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_file.h>
12
13#include "msm_drv.h"
14#include "msm_gpu.h"
15#include "msm_gem.h"
16#include "msm_gpu_trace.h"
17
18/*
19 * Cmdstream submission:
20 */
21
22/* make sure these don't conflict w/ MSM_SUBMIT_BO_x */
23#define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */
24#define BO_LOCKED 0x4000
25#define BO_PINNED 0x2000
26
27static struct msm_gem_submit *submit_create(struct drm_device *dev,
28 struct msm_gpu *gpu, struct msm_gem_address_space *aspace,
29 struct msm_gpu_submitqueue *queue, uint32_t nr_bos,
30 uint32_t nr_cmds)
31{
32 struct msm_gem_submit *submit;
33 uint64_t sz = struct_size(submit, bos, nr_bos) +
34 ((u64)nr_cmds * sizeof(submit->cmd[0]));
35
36 if (sz > SIZE_MAX)
37 return NULL;
38
39 submit = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
40 if (!submit)
41 return NULL;
42
43 submit->dev = dev;
44 submit->aspace = aspace;
45 submit->gpu = gpu;
46 submit->fence = NULL;
47 submit->cmd = (void *)&submit->bos[nr_bos];
48 submit->queue = queue;
49 submit->ring = gpu->rb[queue->prio];
50
51 /* initially, until copy_from_user() and bo lookup succeeds: */
52 submit->nr_bos = 0;
53 submit->nr_cmds = 0;
54
55 INIT_LIST_HEAD(&submit->node);
56 INIT_LIST_HEAD(&submit->bo_list);
57 ww_acquire_init(&submit->ticket, &reservation_ww_class);
58
59 return submit;
60}
61
62void msm_gem_submit_free(struct msm_gem_submit *submit)
63{
64 dma_fence_put(submit->fence);
65 list_del(&submit->node);
66 put_pid(submit->pid);
67 msm_submitqueue_put(submit->queue);
68
69 kfree(submit);
70}
71
72static int submit_lookup_objects(struct msm_gem_submit *submit,
73 struct drm_msm_gem_submit *args, struct drm_file *file)
74{
75 unsigned i;
76 int ret = 0;
77
78 for (i = 0; i < args->nr_bos; i++) {
79 struct drm_msm_gem_submit_bo submit_bo;
80 void __user *userptr =
81 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
82
83 /* make sure we don't have garbage flags, in case we hit
84 * error path before flags is initialized:
85 */
86 submit->bos[i].flags = 0;
87
88 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
89 ret = -EFAULT;
90 i = 0;
91 goto out;
92 }
93
94/* at least one of READ and/or WRITE flags should be set: */
95#define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
96
97 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
98 !(submit_bo.flags & MANDATORY_FLAGS)) {
99 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
100 ret = -EINVAL;
101 i = 0;
102 goto out;
103 }
104
105 submit->bos[i].handle = submit_bo.handle;
106 submit->bos[i].flags = submit_bo.flags;
107 /* in validate_objects() we figure out if this is true: */
108 submit->bos[i].iova = submit_bo.presumed;
109 }
110
111 spin_lock(&file->table_lock);
112
113 for (i = 0; i < args->nr_bos; i++) {
114 struct drm_gem_object *obj;
115 struct msm_gem_object *msm_obj;
116
117 /* normally use drm_gem_object_lookup(), but for bulk lookup
118 * all under single table_lock just hit object_idr directly:
119 */
120 obj = idr_find(&file->object_idr, submit->bos[i].handle);
121 if (!obj) {
122 DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
123 ret = -EINVAL;
124 goto out_unlock;
125 }
126
127 msm_obj = to_msm_bo(obj);
128
129 if (!list_empty(&msm_obj->submit_entry)) {
130 DRM_ERROR("handle %u at index %u already on submit list\n",
131 submit->bos[i].handle, i);
132 ret = -EINVAL;
133 goto out_unlock;
134 }
135
136 drm_gem_object_get(obj);
137
138 submit->bos[i].obj = msm_obj;
139
140 list_add_tail(&msm_obj->submit_entry, &submit->bo_list);
141 }
142
143out_unlock:
144 spin_unlock(&file->table_lock);
145
146out:
147 submit->nr_bos = i;
148
149 return ret;
150}
151
152static void submit_unlock_unpin_bo(struct msm_gem_submit *submit,
153 int i, bool backoff)
154{
155 struct msm_gem_object *msm_obj = submit->bos[i].obj;
156
157 if (submit->bos[i].flags & BO_PINNED)
158 msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
159
160 if (submit->bos[i].flags & BO_LOCKED)
161 ww_mutex_unlock(&msm_obj->base.resv->lock);
162
163 if (backoff && !(submit->bos[i].flags & BO_VALID))
164 submit->bos[i].iova = 0;
165
166 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED);
167}
168
169/* This is where we make sure all the bo's are reserved and pin'd: */
170static int submit_lock_objects(struct msm_gem_submit *submit)
171{
172 int contended, slow_locked = -1, i, ret = 0;
173
174retry:
175 for (i = 0; i < submit->nr_bos; i++) {
176 struct msm_gem_object *msm_obj = submit->bos[i].obj;
177
178 if (slow_locked == i)
179 slow_locked = -1;
180
181 contended = i;
182
183 if (!(submit->bos[i].flags & BO_LOCKED)) {
184 ret = ww_mutex_lock_interruptible(&msm_obj->base.resv->lock,
185 &submit->ticket);
186 if (ret)
187 goto fail;
188 submit->bos[i].flags |= BO_LOCKED;
189 }
190 }
191
192 ww_acquire_done(&submit->ticket);
193
194 return 0;
195
196fail:
197 for (; i >= 0; i--)
198 submit_unlock_unpin_bo(submit, i, true);
199
200 if (slow_locked > 0)
201 submit_unlock_unpin_bo(submit, slow_locked, true);
202
203 if (ret == -EDEADLK) {
204 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
205 /* we lost out in a seqno race, lock and retry.. */
206 ret = ww_mutex_lock_slow_interruptible(&msm_obj->base.resv->lock,
207 &submit->ticket);
208 if (!ret) {
209 submit->bos[contended].flags |= BO_LOCKED;
210 slow_locked = contended;
211 goto retry;
212 }
213 }
214
215 return ret;
216}
217
218static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
219{
220 int i, ret = 0;
221
222 for (i = 0; i < submit->nr_bos; i++) {
223 struct msm_gem_object *msm_obj = submit->bos[i].obj;
224 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
225
226 if (!write) {
227 /* NOTE: _reserve_shared() must happen before
228 * _add_shared_fence(), which makes this a slightly
229 * strange place to call it. OTOH this is a
230 * convenient can-fail point to hook it in.
231 */
232 ret = dma_resv_reserve_shared(msm_obj->base.resv,
233 1);
234 if (ret)
235 return ret;
236 }
237
238 if (no_implicit)
239 continue;
240
241 ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
242 write);
243 if (ret)
244 break;
245 }
246
247 return ret;
248}
249
250static int submit_pin_objects(struct msm_gem_submit *submit)
251{
252 int i, ret = 0;
253
254 submit->valid = true;
255
256 for (i = 0; i < submit->nr_bos; i++) {
257 struct msm_gem_object *msm_obj = submit->bos[i].obj;
258 uint64_t iova;
259
260 /* if locking succeeded, pin bo: */
261 ret = msm_gem_get_and_pin_iova(&msm_obj->base,
262 submit->aspace, &iova);
263
264 if (ret)
265 break;
266
267 submit->bos[i].flags |= BO_PINNED;
268
269 if (iova == submit->bos[i].iova) {
270 submit->bos[i].flags |= BO_VALID;
271 } else {
272 submit->bos[i].iova = iova;
273 /* iova changed, so address in cmdstream is not valid: */
274 submit->bos[i].flags &= ~BO_VALID;
275 submit->valid = false;
276 }
277 }
278
279 return ret;
280}
281
282static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
283 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
284{
285 if (idx >= submit->nr_bos) {
286 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
287 idx, submit->nr_bos);
288 return -EINVAL;
289 }
290
291 if (obj)
292 *obj = submit->bos[idx].obj;
293 if (iova)
294 *iova = submit->bos[idx].iova;
295 if (valid)
296 *valid = !!(submit->bos[idx].flags & BO_VALID);
297
298 return 0;
299}
300
301/* process the reloc's and patch up the cmdstream as needed: */
302static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
303 uint32_t offset, uint32_t nr_relocs, uint64_t relocs)
304{
305 uint32_t i, last_offset = 0;
306 uint32_t *ptr;
307 int ret = 0;
308
309 if (!nr_relocs)
310 return 0;
311
312 if (offset % 4) {
313 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
314 return -EINVAL;
315 }
316
317 /* For now, just map the entire thing. Eventually we probably
318 * to do it page-by-page, w/ kmap() if not vmap()d..
319 */
320 ptr = msm_gem_get_vaddr(&obj->base);
321
322 if (IS_ERR(ptr)) {
323 ret = PTR_ERR(ptr);
324 DBG("failed to map: %d", ret);
325 return ret;
326 }
327
328 for (i = 0; i < nr_relocs; i++) {
329 struct drm_msm_gem_submit_reloc submit_reloc;
330 void __user *userptr =
331 u64_to_user_ptr(relocs + (i * sizeof(submit_reloc)));
332 uint32_t off;
333 uint64_t iova;
334 bool valid;
335
336 if (copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc))) {
337 ret = -EFAULT;
338 goto out;
339 }
340
341 if (submit_reloc.submit_offset % 4) {
342 DRM_ERROR("non-aligned reloc offset: %u\n",
343 submit_reloc.submit_offset);
344 ret = -EINVAL;
345 goto out;
346 }
347
348 /* offset in dwords: */
349 off = submit_reloc.submit_offset / 4;
350
351 if ((off >= (obj->base.size / 4)) ||
352 (off < last_offset)) {
353 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
354 ret = -EINVAL;
355 goto out;
356 }
357
358 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
359 if (ret)
360 goto out;
361
362 if (valid)
363 continue;
364
365 iova += submit_reloc.reloc_offset;
366
367 if (submit_reloc.shift < 0)
368 iova >>= -submit_reloc.shift;
369 else
370 iova <<= submit_reloc.shift;
371
372 ptr[off] = iova | submit_reloc.or;
373
374 last_offset = off;
375 }
376
377out:
378 msm_gem_put_vaddr(&obj->base);
379
380 return ret;
381}
382
383static void submit_cleanup(struct msm_gem_submit *submit)
384{
385 unsigned i;
386
387 for (i = 0; i < submit->nr_bos; i++) {
388 struct msm_gem_object *msm_obj = submit->bos[i].obj;
389 submit_unlock_unpin_bo(submit, i, false);
390 list_del_init(&msm_obj->submit_entry);
391 drm_gem_object_put(&msm_obj->base);
392 }
393
394 ww_acquire_fini(&submit->ticket);
395}
396
397int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
398 struct drm_file *file)
399{
400 static atomic_t ident = ATOMIC_INIT(0);
401 struct msm_drm_private *priv = dev->dev_private;
402 struct drm_msm_gem_submit *args = data;
403 struct msm_file_private *ctx = file->driver_priv;
404 struct msm_gem_submit *submit;
405 struct msm_gpu *gpu = priv->gpu;
406 struct sync_file *sync_file = NULL;
407 struct msm_gpu_submitqueue *queue;
408 struct msm_ringbuffer *ring;
409 int out_fence_fd = -1;
410 struct pid *pid = get_pid(task_pid(current));
411 unsigned i;
412 int ret, submitid;
413 if (!gpu)
414 return -ENXIO;
415
416 /* for now, we just have 3d pipe.. eventually this would need to
417 * be more clever to dispatch to appropriate gpu module:
418 */
419 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
420 return -EINVAL;
421
422 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
423 return -EINVAL;
424
425 if (args->flags & MSM_SUBMIT_SUDO) {
426 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
427 !capable(CAP_SYS_RAWIO))
428 return -EINVAL;
429 }
430
431 queue = msm_submitqueue_get(ctx, args->queueid);
432 if (!queue)
433 return -ENOENT;
434
435 /* Get a unique identifier for the submission for logging purposes */
436 submitid = atomic_inc_return(&ident) - 1;
437
438 ring = gpu->rb[queue->prio];
439 trace_msm_gpu_submit(pid_nr(pid), ring->id, submitid,
440 args->nr_bos, args->nr_cmds);
441
442 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
443 struct dma_fence *in_fence;
444
445 in_fence = sync_file_get_fence(args->fence_fd);
446
447 if (!in_fence)
448 return -EINVAL;
449
450 /*
451 * Wait if the fence is from a foreign context, or if the fence
452 * array contains any fence from a foreign context.
453 */
454 ret = 0;
455 if (!dma_fence_match_context(in_fence, ring->fctx->context))
456 ret = dma_fence_wait(in_fence, true);
457
458 dma_fence_put(in_fence);
459 if (ret)
460 return ret;
461 }
462
463 ret = mutex_lock_interruptible(&dev->struct_mutex);
464 if (ret)
465 return ret;
466
467 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
468 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
469 if (out_fence_fd < 0) {
470 ret = out_fence_fd;
471 goto out_unlock;
472 }
473 }
474
475 submit = submit_create(dev, gpu, ctx->aspace, queue, args->nr_bos,
476 args->nr_cmds);
477 if (!submit) {
478 ret = -ENOMEM;
479 goto out_unlock;
480 }
481
482 submit->pid = pid;
483 submit->ident = submitid;
484
485 if (args->flags & MSM_SUBMIT_SUDO)
486 submit->in_rb = true;
487
488 ret = submit_lookup_objects(submit, args, file);
489 if (ret)
490 goto out;
491
492 ret = submit_lock_objects(submit);
493 if (ret)
494 goto out;
495
496 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
497 if (ret)
498 goto out;
499
500 ret = submit_pin_objects(submit);
501 if (ret)
502 goto out;
503
504 for (i = 0; i < args->nr_cmds; i++) {
505 struct drm_msm_gem_submit_cmd submit_cmd;
506 void __user *userptr =
507 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
508 struct msm_gem_object *msm_obj;
509 uint64_t iova;
510
511 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
512 if (ret) {
513 ret = -EFAULT;
514 goto out;
515 }
516
517 /* validate input from userspace: */
518 switch (submit_cmd.type) {
519 case MSM_SUBMIT_CMD_BUF:
520 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
521 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
522 break;
523 default:
524 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
525 ret = -EINVAL;
526 goto out;
527 }
528
529 ret = submit_bo(submit, submit_cmd.submit_idx,
530 &msm_obj, &iova, NULL);
531 if (ret)
532 goto out;
533
534 if (submit_cmd.size % 4) {
535 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
536 submit_cmd.size);
537 ret = -EINVAL;
538 goto out;
539 }
540
541 if (!submit_cmd.size ||
542 ((submit_cmd.size + submit_cmd.submit_offset) >
543 msm_obj->base.size)) {
544 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size);
545 ret = -EINVAL;
546 goto out;
547 }
548
549 submit->cmd[i].type = submit_cmd.type;
550 submit->cmd[i].size = submit_cmd.size / 4;
551 submit->cmd[i].iova = iova + submit_cmd.submit_offset;
552 submit->cmd[i].idx = submit_cmd.submit_idx;
553
554 if (submit->valid)
555 continue;
556
557 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset,
558 submit_cmd.nr_relocs, submit_cmd.relocs);
559 if (ret)
560 goto out;
561 }
562
563 submit->nr_cmds = i;
564
565 submit->fence = msm_fence_alloc(ring->fctx);
566 if (IS_ERR(submit->fence)) {
567 ret = PTR_ERR(submit->fence);
568 submit->fence = NULL;
569 goto out;
570 }
571
572 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
573 sync_file = sync_file_create(submit->fence);
574 if (!sync_file) {
575 ret = -ENOMEM;
576 goto out;
577 }
578 }
579
580 msm_gpu_submit(gpu, submit, ctx);
581
582 args->fence = submit->fence->seqno;
583
584 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
585 fd_install(out_fence_fd, sync_file->file);
586 args->fence_fd = out_fence_fd;
587 }
588
589out:
590 submit_cleanup(submit);
591 if (ret)
592 msm_gem_submit_free(submit);
593out_unlock:
594 if (ret && (out_fence_fd >= 0))
595 put_unused_fd(out_fence_fd);
596 mutex_unlock(&dev->struct_mutex);
597 return ret;
598}