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 DRM_INFO("Fallback to SW interrupt on ring %s due to HW interrupt time out", ring->name);
278 amdgpu_fence_process(ring);
282 * amdgpu_fence_wait_empty - wait for all fences to signal
284 * @adev: amdgpu device pointer
285 * @ring: ring index the fence is associated with
287 * Wait for all fences on the requested ring to signal (all asics).
288 * Returns 0 if the fences have passed, error for all other cases.
290 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
292 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
293 struct dma_fence *fence, **ptr;
299 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
301 fence = rcu_dereference(*ptr);
302 if (!fence || !dma_fence_get_rcu(fence)) {
308 r = dma_fence_wait(fence, false);
309 dma_fence_put(fence);
314 * amdgpu_fence_wait_polling - busy wait for givn sequence number
316 * @ring: ring index the fence is associated with
317 * @wait_seq: sequence number to wait
318 * @timeout: the timeout for waiting in usecs
320 * Wait for all fences on the requested ring to signal (all asics).
321 * Returns left time if no timeout, 0 or minus if timeout.
323 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
330 seq = amdgpu_fence_read(ring);
333 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
335 return timeout > 0 ? timeout : 0;
338 * amdgpu_fence_count_emitted - get the count of emitted fences
340 * @ring: ring the fence is associated with
342 * Get the number of fences emitted on the requested ring (all asics).
343 * Returns the number of emitted fences on the ring. Used by the
344 * dynpm code to ring track activity.
346 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
350 /* We are not protected by ring lock when reading the last sequence
351 * but it's ok to report slightly wrong fence count here.
353 amdgpu_fence_process(ring);
354 emitted = 0x100000000ull;
355 emitted -= atomic_read(&ring->fence_drv.last_seq);
356 emitted += READ_ONCE(ring->fence_drv.sync_seq);
357 return lower_32_bits(emitted);
361 * amdgpu_fence_driver_start_ring - make the fence driver
362 * ready for use on the requested ring.
364 * @ring: ring to start the fence driver on
365 * @irq_src: interrupt source to use for this ring
366 * @irq_type: interrupt type to use for this ring
368 * Make the fence driver ready for processing (all asics).
369 * Not all asics have all rings, so each asic will only
370 * start the fence driver on the rings it has.
371 * Returns 0 for success, errors for failure.
373 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
374 struct amdgpu_irq_src *irq_src,
377 struct amdgpu_device *adev = ring->adev;
380 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
381 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
382 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
384 /* put fence directly behind firmware */
385 index = ALIGN(adev->uvd.fw->size, 8);
386 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
387 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
389 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
390 amdgpu_irq_get(adev, irq_src, irq_type);
392 ring->fence_drv.irq_src = irq_src;
393 ring->fence_drv.irq_type = irq_type;
394 ring->fence_drv.initialized = true;
396 dev_dbg(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, "
397 "cpu addr 0x%p\n", ring->idx,
398 ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
403 * amdgpu_fence_driver_init_ring - init the fence driver
404 * for the requested ring.
406 * @ring: ring to init the fence driver on
407 * @num_hw_submission: number of entries on the hardware queue
409 * Init the fence driver for the requested ring (all asics).
410 * Helper function for amdgpu_fence_driver_init().
412 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
413 unsigned num_hw_submission)
418 /* Check that num_hw_submission is a power of two */
419 if ((num_hw_submission & (num_hw_submission - 1)) != 0)
422 ring->fence_drv.cpu_addr = NULL;
423 ring->fence_drv.gpu_addr = 0;
424 ring->fence_drv.sync_seq = 0;
425 atomic_set(&ring->fence_drv.last_seq, 0);
426 ring->fence_drv.initialized = false;
428 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
430 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
431 spin_lock_init(&ring->fence_drv.lock);
432 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
434 if (!ring->fence_drv.fences)
437 /* No need to setup the GPU scheduler for KIQ ring */
438 if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) {
439 /* for non-sriov case, no timeout enforce on compute ring */
440 if ((ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
441 && !amdgpu_sriov_vf(ring->adev))
442 timeout = MAX_SCHEDULE_TIMEOUT;
444 timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
446 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
447 num_hw_submission, amdgpu_job_hang_limit,
448 timeout, ring->name);
450 DRM_ERROR("Failed to create scheduler on ring %s.\n",
460 * amdgpu_fence_driver_init - init the fence driver
461 * for all possible rings.
463 * @adev: amdgpu device pointer
465 * Init the fence driver for all possible rings (all asics).
466 * Not all asics have all rings, so each asic will only
467 * start the fence driver on the rings it has using
468 * amdgpu_fence_driver_start_ring().
469 * Returns 0 for success.
471 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
473 if (amdgpu_debugfs_fence_init(adev))
474 dev_err(adev->dev, "fence debugfs file creation failed\n");
480 * amdgpu_fence_driver_fini - tear down the fence driver
481 * for all possible rings.
483 * @adev: amdgpu device pointer
485 * Tear down the fence driver for all possible rings (all asics).
487 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
492 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
493 struct amdgpu_ring *ring = adev->rings[i];
495 if (!ring || !ring->fence_drv.initialized)
497 r = amdgpu_fence_wait_empty(ring);
499 /* no need to trigger GPU reset as we are unloading */
500 amdgpu_fence_driver_force_completion(ring);
502 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
503 ring->fence_drv.irq_type);
504 drm_sched_fini(&ring->sched);
505 del_timer_sync(&ring->fence_drv.fallback_timer);
506 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
507 dma_fence_put(ring->fence_drv.fences[j]);
508 kfree(ring->fence_drv.fences);
509 ring->fence_drv.fences = NULL;
510 ring->fence_drv.initialized = false;
515 * amdgpu_fence_driver_suspend - suspend the fence driver
516 * for all possible rings.
518 * @adev: amdgpu device pointer
520 * Suspend the fence driver for all possible rings (all asics).
522 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
526 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
527 struct amdgpu_ring *ring = adev->rings[i];
528 if (!ring || !ring->fence_drv.initialized)
531 /* wait for gpu to finish processing current batch */
532 r = amdgpu_fence_wait_empty(ring);
534 /* delay GPU reset to resume */
535 amdgpu_fence_driver_force_completion(ring);
538 /* disable the interrupt */
539 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
540 ring->fence_drv.irq_type);
545 * amdgpu_fence_driver_resume - resume the fence driver
546 * for all possible rings.
548 * @adev: amdgpu device pointer
550 * Resume the fence driver for all possible rings (all asics).
551 * Not all asics have all rings, so each asic will only
552 * start the fence driver on the rings it has using
553 * amdgpu_fence_driver_start_ring().
554 * Returns 0 for success.
556 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
560 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
561 struct amdgpu_ring *ring = adev->rings[i];
562 if (!ring || !ring->fence_drv.initialized)
565 /* enable the interrupt */
566 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
567 ring->fence_drv.irq_type);
572 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
574 * @ring: fence of the ring to signal
577 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
579 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
580 amdgpu_fence_process(ring);
584 * Common fence implementation
587 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
592 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
594 struct amdgpu_fence *fence = to_amdgpu_fence(f);
595 return (const char *)fence->ring->name;
599 * amdgpu_fence_enable_signaling - enable signalling on fence
602 * This function is called with fence_queue lock held, and adds a callback
603 * to fence_queue that checks if this fence is signaled, and if so it
604 * signals the fence and removes itself.
606 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
608 struct amdgpu_fence *fence = to_amdgpu_fence(f);
609 struct amdgpu_ring *ring = fence->ring;
611 if (!timer_pending(&ring->fence_drv.fallback_timer))
612 amdgpu_fence_schedule_fallback(ring);
614 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
620 * amdgpu_fence_free - free up the fence memory
622 * @rcu: RCU callback head
624 * Free up the fence memory after the RCU grace period.
626 static void amdgpu_fence_free(struct rcu_head *rcu)
628 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
629 struct amdgpu_fence *fence = to_amdgpu_fence(f);
630 kmem_cache_free(amdgpu_fence_slab, fence);
634 * amdgpu_fence_release - callback that fence can be freed
638 * This function is called when the reference count becomes zero.
639 * It just RCU schedules freeing up the fence.
641 static void amdgpu_fence_release(struct dma_fence *f)
643 call_rcu(&f->rcu, amdgpu_fence_free);
646 static const struct dma_fence_ops amdgpu_fence_ops = {
647 .get_driver_name = amdgpu_fence_get_driver_name,
648 .get_timeline_name = amdgpu_fence_get_timeline_name,
649 .enable_signaling = amdgpu_fence_enable_signaling,
650 .release = amdgpu_fence_release,
656 #if defined(CONFIG_DEBUG_FS)
657 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
659 struct drm_info_node *node = (struct drm_info_node *)m->private;
660 struct drm_device *dev = node->minor->dev;
661 struct amdgpu_device *adev = dev->dev_private;
664 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
665 struct amdgpu_ring *ring = adev->rings[i];
666 if (!ring || !ring->fence_drv.initialized)
669 amdgpu_fence_process(ring);
671 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
672 seq_printf(m, "Last signaled fence 0x%08x\n",
673 atomic_read(&ring->fence_drv.last_seq));
674 seq_printf(m, "Last emitted 0x%08x\n",
675 ring->fence_drv.sync_seq);
677 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
680 /* set in CP_VMID_PREEMPT and preemption occurred */
681 seq_printf(m, "Last preempted 0x%08x\n",
682 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
683 /* set in CP_VMID_RESET and reset occurred */
684 seq_printf(m, "Last reset 0x%08x\n",
685 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
686 /* Both preemption and reset occurred */
687 seq_printf(m, "Last both 0x%08x\n",
688 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
694 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
696 * Manually trigger a gpu reset at the next fence wait.
698 static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
700 struct drm_info_node *node = (struct drm_info_node *) m->private;
701 struct drm_device *dev = node->minor->dev;
702 struct amdgpu_device *adev = dev->dev_private;
704 seq_printf(m, "gpu recover\n");
705 amdgpu_device_gpu_recover(adev, NULL);
710 static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
711 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
712 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
715 static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
716 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
720 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
722 #if defined(CONFIG_DEBUG_FS)
723 if (amdgpu_sriov_vf(adev))
724 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1);
725 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);