xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c (revision 992694ad28584188725751f695a85661e8e3797a)
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(&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  */
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  */
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 	struct amdgpu_fence *vm_af;
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 	unsigned int i;
144 	int r = 0;
145 
146 	if (num_ibs == 0)
147 		return -EINVAL;
148 
149 	/* ring tests don't use a job */
150 	if (job) {
151 		vm = job->vm;
152 		fence_ctx = job->base.s_fence ?
153 			job->base.s_fence->finished.context : 0;
154 		shadow_va = job->shadow_va;
155 		csa_va = job->csa_va;
156 		gds_va = job->gds_va;
157 		init_shadow = job->init_shadow;
158 		af = job->hw_fence;
159 		/* Save the context of the job for reset handling.
160 		 * The driver needs this so it can skip the ring
161 		 * contents for guilty contexts.
162 		 */
163 		af->context = fence_ctx;
164 		/* the vm fence is also part of the job's context */
165 		job->hw_vm_fence->context = fence_ctx;
166 	} else {
167 		vm = NULL;
168 		fence_ctx = 0;
169 		shadow_va = 0;
170 		csa_va = 0;
171 		gds_va = 0;
172 		init_shadow = false;
173 		af = kzalloc_obj(*af, GFP_ATOMIC);
174 		if (!af)
175 			return -ENOMEM;
176 	}
177 
178 	if (!ring->sched.ready) {
179 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
180 		r = -EINVAL;
181 		goto free_fence;
182 	}
183 
184 	if (vm && !job->vmid) {
185 		dev_err(adev->dev, "VM IB without ID\n");
186 		r = -EINVAL;
187 		goto free_fence;
188 	}
189 
190 	if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
191 	    (!ring->funcs->secure_submission_supported)) {
192 		dev_err(adev->dev, "secure submissions not supported on ring <%s>\n", ring->name);
193 		r = -EINVAL;
194 		goto free_fence;
195 	}
196 
197 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
198 		ring->funcs->emit_ib_size;
199 
200 	r = amdgpu_ring_alloc(ring, alloc_size);
201 	if (r) {
202 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
203 		goto free_fence;
204 	}
205 
206 	need_ctx_switch = ring->current_ctx != fence_ctx;
207 	if (ring->funcs->emit_pipeline_sync && job &&
208 	    ((tmp = amdgpu_sync_get_fence(&job->explicit_sync)) ||
209 	     need_ctx_switch || amdgpu_vm_need_pipeline_sync(ring, job))) {
210 
211 		need_pipe_sync = true;
212 
213 		if (tmp)
214 			trace_amdgpu_ib_pipe_sync(job, tmp);
215 
216 		dma_fence_put(tmp);
217 	}
218 
219 	if (job) {
220 		vm_af = job->hw_vm_fence;
221 		/* VM sequence */
222 		vm_af->ib_wptr = ring->wptr;
223 		amdgpu_vm_flush(ring, job, need_pipe_sync);
224 		vm_af->ib_dw_size =
225 			amdgpu_ring_get_dw_distance(ring, vm_af->ib_wptr, ring->wptr);
226 	}
227 
228 	/* IB sequence */
229 	af->ib_wptr = ring->wptr;
230 	amdgpu_ring_ib_begin(ring);
231 
232 	if (ring->funcs->insert_start)
233 		ring->funcs->insert_start(ring);
234 
235 	if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
236 		ring->funcs->emit_mem_sync(ring);
237 
238 	if (ring->funcs->emit_wave_limit &&
239 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
240 		ring->funcs->emit_wave_limit(ring, true);
241 
242 	if (ring->funcs->emit_gfx_shadow && adev->gfx.cp_gfx_shadow)
243 		amdgpu_ring_emit_gfx_shadow(ring, shadow_va, csa_va, gds_va,
244 					    init_shadow, vmid);
245 
246 	if (ring->funcs->init_cond_exec)
247 		cond_exec = amdgpu_ring_init_cond_exec(ring,
248 						       ring->cond_exe_gpu_addr);
249 
250 	/* Skip the IB for guilty contexts */
251 	af->skip_ib_dw_start_offset =
252 		amdgpu_ring_get_dw_distance(ring, af->ib_wptr, ring->wptr);
253 	amdgpu_device_flush_hdp(adev, ring);
254 
255 	if (need_ctx_switch)
256 		status |= AMDGPU_HAVE_CTX_SWITCH;
257 
258 	if (job && ring->funcs->emit_cntxcntl) {
259 		status |= job->preamble_status;
260 		status |= job->preemption_status;
261 		amdgpu_ring_emit_cntxcntl(ring, status);
262 	}
263 
264 	/* Setup initial TMZiness and send it off.
265 	 */
266 	secure = false;
267 	if (job && ring->funcs->emit_frame_cntl) {
268 		secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
269 		amdgpu_ring_emit_frame_cntl(ring, true, secure);
270 	}
271 
272 	for (i = 0; i < num_ibs; ++i) {
273 		ib = &ibs[i];
274 
275 		if (job && ring->funcs->emit_frame_cntl) {
276 			if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) {
277 				amdgpu_ring_emit_frame_cntl(ring, false, secure);
278 				secure = !secure;
279 				amdgpu_ring_emit_frame_cntl(ring, true, secure);
280 			}
281 		}
282 
283 		amdgpu_ring_emit_ib(ring, job, ib, status);
284 		status &= ~AMDGPU_HAVE_CTX_SWITCH;
285 	}
286 
287 	if (job && ring->funcs->emit_frame_cntl)
288 		amdgpu_ring_emit_frame_cntl(ring, false, secure);
289 
290 	amdgpu_device_invalidate_hdp(adev, ring);
291 	/* Skip the IB for guilty contexts */
292 	af->skip_ib_dw_end_offset =
293 		amdgpu_ring_get_dw_distance(ring, af->ib_wptr, ring->wptr);
294 
295 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
296 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
297 
298 	/* wrap the last IB with fence */
299 	if (job && job->uf_addr) {
300 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
301 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
302 	}
303 
304 	if (ring->funcs->emit_gfx_shadow && ring->funcs->init_cond_exec &&
305 	    adev->gfx.cp_gfx_shadow) {
306 		amdgpu_ring_emit_gfx_shadow(ring, 0, 0, 0, false, 0);
307 		amdgpu_ring_init_cond_exec(ring, ring->cond_exe_gpu_addr);
308 	}
309 
310 	amdgpu_fence_emit(ring, af, fence_flags);
311 	*f = &af->base;
312 	/* get a ref for the job */
313 	if (job)
314 		dma_fence_get(*f);
315 
316 	if (ring->funcs->insert_end)
317 		ring->funcs->insert_end(ring);
318 
319 	amdgpu_ring_patch_cond_exec(ring, cond_exec);
320 
321 	ring->current_ctx = fence_ctx;
322 	if (job && ring->funcs->emit_switch_buffer)
323 		amdgpu_ring_emit_switch_buffer(ring);
324 
325 	if (ring->funcs->emit_wave_limit &&
326 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
327 		ring->funcs->emit_wave_limit(ring, false);
328 
329 	amdgpu_ring_ib_end(ring);
330 
331 	af->ib_dw_size = amdgpu_ring_get_dw_distance(ring, af->ib_wptr, ring->wptr);
332 
333 	amdgpu_ring_commit(ring);
334 
335 	return 0;
336 
337 free_fence:
338 	if (!job)
339 		kfree(af);
340 	return r;
341 }
342 
343 /**
344  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
345  *
346  * @adev: amdgpu_device pointer
347  *
348  * Initialize the suballocator to manage a pool of memory
349  * for use as IBs (all asics).
350  * Returns 0 on success, error on failure.
351  */
352 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
353 {
354 	const int sizes[AMDGPU_IB_POOL_MAX] = {
355 		[AMDGPU_IB_POOL_DELAYED] = SZ_1M,
356 		[AMDGPU_IB_POOL_IMMEDIATE] = SZ_128K,
357 		[AMDGPU_IB_POOL_DIRECT] = SZ_512K
358 	};
359 	const gfp_t gfp_flags[AMDGPU_IB_POOL_MAX] = {
360 		/*
361 		 * For normal page table updates and recoverable retry faults
362 		 * (for SVM), further restricted by the VM eviction lock to not
363 		 * wait for memory reclaim.
364 		 */
365 		[AMDGPU_IB_POOL_DELAYED] = GFP_KERNEL,
366 		/*
367 		 * For redirecting unrecoverable retry faults to the dummy page
368 		 * or set the PRT bits. dma_fence submissions might depend on
369 		 * that so we need the emmergency reserves.
370 		 */
371 		[AMDGPU_IB_POOL_IMMEDIATE] = GFP_ATOMIC,
372 		/*
373 		 * For IB tests during GPU resets. Only very small and temporary
374 		 * allocation to allow dma_fences to signal.
375 		 */
376 		[AMDGPU_IB_POOL_DIRECT] = GFP_ATOMIC
377 	};
378 	int r, i;
379 
380 	if (adev->ib_pool_ready)
381 		return 0;
382 
383 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
384 		r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
385 					      sizes[i], gfp_flags[i]);
386 		if (r)
387 			goto error;
388 	}
389 	adev->ib_pool_ready = true;
390 
391 	return 0;
392 
393 error:
394 	while (i--)
395 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
396 	return r;
397 }
398 
399 /**
400  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
401  *
402  * @adev: amdgpu_device pointer
403  *
404  * Tear down the suballocator managing the pool of memory
405  * for use as IBs (all asics).
406  */
407 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
408 {
409 	int i;
410 
411 	if (!adev->ib_pool_ready)
412 		return;
413 
414 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
415 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
416 	adev->ib_pool_ready = false;
417 }
418 
419 /**
420  * amdgpu_ib_pool_gfp_flags - Returns the gfp flags to use for each pool
421  * @adev: amdgpu device pointer
422  * @type: the IB pool type
423  */
424 gfp_t amdgpu_ib_pool_gfp_flags(struct amdgpu_device *adev,
425 			       enum amdgpu_ib_pool_type type)
426 {
427 	return adev->ib_pools[type].gfp_flags;
428 }
429 
430 /**
431  * amdgpu_ib_ring_tests - test IBs on the rings
432  *
433  * @adev: amdgpu_device pointer
434  *
435  * Test an IB (Indirect Buffer) on each ring.
436  * If the test fails, disable the ring.
437  * Returns 0 on success, error if the primary GFX ring
438  * IB test fails.
439  */
440 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
441 {
442 	long tmo_gfx, tmo_mm;
443 	int r, ret = 0;
444 	unsigned int i;
445 
446 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
447 	if (amdgpu_sriov_vf(adev)) {
448 		/* for MM engines in hypervisor side they are not scheduled together
449 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
450 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
451 		 * under SR-IOV should be set to a long time. 8 sec should be enough
452 		 * for the MM comes back to this VF.
453 		 */
454 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
455 	}
456 
457 	if (amdgpu_sriov_runtime(adev)) {
458 		/* for CP & SDMA engines since they are scheduled together so
459 		 * need to make the timeout width enough to cover the time
460 		 * cost waiting for it coming back under RUNTIME only
461 		 */
462 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
463 	} else if (adev->gmc.xgmi.hive_id) {
464 		tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
465 	}
466 
467 	for (i = 0; i < adev->num_rings; ++i) {
468 		struct amdgpu_ring *ring = adev->rings[i];
469 		long tmo;
470 
471 		/* KIQ rings don't have an IB test because we never submit IBs
472 		 * to them and they have no interrupt support.
473 		 */
474 		if (!ring->sched.ready || !ring->funcs->test_ib)
475 			continue;
476 
477 		if (adev->enable_mes &&
478 		    ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
479 			continue;
480 
481 		/* MM engine need more time */
482 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
483 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
484 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
485 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
486 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
487 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
488 			tmo = tmo_mm;
489 		else
490 			tmo = tmo_gfx;
491 
492 		r = amdgpu_ring_test_ib(ring, tmo);
493 		if (!r) {
494 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
495 				      ring->name);
496 			continue;
497 		}
498 
499 		ring->sched.ready = false;
500 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
501 			  ring->name, r);
502 
503 		if (ring == &adev->gfx.gfx_ring[0]) {
504 			/* oh, oh, that's really bad */
505 			adev->accel_working = false;
506 			return r;
507 
508 		} else {
509 			ret = r;
510 		}
511 	}
512 	return ret;
513 }
514 
515 /*
516  * Debugfs info
517  */
518 #if defined(CONFIG_DEBUG_FS)
519 
520 static int amdgpu_debugfs_sa_info_show(struct seq_file *m, void *unused)
521 {
522 	struct amdgpu_device *adev = m->private;
523 
524 	seq_puts(m, "--------------------- DELAYED ---------------------\n");
525 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
526 				     m);
527 	seq_puts(m, "-------------------- IMMEDIATE --------------------\n");
528 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
529 				     m);
530 	seq_puts(m, "--------------------- DIRECT ----------------------\n");
531 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
532 
533 	return 0;
534 }
535 
536 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_sa_info);
537 
538 #endif
539 
540 void amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
541 {
542 #if defined(CONFIG_DEBUG_FS)
543 	struct drm_minor *minor = adev_to_drm(adev)->primary;
544 	struct dentry *root = minor->debugfs_root;
545 
546 	debugfs_create_file("amdgpu_sa_info", 0444, root, adev,
547 			    &amdgpu_debugfs_sa_info_fops);
548 
549 #endif
550 }
551