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 	/* FIXME: spreading the queues across pipes causes perf regressions
160 	 * on POLARIS11 compute workloads */
161 	if (adev->asic_type == CHIP_POLARIS11)
162 		return false;
163 
164 	return adev->gfx.mec.num_mec > 1;
165 }
166 
167 bool amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device *adev,
168 						struct amdgpu_ring *ring)
169 {
170 	int queue = ring->queue;
171 	int pipe = ring->pipe;
172 
173 	/* Policy: use pipe1 queue0 as high priority graphics queue if we
174 	 * have more than one gfx pipe.
175 	 */
176 	if (amdgpu_gfx_is_graphics_multipipe_capable(adev) &&
177 	    adev->gfx.num_gfx_rings > 1 && pipe == 1 && queue == 0) {
178 		int me = ring->me;
179 		int bit;
180 
181 		bit = amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue);
182 		if (ring == &adev->gfx.gfx_ring[bit])
183 			return true;
184 	}
185 
186 	return false;
187 }
188 
189 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
190 					       struct amdgpu_ring *ring)
191 {
192 	/* Policy: use 1st queue as high priority compute queue if we
193 	 * have more than one compute queue.
194 	 */
195 	if (adev->gfx.num_compute_rings > 1 &&
196 	    ring == &adev->gfx.compute_ring[0])
197 		return true;
198 
199 	return false;
200 }
201 
202 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
203 {
204 	int i, queue, pipe;
205 	bool multipipe_policy = amdgpu_gfx_is_compute_multipipe_capable(adev);
206 	int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
207 				     adev->gfx.mec.num_queue_per_pipe,
208 				     adev->gfx.num_compute_rings);
209 
210 	if (multipipe_policy) {
211 		/* policy: make queues evenly cross all pipes on MEC1 only */
212 		for (i = 0; i < max_queues_per_mec; i++) {
213 			pipe = i % adev->gfx.mec.num_pipe_per_mec;
214 			queue = (i / adev->gfx.mec.num_pipe_per_mec) %
215 				adev->gfx.mec.num_queue_per_pipe;
216 
217 			set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
218 					adev->gfx.mec.queue_bitmap);
219 		}
220 	} else {
221 		/* policy: amdgpu owns all queues in the given pipe */
222 		for (i = 0; i < max_queues_per_mec; ++i)
223 			set_bit(i, adev->gfx.mec.queue_bitmap);
224 	}
225 
226 	dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
227 }
228 
229 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
230 {
231 	int i, queue, pipe;
232 	bool multipipe_policy = amdgpu_gfx_is_graphics_multipipe_capable(adev);
233 	int max_queues_per_me = adev->gfx.me.num_pipe_per_me *
234 					adev->gfx.me.num_queue_per_pipe;
235 
236 	if (multipipe_policy) {
237 		/* policy: amdgpu owns the first queue per pipe at this stage
238 		 * will extend to mulitple queues per pipe later */
239 		for (i = 0; i < max_queues_per_me; i++) {
240 			pipe = i % adev->gfx.me.num_pipe_per_me;
241 			queue = (i / adev->gfx.me.num_pipe_per_me) %
242 				adev->gfx.me.num_queue_per_pipe;
243 
244 			set_bit(pipe * adev->gfx.me.num_queue_per_pipe + queue,
245 				adev->gfx.me.queue_bitmap);
246 		}
247 	} else {
248 		for (i = 0; i < max_queues_per_me; ++i)
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 (!adev->enable_mes_kiq && !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 = 0;
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 
481 	if (adev->gfx.kiq.ring.sched.ready && !adev->job_hang)
482 		r = amdgpu_ring_test_helper(kiq_ring);
483 	spin_unlock(&adev->gfx.kiq.ring_lock);
484 
485 	return r;
486 }
487 
488 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
489 					int queue_bit)
490 {
491 	int mec, pipe, queue;
492 	int set_resource_bit = 0;
493 
494 	amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
495 
496 	set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
497 
498 	return set_resource_bit;
499 }
500 
501 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev)
502 {
503 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
504 	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
505 	uint64_t queue_mask = 0;
506 	int r, i;
507 
508 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
509 		return -EINVAL;
510 
511 	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
512 		if (!test_bit(i, adev->gfx.mec.queue_bitmap))
513 			continue;
514 
515 		/* This situation may be hit in the future if a new HW
516 		 * generation exposes more than 64 queues. If so, the
517 		 * definition of queue_mask needs updating */
518 		if (WARN_ON(i > (sizeof(queue_mask)*8))) {
519 			DRM_ERROR("Invalid KCQ enabled: %d\n", i);
520 			break;
521 		}
522 
523 		queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
524 	}
525 
526 	DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
527 							kiq_ring->queue);
528 	spin_lock(&adev->gfx.kiq.ring_lock);
529 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
530 					adev->gfx.num_compute_rings +
531 					kiq->pmf->set_resources_size);
532 	if (r) {
533 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
534 		spin_unlock(&adev->gfx.kiq.ring_lock);
535 		return r;
536 	}
537 
538 	if (adev->enable_mes)
539 		queue_mask = ~0ULL;
540 
541 	kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
542 	for (i = 0; i < adev->gfx.num_compute_rings; i++)
543 		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
544 
545 	r = amdgpu_ring_test_helper(kiq_ring);
546 	spin_unlock(&adev->gfx.kiq.ring_lock);
547 	if (r)
548 		DRM_ERROR("KCQ enable failed\n");
549 
550 	return r;
551 }
552 
553 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
554  *
555  * @adev: amdgpu_device pointer
556  * @bool enable true: enable gfx off feature, false: disable gfx off feature
557  *
558  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
559  * 2. other client can send request to disable gfx off feature, the request should be honored.
560  * 3. other client can cancel their request of disable gfx off feature
561  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
562  */
563 
564 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
565 {
566 	unsigned long delay = GFX_OFF_DELAY_ENABLE;
567 
568 	if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
569 		return;
570 
571 	mutex_lock(&adev->gfx.gfx_off_mutex);
572 
573 	if (enable) {
574 		/* If the count is already 0, it means there's an imbalance bug somewhere.
575 		 * Note that the bug may be in a different caller than the one which triggers the
576 		 * WARN_ON_ONCE.
577 		 */
578 		if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
579 			goto unlock;
580 
581 		adev->gfx.gfx_off_req_count--;
582 
583 		if (adev->gfx.gfx_off_req_count == 0 &&
584 		    !adev->gfx.gfx_off_state) {
585 			/* If going to s2idle, no need to wait */
586 			if (adev->in_s0ix) {
587 				if (!amdgpu_dpm_set_powergating_by_smu(adev,
588 						AMD_IP_BLOCK_TYPE_GFX, true))
589 					adev->gfx.gfx_off_state = true;
590 			} else {
591 				schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
592 					      delay);
593 			}
594 		}
595 	} else {
596 		if (adev->gfx.gfx_off_req_count == 0) {
597 			cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
598 
599 			if (adev->gfx.gfx_off_state &&
600 			    !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
601 				adev->gfx.gfx_off_state = false;
602 
603 				if (adev->gfx.funcs->init_spm_golden) {
604 					dev_dbg(adev->dev,
605 						"GFXOFF is disabled, re-init SPM golden settings\n");
606 					amdgpu_gfx_init_spm_golden(adev);
607 				}
608 			}
609 		}
610 
611 		adev->gfx.gfx_off_req_count++;
612 	}
613 
614 unlock:
615 	mutex_unlock(&adev->gfx.gfx_off_mutex);
616 }
617 
618 int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
619 {
620 	int r = 0;
621 
622 	mutex_lock(&adev->gfx.gfx_off_mutex);
623 
624 	r = amdgpu_dpm_set_residency_gfxoff(adev, value);
625 
626 	mutex_unlock(&adev->gfx.gfx_off_mutex);
627 
628 	return r;
629 }
630 
631 int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
632 {
633 	int r = 0;
634 
635 	mutex_lock(&adev->gfx.gfx_off_mutex);
636 
637 	r = amdgpu_dpm_get_residency_gfxoff(adev, value);
638 
639 	mutex_unlock(&adev->gfx.gfx_off_mutex);
640 
641 	return r;
642 }
643 
644 int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
645 {
646 	int r = 0;
647 
648 	mutex_lock(&adev->gfx.gfx_off_mutex);
649 
650 	r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
651 
652 	mutex_unlock(&adev->gfx.gfx_off_mutex);
653 
654 	return r;
655 }
656 
657 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
658 {
659 
660 	int r = 0;
661 
662 	mutex_lock(&adev->gfx.gfx_off_mutex);
663 
664 	r = amdgpu_dpm_get_status_gfxoff(adev, value);
665 
666 	mutex_unlock(&adev->gfx.gfx_off_mutex);
667 
668 	return r;
669 }
670 
671 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
672 {
673 	int r;
674 
675 	if (amdgpu_ras_is_supported(adev, ras_block->block)) {
676 		if (!amdgpu_persistent_edc_harvesting_supported(adev))
677 			amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
678 
679 		r = amdgpu_ras_block_late_init(adev, ras_block);
680 		if (r)
681 			return r;
682 
683 		r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
684 		if (r)
685 			goto late_fini;
686 	} else {
687 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
688 	}
689 
690 	return 0;
691 late_fini:
692 	amdgpu_ras_block_late_fini(adev, ras_block);
693 	return r;
694 }
695 
696 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
697 		void *err_data,
698 		struct amdgpu_iv_entry *entry)
699 {
700 	/* TODO ue will trigger an interrupt.
701 	 *
702 	 * When “Full RAS” is enabled, the per-IP interrupt sources should
703 	 * be disabled and the driver should only look for the aggregated
704 	 * interrupt via sync flood
705 	 */
706 	if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
707 		kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
708 		if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
709 		    adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
710 			adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
711 		amdgpu_ras_reset_gpu(adev);
712 	}
713 	return AMDGPU_RAS_SUCCESS;
714 }
715 
716 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
717 				  struct amdgpu_irq_src *source,
718 				  struct amdgpu_iv_entry *entry)
719 {
720 	struct ras_common_if *ras_if = adev->gfx.ras_if;
721 	struct ras_dispatch_if ih_data = {
722 		.entry = entry,
723 	};
724 
725 	if (!ras_if)
726 		return 0;
727 
728 	ih_data.head = *ras_if;
729 
730 	DRM_ERROR("CP ECC ERROR IRQ\n");
731 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
732 	return 0;
733 }
734 
735 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
736 {
737 	signed long r, cnt = 0;
738 	unsigned long flags;
739 	uint32_t seq, reg_val_offs = 0, value = 0;
740 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
741 	struct amdgpu_ring *ring = &kiq->ring;
742 
743 	if (amdgpu_device_skip_hw_access(adev))
744 		return 0;
745 
746 	if (adev->mes.ring.sched.ready)
747 		return amdgpu_mes_rreg(adev, reg);
748 
749 	BUG_ON(!ring->funcs->emit_rreg);
750 
751 	spin_lock_irqsave(&kiq->ring_lock, flags);
752 	if (amdgpu_device_wb_get(adev, &reg_val_offs)) {
753 		pr_err("critical bug! too many kiq readers\n");
754 		goto failed_unlock;
755 	}
756 	amdgpu_ring_alloc(ring, 32);
757 	amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
758 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
759 	if (r)
760 		goto failed_undo;
761 
762 	amdgpu_ring_commit(ring);
763 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
764 
765 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
766 
767 	/* don't wait anymore for gpu reset case because this way may
768 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
769 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
770 	 * never return if we keep waiting in virt_kiq_rreg, which cause
771 	 * gpu_recover() hang there.
772 	 *
773 	 * also don't wait anymore for IRQ context
774 	 * */
775 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
776 		goto failed_kiq_read;
777 
778 	might_sleep();
779 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
780 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
781 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
782 	}
783 
784 	if (cnt > MAX_KIQ_REG_TRY)
785 		goto failed_kiq_read;
786 
787 	mb();
788 	value = adev->wb.wb[reg_val_offs];
789 	amdgpu_device_wb_free(adev, reg_val_offs);
790 	return value;
791 
792 failed_undo:
793 	amdgpu_ring_undo(ring);
794 failed_unlock:
795 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
796 failed_kiq_read:
797 	if (reg_val_offs)
798 		amdgpu_device_wb_free(adev, reg_val_offs);
799 	dev_err(adev->dev, "failed to read reg:%x\n", reg);
800 	return ~0;
801 }
802 
803 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
804 {
805 	signed long r, cnt = 0;
806 	unsigned long flags;
807 	uint32_t seq;
808 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
809 	struct amdgpu_ring *ring = &kiq->ring;
810 
811 	BUG_ON(!ring->funcs->emit_wreg);
812 
813 	if (amdgpu_device_skip_hw_access(adev))
814 		return;
815 
816 	if (adev->mes.ring.sched.ready) {
817 		amdgpu_mes_wreg(adev, reg, v);
818 		return;
819 	}
820 
821 	spin_lock_irqsave(&kiq->ring_lock, flags);
822 	amdgpu_ring_alloc(ring, 32);
823 	amdgpu_ring_emit_wreg(ring, reg, v);
824 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
825 	if (r)
826 		goto failed_undo;
827 
828 	amdgpu_ring_commit(ring);
829 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
830 
831 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
832 
833 	/* don't wait anymore for gpu reset case because this way may
834 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
835 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
836 	 * never return if we keep waiting in virt_kiq_rreg, which cause
837 	 * gpu_recover() hang there.
838 	 *
839 	 * also don't wait anymore for IRQ context
840 	 * */
841 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
842 		goto failed_kiq_write;
843 
844 	might_sleep();
845 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
846 
847 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
848 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
849 	}
850 
851 	if (cnt > MAX_KIQ_REG_TRY)
852 		goto failed_kiq_write;
853 
854 	return;
855 
856 failed_undo:
857 	amdgpu_ring_undo(ring);
858 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
859 failed_kiq_write:
860 	dev_err(adev->dev, "failed to write reg:%x\n", reg);
861 }
862 
863 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
864 {
865 	if (amdgpu_num_kcq == -1) {
866 		return 8;
867 	} else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
868 		dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
869 		return 8;
870 	}
871 	return amdgpu_num_kcq;
872 }
873 
874 void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
875 				  uint32_t ucode_id)
876 {
877 	const struct gfx_firmware_header_v1_0 *cp_hdr;
878 	const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
879 	struct amdgpu_firmware_info *info = NULL;
880 	const struct firmware *ucode_fw;
881 	unsigned int fw_size;
882 
883 	switch (ucode_id) {
884 	case AMDGPU_UCODE_ID_CP_PFP:
885 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
886 			adev->gfx.pfp_fw->data;
887 		adev->gfx.pfp_fw_version =
888 			le32_to_cpu(cp_hdr->header.ucode_version);
889 		adev->gfx.pfp_feature_version =
890 			le32_to_cpu(cp_hdr->ucode_feature_version);
891 		ucode_fw = adev->gfx.pfp_fw;
892 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
893 		break;
894 	case AMDGPU_UCODE_ID_CP_RS64_PFP:
895 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
896 			adev->gfx.pfp_fw->data;
897 		adev->gfx.pfp_fw_version =
898 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
899 		adev->gfx.pfp_feature_version =
900 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
901 		ucode_fw = adev->gfx.pfp_fw;
902 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
903 		break;
904 	case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
905 	case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
906 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
907 			adev->gfx.pfp_fw->data;
908 		ucode_fw = adev->gfx.pfp_fw;
909 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
910 		break;
911 	case AMDGPU_UCODE_ID_CP_ME:
912 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
913 			adev->gfx.me_fw->data;
914 		adev->gfx.me_fw_version =
915 			le32_to_cpu(cp_hdr->header.ucode_version);
916 		adev->gfx.me_feature_version =
917 			le32_to_cpu(cp_hdr->ucode_feature_version);
918 		ucode_fw = adev->gfx.me_fw;
919 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
920 		break;
921 	case AMDGPU_UCODE_ID_CP_RS64_ME:
922 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
923 			adev->gfx.me_fw->data;
924 		adev->gfx.me_fw_version =
925 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
926 		adev->gfx.me_feature_version =
927 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
928 		ucode_fw = adev->gfx.me_fw;
929 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
930 		break;
931 	case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
932 	case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
933 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
934 			adev->gfx.me_fw->data;
935 		ucode_fw = adev->gfx.me_fw;
936 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
937 		break;
938 	case AMDGPU_UCODE_ID_CP_CE:
939 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
940 			adev->gfx.ce_fw->data;
941 		adev->gfx.ce_fw_version =
942 			le32_to_cpu(cp_hdr->header.ucode_version);
943 		adev->gfx.ce_feature_version =
944 			le32_to_cpu(cp_hdr->ucode_feature_version);
945 		ucode_fw = adev->gfx.ce_fw;
946 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
947 		break;
948 	case AMDGPU_UCODE_ID_CP_MEC1:
949 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
950 			adev->gfx.mec_fw->data;
951 		adev->gfx.mec_fw_version =
952 			le32_to_cpu(cp_hdr->header.ucode_version);
953 		adev->gfx.mec_feature_version =
954 			le32_to_cpu(cp_hdr->ucode_feature_version);
955 		ucode_fw = adev->gfx.mec_fw;
956 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
957 			  le32_to_cpu(cp_hdr->jt_size) * 4;
958 		break;
959 	case AMDGPU_UCODE_ID_CP_MEC1_JT:
960 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
961 			adev->gfx.mec_fw->data;
962 		ucode_fw = adev->gfx.mec_fw;
963 		fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
964 		break;
965 	case AMDGPU_UCODE_ID_CP_MEC2:
966 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
967 			adev->gfx.mec2_fw->data;
968 		adev->gfx.mec2_fw_version =
969 			le32_to_cpu(cp_hdr->header.ucode_version);
970 		adev->gfx.mec2_feature_version =
971 			le32_to_cpu(cp_hdr->ucode_feature_version);
972 		ucode_fw = adev->gfx.mec2_fw;
973 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
974 			  le32_to_cpu(cp_hdr->jt_size) * 4;
975 		break;
976 	case AMDGPU_UCODE_ID_CP_MEC2_JT:
977 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
978 			adev->gfx.mec2_fw->data;
979 		ucode_fw = adev->gfx.mec2_fw;
980 		fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
981 		break;
982 	case AMDGPU_UCODE_ID_CP_RS64_MEC:
983 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
984 			adev->gfx.mec_fw->data;
985 		adev->gfx.mec_fw_version =
986 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
987 		adev->gfx.mec_feature_version =
988 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
989 		ucode_fw = adev->gfx.mec_fw;
990 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
991 		break;
992 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
993 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
994 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
995 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
996 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
997 			adev->gfx.mec_fw->data;
998 		ucode_fw = adev->gfx.mec_fw;
999 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1000 		break;
1001 	default:
1002 		break;
1003 	}
1004 
1005 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1006 		info = &adev->firmware.ucode[ucode_id];
1007 		info->ucode_id = ucode_id;
1008 		info->fw = ucode_fw;
1009 		adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE);
1010 	}
1011 }
1012