xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c (revision 42b16d3ac371a2fac9b6f08fd75f23f34ba3955a)
1 /*
2  * Copyright 2008 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  * Authors: Dave Airlie
25  *          Alex Deucher
26  *          Jerome Glisse
27  *          Christian König
28  */
29 #include <linux/seq_file.h>
30 #include <linux/slab.h>
31 #include <linux/uaccess.h>
32 #include <linux/debugfs.h>
33 
34 #include <drm/amdgpu_drm.h>
35 #include "amdgpu.h"
36 #include "atom.h"
37 
38 /*
39  * Rings
40  * Most engines on the GPU are fed via ring buffers.  Ring
41  * buffers are areas of GPU accessible memory that the host
42  * writes commands into and the GPU reads commands out of.
43  * There is a rptr (read pointer) that determines where the
44  * GPU is currently reading, and a wptr (write pointer)
45  * which determines where the host has written.  When the
46  * pointers are equal, the ring is idle.  When the host
47  * writes commands to the ring buffer, it increments the
48  * wptr.  The GPU then starts fetching commands and executes
49  * them until the pointers are equal again.
50  */
51 
52 /**
53  * amdgpu_ring_max_ibs - Return max IBs that fit in a single submission.
54  *
55  * @type: ring type for which to return the limit.
56  */
amdgpu_ring_max_ibs(enum amdgpu_ring_type type)57 unsigned int amdgpu_ring_max_ibs(enum amdgpu_ring_type type)
58 {
59 	switch (type) {
60 	case AMDGPU_RING_TYPE_GFX:
61 		/* Need to keep at least 192 on GFX7+ for old radv. */
62 		return 192;
63 	case AMDGPU_RING_TYPE_COMPUTE:
64 		return 125;
65 	case AMDGPU_RING_TYPE_VCN_JPEG:
66 		return 16;
67 	default:
68 		return 49;
69 	}
70 }
71 
72 /**
73  * amdgpu_ring_alloc - allocate space on the ring buffer
74  *
75  * @ring: amdgpu_ring structure holding ring information
76  * @ndw: number of dwords to allocate in the ring buffer
77  *
78  * Allocate @ndw dwords in the ring buffer (all asics).
79  * Returns 0 on success, error on failure.
80  */
amdgpu_ring_alloc(struct amdgpu_ring * ring,unsigned int ndw)81 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned int ndw)
82 {
83 	/* Align requested size with padding so unlock_commit can
84 	 * pad safely */
85 	ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
86 
87 	/* Make sure we aren't trying to allocate more space
88 	 * than the maximum for one submission
89 	 */
90 	if (WARN_ON_ONCE(ndw > ring->max_dw))
91 		return -ENOMEM;
92 
93 	ring->count_dw = ndw;
94 	ring->wptr_old = ring->wptr;
95 
96 	if (ring->funcs->begin_use)
97 		ring->funcs->begin_use(ring);
98 
99 	return 0;
100 }
101 
102 /** amdgpu_ring_insert_nop - insert NOP packets
103  *
104  * @ring: amdgpu_ring structure holding ring information
105  * @count: the number of NOP packets to insert
106  *
107  * This is the generic insert_nop function for rings except SDMA
108  */
amdgpu_ring_insert_nop(struct amdgpu_ring * ring,uint32_t count)109 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
110 {
111 	int i;
112 
113 	for (i = 0; i < count; i++)
114 		amdgpu_ring_write(ring, ring->funcs->nop);
115 }
116 
117 /**
118  * amdgpu_ring_generic_pad_ib - pad IB with NOP packets
119  *
120  * @ring: amdgpu_ring structure holding ring information
121  * @ib: IB to add NOP packets to
122  *
123  * This is the generic pad_ib function for rings except SDMA
124  */
amdgpu_ring_generic_pad_ib(struct amdgpu_ring * ring,struct amdgpu_ib * ib)125 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
126 {
127 	while (ib->length_dw & ring->funcs->align_mask)
128 		ib->ptr[ib->length_dw++] = ring->funcs->nop;
129 }
130 
131 /**
132  * amdgpu_ring_commit - tell the GPU to execute the new
133  * commands on the ring buffer
134  *
135  * @ring: amdgpu_ring structure holding ring information
136  *
137  * Update the wptr (write pointer) to tell the GPU to
138  * execute new commands on the ring buffer (all asics).
139  */
amdgpu_ring_commit(struct amdgpu_ring * ring)140 void amdgpu_ring_commit(struct amdgpu_ring *ring)
141 {
142 	uint32_t count;
143 
144 	/* We pad to match fetch size */
145 	count = ring->funcs->align_mask + 1 -
146 		(ring->wptr & ring->funcs->align_mask);
147 	count &= ring->funcs->align_mask;
148 
149 	if (count != 0)
150 		ring->funcs->insert_nop(ring, count);
151 
152 	mb();
153 	amdgpu_ring_set_wptr(ring);
154 
155 	if (ring->funcs->end_use)
156 		ring->funcs->end_use(ring);
157 }
158 
159 /**
160  * amdgpu_ring_undo - reset the wptr
161  *
162  * @ring: amdgpu_ring structure holding ring information
163  *
164  * Reset the driver's copy of the wptr (all asics).
165  */
amdgpu_ring_undo(struct amdgpu_ring * ring)166 void amdgpu_ring_undo(struct amdgpu_ring *ring)
167 {
168 	ring->wptr = ring->wptr_old;
169 
170 	if (ring->funcs->end_use)
171 		ring->funcs->end_use(ring);
172 }
173 
174 #define amdgpu_ring_get_gpu_addr(ring, offset)				\
175 	(ring->is_mes_queue ?						\
176 	 (ring->mes_ctx->meta_data_gpu_addr + offset) :			\
177 	 (ring->adev->wb.gpu_addr + offset * 4))
178 
179 #define amdgpu_ring_get_cpu_addr(ring, offset)				\
180 	(ring->is_mes_queue ?						\
181 	 (void *)((uint8_t *)(ring->mes_ctx->meta_data_ptr) + offset) : \
182 	 (&ring->adev->wb.wb[offset]))
183 
184 /**
185  * amdgpu_ring_init - init driver ring struct.
186  *
187  * @adev: amdgpu_device pointer
188  * @ring: amdgpu_ring structure holding ring information
189  * @max_dw: maximum number of dw for ring alloc
190  * @irq_src: interrupt source to use for this ring
191  * @irq_type: interrupt type to use for this ring
192  * @hw_prio: ring priority (NORMAL/HIGH)
193  * @sched_score: optional score atomic shared with other schedulers
194  *
195  * Initialize the driver information for the selected ring (all asics).
196  * Returns 0 on success, error on failure.
197  */
amdgpu_ring_init(struct amdgpu_device * adev,struct amdgpu_ring * ring,unsigned int max_dw,struct amdgpu_irq_src * irq_src,unsigned int irq_type,unsigned int hw_prio,atomic_t * sched_score)198 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
199 		     unsigned int max_dw, struct amdgpu_irq_src *irq_src,
200 		     unsigned int irq_type, unsigned int hw_prio,
201 		     atomic_t *sched_score)
202 {
203 	int r;
204 	int sched_hw_submission = amdgpu_sched_hw_submission;
205 	u32 *num_sched;
206 	u32 hw_ip;
207 	unsigned int max_ibs_dw;
208 
209 	/* Set the hw submission limit higher for KIQ because
210 	 * it's used for a number of gfx/compute tasks by both
211 	 * KFD and KGD which may have outstanding fences and
212 	 * it doesn't really use the gpu scheduler anyway;
213 	 * KIQ tasks get submitted directly to the ring.
214 	 */
215 	if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
216 		sched_hw_submission = max(sched_hw_submission, 256);
217 	if (ring->funcs->type == AMDGPU_RING_TYPE_MES)
218 		sched_hw_submission = 8;
219 	else if (ring == &adev->sdma.instance[0].page)
220 		sched_hw_submission = 256;
221 
222 	if (ring->adev == NULL) {
223 		if (adev->num_rings >= AMDGPU_MAX_RINGS)
224 			return -EINVAL;
225 
226 		ring->adev = adev;
227 		ring->num_hw_submission = sched_hw_submission;
228 		ring->sched_score = sched_score;
229 		ring->vmid_wait = dma_fence_get_stub();
230 
231 		if (!ring->is_mes_queue) {
232 			ring->idx = adev->num_rings++;
233 			adev->rings[ring->idx] = ring;
234 		}
235 
236 		r = amdgpu_fence_driver_init_ring(ring);
237 		if (r)
238 			return r;
239 	}
240 
241 	if (ring->is_mes_queue) {
242 		ring->rptr_offs = amdgpu_mes_ctx_get_offs(ring,
243 				AMDGPU_MES_CTX_RPTR_OFFS);
244 		ring->wptr_offs = amdgpu_mes_ctx_get_offs(ring,
245 				AMDGPU_MES_CTX_WPTR_OFFS);
246 		ring->fence_offs = amdgpu_mes_ctx_get_offs(ring,
247 				AMDGPU_MES_CTX_FENCE_OFFS);
248 		ring->trail_fence_offs = amdgpu_mes_ctx_get_offs(ring,
249 				AMDGPU_MES_CTX_TRAIL_FENCE_OFFS);
250 		ring->cond_exe_offs = amdgpu_mes_ctx_get_offs(ring,
251 				AMDGPU_MES_CTX_COND_EXE_OFFS);
252 	} else {
253 		r = amdgpu_device_wb_get(adev, &ring->rptr_offs);
254 		if (r) {
255 			dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
256 			return r;
257 		}
258 
259 		r = amdgpu_device_wb_get(adev, &ring->wptr_offs);
260 		if (r) {
261 			dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
262 			return r;
263 		}
264 
265 		r = amdgpu_device_wb_get(adev, &ring->fence_offs);
266 		if (r) {
267 			dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
268 			return r;
269 		}
270 
271 		r = amdgpu_device_wb_get(adev, &ring->trail_fence_offs);
272 		if (r) {
273 			dev_err(adev->dev, "(%d) ring trail_fence_offs wb alloc failed\n", r);
274 			return r;
275 		}
276 
277 		r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs);
278 		if (r) {
279 			dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
280 			return r;
281 		}
282 	}
283 
284 	ring->fence_gpu_addr =
285 		amdgpu_ring_get_gpu_addr(ring, ring->fence_offs);
286 	ring->fence_cpu_addr =
287 		amdgpu_ring_get_cpu_addr(ring, ring->fence_offs);
288 
289 	ring->rptr_gpu_addr =
290 		amdgpu_ring_get_gpu_addr(ring, ring->rptr_offs);
291 	ring->rptr_cpu_addr =
292 		amdgpu_ring_get_cpu_addr(ring, ring->rptr_offs);
293 
294 	ring->wptr_gpu_addr =
295 		amdgpu_ring_get_gpu_addr(ring, ring->wptr_offs);
296 	ring->wptr_cpu_addr =
297 		amdgpu_ring_get_cpu_addr(ring, ring->wptr_offs);
298 
299 	ring->trail_fence_gpu_addr =
300 		amdgpu_ring_get_gpu_addr(ring, ring->trail_fence_offs);
301 	ring->trail_fence_cpu_addr =
302 		amdgpu_ring_get_cpu_addr(ring, ring->trail_fence_offs);
303 
304 	ring->cond_exe_gpu_addr =
305 		amdgpu_ring_get_gpu_addr(ring, ring->cond_exe_offs);
306 	ring->cond_exe_cpu_addr =
307 		amdgpu_ring_get_cpu_addr(ring, ring->cond_exe_offs);
308 
309 	/* always set cond_exec_polling to CONTINUE */
310 	*ring->cond_exe_cpu_addr = 1;
311 
312 	r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
313 	if (r) {
314 		dev_err(adev->dev, "failed initializing fences (%d).\n", r);
315 		return r;
316 	}
317 
318 	max_ibs_dw = ring->funcs->emit_frame_size +
319 		     amdgpu_ring_max_ibs(ring->funcs->type) * ring->funcs->emit_ib_size;
320 	max_ibs_dw = (max_ibs_dw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
321 
322 	if (WARN_ON(max_ibs_dw > max_dw))
323 		max_dw = max_ibs_dw;
324 
325 	ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
326 
327 	ring->buf_mask = (ring->ring_size / 4) - 1;
328 	ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
329 		0xffffffffffffffff : ring->buf_mask;
330 
331 	/* Allocate ring buffer */
332 	if (ring->is_mes_queue) {
333 		int offset = 0;
334 
335 		BUG_ON(ring->ring_size > PAGE_SIZE*4);
336 
337 		offset = amdgpu_mes_ctx_get_offs(ring,
338 					 AMDGPU_MES_CTX_RING_OFFS);
339 		ring->gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
340 		ring->ring = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
341 		amdgpu_ring_clear_ring(ring);
342 
343 	} else if (ring->ring_obj == NULL) {
344 		r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_dw, PAGE_SIZE,
345 					    AMDGPU_GEM_DOMAIN_GTT,
346 					    &ring->ring_obj,
347 					    &ring->gpu_addr,
348 					    (void **)&ring->ring);
349 		if (r) {
350 			dev_err(adev->dev, "(%d) ring create failed\n", r);
351 			return r;
352 		}
353 		amdgpu_ring_clear_ring(ring);
354 	}
355 
356 	ring->max_dw = max_dw;
357 	ring->hw_prio = hw_prio;
358 
359 	if (!ring->no_scheduler && ring->funcs->type < AMDGPU_HW_IP_NUM) {
360 		hw_ip = ring->funcs->type;
361 		num_sched = &adev->gpu_sched[hw_ip][hw_prio].num_scheds;
362 		adev->gpu_sched[hw_ip][hw_prio].sched[(*num_sched)++] =
363 			&ring->sched;
364 	}
365 
366 	return 0;
367 }
368 
369 /**
370  * amdgpu_ring_fini - tear down the driver ring struct.
371  *
372  * @ring: amdgpu_ring structure holding ring information
373  *
374  * Tear down the driver information for the selected ring (all asics).
375  */
amdgpu_ring_fini(struct amdgpu_ring * ring)376 void amdgpu_ring_fini(struct amdgpu_ring *ring)
377 {
378 
379 	/* Not to finish a ring which is not initialized */
380 	if (!(ring->adev) ||
381 	    (!ring->is_mes_queue && !(ring->adev->rings[ring->idx])))
382 		return;
383 
384 	ring->sched.ready = false;
385 
386 	if (!ring->is_mes_queue) {
387 		amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
388 		amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
389 
390 		amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
391 		amdgpu_device_wb_free(ring->adev, ring->fence_offs);
392 
393 		amdgpu_bo_free_kernel(&ring->ring_obj,
394 				      &ring->gpu_addr,
395 				      (void **)&ring->ring);
396 	} else {
397 		kfree(ring->fence_drv.fences);
398 	}
399 
400 	dma_fence_put(ring->vmid_wait);
401 	ring->vmid_wait = NULL;
402 	ring->me = 0;
403 
404 	if (!ring->is_mes_queue)
405 		ring->adev->rings[ring->idx] = NULL;
406 }
407 
408 /**
409  * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper
410  *
411  * @ring: ring to write to
412  * @reg0: register to write
413  * @reg1: register to wait on
414  * @ref: reference value to write/wait on
415  * @mask: mask to wait on
416  *
417  * Helper for rings that don't support write and wait in a
418  * single oneshot packet.
419  */
amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring * ring,uint32_t reg0,uint32_t reg1,uint32_t ref,uint32_t mask)420 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
421 						uint32_t reg0, uint32_t reg1,
422 						uint32_t ref, uint32_t mask)
423 {
424 	amdgpu_ring_emit_wreg(ring, reg0, ref);
425 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
426 }
427 
428 /**
429  * amdgpu_ring_soft_recovery - try to soft recover a ring lockup
430  *
431  * @ring: ring to try the recovery on
432  * @vmid: VMID we try to get going again
433  * @fence: timedout fence
434  *
435  * Tries to get a ring proceeding again when it is stuck.
436  */
amdgpu_ring_soft_recovery(struct amdgpu_ring * ring,unsigned int vmid,struct dma_fence * fence)437 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
438 			       struct dma_fence *fence)
439 {
440 	unsigned long flags;
441 	ktime_t deadline;
442 
443 	if (unlikely(ring->adev->debug_disable_soft_recovery))
444 		return false;
445 
446 	deadline = ktime_add_us(ktime_get(), 10000);
447 
448 	if (amdgpu_sriov_vf(ring->adev) || !ring->funcs->soft_recovery || !fence)
449 		return false;
450 
451 	spin_lock_irqsave(fence->lock, flags);
452 	if (!dma_fence_is_signaled_locked(fence))
453 		dma_fence_set_error(fence, -ENODATA);
454 	spin_unlock_irqrestore(fence->lock, flags);
455 
456 	atomic_inc(&ring->adev->gpu_reset_counter);
457 	while (!dma_fence_is_signaled(fence) &&
458 	       ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0)
459 		ring->funcs->soft_recovery(ring, vmid);
460 
461 	return dma_fence_is_signaled(fence);
462 }
463 
464 /*
465  * Debugfs info
466  */
467 #if defined(CONFIG_DEBUG_FS)
468 
469 /* Layout of file is 12 bytes consisting of
470  * - rptr
471  * - wptr
472  * - driver's copy of wptr
473  *
474  * followed by n-words of ring data
475  */
amdgpu_debugfs_ring_read(struct file * f,char __user * buf,size_t size,loff_t * pos)476 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
477 					size_t size, loff_t *pos)
478 {
479 	struct amdgpu_ring *ring = file_inode(f)->i_private;
480 	uint32_t value, result, early[3];
481 	loff_t i;
482 	int r;
483 
484 	if (*pos & 3 || size & 3)
485 		return -EINVAL;
486 
487 	result = 0;
488 
489 	if (*pos < 12) {
490 		early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
491 		early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
492 		early[2] = ring->wptr & ring->buf_mask;
493 		for (i = *pos / 4; i < 3 && size; i++) {
494 			r = put_user(early[i], (uint32_t *)buf);
495 			if (r)
496 				return r;
497 			buf += 4;
498 			result += 4;
499 			size -= 4;
500 			*pos += 4;
501 		}
502 	}
503 
504 	while (size) {
505 		if (*pos >= (ring->ring_size + 12))
506 			return result;
507 
508 		value = ring->ring[(*pos - 12)/4];
509 		r = put_user(value, (uint32_t *)buf);
510 		if (r)
511 			return r;
512 		buf += 4;
513 		result += 4;
514 		size -= 4;
515 		*pos += 4;
516 	}
517 
518 	return result;
519 }
520 
521 static const struct file_operations amdgpu_debugfs_ring_fops = {
522 	.owner = THIS_MODULE,
523 	.read = amdgpu_debugfs_ring_read,
524 	.llseek = default_llseek
525 };
526 
amdgpu_debugfs_mqd_read(struct file * f,char __user * buf,size_t size,loff_t * pos)527 static ssize_t amdgpu_debugfs_mqd_read(struct file *f, char __user *buf,
528 				       size_t size, loff_t *pos)
529 {
530 	struct amdgpu_ring *ring = file_inode(f)->i_private;
531 	volatile u32 *mqd;
532 	u32 *kbuf;
533 	int r, i;
534 	uint32_t value, result;
535 
536 	if (*pos & 3 || size & 3)
537 		return -EINVAL;
538 
539 	kbuf = kmalloc(ring->mqd_size, GFP_KERNEL);
540 	if (!kbuf)
541 		return -ENOMEM;
542 
543 	r = amdgpu_bo_reserve(ring->mqd_obj, false);
544 	if (unlikely(r != 0))
545 		goto err_free;
546 
547 	r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&mqd);
548 	if (r)
549 		goto err_unreserve;
550 
551 	/*
552 	 * Copy to local buffer to avoid put_user(), which might fault
553 	 * and acquire mmap_sem, under reservation_ww_class_mutex.
554 	 */
555 	for (i = 0; i < ring->mqd_size/sizeof(u32); i++)
556 		kbuf[i] = mqd[i];
557 
558 	amdgpu_bo_kunmap(ring->mqd_obj);
559 	amdgpu_bo_unreserve(ring->mqd_obj);
560 
561 	result = 0;
562 	while (size) {
563 		if (*pos >= ring->mqd_size)
564 			break;
565 
566 		value = kbuf[*pos/4];
567 		r = put_user(value, (uint32_t *)buf);
568 		if (r)
569 			goto err_free;
570 		buf += 4;
571 		result += 4;
572 		size -= 4;
573 		*pos += 4;
574 	}
575 
576 	kfree(kbuf);
577 	return result;
578 
579 err_unreserve:
580 	amdgpu_bo_unreserve(ring->mqd_obj);
581 err_free:
582 	kfree(kbuf);
583 	return r;
584 }
585 
586 static const struct file_operations amdgpu_debugfs_mqd_fops = {
587 	.owner = THIS_MODULE,
588 	.read = amdgpu_debugfs_mqd_read,
589 	.llseek = default_llseek
590 };
591 
amdgpu_debugfs_ring_error(void * data,u64 val)592 static int amdgpu_debugfs_ring_error(void *data, u64 val)
593 {
594 	struct amdgpu_ring *ring = data;
595 
596 	amdgpu_fence_driver_set_error(ring, val);
597 	return 0;
598 }
599 
600 DEFINE_DEBUGFS_ATTRIBUTE_SIGNED(amdgpu_debugfs_error_fops, NULL,
601 				amdgpu_debugfs_ring_error, "%lld\n");
602 
603 #endif
604 
amdgpu_debugfs_ring_init(struct amdgpu_device * adev,struct amdgpu_ring * ring)605 void amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
606 			      struct amdgpu_ring *ring)
607 {
608 #if defined(CONFIG_DEBUG_FS)
609 	struct drm_minor *minor = adev_to_drm(adev)->primary;
610 	struct dentry *root = minor->debugfs_root;
611 	char name[32];
612 
613 	sprintf(name, "amdgpu_ring_%s", ring->name);
614 	debugfs_create_file_size(name, S_IFREG | 0444, root, ring,
615 				 &amdgpu_debugfs_ring_fops,
616 				 ring->ring_size + 12);
617 
618 	if (ring->mqd_obj) {
619 		sprintf(name, "amdgpu_mqd_%s", ring->name);
620 		debugfs_create_file_size(name, S_IFREG | 0444, root, ring,
621 					 &amdgpu_debugfs_mqd_fops,
622 					 ring->mqd_size);
623 	}
624 
625 	sprintf(name, "amdgpu_error_%s", ring->name);
626 	debugfs_create_file(name, 0200, root, ring,
627 			    &amdgpu_debugfs_error_fops);
628 
629 #endif
630 }
631 
632 /**
633  * amdgpu_ring_test_helper - tests ring and set sched readiness status
634  *
635  * @ring: ring to try the recovery on
636  *
637  * Tests ring and set sched readiness status
638  *
639  * Returns 0 on success, error on failure.
640  */
amdgpu_ring_test_helper(struct amdgpu_ring * ring)641 int amdgpu_ring_test_helper(struct amdgpu_ring *ring)
642 {
643 	struct amdgpu_device *adev = ring->adev;
644 	int r;
645 
646 	r = amdgpu_ring_test_ring(ring);
647 	if (r)
648 		DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n",
649 			      ring->name, r);
650 	else
651 		DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n",
652 			      ring->name);
653 
654 	ring->sched.ready = !r;
655 
656 	return r;
657 }
658 
amdgpu_ring_to_mqd_prop(struct amdgpu_ring * ring,struct amdgpu_mqd_prop * prop)659 static void amdgpu_ring_to_mqd_prop(struct amdgpu_ring *ring,
660 				    struct amdgpu_mqd_prop *prop)
661 {
662 	struct amdgpu_device *adev = ring->adev;
663 	bool is_high_prio_compute = ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE &&
664 				    amdgpu_gfx_is_high_priority_compute_queue(adev, ring);
665 	bool is_high_prio_gfx = ring->funcs->type == AMDGPU_RING_TYPE_GFX &&
666 				amdgpu_gfx_is_high_priority_graphics_queue(adev, ring);
667 
668 	memset(prop, 0, sizeof(*prop));
669 
670 	prop->mqd_gpu_addr = ring->mqd_gpu_addr;
671 	prop->hqd_base_gpu_addr = ring->gpu_addr;
672 	prop->rptr_gpu_addr = ring->rptr_gpu_addr;
673 	prop->wptr_gpu_addr = ring->wptr_gpu_addr;
674 	prop->queue_size = ring->ring_size;
675 	prop->eop_gpu_addr = ring->eop_gpu_addr;
676 	prop->use_doorbell = ring->use_doorbell;
677 	prop->doorbell_index = ring->doorbell_index;
678 
679 	/* map_queues packet doesn't need activate the queue,
680 	 * so only kiq need set this field.
681 	 */
682 	prop->hqd_active = ring->funcs->type == AMDGPU_RING_TYPE_KIQ;
683 
684 	prop->allow_tunneling = is_high_prio_compute;
685 	if (is_high_prio_compute || is_high_prio_gfx) {
686 		prop->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_HIGH;
687 		prop->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MAXIMUM;
688 	}
689 }
690 
amdgpu_ring_init_mqd(struct amdgpu_ring * ring)691 int amdgpu_ring_init_mqd(struct amdgpu_ring *ring)
692 {
693 	struct amdgpu_device *adev = ring->adev;
694 	struct amdgpu_mqd *mqd_mgr;
695 	struct amdgpu_mqd_prop prop;
696 
697 	amdgpu_ring_to_mqd_prop(ring, &prop);
698 
699 	ring->wptr = 0;
700 
701 	if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
702 		mqd_mgr = &adev->mqds[AMDGPU_HW_IP_COMPUTE];
703 	else
704 		mqd_mgr = &adev->mqds[ring->funcs->type];
705 
706 	return mqd_mgr->init_mqd(adev, ring->mqd_ptr, &prop);
707 }
708 
amdgpu_ring_ib_begin(struct amdgpu_ring * ring)709 void amdgpu_ring_ib_begin(struct amdgpu_ring *ring)
710 {
711 	if (ring->is_sw_ring)
712 		amdgpu_sw_ring_ib_begin(ring);
713 }
714 
amdgpu_ring_ib_end(struct amdgpu_ring * ring)715 void amdgpu_ring_ib_end(struct amdgpu_ring *ring)
716 {
717 	if (ring->is_sw_ring)
718 		amdgpu_sw_ring_ib_end(ring);
719 }
720 
amdgpu_ring_ib_on_emit_cntl(struct amdgpu_ring * ring)721 void amdgpu_ring_ib_on_emit_cntl(struct amdgpu_ring *ring)
722 {
723 	if (ring->is_sw_ring)
724 		amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_CONTROL);
725 }
726 
amdgpu_ring_ib_on_emit_ce(struct amdgpu_ring * ring)727 void amdgpu_ring_ib_on_emit_ce(struct amdgpu_ring *ring)
728 {
729 	if (ring->is_sw_ring)
730 		amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_CE);
731 }
732 
amdgpu_ring_ib_on_emit_de(struct amdgpu_ring * ring)733 void amdgpu_ring_ib_on_emit_de(struct amdgpu_ring *ring)
734 {
735 	if (ring->is_sw_ring)
736 		amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_DE);
737 }
738 
amdgpu_ring_sched_ready(struct amdgpu_ring * ring)739 bool amdgpu_ring_sched_ready(struct amdgpu_ring *ring)
740 {
741 	if (!ring)
742 		return false;
743 
744 	if (ring->no_scheduler || !drm_sched_wqueue_ready(&ring->sched))
745 		return false;
746 
747 	return true;
748 }
749