xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
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 
34 static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
35 {
36 	struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
37 	struct amdgpu_job *job = to_amdgpu_job(s_job);
38 	struct amdgpu_task_info *ti;
39 	struct amdgpu_device *adev = ring->adev;
40 	int idx;
41 	int r;
42 
43 	if (!drm_dev_enter(adev_to_drm(adev), &idx)) {
44 		dev_info(adev->dev, "%s - device unplugged skipping recovery on scheduler:%s",
45 			 __func__, s_job->sched->name);
46 
47 		/* Effectively the job is aborted as the device is gone */
48 		return DRM_GPU_SCHED_STAT_ENODEV;
49 	}
50 
51 
52 	adev->job_hang = true;
53 
54 	if (amdgpu_gpu_recovery &&
55 	    amdgpu_ring_soft_recovery(ring, job->vmid, s_job->s_fence->parent)) {
56 		dev_err(adev->dev, "ring %s timeout, but soft recovered\n",
57 			s_job->sched->name);
58 		goto exit;
59 	}
60 
61 	dev_err(adev->dev, "ring %s timeout, signaled seq=%u, emitted seq=%u\n",
62 		job->base.sched->name, atomic_read(&ring->fence_drv.last_seq),
63 		ring->fence_drv.sync_seq);
64 
65 	ti = amdgpu_vm_get_task_info_pasid(ring->adev, job->pasid);
66 	if (ti) {
67 		dev_err(adev->dev,
68 			"Process information: process %s pid %d thread %s pid %d\n",
69 			ti->process_name, ti->tgid, ti->task_name, ti->pid);
70 		amdgpu_vm_put_task_info(ti);
71 	}
72 
73 	dma_fence_set_error(&s_job->s_fence->finished, -ETIME);
74 
75 	if (amdgpu_device_should_recover_gpu(ring->adev)) {
76 		struct amdgpu_reset_context reset_context;
77 		memset(&reset_context, 0, sizeof(reset_context));
78 
79 		reset_context.method = AMD_RESET_METHOD_NONE;
80 		reset_context.reset_req_dev = adev;
81 		reset_context.src = AMDGPU_RESET_SRC_JOB;
82 		clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
83 
84 		r = amdgpu_device_gpu_recover(ring->adev, job, &reset_context);
85 		if (r)
86 			dev_err(adev->dev, "GPU Recovery Failed: %d\n", r);
87 	} else {
88 		drm_sched_suspend_timeout(&ring->sched);
89 		if (amdgpu_sriov_vf(adev))
90 			adev->virt.tdr_debug = true;
91 	}
92 
93 exit:
94 	adev->job_hang = false;
95 	drm_dev_exit(idx);
96 	return DRM_GPU_SCHED_STAT_NOMINAL;
97 }
98 
99 int amdgpu_job_alloc(struct amdgpu_device *adev, struct amdgpu_vm *vm,
100 		     struct drm_sched_entity *entity, void *owner,
101 		     unsigned int num_ibs, struct amdgpu_job **job)
102 {
103 	if (num_ibs == 0)
104 		return -EINVAL;
105 
106 	*job = kzalloc(struct_size(*job, ibs, num_ibs), GFP_KERNEL);
107 	if (!*job)
108 		return -ENOMEM;
109 
110 	/*
111 	 * Initialize the scheduler to at least some ring so that we always
112 	 * have a pointer to adev.
113 	 */
114 	(*job)->base.sched = &adev->rings[0]->sched;
115 	(*job)->vm = vm;
116 
117 	amdgpu_sync_create(&(*job)->explicit_sync);
118 	(*job)->generation = amdgpu_vm_generation(adev, vm);
119 	(*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
120 
121 	if (!entity)
122 		return 0;
123 
124 	return drm_sched_job_init(&(*job)->base, entity, 1, owner);
125 }
126 
127 int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev,
128 			     struct drm_sched_entity *entity, void *owner,
129 			     size_t size, enum amdgpu_ib_pool_type pool_type,
130 			     struct amdgpu_job **job)
131 {
132 	int r;
133 
134 	r = amdgpu_job_alloc(adev, NULL, entity, owner, 1, job);
135 	if (r)
136 		return r;
137 
138 	(*job)->num_ibs = 1;
139 	r = amdgpu_ib_get(adev, NULL, size, pool_type, &(*job)->ibs[0]);
140 	if (r) {
141 		if (entity)
142 			drm_sched_job_cleanup(&(*job)->base);
143 		kfree(*job);
144 	}
145 
146 	return r;
147 }
148 
149 void amdgpu_job_set_resources(struct amdgpu_job *job, struct amdgpu_bo *gds,
150 			      struct amdgpu_bo *gws, struct amdgpu_bo *oa)
151 {
152 	if (gds) {
153 		job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
154 		job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
155 	}
156 	if (gws) {
157 		job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
158 		job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
159 	}
160 	if (oa) {
161 		job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
162 		job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
163 	}
164 }
165 
166 void amdgpu_job_free_resources(struct amdgpu_job *job)
167 {
168 	struct amdgpu_ring *ring = to_amdgpu_ring(job->base.sched);
169 	struct dma_fence *f;
170 	unsigned i;
171 
172 	/* Check if any fences where initialized */
173 	if (job->base.s_fence && job->base.s_fence->finished.ops)
174 		f = &job->base.s_fence->finished;
175 	else if (job->hw_fence.ops)
176 		f = &job->hw_fence;
177 	else
178 		f = NULL;
179 
180 	for (i = 0; i < job->num_ibs; ++i)
181 		amdgpu_ib_free(ring->adev, &job->ibs[i], f);
182 }
183 
184 static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
185 {
186 	struct amdgpu_job *job = to_amdgpu_job(s_job);
187 
188 	drm_sched_job_cleanup(s_job);
189 
190 	amdgpu_sync_free(&job->explicit_sync);
191 
192 	/* only put the hw fence if has embedded fence */
193 	if (!job->hw_fence.ops)
194 		kfree(job);
195 	else
196 		dma_fence_put(&job->hw_fence);
197 }
198 
199 void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
200 				struct amdgpu_job *leader)
201 {
202 	struct dma_fence *fence = &leader->base.s_fence->scheduled;
203 
204 	WARN_ON(job->gang_submit);
205 
206 	/*
207 	 * Don't add a reference when we are the gang leader to avoid circle
208 	 * dependency.
209 	 */
210 	if (job != leader)
211 		dma_fence_get(fence);
212 	job->gang_submit = fence;
213 }
214 
215 void amdgpu_job_free(struct amdgpu_job *job)
216 {
217 	if (job->base.entity)
218 		drm_sched_job_cleanup(&job->base);
219 
220 	amdgpu_job_free_resources(job);
221 	amdgpu_sync_free(&job->explicit_sync);
222 	if (job->gang_submit != &job->base.s_fence->scheduled)
223 		dma_fence_put(job->gang_submit);
224 
225 	if (!job->hw_fence.ops)
226 		kfree(job);
227 	else
228 		dma_fence_put(&job->hw_fence);
229 }
230 
231 struct dma_fence *amdgpu_job_submit(struct amdgpu_job *job)
232 {
233 	struct dma_fence *f;
234 
235 	drm_sched_job_arm(&job->base);
236 	f = dma_fence_get(&job->base.s_fence->finished);
237 	amdgpu_job_free_resources(job);
238 	drm_sched_entity_push_job(&job->base);
239 
240 	return f;
241 }
242 
243 int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
244 			     struct dma_fence **fence)
245 {
246 	int r;
247 
248 	job->base.sched = &ring->sched;
249 	r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job, fence);
250 
251 	if (r)
252 		return r;
253 
254 	amdgpu_job_free(job);
255 	return 0;
256 }
257 
258 static struct dma_fence *
259 amdgpu_job_prepare_job(struct drm_sched_job *sched_job,
260 		      struct drm_sched_entity *s_entity)
261 {
262 	struct amdgpu_ring *ring = to_amdgpu_ring(s_entity->rq->sched);
263 	struct amdgpu_job *job = to_amdgpu_job(sched_job);
264 	struct dma_fence *fence = NULL;
265 	int r;
266 
267 	r = drm_sched_entity_error(s_entity);
268 	if (r)
269 		goto error;
270 
271 	if (!fence && job->gang_submit)
272 		fence = amdgpu_device_switch_gang(ring->adev, job->gang_submit);
273 
274 	while (!fence && job->vm && !job->vmid) {
275 		r = amdgpu_vmid_grab(job->vm, ring, job, &fence);
276 		if (r) {
277 			dev_err(ring->adev->dev, "Error getting VM ID (%d)\n", r);
278 			goto error;
279 		}
280 	}
281 
282 	return fence;
283 
284 error:
285 	dma_fence_set_error(&job->base.s_fence->finished, r);
286 	return NULL;
287 }
288 
289 static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
290 {
291 	struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
292 	struct amdgpu_device *adev = ring->adev;
293 	struct dma_fence *fence = NULL, *finished;
294 	struct amdgpu_job *job;
295 	int r = 0;
296 
297 	job = to_amdgpu_job(sched_job);
298 	finished = &job->base.s_fence->finished;
299 
300 	trace_amdgpu_sched_run_job(job);
301 
302 	/* Skip job if VRAM is lost and never resubmit gangs */
303 	if (job->generation != amdgpu_vm_generation(adev, job->vm) ||
304 	    (job->job_run_counter && job->gang_submit))
305 		dma_fence_set_error(finished, -ECANCELED);
306 
307 	if (finished->error < 0) {
308 		dev_dbg(adev->dev, "Skip scheduling IBs in ring(%s)",
309 			ring->name);
310 	} else {
311 		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
312 				       &fence);
313 		if (r)
314 			dev_err(adev->dev,
315 				"Error scheduling IBs (%d) in ring(%s)", r,
316 				ring->name);
317 	}
318 
319 	job->job_run_counter++;
320 	amdgpu_job_free_resources(job);
321 
322 	fence = r ? ERR_PTR(r) : fence;
323 	return fence;
324 }
325 
326 #define to_drm_sched_job(sched_job)		\
327 		container_of((sched_job), struct drm_sched_job, queue_node)
328 
329 void amdgpu_job_stop_all_jobs_on_sched(struct drm_gpu_scheduler *sched)
330 {
331 	struct drm_sched_job *s_job;
332 	struct drm_sched_entity *s_entity = NULL;
333 	int i;
334 
335 	/* Signal all jobs not yet scheduled */
336 	for (i = DRM_SCHED_PRIORITY_KERNEL; i < sched->num_rqs; i++) {
337 		struct drm_sched_rq *rq = sched->sched_rq[i];
338 		spin_lock(&rq->lock);
339 		list_for_each_entry(s_entity, &rq->entities, list) {
340 			while ((s_job = to_drm_sched_job(spsc_queue_pop(&s_entity->job_queue)))) {
341 				struct drm_sched_fence *s_fence = s_job->s_fence;
342 
343 				dma_fence_signal(&s_fence->scheduled);
344 				dma_fence_set_error(&s_fence->finished, -EHWPOISON);
345 				dma_fence_signal(&s_fence->finished);
346 			}
347 		}
348 		spin_unlock(&rq->lock);
349 	}
350 
351 	/* Signal all jobs already scheduled to HW */
352 	list_for_each_entry(s_job, &sched->pending_list, list) {
353 		struct drm_sched_fence *s_fence = s_job->s_fence;
354 
355 		dma_fence_set_error(&s_fence->finished, -EHWPOISON);
356 		dma_fence_signal(&s_fence->finished);
357 	}
358 }
359 
360 const struct drm_sched_backend_ops amdgpu_sched_ops = {
361 	.prepare_job = amdgpu_job_prepare_job,
362 	.run_job = amdgpu_job_run,
363 	.timedout_job = amdgpu_job_timedout,
364 	.free_job = amdgpu_job_free_cb
365 };
366