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/*
2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published by
7 * the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 *
14 * You should have received a copy of the GNU General Public License along with
15 * this program. If not, see <http://www.gnu.org/licenses/>.
16 */
17
18#include <linux/sync_file.h>
19
20#include "msm_drv.h"
21#include "msm_gpu.h"
22#include "msm_gem.h"
23
24/*
25 * Cmdstream submission:
26 */
27
28/* make sure these don't conflict w/ MSM_SUBMIT_BO_x */
29#define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */
30#define BO_LOCKED 0x4000
31#define BO_PINNED 0x2000
32
33static struct msm_gem_submit *submit_create(struct drm_device *dev,
34 struct msm_gpu *gpu, struct msm_gpu_submitqueue *queue,
35 uint32_t nr_bos, uint32_t nr_cmds)
36{
37 struct msm_gem_submit *submit;
38 uint64_t sz = sizeof(*submit) + ((u64)nr_bos * sizeof(submit->bos[0])) +
39 ((u64)nr_cmds * sizeof(submit->cmd[0]));
40
41 if (sz > SIZE_MAX)
42 return NULL;
43
44 submit = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
45 if (!submit)
46 return NULL;
47
48 submit->dev = dev;
49 submit->gpu = gpu;
50 submit->fence = NULL;
51 submit->pid = get_pid(task_pid(current));
52 submit->cmd = (void *)&submit->bos[nr_bos];
53 submit->queue = queue;
54 submit->ring = gpu->rb[queue->prio];
55
56 /* initially, until copy_from_user() and bo lookup succeeds: */
57 submit->nr_bos = 0;
58 submit->nr_cmds = 0;
59
60 INIT_LIST_HEAD(&submit->node);
61 INIT_LIST_HEAD(&submit->bo_list);
62 ww_acquire_init(&submit->ticket, &reservation_ww_class);
63
64 return submit;
65}
66
67void msm_gem_submit_free(struct msm_gem_submit *submit)
68{
69 dma_fence_put(submit->fence);
70 list_del(&submit->node);
71 put_pid(submit->pid);
72 msm_submitqueue_put(submit->queue);
73
74 kfree(submit);
75}
76
77static inline unsigned long __must_check
78copy_from_user_inatomic(void *to, const void __user *from, unsigned long n)
79{
80 if (access_ok(VERIFY_READ, from, n))
81 return __copy_from_user_inatomic(to, from, n);
82 return -EFAULT;
83}
84
85static int submit_lookup_objects(struct msm_gem_submit *submit,
86 struct drm_msm_gem_submit *args, struct drm_file *file)
87{
88 unsigned i;
89 int ret = 0;
90
91 spin_lock(&file->table_lock);
92 pagefault_disable();
93
94 for (i = 0; i < args->nr_bos; i++) {
95 struct drm_msm_gem_submit_bo submit_bo;
96 struct drm_gem_object *obj;
97 struct msm_gem_object *msm_obj;
98 void __user *userptr =
99 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
100
101 /* make sure we don't have garbage flags, in case we hit
102 * error path before flags is initialized:
103 */
104 submit->bos[i].flags = 0;
105
106 if (copy_from_user_inatomic(&submit_bo, userptr, sizeof(submit_bo))) {
107 pagefault_enable();
108 spin_unlock(&file->table_lock);
109 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
110 ret = -EFAULT;
111 goto out;
112 }
113 spin_lock(&file->table_lock);
114 pagefault_disable();
115 }
116
117 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
118 !(submit_bo.flags & MSM_SUBMIT_BO_FLAGS)) {
119 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
120 ret = -EINVAL;
121 goto out_unlock;
122 }
123
124 submit->bos[i].flags = submit_bo.flags;
125 /* in validate_objects() we figure out if this is true: */
126 submit->bos[i].iova = submit_bo.presumed;
127
128 /* normally use drm_gem_object_lookup(), but for bulk lookup
129 * all under single table_lock just hit object_idr directly:
130 */
131 obj = idr_find(&file->object_idr, submit_bo.handle);
132 if (!obj) {
133 DRM_ERROR("invalid handle %u at index %u\n", submit_bo.handle, i);
134 ret = -EINVAL;
135 goto out_unlock;
136 }
137
138 msm_obj = to_msm_bo(obj);
139
140 if (!list_empty(&msm_obj->submit_entry)) {
141 DRM_ERROR("handle %u at index %u already on submit list\n",
142 submit_bo.handle, i);
143 ret = -EINVAL;
144 goto out_unlock;
145 }
146
147 drm_gem_object_reference(obj);
148
149 submit->bos[i].obj = msm_obj;
150
151 list_add_tail(&msm_obj->submit_entry, &submit->bo_list);
152 }
153
154out_unlock:
155 pagefault_enable();
156 spin_unlock(&file->table_lock);
157
158out:
159 submit->nr_bos = i;
160
161 return ret;
162}
163
164static void submit_unlock_unpin_bo(struct msm_gem_submit *submit,
165 int i, bool backoff)
166{
167 struct msm_gem_object *msm_obj = submit->bos[i].obj;
168
169 if (submit->bos[i].flags & BO_PINNED)
170 msm_gem_put_iova(&msm_obj->base, submit->gpu->aspace);
171
172 if (submit->bos[i].flags & BO_LOCKED)
173 ww_mutex_unlock(&msm_obj->resv->lock);
174
175 if (backoff && !(submit->bos[i].flags & BO_VALID))
176 submit->bos[i].iova = 0;
177
178 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED);
179}
180
181/* This is where we make sure all the bo's are reserved and pin'd: */
182static int submit_lock_objects(struct msm_gem_submit *submit)
183{
184 int contended, slow_locked = -1, i, ret = 0;
185
186retry:
187 for (i = 0; i < submit->nr_bos; i++) {
188 struct msm_gem_object *msm_obj = submit->bos[i].obj;
189
190 if (slow_locked == i)
191 slow_locked = -1;
192
193 contended = i;
194
195 if (!(submit->bos[i].flags & BO_LOCKED)) {
196 ret = ww_mutex_lock_interruptible(&msm_obj->resv->lock,
197 &submit->ticket);
198 if (ret)
199 goto fail;
200 submit->bos[i].flags |= BO_LOCKED;
201 }
202 }
203
204 ww_acquire_done(&submit->ticket);
205
206 return 0;
207
208fail:
209 for (; i >= 0; i--)
210 submit_unlock_unpin_bo(submit, i, true);
211
212 if (slow_locked > 0)
213 submit_unlock_unpin_bo(submit, slow_locked, true);
214
215 if (ret == -EDEADLK) {
216 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
217 /* we lost out in a seqno race, lock and retry.. */
218 ret = ww_mutex_lock_slow_interruptible(&msm_obj->resv->lock,
219 &submit->ticket);
220 if (!ret) {
221 submit->bos[contended].flags |= BO_LOCKED;
222 slow_locked = contended;
223 goto retry;
224 }
225 }
226
227 return ret;
228}
229
230static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
231{
232 int i, ret = 0;
233
234 for (i = 0; i < submit->nr_bos; i++) {
235 struct msm_gem_object *msm_obj = submit->bos[i].obj;
236 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
237
238 if (!write) {
239 /* NOTE: _reserve_shared() must happen before
240 * _add_shared_fence(), which makes this a slightly
241 * strange place to call it. OTOH this is a
242 * convenient can-fail point to hook it in.
243 */
244 ret = reservation_object_reserve_shared(msm_obj->resv);
245 if (ret)
246 return ret;
247 }
248
249 if (no_implicit)
250 continue;
251
252 ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
253 write);
254 if (ret)
255 break;
256 }
257
258 return ret;
259}
260
261static int submit_pin_objects(struct msm_gem_submit *submit)
262{
263 int i, ret = 0;
264
265 submit->valid = true;
266
267 for (i = 0; i < submit->nr_bos; i++) {
268 struct msm_gem_object *msm_obj = submit->bos[i].obj;
269 uint64_t iova;
270
271 /* if locking succeeded, pin bo: */
272 ret = msm_gem_get_iova(&msm_obj->base,
273 submit->gpu->aspace, &iova);
274
275 if (ret)
276 break;
277
278 submit->bos[i].flags |= BO_PINNED;
279
280 if (iova == submit->bos[i].iova) {
281 submit->bos[i].flags |= BO_VALID;
282 } else {
283 submit->bos[i].iova = iova;
284 /* iova changed, so address in cmdstream is not valid: */
285 submit->bos[i].flags &= ~BO_VALID;
286 submit->valid = false;
287 }
288 }
289
290 return ret;
291}
292
293static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
294 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
295{
296 if (idx >= submit->nr_bos) {
297 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
298 idx, submit->nr_bos);
299 return -EINVAL;
300 }
301
302 if (obj)
303 *obj = submit->bos[idx].obj;
304 if (iova)
305 *iova = submit->bos[idx].iova;
306 if (valid)
307 *valid = !!(submit->bos[idx].flags & BO_VALID);
308
309 return 0;
310}
311
312/* process the reloc's and patch up the cmdstream as needed: */
313static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
314 uint32_t offset, uint32_t nr_relocs, uint64_t relocs)
315{
316 uint32_t i, last_offset = 0;
317 uint32_t *ptr;
318 int ret = 0;
319
320 if (offset % 4) {
321 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
322 return -EINVAL;
323 }
324
325 /* For now, just map the entire thing. Eventually we probably
326 * to do it page-by-page, w/ kmap() if not vmap()d..
327 */
328 ptr = msm_gem_get_vaddr(&obj->base);
329
330 if (IS_ERR(ptr)) {
331 ret = PTR_ERR(ptr);
332 DBG("failed to map: %d", ret);
333 return ret;
334 }
335
336 for (i = 0; i < nr_relocs; i++) {
337 struct drm_msm_gem_submit_reloc submit_reloc;
338 void __user *userptr =
339 u64_to_user_ptr(relocs + (i * sizeof(submit_reloc)));
340 uint32_t off;
341 uint64_t iova;
342 bool valid;
343
344 if (copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc))) {
345 ret = -EFAULT;
346 goto out;
347 }
348
349 if (submit_reloc.submit_offset % 4) {
350 DRM_ERROR("non-aligned reloc offset: %u\n",
351 submit_reloc.submit_offset);
352 ret = -EINVAL;
353 goto out;
354 }
355
356 /* offset in dwords: */
357 off = submit_reloc.submit_offset / 4;
358
359 if ((off >= (obj->base.size / 4)) ||
360 (off < last_offset)) {
361 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
362 ret = -EINVAL;
363 goto out;
364 }
365
366 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
367 if (ret)
368 goto out;
369
370 if (valid)
371 continue;
372
373 iova += submit_reloc.reloc_offset;
374
375 if (submit_reloc.shift < 0)
376 iova >>= -submit_reloc.shift;
377 else
378 iova <<= submit_reloc.shift;
379
380 ptr[off] = iova | submit_reloc.or;
381
382 last_offset = off;
383 }
384
385out:
386 msm_gem_put_vaddr(&obj->base);
387
388 return ret;
389}
390
391static void submit_cleanup(struct msm_gem_submit *submit)
392{
393 unsigned i;
394
395 for (i = 0; i < submit->nr_bos; i++) {
396 struct msm_gem_object *msm_obj = submit->bos[i].obj;
397 submit_unlock_unpin_bo(submit, i, false);
398 list_del_init(&msm_obj->submit_entry);
399 drm_gem_object_unreference(&msm_obj->base);
400 }
401
402 ww_acquire_fini(&submit->ticket);
403}
404
405int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
406 struct drm_file *file)
407{
408 struct msm_drm_private *priv = dev->dev_private;
409 struct drm_msm_gem_submit *args = data;
410 struct msm_file_private *ctx = file->driver_priv;
411 struct msm_gem_submit *submit;
412 struct msm_gpu *gpu = priv->gpu;
413 struct dma_fence *in_fence = NULL;
414 struct sync_file *sync_file = NULL;
415 struct msm_gpu_submitqueue *queue;
416 struct msm_ringbuffer *ring;
417 int out_fence_fd = -1;
418 unsigned i;
419 int ret;
420
421 if (!gpu)
422 return -ENXIO;
423
424 /* for now, we just have 3d pipe.. eventually this would need to
425 * be more clever to dispatch to appropriate gpu module:
426 */
427 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
428 return -EINVAL;
429
430 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
431 return -EINVAL;
432
433 if (args->flags & MSM_SUBMIT_SUDO) {
434 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
435 !capable(CAP_SYS_RAWIO))
436 return -EINVAL;
437 }
438
439 queue = msm_submitqueue_get(ctx, args->queueid);
440 if (!queue)
441 return -ENOENT;
442
443 ring = gpu->rb[queue->prio];
444
445 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
446 in_fence = sync_file_get_fence(args->fence_fd);
447
448 if (!in_fence)
449 return -EINVAL;
450
451 /*
452 * Wait if the fence is from a foreign context, or if the fence
453 * array contains any fence from a foreign context.
454 */
455 if (!dma_fence_match_context(in_fence, ring->fctx->context)) {
456 ret = dma_fence_wait(in_fence, true);
457 if (ret)
458 return ret;
459 }
460 }
461
462 ret = mutex_lock_interruptible(&dev->struct_mutex);
463 if (ret)
464 return ret;
465
466 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
467 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
468 if (out_fence_fd < 0) {
469 ret = out_fence_fd;
470 goto out_unlock;
471 }
472 }
473
474 submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds);
475 if (!submit) {
476 ret = -ENOMEM;
477 goto out_unlock;
478 }
479
480 if (args->flags & MSM_SUBMIT_SUDO)
481 submit->in_rb = true;
482
483 ret = submit_lookup_objects(submit, args, file);
484 if (ret)
485 goto out;
486
487 ret = submit_lock_objects(submit);
488 if (ret)
489 goto out;
490
491 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
492 if (ret)
493 goto out;
494
495 ret = submit_pin_objects(submit);
496 if (ret)
497 goto out;
498
499 for (i = 0; i < args->nr_cmds; i++) {
500 struct drm_msm_gem_submit_cmd submit_cmd;
501 void __user *userptr =
502 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
503 struct msm_gem_object *msm_obj;
504 uint64_t iova;
505
506 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
507 if (ret) {
508 ret = -EFAULT;
509 goto out;
510 }
511
512 /* validate input from userspace: */
513 switch (submit_cmd.type) {
514 case MSM_SUBMIT_CMD_BUF:
515 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
516 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
517 break;
518 default:
519 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
520 ret = -EINVAL;
521 goto out;
522 }
523
524 ret = submit_bo(submit, submit_cmd.submit_idx,
525 &msm_obj, &iova, NULL);
526 if (ret)
527 goto out;
528
529 if (submit_cmd.size % 4) {
530 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
531 submit_cmd.size);
532 ret = -EINVAL;
533 goto out;
534 }
535
536 if (!submit_cmd.size ||
537 ((submit_cmd.size + submit_cmd.submit_offset) >
538 msm_obj->base.size)) {
539 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size);
540 ret = -EINVAL;
541 goto out;
542 }
543
544 submit->cmd[i].type = submit_cmd.type;
545 submit->cmd[i].size = submit_cmd.size / 4;
546 submit->cmd[i].iova = iova + submit_cmd.submit_offset;
547 submit->cmd[i].idx = submit_cmd.submit_idx;
548
549 if (submit->valid)
550 continue;
551
552 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset,
553 submit_cmd.nr_relocs, submit_cmd.relocs);
554 if (ret)
555 goto out;
556 }
557
558 submit->nr_cmds = i;
559
560 submit->fence = msm_fence_alloc(ring->fctx);
561 if (IS_ERR(submit->fence)) {
562 ret = PTR_ERR(submit->fence);
563 submit->fence = NULL;
564 goto out;
565 }
566
567 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
568 sync_file = sync_file_create(submit->fence);
569 if (!sync_file) {
570 ret = -ENOMEM;
571 goto out;
572 }
573 }
574
575 msm_gpu_submit(gpu, submit, ctx);
576
577 args->fence = submit->fence->seqno;
578
579 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
580 fd_install(out_fence_fd, sync_file->file);
581 args->fence_fd = out_fence_fd;
582 }
583
584out:
585 if (in_fence)
586 dma_fence_put(in_fence);
587 submit_cleanup(submit);
588 if (ret)
589 msm_gem_submit_free(submit);
590out_unlock:
591 if (ret && (out_fence_fd >= 0))
592 put_unused_fd(out_fence_fd);
593 mutex_unlock(&dev->struct_mutex);
594 return ret;
595}