Commit | Line | Data |
---|---|---|
54cf91dc CW |
1 | /* |
2 | * Copyright © 2008,2010 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 | * Chris Wilson <chris@chris-wilson.co.uk> | |
26 | * | |
27 | */ | |
28 | ||
29 | #include "drmP.h" | |
30 | #include "drm.h" | |
31 | #include "i915_drm.h" | |
32 | #include "i915_drv.h" | |
33 | #include "i915_trace.h" | |
34 | #include "intel_drv.h" | |
35 | ||
36 | struct change_domains { | |
37 | uint32_t invalidate_domains; | |
38 | uint32_t flush_domains; | |
39 | uint32_t flush_rings; | |
c59a333f | 40 | uint32_t flips; |
54cf91dc CW |
41 | }; |
42 | ||
43 | /* | |
44 | * Set the next domain for the specified object. This | |
45 | * may not actually perform the necessary flushing/invaliding though, | |
46 | * as that may want to be batched with other set_domain operations | |
47 | * | |
48 | * This is (we hope) the only really tricky part of gem. The goal | |
49 | * is fairly simple -- track which caches hold bits of the object | |
50 | * and make sure they remain coherent. A few concrete examples may | |
51 | * help to explain how it works. For shorthand, we use the notation | |
52 | * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the | |
53 | * a pair of read and write domain masks. | |
54 | * | |
55 | * Case 1: the batch buffer | |
56 | * | |
57 | * 1. Allocated | |
58 | * 2. Written by CPU | |
59 | * 3. Mapped to GTT | |
60 | * 4. Read by GPU | |
61 | * 5. Unmapped from GTT | |
62 | * 6. Freed | |
63 | * | |
64 | * Let's take these a step at a time | |
65 | * | |
66 | * 1. Allocated | |
67 | * Pages allocated from the kernel may still have | |
68 | * cache contents, so we set them to (CPU, CPU) always. | |
69 | * 2. Written by CPU (using pwrite) | |
70 | * The pwrite function calls set_domain (CPU, CPU) and | |
71 | * this function does nothing (as nothing changes) | |
72 | * 3. Mapped by GTT | |
73 | * This function asserts that the object is not | |
74 | * currently in any GPU-based read or write domains | |
75 | * 4. Read by GPU | |
76 | * i915_gem_execbuffer calls set_domain (COMMAND, 0). | |
77 | * As write_domain is zero, this function adds in the | |
78 | * current read domains (CPU+COMMAND, 0). | |
79 | * flush_domains is set to CPU. | |
80 | * invalidate_domains is set to COMMAND | |
81 | * clflush is run to get data out of the CPU caches | |
82 | * then i915_dev_set_domain calls i915_gem_flush to | |
83 | * emit an MI_FLUSH and drm_agp_chipset_flush | |
84 | * 5. Unmapped from GTT | |
85 | * i915_gem_object_unbind calls set_domain (CPU, CPU) | |
86 | * flush_domains and invalidate_domains end up both zero | |
87 | * so no flushing/invalidating happens | |
88 | * 6. Freed | |
89 | * yay, done | |
90 | * | |
91 | * Case 2: The shared render buffer | |
92 | * | |
93 | * 1. Allocated | |
94 | * 2. Mapped to GTT | |
95 | * 3. Read/written by GPU | |
96 | * 4. set_domain to (CPU,CPU) | |
97 | * 5. Read/written by CPU | |
98 | * 6. Read/written by GPU | |
99 | * | |
100 | * 1. Allocated | |
101 | * Same as last example, (CPU, CPU) | |
102 | * 2. Mapped to GTT | |
103 | * Nothing changes (assertions find that it is not in the GPU) | |
104 | * 3. Read/written by GPU | |
105 | * execbuffer calls set_domain (RENDER, RENDER) | |
106 | * flush_domains gets CPU | |
107 | * invalidate_domains gets GPU | |
108 | * clflush (obj) | |
109 | * MI_FLUSH and drm_agp_chipset_flush | |
110 | * 4. set_domain (CPU, CPU) | |
111 | * flush_domains gets GPU | |
112 | * invalidate_domains gets CPU | |
113 | * wait_rendering (obj) to make sure all drawing is complete. | |
114 | * This will include an MI_FLUSH to get the data from GPU | |
115 | * to memory | |
116 | * clflush (obj) to invalidate the CPU cache | |
117 | * Another MI_FLUSH in i915_gem_flush (eliminate this somehow?) | |
118 | * 5. Read/written by CPU | |
119 | * cache lines are loaded and dirtied | |
120 | * 6. Read written by GPU | |
121 | * Same as last GPU access | |
122 | * | |
123 | * Case 3: The constant buffer | |
124 | * | |
125 | * 1. Allocated | |
126 | * 2. Written by CPU | |
127 | * 3. Read by GPU | |
128 | * 4. Updated (written) by CPU again | |
129 | * 5. Read by GPU | |
130 | * | |
131 | * 1. Allocated | |
132 | * (CPU, CPU) | |
133 | * 2. Written by CPU | |
134 | * (CPU, CPU) | |
135 | * 3. Read by GPU | |
136 | * (CPU+RENDER, 0) | |
137 | * flush_domains = CPU | |
138 | * invalidate_domains = RENDER | |
139 | * clflush (obj) | |
140 | * MI_FLUSH | |
141 | * drm_agp_chipset_flush | |
142 | * 4. Updated (written) by CPU again | |
143 | * (CPU, CPU) | |
144 | * flush_domains = 0 (no previous write domain) | |
145 | * invalidate_domains = 0 (no new read domains) | |
146 | * 5. Read by GPU | |
147 | * (CPU+RENDER, 0) | |
148 | * flush_domains = CPU | |
149 | * invalidate_domains = RENDER | |
150 | * clflush (obj) | |
151 | * MI_FLUSH | |
152 | * drm_agp_chipset_flush | |
153 | */ | |
154 | static void | |
155 | i915_gem_object_set_to_gpu_domain(struct drm_i915_gem_object *obj, | |
156 | struct intel_ring_buffer *ring, | |
157 | struct change_domains *cd) | |
158 | { | |
159 | uint32_t invalidate_domains = 0, flush_domains = 0; | |
160 | ||
161 | /* | |
162 | * If the object isn't moving to a new write domain, | |
163 | * let the object stay in multiple read domains | |
164 | */ | |
165 | if (obj->base.pending_write_domain == 0) | |
166 | obj->base.pending_read_domains |= obj->base.read_domains; | |
167 | ||
168 | /* | |
169 | * Flush the current write domain if | |
170 | * the new read domains don't match. Invalidate | |
171 | * any read domains which differ from the old | |
172 | * write domain | |
173 | */ | |
174 | if (obj->base.write_domain && | |
175 | (((obj->base.write_domain != obj->base.pending_read_domains || | |
176 | obj->ring != ring)) || | |
177 | (obj->fenced_gpu_access && !obj->pending_fenced_gpu_access))) { | |
178 | flush_domains |= obj->base.write_domain; | |
179 | invalidate_domains |= | |
180 | obj->base.pending_read_domains & ~obj->base.write_domain; | |
181 | } | |
182 | /* | |
183 | * Invalidate any read caches which may have | |
184 | * stale data. That is, any new read domains. | |
185 | */ | |
186 | invalidate_domains |= obj->base.pending_read_domains & ~obj->base.read_domains; | |
187 | if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) | |
188 | i915_gem_clflush_object(obj); | |
189 | ||
190 | /* blow away mappings if mapped through GTT */ | |
191 | if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_GTT) | |
192 | i915_gem_release_mmap(obj); | |
193 | ||
c59a333f CW |
194 | if (obj->base.pending_write_domain) |
195 | cd->flips |= atomic_read(&obj->pending_flip); | |
196 | ||
54cf91dc CW |
197 | /* The actual obj->write_domain will be updated with |
198 | * pending_write_domain after we emit the accumulated flush for all | |
199 | * of our domain changes in execbuffers (which clears objects' | |
200 | * write_domains). So if we have a current write domain that we | |
201 | * aren't changing, set pending_write_domain to that. | |
202 | */ | |
203 | if (flush_domains == 0 && obj->base.pending_write_domain == 0) | |
204 | obj->base.pending_write_domain = obj->base.write_domain; | |
205 | ||
206 | cd->invalidate_domains |= invalidate_domains; | |
207 | cd->flush_domains |= flush_domains; | |
208 | if (flush_domains & I915_GEM_GPU_DOMAINS) | |
209 | cd->flush_rings |= obj->ring->id; | |
210 | if (invalidate_domains & I915_GEM_GPU_DOMAINS) | |
211 | cd->flush_rings |= ring->id; | |
212 | } | |
213 | ||
67731b87 CW |
214 | struct eb_objects { |
215 | int and; | |
216 | struct hlist_head buckets[0]; | |
217 | }; | |
218 | ||
219 | static struct eb_objects * | |
220 | eb_create(int size) | |
221 | { | |
222 | struct eb_objects *eb; | |
223 | int count = PAGE_SIZE / sizeof(struct hlist_head) / 2; | |
224 | while (count > size) | |
225 | count >>= 1; | |
226 | eb = kzalloc(count*sizeof(struct hlist_head) + | |
227 | sizeof(struct eb_objects), | |
228 | GFP_KERNEL); | |
229 | if (eb == NULL) | |
230 | return eb; | |
231 | ||
232 | eb->and = count - 1; | |
233 | return eb; | |
234 | } | |
235 | ||
236 | static void | |
237 | eb_reset(struct eb_objects *eb) | |
238 | { | |
239 | memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head)); | |
240 | } | |
241 | ||
242 | static void | |
243 | eb_add_object(struct eb_objects *eb, struct drm_i915_gem_object *obj) | |
244 | { | |
245 | hlist_add_head(&obj->exec_node, | |
246 | &eb->buckets[obj->exec_handle & eb->and]); | |
247 | } | |
248 | ||
249 | static struct drm_i915_gem_object * | |
250 | eb_get_object(struct eb_objects *eb, unsigned long handle) | |
251 | { | |
252 | struct hlist_head *head; | |
253 | struct hlist_node *node; | |
254 | struct drm_i915_gem_object *obj; | |
255 | ||
256 | head = &eb->buckets[handle & eb->and]; | |
257 | hlist_for_each(node, head) { | |
258 | obj = hlist_entry(node, struct drm_i915_gem_object, exec_node); | |
259 | if (obj->exec_handle == handle) | |
260 | return obj; | |
261 | } | |
262 | ||
263 | return NULL; | |
264 | } | |
265 | ||
266 | static void | |
267 | eb_destroy(struct eb_objects *eb) | |
268 | { | |
269 | kfree(eb); | |
270 | } | |
271 | ||
54cf91dc CW |
272 | static int |
273 | i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj, | |
67731b87 | 274 | struct eb_objects *eb, |
54cf91dc CW |
275 | struct drm_i915_gem_relocation_entry *reloc) |
276 | { | |
277 | struct drm_device *dev = obj->base.dev; | |
278 | struct drm_gem_object *target_obj; | |
279 | uint32_t target_offset; | |
280 | int ret = -EINVAL; | |
281 | ||
67731b87 CW |
282 | /* we've already hold a reference to all valid objects */ |
283 | target_obj = &eb_get_object(eb, reloc->target_handle)->base; | |
284 | if (unlikely(target_obj == NULL)) | |
54cf91dc CW |
285 | return -ENOENT; |
286 | ||
287 | target_offset = to_intel_bo(target_obj)->gtt_offset; | |
288 | ||
54cf91dc CW |
289 | /* The target buffer should have appeared before us in the |
290 | * exec_object list, so it should have a GTT space bound by now. | |
291 | */ | |
b8f7ab17 | 292 | if (unlikely(target_offset == 0)) { |
54cf91dc CW |
293 | DRM_ERROR("No GTT space found for object %d\n", |
294 | reloc->target_handle); | |
67731b87 | 295 | return ret; |
54cf91dc CW |
296 | } |
297 | ||
298 | /* Validate that the target is in a valid r/w GPU domain */ | |
b8f7ab17 | 299 | if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) { |
54cf91dc CW |
300 | DRM_ERROR("reloc with multiple write domains: " |
301 | "obj %p target %d offset %d " | |
302 | "read %08x write %08x", | |
303 | obj, reloc->target_handle, | |
304 | (int) reloc->offset, | |
305 | reloc->read_domains, | |
306 | reloc->write_domain); | |
67731b87 | 307 | return ret; |
54cf91dc | 308 | } |
b8f7ab17 | 309 | if (unlikely((reloc->write_domain | reloc->read_domains) & I915_GEM_DOMAIN_CPU)) { |
54cf91dc CW |
310 | DRM_ERROR("reloc with read/write CPU domains: " |
311 | "obj %p target %d offset %d " | |
312 | "read %08x write %08x", | |
313 | obj, reloc->target_handle, | |
314 | (int) reloc->offset, | |
315 | reloc->read_domains, | |
316 | reloc->write_domain); | |
67731b87 | 317 | return ret; |
54cf91dc | 318 | } |
b8f7ab17 CW |
319 | if (unlikely(reloc->write_domain && target_obj->pending_write_domain && |
320 | reloc->write_domain != target_obj->pending_write_domain)) { | |
54cf91dc CW |
321 | DRM_ERROR("Write domain conflict: " |
322 | "obj %p target %d offset %d " | |
323 | "new %08x old %08x\n", | |
324 | obj, reloc->target_handle, | |
325 | (int) reloc->offset, | |
326 | reloc->write_domain, | |
327 | target_obj->pending_write_domain); | |
67731b87 | 328 | return ret; |
54cf91dc CW |
329 | } |
330 | ||
331 | target_obj->pending_read_domains |= reloc->read_domains; | |
332 | target_obj->pending_write_domain |= reloc->write_domain; | |
333 | ||
334 | /* If the relocation already has the right value in it, no | |
335 | * more work needs to be done. | |
336 | */ | |
337 | if (target_offset == reloc->presumed_offset) | |
67731b87 | 338 | return 0; |
54cf91dc CW |
339 | |
340 | /* Check that the relocation address is valid... */ | |
b8f7ab17 | 341 | if (unlikely(reloc->offset > obj->base.size - 4)) { |
54cf91dc CW |
342 | DRM_ERROR("Relocation beyond object bounds: " |
343 | "obj %p target %d offset %d size %d.\n", | |
344 | obj, reloc->target_handle, | |
345 | (int) reloc->offset, | |
346 | (int) obj->base.size); | |
67731b87 | 347 | return ret; |
54cf91dc | 348 | } |
b8f7ab17 | 349 | if (unlikely(reloc->offset & 3)) { |
54cf91dc CW |
350 | DRM_ERROR("Relocation not 4-byte aligned: " |
351 | "obj %p target %d offset %d.\n", | |
352 | obj, reloc->target_handle, | |
353 | (int) reloc->offset); | |
67731b87 | 354 | return ret; |
54cf91dc CW |
355 | } |
356 | ||
54cf91dc CW |
357 | reloc->delta += target_offset; |
358 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) { | |
359 | uint32_t page_offset = reloc->offset & ~PAGE_MASK; | |
360 | char *vaddr; | |
361 | ||
362 | vaddr = kmap_atomic(obj->pages[reloc->offset >> PAGE_SHIFT]); | |
363 | *(uint32_t *)(vaddr + page_offset) = reloc->delta; | |
364 | kunmap_atomic(vaddr); | |
365 | } else { | |
366 | struct drm_i915_private *dev_priv = dev->dev_private; | |
367 | uint32_t __iomem *reloc_entry; | |
368 | void __iomem *reloc_page; | |
369 | ||
370 | ret = i915_gem_object_set_to_gtt_domain(obj, 1); | |
371 | if (ret) | |
67731b87 | 372 | return ret; |
54cf91dc CW |
373 | |
374 | /* Map the page containing the relocation we're going to perform. */ | |
375 | reloc->offset += obj->gtt_offset; | |
376 | reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping, | |
377 | reloc->offset & PAGE_MASK); | |
378 | reloc_entry = (uint32_t __iomem *) | |
379 | (reloc_page + (reloc->offset & ~PAGE_MASK)); | |
380 | iowrite32(reloc->delta, reloc_entry); | |
381 | io_mapping_unmap_atomic(reloc_page); | |
382 | } | |
383 | ||
384 | /* and update the user's relocation entry */ | |
385 | reloc->presumed_offset = target_offset; | |
386 | ||
67731b87 | 387 | return 0; |
54cf91dc CW |
388 | } |
389 | ||
390 | static int | |
391 | i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj, | |
6fe4f140 | 392 | struct eb_objects *eb) |
54cf91dc CW |
393 | { |
394 | struct drm_i915_gem_relocation_entry __user *user_relocs; | |
6fe4f140 | 395 | struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; |
54cf91dc CW |
396 | int i, ret; |
397 | ||
398 | user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr; | |
399 | for (i = 0; i < entry->relocation_count; i++) { | |
400 | struct drm_i915_gem_relocation_entry reloc; | |
401 | ||
402 | if (__copy_from_user_inatomic(&reloc, | |
403 | user_relocs+i, | |
404 | sizeof(reloc))) | |
405 | return -EFAULT; | |
406 | ||
6fe4f140 | 407 | ret = i915_gem_execbuffer_relocate_entry(obj, eb, &reloc); |
54cf91dc CW |
408 | if (ret) |
409 | return ret; | |
410 | ||
411 | if (__copy_to_user_inatomic(&user_relocs[i].presumed_offset, | |
412 | &reloc.presumed_offset, | |
413 | sizeof(reloc.presumed_offset))) | |
414 | return -EFAULT; | |
415 | } | |
416 | ||
417 | return 0; | |
418 | } | |
419 | ||
420 | static int | |
421 | i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj, | |
67731b87 | 422 | struct eb_objects *eb, |
54cf91dc CW |
423 | struct drm_i915_gem_relocation_entry *relocs) |
424 | { | |
6fe4f140 | 425 | const struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; |
54cf91dc CW |
426 | int i, ret; |
427 | ||
428 | for (i = 0; i < entry->relocation_count; i++) { | |
6fe4f140 | 429 | ret = i915_gem_execbuffer_relocate_entry(obj, eb, &relocs[i]); |
54cf91dc CW |
430 | if (ret) |
431 | return ret; | |
432 | } | |
433 | ||
434 | return 0; | |
435 | } | |
436 | ||
437 | static int | |
438 | i915_gem_execbuffer_relocate(struct drm_device *dev, | |
67731b87 | 439 | struct eb_objects *eb, |
6fe4f140 | 440 | struct list_head *objects) |
54cf91dc | 441 | { |
432e58ed CW |
442 | struct drm_i915_gem_object *obj; |
443 | int ret; | |
54cf91dc | 444 | |
432e58ed | 445 | list_for_each_entry(obj, objects, exec_list) { |
6fe4f140 | 446 | ret = i915_gem_execbuffer_relocate_object(obj, eb); |
54cf91dc CW |
447 | if (ret) |
448 | return ret; | |
449 | } | |
450 | ||
451 | return 0; | |
452 | } | |
453 | ||
454 | static int | |
d9e86c0e | 455 | i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring, |
54cf91dc | 456 | struct drm_file *file, |
6fe4f140 | 457 | struct list_head *objects) |
54cf91dc | 458 | { |
432e58ed | 459 | struct drm_i915_gem_object *obj; |
432e58ed | 460 | int ret, retry; |
9b3826bf | 461 | bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4; |
6fe4f140 CW |
462 | struct list_head ordered_objects; |
463 | ||
464 | INIT_LIST_HEAD(&ordered_objects); | |
465 | while (!list_empty(objects)) { | |
466 | struct drm_i915_gem_exec_object2 *entry; | |
467 | bool need_fence, need_mappable; | |
468 | ||
469 | obj = list_first_entry(objects, | |
470 | struct drm_i915_gem_object, | |
471 | exec_list); | |
472 | entry = obj->exec_entry; | |
473 | ||
474 | need_fence = | |
475 | has_fenced_gpu_access && | |
476 | entry->flags & EXEC_OBJECT_NEEDS_FENCE && | |
477 | obj->tiling_mode != I915_TILING_NONE; | |
478 | need_mappable = | |
479 | entry->relocation_count ? true : need_fence; | |
480 | ||
481 | if (need_mappable) | |
482 | list_move(&obj->exec_list, &ordered_objects); | |
483 | else | |
484 | list_move_tail(&obj->exec_list, &ordered_objects); | |
595dad76 CW |
485 | |
486 | obj->base.pending_read_domains = 0; | |
487 | obj->base.pending_write_domain = 0; | |
6fe4f140 CW |
488 | } |
489 | list_splice(&ordered_objects, objects); | |
54cf91dc CW |
490 | |
491 | /* Attempt to pin all of the buffers into the GTT. | |
492 | * This is done in 3 phases: | |
493 | * | |
494 | * 1a. Unbind all objects that do not match the GTT constraints for | |
495 | * the execbuffer (fenceable, mappable, alignment etc). | |
496 | * 1b. Increment pin count for already bound objects. | |
497 | * 2. Bind new objects. | |
498 | * 3. Decrement pin count. | |
499 | * | |
500 | * This avoid unnecessary unbinding of later objects in order to makr | |
501 | * room for the earlier objects *unless* we need to defragment. | |
502 | */ | |
503 | retry = 0; | |
504 | do { | |
505 | ret = 0; | |
506 | ||
507 | /* Unbind any ill-fitting objects or pin. */ | |
432e58ed | 508 | list_for_each_entry(obj, objects, exec_list) { |
6fe4f140 | 509 | struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; |
54cf91dc | 510 | bool need_fence, need_mappable; |
6fe4f140 | 511 | if (!obj->gtt_space) |
54cf91dc CW |
512 | continue; |
513 | ||
514 | need_fence = | |
9b3826bf | 515 | has_fenced_gpu_access && |
54cf91dc CW |
516 | entry->flags & EXEC_OBJECT_NEEDS_FENCE && |
517 | obj->tiling_mode != I915_TILING_NONE; | |
518 | need_mappable = | |
519 | entry->relocation_count ? true : need_fence; | |
520 | ||
521 | if ((entry->alignment && obj->gtt_offset & (entry->alignment - 1)) || | |
522 | (need_mappable && !obj->map_and_fenceable)) | |
523 | ret = i915_gem_object_unbind(obj); | |
524 | else | |
525 | ret = i915_gem_object_pin(obj, | |
526 | entry->alignment, | |
527 | need_mappable); | |
432e58ed | 528 | if (ret) |
54cf91dc | 529 | goto err; |
432e58ed CW |
530 | |
531 | entry++; | |
54cf91dc CW |
532 | } |
533 | ||
534 | /* Bind fresh objects */ | |
432e58ed | 535 | list_for_each_entry(obj, objects, exec_list) { |
6fe4f140 | 536 | struct drm_i915_gem_exec_object2 *entry = obj->exec_entry; |
54cf91dc CW |
537 | bool need_fence; |
538 | ||
539 | need_fence = | |
9b3826bf | 540 | has_fenced_gpu_access && |
54cf91dc CW |
541 | entry->flags & EXEC_OBJECT_NEEDS_FENCE && |
542 | obj->tiling_mode != I915_TILING_NONE; | |
543 | ||
544 | if (!obj->gtt_space) { | |
545 | bool need_mappable = | |
546 | entry->relocation_count ? true : need_fence; | |
547 | ||
548 | ret = i915_gem_object_pin(obj, | |
549 | entry->alignment, | |
550 | need_mappable); | |
551 | if (ret) | |
552 | break; | |
553 | } | |
554 | ||
9b3826bf CW |
555 | if (has_fenced_gpu_access) { |
556 | if (need_fence) { | |
ce453d81 | 557 | ret = i915_gem_object_get_fence(obj, ring); |
9b3826bf CW |
558 | if (ret) |
559 | break; | |
560 | } else if (entry->flags & EXEC_OBJECT_NEEDS_FENCE && | |
561 | obj->tiling_mode == I915_TILING_NONE) { | |
562 | /* XXX pipelined! */ | |
563 | ret = i915_gem_object_put_fence(obj); | |
564 | if (ret) | |
565 | break; | |
566 | } | |
567 | obj->pending_fenced_gpu_access = need_fence; | |
54cf91dc CW |
568 | } |
569 | ||
570 | entry->offset = obj->gtt_offset; | |
571 | } | |
572 | ||
432e58ed CW |
573 | /* Decrement pin count for bound objects */ |
574 | list_for_each_entry(obj, objects, exec_list) { | |
54cf91dc CW |
575 | if (obj->gtt_space) |
576 | i915_gem_object_unpin(obj); | |
577 | } | |
578 | ||
579 | if (ret != -ENOSPC || retry > 1) | |
580 | return ret; | |
581 | ||
582 | /* First attempt, just clear anything that is purgeable. | |
583 | * Second attempt, clear the entire GTT. | |
584 | */ | |
d9e86c0e | 585 | ret = i915_gem_evict_everything(ring->dev, retry == 0); |
54cf91dc CW |
586 | if (ret) |
587 | return ret; | |
588 | ||
589 | retry++; | |
590 | } while (1); | |
432e58ed CW |
591 | |
592 | err: | |
602606a4 CW |
593 | obj = list_entry(obj->exec_list.prev, |
594 | struct drm_i915_gem_object, | |
595 | exec_list); | |
432e58ed CW |
596 | while (objects != &obj->exec_list) { |
597 | if (obj->gtt_space) | |
598 | i915_gem_object_unpin(obj); | |
599 | ||
600 | obj = list_entry(obj->exec_list.prev, | |
601 | struct drm_i915_gem_object, | |
602 | exec_list); | |
603 | } | |
604 | ||
605 | return ret; | |
54cf91dc CW |
606 | } |
607 | ||
608 | static int | |
609 | i915_gem_execbuffer_relocate_slow(struct drm_device *dev, | |
610 | struct drm_file *file, | |
d9e86c0e | 611 | struct intel_ring_buffer *ring, |
432e58ed | 612 | struct list_head *objects, |
67731b87 | 613 | struct eb_objects *eb, |
432e58ed | 614 | struct drm_i915_gem_exec_object2 *exec, |
54cf91dc CW |
615 | int count) |
616 | { | |
617 | struct drm_i915_gem_relocation_entry *reloc; | |
432e58ed | 618 | struct drm_i915_gem_object *obj; |
dd6864a4 | 619 | int *reloc_offset; |
54cf91dc CW |
620 | int i, total, ret; |
621 | ||
67731b87 | 622 | /* We may process another execbuffer during the unlock... */ |
36cf1742 | 623 | while (!list_empty(objects)) { |
67731b87 CW |
624 | obj = list_first_entry(objects, |
625 | struct drm_i915_gem_object, | |
626 | exec_list); | |
627 | list_del_init(&obj->exec_list); | |
628 | drm_gem_object_unreference(&obj->base); | |
629 | } | |
630 | ||
54cf91dc CW |
631 | mutex_unlock(&dev->struct_mutex); |
632 | ||
633 | total = 0; | |
634 | for (i = 0; i < count; i++) | |
432e58ed | 635 | total += exec[i].relocation_count; |
54cf91dc | 636 | |
dd6864a4 | 637 | reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset)); |
54cf91dc | 638 | reloc = drm_malloc_ab(total, sizeof(*reloc)); |
dd6864a4 CW |
639 | if (reloc == NULL || reloc_offset == NULL) { |
640 | drm_free_large(reloc); | |
641 | drm_free_large(reloc_offset); | |
54cf91dc CW |
642 | mutex_lock(&dev->struct_mutex); |
643 | return -ENOMEM; | |
644 | } | |
645 | ||
646 | total = 0; | |
647 | for (i = 0; i < count; i++) { | |
648 | struct drm_i915_gem_relocation_entry __user *user_relocs; | |
649 | ||
432e58ed | 650 | user_relocs = (void __user *)(uintptr_t)exec[i].relocs_ptr; |
54cf91dc CW |
651 | |
652 | if (copy_from_user(reloc+total, user_relocs, | |
432e58ed | 653 | exec[i].relocation_count * sizeof(*reloc))) { |
54cf91dc CW |
654 | ret = -EFAULT; |
655 | mutex_lock(&dev->struct_mutex); | |
656 | goto err; | |
657 | } | |
658 | ||
dd6864a4 | 659 | reloc_offset[i] = total; |
432e58ed | 660 | total += exec[i].relocation_count; |
54cf91dc CW |
661 | } |
662 | ||
663 | ret = i915_mutex_lock_interruptible(dev); | |
664 | if (ret) { | |
665 | mutex_lock(&dev->struct_mutex); | |
666 | goto err; | |
667 | } | |
668 | ||
67731b87 | 669 | /* reacquire the objects */ |
67731b87 CW |
670 | eb_reset(eb); |
671 | for (i = 0; i < count; i++) { | |
67731b87 CW |
672 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, |
673 | exec[i].handle)); | |
c8725226 | 674 | if (&obj->base == NULL) { |
67731b87 CW |
675 | DRM_ERROR("Invalid object handle %d at index %d\n", |
676 | exec[i].handle, i); | |
677 | ret = -ENOENT; | |
678 | goto err; | |
679 | } | |
680 | ||
681 | list_add_tail(&obj->exec_list, objects); | |
682 | obj->exec_handle = exec[i].handle; | |
6fe4f140 | 683 | obj->exec_entry = &exec[i]; |
67731b87 CW |
684 | eb_add_object(eb, obj); |
685 | } | |
686 | ||
6fe4f140 | 687 | ret = i915_gem_execbuffer_reserve(ring, file, objects); |
54cf91dc CW |
688 | if (ret) |
689 | goto err; | |
690 | ||
432e58ed | 691 | list_for_each_entry(obj, objects, exec_list) { |
dd6864a4 | 692 | int offset = obj->exec_entry - exec; |
67731b87 | 693 | ret = i915_gem_execbuffer_relocate_object_slow(obj, eb, |
dd6864a4 | 694 | reloc + reloc_offset[offset]); |
54cf91dc CW |
695 | if (ret) |
696 | goto err; | |
54cf91dc CW |
697 | } |
698 | ||
699 | /* Leave the user relocations as are, this is the painfully slow path, | |
700 | * and we want to avoid the complication of dropping the lock whilst | |
701 | * having buffers reserved in the aperture and so causing spurious | |
702 | * ENOSPC for random operations. | |
703 | */ | |
704 | ||
705 | err: | |
706 | drm_free_large(reloc); | |
dd6864a4 | 707 | drm_free_large(reloc_offset); |
54cf91dc CW |
708 | return ret; |
709 | } | |
710 | ||
88241785 | 711 | static int |
54cf91dc CW |
712 | i915_gem_execbuffer_flush(struct drm_device *dev, |
713 | uint32_t invalidate_domains, | |
714 | uint32_t flush_domains, | |
715 | uint32_t flush_rings) | |
716 | { | |
717 | drm_i915_private_t *dev_priv = dev->dev_private; | |
88241785 | 718 | int i, ret; |
54cf91dc CW |
719 | |
720 | if (flush_domains & I915_GEM_DOMAIN_CPU) | |
721 | intel_gtt_chipset_flush(); | |
722 | ||
63256ec5 CW |
723 | if (flush_domains & I915_GEM_DOMAIN_GTT) |
724 | wmb(); | |
725 | ||
54cf91dc | 726 | if ((flush_domains | invalidate_domains) & I915_GEM_GPU_DOMAINS) { |
1ec14ad3 | 727 | for (i = 0; i < I915_NUM_RINGS; i++) |
88241785 | 728 | if (flush_rings & (1 << i)) { |
db53a302 | 729 | ret = i915_gem_flush_ring(&dev_priv->ring[i], |
88241785 CW |
730 | invalidate_domains, |
731 | flush_domains); | |
732 | if (ret) | |
733 | return ret; | |
734 | } | |
54cf91dc | 735 | } |
88241785 CW |
736 | |
737 | return 0; | |
54cf91dc CW |
738 | } |
739 | ||
1ec14ad3 CW |
740 | static int |
741 | i915_gem_execbuffer_sync_rings(struct drm_i915_gem_object *obj, | |
742 | struct intel_ring_buffer *to) | |
743 | { | |
744 | struct intel_ring_buffer *from = obj->ring; | |
745 | u32 seqno; | |
746 | int ret, idx; | |
747 | ||
748 | if (from == NULL || to == from) | |
749 | return 0; | |
750 | ||
a1656b90 CW |
751 | /* XXX gpu semaphores are implicated in various hard hangs on SNB */ |
752 | if (INTEL_INFO(obj->base.dev)->gen < 6 || !i915_semaphores) | |
ce453d81 | 753 | return i915_gem_object_wait_rendering(obj); |
1ec14ad3 CW |
754 | |
755 | idx = intel_ring_sync_index(from, to); | |
756 | ||
757 | seqno = obj->last_rendering_seqno; | |
758 | if (seqno <= from->sync_seqno[idx]) | |
759 | return 0; | |
760 | ||
761 | if (seqno == from->outstanding_lazy_request) { | |
762 | struct drm_i915_gem_request *request; | |
763 | ||
764 | request = kzalloc(sizeof(*request), GFP_KERNEL); | |
765 | if (request == NULL) | |
766 | return -ENOMEM; | |
767 | ||
db53a302 | 768 | ret = i915_add_request(from, NULL, request); |
1ec14ad3 CW |
769 | if (ret) { |
770 | kfree(request); | |
771 | return ret; | |
772 | } | |
773 | ||
774 | seqno = request->seqno; | |
775 | } | |
776 | ||
777 | from->sync_seqno[idx] = seqno; | |
778 | return intel_ring_sync(to, from, seqno - 1); | |
779 | } | |
54cf91dc | 780 | |
c59a333f CW |
781 | static int |
782 | i915_gem_execbuffer_wait_for_flips(struct intel_ring_buffer *ring, u32 flips) | |
783 | { | |
784 | u32 plane, flip_mask; | |
785 | int ret; | |
786 | ||
787 | /* Check for any pending flips. As we only maintain a flip queue depth | |
788 | * of 1, we can simply insert a WAIT for the next display flip prior | |
789 | * to executing the batch and avoid stalling the CPU. | |
790 | */ | |
791 | ||
792 | for (plane = 0; flips >> plane; plane++) { | |
793 | if (((flips >> plane) & 1) == 0) | |
794 | continue; | |
795 | ||
796 | if (plane) | |
797 | flip_mask = MI_WAIT_FOR_PLANE_B_FLIP; | |
798 | else | |
799 | flip_mask = MI_WAIT_FOR_PLANE_A_FLIP; | |
800 | ||
801 | ret = intel_ring_begin(ring, 2); | |
802 | if (ret) | |
803 | return ret; | |
804 | ||
805 | intel_ring_emit(ring, MI_WAIT_FOR_EVENT | flip_mask); | |
806 | intel_ring_emit(ring, MI_NOOP); | |
807 | intel_ring_advance(ring); | |
808 | } | |
809 | ||
810 | return 0; | |
811 | } | |
812 | ||
813 | ||
54cf91dc | 814 | static int |
432e58ed CW |
815 | i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring, |
816 | struct list_head *objects) | |
54cf91dc | 817 | { |
432e58ed | 818 | struct drm_i915_gem_object *obj; |
54cf91dc | 819 | struct change_domains cd; |
432e58ed | 820 | int ret; |
54cf91dc | 821 | |
c59a333f | 822 | memset(&cd, 0, sizeof(cd)); |
432e58ed CW |
823 | list_for_each_entry(obj, objects, exec_list) |
824 | i915_gem_object_set_to_gpu_domain(obj, ring, &cd); | |
54cf91dc CW |
825 | |
826 | if (cd.invalidate_domains | cd.flush_domains) { | |
88241785 CW |
827 | ret = i915_gem_execbuffer_flush(ring->dev, |
828 | cd.invalidate_domains, | |
829 | cd.flush_domains, | |
830 | cd.flush_rings); | |
831 | if (ret) | |
832 | return ret; | |
54cf91dc CW |
833 | } |
834 | ||
c59a333f CW |
835 | if (cd.flips) { |
836 | ret = i915_gem_execbuffer_wait_for_flips(ring, cd.flips); | |
837 | if (ret) | |
838 | return ret; | |
839 | } | |
840 | ||
432e58ed | 841 | list_for_each_entry(obj, objects, exec_list) { |
1ec14ad3 CW |
842 | ret = i915_gem_execbuffer_sync_rings(obj, ring); |
843 | if (ret) | |
844 | return ret; | |
54cf91dc CW |
845 | } |
846 | ||
847 | return 0; | |
848 | } | |
849 | ||
432e58ed CW |
850 | static bool |
851 | i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec) | |
54cf91dc | 852 | { |
432e58ed | 853 | return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0; |
54cf91dc CW |
854 | } |
855 | ||
856 | static int | |
857 | validate_exec_list(struct drm_i915_gem_exec_object2 *exec, | |
858 | int count) | |
859 | { | |
860 | int i; | |
861 | ||
862 | for (i = 0; i < count; i++) { | |
863 | char __user *ptr = (char __user *)(uintptr_t)exec[i].relocs_ptr; | |
864 | int length; /* limited by fault_in_pages_readable() */ | |
865 | ||
866 | /* First check for malicious input causing overflow */ | |
867 | if (exec[i].relocation_count > | |
868 | INT_MAX / sizeof(struct drm_i915_gem_relocation_entry)) | |
869 | return -EINVAL; | |
870 | ||
871 | length = exec[i].relocation_count * | |
872 | sizeof(struct drm_i915_gem_relocation_entry); | |
873 | if (!access_ok(VERIFY_READ, ptr, length)) | |
874 | return -EFAULT; | |
875 | ||
876 | /* we may also need to update the presumed offsets */ | |
877 | if (!access_ok(VERIFY_WRITE, ptr, length)) | |
878 | return -EFAULT; | |
879 | ||
880 | if (fault_in_pages_readable(ptr, length)) | |
881 | return -EFAULT; | |
882 | } | |
883 | ||
884 | return 0; | |
885 | } | |
886 | ||
432e58ed CW |
887 | static void |
888 | i915_gem_execbuffer_move_to_active(struct list_head *objects, | |
1ec14ad3 CW |
889 | struct intel_ring_buffer *ring, |
890 | u32 seqno) | |
432e58ed CW |
891 | { |
892 | struct drm_i915_gem_object *obj; | |
893 | ||
894 | list_for_each_entry(obj, objects, exec_list) { | |
db53a302 CW |
895 | u32 old_read = obj->base.read_domains; |
896 | u32 old_write = obj->base.write_domain; | |
897 | ||
898 | ||
432e58ed CW |
899 | obj->base.read_domains = obj->base.pending_read_domains; |
900 | obj->base.write_domain = obj->base.pending_write_domain; | |
901 | obj->fenced_gpu_access = obj->pending_fenced_gpu_access; | |
902 | ||
1ec14ad3 | 903 | i915_gem_object_move_to_active(obj, ring, seqno); |
432e58ed CW |
904 | if (obj->base.write_domain) { |
905 | obj->dirty = 1; | |
87ca9c8a | 906 | obj->pending_gpu_write = true; |
432e58ed CW |
907 | list_move_tail(&obj->gpu_write_list, |
908 | &ring->gpu_write_list); | |
909 | intel_mark_busy(ring->dev, obj); | |
910 | } | |
911 | ||
db53a302 | 912 | trace_i915_gem_object_change_domain(obj, old_read, old_write); |
432e58ed CW |
913 | } |
914 | } | |
915 | ||
54cf91dc CW |
916 | static void |
917 | i915_gem_execbuffer_retire_commands(struct drm_device *dev, | |
432e58ed | 918 | struct drm_file *file, |
54cf91dc CW |
919 | struct intel_ring_buffer *ring) |
920 | { | |
432e58ed | 921 | struct drm_i915_gem_request *request; |
b72f3acb | 922 | u32 invalidate; |
54cf91dc | 923 | |
432e58ed CW |
924 | /* |
925 | * Ensure that the commands in the batch buffer are | |
926 | * finished before the interrupt fires. | |
927 | * | |
928 | * The sampler always gets flushed on i965 (sigh). | |
929 | */ | |
b72f3acb | 930 | invalidate = I915_GEM_DOMAIN_COMMAND; |
54cf91dc | 931 | if (INTEL_INFO(dev)->gen >= 4) |
b72f3acb CW |
932 | invalidate |= I915_GEM_DOMAIN_SAMPLER; |
933 | if (ring->flush(ring, invalidate, 0)) { | |
db53a302 | 934 | i915_gem_next_request_seqno(ring); |
b72f3acb CW |
935 | return; |
936 | } | |
54cf91dc | 937 | |
432e58ed CW |
938 | /* Add a breadcrumb for the completion of the batch buffer */ |
939 | request = kzalloc(sizeof(*request), GFP_KERNEL); | |
db53a302 CW |
940 | if (request == NULL || i915_add_request(ring, file, request)) { |
941 | i915_gem_next_request_seqno(ring); | |
432e58ed CW |
942 | kfree(request); |
943 | } | |
944 | } | |
54cf91dc CW |
945 | |
946 | static int | |
947 | i915_gem_do_execbuffer(struct drm_device *dev, void *data, | |
948 | struct drm_file *file, | |
949 | struct drm_i915_gem_execbuffer2 *args, | |
432e58ed | 950 | struct drm_i915_gem_exec_object2 *exec) |
54cf91dc CW |
951 | { |
952 | drm_i915_private_t *dev_priv = dev->dev_private; | |
432e58ed | 953 | struct list_head objects; |
67731b87 | 954 | struct eb_objects *eb; |
54cf91dc CW |
955 | struct drm_i915_gem_object *batch_obj; |
956 | struct drm_clip_rect *cliprects = NULL; | |
54cf91dc | 957 | struct intel_ring_buffer *ring; |
c4e7a414 | 958 | u32 exec_start, exec_len; |
1ec14ad3 | 959 | u32 seqno; |
72bfa19c | 960 | int ret, mode, i; |
54cf91dc | 961 | |
432e58ed CW |
962 | if (!i915_gem_check_execbuffer(args)) { |
963 | DRM_ERROR("execbuf with invalid offset/length\n"); | |
964 | return -EINVAL; | |
965 | } | |
966 | ||
967 | ret = validate_exec_list(exec, args->buffer_count); | |
54cf91dc CW |
968 | if (ret) |
969 | return ret; | |
970 | ||
54cf91dc CW |
971 | switch (args->flags & I915_EXEC_RING_MASK) { |
972 | case I915_EXEC_DEFAULT: | |
973 | case I915_EXEC_RENDER: | |
1ec14ad3 | 974 | ring = &dev_priv->ring[RCS]; |
54cf91dc CW |
975 | break; |
976 | case I915_EXEC_BSD: | |
977 | if (!HAS_BSD(dev)) { | |
978 | DRM_ERROR("execbuf with invalid ring (BSD)\n"); | |
979 | return -EINVAL; | |
980 | } | |
1ec14ad3 | 981 | ring = &dev_priv->ring[VCS]; |
54cf91dc CW |
982 | break; |
983 | case I915_EXEC_BLT: | |
984 | if (!HAS_BLT(dev)) { | |
985 | DRM_ERROR("execbuf with invalid ring (BLT)\n"); | |
986 | return -EINVAL; | |
987 | } | |
1ec14ad3 | 988 | ring = &dev_priv->ring[BCS]; |
54cf91dc CW |
989 | break; |
990 | default: | |
991 | DRM_ERROR("execbuf with unknown ring: %d\n", | |
992 | (int)(args->flags & I915_EXEC_RING_MASK)); | |
993 | return -EINVAL; | |
994 | } | |
995 | ||
72bfa19c CW |
996 | mode = args->flags & I915_EXEC_CONSTANTS_MASK; |
997 | switch (mode) { | |
998 | case I915_EXEC_CONSTANTS_REL_GENERAL: | |
999 | case I915_EXEC_CONSTANTS_ABSOLUTE: | |
1000 | case I915_EXEC_CONSTANTS_REL_SURFACE: | |
1001 | if (ring == &dev_priv->ring[RCS] && | |
1002 | mode != dev_priv->relative_constants_mode) { | |
1003 | if (INTEL_INFO(dev)->gen < 4) | |
1004 | return -EINVAL; | |
1005 | ||
1006 | if (INTEL_INFO(dev)->gen > 5 && | |
1007 | mode == I915_EXEC_CONSTANTS_REL_SURFACE) | |
1008 | return -EINVAL; | |
1009 | ||
1010 | ret = intel_ring_begin(ring, 4); | |
1011 | if (ret) | |
1012 | return ret; | |
1013 | ||
1014 | intel_ring_emit(ring, MI_NOOP); | |
1015 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1)); | |
1016 | intel_ring_emit(ring, INSTPM); | |
1017 | intel_ring_emit(ring, | |
1018 | I915_EXEC_CONSTANTS_MASK << 16 | mode); | |
1019 | intel_ring_advance(ring); | |
1020 | ||
1021 | dev_priv->relative_constants_mode = mode; | |
1022 | } | |
1023 | break; | |
1024 | default: | |
1025 | DRM_ERROR("execbuf with unknown constants: %d\n", mode); | |
1026 | return -EINVAL; | |
1027 | } | |
1028 | ||
54cf91dc CW |
1029 | if (args->buffer_count < 1) { |
1030 | DRM_ERROR("execbuf with %d buffers\n", args->buffer_count); | |
1031 | return -EINVAL; | |
1032 | } | |
54cf91dc CW |
1033 | |
1034 | if (args->num_cliprects != 0) { | |
1ec14ad3 | 1035 | if (ring != &dev_priv->ring[RCS]) { |
c4e7a414 CW |
1036 | DRM_ERROR("clip rectangles are only valid with the render ring\n"); |
1037 | return -EINVAL; | |
1038 | } | |
1039 | ||
432e58ed | 1040 | cliprects = kmalloc(args->num_cliprects * sizeof(*cliprects), |
54cf91dc CW |
1041 | GFP_KERNEL); |
1042 | if (cliprects == NULL) { | |
1043 | ret = -ENOMEM; | |
1044 | goto pre_mutex_err; | |
1045 | } | |
1046 | ||
432e58ed CW |
1047 | if (copy_from_user(cliprects, |
1048 | (struct drm_clip_rect __user *)(uintptr_t) | |
1049 | args->cliprects_ptr, | |
1050 | sizeof(*cliprects)*args->num_cliprects)) { | |
54cf91dc CW |
1051 | ret = -EFAULT; |
1052 | goto pre_mutex_err; | |
1053 | } | |
1054 | } | |
1055 | ||
54cf91dc CW |
1056 | ret = i915_mutex_lock_interruptible(dev); |
1057 | if (ret) | |
1058 | goto pre_mutex_err; | |
1059 | ||
1060 | if (dev_priv->mm.suspended) { | |
1061 | mutex_unlock(&dev->struct_mutex); | |
1062 | ret = -EBUSY; | |
1063 | goto pre_mutex_err; | |
1064 | } | |
1065 | ||
67731b87 CW |
1066 | eb = eb_create(args->buffer_count); |
1067 | if (eb == NULL) { | |
1068 | mutex_unlock(&dev->struct_mutex); | |
1069 | ret = -ENOMEM; | |
1070 | goto pre_mutex_err; | |
1071 | } | |
1072 | ||
54cf91dc | 1073 | /* Look up object handles */ |
432e58ed | 1074 | INIT_LIST_HEAD(&objects); |
54cf91dc CW |
1075 | for (i = 0; i < args->buffer_count; i++) { |
1076 | struct drm_i915_gem_object *obj; | |
1077 | ||
432e58ed CW |
1078 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, |
1079 | exec[i].handle)); | |
c8725226 | 1080 | if (&obj->base == NULL) { |
54cf91dc | 1081 | DRM_ERROR("Invalid object handle %d at index %d\n", |
432e58ed | 1082 | exec[i].handle, i); |
54cf91dc | 1083 | /* prevent error path from reading uninitialized data */ |
54cf91dc CW |
1084 | ret = -ENOENT; |
1085 | goto err; | |
1086 | } | |
54cf91dc | 1087 | |
432e58ed CW |
1088 | if (!list_empty(&obj->exec_list)) { |
1089 | DRM_ERROR("Object %p [handle %d, index %d] appears more than once in object list\n", | |
1090 | obj, exec[i].handle, i); | |
54cf91dc CW |
1091 | ret = -EINVAL; |
1092 | goto err; | |
1093 | } | |
432e58ed CW |
1094 | |
1095 | list_add_tail(&obj->exec_list, &objects); | |
67731b87 | 1096 | obj->exec_handle = exec[i].handle; |
6fe4f140 | 1097 | obj->exec_entry = &exec[i]; |
67731b87 | 1098 | eb_add_object(eb, obj); |
54cf91dc CW |
1099 | } |
1100 | ||
6fe4f140 CW |
1101 | /* take note of the batch buffer before we might reorder the lists */ |
1102 | batch_obj = list_entry(objects.prev, | |
1103 | struct drm_i915_gem_object, | |
1104 | exec_list); | |
1105 | ||
54cf91dc | 1106 | /* Move the objects en-masse into the GTT, evicting if necessary. */ |
6fe4f140 | 1107 | ret = i915_gem_execbuffer_reserve(ring, file, &objects); |
54cf91dc CW |
1108 | if (ret) |
1109 | goto err; | |
1110 | ||
1111 | /* The objects are in their final locations, apply the relocations. */ | |
6fe4f140 | 1112 | ret = i915_gem_execbuffer_relocate(dev, eb, &objects); |
54cf91dc CW |
1113 | if (ret) { |
1114 | if (ret == -EFAULT) { | |
d9e86c0e | 1115 | ret = i915_gem_execbuffer_relocate_slow(dev, file, ring, |
67731b87 CW |
1116 | &objects, eb, |
1117 | exec, | |
54cf91dc CW |
1118 | args->buffer_count); |
1119 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); | |
1120 | } | |
1121 | if (ret) | |
1122 | goto err; | |
1123 | } | |
1124 | ||
1125 | /* Set the pending read domains for the batch buffer to COMMAND */ | |
54cf91dc CW |
1126 | if (batch_obj->base.pending_write_domain) { |
1127 | DRM_ERROR("Attempting to use self-modifying batch buffer\n"); | |
1128 | ret = -EINVAL; | |
1129 | goto err; | |
1130 | } | |
1131 | batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND; | |
1132 | ||
432e58ed CW |
1133 | ret = i915_gem_execbuffer_move_to_gpu(ring, &objects); |
1134 | if (ret) | |
54cf91dc | 1135 | goto err; |
54cf91dc | 1136 | |
db53a302 | 1137 | seqno = i915_gem_next_request_seqno(ring); |
076e2c0e | 1138 | for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++) { |
1ec14ad3 CW |
1139 | if (seqno < ring->sync_seqno[i]) { |
1140 | /* The GPU can not handle its semaphore value wrapping, | |
1141 | * so every billion or so execbuffers, we need to stall | |
1142 | * the GPU in order to reset the counters. | |
1143 | */ | |
1144 | ret = i915_gpu_idle(dev); | |
1145 | if (ret) | |
1146 | goto err; | |
1147 | ||
1148 | BUG_ON(ring->sync_seqno[i]); | |
1149 | } | |
1150 | } | |
1151 | ||
db53a302 CW |
1152 | trace_i915_gem_ring_dispatch(ring, seqno); |
1153 | ||
c4e7a414 CW |
1154 | exec_start = batch_obj->gtt_offset + args->batch_start_offset; |
1155 | exec_len = args->batch_len; | |
1156 | if (cliprects) { | |
1157 | for (i = 0; i < args->num_cliprects; i++) { | |
1158 | ret = i915_emit_box(dev, &cliprects[i], | |
1159 | args->DR1, args->DR4); | |
1160 | if (ret) | |
1161 | goto err; | |
1162 | ||
1163 | ret = ring->dispatch_execbuffer(ring, | |
1164 | exec_start, exec_len); | |
1165 | if (ret) | |
1166 | goto err; | |
1167 | } | |
1168 | } else { | |
1169 | ret = ring->dispatch_execbuffer(ring, exec_start, exec_len); | |
1170 | if (ret) | |
1171 | goto err; | |
1172 | } | |
54cf91dc | 1173 | |
1ec14ad3 | 1174 | i915_gem_execbuffer_move_to_active(&objects, ring, seqno); |
432e58ed | 1175 | i915_gem_execbuffer_retire_commands(dev, file, ring); |
54cf91dc CW |
1176 | |
1177 | err: | |
67731b87 | 1178 | eb_destroy(eb); |
432e58ed CW |
1179 | while (!list_empty(&objects)) { |
1180 | struct drm_i915_gem_object *obj; | |
1181 | ||
1182 | obj = list_first_entry(&objects, | |
1183 | struct drm_i915_gem_object, | |
1184 | exec_list); | |
1185 | list_del_init(&obj->exec_list); | |
1186 | drm_gem_object_unreference(&obj->base); | |
54cf91dc CW |
1187 | } |
1188 | ||
1189 | mutex_unlock(&dev->struct_mutex); | |
1190 | ||
1191 | pre_mutex_err: | |
54cf91dc | 1192 | kfree(cliprects); |
54cf91dc CW |
1193 | return ret; |
1194 | } | |
1195 | ||
1196 | /* | |
1197 | * Legacy execbuffer just creates an exec2 list from the original exec object | |
1198 | * list array and passes it to the real function. | |
1199 | */ | |
1200 | int | |
1201 | i915_gem_execbuffer(struct drm_device *dev, void *data, | |
1202 | struct drm_file *file) | |
1203 | { | |
1204 | struct drm_i915_gem_execbuffer *args = data; | |
1205 | struct drm_i915_gem_execbuffer2 exec2; | |
1206 | struct drm_i915_gem_exec_object *exec_list = NULL; | |
1207 | struct drm_i915_gem_exec_object2 *exec2_list = NULL; | |
1208 | int ret, i; | |
1209 | ||
54cf91dc CW |
1210 | if (args->buffer_count < 1) { |
1211 | DRM_ERROR("execbuf with %d buffers\n", args->buffer_count); | |
1212 | return -EINVAL; | |
1213 | } | |
1214 | ||
1215 | /* Copy in the exec list from userland */ | |
1216 | exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count); | |
1217 | exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count); | |
1218 | if (exec_list == NULL || exec2_list == NULL) { | |
1219 | DRM_ERROR("Failed to allocate exec list for %d buffers\n", | |
1220 | args->buffer_count); | |
1221 | drm_free_large(exec_list); | |
1222 | drm_free_large(exec2_list); | |
1223 | return -ENOMEM; | |
1224 | } | |
1225 | ret = copy_from_user(exec_list, | |
1226 | (struct drm_i915_relocation_entry __user *) | |
1227 | (uintptr_t) args->buffers_ptr, | |
1228 | sizeof(*exec_list) * args->buffer_count); | |
1229 | if (ret != 0) { | |
1230 | DRM_ERROR("copy %d exec entries failed %d\n", | |
1231 | args->buffer_count, ret); | |
1232 | drm_free_large(exec_list); | |
1233 | drm_free_large(exec2_list); | |
1234 | return -EFAULT; | |
1235 | } | |
1236 | ||
1237 | for (i = 0; i < args->buffer_count; i++) { | |
1238 | exec2_list[i].handle = exec_list[i].handle; | |
1239 | exec2_list[i].relocation_count = exec_list[i].relocation_count; | |
1240 | exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr; | |
1241 | exec2_list[i].alignment = exec_list[i].alignment; | |
1242 | exec2_list[i].offset = exec_list[i].offset; | |
1243 | if (INTEL_INFO(dev)->gen < 4) | |
1244 | exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE; | |
1245 | else | |
1246 | exec2_list[i].flags = 0; | |
1247 | } | |
1248 | ||
1249 | exec2.buffers_ptr = args->buffers_ptr; | |
1250 | exec2.buffer_count = args->buffer_count; | |
1251 | exec2.batch_start_offset = args->batch_start_offset; | |
1252 | exec2.batch_len = args->batch_len; | |
1253 | exec2.DR1 = args->DR1; | |
1254 | exec2.DR4 = args->DR4; | |
1255 | exec2.num_cliprects = args->num_cliprects; | |
1256 | exec2.cliprects_ptr = args->cliprects_ptr; | |
1257 | exec2.flags = I915_EXEC_RENDER; | |
1258 | ||
1259 | ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list); | |
1260 | if (!ret) { | |
1261 | /* Copy the new buffer offsets back to the user's exec list. */ | |
1262 | for (i = 0; i < args->buffer_count; i++) | |
1263 | exec_list[i].offset = exec2_list[i].offset; | |
1264 | /* ... and back out to userspace */ | |
1265 | ret = copy_to_user((struct drm_i915_relocation_entry __user *) | |
1266 | (uintptr_t) args->buffers_ptr, | |
1267 | exec_list, | |
1268 | sizeof(*exec_list) * args->buffer_count); | |
1269 | if (ret) { | |
1270 | ret = -EFAULT; | |
1271 | DRM_ERROR("failed to copy %d exec entries " | |
1272 | "back to user (%d)\n", | |
1273 | args->buffer_count, ret); | |
1274 | } | |
1275 | } | |
1276 | ||
1277 | drm_free_large(exec_list); | |
1278 | drm_free_large(exec2_list); | |
1279 | return ret; | |
1280 | } | |
1281 | ||
1282 | int | |
1283 | i915_gem_execbuffer2(struct drm_device *dev, void *data, | |
1284 | struct drm_file *file) | |
1285 | { | |
1286 | struct drm_i915_gem_execbuffer2 *args = data; | |
1287 | struct drm_i915_gem_exec_object2 *exec2_list = NULL; | |
1288 | int ret; | |
1289 | ||
54cf91dc CW |
1290 | if (args->buffer_count < 1) { |
1291 | DRM_ERROR("execbuf2 with %d buffers\n", args->buffer_count); | |
1292 | return -EINVAL; | |
1293 | } | |
1294 | ||
8408c282 CW |
1295 | exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count, |
1296 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); | |
1297 | if (exec2_list == NULL) | |
1298 | exec2_list = drm_malloc_ab(sizeof(*exec2_list), | |
1299 | args->buffer_count); | |
54cf91dc CW |
1300 | if (exec2_list == NULL) { |
1301 | DRM_ERROR("Failed to allocate exec list for %d buffers\n", | |
1302 | args->buffer_count); | |
1303 | return -ENOMEM; | |
1304 | } | |
1305 | ret = copy_from_user(exec2_list, | |
1306 | (struct drm_i915_relocation_entry __user *) | |
1307 | (uintptr_t) args->buffers_ptr, | |
1308 | sizeof(*exec2_list) * args->buffer_count); | |
1309 | if (ret != 0) { | |
1310 | DRM_ERROR("copy %d exec entries failed %d\n", | |
1311 | args->buffer_count, ret); | |
1312 | drm_free_large(exec2_list); | |
1313 | return -EFAULT; | |
1314 | } | |
1315 | ||
1316 | ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list); | |
1317 | if (!ret) { | |
1318 | /* Copy the new buffer offsets back to the user's exec list. */ | |
1319 | ret = copy_to_user((struct drm_i915_relocation_entry __user *) | |
1320 | (uintptr_t) args->buffers_ptr, | |
1321 | exec2_list, | |
1322 | sizeof(*exec2_list) * args->buffer_count); | |
1323 | if (ret) { | |
1324 | ret = -EFAULT; | |
1325 | DRM_ERROR("failed to copy %d exec entries " | |
1326 | "back to user (%d)\n", | |
1327 | args->buffer_count, ret); | |
1328 | } | |
1329 | } | |
1330 | ||
1331 | drm_free_large(exec2_list); | |
1332 | return ret; | |
1333 | } |