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
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>
39 #include "amdgpu_trace.h"
43 * Fences mark an event in the GPUs pipeline and are used
44 * for GPU/CPU synchronization. When the fence is written,
45 * it is expected that all buffers associated with that fence
46 * are no longer in use by the associated ring on the GPU and
47 * that the the relevant GPU caches have been flushed.
51 struct dma_fence base;
54 struct amdgpu_ring *ring;
57 static struct kmem_cache *amdgpu_fence_slab;
59 int amdgpu_fence_slab_init(void)
61 amdgpu_fence_slab = kmem_cache_create(
62 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
63 SLAB_HWCACHE_ALIGN, NULL);
64 if (!amdgpu_fence_slab)
69 void amdgpu_fence_slab_fini(void)
72 kmem_cache_destroy(amdgpu_fence_slab);
77 static const struct dma_fence_ops amdgpu_fence_ops;
78 static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
80 struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
82 if (__f->base.ops == &amdgpu_fence_ops)
89 * amdgpu_fence_write - write a fence value
91 * @ring: ring the fence is associated with
92 * @seq: sequence number to write
94 * Writes a fence value to memory (all asics).
96 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
98 struct amdgpu_fence_driver *drv = &ring->fence_drv;
101 *drv->cpu_addr = cpu_to_le32(seq);
105 * amdgpu_fence_read - read a fence value
107 * @ring: ring the fence is associated with
109 * Reads a fence value from memory (all asics).
110 * Returns the value of the fence read from memory.
112 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
114 struct amdgpu_fence_driver *drv = &ring->fence_drv;
118 seq = le32_to_cpu(*drv->cpu_addr);
120 seq = atomic_read(&drv->last_seq);
126 * amdgpu_fence_emit - emit a fence on the requested ring
128 * @ring: ring the fence is associated with
129 * @f: resulting fence object
131 * Emits a fence command on the requested ring (all asics).
132 * Returns 0 on success, -ENOMEM on failure.
134 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
137 struct amdgpu_device *adev = ring->adev;
138 struct amdgpu_fence *fence;
139 struct dma_fence __rcu **ptr;
143 fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
147 seq = ++ring->fence_drv.sync_seq;
149 dma_fence_init(&fence->base, &amdgpu_fence_ops,
150 &ring->fence_drv.lock,
151 adev->fence_context + ring->idx,
153 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
154 seq, flags | AMDGPU_FENCE_FLAG_INT);
156 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
157 if (unlikely(rcu_dereference_protected(*ptr, 1))) {
158 struct dma_fence *old;
161 old = dma_fence_get_rcu_safe(ptr);
165 r = dma_fence_wait(old, false);
172 /* This function can't be called concurrently anyway, otherwise
173 * emitting the fence would mess up the hardware ring buffer.
175 rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
183 * amdgpu_fence_emit_polling - emit a fence on the requeste ring
185 * @ring: ring the fence is associated with
186 * @s: resulting sequence number
188 * Emits a fence command on the requested ring (all asics).
189 * Used For polling fence.
190 * Returns 0 on success, -ENOMEM on failure.
192 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s)
199 seq = ++ring->fence_drv.sync_seq;
200 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
209 * amdgpu_fence_schedule_fallback - schedule fallback check
211 * @ring: pointer to struct amdgpu_ring
213 * Start a timer as fallback to our interrupts.
215 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
217 mod_timer(&ring->fence_drv.fallback_timer,
218 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
222 * amdgpu_fence_process - check for fence activity
224 * @ring: pointer to struct amdgpu_ring
226 * Checks the current fence value and calculates the last
227 * signalled fence value. Wakes the fence queue if the
228 * sequence number has increased.
230 * Returns true if fence was processed
232 bool amdgpu_fence_process(struct amdgpu_ring *ring)
234 struct amdgpu_fence_driver *drv = &ring->fence_drv;
235 uint32_t seq, last_seq;
239 last_seq = atomic_read(&ring->fence_drv.last_seq);
240 seq = amdgpu_fence_read(ring);
242 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
244 if (del_timer(&ring->fence_drv.fallback_timer) &&
245 seq != ring->fence_drv.sync_seq)
246 amdgpu_fence_schedule_fallback(ring);
248 if (unlikely(seq == last_seq))
251 last_seq &= drv->num_fences_mask;
252 seq &= drv->num_fences_mask;
255 struct dma_fence *fence, **ptr;
258 last_seq &= drv->num_fences_mask;
259 ptr = &drv->fences[last_seq];
261 /* There is always exactly one thread signaling this fence slot */
262 fence = rcu_dereference_protected(*ptr, 1);
263 RCU_INIT_POINTER(*ptr, NULL);
268 r = dma_fence_signal(fence);
270 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
274 dma_fence_put(fence);
275 } while (last_seq != seq);
281 * amdgpu_fence_fallback - fallback for hardware interrupts
283 * @work: delayed work item
285 * Checks for fence activity.
287 static void amdgpu_fence_fallback(struct timer_list *t)
289 struct amdgpu_ring *ring = from_timer(ring, t,
290 fence_drv.fallback_timer);
292 if (amdgpu_fence_process(ring))
293 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
297 * amdgpu_fence_wait_empty - wait for all fences to signal
299 * @adev: amdgpu device pointer
300 * @ring: ring index the fence is associated with
302 * Wait for all fences on the requested ring to signal (all asics).
303 * Returns 0 if the fences have passed, error for all other cases.
305 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
307 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
308 struct dma_fence *fence, **ptr;
314 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
316 fence = rcu_dereference(*ptr);
317 if (!fence || !dma_fence_get_rcu(fence)) {
323 r = dma_fence_wait(fence, false);
324 dma_fence_put(fence);
329 * amdgpu_fence_wait_polling - busy wait for givn sequence number
331 * @ring: ring index the fence is associated with
332 * @wait_seq: sequence number to wait
333 * @timeout: the timeout for waiting in usecs
335 * Wait for all fences on the requested ring to signal (all asics).
336 * Returns left time if no timeout, 0 or minus if timeout.
338 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
345 seq = amdgpu_fence_read(ring);
348 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
350 return timeout > 0 ? timeout : 0;
353 * amdgpu_fence_count_emitted - get the count of emitted fences
355 * @ring: ring the fence is associated with
357 * Get the number of fences emitted on the requested ring (all asics).
358 * Returns the number of emitted fences on the ring. Used by the
359 * dynpm code to ring track activity.
361 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
365 /* We are not protected by ring lock when reading the last sequence
366 * but it's ok to report slightly wrong fence count here.
368 amdgpu_fence_process(ring);
369 emitted = 0x100000000ull;
370 emitted -= atomic_read(&ring->fence_drv.last_seq);
371 emitted += READ_ONCE(ring->fence_drv.sync_seq);
372 return lower_32_bits(emitted);
376 * amdgpu_fence_driver_start_ring - make the fence driver
377 * ready for use on the requested ring.
379 * @ring: ring to start the fence driver on
380 * @irq_src: interrupt source to use for this ring
381 * @irq_type: interrupt type to use for this ring
383 * Make the fence driver ready for processing (all asics).
384 * Not all asics have all rings, so each asic will only
385 * start the fence driver on the rings it has.
386 * Returns 0 for success, errors for failure.
388 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
389 struct amdgpu_irq_src *irq_src,
392 struct amdgpu_device *adev = ring->adev;
395 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
396 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
397 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
399 /* put fence directly behind firmware */
400 index = ALIGN(adev->uvd.fw->size, 8);
401 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
402 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
404 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
405 amdgpu_irq_get(adev, irq_src, irq_type);
407 ring->fence_drv.irq_src = irq_src;
408 ring->fence_drv.irq_type = irq_type;
409 ring->fence_drv.initialized = true;
411 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr "
412 "0x%016llx, cpu addr 0x%p\n", ring->name,
413 ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
418 * amdgpu_fence_driver_init_ring - init the fence driver
419 * for the requested ring.
421 * @ring: ring to init the fence driver on
422 * @num_hw_submission: number of entries on the hardware queue
424 * Init the fence driver for the requested ring (all asics).
425 * Helper function for amdgpu_fence_driver_init().
427 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
428 unsigned num_hw_submission)
433 /* Check that num_hw_submission is a power of two */
434 if ((num_hw_submission & (num_hw_submission - 1)) != 0)
437 ring->fence_drv.cpu_addr = NULL;
438 ring->fence_drv.gpu_addr = 0;
439 ring->fence_drv.sync_seq = 0;
440 atomic_set(&ring->fence_drv.last_seq, 0);
441 ring->fence_drv.initialized = false;
443 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
445 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
446 spin_lock_init(&ring->fence_drv.lock);
447 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
449 if (!ring->fence_drv.fences)
452 /* No need to setup the GPU scheduler for KIQ ring */
453 if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) {
454 /* for non-sriov case, no timeout enforce on compute ring */
455 if ((ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
456 && !amdgpu_sriov_vf(ring->adev))
457 timeout = MAX_SCHEDULE_TIMEOUT;
459 timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
461 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
462 num_hw_submission, amdgpu_job_hang_limit,
463 timeout, ring->name);
465 DRM_ERROR("Failed to create scheduler on ring %s.\n",
475 * amdgpu_fence_driver_init - init the fence driver
476 * for all possible rings.
478 * @adev: amdgpu device pointer
480 * Init the fence driver for all possible rings (all asics).
481 * Not all asics have all rings, so each asic will only
482 * start the fence driver on the rings it has using
483 * amdgpu_fence_driver_start_ring().
484 * Returns 0 for success.
486 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
488 if (amdgpu_debugfs_fence_init(adev))
489 dev_err(adev->dev, "fence debugfs file creation failed\n");
495 * amdgpu_fence_driver_fini - tear down the fence driver
496 * for all possible rings.
498 * @adev: amdgpu device pointer
500 * Tear down the fence driver for all possible rings (all asics).
502 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
507 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
508 struct amdgpu_ring *ring = adev->rings[i];
510 if (!ring || !ring->fence_drv.initialized)
512 r = amdgpu_fence_wait_empty(ring);
514 /* no need to trigger GPU reset as we are unloading */
515 amdgpu_fence_driver_force_completion(ring);
517 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
518 ring->fence_drv.irq_type);
519 drm_sched_fini(&ring->sched);
520 del_timer_sync(&ring->fence_drv.fallback_timer);
521 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
522 dma_fence_put(ring->fence_drv.fences[j]);
523 kfree(ring->fence_drv.fences);
524 ring->fence_drv.fences = NULL;
525 ring->fence_drv.initialized = false;
530 * amdgpu_fence_driver_suspend - suspend the fence driver
531 * for all possible rings.
533 * @adev: amdgpu device pointer
535 * Suspend the fence driver for all possible rings (all asics).
537 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
541 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
542 struct amdgpu_ring *ring = adev->rings[i];
543 if (!ring || !ring->fence_drv.initialized)
546 /* wait for gpu to finish processing current batch */
547 r = amdgpu_fence_wait_empty(ring);
549 /* delay GPU reset to resume */
550 amdgpu_fence_driver_force_completion(ring);
553 /* disable the interrupt */
554 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
555 ring->fence_drv.irq_type);
560 * amdgpu_fence_driver_resume - resume the fence driver
561 * for all possible rings.
563 * @adev: amdgpu device pointer
565 * Resume the fence driver for all possible rings (all asics).
566 * Not all asics have all rings, so each asic will only
567 * start the fence driver on the rings it has using
568 * amdgpu_fence_driver_start_ring().
569 * Returns 0 for success.
571 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
575 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
576 struct amdgpu_ring *ring = adev->rings[i];
577 if (!ring || !ring->fence_drv.initialized)
580 /* enable the interrupt */
581 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
582 ring->fence_drv.irq_type);
587 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
589 * @ring: fence of the ring to signal
592 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
594 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
595 amdgpu_fence_process(ring);
599 * Common fence implementation
602 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
607 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
609 struct amdgpu_fence *fence = to_amdgpu_fence(f);
610 return (const char *)fence->ring->name;
614 * amdgpu_fence_enable_signaling - enable signalling on fence
617 * This function is called with fence_queue lock held, and adds a callback
618 * to fence_queue that checks if this fence is signaled, and if so it
619 * signals the fence and removes itself.
621 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
623 struct amdgpu_fence *fence = to_amdgpu_fence(f);
624 struct amdgpu_ring *ring = fence->ring;
626 if (!timer_pending(&ring->fence_drv.fallback_timer))
627 amdgpu_fence_schedule_fallback(ring);
629 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
635 * amdgpu_fence_free - free up the fence memory
637 * @rcu: RCU callback head
639 * Free up the fence memory after the RCU grace period.
641 static void amdgpu_fence_free(struct rcu_head *rcu)
643 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
644 struct amdgpu_fence *fence = to_amdgpu_fence(f);
645 kmem_cache_free(amdgpu_fence_slab, fence);
649 * amdgpu_fence_release - callback that fence can be freed
653 * This function is called when the reference count becomes zero.
654 * It just RCU schedules freeing up the fence.
656 static void amdgpu_fence_release(struct dma_fence *f)
658 call_rcu(&f->rcu, amdgpu_fence_free);
661 static const struct dma_fence_ops amdgpu_fence_ops = {
662 .get_driver_name = amdgpu_fence_get_driver_name,
663 .get_timeline_name = amdgpu_fence_get_timeline_name,
664 .enable_signaling = amdgpu_fence_enable_signaling,
665 .release = amdgpu_fence_release,
671 #if defined(CONFIG_DEBUG_FS)
672 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
674 struct drm_info_node *node = (struct drm_info_node *)m->private;
675 struct drm_device *dev = node->minor->dev;
676 struct amdgpu_device *adev = dev->dev_private;
679 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
680 struct amdgpu_ring *ring = adev->rings[i];
681 if (!ring || !ring->fence_drv.initialized)
684 amdgpu_fence_process(ring);
686 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
687 seq_printf(m, "Last signaled fence 0x%08x\n",
688 atomic_read(&ring->fence_drv.last_seq));
689 seq_printf(m, "Last emitted 0x%08x\n",
690 ring->fence_drv.sync_seq);
692 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
695 /* set in CP_VMID_PREEMPT and preemption occurred */
696 seq_printf(m, "Last preempted 0x%08x\n",
697 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
698 /* set in CP_VMID_RESET and reset occurred */
699 seq_printf(m, "Last reset 0x%08x\n",
700 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
701 /* Both preemption and reset occurred */
702 seq_printf(m, "Last both 0x%08x\n",
703 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
709 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
711 * Manually trigger a gpu reset at the next fence wait.
713 static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
715 struct drm_info_node *node = (struct drm_info_node *) m->private;
716 struct drm_device *dev = node->minor->dev;
717 struct amdgpu_device *adev = dev->dev_private;
719 seq_printf(m, "gpu recover\n");
720 amdgpu_device_gpu_recover(adev, NULL);
725 static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
726 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
727 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
730 static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
731 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
735 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
737 #if defined(CONFIG_DEBUG_FS)
738 if (amdgpu_sriov_vf(adev))
739 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1);
740 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);