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, int xcc_id)
308{
309 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
310 struct amdgpu_irq_src *irq = &kiq->irq;
311 struct amdgpu_ring *ring = &kiq->ring;
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 snprintf(ring->name, sizeof(ring->name), "kiq_%d.%d.%d.%d",
333 xcc_id, ring->me, ring->pipe, ring->queue);
334 r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
335 AMDGPU_RING_PRIO_DEFAULT, NULL);
336 if (r)
337 dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
338
339 return r;
340}
341
342void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
343{
344 amdgpu_ring_fini(ring);
345}
346
347void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev, int xcc_id)
348{
349 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
350
351 amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
352}
353
354int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
355 unsigned int hpd_size, int xcc_id)
356{
357 int r;
358 u32 *hpd;
359 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
360
361 r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
362 AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
363 &kiq->eop_gpu_addr, (void **)&hpd);
364 if (r) {
365 dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
366 return r;
367 }
368
369 memset(hpd, 0, hpd_size);
370
371 r = amdgpu_bo_reserve(kiq->eop_obj, true);
372 if (unlikely(r != 0))
373 dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
374 amdgpu_bo_kunmap(kiq->eop_obj);
375 amdgpu_bo_unreserve(kiq->eop_obj);
376
377 return 0;
378}
379
380/* create MQD for each compute/gfx queue */
381int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
382 unsigned int mqd_size, int xcc_id)
383{
384 int r, i, j;
385 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
386 struct amdgpu_ring *ring = &kiq->ring;
387 u32 domain = AMDGPU_GEM_DOMAIN_GTT;
388
389#if !defined(CONFIG_ARM) && !defined(CONFIG_ARM64)
390 /* Only enable on gfx10 and 11 for now to avoid changing behavior on older chips */
391 if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(10, 0, 0))
392 domain |= AMDGPU_GEM_DOMAIN_VRAM;
393#endif
394
395 /* create MQD for KIQ */
396 if (!adev->enable_mes_kiq && !ring->mqd_obj) {
397 /* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
398 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
399 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
400 * KIQ MQD no matter SRIOV or Bare-metal
401 */
402 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
403 AMDGPU_GEM_DOMAIN_VRAM |
404 AMDGPU_GEM_DOMAIN_GTT,
405 &ring->mqd_obj,
406 &ring->mqd_gpu_addr,
407 &ring->mqd_ptr);
408 if (r) {
409 dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
410 return r;
411 }
412
413 /* prepare MQD backup */
414 kiq->mqd_backup = kmalloc(mqd_size, GFP_KERNEL);
415 if (!kiq->mqd_backup) {
416 dev_warn(adev->dev,
417 "no memory to create MQD backup for ring %s\n", ring->name);
418 return -ENOMEM;
419 }
420 }
421
422 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
423 /* create MQD for each KGQ */
424 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
425 ring = &adev->gfx.gfx_ring[i];
426 if (!ring->mqd_obj) {
427 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
428 domain, &ring->mqd_obj,
429 &ring->mqd_gpu_addr, &ring->mqd_ptr);
430 if (r) {
431 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
432 return r;
433 }
434
435 ring->mqd_size = mqd_size;
436 /* prepare MQD backup */
437 adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
438 if (!adev->gfx.me.mqd_backup[i]) {
439 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
440 return -ENOMEM;
441 }
442 }
443 }
444 }
445
446 /* create MQD for each KCQ */
447 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
448 j = i + xcc_id * adev->gfx.num_compute_rings;
449 ring = &adev->gfx.compute_ring[j];
450 if (!ring->mqd_obj) {
451 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
452 domain, &ring->mqd_obj,
453 &ring->mqd_gpu_addr, &ring->mqd_ptr);
454 if (r) {
455 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
456 return r;
457 }
458
459 ring->mqd_size = mqd_size;
460 /* prepare MQD backup */
461 adev->gfx.mec.mqd_backup[j] = kmalloc(mqd_size, GFP_KERNEL);
462 if (!adev->gfx.mec.mqd_backup[j]) {
463 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
464 return -ENOMEM;
465 }
466 }
467 }
468
469 return 0;
470}
471
472void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev, int xcc_id)
473{
474 struct amdgpu_ring *ring = NULL;
475 int i, j;
476 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
477
478 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
479 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
480 ring = &adev->gfx.gfx_ring[i];
481 kfree(adev->gfx.me.mqd_backup[i]);
482 amdgpu_bo_free_kernel(&ring->mqd_obj,
483 &ring->mqd_gpu_addr,
484 &ring->mqd_ptr);
485 }
486 }
487
488 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
489 j = i + xcc_id * adev->gfx.num_compute_rings;
490 ring = &adev->gfx.compute_ring[j];
491 kfree(adev->gfx.mec.mqd_backup[j]);
492 amdgpu_bo_free_kernel(&ring->mqd_obj,
493 &ring->mqd_gpu_addr,
494 &ring->mqd_ptr);
495 }
496
497 ring = &kiq->ring;
498 kfree(kiq->mqd_backup);
499 amdgpu_bo_free_kernel(&ring->mqd_obj,
500 &ring->mqd_gpu_addr,
501 &ring->mqd_ptr);
502}
503
504int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
505{
506 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
507 struct amdgpu_ring *kiq_ring = &kiq->ring;
508 struct amdgpu_hive_info *hive;
509 struct amdgpu_ras *ras;
510 int hive_ras_recovery = 0;
511 int i, r = 0;
512 int j;
513
514 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
515 return -EINVAL;
516
517 spin_lock(&kiq->ring_lock);
518 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
519 adev->gfx.num_compute_rings)) {
520 spin_unlock(&kiq->ring_lock);
521 return -ENOMEM;
522 }
523
524 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
525 j = i + xcc_id * adev->gfx.num_compute_rings;
526 kiq->pmf->kiq_unmap_queues(kiq_ring,
527 &adev->gfx.compute_ring[j],
528 RESET_QUEUES, 0, 0);
529 }
530
531 /**
532 * This is workaround: only skip kiq_ring test
533 * during ras recovery in suspend stage for gfx9.4.3
534 */
535 hive = amdgpu_get_xgmi_hive(adev);
536 if (hive) {
537 hive_ras_recovery = atomic_read(&hive->ras_recovery);
538 amdgpu_put_xgmi_hive(hive);
539 }
540
541 ras = amdgpu_ras_get_context(adev);
542 if ((amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) &&
543 ras && (atomic_read(&ras->in_recovery) || hive_ras_recovery)) {
544 spin_unlock(&kiq->ring_lock);
545 return 0;
546 }
547
548 if (kiq_ring->sched.ready && !adev->job_hang)
549 r = amdgpu_ring_test_helper(kiq_ring);
550 spin_unlock(&kiq->ring_lock);
551
552 return r;
553}
554
555int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id)
556{
557 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
558 struct amdgpu_ring *kiq_ring = &kiq->ring;
559 int i, r = 0;
560 int j;
561
562 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
563 return -EINVAL;
564
565 spin_lock(&kiq->ring_lock);
566 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
567 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
568 adev->gfx.num_gfx_rings)) {
569 spin_unlock(&kiq->ring_lock);
570 return -ENOMEM;
571 }
572
573 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
574 j = i + xcc_id * adev->gfx.num_gfx_rings;
575 kiq->pmf->kiq_unmap_queues(kiq_ring,
576 &adev->gfx.gfx_ring[j],
577 PREEMPT_QUEUES, 0, 0);
578 }
579 }
580
581 if (adev->gfx.kiq[0].ring.sched.ready && !adev->job_hang)
582 r = amdgpu_ring_test_helper(kiq_ring);
583 spin_unlock(&kiq->ring_lock);
584
585 return r;
586}
587
588int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
589 int queue_bit)
590{
591 int mec, pipe, queue;
592 int set_resource_bit = 0;
593
594 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
595
596 set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
597
598 return set_resource_bit;
599}
600
601int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
602{
603 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
604 struct amdgpu_ring *kiq_ring = &kiq->ring;
605 uint64_t queue_mask = 0;
606 int r, i, j;
607
608 if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
609 return -EINVAL;
610
611 for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
612 if (!test_bit(i, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
613 continue;
614
615 /* This situation may be hit in the future if a new HW
616 * generation exposes more than 64 queues. If so, the
617 * definition of queue_mask needs updating */
618 if (WARN_ON(i > (sizeof(queue_mask)*8))) {
619 DRM_ERROR("Invalid KCQ enabled: %d\n", i);
620 break;
621 }
622
623 queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
624 }
625
626 DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
627 kiq_ring->queue);
628 amdgpu_device_flush_hdp(adev, NULL);
629
630 spin_lock(&kiq->ring_lock);
631 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
632 adev->gfx.num_compute_rings +
633 kiq->pmf->set_resources_size);
634 if (r) {
635 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
636 spin_unlock(&kiq->ring_lock);
637 return r;
638 }
639
640 if (adev->enable_mes)
641 queue_mask = ~0ULL;
642
643 kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
644 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
645 j = i + xcc_id * adev->gfx.num_compute_rings;
646 kiq->pmf->kiq_map_queues(kiq_ring,
647 &adev->gfx.compute_ring[j]);
648 }
649
650 r = amdgpu_ring_test_helper(kiq_ring);
651 spin_unlock(&kiq->ring_lock);
652 if (r)
653 DRM_ERROR("KCQ enable failed\n");
654
655 return r;
656}
657
658int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
659{
660 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
661 struct amdgpu_ring *kiq_ring = &kiq->ring;
662 int r, i, j;
663
664 if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
665 return -EINVAL;
666
667 amdgpu_device_flush_hdp(adev, NULL);
668
669 spin_lock(&kiq->ring_lock);
670 /* No need to map kcq on the slave */
671 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
672 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
673 adev->gfx.num_gfx_rings);
674 if (r) {
675 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
676 spin_unlock(&kiq->ring_lock);
677 return r;
678 }
679
680 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
681 j = i + xcc_id * adev->gfx.num_gfx_rings;
682 kiq->pmf->kiq_map_queues(kiq_ring,
683 &adev->gfx.gfx_ring[j]);
684 }
685 }
686
687 r = amdgpu_ring_test_helper(kiq_ring);
688 spin_unlock(&kiq->ring_lock);
689 if (r)
690 DRM_ERROR("KGQ enable failed\n");
691
692 return r;
693}
694
695/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
696 *
697 * @adev: amdgpu_device pointer
698 * @bool enable true: enable gfx off feature, false: disable gfx off feature
699 *
700 * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
701 * 2. other client can send request to disable gfx off feature, the request should be honored.
702 * 3. other client can cancel their request of disable gfx off feature
703 * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
704 */
705
706void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
707{
708 unsigned long delay = GFX_OFF_DELAY_ENABLE;
709
710 if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
711 return;
712
713 mutex_lock(&adev->gfx.gfx_off_mutex);
714
715 if (enable) {
716 /* If the count is already 0, it means there's an imbalance bug somewhere.
717 * Note that the bug may be in a different caller than the one which triggers the
718 * WARN_ON_ONCE.
719 */
720 if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
721 goto unlock;
722
723 adev->gfx.gfx_off_req_count--;
724
725 if (adev->gfx.gfx_off_req_count == 0 &&
726 !adev->gfx.gfx_off_state) {
727 /* If going to s2idle, no need to wait */
728 if (adev->in_s0ix) {
729 if (!amdgpu_dpm_set_powergating_by_smu(adev,
730 AMD_IP_BLOCK_TYPE_GFX, true))
731 adev->gfx.gfx_off_state = true;
732 } else {
733 schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
734 delay);
735 }
736 }
737 } else {
738 if (adev->gfx.gfx_off_req_count == 0) {
739 cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
740
741 if (adev->gfx.gfx_off_state &&
742 !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
743 adev->gfx.gfx_off_state = false;
744
745 if (adev->gfx.funcs->init_spm_golden) {
746 dev_dbg(adev->dev,
747 "GFXOFF is disabled, re-init SPM golden settings\n");
748 amdgpu_gfx_init_spm_golden(adev);
749 }
750 }
751 }
752
753 adev->gfx.gfx_off_req_count++;
754 }
755
756unlock:
757 mutex_unlock(&adev->gfx.gfx_off_mutex);
758}
759
760int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
761{
762 int r = 0;
763
764 mutex_lock(&adev->gfx.gfx_off_mutex);
765
766 r = amdgpu_dpm_set_residency_gfxoff(adev, value);
767
768 mutex_unlock(&adev->gfx.gfx_off_mutex);
769
770 return r;
771}
772
773int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
774{
775 int r = 0;
776
777 mutex_lock(&adev->gfx.gfx_off_mutex);
778
779 r = amdgpu_dpm_get_residency_gfxoff(adev, value);
780
781 mutex_unlock(&adev->gfx.gfx_off_mutex);
782
783 return r;
784}
785
786int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
787{
788 int r = 0;
789
790 mutex_lock(&adev->gfx.gfx_off_mutex);
791
792 r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
793
794 mutex_unlock(&adev->gfx.gfx_off_mutex);
795
796 return r;
797}
798
799int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
800{
801
802 int r = 0;
803
804 mutex_lock(&adev->gfx.gfx_off_mutex);
805
806 r = amdgpu_dpm_get_status_gfxoff(adev, value);
807
808 mutex_unlock(&adev->gfx.gfx_off_mutex);
809
810 return r;
811}
812
813int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
814{
815 int r;
816
817 if (amdgpu_ras_is_supported(adev, ras_block->block)) {
818 if (!amdgpu_persistent_edc_harvesting_supported(adev))
819 amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
820
821 r = amdgpu_ras_block_late_init(adev, ras_block);
822 if (r)
823 return r;
824
825 if (adev->gfx.cp_ecc_error_irq.funcs) {
826 r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
827 if (r)
828 goto late_fini;
829 }
830 } else {
831 amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
832 }
833
834 return 0;
835late_fini:
836 amdgpu_ras_block_late_fini(adev, ras_block);
837 return r;
838}
839
840int amdgpu_gfx_ras_sw_init(struct amdgpu_device *adev)
841{
842 int err = 0;
843 struct amdgpu_gfx_ras *ras = NULL;
844
845 /* adev->gfx.ras is NULL, which means gfx does not
846 * support ras function, then do nothing here.
847 */
848 if (!adev->gfx.ras)
849 return 0;
850
851 ras = adev->gfx.ras;
852
853 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
854 if (err) {
855 dev_err(adev->dev, "Failed to register gfx ras block!\n");
856 return err;
857 }
858
859 strcpy(ras->ras_block.ras_comm.name, "gfx");
860 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__GFX;
861 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
862 adev->gfx.ras_if = &ras->ras_block.ras_comm;
863
864 /* If not define special ras_late_init function, use gfx default ras_late_init */
865 if (!ras->ras_block.ras_late_init)
866 ras->ras_block.ras_late_init = amdgpu_gfx_ras_late_init;
867
868 /* If not defined special ras_cb function, use default ras_cb */
869 if (!ras->ras_block.ras_cb)
870 ras->ras_block.ras_cb = amdgpu_gfx_process_ras_data_cb;
871
872 return 0;
873}
874
875int amdgpu_gfx_poison_consumption_handler(struct amdgpu_device *adev,
876 struct amdgpu_iv_entry *entry)
877{
878 if (adev->gfx.ras && adev->gfx.ras->poison_consumption_handler)
879 return adev->gfx.ras->poison_consumption_handler(adev, entry);
880
881 return 0;
882}
883
884int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
885 void *err_data,
886 struct amdgpu_iv_entry *entry)
887{
888 /* TODO ue will trigger an interrupt.
889 *
890 * When “Full RAS” is enabled, the per-IP interrupt sources should
891 * be disabled and the driver should only look for the aggregated
892 * interrupt via sync flood
893 */
894 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
895 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
896 if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
897 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
898 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
899 amdgpu_ras_reset_gpu(adev);
900 }
901 return AMDGPU_RAS_SUCCESS;
902}
903
904int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
905 struct amdgpu_irq_src *source,
906 struct amdgpu_iv_entry *entry)
907{
908 struct ras_common_if *ras_if = adev->gfx.ras_if;
909 struct ras_dispatch_if ih_data = {
910 .entry = entry,
911 };
912
913 if (!ras_if)
914 return 0;
915
916 ih_data.head = *ras_if;
917
918 DRM_ERROR("CP ECC ERROR IRQ\n");
919 amdgpu_ras_interrupt_dispatch(adev, &ih_data);
920 return 0;
921}
922
923void amdgpu_gfx_ras_error_func(struct amdgpu_device *adev,
924 void *ras_error_status,
925 void (*func)(struct amdgpu_device *adev, void *ras_error_status,
926 int xcc_id))
927{
928 int i;
929 int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
930 uint32_t xcc_mask = GENMASK(num_xcc - 1, 0);
931 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
932
933 if (err_data) {
934 err_data->ue_count = 0;
935 err_data->ce_count = 0;
936 }
937
938 for_each_inst(i, xcc_mask)
939 func(adev, ras_error_status, i);
940}
941
942uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg, uint32_t xcc_id)
943{
944 signed long r, cnt = 0;
945 unsigned long flags;
946 uint32_t seq, reg_val_offs = 0, value = 0;
947 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
948 struct amdgpu_ring *ring = &kiq->ring;
949
950 if (amdgpu_device_skip_hw_access(adev))
951 return 0;
952
953 if (adev->mes.ring.sched.ready)
954 return amdgpu_mes_rreg(adev, reg);
955
956 BUG_ON(!ring->funcs->emit_rreg);
957
958 spin_lock_irqsave(&kiq->ring_lock, flags);
959 if (amdgpu_device_wb_get(adev, ®_val_offs)) {
960 pr_err("critical bug! too many kiq readers\n");
961 goto failed_unlock;
962 }
963 amdgpu_ring_alloc(ring, 32);
964 amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
965 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
966 if (r)
967 goto failed_undo;
968
969 amdgpu_ring_commit(ring);
970 spin_unlock_irqrestore(&kiq->ring_lock, flags);
971
972 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
973
974 /* don't wait anymore for gpu reset case because this way may
975 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
976 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
977 * never return if we keep waiting in virt_kiq_rreg, which cause
978 * gpu_recover() hang there.
979 *
980 * also don't wait anymore for IRQ context
981 * */
982 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
983 goto failed_kiq_read;
984
985 might_sleep();
986 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
987 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
988 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
989 }
990
991 if (cnt > MAX_KIQ_REG_TRY)
992 goto failed_kiq_read;
993
994 mb();
995 value = adev->wb.wb[reg_val_offs];
996 amdgpu_device_wb_free(adev, reg_val_offs);
997 return value;
998
999failed_undo:
1000 amdgpu_ring_undo(ring);
1001failed_unlock:
1002 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1003failed_kiq_read:
1004 if (reg_val_offs)
1005 amdgpu_device_wb_free(adev, reg_val_offs);
1006 dev_err(adev->dev, "failed to read reg:%x\n", reg);
1007 return ~0;
1008}
1009
1010void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, uint32_t xcc_id)
1011{
1012 signed long r, cnt = 0;
1013 unsigned long flags;
1014 uint32_t seq;
1015 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
1016 struct amdgpu_ring *ring = &kiq->ring;
1017
1018 BUG_ON(!ring->funcs->emit_wreg);
1019
1020 if (amdgpu_device_skip_hw_access(adev))
1021 return;
1022
1023 if (adev->mes.ring.sched.ready) {
1024 amdgpu_mes_wreg(adev, reg, v);
1025 return;
1026 }
1027
1028 spin_lock_irqsave(&kiq->ring_lock, flags);
1029 amdgpu_ring_alloc(ring, 32);
1030 amdgpu_ring_emit_wreg(ring, reg, v);
1031 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
1032 if (r)
1033 goto failed_undo;
1034
1035 amdgpu_ring_commit(ring);
1036 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1037
1038 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1039
1040 /* don't wait anymore for gpu reset case because this way may
1041 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
1042 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
1043 * never return if we keep waiting in virt_kiq_rreg, which cause
1044 * gpu_recover() hang there.
1045 *
1046 * also don't wait anymore for IRQ context
1047 * */
1048 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
1049 goto failed_kiq_write;
1050
1051 might_sleep();
1052 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
1053
1054 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
1055 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
1056 }
1057
1058 if (cnt > MAX_KIQ_REG_TRY)
1059 goto failed_kiq_write;
1060
1061 return;
1062
1063failed_undo:
1064 amdgpu_ring_undo(ring);
1065 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1066failed_kiq_write:
1067 dev_err(adev->dev, "failed to write reg:%x\n", reg);
1068}
1069
1070int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
1071{
1072 if (amdgpu_num_kcq == -1) {
1073 return 8;
1074 } else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
1075 dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
1076 return 8;
1077 }
1078 return amdgpu_num_kcq;
1079}
1080
1081void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
1082 uint32_t ucode_id)
1083{
1084 const struct gfx_firmware_header_v1_0 *cp_hdr;
1085 const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
1086 struct amdgpu_firmware_info *info = NULL;
1087 const struct firmware *ucode_fw;
1088 unsigned int fw_size;
1089
1090 switch (ucode_id) {
1091 case AMDGPU_UCODE_ID_CP_PFP:
1092 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1093 adev->gfx.pfp_fw->data;
1094 adev->gfx.pfp_fw_version =
1095 le32_to_cpu(cp_hdr->header.ucode_version);
1096 adev->gfx.pfp_feature_version =
1097 le32_to_cpu(cp_hdr->ucode_feature_version);
1098 ucode_fw = adev->gfx.pfp_fw;
1099 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1100 break;
1101 case AMDGPU_UCODE_ID_CP_RS64_PFP:
1102 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1103 adev->gfx.pfp_fw->data;
1104 adev->gfx.pfp_fw_version =
1105 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1106 adev->gfx.pfp_feature_version =
1107 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1108 ucode_fw = adev->gfx.pfp_fw;
1109 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1110 break;
1111 case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
1112 case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
1113 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1114 adev->gfx.pfp_fw->data;
1115 ucode_fw = adev->gfx.pfp_fw;
1116 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1117 break;
1118 case AMDGPU_UCODE_ID_CP_ME:
1119 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1120 adev->gfx.me_fw->data;
1121 adev->gfx.me_fw_version =
1122 le32_to_cpu(cp_hdr->header.ucode_version);
1123 adev->gfx.me_feature_version =
1124 le32_to_cpu(cp_hdr->ucode_feature_version);
1125 ucode_fw = adev->gfx.me_fw;
1126 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1127 break;
1128 case AMDGPU_UCODE_ID_CP_RS64_ME:
1129 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1130 adev->gfx.me_fw->data;
1131 adev->gfx.me_fw_version =
1132 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1133 adev->gfx.me_feature_version =
1134 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1135 ucode_fw = adev->gfx.me_fw;
1136 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1137 break;
1138 case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
1139 case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
1140 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1141 adev->gfx.me_fw->data;
1142 ucode_fw = adev->gfx.me_fw;
1143 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1144 break;
1145 case AMDGPU_UCODE_ID_CP_CE:
1146 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1147 adev->gfx.ce_fw->data;
1148 adev->gfx.ce_fw_version =
1149 le32_to_cpu(cp_hdr->header.ucode_version);
1150 adev->gfx.ce_feature_version =
1151 le32_to_cpu(cp_hdr->ucode_feature_version);
1152 ucode_fw = adev->gfx.ce_fw;
1153 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1154 break;
1155 case AMDGPU_UCODE_ID_CP_MEC1:
1156 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1157 adev->gfx.mec_fw->data;
1158 adev->gfx.mec_fw_version =
1159 le32_to_cpu(cp_hdr->header.ucode_version);
1160 adev->gfx.mec_feature_version =
1161 le32_to_cpu(cp_hdr->ucode_feature_version);
1162 ucode_fw = adev->gfx.mec_fw;
1163 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1164 le32_to_cpu(cp_hdr->jt_size) * 4;
1165 break;
1166 case AMDGPU_UCODE_ID_CP_MEC1_JT:
1167 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1168 adev->gfx.mec_fw->data;
1169 ucode_fw = adev->gfx.mec_fw;
1170 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1171 break;
1172 case AMDGPU_UCODE_ID_CP_MEC2:
1173 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1174 adev->gfx.mec2_fw->data;
1175 adev->gfx.mec2_fw_version =
1176 le32_to_cpu(cp_hdr->header.ucode_version);
1177 adev->gfx.mec2_feature_version =
1178 le32_to_cpu(cp_hdr->ucode_feature_version);
1179 ucode_fw = adev->gfx.mec2_fw;
1180 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1181 le32_to_cpu(cp_hdr->jt_size) * 4;
1182 break;
1183 case AMDGPU_UCODE_ID_CP_MEC2_JT:
1184 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1185 adev->gfx.mec2_fw->data;
1186 ucode_fw = adev->gfx.mec2_fw;
1187 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1188 break;
1189 case AMDGPU_UCODE_ID_CP_RS64_MEC:
1190 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1191 adev->gfx.mec_fw->data;
1192 adev->gfx.mec_fw_version =
1193 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1194 adev->gfx.mec_feature_version =
1195 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1196 ucode_fw = adev->gfx.mec_fw;
1197 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1198 break;
1199 case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
1200 case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
1201 case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
1202 case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
1203 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1204 adev->gfx.mec_fw->data;
1205 ucode_fw = adev->gfx.mec_fw;
1206 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1207 break;
1208 default:
1209 break;
1210 }
1211
1212 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1213 info = &adev->firmware.ucode[ucode_id];
1214 info->ucode_id = ucode_id;
1215 info->fw = ucode_fw;
1216 adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE);
1217 }
1218}
1219
1220bool amdgpu_gfx_is_master_xcc(struct amdgpu_device *adev, int xcc_id)
1221{
1222 return !(xcc_id % (adev->gfx.num_xcc_per_xcp ?
1223 adev->gfx.num_xcc_per_xcp : 1));
1224}
1225
1226static ssize_t amdgpu_gfx_get_current_compute_partition(struct device *dev,
1227 struct device_attribute *addr,
1228 char *buf)
1229{
1230 struct drm_device *ddev = dev_get_drvdata(dev);
1231 struct amdgpu_device *adev = drm_to_adev(ddev);
1232 int mode;
1233
1234 mode = amdgpu_xcp_query_partition_mode(adev->xcp_mgr,
1235 AMDGPU_XCP_FL_NONE);
1236
1237 return sysfs_emit(buf, "%s\n", amdgpu_gfx_compute_mode_desc(mode));
1238}
1239
1240static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev,
1241 struct device_attribute *addr,
1242 const char *buf, size_t count)
1243{
1244 struct drm_device *ddev = dev_get_drvdata(dev);
1245 struct amdgpu_device *adev = drm_to_adev(ddev);
1246 enum amdgpu_gfx_partition mode;
1247 int ret = 0, num_xcc;
1248
1249 num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1250 if (num_xcc % 2 != 0)
1251 return -EINVAL;
1252
1253 if (!strncasecmp("SPX", buf, strlen("SPX"))) {
1254 mode = AMDGPU_SPX_PARTITION_MODE;
1255 } else if (!strncasecmp("DPX", buf, strlen("DPX"))) {
1256 /*
1257 * DPX mode needs AIDs to be in multiple of 2.
1258 * Each AID connects 2 XCCs.
1259 */
1260 if (num_xcc%4)
1261 return -EINVAL;
1262 mode = AMDGPU_DPX_PARTITION_MODE;
1263 } else if (!strncasecmp("TPX", buf, strlen("TPX"))) {
1264 if (num_xcc != 6)
1265 return -EINVAL;
1266 mode = AMDGPU_TPX_PARTITION_MODE;
1267 } else if (!strncasecmp("QPX", buf, strlen("QPX"))) {
1268 if (num_xcc != 8)
1269 return -EINVAL;
1270 mode = AMDGPU_QPX_PARTITION_MODE;
1271 } else if (!strncasecmp("CPX", buf, strlen("CPX"))) {
1272 mode = AMDGPU_CPX_PARTITION_MODE;
1273 } else {
1274 return -EINVAL;
1275 }
1276
1277 ret = amdgpu_xcp_switch_partition_mode(adev->xcp_mgr, mode);
1278
1279 if (ret)
1280 return ret;
1281
1282 return count;
1283}
1284
1285static ssize_t amdgpu_gfx_get_available_compute_partition(struct device *dev,
1286 struct device_attribute *addr,
1287 char *buf)
1288{
1289 struct drm_device *ddev = dev_get_drvdata(dev);
1290 struct amdgpu_device *adev = drm_to_adev(ddev);
1291 char *supported_partition;
1292
1293 /* TBD */
1294 switch (NUM_XCC(adev->gfx.xcc_mask)) {
1295 case 8:
1296 supported_partition = "SPX, DPX, QPX, CPX";
1297 break;
1298 case 6:
1299 supported_partition = "SPX, TPX, CPX";
1300 break;
1301 case 4:
1302 supported_partition = "SPX, DPX, CPX";
1303 break;
1304 /* this seems only existing in emulation phase */
1305 case 2:
1306 supported_partition = "SPX, CPX";
1307 break;
1308 default:
1309 supported_partition = "Not supported";
1310 break;
1311 }
1312
1313 return sysfs_emit(buf, "%s\n", supported_partition);
1314}
1315
1316static DEVICE_ATTR(current_compute_partition, 0644,
1317 amdgpu_gfx_get_current_compute_partition,
1318 amdgpu_gfx_set_compute_partition);
1319
1320static DEVICE_ATTR(available_compute_partition, 0444,
1321 amdgpu_gfx_get_available_compute_partition, NULL);
1322
1323int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev)
1324{
1325 int r;
1326
1327 r = device_create_file(adev->dev, &dev_attr_current_compute_partition);
1328 if (r)
1329 return r;
1330
1331 r = device_create_file(adev->dev, &dev_attr_available_compute_partition);
1332
1333 return r;
1334}
1335
1336void amdgpu_gfx_sysfs_fini(struct amdgpu_device *adev)
1337{
1338 device_remove_file(adev->dev, &dev_attr_current_compute_partition);
1339 device_remove_file(adev->dev, &dev_attr_available_compute_partition);
1340}
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}