Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2017 Etnaviv Project
4 */
5
6#include <linux/moduleparam.h>
7
8#include "etnaviv_drv.h"
9#include "etnaviv_dump.h"
10#include "etnaviv_gem.h"
11#include "etnaviv_gpu.h"
12#include "etnaviv_sched.h"
13#include "state.xml.h"
14#include "state_hi.xml.h"
15
16static int etnaviv_job_hang_limit = 0;
17module_param_named(job_hang_limit, etnaviv_job_hang_limit, int , 0444);
18static int etnaviv_hw_jobs_limit = 4;
19module_param_named(hw_job_limit, etnaviv_hw_jobs_limit, int , 0444);
20
21static struct dma_fence *etnaviv_sched_run_job(struct drm_sched_job *sched_job)
22{
23 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
24 struct dma_fence *fence = NULL;
25
26 if (likely(!sched_job->s_fence->finished.error))
27 fence = etnaviv_gpu_submit(submit);
28 else
29 dev_dbg(submit->gpu->dev, "skipping bad job\n");
30
31 return fence;
32}
33
34static enum drm_gpu_sched_stat etnaviv_sched_timedout_job(struct drm_sched_job
35 *sched_job)
36{
37 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
38 struct etnaviv_gpu *gpu = submit->gpu;
39 u32 dma_addr, primid = 0;
40 int change;
41
42 /*
43 * If the GPU managed to complete this jobs fence, the timout is
44 * spurious. Bail out.
45 */
46 if (dma_fence_is_signaled(submit->out_fence))
47 goto out_no_timeout;
48
49 /*
50 * If the GPU is still making forward progress on the front-end (which
51 * should never loop) we shift out the timeout to give it a chance to
52 * finish the job.
53 */
54 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
55 change = dma_addr - gpu->hangcheck_dma_addr;
56 if (submit->exec_state == ETNA_PIPE_3D) {
57 /* guard against concurrent usage from perfmon_sample */
58 mutex_lock(&gpu->lock);
59 gpu_write(gpu, VIVS_MC_PROFILE_CONFIG0,
60 VIVS_MC_PROFILE_CONFIG0_FE_CURRENT_PRIM <<
61 VIVS_MC_PROFILE_CONFIG0_FE__SHIFT);
62 primid = gpu_read(gpu, VIVS_MC_PROFILE_FE_READ);
63 mutex_unlock(&gpu->lock);
64 }
65 if (gpu->state == ETNA_GPU_STATE_RUNNING &&
66 (gpu->completed_fence != gpu->hangcheck_fence ||
67 change < 0 || change > 16 ||
68 (submit->exec_state == ETNA_PIPE_3D &&
69 gpu->hangcheck_primid != primid))) {
70 gpu->hangcheck_dma_addr = dma_addr;
71 gpu->hangcheck_primid = primid;
72 gpu->hangcheck_fence = gpu->completed_fence;
73 goto out_no_timeout;
74 }
75
76 /* block scheduler */
77 drm_sched_stop(&gpu->sched, sched_job);
78
79 if(sched_job)
80 drm_sched_increase_karma(sched_job);
81
82 /* get the GPU back into the init state */
83 etnaviv_core_dump(submit);
84 etnaviv_gpu_recover_hang(submit);
85
86 drm_sched_resubmit_jobs(&gpu->sched);
87
88 drm_sched_start(&gpu->sched, 0);
89 return DRM_GPU_SCHED_STAT_NOMINAL;
90
91out_no_timeout:
92 list_add(&sched_job->list, &sched_job->sched->pending_list);
93 return DRM_GPU_SCHED_STAT_NOMINAL;
94}
95
96static void etnaviv_sched_free_job(struct drm_sched_job *sched_job)
97{
98 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
99
100 drm_sched_job_cleanup(sched_job);
101
102 etnaviv_submit_put(submit);
103}
104
105static const struct drm_sched_backend_ops etnaviv_sched_ops = {
106 .run_job = etnaviv_sched_run_job,
107 .timedout_job = etnaviv_sched_timedout_job,
108 .free_job = etnaviv_sched_free_job,
109};
110
111int etnaviv_sched_push_job(struct etnaviv_gem_submit *submit)
112{
113 struct etnaviv_gpu *gpu = submit->gpu;
114 int ret;
115
116 /*
117 * Hold the sched lock across the whole operation to avoid jobs being
118 * pushed out of order with regard to their sched fence seqnos as
119 * allocated in drm_sched_job_arm.
120 */
121 mutex_lock(&gpu->sched_lock);
122
123 drm_sched_job_arm(&submit->sched_job);
124
125 submit->out_fence = dma_fence_get(&submit->sched_job.s_fence->finished);
126 ret = xa_alloc_cyclic(&gpu->user_fences, &submit->out_fence_id,
127 submit->out_fence, xa_limit_32b,
128 &gpu->next_user_fence, GFP_KERNEL);
129 if (ret < 0) {
130 drm_sched_job_cleanup(&submit->sched_job);
131 goto out_unlock;
132 }
133
134 /* the scheduler holds on to the job now */
135 kref_get(&submit->refcount);
136
137 drm_sched_entity_push_job(&submit->sched_job);
138
139out_unlock:
140 mutex_unlock(&gpu->sched_lock);
141
142 return ret;
143}
144
145int etnaviv_sched_init(struct etnaviv_gpu *gpu)
146{
147 int ret;
148
149 ret = drm_sched_init(&gpu->sched, &etnaviv_sched_ops, NULL,
150 DRM_SCHED_PRIORITY_COUNT,
151 etnaviv_hw_jobs_limit, etnaviv_job_hang_limit,
152 msecs_to_jiffies(500), NULL, NULL,
153 dev_name(gpu->dev), gpu->dev);
154 if (ret)
155 return ret;
156
157 return 0;
158}
159
160void etnaviv_sched_fini(struct etnaviv_gpu *gpu)
161{
162 drm_sched_fini(&gpu->sched);
163}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2017 Etnaviv Project
4 */
5
6#include <linux/moduleparam.h>
7
8#include "etnaviv_drv.h"
9#include "etnaviv_dump.h"
10#include "etnaviv_gem.h"
11#include "etnaviv_gpu.h"
12#include "etnaviv_sched.h"
13#include "state.xml.h"
14
15static int etnaviv_job_hang_limit = 0;
16module_param_named(job_hang_limit, etnaviv_job_hang_limit, int , 0444);
17static int etnaviv_hw_jobs_limit = 4;
18module_param_named(hw_job_limit, etnaviv_hw_jobs_limit, int , 0444);
19
20static struct dma_fence *etnaviv_sched_run_job(struct drm_sched_job *sched_job)
21{
22 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
23 struct dma_fence *fence = NULL;
24
25 if (likely(!sched_job->s_fence->finished.error))
26 fence = etnaviv_gpu_submit(submit);
27 else
28 dev_dbg(submit->gpu->dev, "skipping bad job\n");
29
30 return fence;
31}
32
33static enum drm_gpu_sched_stat etnaviv_sched_timedout_job(struct drm_sched_job
34 *sched_job)
35{
36 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
37 struct etnaviv_gpu *gpu = submit->gpu;
38 u32 dma_addr;
39 int change;
40
41 /* block scheduler */
42 drm_sched_stop(&gpu->sched, sched_job);
43
44 /*
45 * If the GPU managed to complete this jobs fence, the timout is
46 * spurious. Bail out.
47 */
48 if (dma_fence_is_signaled(submit->out_fence))
49 goto out_no_timeout;
50
51 /*
52 * If the GPU is still making forward progress on the front-end (which
53 * should never loop) we shift out the timeout to give it a chance to
54 * finish the job.
55 */
56 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
57 change = dma_addr - gpu->hangcheck_dma_addr;
58 if (gpu->state == ETNA_GPU_STATE_RUNNING &&
59 (gpu->completed_fence != gpu->hangcheck_fence ||
60 change < 0 || change > 16)) {
61 gpu->hangcheck_dma_addr = dma_addr;
62 gpu->hangcheck_fence = gpu->completed_fence;
63 goto out_no_timeout;
64 }
65
66 if(sched_job)
67 drm_sched_increase_karma(sched_job);
68
69 /* get the GPU back into the init state */
70 etnaviv_core_dump(submit);
71 etnaviv_gpu_recover_hang(submit);
72
73 drm_sched_resubmit_jobs(&gpu->sched);
74
75 drm_sched_start(&gpu->sched, true);
76 return DRM_GPU_SCHED_STAT_NOMINAL;
77
78out_no_timeout:
79 /* restart scheduler after GPU is usable again */
80 drm_sched_start(&gpu->sched, true);
81 return DRM_GPU_SCHED_STAT_NOMINAL;
82}
83
84static void etnaviv_sched_free_job(struct drm_sched_job *sched_job)
85{
86 struct etnaviv_gem_submit *submit = to_etnaviv_submit(sched_job);
87
88 drm_sched_job_cleanup(sched_job);
89
90 etnaviv_submit_put(submit);
91}
92
93static const struct drm_sched_backend_ops etnaviv_sched_ops = {
94 .run_job = etnaviv_sched_run_job,
95 .timedout_job = etnaviv_sched_timedout_job,
96 .free_job = etnaviv_sched_free_job,
97};
98
99int etnaviv_sched_push_job(struct etnaviv_gem_submit *submit)
100{
101 struct etnaviv_gpu *gpu = submit->gpu;
102 int ret;
103
104 /*
105 * Hold the sched lock across the whole operation to avoid jobs being
106 * pushed out of order with regard to their sched fence seqnos as
107 * allocated in drm_sched_job_arm.
108 */
109 mutex_lock(&gpu->sched_lock);
110
111 drm_sched_job_arm(&submit->sched_job);
112
113 submit->out_fence = dma_fence_get(&submit->sched_job.s_fence->finished);
114 ret = xa_alloc_cyclic(&gpu->user_fences, &submit->out_fence_id,
115 submit->out_fence, xa_limit_32b,
116 &gpu->next_user_fence, GFP_KERNEL);
117 if (ret < 0) {
118 drm_sched_job_cleanup(&submit->sched_job);
119 goto out_unlock;
120 }
121
122 /* the scheduler holds on to the job now */
123 kref_get(&submit->refcount);
124
125 drm_sched_entity_push_job(&submit->sched_job);
126
127out_unlock:
128 mutex_unlock(&gpu->sched_lock);
129
130 return ret;
131}
132
133int etnaviv_sched_init(struct etnaviv_gpu *gpu)
134{
135 int ret;
136
137 ret = drm_sched_init(&gpu->sched, &etnaviv_sched_ops, NULL,
138 DRM_SCHED_PRIORITY_COUNT,
139 etnaviv_hw_jobs_limit, etnaviv_job_hang_limit,
140 msecs_to_jiffies(500), NULL, NULL,
141 dev_name(gpu->dev), gpu->dev);
142 if (ret)
143 return ret;
144
145 return 0;
146}
147
148void etnaviv_sched_fini(struct etnaviv_gpu *gpu)
149{
150 drm_sched_fini(&gpu->sched);
151}