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
submit_create(struct drm_device * dev,struct msm_gpu * gpu,struct msm_gpu_submitqueue * queue,uint32_t nr_bos,uint32_t nr_cmds)24 static struct msm_gem_submit *submit_create(struct drm_device *dev,
25 struct msm_gpu *gpu,
26 struct msm_gpu_submitqueue *queue, uint32_t nr_bos,
27 uint32_t nr_cmds)
28 {
29 static atomic_t ident = ATOMIC_INIT(0);
30 struct msm_gem_submit *submit;
31 uint64_t sz;
32 int ret;
33
34 sz = struct_size(submit, bos, nr_bos) +
35 ((u64)nr_cmds * sizeof(submit->cmd[0]));
36
37 if (sz > SIZE_MAX)
38 return ERR_PTR(-ENOMEM);
39
40 submit = kzalloc(sz, GFP_KERNEL);
41 if (!submit)
42 return ERR_PTR(-ENOMEM);
43
44 ret = drm_sched_job_init(&submit->base, queue->entity, queue);
45 if (ret) {
46 kfree(submit);
47 return ERR_PTR(ret);
48 }
49
50 kref_init(&submit->ref);
51 submit->dev = dev;
52 submit->aspace = queue->ctx->aspace;
53 submit->gpu = gpu;
54 submit->cmd = (void *)&submit->bos[nr_bos];
55 submit->queue = queue;
56 submit->pid = get_pid(task_pid(current));
57 submit->ring = gpu->rb[queue->ring_nr];
58 submit->fault_dumped = false;
59
60 /* Get a unique identifier for the submission for logging purposes */
61 submit->ident = atomic_inc_return(&ident) - 1;
62
63 INIT_LIST_HEAD(&submit->node);
64
65 return submit;
66 }
67
__msm_gem_submit_destroy(struct kref * kref)68 void __msm_gem_submit_destroy(struct kref *kref)
69 {
70 struct msm_gem_submit *submit =
71 container_of(kref, struct msm_gem_submit, ref);
72 unsigned i;
73
74 if (submit->fence_id) {
75 mutex_lock(&submit->queue->idr_lock);
76 idr_remove(&submit->queue->fence_idr, submit->fence_id);
77 mutex_unlock(&submit->queue->idr_lock);
78 }
79
80 dma_fence_put(submit->user_fence);
81 dma_fence_put(submit->hw_fence);
82
83 put_pid(submit->pid);
84 msm_submitqueue_put(submit->queue);
85
86 for (i = 0; i < submit->nr_cmds; i++)
87 kfree(submit->cmd[i].relocs);
88
89 kfree(submit);
90 }
91
submit_lookup_objects(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)92 static int submit_lookup_objects(struct msm_gem_submit *submit,
93 struct drm_msm_gem_submit *args, struct drm_file *file)
94 {
95 unsigned i;
96 int ret = 0;
97
98 for (i = 0; i < args->nr_bos; i++) {
99 struct drm_msm_gem_submit_bo submit_bo;
100 void __user *userptr =
101 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
102
103 /* make sure we don't have garbage flags, in case we hit
104 * error path before flags is initialized:
105 */
106 submit->bos[i].flags = 0;
107
108 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
109 ret = -EFAULT;
110 i = 0;
111 goto out;
112 }
113
114 /* at least one of READ and/or WRITE flags should be set: */
115 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
116
117 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
118 !(submit_bo.flags & MANDATORY_FLAGS)) {
119 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
120 ret = -EINVAL;
121 i = 0;
122 goto out;
123 }
124
125 submit->bos[i].handle = submit_bo.handle;
126 submit->bos[i].flags = submit_bo.flags;
127 /* in validate_objects() we figure out if this is true: */
128 submit->bos[i].iova = submit_bo.presumed;
129 }
130
131 spin_lock(&file->table_lock);
132
133 for (i = 0; i < args->nr_bos; i++) {
134 struct drm_gem_object *obj;
135
136 /* normally use drm_gem_object_lookup(), but for bulk lookup
137 * all under single table_lock just hit object_idr directly:
138 */
139 obj = idr_find(&file->object_idr, submit->bos[i].handle);
140 if (!obj) {
141 DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
142 ret = -EINVAL;
143 goto out_unlock;
144 }
145
146 drm_gem_object_get(obj);
147
148 submit->bos[i].obj = to_msm_bo(obj);
149 }
150
151 out_unlock:
152 spin_unlock(&file->table_lock);
153
154 out:
155 submit->nr_bos = i;
156
157 return ret;
158 }
159
submit_lookup_cmds(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)160 static int submit_lookup_cmds(struct msm_gem_submit *submit,
161 struct drm_msm_gem_submit *args, struct drm_file *file)
162 {
163 unsigned i;
164 size_t sz;
165 int ret = 0;
166
167 for (i = 0; i < args->nr_cmds; i++) {
168 struct drm_msm_gem_submit_cmd submit_cmd;
169 void __user *userptr =
170 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
171
172 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
173 if (ret) {
174 ret = -EFAULT;
175 goto out;
176 }
177
178 /* validate input from userspace: */
179 switch (submit_cmd.type) {
180 case MSM_SUBMIT_CMD_BUF:
181 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
182 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
183 break;
184 default:
185 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
186 return -EINVAL;
187 }
188
189 if (submit_cmd.size % 4) {
190 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
191 submit_cmd.size);
192 ret = -EINVAL;
193 goto out;
194 }
195
196 submit->cmd[i].type = submit_cmd.type;
197 submit->cmd[i].size = submit_cmd.size / 4;
198 submit->cmd[i].offset = submit_cmd.submit_offset / 4;
199 submit->cmd[i].idx = submit_cmd.submit_idx;
200 submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
201
202 userptr = u64_to_user_ptr(submit_cmd.relocs);
203
204 sz = array_size(submit_cmd.nr_relocs,
205 sizeof(struct drm_msm_gem_submit_reloc));
206 /* check for overflow: */
207 if (sz == SIZE_MAX) {
208 ret = -ENOMEM;
209 goto out;
210 }
211 submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL);
212 if (!submit->cmd[i].relocs) {
213 ret = -ENOMEM;
214 goto out;
215 }
216 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
217 if (ret) {
218 ret = -EFAULT;
219 goto out;
220 }
221 }
222
223 out:
224 return ret;
225 }
226
227 /* Unwind bo state, according to cleanup_flags. In the success case, only
228 * the lock is dropped at the end of the submit (and active/pin ref is dropped
229 * later when the submit is retired).
230 */
submit_cleanup_bo(struct msm_gem_submit * submit,int i,unsigned cleanup_flags)231 static void submit_cleanup_bo(struct msm_gem_submit *submit, int i,
232 unsigned cleanup_flags)
233 {
234 struct drm_gem_object *obj = &submit->bos[i].obj->base;
235 unsigned flags = submit->bos[i].flags & cleanup_flags;
236
237 /*
238 * Clear flags bit before dropping lock, so that the msm_job_run()
239 * path isn't racing with submit_cleanup() (ie. the read/modify/
240 * write is protected by the obj lock in all paths)
241 */
242 submit->bos[i].flags &= ~cleanup_flags;
243
244 if (flags & BO_VMA_PINNED)
245 msm_gem_unpin_vma(submit->bos[i].vma);
246
247 if (flags & BO_OBJ_PINNED)
248 msm_gem_unpin_locked(obj);
249
250 if (flags & BO_LOCKED)
251 dma_resv_unlock(obj->resv);
252 }
253
submit_unlock_unpin_bo(struct msm_gem_submit * submit,int i)254 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
255 {
256 unsigned cleanup_flags = BO_VMA_PINNED | BO_OBJ_PINNED | BO_LOCKED;
257 submit_cleanup_bo(submit, i, cleanup_flags);
258
259 if (!(submit->bos[i].flags & BO_VALID))
260 submit->bos[i].iova = 0;
261 }
262
263 /* This is where we make sure all the bo's are reserved and pin'd: */
submit_lock_objects(struct msm_gem_submit * submit)264 static int submit_lock_objects(struct msm_gem_submit *submit)
265 {
266 int contended, slow_locked = -1, i, ret = 0;
267
268 retry:
269 for (i = 0; i < submit->nr_bos; i++) {
270 struct msm_gem_object *msm_obj = submit->bos[i].obj;
271
272 if (slow_locked == i)
273 slow_locked = -1;
274
275 contended = i;
276
277 if (!(submit->bos[i].flags & BO_LOCKED)) {
278 ret = dma_resv_lock_interruptible(msm_obj->base.resv,
279 &submit->ticket);
280 if (ret)
281 goto fail;
282 submit->bos[i].flags |= BO_LOCKED;
283 }
284 }
285
286 ww_acquire_done(&submit->ticket);
287
288 return 0;
289
290 fail:
291 if (ret == -EALREADY) {
292 DRM_ERROR("handle %u at index %u already on submit list\n",
293 submit->bos[i].handle, i);
294 ret = -EINVAL;
295 }
296
297 for (; i >= 0; i--)
298 submit_unlock_unpin_bo(submit, i);
299
300 if (slow_locked > 0)
301 submit_unlock_unpin_bo(submit, slow_locked);
302
303 if (ret == -EDEADLK) {
304 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
305 /* we lost out in a seqno race, lock and retry.. */
306 ret = dma_resv_lock_slow_interruptible(msm_obj->base.resv,
307 &submit->ticket);
308 if (!ret) {
309 submit->bos[contended].flags |= BO_LOCKED;
310 slow_locked = contended;
311 goto retry;
312 }
313
314 /* Not expecting -EALREADY here, if the bo was already
315 * locked, we should have gotten -EALREADY already from
316 * the dma_resv_lock_interruptable() call.
317 */
318 WARN_ON_ONCE(ret == -EALREADY);
319 }
320
321 return ret;
322 }
323
submit_fence_sync(struct msm_gem_submit * submit,bool no_implicit)324 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
325 {
326 int i, ret = 0;
327
328 for (i = 0; i < submit->nr_bos; i++) {
329 struct drm_gem_object *obj = &submit->bos[i].obj->base;
330 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
331
332 /* NOTE: _reserve_shared() must happen before
333 * _add_shared_fence(), which makes this a slightly
334 * strange place to call it. OTOH this is a
335 * convenient can-fail point to hook it in.
336 */
337 ret = dma_resv_reserve_fences(obj->resv, 1);
338 if (ret)
339 return ret;
340
341 /* If userspace has determined that explicit fencing is
342 * used, it can disable implicit sync on the entire
343 * submit:
344 */
345 if (no_implicit)
346 continue;
347
348 /* Otherwise userspace can ask for implicit sync to be
349 * disabled on specific buffers. This is useful for internal
350 * usermode driver managed buffers, suballocation, etc.
351 */
352 if (submit->bos[i].flags & MSM_SUBMIT_BO_NO_IMPLICIT)
353 continue;
354
355 ret = drm_sched_job_add_implicit_dependencies(&submit->base,
356 obj,
357 write);
358 if (ret)
359 break;
360 }
361
362 return ret;
363 }
364
submit_pin_objects(struct msm_gem_submit * submit)365 static int submit_pin_objects(struct msm_gem_submit *submit)
366 {
367 int i, ret = 0;
368
369 submit->valid = true;
370
371 for (i = 0; i < submit->nr_bos; i++) {
372 struct drm_gem_object *obj = &submit->bos[i].obj->base;
373 struct msm_gem_vma *vma;
374
375 /* if locking succeeded, pin bo: */
376 vma = msm_gem_get_vma_locked(obj, submit->aspace);
377 if (IS_ERR(vma)) {
378 ret = PTR_ERR(vma);
379 break;
380 }
381
382 ret = msm_gem_pin_vma_locked(obj, vma);
383 if (ret)
384 break;
385
386 submit->bos[i].flags |= BO_OBJ_PINNED | BO_VMA_PINNED;
387 submit->bos[i].vma = vma;
388
389 if (vma->iova == submit->bos[i].iova) {
390 submit->bos[i].flags |= BO_VALID;
391 } else {
392 submit->bos[i].iova = vma->iova;
393 /* iova changed, so address in cmdstream is not valid: */
394 submit->bos[i].flags &= ~BO_VALID;
395 submit->valid = false;
396 }
397 }
398
399 return ret;
400 }
401
submit_attach_object_fences(struct msm_gem_submit * submit)402 static void submit_attach_object_fences(struct msm_gem_submit *submit)
403 {
404 int i;
405
406 for (i = 0; i < submit->nr_bos; i++) {
407 struct drm_gem_object *obj = &submit->bos[i].obj->base;
408
409 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
410 dma_resv_add_fence(obj->resv, submit->user_fence,
411 DMA_RESV_USAGE_WRITE);
412 else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
413 dma_resv_add_fence(obj->resv, submit->user_fence,
414 DMA_RESV_USAGE_READ);
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_VMA_PINNED | BO_OBJ_PINNED;
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 drm_gem_object_put(obj);
540 }
541 }
542
543 struct msm_submit_post_dep {
544 struct drm_syncobj *syncobj;
545 uint64_t point;
546 struct dma_fence_chain *chain;
547 };
548
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)549 static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
550 struct drm_file *file,
551 uint64_t in_syncobjs_addr,
552 uint32_t nr_in_syncobjs,
553 size_t syncobj_stride,
554 struct msm_ringbuffer *ring)
555 {
556 struct drm_syncobj **syncobjs = NULL;
557 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
558 int ret = 0;
559 uint32_t i, j;
560
561 syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
562 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
563 if (!syncobjs)
564 return ERR_PTR(-ENOMEM);
565
566 for (i = 0; i < nr_in_syncobjs; ++i) {
567 uint64_t address = in_syncobjs_addr + i * syncobj_stride;
568 struct dma_fence *fence;
569
570 if (copy_from_user(&syncobj_desc,
571 u64_to_user_ptr(address),
572 min(syncobj_stride, sizeof(syncobj_desc)))) {
573 ret = -EFAULT;
574 break;
575 }
576
577 if (syncobj_desc.point &&
578 !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
579 ret = -EOPNOTSUPP;
580 break;
581 }
582
583 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
584 ret = -EINVAL;
585 break;
586 }
587
588 ret = drm_syncobj_find_fence(file, syncobj_desc.handle,
589 syncobj_desc.point, 0, &fence);
590 if (ret)
591 break;
592
593 ret = drm_sched_job_add_dependency(&submit->base, fence);
594 if (ret)
595 break;
596
597 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
598 syncobjs[i] =
599 drm_syncobj_find(file, syncobj_desc.handle);
600 if (!syncobjs[i]) {
601 ret = -EINVAL;
602 break;
603 }
604 }
605 }
606
607 if (ret) {
608 for (j = 0; j <= i; ++j) {
609 if (syncobjs[j])
610 drm_syncobj_put(syncobjs[j]);
611 }
612 kfree(syncobjs);
613 return ERR_PTR(ret);
614 }
615 return syncobjs;
616 }
617
msm_reset_syncobjs(struct drm_syncobj ** syncobjs,uint32_t nr_syncobjs)618 static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
619 uint32_t nr_syncobjs)
620 {
621 uint32_t i;
622
623 for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
624 if (syncobjs[i])
625 drm_syncobj_replace_fence(syncobjs[i], NULL);
626 }
627 }
628
msm_parse_post_deps(struct drm_device * dev,struct drm_file * file,uint64_t syncobjs_addr,uint32_t nr_syncobjs,size_t syncobj_stride)629 static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
630 struct drm_file *file,
631 uint64_t syncobjs_addr,
632 uint32_t nr_syncobjs,
633 size_t syncobj_stride)
634 {
635 struct msm_submit_post_dep *post_deps;
636 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
637 int ret = 0;
638 uint32_t i, j;
639
640 post_deps = kcalloc(nr_syncobjs, sizeof(*post_deps),
641 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
642 if (!post_deps)
643 return ERR_PTR(-ENOMEM);
644
645 for (i = 0; i < nr_syncobjs; ++i) {
646 uint64_t address = syncobjs_addr + i * syncobj_stride;
647
648 if (copy_from_user(&syncobj_desc,
649 u64_to_user_ptr(address),
650 min(syncobj_stride, sizeof(syncobj_desc)))) {
651 ret = -EFAULT;
652 break;
653 }
654
655 post_deps[i].point = syncobj_desc.point;
656
657 if (syncobj_desc.flags) {
658 ret = -EINVAL;
659 break;
660 }
661
662 if (syncobj_desc.point) {
663 if (!drm_core_check_feature(dev,
664 DRIVER_SYNCOBJ_TIMELINE)) {
665 ret = -EOPNOTSUPP;
666 break;
667 }
668
669 post_deps[i].chain = dma_fence_chain_alloc();
670 if (!post_deps[i].chain) {
671 ret = -ENOMEM;
672 break;
673 }
674 }
675
676 post_deps[i].syncobj =
677 drm_syncobj_find(file, syncobj_desc.handle);
678 if (!post_deps[i].syncobj) {
679 ret = -EINVAL;
680 break;
681 }
682 }
683
684 if (ret) {
685 for (j = 0; j <= i; ++j) {
686 dma_fence_chain_free(post_deps[j].chain);
687 if (post_deps[j].syncobj)
688 drm_syncobj_put(post_deps[j].syncobj);
689 }
690
691 kfree(post_deps);
692 return ERR_PTR(ret);
693 }
694
695 return post_deps;
696 }
697
msm_process_post_deps(struct msm_submit_post_dep * post_deps,uint32_t count,struct dma_fence * fence)698 static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
699 uint32_t count, struct dma_fence *fence)
700 {
701 uint32_t i;
702
703 for (i = 0; post_deps && i < count; ++i) {
704 if (post_deps[i].chain) {
705 drm_syncobj_add_point(post_deps[i].syncobj,
706 post_deps[i].chain,
707 fence, post_deps[i].point);
708 post_deps[i].chain = NULL;
709 } else {
710 drm_syncobj_replace_fence(post_deps[i].syncobj,
711 fence);
712 }
713 }
714 }
715
msm_ioctl_gem_submit(struct drm_device * dev,void * data,struct drm_file * file)716 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
717 struct drm_file *file)
718 {
719 struct msm_drm_private *priv = dev->dev_private;
720 struct drm_msm_gem_submit *args = data;
721 struct msm_file_private *ctx = file->driver_priv;
722 struct msm_gem_submit *submit;
723 struct msm_gpu *gpu = priv->gpu;
724 struct msm_gpu_submitqueue *queue;
725 struct msm_ringbuffer *ring;
726 struct msm_submit_post_dep *post_deps = NULL;
727 struct drm_syncobj **syncobjs_to_reset = NULL;
728 int out_fence_fd = -1;
729 bool has_ww_ticket = false;
730 unsigned i;
731 int ret;
732
733 if (!gpu)
734 return -ENXIO;
735
736 if (args->pad)
737 return -EINVAL;
738
739 if (unlikely(!ctx->aspace) && !capable(CAP_SYS_RAWIO)) {
740 DRM_ERROR_RATELIMITED("IOMMU support or CAP_SYS_RAWIO required!\n");
741 return -EPERM;
742 }
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 ring = gpu->rb[queue->ring_nr];
764
765 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
766 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
767 if (out_fence_fd < 0) {
768 ret = out_fence_fd;
769 return ret;
770 }
771 }
772
773 submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds);
774 if (IS_ERR(submit))
775 return PTR_ERR(submit);
776
777 trace_msm_gpu_submit(pid_nr(submit->pid), ring->id, submit->ident,
778 args->nr_bos, args->nr_cmds);
779
780 ret = mutex_lock_interruptible(&queue->lock);
781 if (ret)
782 goto out_post_unlock;
783
784 if (args->flags & MSM_SUBMIT_SUDO)
785 submit->in_rb = true;
786
787 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
788 struct dma_fence *in_fence;
789
790 in_fence = sync_file_get_fence(args->fence_fd);
791
792 if (!in_fence) {
793 ret = -EINVAL;
794 goto out_unlock;
795 }
796
797 ret = drm_sched_job_add_dependency(&submit->base, in_fence);
798 if (ret)
799 goto out_unlock;
800 }
801
802 if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
803 syncobjs_to_reset = msm_parse_deps(submit, file,
804 args->in_syncobjs,
805 args->nr_in_syncobjs,
806 args->syncobj_stride, ring);
807 if (IS_ERR(syncobjs_to_reset)) {
808 ret = PTR_ERR(syncobjs_to_reset);
809 goto out_unlock;
810 }
811 }
812
813 if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
814 post_deps = msm_parse_post_deps(dev, file,
815 args->out_syncobjs,
816 args->nr_out_syncobjs,
817 args->syncobj_stride);
818 if (IS_ERR(post_deps)) {
819 ret = PTR_ERR(post_deps);
820 goto out_unlock;
821 }
822 }
823
824 ret = submit_lookup_objects(submit, args, file);
825 if (ret)
826 goto out;
827
828 ret = submit_lookup_cmds(submit, args, file);
829 if (ret)
830 goto out;
831
832 /* copy_*_user while holding a ww ticket upsets lockdep */
833 ww_acquire_init(&submit->ticket, &reservation_ww_class);
834 has_ww_ticket = true;
835 ret = submit_lock_objects(submit);
836 if (ret)
837 goto out;
838
839 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
840 if (ret)
841 goto out;
842
843 ret = submit_pin_objects(submit);
844 if (ret)
845 goto out;
846
847 for (i = 0; i < args->nr_cmds; i++) {
848 struct msm_gem_object *msm_obj;
849 uint64_t iova;
850
851 ret = submit_bo(submit, submit->cmd[i].idx,
852 &msm_obj, &iova, NULL);
853 if (ret)
854 goto out;
855
856 if (!submit->cmd[i].size ||
857 ((submit->cmd[i].size + submit->cmd[i].offset) >
858 msm_obj->base.size / 4)) {
859 DRM_ERROR("invalid cmdstream size: %u\n", submit->cmd[i].size * 4);
860 ret = -EINVAL;
861 goto out;
862 }
863
864 submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
865
866 if (submit->valid)
867 continue;
868
869 ret = submit_reloc(submit, msm_obj, submit->cmd[i].offset * 4,
870 submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
871 if (ret)
872 goto out;
873 }
874
875 submit->nr_cmds = i;
876
877 mutex_lock(&queue->idr_lock);
878
879 /*
880 * If using userspace provided seqno fence, validate that the id
881 * is available before arming sched job. Since access to fence_idr
882 * is serialized on the queue lock, the slot should be still avail
883 * after the job is armed
884 */
885 if ((args->flags & MSM_SUBMIT_FENCE_SN_IN) &&
886 idr_find(&queue->fence_idr, args->fence)) {
887 mutex_unlock(&queue->idr_lock);
888 ret = -EINVAL;
889 goto out;
890 }
891
892 drm_sched_job_arm(&submit->base);
893
894 submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
895
896 if (args->flags & MSM_SUBMIT_FENCE_SN_IN) {
897 /*
898 * Userspace has assigned the seqno fence that it wants
899 * us to use. It is an error to pick a fence sequence
900 * number that is not available.
901 */
902 submit->fence_id = args->fence;
903 ret = idr_alloc_u32(&queue->fence_idr, submit->user_fence,
904 &submit->fence_id, submit->fence_id,
905 GFP_KERNEL);
906 /*
907 * We've already validated that the fence_id slot is valid,
908 * so if idr_alloc_u32 failed, it is a kernel bug
909 */
910 WARN_ON(ret);
911 } else {
912 /*
913 * Allocate an id which can be used by WAIT_FENCE ioctl to map
914 * back to the underlying fence.
915 */
916 submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
917 submit->user_fence, 1,
918 INT_MAX, GFP_KERNEL);
919 }
920
921 mutex_unlock(&queue->idr_lock);
922
923 if (submit->fence_id < 0) {
924 ret = submit->fence_id;
925 submit->fence_id = 0;
926 }
927
928 if (ret == 0 && args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
929 struct sync_file *sync_file = sync_file_create(submit->user_fence);
930 if (!sync_file) {
931 ret = -ENOMEM;
932 } else {
933 fd_install(out_fence_fd, sync_file->file);
934 args->fence_fd = out_fence_fd;
935 }
936 }
937
938 submit_attach_object_fences(submit);
939
940 /* The scheduler owns a ref now: */
941 msm_gem_submit_get(submit);
942
943 drm_sched_entity_push_job(&submit->base);
944
945 args->fence = submit->fence_id;
946 queue->last_fence = submit->fence_id;
947
948 msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
949 msm_process_post_deps(post_deps, args->nr_out_syncobjs,
950 submit->user_fence);
951
952
953 out:
954 submit_cleanup(submit, !!ret);
955 if (has_ww_ticket)
956 ww_acquire_fini(&submit->ticket);
957 out_unlock:
958 if (ret && (out_fence_fd >= 0))
959 put_unused_fd(out_fence_fd);
960 mutex_unlock(&queue->lock);
961 out_post_unlock:
962 msm_gem_submit_put(submit);
963 if (!IS_ERR_OR_NULL(post_deps)) {
964 for (i = 0; i < args->nr_out_syncobjs; ++i) {
965 kfree(post_deps[i].chain);
966 drm_syncobj_put(post_deps[i].syncobj);
967 }
968 kfree(post_deps);
969 }
970
971 if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
972 for (i = 0; i < args->nr_in_syncobjs; ++i) {
973 if (syncobjs_to_reset[i])
974 drm_syncobj_put(syncobjs_to_reset[i]);
975 }
976 kfree(syncobjs_to_reset);
977 }
978
979 return ret;
980 }
981