1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
5 */
6
7 #include <linux/file.h>
8 #include <linux/sync_file.h>
9 #include <linux/uaccess.h>
10
11 #include <drm/drm_drv.h>
12 #include <drm/drm_file.h>
13 #include <drm/drm_syncobj.h>
14
15 #include "msm_drv.h"
16 #include "msm_gpu.h"
17 #include "msm_gem.h"
18 #include "msm_gpu_trace.h"
19
20 /*
21 * Cmdstream submission:
22 */
23
24 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */
25 #define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */
26 #define BO_LOCKED 0x4000 /* obj lock is held */
27 #define BO_ACTIVE 0x2000 /* active refcnt is held */
28 #define BO_PINNED 0x1000 /* obj is pinned and on active list */
29
submit_create(struct drm_device * dev,struct msm_gpu * gpu,struct msm_gpu_submitqueue * queue,uint32_t nr_bos,uint32_t nr_cmds)30 static struct msm_gem_submit *submit_create(struct drm_device *dev,
31 struct msm_gpu *gpu,
32 struct msm_gpu_submitqueue *queue, uint32_t nr_bos,
33 uint32_t nr_cmds)
34 {
35 struct msm_gem_submit *submit;
36 uint64_t sz;
37 int ret;
38
39 sz = struct_size(submit, bos, nr_bos) +
40 ((u64)nr_cmds * sizeof(submit->cmd[0]));
41
42 if (sz > SIZE_MAX)
43 return ERR_PTR(-ENOMEM);
44
45 submit = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
46 if (!submit)
47 return ERR_PTR(-ENOMEM);
48
49 ret = drm_sched_job_init(&submit->base, queue->entity, queue);
50 if (ret) {
51 kfree(submit);
52 return ERR_PTR(ret);
53 }
54
55 xa_init_flags(&submit->deps, XA_FLAGS_ALLOC);
56
57 kref_init(&submit->ref);
58 submit->dev = dev;
59 submit->aspace = queue->ctx->aspace;
60 submit->gpu = gpu;
61 submit->cmd = (void *)&submit->bos[nr_bos];
62 submit->queue = queue;
63 submit->ring = gpu->rb[queue->ring_nr];
64 submit->fault_dumped = false;
65
66 INIT_LIST_HEAD(&submit->node);
67
68 return submit;
69 }
70
__msm_gem_submit_destroy(struct kref * kref)71 void __msm_gem_submit_destroy(struct kref *kref)
72 {
73 struct msm_gem_submit *submit =
74 container_of(kref, struct msm_gem_submit, ref);
75 unsigned long index;
76 struct dma_fence *fence;
77 unsigned i;
78
79 if (submit->fence_id) {
80 mutex_lock(&submit->queue->lock);
81 idr_remove(&submit->queue->fence_idr, submit->fence_id);
82 mutex_unlock(&submit->queue->lock);
83 }
84
85 xa_for_each (&submit->deps, index, fence) {
86 dma_fence_put(fence);
87 }
88
89 xa_destroy(&submit->deps);
90
91 dma_fence_put(submit->user_fence);
92 dma_fence_put(submit->hw_fence);
93
94 put_pid(submit->pid);
95 msm_submitqueue_put(submit->queue);
96
97 for (i = 0; i < submit->nr_cmds; i++)
98 kfree(submit->cmd[i].relocs);
99
100 kfree(submit);
101 }
102
submit_lookup_objects(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)103 static int submit_lookup_objects(struct msm_gem_submit *submit,
104 struct drm_msm_gem_submit *args, struct drm_file *file)
105 {
106 unsigned i;
107 int ret = 0;
108
109 for (i = 0; i < args->nr_bos; i++) {
110 struct drm_msm_gem_submit_bo submit_bo;
111 void __user *userptr =
112 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
113
114 /* make sure we don't have garbage flags, in case we hit
115 * error path before flags is initialized:
116 */
117 submit->bos[i].flags = 0;
118
119 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
120 ret = -EFAULT;
121 i = 0;
122 goto out;
123 }
124
125 /* at least one of READ and/or WRITE flags should be set: */
126 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
127
128 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
129 !(submit_bo.flags & MANDATORY_FLAGS)) {
130 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
131 ret = -EINVAL;
132 i = 0;
133 goto out;
134 }
135
136 submit->bos[i].handle = submit_bo.handle;
137 submit->bos[i].flags = submit_bo.flags;
138 /* in validate_objects() we figure out if this is true: */
139 submit->bos[i].iova = submit_bo.presumed;
140 }
141
142 spin_lock(&file->table_lock);
143
144 for (i = 0; i < args->nr_bos; i++) {
145 struct drm_gem_object *obj;
146
147 /* normally use drm_gem_object_lookup(), but for bulk lookup
148 * all under single table_lock just hit object_idr directly:
149 */
150 obj = idr_find(&file->object_idr, submit->bos[i].handle);
151 if (!obj) {
152 DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
153 ret = -EINVAL;
154 goto out_unlock;
155 }
156
157 drm_gem_object_get(obj);
158
159 submit->bos[i].obj = to_msm_bo(obj);
160 }
161
162 out_unlock:
163 spin_unlock(&file->table_lock);
164
165 out:
166 submit->nr_bos = i;
167
168 return ret;
169 }
170
submit_lookup_cmds(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)171 static int submit_lookup_cmds(struct msm_gem_submit *submit,
172 struct drm_msm_gem_submit *args, struct drm_file *file)
173 {
174 unsigned i;
175 size_t sz;
176 int ret = 0;
177
178 for (i = 0; i < args->nr_cmds; i++) {
179 struct drm_msm_gem_submit_cmd submit_cmd;
180 void __user *userptr =
181 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
182
183 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
184 if (ret) {
185 ret = -EFAULT;
186 goto out;
187 }
188
189 /* validate input from userspace: */
190 switch (submit_cmd.type) {
191 case MSM_SUBMIT_CMD_BUF:
192 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
193 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
194 break;
195 default:
196 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
197 return -EINVAL;
198 }
199
200 if (submit_cmd.size % 4) {
201 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
202 submit_cmd.size);
203 ret = -EINVAL;
204 goto out;
205 }
206
207 submit->cmd[i].type = submit_cmd.type;
208 submit->cmd[i].size = submit_cmd.size / 4;
209 submit->cmd[i].offset = submit_cmd.submit_offset / 4;
210 submit->cmd[i].idx = submit_cmd.submit_idx;
211 submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
212
213 userptr = u64_to_user_ptr(submit_cmd.relocs);
214
215 sz = array_size(submit_cmd.nr_relocs,
216 sizeof(struct drm_msm_gem_submit_reloc));
217 /* check for overflow: */
218 if (sz == SIZE_MAX) {
219 ret = -ENOMEM;
220 goto out;
221 }
222 submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL);
223 if (!submit->cmd[i].relocs) {
224 ret = -ENOMEM;
225 goto out;
226 }
227 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
228 if (ret) {
229 ret = -EFAULT;
230 goto out;
231 }
232 }
233
234 out:
235 return ret;
236 }
237
238 /* Unwind bo state, according to cleanup_flags. In the success case, only
239 * the lock is dropped at the end of the submit (and active/pin ref is dropped
240 * later when the submit is retired).
241 */
submit_cleanup_bo(struct msm_gem_submit * submit,int i,unsigned cleanup_flags)242 static void submit_cleanup_bo(struct msm_gem_submit *submit, int i,
243 unsigned cleanup_flags)
244 {
245 struct drm_gem_object *obj = &submit->bos[i].obj->base;
246 unsigned flags = submit->bos[i].flags & cleanup_flags;
247
248 if (flags & BO_PINNED)
249 msm_gem_unpin_iova_locked(obj, submit->aspace);
250
251 if (flags & BO_ACTIVE)
252 msm_gem_active_put(obj);
253
254 if (flags & BO_LOCKED)
255 dma_resv_unlock(obj->resv);
256
257 submit->bos[i].flags &= ~cleanup_flags;
258 }
259
submit_unlock_unpin_bo(struct msm_gem_submit * submit,int i)260 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
261 {
262 submit_cleanup_bo(submit, i, BO_PINNED | BO_ACTIVE | BO_LOCKED);
263
264 if (!(submit->bos[i].flags & BO_VALID))
265 submit->bos[i].iova = 0;
266 }
267
268 /* This is where we make sure all the bo's are reserved and pin'd: */
submit_lock_objects(struct msm_gem_submit * submit)269 static int submit_lock_objects(struct msm_gem_submit *submit)
270 {
271 int contended, slow_locked = -1, i, ret = 0;
272
273 retry:
274 for (i = 0; i < submit->nr_bos; i++) {
275 struct msm_gem_object *msm_obj = submit->bos[i].obj;
276
277 if (slow_locked == i)
278 slow_locked = -1;
279
280 contended = i;
281
282 if (!(submit->bos[i].flags & BO_LOCKED)) {
283 ret = dma_resv_lock_interruptible(msm_obj->base.resv,
284 &submit->ticket);
285 if (ret)
286 goto fail;
287 submit->bos[i].flags |= BO_LOCKED;
288 }
289 }
290
291 ww_acquire_done(&submit->ticket);
292
293 return 0;
294
295 fail:
296 if (ret == -EALREADY) {
297 DRM_ERROR("handle %u at index %u already on submit list\n",
298 submit->bos[i].handle, i);
299 ret = -EINVAL;
300 }
301
302 for (; i >= 0; i--)
303 submit_unlock_unpin_bo(submit, i);
304
305 if (slow_locked > 0)
306 submit_unlock_unpin_bo(submit, slow_locked);
307
308 if (ret == -EDEADLK) {
309 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
310 /* we lost out in a seqno race, lock and retry.. */
311 ret = dma_resv_lock_slow_interruptible(msm_obj->base.resv,
312 &submit->ticket);
313 if (!ret) {
314 submit->bos[contended].flags |= BO_LOCKED;
315 slow_locked = contended;
316 goto retry;
317 }
318
319 /* Not expecting -EALREADY here, if the bo was already
320 * locked, we should have gotten -EALREADY already from
321 * the dma_resv_lock_interruptable() call.
322 */
323 WARN_ON_ONCE(ret == -EALREADY);
324 }
325
326 return ret;
327 }
328
submit_fence_sync(struct msm_gem_submit * submit,bool no_implicit)329 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
330 {
331 int i, ret = 0;
332
333 for (i = 0; i < submit->nr_bos; i++) {
334 struct drm_gem_object *obj = &submit->bos[i].obj->base;
335 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
336
337 if (!write) {
338 /* NOTE: _reserve_shared() must happen before
339 * _add_shared_fence(), which makes this a slightly
340 * strange place to call it. OTOH this is a
341 * convenient can-fail point to hook it in.
342 */
343 ret = dma_resv_reserve_shared(obj->resv, 1);
344 if (ret)
345 return ret;
346 }
347
348 if (no_implicit)
349 continue;
350
351 ret = drm_gem_fence_array_add_implicit(&submit->deps, obj,
352 write);
353 if (ret)
354 break;
355 }
356
357 return ret;
358 }
359
submit_pin_objects(struct msm_gem_submit * submit)360 static int submit_pin_objects(struct msm_gem_submit *submit)
361 {
362 int i, ret = 0;
363
364 submit->valid = true;
365
366 /*
367 * Increment active_count first, so if under memory pressure, we
368 * don't inadvertently evict a bo needed by the submit in order
369 * to pin an earlier bo in the same submit.
370 */
371 for (i = 0; i < submit->nr_bos; i++) {
372 struct drm_gem_object *obj = &submit->bos[i].obj->base;
373
374 msm_gem_active_get(obj, submit->gpu);
375 submit->bos[i].flags |= BO_ACTIVE;
376 }
377
378 for (i = 0; i < submit->nr_bos; i++) {
379 struct drm_gem_object *obj = &submit->bos[i].obj->base;
380 uint64_t iova;
381
382 /* if locking succeeded, pin bo: */
383 ret = msm_gem_get_and_pin_iova_locked(obj,
384 submit->aspace, &iova);
385
386 if (ret)
387 break;
388
389 submit->bos[i].flags |= BO_PINNED;
390
391 if (iova == submit->bos[i].iova) {
392 submit->bos[i].flags |= BO_VALID;
393 } else {
394 submit->bos[i].iova = iova;
395 /* iova changed, so address in cmdstream is not valid: */
396 submit->bos[i].flags &= ~BO_VALID;
397 submit->valid = false;
398 }
399 }
400
401 return ret;
402 }
403
submit_attach_object_fences(struct msm_gem_submit * submit)404 static void submit_attach_object_fences(struct msm_gem_submit *submit)
405 {
406 int i;
407
408 for (i = 0; i < submit->nr_bos; i++) {
409 struct drm_gem_object *obj = &submit->bos[i].obj->base;
410
411 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
412 dma_resv_add_excl_fence(obj->resv, submit->user_fence);
413 else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
414 dma_resv_add_shared_fence(obj->resv, submit->user_fence);
415 }
416 }
417
submit_bo(struct msm_gem_submit * submit,uint32_t idx,struct msm_gem_object ** obj,uint64_t * iova,bool * valid)418 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
419 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
420 {
421 if (idx >= submit->nr_bos) {
422 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
423 idx, submit->nr_bos);
424 return -EINVAL;
425 }
426
427 if (obj)
428 *obj = submit->bos[idx].obj;
429 if (iova)
430 *iova = submit->bos[idx].iova;
431 if (valid)
432 *valid = !!(submit->bos[idx].flags & BO_VALID);
433
434 return 0;
435 }
436
437 /* process the reloc's and patch up the cmdstream as needed: */
submit_reloc(struct msm_gem_submit * submit,struct msm_gem_object * obj,uint32_t offset,uint32_t nr_relocs,struct drm_msm_gem_submit_reloc * relocs)438 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
439 uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs)
440 {
441 uint32_t i, last_offset = 0;
442 uint32_t *ptr;
443 int ret = 0;
444
445 if (!nr_relocs)
446 return 0;
447
448 if (offset % 4) {
449 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
450 return -EINVAL;
451 }
452
453 /* For now, just map the entire thing. Eventually we probably
454 * to do it page-by-page, w/ kmap() if not vmap()d..
455 */
456 ptr = msm_gem_get_vaddr_locked(&obj->base);
457
458 if (IS_ERR(ptr)) {
459 ret = PTR_ERR(ptr);
460 DBG("failed to map: %d", ret);
461 return ret;
462 }
463
464 for (i = 0; i < nr_relocs; i++) {
465 struct drm_msm_gem_submit_reloc submit_reloc = relocs[i];
466 uint32_t off;
467 uint64_t iova;
468 bool valid;
469
470 if (submit_reloc.submit_offset % 4) {
471 DRM_ERROR("non-aligned reloc offset: %u\n",
472 submit_reloc.submit_offset);
473 ret = -EINVAL;
474 goto out;
475 }
476
477 /* offset in dwords: */
478 off = submit_reloc.submit_offset / 4;
479
480 if ((off >= (obj->base.size / 4)) ||
481 (off < last_offset)) {
482 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
483 ret = -EINVAL;
484 goto out;
485 }
486
487 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
488 if (ret)
489 goto out;
490
491 if (valid)
492 continue;
493
494 iova += submit_reloc.reloc_offset;
495
496 if (submit_reloc.shift < 0)
497 iova >>= -submit_reloc.shift;
498 else
499 iova <<= submit_reloc.shift;
500
501 ptr[off] = iova | submit_reloc.or;
502
503 last_offset = off;
504 }
505
506 out:
507 msm_gem_put_vaddr_locked(&obj->base);
508
509 return ret;
510 }
511
512 /* Cleanup submit at end of ioctl. In the error case, this also drops
513 * references, unpins, and drops active refcnt. In the non-error case,
514 * this is done when the submit is retired.
515 */
submit_cleanup(struct msm_gem_submit * submit,bool error)516 static void submit_cleanup(struct msm_gem_submit *submit, bool error)
517 {
518 unsigned cleanup_flags = BO_LOCKED;
519 unsigned i;
520
521 if (error)
522 cleanup_flags |= BO_PINNED | BO_ACTIVE;
523
524 for (i = 0; i < submit->nr_bos; i++) {
525 struct msm_gem_object *msm_obj = submit->bos[i].obj;
526 submit_cleanup_bo(submit, i, cleanup_flags);
527 if (error)
528 drm_gem_object_put(&msm_obj->base);
529 }
530 }
531
msm_submit_retire(struct msm_gem_submit * submit)532 void msm_submit_retire(struct msm_gem_submit *submit)
533 {
534 int i;
535
536 for (i = 0; i < submit->nr_bos; i++) {
537 struct drm_gem_object *obj = &submit->bos[i].obj->base;
538
539 msm_gem_lock(obj);
540 submit_cleanup_bo(submit, i, BO_PINNED | BO_ACTIVE);
541 msm_gem_unlock(obj);
542 drm_gem_object_put(obj);
543 }
544 }
545
546 struct msm_submit_post_dep {
547 struct drm_syncobj *syncobj;
548 uint64_t point;
549 struct dma_fence_chain *chain;
550 };
551
msm_parse_deps(struct msm_gem_submit * submit,struct drm_file * file,uint64_t in_syncobjs_addr,uint32_t nr_in_syncobjs,size_t syncobj_stride,struct msm_ringbuffer * ring)552 static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
553 struct drm_file *file,
554 uint64_t in_syncobjs_addr,
555 uint32_t nr_in_syncobjs,
556 size_t syncobj_stride,
557 struct msm_ringbuffer *ring)
558 {
559 struct drm_syncobj **syncobjs = NULL;
560 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
561 int ret = 0;
562 uint32_t i, j;
563
564 syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
565 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
566 if (!syncobjs)
567 return ERR_PTR(-ENOMEM);
568
569 for (i = 0; i < nr_in_syncobjs; ++i) {
570 uint64_t address = in_syncobjs_addr + i * syncobj_stride;
571 struct dma_fence *fence;
572
573 if (copy_from_user(&syncobj_desc,
574 u64_to_user_ptr(address),
575 min(syncobj_stride, sizeof(syncobj_desc)))) {
576 ret = -EFAULT;
577 break;
578 }
579
580 if (syncobj_desc.point &&
581 !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
582 ret = -EOPNOTSUPP;
583 break;
584 }
585
586 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
587 ret = -EINVAL;
588 break;
589 }
590
591 ret = drm_syncobj_find_fence(file, syncobj_desc.handle,
592 syncobj_desc.point, 0, &fence);
593 if (ret)
594 break;
595
596 ret = drm_gem_fence_array_add(&submit->deps, fence);
597 if (ret)
598 break;
599
600 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
601 syncobjs[i] =
602 drm_syncobj_find(file, syncobj_desc.handle);
603 if (!syncobjs[i]) {
604 ret = -EINVAL;
605 break;
606 }
607 }
608 }
609
610 if (ret) {
611 for (j = 0; j <= i; ++j) {
612 if (syncobjs[j])
613 drm_syncobj_put(syncobjs[j]);
614 }
615 kfree(syncobjs);
616 return ERR_PTR(ret);
617 }
618 return syncobjs;
619 }
620
msm_reset_syncobjs(struct drm_syncobj ** syncobjs,uint32_t nr_syncobjs)621 static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
622 uint32_t nr_syncobjs)
623 {
624 uint32_t i;
625
626 for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
627 if (syncobjs[i])
628 drm_syncobj_replace_fence(syncobjs[i], NULL);
629 }
630 }
631
msm_parse_post_deps(struct drm_device * dev,struct drm_file * file,uint64_t syncobjs_addr,uint32_t nr_syncobjs,size_t syncobj_stride)632 static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
633 struct drm_file *file,
634 uint64_t syncobjs_addr,
635 uint32_t nr_syncobjs,
636 size_t syncobj_stride)
637 {
638 struct msm_submit_post_dep *post_deps;
639 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
640 int ret = 0;
641 uint32_t i, j;
642
643 post_deps = kcalloc(nr_syncobjs, sizeof(*post_deps),
644 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
645 if (!post_deps)
646 return ERR_PTR(-ENOMEM);
647
648 for (i = 0; i < nr_syncobjs; ++i) {
649 uint64_t address = syncobjs_addr + i * syncobj_stride;
650
651 if (copy_from_user(&syncobj_desc,
652 u64_to_user_ptr(address),
653 min(syncobj_stride, sizeof(syncobj_desc)))) {
654 ret = -EFAULT;
655 break;
656 }
657
658 post_deps[i].point = syncobj_desc.point;
659
660 if (syncobj_desc.flags) {
661 ret = -EINVAL;
662 break;
663 }
664
665 if (syncobj_desc.point) {
666 if (!drm_core_check_feature(dev,
667 DRIVER_SYNCOBJ_TIMELINE)) {
668 ret = -EOPNOTSUPP;
669 break;
670 }
671
672 post_deps[i].chain = dma_fence_chain_alloc();
673 if (!post_deps[i].chain) {
674 ret = -ENOMEM;
675 break;
676 }
677 }
678
679 post_deps[i].syncobj =
680 drm_syncobj_find(file, syncobj_desc.handle);
681 if (!post_deps[i].syncobj) {
682 ret = -EINVAL;
683 break;
684 }
685 }
686
687 if (ret) {
688 for (j = 0; j <= i; ++j) {
689 dma_fence_chain_free(post_deps[j].chain);
690 if (post_deps[j].syncobj)
691 drm_syncobj_put(post_deps[j].syncobj);
692 }
693
694 kfree(post_deps);
695 return ERR_PTR(ret);
696 }
697
698 return post_deps;
699 }
700
msm_process_post_deps(struct msm_submit_post_dep * post_deps,uint32_t count,struct dma_fence * fence)701 static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
702 uint32_t count, struct dma_fence *fence)
703 {
704 uint32_t i;
705
706 for (i = 0; post_deps && i < count; ++i) {
707 if (post_deps[i].chain) {
708 drm_syncobj_add_point(post_deps[i].syncobj,
709 post_deps[i].chain,
710 fence, post_deps[i].point);
711 post_deps[i].chain = NULL;
712 } else {
713 drm_syncobj_replace_fence(post_deps[i].syncobj,
714 fence);
715 }
716 }
717 }
718
msm_ioctl_gem_submit(struct drm_device * dev,void * data,struct drm_file * file)719 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
720 struct drm_file *file)
721 {
722 static atomic_t ident = ATOMIC_INIT(0);
723 struct msm_drm_private *priv = dev->dev_private;
724 struct drm_msm_gem_submit *args = data;
725 struct msm_file_private *ctx = file->driver_priv;
726 struct msm_gem_submit *submit = NULL;
727 struct msm_gpu *gpu = priv->gpu;
728 struct msm_gpu_submitqueue *queue;
729 struct msm_ringbuffer *ring;
730 struct msm_submit_post_dep *post_deps = NULL;
731 struct drm_syncobj **syncobjs_to_reset = NULL;
732 int out_fence_fd = -1;
733 struct pid *pid = get_pid(task_pid(current));
734 bool has_ww_ticket = false;
735 unsigned i;
736 int ret, submitid;
737
738 if (!gpu)
739 return -ENXIO;
740
741 if (args->pad)
742 return -EINVAL;
743
744 /* for now, we just have 3d pipe.. eventually this would need to
745 * be more clever to dispatch to appropriate gpu module:
746 */
747 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
748 return -EINVAL;
749
750 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
751 return -EINVAL;
752
753 if (args->flags & MSM_SUBMIT_SUDO) {
754 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
755 !capable(CAP_SYS_RAWIO))
756 return -EINVAL;
757 }
758
759 queue = msm_submitqueue_get(ctx, args->queueid);
760 if (!queue)
761 return -ENOENT;
762
763 /* Get a unique identifier for the submission for logging purposes */
764 submitid = atomic_inc_return(&ident) - 1;
765
766 ring = gpu->rb[queue->ring_nr];
767 trace_msm_gpu_submit(pid_nr(pid), ring->id, submitid,
768 args->nr_bos, args->nr_cmds);
769
770 ret = mutex_lock_interruptible(&queue->lock);
771 if (ret)
772 goto out_post_unlock;
773
774 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
775 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
776 if (out_fence_fd < 0) {
777 ret = out_fence_fd;
778 goto out_unlock;
779 }
780 }
781
782 submit = submit_create(dev, gpu, queue, args->nr_bos,
783 args->nr_cmds);
784 if (IS_ERR(submit)) {
785 ret = PTR_ERR(submit);
786 submit = NULL;
787 goto out_unlock;
788 }
789
790 submit->pid = pid;
791 submit->ident = submitid;
792
793 if (args->flags & MSM_SUBMIT_SUDO)
794 submit->in_rb = true;
795
796 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
797 struct dma_fence *in_fence;
798
799 in_fence = sync_file_get_fence(args->fence_fd);
800
801 if (!in_fence) {
802 ret = -EINVAL;
803 goto out_unlock;
804 }
805
806 ret = drm_gem_fence_array_add(&submit->deps, in_fence);
807 if (ret)
808 goto out_unlock;
809 }
810
811 if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
812 syncobjs_to_reset = msm_parse_deps(submit, file,
813 args->in_syncobjs,
814 args->nr_in_syncobjs,
815 args->syncobj_stride, ring);
816 if (IS_ERR(syncobjs_to_reset)) {
817 ret = PTR_ERR(syncobjs_to_reset);
818 goto out_unlock;
819 }
820 }
821
822 if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
823 post_deps = msm_parse_post_deps(dev, file,
824 args->out_syncobjs,
825 args->nr_out_syncobjs,
826 args->syncobj_stride);
827 if (IS_ERR(post_deps)) {
828 ret = PTR_ERR(post_deps);
829 goto out_unlock;
830 }
831 }
832
833 ret = submit_lookup_objects(submit, args, file);
834 if (ret)
835 goto out;
836
837 ret = submit_lookup_cmds(submit, args, file);
838 if (ret)
839 goto out;
840
841 /* copy_*_user while holding a ww ticket upsets lockdep */
842 ww_acquire_init(&submit->ticket, &reservation_ww_class);
843 has_ww_ticket = true;
844 ret = submit_lock_objects(submit);
845 if (ret)
846 goto out;
847
848 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
849 if (ret)
850 goto out;
851
852 ret = submit_pin_objects(submit);
853 if (ret)
854 goto out;
855
856 for (i = 0; i < args->nr_cmds; i++) {
857 struct msm_gem_object *msm_obj;
858 uint64_t iova;
859
860 ret = submit_bo(submit, submit->cmd[i].idx,
861 &msm_obj, &iova, NULL);
862 if (ret)
863 goto out;
864
865 if (!submit->cmd[i].size ||
866 ((submit->cmd[i].size + submit->cmd[i].offset) >
867 msm_obj->base.size / 4)) {
868 DRM_ERROR("invalid cmdstream size: %u\n", submit->cmd[i].size * 4);
869 ret = -EINVAL;
870 goto out;
871 }
872
873 submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
874
875 if (submit->valid)
876 continue;
877
878 ret = submit_reloc(submit, msm_obj, submit->cmd[i].offset * 4,
879 submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
880 if (ret)
881 goto out;
882 }
883
884 submit->nr_cmds = i;
885
886 submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
887
888 /*
889 * Allocate an id which can be used by WAIT_FENCE ioctl to map back
890 * to the underlying fence.
891 */
892 submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
893 submit->user_fence, 1, INT_MAX, GFP_KERNEL);
894 if (submit->fence_id < 0) {
895 ret = submit->fence_id;
896 submit->fence_id = 0;
897 goto out;
898 }
899
900 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
901 struct sync_file *sync_file = sync_file_create(submit->user_fence);
902 if (!sync_file) {
903 ret = -ENOMEM;
904 goto out;
905 }
906 fd_install(out_fence_fd, sync_file->file);
907 args->fence_fd = out_fence_fd;
908 }
909
910 submit_attach_object_fences(submit);
911
912 /* The scheduler owns a ref now: */
913 msm_gem_submit_get(submit);
914
915 drm_sched_entity_push_job(&submit->base, queue->entity);
916
917 args->fence = submit->fence_id;
918 queue->last_fence = submit->fence_id;
919
920 msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
921 msm_process_post_deps(post_deps, args->nr_out_syncobjs,
922 submit->user_fence);
923
924
925 out:
926 submit_cleanup(submit, !!ret);
927 if (has_ww_ticket)
928 ww_acquire_fini(&submit->ticket);
929 out_unlock:
930 if (ret && (out_fence_fd >= 0))
931 put_unused_fd(out_fence_fd);
932 mutex_unlock(&queue->lock);
933 if (submit)
934 msm_gem_submit_put(submit);
935 out_post_unlock:
936 if (!IS_ERR_OR_NULL(post_deps)) {
937 for (i = 0; i < args->nr_out_syncobjs; ++i) {
938 kfree(post_deps[i].chain);
939 drm_syncobj_put(post_deps[i].syncobj);
940 }
941 kfree(post_deps);
942 }
943
944 if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
945 for (i = 0; i < args->nr_in_syncobjs; ++i) {
946 if (syncobjs_to_reset[i])
947 drm_syncobj_put(syncobjs_to_reset[i]);
948 }
949 kfree(syncobjs_to_reset);
950 }
951
952 return ret;
953 }
954