1 | /* |
2 | * Copyright 2008 Advanced Micro Devices, Inc. |
3 | * Copyright 2008 Red Hat Inc. |
4 | * Copyright 2009 Jerome Glisse. |
5 | * |
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: |
12 | * |
13 | * The above copyright notice and this permission notice shall be included in |
14 | * all copies or substantial portions of the Software. |
15 | * |
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. |
23 | * |
24 | * Authors: Dave Airlie |
25 | * Alex Deucher |
26 | * Jerome Glisse |
27 | * Christian König |
28 | */ |
29 | #include <linux/seq_file.h> |
30 | #include <linux/slab.h> |
31 | |
32 | #include <drm/amdgpu_drm.h> |
33 | |
34 | #include "amdgpu.h" |
35 | #include "atom.h" |
36 | #include "amdgpu_trace.h" |
37 | |
38 | #define AMDGPU_IB_TEST_TIMEOUT msecs_to_jiffies(1000) |
39 | #define AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT msecs_to_jiffies(2000) |
40 | |
41 | /* |
42 | * IB |
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. |
49 | */ |
50 | |
51 | /** |
52 | * amdgpu_ib_get - request an IB (Indirect Buffer) |
53 | * |
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 |
59 | * |
60 | * Request an IB (all asics). IBs are allocated using the |
61 | * suballocator. |
62 | * Returns 0 on success, error on failure. |
63 | */ |
64 | int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm, |
65 | unsigned int size, enum amdgpu_ib_pool_type pool_type, |
66 | struct amdgpu_ib *ib) |
67 | { |
68 | int r; |
69 | |
70 | if (size) { |
71 | r = amdgpu_sa_bo_new(sa_manager: &adev->ib_pools[pool_type], |
72 | sa_bo: &ib->sa_bo, size); |
73 | if (r) { |
74 | dev_err(adev->dev, "failed to get a new IB (%d)\n" , r); |
75 | return r; |
76 | } |
77 | |
78 | ib->ptr = amdgpu_sa_bo_cpu_addr(sa_bo: ib->sa_bo); |
79 | /* flush the cache before commit the IB */ |
80 | ib->flags = AMDGPU_IB_FLAG_EMIT_MEM_SYNC; |
81 | |
82 | if (!vm) |
83 | ib->gpu_addr = amdgpu_sa_bo_gpu_addr(sa_bo: ib->sa_bo); |
84 | } |
85 | |
86 | return 0; |
87 | } |
88 | |
89 | /** |
90 | * amdgpu_ib_free - free an IB (Indirect Buffer) |
91 | * |
92 | * @adev: amdgpu_device pointer |
93 | * @ib: IB object to free |
94 | * @f: the fence SA bo need wait on for the ib alloation |
95 | * |
96 | * Free an IB (all asics). |
97 | */ |
98 | void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib, |
99 | struct dma_fence *f) |
100 | { |
101 | amdgpu_sa_bo_free(adev, sa_bo: &ib->sa_bo, fence: f); |
102 | } |
103 | |
104 | /** |
105 | * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring |
106 | * |
107 | * @ring: ring index the IB is associated with |
108 | * @num_ibs: number of IBs to schedule |
109 | * @ibs: IB objects to schedule |
110 | * @job: job to schedule |
111 | * @f: fence created during this submission |
112 | * |
113 | * Schedule an IB on the associated ring (all asics). |
114 | * Returns 0 on success, error on failure. |
115 | * |
116 | * On SI, there are two parallel engines fed from the primary ring, |
117 | * the CE (Constant Engine) and the DE (Drawing Engine). Since |
118 | * resource descriptors have moved to memory, the CE allows you to |
119 | * prime the caches while the DE is updating register state so that |
120 | * the resource descriptors will be already in cache when the draw is |
121 | * processed. To accomplish this, the userspace driver submits two |
122 | * IBs, one for the CE and one for the DE. If there is a CE IB (called |
123 | * a CONST_IB), it will be put on the ring prior to the DE IB. Prior |
124 | * to SI there was just a DE IB. |
125 | */ |
126 | int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs, |
127 | struct amdgpu_ib *ibs, struct amdgpu_job *job, |
128 | struct dma_fence **f) |
129 | { |
130 | struct amdgpu_device *adev = ring->adev; |
131 | struct amdgpu_ib *ib = &ibs[0]; |
132 | struct dma_fence *tmp = NULL; |
133 | bool need_ctx_switch; |
134 | struct amdgpu_vm *vm; |
135 | uint64_t fence_ctx; |
136 | uint32_t status = 0, alloc_size; |
137 | unsigned int fence_flags = 0; |
138 | bool secure, init_shadow; |
139 | u64 shadow_va, csa_va, gds_va; |
140 | int vmid = AMDGPU_JOB_GET_VMID(job); |
141 | bool need_pipe_sync = false; |
142 | unsigned int cond_exec; |
143 | |
144 | unsigned int i; |
145 | int r = 0; |
146 | |
147 | if (num_ibs == 0) |
148 | return -EINVAL; |
149 | |
150 | /* ring tests don't use a job */ |
151 | if (job) { |
152 | vm = job->vm; |
153 | fence_ctx = job->base.s_fence ? |
154 | job->base.s_fence->scheduled.context : 0; |
155 | shadow_va = job->shadow_va; |
156 | csa_va = job->csa_va; |
157 | gds_va = job->gds_va; |
158 | init_shadow = job->init_shadow; |
159 | } else { |
160 | vm = NULL; |
161 | fence_ctx = 0; |
162 | shadow_va = 0; |
163 | csa_va = 0; |
164 | gds_va = 0; |
165 | init_shadow = false; |
166 | } |
167 | |
168 | if (!ring->sched.ready && !ring->is_mes_queue) { |
169 | dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n" , ring->name); |
170 | return -EINVAL; |
171 | } |
172 | |
173 | if (vm && !job->vmid && !ring->is_mes_queue) { |
174 | dev_err(adev->dev, "VM IB without ID\n" ); |
175 | return -EINVAL; |
176 | } |
177 | |
178 | if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) && |
179 | (!ring->funcs->secure_submission_supported)) { |
180 | dev_err(adev->dev, "secure submissions not supported on ring <%s>\n" , ring->name); |
181 | return -EINVAL; |
182 | } |
183 | |
184 | alloc_size = ring->funcs->emit_frame_size + num_ibs * |
185 | ring->funcs->emit_ib_size; |
186 | |
187 | r = amdgpu_ring_alloc(ring, ndw: alloc_size); |
188 | if (r) { |
189 | dev_err(adev->dev, "scheduling IB failed (%d).\n" , r); |
190 | return r; |
191 | } |
192 | |
193 | need_ctx_switch = ring->current_ctx != fence_ctx; |
194 | if (ring->funcs->emit_pipeline_sync && job && |
195 | ((tmp = amdgpu_sync_get_fence(sync: &job->explicit_sync)) || |
196 | (amdgpu_sriov_vf(adev) && need_ctx_switch) || |
197 | amdgpu_vm_need_pipeline_sync(ring, job))) { |
198 | need_pipe_sync = true; |
199 | |
200 | if (tmp) |
201 | trace_amdgpu_ib_pipe_sync(sched_job: job, fence: tmp); |
202 | |
203 | dma_fence_put(fence: tmp); |
204 | } |
205 | |
206 | if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync) |
207 | ring->funcs->emit_mem_sync(ring); |
208 | |
209 | if (ring->funcs->emit_wave_limit && |
210 | ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH) |
211 | ring->funcs->emit_wave_limit(ring, true); |
212 | |
213 | if (ring->funcs->insert_start) |
214 | ring->funcs->insert_start(ring); |
215 | |
216 | if (job) { |
217 | r = amdgpu_vm_flush(ring, job, need_pipe_sync); |
218 | if (r) { |
219 | amdgpu_ring_undo(ring); |
220 | return r; |
221 | } |
222 | } |
223 | |
224 | amdgpu_ring_ib_begin(ring); |
225 | |
226 | if (ring->funcs->emit_gfx_shadow) |
227 | amdgpu_ring_emit_gfx_shadow(ring, shadow_va, csa_va, gds_va, |
228 | init_shadow, vmid); |
229 | |
230 | if (ring->funcs->init_cond_exec) |
231 | cond_exec = amdgpu_ring_init_cond_exec(ring, |
232 | ring->cond_exe_gpu_addr); |
233 | |
234 | amdgpu_device_flush_hdp(adev, ring); |
235 | |
236 | if (need_ctx_switch) |
237 | status |= AMDGPU_HAVE_CTX_SWITCH; |
238 | |
239 | if (job && ring->funcs->emit_cntxcntl) { |
240 | status |= job->preamble_status; |
241 | status |= job->preemption_status; |
242 | amdgpu_ring_emit_cntxcntl(ring, status); |
243 | } |
244 | |
245 | /* Setup initial TMZiness and send it off. |
246 | */ |
247 | secure = false; |
248 | if (job && ring->funcs->emit_frame_cntl) { |
249 | secure = ib->flags & AMDGPU_IB_FLAGS_SECURE; |
250 | amdgpu_ring_emit_frame_cntl(ring, true, secure); |
251 | } |
252 | |
253 | for (i = 0; i < num_ibs; ++i) { |
254 | ib = &ibs[i]; |
255 | |
256 | if (job && ring->funcs->emit_frame_cntl) { |
257 | if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) { |
258 | amdgpu_ring_emit_frame_cntl(ring, false, secure); |
259 | secure = !secure; |
260 | amdgpu_ring_emit_frame_cntl(ring, true, secure); |
261 | } |
262 | } |
263 | |
264 | amdgpu_ring_emit_ib(ring, job, ib, status); |
265 | status &= ~AMDGPU_HAVE_CTX_SWITCH; |
266 | } |
267 | |
268 | if (job && ring->funcs->emit_frame_cntl) |
269 | amdgpu_ring_emit_frame_cntl(ring, false, secure); |
270 | |
271 | amdgpu_device_invalidate_hdp(adev, ring); |
272 | |
273 | if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE) |
274 | fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY; |
275 | |
276 | /* wrap the last IB with fence */ |
277 | if (job && job->uf_addr) { |
278 | amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence, |
279 | fence_flags | AMDGPU_FENCE_FLAG_64BIT); |
280 | } |
281 | |
282 | if (ring->funcs->emit_gfx_shadow && ring->funcs->init_cond_exec) { |
283 | amdgpu_ring_emit_gfx_shadow(ring, 0, 0, 0, false, 0); |
284 | amdgpu_ring_init_cond_exec(ring, ring->cond_exe_gpu_addr); |
285 | } |
286 | |
287 | r = amdgpu_fence_emit(ring, fence: f, job, flags: fence_flags); |
288 | if (r) { |
289 | dev_err(adev->dev, "failed to emit fence (%d)\n" , r); |
290 | if (job && job->vmid) |
291 | amdgpu_vmid_reset(adev, vmhub: ring->vm_hub, vmid: job->vmid); |
292 | amdgpu_ring_undo(ring); |
293 | return r; |
294 | } |
295 | |
296 | if (ring->funcs->insert_end) |
297 | ring->funcs->insert_end(ring); |
298 | |
299 | amdgpu_ring_patch_cond_exec(ring, offset: cond_exec); |
300 | |
301 | ring->current_ctx = fence_ctx; |
302 | if (vm && ring->funcs->emit_switch_buffer) |
303 | amdgpu_ring_emit_switch_buffer(ring); |
304 | |
305 | if (ring->funcs->emit_wave_limit && |
306 | ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH) |
307 | ring->funcs->emit_wave_limit(ring, false); |
308 | |
309 | amdgpu_ring_ib_end(ring); |
310 | amdgpu_ring_commit(ring); |
311 | return 0; |
312 | } |
313 | |
314 | /** |
315 | * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool |
316 | * |
317 | * @adev: amdgpu_device pointer |
318 | * |
319 | * Initialize the suballocator to manage a pool of memory |
320 | * for use as IBs (all asics). |
321 | * Returns 0 on success, error on failure. |
322 | */ |
323 | int amdgpu_ib_pool_init(struct amdgpu_device *adev) |
324 | { |
325 | int r, i; |
326 | |
327 | if (adev->ib_pool_ready) |
328 | return 0; |
329 | |
330 | for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) { |
331 | r = amdgpu_sa_bo_manager_init(adev, sa_manager: &adev->ib_pools[i], |
332 | AMDGPU_IB_POOL_SIZE, align: 256, |
333 | AMDGPU_GEM_DOMAIN_GTT); |
334 | if (r) |
335 | goto error; |
336 | } |
337 | adev->ib_pool_ready = true; |
338 | |
339 | return 0; |
340 | |
341 | error: |
342 | while (i--) |
343 | amdgpu_sa_bo_manager_fini(adev, sa_manager: &adev->ib_pools[i]); |
344 | return r; |
345 | } |
346 | |
347 | /** |
348 | * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool |
349 | * |
350 | * @adev: amdgpu_device pointer |
351 | * |
352 | * Tear down the suballocator managing the pool of memory |
353 | * for use as IBs (all asics). |
354 | */ |
355 | void amdgpu_ib_pool_fini(struct amdgpu_device *adev) |
356 | { |
357 | int i; |
358 | |
359 | if (!adev->ib_pool_ready) |
360 | return; |
361 | |
362 | for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) |
363 | amdgpu_sa_bo_manager_fini(adev, sa_manager: &adev->ib_pools[i]); |
364 | adev->ib_pool_ready = false; |
365 | } |
366 | |
367 | /** |
368 | * amdgpu_ib_ring_tests - test IBs on the rings |
369 | * |
370 | * @adev: amdgpu_device pointer |
371 | * |
372 | * Test an IB (Indirect Buffer) on each ring. |
373 | * If the test fails, disable the ring. |
374 | * Returns 0 on success, error if the primary GFX ring |
375 | * IB test fails. |
376 | */ |
377 | int amdgpu_ib_ring_tests(struct amdgpu_device *adev) |
378 | { |
379 | long tmo_gfx, tmo_mm; |
380 | int r, ret = 0; |
381 | unsigned int i; |
382 | |
383 | tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT; |
384 | if (amdgpu_sriov_vf(adev)) { |
385 | /* for MM engines in hypervisor side they are not scheduled together |
386 | * with CP and SDMA engines, so even in exclusive mode MM engine could |
387 | * still running on other VF thus the IB TEST TIMEOUT for MM engines |
388 | * under SR-IOV should be set to a long time. 8 sec should be enough |
389 | * for the MM comes back to this VF. |
390 | */ |
391 | tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT; |
392 | } |
393 | |
394 | if (amdgpu_sriov_runtime(adev)) { |
395 | /* for CP & SDMA engines since they are scheduled together so |
396 | * need to make the timeout width enough to cover the time |
397 | * cost waiting for it coming back under RUNTIME only |
398 | */ |
399 | tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT; |
400 | } else if (adev->gmc.xgmi.hive_id) { |
401 | tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT; |
402 | } |
403 | |
404 | for (i = 0; i < adev->num_rings; ++i) { |
405 | struct amdgpu_ring *ring = adev->rings[i]; |
406 | long tmo; |
407 | |
408 | /* KIQ rings don't have an IB test because we never submit IBs |
409 | * to them and they have no interrupt support. |
410 | */ |
411 | if (!ring->sched.ready || !ring->funcs->test_ib) |
412 | continue; |
413 | |
414 | if (adev->enable_mes && |
415 | ring->funcs->type == AMDGPU_RING_TYPE_KIQ) |
416 | continue; |
417 | |
418 | /* MM engine need more time */ |
419 | if (ring->funcs->type == AMDGPU_RING_TYPE_UVD || |
420 | ring->funcs->type == AMDGPU_RING_TYPE_VCE || |
421 | ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC || |
422 | ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC || |
423 | ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC || |
424 | ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG) |
425 | tmo = tmo_mm; |
426 | else |
427 | tmo = tmo_gfx; |
428 | |
429 | r = amdgpu_ring_test_ib(ring, tmo); |
430 | if (!r) { |
431 | DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n" , |
432 | ring->name); |
433 | continue; |
434 | } |
435 | |
436 | ring->sched.ready = false; |
437 | DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n" , |
438 | ring->name, r); |
439 | |
440 | if (ring == &adev->gfx.gfx_ring[0]) { |
441 | /* oh, oh, that's really bad */ |
442 | adev->accel_working = false; |
443 | return r; |
444 | |
445 | } else { |
446 | ret = r; |
447 | } |
448 | } |
449 | return ret; |
450 | } |
451 | |
452 | /* |
453 | * Debugfs info |
454 | */ |
455 | #if defined(CONFIG_DEBUG_FS) |
456 | |
457 | static int amdgpu_debugfs_sa_info_show(struct seq_file *m, void *unused) |
458 | { |
459 | struct amdgpu_device *adev = m->private; |
460 | |
461 | seq_puts(m, s: "--------------------- DELAYED ---------------------\n" ); |
462 | amdgpu_sa_bo_dump_debug_info(sa_manager: &adev->ib_pools[AMDGPU_IB_POOL_DELAYED], |
463 | m); |
464 | seq_puts(m, s: "-------------------- IMMEDIATE --------------------\n" ); |
465 | amdgpu_sa_bo_dump_debug_info(sa_manager: &adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE], |
466 | m); |
467 | seq_puts(m, s: "--------------------- DIRECT ----------------------\n" ); |
468 | amdgpu_sa_bo_dump_debug_info(sa_manager: &adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m); |
469 | |
470 | return 0; |
471 | } |
472 | |
473 | DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_sa_info); |
474 | |
475 | #endif |
476 | |
477 | void amdgpu_debugfs_sa_init(struct amdgpu_device *adev) |
478 | { |
479 | #if defined(CONFIG_DEBUG_FS) |
480 | struct drm_minor *minor = adev_to_drm(adev)->primary; |
481 | struct dentry *root = minor->debugfs_root; |
482 | |
483 | debugfs_create_file(name: "amdgpu_sa_info" , mode: 0444, parent: root, data: adev, |
484 | fops: &amdgpu_debugfs_sa_info_fops); |
485 | |
486 | #endif |
487 | } |
488 | |