Loading...
1/*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26#include <linux/firmware.h>
27#include "amdgpu.h"
28#include "amdgpu_gfx.h"
29#include "amdgpu_rlc.h"
30#include "amdgpu_ras.h"
31#include "amdgpu_xcp.h"
32#include "amdgpu_xgmi.h"
33
34/* delay 0.1 second to enable gfx off feature */
35#define GFX_OFF_DELAY_ENABLE msecs_to_jiffies(100)
36
37#define GFX_OFF_NO_DELAY 0
38
39/*
40 * GPU GFX IP block helpers function.
41 */
42
43int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
44 int pipe, int queue)
45{
46 int bit = 0;
47
48 bit += mec * adev->gfx.mec.num_pipe_per_mec
49 * adev->gfx.mec.num_queue_per_pipe;
50 bit += pipe * adev->gfx.mec.num_queue_per_pipe;
51 bit += queue;
52
53 return bit;
54}
55
56void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
57 int *mec, int *pipe, int *queue)
58{
59 *queue = bit % adev->gfx.mec.num_queue_per_pipe;
60 *pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
61 % adev->gfx.mec.num_pipe_per_mec;
62 *mec = (bit / adev->gfx.mec.num_queue_per_pipe)
63 / adev->gfx.mec.num_pipe_per_mec;
64
65}
66
67bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
68 int xcc_id, int mec, int pipe, int queue)
69{
70 return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
71 adev->gfx.mec_bitmap[xcc_id].queue_bitmap);
72}
73
74int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
75 int me, int pipe, int queue)
76{
77 int bit = 0;
78
79 bit += me * adev->gfx.me.num_pipe_per_me
80 * adev->gfx.me.num_queue_per_pipe;
81 bit += pipe * adev->gfx.me.num_queue_per_pipe;
82 bit += queue;
83
84 return bit;
85}
86
87void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
88 int *me, int *pipe, int *queue)
89{
90 *queue = bit % adev->gfx.me.num_queue_per_pipe;
91 *pipe = (bit / adev->gfx.me.num_queue_per_pipe)
92 % adev->gfx.me.num_pipe_per_me;
93 *me = (bit / adev->gfx.me.num_queue_per_pipe)
94 / adev->gfx.me.num_pipe_per_me;
95}
96
97bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
98 int me, int pipe, int queue)
99{
100 return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
101 adev->gfx.me.queue_bitmap);
102}
103
104/**
105 * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
106 *
107 * @mask: array in which the per-shader array disable masks will be stored
108 * @max_se: number of SEs
109 * @max_sh: number of SHs
110 *
111 * The bitmask of CUs to be disabled in the shader array determined by se and
112 * sh is stored in mask[se * max_sh + sh].
113 */
114void amdgpu_gfx_parse_disable_cu(unsigned int *mask, unsigned int max_se, unsigned int max_sh)
115{
116 unsigned int se, sh, cu;
117 const char *p;
118
119 memset(mask, 0, sizeof(*mask) * max_se * max_sh);
120
121 if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
122 return;
123
124 p = amdgpu_disable_cu;
125 for (;;) {
126 char *next;
127 int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
128
129 if (ret < 3) {
130 DRM_ERROR("amdgpu: could not parse disable_cu\n");
131 return;
132 }
133
134 if (se < max_se && sh < max_sh && cu < 16) {
135 DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
136 mask[se * max_sh + sh] |= 1u << cu;
137 } else {
138 DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
139 se, sh, cu);
140 }
141
142 next = strchr(p, ',');
143 if (!next)
144 break;
145 p = next + 1;
146 }
147}
148
149static bool amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device *adev)
150{
151 return amdgpu_async_gfx_ring && adev->gfx.me.num_pipe_per_me > 1;
152}
153
154static bool amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device *adev)
155{
156 if (amdgpu_compute_multipipe != -1) {
157 DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
158 amdgpu_compute_multipipe);
159 return amdgpu_compute_multipipe == 1;
160 }
161
162 if (amdgpu_ip_version(adev, GC_HWIP, 0) > IP_VERSION(9, 0, 0))
163 return true;
164
165 /* FIXME: spreading the queues across pipes causes perf regressions
166 * on POLARIS11 compute workloads */
167 if (adev->asic_type == CHIP_POLARIS11)
168 return false;
169
170 return adev->gfx.mec.num_mec > 1;
171}
172
173bool amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device *adev,
174 struct amdgpu_ring *ring)
175{
176 int queue = ring->queue;
177 int pipe = ring->pipe;
178
179 /* Policy: use pipe1 queue0 as high priority graphics queue if we
180 * have more than one gfx pipe.
181 */
182 if (amdgpu_gfx_is_graphics_multipipe_capable(adev) &&
183 adev->gfx.num_gfx_rings > 1 && pipe == 1 && queue == 0) {
184 int me = ring->me;
185 int bit;
186
187 bit = amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue);
188 if (ring == &adev->gfx.gfx_ring[bit])
189 return true;
190 }
191
192 return false;
193}
194
195bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
196 struct amdgpu_ring *ring)
197{
198 /* Policy: use 1st queue as high priority compute queue if we
199 * have more than one compute queue.
200 */
201 if (adev->gfx.num_compute_rings > 1 &&
202 ring == &adev->gfx.compute_ring[0])
203 return true;
204
205 return false;
206}
207
208void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
209{
210 int i, j, queue, pipe;
211 bool multipipe_policy = amdgpu_gfx_is_compute_multipipe_capable(adev);
212 int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
213 adev->gfx.mec.num_queue_per_pipe,
214 adev->gfx.num_compute_rings);
215 int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
216
217 if (multipipe_policy) {
218 /* policy: make queues evenly cross all pipes on MEC1 only
219 * for multiple xcc, just use the original policy for simplicity */
220 for (j = 0; j < num_xcc; j++) {
221 for (i = 0; i < max_queues_per_mec; i++) {
222 pipe = i % adev->gfx.mec.num_pipe_per_mec;
223 queue = (i / adev->gfx.mec.num_pipe_per_mec) %
224 adev->gfx.mec.num_queue_per_pipe;
225
226 set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
227 adev->gfx.mec_bitmap[j].queue_bitmap);
228 }
229 }
230 } else {
231 /* policy: amdgpu owns all queues in the given pipe */
232 for (j = 0; j < num_xcc; j++) {
233 for (i = 0; i < max_queues_per_mec; ++i)
234 set_bit(i, adev->gfx.mec_bitmap[j].queue_bitmap);
235 }
236 }
237
238 for (j = 0; j < num_xcc; j++) {
239 dev_dbg(adev->dev, "mec queue bitmap weight=%d\n",
240 bitmap_weight(adev->gfx.mec_bitmap[j].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
241 }
242}
243
244void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
245{
246 int i, queue, pipe;
247 bool multipipe_policy = amdgpu_gfx_is_graphics_multipipe_capable(adev);
248 int max_queues_per_me = adev->gfx.me.num_pipe_per_me *
249 adev->gfx.me.num_queue_per_pipe;
250
251 if (multipipe_policy) {
252 /* policy: amdgpu owns the first queue per pipe at this stage
253 * will extend to mulitple queues per pipe later */
254 for (i = 0; i < max_queues_per_me; i++) {
255 pipe = i % adev->gfx.me.num_pipe_per_me;
256 queue = (i / adev->gfx.me.num_pipe_per_me) %
257 adev->gfx.me.num_queue_per_pipe;
258
259 set_bit(pipe * adev->gfx.me.num_queue_per_pipe + queue,
260 adev->gfx.me.queue_bitmap);
261 }
262 } else {
263 for (i = 0; i < max_queues_per_me; ++i)
264 set_bit(i, adev->gfx.me.queue_bitmap);
265 }
266
267 /* update the number of active graphics rings */
268 adev->gfx.num_gfx_rings =
269 bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
270}
271
272static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
273 struct amdgpu_ring *ring, int xcc_id)
274{
275 int queue_bit;
276 int mec, pipe, queue;
277
278 queue_bit = adev->gfx.mec.num_mec
279 * adev->gfx.mec.num_pipe_per_mec
280 * adev->gfx.mec.num_queue_per_pipe;
281
282 while (--queue_bit >= 0) {
283 if (test_bit(queue_bit, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
284 continue;
285
286 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
287
288 /*
289 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
290 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
291 * only can be issued on queue 0.
292 */
293 if ((mec == 1 && pipe > 1) || queue != 0)
294 continue;
295
296 ring->me = mec + 1;
297 ring->pipe = pipe;
298 ring->queue = queue;
299
300 return 0;
301 }
302
303 dev_err(adev->dev, "Failed to find a queue for KIQ\n");
304 return -EINVAL;
305}
306
307int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
308 struct amdgpu_ring *ring,
309 struct amdgpu_irq_src *irq, int xcc_id)
310{
311 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
312 int r = 0;
313
314 spin_lock_init(&kiq->ring_lock);
315
316 ring->adev = NULL;
317 ring->ring_obj = NULL;
318 ring->use_doorbell = true;
319 ring->xcc_id = xcc_id;
320 ring->vm_hub = AMDGPU_GFXHUB(xcc_id);
321 ring->doorbell_index =
322 (adev->doorbell_index.kiq +
323 xcc_id * adev->doorbell_index.xcc_doorbell_range)
324 << 1;
325
326 r = amdgpu_gfx_kiq_acquire(adev, ring, xcc_id);
327 if (r)
328 return r;
329
330 ring->eop_gpu_addr = kiq->eop_gpu_addr;
331 ring->no_scheduler = true;
332 sprintf(ring->name, "kiq_%d.%d.%d.%d", xcc_id, ring->me, ring->pipe, ring->queue);
333 r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
334 AMDGPU_RING_PRIO_DEFAULT, NULL);
335 if (r)
336 dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
337
338 return r;
339}
340
341void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
342{
343 amdgpu_ring_fini(ring);
344}
345
346void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev, int xcc_id)
347{
348 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
349
350 amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
351}
352
353int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
354 unsigned int hpd_size, int xcc_id)
355{
356 int r;
357 u32 *hpd;
358 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
359
360 r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
361 AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
362 &kiq->eop_gpu_addr, (void **)&hpd);
363 if (r) {
364 dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
365 return r;
366 }
367
368 memset(hpd, 0, hpd_size);
369
370 r = amdgpu_bo_reserve(kiq->eop_obj, true);
371 if (unlikely(r != 0))
372 dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
373 amdgpu_bo_kunmap(kiq->eop_obj);
374 amdgpu_bo_unreserve(kiq->eop_obj);
375
376 return 0;
377}
378
379/* create MQD for each compute/gfx queue */
380int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
381 unsigned int mqd_size, int xcc_id)
382{
383 int r, i, j;
384 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
385 struct amdgpu_ring *ring = &kiq->ring;
386 u32 domain = AMDGPU_GEM_DOMAIN_GTT;
387
388#if !defined(CONFIG_ARM) && !defined(CONFIG_ARM64)
389 /* Only enable on gfx10 and 11 for now to avoid changing behavior on older chips */
390 if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(10, 0, 0))
391 domain |= AMDGPU_GEM_DOMAIN_VRAM;
392#endif
393
394 /* create MQD for KIQ */
395 if (!adev->enable_mes_kiq && !ring->mqd_obj) {
396 /* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
397 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
398 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
399 * KIQ MQD no matter SRIOV or Bare-metal
400 */
401 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
402 AMDGPU_GEM_DOMAIN_VRAM |
403 AMDGPU_GEM_DOMAIN_GTT,
404 &ring->mqd_obj,
405 &ring->mqd_gpu_addr,
406 &ring->mqd_ptr);
407 if (r) {
408 dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
409 return r;
410 }
411
412 /* prepare MQD backup */
413 kiq->mqd_backup = kmalloc(mqd_size, GFP_KERNEL);
414 if (!kiq->mqd_backup) {
415 dev_warn(adev->dev,
416 "no memory to create MQD backup for ring %s\n", ring->name);
417 return -ENOMEM;
418 }
419 }
420
421 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
422 /* create MQD for each KGQ */
423 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
424 ring = &adev->gfx.gfx_ring[i];
425 if (!ring->mqd_obj) {
426 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
427 domain, &ring->mqd_obj,
428 &ring->mqd_gpu_addr, &ring->mqd_ptr);
429 if (r) {
430 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
431 return r;
432 }
433
434 ring->mqd_size = mqd_size;
435 /* prepare MQD backup */
436 adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
437 if (!adev->gfx.me.mqd_backup[i]) {
438 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
439 return -ENOMEM;
440 }
441 }
442 }
443 }
444
445 /* create MQD for each KCQ */
446 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
447 j = i + xcc_id * adev->gfx.num_compute_rings;
448 ring = &adev->gfx.compute_ring[j];
449 if (!ring->mqd_obj) {
450 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
451 domain, &ring->mqd_obj,
452 &ring->mqd_gpu_addr, &ring->mqd_ptr);
453 if (r) {
454 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
455 return r;
456 }
457
458 ring->mqd_size = mqd_size;
459 /* prepare MQD backup */
460 adev->gfx.mec.mqd_backup[j] = kmalloc(mqd_size, GFP_KERNEL);
461 if (!adev->gfx.mec.mqd_backup[j]) {
462 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
463 return -ENOMEM;
464 }
465 }
466 }
467
468 return 0;
469}
470
471void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev, int xcc_id)
472{
473 struct amdgpu_ring *ring = NULL;
474 int i, j;
475 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
476
477 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
478 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
479 ring = &adev->gfx.gfx_ring[i];
480 kfree(adev->gfx.me.mqd_backup[i]);
481 amdgpu_bo_free_kernel(&ring->mqd_obj,
482 &ring->mqd_gpu_addr,
483 &ring->mqd_ptr);
484 }
485 }
486
487 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
488 j = i + xcc_id * adev->gfx.num_compute_rings;
489 ring = &adev->gfx.compute_ring[j];
490 kfree(adev->gfx.mec.mqd_backup[j]);
491 amdgpu_bo_free_kernel(&ring->mqd_obj,
492 &ring->mqd_gpu_addr,
493 &ring->mqd_ptr);
494 }
495
496 ring = &kiq->ring;
497 kfree(kiq->mqd_backup);
498 amdgpu_bo_free_kernel(&ring->mqd_obj,
499 &ring->mqd_gpu_addr,
500 &ring->mqd_ptr);
501}
502
503int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
504{
505 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
506 struct amdgpu_ring *kiq_ring = &kiq->ring;
507 struct amdgpu_hive_info *hive;
508 struct amdgpu_ras *ras;
509 int hive_ras_recovery = 0;
510 int i, r = 0;
511 int j;
512
513 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
514 return -EINVAL;
515
516 spin_lock(&kiq->ring_lock);
517 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
518 adev->gfx.num_compute_rings)) {
519 spin_unlock(&kiq->ring_lock);
520 return -ENOMEM;
521 }
522
523 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
524 j = i + xcc_id * adev->gfx.num_compute_rings;
525 kiq->pmf->kiq_unmap_queues(kiq_ring,
526 &adev->gfx.compute_ring[j],
527 RESET_QUEUES, 0, 0);
528 }
529
530 /**
531 * This is workaround: only skip kiq_ring test
532 * during ras recovery in suspend stage for gfx9.4.3
533 */
534 hive = amdgpu_get_xgmi_hive(adev);
535 if (hive) {
536 hive_ras_recovery = atomic_read(&hive->ras_recovery);
537 amdgpu_put_xgmi_hive(hive);
538 }
539
540 ras = amdgpu_ras_get_context(adev);
541 if ((amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) &&
542 ras && (atomic_read(&ras->in_recovery) || hive_ras_recovery)) {
543 spin_unlock(&kiq->ring_lock);
544 return 0;
545 }
546
547 if (kiq_ring->sched.ready && !adev->job_hang)
548 r = amdgpu_ring_test_helper(kiq_ring);
549 spin_unlock(&kiq->ring_lock);
550
551 return r;
552}
553
554int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id)
555{
556 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
557 struct amdgpu_ring *kiq_ring = &kiq->ring;
558 int i, r = 0;
559 int j;
560
561 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
562 return -EINVAL;
563
564 spin_lock(&kiq->ring_lock);
565 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
566 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
567 adev->gfx.num_gfx_rings)) {
568 spin_unlock(&kiq->ring_lock);
569 return -ENOMEM;
570 }
571
572 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
573 j = i + xcc_id * adev->gfx.num_gfx_rings;
574 kiq->pmf->kiq_unmap_queues(kiq_ring,
575 &adev->gfx.gfx_ring[j],
576 PREEMPT_QUEUES, 0, 0);
577 }
578 }
579
580 if (adev->gfx.kiq[0].ring.sched.ready && !adev->job_hang)
581 r = amdgpu_ring_test_helper(kiq_ring);
582 spin_unlock(&kiq->ring_lock);
583
584 return r;
585}
586
587int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
588 int queue_bit)
589{
590 int mec, pipe, queue;
591 int set_resource_bit = 0;
592
593 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
594
595 set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
596
597 return set_resource_bit;
598}
599
600int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
601{
602 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
603 struct amdgpu_ring *kiq_ring = &kiq->ring;
604 uint64_t queue_mask = 0;
605 int r, i, j;
606
607 if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
608 return -EINVAL;
609
610 for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
611 if (!test_bit(i, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
612 continue;
613
614 /* This situation may be hit in the future if a new HW
615 * generation exposes more than 64 queues. If so, the
616 * definition of queue_mask needs updating */
617 if (WARN_ON(i > (sizeof(queue_mask)*8))) {
618 DRM_ERROR("Invalid KCQ enabled: %d\n", i);
619 break;
620 }
621
622 queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
623 }
624
625 DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
626 kiq_ring->queue);
627 amdgpu_device_flush_hdp(adev, NULL);
628
629 spin_lock(&kiq->ring_lock);
630 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
631 adev->gfx.num_compute_rings +
632 kiq->pmf->set_resources_size);
633 if (r) {
634 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
635 spin_unlock(&kiq->ring_lock);
636 return r;
637 }
638
639 if (adev->enable_mes)
640 queue_mask = ~0ULL;
641
642 kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
643 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
644 j = i + xcc_id * adev->gfx.num_compute_rings;
645 kiq->pmf->kiq_map_queues(kiq_ring,
646 &adev->gfx.compute_ring[j]);
647 }
648
649 r = amdgpu_ring_test_helper(kiq_ring);
650 spin_unlock(&kiq->ring_lock);
651 if (r)
652 DRM_ERROR("KCQ enable failed\n");
653
654 return r;
655}
656
657int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
658{
659 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
660 struct amdgpu_ring *kiq_ring = &kiq->ring;
661 int r, i, j;
662
663 if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
664 return -EINVAL;
665
666 amdgpu_device_flush_hdp(adev, NULL);
667
668 spin_lock(&kiq->ring_lock);
669 /* No need to map kcq on the slave */
670 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
671 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
672 adev->gfx.num_gfx_rings);
673 if (r) {
674 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
675 spin_unlock(&kiq->ring_lock);
676 return r;
677 }
678
679 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
680 j = i + xcc_id * adev->gfx.num_gfx_rings;
681 kiq->pmf->kiq_map_queues(kiq_ring,
682 &adev->gfx.gfx_ring[j]);
683 }
684 }
685
686 r = amdgpu_ring_test_helper(kiq_ring);
687 spin_unlock(&kiq->ring_lock);
688 if (r)
689 DRM_ERROR("KCQ enable failed\n");
690
691 return r;
692}
693
694/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
695 *
696 * @adev: amdgpu_device pointer
697 * @bool enable true: enable gfx off feature, false: disable gfx off feature
698 *
699 * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
700 * 2. other client can send request to disable gfx off feature, the request should be honored.
701 * 3. other client can cancel their request of disable gfx off feature
702 * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
703 */
704
705void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
706{
707 unsigned long delay = GFX_OFF_DELAY_ENABLE;
708
709 if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
710 return;
711
712 mutex_lock(&adev->gfx.gfx_off_mutex);
713
714 if (enable) {
715 /* If the count is already 0, it means there's an imbalance bug somewhere.
716 * Note that the bug may be in a different caller than the one which triggers the
717 * WARN_ON_ONCE.
718 */
719 if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
720 goto unlock;
721
722 adev->gfx.gfx_off_req_count--;
723
724 if (adev->gfx.gfx_off_req_count == 0 &&
725 !adev->gfx.gfx_off_state) {
726 /* If going to s2idle, no need to wait */
727 if (adev->in_s0ix) {
728 if (!amdgpu_dpm_set_powergating_by_smu(adev,
729 AMD_IP_BLOCK_TYPE_GFX, true))
730 adev->gfx.gfx_off_state = true;
731 } else {
732 schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
733 delay);
734 }
735 }
736 } else {
737 if (adev->gfx.gfx_off_req_count == 0) {
738 cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
739
740 if (adev->gfx.gfx_off_state &&
741 !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
742 adev->gfx.gfx_off_state = false;
743
744 if (adev->gfx.funcs->init_spm_golden) {
745 dev_dbg(adev->dev,
746 "GFXOFF is disabled, re-init SPM golden settings\n");
747 amdgpu_gfx_init_spm_golden(adev);
748 }
749 }
750 }
751
752 adev->gfx.gfx_off_req_count++;
753 }
754
755unlock:
756 mutex_unlock(&adev->gfx.gfx_off_mutex);
757}
758
759int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
760{
761 int r = 0;
762
763 mutex_lock(&adev->gfx.gfx_off_mutex);
764
765 r = amdgpu_dpm_set_residency_gfxoff(adev, value);
766
767 mutex_unlock(&adev->gfx.gfx_off_mutex);
768
769 return r;
770}
771
772int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
773{
774 int r = 0;
775
776 mutex_lock(&adev->gfx.gfx_off_mutex);
777
778 r = amdgpu_dpm_get_residency_gfxoff(adev, value);
779
780 mutex_unlock(&adev->gfx.gfx_off_mutex);
781
782 return r;
783}
784
785int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
786{
787 int r = 0;
788
789 mutex_lock(&adev->gfx.gfx_off_mutex);
790
791 r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
792
793 mutex_unlock(&adev->gfx.gfx_off_mutex);
794
795 return r;
796}
797
798int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
799{
800
801 int r = 0;
802
803 mutex_lock(&adev->gfx.gfx_off_mutex);
804
805 r = amdgpu_dpm_get_status_gfxoff(adev, value);
806
807 mutex_unlock(&adev->gfx.gfx_off_mutex);
808
809 return r;
810}
811
812int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
813{
814 int r;
815
816 if (amdgpu_ras_is_supported(adev, ras_block->block)) {
817 if (!amdgpu_persistent_edc_harvesting_supported(adev))
818 amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
819
820 r = amdgpu_ras_block_late_init(adev, ras_block);
821 if (r)
822 return r;
823
824 if (adev->gfx.cp_ecc_error_irq.funcs) {
825 r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
826 if (r)
827 goto late_fini;
828 }
829 } else {
830 amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
831 }
832
833 return 0;
834late_fini:
835 amdgpu_ras_block_late_fini(adev, ras_block);
836 return r;
837}
838
839int amdgpu_gfx_ras_sw_init(struct amdgpu_device *adev)
840{
841 int err = 0;
842 struct amdgpu_gfx_ras *ras = NULL;
843
844 /* adev->gfx.ras is NULL, which means gfx does not
845 * support ras function, then do nothing here.
846 */
847 if (!adev->gfx.ras)
848 return 0;
849
850 ras = adev->gfx.ras;
851
852 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
853 if (err) {
854 dev_err(adev->dev, "Failed to register gfx ras block!\n");
855 return err;
856 }
857
858 strcpy(ras->ras_block.ras_comm.name, "gfx");
859 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__GFX;
860 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
861 adev->gfx.ras_if = &ras->ras_block.ras_comm;
862
863 /* If not define special ras_late_init function, use gfx default ras_late_init */
864 if (!ras->ras_block.ras_late_init)
865 ras->ras_block.ras_late_init = amdgpu_gfx_ras_late_init;
866
867 /* If not defined special ras_cb function, use default ras_cb */
868 if (!ras->ras_block.ras_cb)
869 ras->ras_block.ras_cb = amdgpu_gfx_process_ras_data_cb;
870
871 return 0;
872}
873
874int amdgpu_gfx_poison_consumption_handler(struct amdgpu_device *adev,
875 struct amdgpu_iv_entry *entry)
876{
877 if (adev->gfx.ras && adev->gfx.ras->poison_consumption_handler)
878 return adev->gfx.ras->poison_consumption_handler(adev, entry);
879
880 return 0;
881}
882
883int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
884 void *err_data,
885 struct amdgpu_iv_entry *entry)
886{
887 /* TODO ue will trigger an interrupt.
888 *
889 * When “Full RAS” is enabled, the per-IP interrupt sources should
890 * be disabled and the driver should only look for the aggregated
891 * interrupt via sync flood
892 */
893 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
894 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
895 if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
896 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
897 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
898 amdgpu_ras_reset_gpu(adev);
899 }
900 return AMDGPU_RAS_SUCCESS;
901}
902
903int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
904 struct amdgpu_irq_src *source,
905 struct amdgpu_iv_entry *entry)
906{
907 struct ras_common_if *ras_if = adev->gfx.ras_if;
908 struct ras_dispatch_if ih_data = {
909 .entry = entry,
910 };
911
912 if (!ras_if)
913 return 0;
914
915 ih_data.head = *ras_if;
916
917 DRM_ERROR("CP ECC ERROR IRQ\n");
918 amdgpu_ras_interrupt_dispatch(adev, &ih_data);
919 return 0;
920}
921
922void amdgpu_gfx_ras_error_func(struct amdgpu_device *adev,
923 void *ras_error_status,
924 void (*func)(struct amdgpu_device *adev, void *ras_error_status,
925 int xcc_id))
926{
927 int i;
928 int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
929 uint32_t xcc_mask = GENMASK(num_xcc - 1, 0);
930 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
931
932 if (err_data) {
933 err_data->ue_count = 0;
934 err_data->ce_count = 0;
935 }
936
937 for_each_inst(i, xcc_mask)
938 func(adev, ras_error_status, i);
939}
940
941uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg, uint32_t xcc_id)
942{
943 signed long r, cnt = 0;
944 unsigned long flags;
945 uint32_t seq, reg_val_offs = 0, value = 0;
946 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
947 struct amdgpu_ring *ring = &kiq->ring;
948
949 if (amdgpu_device_skip_hw_access(adev))
950 return 0;
951
952 if (adev->mes.ring.sched.ready)
953 return amdgpu_mes_rreg(adev, reg);
954
955 BUG_ON(!ring->funcs->emit_rreg);
956
957 spin_lock_irqsave(&kiq->ring_lock, flags);
958 if (amdgpu_device_wb_get(adev, ®_val_offs)) {
959 pr_err("critical bug! too many kiq readers\n");
960 goto failed_unlock;
961 }
962 amdgpu_ring_alloc(ring, 32);
963 amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
964 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
965 if (r)
966 goto failed_undo;
967
968 amdgpu_ring_commit(ring);
969 spin_unlock_irqrestore(&kiq->ring_lock, flags);
970
971 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
972
973 /* don't wait anymore for gpu reset case because this way may
974 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
975 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
976 * never return if we keep waiting in virt_kiq_rreg, which cause
977 * gpu_recover() hang there.
978 *
979 * also don't wait anymore for IRQ context
980 * */
981 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
982 goto failed_kiq_read;
983
984 might_sleep();
985 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
986 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
987 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
988 }
989
990 if (cnt > MAX_KIQ_REG_TRY)
991 goto failed_kiq_read;
992
993 mb();
994 value = adev->wb.wb[reg_val_offs];
995 amdgpu_device_wb_free(adev, reg_val_offs);
996 return value;
997
998failed_undo:
999 amdgpu_ring_undo(ring);
1000failed_unlock:
1001 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1002failed_kiq_read:
1003 if (reg_val_offs)
1004 amdgpu_device_wb_free(adev, reg_val_offs);
1005 dev_err(adev->dev, "failed to read reg:%x\n", reg);
1006 return ~0;
1007}
1008
1009void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, uint32_t xcc_id)
1010{
1011 signed long r, cnt = 0;
1012 unsigned long flags;
1013 uint32_t seq;
1014 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
1015 struct amdgpu_ring *ring = &kiq->ring;
1016
1017 BUG_ON(!ring->funcs->emit_wreg);
1018
1019 if (amdgpu_device_skip_hw_access(adev))
1020 return;
1021
1022 if (adev->mes.ring.sched.ready) {
1023 amdgpu_mes_wreg(adev, reg, v);
1024 return;
1025 }
1026
1027 spin_lock_irqsave(&kiq->ring_lock, flags);
1028 amdgpu_ring_alloc(ring, 32);
1029 amdgpu_ring_emit_wreg(ring, reg, v);
1030 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
1031 if (r)
1032 goto failed_undo;
1033
1034 amdgpu_ring_commit(ring);
1035 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1036
1037 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1038
1039 /* don't wait anymore for gpu reset case because this way may
1040 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
1041 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
1042 * never return if we keep waiting in virt_kiq_rreg, which cause
1043 * gpu_recover() hang there.
1044 *
1045 * also don't wait anymore for IRQ context
1046 * */
1047 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
1048 goto failed_kiq_write;
1049
1050 might_sleep();
1051 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
1052
1053 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
1054 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1055 }
1056
1057 if (cnt > MAX_KIQ_REG_TRY)
1058 goto failed_kiq_write;
1059
1060 return;
1061
1062failed_undo:
1063 amdgpu_ring_undo(ring);
1064 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1065failed_kiq_write:
1066 dev_err(adev->dev, "failed to write reg:%x\n", reg);
1067}
1068
1069int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
1070{
1071 if (amdgpu_num_kcq == -1) {
1072 return 8;
1073 } else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
1074 dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
1075 return 8;
1076 }
1077 return amdgpu_num_kcq;
1078}
1079
1080void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
1081 uint32_t ucode_id)
1082{
1083 const struct gfx_firmware_header_v1_0 *cp_hdr;
1084 const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
1085 struct amdgpu_firmware_info *info = NULL;
1086 const struct firmware *ucode_fw;
1087 unsigned int fw_size;
1088
1089 switch (ucode_id) {
1090 case AMDGPU_UCODE_ID_CP_PFP:
1091 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1092 adev->gfx.pfp_fw->data;
1093 adev->gfx.pfp_fw_version =
1094 le32_to_cpu(cp_hdr->header.ucode_version);
1095 adev->gfx.pfp_feature_version =
1096 le32_to_cpu(cp_hdr->ucode_feature_version);
1097 ucode_fw = adev->gfx.pfp_fw;
1098 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1099 break;
1100 case AMDGPU_UCODE_ID_CP_RS64_PFP:
1101 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1102 adev->gfx.pfp_fw->data;
1103 adev->gfx.pfp_fw_version =
1104 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1105 adev->gfx.pfp_feature_version =
1106 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1107 ucode_fw = adev->gfx.pfp_fw;
1108 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1109 break;
1110 case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
1111 case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
1112 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1113 adev->gfx.pfp_fw->data;
1114 ucode_fw = adev->gfx.pfp_fw;
1115 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1116 break;
1117 case AMDGPU_UCODE_ID_CP_ME:
1118 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1119 adev->gfx.me_fw->data;
1120 adev->gfx.me_fw_version =
1121 le32_to_cpu(cp_hdr->header.ucode_version);
1122 adev->gfx.me_feature_version =
1123 le32_to_cpu(cp_hdr->ucode_feature_version);
1124 ucode_fw = adev->gfx.me_fw;
1125 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1126 break;
1127 case AMDGPU_UCODE_ID_CP_RS64_ME:
1128 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1129 adev->gfx.me_fw->data;
1130 adev->gfx.me_fw_version =
1131 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1132 adev->gfx.me_feature_version =
1133 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1134 ucode_fw = adev->gfx.me_fw;
1135 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1136 break;
1137 case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
1138 case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
1139 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1140 adev->gfx.me_fw->data;
1141 ucode_fw = adev->gfx.me_fw;
1142 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1143 break;
1144 case AMDGPU_UCODE_ID_CP_CE:
1145 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1146 adev->gfx.ce_fw->data;
1147 adev->gfx.ce_fw_version =
1148 le32_to_cpu(cp_hdr->header.ucode_version);
1149 adev->gfx.ce_feature_version =
1150 le32_to_cpu(cp_hdr->ucode_feature_version);
1151 ucode_fw = adev->gfx.ce_fw;
1152 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1153 break;
1154 case AMDGPU_UCODE_ID_CP_MEC1:
1155 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1156 adev->gfx.mec_fw->data;
1157 adev->gfx.mec_fw_version =
1158 le32_to_cpu(cp_hdr->header.ucode_version);
1159 adev->gfx.mec_feature_version =
1160 le32_to_cpu(cp_hdr->ucode_feature_version);
1161 ucode_fw = adev->gfx.mec_fw;
1162 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1163 le32_to_cpu(cp_hdr->jt_size) * 4;
1164 break;
1165 case AMDGPU_UCODE_ID_CP_MEC1_JT:
1166 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1167 adev->gfx.mec_fw->data;
1168 ucode_fw = adev->gfx.mec_fw;
1169 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1170 break;
1171 case AMDGPU_UCODE_ID_CP_MEC2:
1172 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1173 adev->gfx.mec2_fw->data;
1174 adev->gfx.mec2_fw_version =
1175 le32_to_cpu(cp_hdr->header.ucode_version);
1176 adev->gfx.mec2_feature_version =
1177 le32_to_cpu(cp_hdr->ucode_feature_version);
1178 ucode_fw = adev->gfx.mec2_fw;
1179 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1180 le32_to_cpu(cp_hdr->jt_size) * 4;
1181 break;
1182 case AMDGPU_UCODE_ID_CP_MEC2_JT:
1183 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1184 adev->gfx.mec2_fw->data;
1185 ucode_fw = adev->gfx.mec2_fw;
1186 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1187 break;
1188 case AMDGPU_UCODE_ID_CP_RS64_MEC:
1189 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1190 adev->gfx.mec_fw->data;
1191 adev->gfx.mec_fw_version =
1192 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1193 adev->gfx.mec_feature_version =
1194 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1195 ucode_fw = adev->gfx.mec_fw;
1196 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1197 break;
1198 case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
1199 case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
1200 case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
1201 case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
1202 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1203 adev->gfx.mec_fw->data;
1204 ucode_fw = adev->gfx.mec_fw;
1205 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1206 break;
1207 default:
1208 break;
1209 }
1210
1211 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1212 info = &adev->firmware.ucode[ucode_id];
1213 info->ucode_id = ucode_id;
1214 info->fw = ucode_fw;
1215 adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE);
1216 }
1217}
1218
1219bool amdgpu_gfx_is_master_xcc(struct amdgpu_device *adev, int xcc_id)
1220{
1221 return !(xcc_id % (adev->gfx.num_xcc_per_xcp ?
1222 adev->gfx.num_xcc_per_xcp : 1));
1223}
1224
1225static ssize_t amdgpu_gfx_get_current_compute_partition(struct device *dev,
1226 struct device_attribute *addr,
1227 char *buf)
1228{
1229 struct drm_device *ddev = dev_get_drvdata(dev);
1230 struct amdgpu_device *adev = drm_to_adev(ddev);
1231 int mode;
1232
1233 mode = amdgpu_xcp_query_partition_mode(adev->xcp_mgr,
1234 AMDGPU_XCP_FL_NONE);
1235
1236 return sysfs_emit(buf, "%s\n", amdgpu_gfx_compute_mode_desc(mode));
1237}
1238
1239static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev,
1240 struct device_attribute *addr,
1241 const char *buf, size_t count)
1242{
1243 struct drm_device *ddev = dev_get_drvdata(dev);
1244 struct amdgpu_device *adev = drm_to_adev(ddev);
1245 enum amdgpu_gfx_partition mode;
1246 int ret = 0, num_xcc;
1247
1248 num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1249 if (num_xcc % 2 != 0)
1250 return -EINVAL;
1251
1252 if (!strncasecmp("SPX", buf, strlen("SPX"))) {
1253 mode = AMDGPU_SPX_PARTITION_MODE;
1254 } else if (!strncasecmp("DPX", buf, strlen("DPX"))) {
1255 /*
1256 * DPX mode needs AIDs to be in multiple of 2.
1257 * Each AID connects 2 XCCs.
1258 */
1259 if (num_xcc%4)
1260 return -EINVAL;
1261 mode = AMDGPU_DPX_PARTITION_MODE;
1262 } else if (!strncasecmp("TPX", buf, strlen("TPX"))) {
1263 if (num_xcc != 6)
1264 return -EINVAL;
1265 mode = AMDGPU_TPX_PARTITION_MODE;
1266 } else if (!strncasecmp("QPX", buf, strlen("QPX"))) {
1267 if (num_xcc != 8)
1268 return -EINVAL;
1269 mode = AMDGPU_QPX_PARTITION_MODE;
1270 } else if (!strncasecmp("CPX", buf, strlen("CPX"))) {
1271 mode = AMDGPU_CPX_PARTITION_MODE;
1272 } else {
1273 return -EINVAL;
1274 }
1275
1276 ret = amdgpu_xcp_switch_partition_mode(adev->xcp_mgr, mode);
1277
1278 if (ret)
1279 return ret;
1280
1281 return count;
1282}
1283
1284static ssize_t amdgpu_gfx_get_available_compute_partition(struct device *dev,
1285 struct device_attribute *addr,
1286 char *buf)
1287{
1288 struct drm_device *ddev = dev_get_drvdata(dev);
1289 struct amdgpu_device *adev = drm_to_adev(ddev);
1290 char *supported_partition;
1291
1292 /* TBD */
1293 switch (NUM_XCC(adev->gfx.xcc_mask)) {
1294 case 8:
1295 supported_partition = "SPX, DPX, QPX, CPX";
1296 break;
1297 case 6:
1298 supported_partition = "SPX, TPX, CPX";
1299 break;
1300 case 4:
1301 supported_partition = "SPX, DPX, CPX";
1302 break;
1303 /* this seems only existing in emulation phase */
1304 case 2:
1305 supported_partition = "SPX, CPX";
1306 break;
1307 default:
1308 supported_partition = "Not supported";
1309 break;
1310 }
1311
1312 return sysfs_emit(buf, "%s\n", supported_partition);
1313}
1314
1315static DEVICE_ATTR(current_compute_partition, 0644,
1316 amdgpu_gfx_get_current_compute_partition,
1317 amdgpu_gfx_set_compute_partition);
1318
1319static DEVICE_ATTR(available_compute_partition, 0444,
1320 amdgpu_gfx_get_available_compute_partition, NULL);
1321
1322int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev)
1323{
1324 int r;
1325
1326 r = device_create_file(adev->dev, &dev_attr_current_compute_partition);
1327 if (r)
1328 return r;
1329
1330 r = device_create_file(adev->dev, &dev_attr_available_compute_partition);
1331
1332 return r;
1333}
1334
1335void amdgpu_gfx_sysfs_fini(struct amdgpu_device *adev)
1336{
1337 device_remove_file(adev->dev, &dev_attr_current_compute_partition);
1338 device_remove_file(adev->dev, &dev_attr_available_compute_partition);
1339}
1/*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26#include <linux/firmware.h>
27#include <linux/pm_runtime.h>
28
29#include "amdgpu.h"
30#include "amdgpu_gfx.h"
31#include "amdgpu_rlc.h"
32#include "amdgpu_ras.h"
33#include "amdgpu_reset.h"
34#include "amdgpu_xcp.h"
35#include "amdgpu_xgmi.h"
36
37/* delay 0.1 second to enable gfx off feature */
38#define GFX_OFF_DELAY_ENABLE msecs_to_jiffies(100)
39
40#define GFX_OFF_NO_DELAY 0
41
42/*
43 * GPU GFX IP block helpers function.
44 */
45
46int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
47 int pipe, int queue)
48{
49 int bit = 0;
50
51 bit += mec * adev->gfx.mec.num_pipe_per_mec
52 * adev->gfx.mec.num_queue_per_pipe;
53 bit += pipe * adev->gfx.mec.num_queue_per_pipe;
54 bit += queue;
55
56 return bit;
57}
58
59void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
60 int *mec, int *pipe, int *queue)
61{
62 *queue = bit % adev->gfx.mec.num_queue_per_pipe;
63 *pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
64 % adev->gfx.mec.num_pipe_per_mec;
65 *mec = (bit / adev->gfx.mec.num_queue_per_pipe)
66 / adev->gfx.mec.num_pipe_per_mec;
67
68}
69
70bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
71 int xcc_id, int mec, int pipe, int queue)
72{
73 return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
74 adev->gfx.mec_bitmap[xcc_id].queue_bitmap);
75}
76
77int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
78 int me, int pipe, int queue)
79{
80 int bit = 0;
81
82 bit += me * adev->gfx.me.num_pipe_per_me
83 * adev->gfx.me.num_queue_per_pipe;
84 bit += pipe * adev->gfx.me.num_queue_per_pipe;
85 bit += queue;
86
87 return bit;
88}
89
90bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
91 int me, int pipe, int queue)
92{
93 return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
94 adev->gfx.me.queue_bitmap);
95}
96
97/**
98 * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
99 *
100 * @mask: array in which the per-shader array disable masks will be stored
101 * @max_se: number of SEs
102 * @max_sh: number of SHs
103 *
104 * The bitmask of CUs to be disabled in the shader array determined by se and
105 * sh is stored in mask[se * max_sh + sh].
106 */
107void amdgpu_gfx_parse_disable_cu(unsigned int *mask, unsigned int max_se, unsigned int max_sh)
108{
109 unsigned int se, sh, cu;
110 const char *p;
111
112 memset(mask, 0, sizeof(*mask) * max_se * max_sh);
113
114 if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
115 return;
116
117 p = amdgpu_disable_cu;
118 for (;;) {
119 char *next;
120 int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
121
122 if (ret < 3) {
123 DRM_ERROR("amdgpu: could not parse disable_cu\n");
124 return;
125 }
126
127 if (se < max_se && sh < max_sh && cu < 16) {
128 DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
129 mask[se * max_sh + sh] |= 1u << cu;
130 } else {
131 DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
132 se, sh, cu);
133 }
134
135 next = strchr(p, ',');
136 if (!next)
137 break;
138 p = next + 1;
139 }
140}
141
142static bool amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device *adev)
143{
144 return amdgpu_async_gfx_ring && adev->gfx.me.num_pipe_per_me > 1;
145}
146
147static bool amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device *adev)
148{
149 if (amdgpu_compute_multipipe != -1) {
150 DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
151 amdgpu_compute_multipipe);
152 return amdgpu_compute_multipipe == 1;
153 }
154
155 if (amdgpu_ip_version(adev, GC_HWIP, 0) > IP_VERSION(9, 0, 0))
156 return true;
157
158 /* FIXME: spreading the queues across pipes causes perf regressions
159 * on POLARIS11 compute workloads */
160 if (adev->asic_type == CHIP_POLARIS11)
161 return false;
162
163 return adev->gfx.mec.num_mec > 1;
164}
165
166bool amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device *adev,
167 struct amdgpu_ring *ring)
168{
169 int queue = ring->queue;
170 int pipe = ring->pipe;
171
172 /* Policy: use pipe1 queue0 as high priority graphics queue if we
173 * have more than one gfx pipe.
174 */
175 if (amdgpu_gfx_is_graphics_multipipe_capable(adev) &&
176 adev->gfx.num_gfx_rings > 1 && pipe == 1 && queue == 0) {
177 int me = ring->me;
178 int bit;
179
180 bit = amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue);
181 if (ring == &adev->gfx.gfx_ring[bit])
182 return true;
183 }
184
185 return false;
186}
187
188bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
189 struct amdgpu_ring *ring)
190{
191 /* Policy: use 1st queue as high priority compute queue if we
192 * have more than one compute queue.
193 */
194 if (adev->gfx.num_compute_rings > 1 &&
195 ring == &adev->gfx.compute_ring[0])
196 return true;
197
198 return false;
199}
200
201void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
202{
203 int i, j, queue, pipe;
204 bool multipipe_policy = amdgpu_gfx_is_compute_multipipe_capable(adev);
205 int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
206 adev->gfx.mec.num_queue_per_pipe,
207 adev->gfx.num_compute_rings);
208 int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
209
210 if (multipipe_policy) {
211 /* policy: make queues evenly cross all pipes on MEC1 only
212 * for multiple xcc, just use the original policy for simplicity */
213 for (j = 0; j < num_xcc; j++) {
214 for (i = 0; i < max_queues_per_mec; i++) {
215 pipe = i % adev->gfx.mec.num_pipe_per_mec;
216 queue = (i / adev->gfx.mec.num_pipe_per_mec) %
217 adev->gfx.mec.num_queue_per_pipe;
218
219 set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
220 adev->gfx.mec_bitmap[j].queue_bitmap);
221 }
222 }
223 } else {
224 /* policy: amdgpu owns all queues in the given pipe */
225 for (j = 0; j < num_xcc; j++) {
226 for (i = 0; i < max_queues_per_mec; ++i)
227 set_bit(i, adev->gfx.mec_bitmap[j].queue_bitmap);
228 }
229 }
230
231 for (j = 0; j < num_xcc; j++) {
232 dev_dbg(adev->dev, "mec queue bitmap weight=%d\n",
233 bitmap_weight(adev->gfx.mec_bitmap[j].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
234 }
235}
236
237void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
238{
239 int i, queue, pipe;
240 bool multipipe_policy = amdgpu_gfx_is_graphics_multipipe_capable(adev);
241 int max_queues_per_me = adev->gfx.me.num_pipe_per_me *
242 adev->gfx.me.num_queue_per_pipe;
243
244 if (multipipe_policy) {
245 /* policy: amdgpu owns the first queue per pipe at this stage
246 * will extend to mulitple queues per pipe later */
247 for (i = 0; i < max_queues_per_me; i++) {
248 pipe = i % adev->gfx.me.num_pipe_per_me;
249 queue = (i / adev->gfx.me.num_pipe_per_me) %
250 adev->gfx.me.num_queue_per_pipe;
251
252 set_bit(pipe * adev->gfx.me.num_queue_per_pipe + queue,
253 adev->gfx.me.queue_bitmap);
254 }
255 } else {
256 for (i = 0; i < max_queues_per_me; ++i)
257 set_bit(i, adev->gfx.me.queue_bitmap);
258 }
259
260 /* update the number of active graphics rings */
261 adev->gfx.num_gfx_rings =
262 bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
263}
264
265static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
266 struct amdgpu_ring *ring, int xcc_id)
267{
268 int queue_bit;
269 int mec, pipe, queue;
270
271 queue_bit = adev->gfx.mec.num_mec
272 * adev->gfx.mec.num_pipe_per_mec
273 * adev->gfx.mec.num_queue_per_pipe;
274
275 while (--queue_bit >= 0) {
276 if (test_bit(queue_bit, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
277 continue;
278
279 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
280
281 /*
282 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
283 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
284 * only can be issued on queue 0.
285 */
286 if ((mec == 1 && pipe > 1) || queue != 0)
287 continue;
288
289 ring->me = mec + 1;
290 ring->pipe = pipe;
291 ring->queue = queue;
292
293 return 0;
294 }
295
296 dev_err(adev->dev, "Failed to find a queue for KIQ\n");
297 return -EINVAL;
298}
299
300int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev, int xcc_id)
301{
302 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
303 struct amdgpu_irq_src *irq = &kiq->irq;
304 struct amdgpu_ring *ring = &kiq->ring;
305 int r = 0;
306
307 spin_lock_init(&kiq->ring_lock);
308
309 ring->adev = NULL;
310 ring->ring_obj = NULL;
311 ring->use_doorbell = true;
312 ring->xcc_id = xcc_id;
313 ring->vm_hub = AMDGPU_GFXHUB(xcc_id);
314 ring->doorbell_index =
315 (adev->doorbell_index.kiq +
316 xcc_id * adev->doorbell_index.xcc_doorbell_range)
317 << 1;
318
319 r = amdgpu_gfx_kiq_acquire(adev, ring, xcc_id);
320 if (r)
321 return r;
322
323 ring->eop_gpu_addr = kiq->eop_gpu_addr;
324 ring->no_scheduler = true;
325 snprintf(ring->name, sizeof(ring->name), "kiq_%hhu.%hhu.%hhu.%hhu",
326 (unsigned char)xcc_id, (unsigned char)ring->me,
327 (unsigned char)ring->pipe, (unsigned char)ring->queue);
328 r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
329 AMDGPU_RING_PRIO_DEFAULT, NULL);
330 if (r)
331 dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
332
333 return r;
334}
335
336void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
337{
338 amdgpu_ring_fini(ring);
339}
340
341void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev, int xcc_id)
342{
343 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
344
345 amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
346}
347
348int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
349 unsigned int hpd_size, int xcc_id)
350{
351 int r;
352 u32 *hpd;
353 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
354
355 r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
356 AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
357 &kiq->eop_gpu_addr, (void **)&hpd);
358 if (r) {
359 dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
360 return r;
361 }
362
363 memset(hpd, 0, hpd_size);
364
365 r = amdgpu_bo_reserve(kiq->eop_obj, true);
366 if (unlikely(r != 0))
367 dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
368 amdgpu_bo_kunmap(kiq->eop_obj);
369 amdgpu_bo_unreserve(kiq->eop_obj);
370
371 return 0;
372}
373
374/* create MQD for each compute/gfx queue */
375int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
376 unsigned int mqd_size, int xcc_id)
377{
378 int r, i, j;
379 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
380 struct amdgpu_ring *ring = &kiq->ring;
381 u32 domain = AMDGPU_GEM_DOMAIN_GTT;
382
383#if !defined(CONFIG_ARM) && !defined(CONFIG_ARM64)
384 /* Only enable on gfx10 and 11 for now to avoid changing behavior on older chips */
385 if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(10, 0, 0))
386 domain |= AMDGPU_GEM_DOMAIN_VRAM;
387#endif
388
389 /* create MQD for KIQ */
390 if (!adev->enable_mes_kiq && !ring->mqd_obj) {
391 /* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
392 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
393 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
394 * KIQ MQD no matter SRIOV or Bare-metal
395 */
396 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
397 AMDGPU_GEM_DOMAIN_VRAM |
398 AMDGPU_GEM_DOMAIN_GTT,
399 &ring->mqd_obj,
400 &ring->mqd_gpu_addr,
401 &ring->mqd_ptr);
402 if (r) {
403 dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
404 return r;
405 }
406
407 /* prepare MQD backup */
408 kiq->mqd_backup = kzalloc(mqd_size, GFP_KERNEL);
409 if (!kiq->mqd_backup) {
410 dev_warn(adev->dev,
411 "no memory to create MQD backup for ring %s\n", ring->name);
412 return -ENOMEM;
413 }
414 }
415
416 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
417 /* create MQD for each KGQ */
418 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
419 ring = &adev->gfx.gfx_ring[i];
420 if (!ring->mqd_obj) {
421 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
422 domain, &ring->mqd_obj,
423 &ring->mqd_gpu_addr, &ring->mqd_ptr);
424 if (r) {
425 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
426 return r;
427 }
428
429 ring->mqd_size = mqd_size;
430 /* prepare MQD backup */
431 adev->gfx.me.mqd_backup[i] = kzalloc(mqd_size, GFP_KERNEL);
432 if (!adev->gfx.me.mqd_backup[i]) {
433 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
434 return -ENOMEM;
435 }
436 }
437 }
438 }
439
440 /* create MQD for each KCQ */
441 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
442 j = i + xcc_id * adev->gfx.num_compute_rings;
443 ring = &adev->gfx.compute_ring[j];
444 if (!ring->mqd_obj) {
445 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
446 domain, &ring->mqd_obj,
447 &ring->mqd_gpu_addr, &ring->mqd_ptr);
448 if (r) {
449 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
450 return r;
451 }
452
453 ring->mqd_size = mqd_size;
454 /* prepare MQD backup */
455 adev->gfx.mec.mqd_backup[j] = kzalloc(mqd_size, GFP_KERNEL);
456 if (!adev->gfx.mec.mqd_backup[j]) {
457 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
458 return -ENOMEM;
459 }
460 }
461 }
462
463 return 0;
464}
465
466void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev, int xcc_id)
467{
468 struct amdgpu_ring *ring = NULL;
469 int i, j;
470 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
471
472 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
473 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
474 ring = &adev->gfx.gfx_ring[i];
475 kfree(adev->gfx.me.mqd_backup[i]);
476 amdgpu_bo_free_kernel(&ring->mqd_obj,
477 &ring->mqd_gpu_addr,
478 &ring->mqd_ptr);
479 }
480 }
481
482 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
483 j = i + xcc_id * adev->gfx.num_compute_rings;
484 ring = &adev->gfx.compute_ring[j];
485 kfree(adev->gfx.mec.mqd_backup[j]);
486 amdgpu_bo_free_kernel(&ring->mqd_obj,
487 &ring->mqd_gpu_addr,
488 &ring->mqd_ptr);
489 }
490
491 ring = &kiq->ring;
492 kfree(kiq->mqd_backup);
493 amdgpu_bo_free_kernel(&ring->mqd_obj,
494 &ring->mqd_gpu_addr,
495 &ring->mqd_ptr);
496}
497
498int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
499{
500 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
501 struct amdgpu_ring *kiq_ring = &kiq->ring;
502 int i, r = 0;
503 int j;
504
505 if (adev->enable_mes) {
506 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
507 j = i + xcc_id * adev->gfx.num_compute_rings;
508 amdgpu_mes_unmap_legacy_queue(adev,
509 &adev->gfx.compute_ring[j],
510 RESET_QUEUES, 0, 0);
511 }
512 return 0;
513 }
514
515 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
516 return -EINVAL;
517
518 if (!kiq_ring->sched.ready || adev->job_hang || amdgpu_in_reset(adev))
519 return 0;
520
521 spin_lock(&kiq->ring_lock);
522 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
523 adev->gfx.num_compute_rings)) {
524 spin_unlock(&kiq->ring_lock);
525 return -ENOMEM;
526 }
527
528 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
529 j = i + xcc_id * adev->gfx.num_compute_rings;
530 kiq->pmf->kiq_unmap_queues(kiq_ring,
531 &adev->gfx.compute_ring[j],
532 RESET_QUEUES, 0, 0);
533 }
534 /* Submit unmap queue packet */
535 amdgpu_ring_commit(kiq_ring);
536 /*
537 * Ring test will do a basic scratch register change check. Just run
538 * this to ensure that unmap queues that is submitted before got
539 * processed successfully before returning.
540 */
541 r = amdgpu_ring_test_helper(kiq_ring);
542
543 spin_unlock(&kiq->ring_lock);
544
545 return r;
546}
547
548int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id)
549{
550 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
551 struct amdgpu_ring *kiq_ring = &kiq->ring;
552 int i, r = 0;
553 int j;
554
555 if (adev->enable_mes) {
556 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
557 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
558 j = i + xcc_id * adev->gfx.num_gfx_rings;
559 amdgpu_mes_unmap_legacy_queue(adev,
560 &adev->gfx.gfx_ring[j],
561 PREEMPT_QUEUES, 0, 0);
562 }
563 }
564 return 0;
565 }
566
567 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
568 return -EINVAL;
569
570 if (!adev->gfx.kiq[0].ring.sched.ready || adev->job_hang)
571 return 0;
572
573 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
574 spin_lock(&kiq->ring_lock);
575 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
576 adev->gfx.num_gfx_rings)) {
577 spin_unlock(&kiq->ring_lock);
578 return -ENOMEM;
579 }
580
581 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
582 j = i + xcc_id * adev->gfx.num_gfx_rings;
583 kiq->pmf->kiq_unmap_queues(kiq_ring,
584 &adev->gfx.gfx_ring[j],
585 PREEMPT_QUEUES, 0, 0);
586 }
587 /* Submit unmap queue packet */
588 amdgpu_ring_commit(kiq_ring);
589
590 /*
591 * Ring test will do a basic scratch register change check.
592 * Just run this to ensure that unmap queues that is submitted
593 * before got processed successfully before returning.
594 */
595 r = amdgpu_ring_test_helper(kiq_ring);
596 spin_unlock(&kiq->ring_lock);
597 }
598
599 return r;
600}
601
602int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
603 int queue_bit)
604{
605 int mec, pipe, queue;
606 int set_resource_bit = 0;
607
608 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
609
610 set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
611
612 return set_resource_bit;
613}
614
615static int amdgpu_gfx_mes_enable_kcq(struct amdgpu_device *adev, int xcc_id)
616{
617 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
618 struct amdgpu_ring *kiq_ring = &kiq->ring;
619 uint64_t queue_mask = ~0ULL;
620 int r, i, j;
621
622 amdgpu_device_flush_hdp(adev, NULL);
623
624 if (!adev->enable_uni_mes) {
625 spin_lock(&kiq->ring_lock);
626 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->set_resources_size);
627 if (r) {
628 dev_err(adev->dev, "Failed to lock KIQ (%d).\n", r);
629 spin_unlock(&kiq->ring_lock);
630 return r;
631 }
632
633 kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
634 r = amdgpu_ring_test_helper(kiq_ring);
635 spin_unlock(&kiq->ring_lock);
636 if (r)
637 dev_err(adev->dev, "KIQ failed to set resources\n");
638 }
639
640 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
641 j = i + xcc_id * adev->gfx.num_compute_rings;
642 r = amdgpu_mes_map_legacy_queue(adev,
643 &adev->gfx.compute_ring[j]);
644 if (r) {
645 dev_err(adev->dev, "failed to map compute queue\n");
646 return r;
647 }
648 }
649
650 return 0;
651}
652
653int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
654{
655 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
656 struct amdgpu_ring *kiq_ring = &kiq->ring;
657 uint64_t queue_mask = 0;
658 int r, i, j;
659
660 if (adev->mes.enable_legacy_queue_map)
661 return amdgpu_gfx_mes_enable_kcq(adev, xcc_id);
662
663 if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
664 return -EINVAL;
665
666 for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
667 if (!test_bit(i, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
668 continue;
669
670 /* This situation may be hit in the future if a new HW
671 * generation exposes more than 64 queues. If so, the
672 * definition of queue_mask needs updating */
673 if (WARN_ON(i > (sizeof(queue_mask)*8))) {
674 DRM_ERROR("Invalid KCQ enabled: %d\n", i);
675 break;
676 }
677
678 queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
679 }
680
681 amdgpu_device_flush_hdp(adev, NULL);
682
683 DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
684 kiq_ring->queue);
685
686 spin_lock(&kiq->ring_lock);
687 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
688 adev->gfx.num_compute_rings +
689 kiq->pmf->set_resources_size);
690 if (r) {
691 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
692 spin_unlock(&kiq->ring_lock);
693 return r;
694 }
695
696 kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
697 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
698 j = i + xcc_id * adev->gfx.num_compute_rings;
699 kiq->pmf->kiq_map_queues(kiq_ring,
700 &adev->gfx.compute_ring[j]);
701 }
702 /* Submit map queue packet */
703 amdgpu_ring_commit(kiq_ring);
704 /*
705 * Ring test will do a basic scratch register change check. Just run
706 * this to ensure that map queues that is submitted before got
707 * processed successfully before returning.
708 */
709 r = amdgpu_ring_test_helper(kiq_ring);
710 spin_unlock(&kiq->ring_lock);
711 if (r)
712 DRM_ERROR("KCQ enable failed\n");
713
714 return r;
715}
716
717int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
718{
719 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
720 struct amdgpu_ring *kiq_ring = &kiq->ring;
721 int r, i, j;
722
723 if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
724 return -EINVAL;
725
726 amdgpu_device_flush_hdp(adev, NULL);
727
728 if (adev->mes.enable_legacy_queue_map) {
729 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
730 j = i + xcc_id * adev->gfx.num_gfx_rings;
731 r = amdgpu_mes_map_legacy_queue(adev,
732 &adev->gfx.gfx_ring[j]);
733 if (r) {
734 DRM_ERROR("failed to map gfx queue\n");
735 return r;
736 }
737 }
738
739 return 0;
740 }
741
742 spin_lock(&kiq->ring_lock);
743 /* No need to map kcq on the slave */
744 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
745 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
746 adev->gfx.num_gfx_rings);
747 if (r) {
748 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
749 spin_unlock(&kiq->ring_lock);
750 return r;
751 }
752
753 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
754 j = i + xcc_id * adev->gfx.num_gfx_rings;
755 kiq->pmf->kiq_map_queues(kiq_ring,
756 &adev->gfx.gfx_ring[j]);
757 }
758 }
759 /* Submit map queue packet */
760 amdgpu_ring_commit(kiq_ring);
761 /*
762 * Ring test will do a basic scratch register change check. Just run
763 * this to ensure that map queues that is submitted before got
764 * processed successfully before returning.
765 */
766 r = amdgpu_ring_test_helper(kiq_ring);
767 spin_unlock(&kiq->ring_lock);
768 if (r)
769 DRM_ERROR("KGQ enable failed\n");
770
771 return r;
772}
773
774/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
775 *
776 * @adev: amdgpu_device pointer
777 * @bool enable true: enable gfx off feature, false: disable gfx off feature
778 *
779 * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
780 * 2. other client can send request to disable gfx off feature, the request should be honored.
781 * 3. other client can cancel their request of disable gfx off feature
782 * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
783 */
784
785void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
786{
787 unsigned long delay = GFX_OFF_DELAY_ENABLE;
788
789 if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
790 return;
791
792 mutex_lock(&adev->gfx.gfx_off_mutex);
793
794 if (enable) {
795 /* If the count is already 0, it means there's an imbalance bug somewhere.
796 * Note that the bug may be in a different caller than the one which triggers the
797 * WARN_ON_ONCE.
798 */
799 if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
800 goto unlock;
801
802 adev->gfx.gfx_off_req_count--;
803
804 if (adev->gfx.gfx_off_req_count == 0 &&
805 !adev->gfx.gfx_off_state) {
806 /* If going to s2idle, no need to wait */
807 if (adev->in_s0ix) {
808 if (!amdgpu_dpm_set_powergating_by_smu(adev,
809 AMD_IP_BLOCK_TYPE_GFX, true))
810 adev->gfx.gfx_off_state = true;
811 } else {
812 schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
813 delay);
814 }
815 }
816 } else {
817 if (adev->gfx.gfx_off_req_count == 0) {
818 cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
819
820 if (adev->gfx.gfx_off_state &&
821 !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
822 adev->gfx.gfx_off_state = false;
823
824 if (adev->gfx.funcs->init_spm_golden) {
825 dev_dbg(adev->dev,
826 "GFXOFF is disabled, re-init SPM golden settings\n");
827 amdgpu_gfx_init_spm_golden(adev);
828 }
829 }
830 }
831
832 adev->gfx.gfx_off_req_count++;
833 }
834
835unlock:
836 mutex_unlock(&adev->gfx.gfx_off_mutex);
837}
838
839int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
840{
841 int r = 0;
842
843 mutex_lock(&adev->gfx.gfx_off_mutex);
844
845 r = amdgpu_dpm_set_residency_gfxoff(adev, value);
846
847 mutex_unlock(&adev->gfx.gfx_off_mutex);
848
849 return r;
850}
851
852int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
853{
854 int r = 0;
855
856 mutex_lock(&adev->gfx.gfx_off_mutex);
857
858 r = amdgpu_dpm_get_residency_gfxoff(adev, value);
859
860 mutex_unlock(&adev->gfx.gfx_off_mutex);
861
862 return r;
863}
864
865int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
866{
867 int r = 0;
868
869 mutex_lock(&adev->gfx.gfx_off_mutex);
870
871 r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
872
873 mutex_unlock(&adev->gfx.gfx_off_mutex);
874
875 return r;
876}
877
878int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
879{
880
881 int r = 0;
882
883 mutex_lock(&adev->gfx.gfx_off_mutex);
884
885 r = amdgpu_dpm_get_status_gfxoff(adev, value);
886
887 mutex_unlock(&adev->gfx.gfx_off_mutex);
888
889 return r;
890}
891
892int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
893{
894 int r;
895
896 if (amdgpu_ras_is_supported(adev, ras_block->block)) {
897 if (!amdgpu_persistent_edc_harvesting_supported(adev)) {
898 r = amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
899 if (r)
900 return r;
901 }
902
903 r = amdgpu_ras_block_late_init(adev, ras_block);
904 if (r)
905 return r;
906
907 if (amdgpu_sriov_vf(adev))
908 return r;
909
910 if (adev->gfx.cp_ecc_error_irq.funcs) {
911 r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
912 if (r)
913 goto late_fini;
914 }
915 } else {
916 amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
917 }
918
919 return 0;
920late_fini:
921 amdgpu_ras_block_late_fini(adev, ras_block);
922 return r;
923}
924
925int amdgpu_gfx_ras_sw_init(struct amdgpu_device *adev)
926{
927 int err = 0;
928 struct amdgpu_gfx_ras *ras = NULL;
929
930 /* adev->gfx.ras is NULL, which means gfx does not
931 * support ras function, then do nothing here.
932 */
933 if (!adev->gfx.ras)
934 return 0;
935
936 ras = adev->gfx.ras;
937
938 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
939 if (err) {
940 dev_err(adev->dev, "Failed to register gfx ras block!\n");
941 return err;
942 }
943
944 strcpy(ras->ras_block.ras_comm.name, "gfx");
945 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__GFX;
946 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
947 adev->gfx.ras_if = &ras->ras_block.ras_comm;
948
949 /* If not define special ras_late_init function, use gfx default ras_late_init */
950 if (!ras->ras_block.ras_late_init)
951 ras->ras_block.ras_late_init = amdgpu_gfx_ras_late_init;
952
953 /* If not defined special ras_cb function, use default ras_cb */
954 if (!ras->ras_block.ras_cb)
955 ras->ras_block.ras_cb = amdgpu_gfx_process_ras_data_cb;
956
957 return 0;
958}
959
960int amdgpu_gfx_poison_consumption_handler(struct amdgpu_device *adev,
961 struct amdgpu_iv_entry *entry)
962{
963 if (adev->gfx.ras && adev->gfx.ras->poison_consumption_handler)
964 return adev->gfx.ras->poison_consumption_handler(adev, entry);
965
966 return 0;
967}
968
969int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
970 void *err_data,
971 struct amdgpu_iv_entry *entry)
972{
973 /* TODO ue will trigger an interrupt.
974 *
975 * When “Full RAS” is enabled, the per-IP interrupt sources should
976 * be disabled and the driver should only look for the aggregated
977 * interrupt via sync flood
978 */
979 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
980 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
981 if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
982 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
983 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
984 amdgpu_ras_reset_gpu(adev);
985 }
986 return AMDGPU_RAS_SUCCESS;
987}
988
989int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
990 struct amdgpu_irq_src *source,
991 struct amdgpu_iv_entry *entry)
992{
993 struct ras_common_if *ras_if = adev->gfx.ras_if;
994 struct ras_dispatch_if ih_data = {
995 .entry = entry,
996 };
997
998 if (!ras_if)
999 return 0;
1000
1001 ih_data.head = *ras_if;
1002
1003 DRM_ERROR("CP ECC ERROR IRQ\n");
1004 amdgpu_ras_interrupt_dispatch(adev, &ih_data);
1005 return 0;
1006}
1007
1008void amdgpu_gfx_ras_error_func(struct amdgpu_device *adev,
1009 void *ras_error_status,
1010 void (*func)(struct amdgpu_device *adev, void *ras_error_status,
1011 int xcc_id))
1012{
1013 int i;
1014 int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
1015 uint32_t xcc_mask = GENMASK(num_xcc - 1, 0);
1016 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
1017
1018 if (err_data) {
1019 err_data->ue_count = 0;
1020 err_data->ce_count = 0;
1021 }
1022
1023 for_each_inst(i, xcc_mask)
1024 func(adev, ras_error_status, i);
1025}
1026
1027uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg, uint32_t xcc_id)
1028{
1029 signed long r, cnt = 0;
1030 unsigned long flags;
1031 uint32_t seq, reg_val_offs = 0, value = 0;
1032 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
1033 struct amdgpu_ring *ring = &kiq->ring;
1034
1035 if (amdgpu_device_skip_hw_access(adev))
1036 return 0;
1037
1038 if (adev->mes.ring[0].sched.ready)
1039 return amdgpu_mes_rreg(adev, reg);
1040
1041 BUG_ON(!ring->funcs->emit_rreg);
1042
1043 spin_lock_irqsave(&kiq->ring_lock, flags);
1044 if (amdgpu_device_wb_get(adev, ®_val_offs)) {
1045 pr_err("critical bug! too many kiq readers\n");
1046 goto failed_unlock;
1047 }
1048 r = amdgpu_ring_alloc(ring, 32);
1049 if (r)
1050 goto failed_unlock;
1051
1052 amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
1053 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
1054 if (r)
1055 goto failed_undo;
1056
1057 amdgpu_ring_commit(ring);
1058 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1059
1060 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1061
1062 /* don't wait anymore for gpu reset case because this way may
1063 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
1064 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
1065 * never return if we keep waiting in virt_kiq_rreg, which cause
1066 * gpu_recover() hang there.
1067 *
1068 * also don't wait anymore for IRQ context
1069 * */
1070 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
1071 goto failed_kiq_read;
1072
1073 might_sleep();
1074 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
1075 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
1076 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1077 }
1078
1079 if (cnt > MAX_KIQ_REG_TRY)
1080 goto failed_kiq_read;
1081
1082 mb();
1083 value = adev->wb.wb[reg_val_offs];
1084 amdgpu_device_wb_free(adev, reg_val_offs);
1085 return value;
1086
1087failed_undo:
1088 amdgpu_ring_undo(ring);
1089failed_unlock:
1090 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1091failed_kiq_read:
1092 if (reg_val_offs)
1093 amdgpu_device_wb_free(adev, reg_val_offs);
1094 dev_err(adev->dev, "failed to read reg:%x\n", reg);
1095 return ~0;
1096}
1097
1098void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, uint32_t xcc_id)
1099{
1100 signed long r, cnt = 0;
1101 unsigned long flags;
1102 uint32_t seq;
1103 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
1104 struct amdgpu_ring *ring = &kiq->ring;
1105
1106 BUG_ON(!ring->funcs->emit_wreg);
1107
1108 if (amdgpu_device_skip_hw_access(adev))
1109 return;
1110
1111 if (adev->mes.ring[0].sched.ready) {
1112 amdgpu_mes_wreg(adev, reg, v);
1113 return;
1114 }
1115
1116 spin_lock_irqsave(&kiq->ring_lock, flags);
1117 r = amdgpu_ring_alloc(ring, 32);
1118 if (r)
1119 goto failed_unlock;
1120
1121 amdgpu_ring_emit_wreg(ring, reg, v);
1122 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
1123 if (r)
1124 goto failed_undo;
1125
1126 amdgpu_ring_commit(ring);
1127 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1128
1129 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1130
1131 /* don't wait anymore for gpu reset case because this way may
1132 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
1133 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
1134 * never return if we keep waiting in virt_kiq_rreg, which cause
1135 * gpu_recover() hang there.
1136 *
1137 * also don't wait anymore for IRQ context
1138 * */
1139 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
1140 goto failed_kiq_write;
1141
1142 might_sleep();
1143 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
1144
1145 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
1146 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1147 }
1148
1149 if (cnt > MAX_KIQ_REG_TRY)
1150 goto failed_kiq_write;
1151
1152 return;
1153
1154failed_undo:
1155 amdgpu_ring_undo(ring);
1156failed_unlock:
1157 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1158failed_kiq_write:
1159 dev_err(adev->dev, "failed to write reg:%x\n", reg);
1160}
1161
1162int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
1163{
1164 if (amdgpu_num_kcq == -1) {
1165 return 8;
1166 } else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
1167 dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
1168 return 8;
1169 }
1170 return amdgpu_num_kcq;
1171}
1172
1173void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
1174 uint32_t ucode_id)
1175{
1176 const struct gfx_firmware_header_v1_0 *cp_hdr;
1177 const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
1178 struct amdgpu_firmware_info *info = NULL;
1179 const struct firmware *ucode_fw;
1180 unsigned int fw_size;
1181
1182 switch (ucode_id) {
1183 case AMDGPU_UCODE_ID_CP_PFP:
1184 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1185 adev->gfx.pfp_fw->data;
1186 adev->gfx.pfp_fw_version =
1187 le32_to_cpu(cp_hdr->header.ucode_version);
1188 adev->gfx.pfp_feature_version =
1189 le32_to_cpu(cp_hdr->ucode_feature_version);
1190 ucode_fw = adev->gfx.pfp_fw;
1191 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1192 break;
1193 case AMDGPU_UCODE_ID_CP_RS64_PFP:
1194 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1195 adev->gfx.pfp_fw->data;
1196 adev->gfx.pfp_fw_version =
1197 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1198 adev->gfx.pfp_feature_version =
1199 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1200 ucode_fw = adev->gfx.pfp_fw;
1201 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1202 break;
1203 case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
1204 case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
1205 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1206 adev->gfx.pfp_fw->data;
1207 ucode_fw = adev->gfx.pfp_fw;
1208 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1209 break;
1210 case AMDGPU_UCODE_ID_CP_ME:
1211 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1212 adev->gfx.me_fw->data;
1213 adev->gfx.me_fw_version =
1214 le32_to_cpu(cp_hdr->header.ucode_version);
1215 adev->gfx.me_feature_version =
1216 le32_to_cpu(cp_hdr->ucode_feature_version);
1217 ucode_fw = adev->gfx.me_fw;
1218 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1219 break;
1220 case AMDGPU_UCODE_ID_CP_RS64_ME:
1221 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1222 adev->gfx.me_fw->data;
1223 adev->gfx.me_fw_version =
1224 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1225 adev->gfx.me_feature_version =
1226 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1227 ucode_fw = adev->gfx.me_fw;
1228 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1229 break;
1230 case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
1231 case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
1232 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1233 adev->gfx.me_fw->data;
1234 ucode_fw = adev->gfx.me_fw;
1235 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1236 break;
1237 case AMDGPU_UCODE_ID_CP_CE:
1238 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1239 adev->gfx.ce_fw->data;
1240 adev->gfx.ce_fw_version =
1241 le32_to_cpu(cp_hdr->header.ucode_version);
1242 adev->gfx.ce_feature_version =
1243 le32_to_cpu(cp_hdr->ucode_feature_version);
1244 ucode_fw = adev->gfx.ce_fw;
1245 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1246 break;
1247 case AMDGPU_UCODE_ID_CP_MEC1:
1248 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1249 adev->gfx.mec_fw->data;
1250 adev->gfx.mec_fw_version =
1251 le32_to_cpu(cp_hdr->header.ucode_version);
1252 adev->gfx.mec_feature_version =
1253 le32_to_cpu(cp_hdr->ucode_feature_version);
1254 ucode_fw = adev->gfx.mec_fw;
1255 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1256 le32_to_cpu(cp_hdr->jt_size) * 4;
1257 break;
1258 case AMDGPU_UCODE_ID_CP_MEC1_JT:
1259 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1260 adev->gfx.mec_fw->data;
1261 ucode_fw = adev->gfx.mec_fw;
1262 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1263 break;
1264 case AMDGPU_UCODE_ID_CP_MEC2:
1265 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1266 adev->gfx.mec2_fw->data;
1267 adev->gfx.mec2_fw_version =
1268 le32_to_cpu(cp_hdr->header.ucode_version);
1269 adev->gfx.mec2_feature_version =
1270 le32_to_cpu(cp_hdr->ucode_feature_version);
1271 ucode_fw = adev->gfx.mec2_fw;
1272 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1273 le32_to_cpu(cp_hdr->jt_size) * 4;
1274 break;
1275 case AMDGPU_UCODE_ID_CP_MEC2_JT:
1276 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1277 adev->gfx.mec2_fw->data;
1278 ucode_fw = adev->gfx.mec2_fw;
1279 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1280 break;
1281 case AMDGPU_UCODE_ID_CP_RS64_MEC:
1282 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1283 adev->gfx.mec_fw->data;
1284 adev->gfx.mec_fw_version =
1285 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1286 adev->gfx.mec_feature_version =
1287 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1288 ucode_fw = adev->gfx.mec_fw;
1289 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1290 break;
1291 case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
1292 case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
1293 case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
1294 case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
1295 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1296 adev->gfx.mec_fw->data;
1297 ucode_fw = adev->gfx.mec_fw;
1298 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1299 break;
1300 default:
1301 dev_err(adev->dev, "Invalid ucode id %u\n", ucode_id);
1302 return;
1303 }
1304
1305 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1306 info = &adev->firmware.ucode[ucode_id];
1307 info->ucode_id = ucode_id;
1308 info->fw = ucode_fw;
1309 adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE);
1310 }
1311}
1312
1313bool amdgpu_gfx_is_master_xcc(struct amdgpu_device *adev, int xcc_id)
1314{
1315 return !(xcc_id % (adev->gfx.num_xcc_per_xcp ?
1316 adev->gfx.num_xcc_per_xcp : 1));
1317}
1318
1319static ssize_t amdgpu_gfx_get_current_compute_partition(struct device *dev,
1320 struct device_attribute *addr,
1321 char *buf)
1322{
1323 struct drm_device *ddev = dev_get_drvdata(dev);
1324 struct amdgpu_device *adev = drm_to_adev(ddev);
1325 int mode;
1326
1327 mode = amdgpu_xcp_query_partition_mode(adev->xcp_mgr,
1328 AMDGPU_XCP_FL_NONE);
1329
1330 return sysfs_emit(buf, "%s\n", amdgpu_gfx_compute_mode_desc(mode));
1331}
1332
1333static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev,
1334 struct device_attribute *addr,
1335 const char *buf, size_t count)
1336{
1337 struct drm_device *ddev = dev_get_drvdata(dev);
1338 struct amdgpu_device *adev = drm_to_adev(ddev);
1339 enum amdgpu_gfx_partition mode;
1340 int ret = 0, num_xcc;
1341
1342 num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1343 if (num_xcc % 2 != 0)
1344 return -EINVAL;
1345
1346 if (!strncasecmp("SPX", buf, strlen("SPX"))) {
1347 mode = AMDGPU_SPX_PARTITION_MODE;
1348 } else if (!strncasecmp("DPX", buf, strlen("DPX"))) {
1349 /*
1350 * DPX mode needs AIDs to be in multiple of 2.
1351 * Each AID connects 2 XCCs.
1352 */
1353 if (num_xcc%4)
1354 return -EINVAL;
1355 mode = AMDGPU_DPX_PARTITION_MODE;
1356 } else if (!strncasecmp("TPX", buf, strlen("TPX"))) {
1357 if (num_xcc != 6)
1358 return -EINVAL;
1359 mode = AMDGPU_TPX_PARTITION_MODE;
1360 } else if (!strncasecmp("QPX", buf, strlen("QPX"))) {
1361 if (num_xcc != 8)
1362 return -EINVAL;
1363 mode = AMDGPU_QPX_PARTITION_MODE;
1364 } else if (!strncasecmp("CPX", buf, strlen("CPX"))) {
1365 mode = AMDGPU_CPX_PARTITION_MODE;
1366 } else {
1367 return -EINVAL;
1368 }
1369
1370 ret = amdgpu_xcp_switch_partition_mode(adev->xcp_mgr, mode);
1371
1372 if (ret)
1373 return ret;
1374
1375 return count;
1376}
1377
1378static const char *xcp_desc[] = {
1379 [AMDGPU_SPX_PARTITION_MODE] = "SPX",
1380 [AMDGPU_DPX_PARTITION_MODE] = "DPX",
1381 [AMDGPU_TPX_PARTITION_MODE] = "TPX",
1382 [AMDGPU_QPX_PARTITION_MODE] = "QPX",
1383 [AMDGPU_CPX_PARTITION_MODE] = "CPX",
1384};
1385
1386static ssize_t amdgpu_gfx_get_available_compute_partition(struct device *dev,
1387 struct device_attribute *addr,
1388 char *buf)
1389{
1390 struct drm_device *ddev = dev_get_drvdata(dev);
1391 struct amdgpu_device *adev = drm_to_adev(ddev);
1392 struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
1393 int size = 0, mode;
1394 char *sep = "";
1395
1396 if (!xcp_mgr || !xcp_mgr->avail_xcp_modes)
1397 return sysfs_emit(buf, "Not supported\n");
1398
1399 for_each_inst(mode, xcp_mgr->avail_xcp_modes) {
1400 size += sysfs_emit_at(buf, size, "%s%s", sep, xcp_desc[mode]);
1401 sep = ", ";
1402 }
1403
1404 size += sysfs_emit_at(buf, size, "\n");
1405
1406 return size;
1407}
1408
1409static int amdgpu_gfx_run_cleaner_shader_job(struct amdgpu_ring *ring)
1410{
1411 struct amdgpu_device *adev = ring->adev;
1412 struct drm_gpu_scheduler *sched = &ring->sched;
1413 struct drm_sched_entity entity;
1414 struct dma_fence *f;
1415 struct amdgpu_job *job;
1416 struct amdgpu_ib *ib;
1417 int i, r;
1418
1419 /* Initialize the scheduler entity */
1420 r = drm_sched_entity_init(&entity, DRM_SCHED_PRIORITY_NORMAL,
1421 &sched, 1, NULL);
1422 if (r) {
1423 dev_err(adev->dev, "Failed setting up GFX kernel entity.\n");
1424 goto err;
1425 }
1426
1427 r = amdgpu_job_alloc_with_ib(ring->adev, &entity, NULL,
1428 64, 0,
1429 &job);
1430 if (r)
1431 goto err;
1432
1433 job->enforce_isolation = true;
1434
1435 ib = &job->ibs[0];
1436 for (i = 0; i <= ring->funcs->align_mask; ++i)
1437 ib->ptr[i] = ring->funcs->nop;
1438 ib->length_dw = ring->funcs->align_mask + 1;
1439
1440 f = amdgpu_job_submit(job);
1441
1442 r = dma_fence_wait(f, false);
1443 if (r)
1444 goto err;
1445
1446 dma_fence_put(f);
1447
1448 /* Clean up the scheduler entity */
1449 drm_sched_entity_destroy(&entity);
1450 return 0;
1451
1452err:
1453 return r;
1454}
1455
1456static int amdgpu_gfx_run_cleaner_shader(struct amdgpu_device *adev, int xcp_id)
1457{
1458 int num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1459 struct amdgpu_ring *ring;
1460 int num_xcc_to_clear;
1461 int i, r, xcc_id;
1462
1463 if (adev->gfx.num_xcc_per_xcp)
1464 num_xcc_to_clear = adev->gfx.num_xcc_per_xcp;
1465 else
1466 num_xcc_to_clear = 1;
1467
1468 for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
1469 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
1470 ring = &adev->gfx.compute_ring[i + xcc_id * adev->gfx.num_compute_rings];
1471 if ((ring->xcp_id == xcp_id) && ring->sched.ready) {
1472 r = amdgpu_gfx_run_cleaner_shader_job(ring);
1473 if (r)
1474 return r;
1475 num_xcc_to_clear--;
1476 break;
1477 }
1478 }
1479 }
1480
1481 if (num_xcc_to_clear)
1482 return -ENOENT;
1483
1484 return 0;
1485}
1486
1487static ssize_t amdgpu_gfx_set_run_cleaner_shader(struct device *dev,
1488 struct device_attribute *attr,
1489 const char *buf,
1490 size_t count)
1491{
1492 struct drm_device *ddev = dev_get_drvdata(dev);
1493 struct amdgpu_device *adev = drm_to_adev(ddev);
1494 int ret;
1495 long value;
1496
1497 if (amdgpu_in_reset(adev))
1498 return -EPERM;
1499 if (adev->in_suspend && !adev->in_runpm)
1500 return -EPERM;
1501
1502 ret = kstrtol(buf, 0, &value);
1503
1504 if (ret)
1505 return -EINVAL;
1506
1507 if (value < 0)
1508 return -EINVAL;
1509
1510 if (adev->xcp_mgr) {
1511 if (value >= adev->xcp_mgr->num_xcps)
1512 return -EINVAL;
1513 } else {
1514 if (value > 1)
1515 return -EINVAL;
1516 }
1517
1518 ret = pm_runtime_get_sync(ddev->dev);
1519 if (ret < 0) {
1520 pm_runtime_put_autosuspend(ddev->dev);
1521 return ret;
1522 }
1523
1524 ret = amdgpu_gfx_run_cleaner_shader(adev, value);
1525
1526 pm_runtime_mark_last_busy(ddev->dev);
1527 pm_runtime_put_autosuspend(ddev->dev);
1528
1529 if (ret)
1530 return ret;
1531
1532 return count;
1533}
1534
1535static ssize_t amdgpu_gfx_get_enforce_isolation(struct device *dev,
1536 struct device_attribute *attr,
1537 char *buf)
1538{
1539 struct drm_device *ddev = dev_get_drvdata(dev);
1540 struct amdgpu_device *adev = drm_to_adev(ddev);
1541 int i;
1542 ssize_t size = 0;
1543
1544 if (adev->xcp_mgr) {
1545 for (i = 0; i < adev->xcp_mgr->num_xcps; i++) {
1546 size += sysfs_emit_at(buf, size, "%u", adev->enforce_isolation[i]);
1547 if (i < (adev->xcp_mgr->num_xcps - 1))
1548 size += sysfs_emit_at(buf, size, " ");
1549 }
1550 buf[size++] = '\n';
1551 } else {
1552 size = sysfs_emit_at(buf, 0, "%u\n", adev->enforce_isolation[0]);
1553 }
1554
1555 return size;
1556}
1557
1558static ssize_t amdgpu_gfx_set_enforce_isolation(struct device *dev,
1559 struct device_attribute *attr,
1560 const char *buf, size_t count)
1561{
1562 struct drm_device *ddev = dev_get_drvdata(dev);
1563 struct amdgpu_device *adev = drm_to_adev(ddev);
1564 long partition_values[MAX_XCP] = {0};
1565 int ret, i, num_partitions;
1566 const char *input_buf = buf;
1567
1568 for (i = 0; i < (adev->xcp_mgr ? adev->xcp_mgr->num_xcps : 1); i++) {
1569 ret = sscanf(input_buf, "%ld", &partition_values[i]);
1570 if (ret <= 0)
1571 break;
1572
1573 /* Move the pointer to the next value in the string */
1574 input_buf = strchr(input_buf, ' ');
1575 if (input_buf) {
1576 input_buf++;
1577 } else {
1578 i++;
1579 break;
1580 }
1581 }
1582 num_partitions = i;
1583
1584 if (adev->xcp_mgr && num_partitions != adev->xcp_mgr->num_xcps)
1585 return -EINVAL;
1586
1587 if (!adev->xcp_mgr && num_partitions != 1)
1588 return -EINVAL;
1589
1590 for (i = 0; i < num_partitions; i++) {
1591 if (partition_values[i] != 0 && partition_values[i] != 1)
1592 return -EINVAL;
1593 }
1594
1595 mutex_lock(&adev->enforce_isolation_mutex);
1596 for (i = 0; i < num_partitions; i++) {
1597 if (adev->enforce_isolation[i] && !partition_values[i])
1598 /* Going from enabled to disabled */
1599 amdgpu_vmid_free_reserved(adev, AMDGPU_GFXHUB(i));
1600 else if (!adev->enforce_isolation[i] && partition_values[i])
1601 /* Going from disabled to enabled */
1602 amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(i));
1603 adev->enforce_isolation[i] = partition_values[i];
1604 }
1605 mutex_unlock(&adev->enforce_isolation_mutex);
1606
1607 amdgpu_mes_update_enforce_isolation(adev);
1608
1609 return count;
1610}
1611
1612static ssize_t amdgpu_gfx_get_gfx_reset_mask(struct device *dev,
1613 struct device_attribute *attr,
1614 char *buf)
1615{
1616 struct drm_device *ddev = dev_get_drvdata(dev);
1617 struct amdgpu_device *adev = drm_to_adev(ddev);
1618
1619 if (!adev)
1620 return -ENODEV;
1621
1622 return amdgpu_show_reset_mask(buf, adev->gfx.gfx_supported_reset);
1623}
1624
1625static ssize_t amdgpu_gfx_get_compute_reset_mask(struct device *dev,
1626 struct device_attribute *attr,
1627 char *buf)
1628{
1629 struct drm_device *ddev = dev_get_drvdata(dev);
1630 struct amdgpu_device *adev = drm_to_adev(ddev);
1631
1632 if (!adev)
1633 return -ENODEV;
1634
1635 return amdgpu_show_reset_mask(buf, adev->gfx.compute_supported_reset);
1636}
1637
1638static DEVICE_ATTR(run_cleaner_shader, 0200,
1639 NULL, amdgpu_gfx_set_run_cleaner_shader);
1640
1641static DEVICE_ATTR(enforce_isolation, 0644,
1642 amdgpu_gfx_get_enforce_isolation,
1643 amdgpu_gfx_set_enforce_isolation);
1644
1645static DEVICE_ATTR(current_compute_partition, 0644,
1646 amdgpu_gfx_get_current_compute_partition,
1647 amdgpu_gfx_set_compute_partition);
1648
1649static DEVICE_ATTR(available_compute_partition, 0444,
1650 amdgpu_gfx_get_available_compute_partition, NULL);
1651static DEVICE_ATTR(gfx_reset_mask, 0444,
1652 amdgpu_gfx_get_gfx_reset_mask, NULL);
1653
1654static DEVICE_ATTR(compute_reset_mask, 0444,
1655 amdgpu_gfx_get_compute_reset_mask, NULL);
1656
1657static int amdgpu_gfx_sysfs_xcp_init(struct amdgpu_device *adev)
1658{
1659 struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
1660 bool xcp_switch_supported;
1661 int r;
1662
1663 if (!xcp_mgr)
1664 return 0;
1665
1666 xcp_switch_supported =
1667 (xcp_mgr->funcs && xcp_mgr->funcs->switch_partition_mode);
1668
1669 if (!xcp_switch_supported)
1670 dev_attr_current_compute_partition.attr.mode &=
1671 ~(S_IWUSR | S_IWGRP | S_IWOTH);
1672
1673 r = device_create_file(adev->dev, &dev_attr_current_compute_partition);
1674 if (r)
1675 return r;
1676
1677 if (xcp_switch_supported)
1678 r = device_create_file(adev->dev,
1679 &dev_attr_available_compute_partition);
1680
1681 return r;
1682}
1683
1684static void amdgpu_gfx_sysfs_xcp_fini(struct amdgpu_device *adev)
1685{
1686 struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
1687 bool xcp_switch_supported;
1688
1689 if (!xcp_mgr)
1690 return;
1691
1692 xcp_switch_supported =
1693 (xcp_mgr->funcs && xcp_mgr->funcs->switch_partition_mode);
1694 device_remove_file(adev->dev, &dev_attr_current_compute_partition);
1695
1696 if (xcp_switch_supported)
1697 device_remove_file(adev->dev,
1698 &dev_attr_available_compute_partition);
1699}
1700
1701static int amdgpu_gfx_sysfs_isolation_shader_init(struct amdgpu_device *adev)
1702{
1703 int r;
1704
1705 r = device_create_file(adev->dev, &dev_attr_enforce_isolation);
1706 if (r)
1707 return r;
1708 if (adev->gfx.enable_cleaner_shader)
1709 r = device_create_file(adev->dev, &dev_attr_run_cleaner_shader);
1710
1711 return r;
1712}
1713
1714static void amdgpu_gfx_sysfs_isolation_shader_fini(struct amdgpu_device *adev)
1715{
1716 device_remove_file(adev->dev, &dev_attr_enforce_isolation);
1717 if (adev->gfx.enable_cleaner_shader)
1718 device_remove_file(adev->dev, &dev_attr_run_cleaner_shader);
1719}
1720
1721static int amdgpu_gfx_sysfs_reset_mask_init(struct amdgpu_device *adev)
1722{
1723 int r = 0;
1724
1725 if (!amdgpu_gpu_recovery)
1726 return r;
1727
1728 if (adev->gfx.num_gfx_rings) {
1729 r = device_create_file(adev->dev, &dev_attr_gfx_reset_mask);
1730 if (r)
1731 return r;
1732 }
1733
1734 if (adev->gfx.num_compute_rings) {
1735 r = device_create_file(adev->dev, &dev_attr_compute_reset_mask);
1736 if (r)
1737 return r;
1738 }
1739
1740 return r;
1741}
1742
1743static void amdgpu_gfx_sysfs_reset_mask_fini(struct amdgpu_device *adev)
1744{
1745 if (!amdgpu_gpu_recovery)
1746 return;
1747
1748 if (adev->gfx.num_gfx_rings)
1749 device_remove_file(adev->dev, &dev_attr_gfx_reset_mask);
1750
1751 if (adev->gfx.num_compute_rings)
1752 device_remove_file(adev->dev, &dev_attr_compute_reset_mask);
1753}
1754
1755int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev)
1756{
1757 int r;
1758
1759 r = amdgpu_gfx_sysfs_xcp_init(adev);
1760 if (r) {
1761 dev_err(adev->dev, "failed to create xcp sysfs files");
1762 return r;
1763 }
1764
1765 r = amdgpu_gfx_sysfs_isolation_shader_init(adev);
1766 if (r)
1767 dev_err(adev->dev, "failed to create isolation sysfs files");
1768
1769 r = amdgpu_gfx_sysfs_reset_mask_init(adev);
1770 if (r)
1771 dev_err(adev->dev, "failed to create reset mask sysfs files");
1772
1773 return r;
1774}
1775
1776void amdgpu_gfx_sysfs_fini(struct amdgpu_device *adev)
1777{
1778 if (adev->dev->kobj.sd) {
1779 amdgpu_gfx_sysfs_xcp_fini(adev);
1780 amdgpu_gfx_sysfs_isolation_shader_fini(adev);
1781 amdgpu_gfx_sysfs_reset_mask_fini(adev);
1782 }
1783}
1784
1785int amdgpu_gfx_cleaner_shader_sw_init(struct amdgpu_device *adev,
1786 unsigned int cleaner_shader_size)
1787{
1788 if (!adev->gfx.enable_cleaner_shader)
1789 return -EOPNOTSUPP;
1790
1791 return amdgpu_bo_create_kernel(adev, cleaner_shader_size, PAGE_SIZE,
1792 AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT,
1793 &adev->gfx.cleaner_shader_obj,
1794 &adev->gfx.cleaner_shader_gpu_addr,
1795 (void **)&adev->gfx.cleaner_shader_cpu_ptr);
1796}
1797
1798void amdgpu_gfx_cleaner_shader_sw_fini(struct amdgpu_device *adev)
1799{
1800 if (!adev->gfx.enable_cleaner_shader)
1801 return;
1802
1803 amdgpu_bo_free_kernel(&adev->gfx.cleaner_shader_obj,
1804 &adev->gfx.cleaner_shader_gpu_addr,
1805 (void **)&adev->gfx.cleaner_shader_cpu_ptr);
1806}
1807
1808void amdgpu_gfx_cleaner_shader_init(struct amdgpu_device *adev,
1809 unsigned int cleaner_shader_size,
1810 const void *cleaner_shader_ptr)
1811{
1812 if (!adev->gfx.enable_cleaner_shader)
1813 return;
1814
1815 if (adev->gfx.cleaner_shader_cpu_ptr && cleaner_shader_ptr)
1816 memcpy_toio(adev->gfx.cleaner_shader_cpu_ptr, cleaner_shader_ptr,
1817 cleaner_shader_size);
1818}
1819
1820/**
1821 * amdgpu_gfx_kfd_sch_ctrl - Control the KFD scheduler from the KGD (Graphics Driver)
1822 * @adev: amdgpu_device pointer
1823 * @idx: Index of the scheduler to control
1824 * @enable: Whether to enable or disable the KFD scheduler
1825 *
1826 * This function is used to control the KFD (Kernel Fusion Driver) scheduler
1827 * from the KGD. It is part of the cleaner shader feature. This function plays
1828 * a key role in enforcing process isolation on the GPU.
1829 *
1830 * The function uses a reference count mechanism (kfd_sch_req_count) to keep
1831 * track of the number of requests to enable the KFD scheduler. When a request
1832 * to enable the KFD scheduler is made, the reference count is decremented.
1833 * When the reference count reaches zero, a delayed work is scheduled to
1834 * enforce isolation after a delay of GFX_SLICE_PERIOD.
1835 *
1836 * When a request to disable the KFD scheduler is made, the function first
1837 * checks if the reference count is zero. If it is, it cancels the delayed work
1838 * for enforcing isolation and checks if the KFD scheduler is active. If the
1839 * KFD scheduler is active, it sends a request to stop the KFD scheduler and
1840 * sets the KFD scheduler state to inactive. Then, it increments the reference
1841 * count.
1842 *
1843 * The function is synchronized using the kfd_sch_mutex to ensure that the KFD
1844 * scheduler state and reference count are updated atomically.
1845 *
1846 * Note: If the reference count is already zero when a request to enable the
1847 * KFD scheduler is made, it means there's an imbalance bug somewhere. The
1848 * function triggers a warning in this case.
1849 */
1850static void amdgpu_gfx_kfd_sch_ctrl(struct amdgpu_device *adev, u32 idx,
1851 bool enable)
1852{
1853 mutex_lock(&adev->gfx.kfd_sch_mutex);
1854
1855 if (enable) {
1856 /* If the count is already 0, it means there's an imbalance bug somewhere.
1857 * Note that the bug may be in a different caller than the one which triggers the
1858 * WARN_ON_ONCE.
1859 */
1860 if (WARN_ON_ONCE(adev->gfx.kfd_sch_req_count[idx] == 0)) {
1861 dev_err(adev->dev, "Attempted to enable KFD scheduler when reference count is already zero\n");
1862 goto unlock;
1863 }
1864
1865 adev->gfx.kfd_sch_req_count[idx]--;
1866
1867 if (adev->gfx.kfd_sch_req_count[idx] == 0 &&
1868 adev->gfx.kfd_sch_inactive[idx]) {
1869 schedule_delayed_work(&adev->gfx.enforce_isolation[idx].work,
1870 msecs_to_jiffies(adev->gfx.enforce_isolation_time[idx]));
1871 }
1872 } else {
1873 if (adev->gfx.kfd_sch_req_count[idx] == 0) {
1874 cancel_delayed_work_sync(&adev->gfx.enforce_isolation[idx].work);
1875 if (!adev->gfx.kfd_sch_inactive[idx]) {
1876 amdgpu_amdkfd_stop_sched(adev, idx);
1877 adev->gfx.kfd_sch_inactive[idx] = true;
1878 }
1879 }
1880
1881 adev->gfx.kfd_sch_req_count[idx]++;
1882 }
1883
1884unlock:
1885 mutex_unlock(&adev->gfx.kfd_sch_mutex);
1886}
1887
1888/**
1889 * amdgpu_gfx_enforce_isolation_handler - work handler for enforcing shader isolation
1890 *
1891 * @work: work_struct.
1892 *
1893 * This function is the work handler for enforcing shader isolation on AMD GPUs.
1894 * It counts the number of emitted fences for each GFX and compute ring. If there
1895 * are any fences, it schedules the `enforce_isolation_work` to be run after a
1896 * delay of `GFX_SLICE_PERIOD`. If there are no fences, it signals the Kernel Fusion
1897 * Driver (KFD) to resume the runqueue. The function is synchronized using the
1898 * `enforce_isolation_mutex`.
1899 */
1900void amdgpu_gfx_enforce_isolation_handler(struct work_struct *work)
1901{
1902 struct amdgpu_isolation_work *isolation_work =
1903 container_of(work, struct amdgpu_isolation_work, work.work);
1904 struct amdgpu_device *adev = isolation_work->adev;
1905 u32 i, idx, fences = 0;
1906
1907 if (isolation_work->xcp_id == AMDGPU_XCP_NO_PARTITION)
1908 idx = 0;
1909 else
1910 idx = isolation_work->xcp_id;
1911
1912 if (idx >= MAX_XCP)
1913 return;
1914
1915 mutex_lock(&adev->enforce_isolation_mutex);
1916 for (i = 0; i < AMDGPU_MAX_GFX_RINGS; ++i) {
1917 if (isolation_work->xcp_id == adev->gfx.gfx_ring[i].xcp_id)
1918 fences += amdgpu_fence_count_emitted(&adev->gfx.gfx_ring[i]);
1919 }
1920 for (i = 0; i < (AMDGPU_MAX_COMPUTE_RINGS * AMDGPU_MAX_GC_INSTANCES); ++i) {
1921 if (isolation_work->xcp_id == adev->gfx.compute_ring[i].xcp_id)
1922 fences += amdgpu_fence_count_emitted(&adev->gfx.compute_ring[i]);
1923 }
1924 if (fences) {
1925 /* we've already had our timeslice, so let's wrap this up */
1926 schedule_delayed_work(&adev->gfx.enforce_isolation[idx].work,
1927 msecs_to_jiffies(1));
1928 } else {
1929 /* Tell KFD to resume the runqueue */
1930 if (adev->kfd.init_complete) {
1931 WARN_ON_ONCE(!adev->gfx.kfd_sch_inactive[idx]);
1932 WARN_ON_ONCE(adev->gfx.kfd_sch_req_count[idx]);
1933 amdgpu_amdkfd_start_sched(adev, idx);
1934 adev->gfx.kfd_sch_inactive[idx] = false;
1935 }
1936 }
1937 mutex_unlock(&adev->enforce_isolation_mutex);
1938}
1939
1940static void
1941amdgpu_gfx_enforce_isolation_wait_for_kfd(struct amdgpu_device *adev,
1942 u32 idx)
1943{
1944 unsigned long cjiffies;
1945 bool wait = false;
1946
1947 mutex_lock(&adev->enforce_isolation_mutex);
1948 if (adev->enforce_isolation[idx]) {
1949 /* set the initial values if nothing is set */
1950 if (!adev->gfx.enforce_isolation_jiffies[idx]) {
1951 adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
1952 adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
1953 }
1954 /* Make sure KFD gets a chance to run */
1955 if (amdgpu_amdkfd_compute_active(adev, idx)) {
1956 cjiffies = jiffies;
1957 if (time_after(cjiffies, adev->gfx.enforce_isolation_jiffies[idx])) {
1958 cjiffies -= adev->gfx.enforce_isolation_jiffies[idx];
1959 if ((jiffies_to_msecs(cjiffies) >= GFX_SLICE_PERIOD_MS)) {
1960 /* if our time is up, let KGD work drain before scheduling more */
1961 wait = true;
1962 /* reset the timer period */
1963 adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
1964 } else {
1965 /* set the timer period to what's left in our time slice */
1966 adev->gfx.enforce_isolation_time[idx] =
1967 GFX_SLICE_PERIOD_MS - jiffies_to_msecs(cjiffies);
1968 }
1969 } else {
1970 /* if jiffies wrap around we will just wait a little longer */
1971 adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
1972 }
1973 } else {
1974 /* if there is no KFD work, then set the full slice period */
1975 adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
1976 adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
1977 }
1978 }
1979 mutex_unlock(&adev->enforce_isolation_mutex);
1980
1981 if (wait)
1982 msleep(GFX_SLICE_PERIOD_MS);
1983}
1984
1985void amdgpu_gfx_enforce_isolation_ring_begin_use(struct amdgpu_ring *ring)
1986{
1987 struct amdgpu_device *adev = ring->adev;
1988 u32 idx;
1989 bool sched_work = false;
1990
1991 if (!adev->gfx.enable_cleaner_shader)
1992 return;
1993
1994 if (ring->xcp_id == AMDGPU_XCP_NO_PARTITION)
1995 idx = 0;
1996 else
1997 idx = ring->xcp_id;
1998
1999 if (idx >= MAX_XCP)
2000 return;
2001
2002 /* Don't submit more work until KFD has had some time */
2003 amdgpu_gfx_enforce_isolation_wait_for_kfd(adev, idx);
2004
2005 mutex_lock(&adev->enforce_isolation_mutex);
2006 if (adev->enforce_isolation[idx]) {
2007 if (adev->kfd.init_complete)
2008 sched_work = true;
2009 }
2010 mutex_unlock(&adev->enforce_isolation_mutex);
2011
2012 if (sched_work)
2013 amdgpu_gfx_kfd_sch_ctrl(adev, idx, false);
2014}
2015
2016void amdgpu_gfx_enforce_isolation_ring_end_use(struct amdgpu_ring *ring)
2017{
2018 struct amdgpu_device *adev = ring->adev;
2019 u32 idx;
2020 bool sched_work = false;
2021
2022 if (!adev->gfx.enable_cleaner_shader)
2023 return;
2024
2025 if (ring->xcp_id == AMDGPU_XCP_NO_PARTITION)
2026 idx = 0;
2027 else
2028 idx = ring->xcp_id;
2029
2030 if (idx >= MAX_XCP)
2031 return;
2032
2033 mutex_lock(&adev->enforce_isolation_mutex);
2034 if (adev->enforce_isolation[idx]) {
2035 if (adev->kfd.init_complete)
2036 sched_work = true;
2037 }
2038 mutex_unlock(&adev->enforce_isolation_mutex);
2039
2040 if (sched_work)
2041 amdgpu_gfx_kfd_sch_ctrl(adev, idx, true);
2042}
2043
2044/*
2045 * debugfs for to enable/disable gfx job submission to specific core.
2046 */
2047#if defined(CONFIG_DEBUG_FS)
2048static int amdgpu_debugfs_gfx_sched_mask_set(void *data, u64 val)
2049{
2050 struct amdgpu_device *adev = (struct amdgpu_device *)data;
2051 u32 i;
2052 u64 mask = 0;
2053 struct amdgpu_ring *ring;
2054
2055 if (!adev)
2056 return -ENODEV;
2057
2058 mask = (1ULL << adev->gfx.num_gfx_rings) - 1;
2059 if ((val & mask) == 0)
2060 return -EINVAL;
2061
2062 for (i = 0; i < adev->gfx.num_gfx_rings; ++i) {
2063 ring = &adev->gfx.gfx_ring[i];
2064 if (val & (1 << i))
2065 ring->sched.ready = true;
2066 else
2067 ring->sched.ready = false;
2068 }
2069 /* publish sched.ready flag update effective immediately across smp */
2070 smp_rmb();
2071 return 0;
2072}
2073
2074static int amdgpu_debugfs_gfx_sched_mask_get(void *data, u64 *val)
2075{
2076 struct amdgpu_device *adev = (struct amdgpu_device *)data;
2077 u32 i;
2078 u64 mask = 0;
2079 struct amdgpu_ring *ring;
2080
2081 if (!adev)
2082 return -ENODEV;
2083 for (i = 0; i < adev->gfx.num_gfx_rings; ++i) {
2084 ring = &adev->gfx.gfx_ring[i];
2085 if (ring->sched.ready)
2086 mask |= 1ULL << i;
2087 }
2088
2089 *val = mask;
2090 return 0;
2091}
2092
2093DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gfx_sched_mask_fops,
2094 amdgpu_debugfs_gfx_sched_mask_get,
2095 amdgpu_debugfs_gfx_sched_mask_set, "%llx\n");
2096
2097#endif
2098
2099void amdgpu_debugfs_gfx_sched_mask_init(struct amdgpu_device *adev)
2100{
2101#if defined(CONFIG_DEBUG_FS)
2102 struct drm_minor *minor = adev_to_drm(adev)->primary;
2103 struct dentry *root = minor->debugfs_root;
2104 char name[32];
2105
2106 if (!(adev->gfx.num_gfx_rings > 1))
2107 return;
2108 sprintf(name, "amdgpu_gfx_sched_mask");
2109 debugfs_create_file(name, 0600, root, adev,
2110 &amdgpu_debugfs_gfx_sched_mask_fops);
2111#endif
2112}
2113
2114/*
2115 * debugfs for to enable/disable compute job submission to specific core.
2116 */
2117#if defined(CONFIG_DEBUG_FS)
2118static int amdgpu_debugfs_compute_sched_mask_set(void *data, u64 val)
2119{
2120 struct amdgpu_device *adev = (struct amdgpu_device *)data;
2121 u32 i;
2122 u64 mask = 0;
2123 struct amdgpu_ring *ring;
2124
2125 if (!adev)
2126 return -ENODEV;
2127
2128 mask = (1ULL << adev->gfx.num_compute_rings) - 1;
2129 if ((val & mask) == 0)
2130 return -EINVAL;
2131
2132 for (i = 0; i < adev->gfx.num_compute_rings; ++i) {
2133 ring = &adev->gfx.compute_ring[i];
2134 if (val & (1 << i))
2135 ring->sched.ready = true;
2136 else
2137 ring->sched.ready = false;
2138 }
2139
2140 /* publish sched.ready flag update effective immediately across smp */
2141 smp_rmb();
2142 return 0;
2143}
2144
2145static int amdgpu_debugfs_compute_sched_mask_get(void *data, u64 *val)
2146{
2147 struct amdgpu_device *adev = (struct amdgpu_device *)data;
2148 u32 i;
2149 u64 mask = 0;
2150 struct amdgpu_ring *ring;
2151
2152 if (!adev)
2153 return -ENODEV;
2154 for (i = 0; i < adev->gfx.num_compute_rings; ++i) {
2155 ring = &adev->gfx.compute_ring[i];
2156 if (ring->sched.ready)
2157 mask |= 1ULL << i;
2158 }
2159
2160 *val = mask;
2161 return 0;
2162}
2163
2164DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_compute_sched_mask_fops,
2165 amdgpu_debugfs_compute_sched_mask_get,
2166 amdgpu_debugfs_compute_sched_mask_set, "%llx\n");
2167
2168#endif
2169
2170void amdgpu_debugfs_compute_sched_mask_init(struct amdgpu_device *adev)
2171{
2172#if defined(CONFIG_DEBUG_FS)
2173 struct drm_minor *minor = adev_to_drm(adev)->primary;
2174 struct dentry *root = minor->debugfs_root;
2175 char name[32];
2176
2177 if (!(adev->gfx.num_compute_rings > 1))
2178 return;
2179 sprintf(name, "amdgpu_compute_sched_mask");
2180 debugfs_create_file(name, 0600, root, adev,
2181 &amdgpu_debugfs_compute_sched_mask_fops);
2182#endif
2183}