1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright 2019 Linaro, Ltd, Rob Herring <robh@kernel.org> */
3 /* Copyright 2019 Collabora ltd. */
4 #include <linux/delay.h>
5 #include <linux/interrupt.h>
7 #include <linux/platform_device.h>
8 #include <linux/pm_runtime.h>
9 #include <linux/dma-resv.h>
10 #include <drm/gpu_scheduler.h>
11 #include <drm/panfrost_drm.h>
13 #include "panfrost_device.h"
14 #include "panfrost_devfreq.h"
15 #include "panfrost_job.h"
16 #include "panfrost_features.h"
17 #include "panfrost_issues.h"
18 #include "panfrost_gem.h"
19 #include "panfrost_regs.h"
20 #include "panfrost_gpu.h"
21 #include "panfrost_mmu.h"
23 #define job_write(dev, reg, data) writel(data, dev->iomem + (reg))
24 #define job_read(dev, reg) readl(dev->iomem + (reg))
26 struct panfrost_queue_state {
27 struct drm_gpu_scheduler sched;
33 struct panfrost_job_slot {
34 struct panfrost_queue_state queue[NUM_JOB_SLOTS];
38 static struct panfrost_job *
39 to_panfrost_job(struct drm_sched_job *sched_job)
41 return container_of(sched_job, struct panfrost_job, base);
44 struct panfrost_fence {
45 struct dma_fence base;
46 struct drm_device *dev;
47 /* panfrost seqno for signaled() test */
52 static inline struct panfrost_fence *
53 to_panfrost_fence(struct dma_fence *fence)
55 return (struct panfrost_fence *)fence;
58 static const char *panfrost_fence_get_driver_name(struct dma_fence *fence)
63 static const char *panfrost_fence_get_timeline_name(struct dma_fence *fence)
65 struct panfrost_fence *f = to_panfrost_fence(fence);
69 return "panfrost-js-0";
71 return "panfrost-js-1";
73 return "panfrost-js-2";
79 static const struct dma_fence_ops panfrost_fence_ops = {
80 .get_driver_name = panfrost_fence_get_driver_name,
81 .get_timeline_name = panfrost_fence_get_timeline_name,
84 static struct dma_fence *panfrost_fence_create(struct panfrost_device *pfdev, int js_num)
86 struct panfrost_fence *fence;
87 struct panfrost_job_slot *js = pfdev->js;
89 fence = kzalloc(sizeof(*fence), GFP_KERNEL);
91 return ERR_PTR(-ENOMEM);
93 fence->dev = pfdev->ddev;
94 fence->queue = js_num;
95 fence->seqno = ++js->queue[js_num].emit_seqno;
96 dma_fence_init(&fence->base, &panfrost_fence_ops, &js->job_lock,
97 js->queue[js_num].fence_context, fence->seqno);
102 static int panfrost_job_get_slot(struct panfrost_job *job)
104 /* JS0: fragment jobs.
105 * JS1: vertex/tiler jobs
108 if (job->requirements & PANFROST_JD_REQ_FS)
111 /* Not exposed to userspace yet */
113 if (job->requirements & PANFROST_JD_REQ_ONLY_COMPUTE) {
114 if ((job->requirements & PANFROST_JD_REQ_CORE_GRP_MASK) &&
115 (job->pfdev->features.nr_core_groups == 2))
117 if (panfrost_has_hw_issue(job->pfdev, HW_ISSUE_8987))
124 static void panfrost_job_write_affinity(struct panfrost_device *pfdev,
131 * Use all cores for now.
132 * Eventually we may need to support tiler only jobs and h/w with
133 * multiple (2) coherent core groups
135 affinity = pfdev->features.shader_present;
137 job_write(pfdev, JS_AFFINITY_NEXT_LO(js), affinity & 0xFFFFFFFF);
138 job_write(pfdev, JS_AFFINITY_NEXT_HI(js), affinity >> 32);
141 static void panfrost_job_hw_submit(struct panfrost_job *job, int js)
143 struct panfrost_device *pfdev = job->pfdev;
145 u64 jc_head = job->jc;
148 ret = pm_runtime_get_sync(pfdev->dev);
152 if (WARN_ON(job_read(pfdev, JS_COMMAND_NEXT(js)))) {
153 pm_runtime_put_sync_autosuspend(pfdev->dev);
157 cfg = panfrost_mmu_as_get(pfdev, &job->file_priv->mmu);
158 panfrost_devfreq_record_busy(pfdev);
160 job_write(pfdev, JS_HEAD_NEXT_LO(js), jc_head & 0xFFFFFFFF);
161 job_write(pfdev, JS_HEAD_NEXT_HI(js), jc_head >> 32);
163 panfrost_job_write_affinity(pfdev, job->requirements, js);
165 /* start MMU, medium priority, cache clean/flush on end, clean/flush on
167 cfg |= JS_CONFIG_THREAD_PRI(8) |
168 JS_CONFIG_START_FLUSH_CLEAN_INVALIDATE |
169 JS_CONFIG_END_FLUSH_CLEAN_INVALIDATE;
171 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION))
172 cfg |= JS_CONFIG_ENABLE_FLUSH_REDUCTION;
174 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10649))
175 cfg |= JS_CONFIG_START_MMU;
177 job_write(pfdev, JS_CONFIG_NEXT(js), cfg);
179 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION))
180 job_write(pfdev, JS_FLUSH_ID_NEXT(js), job->flush_id);
183 dev_dbg(pfdev->dev, "JS: Submitting atom %p to js[%d] with head=0x%llx",
186 job_write(pfdev, JS_COMMAND_NEXT(js), JS_COMMAND_START);
189 static void panfrost_acquire_object_fences(struct drm_gem_object **bos,
191 struct dma_fence **implicit_fences)
195 for (i = 0; i < bo_count; i++)
196 implicit_fences[i] = dma_resv_get_excl_rcu(bos[i]->resv);
199 static void panfrost_attach_object_fences(struct drm_gem_object **bos,
201 struct dma_fence *fence)
205 for (i = 0; i < bo_count; i++)
206 dma_resv_add_excl_fence(bos[i]->resv, fence);
209 int panfrost_job_push(struct panfrost_job *job)
211 struct panfrost_device *pfdev = job->pfdev;
212 int slot = panfrost_job_get_slot(job);
213 struct drm_sched_entity *entity = &job->file_priv->sched_entity[slot];
214 struct ww_acquire_ctx acquire_ctx;
217 mutex_lock(&pfdev->sched_lock);
219 ret = drm_gem_lock_reservations(job->bos, job->bo_count,
222 mutex_unlock(&pfdev->sched_lock);
226 ret = drm_sched_job_init(&job->base, entity, NULL);
228 mutex_unlock(&pfdev->sched_lock);
232 job->render_done_fence = dma_fence_get(&job->base.s_fence->finished);
234 kref_get(&job->refcount); /* put by scheduler job completion */
236 panfrost_acquire_object_fences(job->bos, job->bo_count,
237 job->implicit_fences);
239 drm_sched_entity_push_job(&job->base, entity);
241 mutex_unlock(&pfdev->sched_lock);
243 panfrost_attach_object_fences(job->bos, job->bo_count,
244 job->render_done_fence);
247 drm_gem_unlock_reservations(job->bos, job->bo_count, &acquire_ctx);
252 static void panfrost_job_cleanup(struct kref *ref)
254 struct panfrost_job *job = container_of(ref, struct panfrost_job,
258 if (job->in_fences) {
259 for (i = 0; i < job->in_fence_count; i++)
260 dma_fence_put(job->in_fences[i]);
261 kvfree(job->in_fences);
263 if (job->implicit_fences) {
264 for (i = 0; i < job->bo_count; i++)
265 dma_fence_put(job->implicit_fences[i]);
266 kvfree(job->implicit_fences);
268 dma_fence_put(job->done_fence);
269 dma_fence_put(job->render_done_fence);
272 for (i = 0; i < job->bo_count; i++)
273 panfrost_gem_mapping_put(job->mappings[i]);
274 kvfree(job->mappings);
278 struct panfrost_gem_object *bo;
280 for (i = 0; i < job->bo_count; i++) {
281 bo = to_panfrost_bo(job->bos[i]);
282 drm_gem_object_put_unlocked(job->bos[i]);
291 void panfrost_job_put(struct panfrost_job *job)
293 kref_put(&job->refcount, panfrost_job_cleanup);
296 static void panfrost_job_free(struct drm_sched_job *sched_job)
298 struct panfrost_job *job = to_panfrost_job(sched_job);
300 drm_sched_job_cleanup(sched_job);
302 panfrost_job_put(job);
305 static struct dma_fence *panfrost_job_dependency(struct drm_sched_job *sched_job,
306 struct drm_sched_entity *s_entity)
308 struct panfrost_job *job = to_panfrost_job(sched_job);
309 struct dma_fence *fence;
312 /* Explicit fences */
313 for (i = 0; i < job->in_fence_count; i++) {
314 if (job->in_fences[i]) {
315 fence = job->in_fences[i];
316 job->in_fences[i] = NULL;
321 /* Implicit fences, max. one per BO */
322 for (i = 0; i < job->bo_count; i++) {
323 if (job->implicit_fences[i]) {
324 fence = job->implicit_fences[i];
325 job->implicit_fences[i] = NULL;
333 static struct dma_fence *panfrost_job_run(struct drm_sched_job *sched_job)
335 struct panfrost_job *job = to_panfrost_job(sched_job);
336 struct panfrost_device *pfdev = job->pfdev;
337 int slot = panfrost_job_get_slot(job);
338 struct dma_fence *fence = NULL;
340 if (unlikely(job->base.s_fence->finished.error))
343 pfdev->jobs[slot] = job;
345 fence = panfrost_fence_create(pfdev, slot);
350 dma_fence_put(job->done_fence);
351 job->done_fence = dma_fence_get(fence);
353 panfrost_job_hw_submit(job, slot);
358 void panfrost_job_enable_interrupts(struct panfrost_device *pfdev)
363 for (j = 0; j < NUM_JOB_SLOTS; j++) {
364 irq_mask |= MK_JS_MASK(j);
367 job_write(pfdev, JOB_INT_CLEAR, irq_mask);
368 job_write(pfdev, JOB_INT_MASK, irq_mask);
371 static void panfrost_job_timedout(struct drm_sched_job *sched_job)
373 struct panfrost_job *job = to_panfrost_job(sched_job);
374 struct panfrost_device *pfdev = job->pfdev;
375 int js = panfrost_job_get_slot(job);
380 * If the GPU managed to complete this jobs fence, the timeout is
381 * spurious. Bail out.
383 if (dma_fence_is_signaled(job->done_fence))
386 dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",
388 job_read(pfdev, JS_CONFIG(js)),
389 job_read(pfdev, JS_STATUS(js)),
390 job_read(pfdev, JS_HEAD_LO(js)),
391 job_read(pfdev, JS_TAIL_LO(js)),
394 if (!mutex_trylock(&pfdev->reset_lock))
397 for (i = 0; i < NUM_JOB_SLOTS; i++) {
398 struct drm_gpu_scheduler *sched = &pfdev->js->queue[i].sched;
400 drm_sched_stop(sched, sched_job);
402 /* Ensure any timeouts on other slots have finished */
403 cancel_delayed_work_sync(&sched->work_tdr);
406 drm_sched_increase_karma(sched_job);
408 spin_lock_irqsave(&pfdev->js->job_lock, flags);
409 for (i = 0; i < NUM_JOB_SLOTS; i++) {
410 if (pfdev->jobs[i]) {
411 pm_runtime_put_noidle(pfdev->dev);
412 pfdev->jobs[i] = NULL;
415 spin_unlock_irqrestore(&pfdev->js->job_lock, flags);
417 panfrost_devfreq_record_idle(pfdev);
418 panfrost_device_reset(pfdev);
420 for (i = 0; i < NUM_JOB_SLOTS; i++)
421 drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched);
423 /* restart scheduler after GPU is usable again */
424 for (i = 0; i < NUM_JOB_SLOTS; i++)
425 drm_sched_start(&pfdev->js->queue[i].sched, true);
427 mutex_unlock(&pfdev->reset_lock);
430 static const struct drm_sched_backend_ops panfrost_sched_ops = {
431 .dependency = panfrost_job_dependency,
432 .run_job = panfrost_job_run,
433 .timedout_job = panfrost_job_timedout,
434 .free_job = panfrost_job_free
437 static irqreturn_t panfrost_job_irq_handler(int irq, void *data)
439 struct panfrost_device *pfdev = data;
440 u32 status = job_read(pfdev, JOB_INT_STAT);
443 dev_dbg(pfdev->dev, "jobslot irq status=%x\n", status);
448 pm_runtime_mark_last_busy(pfdev->dev);
450 for (j = 0; status; j++) {
451 u32 mask = MK_JS_MASK(j);
453 if (!(status & mask))
456 job_write(pfdev, JOB_INT_CLEAR, mask);
458 if (status & JOB_INT_MASK_ERR(j)) {
459 job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_NOP);
461 dev_err(pfdev->dev, "js fault, js=%d, status=%s, head=0x%x, tail=0x%x",
463 panfrost_exception_name(pfdev, job_read(pfdev, JS_STATUS(j))),
464 job_read(pfdev, JS_HEAD_LO(j)),
465 job_read(pfdev, JS_TAIL_LO(j)));
467 drm_sched_fault(&pfdev->js->queue[j].sched);
470 if (status & JOB_INT_MASK_DONE(j)) {
471 struct panfrost_job *job;
473 spin_lock(&pfdev->js->job_lock);
474 job = pfdev->jobs[j];
475 /* Only NULL if job timeout occurred */
477 pfdev->jobs[j] = NULL;
479 panfrost_mmu_as_put(pfdev, &job->file_priv->mmu);
480 panfrost_devfreq_record_idle(pfdev);
482 dma_fence_signal_locked(job->done_fence);
483 pm_runtime_put_autosuspend(pfdev->dev);
485 spin_unlock(&pfdev->js->job_lock);
494 int panfrost_job_init(struct panfrost_device *pfdev)
496 struct panfrost_job_slot *js;
499 pfdev->js = js = devm_kzalloc(pfdev->dev, sizeof(*js), GFP_KERNEL);
503 spin_lock_init(&js->job_lock);
505 irq = platform_get_irq_byname(to_platform_device(pfdev->dev), "job");
509 ret = devm_request_irq(pfdev->dev, irq, panfrost_job_irq_handler,
510 IRQF_SHARED, "job", pfdev);
512 dev_err(pfdev->dev, "failed to request job irq");
516 for (j = 0; j < NUM_JOB_SLOTS; j++) {
517 js->queue[j].fence_context = dma_fence_context_alloc(1);
519 ret = drm_sched_init(&js->queue[j].sched,
521 1, 0, msecs_to_jiffies(500),
524 dev_err(pfdev->dev, "Failed to create scheduler: %d.", ret);
529 panfrost_job_enable_interrupts(pfdev);
534 for (j--; j >= 0; j--)
535 drm_sched_fini(&js->queue[j].sched);
540 void panfrost_job_fini(struct panfrost_device *pfdev)
542 struct panfrost_job_slot *js = pfdev->js;
545 job_write(pfdev, JOB_INT_MASK, 0);
547 for (j = 0; j < NUM_JOB_SLOTS; j++)
548 drm_sched_fini(&js->queue[j].sched);
552 int panfrost_job_open(struct panfrost_file_priv *panfrost_priv)
554 struct panfrost_device *pfdev = panfrost_priv->pfdev;
555 struct panfrost_job_slot *js = pfdev->js;
556 struct drm_gpu_scheduler *sched;
559 for (i = 0; i < NUM_JOB_SLOTS; i++) {
560 sched = &js->queue[i].sched;
561 ret = drm_sched_entity_init(&panfrost_priv->sched_entity[i],
562 DRM_SCHED_PRIORITY_NORMAL, &sched,
570 void panfrost_job_close(struct panfrost_file_priv *panfrost_priv)
574 for (i = 0; i < NUM_JOB_SLOTS; i++)
575 drm_sched_entity_destroy(&panfrost_priv->sched_entity[i]);
578 int panfrost_job_is_idle(struct panfrost_device *pfdev)
580 struct panfrost_job_slot *js = pfdev->js;
583 /* Check whether the hardware is idle */
584 if (atomic_read(&pfdev->devfreq.busy_count))
587 for (i = 0; i < NUM_JOB_SLOTS; i++) {
588 /* If there are any jobs in the HW queue, we're not idle */
589 if (atomic_read(&js->queue[i].sched.hw_rq_count))