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 DH |
28 | #include <drm/drmP.h> |
29 | #include <drm/i915_drm.h> | |
673a394b | 30 | #include "i915_drv.h" |
1c5d22f7 | 31 | #include "i915_trace.h" |
652c393a | 32 | #include "intel_drv.h" |
5949eac4 | 33 | #include <linux/shmem_fs.h> |
5a0e3ad6 | 34 | #include <linux/slab.h> |
673a394b | 35 | #include <linux/swap.h> |
79e53945 | 36 | #include <linux/pci.h> |
1286ff73 | 37 | #include <linux/dma-buf.h> |
673a394b | 38 | |
88241785 | 39 | static __must_check int i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj); |
05394f39 CW |
40 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); |
41 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj); | |
88241785 CW |
42 | static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj, |
43 | unsigned alignment, | |
44 | bool map_and_fenceable); | |
05394f39 CW |
45 | static int i915_gem_phys_pwrite(struct drm_device *dev, |
46 | struct drm_i915_gem_object *obj, | |
71acb5eb | 47 | struct drm_i915_gem_pwrite *args, |
05394f39 | 48 | struct drm_file *file); |
673a394b | 49 | |
61050808 CW |
50 | static void i915_gem_write_fence(struct drm_device *dev, int reg, |
51 | struct drm_i915_gem_object *obj); | |
52 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, | |
53 | struct drm_i915_fence_reg *fence, | |
54 | bool enable); | |
55 | ||
17250b71 | 56 | static int i915_gem_inactive_shrink(struct shrinker *shrinker, |
1495f230 | 57 | struct shrink_control *sc); |
8c59967c | 58 | static void i915_gem_object_truncate(struct drm_i915_gem_object *obj); |
31169714 | 59 | |
61050808 CW |
60 | static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj) |
61 | { | |
62 | if (obj->tiling_mode) | |
63 | i915_gem_release_mmap(obj); | |
64 | ||
65 | /* As we do not have an associated fence register, we will force | |
66 | * a tiling change if we ever need to acquire one. | |
67 | */ | |
5d82e3e6 | 68 | obj->fence_dirty = false; |
61050808 CW |
69 | obj->fence_reg = I915_FENCE_REG_NONE; |
70 | } | |
71 | ||
73aa808f CW |
72 | /* some bookkeeping */ |
73 | static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv, | |
74 | size_t size) | |
75 | { | |
76 | dev_priv->mm.object_count++; | |
77 | dev_priv->mm.object_memory += size; | |
78 | } | |
79 | ||
80 | static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv, | |
81 | size_t size) | |
82 | { | |
83 | dev_priv->mm.object_count--; | |
84 | dev_priv->mm.object_memory -= size; | |
85 | } | |
86 | ||
21dd3734 CW |
87 | static int |
88 | i915_gem_wait_for_error(struct drm_device *dev) | |
30dbf0c0 CW |
89 | { |
90 | struct drm_i915_private *dev_priv = dev->dev_private; | |
91 | struct completion *x = &dev_priv->error_completion; | |
92 | unsigned long flags; | |
93 | int ret; | |
94 | ||
95 | if (!atomic_read(&dev_priv->mm.wedged)) | |
96 | return 0; | |
97 | ||
0a6759c6 DV |
98 | /* |
99 | * Only wait 10 seconds for the gpu reset to complete to avoid hanging | |
100 | * userspace. If it takes that long something really bad is going on and | |
101 | * we should simply try to bail out and fail as gracefully as possible. | |
102 | */ | |
103 | ret = wait_for_completion_interruptible_timeout(x, 10*HZ); | |
104 | if (ret == 0) { | |
105 | DRM_ERROR("Timed out waiting for the gpu reset to complete\n"); | |
106 | return -EIO; | |
107 | } else if (ret < 0) { | |
30dbf0c0 | 108 | return ret; |
0a6759c6 | 109 | } |
30dbf0c0 | 110 | |
21dd3734 CW |
111 | if (atomic_read(&dev_priv->mm.wedged)) { |
112 | /* GPU is hung, bump the completion count to account for | |
113 | * the token we just consumed so that we never hit zero and | |
114 | * end up waiting upon a subsequent completion event that | |
115 | * will never happen. | |
116 | */ | |
117 | spin_lock_irqsave(&x->wait.lock, flags); | |
118 | x->done++; | |
119 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
120 | } | |
121 | return 0; | |
30dbf0c0 CW |
122 | } |
123 | ||
54cf91dc | 124 | int i915_mutex_lock_interruptible(struct drm_device *dev) |
76c1dec1 | 125 | { |
76c1dec1 CW |
126 | int ret; |
127 | ||
21dd3734 | 128 | ret = i915_gem_wait_for_error(dev); |
76c1dec1 CW |
129 | if (ret) |
130 | return ret; | |
131 | ||
132 | ret = mutex_lock_interruptible(&dev->struct_mutex); | |
133 | if (ret) | |
134 | return ret; | |
135 | ||
23bc5982 | 136 | WARN_ON(i915_verify_lists(dev)); |
76c1dec1 CW |
137 | return 0; |
138 | } | |
30dbf0c0 | 139 | |
7d1c4804 | 140 | static inline bool |
05394f39 | 141 | i915_gem_object_is_inactive(struct drm_i915_gem_object *obj) |
7d1c4804 | 142 | { |
1b50247a | 143 | return !obj->active; |
7d1c4804 CW |
144 | } |
145 | ||
79e53945 JB |
146 | int |
147 | i915_gem_init_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 148 | struct drm_file *file) |
79e53945 JB |
149 | { |
150 | struct drm_i915_gem_init *args = data; | |
2021746e | 151 | |
7bb6fb8d DV |
152 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
153 | return -ENODEV; | |
154 | ||
2021746e CW |
155 | if (args->gtt_start >= args->gtt_end || |
156 | (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1)) | |
157 | return -EINVAL; | |
79e53945 | 158 | |
f534bc0b DV |
159 | /* GEM with user mode setting was never supported on ilk and later. */ |
160 | if (INTEL_INFO(dev)->gen >= 5) | |
161 | return -ENODEV; | |
162 | ||
79e53945 | 163 | mutex_lock(&dev->struct_mutex); |
644ec02b DV |
164 | i915_gem_init_global_gtt(dev, args->gtt_start, |
165 | args->gtt_end, args->gtt_end); | |
673a394b EA |
166 | mutex_unlock(&dev->struct_mutex); |
167 | ||
2021746e | 168 | return 0; |
673a394b EA |
169 | } |
170 | ||
5a125c3c EA |
171 | int |
172 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 173 | struct drm_file *file) |
5a125c3c | 174 | { |
73aa808f | 175 | struct drm_i915_private *dev_priv = dev->dev_private; |
5a125c3c | 176 | struct drm_i915_gem_get_aperture *args = data; |
6299f992 CW |
177 | struct drm_i915_gem_object *obj; |
178 | size_t pinned; | |
5a125c3c | 179 | |
6299f992 | 180 | pinned = 0; |
73aa808f | 181 | mutex_lock(&dev->struct_mutex); |
1b50247a CW |
182 | list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list) |
183 | if (obj->pin_count) | |
184 | pinned += obj->gtt_space->size; | |
73aa808f | 185 | mutex_unlock(&dev->struct_mutex); |
5a125c3c | 186 | |
6299f992 | 187 | args->aper_size = dev_priv->mm.gtt_total; |
0206e353 | 188 | args->aper_available_size = args->aper_size - pinned; |
6299f992 | 189 | |
5a125c3c EA |
190 | return 0; |
191 | } | |
192 | ||
ff72145b DA |
193 | static int |
194 | i915_gem_create(struct drm_file *file, | |
195 | struct drm_device *dev, | |
196 | uint64_t size, | |
197 | uint32_t *handle_p) | |
673a394b | 198 | { |
05394f39 | 199 | struct drm_i915_gem_object *obj; |
a1a2d1d3 PP |
200 | int ret; |
201 | u32 handle; | |
673a394b | 202 | |
ff72145b | 203 | size = roundup(size, PAGE_SIZE); |
8ffc0246 CW |
204 | if (size == 0) |
205 | return -EINVAL; | |
673a394b EA |
206 | |
207 | /* Allocate the new object */ | |
ff72145b | 208 | obj = i915_gem_alloc_object(dev, size); |
673a394b EA |
209 | if (obj == NULL) |
210 | return -ENOMEM; | |
211 | ||
05394f39 | 212 | ret = drm_gem_handle_create(file, &obj->base, &handle); |
1dfd9754 | 213 | if (ret) { |
05394f39 CW |
214 | drm_gem_object_release(&obj->base); |
215 | i915_gem_info_remove_obj(dev->dev_private, obj->base.size); | |
202f2fef | 216 | kfree(obj); |
673a394b | 217 | return ret; |
1dfd9754 | 218 | } |
673a394b | 219 | |
202f2fef | 220 | /* drop reference from allocate - handle holds it now */ |
05394f39 | 221 | drm_gem_object_unreference(&obj->base); |
202f2fef CW |
222 | trace_i915_gem_object_create(obj); |
223 | ||
ff72145b | 224 | *handle_p = handle; |
673a394b EA |
225 | return 0; |
226 | } | |
227 | ||
ff72145b DA |
228 | int |
229 | i915_gem_dumb_create(struct drm_file *file, | |
230 | struct drm_device *dev, | |
231 | struct drm_mode_create_dumb *args) | |
232 | { | |
233 | /* have to work out size/pitch and return them */ | |
ed0291fd | 234 | args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64); |
ff72145b DA |
235 | args->size = args->pitch * args->height; |
236 | return i915_gem_create(file, dev, | |
237 | args->size, &args->handle); | |
238 | } | |
239 | ||
240 | int i915_gem_dumb_destroy(struct drm_file *file, | |
241 | struct drm_device *dev, | |
242 | uint32_t handle) | |
243 | { | |
244 | return drm_gem_handle_delete(file, handle); | |
245 | } | |
246 | ||
247 | /** | |
248 | * Creates a new mm object and returns a handle to it. | |
249 | */ | |
250 | int | |
251 | i915_gem_create_ioctl(struct drm_device *dev, void *data, | |
252 | struct drm_file *file) | |
253 | { | |
254 | struct drm_i915_gem_create *args = data; | |
63ed2cb2 | 255 | |
ff72145b DA |
256 | return i915_gem_create(file, dev, |
257 | args->size, &args->handle); | |
258 | } | |
259 | ||
05394f39 | 260 | static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj) |
280b713b | 261 | { |
05394f39 | 262 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
280b713b EA |
263 | |
264 | return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 && | |
05394f39 | 265 | obj->tiling_mode != I915_TILING_NONE; |
280b713b EA |
266 | } |
267 | ||
8461d226 DV |
268 | static inline int |
269 | __copy_to_user_swizzled(char __user *cpu_vaddr, | |
270 | const char *gpu_vaddr, int gpu_offset, | |
271 | int length) | |
272 | { | |
273 | int ret, cpu_offset = 0; | |
274 | ||
275 | while (length > 0) { | |
276 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
277 | int this_length = min(cacheline_end - gpu_offset, length); | |
278 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
279 | ||
280 | ret = __copy_to_user(cpu_vaddr + cpu_offset, | |
281 | gpu_vaddr + swizzled_gpu_offset, | |
282 | this_length); | |
283 | if (ret) | |
284 | return ret + length; | |
285 | ||
286 | cpu_offset += this_length; | |
287 | gpu_offset += this_length; | |
288 | length -= this_length; | |
289 | } | |
290 | ||
291 | return 0; | |
292 | } | |
293 | ||
8c59967c | 294 | static inline int |
4f0c7cfb BW |
295 | __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, |
296 | const char __user *cpu_vaddr, | |
8c59967c DV |
297 | int length) |
298 | { | |
299 | int ret, cpu_offset = 0; | |
300 | ||
301 | while (length > 0) { | |
302 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
303 | int this_length = min(cacheline_end - gpu_offset, length); | |
304 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
305 | ||
306 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, | |
307 | cpu_vaddr + cpu_offset, | |
308 | this_length); | |
309 | if (ret) | |
310 | return ret + length; | |
311 | ||
312 | cpu_offset += this_length; | |
313 | gpu_offset += this_length; | |
314 | length -= this_length; | |
315 | } | |
316 | ||
317 | return 0; | |
318 | } | |
319 | ||
d174bd64 DV |
320 | /* Per-page copy function for the shmem pread fastpath. |
321 | * Flushes invalid cachelines before reading the target if | |
322 | * needs_clflush is set. */ | |
eb01459f | 323 | static int |
d174bd64 DV |
324 | shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length, |
325 | char __user *user_data, | |
326 | bool page_do_bit17_swizzling, bool needs_clflush) | |
327 | { | |
328 | char *vaddr; | |
329 | int ret; | |
330 | ||
e7e58eb5 | 331 | if (unlikely(page_do_bit17_swizzling)) |
d174bd64 DV |
332 | return -EINVAL; |
333 | ||
334 | vaddr = kmap_atomic(page); | |
335 | if (needs_clflush) | |
336 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
337 | page_length); | |
338 | ret = __copy_to_user_inatomic(user_data, | |
339 | vaddr + shmem_page_offset, | |
340 | page_length); | |
341 | kunmap_atomic(vaddr); | |
342 | ||
343 | return ret; | |
344 | } | |
345 | ||
23c18c71 DV |
346 | static void |
347 | shmem_clflush_swizzled_range(char *addr, unsigned long length, | |
348 | bool swizzled) | |
349 | { | |
e7e58eb5 | 350 | if (unlikely(swizzled)) { |
23c18c71 DV |
351 | unsigned long start = (unsigned long) addr; |
352 | unsigned long end = (unsigned long) addr + length; | |
353 | ||
354 | /* For swizzling simply ensure that we always flush both | |
355 | * channels. Lame, but simple and it works. Swizzled | |
356 | * pwrite/pread is far from a hotpath - current userspace | |
357 | * doesn't use it at all. */ | |
358 | start = round_down(start, 128); | |
359 | end = round_up(end, 128); | |
360 | ||
361 | drm_clflush_virt_range((void *)start, end - start); | |
362 | } else { | |
363 | drm_clflush_virt_range(addr, length); | |
364 | } | |
365 | ||
366 | } | |
367 | ||
d174bd64 DV |
368 | /* Only difference to the fast-path function is that this can handle bit17 |
369 | * and uses non-atomic copy and kmap functions. */ | |
370 | static int | |
371 | shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length, | |
372 | char __user *user_data, | |
373 | bool page_do_bit17_swizzling, bool needs_clflush) | |
374 | { | |
375 | char *vaddr; | |
376 | int ret; | |
377 | ||
378 | vaddr = kmap(page); | |
379 | if (needs_clflush) | |
23c18c71 DV |
380 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
381 | page_length, | |
382 | page_do_bit17_swizzling); | |
d174bd64 DV |
383 | |
384 | if (page_do_bit17_swizzling) | |
385 | ret = __copy_to_user_swizzled(user_data, | |
386 | vaddr, shmem_page_offset, | |
387 | page_length); | |
388 | else | |
389 | ret = __copy_to_user(user_data, | |
390 | vaddr + shmem_page_offset, | |
391 | page_length); | |
392 | kunmap(page); | |
393 | ||
394 | return ret; | |
395 | } | |
396 | ||
eb01459f | 397 | static int |
dbf7bff0 DV |
398 | i915_gem_shmem_pread(struct drm_device *dev, |
399 | struct drm_i915_gem_object *obj, | |
400 | struct drm_i915_gem_pread *args, | |
401 | struct drm_file *file) | |
eb01459f | 402 | { |
05394f39 | 403 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
8461d226 | 404 | char __user *user_data; |
eb01459f | 405 | ssize_t remain; |
8461d226 | 406 | loff_t offset; |
eb2c0c81 | 407 | int shmem_page_offset, page_length, ret = 0; |
8461d226 | 408 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
dbf7bff0 | 409 | int hit_slowpath = 0; |
96d79b52 | 410 | int prefaulted = 0; |
8489731c | 411 | int needs_clflush = 0; |
692a576b | 412 | int release_page; |
eb01459f | 413 | |
8461d226 | 414 | user_data = (char __user *) (uintptr_t) args->data_ptr; |
eb01459f EA |
415 | remain = args->size; |
416 | ||
8461d226 | 417 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
eb01459f | 418 | |
8489731c DV |
419 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) { |
420 | /* If we're not in the cpu read domain, set ourself into the gtt | |
421 | * read domain and manually flush cachelines (if required). This | |
422 | * optimizes for the case when the gpu will dirty the data | |
423 | * anyway again before the next pread happens. */ | |
424 | if (obj->cache_level == I915_CACHE_NONE) | |
425 | needs_clflush = 1; | |
426 | ret = i915_gem_object_set_to_gtt_domain(obj, false); | |
427 | if (ret) | |
428 | return ret; | |
429 | } | |
eb01459f | 430 | |
8461d226 | 431 | offset = args->offset; |
eb01459f EA |
432 | |
433 | while (remain > 0) { | |
e5281ccd CW |
434 | struct page *page; |
435 | ||
eb01459f EA |
436 | /* Operation in this page |
437 | * | |
eb01459f | 438 | * shmem_page_offset = offset within page in shmem file |
eb01459f EA |
439 | * page_length = bytes to copy for this page |
440 | */ | |
c8cbbb8b | 441 | shmem_page_offset = offset_in_page(offset); |
eb01459f EA |
442 | page_length = remain; |
443 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
444 | page_length = PAGE_SIZE - shmem_page_offset; | |
eb01459f | 445 | |
692a576b DV |
446 | if (obj->pages) { |
447 | page = obj->pages[offset >> PAGE_SHIFT]; | |
448 | release_page = 0; | |
449 | } else { | |
450 | page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT); | |
451 | if (IS_ERR(page)) { | |
452 | ret = PTR_ERR(page); | |
453 | goto out; | |
454 | } | |
455 | release_page = 1; | |
b65552f0 | 456 | } |
e5281ccd | 457 | |
8461d226 DV |
458 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
459 | (page_to_phys(page) & (1 << 17)) != 0; | |
460 | ||
d174bd64 DV |
461 | ret = shmem_pread_fast(page, shmem_page_offset, page_length, |
462 | user_data, page_do_bit17_swizzling, | |
463 | needs_clflush); | |
464 | if (ret == 0) | |
465 | goto next_page; | |
dbf7bff0 DV |
466 | |
467 | hit_slowpath = 1; | |
692a576b | 468 | page_cache_get(page); |
dbf7bff0 DV |
469 | mutex_unlock(&dev->struct_mutex); |
470 | ||
96d79b52 | 471 | if (!prefaulted) { |
f56f821f | 472 | ret = fault_in_multipages_writeable(user_data, remain); |
96d79b52 DV |
473 | /* Userspace is tricking us, but we've already clobbered |
474 | * its pages with the prefault and promised to write the | |
475 | * data up to the first fault. Hence ignore any errors | |
476 | * and just continue. */ | |
477 | (void)ret; | |
478 | prefaulted = 1; | |
479 | } | |
eb01459f | 480 | |
d174bd64 DV |
481 | ret = shmem_pread_slow(page, shmem_page_offset, page_length, |
482 | user_data, page_do_bit17_swizzling, | |
483 | needs_clflush); | |
eb01459f | 484 | |
dbf7bff0 | 485 | mutex_lock(&dev->struct_mutex); |
e5281ccd | 486 | page_cache_release(page); |
dbf7bff0 | 487 | next_page: |
e5281ccd | 488 | mark_page_accessed(page); |
692a576b DV |
489 | if (release_page) |
490 | page_cache_release(page); | |
e5281ccd | 491 | |
8461d226 DV |
492 | if (ret) { |
493 | ret = -EFAULT; | |
494 | goto out; | |
495 | } | |
496 | ||
eb01459f | 497 | remain -= page_length; |
8461d226 | 498 | user_data += page_length; |
eb01459f EA |
499 | offset += page_length; |
500 | } | |
501 | ||
4f27b75d | 502 | out: |
dbf7bff0 DV |
503 | if (hit_slowpath) { |
504 | /* Fixup: Kill any reinstated backing storage pages */ | |
505 | if (obj->madv == __I915_MADV_PURGED) | |
506 | i915_gem_object_truncate(obj); | |
507 | } | |
eb01459f EA |
508 | |
509 | return ret; | |
510 | } | |
511 | ||
673a394b EA |
512 | /** |
513 | * Reads data from the object referenced by handle. | |
514 | * | |
515 | * On error, the contents of *data are undefined. | |
516 | */ | |
517 | int | |
518 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 519 | struct drm_file *file) |
673a394b EA |
520 | { |
521 | struct drm_i915_gem_pread *args = data; | |
05394f39 | 522 | struct drm_i915_gem_object *obj; |
35b62a89 | 523 | int ret = 0; |
673a394b | 524 | |
51311d0a CW |
525 | if (args->size == 0) |
526 | return 0; | |
527 | ||
528 | if (!access_ok(VERIFY_WRITE, | |
529 | (char __user *)(uintptr_t)args->data_ptr, | |
530 | args->size)) | |
531 | return -EFAULT; | |
532 | ||
4f27b75d | 533 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 534 | if (ret) |
4f27b75d | 535 | return ret; |
673a394b | 536 | |
05394f39 | 537 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 538 | if (&obj->base == NULL) { |
1d7cfea1 CW |
539 | ret = -ENOENT; |
540 | goto unlock; | |
4f27b75d | 541 | } |
673a394b | 542 | |
7dcd2499 | 543 | /* Bounds check source. */ |
05394f39 CW |
544 | if (args->offset > obj->base.size || |
545 | args->size > obj->base.size - args->offset) { | |
ce9d419d | 546 | ret = -EINVAL; |
35b62a89 | 547 | goto out; |
ce9d419d CW |
548 | } |
549 | ||
1286ff73 DV |
550 | /* prime objects have no backing filp to GEM pread/pwrite |
551 | * pages from. | |
552 | */ | |
553 | if (!obj->base.filp) { | |
554 | ret = -EINVAL; | |
555 | goto out; | |
556 | } | |
557 | ||
db53a302 CW |
558 | trace_i915_gem_object_pread(obj, args->offset, args->size); |
559 | ||
dbf7bff0 | 560 | ret = i915_gem_shmem_pread(dev, obj, args, file); |
673a394b | 561 | |
35b62a89 | 562 | out: |
05394f39 | 563 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 564 | unlock: |
4f27b75d | 565 | mutex_unlock(&dev->struct_mutex); |
eb01459f | 566 | return ret; |
673a394b EA |
567 | } |
568 | ||
0839ccb8 KP |
569 | /* This is the fast write path which cannot handle |
570 | * page faults in the source data | |
9b7530cc | 571 | */ |
0839ccb8 KP |
572 | |
573 | static inline int | |
574 | fast_user_write(struct io_mapping *mapping, | |
575 | loff_t page_base, int page_offset, | |
576 | char __user *user_data, | |
577 | int length) | |
9b7530cc | 578 | { |
4f0c7cfb BW |
579 | void __iomem *vaddr_atomic; |
580 | void *vaddr; | |
0839ccb8 | 581 | unsigned long unwritten; |
9b7530cc | 582 | |
3e4d3af5 | 583 | vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base); |
4f0c7cfb BW |
584 | /* We can use the cpu mem copy function because this is X86. */ |
585 | vaddr = (void __force*)vaddr_atomic + page_offset; | |
586 | unwritten = __copy_from_user_inatomic_nocache(vaddr, | |
0839ccb8 | 587 | user_data, length); |
3e4d3af5 | 588 | io_mapping_unmap_atomic(vaddr_atomic); |
fbd5a26d | 589 | return unwritten; |
0839ccb8 KP |
590 | } |
591 | ||
3de09aa3 EA |
592 | /** |
593 | * This is the fast pwrite path, where we copy the data directly from the | |
594 | * user into the GTT, uncached. | |
595 | */ | |
673a394b | 596 | static int |
05394f39 CW |
597 | i915_gem_gtt_pwrite_fast(struct drm_device *dev, |
598 | struct drm_i915_gem_object *obj, | |
3de09aa3 | 599 | struct drm_i915_gem_pwrite *args, |
05394f39 | 600 | struct drm_file *file) |
673a394b | 601 | { |
0839ccb8 | 602 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b | 603 | ssize_t remain; |
0839ccb8 | 604 | loff_t offset, page_base; |
673a394b | 605 | char __user *user_data; |
935aaa69 DV |
606 | int page_offset, page_length, ret; |
607 | ||
608 | ret = i915_gem_object_pin(obj, 0, true); | |
609 | if (ret) | |
610 | goto out; | |
611 | ||
612 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
613 | if (ret) | |
614 | goto out_unpin; | |
615 | ||
616 | ret = i915_gem_object_put_fence(obj); | |
617 | if (ret) | |
618 | goto out_unpin; | |
673a394b EA |
619 | |
620 | user_data = (char __user *) (uintptr_t) args->data_ptr; | |
621 | remain = args->size; | |
673a394b | 622 | |
05394f39 | 623 | offset = obj->gtt_offset + args->offset; |
673a394b EA |
624 | |
625 | while (remain > 0) { | |
626 | /* Operation in this page | |
627 | * | |
0839ccb8 KP |
628 | * page_base = page offset within aperture |
629 | * page_offset = offset within page | |
630 | * page_length = bytes to copy for this page | |
673a394b | 631 | */ |
c8cbbb8b CW |
632 | page_base = offset & PAGE_MASK; |
633 | page_offset = offset_in_page(offset); | |
0839ccb8 KP |
634 | page_length = remain; |
635 | if ((page_offset + remain) > PAGE_SIZE) | |
636 | page_length = PAGE_SIZE - page_offset; | |
637 | ||
0839ccb8 | 638 | /* If we get a fault while copying data, then (presumably) our |
3de09aa3 EA |
639 | * source page isn't available. Return the error and we'll |
640 | * retry in the slow path. | |
0839ccb8 | 641 | */ |
fbd5a26d | 642 | if (fast_user_write(dev_priv->mm.gtt_mapping, page_base, |
935aaa69 DV |
643 | page_offset, user_data, page_length)) { |
644 | ret = -EFAULT; | |
645 | goto out_unpin; | |
646 | } | |
673a394b | 647 | |
0839ccb8 KP |
648 | remain -= page_length; |
649 | user_data += page_length; | |
650 | offset += page_length; | |
673a394b | 651 | } |
673a394b | 652 | |
935aaa69 DV |
653 | out_unpin: |
654 | i915_gem_object_unpin(obj); | |
655 | out: | |
3de09aa3 | 656 | return ret; |
673a394b EA |
657 | } |
658 | ||
d174bd64 DV |
659 | /* Per-page copy function for the shmem pwrite fastpath. |
660 | * Flushes invalid cachelines before writing to the target if | |
661 | * needs_clflush_before is set and flushes out any written cachelines after | |
662 | * writing if needs_clflush is set. */ | |
3043c60c | 663 | static int |
d174bd64 DV |
664 | shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length, |
665 | char __user *user_data, | |
666 | bool page_do_bit17_swizzling, | |
667 | bool needs_clflush_before, | |
668 | bool needs_clflush_after) | |
673a394b | 669 | { |
d174bd64 | 670 | char *vaddr; |
673a394b | 671 | int ret; |
3de09aa3 | 672 | |
e7e58eb5 | 673 | if (unlikely(page_do_bit17_swizzling)) |
d174bd64 | 674 | return -EINVAL; |
3de09aa3 | 675 | |
d174bd64 DV |
676 | vaddr = kmap_atomic(page); |
677 | if (needs_clflush_before) | |
678 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
679 | page_length); | |
680 | ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset, | |
681 | user_data, | |
682 | page_length); | |
683 | if (needs_clflush_after) | |
684 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
685 | page_length); | |
686 | kunmap_atomic(vaddr); | |
3de09aa3 EA |
687 | |
688 | return ret; | |
689 | } | |
690 | ||
d174bd64 DV |
691 | /* Only difference to the fast-path function is that this can handle bit17 |
692 | * and uses non-atomic copy and kmap functions. */ | |
3043c60c | 693 | static int |
d174bd64 DV |
694 | shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length, |
695 | char __user *user_data, | |
696 | bool page_do_bit17_swizzling, | |
697 | bool needs_clflush_before, | |
698 | bool needs_clflush_after) | |
673a394b | 699 | { |
d174bd64 DV |
700 | char *vaddr; |
701 | int ret; | |
e5281ccd | 702 | |
d174bd64 | 703 | vaddr = kmap(page); |
e7e58eb5 | 704 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) |
23c18c71 DV |
705 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
706 | page_length, | |
707 | page_do_bit17_swizzling); | |
d174bd64 DV |
708 | if (page_do_bit17_swizzling) |
709 | ret = __copy_from_user_swizzled(vaddr, shmem_page_offset, | |
e5281ccd CW |
710 | user_data, |
711 | page_length); | |
d174bd64 DV |
712 | else |
713 | ret = __copy_from_user(vaddr + shmem_page_offset, | |
714 | user_data, | |
715 | page_length); | |
716 | if (needs_clflush_after) | |
23c18c71 DV |
717 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
718 | page_length, | |
719 | page_do_bit17_swizzling); | |
d174bd64 | 720 | kunmap(page); |
40123c1f | 721 | |
d174bd64 | 722 | return ret; |
40123c1f EA |
723 | } |
724 | ||
40123c1f | 725 | static int |
e244a443 DV |
726 | i915_gem_shmem_pwrite(struct drm_device *dev, |
727 | struct drm_i915_gem_object *obj, | |
728 | struct drm_i915_gem_pwrite *args, | |
729 | struct drm_file *file) | |
40123c1f | 730 | { |
05394f39 | 731 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
40123c1f | 732 | ssize_t remain; |
8c59967c DV |
733 | loff_t offset; |
734 | char __user *user_data; | |
eb2c0c81 | 735 | int shmem_page_offset, page_length, ret = 0; |
8c59967c | 736 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
e244a443 | 737 | int hit_slowpath = 0; |
58642885 DV |
738 | int needs_clflush_after = 0; |
739 | int needs_clflush_before = 0; | |
692a576b | 740 | int release_page; |
40123c1f | 741 | |
8c59967c | 742 | user_data = (char __user *) (uintptr_t) args->data_ptr; |
40123c1f EA |
743 | remain = args->size; |
744 | ||
8c59967c | 745 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
40123c1f | 746 | |
58642885 DV |
747 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
748 | /* If we're not in the cpu write domain, set ourself into the gtt | |
749 | * write domain and manually flush cachelines (if required). This | |
750 | * optimizes for the case when the gpu will use the data | |
751 | * right away and we therefore have to clflush anyway. */ | |
752 | if (obj->cache_level == I915_CACHE_NONE) | |
753 | needs_clflush_after = 1; | |
754 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
755 | if (ret) | |
756 | return ret; | |
757 | } | |
758 | /* Same trick applies for invalidate partially written cachelines before | |
759 | * writing. */ | |
760 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU) | |
761 | && obj->cache_level == I915_CACHE_NONE) | |
762 | needs_clflush_before = 1; | |
763 | ||
673a394b | 764 | offset = args->offset; |
05394f39 | 765 | obj->dirty = 1; |
673a394b | 766 | |
40123c1f | 767 | while (remain > 0) { |
e5281ccd | 768 | struct page *page; |
58642885 | 769 | int partial_cacheline_write; |
e5281ccd | 770 | |
40123c1f EA |
771 | /* Operation in this page |
772 | * | |
40123c1f | 773 | * shmem_page_offset = offset within page in shmem file |
40123c1f EA |
774 | * page_length = bytes to copy for this page |
775 | */ | |
c8cbbb8b | 776 | shmem_page_offset = offset_in_page(offset); |
40123c1f EA |
777 | |
778 | page_length = remain; | |
779 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
780 | page_length = PAGE_SIZE - shmem_page_offset; | |
40123c1f | 781 | |
58642885 DV |
782 | /* If we don't overwrite a cacheline completely we need to be |
783 | * careful to have up-to-date data by first clflushing. Don't | |
784 | * overcomplicate things and flush the entire patch. */ | |
785 | partial_cacheline_write = needs_clflush_before && | |
786 | ((shmem_page_offset | page_length) | |
787 | & (boot_cpu_data.x86_clflush_size - 1)); | |
788 | ||
692a576b DV |
789 | if (obj->pages) { |
790 | page = obj->pages[offset >> PAGE_SHIFT]; | |
791 | release_page = 0; | |
792 | } else { | |
793 | page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT); | |
794 | if (IS_ERR(page)) { | |
795 | ret = PTR_ERR(page); | |
796 | goto out; | |
797 | } | |
798 | release_page = 1; | |
e5281ccd CW |
799 | } |
800 | ||
8c59967c DV |
801 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
802 | (page_to_phys(page) & (1 << 17)) != 0; | |
803 | ||
d174bd64 DV |
804 | ret = shmem_pwrite_fast(page, shmem_page_offset, page_length, |
805 | user_data, page_do_bit17_swizzling, | |
806 | partial_cacheline_write, | |
807 | needs_clflush_after); | |
808 | if (ret == 0) | |
809 | goto next_page; | |
e244a443 DV |
810 | |
811 | hit_slowpath = 1; | |
692a576b | 812 | page_cache_get(page); |
e244a443 DV |
813 | mutex_unlock(&dev->struct_mutex); |
814 | ||
d174bd64 DV |
815 | ret = shmem_pwrite_slow(page, shmem_page_offset, page_length, |
816 | user_data, page_do_bit17_swizzling, | |
817 | partial_cacheline_write, | |
818 | needs_clflush_after); | |
40123c1f | 819 | |
e244a443 | 820 | mutex_lock(&dev->struct_mutex); |
692a576b | 821 | page_cache_release(page); |
e244a443 | 822 | next_page: |
e5281ccd CW |
823 | set_page_dirty(page); |
824 | mark_page_accessed(page); | |
692a576b DV |
825 | if (release_page) |
826 | page_cache_release(page); | |
e5281ccd | 827 | |
8c59967c DV |
828 | if (ret) { |
829 | ret = -EFAULT; | |
830 | goto out; | |
831 | } | |
832 | ||
40123c1f | 833 | remain -= page_length; |
8c59967c | 834 | user_data += page_length; |
40123c1f | 835 | offset += page_length; |
673a394b EA |
836 | } |
837 | ||
fbd5a26d | 838 | out: |
e244a443 DV |
839 | if (hit_slowpath) { |
840 | /* Fixup: Kill any reinstated backing storage pages */ | |
841 | if (obj->madv == __I915_MADV_PURGED) | |
842 | i915_gem_object_truncate(obj); | |
843 | /* and flush dirty cachelines in case the object isn't in the cpu write | |
844 | * domain anymore. */ | |
845 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { | |
846 | i915_gem_clflush_object(obj); | |
847 | intel_gtt_chipset_flush(); | |
848 | } | |
8c59967c | 849 | } |
673a394b | 850 | |
58642885 DV |
851 | if (needs_clflush_after) |
852 | intel_gtt_chipset_flush(); | |
853 | ||
40123c1f | 854 | return ret; |
673a394b EA |
855 | } |
856 | ||
857 | /** | |
858 | * Writes data to the object referenced by handle. | |
859 | * | |
860 | * On error, the contents of the buffer that were to be modified are undefined. | |
861 | */ | |
862 | int | |
863 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, | |
fbd5a26d | 864 | struct drm_file *file) |
673a394b EA |
865 | { |
866 | struct drm_i915_gem_pwrite *args = data; | |
05394f39 | 867 | struct drm_i915_gem_object *obj; |
51311d0a CW |
868 | int ret; |
869 | ||
870 | if (args->size == 0) | |
871 | return 0; | |
872 | ||
873 | if (!access_ok(VERIFY_READ, | |
874 | (char __user *)(uintptr_t)args->data_ptr, | |
875 | args->size)) | |
876 | return -EFAULT; | |
877 | ||
f56f821f DV |
878 | ret = fault_in_multipages_readable((char __user *)(uintptr_t)args->data_ptr, |
879 | args->size); | |
51311d0a CW |
880 | if (ret) |
881 | return -EFAULT; | |
673a394b | 882 | |
fbd5a26d | 883 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 884 | if (ret) |
fbd5a26d | 885 | return ret; |
1d7cfea1 | 886 | |
05394f39 | 887 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 888 | if (&obj->base == NULL) { |
1d7cfea1 CW |
889 | ret = -ENOENT; |
890 | goto unlock; | |
fbd5a26d | 891 | } |
673a394b | 892 | |
7dcd2499 | 893 | /* Bounds check destination. */ |
05394f39 CW |
894 | if (args->offset > obj->base.size || |
895 | args->size > obj->base.size - args->offset) { | |
ce9d419d | 896 | ret = -EINVAL; |
35b62a89 | 897 | goto out; |
ce9d419d CW |
898 | } |
899 | ||
1286ff73 DV |
900 | /* prime objects have no backing filp to GEM pread/pwrite |
901 | * pages from. | |
902 | */ | |
903 | if (!obj->base.filp) { | |
904 | ret = -EINVAL; | |
905 | goto out; | |
906 | } | |
907 | ||
db53a302 CW |
908 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); |
909 | ||
935aaa69 | 910 | ret = -EFAULT; |
673a394b EA |
911 | /* We can only do the GTT pwrite on untiled buffers, as otherwise |
912 | * it would end up going through the fenced access, and we'll get | |
913 | * different detiling behavior between reading and writing. | |
914 | * pread/pwrite currently are reading and writing from the CPU | |
915 | * perspective, requiring manual detiling by the client. | |
916 | */ | |
5c0480f2 | 917 | if (obj->phys_obj) { |
fbd5a26d | 918 | ret = i915_gem_phys_pwrite(dev, obj, args, file); |
5c0480f2 DV |
919 | goto out; |
920 | } | |
921 | ||
922 | if (obj->gtt_space && | |
3ae53783 | 923 | obj->cache_level == I915_CACHE_NONE && |
c07496fa | 924 | obj->tiling_mode == I915_TILING_NONE && |
ffc62976 | 925 | obj->map_and_fenceable && |
5c0480f2 | 926 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
fbd5a26d | 927 | ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); |
935aaa69 DV |
928 | /* Note that the gtt paths might fail with non-page-backed user |
929 | * pointers (e.g. gtt mappings when moving data between | |
930 | * textures). Fallback to the shmem path in that case. */ | |
fbd5a26d | 931 | } |
673a394b | 932 | |
5c0480f2 | 933 | if (ret == -EFAULT) |
935aaa69 | 934 | ret = i915_gem_shmem_pwrite(dev, obj, args, file); |
5c0480f2 | 935 | |
35b62a89 | 936 | out: |
05394f39 | 937 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 938 | unlock: |
fbd5a26d | 939 | mutex_unlock(&dev->struct_mutex); |
673a394b EA |
940 | return ret; |
941 | } | |
942 | ||
943 | /** | |
2ef7eeaa EA |
944 | * Called when user space prepares to use an object with the CPU, either |
945 | * through the mmap ioctl's mapping or a GTT mapping. | |
673a394b EA |
946 | */ |
947 | int | |
948 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 949 | struct drm_file *file) |
673a394b EA |
950 | { |
951 | struct drm_i915_gem_set_domain *args = data; | |
05394f39 | 952 | struct drm_i915_gem_object *obj; |
2ef7eeaa EA |
953 | uint32_t read_domains = args->read_domains; |
954 | uint32_t write_domain = args->write_domain; | |
673a394b EA |
955 | int ret; |
956 | ||
2ef7eeaa | 957 | /* Only handle setting domains to types used by the CPU. */ |
21d509e3 | 958 | if (write_domain & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
959 | return -EINVAL; |
960 | ||
21d509e3 | 961 | if (read_domains & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
962 | return -EINVAL; |
963 | ||
964 | /* Having something in the write domain implies it's in the read | |
965 | * domain, and only that read domain. Enforce that in the request. | |
966 | */ | |
967 | if (write_domain != 0 && read_domains != write_domain) | |
968 | return -EINVAL; | |
969 | ||
76c1dec1 | 970 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 971 | if (ret) |
76c1dec1 | 972 | return ret; |
1d7cfea1 | 973 | |
05394f39 | 974 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 975 | if (&obj->base == NULL) { |
1d7cfea1 CW |
976 | ret = -ENOENT; |
977 | goto unlock; | |
76c1dec1 | 978 | } |
673a394b | 979 | |
2ef7eeaa EA |
980 | if (read_domains & I915_GEM_DOMAIN_GTT) { |
981 | ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); | |
02354392 EA |
982 | |
983 | /* Silently promote "you're not bound, there was nothing to do" | |
984 | * to success, since the client was just asking us to | |
985 | * make sure everything was done. | |
986 | */ | |
987 | if (ret == -EINVAL) | |
988 | ret = 0; | |
2ef7eeaa | 989 | } else { |
e47c68e9 | 990 | ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); |
2ef7eeaa EA |
991 | } |
992 | ||
05394f39 | 993 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 994 | unlock: |
673a394b EA |
995 | mutex_unlock(&dev->struct_mutex); |
996 | return ret; | |
997 | } | |
998 | ||
999 | /** | |
1000 | * Called when user space has done writes to this buffer | |
1001 | */ | |
1002 | int | |
1003 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1004 | struct drm_file *file) |
673a394b EA |
1005 | { |
1006 | struct drm_i915_gem_sw_finish *args = data; | |
05394f39 | 1007 | struct drm_i915_gem_object *obj; |
673a394b EA |
1008 | int ret = 0; |
1009 | ||
76c1dec1 | 1010 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1011 | if (ret) |
76c1dec1 | 1012 | return ret; |
1d7cfea1 | 1013 | |
05394f39 | 1014 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 1015 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1016 | ret = -ENOENT; |
1017 | goto unlock; | |
673a394b EA |
1018 | } |
1019 | ||
673a394b | 1020 | /* Pinned buffers may be scanout, so flush the cache */ |
05394f39 | 1021 | if (obj->pin_count) |
e47c68e9 EA |
1022 | i915_gem_object_flush_cpu_write_domain(obj); |
1023 | ||
05394f39 | 1024 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1025 | unlock: |
673a394b EA |
1026 | mutex_unlock(&dev->struct_mutex); |
1027 | return ret; | |
1028 | } | |
1029 | ||
1030 | /** | |
1031 | * Maps the contents of an object, returning the address it is mapped | |
1032 | * into. | |
1033 | * | |
1034 | * While the mapping holds a reference on the contents of the object, it doesn't | |
1035 | * imply a ref on the object itself. | |
1036 | */ | |
1037 | int | |
1038 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1039 | struct drm_file *file) |
673a394b EA |
1040 | { |
1041 | struct drm_i915_gem_mmap *args = data; | |
1042 | struct drm_gem_object *obj; | |
673a394b EA |
1043 | unsigned long addr; |
1044 | ||
05394f39 | 1045 | obj = drm_gem_object_lookup(dev, file, args->handle); |
673a394b | 1046 | if (obj == NULL) |
bf79cb91 | 1047 | return -ENOENT; |
673a394b | 1048 | |
1286ff73 DV |
1049 | /* prime objects have no backing filp to GEM mmap |
1050 | * pages from. | |
1051 | */ | |
1052 | if (!obj->filp) { | |
1053 | drm_gem_object_unreference_unlocked(obj); | |
1054 | return -EINVAL; | |
1055 | } | |
1056 | ||
6be5ceb0 | 1057 | addr = vm_mmap(obj->filp, 0, args->size, |
673a394b EA |
1058 | PROT_READ | PROT_WRITE, MAP_SHARED, |
1059 | args->offset); | |
bc9025bd | 1060 | drm_gem_object_unreference_unlocked(obj); |
673a394b EA |
1061 | if (IS_ERR((void *)addr)) |
1062 | return addr; | |
1063 | ||
1064 | args->addr_ptr = (uint64_t) addr; | |
1065 | ||
1066 | return 0; | |
1067 | } | |
1068 | ||
de151cf6 JB |
1069 | /** |
1070 | * i915_gem_fault - fault a page into the GTT | |
1071 | * vma: VMA in question | |
1072 | * vmf: fault info | |
1073 | * | |
1074 | * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped | |
1075 | * from userspace. The fault handler takes care of binding the object to | |
1076 | * the GTT (if needed), allocating and programming a fence register (again, | |
1077 | * only if needed based on whether the old reg is still valid or the object | |
1078 | * is tiled) and inserting a new PTE into the faulting process. | |
1079 | * | |
1080 | * Note that the faulting process may involve evicting existing objects | |
1081 | * from the GTT and/or fence registers to make room. So performance may | |
1082 | * suffer if the GTT working set is large or there are few fence registers | |
1083 | * left. | |
1084 | */ | |
1085 | int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
1086 | { | |
05394f39 CW |
1087 | struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data); |
1088 | struct drm_device *dev = obj->base.dev; | |
7d1c4804 | 1089 | drm_i915_private_t *dev_priv = dev->dev_private; |
de151cf6 JB |
1090 | pgoff_t page_offset; |
1091 | unsigned long pfn; | |
1092 | int ret = 0; | |
0f973f27 | 1093 | bool write = !!(vmf->flags & FAULT_FLAG_WRITE); |
de151cf6 JB |
1094 | |
1095 | /* We don't use vmf->pgoff since that has the fake offset */ | |
1096 | page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >> | |
1097 | PAGE_SHIFT; | |
1098 | ||
d9bc7e9f CW |
1099 | ret = i915_mutex_lock_interruptible(dev); |
1100 | if (ret) | |
1101 | goto out; | |
a00b10c3 | 1102 | |
db53a302 CW |
1103 | trace_i915_gem_object_fault(obj, page_offset, true, write); |
1104 | ||
d9bc7e9f | 1105 | /* Now bind it into the GTT if needed */ |
919926ae CW |
1106 | if (!obj->map_and_fenceable) { |
1107 | ret = i915_gem_object_unbind(obj); | |
1108 | if (ret) | |
1109 | goto unlock; | |
a00b10c3 | 1110 | } |
05394f39 | 1111 | if (!obj->gtt_space) { |
75e9e915 | 1112 | ret = i915_gem_object_bind_to_gtt(obj, 0, true); |
c715089f CW |
1113 | if (ret) |
1114 | goto unlock; | |
de151cf6 | 1115 | |
e92d03bf EA |
1116 | ret = i915_gem_object_set_to_gtt_domain(obj, write); |
1117 | if (ret) | |
1118 | goto unlock; | |
1119 | } | |
4a684a41 | 1120 | |
74898d7e DV |
1121 | if (!obj->has_global_gtt_mapping) |
1122 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
1123 | ||
06d98131 | 1124 | ret = i915_gem_object_get_fence(obj); |
d9e86c0e CW |
1125 | if (ret) |
1126 | goto unlock; | |
de151cf6 | 1127 | |
05394f39 CW |
1128 | if (i915_gem_object_is_inactive(obj)) |
1129 | list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list); | |
7d1c4804 | 1130 | |
6299f992 CW |
1131 | obj->fault_mappable = true; |
1132 | ||
dd2757f8 | 1133 | pfn = ((dev_priv->mm.gtt_base_addr + obj->gtt_offset) >> PAGE_SHIFT) + |
de151cf6 JB |
1134 | page_offset; |
1135 | ||
1136 | /* Finally, remap it using the new GTT offset */ | |
1137 | ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn); | |
c715089f | 1138 | unlock: |
de151cf6 | 1139 | mutex_unlock(&dev->struct_mutex); |
d9bc7e9f | 1140 | out: |
de151cf6 | 1141 | switch (ret) { |
d9bc7e9f | 1142 | case -EIO: |
a9340cca DV |
1143 | /* If this -EIO is due to a gpu hang, give the reset code a |
1144 | * chance to clean up the mess. Otherwise return the proper | |
1145 | * SIGBUS. */ | |
1146 | if (!atomic_read(&dev_priv->mm.wedged)) | |
1147 | return VM_FAULT_SIGBUS; | |
045e769a | 1148 | case -EAGAIN: |
d9bc7e9f CW |
1149 | /* Give the error handler a chance to run and move the |
1150 | * objects off the GPU active list. Next time we service the | |
1151 | * fault, we should be able to transition the page into the | |
1152 | * GTT without touching the GPU (and so avoid further | |
1153 | * EIO/EGAIN). If the GPU is wedged, then there is no issue | |
1154 | * with coherency, just lost writes. | |
1155 | */ | |
045e769a | 1156 | set_need_resched(); |
c715089f CW |
1157 | case 0: |
1158 | case -ERESTARTSYS: | |
bed636ab | 1159 | case -EINTR: |
c715089f | 1160 | return VM_FAULT_NOPAGE; |
de151cf6 | 1161 | case -ENOMEM: |
de151cf6 | 1162 | return VM_FAULT_OOM; |
de151cf6 | 1163 | default: |
c715089f | 1164 | return VM_FAULT_SIGBUS; |
de151cf6 JB |
1165 | } |
1166 | } | |
1167 | ||
901782b2 CW |
1168 | /** |
1169 | * i915_gem_release_mmap - remove physical page mappings | |
1170 | * @obj: obj in question | |
1171 | * | |
af901ca1 | 1172 | * Preserve the reservation of the mmapping with the DRM core code, but |
901782b2 CW |
1173 | * relinquish ownership of the pages back to the system. |
1174 | * | |
1175 | * It is vital that we remove the page mapping if we have mapped a tiled | |
1176 | * object through the GTT and then lose the fence register due to | |
1177 | * resource pressure. Similarly if the object has been moved out of the | |
1178 | * aperture, than pages mapped into userspace must be revoked. Removing the | |
1179 | * mapping will then trigger a page fault on the next user access, allowing | |
1180 | * fixup by i915_gem_fault(). | |
1181 | */ | |
d05ca301 | 1182 | void |
05394f39 | 1183 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) |
901782b2 | 1184 | { |
6299f992 CW |
1185 | if (!obj->fault_mappable) |
1186 | return; | |
901782b2 | 1187 | |
f6e47884 CW |
1188 | if (obj->base.dev->dev_mapping) |
1189 | unmap_mapping_range(obj->base.dev->dev_mapping, | |
1190 | (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT, | |
1191 | obj->base.size, 1); | |
fb7d516a | 1192 | |
6299f992 | 1193 | obj->fault_mappable = false; |
901782b2 CW |
1194 | } |
1195 | ||
92b88aeb | 1196 | static uint32_t |
e28f8711 | 1197 | i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode) |
92b88aeb | 1198 | { |
e28f8711 | 1199 | uint32_t gtt_size; |
92b88aeb CW |
1200 | |
1201 | if (INTEL_INFO(dev)->gen >= 4 || | |
e28f8711 CW |
1202 | tiling_mode == I915_TILING_NONE) |
1203 | return size; | |
92b88aeb CW |
1204 | |
1205 | /* Previous chips need a power-of-two fence region when tiling */ | |
1206 | if (INTEL_INFO(dev)->gen == 3) | |
e28f8711 | 1207 | gtt_size = 1024*1024; |
92b88aeb | 1208 | else |
e28f8711 | 1209 | gtt_size = 512*1024; |
92b88aeb | 1210 | |
e28f8711 CW |
1211 | while (gtt_size < size) |
1212 | gtt_size <<= 1; | |
92b88aeb | 1213 | |
e28f8711 | 1214 | return gtt_size; |
92b88aeb CW |
1215 | } |
1216 | ||
de151cf6 JB |
1217 | /** |
1218 | * i915_gem_get_gtt_alignment - return required GTT alignment for an object | |
1219 | * @obj: object to check | |
1220 | * | |
1221 | * Return the required GTT alignment for an object, taking into account | |
5e783301 | 1222 | * potential fence register mapping. |
de151cf6 JB |
1223 | */ |
1224 | static uint32_t | |
e28f8711 CW |
1225 | i915_gem_get_gtt_alignment(struct drm_device *dev, |
1226 | uint32_t size, | |
1227 | int tiling_mode) | |
de151cf6 | 1228 | { |
de151cf6 JB |
1229 | /* |
1230 | * Minimum alignment is 4k (GTT page size), but might be greater | |
1231 | * if a fence register is needed for the object. | |
1232 | */ | |
a00b10c3 | 1233 | if (INTEL_INFO(dev)->gen >= 4 || |
e28f8711 | 1234 | tiling_mode == I915_TILING_NONE) |
de151cf6 JB |
1235 | return 4096; |
1236 | ||
a00b10c3 CW |
1237 | /* |
1238 | * Previous chips need to be aligned to the size of the smallest | |
1239 | * fence register that can contain the object. | |
1240 | */ | |
e28f8711 | 1241 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
a00b10c3 CW |
1242 | } |
1243 | ||
5e783301 DV |
1244 | /** |
1245 | * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an | |
1246 | * unfenced object | |
e28f8711 CW |
1247 | * @dev: the device |
1248 | * @size: size of the object | |
1249 | * @tiling_mode: tiling mode of the object | |
5e783301 DV |
1250 | * |
1251 | * Return the required GTT alignment for an object, only taking into account | |
1252 | * unfenced tiled surface requirements. | |
1253 | */ | |
467cffba | 1254 | uint32_t |
e28f8711 CW |
1255 | i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev, |
1256 | uint32_t size, | |
1257 | int tiling_mode) | |
5e783301 | 1258 | { |
5e783301 DV |
1259 | /* |
1260 | * Minimum alignment is 4k (GTT page size) for sane hw. | |
1261 | */ | |
1262 | if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) || | |
e28f8711 | 1263 | tiling_mode == I915_TILING_NONE) |
5e783301 DV |
1264 | return 4096; |
1265 | ||
e28f8711 CW |
1266 | /* Previous hardware however needs to be aligned to a power-of-two |
1267 | * tile height. The simplest method for determining this is to reuse | |
1268 | * the power-of-tile object size. | |
5e783301 | 1269 | */ |
e28f8711 | 1270 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
5e783301 DV |
1271 | } |
1272 | ||
de151cf6 | 1273 | int |
ff72145b DA |
1274 | i915_gem_mmap_gtt(struct drm_file *file, |
1275 | struct drm_device *dev, | |
1276 | uint32_t handle, | |
1277 | uint64_t *offset) | |
de151cf6 | 1278 | { |
da761a6e | 1279 | struct drm_i915_private *dev_priv = dev->dev_private; |
05394f39 | 1280 | struct drm_i915_gem_object *obj; |
de151cf6 JB |
1281 | int ret; |
1282 | ||
76c1dec1 | 1283 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1284 | if (ret) |
76c1dec1 | 1285 | return ret; |
de151cf6 | 1286 | |
ff72145b | 1287 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle)); |
c8725226 | 1288 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1289 | ret = -ENOENT; |
1290 | goto unlock; | |
1291 | } | |
de151cf6 | 1292 | |
05394f39 | 1293 | if (obj->base.size > dev_priv->mm.gtt_mappable_end) { |
da761a6e | 1294 | ret = -E2BIG; |
ff56b0bc | 1295 | goto out; |
da761a6e CW |
1296 | } |
1297 | ||
05394f39 | 1298 | if (obj->madv != I915_MADV_WILLNEED) { |
ab18282d | 1299 | DRM_ERROR("Attempting to mmap a purgeable buffer\n"); |
1d7cfea1 CW |
1300 | ret = -EINVAL; |
1301 | goto out; | |
ab18282d CW |
1302 | } |
1303 | ||
05394f39 | 1304 | if (!obj->base.map_list.map) { |
b464e9a2 | 1305 | ret = drm_gem_create_mmap_offset(&obj->base); |
1d7cfea1 CW |
1306 | if (ret) |
1307 | goto out; | |
de151cf6 JB |
1308 | } |
1309 | ||
ff72145b | 1310 | *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT; |
de151cf6 | 1311 | |
1d7cfea1 | 1312 | out: |
05394f39 | 1313 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1314 | unlock: |
de151cf6 | 1315 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 1316 | return ret; |
de151cf6 JB |
1317 | } |
1318 | ||
ff72145b DA |
1319 | /** |
1320 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing | |
1321 | * @dev: DRM device | |
1322 | * @data: GTT mapping ioctl data | |
1323 | * @file: GEM object info | |
1324 | * | |
1325 | * Simply returns the fake offset to userspace so it can mmap it. | |
1326 | * The mmap call will end up in drm_gem_mmap(), which will set things | |
1327 | * up so we can get faults in the handler above. | |
1328 | * | |
1329 | * The fault handler will take care of binding the object into the GTT | |
1330 | * (since it may have been evicted to make room for something), allocating | |
1331 | * a fence register, and mapping the appropriate aperture address into | |
1332 | * userspace. | |
1333 | */ | |
1334 | int | |
1335 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, | |
1336 | struct drm_file *file) | |
1337 | { | |
1338 | struct drm_i915_gem_mmap_gtt *args = data; | |
1339 | ||
ff72145b DA |
1340 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); |
1341 | } | |
1342 | ||
1286ff73 | 1343 | int |
05394f39 | 1344 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj, |
e5281ccd CW |
1345 | gfp_t gfpmask) |
1346 | { | |
e5281ccd CW |
1347 | int page_count, i; |
1348 | struct address_space *mapping; | |
1349 | struct inode *inode; | |
1350 | struct page *page; | |
1351 | ||
1286ff73 DV |
1352 | if (obj->pages || obj->sg_table) |
1353 | return 0; | |
1354 | ||
e5281ccd CW |
1355 | /* Get the list of pages out of our struct file. They'll be pinned |
1356 | * at this point until we release them. | |
1357 | */ | |
05394f39 CW |
1358 | page_count = obj->base.size / PAGE_SIZE; |
1359 | BUG_ON(obj->pages != NULL); | |
1360 | obj->pages = drm_malloc_ab(page_count, sizeof(struct page *)); | |
1361 | if (obj->pages == NULL) | |
e5281ccd CW |
1362 | return -ENOMEM; |
1363 | ||
05394f39 | 1364 | inode = obj->base.filp->f_path.dentry->d_inode; |
e5281ccd | 1365 | mapping = inode->i_mapping; |
5949eac4 HD |
1366 | gfpmask |= mapping_gfp_mask(mapping); |
1367 | ||
e5281ccd | 1368 | for (i = 0; i < page_count; i++) { |
5949eac4 | 1369 | page = shmem_read_mapping_page_gfp(mapping, i, gfpmask); |
e5281ccd CW |
1370 | if (IS_ERR(page)) |
1371 | goto err_pages; | |
1372 | ||
05394f39 | 1373 | obj->pages[i] = page; |
e5281ccd CW |
1374 | } |
1375 | ||
6dacfd2f | 1376 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
e5281ccd CW |
1377 | i915_gem_object_do_bit_17_swizzle(obj); |
1378 | ||
1379 | return 0; | |
1380 | ||
1381 | err_pages: | |
1382 | while (i--) | |
05394f39 | 1383 | page_cache_release(obj->pages[i]); |
e5281ccd | 1384 | |
05394f39 CW |
1385 | drm_free_large(obj->pages); |
1386 | obj->pages = NULL; | |
e5281ccd CW |
1387 | return PTR_ERR(page); |
1388 | } | |
1389 | ||
5cdf5881 | 1390 | static void |
05394f39 | 1391 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj) |
673a394b | 1392 | { |
05394f39 | 1393 | int page_count = obj->base.size / PAGE_SIZE; |
673a394b EA |
1394 | int i; |
1395 | ||
1286ff73 DV |
1396 | if (!obj->pages) |
1397 | return; | |
1398 | ||
05394f39 | 1399 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
673a394b | 1400 | |
6dacfd2f | 1401 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
280b713b EA |
1402 | i915_gem_object_save_bit_17_swizzle(obj); |
1403 | ||
05394f39 CW |
1404 | if (obj->madv == I915_MADV_DONTNEED) |
1405 | obj->dirty = 0; | |
3ef94daa CW |
1406 | |
1407 | for (i = 0; i < page_count; i++) { | |
05394f39 CW |
1408 | if (obj->dirty) |
1409 | set_page_dirty(obj->pages[i]); | |
3ef94daa | 1410 | |
05394f39 CW |
1411 | if (obj->madv == I915_MADV_WILLNEED) |
1412 | mark_page_accessed(obj->pages[i]); | |
3ef94daa | 1413 | |
05394f39 | 1414 | page_cache_release(obj->pages[i]); |
3ef94daa | 1415 | } |
05394f39 | 1416 | obj->dirty = 0; |
673a394b | 1417 | |
05394f39 CW |
1418 | drm_free_large(obj->pages); |
1419 | obj->pages = NULL; | |
673a394b EA |
1420 | } |
1421 | ||
54cf91dc | 1422 | void |
05394f39 | 1423 | i915_gem_object_move_to_active(struct drm_i915_gem_object *obj, |
1ec14ad3 CW |
1424 | struct intel_ring_buffer *ring, |
1425 | u32 seqno) | |
673a394b | 1426 | { |
05394f39 | 1427 | struct drm_device *dev = obj->base.dev; |
69dc4987 | 1428 | struct drm_i915_private *dev_priv = dev->dev_private; |
617dbe27 | 1429 | |
852835f3 | 1430 | BUG_ON(ring == NULL); |
05394f39 | 1431 | obj->ring = ring; |
673a394b EA |
1432 | |
1433 | /* Add a reference if we're newly entering the active list. */ | |
05394f39 CW |
1434 | if (!obj->active) { |
1435 | drm_gem_object_reference(&obj->base); | |
1436 | obj->active = 1; | |
673a394b | 1437 | } |
e35a41de | 1438 | |
673a394b | 1439 | /* Move from whatever list we were on to the tail of execution. */ |
05394f39 CW |
1440 | list_move_tail(&obj->mm_list, &dev_priv->mm.active_list); |
1441 | list_move_tail(&obj->ring_list, &ring->active_list); | |
caea7476 | 1442 | |
05394f39 | 1443 | obj->last_rendering_seqno = seqno; |
caea7476 | 1444 | |
7dd49065 | 1445 | if (obj->fenced_gpu_access) { |
caea7476 | 1446 | obj->last_fenced_seqno = seqno; |
caea7476 | 1447 | |
7dd49065 CW |
1448 | /* Bump MRU to take account of the delayed flush */ |
1449 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
1450 | struct drm_i915_fence_reg *reg; | |
1451 | ||
1452 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
1453 | list_move_tail(®->lru_list, | |
1454 | &dev_priv->mm.fence_list); | |
1455 | } | |
caea7476 CW |
1456 | } |
1457 | } | |
1458 | ||
1459 | static void | |
1460 | i915_gem_object_move_off_active(struct drm_i915_gem_object *obj) | |
1461 | { | |
1462 | list_del_init(&obj->ring_list); | |
1463 | obj->last_rendering_seqno = 0; | |
15a13bbd | 1464 | obj->last_fenced_seqno = 0; |
673a394b EA |
1465 | } |
1466 | ||
ce44b0ea | 1467 | static void |
05394f39 | 1468 | i915_gem_object_move_to_flushing(struct drm_i915_gem_object *obj) |
ce44b0ea | 1469 | { |
05394f39 | 1470 | struct drm_device *dev = obj->base.dev; |
ce44b0ea | 1471 | drm_i915_private_t *dev_priv = dev->dev_private; |
ce44b0ea | 1472 | |
05394f39 CW |
1473 | BUG_ON(!obj->active); |
1474 | list_move_tail(&obj->mm_list, &dev_priv->mm.flushing_list); | |
caea7476 CW |
1475 | |
1476 | i915_gem_object_move_off_active(obj); | |
1477 | } | |
1478 | ||
1479 | static void | |
1480 | i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj) | |
1481 | { | |
1482 | struct drm_device *dev = obj->base.dev; | |
1483 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1484 | ||
1b50247a | 1485 | list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list); |
caea7476 CW |
1486 | |
1487 | BUG_ON(!list_empty(&obj->gpu_write_list)); | |
1488 | BUG_ON(!obj->active); | |
1489 | obj->ring = NULL; | |
1490 | ||
1491 | i915_gem_object_move_off_active(obj); | |
1492 | obj->fenced_gpu_access = false; | |
caea7476 CW |
1493 | |
1494 | obj->active = 0; | |
87ca9c8a | 1495 | obj->pending_gpu_write = false; |
caea7476 CW |
1496 | drm_gem_object_unreference(&obj->base); |
1497 | ||
1498 | WARN_ON(i915_verify_lists(dev)); | |
ce44b0ea | 1499 | } |
673a394b | 1500 | |
963b4836 CW |
1501 | /* Immediately discard the backing storage */ |
1502 | static void | |
05394f39 | 1503 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) |
963b4836 | 1504 | { |
bb6baf76 | 1505 | struct inode *inode; |
963b4836 | 1506 | |
ae9fed6b CW |
1507 | /* Our goal here is to return as much of the memory as |
1508 | * is possible back to the system as we are called from OOM. | |
1509 | * To do this we must instruct the shmfs to drop all of its | |
e2377fe0 | 1510 | * backing pages, *now*. |
ae9fed6b | 1511 | */ |
05394f39 | 1512 | inode = obj->base.filp->f_path.dentry->d_inode; |
e2377fe0 | 1513 | shmem_truncate_range(inode, 0, (loff_t)-1); |
bb6baf76 | 1514 | |
a14917ee CW |
1515 | if (obj->base.map_list.map) |
1516 | drm_gem_free_mmap_offset(&obj->base); | |
1517 | ||
05394f39 | 1518 | obj->madv = __I915_MADV_PURGED; |
963b4836 CW |
1519 | } |
1520 | ||
1521 | static inline int | |
05394f39 | 1522 | i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj) |
963b4836 | 1523 | { |
05394f39 | 1524 | return obj->madv == I915_MADV_DONTNEED; |
963b4836 CW |
1525 | } |
1526 | ||
63560396 | 1527 | static void |
db53a302 CW |
1528 | i915_gem_process_flushing_list(struct intel_ring_buffer *ring, |
1529 | uint32_t flush_domains) | |
63560396 | 1530 | { |
05394f39 | 1531 | struct drm_i915_gem_object *obj, *next; |
63560396 | 1532 | |
05394f39 | 1533 | list_for_each_entry_safe(obj, next, |
64193406 | 1534 | &ring->gpu_write_list, |
63560396 | 1535 | gpu_write_list) { |
05394f39 CW |
1536 | if (obj->base.write_domain & flush_domains) { |
1537 | uint32_t old_write_domain = obj->base.write_domain; | |
63560396 | 1538 | |
05394f39 CW |
1539 | obj->base.write_domain = 0; |
1540 | list_del_init(&obj->gpu_write_list); | |
1ec14ad3 | 1541 | i915_gem_object_move_to_active(obj, ring, |
db53a302 | 1542 | i915_gem_next_request_seqno(ring)); |
63560396 | 1543 | |
63560396 | 1544 | trace_i915_gem_object_change_domain(obj, |
05394f39 | 1545 | obj->base.read_domains, |
63560396 DV |
1546 | old_write_domain); |
1547 | } | |
1548 | } | |
1549 | } | |
8187a2b7 | 1550 | |
53d227f2 DV |
1551 | static u32 |
1552 | i915_gem_get_seqno(struct drm_device *dev) | |
1553 | { | |
1554 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1555 | u32 seqno = dev_priv->next_seqno; | |
1556 | ||
1557 | /* reserve 0 for non-seqno */ | |
1558 | if (++dev_priv->next_seqno == 0) | |
1559 | dev_priv->next_seqno = 1; | |
1560 | ||
1561 | return seqno; | |
1562 | } | |
1563 | ||
1564 | u32 | |
1565 | i915_gem_next_request_seqno(struct intel_ring_buffer *ring) | |
1566 | { | |
1567 | if (ring->outstanding_lazy_request == 0) | |
1568 | ring->outstanding_lazy_request = i915_gem_get_seqno(ring->dev); | |
1569 | ||
1570 | return ring->outstanding_lazy_request; | |
1571 | } | |
1572 | ||
3cce469c | 1573 | int |
db53a302 | 1574 | i915_add_request(struct intel_ring_buffer *ring, |
f787a5f5 | 1575 | struct drm_file *file, |
db53a302 | 1576 | struct drm_i915_gem_request *request) |
673a394b | 1577 | { |
db53a302 | 1578 | drm_i915_private_t *dev_priv = ring->dev->dev_private; |
673a394b | 1579 | uint32_t seqno; |
a71d8d94 | 1580 | u32 request_ring_position; |
673a394b | 1581 | int was_empty; |
3cce469c CW |
1582 | int ret; |
1583 | ||
cc889e0f DV |
1584 | /* |
1585 | * Emit any outstanding flushes - execbuf can fail to emit the flush | |
1586 | * after having emitted the batchbuffer command. Hence we need to fix | |
1587 | * things up similar to emitting the lazy request. The difference here | |
1588 | * is that the flush _must_ happen before the next request, no matter | |
1589 | * what. | |
1590 | */ | |
1591 | if (ring->gpu_caches_dirty) { | |
1592 | ret = i915_gem_flush_ring(ring, 0, I915_GEM_GPU_DOMAINS); | |
1593 | if (ret) | |
1594 | return ret; | |
1595 | ||
1596 | ring->gpu_caches_dirty = false; | |
1597 | } | |
1598 | ||
3cce469c | 1599 | BUG_ON(request == NULL); |
53d227f2 | 1600 | seqno = i915_gem_next_request_seqno(ring); |
673a394b | 1601 | |
a71d8d94 CW |
1602 | /* Record the position of the start of the request so that |
1603 | * should we detect the updated seqno part-way through the | |
1604 | * GPU processing the request, we never over-estimate the | |
1605 | * position of the head. | |
1606 | */ | |
1607 | request_ring_position = intel_ring_get_tail(ring); | |
1608 | ||
3cce469c CW |
1609 | ret = ring->add_request(ring, &seqno); |
1610 | if (ret) | |
1611 | return ret; | |
673a394b | 1612 | |
db53a302 | 1613 | trace_i915_gem_request_add(ring, seqno); |
673a394b EA |
1614 | |
1615 | request->seqno = seqno; | |
852835f3 | 1616 | request->ring = ring; |
a71d8d94 | 1617 | request->tail = request_ring_position; |
673a394b | 1618 | request->emitted_jiffies = jiffies; |
852835f3 ZN |
1619 | was_empty = list_empty(&ring->request_list); |
1620 | list_add_tail(&request->list, &ring->request_list); | |
1621 | ||
db53a302 CW |
1622 | if (file) { |
1623 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
1624 | ||
1c25595f | 1625 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 1626 | request->file_priv = file_priv; |
b962442e | 1627 | list_add_tail(&request->client_list, |
f787a5f5 | 1628 | &file_priv->mm.request_list); |
1c25595f | 1629 | spin_unlock(&file_priv->mm.lock); |
b962442e | 1630 | } |
673a394b | 1631 | |
5391d0cf | 1632 | ring->outstanding_lazy_request = 0; |
db53a302 | 1633 | |
f65d9421 | 1634 | if (!dev_priv->mm.suspended) { |
3e0dc6b0 BW |
1635 | if (i915_enable_hangcheck) { |
1636 | mod_timer(&dev_priv->hangcheck_timer, | |
1637 | jiffies + | |
1638 | msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD)); | |
1639 | } | |
f65d9421 | 1640 | if (was_empty) |
b3b079db CW |
1641 | queue_delayed_work(dev_priv->wq, |
1642 | &dev_priv->mm.retire_work, HZ); | |
f65d9421 | 1643 | } |
cc889e0f DV |
1644 | |
1645 | WARN_ON(!list_empty(&ring->gpu_write_list)); | |
1646 | ||
3cce469c | 1647 | return 0; |
673a394b EA |
1648 | } |
1649 | ||
f787a5f5 CW |
1650 | static inline void |
1651 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) | |
673a394b | 1652 | { |
1c25595f | 1653 | struct drm_i915_file_private *file_priv = request->file_priv; |
673a394b | 1654 | |
1c25595f CW |
1655 | if (!file_priv) |
1656 | return; | |
1c5d22f7 | 1657 | |
1c25595f | 1658 | spin_lock(&file_priv->mm.lock); |
09bfa517 HRK |
1659 | if (request->file_priv) { |
1660 | list_del(&request->client_list); | |
1661 | request->file_priv = NULL; | |
1662 | } | |
1c25595f | 1663 | spin_unlock(&file_priv->mm.lock); |
673a394b | 1664 | } |
673a394b | 1665 | |
dfaae392 CW |
1666 | static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv, |
1667 | struct intel_ring_buffer *ring) | |
9375e446 | 1668 | { |
dfaae392 CW |
1669 | while (!list_empty(&ring->request_list)) { |
1670 | struct drm_i915_gem_request *request; | |
673a394b | 1671 | |
dfaae392 CW |
1672 | request = list_first_entry(&ring->request_list, |
1673 | struct drm_i915_gem_request, | |
1674 | list); | |
de151cf6 | 1675 | |
dfaae392 | 1676 | list_del(&request->list); |
f787a5f5 | 1677 | i915_gem_request_remove_from_client(request); |
dfaae392 CW |
1678 | kfree(request); |
1679 | } | |
673a394b | 1680 | |
dfaae392 | 1681 | while (!list_empty(&ring->active_list)) { |
05394f39 | 1682 | struct drm_i915_gem_object *obj; |
9375e446 | 1683 | |
05394f39 CW |
1684 | obj = list_first_entry(&ring->active_list, |
1685 | struct drm_i915_gem_object, | |
1686 | ring_list); | |
9375e446 | 1687 | |
05394f39 CW |
1688 | obj->base.write_domain = 0; |
1689 | list_del_init(&obj->gpu_write_list); | |
1690 | i915_gem_object_move_to_inactive(obj); | |
673a394b EA |
1691 | } |
1692 | } | |
1693 | ||
312817a3 CW |
1694 | static void i915_gem_reset_fences(struct drm_device *dev) |
1695 | { | |
1696 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1697 | int i; | |
1698 | ||
4b9de737 | 1699 | for (i = 0; i < dev_priv->num_fence_regs; i++) { |
312817a3 | 1700 | struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i]; |
7d2cb39c | 1701 | |
ada726c7 | 1702 | i915_gem_write_fence(dev, i, NULL); |
7d2cb39c | 1703 | |
ada726c7 CW |
1704 | if (reg->obj) |
1705 | i915_gem_object_fence_lost(reg->obj); | |
7d2cb39c | 1706 | |
ada726c7 CW |
1707 | reg->pin_count = 0; |
1708 | reg->obj = NULL; | |
1709 | INIT_LIST_HEAD(®->lru_list); | |
312817a3 | 1710 | } |
ada726c7 CW |
1711 | |
1712 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); | |
312817a3 CW |
1713 | } |
1714 | ||
069efc1d | 1715 | void i915_gem_reset(struct drm_device *dev) |
673a394b | 1716 | { |
77f01230 | 1717 | struct drm_i915_private *dev_priv = dev->dev_private; |
05394f39 | 1718 | struct drm_i915_gem_object *obj; |
b4519513 | 1719 | struct intel_ring_buffer *ring; |
1ec14ad3 | 1720 | int i; |
673a394b | 1721 | |
b4519513 CW |
1722 | for_each_ring(ring, dev_priv, i) |
1723 | i915_gem_reset_ring_lists(dev_priv, ring); | |
dfaae392 CW |
1724 | |
1725 | /* Remove anything from the flushing lists. The GPU cache is likely | |
1726 | * to be lost on reset along with the data, so simply move the | |
1727 | * lost bo to the inactive list. | |
1728 | */ | |
1729 | while (!list_empty(&dev_priv->mm.flushing_list)) { | |
0206e353 | 1730 | obj = list_first_entry(&dev_priv->mm.flushing_list, |
05394f39 CW |
1731 | struct drm_i915_gem_object, |
1732 | mm_list); | |
dfaae392 | 1733 | |
05394f39 CW |
1734 | obj->base.write_domain = 0; |
1735 | list_del_init(&obj->gpu_write_list); | |
1736 | i915_gem_object_move_to_inactive(obj); | |
dfaae392 CW |
1737 | } |
1738 | ||
1739 | /* Move everything out of the GPU domains to ensure we do any | |
1740 | * necessary invalidation upon reuse. | |
1741 | */ | |
05394f39 | 1742 | list_for_each_entry(obj, |
77f01230 | 1743 | &dev_priv->mm.inactive_list, |
69dc4987 | 1744 | mm_list) |
77f01230 | 1745 | { |
05394f39 | 1746 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; |
77f01230 | 1747 | } |
069efc1d CW |
1748 | |
1749 | /* The fence registers are invalidated so clear them out */ | |
312817a3 | 1750 | i915_gem_reset_fences(dev); |
673a394b EA |
1751 | } |
1752 | ||
1753 | /** | |
1754 | * This function clears the request list as sequence numbers are passed. | |
1755 | */ | |
a71d8d94 | 1756 | void |
db53a302 | 1757 | i915_gem_retire_requests_ring(struct intel_ring_buffer *ring) |
673a394b | 1758 | { |
673a394b | 1759 | uint32_t seqno; |
1ec14ad3 | 1760 | int i; |
673a394b | 1761 | |
db53a302 | 1762 | if (list_empty(&ring->request_list)) |
6c0594a3 KW |
1763 | return; |
1764 | ||
db53a302 | 1765 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b | 1766 | |
78501eac | 1767 | seqno = ring->get_seqno(ring); |
1ec14ad3 | 1768 | |
076e2c0e | 1769 | for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++) |
1ec14ad3 CW |
1770 | if (seqno >= ring->sync_seqno[i]) |
1771 | ring->sync_seqno[i] = 0; | |
1772 | ||
852835f3 | 1773 | while (!list_empty(&ring->request_list)) { |
673a394b | 1774 | struct drm_i915_gem_request *request; |
673a394b | 1775 | |
852835f3 | 1776 | request = list_first_entry(&ring->request_list, |
673a394b EA |
1777 | struct drm_i915_gem_request, |
1778 | list); | |
673a394b | 1779 | |
dfaae392 | 1780 | if (!i915_seqno_passed(seqno, request->seqno)) |
b84d5f0c CW |
1781 | break; |
1782 | ||
db53a302 | 1783 | trace_i915_gem_request_retire(ring, request->seqno); |
a71d8d94 CW |
1784 | /* We know the GPU must have read the request to have |
1785 | * sent us the seqno + interrupt, so use the position | |
1786 | * of tail of the request to update the last known position | |
1787 | * of the GPU head. | |
1788 | */ | |
1789 | ring->last_retired_head = request->tail; | |
b84d5f0c CW |
1790 | |
1791 | list_del(&request->list); | |
f787a5f5 | 1792 | i915_gem_request_remove_from_client(request); |
b84d5f0c CW |
1793 | kfree(request); |
1794 | } | |
673a394b | 1795 | |
b84d5f0c CW |
1796 | /* Move any buffers on the active list that are no longer referenced |
1797 | * by the ringbuffer to the flushing/inactive lists as appropriate. | |
1798 | */ | |
1799 | while (!list_empty(&ring->active_list)) { | |
05394f39 | 1800 | struct drm_i915_gem_object *obj; |
b84d5f0c | 1801 | |
0206e353 | 1802 | obj = list_first_entry(&ring->active_list, |
05394f39 CW |
1803 | struct drm_i915_gem_object, |
1804 | ring_list); | |
673a394b | 1805 | |
05394f39 | 1806 | if (!i915_seqno_passed(seqno, obj->last_rendering_seqno)) |
673a394b | 1807 | break; |
b84d5f0c | 1808 | |
05394f39 | 1809 | if (obj->base.write_domain != 0) |
b84d5f0c CW |
1810 | i915_gem_object_move_to_flushing(obj); |
1811 | else | |
1812 | i915_gem_object_move_to_inactive(obj); | |
673a394b | 1813 | } |
9d34e5db | 1814 | |
db53a302 CW |
1815 | if (unlikely(ring->trace_irq_seqno && |
1816 | i915_seqno_passed(seqno, ring->trace_irq_seqno))) { | |
1ec14ad3 | 1817 | ring->irq_put(ring); |
db53a302 | 1818 | ring->trace_irq_seqno = 0; |
9d34e5db | 1819 | } |
23bc5982 | 1820 | |
db53a302 | 1821 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b EA |
1822 | } |
1823 | ||
b09a1fec CW |
1824 | void |
1825 | i915_gem_retire_requests(struct drm_device *dev) | |
1826 | { | |
1827 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 1828 | struct intel_ring_buffer *ring; |
1ec14ad3 | 1829 | int i; |
b09a1fec | 1830 | |
b4519513 CW |
1831 | for_each_ring(ring, dev_priv, i) |
1832 | i915_gem_retire_requests_ring(ring); | |
b09a1fec CW |
1833 | } |
1834 | ||
75ef9da2 | 1835 | static void |
673a394b EA |
1836 | i915_gem_retire_work_handler(struct work_struct *work) |
1837 | { | |
1838 | drm_i915_private_t *dev_priv; | |
1839 | struct drm_device *dev; | |
b4519513 | 1840 | struct intel_ring_buffer *ring; |
0a58705b CW |
1841 | bool idle; |
1842 | int i; | |
673a394b EA |
1843 | |
1844 | dev_priv = container_of(work, drm_i915_private_t, | |
1845 | mm.retire_work.work); | |
1846 | dev = dev_priv->dev; | |
1847 | ||
891b48cf CW |
1848 | /* Come back later if the device is busy... */ |
1849 | if (!mutex_trylock(&dev->struct_mutex)) { | |
1850 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ); | |
1851 | return; | |
1852 | } | |
1853 | ||
b09a1fec | 1854 | i915_gem_retire_requests(dev); |
d1b851fc | 1855 | |
0a58705b CW |
1856 | /* Send a periodic flush down the ring so we don't hold onto GEM |
1857 | * objects indefinitely. | |
1858 | */ | |
1859 | idle = true; | |
b4519513 | 1860 | for_each_ring(ring, dev_priv, i) { |
cc889e0f | 1861 | if (ring->gpu_caches_dirty) { |
0a58705b | 1862 | struct drm_i915_gem_request *request; |
0a58705b | 1863 | |
0a58705b | 1864 | request = kzalloc(sizeof(*request), GFP_KERNEL); |
cc889e0f | 1865 | if (request == NULL || |
db53a302 | 1866 | i915_add_request(ring, NULL, request)) |
0a58705b CW |
1867 | kfree(request); |
1868 | } | |
1869 | ||
1870 | idle &= list_empty(&ring->request_list); | |
1871 | } | |
1872 | ||
1873 | if (!dev_priv->mm.suspended && !idle) | |
9c9fe1f8 | 1874 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ); |
0a58705b | 1875 | |
673a394b EA |
1876 | mutex_unlock(&dev->struct_mutex); |
1877 | } | |
1878 | ||
d6b2c790 DV |
1879 | int |
1880 | i915_gem_check_wedge(struct drm_i915_private *dev_priv, | |
1881 | bool interruptible) | |
b4aca010 | 1882 | { |
b4aca010 BW |
1883 | if (atomic_read(&dev_priv->mm.wedged)) { |
1884 | struct completion *x = &dev_priv->error_completion; | |
1885 | bool recovery_complete; | |
1886 | unsigned long flags; | |
1887 | ||
1888 | /* Give the error handler a chance to run. */ | |
1889 | spin_lock_irqsave(&x->wait.lock, flags); | |
1890 | recovery_complete = x->done > 0; | |
1891 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
1892 | ||
d6b2c790 DV |
1893 | /* Non-interruptible callers can't handle -EAGAIN, hence return |
1894 | * -EIO unconditionally for these. */ | |
1895 | if (!interruptible) | |
1896 | return -EIO; | |
1897 | ||
1898 | /* Recovery complete, but still wedged means reset failure. */ | |
1899 | if (recovery_complete) | |
1900 | return -EIO; | |
1901 | ||
1902 | return -EAGAIN; | |
b4aca010 BW |
1903 | } |
1904 | ||
1905 | return 0; | |
1906 | } | |
1907 | ||
1908 | /* | |
1909 | * Compare seqno against outstanding lazy request. Emit a request if they are | |
1910 | * equal. | |
1911 | */ | |
1912 | static int | |
1913 | i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno) | |
1914 | { | |
1915 | int ret = 0; | |
1916 | ||
1917 | BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex)); | |
1918 | ||
1919 | if (seqno == ring->outstanding_lazy_request) { | |
1920 | struct drm_i915_gem_request *request; | |
1921 | ||
1922 | request = kzalloc(sizeof(*request), GFP_KERNEL); | |
1923 | if (request == NULL) | |
1924 | return -ENOMEM; | |
1925 | ||
1926 | ret = i915_add_request(ring, NULL, request); | |
1927 | if (ret) { | |
1928 | kfree(request); | |
1929 | return ret; | |
1930 | } | |
1931 | ||
1932 | BUG_ON(seqno != request->seqno); | |
1933 | } | |
1934 | ||
1935 | return ret; | |
1936 | } | |
1937 | ||
5c81fe85 BW |
1938 | /** |
1939 | * __wait_seqno - wait until execution of seqno has finished | |
1940 | * @ring: the ring expected to report seqno | |
1941 | * @seqno: duh! | |
1942 | * @interruptible: do an interruptible wait (normally yes) | |
1943 | * @timeout: in - how long to wait (NULL forever); out - how much time remaining | |
1944 | * | |
1945 | * Returns 0 if the seqno was found within the alloted time. Else returns the | |
1946 | * errno with remaining time filled in timeout argument. | |
1947 | */ | |
604dd3ec | 1948 | static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno, |
5c81fe85 | 1949 | bool interruptible, struct timespec *timeout) |
604dd3ec BW |
1950 | { |
1951 | drm_i915_private_t *dev_priv = ring->dev->dev_private; | |
5c81fe85 BW |
1952 | struct timespec before, now, wait_time={1,0}; |
1953 | unsigned long timeout_jiffies; | |
1954 | long end; | |
1955 | bool wait_forever = true; | |
d6b2c790 | 1956 | int ret; |
604dd3ec BW |
1957 | |
1958 | if (i915_seqno_passed(ring->get_seqno(ring), seqno)) | |
1959 | return 0; | |
1960 | ||
1961 | trace_i915_gem_request_wait_begin(ring, seqno); | |
5c81fe85 BW |
1962 | |
1963 | if (timeout != NULL) { | |
1964 | wait_time = *timeout; | |
1965 | wait_forever = false; | |
1966 | } | |
1967 | ||
1968 | timeout_jiffies = timespec_to_jiffies(&wait_time); | |
1969 | ||
604dd3ec BW |
1970 | if (WARN_ON(!ring->irq_get(ring))) |
1971 | return -ENODEV; | |
1972 | ||
5c81fe85 BW |
1973 | /* Record current time in case interrupted by signal, or wedged * */ |
1974 | getrawmonotonic(&before); | |
1975 | ||
604dd3ec BW |
1976 | #define EXIT_COND \ |
1977 | (i915_seqno_passed(ring->get_seqno(ring), seqno) || \ | |
1978 | atomic_read(&dev_priv->mm.wedged)) | |
5c81fe85 BW |
1979 | do { |
1980 | if (interruptible) | |
1981 | end = wait_event_interruptible_timeout(ring->irq_queue, | |
1982 | EXIT_COND, | |
1983 | timeout_jiffies); | |
1984 | else | |
1985 | end = wait_event_timeout(ring->irq_queue, EXIT_COND, | |
1986 | timeout_jiffies); | |
604dd3ec | 1987 | |
d6b2c790 DV |
1988 | ret = i915_gem_check_wedge(dev_priv, interruptible); |
1989 | if (ret) | |
1990 | end = ret; | |
5c81fe85 BW |
1991 | } while (end == 0 && wait_forever); |
1992 | ||
1993 | getrawmonotonic(&now); | |
604dd3ec BW |
1994 | |
1995 | ring->irq_put(ring); | |
1996 | trace_i915_gem_request_wait_end(ring, seqno); | |
1997 | #undef EXIT_COND | |
1998 | ||
5c81fe85 BW |
1999 | if (timeout) { |
2000 | struct timespec sleep_time = timespec_sub(now, before); | |
2001 | *timeout = timespec_sub(*timeout, sleep_time); | |
2002 | } | |
2003 | ||
2004 | switch (end) { | |
eeef9b38 | 2005 | case -EIO: |
5c81fe85 BW |
2006 | case -EAGAIN: /* Wedged */ |
2007 | case -ERESTARTSYS: /* Signal */ | |
2008 | return (int)end; | |
2009 | case 0: /* Timeout */ | |
2010 | if (timeout) | |
2011 | set_normalized_timespec(timeout, 0, 0); | |
2012 | return -ETIME; | |
2013 | default: /* Completed */ | |
2014 | WARN_ON(end < 0); /* We're not aware of other errors */ | |
2015 | return 0; | |
2016 | } | |
604dd3ec BW |
2017 | } |
2018 | ||
db53a302 CW |
2019 | /** |
2020 | * Waits for a sequence number to be signaled, and cleans up the | |
2021 | * request and object lists appropriately for that event. | |
2022 | */ | |
5a5a0c64 | 2023 | int |
199b2bc2 | 2024 | i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno) |
673a394b | 2025 | { |
db53a302 | 2026 | drm_i915_private_t *dev_priv = ring->dev->dev_private; |
673a394b EA |
2027 | int ret = 0; |
2028 | ||
2029 | BUG_ON(seqno == 0); | |
2030 | ||
d6b2c790 | 2031 | ret = i915_gem_check_wedge(dev_priv, dev_priv->mm.interruptible); |
b4aca010 BW |
2032 | if (ret) |
2033 | return ret; | |
3cce469c | 2034 | |
b4aca010 BW |
2035 | ret = i915_gem_check_olr(ring, seqno); |
2036 | if (ret) | |
2037 | return ret; | |
ffed1d09 | 2038 | |
5c81fe85 | 2039 | ret = __wait_seqno(ring, seqno, dev_priv->mm.interruptible, NULL); |
673a394b | 2040 | |
673a394b EA |
2041 | return ret; |
2042 | } | |
2043 | ||
673a394b EA |
2044 | /** |
2045 | * Ensures that all rendering to the object has completed and the object is | |
2046 | * safe to unbind from the GTT or access from the CPU. | |
2047 | */ | |
54cf91dc | 2048 | int |
ce453d81 | 2049 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj) |
673a394b | 2050 | { |
673a394b EA |
2051 | int ret; |
2052 | ||
e47c68e9 EA |
2053 | /* This function only exists to support waiting for existing rendering, |
2054 | * not for emitting required flushes. | |
673a394b | 2055 | */ |
05394f39 | 2056 | BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0); |
673a394b EA |
2057 | |
2058 | /* If there is rendering queued on the buffer being evicted, wait for | |
2059 | * it. | |
2060 | */ | |
05394f39 | 2061 | if (obj->active) { |
199b2bc2 | 2062 | ret = i915_wait_seqno(obj->ring, obj->last_rendering_seqno); |
2cf34d7b | 2063 | if (ret) |
673a394b | 2064 | return ret; |
b2da9fe5 | 2065 | i915_gem_retire_requests_ring(obj->ring); |
673a394b EA |
2066 | } |
2067 | ||
2068 | return 0; | |
2069 | } | |
2070 | ||
30dfebf3 DV |
2071 | /** |
2072 | * Ensures that an object will eventually get non-busy by flushing any required | |
2073 | * write domains, emitting any outstanding lazy request and retiring and | |
2074 | * completed requests. | |
2075 | */ | |
2076 | static int | |
2077 | i915_gem_object_flush_active(struct drm_i915_gem_object *obj) | |
2078 | { | |
2079 | int ret; | |
2080 | ||
2081 | if (obj->active) { | |
2082 | ret = i915_gem_object_flush_gpu_write_domain(obj); | |
2083 | if (ret) | |
2084 | return ret; | |
2085 | ||
2086 | ret = i915_gem_check_olr(obj->ring, | |
2087 | obj->last_rendering_seqno); | |
2088 | if (ret) | |
2089 | return ret; | |
2090 | i915_gem_retire_requests_ring(obj->ring); | |
2091 | } | |
2092 | ||
2093 | return 0; | |
2094 | } | |
2095 | ||
23ba4fd0 BW |
2096 | /** |
2097 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT | |
2098 | * @DRM_IOCTL_ARGS: standard ioctl arguments | |
2099 | * | |
2100 | * Returns 0 if successful, else an error is returned with the remaining time in | |
2101 | * the timeout parameter. | |
2102 | * -ETIME: object is still busy after timeout | |
2103 | * -ERESTARTSYS: signal interrupted the wait | |
2104 | * -ENONENT: object doesn't exist | |
2105 | * Also possible, but rare: | |
2106 | * -EAGAIN: GPU wedged | |
2107 | * -ENOMEM: damn | |
2108 | * -ENODEV: Internal IRQ fail | |
2109 | * -E?: The add request failed | |
2110 | * | |
2111 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any | |
2112 | * non-zero timeout parameter the wait ioctl will wait for the given number of | |
2113 | * nanoseconds on an object becoming unbusy. Since the wait itself does so | |
2114 | * without holding struct_mutex the object may become re-busied before this | |
2115 | * function completes. A similar but shorter * race condition exists in the busy | |
2116 | * ioctl | |
2117 | */ | |
2118 | int | |
2119 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
2120 | { | |
2121 | struct drm_i915_gem_wait *args = data; | |
2122 | struct drm_i915_gem_object *obj; | |
2123 | struct intel_ring_buffer *ring = NULL; | |
eac1f14f | 2124 | struct timespec timeout_stack, *timeout = NULL; |
23ba4fd0 BW |
2125 | u32 seqno = 0; |
2126 | int ret = 0; | |
2127 | ||
eac1f14f BW |
2128 | if (args->timeout_ns >= 0) { |
2129 | timeout_stack = ns_to_timespec(args->timeout_ns); | |
2130 | timeout = &timeout_stack; | |
2131 | } | |
23ba4fd0 BW |
2132 | |
2133 | ret = i915_mutex_lock_interruptible(dev); | |
2134 | if (ret) | |
2135 | return ret; | |
2136 | ||
2137 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle)); | |
2138 | if (&obj->base == NULL) { | |
2139 | mutex_unlock(&dev->struct_mutex); | |
2140 | return -ENOENT; | |
2141 | } | |
2142 | ||
30dfebf3 DV |
2143 | /* Need to make sure the object gets inactive eventually. */ |
2144 | ret = i915_gem_object_flush_active(obj); | |
23ba4fd0 BW |
2145 | if (ret) |
2146 | goto out; | |
2147 | ||
2148 | if (obj->active) { | |
2149 | seqno = obj->last_rendering_seqno; | |
2150 | ring = obj->ring; | |
2151 | } | |
2152 | ||
2153 | if (seqno == 0) | |
2154 | goto out; | |
2155 | ||
23ba4fd0 BW |
2156 | /* Do this after OLR check to make sure we make forward progress polling |
2157 | * on this IOCTL with a 0 timeout (like busy ioctl) | |
2158 | */ | |
2159 | if (!args->timeout_ns) { | |
2160 | ret = -ETIME; | |
2161 | goto out; | |
2162 | } | |
2163 | ||
2164 | drm_gem_object_unreference(&obj->base); | |
2165 | mutex_unlock(&dev->struct_mutex); | |
2166 | ||
eac1f14f BW |
2167 | ret = __wait_seqno(ring, seqno, true, timeout); |
2168 | if (timeout) { | |
2169 | WARN_ON(!timespec_valid(timeout)); | |
2170 | args->timeout_ns = timespec_to_ns(timeout); | |
2171 | } | |
23ba4fd0 BW |
2172 | return ret; |
2173 | ||
2174 | out: | |
2175 | drm_gem_object_unreference(&obj->base); | |
2176 | mutex_unlock(&dev->struct_mutex); | |
2177 | return ret; | |
2178 | } | |
2179 | ||
5816d648 BW |
2180 | /** |
2181 | * i915_gem_object_sync - sync an object to a ring. | |
2182 | * | |
2183 | * @obj: object which may be in use on another ring. | |
2184 | * @to: ring we wish to use the object on. May be NULL. | |
2185 | * | |
2186 | * This code is meant to abstract object synchronization with the GPU. | |
2187 | * Calling with NULL implies synchronizing the object with the CPU | |
2188 | * rather than a particular GPU ring. | |
2189 | * | |
2190 | * Returns 0 if successful, else propagates up the lower layer error. | |
2191 | */ | |
2911a35b BW |
2192 | int |
2193 | i915_gem_object_sync(struct drm_i915_gem_object *obj, | |
2194 | struct intel_ring_buffer *to) | |
2195 | { | |
2196 | struct intel_ring_buffer *from = obj->ring; | |
2197 | u32 seqno; | |
2198 | int ret, idx; | |
2199 | ||
2200 | if (from == NULL || to == from) | |
2201 | return 0; | |
2202 | ||
5816d648 | 2203 | if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev)) |
2911a35b BW |
2204 | return i915_gem_object_wait_rendering(obj); |
2205 | ||
2206 | idx = intel_ring_sync_index(from, to); | |
2207 | ||
2208 | seqno = obj->last_rendering_seqno; | |
2209 | if (seqno <= from->sync_seqno[idx]) | |
2210 | return 0; | |
2211 | ||
b4aca010 BW |
2212 | ret = i915_gem_check_olr(obj->ring, seqno); |
2213 | if (ret) | |
2214 | return ret; | |
2911a35b | 2215 | |
1500f7ea | 2216 | ret = to->sync_to(to, from, seqno); |
e3a5a225 BW |
2217 | if (!ret) |
2218 | from->sync_seqno[idx] = seqno; | |
2911a35b | 2219 | |
e3a5a225 | 2220 | return ret; |
2911a35b BW |
2221 | } |
2222 | ||
b5ffc9bc CW |
2223 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) |
2224 | { | |
2225 | u32 old_write_domain, old_read_domains; | |
2226 | ||
b5ffc9bc CW |
2227 | /* Act a barrier for all accesses through the GTT */ |
2228 | mb(); | |
2229 | ||
2230 | /* Force a pagefault for domain tracking on next user access */ | |
2231 | i915_gem_release_mmap(obj); | |
2232 | ||
b97c3d9c KP |
2233 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
2234 | return; | |
2235 | ||
b5ffc9bc CW |
2236 | old_read_domains = obj->base.read_domains; |
2237 | old_write_domain = obj->base.write_domain; | |
2238 | ||
2239 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; | |
2240 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; | |
2241 | ||
2242 | trace_i915_gem_object_change_domain(obj, | |
2243 | old_read_domains, | |
2244 | old_write_domain); | |
2245 | } | |
2246 | ||
673a394b EA |
2247 | /** |
2248 | * Unbinds an object from the GTT aperture. | |
2249 | */ | |
0f973f27 | 2250 | int |
05394f39 | 2251 | i915_gem_object_unbind(struct drm_i915_gem_object *obj) |
673a394b | 2252 | { |
7bddb01f | 2253 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
673a394b EA |
2254 | int ret = 0; |
2255 | ||
05394f39 | 2256 | if (obj->gtt_space == NULL) |
673a394b EA |
2257 | return 0; |
2258 | ||
31d8d651 CW |
2259 | if (obj->pin_count) |
2260 | return -EBUSY; | |
673a394b | 2261 | |
a8198eea | 2262 | ret = i915_gem_object_finish_gpu(obj); |
1488fc08 | 2263 | if (ret) |
a8198eea CW |
2264 | return ret; |
2265 | /* Continue on if we fail due to EIO, the GPU is hung so we | |
2266 | * should be safe and we need to cleanup or else we might | |
2267 | * cause memory corruption through use-after-free. | |
2268 | */ | |
2269 | ||
b5ffc9bc | 2270 | i915_gem_object_finish_gtt(obj); |
5323fd04 | 2271 | |
673a394b EA |
2272 | /* Move the object to the CPU domain to ensure that |
2273 | * any possible CPU writes while it's not in the GTT | |
a8198eea | 2274 | * are flushed when we go to remap it. |
673a394b | 2275 | */ |
a8198eea CW |
2276 | if (ret == 0) |
2277 | ret = i915_gem_object_set_to_cpu_domain(obj, 1); | |
8dc1775d | 2278 | if (ret == -ERESTARTSYS) |
673a394b | 2279 | return ret; |
812ed492 | 2280 | if (ret) { |
a8198eea CW |
2281 | /* In the event of a disaster, abandon all caches and |
2282 | * hope for the best. | |
2283 | */ | |
812ed492 | 2284 | i915_gem_clflush_object(obj); |
05394f39 | 2285 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
812ed492 | 2286 | } |
673a394b | 2287 | |
96b47b65 | 2288 | /* release the fence reg _after_ flushing */ |
d9e86c0e | 2289 | ret = i915_gem_object_put_fence(obj); |
1488fc08 | 2290 | if (ret) |
d9e86c0e | 2291 | return ret; |
96b47b65 | 2292 | |
db53a302 CW |
2293 | trace_i915_gem_object_unbind(obj); |
2294 | ||
74898d7e DV |
2295 | if (obj->has_global_gtt_mapping) |
2296 | i915_gem_gtt_unbind_object(obj); | |
7bddb01f DV |
2297 | if (obj->has_aliasing_ppgtt_mapping) { |
2298 | i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj); | |
2299 | obj->has_aliasing_ppgtt_mapping = 0; | |
2300 | } | |
74163907 | 2301 | i915_gem_gtt_finish_object(obj); |
7bddb01f | 2302 | |
e5281ccd | 2303 | i915_gem_object_put_pages_gtt(obj); |
673a394b | 2304 | |
6299f992 | 2305 | list_del_init(&obj->gtt_list); |
05394f39 | 2306 | list_del_init(&obj->mm_list); |
75e9e915 | 2307 | /* Avoid an unnecessary call to unbind on rebind. */ |
05394f39 | 2308 | obj->map_and_fenceable = true; |
673a394b | 2309 | |
05394f39 CW |
2310 | drm_mm_put_block(obj->gtt_space); |
2311 | obj->gtt_space = NULL; | |
2312 | obj->gtt_offset = 0; | |
673a394b | 2313 | |
05394f39 | 2314 | if (i915_gem_object_is_purgeable(obj)) |
963b4836 CW |
2315 | i915_gem_object_truncate(obj); |
2316 | ||
8dc1775d | 2317 | return ret; |
673a394b EA |
2318 | } |
2319 | ||
88241785 | 2320 | int |
db53a302 | 2321 | i915_gem_flush_ring(struct intel_ring_buffer *ring, |
54cf91dc CW |
2322 | uint32_t invalidate_domains, |
2323 | uint32_t flush_domains) | |
2324 | { | |
88241785 CW |
2325 | int ret; |
2326 | ||
36d527de CW |
2327 | if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0) |
2328 | return 0; | |
2329 | ||
db53a302 CW |
2330 | trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains); |
2331 | ||
88241785 CW |
2332 | ret = ring->flush(ring, invalidate_domains, flush_domains); |
2333 | if (ret) | |
2334 | return ret; | |
2335 | ||
36d527de CW |
2336 | if (flush_domains & I915_GEM_GPU_DOMAINS) |
2337 | i915_gem_process_flushing_list(ring, flush_domains); | |
2338 | ||
88241785 | 2339 | return 0; |
54cf91dc CW |
2340 | } |
2341 | ||
b2da9fe5 | 2342 | static int i915_ring_idle(struct intel_ring_buffer *ring) |
a56ba56c | 2343 | { |
88241785 CW |
2344 | int ret; |
2345 | ||
395b70be | 2346 | if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list)) |
64193406 CW |
2347 | return 0; |
2348 | ||
88241785 | 2349 | if (!list_empty(&ring->gpu_write_list)) { |
db53a302 | 2350 | ret = i915_gem_flush_ring(ring, |
0ac74c6b | 2351 | I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS); |
88241785 CW |
2352 | if (ret) |
2353 | return ret; | |
2354 | } | |
2355 | ||
199b2bc2 | 2356 | return i915_wait_seqno(ring, i915_gem_next_request_seqno(ring)); |
a56ba56c CW |
2357 | } |
2358 | ||
b2da9fe5 | 2359 | int i915_gpu_idle(struct drm_device *dev) |
4df2faf4 DV |
2360 | { |
2361 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 2362 | struct intel_ring_buffer *ring; |
1ec14ad3 | 2363 | int ret, i; |
4df2faf4 | 2364 | |
4df2faf4 | 2365 | /* Flush everything onto the inactive list. */ |
b4519513 | 2366 | for_each_ring(ring, dev_priv, i) { |
b6c7488d BW |
2367 | ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID); |
2368 | if (ret) | |
2369 | return ret; | |
2370 | ||
b4519513 | 2371 | ret = i915_ring_idle(ring); |
1ec14ad3 CW |
2372 | if (ret) |
2373 | return ret; | |
b4519513 CW |
2374 | |
2375 | /* Is the device fubar? */ | |
2376 | if (WARN_ON(!list_empty(&ring->gpu_write_list))) | |
2377 | return -EBUSY; | |
1ec14ad3 | 2378 | } |
4df2faf4 | 2379 | |
8a1a49f9 | 2380 | return 0; |
4df2faf4 DV |
2381 | } |
2382 | ||
9ce079e4 CW |
2383 | static void sandybridge_write_fence_reg(struct drm_device *dev, int reg, |
2384 | struct drm_i915_gem_object *obj) | |
4e901fdc | 2385 | { |
4e901fdc | 2386 | drm_i915_private_t *dev_priv = dev->dev_private; |
4e901fdc EA |
2387 | uint64_t val; |
2388 | ||
9ce079e4 CW |
2389 | if (obj) { |
2390 | u32 size = obj->gtt_space->size; | |
4e901fdc | 2391 | |
9ce079e4 CW |
2392 | val = (uint64_t)((obj->gtt_offset + size - 4096) & |
2393 | 0xfffff000) << 32; | |
2394 | val |= obj->gtt_offset & 0xfffff000; | |
2395 | val |= (uint64_t)((obj->stride / 128) - 1) << | |
2396 | SANDYBRIDGE_FENCE_PITCH_SHIFT; | |
4e901fdc | 2397 | |
9ce079e4 CW |
2398 | if (obj->tiling_mode == I915_TILING_Y) |
2399 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; | |
2400 | val |= I965_FENCE_REG_VALID; | |
2401 | } else | |
2402 | val = 0; | |
c6642782 | 2403 | |
9ce079e4 CW |
2404 | I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + reg * 8, val); |
2405 | POSTING_READ(FENCE_REG_SANDYBRIDGE_0 + reg * 8); | |
4e901fdc EA |
2406 | } |
2407 | ||
9ce079e4 CW |
2408 | static void i965_write_fence_reg(struct drm_device *dev, int reg, |
2409 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2410 | { |
de151cf6 | 2411 | drm_i915_private_t *dev_priv = dev->dev_private; |
de151cf6 JB |
2412 | uint64_t val; |
2413 | ||
9ce079e4 CW |
2414 | if (obj) { |
2415 | u32 size = obj->gtt_space->size; | |
de151cf6 | 2416 | |
9ce079e4 CW |
2417 | val = (uint64_t)((obj->gtt_offset + size - 4096) & |
2418 | 0xfffff000) << 32; | |
2419 | val |= obj->gtt_offset & 0xfffff000; | |
2420 | val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT; | |
2421 | if (obj->tiling_mode == I915_TILING_Y) | |
2422 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; | |
2423 | val |= I965_FENCE_REG_VALID; | |
2424 | } else | |
2425 | val = 0; | |
c6642782 | 2426 | |
9ce079e4 CW |
2427 | I915_WRITE64(FENCE_REG_965_0 + reg * 8, val); |
2428 | POSTING_READ(FENCE_REG_965_0 + reg * 8); | |
de151cf6 JB |
2429 | } |
2430 | ||
9ce079e4 CW |
2431 | static void i915_write_fence_reg(struct drm_device *dev, int reg, |
2432 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2433 | { |
de151cf6 | 2434 | drm_i915_private_t *dev_priv = dev->dev_private; |
9ce079e4 | 2435 | u32 val; |
de151cf6 | 2436 | |
9ce079e4 CW |
2437 | if (obj) { |
2438 | u32 size = obj->gtt_space->size; | |
2439 | int pitch_val; | |
2440 | int tile_width; | |
c6642782 | 2441 | |
9ce079e4 CW |
2442 | WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) || |
2443 | (size & -size) != size || | |
2444 | (obj->gtt_offset & (size - 1)), | |
2445 | "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n", | |
2446 | obj->gtt_offset, obj->map_and_fenceable, size); | |
c6642782 | 2447 | |
9ce079e4 CW |
2448 | if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)) |
2449 | tile_width = 128; | |
2450 | else | |
2451 | tile_width = 512; | |
2452 | ||
2453 | /* Note: pitch better be a power of two tile widths */ | |
2454 | pitch_val = obj->stride / tile_width; | |
2455 | pitch_val = ffs(pitch_val) - 1; | |
2456 | ||
2457 | val = obj->gtt_offset; | |
2458 | if (obj->tiling_mode == I915_TILING_Y) | |
2459 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
2460 | val |= I915_FENCE_SIZE_BITS(size); | |
2461 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
2462 | val |= I830_FENCE_REG_VALID; | |
2463 | } else | |
2464 | val = 0; | |
2465 | ||
2466 | if (reg < 8) | |
2467 | reg = FENCE_REG_830_0 + reg * 4; | |
2468 | else | |
2469 | reg = FENCE_REG_945_8 + (reg - 8) * 4; | |
2470 | ||
2471 | I915_WRITE(reg, val); | |
2472 | POSTING_READ(reg); | |
de151cf6 JB |
2473 | } |
2474 | ||
9ce079e4 CW |
2475 | static void i830_write_fence_reg(struct drm_device *dev, int reg, |
2476 | struct drm_i915_gem_object *obj) | |
de151cf6 | 2477 | { |
de151cf6 | 2478 | drm_i915_private_t *dev_priv = dev->dev_private; |
de151cf6 | 2479 | uint32_t val; |
de151cf6 | 2480 | |
9ce079e4 CW |
2481 | if (obj) { |
2482 | u32 size = obj->gtt_space->size; | |
2483 | uint32_t pitch_val; | |
de151cf6 | 2484 | |
9ce079e4 CW |
2485 | WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) || |
2486 | (size & -size) != size || | |
2487 | (obj->gtt_offset & (size - 1)), | |
2488 | "object 0x%08x not 512K or pot-size 0x%08x aligned\n", | |
2489 | obj->gtt_offset, size); | |
e76a16de | 2490 | |
9ce079e4 CW |
2491 | pitch_val = obj->stride / 128; |
2492 | pitch_val = ffs(pitch_val) - 1; | |
de151cf6 | 2493 | |
9ce079e4 CW |
2494 | val = obj->gtt_offset; |
2495 | if (obj->tiling_mode == I915_TILING_Y) | |
2496 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
2497 | val |= I830_FENCE_SIZE_BITS(size); | |
2498 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
2499 | val |= I830_FENCE_REG_VALID; | |
2500 | } else | |
2501 | val = 0; | |
c6642782 | 2502 | |
9ce079e4 CW |
2503 | I915_WRITE(FENCE_REG_830_0 + reg * 4, val); |
2504 | POSTING_READ(FENCE_REG_830_0 + reg * 4); | |
2505 | } | |
2506 | ||
2507 | static void i915_gem_write_fence(struct drm_device *dev, int reg, | |
2508 | struct drm_i915_gem_object *obj) | |
2509 | { | |
2510 | switch (INTEL_INFO(dev)->gen) { | |
2511 | case 7: | |
2512 | case 6: sandybridge_write_fence_reg(dev, reg, obj); break; | |
2513 | case 5: | |
2514 | case 4: i965_write_fence_reg(dev, reg, obj); break; | |
2515 | case 3: i915_write_fence_reg(dev, reg, obj); break; | |
2516 | case 2: i830_write_fence_reg(dev, reg, obj); break; | |
2517 | default: break; | |
2518 | } | |
de151cf6 JB |
2519 | } |
2520 | ||
61050808 CW |
2521 | static inline int fence_number(struct drm_i915_private *dev_priv, |
2522 | struct drm_i915_fence_reg *fence) | |
2523 | { | |
2524 | return fence - dev_priv->fence_regs; | |
2525 | } | |
2526 | ||
2527 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, | |
2528 | struct drm_i915_fence_reg *fence, | |
2529 | bool enable) | |
2530 | { | |
2531 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
2532 | int reg = fence_number(dev_priv, fence); | |
2533 | ||
2534 | i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL); | |
2535 | ||
2536 | if (enable) { | |
2537 | obj->fence_reg = reg; | |
2538 | fence->obj = obj; | |
2539 | list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list); | |
2540 | } else { | |
2541 | obj->fence_reg = I915_FENCE_REG_NONE; | |
2542 | fence->obj = NULL; | |
2543 | list_del_init(&fence->lru_list); | |
2544 | } | |
2545 | } | |
2546 | ||
d9e86c0e | 2547 | static int |
a360bb1a | 2548 | i915_gem_object_flush_fence(struct drm_i915_gem_object *obj) |
d9e86c0e CW |
2549 | { |
2550 | int ret; | |
2551 | ||
2552 | if (obj->fenced_gpu_access) { | |
88241785 | 2553 | if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) { |
1c293ea3 | 2554 | ret = i915_gem_flush_ring(obj->ring, |
88241785 CW |
2555 | 0, obj->base.write_domain); |
2556 | if (ret) | |
2557 | return ret; | |
2558 | } | |
d9e86c0e CW |
2559 | |
2560 | obj->fenced_gpu_access = false; | |
2561 | } | |
2562 | ||
1c293ea3 | 2563 | if (obj->last_fenced_seqno) { |
199b2bc2 | 2564 | ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno); |
18991845 CW |
2565 | if (ret) |
2566 | return ret; | |
d9e86c0e CW |
2567 | |
2568 | obj->last_fenced_seqno = 0; | |
d9e86c0e CW |
2569 | } |
2570 | ||
63256ec5 CW |
2571 | /* Ensure that all CPU reads are completed before installing a fence |
2572 | * and all writes before removing the fence. | |
2573 | */ | |
2574 | if (obj->base.read_domains & I915_GEM_DOMAIN_GTT) | |
2575 | mb(); | |
2576 | ||
d9e86c0e CW |
2577 | return 0; |
2578 | } | |
2579 | ||
2580 | int | |
2581 | i915_gem_object_put_fence(struct drm_i915_gem_object *obj) | |
2582 | { | |
61050808 | 2583 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; |
d9e86c0e CW |
2584 | int ret; |
2585 | ||
a360bb1a | 2586 | ret = i915_gem_object_flush_fence(obj); |
d9e86c0e CW |
2587 | if (ret) |
2588 | return ret; | |
2589 | ||
61050808 CW |
2590 | if (obj->fence_reg == I915_FENCE_REG_NONE) |
2591 | return 0; | |
d9e86c0e | 2592 | |
61050808 CW |
2593 | i915_gem_object_update_fence(obj, |
2594 | &dev_priv->fence_regs[obj->fence_reg], | |
2595 | false); | |
2596 | i915_gem_object_fence_lost(obj); | |
d9e86c0e CW |
2597 | |
2598 | return 0; | |
2599 | } | |
2600 | ||
2601 | static struct drm_i915_fence_reg * | |
a360bb1a | 2602 | i915_find_fence_reg(struct drm_device *dev) |
ae3db24a | 2603 | { |
ae3db24a | 2604 | struct drm_i915_private *dev_priv = dev->dev_private; |
8fe301ad | 2605 | struct drm_i915_fence_reg *reg, *avail; |
d9e86c0e | 2606 | int i; |
ae3db24a DV |
2607 | |
2608 | /* First try to find a free reg */ | |
d9e86c0e | 2609 | avail = NULL; |
ae3db24a DV |
2610 | for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) { |
2611 | reg = &dev_priv->fence_regs[i]; | |
2612 | if (!reg->obj) | |
d9e86c0e | 2613 | return reg; |
ae3db24a | 2614 | |
1690e1eb | 2615 | if (!reg->pin_count) |
d9e86c0e | 2616 | avail = reg; |
ae3db24a DV |
2617 | } |
2618 | ||
d9e86c0e CW |
2619 | if (avail == NULL) |
2620 | return NULL; | |
ae3db24a DV |
2621 | |
2622 | /* None available, try to steal one or wait for a user to finish */ | |
d9e86c0e | 2623 | list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) { |
1690e1eb | 2624 | if (reg->pin_count) |
ae3db24a DV |
2625 | continue; |
2626 | ||
8fe301ad | 2627 | return reg; |
ae3db24a DV |
2628 | } |
2629 | ||
8fe301ad | 2630 | return NULL; |
ae3db24a DV |
2631 | } |
2632 | ||
de151cf6 | 2633 | /** |
9a5a53b3 | 2634 | * i915_gem_object_get_fence - set up fencing for an object |
de151cf6 JB |
2635 | * @obj: object to map through a fence reg |
2636 | * | |
2637 | * When mapping objects through the GTT, userspace wants to be able to write | |
2638 | * to them without having to worry about swizzling if the object is tiled. | |
de151cf6 JB |
2639 | * This function walks the fence regs looking for a free one for @obj, |
2640 | * stealing one if it can't find any. | |
2641 | * | |
2642 | * It then sets up the reg based on the object's properties: address, pitch | |
2643 | * and tiling format. | |
9a5a53b3 CW |
2644 | * |
2645 | * For an untiled surface, this removes any existing fence. | |
de151cf6 | 2646 | */ |
8c4b8c3f | 2647 | int |
06d98131 | 2648 | i915_gem_object_get_fence(struct drm_i915_gem_object *obj) |
de151cf6 | 2649 | { |
05394f39 | 2650 | struct drm_device *dev = obj->base.dev; |
79e53945 | 2651 | struct drm_i915_private *dev_priv = dev->dev_private; |
14415745 | 2652 | bool enable = obj->tiling_mode != I915_TILING_NONE; |
d9e86c0e | 2653 | struct drm_i915_fence_reg *reg; |
ae3db24a | 2654 | int ret; |
de151cf6 | 2655 | |
14415745 CW |
2656 | /* Have we updated the tiling parameters upon the object and so |
2657 | * will need to serialise the write to the associated fence register? | |
2658 | */ | |
5d82e3e6 | 2659 | if (obj->fence_dirty) { |
14415745 CW |
2660 | ret = i915_gem_object_flush_fence(obj); |
2661 | if (ret) | |
2662 | return ret; | |
2663 | } | |
9a5a53b3 | 2664 | |
d9e86c0e | 2665 | /* Just update our place in the LRU if our fence is getting reused. */ |
05394f39 CW |
2666 | if (obj->fence_reg != I915_FENCE_REG_NONE) { |
2667 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
5d82e3e6 | 2668 | if (!obj->fence_dirty) { |
14415745 CW |
2669 | list_move_tail(®->lru_list, |
2670 | &dev_priv->mm.fence_list); | |
2671 | return 0; | |
2672 | } | |
2673 | } else if (enable) { | |
2674 | reg = i915_find_fence_reg(dev); | |
2675 | if (reg == NULL) | |
2676 | return -EDEADLK; | |
d9e86c0e | 2677 | |
14415745 CW |
2678 | if (reg->obj) { |
2679 | struct drm_i915_gem_object *old = reg->obj; | |
2680 | ||
2681 | ret = i915_gem_object_flush_fence(old); | |
29c5a587 CW |
2682 | if (ret) |
2683 | return ret; | |
2684 | ||
14415745 | 2685 | i915_gem_object_fence_lost(old); |
29c5a587 | 2686 | } |
14415745 | 2687 | } else |
a09ba7fa | 2688 | return 0; |
a09ba7fa | 2689 | |
14415745 | 2690 | i915_gem_object_update_fence(obj, reg, enable); |
5d82e3e6 | 2691 | obj->fence_dirty = false; |
14415745 | 2692 | |
9ce079e4 | 2693 | return 0; |
de151cf6 JB |
2694 | } |
2695 | ||
673a394b EA |
2696 | /** |
2697 | * Finds free space in the GTT aperture and binds the object there. | |
2698 | */ | |
2699 | static int | |
05394f39 | 2700 | i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj, |
920afa77 | 2701 | unsigned alignment, |
75e9e915 | 2702 | bool map_and_fenceable) |
673a394b | 2703 | { |
05394f39 | 2704 | struct drm_device *dev = obj->base.dev; |
673a394b | 2705 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b | 2706 | struct drm_mm_node *free_space; |
a00b10c3 | 2707 | gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN; |
5e783301 | 2708 | u32 size, fence_size, fence_alignment, unfenced_alignment; |
75e9e915 | 2709 | bool mappable, fenceable; |
07f73f69 | 2710 | int ret; |
673a394b | 2711 | |
05394f39 | 2712 | if (obj->madv != I915_MADV_WILLNEED) { |
3ef94daa CW |
2713 | DRM_ERROR("Attempting to bind a purgeable object\n"); |
2714 | return -EINVAL; | |
2715 | } | |
2716 | ||
e28f8711 CW |
2717 | fence_size = i915_gem_get_gtt_size(dev, |
2718 | obj->base.size, | |
2719 | obj->tiling_mode); | |
2720 | fence_alignment = i915_gem_get_gtt_alignment(dev, | |
2721 | obj->base.size, | |
2722 | obj->tiling_mode); | |
2723 | unfenced_alignment = | |
2724 | i915_gem_get_unfenced_gtt_alignment(dev, | |
2725 | obj->base.size, | |
2726 | obj->tiling_mode); | |
a00b10c3 | 2727 | |
673a394b | 2728 | if (alignment == 0) |
5e783301 DV |
2729 | alignment = map_and_fenceable ? fence_alignment : |
2730 | unfenced_alignment; | |
75e9e915 | 2731 | if (map_and_fenceable && alignment & (fence_alignment - 1)) { |
673a394b EA |
2732 | DRM_ERROR("Invalid object alignment requested %u\n", alignment); |
2733 | return -EINVAL; | |
2734 | } | |
2735 | ||
05394f39 | 2736 | size = map_and_fenceable ? fence_size : obj->base.size; |
a00b10c3 | 2737 | |
654fc607 CW |
2738 | /* If the object is bigger than the entire aperture, reject it early |
2739 | * before evicting everything in a vain attempt to find space. | |
2740 | */ | |
05394f39 | 2741 | if (obj->base.size > |
75e9e915 | 2742 | (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) { |
654fc607 CW |
2743 | DRM_ERROR("Attempting to bind an object larger than the aperture\n"); |
2744 | return -E2BIG; | |
2745 | } | |
2746 | ||
673a394b | 2747 | search_free: |
75e9e915 | 2748 | if (map_and_fenceable) |
920afa77 DV |
2749 | free_space = |
2750 | drm_mm_search_free_in_range(&dev_priv->mm.gtt_space, | |
6b9d89b4 CW |
2751 | size, alignment, |
2752 | 0, dev_priv->mm.gtt_mappable_end, | |
920afa77 DV |
2753 | 0); |
2754 | else | |
2755 | free_space = drm_mm_search_free(&dev_priv->mm.gtt_space, | |
a00b10c3 | 2756 | size, alignment, 0); |
920afa77 DV |
2757 | |
2758 | if (free_space != NULL) { | |
75e9e915 | 2759 | if (map_and_fenceable) |
05394f39 | 2760 | obj->gtt_space = |
920afa77 | 2761 | drm_mm_get_block_range_generic(free_space, |
a00b10c3 | 2762 | size, alignment, 0, |
6b9d89b4 | 2763 | 0, dev_priv->mm.gtt_mappable_end, |
920afa77 DV |
2764 | 0); |
2765 | else | |
05394f39 | 2766 | obj->gtt_space = |
a00b10c3 | 2767 | drm_mm_get_block(free_space, size, alignment); |
920afa77 | 2768 | } |
05394f39 | 2769 | if (obj->gtt_space == NULL) { |
673a394b EA |
2770 | /* If the gtt is empty and we're still having trouble |
2771 | * fitting our object in, we're out of memory. | |
2772 | */ | |
75e9e915 DV |
2773 | ret = i915_gem_evict_something(dev, size, alignment, |
2774 | map_and_fenceable); | |
9731129c | 2775 | if (ret) |
673a394b | 2776 | return ret; |
9731129c | 2777 | |
673a394b EA |
2778 | goto search_free; |
2779 | } | |
2780 | ||
e5281ccd | 2781 | ret = i915_gem_object_get_pages_gtt(obj, gfpmask); |
673a394b | 2782 | if (ret) { |
05394f39 CW |
2783 | drm_mm_put_block(obj->gtt_space); |
2784 | obj->gtt_space = NULL; | |
07f73f69 CW |
2785 | |
2786 | if (ret == -ENOMEM) { | |
809b6334 CW |
2787 | /* first try to reclaim some memory by clearing the GTT */ |
2788 | ret = i915_gem_evict_everything(dev, false); | |
07f73f69 | 2789 | if (ret) { |
07f73f69 | 2790 | /* now try to shrink everyone else */ |
4bdadb97 CW |
2791 | if (gfpmask) { |
2792 | gfpmask = 0; | |
2793 | goto search_free; | |
07f73f69 CW |
2794 | } |
2795 | ||
809b6334 | 2796 | return -ENOMEM; |
07f73f69 CW |
2797 | } |
2798 | ||
2799 | goto search_free; | |
2800 | } | |
2801 | ||
673a394b EA |
2802 | return ret; |
2803 | } | |
2804 | ||
74163907 | 2805 | ret = i915_gem_gtt_prepare_object(obj); |
7c2e6fdf | 2806 | if (ret) { |
e5281ccd | 2807 | i915_gem_object_put_pages_gtt(obj); |
05394f39 CW |
2808 | drm_mm_put_block(obj->gtt_space); |
2809 | obj->gtt_space = NULL; | |
07f73f69 | 2810 | |
809b6334 | 2811 | if (i915_gem_evict_everything(dev, false)) |
07f73f69 | 2812 | return ret; |
07f73f69 CW |
2813 | |
2814 | goto search_free; | |
673a394b | 2815 | } |
673a394b | 2816 | |
0ebb9829 DV |
2817 | if (!dev_priv->mm.aliasing_ppgtt) |
2818 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
673a394b | 2819 | |
6299f992 | 2820 | list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list); |
05394f39 | 2821 | list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list); |
bf1a1092 | 2822 | |
673a394b EA |
2823 | /* Assert that the object is not currently in any GPU domain. As it |
2824 | * wasn't in the GTT, there shouldn't be any way it could have been in | |
2825 | * a GPU cache | |
2826 | */ | |
05394f39 CW |
2827 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); |
2828 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); | |
673a394b | 2829 | |
6299f992 | 2830 | obj->gtt_offset = obj->gtt_space->start; |
1c5d22f7 | 2831 | |
75e9e915 | 2832 | fenceable = |
05394f39 | 2833 | obj->gtt_space->size == fence_size && |
0206e353 | 2834 | (obj->gtt_space->start & (fence_alignment - 1)) == 0; |
a00b10c3 | 2835 | |
75e9e915 | 2836 | mappable = |
05394f39 | 2837 | obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end; |
a00b10c3 | 2838 | |
05394f39 | 2839 | obj->map_and_fenceable = mappable && fenceable; |
75e9e915 | 2840 | |
db53a302 | 2841 | trace_i915_gem_object_bind(obj, map_and_fenceable); |
673a394b EA |
2842 | return 0; |
2843 | } | |
2844 | ||
2845 | void | |
05394f39 | 2846 | i915_gem_clflush_object(struct drm_i915_gem_object *obj) |
673a394b | 2847 | { |
673a394b EA |
2848 | /* If we don't have a page list set up, then we're not pinned |
2849 | * to GPU, and we can ignore the cache flush because it'll happen | |
2850 | * again at bind time. | |
2851 | */ | |
05394f39 | 2852 | if (obj->pages == NULL) |
673a394b EA |
2853 | return; |
2854 | ||
9c23f7fc CW |
2855 | /* If the GPU is snooping the contents of the CPU cache, |
2856 | * we do not need to manually clear the CPU cache lines. However, | |
2857 | * the caches are only snooped when the render cache is | |
2858 | * flushed/invalidated. As we always have to emit invalidations | |
2859 | * and flushes when moving into and out of the RENDER domain, correct | |
2860 | * snooping behaviour occurs naturally as the result of our domain | |
2861 | * tracking. | |
2862 | */ | |
2863 | if (obj->cache_level != I915_CACHE_NONE) | |
2864 | return; | |
2865 | ||
1c5d22f7 | 2866 | trace_i915_gem_object_clflush(obj); |
cfa16a0d | 2867 | |
05394f39 | 2868 | drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE); |
673a394b EA |
2869 | } |
2870 | ||
e47c68e9 | 2871 | /** Flushes any GPU write domain for the object if it's dirty. */ |
88241785 | 2872 | static int |
3619df03 | 2873 | i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 2874 | { |
05394f39 | 2875 | if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0) |
88241785 | 2876 | return 0; |
e47c68e9 EA |
2877 | |
2878 | /* Queue the GPU write cache flushing we need. */ | |
db53a302 | 2879 | return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain); |
e47c68e9 EA |
2880 | } |
2881 | ||
2882 | /** Flushes the GTT write domain for the object if it's dirty. */ | |
2883 | static void | |
05394f39 | 2884 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 2885 | { |
1c5d22f7 CW |
2886 | uint32_t old_write_domain; |
2887 | ||
05394f39 | 2888 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
e47c68e9 EA |
2889 | return; |
2890 | ||
63256ec5 | 2891 | /* No actual flushing is required for the GTT write domain. Writes |
e47c68e9 EA |
2892 | * to it immediately go to main memory as far as we know, so there's |
2893 | * no chipset flush. It also doesn't land in render cache. | |
63256ec5 CW |
2894 | * |
2895 | * However, we do have to enforce the order so that all writes through | |
2896 | * the GTT land before any writes to the device, such as updates to | |
2897 | * the GATT itself. | |
e47c68e9 | 2898 | */ |
63256ec5 CW |
2899 | wmb(); |
2900 | ||
05394f39 CW |
2901 | old_write_domain = obj->base.write_domain; |
2902 | obj->base.write_domain = 0; | |
1c5d22f7 CW |
2903 | |
2904 | trace_i915_gem_object_change_domain(obj, | |
05394f39 | 2905 | obj->base.read_domains, |
1c5d22f7 | 2906 | old_write_domain); |
e47c68e9 EA |
2907 | } |
2908 | ||
2909 | /** Flushes the CPU write domain for the object if it's dirty. */ | |
2910 | static void | |
05394f39 | 2911 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 2912 | { |
1c5d22f7 | 2913 | uint32_t old_write_domain; |
e47c68e9 | 2914 | |
05394f39 | 2915 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
e47c68e9 EA |
2916 | return; |
2917 | ||
2918 | i915_gem_clflush_object(obj); | |
40ce6575 | 2919 | intel_gtt_chipset_flush(); |
05394f39 CW |
2920 | old_write_domain = obj->base.write_domain; |
2921 | obj->base.write_domain = 0; | |
1c5d22f7 CW |
2922 | |
2923 | trace_i915_gem_object_change_domain(obj, | |
05394f39 | 2924 | obj->base.read_domains, |
1c5d22f7 | 2925 | old_write_domain); |
e47c68e9 EA |
2926 | } |
2927 | ||
2ef7eeaa EA |
2928 | /** |
2929 | * Moves a single object to the GTT read, and possibly write domain. | |
2930 | * | |
2931 | * This function returns when the move is complete, including waiting on | |
2932 | * flushes to occur. | |
2933 | */ | |
79e53945 | 2934 | int |
2021746e | 2935 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
2ef7eeaa | 2936 | { |
8325a09d | 2937 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
1c5d22f7 | 2938 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 | 2939 | int ret; |
2ef7eeaa | 2940 | |
02354392 | 2941 | /* Not valid to be called on unbound objects. */ |
05394f39 | 2942 | if (obj->gtt_space == NULL) |
02354392 EA |
2943 | return -EINVAL; |
2944 | ||
8d7e3de1 CW |
2945 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
2946 | return 0; | |
2947 | ||
88241785 CW |
2948 | ret = i915_gem_object_flush_gpu_write_domain(obj); |
2949 | if (ret) | |
2950 | return ret; | |
2951 | ||
87ca9c8a | 2952 | if (obj->pending_gpu_write || write) { |
ce453d81 | 2953 | ret = i915_gem_object_wait_rendering(obj); |
87ca9c8a CW |
2954 | if (ret) |
2955 | return ret; | |
2956 | } | |
2dafb1e0 | 2957 | |
7213342d | 2958 | i915_gem_object_flush_cpu_write_domain(obj); |
1c5d22f7 | 2959 | |
05394f39 CW |
2960 | old_write_domain = obj->base.write_domain; |
2961 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 2962 | |
e47c68e9 EA |
2963 | /* It should now be out of any other write domains, and we can update |
2964 | * the domain values for our changes. | |
2965 | */ | |
05394f39 CW |
2966 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
2967 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; | |
e47c68e9 | 2968 | if (write) { |
05394f39 CW |
2969 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
2970 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; | |
2971 | obj->dirty = 1; | |
2ef7eeaa EA |
2972 | } |
2973 | ||
1c5d22f7 CW |
2974 | trace_i915_gem_object_change_domain(obj, |
2975 | old_read_domains, | |
2976 | old_write_domain); | |
2977 | ||
8325a09d CW |
2978 | /* And bump the LRU for this access */ |
2979 | if (i915_gem_object_is_inactive(obj)) | |
2980 | list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list); | |
2981 | ||
e47c68e9 EA |
2982 | return 0; |
2983 | } | |
2984 | ||
e4ffd173 CW |
2985 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
2986 | enum i915_cache_level cache_level) | |
2987 | { | |
7bddb01f DV |
2988 | struct drm_device *dev = obj->base.dev; |
2989 | drm_i915_private_t *dev_priv = dev->dev_private; | |
e4ffd173 CW |
2990 | int ret; |
2991 | ||
2992 | if (obj->cache_level == cache_level) | |
2993 | return 0; | |
2994 | ||
2995 | if (obj->pin_count) { | |
2996 | DRM_DEBUG("can not change the cache level of pinned objects\n"); | |
2997 | return -EBUSY; | |
2998 | } | |
2999 | ||
3000 | if (obj->gtt_space) { | |
3001 | ret = i915_gem_object_finish_gpu(obj); | |
3002 | if (ret) | |
3003 | return ret; | |
3004 | ||
3005 | i915_gem_object_finish_gtt(obj); | |
3006 | ||
3007 | /* Before SandyBridge, you could not use tiling or fence | |
3008 | * registers with snooped memory, so relinquish any fences | |
3009 | * currently pointing to our region in the aperture. | |
3010 | */ | |
3011 | if (INTEL_INFO(obj->base.dev)->gen < 6) { | |
3012 | ret = i915_gem_object_put_fence(obj); | |
3013 | if (ret) | |
3014 | return ret; | |
3015 | } | |
3016 | ||
74898d7e DV |
3017 | if (obj->has_global_gtt_mapping) |
3018 | i915_gem_gtt_bind_object(obj, cache_level); | |
7bddb01f DV |
3019 | if (obj->has_aliasing_ppgtt_mapping) |
3020 | i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt, | |
3021 | obj, cache_level); | |
e4ffd173 CW |
3022 | } |
3023 | ||
3024 | if (cache_level == I915_CACHE_NONE) { | |
3025 | u32 old_read_domains, old_write_domain; | |
3026 | ||
3027 | /* If we're coming from LLC cached, then we haven't | |
3028 | * actually been tracking whether the data is in the | |
3029 | * CPU cache or not, since we only allow one bit set | |
3030 | * in obj->write_domain and have been skipping the clflushes. | |
3031 | * Just set it to the CPU cache for now. | |
3032 | */ | |
3033 | WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU); | |
3034 | WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU); | |
3035 | ||
3036 | old_read_domains = obj->base.read_domains; | |
3037 | old_write_domain = obj->base.write_domain; | |
3038 | ||
3039 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
3040 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
3041 | ||
3042 | trace_i915_gem_object_change_domain(obj, | |
3043 | old_read_domains, | |
3044 | old_write_domain); | |
3045 | } | |
3046 | ||
3047 | obj->cache_level = cache_level; | |
3048 | return 0; | |
3049 | } | |
3050 | ||
b9241ea3 | 3051 | /* |
2da3b9b9 CW |
3052 | * Prepare buffer for display plane (scanout, cursors, etc). |
3053 | * Can be called from an uninterruptible phase (modesetting) and allows | |
3054 | * any flushes to be pipelined (for pageflips). | |
b9241ea3 ZW |
3055 | */ |
3056 | int | |
2da3b9b9 CW |
3057 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
3058 | u32 alignment, | |
919926ae | 3059 | struct intel_ring_buffer *pipelined) |
b9241ea3 | 3060 | { |
2da3b9b9 | 3061 | u32 old_read_domains, old_write_domain; |
b9241ea3 ZW |
3062 | int ret; |
3063 | ||
88241785 CW |
3064 | ret = i915_gem_object_flush_gpu_write_domain(obj); |
3065 | if (ret) | |
3066 | return ret; | |
3067 | ||
0be73284 | 3068 | if (pipelined != obj->ring) { |
2911a35b BW |
3069 | ret = i915_gem_object_sync(obj, pipelined); |
3070 | if (ret) | |
b9241ea3 ZW |
3071 | return ret; |
3072 | } | |
3073 | ||
a7ef0640 EA |
3074 | /* The display engine is not coherent with the LLC cache on gen6. As |
3075 | * a result, we make sure that the pinning that is about to occur is | |
3076 | * done with uncached PTEs. This is lowest common denominator for all | |
3077 | * chipsets. | |
3078 | * | |
3079 | * However for gen6+, we could do better by using the GFDT bit instead | |
3080 | * of uncaching, which would allow us to flush all the LLC-cached data | |
3081 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. | |
3082 | */ | |
3083 | ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE); | |
3084 | if (ret) | |
3085 | return ret; | |
3086 | ||
2da3b9b9 CW |
3087 | /* As the user may map the buffer once pinned in the display plane |
3088 | * (e.g. libkms for the bootup splash), we have to ensure that we | |
3089 | * always use map_and_fenceable for all scanout buffers. | |
3090 | */ | |
3091 | ret = i915_gem_object_pin(obj, alignment, true); | |
3092 | if (ret) | |
3093 | return ret; | |
3094 | ||
b118c1e3 CW |
3095 | i915_gem_object_flush_cpu_write_domain(obj); |
3096 | ||
2da3b9b9 | 3097 | old_write_domain = obj->base.write_domain; |
05394f39 | 3098 | old_read_domains = obj->base.read_domains; |
2da3b9b9 CW |
3099 | |
3100 | /* It should now be out of any other write domains, and we can update | |
3101 | * the domain values for our changes. | |
3102 | */ | |
3103 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); | |
05394f39 | 3104 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
b9241ea3 ZW |
3105 | |
3106 | trace_i915_gem_object_change_domain(obj, | |
3107 | old_read_domains, | |
2da3b9b9 | 3108 | old_write_domain); |
b9241ea3 ZW |
3109 | |
3110 | return 0; | |
3111 | } | |
3112 | ||
85345517 | 3113 | int |
a8198eea | 3114 | i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj) |
85345517 | 3115 | { |
88241785 CW |
3116 | int ret; |
3117 | ||
a8198eea | 3118 | if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0) |
85345517 CW |
3119 | return 0; |
3120 | ||
88241785 | 3121 | if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) { |
db53a302 | 3122 | ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain); |
88241785 CW |
3123 | if (ret) |
3124 | return ret; | |
3125 | } | |
85345517 | 3126 | |
c501ae7f CW |
3127 | ret = i915_gem_object_wait_rendering(obj); |
3128 | if (ret) | |
3129 | return ret; | |
3130 | ||
a8198eea CW |
3131 | /* Ensure that we invalidate the GPU's caches and TLBs. */ |
3132 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; | |
c501ae7f | 3133 | return 0; |
85345517 CW |
3134 | } |
3135 | ||
e47c68e9 EA |
3136 | /** |
3137 | * Moves a single object to the CPU read, and possibly write domain. | |
3138 | * | |
3139 | * This function returns when the move is complete, including waiting on | |
3140 | * flushes to occur. | |
3141 | */ | |
dabdfe02 | 3142 | int |
919926ae | 3143 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
e47c68e9 | 3144 | { |
1c5d22f7 | 3145 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 EA |
3146 | int ret; |
3147 | ||
8d7e3de1 CW |
3148 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3149 | return 0; | |
3150 | ||
88241785 CW |
3151 | ret = i915_gem_object_flush_gpu_write_domain(obj); |
3152 | if (ret) | |
3153 | return ret; | |
3154 | ||
f8413190 CW |
3155 | if (write || obj->pending_gpu_write) { |
3156 | ret = i915_gem_object_wait_rendering(obj); | |
3157 | if (ret) | |
3158 | return ret; | |
3159 | } | |
2ef7eeaa | 3160 | |
e47c68e9 | 3161 | i915_gem_object_flush_gtt_write_domain(obj); |
2ef7eeaa | 3162 | |
05394f39 CW |
3163 | old_write_domain = obj->base.write_domain; |
3164 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 3165 | |
e47c68e9 | 3166 | /* Flush the CPU cache if it's still invalid. */ |
05394f39 | 3167 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
2ef7eeaa | 3168 | i915_gem_clflush_object(obj); |
2ef7eeaa | 3169 | |
05394f39 | 3170 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
2ef7eeaa EA |
3171 | } |
3172 | ||
3173 | /* It should now be out of any other write domains, and we can update | |
3174 | * the domain values for our changes. | |
3175 | */ | |
05394f39 | 3176 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
e47c68e9 EA |
3177 | |
3178 | /* If we're writing through the CPU, then the GPU read domains will | |
3179 | * need to be invalidated at next use. | |
3180 | */ | |
3181 | if (write) { | |
05394f39 CW |
3182 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3183 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
e47c68e9 | 3184 | } |
2ef7eeaa | 3185 | |
1c5d22f7 CW |
3186 | trace_i915_gem_object_change_domain(obj, |
3187 | old_read_domains, | |
3188 | old_write_domain); | |
3189 | ||
2ef7eeaa EA |
3190 | return 0; |
3191 | } | |
3192 | ||
673a394b EA |
3193 | /* Throttle our rendering by waiting until the ring has completed our requests |
3194 | * emitted over 20 msec ago. | |
3195 | * | |
b962442e EA |
3196 | * Note that if we were to use the current jiffies each time around the loop, |
3197 | * we wouldn't escape the function with any frames outstanding if the time to | |
3198 | * render a frame was over 20ms. | |
3199 | * | |
673a394b EA |
3200 | * This should get us reasonable parallelism between CPU and GPU but also |
3201 | * relatively low latency when blocking on a particular request to finish. | |
3202 | */ | |
40a5f0de | 3203 | static int |
f787a5f5 | 3204 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
40a5f0de | 3205 | { |
f787a5f5 CW |
3206 | struct drm_i915_private *dev_priv = dev->dev_private; |
3207 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
b962442e | 3208 | unsigned long recent_enough = jiffies - msecs_to_jiffies(20); |
f787a5f5 CW |
3209 | struct drm_i915_gem_request *request; |
3210 | struct intel_ring_buffer *ring = NULL; | |
3211 | u32 seqno = 0; | |
3212 | int ret; | |
93533c29 | 3213 | |
e110e8d6 CW |
3214 | if (atomic_read(&dev_priv->mm.wedged)) |
3215 | return -EIO; | |
3216 | ||
1c25595f | 3217 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 3218 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { |
b962442e EA |
3219 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
3220 | break; | |
40a5f0de | 3221 | |
f787a5f5 CW |
3222 | ring = request->ring; |
3223 | seqno = request->seqno; | |
b962442e | 3224 | } |
1c25595f | 3225 | spin_unlock(&file_priv->mm.lock); |
40a5f0de | 3226 | |
f787a5f5 CW |
3227 | if (seqno == 0) |
3228 | return 0; | |
2bc43b5c | 3229 | |
5c81fe85 | 3230 | ret = __wait_seqno(ring, seqno, true, NULL); |
f787a5f5 CW |
3231 | if (ret == 0) |
3232 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); | |
40a5f0de EA |
3233 | |
3234 | return ret; | |
3235 | } | |
3236 | ||
673a394b | 3237 | int |
05394f39 CW |
3238 | i915_gem_object_pin(struct drm_i915_gem_object *obj, |
3239 | uint32_t alignment, | |
75e9e915 | 3240 | bool map_and_fenceable) |
673a394b | 3241 | { |
673a394b EA |
3242 | int ret; |
3243 | ||
7e81a42e CW |
3244 | if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT)) |
3245 | return -EBUSY; | |
ac0c6b5a | 3246 | |
05394f39 CW |
3247 | if (obj->gtt_space != NULL) { |
3248 | if ((alignment && obj->gtt_offset & (alignment - 1)) || | |
3249 | (map_and_fenceable && !obj->map_and_fenceable)) { | |
3250 | WARN(obj->pin_count, | |
ae7d49d8 | 3251 | "bo is already pinned with incorrect alignment:" |
75e9e915 DV |
3252 | " offset=%x, req.alignment=%x, req.map_and_fenceable=%d," |
3253 | " obj->map_and_fenceable=%d\n", | |
05394f39 | 3254 | obj->gtt_offset, alignment, |
75e9e915 | 3255 | map_and_fenceable, |
05394f39 | 3256 | obj->map_and_fenceable); |
ac0c6b5a CW |
3257 | ret = i915_gem_object_unbind(obj); |
3258 | if (ret) | |
3259 | return ret; | |
3260 | } | |
3261 | } | |
3262 | ||
05394f39 | 3263 | if (obj->gtt_space == NULL) { |
a00b10c3 | 3264 | ret = i915_gem_object_bind_to_gtt(obj, alignment, |
75e9e915 | 3265 | map_and_fenceable); |
9731129c | 3266 | if (ret) |
673a394b | 3267 | return ret; |
22c344e9 | 3268 | } |
76446cac | 3269 | |
74898d7e DV |
3270 | if (!obj->has_global_gtt_mapping && map_and_fenceable) |
3271 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
3272 | ||
1b50247a | 3273 | obj->pin_count++; |
6299f992 | 3274 | obj->pin_mappable |= map_and_fenceable; |
673a394b EA |
3275 | |
3276 | return 0; | |
3277 | } | |
3278 | ||
3279 | void | |
05394f39 | 3280 | i915_gem_object_unpin(struct drm_i915_gem_object *obj) |
673a394b | 3281 | { |
05394f39 CW |
3282 | BUG_ON(obj->pin_count == 0); |
3283 | BUG_ON(obj->gtt_space == NULL); | |
673a394b | 3284 | |
1b50247a | 3285 | if (--obj->pin_count == 0) |
6299f992 | 3286 | obj->pin_mappable = false; |
673a394b EA |
3287 | } |
3288 | ||
3289 | int | |
3290 | i915_gem_pin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3291 | struct drm_file *file) |
673a394b EA |
3292 | { |
3293 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 3294 | struct drm_i915_gem_object *obj; |
673a394b EA |
3295 | int ret; |
3296 | ||
1d7cfea1 CW |
3297 | ret = i915_mutex_lock_interruptible(dev); |
3298 | if (ret) | |
3299 | return ret; | |
673a394b | 3300 | |
05394f39 | 3301 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3302 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3303 | ret = -ENOENT; |
3304 | goto unlock; | |
673a394b | 3305 | } |
673a394b | 3306 | |
05394f39 | 3307 | if (obj->madv != I915_MADV_WILLNEED) { |
bb6baf76 | 3308 | DRM_ERROR("Attempting to pin a purgeable buffer\n"); |
1d7cfea1 CW |
3309 | ret = -EINVAL; |
3310 | goto out; | |
3ef94daa CW |
3311 | } |
3312 | ||
05394f39 | 3313 | if (obj->pin_filp != NULL && obj->pin_filp != file) { |
79e53945 JB |
3314 | DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n", |
3315 | args->handle); | |
1d7cfea1 CW |
3316 | ret = -EINVAL; |
3317 | goto out; | |
79e53945 JB |
3318 | } |
3319 | ||
05394f39 CW |
3320 | obj->user_pin_count++; |
3321 | obj->pin_filp = file; | |
3322 | if (obj->user_pin_count == 1) { | |
75e9e915 | 3323 | ret = i915_gem_object_pin(obj, args->alignment, true); |
1d7cfea1 CW |
3324 | if (ret) |
3325 | goto out; | |
673a394b EA |
3326 | } |
3327 | ||
3328 | /* XXX - flush the CPU caches for pinned objects | |
3329 | * as the X server doesn't manage domains yet | |
3330 | */ | |
e47c68e9 | 3331 | i915_gem_object_flush_cpu_write_domain(obj); |
05394f39 | 3332 | args->offset = obj->gtt_offset; |
1d7cfea1 | 3333 | out: |
05394f39 | 3334 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3335 | unlock: |
673a394b | 3336 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3337 | return ret; |
673a394b EA |
3338 | } |
3339 | ||
3340 | int | |
3341 | i915_gem_unpin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3342 | struct drm_file *file) |
673a394b EA |
3343 | { |
3344 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 3345 | struct drm_i915_gem_object *obj; |
76c1dec1 | 3346 | int ret; |
673a394b | 3347 | |
1d7cfea1 CW |
3348 | ret = i915_mutex_lock_interruptible(dev); |
3349 | if (ret) | |
3350 | return ret; | |
673a394b | 3351 | |
05394f39 | 3352 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3353 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3354 | ret = -ENOENT; |
3355 | goto unlock; | |
673a394b | 3356 | } |
76c1dec1 | 3357 | |
05394f39 | 3358 | if (obj->pin_filp != file) { |
79e53945 JB |
3359 | DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n", |
3360 | args->handle); | |
1d7cfea1 CW |
3361 | ret = -EINVAL; |
3362 | goto out; | |
79e53945 | 3363 | } |
05394f39 CW |
3364 | obj->user_pin_count--; |
3365 | if (obj->user_pin_count == 0) { | |
3366 | obj->pin_filp = NULL; | |
79e53945 JB |
3367 | i915_gem_object_unpin(obj); |
3368 | } | |
673a394b | 3369 | |
1d7cfea1 | 3370 | out: |
05394f39 | 3371 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3372 | unlock: |
673a394b | 3373 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3374 | return ret; |
673a394b EA |
3375 | } |
3376 | ||
3377 | int | |
3378 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3379 | struct drm_file *file) |
673a394b EA |
3380 | { |
3381 | struct drm_i915_gem_busy *args = data; | |
05394f39 | 3382 | struct drm_i915_gem_object *obj; |
30dbf0c0 CW |
3383 | int ret; |
3384 | ||
76c1dec1 | 3385 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 3386 | if (ret) |
76c1dec1 | 3387 | return ret; |
673a394b | 3388 | |
05394f39 | 3389 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3390 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3391 | ret = -ENOENT; |
3392 | goto unlock; | |
673a394b | 3393 | } |
d1b851fc | 3394 | |
0be555b6 CW |
3395 | /* Count all active objects as busy, even if they are currently not used |
3396 | * by the gpu. Users of this interface expect objects to eventually | |
3397 | * become non-busy without any further actions, therefore emit any | |
3398 | * necessary flushes here. | |
c4de0a5d | 3399 | */ |
30dfebf3 | 3400 | ret = i915_gem_object_flush_active(obj); |
0be555b6 | 3401 | |
30dfebf3 | 3402 | args->busy = obj->active; |
673a394b | 3403 | |
05394f39 | 3404 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3405 | unlock: |
673a394b | 3406 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3407 | return ret; |
673a394b EA |
3408 | } |
3409 | ||
3410 | int | |
3411 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, | |
3412 | struct drm_file *file_priv) | |
3413 | { | |
0206e353 | 3414 | return i915_gem_ring_throttle(dev, file_priv); |
673a394b EA |
3415 | } |
3416 | ||
3ef94daa CW |
3417 | int |
3418 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, | |
3419 | struct drm_file *file_priv) | |
3420 | { | |
3421 | struct drm_i915_gem_madvise *args = data; | |
05394f39 | 3422 | struct drm_i915_gem_object *obj; |
76c1dec1 | 3423 | int ret; |
3ef94daa CW |
3424 | |
3425 | switch (args->madv) { | |
3426 | case I915_MADV_DONTNEED: | |
3427 | case I915_MADV_WILLNEED: | |
3428 | break; | |
3429 | default: | |
3430 | return -EINVAL; | |
3431 | } | |
3432 | ||
1d7cfea1 CW |
3433 | ret = i915_mutex_lock_interruptible(dev); |
3434 | if (ret) | |
3435 | return ret; | |
3436 | ||
05394f39 | 3437 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); |
c8725226 | 3438 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3439 | ret = -ENOENT; |
3440 | goto unlock; | |
3ef94daa | 3441 | } |
3ef94daa | 3442 | |
05394f39 | 3443 | if (obj->pin_count) { |
1d7cfea1 CW |
3444 | ret = -EINVAL; |
3445 | goto out; | |
3ef94daa CW |
3446 | } |
3447 | ||
05394f39 CW |
3448 | if (obj->madv != __I915_MADV_PURGED) |
3449 | obj->madv = args->madv; | |
3ef94daa | 3450 | |
2d7ef395 | 3451 | /* if the object is no longer bound, discard its backing storage */ |
05394f39 CW |
3452 | if (i915_gem_object_is_purgeable(obj) && |
3453 | obj->gtt_space == NULL) | |
2d7ef395 CW |
3454 | i915_gem_object_truncate(obj); |
3455 | ||
05394f39 | 3456 | args->retained = obj->madv != __I915_MADV_PURGED; |
bb6baf76 | 3457 | |
1d7cfea1 | 3458 | out: |
05394f39 | 3459 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3460 | unlock: |
3ef94daa | 3461 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3462 | return ret; |
3ef94daa CW |
3463 | } |
3464 | ||
05394f39 CW |
3465 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, |
3466 | size_t size) | |
ac52bc56 | 3467 | { |
73aa808f | 3468 | struct drm_i915_private *dev_priv = dev->dev_private; |
c397b908 | 3469 | struct drm_i915_gem_object *obj; |
5949eac4 | 3470 | struct address_space *mapping; |
bed1ea95 | 3471 | u32 mask; |
ac52bc56 | 3472 | |
c397b908 DV |
3473 | obj = kzalloc(sizeof(*obj), GFP_KERNEL); |
3474 | if (obj == NULL) | |
3475 | return NULL; | |
673a394b | 3476 | |
c397b908 DV |
3477 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { |
3478 | kfree(obj); | |
3479 | return NULL; | |
3480 | } | |
673a394b | 3481 | |
bed1ea95 CW |
3482 | mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; |
3483 | if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) { | |
3484 | /* 965gm cannot relocate objects above 4GiB. */ | |
3485 | mask &= ~__GFP_HIGHMEM; | |
3486 | mask |= __GFP_DMA32; | |
3487 | } | |
3488 | ||
5949eac4 | 3489 | mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
bed1ea95 | 3490 | mapping_set_gfp_mask(mapping, mask); |
5949eac4 | 3491 | |
73aa808f CW |
3492 | i915_gem_info_add_obj(dev_priv, size); |
3493 | ||
c397b908 DV |
3494 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
3495 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
673a394b | 3496 | |
3d29b842 ED |
3497 | if (HAS_LLC(dev)) { |
3498 | /* On some devices, we can have the GPU use the LLC (the CPU | |
a1871112 EA |
3499 | * cache) for about a 10% performance improvement |
3500 | * compared to uncached. Graphics requests other than | |
3501 | * display scanout are coherent with the CPU in | |
3502 | * accessing this cache. This means in this mode we | |
3503 | * don't need to clflush on the CPU side, and on the | |
3504 | * GPU side we only need to flush internal caches to | |
3505 | * get data visible to the CPU. | |
3506 | * | |
3507 | * However, we maintain the display planes as UC, and so | |
3508 | * need to rebind when first used as such. | |
3509 | */ | |
3510 | obj->cache_level = I915_CACHE_LLC; | |
3511 | } else | |
3512 | obj->cache_level = I915_CACHE_NONE; | |
3513 | ||
62b8b215 | 3514 | obj->base.driver_private = NULL; |
c397b908 | 3515 | obj->fence_reg = I915_FENCE_REG_NONE; |
69dc4987 | 3516 | INIT_LIST_HEAD(&obj->mm_list); |
93a37f20 | 3517 | INIT_LIST_HEAD(&obj->gtt_list); |
69dc4987 | 3518 | INIT_LIST_HEAD(&obj->ring_list); |
432e58ed | 3519 | INIT_LIST_HEAD(&obj->exec_list); |
c397b908 | 3520 | INIT_LIST_HEAD(&obj->gpu_write_list); |
c397b908 | 3521 | obj->madv = I915_MADV_WILLNEED; |
75e9e915 DV |
3522 | /* Avoid an unnecessary call to unbind on the first bind. */ |
3523 | obj->map_and_fenceable = true; | |
de151cf6 | 3524 | |
05394f39 | 3525 | return obj; |
c397b908 DV |
3526 | } |
3527 | ||
3528 | int i915_gem_init_object(struct drm_gem_object *obj) | |
3529 | { | |
3530 | BUG(); | |
de151cf6 | 3531 | |
673a394b EA |
3532 | return 0; |
3533 | } | |
3534 | ||
1488fc08 | 3535 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
673a394b | 3536 | { |
1488fc08 | 3537 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); |
05394f39 | 3538 | struct drm_device *dev = obj->base.dev; |
be72615b | 3539 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b | 3540 | |
26e12f89 CW |
3541 | trace_i915_gem_object_destroy(obj); |
3542 | ||
1286ff73 DV |
3543 | if (gem_obj->import_attach) |
3544 | drm_prime_gem_destroy(gem_obj, obj->sg_table); | |
3545 | ||
1488fc08 CW |
3546 | if (obj->phys_obj) |
3547 | i915_gem_detach_phys_object(dev, obj); | |
3548 | ||
3549 | obj->pin_count = 0; | |
3550 | if (WARN_ON(i915_gem_object_unbind(obj) == -ERESTARTSYS)) { | |
3551 | bool was_interruptible; | |
3552 | ||
3553 | was_interruptible = dev_priv->mm.interruptible; | |
3554 | dev_priv->mm.interruptible = false; | |
3555 | ||
3556 | WARN_ON(i915_gem_object_unbind(obj)); | |
3557 | ||
3558 | dev_priv->mm.interruptible = was_interruptible; | |
3559 | } | |
3560 | ||
05394f39 | 3561 | if (obj->base.map_list.map) |
b464e9a2 | 3562 | drm_gem_free_mmap_offset(&obj->base); |
de151cf6 | 3563 | |
05394f39 CW |
3564 | drm_gem_object_release(&obj->base); |
3565 | i915_gem_info_remove_obj(dev_priv, obj->base.size); | |
c397b908 | 3566 | |
05394f39 CW |
3567 | kfree(obj->bit_17); |
3568 | kfree(obj); | |
673a394b EA |
3569 | } |
3570 | ||
29105ccc CW |
3571 | int |
3572 | i915_gem_idle(struct drm_device *dev) | |
3573 | { | |
3574 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3575 | int ret; | |
28dfe52a | 3576 | |
29105ccc | 3577 | mutex_lock(&dev->struct_mutex); |
1c5d22f7 | 3578 | |
87acb0a5 | 3579 | if (dev_priv->mm.suspended) { |
29105ccc CW |
3580 | mutex_unlock(&dev->struct_mutex); |
3581 | return 0; | |
28dfe52a EA |
3582 | } |
3583 | ||
b2da9fe5 | 3584 | ret = i915_gpu_idle(dev); |
6dbe2772 KP |
3585 | if (ret) { |
3586 | mutex_unlock(&dev->struct_mutex); | |
673a394b | 3587 | return ret; |
6dbe2772 | 3588 | } |
b2da9fe5 | 3589 | i915_gem_retire_requests(dev); |
673a394b | 3590 | |
29105ccc | 3591 | /* Under UMS, be paranoid and evict. */ |
a39d7efc CW |
3592 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
3593 | i915_gem_evict_everything(dev, false); | |
29105ccc | 3594 | |
312817a3 CW |
3595 | i915_gem_reset_fences(dev); |
3596 | ||
29105ccc CW |
3597 | /* Hack! Don't let anybody do execbuf while we don't control the chip. |
3598 | * We need to replace this with a semaphore, or something. | |
3599 | * And not confound mm.suspended! | |
3600 | */ | |
3601 | dev_priv->mm.suspended = 1; | |
bc0c7f14 | 3602 | del_timer_sync(&dev_priv->hangcheck_timer); |
29105ccc CW |
3603 | |
3604 | i915_kernel_lost_context(dev); | |
6dbe2772 | 3605 | i915_gem_cleanup_ringbuffer(dev); |
29105ccc | 3606 | |
6dbe2772 KP |
3607 | mutex_unlock(&dev->struct_mutex); |
3608 | ||
29105ccc CW |
3609 | /* Cancel the retire work handler, which should be idle now. */ |
3610 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); | |
3611 | ||
673a394b EA |
3612 | return 0; |
3613 | } | |
3614 | ||
b9524a1e BW |
3615 | void i915_gem_l3_remap(struct drm_device *dev) |
3616 | { | |
3617 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3618 | u32 misccpctl; | |
3619 | int i; | |
3620 | ||
3621 | if (!IS_IVYBRIDGE(dev)) | |
3622 | return; | |
3623 | ||
3624 | if (!dev_priv->mm.l3_remap_info) | |
3625 | return; | |
3626 | ||
3627 | misccpctl = I915_READ(GEN7_MISCCPCTL); | |
3628 | I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE); | |
3629 | POSTING_READ(GEN7_MISCCPCTL); | |
3630 | ||
3631 | for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) { | |
3632 | u32 remap = I915_READ(GEN7_L3LOG_BASE + i); | |
3633 | if (remap && remap != dev_priv->mm.l3_remap_info[i/4]) | |
3634 | DRM_DEBUG("0x%x was already programmed to %x\n", | |
3635 | GEN7_L3LOG_BASE + i, remap); | |
3636 | if (remap && !dev_priv->mm.l3_remap_info[i/4]) | |
3637 | DRM_DEBUG_DRIVER("Clearing remapped register\n"); | |
3638 | I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->mm.l3_remap_info[i/4]); | |
3639 | } | |
3640 | ||
3641 | /* Make sure all the writes land before disabling dop clock gating */ | |
3642 | POSTING_READ(GEN7_L3LOG_BASE); | |
3643 | ||
3644 | I915_WRITE(GEN7_MISCCPCTL, misccpctl); | |
3645 | } | |
3646 | ||
f691e2f4 DV |
3647 | void i915_gem_init_swizzling(struct drm_device *dev) |
3648 | { | |
3649 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3650 | ||
11782b02 | 3651 | if (INTEL_INFO(dev)->gen < 5 || |
f691e2f4 DV |
3652 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
3653 | return; | |
3654 | ||
3655 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | | |
3656 | DISP_TILE_SURFACE_SWIZZLING); | |
3657 | ||
11782b02 DV |
3658 | if (IS_GEN5(dev)) |
3659 | return; | |
3660 | ||
f691e2f4 DV |
3661 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
3662 | if (IS_GEN6(dev)) | |
6b26c86d | 3663 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); |
f691e2f4 | 3664 | else |
6b26c86d | 3665 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); |
f691e2f4 | 3666 | } |
e21af88d DV |
3667 | |
3668 | void i915_gem_init_ppgtt(struct drm_device *dev) | |
3669 | { | |
3670 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3671 | uint32_t pd_offset; | |
3672 | struct intel_ring_buffer *ring; | |
55a254ac DV |
3673 | struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt; |
3674 | uint32_t __iomem *pd_addr; | |
3675 | uint32_t pd_entry; | |
e21af88d DV |
3676 | int i; |
3677 | ||
3678 | if (!dev_priv->mm.aliasing_ppgtt) | |
3679 | return; | |
3680 | ||
55a254ac DV |
3681 | |
3682 | pd_addr = dev_priv->mm.gtt->gtt + ppgtt->pd_offset/sizeof(uint32_t); | |
3683 | for (i = 0; i < ppgtt->num_pd_entries; i++) { | |
3684 | dma_addr_t pt_addr; | |
3685 | ||
3686 | if (dev_priv->mm.gtt->needs_dmar) | |
3687 | pt_addr = ppgtt->pt_dma_addr[i]; | |
3688 | else | |
3689 | pt_addr = page_to_phys(ppgtt->pt_pages[i]); | |
3690 | ||
3691 | pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr); | |
3692 | pd_entry |= GEN6_PDE_VALID; | |
3693 | ||
3694 | writel(pd_entry, pd_addr + i); | |
3695 | } | |
3696 | readl(pd_addr); | |
3697 | ||
3698 | pd_offset = ppgtt->pd_offset; | |
e21af88d DV |
3699 | pd_offset /= 64; /* in cachelines, */ |
3700 | pd_offset <<= 16; | |
3701 | ||
3702 | if (INTEL_INFO(dev)->gen == 6) { | |
48ecfa10 DV |
3703 | uint32_t ecochk, gab_ctl, ecobits; |
3704 | ||
3705 | ecobits = I915_READ(GAC_ECO_BITS); | |
3706 | I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B); | |
be901a5a DV |
3707 | |
3708 | gab_ctl = I915_READ(GAB_CTL); | |
3709 | I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT); | |
3710 | ||
3711 | ecochk = I915_READ(GAM_ECOCHK); | |
e21af88d DV |
3712 | I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | |
3713 | ECOCHK_PPGTT_CACHE64B); | |
6b26c86d | 3714 | I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE)); |
e21af88d DV |
3715 | } else if (INTEL_INFO(dev)->gen >= 7) { |
3716 | I915_WRITE(GAM_ECOCHK, ECOCHK_PPGTT_CACHE64B); | |
3717 | /* GFX_MODE is per-ring on gen7+ */ | |
3718 | } | |
3719 | ||
b4519513 | 3720 | for_each_ring(ring, dev_priv, i) { |
e21af88d DV |
3721 | if (INTEL_INFO(dev)->gen >= 7) |
3722 | I915_WRITE(RING_MODE_GEN7(ring), | |
6b26c86d | 3723 | _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE)); |
e21af88d DV |
3724 | |
3725 | I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G); | |
3726 | I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset); | |
3727 | } | |
3728 | } | |
3729 | ||
67b1b571 CW |
3730 | static bool |
3731 | intel_enable_blt(struct drm_device *dev) | |
3732 | { | |
3733 | if (!HAS_BLT(dev)) | |
3734 | return false; | |
3735 | ||
3736 | /* The blitter was dysfunctional on early prototypes */ | |
3737 | if (IS_GEN6(dev) && dev->pdev->revision < 8) { | |
3738 | DRM_INFO("BLT not supported on this pre-production hardware;" | |
3739 | " graphics performance will be degraded.\n"); | |
3740 | return false; | |
3741 | } | |
3742 | ||
3743 | return true; | |
3744 | } | |
3745 | ||
8187a2b7 | 3746 | int |
f691e2f4 | 3747 | i915_gem_init_hw(struct drm_device *dev) |
8187a2b7 ZN |
3748 | { |
3749 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3750 | int ret; | |
68f95ba9 | 3751 | |
8ecd1a66 DV |
3752 | if (!intel_enable_gtt()) |
3753 | return -EIO; | |
3754 | ||
b9524a1e BW |
3755 | i915_gem_l3_remap(dev); |
3756 | ||
f691e2f4 DV |
3757 | i915_gem_init_swizzling(dev); |
3758 | ||
5c1143bb | 3759 | ret = intel_init_render_ring_buffer(dev); |
68f95ba9 | 3760 | if (ret) |
b6913e4b | 3761 | return ret; |
68f95ba9 CW |
3762 | |
3763 | if (HAS_BSD(dev)) { | |
5c1143bb | 3764 | ret = intel_init_bsd_ring_buffer(dev); |
68f95ba9 CW |
3765 | if (ret) |
3766 | goto cleanup_render_ring; | |
d1b851fc | 3767 | } |
68f95ba9 | 3768 | |
67b1b571 | 3769 | if (intel_enable_blt(dev)) { |
549f7365 CW |
3770 | ret = intel_init_blt_ring_buffer(dev); |
3771 | if (ret) | |
3772 | goto cleanup_bsd_ring; | |
3773 | } | |
3774 | ||
6f392d54 CW |
3775 | dev_priv->next_seqno = 1; |
3776 | ||
254f965c BW |
3777 | /* |
3778 | * XXX: There was some w/a described somewhere suggesting loading | |
3779 | * contexts before PPGTT. | |
3780 | */ | |
3781 | i915_gem_context_init(dev); | |
e21af88d DV |
3782 | i915_gem_init_ppgtt(dev); |
3783 | ||
68f95ba9 CW |
3784 | return 0; |
3785 | ||
549f7365 | 3786 | cleanup_bsd_ring: |
1ec14ad3 | 3787 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); |
68f95ba9 | 3788 | cleanup_render_ring: |
1ec14ad3 | 3789 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); |
8187a2b7 ZN |
3790 | return ret; |
3791 | } | |
3792 | ||
1070a42b CW |
3793 | static bool |
3794 | intel_enable_ppgtt(struct drm_device *dev) | |
3795 | { | |
3796 | if (i915_enable_ppgtt >= 0) | |
3797 | return i915_enable_ppgtt; | |
3798 | ||
3799 | #ifdef CONFIG_INTEL_IOMMU | |
3800 | /* Disable ppgtt on SNB if VT-d is on. */ | |
3801 | if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) | |
3802 | return false; | |
3803 | #endif | |
3804 | ||
3805 | return true; | |
3806 | } | |
3807 | ||
3808 | int i915_gem_init(struct drm_device *dev) | |
3809 | { | |
3810 | struct drm_i915_private *dev_priv = dev->dev_private; | |
3811 | unsigned long gtt_size, mappable_size; | |
3812 | int ret; | |
3813 | ||
3814 | gtt_size = dev_priv->mm.gtt->gtt_total_entries << PAGE_SHIFT; | |
3815 | mappable_size = dev_priv->mm.gtt->gtt_mappable_entries << PAGE_SHIFT; | |
3816 | ||
3817 | mutex_lock(&dev->struct_mutex); | |
3818 | if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) { | |
3819 | /* PPGTT pdes are stolen from global gtt ptes, so shrink the | |
3820 | * aperture accordingly when using aliasing ppgtt. */ | |
3821 | gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE; | |
3822 | ||
3823 | i915_gem_init_global_gtt(dev, 0, mappable_size, gtt_size); | |
3824 | ||
3825 | ret = i915_gem_init_aliasing_ppgtt(dev); | |
3826 | if (ret) { | |
3827 | mutex_unlock(&dev->struct_mutex); | |
3828 | return ret; | |
3829 | } | |
3830 | } else { | |
3831 | /* Let GEM Manage all of the aperture. | |
3832 | * | |
3833 | * However, leave one page at the end still bound to the scratch | |
3834 | * page. There are a number of places where the hardware | |
3835 | * apparently prefetches past the end of the object, and we've | |
3836 | * seen multiple hangs with the GPU head pointer stuck in a | |
3837 | * batchbuffer bound at the last page of the aperture. One page | |
3838 | * should be enough to keep any prefetching inside of the | |
3839 | * aperture. | |
3840 | */ | |
3841 | i915_gem_init_global_gtt(dev, 0, mappable_size, | |
3842 | gtt_size); | |
3843 | } | |
3844 | ||
3845 | ret = i915_gem_init_hw(dev); | |
3846 | mutex_unlock(&dev->struct_mutex); | |
3847 | if (ret) { | |
3848 | i915_gem_cleanup_aliasing_ppgtt(dev); | |
3849 | return ret; | |
3850 | } | |
3851 | ||
53ca26ca DV |
3852 | /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */ |
3853 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) | |
3854 | dev_priv->dri1.allow_batchbuffer = 1; | |
1070a42b CW |
3855 | return 0; |
3856 | } | |
3857 | ||
8187a2b7 ZN |
3858 | void |
3859 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) | |
3860 | { | |
3861 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 3862 | struct intel_ring_buffer *ring; |
1ec14ad3 | 3863 | int i; |
8187a2b7 | 3864 | |
b4519513 CW |
3865 | for_each_ring(ring, dev_priv, i) |
3866 | intel_cleanup_ring_buffer(ring); | |
8187a2b7 ZN |
3867 | } |
3868 | ||
673a394b EA |
3869 | int |
3870 | i915_gem_entervt_ioctl(struct drm_device *dev, void *data, | |
3871 | struct drm_file *file_priv) | |
3872 | { | |
3873 | drm_i915_private_t *dev_priv = dev->dev_private; | |
b4519513 | 3874 | int ret; |
673a394b | 3875 | |
79e53945 JB |
3876 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
3877 | return 0; | |
3878 | ||
ba1234d1 | 3879 | if (atomic_read(&dev_priv->mm.wedged)) { |
673a394b | 3880 | DRM_ERROR("Reenabling wedged hardware, good luck\n"); |
ba1234d1 | 3881 | atomic_set(&dev_priv->mm.wedged, 0); |
673a394b EA |
3882 | } |
3883 | ||
673a394b | 3884 | mutex_lock(&dev->struct_mutex); |
9bb2d6f9 EA |
3885 | dev_priv->mm.suspended = 0; |
3886 | ||
f691e2f4 | 3887 | ret = i915_gem_init_hw(dev); |
d816f6ac WF |
3888 | if (ret != 0) { |
3889 | mutex_unlock(&dev->struct_mutex); | |
9bb2d6f9 | 3890 | return ret; |
d816f6ac | 3891 | } |
9bb2d6f9 | 3892 | |
69dc4987 | 3893 | BUG_ON(!list_empty(&dev_priv->mm.active_list)); |
673a394b EA |
3894 | BUG_ON(!list_empty(&dev_priv->mm.flushing_list)); |
3895 | BUG_ON(!list_empty(&dev_priv->mm.inactive_list)); | |
673a394b | 3896 | mutex_unlock(&dev->struct_mutex); |
dbb19d30 | 3897 | |
5f35308b CW |
3898 | ret = drm_irq_install(dev); |
3899 | if (ret) | |
3900 | goto cleanup_ringbuffer; | |
dbb19d30 | 3901 | |
673a394b | 3902 | return 0; |
5f35308b CW |
3903 | |
3904 | cleanup_ringbuffer: | |
3905 | mutex_lock(&dev->struct_mutex); | |
3906 | i915_gem_cleanup_ringbuffer(dev); | |
3907 | dev_priv->mm.suspended = 1; | |
3908 | mutex_unlock(&dev->struct_mutex); | |
3909 | ||
3910 | return ret; | |
673a394b EA |
3911 | } |
3912 | ||
3913 | int | |
3914 | i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, | |
3915 | struct drm_file *file_priv) | |
3916 | { | |
79e53945 JB |
3917 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
3918 | return 0; | |
3919 | ||
dbb19d30 | 3920 | drm_irq_uninstall(dev); |
e6890f6f | 3921 | return i915_gem_idle(dev); |
673a394b EA |
3922 | } |
3923 | ||
3924 | void | |
3925 | i915_gem_lastclose(struct drm_device *dev) | |
3926 | { | |
3927 | int ret; | |
673a394b | 3928 | |
e806b495 EA |
3929 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
3930 | return; | |
3931 | ||
6dbe2772 KP |
3932 | ret = i915_gem_idle(dev); |
3933 | if (ret) | |
3934 | DRM_ERROR("failed to idle hardware: %d\n", ret); | |
673a394b EA |
3935 | } |
3936 | ||
64193406 CW |
3937 | static void |
3938 | init_ring_lists(struct intel_ring_buffer *ring) | |
3939 | { | |
3940 | INIT_LIST_HEAD(&ring->active_list); | |
3941 | INIT_LIST_HEAD(&ring->request_list); | |
3942 | INIT_LIST_HEAD(&ring->gpu_write_list); | |
3943 | } | |
3944 | ||
673a394b EA |
3945 | void |
3946 | i915_gem_load(struct drm_device *dev) | |
3947 | { | |
b5aa8a0f | 3948 | int i; |
673a394b EA |
3949 | drm_i915_private_t *dev_priv = dev->dev_private; |
3950 | ||
69dc4987 | 3951 | INIT_LIST_HEAD(&dev_priv->mm.active_list); |
673a394b EA |
3952 | INIT_LIST_HEAD(&dev_priv->mm.flushing_list); |
3953 | INIT_LIST_HEAD(&dev_priv->mm.inactive_list); | |
a09ba7fa | 3954 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
93a37f20 | 3955 | INIT_LIST_HEAD(&dev_priv->mm.gtt_list); |
1ec14ad3 CW |
3956 | for (i = 0; i < I915_NUM_RINGS; i++) |
3957 | init_ring_lists(&dev_priv->ring[i]); | |
4b9de737 | 3958 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) |
007cc8ac | 3959 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); |
673a394b EA |
3960 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, |
3961 | i915_gem_retire_work_handler); | |
30dbf0c0 | 3962 | init_completion(&dev_priv->error_completion); |
31169714 | 3963 | |
94400120 DA |
3964 | /* On GEN3 we really need to make sure the ARB C3 LP bit is set */ |
3965 | if (IS_GEN3(dev)) { | |
50743298 DV |
3966 | I915_WRITE(MI_ARB_STATE, |
3967 | _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE)); | |
94400120 DA |
3968 | } |
3969 | ||
72bfa19c CW |
3970 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; |
3971 | ||
de151cf6 | 3972 | /* Old X drivers will take 0-2 for front, back, depth buffers */ |
b397c836 EA |
3973 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
3974 | dev_priv->fence_reg_start = 3; | |
de151cf6 | 3975 | |
a6c45cf0 | 3976 | if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) |
de151cf6 JB |
3977 | dev_priv->num_fence_regs = 16; |
3978 | else | |
3979 | dev_priv->num_fence_regs = 8; | |
3980 | ||
b5aa8a0f | 3981 | /* Initialize fence registers to zero */ |
ada726c7 | 3982 | i915_gem_reset_fences(dev); |
10ed13e4 | 3983 | |
673a394b | 3984 | i915_gem_detect_bit_6_swizzle(dev); |
6b95a207 | 3985 | init_waitqueue_head(&dev_priv->pending_flip_queue); |
17250b71 | 3986 | |
ce453d81 CW |
3987 | dev_priv->mm.interruptible = true; |
3988 | ||
17250b71 CW |
3989 | dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink; |
3990 | dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS; | |
3991 | register_shrinker(&dev_priv->mm.inactive_shrinker); | |
673a394b | 3992 | } |
71acb5eb DA |
3993 | |
3994 | /* | |
3995 | * Create a physically contiguous memory object for this object | |
3996 | * e.g. for cursor + overlay regs | |
3997 | */ | |
995b6762 CW |
3998 | static int i915_gem_init_phys_object(struct drm_device *dev, |
3999 | int id, int size, int align) | |
71acb5eb DA |
4000 | { |
4001 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4002 | struct drm_i915_gem_phys_object *phys_obj; | |
4003 | int ret; | |
4004 | ||
4005 | if (dev_priv->mm.phys_objs[id - 1] || !size) | |
4006 | return 0; | |
4007 | ||
9a298b2a | 4008 | phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL); |
71acb5eb DA |
4009 | if (!phys_obj) |
4010 | return -ENOMEM; | |
4011 | ||
4012 | phys_obj->id = id; | |
4013 | ||
6eeefaf3 | 4014 | phys_obj->handle = drm_pci_alloc(dev, size, align); |
71acb5eb DA |
4015 | if (!phys_obj->handle) { |
4016 | ret = -ENOMEM; | |
4017 | goto kfree_obj; | |
4018 | } | |
4019 | #ifdef CONFIG_X86 | |
4020 | set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
4021 | #endif | |
4022 | ||
4023 | dev_priv->mm.phys_objs[id - 1] = phys_obj; | |
4024 | ||
4025 | return 0; | |
4026 | kfree_obj: | |
9a298b2a | 4027 | kfree(phys_obj); |
71acb5eb DA |
4028 | return ret; |
4029 | } | |
4030 | ||
995b6762 | 4031 | static void i915_gem_free_phys_object(struct drm_device *dev, int id) |
71acb5eb DA |
4032 | { |
4033 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4034 | struct drm_i915_gem_phys_object *phys_obj; | |
4035 | ||
4036 | if (!dev_priv->mm.phys_objs[id - 1]) | |
4037 | return; | |
4038 | ||
4039 | phys_obj = dev_priv->mm.phys_objs[id - 1]; | |
4040 | if (phys_obj->cur_obj) { | |
4041 | i915_gem_detach_phys_object(dev, phys_obj->cur_obj); | |
4042 | } | |
4043 | ||
4044 | #ifdef CONFIG_X86 | |
4045 | set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
4046 | #endif | |
4047 | drm_pci_free(dev, phys_obj->handle); | |
4048 | kfree(phys_obj); | |
4049 | dev_priv->mm.phys_objs[id - 1] = NULL; | |
4050 | } | |
4051 | ||
4052 | void i915_gem_free_all_phys_object(struct drm_device *dev) | |
4053 | { | |
4054 | int i; | |
4055 | ||
260883c8 | 4056 | for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++) |
71acb5eb DA |
4057 | i915_gem_free_phys_object(dev, i); |
4058 | } | |
4059 | ||
4060 | void i915_gem_detach_phys_object(struct drm_device *dev, | |
05394f39 | 4061 | struct drm_i915_gem_object *obj) |
71acb5eb | 4062 | { |
05394f39 | 4063 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
e5281ccd | 4064 | char *vaddr; |
71acb5eb | 4065 | int i; |
71acb5eb DA |
4066 | int page_count; |
4067 | ||
05394f39 | 4068 | if (!obj->phys_obj) |
71acb5eb | 4069 | return; |
05394f39 | 4070 | vaddr = obj->phys_obj->handle->vaddr; |
71acb5eb | 4071 | |
05394f39 | 4072 | page_count = obj->base.size / PAGE_SIZE; |
71acb5eb | 4073 | for (i = 0; i < page_count; i++) { |
5949eac4 | 4074 | struct page *page = shmem_read_mapping_page(mapping, i); |
e5281ccd CW |
4075 | if (!IS_ERR(page)) { |
4076 | char *dst = kmap_atomic(page); | |
4077 | memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE); | |
4078 | kunmap_atomic(dst); | |
4079 | ||
4080 | drm_clflush_pages(&page, 1); | |
4081 | ||
4082 | set_page_dirty(page); | |
4083 | mark_page_accessed(page); | |
4084 | page_cache_release(page); | |
4085 | } | |
71acb5eb | 4086 | } |
40ce6575 | 4087 | intel_gtt_chipset_flush(); |
d78b47b9 | 4088 | |
05394f39 CW |
4089 | obj->phys_obj->cur_obj = NULL; |
4090 | obj->phys_obj = NULL; | |
71acb5eb DA |
4091 | } |
4092 | ||
4093 | int | |
4094 | i915_gem_attach_phys_object(struct drm_device *dev, | |
05394f39 | 4095 | struct drm_i915_gem_object *obj, |
6eeefaf3 CW |
4096 | int id, |
4097 | int align) | |
71acb5eb | 4098 | { |
05394f39 | 4099 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
71acb5eb | 4100 | drm_i915_private_t *dev_priv = dev->dev_private; |
71acb5eb DA |
4101 | int ret = 0; |
4102 | int page_count; | |
4103 | int i; | |
4104 | ||
4105 | if (id > I915_MAX_PHYS_OBJECT) | |
4106 | return -EINVAL; | |
4107 | ||
05394f39 CW |
4108 | if (obj->phys_obj) { |
4109 | if (obj->phys_obj->id == id) | |
71acb5eb DA |
4110 | return 0; |
4111 | i915_gem_detach_phys_object(dev, obj); | |
4112 | } | |
4113 | ||
71acb5eb DA |
4114 | /* create a new object */ |
4115 | if (!dev_priv->mm.phys_objs[id - 1]) { | |
4116 | ret = i915_gem_init_phys_object(dev, id, | |
05394f39 | 4117 | obj->base.size, align); |
71acb5eb | 4118 | if (ret) { |
05394f39 CW |
4119 | DRM_ERROR("failed to init phys object %d size: %zu\n", |
4120 | id, obj->base.size); | |
e5281ccd | 4121 | return ret; |
71acb5eb DA |
4122 | } |
4123 | } | |
4124 | ||
4125 | /* bind to the object */ | |
05394f39 CW |
4126 | obj->phys_obj = dev_priv->mm.phys_objs[id - 1]; |
4127 | obj->phys_obj->cur_obj = obj; | |
71acb5eb | 4128 | |
05394f39 | 4129 | page_count = obj->base.size / PAGE_SIZE; |
71acb5eb DA |
4130 | |
4131 | for (i = 0; i < page_count; i++) { | |
e5281ccd CW |
4132 | struct page *page; |
4133 | char *dst, *src; | |
4134 | ||
5949eac4 | 4135 | page = shmem_read_mapping_page(mapping, i); |
e5281ccd CW |
4136 | if (IS_ERR(page)) |
4137 | return PTR_ERR(page); | |
71acb5eb | 4138 | |
ff75b9bc | 4139 | src = kmap_atomic(page); |
05394f39 | 4140 | dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE); |
71acb5eb | 4141 | memcpy(dst, src, PAGE_SIZE); |
3e4d3af5 | 4142 | kunmap_atomic(src); |
71acb5eb | 4143 | |
e5281ccd CW |
4144 | mark_page_accessed(page); |
4145 | page_cache_release(page); | |
4146 | } | |
d78b47b9 | 4147 | |
71acb5eb | 4148 | return 0; |
71acb5eb DA |
4149 | } |
4150 | ||
4151 | static int | |
05394f39 CW |
4152 | i915_gem_phys_pwrite(struct drm_device *dev, |
4153 | struct drm_i915_gem_object *obj, | |
71acb5eb DA |
4154 | struct drm_i915_gem_pwrite *args, |
4155 | struct drm_file *file_priv) | |
4156 | { | |
05394f39 | 4157 | void *vaddr = obj->phys_obj->handle->vaddr + args->offset; |
b47b30cc | 4158 | char __user *user_data = (char __user *) (uintptr_t) args->data_ptr; |
71acb5eb | 4159 | |
b47b30cc CW |
4160 | if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) { |
4161 | unsigned long unwritten; | |
4162 | ||
4163 | /* The physical object once assigned is fixed for the lifetime | |
4164 | * of the obj, so we can safely drop the lock and continue | |
4165 | * to access vaddr. | |
4166 | */ | |
4167 | mutex_unlock(&dev->struct_mutex); | |
4168 | unwritten = copy_from_user(vaddr, user_data, args->size); | |
4169 | mutex_lock(&dev->struct_mutex); | |
4170 | if (unwritten) | |
4171 | return -EFAULT; | |
4172 | } | |
71acb5eb | 4173 | |
40ce6575 | 4174 | intel_gtt_chipset_flush(); |
71acb5eb DA |
4175 | return 0; |
4176 | } | |
b962442e | 4177 | |
f787a5f5 | 4178 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
b962442e | 4179 | { |
f787a5f5 | 4180 | struct drm_i915_file_private *file_priv = file->driver_priv; |
b962442e EA |
4181 | |
4182 | /* Clean up our request list when the client is going away, so that | |
4183 | * later retire_requests won't dereference our soon-to-be-gone | |
4184 | * file_priv. | |
4185 | */ | |
1c25595f | 4186 | spin_lock(&file_priv->mm.lock); |
f787a5f5 CW |
4187 | while (!list_empty(&file_priv->mm.request_list)) { |
4188 | struct drm_i915_gem_request *request; | |
4189 | ||
4190 | request = list_first_entry(&file_priv->mm.request_list, | |
4191 | struct drm_i915_gem_request, | |
4192 | client_list); | |
4193 | list_del(&request->client_list); | |
4194 | request->file_priv = NULL; | |
4195 | } | |
1c25595f | 4196 | spin_unlock(&file_priv->mm.lock); |
b962442e | 4197 | } |
31169714 | 4198 | |
1637ef41 CW |
4199 | static int |
4200 | i915_gpu_is_active(struct drm_device *dev) | |
4201 | { | |
4202 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4203 | int lists_empty; | |
4204 | ||
1637ef41 | 4205 | lists_empty = list_empty(&dev_priv->mm.flushing_list) && |
17250b71 | 4206 | list_empty(&dev_priv->mm.active_list); |
1637ef41 CW |
4207 | |
4208 | return !lists_empty; | |
4209 | } | |
4210 | ||
31169714 | 4211 | static int |
1495f230 | 4212 | i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc) |
31169714 | 4213 | { |
17250b71 CW |
4214 | struct drm_i915_private *dev_priv = |
4215 | container_of(shrinker, | |
4216 | struct drm_i915_private, | |
4217 | mm.inactive_shrinker); | |
4218 | struct drm_device *dev = dev_priv->dev; | |
4219 | struct drm_i915_gem_object *obj, *next; | |
1495f230 | 4220 | int nr_to_scan = sc->nr_to_scan; |
17250b71 CW |
4221 | int cnt; |
4222 | ||
4223 | if (!mutex_trylock(&dev->struct_mutex)) | |
bbe2e11a | 4224 | return 0; |
31169714 CW |
4225 | |
4226 | /* "fast-path" to count number of available objects */ | |
4227 | if (nr_to_scan == 0) { | |
17250b71 CW |
4228 | cnt = 0; |
4229 | list_for_each_entry(obj, | |
4230 | &dev_priv->mm.inactive_list, | |
4231 | mm_list) | |
4232 | cnt++; | |
4233 | mutex_unlock(&dev->struct_mutex); | |
4234 | return cnt / 100 * sysctl_vfs_cache_pressure; | |
31169714 CW |
4235 | } |
4236 | ||
1637ef41 | 4237 | rescan: |
31169714 | 4238 | /* first scan for clean buffers */ |
17250b71 | 4239 | i915_gem_retire_requests(dev); |
31169714 | 4240 | |
17250b71 CW |
4241 | list_for_each_entry_safe(obj, next, |
4242 | &dev_priv->mm.inactive_list, | |
4243 | mm_list) { | |
4244 | if (i915_gem_object_is_purgeable(obj)) { | |
2021746e CW |
4245 | if (i915_gem_object_unbind(obj) == 0 && |
4246 | --nr_to_scan == 0) | |
17250b71 | 4247 | break; |
31169714 | 4248 | } |
31169714 CW |
4249 | } |
4250 | ||
4251 | /* second pass, evict/count anything still on the inactive list */ | |
17250b71 CW |
4252 | cnt = 0; |
4253 | list_for_each_entry_safe(obj, next, | |
4254 | &dev_priv->mm.inactive_list, | |
4255 | mm_list) { | |
2021746e CW |
4256 | if (nr_to_scan && |
4257 | i915_gem_object_unbind(obj) == 0) | |
17250b71 | 4258 | nr_to_scan--; |
2021746e | 4259 | else |
17250b71 CW |
4260 | cnt++; |
4261 | } | |
4262 | ||
4263 | if (nr_to_scan && i915_gpu_is_active(dev)) { | |
1637ef41 CW |
4264 | /* |
4265 | * We are desperate for pages, so as a last resort, wait | |
4266 | * for the GPU to finish and discard whatever we can. | |
4267 | * This has a dramatic impact to reduce the number of | |
4268 | * OOM-killer events whilst running the GPU aggressively. | |
4269 | */ | |
b2da9fe5 | 4270 | if (i915_gpu_idle(dev) == 0) |
1637ef41 CW |
4271 | goto rescan; |
4272 | } | |
17250b71 CW |
4273 | mutex_unlock(&dev->struct_mutex); |
4274 | return cnt / 100 * sysctl_vfs_cache_pressure; | |
31169714 | 4275 | } |