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