Loading...
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
27#define BO_PINNED 0x2000
28
29static struct msm_gem_submit *submit_create(struct drm_device *dev,
30 struct msm_gpu *gpu, struct msm_gem_address_space *aspace,
31 struct msm_gpu_submitqueue *queue, uint32_t nr_bos,
32 uint32_t nr_cmds)
33{
34 struct msm_gem_submit *submit;
35 uint64_t sz = struct_size(submit, bos, nr_bos) +
36 ((u64)nr_cmds * sizeof(submit->cmd[0]));
37
38 if (sz > SIZE_MAX)
39 return NULL;
40
41 submit = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
42 if (!submit)
43 return NULL;
44
45 submit->dev = dev;
46 submit->aspace = aspace;
47 submit->gpu = gpu;
48 submit->fence = NULL;
49 submit->cmd = (void *)&submit->bos[nr_bos];
50 submit->queue = queue;
51 submit->ring = gpu->rb[queue->prio];
52
53 /* initially, until copy_from_user() and bo lookup succeeds: */
54 submit->nr_bos = 0;
55 submit->nr_cmds = 0;
56
57 INIT_LIST_HEAD(&submit->node);
58 INIT_LIST_HEAD(&submit->bo_list);
59
60 return submit;
61}
62
63void msm_gem_submit_free(struct msm_gem_submit *submit)
64{
65 dma_fence_put(submit->fence);
66 list_del(&submit->node);
67 put_pid(submit->pid);
68 msm_submitqueue_put(submit->queue);
69
70 kfree(submit);
71}
72
73static int submit_lookup_objects(struct msm_gem_submit *submit,
74 struct drm_msm_gem_submit *args, struct drm_file *file)
75{
76 unsigned i;
77 int ret = 0;
78
79 for (i = 0; i < args->nr_bos; i++) {
80 struct drm_msm_gem_submit_bo submit_bo;
81 void __user *userptr =
82 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
83
84 /* make sure we don't have garbage flags, in case we hit
85 * error path before flags is initialized:
86 */
87 submit->bos[i].flags = 0;
88
89 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
90 ret = -EFAULT;
91 i = 0;
92 goto out;
93 }
94
95/* at least one of READ and/or WRITE flags should be set: */
96#define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
97
98 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
99 !(submit_bo.flags & MANDATORY_FLAGS)) {
100 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
101 ret = -EINVAL;
102 i = 0;
103 goto out;
104 }
105
106 submit->bos[i].handle = submit_bo.handle;
107 submit->bos[i].flags = submit_bo.flags;
108 /* in validate_objects() we figure out if this is true: */
109 submit->bos[i].iova = submit_bo.presumed;
110 }
111
112 spin_lock(&file->table_lock);
113
114 for (i = 0; i < args->nr_bos; i++) {
115 struct drm_gem_object *obj;
116 struct msm_gem_object *msm_obj;
117
118 /* normally use drm_gem_object_lookup(), but for bulk lookup
119 * all under single table_lock just hit object_idr directly:
120 */
121 obj = idr_find(&file->object_idr, submit->bos[i].handle);
122 if (!obj) {
123 DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
124 ret = -EINVAL;
125 goto out_unlock;
126 }
127
128 msm_obj = to_msm_bo(obj);
129
130 if (!list_empty(&msm_obj->submit_entry)) {
131 DRM_ERROR("handle %u at index %u already on submit list\n",
132 submit->bos[i].handle, i);
133 ret = -EINVAL;
134 goto out_unlock;
135 }
136
137 drm_gem_object_get(obj);
138
139 submit->bos[i].obj = msm_obj;
140
141 list_add_tail(&msm_obj->submit_entry, &submit->bo_list);
142 }
143
144out_unlock:
145 spin_unlock(&file->table_lock);
146
147out:
148 submit->nr_bos = i;
149
150 return ret;
151}
152
153static void submit_unlock_unpin_bo(struct msm_gem_submit *submit,
154 int i, bool backoff)
155{
156 struct msm_gem_object *msm_obj = submit->bos[i].obj;
157
158 if (submit->bos[i].flags & BO_PINNED)
159 msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
160
161 if (submit->bos[i].flags & BO_LOCKED)
162 dma_resv_unlock(msm_obj->base.resv);
163
164 if (backoff && !(submit->bos[i].flags & BO_VALID))
165 submit->bos[i].iova = 0;
166
167 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED);
168}
169
170/* This is where we make sure all the bo's are reserved and pin'd: */
171static int submit_lock_objects(struct msm_gem_submit *submit)
172{
173 int contended, slow_locked = -1, i, ret = 0;
174
175retry:
176 for (i = 0; i < submit->nr_bos; i++) {
177 struct msm_gem_object *msm_obj = submit->bos[i].obj;
178
179 if (slow_locked == i)
180 slow_locked = -1;
181
182 contended = i;
183
184 if (!(submit->bos[i].flags & BO_LOCKED)) {
185 ret = dma_resv_lock_interruptible(msm_obj->base.resv,
186 &submit->ticket);
187 if (ret)
188 goto fail;
189 submit->bos[i].flags |= BO_LOCKED;
190 }
191 }
192
193 ww_acquire_done(&submit->ticket);
194
195 return 0;
196
197fail:
198 for (; i >= 0; i--)
199 submit_unlock_unpin_bo(submit, i, true);
200
201 if (slow_locked > 0)
202 submit_unlock_unpin_bo(submit, slow_locked, true);
203
204 if (ret == -EDEADLK) {
205 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
206 /* we lost out in a seqno race, lock and retry.. */
207 ret = dma_resv_lock_slow_interruptible(msm_obj->base.resv,
208 &submit->ticket);
209 if (!ret) {
210 submit->bos[contended].flags |= BO_LOCKED;
211 slow_locked = contended;
212 goto retry;
213 }
214 }
215
216 return ret;
217}
218
219static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
220{
221 int i, ret = 0;
222
223 for (i = 0; i < submit->nr_bos; i++) {
224 struct msm_gem_object *msm_obj = submit->bos[i].obj;
225 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
226
227 if (!write) {
228 /* NOTE: _reserve_shared() must happen before
229 * _add_shared_fence(), which makes this a slightly
230 * strange place to call it. OTOH this is a
231 * convenient can-fail point to hook it in.
232 */
233 ret = dma_resv_reserve_shared(msm_obj->base.resv,
234 1);
235 if (ret)
236 return ret;
237 }
238
239 if (no_implicit)
240 continue;
241
242 ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
243 write);
244 if (ret)
245 break;
246 }
247
248 return ret;
249}
250
251static int submit_pin_objects(struct msm_gem_submit *submit)
252{
253 int i, ret = 0;
254
255 submit->valid = true;
256
257 for (i = 0; i < submit->nr_bos; i++) {
258 struct msm_gem_object *msm_obj = submit->bos[i].obj;
259 uint64_t iova;
260
261 /* if locking succeeded, pin bo: */
262 ret = msm_gem_get_and_pin_iova(&msm_obj->base,
263 submit->aspace, &iova);
264
265 if (ret)
266 break;
267
268 submit->bos[i].flags |= BO_PINNED;
269
270 if (iova == submit->bos[i].iova) {
271 submit->bos[i].flags |= BO_VALID;
272 } else {
273 submit->bos[i].iova = iova;
274 /* iova changed, so address in cmdstream is not valid: */
275 submit->bos[i].flags &= ~BO_VALID;
276 submit->valid = false;
277 }
278 }
279
280 return ret;
281}
282
283static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
284 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
285{
286 if (idx >= submit->nr_bos) {
287 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
288 idx, submit->nr_bos);
289 return -EINVAL;
290 }
291
292 if (obj)
293 *obj = submit->bos[idx].obj;
294 if (iova)
295 *iova = submit->bos[idx].iova;
296 if (valid)
297 *valid = !!(submit->bos[idx].flags & BO_VALID);
298
299 return 0;
300}
301
302/* process the reloc's and patch up the cmdstream as needed: */
303static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
304 uint32_t offset, uint32_t nr_relocs, uint64_t relocs)
305{
306 uint32_t i, last_offset = 0;
307 uint32_t *ptr;
308 int ret = 0;
309
310 if (!nr_relocs)
311 return 0;
312
313 if (offset % 4) {
314 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
315 return -EINVAL;
316 }
317
318 /* For now, just map the entire thing. Eventually we probably
319 * to do it page-by-page, w/ kmap() if not vmap()d..
320 */
321 ptr = msm_gem_get_vaddr(&obj->base);
322
323 if (IS_ERR(ptr)) {
324 ret = PTR_ERR(ptr);
325 DBG("failed to map: %d", ret);
326 return ret;
327 }
328
329 for (i = 0; i < nr_relocs; i++) {
330 struct drm_msm_gem_submit_reloc submit_reloc;
331 void __user *userptr =
332 u64_to_user_ptr(relocs + (i * sizeof(submit_reloc)));
333 uint32_t off;
334 uint64_t iova;
335 bool valid;
336
337 if (copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc))) {
338 ret = -EFAULT;
339 goto out;
340 }
341
342 if (submit_reloc.submit_offset % 4) {
343 DRM_ERROR("non-aligned reloc offset: %u\n",
344 submit_reloc.submit_offset);
345 ret = -EINVAL;
346 goto out;
347 }
348
349 /* offset in dwords: */
350 off = submit_reloc.submit_offset / 4;
351
352 if ((off >= (obj->base.size / 4)) ||
353 (off < last_offset)) {
354 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
355 ret = -EINVAL;
356 goto out;
357 }
358
359 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
360 if (ret)
361 goto out;
362
363 if (valid)
364 continue;
365
366 iova += submit_reloc.reloc_offset;
367
368 if (submit_reloc.shift < 0)
369 iova >>= -submit_reloc.shift;
370 else
371 iova <<= submit_reloc.shift;
372
373 ptr[off] = iova | submit_reloc.or;
374
375 last_offset = off;
376 }
377
378out:
379 msm_gem_put_vaddr(&obj->base);
380
381 return ret;
382}
383
384static void submit_cleanup(struct msm_gem_submit *submit)
385{
386 unsigned i;
387
388 for (i = 0; i < submit->nr_bos; i++) {
389 struct msm_gem_object *msm_obj = submit->bos[i].obj;
390 submit_unlock_unpin_bo(submit, i, false);
391 list_del_init(&msm_obj->submit_entry);
392 drm_gem_object_put_locked(&msm_obj->base);
393 }
394}
395
396
397struct msm_submit_post_dep {
398 struct drm_syncobj *syncobj;
399 uint64_t point;
400 struct dma_fence_chain *chain;
401};
402
403static struct drm_syncobj **msm_wait_deps(struct drm_device *dev,
404 struct drm_file *file,
405 uint64_t in_syncobjs_addr,
406 uint32_t nr_in_syncobjs,
407 size_t syncobj_stride,
408 struct msm_ringbuffer *ring)
409{
410 struct drm_syncobj **syncobjs = NULL;
411 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
412 int ret = 0;
413 uint32_t i, j;
414
415 syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
416 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
417 if (!syncobjs)
418 return ERR_PTR(-ENOMEM);
419
420 for (i = 0; i < nr_in_syncobjs; ++i) {
421 uint64_t address = in_syncobjs_addr + i * syncobj_stride;
422 struct dma_fence *fence;
423
424 if (copy_from_user(&syncobj_desc,
425 u64_to_user_ptr(address),
426 min(syncobj_stride, sizeof(syncobj_desc)))) {
427 ret = -EFAULT;
428 break;
429 }
430
431 if (syncobj_desc.point &&
432 !drm_core_check_feature(dev, DRIVER_SYNCOBJ_TIMELINE)) {
433 ret = -EOPNOTSUPP;
434 break;
435 }
436
437 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
438 ret = -EINVAL;
439 break;
440 }
441
442 ret = drm_syncobj_find_fence(file, syncobj_desc.handle,
443 syncobj_desc.point, 0, &fence);
444 if (ret)
445 break;
446
447 if (!dma_fence_match_context(fence, ring->fctx->context))
448 ret = dma_fence_wait(fence, true);
449
450 dma_fence_put(fence);
451 if (ret)
452 break;
453
454 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
455 syncobjs[i] =
456 drm_syncobj_find(file, syncobj_desc.handle);
457 if (!syncobjs[i]) {
458 ret = -EINVAL;
459 break;
460 }
461 }
462 }
463
464 if (ret) {
465 for (j = 0; j <= i; ++j) {
466 if (syncobjs[j])
467 drm_syncobj_put(syncobjs[j]);
468 }
469 kfree(syncobjs);
470 return ERR_PTR(ret);
471 }
472 return syncobjs;
473}
474
475static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
476 uint32_t nr_syncobjs)
477{
478 uint32_t i;
479
480 for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
481 if (syncobjs[i])
482 drm_syncobj_replace_fence(syncobjs[i], NULL);
483 }
484}
485
486static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
487 struct drm_file *file,
488 uint64_t syncobjs_addr,
489 uint32_t nr_syncobjs,
490 size_t syncobj_stride)
491{
492 struct msm_submit_post_dep *post_deps;
493 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
494 int ret = 0;
495 uint32_t i, j;
496
497 post_deps = kmalloc_array(nr_syncobjs, sizeof(*post_deps),
498 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
499 if (!post_deps)
500 return ERR_PTR(-ENOMEM);
501
502 for (i = 0; i < nr_syncobjs; ++i) {
503 uint64_t address = syncobjs_addr + i * syncobj_stride;
504
505 if (copy_from_user(&syncobj_desc,
506 u64_to_user_ptr(address),
507 min(syncobj_stride, sizeof(syncobj_desc)))) {
508 ret = -EFAULT;
509 break;
510 }
511
512 post_deps[i].point = syncobj_desc.point;
513 post_deps[i].chain = NULL;
514
515 if (syncobj_desc.flags) {
516 ret = -EINVAL;
517 break;
518 }
519
520 if (syncobj_desc.point) {
521 if (!drm_core_check_feature(dev,
522 DRIVER_SYNCOBJ_TIMELINE)) {
523 ret = -EOPNOTSUPP;
524 break;
525 }
526
527 post_deps[i].chain =
528 kmalloc(sizeof(*post_deps[i].chain),
529 GFP_KERNEL);
530 if (!post_deps[i].chain) {
531 ret = -ENOMEM;
532 break;
533 }
534 }
535
536 post_deps[i].syncobj =
537 drm_syncobj_find(file, syncobj_desc.handle);
538 if (!post_deps[i].syncobj) {
539 ret = -EINVAL;
540 break;
541 }
542 }
543
544 if (ret) {
545 for (j = 0; j <= i; ++j) {
546 kfree(post_deps[j].chain);
547 if (post_deps[j].syncobj)
548 drm_syncobj_put(post_deps[j].syncobj);
549 }
550
551 kfree(post_deps);
552 return ERR_PTR(ret);
553 }
554
555 return post_deps;
556}
557
558static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
559 uint32_t count, struct dma_fence *fence)
560{
561 uint32_t i;
562
563 for (i = 0; post_deps && i < count; ++i) {
564 if (post_deps[i].chain) {
565 drm_syncobj_add_point(post_deps[i].syncobj,
566 post_deps[i].chain,
567 fence, post_deps[i].point);
568 post_deps[i].chain = NULL;
569 } else {
570 drm_syncobj_replace_fence(post_deps[i].syncobj,
571 fence);
572 }
573 }
574}
575
576int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
577 struct drm_file *file)
578{
579 static atomic_t ident = ATOMIC_INIT(0);
580 struct msm_drm_private *priv = dev->dev_private;
581 struct drm_msm_gem_submit *args = data;
582 struct msm_file_private *ctx = file->driver_priv;
583 struct msm_gem_submit *submit;
584 struct msm_gpu *gpu = priv->gpu;
585 struct sync_file *sync_file = NULL;
586 struct msm_gpu_submitqueue *queue;
587 struct msm_ringbuffer *ring;
588 struct msm_submit_post_dep *post_deps = NULL;
589 struct drm_syncobj **syncobjs_to_reset = NULL;
590 int out_fence_fd = -1;
591 struct pid *pid = get_pid(task_pid(current));
592 bool has_ww_ticket = false;
593 unsigned i;
594 int ret, submitid;
595 if (!gpu)
596 return -ENXIO;
597
598 if (args->pad)
599 return -EINVAL;
600
601 /* for now, we just have 3d pipe.. eventually this would need to
602 * be more clever to dispatch to appropriate gpu module:
603 */
604 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
605 return -EINVAL;
606
607 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
608 return -EINVAL;
609
610 if (args->flags & MSM_SUBMIT_SUDO) {
611 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
612 !capable(CAP_SYS_RAWIO))
613 return -EINVAL;
614 }
615
616 queue = msm_submitqueue_get(ctx, args->queueid);
617 if (!queue)
618 return -ENOENT;
619
620 /* Get a unique identifier for the submission for logging purposes */
621 submitid = atomic_inc_return(&ident) - 1;
622
623 ring = gpu->rb[queue->prio];
624 trace_msm_gpu_submit(pid_nr(pid), ring->id, submitid,
625 args->nr_bos, args->nr_cmds);
626
627 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
628 struct dma_fence *in_fence;
629
630 in_fence = sync_file_get_fence(args->fence_fd);
631
632 if (!in_fence)
633 return -EINVAL;
634
635 /*
636 * Wait if the fence is from a foreign context, or if the fence
637 * array contains any fence from a foreign context.
638 */
639 ret = 0;
640 if (!dma_fence_match_context(in_fence, ring->fctx->context))
641 ret = dma_fence_wait(in_fence, true);
642
643 dma_fence_put(in_fence);
644 if (ret)
645 return ret;
646 }
647
648 if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
649 syncobjs_to_reset = msm_wait_deps(dev, file,
650 args->in_syncobjs,
651 args->nr_in_syncobjs,
652 args->syncobj_stride, ring);
653 if (IS_ERR(syncobjs_to_reset))
654 return PTR_ERR(syncobjs_to_reset);
655 }
656
657 if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
658 post_deps = msm_parse_post_deps(dev, file,
659 args->out_syncobjs,
660 args->nr_out_syncobjs,
661 args->syncobj_stride);
662 if (IS_ERR(post_deps)) {
663 ret = PTR_ERR(post_deps);
664 goto out_post_unlock;
665 }
666 }
667
668 ret = mutex_lock_interruptible(&dev->struct_mutex);
669 if (ret)
670 goto out_post_unlock;
671
672 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
673 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
674 if (out_fence_fd < 0) {
675 ret = out_fence_fd;
676 goto out_unlock;
677 }
678 }
679
680 submit = submit_create(dev, gpu, ctx->aspace, queue, args->nr_bos,
681 args->nr_cmds);
682 if (!submit) {
683 ret = -ENOMEM;
684 goto out_unlock;
685 }
686
687 submit->pid = pid;
688 submit->ident = submitid;
689
690 if (args->flags & MSM_SUBMIT_SUDO)
691 submit->in_rb = true;
692
693 ret = submit_lookup_objects(submit, args, file);
694 if (ret)
695 goto out;
696
697 /* copy_*_user while holding a ww ticket upsets lockdep */
698 ww_acquire_init(&submit->ticket, &reservation_ww_class);
699 has_ww_ticket = true;
700 ret = submit_lock_objects(submit);
701 if (ret)
702 goto out;
703
704 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
705 if (ret)
706 goto out;
707
708 ret = submit_pin_objects(submit);
709 if (ret)
710 goto out;
711
712 for (i = 0; i < args->nr_cmds; i++) {
713 struct drm_msm_gem_submit_cmd submit_cmd;
714 void __user *userptr =
715 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
716 struct msm_gem_object *msm_obj;
717 uint64_t iova;
718
719 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
720 if (ret) {
721 ret = -EFAULT;
722 goto out;
723 }
724
725 /* validate input from userspace: */
726 switch (submit_cmd.type) {
727 case MSM_SUBMIT_CMD_BUF:
728 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
729 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
730 break;
731 default:
732 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
733 ret = -EINVAL;
734 goto out;
735 }
736
737 ret = submit_bo(submit, submit_cmd.submit_idx,
738 &msm_obj, &iova, NULL);
739 if (ret)
740 goto out;
741
742 if (submit_cmd.size % 4) {
743 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
744 submit_cmd.size);
745 ret = -EINVAL;
746 goto out;
747 }
748
749 if (!submit_cmd.size ||
750 ((submit_cmd.size + submit_cmd.submit_offset) >
751 msm_obj->base.size)) {
752 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size);
753 ret = -EINVAL;
754 goto out;
755 }
756
757 submit->cmd[i].type = submit_cmd.type;
758 submit->cmd[i].size = submit_cmd.size / 4;
759 submit->cmd[i].iova = iova + submit_cmd.submit_offset;
760 submit->cmd[i].idx = submit_cmd.submit_idx;
761
762 if (submit->valid)
763 continue;
764
765 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset,
766 submit_cmd.nr_relocs, submit_cmd.relocs);
767 if (ret)
768 goto out;
769 }
770
771 submit->nr_cmds = i;
772
773 submit->fence = msm_fence_alloc(ring->fctx);
774 if (IS_ERR(submit->fence)) {
775 ret = PTR_ERR(submit->fence);
776 submit->fence = NULL;
777 goto out;
778 }
779
780 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
781 sync_file = sync_file_create(submit->fence);
782 if (!sync_file) {
783 ret = -ENOMEM;
784 goto out;
785 }
786 }
787
788 msm_gpu_submit(gpu, submit, ctx);
789
790 args->fence = submit->fence->seqno;
791
792 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
793 fd_install(out_fence_fd, sync_file->file);
794 args->fence_fd = out_fence_fd;
795 }
796
797 msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
798 msm_process_post_deps(post_deps, args->nr_out_syncobjs,
799 submit->fence);
800
801
802out:
803 submit_cleanup(submit);
804 if (has_ww_ticket)
805 ww_acquire_fini(&submit->ticket);
806 if (ret)
807 msm_gem_submit_free(submit);
808out_unlock:
809 if (ret && (out_fence_fd >= 0))
810 put_unused_fd(out_fence_fd);
811 mutex_unlock(&dev->struct_mutex);
812
813out_post_unlock:
814 if (!IS_ERR_OR_NULL(post_deps)) {
815 for (i = 0; i < args->nr_out_syncobjs; ++i) {
816 kfree(post_deps[i].chain);
817 drm_syncobj_put(post_deps[i].syncobj);
818 }
819 kfree(post_deps);
820 }
821
822 if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
823 for (i = 0; i < args->nr_in_syncobjs; ++i) {
824 if (syncobjs_to_reset[i])
825 drm_syncobj_put(syncobjs_to_reset[i]);
826 }
827 kfree(syncobjs_to_reset);
828 }
829
830 return ret;
831}
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
24static 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
68void __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
92static 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
151out_unlock:
152 spin_unlock(&file->table_lock);
153
154out:
155 submit->nr_bos = i;
156
157 return ret;
158}
159
160static 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 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
213 if (ret) {
214 ret = -EFAULT;
215 goto out;
216 }
217 }
218
219out:
220 return ret;
221}
222
223/* Unwind bo state, according to cleanup_flags. In the success case, only
224 * the lock is dropped at the end of the submit (and active/pin ref is dropped
225 * later when the submit is retired).
226 */
227static void submit_cleanup_bo(struct msm_gem_submit *submit, int i,
228 unsigned cleanup_flags)
229{
230 struct drm_gem_object *obj = &submit->bos[i].obj->base;
231 unsigned flags = submit->bos[i].flags & cleanup_flags;
232
233 /*
234 * Clear flags bit before dropping lock, so that the msm_job_run()
235 * path isn't racing with submit_cleanup() (ie. the read/modify/
236 * write is protected by the obj lock in all paths)
237 */
238 submit->bos[i].flags &= ~cleanup_flags;
239
240 if (flags & BO_VMA_PINNED)
241 msm_gem_unpin_vma(submit->bos[i].vma);
242
243 if (flags & BO_OBJ_PINNED)
244 msm_gem_unpin_locked(obj);
245
246 if (flags & BO_LOCKED)
247 dma_resv_unlock(obj->resv);
248}
249
250static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
251{
252 unsigned cleanup_flags = BO_VMA_PINNED | BO_OBJ_PINNED | BO_LOCKED;
253 submit_cleanup_bo(submit, i, cleanup_flags);
254
255 if (!(submit->bos[i].flags & BO_VALID))
256 submit->bos[i].iova = 0;
257}
258
259/* This is where we make sure all the bo's are reserved and pin'd: */
260static int submit_lock_objects(struct msm_gem_submit *submit)
261{
262 int contended, slow_locked = -1, i, ret = 0;
263
264retry:
265 for (i = 0; i < submit->nr_bos; i++) {
266 struct msm_gem_object *msm_obj = submit->bos[i].obj;
267
268 if (slow_locked == i)
269 slow_locked = -1;
270
271 contended = i;
272
273 if (!(submit->bos[i].flags & BO_LOCKED)) {
274 ret = dma_resv_lock_interruptible(msm_obj->base.resv,
275 &submit->ticket);
276 if (ret)
277 goto fail;
278 submit->bos[i].flags |= BO_LOCKED;
279 }
280 }
281
282 ww_acquire_done(&submit->ticket);
283
284 return 0;
285
286fail:
287 if (ret == -EALREADY) {
288 DRM_ERROR("handle %u at index %u already on submit list\n",
289 submit->bos[i].handle, i);
290 ret = -EINVAL;
291 }
292
293 for (; i >= 0; i--)
294 submit_unlock_unpin_bo(submit, i);
295
296 if (slow_locked > 0)
297 submit_unlock_unpin_bo(submit, slow_locked);
298
299 if (ret == -EDEADLK) {
300 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
301 /* we lost out in a seqno race, lock and retry.. */
302 ret = dma_resv_lock_slow_interruptible(msm_obj->base.resv,
303 &submit->ticket);
304 if (!ret) {
305 submit->bos[contended].flags |= BO_LOCKED;
306 slow_locked = contended;
307 goto retry;
308 }
309
310 /* Not expecting -EALREADY here, if the bo was already
311 * locked, we should have gotten -EALREADY already from
312 * the dma_resv_lock_interruptable() call.
313 */
314 WARN_ON_ONCE(ret == -EALREADY);
315 }
316
317 return ret;
318}
319
320static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
321{
322 int i, ret = 0;
323
324 for (i = 0; i < submit->nr_bos; i++) {
325 struct drm_gem_object *obj = &submit->bos[i].obj->base;
326 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
327
328 /* NOTE: _reserve_shared() must happen before
329 * _add_shared_fence(), which makes this a slightly
330 * strange place to call it. OTOH this is a
331 * convenient can-fail point to hook it in.
332 */
333 ret = dma_resv_reserve_fences(obj->resv, 1);
334 if (ret)
335 return ret;
336
337 if (no_implicit)
338 continue;
339
340 ret = drm_sched_job_add_implicit_dependencies(&submit->base,
341 obj,
342 write);
343 if (ret)
344 break;
345 }
346
347 return ret;
348}
349
350static int submit_pin_objects(struct msm_gem_submit *submit)
351{
352 int i, ret = 0;
353
354 submit->valid = true;
355
356 for (i = 0; i < submit->nr_bos; i++) {
357 struct drm_gem_object *obj = &submit->bos[i].obj->base;
358 struct msm_gem_vma *vma;
359
360 /* if locking succeeded, pin bo: */
361 vma = msm_gem_get_vma_locked(obj, submit->aspace);
362 if (IS_ERR(vma)) {
363 ret = PTR_ERR(vma);
364 break;
365 }
366
367 ret = msm_gem_pin_vma_locked(obj, vma);
368 if (ret)
369 break;
370
371 submit->bos[i].flags |= BO_OBJ_PINNED | BO_VMA_PINNED;
372 submit->bos[i].vma = vma;
373
374 if (vma->iova == submit->bos[i].iova) {
375 submit->bos[i].flags |= BO_VALID;
376 } else {
377 submit->bos[i].iova = vma->iova;
378 /* iova changed, so address in cmdstream is not valid: */
379 submit->bos[i].flags &= ~BO_VALID;
380 submit->valid = false;
381 }
382 }
383
384 return ret;
385}
386
387static void submit_attach_object_fences(struct msm_gem_submit *submit)
388{
389 int i;
390
391 for (i = 0; i < submit->nr_bos; i++) {
392 struct drm_gem_object *obj = &submit->bos[i].obj->base;
393
394 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
395 dma_resv_add_fence(obj->resv, submit->user_fence,
396 DMA_RESV_USAGE_WRITE);
397 else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
398 dma_resv_add_fence(obj->resv, submit->user_fence,
399 DMA_RESV_USAGE_READ);
400 }
401}
402
403static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
404 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
405{
406 if (idx >= submit->nr_bos) {
407 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
408 idx, submit->nr_bos);
409 return -EINVAL;
410 }
411
412 if (obj)
413 *obj = submit->bos[idx].obj;
414 if (iova)
415 *iova = submit->bos[idx].iova;
416 if (valid)
417 *valid = !!(submit->bos[idx].flags & BO_VALID);
418
419 return 0;
420}
421
422/* process the reloc's and patch up the cmdstream as needed: */
423static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
424 uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs)
425{
426 uint32_t i, last_offset = 0;
427 uint32_t *ptr;
428 int ret = 0;
429
430 if (!nr_relocs)
431 return 0;
432
433 if (offset % 4) {
434 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
435 return -EINVAL;
436 }
437
438 /* For now, just map the entire thing. Eventually we probably
439 * to do it page-by-page, w/ kmap() if not vmap()d..
440 */
441 ptr = msm_gem_get_vaddr_locked(&obj->base);
442
443 if (IS_ERR(ptr)) {
444 ret = PTR_ERR(ptr);
445 DBG("failed to map: %d", ret);
446 return ret;
447 }
448
449 for (i = 0; i < nr_relocs; i++) {
450 struct drm_msm_gem_submit_reloc submit_reloc = relocs[i];
451 uint32_t off;
452 uint64_t iova;
453 bool valid;
454
455 if (submit_reloc.submit_offset % 4) {
456 DRM_ERROR("non-aligned reloc offset: %u\n",
457 submit_reloc.submit_offset);
458 ret = -EINVAL;
459 goto out;
460 }
461
462 /* offset in dwords: */
463 off = submit_reloc.submit_offset / 4;
464
465 if ((off >= (obj->base.size / 4)) ||
466 (off < last_offset)) {
467 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
468 ret = -EINVAL;
469 goto out;
470 }
471
472 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
473 if (ret)
474 goto out;
475
476 if (valid)
477 continue;
478
479 iova += submit_reloc.reloc_offset;
480
481 if (submit_reloc.shift < 0)
482 iova >>= -submit_reloc.shift;
483 else
484 iova <<= submit_reloc.shift;
485
486 ptr[off] = iova | submit_reloc.or;
487
488 last_offset = off;
489 }
490
491out:
492 msm_gem_put_vaddr_locked(&obj->base);
493
494 return ret;
495}
496
497/* Cleanup submit at end of ioctl. In the error case, this also drops
498 * references, unpins, and drops active refcnt. In the non-error case,
499 * this is done when the submit is retired.
500 */
501static void submit_cleanup(struct msm_gem_submit *submit, bool error)
502{
503 unsigned cleanup_flags = BO_LOCKED;
504 unsigned i;
505
506 if (error)
507 cleanup_flags |= BO_VMA_PINNED | BO_OBJ_PINNED;
508
509 for (i = 0; i < submit->nr_bos; i++) {
510 struct msm_gem_object *msm_obj = submit->bos[i].obj;
511 submit_cleanup_bo(submit, i, cleanup_flags);
512 if (error)
513 drm_gem_object_put(&msm_obj->base);
514 }
515}
516
517void msm_submit_retire(struct msm_gem_submit *submit)
518{
519 int i;
520
521 for (i = 0; i < submit->nr_bos; i++) {
522 struct drm_gem_object *obj = &submit->bos[i].obj->base;
523
524 drm_gem_object_put(obj);
525 }
526}
527
528struct msm_submit_post_dep {
529 struct drm_syncobj *syncobj;
530 uint64_t point;
531 struct dma_fence_chain *chain;
532};
533
534static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
535 struct drm_file *file,
536 uint64_t in_syncobjs_addr,
537 uint32_t nr_in_syncobjs,
538 size_t syncobj_stride,
539 struct msm_ringbuffer *ring)
540{
541 struct drm_syncobj **syncobjs = NULL;
542 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
543 int ret = 0;
544 uint32_t i, j;
545
546 syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
547 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
548 if (!syncobjs)
549 return ERR_PTR(-ENOMEM);
550
551 for (i = 0; i < nr_in_syncobjs; ++i) {
552 uint64_t address = in_syncobjs_addr + i * syncobj_stride;
553 struct dma_fence *fence;
554
555 if (copy_from_user(&syncobj_desc,
556 u64_to_user_ptr(address),
557 min(syncobj_stride, sizeof(syncobj_desc)))) {
558 ret = -EFAULT;
559 break;
560 }
561
562 if (syncobj_desc.point &&
563 !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
564 ret = -EOPNOTSUPP;
565 break;
566 }
567
568 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
569 ret = -EINVAL;
570 break;
571 }
572
573 ret = drm_syncobj_find_fence(file, syncobj_desc.handle,
574 syncobj_desc.point, 0, &fence);
575 if (ret)
576 break;
577
578 ret = drm_sched_job_add_dependency(&submit->base, fence);
579 if (ret)
580 break;
581
582 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
583 syncobjs[i] =
584 drm_syncobj_find(file, syncobj_desc.handle);
585 if (!syncobjs[i]) {
586 ret = -EINVAL;
587 break;
588 }
589 }
590 }
591
592 if (ret) {
593 for (j = 0; j <= i; ++j) {
594 if (syncobjs[j])
595 drm_syncobj_put(syncobjs[j]);
596 }
597 kfree(syncobjs);
598 return ERR_PTR(ret);
599 }
600 return syncobjs;
601}
602
603static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
604 uint32_t nr_syncobjs)
605{
606 uint32_t i;
607
608 for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
609 if (syncobjs[i])
610 drm_syncobj_replace_fence(syncobjs[i], NULL);
611 }
612}
613
614static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
615 struct drm_file *file,
616 uint64_t syncobjs_addr,
617 uint32_t nr_syncobjs,
618 size_t syncobj_stride)
619{
620 struct msm_submit_post_dep *post_deps;
621 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
622 int ret = 0;
623 uint32_t i, j;
624
625 post_deps = kmalloc_array(nr_syncobjs, sizeof(*post_deps),
626 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
627 if (!post_deps)
628 return ERR_PTR(-ENOMEM);
629
630 for (i = 0; i < nr_syncobjs; ++i) {
631 uint64_t address = syncobjs_addr + i * syncobj_stride;
632
633 if (copy_from_user(&syncobj_desc,
634 u64_to_user_ptr(address),
635 min(syncobj_stride, sizeof(syncobj_desc)))) {
636 ret = -EFAULT;
637 break;
638 }
639
640 post_deps[i].point = syncobj_desc.point;
641 post_deps[i].chain = NULL;
642
643 if (syncobj_desc.flags) {
644 ret = -EINVAL;
645 break;
646 }
647
648 if (syncobj_desc.point) {
649 if (!drm_core_check_feature(dev,
650 DRIVER_SYNCOBJ_TIMELINE)) {
651 ret = -EOPNOTSUPP;
652 break;
653 }
654
655 post_deps[i].chain = dma_fence_chain_alloc();
656 if (!post_deps[i].chain) {
657 ret = -ENOMEM;
658 break;
659 }
660 }
661
662 post_deps[i].syncobj =
663 drm_syncobj_find(file, syncobj_desc.handle);
664 if (!post_deps[i].syncobj) {
665 ret = -EINVAL;
666 break;
667 }
668 }
669
670 if (ret) {
671 for (j = 0; j <= i; ++j) {
672 dma_fence_chain_free(post_deps[j].chain);
673 if (post_deps[j].syncobj)
674 drm_syncobj_put(post_deps[j].syncobj);
675 }
676
677 kfree(post_deps);
678 return ERR_PTR(ret);
679 }
680
681 return post_deps;
682}
683
684static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
685 uint32_t count, struct dma_fence *fence)
686{
687 uint32_t i;
688
689 for (i = 0; post_deps && i < count; ++i) {
690 if (post_deps[i].chain) {
691 drm_syncobj_add_point(post_deps[i].syncobj,
692 post_deps[i].chain,
693 fence, post_deps[i].point);
694 post_deps[i].chain = NULL;
695 } else {
696 drm_syncobj_replace_fence(post_deps[i].syncobj,
697 fence);
698 }
699 }
700}
701
702int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
703 struct drm_file *file)
704{
705 struct msm_drm_private *priv = dev->dev_private;
706 struct drm_msm_gem_submit *args = data;
707 struct msm_file_private *ctx = file->driver_priv;
708 struct msm_gem_submit *submit;
709 struct msm_gpu *gpu = priv->gpu;
710 struct msm_gpu_submitqueue *queue;
711 struct msm_ringbuffer *ring;
712 struct msm_submit_post_dep *post_deps = NULL;
713 struct drm_syncobj **syncobjs_to_reset = NULL;
714 int out_fence_fd = -1;
715 bool has_ww_ticket = false;
716 unsigned i;
717 int ret;
718
719 if (!gpu)
720 return -ENXIO;
721
722 if (args->pad)
723 return -EINVAL;
724
725 if (unlikely(!ctx->aspace) && !capable(CAP_SYS_RAWIO)) {
726 DRM_ERROR_RATELIMITED("IOMMU support or CAP_SYS_RAWIO required!\n");
727 return -EPERM;
728 }
729
730 /* for now, we just have 3d pipe.. eventually this would need to
731 * be more clever to dispatch to appropriate gpu module:
732 */
733 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
734 return -EINVAL;
735
736 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
737 return -EINVAL;
738
739 if (args->flags & MSM_SUBMIT_SUDO) {
740 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
741 !capable(CAP_SYS_RAWIO))
742 return -EINVAL;
743 }
744
745 queue = msm_submitqueue_get(ctx, args->queueid);
746 if (!queue)
747 return -ENOENT;
748
749 ring = gpu->rb[queue->ring_nr];
750
751 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
752 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
753 if (out_fence_fd < 0) {
754 ret = out_fence_fd;
755 return ret;
756 }
757 }
758
759 submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds);
760 if (IS_ERR(submit))
761 return PTR_ERR(submit);
762
763 trace_msm_gpu_submit(pid_nr(submit->pid), ring->id, submit->ident,
764 args->nr_bos, args->nr_cmds);
765
766 ret = mutex_lock_interruptible(&queue->lock);
767 if (ret)
768 goto out_post_unlock;
769
770 if (args->flags & MSM_SUBMIT_SUDO)
771 submit->in_rb = true;
772
773 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
774 struct dma_fence *in_fence;
775
776 in_fence = sync_file_get_fence(args->fence_fd);
777
778 if (!in_fence) {
779 ret = -EINVAL;
780 goto out_unlock;
781 }
782
783 ret = drm_sched_job_add_dependency(&submit->base, in_fence);
784 if (ret)
785 goto out_unlock;
786 }
787
788 if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
789 syncobjs_to_reset = msm_parse_deps(submit, file,
790 args->in_syncobjs,
791 args->nr_in_syncobjs,
792 args->syncobj_stride, ring);
793 if (IS_ERR(syncobjs_to_reset)) {
794 ret = PTR_ERR(syncobjs_to_reset);
795 goto out_unlock;
796 }
797 }
798
799 if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
800 post_deps = msm_parse_post_deps(dev, file,
801 args->out_syncobjs,
802 args->nr_out_syncobjs,
803 args->syncobj_stride);
804 if (IS_ERR(post_deps)) {
805 ret = PTR_ERR(post_deps);
806 goto out_unlock;
807 }
808 }
809
810 ret = submit_lookup_objects(submit, args, file);
811 if (ret)
812 goto out;
813
814 ret = submit_lookup_cmds(submit, args, file);
815 if (ret)
816 goto out;
817
818 /* copy_*_user while holding a ww ticket upsets lockdep */
819 ww_acquire_init(&submit->ticket, &reservation_ww_class);
820 has_ww_ticket = true;
821 ret = submit_lock_objects(submit);
822 if (ret)
823 goto out;
824
825 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
826 if (ret)
827 goto out;
828
829 ret = submit_pin_objects(submit);
830 if (ret)
831 goto out;
832
833 for (i = 0; i < args->nr_cmds; i++) {
834 struct msm_gem_object *msm_obj;
835 uint64_t iova;
836
837 ret = submit_bo(submit, submit->cmd[i].idx,
838 &msm_obj, &iova, NULL);
839 if (ret)
840 goto out;
841
842 if (!submit->cmd[i].size ||
843 ((submit->cmd[i].size + submit->cmd[i].offset) >
844 msm_obj->base.size / 4)) {
845 DRM_ERROR("invalid cmdstream size: %u\n", submit->cmd[i].size * 4);
846 ret = -EINVAL;
847 goto out;
848 }
849
850 submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
851
852 if (submit->valid)
853 continue;
854
855 ret = submit_reloc(submit, msm_obj, submit->cmd[i].offset * 4,
856 submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
857 if (ret)
858 goto out;
859 }
860
861 submit->nr_cmds = i;
862
863 mutex_lock(&queue->idr_lock);
864
865 /*
866 * If using userspace provided seqno fence, validate that the id
867 * is available before arming sched job. Since access to fence_idr
868 * is serialized on the queue lock, the slot should be still avail
869 * after the job is armed
870 */
871 if ((args->flags & MSM_SUBMIT_FENCE_SN_IN) &&
872 idr_find(&queue->fence_idr, args->fence)) {
873 mutex_unlock(&queue->idr_lock);
874 ret = -EINVAL;
875 goto out;
876 }
877
878 drm_sched_job_arm(&submit->base);
879
880 submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
881
882 if (args->flags & MSM_SUBMIT_FENCE_SN_IN) {
883 /*
884 * Userspace has assigned the seqno fence that it wants
885 * us to use. It is an error to pick a fence sequence
886 * number that is not available.
887 */
888 submit->fence_id = args->fence;
889 ret = idr_alloc_u32(&queue->fence_idr, submit->user_fence,
890 &submit->fence_id, submit->fence_id,
891 GFP_KERNEL);
892 /*
893 * We've already validated that the fence_id slot is valid,
894 * so if idr_alloc_u32 failed, it is a kernel bug
895 */
896 WARN_ON(ret);
897 } else {
898 /*
899 * Allocate an id which can be used by WAIT_FENCE ioctl to map
900 * back to the underlying fence.
901 */
902 submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
903 submit->user_fence, 1,
904 INT_MAX, GFP_KERNEL);
905 }
906
907 mutex_unlock(&queue->idr_lock);
908
909 if (submit->fence_id < 0) {
910 ret = submit->fence_id;
911 submit->fence_id = 0;
912 }
913
914 if (ret == 0 && args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
915 struct sync_file *sync_file = sync_file_create(submit->user_fence);
916 if (!sync_file) {
917 ret = -ENOMEM;
918 } else {
919 fd_install(out_fence_fd, sync_file->file);
920 args->fence_fd = out_fence_fd;
921 }
922 }
923
924 submit_attach_object_fences(submit);
925
926 /* The scheduler owns a ref now: */
927 msm_gem_submit_get(submit);
928
929 drm_sched_entity_push_job(&submit->base);
930
931 args->fence = submit->fence_id;
932 queue->last_fence = submit->fence_id;
933
934 msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
935 msm_process_post_deps(post_deps, args->nr_out_syncobjs,
936 submit->user_fence);
937
938
939out:
940 submit_cleanup(submit, !!ret);
941 if (has_ww_ticket)
942 ww_acquire_fini(&submit->ticket);
943out_unlock:
944 if (ret && (out_fence_fd >= 0))
945 put_unused_fd(out_fence_fd);
946 mutex_unlock(&queue->lock);
947out_post_unlock:
948 msm_gem_submit_put(submit);
949 if (!IS_ERR_OR_NULL(post_deps)) {
950 for (i = 0; i < args->nr_out_syncobjs; ++i) {
951 kfree(post_deps[i].chain);
952 drm_syncobj_put(post_deps[i].syncobj);
953 }
954 kfree(post_deps);
955 }
956
957 if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
958 for (i = 0; i < args->nr_in_syncobjs; ++i) {
959 if (syncobjs_to_reset[i])
960 drm_syncobj_put(syncobjs_to_reset[i]);
961 }
962 kfree(syncobjs_to_reset);
963 }
964
965 return ret;
966}