drm/panfrost: Make sure the shrinker does not reclaim referenced BOs
[linux/fpc-iii.git] / drivers / gpu / drm / panfrost / panfrost_job.c
blob4d383831c1fc2e09ead48d16fe23dcf513a86d5e
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>
6 #include <linux/io.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;
29 u64 fence_context;
30 u64 emit_seqno;
33 struct panfrost_job_slot {
34 struct panfrost_queue_state queue[NUM_JOB_SLOTS];
35 spinlock_t job_lock;
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 */
48 u64 seqno;
49 int queue;
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)
60 return "panfrost";
63 static const char *panfrost_fence_get_timeline_name(struct dma_fence *fence)
65 struct panfrost_fence *f = to_panfrost_fence(fence);
67 switch (f->queue) {
68 case 0:
69 return "panfrost-js-0";
70 case 1:
71 return "panfrost-js-1";
72 case 2:
73 return "panfrost-js-2";
74 default:
75 return NULL;
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);
90 if (!fence)
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);
99 return &fence->base;
102 static int panfrost_job_get_slot(struct panfrost_job *job)
104 /* JS0: fragment jobs.
105 * JS1: vertex/tiler jobs
106 * JS2: compute jobs
108 if (job->requirements & PANFROST_JD_REQ_FS)
109 return 0;
111 /* Not exposed to userspace yet */
112 #if 0
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))
116 return 2;
117 if (panfrost_has_hw_issue(job->pfdev, HW_ISSUE_8987))
118 return 2;
120 #endif
121 return 1;
124 static void panfrost_job_write_affinity(struct panfrost_device *pfdev,
125 u32 requirements,
126 int js)
128 u64 affinity;
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;
144 u32 cfg;
145 u64 jc_head = job->jc;
146 int ret;
148 ret = pm_runtime_get_sync(pfdev->dev);
149 if (ret < 0)
150 return;
152 if (WARN_ON(job_read(pfdev, JS_COMMAND_NEXT(js)))) {
153 pm_runtime_put_sync_autosuspend(pfdev->dev);
154 return;
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
166 * start */
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);
182 /* GO ! */
183 dev_dbg(pfdev->dev, "JS: Submitting atom %p to js[%d] with head=0x%llx",
184 job, js, jc_head);
186 job_write(pfdev, JS_COMMAND_NEXT(js), JS_COMMAND_START);
189 static void panfrost_acquire_object_fences(struct drm_gem_object **bos,
190 int bo_count,
191 struct dma_fence **implicit_fences)
193 int i;
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,
200 int bo_count,
201 struct dma_fence *fence)
203 int i;
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;
215 int ret = 0;
217 mutex_lock(&pfdev->sched_lock);
219 ret = drm_gem_lock_reservations(job->bos, job->bo_count,
220 &acquire_ctx);
221 if (ret) {
222 mutex_unlock(&pfdev->sched_lock);
223 return ret;
226 ret = drm_sched_job_init(&job->base, entity, NULL);
227 if (ret) {
228 mutex_unlock(&pfdev->sched_lock);
229 goto unlock;
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);
246 unlock:
247 drm_gem_unlock_reservations(job->bos, job->bo_count, &acquire_ctx);
249 return ret;
252 static void panfrost_job_cleanup(struct kref *ref)
254 struct panfrost_job *job = container_of(ref, struct panfrost_job,
255 refcount);
256 unsigned int i;
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);
271 if (job->mappings) {
272 for (i = 0; i < job->bo_count; i++) {
273 if (!job->mappings[i])
274 break;
276 atomic_dec(&job->mappings[i]->obj->gpu_usecount);
277 panfrost_gem_mapping_put(job->mappings[i]);
279 kvfree(job->mappings);
282 if (job->bos) {
283 struct panfrost_gem_object *bo;
285 for (i = 0; i < job->bo_count; i++) {
286 bo = to_panfrost_bo(job->bos[i]);
287 drm_gem_object_put_unlocked(job->bos[i]);
290 kvfree(job->bos);
293 kfree(job);
296 void panfrost_job_put(struct panfrost_job *job)
298 kref_put(&job->refcount, panfrost_job_cleanup);
301 static void panfrost_job_free(struct drm_sched_job *sched_job)
303 struct panfrost_job *job = to_panfrost_job(sched_job);
305 drm_sched_job_cleanup(sched_job);
307 panfrost_job_put(job);
310 static struct dma_fence *panfrost_job_dependency(struct drm_sched_job *sched_job,
311 struct drm_sched_entity *s_entity)
313 struct panfrost_job *job = to_panfrost_job(sched_job);
314 struct dma_fence *fence;
315 unsigned int i;
317 /* Explicit fences */
318 for (i = 0; i < job->in_fence_count; i++) {
319 if (job->in_fences[i]) {
320 fence = job->in_fences[i];
321 job->in_fences[i] = NULL;
322 return fence;
326 /* Implicit fences, max. one per BO */
327 for (i = 0; i < job->bo_count; i++) {
328 if (job->implicit_fences[i]) {
329 fence = job->implicit_fences[i];
330 job->implicit_fences[i] = NULL;
331 return fence;
335 return NULL;
338 static struct dma_fence *panfrost_job_run(struct drm_sched_job *sched_job)
340 struct panfrost_job *job = to_panfrost_job(sched_job);
341 struct panfrost_device *pfdev = job->pfdev;
342 int slot = panfrost_job_get_slot(job);
343 struct dma_fence *fence = NULL;
345 if (unlikely(job->base.s_fence->finished.error))
346 return NULL;
348 pfdev->jobs[slot] = job;
350 fence = panfrost_fence_create(pfdev, slot);
351 if (IS_ERR(fence))
352 return NULL;
354 if (job->done_fence)
355 dma_fence_put(job->done_fence);
356 job->done_fence = dma_fence_get(fence);
358 panfrost_job_hw_submit(job, slot);
360 return fence;
363 void panfrost_job_enable_interrupts(struct panfrost_device *pfdev)
365 int j;
366 u32 irq_mask = 0;
368 for (j = 0; j < NUM_JOB_SLOTS; j++) {
369 irq_mask |= MK_JS_MASK(j);
372 job_write(pfdev, JOB_INT_CLEAR, irq_mask);
373 job_write(pfdev, JOB_INT_MASK, irq_mask);
376 static void panfrost_job_timedout(struct drm_sched_job *sched_job)
378 struct panfrost_job *job = to_panfrost_job(sched_job);
379 struct panfrost_device *pfdev = job->pfdev;
380 int js = panfrost_job_get_slot(job);
381 unsigned long flags;
382 int i;
385 * If the GPU managed to complete this jobs fence, the timeout is
386 * spurious. Bail out.
388 if (dma_fence_is_signaled(job->done_fence))
389 return;
391 dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",
393 job_read(pfdev, JS_CONFIG(js)),
394 job_read(pfdev, JS_STATUS(js)),
395 job_read(pfdev, JS_HEAD_LO(js)),
396 job_read(pfdev, JS_TAIL_LO(js)),
397 sched_job);
399 if (!mutex_trylock(&pfdev->reset_lock))
400 return;
402 for (i = 0; i < NUM_JOB_SLOTS; i++) {
403 struct drm_gpu_scheduler *sched = &pfdev->js->queue[i].sched;
405 drm_sched_stop(sched, sched_job);
406 if (js != i)
407 /* Ensure any timeouts on other slots have finished */
408 cancel_delayed_work_sync(&sched->work_tdr);
411 drm_sched_increase_karma(sched_job);
413 spin_lock_irqsave(&pfdev->js->job_lock, flags);
414 for (i = 0; i < NUM_JOB_SLOTS; i++) {
415 if (pfdev->jobs[i]) {
416 pm_runtime_put_noidle(pfdev->dev);
417 pfdev->jobs[i] = NULL;
420 spin_unlock_irqrestore(&pfdev->js->job_lock, flags);
422 panfrost_devfreq_record_idle(pfdev);
423 panfrost_device_reset(pfdev);
425 for (i = 0; i < NUM_JOB_SLOTS; i++)
426 drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched);
428 /* restart scheduler after GPU is usable again */
429 for (i = 0; i < NUM_JOB_SLOTS; i++)
430 drm_sched_start(&pfdev->js->queue[i].sched, true);
432 mutex_unlock(&pfdev->reset_lock);
435 static const struct drm_sched_backend_ops panfrost_sched_ops = {
436 .dependency = panfrost_job_dependency,
437 .run_job = panfrost_job_run,
438 .timedout_job = panfrost_job_timedout,
439 .free_job = panfrost_job_free
442 static irqreturn_t panfrost_job_irq_handler(int irq, void *data)
444 struct panfrost_device *pfdev = data;
445 u32 status = job_read(pfdev, JOB_INT_STAT);
446 int j;
448 dev_dbg(pfdev->dev, "jobslot irq status=%x\n", status);
450 if (!status)
451 return IRQ_NONE;
453 pm_runtime_mark_last_busy(pfdev->dev);
455 for (j = 0; status; j++) {
456 u32 mask = MK_JS_MASK(j);
458 if (!(status & mask))
459 continue;
461 job_write(pfdev, JOB_INT_CLEAR, mask);
463 if (status & JOB_INT_MASK_ERR(j)) {
464 job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_NOP);
466 dev_err(pfdev->dev, "js fault, js=%d, status=%s, head=0x%x, tail=0x%x",
468 panfrost_exception_name(pfdev, job_read(pfdev, JS_STATUS(j))),
469 job_read(pfdev, JS_HEAD_LO(j)),
470 job_read(pfdev, JS_TAIL_LO(j)));
472 drm_sched_fault(&pfdev->js->queue[j].sched);
475 if (status & JOB_INT_MASK_DONE(j)) {
476 struct panfrost_job *job;
478 spin_lock(&pfdev->js->job_lock);
479 job = pfdev->jobs[j];
480 /* Only NULL if job timeout occurred */
481 if (job) {
482 pfdev->jobs[j] = NULL;
484 panfrost_mmu_as_put(pfdev, &job->file_priv->mmu);
485 panfrost_devfreq_record_idle(pfdev);
487 dma_fence_signal_locked(job->done_fence);
488 pm_runtime_put_autosuspend(pfdev->dev);
490 spin_unlock(&pfdev->js->job_lock);
493 status &= ~mask;
496 return IRQ_HANDLED;
499 int panfrost_job_init(struct panfrost_device *pfdev)
501 struct panfrost_job_slot *js;
502 int ret, j, irq;
504 pfdev->js = js = devm_kzalloc(pfdev->dev, sizeof(*js), GFP_KERNEL);
505 if (!js)
506 return -ENOMEM;
508 spin_lock_init(&js->job_lock);
510 irq = platform_get_irq_byname(to_platform_device(pfdev->dev), "job");
511 if (irq <= 0)
512 return -ENODEV;
514 ret = devm_request_irq(pfdev->dev, irq, panfrost_job_irq_handler,
515 IRQF_SHARED, "job", pfdev);
516 if (ret) {
517 dev_err(pfdev->dev, "failed to request job irq");
518 return ret;
521 for (j = 0; j < NUM_JOB_SLOTS; j++) {
522 js->queue[j].fence_context = dma_fence_context_alloc(1);
524 ret = drm_sched_init(&js->queue[j].sched,
525 &panfrost_sched_ops,
526 1, 0, msecs_to_jiffies(500),
527 "pan_js");
528 if (ret) {
529 dev_err(pfdev->dev, "Failed to create scheduler: %d.", ret);
530 goto err_sched;
534 panfrost_job_enable_interrupts(pfdev);
536 return 0;
538 err_sched:
539 for (j--; j >= 0; j--)
540 drm_sched_fini(&js->queue[j].sched);
542 return ret;
545 void panfrost_job_fini(struct panfrost_device *pfdev)
547 struct panfrost_job_slot *js = pfdev->js;
548 int j;
550 job_write(pfdev, JOB_INT_MASK, 0);
552 for (j = 0; j < NUM_JOB_SLOTS; j++)
553 drm_sched_fini(&js->queue[j].sched);
557 int panfrost_job_open(struct panfrost_file_priv *panfrost_priv)
559 struct panfrost_device *pfdev = panfrost_priv->pfdev;
560 struct panfrost_job_slot *js = pfdev->js;
561 struct drm_sched_rq *rq;
562 int ret, i;
564 for (i = 0; i < NUM_JOB_SLOTS; i++) {
565 rq = &js->queue[i].sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL];
566 ret = drm_sched_entity_init(&panfrost_priv->sched_entity[i], &rq, 1, NULL);
567 if (WARN_ON(ret))
568 return ret;
570 return 0;
573 void panfrost_job_close(struct panfrost_file_priv *panfrost_priv)
575 int i;
577 for (i = 0; i < NUM_JOB_SLOTS; i++)
578 drm_sched_entity_destroy(&panfrost_priv->sched_entity[i]);
581 int panfrost_job_is_idle(struct panfrost_device *pfdev)
583 struct panfrost_job_slot *js = pfdev->js;
584 int i;
586 /* Check whether the hardware is idle */
587 if (atomic_read(&pfdev->devfreq.busy_count))
588 return false;
590 for (i = 0; i < NUM_JOB_SLOTS; i++) {
591 /* If there are any jobs in the HW queue, we're not idle */
592 if (atomic_read(&js->queue[i].sched.hw_rq_count))
593 return false;
596 return true;