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