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 "amdgpu.h"
27 #include "amdgpu_gfx.h"
28 #include "amdgpu_rlc.h"
29 #include "amdgpu_ras.h"
30 
31 /* delay 0.1 second to enable gfx off feature */
32 #define GFX_OFF_DELAY_ENABLE         msecs_to_jiffies(100)
33 
34 /*
35  * GPU GFX IP block helpers function.
36  */
37 
38 int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
39 				int pipe, int queue)
40 {
41 	int bit = 0;
42 
43 	bit += mec * adev->gfx.mec.num_pipe_per_mec
44 		* adev->gfx.mec.num_queue_per_pipe;
45 	bit += pipe * adev->gfx.mec.num_queue_per_pipe;
46 	bit += queue;
47 
48 	return bit;
49 }
50 
51 void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
52 				 int *mec, int *pipe, int *queue)
53 {
54 	*queue = bit % adev->gfx.mec.num_queue_per_pipe;
55 	*pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
56 		% adev->gfx.mec.num_pipe_per_mec;
57 	*mec = (bit / adev->gfx.mec.num_queue_per_pipe)
58 	       / adev->gfx.mec.num_pipe_per_mec;
59 
60 }
61 
62 bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
63 				     int mec, int pipe, int queue)
64 {
65 	return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
66 			adev->gfx.mec.queue_bitmap);
67 }
68 
69 int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
70 			       int me, int pipe, int queue)
71 {
72 	int bit = 0;
73 
74 	bit += me * adev->gfx.me.num_pipe_per_me
75 		* adev->gfx.me.num_queue_per_pipe;
76 	bit += pipe * adev->gfx.me.num_queue_per_pipe;
77 	bit += queue;
78 
79 	return bit;
80 }
81 
82 void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
83 				int *me, int *pipe, int *queue)
84 {
85 	*queue = bit % adev->gfx.me.num_queue_per_pipe;
86 	*pipe = (bit / adev->gfx.me.num_queue_per_pipe)
87 		% adev->gfx.me.num_pipe_per_me;
88 	*me = (bit / adev->gfx.me.num_queue_per_pipe)
89 		/ adev->gfx.me.num_pipe_per_me;
90 }
91 
92 bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
93 				    int me, int pipe, int queue)
94 {
95 	return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
96 			adev->gfx.me.queue_bitmap);
97 }
98 
99 /**
100  * amdgpu_gfx_scratch_get - Allocate a scratch register
101  *
102  * @adev: amdgpu_device pointer
103  * @reg: scratch register mmio offset
104  *
105  * Allocate a CP scratch register for use by the driver (all asics).
106  * Returns 0 on success or -EINVAL on failure.
107  */
108 int amdgpu_gfx_scratch_get(struct amdgpu_device *adev, uint32_t *reg)
109 {
110 	int i;
111 
112 	i = ffs(adev->gfx.scratch.free_mask);
113 	if (i != 0 && i <= adev->gfx.scratch.num_reg) {
114 		i--;
115 		adev->gfx.scratch.free_mask &= ~(1u << i);
116 		*reg = adev->gfx.scratch.reg_base + i;
117 		return 0;
118 	}
119 	return -EINVAL;
120 }
121 
122 /**
123  * amdgpu_gfx_scratch_free - Free a scratch register
124  *
125  * @adev: amdgpu_device pointer
126  * @reg: scratch register mmio offset
127  *
128  * Free a CP scratch register allocated for use by the driver (all asics)
129  */
130 void amdgpu_gfx_scratch_free(struct amdgpu_device *adev, uint32_t reg)
131 {
132 	adev->gfx.scratch.free_mask |= 1u << (reg - adev->gfx.scratch.reg_base);
133 }
134 
135 /**
136  * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
137  *
138  * @mask: array in which the per-shader array disable masks will be stored
139  * @max_se: number of SEs
140  * @max_sh: number of SHs
141  *
142  * The bitmask of CUs to be disabled in the shader array determined by se and
143  * sh is stored in mask[se * max_sh + sh].
144  */
145 void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
146 {
147 	unsigned se, sh, cu;
148 	const char *p;
149 
150 	memset(mask, 0, sizeof(*mask) * max_se * max_sh);
151 
152 	if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
153 		return;
154 
155 	p = amdgpu_disable_cu;
156 	for (;;) {
157 		char *next;
158 		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
159 		if (ret < 3) {
160 			DRM_ERROR("amdgpu: could not parse disable_cu\n");
161 			return;
162 		}
163 
164 		if (se < max_se && sh < max_sh && cu < 16) {
165 			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
166 			mask[se * max_sh + sh] |= 1u << cu;
167 		} else {
168 			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
169 				  se, sh, cu);
170 		}
171 
172 		next = strchr(p, ',');
173 		if (!next)
174 			break;
175 		p = next + 1;
176 	}
177 }
178 
179 static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
180 {
181 	if (amdgpu_compute_multipipe != -1) {
182 		DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
183 			 amdgpu_compute_multipipe);
184 		return amdgpu_compute_multipipe == 1;
185 	}
186 
187 	/* FIXME: spreading the queues across pipes causes perf regressions
188 	 * on POLARIS11 compute workloads */
189 	if (adev->asic_type == CHIP_POLARIS11)
190 		return false;
191 
192 	return adev->gfx.mec.num_mec > 1;
193 }
194 
195 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
196 					       struct amdgpu_ring *ring)
197 {
198 	/* Policy: use 1st queue as high priority compute queue if we
199 	 * have more than one compute queue.
200 	 */
201 	if (adev->gfx.num_compute_rings > 1 &&
202 	    ring == &adev->gfx.compute_ring[0])
203 		return true;
204 
205 	return false;
206 }
207 
208 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
209 {
210 	int i, queue, pipe;
211 	bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev);
212 	int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
213 				     adev->gfx.mec.num_queue_per_pipe,
214 				     adev->gfx.num_compute_rings);
215 
216 	if (multipipe_policy) {
217 		/* policy: make queues evenly cross all pipes on MEC1 only */
218 		for (i = 0; i < max_queues_per_mec; i++) {
219 			pipe = i % adev->gfx.mec.num_pipe_per_mec;
220 			queue = (i / adev->gfx.mec.num_pipe_per_mec) %
221 				adev->gfx.mec.num_queue_per_pipe;
222 
223 			set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
224 					adev->gfx.mec.queue_bitmap);
225 		}
226 	} else {
227 		/* policy: amdgpu owns all queues in the given pipe */
228 		for (i = 0; i < max_queues_per_mec; ++i)
229 			set_bit(i, adev->gfx.mec.queue_bitmap);
230 	}
231 
232 	dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
233 }
234 
235 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
236 {
237 	int i, queue, me;
238 
239 	for (i = 0; i < AMDGPU_MAX_GFX_QUEUES; ++i) {
240 		queue = i % adev->gfx.me.num_queue_per_pipe;
241 		me = (i / adev->gfx.me.num_queue_per_pipe)
242 		      / adev->gfx.me.num_pipe_per_me;
243 
244 		if (me >= adev->gfx.me.num_me)
245 			break;
246 		/* policy: amdgpu owns the first queue per pipe at this stage
247 		 * will extend to mulitple queues per pipe later */
248 		if (me == 0 && queue < 1)
249 			set_bit(i, adev->gfx.me.queue_bitmap);
250 	}
251 
252 	/* update the number of active graphics rings */
253 	adev->gfx.num_gfx_rings =
254 		bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
255 }
256 
257 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
258 				  struct amdgpu_ring *ring)
259 {
260 	int queue_bit;
261 	int mec, pipe, queue;
262 
263 	queue_bit = adev->gfx.mec.num_mec
264 		    * adev->gfx.mec.num_pipe_per_mec
265 		    * adev->gfx.mec.num_queue_per_pipe;
266 
267 	while (queue_bit-- >= 0) {
268 		if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
269 			continue;
270 
271 		amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
272 
273 		/*
274 		 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
275 		 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
276 		 * only can be issued on queue 0.
277 		 */
278 		if ((mec == 1 && pipe > 1) || queue != 0)
279 			continue;
280 
281 		ring->me = mec + 1;
282 		ring->pipe = pipe;
283 		ring->queue = queue;
284 
285 		return 0;
286 	}
287 
288 	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
289 	return -EINVAL;
290 }
291 
292 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
293 			     struct amdgpu_ring *ring,
294 			     struct amdgpu_irq_src *irq)
295 {
296 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
297 	int r = 0;
298 
299 	spin_lock_init(&kiq->ring_lock);
300 
301 	ring->adev = NULL;
302 	ring->ring_obj = NULL;
303 	ring->use_doorbell = true;
304 	ring->doorbell_index = adev->doorbell_index.kiq;
305 
306 	r = amdgpu_gfx_kiq_acquire(adev, ring);
307 	if (r)
308 		return r;
309 
310 	ring->eop_gpu_addr = kiq->eop_gpu_addr;
311 	ring->no_scheduler = true;
312 	sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
313 	r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
314 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
315 	if (r)
316 		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
317 
318 	return r;
319 }
320 
321 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
322 {
323 	amdgpu_ring_fini(ring);
324 }
325 
326 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
327 {
328 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
329 
330 	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
331 }
332 
333 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
334 			unsigned hpd_size)
335 {
336 	int r;
337 	u32 *hpd;
338 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
339 
340 	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
341 				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
342 				    &kiq->eop_gpu_addr, (void **)&hpd);
343 	if (r) {
344 		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
345 		return r;
346 	}
347 
348 	memset(hpd, 0, hpd_size);
349 
350 	r = amdgpu_bo_reserve(kiq->eop_obj, true);
351 	if (unlikely(r != 0))
352 		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
353 	amdgpu_bo_kunmap(kiq->eop_obj);
354 	amdgpu_bo_unreserve(kiq->eop_obj);
355 
356 	return 0;
357 }
358 
359 /* create MQD for each compute/gfx queue */
360 int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
361 			   unsigned mqd_size)
362 {
363 	struct amdgpu_ring *ring = NULL;
364 	int r, i;
365 
366 	/* create MQD for KIQ */
367 	ring = &adev->gfx.kiq.ring;
368 	if (!ring->mqd_obj) {
369 		/* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
370 		 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
371 		 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
372 		 * KIQ MQD no matter SRIOV or Bare-metal
373 		 */
374 		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
375 					    AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj,
376 					    &ring->mqd_gpu_addr, &ring->mqd_ptr);
377 		if (r) {
378 			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
379 			return r;
380 		}
381 
382 		/* prepare MQD backup */
383 		adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
384 		if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
385 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
386 	}
387 
388 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
389 		/* create MQD for each KGQ */
390 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
391 			ring = &adev->gfx.gfx_ring[i];
392 			if (!ring->mqd_obj) {
393 				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
394 							    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
395 							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
396 				if (r) {
397 					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
398 					return r;
399 				}
400 
401 				/* prepare MQD backup */
402 				adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
403 				if (!adev->gfx.me.mqd_backup[i])
404 					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
405 			}
406 		}
407 	}
408 
409 	/* create MQD for each KCQ */
410 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
411 		ring = &adev->gfx.compute_ring[i];
412 		if (!ring->mqd_obj) {
413 			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
414 						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
415 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
416 			if (r) {
417 				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
418 				return r;
419 			}
420 
421 			/* prepare MQD backup */
422 			adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
423 			if (!adev->gfx.mec.mqd_backup[i])
424 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
425 		}
426 	}
427 
428 	return 0;
429 }
430 
431 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev)
432 {
433 	struct amdgpu_ring *ring = NULL;
434 	int i;
435 
436 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
437 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
438 			ring = &adev->gfx.gfx_ring[i];
439 			kfree(adev->gfx.me.mqd_backup[i]);
440 			amdgpu_bo_free_kernel(&ring->mqd_obj,
441 					      &ring->mqd_gpu_addr,
442 					      &ring->mqd_ptr);
443 		}
444 	}
445 
446 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
447 		ring = &adev->gfx.compute_ring[i];
448 		kfree(adev->gfx.mec.mqd_backup[i]);
449 		amdgpu_bo_free_kernel(&ring->mqd_obj,
450 				      &ring->mqd_gpu_addr,
451 				      &ring->mqd_ptr);
452 	}
453 
454 	ring = &adev->gfx.kiq.ring;
455 	kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
456 	amdgpu_bo_free_kernel(&ring->mqd_obj,
457 			      &ring->mqd_gpu_addr,
458 			      &ring->mqd_ptr);
459 }
460 
461 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev)
462 {
463 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
464 	struct amdgpu_ring *kiq_ring = &kiq->ring;
465 	int i, r;
466 
467 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
468 		return -EINVAL;
469 
470 	spin_lock(&adev->gfx.kiq.ring_lock);
471 	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
472 					adev->gfx.num_compute_rings)) {
473 		spin_unlock(&adev->gfx.kiq.ring_lock);
474 		return -ENOMEM;
475 	}
476 
477 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
478 		kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i],
479 					   RESET_QUEUES, 0, 0);
480 	r = amdgpu_ring_test_helper(kiq_ring);
481 	spin_unlock(&adev->gfx.kiq.ring_lock);
482 
483 	return r;
484 }
485 
486 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
487 					int queue_bit)
488 {
489 	int mec, pipe, queue;
490 	int set_resource_bit = 0;
491 
492 	amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
493 
494 	set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
495 
496 	return set_resource_bit;
497 }
498 
499 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev)
500 {
501 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
502 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
503 	uint64_t queue_mask = 0;
504 	int r, i;
505 
506 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
507 		return -EINVAL;
508 
509 	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
510 		if (!test_bit(i, adev->gfx.mec.queue_bitmap))
511 			continue;
512 
513 		/* This situation may be hit in the future if a new HW
514 		 * generation exposes more than 64 queues. If so, the
515 		 * definition of queue_mask needs updating */
516 		if (WARN_ON(i > (sizeof(queue_mask)*8))) {
517 			DRM_ERROR("Invalid KCQ enabled: %d\n", i);
518 			break;
519 		}
520 
521 		queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
522 	}
523 
524 	DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
525 							kiq_ring->queue);
526 	spin_lock(&adev->gfx.kiq.ring_lock);
527 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
528 					adev->gfx.num_compute_rings +
529 					kiq->pmf->set_resources_size);
530 	if (r) {
531 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
532 		spin_unlock(&adev->gfx.kiq.ring_lock);
533 		return r;
534 	}
535 
536 	kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
537 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
538 		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
539 
540 	r = amdgpu_ring_test_helper(kiq_ring);
541 	spin_unlock(&adev->gfx.kiq.ring_lock);
542 	if (r)
543 		DRM_ERROR("KCQ enable failed\n");
544 
545 	return r;
546 }
547 
548 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
549  *
550  * @adev: amdgpu_device pointer
551  * @bool enable true: enable gfx off feature, false: disable gfx off feature
552  *
553  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
554  * 2. other client can send request to disable gfx off feature, the request should be honored.
555  * 3. other client can cancel their request of disable gfx off feature
556  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
557  */
558 
559 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
560 {
561 	if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
562 		return;
563 
564 	mutex_lock(&adev->gfx.gfx_off_mutex);
565 
566 	if (!enable)
567 		adev->gfx.gfx_off_req_count++;
568 	else if (adev->gfx.gfx_off_req_count > 0)
569 		adev->gfx.gfx_off_req_count--;
570 
571 	if (enable && !adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
572 		schedule_delayed_work(&adev->gfx.gfx_off_delay_work, GFX_OFF_DELAY_ENABLE);
573 	} else if (!enable && adev->gfx.gfx_off_state) {
574 		if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
575 			adev->gfx.gfx_off_state = false;
576 
577 			if (adev->gfx.funcs->init_spm_golden) {
578 				dev_dbg(adev->dev, "GFXOFF is disabled, re-init SPM golden settings\n");
579 				amdgpu_gfx_init_spm_golden(adev);
580 			}
581 		}
582 	}
583 
584 	mutex_unlock(&adev->gfx.gfx_off_mutex);
585 }
586 
587 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
588 {
589 
590 	int r = 0;
591 
592 	mutex_lock(&adev->gfx.gfx_off_mutex);
593 
594 	r = smu_get_status_gfxoff(adev, value);
595 
596 	mutex_unlock(&adev->gfx.gfx_off_mutex);
597 
598 	return r;
599 }
600 
601 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev)
602 {
603 	int r;
604 	struct ras_fs_if fs_info = {
605 		.sysfs_name = "gfx_err_count",
606 	};
607 	struct ras_ih_if ih_info = {
608 		.cb = amdgpu_gfx_process_ras_data_cb,
609 	};
610 	struct ras_query_if info = { 0 };
611 
612 	if (!adev->gfx.ras_if) {
613 		adev->gfx.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL);
614 		if (!adev->gfx.ras_if)
615 			return -ENOMEM;
616 		adev->gfx.ras_if->block = AMDGPU_RAS_BLOCK__GFX;
617 		adev->gfx.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
618 		adev->gfx.ras_if->sub_block_index = 0;
619 		strcpy(adev->gfx.ras_if->name, "gfx");
620 	}
621 	fs_info.head = ih_info.head = *adev->gfx.ras_if;
622 	r = amdgpu_ras_late_init(adev, adev->gfx.ras_if,
623 				 &fs_info, &ih_info);
624 	if (r)
625 		goto free;
626 
627 	if (amdgpu_ras_is_supported(adev, adev->gfx.ras_if->block)) {
628 		if (adev->gmc.xgmi.connected_to_cpu) {
629 			info.head = *adev->gfx.ras_if;
630 			amdgpu_ras_query_error_status(adev, &info);
631 		} else {
632 			amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
633 		}
634 
635 		r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
636 		if (r)
637 			goto late_fini;
638 	} else {
639 		/* free gfx ras_if if ras is not supported */
640 		r = 0;
641 		goto free;
642 	}
643 
644 	return 0;
645 late_fini:
646 	amdgpu_ras_late_fini(adev, adev->gfx.ras_if, &ih_info);
647 free:
648 	kfree(adev->gfx.ras_if);
649 	adev->gfx.ras_if = NULL;
650 	return r;
651 }
652 
653 void amdgpu_gfx_ras_fini(struct amdgpu_device *adev)
654 {
655 	if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX) &&
656 			adev->gfx.ras_if) {
657 		struct ras_common_if *ras_if = adev->gfx.ras_if;
658 		struct ras_ih_if ih_info = {
659 			.head = *ras_if,
660 			.cb = amdgpu_gfx_process_ras_data_cb,
661 		};
662 
663 		amdgpu_ras_late_fini(adev, ras_if, &ih_info);
664 		kfree(ras_if);
665 	}
666 }
667 
668 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
669 		void *err_data,
670 		struct amdgpu_iv_entry *entry)
671 {
672 	/* TODO ue will trigger an interrupt.
673 	 *
674 	 * When “Full RAS” is enabled, the per-IP interrupt sources should
675 	 * be disabled and the driver should only look for the aggregated
676 	 * interrupt via sync flood
677 	 */
678 	if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
679 		kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
680 		if (adev->gfx.ras_funcs &&
681 		    adev->gfx.ras_funcs->query_ras_error_count)
682 			adev->gfx.ras_funcs->query_ras_error_count(adev, err_data);
683 		amdgpu_ras_reset_gpu(adev);
684 	}
685 	return AMDGPU_RAS_SUCCESS;
686 }
687 
688 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
689 				  struct amdgpu_irq_src *source,
690 				  struct amdgpu_iv_entry *entry)
691 {
692 	struct ras_common_if *ras_if = adev->gfx.ras_if;
693 	struct ras_dispatch_if ih_data = {
694 		.entry = entry,
695 	};
696 
697 	if (!ras_if)
698 		return 0;
699 
700 	ih_data.head = *ras_if;
701 
702 	DRM_ERROR("CP ECC ERROR IRQ\n");
703 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
704 	return 0;
705 }
706 
707 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
708 {
709 	signed long r, cnt = 0;
710 	unsigned long flags;
711 	uint32_t seq, reg_val_offs = 0, value = 0;
712 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
713 	struct amdgpu_ring *ring = &kiq->ring;
714 
715 	if (amdgpu_device_skip_hw_access(adev))
716 		return 0;
717 
718 	BUG_ON(!ring->funcs->emit_rreg);
719 
720 	spin_lock_irqsave(&kiq->ring_lock, flags);
721 	if (amdgpu_device_wb_get(adev, &reg_val_offs)) {
722 		pr_err("critical bug! too many kiq readers\n");
723 		goto failed_unlock;
724 	}
725 	amdgpu_ring_alloc(ring, 32);
726 	amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
727 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
728 	if (r)
729 		goto failed_undo;
730 
731 	amdgpu_ring_commit(ring);
732 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
733 
734 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
735 
736 	/* don't wait anymore for gpu reset case because this way may
737 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
738 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
739 	 * never return if we keep waiting in virt_kiq_rreg, which cause
740 	 * gpu_recover() hang there.
741 	 *
742 	 * also don't wait anymore for IRQ context
743 	 * */
744 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
745 		goto failed_kiq_read;
746 
747 	might_sleep();
748 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
749 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
750 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
751 	}
752 
753 	if (cnt > MAX_KIQ_REG_TRY)
754 		goto failed_kiq_read;
755 
756 	mb();
757 	value = adev->wb.wb[reg_val_offs];
758 	amdgpu_device_wb_free(adev, reg_val_offs);
759 	return value;
760 
761 failed_undo:
762 	amdgpu_ring_undo(ring);
763 failed_unlock:
764 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
765 failed_kiq_read:
766 	if (reg_val_offs)
767 		amdgpu_device_wb_free(adev, reg_val_offs);
768 	dev_err(adev->dev, "failed to read reg:%x\n", reg);
769 	return ~0;
770 }
771 
772 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
773 {
774 	signed long r, cnt = 0;
775 	unsigned long flags;
776 	uint32_t seq;
777 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
778 	struct amdgpu_ring *ring = &kiq->ring;
779 
780 	BUG_ON(!ring->funcs->emit_wreg);
781 
782 	if (amdgpu_device_skip_hw_access(adev))
783 		return;
784 
785 	spin_lock_irqsave(&kiq->ring_lock, flags);
786 	amdgpu_ring_alloc(ring, 32);
787 	amdgpu_ring_emit_wreg(ring, reg, v);
788 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
789 	if (r)
790 		goto failed_undo;
791 
792 	amdgpu_ring_commit(ring);
793 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
794 
795 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
796 
797 	/* don't wait anymore for gpu reset case because this way may
798 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
799 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
800 	 * never return if we keep waiting in virt_kiq_rreg, which cause
801 	 * gpu_recover() hang there.
802 	 *
803 	 * also don't wait anymore for IRQ context
804 	 * */
805 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
806 		goto failed_kiq_write;
807 
808 	might_sleep();
809 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
810 
811 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
812 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
813 	}
814 
815 	if (cnt > MAX_KIQ_REG_TRY)
816 		goto failed_kiq_write;
817 
818 	return;
819 
820 failed_undo:
821 	amdgpu_ring_undo(ring);
822 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
823 failed_kiq_write:
824 	dev_err(adev->dev, "failed to write reg:%x\n", reg);
825 }
826 
827 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
828 {
829 	if (amdgpu_num_kcq == -1) {
830 		return 8;
831 	} else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
832 		dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
833 		return 8;
834 	}
835 	return amdgpu_num_kcq;
836 }
837 
838 /* amdgpu_gfx_state_change_set - Handle gfx power state change set
839  * @adev: amdgpu_device pointer
840  * @state: gfx power state(1 -sGpuChangeState_D0Entry and 2 -sGpuChangeState_D3Entry)
841  *
842  */
843 
844 void amdgpu_gfx_state_change_set(struct amdgpu_device *adev, enum gfx_change_state state)
845 {
846 	mutex_lock(&adev->pm.mutex);
847 	if (adev->powerplay.pp_funcs &&
848 	    adev->powerplay.pp_funcs->gfx_state_change_set)
849 		((adev)->powerplay.pp_funcs->gfx_state_change_set(
850 			(adev)->powerplay.pp_handle, state));
851 	mutex_unlock(&adev->pm.mutex);
852 }
853