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 * IBs (Indirect Buffers) and areas of GPU accessible memory where
39 * commands are stored. You can put a pointer to the IB in the
40 * command ring and the hw will fetch the commands from the IB
41 * and execute them. Generally userspace acceleration drivers
42 * produce command buffers which are send to the kernel and
43 * put in IBs for execution by the requested ring.
45 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev);
48 * amdgpu_ib_get - request an IB (Indirect Buffer)
50 * @ring: ring index the IB is associated with
51 * @size: requested IB size
52 * @ib: IB object returned
54 * Request an IB (all asics). IBs are allocated using the
56 * Returns 0 on success, error on failure.
58 int amdgpu_ib_get(struct amdgpu_ring *ring, struct amdgpu_vm *vm,
59 unsigned size, struct amdgpu_ib *ib)
61 struct amdgpu_device *adev = ring->adev;
65 r = amdgpu_sa_bo_new(adev, &adev->ring_tmp_bo,
66 &ib->sa_bo, size, 256);
68 dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
72 ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
75 ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
85 amdgpu_sync_create(&ib->sync);
103 * amdgpu_ib_free - free an IB (Indirect Buffer)
105 * @adev: amdgpu_device pointer
106 * @ib: IB object to free
108 * Free an IB (all asics).
110 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib)
112 amdgpu_sync_free(adev, &ib->sync, ib->fence);
113 amdgpu_sa_bo_free(adev, &ib->sa_bo, ib->fence);
114 amdgpu_fence_unref(&ib->fence);
118 * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
120 * @adev: amdgpu_device pointer
121 * @num_ibs: number of IBs to schedule
122 * @ibs: IB objects to schedule
123 * @owner: owner for creating the fences
125 * Schedule an IB on the associated ring (all asics).
126 * Returns 0 on success, error on failure.
128 * On SI, there are two parallel engines fed from the primary ring,
129 * the CE (Constant Engine) and the DE (Drawing Engine). Since
130 * resource descriptors have moved to memory, the CE allows you to
131 * prime the caches while the DE is updating register state so that
132 * the resource descriptors will be already in cache when the draw is
133 * processed. To accomplish this, the userspace driver submits two
134 * IBs, one for the CE and one for the DE. If there is a CE IB (called
135 * a CONST_IB), it will be put on the ring prior to the DE IB. Prior
136 * to SI there was just a DE IB.
138 int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
139 struct amdgpu_ib *ibs, void *owner)
141 struct amdgpu_ib *ib = &ibs[0];
142 struct amdgpu_ring *ring;
143 struct amdgpu_ctx *ctx, *old_ctx;
144 struct amdgpu_vm *vm;
156 dev_err(adev->dev, "couldn't schedule ib\n");
160 r = amdgpu_ring_lock(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs);
162 dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
167 /* grab a vm id if necessary */
168 struct amdgpu_fence *vm_id_fence = NULL;
169 vm_id_fence = amdgpu_vm_grab_id(ibs->ring, ibs->vm);
170 amdgpu_sync_fence(&ibs->sync, vm_id_fence);
173 r = amdgpu_sync_rings(&ibs->sync, ring);
175 amdgpu_ring_unlock_undo(ring);
176 dev_err(adev->dev, "failed to sync rings (%d)\n", r);
181 /* do context switch */
182 amdgpu_vm_flush(ring, vm, ib->sync.last_vm_update);
184 if (ring->funcs->emit_gds_switch)
185 amdgpu_ring_emit_gds_switch(ring, ib->vm->ids[ring->idx].id,
186 ib->gds_base, ib->gds_size,
187 ib->gws_base, ib->gws_size,
188 ib->oa_base, ib->oa_size);
190 if (ring->funcs->emit_hdp_flush)
191 amdgpu_ring_emit_hdp_flush(ring);
194 old_ctx = ring->current_ctx;
195 for (i = 0; i < num_ibs; ++i) {
198 if (ib->ring != ring || ib->ctx != ctx || ib->vm != vm) {
199 ring->current_ctx = old_ctx;
200 amdgpu_ring_unlock_undo(ring);
203 amdgpu_ring_emit_ib(ring, ib);
204 ring->current_ctx = ctx;
207 r = amdgpu_fence_emit(ring, owner, &ib->fence);
209 dev_err(adev->dev, "failed to emit fence (%d)\n", r);
210 ring->current_ctx = old_ctx;
211 amdgpu_ring_unlock_undo(ring);
215 /* wrap the last IB with fence */
217 uint64_t addr = amdgpu_bo_gpu_offset(ib->user->bo);
218 addr += ib->user->offset;
219 amdgpu_ring_emit_fence(ring, addr, ib->fence->seq,
220 AMDGPU_FENCE_FLAG_64BIT);
224 amdgpu_vm_fence(adev, ib->vm, ib->fence);
226 amdgpu_ring_unlock_commit(ring);
231 * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
233 * @adev: amdgpu_device pointer
235 * Initialize the suballocator to manage a pool of memory
236 * for use as IBs (all asics).
237 * Returns 0 on success, error on failure.
239 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
243 if (adev->ib_pool_ready) {
246 r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
247 AMDGPU_IB_POOL_SIZE*64*1024,
248 AMDGPU_GPU_PAGE_SIZE,
249 AMDGPU_GEM_DOMAIN_GTT);
254 r = amdgpu_sa_bo_manager_start(adev, &adev->ring_tmp_bo);
259 adev->ib_pool_ready = true;
260 if (amdgpu_debugfs_sa_init(adev)) {
261 dev_err(adev->dev, "failed to register debugfs file for SA\n");
267 * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
269 * @adev: amdgpu_device pointer
271 * Tear down the suballocator managing the pool of memory
272 * for use as IBs (all asics).
274 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
276 if (adev->ib_pool_ready) {
277 amdgpu_sa_bo_manager_suspend(adev, &adev->ring_tmp_bo);
278 amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
279 adev->ib_pool_ready = false;
284 * amdgpu_ib_ring_tests - test IBs on the rings
286 * @adev: amdgpu_device pointer
288 * Test an IB (Indirect Buffer) on each ring.
289 * If the test fails, disable the ring.
290 * Returns 0 on success, error if the primary GFX ring
293 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
298 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
299 struct amdgpu_ring *ring = adev->rings[i];
301 if (!ring || !ring->ready)
304 r = amdgpu_ring_test_ib(ring);
307 adev->needs_reset = false;
309 if (ring == &adev->gfx.gfx_ring[0]) {
310 /* oh, oh, that's really bad */
311 DRM_ERROR("amdgpu: failed testing IB on GFX ring (%d).\n", r);
312 adev->accel_working = false;
316 /* still not good, but we can live with it */
317 DRM_ERROR("amdgpu: failed testing IB on ring %d (%d).\n", i, r);
327 #if defined(CONFIG_DEBUG_FS)
329 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
331 struct drm_info_node *node = (struct drm_info_node *) m->private;
332 struct drm_device *dev = node->minor->dev;
333 struct amdgpu_device *adev = dev->dev_private;
335 amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
341 static struct drm_info_list amdgpu_debugfs_sa_list[] = {
342 {"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
347 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
349 #if defined(CONFIG_DEBUG_FS)
350 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);