Merge tag 'for-linus' of git://git.armlinux.org.uk/~rmk/linux-arm
[platform/kernel/linux-rpi.git] / drivers / gpu / drm / msm / msm_gem_submit.c
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   /* obj lock is held */
27 #define BO_ACTIVE   0x2000   /* active refcnt is held */
28 #define BO_PINNED   0x1000   /* obj is pinned and on active list */
29
30 static struct msm_gem_submit *submit_create(struct drm_device *dev,
31                 struct msm_gpu *gpu,
32                 struct msm_gpu_submitqueue *queue, uint32_t nr_bos,
33                 uint32_t nr_cmds)
34 {
35         struct msm_gem_submit *submit;
36         uint64_t sz;
37         int ret;
38
39         sz = struct_size(submit, bos, nr_bos) +
40                         ((u64)nr_cmds * sizeof(submit->cmd[0]));
41
42         if (sz > SIZE_MAX)
43                 return ERR_PTR(-ENOMEM);
44
45         submit = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
46         if (!submit)
47                 return ERR_PTR(-ENOMEM);
48
49         ret = drm_sched_job_init(&submit->base, queue->entity, queue);
50         if (ret) {
51                 kfree(submit);
52                 return ERR_PTR(ret);
53         }
54
55         xa_init_flags(&submit->deps, XA_FLAGS_ALLOC);
56
57         kref_init(&submit->ref);
58         submit->dev = dev;
59         submit->aspace = queue->ctx->aspace;
60         submit->gpu = gpu;
61         submit->cmd = (void *)&submit->bos[nr_bos];
62         submit->queue = queue;
63         submit->ring = gpu->rb[queue->ring_nr];
64         submit->fault_dumped = false;
65
66         INIT_LIST_HEAD(&submit->node);
67
68         return submit;
69 }
70
71 void __msm_gem_submit_destroy(struct kref *kref)
72 {
73         struct msm_gem_submit *submit =
74                         container_of(kref, struct msm_gem_submit, ref);
75         unsigned long index;
76         struct dma_fence *fence;
77         unsigned i;
78
79         if (submit->fence_id) {
80                 mutex_lock(&submit->queue->lock);
81                 idr_remove(&submit->queue->fence_idr, submit->fence_id);
82                 mutex_unlock(&submit->queue->lock);
83         }
84
85         xa_for_each (&submit->deps, index, fence) {
86                 dma_fence_put(fence);
87         }
88
89         xa_destroy(&submit->deps);
90
91         dma_fence_put(submit->user_fence);
92         dma_fence_put(submit->hw_fence);
93
94         put_pid(submit->pid);
95         msm_submitqueue_put(submit->queue);
96
97         for (i = 0; i < submit->nr_cmds; i++)
98                 kfree(submit->cmd[i].relocs);
99
100         kfree(submit);
101 }
102
103 static int submit_lookup_objects(struct msm_gem_submit *submit,
104                 struct drm_msm_gem_submit *args, struct drm_file *file)
105 {
106         unsigned i;
107         int ret = 0;
108
109         for (i = 0; i < args->nr_bos; i++) {
110                 struct drm_msm_gem_submit_bo submit_bo;
111                 void __user *userptr =
112                         u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
113
114                 /* make sure we don't have garbage flags, in case we hit
115                  * error path before flags is initialized:
116                  */
117                 submit->bos[i].flags = 0;
118
119                 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
120                         ret = -EFAULT;
121                         i = 0;
122                         goto out;
123                 }
124
125 /* at least one of READ and/or WRITE flags should be set: */
126 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
127
128                 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
129                         !(submit_bo.flags & MANDATORY_FLAGS)) {
130                         DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
131                         ret = -EINVAL;
132                         i = 0;
133                         goto out;
134                 }
135
136                 submit->bos[i].handle = submit_bo.handle;
137                 submit->bos[i].flags = submit_bo.flags;
138                 /* in validate_objects() we figure out if this is true: */
139                 submit->bos[i].iova  = submit_bo.presumed;
140         }
141
142         spin_lock(&file->table_lock);
143
144         for (i = 0; i < args->nr_bos; i++) {
145                 struct drm_gem_object *obj;
146
147                 /* normally use drm_gem_object_lookup(), but for bulk lookup
148                  * all under single table_lock just hit object_idr directly:
149                  */
150                 obj = idr_find(&file->object_idr, submit->bos[i].handle);
151                 if (!obj) {
152                         DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
153                         ret = -EINVAL;
154                         goto out_unlock;
155                 }
156
157                 drm_gem_object_get(obj);
158
159                 submit->bos[i].obj = to_msm_bo(obj);
160         }
161
162 out_unlock:
163         spin_unlock(&file->table_lock);
164
165 out:
166         submit->nr_bos = i;
167
168         return ret;
169 }
170
171 static int submit_lookup_cmds(struct msm_gem_submit *submit,
172                 struct drm_msm_gem_submit *args, struct drm_file *file)
173 {
174         unsigned i;
175         size_t sz;
176         int ret = 0;
177
178         for (i = 0; i < args->nr_cmds; i++) {
179                 struct drm_msm_gem_submit_cmd submit_cmd;
180                 void __user *userptr =
181                         u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
182
183                 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
184                 if (ret) {
185                         ret = -EFAULT;
186                         goto out;
187                 }
188
189                 /* validate input from userspace: */
190                 switch (submit_cmd.type) {
191                 case MSM_SUBMIT_CMD_BUF:
192                 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
193                 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
194                         break;
195                 default:
196                         DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
197                         return -EINVAL;
198                 }
199
200                 if (submit_cmd.size % 4) {
201                         DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
202                                         submit_cmd.size);
203                         ret = -EINVAL;
204                         goto out;
205                 }
206
207                 submit->cmd[i].type = submit_cmd.type;
208                 submit->cmd[i].size = submit_cmd.size / 4;
209                 submit->cmd[i].offset = submit_cmd.submit_offset / 4;
210                 submit->cmd[i].idx  = submit_cmd.submit_idx;
211                 submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
212
213                 userptr = u64_to_user_ptr(submit_cmd.relocs);
214
215                 sz = array_size(submit_cmd.nr_relocs,
216                                 sizeof(struct drm_msm_gem_submit_reloc));
217                 /* check for overflow: */
218                 if (sz == SIZE_MAX) {
219                         ret = -ENOMEM;
220                         goto out;
221                 }
222                 submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL);
223                 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
224                 if (ret) {
225                         ret = -EFAULT;
226                         goto out;
227                 }
228         }
229
230 out:
231         return ret;
232 }
233
234 /* Unwind bo state, according to cleanup_flags.  In the success case, only
235  * the lock is dropped at the end of the submit (and active/pin ref is dropped
236  * later when the submit is retired).
237  */
238 static void submit_cleanup_bo(struct msm_gem_submit *submit, int i,
239                 unsigned cleanup_flags)
240 {
241         struct drm_gem_object *obj = &submit->bos[i].obj->base;
242         unsigned flags = submit->bos[i].flags & cleanup_flags;
243
244         if (flags & BO_PINNED)
245                 msm_gem_unpin_iova_locked(obj, submit->aspace);
246
247         if (flags & BO_ACTIVE)
248                 msm_gem_active_put(obj);
249
250         if (flags & BO_LOCKED)
251                 dma_resv_unlock(obj->resv);
252
253         submit->bos[i].flags &= ~cleanup_flags;
254 }
255
256 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
257 {
258         submit_cleanup_bo(submit, i, BO_PINNED | BO_ACTIVE | BO_LOCKED);
259
260         if (!(submit->bos[i].flags & BO_VALID))
261                 submit->bos[i].iova = 0;
262 }
263
264 /* This is where we make sure all the bo's are reserved and pin'd: */
265 static int submit_lock_objects(struct msm_gem_submit *submit)
266 {
267         int contended, slow_locked = -1, i, ret = 0;
268
269 retry:
270         for (i = 0; i < submit->nr_bos; i++) {
271                 struct msm_gem_object *msm_obj = submit->bos[i].obj;
272
273                 if (slow_locked == i)
274                         slow_locked = -1;
275
276                 contended = i;
277
278                 if (!(submit->bos[i].flags & BO_LOCKED)) {
279                         ret = dma_resv_lock_interruptible(msm_obj->base.resv,
280                                                           &submit->ticket);
281                         if (ret)
282                                 goto fail;
283                         submit->bos[i].flags |= BO_LOCKED;
284                 }
285         }
286
287         ww_acquire_done(&submit->ticket);
288
289         return 0;
290
291 fail:
292         if (ret == -EALREADY) {
293                 DRM_ERROR("handle %u at index %u already on submit list\n",
294                                 submit->bos[i].handle, i);
295                 ret = -EINVAL;
296         }
297
298         for (; i >= 0; i--)
299                 submit_unlock_unpin_bo(submit, i);
300
301         if (slow_locked > 0)
302                 submit_unlock_unpin_bo(submit, slow_locked);
303
304         if (ret == -EDEADLK) {
305                 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
306                 /* we lost out in a seqno race, lock and retry.. */
307                 ret = dma_resv_lock_slow_interruptible(msm_obj->base.resv,
308                                                        &submit->ticket);
309                 if (!ret) {
310                         submit->bos[contended].flags |= BO_LOCKED;
311                         slow_locked = contended;
312                         goto retry;
313                 }
314
315                 /* Not expecting -EALREADY here, if the bo was already
316                  * locked, we should have gotten -EALREADY already from
317                  * the dma_resv_lock_interruptable() call.
318                  */
319                 WARN_ON_ONCE(ret == -EALREADY);
320         }
321
322         return ret;
323 }
324
325 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
326 {
327         int i, ret = 0;
328
329         for (i = 0; i < submit->nr_bos; i++) {
330                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
331                 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
332
333                 if (!write) {
334                         /* NOTE: _reserve_shared() must happen before
335                          * _add_shared_fence(), which makes this a slightly
336                          * strange place to call it.  OTOH this is a
337                          * convenient can-fail point to hook it in.
338                          */
339                         ret = dma_resv_reserve_shared(obj->resv, 1);
340                         if (ret)
341                                 return ret;
342                 }
343
344                 if (no_implicit)
345                         continue;
346
347                 ret = drm_gem_fence_array_add_implicit(&submit->deps, obj,
348                         write);
349                 if (ret)
350                         break;
351         }
352
353         return ret;
354 }
355
356 static int submit_pin_objects(struct msm_gem_submit *submit)
357 {
358         int i, ret = 0;
359
360         submit->valid = true;
361
362         /*
363          * Increment active_count first, so if under memory pressure, we
364          * don't inadvertently evict a bo needed by the submit in order
365          * to pin an earlier bo in the same submit.
366          */
367         for (i = 0; i < submit->nr_bos; i++) {
368                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
369
370                 msm_gem_active_get(obj, submit->gpu);
371                 submit->bos[i].flags |= BO_ACTIVE;
372         }
373
374         for (i = 0; i < submit->nr_bos; i++) {
375                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
376                 uint64_t iova;
377
378                 /* if locking succeeded, pin bo: */
379                 ret = msm_gem_get_and_pin_iova_locked(obj,
380                                 submit->aspace, &iova);
381
382                 if (ret)
383                         break;
384
385                 submit->bos[i].flags |= BO_PINNED;
386
387                 if (iova == submit->bos[i].iova) {
388                         submit->bos[i].flags |= BO_VALID;
389                 } else {
390                         submit->bos[i].iova = iova;
391                         /* iova changed, so address in cmdstream is not valid: */
392                         submit->bos[i].flags &= ~BO_VALID;
393                         submit->valid = false;
394                 }
395         }
396
397         return ret;
398 }
399
400 static void submit_attach_object_fences(struct msm_gem_submit *submit)
401 {
402         int i;
403
404         for (i = 0; i < submit->nr_bos; i++) {
405                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
406
407                 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
408                         dma_resv_add_excl_fence(obj->resv, submit->user_fence);
409                 else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
410                         dma_resv_add_shared_fence(obj->resv, submit->user_fence);
411         }
412 }
413
414 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
415                 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
416 {
417         if (idx >= submit->nr_bos) {
418                 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
419                                 idx, submit->nr_bos);
420                 return -EINVAL;
421         }
422
423         if (obj)
424                 *obj = submit->bos[idx].obj;
425         if (iova)
426                 *iova = submit->bos[idx].iova;
427         if (valid)
428                 *valid = !!(submit->bos[idx].flags & BO_VALID);
429
430         return 0;
431 }
432
433 /* process the reloc's and patch up the cmdstream as needed: */
434 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
435                 uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs)
436 {
437         uint32_t i, last_offset = 0;
438         uint32_t *ptr;
439         int ret = 0;
440
441         if (!nr_relocs)
442                 return 0;
443
444         if (offset % 4) {
445                 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
446                 return -EINVAL;
447         }
448
449         /* For now, just map the entire thing.  Eventually we probably
450          * to do it page-by-page, w/ kmap() if not vmap()d..
451          */
452         ptr = msm_gem_get_vaddr_locked(&obj->base);
453
454         if (IS_ERR(ptr)) {
455                 ret = PTR_ERR(ptr);
456                 DBG("failed to map: %d", ret);
457                 return ret;
458         }
459
460         for (i = 0; i < nr_relocs; i++) {
461                 struct drm_msm_gem_submit_reloc submit_reloc = relocs[i];
462                 uint32_t off;
463                 uint64_t iova;
464                 bool valid;
465
466                 if (submit_reloc.submit_offset % 4) {
467                         DRM_ERROR("non-aligned reloc offset: %u\n",
468                                         submit_reloc.submit_offset);
469                         ret = -EINVAL;
470                         goto out;
471                 }
472
473                 /* offset in dwords: */
474                 off = submit_reloc.submit_offset / 4;
475
476                 if ((off >= (obj->base.size / 4)) ||
477                                 (off < last_offset)) {
478                         DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
479                         ret = -EINVAL;
480                         goto out;
481                 }
482
483                 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
484                 if (ret)
485                         goto out;
486
487                 if (valid)
488                         continue;
489
490                 iova += submit_reloc.reloc_offset;
491
492                 if (submit_reloc.shift < 0)
493                         iova >>= -submit_reloc.shift;
494                 else
495                         iova <<= submit_reloc.shift;
496
497                 ptr[off] = iova | submit_reloc.or;
498
499                 last_offset = off;
500         }
501
502 out:
503         msm_gem_put_vaddr_locked(&obj->base);
504
505         return ret;
506 }
507
508 /* Cleanup submit at end of ioctl.  In the error case, this also drops
509  * references, unpins, and drops active refcnt.  In the non-error case,
510  * this is done when the submit is retired.
511  */
512 static void submit_cleanup(struct msm_gem_submit *submit, bool error)
513 {
514         unsigned cleanup_flags = BO_LOCKED;
515         unsigned i;
516
517         if (error)
518                 cleanup_flags |= BO_PINNED | BO_ACTIVE;
519
520         for (i = 0; i < submit->nr_bos; i++) {
521                 struct msm_gem_object *msm_obj = submit->bos[i].obj;
522                 submit_cleanup_bo(submit, i, cleanup_flags);
523                 if (error)
524                         drm_gem_object_put(&msm_obj->base);
525         }
526 }
527
528 void msm_submit_retire(struct msm_gem_submit *submit)
529 {
530         int i;
531
532         for (i = 0; i < submit->nr_bos; i++) {
533                 struct drm_gem_object *obj = &submit->bos[i].obj->base;
534
535                 msm_gem_lock(obj);
536                 submit_cleanup_bo(submit, i, BO_PINNED | BO_ACTIVE);
537                 msm_gem_unlock(obj);
538                 drm_gem_object_put(obj);
539         }
540 }
541
542 struct msm_submit_post_dep {
543         struct drm_syncobj *syncobj;
544         uint64_t point;
545         struct dma_fence_chain *chain;
546 };
547
548 static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
549                                            struct drm_file *file,
550                                            uint64_t in_syncobjs_addr,
551                                            uint32_t nr_in_syncobjs,
552                                            size_t syncobj_stride,
553                                            struct msm_ringbuffer *ring)
554 {
555         struct drm_syncobj **syncobjs = NULL;
556         struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
557         int ret = 0;
558         uint32_t i, j;
559
560         syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
561                            GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
562         if (!syncobjs)
563                 return ERR_PTR(-ENOMEM);
564
565         for (i = 0; i < nr_in_syncobjs; ++i) {
566                 uint64_t address = in_syncobjs_addr + i * syncobj_stride;
567                 struct dma_fence *fence;
568
569                 if (copy_from_user(&syncobj_desc,
570                                    u64_to_user_ptr(address),
571                                    min(syncobj_stride, sizeof(syncobj_desc)))) {
572                         ret = -EFAULT;
573                         break;
574                 }
575
576                 if (syncobj_desc.point &&
577                     !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
578                         ret = -EOPNOTSUPP;
579                         break;
580                 }
581
582                 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
583                         ret = -EINVAL;
584                         break;
585                 }
586
587                 ret = drm_syncobj_find_fence(file, syncobj_desc.handle,
588                                              syncobj_desc.point, 0, &fence);
589                 if (ret)
590                         break;
591
592                 ret = drm_gem_fence_array_add(&submit->deps, fence);
593                 if (ret)
594                         break;
595
596                 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
597                         syncobjs[i] =
598                                 drm_syncobj_find(file, syncobj_desc.handle);
599                         if (!syncobjs[i]) {
600                                 ret = -EINVAL;
601                                 break;
602                         }
603                 }
604         }
605
606         if (ret) {
607                 for (j = 0; j <= i; ++j) {
608                         if (syncobjs[j])
609                                 drm_syncobj_put(syncobjs[j]);
610                 }
611                 kfree(syncobjs);
612                 return ERR_PTR(ret);
613         }
614         return syncobjs;
615 }
616
617 static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
618                                uint32_t nr_syncobjs)
619 {
620         uint32_t i;
621
622         for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
623                 if (syncobjs[i])
624                         drm_syncobj_replace_fence(syncobjs[i], NULL);
625         }
626 }
627
628 static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
629                                                        struct drm_file *file,
630                                                        uint64_t syncobjs_addr,
631                                                        uint32_t nr_syncobjs,
632                                                        size_t syncobj_stride)
633 {
634         struct msm_submit_post_dep *post_deps;
635         struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
636         int ret = 0;
637         uint32_t i, j;
638
639         post_deps = kmalloc_array(nr_syncobjs, sizeof(*post_deps),
640                                   GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
641         if (!post_deps)
642                 return ERR_PTR(-ENOMEM);
643
644         for (i = 0; i < nr_syncobjs; ++i) {
645                 uint64_t address = syncobjs_addr + i * syncobj_stride;
646
647                 if (copy_from_user(&syncobj_desc,
648                                    u64_to_user_ptr(address),
649                                    min(syncobj_stride, sizeof(syncobj_desc)))) {
650                         ret = -EFAULT;
651                         break;
652                 }
653
654                 post_deps[i].point = syncobj_desc.point;
655                 post_deps[i].chain = NULL;
656
657                 if (syncobj_desc.flags) {
658                         ret = -EINVAL;
659                         break;
660                 }
661
662                 if (syncobj_desc.point) {
663                         if (!drm_core_check_feature(dev,
664                                                     DRIVER_SYNCOBJ_TIMELINE)) {
665                                 ret = -EOPNOTSUPP;
666                                 break;
667                         }
668
669                         post_deps[i].chain = dma_fence_chain_alloc();
670                         if (!post_deps[i].chain) {
671                                 ret = -ENOMEM;
672                                 break;
673                         }
674                 }
675
676                 post_deps[i].syncobj =
677                         drm_syncobj_find(file, syncobj_desc.handle);
678                 if (!post_deps[i].syncobj) {
679                         ret = -EINVAL;
680                         break;
681                 }
682         }
683
684         if (ret) {
685                 for (j = 0; j <= i; ++j) {
686                         dma_fence_chain_free(post_deps[j].chain);
687                         if (post_deps[j].syncobj)
688                                 drm_syncobj_put(post_deps[j].syncobj);
689                 }
690
691                 kfree(post_deps);
692                 return ERR_PTR(ret);
693         }
694
695         return post_deps;
696 }
697
698 static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
699                                   uint32_t count, struct dma_fence *fence)
700 {
701         uint32_t i;
702
703         for (i = 0; post_deps && i < count; ++i) {
704                 if (post_deps[i].chain) {
705                         drm_syncobj_add_point(post_deps[i].syncobj,
706                                               post_deps[i].chain,
707                                               fence, post_deps[i].point);
708                         post_deps[i].chain = NULL;
709                 } else {
710                         drm_syncobj_replace_fence(post_deps[i].syncobj,
711                                                   fence);
712                 }
713         }
714 }
715
716 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
717                 struct drm_file *file)
718 {
719         static atomic_t ident = ATOMIC_INIT(0);
720         struct msm_drm_private *priv = dev->dev_private;
721         struct drm_msm_gem_submit *args = data;
722         struct msm_file_private *ctx = file->driver_priv;
723         struct msm_gem_submit *submit = NULL;
724         struct msm_gpu *gpu = priv->gpu;
725         struct msm_gpu_submitqueue *queue;
726         struct msm_ringbuffer *ring;
727         struct msm_submit_post_dep *post_deps = NULL;
728         struct drm_syncobj **syncobjs_to_reset = NULL;
729         int out_fence_fd = -1;
730         struct pid *pid = get_pid(task_pid(current));
731         bool has_ww_ticket = false;
732         unsigned i;
733         int ret, submitid;
734
735         if (!gpu)
736                 return -ENXIO;
737
738         if (args->pad)
739                 return -EINVAL;
740
741         /* for now, we just have 3d pipe.. eventually this would need to
742          * be more clever to dispatch to appropriate gpu module:
743          */
744         if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
745                 return -EINVAL;
746
747         if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
748                 return -EINVAL;
749
750         if (args->flags & MSM_SUBMIT_SUDO) {
751                 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
752                     !capable(CAP_SYS_RAWIO))
753                         return -EINVAL;
754         }
755
756         queue = msm_submitqueue_get(ctx, args->queueid);
757         if (!queue)
758                 return -ENOENT;
759
760         /* Get a unique identifier for the submission for logging purposes */
761         submitid = atomic_inc_return(&ident) - 1;
762
763         ring = gpu->rb[queue->ring_nr];
764         trace_msm_gpu_submit(pid_nr(pid), ring->id, submitid,
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_FENCE_FD_OUT) {
772                 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
773                 if (out_fence_fd < 0) {
774                         ret = out_fence_fd;
775                         goto out_unlock;
776                 }
777         }
778
779         submit = submit_create(dev, gpu, queue, args->nr_bos,
780                 args->nr_cmds);
781         if (IS_ERR(submit)) {
782                 ret = PTR_ERR(submit);
783                 goto out_unlock;
784         }
785
786         submit->pid = pid;
787         submit->ident = submitid;
788
789         if (args->flags & MSM_SUBMIT_SUDO)
790                 submit->in_rb = true;
791
792         if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
793                 struct dma_fence *in_fence;
794
795                 in_fence = sync_file_get_fence(args->fence_fd);
796
797                 if (!in_fence) {
798                         ret = -EINVAL;
799                         goto out_unlock;
800                 }
801
802                 ret = drm_gem_fence_array_add(&submit->deps, in_fence);
803                 if (ret)
804                         goto out_unlock;
805         }
806
807         if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
808                 syncobjs_to_reset = msm_parse_deps(submit, file,
809                                                    args->in_syncobjs,
810                                                    args->nr_in_syncobjs,
811                                                    args->syncobj_stride, ring);
812                 if (IS_ERR(syncobjs_to_reset)) {
813                         ret = PTR_ERR(syncobjs_to_reset);
814                         goto out_unlock;
815                 }
816         }
817
818         if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
819                 post_deps = msm_parse_post_deps(dev, file,
820                                                 args->out_syncobjs,
821                                                 args->nr_out_syncobjs,
822                                                 args->syncobj_stride);
823                 if (IS_ERR(post_deps)) {
824                         ret = PTR_ERR(post_deps);
825                         goto out_unlock;
826                 }
827         }
828
829         ret = submit_lookup_objects(submit, args, file);
830         if (ret)
831                 goto out;
832
833         ret = submit_lookup_cmds(submit, args, file);
834         if (ret)
835                 goto out;
836
837         /* copy_*_user while holding a ww ticket upsets lockdep */
838         ww_acquire_init(&submit->ticket, &reservation_ww_class);
839         has_ww_ticket = true;
840         ret = submit_lock_objects(submit);
841         if (ret)
842                 goto out;
843
844         ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
845         if (ret)
846                 goto out;
847
848         ret = submit_pin_objects(submit);
849         if (ret)
850                 goto out;
851
852         for (i = 0; i < args->nr_cmds; i++) {
853                 struct msm_gem_object *msm_obj;
854                 uint64_t iova;
855
856                 ret = submit_bo(submit, submit->cmd[i].idx,
857                                 &msm_obj, &iova, NULL);
858                 if (ret)
859                         goto out;
860
861                 if (!submit->cmd[i].size ||
862                         ((submit->cmd[i].size + submit->cmd[i].offset) >
863                                 msm_obj->base.size / 4)) {
864                         DRM_ERROR("invalid cmdstream size: %u\n", submit->cmd[i].size * 4);
865                         ret = -EINVAL;
866                         goto out;
867                 }
868
869                 submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
870
871                 if (submit->valid)
872                         continue;
873
874                 ret = submit_reloc(submit, msm_obj, submit->cmd[i].offset * 4,
875                                 submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
876                 if (ret)
877                         goto out;
878         }
879
880         submit->nr_cmds = i;
881
882         submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
883
884         /*
885          * Allocate an id which can be used by WAIT_FENCE ioctl to map back
886          * to the underlying fence.
887          */
888         submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
889                         submit->user_fence, 0, INT_MAX, GFP_KERNEL);
890         if (submit->fence_id < 0) {
891                 ret = submit->fence_id = 0;
892                 submit->fence_id = 0;
893                 goto out;
894         }
895
896         if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
897                 struct sync_file *sync_file = sync_file_create(submit->user_fence);
898                 if (!sync_file) {
899                         ret = -ENOMEM;
900                         goto out;
901                 }
902                 fd_install(out_fence_fd, sync_file->file);
903                 args->fence_fd = out_fence_fd;
904         }
905
906         submit_attach_object_fences(submit);
907
908         /* The scheduler owns a ref now: */
909         msm_gem_submit_get(submit);
910
911         drm_sched_entity_push_job(&submit->base, queue->entity);
912
913         args->fence = submit->fence_id;
914
915         msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
916         msm_process_post_deps(post_deps, args->nr_out_syncobjs,
917                               submit->user_fence);
918
919
920 out:
921         submit_cleanup(submit, !!ret);
922         if (has_ww_ticket)
923                 ww_acquire_fini(&submit->ticket);
924 out_unlock:
925         if (ret && (out_fence_fd >= 0))
926                 put_unused_fd(out_fence_fd);
927         mutex_unlock(&queue->lock);
928         if (submit)
929                 msm_gem_submit_put(submit);
930 out_post_unlock:
931         if (!IS_ERR_OR_NULL(post_deps)) {
932                 for (i = 0; i < args->nr_out_syncobjs; ++i) {
933                         kfree(post_deps[i].chain);
934                         drm_syncobj_put(post_deps[i].syncobj);
935                 }
936                 kfree(post_deps);
937         }
938
939         if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
940                 for (i = 0; i < args->nr_in_syncobjs; ++i) {
941                         if (syncobjs_to_reset[i])
942                                 drm_syncobj_put(syncobjs_to_reset[i]);
943                 }
944                 kfree(syncobjs_to_reset);
945         }
946
947         return ret;
948 }