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
 27static struct v3d_job *
 28to_v3d_job(struct drm_sched_job *sched_job)
 29{
 30	return container_of(sched_job, struct v3d_job, base);
 31}
 32
 33static struct v3d_bin_job *
 34to_bin_job(struct drm_sched_job *sched_job)
 35{
 36	return container_of(sched_job, struct v3d_bin_job, base.base);
 37}
 38
 39static struct v3d_render_job *
 40to_render_job(struct drm_sched_job *sched_job)
 41{
 42	return container_of(sched_job, struct v3d_render_job, base.base);
 43}
 44
 45static struct v3d_tfu_job *
 46to_tfu_job(struct drm_sched_job *sched_job)
 47{
 48	return container_of(sched_job, struct v3d_tfu_job, base.base);
 49}
 50
 51static struct v3d_csd_job *
 52to_csd_job(struct drm_sched_job *sched_job)
 53{
 54	return container_of(sched_job, struct v3d_csd_job, base.base);
 55}
 56
 57static void
 58v3d_job_free(struct drm_sched_job *sched_job)
 59{
 60	struct v3d_job *job = to_v3d_job(sched_job);
 61
 62	drm_sched_job_cleanup(sched_job);
 63	v3d_job_put(job);
 64}
 65
 66/**
 67 * Returns the fences that the job depends on, one by one.
 68 *
 69 * If placed in the scheduler's .dependency method, the corresponding
 70 * .run_job won't be called until all of them have been signaled.
 71 */
 72static struct dma_fence *
 73v3d_job_dependency(struct drm_sched_job *sched_job,
 74		   struct drm_sched_entity *s_entity)
 75{
 76	struct v3d_job *job = to_v3d_job(sched_job);
 77
 78	/* XXX: Wait on a fence for switching the GMP if necessary,
 79	 * and then do so.
 80	 */
 81
 82	if (!xa_empty(&job->deps))
 83		return xa_erase(&job->deps, job->last_dep++);
 84
 85	return NULL;
 86}
 87
 88static struct dma_fence *v3d_bin_job_run(struct drm_sched_job *sched_job)
 89{
 90	struct v3d_bin_job *job = to_bin_job(sched_job);
 91	struct v3d_dev *v3d = job->base.v3d;
 92	struct drm_device *dev = &v3d->drm;
 93	struct dma_fence *fence;
 94	unsigned long irqflags;
 95
 96	if (unlikely(job->base.base.s_fence->finished.error))
 97		return NULL;
 98
 99	/* Lock required around bin_job update vs
100	 * v3d_overflow_mem_work().
101	 */
102	spin_lock_irqsave(&v3d->job_lock, irqflags);
103	v3d->bin_job = job;
104	/* Clear out the overflow allocation, so we don't
105	 * reuse the overflow attached to a previous job.
106	 */
107	V3D_CORE_WRITE(0, V3D_PTB_BPOS, 0);
108	spin_unlock_irqrestore(&v3d->job_lock, irqflags);
109
110	v3d_invalidate_caches(v3d);
111
112	fence = v3d_fence_create(v3d, V3D_BIN);
113	if (IS_ERR(fence))
114		return NULL;
115
116	if (job->base.irq_fence)
117		dma_fence_put(job->base.irq_fence);
118	job->base.irq_fence = dma_fence_get(fence);
119
120	trace_v3d_submit_cl(dev, false, to_v3d_fence(fence)->seqno,
121			    job->start, job->end);
122
123	/* Set the current and end address of the control list.
124	 * Writing the end register is what starts the job.
125	 */
126	if (job->qma) {
127		V3D_CORE_WRITE(0, V3D_CLE_CT0QMA, job->qma);
128		V3D_CORE_WRITE(0, V3D_CLE_CT0QMS, job->qms);
129	}
130	if (job->qts) {
131		V3D_CORE_WRITE(0, V3D_CLE_CT0QTS,
132			       V3D_CLE_CT0QTS_ENABLE |
133			       job->qts);
134	}
135	V3D_CORE_WRITE(0, V3D_CLE_CT0QBA, job->start);
136	V3D_CORE_WRITE(0, V3D_CLE_CT0QEA, job->end);
137
138	return fence;
139}
140
141static struct dma_fence *v3d_render_job_run(struct drm_sched_job *sched_job)
142{
143	struct v3d_render_job *job = to_render_job(sched_job);
144	struct v3d_dev *v3d = job->base.v3d;
145	struct drm_device *dev = &v3d->drm;
146	struct dma_fence *fence;
147
148	if (unlikely(job->base.base.s_fence->finished.error))
149		return NULL;
150
151	v3d->render_job = job;
152
153	/* Can we avoid this flush?  We need to be careful of
154	 * scheduling, though -- imagine job0 rendering to texture and
155	 * job1 reading, and them being executed as bin0, bin1,
156	 * render0, render1, so that render1's flush at bin time
157	 * wasn't enough.
158	 */
159	v3d_invalidate_caches(v3d);
160
161	fence = v3d_fence_create(v3d, V3D_RENDER);
162	if (IS_ERR(fence))
163		return NULL;
164
165	if (job->base.irq_fence)
166		dma_fence_put(job->base.irq_fence);
167	job->base.irq_fence = dma_fence_get(fence);
168
169	trace_v3d_submit_cl(dev, true, to_v3d_fence(fence)->seqno,
170			    job->start, job->end);
171
172	/* XXX: Set the QCFG */
173
174	/* Set the current and end address of the control list.
175	 * Writing the end register is what starts the job.
176	 */
177	V3D_CORE_WRITE(0, V3D_CLE_CT1QBA, job->start);
178	V3D_CORE_WRITE(0, V3D_CLE_CT1QEA, job->end);
179
180	return fence;
181}
182
183static struct dma_fence *
184v3d_tfu_job_run(struct drm_sched_job *sched_job)
185{
186	struct v3d_tfu_job *job = to_tfu_job(sched_job);
187	struct v3d_dev *v3d = job->base.v3d;
188	struct drm_device *dev = &v3d->drm;
189	struct dma_fence *fence;
190
191	fence = v3d_fence_create(v3d, V3D_TFU);
192	if (IS_ERR(fence))
193		return NULL;
194
195	v3d->tfu_job = job;
196	if (job->base.irq_fence)
197		dma_fence_put(job->base.irq_fence);
198	job->base.irq_fence = dma_fence_get(fence);
199
200	trace_v3d_submit_tfu(dev, to_v3d_fence(fence)->seqno);
201
202	V3D_WRITE(V3D_TFU_IIA, job->args.iia);
203	V3D_WRITE(V3D_TFU_IIS, job->args.iis);
204	V3D_WRITE(V3D_TFU_ICA, job->args.ica);
205	V3D_WRITE(V3D_TFU_IUA, job->args.iua);
206	V3D_WRITE(V3D_TFU_IOA, job->args.ioa);
207	V3D_WRITE(V3D_TFU_IOS, job->args.ios);
208	V3D_WRITE(V3D_TFU_COEF0, job->args.coef[0]);
209	if (job->args.coef[0] & V3D_TFU_COEF0_USECOEF) {
210		V3D_WRITE(V3D_TFU_COEF1, job->args.coef[1]);
211		V3D_WRITE(V3D_TFU_COEF2, job->args.coef[2]);
212		V3D_WRITE(V3D_TFU_COEF3, job->args.coef[3]);
213	}
214	/* ICFG kicks off the job. */
215	V3D_WRITE(V3D_TFU_ICFG, job->args.icfg | V3D_TFU_ICFG_IOC);
216
217	return fence;
218}
219
220static struct dma_fence *
221v3d_csd_job_run(struct drm_sched_job *sched_job)
222{
223	struct v3d_csd_job *job = to_csd_job(sched_job);
224	struct v3d_dev *v3d = job->base.v3d;
225	struct drm_device *dev = &v3d->drm;
226	struct dma_fence *fence;
227	int i;
228
229	v3d->csd_job = job;
230
231	v3d_invalidate_caches(v3d);
232
233	fence = v3d_fence_create(v3d, V3D_CSD);
234	if (IS_ERR(fence))
235		return NULL;
236
237	if (job->base.irq_fence)
238		dma_fence_put(job->base.irq_fence);
239	job->base.irq_fence = dma_fence_get(fence);
240
241	trace_v3d_submit_csd(dev, to_v3d_fence(fence)->seqno);
242
243	for (i = 1; i <= 6; i++)
244		V3D_CORE_WRITE(0, V3D_CSD_QUEUED_CFG0 + 4 * i, job->args.cfg[i]);
245	/* CFG0 write kicks off the job. */
246	V3D_CORE_WRITE(0, V3D_CSD_QUEUED_CFG0, job->args.cfg[0]);
247
248	return fence;
249}
250
251static struct dma_fence *
252v3d_cache_clean_job_run(struct drm_sched_job *sched_job)
253{
254	struct v3d_job *job = to_v3d_job(sched_job);
255	struct v3d_dev *v3d = job->v3d;
256
257	v3d_clean_caches(v3d);
258
259	return NULL;
260}
261
262static void
263v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job *sched_job)
264{
265	enum v3d_queue q;
266
267	mutex_lock(&v3d->reset_lock);
268
269	/* block scheduler */
270	for (q = 0; q < V3D_MAX_QUEUES; q++)
271		drm_sched_stop(&v3d->queue[q].sched, sched_job);
272
273	if (sched_job)
274		drm_sched_increase_karma(sched_job);
275
276	/* get the GPU back into the init state */
277	v3d_reset(v3d);
278
279	for (q = 0; q < V3D_MAX_QUEUES; q++)
280		drm_sched_resubmit_jobs(&v3d->queue[q].sched);
281
282	/* Unblock schedulers and restart their jobs. */
283	for (q = 0; q < V3D_MAX_QUEUES; q++) {
284		drm_sched_start(&v3d->queue[q].sched, true);
285	}
286
287	mutex_unlock(&v3d->reset_lock);
288}
289
290/* If the current address or return address have changed, then the GPU
291 * has probably made progress and we should delay the reset.  This
292 * could fail if the GPU got in an infinite loop in the CL, but that
293 * is pretty unlikely outside of an i-g-t testcase.
294 */
295static void
296v3d_cl_job_timedout(struct drm_sched_job *sched_job, enum v3d_queue q,
297		    u32 *timedout_ctca, u32 *timedout_ctra)
298{
299	struct v3d_job *job = to_v3d_job(sched_job);
300	struct v3d_dev *v3d = job->v3d;
301	u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(q));
302	u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(q));
303
304	if (*timedout_ctca != ctca || *timedout_ctra != ctra) {
305		*timedout_ctca = ctca;
306		*timedout_ctra = ctra;
307		return;
308	}
309
310	v3d_gpu_reset_for_timeout(v3d, sched_job);
311}
312
313static void
314v3d_bin_job_timedout(struct drm_sched_job *sched_job)
315{
316	struct v3d_bin_job *job = to_bin_job(sched_job);
317
318	v3d_cl_job_timedout(sched_job, V3D_BIN,
319			    &job->timedout_ctca, &job->timedout_ctra);
320}
321
322static void
323v3d_render_job_timedout(struct drm_sched_job *sched_job)
324{
325	struct v3d_render_job *job = to_render_job(sched_job);
326
327	v3d_cl_job_timedout(sched_job, V3D_RENDER,
328			    &job->timedout_ctca, &job->timedout_ctra);
329}
330
331static void
332v3d_generic_job_timedout(struct drm_sched_job *sched_job)
333{
334	struct v3d_job *job = to_v3d_job(sched_job);
335
336	v3d_gpu_reset_for_timeout(job->v3d, sched_job);
337}
338
339static void
340v3d_csd_job_timedout(struct drm_sched_job *sched_job)
341{
342	struct v3d_csd_job *job = to_csd_job(sched_job);
343	struct v3d_dev *v3d = job->base.v3d;
344	u32 batches = V3D_CORE_READ(0, V3D_CSD_CURRENT_CFG4);
345
346	/* If we've made progress, skip reset and let the timer get
347	 * rearmed.
348	 */
349	if (job->timedout_batches != batches) {
350		job->timedout_batches = batches;
351		return;
352	}
353
354	v3d_gpu_reset_for_timeout(v3d, sched_job);
355}
356
357static const struct drm_sched_backend_ops v3d_bin_sched_ops = {
358	.dependency = v3d_job_dependency,
359	.run_job = v3d_bin_job_run,
360	.timedout_job = v3d_bin_job_timedout,
361	.free_job = v3d_job_free,
362};
363
364static const struct drm_sched_backend_ops v3d_render_sched_ops = {
365	.dependency = v3d_job_dependency,
366	.run_job = v3d_render_job_run,
367	.timedout_job = v3d_render_job_timedout,
368	.free_job = v3d_job_free,
369};
370
371static const struct drm_sched_backend_ops v3d_tfu_sched_ops = {
372	.dependency = v3d_job_dependency,
373	.run_job = v3d_tfu_job_run,
374	.timedout_job = v3d_generic_job_timedout,
375	.free_job = v3d_job_free,
376};
377
378static const struct drm_sched_backend_ops v3d_csd_sched_ops = {
379	.dependency = v3d_job_dependency,
380	.run_job = v3d_csd_job_run,
381	.timedout_job = v3d_csd_job_timedout,
382	.free_job = v3d_job_free
383};
384
385static const struct drm_sched_backend_ops v3d_cache_clean_sched_ops = {
386	.dependency = v3d_job_dependency,
387	.run_job = v3d_cache_clean_job_run,
388	.timedout_job = v3d_generic_job_timedout,
389	.free_job = v3d_job_free
390};
391
392int
393v3d_sched_init(struct v3d_dev *v3d)
394{
395	int hw_jobs_limit = 1;
396	int job_hang_limit = 0;
397	int hang_limit_ms = 500;
398	int ret;
399
400	ret = drm_sched_init(&v3d->queue[V3D_BIN].sched,
401			     &v3d_bin_sched_ops,
402			     hw_jobs_limit, job_hang_limit,
403			     msecs_to_jiffies(hang_limit_ms),
404			     "v3d_bin");
405	if (ret) {
406		dev_err(v3d->drm.dev, "Failed to create bin scheduler: %d.", ret);
407		return ret;
408	}
409
410	ret = drm_sched_init(&v3d->queue[V3D_RENDER].sched,
411			     &v3d_render_sched_ops,
412			     hw_jobs_limit, job_hang_limit,
413			     msecs_to_jiffies(hang_limit_ms),
414			     "v3d_render");
415	if (ret) {
416		dev_err(v3d->drm.dev, "Failed to create render scheduler: %d.",
417			ret);
418		v3d_sched_fini(v3d);
419		return ret;
420	}
421
422	ret = drm_sched_init(&v3d->queue[V3D_TFU].sched,
423			     &v3d_tfu_sched_ops,
424			     hw_jobs_limit, job_hang_limit,
425			     msecs_to_jiffies(hang_limit_ms),
426			     "v3d_tfu");
427	if (ret) {
428		dev_err(v3d->drm.dev, "Failed to create TFU scheduler: %d.",
429			ret);
430		v3d_sched_fini(v3d);
431		return ret;
432	}
433
434	if (v3d_has_csd(v3d)) {
435		ret = drm_sched_init(&v3d->queue[V3D_CSD].sched,
436				     &v3d_csd_sched_ops,
437				     hw_jobs_limit, job_hang_limit,
438				     msecs_to_jiffies(hang_limit_ms),
439				     "v3d_csd");
440		if (ret) {
441			dev_err(v3d->drm.dev, "Failed to create CSD scheduler: %d.",
442				ret);
443			v3d_sched_fini(v3d);
444			return ret;
445		}
446
447		ret = drm_sched_init(&v3d->queue[V3D_CACHE_CLEAN].sched,
448				     &v3d_cache_clean_sched_ops,
449				     hw_jobs_limit, job_hang_limit,
450				     msecs_to_jiffies(hang_limit_ms),
451				     "v3d_cache_clean");
452		if (ret) {
453			dev_err(v3d->drm.dev, "Failed to create CACHE_CLEAN scheduler: %d.",
454				ret);
455			v3d_sched_fini(v3d);
456			return ret;
457		}
458	}
459
460	return 0;
461}
462
463void
464v3d_sched_fini(struct v3d_dev *v3d)
465{
466	enum v3d_queue q;
467
468	for (q = 0; q < V3D_MAX_QUEUES; q++) {
469		if (v3d->queue[q].sched.ready)
470			drm_sched_fini(&v3d->queue[q].sched);
471	}
472}