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 */
amdgpu_ib_get(struct amdgpu_device * adev,struct amdgpu_vm * vm,unsigned int size,enum amdgpu_ib_pool_type pool_type,struct amdgpu_ib * ib)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(&adev->ib_pools[pool_type],
72 &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(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(ib->sa_bo);
84 }
85
86 return 0;
87 }
88
89 /**
90 * amdgpu_ib_free - free an IB (Indirect Buffer)
91 *
92 * @ib: IB object to free
93 * @f: the fence SA bo need wait on for the ib alloation
94 *
95 * Free an IB (all asics).
96 */
amdgpu_ib_free(struct amdgpu_ib * ib,struct dma_fence * f)97 void amdgpu_ib_free(struct amdgpu_ib *ib, struct dma_fence *f)
98 {
99 amdgpu_sa_bo_free(&ib->sa_bo, f);
100 }
101
102 /**
103 * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
104 *
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
110 *
111 * Schedule an IB on the associated ring (all asics).
112 * Returns 0 on success, error on failure.
113 *
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.
123 */
amdgpu_ib_schedule(struct amdgpu_ring * ring,unsigned int num_ibs,struct amdgpu_ib * ibs,struct amdgpu_job * job,struct dma_fence ** f)124 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs,
125 struct amdgpu_ib *ibs, struct amdgpu_job *job,
126 struct dma_fence **f)
127 {
128 struct amdgpu_device *adev = ring->adev;
129 struct amdgpu_ib *ib = &ibs[0];
130 struct dma_fence *tmp = NULL;
131 struct amdgpu_fence *af;
132 bool need_ctx_switch;
133 struct amdgpu_vm *vm;
134 uint64_t fence_ctx;
135 uint32_t status = 0, alloc_size;
136 unsigned int fence_flags = 0;
137 bool secure, init_shadow;
138 u64 shadow_va, csa_va, gds_va;
139 int vmid = AMDGPU_JOB_GET_VMID(job);
140 bool need_pipe_sync = false;
141 unsigned int cond_exec;
142 unsigned int i;
143 int r = 0;
144
145 if (num_ibs == 0)
146 return -EINVAL;
147
148 /* ring tests don't use a job */
149 if (job) {
150 vm = job->vm;
151 fence_ctx = job->base.s_fence ?
152 job->base.s_fence->finished.context : 0;
153 shadow_va = job->shadow_va;
154 csa_va = job->csa_va;
155 gds_va = job->gds_va;
156 init_shadow = job->init_shadow;
157 af = job->hw_fence;
158 /* Save the context of the job for reset handling.
159 * The driver needs this so it can skip the ring
160 * contents for guilty contexts.
161 */
162 af->context = fence_ctx;
163 /* the vm fence is also part of the job's context */
164 job->hw_vm_fence->context = fence_ctx;
165 } else {
166 vm = NULL;
167 fence_ctx = 0;
168 shadow_va = 0;
169 csa_va = 0;
170 gds_va = 0;
171 init_shadow = false;
172 af = kzalloc(sizeof(*af), GFP_ATOMIC);
173 if (!af)
174 return -ENOMEM;
175 }
176
177 if (!ring->sched.ready) {
178 dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
179 r = -EINVAL;
180 goto free_fence;
181 }
182
183 if (vm && !job->vmid) {
184 dev_err(adev->dev, "VM IB without ID\n");
185 r = -EINVAL;
186 goto free_fence;
187 }
188
189 if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
190 (!ring->funcs->secure_submission_supported)) {
191 dev_err(adev->dev, "secure submissions not supported on ring <%s>\n", ring->name);
192 r = -EINVAL;
193 goto free_fence;
194 }
195
196 alloc_size = ring->funcs->emit_frame_size + num_ibs *
197 ring->funcs->emit_ib_size;
198
199 r = amdgpu_ring_alloc(ring, alloc_size);
200 if (r) {
201 dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
202 goto free_fence;
203 }
204
205 need_ctx_switch = ring->current_ctx != fence_ctx;
206 if (ring->funcs->emit_pipeline_sync && job &&
207 ((tmp = amdgpu_sync_get_fence(&job->explicit_sync)) ||
208 need_ctx_switch || amdgpu_vm_need_pipeline_sync(ring, job))) {
209
210 need_pipe_sync = true;
211
212 if (tmp)
213 trace_amdgpu_ib_pipe_sync(job, tmp);
214
215 dma_fence_put(tmp);
216 }
217
218 if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
219 ring->funcs->emit_mem_sync(ring);
220
221 if (ring->funcs->emit_wave_limit &&
222 ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
223 ring->funcs->emit_wave_limit(ring, true);
224
225 if (ring->funcs->insert_start)
226 ring->funcs->insert_start(ring);
227
228 if (job) {
229 r = amdgpu_vm_flush(ring, job, need_pipe_sync);
230 if (r) {
231 amdgpu_ring_undo(ring);
232 return r;
233 }
234 }
235
236 amdgpu_ring_ib_begin(ring);
237
238 if (ring->funcs->emit_gfx_shadow && adev->gfx.cp_gfx_shadow)
239 amdgpu_ring_emit_gfx_shadow(ring, shadow_va, csa_va, gds_va,
240 init_shadow, vmid);
241
242 if (ring->funcs->init_cond_exec)
243 cond_exec = amdgpu_ring_init_cond_exec(ring,
244 ring->cond_exe_gpu_addr);
245
246 amdgpu_device_flush_hdp(adev, ring);
247
248 if (need_ctx_switch)
249 status |= AMDGPU_HAVE_CTX_SWITCH;
250
251 if (job && ring->funcs->emit_cntxcntl) {
252 status |= job->preamble_status;
253 status |= job->preemption_status;
254 amdgpu_ring_emit_cntxcntl(ring, status);
255 }
256
257 /* Setup initial TMZiness and send it off.
258 */
259 secure = false;
260 if (job && ring->funcs->emit_frame_cntl) {
261 secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
262 amdgpu_ring_emit_frame_cntl(ring, true, secure);
263 }
264
265 for (i = 0; i < num_ibs; ++i) {
266 ib = &ibs[i];
267
268 if (job && ring->funcs->emit_frame_cntl) {
269 if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) {
270 amdgpu_ring_emit_frame_cntl(ring, false, secure);
271 secure = !secure;
272 amdgpu_ring_emit_frame_cntl(ring, true, secure);
273 }
274 }
275
276 amdgpu_ring_emit_ib(ring, job, ib, status);
277 status &= ~AMDGPU_HAVE_CTX_SWITCH;
278 }
279
280 if (job && ring->funcs->emit_frame_cntl)
281 amdgpu_ring_emit_frame_cntl(ring, false, secure);
282
283 amdgpu_device_invalidate_hdp(adev, ring);
284
285 if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
286 fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
287
288 /* wrap the last IB with fence */
289 if (job && job->uf_addr) {
290 amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
291 fence_flags | AMDGPU_FENCE_FLAG_64BIT);
292 }
293
294 if (ring->funcs->emit_gfx_shadow && ring->funcs->init_cond_exec &&
295 adev->gfx.cp_gfx_shadow) {
296 amdgpu_ring_emit_gfx_shadow(ring, 0, 0, 0, false, 0);
297 amdgpu_ring_init_cond_exec(ring, ring->cond_exe_gpu_addr);
298 }
299
300 r = amdgpu_fence_emit(ring, af, fence_flags);
301 if (r) {
302 dev_err(adev->dev, "failed to emit fence (%d)\n", r);
303 if (job && job->vmid)
304 amdgpu_vmid_reset(adev, ring->vm_hub, job->vmid);
305 amdgpu_ring_undo(ring);
306 goto free_fence;
307 }
308 *f = &af->base;
309 /* get a ref for the job */
310 if (job)
311 dma_fence_get(*f);
312
313 if (ring->funcs->insert_end)
314 ring->funcs->insert_end(ring);
315
316 amdgpu_ring_patch_cond_exec(ring, cond_exec);
317
318 ring->current_ctx = fence_ctx;
319 if (job && ring->funcs->emit_switch_buffer)
320 amdgpu_ring_emit_switch_buffer(ring);
321
322 if (ring->funcs->emit_wave_limit &&
323 ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
324 ring->funcs->emit_wave_limit(ring, false);
325
326 /* Save the wptr associated with this fence.
327 * This must be last for resets to work properly
328 * as we need to save the wptr associated with this
329 * fence so we know what rings contents to backup
330 * after we reset the queue.
331 */
332 amdgpu_fence_save_wptr(af);
333
334 amdgpu_ring_ib_end(ring);
335 amdgpu_ring_commit(ring);
336
337 return 0;
338
339 free_fence:
340 if (!job)
341 kfree(af);
342 return r;
343 }
344
345 /**
346 * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
347 *
348 * @adev: amdgpu_device pointer
349 *
350 * Initialize the suballocator to manage a pool of memory
351 * for use as IBs (all asics).
352 * Returns 0 on success, error on failure.
353 */
amdgpu_ib_pool_init(struct amdgpu_device * adev)354 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
355 {
356 int r, i;
357
358 if (adev->ib_pool_ready)
359 return 0;
360
361 for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
362 r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
363 AMDGPU_IB_POOL_SIZE, 256,
364 AMDGPU_GEM_DOMAIN_GTT);
365 if (r)
366 goto error;
367 }
368 adev->ib_pool_ready = true;
369
370 return 0;
371
372 error:
373 while (i--)
374 amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
375 return r;
376 }
377
378 /**
379 * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
380 *
381 * @adev: amdgpu_device pointer
382 *
383 * Tear down the suballocator managing the pool of memory
384 * for use as IBs (all asics).
385 */
amdgpu_ib_pool_fini(struct amdgpu_device * adev)386 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
387 {
388 int i;
389
390 if (!adev->ib_pool_ready)
391 return;
392
393 for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
394 amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
395 adev->ib_pool_ready = false;
396 }
397
398 /**
399 * amdgpu_ib_ring_tests - test IBs on the rings
400 *
401 * @adev: amdgpu_device pointer
402 *
403 * Test an IB (Indirect Buffer) on each ring.
404 * If the test fails, disable the ring.
405 * Returns 0 on success, error if the primary GFX ring
406 * IB test fails.
407 */
amdgpu_ib_ring_tests(struct amdgpu_device * adev)408 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
409 {
410 long tmo_gfx, tmo_mm;
411 int r, ret = 0;
412 unsigned int i;
413
414 tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
415 if (amdgpu_sriov_vf(adev)) {
416 /* for MM engines in hypervisor side they are not scheduled together
417 * with CP and SDMA engines, so even in exclusive mode MM engine could
418 * still running on other VF thus the IB TEST TIMEOUT for MM engines
419 * under SR-IOV should be set to a long time. 8 sec should be enough
420 * for the MM comes back to this VF.
421 */
422 tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
423 }
424
425 if (amdgpu_sriov_runtime(adev)) {
426 /* for CP & SDMA engines since they are scheduled together so
427 * need to make the timeout width enough to cover the time
428 * cost waiting for it coming back under RUNTIME only
429 */
430 tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
431 } else if (adev->gmc.xgmi.hive_id) {
432 tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
433 }
434
435 for (i = 0; i < adev->num_rings; ++i) {
436 struct amdgpu_ring *ring = adev->rings[i];
437 long tmo;
438
439 /* KIQ rings don't have an IB test because we never submit IBs
440 * to them and they have no interrupt support.
441 */
442 if (!ring->sched.ready || !ring->funcs->test_ib)
443 continue;
444
445 if (adev->enable_mes &&
446 ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
447 continue;
448
449 /* MM engine need more time */
450 if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
451 ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
452 ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
453 ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
454 ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
455 ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
456 tmo = tmo_mm;
457 else
458 tmo = tmo_gfx;
459
460 r = amdgpu_ring_test_ib(ring, tmo);
461 if (!r) {
462 DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
463 ring->name);
464 continue;
465 }
466
467 ring->sched.ready = false;
468 DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
469 ring->name, r);
470
471 if (ring == &adev->gfx.gfx_ring[0]) {
472 /* oh, oh, that's really bad */
473 adev->accel_working = false;
474 return r;
475
476 } else {
477 ret = r;
478 }
479 }
480 return ret;
481 }
482
483 /*
484 * Debugfs info
485 */
486 #if defined(CONFIG_DEBUG_FS)
487
amdgpu_debugfs_sa_info_show(struct seq_file * m,void * unused)488 static int amdgpu_debugfs_sa_info_show(struct seq_file *m, void *unused)
489 {
490 struct amdgpu_device *adev = m->private;
491
492 seq_puts(m, "--------------------- DELAYED ---------------------\n");
493 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
494 m);
495 seq_puts(m, "-------------------- IMMEDIATE --------------------\n");
496 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
497 m);
498 seq_puts(m, "--------------------- DIRECT ----------------------\n");
499 amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
500
501 return 0;
502 }
503
504 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_sa_info);
505
506 #endif
507
amdgpu_debugfs_sa_init(struct amdgpu_device * adev)508 void amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
509 {
510 #if defined(CONFIG_DEBUG_FS)
511 struct drm_minor *minor = adev_to_drm(adev)->primary;
512 struct dentry *root = minor->debugfs_root;
513
514 debugfs_create_file("amdgpu_sa_info", 0444, root, adev,
515 &amdgpu_debugfs_sa_info_fops);
516
517 #endif
518 }
519