioctl arguments | |
2481 | * | |
2482 | * Returns 0 if successful, else an error is returned with the remaining time in | |
2483 | * the timeout parameter. | |
2484 | * -ETIME: object is still busy after timeout | |
2485 | * -ERESTARTSYS: signal interrupted the wait | |
2486 | * -ENONENT: object doesn't exist | |
2487 | * Also possible, but rare: | |
2488 | * -EAGAIN: GPU wedged | |
2489 | * -ENOMEM: damn | |
2490 | * -ENODEV: Internal IRQ fail | |
2491 | * -E?: The add request failed | |
2492 | * | |
2493 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any | |
2494 | * non-zero timeout parameter the wait ioctl will wait for the given number of | |
2495 | * nanoseconds on an object becoming unbusy. Since the wait itself does so | |
2496 | * without holding struct_mutex the object may become re-busied before this | |
2497 | * function completes. A similar but shorter * race condition exists in the busy | |
2498 | * ioctl | |
2499 | */ | |
2500 | int | |
2501 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
2502 | { | |
f69061be | 2503 | drm_i915_private_t *dev_priv = dev->dev_private; |
23ba4fd0 BW |
2504 | struct drm_i915_gem_wait *args = data; |
2505 | struct drm_i915_gem_object *obj; | |
2506 | struct intel_ring_buffer *ring = NULL; | |
eac1f14f | 2507 | struct timespec timeout_stack, *timeout = NULL; |
f69061be | 2508 | unsigned reset_counter; |
23ba4fd0 BW |
2509 | u32 seqno = 0; |
2510 | int ret = 0; | |
2511 | ||
eac1f14f BW |
2512 | if (args->timeout_ns >= 0) { |
2513 | timeout_stack = ns_to_timespec(args->timeout_ns); | |
2514 | timeout = &timeout_stack; | |
2515 | } | |
23ba4fd0 BW |
2516 | |
2517 | ret = i915_mutex_lock_interruptible(dev); | |
2518 | if (ret) | |
2519 | return ret; | |
2520 | ||
2521 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle)); | |
2522 | if (&obj->base == NULL) { | |
2523 | mutex_unlock(&dev->struct_mutex); | |
2524 | return -ENOENT; | |
2525 | } | |
2526 | ||
30dfebf3 DV |
2527 | /* Need to make sure the object gets inactive eventually. */ |
2528 | ret = i915_gem_object_flush_active(obj); | |
23ba4fd0 BW |
2529 | if (ret) |
2530 | goto out; | |
2531 | ||
2532 | if (obj->active) { | |
0201f1ec | 2533 | seqno = obj->last_read_seqno; |
23ba4fd0 BW |
2534 | ring = obj->ring; |
2535 | } | |
2536 | ||
2537 | if (seqno == 0) | |
2538 | goto out; | |
2539 | ||
23ba4fd0 BW |
2540 | /* Do this after OLR check to make sure we make forward progress polling |
2541 | * on this IOCTL with a 0 timeout (like busy ioctl) | |
2542 | */ | |
2543 | if (!args->timeout_ns) { | |
2544 | ret = -ETIME; | |
2545 | goto out; | |
2546 | } | |
2547 | ||
2548 | drm_gem_object_unreference(&obj->base); | |
f69061be | 2549 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
23ba4fd0 BW |
2550 | mutex_unlock(&dev->struct_mutex); |
2551 | ||
f69061be | 2552 | ret = __wait_seqno(ring, seqno, reset_counter, true, timeout); |
4f42f4ef | 2553 | if (timeout) |
eac1f14f | 2554 | args->timeout_ns = timespec_to_ns(timeout); |
23ba4fd0 BW |
2555 | return ret; |
2556 | ||
2557 | out: | |
2558 | drm_gem_object_unreference(&obj->base); | |
2559 | mutex_unlock(&dev->struct_mutex); | |
2560 | return ret; | |
2561 | } | |
2562 | ||
5816d648 BW |
2563 | /** |
2564 | * i915_gem_object_sync - sync an object to a ring. | |
2565 | * | |
2566 | * @obj: object which may be in use on another ring. | |
2567 | * @to: ring we wish to use the object on. May be NULL. | |
2568 | * | |
2569 | * This code is meant to abstract object synchronization with the GPU. | |
2570 | * Calling with NULL implies synchronizing the object with the CPU | |
2571 | * rather than a particular GPU ring. | |
2572 | * | |
2573 | * Returns 0 if successful, else propagates up the lower layer error. | |
2574 | */ | |
2911a35b BW |
2575 | int |
2576 | i915_gem_object_sync(struct drm_i915_gem_object *obj, | |
2577 | struct intel_ring_buffer *to) | |
2578 | { | |
2579 | struct intel_ring_buffer *from = obj->ring; | |
2580 | u32 seqno; | |
2581 | int ret, idx; | |
2582 | ||
2583 | if (from == NULL || to == from) | |
2584 | return 0; | |
2585 | ||
5816d648 | 2586 | if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev)) |
0201f1ec | 2587 | return i915_gem_object_wait_rendering(obj, false); |
2911a35b BW |
2588 | |
2589 | idx = intel_ring_sync_index(from, to); | |
2590 | ||
0201f1ec | 2591 | seqno = obj->last_read_seqno; |
2911a35b BW |
2592 | if (seqno <= from->sync_seqno[idx]) |
2593 | return 0; | |
2594 | ||
b4aca010 BW |
2595 | ret = i915_gem_check_olr(obj->ring, seqno); |
2596 | if (ret) | |
2597 | return ret; | |
2911a35b | 2598 | |
1500f7ea | 2599 | ret = to->sync_to(to, from, seqno); |
e3a5a225 | 2600 | if (!ret) |
7b01e260 MK |
2601 | /* We use last_read_seqno because sync_to() |
2602 | * might have just caused seqno wrap under | |
2603 | * the radar. | |
2604 | */ | |
2605 | from->sync_seqno[idx] = obj->last_read_seqno; | |
2911a35b | 2606 | |
e3a5a225 | 2607 | return ret; |
2911a35b BW |
2608 | } |
2609 | ||
b5ffc9bc CW |
2610 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) |
2611 | { | |
2612 | u32 old_write_domain, old_read_domains; | |
2613 | ||
b5ffc9bc CW |
2614 | /* Force a pagefault for domain tracking on next user access */ |
2615 | i915_gem_release_mmap(obj); | |
2616 | ||
b97c3d9c KP |
2617 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
2618 | return; | |
2619 | ||
97c809fd CW |
2620 | /* Wait for any direct GTT access to complete */ |
2621 | mb(); | |
2622 | ||
b5ffc9bc CW |
2623 | old_read_domains = obj->base.read_domains; |
2624 | old_write_domain = obj->base.write_domain; | |
2625 | ||
2626 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; | |
2627 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; | |
2628 | ||
2629 | trace_i915_gem_object_change_domain(obj, | |
2630 | old_read_domains, | |
2631 | old_write_domain); | |
2632 | } | |
2633 | ||
07fe0b12 | 2634 | int i915_vma_unbind(struct i915_vma *vma) |
673a394b | 2635 | { |
07fe0b12 | 2636 | struct drm_i915_gem_object *obj = vma->obj; |
7bddb01f | 2637 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
43e28f09 | 2638 | int ret; |
673a394b | 2639 | |
07fe0b12 | 2640 | if (list_empty(&vma->vma_link)) |
673a394b EA |
2641 | return 0; |
2642 | ||
433544bd BW |
2643 | if (!drm_mm_node_allocated(&vma->node)) |
2644 | goto destroy; | |
2645 | ||
31d8d651 CW |
2646 | if (obj->pin_count) |
2647 | return -EBUSY; | |
673a394b | 2648 | |
c4670ad0 CW |
2649 | BUG_ON(obj->pages == NULL); |
2650 | ||
a8198eea | 2651 | ret = i915_gem_object_finish_gpu(obj); |
1488fc08 | 2652 | if (ret) |
a8198eea CW |
2653 | return ret; |
2654 | /* Continue on if we fail due to EIO, the GPU is hung so we | |
2655 | * should be safe and we need to cleanup or else we might | |
2656 | * cause memory corruption through use-after-free. | |
2657 | */ | |
2658 | ||
b5ffc9bc | 2659 | i915_gem_object_finish_gtt(obj); |
5323fd04 | 2660 | |
96b47b65 | 2661 | /* release the fence reg _after_ flushing */ |
d9e86c0e | 2662 | ret = i915_gem_object_put_fence(obj); |
1488fc08 | 2663 | if (ret) |
d9e86c0e | 2664 | return ret; |
96b47b65 | 2665 | |
07fe0b12 | 2666 | trace_i915_vma_unbind(vma); |
db53a302 | 2667 | |
74898d7e DV |
2668 | if (obj->has_global_gtt_mapping) |
2669 | i915_gem_gtt_unbind_object(obj); | |
7bddb01f DV |
2670 | if (obj->has_aliasing_ppgtt_mapping) { |
2671 | i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj); | |
2672 | obj->has_aliasing_ppgtt_mapping = 0; | |
2673 | } | |
74163907 | 2674 | i915_gem_gtt_finish_object(obj); |
401c29f6 | 2675 | i915_gem_object_unpin_pages(obj); |
7bddb01f | 2676 | |
ca191b13 | 2677 | list_del(&vma->mm_list); |
75e9e915 | 2678 | /* Avoid an unnecessary call to unbind on rebind. */ |
5cacaac7 BW |
2679 | if (i915_is_ggtt(vma->vm)) |
2680 | obj->map_and_fenceable = true; | |
673a394b | 2681 | |
2f633156 | 2682 | drm_mm_remove_node(&vma->node); |
433544bd BW |
2683 | |
2684 | destroy: | |
2f633156 BW |
2685 | i915_gem_vma_destroy(vma); |
2686 | ||
2687 | /* Since the unbound list is global, only move to that list if | |
2688 | * no more VMAs exist. | |
2689 | * NB: Until we have real VMAs there will only ever be one */ | |
2690 | WARN_ON(!list_empty(&obj->vma_list)); | |
2691 | if (list_empty(&obj->vma_list)) | |
2692 | list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list); | |
673a394b | 2693 | |
88241785 | 2694 | return 0; |
54cf91dc CW |
2695 | } |
2696 | ||
07fe0b12 BW |
2697 | /** |
2698 | * Unbinds an object from the global GTT aperture. | |
2699 | */ | |
2700 | int | |
2701 | i915_gem_object_ggtt_unbind(struct drm_i915_gem_object *obj) | |
2702 | { | |
2703 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
2704 | struct i915_address_space *ggtt = &dev_priv->gtt.base; | |
2705 | ||
58e73e15 | 2706 | if (!i915_gem_obj_ggtt_bound(obj)) |
07fe0b12 BW |
2707 | return 0; |
2708 | ||
2709 | if (obj->pin_count) | |
2710 | return -EBUSY; | |
2711 | ||
2712 | BUG_ON(obj->pages == NULL); | |
2713 | ||
2714 | return i915_vma_unbind(i915_gem_obj_to_vma(obj, ggtt)); | |
2715 | } | |
2716 | ||
b2da9fe5 | 2717 | int i915_gpu_idle(struct drm_device *dev) |
4df2faf4 DV |
2718 | { |
2719 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 2720 | struct intel_ring_buffer *ring; |
1ec14ad3 | 2721 | int ret, i; |
4df2faf4 | 2722 | |
4df2faf4 | 2723 | /* Flush everything onto the inactive list. */ |
b4519513 | 2724 | for_each_ring(ring, dev_priv, i) { |
b6c7488d BW |
2725 | ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID); |
2726 | if (ret) | |
2727 | return ret; | |
2728 | ||
3e960501 | 2729 | ret = intel_ring_idle(ring); |
1ec14ad3 CW |
2730 | if (ret) |
2731 | return ret; | |
2732 | } | |
4df2faf4 | 2733 | |
8a1a49f9 | 2734 | return 0; |
4df2faf4 DV |
2735 | } |
2736 | ||
9ce079e4 CW |
2737 | static void i965_write_fence_reg(struct drm_device *dev, int reg, |
2738 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2739 | { |
de151cf6 | 2740 | drm_i915_private_t *dev_priv = dev->dev_private; |
56c844e5 ID |
2741 | int fence_reg; |
2742 | int fence_pitch_shift; | |
de151cf6 | 2743 | |
56c844e5 ID |
2744 | if (INTEL_INFO(dev)->gen >= 6) { |
2745 | fence_reg = FENCE_REG_SANDYBRIDGE_0; | |
2746 | fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT; | |
2747 | } else { | |
2748 | fence_reg = FENCE_REG_965_0; | |
2749 | fence_pitch_shift = I965_FENCE_PITCH_SHIFT; | |
2750 | } | |
2751 | ||
d18b9619 CW |
2752 | fence_reg += reg * 8; |
2753 | ||
2754 | /* To w/a incoherency with non-atomic 64-bit register updates, | |
2755 | * we split the 64-bit update into two 32-bit writes. In order | |
2756 | * for a partial fence not to be evaluated between writes, we | |
2757 | * precede the update with write to turn off the fence register, | |
2758 | * and only enable the fence as the last step. | |
2759 | * | |
2760 | * For extra levels of paranoia, we make sure each step lands | |
2761 | * before applying the next step. | |
2762 | */ | |
2763 | I915_WRITE(fence_reg, 0); | |
2764 | POSTING_READ(fence_reg); | |
2765 | ||
9ce079e4 | 2766 | if (obj) { |
f343c5f6 | 2767 | u32 size = i915_gem_obj_ggtt_size(obj); |
d18b9619 | 2768 | uint64_t val; |
de151cf6 | 2769 | |
f343c5f6 | 2770 | val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) & |
9ce079e4 | 2771 | 0xfffff000) << 32; |
f343c5f6 | 2772 | val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000; |
56c844e5 | 2773 | val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift; |
9ce079e4 CW |
2774 | if (obj->tiling_mode == I915_TILING_Y) |
2775 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; | |
2776 | val |= I965_FENCE_REG_VALID; | |
c6642782 | 2777 | |
d18b9619 CW |
2778 | I915_WRITE(fence_reg + 4, val >> 32); |
2779 | POSTING_READ(fence_reg + 4); | |
2780 | ||
2781 | I915_WRITE(fence_reg + 0, val); | |
2782 | POSTING_READ(fence_reg); | |
2783 | } else { | |
2784 | I915_WRITE(fence_reg + 4, 0); | |
2785 | POSTING_READ(fence_reg + 4); | |
2786 | } | |
de151cf6 JB |
2787 | } |
2788 | ||
9ce079e4 CW |
2789 | static void i915_write_fence_reg(struct drm_device *dev, int reg, |
2790 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2791 | { |
de151cf6 | 2792 | drm_i915_private_t *dev_priv = dev->dev_private; |
9ce079e4 | 2793 | u32 val; |
de151cf6 | 2794 | |
9ce079e4 | 2795 | if (obj) { |
f343c5f6 | 2796 | u32 size = i915_gem_obj_ggtt_size(obj); |
9ce079e4 CW |
2797 | int pitch_val; |
2798 | int tile_width; | |
c6642782 | 2799 | |
f343c5f6 | 2800 | WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) || |
9ce079e4 | 2801 | (size & -size) != size || |
f343c5f6 BW |
2802 | (i915_gem_obj_ggtt_offset(obj) & (size - 1)), |
2803 | "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n", | |
2804 | i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size); | |
c6642782 | 2805 | |
9ce079e4 CW |
2806 | if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)) |
2807 | tile_width = 128; | |
2808 | else | |
2809 | tile_width = 512; | |
2810 | ||
2811 | /* Note: pitch better be a power of two tile widths */ | |
2812 | pitch_val = obj->stride / tile_width; | |
2813 | pitch_val = ffs(pitch_val) - 1; | |
2814 | ||
f343c5f6 | 2815 | val = i915_gem_obj_ggtt_offset(obj); |
9ce079e4 CW |
2816 | if (obj->tiling_mode == I915_TILING_Y) |
2817 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
2818 | val |= I915_FENCE_SIZE_BITS(size); | |
2819 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
2820 | val |= I830_FENCE_REG_VALID; | |
2821 | } else | |
2822 | val = 0; | |
2823 | ||
2824 | if (reg < 8) | |
2825 | reg = FENCE_REG_830_0 + reg * 4; | |
2826 | else | |
2827 | reg = FENCE_REG_945_8 + (reg - 8) * 4; | |
2828 | ||
2829 | I915_WRITE(reg, val); | |
2830 | POSTING_READ(reg); | |
de151cf6 JB |
2831 | } |
2832 | ||
9ce079e4 CW |
2833 | static void i830_write_fence_reg(struct drm_device *dev, int reg, |
2834 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2835 | { |
de151cf6 | 2836 | drm_i915_private_t *dev_priv = dev->dev_private; |
de151cf6 | 2837 | uint32_t val; |
de151cf6 | 2838 | |
9ce079e4 | 2839 | if (obj) { |
f343c5f6 | 2840 | u32 size = i915_gem_obj_ggtt_size(obj); |
9ce079e4 | 2841 | uint32_t pitch_val; |
de151cf6 | 2842 | |
f343c5f6 | 2843 | WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) || |
9ce079e4 | 2844 | (size & -size) != size || |
f343c5f6 BW |
2845 | (i915_gem_obj_ggtt_offset(obj) & (size - 1)), |
2846 | "object 0x%08lx not 512K or pot-size 0x%08x aligned\n", | |
2847 | i915_gem_obj_ggtt_offset(obj), size); | |
e76a16de | 2848 | |
9ce079e4 CW |
2849 | pitch_val = obj->stride / 128; |
2850 | pitch_val = ffs(pitch_val) - 1; | |
de151cf6 | 2851 | |
f343c5f6 | 2852 | val = i915_gem_obj_ggtt_offset(obj); |
9ce079e4 CW |
2853 | if (obj->tiling_mode == I915_TILING_Y) |
2854 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
2855 | val |= I830_FENCE_SIZE_BITS(size); | |
2856 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
2857 | val |= I830_FENCE_REG_VALID; | |
2858 | } else | |
2859 | val = 0; | |
c6642782 | 2860 | |
9ce079e4 CW |
2861 | I915_WRITE(FENCE_REG_830_0 + reg * 4, val); |
2862 | POSTING_READ(FENCE_REG_830_0 + reg * 4); | |
2863 | } | |
2864 | ||
d0a57789 CW |
2865 | inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj) |
2866 | { | |
2867 | return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT; | |
2868 | } | |
2869 | ||
9ce079e4 CW |
2870 | static void i915_gem_write_fence(struct drm_device *dev, int reg, |
2871 | struct drm_i915_gem_object *obj) | |
2872 | { | |
d0a57789 CW |
2873 | struct drm_i915_private *dev_priv = dev->dev_private; |
2874 | ||
2875 | /* Ensure that all CPU reads are completed before installing a fence | |
2876 | * and all writes before removing the fence. | |
2877 | */ | |
2878 | if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj)) | |
2879 | mb(); | |
2880 | ||
94a335db DV |
2881 | WARN(obj && (!obj->stride || !obj->tiling_mode), |
2882 | "bogus fence setup with stride: 0x%x, tiling mode: %i\n", | |
2883 | obj->stride, obj->tiling_mode); | |
2884 | ||
9ce079e4 CW |
2885 | switch (INTEL_INFO(dev)->gen) { |
2886 | case 7: | |
56c844e5 | 2887 | case 6: |
9ce079e4 CW |
2888 | case 5: |
2889 | case 4: i965_write_fence_reg(dev, reg, obj); break; | |
2890 | case 3: i915_write_fence_reg(dev, reg, obj); break; | |
2891 | case 2: i830_write_fence_reg(dev, reg, obj); break; | |
7dbf9d6e | 2892 | default: BUG(); |
9ce079e4 | 2893 | } |
d0a57789 CW |
2894 | |
2895 | /* And similarly be paranoid that no direct access to this region | |
2896 | * is reordered to before the fence is installed. | |
2897 | */ | |
2898 | if (i915_gem_object_needs_mb(obj)) | |
2899 | mb(); | |
de151cf6 JB |
2900 | } |
2901 | ||
61050808 CW |
2902 | static inline int fence_number(struct drm_i915_private *dev_priv, |
2903 | struct drm_i915_fence_reg *fence) | |
2904 | { | |
2905 | return fence - dev_priv->fence_regs; | |
2906 | } | |
2907 | ||
2908 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, | |
2909 | struct drm_i915_fence_reg *fence, | |
2910 | bool enable) | |
2911 | { | |
2dc8aae0 | 2912 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
46a0b638 CW |
2913 | int reg = fence_number(dev_priv, fence); |
2914 | ||
2915 | i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL); | |
61050808 CW |
2916 | |
2917 | if (enable) { | |
46a0b638 | 2918 | obj->fence_reg = reg; |
61050808 CW |
2919 | fence->obj = obj; |
2920 | list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list); | |
2921 | } else { | |
2922 | obj->fence_reg = I915_FENCE_REG_NONE; | |
2923 | fence->obj = NULL; | |
2924 | list_del_init(&fence->lru_list); | |
2925 | } | |
94a335db | 2926 | obj->fence_dirty = false; |
61050808 CW |
2927 | } |
2928 | ||
d9e86c0e | 2929 | static int |
d0a57789 | 2930 | i915_gem_object_wait_fence(struct drm_i915_gem_object *obj) |
d9e86c0e | 2931 | { |
1c293ea3 | 2932 | if (obj->last_fenced_seqno) { |
86d5bc37 | 2933 | int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno); |
18991845 CW |
2934 | if (ret) |
2935 | return ret; | |
d9e86c0e CW |
2936 | |
2937 | obj->last_fenced_seqno = 0; | |
d9e86c0e CW |
2938 | } |
2939 | ||
86d5bc37 | 2940 | obj->fenced_gpu_access = false; |
d9e86c0e CW |
2941 | return 0; |
2942 | } | |
2943 | ||
2944 | int | |
2945 | i915_gem_object_put_fence(struct drm_i915_gem_object *obj) | |
2946 | { | |
61050808 | 2947 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
f9c513e9 | 2948 | struct drm_i915_fence_reg *fence; |
d9e86c0e CW |
2949 | int ret; |
2950 | ||
d0a57789 | 2951 | ret = i915_gem_object_wait_fence(obj); |
d9e86c0e CW |
2952 | if (ret) |
2953 | return ret; | |
2954 | ||
61050808 CW |
2955 | if (obj->fence_reg == I915_FENCE_REG_NONE) |
2956 | return 0; | |
d9e86c0e | 2957 | |
f9c513e9 CW |
2958 | fence = &dev_priv->fence_regs[obj->fence_reg]; |
2959 | ||
61050808 | 2960 | i915_gem_object_fence_lost(obj); |
f9c513e9 | 2961 | i915_gem_object_update_fence(obj, fence, false); |
d9e86c0e CW |
2962 | |
2963 | return 0; | |
2964 | } | |
2965 | ||
2966 | static struct drm_i915_fence_reg * | |
a360bb1a | 2967 | i915_find_fence_reg(struct drm_device *dev) |
ae3db24a | 2968 | { |
ae3db24a | 2969 | struct drm_i915_private *dev_priv = dev->dev_private; |
8fe301ad | 2970 | struct drm_i915_fence_reg *reg, *avail; |
d9e86c0e | 2971 | int i; |
ae3db24a DV |
2972 | |
2973 | /* First try to find a free reg */ | |
d9e86c0e | 2974 | avail = NULL; |
ae3db24a DV |
2975 | for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) { |
2976 | reg = &dev_priv->fence_regs[i]; | |
2977 | if (!reg->obj) | |
d9e86c0e | 2978 | return reg; |
ae3db24a | 2979 | |
1690e1eb | 2980 | if (!reg->pin_count) |
d9e86c0e | 2981 | avail = reg; |
ae3db24a DV |
2982 | } |
2983 | ||
d9e86c0e CW |
2984 | if (avail == NULL) |
2985 | return NULL; | |
ae3db24a DV |
2986 | |
2987 | /* None available, try to steal one or wait for a user to finish */ | |
d9e86c0e | 2988 | list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) { |
1690e1eb | 2989 | if (reg->pin_count) |
ae3db24a DV |
2990 | continue; |
2991 | ||
8fe301ad | 2992 | return reg; |
ae3db24a DV |
2993 | } |
2994 | ||
8fe301ad | 2995 | return NULL; |
ae3db24a DV |
2996 | } |
2997 | ||
de151cf6 | 2998 | /** |
9a5a53b3 | 2999 | * i915_gem_object_get_fence - set up fencing for an object |
de151cf6 JB |
3000 | * @obj: object to map through a fence reg |
3001 | * | |
3002 | * When mapping objects through the GTT, userspace wants to be able to write | |
3003 | * to them without having to worry about swizzling if the object is tiled. | |
de151cf6 JB |
3004 | * This function walks the fence regs looking for a free one for @obj, |
3005 | * stealing one if it can't find any. | |
3006 | * | |
3007 | * It then sets up the reg based on the object's properties: address, pitch | |
3008 | * and tiling format. | |
9a5a53b3 CW |
3009 | * |
3010 | * For an untiled surface, this removes any existing fence. | |
de151cf6 | 3011 | */ |
8c4b8c3f | 3012 | int |
06d98131 | 3013 | i915_gem_object_get_fence(struct drm_i915_gem_object *obj) |
de151cf6 | 3014 | { |
05394f39 | 3015 | struct drm_device *dev = obj->base.dev; |
79e53945 | 3016 | struct drm_i915_private *dev_priv = dev->dev_private; |
14415745 | 3017 | bool enable = obj->tiling_mode != I915_TILING_NONE; |
d9e86c0e | 3018 | struct drm_i915_fence_reg *reg; |
ae3db24a | 3019 | int ret; |
de151cf6 | 3020 | |
14415745 CW |
3021 | /* Have we updated the tiling parameters upon the object and so |
3022 | * will need to serialise the write to the associated fence register? | |
3023 | */ | |
5d82e3e6 | 3024 | if (obj->fence_dirty) { |
d0a57789 | 3025 | ret = i915_gem_object_wait_fence(obj); |
14415745 CW |
3026 | if (ret) |
3027 | return ret; | |
3028 | } | |
9a5a53b3 | 3029 | |
d9e86c0e | 3030 | /* Just update our place in the LRU if our fence is getting reused. */ |
05394f39 CW |
3031 | if (obj->fence_reg != I915_FENCE_REG_NONE) { |
3032 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
5d82e3e6 | 3033 | if (!obj->fence_dirty) { |
14415745 CW |
3034 | list_move_tail(®->lru_list, |
3035 | &dev_priv->mm.fence_list); | |
3036 | return 0; | |
3037 | } | |
3038 | } else if (enable) { | |
3039 | reg = i915_find_fence_reg(dev); | |
3040 | if (reg == NULL) | |
3041 | return -EDEADLK; | |
d9e86c0e | 3042 | |
14415745 CW |
3043 | if (reg->obj) { |
3044 | struct drm_i915_gem_object *old = reg->obj; | |
3045 | ||
d0a57789 | 3046 | ret = i915_gem_object_wait_fence(old); |
29c5a587 CW |
3047 | if (ret) |
3048 | return ret; | |
3049 | ||
14415745 | 3050 | i915_gem_object_fence_lost(old); |
29c5a587 | 3051 | } |
14415745 | 3052 | } else |
a09ba7fa | 3053 | return 0; |
a09ba7fa | 3054 | |
14415745 | 3055 | i915_gem_object_update_fence(obj, reg, enable); |
14415745 | 3056 | |
9ce079e4 | 3057 | return 0; |
de151cf6 JB |
3058 | } |
3059 | ||
42d6ab48 CW |
3060 | static bool i915_gem_valid_gtt_space(struct drm_device *dev, |
3061 | struct drm_mm_node *gtt_space, | |
3062 | unsigned long cache_level) | |
3063 | { | |
3064 | struct drm_mm_node *other; | |
3065 | ||
3066 | /* On non-LLC machines we have to be careful when putting differing | |
3067 | * types of snoopable memory together to avoid the prefetcher | |
4239ca77 | 3068 | * crossing memory domains and dying. |
42d6ab48 CW |
3069 | */ |
3070 | if (HAS_LLC(dev)) | |
3071 | return true; | |
3072 | ||
c6cfb325 | 3073 | if (!drm_mm_node_allocated(gtt_space)) |
42d6ab48 CW |
3074 | return true; |
3075 | ||
3076 | if (list_empty(>t_space->node_list)) | |
3077 | return true; | |
3078 | ||
3079 | other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list); | |
3080 | if (other->allocated && !other->hole_follows && other->color != cache_level) | |
3081 | return false; | |
3082 | ||
3083 | other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list); | |
3084 | if (other->allocated && !gtt_space->hole_follows && other->color != cache_level) | |
3085 | return false; | |
3086 | ||
3087 | return true; | |
3088 | } | |
3089 | ||
3090 | static void i915_gem_verify_gtt(struct drm_device *dev) | |
3091 | { | |
3092 | #if WATCH_GTT | |
3093 | struct drm_i915_private *dev_priv = dev->dev_private; | |
3094 | struct drm_i915_gem_object *obj; | |
3095 | int err = 0; | |
3096 | ||
35c20a60 | 3097 | list_for_each_entry(obj, &dev_priv->mm.gtt_list, global_list) { |
42d6ab48 CW |
3098 | if (obj->gtt_space == NULL) { |
3099 | printk(KERN_ERR "object found on GTT list with no space reserved\n"); | |
3100 | err++; | |
3101 | continue; | |
3102 | } | |
3103 | ||
3104 | if (obj->cache_level != obj->gtt_space->color) { | |
3105 | printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n", | |
f343c5f6 BW |
3106 | i915_gem_obj_ggtt_offset(obj), |
3107 | i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj), | |
42d6ab48 CW |
3108 | obj->cache_level, |
3109 | obj->gtt_space->color); | |
3110 | err++; | |
3111 | continue; | |
3112 | } | |
3113 | ||
3114 | if (!i915_gem_valid_gtt_space(dev, | |
3115 | obj->gtt_space, | |
3116 | obj->cache_level)) { | |
3117 | printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n", | |
f343c5f6 BW |
3118 | i915_gem_obj_ggtt_offset(obj), |
3119 | i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj), | |
42d6ab48 CW |
3120 | obj->cache_level); |
3121 | err++; | |
3122 | continue; | |
3123 | } | |
3124 | } | |
3125 | ||
3126 | WARN_ON(err); | |
3127 | #endif | |
3128 | } | |
3129 | ||
673a394b EA |
3130 | /** |
3131 | * Finds free space in the GTT aperture and binds the object there. | |
3132 | */ | |
3133 | static int | |
07fe0b12 BW |
3134 | i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, |
3135 | struct i915_address_space *vm, | |
3136 | unsigned alignment, | |
3137 | bool map_and_fenceable, | |
3138 | bool nonblocking) | |
673a394b | 3139 | { |
05394f39 | 3140 | struct drm_device *dev = obj->base.dev; |
673a394b | 3141 | drm_i915_private_t *dev_priv = dev->dev_private; |
5e783301 | 3142 | u32 size, fence_size, fence_alignment, unfenced_alignment; |
07fe0b12 BW |
3143 | size_t gtt_max = |
3144 | map_and_fenceable ? dev_priv->gtt.mappable_end : vm->total; | |
2f633156 | 3145 | struct i915_vma *vma; |
07f73f69 | 3146 | int ret; |
673a394b | 3147 | |
e28f8711 CW |
3148 | fence_size = i915_gem_get_gtt_size(dev, |
3149 | obj->base.size, | |
3150 | obj->tiling_mode); | |
3151 | fence_alignment = i915_gem_get_gtt_alignment(dev, | |
3152 | obj->base.size, | |
d865110c | 3153 | obj->tiling_mode, true); |
e28f8711 | 3154 | unfenced_alignment = |
d865110c | 3155 | i915_gem_get_gtt_alignment(dev, |
e28f8711 | 3156 | obj->base.size, |
d865110c | 3157 | obj->tiling_mode, false); |
a00b10c3 | 3158 | |
673a394b | 3159 | if (alignment == 0) |
5e783301 DV |
3160 | alignment = map_and_fenceable ? fence_alignment : |
3161 | unfenced_alignment; | |
75e9e915 | 3162 | if (map_and_fenceable && alignment & (fence_alignment - 1)) { |
673a394b EA |
3163 | DRM_ERROR("Invalid object alignment requested %u\n", alignment); |
3164 | return -EINVAL; | |
3165 | } | |
3166 | ||
05394f39 | 3167 | size = map_and_fenceable ? fence_size : obj->base.size; |
a00b10c3 | 3168 | |
654fc607 CW |
3169 | /* If the object is bigger than the entire aperture, reject it early |
3170 | * before evicting everything in a vain attempt to find space. | |
3171 | */ | |
0a9ae0d7 | 3172 | if (obj->base.size > gtt_max) { |
3765f304 | 3173 | DRM_ERROR("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%zu\n", |
a36689cb CW |
3174 | obj->base.size, |
3175 | map_and_fenceable ? "mappable" : "total", | |
0a9ae0d7 | 3176 | gtt_max); |
654fc607 CW |
3177 | return -E2BIG; |
3178 | } | |
3179 | ||
37e680a1 | 3180 | ret = i915_gem_object_get_pages(obj); |
6c085a72 CW |
3181 | if (ret) |
3182 | return ret; | |
3183 | ||
fbdda6fb CW |
3184 | i915_gem_object_pin_pages(obj); |
3185 | ||
07fe0b12 | 3186 | BUG_ON(!i915_is_ggtt(vm)); |
07fe0b12 | 3187 | |
accfef2e | 3188 | vma = i915_gem_obj_lookup_or_create_vma(obj, vm); |
db473b36 | 3189 | if (IS_ERR(vma)) { |
bc6bc15b DV |
3190 | ret = PTR_ERR(vma); |
3191 | goto err_unpin; | |
2f633156 BW |
3192 | } |
3193 | ||
accfef2e BW |
3194 | /* For now we only ever use 1 vma per object */ |
3195 | WARN_ON(!list_is_singular(&obj->vma_list)); | |
3196 | ||
0a9ae0d7 | 3197 | search_free: |
07fe0b12 | 3198 | ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node, |
0a9ae0d7 | 3199 | size, alignment, |
31e5d7c6 DH |
3200 | obj->cache_level, 0, gtt_max, |
3201 | DRM_MM_SEARCH_DEFAULT); | |
dc9dd7a2 | 3202 | if (ret) { |
f6cd1f15 | 3203 | ret = i915_gem_evict_something(dev, vm, size, alignment, |
42d6ab48 | 3204 | obj->cache_level, |
86a1ee26 CW |
3205 | map_and_fenceable, |
3206 | nonblocking); | |
dc9dd7a2 CW |
3207 | if (ret == 0) |
3208 | goto search_free; | |
9731129c | 3209 | |
bc6bc15b | 3210 | goto err_free_vma; |
673a394b | 3211 | } |
2f633156 | 3212 | if (WARN_ON(!i915_gem_valid_gtt_space(dev, &vma->node, |
c6cfb325 | 3213 | obj->cache_level))) { |
2f633156 | 3214 | ret = -EINVAL; |
bc6bc15b | 3215 | goto err_remove_node; |
673a394b EA |
3216 | } |
3217 | ||
74163907 | 3218 | ret = i915_gem_gtt_prepare_object(obj); |
2f633156 | 3219 | if (ret) |
bc6bc15b | 3220 | goto err_remove_node; |
673a394b | 3221 | |
35c20a60 | 3222 | list_move_tail(&obj->global_list, &dev_priv->mm.bound_list); |
ca191b13 | 3223 | list_add_tail(&vma->mm_list, &vm->inactive_list); |
bf1a1092 | 3224 | |
4bd561b3 BW |
3225 | if (i915_is_ggtt(vm)) { |
3226 | bool mappable, fenceable; | |
a00b10c3 | 3227 | |
49987099 DV |
3228 | fenceable = (vma->node.size == fence_size && |
3229 | (vma->node.start & (fence_alignment - 1)) == 0); | |
4bd561b3 | 3230 | |
49987099 DV |
3231 | mappable = (vma->node.start + obj->base.size <= |
3232 | dev_priv->gtt.mappable_end); | |
a00b10c3 | 3233 | |
5cacaac7 | 3234 | obj->map_and_fenceable = mappable && fenceable; |
4bd561b3 | 3235 | } |
75e9e915 | 3236 | |
7ace7ef2 | 3237 | WARN_ON(map_and_fenceable && !obj->map_and_fenceable); |
75e9e915 | 3238 | |
07fe0b12 | 3239 | trace_i915_vma_bind(vma, map_and_fenceable); |
42d6ab48 | 3240 | i915_gem_verify_gtt(dev); |
673a394b | 3241 | return 0; |
2f633156 | 3242 | |
bc6bc15b | 3243 | err_remove_node: |
6286ef9b | 3244 | drm_mm_remove_node(&vma->node); |
bc6bc15b | 3245 | err_free_vma: |
2f633156 | 3246 | i915_gem_vma_destroy(vma); |
bc6bc15b | 3247 | err_unpin: |
2f633156 | 3248 | i915_gem_object_unpin_pages(obj); |
2f633156 | 3249 | return ret; |
673a394b EA |
3250 | } |
3251 | ||
000433b6 | 3252 | bool |
2c22569b CW |
3253 | i915_gem_clflush_object(struct drm_i915_gem_object *obj, |
3254 | bool force) | |
673a394b | 3255 | { |
673a394b EA |
3256 | /* If we don't have a page list set up, then we're not pinned |
3257 | * to GPU, and we can ignore the cache flush because it'll happen | |
3258 | * again at bind time. | |
3259 | */ | |
05394f39 | 3260 | if (obj->pages == NULL) |
000433b6 | 3261 | return false; |
673a394b | 3262 | |
769ce464 ID |
3263 | /* |
3264 | * Stolen memory is always coherent with the GPU as it is explicitly | |
3265 | * marked as wc by the system, or the system is cache-coherent. | |
3266 | */ | |
3267 | if (obj->stolen) | |
000433b6 | 3268 | return false; |
769ce464 | 3269 | |
9c23f7fc CW |
3270 | /* If the GPU is snooping the contents of the CPU cache, |
3271 | * we do not need to manually clear the CPU cache lines. However, | |
3272 | * the caches are only snooped when the render cache is | |
3273 | * flushed/invalidated. As we always have to emit invalidations | |
3274 | * and flushes when moving into and out of the RENDER domain, correct | |
3275 | * snooping behaviour occurs naturally as the result of our domain | |
3276 | * tracking. | |
3277 | */ | |
2c22569b | 3278 | if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level)) |
000433b6 | 3279 | return false; |
9c23f7fc | 3280 | |
1c5d22f7 | 3281 | trace_i915_gem_object_clflush(obj); |
9da3da66 | 3282 | drm_clflush_sg(obj->pages); |
000433b6 CW |
3283 | |
3284 | return true; | |
e47c68e9 EA |
3285 | } |
3286 | ||
3287 | /** Flushes the GTT write domain for the object if it's dirty. */ | |
3288 | static void | |
05394f39 | 3289 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 3290 | { |
1c5d22f7 CW |
3291 | uint32_t old_write_domain; |
3292 | ||
05394f39 | 3293 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
e47c68e9 EA |
3294 | return; |
3295 | ||
63256ec5 | 3296 | /* No actual flushing is required for the GTT write domain. Writes |
e47c68e9 EA |
3297 | * to it immediately go to main memory as far as we know, so there's |
3298 | * no chipset flush. It also doesn't land in render cache. | |
63256ec5 CW |
3299 | * |
3300 | * However, we do have to enforce the order so that all writes through | |
3301 | * the GTT land before any writes to the device, such as updates to | |
3302 | * the GATT itself. | |
e47c68e9 | 3303 | */ |
63256ec5 CW |
3304 | wmb(); |
3305 | ||
05394f39 CW |
3306 | old_write_domain = obj->base.write_domain; |
3307 | obj->base.write_domain = 0; | |
1c5d22f7 CW |
3308 | |
3309 | trace_i915_gem_object_change_domain(obj, | |
05394f39 | 3310 | obj->base.read_domains, |
1c5d22f7 | 3311 | old_write_domain); |
e47c68e9 EA |
3312 | } |
3313 | ||
3314 | /** Flushes the CPU write domain for the object if it's dirty. */ | |
3315 | static void | |
2c22569b CW |
3316 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj, |
3317 | bool force) | |
e47c68e9 | 3318 | { |
1c5d22f7 | 3319 | uint32_t old_write_domain; |
e47c68e9 | 3320 | |
05394f39 | 3321 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
e47c68e9 EA |
3322 | return; |
3323 | ||
000433b6 CW |
3324 | if (i915_gem_clflush_object(obj, force)) |
3325 | i915_gem_chipset_flush(obj->base.dev); | |
3326 | ||
05394f39 CW |
3327 | old_write_domain = obj->base.write_domain; |
3328 | obj->base.write_domain = 0; | |
1c5d22f7 CW |
3329 | |
3330 | trace_i915_gem_object_change_domain(obj, | |
05394f39 | 3331 | obj->base.read_domains, |
1c5d22f7 | 3332 | old_write_domain); |
e47c68e9 EA |
3333 | } |
3334 | ||
2ef7eeaa EA |
3335 | /** |
3336 | * Moves a single object to the GTT read, and possibly write domain. | |
3337 | * | |
3338 | * This function returns when the move is complete, including waiting on | |
3339 | * flushes to occur. | |
3340 | */ | |
79e53945 | 3341 | int |
2021746e | 3342 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
2ef7eeaa | 3343 | { |
8325a09d | 3344 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
1c5d22f7 | 3345 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 | 3346 | int ret; |
2ef7eeaa | 3347 | |
02354392 | 3348 | /* Not valid to be called on unbound objects. */ |
9843877d | 3349 | if (!i915_gem_obj_bound_any(obj)) |
02354392 EA |
3350 | return -EINVAL; |
3351 | ||
8d7e3de1 CW |
3352 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
3353 | return 0; | |
3354 | ||
0201f1ec | 3355 | ret = i915_gem_object_wait_rendering(obj, !write); |
88241785 CW |
3356 | if (ret) |
3357 | return ret; | |
3358 | ||
2c22569b | 3359 | i915_gem_object_flush_cpu_write_domain(obj, false); |
1c5d22f7 | 3360 | |
d0a57789 CW |
3361 | /* Serialise direct access to this object with the barriers for |
3362 | * coherent writes from the GPU, by effectively invalidating the | |
3363 | * GTT domain upon first access. | |
3364 | */ | |
3365 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) | |
3366 | mb(); | |
3367 | ||
05394f39 CW |
3368 | old_write_domain = obj->base.write_domain; |
3369 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 3370 | |
e47c68e9 EA |
3371 | /* It should now be out of any other write domains, and we can update |
3372 | * the domain values for our changes. | |
3373 | */ | |
05394f39 CW |
3374 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
3375 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; | |
e47c68e9 | 3376 | if (write) { |
05394f39 CW |
3377 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
3378 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; | |
3379 | obj->dirty = 1; | |
2ef7eeaa EA |
3380 | } |
3381 | ||
1c5d22f7 CW |
3382 | trace_i915_gem_object_change_domain(obj, |
3383 | old_read_domains, | |
3384 | old_write_domain); | |
3385 | ||
8325a09d | 3386 | /* And bump the LRU for this access */ |
ca191b13 BW |
3387 | if (i915_gem_object_is_inactive(obj)) { |
3388 | struct i915_vma *vma = i915_gem_obj_to_vma(obj, | |
3389 | &dev_priv->gtt.base); | |
3390 | if (vma) | |
3391 | list_move_tail(&vma->mm_list, | |
3392 | &dev_priv->gtt.base.inactive_list); | |
3393 | ||
3394 | } | |
8325a09d | 3395 | |
e47c68e9 EA |
3396 | return 0; |
3397 | } | |
3398 | ||
e4ffd173 CW |
3399 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
3400 | enum i915_cache_level cache_level) | |
3401 | { | |
7bddb01f DV |
3402 | struct drm_device *dev = obj->base.dev; |
3403 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3089c6f2 | 3404 | struct i915_vma *vma; |
e4ffd173 CW |
3405 | int ret; |
3406 | ||
3407 | if (obj->cache_level == cache_level) | |
3408 | return 0; | |
3409 | ||
3410 | if (obj->pin_count) { | |
3411 | DRM_DEBUG("can not change the cache level of pinned objects\n"); | |
3412 | return -EBUSY; | |
3413 | } | |
3414 | ||
3089c6f2 BW |
3415 | list_for_each_entry(vma, &obj->vma_list, vma_link) { |
3416 | if (!i915_gem_valid_gtt_space(dev, &vma->node, cache_level)) { | |
07fe0b12 | 3417 | ret = i915_vma_unbind(vma); |
3089c6f2 BW |
3418 | if (ret) |
3419 | return ret; | |
3420 | ||
3421 | break; | |
3422 | } | |
42d6ab48 CW |
3423 | } |
3424 | ||
3089c6f2 | 3425 | if (i915_gem_obj_bound_any(obj)) { |
e4ffd173 CW |
3426 | ret = i915_gem_object_finish_gpu(obj); |
3427 | if (ret) | |
3428 | return ret; | |
3429 | ||
3430 | i915_gem_object_finish_gtt(obj); | |
3431 | ||
3432 | /* Before SandyBridge, you could not use tiling or fence | |
3433 | * registers with snooped memory, so relinquish any fences | |
3434 | * currently pointing to our region in the aperture. | |
3435 | */ | |
42d6ab48 | 3436 | if (INTEL_INFO(dev)->gen < 6) { |
e4ffd173 CW |
3437 | ret = i915_gem_object_put_fence(obj); |
3438 | if (ret) | |
3439 | return ret; | |
3440 | } | |
3441 | ||
74898d7e DV |
3442 | if (obj->has_global_gtt_mapping) |
3443 | i915_gem_gtt_bind_object(obj, cache_level); | |
7bddb01f DV |
3444 | if (obj->has_aliasing_ppgtt_mapping) |
3445 | i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt, | |
3446 | obj, cache_level); | |
e4ffd173 CW |
3447 | } |
3448 | ||
2c22569b CW |
3449 | list_for_each_entry(vma, &obj->vma_list, vma_link) |
3450 | vma->node.color = cache_level; | |
3451 | obj->cache_level = cache_level; | |
3452 | ||
3453 | if (cpu_write_needs_clflush(obj)) { | |
e4ffd173 CW |
3454 | u32 old_read_domains, old_write_domain; |
3455 | ||
3456 | /* If we're coming from LLC cached, then we haven't | |
3457 | * actually been tracking whether the data is in the | |
3458 | * CPU cache or not, since we only allow one bit set | |
3459 | * in obj->write_domain and have been skipping the clflushes. | |
3460 | * Just set it to the CPU cache for now. | |
3461 | */ | |
3462 | WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU); | |
e4ffd173 CW |
3463 | |
3464 | old_read_domains = obj->base.read_domains; | |
3465 | old_write_domain = obj->base.write_domain; | |
3466 | ||
3467 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
3468 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
3469 | ||
3470 | trace_i915_gem_object_change_domain(obj, | |
3471 | old_read_domains, | |
3472 | old_write_domain); | |
3473 | } | |
3474 | ||
42d6ab48 | 3475 | i915_gem_verify_gtt(dev); |
e4ffd173 CW |
3476 | return 0; |
3477 | } | |
3478 | ||
199adf40 BW |
3479 | int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, |
3480 | struct drm_file *file) | |
e6994aee | 3481 | { |
199adf40 | 3482 | struct drm_i915_gem_caching *args = data; |
e6994aee CW |
3483 | struct drm_i915_gem_object *obj; |
3484 | int ret; | |
3485 | ||
3486 | ret = i915_mutex_lock_interruptible(dev); | |
3487 | if (ret) | |
3488 | return ret; | |
3489 | ||
3490 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
3491 | if (&obj->base == NULL) { | |
3492 | ret = -ENOENT; | |
3493 | goto unlock; | |
3494 | } | |
3495 | ||
651d794f CW |
3496 | switch (obj->cache_level) { |
3497 | case I915_CACHE_LLC: | |
3498 | case I915_CACHE_L3_LLC: | |
3499 | args->caching = I915_CACHING_CACHED; | |
3500 | break; | |
3501 | ||
4257d3ba CW |
3502 | case I915_CACHE_WT: |
3503 | args->caching = I915_CACHING_DISPLAY; | |
3504 | break; | |
3505 | ||
651d794f CW |
3506 | default: |
3507 | args->caching = I915_CACHING_NONE; | |
3508 | break; | |
3509 | } | |
e6994aee CW |
3510 | |
3511 | drm_gem_object_unreference(&obj->base); | |
3512 | unlock: | |
3513 | mutex_unlock(&dev->struct_mutex); | |
3514 | return ret; | |
3515 | } | |
3516 | ||
199adf40 BW |
3517 | int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data, |
3518 | struct drm_file *file) | |
e6994aee | 3519 | { |
199adf40 | 3520 | struct drm_i915_gem_caching *args = data; |
e6994aee CW |
3521 | struct drm_i915_gem_object *obj; |
3522 | enum i915_cache_level level; | |
3523 | int ret; | |
3524 | ||
199adf40 BW |
3525 | switch (args->caching) { |
3526 | case I915_CACHING_NONE: | |
e6994aee CW |
3527 | level = I915_CACHE_NONE; |
3528 | break; | |
199adf40 | 3529 | case I915_CACHING_CACHED: |
e6994aee CW |
3530 | level = I915_CACHE_LLC; |
3531 | break; | |
4257d3ba CW |
3532 | case I915_CACHING_DISPLAY: |
3533 | level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE; | |
3534 | break; | |
e6994aee CW |
3535 | default: |
3536 | return -EINVAL; | |
3537 | } | |
3538 | ||
3bc2913e BW |
3539 | ret = i915_mutex_lock_interruptible(dev); |
3540 | if (ret) | |
3541 | return ret; | |
3542 | ||
e6994aee CW |
3543 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
3544 | if (&obj->base == NULL) { | |
3545 | ret = -ENOENT; | |
3546 | goto unlock; | |
3547 | } | |
3548 | ||
3549 | ret = i915_gem_object_set_cache_level(obj, level); | |
3550 | ||
3551 | drm_gem_object_unreference(&obj->base); | |
3552 | unlock: | |
3553 | mutex_unlock(&dev->struct_mutex); | |
3554 | return ret; | |
3555 | } | |
3556 | ||
cc98b413 CW |
3557 | static bool is_pin_display(struct drm_i915_gem_object *obj) |
3558 | { | |
3559 | /* There are 3 sources that pin objects: | |
3560 | * 1. The display engine (scanouts, sprites, cursors); | |
3561 | * 2. Reservations for execbuffer; | |
3562 | * 3. The user. | |
3563 | * | |
3564 | * We can ignore reservations as we hold the struct_mutex and | |
3565 | * are only called outside of the reservation path. The user | |
3566 | * can only increment pin_count once, and so if after | |
3567 | * subtracting the potential reference by the user, any pin_count | |
3568 | * remains, it must be due to another use by the display engine. | |
3569 | */ | |
3570 | return obj->pin_count - !!obj->user_pin_count; | |
3571 | } | |
3572 | ||
b9241ea3 | 3573 | /* |
2da3b9b9 CW |
3574 | * Prepare buffer for display plane (scanout, cursors, etc). |
3575 | * Can be called from an uninterruptible phase (modesetting) and allows | |
3576 | * any flushes to be pipelined (for pageflips). | |
b9241ea3 ZW |
3577 | */ |
3578 | int | |
2da3b9b9 CW |
3579 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
3580 | u32 alignment, | |
919926ae | 3581 | struct intel_ring_buffer *pipelined) |
b9241ea3 | 3582 | { |
2da3b9b9 | 3583 | u32 old_read_domains, old_write_domain; |
b9241ea3 ZW |
3584 | int ret; |
3585 | ||
0be73284 | 3586 | if (pipelined != obj->ring) { |
2911a35b BW |
3587 | ret = i915_gem_object_sync(obj, pipelined); |
3588 | if (ret) | |
b9241ea3 ZW |
3589 | return ret; |
3590 | } | |
3591 | ||
cc98b413 CW |
3592 | /* Mark the pin_display early so that we account for the |
3593 | * display coherency whilst setting up the cache domains. | |
3594 | */ | |
3595 | obj->pin_display = true; | |
3596 | ||
a7ef0640 EA |
3597 | /* The display engine is not coherent with the LLC cache on gen6. As |
3598 | * a result, we make sure that the pinning that is about to occur is | |
3599 | * done with uncached PTEs. This is lowest common denominator for all | |
3600 | * chipsets. | |
3601 | * | |
3602 | * However for gen6+, we could do better by using the GFDT bit instead | |
3603 | * of uncaching, which would allow us to flush all the LLC-cached data | |
3604 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. | |
3605 | */ | |
651d794f CW |
3606 | ret = i915_gem_object_set_cache_level(obj, |
3607 | HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE); | |
a7ef0640 | 3608 | if (ret) |
cc98b413 | 3609 | goto err_unpin_display; |
a7ef0640 | 3610 | |
2da3b9b9 CW |
3611 | /* As the user may map the buffer once pinned in the display plane |
3612 | * (e.g. libkms for the bootup splash), we have to ensure that we | |
3613 | * always use map_and_fenceable for all scanout buffers. | |
3614 | */ | |
c37e2204 | 3615 | ret = i915_gem_obj_ggtt_pin(obj, alignment, true, false); |
2da3b9b9 | 3616 | if (ret) |
cc98b413 | 3617 | goto err_unpin_display; |
2da3b9b9 | 3618 | |
2c22569b | 3619 | i915_gem_object_flush_cpu_write_domain(obj, true); |
b118c1e3 | 3620 | |
2da3b9b9 | 3621 | old_write_domain = obj->base.write_domain; |
05394f39 | 3622 | old_read_domains = obj->base.read_domains; |
2da3b9b9 CW |
3623 | |
3624 | /* It should now be out of any other write domains, and we can update | |
3625 | * the domain values for our changes. | |
3626 | */ | |
e5f1d962 | 3627 | obj->base.write_domain = 0; |
05394f39 | 3628 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
b9241ea3 ZW |
3629 | |
3630 | trace_i915_gem_object_change_domain(obj, | |
3631 | old_read_domains, | |
2da3b9b9 | 3632 | old_write_domain); |
b9241ea3 ZW |
3633 | |
3634 | return 0; | |
cc98b413 CW |
3635 | |
3636 | err_unpin_display: | |
3637 | obj->pin_display = is_pin_display(obj); | |
3638 | return ret; | |
3639 | } | |
3640 | ||
3641 | void | |
3642 | i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj) | |
3643 | { | |
3644 | i915_gem_object_unpin(obj); | |
3645 | obj->pin_display = is_pin_display(obj); | |
b9241ea3 ZW |
3646 | } |
3647 | ||
85345517 | 3648 | int |
a8198eea | 3649 | i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj) |
85345517 | 3650 | { |
88241785 CW |
3651 | int ret; |
3652 | ||
a8198eea | 3653 | if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0) |
85345517 CW |
3654 | return 0; |
3655 | ||
0201f1ec | 3656 | ret = i915_gem_object_wait_rendering(obj, false); |
c501ae7f CW |
3657 | if (ret) |
3658 | return ret; | |
3659 | ||
a8198eea CW |
3660 | /* Ensure that we invalidate the GPU's caches and TLBs. */ |
3661 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; | |
c501ae7f | 3662 | return 0; |
85345517 CW |
3663 | } |
3664 | ||
e47c68e9 EA |
3665 | /** |
3666 | * Moves a single object to the CPU read, and possibly write domain. | |
3667 | * | |
3668 | * This function returns when the move is complete, including waiting on | |
3669 | * flushes to occur. | |
3670 | */ | |
dabdfe02 | 3671 | int |
919926ae | 3672 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
e47c68e9 | 3673 | { |
1c5d22f7 | 3674 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 EA |
3675 | int ret; |
3676 | ||
8d7e3de1 CW |
3677 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3678 | return 0; | |
3679 | ||
0201f1ec | 3680 | ret = i915_gem_object_wait_rendering(obj, !write); |
88241785 CW |
3681 | if (ret) |
3682 | return ret; | |
3683 | ||
e47c68e9 | 3684 | i915_gem_object_flush_gtt_write_domain(obj); |
2ef7eeaa | 3685 | |
05394f39 CW |
3686 | old_write_domain = obj->base.write_domain; |
3687 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 3688 | |
e47c68e9 | 3689 | /* Flush the CPU cache if it's still invalid. */ |
05394f39 | 3690 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
2c22569b | 3691 | i915_gem_clflush_object(obj, false); |
2ef7eeaa | 3692 | |
05394f39 | 3693 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
2ef7eeaa EA |
3694 | } |
3695 | ||
3696 | /* It should now be out of any other write domains, and we can update | |
3697 | * the domain values for our changes. | |
3698 | */ | |
05394f39 | 3699 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
e47c68e9 EA |
3700 | |
3701 | /* If we're writing through the CPU, then the GPU read domains will | |
3702 | * need to be invalidated at next use. | |
3703 | */ | |
3704 | if (write) { | |
05394f39 CW |
3705 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3706 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
e47c68e9 | 3707 | } |
2ef7eeaa | 3708 | |
1c5d22f7 CW |
3709 | trace_i915_gem_object_change_domain(obj, |
3710 | old_read_domains, | |
3711 | old_write_domain); | |
3712 | ||
2ef7eeaa EA |
3713 | return 0; |
3714 | } | |
3715 | ||
673a394b EA |
3716 | /* Throttle our rendering by waiting until the ring has completed our requests |
3717 | * emitted over 20 msec ago. | |
3718 | * | |
b962442e EA |
3719 | * Note that if we were to use the current jiffies each time around the loop, |
3720 | * we wouldn't escape the function with any frames outstanding if the time to | |
3721 | * render a frame was over 20ms. | |
3722 | * | |
673a394b EA |
3723 | * This should get us reasonable parallelism between CPU and GPU but also |
3724 | * relatively low latency when blocking on a particular request to finish. | |
3725 | */ | |
40a5f0de | 3726 | static int |
f787a5f5 | 3727 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
40a5f0de | 3728 | { |
f787a5f5 CW |
3729 | struct drm_i915_private *dev_priv = dev->dev_private; |
3730 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
b962442e | 3731 | unsigned long recent_enough = jiffies - msecs_to_jiffies(20); |
f787a5f5 CW |
3732 | struct drm_i915_gem_request *request; |
3733 | struct intel_ring_buffer *ring = NULL; | |
f69061be | 3734 | unsigned reset_counter; |
f787a5f5 CW |
3735 | u32 seqno = 0; |
3736 | int ret; | |
93533c29 | 3737 | |
308887aa DV |
3738 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
3739 | if (ret) | |
3740 | return ret; | |
3741 | ||
3742 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, false); | |
3743 | if (ret) | |
3744 | return ret; | |
e110e8d6 | 3745 | |
1c25595f | 3746 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 3747 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { |
b962442e EA |
3748 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
3749 | break; | |
40a5f0de | 3750 | |
f787a5f5 CW |
3751 | ring = request->ring; |
3752 | seqno = request->seqno; | |
b962442e | 3753 | } |
f69061be | 3754 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); |
1c25595f | 3755 | spin_unlock(&file_priv->mm.lock); |
40a5f0de | 3756 | |
f787a5f5 CW |
3757 | if (seqno == 0) |
3758 | return 0; | |
2bc43b5c | 3759 | |
f69061be | 3760 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL); |
f787a5f5 CW |
3761 | if (ret == 0) |
3762 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); | |
40a5f0de EA |
3763 | |
3764 | return ret; | |
3765 | } | |
3766 | ||
673a394b | 3767 | int |
05394f39 | 3768 | i915_gem_object_pin(struct drm_i915_gem_object *obj, |
c37e2204 | 3769 | struct i915_address_space *vm, |
05394f39 | 3770 | uint32_t alignment, |
86a1ee26 CW |
3771 | bool map_and_fenceable, |
3772 | bool nonblocking) | |
673a394b | 3773 | { |
07fe0b12 | 3774 | struct i915_vma *vma; |
673a394b EA |
3775 | int ret; |
3776 | ||
7e81a42e CW |
3777 | if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT)) |
3778 | return -EBUSY; | |
ac0c6b5a | 3779 | |
07fe0b12 BW |
3780 | WARN_ON(map_and_fenceable && !i915_is_ggtt(vm)); |
3781 | ||
3782 | vma = i915_gem_obj_to_vma(obj, vm); | |
3783 | ||
3784 | if (vma) { | |
3785 | if ((alignment && | |
3786 | vma->node.start & (alignment - 1)) || | |
05394f39 CW |
3787 | (map_and_fenceable && !obj->map_and_fenceable)) { |
3788 | WARN(obj->pin_count, | |
ae7d49d8 | 3789 | "bo is already pinned with incorrect alignment:" |
f343c5f6 | 3790 | " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d," |
75e9e915 | 3791 | " obj->map_and_fenceable=%d\n", |
07fe0b12 | 3792 | i915_gem_obj_offset(obj, vm), alignment, |
75e9e915 | 3793 | map_and_fenceable, |
05394f39 | 3794 | obj->map_and_fenceable); |
07fe0b12 | 3795 | ret = i915_vma_unbind(vma); |
ac0c6b5a CW |
3796 | if (ret) |
3797 | return ret; | |
3798 | } | |
3799 | } | |
3800 | ||
07fe0b12 | 3801 | if (!i915_gem_obj_bound(obj, vm)) { |
8742267a CW |
3802 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
3803 | ||
07fe0b12 BW |
3804 | ret = i915_gem_object_bind_to_vm(obj, vm, alignment, |
3805 | map_and_fenceable, | |
3806 | nonblocking); | |
9731129c | 3807 | if (ret) |
673a394b | 3808 | return ret; |
8742267a CW |
3809 | |
3810 | if (!dev_priv->mm.aliasing_ppgtt) | |
3811 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
22c344e9 | 3812 | } |
76446cac | 3813 | |
74898d7e DV |
3814 | if (!obj->has_global_gtt_mapping && map_and_fenceable) |
3815 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
3816 | ||
1b50247a | 3817 | obj->pin_count++; |
6299f992 | 3818 | obj->pin_mappable |= map_and_fenceable; |
673a394b EA |
3819 | |
3820 | return 0; | |
3821 | } | |
3822 | ||
3823 | void | |
05394f39 | 3824 | i915_gem_object_unpin(struct drm_i915_gem_object *obj) |
673a394b | 3825 | { |
05394f39 | 3826 | BUG_ON(obj->pin_count == 0); |
9843877d | 3827 | BUG_ON(!i915_gem_obj_bound_any(obj)); |
673a394b | 3828 | |
1b50247a | 3829 | if (--obj->pin_count == 0) |
6299f992 | 3830 | obj->pin_mappable = false; |
673a394b EA |
3831 | } |
3832 | ||
3833 | int | |
3834 | i915_gem_pin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3835 | struct drm_file *file) |
673a394b EA |
3836 | { |
3837 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 3838 | struct drm_i915_gem_object *obj; |
673a394b EA |
3839 | int ret; |
3840 | ||
1d7cfea1 CW |
3841 | ret = i915_mutex_lock_interruptible(dev); |
3842 | if (ret) | |
3843 | return ret; | |
673a394b | 3844 | |
05394f39 | 3845 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3846 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3847 | ret = -ENOENT; |
3848 | goto unlock; | |
673a394b | 3849 | } |
673a394b | 3850 | |
05394f39 | 3851 | if (obj->madv != I915_MADV_WILLNEED) { |
bb6baf76 | 3852 | DRM_ERROR("Attempting to pin a purgeable buffer\n"); |
1d7cfea1 CW |
3853 | ret = -EINVAL; |
3854 | goto out; | |
3ef94daa CW |
3855 | } |
3856 | ||
05394f39 | 3857 | if (obj->pin_filp != NULL && obj->pin_filp != file) { |
79e53945 JB |
3858 | DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n", |
3859 | args->handle); | |
1d7cfea1 CW |
3860 | ret = -EINVAL; |
3861 | goto out; | |
79e53945 JB |
3862 | } |
3863 | ||
93be8788 | 3864 | if (obj->user_pin_count == 0) { |
c37e2204 | 3865 | ret = i915_gem_obj_ggtt_pin(obj, args->alignment, true, false); |
1d7cfea1 CW |
3866 | if (ret) |
3867 | goto out; | |
673a394b EA |
3868 | } |
3869 | ||
93be8788 CW |
3870 | obj->user_pin_count++; |
3871 | obj->pin_filp = file; | |
3872 | ||
f343c5f6 | 3873 | args->offset = i915_gem_obj_ggtt_offset(obj); |
1d7cfea1 | 3874 | out: |
05394f39 | 3875 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3876 | unlock: |
673a394b | 3877 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3878 | return ret; |
673a394b EA |
3879 | } |
3880 | ||
3881 | int | |
3882 | i915_gem_unpin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3883 | struct drm_file *file) |
673a394b EA |
3884 | { |
3885 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 3886 | struct drm_i915_gem_object *obj; |
76c1dec1 | 3887 | int ret; |
673a394b | 3888 | |
1d7cfea1 CW |
3889 | ret = i915_mutex_lock_interruptible(dev); |
3890 | if (ret) | |
3891 | return ret; | |
673a394b | 3892 | |
05394f39 | 3893 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3894 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3895 | ret = -ENOENT; |
3896 | goto unlock; | |
673a394b | 3897 | } |
76c1dec1 | 3898 | |
05394f39 | 3899 | if (obj->pin_filp != file) { |
79e53945 JB |
3900 | DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n", |
3901 | args->handle); | |
1d7cfea1 CW |
3902 | ret = -EINVAL; |
3903 | goto out; | |
79e53945 | 3904 | } |
05394f39 CW |
3905 | obj->user_pin_count--; |
3906 | if (obj->user_pin_count == 0) { | |
3907 | obj->pin_filp = NULL; | |
79e53945 JB |
3908 | i915_gem_object_unpin(obj); |
3909 | } | |
673a394b | 3910 | |
1d7cfea1 | 3911 | out: |
05394f39 | 3912 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3913 | unlock: |
673a394b | 3914 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3915 | return ret; |
673a394b EA |
3916 | } |
3917 | ||
3918 | int | |
3919 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3920 | struct drm_file *file) |
673a394b EA |
3921 | { |
3922 | struct drm_i915_gem_busy *args = data; | |
05394f39 | 3923 | struct drm_i915_gem_object *obj; |
30dbf0c0 CW |
3924 | int ret; |
3925 | ||
76c1dec1 | 3926 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 3927 | if (ret) |
76c1dec1 | 3928 | return ret; |
673a394b | 3929 | |
05394f39 | 3930 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3931 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3932 | ret = -ENOENT; |
3933 | goto unlock; | |
673a394b | 3934 | } |
d1b851fc | 3935 | |
0be555b6 CW |
3936 | /* Count all active objects as busy, even if they are currently not used |
3937 | * by the gpu. Users of this interface expect objects to eventually | |
3938 | * become non-busy without any further actions, therefore emit any | |
3939 | * necessary flushes here. | |
c4de0a5d | 3940 | */ |
30dfebf3 | 3941 | ret = i915_gem_object_flush_active(obj); |
0be555b6 | 3942 | |
30dfebf3 | 3943 | args->busy = obj->active; |
e9808edd CW |
3944 | if (obj->ring) { |
3945 | BUILD_BUG_ON(I915_NUM_RINGS > 16); | |
3946 | args->busy |= intel_ring_flag(obj->ring) << 16; | |
3947 | } | |
673a394b | 3948 | |
05394f39 | 3949 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3950 | unlock: |
673a394b | 3951 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3952 | return ret; |
673a394b EA |
3953 | } |
3954 | ||
3955 | int | |
3956 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, | |
3957 | struct drm_file *file_priv) | |
3958 | { | |
0206e353 | 3959 | return i915_gem_ring_throttle(dev, file_priv); |
673a394b EA |
3960 | } |
3961 | ||
3ef94daa CW |
3962 | int |
3963 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, | |
3964 | struct drm_file *file_priv) | |
3965 | { | |
3966 | struct drm_i915_gem_madvise *args = data; | |
05394f39 | 3967 | struct drm_i915_gem_object *obj; |
76c1dec1 | 3968 | int ret; |
3ef94daa CW |
3969 | |
3970 | switch (args->madv) { | |
3971 | case I915_MADV_DONTNEED: | |
3972 | case I915_MADV_WILLNEED: | |
3973 | break; | |
3974 | default: | |
3975 | return -EINVAL; | |
3976 | } | |
3977 | ||
1d7cfea1 CW |
3978 | ret = i915_mutex_lock_interruptible(dev); |
3979 | if (ret) | |
3980 | return ret; | |
3981 | ||
05394f39 | 3982 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); |
c8725226 | 3983 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3984 | ret = -ENOENT; |
3985 | goto unlock; | |
3ef94daa | 3986 | } |
3ef94daa | 3987 | |
05394f39 | 3988 | if (obj->pin_count) { |
1d7cfea1 CW |
3989 | ret = -EINVAL; |
3990 | goto out; | |
3ef94daa CW |
3991 | } |
3992 | ||
05394f39 CW |
3993 | if (obj->madv != __I915_MADV_PURGED) |
3994 | obj->madv = args->madv; | |
3ef94daa | 3995 | |
6c085a72 CW |
3996 | /* if the object is no longer attached, discard its backing storage */ |
3997 | if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL) | |
2d7ef395 CW |
3998 | i915_gem_object_truncate(obj); |
3999 | ||
05394f39 | 4000 | args->retained = obj->madv != __I915_MADV_PURGED; |
bb6baf76 | 4001 | |
1d7cfea1 | 4002 | out: |
05394f39 | 4003 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 4004 | unlock: |
3ef94daa | 4005 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 4006 | return ret; |
3ef94daa CW |
4007 | } |
4008 | ||
37e680a1 CW |
4009 | void i915_gem_object_init(struct drm_i915_gem_object *obj, |
4010 | const struct drm_i915_gem_object_ops *ops) | |
0327d6ba | 4011 | { |
35c20a60 | 4012 | INIT_LIST_HEAD(&obj->global_list); |
0327d6ba CW |
4013 | INIT_LIST_HEAD(&obj->ring_list); |
4014 | INIT_LIST_HEAD(&obj->exec_list); | |
b25cb2f8 | 4015 | INIT_LIST_HEAD(&obj->obj_exec_link); |
2f633156 | 4016 | INIT_LIST_HEAD(&obj->vma_list); |
0327d6ba | 4017 | |
37e680a1 CW |
4018 | obj->ops = ops; |
4019 | ||
0327d6ba CW |
4020 | obj->fence_reg = I915_FENCE_REG_NONE; |
4021 | obj->madv = I915_MADV_WILLNEED; | |
4022 | /* Avoid an unnecessary call to unbind on the first bind. */ | |
4023 | obj->map_and_fenceable = true; | |
4024 | ||
4025 | i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size); | |
4026 | } | |
4027 | ||
37e680a1 CW |
4028 | static const struct drm_i915_gem_object_ops i915_gem_object_ops = { |
4029 | .get_pages = i915_gem_object_get_pages_gtt, | |
4030 | .put_pages = i915_gem_object_put_pages_gtt, | |
4031 | }; | |
4032 | ||
05394f39 CW |
4033 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, |
4034 | size_t size) | |
ac52bc56 | 4035 | { |
c397b908 | 4036 | struct drm_i915_gem_object *obj; |
5949eac4 | 4037 | struct address_space *mapping; |
1a240d4d | 4038 | gfp_t mask; |
ac52bc56 | 4039 | |
42dcedd4 | 4040 | obj = i915_gem_object_alloc(dev); |
c397b908 DV |
4041 | if (obj == NULL) |
4042 | return NULL; | |
673a394b | 4043 | |
c397b908 | 4044 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { |
42dcedd4 | 4045 | i915_gem_object_free(obj); |
c397b908 DV |
4046 | return NULL; |
4047 | } | |
673a394b | 4048 | |
bed1ea95 CW |
4049 | mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; |
4050 | if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) { | |
4051 | /* 965gm cannot relocate objects above 4GiB. */ | |
4052 | mask &= ~__GFP_HIGHMEM; | |
4053 | mask |= __GFP_DMA32; | |
4054 | } | |
4055 | ||
496ad9aa | 4056 | mapping = file_inode(obj->base.filp)->i_mapping; |
bed1ea95 | 4057 | mapping_set_gfp_mask(mapping, mask); |
5949eac4 | 4058 | |
37e680a1 | 4059 | i915_gem_object_init(obj, &i915_gem_object_ops); |
73aa808f | 4060 | |
c397b908 DV |
4061 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
4062 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
673a394b | 4063 | |
3d29b842 ED |
4064 | if (HAS_LLC(dev)) { |
4065 | /* On some devices, we can have the GPU use the LLC (the CPU | |
a1871112 EA |
4066 | * cache) for about a 10% performance improvement |
4067 | * compared to uncached. Graphics requests other than | |
4068 | * display scanout are coherent with the CPU in | |
4069 | * accessing this cache. This means in this mode we | |
4070 | * don't need to clflush on the CPU side, and on the | |
4071 | * GPU side we only need to flush internal caches to | |
4072 | * get data visible to the CPU. | |
4073 | * | |
4074 | * However, we maintain the display planes as UC, and so | |
4075 | * need to rebind when first used as such. | |
4076 | */ | |
4077 | obj->cache_level = I915_CACHE_LLC; | |
4078 | } else | |
4079 | obj->cache_level = I915_CACHE_NONE; | |
4080 | ||
d861e338 DV |
4081 | trace_i915_gem_object_create(obj); |
4082 | ||
05394f39 | 4083 | return obj; |
c397b908 DV |
4084 | } |
4085 | ||
4086 | int i915_gem_init_object(struct drm_gem_object *obj) | |
4087 | { | |
4088 | BUG(); | |
de151cf6 | 4089 | |
673a394b EA |
4090 | return 0; |
4091 | } | |
4092 | ||
1488fc08 | 4093 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
673a394b | 4094 | { |
1488fc08 | 4095 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); |
05394f39 | 4096 | struct drm_device *dev = obj->base.dev; |
be72615b | 4097 | drm_i915_private_t *dev_priv = dev->dev_private; |
07fe0b12 | 4098 | struct i915_vma *vma, *next; |
673a394b | 4099 | |
26e12f89 CW |
4100 | trace_i915_gem_object_destroy(obj); |
4101 | ||
1488fc08 CW |
4102 | if (obj->phys_obj) |
4103 | i915_gem_detach_phys_object(dev, obj); | |
4104 | ||
4105 | obj->pin_count = 0; | |
07fe0b12 BW |
4106 | /* NB: 0 or 1 elements */ |
4107 | WARN_ON(!list_empty(&obj->vma_list) && | |
4108 | !list_is_singular(&obj->vma_list)); | |
4109 | list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) { | |
4110 | int ret = i915_vma_unbind(vma); | |
4111 | if (WARN_ON(ret == -ERESTARTSYS)) { | |
4112 | bool was_interruptible; | |
1488fc08 | 4113 | |
07fe0b12 BW |
4114 | was_interruptible = dev_priv->mm.interruptible; |
4115 | dev_priv->mm.interruptible = false; | |
1488fc08 | 4116 | |
07fe0b12 | 4117 | WARN_ON(i915_vma_unbind(vma)); |
1488fc08 | 4118 | |
07fe0b12 BW |
4119 | dev_priv->mm.interruptible = was_interruptible; |
4120 | } | |
1488fc08 CW |
4121 | } |
4122 | ||
1d64ae71 BW |
4123 | /* Stolen objects don't hold a ref, but do hold pin count. Fix that up |
4124 | * before progressing. */ | |
4125 | if (obj->stolen) | |
4126 | i915_gem_object_unpin_pages(obj); | |
4127 | ||
401c29f6 BW |
4128 | if (WARN_ON(obj->pages_pin_count)) |
4129 | obj->pages_pin_count = 0; | |
37e680a1 | 4130 | i915_gem_object_put_pages(obj); |
d8cb5086 | 4131 | i915_gem_object_free_mmap_offset(obj); |
0104fdbb | 4132 | i915_gem_object_release_stolen(obj); |
de151cf6 | 4133 | |
9da3da66 CW |
4134 | BUG_ON(obj->pages); |
4135 | ||
2f745ad3 CW |
4136 | if (obj->base.import_attach) |
4137 | drm_prime_gem_destroy(&obj->base, NULL); | |
de151cf6 | 4138 | |
05394f39 CW |
4139 | drm_gem_object_release(&obj->base); |
4140 | i915_gem_info_remove_obj(dev_priv, obj->base.size); | |
c397b908 | 4141 | |
05394f39 | 4142 | kfree(obj->bit_17); |
42dcedd4 | 4143 | i915_gem_object_free(obj); |
673a394b EA |
4144 | } |
4145 | ||
2f633156 BW |
4146 | struct i915_vma *i915_gem_vma_create(struct drm_i915_gem_object *obj, |
4147 | struct i915_address_space *vm) | |
4148 | { | |
4149 | struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL); | |
4150 | if (vma == NULL) | |
4151 | return ERR_PTR(-ENOMEM); | |
4152 | ||
4153 | INIT_LIST_HEAD(&vma->vma_link); | |
ca191b13 | 4154 | INIT_LIST_HEAD(&vma->mm_list); |
82a55ad1 | 4155 | INIT_LIST_HEAD(&vma->exec_list); |
2f633156 BW |
4156 | vma->vm = vm; |
4157 | vma->obj = obj; | |
4158 | ||
8b9c2b94 BW |
4159 | /* Keep GGTT vmas first to make debug easier */ |
4160 | if (i915_is_ggtt(vm)) | |
4161 | list_add(&vma->vma_link, &obj->vma_list); | |
4162 | else | |
4163 | list_add_tail(&vma->vma_link, &obj->vma_list); | |
4164 | ||
2f633156 BW |
4165 | return vma; |
4166 | } | |
4167 | ||
4168 | void i915_gem_vma_destroy(struct i915_vma *vma) | |
4169 | { | |
4170 | WARN_ON(vma->node.allocated); | |
8b9c2b94 | 4171 | list_del(&vma->vma_link); |
2f633156 BW |
4172 | kfree(vma); |
4173 | } | |
4174 | ||
29105ccc CW |
4175 | int |
4176 | i915_gem_idle(struct drm_device *dev) | |
4177 | { | |
4178 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4179 | int ret; | |
28dfe52a | 4180 | |
db1b76ca | 4181 | if (dev_priv->ums.mm_suspended) { |
29105ccc CW |
4182 | mutex_unlock(&dev->struct_mutex); |
4183 | return 0; | |
28dfe52a EA |
4184 | } |
4185 | ||
b2da9fe5 | 4186 | ret = i915_gpu_idle(dev); |
6dbe2772 KP |
4187 | if (ret) { |
4188 | mutex_unlock(&dev->struct_mutex); | |
673a394b | 4189 | return ret; |
6dbe2772 | 4190 | } |
b2da9fe5 | 4191 | i915_gem_retire_requests(dev); |
673a394b | 4192 | |
29105ccc | 4193 | /* Under UMS, be paranoid and evict. */ |
a39d7efc | 4194 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
6c085a72 | 4195 | i915_gem_evict_everything(dev); |
29105ccc | 4196 | |
99584db3 | 4197 | del_timer_sync(&dev_priv->gpu_error.hangcheck_timer); |
29105ccc CW |
4198 | |
4199 | i915_kernel_lost_context(dev); | |
6dbe2772 | 4200 | i915_gem_cleanup_ringbuffer(dev); |
29105ccc | 4201 | |
29105ccc CW |
4202 | /* Cancel the retire work handler, which should be idle now. */ |
4203 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); | |
4204 | ||
673a394b EA |
4205 | return 0; |
4206 | } | |
4207 | ||
b9524a1e BW |
4208 | void i915_gem_l3_remap(struct drm_device *dev) |
4209 | { | |
4210 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4211 | u32 misccpctl; | |
4212 | int i; | |
4213 | ||
eb32e458 | 4214 | if (!HAS_L3_GPU_CACHE(dev)) |
b9524a1e BW |
4215 | return; |
4216 | ||
a4da4fa4 | 4217 | if (!dev_priv->l3_parity.remap_info) |
b9524a1e BW |
4218 | return; |
4219 | ||
4220 | misccpctl = I915_READ(GEN7_MISCCPCTL); | |
4221 | I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE); | |
4222 | POSTING_READ(GEN7_MISCCPCTL); | |
4223 | ||
4224 | for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) { | |
4225 | u32 remap = I915_READ(GEN7_L3LOG_BASE + i); | |
a4da4fa4 | 4226 | if (remap && remap != dev_priv->l3_parity.remap_info[i/4]) |
b9524a1e BW |
4227 | DRM_DEBUG("0x%x was already programmed to %x\n", |
4228 | GEN7_L3LOG_BASE + i, remap); | |
a4da4fa4 | 4229 | if (remap && !dev_priv->l3_parity.remap_info[i/4]) |
b9524a1e | 4230 | DRM_DEBUG_DRIVER("Clearing remapped register\n"); |
a4da4fa4 | 4231 | I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]); |
b9524a1e BW |
4232 | } |
4233 | ||
4234 | /* Make sure all the writes land before disabling dop clock gating */ | |
4235 | POSTING_READ(GEN7_L3LOG_BASE); | |
4236 | ||
4237 | I915_WRITE(GEN7_MISCCPCTL, misccpctl); | |
4238 | } | |
4239 | ||
f691e2f4 DV |
4240 | void i915_gem_init_swizzling(struct drm_device *dev) |
4241 | { | |
4242 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4243 | ||
11782b02 | 4244 | if (INTEL_INFO(dev)->gen < 5 || |
f691e2f4 DV |
4245 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
4246 | return; | |
4247 | ||
4248 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | | |
4249 | DISP_TILE_SURFACE_SWIZZLING); | |
4250 | ||
11782b02 DV |
4251 | if (IS_GEN5(dev)) |
4252 | return; | |
4253 | ||
f691e2f4 DV |
4254 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
4255 | if (IS_GEN6(dev)) | |
6b26c86d | 4256 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); |
8782e26c | 4257 | else if (IS_GEN7(dev)) |
6b26c86d | 4258 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); |
8782e26c BW |
4259 | else |
4260 | BUG(); | |
f691e2f4 | 4261 | } |
e21af88d | 4262 | |
67b1b571 CW |
4263 | static bool |
4264 | intel_enable_blt(struct drm_device *dev) | |
4265 | { | |
4266 | if (!HAS_BLT(dev)) | |
4267 | return false; | |
4268 | ||
4269 | /* The blitter was dysfunctional on early prototypes */ | |
4270 | if (IS_GEN6(dev) && dev->pdev->revision < 8) { | |
4271 | DRM_INFO("BLT not supported on this pre-production hardware;" | |
4272 | " graphics performance will be degraded.\n"); | |
4273 | return false; | |
4274 | } | |
4275 | ||
4276 | return true; | |
4277 | } | |
4278 | ||
4fc7c971 | 4279 | static int i915_gem_init_rings(struct drm_device *dev) |
8187a2b7 | 4280 | { |
4fc7c971 | 4281 | struct drm_i915_private *dev_priv = dev->dev_private; |
8187a2b7 | 4282 | int ret; |
68f95ba9 | 4283 | |
5c1143bb | 4284 | ret = intel_init_render_ring_buffer(dev); |
68f95ba9 | 4285 | if (ret) |
b6913e4b | 4286 | return ret; |
68f95ba9 CW |
4287 | |
4288 | if (HAS_BSD(dev)) { | |
5c1143bb | 4289 | ret = intel_init_bsd_ring_buffer(dev); |
68f95ba9 CW |
4290 | if (ret) |
4291 | goto cleanup_render_ring; | |
d1b851fc | 4292 | } |
68f95ba9 | 4293 | |
67b1b571 | 4294 | if (intel_enable_blt(dev)) { |
549f7365 CW |
4295 | ret = intel_init_blt_ring_buffer(dev); |
4296 | if (ret) | |
4297 | goto cleanup_bsd_ring; | |
4298 | } | |
4299 | ||
9a8a2213 BW |
4300 | if (HAS_VEBOX(dev)) { |
4301 | ret = intel_init_vebox_ring_buffer(dev); | |
4302 | if (ret) | |
4303 | goto cleanup_blt_ring; | |
4304 | } | |
4305 | ||
4306 | ||
99433931 | 4307 | ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000)); |
4fc7c971 | 4308 | if (ret) |
9a8a2213 | 4309 | goto cleanup_vebox_ring; |
4fc7c971 BW |
4310 | |
4311 | return 0; | |
4312 | ||
9a8a2213 BW |
4313 | cleanup_vebox_ring: |
4314 | intel_cleanup_ring_buffer(&dev_priv->ring[VECS]); | |
4fc7c971 BW |
4315 | cleanup_blt_ring: |
4316 | intel_cleanup_ring_buffer(&dev_priv->ring[BCS]); | |
4317 | cleanup_bsd_ring: | |
4318 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); | |
4319 | cleanup_render_ring: | |
4320 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); | |
4321 | ||
4322 | return ret; | |
4323 | } | |
4324 | ||
4325 | int | |
4326 | i915_gem_init_hw(struct drm_device *dev) | |
4327 | { | |
4328 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4329 | int ret; | |
4330 | ||
4331 | if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt()) | |
4332 | return -EIO; | |
4333 | ||
59124506 | 4334 | if (dev_priv->ellc_size) |
05e21cc4 | 4335 | I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf)); |
4fc7c971 | 4336 | |
88a2b2a3 BW |
4337 | if (HAS_PCH_NOP(dev)) { |
4338 | u32 temp = I915_READ(GEN7_MSG_CTL); | |
4339 | temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK); | |
4340 | I915_WRITE(GEN7_MSG_CTL, temp); | |
4341 | } | |
4342 | ||
4fc7c971 BW |
4343 | i915_gem_l3_remap(dev); |
4344 | ||
4345 | i915_gem_init_swizzling(dev); | |
4346 | ||
4347 | ret = i915_gem_init_rings(dev); | |
99433931 MK |
4348 | if (ret) |
4349 | return ret; | |
4350 | ||
254f965c BW |
4351 | /* |
4352 | * XXX: There was some w/a described somewhere suggesting loading | |
4353 | * contexts before PPGTT. | |
4354 | */ | |
4355 | i915_gem_context_init(dev); | |
b7c36d25 BW |
4356 | if (dev_priv->mm.aliasing_ppgtt) { |
4357 | ret = dev_priv->mm.aliasing_ppgtt->enable(dev); | |
4358 | if (ret) { | |
4359 | i915_gem_cleanup_aliasing_ppgtt(dev); | |
4360 | DRM_INFO("PPGTT enable failed. This is not fatal, but unexpected\n"); | |
4361 | } | |
4362 | } | |
e21af88d | 4363 | |
68f95ba9 | 4364 | return 0; |
8187a2b7 ZN |
4365 | } |
4366 | ||
1070a42b CW |
4367 | int i915_gem_init(struct drm_device *dev) |
4368 | { | |
4369 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1070a42b CW |
4370 | int ret; |
4371 | ||
1070a42b | 4372 | mutex_lock(&dev->struct_mutex); |
d62b4892 JB |
4373 | |
4374 | if (IS_VALLEYVIEW(dev)) { | |
4375 | /* VLVA0 (potential hack), BIOS isn't actually waking us */ | |
4376 | I915_WRITE(VLV_GTLC_WAKE_CTRL, 1); | |
4377 | if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) & 1) == 1, 10)) | |
4378 | DRM_DEBUG_DRIVER("allow wake ack timed out\n"); | |
4379 | } | |
4380 | ||
d7e5008f | 4381 | i915_gem_init_global_gtt(dev); |
d62b4892 | 4382 | |
1070a42b CW |
4383 | ret = i915_gem_init_hw(dev); |
4384 | mutex_unlock(&dev->struct_mutex); | |
4385 | if (ret) { | |
4386 | i915_gem_cleanup_aliasing_ppgtt(dev); | |
4387 | return ret; | |
4388 | } | |
4389 | ||
53ca26ca DV |
4390 | /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */ |
4391 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) | |
4392 | dev_priv->dri1.allow_batchbuffer = 1; | |
1070a42b CW |
4393 | return 0; |
4394 | } | |
4395 | ||
8187a2b7 ZN |
4396 | void |
4397 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) | |
4398 | { | |
4399 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 4400 | struct intel_ring_buffer *ring; |
1ec14ad3 | 4401 | int i; |
8187a2b7 | 4402 | |
b4519513 CW |
4403 | for_each_ring(ring, dev_priv, i) |
4404 | intel_cleanup_ring_buffer(ring); | |
8187a2b7 ZN |
4405 | } |
4406 | ||
673a394b EA |
4407 | int |
4408 | i915_gem_entervt_ioctl(struct drm_device *dev, void *data, | |
4409 | struct drm_file *file_priv) | |
4410 | { | |
db1b76ca | 4411 | struct drm_i915_private *dev_priv = dev->dev_private; |
b4519513 | 4412 | int ret; |
673a394b | 4413 | |
79e53945 JB |
4414 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4415 | return 0; | |
4416 | ||
1f83fee0 | 4417 | if (i915_reset_in_progress(&dev_priv->gpu_error)) { |
673a394b | 4418 | DRM_ERROR("Reenabling wedged hardware, good luck\n"); |
1f83fee0 | 4419 | atomic_set(&dev_priv->gpu_error.reset_counter, 0); |
673a394b EA |
4420 | } |
4421 | ||
673a394b | 4422 | mutex_lock(&dev->struct_mutex); |
db1b76ca | 4423 | dev_priv->ums.mm_suspended = 0; |
9bb2d6f9 | 4424 | |
f691e2f4 | 4425 | ret = i915_gem_init_hw(dev); |
d816f6ac WF |
4426 | if (ret != 0) { |
4427 | mutex_unlock(&dev->struct_mutex); | |
9bb2d6f9 | 4428 | return ret; |
d816f6ac | 4429 | } |
9bb2d6f9 | 4430 | |
5cef07e1 | 4431 | BUG_ON(!list_empty(&dev_priv->gtt.base.active_list)); |
673a394b | 4432 | mutex_unlock(&dev->struct_mutex); |
dbb19d30 | 4433 | |
5f35308b CW |
4434 | ret = drm_irq_install(dev); |
4435 | if (ret) | |
4436 | goto cleanup_ringbuffer; | |
dbb19d30 | 4437 | |
673a394b | 4438 | return 0; |
5f35308b CW |
4439 | |
4440 | cleanup_ringbuffer: | |
4441 | mutex_lock(&dev->struct_mutex); | |
4442 | i915_gem_cleanup_ringbuffer(dev); | |
db1b76ca | 4443 | dev_priv->ums.mm_suspended = 1; |
5f35308b CW |
4444 | mutex_unlock(&dev->struct_mutex); |
4445 | ||
4446 | return ret; | |
673a394b EA |
4447 | } |
4448 | ||
4449 | int | |
4450 | i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, | |
4451 | struct drm_file *file_priv) | |
4452 | { | |
db1b76ca DV |
4453 | struct drm_i915_private *dev_priv = dev->dev_private; |
4454 | int ret; | |
4455 | ||
79e53945 JB |
4456 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4457 | return 0; | |
4458 | ||
dbb19d30 | 4459 | drm_irq_uninstall(dev); |
db1b76ca DV |
4460 | |
4461 | mutex_lock(&dev->struct_mutex); | |
4462 | ret = i915_gem_idle(dev); | |
4463 | ||
4464 | /* Hack! Don't let anybody do execbuf while we don't control the chip. | |
4465 | * We need to replace this with a semaphore, or something. | |
4466 | * And not confound ums.mm_suspended! | |
4467 | */ | |
4468 | if (ret != 0) | |
4469 | dev_priv->ums.mm_suspended = 1; | |
4470 | mutex_unlock(&dev->struct_mutex); | |
4471 | ||
4472 | return ret; | |
673a394b EA |
4473 | } |
4474 | ||
4475 | void | |
4476 | i915_gem_lastclose(struct drm_device *dev) | |
4477 | { | |
4478 | int ret; | |
673a394b | 4479 | |
e806b495 EA |
4480 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
4481 | return; | |
4482 | ||
db1b76ca | 4483 | mutex_lock(&dev->struct_mutex); |
6dbe2772 KP |
4484 | ret = i915_gem_idle(dev); |
4485 | if (ret) | |
4486 | DRM_ERROR("failed to idle hardware: %d\n", ret); | |
db1b76ca | 4487 | mutex_unlock(&dev->struct_mutex); |
673a394b EA |
4488 | } |
4489 | ||
64193406 CW |
4490 | static void |
4491 | init_ring_lists(struct intel_ring_buffer *ring) | |
4492 | { | |
4493 | INIT_LIST_HEAD(&ring->active_list); | |
4494 | INIT_LIST_HEAD(&ring->request_list); | |
64193406 CW |
4495 | } |
4496 | ||
fc8c067e BW |
4497 | static void i915_init_vm(struct drm_i915_private *dev_priv, |
4498 | struct i915_address_space *vm) | |
4499 | { | |
4500 | vm->dev = dev_priv->dev; | |
4501 | INIT_LIST_HEAD(&vm->active_list); | |
4502 | INIT_LIST_HEAD(&vm->inactive_list); | |
4503 | INIT_LIST_HEAD(&vm->global_link); | |
4504 | list_add(&vm->global_link, &dev_priv->vm_list); | |
4505 | } | |
4506 | ||
673a394b EA |
4507 | void |
4508 | i915_gem_load(struct drm_device *dev) | |
4509 | { | |
4510 | drm_i915_private_t *dev_priv = dev->dev_private; | |
42dcedd4 CW |
4511 | int i; |
4512 | ||
4513 | dev_priv->slab = | |
4514 | kmem_cache_create("i915_gem_object", | |
4515 | sizeof(struct drm_i915_gem_object), 0, | |
4516 | SLAB_HWCACHE_ALIGN, | |
4517 | NULL); | |
673a394b | 4518 | |
fc8c067e BW |
4519 | INIT_LIST_HEAD(&dev_priv->vm_list); |
4520 | i915_init_vm(dev_priv, &dev_priv->gtt.base); | |
4521 | ||
6c085a72 CW |
4522 | INIT_LIST_HEAD(&dev_priv->mm.unbound_list); |
4523 | INIT_LIST_HEAD(&dev_priv->mm.bound_list); | |
a09ba7fa | 4524 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
1ec14ad3 CW |
4525 | for (i = 0; i < I915_NUM_RINGS; i++) |
4526 | init_ring_lists(&dev_priv->ring[i]); | |
4b9de737 | 4527 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) |
007cc8ac | 4528 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); |
673a394b EA |
4529 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, |
4530 | i915_gem_retire_work_handler); | |
1f83fee0 | 4531 | init_waitqueue_head(&dev_priv->gpu_error.reset_queue); |
31169714 | 4532 | |
94400120 DA |
4533 | /* On GEN3 we really need to make sure the ARB C3 LP bit is set */ |
4534 | if (IS_GEN3(dev)) { | |
50743298 DV |
4535 | I915_WRITE(MI_ARB_STATE, |
4536 | _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE)); | |
94400120 DA |
4537 | } |
4538 | ||
72bfa19c CW |
4539 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; |
4540 | ||
de151cf6 | 4541 | /* Old X drivers will take 0-2 for front, back, depth buffers */ |
b397c836 EA |
4542 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
4543 | dev_priv->fence_reg_start = 3; | |
de151cf6 | 4544 | |
42b5aeab VS |
4545 | if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev)) |
4546 | dev_priv->num_fence_regs = 32; | |
4547 | else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) | |
de151cf6 JB |
4548 | dev_priv->num_fence_regs = 16; |
4549 | else | |
4550 | dev_priv->num_fence_regs = 8; | |
4551 | ||
b5aa8a0f | 4552 | /* Initialize fence registers to zero */ |
19b2dbde CW |
4553 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
4554 | i915_gem_restore_fences(dev); | |
10ed13e4 | 4555 | |
673a394b | 4556 | i915_gem_detect_bit_6_swizzle(dev); |
6b95a207 | 4557 | init_waitqueue_head(&dev_priv->pending_flip_queue); |
17250b71 | 4558 | |
ce453d81 CW |
4559 | dev_priv->mm.interruptible = true; |
4560 | ||
17250b71 CW |
4561 | dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink; |
4562 | dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS; | |
4563 | register_shrinker(&dev_priv->mm.inactive_shrinker); | |
673a394b | 4564 | } |
71acb5eb DA |
4565 | |
4566 | /* | |
4567 | * Create a physically contiguous memory object for this object | |
4568 | * e.g. for cursor + overlay regs | |
4569 | */ | |
995b6762 CW |
4570 | static int i915_gem_init_phys_object(struct drm_device *dev, |
4571 | int id, int size, int align) | |
71acb5eb DA |
4572 | { |
4573 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4574 | struct drm_i915_gem_phys_object *phys_obj; | |
4575 | int ret; | |
4576 | ||
4577 | if (dev_priv->mm.phys_objs[id - 1] || !size) | |
4578 | return 0; | |
4579 | ||
9a298b2a | 4580 | phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL); |
71acb5eb DA |
4581 | if (!phys_obj) |
4582 | return -ENOMEM; | |
4583 | ||
4584 | phys_obj->id = id; | |
4585 | ||
6eeefaf3 | 4586 | phys_obj->handle = drm_pci_alloc(dev, size, align); |
71acb5eb DA |
4587 | if (!phys_obj->handle) { |
4588 | ret = -ENOMEM; | |
4589 | goto kfree_obj; | |
4590 | } | |
4591 | #ifdef CONFIG_X86 | |
4592 | set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
4593 | #endif | |
4594 | ||
4595 | dev_priv->mm.phys_objs[id - 1] = phys_obj; | |
4596 | ||
4597 | return 0; | |
4598 | kfree_obj: | |
9a298b2a | 4599 | kfree(phys_obj); |
71acb5eb DA |
4600 | return ret; |
4601 | } | |
4602 | ||
995b6762 | 4603 | static void i915_gem_free_phys_object(struct drm_device *dev, int id) |
71acb5eb DA |
4604 | { |
4605 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4606 | struct drm_i915_gem_phys_object *phys_obj; | |
4607 | ||
4608 | if (!dev_priv->mm.phys_objs[id - 1]) | |
4609 | return; | |
4610 | ||
4611 | phys_obj = dev_priv->mm.phys_objs[id - 1]; | |
4612 | if (phys_obj->cur_obj) { | |
4613 | i915_gem_detach_phys_object(dev, phys_obj->cur_obj); | |
4614 | } | |
4615 | ||
4616 | #ifdef CONFIG_X86 | |
4617 | set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
4618 | #endif | |
4619 | drm_pci_free(dev, phys_obj->handle); | |
4620 | kfree(phys_obj); | |
4621 | dev_priv->mm.phys_objs[id - 1] = NULL; | |
4622 | } | |
4623 | ||
4624 | void i915_gem_free_all_phys_object(struct drm_device *dev) | |
4625 | { | |
4626 | int i; | |
4627 | ||
260883c8 | 4628 | for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++) |
71acb5eb DA |
4629 | i915_gem_free_phys_object(dev, i); |
4630 | } | |
4631 | ||
4632 | void i915_gem_detach_phys_object(struct drm_device *dev, | |
05394f39 | 4633 | struct drm_i915_gem_object *obj) |
71acb5eb | 4634 | { |
496ad9aa | 4635 | struct address_space *mapping = file_inode(obj->base.filp)->i_mapping; |
e5281ccd | 4636 | char *vaddr; |
71acb5eb | 4637 | int i; |
71acb5eb DA |
4638 | int page_count; |
4639 | ||
05394f39 | 4640 | if (!obj->phys_obj) |
71acb5eb | 4641 | return; |
05394f39 | 4642 | vaddr = obj->phys_obj->handle->vaddr; |
71acb5eb | 4643 | |
05394f39 | 4644 | page_count = obj->base.size / PAGE_SIZE; |
71acb5eb | 4645 | for (i = 0; i < page_count; i++) { |
5949eac4 | 4646 | struct page *page = shmem_read_mapping_page(mapping, i); |
e5281ccd CW |
4647 | if (!IS_ERR(page)) { |
4648 | char *dst = kmap_atomic(page); | |
4649 | memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE); | |
4650 | kunmap_atomic(dst); | |
4651 | ||
4652 | drm_clflush_pages(&page, 1); | |
4653 | ||
4654 | set_page_dirty(page); | |
4655 | mark_page_accessed(page); | |
4656 | page_cache_release(page); | |
4657 | } | |
71acb5eb | 4658 | } |
e76e9aeb | 4659 | i915_gem_chipset_flush(dev); |
d78b47b9 | 4660 | |
05394f39 CW |
4661 | obj->phys_obj->cur_obj = NULL; |
4662 | obj->phys_obj = NULL; | |
71acb5eb DA |
4663 | } |
4664 | ||
4665 | int | |
4666 | i915_gem_attach_phys_object(struct drm_device *dev, | |
05394f39 | 4667 | struct drm_i915_gem_object *obj, |
6eeefaf3 CW |
4668 | int id, |
4669 | int align) | |
71acb5eb | 4670 | { |
496ad9aa | 4671 | struct address_space *mapping = file_inode(obj->base.filp)->i_mapping; |
71acb5eb | 4672 | drm_i915_private_t *dev_priv = dev->dev_private; |
71acb5eb DA |
4673 | int ret = 0; |
4674 | int page_count; | |
4675 | int i; | |
4676 | ||
4677 | if (id > I915_MAX_PHYS_OBJECT) | |
4678 | return -EINVAL; | |
4679 | ||
05394f39 CW |
4680 | if (obj->phys_obj) { |
4681 | if (obj->phys_obj->id == id) | |
71acb5eb DA |
4682 | return 0; |
4683 | i915_gem_detach_phys_object(dev, obj); | |
4684 | } | |
4685 | ||
71acb5eb DA |
4686 | /* create a new object */ |
4687 | if (!dev_priv->mm.phys_objs[id - 1]) { | |
4688 | ret = i915_gem_init_phys_object(dev, id, | |
05394f39 | 4689 | obj->base.size, align); |
71acb5eb | 4690 | if (ret) { |
05394f39 CW |
4691 | DRM_ERROR("failed to init phys object %d size: %zu\n", |
4692 | id, obj->base.size); | |
e5281ccd | 4693 | return ret; |
71acb5eb DA |
4694 | } |
4695 | } | |
4696 | ||
4697 | /* bind to the object */ | |
05394f39 CW |
4698 | obj->phys_obj = dev_priv->mm.phys_objs[id - 1]; |
4699 | obj->phys_obj->cur_obj = obj; | |
71acb5eb | 4700 | |
05394f39 | 4701 | page_count = obj->base.size / PAGE_SIZE; |
71acb5eb DA |
4702 | |
4703 | for (i = 0; i < page_count; i++) { | |
e5281ccd CW |
4704 | struct page *page; |
4705 | char *dst, *src; | |
4706 | ||
5949eac4 | 4707 | page = shmem_read_mapping_page(mapping, i); |
e5281ccd CW |
4708 | if (IS_ERR(page)) |
4709 | return PTR_ERR(page); | |
71acb5eb | 4710 | |
ff75b9bc | 4711 | src = kmap_atomic(page); |
05394f39 | 4712 | dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE); |
71acb5eb | 4713 | memcpy(dst, src, PAGE_SIZE); |
3e4d3af5 | 4714 | kunmap_atomic(src); |
71acb5eb | 4715 | |
e5281ccd CW |
4716 | mark_page_accessed(page); |
4717 | page_cache_release(page); | |
4718 | } | |
d78b47b9 | 4719 | |
71acb5eb | 4720 | return 0; |
71acb5eb DA |
4721 | } |
4722 | ||
4723 | static int | |
05394f39 CW |
4724 | i915_gem_phys_pwrite(struct drm_device *dev, |
4725 | struct drm_i915_gem_object *obj, | |
71acb5eb DA |
4726 | struct drm_i915_gem_pwrite *args, |
4727 | struct drm_file *file_priv) | |
4728 | { | |
05394f39 | 4729 | void *vaddr = obj->phys_obj->handle->vaddr + args->offset; |
2bb4629a | 4730 | char __user *user_data = to_user_ptr(args->data_ptr); |
71acb5eb | 4731 | |
b47b30cc CW |
4732 | if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) { |
4733 | unsigned long unwritten; | |
4734 | ||
4735 | /* The physical object once assigned is fixed for the lifetime | |
4736 | * of the obj, so we can safely drop the lock and continue | |
4737 | * to access vaddr. | |
4738 | */ | |
4739 | mutex_unlock(&dev->struct_mutex); | |
4740 | unwritten = copy_from_user(vaddr, user_data, args->size); | |
4741 | mutex_lock(&dev->struct_mutex); | |
4742 | if (unwritten) | |
4743 | return -EFAULT; | |
4744 | } | |
71acb5eb | 4745 | |
e76e9aeb | 4746 | i915_gem_chipset_flush(dev); |
71acb5eb DA |
4747 | return 0; |
4748 | } | |
b962442e | 4749 | |
f787a5f5 | 4750 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
b962442e | 4751 | { |
f787a5f5 | 4752 | struct drm_i915_file_private *file_priv = file->driver_priv; |
b962442e EA |
4753 | |
4754 | /* Clean up our request list when the client is going away, so that | |
4755 | * later retire_requests won't dereference our soon-to-be-gone | |
4756 | * file_priv. | |
4757 | */ | |
1c25595f | 4758 | spin_lock(&file_priv->mm.lock); |
f787a5f5 CW |
4759 | while (!list_empty(&file_priv->mm.request_list)) { |
4760 | struct drm_i915_gem_request *request; | |
4761 | ||
4762 | request = list_first_entry(&file_priv->mm.request_list, | |
4763 | struct drm_i915_gem_request, | |
4764 | client_list); | |
4765 | list_del(&request->client_list); | |
4766 | request->file_priv = NULL; | |
4767 | } | |
1c25595f | 4768 | spin_unlock(&file_priv->mm.lock); |
b962442e | 4769 | } |
31169714 | 4770 | |
5774506f CW |
4771 | static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task) |
4772 | { | |
4773 | if (!mutex_is_locked(mutex)) | |
4774 | return false; | |
4775 | ||
4776 | #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES) | |
4777 | return mutex->owner == task; | |
4778 | #else | |
4779 | /* Since UP may be pre-empted, we cannot assume that we own the lock */ | |
4780 | return false; | |
4781 | #endif | |
4782 | } | |
4783 | ||
31169714 | 4784 | static int |
1495f230 | 4785 | i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc) |
31169714 | 4786 | { |
17250b71 CW |
4787 | struct drm_i915_private *dev_priv = |
4788 | container_of(shrinker, | |
4789 | struct drm_i915_private, | |
4790 | mm.inactive_shrinker); | |
4791 | struct drm_device *dev = dev_priv->dev; | |
6c085a72 | 4792 | struct drm_i915_gem_object *obj; |
1495f230 | 4793 | int nr_to_scan = sc->nr_to_scan; |
5774506f | 4794 | bool unlock = true; |
17250b71 CW |
4795 | int cnt; |
4796 | ||
5774506f CW |
4797 | if (!mutex_trylock(&dev->struct_mutex)) { |
4798 | if (!mutex_is_locked_by(&dev->struct_mutex, current)) | |
4799 | return 0; | |
4800 | ||
677feac2 DV |
4801 | if (dev_priv->mm.shrinker_no_lock_stealing) |
4802 | return 0; | |
4803 | ||
5774506f CW |
4804 | unlock = false; |
4805 | } | |
31169714 | 4806 | |
6c085a72 CW |
4807 | if (nr_to_scan) { |
4808 | nr_to_scan -= i915_gem_purge(dev_priv, nr_to_scan); | |
93927ca5 DV |
4809 | if (nr_to_scan > 0) |
4810 | nr_to_scan -= __i915_gem_shrink(dev_priv, nr_to_scan, | |
4811 | false); | |
6c085a72 CW |
4812 | if (nr_to_scan > 0) |
4813 | i915_gem_shrink_all(dev_priv); | |
31169714 CW |
4814 | } |
4815 | ||
17250b71 | 4816 | cnt = 0; |
35c20a60 | 4817 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) |
a5570178 CW |
4818 | if (obj->pages_pin_count == 0) |
4819 | cnt += obj->base.size >> PAGE_SHIFT; | |
fcb4a578 BW |
4820 | |
4821 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { | |
4822 | if (obj->active) | |
4823 | continue; | |
4824 | ||
a5570178 | 4825 | if (obj->pin_count == 0 && obj->pages_pin_count == 0) |
6c085a72 | 4826 | cnt += obj->base.size >> PAGE_SHIFT; |
fcb4a578 | 4827 | } |
17250b71 | 4828 | |
5774506f CW |
4829 | if (unlock) |
4830 | mutex_unlock(&dev->struct_mutex); | |
6c085a72 | 4831 | return cnt; |
31169714 | 4832 | } |
a70a3148 BW |
4833 | |
4834 | /* All the new VM stuff */ | |
4835 | unsigned long i915_gem_obj_offset(struct drm_i915_gem_object *o, | |
4836 | struct i915_address_space *vm) | |
4837 | { | |
4838 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; | |
4839 | struct i915_vma *vma; | |
4840 | ||
4841 | if (vm == &dev_priv->mm.aliasing_ppgtt->base) | |
4842 | vm = &dev_priv->gtt.base; | |
4843 | ||
4844 | BUG_ON(list_empty(&o->vma_list)); | |
4845 | list_for_each_entry(vma, &o->vma_list, vma_link) { | |
4846 | if (vma->vm == vm) | |
4847 | return vma->node.start; | |
4848 | ||
4849 | } | |
4850 | return -1; | |
4851 | } | |
4852 | ||
4853 | bool i915_gem_obj_bound(struct drm_i915_gem_object *o, | |
4854 | struct i915_address_space *vm) | |
4855 | { | |
4856 | struct i915_vma *vma; | |
4857 | ||
4858 | list_for_each_entry(vma, &o->vma_list, vma_link) | |
8b9c2b94 | 4859 | if (vma->vm == vm && drm_mm_node_allocated(&vma->node)) |
a70a3148 BW |
4860 | return true; |
4861 | ||
4862 | return false; | |
4863 | } | |
4864 | ||
4865 | bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o) | |
4866 | { | |
4867 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; | |
4868 | struct i915_address_space *vm; | |
4869 | ||
4870 | list_for_each_entry(vm, &dev_priv->vm_list, global_link) | |
4871 | if (i915_gem_obj_bound(o, vm)) | |
4872 | return true; | |
4873 | ||
4874 | return false; | |
4875 | } | |
4876 | ||
4877 | unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o, | |
4878 | struct i915_address_space *vm) | |
4879 | { | |
4880 | struct drm_i915_private *dev_priv = o->base.dev->dev_private; | |
4881 | struct i915_vma *vma; | |
4882 | ||
4883 | if (vm == &dev_priv->mm.aliasing_ppgtt->base) | |
4884 | vm = &dev_priv->gtt.base; | |
4885 | ||
4886 | BUG_ON(list_empty(&o->vma_list)); | |
4887 | ||
4888 | list_for_each_entry(vma, &o->vma_list, vma_link) | |
4889 | if (vma->vm == vm) | |
4890 | return vma->node.size; | |
4891 | ||
4892 | return 0; | |
4893 | } | |
4894 | ||
4895 | struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj, | |
4896 | struct i915_address_space *vm) | |
4897 | { | |
4898 | struct i915_vma *vma; | |
4899 | list_for_each_entry(vma, &obj->vma_list, vma_link) | |
4900 | if (vma->vm == vm) | |
4901 | return vma; | |
4902 | ||
4903 | return NULL; | |
4904 | } | |
accfef2e BW |
4905 | |
4906 | struct i915_vma * | |
4907 | i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj, | |
4908 | struct i915_address_space *vm) | |
4909 | { | |
4910 | struct i915_vma *vma; | |
4911 | ||
4912 | vma = i915_gem_obj_to_vma(obj, vm); | |
4913 | if (!vma) | |
4914 | vma = i915_gem_vma_create(obj, vm); | |
4915 | ||
4916 | return vma; | |
4917 | } |