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 <drm/drmP.h>
30 #include <drm/i915_drm.h>
31 #include "i915_drv.h"
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/dma_remapping.h>
35
36 #define __EXEC_OBJECT_HAS_PIN (1<<31)
37 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
38 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
39 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
40
41 #define BATCH_OFFSET_BIAS (256*1024)
42
43 struct eb_vmas {
44 struct list_head vmas;
45 int and;
46 union {
47 struct i915_vma *lut[0];
48 struct hlist_head buckets[0];
49 };
50 };
51
52 static struct eb_vmas *
eb_create(struct drm_i915_gem_execbuffer2 * args)53 eb_create(struct drm_i915_gem_execbuffer2 *args)
54 {
55 struct eb_vmas *eb = NULL;
56
57 if (args->flags & I915_EXEC_HANDLE_LUT) {
58 unsigned size = args->buffer_count;
59 size *= sizeof(struct i915_vma *);
60 size += sizeof(struct eb_vmas);
61 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
62 }
63
64 if (eb == NULL) {
65 unsigned size = args->buffer_count;
66 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
67 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
68 while (count > 2*size)
69 count >>= 1;
70 eb = kzalloc(count*sizeof(struct hlist_head) +
71 sizeof(struct eb_vmas),
72 GFP_TEMPORARY);
73 if (eb == NULL)
74 return eb;
75
76 eb->and = count - 1;
77 } else
78 eb->and = -args->buffer_count;
79
80 INIT_LIST_HEAD(&eb->vmas);
81 return eb;
82 }
83
84 static void
eb_reset(struct eb_vmas * eb)85 eb_reset(struct eb_vmas *eb)
86 {
87 if (eb->and >= 0)
88 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
89 }
90
91 static int
eb_lookup_vmas(struct eb_vmas * eb,struct drm_i915_gem_exec_object2 * exec,const struct drm_i915_gem_execbuffer2 * args,struct i915_address_space * vm,struct drm_file * file)92 eb_lookup_vmas(struct eb_vmas *eb,
93 struct drm_i915_gem_exec_object2 *exec,
94 const struct drm_i915_gem_execbuffer2 *args,
95 struct i915_address_space *vm,
96 struct drm_file *file)
97 {
98 struct drm_i915_gem_object *obj;
99 struct list_head objects;
100 int i, ret;
101
102 INIT_LIST_HEAD(&objects);
103 spin_lock(&file->table_lock);
104 /* Grab a reference to the object and release the lock so we can lookup
105 * or create the VMA without using GFP_ATOMIC */
106 for (i = 0; i < args->buffer_count; i++) {
107 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
108 if (obj == NULL) {
109 spin_unlock(&file->table_lock);
110 DRM_DEBUG("Invalid object handle %d at index %d\n",
111 exec[i].handle, i);
112 ret = -ENOENT;
113 goto err;
114 }
115
116 if (!list_empty(&obj->obj_exec_link)) {
117 spin_unlock(&file->table_lock);
118 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
119 obj, exec[i].handle, i);
120 ret = -EINVAL;
121 goto err;
122 }
123
124 drm_gem_object_reference(&obj->base);
125 list_add_tail(&obj->obj_exec_link, &objects);
126 }
127 spin_unlock(&file->table_lock);
128
129 i = 0;
130 while (!list_empty(&objects)) {
131 struct i915_vma *vma;
132
133 obj = list_first_entry(&objects,
134 struct drm_i915_gem_object,
135 obj_exec_link);
136
137 /*
138 * NOTE: We can leak any vmas created here when something fails
139 * later on. But that's no issue since vma_unbind can deal with
140 * vmas which are not actually bound. And since only
141 * lookup_or_create exists as an interface to get at the vma
142 * from the (obj, vm) we don't run the risk of creating
143 * duplicated vmas for the same vm.
144 */
145 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
146 if (IS_ERR(vma)) {
147 DRM_DEBUG("Failed to lookup VMA\n");
148 ret = PTR_ERR(vma);
149 goto err;
150 }
151
152 /* Transfer ownership from the objects list to the vmas list. */
153 list_add_tail(&vma->exec_list, &eb->vmas);
154 list_del_init(&obj->obj_exec_link);
155
156 vma->exec_entry = &exec[i];
157 if (eb->and < 0) {
158 eb->lut[i] = vma;
159 } else {
160 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
161 vma->exec_handle = handle;
162 hlist_add_head(&vma->exec_node,
163 &eb->buckets[handle & eb->and]);
164 }
165 ++i;
166 }
167
168 return 0;
169
170
171 err:
172 while (!list_empty(&objects)) {
173 obj = list_first_entry(&objects,
174 struct drm_i915_gem_object,
175 obj_exec_link);
176 list_del_init(&obj->obj_exec_link);
177 drm_gem_object_unreference(&obj->base);
178 }
179 /*
180 * Objects already transfered to the vmas list will be unreferenced by
181 * eb_destroy.
182 */
183
184 return ret;
185 }
186
eb_get_vma(struct eb_vmas * eb,unsigned long handle)187 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
188 {
189 if (eb->and < 0) {
190 if (handle >= -eb->and)
191 return NULL;
192 return eb->lut[handle];
193 } else {
194 struct hlist_head *head;
195 struct hlist_node *node;
196
197 head = &eb->buckets[handle & eb->and];
198 hlist_for_each(node, head) {
199 struct i915_vma *vma;
200
201 vma = hlist_entry(node, struct i915_vma, exec_node);
202 if (vma->exec_handle == handle)
203 return vma;
204 }
205 return NULL;
206 }
207 }
208
209 static void
i915_gem_execbuffer_unreserve_vma(struct i915_vma * vma)210 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
211 {
212 struct drm_i915_gem_exec_object2 *entry;
213 struct drm_i915_gem_object *obj = vma->obj;
214
215 if (!drm_mm_node_allocated(&vma->node))
216 return;
217
218 entry = vma->exec_entry;
219
220 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
221 i915_gem_object_unpin_fence(obj);
222
223 if (entry->flags & __EXEC_OBJECT_HAS_PIN)
224 vma->pin_count--;
225
226 entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
227 }
228
eb_destroy(struct eb_vmas * eb)229 static void eb_destroy(struct eb_vmas *eb)
230 {
231 while (!list_empty(&eb->vmas)) {
232 struct i915_vma *vma;
233
234 vma = list_first_entry(&eb->vmas,
235 struct i915_vma,
236 exec_list);
237 list_del_init(&vma->exec_list);
238 i915_gem_execbuffer_unreserve_vma(vma);
239 drm_gem_object_unreference(&vma->obj->base);
240 }
241 kfree(eb);
242 }
243
use_cpu_reloc(struct drm_i915_gem_object * obj)244 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
245 {
246 return (HAS_LLC(obj->base.dev) ||
247 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
248 !obj->map_and_fenceable ||
249 obj->cache_level != I915_CACHE_NONE);
250 }
251
252 static int
relocate_entry_cpu(struct drm_i915_gem_object * obj,struct drm_i915_gem_relocation_entry * reloc,uint64_t target_offset)253 relocate_entry_cpu(struct drm_i915_gem_object *obj,
254 struct drm_i915_gem_relocation_entry *reloc,
255 uint64_t target_offset)
256 {
257 struct drm_device *dev = obj->base.dev;
258 uint32_t page_offset = offset_in_page(reloc->offset);
259 uint64_t delta = reloc->delta + target_offset;
260 char *vaddr;
261 int ret;
262
263 ret = i915_gem_object_set_to_cpu_domain(obj, true);
264 if (ret)
265 return ret;
266
267 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
268 reloc->offset >> PAGE_SHIFT));
269 *(uint32_t *)(vaddr + page_offset) = lower_32_bits(delta);
270
271 if (INTEL_INFO(dev)->gen >= 8) {
272 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
273
274 if (page_offset == 0) {
275 kunmap_atomic(vaddr);
276 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
277 (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
278 }
279
280 *(uint32_t *)(vaddr + page_offset) = upper_32_bits(delta);
281 }
282
283 kunmap_atomic(vaddr);
284
285 return 0;
286 }
287
288 static int
relocate_entry_gtt(struct drm_i915_gem_object * obj,struct drm_i915_gem_relocation_entry * reloc,uint64_t target_offset)289 relocate_entry_gtt(struct drm_i915_gem_object *obj,
290 struct drm_i915_gem_relocation_entry *reloc,
291 uint64_t target_offset)
292 {
293 struct drm_device *dev = obj->base.dev;
294 struct drm_i915_private *dev_priv = dev->dev_private;
295 uint64_t delta = reloc->delta + target_offset;
296 uint64_t offset;
297 void __iomem *reloc_page;
298 int ret;
299
300 ret = i915_gem_object_set_to_gtt_domain(obj, true);
301 if (ret)
302 return ret;
303
304 ret = i915_gem_object_put_fence(obj);
305 if (ret)
306 return ret;
307
308 /* Map the page containing the relocation we're going to perform. */
309 offset = i915_gem_obj_ggtt_offset(obj);
310 offset += reloc->offset;
311 reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
312 offset & PAGE_MASK);
313 iowrite32(lower_32_bits(delta), reloc_page + offset_in_page(offset));
314
315 if (INTEL_INFO(dev)->gen >= 8) {
316 offset += sizeof(uint32_t);
317
318 if (offset_in_page(offset) == 0) {
319 io_mapping_unmap_atomic(reloc_page);
320 reloc_page =
321 io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
322 offset);
323 }
324
325 iowrite32(upper_32_bits(delta),
326 reloc_page + offset_in_page(offset));
327 }
328
329 io_mapping_unmap_atomic(reloc_page);
330
331 return 0;
332 }
333
334 static int
i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object * obj,struct eb_vmas * eb,struct drm_i915_gem_relocation_entry * reloc)335 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
336 struct eb_vmas *eb,
337 struct drm_i915_gem_relocation_entry *reloc)
338 {
339 struct drm_device *dev = obj->base.dev;
340 struct drm_gem_object *target_obj;
341 struct drm_i915_gem_object *target_i915_obj;
342 struct i915_vma *target_vma;
343 uint64_t target_offset;
344 int ret;
345
346 /* we've already hold a reference to all valid objects */
347 target_vma = eb_get_vma(eb, reloc->target_handle);
348 if (unlikely(target_vma == NULL))
349 return -ENOENT;
350 target_i915_obj = target_vma->obj;
351 target_obj = &target_vma->obj->base;
352
353 target_offset = target_vma->node.start;
354
355 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
356 * pipe_control writes because the gpu doesn't properly redirect them
357 * through the ppgtt for non_secure batchbuffers. */
358 if (unlikely(IS_GEN6(dev) &&
359 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
360 !target_i915_obj->has_global_gtt_mapping)) {
361 struct i915_vma *vma =
362 list_first_entry(&target_i915_obj->vma_list,
363 typeof(*vma), vma_link);
364 vma->bind_vma(vma, target_i915_obj->cache_level, GLOBAL_BIND);
365 }
366
367 /* Validate that the target is in a valid r/w GPU domain */
368 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
369 DRM_DEBUG("reloc with multiple write domains: "
370 "obj %p target %d offset %d "
371 "read %08x write %08x",
372 obj, reloc->target_handle,
373 (int) reloc->offset,
374 reloc->read_domains,
375 reloc->write_domain);
376 return -EINVAL;
377 }
378 if (unlikely((reloc->write_domain | reloc->read_domains)
379 & ~I915_GEM_GPU_DOMAINS)) {
380 DRM_DEBUG("reloc with read/write non-GPU domains: "
381 "obj %p target %d offset %d "
382 "read %08x write %08x",
383 obj, reloc->target_handle,
384 (int) reloc->offset,
385 reloc->read_domains,
386 reloc->write_domain);
387 return -EINVAL;
388 }
389
390 target_obj->pending_read_domains |= reloc->read_domains;
391 target_obj->pending_write_domain |= reloc->write_domain;
392
393 /* If the relocation already has the right value in it, no
394 * more work needs to be done.
395 */
396 if (target_offset == reloc->presumed_offset)
397 return 0;
398
399 /* Check that the relocation address is valid... */
400 if (unlikely(reloc->offset >
401 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
402 DRM_DEBUG("Relocation beyond object bounds: "
403 "obj %p target %d offset %d size %d.\n",
404 obj, reloc->target_handle,
405 (int) reloc->offset,
406 (int) obj->base.size);
407 return -EINVAL;
408 }
409 if (unlikely(reloc->offset & 3)) {
410 DRM_DEBUG("Relocation not 4-byte aligned: "
411 "obj %p target %d offset %d.\n",
412 obj, reloc->target_handle,
413 (int) reloc->offset);
414 return -EINVAL;
415 }
416
417 /* We can't wait for rendering with pagefaults disabled */
418 if (obj->active && in_atomic())
419 return -EFAULT;
420
421 if (use_cpu_reloc(obj))
422 ret = relocate_entry_cpu(obj, reloc, target_offset);
423 else
424 ret = relocate_entry_gtt(obj, reloc, target_offset);
425
426 if (ret)
427 return ret;
428
429 /* and update the user's relocation entry */
430 reloc->presumed_offset = target_offset;
431
432 return 0;
433 }
434
435 static int
i915_gem_execbuffer_relocate_vma(struct i915_vma * vma,struct eb_vmas * eb)436 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
437 struct eb_vmas *eb)
438 {
439 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
440 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
441 struct drm_i915_gem_relocation_entry __user *user_relocs;
442 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
443 int remain, ret;
444
445 user_relocs = to_user_ptr(entry->relocs_ptr);
446
447 remain = entry->relocation_count;
448 while (remain) {
449 struct drm_i915_gem_relocation_entry *r = stack_reloc;
450 int count = remain;
451 if (count > ARRAY_SIZE(stack_reloc))
452 count = ARRAY_SIZE(stack_reloc);
453 remain -= count;
454
455 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
456 return -EFAULT;
457
458 do {
459 u64 offset = r->presumed_offset;
460
461 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r);
462 if (ret)
463 return ret;
464
465 if (r->presumed_offset != offset &&
466 __copy_to_user_inatomic(&user_relocs->presumed_offset,
467 &r->presumed_offset,
468 sizeof(r->presumed_offset))) {
469 return -EFAULT;
470 }
471
472 user_relocs++;
473 r++;
474 } while (--count);
475 }
476
477 return 0;
478 #undef N_RELOC
479 }
480
481 static int
i915_gem_execbuffer_relocate_vma_slow(struct i915_vma * vma,struct eb_vmas * eb,struct drm_i915_gem_relocation_entry * relocs)482 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
483 struct eb_vmas *eb,
484 struct drm_i915_gem_relocation_entry *relocs)
485 {
486 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
487 int i, ret;
488
489 for (i = 0; i < entry->relocation_count; i++) {
490 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i]);
491 if (ret)
492 return ret;
493 }
494
495 return 0;
496 }
497
498 static int
i915_gem_execbuffer_relocate(struct eb_vmas * eb)499 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
500 {
501 struct i915_vma *vma;
502 int ret = 0;
503
504 /* This is the fast path and we cannot handle a pagefault whilst
505 * holding the struct mutex lest the user pass in the relocations
506 * contained within a mmaped bo. For in such a case we, the page
507 * fault handler would call i915_gem_fault() and we would try to
508 * acquire the struct mutex again. Obviously this is bad and so
509 * lockdep complains vehemently.
510 */
511 pagefault_disable();
512 list_for_each_entry(vma, &eb->vmas, exec_list) {
513 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
514 if (ret)
515 break;
516 }
517 pagefault_enable();
518
519 return ret;
520 }
521
522 static int
i915_gem_execbuffer_reserve_vma(struct i915_vma * vma,struct intel_engine_cs * ring,bool * need_reloc)523 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
524 struct intel_engine_cs *ring,
525 bool *need_reloc)
526 {
527 struct drm_i915_gem_object *obj = vma->obj;
528 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
529 uint64_t flags;
530 int ret;
531
532 flags = 0;
533 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
534 flags |= PIN_MAPPABLE;
535 if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
536 flags |= PIN_GLOBAL;
537 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
538 flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
539
540 ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
541 if (ret)
542 return ret;
543
544 entry->flags |= __EXEC_OBJECT_HAS_PIN;
545
546 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
547 ret = i915_gem_object_get_fence(obj);
548 if (ret)
549 return ret;
550
551 if (i915_gem_object_pin_fence(obj))
552 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
553 }
554
555 if (entry->offset != vma->node.start) {
556 entry->offset = vma->node.start;
557 *need_reloc = true;
558 }
559
560 if (entry->flags & EXEC_OBJECT_WRITE) {
561 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
562 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
563 }
564
565 return 0;
566 }
567
568 static bool
need_reloc_mappable(struct i915_vma * vma)569 need_reloc_mappable(struct i915_vma *vma)
570 {
571 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
572
573 if (entry->relocation_count == 0)
574 return false;
575
576 if (!i915_is_ggtt(vma->vm))
577 return false;
578
579 /* See also use_cpu_reloc() */
580 if (HAS_LLC(vma->obj->base.dev))
581 return false;
582
583 if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
584 return false;
585
586 return true;
587 }
588
589 static bool
eb_vma_misplaced(struct i915_vma * vma)590 eb_vma_misplaced(struct i915_vma *vma)
591 {
592 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
593 struct drm_i915_gem_object *obj = vma->obj;
594
595 WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
596 !i915_is_ggtt(vma->vm));
597
598 if (entry->alignment &&
599 vma->node.start & (entry->alignment - 1))
600 return true;
601
602 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
603 return true;
604
605 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
606 vma->node.start < BATCH_OFFSET_BIAS)
607 return true;
608
609 return false;
610 }
611
612 static int
i915_gem_execbuffer_reserve(struct intel_engine_cs * ring,struct list_head * vmas,bool * need_relocs)613 i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
614 struct list_head *vmas,
615 bool *need_relocs)
616 {
617 struct drm_i915_gem_object *obj;
618 struct i915_vma *vma;
619 struct i915_address_space *vm;
620 struct list_head ordered_vmas;
621 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
622 int retry;
623
624 i915_gem_retire_requests_ring(ring);
625
626 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
627
628 INIT_LIST_HEAD(&ordered_vmas);
629 while (!list_empty(vmas)) {
630 struct drm_i915_gem_exec_object2 *entry;
631 bool need_fence, need_mappable;
632
633 vma = list_first_entry(vmas, struct i915_vma, exec_list);
634 obj = vma->obj;
635 entry = vma->exec_entry;
636
637 if (!has_fenced_gpu_access)
638 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
639 need_fence =
640 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
641 obj->tiling_mode != I915_TILING_NONE;
642 need_mappable = need_fence || need_reloc_mappable(vma);
643
644 if (need_mappable) {
645 entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
646 list_move(&vma->exec_list, &ordered_vmas);
647 } else
648 list_move_tail(&vma->exec_list, &ordered_vmas);
649
650 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
651 obj->base.pending_write_domain = 0;
652 }
653 list_splice(&ordered_vmas, vmas);
654
655 /* Attempt to pin all of the buffers into the GTT.
656 * This is done in 3 phases:
657 *
658 * 1a. Unbind all objects that do not match the GTT constraints for
659 * the execbuffer (fenceable, mappable, alignment etc).
660 * 1b. Increment pin count for already bound objects.
661 * 2. Bind new objects.
662 * 3. Decrement pin count.
663 *
664 * This avoid unnecessary unbinding of later objects in order to make
665 * room for the earlier objects *unless* we need to defragment.
666 */
667 retry = 0;
668 do {
669 int ret = 0;
670
671 /* Unbind any ill-fitting objects or pin. */
672 list_for_each_entry(vma, vmas, exec_list) {
673 if (!drm_mm_node_allocated(&vma->node))
674 continue;
675
676 if (eb_vma_misplaced(vma))
677 ret = i915_vma_unbind(vma);
678 else
679 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
680 if (ret)
681 goto err;
682 }
683
684 /* Bind fresh objects */
685 list_for_each_entry(vma, vmas, exec_list) {
686 if (drm_mm_node_allocated(&vma->node))
687 continue;
688
689 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
690 if (ret)
691 goto err;
692 }
693
694 err:
695 if (ret != -ENOSPC || retry++)
696 return ret;
697
698 /* Decrement pin count for bound objects */
699 list_for_each_entry(vma, vmas, exec_list)
700 i915_gem_execbuffer_unreserve_vma(vma);
701
702 ret = i915_gem_evict_vm(vm, true);
703 if (ret)
704 return ret;
705 } while (1);
706 }
707
708 static int
i915_gem_execbuffer_relocate_slow(struct drm_device * dev,struct drm_i915_gem_execbuffer2 * args,struct drm_file * file,struct intel_engine_cs * ring,struct eb_vmas * eb,struct drm_i915_gem_exec_object2 * exec)709 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
710 struct drm_i915_gem_execbuffer2 *args,
711 struct drm_file *file,
712 struct intel_engine_cs *ring,
713 struct eb_vmas *eb,
714 struct drm_i915_gem_exec_object2 *exec)
715 {
716 struct drm_i915_gem_relocation_entry *reloc;
717 struct i915_address_space *vm;
718 struct i915_vma *vma;
719 bool need_relocs;
720 int *reloc_offset;
721 int i, total, ret;
722 unsigned count = args->buffer_count;
723
724 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
725
726 /* We may process another execbuffer during the unlock... */
727 while (!list_empty(&eb->vmas)) {
728 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
729 list_del_init(&vma->exec_list);
730 i915_gem_execbuffer_unreserve_vma(vma);
731 drm_gem_object_unreference(&vma->obj->base);
732 }
733
734 mutex_unlock(&dev->struct_mutex);
735
736 total = 0;
737 for (i = 0; i < count; i++)
738 total += exec[i].relocation_count;
739
740 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
741 reloc = drm_malloc_ab(total, sizeof(*reloc));
742 if (reloc == NULL || reloc_offset == NULL) {
743 drm_free_large(reloc);
744 drm_free_large(reloc_offset);
745 mutex_lock(&dev->struct_mutex);
746 return -ENOMEM;
747 }
748
749 total = 0;
750 for (i = 0; i < count; i++) {
751 struct drm_i915_gem_relocation_entry __user *user_relocs;
752 u64 invalid_offset = (u64)-1;
753 int j;
754
755 user_relocs = to_user_ptr(exec[i].relocs_ptr);
756
757 if (copy_from_user(reloc+total, user_relocs,
758 exec[i].relocation_count * sizeof(*reloc))) {
759 ret = -EFAULT;
760 mutex_lock(&dev->struct_mutex);
761 goto err;
762 }
763
764 /* As we do not update the known relocation offsets after
765 * relocating (due to the complexities in lock handling),
766 * we need to mark them as invalid now so that we force the
767 * relocation processing next time. Just in case the target
768 * object is evicted and then rebound into its old
769 * presumed_offset before the next execbuffer - if that
770 * happened we would make the mistake of assuming that the
771 * relocations were valid.
772 */
773 for (j = 0; j < exec[i].relocation_count; j++) {
774 if (__copy_to_user(&user_relocs[j].presumed_offset,
775 &invalid_offset,
776 sizeof(invalid_offset))) {
777 ret = -EFAULT;
778 mutex_lock(&dev->struct_mutex);
779 goto err;
780 }
781 }
782
783 reloc_offset[i] = total;
784 total += exec[i].relocation_count;
785 }
786
787 ret = i915_mutex_lock_interruptible(dev);
788 if (ret) {
789 mutex_lock(&dev->struct_mutex);
790 goto err;
791 }
792
793 /* reacquire the objects */
794 eb_reset(eb);
795 ret = eb_lookup_vmas(eb, exec, args, vm, file);
796 if (ret)
797 goto err;
798
799 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
800 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
801 if (ret)
802 goto err;
803
804 list_for_each_entry(vma, &eb->vmas, exec_list) {
805 int offset = vma->exec_entry - exec;
806 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
807 reloc + reloc_offset[offset]);
808 if (ret)
809 goto err;
810 }
811
812 /* Leave the user relocations as are, this is the painfully slow path,
813 * and we want to avoid the complication of dropping the lock whilst
814 * having buffers reserved in the aperture and so causing spurious
815 * ENOSPC for random operations.
816 */
817
818 err:
819 drm_free_large(reloc);
820 drm_free_large(reloc_offset);
821 return ret;
822 }
823
824 static int
i915_gem_execbuffer_move_to_gpu(struct intel_engine_cs * ring,struct list_head * vmas)825 i915_gem_execbuffer_move_to_gpu(struct intel_engine_cs *ring,
826 struct list_head *vmas)
827 {
828 struct i915_vma *vma;
829 uint32_t flush_domains = 0;
830 bool flush_chipset = false;
831 int ret;
832
833 list_for_each_entry(vma, vmas, exec_list) {
834 struct drm_i915_gem_object *obj = vma->obj;
835 ret = i915_gem_object_sync(obj, ring);
836 if (ret)
837 return ret;
838
839 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
840 flush_chipset |= i915_gem_clflush_object(obj, false);
841
842 flush_domains |= obj->base.write_domain;
843 }
844
845 if (flush_chipset)
846 i915_gem_chipset_flush(ring->dev);
847
848 if (flush_domains & I915_GEM_DOMAIN_GTT)
849 wmb();
850
851 /* Unconditionally invalidate gpu caches and ensure that we do flush
852 * any residual writes from the previous batch.
853 */
854 return intel_ring_invalidate_all_caches(ring);
855 }
856
857 static bool
i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 * exec)858 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
859 {
860 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
861 return false;
862
863 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
864 }
865
866 static int
validate_exec_list(struct drm_device * dev,struct drm_i915_gem_exec_object2 * exec,int count)867 validate_exec_list(struct drm_device *dev,
868 struct drm_i915_gem_exec_object2 *exec,
869 int count)
870 {
871 unsigned relocs_total = 0;
872 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
873 unsigned invalid_flags;
874 int i;
875
876 invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
877 if (USES_FULL_PPGTT(dev))
878 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
879
880 for (i = 0; i < count; i++) {
881 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
882 int length; /* limited by fault_in_pages_readable() */
883
884 if (exec[i].flags & invalid_flags)
885 return -EINVAL;
886
887 /* First check for malicious input causing overflow in
888 * the worst case where we need to allocate the entire
889 * relocation tree as a single array.
890 */
891 if (exec[i].relocation_count > relocs_max - relocs_total)
892 return -EINVAL;
893 relocs_total += exec[i].relocation_count;
894
895 length = exec[i].relocation_count *
896 sizeof(struct drm_i915_gem_relocation_entry);
897 /*
898 * We must check that the entire relocation array is safe
899 * to read, but since we may need to update the presumed
900 * offsets during execution, check for full write access.
901 */
902 if (!access_ok(VERIFY_WRITE, ptr, length))
903 return -EFAULT;
904
905 if (likely(!i915.prefault_disable)) {
906 if (fault_in_multipages_readable(ptr, length))
907 return -EFAULT;
908 }
909 }
910
911 return 0;
912 }
913
914 static struct intel_context *
i915_gem_validate_context(struct drm_device * dev,struct drm_file * file,struct intel_engine_cs * ring,const u32 ctx_id)915 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
916 struct intel_engine_cs *ring, const u32 ctx_id)
917 {
918 struct intel_context *ctx = NULL;
919 struct i915_ctx_hang_stats *hs;
920
921 if (ring->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
922 return ERR_PTR(-EINVAL);
923
924 ctx = i915_gem_context_get(file->driver_priv, ctx_id);
925 if (IS_ERR(ctx))
926 return ctx;
927
928 hs = &ctx->hang_stats;
929 if (hs->banned) {
930 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
931 return ERR_PTR(-EIO);
932 }
933
934 if (i915.enable_execlists && !ctx->engine[ring->id].state) {
935 int ret = intel_lr_context_deferred_create(ctx, ring);
936 if (ret) {
937 DRM_DEBUG("Could not create LRC %u: %d\n", ctx_id, ret);
938 return ERR_PTR(ret);
939 }
940 }
941
942 return ctx;
943 }
944
945 void
i915_gem_execbuffer_move_to_active(struct list_head * vmas,struct intel_engine_cs * ring)946 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
947 struct intel_engine_cs *ring)
948 {
949 u32 seqno = intel_ring_get_seqno(ring);
950 struct i915_vma *vma;
951
952 list_for_each_entry(vma, vmas, exec_list) {
953 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
954 struct drm_i915_gem_object *obj = vma->obj;
955 u32 old_read = obj->base.read_domains;
956 u32 old_write = obj->base.write_domain;
957
958 obj->base.write_domain = obj->base.pending_write_domain;
959 if (obj->base.write_domain == 0)
960 obj->base.pending_read_domains |= obj->base.read_domains;
961 obj->base.read_domains = obj->base.pending_read_domains;
962
963 i915_vma_move_to_active(vma, ring);
964 if (obj->base.write_domain) {
965 obj->dirty = 1;
966 obj->last_write_seqno = seqno;
967
968 intel_fb_obj_invalidate(obj, ring);
969
970 /* update for the implicit flush after a batch */
971 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
972 }
973 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
974 obj->last_fenced_seqno = seqno;
975 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
976 struct drm_i915_private *dev_priv = to_i915(ring->dev);
977 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
978 &dev_priv->mm.fence_list);
979 }
980 }
981
982 trace_i915_gem_object_change_domain(obj, old_read, old_write);
983 }
984 }
985
986 void
i915_gem_execbuffer_retire_commands(struct drm_device * dev,struct drm_file * file,struct intel_engine_cs * ring,struct drm_i915_gem_object * obj)987 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
988 struct drm_file *file,
989 struct intel_engine_cs *ring,
990 struct drm_i915_gem_object *obj)
991 {
992 /* Unconditionally force add_request to emit a full flush. */
993 ring->gpu_caches_dirty = true;
994
995 /* Add a breadcrumb for the completion of the batch buffer */
996 (void)__i915_add_request(ring, file, obj, NULL);
997 }
998
999 static int
i915_reset_gen7_sol_offsets(struct drm_device * dev,struct intel_engine_cs * ring)1000 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1001 struct intel_engine_cs *ring)
1002 {
1003 struct drm_i915_private *dev_priv = dev->dev_private;
1004 int ret, i;
1005
1006 if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS]) {
1007 DRM_DEBUG("sol reset is gen7/rcs only\n");
1008 return -EINVAL;
1009 }
1010
1011 ret = intel_ring_begin(ring, 4 * 3);
1012 if (ret)
1013 return ret;
1014
1015 for (i = 0; i < 4; i++) {
1016 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1017 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1018 intel_ring_emit(ring, 0);
1019 }
1020
1021 intel_ring_advance(ring);
1022
1023 return 0;
1024 }
1025
1026 int
i915_gem_ringbuffer_submission(struct drm_device * dev,struct drm_file * file,struct intel_engine_cs * ring,struct intel_context * ctx,struct drm_i915_gem_execbuffer2 * args,struct list_head * vmas,struct drm_i915_gem_object * batch_obj,u64 exec_start,u32 flags)1027 i915_gem_ringbuffer_submission(struct drm_device *dev, struct drm_file *file,
1028 struct intel_engine_cs *ring,
1029 struct intel_context *ctx,
1030 struct drm_i915_gem_execbuffer2 *args,
1031 struct list_head *vmas,
1032 struct drm_i915_gem_object *batch_obj,
1033 u64 exec_start, u32 flags)
1034 {
1035 struct drm_clip_rect *cliprects = NULL;
1036 struct drm_i915_private *dev_priv = dev->dev_private;
1037 u64 exec_len;
1038 int instp_mode;
1039 u32 instp_mask;
1040 int i, ret = 0;
1041
1042 if (args->num_cliprects != 0) {
1043 if (ring != &dev_priv->ring[RCS]) {
1044 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1045 return -EINVAL;
1046 }
1047
1048 if (INTEL_INFO(dev)->gen >= 5) {
1049 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1050 return -EINVAL;
1051 }
1052
1053 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1054 DRM_DEBUG("execbuf with %u cliprects\n",
1055 args->num_cliprects);
1056 return -EINVAL;
1057 }
1058
1059 cliprects = kcalloc(args->num_cliprects,
1060 sizeof(*cliprects),
1061 GFP_KERNEL);
1062 if (cliprects == NULL) {
1063 ret = -ENOMEM;
1064 goto error;
1065 }
1066
1067 if (copy_from_user(cliprects,
1068 to_user_ptr(args->cliprects_ptr),
1069 sizeof(*cliprects)*args->num_cliprects)) {
1070 ret = -EFAULT;
1071 goto error;
1072 }
1073 } else {
1074 if (args->DR4 == 0xffffffff) {
1075 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1076 args->DR4 = 0;
1077 }
1078
1079 if (args->DR1 || args->DR4 || args->cliprects_ptr) {
1080 DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
1081 return -EINVAL;
1082 }
1083 }
1084
1085 ret = i915_gem_execbuffer_move_to_gpu(ring, vmas);
1086 if (ret)
1087 goto error;
1088
1089 ret = i915_switch_context(ring, ctx);
1090 if (ret)
1091 goto error;
1092
1093 instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1094 instp_mask = I915_EXEC_CONSTANTS_MASK;
1095 switch (instp_mode) {
1096 case I915_EXEC_CONSTANTS_REL_GENERAL:
1097 case I915_EXEC_CONSTANTS_ABSOLUTE:
1098 case I915_EXEC_CONSTANTS_REL_SURFACE:
1099 if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
1100 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1101 ret = -EINVAL;
1102 goto error;
1103 }
1104
1105 if (instp_mode != dev_priv->relative_constants_mode) {
1106 if (INTEL_INFO(dev)->gen < 4) {
1107 DRM_DEBUG("no rel constants on pre-gen4\n");
1108 ret = -EINVAL;
1109 goto error;
1110 }
1111
1112 if (INTEL_INFO(dev)->gen > 5 &&
1113 instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1114 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1115 ret = -EINVAL;
1116 goto error;
1117 }
1118
1119 /* The HW changed the meaning on this bit on gen6 */
1120 if (INTEL_INFO(dev)->gen >= 6)
1121 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1122 }
1123 break;
1124 default:
1125 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1126 ret = -EINVAL;
1127 goto error;
1128 }
1129
1130 if (ring == &dev_priv->ring[RCS] &&
1131 instp_mode != dev_priv->relative_constants_mode) {
1132 ret = intel_ring_begin(ring, 4);
1133 if (ret)
1134 goto error;
1135
1136 intel_ring_emit(ring, MI_NOOP);
1137 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1138 intel_ring_emit(ring, INSTPM);
1139 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1140 intel_ring_advance(ring);
1141
1142 dev_priv->relative_constants_mode = instp_mode;
1143 }
1144
1145 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1146 ret = i915_reset_gen7_sol_offsets(dev, ring);
1147 if (ret)
1148 goto error;
1149 }
1150
1151 exec_len = args->batch_len;
1152 if (cliprects) {
1153 for (i = 0; i < args->num_cliprects; i++) {
1154 ret = i915_emit_box(dev, &cliprects[i],
1155 args->DR1, args->DR4);
1156 if (ret)
1157 goto error;
1158
1159 ret = ring->dispatch_execbuffer(ring,
1160 exec_start, exec_len,
1161 flags);
1162 if (ret)
1163 goto error;
1164 }
1165 } else {
1166 ret = ring->dispatch_execbuffer(ring,
1167 exec_start, exec_len,
1168 flags);
1169 if (ret)
1170 return ret;
1171 }
1172
1173 trace_i915_gem_ring_dispatch(ring, intel_ring_get_seqno(ring), flags);
1174
1175 i915_gem_execbuffer_move_to_active(vmas, ring);
1176 i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
1177
1178 error:
1179 kfree(cliprects);
1180 return ret;
1181 }
1182
1183 /**
1184 * Find one BSD ring to dispatch the corresponding BSD command.
1185 * The Ring ID is returned.
1186 */
gen8_dispatch_bsd_ring(struct drm_device * dev,struct drm_file * file)1187 static int gen8_dispatch_bsd_ring(struct drm_device *dev,
1188 struct drm_file *file)
1189 {
1190 struct drm_i915_private *dev_priv = dev->dev_private;
1191 struct drm_i915_file_private *file_priv = file->driver_priv;
1192
1193 /* Check whether the file_priv is using one ring */
1194 if (file_priv->bsd_ring)
1195 return file_priv->bsd_ring->id;
1196 else {
1197 /* If no, use the ping-pong mechanism to select one ring */
1198 int ring_id;
1199
1200 mutex_lock(&dev->struct_mutex);
1201 if (dev_priv->mm.bsd_ring_dispatch_index == 0) {
1202 ring_id = VCS;
1203 dev_priv->mm.bsd_ring_dispatch_index = 1;
1204 } else {
1205 ring_id = VCS2;
1206 dev_priv->mm.bsd_ring_dispatch_index = 0;
1207 }
1208 file_priv->bsd_ring = &dev_priv->ring[ring_id];
1209 mutex_unlock(&dev->struct_mutex);
1210 return ring_id;
1211 }
1212 }
1213
1214 static struct drm_i915_gem_object *
eb_get_batch(struct eb_vmas * eb)1215 eb_get_batch(struct eb_vmas *eb)
1216 {
1217 struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
1218
1219 /*
1220 * SNA is doing fancy tricks with compressing batch buffers, which leads
1221 * to negative relocation deltas. Usually that works out ok since the
1222 * relocate address is still positive, except when the batch is placed
1223 * very low in the GTT. Ensure this doesn't happen.
1224 *
1225 * Note that actual hangs have only been observed on gen7, but for
1226 * paranoia do it everywhere.
1227 */
1228 vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
1229
1230 return vma->obj;
1231 }
1232
1233 static int
i915_gem_do_execbuffer(struct drm_device * dev,void * data,struct drm_file * file,struct drm_i915_gem_execbuffer2 * args,struct drm_i915_gem_exec_object2 * exec)1234 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1235 struct drm_file *file,
1236 struct drm_i915_gem_execbuffer2 *args,
1237 struct drm_i915_gem_exec_object2 *exec)
1238 {
1239 struct drm_i915_private *dev_priv = dev->dev_private;
1240 struct eb_vmas *eb;
1241 struct drm_i915_gem_object *batch_obj;
1242 struct intel_engine_cs *ring;
1243 struct intel_context *ctx;
1244 struct i915_address_space *vm;
1245 const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1246 u64 exec_start = args->batch_start_offset;
1247 u32 flags;
1248 int ret;
1249 bool need_relocs;
1250
1251 if (!i915_gem_check_execbuffer(args))
1252 return -EINVAL;
1253
1254 ret = validate_exec_list(dev, exec, args->buffer_count);
1255 if (ret)
1256 return ret;
1257
1258 flags = 0;
1259 if (args->flags & I915_EXEC_SECURE) {
1260 if (!file->is_master || !capable(CAP_SYS_ADMIN))
1261 return -EPERM;
1262
1263 flags |= I915_DISPATCH_SECURE;
1264 }
1265 if (args->flags & I915_EXEC_IS_PINNED)
1266 flags |= I915_DISPATCH_PINNED;
1267
1268 if ((args->flags & I915_EXEC_RING_MASK) > LAST_USER_RING) {
1269 DRM_DEBUG("execbuf with unknown ring: %d\n",
1270 (int)(args->flags & I915_EXEC_RING_MASK));
1271 return -EINVAL;
1272 }
1273
1274 if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_DEFAULT)
1275 ring = &dev_priv->ring[RCS];
1276 else if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_BSD) {
1277 if (HAS_BSD2(dev)) {
1278 int ring_id;
1279 ring_id = gen8_dispatch_bsd_ring(dev, file);
1280 ring = &dev_priv->ring[ring_id];
1281 } else
1282 ring = &dev_priv->ring[VCS];
1283 } else
1284 ring = &dev_priv->ring[(args->flags & I915_EXEC_RING_MASK) - 1];
1285
1286 if (!intel_ring_initialized(ring)) {
1287 DRM_DEBUG("execbuf with invalid ring: %d\n",
1288 (int)(args->flags & I915_EXEC_RING_MASK));
1289 return -EINVAL;
1290 }
1291
1292 if (args->buffer_count < 1) {
1293 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1294 return -EINVAL;
1295 }
1296
1297 intel_runtime_pm_get(dev_priv);
1298
1299 ret = i915_mutex_lock_interruptible(dev);
1300 if (ret)
1301 goto pre_mutex_err;
1302
1303 if (dev_priv->ums.mm_suspended) {
1304 mutex_unlock(&dev->struct_mutex);
1305 ret = -EBUSY;
1306 goto pre_mutex_err;
1307 }
1308
1309 ctx = i915_gem_validate_context(dev, file, ring, ctx_id);
1310 if (IS_ERR(ctx)) {
1311 mutex_unlock(&dev->struct_mutex);
1312 ret = PTR_ERR(ctx);
1313 goto pre_mutex_err;
1314 }
1315
1316 i915_gem_context_reference(ctx);
1317
1318 if (ctx->ppgtt)
1319 vm = &ctx->ppgtt->base;
1320 else
1321 vm = &dev_priv->gtt.base;
1322
1323 eb = eb_create(args);
1324 if (eb == NULL) {
1325 i915_gem_context_unreference(ctx);
1326 mutex_unlock(&dev->struct_mutex);
1327 ret = -ENOMEM;
1328 goto pre_mutex_err;
1329 }
1330
1331 /* Look up object handles */
1332 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1333 if (ret)
1334 goto err;
1335
1336 /* take note of the batch buffer before we might reorder the lists */
1337 batch_obj = eb_get_batch(eb);
1338
1339 /* Move the objects en-masse into the GTT, evicting if necessary. */
1340 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1341 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
1342 if (ret)
1343 goto err;
1344
1345 /* The objects are in their final locations, apply the relocations. */
1346 if (need_relocs)
1347 ret = i915_gem_execbuffer_relocate(eb);
1348 if (ret) {
1349 if (ret == -EFAULT) {
1350 ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1351 eb, exec);
1352 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1353 }
1354 if (ret)
1355 goto err;
1356 }
1357
1358 /* Set the pending read domains for the batch buffer to COMMAND */
1359 if (batch_obj->base.pending_write_domain) {
1360 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1361 ret = -EINVAL;
1362 goto err;
1363 }
1364 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1365
1366 if (i915_needs_cmd_parser(ring)) {
1367 ret = i915_parse_cmds(ring,
1368 batch_obj,
1369 args->batch_start_offset,
1370 file->is_master);
1371 if (ret)
1372 goto err;
1373
1374 /*
1375 * XXX: Actually do this when enabling batch copy...
1376 *
1377 * Set the DISPATCH_SECURE bit to remove the NON_SECURE bit
1378 * from MI_BATCH_BUFFER_START commands issued in the
1379 * dispatch_execbuffer implementations. We specifically don't
1380 * want that set when the command parser is enabled.
1381 */
1382 }
1383
1384 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1385 * batch" bit. Hence we need to pin secure batches into the global gtt.
1386 * hsw should have this fixed, but bdw mucks it up again. */
1387 if (flags & I915_DISPATCH_SECURE) {
1388 /*
1389 * So on first glance it looks freaky that we pin the batch here
1390 * outside of the reservation loop. But:
1391 * - The batch is already pinned into the relevant ppgtt, so we
1392 * already have the backing storage fully allocated.
1393 * - No other BO uses the global gtt (well contexts, but meh),
1394 * so we don't really have issues with mutliple objects not
1395 * fitting due to fragmentation.
1396 * So this is actually safe.
1397 */
1398 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1399 if (ret)
1400 goto err;
1401
1402 exec_start += i915_gem_obj_ggtt_offset(batch_obj);
1403 } else
1404 exec_start += i915_gem_obj_offset(batch_obj, vm);
1405
1406 ret = dev_priv->gt.do_execbuf(dev, file, ring, ctx, args,
1407 &eb->vmas, batch_obj, exec_start, flags);
1408
1409 /*
1410 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1411 * batch vma for correctness. For less ugly and less fragility this
1412 * needs to be adjusted to also track the ggtt batch vma properly as
1413 * active.
1414 */
1415 if (flags & I915_DISPATCH_SECURE)
1416 i915_gem_object_ggtt_unpin(batch_obj);
1417 err:
1418 /* the request owns the ref now */
1419 i915_gem_context_unreference(ctx);
1420 eb_destroy(eb);
1421
1422 mutex_unlock(&dev->struct_mutex);
1423
1424 pre_mutex_err:
1425 /* intel_gpu_busy should also get a ref, so it will free when the device
1426 * is really idle. */
1427 intel_runtime_pm_put(dev_priv);
1428 return ret;
1429 }
1430
1431 /*
1432 * Legacy execbuffer just creates an exec2 list from the original exec object
1433 * list array and passes it to the real function.
1434 */
1435 int
i915_gem_execbuffer(struct drm_device * dev,void * data,struct drm_file * file)1436 i915_gem_execbuffer(struct drm_device *dev, void *data,
1437 struct drm_file *file)
1438 {
1439 struct drm_i915_gem_execbuffer *args = data;
1440 struct drm_i915_gem_execbuffer2 exec2;
1441 struct drm_i915_gem_exec_object *exec_list = NULL;
1442 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1443 int ret, i;
1444
1445 if (args->buffer_count < 1) {
1446 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1447 return -EINVAL;
1448 }
1449
1450 /* Copy in the exec list from userland */
1451 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1452 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1453 if (exec_list == NULL || exec2_list == NULL) {
1454 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1455 args->buffer_count);
1456 drm_free_large(exec_list);
1457 drm_free_large(exec2_list);
1458 return -ENOMEM;
1459 }
1460 ret = copy_from_user(exec_list,
1461 to_user_ptr(args->buffers_ptr),
1462 sizeof(*exec_list) * args->buffer_count);
1463 if (ret != 0) {
1464 DRM_DEBUG("copy %d exec entries failed %d\n",
1465 args->buffer_count, ret);
1466 drm_free_large(exec_list);
1467 drm_free_large(exec2_list);
1468 return -EFAULT;
1469 }
1470
1471 for (i = 0; i < args->buffer_count; i++) {
1472 exec2_list[i].handle = exec_list[i].handle;
1473 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1474 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1475 exec2_list[i].alignment = exec_list[i].alignment;
1476 exec2_list[i].offset = exec_list[i].offset;
1477 if (INTEL_INFO(dev)->gen < 4)
1478 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1479 else
1480 exec2_list[i].flags = 0;
1481 }
1482
1483 exec2.buffers_ptr = args->buffers_ptr;
1484 exec2.buffer_count = args->buffer_count;
1485 exec2.batch_start_offset = args->batch_start_offset;
1486 exec2.batch_len = args->batch_len;
1487 exec2.DR1 = args->DR1;
1488 exec2.DR4 = args->DR4;
1489 exec2.num_cliprects = args->num_cliprects;
1490 exec2.cliprects_ptr = args->cliprects_ptr;
1491 exec2.flags = I915_EXEC_RENDER;
1492 i915_execbuffer2_set_context_id(exec2, 0);
1493
1494 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1495 if (!ret) {
1496 struct drm_i915_gem_exec_object __user *user_exec_list =
1497 to_user_ptr(args->buffers_ptr);
1498
1499 /* Copy the new buffer offsets back to the user's exec list. */
1500 for (i = 0; i < args->buffer_count; i++) {
1501 ret = __copy_to_user(&user_exec_list[i].offset,
1502 &exec2_list[i].offset,
1503 sizeof(user_exec_list[i].offset));
1504 if (ret) {
1505 ret = -EFAULT;
1506 DRM_DEBUG("failed to copy %d exec entries "
1507 "back to user (%d)\n",
1508 args->buffer_count, ret);
1509 break;
1510 }
1511 }
1512 }
1513
1514 drm_free_large(exec_list);
1515 drm_free_large(exec2_list);
1516 return ret;
1517 }
1518
1519 int
i915_gem_execbuffer2(struct drm_device * dev,void * data,struct drm_file * file)1520 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1521 struct drm_file *file)
1522 {
1523 struct drm_i915_gem_execbuffer2 *args = data;
1524 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1525 int ret;
1526
1527 if (args->buffer_count < 1 ||
1528 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1529 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1530 return -EINVAL;
1531 }
1532
1533 if (args->rsvd2 != 0) {
1534 DRM_DEBUG("dirty rvsd2 field\n");
1535 return -EINVAL;
1536 }
1537
1538 exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1539 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
1540 if (exec2_list == NULL)
1541 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1542 args->buffer_count);
1543 if (exec2_list == NULL) {
1544 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1545 args->buffer_count);
1546 return -ENOMEM;
1547 }
1548 ret = copy_from_user(exec2_list,
1549 to_user_ptr(args->buffers_ptr),
1550 sizeof(*exec2_list) * args->buffer_count);
1551 if (ret != 0) {
1552 DRM_DEBUG("copy %d exec entries failed %d\n",
1553 args->buffer_count, ret);
1554 drm_free_large(exec2_list);
1555 return -EFAULT;
1556 }
1557
1558 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1559 if (!ret) {
1560 /* Copy the new buffer offsets back to the user's exec list. */
1561 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1562 to_user_ptr(args->buffers_ptr);
1563 int i;
1564
1565 for (i = 0; i < args->buffer_count; i++) {
1566 ret = __copy_to_user(&user_exec_list[i].offset,
1567 &exec2_list[i].offset,
1568 sizeof(user_exec_list[i].offset));
1569 if (ret) {
1570 ret = -EFAULT;
1571 DRM_DEBUG("failed to copy %d exec entries "
1572 "back to user\n",
1573 args->buffer_count);
1574 break;
1575 }
1576 }
1577 }
1578
1579 drm_free_large(exec2_list);
1580 return ret;
1581 }
1582