xref: /openbsd-src/sys/dev/pci/drm/amd/amdgpu/amdgpu_gfx.c (revision ff0e7be1ebbcc809ea8ad2b6dafe215824da9e46)
1 /*
2  * Copyright 2014 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  */
25 
26 #include <linux/firmware.h>
27 #include "amdgpu.h"
28 #include "amdgpu_gfx.h"
29 #include "amdgpu_rlc.h"
30 #include "amdgpu_ras.h"
31 
32 /* delay 0.1 second to enable gfx off feature */
33 #define GFX_OFF_DELAY_ENABLE         msecs_to_jiffies(100)
34 
35 #define GFX_OFF_NO_DELAY 0
36 
37 /*
38  * GPU GFX IP block helpers function.
39  */
40 
41 int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
42 				int pipe, int queue)
43 {
44 	int bit = 0;
45 
46 	bit += mec * adev->gfx.mec.num_pipe_per_mec
47 		* adev->gfx.mec.num_queue_per_pipe;
48 	bit += pipe * adev->gfx.mec.num_queue_per_pipe;
49 	bit += queue;
50 
51 	return bit;
52 }
53 
54 void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
55 				 int *mec, int *pipe, int *queue)
56 {
57 	*queue = bit % adev->gfx.mec.num_queue_per_pipe;
58 	*pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
59 		% adev->gfx.mec.num_pipe_per_mec;
60 	*mec = (bit / adev->gfx.mec.num_queue_per_pipe)
61 	       / adev->gfx.mec.num_pipe_per_mec;
62 
63 }
64 
65 bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
66 				     int mec, int pipe, int queue)
67 {
68 	return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
69 			adev->gfx.mec.queue_bitmap);
70 }
71 
72 int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
73 			       int me, int pipe, int queue)
74 {
75 	int bit = 0;
76 
77 	bit += me * adev->gfx.me.num_pipe_per_me
78 		* adev->gfx.me.num_queue_per_pipe;
79 	bit += pipe * adev->gfx.me.num_queue_per_pipe;
80 	bit += queue;
81 
82 	return bit;
83 }
84 
85 void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
86 				int *me, int *pipe, int *queue)
87 {
88 	*queue = bit % adev->gfx.me.num_queue_per_pipe;
89 	*pipe = (bit / adev->gfx.me.num_queue_per_pipe)
90 		% adev->gfx.me.num_pipe_per_me;
91 	*me = (bit / adev->gfx.me.num_queue_per_pipe)
92 		/ adev->gfx.me.num_pipe_per_me;
93 }
94 
95 bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
96 				    int me, int pipe, int queue)
97 {
98 	return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
99 			adev->gfx.me.queue_bitmap);
100 }
101 
102 /**
103  * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
104  *
105  * @mask: array in which the per-shader array disable masks will be stored
106  * @max_se: number of SEs
107  * @max_sh: number of SHs
108  *
109  * The bitmask of CUs to be disabled in the shader array determined by se and
110  * sh is stored in mask[se * max_sh + sh].
111  */
112 void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
113 {
114 	unsigned se, sh, cu;
115 	const char *p;
116 
117 	memset(mask, 0, sizeof(*mask) * max_se * max_sh);
118 
119 	if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
120 		return;
121 
122 #ifdef notyet
123 	p = amdgpu_disable_cu;
124 	for (;;) {
125 		char *next;
126 		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
127 		if (ret < 3) {
128 			DRM_ERROR("amdgpu: could not parse disable_cu\n");
129 			return;
130 		}
131 
132 		if (se < max_se && sh < max_sh && cu < 16) {
133 			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
134 			mask[se * max_sh + sh] |= 1u << cu;
135 		} else {
136 			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
137 				  se, sh, cu);
138 		}
139 
140 		next = strchr(p, ',');
141 		if (!next)
142 			break;
143 		p = next + 1;
144 	}
145 #endif
146 }
147 
148 static bool amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device *adev)
149 {
150 	return amdgpu_async_gfx_ring && adev->gfx.me.num_pipe_per_me > 1;
151 }
152 
153 static bool amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device *adev)
154 {
155 	if (amdgpu_compute_multipipe != -1) {
156 		DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
157 			 amdgpu_compute_multipipe);
158 		return amdgpu_compute_multipipe == 1;
159 	}
160 
161 	if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))
162 		return true;
163 
164 	/* FIXME: spreading the queues across pipes causes perf regressions
165 	 * on POLARIS11 compute workloads */
166 	if (adev->asic_type == CHIP_POLARIS11)
167 		return false;
168 
169 	return adev->gfx.mec.num_mec > 1;
170 }
171 
172 bool amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device *adev,
173 						struct amdgpu_ring *ring)
174 {
175 	int queue = ring->queue;
176 	int pipe = ring->pipe;
177 
178 	/* Policy: use pipe1 queue0 as high priority graphics queue if we
179 	 * have more than one gfx pipe.
180 	 */
181 	if (amdgpu_gfx_is_graphics_multipipe_capable(adev) &&
182 	    adev->gfx.num_gfx_rings > 1 && pipe == 1 && queue == 0) {
183 		int me = ring->me;
184 		int bit;
185 
186 		bit = amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue);
187 		if (ring == &adev->gfx.gfx_ring[bit])
188 			return true;
189 	}
190 
191 	return false;
192 }
193 
194 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
195 					       struct amdgpu_ring *ring)
196 {
197 	/* Policy: use 1st queue as high priority compute queue if we
198 	 * have more than one compute queue.
199 	 */
200 	if (adev->gfx.num_compute_rings > 1 &&
201 	    ring == &adev->gfx.compute_ring[0])
202 		return true;
203 
204 	return false;
205 }
206 
207 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
208 {
209 	int i, queue, pipe;
210 	bool multipipe_policy = amdgpu_gfx_is_compute_multipipe_capable(adev);
211 	int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
212 				     adev->gfx.mec.num_queue_per_pipe,
213 				     adev->gfx.num_compute_rings);
214 
215 	if (multipipe_policy) {
216 		/* policy: make queues evenly cross all pipes on MEC1 only */
217 		for (i = 0; i < max_queues_per_mec; i++) {
218 			pipe = i % adev->gfx.mec.num_pipe_per_mec;
219 			queue = (i / adev->gfx.mec.num_pipe_per_mec) %
220 				adev->gfx.mec.num_queue_per_pipe;
221 
222 			set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
223 					adev->gfx.mec.queue_bitmap);
224 		}
225 	} else {
226 		/* policy: amdgpu owns all queues in the given pipe */
227 		for (i = 0; i < max_queues_per_mec; ++i)
228 			set_bit(i, adev->gfx.mec.queue_bitmap);
229 	}
230 
231 	dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
232 }
233 
234 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
235 {
236 	int i, queue, pipe;
237 	bool multipipe_policy = amdgpu_gfx_is_graphics_multipipe_capable(adev);
238 	int max_queues_per_me = adev->gfx.me.num_pipe_per_me *
239 					adev->gfx.me.num_queue_per_pipe;
240 
241 	if (multipipe_policy) {
242 		/* policy: amdgpu owns the first queue per pipe at this stage
243 		 * will extend to mulitple queues per pipe later */
244 		for (i = 0; i < max_queues_per_me; i++) {
245 			pipe = i % adev->gfx.me.num_pipe_per_me;
246 			queue = (i / adev->gfx.me.num_pipe_per_me) %
247 				adev->gfx.me.num_queue_per_pipe;
248 
249 			set_bit(pipe * adev->gfx.me.num_queue_per_pipe + queue,
250 				adev->gfx.me.queue_bitmap);
251 		}
252 	} else {
253 		for (i = 0; i < max_queues_per_me; ++i)
254 			set_bit(i, adev->gfx.me.queue_bitmap);
255 	}
256 
257 	/* update the number of active graphics rings */
258 	adev->gfx.num_gfx_rings =
259 		bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
260 }
261 
262 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
263 				  struct amdgpu_ring *ring)
264 {
265 	int queue_bit;
266 	int mec, pipe, queue;
267 
268 	queue_bit = adev->gfx.mec.num_mec
269 		    * adev->gfx.mec.num_pipe_per_mec
270 		    * adev->gfx.mec.num_queue_per_pipe;
271 
272 	while (--queue_bit >= 0) {
273 		if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
274 			continue;
275 
276 		amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
277 
278 		/*
279 		 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
280 		 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
281 		 * only can be issued on queue 0.
282 		 */
283 		if ((mec == 1 && pipe > 1) || queue != 0)
284 			continue;
285 
286 		ring->me = mec + 1;
287 		ring->pipe = pipe;
288 		ring->queue = queue;
289 
290 		return 0;
291 	}
292 
293 	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
294 	return -EINVAL;
295 }
296 
297 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
298 			     struct amdgpu_ring *ring,
299 			     struct amdgpu_irq_src *irq)
300 {
301 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
302 	int r = 0;
303 
304 	mtx_init(&kiq->ring_lock, IPL_TTY);
305 
306 	ring->adev = NULL;
307 	ring->ring_obj = NULL;
308 	ring->use_doorbell = true;
309 	ring->doorbell_index = adev->doorbell_index.kiq;
310 
311 	r = amdgpu_gfx_kiq_acquire(adev, ring);
312 	if (r)
313 		return r;
314 
315 	ring->eop_gpu_addr = kiq->eop_gpu_addr;
316 	ring->no_scheduler = true;
317 	snprintf(ring->name, sizeof(ring->name), "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
318 	r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
319 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
320 	if (r)
321 		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
322 
323 	return r;
324 }
325 
326 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
327 {
328 	amdgpu_ring_fini(ring);
329 }
330 
331 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
332 {
333 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
334 
335 	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
336 }
337 
338 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
339 			unsigned hpd_size)
340 {
341 	int r;
342 	u32 *hpd;
343 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
344 
345 	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
346 				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
347 				    &kiq->eop_gpu_addr, (void **)&hpd);
348 	if (r) {
349 		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
350 		return r;
351 	}
352 
353 	memset(hpd, 0, hpd_size);
354 
355 	r = amdgpu_bo_reserve(kiq->eop_obj, true);
356 	if (unlikely(r != 0))
357 		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
358 	amdgpu_bo_kunmap(kiq->eop_obj);
359 	amdgpu_bo_unreserve(kiq->eop_obj);
360 
361 	return 0;
362 }
363 
364 /* create MQD for each compute/gfx queue */
365 int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
366 			   unsigned mqd_size)
367 {
368 	struct amdgpu_ring *ring = NULL;
369 	int r, i;
370 
371 	/* create MQD for KIQ */
372 	ring = &adev->gfx.kiq.ring;
373 	if (!adev->enable_mes_kiq && !ring->mqd_obj) {
374 		/* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
375 		 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
376 		 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
377 		 * KIQ MQD no matter SRIOV or Bare-metal
378 		 */
379 		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
380 					    AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj,
381 					    &ring->mqd_gpu_addr, &ring->mqd_ptr);
382 		if (r) {
383 			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
384 			return r;
385 		}
386 
387 		/* prepare MQD backup */
388 		adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
389 		if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
390 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
391 	}
392 
393 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
394 		/* create MQD for each KGQ */
395 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
396 			ring = &adev->gfx.gfx_ring[i];
397 			if (!ring->mqd_obj) {
398 				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
399 							    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
400 							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
401 				if (r) {
402 					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
403 					return r;
404 				}
405 
406 				/* prepare MQD backup */
407 				adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
408 				if (!adev->gfx.me.mqd_backup[i])
409 					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
410 			}
411 		}
412 	}
413 
414 	/* create MQD for each KCQ */
415 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
416 		ring = &adev->gfx.compute_ring[i];
417 		if (!ring->mqd_obj) {
418 			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
419 						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
420 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
421 			if (r) {
422 				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
423 				return r;
424 			}
425 
426 			/* prepare MQD backup */
427 			adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
428 			if (!adev->gfx.mec.mqd_backup[i])
429 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
430 		}
431 	}
432 
433 	return 0;
434 }
435 
436 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev)
437 {
438 	struct amdgpu_ring *ring = NULL;
439 	int i;
440 
441 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
442 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
443 			ring = &adev->gfx.gfx_ring[i];
444 			kfree(adev->gfx.me.mqd_backup[i]);
445 			amdgpu_bo_free_kernel(&ring->mqd_obj,
446 					      &ring->mqd_gpu_addr,
447 					      &ring->mqd_ptr);
448 		}
449 	}
450 
451 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
452 		ring = &adev->gfx.compute_ring[i];
453 		kfree(adev->gfx.mec.mqd_backup[i]);
454 		amdgpu_bo_free_kernel(&ring->mqd_obj,
455 				      &ring->mqd_gpu_addr,
456 				      &ring->mqd_ptr);
457 	}
458 
459 	ring = &adev->gfx.kiq.ring;
460 	kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
461 	amdgpu_bo_free_kernel(&ring->mqd_obj,
462 			      &ring->mqd_gpu_addr,
463 			      &ring->mqd_ptr);
464 }
465 
466 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev)
467 {
468 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
469 	struct amdgpu_ring *kiq_ring = &kiq->ring;
470 	int i, r = 0;
471 
472 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
473 		return -EINVAL;
474 
475 	spin_lock(&adev->gfx.kiq.ring_lock);
476 	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
477 					adev->gfx.num_compute_rings)) {
478 		spin_unlock(&adev->gfx.kiq.ring_lock);
479 		return -ENOMEM;
480 	}
481 
482 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
483 		kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i],
484 					   RESET_QUEUES, 0, 0);
485 
486 	if (adev->gfx.kiq.ring.sched.ready && !adev->job_hang)
487 		r = amdgpu_ring_test_helper(kiq_ring);
488 	spin_unlock(&adev->gfx.kiq.ring_lock);
489 
490 	return r;
491 }
492 
493 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
494 					int queue_bit)
495 {
496 	int mec, pipe, queue;
497 	int set_resource_bit = 0;
498 
499 	amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
500 
501 	set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
502 
503 	return set_resource_bit;
504 }
505 
506 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev)
507 {
508 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
509 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
510 	uint64_t queue_mask = 0;
511 	int r, i;
512 
513 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
514 		return -EINVAL;
515 
516 	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
517 		if (!test_bit(i, adev->gfx.mec.queue_bitmap))
518 			continue;
519 
520 		/* This situation may be hit in the future if a new HW
521 		 * generation exposes more than 64 queues. If so, the
522 		 * definition of queue_mask needs updating */
523 		if (WARN_ON(i > (sizeof(queue_mask)*8))) {
524 			DRM_ERROR("Invalid KCQ enabled: %d\n", i);
525 			break;
526 		}
527 
528 		queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
529 	}
530 
531 	DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
532 							kiq_ring->queue);
533 	spin_lock(&adev->gfx.kiq.ring_lock);
534 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
535 					adev->gfx.num_compute_rings +
536 					kiq->pmf->set_resources_size);
537 	if (r) {
538 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
539 		spin_unlock(&adev->gfx.kiq.ring_lock);
540 		return r;
541 	}
542 
543 	if (adev->enable_mes)
544 		queue_mask = ~0ULL;
545 
546 	kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
547 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
548 		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
549 
550 	r = amdgpu_ring_test_helper(kiq_ring);
551 	spin_unlock(&adev->gfx.kiq.ring_lock);
552 	if (r)
553 		DRM_ERROR("KCQ enable failed\n");
554 
555 	return r;
556 }
557 
558 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
559  *
560  * @adev: amdgpu_device pointer
561  * @bool enable true: enable gfx off feature, false: disable gfx off feature
562  *
563  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
564  * 2. other client can send request to disable gfx off feature, the request should be honored.
565  * 3. other client can cancel their request of disable gfx off feature
566  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
567  */
568 
569 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
570 {
571 	unsigned long delay = GFX_OFF_DELAY_ENABLE;
572 
573 	if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
574 		return;
575 
576 	mutex_lock(&adev->gfx.gfx_off_mutex);
577 
578 	if (enable) {
579 		/* If the count is already 0, it means there's an imbalance bug somewhere.
580 		 * Note that the bug may be in a different caller than the one which triggers the
581 		 * WARN_ON_ONCE.
582 		 */
583 		if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
584 			goto unlock;
585 
586 		adev->gfx.gfx_off_req_count--;
587 
588 		if (adev->gfx.gfx_off_req_count == 0 &&
589 		    !adev->gfx.gfx_off_state) {
590 			/* If going to s2idle, no need to wait */
591 			if (adev->in_s0ix) {
592 				if (!amdgpu_dpm_set_powergating_by_smu(adev,
593 						AMD_IP_BLOCK_TYPE_GFX, true))
594 					adev->gfx.gfx_off_state = true;
595 			} else {
596 				schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
597 					      delay);
598 			}
599 		}
600 	} else {
601 		if (adev->gfx.gfx_off_req_count == 0) {
602 			cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
603 
604 			if (adev->gfx.gfx_off_state &&
605 			    !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
606 				adev->gfx.gfx_off_state = false;
607 
608 				if (adev->gfx.funcs->init_spm_golden) {
609 					dev_dbg(adev->dev,
610 						"GFXOFF is disabled, re-init SPM golden settings\n");
611 					amdgpu_gfx_init_spm_golden(adev);
612 				}
613 			}
614 		}
615 
616 		adev->gfx.gfx_off_req_count++;
617 	}
618 
619 unlock:
620 	mutex_unlock(&adev->gfx.gfx_off_mutex);
621 }
622 
623 int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
624 {
625 	int r = 0;
626 
627 	mutex_lock(&adev->gfx.gfx_off_mutex);
628 
629 	r = amdgpu_dpm_set_residency_gfxoff(adev, value);
630 
631 	mutex_unlock(&adev->gfx.gfx_off_mutex);
632 
633 	return r;
634 }
635 
636 int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
637 {
638 	int r = 0;
639 
640 	mutex_lock(&adev->gfx.gfx_off_mutex);
641 
642 	r = amdgpu_dpm_get_residency_gfxoff(adev, value);
643 
644 	mutex_unlock(&adev->gfx.gfx_off_mutex);
645 
646 	return r;
647 }
648 
649 int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
650 {
651 	int r = 0;
652 
653 	mutex_lock(&adev->gfx.gfx_off_mutex);
654 
655 	r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
656 
657 	mutex_unlock(&adev->gfx.gfx_off_mutex);
658 
659 	return r;
660 }
661 
662 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
663 {
664 
665 	int r = 0;
666 
667 	mutex_lock(&adev->gfx.gfx_off_mutex);
668 
669 	r = amdgpu_dpm_get_status_gfxoff(adev, value);
670 
671 	mutex_unlock(&adev->gfx.gfx_off_mutex);
672 
673 	return r;
674 }
675 
676 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
677 {
678 	int r;
679 
680 	if (amdgpu_ras_is_supported(adev, ras_block->block)) {
681 		if (!amdgpu_persistent_edc_harvesting_supported(adev))
682 			amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
683 
684 		r = amdgpu_ras_block_late_init(adev, ras_block);
685 		if (r)
686 			return r;
687 
688 		r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
689 		if (r)
690 			goto late_fini;
691 	} else {
692 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
693 	}
694 
695 	return 0;
696 late_fini:
697 	amdgpu_ras_block_late_fini(adev, ras_block);
698 	return r;
699 }
700 
701 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
702 		void *err_data,
703 		struct amdgpu_iv_entry *entry)
704 {
705 	/* TODO ue will trigger an interrupt.
706 	 *
707 	 * When “Full RAS” is enabled, the per-IP interrupt sources should
708 	 * be disabled and the driver should only look for the aggregated
709 	 * interrupt via sync flood
710 	 */
711 	if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
712 		kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
713 		if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
714 		    adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
715 			adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
716 		amdgpu_ras_reset_gpu(adev);
717 	}
718 	return AMDGPU_RAS_SUCCESS;
719 }
720 
721 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
722 				  struct amdgpu_irq_src *source,
723 				  struct amdgpu_iv_entry *entry)
724 {
725 	struct ras_common_if *ras_if = adev->gfx.ras_if;
726 	struct ras_dispatch_if ih_data = {
727 		.entry = entry,
728 	};
729 
730 	if (!ras_if)
731 		return 0;
732 
733 	ih_data.head = *ras_if;
734 
735 	DRM_ERROR("CP ECC ERROR IRQ\n");
736 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
737 	return 0;
738 }
739 
740 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
741 {
742 	signed long r, cnt = 0;
743 	unsigned long flags;
744 	uint32_t seq, reg_val_offs = 0, value = 0;
745 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
746 	struct amdgpu_ring *ring = &kiq->ring;
747 
748 	if (amdgpu_device_skip_hw_access(adev))
749 		return 0;
750 
751 	if (adev->mes.ring.sched.ready)
752 		return amdgpu_mes_rreg(adev, reg);
753 
754 	BUG_ON(!ring->funcs->emit_rreg);
755 
756 	spin_lock_irqsave(&kiq->ring_lock, flags);
757 	if (amdgpu_device_wb_get(adev, &reg_val_offs)) {
758 		pr_err("critical bug! too many kiq readers\n");
759 		goto failed_unlock;
760 	}
761 	amdgpu_ring_alloc(ring, 32);
762 	amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
763 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
764 	if (r)
765 		goto failed_undo;
766 
767 	amdgpu_ring_commit(ring);
768 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
769 
770 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
771 
772 	/* don't wait anymore for gpu reset case because this way may
773 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
774 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
775 	 * never return if we keep waiting in virt_kiq_rreg, which cause
776 	 * gpu_recover() hang there.
777 	 *
778 	 * also don't wait anymore for IRQ context
779 	 * */
780 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
781 		goto failed_kiq_read;
782 
783 	might_sleep();
784 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
785 		drm_msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
786 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
787 	}
788 
789 	if (cnt > MAX_KIQ_REG_TRY)
790 		goto failed_kiq_read;
791 
792 	mb();
793 	value = adev->wb.wb[reg_val_offs];
794 	amdgpu_device_wb_free(adev, reg_val_offs);
795 	return value;
796 
797 failed_undo:
798 	amdgpu_ring_undo(ring);
799 failed_unlock:
800 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
801 failed_kiq_read:
802 	if (reg_val_offs)
803 		amdgpu_device_wb_free(adev, reg_val_offs);
804 	dev_err(adev->dev, "failed to read reg:%x\n", reg);
805 	return ~0;
806 }
807 
808 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
809 {
810 	signed long r, cnt = 0;
811 	unsigned long flags;
812 	uint32_t seq;
813 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
814 	struct amdgpu_ring *ring = &kiq->ring;
815 
816 	BUG_ON(!ring->funcs->emit_wreg);
817 
818 	if (amdgpu_device_skip_hw_access(adev))
819 		return;
820 
821 	if (adev->mes.ring.sched.ready) {
822 		amdgpu_mes_wreg(adev, reg, v);
823 		return;
824 	}
825 
826 	spin_lock_irqsave(&kiq->ring_lock, flags);
827 	amdgpu_ring_alloc(ring, 32);
828 	amdgpu_ring_emit_wreg(ring, reg, v);
829 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
830 	if (r)
831 		goto failed_undo;
832 
833 	amdgpu_ring_commit(ring);
834 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
835 
836 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
837 
838 	/* don't wait anymore for gpu reset case because this way may
839 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
840 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
841 	 * never return if we keep waiting in virt_kiq_rreg, which cause
842 	 * gpu_recover() hang there.
843 	 *
844 	 * also don't wait anymore for IRQ context
845 	 * */
846 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
847 		goto failed_kiq_write;
848 
849 	might_sleep();
850 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
851 
852 		drm_msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
853 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
854 	}
855 
856 	if (cnt > MAX_KIQ_REG_TRY)
857 		goto failed_kiq_write;
858 
859 	return;
860 
861 failed_undo:
862 	amdgpu_ring_undo(ring);
863 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
864 failed_kiq_write:
865 	dev_err(adev->dev, "failed to write reg:%x\n", reg);
866 }
867 
868 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
869 {
870 	if (amdgpu_num_kcq == -1) {
871 		return 8;
872 	} else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
873 		dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
874 		return 8;
875 	}
876 	return amdgpu_num_kcq;
877 }
878 
879 void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
880 				  uint32_t ucode_id)
881 {
882 	const struct gfx_firmware_header_v1_0 *cp_hdr;
883 	const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
884 	struct amdgpu_firmware_info *info = NULL;
885 	const struct firmware *ucode_fw;
886 	unsigned int fw_size;
887 
888 	switch (ucode_id) {
889 	case AMDGPU_UCODE_ID_CP_PFP:
890 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
891 			adev->gfx.pfp_fw->data;
892 		adev->gfx.pfp_fw_version =
893 			le32_to_cpu(cp_hdr->header.ucode_version);
894 		adev->gfx.pfp_feature_version =
895 			le32_to_cpu(cp_hdr->ucode_feature_version);
896 		ucode_fw = adev->gfx.pfp_fw;
897 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
898 		break;
899 	case AMDGPU_UCODE_ID_CP_RS64_PFP:
900 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
901 			adev->gfx.pfp_fw->data;
902 		adev->gfx.pfp_fw_version =
903 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
904 		adev->gfx.pfp_feature_version =
905 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
906 		ucode_fw = adev->gfx.pfp_fw;
907 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
908 		break;
909 	case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
910 	case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
911 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
912 			adev->gfx.pfp_fw->data;
913 		ucode_fw = adev->gfx.pfp_fw;
914 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
915 		break;
916 	case AMDGPU_UCODE_ID_CP_ME:
917 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
918 			adev->gfx.me_fw->data;
919 		adev->gfx.me_fw_version =
920 			le32_to_cpu(cp_hdr->header.ucode_version);
921 		adev->gfx.me_feature_version =
922 			le32_to_cpu(cp_hdr->ucode_feature_version);
923 		ucode_fw = adev->gfx.me_fw;
924 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
925 		break;
926 	case AMDGPU_UCODE_ID_CP_RS64_ME:
927 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
928 			adev->gfx.me_fw->data;
929 		adev->gfx.me_fw_version =
930 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
931 		adev->gfx.me_feature_version =
932 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
933 		ucode_fw = adev->gfx.me_fw;
934 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
935 		break;
936 	case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
937 	case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
938 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
939 			adev->gfx.me_fw->data;
940 		ucode_fw = adev->gfx.me_fw;
941 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
942 		break;
943 	case AMDGPU_UCODE_ID_CP_CE:
944 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
945 			adev->gfx.ce_fw->data;
946 		adev->gfx.ce_fw_version =
947 			le32_to_cpu(cp_hdr->header.ucode_version);
948 		adev->gfx.ce_feature_version =
949 			le32_to_cpu(cp_hdr->ucode_feature_version);
950 		ucode_fw = adev->gfx.ce_fw;
951 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
952 		break;
953 	case AMDGPU_UCODE_ID_CP_MEC1:
954 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
955 			adev->gfx.mec_fw->data;
956 		adev->gfx.mec_fw_version =
957 			le32_to_cpu(cp_hdr->header.ucode_version);
958 		adev->gfx.mec_feature_version =
959 			le32_to_cpu(cp_hdr->ucode_feature_version);
960 		ucode_fw = adev->gfx.mec_fw;
961 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
962 			  le32_to_cpu(cp_hdr->jt_size) * 4;
963 		break;
964 	case AMDGPU_UCODE_ID_CP_MEC1_JT:
965 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
966 			adev->gfx.mec_fw->data;
967 		ucode_fw = adev->gfx.mec_fw;
968 		fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
969 		break;
970 	case AMDGPU_UCODE_ID_CP_MEC2:
971 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
972 			adev->gfx.mec2_fw->data;
973 		adev->gfx.mec2_fw_version =
974 			le32_to_cpu(cp_hdr->header.ucode_version);
975 		adev->gfx.mec2_feature_version =
976 			le32_to_cpu(cp_hdr->ucode_feature_version);
977 		ucode_fw = adev->gfx.mec2_fw;
978 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
979 			  le32_to_cpu(cp_hdr->jt_size) * 4;
980 		break;
981 	case AMDGPU_UCODE_ID_CP_MEC2_JT:
982 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
983 			adev->gfx.mec2_fw->data;
984 		ucode_fw = adev->gfx.mec2_fw;
985 		fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
986 		break;
987 	case AMDGPU_UCODE_ID_CP_RS64_MEC:
988 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
989 			adev->gfx.mec_fw->data;
990 		adev->gfx.mec_fw_version =
991 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
992 		adev->gfx.mec_feature_version =
993 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
994 		ucode_fw = adev->gfx.mec_fw;
995 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
996 		break;
997 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
998 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
999 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
1000 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
1001 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1002 			adev->gfx.mec_fw->data;
1003 		ucode_fw = adev->gfx.mec_fw;
1004 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1005 		break;
1006 	default:
1007 		break;
1008 	}
1009 
1010 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1011 		info = &adev->firmware.ucode[ucode_id];
1012 		info->ucode_id = ucode_id;
1013 		info->fw = ucode_fw;
1014 		adev->firmware.fw_size += roundup2(fw_size, PAGE_SIZE);
1015 	}
1016 }
1017