1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
5 */
6
7 #include <linux/file.h>
8 #include <linux/sync_file.h>
9 #include <linux/uaccess.h>
10
11 #include <drm/drm_drv.h>
12 #include <drm/drm_file.h>
13 #include <drm/drm_syncobj.h>
14
15 #include "msm_drv.h"
16 #include "msm_gpu.h"
17 #include "msm_gem.h"
18 #include "msm_gpu_trace.h"
19
20 /*
21 * Cmdstream submission:
22 */
23
submit_create(struct drm_device * dev,struct msm_gpu * gpu,struct msm_gpu_submitqueue * queue,uint32_t nr_bos,uint32_t nr_cmds)24 static struct msm_gem_submit *submit_create(struct drm_device *dev,
25 struct msm_gpu *gpu,
26 struct msm_gpu_submitqueue *queue, uint32_t nr_bos,
27 uint32_t nr_cmds)
28 {
29 static atomic_t ident = ATOMIC_INIT(0);
30 struct msm_gem_submit *submit;
31 uint64_t sz;
32 int ret;
33
34 sz = struct_size(submit, bos, nr_bos) +
35 ((u64)nr_cmds * sizeof(submit->cmd[0]));
36
37 if (sz > SIZE_MAX)
38 return ERR_PTR(-ENOMEM);
39
40 submit = kzalloc(sz, GFP_KERNEL);
41 if (!submit)
42 return ERR_PTR(-ENOMEM);
43
44 ret = drm_sched_job_init(&submit->base, queue->entity, queue);
45 if (ret) {
46 kfree(submit);
47 return ERR_PTR(ret);
48 }
49
50 kref_init(&submit->ref);
51 submit->dev = dev;
52 submit->aspace = queue->ctx->aspace;
53 submit->gpu = gpu;
54 submit->cmd = (void *)&submit->bos[nr_bos];
55 submit->queue = queue;
56 submit->pid = get_pid(task_pid(current));
57 submit->ring = gpu->rb[queue->ring_nr];
58 submit->fault_dumped = false;
59
60 /* Get a unique identifier for the submission for logging purposes */
61 submit->ident = atomic_inc_return(&ident) - 1;
62
63 INIT_LIST_HEAD(&submit->node);
64
65 return submit;
66 }
67
__msm_gem_submit_destroy(struct kref * kref)68 void __msm_gem_submit_destroy(struct kref *kref)
69 {
70 struct msm_gem_submit *submit =
71 container_of(kref, struct msm_gem_submit, ref);
72 unsigned i;
73
74 if (submit->fence_id) {
75 mutex_lock(&submit->queue->idr_lock);
76 idr_remove(&submit->queue->fence_idr, submit->fence_id);
77 mutex_unlock(&submit->queue->idr_lock);
78 }
79
80 dma_fence_put(submit->user_fence);
81 dma_fence_put(submit->hw_fence);
82
83 put_pid(submit->pid);
84 msm_submitqueue_put(submit->queue);
85
86 for (i = 0; i < submit->nr_cmds; i++)
87 kfree(submit->cmd[i].relocs);
88
89 kfree(submit);
90 }
91
submit_lookup_objects(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)92 static int submit_lookup_objects(struct msm_gem_submit *submit,
93 struct drm_msm_gem_submit *args, struct drm_file *file)
94 {
95 unsigned i;
96 int ret = 0;
97
98 for (i = 0; i < args->nr_bos; i++) {
99 struct drm_msm_gem_submit_bo submit_bo;
100 void __user *userptr =
101 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
102
103 /* make sure we don't have garbage flags, in case we hit
104 * error path before flags is initialized:
105 */
106 submit->bos[i].flags = 0;
107
108 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
109 ret = -EFAULT;
110 i = 0;
111 goto out;
112 }
113
114 /* at least one of READ and/or WRITE flags should be set: */
115 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
116
117 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
118 !(submit_bo.flags & MANDATORY_FLAGS)) {
119 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
120 ret = -EINVAL;
121 i = 0;
122 goto out;
123 }
124
125 submit->bos[i].handle = submit_bo.handle;
126 submit->bos[i].flags = submit_bo.flags;
127 /* in validate_objects() we figure out if this is true: */
128 submit->bos[i].iova = submit_bo.presumed;
129 }
130
131 spin_lock(&file->table_lock);
132
133 for (i = 0; i < args->nr_bos; i++) {
134 struct drm_gem_object *obj;
135
136 /* normally use drm_gem_object_lookup(), but for bulk lookup
137 * all under single table_lock just hit object_idr directly:
138 */
139 obj = idr_find(&file->object_idr, submit->bos[i].handle);
140 if (!obj) {
141 DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
142 ret = -EINVAL;
143 goto out_unlock;
144 }
145
146 drm_gem_object_get(obj);
147
148 submit->bos[i].obj = to_msm_bo(obj);
149 }
150
151 out_unlock:
152 spin_unlock(&file->table_lock);
153
154 out:
155 submit->nr_bos = i;
156
157 return ret;
158 }
159
submit_lookup_cmds(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)160 static int submit_lookup_cmds(struct msm_gem_submit *submit,
161 struct drm_msm_gem_submit *args, struct drm_file *file)
162 {
163 unsigned i;
164 size_t sz;
165 int ret = 0;
166
167 for (i = 0; i < args->nr_cmds; i++) {
168 struct drm_msm_gem_submit_cmd submit_cmd;
169 void __user *userptr =
170 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
171
172 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
173 if (ret) {
174 ret = -EFAULT;
175 goto out;
176 }
177
178 /* validate input from userspace: */
179 switch (submit_cmd.type) {
180 case MSM_SUBMIT_CMD_BUF:
181 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
182 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
183 break;
184 default:
185 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
186 return -EINVAL;
187 }
188
189 if (submit_cmd.size % 4) {
190 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
191 submit_cmd.size);
192 ret = -EINVAL;
193 goto out;
194 }
195
196 submit->cmd[i].type = submit_cmd.type;
197 submit->cmd[i].size = submit_cmd.size / 4;
198 submit->cmd[i].offset = submit_cmd.submit_offset / 4;
199 submit->cmd[i].idx = submit_cmd.submit_idx;
200 submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
201
202 userptr = u64_to_user_ptr(submit_cmd.relocs);
203
204 sz = array_size(submit_cmd.nr_relocs,
205 sizeof(struct drm_msm_gem_submit_reloc));
206 /* check for overflow: */
207 if (sz == SIZE_MAX) {
208 ret = -ENOMEM;
209 goto out;
210 }
211 submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL);
212 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
213 if (ret) {
214 ret = -EFAULT;
215 goto out;
216 }
217 }
218
219 out:
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 */
submit_cleanup_bo(struct msm_gem_submit * submit,int i,unsigned cleanup_flags)227 static 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
submit_unlock_unpin_bo(struct msm_gem_submit * submit,int i)250 static 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: */
submit_lock_objects(struct msm_gem_submit * submit)260 static int submit_lock_objects(struct msm_gem_submit *submit)
261 {
262 int contended, slow_locked = -1, i, ret = 0;
263
264 retry:
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
286 fail:
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
submit_fence_sync(struct msm_gem_submit * submit,bool no_implicit)320 static 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 /* exclusive fences must be ordered */
338 if (no_implicit && !write)
339 continue;
340
341 ret = drm_sched_job_add_implicit_dependencies(&submit->base,
342 obj,
343 write);
344 if (ret)
345 break;
346 }
347
348 return ret;
349 }
350
submit_pin_objects(struct msm_gem_submit * submit)351 static int submit_pin_objects(struct msm_gem_submit *submit)
352 {
353 int i, ret = 0;
354
355 submit->valid = true;
356
357 for (i = 0; i < submit->nr_bos; i++) {
358 struct drm_gem_object *obj = &submit->bos[i].obj->base;
359 struct msm_gem_vma *vma;
360
361 /* if locking succeeded, pin bo: */
362 vma = msm_gem_get_vma_locked(obj, submit->aspace);
363 if (IS_ERR(vma)) {
364 ret = PTR_ERR(vma);
365 break;
366 }
367
368 ret = msm_gem_pin_vma_locked(obj, vma);
369 if (ret)
370 break;
371
372 submit->bos[i].flags |= BO_OBJ_PINNED | BO_VMA_PINNED;
373 submit->bos[i].vma = vma;
374
375 if (vma->iova == submit->bos[i].iova) {
376 submit->bos[i].flags |= BO_VALID;
377 } else {
378 submit->bos[i].iova = vma->iova;
379 /* iova changed, so address in cmdstream is not valid: */
380 submit->bos[i].flags &= ~BO_VALID;
381 submit->valid = false;
382 }
383 }
384
385 return ret;
386 }
387
submit_attach_object_fences(struct msm_gem_submit * submit)388 static void submit_attach_object_fences(struct msm_gem_submit *submit)
389 {
390 int i;
391
392 for (i = 0; i < submit->nr_bos; i++) {
393 struct drm_gem_object *obj = &submit->bos[i].obj->base;
394
395 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
396 dma_resv_add_fence(obj->resv, submit->user_fence,
397 DMA_RESV_USAGE_WRITE);
398 else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
399 dma_resv_add_fence(obj->resv, submit->user_fence,
400 DMA_RESV_USAGE_READ);
401 }
402 }
403
submit_bo(struct msm_gem_submit * submit,uint32_t idx,struct msm_gem_object ** obj,uint64_t * iova,bool * valid)404 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
405 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
406 {
407 if (idx >= submit->nr_bos) {
408 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
409 idx, submit->nr_bos);
410 return -EINVAL;
411 }
412
413 if (obj)
414 *obj = submit->bos[idx].obj;
415 if (iova)
416 *iova = submit->bos[idx].iova;
417 if (valid)
418 *valid = !!(submit->bos[idx].flags & BO_VALID);
419
420 return 0;
421 }
422
423 /* process the reloc's and patch up the cmdstream as needed: */
submit_reloc(struct msm_gem_submit * submit,struct msm_gem_object * obj,uint32_t offset,uint32_t nr_relocs,struct drm_msm_gem_submit_reloc * relocs)424 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
425 uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs)
426 {
427 uint32_t i, last_offset = 0;
428 uint32_t *ptr;
429 int ret = 0;
430
431 if (!nr_relocs)
432 return 0;
433
434 if (offset % 4) {
435 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
436 return -EINVAL;
437 }
438
439 /* For now, just map the entire thing. Eventually we probably
440 * to do it page-by-page, w/ kmap() if not vmap()d..
441 */
442 ptr = msm_gem_get_vaddr_locked(&obj->base);
443
444 if (IS_ERR(ptr)) {
445 ret = PTR_ERR(ptr);
446 DBG("failed to map: %d", ret);
447 return ret;
448 }
449
450 for (i = 0; i < nr_relocs; i++) {
451 struct drm_msm_gem_submit_reloc submit_reloc = relocs[i];
452 uint32_t off;
453 uint64_t iova;
454 bool valid;
455
456 if (submit_reloc.submit_offset % 4) {
457 DRM_ERROR("non-aligned reloc offset: %u\n",
458 submit_reloc.submit_offset);
459 ret = -EINVAL;
460 goto out;
461 }
462
463 /* offset in dwords: */
464 off = submit_reloc.submit_offset / 4;
465
466 if ((off >= (obj->base.size / 4)) ||
467 (off < last_offset)) {
468 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
469 ret = -EINVAL;
470 goto out;
471 }
472
473 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
474 if (ret)
475 goto out;
476
477 if (valid)
478 continue;
479
480 iova += submit_reloc.reloc_offset;
481
482 if (submit_reloc.shift < 0)
483 iova >>= -submit_reloc.shift;
484 else
485 iova <<= submit_reloc.shift;
486
487 ptr[off] = iova | submit_reloc.or;
488
489 last_offset = off;
490 }
491
492 out:
493 msm_gem_put_vaddr_locked(&obj->base);
494
495 return ret;
496 }
497
498 /* Cleanup submit at end of ioctl. In the error case, this also drops
499 * references, unpins, and drops active refcnt. In the non-error case,
500 * this is done when the submit is retired.
501 */
submit_cleanup(struct msm_gem_submit * submit,bool error)502 static void submit_cleanup(struct msm_gem_submit *submit, bool error)
503 {
504 unsigned cleanup_flags = BO_LOCKED;
505 unsigned i;
506
507 if (error)
508 cleanup_flags |= BO_VMA_PINNED | BO_OBJ_PINNED;
509
510 for (i = 0; i < submit->nr_bos; i++) {
511 struct msm_gem_object *msm_obj = submit->bos[i].obj;
512 submit_cleanup_bo(submit, i, cleanup_flags);
513 if (error)
514 drm_gem_object_put(&msm_obj->base);
515 }
516 }
517
msm_submit_retire(struct msm_gem_submit * submit)518 void msm_submit_retire(struct msm_gem_submit *submit)
519 {
520 int i;
521
522 for (i = 0; i < submit->nr_bos; i++) {
523 struct drm_gem_object *obj = &submit->bos[i].obj->base;
524
525 drm_gem_object_put(obj);
526 }
527 }
528
529 struct msm_submit_post_dep {
530 struct drm_syncobj *syncobj;
531 uint64_t point;
532 struct dma_fence_chain *chain;
533 };
534
msm_parse_deps(struct msm_gem_submit * submit,struct drm_file * file,uint64_t in_syncobjs_addr,uint32_t nr_in_syncobjs,size_t syncobj_stride,struct msm_ringbuffer * ring)535 static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
536 struct drm_file *file,
537 uint64_t in_syncobjs_addr,
538 uint32_t nr_in_syncobjs,
539 size_t syncobj_stride,
540 struct msm_ringbuffer *ring)
541 {
542 struct drm_syncobj **syncobjs = NULL;
543 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
544 int ret = 0;
545 uint32_t i, j;
546
547 syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
548 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
549 if (!syncobjs)
550 return ERR_PTR(-ENOMEM);
551
552 for (i = 0; i < nr_in_syncobjs; ++i) {
553 uint64_t address = in_syncobjs_addr + i * syncobj_stride;
554 struct dma_fence *fence;
555
556 if (copy_from_user(&syncobj_desc,
557 u64_to_user_ptr(address),
558 min(syncobj_stride, sizeof(syncobj_desc)))) {
559 ret = -EFAULT;
560 break;
561 }
562
563 if (syncobj_desc.point &&
564 !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
565 ret = -EOPNOTSUPP;
566 break;
567 }
568
569 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
570 ret = -EINVAL;
571 break;
572 }
573
574 ret = drm_syncobj_find_fence(file, syncobj_desc.handle,
575 syncobj_desc.point, 0, &fence);
576 if (ret)
577 break;
578
579 ret = drm_sched_job_add_dependency(&submit->base, fence);
580 if (ret)
581 break;
582
583 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
584 syncobjs[i] =
585 drm_syncobj_find(file, syncobj_desc.handle);
586 if (!syncobjs[i]) {
587 ret = -EINVAL;
588 break;
589 }
590 }
591 }
592
593 if (ret) {
594 for (j = 0; j <= i; ++j) {
595 if (syncobjs[j])
596 drm_syncobj_put(syncobjs[j]);
597 }
598 kfree(syncobjs);
599 return ERR_PTR(ret);
600 }
601 return syncobjs;
602 }
603
msm_reset_syncobjs(struct drm_syncobj ** syncobjs,uint32_t nr_syncobjs)604 static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
605 uint32_t nr_syncobjs)
606 {
607 uint32_t i;
608
609 for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
610 if (syncobjs[i])
611 drm_syncobj_replace_fence(syncobjs[i], NULL);
612 }
613 }
614
msm_parse_post_deps(struct drm_device * dev,struct drm_file * file,uint64_t syncobjs_addr,uint32_t nr_syncobjs,size_t syncobj_stride)615 static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
616 struct drm_file *file,
617 uint64_t syncobjs_addr,
618 uint32_t nr_syncobjs,
619 size_t syncobj_stride)
620 {
621 struct msm_submit_post_dep *post_deps;
622 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
623 int ret = 0;
624 uint32_t i, j;
625
626 post_deps = kmalloc_array(nr_syncobjs, sizeof(*post_deps),
627 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
628 if (!post_deps)
629 return ERR_PTR(-ENOMEM);
630
631 for (i = 0; i < nr_syncobjs; ++i) {
632 uint64_t address = syncobjs_addr + i * syncobj_stride;
633
634 if (copy_from_user(&syncobj_desc,
635 u64_to_user_ptr(address),
636 min(syncobj_stride, sizeof(syncobj_desc)))) {
637 ret = -EFAULT;
638 break;
639 }
640
641 post_deps[i].point = syncobj_desc.point;
642 post_deps[i].chain = NULL;
643
644 if (syncobj_desc.flags) {
645 ret = -EINVAL;
646 break;
647 }
648
649 if (syncobj_desc.point) {
650 if (!drm_core_check_feature(dev,
651 DRIVER_SYNCOBJ_TIMELINE)) {
652 ret = -EOPNOTSUPP;
653 break;
654 }
655
656 post_deps[i].chain = dma_fence_chain_alloc();
657 if (!post_deps[i].chain) {
658 ret = -ENOMEM;
659 break;
660 }
661 }
662
663 post_deps[i].syncobj =
664 drm_syncobj_find(file, syncobj_desc.handle);
665 if (!post_deps[i].syncobj) {
666 ret = -EINVAL;
667 break;
668 }
669 }
670
671 if (ret) {
672 for (j = 0; j <= i; ++j) {
673 dma_fence_chain_free(post_deps[j].chain);
674 if (post_deps[j].syncobj)
675 drm_syncobj_put(post_deps[j].syncobj);
676 }
677
678 kfree(post_deps);
679 return ERR_PTR(ret);
680 }
681
682 return post_deps;
683 }
684
msm_process_post_deps(struct msm_submit_post_dep * post_deps,uint32_t count,struct dma_fence * fence)685 static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
686 uint32_t count, struct dma_fence *fence)
687 {
688 uint32_t i;
689
690 for (i = 0; post_deps && i < count; ++i) {
691 if (post_deps[i].chain) {
692 drm_syncobj_add_point(post_deps[i].syncobj,
693 post_deps[i].chain,
694 fence, post_deps[i].point);
695 post_deps[i].chain = NULL;
696 } else {
697 drm_syncobj_replace_fence(post_deps[i].syncobj,
698 fence);
699 }
700 }
701 }
702
msm_ioctl_gem_submit(struct drm_device * dev,void * data,struct drm_file * file)703 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
704 struct drm_file *file)
705 {
706 struct msm_drm_private *priv = dev->dev_private;
707 struct drm_msm_gem_submit *args = data;
708 struct msm_file_private *ctx = file->driver_priv;
709 struct msm_gem_submit *submit;
710 struct msm_gpu *gpu = priv->gpu;
711 struct msm_gpu_submitqueue *queue;
712 struct msm_ringbuffer *ring;
713 struct msm_submit_post_dep *post_deps = NULL;
714 struct drm_syncobj **syncobjs_to_reset = NULL;
715 int out_fence_fd = -1;
716 bool has_ww_ticket = false;
717 unsigned i;
718 int ret;
719
720 if (!gpu)
721 return -ENXIO;
722
723 if (args->pad)
724 return -EINVAL;
725
726 if (unlikely(!ctx->aspace) && !capable(CAP_SYS_RAWIO)) {
727 DRM_ERROR_RATELIMITED("IOMMU support or CAP_SYS_RAWIO required!\n");
728 return -EPERM;
729 }
730
731 /* for now, we just have 3d pipe.. eventually this would need to
732 * be more clever to dispatch to appropriate gpu module:
733 */
734 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
735 return -EINVAL;
736
737 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
738 return -EINVAL;
739
740 if (args->flags & MSM_SUBMIT_SUDO) {
741 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
742 !capable(CAP_SYS_RAWIO))
743 return -EINVAL;
744 }
745
746 queue = msm_submitqueue_get(ctx, args->queueid);
747 if (!queue)
748 return -ENOENT;
749
750 ring = gpu->rb[queue->ring_nr];
751
752 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
753 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
754 if (out_fence_fd < 0) {
755 ret = out_fence_fd;
756 return ret;
757 }
758 }
759
760 submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds);
761 if (IS_ERR(submit))
762 return PTR_ERR(submit);
763
764 trace_msm_gpu_submit(pid_nr(submit->pid), ring->id, submit->ident,
765 args->nr_bos, args->nr_cmds);
766
767 ret = mutex_lock_interruptible(&queue->lock);
768 if (ret)
769 goto out_post_unlock;
770
771 if (args->flags & MSM_SUBMIT_SUDO)
772 submit->in_rb = true;
773
774 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
775 struct dma_fence *in_fence;
776
777 in_fence = sync_file_get_fence(args->fence_fd);
778
779 if (!in_fence) {
780 ret = -EINVAL;
781 goto out_unlock;
782 }
783
784 ret = drm_sched_job_add_dependency(&submit->base, in_fence);
785 if (ret)
786 goto out_unlock;
787 }
788
789 if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
790 syncobjs_to_reset = msm_parse_deps(submit, file,
791 args->in_syncobjs,
792 args->nr_in_syncobjs,
793 args->syncobj_stride, ring);
794 if (IS_ERR(syncobjs_to_reset)) {
795 ret = PTR_ERR(syncobjs_to_reset);
796 goto out_unlock;
797 }
798 }
799
800 if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
801 post_deps = msm_parse_post_deps(dev, file,
802 args->out_syncobjs,
803 args->nr_out_syncobjs,
804 args->syncobj_stride);
805 if (IS_ERR(post_deps)) {
806 ret = PTR_ERR(post_deps);
807 goto out_unlock;
808 }
809 }
810
811 ret = submit_lookup_objects(submit, args, file);
812 if (ret)
813 goto out;
814
815 ret = submit_lookup_cmds(submit, args, file);
816 if (ret)
817 goto out;
818
819 /* copy_*_user while holding a ww ticket upsets lockdep */
820 ww_acquire_init(&submit->ticket, &reservation_ww_class);
821 has_ww_ticket = true;
822 ret = submit_lock_objects(submit);
823 if (ret)
824 goto out;
825
826 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
827 if (ret)
828 goto out;
829
830 ret = submit_pin_objects(submit);
831 if (ret)
832 goto out;
833
834 for (i = 0; i < args->nr_cmds; i++) {
835 struct msm_gem_object *msm_obj;
836 uint64_t iova;
837
838 ret = submit_bo(submit, submit->cmd[i].idx,
839 &msm_obj, &iova, NULL);
840 if (ret)
841 goto out;
842
843 if (!submit->cmd[i].size ||
844 ((submit->cmd[i].size + submit->cmd[i].offset) >
845 msm_obj->base.size / 4)) {
846 DRM_ERROR("invalid cmdstream size: %u\n", submit->cmd[i].size * 4);
847 ret = -EINVAL;
848 goto out;
849 }
850
851 submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
852
853 if (submit->valid)
854 continue;
855
856 ret = submit_reloc(submit, msm_obj, submit->cmd[i].offset * 4,
857 submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
858 if (ret)
859 goto out;
860 }
861
862 submit->nr_cmds = i;
863
864 mutex_lock(&queue->idr_lock);
865
866 /*
867 * If using userspace provided seqno fence, validate that the id
868 * is available before arming sched job. Since access to fence_idr
869 * is serialized on the queue lock, the slot should be still avail
870 * after the job is armed
871 */
872 if ((args->flags & MSM_SUBMIT_FENCE_SN_IN) &&
873 idr_find(&queue->fence_idr, args->fence)) {
874 mutex_unlock(&queue->idr_lock);
875 ret = -EINVAL;
876 goto out;
877 }
878
879 drm_sched_job_arm(&submit->base);
880
881 submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
882
883 if (args->flags & MSM_SUBMIT_FENCE_SN_IN) {
884 /*
885 * Userspace has assigned the seqno fence that it wants
886 * us to use. It is an error to pick a fence sequence
887 * number that is not available.
888 */
889 submit->fence_id = args->fence;
890 ret = idr_alloc_u32(&queue->fence_idr, submit->user_fence,
891 &submit->fence_id, submit->fence_id,
892 GFP_KERNEL);
893 /*
894 * We've already validated that the fence_id slot is valid,
895 * so if idr_alloc_u32 failed, it is a kernel bug
896 */
897 WARN_ON(ret);
898 } else {
899 /*
900 * Allocate an id which can be used by WAIT_FENCE ioctl to map
901 * back to the underlying fence.
902 */
903 submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
904 submit->user_fence, 1,
905 INT_MAX, GFP_KERNEL);
906 }
907
908 mutex_unlock(&queue->idr_lock);
909
910 if (submit->fence_id < 0) {
911 ret = submit->fence_id;
912 submit->fence_id = 0;
913 }
914
915 if (ret == 0 && args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
916 struct sync_file *sync_file = sync_file_create(submit->user_fence);
917 if (!sync_file) {
918 ret = -ENOMEM;
919 } else {
920 fd_install(out_fence_fd, sync_file->file);
921 args->fence_fd = out_fence_fd;
922 }
923 }
924
925 submit_attach_object_fences(submit);
926
927 /* The scheduler owns a ref now: */
928 msm_gem_submit_get(submit);
929
930 drm_sched_entity_push_job(&submit->base);
931
932 args->fence = submit->fence_id;
933 queue->last_fence = submit->fence_id;
934
935 msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
936 msm_process_post_deps(post_deps, args->nr_out_syncobjs,
937 submit->user_fence);
938
939
940 out:
941 submit_cleanup(submit, !!ret);
942 if (has_ww_ticket)
943 ww_acquire_fini(&submit->ticket);
944 out_unlock:
945 if (ret && (out_fence_fd >= 0))
946 put_unused_fd(out_fence_fd);
947 mutex_unlock(&queue->lock);
948 out_post_unlock:
949 msm_gem_submit_put(submit);
950 if (!IS_ERR_OR_NULL(post_deps)) {
951 for (i = 0; i < args->nr_out_syncobjs; ++i) {
952 kfree(post_deps[i].chain);
953 drm_syncobj_put(post_deps[i].syncobj);
954 }
955 kfree(post_deps);
956 }
957
958 if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
959 for (i = 0; i < args->nr_in_syncobjs; ++i) {
960 if (syncobjs_to_reset[i])
961 drm_syncobj_put(syncobjs_to_reset[i]);
962 }
963 kfree(syncobjs_to_reset);
964 }
965
966 return ret;
967 }
968