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_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 
43 int 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 
69 int 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 
87 void 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 
99 static 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 
110 void 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 
120 int 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 
144 static 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 
176 static 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 
214 const 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 };
220