]>
Commit | Line | Data |
---|---|---|
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 | ||
28 | #include <drm/drmP.h> | |
29 | #include <drm/i915_drm.h> | |
30 | #include "i915_drv.h" | |
31 | #include "i915_trace.h" | |
32 | #include "intel_drv.h" | |
33 | #include <linux/shmem_fs.h> | |
34 | #include <linux/slab.h> | |
35 | #include <linux/swap.h> | |
36 | #include <linux/pci.h> | |
37 | #include <linux/dma-buf.h> | |
38 | ||
39 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); | |
40 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj); | |
41 | static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj, | |
42 | unsigned alignment, | |
43 | bool map_and_fenceable, | |
44 | bool nonblocking); | |
45 | static int i915_gem_phys_pwrite(struct drm_device *dev, | |
46 | struct drm_i915_gem_object *obj, | |
47 | struct drm_i915_gem_pwrite *args, | |
48 | struct drm_file *file); | |
49 | ||
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 | ||
56 | static int i915_gem_inactive_shrink(struct shrinker *shrinker, | |
57 | struct shrink_control *sc); | |
58 | static long i915_gem_purge(struct drm_i915_private *dev_priv, long target); | |
59 | static void i915_gem_shrink_all(struct drm_i915_private *dev_priv); | |
60 | static void i915_gem_object_truncate(struct drm_i915_gem_object *obj); | |
61 | ||
62 | static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj) | |
63 | { | |
64 | if (obj->tiling_mode) | |
65 | i915_gem_release_mmap(obj); | |
66 | ||
67 | /* As we do not have an associated fence register, we will force | |
68 | * a tiling change if we ever need to acquire one. | |
69 | */ | |
70 | obj->fence_dirty = false; | |
71 | obj->fence_reg = I915_FENCE_REG_NONE; | |
72 | } | |
73 | ||
74 | /* some bookkeeping */ | |
75 | static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv, | |
76 | size_t size) | |
77 | { | |
78 | dev_priv->mm.object_count++; | |
79 | dev_priv->mm.object_memory += size; | |
80 | } | |
81 | ||
82 | static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv, | |
83 | size_t size) | |
84 | { | |
85 | dev_priv->mm.object_count--; | |
86 | dev_priv->mm.object_memory -= size; | |
87 | } | |
88 | ||
89 | static int | |
90 | i915_gem_wait_for_error(struct i915_gpu_error *error) | |
91 | { | |
92 | int ret; | |
93 | ||
94 | #define EXIT_COND (!i915_reset_in_progress(error)) | |
95 | if (EXIT_COND) | |
96 | return 0; | |
97 | ||
98 | /* GPU is already declared terminally dead, give up. */ | |
99 | if (i915_terminally_wedged(error)) | |
100 | return -EIO; | |
101 | ||
102 | /* | |
103 | * Only wait 10 seconds for the gpu reset to complete to avoid hanging | |
104 | * userspace. If it takes that long something really bad is going on and | |
105 | * we should simply try to bail out and fail as gracefully as possible. | |
106 | */ | |
107 | ret = wait_event_interruptible_timeout(error->reset_queue, | |
108 | EXIT_COND, | |
109 | 10*HZ); | |
110 | if (ret == 0) { | |
111 | DRM_ERROR("Timed out waiting for the gpu reset to complete\n"); | |
112 | return -EIO; | |
113 | } else if (ret < 0) { | |
114 | return ret; | |
115 | } | |
116 | #undef EXIT_COND | |
117 | ||
118 | return 0; | |
119 | } | |
120 | ||
121 | int i915_mutex_lock_interruptible(struct drm_device *dev) | |
122 | { | |
123 | struct drm_i915_private *dev_priv = dev->dev_private; | |
124 | int ret; | |
125 | ||
126 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); | |
127 | if (ret) | |
128 | return ret; | |
129 | ||
130 | ret = mutex_lock_interruptible(&dev->struct_mutex); | |
131 | if (ret) | |
132 | return ret; | |
133 | ||
134 | WARN_ON(i915_verify_lists(dev)); | |
135 | return 0; | |
136 | } | |
137 | ||
138 | static inline bool | |
139 | i915_gem_object_is_inactive(struct drm_i915_gem_object *obj) | |
140 | { | |
141 | return obj->gtt_space && !obj->active; | |
142 | } | |
143 | ||
144 | int | |
145 | i915_gem_init_ioctl(struct drm_device *dev, void *data, | |
146 | struct drm_file *file) | |
147 | { | |
148 | struct drm_i915_private *dev_priv = dev->dev_private; | |
149 | struct drm_i915_gem_init *args = data; | |
150 | ||
151 | if (drm_core_check_feature(dev, DRIVER_MODESET)) | |
152 | return -ENODEV; | |
153 | ||
154 | if (args->gtt_start >= args->gtt_end || | |
155 | (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1)) | |
156 | return -EINVAL; | |
157 | ||
158 | /* GEM with user mode setting was never supported on ilk and later. */ | |
159 | if (INTEL_INFO(dev)->gen >= 5) | |
160 | return -ENODEV; | |
161 | ||
162 | mutex_lock(&dev->struct_mutex); | |
163 | i915_gem_setup_global_gtt(dev, args->gtt_start, args->gtt_end, | |
164 | args->gtt_end); | |
165 | dev_priv->gtt.mappable_end = args->gtt_end; | |
166 | mutex_unlock(&dev->struct_mutex); | |
167 | ||
168 | return 0; | |
169 | } | |
170 | ||
171 | int | |
172 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, | |
173 | struct drm_file *file) | |
174 | { | |
175 | struct drm_i915_private *dev_priv = dev->dev_private; | |
176 | struct drm_i915_gem_get_aperture *args = data; | |
177 | struct drm_i915_gem_object *obj; | |
178 | size_t pinned; | |
179 | ||
180 | pinned = 0; | |
181 | mutex_lock(&dev->struct_mutex); | |
182 | list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list) | |
183 | if (obj->pin_count) | |
184 | pinned += obj->gtt_space->size; | |
185 | mutex_unlock(&dev->struct_mutex); | |
186 | ||
187 | args->aper_size = dev_priv->gtt.total; | |
188 | args->aper_available_size = args->aper_size - pinned; | |
189 | ||
190 | return 0; | |
191 | } | |
192 | ||
193 | void *i915_gem_object_alloc(struct drm_device *dev) | |
194 | { | |
195 | struct drm_i915_private *dev_priv = dev->dev_private; | |
196 | return kmem_cache_alloc(dev_priv->slab, GFP_KERNEL | __GFP_ZERO); | |
197 | } | |
198 | ||
199 | void i915_gem_object_free(struct drm_i915_gem_object *obj) | |
200 | { | |
201 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
202 | kmem_cache_free(dev_priv->slab, obj); | |
203 | } | |
204 | ||
205 | static int | |
206 | i915_gem_create(struct drm_file *file, | |
207 | struct drm_device *dev, | |
208 | uint64_t size, | |
209 | uint32_t *handle_p) | |
210 | { | |
211 | struct drm_i915_gem_object *obj; | |
212 | int ret; | |
213 | u32 handle; | |
214 | ||
215 | size = roundup(size, PAGE_SIZE); | |
216 | if (size == 0) | |
217 | return -EINVAL; | |
218 | ||
219 | /* Allocate the new object */ | |
220 | obj = i915_gem_alloc_object(dev, size); | |
221 | if (obj == NULL) | |
222 | return -ENOMEM; | |
223 | ||
224 | ret = drm_gem_handle_create(file, &obj->base, &handle); | |
225 | if (ret) { | |
226 | drm_gem_object_release(&obj->base); | |
227 | i915_gem_info_remove_obj(dev->dev_private, obj->base.size); | |
228 | i915_gem_object_free(obj); | |
229 | return ret; | |
230 | } | |
231 | ||
232 | /* drop reference from allocate - handle holds it now */ | |
233 | drm_gem_object_unreference(&obj->base); | |
234 | trace_i915_gem_object_create(obj); | |
235 | ||
236 | *handle_p = handle; | |
237 | return 0; | |
238 | } | |
239 | ||
240 | int | |
241 | i915_gem_dumb_create(struct drm_file *file, | |
242 | struct drm_device *dev, | |
243 | struct drm_mode_create_dumb *args) | |
244 | { | |
245 | /* have to work out size/pitch and return them */ | |
246 | args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64); | |
247 | args->size = args->pitch * args->height; | |
248 | return i915_gem_create(file, dev, | |
249 | args->size, &args->handle); | |
250 | } | |
251 | ||
252 | int i915_gem_dumb_destroy(struct drm_file *file, | |
253 | struct drm_device *dev, | |
254 | uint32_t handle) | |
255 | { | |
256 | return drm_gem_handle_delete(file, handle); | |
257 | } | |
258 | ||
259 | /** | |
260 | * Creates a new mm object and returns a handle to it. | |
261 | */ | |
262 | int | |
263 | i915_gem_create_ioctl(struct drm_device *dev, void *data, | |
264 | struct drm_file *file) | |
265 | { | |
266 | struct drm_i915_gem_create *args = data; | |
267 | ||
268 | return i915_gem_create(file, dev, | |
269 | args->size, &args->handle); | |
270 | } | |
271 | ||
272 | static inline int | |
273 | __copy_to_user_swizzled(char __user *cpu_vaddr, | |
274 | const char *gpu_vaddr, int gpu_offset, | |
275 | int length) | |
276 | { | |
277 | int ret, cpu_offset = 0; | |
278 | ||
279 | while (length > 0) { | |
280 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
281 | int this_length = min(cacheline_end - gpu_offset, length); | |
282 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
283 | ||
284 | ret = __copy_to_user(cpu_vaddr + cpu_offset, | |
285 | gpu_vaddr + swizzled_gpu_offset, | |
286 | this_length); | |
287 | if (ret) | |
288 | return ret + length; | |
289 | ||
290 | cpu_offset += this_length; | |
291 | gpu_offset += this_length; | |
292 | length -= this_length; | |
293 | } | |
294 | ||
295 | return 0; | |
296 | } | |
297 | ||
298 | static inline int | |
299 | __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, | |
300 | const char __user *cpu_vaddr, | |
301 | int length) | |
302 | { | |
303 | int ret, cpu_offset = 0; | |
304 | ||
305 | while (length > 0) { | |
306 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
307 | int this_length = min(cacheline_end - gpu_offset, length); | |
308 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
309 | ||
310 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, | |
311 | cpu_vaddr + cpu_offset, | |
312 | this_length); | |
313 | if (ret) | |
314 | return ret + length; | |
315 | ||
316 | cpu_offset += this_length; | |
317 | gpu_offset += this_length; | |
318 | length -= this_length; | |
319 | } | |
320 | ||
321 | return 0; | |
322 | } | |
323 | ||
324 | /* Per-page copy function for the shmem pread fastpath. | |
325 | * Flushes invalid cachelines before reading the target if | |
326 | * needs_clflush is set. */ | |
327 | static int | |
328 | shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length, | |
329 | char __user *user_data, | |
330 | bool page_do_bit17_swizzling, bool needs_clflush) | |
331 | { | |
332 | char *vaddr; | |
333 | int ret; | |
334 | ||
335 | if (unlikely(page_do_bit17_swizzling)) | |
336 | return -EINVAL; | |
337 | ||
338 | vaddr = kmap_atomic(page); | |
339 | if (needs_clflush) | |
340 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
341 | page_length); | |
342 | ret = __copy_to_user_inatomic(user_data, | |
343 | vaddr + shmem_page_offset, | |
344 | page_length); | |
345 | kunmap_atomic(vaddr); | |
346 | ||
347 | return ret ? -EFAULT : 0; | |
348 | } | |
349 | ||
350 | static void | |
351 | shmem_clflush_swizzled_range(char *addr, unsigned long length, | |
352 | bool swizzled) | |
353 | { | |
354 | if (unlikely(swizzled)) { | |
355 | unsigned long start = (unsigned long) addr; | |
356 | unsigned long end = (unsigned long) addr + length; | |
357 | ||
358 | /* For swizzling simply ensure that we always flush both | |
359 | * channels. Lame, but simple and it works. Swizzled | |
360 | * pwrite/pread is far from a hotpath - current userspace | |
361 | * doesn't use it at all. */ | |
362 | start = round_down(start, 128); | |
363 | end = round_up(end, 128); | |
364 | ||
365 | drm_clflush_virt_range((void *)start, end - start); | |
366 | } else { | |
367 | drm_clflush_virt_range(addr, length); | |
368 | } | |
369 | ||
370 | } | |
371 | ||
372 | /* Only difference to the fast-path function is that this can handle bit17 | |
373 | * and uses non-atomic copy and kmap functions. */ | |
374 | static int | |
375 | shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length, | |
376 | char __user *user_data, | |
377 | bool page_do_bit17_swizzling, bool needs_clflush) | |
378 | { | |
379 | char *vaddr; | |
380 | int ret; | |
381 | ||
382 | vaddr = kmap(page); | |
383 | if (needs_clflush) | |
384 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, | |
385 | page_length, | |
386 | page_do_bit17_swizzling); | |
387 | ||
388 | if (page_do_bit17_swizzling) | |
389 | ret = __copy_to_user_swizzled(user_data, | |
390 | vaddr, shmem_page_offset, | |
391 | page_length); | |
392 | else | |
393 | ret = __copy_to_user(user_data, | |
394 | vaddr + shmem_page_offset, | |
395 | page_length); | |
396 | kunmap(page); | |
397 | ||
398 | return ret ? - EFAULT : 0; | |
399 | } | |
400 | ||
401 | static int | |
402 | i915_gem_shmem_pread(struct drm_device *dev, | |
403 | struct drm_i915_gem_object *obj, | |
404 | struct drm_i915_gem_pread *args, | |
405 | struct drm_file *file) | |
406 | { | |
407 | char __user *user_data; | |
408 | ssize_t remain; | |
409 | loff_t offset; | |
410 | int shmem_page_offset, page_length, ret = 0; | |
411 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; | |
412 | int prefaulted = 0; | |
413 | int needs_clflush = 0; | |
414 | struct scatterlist *sg; | |
415 | int i; | |
416 | ||
417 | user_data = to_user_ptr(args->data_ptr); | |
418 | remain = args->size; | |
419 | ||
420 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); | |
421 | ||
422 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) { | |
423 | /* If we're not in the cpu read domain, set ourself into the gtt | |
424 | * read domain and manually flush cachelines (if required). This | |
425 | * optimizes for the case when the gpu will dirty the data | |
426 | * anyway again before the next pread happens. */ | |
427 | if (obj->cache_level == I915_CACHE_NONE) | |
428 | needs_clflush = 1; | |
429 | if (obj->gtt_space) { | |
430 | ret = i915_gem_object_set_to_gtt_domain(obj, false); | |
431 | if (ret) | |
432 | return ret; | |
433 | } | |
434 | } | |
435 | ||
436 | ret = i915_gem_object_get_pages(obj); | |
437 | if (ret) | |
438 | return ret; | |
439 | ||
440 | i915_gem_object_pin_pages(obj); | |
441 | ||
442 | offset = args->offset; | |
443 | ||
444 | for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) { | |
445 | struct page *page; | |
446 | ||
447 | if (i < offset >> PAGE_SHIFT) | |
448 | continue; | |
449 | ||
450 | if (remain <= 0) | |
451 | break; | |
452 | ||
453 | /* Operation in this page | |
454 | * | |
455 | * shmem_page_offset = offset within page in shmem file | |
456 | * page_length = bytes to copy for this page | |
457 | */ | |
458 | shmem_page_offset = offset_in_page(offset); | |
459 | page_length = remain; | |
460 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
461 | page_length = PAGE_SIZE - shmem_page_offset; | |
462 | ||
463 | page = sg_page(sg); | |
464 | page_do_bit17_swizzling = obj_do_bit17_swizzling && | |
465 | (page_to_phys(page) & (1 << 17)) != 0; | |
466 | ||
467 | ret = shmem_pread_fast(page, shmem_page_offset, page_length, | |
468 | user_data, page_do_bit17_swizzling, | |
469 | needs_clflush); | |
470 | if (ret == 0) | |
471 | goto next_page; | |
472 | ||
473 | mutex_unlock(&dev->struct_mutex); | |
474 | ||
475 | if (!prefaulted) { | |
476 | ret = fault_in_multipages_writeable(user_data, remain); | |
477 | /* Userspace is tricking us, but we've already clobbered | |
478 | * its pages with the prefault and promised to write the | |
479 | * data up to the first fault. Hence ignore any errors | |
480 | * and just continue. */ | |
481 | (void)ret; | |
482 | prefaulted = 1; | |
483 | } | |
484 | ||
485 | ret = shmem_pread_slow(page, shmem_page_offset, page_length, | |
486 | user_data, page_do_bit17_swizzling, | |
487 | needs_clflush); | |
488 | ||
489 | mutex_lock(&dev->struct_mutex); | |
490 | ||
491 | next_page: | |
492 | mark_page_accessed(page); | |
493 | ||
494 | if (ret) | |
495 | goto out; | |
496 | ||
497 | remain -= page_length; | |
498 | user_data += page_length; | |
499 | offset += page_length; | |
500 | } | |
501 | ||
502 | out: | |
503 | i915_gem_object_unpin_pages(obj); | |
504 | ||
505 | return ret; | |
506 | } | |
507 | ||
508 | /** | |
509 | * Reads data from the object referenced by handle. | |
510 | * | |
511 | * On error, the contents of *data are undefined. | |
512 | */ | |
513 | int | |
514 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, | |
515 | struct drm_file *file) | |
516 | { | |
517 | struct drm_i915_gem_pread *args = data; | |
518 | struct drm_i915_gem_object *obj; | |
519 | int ret = 0; | |
520 | ||
521 | if (args->size == 0) | |
522 | return 0; | |
523 | ||
524 | if (!access_ok(VERIFY_WRITE, | |
525 | to_user_ptr(args->data_ptr), | |
526 | args->size)) | |
527 | return -EFAULT; | |
528 | ||
529 | ret = i915_mutex_lock_interruptible(dev); | |
530 | if (ret) | |
531 | return ret; | |
532 | ||
533 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
534 | if (&obj->base == NULL) { | |
535 | ret = -ENOENT; | |
536 | goto unlock; | |
537 | } | |
538 | ||
539 | /* Bounds check source. */ | |
540 | if (args->offset > obj->base.size || | |
541 | args->size > obj->base.size - args->offset) { | |
542 | ret = -EINVAL; | |
543 | goto out; | |
544 | } | |
545 | ||
546 | /* prime objects have no backing filp to GEM pread/pwrite | |
547 | * pages from. | |
548 | */ | |
549 | if (!obj->base.filp) { | |
550 | ret = -EINVAL; | |
551 | goto out; | |
552 | } | |
553 | ||
554 | trace_i915_gem_object_pread(obj, args->offset, args->size); | |
555 | ||
556 | ret = i915_gem_shmem_pread(dev, obj, args, file); | |
557 | ||
558 | out: | |
559 | drm_gem_object_unreference(&obj->base); | |
560 | unlock: | |
561 | mutex_unlock(&dev->struct_mutex); | |
562 | return ret; | |
563 | } | |
564 | ||
565 | /* This is the fast write path which cannot handle | |
566 | * page faults in the source data | |
567 | */ | |
568 | ||
569 | static inline int | |
570 | fast_user_write(struct io_mapping *mapping, | |
571 | loff_t page_base, int page_offset, | |
572 | char __user *user_data, | |
573 | int length) | |
574 | { | |
575 | void __iomem *vaddr_atomic; | |
576 | void *vaddr; | |
577 | unsigned long unwritten; | |
578 | ||
579 | vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base); | |
580 | /* We can use the cpu mem copy function because this is X86. */ | |
581 | vaddr = (void __force*)vaddr_atomic + page_offset; | |
582 | unwritten = __copy_from_user_inatomic_nocache(vaddr, | |
583 | user_data, length); | |
584 | io_mapping_unmap_atomic(vaddr_atomic); | |
585 | return unwritten; | |
586 | } | |
587 | ||
588 | /** | |
589 | * This is the fast pwrite path, where we copy the data directly from the | |
590 | * user into the GTT, uncached. | |
591 | */ | |
592 | static int | |
593 | i915_gem_gtt_pwrite_fast(struct drm_device *dev, | |
594 | struct drm_i915_gem_object *obj, | |
595 | struct drm_i915_gem_pwrite *args, | |
596 | struct drm_file *file) | |
597 | { | |
598 | drm_i915_private_t *dev_priv = dev->dev_private; | |
599 | ssize_t remain; | |
600 | loff_t offset, page_base; | |
601 | char __user *user_data; | |
602 | int page_offset, page_length, ret; | |
603 | ||
604 | ret = i915_gem_object_pin(obj, 0, true, true); | |
605 | if (ret) | |
606 | goto out; | |
607 | ||
608 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
609 | if (ret) | |
610 | goto out_unpin; | |
611 | ||
612 | ret = i915_gem_object_put_fence(obj); | |
613 | if (ret) | |
614 | goto out_unpin; | |
615 | ||
616 | user_data = to_user_ptr(args->data_ptr); | |
617 | remain = args->size; | |
618 | ||
619 | offset = obj->gtt_offset + args->offset; | |
620 | ||
621 | while (remain > 0) { | |
622 | /* Operation in this page | |
623 | * | |
624 | * page_base = page offset within aperture | |
625 | * page_offset = offset within page | |
626 | * page_length = bytes to copy for this page | |
627 | */ | |
628 | page_base = offset & PAGE_MASK; | |
629 | page_offset = offset_in_page(offset); | |
630 | page_length = remain; | |
631 | if ((page_offset + remain) > PAGE_SIZE) | |
632 | page_length = PAGE_SIZE - page_offset; | |
633 | ||
634 | /* If we get a fault while copying data, then (presumably) our | |
635 | * source page isn't available. Return the error and we'll | |
636 | * retry in the slow path. | |
637 | */ | |
638 | if (fast_user_write(dev_priv->gtt.mappable, page_base, | |
639 | page_offset, user_data, page_length)) { | |
640 | ret = -EFAULT; | |
641 | goto out_unpin; | |
642 | } | |
643 | ||
644 | remain -= page_length; | |
645 | user_data += page_length; | |
646 | offset += page_length; | |
647 | } | |
648 | ||
649 | out_unpin: | |
650 | i915_gem_object_unpin(obj); | |
651 | out: | |
652 | return ret; | |
653 | } | |
654 | ||
655 | /* Per-page copy function for the shmem pwrite fastpath. | |
656 | * Flushes invalid cachelines before writing to the target if | |
657 | * needs_clflush_before is set and flushes out any written cachelines after | |
658 | * writing if needs_clflush is set. */ | |
659 | static int | |
660 | shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length, | |
661 | char __user *user_data, | |
662 | bool page_do_bit17_swizzling, | |
663 | bool needs_clflush_before, | |
664 | bool needs_clflush_after) | |
665 | { | |
666 | char *vaddr; | |
667 | int ret; | |
668 | ||
669 | if (unlikely(page_do_bit17_swizzling)) | |
670 | return -EINVAL; | |
671 | ||
672 | vaddr = kmap_atomic(page); | |
673 | if (needs_clflush_before) | |
674 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
675 | page_length); | |
676 | ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset, | |
677 | user_data, | |
678 | page_length); | |
679 | if (needs_clflush_after) | |
680 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
681 | page_length); | |
682 | kunmap_atomic(vaddr); | |
683 | ||
684 | return ret ? -EFAULT : 0; | |
685 | } | |
686 | ||
687 | /* Only difference to the fast-path function is that this can handle bit17 | |
688 | * and uses non-atomic copy and kmap functions. */ | |
689 | static int | |
690 | shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length, | |
691 | char __user *user_data, | |
692 | bool page_do_bit17_swizzling, | |
693 | bool needs_clflush_before, | |
694 | bool needs_clflush_after) | |
695 | { | |
696 | char *vaddr; | |
697 | int ret; | |
698 | ||
699 | vaddr = kmap(page); | |
700 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) | |
701 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, | |
702 | page_length, | |
703 | page_do_bit17_swizzling); | |
704 | if (page_do_bit17_swizzling) | |
705 | ret = __copy_from_user_swizzled(vaddr, shmem_page_offset, | |
706 | user_data, | |
707 | page_length); | |
708 | else | |
709 | ret = __copy_from_user(vaddr + shmem_page_offset, | |
710 | user_data, | |
711 | page_length); | |
712 | if (needs_clflush_after) | |
713 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, | |
714 | page_length, | |
715 | page_do_bit17_swizzling); | |
716 | kunmap(page); | |
717 | ||
718 | return ret ? -EFAULT : 0; | |
719 | } | |
720 | ||
721 | static int | |
722 | i915_gem_shmem_pwrite(struct drm_device *dev, | |
723 | struct drm_i915_gem_object *obj, | |
724 | struct drm_i915_gem_pwrite *args, | |
725 | struct drm_file *file) | |
726 | { | |
727 | ssize_t remain; | |
728 | loff_t offset; | |
729 | char __user *user_data; | |
730 | int shmem_page_offset, page_length, ret = 0; | |
731 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; | |
732 | int hit_slowpath = 0; | |
733 | int needs_clflush_after = 0; | |
734 | int needs_clflush_before = 0; | |
735 | int i; | |
736 | struct scatterlist *sg; | |
737 | ||
738 | user_data = to_user_ptr(args->data_ptr); | |
739 | remain = args->size; | |
740 | ||
741 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); | |
742 | ||
743 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { | |
744 | /* If we're not in the cpu write domain, set ourself into the gtt | |
745 | * write domain and manually flush cachelines (if required). This | |
746 | * optimizes for the case when the gpu will use the data | |
747 | * right away and we therefore have to clflush anyway. */ | |
748 | if (obj->cache_level == I915_CACHE_NONE) | |
749 | needs_clflush_after = 1; | |
750 | if (obj->gtt_space) { | |
751 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
752 | if (ret) | |
753 | return ret; | |
754 | } | |
755 | } | |
756 | /* Same trick applies for invalidate partially written cachelines before | |
757 | * writing. */ | |
758 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU) | |
759 | && obj->cache_level == I915_CACHE_NONE) | |
760 | needs_clflush_before = 1; | |
761 | ||
762 | ret = i915_gem_object_get_pages(obj); | |
763 | if (ret) | |
764 | return ret; | |
765 | ||
766 | i915_gem_object_pin_pages(obj); | |
767 | ||
768 | offset = args->offset; | |
769 | obj->dirty = 1; | |
770 | ||
771 | for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) { | |
772 | struct page *page; | |
773 | int partial_cacheline_write; | |
774 | ||
775 | if (i < offset >> PAGE_SHIFT) | |
776 | continue; | |
777 | ||
778 | if (remain <= 0) | |
779 | break; | |
780 | ||
781 | /* Operation in this page | |
782 | * | |
783 | * shmem_page_offset = offset within page in shmem file | |
784 | * page_length = bytes to copy for this page | |
785 | */ | |
786 | shmem_page_offset = offset_in_page(offset); | |
787 | ||
788 | page_length = remain; | |
789 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
790 | page_length = PAGE_SIZE - shmem_page_offset; | |
791 | ||
792 | /* If we don't overwrite a cacheline completely we need to be | |
793 | * careful to have up-to-date data by first clflushing. Don't | |
794 | * overcomplicate things and flush the entire patch. */ | |
795 | partial_cacheline_write = needs_clflush_before && | |
796 | ((shmem_page_offset | page_length) | |
797 | & (boot_cpu_data.x86_clflush_size - 1)); | |
798 | ||
799 | page = sg_page(sg); | |
800 | page_do_bit17_swizzling = obj_do_bit17_swizzling && | |
801 | (page_to_phys(page) & (1 << 17)) != 0; | |
802 | ||
803 | ret = shmem_pwrite_fast(page, shmem_page_offset, page_length, | |
804 | user_data, page_do_bit17_swizzling, | |
805 | partial_cacheline_write, | |
806 | needs_clflush_after); | |
807 | if (ret == 0) | |
808 | goto next_page; | |
809 | ||
810 | hit_slowpath = 1; | |
811 | mutex_unlock(&dev->struct_mutex); | |
812 | ret = shmem_pwrite_slow(page, shmem_page_offset, page_length, | |
813 | user_data, page_do_bit17_swizzling, | |
814 | partial_cacheline_write, | |
815 | needs_clflush_after); | |
816 | ||
817 | mutex_lock(&dev->struct_mutex); | |
818 | ||
819 | next_page: | |
820 | set_page_dirty(page); | |
821 | mark_page_accessed(page); | |
822 | ||
823 | if (ret) | |
824 | goto out; | |
825 | ||
826 | remain -= page_length; | |
827 | user_data += page_length; | |
828 | offset += page_length; | |
829 | } | |
830 | ||
831 | out: | |
832 | i915_gem_object_unpin_pages(obj); | |
833 | ||
834 | if (hit_slowpath) { | |
835 | /* | |
836 | * Fixup: Flush cpu caches in case we didn't flush the dirty | |
837 | * cachelines in-line while writing and the object moved | |
838 | * out of the cpu write domain while we've dropped the lock. | |
839 | */ | |
840 | if (!needs_clflush_after && | |
841 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { | |
842 | i915_gem_clflush_object(obj); | |
843 | i915_gem_chipset_flush(dev); | |
844 | } | |
845 | } | |
846 | ||
847 | if (needs_clflush_after) | |
848 | i915_gem_chipset_flush(dev); | |
849 | ||
850 | return ret; | |
851 | } | |
852 | ||
853 | /** | |
854 | * Writes data to the object referenced by handle. | |
855 | * | |
856 | * On error, the contents of the buffer that were to be modified are undefined. | |
857 | */ | |
858 | int | |
859 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, | |
860 | struct drm_file *file) | |
861 | { | |
862 | struct drm_i915_gem_pwrite *args = data; | |
863 | struct drm_i915_gem_object *obj; | |
864 | int ret; | |
865 | ||
866 | if (args->size == 0) | |
867 | return 0; | |
868 | ||
869 | if (!access_ok(VERIFY_READ, | |
870 | to_user_ptr(args->data_ptr), | |
871 | args->size)) | |
872 | return -EFAULT; | |
873 | ||
874 | ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr), | |
875 | args->size); | |
876 | if (ret) | |
877 | return -EFAULT; | |
878 | ||
879 | ret = i915_mutex_lock_interruptible(dev); | |
880 | if (ret) | |
881 | return ret; | |
882 | ||
883 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
884 | if (&obj->base == NULL) { | |
885 | ret = -ENOENT; | |
886 | goto unlock; | |
887 | } | |
888 | ||
889 | /* Bounds check destination. */ | |
890 | if (args->offset > obj->base.size || | |
891 | args->size > obj->base.size - args->offset) { | |
892 | ret = -EINVAL; | |
893 | goto out; | |
894 | } | |
895 | ||
896 | /* prime objects have no backing filp to GEM pread/pwrite | |
897 | * pages from. | |
898 | */ | |
899 | if (!obj->base.filp) { | |
900 | ret = -EINVAL; | |
901 | goto out; | |
902 | } | |
903 | ||
904 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); | |
905 | ||
906 | ret = -EFAULT; | |
907 | /* We can only do the GTT pwrite on untiled buffers, as otherwise | |
908 | * it would end up going through the fenced access, and we'll get | |
909 | * different detiling behavior between reading and writing. | |
910 | * pread/pwrite currently are reading and writing from the CPU | |
911 | * perspective, requiring manual detiling by the client. | |
912 | */ | |
913 | if (obj->phys_obj) { | |
914 | ret = i915_gem_phys_pwrite(dev, obj, args, file); | |
915 | goto out; | |
916 | } | |
917 | ||
918 | if (obj->cache_level == I915_CACHE_NONE && | |
919 | obj->tiling_mode == I915_TILING_NONE && | |
920 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { | |
921 | ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); | |
922 | /* Note that the gtt paths might fail with non-page-backed user | |
923 | * pointers (e.g. gtt mappings when moving data between | |
924 | * textures). Fallback to the shmem path in that case. */ | |
925 | } | |
926 | ||
927 | if (ret == -EFAULT || ret == -ENOSPC) | |
928 | ret = i915_gem_shmem_pwrite(dev, obj, args, file); | |
929 | ||
930 | out: | |
931 | drm_gem_object_unreference(&obj->base); | |
932 | unlock: | |
933 | mutex_unlock(&dev->struct_mutex); | |
934 | return ret; | |
935 | } | |
936 | ||
937 | int | |
938 | i915_gem_check_wedge(struct i915_gpu_error *error, | |
939 | bool interruptible) | |
940 | { | |
941 | if (i915_reset_in_progress(error)) { | |
942 | /* Non-interruptible callers can't handle -EAGAIN, hence return | |
943 | * -EIO unconditionally for these. */ | |
944 | if (!interruptible) | |
945 | return -EIO; | |
946 | ||
947 | /* Recovery complete, but the reset failed ... */ | |
948 | if (i915_terminally_wedged(error)) | |
949 | return -EIO; | |
950 | ||
951 | return -EAGAIN; | |
952 | } | |
953 | ||
954 | return 0; | |
955 | } | |
956 | ||
957 | /* | |
958 | * Compare seqno against outstanding lazy request. Emit a request if they are | |
959 | * equal. | |
960 | */ | |
961 | static int | |
962 | i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno) | |
963 | { | |
964 | int ret; | |
965 | ||
966 | BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex)); | |
967 | ||
968 | ret = 0; | |
969 | if (seqno == ring->outstanding_lazy_request) | |
970 | ret = i915_add_request(ring, NULL, NULL); | |
971 | ||
972 | return ret; | |
973 | } | |
974 | ||
975 | /** | |
976 | * __wait_seqno - wait until execution of seqno has finished | |
977 | * @ring: the ring expected to report seqno | |
978 | * @seqno: duh! | |
979 | * @reset_counter: reset sequence associated with the given seqno | |
980 | * @interruptible: do an interruptible wait (normally yes) | |
981 | * @timeout: in - how long to wait (NULL forever); out - how much time remaining | |
982 | * | |
983 | * Note: It is of utmost importance that the passed in seqno and reset_counter | |
984 | * values have been read by the caller in an smp safe manner. Where read-side | |
985 | * locks are involved, it is sufficient to read the reset_counter before | |
986 | * unlocking the lock that protects the seqno. For lockless tricks, the | |
987 | * reset_counter _must_ be read before, and an appropriate smp_rmb must be | |
988 | * inserted. | |
989 | * | |
990 | * Returns 0 if the seqno was found within the alloted time. Else returns the | |
991 | * errno with remaining time filled in timeout argument. | |
992 | */ | |
993 | static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno, | |
994 | unsigned reset_counter, | |
995 | bool interruptible, struct timespec *timeout) | |
996 | { | |
997 | drm_i915_private_t *dev_priv = ring->dev->dev_private; | |
998 | struct timespec before, now, wait_time={1,0}; | |
999 | unsigned long timeout_jiffies; | |
1000 | long end; | |
1001 | bool wait_forever = true; | |
1002 | int ret; | |
1003 | ||
1004 | if (i915_seqno_passed(ring->get_seqno(ring, true), seqno)) | |
1005 | return 0; | |
1006 | ||
1007 | trace_i915_gem_request_wait_begin(ring, seqno); | |
1008 | ||
1009 | if (timeout != NULL) { | |
1010 | wait_time = *timeout; | |
1011 | wait_forever = false; | |
1012 | } | |
1013 | ||
1014 | timeout_jiffies = timespec_to_jiffies(&wait_time); | |
1015 | ||
1016 | if (WARN_ON(!ring->irq_get(ring))) | |
1017 | return -ENODEV; | |
1018 | ||
1019 | /* Record current time in case interrupted by signal, or wedged * */ | |
1020 | getrawmonotonic(&before); | |
1021 | ||
1022 | #define EXIT_COND \ | |
1023 | (i915_seqno_passed(ring->get_seqno(ring, false), seqno) || \ | |
1024 | i915_reset_in_progress(&dev_priv->gpu_error) || \ | |
1025 | reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) | |
1026 | do { | |
1027 | if (interruptible) | |
1028 | end = wait_event_interruptible_timeout(ring->irq_queue, | |
1029 | EXIT_COND, | |
1030 | timeout_jiffies); | |
1031 | else | |
1032 | end = wait_event_timeout(ring->irq_queue, EXIT_COND, | |
1033 | timeout_jiffies); | |
1034 | ||
1035 | /* We need to check whether any gpu reset happened in between | |
1036 | * the caller grabbing the seqno and now ... */ | |
1037 | if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) | |
1038 | end = -EAGAIN; | |
1039 | ||
1040 | /* ... but upgrade the -EGAIN to an -EIO if the gpu is truely | |
1041 | * gone. */ | |
1042 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); | |
1043 | if (ret) | |
1044 | end = ret; | |
1045 | } while (end == 0 && wait_forever); | |
1046 | ||
1047 | getrawmonotonic(&now); | |
1048 | ||
1049 | ring->irq_put(ring); | |
1050 | trace_i915_gem_request_wait_end(ring, seqno); | |
1051 | #undef EXIT_COND | |
1052 | ||
1053 | if (timeout) { | |
1054 | struct timespec sleep_time = timespec_sub(now, before); | |
1055 | *timeout = timespec_sub(*timeout, sleep_time); | |
1056 | } | |
1057 | ||
1058 | switch (end) { | |
1059 | case -EIO: | |
1060 | case -EAGAIN: /* Wedged */ | |
1061 | case -ERESTARTSYS: /* Signal */ | |
1062 | return (int)end; | |
1063 | case 0: /* Timeout */ | |
1064 | if (timeout) | |
1065 | set_normalized_timespec(timeout, 0, 0); | |
1066 | return -ETIME; | |
1067 | default: /* Completed */ | |
1068 | WARN_ON(end < 0); /* We're not aware of other errors */ | |
1069 | return 0; | |
1070 | } | |
1071 | } | |
1072 | ||
1073 | /** | |
1074 | * Waits for a sequence number to be signaled, and cleans up the | |
1075 | * request and object lists appropriately for that event. | |
1076 | */ | |
1077 | int | |
1078 | i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno) | |
1079 | { | |
1080 | struct drm_device *dev = ring->dev; | |
1081 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1082 | bool interruptible = dev_priv->mm.interruptible; | |
1083 | int ret; | |
1084 | ||
1085 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); | |
1086 | BUG_ON(seqno == 0); | |
1087 | ||
1088 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible); | |
1089 | if (ret) | |
1090 | return ret; | |
1091 | ||
1092 | ret = i915_gem_check_olr(ring, seqno); | |
1093 | if (ret) | |
1094 | return ret; | |
1095 | ||
1096 | return __wait_seqno(ring, seqno, | |
1097 | atomic_read(&dev_priv->gpu_error.reset_counter), | |
1098 | interruptible, NULL); | |
1099 | } | |
1100 | ||
1101 | /** | |
1102 | * Ensures that all rendering to the object has completed and the object is | |
1103 | * safe to unbind from the GTT or access from the CPU. | |
1104 | */ | |
1105 | static __must_check int | |
1106 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj, | |
1107 | bool readonly) | |
1108 | { | |
1109 | struct intel_ring_buffer *ring = obj->ring; | |
1110 | u32 seqno; | |
1111 | int ret; | |
1112 | ||
1113 | seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno; | |
1114 | if (seqno == 0) | |
1115 | return 0; | |
1116 | ||
1117 | ret = i915_wait_seqno(ring, seqno); | |
1118 | if (ret) | |
1119 | return ret; | |
1120 | ||
1121 | i915_gem_retire_requests_ring(ring); | |
1122 | ||
1123 | /* Manually manage the write flush as we may have not yet | |
1124 | * retired the buffer. | |
1125 | */ | |
1126 | if (obj->last_write_seqno && | |
1127 | i915_seqno_passed(seqno, obj->last_write_seqno)) { | |
1128 | obj->last_write_seqno = 0; | |
1129 | obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS; | |
1130 | } | |
1131 | ||
1132 | return 0; | |
1133 | } | |
1134 | ||
1135 | /* A nonblocking variant of the above wait. This is a highly dangerous routine | |
1136 | * as the object state may change during this call. | |
1137 | */ | |
1138 | static __must_check int | |
1139 | i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj, | |
1140 | bool readonly) | |
1141 | { | |
1142 | struct drm_device *dev = obj->base.dev; | |
1143 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1144 | struct intel_ring_buffer *ring = obj->ring; | |
1145 | unsigned reset_counter; | |
1146 | u32 seqno; | |
1147 | int ret; | |
1148 | ||
1149 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); | |
1150 | BUG_ON(!dev_priv->mm.interruptible); | |
1151 | ||
1152 | seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno; | |
1153 | if (seqno == 0) | |
1154 | return 0; | |
1155 | ||
1156 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, true); | |
1157 | if (ret) | |
1158 | return ret; | |
1159 | ||
1160 | ret = i915_gem_check_olr(ring, seqno); | |
1161 | if (ret) | |
1162 | return ret; | |
1163 | ||
1164 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); | |
1165 | mutex_unlock(&dev->struct_mutex); | |
1166 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL); | |
1167 | mutex_lock(&dev->struct_mutex); | |
1168 | ||
1169 | i915_gem_retire_requests_ring(ring); | |
1170 | ||
1171 | /* Manually manage the write flush as we may have not yet | |
1172 | * retired the buffer. | |
1173 | */ | |
1174 | if (obj->last_write_seqno && | |
1175 | i915_seqno_passed(seqno, obj->last_write_seqno)) { | |
1176 | obj->last_write_seqno = 0; | |
1177 | obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS; | |
1178 | } | |
1179 | ||
1180 | return ret; | |
1181 | } | |
1182 | ||
1183 | /** | |
1184 | * Called when user space prepares to use an object with the CPU, either | |
1185 | * through the mmap ioctl's mapping or a GTT mapping. | |
1186 | */ | |
1187 | int | |
1188 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, | |
1189 | struct drm_file *file) | |
1190 | { | |
1191 | struct drm_i915_gem_set_domain *args = data; | |
1192 | struct drm_i915_gem_object *obj; | |
1193 | uint32_t read_domains = args->read_domains; | |
1194 | uint32_t write_domain = args->write_domain; | |
1195 | int ret; | |
1196 | ||
1197 | /* Only handle setting domains to types used by the CPU. */ | |
1198 | if (write_domain & I915_GEM_GPU_DOMAINS) | |
1199 | return -EINVAL; | |
1200 | ||
1201 | if (read_domains & I915_GEM_GPU_DOMAINS) | |
1202 | return -EINVAL; | |
1203 | ||
1204 | /* Having something in the write domain implies it's in the read | |
1205 | * domain, and only that read domain. Enforce that in the request. | |
1206 | */ | |
1207 | if (write_domain != 0 && read_domains != write_domain) | |
1208 | return -EINVAL; | |
1209 | ||
1210 | ret = i915_mutex_lock_interruptible(dev); | |
1211 | if (ret) | |
1212 | return ret; | |
1213 | ||
1214 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
1215 | if (&obj->base == NULL) { | |
1216 | ret = -ENOENT; | |
1217 | goto unlock; | |
1218 | } | |
1219 | ||
1220 | /* Try to flush the object off the GPU without holding the lock. | |
1221 | * We will repeat the flush holding the lock in the normal manner | |
1222 | * to catch cases where we are gazumped. | |
1223 | */ | |
1224 | ret = i915_gem_object_wait_rendering__nonblocking(obj, !write_domain); | |
1225 | if (ret) | |
1226 | goto unref; | |
1227 | ||
1228 | if (read_domains & I915_GEM_DOMAIN_GTT) { | |
1229 | ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); | |
1230 | ||
1231 | /* Silently promote "you're not bound, there was nothing to do" | |
1232 | * to success, since the client was just asking us to | |
1233 | * make sure everything was done. | |
1234 | */ | |
1235 | if (ret == -EINVAL) | |
1236 | ret = 0; | |
1237 | } else { | |
1238 | ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); | |
1239 | } | |
1240 | ||
1241 | unref: | |
1242 | drm_gem_object_unreference(&obj->base); | |
1243 | unlock: | |
1244 | mutex_unlock(&dev->struct_mutex); | |
1245 | return ret; | |
1246 | } | |
1247 | ||
1248 | /** | |
1249 | * Called when user space has done writes to this buffer | |
1250 | */ | |
1251 | int | |
1252 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, | |
1253 | struct drm_file *file) | |
1254 | { | |
1255 | struct drm_i915_gem_sw_finish *args = data; | |
1256 | struct drm_i915_gem_object *obj; | |
1257 | int ret = 0; | |
1258 | ||
1259 | ret = i915_mutex_lock_interruptible(dev); | |
1260 | if (ret) | |
1261 | return ret; | |
1262 | ||
1263 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
1264 | if (&obj->base == NULL) { | |
1265 | ret = -ENOENT; | |
1266 | goto unlock; | |
1267 | } | |
1268 | ||
1269 | /* Pinned buffers may be scanout, so flush the cache */ | |
1270 | if (obj->pin_count) | |
1271 | i915_gem_object_flush_cpu_write_domain(obj); | |
1272 | ||
1273 | drm_gem_object_unreference(&obj->base); | |
1274 | unlock: | |
1275 | mutex_unlock(&dev->struct_mutex); | |
1276 | return ret; | |
1277 | } | |
1278 | ||
1279 | /** | |
1280 | * Maps the contents of an object, returning the address it is mapped | |
1281 | * into. | |
1282 | * | |
1283 | * While the mapping holds a reference on the contents of the object, it doesn't | |
1284 | * imply a ref on the object itself. | |
1285 | */ | |
1286 | int | |
1287 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, | |
1288 | struct drm_file *file) | |
1289 | { | |
1290 | struct drm_i915_gem_mmap *args = data; | |
1291 | struct drm_gem_object *obj; | |
1292 | unsigned long addr; | |
1293 | ||
1294 | obj = drm_gem_object_lookup(dev, file, args->handle); | |
1295 | if (obj == NULL) | |
1296 | return -ENOENT; | |
1297 | ||
1298 | /* prime objects have no backing filp to GEM mmap | |
1299 | * pages from. | |
1300 | */ | |
1301 | if (!obj->filp) { | |
1302 | drm_gem_object_unreference_unlocked(obj); | |
1303 | return -EINVAL; | |
1304 | } | |
1305 | ||
1306 | addr = vm_mmap(obj->filp, 0, args->size, | |
1307 | PROT_READ | PROT_WRITE, MAP_SHARED, | |
1308 | args->offset); | |
1309 | drm_gem_object_unreference_unlocked(obj); | |
1310 | if (IS_ERR((void *)addr)) | |
1311 | return addr; | |
1312 | ||
1313 | args->addr_ptr = (uint64_t) addr; | |
1314 | ||
1315 | return 0; | |
1316 | } | |
1317 | ||
1318 | /** | |
1319 | * i915_gem_fault - fault a page into the GTT | |
1320 | * vma: VMA in question | |
1321 | * vmf: fault info | |
1322 | * | |
1323 | * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped | |
1324 | * from userspace. The fault handler takes care of binding the object to | |
1325 | * the GTT (if needed), allocating and programming a fence register (again, | |
1326 | * only if needed based on whether the old reg is still valid or the object | |
1327 | * is tiled) and inserting a new PTE into the faulting process. | |
1328 | * | |
1329 | * Note that the faulting process may involve evicting existing objects | |
1330 | * from the GTT and/or fence registers to make room. So performance may | |
1331 | * suffer if the GTT working set is large or there are few fence registers | |
1332 | * left. | |
1333 | */ | |
1334 | int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
1335 | { | |
1336 | struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data); | |
1337 | struct drm_device *dev = obj->base.dev; | |
1338 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1339 | pgoff_t page_offset; | |
1340 | unsigned long pfn; | |
1341 | int ret = 0; | |
1342 | bool write = !!(vmf->flags & FAULT_FLAG_WRITE); | |
1343 | ||
1344 | /* We don't use vmf->pgoff since that has the fake offset */ | |
1345 | page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >> | |
1346 | PAGE_SHIFT; | |
1347 | ||
1348 | ret = i915_mutex_lock_interruptible(dev); | |
1349 | if (ret) | |
1350 | goto out; | |
1351 | ||
1352 | trace_i915_gem_object_fault(obj, page_offset, true, write); | |
1353 | ||
1354 | /* Access to snoopable pages through the GTT is incoherent. */ | |
1355 | if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) { | |
1356 | ret = -EINVAL; | |
1357 | goto unlock; | |
1358 | } | |
1359 | ||
1360 | /* Now bind it into the GTT if needed */ | |
1361 | ret = i915_gem_object_pin(obj, 0, true, false); | |
1362 | if (ret) | |
1363 | goto unlock; | |
1364 | ||
1365 | ret = i915_gem_object_set_to_gtt_domain(obj, write); | |
1366 | if (ret) | |
1367 | goto unpin; | |
1368 | ||
1369 | ret = i915_gem_object_get_fence(obj); | |
1370 | if (ret) | |
1371 | goto unpin; | |
1372 | ||
1373 | obj->fault_mappable = true; | |
1374 | ||
1375 | pfn = ((dev_priv->gtt.mappable_base + obj->gtt_offset) >> PAGE_SHIFT) + | |
1376 | page_offset; | |
1377 | ||
1378 | /* Finally, remap it using the new GTT offset */ | |
1379 | ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn); | |
1380 | unpin: | |
1381 | i915_gem_object_unpin(obj); | |
1382 | unlock: | |
1383 | mutex_unlock(&dev->struct_mutex); | |
1384 | out: | |
1385 | switch (ret) { | |
1386 | case -EIO: | |
1387 | /* If this -EIO is due to a gpu hang, give the reset code a | |
1388 | * chance to clean up the mess. Otherwise return the proper | |
1389 | * SIGBUS. */ | |
1390 | if (i915_terminally_wedged(&dev_priv->gpu_error)) | |
1391 | return VM_FAULT_SIGBUS; | |
1392 | case -EAGAIN: | |
1393 | /* Give the error handler a chance to run and move the | |
1394 | * objects off the GPU active list. Next time we service the | |
1395 | * fault, we should be able to transition the page into the | |
1396 | * GTT without touching the GPU (and so avoid further | |
1397 | * EIO/EGAIN). If the GPU is wedged, then there is no issue | |
1398 | * with coherency, just lost writes. | |
1399 | */ | |
1400 | set_need_resched(); | |
1401 | case 0: | |
1402 | case -ERESTARTSYS: | |
1403 | case -EINTR: | |
1404 | case -EBUSY: | |
1405 | /* | |
1406 | * EBUSY is ok: this just means that another thread | |
1407 | * already did the job. | |
1408 | */ | |
1409 | return VM_FAULT_NOPAGE; | |
1410 | case -ENOMEM: | |
1411 | return VM_FAULT_OOM; | |
1412 | case -ENOSPC: | |
1413 | return VM_FAULT_SIGBUS; | |
1414 | default: | |
1415 | WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret); | |
1416 | return VM_FAULT_SIGBUS; | |
1417 | } | |
1418 | } | |
1419 | ||
1420 | /** | |
1421 | * i915_gem_release_mmap - remove physical page mappings | |
1422 | * @obj: obj in question | |
1423 | * | |
1424 | * Preserve the reservation of the mmapping with the DRM core code, but | |
1425 | * relinquish ownership of the pages back to the system. | |
1426 | * | |
1427 | * It is vital that we remove the page mapping if we have mapped a tiled | |
1428 | * object through the GTT and then lose the fence register due to | |
1429 | * resource pressure. Similarly if the object has been moved out of the | |
1430 | * aperture, than pages mapped into userspace must be revoked. Removing the | |
1431 | * mapping will then trigger a page fault on the next user access, allowing | |
1432 | * fixup by i915_gem_fault(). | |
1433 | */ | |
1434 | void | |
1435 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) | |
1436 | { | |
1437 | if (!obj->fault_mappable) | |
1438 | return; | |
1439 | ||
1440 | if (obj->base.dev->dev_mapping) | |
1441 | unmap_mapping_range(obj->base.dev->dev_mapping, | |
1442 | (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT, | |
1443 | obj->base.size, 1); | |
1444 | ||
1445 | obj->fault_mappable = false; | |
1446 | } | |
1447 | ||
1448 | uint32_t | |
1449 | i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode) | |
1450 | { | |
1451 | uint32_t gtt_size; | |
1452 | ||
1453 | if (INTEL_INFO(dev)->gen >= 4 || | |
1454 | tiling_mode == I915_TILING_NONE) | |
1455 | return size; | |
1456 | ||
1457 | /* Previous chips need a power-of-two fence region when tiling */ | |
1458 | if (INTEL_INFO(dev)->gen == 3) | |
1459 | gtt_size = 1024*1024; | |
1460 | else | |
1461 | gtt_size = 512*1024; | |
1462 | ||
1463 | while (gtt_size < size) | |
1464 | gtt_size <<= 1; | |
1465 | ||
1466 | return gtt_size; | |
1467 | } | |
1468 | ||
1469 | /** | |
1470 | * i915_gem_get_gtt_alignment - return required GTT alignment for an object | |
1471 | * @obj: object to check | |
1472 | * | |
1473 | * Return the required GTT alignment for an object, taking into account | |
1474 | * potential fence register mapping. | |
1475 | */ | |
1476 | uint32_t | |
1477 | i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size, | |
1478 | int tiling_mode, bool fenced) | |
1479 | { | |
1480 | /* | |
1481 | * Minimum alignment is 4k (GTT page size), but might be greater | |
1482 | * if a fence register is needed for the object. | |
1483 | */ | |
1484 | if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) || | |
1485 | tiling_mode == I915_TILING_NONE) | |
1486 | return 4096; | |
1487 | ||
1488 | /* | |
1489 | * Previous chips need to be aligned to the size of the smallest | |
1490 | * fence register that can contain the object. | |
1491 | */ | |
1492 | return i915_gem_get_gtt_size(dev, size, tiling_mode); | |
1493 | } | |
1494 | ||
1495 | static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj) | |
1496 | { | |
1497 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
1498 | int ret; | |
1499 | ||
1500 | if (obj->base.map_list.map) | |
1501 | return 0; | |
1502 | ||
1503 | dev_priv->mm.shrinker_no_lock_stealing = true; | |
1504 | ||
1505 | ret = drm_gem_create_mmap_offset(&obj->base); | |
1506 | if (ret != -ENOSPC) | |
1507 | goto out; | |
1508 | ||
1509 | /* Badly fragmented mmap space? The only way we can recover | |
1510 | * space is by destroying unwanted objects. We can't randomly release | |
1511 | * mmap_offsets as userspace expects them to be persistent for the | |
1512 | * lifetime of the objects. The closest we can is to release the | |
1513 | * offsets on purgeable objects by truncating it and marking it purged, | |
1514 | * which prevents userspace from ever using that object again. | |
1515 | */ | |
1516 | i915_gem_purge(dev_priv, obj->base.size >> PAGE_SHIFT); | |
1517 | ret = drm_gem_create_mmap_offset(&obj->base); | |
1518 | if (ret != -ENOSPC) | |
1519 | goto out; | |
1520 | ||
1521 | i915_gem_shrink_all(dev_priv); | |
1522 | ret = drm_gem_create_mmap_offset(&obj->base); | |
1523 | out: | |
1524 | dev_priv->mm.shrinker_no_lock_stealing = false; | |
1525 | ||
1526 | return ret; | |
1527 | } | |
1528 | ||
1529 | static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj) | |
1530 | { | |
1531 | if (!obj->base.map_list.map) | |
1532 | return; | |
1533 | ||
1534 | drm_gem_free_mmap_offset(&obj->base); | |
1535 | } | |
1536 | ||
1537 | int | |
1538 | i915_gem_mmap_gtt(struct drm_file *file, | |
1539 | struct drm_device *dev, | |
1540 | uint32_t handle, | |
1541 | uint64_t *offset) | |
1542 | { | |
1543 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1544 | struct drm_i915_gem_object *obj; | |
1545 | int ret; | |
1546 | ||
1547 | ret = i915_mutex_lock_interruptible(dev); | |
1548 | if (ret) | |
1549 | return ret; | |
1550 | ||
1551 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle)); | |
1552 | if (&obj->base == NULL) { | |
1553 | ret = -ENOENT; | |
1554 | goto unlock; | |
1555 | } | |
1556 | ||
1557 | if (obj->base.size > dev_priv->gtt.mappable_end) { | |
1558 | ret = -E2BIG; | |
1559 | goto out; | |
1560 | } | |
1561 | ||
1562 | if (obj->madv != I915_MADV_WILLNEED) { | |
1563 | DRM_ERROR("Attempting to mmap a purgeable buffer\n"); | |
1564 | ret = -EINVAL; | |
1565 | goto out; | |
1566 | } | |
1567 | ||
1568 | ret = i915_gem_object_create_mmap_offset(obj); | |
1569 | if (ret) | |
1570 | goto out; | |
1571 | ||
1572 | *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT; | |
1573 | ||
1574 | out: | |
1575 | drm_gem_object_unreference(&obj->base); | |
1576 | unlock: | |
1577 | mutex_unlock(&dev->struct_mutex); | |
1578 | return ret; | |
1579 | } | |
1580 | ||
1581 | /** | |
1582 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing | |
1583 | * @dev: DRM device | |
1584 | * @data: GTT mapping ioctl data | |
1585 | * @file: GEM object info | |
1586 | * | |
1587 | * Simply returns the fake offset to userspace so it can mmap it. | |
1588 | * The mmap call will end up in drm_gem_mmap(), which will set things | |
1589 | * up so we can get faults in the handler above. | |
1590 | * | |
1591 | * The fault handler will take care of binding the object into the GTT | |
1592 | * (since it may have been evicted to make room for something), allocating | |
1593 | * a fence register, and mapping the appropriate aperture address into | |
1594 | * userspace. | |
1595 | */ | |
1596 | int | |
1597 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, | |
1598 | struct drm_file *file) | |
1599 | { | |
1600 | struct drm_i915_gem_mmap_gtt *args = data; | |
1601 | ||
1602 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); | |
1603 | } | |
1604 | ||
1605 | /* Immediately discard the backing storage */ | |
1606 | static void | |
1607 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) | |
1608 | { | |
1609 | struct inode *inode; | |
1610 | ||
1611 | i915_gem_object_free_mmap_offset(obj); | |
1612 | ||
1613 | if (obj->base.filp == NULL) | |
1614 | return; | |
1615 | ||
1616 | /* Our goal here is to return as much of the memory as | |
1617 | * is possible back to the system as we are called from OOM. | |
1618 | * To do this we must instruct the shmfs to drop all of its | |
1619 | * backing pages, *now*. | |
1620 | */ | |
1621 | inode = obj->base.filp->f_path.dentry->d_inode; | |
1622 | shmem_truncate_range(inode, 0, (loff_t)-1); | |
1623 | ||
1624 | obj->madv = __I915_MADV_PURGED; | |
1625 | } | |
1626 | ||
1627 | static inline int | |
1628 | i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj) | |
1629 | { | |
1630 | return obj->madv == I915_MADV_DONTNEED; | |
1631 | } | |
1632 | ||
1633 | static void | |
1634 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj) | |
1635 | { | |
1636 | int page_count = obj->base.size / PAGE_SIZE; | |
1637 | struct scatterlist *sg; | |
1638 | int ret, i; | |
1639 | ||
1640 | BUG_ON(obj->madv == __I915_MADV_PURGED); | |
1641 | ||
1642 | ret = i915_gem_object_set_to_cpu_domain(obj, true); | |
1643 | if (ret) { | |
1644 | /* In the event of a disaster, abandon all caches and | |
1645 | * hope for the best. | |
1646 | */ | |
1647 | WARN_ON(ret != -EIO); | |
1648 | i915_gem_clflush_object(obj); | |
1649 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
1650 | } | |
1651 | ||
1652 | if (i915_gem_object_needs_bit17_swizzle(obj)) | |
1653 | i915_gem_object_save_bit_17_swizzle(obj); | |
1654 | ||
1655 | if (obj->madv == I915_MADV_DONTNEED) | |
1656 | obj->dirty = 0; | |
1657 | ||
1658 | for_each_sg(obj->pages->sgl, sg, page_count, i) { | |
1659 | struct page *page = sg_page(sg); | |
1660 | ||
1661 | if (obj->dirty) | |
1662 | set_page_dirty(page); | |
1663 | ||
1664 | if (obj->madv == I915_MADV_WILLNEED) | |
1665 | mark_page_accessed(page); | |
1666 | ||
1667 | page_cache_release(page); | |
1668 | } | |
1669 | obj->dirty = 0; | |
1670 | ||
1671 | sg_free_table(obj->pages); | |
1672 | kfree(obj->pages); | |
1673 | } | |
1674 | ||
1675 | int | |
1676 | i915_gem_object_put_pages(struct drm_i915_gem_object *obj) | |
1677 | { | |
1678 | const struct drm_i915_gem_object_ops *ops = obj->ops; | |
1679 | ||
1680 | if (obj->pages == NULL) | |
1681 | return 0; | |
1682 | ||
1683 | BUG_ON(obj->gtt_space); | |
1684 | ||
1685 | if (obj->pages_pin_count) | |
1686 | return -EBUSY; | |
1687 | ||
1688 | /* ->put_pages might need to allocate memory for the bit17 swizzle | |
1689 | * array, hence protect them from being reaped by removing them from gtt | |
1690 | * lists early. */ | |
1691 | list_del(&obj->gtt_list); | |
1692 | ||
1693 | ops->put_pages(obj); | |
1694 | obj->pages = NULL; | |
1695 | ||
1696 | if (i915_gem_object_is_purgeable(obj)) | |
1697 | i915_gem_object_truncate(obj); | |
1698 | ||
1699 | return 0; | |
1700 | } | |
1701 | ||
1702 | static long | |
1703 | __i915_gem_shrink(struct drm_i915_private *dev_priv, long target, | |
1704 | bool purgeable_only) | |
1705 | { | |
1706 | struct drm_i915_gem_object *obj, *next; | |
1707 | long count = 0; | |
1708 | ||
1709 | list_for_each_entry_safe(obj, next, | |
1710 | &dev_priv->mm.unbound_list, | |
1711 | gtt_list) { | |
1712 | if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) && | |
1713 | i915_gem_object_put_pages(obj) == 0) { | |
1714 | count += obj->base.size >> PAGE_SHIFT; | |
1715 | if (count >= target) | |
1716 | return count; | |
1717 | } | |
1718 | } | |
1719 | ||
1720 | list_for_each_entry_safe(obj, next, | |
1721 | &dev_priv->mm.inactive_list, | |
1722 | mm_list) { | |
1723 | if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) && | |
1724 | i915_gem_object_unbind(obj) == 0 && | |
1725 | i915_gem_object_put_pages(obj) == 0) { | |
1726 | count += obj->base.size >> PAGE_SHIFT; | |
1727 | if (count >= target) | |
1728 | return count; | |
1729 | } | |
1730 | } | |
1731 | ||
1732 | return count; | |
1733 | } | |
1734 | ||
1735 | static long | |
1736 | i915_gem_purge(struct drm_i915_private *dev_priv, long target) | |
1737 | { | |
1738 | return __i915_gem_shrink(dev_priv, target, true); | |
1739 | } | |
1740 | ||
1741 | static void | |
1742 | i915_gem_shrink_all(struct drm_i915_private *dev_priv) | |
1743 | { | |
1744 | struct drm_i915_gem_object *obj, *next; | |
1745 | ||
1746 | i915_gem_evict_everything(dev_priv->dev); | |
1747 | ||
1748 | list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list, gtt_list) | |
1749 | i915_gem_object_put_pages(obj); | |
1750 | } | |
1751 | ||
1752 | static int | |
1753 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj) | |
1754 | { | |
1755 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
1756 | int page_count, i; | |
1757 | struct address_space *mapping; | |
1758 | struct sg_table *st; | |
1759 | struct scatterlist *sg; | |
1760 | struct page *page; | |
1761 | gfp_t gfp; | |
1762 | ||
1763 | /* Assert that the object is not currently in any GPU domain. As it | |
1764 | * wasn't in the GTT, there shouldn't be any way it could have been in | |
1765 | * a GPU cache | |
1766 | */ | |
1767 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); | |
1768 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); | |
1769 | ||
1770 | st = kmalloc(sizeof(*st), GFP_KERNEL); | |
1771 | if (st == NULL) | |
1772 | return -ENOMEM; | |
1773 | ||
1774 | page_count = obj->base.size / PAGE_SIZE; | |
1775 | if (sg_alloc_table(st, page_count, GFP_KERNEL)) { | |
1776 | sg_free_table(st); | |
1777 | kfree(st); | |
1778 | return -ENOMEM; | |
1779 | } | |
1780 | ||
1781 | /* Get the list of pages out of our struct file. They'll be pinned | |
1782 | * at this point until we release them. | |
1783 | * | |
1784 | * Fail silently without starting the shrinker | |
1785 | */ | |
1786 | mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; | |
1787 | gfp = mapping_gfp_mask(mapping); | |
1788 | gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD; | |
1789 | gfp &= ~(__GFP_IO | __GFP_WAIT); | |
1790 | for_each_sg(st->sgl, sg, page_count, i) { | |
1791 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); | |
1792 | if (IS_ERR(page)) { | |
1793 | i915_gem_purge(dev_priv, page_count); | |
1794 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); | |
1795 | } | |
1796 | if (IS_ERR(page)) { | |
1797 | /* We've tried hard to allocate the memory by reaping | |
1798 | * our own buffer, now let the real VM do its job and | |
1799 | * go down in flames if truly OOM. | |
1800 | */ | |
1801 | gfp &= ~(__GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD); | |
1802 | gfp |= __GFP_IO | __GFP_WAIT; | |
1803 | ||
1804 | i915_gem_shrink_all(dev_priv); | |
1805 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); | |
1806 | if (IS_ERR(page)) | |
1807 | goto err_pages; | |
1808 | ||
1809 | gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD; | |
1810 | gfp &= ~(__GFP_IO | __GFP_WAIT); | |
1811 | } | |
1812 | ||
1813 | sg_set_page(sg, page, PAGE_SIZE, 0); | |
1814 | } | |
1815 | ||
1816 | obj->pages = st; | |
1817 | ||
1818 | if (i915_gem_object_needs_bit17_swizzle(obj)) | |
1819 | i915_gem_object_do_bit_17_swizzle(obj); | |
1820 | ||
1821 | return 0; | |
1822 | ||
1823 | err_pages: | |
1824 | for_each_sg(st->sgl, sg, i, page_count) | |
1825 | page_cache_release(sg_page(sg)); | |
1826 | sg_free_table(st); | |
1827 | kfree(st); | |
1828 | return PTR_ERR(page); | |
1829 | } | |
1830 | ||
1831 | /* Ensure that the associated pages are gathered from the backing storage | |
1832 | * and pinned into our object. i915_gem_object_get_pages() may be called | |
1833 | * multiple times before they are released by a single call to | |
1834 | * i915_gem_object_put_pages() - once the pages are no longer referenced | |
1835 | * either as a result of memory pressure (reaping pages under the shrinker) | |
1836 | * or as the object is itself released. | |
1837 | */ | |
1838 | int | |
1839 | i915_gem_object_get_pages(struct drm_i915_gem_object *obj) | |
1840 | { | |
1841 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
1842 | const struct drm_i915_gem_object_ops *ops = obj->ops; | |
1843 | int ret; | |
1844 | ||
1845 | if (obj->pages) | |
1846 | return 0; | |
1847 | ||
1848 | if (obj->madv != I915_MADV_WILLNEED) { | |
1849 | DRM_ERROR("Attempting to obtain a purgeable object\n"); | |
1850 | return -EINVAL; | |
1851 | } | |
1852 | ||
1853 | BUG_ON(obj->pages_pin_count); | |
1854 | ||
1855 | ret = ops->get_pages(obj); | |
1856 | if (ret) | |
1857 | return ret; | |
1858 | ||
1859 | list_add_tail(&obj->gtt_list, &dev_priv->mm.unbound_list); | |
1860 | return 0; | |
1861 | } | |
1862 | ||
1863 | void | |
1864 | i915_gem_object_move_to_active(struct drm_i915_gem_object *obj, | |
1865 | struct intel_ring_buffer *ring) | |
1866 | { | |
1867 | struct drm_device *dev = obj->base.dev; | |
1868 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1869 | u32 seqno = intel_ring_get_seqno(ring); | |
1870 | ||
1871 | BUG_ON(ring == NULL); | |
1872 | obj->ring = ring; | |
1873 | ||
1874 | /* Add a reference if we're newly entering the active list. */ | |
1875 | if (!obj->active) { | |
1876 | drm_gem_object_reference(&obj->base); | |
1877 | obj->active = 1; | |
1878 | } | |
1879 | ||
1880 | /* Move from whatever list we were on to the tail of execution. */ | |
1881 | list_move_tail(&obj->mm_list, &dev_priv->mm.active_list); | |
1882 | list_move_tail(&obj->ring_list, &ring->active_list); | |
1883 | ||
1884 | obj->last_read_seqno = seqno; | |
1885 | ||
1886 | if (obj->fenced_gpu_access) { | |
1887 | obj->last_fenced_seqno = seqno; | |
1888 | ||
1889 | /* Bump MRU to take account of the delayed flush */ | |
1890 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
1891 | struct drm_i915_fence_reg *reg; | |
1892 | ||
1893 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
1894 | list_move_tail(®->lru_list, | |
1895 | &dev_priv->mm.fence_list); | |
1896 | } | |
1897 | } | |
1898 | } | |
1899 | ||
1900 | static void | |
1901 | i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj) | |
1902 | { | |
1903 | struct drm_device *dev = obj->base.dev; | |
1904 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1905 | ||
1906 | BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS); | |
1907 | BUG_ON(!obj->active); | |
1908 | ||
1909 | list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list); | |
1910 | ||
1911 | list_del_init(&obj->ring_list); | |
1912 | obj->ring = NULL; | |
1913 | ||
1914 | obj->last_read_seqno = 0; | |
1915 | obj->last_write_seqno = 0; | |
1916 | obj->base.write_domain = 0; | |
1917 | ||
1918 | obj->last_fenced_seqno = 0; | |
1919 | obj->fenced_gpu_access = false; | |
1920 | ||
1921 | obj->active = 0; | |
1922 | drm_gem_object_unreference(&obj->base); | |
1923 | ||
1924 | WARN_ON(i915_verify_lists(dev)); | |
1925 | } | |
1926 | ||
1927 | static int | |
1928 | i915_gem_init_seqno(struct drm_device *dev, u32 seqno) | |
1929 | { | |
1930 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1931 | struct intel_ring_buffer *ring; | |
1932 | int ret, i, j; | |
1933 | ||
1934 | /* Carefully retire all requests without writing to the rings */ | |
1935 | for_each_ring(ring, dev_priv, i) { | |
1936 | ret = intel_ring_idle(ring); | |
1937 | if (ret) | |
1938 | return ret; | |
1939 | } | |
1940 | i915_gem_retire_requests(dev); | |
1941 | ||
1942 | /* Finally reset hw state */ | |
1943 | for_each_ring(ring, dev_priv, i) { | |
1944 | intel_ring_init_seqno(ring, seqno); | |
1945 | ||
1946 | for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++) | |
1947 | ring->sync_seqno[j] = 0; | |
1948 | } | |
1949 | ||
1950 | return 0; | |
1951 | } | |
1952 | ||
1953 | int i915_gem_set_seqno(struct drm_device *dev, u32 seqno) | |
1954 | { | |
1955 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1956 | int ret; | |
1957 | ||
1958 | if (seqno == 0) | |
1959 | return -EINVAL; | |
1960 | ||
1961 | /* HWS page needs to be set less than what we | |
1962 | * will inject to ring | |
1963 | */ | |
1964 | ret = i915_gem_init_seqno(dev, seqno - 1); | |
1965 | if (ret) | |
1966 | return ret; | |
1967 | ||
1968 | /* Carefully set the last_seqno value so that wrap | |
1969 | * detection still works | |
1970 | */ | |
1971 | dev_priv->next_seqno = seqno; | |
1972 | dev_priv->last_seqno = seqno - 1; | |
1973 | if (dev_priv->last_seqno == 0) | |
1974 | dev_priv->last_seqno--; | |
1975 | ||
1976 | return 0; | |
1977 | } | |
1978 | ||
1979 | int | |
1980 | i915_gem_get_seqno(struct drm_device *dev, u32 *seqno) | |
1981 | { | |
1982 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1983 | ||
1984 | /* reserve 0 for non-seqno */ | |
1985 | if (dev_priv->next_seqno == 0) { | |
1986 | int ret = i915_gem_init_seqno(dev, 0); | |
1987 | if (ret) | |
1988 | return ret; | |
1989 | ||
1990 | dev_priv->next_seqno = 1; | |
1991 | } | |
1992 | ||
1993 | *seqno = dev_priv->last_seqno = dev_priv->next_seqno++; | |
1994 | return 0; | |
1995 | } | |
1996 | ||
1997 | int | |
1998 | i915_add_request(struct intel_ring_buffer *ring, | |
1999 | struct drm_file *file, | |
2000 | u32 *out_seqno) | |
2001 | { | |
2002 | drm_i915_private_t *dev_priv = ring->dev->dev_private; | |
2003 | struct drm_i915_gem_request *request; | |
2004 | u32 request_ring_position; | |
2005 | int was_empty; | |
2006 | int ret; | |
2007 | ||
2008 | /* | |
2009 | * Emit any outstanding flushes - execbuf can fail to emit the flush | |
2010 | * after having emitted the batchbuffer command. Hence we need to fix | |
2011 | * things up similar to emitting the lazy request. The difference here | |
2012 | * is that the flush _must_ happen before the next request, no matter | |
2013 | * what. | |
2014 | */ | |
2015 | ret = intel_ring_flush_all_caches(ring); | |
2016 | if (ret) | |
2017 | return ret; | |
2018 | ||
2019 | request = kmalloc(sizeof(*request), GFP_KERNEL); | |
2020 | if (request == NULL) | |
2021 | return -ENOMEM; | |
2022 | ||
2023 | ||
2024 | /* Record the position of the start of the request so that | |
2025 | * should we detect the updated seqno part-way through the | |
2026 | * GPU processing the request, we never over-estimate the | |
2027 | * position of the head. | |
2028 | */ | |
2029 | request_ring_position = intel_ring_get_tail(ring); | |
2030 | ||
2031 | ret = ring->add_request(ring); | |
2032 | if (ret) { | |
2033 | kfree(request); | |
2034 | return ret; | |
2035 | } | |
2036 | ||
2037 | request->seqno = intel_ring_get_seqno(ring); | |
2038 | request->ring = ring; | |
2039 | request->tail = request_ring_position; | |
2040 | request->emitted_jiffies = jiffies; | |
2041 | was_empty = list_empty(&ring->request_list); | |
2042 | list_add_tail(&request->list, &ring->request_list); | |
2043 | request->file_priv = NULL; | |
2044 | ||
2045 | if (file) { | |
2046 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
2047 | ||
2048 | spin_lock(&file_priv->mm.lock); | |
2049 | request->file_priv = file_priv; | |
2050 | list_add_tail(&request->client_list, | |
2051 | &file_priv->mm.request_list); | |
2052 | spin_unlock(&file_priv->mm.lock); | |
2053 | } | |
2054 | ||
2055 | trace_i915_gem_request_add(ring, request->seqno); | |
2056 | ring->outstanding_lazy_request = 0; | |
2057 | ||
2058 | if (!dev_priv->mm.suspended) { | |
2059 | if (i915_enable_hangcheck) { | |
2060 | mod_timer(&dev_priv->gpu_error.hangcheck_timer, | |
2061 | round_jiffies_up(jiffies + DRM_I915_HANGCHECK_JIFFIES)); | |
2062 | } | |
2063 | if (was_empty) { | |
2064 | queue_delayed_work(dev_priv->wq, | |
2065 | &dev_priv->mm.retire_work, | |
2066 | round_jiffies_up_relative(HZ)); | |
2067 | intel_mark_busy(dev_priv->dev); | |
2068 | } | |
2069 | } | |
2070 | ||
2071 | if (out_seqno) | |
2072 | *out_seqno = request->seqno; | |
2073 | return 0; | |
2074 | } | |
2075 | ||
2076 | static inline void | |
2077 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) | |
2078 | { | |
2079 | struct drm_i915_file_private *file_priv = request->file_priv; | |
2080 | ||
2081 | if (!file_priv) | |
2082 | return; | |
2083 | ||
2084 | spin_lock(&file_priv->mm.lock); | |
2085 | if (request->file_priv) { | |
2086 | list_del(&request->client_list); | |
2087 | request->file_priv = NULL; | |
2088 | } | |
2089 | spin_unlock(&file_priv->mm.lock); | |
2090 | } | |
2091 | ||
2092 | static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv, | |
2093 | struct intel_ring_buffer *ring) | |
2094 | { | |
2095 | while (!list_empty(&ring->request_list)) { | |
2096 | struct drm_i915_gem_request *request; | |
2097 | ||
2098 | request = list_first_entry(&ring->request_list, | |
2099 | struct drm_i915_gem_request, | |
2100 | list); | |
2101 | ||
2102 | list_del(&request->list); | |
2103 | i915_gem_request_remove_from_client(request); | |
2104 | kfree(request); | |
2105 | } | |
2106 | ||
2107 | while (!list_empty(&ring->active_list)) { | |
2108 | struct drm_i915_gem_object *obj; | |
2109 | ||
2110 | obj = list_first_entry(&ring->active_list, | |
2111 | struct drm_i915_gem_object, | |
2112 | ring_list); | |
2113 | ||
2114 | i915_gem_object_move_to_inactive(obj); | |
2115 | } | |
2116 | } | |
2117 | ||
2118 | static void i915_gem_reset_fences(struct drm_device *dev) | |
2119 | { | |
2120 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2121 | int i; | |
2122 | ||
2123 | for (i = 0; i < dev_priv->num_fence_regs; i++) { | |
2124 | struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i]; | |
2125 | ||
2126 | i915_gem_write_fence(dev, i, NULL); | |
2127 | ||
2128 | if (reg->obj) | |
2129 | i915_gem_object_fence_lost(reg->obj); | |
2130 | ||
2131 | reg->pin_count = 0; | |
2132 | reg->obj = NULL; | |
2133 | INIT_LIST_HEAD(®->lru_list); | |
2134 | } | |
2135 | ||
2136 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); | |
2137 | } | |
2138 | ||
2139 | void i915_gem_reset(struct drm_device *dev) | |
2140 | { | |
2141 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2142 | struct drm_i915_gem_object *obj; | |
2143 | struct intel_ring_buffer *ring; | |
2144 | int i; | |
2145 | ||
2146 | for_each_ring(ring, dev_priv, i) | |
2147 | i915_gem_reset_ring_lists(dev_priv, ring); | |
2148 | ||
2149 | /* Move everything out of the GPU domains to ensure we do any | |
2150 | * necessary invalidation upon reuse. | |
2151 | */ | |
2152 | list_for_each_entry(obj, | |
2153 | &dev_priv->mm.inactive_list, | |
2154 | mm_list) | |
2155 | { | |
2156 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; | |
2157 | } | |
2158 | ||
2159 | /* The fence registers are invalidated so clear them out */ | |
2160 | i915_gem_reset_fences(dev); | |
2161 | } | |
2162 | ||
2163 | /** | |
2164 | * This function clears the request list as sequence numbers are passed. | |
2165 | */ | |
2166 | void | |
2167 | i915_gem_retire_requests_ring(struct intel_ring_buffer *ring) | |
2168 | { | |
2169 | uint32_t seqno; | |
2170 | ||
2171 | if (list_empty(&ring->request_list)) | |
2172 | return; | |
2173 | ||
2174 | WARN_ON(i915_verify_lists(ring->dev)); | |
2175 | ||
2176 | seqno = ring->get_seqno(ring, true); | |
2177 | ||
2178 | while (!list_empty(&ring->request_list)) { | |
2179 | struct drm_i915_gem_request *request; | |
2180 | ||
2181 | request = list_first_entry(&ring->request_list, | |
2182 | struct drm_i915_gem_request, | |
2183 | list); | |
2184 | ||
2185 | if (!i915_seqno_passed(seqno, request->seqno)) | |
2186 | break; | |
2187 | ||
2188 | trace_i915_gem_request_retire(ring, request->seqno); | |
2189 | /* We know the GPU must have read the request to have | |
2190 | * sent us the seqno + interrupt, so use the position | |
2191 | * of tail of the request to update the last known position | |
2192 | * of the GPU head. | |
2193 | */ | |
2194 | ring->last_retired_head = request->tail; | |
2195 | ||
2196 | list_del(&request->list); | |
2197 | i915_gem_request_remove_from_client(request); | |
2198 | kfree(request); | |
2199 | } | |
2200 | ||
2201 | /* Move any buffers on the active list that are no longer referenced | |
2202 | * by the ringbuffer to the flushing/inactive lists as appropriate. | |
2203 | */ | |
2204 | while (!list_empty(&ring->active_list)) { | |
2205 | struct drm_i915_gem_object *obj; | |
2206 | ||
2207 | obj = list_first_entry(&ring->active_list, | |
2208 | struct drm_i915_gem_object, | |
2209 | ring_list); | |
2210 | ||
2211 | if (!i915_seqno_passed(seqno, obj->last_read_seqno)) | |
2212 | break; | |
2213 | ||
2214 | i915_gem_object_move_to_inactive(obj); | |
2215 | } | |
2216 | ||
2217 | if (unlikely(ring->trace_irq_seqno && | |
2218 | i915_seqno_passed(seqno, ring->trace_irq_seqno))) { | |
2219 | ring->irq_put(ring); | |
2220 | ring->trace_irq_seqno = 0; | |
2221 | } | |
2222 | ||
2223 | WARN_ON(i915_verify_lists(ring->dev)); | |
2224 | } | |
2225 | ||
2226 | void | |
2227 | i915_gem_retire_requests(struct drm_device *dev) | |
2228 | { | |
2229 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2230 | struct intel_ring_buffer *ring; | |
2231 | int i; | |
2232 | ||
2233 | for_each_ring(ring, dev_priv, i) | |
2234 | i915_gem_retire_requests_ring(ring); | |
2235 | } | |
2236 | ||
2237 | static void | |
2238 | i915_gem_retire_work_handler(struct work_struct *work) | |
2239 | { | |
2240 | drm_i915_private_t *dev_priv; | |
2241 | struct drm_device *dev; | |
2242 | struct intel_ring_buffer *ring; | |
2243 | bool idle; | |
2244 | int i; | |
2245 | ||
2246 | dev_priv = container_of(work, drm_i915_private_t, | |
2247 | mm.retire_work.work); | |
2248 | dev = dev_priv->dev; | |
2249 | ||
2250 | /* Come back later if the device is busy... */ | |
2251 | if (!mutex_trylock(&dev->struct_mutex)) { | |
2252 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, | |
2253 | round_jiffies_up_relative(HZ)); | |
2254 | return; | |
2255 | } | |
2256 | ||
2257 | i915_gem_retire_requests(dev); | |
2258 | ||
2259 | /* Send a periodic flush down the ring so we don't hold onto GEM | |
2260 | * objects indefinitely. | |
2261 | */ | |
2262 | idle = true; | |
2263 | for_each_ring(ring, dev_priv, i) { | |
2264 | if (ring->gpu_caches_dirty) | |
2265 | i915_add_request(ring, NULL, NULL); | |
2266 | ||
2267 | idle &= list_empty(&ring->request_list); | |
2268 | } | |
2269 | ||
2270 | if (!dev_priv->mm.suspended && !idle) | |
2271 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, | |
2272 | round_jiffies_up_relative(HZ)); | |
2273 | if (idle) | |
2274 | intel_mark_idle(dev); | |
2275 | ||
2276 | mutex_unlock(&dev->struct_mutex); | |
2277 | } | |
2278 | ||
2279 | /** | |
2280 | * Ensures that an object will eventually get non-busy by flushing any required | |
2281 | * write domains, emitting any outstanding lazy request and retiring and | |
2282 | * completed requests. | |
2283 | */ | |
2284 | static int | |
2285 | i915_gem_object_flush_active(struct drm_i915_gem_object *obj) | |
2286 | { | |
2287 | int ret; | |
2288 | ||
2289 | if (obj->active) { | |
2290 | ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno); | |
2291 | if (ret) | |
2292 | return ret; | |
2293 | ||
2294 | i915_gem_retire_requests_ring(obj->ring); | |
2295 | } | |
2296 | ||
2297 | return 0; | |
2298 | } | |
2299 | ||
2300 | /** | |
2301 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT | |
2302 | * @DRM_IOCTL_ARGS: standard ioctl arguments | |
2303 | * | |
2304 | * Returns 0 if successful, else an error is returned with the remaining time in | |
2305 | * the timeout parameter. | |
2306 | * -ETIME: object is still busy after timeout | |
2307 | * -ERESTARTSYS: signal interrupted the wait | |
2308 | * -ENONENT: object doesn't exist | |
2309 | * Also possible, but rare: | |
2310 | * -EAGAIN: GPU wedged | |
2311 | * -ENOMEM: damn | |
2312 | * -ENODEV: Internal IRQ fail | |
2313 | * -E?: The add request failed | |
2314 | * | |
2315 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any | |
2316 | * non-zero timeout parameter the wait ioctl will wait for the given number of | |
2317 | * nanoseconds on an object becoming unbusy. Since the wait itself does so | |
2318 | * without holding struct_mutex the object may become re-busied before this | |
2319 | * function completes. A similar but shorter * race condition exists in the busy | |
2320 | * ioctl | |
2321 | */ | |
2322 | int | |
2323 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
2324 | { | |
2325 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2326 | struct drm_i915_gem_wait *args = data; | |
2327 | struct drm_i915_gem_object *obj; | |
2328 | struct intel_ring_buffer *ring = NULL; | |
2329 | struct timespec timeout_stack, *timeout = NULL; | |
2330 | unsigned reset_counter; | |
2331 | u32 seqno = 0; | |
2332 | int ret = 0; | |
2333 | ||
2334 | if (args->timeout_ns >= 0) { | |
2335 | timeout_stack = ns_to_timespec(args->timeout_ns); | |
2336 | timeout = &timeout_stack; | |
2337 | } | |
2338 | ||
2339 | ret = i915_mutex_lock_interruptible(dev); | |
2340 | if (ret) | |
2341 | return ret; | |
2342 | ||
2343 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle)); | |
2344 | if (&obj->base == NULL) { | |
2345 | mutex_unlock(&dev->struct_mutex); | |
2346 | return -ENOENT; | |
2347 | } | |
2348 | ||
2349 | /* Need to make sure the object gets inactive eventually. */ | |
2350 | ret = i915_gem_object_flush_active(obj); | |
2351 | if (ret) | |
2352 | goto out; | |
2353 | ||
2354 | if (obj->active) { | |
2355 | seqno = obj->last_read_seqno; | |
2356 | ring = obj->ring; | |
2357 | } | |
2358 | ||
2359 | if (seqno == 0) | |
2360 | goto out; | |
2361 | ||
2362 | /* Do this after OLR check to make sure we make forward progress polling | |
2363 | * on this IOCTL with a 0 timeout (like busy ioctl) | |
2364 | */ | |
2365 | if (!args->timeout_ns) { | |
2366 | ret = -ETIME; | |
2367 | goto out; | |
2368 | } | |
2369 | ||
2370 | drm_gem_object_unreference(&obj->base); | |
2371 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); | |
2372 | mutex_unlock(&dev->struct_mutex); | |
2373 | ||
2374 | ret = __wait_seqno(ring, seqno, reset_counter, true, timeout); | |
2375 | if (timeout) { | |
2376 | WARN_ON(!timespec_valid(timeout)); | |
2377 | args->timeout_ns = timespec_to_ns(timeout); | |
2378 | } | |
2379 | return ret; | |
2380 | ||
2381 | out: | |
2382 | drm_gem_object_unreference(&obj->base); | |
2383 | mutex_unlock(&dev->struct_mutex); | |
2384 | return ret; | |
2385 | } | |
2386 | ||
2387 | /** | |
2388 | * i915_gem_object_sync - sync an object to a ring. | |
2389 | * | |
2390 | * @obj: object which may be in use on another ring. | |
2391 | * @to: ring we wish to use the object on. May be NULL. | |
2392 | * | |
2393 | * This code is meant to abstract object synchronization with the GPU. | |
2394 | * Calling with NULL implies synchronizing the object with the CPU | |
2395 | * rather than a particular GPU ring. | |
2396 | * | |
2397 | * Returns 0 if successful, else propagates up the lower layer error. | |
2398 | */ | |
2399 | int | |
2400 | i915_gem_object_sync(struct drm_i915_gem_object *obj, | |
2401 | struct intel_ring_buffer *to) | |
2402 | { | |
2403 | struct intel_ring_buffer *from = obj->ring; | |
2404 | u32 seqno; | |
2405 | int ret, idx; | |
2406 | ||
2407 | if (from == NULL || to == from) | |
2408 | return 0; | |
2409 | ||
2410 | if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev)) | |
2411 | return i915_gem_object_wait_rendering(obj, false); | |
2412 | ||
2413 | idx = intel_ring_sync_index(from, to); | |
2414 | ||
2415 | seqno = obj->last_read_seqno; | |
2416 | if (seqno <= from->sync_seqno[idx]) | |
2417 | return 0; | |
2418 | ||
2419 | ret = i915_gem_check_olr(obj->ring, seqno); | |
2420 | if (ret) | |
2421 | return ret; | |
2422 | ||
2423 | ret = to->sync_to(to, from, seqno); | |
2424 | if (!ret) | |
2425 | /* We use last_read_seqno because sync_to() | |
2426 | * might have just caused seqno wrap under | |
2427 | * the radar. | |
2428 | */ | |
2429 | from->sync_seqno[idx] = obj->last_read_seqno; | |
2430 | ||
2431 | return ret; | |
2432 | } | |
2433 | ||
2434 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) | |
2435 | { | |
2436 | u32 old_write_domain, old_read_domains; | |
2437 | ||
2438 | /* Force a pagefault for domain tracking on next user access */ | |
2439 | i915_gem_release_mmap(obj); | |
2440 | ||
2441 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) | |
2442 | return; | |
2443 | ||
2444 | /* Wait for any direct GTT access to complete */ | |
2445 | mb(); | |
2446 | ||
2447 | old_read_domains = obj->base.read_domains; | |
2448 | old_write_domain = obj->base.write_domain; | |
2449 | ||
2450 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; | |
2451 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; | |
2452 | ||
2453 | trace_i915_gem_object_change_domain(obj, | |
2454 | old_read_domains, | |
2455 | old_write_domain); | |
2456 | } | |
2457 | ||
2458 | /** | |
2459 | * Unbinds an object from the GTT aperture. | |
2460 | */ | |
2461 | int | |
2462 | i915_gem_object_unbind(struct drm_i915_gem_object *obj) | |
2463 | { | |
2464 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; | |
2465 | int ret; | |
2466 | ||
2467 | if (obj->gtt_space == NULL) | |
2468 | return 0; | |
2469 | ||
2470 | if (obj->pin_count) | |
2471 | return -EBUSY; | |
2472 | ||
2473 | BUG_ON(obj->pages == NULL); | |
2474 | ||
2475 | ret = i915_gem_object_finish_gpu(obj); | |
2476 | if (ret) | |
2477 | return ret; | |
2478 | /* Continue on if we fail due to EIO, the GPU is hung so we | |
2479 | * should be safe and we need to cleanup or else we might | |
2480 | * cause memory corruption through use-after-free. | |
2481 | */ | |
2482 | ||
2483 | i915_gem_object_finish_gtt(obj); | |
2484 | ||
2485 | /* release the fence reg _after_ flushing */ | |
2486 | ret = i915_gem_object_put_fence(obj); | |
2487 | if (ret) | |
2488 | return ret; | |
2489 | ||
2490 | trace_i915_gem_object_unbind(obj); | |
2491 | ||
2492 | if (obj->has_global_gtt_mapping) | |
2493 | i915_gem_gtt_unbind_object(obj); | |
2494 | if (obj->has_aliasing_ppgtt_mapping) { | |
2495 | i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj); | |
2496 | obj->has_aliasing_ppgtt_mapping = 0; | |
2497 | } | |
2498 | i915_gem_gtt_finish_object(obj); | |
2499 | ||
2500 | list_del(&obj->mm_list); | |
2501 | list_move_tail(&obj->gtt_list, &dev_priv->mm.unbound_list); | |
2502 | /* Avoid an unnecessary call to unbind on rebind. */ | |
2503 | obj->map_and_fenceable = true; | |
2504 | ||
2505 | drm_mm_put_block(obj->gtt_space); | |
2506 | obj->gtt_space = NULL; | |
2507 | obj->gtt_offset = 0; | |
2508 | ||
2509 | return 0; | |
2510 | } | |
2511 | ||
2512 | int i915_gpu_idle(struct drm_device *dev) | |
2513 | { | |
2514 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2515 | struct intel_ring_buffer *ring; | |
2516 | int ret, i; | |
2517 | ||
2518 | /* Flush everything onto the inactive list. */ | |
2519 | for_each_ring(ring, dev_priv, i) { | |
2520 | ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID); | |
2521 | if (ret) | |
2522 | return ret; | |
2523 | ||
2524 | ret = intel_ring_idle(ring); | |
2525 | if (ret) | |
2526 | return ret; | |
2527 | } | |
2528 | ||
2529 | return 0; | |
2530 | } | |
2531 | ||
2532 | static void i965_write_fence_reg(struct drm_device *dev, int reg, | |
2533 | struct drm_i915_gem_object *obj) | |
2534 | { | |
2535 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2536 | int fence_reg; | |
2537 | int fence_pitch_shift; | |
2538 | uint64_t val; | |
2539 | ||
2540 | if (INTEL_INFO(dev)->gen >= 6) { | |
2541 | fence_reg = FENCE_REG_SANDYBRIDGE_0; | |
2542 | fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT; | |
2543 | } else { | |
2544 | fence_reg = FENCE_REG_965_0; | |
2545 | fence_pitch_shift = I965_FENCE_PITCH_SHIFT; | |
2546 | } | |
2547 | ||
2548 | if (obj) { | |
2549 | u32 size = obj->gtt_space->size; | |
2550 | ||
2551 | val = (uint64_t)((obj->gtt_offset + size - 4096) & | |
2552 | 0xfffff000) << 32; | |
2553 | val |= obj->gtt_offset & 0xfffff000; | |
2554 | val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift; | |
2555 | if (obj->tiling_mode == I915_TILING_Y) | |
2556 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; | |
2557 | val |= I965_FENCE_REG_VALID; | |
2558 | } else | |
2559 | val = 0; | |
2560 | ||
2561 | fence_reg += reg * 8; | |
2562 | I915_WRITE64(fence_reg, val); | |
2563 | POSTING_READ(fence_reg); | |
2564 | } | |
2565 | ||
2566 | static void i915_write_fence_reg(struct drm_device *dev, int reg, | |
2567 | struct drm_i915_gem_object *obj) | |
2568 | { | |
2569 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2570 | u32 val; | |
2571 | ||
2572 | if (obj) { | |
2573 | u32 size = obj->gtt_space->size; | |
2574 | int pitch_val; | |
2575 | int tile_width; | |
2576 | ||
2577 | WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) || | |
2578 | (size & -size) != size || | |
2579 | (obj->gtt_offset & (size - 1)), | |
2580 | "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n", | |
2581 | obj->gtt_offset, obj->map_and_fenceable, size); | |
2582 | ||
2583 | if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)) | |
2584 | tile_width = 128; | |
2585 | else | |
2586 | tile_width = 512; | |
2587 | ||
2588 | /* Note: pitch better be a power of two tile widths */ | |
2589 | pitch_val = obj->stride / tile_width; | |
2590 | pitch_val = ffs(pitch_val) - 1; | |
2591 | ||
2592 | val = obj->gtt_offset; | |
2593 | if (obj->tiling_mode == I915_TILING_Y) | |
2594 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
2595 | val |= I915_FENCE_SIZE_BITS(size); | |
2596 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
2597 | val |= I830_FENCE_REG_VALID; | |
2598 | } else | |
2599 | val = 0; | |
2600 | ||
2601 | if (reg < 8) | |
2602 | reg = FENCE_REG_830_0 + reg * 4; | |
2603 | else | |
2604 | reg = FENCE_REG_945_8 + (reg - 8) * 4; | |
2605 | ||
2606 | I915_WRITE(reg, val); | |
2607 | POSTING_READ(reg); | |
2608 | } | |
2609 | ||
2610 | static void i830_write_fence_reg(struct drm_device *dev, int reg, | |
2611 | struct drm_i915_gem_object *obj) | |
2612 | { | |
2613 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2614 | uint32_t val; | |
2615 | ||
2616 | if (obj) { | |
2617 | u32 size = obj->gtt_space->size; | |
2618 | uint32_t pitch_val; | |
2619 | ||
2620 | WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) || | |
2621 | (size & -size) != size || | |
2622 | (obj->gtt_offset & (size - 1)), | |
2623 | "object 0x%08x not 512K or pot-size 0x%08x aligned\n", | |
2624 | obj->gtt_offset, size); | |
2625 | ||
2626 | pitch_val = obj->stride / 128; | |
2627 | pitch_val = ffs(pitch_val) - 1; | |
2628 | ||
2629 | val = obj->gtt_offset; | |
2630 | if (obj->tiling_mode == I915_TILING_Y) | |
2631 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; | |
2632 | val |= I830_FENCE_SIZE_BITS(size); | |
2633 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; | |
2634 | val |= I830_FENCE_REG_VALID; | |
2635 | } else | |
2636 | val = 0; | |
2637 | ||
2638 | I915_WRITE(FENCE_REG_830_0 + reg * 4, val); | |
2639 | POSTING_READ(FENCE_REG_830_0 + reg * 4); | |
2640 | } | |
2641 | ||
2642 | inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj) | |
2643 | { | |
2644 | return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT; | |
2645 | } | |
2646 | ||
2647 | static void i915_gem_write_fence(struct drm_device *dev, int reg, | |
2648 | struct drm_i915_gem_object *obj) | |
2649 | { | |
2650 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2651 | ||
2652 | /* Ensure that all CPU reads are completed before installing a fence | |
2653 | * and all writes before removing the fence. | |
2654 | */ | |
2655 | if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj)) | |
2656 | mb(); | |
2657 | ||
2658 | switch (INTEL_INFO(dev)->gen) { | |
2659 | case 7: | |
2660 | case 6: | |
2661 | case 5: | |
2662 | case 4: i965_write_fence_reg(dev, reg, obj); break; | |
2663 | case 3: i915_write_fence_reg(dev, reg, obj); break; | |
2664 | case 2: i830_write_fence_reg(dev, reg, obj); break; | |
2665 | default: BUG(); | |
2666 | } | |
2667 | ||
2668 | /* And similarly be paranoid that no direct access to this region | |
2669 | * is reordered to before the fence is installed. | |
2670 | */ | |
2671 | if (i915_gem_object_needs_mb(obj)) | |
2672 | mb(); | |
2673 | } | |
2674 | ||
2675 | static inline int fence_number(struct drm_i915_private *dev_priv, | |
2676 | struct drm_i915_fence_reg *fence) | |
2677 | { | |
2678 | return fence - dev_priv->fence_regs; | |
2679 | } | |
2680 | ||
2681 | static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj, | |
2682 | struct drm_i915_fence_reg *fence, | |
2683 | bool enable) | |
2684 | { | |
2685 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
2686 | int reg = fence_number(dev_priv, fence); | |
2687 | ||
2688 | i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL); | |
2689 | ||
2690 | if (enable) { | |
2691 | obj->fence_reg = reg; | |
2692 | fence->obj = obj; | |
2693 | list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list); | |
2694 | } else { | |
2695 | obj->fence_reg = I915_FENCE_REG_NONE; | |
2696 | fence->obj = NULL; | |
2697 | list_del_init(&fence->lru_list); | |
2698 | } | |
2699 | } | |
2700 | ||
2701 | static int | |
2702 | i915_gem_object_wait_fence(struct drm_i915_gem_object *obj) | |
2703 | { | |
2704 | if (obj->last_fenced_seqno) { | |
2705 | int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno); | |
2706 | if (ret) | |
2707 | return ret; | |
2708 | ||
2709 | obj->last_fenced_seqno = 0; | |
2710 | } | |
2711 | ||
2712 | obj->fenced_gpu_access = false; | |
2713 | return 0; | |
2714 | } | |
2715 | ||
2716 | int | |
2717 | i915_gem_object_put_fence(struct drm_i915_gem_object *obj) | |
2718 | { | |
2719 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
2720 | int ret; | |
2721 | ||
2722 | ret = i915_gem_object_wait_fence(obj); | |
2723 | if (ret) | |
2724 | return ret; | |
2725 | ||
2726 | if (obj->fence_reg == I915_FENCE_REG_NONE) | |
2727 | return 0; | |
2728 | ||
2729 | i915_gem_object_update_fence(obj, | |
2730 | &dev_priv->fence_regs[obj->fence_reg], | |
2731 | false); | |
2732 | i915_gem_object_fence_lost(obj); | |
2733 | ||
2734 | return 0; | |
2735 | } | |
2736 | ||
2737 | static struct drm_i915_fence_reg * | |
2738 | i915_find_fence_reg(struct drm_device *dev) | |
2739 | { | |
2740 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2741 | struct drm_i915_fence_reg *reg, *avail; | |
2742 | int i; | |
2743 | ||
2744 | /* First try to find a free reg */ | |
2745 | avail = NULL; | |
2746 | for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) { | |
2747 | reg = &dev_priv->fence_regs[i]; | |
2748 | if (!reg->obj) | |
2749 | return reg; | |
2750 | ||
2751 | if (!reg->pin_count) | |
2752 | avail = reg; | |
2753 | } | |
2754 | ||
2755 | if (avail == NULL) | |
2756 | return NULL; | |
2757 | ||
2758 | /* None available, try to steal one or wait for a user to finish */ | |
2759 | list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) { | |
2760 | if (reg->pin_count) | |
2761 | continue; | |
2762 | ||
2763 | return reg; | |
2764 | } | |
2765 | ||
2766 | return NULL; | |
2767 | } | |
2768 | ||
2769 | /** | |
2770 | * i915_gem_object_get_fence - set up fencing for an object | |
2771 | * @obj: object to map through a fence reg | |
2772 | * | |
2773 | * When mapping objects through the GTT, userspace wants to be able to write | |
2774 | * to them without having to worry about swizzling if the object is tiled. | |
2775 | * This function walks the fence regs looking for a free one for @obj, | |
2776 | * stealing one if it can't find any. | |
2777 | * | |
2778 | * It then sets up the reg based on the object's properties: address, pitch | |
2779 | * and tiling format. | |
2780 | * | |
2781 | * For an untiled surface, this removes any existing fence. | |
2782 | */ | |
2783 | int | |
2784 | i915_gem_object_get_fence(struct drm_i915_gem_object *obj) | |
2785 | { | |
2786 | struct drm_device *dev = obj->base.dev; | |
2787 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2788 | bool enable = obj->tiling_mode != I915_TILING_NONE; | |
2789 | struct drm_i915_fence_reg *reg; | |
2790 | int ret; | |
2791 | ||
2792 | /* Have we updated the tiling parameters upon the object and so | |
2793 | * will need to serialise the write to the associated fence register? | |
2794 | */ | |
2795 | if (obj->fence_dirty) { | |
2796 | ret = i915_gem_object_wait_fence(obj); | |
2797 | if (ret) | |
2798 | return ret; | |
2799 | } | |
2800 | ||
2801 | /* Just update our place in the LRU if our fence is getting reused. */ | |
2802 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
2803 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
2804 | if (!obj->fence_dirty) { | |
2805 | list_move_tail(®->lru_list, | |
2806 | &dev_priv->mm.fence_list); | |
2807 | return 0; | |
2808 | } | |
2809 | } else if (enable) { | |
2810 | reg = i915_find_fence_reg(dev); | |
2811 | if (reg == NULL) | |
2812 | return -EDEADLK; | |
2813 | ||
2814 | if (reg->obj) { | |
2815 | struct drm_i915_gem_object *old = reg->obj; | |
2816 | ||
2817 | ret = i915_gem_object_wait_fence(old); | |
2818 | if (ret) | |
2819 | return ret; | |
2820 | ||
2821 | i915_gem_object_fence_lost(old); | |
2822 | } | |
2823 | } else | |
2824 | return 0; | |
2825 | ||
2826 | i915_gem_object_update_fence(obj, reg, enable); | |
2827 | obj->fence_dirty = false; | |
2828 | ||
2829 | return 0; | |
2830 | } | |
2831 | ||
2832 | static bool i915_gem_valid_gtt_space(struct drm_device *dev, | |
2833 | struct drm_mm_node *gtt_space, | |
2834 | unsigned long cache_level) | |
2835 | { | |
2836 | struct drm_mm_node *other; | |
2837 | ||
2838 | /* On non-LLC machines we have to be careful when putting differing | |
2839 | * types of snoopable memory together to avoid the prefetcher | |
2840 | * crossing memory domains and dying. | |
2841 | */ | |
2842 | if (HAS_LLC(dev)) | |
2843 | return true; | |
2844 | ||
2845 | if (gtt_space == NULL) | |
2846 | return true; | |
2847 | ||
2848 | if (list_empty(>t_space->node_list)) | |
2849 | return true; | |
2850 | ||
2851 | other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list); | |
2852 | if (other->allocated && !other->hole_follows && other->color != cache_level) | |
2853 | return false; | |
2854 | ||
2855 | other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list); | |
2856 | if (other->allocated && !gtt_space->hole_follows && other->color != cache_level) | |
2857 | return false; | |
2858 | ||
2859 | return true; | |
2860 | } | |
2861 | ||
2862 | static void i915_gem_verify_gtt(struct drm_device *dev) | |
2863 | { | |
2864 | #if WATCH_GTT | |
2865 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2866 | struct drm_i915_gem_object *obj; | |
2867 | int err = 0; | |
2868 | ||
2869 | list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list) { | |
2870 | if (obj->gtt_space == NULL) { | |
2871 | printk(KERN_ERR "object found on GTT list with no space reserved\n"); | |
2872 | err++; | |
2873 | continue; | |
2874 | } | |
2875 | ||
2876 | if (obj->cache_level != obj->gtt_space->color) { | |
2877 | printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n", | |
2878 | obj->gtt_space->start, | |
2879 | obj->gtt_space->start + obj->gtt_space->size, | |
2880 | obj->cache_level, | |
2881 | obj->gtt_space->color); | |
2882 | err++; | |
2883 | continue; | |
2884 | } | |
2885 | ||
2886 | if (!i915_gem_valid_gtt_space(dev, | |
2887 | obj->gtt_space, | |
2888 | obj->cache_level)) { | |
2889 | printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n", | |
2890 | obj->gtt_space->start, | |
2891 | obj->gtt_space->start + obj->gtt_space->size, | |
2892 | obj->cache_level); | |
2893 | err++; | |
2894 | continue; | |
2895 | } | |
2896 | } | |
2897 | ||
2898 | WARN_ON(err); | |
2899 | #endif | |
2900 | } | |
2901 | ||
2902 | /** | |
2903 | * Finds free space in the GTT aperture and binds the object there. | |
2904 | */ | |
2905 | static int | |
2906 | i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj, | |
2907 | unsigned alignment, | |
2908 | bool map_and_fenceable, | |
2909 | bool nonblocking) | |
2910 | { | |
2911 | struct drm_device *dev = obj->base.dev; | |
2912 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2913 | struct drm_mm_node *node; | |
2914 | u32 size, fence_size, fence_alignment, unfenced_alignment; | |
2915 | bool mappable, fenceable; | |
2916 | int ret; | |
2917 | ||
2918 | fence_size = i915_gem_get_gtt_size(dev, | |
2919 | obj->base.size, | |
2920 | obj->tiling_mode); | |
2921 | fence_alignment = i915_gem_get_gtt_alignment(dev, | |
2922 | obj->base.size, | |
2923 | obj->tiling_mode, true); | |
2924 | unfenced_alignment = | |
2925 | i915_gem_get_gtt_alignment(dev, | |
2926 | obj->base.size, | |
2927 | obj->tiling_mode, false); | |
2928 | ||
2929 | if (alignment == 0) | |
2930 | alignment = map_and_fenceable ? fence_alignment : | |
2931 | unfenced_alignment; | |
2932 | if (map_and_fenceable && alignment & (fence_alignment - 1)) { | |
2933 | DRM_ERROR("Invalid object alignment requested %u\n", alignment); | |
2934 | return -EINVAL; | |
2935 | } | |
2936 | ||
2937 | size = map_and_fenceable ? fence_size : obj->base.size; | |
2938 | ||
2939 | /* If the object is bigger than the entire aperture, reject it early | |
2940 | * before evicting everything in a vain attempt to find space. | |
2941 | */ | |
2942 | if (obj->base.size > | |
2943 | (map_and_fenceable ? dev_priv->gtt.mappable_end : dev_priv->gtt.total)) { | |
2944 | DRM_ERROR("Attempting to bind an object larger than the aperture\n"); | |
2945 | return -E2BIG; | |
2946 | } | |
2947 | ||
2948 | ret = i915_gem_object_get_pages(obj); | |
2949 | if (ret) | |
2950 | return ret; | |
2951 | ||
2952 | i915_gem_object_pin_pages(obj); | |
2953 | ||
2954 | node = kzalloc(sizeof(*node), GFP_KERNEL); | |
2955 | if (node == NULL) { | |
2956 | i915_gem_object_unpin_pages(obj); | |
2957 | return -ENOMEM; | |
2958 | } | |
2959 | ||
2960 | search_free: | |
2961 | if (map_and_fenceable) | |
2962 | ret = drm_mm_insert_node_in_range_generic(&dev_priv->mm.gtt_space, node, | |
2963 | size, alignment, obj->cache_level, | |
2964 | 0, dev_priv->gtt.mappable_end); | |
2965 | else | |
2966 | ret = drm_mm_insert_node_generic(&dev_priv->mm.gtt_space, node, | |
2967 | size, alignment, obj->cache_level); | |
2968 | if (ret) { | |
2969 | ret = i915_gem_evict_something(dev, size, alignment, | |
2970 | obj->cache_level, | |
2971 | map_and_fenceable, | |
2972 | nonblocking); | |
2973 | if (ret == 0) | |
2974 | goto search_free; | |
2975 | ||
2976 | i915_gem_object_unpin_pages(obj); | |
2977 | kfree(node); | |
2978 | return ret; | |
2979 | } | |
2980 | if (WARN_ON(!i915_gem_valid_gtt_space(dev, node, obj->cache_level))) { | |
2981 | i915_gem_object_unpin_pages(obj); | |
2982 | drm_mm_put_block(node); | |
2983 | return -EINVAL; | |
2984 | } | |
2985 | ||
2986 | ret = i915_gem_gtt_prepare_object(obj); | |
2987 | if (ret) { | |
2988 | i915_gem_object_unpin_pages(obj); | |
2989 | drm_mm_put_block(node); | |
2990 | return ret; | |
2991 | } | |
2992 | ||
2993 | list_move_tail(&obj->gtt_list, &dev_priv->mm.bound_list); | |
2994 | list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list); | |
2995 | ||
2996 | obj->gtt_space = node; | |
2997 | obj->gtt_offset = node->start; | |
2998 | ||
2999 | fenceable = | |
3000 | node->size == fence_size && | |
3001 | (node->start & (fence_alignment - 1)) == 0; | |
3002 | ||
3003 | mappable = | |
3004 | obj->gtt_offset + obj->base.size <= dev_priv->gtt.mappable_end; | |
3005 | ||
3006 | obj->map_and_fenceable = mappable && fenceable; | |
3007 | ||
3008 | i915_gem_object_unpin_pages(obj); | |
3009 | trace_i915_gem_object_bind(obj, map_and_fenceable); | |
3010 | i915_gem_verify_gtt(dev); | |
3011 | return 0; | |
3012 | } | |
3013 | ||
3014 | void | |
3015 | i915_gem_clflush_object(struct drm_i915_gem_object *obj) | |
3016 | { | |
3017 | /* If we don't have a page list set up, then we're not pinned | |
3018 | * to GPU, and we can ignore the cache flush because it'll happen | |
3019 | * again at bind time. | |
3020 | */ | |
3021 | if (obj->pages == NULL) | |
3022 | return; | |
3023 | ||
3024 | /* | |
3025 | * Stolen memory is always coherent with the GPU as it is explicitly | |
3026 | * marked as wc by the system, or the system is cache-coherent. | |
3027 | */ | |
3028 | if (obj->stolen) | |
3029 | return; | |
3030 | ||
3031 | /* If the GPU is snooping the contents of the CPU cache, | |
3032 | * we do not need to manually clear the CPU cache lines. However, | |
3033 | * the caches are only snooped when the render cache is | |
3034 | * flushed/invalidated. As we always have to emit invalidations | |
3035 | * and flushes when moving into and out of the RENDER domain, correct | |
3036 | * snooping behaviour occurs naturally as the result of our domain | |
3037 | * tracking. | |
3038 | */ | |
3039 | if (obj->cache_level != I915_CACHE_NONE) | |
3040 | return; | |
3041 | ||
3042 | trace_i915_gem_object_clflush(obj); | |
3043 | ||
3044 | drm_clflush_sg(obj->pages); | |
3045 | } | |
3046 | ||
3047 | /** Flushes the GTT write domain for the object if it's dirty. */ | |
3048 | static void | |
3049 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) | |
3050 | { | |
3051 | uint32_t old_write_domain; | |
3052 | ||
3053 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) | |
3054 | return; | |
3055 | ||
3056 | /* No actual flushing is required for the GTT write domain. Writes | |
3057 | * to it immediately go to main memory as far as we know, so there's | |
3058 | * no chipset flush. It also doesn't land in render cache. | |
3059 | * | |
3060 | * However, we do have to enforce the order so that all writes through | |
3061 | * the GTT land before any writes to the device, such as updates to | |
3062 | * the GATT itself. | |
3063 | */ | |
3064 | wmb(); | |
3065 | ||
3066 | old_write_domain = obj->base.write_domain; | |
3067 | obj->base.write_domain = 0; | |
3068 | ||
3069 | trace_i915_gem_object_change_domain(obj, | |
3070 | obj->base.read_domains, | |
3071 | old_write_domain); | |
3072 | } | |
3073 | ||
3074 | /** Flushes the CPU write domain for the object if it's dirty. */ | |
3075 | static void | |
3076 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj) | |
3077 | { | |
3078 | uint32_t old_write_domain; | |
3079 | ||
3080 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) | |
3081 | return; | |
3082 | ||
3083 | i915_gem_clflush_object(obj); | |
3084 | i915_gem_chipset_flush(obj->base.dev); | |
3085 | old_write_domain = obj->base.write_domain; | |
3086 | obj->base.write_domain = 0; | |
3087 | ||
3088 | trace_i915_gem_object_change_domain(obj, | |
3089 | obj->base.read_domains, | |
3090 | old_write_domain); | |
3091 | } | |
3092 | ||
3093 | /** | |
3094 | * Moves a single object to the GTT read, and possibly write domain. | |
3095 | * | |
3096 | * This function returns when the move is complete, including waiting on | |
3097 | * flushes to occur. | |
3098 | */ | |
3099 | int | |
3100 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) | |
3101 | { | |
3102 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; | |
3103 | uint32_t old_write_domain, old_read_domains; | |
3104 | int ret; | |
3105 | ||
3106 | /* Not valid to be called on unbound objects. */ | |
3107 | if (obj->gtt_space == NULL) | |
3108 | return -EINVAL; | |
3109 | ||
3110 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) | |
3111 | return 0; | |
3112 | ||
3113 | ret = i915_gem_object_wait_rendering(obj, !write); | |
3114 | if (ret) | |
3115 | return ret; | |
3116 | ||
3117 | i915_gem_object_flush_cpu_write_domain(obj); | |
3118 | ||
3119 | /* Serialise direct access to this object with the barriers for | |
3120 | * coherent writes from the GPU, by effectively invalidating the | |
3121 | * GTT domain upon first access. | |
3122 | */ | |
3123 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) | |
3124 | mb(); | |
3125 | ||
3126 | old_write_domain = obj->base.write_domain; | |
3127 | old_read_domains = obj->base.read_domains; | |
3128 | ||
3129 | /* It should now be out of any other write domains, and we can update | |
3130 | * the domain values for our changes. | |
3131 | */ | |
3132 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); | |
3133 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; | |
3134 | if (write) { | |
3135 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; | |
3136 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; | |
3137 | obj->dirty = 1; | |
3138 | } | |
3139 | ||
3140 | trace_i915_gem_object_change_domain(obj, | |
3141 | old_read_domains, | |
3142 | old_write_domain); | |
3143 | ||
3144 | /* And bump the LRU for this access */ | |
3145 | if (i915_gem_object_is_inactive(obj)) | |
3146 | list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list); | |
3147 | ||
3148 | return 0; | |
3149 | } | |
3150 | ||
3151 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, | |
3152 | enum i915_cache_level cache_level) | |
3153 | { | |
3154 | struct drm_device *dev = obj->base.dev; | |
3155 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3156 | int ret; | |
3157 | ||
3158 | if (obj->cache_level == cache_level) | |
3159 | return 0; | |
3160 | ||
3161 | if (obj->pin_count) { | |
3162 | DRM_DEBUG("can not change the cache level of pinned objects\n"); | |
3163 | return -EBUSY; | |
3164 | } | |
3165 | ||
3166 | if (!i915_gem_valid_gtt_space(dev, obj->gtt_space, cache_level)) { | |
3167 | ret = i915_gem_object_unbind(obj); | |
3168 | if (ret) | |
3169 | return ret; | |
3170 | } | |
3171 | ||
3172 | if (obj->gtt_space) { | |
3173 | ret = i915_gem_object_finish_gpu(obj); | |
3174 | if (ret) | |
3175 | return ret; | |
3176 | ||
3177 | i915_gem_object_finish_gtt(obj); | |
3178 | ||
3179 | /* Before SandyBridge, you could not use tiling or fence | |
3180 | * registers with snooped memory, so relinquish any fences | |
3181 | * currently pointing to our region in the aperture. | |
3182 | */ | |
3183 | if (INTEL_INFO(dev)->gen < 6) { | |
3184 | ret = i915_gem_object_put_fence(obj); | |
3185 | if (ret) | |
3186 | return ret; | |
3187 | } | |
3188 | ||
3189 | if (obj->has_global_gtt_mapping) | |
3190 | i915_gem_gtt_bind_object(obj, cache_level); | |
3191 | if (obj->has_aliasing_ppgtt_mapping) | |
3192 | i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt, | |
3193 | obj, cache_level); | |
3194 | ||
3195 | obj->gtt_space->color = cache_level; | |
3196 | } | |
3197 | ||
3198 | if (cache_level == I915_CACHE_NONE) { | |
3199 | u32 old_read_domains, old_write_domain; | |
3200 | ||
3201 | /* If we're coming from LLC cached, then we haven't | |
3202 | * actually been tracking whether the data is in the | |
3203 | * CPU cache or not, since we only allow one bit set | |
3204 | * in obj->write_domain and have been skipping the clflushes. | |
3205 | * Just set it to the CPU cache for now. | |
3206 | */ | |
3207 | WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU); | |
3208 | WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU); | |
3209 | ||
3210 | old_read_domains = obj->base.read_domains; | |
3211 | old_write_domain = obj->base.write_domain; | |
3212 | ||
3213 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
3214 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
3215 | ||
3216 | trace_i915_gem_object_change_domain(obj, | |
3217 | old_read_domains, | |
3218 | old_write_domain); | |
3219 | } | |
3220 | ||
3221 | obj->cache_level = cache_level; | |
3222 | i915_gem_verify_gtt(dev); | |
3223 | return 0; | |
3224 | } | |
3225 | ||
3226 | int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, | |
3227 | struct drm_file *file) | |
3228 | { | |
3229 | struct drm_i915_gem_caching *args = data; | |
3230 | struct drm_i915_gem_object *obj; | |
3231 | int ret; | |
3232 | ||
3233 | ret = i915_mutex_lock_interruptible(dev); | |
3234 | if (ret) | |
3235 | return ret; | |
3236 | ||
3237 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
3238 | if (&obj->base == NULL) { | |
3239 | ret = -ENOENT; | |
3240 | goto unlock; | |
3241 | } | |
3242 | ||
3243 | args->caching = obj->cache_level != I915_CACHE_NONE; | |
3244 | ||
3245 | drm_gem_object_unreference(&obj->base); | |
3246 | unlock: | |
3247 | mutex_unlock(&dev->struct_mutex); | |
3248 | return ret; | |
3249 | } | |
3250 | ||
3251 | int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data, | |
3252 | struct drm_file *file) | |
3253 | { | |
3254 | struct drm_i915_gem_caching *args = data; | |
3255 | struct drm_i915_gem_object *obj; | |
3256 | enum i915_cache_level level; | |
3257 | int ret; | |
3258 | ||
3259 | switch (args->caching) { | |
3260 | case I915_CACHING_NONE: | |
3261 | level = I915_CACHE_NONE; | |
3262 | break; | |
3263 | case I915_CACHING_CACHED: | |
3264 | level = I915_CACHE_LLC; | |
3265 | break; | |
3266 | default: | |
3267 | return -EINVAL; | |
3268 | } | |
3269 | ||
3270 | ret = i915_mutex_lock_interruptible(dev); | |
3271 | if (ret) | |
3272 | return ret; | |
3273 | ||
3274 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
3275 | if (&obj->base == NULL) { | |
3276 | ret = -ENOENT; | |
3277 | goto unlock; | |
3278 | } | |
3279 | ||
3280 | ret = i915_gem_object_set_cache_level(obj, level); | |
3281 | ||
3282 | drm_gem_object_unreference(&obj->base); | |
3283 | unlock: | |
3284 | mutex_unlock(&dev->struct_mutex); | |
3285 | return ret; | |
3286 | } | |
3287 | ||
3288 | /* | |
3289 | * Prepare buffer for display plane (scanout, cursors, etc). | |
3290 | * Can be called from an uninterruptible phase (modesetting) and allows | |
3291 | * any flushes to be pipelined (for pageflips). | |
3292 | */ | |
3293 | int | |
3294 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, | |
3295 | u32 alignment, | |
3296 | struct intel_ring_buffer *pipelined) | |
3297 | { | |
3298 | u32 old_read_domains, old_write_domain; | |
3299 | int ret; | |
3300 | ||
3301 | if (pipelined != obj->ring) { | |
3302 | ret = i915_gem_object_sync(obj, pipelined); | |
3303 | if (ret) | |
3304 | return ret; | |
3305 | } | |
3306 | ||
3307 | /* The display engine is not coherent with the LLC cache on gen6. As | |
3308 | * a result, we make sure that the pinning that is about to occur is | |
3309 | * done with uncached PTEs. This is lowest common denominator for all | |
3310 | * chipsets. | |
3311 | * | |
3312 | * However for gen6+, we could do better by using the GFDT bit instead | |
3313 | * of uncaching, which would allow us to flush all the LLC-cached data | |
3314 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. | |
3315 | */ | |
3316 | ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE); | |
3317 | if (ret) | |
3318 | return ret; | |
3319 | ||
3320 | /* As the user may map the buffer once pinned in the display plane | |
3321 | * (e.g. libkms for the bootup splash), we have to ensure that we | |
3322 | * always use map_and_fenceable for all scanout buffers. | |
3323 | */ | |
3324 | ret = i915_gem_object_pin(obj, alignment, true, false); | |
3325 | if (ret) | |
3326 | return ret; | |
3327 | ||
3328 | i915_gem_object_flush_cpu_write_domain(obj); | |
3329 | ||
3330 | old_write_domain = obj->base.write_domain; | |
3331 | old_read_domains = obj->base.read_domains; | |
3332 | ||
3333 | /* It should now be out of any other write domains, and we can update | |
3334 | * the domain values for our changes. | |
3335 | */ | |
3336 | obj->base.write_domain = 0; | |
3337 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; | |
3338 | ||
3339 | trace_i915_gem_object_change_domain(obj, | |
3340 | old_read_domains, | |
3341 | old_write_domain); | |
3342 | ||
3343 | return 0; | |
3344 | } | |
3345 | ||
3346 | int | |
3347 | i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj) | |
3348 | { | |
3349 | int ret; | |
3350 | ||
3351 | if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0) | |
3352 | return 0; | |
3353 | ||
3354 | ret = i915_gem_object_wait_rendering(obj, false); | |
3355 | if (ret) | |
3356 | return ret; | |
3357 | ||
3358 | /* Ensure that we invalidate the GPU's caches and TLBs. */ | |
3359 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; | |
3360 | return 0; | |
3361 | } | |
3362 | ||
3363 | /** | |
3364 | * Moves a single object to the CPU read, and possibly write domain. | |
3365 | * | |
3366 | * This function returns when the move is complete, including waiting on | |
3367 | * flushes to occur. | |
3368 | */ | |
3369 | int | |
3370 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) | |
3371 | { | |
3372 | uint32_t old_write_domain, old_read_domains; | |
3373 | int ret; | |
3374 | ||
3375 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) | |
3376 | return 0; | |
3377 | ||
3378 | ret = i915_gem_object_wait_rendering(obj, !write); | |
3379 | if (ret) | |
3380 | return ret; | |
3381 | ||
3382 | i915_gem_object_flush_gtt_write_domain(obj); | |
3383 | ||
3384 | old_write_domain = obj->base.write_domain; | |
3385 | old_read_domains = obj->base.read_domains; | |
3386 | ||
3387 | /* Flush the CPU cache if it's still invalid. */ | |
3388 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { | |
3389 | i915_gem_clflush_object(obj); | |
3390 | ||
3391 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; | |
3392 | } | |
3393 | ||
3394 | /* It should now be out of any other write domains, and we can update | |
3395 | * the domain values for our changes. | |
3396 | */ | |
3397 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); | |
3398 | ||
3399 | /* If we're writing through the CPU, then the GPU read domains will | |
3400 | * need to be invalidated at next use. | |
3401 | */ | |
3402 | if (write) { | |
3403 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
3404 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
3405 | } | |
3406 | ||
3407 | trace_i915_gem_object_change_domain(obj, | |
3408 | old_read_domains, | |
3409 | old_write_domain); | |
3410 | ||
3411 | return 0; | |
3412 | } | |
3413 | ||
3414 | /* Throttle our rendering by waiting until the ring has completed our requests | |
3415 | * emitted over 20 msec ago. | |
3416 | * | |
3417 | * Note that if we were to use the current jiffies each time around the loop, | |
3418 | * we wouldn't escape the function with any frames outstanding if the time to | |
3419 | * render a frame was over 20ms. | |
3420 | * | |
3421 | * This should get us reasonable parallelism between CPU and GPU but also | |
3422 | * relatively low latency when blocking on a particular request to finish. | |
3423 | */ | |
3424 | static int | |
3425 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) | |
3426 | { | |
3427 | struct drm_i915_private *dev_priv = dev->dev_private; | |
3428 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
3429 | unsigned long recent_enough = jiffies - msecs_to_jiffies(20); | |
3430 | struct drm_i915_gem_request *request; | |
3431 | struct intel_ring_buffer *ring = NULL; | |
3432 | unsigned reset_counter; | |
3433 | u32 seqno = 0; | |
3434 | int ret; | |
3435 | ||
3436 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); | |
3437 | if (ret) | |
3438 | return ret; | |
3439 | ||
3440 | ret = i915_gem_check_wedge(&dev_priv->gpu_error, false); | |
3441 | if (ret) | |
3442 | return ret; | |
3443 | ||
3444 | spin_lock(&file_priv->mm.lock); | |
3445 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { | |
3446 | if (time_after_eq(request->emitted_jiffies, recent_enough)) | |
3447 | break; | |
3448 | ||
3449 | ring = request->ring; | |
3450 | seqno = request->seqno; | |
3451 | } | |
3452 | reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter); | |
3453 | spin_unlock(&file_priv->mm.lock); | |
3454 | ||
3455 | if (seqno == 0) | |
3456 | return 0; | |
3457 | ||
3458 | ret = __wait_seqno(ring, seqno, reset_counter, true, NULL); | |
3459 | if (ret == 0) | |
3460 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); | |
3461 | ||
3462 | return ret; | |
3463 | } | |
3464 | ||
3465 | int | |
3466 | i915_gem_object_pin(struct drm_i915_gem_object *obj, | |
3467 | uint32_t alignment, | |
3468 | bool map_and_fenceable, | |
3469 | bool nonblocking) | |
3470 | { | |
3471 | int ret; | |
3472 | ||
3473 | if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT)) | |
3474 | return -EBUSY; | |
3475 | ||
3476 | if (obj->gtt_space != NULL) { | |
3477 | if ((alignment && obj->gtt_offset & (alignment - 1)) || | |
3478 | (map_and_fenceable && !obj->map_and_fenceable)) { | |
3479 | WARN(obj->pin_count, | |
3480 | "bo is already pinned with incorrect alignment:" | |
3481 | " offset=%x, req.alignment=%x, req.map_and_fenceable=%d," | |
3482 | " obj->map_and_fenceable=%d\n", | |
3483 | obj->gtt_offset, alignment, | |
3484 | map_and_fenceable, | |
3485 | obj->map_and_fenceable); | |
3486 | ret = i915_gem_object_unbind(obj); | |
3487 | if (ret) | |
3488 | return ret; | |
3489 | } | |
3490 | } | |
3491 | ||
3492 | if (obj->gtt_space == NULL) { | |
3493 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
3494 | ||
3495 | ret = i915_gem_object_bind_to_gtt(obj, alignment, | |
3496 | map_and_fenceable, | |
3497 | nonblocking); | |
3498 | if (ret) | |
3499 | return ret; | |
3500 | ||
3501 | if (!dev_priv->mm.aliasing_ppgtt) | |
3502 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
3503 | } | |
3504 | ||
3505 | if (!obj->has_global_gtt_mapping && map_and_fenceable) | |
3506 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
3507 | ||
3508 | obj->pin_count++; | |
3509 | obj->pin_mappable |= map_and_fenceable; | |
3510 | ||
3511 | return 0; | |
3512 | } | |
3513 | ||
3514 | void | |
3515 | i915_gem_object_unpin(struct drm_i915_gem_object *obj) | |
3516 | { | |
3517 | BUG_ON(obj->pin_count == 0); | |
3518 | BUG_ON(obj->gtt_space == NULL); | |
3519 | ||
3520 | if (--obj->pin_count == 0) | |
3521 | obj->pin_mappable = false; | |
3522 | } | |
3523 | ||
3524 | int | |
3525 | i915_gem_pin_ioctl(struct drm_device *dev, void *data, | |
3526 | struct drm_file *file) | |
3527 | { | |
3528 | struct drm_i915_gem_pin *args = data; | |
3529 | struct drm_i915_gem_object *obj; | |
3530 | int ret; | |
3531 | ||
3532 | ret = i915_mutex_lock_interruptible(dev); | |
3533 | if (ret) | |
3534 | return ret; | |
3535 | ||
3536 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
3537 | if (&obj->base == NULL) { | |
3538 | ret = -ENOENT; | |
3539 | goto unlock; | |
3540 | } | |
3541 | ||
3542 | if (obj->madv != I915_MADV_WILLNEED) { | |
3543 | DRM_ERROR("Attempting to pin a purgeable buffer\n"); | |
3544 | ret = -EINVAL; | |
3545 | goto out; | |
3546 | } | |
3547 | ||
3548 | if (obj->pin_filp != NULL && obj->pin_filp != file) { | |
3549 | DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n", | |
3550 | args->handle); | |
3551 | ret = -EINVAL; | |
3552 | goto out; | |
3553 | } | |
3554 | ||
3555 | if (obj->user_pin_count == 0) { | |
3556 | ret = i915_gem_object_pin(obj, args->alignment, true, false); | |
3557 | if (ret) | |
3558 | goto out; | |
3559 | } | |
3560 | ||
3561 | obj->user_pin_count++; | |
3562 | obj->pin_filp = file; | |
3563 | ||
3564 | /* XXX - flush the CPU caches for pinned objects | |
3565 | * as the X server doesn't manage domains yet | |
3566 | */ | |
3567 | i915_gem_object_flush_cpu_write_domain(obj); | |
3568 | args->offset = obj->gtt_offset; | |
3569 | out: | |
3570 | drm_gem_object_unreference(&obj->base); | |
3571 | unlock: | |
3572 | mutex_unlock(&dev->struct_mutex); | |
3573 | return ret; | |
3574 | } | |
3575 | ||
3576 | int | |
3577 | i915_gem_unpin_ioctl(struct drm_device *dev, void *data, | |
3578 | struct drm_file *file) | |
3579 | { | |
3580 | struct drm_i915_gem_pin *args = data; | |
3581 | struct drm_i915_gem_object *obj; | |
3582 | int ret; | |
3583 | ||
3584 | ret = i915_mutex_lock_interruptible(dev); | |
3585 | if (ret) | |
3586 | return ret; | |
3587 | ||
3588 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
3589 | if (&obj->base == NULL) { | |
3590 | ret = -ENOENT; | |
3591 | goto unlock; | |
3592 | } | |
3593 | ||
3594 | if (obj->pin_filp != file) { | |
3595 | DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n", | |
3596 | args->handle); | |
3597 | ret = -EINVAL; | |
3598 | goto out; | |
3599 | } | |
3600 | obj->user_pin_count--; | |
3601 | if (obj->user_pin_count == 0) { | |
3602 | obj->pin_filp = NULL; | |
3603 | i915_gem_object_unpin(obj); | |
3604 | } | |
3605 | ||
3606 | out: | |
3607 | drm_gem_object_unreference(&obj->base); | |
3608 | unlock: | |
3609 | mutex_unlock(&dev->struct_mutex); | |
3610 | return ret; | |
3611 | } | |
3612 | ||
3613 | int | |
3614 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, | |
3615 | struct drm_file *file) | |
3616 | { | |
3617 | struct drm_i915_gem_busy *args = data; | |
3618 | struct drm_i915_gem_object *obj; | |
3619 | int ret; | |
3620 | ||
3621 | ret = i915_mutex_lock_interruptible(dev); | |
3622 | if (ret) | |
3623 | return ret; | |
3624 | ||
3625 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); | |
3626 | if (&obj->base == NULL) { | |
3627 | ret = -ENOENT; | |
3628 | goto unlock; | |
3629 | } | |
3630 | ||
3631 | /* Count all active objects as busy, even if they are currently not used | |
3632 | * by the gpu. Users of this interface expect objects to eventually | |
3633 | * become non-busy without any further actions, therefore emit any | |
3634 | * necessary flushes here. | |
3635 | */ | |
3636 | ret = i915_gem_object_flush_active(obj); | |
3637 | ||
3638 | args->busy = obj->active; | |
3639 | if (obj->ring) { | |
3640 | BUILD_BUG_ON(I915_NUM_RINGS > 16); | |
3641 | args->busy |= intel_ring_flag(obj->ring) << 16; | |
3642 | } | |
3643 | ||
3644 | drm_gem_object_unreference(&obj->base); | |
3645 | unlock: | |
3646 | mutex_unlock(&dev->struct_mutex); | |
3647 | return ret; | |
3648 | } | |
3649 | ||
3650 | int | |
3651 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, | |
3652 | struct drm_file *file_priv) | |
3653 | { | |
3654 | return i915_gem_ring_throttle(dev, file_priv); | |
3655 | } | |
3656 | ||
3657 | int | |
3658 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, | |
3659 | struct drm_file *file_priv) | |
3660 | { | |
3661 | struct drm_i915_gem_madvise *args = data; | |
3662 | struct drm_i915_gem_object *obj; | |
3663 | int ret; | |
3664 | ||
3665 | switch (args->madv) { | |
3666 | case I915_MADV_DONTNEED: | |
3667 | case I915_MADV_WILLNEED: | |
3668 | break; | |
3669 | default: | |
3670 | return -EINVAL; | |
3671 | } | |
3672 | ||
3673 | ret = i915_mutex_lock_interruptible(dev); | |
3674 | if (ret) | |
3675 | return ret; | |
3676 | ||
3677 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); | |
3678 | if (&obj->base == NULL) { | |
3679 | ret = -ENOENT; | |
3680 | goto unlock; | |
3681 | } | |
3682 | ||
3683 | if (obj->pin_count) { | |
3684 | ret = -EINVAL; | |
3685 | goto out; | |
3686 | } | |
3687 | ||
3688 | if (obj->madv != __I915_MADV_PURGED) | |
3689 | obj->madv = args->madv; | |
3690 | ||
3691 | /* if the object is no longer attached, discard its backing storage */ | |
3692 | if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL) | |
3693 | i915_gem_object_truncate(obj); | |
3694 | ||
3695 | args->retained = obj->madv != __I915_MADV_PURGED; | |
3696 | ||
3697 | out: | |
3698 | drm_gem_object_unreference(&obj->base); | |
3699 | unlock: | |
3700 | mutex_unlock(&dev->struct_mutex); | |
3701 | return ret; | |
3702 | } | |
3703 | ||
3704 | void i915_gem_object_init(struct drm_i915_gem_object *obj, | |
3705 | const struct drm_i915_gem_object_ops *ops) | |
3706 | { | |
3707 | INIT_LIST_HEAD(&obj->mm_list); | |
3708 | INIT_LIST_HEAD(&obj->gtt_list); | |
3709 | INIT_LIST_HEAD(&obj->ring_list); | |
3710 | INIT_LIST_HEAD(&obj->exec_list); | |
3711 | ||
3712 | obj->ops = ops; | |
3713 | ||
3714 | obj->fence_reg = I915_FENCE_REG_NONE; | |
3715 | obj->madv = I915_MADV_WILLNEED; | |
3716 | /* Avoid an unnecessary call to unbind on the first bind. */ | |
3717 | obj->map_and_fenceable = true; | |
3718 | ||
3719 | i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size); | |
3720 | } | |
3721 | ||
3722 | static const struct drm_i915_gem_object_ops i915_gem_object_ops = { | |
3723 | .get_pages = i915_gem_object_get_pages_gtt, | |
3724 | .put_pages = i915_gem_object_put_pages_gtt, | |
3725 | }; | |
3726 | ||
3727 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, | |
3728 | size_t size) | |
3729 | { | |
3730 | struct drm_i915_gem_object *obj; | |
3731 | struct address_space *mapping; | |
3732 | gfp_t mask; | |
3733 | ||
3734 | obj = i915_gem_object_alloc(dev); | |
3735 | if (obj == NULL) | |
3736 | return NULL; | |
3737 | ||
3738 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { | |
3739 | i915_gem_object_free(obj); | |
3740 | return NULL; | |
3741 | } | |
3742 | ||
3743 | mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; | |
3744 | if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) { | |
3745 | /* 965gm cannot relocate objects above 4GiB. */ | |
3746 | mask &= ~__GFP_HIGHMEM; | |
3747 | mask |= __GFP_DMA32; | |
3748 | } | |
3749 | ||
3750 | mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; | |
3751 | mapping_set_gfp_mask(mapping, mask); | |
3752 | ||
3753 | i915_gem_object_init(obj, &i915_gem_object_ops); | |
3754 | ||
3755 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
3756 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
3757 | ||
3758 | if (HAS_LLC(dev)) { | |
3759 | /* On some devices, we can have the GPU use the LLC (the CPU | |
3760 | * cache) for about a 10% performance improvement | |
3761 | * compared to uncached. Graphics requests other than | |
3762 | * display scanout are coherent with the CPU in | |
3763 | * accessing this cache. This means in this mode we | |
3764 | * don't need to clflush on the CPU side, and on the | |
3765 | * GPU side we only need to flush internal caches to | |
3766 | * get data visible to the CPU. | |
3767 | * | |
3768 | * However, we maintain the display planes as UC, and so | |
3769 | * need to rebind when first used as such. | |
3770 | */ | |
3771 | obj->cache_level = I915_CACHE_LLC; | |
3772 | } else | |
3773 | obj->cache_level = I915_CACHE_NONE; | |
3774 | ||
3775 | return obj; | |
3776 | } | |
3777 | ||
3778 | int i915_gem_init_object(struct drm_gem_object *obj) | |
3779 | { | |
3780 | BUG(); | |
3781 | ||
3782 | return 0; | |
3783 | } | |
3784 | ||
3785 | void i915_gem_free_object(struct drm_gem_object *gem_obj) | |
3786 | { | |
3787 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); | |
3788 | struct drm_device *dev = obj->base.dev; | |
3789 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3790 | ||
3791 | trace_i915_gem_object_destroy(obj); | |
3792 | ||
3793 | if (obj->phys_obj) | |
3794 | i915_gem_detach_phys_object(dev, obj); | |
3795 | ||
3796 | obj->pin_count = 0; | |
3797 | if (WARN_ON(i915_gem_object_unbind(obj) == -ERESTARTSYS)) { | |
3798 | bool was_interruptible; | |
3799 | ||
3800 | was_interruptible = dev_priv->mm.interruptible; | |
3801 | dev_priv->mm.interruptible = false; | |
3802 | ||
3803 | WARN_ON(i915_gem_object_unbind(obj)); | |
3804 | ||
3805 | dev_priv->mm.interruptible = was_interruptible; | |
3806 | } | |
3807 | ||
3808 | obj->pages_pin_count = 0; | |
3809 | i915_gem_object_put_pages(obj); | |
3810 | i915_gem_object_free_mmap_offset(obj); | |
3811 | i915_gem_object_release_stolen(obj); | |
3812 | ||
3813 | BUG_ON(obj->pages); | |
3814 | ||
3815 | if (obj->base.import_attach) | |
3816 | drm_prime_gem_destroy(&obj->base, NULL); | |
3817 | ||
3818 | drm_gem_object_release(&obj->base); | |
3819 | i915_gem_info_remove_obj(dev_priv, obj->base.size); | |
3820 | ||
3821 | kfree(obj->bit_17); | |
3822 | i915_gem_object_free(obj); | |
3823 | } | |
3824 | ||
3825 | int | |
3826 | i915_gem_idle(struct drm_device *dev) | |
3827 | { | |
3828 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3829 | int ret; | |
3830 | ||
3831 | mutex_lock(&dev->struct_mutex); | |
3832 | ||
3833 | if (dev_priv->mm.suspended) { | |
3834 | mutex_unlock(&dev->struct_mutex); | |
3835 | return 0; | |
3836 | } | |
3837 | ||
3838 | ret = i915_gpu_idle(dev); | |
3839 | if (ret) { | |
3840 | mutex_unlock(&dev->struct_mutex); | |
3841 | return ret; | |
3842 | } | |
3843 | i915_gem_retire_requests(dev); | |
3844 | ||
3845 | /* Under UMS, be paranoid and evict. */ | |
3846 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) | |
3847 | i915_gem_evict_everything(dev); | |
3848 | ||
3849 | i915_gem_reset_fences(dev); | |
3850 | ||
3851 | /* Hack! Don't let anybody do execbuf while we don't control the chip. | |
3852 | * We need to replace this with a semaphore, or something. | |
3853 | * And not confound mm.suspended! | |
3854 | */ | |
3855 | dev_priv->mm.suspended = 1; | |
3856 | del_timer_sync(&dev_priv->gpu_error.hangcheck_timer); | |
3857 | ||
3858 | i915_kernel_lost_context(dev); | |
3859 | i915_gem_cleanup_ringbuffer(dev); | |
3860 | ||
3861 | mutex_unlock(&dev->struct_mutex); | |
3862 | ||
3863 | /* Cancel the retire work handler, which should be idle now. */ | |
3864 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); | |
3865 | ||
3866 | return 0; | |
3867 | } | |
3868 | ||
3869 | void i915_gem_l3_remap(struct drm_device *dev) | |
3870 | { | |
3871 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3872 | u32 misccpctl; | |
3873 | int i; | |
3874 | ||
3875 | if (!HAS_L3_GPU_CACHE(dev)) | |
3876 | return; | |
3877 | ||
3878 | if (!dev_priv->l3_parity.remap_info) | |
3879 | return; | |
3880 | ||
3881 | misccpctl = I915_READ(GEN7_MISCCPCTL); | |
3882 | I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE); | |
3883 | POSTING_READ(GEN7_MISCCPCTL); | |
3884 | ||
3885 | for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) { | |
3886 | u32 remap = I915_READ(GEN7_L3LOG_BASE + i); | |
3887 | if (remap && remap != dev_priv->l3_parity.remap_info[i/4]) | |
3888 | DRM_DEBUG("0x%x was already programmed to %x\n", | |
3889 | GEN7_L3LOG_BASE + i, remap); | |
3890 | if (remap && !dev_priv->l3_parity.remap_info[i/4]) | |
3891 | DRM_DEBUG_DRIVER("Clearing remapped register\n"); | |
3892 | I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]); | |
3893 | } | |
3894 | ||
3895 | /* Make sure all the writes land before disabling dop clock gating */ | |
3896 | POSTING_READ(GEN7_L3LOG_BASE); | |
3897 | ||
3898 | I915_WRITE(GEN7_MISCCPCTL, misccpctl); | |
3899 | } | |
3900 | ||
3901 | void i915_gem_init_swizzling(struct drm_device *dev) | |
3902 | { | |
3903 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3904 | ||
3905 | if (INTEL_INFO(dev)->gen < 5 || | |
3906 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) | |
3907 | return; | |
3908 | ||
3909 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | | |
3910 | DISP_TILE_SURFACE_SWIZZLING); | |
3911 | ||
3912 | if (IS_GEN5(dev)) | |
3913 | return; | |
3914 | ||
3915 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); | |
3916 | if (IS_GEN6(dev)) | |
3917 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); | |
3918 | else if (IS_GEN7(dev)) | |
3919 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); | |
3920 | else | |
3921 | BUG(); | |
3922 | } | |
3923 | ||
3924 | static bool | |
3925 | intel_enable_blt(struct drm_device *dev) | |
3926 | { | |
3927 | if (!HAS_BLT(dev)) | |
3928 | return false; | |
3929 | ||
3930 | /* The blitter was dysfunctional on early prototypes */ | |
3931 | if (IS_GEN6(dev) && dev->pdev->revision < 8) { | |
3932 | DRM_INFO("BLT not supported on this pre-production hardware;" | |
3933 | " graphics performance will be degraded.\n"); | |
3934 | return false; | |
3935 | } | |
3936 | ||
3937 | return true; | |
3938 | } | |
3939 | ||
3940 | static int i915_gem_init_rings(struct drm_device *dev) | |
3941 | { | |
3942 | struct drm_i915_private *dev_priv = dev->dev_private; | |
3943 | int ret; | |
3944 | ||
3945 | ret = intel_init_render_ring_buffer(dev); | |
3946 | if (ret) | |
3947 | return ret; | |
3948 | ||
3949 | if (HAS_BSD(dev)) { | |
3950 | ret = intel_init_bsd_ring_buffer(dev); | |
3951 | if (ret) | |
3952 | goto cleanup_render_ring; | |
3953 | } | |
3954 | ||
3955 | if (intel_enable_blt(dev)) { | |
3956 | ret = intel_init_blt_ring_buffer(dev); | |
3957 | if (ret) | |
3958 | goto cleanup_bsd_ring; | |
3959 | } | |
3960 | ||
3961 | ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000)); | |
3962 | if (ret) | |
3963 | goto cleanup_blt_ring; | |
3964 | ||
3965 | return 0; | |
3966 | ||
3967 | cleanup_blt_ring: | |
3968 | intel_cleanup_ring_buffer(&dev_priv->ring[BCS]); | |
3969 | cleanup_bsd_ring: | |
3970 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); | |
3971 | cleanup_render_ring: | |
3972 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); | |
3973 | ||
3974 | return ret; | |
3975 | } | |
3976 | ||
3977 | int | |
3978 | i915_gem_init_hw(struct drm_device *dev) | |
3979 | { | |
3980 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3981 | int ret; | |
3982 | ||
3983 | if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt()) | |
3984 | return -EIO; | |
3985 | ||
3986 | if (IS_HASWELL(dev) && (I915_READ(0x120010) == 1)) | |
3987 | I915_WRITE(0x9008, I915_READ(0x9008) | 0xf0000); | |
3988 | ||
3989 | i915_gem_l3_remap(dev); | |
3990 | ||
3991 | i915_gem_init_swizzling(dev); | |
3992 | ||
3993 | ret = i915_gem_init_rings(dev); | |
3994 | if (ret) | |
3995 | return ret; | |
3996 | ||
3997 | /* | |
3998 | * XXX: There was some w/a described somewhere suggesting loading | |
3999 | * contexts before PPGTT. | |
4000 | */ | |
4001 | i915_gem_context_init(dev); | |
4002 | i915_gem_init_ppgtt(dev); | |
4003 | ||
4004 | return 0; | |
4005 | } | |
4006 | ||
4007 | int i915_gem_init(struct drm_device *dev) | |
4008 | { | |
4009 | struct drm_i915_private *dev_priv = dev->dev_private; | |
4010 | int ret; | |
4011 | ||
4012 | mutex_lock(&dev->struct_mutex); | |
4013 | i915_gem_init_global_gtt(dev); | |
4014 | ret = i915_gem_init_hw(dev); | |
4015 | mutex_unlock(&dev->struct_mutex); | |
4016 | if (ret) { | |
4017 | i915_gem_cleanup_aliasing_ppgtt(dev); | |
4018 | return ret; | |
4019 | } | |
4020 | ||
4021 | /* Allow hardware batchbuffers unless told otherwise, but not for KMS. */ | |
4022 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) | |
4023 | dev_priv->dri1.allow_batchbuffer = 1; | |
4024 | return 0; | |
4025 | } | |
4026 | ||
4027 | void | |
4028 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) | |
4029 | { | |
4030 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4031 | struct intel_ring_buffer *ring; | |
4032 | int i; | |
4033 | ||
4034 | for_each_ring(ring, dev_priv, i) | |
4035 | intel_cleanup_ring_buffer(ring); | |
4036 | } | |
4037 | ||
4038 | int | |
4039 | i915_gem_entervt_ioctl(struct drm_device *dev, void *data, | |
4040 | struct drm_file *file_priv) | |
4041 | { | |
4042 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4043 | int ret; | |
4044 | ||
4045 | if (drm_core_check_feature(dev, DRIVER_MODESET)) | |
4046 | return 0; | |
4047 | ||
4048 | if (i915_reset_in_progress(&dev_priv->gpu_error)) { | |
4049 | DRM_ERROR("Reenabling wedged hardware, good luck\n"); | |
4050 | atomic_set(&dev_priv->gpu_error.reset_counter, 0); | |
4051 | } | |
4052 | ||
4053 | mutex_lock(&dev->struct_mutex); | |
4054 | dev_priv->mm.suspended = 0; | |
4055 | ||
4056 | ret = i915_gem_init_hw(dev); | |
4057 | if (ret != 0) { | |
4058 | mutex_unlock(&dev->struct_mutex); | |
4059 | return ret; | |
4060 | } | |
4061 | ||
4062 | BUG_ON(!list_empty(&dev_priv->mm.active_list)); | |
4063 | mutex_unlock(&dev->struct_mutex); | |
4064 | ||
4065 | ret = drm_irq_install(dev); | |
4066 | if (ret) | |
4067 | goto cleanup_ringbuffer; | |
4068 | ||
4069 | return 0; | |
4070 | ||
4071 | cleanup_ringbuffer: | |
4072 | mutex_lock(&dev->struct_mutex); | |
4073 | i915_gem_cleanup_ringbuffer(dev); | |
4074 | dev_priv->mm.suspended = 1; | |
4075 | mutex_unlock(&dev->struct_mutex); | |
4076 | ||
4077 | return ret; | |
4078 | } | |
4079 | ||
4080 | int | |
4081 | i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, | |
4082 | struct drm_file *file_priv) | |
4083 | { | |
4084 | if (drm_core_check_feature(dev, DRIVER_MODESET)) | |
4085 | return 0; | |
4086 | ||
4087 | drm_irq_uninstall(dev); | |
4088 | return i915_gem_idle(dev); | |
4089 | } | |
4090 | ||
4091 | void | |
4092 | i915_gem_lastclose(struct drm_device *dev) | |
4093 | { | |
4094 | int ret; | |
4095 | ||
4096 | if (drm_core_check_feature(dev, DRIVER_MODESET)) | |
4097 | return; | |
4098 | ||
4099 | ret = i915_gem_idle(dev); | |
4100 | if (ret) | |
4101 | DRM_ERROR("failed to idle hardware: %d\n", ret); | |
4102 | } | |
4103 | ||
4104 | static void | |
4105 | init_ring_lists(struct intel_ring_buffer *ring) | |
4106 | { | |
4107 | INIT_LIST_HEAD(&ring->active_list); | |
4108 | INIT_LIST_HEAD(&ring->request_list); | |
4109 | } | |
4110 | ||
4111 | void | |
4112 | i915_gem_load(struct drm_device *dev) | |
4113 | { | |
4114 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4115 | int i; | |
4116 | ||
4117 | dev_priv->slab = | |
4118 | kmem_cache_create("i915_gem_object", | |
4119 | sizeof(struct drm_i915_gem_object), 0, | |
4120 | SLAB_HWCACHE_ALIGN, | |
4121 | NULL); | |
4122 | ||
4123 | INIT_LIST_HEAD(&dev_priv->mm.active_list); | |
4124 | INIT_LIST_HEAD(&dev_priv->mm.inactive_list); | |
4125 | INIT_LIST_HEAD(&dev_priv->mm.unbound_list); | |
4126 | INIT_LIST_HEAD(&dev_priv->mm.bound_list); | |
4127 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); | |
4128 | for (i = 0; i < I915_NUM_RINGS; i++) | |
4129 | init_ring_lists(&dev_priv->ring[i]); | |
4130 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) | |
4131 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); | |
4132 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, | |
4133 | i915_gem_retire_work_handler); | |
4134 | init_waitqueue_head(&dev_priv->gpu_error.reset_queue); | |
4135 | ||
4136 | /* On GEN3 we really need to make sure the ARB C3 LP bit is set */ | |
4137 | if (IS_GEN3(dev)) { | |
4138 | I915_WRITE(MI_ARB_STATE, | |
4139 | _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE)); | |
4140 | } | |
4141 | ||
4142 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; | |
4143 | ||
4144 | /* Old X drivers will take 0-2 for front, back, depth buffers */ | |
4145 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) | |
4146 | dev_priv->fence_reg_start = 3; | |
4147 | ||
4148 | if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) | |
4149 | dev_priv->num_fence_regs = 16; | |
4150 | else | |
4151 | dev_priv->num_fence_regs = 8; | |
4152 | ||
4153 | /* Initialize fence registers to zero */ | |
4154 | i915_gem_reset_fences(dev); | |
4155 | ||
4156 | i915_gem_detect_bit_6_swizzle(dev); | |
4157 | init_waitqueue_head(&dev_priv->pending_flip_queue); | |
4158 | ||
4159 | dev_priv->mm.interruptible = true; | |
4160 | ||
4161 | dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink; | |
4162 | dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS; | |
4163 | register_shrinker(&dev_priv->mm.inactive_shrinker); | |
4164 | } | |
4165 | ||
4166 | /* | |
4167 | * Create a physically contiguous memory object for this object | |
4168 | * e.g. for cursor + overlay regs | |
4169 | */ | |
4170 | static int i915_gem_init_phys_object(struct drm_device *dev, | |
4171 | int id, int size, int align) | |
4172 | { | |
4173 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4174 | struct drm_i915_gem_phys_object *phys_obj; | |
4175 | int ret; | |
4176 | ||
4177 | if (dev_priv->mm.phys_objs[id - 1] || !size) | |
4178 | return 0; | |
4179 | ||
4180 | phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL); | |
4181 | if (!phys_obj) | |
4182 | return -ENOMEM; | |
4183 | ||
4184 | phys_obj->id = id; | |
4185 | ||
4186 | phys_obj->handle = drm_pci_alloc(dev, size, align); | |
4187 | if (!phys_obj->handle) { | |
4188 | ret = -ENOMEM; | |
4189 | goto kfree_obj; | |
4190 | } | |
4191 | #ifdef CONFIG_X86 | |
4192 | set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
4193 | #endif | |
4194 | ||
4195 | dev_priv->mm.phys_objs[id - 1] = phys_obj; | |
4196 | ||
4197 | return 0; | |
4198 | kfree_obj: | |
4199 | kfree(phys_obj); | |
4200 | return ret; | |
4201 | } | |
4202 | ||
4203 | static void i915_gem_free_phys_object(struct drm_device *dev, int id) | |
4204 | { | |
4205 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4206 | struct drm_i915_gem_phys_object *phys_obj; | |
4207 | ||
4208 | if (!dev_priv->mm.phys_objs[id - 1]) | |
4209 | return; | |
4210 | ||
4211 | phys_obj = dev_priv->mm.phys_objs[id - 1]; | |
4212 | if (phys_obj->cur_obj) { | |
4213 | i915_gem_detach_phys_object(dev, phys_obj->cur_obj); | |
4214 | } | |
4215 | ||
4216 | #ifdef CONFIG_X86 | |
4217 | set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
4218 | #endif | |
4219 | drm_pci_free(dev, phys_obj->handle); | |
4220 | kfree(phys_obj); | |
4221 | dev_priv->mm.phys_objs[id - 1] = NULL; | |
4222 | } | |
4223 | ||
4224 | void i915_gem_free_all_phys_object(struct drm_device *dev) | |
4225 | { | |
4226 | int i; | |
4227 | ||
4228 | for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++) | |
4229 | i915_gem_free_phys_object(dev, i); | |
4230 | } | |
4231 | ||
4232 | void i915_gem_detach_phys_object(struct drm_device *dev, | |
4233 | struct drm_i915_gem_object *obj) | |
4234 | { | |
4235 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; | |
4236 | char *vaddr; | |
4237 | int i; | |
4238 | int page_count; | |
4239 | ||
4240 | if (!obj->phys_obj) | |
4241 | return; | |
4242 | vaddr = obj->phys_obj->handle->vaddr; | |
4243 | ||
4244 | page_count = obj->base.size / PAGE_SIZE; | |
4245 | for (i = 0; i < page_count; i++) { | |
4246 | struct page *page = shmem_read_mapping_page(mapping, i); | |
4247 | if (!IS_ERR(page)) { | |
4248 | char *dst = kmap_atomic(page); | |
4249 | memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE); | |
4250 | kunmap_atomic(dst); | |
4251 | ||
4252 | drm_clflush_pages(&page, 1); | |
4253 | ||
4254 | set_page_dirty(page); | |
4255 | mark_page_accessed(page); | |
4256 | page_cache_release(page); | |
4257 | } | |
4258 | } | |
4259 | i915_gem_chipset_flush(dev); | |
4260 | ||
4261 | obj->phys_obj->cur_obj = NULL; | |
4262 | obj->phys_obj = NULL; | |
4263 | } | |
4264 | ||
4265 | int | |
4266 | i915_gem_attach_phys_object(struct drm_device *dev, | |
4267 | struct drm_i915_gem_object *obj, | |
4268 | int id, | |
4269 | int align) | |
4270 | { | |
4271 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; | |
4272 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4273 | int ret = 0; | |
4274 | int page_count; | |
4275 | int i; | |
4276 | ||
4277 | if (id > I915_MAX_PHYS_OBJECT) | |
4278 | return -EINVAL; | |
4279 | ||
4280 | if (obj->phys_obj) { | |
4281 | if (obj->phys_obj->id == id) | |
4282 | return 0; | |
4283 | i915_gem_detach_phys_object(dev, obj); | |
4284 | } | |
4285 | ||
4286 | /* create a new object */ | |
4287 | if (!dev_priv->mm.phys_objs[id - 1]) { | |
4288 | ret = i915_gem_init_phys_object(dev, id, | |
4289 | obj->base.size, align); | |
4290 | if (ret) { | |
4291 | DRM_ERROR("failed to init phys object %d size: %zu\n", | |
4292 | id, obj->base.size); | |
4293 | return ret; | |
4294 | } | |
4295 | } | |
4296 | ||
4297 | /* bind to the object */ | |
4298 | obj->phys_obj = dev_priv->mm.phys_objs[id - 1]; | |
4299 | obj->phys_obj->cur_obj = obj; | |
4300 | ||
4301 | page_count = obj->base.size / PAGE_SIZE; | |
4302 | ||
4303 | for (i = 0; i < page_count; i++) { | |
4304 | struct page *page; | |
4305 | char *dst, *src; | |
4306 | ||
4307 | page = shmem_read_mapping_page(mapping, i); | |
4308 | if (IS_ERR(page)) | |
4309 | return PTR_ERR(page); | |
4310 | ||
4311 | src = kmap_atomic(page); | |
4312 | dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE); | |
4313 | memcpy(dst, src, PAGE_SIZE); | |
4314 | kunmap_atomic(src); | |
4315 | ||
4316 | mark_page_accessed(page); | |
4317 | page_cache_release(page); | |
4318 | } | |
4319 | ||
4320 | return 0; | |
4321 | } | |
4322 | ||
4323 | static int | |
4324 | i915_gem_phys_pwrite(struct drm_device *dev, | |
4325 | struct drm_i915_gem_object *obj, | |
4326 | struct drm_i915_gem_pwrite *args, | |
4327 | struct drm_file *file_priv) | |
4328 | { | |
4329 | void *vaddr = obj->phys_obj->handle->vaddr + args->offset; | |
4330 | char __user *user_data = to_user_ptr(args->data_ptr); | |
4331 | ||
4332 | if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) { | |
4333 | unsigned long unwritten; | |
4334 | ||
4335 | /* The physical object once assigned is fixed for the lifetime | |
4336 | * of the obj, so we can safely drop the lock and continue | |
4337 | * to access vaddr. | |
4338 | */ | |
4339 | mutex_unlock(&dev->struct_mutex); | |
4340 | unwritten = copy_from_user(vaddr, user_data, args->size); | |
4341 | mutex_lock(&dev->struct_mutex); | |
4342 | if (unwritten) | |
4343 | return -EFAULT; | |
4344 | } | |
4345 | ||
4346 | i915_gem_chipset_flush(dev); | |
4347 | return 0; | |
4348 | } | |
4349 | ||
4350 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) | |
4351 | { | |
4352 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
4353 | ||
4354 | /* Clean up our request list when the client is going away, so that | |
4355 | * later retire_requests won't dereference our soon-to-be-gone | |
4356 | * file_priv. | |
4357 | */ | |
4358 | spin_lock(&file_priv->mm.lock); | |
4359 | while (!list_empty(&file_priv->mm.request_list)) { | |
4360 | struct drm_i915_gem_request *request; | |
4361 | ||
4362 | request = list_first_entry(&file_priv->mm.request_list, | |
4363 | struct drm_i915_gem_request, | |
4364 | client_list); | |
4365 | list_del(&request->client_list); | |
4366 | request->file_priv = NULL; | |
4367 | } | |
4368 | spin_unlock(&file_priv->mm.lock); | |
4369 | } | |
4370 | ||
4371 | static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task) | |
4372 | { | |
4373 | if (!mutex_is_locked(mutex)) | |
4374 | return false; | |
4375 | ||
4376 | #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES) | |
4377 | return mutex->owner == task; | |
4378 | #else | |
4379 | /* Since UP may be pre-empted, we cannot assume that we own the lock */ | |
4380 | return false; | |
4381 | #endif | |
4382 | } | |
4383 | ||
4384 | static int | |
4385 | i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc) | |
4386 | { | |
4387 | struct drm_i915_private *dev_priv = | |
4388 | container_of(shrinker, | |
4389 | struct drm_i915_private, | |
4390 | mm.inactive_shrinker); | |
4391 | struct drm_device *dev = dev_priv->dev; | |
4392 | struct drm_i915_gem_object *obj; | |
4393 | int nr_to_scan = sc->nr_to_scan; | |
4394 | bool unlock = true; | |
4395 | int cnt; | |
4396 | ||
4397 | if (!mutex_trylock(&dev->struct_mutex)) { | |
4398 | if (!mutex_is_locked_by(&dev->struct_mutex, current)) | |
4399 | return 0; | |
4400 | ||
4401 | if (dev_priv->mm.shrinker_no_lock_stealing) | |
4402 | return 0; | |
4403 | ||
4404 | unlock = false; | |
4405 | } | |
4406 | ||
4407 | if (nr_to_scan) { | |
4408 | nr_to_scan -= i915_gem_purge(dev_priv, nr_to_scan); | |
4409 | if (nr_to_scan > 0) | |
4410 | nr_to_scan -= __i915_gem_shrink(dev_priv, nr_to_scan, | |
4411 | false); | |
4412 | if (nr_to_scan > 0) | |
4413 | i915_gem_shrink_all(dev_priv); | |
4414 | } | |
4415 | ||
4416 | cnt = 0; | |
4417 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, gtt_list) | |
4418 | if (obj->pages_pin_count == 0) | |
4419 | cnt += obj->base.size >> PAGE_SHIFT; | |
4420 | list_for_each_entry(obj, &dev_priv->mm.inactive_list, gtt_list) | |
4421 | if (obj->pin_count == 0 && obj->pages_pin_count == 0) | |
4422 | cnt += obj->base.size >> PAGE_SHIFT; | |
4423 | ||
4424 | if (unlock) | |
4425 | mutex_unlock(&dev->struct_mutex); | |
4426 | return cnt; | |
4427 | } |