Linux Audio

Check our new training course

Yocto / OpenEmbedded training

Feb 10-13, 2025
Register
Loading...
v5.4
  1/*
  2 * Copyright 2014 Advanced Micro Devices, Inc.
  3 * Copyright 2008 Red Hat Inc.
  4 * Copyright 2009 Jerome Glisse.
  5 *
  6 * Permission is hereby granted, free of charge, to any person obtaining a
  7 * copy of this software and associated documentation files (the "Software"),
  8 * to deal in the Software without restriction, including without limitation
  9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 10 * and/or sell copies of the Software, and to permit persons to whom the
 11 * Software is furnished to do so, subject to the following conditions:
 12 *
 13 * The above copyright notice and this permission notice shall be included in
 14 * all copies or substantial portions of the Software.
 15 *
 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 22 * OTHER DEALINGS IN THE SOFTWARE.
 23 *
 24 */
 25
 26#include "amdgpu.h"
 27#include "amdgpu_gfx.h"
 28#include "amdgpu_rlc.h"
 
 29
 30/* delay 0.1 second to enable gfx off feature */
 31#define GFX_OFF_DELAY_ENABLE         msecs_to_jiffies(100)
 32
 
 
 33/*
 34 * GPU GFX IP block helpers function.
 35 */
 36
 37int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
 38				int pipe, int queue)
 39{
 40	int bit = 0;
 41
 42	bit += mec * adev->gfx.mec.num_pipe_per_mec
 43		* adev->gfx.mec.num_queue_per_pipe;
 44	bit += pipe * adev->gfx.mec.num_queue_per_pipe;
 45	bit += queue;
 46
 47	return bit;
 48}
 49
 50void amdgpu_gfx_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
 51				 int *mec, int *pipe, int *queue)
 52{
 53	*queue = bit % adev->gfx.mec.num_queue_per_pipe;
 54	*pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
 55		% adev->gfx.mec.num_pipe_per_mec;
 56	*mec = (bit / adev->gfx.mec.num_queue_per_pipe)
 57	       / adev->gfx.mec.num_pipe_per_mec;
 58
 59}
 60
 61bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
 62				     int mec, int pipe, int queue)
 63{
 64	return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
 65			adev->gfx.mec.queue_bitmap);
 66}
 67
 68int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
 69			       int me, int pipe, int queue)
 70{
 71	int bit = 0;
 72
 73	bit += me * adev->gfx.me.num_pipe_per_me
 74		* adev->gfx.me.num_queue_per_pipe;
 75	bit += pipe * adev->gfx.me.num_queue_per_pipe;
 76	bit += queue;
 77
 78	return bit;
 79}
 80
 81void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
 82				int *me, int *pipe, int *queue)
 83{
 84	*queue = bit % adev->gfx.me.num_queue_per_pipe;
 85	*pipe = (bit / adev->gfx.me.num_queue_per_pipe)
 86		% adev->gfx.me.num_pipe_per_me;
 87	*me = (bit / adev->gfx.me.num_queue_per_pipe)
 88		/ adev->gfx.me.num_pipe_per_me;
 89}
 90
 91bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
 92				    int me, int pipe, int queue)
 93{
 94	return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
 95			adev->gfx.me.queue_bitmap);
 96}
 97
 98/**
 99 * amdgpu_gfx_scratch_get - Allocate a scratch register
100 *
101 * @adev: amdgpu_device pointer
102 * @reg: scratch register mmio offset
103 *
104 * Allocate a CP scratch register for use by the driver (all asics).
105 * Returns 0 on success or -EINVAL on failure.
106 */
107int amdgpu_gfx_scratch_get(struct amdgpu_device *adev, uint32_t *reg)
108{
109	int i;
110
111	i = ffs(adev->gfx.scratch.free_mask);
112	if (i != 0 && i <= adev->gfx.scratch.num_reg) {
113		i--;
114		adev->gfx.scratch.free_mask &= ~(1u << i);
115		*reg = adev->gfx.scratch.reg_base + i;
116		return 0;
117	}
118	return -EINVAL;
119}
120
121/**
122 * amdgpu_gfx_scratch_free - Free a scratch register
123 *
124 * @adev: amdgpu_device pointer
125 * @reg: scratch register mmio offset
126 *
127 * Free a CP scratch register allocated for use by the driver (all asics)
128 */
129void amdgpu_gfx_scratch_free(struct amdgpu_device *adev, uint32_t reg)
130{
131	adev->gfx.scratch.free_mask |= 1u << (reg - adev->gfx.scratch.reg_base);
132}
133
134/**
135 * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
136 *
137 * @mask: array in which the per-shader array disable masks will be stored
138 * @max_se: number of SEs
139 * @max_sh: number of SHs
140 *
141 * The bitmask of CUs to be disabled in the shader array determined by se and
142 * sh is stored in mask[se * max_sh + sh].
143 */
144void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
145{
146	unsigned se, sh, cu;
147	const char *p;
148
149	memset(mask, 0, sizeof(*mask) * max_se * max_sh);
150
151	if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
152		return;
153
154	p = amdgpu_disable_cu;
155	for (;;) {
156		char *next;
157		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
158		if (ret < 3) {
159			DRM_ERROR("amdgpu: could not parse disable_cu\n");
160			return;
161		}
162
163		if (se < max_se && sh < max_sh && cu < 16) {
164			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
165			mask[se * max_sh + sh] |= 1u << cu;
166		} else {
167			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
168				  se, sh, cu);
169		}
170
171		next = strchr(p, ',');
172		if (!next)
173			break;
174		p = next + 1;
175	}
176}
177
178static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
179{
180	if (amdgpu_compute_multipipe != -1) {
181		DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
182			 amdgpu_compute_multipipe);
183		return amdgpu_compute_multipipe == 1;
184	}
185
186	/* FIXME: spreading the queues across pipes causes perf regressions
187	 * on POLARIS11 compute workloads */
188	if (adev->asic_type == CHIP_POLARIS11)
189		return false;
190
191	return adev->gfx.mec.num_mec > 1;
192}
193
194void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
 
195{
196	int i, queue, pipe, mec;
197	bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev);
 
 
 
 
198
199	/* policy for amdgpu compute queue ownership */
200	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
201		queue = i % adev->gfx.mec.num_queue_per_pipe;
202		pipe = (i / adev->gfx.mec.num_queue_per_pipe)
203			% adev->gfx.mec.num_pipe_per_mec;
204		mec = (i / adev->gfx.mec.num_queue_per_pipe)
205			/ adev->gfx.mec.num_pipe_per_mec;
206
207		/* we've run out of HW */
208		if (mec >= adev->gfx.mec.num_mec)
209			break;
 
 
 
 
 
 
 
 
 
 
 
210
211		if (multipipe_policy) {
212			/* policy: amdgpu owns the first two queues of the first MEC */
213			if (mec == 0 && queue < 2)
214				set_bit(i, adev->gfx.mec.queue_bitmap);
215		} else {
216			/* policy: amdgpu owns all queues in the first pipe */
217			if (mec == 0 && pipe == 0)
218				set_bit(i, adev->gfx.mec.queue_bitmap);
219		}
 
 
 
 
220	}
221
222	/* update the number of active compute rings */
223	adev->gfx.num_compute_rings =
224		bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
225
226	/* If you hit this case and edited the policy, you probably just
227	 * need to increase AMDGPU_MAX_COMPUTE_RINGS */
228	if (WARN_ON(adev->gfx.num_compute_rings > AMDGPU_MAX_COMPUTE_RINGS))
229		adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
230}
231
232void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
233{
234	int i, queue, pipe, me;
235
236	for (i = 0; i < AMDGPU_MAX_GFX_QUEUES; ++i) {
237		queue = i % adev->gfx.me.num_queue_per_pipe;
238		pipe = (i / adev->gfx.me.num_queue_per_pipe)
239			% adev->gfx.me.num_pipe_per_me;
240		me = (i / adev->gfx.me.num_queue_per_pipe)
241		      / adev->gfx.me.num_pipe_per_me;
242
243		if (me >= adev->gfx.me.num_me)
244			break;
245		/* policy: amdgpu owns the first queue per pipe at this stage
246		 * will extend to mulitple queues per pipe later */
247		if (me == 0 && queue < 1)
248			set_bit(i, adev->gfx.me.queue_bitmap);
249	}
250
251	/* update the number of active graphics rings */
252	adev->gfx.num_gfx_rings =
253		bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
254}
255
256static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
257				  struct amdgpu_ring *ring)
258{
259	int queue_bit;
260	int mec, pipe, queue;
261
262	queue_bit = adev->gfx.mec.num_mec
263		    * adev->gfx.mec.num_pipe_per_mec
264		    * adev->gfx.mec.num_queue_per_pipe;
265
266	while (queue_bit-- >= 0) {
267		if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
268			continue;
269
270		amdgpu_gfx_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
271
272		/*
273		 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
274		 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
275		 * only can be issued on queue 0.
276		 */
277		if ((mec == 1 && pipe > 1) || queue != 0)
278			continue;
279
280		ring->me = mec + 1;
281		ring->pipe = pipe;
282		ring->queue = queue;
283
284		return 0;
285	}
286
287	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
288	return -EINVAL;
289}
290
291int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
292			     struct amdgpu_ring *ring,
293			     struct amdgpu_irq_src *irq)
294{
295	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
296	int r = 0;
297
298	spin_lock_init(&kiq->ring_lock);
299
300	r = amdgpu_device_wb_get(adev, &adev->virt.reg_val_offs);
301	if (r)
302		return r;
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	sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
315	r = amdgpu_ring_init(adev, ring, 1024,
316			     irq, AMDGPU_CP_KIQ_IRQ_DRIVER0);
317	if (r)
318		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
319
320	return r;
321}
322
323void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring,
324			      struct amdgpu_irq_src *irq)
325{
326	amdgpu_device_wb_free(ring->adev, ring->adev->virt.reg_val_offs);
327	amdgpu_ring_fini(ring);
328}
329
330void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
331{
332	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
333
334	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
335}
336
337int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
338			unsigned hpd_size)
339{
340	int r;
341	u32 *hpd;
342	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
343
344	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
345				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
346				    &kiq->eop_gpu_addr, (void **)&hpd);
347	if (r) {
348		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
349		return r;
350	}
351
352	memset(hpd, 0, hpd_size);
353
354	r = amdgpu_bo_reserve(kiq->eop_obj, true);
355	if (unlikely(r != 0))
356		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
357	amdgpu_bo_kunmap(kiq->eop_obj);
358	amdgpu_bo_unreserve(kiq->eop_obj);
359
360	return 0;
361}
362
363/* create MQD for each compute/gfx queue */
364int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
365			   unsigned mqd_size)
366{
367	struct amdgpu_ring *ring = NULL;
368	int r, i;
369
370	/* create MQD for KIQ */
371	ring = &adev->gfx.kiq.ring;
372	if (!ring->mqd_obj) {
373		/* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
374		 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
375		 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
376		 * KIQ MQD no matter SRIOV or Bare-metal
377		 */
378		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
379					    AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj,
380					    &ring->mqd_gpu_addr, &ring->mqd_ptr);
381		if (r) {
382			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
383			return r;
384		}
385
386		/* prepare MQD backup */
387		adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
388		if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
389				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
390	}
391
392	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
393		/* create MQD for each KGQ */
394		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
395			ring = &adev->gfx.gfx_ring[i];
396			if (!ring->mqd_obj) {
397				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
398							    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
399							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
400				if (r) {
401					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
402					return r;
403				}
404
405				/* prepare MQD backup */
406				adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
407				if (!adev->gfx.me.mqd_backup[i])
408					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
409			}
410		}
411	}
412
413	/* create MQD for each KCQ */
414	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
415		ring = &adev->gfx.compute_ring[i];
416		if (!ring->mqd_obj) {
417			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
418						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
419						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
420			if (r) {
421				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
422				return r;
423			}
424
425			/* prepare MQD backup */
426			adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
427			if (!adev->gfx.mec.mqd_backup[i])
428				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
429		}
430	}
431
432	return 0;
433}
434
435void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev)
436{
437	struct amdgpu_ring *ring = NULL;
438	int i;
439
440	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
441		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
442			ring = &adev->gfx.gfx_ring[i];
443			kfree(adev->gfx.me.mqd_backup[i]);
444			amdgpu_bo_free_kernel(&ring->mqd_obj,
445					      &ring->mqd_gpu_addr,
446					      &ring->mqd_ptr);
447		}
448	}
449
450	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
451		ring = &adev->gfx.compute_ring[i];
452		kfree(adev->gfx.mec.mqd_backup[i]);
453		amdgpu_bo_free_kernel(&ring->mqd_obj,
454				      &ring->mqd_gpu_addr,
455				      &ring->mqd_ptr);
456	}
457
458	ring = &adev->gfx.kiq.ring;
459	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring)
460		kfree(adev->gfx.me.mqd_backup[AMDGPU_MAX_GFX_RINGS]);
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
467int 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;
472
473	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
474		return -EINVAL;
475
 
476	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
477					adev->gfx.num_compute_rings))
 
478		return -ENOMEM;
 
479
480	for (i = 0; i < adev->gfx.num_compute_rings; i++)
481		kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i],
482					   RESET_QUEUES, 0, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
483
484	return amdgpu_ring_test_ring(kiq_ring);
 
 
485}
486
487int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev)
488{
489	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
490	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
491	uint64_t queue_mask = 0;
492	int r, i;
493
494	if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
495		return -EINVAL;
496
497	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
498		if (!test_bit(i, adev->gfx.mec.queue_bitmap))
499			continue;
500
501		/* This situation may be hit in the future if a new HW
502		 * generation exposes more than 64 queues. If so, the
503		 * definition of queue_mask needs updating */
504		if (WARN_ON(i > (sizeof(queue_mask)*8))) {
505			DRM_ERROR("Invalid KCQ enabled: %d\n", i);
506			break;
507		}
508
509		queue_mask |= (1ull << i);
510	}
511
512	DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
513							kiq_ring->queue);
514
515	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
516					adev->gfx.num_compute_rings +
517					kiq->pmf->set_resources_size);
518	if (r) {
519		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
 
520		return r;
521	}
522
523	kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
524	for (i = 0; i < adev->gfx.num_compute_rings; i++)
525		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
526
527	r = amdgpu_ring_test_helper(kiq_ring);
 
528	if (r)
529		DRM_ERROR("KCQ enable failed\n");
530
531	return r;
532}
533
534/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
535 *
536 * @adev: amdgpu_device pointer
537 * @bool enable true: enable gfx off feature, false: disable gfx off feature
538 *
539 * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
540 * 2. other client can send request to disable gfx off feature, the request should be honored.
541 * 3. other client can cancel their request of disable gfx off feature
542 * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
543 */
544
545void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
546{
547	if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
548		return;
549
550	if (!is_support_sw_smu(adev) &&
551	    (!adev->powerplay.pp_funcs ||
552	     !adev->powerplay.pp_funcs->set_powergating_by_smu))
553		return;
554
555
556	mutex_lock(&adev->gfx.gfx_off_mutex);
557
558	if (!enable)
559		adev->gfx.gfx_off_req_count++;
560	else if (adev->gfx.gfx_off_req_count > 0)
 
 
 
 
 
561		adev->gfx.gfx_off_req_count--;
562
563	if (enable && !adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
564		schedule_delayed_work(&adev->gfx.gfx_off_delay_work, GFX_OFF_DELAY_ENABLE);
565	} else if (!enable && adev->gfx.gfx_off_state) {
566		if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false))
567			adev->gfx.gfx_off_state = false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
568	}
569
 
570	mutex_unlock(&adev->gfx.gfx_off_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
571}
v5.14.15
  1/*
  2 * Copyright 2014 Advanced Micro Devices, Inc.
  3 * Copyright 2008 Red Hat Inc.
  4 * Copyright 2009 Jerome Glisse.
  5 *
  6 * Permission is hereby granted, free of charge, to any person obtaining a
  7 * copy of this software and associated documentation files (the "Software"),
  8 * to deal in the Software without restriction, including without limitation
  9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 10 * and/or sell copies of the Software, and to permit persons to whom the
 11 * Software is furnished to do so, subject to the following conditions:
 12 *
 13 * The above copyright notice and this permission notice shall be included in
 14 * all copies or substantial portions of the Software.
 15 *
 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 22 * OTHER DEALINGS IN THE SOFTWARE.
 23 *
 24 */
 25
 26#include "amdgpu.h"
 27#include "amdgpu_gfx.h"
 28#include "amdgpu_rlc.h"
 29#include "amdgpu_ras.h"
 30
 31/* delay 0.1 second to enable gfx off feature */
 32#define GFX_OFF_DELAY_ENABLE         msecs_to_jiffies(100)
 33
 34#define GFX_OFF_NO_DELAY 0
 35
 36/*
 37 * GPU GFX IP block helpers function.
 38 */
 39
 40int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
 41				int pipe, int queue)
 42{
 43	int bit = 0;
 44
 45	bit += mec * adev->gfx.mec.num_pipe_per_mec
 46		* adev->gfx.mec.num_queue_per_pipe;
 47	bit += pipe * adev->gfx.mec.num_queue_per_pipe;
 48	bit += queue;
 49
 50	return bit;
 51}
 52
 53void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
 54				 int *mec, int *pipe, int *queue)
 55{
 56	*queue = bit % adev->gfx.mec.num_queue_per_pipe;
 57	*pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
 58		% adev->gfx.mec.num_pipe_per_mec;
 59	*mec = (bit / adev->gfx.mec.num_queue_per_pipe)
 60	       / adev->gfx.mec.num_pipe_per_mec;
 61
 62}
 63
 64bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
 65				     int mec, int pipe, int queue)
 66{
 67	return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
 68			adev->gfx.mec.queue_bitmap);
 69}
 70
 71int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
 72			       int me, int pipe, int queue)
 73{
 74	int bit = 0;
 75
 76	bit += me * adev->gfx.me.num_pipe_per_me
 77		* adev->gfx.me.num_queue_per_pipe;
 78	bit += pipe * adev->gfx.me.num_queue_per_pipe;
 79	bit += queue;
 80
 81	return bit;
 82}
 83
 84void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
 85				int *me, int *pipe, int *queue)
 86{
 87	*queue = bit % adev->gfx.me.num_queue_per_pipe;
 88	*pipe = (bit / adev->gfx.me.num_queue_per_pipe)
 89		% adev->gfx.me.num_pipe_per_me;
 90	*me = (bit / adev->gfx.me.num_queue_per_pipe)
 91		/ adev->gfx.me.num_pipe_per_me;
 92}
 93
 94bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
 95				    int me, int pipe, int queue)
 96{
 97	return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
 98			adev->gfx.me.queue_bitmap);
 99}
100
101/**
102 * amdgpu_gfx_scratch_get - Allocate a scratch register
103 *
104 * @adev: amdgpu_device pointer
105 * @reg: scratch register mmio offset
106 *
107 * Allocate a CP scratch register for use by the driver (all asics).
108 * Returns 0 on success or -EINVAL on failure.
109 */
110int amdgpu_gfx_scratch_get(struct amdgpu_device *adev, uint32_t *reg)
111{
112	int i;
113
114	i = ffs(adev->gfx.scratch.free_mask);
115	if (i != 0 && i <= adev->gfx.scratch.num_reg) {
116		i--;
117		adev->gfx.scratch.free_mask &= ~(1u << i);
118		*reg = adev->gfx.scratch.reg_base + i;
119		return 0;
120	}
121	return -EINVAL;
122}
123
124/**
125 * amdgpu_gfx_scratch_free - Free a scratch register
126 *
127 * @adev: amdgpu_device pointer
128 * @reg: scratch register mmio offset
129 *
130 * Free a CP scratch register allocated for use by the driver (all asics)
131 */
132void amdgpu_gfx_scratch_free(struct amdgpu_device *adev, uint32_t reg)
133{
134	adev->gfx.scratch.free_mask |= 1u << (reg - adev->gfx.scratch.reg_base);
135}
136
137/**
138 * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
139 *
140 * @mask: array in which the per-shader array disable masks will be stored
141 * @max_se: number of SEs
142 * @max_sh: number of SHs
143 *
144 * The bitmask of CUs to be disabled in the shader array determined by se and
145 * sh is stored in mask[se * max_sh + sh].
146 */
147void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
148{
149	unsigned se, sh, cu;
150	const char *p;
151
152	memset(mask, 0, sizeof(*mask) * max_se * max_sh);
153
154	if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
155		return;
156
157	p = amdgpu_disable_cu;
158	for (;;) {
159		char *next;
160		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
161		if (ret < 3) {
162			DRM_ERROR("amdgpu: could not parse disable_cu\n");
163			return;
164		}
165
166		if (se < max_se && sh < max_sh && cu < 16) {
167			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
168			mask[se * max_sh + sh] |= 1u << cu;
169		} else {
170			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
171				  se, sh, cu);
172		}
173
174		next = strchr(p, ',');
175		if (!next)
176			break;
177		p = next + 1;
178	}
179}
180
181static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
182{
183	if (amdgpu_compute_multipipe != -1) {
184		DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
185			 amdgpu_compute_multipipe);
186		return amdgpu_compute_multipipe == 1;
187	}
188
189	/* FIXME: spreading the queues across pipes causes perf regressions
190	 * on POLARIS11 compute workloads */
191	if (adev->asic_type == CHIP_POLARIS11)
192		return false;
193
194	return adev->gfx.mec.num_mec > 1;
195}
196
197bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
198					       struct amdgpu_ring *ring)
199{
200	/* Policy: use 1st queue as high priority compute queue if we
201	 * have more than one compute queue.
202	 */
203	if (adev->gfx.num_compute_rings > 1 &&
204	    ring == &adev->gfx.compute_ring[0])
205		return true;
206
207	return false;
208}
 
 
 
 
 
209
210void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
211{
212	int i, queue, pipe;
213	bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev);
214	int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
215				     adev->gfx.mec.num_queue_per_pipe,
216				     adev->gfx.num_compute_rings);
217
218	if (multipipe_policy) {
219		/* policy: make queues evenly cross all pipes on MEC1 only */
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.queue_bitmap);
 
 
 
 
 
 
227		}
228	} else {
229		/* policy: amdgpu owns all queues in the given pipe */
230		for (i = 0; i < max_queues_per_mec; ++i)
231			set_bit(i, adev->gfx.mec.queue_bitmap);
232	}
233
234	dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
 
 
 
 
 
 
 
235}
236
237void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
238{
239	int i, queue, me;
240
241	for (i = 0; i < AMDGPU_MAX_GFX_QUEUES; ++i) {
242		queue = i % adev->gfx.me.num_queue_per_pipe;
 
 
243		me = (i / adev->gfx.me.num_queue_per_pipe)
244		      / adev->gfx.me.num_pipe_per_me;
245
246		if (me >= adev->gfx.me.num_me)
247			break;
248		/* policy: amdgpu owns the first queue per pipe at this stage
249		 * will extend to mulitple queues per pipe later */
250		if (me == 0 && queue < 1)
251			set_bit(i, adev->gfx.me.queue_bitmap);
252	}
253
254	/* update the number of active graphics rings */
255	adev->gfx.num_gfx_rings =
256		bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
257}
258
259static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
260				  struct amdgpu_ring *ring)
261{
262	int queue_bit;
263	int mec, pipe, queue;
264
265	queue_bit = adev->gfx.mec.num_mec
266		    * adev->gfx.mec.num_pipe_per_mec
267		    * adev->gfx.mec.num_queue_per_pipe;
268
269	while (queue_bit-- >= 0) {
270		if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
271			continue;
272
273		amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
274
275		/*
276		 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
277		 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
278		 * only can be issued on queue 0.
279		 */
280		if ((mec == 1 && pipe > 1) || queue != 0)
281			continue;
282
283		ring->me = mec + 1;
284		ring->pipe = pipe;
285		ring->queue = queue;
286
287		return 0;
288	}
289
290	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
291	return -EINVAL;
292}
293
294int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
295			     struct amdgpu_ring *ring,
296			     struct amdgpu_irq_src *irq)
297{
298	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
299	int r = 0;
300
301	spin_lock_init(&kiq->ring_lock);
302
 
 
 
 
303	ring->adev = NULL;
304	ring->ring_obj = NULL;
305	ring->use_doorbell = true;
306	ring->doorbell_index = adev->doorbell_index.kiq;
307
308	r = amdgpu_gfx_kiq_acquire(adev, ring);
309	if (r)
310		return r;
311
312	ring->eop_gpu_addr = kiq->eop_gpu_addr;
313	ring->no_scheduler = true;
314	sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
315	r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
316			     AMDGPU_RING_PRIO_DEFAULT, NULL);
317	if (r)
318		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
319
320	return r;
321}
322
323void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
 
324{
 
325	amdgpu_ring_fini(ring);
326}
327
328void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
329{
330	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
331
332	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
333}
334
335int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
336			unsigned hpd_size)
337{
338	int r;
339	u32 *hpd;
340	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
341
342	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
343				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
344				    &kiq->eop_gpu_addr, (void **)&hpd);
345	if (r) {
346		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
347		return r;
348	}
349
350	memset(hpd, 0, hpd_size);
351
352	r = amdgpu_bo_reserve(kiq->eop_obj, true);
353	if (unlikely(r != 0))
354		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
355	amdgpu_bo_kunmap(kiq->eop_obj);
356	amdgpu_bo_unreserve(kiq->eop_obj);
357
358	return 0;
359}
360
361/* create MQD for each compute/gfx queue */
362int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
363			   unsigned mqd_size)
364{
365	struct amdgpu_ring *ring = NULL;
366	int r, i;
367
368	/* create MQD for KIQ */
369	ring = &adev->gfx.kiq.ring;
370	if (!ring->mqd_obj) {
371		/* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
372		 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
373		 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
374		 * KIQ MQD no matter SRIOV or Bare-metal
375		 */
376		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
377					    AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj,
378					    &ring->mqd_gpu_addr, &ring->mqd_ptr);
379		if (r) {
380			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
381			return r;
382		}
383
384		/* prepare MQD backup */
385		adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
386		if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
387				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
388	}
389
390	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
391		/* create MQD for each KGQ */
392		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
393			ring = &adev->gfx.gfx_ring[i];
394			if (!ring->mqd_obj) {
395				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
396							    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
397							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
398				if (r) {
399					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
400					return r;
401				}
402
403				/* prepare MQD backup */
404				adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
405				if (!adev->gfx.me.mqd_backup[i])
406					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
407			}
408		}
409	}
410
411	/* create MQD for each KCQ */
412	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
413		ring = &adev->gfx.compute_ring[i];
414		if (!ring->mqd_obj) {
415			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
416						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
417						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
418			if (r) {
419				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
420				return r;
421			}
422
423			/* prepare MQD backup */
424			adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
425			if (!adev->gfx.mec.mqd_backup[i])
426				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
427		}
428	}
429
430	return 0;
431}
432
433void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev)
434{
435	struct amdgpu_ring *ring = NULL;
436	int i;
437
438	if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
439		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
440			ring = &adev->gfx.gfx_ring[i];
441			kfree(adev->gfx.me.mqd_backup[i]);
442			amdgpu_bo_free_kernel(&ring->mqd_obj,
443					      &ring->mqd_gpu_addr,
444					      &ring->mqd_ptr);
445		}
446	}
447
448	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
449		ring = &adev->gfx.compute_ring[i];
450		kfree(adev->gfx.mec.mqd_backup[i]);
451		amdgpu_bo_free_kernel(&ring->mqd_obj,
452				      &ring->mqd_gpu_addr,
453				      &ring->mqd_ptr);
454	}
455
456	ring = &adev->gfx.kiq.ring;
 
 
457	kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
458	amdgpu_bo_free_kernel(&ring->mqd_obj,
459			      &ring->mqd_gpu_addr,
460			      &ring->mqd_ptr);
461}
462
463int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev)
464{
465	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
466	struct amdgpu_ring *kiq_ring = &kiq->ring;
467	int i, r;
468
469	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
470		return -EINVAL;
471
472	spin_lock(&adev->gfx.kiq.ring_lock);
473	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
474					adev->gfx.num_compute_rings)) {
475		spin_unlock(&adev->gfx.kiq.ring_lock);
476		return -ENOMEM;
477	}
478
479	for (i = 0; i < adev->gfx.num_compute_rings; i++)
480		kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i],
481					   RESET_QUEUES, 0, 0);
482	r = amdgpu_ring_test_helper(kiq_ring);
483	spin_unlock(&adev->gfx.kiq.ring_lock);
484
485	return r;
486}
487
488int 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
501int 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	kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
539	for (i = 0; i < adev->gfx.num_compute_rings; i++)
540		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
541
542	r = amdgpu_ring_test_helper(kiq_ring);
543	spin_unlock(&adev->gfx.kiq.ring_lock);
544	if (r)
545		DRM_ERROR("KCQ enable failed\n");
546
547	return r;
548}
549
550/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
551 *
552 * @adev: amdgpu_device pointer
553 * @bool enable true: enable gfx off feature, false: disable gfx off feature
554 *
555 * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
556 * 2. other client can send request to disable gfx off feature, the request should be honored.
557 * 3. other client can cancel their request of disable gfx off feature
558 * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
559 */
560
561void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
562{
563	unsigned long delay = GFX_OFF_DELAY_ENABLE;
 
564
565	if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
 
 
566		return;
567
 
568	mutex_lock(&adev->gfx.gfx_off_mutex);
569
570	if (enable) {
571		/* If the count is already 0, it means there's an imbalance bug somewhere.
572		 * Note that the bug may be in a different caller than the one which triggers the
573		 * WARN_ON_ONCE.
574		 */
575		if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
576			goto unlock;
577
578		adev->gfx.gfx_off_req_count--;
579
580		if (adev->gfx.gfx_off_req_count == 0 &&
581		    !adev->gfx.gfx_off_state) {
582			/* If going to s2idle, no need to wait */
583			if (adev->in_s0ix)
584				delay = GFX_OFF_NO_DELAY;
585			schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
586					      delay);
587		}
588	} else {
589		if (adev->gfx.gfx_off_req_count == 0) {
590			cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
591
592			if (adev->gfx.gfx_off_state &&
593			    !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
594				adev->gfx.gfx_off_state = false;
595
596				if (adev->gfx.funcs->init_spm_golden) {
597					dev_dbg(adev->dev,
598						"GFXOFF is disabled, re-init SPM golden settings\n");
599					amdgpu_gfx_init_spm_golden(adev);
600				}
601			}
602		}
603
604		adev->gfx.gfx_off_req_count++;
605	}
606
607unlock:
608	mutex_unlock(&adev->gfx.gfx_off_mutex);
609}
610
611int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
612{
613
614	int r = 0;
615
616	mutex_lock(&adev->gfx.gfx_off_mutex);
617
618	r = smu_get_status_gfxoff(adev, value);
619
620	mutex_unlock(&adev->gfx.gfx_off_mutex);
621
622	return r;
623}
624
625int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev)
626{
627	int r;
628	struct ras_fs_if fs_info = {
629		.sysfs_name = "gfx_err_count",
630	};
631	struct ras_ih_if ih_info = {
632		.cb = amdgpu_gfx_process_ras_data_cb,
633	};
634
635	if (!adev->gfx.ras_if) {
636		adev->gfx.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL);
637		if (!adev->gfx.ras_if)
638			return -ENOMEM;
639		adev->gfx.ras_if->block = AMDGPU_RAS_BLOCK__GFX;
640		adev->gfx.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
641		adev->gfx.ras_if->sub_block_index = 0;
642		strcpy(adev->gfx.ras_if->name, "gfx");
643	}
644	fs_info.head = ih_info.head = *adev->gfx.ras_if;
645	r = amdgpu_ras_late_init(adev, adev->gfx.ras_if,
646				 &fs_info, &ih_info);
647	if (r)
648		goto free;
649
650	if (amdgpu_ras_is_supported(adev, adev->gfx.ras_if->block)) {
651		if (!amdgpu_persistent_edc_harvesting_supported(adev))
652			amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
653
654		r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
655		if (r)
656			goto late_fini;
657	} else {
658		/* free gfx ras_if if ras is not supported */
659		r = 0;
660		goto free;
661	}
662
663	return 0;
664late_fini:
665	amdgpu_ras_late_fini(adev, adev->gfx.ras_if, &ih_info);
666free:
667	kfree(adev->gfx.ras_if);
668	adev->gfx.ras_if = NULL;
669	return r;
670}
671
672void amdgpu_gfx_ras_fini(struct amdgpu_device *adev)
673{
674	if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX) &&
675			adev->gfx.ras_if) {
676		struct ras_common_if *ras_if = adev->gfx.ras_if;
677		struct ras_ih_if ih_info = {
678			.head = *ras_if,
679			.cb = amdgpu_gfx_process_ras_data_cb,
680		};
681
682		amdgpu_ras_late_fini(adev, ras_if, &ih_info);
683		kfree(ras_if);
684	}
685}
686
687int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
688		void *err_data,
689		struct amdgpu_iv_entry *entry)
690{
691	/* TODO ue will trigger an interrupt.
692	 *
693	 * When “Full RAS” is enabled, the per-IP interrupt sources should
694	 * be disabled and the driver should only look for the aggregated
695	 * interrupt via sync flood
696	 */
697	if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
698		kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
699		if (adev->gfx.ras_funcs &&
700		    adev->gfx.ras_funcs->query_ras_error_count)
701			adev->gfx.ras_funcs->query_ras_error_count(adev, err_data);
702		amdgpu_ras_reset_gpu(adev);
703	}
704	return AMDGPU_RAS_SUCCESS;
705}
706
707int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
708				  struct amdgpu_irq_src *source,
709				  struct amdgpu_iv_entry *entry)
710{
711	struct ras_common_if *ras_if = adev->gfx.ras_if;
712	struct ras_dispatch_if ih_data = {
713		.entry = entry,
714	};
715
716	if (!ras_if)
717		return 0;
718
719	ih_data.head = *ras_if;
720
721	DRM_ERROR("CP ECC ERROR IRQ\n");
722	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
723	return 0;
724}
725
726uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
727{
728	signed long r, cnt = 0;
729	unsigned long flags;
730	uint32_t seq, reg_val_offs = 0, value = 0;
731	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
732	struct amdgpu_ring *ring = &kiq->ring;
733
734	if (amdgpu_device_skip_hw_access(adev))
735		return 0;
736
737	BUG_ON(!ring->funcs->emit_rreg);
738
739	spin_lock_irqsave(&kiq->ring_lock, flags);
740	if (amdgpu_device_wb_get(adev, &reg_val_offs)) {
741		pr_err("critical bug! too many kiq readers\n");
742		goto failed_unlock;
743	}
744	amdgpu_ring_alloc(ring, 32);
745	amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
746	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
747	if (r)
748		goto failed_undo;
749
750	amdgpu_ring_commit(ring);
751	spin_unlock_irqrestore(&kiq->ring_lock, flags);
752
753	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
754
755	/* don't wait anymore for gpu reset case because this way may
756	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
757	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
758	 * never return if we keep waiting in virt_kiq_rreg, which cause
759	 * gpu_recover() hang there.
760	 *
761	 * also don't wait anymore for IRQ context
762	 * */
763	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
764		goto failed_kiq_read;
765
766	might_sleep();
767	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
768		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
769		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
770	}
771
772	if (cnt > MAX_KIQ_REG_TRY)
773		goto failed_kiq_read;
774
775	mb();
776	value = adev->wb.wb[reg_val_offs];
777	amdgpu_device_wb_free(adev, reg_val_offs);
778	return value;
779
780failed_undo:
781	amdgpu_ring_undo(ring);
782failed_unlock:
783	spin_unlock_irqrestore(&kiq->ring_lock, flags);
784failed_kiq_read:
785	if (reg_val_offs)
786		amdgpu_device_wb_free(adev, reg_val_offs);
787	dev_err(adev->dev, "failed to read reg:%x\n", reg);
788	return ~0;
789}
790
791void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
792{
793	signed long r, cnt = 0;
794	unsigned long flags;
795	uint32_t seq;
796	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
797	struct amdgpu_ring *ring = &kiq->ring;
798
799	BUG_ON(!ring->funcs->emit_wreg);
800
801	if (amdgpu_device_skip_hw_access(adev))
802		return;
803
804	spin_lock_irqsave(&kiq->ring_lock, flags);
805	amdgpu_ring_alloc(ring, 32);
806	amdgpu_ring_emit_wreg(ring, reg, v);
807	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
808	if (r)
809		goto failed_undo;
810
811	amdgpu_ring_commit(ring);
812	spin_unlock_irqrestore(&kiq->ring_lock, flags);
813
814	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
815
816	/* don't wait anymore for gpu reset case because this way may
817	 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
818	 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
819	 * never return if we keep waiting in virt_kiq_rreg, which cause
820	 * gpu_recover() hang there.
821	 *
822	 * also don't wait anymore for IRQ context
823	 * */
824	if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
825		goto failed_kiq_write;
826
827	might_sleep();
828	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
829
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_write;
836
837	return;
838
839failed_undo:
840	amdgpu_ring_undo(ring);
841	spin_unlock_irqrestore(&kiq->ring_lock, flags);
842failed_kiq_write:
843	dev_err(adev->dev, "failed to write reg:%x\n", reg);
844}
845
846int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
847{
848	if (amdgpu_num_kcq == -1) {
849		return 8;
850	} else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
851		dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
852		return 8;
853	}
854	return amdgpu_num_kcq;
855}
856
857/* amdgpu_gfx_state_change_set - Handle gfx power state change set
858 * @adev: amdgpu_device pointer
859 * @state: gfx power state(1 -sGpuChangeState_D0Entry and 2 -sGpuChangeState_D3Entry)
860 *
861 */
862
863void amdgpu_gfx_state_change_set(struct amdgpu_device *adev, enum gfx_change_state state)
864{
865	mutex_lock(&adev->pm.mutex);
866	if (adev->powerplay.pp_funcs &&
867	    adev->powerplay.pp_funcs->gfx_state_change_set)
868		((adev)->powerplay.pp_funcs->gfx_state_change_set(
869			(adev)->powerplay.pp_handle, state));
870	mutex_unlock(&adev->pm.mutex);
871}