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 	p = amdgpu_disable_cu;
123 	for (;;) {
124 		char *next;
125 		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
126 		if (ret < 3) {
127 			DRM_ERROR("amdgpu: could not parse disable_cu\n");
128 			return;
129 		}
130 
131 		if (se < max_se && sh < max_sh && cu < 16) {
132 			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
133 			mask[se * max_sh + sh] |= 1u << cu;
134 		} else {
135 			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
136 				  se, sh, cu);
137 		}
138 
139 		next = strchr(p, ',');
140 		if (!next)
141 			break;
142 		p = next + 1;
143 	}
144 }
145 
146 static bool amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device *adev)
147 {
148 	return amdgpu_async_gfx_ring && adev->gfx.me.num_pipe_per_me > 1;
149 }
150 
151 static bool amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device *adev)
152 {
153 	if (amdgpu_compute_multipipe != -1) {
154 		DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
155 			 amdgpu_compute_multipipe);
156 		return amdgpu_compute_multipipe == 1;
157 	}
158 
159 	if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))
160 		return true;
161 
162 	/* FIXME: spreading the queues across pipes causes perf regressions
163 	 * on POLARIS11 compute workloads */
164 	if (adev->asic_type == CHIP_POLARIS11)
165 		return false;
166 
167 	return adev->gfx.mec.num_mec > 1;
168 }
169 
170 bool amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device *adev,
171 						struct amdgpu_ring *ring)
172 {
173 	int queue = ring->queue;
174 	int pipe = ring->pipe;
175 
176 	/* Policy: use pipe1 queue0 as high priority graphics queue if we
177 	 * have more than one gfx pipe.
178 	 */
179 	if (amdgpu_gfx_is_graphics_multipipe_capable(adev) &&
180 	    adev->gfx.num_gfx_rings > 1 && pipe == 1 && queue == 0) {
181 		int me = ring->me;
182 		int bit;
183 
184 		bit = amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue);
185 		if (ring == &adev->gfx.gfx_ring[bit])
186 			return true;
187 	}
188 
189 	return false;
190 }
191 
192 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
193 					       struct amdgpu_ring *ring)
194 {
195 	/* Policy: use 1st queue as high priority compute queue if we
196 	 * have more than one compute queue.
197 	 */
198 	if (adev->gfx.num_compute_rings > 1 &&
199 	    ring == &adev->gfx.compute_ring[0])
200 		return true;
201 
202 	return false;
203 }
204 
205 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
206 {
207 	int i, queue, pipe;
208 	bool multipipe_policy = amdgpu_gfx_is_compute_multipipe_capable(adev);
209 	int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
210 				     adev->gfx.mec.num_queue_per_pipe,
211 				     adev->gfx.num_compute_rings);
212 
213 	if (multipipe_policy) {
214 		/* policy: make queues evenly cross all pipes on MEC1 only */
215 		for (i = 0; i < max_queues_per_mec; i++) {
216 			pipe = i % adev->gfx.mec.num_pipe_per_mec;
217 			queue = (i / adev->gfx.mec.num_pipe_per_mec) %
218 				adev->gfx.mec.num_queue_per_pipe;
219 
220 			set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
221 					adev->gfx.mec.queue_bitmap);
222 		}
223 	} else {
224 		/* policy: amdgpu owns all queues in the given pipe */
225 		for (i = 0; i < max_queues_per_mec; ++i)
226 			set_bit(i, adev->gfx.mec.queue_bitmap);
227 	}
228 
229 	dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
230 }
231 
232 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
233 {
234 	int i, queue, pipe;
235 	bool multipipe_policy = amdgpu_gfx_is_graphics_multipipe_capable(adev);
236 	int max_queues_per_me = adev->gfx.me.num_pipe_per_me *
237 					adev->gfx.me.num_queue_per_pipe;
238 
239 	if (multipipe_policy) {
240 		/* policy: amdgpu owns the first queue per pipe at this stage
241 		 * will extend to mulitple queues per pipe later */
242 		for (i = 0; i < max_queues_per_me; i++) {
243 			pipe = i % adev->gfx.me.num_pipe_per_me;
244 			queue = (i / adev->gfx.me.num_pipe_per_me) %
245 				adev->gfx.me.num_queue_per_pipe;
246 
247 			set_bit(pipe * adev->gfx.me.num_queue_per_pipe + queue,
248 				adev->gfx.me.queue_bitmap);
249 		}
250 	} else {
251 		for (i = 0; i < max_queues_per_me; ++i)
252 			set_bit(i, adev->gfx.me.queue_bitmap);
253 	}
254 
255 	/* update the number of active graphics rings */
256 	adev->gfx.num_gfx_rings =
257 		bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
258 }
259 
260 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
261 				  struct amdgpu_ring *ring)
262 {
263 	int queue_bit;
264 	int mec, pipe, queue;
265 
266 	queue_bit = adev->gfx.mec.num_mec
267 		    * adev->gfx.mec.num_pipe_per_mec
268 		    * adev->gfx.mec.num_queue_per_pipe;
269 
270 	while (--queue_bit >= 0) {
271 		if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
272 			continue;
273 
274 		amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
275 
276 		/*
277 		 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
278 		 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
279 		 * only can be issued on queue 0.
280 		 */
281 		if ((mec == 1 && pipe > 1) || queue != 0)
282 			continue;
283 
284 		ring->me = mec + 1;
285 		ring->pipe = pipe;
286 		ring->queue = queue;
287 
288 		return 0;
289 	}
290 
291 	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
292 	return -EINVAL;
293 }
294 
295 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
296 			     struct amdgpu_ring *ring,
297 			     struct amdgpu_irq_src *irq)
298 {
299 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
300 	int r = 0;
301 
302 	spin_lock_init(&kiq->ring_lock);
303 
304 	ring->adev = NULL;
305 	ring->ring_obj = NULL;
306 	ring->use_doorbell = true;
307 	ring->doorbell_index = adev->doorbell_index.kiq;
308 	ring->vm_hub = AMDGPU_GFXHUB_0;
309 
310 	r = amdgpu_gfx_kiq_acquire(adev, ring);
311 	if (r)
312 		return r;
313 
314 	ring->eop_gpu_addr = kiq->eop_gpu_addr;
315 	ring->no_scheduler = true;
316 	sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
317 	r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
318 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
319 	if (r)
320 		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
321 
322 	return r;
323 }
324 
325 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
326 {
327 	amdgpu_ring_fini(ring);
328 }
329 
330 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
331 {
332 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
333 
334 	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
335 }
336 
337 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
338 			unsigned hpd_size)
339 {
340 	int r;
341 	u32 *hpd;
342 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
343 
344 	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
345 				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
346 				    &kiq->eop_gpu_addr, (void **)&hpd);
347 	if (r) {
348 		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
349 		return r;
350 	}
351 
352 	memset(hpd, 0, hpd_size);
353 
354 	r = amdgpu_bo_reserve(kiq->eop_obj, true);
355 	if (unlikely(r != 0))
356 		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
357 	amdgpu_bo_kunmap(kiq->eop_obj);
358 	amdgpu_bo_unreserve(kiq->eop_obj);
359 
360 	return 0;
361 }
362 
363 /* create MQD for each compute/gfx queue */
364 int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
365 			   unsigned mqd_size)
366 {
367 	struct amdgpu_ring *ring = NULL;
368 	int r, i;
369 
370 	/* create MQD for KIQ */
371 	ring = &adev->gfx.kiq.ring;
372 	if (!adev->enable_mes_kiq && !ring->mqd_obj) {
373 		/* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
374 		 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
375 		 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
376 		 * KIQ MQD no matter SRIOV or Bare-metal
377 		 */
378 		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
379 					    AMDGPU_GEM_DOMAIN_VRAM |
380 					    AMDGPU_GEM_DOMAIN_GTT,
381 					    &ring->mqd_obj,
382 					    &ring->mqd_gpu_addr,
383 					    &ring->mqd_ptr);
384 		if (r) {
385 			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
386 			return r;
387 		}
388 
389 		/* prepare MQD backup */
390 		adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
391 		if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
392 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
393 	}
394 
395 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
396 		/* create MQD for each KGQ */
397 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
398 			ring = &adev->gfx.gfx_ring[i];
399 			if (!ring->mqd_obj) {
400 				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
401 							    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
402 							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
403 				if (r) {
404 					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
405 					return r;
406 				}
407 
408 				/* prepare MQD backup */
409 				adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
410 				if (!adev->gfx.me.mqd_backup[i])
411 					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
412 			}
413 		}
414 	}
415 
416 	/* create MQD for each KCQ */
417 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
418 		ring = &adev->gfx.compute_ring[i];
419 		if (!ring->mqd_obj) {
420 			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
421 						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
422 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
423 			if (r) {
424 				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
425 				return r;
426 			}
427 
428 			/* prepare MQD backup */
429 			adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
430 			if (!adev->gfx.mec.mqd_backup[i])
431 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
432 		}
433 	}
434 
435 	return 0;
436 }
437 
438 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev)
439 {
440 	struct amdgpu_ring *ring = NULL;
441 	int i;
442 
443 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
444 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
445 			ring = &adev->gfx.gfx_ring[i];
446 			kfree(adev->gfx.me.mqd_backup[i]);
447 			amdgpu_bo_free_kernel(&ring->mqd_obj,
448 					      &ring->mqd_gpu_addr,
449 					      &ring->mqd_ptr);
450 		}
451 	}
452 
453 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
454 		ring = &adev->gfx.compute_ring[i];
455 		kfree(adev->gfx.mec.mqd_backup[i]);
456 		amdgpu_bo_free_kernel(&ring->mqd_obj,
457 				      &ring->mqd_gpu_addr,
458 				      &ring->mqd_ptr);
459 	}
460 
461 	ring = &adev->gfx.kiq.ring;
462 	kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
463 	amdgpu_bo_free_kernel(&ring->mqd_obj,
464 			      &ring->mqd_gpu_addr,
465 			      &ring->mqd_ptr);
466 }
467 
468 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev)
469 {
470 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
471 	struct amdgpu_ring *kiq_ring = &kiq->ring;
472 	int i, r = 0;
473 
474 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
475 		return -EINVAL;
476 
477 	spin_lock(&adev->gfx.kiq.ring_lock);
478 	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
479 					adev->gfx.num_compute_rings)) {
480 		spin_unlock(&adev->gfx.kiq.ring_lock);
481 		return -ENOMEM;
482 	}
483 
484 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
485 		kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i],
486 					   RESET_QUEUES, 0, 0);
487 
488 	if (adev->gfx.kiq.ring.sched.ready && !adev->job_hang)
489 		r = amdgpu_ring_test_helper(kiq_ring);
490 	spin_unlock(&adev->gfx.kiq.ring_lock);
491 
492 	return r;
493 }
494 
495 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
496 					int queue_bit)
497 {
498 	int mec, pipe, queue;
499 	int set_resource_bit = 0;
500 
501 	amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
502 
503 	set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
504 
505 	return set_resource_bit;
506 }
507 
508 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev)
509 {
510 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
511 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
512 	uint64_t queue_mask = 0;
513 	int r, i;
514 
515 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
516 		return -EINVAL;
517 
518 	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
519 		if (!test_bit(i, adev->gfx.mec.queue_bitmap))
520 			continue;
521 
522 		/* This situation may be hit in the future if a new HW
523 		 * generation exposes more than 64 queues. If so, the
524 		 * definition of queue_mask needs updating */
525 		if (WARN_ON(i > (sizeof(queue_mask)*8))) {
526 			DRM_ERROR("Invalid KCQ enabled: %d\n", i);
527 			break;
528 		}
529 
530 		queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
531 	}
532 
533 	DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
534 							kiq_ring->queue);
535 	spin_lock(&adev->gfx.kiq.ring_lock);
536 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
537 					adev->gfx.num_compute_rings +
538 					kiq->pmf->set_resources_size);
539 	if (r) {
540 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
541 		spin_unlock(&adev->gfx.kiq.ring_lock);
542 		return r;
543 	}
544 
545 	if (adev->enable_mes)
546 		queue_mask = ~0ULL;
547 
548 	kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
549 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
550 		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
551 
552 	r = amdgpu_ring_test_helper(kiq_ring);
553 	spin_unlock(&adev->gfx.kiq.ring_lock);
554 	if (r)
555 		DRM_ERROR("KCQ enable failed\n");
556 
557 	return r;
558 }
559 
560 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
561  *
562  * @adev: amdgpu_device pointer
563  * @bool enable true: enable gfx off feature, false: disable gfx off feature
564  *
565  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
566  * 2. other client can send request to disable gfx off feature, the request should be honored.
567  * 3. other client can cancel their request of disable gfx off feature
568  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
569  */
570 
571 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
572 {
573 	unsigned long delay = GFX_OFF_DELAY_ENABLE;
574 
575 	if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
576 		return;
577 
578 	mutex_lock(&adev->gfx.gfx_off_mutex);
579 
580 	if (enable) {
581 		/* If the count is already 0, it means there's an imbalance bug somewhere.
582 		 * Note that the bug may be in a different caller than the one which triggers the
583 		 * WARN_ON_ONCE.
584 		 */
585 		if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
586 			goto unlock;
587 
588 		adev->gfx.gfx_off_req_count--;
589 
590 		if (adev->gfx.gfx_off_req_count == 0 &&
591 		    !adev->gfx.gfx_off_state) {
592 			/* If going to s2idle, no need to wait */
593 			if (adev->in_s0ix) {
594 				if (!amdgpu_dpm_set_powergating_by_smu(adev,
595 						AMD_IP_BLOCK_TYPE_GFX, true))
596 					adev->gfx.gfx_off_state = true;
597 			} else {
598 				schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
599 					      delay);
600 			}
601 		}
602 	} else {
603 		if (adev->gfx.gfx_off_req_count == 0) {
604 			cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
605 
606 			if (adev->gfx.gfx_off_state &&
607 			    !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
608 				adev->gfx.gfx_off_state = false;
609 
610 				if (adev->gfx.funcs->init_spm_golden) {
611 					dev_dbg(adev->dev,
612 						"GFXOFF is disabled, re-init SPM golden settings\n");
613 					amdgpu_gfx_init_spm_golden(adev);
614 				}
615 			}
616 		}
617 
618 		adev->gfx.gfx_off_req_count++;
619 	}
620 
621 unlock:
622 	mutex_unlock(&adev->gfx.gfx_off_mutex);
623 }
624 
625 int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
626 {
627 	int r = 0;
628 
629 	mutex_lock(&adev->gfx.gfx_off_mutex);
630 
631 	r = amdgpu_dpm_set_residency_gfxoff(adev, value);
632 
633 	mutex_unlock(&adev->gfx.gfx_off_mutex);
634 
635 	return r;
636 }
637 
638 int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
639 {
640 	int r = 0;
641 
642 	mutex_lock(&adev->gfx.gfx_off_mutex);
643 
644 	r = amdgpu_dpm_get_residency_gfxoff(adev, value);
645 
646 	mutex_unlock(&adev->gfx.gfx_off_mutex);
647 
648 	return r;
649 }
650 
651 int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
652 {
653 	int r = 0;
654 
655 	mutex_lock(&adev->gfx.gfx_off_mutex);
656 
657 	r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
658 
659 	mutex_unlock(&adev->gfx.gfx_off_mutex);
660 
661 	return r;
662 }
663 
664 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
665 {
666 
667 	int r = 0;
668 
669 	mutex_lock(&adev->gfx.gfx_off_mutex);
670 
671 	r = amdgpu_dpm_get_status_gfxoff(adev, value);
672 
673 	mutex_unlock(&adev->gfx.gfx_off_mutex);
674 
675 	return r;
676 }
677 
678 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
679 {
680 	int r;
681 
682 	if (amdgpu_ras_is_supported(adev, ras_block->block)) {
683 		if (!amdgpu_persistent_edc_harvesting_supported(adev))
684 			amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
685 
686 		r = amdgpu_ras_block_late_init(adev, ras_block);
687 		if (r)
688 			return r;
689 
690 		if (adev->gfx.cp_ecc_error_irq.funcs) {
691 			r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
692 			if (r)
693 				goto late_fini;
694 		}
695 	} else {
696 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
697 	}
698 
699 	return 0;
700 late_fini:
701 	amdgpu_ras_block_late_fini(adev, ras_block);
702 	return r;
703 }
704 
705 int amdgpu_gfx_ras_sw_init(struct amdgpu_device *adev)
706 {
707 	int err = 0;
708 	struct amdgpu_gfx_ras *ras = NULL;
709 
710 	/* adev->gfx.ras is NULL, which means gfx does not
711 	 * support ras function, then do nothing here.
712 	 */
713 	if (!adev->gfx.ras)
714 		return 0;
715 
716 	ras = adev->gfx.ras;
717 
718 	err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
719 	if (err) {
720 		dev_err(adev->dev, "Failed to register gfx ras block!\n");
721 		return err;
722 	}
723 
724 	strcpy(ras->ras_block.ras_comm.name, "gfx");
725 	ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__GFX;
726 	ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
727 	adev->gfx.ras_if = &ras->ras_block.ras_comm;
728 
729 	/* If not define special ras_late_init function, use gfx default ras_late_init */
730 	if (!ras->ras_block.ras_late_init)
731 		ras->ras_block.ras_late_init = amdgpu_gfx_ras_late_init;
732 
733 	/* If not defined special ras_cb function, use default ras_cb */
734 	if (!ras->ras_block.ras_cb)
735 		ras->ras_block.ras_cb = amdgpu_gfx_process_ras_data_cb;
736 
737 	return 0;
738 }
739 
740 int amdgpu_gfx_poison_consumption_handler(struct amdgpu_device *adev,
741 						struct amdgpu_iv_entry *entry)
742 {
743 	if (adev->gfx.ras && adev->gfx.ras->poison_consumption_handler)
744 		return adev->gfx.ras->poison_consumption_handler(adev, entry);
745 
746 	return 0;
747 }
748 
749 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
750 		void *err_data,
751 		struct amdgpu_iv_entry *entry)
752 {
753 	/* TODO ue will trigger an interrupt.
754 	 *
755 	 * When “Full RAS” is enabled, the per-IP interrupt sources should
756 	 * be disabled and the driver should only look for the aggregated
757 	 * interrupt via sync flood
758 	 */
759 	if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
760 		kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
761 		if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
762 		    adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
763 			adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
764 		amdgpu_ras_reset_gpu(adev);
765 	}
766 	return AMDGPU_RAS_SUCCESS;
767 }
768 
769 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
770 				  struct amdgpu_irq_src *source,
771 				  struct amdgpu_iv_entry *entry)
772 {
773 	struct ras_common_if *ras_if = adev->gfx.ras_if;
774 	struct ras_dispatch_if ih_data = {
775 		.entry = entry,
776 	};
777 
778 	if (!ras_if)
779 		return 0;
780 
781 	ih_data.head = *ras_if;
782 
783 	DRM_ERROR("CP ECC ERROR IRQ\n");
784 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
785 	return 0;
786 }
787 
788 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
789 {
790 	signed long r, cnt = 0;
791 	unsigned long flags;
792 	uint32_t seq, reg_val_offs = 0, value = 0;
793 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
794 	struct amdgpu_ring *ring = &kiq->ring;
795 
796 	if (amdgpu_device_skip_hw_access(adev))
797 		return 0;
798 
799 	if (adev->mes.ring.sched.ready)
800 		return amdgpu_mes_rreg(adev, reg);
801 
802 	BUG_ON(!ring->funcs->emit_rreg);
803 
804 	spin_lock_irqsave(&kiq->ring_lock, flags);
805 	if (amdgpu_device_wb_get(adev, &reg_val_offs)) {
806 		pr_err("critical bug! too many kiq readers\n");
807 		goto failed_unlock;
808 	}
809 	amdgpu_ring_alloc(ring, 32);
810 	amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
811 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
812 	if (r)
813 		goto failed_undo;
814 
815 	amdgpu_ring_commit(ring);
816 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
817 
818 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
819 
820 	/* don't wait anymore for gpu reset case because this way may
821 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
822 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
823 	 * never return if we keep waiting in virt_kiq_rreg, which cause
824 	 * gpu_recover() hang there.
825 	 *
826 	 * also don't wait anymore for IRQ context
827 	 * */
828 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
829 		goto failed_kiq_read;
830 
831 	might_sleep();
832 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
833 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
834 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
835 	}
836 
837 	if (cnt > MAX_KIQ_REG_TRY)
838 		goto failed_kiq_read;
839 
840 	mb();
841 	value = adev->wb.wb[reg_val_offs];
842 	amdgpu_device_wb_free(adev, reg_val_offs);
843 	return value;
844 
845 failed_undo:
846 	amdgpu_ring_undo(ring);
847 failed_unlock:
848 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
849 failed_kiq_read:
850 	if (reg_val_offs)
851 		amdgpu_device_wb_free(adev, reg_val_offs);
852 	dev_err(adev->dev, "failed to read reg:%x\n", reg);
853 	return ~0;
854 }
855 
856 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
857 {
858 	signed long r, cnt = 0;
859 	unsigned long flags;
860 	uint32_t seq;
861 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
862 	struct amdgpu_ring *ring = &kiq->ring;
863 
864 	BUG_ON(!ring->funcs->emit_wreg);
865 
866 	if (amdgpu_device_skip_hw_access(adev))
867 		return;
868 
869 	if (adev->mes.ring.sched.ready) {
870 		amdgpu_mes_wreg(adev, reg, v);
871 		return;
872 	}
873 
874 	spin_lock_irqsave(&kiq->ring_lock, flags);
875 	amdgpu_ring_alloc(ring, 32);
876 	amdgpu_ring_emit_wreg(ring, reg, v);
877 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
878 	if (r)
879 		goto failed_undo;
880 
881 	amdgpu_ring_commit(ring);
882 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
883 
884 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
885 
886 	/* don't wait anymore for gpu reset case because this way may
887 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
888 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
889 	 * never return if we keep waiting in virt_kiq_rreg, which cause
890 	 * gpu_recover() hang there.
891 	 *
892 	 * also don't wait anymore for IRQ context
893 	 * */
894 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
895 		goto failed_kiq_write;
896 
897 	might_sleep();
898 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
899 
900 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
901 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
902 	}
903 
904 	if (cnt > MAX_KIQ_REG_TRY)
905 		goto failed_kiq_write;
906 
907 	return;
908 
909 failed_undo:
910 	amdgpu_ring_undo(ring);
911 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
912 failed_kiq_write:
913 	dev_err(adev->dev, "failed to write reg:%x\n", reg);
914 }
915 
916 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
917 {
918 	if (amdgpu_num_kcq == -1) {
919 		return 8;
920 	} else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
921 		dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
922 		return 8;
923 	}
924 	return amdgpu_num_kcq;
925 }
926 
927 void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
928 				  uint32_t ucode_id)
929 {
930 	const struct gfx_firmware_header_v1_0 *cp_hdr;
931 	const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
932 	struct amdgpu_firmware_info *info = NULL;
933 	const struct firmware *ucode_fw;
934 	unsigned int fw_size;
935 
936 	switch (ucode_id) {
937 	case AMDGPU_UCODE_ID_CP_PFP:
938 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
939 			adev->gfx.pfp_fw->data;
940 		adev->gfx.pfp_fw_version =
941 			le32_to_cpu(cp_hdr->header.ucode_version);
942 		adev->gfx.pfp_feature_version =
943 			le32_to_cpu(cp_hdr->ucode_feature_version);
944 		ucode_fw = adev->gfx.pfp_fw;
945 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
946 		break;
947 	case AMDGPU_UCODE_ID_CP_RS64_PFP:
948 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
949 			adev->gfx.pfp_fw->data;
950 		adev->gfx.pfp_fw_version =
951 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
952 		adev->gfx.pfp_feature_version =
953 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
954 		ucode_fw = adev->gfx.pfp_fw;
955 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
956 		break;
957 	case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
958 	case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
959 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
960 			adev->gfx.pfp_fw->data;
961 		ucode_fw = adev->gfx.pfp_fw;
962 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
963 		break;
964 	case AMDGPU_UCODE_ID_CP_ME:
965 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
966 			adev->gfx.me_fw->data;
967 		adev->gfx.me_fw_version =
968 			le32_to_cpu(cp_hdr->header.ucode_version);
969 		adev->gfx.me_feature_version =
970 			le32_to_cpu(cp_hdr->ucode_feature_version);
971 		ucode_fw = adev->gfx.me_fw;
972 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
973 		break;
974 	case AMDGPU_UCODE_ID_CP_RS64_ME:
975 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
976 			adev->gfx.me_fw->data;
977 		adev->gfx.me_fw_version =
978 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
979 		adev->gfx.me_feature_version =
980 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
981 		ucode_fw = adev->gfx.me_fw;
982 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
983 		break;
984 	case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
985 	case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
986 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
987 			adev->gfx.me_fw->data;
988 		ucode_fw = adev->gfx.me_fw;
989 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
990 		break;
991 	case AMDGPU_UCODE_ID_CP_CE:
992 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
993 			adev->gfx.ce_fw->data;
994 		adev->gfx.ce_fw_version =
995 			le32_to_cpu(cp_hdr->header.ucode_version);
996 		adev->gfx.ce_feature_version =
997 			le32_to_cpu(cp_hdr->ucode_feature_version);
998 		ucode_fw = adev->gfx.ce_fw;
999 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1000 		break;
1001 	case AMDGPU_UCODE_ID_CP_MEC1:
1002 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1003 			adev->gfx.mec_fw->data;
1004 		adev->gfx.mec_fw_version =
1005 			le32_to_cpu(cp_hdr->header.ucode_version);
1006 		adev->gfx.mec_feature_version =
1007 			le32_to_cpu(cp_hdr->ucode_feature_version);
1008 		ucode_fw = adev->gfx.mec_fw;
1009 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1010 			  le32_to_cpu(cp_hdr->jt_size) * 4;
1011 		break;
1012 	case AMDGPU_UCODE_ID_CP_MEC1_JT:
1013 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1014 			adev->gfx.mec_fw->data;
1015 		ucode_fw = adev->gfx.mec_fw;
1016 		fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1017 		break;
1018 	case AMDGPU_UCODE_ID_CP_MEC2:
1019 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1020 			adev->gfx.mec2_fw->data;
1021 		adev->gfx.mec2_fw_version =
1022 			le32_to_cpu(cp_hdr->header.ucode_version);
1023 		adev->gfx.mec2_feature_version =
1024 			le32_to_cpu(cp_hdr->ucode_feature_version);
1025 		ucode_fw = adev->gfx.mec2_fw;
1026 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1027 			  le32_to_cpu(cp_hdr->jt_size) * 4;
1028 		break;
1029 	case AMDGPU_UCODE_ID_CP_MEC2_JT:
1030 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1031 			adev->gfx.mec2_fw->data;
1032 		ucode_fw = adev->gfx.mec2_fw;
1033 		fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1034 		break;
1035 	case AMDGPU_UCODE_ID_CP_RS64_MEC:
1036 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1037 			adev->gfx.mec_fw->data;
1038 		adev->gfx.mec_fw_version =
1039 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1040 		adev->gfx.mec_feature_version =
1041 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1042 		ucode_fw = adev->gfx.mec_fw;
1043 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1044 		break;
1045 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
1046 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
1047 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
1048 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
1049 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1050 			adev->gfx.mec_fw->data;
1051 		ucode_fw = adev->gfx.mec_fw;
1052 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1053 		break;
1054 	default:
1055 		break;
1056 	}
1057 
1058 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1059 		info = &adev->firmware.ucode[ucode_id];
1060 		info->ucode_id = ucode_id;
1061 		info->fw = ucode_fw;
1062 		adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE);
1063 	}
1064 }
1065