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 #define GFX_OFF_NO_DELAY 0
35 
36 /*
37  * GPU GFX IP block helpers function.
38  */
39 
40 int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
41 				int pipe, int queue)
42 {
43 	int bit = 0;
44 
45 	bit += mec * adev->gfx.mec.num_pipe_per_mec
46 		* adev->gfx.mec.num_queue_per_pipe;
47 	bit += pipe * adev->gfx.mec.num_queue_per_pipe;
48 	bit += queue;
49 
50 	return bit;
51 }
52 
53 void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
54 				 int *mec, int *pipe, int *queue)
55 {
56 	*queue = bit % adev->gfx.mec.num_queue_per_pipe;
57 	*pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
58 		% adev->gfx.mec.num_pipe_per_mec;
59 	*mec = (bit / adev->gfx.mec.num_queue_per_pipe)
60 	       / adev->gfx.mec.num_pipe_per_mec;
61 
62 }
63 
64 bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
65 				     int mec, int pipe, int queue)
66 {
67 	return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
68 			adev->gfx.mec.queue_bitmap);
69 }
70 
71 int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
72 			       int me, int pipe, int queue)
73 {
74 	int bit = 0;
75 
76 	bit += me * adev->gfx.me.num_pipe_per_me
77 		* adev->gfx.me.num_queue_per_pipe;
78 	bit += pipe * adev->gfx.me.num_queue_per_pipe;
79 	bit += queue;
80 
81 	return bit;
82 }
83 
84 void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
85 				int *me, int *pipe, int *queue)
86 {
87 	*queue = bit % adev->gfx.me.num_queue_per_pipe;
88 	*pipe = (bit / adev->gfx.me.num_queue_per_pipe)
89 		% adev->gfx.me.num_pipe_per_me;
90 	*me = (bit / adev->gfx.me.num_queue_per_pipe)
91 		/ adev->gfx.me.num_pipe_per_me;
92 }
93 
94 bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
95 				    int me, int pipe, int queue)
96 {
97 	return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
98 			adev->gfx.me.queue_bitmap);
99 }
100 
101 /**
102  * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
103  *
104  * @mask: array in which the per-shader array disable masks will be stored
105  * @max_se: number of SEs
106  * @max_sh: number of SHs
107  *
108  * The bitmask of CUs to be disabled in the shader array determined by se and
109  * sh is stored in mask[se * max_sh + sh].
110  */
111 void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
112 {
113 	unsigned se, sh, cu;
114 	const char *p;
115 
116 	memset(mask, 0, sizeof(*mask) * max_se * max_sh);
117 
118 	if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
119 		return;
120 
121 	p = amdgpu_disable_cu;
122 	for (;;) {
123 		char *next;
124 		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
125 		if (ret < 3) {
126 			DRM_ERROR("amdgpu: could not parse disable_cu\n");
127 			return;
128 		}
129 
130 		if (se < max_se && sh < max_sh && cu < 16) {
131 			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
132 			mask[se * max_sh + sh] |= 1u << cu;
133 		} else {
134 			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
135 				  se, sh, cu);
136 		}
137 
138 		next = strchr(p, ',');
139 		if (!next)
140 			break;
141 		p = next + 1;
142 	}
143 }
144 
145 static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
146 {
147 	if (amdgpu_compute_multipipe != -1) {
148 		DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
149 			 amdgpu_compute_multipipe);
150 		return amdgpu_compute_multipipe == 1;
151 	}
152 
153 	/* FIXME: spreading the queues across pipes causes perf regressions
154 	 * on POLARIS11 compute workloads */
155 	if (adev->asic_type == CHIP_POLARIS11)
156 		return false;
157 
158 	return adev->gfx.mec.num_mec > 1;
159 }
160 
161 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
162 					       struct amdgpu_ring *ring)
163 {
164 	/* Policy: use 1st queue as high priority compute queue if we
165 	 * have more than one compute queue.
166 	 */
167 	if (adev->gfx.num_compute_rings > 1 &&
168 	    ring == &adev->gfx.compute_ring[0])
169 		return true;
170 
171 	return false;
172 }
173 
174 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
175 {
176 	int i, queue, pipe;
177 	bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev);
178 	int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
179 				     adev->gfx.mec.num_queue_per_pipe,
180 				     adev->gfx.num_compute_rings);
181 
182 	if (multipipe_policy) {
183 		/* policy: make queues evenly cross all pipes on MEC1 only */
184 		for (i = 0; i < max_queues_per_mec; i++) {
185 			pipe = i % adev->gfx.mec.num_pipe_per_mec;
186 			queue = (i / adev->gfx.mec.num_pipe_per_mec) %
187 				adev->gfx.mec.num_queue_per_pipe;
188 
189 			set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
190 					adev->gfx.mec.queue_bitmap);
191 		}
192 	} else {
193 		/* policy: amdgpu owns all queues in the given pipe */
194 		for (i = 0; i < max_queues_per_mec; ++i)
195 			set_bit(i, adev->gfx.mec.queue_bitmap);
196 	}
197 
198 	dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
199 }
200 
201 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
202 {
203 	int i, queue, me;
204 
205 	for (i = 0; i < AMDGPU_MAX_GFX_QUEUES; ++i) {
206 		queue = i % adev->gfx.me.num_queue_per_pipe;
207 		me = (i / adev->gfx.me.num_queue_per_pipe)
208 		      / adev->gfx.me.num_pipe_per_me;
209 
210 		if (me >= adev->gfx.me.num_me)
211 			break;
212 		/* policy: amdgpu owns the first queue per pipe at this stage
213 		 * will extend to mulitple queues per pipe later */
214 		if (me == 0 && queue < 1)
215 			set_bit(i, adev->gfx.me.queue_bitmap);
216 	}
217 
218 	/* update the number of active graphics rings */
219 	adev->gfx.num_gfx_rings =
220 		bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
221 }
222 
223 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
224 				  struct amdgpu_ring *ring)
225 {
226 	int queue_bit;
227 	int mec, pipe, queue;
228 
229 	queue_bit = adev->gfx.mec.num_mec
230 		    * adev->gfx.mec.num_pipe_per_mec
231 		    * adev->gfx.mec.num_queue_per_pipe;
232 
233 	while (--queue_bit >= 0) {
234 		if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
235 			continue;
236 
237 		amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
238 
239 		/*
240 		 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
241 		 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
242 		 * only can be issued on queue 0.
243 		 */
244 		if ((mec == 1 && pipe > 1) || queue != 0)
245 			continue;
246 
247 		ring->me = mec + 1;
248 		ring->pipe = pipe;
249 		ring->queue = queue;
250 
251 		return 0;
252 	}
253 
254 	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
255 	return -EINVAL;
256 }
257 
258 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
259 			     struct amdgpu_ring *ring,
260 			     struct amdgpu_irq_src *irq)
261 {
262 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
263 	int r = 0;
264 
265 	spin_lock_init(&kiq->ring_lock);
266 
267 	ring->adev = NULL;
268 	ring->ring_obj = NULL;
269 	ring->use_doorbell = true;
270 	ring->doorbell_index = adev->doorbell_index.kiq;
271 
272 	r = amdgpu_gfx_kiq_acquire(adev, ring);
273 	if (r)
274 		return r;
275 
276 	ring->eop_gpu_addr = kiq->eop_gpu_addr;
277 	ring->no_scheduler = true;
278 	sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
279 	r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
280 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
281 	if (r)
282 		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
283 
284 	return r;
285 }
286 
287 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
288 {
289 	amdgpu_ring_fini(ring);
290 }
291 
292 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
293 {
294 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
295 
296 	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
297 }
298 
299 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
300 			unsigned hpd_size)
301 {
302 	int r;
303 	u32 *hpd;
304 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
305 
306 	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
307 				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
308 				    &kiq->eop_gpu_addr, (void **)&hpd);
309 	if (r) {
310 		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
311 		return r;
312 	}
313 
314 	memset(hpd, 0, hpd_size);
315 
316 	r = amdgpu_bo_reserve(kiq->eop_obj, true);
317 	if (unlikely(r != 0))
318 		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
319 	amdgpu_bo_kunmap(kiq->eop_obj);
320 	amdgpu_bo_unreserve(kiq->eop_obj);
321 
322 	return 0;
323 }
324 
325 /* create MQD for each compute/gfx queue */
326 int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
327 			   unsigned mqd_size)
328 {
329 	struct amdgpu_ring *ring = NULL;
330 	int r, i;
331 
332 	/* create MQD for KIQ */
333 	ring = &adev->gfx.kiq.ring;
334 	if (!adev->enable_mes_kiq && !ring->mqd_obj) {
335 		/* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
336 		 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
337 		 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
338 		 * KIQ MQD no matter SRIOV or Bare-metal
339 		 */
340 		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
341 					    AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj,
342 					    &ring->mqd_gpu_addr, &ring->mqd_ptr);
343 		if (r) {
344 			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
345 			return r;
346 		}
347 
348 		/* prepare MQD backup */
349 		adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
350 		if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
351 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
352 	}
353 
354 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
355 		/* create MQD for each KGQ */
356 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
357 			ring = &adev->gfx.gfx_ring[i];
358 			if (!ring->mqd_obj) {
359 				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
360 							    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
361 							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
362 				if (r) {
363 					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
364 					return r;
365 				}
366 
367 				/* prepare MQD backup */
368 				adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
369 				if (!adev->gfx.me.mqd_backup[i])
370 					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
371 			}
372 		}
373 	}
374 
375 	/* create MQD for each KCQ */
376 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
377 		ring = &adev->gfx.compute_ring[i];
378 		if (!ring->mqd_obj) {
379 			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
380 						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
381 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
382 			if (r) {
383 				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
384 				return r;
385 			}
386 
387 			/* prepare MQD backup */
388 			adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
389 			if (!adev->gfx.mec.mqd_backup[i])
390 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
391 		}
392 	}
393 
394 	return 0;
395 }
396 
397 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev)
398 {
399 	struct amdgpu_ring *ring = NULL;
400 	int i;
401 
402 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
403 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
404 			ring = &adev->gfx.gfx_ring[i];
405 			kfree(adev->gfx.me.mqd_backup[i]);
406 			amdgpu_bo_free_kernel(&ring->mqd_obj,
407 					      &ring->mqd_gpu_addr,
408 					      &ring->mqd_ptr);
409 		}
410 	}
411 
412 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
413 		ring = &adev->gfx.compute_ring[i];
414 		kfree(adev->gfx.mec.mqd_backup[i]);
415 		amdgpu_bo_free_kernel(&ring->mqd_obj,
416 				      &ring->mqd_gpu_addr,
417 				      &ring->mqd_ptr);
418 	}
419 
420 	ring = &adev->gfx.kiq.ring;
421 	kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
422 	amdgpu_bo_free_kernel(&ring->mqd_obj,
423 			      &ring->mqd_gpu_addr,
424 			      &ring->mqd_ptr);
425 }
426 
427 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev)
428 {
429 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
430 	struct amdgpu_ring *kiq_ring = &kiq->ring;
431 	int i, r = 0;
432 
433 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
434 		return -EINVAL;
435 
436 	spin_lock(&adev->gfx.kiq.ring_lock);
437 	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
438 					adev->gfx.num_compute_rings)) {
439 		spin_unlock(&adev->gfx.kiq.ring_lock);
440 		return -ENOMEM;
441 	}
442 
443 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
444 		kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i],
445 					   RESET_QUEUES, 0, 0);
446 
447 	if (adev->gfx.kiq.ring.sched.ready)
448 		r = amdgpu_ring_test_helper(kiq_ring);
449 	spin_unlock(&adev->gfx.kiq.ring_lock);
450 
451 	return r;
452 }
453 
454 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
455 					int queue_bit)
456 {
457 	int mec, pipe, queue;
458 	int set_resource_bit = 0;
459 
460 	amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
461 
462 	set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
463 
464 	return set_resource_bit;
465 }
466 
467 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev)
468 {
469 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
470 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
471 	uint64_t queue_mask = 0;
472 	int r, i;
473 
474 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
475 		return -EINVAL;
476 
477 	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
478 		if (!test_bit(i, adev->gfx.mec.queue_bitmap))
479 			continue;
480 
481 		/* This situation may be hit in the future if a new HW
482 		 * generation exposes more than 64 queues. If so, the
483 		 * definition of queue_mask needs updating */
484 		if (WARN_ON(i > (sizeof(queue_mask)*8))) {
485 			DRM_ERROR("Invalid KCQ enabled: %d\n", i);
486 			break;
487 		}
488 
489 		queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
490 	}
491 
492 	DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
493 							kiq_ring->queue);
494 	spin_lock(&adev->gfx.kiq.ring_lock);
495 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
496 					adev->gfx.num_compute_rings +
497 					kiq->pmf->set_resources_size);
498 	if (r) {
499 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
500 		spin_unlock(&adev->gfx.kiq.ring_lock);
501 		return r;
502 	}
503 
504 	if (adev->enable_mes)
505 		queue_mask = ~0ULL;
506 
507 	kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
508 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
509 		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
510 
511 	r = amdgpu_ring_test_helper(kiq_ring);
512 	spin_unlock(&adev->gfx.kiq.ring_lock);
513 	if (r)
514 		DRM_ERROR("KCQ enable failed\n");
515 
516 	return r;
517 }
518 
519 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
520  *
521  * @adev: amdgpu_device pointer
522  * @bool enable true: enable gfx off feature, false: disable gfx off feature
523  *
524  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
525  * 2. other client can send request to disable gfx off feature, the request should be honored.
526  * 3. other client can cancel their request of disable gfx off feature
527  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
528  */
529 
530 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
531 {
532 	unsigned long delay = GFX_OFF_DELAY_ENABLE;
533 
534 	if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
535 		return;
536 
537 	mutex_lock(&adev->gfx.gfx_off_mutex);
538 
539 	if (enable) {
540 		/* If the count is already 0, it means there's an imbalance bug somewhere.
541 		 * Note that the bug may be in a different caller than the one which triggers the
542 		 * WARN_ON_ONCE.
543 		 */
544 		if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
545 			goto unlock;
546 
547 		adev->gfx.gfx_off_req_count--;
548 
549 		if (adev->gfx.gfx_off_req_count == 0 &&
550 		    !adev->gfx.gfx_off_state) {
551 			/* If going to s2idle, no need to wait */
552 			if (adev->in_s0ix)
553 				delay = GFX_OFF_NO_DELAY;
554 			schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
555 					      delay);
556 		}
557 	} else {
558 		if (adev->gfx.gfx_off_req_count == 0) {
559 			cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
560 
561 			if (adev->gfx.gfx_off_state &&
562 			    !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
563 				adev->gfx.gfx_off_state = false;
564 
565 				if (adev->gfx.funcs->init_spm_golden) {
566 					dev_dbg(adev->dev,
567 						"GFXOFF is disabled, re-init SPM golden settings\n");
568 					amdgpu_gfx_init_spm_golden(adev);
569 				}
570 			}
571 		}
572 
573 		adev->gfx.gfx_off_req_count++;
574 	}
575 
576 unlock:
577 	mutex_unlock(&adev->gfx.gfx_off_mutex);
578 }
579 
580 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
581 {
582 
583 	int r = 0;
584 
585 	mutex_lock(&adev->gfx.gfx_off_mutex);
586 
587 	r = amdgpu_dpm_get_status_gfxoff(adev, value);
588 
589 	mutex_unlock(&adev->gfx.gfx_off_mutex);
590 
591 	return r;
592 }
593 
594 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
595 {
596 	int r;
597 	r = amdgpu_ras_block_late_init(adev, ras_block);
598 	if (r)
599 		return r;
600 
601 	if (amdgpu_ras_is_supported(adev, ras_block->block)) {
602 		if (!amdgpu_persistent_edc_harvesting_supported(adev))
603 			amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
604 
605 		r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
606 		if (r)
607 			goto late_fini;
608 	}
609 
610 	return 0;
611 late_fini:
612 	amdgpu_ras_block_late_fini(adev, ras_block);
613 	return r;
614 }
615 
616 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
617 		void *err_data,
618 		struct amdgpu_iv_entry *entry)
619 {
620 	/* TODO ue will trigger an interrupt.
621 	 *
622 	 * When “Full RAS” is enabled, the per-IP interrupt sources should
623 	 * be disabled and the driver should only look for the aggregated
624 	 * interrupt via sync flood
625 	 */
626 	if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
627 		kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
628 		if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
629 		    adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
630 			adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
631 		amdgpu_ras_reset_gpu(adev);
632 	}
633 	return AMDGPU_RAS_SUCCESS;
634 }
635 
636 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
637 				  struct amdgpu_irq_src *source,
638 				  struct amdgpu_iv_entry *entry)
639 {
640 	struct ras_common_if *ras_if = adev->gfx.ras_if;
641 	struct ras_dispatch_if ih_data = {
642 		.entry = entry,
643 	};
644 
645 	if (!ras_if)
646 		return 0;
647 
648 	ih_data.head = *ras_if;
649 
650 	DRM_ERROR("CP ECC ERROR IRQ\n");
651 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
652 	return 0;
653 }
654 
655 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
656 {
657 	signed long r, cnt = 0;
658 	unsigned long flags;
659 	uint32_t seq, reg_val_offs = 0, value = 0;
660 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
661 	struct amdgpu_ring *ring = &kiq->ring;
662 
663 	if (amdgpu_device_skip_hw_access(adev))
664 		return 0;
665 
666 	BUG_ON(!ring->funcs->emit_rreg);
667 
668 	spin_lock_irqsave(&kiq->ring_lock, flags);
669 	if (amdgpu_device_wb_get(adev, &reg_val_offs)) {
670 		pr_err("critical bug! too many kiq readers\n");
671 		goto failed_unlock;
672 	}
673 	amdgpu_ring_alloc(ring, 32);
674 	amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
675 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
676 	if (r)
677 		goto failed_undo;
678 
679 	amdgpu_ring_commit(ring);
680 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
681 
682 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
683 
684 	/* don't wait anymore for gpu reset case because this way may
685 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
686 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
687 	 * never return if we keep waiting in virt_kiq_rreg, which cause
688 	 * gpu_recover() hang there.
689 	 *
690 	 * also don't wait anymore for IRQ context
691 	 * */
692 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
693 		goto failed_kiq_read;
694 
695 	might_sleep();
696 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
697 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
698 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
699 	}
700 
701 	if (cnt > MAX_KIQ_REG_TRY)
702 		goto failed_kiq_read;
703 
704 	mb();
705 	value = adev->wb.wb[reg_val_offs];
706 	amdgpu_device_wb_free(adev, reg_val_offs);
707 	return value;
708 
709 failed_undo:
710 	amdgpu_ring_undo(ring);
711 failed_unlock:
712 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
713 failed_kiq_read:
714 	if (reg_val_offs)
715 		amdgpu_device_wb_free(adev, reg_val_offs);
716 	dev_err(adev->dev, "failed to read reg:%x\n", reg);
717 	return ~0;
718 }
719 
720 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
721 {
722 	signed long r, cnt = 0;
723 	unsigned long flags;
724 	uint32_t seq;
725 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
726 	struct amdgpu_ring *ring = &kiq->ring;
727 
728 	BUG_ON(!ring->funcs->emit_wreg);
729 
730 	if (amdgpu_device_skip_hw_access(adev))
731 		return;
732 
733 	spin_lock_irqsave(&kiq->ring_lock, flags);
734 	amdgpu_ring_alloc(ring, 32);
735 	amdgpu_ring_emit_wreg(ring, reg, v);
736 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
737 	if (r)
738 		goto failed_undo;
739 
740 	amdgpu_ring_commit(ring);
741 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
742 
743 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
744 
745 	/* don't wait anymore for gpu reset case because this way may
746 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
747 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
748 	 * never return if we keep waiting in virt_kiq_rreg, which cause
749 	 * gpu_recover() hang there.
750 	 *
751 	 * also don't wait anymore for IRQ context
752 	 * */
753 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
754 		goto failed_kiq_write;
755 
756 	might_sleep();
757 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
758 
759 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
760 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
761 	}
762 
763 	if (cnt > MAX_KIQ_REG_TRY)
764 		goto failed_kiq_write;
765 
766 	return;
767 
768 failed_undo:
769 	amdgpu_ring_undo(ring);
770 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
771 failed_kiq_write:
772 	dev_err(adev->dev, "failed to write reg:%x\n", reg);
773 }
774 
775 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
776 {
777 	if (amdgpu_num_kcq == -1) {
778 		return 8;
779 	} else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
780 		dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
781 		return 8;
782 	}
783 	return amdgpu_num_kcq;
784 }
785