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 #include "amdgpu_xcp.h"
32 
33 /* delay 0.1 second to enable gfx off feature */
34 #define GFX_OFF_DELAY_ENABLE         msecs_to_jiffies(100)
35 
36 #define GFX_OFF_NO_DELAY 0
37 
38 /*
39  * GPU GFX IP block helpers function.
40  */
41 
amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device * adev,int mec,int pipe,int queue)42 int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
43 				int pipe, int queue)
44 {
45 	int bit = 0;
46 
47 	bit += mec * adev->gfx.mec.num_pipe_per_mec
48 		* adev->gfx.mec.num_queue_per_pipe;
49 	bit += pipe * adev->gfx.mec.num_queue_per_pipe;
50 	bit += queue;
51 
52 	return bit;
53 }
54 
amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device * adev,int bit,int * mec,int * pipe,int * queue)55 void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
56 				 int *mec, int *pipe, int *queue)
57 {
58 	*queue = bit % adev->gfx.mec.num_queue_per_pipe;
59 	*pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
60 		% adev->gfx.mec.num_pipe_per_mec;
61 	*mec = (bit / adev->gfx.mec.num_queue_per_pipe)
62 	       / adev->gfx.mec.num_pipe_per_mec;
63 
64 }
65 
amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device * adev,int xcc_id,int mec,int pipe,int queue)66 bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
67 				     int xcc_id, int mec, int pipe, int queue)
68 {
69 	return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
70 			adev->gfx.mec_bitmap[xcc_id].queue_bitmap);
71 }
72 
amdgpu_gfx_me_queue_to_bit(struct amdgpu_device * adev,int me,int pipe,int queue)73 int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
74 			       int me, int pipe, int queue)
75 {
76 	int bit = 0;
77 
78 	bit += me * adev->gfx.me.num_pipe_per_me
79 		* adev->gfx.me.num_queue_per_pipe;
80 	bit += pipe * adev->gfx.me.num_queue_per_pipe;
81 	bit += queue;
82 
83 	return bit;
84 }
85 
amdgpu_gfx_bit_to_me_queue(struct amdgpu_device * adev,int bit,int * me,int * pipe,int * queue)86 void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
87 				int *me, int *pipe, int *queue)
88 {
89 	*queue = bit % adev->gfx.me.num_queue_per_pipe;
90 	*pipe = (bit / adev->gfx.me.num_queue_per_pipe)
91 		% adev->gfx.me.num_pipe_per_me;
92 	*me = (bit / adev->gfx.me.num_queue_per_pipe)
93 		/ adev->gfx.me.num_pipe_per_me;
94 }
95 
amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device * adev,int me,int pipe,int queue)96 bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
97 				    int me, int pipe, int queue)
98 {
99 	return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
100 			adev->gfx.me.queue_bitmap);
101 }
102 
103 /**
104  * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
105  *
106  * @mask: array in which the per-shader array disable masks will be stored
107  * @max_se: number of SEs
108  * @max_sh: number of SHs
109  *
110  * The bitmask of CUs to be disabled in the shader array determined by se and
111  * sh is stored in mask[se * max_sh + sh].
112  */
amdgpu_gfx_parse_disable_cu(unsigned int * mask,unsigned int max_se,unsigned int max_sh)113 void amdgpu_gfx_parse_disable_cu(unsigned int *mask, unsigned int max_se, unsigned int max_sh)
114 {
115 	unsigned int se, sh, cu;
116 	const char *p;
117 
118 	memset(mask, 0, sizeof(*mask) * max_se * max_sh);
119 
120 	if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
121 		return;
122 
123 	p = amdgpu_disable_cu;
124 	for (;;) {
125 		char *next;
126 		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
127 
128 		if (ret < 3) {
129 			DRM_ERROR("amdgpu: could not parse disable_cu\n");
130 			return;
131 		}
132 
133 		if (se < max_se && sh < max_sh && cu < 16) {
134 			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
135 			mask[se * max_sh + sh] |= 1u << cu;
136 		} else {
137 			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
138 				  se, sh, cu);
139 		}
140 
141 		next = strchr(p, ',');
142 		if (!next)
143 			break;
144 		p = next + 1;
145 	}
146 }
147 
amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device * adev)148 static bool amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device *adev)
149 {
150 	return amdgpu_async_gfx_ring && adev->gfx.me.num_pipe_per_me > 1;
151 }
152 
amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device * adev)153 static bool amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device *adev)
154 {
155 	if (amdgpu_compute_multipipe != -1) {
156 		DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
157 			 amdgpu_compute_multipipe);
158 		return amdgpu_compute_multipipe == 1;
159 	}
160 
161 	if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))
162 		return true;
163 
164 	/* FIXME: spreading the queues across pipes causes perf regressions
165 	 * on POLARIS11 compute workloads */
166 	if (adev->asic_type == CHIP_POLARIS11)
167 		return false;
168 
169 	return adev->gfx.mec.num_mec > 1;
170 }
171 
amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device * adev,struct amdgpu_ring * ring)172 bool amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device *adev,
173 						struct amdgpu_ring *ring)
174 {
175 	int queue = ring->queue;
176 	int pipe = ring->pipe;
177 
178 	/* Policy: use pipe1 queue0 as high priority graphics queue if we
179 	 * have more than one gfx pipe.
180 	 */
181 	if (amdgpu_gfx_is_graphics_multipipe_capable(adev) &&
182 	    adev->gfx.num_gfx_rings > 1 && pipe == 1 && queue == 0) {
183 		int me = ring->me;
184 		int bit;
185 
186 		bit = amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue);
187 		if (ring == &adev->gfx.gfx_ring[bit])
188 			return true;
189 	}
190 
191 	return false;
192 }
193 
amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device * adev,struct amdgpu_ring * ring)194 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
195 					       struct amdgpu_ring *ring)
196 {
197 	/* Policy: use 1st queue as high priority compute queue if we
198 	 * have more than one compute queue.
199 	 */
200 	if (adev->gfx.num_compute_rings > 1 &&
201 	    ring == &adev->gfx.compute_ring[0])
202 		return true;
203 
204 	return false;
205 }
206 
amdgpu_gfx_compute_queue_acquire(struct amdgpu_device * adev)207 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
208 {
209 	int i, j, queue, pipe;
210 	bool multipipe_policy = amdgpu_gfx_is_compute_multipipe_capable(adev);
211 	int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
212 				     adev->gfx.mec.num_queue_per_pipe,
213 				     adev->gfx.num_compute_rings);
214 	int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
215 
216 	if (multipipe_policy) {
217 		/* policy: make queues evenly cross all pipes on MEC1 only
218 		 * for multiple xcc, just use the original policy for simplicity */
219 		for (j = 0; j < num_xcc; j++) {
220 			for (i = 0; i < max_queues_per_mec; i++) {
221 				pipe = i % adev->gfx.mec.num_pipe_per_mec;
222 				queue = (i / adev->gfx.mec.num_pipe_per_mec) %
223 					 adev->gfx.mec.num_queue_per_pipe;
224 
225 				set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
226 					adev->gfx.mec_bitmap[j].queue_bitmap);
227 			}
228 		}
229 	} else {
230 		/* policy: amdgpu owns all queues in the given pipe */
231 		for (j = 0; j < num_xcc; j++) {
232 			for (i = 0; i < max_queues_per_mec; ++i)
233 				set_bit(i, adev->gfx.mec_bitmap[j].queue_bitmap);
234 		}
235 	}
236 
237 	for (j = 0; j < num_xcc; j++) {
238 		dev_dbg(adev->dev, "mec queue bitmap weight=%d\n",
239 			bitmap_weight(adev->gfx.mec_bitmap[j].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
240 	}
241 }
242 
amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device * adev)243 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
244 {
245 	int i, queue, pipe;
246 	bool multipipe_policy = amdgpu_gfx_is_graphics_multipipe_capable(adev);
247 	int max_queues_per_me = adev->gfx.me.num_pipe_per_me *
248 					adev->gfx.me.num_queue_per_pipe;
249 
250 	if (multipipe_policy) {
251 		/* policy: amdgpu owns the first queue per pipe at this stage
252 		 * will extend to mulitple queues per pipe later */
253 		for (i = 0; i < max_queues_per_me; i++) {
254 			pipe = i % adev->gfx.me.num_pipe_per_me;
255 			queue = (i / adev->gfx.me.num_pipe_per_me) %
256 				adev->gfx.me.num_queue_per_pipe;
257 
258 			set_bit(pipe * adev->gfx.me.num_queue_per_pipe + queue,
259 				adev->gfx.me.queue_bitmap);
260 		}
261 	} else {
262 		for (i = 0; i < max_queues_per_me; ++i)
263 			set_bit(i, adev->gfx.me.queue_bitmap);
264 	}
265 
266 	/* update the number of active graphics rings */
267 	adev->gfx.num_gfx_rings =
268 		bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
269 }
270 
amdgpu_gfx_kiq_acquire(struct amdgpu_device * adev,struct amdgpu_ring * ring,int xcc_id)271 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
272 				  struct amdgpu_ring *ring, int xcc_id)
273 {
274 	int queue_bit;
275 	int mec, pipe, queue;
276 
277 	queue_bit = adev->gfx.mec.num_mec
278 		    * adev->gfx.mec.num_pipe_per_mec
279 		    * adev->gfx.mec.num_queue_per_pipe;
280 
281 	while (--queue_bit >= 0) {
282 		if (test_bit(queue_bit, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
283 			continue;
284 
285 		amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
286 
287 		/*
288 		 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
289 		 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
290 		 * only can be issued on queue 0.
291 		 */
292 		if ((mec == 1 && pipe > 1) || queue != 0)
293 			continue;
294 
295 		ring->me = mec + 1;
296 		ring->pipe = pipe;
297 		ring->queue = queue;
298 
299 		return 0;
300 	}
301 
302 	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
303 	return -EINVAL;
304 }
305 
amdgpu_gfx_kiq_init_ring(struct amdgpu_device * adev,struct amdgpu_ring * ring,struct amdgpu_irq_src * irq,int xcc_id)306 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
307 			     struct amdgpu_ring *ring,
308 			     struct amdgpu_irq_src *irq, int xcc_id)
309 {
310 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
311 	int r = 0;
312 
313 	spin_lock_init(&kiq->ring_lock);
314 
315 	ring->adev = NULL;
316 	ring->ring_obj = NULL;
317 	ring->use_doorbell = true;
318 	ring->xcc_id = xcc_id;
319 	ring->vm_hub = AMDGPU_GFXHUB(xcc_id);
320 	ring->doorbell_index =
321 		(adev->doorbell_index.kiq +
322 		 xcc_id * adev->doorbell_index.xcc_doorbell_range)
323 		<< 1;
324 
325 	r = amdgpu_gfx_kiq_acquire(adev, ring, xcc_id);
326 	if (r)
327 		return r;
328 
329 	ring->eop_gpu_addr = kiq->eop_gpu_addr;
330 	ring->no_scheduler = true;
331 	sprintf(ring->name, "kiq_%d.%d.%d.%d", xcc_id, ring->me, ring->pipe, ring->queue);
332 	r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
333 			     AMDGPU_RING_PRIO_DEFAULT, NULL);
334 	if (r)
335 		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
336 
337 	return r;
338 }
339 
amdgpu_gfx_kiq_free_ring(struct amdgpu_ring * ring)340 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
341 {
342 	amdgpu_ring_fini(ring);
343 }
344 
amdgpu_gfx_kiq_fini(struct amdgpu_device * adev,int xcc_id)345 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev, int xcc_id)
346 {
347 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
348 
349 	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
350 }
351 
amdgpu_gfx_kiq_init(struct amdgpu_device * adev,unsigned int hpd_size,int xcc_id)352 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
353 			unsigned int hpd_size, int xcc_id)
354 {
355 	int r;
356 	u32 *hpd;
357 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
358 
359 	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
360 				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
361 				    &kiq->eop_gpu_addr, (void **)&hpd);
362 	if (r) {
363 		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
364 		return r;
365 	}
366 
367 	memset(hpd, 0, hpd_size);
368 
369 	r = amdgpu_bo_reserve(kiq->eop_obj, true);
370 	if (unlikely(r != 0))
371 		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
372 	amdgpu_bo_kunmap(kiq->eop_obj);
373 	amdgpu_bo_unreserve(kiq->eop_obj);
374 
375 	return 0;
376 }
377 
378 /* create MQD for each compute/gfx queue */
amdgpu_gfx_mqd_sw_init(struct amdgpu_device * adev,unsigned int mqd_size,int xcc_id)379 int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
380 			   unsigned int mqd_size, int xcc_id)
381 {
382 	int r, i, j;
383 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
384 	struct amdgpu_ring *ring = &kiq->ring;
385 	u32 domain = AMDGPU_GEM_DOMAIN_GTT;
386 
387 #if !defined(CONFIG_ARM) && !defined(CONFIG_ARM64)
388 	/* Only enable on gfx10 and 11 for now to avoid changing behavior on older chips */
389 	if (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(10, 0, 0))
390 		domain |= AMDGPU_GEM_DOMAIN_VRAM;
391 #endif
392 
393 	/* create MQD for KIQ */
394 	if (!adev->enable_mes_kiq && !ring->mqd_obj) {
395 		/* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
396 		 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
397 		 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
398 		 * KIQ MQD no matter SRIOV or Bare-metal
399 		 */
400 		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
401 					    AMDGPU_GEM_DOMAIN_VRAM |
402 					    AMDGPU_GEM_DOMAIN_GTT,
403 					    &ring->mqd_obj,
404 					    &ring->mqd_gpu_addr,
405 					    &ring->mqd_ptr);
406 		if (r) {
407 			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
408 			return r;
409 		}
410 
411 		/* prepare MQD backup */
412 		kiq->mqd_backup = kmalloc(mqd_size, GFP_KERNEL);
413 		if (!kiq->mqd_backup) {
414 			dev_warn(adev->dev,
415 				 "no memory to create MQD backup for ring %s\n", ring->name);
416 			return -ENOMEM;
417 		}
418 	}
419 
420 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
421 		/* create MQD for each KGQ */
422 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
423 			ring = &adev->gfx.gfx_ring[i];
424 			if (!ring->mqd_obj) {
425 				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
426 							    domain, &ring->mqd_obj,
427 							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
428 				if (r) {
429 					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
430 					return r;
431 				}
432 
433 				ring->mqd_size = mqd_size;
434 				/* prepare MQD backup */
435 				adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
436 				if (!adev->gfx.me.mqd_backup[i]) {
437 					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
438 					return -ENOMEM;
439 				}
440 			}
441 		}
442 	}
443 
444 	/* create MQD for each KCQ */
445 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
446 		j = i + xcc_id * adev->gfx.num_compute_rings;
447 		ring = &adev->gfx.compute_ring[j];
448 		if (!ring->mqd_obj) {
449 			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
450 						    domain, &ring->mqd_obj,
451 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
452 			if (r) {
453 				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
454 				return r;
455 			}
456 
457 			ring->mqd_size = mqd_size;
458 			/* prepare MQD backup */
459 			adev->gfx.mec.mqd_backup[j] = kmalloc(mqd_size, GFP_KERNEL);
460 			if (!adev->gfx.mec.mqd_backup[j]) {
461 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
462 				return -ENOMEM;
463 			}
464 		}
465 	}
466 
467 	return 0;
468 }
469 
amdgpu_gfx_mqd_sw_fini(struct amdgpu_device * adev,int xcc_id)470 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev, int xcc_id)
471 {
472 	struct amdgpu_ring *ring = NULL;
473 	int i, j;
474 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
475 
476 	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
477 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
478 			ring = &adev->gfx.gfx_ring[i];
479 			kfree(adev->gfx.me.mqd_backup[i]);
480 			amdgpu_bo_free_kernel(&ring->mqd_obj,
481 					      &ring->mqd_gpu_addr,
482 					      &ring->mqd_ptr);
483 		}
484 	}
485 
486 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
487 		j = i + xcc_id * adev->gfx.num_compute_rings;
488 		ring = &adev->gfx.compute_ring[j];
489 		kfree(adev->gfx.mec.mqd_backup[j]);
490 		amdgpu_bo_free_kernel(&ring->mqd_obj,
491 				      &ring->mqd_gpu_addr,
492 				      &ring->mqd_ptr);
493 	}
494 
495 	ring = &kiq->ring;
496 	kfree(kiq->mqd_backup);
497 	amdgpu_bo_free_kernel(&ring->mqd_obj,
498 			      &ring->mqd_gpu_addr,
499 			      &ring->mqd_ptr);
500 }
501 
amdgpu_gfx_disable_kcq(struct amdgpu_device * adev,int xcc_id)502 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
503 {
504 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
505 	struct amdgpu_ring *kiq_ring = &kiq->ring;
506 	int i, r = 0;
507 	int j;
508 
509 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
510 		return -EINVAL;
511 
512 	spin_lock(&kiq->ring_lock);
513 	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
514 					adev->gfx.num_compute_rings)) {
515 		spin_unlock(&kiq->ring_lock);
516 		return -ENOMEM;
517 	}
518 
519 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
520 		j = i + xcc_id * adev->gfx.num_compute_rings;
521 		kiq->pmf->kiq_unmap_queues(kiq_ring,
522 					   &adev->gfx.compute_ring[j],
523 					   RESET_QUEUES, 0, 0);
524 	}
525 
526 	if (kiq_ring->sched.ready && !adev->job_hang)
527 		r = amdgpu_ring_test_helper(kiq_ring);
528 	spin_unlock(&kiq->ring_lock);
529 
530 	return r;
531 }
532 
amdgpu_gfx_disable_kgq(struct amdgpu_device * adev,int xcc_id)533 int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id)
534 {
535 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
536 	struct amdgpu_ring *kiq_ring = &kiq->ring;
537 	int i, r = 0;
538 	int j;
539 
540 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
541 		return -EINVAL;
542 
543 	spin_lock(&kiq->ring_lock);
544 	if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
545 		if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
546 						adev->gfx.num_gfx_rings)) {
547 			spin_unlock(&kiq->ring_lock);
548 			return -ENOMEM;
549 		}
550 
551 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
552 			j = i + xcc_id * adev->gfx.num_gfx_rings;
553 			kiq->pmf->kiq_unmap_queues(kiq_ring,
554 						   &adev->gfx.gfx_ring[j],
555 						   PREEMPT_QUEUES, 0, 0);
556 		}
557 	}
558 
559 	if (adev->gfx.kiq[0].ring.sched.ready && !adev->job_hang)
560 		r = amdgpu_ring_test_helper(kiq_ring);
561 	spin_unlock(&kiq->ring_lock);
562 
563 	return r;
564 }
565 
amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device * adev,int queue_bit)566 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
567 					int queue_bit)
568 {
569 	int mec, pipe, queue;
570 	int set_resource_bit = 0;
571 
572 	amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
573 
574 	set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
575 
576 	return set_resource_bit;
577 }
578 
amdgpu_gfx_enable_kcq(struct amdgpu_device * adev,int xcc_id)579 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
580 {
581 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
582 	struct amdgpu_ring *kiq_ring = &kiq->ring;
583 	uint64_t queue_mask = 0;
584 	int r, i, j;
585 
586 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
587 		return -EINVAL;
588 
589 	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
590 		if (!test_bit(i, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
591 			continue;
592 
593 		/* This situation may be hit in the future if a new HW
594 		 * generation exposes more than 64 queues. If so, the
595 		 * definition of queue_mask needs updating */
596 		if (WARN_ON(i > (sizeof(queue_mask)*8))) {
597 			DRM_ERROR("Invalid KCQ enabled: %d\n", i);
598 			break;
599 		}
600 
601 		queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
602 	}
603 
604 	DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
605 							kiq_ring->queue);
606 	amdgpu_device_flush_hdp(adev, NULL);
607 
608 	spin_lock(&kiq->ring_lock);
609 	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
610 					adev->gfx.num_compute_rings +
611 					kiq->pmf->set_resources_size);
612 	if (r) {
613 		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
614 		spin_unlock(&kiq->ring_lock);
615 		return r;
616 	}
617 
618 	if (adev->enable_mes)
619 		queue_mask = ~0ULL;
620 
621 	kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
622 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
623 		j = i + xcc_id * adev->gfx.num_compute_rings;
624 			kiq->pmf->kiq_map_queues(kiq_ring,
625 						 &adev->gfx.compute_ring[j]);
626 	}
627 
628 	r = amdgpu_ring_test_helper(kiq_ring);
629 	spin_unlock(&kiq->ring_lock);
630 	if (r)
631 		DRM_ERROR("KCQ enable failed\n");
632 
633 	return r;
634 }
635 
amdgpu_gfx_enable_kgq(struct amdgpu_device * adev,int xcc_id)636 int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
637 {
638 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
639 	struct amdgpu_ring *kiq_ring = &kiq->ring;
640 	int r, i, j;
641 
642 	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
643 		return -EINVAL;
644 
645 	amdgpu_device_flush_hdp(adev, NULL);
646 
647 	spin_lock(&kiq->ring_lock);
648 	/* No need to map kcq on the slave */
649 	if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
650 		r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
651 						adev->gfx.num_gfx_rings);
652 		if (r) {
653 			DRM_ERROR("Failed to lock KIQ (%d).\n", r);
654 			spin_unlock(&kiq->ring_lock);
655 			return r;
656 		}
657 
658 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
659 			j = i + xcc_id * adev->gfx.num_gfx_rings;
660 			kiq->pmf->kiq_map_queues(kiq_ring,
661 						 &adev->gfx.gfx_ring[j]);
662 		}
663 	}
664 
665 	r = amdgpu_ring_test_helper(kiq_ring);
666 	spin_unlock(&kiq->ring_lock);
667 	if (r)
668 		DRM_ERROR("KCQ enable failed\n");
669 
670 	return r;
671 }
672 
673 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
674  *
675  * @adev: amdgpu_device pointer
676  * @bool enable true: enable gfx off feature, false: disable gfx off feature
677  *
678  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
679  * 2. other client can send request to disable gfx off feature, the request should be honored.
680  * 3. other client can cancel their request of disable gfx off feature
681  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
682  */
683 
amdgpu_gfx_off_ctrl(struct amdgpu_device * adev,bool enable)684 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
685 {
686 	unsigned long delay = GFX_OFF_DELAY_ENABLE;
687 
688 	if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
689 		return;
690 
691 	mutex_lock(&adev->gfx.gfx_off_mutex);
692 
693 	if (enable) {
694 		/* If the count is already 0, it means there's an imbalance bug somewhere.
695 		 * Note that the bug may be in a different caller than the one which triggers the
696 		 * WARN_ON_ONCE.
697 		 */
698 		if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
699 			goto unlock;
700 
701 		adev->gfx.gfx_off_req_count--;
702 
703 		if (adev->gfx.gfx_off_req_count == 0 &&
704 		    !adev->gfx.gfx_off_state) {
705 			/* If going to s2idle, no need to wait */
706 			if (adev->in_s0ix) {
707 				if (!amdgpu_dpm_set_powergating_by_smu(adev,
708 						AMD_IP_BLOCK_TYPE_GFX, true))
709 					adev->gfx.gfx_off_state = true;
710 			} else {
711 				schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
712 					      delay);
713 			}
714 		}
715 	} else {
716 		if (adev->gfx.gfx_off_req_count == 0) {
717 			cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
718 
719 			if (adev->gfx.gfx_off_state &&
720 			    !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
721 				adev->gfx.gfx_off_state = false;
722 
723 				if (adev->gfx.funcs->init_spm_golden) {
724 					dev_dbg(adev->dev,
725 						"GFXOFF is disabled, re-init SPM golden settings\n");
726 					amdgpu_gfx_init_spm_golden(adev);
727 				}
728 			}
729 		}
730 
731 		adev->gfx.gfx_off_req_count++;
732 	}
733 
734 unlock:
735 	mutex_unlock(&adev->gfx.gfx_off_mutex);
736 }
737 
amdgpu_set_gfx_off_residency(struct amdgpu_device * adev,bool value)738 int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
739 {
740 	int r = 0;
741 
742 	mutex_lock(&adev->gfx.gfx_off_mutex);
743 
744 	r = amdgpu_dpm_set_residency_gfxoff(adev, value);
745 
746 	mutex_unlock(&adev->gfx.gfx_off_mutex);
747 
748 	return r;
749 }
750 
amdgpu_get_gfx_off_residency(struct amdgpu_device * adev,u32 * value)751 int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
752 {
753 	int r = 0;
754 
755 	mutex_lock(&adev->gfx.gfx_off_mutex);
756 
757 	r = amdgpu_dpm_get_residency_gfxoff(adev, value);
758 
759 	mutex_unlock(&adev->gfx.gfx_off_mutex);
760 
761 	return r;
762 }
763 
amdgpu_get_gfx_off_entrycount(struct amdgpu_device * adev,u64 * value)764 int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
765 {
766 	int r = 0;
767 
768 	mutex_lock(&adev->gfx.gfx_off_mutex);
769 
770 	r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
771 
772 	mutex_unlock(&adev->gfx.gfx_off_mutex);
773 
774 	return r;
775 }
776 
amdgpu_get_gfx_off_status(struct amdgpu_device * adev,uint32_t * value)777 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
778 {
779 
780 	int r = 0;
781 
782 	mutex_lock(&adev->gfx.gfx_off_mutex);
783 
784 	r = amdgpu_dpm_get_status_gfxoff(adev, value);
785 
786 	mutex_unlock(&adev->gfx.gfx_off_mutex);
787 
788 	return r;
789 }
790 
amdgpu_gfx_ras_late_init(struct amdgpu_device * adev,struct ras_common_if * ras_block)791 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
792 {
793 	int r;
794 
795 	if (amdgpu_ras_is_supported(adev, ras_block->block)) {
796 		if (!amdgpu_persistent_edc_harvesting_supported(adev)) {
797 			r = amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
798 			if (r)
799 				return r;
800 		}
801 
802 		r = amdgpu_ras_block_late_init(adev, ras_block);
803 		if (r)
804 			return r;
805 
806 		if (adev->gfx.cp_ecc_error_irq.funcs) {
807 			r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
808 			if (r)
809 				goto late_fini;
810 		}
811 	} else {
812 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
813 	}
814 
815 	return 0;
816 late_fini:
817 	amdgpu_ras_block_late_fini(adev, ras_block);
818 	return r;
819 }
820 
amdgpu_gfx_ras_sw_init(struct amdgpu_device * adev)821 int amdgpu_gfx_ras_sw_init(struct amdgpu_device *adev)
822 {
823 	int err = 0;
824 	struct amdgpu_gfx_ras *ras = NULL;
825 
826 	/* adev->gfx.ras is NULL, which means gfx does not
827 	 * support ras function, then do nothing here.
828 	 */
829 	if (!adev->gfx.ras)
830 		return 0;
831 
832 	ras = adev->gfx.ras;
833 
834 	err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
835 	if (err) {
836 		dev_err(adev->dev, "Failed to register gfx ras block!\n");
837 		return err;
838 	}
839 
840 	strcpy(ras->ras_block.ras_comm.name, "gfx");
841 	ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__GFX;
842 	ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
843 	adev->gfx.ras_if = &ras->ras_block.ras_comm;
844 
845 	/* If not define special ras_late_init function, use gfx default ras_late_init */
846 	if (!ras->ras_block.ras_late_init)
847 		ras->ras_block.ras_late_init = amdgpu_gfx_ras_late_init;
848 
849 	/* If not defined special ras_cb function, use default ras_cb */
850 	if (!ras->ras_block.ras_cb)
851 		ras->ras_block.ras_cb = amdgpu_gfx_process_ras_data_cb;
852 
853 	return 0;
854 }
855 
amdgpu_gfx_poison_consumption_handler(struct amdgpu_device * adev,struct amdgpu_iv_entry * entry)856 int amdgpu_gfx_poison_consumption_handler(struct amdgpu_device *adev,
857 						struct amdgpu_iv_entry *entry)
858 {
859 	if (adev->gfx.ras && adev->gfx.ras->poison_consumption_handler)
860 		return adev->gfx.ras->poison_consumption_handler(adev, entry);
861 
862 	return 0;
863 }
864 
amdgpu_gfx_process_ras_data_cb(struct amdgpu_device * adev,void * err_data,struct amdgpu_iv_entry * entry)865 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
866 		void *err_data,
867 		struct amdgpu_iv_entry *entry)
868 {
869 	/* TODO ue will trigger an interrupt.
870 	 *
871 	 * When “Full RAS” is enabled, the per-IP interrupt sources should
872 	 * be disabled and the driver should only look for the aggregated
873 	 * interrupt via sync flood
874 	 */
875 	if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
876 		kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
877 		if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
878 		    adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
879 			adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
880 		amdgpu_ras_reset_gpu(adev);
881 	}
882 	return AMDGPU_RAS_SUCCESS;
883 }
884 
amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)885 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
886 				  struct amdgpu_irq_src *source,
887 				  struct amdgpu_iv_entry *entry)
888 {
889 	struct ras_common_if *ras_if = adev->gfx.ras_if;
890 	struct ras_dispatch_if ih_data = {
891 		.entry = entry,
892 	};
893 
894 	if (!ras_if)
895 		return 0;
896 
897 	ih_data.head = *ras_if;
898 
899 	DRM_ERROR("CP ECC ERROR IRQ\n");
900 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
901 	return 0;
902 }
903 
amdgpu_gfx_ras_error_func(struct amdgpu_device * adev,void * ras_error_status,void (* func)(struct amdgpu_device * adev,void * ras_error_status,int xcc_id))904 void amdgpu_gfx_ras_error_func(struct amdgpu_device *adev,
905 		void *ras_error_status,
906 		void (*func)(struct amdgpu_device *adev, void *ras_error_status,
907 				int xcc_id))
908 {
909 	int i;
910 	int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
911 	uint32_t xcc_mask = GENMASK(num_xcc - 1, 0);
912 	struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
913 
914 	if (err_data) {
915 		err_data->ue_count = 0;
916 		err_data->ce_count = 0;
917 	}
918 
919 	for_each_inst(i, xcc_mask)
920 		func(adev, ras_error_status, i);
921 }
922 
amdgpu_kiq_rreg(struct amdgpu_device * adev,uint32_t reg)923 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
924 {
925 	signed long r, cnt = 0;
926 	unsigned long flags;
927 	uint32_t seq, reg_val_offs = 0, value = 0;
928 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
929 	struct amdgpu_ring *ring = &kiq->ring;
930 
931 	if (amdgpu_device_skip_hw_access(adev))
932 		return 0;
933 
934 	if (adev->mes.ring.sched.ready)
935 		return amdgpu_mes_rreg(adev, reg);
936 
937 	BUG_ON(!ring->funcs->emit_rreg);
938 
939 	spin_lock_irqsave(&kiq->ring_lock, flags);
940 	if (amdgpu_device_wb_get(adev, &reg_val_offs)) {
941 		pr_err("critical bug! too many kiq readers\n");
942 		goto failed_unlock;
943 	}
944 	r = amdgpu_ring_alloc(ring, 32);
945 	if (r)
946 		goto failed_unlock;
947 
948 	amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
949 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
950 	if (r)
951 		goto failed_undo;
952 
953 	amdgpu_ring_commit(ring);
954 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
955 
956 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
957 
958 	/* don't wait anymore for gpu reset case because this way may
959 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
960 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
961 	 * never return if we keep waiting in virt_kiq_rreg, which cause
962 	 * gpu_recover() hang there.
963 	 *
964 	 * also don't wait anymore for IRQ context
965 	 * */
966 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
967 		goto failed_kiq_read;
968 
969 	might_sleep();
970 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
971 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
972 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
973 	}
974 
975 	if (cnt > MAX_KIQ_REG_TRY)
976 		goto failed_kiq_read;
977 
978 	mb();
979 	value = adev->wb.wb[reg_val_offs];
980 	amdgpu_device_wb_free(adev, reg_val_offs);
981 	return value;
982 
983 failed_undo:
984 	amdgpu_ring_undo(ring);
985 failed_unlock:
986 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
987 failed_kiq_read:
988 	if (reg_val_offs)
989 		amdgpu_device_wb_free(adev, reg_val_offs);
990 	dev_err(adev->dev, "failed to read reg:%x\n", reg);
991 	return ~0;
992 }
993 
amdgpu_kiq_wreg(struct amdgpu_device * adev,uint32_t reg,uint32_t v)994 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
995 {
996 	signed long r, cnt = 0;
997 	unsigned long flags;
998 	uint32_t seq;
999 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
1000 	struct amdgpu_ring *ring = &kiq->ring;
1001 
1002 	BUG_ON(!ring->funcs->emit_wreg);
1003 
1004 	if (amdgpu_device_skip_hw_access(adev))
1005 		return;
1006 
1007 	if (adev->mes.ring.sched.ready) {
1008 		amdgpu_mes_wreg(adev, reg, v);
1009 		return;
1010 	}
1011 
1012 	spin_lock_irqsave(&kiq->ring_lock, flags);
1013 	r = amdgpu_ring_alloc(ring, 32);
1014 	if (r)
1015 		goto failed_unlock;
1016 
1017 	amdgpu_ring_emit_wreg(ring, reg, v);
1018 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
1019 	if (r)
1020 		goto failed_undo;
1021 
1022 	amdgpu_ring_commit(ring);
1023 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
1024 
1025 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1026 
1027 	/* don't wait anymore for gpu reset case because this way may
1028 	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
1029 	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
1030 	 * never return if we keep waiting in virt_kiq_rreg, which cause
1031 	 * gpu_recover() hang there.
1032 	 *
1033 	 * also don't wait anymore for IRQ context
1034 	 * */
1035 	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
1036 		goto failed_kiq_write;
1037 
1038 	might_sleep();
1039 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
1040 
1041 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
1042 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1043 	}
1044 
1045 	if (cnt > MAX_KIQ_REG_TRY)
1046 		goto failed_kiq_write;
1047 
1048 	return;
1049 
1050 failed_undo:
1051 	amdgpu_ring_undo(ring);
1052 failed_unlock:
1053 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
1054 failed_kiq_write:
1055 	dev_err(adev->dev, "failed to write reg:%x\n", reg);
1056 }
1057 
amdgpu_gfx_get_num_kcq(struct amdgpu_device * adev)1058 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
1059 {
1060 	if (amdgpu_num_kcq == -1) {
1061 		return 8;
1062 	} else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
1063 		dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
1064 		return 8;
1065 	}
1066 	return amdgpu_num_kcq;
1067 }
1068 
amdgpu_gfx_cp_init_microcode(struct amdgpu_device * adev,uint32_t ucode_id)1069 void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
1070 				  uint32_t ucode_id)
1071 {
1072 	const struct gfx_firmware_header_v1_0 *cp_hdr;
1073 	const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
1074 	struct amdgpu_firmware_info *info = NULL;
1075 	const struct firmware *ucode_fw;
1076 	unsigned int fw_size;
1077 
1078 	switch (ucode_id) {
1079 	case AMDGPU_UCODE_ID_CP_PFP:
1080 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1081 			adev->gfx.pfp_fw->data;
1082 		adev->gfx.pfp_fw_version =
1083 			le32_to_cpu(cp_hdr->header.ucode_version);
1084 		adev->gfx.pfp_feature_version =
1085 			le32_to_cpu(cp_hdr->ucode_feature_version);
1086 		ucode_fw = adev->gfx.pfp_fw;
1087 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1088 		break;
1089 	case AMDGPU_UCODE_ID_CP_RS64_PFP:
1090 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1091 			adev->gfx.pfp_fw->data;
1092 		adev->gfx.pfp_fw_version =
1093 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1094 		adev->gfx.pfp_feature_version =
1095 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1096 		ucode_fw = adev->gfx.pfp_fw;
1097 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1098 		break;
1099 	case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
1100 	case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
1101 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1102 			adev->gfx.pfp_fw->data;
1103 		ucode_fw = adev->gfx.pfp_fw;
1104 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1105 		break;
1106 	case AMDGPU_UCODE_ID_CP_ME:
1107 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1108 			adev->gfx.me_fw->data;
1109 		adev->gfx.me_fw_version =
1110 			le32_to_cpu(cp_hdr->header.ucode_version);
1111 		adev->gfx.me_feature_version =
1112 			le32_to_cpu(cp_hdr->ucode_feature_version);
1113 		ucode_fw = adev->gfx.me_fw;
1114 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1115 		break;
1116 	case AMDGPU_UCODE_ID_CP_RS64_ME:
1117 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1118 			adev->gfx.me_fw->data;
1119 		adev->gfx.me_fw_version =
1120 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1121 		adev->gfx.me_feature_version =
1122 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1123 		ucode_fw = adev->gfx.me_fw;
1124 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1125 		break;
1126 	case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
1127 	case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
1128 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1129 			adev->gfx.me_fw->data;
1130 		ucode_fw = adev->gfx.me_fw;
1131 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1132 		break;
1133 	case AMDGPU_UCODE_ID_CP_CE:
1134 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1135 			adev->gfx.ce_fw->data;
1136 		adev->gfx.ce_fw_version =
1137 			le32_to_cpu(cp_hdr->header.ucode_version);
1138 		adev->gfx.ce_feature_version =
1139 			le32_to_cpu(cp_hdr->ucode_feature_version);
1140 		ucode_fw = adev->gfx.ce_fw;
1141 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1142 		break;
1143 	case AMDGPU_UCODE_ID_CP_MEC1:
1144 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1145 			adev->gfx.mec_fw->data;
1146 		adev->gfx.mec_fw_version =
1147 			le32_to_cpu(cp_hdr->header.ucode_version);
1148 		adev->gfx.mec_feature_version =
1149 			le32_to_cpu(cp_hdr->ucode_feature_version);
1150 		ucode_fw = adev->gfx.mec_fw;
1151 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1152 			  le32_to_cpu(cp_hdr->jt_size) * 4;
1153 		break;
1154 	case AMDGPU_UCODE_ID_CP_MEC1_JT:
1155 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1156 			adev->gfx.mec_fw->data;
1157 		ucode_fw = adev->gfx.mec_fw;
1158 		fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1159 		break;
1160 	case AMDGPU_UCODE_ID_CP_MEC2:
1161 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1162 			adev->gfx.mec2_fw->data;
1163 		adev->gfx.mec2_fw_version =
1164 			le32_to_cpu(cp_hdr->header.ucode_version);
1165 		adev->gfx.mec2_feature_version =
1166 			le32_to_cpu(cp_hdr->ucode_feature_version);
1167 		ucode_fw = adev->gfx.mec2_fw;
1168 		fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1169 			  le32_to_cpu(cp_hdr->jt_size) * 4;
1170 		break;
1171 	case AMDGPU_UCODE_ID_CP_MEC2_JT:
1172 		cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1173 			adev->gfx.mec2_fw->data;
1174 		ucode_fw = adev->gfx.mec2_fw;
1175 		fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1176 		break;
1177 	case AMDGPU_UCODE_ID_CP_RS64_MEC:
1178 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1179 			adev->gfx.mec_fw->data;
1180 		adev->gfx.mec_fw_version =
1181 			le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1182 		adev->gfx.mec_feature_version =
1183 			le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1184 		ucode_fw = adev->gfx.mec_fw;
1185 		fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1186 		break;
1187 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
1188 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
1189 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
1190 	case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
1191 		cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1192 			adev->gfx.mec_fw->data;
1193 		ucode_fw = adev->gfx.mec_fw;
1194 		fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1195 		break;
1196 	default:
1197 		dev_err(adev->dev, "Invalid ucode id %u\n", ucode_id);
1198 		return;
1199 	}
1200 
1201 	if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1202 		info = &adev->firmware.ucode[ucode_id];
1203 		info->ucode_id = ucode_id;
1204 		info->fw = ucode_fw;
1205 		adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE);
1206 	}
1207 }
1208 
amdgpu_gfx_is_master_xcc(struct amdgpu_device * adev,int xcc_id)1209 bool amdgpu_gfx_is_master_xcc(struct amdgpu_device *adev, int xcc_id)
1210 {
1211 	return !(xcc_id % (adev->gfx.num_xcc_per_xcp ?
1212 			adev->gfx.num_xcc_per_xcp : 1));
1213 }
1214 
amdgpu_gfx_get_current_compute_partition(struct device * dev,struct device_attribute * addr,char * buf)1215 static ssize_t amdgpu_gfx_get_current_compute_partition(struct device *dev,
1216 						struct device_attribute *addr,
1217 						char *buf)
1218 {
1219 	struct drm_device *ddev = dev_get_drvdata(dev);
1220 	struct amdgpu_device *adev = drm_to_adev(ddev);
1221 	int mode;
1222 
1223 	mode = amdgpu_xcp_query_partition_mode(adev->xcp_mgr,
1224 					       AMDGPU_XCP_FL_NONE);
1225 
1226 	return sysfs_emit(buf, "%s\n", amdgpu_gfx_compute_mode_desc(mode));
1227 }
1228 
amdgpu_gfx_set_compute_partition(struct device * dev,struct device_attribute * addr,const char * buf,size_t count)1229 static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev,
1230 						struct device_attribute *addr,
1231 						const char *buf, size_t count)
1232 {
1233 	struct drm_device *ddev = dev_get_drvdata(dev);
1234 	struct amdgpu_device *adev = drm_to_adev(ddev);
1235 	enum amdgpu_gfx_partition mode;
1236 	int ret = 0, num_xcc;
1237 
1238 	num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1239 	if (num_xcc % 2 != 0)
1240 		return -EINVAL;
1241 
1242 	if (!strncasecmp("SPX", buf, strlen("SPX"))) {
1243 		mode = AMDGPU_SPX_PARTITION_MODE;
1244 	} else if (!strncasecmp("DPX", buf, strlen("DPX"))) {
1245 		/*
1246 		 * DPX mode needs AIDs to be in multiple of 2.
1247 		 * Each AID connects 2 XCCs.
1248 		 */
1249 		if (num_xcc%4)
1250 			return -EINVAL;
1251 		mode = AMDGPU_DPX_PARTITION_MODE;
1252 	} else if (!strncasecmp("TPX", buf, strlen("TPX"))) {
1253 		if (num_xcc != 6)
1254 			return -EINVAL;
1255 		mode = AMDGPU_TPX_PARTITION_MODE;
1256 	} else if (!strncasecmp("QPX", buf, strlen("QPX"))) {
1257 		if (num_xcc != 8)
1258 			return -EINVAL;
1259 		mode = AMDGPU_QPX_PARTITION_MODE;
1260 	} else if (!strncasecmp("CPX", buf, strlen("CPX"))) {
1261 		mode = AMDGPU_CPX_PARTITION_MODE;
1262 	} else {
1263 		return -EINVAL;
1264 	}
1265 
1266 	ret = amdgpu_xcp_switch_partition_mode(adev->xcp_mgr, mode);
1267 
1268 	if (ret)
1269 		return ret;
1270 
1271 	return count;
1272 }
1273 
amdgpu_gfx_get_available_compute_partition(struct device * dev,struct device_attribute * addr,char * buf)1274 static ssize_t amdgpu_gfx_get_available_compute_partition(struct device *dev,
1275 						struct device_attribute *addr,
1276 						char *buf)
1277 {
1278 	struct drm_device *ddev = dev_get_drvdata(dev);
1279 	struct amdgpu_device *adev = drm_to_adev(ddev);
1280 	char *supported_partition;
1281 
1282 	/* TBD */
1283 	switch (NUM_XCC(adev->gfx.xcc_mask)) {
1284 	case 8:
1285 		supported_partition = "SPX, DPX, QPX, CPX";
1286 		break;
1287 	case 6:
1288 		supported_partition = "SPX, TPX, CPX";
1289 		break;
1290 	case 4:
1291 		supported_partition = "SPX, DPX, CPX";
1292 		break;
1293 	/* this seems only existing in emulation phase */
1294 	case 2:
1295 		supported_partition = "SPX, CPX";
1296 		break;
1297 	default:
1298 		supported_partition = "Not supported";
1299 		break;
1300 	}
1301 
1302 	return sysfs_emit(buf, "%s\n", supported_partition);
1303 }
1304 
1305 static DEVICE_ATTR(current_compute_partition, 0644,
1306 		   amdgpu_gfx_get_current_compute_partition,
1307 		   amdgpu_gfx_set_compute_partition);
1308 
1309 static DEVICE_ATTR(available_compute_partition, 0444,
1310 		   amdgpu_gfx_get_available_compute_partition, NULL);
1311 
amdgpu_gfx_sysfs_init(struct amdgpu_device * adev)1312 int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev)
1313 {
1314 	int r;
1315 
1316 	r = device_create_file(adev->dev, &dev_attr_current_compute_partition);
1317 	if (r)
1318 		return r;
1319 
1320 	r = device_create_file(adev->dev, &dev_attr_available_compute_partition);
1321 
1322 	return r;
1323 }
1324 
amdgpu_gfx_sysfs_fini(struct amdgpu_device * adev)1325 void amdgpu_gfx_sysfs_fini(struct amdgpu_device *adev)
1326 {
1327 	device_remove_file(adev->dev, &dev_attr_current_compute_partition);
1328 	device_remove_file(adev->dev, &dev_attr_available_compute_partition);
1329 }
1330