2 * Copyright 2009 Jerome Glisse.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
28 * Jerome Glisse <glisse@freedesktop.org>
31 #include <linux/seq_file.h>
32 #include <linux/atomic.h>
33 #include <linux/wait.h>
34 #include <linux/kref.h>
35 #include <linux/slab.h>
36 #include <linux/firmware.h>
37 #include <linux/pm_runtime.h>
39 #include <drm/drm_drv.h>
41 #include "amdgpu_trace.h"
45 * Fences mark an event in the GPUs pipeline and are used
46 * for GPU/CPU synchronization. When the fence is written,
47 * it is expected that all buffers associated with that fence
48 * are no longer in use by the associated ring on the GPU and
49 * that the the relevant GPU caches have been flushed.
53 struct dma_fence base;
56 struct amdgpu_ring *ring;
59 static struct kmem_cache *amdgpu_fence_slab;
61 int amdgpu_fence_slab_init(void)
63 amdgpu_fence_slab = kmem_cache_create(
64 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
65 SLAB_HWCACHE_ALIGN, NULL);
66 if (!amdgpu_fence_slab)
71 void amdgpu_fence_slab_fini(void)
74 kmem_cache_destroy(amdgpu_fence_slab);
79 static const struct dma_fence_ops amdgpu_fence_ops;
80 static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
82 struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
84 if (__f->base.ops == &amdgpu_fence_ops)
91 * amdgpu_fence_write - write a fence value
93 * @ring: ring the fence is associated with
94 * @seq: sequence number to write
96 * Writes a fence value to memory (all asics).
98 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
100 struct amdgpu_fence_driver *drv = &ring->fence_drv;
103 *drv->cpu_addr = cpu_to_le32(seq);
107 * amdgpu_fence_read - read a fence value
109 * @ring: ring the fence is associated with
111 * Reads a fence value from memory (all asics).
112 * Returns the value of the fence read from memory.
114 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
116 struct amdgpu_fence_driver *drv = &ring->fence_drv;
120 seq = le32_to_cpu(*drv->cpu_addr);
122 seq = atomic_read(&drv->last_seq);
128 * amdgpu_fence_emit - emit a fence on the requested ring
130 * @ring: ring the fence is associated with
131 * @f: resulting fence object
132 * @flags: flags to pass into the subordinate .emit_fence() call
134 * Emits a fence command on the requested ring (all asics).
135 * Returns 0 on success, -ENOMEM on failure.
137 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
140 struct amdgpu_device *adev = ring->adev;
141 struct amdgpu_fence *fence;
142 struct dma_fence __rcu **ptr;
146 fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
150 seq = ++ring->fence_drv.sync_seq;
152 dma_fence_init(&fence->base, &amdgpu_fence_ops,
153 &ring->fence_drv.lock,
154 adev->fence_context + ring->idx,
156 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
157 seq, flags | AMDGPU_FENCE_FLAG_INT);
158 pm_runtime_get_noresume(adev_to_drm(adev)->dev);
159 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
160 if (unlikely(rcu_dereference_protected(*ptr, 1))) {
161 struct dma_fence *old;
164 old = dma_fence_get_rcu_safe(ptr);
168 r = dma_fence_wait(old, false);
175 /* This function can't be called concurrently anyway, otherwise
176 * emitting the fence would mess up the hardware ring buffer.
178 rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
186 * amdgpu_fence_emit_polling - emit a fence on the requeste ring
188 * @ring: ring the fence is associated with
189 * @s: resulting sequence number
190 * @timeout: the timeout for waiting in usecs
192 * Emits a fence command on the requested ring (all asics).
193 * Used For polling fence.
194 * Returns 0 on success, -ENOMEM on failure.
196 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
205 seq = ++ring->fence_drv.sync_seq;
206 r = amdgpu_fence_wait_polling(ring,
207 seq - ring->fence_drv.num_fences_mask,
212 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
221 * amdgpu_fence_schedule_fallback - schedule fallback check
223 * @ring: pointer to struct amdgpu_ring
225 * Start a timer as fallback to our interrupts.
227 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
229 mod_timer(&ring->fence_drv.fallback_timer,
230 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
234 * amdgpu_fence_process - check for fence activity
236 * @ring: pointer to struct amdgpu_ring
238 * Checks the current fence value and calculates the last
239 * signalled fence value. Wakes the fence queue if the
240 * sequence number has increased.
242 * Returns true if fence was processed
244 bool amdgpu_fence_process(struct amdgpu_ring *ring)
246 struct amdgpu_fence_driver *drv = &ring->fence_drv;
247 struct amdgpu_device *adev = ring->adev;
248 uint32_t seq, last_seq;
252 last_seq = atomic_read(&ring->fence_drv.last_seq);
253 seq = amdgpu_fence_read(ring);
255 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
257 if (del_timer(&ring->fence_drv.fallback_timer) &&
258 seq != ring->fence_drv.sync_seq)
259 amdgpu_fence_schedule_fallback(ring);
261 if (unlikely(seq == last_seq))
264 last_seq &= drv->num_fences_mask;
265 seq &= drv->num_fences_mask;
268 struct dma_fence *fence, **ptr;
271 last_seq &= drv->num_fences_mask;
272 ptr = &drv->fences[last_seq];
274 /* There is always exactly one thread signaling this fence slot */
275 fence = rcu_dereference_protected(*ptr, 1);
276 RCU_INIT_POINTER(*ptr, NULL);
281 r = dma_fence_signal(fence);
283 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
287 dma_fence_put(fence);
288 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
289 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
290 } while (last_seq != seq);
296 * amdgpu_fence_fallback - fallback for hardware interrupts
298 * @t: timer context used to obtain the pointer to ring structure
300 * Checks for fence activity.
302 static void amdgpu_fence_fallback(struct timer_list *t)
304 struct amdgpu_ring *ring = from_timer(ring, t,
305 fence_drv.fallback_timer);
307 if (amdgpu_fence_process(ring))
308 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
312 * amdgpu_fence_wait_empty - wait for all fences to signal
314 * @ring: ring index the fence is associated with
316 * Wait for all fences on the requested ring to signal (all asics).
317 * Returns 0 if the fences have passed, error for all other cases.
319 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
321 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
322 struct dma_fence *fence, **ptr;
328 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
330 fence = rcu_dereference(*ptr);
331 if (!fence || !dma_fence_get_rcu(fence)) {
337 r = dma_fence_wait(fence, false);
338 dma_fence_put(fence);
343 * amdgpu_fence_wait_polling - busy wait for givn sequence number
345 * @ring: ring index the fence is associated with
346 * @wait_seq: sequence number to wait
347 * @timeout: the timeout for waiting in usecs
349 * Wait for all fences on the requested ring to signal (all asics).
350 * Returns left time if no timeout, 0 or minus if timeout.
352 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
359 seq = amdgpu_fence_read(ring);
362 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
364 return timeout > 0 ? timeout : 0;
367 * amdgpu_fence_count_emitted - get the count of emitted fences
369 * @ring: ring the fence is associated with
371 * Get the number of fences emitted on the requested ring (all asics).
372 * Returns the number of emitted fences on the ring. Used by the
373 * dynpm code to ring track activity.
375 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
379 /* We are not protected by ring lock when reading the last sequence
380 * but it's ok to report slightly wrong fence count here.
382 amdgpu_fence_process(ring);
383 emitted = 0x100000000ull;
384 emitted -= atomic_read(&ring->fence_drv.last_seq);
385 emitted += READ_ONCE(ring->fence_drv.sync_seq);
386 return lower_32_bits(emitted);
390 * amdgpu_fence_driver_start_ring - make the fence driver
391 * ready for use on the requested ring.
393 * @ring: ring to start the fence driver on
394 * @irq_src: interrupt source to use for this ring
395 * @irq_type: interrupt type to use for this ring
397 * Make the fence driver ready for processing (all asics).
398 * Not all asics have all rings, so each asic will only
399 * start the fence driver on the rings it has.
400 * Returns 0 for success, errors for failure.
402 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
403 struct amdgpu_irq_src *irq_src,
406 struct amdgpu_device *adev = ring->adev;
409 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
410 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
411 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
413 /* put fence directly behind firmware */
414 index = ALIGN(adev->uvd.fw->size, 8);
415 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
416 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
418 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
420 ring->fence_drv.irq_src = irq_src;
421 ring->fence_drv.irq_type = irq_type;
422 ring->fence_drv.initialized = true;
424 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
425 ring->name, ring->fence_drv.gpu_addr);
430 * amdgpu_fence_driver_init_ring - init the fence driver
431 * for the requested ring.
433 * @ring: ring to init the fence driver on
434 * @num_hw_submission: number of entries on the hardware queue
435 * @sched_score: optional score atomic shared with other schedulers
437 * Init the fence driver for the requested ring (all asics).
438 * Helper function for amdgpu_fence_driver_init().
440 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
441 unsigned num_hw_submission,
442 atomic_t *sched_score)
444 struct amdgpu_device *adev = ring->adev;
451 if (!is_power_of_2(num_hw_submission))
454 ring->fence_drv.cpu_addr = NULL;
455 ring->fence_drv.gpu_addr = 0;
456 ring->fence_drv.sync_seq = 0;
457 atomic_set(&ring->fence_drv.last_seq, 0);
458 ring->fence_drv.initialized = false;
460 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
462 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
463 spin_lock_init(&ring->fence_drv.lock);
464 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
466 if (!ring->fence_drv.fences)
469 /* No need to setup the GPU scheduler for rings that don't need it */
470 if (ring->no_scheduler)
473 switch (ring->funcs->type) {
474 case AMDGPU_RING_TYPE_GFX:
475 timeout = adev->gfx_timeout;
477 case AMDGPU_RING_TYPE_COMPUTE:
478 timeout = adev->compute_timeout;
480 case AMDGPU_RING_TYPE_SDMA:
481 timeout = adev->sdma_timeout;
484 timeout = adev->video_timeout;
488 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
489 num_hw_submission, amdgpu_job_hang_limit,
490 timeout, NULL, sched_score, ring->name);
492 DRM_ERROR("Failed to create scheduler on ring %s.\n",
501 * amdgpu_fence_driver_sw_init - init the fence driver
502 * for all possible rings.
504 * @adev: amdgpu device pointer
506 * Init the fence driver for all possible rings (all asics).
507 * Not all asics have all rings, so each asic will only
508 * start the fence driver on the rings it has using
509 * amdgpu_fence_driver_start_ring().
510 * Returns 0 for success.
512 int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev)
518 * amdgpu_fence_driver_hw_fini - tear down the fence driver
519 * for all possible rings.
521 * @adev: amdgpu device pointer
523 * Tear down the fence driver for all possible rings (all asics).
525 void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev)
529 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
530 struct amdgpu_ring *ring = adev->rings[i];
532 if (!ring || !ring->fence_drv.initialized)
535 /* You can't wait for HW to signal if it's gone */
536 if (!drm_dev_is_unplugged(&adev->ddev))
537 r = amdgpu_fence_wait_empty(ring);
540 /* no need to trigger GPU reset as we are unloading */
542 amdgpu_fence_driver_force_completion(ring);
544 if (ring->fence_drv.irq_src)
545 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
546 ring->fence_drv.irq_type);
548 del_timer_sync(&ring->fence_drv.fallback_timer);
552 void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev)
556 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
557 struct amdgpu_ring *ring = adev->rings[i];
559 if (!ring || !ring->fence_drv.initialized)
562 if (!ring->no_scheduler)
563 drm_sched_fini(&ring->sched);
565 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
566 dma_fence_put(ring->fence_drv.fences[j]);
567 kfree(ring->fence_drv.fences);
568 ring->fence_drv.fences = NULL;
569 ring->fence_drv.initialized = false;
574 * amdgpu_fence_driver_hw_init - enable the fence driver
575 * for all possible rings.
577 * @adev: amdgpu device pointer
579 * Enable the fence driver for all possible rings (all asics).
580 * Not all asics have all rings, so each asic will only
581 * start the fence driver on the rings it has using
582 * amdgpu_fence_driver_start_ring().
583 * Returns 0 for success.
585 void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev)
589 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
590 struct amdgpu_ring *ring = adev->rings[i];
591 if (!ring || !ring->fence_drv.initialized)
594 /* enable the interrupt */
595 if (ring->fence_drv.irq_src)
596 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
597 ring->fence_drv.irq_type);
602 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
604 * @ring: fence of the ring to signal
607 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
609 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
610 amdgpu_fence_process(ring);
614 * Common fence implementation
617 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
622 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
624 struct amdgpu_fence *fence = to_amdgpu_fence(f);
625 return (const char *)fence->ring->name;
629 * amdgpu_fence_enable_signaling - enable signalling on fence
632 * This function is called with fence_queue lock held, and adds a callback
633 * to fence_queue that checks if this fence is signaled, and if so it
634 * signals the fence and removes itself.
636 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
638 struct amdgpu_fence *fence = to_amdgpu_fence(f);
639 struct amdgpu_ring *ring = fence->ring;
641 if (!timer_pending(&ring->fence_drv.fallback_timer))
642 amdgpu_fence_schedule_fallback(ring);
644 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
650 * amdgpu_fence_free - free up the fence memory
652 * @rcu: RCU callback head
654 * Free up the fence memory after the RCU grace period.
656 static void amdgpu_fence_free(struct rcu_head *rcu)
658 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
659 struct amdgpu_fence *fence = to_amdgpu_fence(f);
660 kmem_cache_free(amdgpu_fence_slab, fence);
664 * amdgpu_fence_release - callback that fence can be freed
668 * This function is called when the reference count becomes zero.
669 * It just RCU schedules freeing up the fence.
671 static void amdgpu_fence_release(struct dma_fence *f)
673 call_rcu(&f->rcu, amdgpu_fence_free);
676 static const struct dma_fence_ops amdgpu_fence_ops = {
677 .get_driver_name = amdgpu_fence_get_driver_name,
678 .get_timeline_name = amdgpu_fence_get_timeline_name,
679 .enable_signaling = amdgpu_fence_enable_signaling,
680 .release = amdgpu_fence_release,
686 #if defined(CONFIG_DEBUG_FS)
687 static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
689 struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
692 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
693 struct amdgpu_ring *ring = adev->rings[i];
694 if (!ring || !ring->fence_drv.initialized)
697 amdgpu_fence_process(ring);
699 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
700 seq_printf(m, "Last signaled fence 0x%08x\n",
701 atomic_read(&ring->fence_drv.last_seq));
702 seq_printf(m, "Last emitted 0x%08x\n",
703 ring->fence_drv.sync_seq);
705 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
706 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
707 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
708 le32_to_cpu(*ring->trail_fence_cpu_addr));
709 seq_printf(m, "Last emitted 0x%08x\n",
713 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
716 /* set in CP_VMID_PREEMPT and preemption occurred */
717 seq_printf(m, "Last preempted 0x%08x\n",
718 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
719 /* set in CP_VMID_RESET and reset occurred */
720 seq_printf(m, "Last reset 0x%08x\n",
721 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
722 /* Both preemption and reset occurred */
723 seq_printf(m, "Last both 0x%08x\n",
724 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
730 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
732 * Manually trigger a gpu reset at the next fence wait.
734 static int gpu_recover_get(void *data, u64 *val)
736 struct amdgpu_device *adev = (struct amdgpu_device *)data;
737 struct drm_device *dev = adev_to_drm(adev);
740 r = pm_runtime_get_sync(dev->dev);
742 pm_runtime_put_autosuspend(dev->dev);
746 *val = amdgpu_device_gpu_recover(adev, NULL);
748 pm_runtime_mark_last_busy(dev->dev);
749 pm_runtime_put_autosuspend(dev->dev);
754 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
755 DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
760 void amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
762 #if defined(CONFIG_DEBUG_FS)
763 struct drm_minor *minor = adev_to_drm(adev)->primary;
764 struct dentry *root = minor->debugfs_root;
766 debugfs_create_file("amdgpu_fence_info", 0444, root, adev,
767 &amdgpu_debugfs_fence_info_fops);
769 if (!amdgpu_sriov_vf(adev))
770 debugfs_create_file("amdgpu_gpu_recover", 0444, root, adev,
771 &amdgpu_debugfs_gpu_recover_fops);