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)
136 struct amdgpu_device *adev = ring->adev;
137 struct amdgpu_fence *fence;
138 struct dma_fence *old, **ptr;
141 fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
145 seq = ++ring->fence_drv.sync_seq;
147 dma_fence_init(&fence->base, &amdgpu_fence_ops,
148 &ring->fence_drv.lock,
149 adev->fence_context + ring->idx,
151 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
152 seq, AMDGPU_FENCE_FLAG_INT);
154 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
155 /* This function can't be called concurrently anyway, otherwise
156 * emitting the fence would mess up the hardware ring buffer.
158 old = rcu_dereference_protected(*ptr, 1);
159 if (old && !dma_fence_is_signaled(old)) {
160 DRM_INFO("rcu slot is busy\n");
161 dma_fence_wait(old, false);
164 rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
172 * amdgpu_fence_emit_polling - emit a fence on the requeste ring
174 * @ring: ring the fence is associated with
175 * @s: resulting sequence number
177 * Emits a fence command on the requested ring (all asics).
178 * Used For polling fence.
179 * Returns 0 on success, -ENOMEM on failure.
181 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s)
188 seq = ++ring->fence_drv.sync_seq;
189 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
198 * amdgpu_fence_schedule_fallback - schedule fallback check
200 * @ring: pointer to struct amdgpu_ring
202 * Start a timer as fallback to our interrupts.
204 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
206 mod_timer(&ring->fence_drv.fallback_timer,
207 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
211 * amdgpu_fence_process - check for fence activity
213 * @ring: pointer to struct amdgpu_ring
215 * Checks the current fence value and calculates the last
216 * signalled fence value. Wakes the fence queue if the
217 * sequence number has increased.
219 void amdgpu_fence_process(struct amdgpu_ring *ring)
221 struct amdgpu_fence_driver *drv = &ring->fence_drv;
222 uint32_t seq, last_seq;
226 last_seq = atomic_read(&ring->fence_drv.last_seq);
227 seq = amdgpu_fence_read(ring);
229 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
231 if (seq != ring->fence_drv.sync_seq)
232 amdgpu_fence_schedule_fallback(ring);
234 if (unlikely(seq == last_seq))
237 last_seq &= drv->num_fences_mask;
238 seq &= drv->num_fences_mask;
241 struct dma_fence *fence, **ptr;
244 last_seq &= drv->num_fences_mask;
245 ptr = &drv->fences[last_seq];
247 /* There is always exactly one thread signaling this fence slot */
248 fence = rcu_dereference_protected(*ptr, 1);
249 RCU_INIT_POINTER(*ptr, NULL);
254 r = dma_fence_signal(fence);
256 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
260 dma_fence_put(fence);
261 } while (last_seq != seq);
265 * amdgpu_fence_fallback - fallback for hardware interrupts
267 * @work: delayed work item
269 * Checks for fence activity.
271 static void amdgpu_fence_fallback(struct timer_list *t)
273 struct amdgpu_ring *ring = from_timer(ring, t,
274 fence_drv.fallback_timer);
276 amdgpu_fence_process(ring);
280 * amdgpu_fence_wait_empty - wait for all fences to signal
282 * @adev: amdgpu device pointer
283 * @ring: ring index the fence is associated with
285 * Wait for all fences on the requested ring to signal (all asics).
286 * Returns 0 if the fences have passed, error for all other cases.
288 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
290 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
291 struct dma_fence *fence, **ptr;
297 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
299 fence = rcu_dereference(*ptr);
300 if (!fence || !dma_fence_get_rcu(fence)) {
306 r = dma_fence_wait(fence, false);
307 dma_fence_put(fence);
312 * amdgpu_fence_wait_polling - busy wait for givn sequence number
314 * @ring: ring index the fence is associated with
315 * @wait_seq: sequence number to wait
316 * @timeout: the timeout for waiting in usecs
318 * Wait for all fences on the requested ring to signal (all asics).
319 * Returns left time if no timeout, 0 or minus if timeout.
321 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
328 seq = amdgpu_fence_read(ring);
331 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
333 return timeout > 0 ? timeout : 0;
336 * amdgpu_fence_count_emitted - get the count of emitted fences
338 * @ring: ring the fence is associated with
340 * Get the number of fences emitted on the requested ring (all asics).
341 * Returns the number of emitted fences on the ring. Used by the
342 * dynpm code to ring track activity.
344 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
348 /* We are not protected by ring lock when reading the last sequence
349 * but it's ok to report slightly wrong fence count here.
351 amdgpu_fence_process(ring);
352 emitted = 0x100000000ull;
353 emitted -= atomic_read(&ring->fence_drv.last_seq);
354 emitted += READ_ONCE(ring->fence_drv.sync_seq);
355 return lower_32_bits(emitted);
359 * amdgpu_fence_driver_start_ring - make the fence driver
360 * ready for use on the requested ring.
362 * @ring: ring to start the fence driver on
363 * @irq_src: interrupt source to use for this ring
364 * @irq_type: interrupt type to use for this ring
366 * Make the fence driver ready for processing (all asics).
367 * Not all asics have all rings, so each asic will only
368 * start the fence driver on the rings it has.
369 * Returns 0 for success, errors for failure.
371 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
372 struct amdgpu_irq_src *irq_src,
375 struct amdgpu_device *adev = ring->adev;
378 if (ring != &adev->uvd.ring) {
379 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
380 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
382 /* put fence directly behind firmware */
383 index = ALIGN(adev->uvd.fw->size, 8);
384 ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index;
385 ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index;
387 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
388 amdgpu_irq_get(adev, irq_src, irq_type);
390 ring->fence_drv.irq_src = irq_src;
391 ring->fence_drv.irq_type = irq_type;
392 ring->fence_drv.initialized = true;
394 dev_dbg(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, "
395 "cpu addr 0x%p\n", ring->idx,
396 ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
401 * amdgpu_fence_driver_init_ring - init the fence driver
402 * for the requested ring.
404 * @ring: ring to init the fence driver on
405 * @num_hw_submission: number of entries on the hardware queue
407 * Init the fence driver for the requested ring (all asics).
408 * Helper function for amdgpu_fence_driver_init().
410 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
411 unsigned num_hw_submission)
415 /* Check that num_hw_submission is a power of two */
416 if ((num_hw_submission & (num_hw_submission - 1)) != 0)
419 ring->fence_drv.cpu_addr = NULL;
420 ring->fence_drv.gpu_addr = 0;
421 ring->fence_drv.sync_seq = 0;
422 atomic_set(&ring->fence_drv.last_seq, 0);
423 ring->fence_drv.initialized = false;
425 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
427 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
428 spin_lock_init(&ring->fence_drv.lock);
429 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
431 if (!ring->fence_drv.fences)
434 /* No need to setup the GPU scheduler for KIQ ring */
435 if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) {
436 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
437 num_hw_submission, amdgpu_job_hang_limit,
438 msecs_to_jiffies(amdgpu_lockup_timeout), ring->name);
440 DRM_ERROR("Failed to create scheduler on ring %s.\n",
450 * amdgpu_fence_driver_init - init the fence driver
451 * for all possible rings.
453 * @adev: amdgpu device pointer
455 * Init the fence driver for all possible rings (all asics).
456 * Not all asics have all rings, so each asic will only
457 * start the fence driver on the rings it has using
458 * amdgpu_fence_driver_start_ring().
459 * Returns 0 for success.
461 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
463 if (amdgpu_debugfs_fence_init(adev))
464 dev_err(adev->dev, "fence debugfs file creation failed\n");
470 * amdgpu_fence_driver_fini - tear down the fence driver
471 * for all possible rings.
473 * @adev: amdgpu device pointer
475 * Tear down the fence driver for all possible rings (all asics).
477 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
482 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
483 struct amdgpu_ring *ring = adev->rings[i];
485 if (!ring || !ring->fence_drv.initialized)
487 r = amdgpu_fence_wait_empty(ring);
489 /* no need to trigger GPU reset as we are unloading */
490 amdgpu_fence_driver_force_completion(ring);
492 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
493 ring->fence_drv.irq_type);
494 drm_sched_fini(&ring->sched);
495 del_timer_sync(&ring->fence_drv.fallback_timer);
496 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
497 dma_fence_put(ring->fence_drv.fences[j]);
498 kfree(ring->fence_drv.fences);
499 ring->fence_drv.fences = NULL;
500 ring->fence_drv.initialized = false;
505 * amdgpu_fence_driver_suspend - suspend the fence driver
506 * for all possible rings.
508 * @adev: amdgpu device pointer
510 * Suspend the fence driver for all possible rings (all asics).
512 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
516 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
517 struct amdgpu_ring *ring = adev->rings[i];
518 if (!ring || !ring->fence_drv.initialized)
521 /* wait for gpu to finish processing current batch */
522 r = amdgpu_fence_wait_empty(ring);
524 /* delay GPU reset to resume */
525 amdgpu_fence_driver_force_completion(ring);
528 /* disable the interrupt */
529 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
530 ring->fence_drv.irq_type);
535 * amdgpu_fence_driver_resume - resume the fence driver
536 * for all possible rings.
538 * @adev: amdgpu device pointer
540 * Resume the fence driver for all possible rings (all asics).
541 * Not all asics have all rings, so each asic will only
542 * start the fence driver on the rings it has using
543 * amdgpu_fence_driver_start_ring().
544 * Returns 0 for success.
546 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
550 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
551 struct amdgpu_ring *ring = adev->rings[i];
552 if (!ring || !ring->fence_drv.initialized)
555 /* enable the interrupt */
556 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
557 ring->fence_drv.irq_type);
562 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
564 * @ring: fence of the ring to signal
567 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
569 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
570 amdgpu_fence_process(ring);
574 * Common fence implementation
577 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
582 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
584 struct amdgpu_fence *fence = to_amdgpu_fence(f);
585 return (const char *)fence->ring->name;
589 * amdgpu_fence_enable_signaling - enable signalling on fence
592 * This function is called with fence_queue lock held, and adds a callback
593 * to fence_queue that checks if this fence is signaled, and if so it
594 * signals the fence and removes itself.
596 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
598 struct amdgpu_fence *fence = to_amdgpu_fence(f);
599 struct amdgpu_ring *ring = fence->ring;
601 if (!timer_pending(&ring->fence_drv.fallback_timer))
602 amdgpu_fence_schedule_fallback(ring);
604 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
610 * amdgpu_fence_free - free up the fence memory
612 * @rcu: RCU callback head
614 * Free up the fence memory after the RCU grace period.
616 static void amdgpu_fence_free(struct rcu_head *rcu)
618 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
619 struct amdgpu_fence *fence = to_amdgpu_fence(f);
620 kmem_cache_free(amdgpu_fence_slab, fence);
624 * amdgpu_fence_release - callback that fence can be freed
628 * This function is called when the reference count becomes zero.
629 * It just RCU schedules freeing up the fence.
631 static void amdgpu_fence_release(struct dma_fence *f)
633 call_rcu(&f->rcu, amdgpu_fence_free);
636 static const struct dma_fence_ops amdgpu_fence_ops = {
637 .get_driver_name = amdgpu_fence_get_driver_name,
638 .get_timeline_name = amdgpu_fence_get_timeline_name,
639 .enable_signaling = amdgpu_fence_enable_signaling,
640 .wait = dma_fence_default_wait,
641 .release = amdgpu_fence_release,
647 #if defined(CONFIG_DEBUG_FS)
648 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
650 struct drm_info_node *node = (struct drm_info_node *)m->private;
651 struct drm_device *dev = node->minor->dev;
652 struct amdgpu_device *adev = dev->dev_private;
655 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
656 struct amdgpu_ring *ring = adev->rings[i];
657 if (!ring || !ring->fence_drv.initialized)
660 amdgpu_fence_process(ring);
662 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
663 seq_printf(m, "Last signaled fence 0x%08x\n",
664 atomic_read(&ring->fence_drv.last_seq));
665 seq_printf(m, "Last emitted 0x%08x\n",
666 ring->fence_drv.sync_seq);
668 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
671 /* set in CP_VMID_PREEMPT and preemption occurred */
672 seq_printf(m, "Last preempted 0x%08x\n",
673 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
674 /* set in CP_VMID_RESET and reset occurred */
675 seq_printf(m, "Last reset 0x%08x\n",
676 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
677 /* Both preemption and reset occurred */
678 seq_printf(m, "Last both 0x%08x\n",
679 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
685 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
687 * Manually trigger a gpu reset at the next fence wait.
689 static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
691 struct drm_info_node *node = (struct drm_info_node *) m->private;
692 struct drm_device *dev = node->minor->dev;
693 struct amdgpu_device *adev = dev->dev_private;
695 seq_printf(m, "gpu recover\n");
696 amdgpu_device_gpu_recover(adev, NULL, true);
701 static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
702 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
703 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
706 static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
707 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
711 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
713 #if defined(CONFIG_DEBUG_FS)
714 if (amdgpu_sriov_vf(adev))
715 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1);
716 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);