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 *old, **ptr;
142 fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
146 seq = ++ring->fence_drv.sync_seq;
148 dma_fence_init(&fence->base, &amdgpu_fence_ops,
149 &ring->fence_drv.lock,
150 adev->fence_context + ring->idx,
152 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
153 seq, flags | AMDGPU_FENCE_FLAG_INT);
155 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
156 /* This function can't be called concurrently anyway, otherwise
157 * emitting the fence would mess up the hardware ring buffer.
159 old = rcu_dereference_protected(*ptr, 1);
160 if (old && !dma_fence_is_signaled(old)) {
161 DRM_INFO("rcu slot is busy\n");
162 dma_fence_wait(old, false);
165 rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
173 * amdgpu_fence_emit_polling - emit a fence on the requeste ring
175 * @ring: ring the fence is associated with
176 * @s: resulting sequence number
178 * Emits a fence command on the requested ring (all asics).
179 * Used For polling fence.
180 * Returns 0 on success, -ENOMEM on failure.
182 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s)
189 seq = ++ring->fence_drv.sync_seq;
190 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
199 * amdgpu_fence_schedule_fallback - schedule fallback check
201 * @ring: pointer to struct amdgpu_ring
203 * Start a timer as fallback to our interrupts.
205 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
207 mod_timer(&ring->fence_drv.fallback_timer,
208 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
212 * amdgpu_fence_process - check for fence activity
214 * @ring: pointer to struct amdgpu_ring
216 * Checks the current fence value and calculates the last
217 * signalled fence value. Wakes the fence queue if the
218 * sequence number has increased.
220 void amdgpu_fence_process(struct amdgpu_ring *ring)
222 struct amdgpu_fence_driver *drv = &ring->fence_drv;
223 uint32_t seq, last_seq;
227 last_seq = atomic_read(&ring->fence_drv.last_seq);
228 seq = amdgpu_fence_read(ring);
230 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
232 if (seq != ring->fence_drv.sync_seq)
233 amdgpu_fence_schedule_fallback(ring);
235 if (unlikely(seq == last_seq))
238 last_seq &= drv->num_fences_mask;
239 seq &= drv->num_fences_mask;
242 struct dma_fence *fence, **ptr;
245 last_seq &= drv->num_fences_mask;
246 ptr = &drv->fences[last_seq];
248 /* There is always exactly one thread signaling this fence slot */
249 fence = rcu_dereference_protected(*ptr, 1);
250 RCU_INIT_POINTER(*ptr, NULL);
255 r = dma_fence_signal(fence);
257 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
261 dma_fence_put(fence);
262 } while (last_seq != seq);
266 * amdgpu_fence_fallback - fallback for hardware interrupts
268 * @work: delayed work item
270 * Checks for fence activity.
272 static void amdgpu_fence_fallback(struct timer_list *t)
274 struct amdgpu_ring *ring = from_timer(ring, t,
275 fence_drv.fallback_timer);
277 amdgpu_fence_process(ring);
281 * amdgpu_fence_wait_empty - wait for all fences to signal
283 * @adev: amdgpu device pointer
284 * @ring: ring index the fence is associated with
286 * Wait for all fences on the requested ring to signal (all asics).
287 * Returns 0 if the fences have passed, error for all other cases.
289 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
291 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
292 struct dma_fence *fence, **ptr;
298 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
300 fence = rcu_dereference(*ptr);
301 if (!fence || !dma_fence_get_rcu(fence)) {
307 r = dma_fence_wait(fence, false);
308 dma_fence_put(fence);
313 * amdgpu_fence_wait_polling - busy wait for givn sequence number
315 * @ring: ring index the fence is associated with
316 * @wait_seq: sequence number to wait
317 * @timeout: the timeout for waiting in usecs
319 * Wait for all fences on the requested ring to signal (all asics).
320 * Returns left time if no timeout, 0 or minus if timeout.
322 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
329 seq = amdgpu_fence_read(ring);
332 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
334 return timeout > 0 ? timeout : 0;
337 * amdgpu_fence_count_emitted - get the count of emitted fences
339 * @ring: ring the fence is associated with
341 * Get the number of fences emitted on the requested ring (all asics).
342 * Returns the number of emitted fences on the ring. Used by the
343 * dynpm code to ring track activity.
345 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
349 /* We are not protected by ring lock when reading the last sequence
350 * but it's ok to report slightly wrong fence count here.
352 amdgpu_fence_process(ring);
353 emitted = 0x100000000ull;
354 emitted -= atomic_read(&ring->fence_drv.last_seq);
355 emitted += READ_ONCE(ring->fence_drv.sync_seq);
356 return lower_32_bits(emitted);
360 * amdgpu_fence_driver_start_ring - make the fence driver
361 * ready for use on the requested ring.
363 * @ring: ring to start the fence driver on
364 * @irq_src: interrupt source to use for this ring
365 * @irq_type: interrupt type to use for this ring
367 * Make the fence driver ready for processing (all asics).
368 * Not all asics have all rings, so each asic will only
369 * start the fence driver on the rings it has.
370 * Returns 0 for success, errors for failure.
372 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
373 struct amdgpu_irq_src *irq_src,
376 struct amdgpu_device *adev = ring->adev;
379 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
380 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
381 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
383 /* put fence directly behind firmware */
384 index = ALIGN(adev->uvd.fw->size, 8);
385 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
386 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
388 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
389 amdgpu_irq_get(adev, irq_src, irq_type);
391 ring->fence_drv.irq_src = irq_src;
392 ring->fence_drv.irq_type = irq_type;
393 ring->fence_drv.initialized = true;
395 dev_dbg(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, "
396 "cpu addr 0x%p\n", ring->idx,
397 ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
402 * amdgpu_fence_driver_init_ring - init the fence driver
403 * for the requested ring.
405 * @ring: ring to init the fence driver on
406 * @num_hw_submission: number of entries on the hardware queue
408 * Init the fence driver for the requested ring (all asics).
409 * Helper function for amdgpu_fence_driver_init().
411 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
412 unsigned num_hw_submission)
417 /* Check that num_hw_submission is a power of two */
418 if ((num_hw_submission & (num_hw_submission - 1)) != 0)
421 ring->fence_drv.cpu_addr = NULL;
422 ring->fence_drv.gpu_addr = 0;
423 ring->fence_drv.sync_seq = 0;
424 atomic_set(&ring->fence_drv.last_seq, 0);
425 ring->fence_drv.initialized = false;
427 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
429 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
430 spin_lock_init(&ring->fence_drv.lock);
431 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
433 if (!ring->fence_drv.fences)
436 /* No need to setup the GPU scheduler for KIQ ring */
437 if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) {
438 /* for non-sriov case, no timeout enforce on compute ring */
439 if ((ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
440 && !amdgpu_sriov_vf(ring->adev))
441 timeout = MAX_SCHEDULE_TIMEOUT;
443 timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
445 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
446 num_hw_submission, amdgpu_job_hang_limit,
447 timeout, ring->name);
449 DRM_ERROR("Failed to create scheduler on ring %s.\n",
459 * amdgpu_fence_driver_init - init the fence driver
460 * for all possible rings.
462 * @adev: amdgpu device pointer
464 * Init the fence driver for all possible rings (all asics).
465 * Not all asics have all rings, so each asic will only
466 * start the fence driver on the rings it has using
467 * amdgpu_fence_driver_start_ring().
468 * Returns 0 for success.
470 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
472 if (amdgpu_debugfs_fence_init(adev))
473 dev_err(adev->dev, "fence debugfs file creation failed\n");
479 * amdgpu_fence_driver_fini - tear down the fence driver
480 * for all possible rings.
482 * @adev: amdgpu device pointer
484 * Tear down the fence driver for all possible rings (all asics).
486 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
491 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
492 struct amdgpu_ring *ring = adev->rings[i];
494 if (!ring || !ring->fence_drv.initialized)
496 r = amdgpu_fence_wait_empty(ring);
498 /* no need to trigger GPU reset as we are unloading */
499 amdgpu_fence_driver_force_completion(ring);
501 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
502 ring->fence_drv.irq_type);
503 drm_sched_fini(&ring->sched);
504 del_timer_sync(&ring->fence_drv.fallback_timer);
505 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
506 dma_fence_put(ring->fence_drv.fences[j]);
507 kfree(ring->fence_drv.fences);
508 ring->fence_drv.fences = NULL;
509 ring->fence_drv.initialized = false;
514 * amdgpu_fence_driver_suspend - suspend the fence driver
515 * for all possible rings.
517 * @adev: amdgpu device pointer
519 * Suspend the fence driver for all possible rings (all asics).
521 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
525 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
526 struct amdgpu_ring *ring = adev->rings[i];
527 if (!ring || !ring->fence_drv.initialized)
530 /* wait for gpu to finish processing current batch */
531 r = amdgpu_fence_wait_empty(ring);
533 /* delay GPU reset to resume */
534 amdgpu_fence_driver_force_completion(ring);
537 /* disable the interrupt */
538 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
539 ring->fence_drv.irq_type);
544 * amdgpu_fence_driver_resume - resume the fence driver
545 * for all possible rings.
547 * @adev: amdgpu device pointer
549 * Resume the fence driver for all possible rings (all asics).
550 * Not all asics have all rings, so each asic will only
551 * start the fence driver on the rings it has using
552 * amdgpu_fence_driver_start_ring().
553 * Returns 0 for success.
555 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
559 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
560 struct amdgpu_ring *ring = adev->rings[i];
561 if (!ring || !ring->fence_drv.initialized)
564 /* enable the interrupt */
565 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
566 ring->fence_drv.irq_type);
571 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
573 * @ring: fence of the ring to signal
576 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
578 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
579 amdgpu_fence_process(ring);
583 * Common fence implementation
586 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
591 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
593 struct amdgpu_fence *fence = to_amdgpu_fence(f);
594 return (const char *)fence->ring->name;
598 * amdgpu_fence_enable_signaling - enable signalling on fence
601 * This function is called with fence_queue lock held, and adds a callback
602 * to fence_queue that checks if this fence is signaled, and if so it
603 * signals the fence and removes itself.
605 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
607 struct amdgpu_fence *fence = to_amdgpu_fence(f);
608 struct amdgpu_ring *ring = fence->ring;
610 if (!timer_pending(&ring->fence_drv.fallback_timer))
611 amdgpu_fence_schedule_fallback(ring);
613 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
619 * amdgpu_fence_free - free up the fence memory
621 * @rcu: RCU callback head
623 * Free up the fence memory after the RCU grace period.
625 static void amdgpu_fence_free(struct rcu_head *rcu)
627 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
628 struct amdgpu_fence *fence = to_amdgpu_fence(f);
629 kmem_cache_free(amdgpu_fence_slab, fence);
633 * amdgpu_fence_release - callback that fence can be freed
637 * This function is called when the reference count becomes zero.
638 * It just RCU schedules freeing up the fence.
640 static void amdgpu_fence_release(struct dma_fence *f)
642 call_rcu(&f->rcu, amdgpu_fence_free);
645 static const struct dma_fence_ops amdgpu_fence_ops = {
646 .get_driver_name = amdgpu_fence_get_driver_name,
647 .get_timeline_name = amdgpu_fence_get_timeline_name,
648 .enable_signaling = amdgpu_fence_enable_signaling,
649 .release = amdgpu_fence_release,
655 #if defined(CONFIG_DEBUG_FS)
656 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
658 struct drm_info_node *node = (struct drm_info_node *)m->private;
659 struct drm_device *dev = node->minor->dev;
660 struct amdgpu_device *adev = dev->dev_private;
663 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
664 struct amdgpu_ring *ring = adev->rings[i];
665 if (!ring || !ring->fence_drv.initialized)
668 amdgpu_fence_process(ring);
670 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
671 seq_printf(m, "Last signaled fence 0x%08x\n",
672 atomic_read(&ring->fence_drv.last_seq));
673 seq_printf(m, "Last emitted 0x%08x\n",
674 ring->fence_drv.sync_seq);
676 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
679 /* set in CP_VMID_PREEMPT and preemption occurred */
680 seq_printf(m, "Last preempted 0x%08x\n",
681 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
682 /* set in CP_VMID_RESET and reset occurred */
683 seq_printf(m, "Last reset 0x%08x\n",
684 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
685 /* Both preemption and reset occurred */
686 seq_printf(m, "Last both 0x%08x\n",
687 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
693 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
695 * Manually trigger a gpu reset at the next fence wait.
697 static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
699 struct drm_info_node *node = (struct drm_info_node *) m->private;
700 struct drm_device *dev = node->minor->dev;
701 struct amdgpu_device *adev = dev->dev_private;
703 seq_printf(m, "gpu recover\n");
704 amdgpu_device_gpu_recover(adev, NULL, true);
709 static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
710 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
711 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
714 static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
715 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
719 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
721 #if defined(CONFIG_DEBUG_FS)
722 if (amdgpu_sriov_vf(adev))
723 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1);
724 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);