xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c (revision cb8fef4b07f71da01843457b66e579ad248a625e)
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 <drm/amdgpu_drm.h>
31 
32 #include "amdgpu.h"
33 #include "atom.h"
34 #include "amdgpu_trace.h"
35 
36 #define AMDGPU_IB_TEST_TIMEOUT	msecs_to_jiffies(1000)
37 #define AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT	msecs_to_jiffies(2000)
38 
39 /*
40  * IB
41  * IBs (Indirect Buffers) and areas of GPU accessible memory where
42  * commands are stored.  You can put a pointer to the IB in the
43  * command ring and the hw will fetch the commands from the IB
44  * and execute them.  Generally userspace acceleration drivers
45  * produce command buffers which are send to the kernel and
46  * put in IBs for execution by the requested ring.
47  */
48 
49 /**
50  * amdgpu_ib_get - request an IB (Indirect Buffer)
51  *
52  * @adev: amdgpu_device pointer
53  * @vm: amdgpu_vm pointer
54  * @size: requested IB size
55  * @pool_type: IB pool type (delayed, immediate, direct)
56  * @ib: IB object returned
57  *
58  * Request an IB (all asics).  IBs are allocated using the
59  * suballocator.
60  * Returns 0 on success, error on failure.
61  */
62 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
63 		  unsigned int size, enum amdgpu_ib_pool_type pool_type,
64 		  struct amdgpu_ib *ib)
65 {
66 	int r;
67 
68 	if (size) {
69 		r = amdgpu_sa_bo_new(&adev->ib_pools[pool_type],
70 				     &ib->sa_bo, size);
71 		if (r) {
72 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
73 			return r;
74 		}
75 
76 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
77 		/* flush the cache before commit the IB */
78 		ib->flags = AMDGPU_IB_FLAG_EMIT_MEM_SYNC;
79 
80 		if (!vm)
81 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
82 	}
83 
84 	return 0;
85 }
86 
87 /**
88  * amdgpu_ib_free - free an IB (Indirect Buffer)
89  *
90  * @ib: IB object to free
91  * @f: the fence SA bo need wait on for the ib alloation
92  *
93  * Free an IB (all asics).
94  */
95 void amdgpu_ib_free(struct amdgpu_ib *ib, struct dma_fence *f)
96 {
97 	amdgpu_sa_bo_free(&ib->sa_bo, f);
98 }
99 
100 /**
101  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
102  *
103  * @ring: ring structure the IB is associated with
104  * @job: job to schedule
105  * @f: fence referenced during this submission
106  *
107  * Schedule an IB on the associated ring (all asics).
108  * Returns 0 on success, error on failure.
109  *
110  * On SI, there are two parallel engines fed from the primary ring,
111  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
112  * resource descriptors have moved to memory, the CE allows you to
113  * prime the caches while the DE is updating register state so that
114  * the resource descriptors will be already in cache when the draw is
115  * processed.  To accomplish this, the userspace driver submits two
116  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
117  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
118  * to SI there was just a DE IB.
119  */
120 int amdgpu_ib_schedule(struct amdgpu_ring *ring, struct amdgpu_job *job,
121 		       struct dma_fence **f)
122 {
123 	struct amdgpu_device *adev = ring->adev;
124 	struct amdgpu_ib *ib;
125 	struct dma_fence *tmp = NULL;
126 	struct amdgpu_fence *af;
127 	struct amdgpu_fence *vm_af;
128 	bool need_ctx_switch;
129 	bool emit_spm_needed = false;
130 	bool emit_gds_needed = false;
131 	uint64_t fence_ctx;
132 	uint32_t status = 0, alloc_size;
133 	unsigned int fence_flags = 0;
134 	bool secure;
135 	unsigned int vmid;
136 	bool need_pipe_sync = false;
137 	unsigned int cond_exec;
138 	unsigned int i;
139 	int r;
140 
141 	if (!job || !job->num_ibs)
142 		return -EINVAL;
143 
144 	vmid = AMDGPU_JOB_GET_VMID(job);
145 	ib = &job->ibs[0];
146 	fence_ctx = job->base.s_fence ?
147 		job->base.s_fence->finished.context : 0;
148 	af = job->hw_fence;
149 	vm_af = job->hw_vm_fence;
150 	/* Save the context of the job for reset handling.
151 	 * The driver needs this so it can skip the ring
152 	 * contents for guilty contexts.
153 	 */
154 	af->context = fence_ctx;
155 	/* the vm fence is also part of the job's context */
156 	vm_af->context = fence_ctx;
157 
158 	if (!ring->sched.ready) {
159 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
160 		return -EINVAL;
161 	}
162 
163 	if (job->vm && !job->vmid) {
164 		dev_err(adev->dev, "VM IB without ID\n");
165 		return -EINVAL;
166 	}
167 
168 	if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
169 	    (!ring->funcs->secure_submission_supported)) {
170 		dev_err(adev->dev, "secure submissions not supported on ring <%s>\n", ring->name);
171 		return -EINVAL;
172 	}
173 
174 	alloc_size = ring->funcs->emit_frame_size + job->num_ibs *
175 		ring->funcs->emit_ib_size;
176 
177 	r = amdgpu_ring_alloc(ring, alloc_size);
178 	if (r) {
179 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
180 		return r;
181 	}
182 
183 	need_ctx_switch = ring->current_ctx != fence_ctx;
184 	if (ring->funcs->emit_pipeline_sync &&
185 	    ((tmp = amdgpu_sync_get_fence(&job->explicit_sync)) ||
186 	     need_ctx_switch || amdgpu_vm_need_pipeline_sync(ring, job))) {
187 
188 		need_pipe_sync = true;
189 
190 		if (tmp)
191 			trace_amdgpu_ib_pipe_sync(job, tmp);
192 
193 		dma_fence_put(tmp);
194 	}
195 
196 	/* VM sequence */
197 	vm_af->ib_wptr = ring->wptr;
198 	amdgpu_vm_flush(ring, job, &need_pipe_sync, &emit_spm_needed,
199 			&emit_gds_needed);
200 	vm_af->ib_dw_size =
201 		amdgpu_ring_get_dw_distance(ring, vm_af->ib_wptr, ring->wptr);
202 
203 	/* IB sequence */
204 	af->ib_wptr = ring->wptr;
205 	amdgpu_ring_ib_begin(ring);
206 
207 	if (ring->funcs->insert_start)
208 		ring->funcs->insert_start(ring);
209 
210 	/* this may have been handled by amdgpu_vm_flush */
211 	if (need_pipe_sync)
212 		amdgpu_ring_emit_pipeline_sync(ring);
213 
214 	if (emit_spm_needed)
215 		adev->gfx.rlc.funcs->update_spm_vmid(adev, ring->xcc_id, ring, job->vmid);
216 
217 	if (emit_gds_needed)
218 		amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
219 					    job->gds_size, job->gws_base,
220 					    job->gws_size, job->oa_base,
221 					    job->oa_size);
222 
223 	if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
224 		ring->funcs->emit_mem_sync(ring);
225 
226 	if (ring->funcs->emit_wave_limit &&
227 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
228 		ring->funcs->emit_wave_limit(ring, true);
229 
230 	if (ring->funcs->emit_gfx_shadow && adev->gfx.cp_gfx_shadow)
231 		amdgpu_ring_emit_gfx_shadow(ring, job->shadow_va, job->csa_va, job->gds_va,
232 					    job->init_shadow, vmid);
233 
234 	if (ring->funcs->init_cond_exec)
235 		cond_exec = amdgpu_ring_init_cond_exec(ring,
236 						       ring->cond_exe_gpu_addr);
237 
238 	/* Skip the IB for guilty contexts */
239 	af->skip_ib_dw_start_offset =
240 		amdgpu_ring_get_dw_distance(ring, af->ib_wptr, ring->wptr);
241 	amdgpu_device_flush_hdp(adev, ring);
242 
243 	if (need_ctx_switch)
244 		status |= AMDGPU_HAVE_CTX_SWITCH;
245 
246 	if (ring->funcs->emit_cntxcntl) {
247 		status |= job->preamble_status;
248 		status |= job->preemption_status;
249 		amdgpu_ring_emit_cntxcntl(ring, status);
250 	}
251 
252 	/* Setup initial TMZiness and send it off.
253 	 */
254 	secure = false;
255 	if (ring->funcs->emit_frame_cntl) {
256 		secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
257 		amdgpu_ring_emit_frame_cntl(ring, true, secure);
258 	}
259 
260 	for (i = 0; i < job->num_ibs; ++i) {
261 		ib = &job->ibs[i];
262 
263 		if (ring->funcs->emit_frame_cntl) {
264 			if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) {
265 				amdgpu_ring_emit_frame_cntl(ring, false, secure);
266 				secure = !secure;
267 				amdgpu_ring_emit_frame_cntl(ring, true, secure);
268 			}
269 		}
270 
271 		amdgpu_ring_emit_ib(ring, job, ib, status);
272 		status &= ~AMDGPU_HAVE_CTX_SWITCH;
273 	}
274 
275 	if (ring->funcs->emit_frame_cntl)
276 		amdgpu_ring_emit_frame_cntl(ring, false, secure);
277 
278 	amdgpu_device_invalidate_hdp(adev, ring);
279 	/* Skip the IB for guilty contexts */
280 	af->skip_ib_dw_end_offset =
281 		amdgpu_ring_get_dw_distance(ring, af->ib_wptr, ring->wptr);
282 
283 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
284 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
285 
286 	/* wrap the last IB with fence */
287 	if (job->uf_addr) {
288 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
289 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
290 	}
291 
292 	if (ring->funcs->emit_gfx_shadow && ring->funcs->init_cond_exec &&
293 	    adev->gfx.cp_gfx_shadow) {
294 		amdgpu_ring_emit_gfx_shadow(ring, 0, 0, 0, false, 0);
295 		amdgpu_ring_init_cond_exec(ring, ring->cond_exe_gpu_addr);
296 	}
297 
298 	amdgpu_fence_emit(ring, af, fence_flags);
299 	*f = &af->base;
300 	/* get a ref for the job */
301 	dma_fence_get(*f);
302 
303 	if (ring->funcs->insert_end)
304 		ring->funcs->insert_end(ring);
305 
306 	amdgpu_ring_patch_cond_exec(ring, cond_exec);
307 
308 	ring->current_ctx = fence_ctx;
309 	if (ring->funcs->emit_switch_buffer)
310 		amdgpu_ring_emit_switch_buffer(ring);
311 
312 	if (ring->funcs->emit_wave_limit &&
313 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
314 		ring->funcs->emit_wave_limit(ring, false);
315 
316 	amdgpu_ring_ib_end(ring);
317 
318 	af->ib_dw_size = amdgpu_ring_get_dw_distance(ring, af->ib_wptr, ring->wptr);
319 
320 	amdgpu_ring_commit(ring);
321 
322 	return 0;
323 }
324 
325 /**
326  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
327  *
328  * @adev: amdgpu_device pointer
329  *
330  * Initialize the suballocator to manage a pool of memory
331  * for use as IBs (all asics).
332  * Returns 0 on success, error on failure.
333  */
334 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
335 {
336 	const int sizes[AMDGPU_IB_POOL_MAX] = {
337 		[AMDGPU_IB_POOL_DELAYED] = SZ_1M,
338 		[AMDGPU_IB_POOL_IMMEDIATE] = SZ_128K,
339 		[AMDGPU_IB_POOL_DIRECT] = SZ_512K
340 	};
341 	const gfp_t gfp_flags[AMDGPU_IB_POOL_MAX] = {
342 		/*
343 		 * For normal page table updates and recoverable retry faults
344 		 * (for SVM), further restricted by the VM eviction lock to not
345 		 * wait for memory reclaim.
346 		 */
347 		[AMDGPU_IB_POOL_DELAYED] = GFP_KERNEL,
348 		/*
349 		 * For redirecting unrecoverable retry faults to the dummy page
350 		 * or set the PRT bits. dma_fence submissions might depend on
351 		 * that so we need the emmergency reserves.
352 		 */
353 		[AMDGPU_IB_POOL_IMMEDIATE] = GFP_ATOMIC,
354 		/*
355 		 * For IB tests during GPU resets. Only very small and temporary
356 		 * allocation to allow dma_fences to signal.
357 		 */
358 		[AMDGPU_IB_POOL_DIRECT] = GFP_ATOMIC
359 	};
360 	int r, i;
361 
362 	if (adev->ib_pool_ready)
363 		return 0;
364 
365 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
366 		r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
367 					      sizes[i], gfp_flags[i]);
368 		if (r)
369 			goto error;
370 	}
371 	adev->ib_pool_ready = true;
372 
373 	return 0;
374 
375 error:
376 	while (i--)
377 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
378 	return r;
379 }
380 
381 /**
382  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
383  *
384  * @adev: amdgpu_device pointer
385  *
386  * Tear down the suballocator managing the pool of memory
387  * for use as IBs (all asics).
388  */
389 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
390 {
391 	int i;
392 
393 	if (!adev->ib_pool_ready)
394 		return;
395 
396 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
397 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
398 	adev->ib_pool_ready = false;
399 }
400 
401 /**
402  * amdgpu_ib_pool_gfp_flags - Returns the gfp flags to use for each pool
403  * @adev: amdgpu device pointer
404  * @type: the IB pool type
405  */
406 gfp_t amdgpu_ib_pool_gfp_flags(struct amdgpu_device *adev,
407 			       enum amdgpu_ib_pool_type type)
408 {
409 	return adev->ib_pools[type].gfp_flags;
410 }
411 
412 /**
413  * amdgpu_ib_ring_tests - test IBs on the rings
414  *
415  * @adev: amdgpu_device pointer
416  *
417  * Test an IB (Indirect Buffer) on each ring.
418  * If the test fails, disable the ring.
419  * Returns 0 on success, error if the primary GFX ring
420  * IB test fails.
421  */
422 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
423 {
424 	long tmo_gfx, tmo_mm;
425 	int r, ret = 0;
426 	unsigned int i;
427 
428 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
429 	if (amdgpu_sriov_vf(adev)) {
430 		/* for MM engines in hypervisor side they are not scheduled together
431 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
432 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
433 		 * under SR-IOV should be set to a long time. 8 sec should be enough
434 		 * for the MM comes back to this VF.
435 		 */
436 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
437 	}
438 
439 	if (amdgpu_sriov_runtime(adev)) {
440 		/* for CP & SDMA engines since they are scheduled together so
441 		 * need to make the timeout width enough to cover the time
442 		 * cost waiting for it coming back under RUNTIME only
443 		 */
444 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
445 	} else if (adev->gmc.xgmi.hive_id) {
446 		tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
447 	}
448 
449 	for (i = 0; i < adev->num_rings; ++i) {
450 		struct amdgpu_ring *ring = adev->rings[i];
451 		long tmo;
452 
453 		/* KIQ rings don't have an IB test because we never submit IBs
454 		 * to them and they have no interrupt support.
455 		 */
456 		if (!ring->sched.ready || !ring->funcs->test_ib)
457 			continue;
458 
459 		if (adev->enable_mes &&
460 		    ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
461 			continue;
462 
463 		/* MM engine need more time */
464 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
465 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
466 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
467 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
468 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
469 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
470 			tmo = tmo_mm;
471 		else
472 			tmo = tmo_gfx;
473 
474 		r = amdgpu_ring_test_ib(ring, tmo);
475 		if (!r) {
476 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
477 				      ring->name);
478 			continue;
479 		}
480 
481 		ring->sched.ready = false;
482 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
483 			  ring->name, r);
484 
485 		if (ring == &adev->gfx.gfx_ring[0]) {
486 			/* oh, oh, that's really bad */
487 			adev->accel_working = false;
488 			return r;
489 
490 		} else {
491 			ret = r;
492 		}
493 	}
494 	return ret;
495 }
496 
497 /*
498  * Debugfs info
499  */
500 #if defined(CONFIG_DEBUG_FS)
501 
502 static int amdgpu_debugfs_sa_info_show(struct seq_file *m, void *unused)
503 {
504 	struct amdgpu_device *adev = m->private;
505 
506 	seq_puts(m, "--------------------- DELAYED ---------------------\n");
507 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
508 				     m);
509 	seq_puts(m, "-------------------- IMMEDIATE --------------------\n");
510 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
511 				     m);
512 	seq_puts(m, "--------------------- DIRECT ----------------------\n");
513 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
514 
515 	return 0;
516 }
517 
518 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_sa_info);
519 
520 #endif
521 
522 void amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
523 {
524 #if defined(CONFIG_DEBUG_FS)
525 	struct drm_minor *minor = adev_to_drm(adev)->primary;
526 	struct dentry *root = minor->debugfs_root;
527 
528 	debugfs_create_file("amdgpu_sa_info", 0444, root, adev,
529 			    &amdgpu_debugfs_sa_info_fops);
530 
531 #endif
532 }
533