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 
31 static void amdgpu_job_timedout(struct drm_sched_job *s_job)
32 {
33 	struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
34 	struct amdgpu_job *job = to_amdgpu_job(s_job);
35 
36 	if (amdgpu_ring_soft_recovery(ring, job->vmid, s_job->s_fence->parent)) {
37 		DRM_ERROR("ring %s timeout, but soft recovered\n",
38 			  s_job->sched->name);
39 		return;
40 	}
41 
42 	DRM_ERROR("ring %s timeout, signaled seq=%u, emitted seq=%u\n",
43 		  job->base.sched->name, atomic_read(&ring->fence_drv.last_seq),
44 		  ring->fence_drv.sync_seq);
45 
46 	if (amdgpu_device_should_recover_gpu(ring->adev))
47 		amdgpu_device_gpu_recover(ring->adev, job);
48 }
49 
50 int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs,
51 		     struct amdgpu_job **job, struct amdgpu_vm *vm)
52 {
53 	size_t size = sizeof(struct amdgpu_job);
54 
55 	if (num_ibs == 0)
56 		return -EINVAL;
57 
58 	size += sizeof(struct amdgpu_ib) * num_ibs;
59 
60 	*job = kzalloc(size, GFP_KERNEL);
61 	if (!*job)
62 		return -ENOMEM;
63 
64 	/*
65 	 * Initialize the scheduler to at least some ring so that we always
66 	 * have a pointer to adev.
67 	 */
68 	(*job)->base.sched = &adev->rings[0]->sched;
69 	(*job)->vm = vm;
70 	(*job)->ibs = (void *)&(*job)[1];
71 	(*job)->num_ibs = num_ibs;
72 
73 	amdgpu_sync_create(&(*job)->sync);
74 	amdgpu_sync_create(&(*job)->sched_sync);
75 	(*job)->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
76 	(*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
77 
78 	return 0;
79 }
80 
81 int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
82 			     struct amdgpu_job **job)
83 {
84 	int r;
85 
86 	r = amdgpu_job_alloc(adev, 1, job, NULL);
87 	if (r)
88 		return r;
89 
90 	r = amdgpu_ib_get(adev, NULL, size, &(*job)->ibs[0]);
91 	if (r)
92 		kfree(*job);
93 
94 	return r;
95 }
96 
97 void amdgpu_job_free_resources(struct amdgpu_job *job)
98 {
99 	struct amdgpu_ring *ring = to_amdgpu_ring(job->base.sched);
100 	struct dma_fence *f;
101 	unsigned i;
102 
103 	/* use sched fence if available */
104 	f = job->base.s_fence ? &job->base.s_fence->finished : job->fence;
105 
106 	for (i = 0; i < job->num_ibs; ++i)
107 		amdgpu_ib_free(ring->adev, &job->ibs[i], f);
108 }
109 
110 static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
111 {
112 	struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
113 	struct amdgpu_job *job = to_amdgpu_job(s_job);
114 
115 	drm_sched_job_cleanup(s_job);
116 
117 	amdgpu_ring_priority_put(ring, s_job->s_priority);
118 	dma_fence_put(job->fence);
119 	amdgpu_sync_free(&job->sync);
120 	amdgpu_sync_free(&job->sched_sync);
121 	kfree(job);
122 }
123 
124 void amdgpu_job_free(struct amdgpu_job *job)
125 {
126 	amdgpu_job_free_resources(job);
127 
128 	dma_fence_put(job->fence);
129 	amdgpu_sync_free(&job->sync);
130 	amdgpu_sync_free(&job->sched_sync);
131 	kfree(job);
132 }
133 
134 int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,
135 		      void *owner, struct dma_fence **f)
136 {
137 	enum drm_sched_priority priority;
138 	struct amdgpu_ring *ring;
139 	int r;
140 
141 	if (!f)
142 		return -EINVAL;
143 
144 	r = drm_sched_job_init(&job->base, entity, owner);
145 	if (r)
146 		return r;
147 
148 	job->owner = owner;
149 	*f = dma_fence_get(&job->base.s_fence->finished);
150 	amdgpu_job_free_resources(job);
151 	priority = job->base.s_priority;
152 	drm_sched_entity_push_job(&job->base, entity);
153 
154 	ring = to_amdgpu_ring(entity->rq->sched);
155 	amdgpu_ring_priority_get(ring, priority);
156 
157 	return 0;
158 }
159 
160 int amdgpu_job_submit_direct(struct amdgpu_job *job, struct amdgpu_ring *ring,
161 			     struct dma_fence **fence)
162 {
163 	int r;
164 
165 	job->base.sched = &ring->sched;
166 	r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, NULL, fence);
167 	job->fence = dma_fence_get(*fence);
168 	if (r)
169 		return r;
170 
171 	amdgpu_job_free(job);
172 	return 0;
173 }
174 
175 static struct dma_fence *amdgpu_job_dependency(struct drm_sched_job *sched_job,
176 					       struct drm_sched_entity *s_entity)
177 {
178 	struct amdgpu_ring *ring = to_amdgpu_ring(s_entity->rq->sched);
179 	struct amdgpu_job *job = to_amdgpu_job(sched_job);
180 	struct amdgpu_vm *vm = job->vm;
181 	struct dma_fence *fence;
182 	bool explicit = false;
183 	int r;
184 
185 	fence = amdgpu_sync_get_fence(&job->sync, &explicit);
186 	if (fence && explicit) {
187 		if (drm_sched_dependency_optimized(fence, s_entity)) {
188 			r = amdgpu_sync_fence(ring->adev, &job->sched_sync,
189 					      fence, false);
190 			if (r)
191 				DRM_ERROR("Error adding fence (%d)\n", r);
192 		}
193 	}
194 
195 	while (fence == NULL && vm && !job->vmid) {
196 		r = amdgpu_vmid_grab(vm, ring, &job->sync,
197 				     &job->base.s_fence->finished,
198 				     job);
199 		if (r)
200 			DRM_ERROR("Error getting VM ID (%d)\n", r);
201 
202 		fence = amdgpu_sync_get_fence(&job->sync, NULL);
203 	}
204 
205 	return fence;
206 }
207 
208 static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
209 {
210 	struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
211 	struct dma_fence *fence = NULL, *finished;
212 	struct amdgpu_job *job;
213 	int r;
214 
215 	job = to_amdgpu_job(sched_job);
216 	finished = &job->base.s_fence->finished;
217 
218 	BUG_ON(amdgpu_sync_peek_fence(&job->sync, NULL));
219 
220 	trace_amdgpu_sched_run_job(job);
221 
222 	if (job->vram_lost_counter != atomic_read(&ring->adev->vram_lost_counter))
223 		dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */
224 
225 	if (finished->error < 0) {
226 		DRM_INFO("Skip scheduling IBs!\n");
227 	} else {
228 		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs, job,
229 				       &fence);
230 		if (r)
231 			DRM_ERROR("Error scheduling IBs (%d)\n", r);
232 	}
233 	/* if gpu reset, hw fence will be replaced here */
234 	dma_fence_put(job->fence);
235 	job->fence = dma_fence_get(fence);
236 
237 	amdgpu_job_free_resources(job);
238 	return fence;
239 }
240 
241 const struct drm_sched_backend_ops amdgpu_sched_ops = {
242 	.dependency = amdgpu_job_dependency,
243 	.run_job = amdgpu_job_run,
244 	.timedout_job = amdgpu_job_timedout,
245 	.free_job = amdgpu_job_free_cb
246 };
247