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>
36 #include "amdgpu_trace.h"
38 #define AMDGPU_IB_TEST_TIMEOUT msecs_to_jiffies(1000)
39 #define AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT msecs_to_jiffies(2000)
43 * IBs (Indirect Buffers) and areas of GPU accessible memory where
44 * commands are stored. You can put a pointer to the IB in the
45 * command ring and the hw will fetch the commands from the IB
46 * and execute them. Generally userspace acceleration drivers
47 * produce command buffers which are send to the kernel and
48 * put in IBs for execution by the requested ring.
52 * amdgpu_ib_get - request an IB (Indirect Buffer)
54 * @adev: amdgpu_device pointer
55 * @vm: amdgpu_vm pointer
56 * @size: requested IB size
57 * @pool_type: IB pool type (delayed, immediate, direct)
58 * @ib: IB object returned
60 * Request an IB (all asics). IBs are allocated using the
62 * Returns 0 on success, error on failure.
64 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
65 unsigned size, enum amdgpu_ib_pool_type pool_type,
71 r = amdgpu_sa_bo_new(&adev->ib_pools[pool_type],
72 &ib->sa_bo, size, 256);
74 dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
78 ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
81 ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
88 * amdgpu_ib_free - free an IB (Indirect Buffer)
90 * @adev: amdgpu_device pointer
91 * @ib: IB object to free
92 * @f: the fence SA bo need wait on for the ib alloation
94 * Free an IB (all asics).
96 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
99 amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
103 * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
105 * @ring: ring index the IB is associated with
106 * @num_ibs: number of IBs to schedule
107 * @ibs: IB objects to schedule
108 * @job: job to schedule
109 * @f: fence created during this submission
111 * Schedule an IB on the associated ring (all asics).
112 * Returns 0 on success, error on failure.
114 * On SI, there are two parallel engines fed from the primary ring,
115 * the CE (Constant Engine) and the DE (Drawing Engine). Since
116 * resource descriptors have moved to memory, the CE allows you to
117 * prime the caches while the DE is updating register state so that
118 * the resource descriptors will be already in cache when the draw is
119 * processed. To accomplish this, the userspace driver submits two
120 * IBs, one for the CE and one for the DE. If there is a CE IB (called
121 * a CONST_IB), it will be put on the ring prior to the DE IB. Prior
122 * to SI there was just a DE IB.
124 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
125 struct amdgpu_ib *ibs, struct amdgpu_job *job,
126 struct dma_fence **f)
128 struct amdgpu_device *adev = ring->adev;
129 struct amdgpu_ib *ib = &ibs[0];
130 struct dma_fence *tmp = NULL;
131 bool skip_preamble, need_ctx_switch;
132 unsigned patch_offset = ~0;
133 struct amdgpu_vm *vm;
135 uint32_t status = 0, alloc_size;
136 unsigned fence_flags = 0;
141 bool need_pipe_sync = false;
146 /* ring tests don't use a job */
149 fence_ctx = job->base.s_fence ?
150 job->base.s_fence->scheduled.context : 0;
156 if (!ring->sched.ready) {
157 dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
161 if (vm && !job->vmid) {
162 dev_err(adev->dev, "VM IB without ID\n");
166 if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
167 (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)) {
168 dev_err(adev->dev, "secure submissions not supported on compute rings\n");
172 alloc_size = ring->funcs->emit_frame_size + num_ibs *
173 ring->funcs->emit_ib_size;
175 r = amdgpu_ring_alloc(ring, alloc_size);
177 dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
181 need_ctx_switch = ring->current_ctx != fence_ctx;
182 if (ring->funcs->emit_pipeline_sync && job &&
183 ((tmp = amdgpu_sync_get_fence(&job->sched_sync)) ||
184 (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
185 amdgpu_vm_need_pipeline_sync(ring, job))) {
186 need_pipe_sync = true;
189 trace_amdgpu_ib_pipe_sync(job, tmp);
194 if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
195 ring->funcs->emit_mem_sync(ring);
197 if (ring->funcs->emit_wave_limit &&
198 ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
199 ring->funcs->emit_wave_limit(ring, true);
201 if (ring->funcs->insert_start)
202 ring->funcs->insert_start(ring);
205 r = amdgpu_vm_flush(ring, job, need_pipe_sync);
207 amdgpu_ring_undo(ring);
212 if (job && ring->funcs->init_cond_exec)
213 patch_offset = amdgpu_ring_init_cond_exec(ring);
216 if (!(adev->flags & AMD_IS_APU))
219 if (ring->funcs->emit_hdp_flush)
220 amdgpu_ring_emit_hdp_flush(ring);
222 amdgpu_asic_flush_hdp(adev, ring);
226 status |= AMDGPU_HAVE_CTX_SWITCH;
228 skip_preamble = ring->current_ctx == fence_ctx;
229 if (job && ring->funcs->emit_cntxcntl) {
230 status |= job->preamble_status;
231 status |= job->preemption_status;
232 amdgpu_ring_emit_cntxcntl(ring, status);
235 /* Setup initial TMZiness and send it off.
238 if (job && ring->funcs->emit_frame_cntl) {
239 secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
240 amdgpu_ring_emit_frame_cntl(ring, true, secure);
243 for (i = 0; i < num_ibs; ++i) {
246 /* drop preamble IBs if we don't have a context switch */
247 if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
249 !(status & AMDGPU_PREAMBLE_IB_PRESENT_FIRST) &&
251 !amdgpu_sriov_vf(adev)) /* for SRIOV preemption, Preamble CE ib must be inserted anyway */
254 if (job && ring->funcs->emit_frame_cntl) {
255 if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) {
256 amdgpu_ring_emit_frame_cntl(ring, false, secure);
258 amdgpu_ring_emit_frame_cntl(ring, true, secure);
262 amdgpu_ring_emit_ib(ring, job, ib, status);
263 status &= ~AMDGPU_HAVE_CTX_SWITCH;
266 if (job && ring->funcs->emit_frame_cntl)
267 amdgpu_ring_emit_frame_cntl(ring, false, secure);
270 if (!(adev->flags & AMD_IS_APU))
272 amdgpu_asic_invalidate_hdp(adev, ring);
274 if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
275 fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
277 /* wrap the last IB with fence */
278 if (job && job->uf_addr) {
279 amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
280 fence_flags | AMDGPU_FENCE_FLAG_64BIT);
283 r = amdgpu_fence_emit(ring, f, fence_flags);
285 dev_err(adev->dev, "failed to emit fence (%d)\n", r);
286 if (job && job->vmid)
287 amdgpu_vmid_reset(adev, ring->funcs->vmhub, job->vmid);
288 amdgpu_ring_undo(ring);
292 if (ring->funcs->insert_end)
293 ring->funcs->insert_end(ring);
295 if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
296 amdgpu_ring_patch_cond_exec(ring, patch_offset);
298 ring->current_ctx = fence_ctx;
299 if (vm && ring->funcs->emit_switch_buffer)
300 amdgpu_ring_emit_switch_buffer(ring);
302 if (ring->funcs->emit_wave_limit &&
303 ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
304 ring->funcs->emit_wave_limit(ring, false);
306 amdgpu_ring_commit(ring);
311 * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
313 * @adev: amdgpu_device pointer
315 * Initialize the suballocator to manage a pool of memory
316 * for use as IBs (all asics).
317 * Returns 0 on success, error on failure.
319 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
324 if (adev->ib_pool_ready)
327 for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
328 if (i == AMDGPU_IB_POOL_DIRECT)
329 size = PAGE_SIZE * 2;
331 size = AMDGPU_IB_POOL_SIZE;
333 r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
334 size, AMDGPU_GPU_PAGE_SIZE,
335 AMDGPU_GEM_DOMAIN_GTT);
339 adev->ib_pool_ready = true;
345 amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
350 * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
352 * @adev: amdgpu_device pointer
354 * Tear down the suballocator managing the pool of memory
355 * for use as IBs (all asics).
357 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
361 if (!adev->ib_pool_ready)
364 for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
365 amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
366 adev->ib_pool_ready = false;
370 * amdgpu_ib_ring_tests - test IBs on the rings
372 * @adev: amdgpu_device pointer
374 * Test an IB (Indirect Buffer) on each ring.
375 * If the test fails, disable the ring.
376 * Returns 0 on success, error if the primary GFX ring
379 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
381 long tmo_gfx, tmo_mm;
385 tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
386 if (amdgpu_sriov_vf(adev)) {
387 /* for MM engines in hypervisor side they are not scheduled together
388 * with CP and SDMA engines, so even in exclusive mode MM engine could
389 * still running on other VF thus the IB TEST TIMEOUT for MM engines
390 * under SR-IOV should be set to a long time. 8 sec should be enough
391 * for the MM comes back to this VF.
393 tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
396 if (amdgpu_sriov_runtime(adev)) {
397 /* for CP & SDMA engines since they are scheduled together so
398 * need to make the timeout width enough to cover the time
399 * cost waiting for it coming back under RUNTIME only
401 tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
402 } else if (adev->gmc.xgmi.hive_id) {
403 tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
406 for (i = 0; i < adev->num_rings; ++i) {
407 struct amdgpu_ring *ring = adev->rings[i];
410 /* KIQ rings don't have an IB test because we never submit IBs
411 * to them and they have no interrupt support.
413 if (!ring->sched.ready || !ring->funcs->test_ib)
416 /* MM engine need more time */
417 if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
418 ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
419 ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
420 ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
421 ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
422 ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
427 r = amdgpu_ring_test_ib(ring, tmo);
429 DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
434 ring->sched.ready = false;
435 DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
438 if (ring == &adev->gfx.gfx_ring[0]) {
439 /* oh, oh, that's really bad */
440 adev->accel_working = false;
453 #if defined(CONFIG_DEBUG_FS)
455 static int amdgpu_debugfs_sa_info_show(struct seq_file *m, void *unused)
457 struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
459 seq_printf(m, "--------------------- DELAYED --------------------- \n");
460 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
462 seq_printf(m, "-------------------- IMMEDIATE -------------------- \n");
463 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
465 seq_printf(m, "--------------------- DIRECT ---------------------- \n");
466 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
471 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_sa_info);
475 void amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
477 #if defined(CONFIG_DEBUG_FS)
478 struct drm_minor *minor = adev_to_drm(adev)->primary;
479 struct dentry *root = minor->debugfs_root;
481 debugfs_create_file("amdgpu_sa_info", 0444, root, adev,
482 &amdgpu_debugfs_sa_info_fops);