xref: /linux/drivers/gpu/drm/msm/msm_gem_submit.c (revision 2c1ed907520c50326b8f604907a8478b27881a2e)
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 /* For userspace errors, use DRM_UT_DRIVER.. so that userspace can enable
21  * error msgs for debugging, but we don't spam dmesg by default
22  */
23 #define SUBMIT_ERROR(err, submit, fmt, ...) \
24 	UERR(err, (submit)->dev, fmt, ##__VA_ARGS__)
25 
26 /*
27  * Cmdstream submission:
28  */
29 
submit_create(struct drm_device * dev,struct msm_gpu * gpu,struct msm_gpu_submitqueue * queue,uint32_t nr_bos,uint32_t nr_cmds)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 	static atomic_t ident = ATOMIC_INIT(0);
36 	struct msm_gem_submit *submit;
37 	uint64_t sz;
38 	int ret;
39 
40 	sz = struct_size(submit, bos, nr_bos) +
41 			((u64)nr_cmds * sizeof(submit->cmd[0]));
42 
43 	if (sz > SIZE_MAX)
44 		return ERR_PTR(-ENOMEM);
45 
46 	submit = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
47 	if (!submit)
48 		return ERR_PTR(-ENOMEM);
49 
50 	submit->hw_fence = msm_fence_alloc();
51 	if (IS_ERR(submit->hw_fence)) {
52 		ret = PTR_ERR(submit->hw_fence);
53 		kfree(submit);
54 		return ERR_PTR(ret);
55 	}
56 
57 	ret = drm_sched_job_init(&submit->base, queue->entity, 1, queue);
58 	if (ret) {
59 		kfree(submit->hw_fence);
60 		kfree(submit);
61 		return ERR_PTR(ret);
62 	}
63 
64 	kref_init(&submit->ref);
65 	submit->dev = dev;
66 	submit->aspace = queue->ctx->aspace;
67 	submit->gpu = gpu;
68 	submit->cmd = (void *)&submit->bos[nr_bos];
69 	submit->queue = queue;
70 	submit->pid = get_pid(task_pid(current));
71 	submit->ring = gpu->rb[queue->ring_nr];
72 	submit->fault_dumped = false;
73 
74 	/* Get a unique identifier for the submission for logging purposes */
75 	submit->ident = atomic_inc_return(&ident) - 1;
76 
77 	INIT_LIST_HEAD(&submit->node);
78 
79 	return submit;
80 }
81 
__msm_gem_submit_destroy(struct kref * kref)82 void __msm_gem_submit_destroy(struct kref *kref)
83 {
84 	struct msm_gem_submit *submit =
85 			container_of(kref, struct msm_gem_submit, ref);
86 	unsigned i;
87 
88 	if (submit->fence_id) {
89 		spin_lock(&submit->queue->idr_lock);
90 		idr_remove(&submit->queue->fence_idr, submit->fence_id);
91 		spin_unlock(&submit->queue->idr_lock);
92 	}
93 
94 	dma_fence_put(submit->user_fence);
95 
96 	/*
97 	 * If the submit is freed before msm_job_run(), then hw_fence is
98 	 * just some pre-allocated memory, not a reference counted fence.
99 	 * Once the job runs and the hw_fence is initialized, it will
100 	 * have a refcount of at least one, since the submit holds a ref
101 	 * to the hw_fence.
102 	 */
103 	if (kref_read(&submit->hw_fence->refcount) == 0) {
104 		kfree(submit->hw_fence);
105 	} else {
106 		dma_fence_put(submit->hw_fence);
107 	}
108 
109 	put_pid(submit->pid);
110 	msm_submitqueue_put(submit->queue);
111 
112 	for (i = 0; i < submit->nr_cmds; i++)
113 		kfree(submit->cmd[i].relocs);
114 
115 	kfree(submit);
116 }
117 
submit_lookup_objects(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)118 static int submit_lookup_objects(struct msm_gem_submit *submit,
119 		struct drm_msm_gem_submit *args, struct drm_file *file)
120 {
121 	unsigned i;
122 	int ret = 0;
123 
124 	for (i = 0; i < args->nr_bos; i++) {
125 		struct drm_msm_gem_submit_bo submit_bo;
126 		void __user *userptr =
127 			u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
128 
129 		/* make sure we don't have garbage flags, in case we hit
130 		 * error path before flags is initialized:
131 		 */
132 		submit->bos[i].flags = 0;
133 
134 		if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
135 			ret = -EFAULT;
136 			i = 0;
137 			goto out;
138 		}
139 
140 /* at least one of READ and/or WRITE flags should be set: */
141 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
142 
143 		if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
144 			!(submit_bo.flags & MANDATORY_FLAGS)) {
145 			ret = SUBMIT_ERROR(EINVAL, submit, "invalid flags: %x\n", submit_bo.flags);
146 			i = 0;
147 			goto out;
148 		}
149 
150 		submit->bos[i].handle = submit_bo.handle;
151 		submit->bos[i].flags = submit_bo.flags;
152 	}
153 
154 	spin_lock(&file->table_lock);
155 
156 	for (i = 0; i < args->nr_bos; i++) {
157 		struct drm_gem_object *obj;
158 
159 		/* normally use drm_gem_object_lookup(), but for bulk lookup
160 		 * all under single table_lock just hit object_idr directly:
161 		 */
162 		obj = idr_find(&file->object_idr, submit->bos[i].handle);
163 		if (!obj) {
164 			ret = SUBMIT_ERROR(EINVAL, submit, "invalid handle %u at index %u\n", submit->bos[i].handle, i);
165 			goto out_unlock;
166 		}
167 
168 		drm_gem_object_get(obj);
169 
170 		submit->bos[i].obj = obj;
171 	}
172 
173 out_unlock:
174 	spin_unlock(&file->table_lock);
175 
176 out:
177 	submit->nr_bos = i;
178 
179 	return ret;
180 }
181 
submit_lookup_cmds(struct msm_gem_submit * submit,struct drm_msm_gem_submit * args,struct drm_file * file)182 static int submit_lookup_cmds(struct msm_gem_submit *submit,
183 		struct drm_msm_gem_submit *args, struct drm_file *file)
184 {
185 	unsigned i;
186 	size_t sz;
187 	int ret = 0;
188 
189 	for (i = 0; i < args->nr_cmds; i++) {
190 		struct drm_msm_gem_submit_cmd submit_cmd;
191 		void __user *userptr =
192 			u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
193 
194 		ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
195 		if (ret) {
196 			ret = -EFAULT;
197 			goto out;
198 		}
199 
200 		/* validate input from userspace: */
201 		switch (submit_cmd.type) {
202 		case MSM_SUBMIT_CMD_BUF:
203 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
204 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
205 			break;
206 		default:
207 			return SUBMIT_ERROR(EINVAL, submit, "invalid type: %08x\n", submit_cmd.type);
208 		}
209 
210 		if (submit_cmd.size % 4) {
211 			ret = SUBMIT_ERROR(EINVAL, submit, "non-aligned cmdstream buffer size: %u\n",
212 					   submit_cmd.size);
213 			goto out;
214 		}
215 
216 		submit->cmd[i].type = submit_cmd.type;
217 		submit->cmd[i].size = submit_cmd.size / 4;
218 		submit->cmd[i].offset = submit_cmd.submit_offset / 4;
219 		submit->cmd[i].idx  = submit_cmd.submit_idx;
220 		submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
221 
222 		userptr = u64_to_user_ptr(submit_cmd.relocs);
223 
224 		sz = array_size(submit_cmd.nr_relocs,
225 				sizeof(struct drm_msm_gem_submit_reloc));
226 		/* check for overflow: */
227 		if (sz == SIZE_MAX) {
228 			ret = -ENOMEM;
229 			goto out;
230 		}
231 		submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN);
232 		if (!submit->cmd[i].relocs) {
233 			ret = -ENOMEM;
234 			goto out;
235 		}
236 		ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
237 		if (ret) {
238 			ret = -EFAULT;
239 			goto out;
240 		}
241 	}
242 
243 out:
244 	return ret;
245 }
246 
247 /* This is where we make sure all the bo's are reserved and pin'd: */
submit_lock_objects(struct msm_gem_submit * submit)248 static int submit_lock_objects(struct msm_gem_submit *submit)
249 {
250 	int ret;
251 
252 	drm_exec_init(&submit->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, submit->nr_bos);
253 
254 	drm_exec_until_all_locked (&submit->exec) {
255 		for (unsigned i = 0; i < submit->nr_bos; i++) {
256 			struct drm_gem_object *obj = submit->bos[i].obj;
257 			ret = drm_exec_prepare_obj(&submit->exec, obj, 1);
258 			drm_exec_retry_on_contention(&submit->exec);
259 			if (ret)
260 				goto error;
261 		}
262 	}
263 
264 	return 0;
265 
266 error:
267 	return ret;
268 }
269 
submit_fence_sync(struct msm_gem_submit * submit)270 static int submit_fence_sync(struct msm_gem_submit *submit)
271 {
272 	int i, ret = 0;
273 
274 	for (i = 0; i < submit->nr_bos; i++) {
275 		struct drm_gem_object *obj = submit->bos[i].obj;
276 		bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
277 
278 		/* Otherwise userspace can ask for implicit sync to be
279 		 * disabled on specific buffers.  This is useful for internal
280 		 * usermode driver managed buffers, suballocation, etc.
281 		 */
282 		if (submit->bos[i].flags & MSM_SUBMIT_BO_NO_IMPLICIT)
283 			continue;
284 
285 		ret = drm_sched_job_add_implicit_dependencies(&submit->base,
286 							      obj,
287 							      write);
288 		if (ret)
289 			break;
290 	}
291 
292 	return ret;
293 }
294 
submit_pin_objects(struct msm_gem_submit * submit)295 static int submit_pin_objects(struct msm_gem_submit *submit)
296 {
297 	struct msm_drm_private *priv = submit->dev->dev_private;
298 	int i, ret = 0;
299 
300 	for (i = 0; i < submit->nr_bos; i++) {
301 		struct drm_gem_object *obj = submit->bos[i].obj;
302 		struct msm_gem_vma *vma;
303 
304 		/* if locking succeeded, pin bo: */
305 		vma = msm_gem_get_vma_locked(obj, submit->aspace);
306 		if (IS_ERR(vma)) {
307 			ret = PTR_ERR(vma);
308 			break;
309 		}
310 
311 		ret = msm_gem_pin_vma_locked(obj, vma);
312 		if (ret)
313 			break;
314 
315 		submit->bos[i].iova = vma->iova;
316 	}
317 
318 	/*
319 	 * A second loop while holding the LRU lock (a) avoids acquiring/dropping
320 	 * the LRU lock for each individual bo, while (b) avoiding holding the
321 	 * LRU lock while calling msm_gem_pin_vma_locked() (which could trigger
322 	 * get_pages() which could trigger reclaim.. and if we held the LRU lock
323 	 * could trigger deadlock with the shrinker).
324 	 */
325 	mutex_lock(&priv->lru.lock);
326 	for (i = 0; i < submit->nr_bos; i++) {
327 		msm_gem_pin_obj_locked(submit->bos[i].obj);
328 	}
329 	mutex_unlock(&priv->lru.lock);
330 
331 	submit->bos_pinned = true;
332 
333 	return ret;
334 }
335 
submit_unpin_objects(struct msm_gem_submit * submit)336 static void submit_unpin_objects(struct msm_gem_submit *submit)
337 {
338 	if (!submit->bos_pinned)
339 		return;
340 
341 	for (int i = 0; i < submit->nr_bos; i++) {
342 		struct drm_gem_object *obj = submit->bos[i].obj;
343 
344 		msm_gem_unpin_locked(obj);
345 	}
346 
347 	submit->bos_pinned = false;
348 }
349 
submit_attach_object_fences(struct msm_gem_submit * submit)350 static void submit_attach_object_fences(struct msm_gem_submit *submit)
351 {
352 	int i;
353 
354 	for (i = 0; i < submit->nr_bos; i++) {
355 		struct drm_gem_object *obj = submit->bos[i].obj;
356 
357 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
358 			dma_resv_add_fence(obj->resv, submit->user_fence,
359 					   DMA_RESV_USAGE_WRITE);
360 		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
361 			dma_resv_add_fence(obj->resv, submit->user_fence,
362 					   DMA_RESV_USAGE_READ);
363 	}
364 }
365 
submit_bo(struct msm_gem_submit * submit,uint32_t idx,struct drm_gem_object ** obj,uint64_t * iova)366 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
367 		struct drm_gem_object **obj, uint64_t *iova)
368 {
369 	if (idx >= submit->nr_bos) {
370 		return SUBMIT_ERROR(EINVAL, submit, "invalid buffer index: %u (out of %u)\n",
371 				    idx, submit->nr_bos);
372 	}
373 
374 	if (obj)
375 		*obj = submit->bos[idx].obj;
376 	if (iova)
377 		*iova = submit->bos[idx].iova;
378 
379 	return 0;
380 }
381 
382 /* process the reloc's and patch up the cmdstream as needed: */
submit_reloc(struct msm_gem_submit * submit,struct drm_gem_object * obj,uint32_t offset,uint32_t nr_relocs,struct drm_msm_gem_submit_reloc * relocs)383 static int submit_reloc(struct msm_gem_submit *submit, struct drm_gem_object *obj,
384 		uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs)
385 {
386 	uint32_t i, last_offset = 0;
387 	uint32_t *ptr;
388 	int ret = 0;
389 
390 	if (offset % 4)
391 		return SUBMIT_ERROR(EINVAL, submit, "non-aligned cmdstream buffer: %u\n", offset);
392 
393 	/* For now, just map the entire thing.  Eventually we probably
394 	 * to do it page-by-page, w/ kmap() if not vmap()d..
395 	 */
396 	ptr = msm_gem_get_vaddr_locked(obj);
397 
398 	if (IS_ERR(ptr)) {
399 		ret = PTR_ERR(ptr);
400 		DBG("failed to map: %d", ret);
401 		return ret;
402 	}
403 
404 	for (i = 0; i < nr_relocs; i++) {
405 		struct drm_msm_gem_submit_reloc submit_reloc = relocs[i];
406 		uint32_t off;
407 		uint64_t iova;
408 
409 		if (submit_reloc.submit_offset % 4) {
410 			ret = SUBMIT_ERROR(EINVAL, submit, "non-aligned reloc offset: %u\n",
411 					   submit_reloc.submit_offset);
412 			goto out;
413 		}
414 
415 		/* offset in dwords: */
416 		off = submit_reloc.submit_offset / 4;
417 
418 		if ((off >= (obj->size / 4)) ||
419 				(off < last_offset)) {
420 			ret = SUBMIT_ERROR(EINVAL, submit, "invalid offset %u at reloc %u\n", off, i);
421 			goto out;
422 		}
423 
424 		ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova);
425 		if (ret)
426 			goto out;
427 
428 		iova += submit_reloc.reloc_offset;
429 
430 		if (submit_reloc.shift < 0)
431 			iova >>= -submit_reloc.shift;
432 		else
433 			iova <<= submit_reloc.shift;
434 
435 		ptr[off] = iova | submit_reloc.or;
436 
437 		last_offset = off;
438 	}
439 
440 out:
441 	msm_gem_put_vaddr_locked(obj);
442 
443 	return ret;
444 }
445 
446 /* Cleanup submit at end of ioctl.  In the error case, this also drops
447  * references, unpins, and drops active refcnt.  In the non-error case,
448  * this is done when the submit is retired.
449  */
submit_cleanup(struct msm_gem_submit * submit,bool error)450 static void submit_cleanup(struct msm_gem_submit *submit, bool error)
451 {
452 	if (error) {
453 		submit_unpin_objects(submit);
454 		/* job wasn't enqueued to scheduler, so early retirement: */
455 		msm_submit_retire(submit);
456 	}
457 
458 	if (submit->exec.objects)
459 		drm_exec_fini(&submit->exec);
460 }
461 
msm_submit_retire(struct msm_gem_submit * submit)462 void msm_submit_retire(struct msm_gem_submit *submit)
463 {
464 	int i;
465 
466 	for (i = 0; i < submit->nr_bos; i++) {
467 		struct drm_gem_object *obj = submit->bos[i].obj;
468 
469 		drm_gem_object_put(obj);
470 	}
471 }
472 
473 struct msm_submit_post_dep {
474 	struct drm_syncobj *syncobj;
475 	uint64_t point;
476 	struct dma_fence_chain *chain;
477 };
478 
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)479 static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
480                                            struct drm_file *file,
481                                            uint64_t in_syncobjs_addr,
482                                            uint32_t nr_in_syncobjs,
483                                            size_t syncobj_stride)
484 {
485 	struct drm_syncobj **syncobjs = NULL;
486 	struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
487 	int ret = 0;
488 	uint32_t i, j;
489 
490 	syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
491 	                   GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
492 	if (!syncobjs)
493 		return ERR_PTR(-ENOMEM);
494 
495 	for (i = 0; i < nr_in_syncobjs; ++i) {
496 		uint64_t address = in_syncobjs_addr + i * syncobj_stride;
497 
498 		if (copy_from_user(&syncobj_desc,
499 			           u64_to_user_ptr(address),
500 			           min(syncobj_stride, sizeof(syncobj_desc)))) {
501 			ret = -EFAULT;
502 			break;
503 		}
504 
505 		if (syncobj_desc.point &&
506 		    !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
507 			ret = SUBMIT_ERROR(EOPNOTSUPP, submit, "syncobj timeline unsupported");
508 			break;
509 		}
510 
511 		if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
512 			ret = -SUBMIT_ERROR(EINVAL, submit, "invalid syncobj flags: %x", syncobj_desc.flags);
513 			break;
514 		}
515 
516 		ret = drm_sched_job_add_syncobj_dependency(&submit->base, file,
517 							   syncobj_desc.handle, syncobj_desc.point);
518 		if (ret)
519 			break;
520 
521 		if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
522 			syncobjs[i] =
523 				drm_syncobj_find(file, syncobj_desc.handle);
524 			if (!syncobjs[i]) {
525 				ret = SUBMIT_ERROR(EINVAL, submit, "invalid syncobj handle: %u", i);
526 				break;
527 			}
528 		}
529 	}
530 
531 	if (ret) {
532 		for (j = 0; j <= i; ++j) {
533 			if (syncobjs[j])
534 				drm_syncobj_put(syncobjs[j]);
535 		}
536 		kfree(syncobjs);
537 		return ERR_PTR(ret);
538 	}
539 	return syncobjs;
540 }
541 
msm_reset_syncobjs(struct drm_syncobj ** syncobjs,uint32_t nr_syncobjs)542 static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
543                                uint32_t nr_syncobjs)
544 {
545 	uint32_t i;
546 
547 	for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
548 		if (syncobjs[i])
549 			drm_syncobj_replace_fence(syncobjs[i], NULL);
550 	}
551 }
552 
msm_parse_post_deps(struct drm_device * dev,struct drm_file * file,uint64_t syncobjs_addr,uint32_t nr_syncobjs,size_t syncobj_stride)553 static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
554                                                        struct drm_file *file,
555                                                        uint64_t syncobjs_addr,
556                                                        uint32_t nr_syncobjs,
557                                                        size_t syncobj_stride)
558 {
559 	struct msm_submit_post_dep *post_deps;
560 	struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
561 	int ret = 0;
562 	uint32_t i, j;
563 
564 	post_deps = kcalloc(nr_syncobjs, sizeof(*post_deps),
565 			    GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
566 	if (!post_deps)
567 		return ERR_PTR(-ENOMEM);
568 
569 	for (i = 0; i < nr_syncobjs; ++i) {
570 		uint64_t address = syncobjs_addr + i * syncobj_stride;
571 
572 		if (copy_from_user(&syncobj_desc,
573 			           u64_to_user_ptr(address),
574 			           min(syncobj_stride, sizeof(syncobj_desc)))) {
575 			ret = -EFAULT;
576 			break;
577 		}
578 
579 		post_deps[i].point = syncobj_desc.point;
580 
581 		if (syncobj_desc.flags) {
582 			ret = UERR(EINVAL, dev, "invalid syncobj flags");
583 			break;
584 		}
585 
586 		if (syncobj_desc.point) {
587 			if (!drm_core_check_feature(dev,
588 			                            DRIVER_SYNCOBJ_TIMELINE)) {
589 				ret = UERR(EOPNOTSUPP, dev, "syncobj timeline unsupported");
590 				break;
591 			}
592 
593 			post_deps[i].chain = dma_fence_chain_alloc();
594 			if (!post_deps[i].chain) {
595 				ret = -ENOMEM;
596 				break;
597 			}
598 		}
599 
600 		post_deps[i].syncobj =
601 			drm_syncobj_find(file, syncobj_desc.handle);
602 		if (!post_deps[i].syncobj) {
603 			ret = UERR(EINVAL, dev, "invalid syncobj handle");
604 			break;
605 		}
606 	}
607 
608 	if (ret) {
609 		for (j = 0; j <= i; ++j) {
610 			dma_fence_chain_free(post_deps[j].chain);
611 			if (post_deps[j].syncobj)
612 				drm_syncobj_put(post_deps[j].syncobj);
613 		}
614 
615 		kfree(post_deps);
616 		return ERR_PTR(ret);
617 	}
618 
619 	return post_deps;
620 }
621 
msm_process_post_deps(struct msm_submit_post_dep * post_deps,uint32_t count,struct dma_fence * fence)622 static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
623                                   uint32_t count, struct dma_fence *fence)
624 {
625 	uint32_t i;
626 
627 	for (i = 0; post_deps && i < count; ++i) {
628 		if (post_deps[i].chain) {
629 			drm_syncobj_add_point(post_deps[i].syncobj,
630 			                      post_deps[i].chain,
631 			                      fence, post_deps[i].point);
632 			post_deps[i].chain = NULL;
633 		} else {
634 			drm_syncobj_replace_fence(post_deps[i].syncobj,
635 			                          fence);
636 		}
637 	}
638 }
639 
msm_ioctl_gem_submit(struct drm_device * dev,void * data,struct drm_file * file)640 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
641 		struct drm_file *file)
642 {
643 	struct msm_drm_private *priv = dev->dev_private;
644 	struct drm_msm_gem_submit *args = data;
645 	struct msm_file_private *ctx = file->driver_priv;
646 	struct msm_gem_submit *submit = NULL;
647 	struct msm_gpu *gpu = priv->gpu;
648 	struct msm_gpu_submitqueue *queue;
649 	struct msm_ringbuffer *ring;
650 	struct msm_submit_post_dep *post_deps = NULL;
651 	struct drm_syncobj **syncobjs_to_reset = NULL;
652 	int out_fence_fd = -1;
653 	unsigned i;
654 	int ret;
655 
656 	if (!gpu)
657 		return -ENXIO;
658 
659 	if (args->pad)
660 		return -EINVAL;
661 
662 	if (unlikely(!ctx->aspace) && !capable(CAP_SYS_RAWIO)) {
663 		DRM_ERROR_RATELIMITED("IOMMU support or CAP_SYS_RAWIO required!\n");
664 		return -EPERM;
665 	}
666 
667 	/* for now, we just have 3d pipe.. eventually this would need to
668 	 * be more clever to dispatch to appropriate gpu module:
669 	 */
670 	if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
671 		return UERR(EINVAL, dev, "invalid pipe");
672 
673 	if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
674 		return UERR(EINVAL, dev, "invalid flags");
675 
676 	if (args->flags & MSM_SUBMIT_SUDO) {
677 		if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
678 		    !capable(CAP_SYS_RAWIO))
679 			return -EINVAL;
680 	}
681 
682 	queue = msm_submitqueue_get(ctx, args->queueid);
683 	if (!queue)
684 		return -ENOENT;
685 
686 	ring = gpu->rb[queue->ring_nr];
687 
688 	if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
689 		out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
690 		if (out_fence_fd < 0) {
691 			ret = out_fence_fd;
692 			goto out_post_unlock;
693 		}
694 	}
695 
696 	submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds);
697 	if (IS_ERR(submit)) {
698 		ret = PTR_ERR(submit);
699 		goto out_post_unlock;
700 	}
701 
702 	trace_msm_gpu_submit(pid_nr(submit->pid), ring->id, submit->ident,
703 		args->nr_bos, args->nr_cmds);
704 
705 	ret = mutex_lock_interruptible(&queue->lock);
706 	if (ret)
707 		goto out_post_unlock;
708 
709 	if (args->flags & MSM_SUBMIT_SUDO)
710 		submit->in_rb = true;
711 
712 	if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
713 		struct dma_fence *in_fence;
714 
715 		in_fence = sync_file_get_fence(args->fence_fd);
716 
717 		if (!in_fence) {
718 			ret = UERR(EINVAL, dev, "invalid in-fence");
719 			goto out_unlock;
720 		}
721 
722 		ret = drm_sched_job_add_dependency(&submit->base, in_fence);
723 		if (ret)
724 			goto out_unlock;
725 	}
726 
727 	if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
728 		syncobjs_to_reset = msm_parse_deps(submit, file,
729 		                                   args->in_syncobjs,
730 		                                   args->nr_in_syncobjs,
731 		                                   args->syncobj_stride);
732 		if (IS_ERR(syncobjs_to_reset)) {
733 			ret = PTR_ERR(syncobjs_to_reset);
734 			goto out_unlock;
735 		}
736 	}
737 
738 	if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
739 		post_deps = msm_parse_post_deps(dev, file,
740 		                                args->out_syncobjs,
741 		                                args->nr_out_syncobjs,
742 		                                args->syncobj_stride);
743 		if (IS_ERR(post_deps)) {
744 			ret = PTR_ERR(post_deps);
745 			goto out_unlock;
746 		}
747 	}
748 
749 	ret = submit_lookup_objects(submit, args, file);
750 	if (ret)
751 		goto out;
752 
753 	ret = submit_lookup_cmds(submit, args, file);
754 	if (ret)
755 		goto out;
756 
757 	/* copy_*_user while holding a ww ticket upsets lockdep */
758 	ret = submit_lock_objects(submit);
759 	if (ret)
760 		goto out;
761 
762 	if (!(args->flags & MSM_SUBMIT_NO_IMPLICIT)) {
763 		ret = submit_fence_sync(submit);
764 		if (ret)
765 			goto out;
766 	}
767 
768 	ret = submit_pin_objects(submit);
769 	if (ret)
770 		goto out;
771 
772 	for (i = 0; i < args->nr_cmds; i++) {
773 		struct drm_gem_object *obj;
774 		uint64_t iova;
775 
776 		ret = submit_bo(submit, submit->cmd[i].idx, &obj, &iova);
777 		if (ret)
778 			goto out;
779 
780 		if (!submit->cmd[i].size ||
781 		    (size_add(submit->cmd[i].size, submit->cmd[i].offset) > obj->size / 4)) {
782 			ret = UERR(EINVAL, dev, "invalid cmdstream size: %u\n",
783 				   submit->cmd[i].size * 4);
784 			goto out;
785 		}
786 
787 		submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
788 
789 		if (likely(!submit->cmd[i].nr_relocs))
790 			continue;
791 
792 		if (!gpu->allow_relocs) {
793 			ret = UERR(EINVAL, dev, "relocs not allowed\n");
794 			goto out;
795 		}
796 
797 		ret = submit_reloc(submit, obj, submit->cmd[i].offset * 4,
798 				submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
799 		if (ret)
800 			goto out;
801 	}
802 
803 	submit->nr_cmds = i;
804 
805 	idr_preload(GFP_KERNEL);
806 
807 	spin_lock(&queue->idr_lock);
808 
809 	/*
810 	 * If using userspace provided seqno fence, validate that the id
811 	 * is available before arming sched job.  Since access to fence_idr
812 	 * is serialized on the queue lock, the slot should be still avail
813 	 * after the job is armed
814 	 */
815 	if ((args->flags & MSM_SUBMIT_FENCE_SN_IN) &&
816 			(!args->fence || idr_find(&queue->fence_idr, args->fence))) {
817 		spin_unlock(&queue->idr_lock);
818 		idr_preload_end();
819 		ret = UERR(EINVAL, dev, "invalid in-fence-sn");
820 		goto out;
821 	}
822 
823 	drm_sched_job_arm(&submit->base);
824 
825 	submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
826 
827 	if (args->flags & MSM_SUBMIT_FENCE_SN_IN) {
828 		/*
829 		 * Userspace has assigned the seqno fence that it wants
830 		 * us to use.  It is an error to pick a fence sequence
831 		 * number that is not available.
832 		 */
833 		submit->fence_id = args->fence;
834 		ret = idr_alloc_u32(&queue->fence_idr, submit->user_fence,
835 				    &submit->fence_id, submit->fence_id,
836 				    GFP_NOWAIT);
837 		/*
838 		 * We've already validated that the fence_id slot is valid,
839 		 * so if idr_alloc_u32 failed, it is a kernel bug
840 		 */
841 		WARN_ON(ret);
842 	} else {
843 		/*
844 		 * Allocate an id which can be used by WAIT_FENCE ioctl to map
845 		 * back to the underlying fence.
846 		 */
847 		submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
848 						    submit->user_fence, 1,
849 						    INT_MAX, GFP_NOWAIT);
850 	}
851 
852 	spin_unlock(&queue->idr_lock);
853 	idr_preload_end();
854 
855 	if (submit->fence_id < 0) {
856 		ret = submit->fence_id;
857 		submit->fence_id = 0;
858 	}
859 
860 	if (ret == 0 && args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
861 		struct sync_file *sync_file = sync_file_create(submit->user_fence);
862 		if (!sync_file) {
863 			ret = -ENOMEM;
864 		} else {
865 			fd_install(out_fence_fd, sync_file->file);
866 			args->fence_fd = out_fence_fd;
867 		}
868 	}
869 
870 	if (ret)
871 		goto out;
872 
873 	submit_attach_object_fences(submit);
874 
875 	/* The scheduler owns a ref now: */
876 	msm_gem_submit_get(submit);
877 
878 	msm_rd_dump_submit(priv->rd, submit, NULL);
879 
880 	drm_sched_entity_push_job(&submit->base);
881 
882 	args->fence = submit->fence_id;
883 	queue->last_fence = submit->fence_id;
884 
885 	msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
886 	msm_process_post_deps(post_deps, args->nr_out_syncobjs,
887 	                      submit->user_fence);
888 
889 
890 out:
891 	submit_cleanup(submit, !!ret);
892 out_unlock:
893 	mutex_unlock(&queue->lock);
894 out_post_unlock:
895 	if (ret && (out_fence_fd >= 0))
896 		put_unused_fd(out_fence_fd);
897 
898 	if (!IS_ERR_OR_NULL(submit)) {
899 		msm_gem_submit_put(submit);
900 	} else {
901 		/*
902 		 * If the submit hasn't yet taken ownership of the queue
903 		 * then we need to drop the reference ourself:
904 		 */
905 		msm_submitqueue_put(queue);
906 	}
907 	if (!IS_ERR_OR_NULL(post_deps)) {
908 		for (i = 0; i < args->nr_out_syncobjs; ++i) {
909 			kfree(post_deps[i].chain);
910 			drm_syncobj_put(post_deps[i].syncobj);
911 		}
912 		kfree(post_deps);
913 	}
914 
915 	if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
916 		for (i = 0; i < args->nr_in_syncobjs; ++i) {
917 			if (syncobjs_to_reset[i])
918 				drm_syncobj_put(syncobjs_to_reset[i]);
919 		}
920 		kfree(syncobjs_to_reset);
921 	}
922 
923 	return ret;
924 }
925