2 * Copyright 2014 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
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:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
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.
26 #include <linux/firmware.h>
28 #include "amdgpu_gfx.h"
29 #include "amdgpu_rlc.h"
30 #include "amdgpu_ras.h"
32 /* delay 0.1 second to enable gfx off feature */
33 #define GFX_OFF_DELAY_ENABLE msecs_to_jiffies(100)
35 #define GFX_OFF_NO_DELAY 0
38 * GPU GFX IP block helpers function.
41 int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec,
46 bit += mec * adev->gfx.mec.num_pipe_per_mec
47 * adev->gfx.mec.num_queue_per_pipe;
48 bit += pipe * adev->gfx.mec.num_queue_per_pipe;
54 void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit,
55 int *mec, int *pipe, int *queue)
57 *queue = bit % adev->gfx.mec.num_queue_per_pipe;
58 *pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
59 % adev->gfx.mec.num_pipe_per_mec;
60 *mec = (bit / adev->gfx.mec.num_queue_per_pipe)
61 / adev->gfx.mec.num_pipe_per_mec;
65 bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
66 int xcc_id, int mec, int pipe, int queue)
68 return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue),
69 adev->gfx.mec_bitmap[xcc_id].queue_bitmap);
72 int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev,
73 int me, int pipe, int queue)
77 bit += me * adev->gfx.me.num_pipe_per_me
78 * adev->gfx.me.num_queue_per_pipe;
79 bit += pipe * adev->gfx.me.num_queue_per_pipe;
85 void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit,
86 int *me, int *pipe, int *queue)
88 *queue = bit % adev->gfx.me.num_queue_per_pipe;
89 *pipe = (bit / adev->gfx.me.num_queue_per_pipe)
90 % adev->gfx.me.num_pipe_per_me;
91 *me = (bit / adev->gfx.me.num_queue_per_pipe)
92 / adev->gfx.me.num_pipe_per_me;
95 bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev,
96 int me, int pipe, int queue)
98 return test_bit(amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue),
99 adev->gfx.me.queue_bitmap);
103 * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
105 * @mask: array in which the per-shader array disable masks will be stored
106 * @max_se: number of SEs
107 * @max_sh: number of SHs
109 * The bitmask of CUs to be disabled in the shader array determined by se and
110 * sh is stored in mask[se * max_sh + sh].
112 void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
117 memset(mask, 0, sizeof(*mask) * max_se * max_sh);
119 if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
122 p = amdgpu_disable_cu;
125 int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
127 DRM_ERROR("amdgpu: could not parse disable_cu\n");
131 if (se < max_se && sh < max_sh && cu < 16) {
132 DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
133 mask[se * max_sh + sh] |= 1u << cu;
135 DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
139 next = strchr(p, ',');
146 static bool amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device *adev)
148 return amdgpu_async_gfx_ring && adev->gfx.me.num_pipe_per_me > 1;
151 static bool amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device *adev)
153 if (amdgpu_compute_multipipe != -1) {
154 DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
155 amdgpu_compute_multipipe);
156 return amdgpu_compute_multipipe == 1;
159 if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))
162 /* FIXME: spreading the queues across pipes causes perf regressions
163 * on POLARIS11 compute workloads */
164 if (adev->asic_type == CHIP_POLARIS11)
167 return adev->gfx.mec.num_mec > 1;
170 bool amdgpu_gfx_is_high_priority_graphics_queue(struct amdgpu_device *adev,
171 struct amdgpu_ring *ring)
173 int queue = ring->queue;
174 int pipe = ring->pipe;
176 /* Policy: use pipe1 queue0 as high priority graphics queue if we
177 * have more than one gfx pipe.
179 if (amdgpu_gfx_is_graphics_multipipe_capable(adev) &&
180 adev->gfx.num_gfx_rings > 1 && pipe == 1 && queue == 0) {
184 bit = amdgpu_gfx_me_queue_to_bit(adev, me, pipe, queue);
185 if (ring == &adev->gfx.gfx_ring[bit])
192 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
193 struct amdgpu_ring *ring)
195 /* Policy: use 1st queue as high priority compute queue if we
196 * have more than one compute queue.
198 if (adev->gfx.num_compute_rings > 1 &&
199 ring == &adev->gfx.compute_ring[0])
205 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
207 int i, j, queue, pipe;
208 bool multipipe_policy = amdgpu_gfx_is_compute_multipipe_capable(adev);
209 int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec *
210 adev->gfx.mec.num_queue_per_pipe,
211 adev->gfx.num_compute_rings);
212 int num_xcd = (adev->gfx.num_xcd > 1) ? adev->gfx.num_xcd : 1;
214 if (multipipe_policy) {
215 /* policy: make queues evenly cross all pipes on MEC1 only
216 * for multiple xcc, just use the original policy for simplicity */
217 for (j = 0; j < num_xcd; j++) {
218 for (i = 0; i < max_queues_per_mec; i++) {
219 pipe = i % adev->gfx.mec.num_pipe_per_mec;
220 queue = (i / adev->gfx.mec.num_pipe_per_mec) %
221 adev->gfx.mec.num_queue_per_pipe;
223 set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue,
224 adev->gfx.mec_bitmap[j].queue_bitmap);
228 /* policy: amdgpu owns all queues in the given pipe */
229 for (j = 0; j < num_xcd; j++) {
230 for (i = 0; i < max_queues_per_mec; ++i)
231 set_bit(i, adev->gfx.mec_bitmap[j].queue_bitmap);
235 for (j = 0; j < num_xcd; j++) {
236 dev_dbg(adev->dev, "mec queue bitmap weight=%d\n",
237 bitmap_weight(adev->gfx.mec_bitmap[j].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES));
241 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev)
244 bool multipipe_policy = amdgpu_gfx_is_graphics_multipipe_capable(adev);
245 int max_queues_per_me = adev->gfx.me.num_pipe_per_me *
246 adev->gfx.me.num_queue_per_pipe;
248 if (multipipe_policy) {
249 /* policy: amdgpu owns the first queue per pipe at this stage
250 * will extend to mulitple queues per pipe later */
251 for (i = 0; i < max_queues_per_me; i++) {
252 pipe = i % adev->gfx.me.num_pipe_per_me;
253 queue = (i / adev->gfx.me.num_pipe_per_me) %
254 adev->gfx.me.num_queue_per_pipe;
256 set_bit(pipe * adev->gfx.me.num_queue_per_pipe + queue,
257 adev->gfx.me.queue_bitmap);
260 for (i = 0; i < max_queues_per_me; ++i)
261 set_bit(i, adev->gfx.me.queue_bitmap);
264 /* update the number of active graphics rings */
265 adev->gfx.num_gfx_rings =
266 bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
269 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
270 struct amdgpu_ring *ring, int xcc_id)
273 int mec, pipe, queue;
275 queue_bit = adev->gfx.mec.num_mec
276 * adev->gfx.mec.num_pipe_per_mec
277 * adev->gfx.mec.num_queue_per_pipe;
279 while (--queue_bit >= 0) {
280 if (test_bit(queue_bit, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
283 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
286 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
287 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
288 * only can be issued on queue 0.
290 if ((mec == 1 && pipe > 1) || queue != 0)
300 dev_err(adev->dev, "Failed to find a queue for KIQ\n");
304 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
305 struct amdgpu_ring *ring,
306 struct amdgpu_irq_src *irq, int xcc_id)
308 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
311 spin_lock_init(&kiq->ring_lock);
314 ring->ring_obj = NULL;
315 ring->use_doorbell = true;
316 ring->doorbell_index = adev->doorbell_index.kiq;
317 ring->xcc_id = xcc_id;
318 ring->vm_hub = AMDGPU_GFXHUB_0;
320 ring->doorbell_index = adev->doorbell_index.xcc1_kiq_start +
323 ring->doorbell_index = adev->doorbell_index.kiq;
325 r = amdgpu_gfx_kiq_acquire(adev, ring, xcc_id);
329 ring->eop_gpu_addr = kiq->eop_gpu_addr;
330 ring->no_scheduler = true;
331 sprintf(ring->name, "kiq_%d.%d.%d.%d", xcc_id, ring->me, ring->pipe, ring->queue);
332 r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
333 AMDGPU_RING_PRIO_DEFAULT, NULL);
335 dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
340 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
342 amdgpu_ring_fini(ring);
345 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev, int xcc_id)
347 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
349 amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
352 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
353 unsigned hpd_size, int xcc_id)
357 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
359 r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
360 AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
361 &kiq->eop_gpu_addr, (void **)&hpd);
363 dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
367 memset(hpd, 0, hpd_size);
369 r = amdgpu_bo_reserve(kiq->eop_obj, true);
370 if (unlikely(r != 0))
371 dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
372 amdgpu_bo_kunmap(kiq->eop_obj);
373 amdgpu_bo_unreserve(kiq->eop_obj);
378 /* create MQD for each compute/gfx queue */
379 int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
380 unsigned mqd_size, int xcc_id)
383 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
384 struct amdgpu_ring *ring = &kiq->ring;
386 /* create MQD for KIQ */
387 if (!adev->enable_mes_kiq && !ring->mqd_obj) {
388 /* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
389 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
390 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
391 * KIQ MQD no matter SRIOV or Bare-metal
393 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
394 AMDGPU_GEM_DOMAIN_VRAM |
395 AMDGPU_GEM_DOMAIN_GTT,
400 dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
404 /* prepare MQD backup */
405 kiq->mqd_backup = kmalloc(mqd_size, GFP_KERNEL);
406 if (!kiq->mqd_backup)
407 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
410 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
411 /* create MQD for each KGQ */
412 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
413 ring = &adev->gfx.gfx_ring[i];
414 if (!ring->mqd_obj) {
415 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
416 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
417 &ring->mqd_gpu_addr, &ring->mqd_ptr);
419 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
423 ring->mqd_size = mqd_size;
424 /* prepare MQD backup */
425 adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
426 if (!adev->gfx.me.mqd_backup[i])
427 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
432 /* create MQD for each KCQ */
433 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
434 ring = &adev->gfx.compute_ring[i + xcc_id * adev->gfx.num_compute_rings];
435 if (!ring->mqd_obj) {
436 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
437 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
438 &ring->mqd_gpu_addr, &ring->mqd_ptr);
440 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
444 ring->mqd_size = mqd_size;
445 /* prepare MQD backup */
446 adev->gfx.mec.mqd_backup[i + xcc_id * adev->gfx.num_compute_rings] = kmalloc(mqd_size, GFP_KERNEL);
447 if (!adev->gfx.mec.mqd_backup[i])
448 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
455 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev, int xcc_id)
457 struct amdgpu_ring *ring = NULL;
459 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
461 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
462 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
463 ring = &adev->gfx.gfx_ring[i];
464 kfree(adev->gfx.me.mqd_backup[i]);
465 amdgpu_bo_free_kernel(&ring->mqd_obj,
471 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
472 j = i + xcc_id * adev->gfx.num_compute_rings;
473 ring = &adev->gfx.compute_ring[i];
474 kfree(adev->gfx.mec.mqd_backup[i]);
475 amdgpu_bo_free_kernel(&ring->mqd_obj,
481 kfree(kiq->mqd_backup);
482 kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
483 amdgpu_bo_free_kernel(&ring->mqd_obj,
488 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
490 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
491 struct amdgpu_ring *kiq_ring = &kiq->ring;
495 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
498 spin_lock(&kiq->ring_lock);
499 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
500 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
501 adev->gfx.num_compute_rings)) {
502 spin_unlock(&kiq->ring_lock);
506 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
507 j = i + xcc_id * adev->gfx.num_compute_rings;
508 kiq->pmf->kiq_unmap_queues(kiq_ring,
509 &adev->gfx.compute_ring[i],
514 if (adev->gfx.kiq[0].ring.sched.ready && !adev->job_hang)
515 r = amdgpu_ring_test_helper(kiq_ring);
516 spin_unlock(&kiq->ring_lock);
521 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
524 int mec, pipe, queue;
525 int set_resource_bit = 0;
527 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
529 set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
531 return set_resource_bit;
534 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
536 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
537 struct amdgpu_ring *kiq_ring = &kiq->ring;
538 uint64_t queue_mask = 0;
541 if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
544 for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
545 if (!test_bit(i, adev->gfx.mec_bitmap[xcc_id].queue_bitmap))
548 /* This situation may be hit in the future if a new HW
549 * generation exposes more than 64 queues. If so, the
550 * definition of queue_mask needs updating */
551 if (WARN_ON(i > (sizeof(queue_mask)*8))) {
552 DRM_ERROR("Invalid KCQ enabled: %d\n", i);
556 queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
559 DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
561 spin_lock(&kiq->ring_lock);
562 /* No need to map kcq on the slave */
563 if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
564 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
565 adev->gfx.num_compute_rings +
566 kiq->pmf->set_resources_size);
568 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
569 spin_unlock(&adev->gfx.kiq[0].ring_lock);
573 if (adev->enable_mes)
576 kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
577 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
578 j = i + xcc_id * adev->gfx.num_compute_rings;
579 kiq->pmf->kiq_map_queues(kiq_ring,
580 &adev->gfx.compute_ring[i]);
584 r = amdgpu_ring_test_helper(kiq_ring);
585 spin_unlock(&kiq->ring_lock);
587 DRM_ERROR("KCQ enable failed\n");
592 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
594 * @adev: amdgpu_device pointer
595 * @bool enable true: enable gfx off feature, false: disable gfx off feature
597 * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
598 * 2. other client can send request to disable gfx off feature, the request should be honored.
599 * 3. other client can cancel their request of disable gfx off feature
600 * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
603 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
605 unsigned long delay = GFX_OFF_DELAY_ENABLE;
607 if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
610 mutex_lock(&adev->gfx.gfx_off_mutex);
613 /* If the count is already 0, it means there's an imbalance bug somewhere.
614 * Note that the bug may be in a different caller than the one which triggers the
617 if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
620 adev->gfx.gfx_off_req_count--;
622 if (adev->gfx.gfx_off_req_count == 0 &&
623 !adev->gfx.gfx_off_state) {
624 /* If going to s2idle, no need to wait */
626 if (!amdgpu_dpm_set_powergating_by_smu(adev,
627 AMD_IP_BLOCK_TYPE_GFX, true))
628 adev->gfx.gfx_off_state = true;
630 schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
635 if (adev->gfx.gfx_off_req_count == 0) {
636 cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
638 if (adev->gfx.gfx_off_state &&
639 !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
640 adev->gfx.gfx_off_state = false;
642 if (adev->gfx.funcs->init_spm_golden) {
644 "GFXOFF is disabled, re-init SPM golden settings\n");
645 amdgpu_gfx_init_spm_golden(adev);
650 adev->gfx.gfx_off_req_count++;
654 mutex_unlock(&adev->gfx.gfx_off_mutex);
657 int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
661 mutex_lock(&adev->gfx.gfx_off_mutex);
663 r = amdgpu_dpm_set_residency_gfxoff(adev, value);
665 mutex_unlock(&adev->gfx.gfx_off_mutex);
670 int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
674 mutex_lock(&adev->gfx.gfx_off_mutex);
676 r = amdgpu_dpm_get_residency_gfxoff(adev, value);
678 mutex_unlock(&adev->gfx.gfx_off_mutex);
683 int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
687 mutex_lock(&adev->gfx.gfx_off_mutex);
689 r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
691 mutex_unlock(&adev->gfx.gfx_off_mutex);
696 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
701 mutex_lock(&adev->gfx.gfx_off_mutex);
703 r = amdgpu_dpm_get_status_gfxoff(adev, value);
705 mutex_unlock(&adev->gfx.gfx_off_mutex);
710 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
714 if (amdgpu_ras_is_supported(adev, ras_block->block)) {
715 if (!amdgpu_persistent_edc_harvesting_supported(adev))
716 amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
718 r = amdgpu_ras_block_late_init(adev, ras_block);
722 r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
726 amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
731 amdgpu_ras_block_late_fini(adev, ras_block);
735 int amdgpu_gfx_ras_sw_init(struct amdgpu_device *adev)
738 struct amdgpu_gfx_ras *ras = NULL;
740 /* adev->gfx.ras is NULL, which means gfx does not
741 * support ras function, then do nothing here.
748 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
750 dev_err(adev->dev, "Failed to register gfx ras block!\n");
754 strcpy(ras->ras_block.ras_comm.name, "gfx");
755 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__GFX;
756 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
757 adev->gfx.ras_if = &ras->ras_block.ras_comm;
759 /* If not define special ras_late_init function, use gfx default ras_late_init */
760 if (!ras->ras_block.ras_late_init)
761 ras->ras_block.ras_late_init = amdgpu_gfx_ras_late_init;
763 /* If not defined special ras_cb function, use default ras_cb */
764 if (!ras->ras_block.ras_cb)
765 ras->ras_block.ras_cb = amdgpu_gfx_process_ras_data_cb;
770 int amdgpu_gfx_poison_consumption_handler(struct amdgpu_device *adev,
771 struct amdgpu_iv_entry *entry)
773 if (adev->gfx.ras && adev->gfx.ras->poison_consumption_handler)
774 return adev->gfx.ras->poison_consumption_handler(adev, entry);
779 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
781 struct amdgpu_iv_entry *entry)
783 /* TODO ue will trigger an interrupt.
785 * When “Full RAS” is enabled, the per-IP interrupt sources should
786 * be disabled and the driver should only look for the aggregated
787 * interrupt via sync flood
789 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
790 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
791 if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
792 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
793 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
794 amdgpu_ras_reset_gpu(adev);
796 return AMDGPU_RAS_SUCCESS;
799 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
800 struct amdgpu_irq_src *source,
801 struct amdgpu_iv_entry *entry)
803 struct ras_common_if *ras_if = adev->gfx.ras_if;
804 struct ras_dispatch_if ih_data = {
811 ih_data.head = *ras_if;
813 DRM_ERROR("CP ECC ERROR IRQ\n");
814 amdgpu_ras_interrupt_dispatch(adev, &ih_data);
818 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
820 signed long r, cnt = 0;
822 uint32_t seq, reg_val_offs = 0, value = 0;
823 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
824 struct amdgpu_ring *ring = &kiq->ring;
826 if (amdgpu_device_skip_hw_access(adev))
829 if (adev->mes.ring.sched.ready)
830 return amdgpu_mes_rreg(adev, reg);
832 BUG_ON(!ring->funcs->emit_rreg);
834 spin_lock_irqsave(&kiq->ring_lock, flags);
835 if (amdgpu_device_wb_get(adev, ®_val_offs)) {
836 pr_err("critical bug! too many kiq readers\n");
839 amdgpu_ring_alloc(ring, 32);
840 amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
841 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
845 amdgpu_ring_commit(ring);
846 spin_unlock_irqrestore(&kiq->ring_lock, flags);
848 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
850 /* don't wait anymore for gpu reset case because this way may
851 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
852 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
853 * never return if we keep waiting in virt_kiq_rreg, which cause
854 * gpu_recover() hang there.
856 * also don't wait anymore for IRQ context
858 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
859 goto failed_kiq_read;
862 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
863 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
864 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
867 if (cnt > MAX_KIQ_REG_TRY)
868 goto failed_kiq_read;
871 value = adev->wb.wb[reg_val_offs];
872 amdgpu_device_wb_free(adev, reg_val_offs);
876 amdgpu_ring_undo(ring);
878 spin_unlock_irqrestore(&kiq->ring_lock, flags);
881 amdgpu_device_wb_free(adev, reg_val_offs);
882 dev_err(adev->dev, "failed to read reg:%x\n", reg);
886 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
888 signed long r, cnt = 0;
891 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
892 struct amdgpu_ring *ring = &kiq->ring;
894 BUG_ON(!ring->funcs->emit_wreg);
896 if (amdgpu_device_skip_hw_access(adev))
899 if (adev->mes.ring.sched.ready) {
900 amdgpu_mes_wreg(adev, reg, v);
904 spin_lock_irqsave(&kiq->ring_lock, flags);
905 amdgpu_ring_alloc(ring, 32);
906 amdgpu_ring_emit_wreg(ring, reg, v);
907 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
911 amdgpu_ring_commit(ring);
912 spin_unlock_irqrestore(&kiq->ring_lock, flags);
914 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
916 /* don't wait anymore for gpu reset case because this way may
917 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
918 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
919 * never return if we keep waiting in virt_kiq_rreg, which cause
920 * gpu_recover() hang there.
922 * also don't wait anymore for IRQ context
924 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
925 goto failed_kiq_write;
928 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
930 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
931 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
934 if (cnt > MAX_KIQ_REG_TRY)
935 goto failed_kiq_write;
940 amdgpu_ring_undo(ring);
941 spin_unlock_irqrestore(&kiq->ring_lock, flags);
943 dev_err(adev->dev, "failed to write reg:%x\n", reg);
946 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
948 if (amdgpu_num_kcq == -1) {
950 } else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
951 dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
954 return amdgpu_num_kcq;
957 void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
960 const struct gfx_firmware_header_v1_0 *cp_hdr;
961 const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
962 struct amdgpu_firmware_info *info = NULL;
963 const struct firmware *ucode_fw;
964 unsigned int fw_size;
967 case AMDGPU_UCODE_ID_CP_PFP:
968 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
969 adev->gfx.pfp_fw->data;
970 adev->gfx.pfp_fw_version =
971 le32_to_cpu(cp_hdr->header.ucode_version);
972 adev->gfx.pfp_feature_version =
973 le32_to_cpu(cp_hdr->ucode_feature_version);
974 ucode_fw = adev->gfx.pfp_fw;
975 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
977 case AMDGPU_UCODE_ID_CP_RS64_PFP:
978 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
979 adev->gfx.pfp_fw->data;
980 adev->gfx.pfp_fw_version =
981 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
982 adev->gfx.pfp_feature_version =
983 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
984 ucode_fw = adev->gfx.pfp_fw;
985 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
987 case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
988 case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
989 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
990 adev->gfx.pfp_fw->data;
991 ucode_fw = adev->gfx.pfp_fw;
992 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
994 case AMDGPU_UCODE_ID_CP_ME:
995 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
996 adev->gfx.me_fw->data;
997 adev->gfx.me_fw_version =
998 le32_to_cpu(cp_hdr->header.ucode_version);
999 adev->gfx.me_feature_version =
1000 le32_to_cpu(cp_hdr->ucode_feature_version);
1001 ucode_fw = adev->gfx.me_fw;
1002 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1004 case AMDGPU_UCODE_ID_CP_RS64_ME:
1005 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1006 adev->gfx.me_fw->data;
1007 adev->gfx.me_fw_version =
1008 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1009 adev->gfx.me_feature_version =
1010 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1011 ucode_fw = adev->gfx.me_fw;
1012 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1014 case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
1015 case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
1016 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1017 adev->gfx.me_fw->data;
1018 ucode_fw = adev->gfx.me_fw;
1019 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1021 case AMDGPU_UCODE_ID_CP_CE:
1022 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1023 adev->gfx.ce_fw->data;
1024 adev->gfx.ce_fw_version =
1025 le32_to_cpu(cp_hdr->header.ucode_version);
1026 adev->gfx.ce_feature_version =
1027 le32_to_cpu(cp_hdr->ucode_feature_version);
1028 ucode_fw = adev->gfx.ce_fw;
1029 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1031 case AMDGPU_UCODE_ID_CP_MEC1:
1032 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1033 adev->gfx.mec_fw->data;
1034 adev->gfx.mec_fw_version =
1035 le32_to_cpu(cp_hdr->header.ucode_version);
1036 adev->gfx.mec_feature_version =
1037 le32_to_cpu(cp_hdr->ucode_feature_version);
1038 ucode_fw = adev->gfx.mec_fw;
1039 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1040 le32_to_cpu(cp_hdr->jt_size) * 4;
1042 case AMDGPU_UCODE_ID_CP_MEC1_JT:
1043 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1044 adev->gfx.mec_fw->data;
1045 ucode_fw = adev->gfx.mec_fw;
1046 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1048 case AMDGPU_UCODE_ID_CP_MEC2:
1049 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1050 adev->gfx.mec2_fw->data;
1051 adev->gfx.mec2_fw_version =
1052 le32_to_cpu(cp_hdr->header.ucode_version);
1053 adev->gfx.mec2_feature_version =
1054 le32_to_cpu(cp_hdr->ucode_feature_version);
1055 ucode_fw = adev->gfx.mec2_fw;
1056 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1057 le32_to_cpu(cp_hdr->jt_size) * 4;
1059 case AMDGPU_UCODE_ID_CP_MEC2_JT:
1060 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1061 adev->gfx.mec2_fw->data;
1062 ucode_fw = adev->gfx.mec2_fw;
1063 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1065 case AMDGPU_UCODE_ID_CP_RS64_MEC:
1066 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1067 adev->gfx.mec_fw->data;
1068 adev->gfx.mec_fw_version =
1069 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1070 adev->gfx.mec_feature_version =
1071 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1072 ucode_fw = adev->gfx.mec_fw;
1073 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1075 case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
1076 case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
1077 case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
1078 case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
1079 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1080 adev->gfx.mec_fw->data;
1081 ucode_fw = adev->gfx.mec_fw;
1082 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1088 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1089 info = &adev->firmware.ucode[ucode_id];
1090 info->ucode_id = ucode_id;
1091 info->fw = ucode_fw;
1092 adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE);
1096 bool amdgpu_gfx_is_master_xcc(struct amdgpu_device *adev, int xcc_id)
1098 return !(xcc_id % (adev->gfx.num_xcc_per_xcp ?
1099 adev->gfx.num_xcc_per_xcp : 1));