xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c (revision face6a3615a649456eb4549f6d474221d877d604)
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 "amdgpu_ras_mgr.h"
37 #include "atom.h"
38 
39 /*
40  * Rings
41  * Most engines on the GPU are fed via ring buffers.  Ring
42  * buffers are areas of GPU accessible memory that the host
43  * writes commands into and the GPU reads commands out of.
44  * There is a rptr (read pointer) that determines where the
45  * GPU is currently reading, and a wptr (write pointer)
46  * which determines where the host has written.  When the
47  * pointers are equal, the ring is idle.  When the host
48  * writes commands to the ring buffer, it increments the
49  * wptr.  The GPU then starts fetching commands and executes
50  * them until the pointers are equal again.
51  */
52 
53 /**
54  * amdgpu_ring_max_ibs - Return max IBs that fit in a single submission.
55  *
56  * @type: ring type for which to return the limit.
57  */
58 unsigned int amdgpu_ring_max_ibs(enum amdgpu_ring_type type)
59 {
60 	switch (type) {
61 	case AMDGPU_RING_TYPE_GFX:
62 		/* Need to keep at least 192 on GFX7+ for old radv. */
63 		return 192;
64 	case AMDGPU_RING_TYPE_COMPUTE:
65 		return 125;
66 	case AMDGPU_RING_TYPE_VCN_JPEG:
67 		return 16;
68 	default:
69 		return 49;
70 	}
71 }
72 
73 /**
74  * amdgpu_ring_alloc - allocate space on the ring buffer
75  *
76  * @ring: amdgpu_ring structure holding ring information
77  * @ndw: number of dwords to allocate in the ring buffer
78  *
79  * Allocate @ndw dwords in the ring buffer (all asics).
80  * Returns 0 on success, error on failure.
81  */
82 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned int ndw)
83 {
84 	/* Align requested size with padding so unlock_commit can
85 	 * pad safely */
86 	ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
87 
88 	/* Make sure we aren't trying to allocate more space
89 	 * than the maximum for one submission
90 	 */
91 	if (WARN_ON_ONCE(ndw > ring->max_dw))
92 		return -ENOMEM;
93 
94 	ring->count_dw = ndw;
95 	ring->wptr_old = ring->wptr;
96 
97 	if (ring->funcs->begin_use)
98 		ring->funcs->begin_use(ring);
99 
100 	return 0;
101 }
102 
103 /**
104  * amdgpu_ring_alloc_reemit - allocate space on the ring buffer for reemit
105  *
106  * @ring: amdgpu_ring structure holding ring information
107  * @ndw: number of dwords to allocate in the ring buffer
108  *
109  * Allocate @ndw dwords in the ring buffer (all asics).
110  * doesn't check the max_dw limit as we may be reemitting
111  * several submissions.
112  */
113 static void amdgpu_ring_alloc_reemit(struct amdgpu_ring *ring, unsigned int ndw)
114 {
115 	/* Align requested size with padding so unlock_commit can
116 	 * pad safely */
117 	ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
118 
119 	ring->count_dw = ndw;
120 	ring->wptr_old = ring->wptr;
121 
122 	if (ring->funcs->begin_use)
123 		ring->funcs->begin_use(ring);
124 }
125 
126 /** amdgpu_ring_insert_nop - insert NOP packets
127  *
128  * @ring: amdgpu_ring structure holding ring information
129  * @count: the number of NOP packets to insert
130  *
131  * This is the generic insert_nop function for rings except SDMA
132  */
133 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
134 {
135 	uint32_t occupied, chunk1, chunk2;
136 
137 	occupied = ring->wptr & ring->buf_mask;
138 	chunk1 = ring->buf_mask + 1 - occupied;
139 	chunk1 = (chunk1 >= count) ? count : chunk1;
140 	chunk2 = count - chunk1;
141 
142 	if (chunk1)
143 		memset32(&ring->ring[occupied], ring->funcs->nop, chunk1);
144 
145 	if (chunk2)
146 		memset32(ring->ring, ring->funcs->nop, chunk2);
147 
148 	ring->wptr += count;
149 	ring->wptr &= ring->ptr_mask;
150 	ring->count_dw -= count;
151 }
152 
153 /**
154  * amdgpu_ring_generic_pad_ib - pad IB with NOP packets
155  *
156  * @ring: amdgpu_ring structure holding ring information
157  * @ib: IB to add NOP packets to
158  *
159  * This is the generic pad_ib function for rings except SDMA
160  */
161 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
162 {
163 	u32 align_mask = ring->funcs->align_mask;
164 	u32 count = ib->length_dw & align_mask;
165 
166 	if (count) {
167 		count = align_mask + 1 - count;
168 
169 		memset32(&ib->ptr[ib->length_dw], ring->funcs->nop, count);
170 
171 		ib->length_dw += count;
172 	}
173 }
174 
175 /**
176  * amdgpu_ring_commit - tell the GPU to execute the new
177  * commands on the ring buffer
178  *
179  * @ring: amdgpu_ring structure holding ring information
180  *
181  * Update the wptr (write pointer) to tell the GPU to
182  * execute new commands on the ring buffer (all asics).
183  */
184 void amdgpu_ring_commit(struct amdgpu_ring *ring)
185 {
186 	uint32_t count;
187 
188 	if (ring->count_dw < 0)
189 		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
190 
191 	/* We pad to match fetch size */
192 	count = ring->funcs->align_mask + 1 -
193 		(ring->wptr & ring->funcs->align_mask);
194 	count &= ring->funcs->align_mask;
195 
196 	if (count != 0)
197 		ring->funcs->insert_nop(ring, count);
198 
199 	mb();
200 	amdgpu_ring_set_wptr(ring);
201 
202 	if (ring->funcs->end_use)
203 		ring->funcs->end_use(ring);
204 }
205 
206 /**
207  * amdgpu_ring_undo - reset the wptr
208  *
209  * @ring: amdgpu_ring structure holding ring information
210  *
211  * Reset the driver's copy of the wptr (all asics).
212  */
213 void amdgpu_ring_undo(struct amdgpu_ring *ring)
214 {
215 	ring->wptr = ring->wptr_old;
216 
217 	if (ring->funcs->end_use)
218 		ring->funcs->end_use(ring);
219 }
220 
221 #define amdgpu_ring_get_gpu_addr(ring, offset)				\
222 	 (ring->adev->wb.gpu_addr + offset * 4)
223 
224 #define amdgpu_ring_get_cpu_addr(ring, offset)				\
225 	 (&ring->adev->wb.wb[offset])
226 
227 /**
228  * amdgpu_ring_init - init driver ring struct.
229  *
230  * @adev: amdgpu_device pointer
231  * @ring: amdgpu_ring structure holding ring information
232  * @max_dw: maximum number of dw for ring alloc
233  * @irq_src: interrupt source to use for this ring
234  * @irq_type: interrupt type to use for this ring
235  * @hw_prio: ring priority (NORMAL/HIGH)
236  * @sched_score: optional score atomic shared with other schedulers
237  *
238  * Initialize the driver information for the selected ring (all asics).
239  * Returns 0 on success, error on failure.
240  */
241 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
242 		     unsigned int max_dw, struct amdgpu_irq_src *irq_src,
243 		     unsigned int irq_type, unsigned int hw_prio,
244 		     atomic_t *sched_score)
245 {
246 	int r;
247 	int sched_hw_submission = amdgpu_sched_hw_submission;
248 	u32 *num_sched;
249 	u32 hw_ip;
250 	unsigned int max_ibs_dw;
251 
252 	/* Set the hw submission limit higher for KIQ because
253 	 * it's used for a number of gfx/compute tasks by both
254 	 * KFD and KGD which may have outstanding fences and
255 	 * it doesn't really use the gpu scheduler anyway;
256 	 * KIQ tasks get submitted directly to the ring.
257 	 */
258 	if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
259 		sched_hw_submission = max(sched_hw_submission, 256);
260 	if (ring->funcs->type == AMDGPU_RING_TYPE_MES)
261 		sched_hw_submission = 8;
262 	else if (ring == &adev->sdma.instance[0].page)
263 		sched_hw_submission = 256;
264 
265 	if (ring->adev == NULL) {
266 		if (adev->num_rings >= AMDGPU_MAX_RINGS)
267 			return -EINVAL;
268 
269 		ring->adev = adev;
270 		ring->num_hw_submission = sched_hw_submission;
271 		ring->sched_score = sched_score;
272 		ring->vmid_wait = dma_fence_get_stub();
273 
274 		ring->idx = adev->num_rings++;
275 		adev->rings[ring->idx] = ring;
276 
277 		r = amdgpu_fence_driver_init_ring(ring);
278 		if (r)
279 			return r;
280 	}
281 
282 	r = amdgpu_device_wb_get(adev, &ring->rptr_offs);
283 	if (r) {
284 		dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
285 		return r;
286 	}
287 
288 	r = amdgpu_device_wb_get(adev, &ring->wptr_offs);
289 	if (r) {
290 		dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
291 		return r;
292 	}
293 
294 	r = amdgpu_device_wb_get(adev, &ring->fence_offs);
295 	if (r) {
296 		dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
297 		return r;
298 	}
299 
300 	r = amdgpu_device_wb_get(adev, &ring->trail_fence_offs);
301 	if (r) {
302 		dev_err(adev->dev, "(%d) ring trail_fence_offs wb alloc failed\n", r);
303 		return r;
304 	}
305 
306 	r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs);
307 	if (r) {
308 		dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
309 		return r;
310 	}
311 
312 	ring->fence_gpu_addr =
313 		amdgpu_ring_get_gpu_addr(ring, ring->fence_offs);
314 	ring->fence_cpu_addr =
315 		amdgpu_ring_get_cpu_addr(ring, ring->fence_offs);
316 
317 	ring->rptr_gpu_addr =
318 		amdgpu_ring_get_gpu_addr(ring, ring->rptr_offs);
319 	ring->rptr_cpu_addr =
320 		amdgpu_ring_get_cpu_addr(ring, ring->rptr_offs);
321 
322 	ring->wptr_gpu_addr =
323 		amdgpu_ring_get_gpu_addr(ring, ring->wptr_offs);
324 	ring->wptr_cpu_addr =
325 		amdgpu_ring_get_cpu_addr(ring, ring->wptr_offs);
326 
327 	ring->trail_fence_gpu_addr =
328 		amdgpu_ring_get_gpu_addr(ring, ring->trail_fence_offs);
329 	ring->trail_fence_cpu_addr =
330 		amdgpu_ring_get_cpu_addr(ring, ring->trail_fence_offs);
331 
332 	ring->cond_exe_gpu_addr =
333 		amdgpu_ring_get_gpu_addr(ring, ring->cond_exe_offs);
334 	ring->cond_exe_cpu_addr =
335 		amdgpu_ring_get_cpu_addr(ring, ring->cond_exe_offs);
336 
337 	/* always set cond_exec_polling to CONTINUE */
338 	*ring->cond_exe_cpu_addr = 1;
339 
340 	if (ring->funcs->type != AMDGPU_RING_TYPE_CPER) {
341 		r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
342 		if (r) {
343 			dev_err(adev->dev, "failed initializing fences (%d).\n", r);
344 			return r;
345 		}
346 
347 		max_ibs_dw = ring->funcs->emit_frame_size +
348 			     amdgpu_ring_max_ibs(ring->funcs->type) * ring->funcs->emit_ib_size;
349 		max_ibs_dw = (max_ibs_dw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
350 
351 		if (WARN_ON(max_ibs_dw > max_dw))
352 			max_dw = max_ibs_dw;
353 
354 		ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
355 	} else {
356 		ring->ring_size = roundup_pow_of_two(max_dw * 4);
357 		ring->count_dw = (ring->ring_size - 4) >> 2;
358 		/* ring buffer is empty now */
359 		ring->wptr = *ring->rptr_cpu_addr = 0;
360 	}
361 
362 	ring->buf_mask = (ring->ring_size / 4) - 1;
363 	ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
364 		0xffffffffffffffff : ring->buf_mask;
365 	/*  Initialize cached_rptr to 0 */
366 	ring->cached_rptr = 0;
367 
368 	if (!ring->ring_backup) {
369 		ring->ring_backup = kvzalloc(ring->ring_size, GFP_KERNEL);
370 		if (!ring->ring_backup)
371 			return -ENOMEM;
372 	}
373 
374 	/* Allocate ring buffer */
375 	if (ring->ring_obj == NULL) {
376 		r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_bytes,
377 					    PAGE_SIZE,
378 					    AMDGPU_GEM_DOMAIN_GTT,
379 					    &ring->ring_obj,
380 					    &ring->gpu_addr,
381 					    (void **)&ring->ring);
382 		if (r) {
383 			dev_err(adev->dev, "(%d) ring create failed\n", r);
384 			kvfree(ring->ring_backup);
385 			return r;
386 		}
387 		amdgpu_ring_clear_ring(ring);
388 	}
389 
390 	ring->max_dw = max_dw;
391 	ring->hw_prio = hw_prio;
392 
393 	if (!ring->no_scheduler && ring->funcs->type < AMDGPU_HW_IP_NUM) {
394 		hw_ip = ring->funcs->type;
395 		num_sched = &adev->gpu_sched[hw_ip][hw_prio].num_scheds;
396 		adev->gpu_sched[hw_ip][hw_prio].sched[(*num_sched)++] =
397 			&ring->sched;
398 	}
399 
400 	return 0;
401 }
402 
403 /**
404  * amdgpu_ring_fini - tear down the driver ring struct.
405  *
406  * @ring: amdgpu_ring structure holding ring information
407  *
408  * Tear down the driver information for the selected ring (all asics).
409  */
410 void amdgpu_ring_fini(struct amdgpu_ring *ring)
411 {
412 
413 	/* Not to finish a ring which is not initialized */
414 	if (!(ring->adev) || !(ring->adev->rings[ring->idx]))
415 		return;
416 
417 	ring->sched.ready = false;
418 
419 	amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
420 	amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
421 
422 	amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
423 	amdgpu_device_wb_free(ring->adev, ring->fence_offs);
424 
425 	amdgpu_bo_free_kernel(&ring->ring_obj,
426 			      &ring->gpu_addr,
427 			      (void **)&ring->ring);
428 	kvfree(ring->ring_backup);
429 	ring->ring_backup = NULL;
430 
431 	dma_fence_put(ring->vmid_wait);
432 	ring->vmid_wait = NULL;
433 	ring->me = 0;
434 }
435 
436 /**
437  * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper
438  *
439  * @ring: ring to write to
440  * @reg0: register to write
441  * @reg1: register to wait on
442  * @ref: reference value to write/wait on
443  * @mask: mask to wait on
444  *
445  * Helper for rings that don't support write and wait in a
446  * single oneshot packet.
447  */
448 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
449 						uint32_t reg0, uint32_t reg1,
450 						uint32_t ref, uint32_t mask)
451 {
452 	amdgpu_ring_emit_wreg(ring, reg0, ref);
453 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
454 }
455 
456 /**
457  * amdgpu_ring_soft_recovery - try to soft recover a ring lockup
458  *
459  * @ring: ring to try the recovery on
460  * @vmid: VMID we try to get going again
461  * @fence: timedout fence
462  *
463  * Tries to get a ring proceeding again when it is stuck.
464  */
465 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
466 			       struct dma_fence *fence)
467 {
468 	unsigned long flags;
469 	ktime_t deadline;
470 	bool ret;
471 
472 	deadline = ktime_add_us(ktime_get(), 10000);
473 
474 	if (amdgpu_sriov_vf(ring->adev) || !ring->funcs->soft_recovery || !fence)
475 		return false;
476 
477 	spin_lock_irqsave(fence->lock, flags);
478 	if (!dma_fence_is_signaled_locked(fence))
479 		dma_fence_set_error(fence, -ENODATA);
480 	spin_unlock_irqrestore(fence->lock, flags);
481 
482 	while (!dma_fence_is_signaled(fence) &&
483 	       ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0)
484 		ring->funcs->soft_recovery(ring, vmid);
485 
486 	ret = dma_fence_is_signaled(fence);
487 	/* increment the counter only if soft reset worked */
488 	if (ret)
489 		atomic_inc(&ring->adev->gpu_reset_counter);
490 
491 	return ret;
492 }
493 
494 /*
495  * Debugfs info
496  */
497 #if defined(CONFIG_DEBUG_FS)
498 
499 static ssize_t amdgpu_ras_cper_debugfs_read(struct file *f, char __user *buf,
500 					    size_t size, loff_t *offset)
501 {
502 	const uint8_t ring_header_size = 12;
503 	struct amdgpu_ring *ring = file_inode(f)->i_private;
504 	struct ras_cmd_cper_snapshot_req *snapshot_req __free(kfree) =
505 		kzalloc(sizeof(struct ras_cmd_cper_snapshot_req), GFP_KERNEL);
506 	struct ras_cmd_cper_snapshot_rsp *snapshot_rsp __free(kfree) =
507 		kzalloc(sizeof(struct ras_cmd_cper_snapshot_rsp), GFP_KERNEL);
508 	struct ras_cmd_cper_record_req *record_req __free(kfree) =
509 		kzalloc(sizeof(struct ras_cmd_cper_record_req), GFP_KERNEL);
510 	struct ras_cmd_cper_record_rsp *record_rsp __free(kfree) =
511 		kzalloc(sizeof(struct ras_cmd_cper_record_rsp), GFP_KERNEL);
512 	uint8_t *ring_header __free(kfree) =
513 		kzalloc(ring_header_size, GFP_KERNEL);
514 	uint32_t total_cper_num;
515 	uint64_t start_cper_id;
516 	int r;
517 
518 	if (!snapshot_req || !snapshot_rsp || !record_req || !record_rsp ||
519 	    !ring_header)
520 		return -ENOMEM;
521 
522 	if (!(*offset)) {
523 		if (copy_to_user(buf, ring_header, ring_header_size))
524 			return -EFAULT;
525 		buf += ring_header_size;
526 	}
527 
528 	r = amdgpu_ras_mgr_handle_ras_cmd(ring->adev,
529 					  RAS_CMD__GET_CPER_SNAPSHOT,
530 					  snapshot_req, sizeof(struct ras_cmd_cper_snapshot_req),
531 					  snapshot_rsp, sizeof(struct ras_cmd_cper_snapshot_rsp));
532 	if (r || !snapshot_rsp->total_cper_num)
533 		return r;
534 
535 	start_cper_id = snapshot_rsp->start_cper_id;
536 	total_cper_num = snapshot_rsp->total_cper_num;
537 
538 	record_req->buf_ptr = (uint64_t)(uintptr_t)buf;
539 	record_req->buf_size = size;
540 	record_req->cper_start_id = start_cper_id + *offset;
541 	record_req->cper_num = total_cper_num;
542 	r = amdgpu_ras_mgr_handle_ras_cmd(ring->adev, RAS_CMD__GET_CPER_RECORD,
543 					  record_req, sizeof(struct ras_cmd_cper_record_req),
544 					  record_rsp, sizeof(struct ras_cmd_cper_record_rsp));
545 	if (r)
546 		return r;
547 
548 	r = *offset ? record_rsp->real_data_size : record_rsp->real_data_size + ring_header_size;
549 	(*offset) += record_rsp->real_cper_num;
550 
551 	return r;
552 }
553 
554 /* Layout of file is 12 bytes consisting of
555  * - rptr
556  * - wptr
557  * - driver's copy of wptr
558  *
559  * followed by n-words of ring data
560  */
561 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
562 					size_t size, loff_t *pos)
563 {
564 	struct amdgpu_ring *ring = file_inode(f)->i_private;
565 	uint32_t value, result, early[3];
566 	uint64_t p;
567 	loff_t i;
568 	int r;
569 
570 	if (ring->funcs->type == AMDGPU_RING_TYPE_CPER && amdgpu_uniras_enabled(ring->adev))
571 		return amdgpu_ras_cper_debugfs_read(f, buf, size, pos);
572 
573 	if (*pos & 3 || size & 3)
574 		return -EINVAL;
575 
576 	result = 0;
577 
578 	if (*pos < 12) {
579 		if (ring->funcs->type == AMDGPU_RING_TYPE_CPER)
580 			mutex_lock(&ring->adev->cper.ring_lock);
581 
582 		early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
583 		early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
584 		early[2] = ring->wptr & ring->buf_mask;
585 		for (i = *pos / 4; i < 3 && size; i++) {
586 			r = put_user(early[i], (uint32_t *)buf);
587 			if (r) {
588 				result = r;
589 				goto out;
590 			}
591 			buf += 4;
592 			result += 4;
593 			size -= 4;
594 			*pos += 4;
595 		}
596 	}
597 
598 	if (ring->funcs->type != AMDGPU_RING_TYPE_CPER) {
599 		while (size) {
600 			if (*pos >= (ring->ring_size + 12))
601 				return result;
602 
603 			value = ring->ring[(*pos - 12)/4];
604 			r = put_user(value, (uint32_t *)buf);
605 			if (r)
606 				return r;
607 			buf += 4;
608 			result += 4;
609 			size -= 4;
610 			*pos += 4;
611 		}
612 	} else {
613 		p = early[0];
614 		if (early[0] <= early[1])
615 			size = (early[1] - early[0]);
616 		else
617 			size = ring->ring_size - (early[0] - early[1]);
618 
619 		while (size) {
620 			if (p == early[1])
621 				goto out;
622 
623 			value = ring->ring[p];
624 			r = put_user(value, (uint32_t *)buf);
625 			if (r) {
626 				result = r;
627 				goto out;
628 			}
629 
630 			buf += 4;
631 			result += 4;
632 			size--;
633 			p++;
634 			p &= ring->ptr_mask;
635 		}
636 	}
637 
638 out:
639 	if (ring->funcs->type == AMDGPU_RING_TYPE_CPER)
640 		mutex_unlock(&ring->adev->cper.ring_lock);
641 
642 	return result;
643 }
644 
645 static ssize_t amdgpu_debugfs_virt_ring_read(struct file *f, char __user *buf,
646 	size_t size, loff_t *pos)
647 {
648 	struct amdgpu_ring *ring = file_inode(f)->i_private;
649 
650 	if (*pos & 3 || size & 3)
651 		return -EINVAL;
652 
653 	if (ring->funcs->type == AMDGPU_RING_TYPE_CPER)
654 		amdgpu_virt_req_ras_cper_dump(ring->adev, false);
655 
656 	return amdgpu_debugfs_ring_read(f, buf, size, pos);
657 }
658 
659 static const struct file_operations amdgpu_debugfs_ring_fops = {
660 	.owner = THIS_MODULE,
661 	.read = amdgpu_debugfs_ring_read,
662 	.llseek = default_llseek
663 };
664 
665 static const struct file_operations amdgpu_debugfs_virt_ring_fops = {
666 	.owner = THIS_MODULE,
667 	.read = amdgpu_debugfs_virt_ring_read,
668 	.llseek = default_llseek
669 };
670 
671 static ssize_t amdgpu_debugfs_mqd_read(struct file *f, char __user *buf,
672 				       size_t size, loff_t *pos)
673 {
674 	struct amdgpu_ring *ring = file_inode(f)->i_private;
675 	ssize_t bytes = min_t(ssize_t, ring->mqd_size - *pos, size);
676 	void *from = ((u8 *)ring->mqd_ptr) + *pos;
677 
678 	if (*pos > ring->mqd_size)
679 		return 0;
680 
681 	if (copy_to_user(buf, from, bytes))
682 		return -EFAULT;
683 
684 	*pos += bytes;
685 	return bytes;
686 }
687 
688 static const struct file_operations amdgpu_debugfs_mqd_fops = {
689 	.owner = THIS_MODULE,
690 	.read = amdgpu_debugfs_mqd_read,
691 	.llseek = default_llseek
692 };
693 
694 static int amdgpu_debugfs_ring_error(void *data, u64 val)
695 {
696 	struct amdgpu_ring *ring = data;
697 
698 	amdgpu_fence_driver_set_error(ring, val);
699 	return 0;
700 }
701 
702 DEFINE_DEBUGFS_ATTRIBUTE_SIGNED(amdgpu_debugfs_error_fops, NULL,
703 				amdgpu_debugfs_ring_error, "%lld\n");
704 
705 #endif
706 
707 void amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
708 			      struct amdgpu_ring *ring)
709 {
710 #if defined(CONFIG_DEBUG_FS)
711 	struct drm_minor *minor = adev_to_drm(adev)->primary;
712 	struct dentry *root = minor->debugfs_root;
713 	char name[32];
714 
715 	sprintf(name, "amdgpu_ring_%s", ring->name);
716 	if (amdgpu_sriov_vf(adev))
717 		debugfs_create_file_size(name, S_IFREG | 0444, root, ring,
718 					 &amdgpu_debugfs_virt_ring_fops,
719 					 ring->ring_size + 12);
720 	else
721 		debugfs_create_file_size(name, S_IFREG | 0444, root, ring,
722 					 &amdgpu_debugfs_ring_fops,
723 					 ring->ring_size + 12);
724 
725 	if (ring->mqd_obj) {
726 		sprintf(name, "amdgpu_mqd_%s", ring->name);
727 		debugfs_create_file_size(name, S_IFREG | 0444, root, ring,
728 					 &amdgpu_debugfs_mqd_fops,
729 					 ring->mqd_size);
730 	}
731 
732 	sprintf(name, "amdgpu_error_%s", ring->name);
733 	debugfs_create_file(name, 0200, root, ring,
734 			    &amdgpu_debugfs_error_fops);
735 
736 #endif
737 }
738 
739 /**
740  * amdgpu_ring_test_helper - tests ring and set sched readiness status
741  *
742  * @ring: ring to try the recovery on
743  *
744  * Tests ring and set sched readiness status
745  *
746  * Returns 0 on success, error on failure.
747  */
748 int amdgpu_ring_test_helper(struct amdgpu_ring *ring)
749 {
750 	struct amdgpu_device *adev = ring->adev;
751 	int r;
752 
753 	r = amdgpu_ring_test_ring(ring);
754 	if (r)
755 		DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n",
756 			      ring->name, r);
757 	else
758 		DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n",
759 			      ring->name);
760 
761 	ring->sched.ready = !r;
762 
763 	return r;
764 }
765 
766 static void amdgpu_ring_to_mqd_prop(struct amdgpu_ring *ring,
767 				    struct amdgpu_mqd_prop *prop)
768 {
769 	struct amdgpu_device *adev = ring->adev;
770 	bool is_high_prio_compute = ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE &&
771 				    amdgpu_gfx_is_high_priority_compute_queue(adev, ring);
772 	bool is_high_prio_gfx = ring->funcs->type == AMDGPU_RING_TYPE_GFX &&
773 				amdgpu_gfx_is_high_priority_graphics_queue(adev, ring);
774 
775 	memset(prop, 0, sizeof(*prop));
776 
777 	prop->mqd_gpu_addr = ring->mqd_gpu_addr;
778 	prop->hqd_base_gpu_addr = ring->gpu_addr;
779 	prop->rptr_gpu_addr = ring->rptr_gpu_addr;
780 	prop->wptr_gpu_addr = ring->wptr_gpu_addr;
781 	prop->queue_size = ring->ring_size;
782 	prop->eop_gpu_addr = ring->eop_gpu_addr;
783 	prop->use_doorbell = ring->use_doorbell;
784 	prop->doorbell_index = ring->doorbell_index;
785 	prop->kernel_queue = true;
786 
787 	/* map_queues packet doesn't need activate the queue,
788 	 * so only kiq need set this field.
789 	 */
790 	prop->hqd_active = ring->funcs->type == AMDGPU_RING_TYPE_KIQ;
791 
792 	prop->allow_tunneling = is_high_prio_compute;
793 	if (is_high_prio_compute || is_high_prio_gfx) {
794 		prop->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_HIGH;
795 		prop->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MAXIMUM;
796 	}
797 }
798 
799 int amdgpu_ring_init_mqd(struct amdgpu_ring *ring)
800 {
801 	struct amdgpu_device *adev = ring->adev;
802 	struct amdgpu_mqd *mqd_mgr;
803 	struct amdgpu_mqd_prop prop;
804 
805 	amdgpu_ring_to_mqd_prop(ring, &prop);
806 
807 	ring->wptr = 0;
808 
809 	if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
810 		mqd_mgr = &adev->mqds[AMDGPU_HW_IP_COMPUTE];
811 	else
812 		mqd_mgr = &adev->mqds[ring->funcs->type];
813 
814 	return mqd_mgr->init_mqd(adev, ring->mqd_ptr, &prop);
815 }
816 
817 void amdgpu_ring_ib_begin(struct amdgpu_ring *ring)
818 {
819 	if (ring->is_sw_ring)
820 		amdgpu_sw_ring_ib_begin(ring);
821 }
822 
823 void amdgpu_ring_ib_end(struct amdgpu_ring *ring)
824 {
825 	if (ring->is_sw_ring)
826 		amdgpu_sw_ring_ib_end(ring);
827 }
828 
829 void amdgpu_ring_ib_on_emit_cntl(struct amdgpu_ring *ring)
830 {
831 	if (ring->is_sw_ring)
832 		amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_CONTROL);
833 }
834 
835 void amdgpu_ring_ib_on_emit_ce(struct amdgpu_ring *ring)
836 {
837 	if (ring->is_sw_ring)
838 		amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_CE);
839 }
840 
841 void amdgpu_ring_ib_on_emit_de(struct amdgpu_ring *ring)
842 {
843 	if (ring->is_sw_ring)
844 		amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_DE);
845 }
846 
847 bool amdgpu_ring_sched_ready(struct amdgpu_ring *ring)
848 {
849 	if (!ring)
850 		return false;
851 
852 	if (ring->no_scheduler || !drm_sched_wqueue_ready(&ring->sched))
853 		return false;
854 
855 	return true;
856 }
857 
858 void amdgpu_ring_reset_helper_begin(struct amdgpu_ring *ring,
859 				    struct amdgpu_fence *guilty_fence)
860 {
861 	/* Stop the scheduler to prevent anybody else from touching the ring buffer. */
862 	drm_sched_wqueue_stop(&ring->sched);
863 	/* back up the non-guilty commands */
864 	amdgpu_ring_backup_unprocessed_commands(ring, guilty_fence);
865 }
866 
867 int amdgpu_ring_reset_helper_end(struct amdgpu_ring *ring,
868 				 struct amdgpu_fence *guilty_fence)
869 {
870 	unsigned int i;
871 	int r;
872 
873 	/* verify that the ring is functional */
874 	r = amdgpu_ring_test_ring(ring);
875 	if (r)
876 		return r;
877 
878 	/* signal the guilty fence and set an error on all fences from the context */
879 	if (guilty_fence)
880 		amdgpu_fence_driver_guilty_force_completion(guilty_fence);
881 	/* Re-emit the non-guilty commands */
882 	if (ring->ring_backup_entries_to_copy) {
883 		amdgpu_ring_alloc_reemit(ring, ring->ring_backup_entries_to_copy);
884 		for (i = 0; i < ring->ring_backup_entries_to_copy; i++)
885 			amdgpu_ring_write(ring, ring->ring_backup[i]);
886 		amdgpu_ring_commit(ring);
887 	}
888 	/* Start the scheduler again */
889 	drm_sched_wqueue_start(&ring->sched);
890 	return 0;
891 }
892 
893 bool amdgpu_ring_is_reset_type_supported(struct amdgpu_ring *ring,
894 					 u32 reset_type)
895 {
896 	switch (ring->funcs->type) {
897 	case AMDGPU_RING_TYPE_GFX:
898 		if (ring->adev->gfx.gfx_supported_reset & reset_type)
899 			return true;
900 		break;
901 	case AMDGPU_RING_TYPE_COMPUTE:
902 		if (ring->adev->gfx.compute_supported_reset & reset_type)
903 			return true;
904 		break;
905 	case AMDGPU_RING_TYPE_SDMA:
906 		if (ring->adev->sdma.supported_reset & reset_type)
907 			return true;
908 		break;
909 	case AMDGPU_RING_TYPE_VCN_DEC:
910 	case AMDGPU_RING_TYPE_VCN_ENC:
911 		if (ring->adev->vcn.supported_reset & reset_type)
912 			return true;
913 		break;
914 	case AMDGPU_RING_TYPE_VCN_JPEG:
915 		if (ring->adev->jpeg.supported_reset & reset_type)
916 			return true;
917 		break;
918 	default:
919 		break;
920 	}
921 	return false;
922 }
923