xref: /openbmc/linux/drivers/gpu/drm/v3d/v3d_sched.c (revision f1575595)
1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2018 Broadcom */
3 
4 /**
5  * DOC: Broadcom V3D scheduling
6  *
7  * The shared DRM GPU scheduler is used to coordinate submitting jobs
8  * to the hardware.  Each DRM fd (roughly a client process) gets its
9  * own scheduler entity, which will process jobs in order.  The GPU
10  * scheduler will round-robin between clients to submit the next job.
11  *
12  * For simplicity, and in order to keep latency low for interactive
13  * jobs when bulk background jobs are queued up, we submit a new job
14  * to the HW only when it has completed the last one, instead of
15  * filling up the CT[01]Q FIFOs with jobs.  Similarly, we use
16  * v3d_job_dependency() to manage the dependency between bin and
17  * render, instead of having the clients submit jobs using the HW's
18  * semaphores to interlock between them.
19  */
20 
21 #include <linux/kthread.h>
22 
23 #include "v3d_drv.h"
24 #include "v3d_regs.h"
25 #include "v3d_trace.h"
26 
27 static struct v3d_job *
28 to_v3d_job(struct drm_sched_job *sched_job)
29 {
30 	return container_of(sched_job, struct v3d_job, base);
31 }
32 
33 static void
34 v3d_job_free(struct drm_sched_job *sched_job)
35 {
36 	struct v3d_job *job = to_v3d_job(sched_job);
37 
38 	v3d_exec_put(job->exec);
39 }
40 
41 /**
42  * Returns the fences that the bin job depends on, one by one.
43  * v3d_job_run() won't be called until all of them have been signaled.
44  */
45 static struct dma_fence *
46 v3d_job_dependency(struct drm_sched_job *sched_job,
47 		   struct drm_sched_entity *s_entity)
48 {
49 	struct v3d_job *job = to_v3d_job(sched_job);
50 	struct v3d_exec_info *exec = job->exec;
51 	enum v3d_queue q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
52 	struct dma_fence *fence;
53 
54 	fence = job->in_fence;
55 	if (fence) {
56 		job->in_fence = NULL;
57 		return fence;
58 	}
59 
60 	if (q == V3D_RENDER) {
61 		/* If we had a bin job, the render job definitely depends on
62 		 * it. We first have to wait for bin to be scheduled, so that
63 		 * its done_fence is created.
64 		 */
65 		fence = exec->bin_done_fence;
66 		if (fence) {
67 			exec->bin_done_fence = NULL;
68 			return fence;
69 		}
70 	}
71 
72 	/* XXX: Wait on a fence for switching the GMP if necessary,
73 	 * and then do so.
74 	 */
75 
76 	return fence;
77 }
78 
79 static struct dma_fence *v3d_job_run(struct drm_sched_job *sched_job)
80 {
81 	struct v3d_job *job = to_v3d_job(sched_job);
82 	struct v3d_exec_info *exec = job->exec;
83 	enum v3d_queue q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
84 	struct v3d_dev *v3d = exec->v3d;
85 	struct drm_device *dev = &v3d->drm;
86 	struct dma_fence *fence;
87 	unsigned long irqflags;
88 
89 	if (unlikely(job->base.s_fence->finished.error))
90 		return NULL;
91 
92 	/* Lock required around bin_job update vs
93 	 * v3d_overflow_mem_work().
94 	 */
95 	spin_lock_irqsave(&v3d->job_lock, irqflags);
96 	if (q == V3D_BIN) {
97 		v3d->bin_job = job->exec;
98 
99 		/* Clear out the overflow allocation, so we don't
100 		 * reuse the overflow attached to a previous job.
101 		 */
102 		V3D_CORE_WRITE(0, V3D_PTB_BPOS, 0);
103 	} else {
104 		v3d->render_job = job->exec;
105 	}
106 	spin_unlock_irqrestore(&v3d->job_lock, irqflags);
107 
108 	/* Can we avoid this flush when q==RENDER?  We need to be
109 	 * careful of scheduling, though -- imagine job0 rendering to
110 	 * texture and job1 reading, and them being executed as bin0,
111 	 * bin1, render0, render1, so that render1's flush at bin time
112 	 * wasn't enough.
113 	 */
114 	v3d_invalidate_caches(v3d);
115 
116 	fence = v3d_fence_create(v3d, q);
117 	if (IS_ERR(fence))
118 		return NULL;
119 
120 	if (job->done_fence)
121 		dma_fence_put(job->done_fence);
122 	job->done_fence = dma_fence_get(fence);
123 
124 	trace_v3d_submit_cl(dev, q == V3D_RENDER, to_v3d_fence(fence)->seqno,
125 			    job->start, job->end);
126 
127 	if (q == V3D_BIN) {
128 		if (exec->qma) {
129 			V3D_CORE_WRITE(0, V3D_CLE_CT0QMA, exec->qma);
130 			V3D_CORE_WRITE(0, V3D_CLE_CT0QMS, exec->qms);
131 		}
132 		if (exec->qts) {
133 			V3D_CORE_WRITE(0, V3D_CLE_CT0QTS,
134 				       V3D_CLE_CT0QTS_ENABLE |
135 				       exec->qts);
136 		}
137 	} else {
138 		/* XXX: Set the QCFG */
139 	}
140 
141 	/* Set the current and end address of the control list.
142 	 * Writing the end register is what starts the job.
143 	 */
144 	V3D_CORE_WRITE(0, V3D_CLE_CTNQBA(q), job->start);
145 	V3D_CORE_WRITE(0, V3D_CLE_CTNQEA(q), job->end);
146 
147 	return fence;
148 }
149 
150 static void
151 v3d_job_timedout(struct drm_sched_job *sched_job)
152 {
153 	struct v3d_job *job = to_v3d_job(sched_job);
154 	struct v3d_exec_info *exec = job->exec;
155 	struct v3d_dev *v3d = exec->v3d;
156 	enum v3d_queue job_q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
157 	enum v3d_queue q;
158 	u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(job_q));
159 	u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(job_q));
160 
161 	/* If the current address or return address have changed, then
162 	 * the GPU has probably made progress and we should delay the
163 	 * reset.  This could fail if the GPU got in an infinite loop
164 	 * in the CL, but that is pretty unlikely outside of an i-g-t
165 	 * testcase.
166 	 */
167 	if (job->timedout_ctca != ctca || job->timedout_ctra != ctra) {
168 		job->timedout_ctca = ctca;
169 		job->timedout_ctra = ctra;
170 
171 		schedule_delayed_work(&job->base.work_tdr,
172 				      job->base.sched->timeout);
173 		return;
174 	}
175 
176 	mutex_lock(&v3d->reset_lock);
177 
178 	/* block scheduler */
179 	for (q = 0; q < V3D_MAX_QUEUES; q++) {
180 		struct drm_gpu_scheduler *sched = &v3d->queue[q].sched;
181 
182 		kthread_park(sched->thread);
183 		drm_sched_hw_job_reset(sched, (sched_job->sched == sched ?
184 					       sched_job : NULL));
185 	}
186 
187 	/* get the GPU back into the init state */
188 	v3d_reset(v3d);
189 
190 	/* Unblock schedulers and restart their jobs. */
191 	for (q = 0; q < V3D_MAX_QUEUES; q++) {
192 		drm_sched_job_recovery(&v3d->queue[q].sched);
193 		kthread_unpark(v3d->queue[q].sched.thread);
194 	}
195 
196 	mutex_unlock(&v3d->reset_lock);
197 }
198 
199 static const struct drm_sched_backend_ops v3d_sched_ops = {
200 	.dependency = v3d_job_dependency,
201 	.run_job = v3d_job_run,
202 	.timedout_job = v3d_job_timedout,
203 	.free_job = v3d_job_free
204 };
205 
206 int
207 v3d_sched_init(struct v3d_dev *v3d)
208 {
209 	int hw_jobs_limit = 1;
210 	int job_hang_limit = 0;
211 	int hang_limit_ms = 500;
212 	int ret;
213 
214 	ret = drm_sched_init(&v3d->queue[V3D_BIN].sched,
215 			     &v3d_sched_ops,
216 			     hw_jobs_limit, job_hang_limit,
217 			     msecs_to_jiffies(hang_limit_ms),
218 			     "v3d_bin");
219 	if (ret) {
220 		dev_err(v3d->dev, "Failed to create bin scheduler: %d.", ret);
221 		return ret;
222 	}
223 
224 	ret = drm_sched_init(&v3d->queue[V3D_RENDER].sched,
225 			     &v3d_sched_ops,
226 			     hw_jobs_limit, job_hang_limit,
227 			     msecs_to_jiffies(hang_limit_ms),
228 			     "v3d_render");
229 	if (ret) {
230 		dev_err(v3d->dev, "Failed to create render scheduler: %d.",
231 			ret);
232 		drm_sched_fini(&v3d->queue[V3D_BIN].sched);
233 		return ret;
234 	}
235 
236 	return 0;
237 }
238 
239 void
240 v3d_sched_fini(struct v3d_dev *v3d)
241 {
242 	enum v3d_queue q;
243 
244 	for (q = 0; q < V3D_MAX_QUEUES; q++)
245 		drm_sched_fini(&v3d->queue[q].sched);
246 }
247