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 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 	submit->hw_fence = msm_fence_alloc();
45 	if (IS_ERR(submit->hw_fence)) {
46 		ret = PTR_ERR(submit->hw_fence);
47 		kfree(submit);
48 		return ERR_PTR(ret);
49 	}
50 
51 	ret = drm_sched_job_init(&submit->base, queue->entity, queue);
52 	if (ret) {
53 		kfree(submit->hw_fence);
54 		kfree(submit);
55 		return ERR_PTR(ret);
56 	}
57 
58 	kref_init(&submit->ref);
59 	submit->dev = dev;
60 	submit->aspace = queue->ctx->aspace;
61 	submit->gpu = gpu;
62 	submit->cmd = (void *)&submit->bos[nr_bos];
63 	submit->queue = queue;
64 	submit->pid = get_pid(task_pid(current));
65 	submit->ring = gpu->rb[queue->ring_nr];
66 	submit->fault_dumped = false;
67 
68 	/* Get a unique identifier for the submission for logging purposes */
69 	submit->ident = atomic_inc_return(&ident) - 1;
70 
71 	INIT_LIST_HEAD(&submit->node);
72 
73 	return submit;
74 }
75 
76 void __msm_gem_submit_destroy(struct kref *kref)
77 {
78 	struct msm_gem_submit *submit =
79 			container_of(kref, struct msm_gem_submit, ref);
80 	unsigned i;
81 
82 	if (submit->fence_id) {
83 		spin_lock(&submit->queue->idr_lock);
84 		idr_remove(&submit->queue->fence_idr, submit->fence_id);
85 		spin_unlock(&submit->queue->idr_lock);
86 	}
87 
88 	dma_fence_put(submit->user_fence);
89 	dma_fence_put(submit->hw_fence);
90 
91 	put_pid(submit->pid);
92 	msm_submitqueue_put(submit->queue);
93 
94 	for (i = 0; i < submit->nr_cmds; i++)
95 		kfree(submit->cmd[i].relocs);
96 
97 	kfree(submit);
98 }
99 
100 static int submit_lookup_objects(struct msm_gem_submit *submit,
101 		struct drm_msm_gem_submit *args, struct drm_file *file)
102 {
103 	unsigned i;
104 	int ret = 0;
105 
106 	for (i = 0; i < args->nr_bos; i++) {
107 		struct drm_msm_gem_submit_bo submit_bo;
108 		void __user *userptr =
109 			u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
110 
111 		/* make sure we don't have garbage flags, in case we hit
112 		 * error path before flags is initialized:
113 		 */
114 		submit->bos[i].flags = 0;
115 
116 		if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
117 			ret = -EFAULT;
118 			i = 0;
119 			goto out;
120 		}
121 
122 /* at least one of READ and/or WRITE flags should be set: */
123 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
124 
125 		if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
126 			!(submit_bo.flags & MANDATORY_FLAGS)) {
127 			DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
128 			ret = -EINVAL;
129 			i = 0;
130 			goto out;
131 		}
132 
133 		submit->bos[i].handle = submit_bo.handle;
134 		submit->bos[i].flags = submit_bo.flags;
135 		/* in validate_objects() we figure out if this is true: */
136 		submit->bos[i].iova  = submit_bo.presumed;
137 	}
138 
139 	spin_lock(&file->table_lock);
140 
141 	for (i = 0; i < args->nr_bos; i++) {
142 		struct drm_gem_object *obj;
143 
144 		/* normally use drm_gem_object_lookup(), but for bulk lookup
145 		 * all under single table_lock just hit object_idr directly:
146 		 */
147 		obj = idr_find(&file->object_idr, submit->bos[i].handle);
148 		if (!obj) {
149 			DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
150 			ret = -EINVAL;
151 			goto out_unlock;
152 		}
153 
154 		drm_gem_object_get(obj);
155 
156 		submit->bos[i].obj = to_msm_bo(obj);
157 	}
158 
159 out_unlock:
160 	spin_unlock(&file->table_lock);
161 
162 out:
163 	submit->nr_bos = i;
164 
165 	return ret;
166 }
167 
168 static int submit_lookup_cmds(struct msm_gem_submit *submit,
169 		struct drm_msm_gem_submit *args, struct drm_file *file)
170 {
171 	unsigned i;
172 	size_t sz;
173 	int ret = 0;
174 
175 	for (i = 0; i < args->nr_cmds; i++) {
176 		struct drm_msm_gem_submit_cmd submit_cmd;
177 		void __user *userptr =
178 			u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
179 
180 		ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
181 		if (ret) {
182 			ret = -EFAULT;
183 			goto out;
184 		}
185 
186 		/* validate input from userspace: */
187 		switch (submit_cmd.type) {
188 		case MSM_SUBMIT_CMD_BUF:
189 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
190 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
191 			break;
192 		default:
193 			DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
194 			return -EINVAL;
195 		}
196 
197 		if (submit_cmd.size % 4) {
198 			DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
199 					submit_cmd.size);
200 			ret = -EINVAL;
201 			goto out;
202 		}
203 
204 		submit->cmd[i].type = submit_cmd.type;
205 		submit->cmd[i].size = submit_cmd.size / 4;
206 		submit->cmd[i].offset = submit_cmd.submit_offset / 4;
207 		submit->cmd[i].idx  = submit_cmd.submit_idx;
208 		submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
209 
210 		userptr = u64_to_user_ptr(submit_cmd.relocs);
211 
212 		sz = array_size(submit_cmd.nr_relocs,
213 				sizeof(struct drm_msm_gem_submit_reloc));
214 		/* check for overflow: */
215 		if (sz == SIZE_MAX) {
216 			ret = -ENOMEM;
217 			goto out;
218 		}
219 		submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL);
220 		if (!submit->cmd[i].relocs) {
221 			ret = -ENOMEM;
222 			goto out;
223 		}
224 		ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
225 		if (ret) {
226 			ret = -EFAULT;
227 			goto out;
228 		}
229 	}
230 
231 out:
232 	return ret;
233 }
234 
235 /* Unwind bo state, according to cleanup_flags.  In the success case, only
236  * the lock is dropped at the end of the submit (and active/pin ref is dropped
237  * later when the submit is retired).
238  */
239 static void submit_cleanup_bo(struct msm_gem_submit *submit, int i,
240 		unsigned cleanup_flags)
241 {
242 	struct drm_gem_object *obj = &submit->bos[i].obj->base;
243 	unsigned flags = submit->bos[i].flags & cleanup_flags;
244 
245 	/*
246 	 * Clear flags bit before dropping lock, so that the msm_job_run()
247 	 * path isn't racing with submit_cleanup() (ie. the read/modify/
248 	 * write is protected by the obj lock in all paths)
249 	 */
250 	submit->bos[i].flags &= ~cleanup_flags;
251 
252 	if (flags & BO_VMA_PINNED)
253 		msm_gem_vma_unpin(submit->bos[i].vma);
254 
255 	if (flags & BO_OBJ_PINNED)
256 		msm_gem_unpin_locked(obj);
257 
258 	if (flags & BO_LOCKED)
259 		dma_resv_unlock(obj->resv);
260 }
261 
262 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
263 {
264 	unsigned cleanup_flags = BO_VMA_PINNED | BO_OBJ_PINNED | BO_LOCKED;
265 	submit_cleanup_bo(submit, i, cleanup_flags);
266 
267 	if (!(submit->bos[i].flags & BO_VALID))
268 		submit->bos[i].iova = 0;
269 }
270 
271 /* This is where we make sure all the bo's are reserved and pin'd: */
272 static int submit_lock_objects(struct msm_gem_submit *submit)
273 {
274 	int contended, slow_locked = -1, i, ret = 0;
275 
276 retry:
277 	for (i = 0; i < submit->nr_bos; i++) {
278 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
279 
280 		if (slow_locked == i)
281 			slow_locked = -1;
282 
283 		contended = i;
284 
285 		if (!(submit->bos[i].flags & BO_LOCKED)) {
286 			ret = dma_resv_lock_interruptible(msm_obj->base.resv,
287 							  &submit->ticket);
288 			if (ret)
289 				goto fail;
290 			submit->bos[i].flags |= BO_LOCKED;
291 		}
292 	}
293 
294 	ww_acquire_done(&submit->ticket);
295 
296 	return 0;
297 
298 fail:
299 	if (ret == -EALREADY) {
300 		DRM_ERROR("handle %u at index %u already on submit list\n",
301 				submit->bos[i].handle, i);
302 		ret = -EINVAL;
303 	}
304 
305 	for (; i >= 0; i--)
306 		submit_unlock_unpin_bo(submit, i);
307 
308 	if (slow_locked > 0)
309 		submit_unlock_unpin_bo(submit, slow_locked);
310 
311 	if (ret == -EDEADLK) {
312 		struct msm_gem_object *msm_obj = submit->bos[contended].obj;
313 		/* we lost out in a seqno race, lock and retry.. */
314 		ret = dma_resv_lock_slow_interruptible(msm_obj->base.resv,
315 						       &submit->ticket);
316 		if (!ret) {
317 			submit->bos[contended].flags |= BO_LOCKED;
318 			slow_locked = contended;
319 			goto retry;
320 		}
321 
322 		/* Not expecting -EALREADY here, if the bo was already
323 		 * locked, we should have gotten -EALREADY already from
324 		 * the dma_resv_lock_interruptable() call.
325 		 */
326 		WARN_ON_ONCE(ret == -EALREADY);
327 	}
328 
329 	return ret;
330 }
331 
332 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
333 {
334 	int i, ret = 0;
335 
336 	for (i = 0; i < submit->nr_bos; i++) {
337 		struct drm_gem_object *obj = &submit->bos[i].obj->base;
338 		bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
339 
340 		/* NOTE: _reserve_shared() must happen before
341 		 * _add_shared_fence(), which makes this a slightly
342 		 * strange place to call it.  OTOH this is a
343 		 * convenient can-fail point to hook it in.
344 		 */
345 		ret = dma_resv_reserve_fences(obj->resv, 1);
346 		if (ret)
347 			return ret;
348 
349 		/* If userspace has determined that explicit fencing is
350 		 * used, it can disable implicit sync on the entire
351 		 * submit:
352 		 */
353 		if (no_implicit)
354 			continue;
355 
356 		/* Otherwise userspace can ask for implicit sync to be
357 		 * disabled on specific buffers.  This is useful for internal
358 		 * usermode driver managed buffers, suballocation, etc.
359 		 */
360 		if (submit->bos[i].flags & MSM_SUBMIT_BO_NO_IMPLICIT)
361 			continue;
362 
363 		ret = drm_sched_job_add_implicit_dependencies(&submit->base,
364 							      obj,
365 							      write);
366 		if (ret)
367 			break;
368 	}
369 
370 	return ret;
371 }
372 
373 static int submit_pin_objects(struct msm_gem_submit *submit)
374 {
375 	int i, ret = 0;
376 
377 	submit->valid = true;
378 
379 	for (i = 0; i < submit->nr_bos; i++) {
380 		struct drm_gem_object *obj = &submit->bos[i].obj->base;
381 		struct msm_gem_vma *vma;
382 
383 		/* if locking succeeded, pin bo: */
384 		vma = msm_gem_get_vma_locked(obj, submit->aspace);
385 		if (IS_ERR(vma)) {
386 			ret = PTR_ERR(vma);
387 			break;
388 		}
389 
390 		ret = msm_gem_pin_vma_locked(obj, vma);
391 		if (ret)
392 			break;
393 
394 		submit->bos[i].flags |= BO_OBJ_PINNED | BO_VMA_PINNED;
395 		submit->bos[i].vma = vma;
396 
397 		if (vma->iova == submit->bos[i].iova) {
398 			submit->bos[i].flags |= BO_VALID;
399 		} else {
400 			submit->bos[i].iova = vma->iova;
401 			/* iova changed, so address in cmdstream is not valid: */
402 			submit->bos[i].flags &= ~BO_VALID;
403 			submit->valid = false;
404 		}
405 	}
406 
407 	return ret;
408 }
409 
410 static void submit_attach_object_fences(struct msm_gem_submit *submit)
411 {
412 	int i;
413 
414 	for (i = 0; i < submit->nr_bos; i++) {
415 		struct drm_gem_object *obj = &submit->bos[i].obj->base;
416 
417 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
418 			dma_resv_add_fence(obj->resv, submit->user_fence,
419 					   DMA_RESV_USAGE_WRITE);
420 		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
421 			dma_resv_add_fence(obj->resv, submit->user_fence,
422 					   DMA_RESV_USAGE_READ);
423 	}
424 }
425 
426 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
427 		struct msm_gem_object **obj, uint64_t *iova, bool *valid)
428 {
429 	if (idx >= submit->nr_bos) {
430 		DRM_ERROR("invalid buffer index: %u (out of %u)\n",
431 				idx, submit->nr_bos);
432 		return -EINVAL;
433 	}
434 
435 	if (obj)
436 		*obj = submit->bos[idx].obj;
437 	if (iova)
438 		*iova = submit->bos[idx].iova;
439 	if (valid)
440 		*valid = !!(submit->bos[idx].flags & BO_VALID);
441 
442 	return 0;
443 }
444 
445 /* process the reloc's and patch up the cmdstream as needed: */
446 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
447 		uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs)
448 {
449 	uint32_t i, last_offset = 0;
450 	uint32_t *ptr;
451 	int ret = 0;
452 
453 	if (!nr_relocs)
454 		return 0;
455 
456 	if (offset % 4) {
457 		DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
458 		return -EINVAL;
459 	}
460 
461 	/* For now, just map the entire thing.  Eventually we probably
462 	 * to do it page-by-page, w/ kmap() if not vmap()d..
463 	 */
464 	ptr = msm_gem_get_vaddr_locked(&obj->base);
465 
466 	if (IS_ERR(ptr)) {
467 		ret = PTR_ERR(ptr);
468 		DBG("failed to map: %d", ret);
469 		return ret;
470 	}
471 
472 	for (i = 0; i < nr_relocs; i++) {
473 		struct drm_msm_gem_submit_reloc submit_reloc = relocs[i];
474 		uint32_t off;
475 		uint64_t iova;
476 		bool valid;
477 
478 		if (submit_reloc.submit_offset % 4) {
479 			DRM_ERROR("non-aligned reloc offset: %u\n",
480 					submit_reloc.submit_offset);
481 			ret = -EINVAL;
482 			goto out;
483 		}
484 
485 		/* offset in dwords: */
486 		off = submit_reloc.submit_offset / 4;
487 
488 		if ((off >= (obj->base.size / 4)) ||
489 				(off < last_offset)) {
490 			DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
491 			ret = -EINVAL;
492 			goto out;
493 		}
494 
495 		ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
496 		if (ret)
497 			goto out;
498 
499 		if (valid)
500 			continue;
501 
502 		iova += submit_reloc.reloc_offset;
503 
504 		if (submit_reloc.shift < 0)
505 			iova >>= -submit_reloc.shift;
506 		else
507 			iova <<= submit_reloc.shift;
508 
509 		ptr[off] = iova | submit_reloc.or;
510 
511 		last_offset = off;
512 	}
513 
514 out:
515 	msm_gem_put_vaddr_locked(&obj->base);
516 
517 	return ret;
518 }
519 
520 /* Cleanup submit at end of ioctl.  In the error case, this also drops
521  * references, unpins, and drops active refcnt.  In the non-error case,
522  * this is done when the submit is retired.
523  */
524 static void submit_cleanup(struct msm_gem_submit *submit, bool error)
525 {
526 	unsigned cleanup_flags = BO_LOCKED;
527 	unsigned i;
528 
529 	if (error)
530 		cleanup_flags |= BO_VMA_PINNED | BO_OBJ_PINNED;
531 
532 	for (i = 0; i < submit->nr_bos; i++) {
533 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
534 		submit_cleanup_bo(submit, i, cleanup_flags);
535 		if (error)
536 			drm_gem_object_put(&msm_obj->base);
537 	}
538 }
539 
540 void msm_submit_retire(struct msm_gem_submit *submit)
541 {
542 	int i;
543 
544 	for (i = 0; i < submit->nr_bos; i++) {
545 		struct drm_gem_object *obj = &submit->bos[i].obj->base;
546 
547 		drm_gem_object_put(obj);
548 	}
549 }
550 
551 struct msm_submit_post_dep {
552 	struct drm_syncobj *syncobj;
553 	uint64_t point;
554 	struct dma_fence_chain *chain;
555 };
556 
557 static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
558                                            struct drm_file *file,
559                                            uint64_t in_syncobjs_addr,
560                                            uint32_t nr_in_syncobjs,
561                                            size_t syncobj_stride)
562 {
563 	struct drm_syncobj **syncobjs = NULL;
564 	struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
565 	int ret = 0;
566 	uint32_t i, j;
567 
568 	syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
569 	                   GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
570 	if (!syncobjs)
571 		return ERR_PTR(-ENOMEM);
572 
573 	for (i = 0; i < nr_in_syncobjs; ++i) {
574 		uint64_t address = in_syncobjs_addr + i * syncobj_stride;
575 
576 		if (copy_from_user(&syncobj_desc,
577 			           u64_to_user_ptr(address),
578 			           min(syncobj_stride, sizeof(syncobj_desc)))) {
579 			ret = -EFAULT;
580 			break;
581 		}
582 
583 		if (syncobj_desc.point &&
584 		    !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
585 			ret = -EOPNOTSUPP;
586 			break;
587 		}
588 
589 		if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
590 			ret = -EINVAL;
591 			break;
592 		}
593 
594 		ret = drm_sched_job_add_syncobj_dependency(&submit->base, file,
595 							   syncobj_desc.handle, syncobj_desc.point);
596 		if (ret)
597 			break;
598 
599 		if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
600 			syncobjs[i] =
601 				drm_syncobj_find(file, syncobj_desc.handle);
602 			if (!syncobjs[i]) {
603 				ret = -EINVAL;
604 				break;
605 			}
606 		}
607 	}
608 
609 	if (ret) {
610 		for (j = 0; j <= i; ++j) {
611 			if (syncobjs[j])
612 				drm_syncobj_put(syncobjs[j]);
613 		}
614 		kfree(syncobjs);
615 		return ERR_PTR(ret);
616 	}
617 	return syncobjs;
618 }
619 
620 static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
621                                uint32_t nr_syncobjs)
622 {
623 	uint32_t i;
624 
625 	for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
626 		if (syncobjs[i])
627 			drm_syncobj_replace_fence(syncobjs[i], NULL);
628 	}
629 }
630 
631 static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
632                                                        struct drm_file *file,
633                                                        uint64_t syncobjs_addr,
634                                                        uint32_t nr_syncobjs,
635                                                        size_t syncobj_stride)
636 {
637 	struct msm_submit_post_dep *post_deps;
638 	struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
639 	int ret = 0;
640 	uint32_t i, j;
641 
642 	post_deps = kcalloc(nr_syncobjs, sizeof(*post_deps),
643 			    GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
644 	if (!post_deps)
645 		return ERR_PTR(-ENOMEM);
646 
647 	for (i = 0; i < nr_syncobjs; ++i) {
648 		uint64_t address = syncobjs_addr + i * syncobj_stride;
649 
650 		if (copy_from_user(&syncobj_desc,
651 			           u64_to_user_ptr(address),
652 			           min(syncobj_stride, sizeof(syncobj_desc)))) {
653 			ret = -EFAULT;
654 			break;
655 		}
656 
657 		post_deps[i].point = syncobj_desc.point;
658 
659 		if (syncobj_desc.flags) {
660 			ret = -EINVAL;
661 			break;
662 		}
663 
664 		if (syncobj_desc.point) {
665 			if (!drm_core_check_feature(dev,
666 			                            DRIVER_SYNCOBJ_TIMELINE)) {
667 				ret = -EOPNOTSUPP;
668 				break;
669 			}
670 
671 			post_deps[i].chain = dma_fence_chain_alloc();
672 			if (!post_deps[i].chain) {
673 				ret = -ENOMEM;
674 				break;
675 			}
676 		}
677 
678 		post_deps[i].syncobj =
679 			drm_syncobj_find(file, syncobj_desc.handle);
680 		if (!post_deps[i].syncobj) {
681 			ret = -EINVAL;
682 			break;
683 		}
684 	}
685 
686 	if (ret) {
687 		for (j = 0; j <= i; ++j) {
688 			dma_fence_chain_free(post_deps[j].chain);
689 			if (post_deps[j].syncobj)
690 				drm_syncobj_put(post_deps[j].syncobj);
691 		}
692 
693 		kfree(post_deps);
694 		return ERR_PTR(ret);
695 	}
696 
697 	return post_deps;
698 }
699 
700 static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
701                                   uint32_t count, struct dma_fence *fence)
702 {
703 	uint32_t i;
704 
705 	for (i = 0; post_deps && i < count; ++i) {
706 		if (post_deps[i].chain) {
707 			drm_syncobj_add_point(post_deps[i].syncobj,
708 			                      post_deps[i].chain,
709 			                      fence, post_deps[i].point);
710 			post_deps[i].chain = NULL;
711 		} else {
712 			drm_syncobj_replace_fence(post_deps[i].syncobj,
713 			                          fence);
714 		}
715 	}
716 }
717 
718 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
719 		struct drm_file *file)
720 {
721 	struct msm_drm_private *priv = dev->dev_private;
722 	struct drm_msm_gem_submit *args = data;
723 	struct msm_file_private *ctx = file->driver_priv;
724 	struct msm_gem_submit *submit = NULL;
725 	struct msm_gpu *gpu = priv->gpu;
726 	struct msm_gpu_submitqueue *queue;
727 	struct msm_ringbuffer *ring;
728 	struct msm_submit_post_dep *post_deps = NULL;
729 	struct drm_syncobj **syncobjs_to_reset = NULL;
730 	int out_fence_fd = -1;
731 	bool has_ww_ticket = false;
732 	unsigned i;
733 	int ret;
734 
735 	if (!gpu)
736 		return -ENXIO;
737 
738 	if (args->pad)
739 		return -EINVAL;
740 
741 	if (unlikely(!ctx->aspace) && !capable(CAP_SYS_RAWIO)) {
742 		DRM_ERROR_RATELIMITED("IOMMU support or CAP_SYS_RAWIO required!\n");
743 		return -EPERM;
744 	}
745 
746 	/* for now, we just have 3d pipe.. eventually this would need to
747 	 * be more clever to dispatch to appropriate gpu module:
748 	 */
749 	if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
750 		return -EINVAL;
751 
752 	if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
753 		return -EINVAL;
754 
755 	if (args->flags & MSM_SUBMIT_SUDO) {
756 		if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
757 		    !capable(CAP_SYS_RAWIO))
758 			return -EINVAL;
759 	}
760 
761 	queue = msm_submitqueue_get(ctx, args->queueid);
762 	if (!queue)
763 		return -ENOENT;
764 
765 	ring = gpu->rb[queue->ring_nr];
766 
767 	if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
768 		out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
769 		if (out_fence_fd < 0) {
770 			ret = out_fence_fd;
771 			goto out_post_unlock;
772 		}
773 	}
774 
775 	submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds);
776 	if (IS_ERR(submit)) {
777 		ret = PTR_ERR(submit);
778 		goto out_post_unlock;
779 	}
780 
781 	trace_msm_gpu_submit(pid_nr(submit->pid), ring->id, submit->ident,
782 		args->nr_bos, args->nr_cmds);
783 
784 	ret = mutex_lock_interruptible(&queue->lock);
785 	if (ret)
786 		goto out_post_unlock;
787 
788 	if (args->flags & MSM_SUBMIT_SUDO)
789 		submit->in_rb = true;
790 
791 	if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
792 		struct dma_fence *in_fence;
793 
794 		in_fence = sync_file_get_fence(args->fence_fd);
795 
796 		if (!in_fence) {
797 			ret = -EINVAL;
798 			goto out_unlock;
799 		}
800 
801 		ret = drm_sched_job_add_dependency(&submit->base, in_fence);
802 		if (ret)
803 			goto out_unlock;
804 	}
805 
806 	if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
807 		syncobjs_to_reset = msm_parse_deps(submit, file,
808 		                                   args->in_syncobjs,
809 		                                   args->nr_in_syncobjs,
810 		                                   args->syncobj_stride);
811 		if (IS_ERR(syncobjs_to_reset)) {
812 			ret = PTR_ERR(syncobjs_to_reset);
813 			goto out_unlock;
814 		}
815 	}
816 
817 	if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
818 		post_deps = msm_parse_post_deps(dev, file,
819 		                                args->out_syncobjs,
820 		                                args->nr_out_syncobjs,
821 		                                args->syncobj_stride);
822 		if (IS_ERR(post_deps)) {
823 			ret = PTR_ERR(post_deps);
824 			goto out_unlock;
825 		}
826 	}
827 
828 	ret = submit_lookup_objects(submit, args, file);
829 	if (ret)
830 		goto out;
831 
832 	ret = submit_lookup_cmds(submit, args, file);
833 	if (ret)
834 		goto out;
835 
836 	/* copy_*_user while holding a ww ticket upsets lockdep */
837 	ww_acquire_init(&submit->ticket, &reservation_ww_class);
838 	has_ww_ticket = true;
839 	ret = submit_lock_objects(submit);
840 	if (ret)
841 		goto out;
842 
843 	ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
844 	if (ret)
845 		goto out;
846 
847 	ret = submit_pin_objects(submit);
848 	if (ret)
849 		goto out;
850 
851 	for (i = 0; i < args->nr_cmds; i++) {
852 		struct msm_gem_object *msm_obj;
853 		uint64_t iova;
854 
855 		ret = submit_bo(submit, submit->cmd[i].idx,
856 				&msm_obj, &iova, NULL);
857 		if (ret)
858 			goto out;
859 
860 		if (!submit->cmd[i].size ||
861 			((submit->cmd[i].size + submit->cmd[i].offset) >
862 				msm_obj->base.size / 4)) {
863 			DRM_ERROR("invalid cmdstream size: %u\n", submit->cmd[i].size * 4);
864 			ret = -EINVAL;
865 			goto out;
866 		}
867 
868 		submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
869 
870 		if (submit->valid)
871 			continue;
872 
873 		ret = submit_reloc(submit, msm_obj, submit->cmd[i].offset * 4,
874 				submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
875 		if (ret)
876 			goto out;
877 	}
878 
879 	submit->nr_cmds = i;
880 
881 	idr_preload(GFP_KERNEL);
882 
883 	spin_lock(&queue->idr_lock);
884 
885 	/*
886 	 * If using userspace provided seqno fence, validate that the id
887 	 * is available before arming sched job.  Since access to fence_idr
888 	 * is serialized on the queue lock, the slot should be still avail
889 	 * after the job is armed
890 	 */
891 	if ((args->flags & MSM_SUBMIT_FENCE_SN_IN) &&
892 			idr_find(&queue->fence_idr, args->fence)) {
893 		spin_unlock(&queue->idr_lock);
894 		idr_preload_end();
895 		ret = -EINVAL;
896 		goto out;
897 	}
898 
899 	drm_sched_job_arm(&submit->base);
900 
901 	submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
902 
903 	if (args->flags & MSM_SUBMIT_FENCE_SN_IN) {
904 		/*
905 		 * Userspace has assigned the seqno fence that it wants
906 		 * us to use.  It is an error to pick a fence sequence
907 		 * number that is not available.
908 		 */
909 		submit->fence_id = args->fence;
910 		ret = idr_alloc_u32(&queue->fence_idr, submit->user_fence,
911 				    &submit->fence_id, submit->fence_id,
912 				    GFP_NOWAIT);
913 		/*
914 		 * We've already validated that the fence_id slot is valid,
915 		 * so if idr_alloc_u32 failed, it is a kernel bug
916 		 */
917 		WARN_ON(ret);
918 	} else {
919 		/*
920 		 * Allocate an id which can be used by WAIT_FENCE ioctl to map
921 		 * back to the underlying fence.
922 		 */
923 		submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
924 						    submit->user_fence, 1,
925 						    INT_MAX, GFP_NOWAIT);
926 	}
927 
928 	spin_unlock(&queue->idr_lock);
929 	idr_preload_end();
930 
931 	if (submit->fence_id < 0) {
932 		ret = submit->fence_id;
933 		submit->fence_id = 0;
934 	}
935 
936 	if (ret == 0 && args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
937 		struct sync_file *sync_file = sync_file_create(submit->user_fence);
938 		if (!sync_file) {
939 			ret = -ENOMEM;
940 		} else {
941 			fd_install(out_fence_fd, sync_file->file);
942 			args->fence_fd = out_fence_fd;
943 		}
944 	}
945 
946 	submit_attach_object_fences(submit);
947 
948 	/* The scheduler owns a ref now: */
949 	msm_gem_submit_get(submit);
950 
951 	msm_rd_dump_submit(priv->rd, submit, NULL);
952 
953 	drm_sched_entity_push_job(&submit->base);
954 
955 	args->fence = submit->fence_id;
956 	queue->last_fence = submit->fence_id;
957 
958 	msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
959 	msm_process_post_deps(post_deps, args->nr_out_syncobjs,
960 	                      submit->user_fence);
961 
962 
963 out:
964 	submit_cleanup(submit, !!ret);
965 	if (has_ww_ticket)
966 		ww_acquire_fini(&submit->ticket);
967 out_unlock:
968 	mutex_unlock(&queue->lock);
969 out_post_unlock:
970 	if (ret && (out_fence_fd >= 0))
971 		put_unused_fd(out_fence_fd);
972 
973 	if (!IS_ERR_OR_NULL(submit)) {
974 		msm_gem_submit_put(submit);
975 	} else {
976 		/*
977 		 * If the submit hasn't yet taken ownership of the queue
978 		 * then we need to drop the reference ourself:
979 		 */
980 		msm_submitqueue_put(queue);
981 	}
982 	if (!IS_ERR_OR_NULL(post_deps)) {
983 		for (i = 0; i < args->nr_out_syncobjs; ++i) {
984 			kfree(post_deps[i].chain);
985 			drm_syncobj_put(post_deps[i].syncobj);
986 		}
987 		kfree(post_deps);
988 	}
989 
990 	if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
991 		for (i = 0; i < args->nr_in_syncobjs; ++i) {
992 			if (syncobjs_to_reset[i])
993 				drm_syncobj_put(syncobjs_to_reset[i]);
994 		}
995 		kfree(syncobjs_to_reset);
996 	}
997 
998 	return ret;
999 }
1000