Linux Audio

Check our new training course

Buildroot integration, development and maintenance

Need a Buildroot system for your embedded project?
Loading...
v4.17
  1/*
  2 * Copyright 2015 Advanced Micro Devices, Inc.
  3 *
  4 * Permission is hereby granted, free of charge, to any person obtaining a
  5 * copy of this software and associated documentation files (the "Software"),
  6 * to deal in the Software without restriction, including without limitation
  7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  8 * and/or sell copies of the Software, and to permit persons to whom the
  9 * Software is furnished to do so, subject to the following conditions:
 10 *
 11 * The above copyright notice and this permission notice shall be included in
 12 * all copies or substantial portions of the Software.
 13 *
 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 20 * OTHER DEALINGS IN THE SOFTWARE.
 21 *
 22 *
 23 */
 24#include <linux/kthread.h>
 25#include <linux/wait.h>
 26#include <linux/sched.h>
 27#include <drm/drmP.h>
 
 
 28#include "amdgpu.h"
 29#include "amdgpu_trace.h"
 
 
 
 30
 31static void amdgpu_job_timedout(struct drm_sched_job *s_job)
 
 32{
 33	struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base);
 34
 35	DRM_ERROR("ring %s timeout, last signaled seq=%u, last emitted seq=%u\n",
 36		  job->base.sched->name,
 37		  atomic_read(&job->ring->fence_drv.last_seq),
 38		  job->ring->fence_drv.sync_seq);
 
 
 39
 40	amdgpu_device_gpu_recover(job->adev, job, false);
 41}
 42
 43int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs,
 44		     struct amdgpu_job **job, struct amdgpu_vm *vm)
 45{
 46	size_t size = sizeof(struct amdgpu_job);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 47
 
 
 
 
 
 
 
 
 
 
 48	if (num_ibs == 0)
 49		return -EINVAL;
 50
 51	size += sizeof(struct amdgpu_ib) * num_ibs;
 52
 53	*job = kzalloc(size, GFP_KERNEL);
 54	if (!*job)
 55		return -ENOMEM;
 56
 57	(*job)->adev = adev;
 
 
 
 
 58	(*job)->vm = vm;
 59	(*job)->ibs = (void *)&(*job)[1];
 60	(*job)->num_ibs = num_ibs;
 61
 62	amdgpu_sync_create(&(*job)->sync);
 63	amdgpu_sync_create(&(*job)->sched_sync);
 64	(*job)->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
 65
 66	return 0;
 
 
 
 67}
 68
 69int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
 
 
 70			     struct amdgpu_job **job)
 71{
 72	int r;
 73
 74	r = amdgpu_job_alloc(adev, 1, job, NULL);
 75	if (r)
 76		return r;
 77
 78	r = amdgpu_ib_get(adev, NULL, size, &(*job)->ibs[0]);
 79	if (r)
 
 
 
 80		kfree(*job);
 81	else
 82		(*job)->vm_pd_addr = adev->gart.table_addr;
 83
 84	return r;
 85}
 86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 87void amdgpu_job_free_resources(struct amdgpu_job *job)
 88{
 89	struct dma_fence *f;
 90	unsigned i;
 91
 92	/* use sched fence if available */
 93	f = job->base.s_fence ? &job->base.s_fence->finished : job->fence;
 
 
 
 
 
 94
 95	for (i = 0; i < job->num_ibs; ++i)
 96		amdgpu_ib_free(job->adev, &job->ibs[i], f);
 97}
 98
 99static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
100{
101	struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base);
 
 
 
 
102
103	amdgpu_ring_priority_put(job->ring, s_job->s_priority);
104	dma_fence_put(job->fence);
105	amdgpu_sync_free(&job->sync);
106	amdgpu_sync_free(&job->sched_sync);
107	kfree(job);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
108}
109
110void amdgpu_job_free(struct amdgpu_job *job)
111{
 
 
 
112	amdgpu_job_free_resources(job);
 
 
 
113
114	dma_fence_put(job->fence);
115	amdgpu_sync_free(&job->sync);
116	amdgpu_sync_free(&job->sched_sync);
117	kfree(job);
118}
119
120int amdgpu_job_submit(struct amdgpu_job *job, struct amdgpu_ring *ring,
121		      struct drm_sched_entity *entity, void *owner,
122		      struct dma_fence **f)
 
 
 
 
 
 
 
 
 
 
 
123{
124	int r;
125	job->ring = ring;
126
127	if (!f)
128		return -EINVAL;
129
130	r = drm_sched_job_init(&job->base, &ring->sched, entity, owner);
131	if (r)
132		return r;
133
134	job->owner = owner;
135	job->fence_ctx = entity->fence_context;
136	*f = dma_fence_get(&job->base.s_fence->finished);
137	amdgpu_job_free_resources(job);
138	amdgpu_ring_priority_get(job->ring, job->base.s_priority);
139	drm_sched_entity_push_job(&job->base, entity);
140
141	return 0;
142}
143
144static struct dma_fence *amdgpu_job_dependency(struct drm_sched_job *sched_job,
145					       struct drm_sched_entity *s_entity)
 
146{
 
147	struct amdgpu_job *job = to_amdgpu_job(sched_job);
148	struct amdgpu_vm *vm = job->vm;
149	bool explicit = false;
150	int r;
151	struct dma_fence *fence = amdgpu_sync_get_fence(&job->sync, &explicit);
152
153	if (fence && explicit) {
154		if (drm_sched_dependency_optimized(fence, s_entity)) {
155			r = amdgpu_sync_fence(job->adev, &job->sched_sync, fence, false);
156			if (r)
157				DRM_ERROR("Error adding fence to sync (%d)\n", r);
158		}
159	}
160
161	while (fence == NULL && vm && !job->vmid) {
162		struct amdgpu_ring *ring = job->ring;
163
164		r = amdgpu_vmid_grab(vm, ring, &job->sync,
165				     &job->base.s_fence->finished,
166				     job);
167		if (r)
168			DRM_ERROR("Error getting VM ID (%d)\n", r);
169
170		fence = amdgpu_sync_get_fence(&job->sync, NULL);
 
 
 
 
 
171	}
172
173	return fence;
 
 
 
 
174}
175
176static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
177{
 
 
178	struct dma_fence *fence = NULL, *finished;
179	struct amdgpu_device *adev;
180	struct amdgpu_job *job;
181	int r;
182
183	if (!sched_job) {
184		DRM_ERROR("job is null\n");
185		return NULL;
186	}
187	job = to_amdgpu_job(sched_job);
188	finished = &job->base.s_fence->finished;
189	adev = job->adev;
190
191	BUG_ON(amdgpu_sync_peek_fence(&job->sync, NULL));
192
193	trace_amdgpu_sched_run_job(job);
194
195	if (job->vram_lost_counter != atomic_read(&adev->vram_lost_counter))
196		dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */
 
 
197
198	if (finished->error < 0) {
199		DRM_INFO("Skip scheduling IBs!\n");
 
200	} else {
201		r = amdgpu_ib_schedule(job->ring, job->num_ibs, job->ibs, job,
202				       &fence);
203		if (r)
204			DRM_ERROR("Error scheduling IBs (%d)\n", r);
 
 
205	}
206	/* if gpu reset, hw fence will be replaced here */
207	dma_fence_put(job->fence);
208	job->fence = dma_fence_get(fence);
209
 
210	amdgpu_job_free_resources(job);
 
 
211	return fence;
212}
213
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
214const struct drm_sched_backend_ops amdgpu_sched_ops = {
215	.dependency = amdgpu_job_dependency,
216	.run_job = amdgpu_job_run,
217	.timedout_job = amdgpu_job_timedout,
218	.free_job = amdgpu_job_free_cb
219};
v6.13.7
  1/*
  2 * Copyright 2015 Advanced Micro Devices, Inc.
  3 *
  4 * Permission is hereby granted, free of charge, to any person obtaining a
  5 * copy of this software and associated documentation files (the "Software"),
  6 * to deal in the Software without restriction, including without limitation
  7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  8 * and/or sell copies of the Software, and to permit persons to whom the
  9 * Software is furnished to do so, subject to the following conditions:
 10 *
 11 * The above copyright notice and this permission notice shall be included in
 12 * all copies or substantial portions of the Software.
 13 *
 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 20 * OTHER DEALINGS IN THE SOFTWARE.
 21 *
 22 *
 23 */
 24#include <linux/kthread.h>
 25#include <linux/wait.h>
 26#include <linux/sched.h>
 27
 28#include <drm/drm_drv.h>
 29
 30#include "amdgpu.h"
 31#include "amdgpu_trace.h"
 32#include "amdgpu_reset.h"
 33#include "amdgpu_dev_coredump.h"
 34#include "amdgpu_xgmi.h"
 35
 36static void amdgpu_job_do_core_dump(struct amdgpu_device *adev,
 37				    struct amdgpu_job *job)
 38{
 39	int i;
 40
 41	dev_info(adev->dev, "Dumping IP State\n");
 42	for (i = 0; i < adev->num_ip_blocks; i++)
 43		if (adev->ip_blocks[i].version->funcs->dump_ip_state)
 44			adev->ip_blocks[i].version->funcs
 45				->dump_ip_state((void *)&adev->ip_blocks[i]);
 46	dev_info(adev->dev, "Dumping IP State Completed\n");
 47
 48	amdgpu_coredump(adev, true, false, job);
 49}
 50
 51static void amdgpu_job_core_dump(struct amdgpu_device *adev,
 52				 struct amdgpu_job *job)
 53{
 54	struct list_head device_list, *device_list_handle =  NULL;
 55	struct amdgpu_device *tmp_adev = NULL;
 56	struct amdgpu_hive_info *hive = NULL;
 57
 58	if (!amdgpu_sriov_vf(adev))
 59		hive = amdgpu_get_xgmi_hive(adev);
 60	if (hive)
 61		mutex_lock(&hive->hive_lock);
 62	/*
 63	 * Reuse the logic in amdgpu_device_gpu_recover() to build list of
 64	 * devices for code dump
 65	 */
 66	INIT_LIST_HEAD(&device_list);
 67	if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1) && hive) {
 68		list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head)
 69			list_add_tail(&tmp_adev->reset_list, &device_list);
 70		if (!list_is_first(&adev->reset_list, &device_list))
 71			list_rotate_to_front(&adev->reset_list, &device_list);
 72		device_list_handle = &device_list;
 73	} else {
 74		list_add_tail(&adev->reset_list, &device_list);
 75		device_list_handle = &device_list;
 76	}
 77
 78	/* Do the coredump for each device */
 79	list_for_each_entry(tmp_adev, device_list_handle, reset_list)
 80		amdgpu_job_do_core_dump(tmp_adev, job);
 81
 82	if (hive) {
 83		mutex_unlock(&hive->hive_lock);
 84		amdgpu_put_xgmi_hive(hive);
 85	}
 86}
 87
 88static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
 89{
 90	struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
 91	struct amdgpu_job *job = to_amdgpu_job(s_job);
 92	struct amdgpu_task_info *ti;
 93	struct amdgpu_device *adev = ring->adev;
 94	int idx;
 95	int r;
 96
 97	if (!drm_dev_enter(adev_to_drm(adev), &idx)) {
 98		dev_info(adev->dev, "%s - device unplugged skipping recovery on scheduler:%s",
 99			 __func__, s_job->sched->name);
100
101		/* Effectively the job is aborted as the device is gone */
102		return DRM_GPU_SCHED_STAT_ENODEV;
103	}
104
105	adev->job_hang = true;
106
107	/*
108	 * Do the coredump immediately after a job timeout to get a very
109	 * close dump/snapshot/representation of GPU's current error status
110	 * Skip it for SRIOV, since VF FLR will be triggered by host driver
111	 * before job timeout
112	 */
113	if (!amdgpu_sriov_vf(adev))
114		amdgpu_job_core_dump(adev, job);
115
116	if (amdgpu_gpu_recovery &&
117	    amdgpu_ring_soft_recovery(ring, job->vmid, s_job->s_fence->parent)) {
118		dev_err(adev->dev, "ring %s timeout, but soft recovered\n",
119			s_job->sched->name);
120		goto exit;
121	}
122
123	dev_err(adev->dev, "ring %s timeout, signaled seq=%u, emitted seq=%u\n",
124		job->base.sched->name, atomic_read(&ring->fence_drv.last_seq),
125		ring->fence_drv.sync_seq);
126
127	ti = amdgpu_vm_get_task_info_pasid(ring->adev, job->pasid);
128	if (ti) {
129		dev_err(adev->dev,
130			"Process information: process %s pid %d thread %s pid %d\n",
131			ti->process_name, ti->tgid, ti->task_name, ti->pid);
132		amdgpu_vm_put_task_info(ti);
133	}
134
135	dma_fence_set_error(&s_job->s_fence->finished, -ETIME);
136
137	/* attempt a per ring reset */
138	if (amdgpu_gpu_recovery &&
139	    ring->funcs->reset) {
140		dev_err(adev->dev, "Starting %s ring reset\n", s_job->sched->name);
141		/* stop the scheduler, but don't mess with the
142		 * bad job yet because if ring reset fails
143		 * we'll fall back to full GPU reset.
144		 */
145		drm_sched_wqueue_stop(&ring->sched);
146		r = amdgpu_ring_reset(ring, job->vmid);
147		if (!r) {
148			if (amdgpu_ring_sched_ready(ring))
149				drm_sched_stop(&ring->sched, s_job);
150			atomic_inc(&ring->adev->gpu_reset_counter);
151			amdgpu_fence_driver_force_completion(ring);
152			if (amdgpu_ring_sched_ready(ring))
153				drm_sched_start(&ring->sched, 0);
154			goto exit;
155		}
156		dev_err(adev->dev, "Ring %s reset failure\n", ring->sched.name);
157	}
158
159	if (amdgpu_device_should_recover_gpu(ring->adev)) {
160		struct amdgpu_reset_context reset_context;
161		memset(&reset_context, 0, sizeof(reset_context));
162
163		reset_context.method = AMD_RESET_METHOD_NONE;
164		reset_context.reset_req_dev = adev;
165		reset_context.src = AMDGPU_RESET_SRC_JOB;
166		clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
167
168		/*
169		 * To avoid an unnecessary extra coredump, as we have already
170		 * got the very close representation of GPU's error status
171		 */
172		set_bit(AMDGPU_SKIP_COREDUMP, &reset_context.flags);
173
174		r = amdgpu_device_gpu_recover(ring->adev, job, &reset_context);
175		if (r)
176			dev_err(adev->dev, "GPU Recovery Failed: %d\n", r);
177	} else {
178		drm_sched_suspend_timeout(&ring->sched);
179		if (amdgpu_sriov_vf(adev))
180			adev->virt.tdr_debug = true;
181	}
182
183exit:
184	adev->job_hang = false;
185	drm_dev_exit(idx);
186	return DRM_GPU_SCHED_STAT_NOMINAL;
187}
188
189int amdgpu_job_alloc(struct amdgpu_device *adev, struct amdgpu_vm *vm,
190		     struct drm_sched_entity *entity, void *owner,
191		     unsigned int num_ibs, struct amdgpu_job **job)
192{
193	if (num_ibs == 0)
194		return -EINVAL;
195
196	*job = kzalloc(struct_size(*job, ibs, num_ibs), GFP_KERNEL);
 
 
197	if (!*job)
198		return -ENOMEM;
199
200	/*
201	 * Initialize the scheduler to at least some ring so that we always
202	 * have a pointer to adev.
203	 */
204	(*job)->base.sched = &adev->rings[0]->sched;
205	(*job)->vm = vm;
 
 
206
207	amdgpu_sync_create(&(*job)->explicit_sync);
208	(*job)->generation = amdgpu_vm_generation(adev, vm);
209	(*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
210
211	if (!entity)
212		return 0;
213
214	return drm_sched_job_init(&(*job)->base, entity, 1, owner);
215}
216
217int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev,
218			     struct drm_sched_entity *entity, void *owner,
219			     size_t size, enum amdgpu_ib_pool_type pool_type,
220			     struct amdgpu_job **job)
221{
222	int r;
223
224	r = amdgpu_job_alloc(adev, NULL, entity, owner, 1, job);
225	if (r)
226		return r;
227
228	(*job)->num_ibs = 1;
229	r = amdgpu_ib_get(adev, NULL, size, pool_type, &(*job)->ibs[0]);
230	if (r) {
231		if (entity)
232			drm_sched_job_cleanup(&(*job)->base);
233		kfree(*job);
234	}
 
235
236	return r;
237}
238
239void amdgpu_job_set_resources(struct amdgpu_job *job, struct amdgpu_bo *gds,
240			      struct amdgpu_bo *gws, struct amdgpu_bo *oa)
241{
242	if (gds) {
243		job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
244		job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
245	}
246	if (gws) {
247		job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
248		job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
249	}
250	if (oa) {
251		job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
252		job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
253	}
254}
255
256void amdgpu_job_free_resources(struct amdgpu_job *job)
257{
258	struct dma_fence *f;
259	unsigned i;
260
261	/* Check if any fences where initialized */
262	if (job->base.s_fence && job->base.s_fence->finished.ops)
263		f = &job->base.s_fence->finished;
264	else if (job->hw_fence.ops)
265		f = &job->hw_fence;
266	else
267		f = NULL;
268
269	for (i = 0; i < job->num_ibs; ++i)
270		amdgpu_ib_free(NULL, &job->ibs[i], f);
271}
272
273static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
274{
275	struct amdgpu_job *job = to_amdgpu_job(s_job);
276
277	drm_sched_job_cleanup(s_job);
278
279	amdgpu_sync_free(&job->explicit_sync);
280
281	/* only put the hw fence if has embedded fence */
282	if (!job->hw_fence.ops)
283		kfree(job);
284	else
285		dma_fence_put(&job->hw_fence);
286}
287
288void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
289				struct amdgpu_job *leader)
290{
291	struct dma_fence *fence = &leader->base.s_fence->scheduled;
292
293	WARN_ON(job->gang_submit);
294
295	/*
296	 * Don't add a reference when we are the gang leader to avoid circle
297	 * dependency.
298	 */
299	if (job != leader)
300		dma_fence_get(fence);
301	job->gang_submit = fence;
302}
303
304void amdgpu_job_free(struct amdgpu_job *job)
305{
306	if (job->base.entity)
307		drm_sched_job_cleanup(&job->base);
308
309	amdgpu_job_free_resources(job);
310	amdgpu_sync_free(&job->explicit_sync);
311	if (job->gang_submit != &job->base.s_fence->scheduled)
312		dma_fence_put(job->gang_submit);
313
314	if (!job->hw_fence.ops)
315		kfree(job);
316	else
317		dma_fence_put(&job->hw_fence);
318}
319
320struct dma_fence *amdgpu_job_submit(struct amdgpu_job *job)
321{
322	struct dma_fence *f;
323
324	drm_sched_job_arm(&job->base);
325	f = dma_fence_get(&job->base.s_fence->finished);
326	amdgpu_job_free_resources(job);
327	drm_sched_entity_push_job(&job->base);
328
329	return f;
330}
331
332int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
333			     struct dma_fence **fence)
334{
335	int r;
 
336
337	job->base.sched = &ring->sched;
338	r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job, fence);
339
 
340	if (r)
341		return r;
342
343	amdgpu_job_free(job);
 
 
 
 
 
 
344	return 0;
345}
346
347static struct dma_fence *
348amdgpu_job_prepare_job(struct drm_sched_job *sched_job,
349		      struct drm_sched_entity *s_entity)
350{
351	struct amdgpu_ring *ring = to_amdgpu_ring(s_entity->rq->sched);
352	struct amdgpu_job *job = to_amdgpu_job(sched_job);
353	struct dma_fence *fence = NULL;
 
354	int r;
 
355
356	r = drm_sched_entity_error(s_entity);
357	if (r)
358		goto error;
 
 
 
 
 
 
 
359
360	if (job->gang_submit)
361		fence = amdgpu_device_switch_gang(ring->adev, job->gang_submit);
 
 
 
362
363	if (!fence && job->vm && !job->vmid) {
364		r = amdgpu_vmid_grab(job->vm, ring, job, &fence);
365		if (r) {
366			dev_err(ring->adev->dev, "Error getting VM ID (%d)\n", r);
367			goto error;
368		}
369	}
370
371	return fence;
372
373error:
374	dma_fence_set_error(&job->base.s_fence->finished, r);
375	return NULL;
376}
377
378static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
379{
380	struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
381	struct amdgpu_device *adev = ring->adev;
382	struct dma_fence *fence = NULL, *finished;
 
383	struct amdgpu_job *job;
384	int r = 0;
385
 
 
 
 
386	job = to_amdgpu_job(sched_job);
387	finished = &job->base.s_fence->finished;
 
 
 
388
389	trace_amdgpu_sched_run_job(job);
390
391	/* Skip job if VRAM is lost and never resubmit gangs */
392	if (job->generation != amdgpu_vm_generation(adev, job->vm) ||
393	    (job->job_run_counter && job->gang_submit))
394		dma_fence_set_error(finished, -ECANCELED);
395
396	if (finished->error < 0) {
397		dev_dbg(adev->dev, "Skip scheduling IBs in ring(%s)",
398			ring->name);
399	} else {
400		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
401				       &fence);
402		if (r)
403			dev_err(adev->dev,
404				"Error scheduling IBs (%d) in ring(%s)", r,
405				ring->name);
406	}
 
 
 
407
408	job->job_run_counter++;
409	amdgpu_job_free_resources(job);
410
411	fence = r ? ERR_PTR(r) : fence;
412	return fence;
413}
414
415#define to_drm_sched_job(sched_job)		\
416		container_of((sched_job), struct drm_sched_job, queue_node)
417
418void amdgpu_job_stop_all_jobs_on_sched(struct drm_gpu_scheduler *sched)
419{
420	struct drm_sched_job *s_job;
421	struct drm_sched_entity *s_entity = NULL;
422	int i;
423
424	/* Signal all jobs not yet scheduled */
425	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
426		struct drm_sched_rq *rq = sched->sched_rq[i];
427		spin_lock(&rq->lock);
428		list_for_each_entry(s_entity, &rq->entities, list) {
429			while ((s_job = to_drm_sched_job(spsc_queue_pop(&s_entity->job_queue)))) {
430				struct drm_sched_fence *s_fence = s_job->s_fence;
431
432				dma_fence_signal(&s_fence->scheduled);
433				dma_fence_set_error(&s_fence->finished, -EHWPOISON);
434				dma_fence_signal(&s_fence->finished);
435			}
436		}
437		spin_unlock(&rq->lock);
438	}
439
440	/* Signal all jobs already scheduled to HW */
441	list_for_each_entry(s_job, &sched->pending_list, list) {
442		struct drm_sched_fence *s_fence = s_job->s_fence;
443
444		dma_fence_set_error(&s_fence->finished, -EHWPOISON);
445		dma_fence_signal(&s_fence->finished);
446	}
447}
448
449const struct drm_sched_backend_ops amdgpu_sched_ops = {
450	.prepare_job = amdgpu_job_prepare_job,
451	.run_job = amdgpu_job_run,
452	.timedout_job = amdgpu_job_timedout,
453	.free_job = amdgpu_job_free_cb
454};