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
amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device * adev,int mec,int pipe,int queue)46 int 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
amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device * adev,int bit,int * mec,int * pipe,int * queue)59 void 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
amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device * adev,int xcc_id,int mec,int pipe,int queue)70 bool 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
amdgpu_gfx_me_queue_to_bit(struct amdgpu_device * adev,int me,int pipe,int queue)77 int 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
amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device * adev,int me,int pipe,int queue)90 bool 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 */
amdgpu_gfx_parse_disable_cu(unsigned int * mask,unsigned int max_se,unsigned int max_sh)107 void 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
amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device * adev)142 static 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
amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device * adev)147 static 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
amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device * adev,struct amdgpu_ring * ring)166 bool 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
amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device * adev,struct amdgpu_ring * ring)188 bool 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
amdgpu_gfx_compute_queue_acquire(struct amdgpu_device * adev)201 void 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
amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device * adev)237 void 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
amdgpu_gfx_kiq_acquire(struct amdgpu_device * adev,struct amdgpu_ring * ring,int xcc_id)265 static 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
amdgpu_gfx_kiq_init_ring(struct amdgpu_device * adev,int xcc_id)300 int 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
amdgpu_gfx_kiq_free_ring(struct amdgpu_ring * ring)336 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
337 {
338 amdgpu_ring_fini(ring);
339 }
340
amdgpu_gfx_kiq_fini(struct amdgpu_device * adev,int xcc_id)341 void 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
amdgpu_gfx_kiq_init(struct amdgpu_device * adev,unsigned int hpd_size,int xcc_id)348 int 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 */
amdgpu_gfx_mqd_sw_init(struct amdgpu_device * adev,unsigned int mqd_size,int xcc_id)375 int 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
amdgpu_gfx_mqd_sw_fini(struct amdgpu_device * adev,int xcc_id)466 void 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
amdgpu_gfx_disable_kcq(struct amdgpu_device * adev,int xcc_id)498 int 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
amdgpu_gfx_disable_kgq(struct amdgpu_device * adev,int xcc_id)548 int 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
amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device * adev,int queue_bit)602 int 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
amdgpu_gfx_mes_enable_kcq(struct amdgpu_device * adev,int xcc_id)615 static 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
amdgpu_gfx_enable_kcq(struct amdgpu_device * adev,int xcc_id)653 int 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
amdgpu_gfx_enable_kgq(struct amdgpu_device * adev,int xcc_id)717 int 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
amdgpu_gfx_off_ctrl(struct amdgpu_device * adev,bool enable)785 void 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
835 unlock:
836 mutex_unlock(&adev->gfx.gfx_off_mutex);
837 }
838
amdgpu_set_gfx_off_residency(struct amdgpu_device * adev,bool value)839 int 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
amdgpu_get_gfx_off_residency(struct amdgpu_device * adev,u32 * value)852 int 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
amdgpu_get_gfx_off_entrycount(struct amdgpu_device * adev,u64 * value)865 int 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
amdgpu_get_gfx_off_status(struct amdgpu_device * adev,uint32_t * value)878 int 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
amdgpu_gfx_ras_late_init(struct amdgpu_device * adev,struct ras_common_if * ras_block)892 int 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;
920 late_fini:
921 amdgpu_ras_block_late_fini(adev, ras_block);
922 return r;
923 }
924
amdgpu_gfx_ras_sw_init(struct amdgpu_device * adev)925 int 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
amdgpu_gfx_poison_consumption_handler(struct amdgpu_device * adev,struct amdgpu_iv_entry * entry)960 int 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
amdgpu_gfx_process_ras_data_cb(struct amdgpu_device * adev,void * err_data,struct amdgpu_iv_entry * entry)969 int 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
amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)989 int 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
amdgpu_gfx_ras_error_func(struct amdgpu_device * adev,void * ras_error_status,void (* func)(struct amdgpu_device * adev,void * ras_error_status,int xcc_id))1008 void 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
amdgpu_kiq_rreg(struct amdgpu_device * adev,uint32_t reg,uint32_t xcc_id)1027 uint32_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
1087 failed_undo:
1088 amdgpu_ring_undo(ring);
1089 failed_unlock:
1090 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1091 failed_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
amdgpu_kiq_wreg(struct amdgpu_device * adev,uint32_t reg,uint32_t v,uint32_t xcc_id)1098 void 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
1154 failed_undo:
1155 amdgpu_ring_undo(ring);
1156 failed_unlock:
1157 spin_unlock_irqrestore(&kiq->ring_lock, flags);
1158 failed_kiq_write:
1159 dev_err(adev->dev, "failed to write reg:%x\n", reg);
1160 }
1161
amdgpu_gfx_get_num_kcq(struct amdgpu_device * adev)1162 int 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
amdgpu_gfx_cp_init_microcode(struct amdgpu_device * adev,uint32_t ucode_id)1173 void 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
amdgpu_gfx_is_master_xcc(struct amdgpu_device * adev,int xcc_id)1313 bool 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
amdgpu_gfx_get_current_compute_partition(struct device * dev,struct device_attribute * addr,char * buf)1319 static 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
amdgpu_gfx_set_compute_partition(struct device * dev,struct device_attribute * addr,const char * buf,size_t count)1333 static 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
1378 static 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
amdgpu_gfx_get_available_compute_partition(struct device * dev,struct device_attribute * addr,char * buf)1386 static 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
amdgpu_gfx_run_cleaner_shader_job(struct amdgpu_ring * ring)1409 static 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
1452 err:
1453 return r;
1454 }
1455
amdgpu_gfx_run_cleaner_shader(struct amdgpu_device * adev,int xcp_id)1456 static 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
amdgpu_gfx_set_run_cleaner_shader(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)1487 static 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
amdgpu_gfx_get_enforce_isolation(struct device * dev,struct device_attribute * attr,char * buf)1535 static 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
amdgpu_gfx_set_enforce_isolation(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)1558 static 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
1597 for (i = 0; i < num_partitions; i++) {
1598 if (adev->enforce_isolation[i] && !partition_values[i]) {
1599 /* Going from enabled to disabled */
1600 amdgpu_vmid_free_reserved(adev, AMDGPU_GFXHUB(i));
1601 amdgpu_mes_set_enforce_isolation(adev, i, false);
1602 } else if (!adev->enforce_isolation[i] && partition_values[i]) {
1603 /* Going from disabled to enabled */
1604 amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(i));
1605 amdgpu_mes_set_enforce_isolation(adev, i, true);
1606 }
1607 adev->enforce_isolation[i] = partition_values[i];
1608 }
1609
1610 mutex_unlock(&adev->enforce_isolation_mutex);
1611
1612 return count;
1613 }
1614
amdgpu_gfx_get_gfx_reset_mask(struct device * dev,struct device_attribute * attr,char * buf)1615 static ssize_t amdgpu_gfx_get_gfx_reset_mask(struct device *dev,
1616 struct device_attribute *attr,
1617 char *buf)
1618 {
1619 struct drm_device *ddev = dev_get_drvdata(dev);
1620 struct amdgpu_device *adev = drm_to_adev(ddev);
1621
1622 if (!adev)
1623 return -ENODEV;
1624
1625 return amdgpu_show_reset_mask(buf, adev->gfx.gfx_supported_reset);
1626 }
1627
amdgpu_gfx_get_compute_reset_mask(struct device * dev,struct device_attribute * attr,char * buf)1628 static ssize_t amdgpu_gfx_get_compute_reset_mask(struct device *dev,
1629 struct device_attribute *attr,
1630 char *buf)
1631 {
1632 struct drm_device *ddev = dev_get_drvdata(dev);
1633 struct amdgpu_device *adev = drm_to_adev(ddev);
1634
1635 if (!adev)
1636 return -ENODEV;
1637
1638 return amdgpu_show_reset_mask(buf, adev->gfx.compute_supported_reset);
1639 }
1640
1641 static DEVICE_ATTR(run_cleaner_shader, 0200,
1642 NULL, amdgpu_gfx_set_run_cleaner_shader);
1643
1644 static DEVICE_ATTR(enforce_isolation, 0644,
1645 amdgpu_gfx_get_enforce_isolation,
1646 amdgpu_gfx_set_enforce_isolation);
1647
1648 static DEVICE_ATTR(current_compute_partition, 0644,
1649 amdgpu_gfx_get_current_compute_partition,
1650 amdgpu_gfx_set_compute_partition);
1651
1652 static DEVICE_ATTR(available_compute_partition, 0444,
1653 amdgpu_gfx_get_available_compute_partition, NULL);
1654 static DEVICE_ATTR(gfx_reset_mask, 0444,
1655 amdgpu_gfx_get_gfx_reset_mask, NULL);
1656
1657 static DEVICE_ATTR(compute_reset_mask, 0444,
1658 amdgpu_gfx_get_compute_reset_mask, NULL);
1659
amdgpu_gfx_sysfs_xcp_init(struct amdgpu_device * adev)1660 static int amdgpu_gfx_sysfs_xcp_init(struct amdgpu_device *adev)
1661 {
1662 struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
1663 bool xcp_switch_supported;
1664 int r;
1665
1666 if (!xcp_mgr)
1667 return 0;
1668
1669 xcp_switch_supported =
1670 (xcp_mgr->funcs && xcp_mgr->funcs->switch_partition_mode);
1671
1672 if (!xcp_switch_supported)
1673 dev_attr_current_compute_partition.attr.mode &=
1674 ~(S_IWUSR | S_IWGRP | S_IWOTH);
1675
1676 r = device_create_file(adev->dev, &dev_attr_current_compute_partition);
1677 if (r)
1678 return r;
1679
1680 if (xcp_switch_supported)
1681 r = device_create_file(adev->dev,
1682 &dev_attr_available_compute_partition);
1683
1684 return r;
1685 }
1686
amdgpu_gfx_sysfs_xcp_fini(struct amdgpu_device * adev)1687 static void amdgpu_gfx_sysfs_xcp_fini(struct amdgpu_device *adev)
1688 {
1689 struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
1690 bool xcp_switch_supported;
1691
1692 if (!xcp_mgr)
1693 return;
1694
1695 xcp_switch_supported =
1696 (xcp_mgr->funcs && xcp_mgr->funcs->switch_partition_mode);
1697 device_remove_file(adev->dev, &dev_attr_current_compute_partition);
1698
1699 if (xcp_switch_supported)
1700 device_remove_file(adev->dev,
1701 &dev_attr_available_compute_partition);
1702 }
1703
amdgpu_gfx_sysfs_isolation_shader_init(struct amdgpu_device * adev)1704 static int amdgpu_gfx_sysfs_isolation_shader_init(struct amdgpu_device *adev)
1705 {
1706 int r;
1707
1708 r = device_create_file(adev->dev, &dev_attr_enforce_isolation);
1709 if (r)
1710 return r;
1711 if (adev->gfx.enable_cleaner_shader)
1712 r = device_create_file(adev->dev, &dev_attr_run_cleaner_shader);
1713
1714 return r;
1715 }
1716
amdgpu_gfx_sysfs_isolation_shader_fini(struct amdgpu_device * adev)1717 static void amdgpu_gfx_sysfs_isolation_shader_fini(struct amdgpu_device *adev)
1718 {
1719 device_remove_file(adev->dev, &dev_attr_enforce_isolation);
1720 if (adev->gfx.enable_cleaner_shader)
1721 device_remove_file(adev->dev, &dev_attr_run_cleaner_shader);
1722 }
1723
amdgpu_gfx_sysfs_reset_mask_init(struct amdgpu_device * adev)1724 static int amdgpu_gfx_sysfs_reset_mask_init(struct amdgpu_device *adev)
1725 {
1726 int r = 0;
1727
1728 if (!amdgpu_gpu_recovery)
1729 return r;
1730
1731 if (adev->gfx.num_gfx_rings) {
1732 r = device_create_file(adev->dev, &dev_attr_gfx_reset_mask);
1733 if (r)
1734 return r;
1735 }
1736
1737 if (adev->gfx.num_compute_rings) {
1738 r = device_create_file(adev->dev, &dev_attr_compute_reset_mask);
1739 if (r)
1740 return r;
1741 }
1742
1743 return r;
1744 }
1745
amdgpu_gfx_sysfs_reset_mask_fini(struct amdgpu_device * adev)1746 static void amdgpu_gfx_sysfs_reset_mask_fini(struct amdgpu_device *adev)
1747 {
1748 if (!amdgpu_gpu_recovery)
1749 return;
1750
1751 if (adev->gfx.num_gfx_rings)
1752 device_remove_file(adev->dev, &dev_attr_gfx_reset_mask);
1753
1754 if (adev->gfx.num_compute_rings)
1755 device_remove_file(adev->dev, &dev_attr_compute_reset_mask);
1756 }
1757
amdgpu_gfx_sysfs_init(struct amdgpu_device * adev)1758 int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev)
1759 {
1760 int r;
1761
1762 r = amdgpu_gfx_sysfs_xcp_init(adev);
1763 if (r) {
1764 dev_err(adev->dev, "failed to create xcp sysfs files");
1765 return r;
1766 }
1767
1768 r = amdgpu_gfx_sysfs_isolation_shader_init(adev);
1769 if (r)
1770 dev_err(adev->dev, "failed to create isolation sysfs files");
1771
1772 r = amdgpu_gfx_sysfs_reset_mask_init(adev);
1773 if (r)
1774 dev_err(adev->dev, "failed to create reset mask sysfs files");
1775
1776 return r;
1777 }
1778
amdgpu_gfx_sysfs_fini(struct amdgpu_device * adev)1779 void amdgpu_gfx_sysfs_fini(struct amdgpu_device *adev)
1780 {
1781 if (adev->dev->kobj.sd) {
1782 amdgpu_gfx_sysfs_xcp_fini(adev);
1783 amdgpu_gfx_sysfs_isolation_shader_fini(adev);
1784 amdgpu_gfx_sysfs_reset_mask_fini(adev);
1785 }
1786 }
1787
amdgpu_gfx_cleaner_shader_sw_init(struct amdgpu_device * adev,unsigned int cleaner_shader_size)1788 int amdgpu_gfx_cleaner_shader_sw_init(struct amdgpu_device *adev,
1789 unsigned int cleaner_shader_size)
1790 {
1791 if (!adev->gfx.enable_cleaner_shader)
1792 return -EOPNOTSUPP;
1793
1794 return amdgpu_bo_create_kernel(adev, cleaner_shader_size, PAGE_SIZE,
1795 AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT,
1796 &adev->gfx.cleaner_shader_obj,
1797 &adev->gfx.cleaner_shader_gpu_addr,
1798 (void **)&adev->gfx.cleaner_shader_cpu_ptr);
1799 }
1800
amdgpu_gfx_cleaner_shader_sw_fini(struct amdgpu_device * adev)1801 void amdgpu_gfx_cleaner_shader_sw_fini(struct amdgpu_device *adev)
1802 {
1803 if (!adev->gfx.enable_cleaner_shader)
1804 return;
1805
1806 amdgpu_bo_free_kernel(&adev->gfx.cleaner_shader_obj,
1807 &adev->gfx.cleaner_shader_gpu_addr,
1808 (void **)&adev->gfx.cleaner_shader_cpu_ptr);
1809 }
1810
amdgpu_gfx_cleaner_shader_init(struct amdgpu_device * adev,unsigned int cleaner_shader_size,const void * cleaner_shader_ptr)1811 void amdgpu_gfx_cleaner_shader_init(struct amdgpu_device *adev,
1812 unsigned int cleaner_shader_size,
1813 const void *cleaner_shader_ptr)
1814 {
1815 if (!adev->gfx.enable_cleaner_shader)
1816 return;
1817
1818 if (adev->gfx.cleaner_shader_cpu_ptr && cleaner_shader_ptr)
1819 memcpy_toio(adev->gfx.cleaner_shader_cpu_ptr, cleaner_shader_ptr,
1820 cleaner_shader_size);
1821 }
1822
1823 /**
1824 * amdgpu_gfx_kfd_sch_ctrl - Control the KFD scheduler from the KGD (Graphics Driver)
1825 * @adev: amdgpu_device pointer
1826 * @idx: Index of the scheduler to control
1827 * @enable: Whether to enable or disable the KFD scheduler
1828 *
1829 * This function is used to control the KFD (Kernel Fusion Driver) scheduler
1830 * from the KGD. It is part of the cleaner shader feature. This function plays
1831 * a key role in enforcing process isolation on the GPU.
1832 *
1833 * The function uses a reference count mechanism (kfd_sch_req_count) to keep
1834 * track of the number of requests to enable the KFD scheduler. When a request
1835 * to enable the KFD scheduler is made, the reference count is decremented.
1836 * When the reference count reaches zero, a delayed work is scheduled to
1837 * enforce isolation after a delay of GFX_SLICE_PERIOD.
1838 *
1839 * When a request to disable the KFD scheduler is made, the function first
1840 * checks if the reference count is zero. If it is, it cancels the delayed work
1841 * for enforcing isolation and checks if the KFD scheduler is active. If the
1842 * KFD scheduler is active, it sends a request to stop the KFD scheduler and
1843 * sets the KFD scheduler state to inactive. Then, it increments the reference
1844 * count.
1845 *
1846 * The function is synchronized using the kfd_sch_mutex to ensure that the KFD
1847 * scheduler state and reference count are updated atomically.
1848 *
1849 * Note: If the reference count is already zero when a request to enable the
1850 * KFD scheduler is made, it means there's an imbalance bug somewhere. The
1851 * function triggers a warning in this case.
1852 */
amdgpu_gfx_kfd_sch_ctrl(struct amdgpu_device * adev,u32 idx,bool enable)1853 static void amdgpu_gfx_kfd_sch_ctrl(struct amdgpu_device *adev, u32 idx,
1854 bool enable)
1855 {
1856 mutex_lock(&adev->gfx.kfd_sch_mutex);
1857
1858 if (enable) {
1859 /* If the count is already 0, it means there's an imbalance bug somewhere.
1860 * Note that the bug may be in a different caller than the one which triggers the
1861 * WARN_ON_ONCE.
1862 */
1863 if (WARN_ON_ONCE(adev->gfx.kfd_sch_req_count[idx] == 0)) {
1864 dev_err(adev->dev, "Attempted to enable KFD scheduler when reference count is already zero\n");
1865 goto unlock;
1866 }
1867
1868 adev->gfx.kfd_sch_req_count[idx]--;
1869
1870 if (adev->gfx.kfd_sch_req_count[idx] == 0 &&
1871 adev->gfx.kfd_sch_inactive[idx]) {
1872 schedule_delayed_work(&adev->gfx.enforce_isolation[idx].work,
1873 msecs_to_jiffies(adev->gfx.enforce_isolation_time[idx]));
1874 }
1875 } else {
1876 if (adev->gfx.kfd_sch_req_count[idx] == 0) {
1877 cancel_delayed_work_sync(&adev->gfx.enforce_isolation[idx].work);
1878 if (!adev->gfx.kfd_sch_inactive[idx]) {
1879 amdgpu_amdkfd_stop_sched(adev, idx);
1880 adev->gfx.kfd_sch_inactive[idx] = true;
1881 }
1882 }
1883
1884 adev->gfx.kfd_sch_req_count[idx]++;
1885 }
1886
1887 unlock:
1888 mutex_unlock(&adev->gfx.kfd_sch_mutex);
1889 }
1890
1891 /**
1892 * amdgpu_gfx_enforce_isolation_handler - work handler for enforcing shader isolation
1893 *
1894 * @work: work_struct.
1895 *
1896 * This function is the work handler for enforcing shader isolation on AMD GPUs.
1897 * It counts the number of emitted fences for each GFX and compute ring. If there
1898 * are any fences, it schedules the `enforce_isolation_work` to be run after a
1899 * delay of `GFX_SLICE_PERIOD`. If there are no fences, it signals the Kernel Fusion
1900 * Driver (KFD) to resume the runqueue. The function is synchronized using the
1901 * `enforce_isolation_mutex`.
1902 */
amdgpu_gfx_enforce_isolation_handler(struct work_struct * work)1903 void amdgpu_gfx_enforce_isolation_handler(struct work_struct *work)
1904 {
1905 struct amdgpu_isolation_work *isolation_work =
1906 container_of(work, struct amdgpu_isolation_work, work.work);
1907 struct amdgpu_device *adev = isolation_work->adev;
1908 u32 i, idx, fences = 0;
1909
1910 if (isolation_work->xcp_id == AMDGPU_XCP_NO_PARTITION)
1911 idx = 0;
1912 else
1913 idx = isolation_work->xcp_id;
1914
1915 if (idx >= MAX_XCP)
1916 return;
1917
1918 mutex_lock(&adev->enforce_isolation_mutex);
1919 for (i = 0; i < AMDGPU_MAX_GFX_RINGS; ++i) {
1920 if (isolation_work->xcp_id == adev->gfx.gfx_ring[i].xcp_id)
1921 fences += amdgpu_fence_count_emitted(&adev->gfx.gfx_ring[i]);
1922 }
1923 for (i = 0; i < (AMDGPU_MAX_COMPUTE_RINGS * AMDGPU_MAX_GC_INSTANCES); ++i) {
1924 if (isolation_work->xcp_id == adev->gfx.compute_ring[i].xcp_id)
1925 fences += amdgpu_fence_count_emitted(&adev->gfx.compute_ring[i]);
1926 }
1927 if (fences) {
1928 /* we've already had our timeslice, so let's wrap this up */
1929 schedule_delayed_work(&adev->gfx.enforce_isolation[idx].work,
1930 msecs_to_jiffies(1));
1931 } else {
1932 /* Tell KFD to resume the runqueue */
1933 if (adev->kfd.init_complete) {
1934 WARN_ON_ONCE(!adev->gfx.kfd_sch_inactive[idx]);
1935 WARN_ON_ONCE(adev->gfx.kfd_sch_req_count[idx]);
1936 amdgpu_amdkfd_start_sched(adev, idx);
1937 adev->gfx.kfd_sch_inactive[idx] = false;
1938 }
1939 }
1940 mutex_unlock(&adev->enforce_isolation_mutex);
1941 }
1942
1943 static void
amdgpu_gfx_enforce_isolation_wait_for_kfd(struct amdgpu_device * adev,u32 idx)1944 amdgpu_gfx_enforce_isolation_wait_for_kfd(struct amdgpu_device *adev,
1945 u32 idx)
1946 {
1947 unsigned long cjiffies;
1948 bool wait = false;
1949
1950 mutex_lock(&adev->enforce_isolation_mutex);
1951 if (adev->enforce_isolation[idx]) {
1952 /* set the initial values if nothing is set */
1953 if (!adev->gfx.enforce_isolation_jiffies[idx]) {
1954 adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
1955 adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
1956 }
1957 /* Make sure KFD gets a chance to run */
1958 if (amdgpu_amdkfd_compute_active(adev, idx)) {
1959 cjiffies = jiffies;
1960 if (time_after(cjiffies, adev->gfx.enforce_isolation_jiffies[idx])) {
1961 cjiffies -= adev->gfx.enforce_isolation_jiffies[idx];
1962 if ((jiffies_to_msecs(cjiffies) >= GFX_SLICE_PERIOD_MS)) {
1963 /* if our time is up, let KGD work drain before scheduling more */
1964 wait = true;
1965 /* reset the timer period */
1966 adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
1967 } else {
1968 /* set the timer period to what's left in our time slice */
1969 adev->gfx.enforce_isolation_time[idx] =
1970 GFX_SLICE_PERIOD_MS - jiffies_to_msecs(cjiffies);
1971 }
1972 } else {
1973 /* if jiffies wrap around we will just wait a little longer */
1974 adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
1975 }
1976 } else {
1977 /* if there is no KFD work, then set the full slice period */
1978 adev->gfx.enforce_isolation_jiffies[idx] = jiffies;
1979 adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS;
1980 }
1981 }
1982 mutex_unlock(&adev->enforce_isolation_mutex);
1983
1984 if (wait)
1985 msleep(GFX_SLICE_PERIOD_MS);
1986 }
1987
amdgpu_gfx_enforce_isolation_ring_begin_use(struct amdgpu_ring * ring)1988 void amdgpu_gfx_enforce_isolation_ring_begin_use(struct amdgpu_ring *ring)
1989 {
1990 struct amdgpu_device *adev = ring->adev;
1991 u32 idx;
1992
1993 if (!adev->gfx.enable_cleaner_shader)
1994 return;
1995
1996 if (ring->xcp_id == AMDGPU_XCP_NO_PARTITION)
1997 idx = 0;
1998 else
1999 idx = ring->xcp_id;
2000
2001 if (idx >= MAX_XCP)
2002 return;
2003
2004 /* Don't submit more work until KFD has had some time */
2005 amdgpu_gfx_enforce_isolation_wait_for_kfd(adev, idx);
2006
2007 mutex_lock(&adev->enforce_isolation_mutex);
2008 if (adev->enforce_isolation[idx]) {
2009 if (adev->kfd.init_complete)
2010 amdgpu_gfx_kfd_sch_ctrl(adev, idx, false);
2011 }
2012 mutex_unlock(&adev->enforce_isolation_mutex);
2013 }
2014
amdgpu_gfx_enforce_isolation_ring_end_use(struct amdgpu_ring * ring)2015 void amdgpu_gfx_enforce_isolation_ring_end_use(struct amdgpu_ring *ring)
2016 {
2017 struct amdgpu_device *adev = ring->adev;
2018 u32 idx;
2019
2020 if (!adev->gfx.enable_cleaner_shader)
2021 return;
2022
2023 if (ring->xcp_id == AMDGPU_XCP_NO_PARTITION)
2024 idx = 0;
2025 else
2026 idx = ring->xcp_id;
2027
2028 if (idx >= MAX_XCP)
2029 return;
2030
2031 mutex_lock(&adev->enforce_isolation_mutex);
2032 if (adev->enforce_isolation[idx]) {
2033 if (adev->kfd.init_complete)
2034 amdgpu_gfx_kfd_sch_ctrl(adev, idx, true);
2035 }
2036 mutex_unlock(&adev->enforce_isolation_mutex);
2037 }
2038
2039 /*
2040 * debugfs for to enable/disable gfx job submission to specific core.
2041 */
2042 #if defined(CONFIG_DEBUG_FS)
amdgpu_debugfs_gfx_sched_mask_set(void * data,u64 val)2043 static int amdgpu_debugfs_gfx_sched_mask_set(void *data, u64 val)
2044 {
2045 struct amdgpu_device *adev = (struct amdgpu_device *)data;
2046 u32 i;
2047 u64 mask = 0;
2048 struct amdgpu_ring *ring;
2049
2050 if (!adev)
2051 return -ENODEV;
2052
2053 mask = (1 << adev->gfx.num_gfx_rings) - 1;
2054 if ((val & mask) == 0)
2055 return -EINVAL;
2056
2057 for (i = 0; i < adev->gfx.num_gfx_rings; ++i) {
2058 ring = &adev->gfx.gfx_ring[i];
2059 if (val & (1 << i))
2060 ring->sched.ready = true;
2061 else
2062 ring->sched.ready = false;
2063 }
2064 /* publish sched.ready flag update effective immediately across smp */
2065 smp_rmb();
2066 return 0;
2067 }
2068
amdgpu_debugfs_gfx_sched_mask_get(void * data,u64 * val)2069 static int amdgpu_debugfs_gfx_sched_mask_get(void *data, u64 *val)
2070 {
2071 struct amdgpu_device *adev = (struct amdgpu_device *)data;
2072 u32 i;
2073 u64 mask = 0;
2074 struct amdgpu_ring *ring;
2075
2076 if (!adev)
2077 return -ENODEV;
2078 for (i = 0; i < adev->gfx.num_gfx_rings; ++i) {
2079 ring = &adev->gfx.gfx_ring[i];
2080 if (ring->sched.ready)
2081 mask |= 1 << i;
2082 }
2083
2084 *val = mask;
2085 return 0;
2086 }
2087
2088 DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gfx_sched_mask_fops,
2089 amdgpu_debugfs_gfx_sched_mask_get,
2090 amdgpu_debugfs_gfx_sched_mask_set, "%llx\n");
2091
2092 #endif
2093
amdgpu_debugfs_gfx_sched_mask_init(struct amdgpu_device * adev)2094 void amdgpu_debugfs_gfx_sched_mask_init(struct amdgpu_device *adev)
2095 {
2096 #if defined(CONFIG_DEBUG_FS)
2097 struct drm_minor *minor = adev_to_drm(adev)->primary;
2098 struct dentry *root = minor->debugfs_root;
2099 char name[32];
2100
2101 if (!(adev->gfx.num_gfx_rings > 1))
2102 return;
2103 sprintf(name, "amdgpu_gfx_sched_mask");
2104 debugfs_create_file(name, 0600, root, adev,
2105 &amdgpu_debugfs_gfx_sched_mask_fops);
2106 #endif
2107 }
2108
2109 /*
2110 * debugfs for to enable/disable compute job submission to specific core.
2111 */
2112 #if defined(CONFIG_DEBUG_FS)
amdgpu_debugfs_compute_sched_mask_set(void * data,u64 val)2113 static int amdgpu_debugfs_compute_sched_mask_set(void *data, u64 val)
2114 {
2115 struct amdgpu_device *adev = (struct amdgpu_device *)data;
2116 u32 i;
2117 u64 mask = 0;
2118 struct amdgpu_ring *ring;
2119
2120 if (!adev)
2121 return -ENODEV;
2122
2123 mask = (1 << adev->gfx.num_compute_rings) - 1;
2124 if ((val & mask) == 0)
2125 return -EINVAL;
2126
2127 for (i = 0; i < adev->gfx.num_compute_rings; ++i) {
2128 ring = &adev->gfx.compute_ring[i];
2129 if (val & (1 << i))
2130 ring->sched.ready = true;
2131 else
2132 ring->sched.ready = false;
2133 }
2134
2135 /* publish sched.ready flag update effective immediately across smp */
2136 smp_rmb();
2137 return 0;
2138 }
2139
amdgpu_debugfs_compute_sched_mask_get(void * data,u64 * val)2140 static int amdgpu_debugfs_compute_sched_mask_get(void *data, u64 *val)
2141 {
2142 struct amdgpu_device *adev = (struct amdgpu_device *)data;
2143 u32 i;
2144 u64 mask = 0;
2145 struct amdgpu_ring *ring;
2146
2147 if (!adev)
2148 return -ENODEV;
2149 for (i = 0; i < adev->gfx.num_compute_rings; ++i) {
2150 ring = &adev->gfx.compute_ring[i];
2151 if (ring->sched.ready)
2152 mask |= 1 << i;
2153 }
2154
2155 *val = mask;
2156 return 0;
2157 }
2158
2159 DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_compute_sched_mask_fops,
2160 amdgpu_debugfs_compute_sched_mask_get,
2161 amdgpu_debugfs_compute_sched_mask_set, "%llx\n");
2162
2163 #endif
2164
amdgpu_debugfs_compute_sched_mask_init(struct amdgpu_device * adev)2165 void amdgpu_debugfs_compute_sched_mask_init(struct amdgpu_device *adev)
2166 {
2167 #if defined(CONFIG_DEBUG_FS)
2168 struct drm_minor *minor = adev_to_drm(adev)->primary;
2169 struct dentry *root = minor->debugfs_root;
2170 char name[32];
2171
2172 if (!(adev->gfx.num_compute_rings > 1))
2173 return;
2174 sprintf(name, "amdgpu_compute_sched_mask");
2175 debugfs_create_file(name, 0600, root, adev,
2176 &amdgpu_debugfs_compute_sched_mask_fops);
2177 #endif
2178 }
2179