2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
24 * Authors: Dave Airlie
29 #include <linux/seq_file.h>
30 #include <linux/slab.h>
32 #include <drm/amdgpu_drm.h>
38 * Most engines on the GPU are fed via ring buffers. Ring
39 * buffers are areas of GPU accessible memory that the host
40 * writes commands into and the GPU reads commands out of.
41 * There is a rptr (read pointer) that determines where the
42 * GPU is currently reading, and a wptr (write pointer)
43 * which determines where the host has written. When the
44 * pointers are equal, the ring is idle. When the host
45 * writes commands to the ring buffer, it increments the
46 * wptr. The GPU then starts fetching commands and executes
47 * them until the pointers are equal again.
49 static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring);
52 * amdgpu_ring_free_size - update the free size
54 * @adev: amdgpu_device pointer
55 * @ring: amdgpu_ring structure holding ring information
57 * Update the free dw slots in the ring buffer (all asics).
59 void amdgpu_ring_free_size(struct amdgpu_ring *ring)
61 uint32_t rptr = amdgpu_ring_get_rptr(ring);
63 /* This works because ring_size is a power of 2 */
64 ring->ring_free_dw = rptr + (ring->ring_size / 4);
65 ring->ring_free_dw -= ring->wptr;
66 ring->ring_free_dw &= ring->ptr_mask;
67 if (!ring->ring_free_dw) {
68 /* this is an empty ring */
69 ring->ring_free_dw = ring->ring_size / 4;
70 /* update lockup info to avoid false positive */
71 amdgpu_ring_lockup_update(ring);
76 * amdgpu_ring_alloc - allocate space on the ring buffer
78 * @adev: amdgpu_device pointer
79 * @ring: amdgpu_ring structure holding ring information
80 * @ndw: number of dwords to allocate in the ring buffer
82 * Allocate @ndw dwords in the ring buffer (all asics).
83 * Returns 0 on success, error on failure.
85 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw)
89 /* make sure we aren't trying to allocate more space than there is on the ring */
90 if (ndw > (ring->ring_size / 4))
92 /* Align requested size with padding so unlock_commit can
94 amdgpu_ring_free_size(ring);
95 ndw = (ndw + ring->align_mask) & ~ring->align_mask;
96 while (ndw > (ring->ring_free_dw - 1)) {
97 amdgpu_ring_free_size(ring);
98 if (ndw < ring->ring_free_dw) {
101 r = amdgpu_fence_wait_next(ring);
105 ring->count_dw = ndw;
106 ring->wptr_old = ring->wptr;
111 * amdgpu_ring_lock - lock the ring and allocate space on it
113 * @adev: amdgpu_device pointer
114 * @ring: amdgpu_ring structure holding ring information
115 * @ndw: number of dwords to allocate in the ring buffer
117 * Lock the ring and allocate @ndw dwords in the ring buffer
119 * Returns 0 on success, error on failure.
121 int amdgpu_ring_lock(struct amdgpu_ring *ring, unsigned ndw)
125 mutex_lock(ring->ring_lock);
126 r = amdgpu_ring_alloc(ring, ndw);
128 mutex_unlock(ring->ring_lock);
135 * amdgpu_ring_commit - tell the GPU to execute the new
136 * commands on the ring buffer
138 * @adev: amdgpu_device pointer
139 * @ring: amdgpu_ring structure holding ring information
141 * Update the wptr (write pointer) to tell the GPU to
142 * execute new commands on the ring buffer (all asics).
144 void amdgpu_ring_commit(struct amdgpu_ring *ring)
146 /* We pad to match fetch size */
147 while (ring->wptr & ring->align_mask) {
148 amdgpu_ring_write(ring, ring->nop);
151 amdgpu_ring_set_wptr(ring);
155 * amdgpu_ring_unlock_commit - tell the GPU to execute the new
156 * commands on the ring buffer and unlock it
158 * @ring: amdgpu_ring structure holding ring information
160 * Call amdgpu_ring_commit() then unlock the ring (all asics).
162 void amdgpu_ring_unlock_commit(struct amdgpu_ring *ring)
164 amdgpu_ring_commit(ring);
165 mutex_unlock(ring->ring_lock);
169 * amdgpu_ring_undo - reset the wptr
171 * @ring: amdgpu_ring structure holding ring information
173 * Reset the driver's copy of the wptr (all asics).
175 void amdgpu_ring_undo(struct amdgpu_ring *ring)
177 ring->wptr = ring->wptr_old;
181 * amdgpu_ring_unlock_undo - reset the wptr and unlock the ring
183 * @ring: amdgpu_ring structure holding ring information
185 * Call amdgpu_ring_undo() then unlock the ring (all asics).
187 void amdgpu_ring_unlock_undo(struct amdgpu_ring *ring)
189 amdgpu_ring_undo(ring);
190 mutex_unlock(ring->ring_lock);
194 * amdgpu_ring_lockup_update - update lockup variables
196 * @ring: amdgpu_ring structure holding ring information
198 * Update the last rptr value and timestamp (all asics).
200 void amdgpu_ring_lockup_update(struct amdgpu_ring *ring)
202 atomic_set(&ring->last_rptr, amdgpu_ring_get_rptr(ring));
203 atomic64_set(&ring->last_activity, jiffies_64);
207 * amdgpu_ring_test_lockup() - check if ring is lockedup by recording information
208 * @ring: amdgpu_ring structure holding ring information
211 bool amdgpu_ring_test_lockup(struct amdgpu_ring *ring)
213 uint32_t rptr = amdgpu_ring_get_rptr(ring);
214 uint64_t last = atomic64_read(&ring->last_activity);
217 if (rptr != atomic_read(&ring->last_rptr)) {
218 /* ring is still working, no lockup */
219 amdgpu_ring_lockup_update(ring);
223 elapsed = jiffies_to_msecs(jiffies_64 - last);
224 if (amdgpu_lockup_timeout && elapsed >= amdgpu_lockup_timeout) {
225 dev_err(ring->adev->dev, "ring %d stalled for more than %llumsec\n",
229 /* give a chance to the GPU ... */
234 * amdgpu_ring_backup - Back up the content of a ring
236 * @ring: the ring we want to back up
238 * Saves all unprocessed commits from a ring, returns the number of dwords saved.
240 unsigned amdgpu_ring_backup(struct amdgpu_ring *ring,
243 unsigned size, ptr, i;
245 /* just in case lock the ring */
246 mutex_lock(ring->ring_lock);
249 if (ring->ring_obj == NULL) {
250 mutex_unlock(ring->ring_lock);
254 /* it doesn't make sense to save anything if all fences are signaled */
255 if (!amdgpu_fence_count_emitted(ring)) {
256 mutex_unlock(ring->ring_lock);
260 ptr = le32_to_cpu(*ring->next_rptr_cpu_addr);
262 size = ring->wptr + (ring->ring_size / 4);
264 size &= ring->ptr_mask;
266 mutex_unlock(ring->ring_lock);
270 /* and then save the content of the ring */
271 *data = kmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
273 mutex_unlock(ring->ring_lock);
276 for (i = 0; i < size; ++i) {
277 (*data)[i] = ring->ring[ptr++];
278 ptr &= ring->ptr_mask;
281 mutex_unlock(ring->ring_lock);
286 * amdgpu_ring_restore - append saved commands to the ring again
288 * @ring: ring to append commands to
289 * @size: number of dwords we want to write
290 * @data: saved commands
292 * Allocates space on the ring and restore the previously saved commands.
294 int amdgpu_ring_restore(struct amdgpu_ring *ring,
295 unsigned size, uint32_t *data)
302 /* restore the saved ring content */
303 r = amdgpu_ring_lock(ring, size);
307 for (i = 0; i < size; ++i) {
308 amdgpu_ring_write(ring, data[i]);
311 amdgpu_ring_unlock_commit(ring);
317 * amdgpu_ring_init - init driver ring struct.
319 * @adev: amdgpu_device pointer
320 * @ring: amdgpu_ring structure holding ring information
321 * @ring_size: size of the ring
322 * @nop: nop packet for this ring
324 * Initialize the driver information for the selected ring (all asics).
325 * Returns 0 on success, error on failure.
327 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
328 unsigned ring_size, u32 nop, u32 align_mask,
329 struct amdgpu_irq_src *irq_src, unsigned irq_type,
330 enum amdgpu_ring_type ring_type)
335 if (ring->adev == NULL) {
336 if (adev->num_rings >= AMDGPU_MAX_RINGS)
340 ring->idx = adev->num_rings++;
341 adev->rings[ring->idx] = ring;
342 amdgpu_fence_driver_init_ring(ring);
345 init_waitqueue_head(&ring->fence_drv.fence_queue);
347 r = amdgpu_wb_get(adev, &ring->rptr_offs);
349 dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
353 r = amdgpu_wb_get(adev, &ring->wptr_offs);
355 dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
359 r = amdgpu_wb_get(adev, &ring->fence_offs);
361 dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
365 r = amdgpu_wb_get(adev, &ring->next_rptr_offs);
367 dev_err(adev->dev, "(%d) ring next_rptr wb alloc failed\n", r);
370 ring->next_rptr_gpu_addr = adev->wb.gpu_addr + (ring->next_rptr_offs * 4);
371 ring->next_rptr_cpu_addr = &adev->wb.wb[ring->next_rptr_offs];
372 spin_lock_init(&ring->fence_lock);
373 r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
375 dev_err(adev->dev, "failed initializing fences (%d).\n", r);
379 ring->ring_lock = &adev->ring_lock;
380 /* Align ring size */
381 rb_bufsz = order_base_2(ring_size / 8);
382 ring_size = (1 << (rb_bufsz + 1)) * 4;
383 ring->ring_size = ring_size;
384 ring->align_mask = align_mask;
386 ring->type = ring_type;
388 /* Allocate ring buffer */
389 if (ring->ring_obj == NULL) {
390 r = amdgpu_bo_create(adev, ring->ring_size, PAGE_SIZE, true,
391 AMDGPU_GEM_DOMAIN_GTT, 0,
392 NULL, &ring->ring_obj);
394 dev_err(adev->dev, "(%d) ring create failed\n", r);
397 r = amdgpu_bo_reserve(ring->ring_obj, false);
398 if (unlikely(r != 0))
400 r = amdgpu_bo_pin(ring->ring_obj, AMDGPU_GEM_DOMAIN_GTT,
403 amdgpu_bo_unreserve(ring->ring_obj);
404 dev_err(adev->dev, "(%d) ring pin failed\n", r);
407 r = amdgpu_bo_kmap(ring->ring_obj,
408 (void **)&ring->ring);
409 amdgpu_bo_unreserve(ring->ring_obj);
411 dev_err(adev->dev, "(%d) ring map failed\n", r);
415 ring->ptr_mask = (ring->ring_size / 4) - 1;
416 ring->ring_free_dw = ring->ring_size / 4;
418 if (amdgpu_debugfs_ring_init(adev, ring)) {
419 DRM_ERROR("Failed to register debugfs file for rings !\n");
421 amdgpu_ring_lockup_update(ring);
426 * amdgpu_ring_fini - tear down the driver ring struct.
428 * @adev: amdgpu_device pointer
429 * @ring: amdgpu_ring structure holding ring information
431 * Tear down the driver information for the selected ring (all asics).
433 void amdgpu_ring_fini(struct amdgpu_ring *ring)
436 struct amdgpu_bo *ring_obj;
438 if (ring->ring_lock == NULL)
441 mutex_lock(ring->ring_lock);
442 ring_obj = ring->ring_obj;
445 ring->ring_obj = NULL;
446 mutex_unlock(ring->ring_lock);
448 amdgpu_wb_free(ring->adev, ring->fence_offs);
449 amdgpu_wb_free(ring->adev, ring->rptr_offs);
450 amdgpu_wb_free(ring->adev, ring->wptr_offs);
451 amdgpu_wb_free(ring->adev, ring->next_rptr_offs);
454 r = amdgpu_bo_reserve(ring_obj, false);
455 if (likely(r == 0)) {
456 amdgpu_bo_kunmap(ring_obj);
457 amdgpu_bo_unpin(ring_obj);
458 amdgpu_bo_unreserve(ring_obj);
460 amdgpu_bo_unref(&ring_obj);
467 #if defined(CONFIG_DEBUG_FS)
469 static int amdgpu_debugfs_ring_info(struct seq_file *m, void *data)
471 struct drm_info_node *node = (struct drm_info_node *) m->private;
472 struct drm_device *dev = node->minor->dev;
473 struct amdgpu_device *adev = dev->dev_private;
474 int roffset = *(int*)node->info_ent->data;
475 struct amdgpu_ring *ring = (void *)(((uint8_t*)adev) + roffset);
477 uint32_t rptr, wptr, rptr_next;
478 unsigned count, i, j;
480 amdgpu_ring_free_size(ring);
481 count = (ring->ring_size / 4) - ring->ring_free_dw;
483 wptr = amdgpu_ring_get_wptr(ring);
484 seq_printf(m, "wptr: 0x%08x [%5d]\n",
487 rptr = amdgpu_ring_get_rptr(ring);
488 seq_printf(m, "rptr: 0x%08x [%5d]\n",
493 seq_printf(m, "driver's copy of the wptr: 0x%08x [%5d]\n",
494 ring->wptr, ring->wptr);
495 seq_printf(m, "last semaphore signal addr : 0x%016llx\n",
496 ring->last_semaphore_signal_addr);
497 seq_printf(m, "last semaphore wait addr : 0x%016llx\n",
498 ring->last_semaphore_wait_addr);
499 seq_printf(m, "%u free dwords in ring\n", ring->ring_free_dw);
500 seq_printf(m, "%u dwords in ring\n", count);
505 /* print 8 dw before current rptr as often it's the last executed
506 * packet that is the root issue
508 i = (rptr + ring->ptr_mask + 1 - 32) & ring->ptr_mask;
509 for (j = 0; j <= (count + 32); j++) {
510 seq_printf(m, "r[%5d]=0x%08x", i, ring->ring[i]);
516 i = (i + 1) & ring->ptr_mask;
521 /* TODO: clean this up !*/
522 static int amdgpu_gfx_index = offsetof(struct amdgpu_device, gfx.gfx_ring[0]);
523 static int cayman_cp1_index = offsetof(struct amdgpu_device, gfx.compute_ring[0]);
524 static int cayman_cp2_index = offsetof(struct amdgpu_device, gfx.compute_ring[1]);
525 static int amdgpu_dma1_index = offsetof(struct amdgpu_device, sdma[0].ring);
526 static int amdgpu_dma2_index = offsetof(struct amdgpu_device, sdma[1].ring);
527 static int r600_uvd_index = offsetof(struct amdgpu_device, uvd.ring);
528 static int si_vce1_index = offsetof(struct amdgpu_device, vce.ring[0]);
529 static int si_vce2_index = offsetof(struct amdgpu_device, vce.ring[1]);
531 static struct drm_info_list amdgpu_debugfs_ring_info_list[] = {
532 {"amdgpu_ring_gfx", amdgpu_debugfs_ring_info, 0, &amdgpu_gfx_index},
533 {"amdgpu_ring_cp1", amdgpu_debugfs_ring_info, 0, &cayman_cp1_index},
534 {"amdgpu_ring_cp2", amdgpu_debugfs_ring_info, 0, &cayman_cp2_index},
535 {"amdgpu_ring_dma1", amdgpu_debugfs_ring_info, 0, &amdgpu_dma1_index},
536 {"amdgpu_ring_dma2", amdgpu_debugfs_ring_info, 0, &amdgpu_dma2_index},
537 {"amdgpu_ring_uvd", amdgpu_debugfs_ring_info, 0, &r600_uvd_index},
538 {"amdgpu_ring_vce1", amdgpu_debugfs_ring_info, 0, &si_vce1_index},
539 {"amdgpu_ring_vce2", amdgpu_debugfs_ring_info, 0, &si_vce2_index},
544 static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring)
546 #if defined(CONFIG_DEBUG_FS)
548 for (i = 0; i < ARRAY_SIZE(amdgpu_debugfs_ring_info_list); ++i) {
549 struct drm_info_list *info = &amdgpu_debugfs_ring_info_list[i];
550 int roffset = *(int*)amdgpu_debugfs_ring_info_list[i].data;
551 struct amdgpu_ring *other = (void *)(((uint8_t*)adev) + roffset);
557 r = amdgpu_debugfs_add_files(adev, info, 1);