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>
33 #include <drm/drm_debugfs.h>
37 #include "amdgpu_trace.h"
39 #define AMDGPU_IB_TEST_TIMEOUT msecs_to_jiffies(1000)
40 #define AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT msecs_to_jiffies(2000)
44 * IBs (Indirect Buffers) and areas of GPU accessible memory where
45 * commands are stored. You can put a pointer to the IB in the
46 * command ring and the hw will fetch the commands from the IB
47 * and execute them. Generally userspace acceleration drivers
48 * produce command buffers which are send to the kernel and
49 * put in IBs for execution by the requested ring.
53 * amdgpu_ib_get - request an IB (Indirect Buffer)
55 * @ring: ring index the IB is associated with
56 * @size: requested IB size
57 * @ib: IB object returned
59 * Request an IB (all asics). IBs are allocated using the
61 * Returns 0 on success, error on failure.
63 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
64 unsigned size, enum amdgpu_ib_pool_type pool_type,
70 r = amdgpu_sa_bo_new(&adev->ib_pools[pool_type],
71 &ib->sa_bo, size, 256);
73 dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
77 ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
80 ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
87 * amdgpu_ib_free - free an IB (Indirect Buffer)
89 * @adev: amdgpu_device pointer
90 * @ib: IB object to free
91 * @f: the fence SA bo need wait on for the ib alloation
93 * Free an IB (all asics).
95 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
98 amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
102 * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
104 * @adev: amdgpu_device pointer
105 * @num_ibs: number of IBs to schedule
106 * @ibs: IB objects to schedule
107 * @f: fence created during this submission
109 * Schedule an IB on the associated ring (all asics).
110 * Returns 0 on success, error on failure.
112 * On SI, there are two parallel engines fed from the primary ring,
113 * the CE (Constant Engine) and the DE (Drawing Engine). Since
114 * resource descriptors have moved to memory, the CE allows you to
115 * prime the caches while the DE is updating register state so that
116 * the resource descriptors will be already in cache when the draw is
117 * processed. To accomplish this, the userspace driver submits two
118 * IBs, one for the CE and one for the DE. If there is a CE IB (called
119 * a CONST_IB), it will be put on the ring prior to the DE IB. Prior
120 * to SI there was just a DE IB.
122 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
123 struct amdgpu_ib *ibs, struct amdgpu_job *job,
124 struct dma_fence **f)
126 struct amdgpu_device *adev = ring->adev;
127 struct amdgpu_ib *ib = &ibs[0];
128 struct dma_fence *tmp = NULL;
129 bool skip_preamble, need_ctx_switch;
130 unsigned patch_offset = ~0;
131 struct amdgpu_vm *vm;
133 uint32_t status = 0, alloc_size;
134 unsigned fence_flags = 0;
139 bool need_pipe_sync = false;
144 /* ring tests don't use a job */
147 fence_ctx = job->base.s_fence ?
148 job->base.s_fence->scheduled.context : 0;
154 if (!ring->sched.ready) {
155 dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
159 if (vm && !job->vmid) {
160 dev_err(adev->dev, "VM IB without ID\n");
164 if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
165 (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)) {
166 dev_err(adev->dev, "secure submissions not supported on compute rings\n");
170 alloc_size = ring->funcs->emit_frame_size + num_ibs *
171 ring->funcs->emit_ib_size;
173 r = amdgpu_ring_alloc(ring, alloc_size);
175 dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
179 need_ctx_switch = ring->current_ctx != fence_ctx;
180 if (ring->funcs->emit_pipeline_sync && job &&
181 ((tmp = amdgpu_sync_get_fence(&job->sched_sync)) ||
182 (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
183 amdgpu_vm_need_pipeline_sync(ring, job))) {
184 need_pipe_sync = true;
187 trace_amdgpu_ib_pipe_sync(job, tmp);
192 if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
193 ring->funcs->emit_mem_sync(ring);
195 if (ring->funcs->insert_start)
196 ring->funcs->insert_start(ring);
199 r = amdgpu_vm_flush(ring, job, need_pipe_sync);
201 amdgpu_ring_undo(ring);
206 if (job && ring->funcs->init_cond_exec)
207 patch_offset = amdgpu_ring_init_cond_exec(ring);
210 if (!(adev->flags & AMD_IS_APU))
213 if (ring->funcs->emit_hdp_flush)
214 amdgpu_ring_emit_hdp_flush(ring);
216 amdgpu_asic_flush_hdp(adev, ring);
220 status |= AMDGPU_HAVE_CTX_SWITCH;
222 skip_preamble = ring->current_ctx == fence_ctx;
223 if (job && ring->funcs->emit_cntxcntl) {
224 status |= job->preamble_status;
225 status |= job->preemption_status;
226 amdgpu_ring_emit_cntxcntl(ring, status);
229 /* Setup initial TMZiness and send it off.
232 if (job && ring->funcs->emit_frame_cntl) {
233 secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
234 amdgpu_ring_emit_frame_cntl(ring, true, secure);
237 for (i = 0; i < num_ibs; ++i) {
240 /* drop preamble IBs if we don't have a context switch */
241 if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
243 !(status & AMDGPU_PREAMBLE_IB_PRESENT_FIRST) &&
245 !amdgpu_sriov_vf(adev)) /* for SRIOV preemption, Preamble CE ib must be inserted anyway */
248 if (job && ring->funcs->emit_frame_cntl) {
249 if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) {
250 amdgpu_ring_emit_frame_cntl(ring, false, secure);
252 amdgpu_ring_emit_frame_cntl(ring, true, secure);
256 amdgpu_ring_emit_ib(ring, job, ib, status);
257 status &= ~AMDGPU_HAVE_CTX_SWITCH;
260 if (job && ring->funcs->emit_frame_cntl)
261 amdgpu_ring_emit_frame_cntl(ring, false, secure);
264 if (!(adev->flags & AMD_IS_APU))
266 amdgpu_asic_invalidate_hdp(adev, ring);
268 if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
269 fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
271 /* wrap the last IB with fence */
272 if (job && job->uf_addr) {
273 amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
274 fence_flags | AMDGPU_FENCE_FLAG_64BIT);
277 r = amdgpu_fence_emit(ring, f, fence_flags);
279 dev_err(adev->dev, "failed to emit fence (%d)\n", r);
280 if (job && job->vmid)
281 amdgpu_vmid_reset(adev, ring->funcs->vmhub, job->vmid);
282 amdgpu_ring_undo(ring);
286 if (ring->funcs->insert_end)
287 ring->funcs->insert_end(ring);
289 if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
290 amdgpu_ring_patch_cond_exec(ring, patch_offset);
292 ring->current_ctx = fence_ctx;
293 if (vm && ring->funcs->emit_switch_buffer)
294 amdgpu_ring_emit_switch_buffer(ring);
295 amdgpu_ring_commit(ring);
300 * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
302 * @adev: amdgpu_device pointer
304 * Initialize the suballocator to manage a pool of memory
305 * for use as IBs (all asics).
306 * Returns 0 on success, error on failure.
308 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
313 if (adev->ib_pool_ready)
316 for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
317 if (i == AMDGPU_IB_POOL_DIRECT)
318 size = PAGE_SIZE * 2;
320 size = AMDGPU_IB_POOL_SIZE;
322 r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
323 size, AMDGPU_GPU_PAGE_SIZE,
324 AMDGPU_GEM_DOMAIN_GTT);
328 adev->ib_pool_ready = true;
334 amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
339 * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
341 * @adev: amdgpu_device pointer
343 * Tear down the suballocator managing the pool of memory
344 * for use as IBs (all asics).
346 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
350 if (!adev->ib_pool_ready)
353 for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
354 amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
355 adev->ib_pool_ready = false;
359 * amdgpu_ib_ring_tests - test IBs on the rings
361 * @adev: amdgpu_device pointer
363 * Test an IB (Indirect Buffer) on each ring.
364 * If the test fails, disable the ring.
365 * Returns 0 on success, error if the primary GFX ring
368 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
370 long tmo_gfx, tmo_mm;
374 tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
375 if (amdgpu_sriov_vf(adev)) {
376 /* for MM engines in hypervisor side they are not scheduled together
377 * with CP and SDMA engines, so even in exclusive mode MM engine could
378 * still running on other VF thus the IB TEST TIMEOUT for MM engines
379 * under SR-IOV should be set to a long time. 8 sec should be enough
380 * for the MM comes back to this VF.
382 tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
385 if (amdgpu_sriov_runtime(adev)) {
386 /* for CP & SDMA engines since they are scheduled together so
387 * need to make the timeout width enough to cover the time
388 * cost waiting for it coming back under RUNTIME only
390 tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
391 } else if (adev->gmc.xgmi.hive_id) {
392 tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
395 for (i = 0; i < adev->num_rings; ++i) {
396 struct amdgpu_ring *ring = adev->rings[i];
399 /* KIQ rings don't have an IB test because we never submit IBs
400 * to them and they have no interrupt support.
402 if (!ring->sched.ready || !ring->funcs->test_ib)
405 /* MM engine need more time */
406 if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
407 ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
408 ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
409 ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
410 ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
411 ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
416 r = amdgpu_ring_test_ib(ring, tmo);
418 DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
423 ring->sched.ready = false;
424 DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
427 if (ring == &adev->gfx.gfx_ring[0]) {
428 /* oh, oh, that's really bad */
429 adev->accel_working = false;
442 #if defined(CONFIG_DEBUG_FS)
444 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
446 struct drm_info_node *node = (struct drm_info_node *) m->private;
447 struct drm_device *dev = node->minor->dev;
448 struct amdgpu_device *adev = dev->dev_private;
450 seq_printf(m, "--------------------- DELAYED --------------------- \n");
451 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
453 seq_printf(m, "-------------------- IMMEDIATE -------------------- \n");
454 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
456 seq_printf(m, "--------------------- DIRECT ---------------------- \n");
457 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
462 static const struct drm_info_list amdgpu_debugfs_sa_list[] = {
463 {"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
468 int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
470 #if defined(CONFIG_DEBUG_FS)
471 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list,
472 ARRAY_SIZE(amdgpu_debugfs_sa_list));