xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c (revision ab93e0dd72c37d378dd936f031ffb83ff2bd87ce)
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->scheduled.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 = job->base.s_fence ? job->base.s_fence->finished.context : 0;
163 	} else {
164 		vm = NULL;
165 		fence_ctx = 0;
166 		shadow_va = 0;
167 		csa_va = 0;
168 		gds_va = 0;
169 		init_shadow = false;
170 		af = NULL;
171 	}
172 
173 	if (!ring->sched.ready) {
174 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
175 		return -EINVAL;
176 	}
177 
178 	if (vm && !job->vmid) {
179 		dev_err(adev->dev, "VM IB without ID\n");
180 		return -EINVAL;
181 	}
182 
183 	if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
184 	    (!ring->funcs->secure_submission_supported)) {
185 		dev_err(adev->dev, "secure submissions not supported on ring <%s>\n", ring->name);
186 		return -EINVAL;
187 	}
188 
189 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
190 		ring->funcs->emit_ib_size;
191 
192 	r = amdgpu_ring_alloc(ring, alloc_size);
193 	if (r) {
194 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
195 		return r;
196 	}
197 
198 	need_ctx_switch = ring->current_ctx != fence_ctx;
199 	if (ring->funcs->emit_pipeline_sync && job &&
200 	    ((tmp = amdgpu_sync_get_fence(&job->explicit_sync)) ||
201 	     need_ctx_switch || amdgpu_vm_need_pipeline_sync(ring, job))) {
202 
203 		need_pipe_sync = true;
204 
205 		if (tmp)
206 			trace_amdgpu_ib_pipe_sync(job, tmp);
207 
208 		dma_fence_put(tmp);
209 	}
210 
211 	if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
212 		ring->funcs->emit_mem_sync(ring);
213 
214 	if (ring->funcs->emit_wave_limit &&
215 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
216 		ring->funcs->emit_wave_limit(ring, true);
217 
218 	if (ring->funcs->insert_start)
219 		ring->funcs->insert_start(ring);
220 
221 	if (job) {
222 		r = amdgpu_vm_flush(ring, job, need_pipe_sync);
223 		if (r) {
224 			amdgpu_ring_undo(ring);
225 			return r;
226 		}
227 	}
228 
229 	amdgpu_ring_ib_begin(ring);
230 
231 	if (ring->funcs->emit_gfx_shadow)
232 		amdgpu_ring_emit_gfx_shadow(ring, shadow_va, csa_va, gds_va,
233 					    init_shadow, vmid);
234 
235 	if (ring->funcs->init_cond_exec)
236 		cond_exec = amdgpu_ring_init_cond_exec(ring,
237 						       ring->cond_exe_gpu_addr);
238 
239 	amdgpu_device_flush_hdp(adev, ring);
240 
241 	if (need_ctx_switch)
242 		status |= AMDGPU_HAVE_CTX_SWITCH;
243 
244 	if (job && ring->funcs->emit_cntxcntl) {
245 		status |= job->preamble_status;
246 		status |= job->preemption_status;
247 		amdgpu_ring_emit_cntxcntl(ring, status);
248 	}
249 
250 	/* Setup initial TMZiness and send it off.
251 	 */
252 	secure = false;
253 	if (job && ring->funcs->emit_frame_cntl) {
254 		secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
255 		amdgpu_ring_emit_frame_cntl(ring, true, secure);
256 	}
257 
258 	for (i = 0; i < num_ibs; ++i) {
259 		ib = &ibs[i];
260 
261 		if (job && ring->funcs->emit_frame_cntl) {
262 			if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) {
263 				amdgpu_ring_emit_frame_cntl(ring, false, secure);
264 				secure = !secure;
265 				amdgpu_ring_emit_frame_cntl(ring, true, secure);
266 			}
267 		}
268 
269 		amdgpu_ring_emit_ib(ring, job, ib, status);
270 		status &= ~AMDGPU_HAVE_CTX_SWITCH;
271 	}
272 
273 	if (job && ring->funcs->emit_frame_cntl)
274 		amdgpu_ring_emit_frame_cntl(ring, false, secure);
275 
276 	amdgpu_device_invalidate_hdp(adev, ring);
277 
278 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
279 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
280 
281 	/* wrap the last IB with fence */
282 	if (job && job->uf_addr) {
283 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
284 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
285 	}
286 
287 	if (ring->funcs->emit_gfx_shadow && ring->funcs->init_cond_exec) {
288 		amdgpu_ring_emit_gfx_shadow(ring, 0, 0, 0, false, 0);
289 		amdgpu_ring_init_cond_exec(ring, ring->cond_exe_gpu_addr);
290 	}
291 
292 	r = amdgpu_fence_emit(ring, f, af, fence_flags);
293 	if (r) {
294 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
295 		if (job && job->vmid)
296 			amdgpu_vmid_reset(adev, ring->vm_hub, job->vmid);
297 		amdgpu_ring_undo(ring);
298 		return r;
299 	}
300 
301 	if (ring->funcs->insert_end)
302 		ring->funcs->insert_end(ring);
303 
304 	amdgpu_ring_patch_cond_exec(ring, cond_exec);
305 
306 	ring->current_ctx = fence_ctx;
307 	if (job && ring->funcs->emit_switch_buffer)
308 		amdgpu_ring_emit_switch_buffer(ring);
309 
310 	if (ring->funcs->emit_wave_limit &&
311 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
312 		ring->funcs->emit_wave_limit(ring, false);
313 
314 	/* Save the wptr associated with this fence.
315 	 * This must be last for resets to work properly
316 	 * as we need to save the wptr associated with this
317 	 * fence so we know what rings contents to backup
318 	 * after we reset the queue.
319 	 */
320 	amdgpu_fence_save_wptr(*f);
321 
322 	amdgpu_ring_ib_end(ring);
323 	amdgpu_ring_commit(ring);
324 
325 	return 0;
326 }
327 
328 /**
329  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
330  *
331  * @adev: amdgpu_device pointer
332  *
333  * Initialize the suballocator to manage a pool of memory
334  * for use as IBs (all asics).
335  * Returns 0 on success, error on failure.
336  */
amdgpu_ib_pool_init(struct amdgpu_device * adev)337 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
338 {
339 	int r, i;
340 
341 	if (adev->ib_pool_ready)
342 		return 0;
343 
344 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
345 		r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
346 					      AMDGPU_IB_POOL_SIZE, 256,
347 					      AMDGPU_GEM_DOMAIN_GTT);
348 		if (r)
349 			goto error;
350 	}
351 	adev->ib_pool_ready = true;
352 
353 	return 0;
354 
355 error:
356 	while (i--)
357 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
358 	return r;
359 }
360 
361 /**
362  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
363  *
364  * @adev: amdgpu_device pointer
365  *
366  * Tear down the suballocator managing the pool of memory
367  * for use as IBs (all asics).
368  */
amdgpu_ib_pool_fini(struct amdgpu_device * adev)369 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
370 {
371 	int i;
372 
373 	if (!adev->ib_pool_ready)
374 		return;
375 
376 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
377 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
378 	adev->ib_pool_ready = false;
379 }
380 
381 /**
382  * amdgpu_ib_ring_tests - test IBs on the rings
383  *
384  * @adev: amdgpu_device pointer
385  *
386  * Test an IB (Indirect Buffer) on each ring.
387  * If the test fails, disable the ring.
388  * Returns 0 on success, error if the primary GFX ring
389  * IB test fails.
390  */
amdgpu_ib_ring_tests(struct amdgpu_device * adev)391 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
392 {
393 	long tmo_gfx, tmo_mm;
394 	int r, ret = 0;
395 	unsigned int i;
396 
397 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
398 	if (amdgpu_sriov_vf(adev)) {
399 		/* for MM engines in hypervisor side they are not scheduled together
400 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
401 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
402 		 * under SR-IOV should be set to a long time. 8 sec should be enough
403 		 * for the MM comes back to this VF.
404 		 */
405 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
406 	}
407 
408 	if (amdgpu_sriov_runtime(adev)) {
409 		/* for CP & SDMA engines since they are scheduled together so
410 		 * need to make the timeout width enough to cover the time
411 		 * cost waiting for it coming back under RUNTIME only
412 		 */
413 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
414 	} else if (adev->gmc.xgmi.hive_id) {
415 		tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
416 	}
417 
418 	for (i = 0; i < adev->num_rings; ++i) {
419 		struct amdgpu_ring *ring = adev->rings[i];
420 		long tmo;
421 
422 		/* KIQ rings don't have an IB test because we never submit IBs
423 		 * to them and they have no interrupt support.
424 		 */
425 		if (!ring->sched.ready || !ring->funcs->test_ib)
426 			continue;
427 
428 		if (adev->enable_mes &&
429 		    ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
430 			continue;
431 
432 		/* MM engine need more time */
433 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
434 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
435 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
436 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
437 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
438 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
439 			tmo = tmo_mm;
440 		else
441 			tmo = tmo_gfx;
442 
443 		r = amdgpu_ring_test_ib(ring, tmo);
444 		if (!r) {
445 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
446 				      ring->name);
447 			continue;
448 		}
449 
450 		ring->sched.ready = false;
451 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
452 			  ring->name, r);
453 
454 		if (ring == &adev->gfx.gfx_ring[0]) {
455 			/* oh, oh, that's really bad */
456 			adev->accel_working = false;
457 			return r;
458 
459 		} else {
460 			ret = r;
461 		}
462 	}
463 	return ret;
464 }
465 
466 /*
467  * Debugfs info
468  */
469 #if defined(CONFIG_DEBUG_FS)
470 
amdgpu_debugfs_sa_info_show(struct seq_file * m,void * unused)471 static int amdgpu_debugfs_sa_info_show(struct seq_file *m, void *unused)
472 {
473 	struct amdgpu_device *adev = m->private;
474 
475 	seq_puts(m, "--------------------- DELAYED ---------------------\n");
476 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
477 				     m);
478 	seq_puts(m, "-------------------- IMMEDIATE --------------------\n");
479 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
480 				     m);
481 	seq_puts(m, "--------------------- DIRECT ----------------------\n");
482 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
483 
484 	return 0;
485 }
486 
487 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_sa_info);
488 
489 #endif
490 
amdgpu_debugfs_sa_init(struct amdgpu_device * adev)491 void amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
492 {
493 #if defined(CONFIG_DEBUG_FS)
494 	struct drm_minor *minor = adev_to_drm(adev)->primary;
495 	struct dentry *root = minor->debugfs_root;
496 
497 	debugfs_create_file("amdgpu_sa_info", 0444, root, adev,
498 			    &amdgpu_debugfs_sa_info_fops);
499 
500 #endif
501 }
502