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_process - check for fence activity
201 * @ring: pointer to struct amdgpu_ring
203 * Checks the current fence value and calculates the last
204 * signalled fence value. Wakes the fence queue if the
205 * sequence number has increased.
207 void amdgpu_fence_process(struct amdgpu_ring *ring)
209 struct amdgpu_fence_driver *drv = &ring->fence_drv;
210 uint32_t seq, last_seq;
214 last_seq = atomic_read(&ring->fence_drv.last_seq);
215 seq = amdgpu_fence_read(ring);
217 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
219 if (unlikely(seq == last_seq))
222 last_seq &= drv->num_fences_mask;
223 seq &= drv->num_fences_mask;
226 struct dma_fence *fence, **ptr;
229 last_seq &= drv->num_fences_mask;
230 ptr = &drv->fences[last_seq];
232 /* There is always exactly one thread signaling this fence slot */
233 fence = rcu_dereference_protected(*ptr, 1);
234 RCU_INIT_POINTER(*ptr, NULL);
239 r = dma_fence_signal(fence);
241 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
245 dma_fence_put(fence);
246 } while (last_seq != seq);
250 * amdgpu_fence_wait_empty - wait for all fences to signal
252 * @adev: amdgpu device pointer
253 * @ring: ring index the fence is associated with
255 * Wait for all fences on the requested ring to signal (all asics).
256 * Returns 0 if the fences have passed, error for all other cases.
258 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
260 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
261 struct dma_fence *fence, **ptr;
267 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
269 fence = rcu_dereference(*ptr);
270 if (!fence || !dma_fence_get_rcu(fence)) {
276 r = dma_fence_wait(fence, false);
277 dma_fence_put(fence);
282 * amdgpu_fence_wait_polling - busy wait for givn sequence number
284 * @ring: ring index the fence is associated with
285 * @wait_seq: sequence number to wait
286 * @timeout: the timeout for waiting in usecs
288 * Wait for all fences on the requested ring to signal (all asics).
289 * Returns left time if no timeout, 0 or minus if timeout.
291 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
298 seq = amdgpu_fence_read(ring);
301 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
303 return timeout > 0 ? timeout : 0;
306 * amdgpu_fence_count_emitted - get the count of emitted fences
308 * @ring: ring the fence is associated with
310 * Get the number of fences emitted on the requested ring (all asics).
311 * Returns the number of emitted fences on the ring. Used by the
312 * dynpm code to ring track activity.
314 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
318 /* We are not protected by ring lock when reading the last sequence
319 * but it's ok to report slightly wrong fence count here.
321 amdgpu_fence_process(ring);
322 emitted = 0x100000000ull;
323 emitted -= atomic_read(&ring->fence_drv.last_seq);
324 emitted += READ_ONCE(ring->fence_drv.sync_seq);
325 return lower_32_bits(emitted);
329 * amdgpu_fence_driver_start_ring - make the fence driver
330 * ready for use on the requested ring.
332 * @ring: ring to start the fence driver on
333 * @irq_src: interrupt source to use for this ring
334 * @irq_type: interrupt type to use for this ring
336 * Make the fence driver ready for processing (all asics).
337 * Not all asics have all rings, so each asic will only
338 * start the fence driver on the rings it has.
339 * Returns 0 for success, errors for failure.
341 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
342 struct amdgpu_irq_src *irq_src,
345 struct amdgpu_device *adev = ring->adev;
348 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
349 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
350 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
352 /* put fence directly behind firmware */
353 index = ALIGN(adev->uvd.fw->size, 8);
354 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
355 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
357 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
358 amdgpu_irq_get(adev, irq_src, irq_type);
360 ring->fence_drv.irq_src = irq_src;
361 ring->fence_drv.irq_type = irq_type;
362 ring->fence_drv.initialized = true;
364 dev_dbg(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, "
365 "cpu addr 0x%p\n", ring->idx,
366 ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
371 * amdgpu_fence_driver_init_ring - init the fence driver
372 * for the requested ring.
374 * @ring: ring to init the fence driver on
375 * @num_hw_submission: number of entries on the hardware queue
377 * Init the fence driver for the requested ring (all asics).
378 * Helper function for amdgpu_fence_driver_init().
380 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
381 unsigned num_hw_submission)
386 /* Check that num_hw_submission is a power of two */
387 if ((num_hw_submission & (num_hw_submission - 1)) != 0)
390 ring->fence_drv.cpu_addr = NULL;
391 ring->fence_drv.gpu_addr = 0;
392 ring->fence_drv.sync_seq = 0;
393 atomic_set(&ring->fence_drv.last_seq, 0);
394 ring->fence_drv.initialized = false;
396 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
397 spin_lock_init(&ring->fence_drv.lock);
398 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
400 if (!ring->fence_drv.fences)
403 /* No need to setup the GPU scheduler for KIQ ring */
404 if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) {
405 /* for non-sriov case, no timeout enforce on compute ring */
406 if ((ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
407 && !amdgpu_sriov_vf(ring->adev))
408 timeout = MAX_SCHEDULE_TIMEOUT;
410 timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
412 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
413 num_hw_submission, amdgpu_job_hang_limit,
414 timeout, ring->name);
416 DRM_ERROR("Failed to create scheduler on ring %s.\n",
426 * amdgpu_fence_driver_init - init the fence driver
427 * for all possible rings.
429 * @adev: amdgpu device pointer
431 * Init the fence driver for all possible rings (all asics).
432 * Not all asics have all rings, so each asic will only
433 * start the fence driver on the rings it has using
434 * amdgpu_fence_driver_start_ring().
435 * Returns 0 for success.
437 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
439 if (amdgpu_debugfs_fence_init(adev))
440 dev_err(adev->dev, "fence debugfs file creation failed\n");
446 * amdgpu_fence_driver_fini - tear down the fence driver
447 * for all possible rings.
449 * @adev: amdgpu device pointer
451 * Tear down the fence driver for all possible rings (all asics).
453 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
458 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
459 struct amdgpu_ring *ring = adev->rings[i];
461 if (!ring || !ring->fence_drv.initialized)
463 r = amdgpu_fence_wait_empty(ring);
465 /* no need to trigger GPU reset as we are unloading */
466 amdgpu_fence_driver_force_completion(ring);
468 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
469 ring->fence_drv.irq_type);
470 drm_sched_fini(&ring->sched);
471 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
472 dma_fence_put(ring->fence_drv.fences[j]);
473 kfree(ring->fence_drv.fences);
474 ring->fence_drv.fences = NULL;
475 ring->fence_drv.initialized = false;
480 * amdgpu_fence_driver_suspend - suspend the fence driver
481 * for all possible rings.
483 * @adev: amdgpu device pointer
485 * Suspend the fence driver for all possible rings (all asics).
487 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
491 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
492 struct amdgpu_ring *ring = adev->rings[i];
493 if (!ring || !ring->fence_drv.initialized)
496 /* wait for gpu to finish processing current batch */
497 r = amdgpu_fence_wait_empty(ring);
499 /* delay GPU reset to resume */
500 amdgpu_fence_driver_force_completion(ring);
503 /* disable the interrupt */
504 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
505 ring->fence_drv.irq_type);
510 * amdgpu_fence_driver_resume - resume the fence driver
511 * for all possible rings.
513 * @adev: amdgpu device pointer
515 * Resume the fence driver for all possible rings (all asics).
516 * Not all asics have all rings, so each asic will only
517 * start the fence driver on the rings it has using
518 * amdgpu_fence_driver_start_ring().
519 * Returns 0 for success.
521 void amdgpu_fence_driver_resume(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 /* enable the interrupt */
531 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
532 ring->fence_drv.irq_type);
537 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
539 * @ring: fence of the ring to signal
542 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
544 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
545 amdgpu_fence_process(ring);
549 * Common fence implementation
552 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
557 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
559 struct amdgpu_fence *fence = to_amdgpu_fence(f);
560 return (const char *)fence->ring->name;
564 * amdgpu_fence_free - free up the fence memory
566 * @rcu: RCU callback head
568 * Free up the fence memory after the RCU grace period.
570 static void amdgpu_fence_free(struct rcu_head *rcu)
572 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
573 struct amdgpu_fence *fence = to_amdgpu_fence(f);
574 kmem_cache_free(amdgpu_fence_slab, fence);
578 * amdgpu_fence_release - callback that fence can be freed
582 * This function is called when the reference count becomes zero.
583 * It just RCU schedules freeing up the fence.
585 static void amdgpu_fence_release(struct dma_fence *f)
587 call_rcu(&f->rcu, amdgpu_fence_free);
590 static const struct dma_fence_ops amdgpu_fence_ops = {
591 .get_driver_name = amdgpu_fence_get_driver_name,
592 .get_timeline_name = amdgpu_fence_get_timeline_name,
593 .release = amdgpu_fence_release,
599 #if defined(CONFIG_DEBUG_FS)
600 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
602 struct drm_info_node *node = (struct drm_info_node *)m->private;
603 struct drm_device *dev = node->minor->dev;
604 struct amdgpu_device *adev = dev->dev_private;
607 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
608 struct amdgpu_ring *ring = adev->rings[i];
609 if (!ring || !ring->fence_drv.initialized)
612 amdgpu_fence_process(ring);
614 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
615 seq_printf(m, "Last signaled fence 0x%08x\n",
616 atomic_read(&ring->fence_drv.last_seq));
617 seq_printf(m, "Last emitted 0x%08x\n",
618 ring->fence_drv.sync_seq);
620 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
623 /* set in CP_VMID_PREEMPT and preemption occurred */
624 seq_printf(m, "Last preempted 0x%08x\n",
625 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
626 /* set in CP_VMID_RESET and reset occurred */
627 seq_printf(m, "Last reset 0x%08x\n",
628 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
629 /* Both preemption and reset occurred */
630 seq_printf(m, "Last both 0x%08x\n",
631 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
637 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
639 * Manually trigger a gpu reset at the next fence wait.
641 static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
643 struct drm_info_node *node = (struct drm_info_node *) m->private;
644 struct drm_device *dev = node->minor->dev;
645 struct amdgpu_device *adev = dev->dev_private;
647 seq_printf(m, "gpu recover\n");
648 amdgpu_device_gpu_recover(adev, NULL);
653 static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
654 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
655 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
658 static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
659 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
663 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
665 #if defined(CONFIG_DEBUG_FS)
666 if (amdgpu_sriov_vf(adev))
667 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1);
668 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);