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_alloc - allocate space on the ring buffer
54 * @adev: amdgpu_device pointer
55 * @ring: amdgpu_ring structure holding ring information
56 * @ndw: number of dwords to allocate in the ring buffer
58 * Allocate @ndw dwords in the ring buffer (all asics).
59 * Returns 0 on success, error on failure.
61 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw)
63 /* Align requested size with padding so unlock_commit can
65 ndw = (ndw + ring->align_mask) & ~ring->align_mask;
67 /* Make sure we aren't trying to allocate more space
68 * than the maximum for one submission
70 if (WARN_ON_ONCE(ndw > ring->max_dw))
74 ring->wptr_old = ring->wptr;
78 /** amdgpu_ring_insert_nop - insert NOP packets
80 * @ring: amdgpu_ring structure holding ring information
81 * @count: the number of NOP packets to insert
83 * This is the generic insert_nop function for rings except SDMA
85 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
89 for (i = 0; i < count; i++)
90 amdgpu_ring_write(ring, ring->nop);
93 /** amdgpu_ring_generic_pad_ib - pad IB with NOP packets
95 * @ring: amdgpu_ring structure holding ring information
96 * @ib: IB to add NOP packets to
98 * This is the generic pad_ib function for rings except SDMA
100 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
102 while (ib->length_dw & ring->align_mask)
103 ib->ptr[ib->length_dw++] = ring->nop;
107 * amdgpu_ring_commit - tell the GPU to execute the new
108 * commands on the ring buffer
110 * @adev: amdgpu_device pointer
111 * @ring: amdgpu_ring structure holding ring information
113 * Update the wptr (write pointer) to tell the GPU to
114 * execute new commands on the ring buffer (all asics).
116 void amdgpu_ring_commit(struct amdgpu_ring *ring)
120 /* We pad to match fetch size */
121 count = ring->align_mask + 1 - (ring->wptr & ring->align_mask);
122 count %= ring->align_mask + 1;
123 ring->funcs->insert_nop(ring, count);
126 amdgpu_ring_set_wptr(ring);
130 * amdgpu_ring_undo - reset the wptr
132 * @ring: amdgpu_ring structure holding ring information
134 * Reset the driver's copy of the wptr (all asics).
136 void amdgpu_ring_undo(struct amdgpu_ring *ring)
138 ring->wptr = ring->wptr_old;
142 * amdgpu_ring_backup - Back up the content of a ring
144 * @ring: the ring we want to back up
146 * Saves all unprocessed commits from a ring, returns the number of dwords saved.
148 unsigned amdgpu_ring_backup(struct amdgpu_ring *ring,
151 unsigned size, ptr, i;
155 if (ring->ring_obj == NULL)
158 /* it doesn't make sense to save anything if all fences are signaled */
159 if (!amdgpu_fence_count_emitted(ring))
162 ptr = le32_to_cpu(*ring->next_rptr_cpu_addr);
164 size = ring->wptr + (ring->ring_size / 4);
166 size &= ring->ptr_mask;
170 /* and then save the content of the ring */
171 *data = kmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
174 for (i = 0; i < size; ++i) {
175 (*data)[i] = ring->ring[ptr++];
176 ptr &= ring->ptr_mask;
183 * amdgpu_ring_restore - append saved commands to the ring again
185 * @ring: ring to append commands to
186 * @size: number of dwords we want to write
187 * @data: saved commands
189 * Allocates space on the ring and restore the previously saved commands.
191 int amdgpu_ring_restore(struct amdgpu_ring *ring,
192 unsigned size, uint32_t *data)
199 /* restore the saved ring content */
200 r = amdgpu_ring_alloc(ring, size);
204 for (i = 0; i < size; ++i) {
205 amdgpu_ring_write(ring, data[i]);
208 amdgpu_ring_commit(ring);
214 * amdgpu_ring_init - init driver ring struct.
216 * @adev: amdgpu_device pointer
217 * @ring: amdgpu_ring structure holding ring information
218 * @ring_size: size of the ring
219 * @nop: nop packet for this ring
221 * Initialize the driver information for the selected ring (all asics).
222 * Returns 0 on success, error on failure.
224 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
225 unsigned ring_size, u32 nop, u32 align_mask,
226 struct amdgpu_irq_src *irq_src, unsigned irq_type,
227 enum amdgpu_ring_type ring_type)
232 if (ring->adev == NULL) {
233 if (adev->num_rings >= AMDGPU_MAX_RINGS)
237 ring->idx = adev->num_rings++;
238 adev->rings[ring->idx] = ring;
239 r = amdgpu_fence_driver_init_ring(ring);
244 r = amdgpu_wb_get(adev, &ring->rptr_offs);
246 dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
250 r = amdgpu_wb_get(adev, &ring->wptr_offs);
252 dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
256 r = amdgpu_wb_get(adev, &ring->fence_offs);
258 dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
262 r = amdgpu_wb_get(adev, &ring->next_rptr_offs);
264 dev_err(adev->dev, "(%d) ring next_rptr wb alloc failed\n", r);
267 ring->next_rptr_gpu_addr = adev->wb.gpu_addr + (ring->next_rptr_offs * 4);
268 ring->next_rptr_cpu_addr = &adev->wb.wb[ring->next_rptr_offs];
269 spin_lock_init(&ring->fence_lock);
270 r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
272 dev_err(adev->dev, "failed initializing fences (%d).\n", r);
276 /* Align ring size */
277 rb_bufsz = order_base_2(ring_size / 8);
278 ring_size = (1 << (rb_bufsz + 1)) * 4;
279 ring->ring_size = ring_size;
280 ring->align_mask = align_mask;
282 ring->type = ring_type;
284 /* Allocate ring buffer */
285 if (ring->ring_obj == NULL) {
286 r = amdgpu_bo_create(adev, ring->ring_size, PAGE_SIZE, true,
287 AMDGPU_GEM_DOMAIN_GTT, 0,
288 NULL, NULL, &ring->ring_obj);
290 dev_err(adev->dev, "(%d) ring create failed\n", r);
293 r = amdgpu_bo_reserve(ring->ring_obj, false);
294 if (unlikely(r != 0))
296 r = amdgpu_bo_pin(ring->ring_obj, AMDGPU_GEM_DOMAIN_GTT,
299 amdgpu_bo_unreserve(ring->ring_obj);
300 dev_err(adev->dev, "(%d) ring pin failed\n", r);
303 r = amdgpu_bo_kmap(ring->ring_obj,
304 (void **)&ring->ring);
305 amdgpu_bo_unreserve(ring->ring_obj);
307 dev_err(adev->dev, "(%d) ring map failed\n", r);
311 ring->ptr_mask = (ring->ring_size / 4) - 1;
312 ring->max_dw = DIV_ROUND_UP(ring->ring_size / 4,
313 amdgpu_sched_hw_submission);
315 if (amdgpu_debugfs_ring_init(adev, ring)) {
316 DRM_ERROR("Failed to register debugfs file for rings !\n");
322 * amdgpu_ring_fini - tear down the driver ring struct.
324 * @adev: amdgpu_device pointer
325 * @ring: amdgpu_ring structure holding ring information
327 * Tear down the driver information for the selected ring (all asics).
329 void amdgpu_ring_fini(struct amdgpu_ring *ring)
332 struct amdgpu_bo *ring_obj;
334 ring_obj = ring->ring_obj;
337 ring->ring_obj = NULL;
339 amdgpu_wb_free(ring->adev, ring->fence_offs);
340 amdgpu_wb_free(ring->adev, ring->rptr_offs);
341 amdgpu_wb_free(ring->adev, ring->wptr_offs);
342 amdgpu_wb_free(ring->adev, ring->next_rptr_offs);
345 r = amdgpu_bo_reserve(ring_obj, false);
346 if (likely(r == 0)) {
347 amdgpu_bo_kunmap(ring_obj);
348 amdgpu_bo_unpin(ring_obj);
349 amdgpu_bo_unreserve(ring_obj);
351 amdgpu_bo_unref(&ring_obj);
356 * amdgpu_ring_from_fence - get ring from fence
358 * @f: fence structure
360 * Extract the ring a fence belongs to. Handles both scheduler as
361 * well as hardware fences.
363 struct amdgpu_ring *amdgpu_ring_from_fence(struct fence *f)
365 struct amdgpu_fence *a_fence;
366 struct amd_sched_fence *s_fence;
368 s_fence = to_amd_sched_fence(f);
370 return container_of(s_fence->sched, struct amdgpu_ring, sched);
372 a_fence = to_amdgpu_fence(f);
374 return a_fence->ring;
382 #if defined(CONFIG_DEBUG_FS)
384 static int amdgpu_debugfs_ring_info(struct seq_file *m, void *data)
386 struct drm_info_node *node = (struct drm_info_node *) m->private;
387 struct drm_device *dev = node->minor->dev;
388 struct amdgpu_device *adev = dev->dev_private;
389 int roffset = *(int*)node->info_ent->data;
390 struct amdgpu_ring *ring = (void *)(((uint8_t*)adev) + roffset);
392 uint32_t rptr, wptr, rptr_next;
395 wptr = amdgpu_ring_get_wptr(ring);
396 seq_printf(m, "wptr: 0x%08x [%5d]\n", wptr, wptr);
398 rptr = amdgpu_ring_get_rptr(ring);
399 rptr_next = le32_to_cpu(*ring->next_rptr_cpu_addr);
401 seq_printf(m, "rptr: 0x%08x [%5d]\n", rptr, rptr);
403 seq_printf(m, "driver's copy of the wptr: 0x%08x [%5d]\n",
404 ring->wptr, ring->wptr);
409 /* print 8 dw before current rptr as often it's the last executed
410 * packet that is the root issue
412 i = (rptr + ring->ptr_mask + 1 - 32) & ring->ptr_mask;
414 seq_printf(m, "r[%5d]=0x%08x", i, ring->ring[i]);
420 i = (i + 1) & ring->ptr_mask;
423 seq_printf(m, "r[%5d]=0x%08x", i, ring->ring[i]);
429 i = (i + 1) & ring->ptr_mask;
434 /* TODO: clean this up !*/
435 static int amdgpu_gfx_index = offsetof(struct amdgpu_device, gfx.gfx_ring[0]);
436 static int cayman_cp1_index = offsetof(struct amdgpu_device, gfx.compute_ring[0]);
437 static int cayman_cp2_index = offsetof(struct amdgpu_device, gfx.compute_ring[1]);
438 static int amdgpu_dma1_index = offsetof(struct amdgpu_device, sdma.instance[0].ring);
439 static int amdgpu_dma2_index = offsetof(struct amdgpu_device, sdma.instance[1].ring);
440 static int r600_uvd_index = offsetof(struct amdgpu_device, uvd.ring);
441 static int si_vce1_index = offsetof(struct amdgpu_device, vce.ring[0]);
442 static int si_vce2_index = offsetof(struct amdgpu_device, vce.ring[1]);
444 static struct drm_info_list amdgpu_debugfs_ring_info_list[] = {
445 {"amdgpu_ring_gfx", amdgpu_debugfs_ring_info, 0, &amdgpu_gfx_index},
446 {"amdgpu_ring_cp1", amdgpu_debugfs_ring_info, 0, &cayman_cp1_index},
447 {"amdgpu_ring_cp2", amdgpu_debugfs_ring_info, 0, &cayman_cp2_index},
448 {"amdgpu_ring_dma1", amdgpu_debugfs_ring_info, 0, &amdgpu_dma1_index},
449 {"amdgpu_ring_dma2", amdgpu_debugfs_ring_info, 0, &amdgpu_dma2_index},
450 {"amdgpu_ring_uvd", amdgpu_debugfs_ring_info, 0, &r600_uvd_index},
451 {"amdgpu_ring_vce1", amdgpu_debugfs_ring_info, 0, &si_vce1_index},
452 {"amdgpu_ring_vce2", amdgpu_debugfs_ring_info, 0, &si_vce2_index},
457 static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring)
459 #if defined(CONFIG_DEBUG_FS)
461 for (i = 0; i < ARRAY_SIZE(amdgpu_debugfs_ring_info_list); ++i) {
462 struct drm_info_list *info = &amdgpu_debugfs_ring_info_list[i];
463 int roffset = *(int*)amdgpu_debugfs_ring_info_list[i].data;
464 struct amdgpu_ring *other = (void *)(((uint8_t*)adev) + roffset);
470 r = amdgpu_debugfs_add_files(adev, info, 1);