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);
159 panfrost_devfreq_record_transition(pfdev, js);
161 job_write(pfdev, JS_HEAD_NEXT_LO(js), jc_head & 0xFFFFFFFF);
162 job_write(pfdev, JS_HEAD_NEXT_HI(js), jc_head >> 32);
164 panfrost_job_write_affinity(pfdev, job->requirements, js);
166 /* start MMU, medium priority, cache clean/flush on end, clean/flush on
168 cfg |= JS_CONFIG_THREAD_PRI(8) |
169 JS_CONFIG_START_FLUSH_CLEAN_INVALIDATE |
170 JS_CONFIG_END_FLUSH_CLEAN_INVALIDATE;
172 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION))
173 cfg |= JS_CONFIG_ENABLE_FLUSH_REDUCTION;
175 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10649))
176 cfg |= JS_CONFIG_START_MMU;
178 job_write(pfdev, JS_CONFIG_NEXT(js), cfg);
180 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION))
181 job_write(pfdev, JS_FLUSH_ID_NEXT(js), job->flush_id);
184 dev_dbg(pfdev->dev, "JS: Submitting atom %p to js[%d] with head=0x%llx",
187 job_write(pfdev, JS_COMMAND_NEXT(js), JS_COMMAND_START);
190 static void panfrost_acquire_object_fences(struct drm_gem_object **bos,
192 struct dma_fence **implicit_fences)
196 for (i = 0; i < bo_count; i++)
197 implicit_fences[i] = dma_resv_get_excl_rcu(bos[i]->resv);
200 static void panfrost_attach_object_fences(struct drm_gem_object **bos,
202 struct dma_fence *fence)
206 for (i = 0; i < bo_count; i++)
207 dma_resv_add_excl_fence(bos[i]->resv, fence);
210 int panfrost_job_push(struct panfrost_job *job)
212 struct panfrost_device *pfdev = job->pfdev;
213 int slot = panfrost_job_get_slot(job);
214 struct drm_sched_entity *entity = &job->file_priv->sched_entity[slot];
215 struct ww_acquire_ctx acquire_ctx;
218 mutex_lock(&pfdev->sched_lock);
220 ret = drm_gem_lock_reservations(job->bos, job->bo_count,
223 mutex_unlock(&pfdev->sched_lock);
227 ret = drm_sched_job_init(&job->base, entity, NULL);
229 mutex_unlock(&pfdev->sched_lock);
233 job->render_done_fence = dma_fence_get(&job->base.s_fence->finished);
235 kref_get(&job->refcount); /* put by scheduler job completion */
237 panfrost_acquire_object_fences(job->bos, job->bo_count,
238 job->implicit_fences);
240 drm_sched_entity_push_job(&job->base, entity);
242 mutex_unlock(&pfdev->sched_lock);
244 panfrost_attach_object_fences(job->bos, job->bo_count,
245 job->render_done_fence);
248 drm_gem_unlock_reservations(job->bos, job->bo_count, &acquire_ctx);
253 static void panfrost_job_cleanup(struct kref *ref)
255 struct panfrost_job *job = container_of(ref, struct panfrost_job,
259 if (job->in_fences) {
260 for (i = 0; i < job->in_fence_count; i++)
261 dma_fence_put(job->in_fences[i]);
262 kvfree(job->in_fences);
264 if (job->implicit_fences) {
265 for (i = 0; i < job->bo_count; i++)
266 dma_fence_put(job->implicit_fences[i]);
267 kvfree(job->implicit_fences);
269 dma_fence_put(job->done_fence);
270 dma_fence_put(job->render_done_fence);
273 for (i = 0; i < job->bo_count; i++)
274 drm_gem_object_put_unlocked(job->bos[i]);
281 void panfrost_job_put(struct panfrost_job *job)
283 kref_put(&job->refcount, panfrost_job_cleanup);
286 static void panfrost_job_free(struct drm_sched_job *sched_job)
288 struct panfrost_job *job = to_panfrost_job(sched_job);
290 drm_sched_job_cleanup(sched_job);
292 panfrost_job_put(job);
295 static struct dma_fence *panfrost_job_dependency(struct drm_sched_job *sched_job,
296 struct drm_sched_entity *s_entity)
298 struct panfrost_job *job = to_panfrost_job(sched_job);
299 struct dma_fence *fence;
302 /* Explicit fences */
303 for (i = 0; i < job->in_fence_count; i++) {
304 if (job->in_fences[i]) {
305 fence = job->in_fences[i];
306 job->in_fences[i] = NULL;
311 /* Implicit fences, max. one per BO */
312 for (i = 0; i < job->bo_count; i++) {
313 if (job->implicit_fences[i]) {
314 fence = job->implicit_fences[i];
315 job->implicit_fences[i] = NULL;
323 static struct dma_fence *panfrost_job_run(struct drm_sched_job *sched_job)
325 struct panfrost_job *job = to_panfrost_job(sched_job);
326 struct panfrost_device *pfdev = job->pfdev;
327 int slot = panfrost_job_get_slot(job);
328 struct dma_fence *fence = NULL;
330 if (unlikely(job->base.s_fence->finished.error))
333 pfdev->jobs[slot] = job;
335 fence = panfrost_fence_create(pfdev, slot);
340 dma_fence_put(job->done_fence);
341 job->done_fence = dma_fence_get(fence);
343 panfrost_job_hw_submit(job, slot);
348 void panfrost_job_enable_interrupts(struct panfrost_device *pfdev)
353 for (j = 0; j < NUM_JOB_SLOTS; j++) {
354 irq_mask |= MK_JS_MASK(j);
357 job_write(pfdev, JOB_INT_CLEAR, irq_mask);
358 job_write(pfdev, JOB_INT_MASK, irq_mask);
361 static void panfrost_job_timedout(struct drm_sched_job *sched_job)
363 struct panfrost_job *job = to_panfrost_job(sched_job);
364 struct panfrost_device *pfdev = job->pfdev;
365 int js = panfrost_job_get_slot(job);
370 * If the GPU managed to complete this jobs fence, the timeout is
371 * spurious. Bail out.
373 if (dma_fence_is_signaled(job->done_fence))
376 dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",
378 job_read(pfdev, JS_CONFIG(js)),
379 job_read(pfdev, JS_STATUS(js)),
380 job_read(pfdev, JS_HEAD_LO(js)),
381 job_read(pfdev, JS_TAIL_LO(js)),
384 mutex_lock(&pfdev->reset_lock);
386 for (i = 0; i < NUM_JOB_SLOTS; i++)
387 drm_sched_stop(&pfdev->js->queue[i].sched, sched_job);
390 drm_sched_increase_karma(sched_job);
392 spin_lock_irqsave(&pfdev->js->job_lock, flags);
393 for (i = 0; i < NUM_JOB_SLOTS; i++) {
394 if (pfdev->jobs[i]) {
395 pm_runtime_put_noidle(pfdev->dev);
396 pfdev->jobs[i] = NULL;
399 spin_unlock_irqrestore(&pfdev->js->job_lock, flags);
401 /* panfrost_core_dump(pfdev); */
403 panfrost_devfreq_record_transition(pfdev, js);
404 panfrost_device_reset(pfdev);
406 for (i = 0; i < NUM_JOB_SLOTS; i++)
407 drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched);
409 /* restart scheduler after GPU is usable again */
410 for (i = 0; i < NUM_JOB_SLOTS; i++)
411 drm_sched_start(&pfdev->js->queue[i].sched, true);
413 mutex_unlock(&pfdev->reset_lock);
416 static const struct drm_sched_backend_ops panfrost_sched_ops = {
417 .dependency = panfrost_job_dependency,
418 .run_job = panfrost_job_run,
419 .timedout_job = panfrost_job_timedout,
420 .free_job = panfrost_job_free
423 static irqreturn_t panfrost_job_irq_handler(int irq, void *data)
425 struct panfrost_device *pfdev = data;
426 u32 status = job_read(pfdev, JOB_INT_STAT);
429 dev_dbg(pfdev->dev, "jobslot irq status=%x\n", status);
434 pm_runtime_mark_last_busy(pfdev->dev);
436 for (j = 0; status; j++) {
437 u32 mask = MK_JS_MASK(j);
439 if (!(status & mask))
442 job_write(pfdev, JOB_INT_CLEAR, mask);
444 if (status & JOB_INT_MASK_ERR(j)) {
445 job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_NOP);
447 dev_err(pfdev->dev, "js fault, js=%d, status=%s, head=0x%x, tail=0x%x",
449 panfrost_exception_name(pfdev, job_read(pfdev, JS_STATUS(j))),
450 job_read(pfdev, JS_HEAD_LO(j)),
451 job_read(pfdev, JS_TAIL_LO(j)));
453 drm_sched_fault(&pfdev->js->queue[j].sched);
456 if (status & JOB_INT_MASK_DONE(j)) {
457 struct panfrost_job *job;
459 spin_lock(&pfdev->js->job_lock);
460 job = pfdev->jobs[j];
461 /* Only NULL if job timeout occurred */
463 pfdev->jobs[j] = NULL;
465 panfrost_mmu_as_put(pfdev, &job->file_priv->mmu);
466 panfrost_devfreq_record_transition(pfdev, j);
468 dma_fence_signal_locked(job->done_fence);
469 pm_runtime_put_autosuspend(pfdev->dev);
471 spin_unlock(&pfdev->js->job_lock);
480 int panfrost_job_init(struct panfrost_device *pfdev)
482 struct panfrost_job_slot *js;
485 pfdev->js = js = devm_kzalloc(pfdev->dev, sizeof(*js), GFP_KERNEL);
489 spin_lock_init(&js->job_lock);
491 irq = platform_get_irq_byname(to_platform_device(pfdev->dev), "job");
495 ret = devm_request_irq(pfdev->dev, irq, panfrost_job_irq_handler,
496 IRQF_SHARED, "job", pfdev);
498 dev_err(pfdev->dev, "failed to request job irq");
502 for (j = 0; j < NUM_JOB_SLOTS; j++) {
503 js->queue[j].fence_context = dma_fence_context_alloc(1);
505 ret = drm_sched_init(&js->queue[j].sched,
507 1, 0, msecs_to_jiffies(500),
510 dev_err(pfdev->dev, "Failed to create scheduler: %d.", ret);
515 panfrost_job_enable_interrupts(pfdev);
520 for (j--; j >= 0; j--)
521 drm_sched_fini(&js->queue[j].sched);
526 void panfrost_job_fini(struct panfrost_device *pfdev)
528 struct panfrost_job_slot *js = pfdev->js;
531 job_write(pfdev, JOB_INT_MASK, 0);
533 for (j = 0; j < NUM_JOB_SLOTS; j++)
534 drm_sched_fini(&js->queue[j].sched);
538 int panfrost_job_open(struct panfrost_file_priv *panfrost_priv)
540 struct panfrost_device *pfdev = panfrost_priv->pfdev;
541 struct panfrost_job_slot *js = pfdev->js;
542 struct drm_sched_rq *rq;
545 for (i = 0; i < NUM_JOB_SLOTS; i++) {
546 rq = &js->queue[i].sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL];
547 ret = drm_sched_entity_init(&panfrost_priv->sched_entity[i], &rq, 1, NULL);
554 void panfrost_job_close(struct panfrost_file_priv *panfrost_priv)
558 for (i = 0; i < NUM_JOB_SLOTS; i++)
559 drm_sched_entity_destroy(&panfrost_priv->sched_entity[i]);
562 int panfrost_job_is_idle(struct panfrost_device *pfdev)
564 struct panfrost_job_slot *js = pfdev->js;
567 for (i = 0; i < NUM_JOB_SLOTS; i++) {
568 /* If there are any jobs in the HW queue, we're not idle */
569 if (atomic_read(&js->queue[i].sched.hw_rq_count))
572 /* Check whether the hardware is idle */
573 if (pfdev->devfreq.slot[i].busy)