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)
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[0].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)
308 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
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->vm_hub = AMDGPU_GFXHUB_0;
319 r = amdgpu_gfx_kiq_acquire(adev, ring);
323 ring->eop_gpu_addr = kiq->eop_gpu_addr;
324 ring->no_scheduler = true;
325 sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
326 r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0,
327 AMDGPU_RING_PRIO_DEFAULT, NULL);
329 dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
334 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring)
336 amdgpu_ring_fini(ring);
339 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
341 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
343 amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
346 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
351 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
353 r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
354 AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
355 &kiq->eop_gpu_addr, (void **)&hpd);
357 dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
361 memset(hpd, 0, hpd_size);
363 r = amdgpu_bo_reserve(kiq->eop_obj, true);
364 if (unlikely(r != 0))
365 dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
366 amdgpu_bo_kunmap(kiq->eop_obj);
367 amdgpu_bo_unreserve(kiq->eop_obj);
372 /* create MQD for each compute/gfx queue */
373 int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
376 struct amdgpu_ring *ring = NULL;
379 /* create MQD for KIQ */
380 ring = &adev->gfx.kiq[0].ring;
381 if (!adev->enable_mes_kiq && !ring->mqd_obj) {
382 /* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
383 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
384 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
385 * KIQ MQD no matter SRIOV or Bare-metal
387 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
388 AMDGPU_GEM_DOMAIN_VRAM |
389 AMDGPU_GEM_DOMAIN_GTT,
394 dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
398 /* prepare MQD backup */
399 adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
400 if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
401 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
404 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
405 /* create MQD for each KGQ */
406 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
407 ring = &adev->gfx.gfx_ring[i];
408 if (!ring->mqd_obj) {
409 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
410 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
411 &ring->mqd_gpu_addr, &ring->mqd_ptr);
413 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
417 /* prepare MQD backup */
418 adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
419 if (!adev->gfx.me.mqd_backup[i])
420 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
425 /* create MQD for each KCQ */
426 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
427 ring = &adev->gfx.compute_ring[i];
428 if (!ring->mqd_obj) {
429 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
430 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
431 &ring->mqd_gpu_addr, &ring->mqd_ptr);
433 dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
437 /* prepare MQD backup */
438 adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
439 if (!adev->gfx.mec.mqd_backup[i])
440 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
447 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev)
449 struct amdgpu_ring *ring = NULL;
452 if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) {
453 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
454 ring = &adev->gfx.gfx_ring[i];
455 kfree(adev->gfx.me.mqd_backup[i]);
456 amdgpu_bo_free_kernel(&ring->mqd_obj,
462 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
463 ring = &adev->gfx.compute_ring[i];
464 kfree(adev->gfx.mec.mqd_backup[i]);
465 amdgpu_bo_free_kernel(&ring->mqd_obj,
470 ring = &adev->gfx.kiq[0].ring;
471 kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
472 amdgpu_bo_free_kernel(&ring->mqd_obj,
477 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev)
479 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
480 struct amdgpu_ring *kiq_ring = &kiq->ring;
483 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
486 spin_lock(&adev->gfx.kiq[0].ring_lock);
487 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
488 adev->gfx.num_compute_rings)) {
489 spin_unlock(&adev->gfx.kiq[0].ring_lock);
493 for (i = 0; i < adev->gfx.num_compute_rings; i++)
494 kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i],
497 if (adev->gfx.kiq[0].ring.sched.ready && !adev->job_hang)
498 r = amdgpu_ring_test_helper(kiq_ring);
499 spin_unlock(&adev->gfx.kiq[0].ring_lock);
504 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
507 int mec, pipe, queue;
508 int set_resource_bit = 0;
510 amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue);
512 set_resource_bit = mec * 4 * 8 + pipe * 8 + queue;
514 return set_resource_bit;
517 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev)
519 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
520 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
521 uint64_t queue_mask = 0;
524 if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources)
527 for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
528 if (!test_bit(i, adev->gfx.mec_bitmap[0].queue_bitmap))
531 /* This situation may be hit in the future if a new HW
532 * generation exposes more than 64 queues. If so, the
533 * definition of queue_mask needs updating */
534 if (WARN_ON(i > (sizeof(queue_mask)*8))) {
535 DRM_ERROR("Invalid KCQ enabled: %d\n", i);
539 queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i));
542 DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe,
544 spin_lock(&adev->gfx.kiq[0].ring_lock);
545 r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
546 adev->gfx.num_compute_rings +
547 kiq->pmf->set_resources_size);
549 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
550 spin_unlock(&adev->gfx.kiq[0].ring_lock);
554 if (adev->enable_mes)
557 kiq->pmf->kiq_set_resources(kiq_ring, queue_mask);
558 for (i = 0; i < adev->gfx.num_compute_rings; i++)
559 kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]);
561 r = amdgpu_ring_test_helper(kiq_ring);
562 spin_unlock(&adev->gfx.kiq[0].ring_lock);
564 DRM_ERROR("KCQ enable failed\n");
569 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
571 * @adev: amdgpu_device pointer
572 * @bool enable true: enable gfx off feature, false: disable gfx off feature
574 * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
575 * 2. other client can send request to disable gfx off feature, the request should be honored.
576 * 3. other client can cancel their request of disable gfx off feature
577 * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
580 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
582 unsigned long delay = GFX_OFF_DELAY_ENABLE;
584 if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
587 mutex_lock(&adev->gfx.gfx_off_mutex);
590 /* If the count is already 0, it means there's an imbalance bug somewhere.
591 * Note that the bug may be in a different caller than the one which triggers the
594 if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
597 adev->gfx.gfx_off_req_count--;
599 if (adev->gfx.gfx_off_req_count == 0 &&
600 !adev->gfx.gfx_off_state) {
601 /* If going to s2idle, no need to wait */
603 if (!amdgpu_dpm_set_powergating_by_smu(adev,
604 AMD_IP_BLOCK_TYPE_GFX, true))
605 adev->gfx.gfx_off_state = true;
607 schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
612 if (adev->gfx.gfx_off_req_count == 0) {
613 cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
615 if (adev->gfx.gfx_off_state &&
616 !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
617 adev->gfx.gfx_off_state = false;
619 if (adev->gfx.funcs->init_spm_golden) {
621 "GFXOFF is disabled, re-init SPM golden settings\n");
622 amdgpu_gfx_init_spm_golden(adev);
627 adev->gfx.gfx_off_req_count++;
631 mutex_unlock(&adev->gfx.gfx_off_mutex);
634 int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
638 mutex_lock(&adev->gfx.gfx_off_mutex);
640 r = amdgpu_dpm_set_residency_gfxoff(adev, value);
642 mutex_unlock(&adev->gfx.gfx_off_mutex);
647 int amdgpu_get_gfx_off_residency(struct amdgpu_device *adev, u32 *value)
651 mutex_lock(&adev->gfx.gfx_off_mutex);
653 r = amdgpu_dpm_get_residency_gfxoff(adev, value);
655 mutex_unlock(&adev->gfx.gfx_off_mutex);
660 int amdgpu_get_gfx_off_entrycount(struct amdgpu_device *adev, u64 *value)
664 mutex_lock(&adev->gfx.gfx_off_mutex);
666 r = amdgpu_dpm_get_entrycount_gfxoff(adev, value);
668 mutex_unlock(&adev->gfx.gfx_off_mutex);
673 int amdgpu_get_gfx_off_status(struct amdgpu_device *adev, uint32_t *value)
678 mutex_lock(&adev->gfx.gfx_off_mutex);
680 r = amdgpu_dpm_get_status_gfxoff(adev, value);
682 mutex_unlock(&adev->gfx.gfx_off_mutex);
687 int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
691 if (amdgpu_ras_is_supported(adev, ras_block->block)) {
692 if (!amdgpu_persistent_edc_harvesting_supported(adev))
693 amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX);
695 r = amdgpu_ras_block_late_init(adev, ras_block);
699 r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0);
703 amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
708 amdgpu_ras_block_late_fini(adev, ras_block);
712 int amdgpu_gfx_ras_sw_init(struct amdgpu_device *adev)
715 struct amdgpu_gfx_ras *ras = NULL;
717 /* adev->gfx.ras is NULL, which means gfx does not
718 * support ras function, then do nothing here.
725 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
727 dev_err(adev->dev, "Failed to register gfx ras block!\n");
731 strcpy(ras->ras_block.ras_comm.name, "gfx");
732 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__GFX;
733 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
734 adev->gfx.ras_if = &ras->ras_block.ras_comm;
736 /* If not define special ras_late_init function, use gfx default ras_late_init */
737 if (!ras->ras_block.ras_late_init)
738 ras->ras_block.ras_late_init = amdgpu_gfx_ras_late_init;
740 /* If not defined special ras_cb function, use default ras_cb */
741 if (!ras->ras_block.ras_cb)
742 ras->ras_block.ras_cb = amdgpu_gfx_process_ras_data_cb;
747 int amdgpu_gfx_poison_consumption_handler(struct amdgpu_device *adev,
748 struct amdgpu_iv_entry *entry)
750 if (adev->gfx.ras && adev->gfx.ras->poison_consumption_handler)
751 return adev->gfx.ras->poison_consumption_handler(adev, entry);
756 int amdgpu_gfx_process_ras_data_cb(struct amdgpu_device *adev,
758 struct amdgpu_iv_entry *entry)
760 /* TODO ue will trigger an interrupt.
762 * When “Full RAS” is enabled, the per-IP interrupt sources should
763 * be disabled and the driver should only look for the aggregated
764 * interrupt via sync flood
766 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__GFX)) {
767 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
768 if (adev->gfx.ras && adev->gfx.ras->ras_block.hw_ops &&
769 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count)
770 adev->gfx.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
771 amdgpu_ras_reset_gpu(adev);
773 return AMDGPU_RAS_SUCCESS;
776 int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev,
777 struct amdgpu_irq_src *source,
778 struct amdgpu_iv_entry *entry)
780 struct ras_common_if *ras_if = adev->gfx.ras_if;
781 struct ras_dispatch_if ih_data = {
788 ih_data.head = *ras_if;
790 DRM_ERROR("CP ECC ERROR IRQ\n");
791 amdgpu_ras_interrupt_dispatch(adev, &ih_data);
795 uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg)
797 signed long r, cnt = 0;
799 uint32_t seq, reg_val_offs = 0, value = 0;
800 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
801 struct amdgpu_ring *ring = &kiq->ring;
803 if (amdgpu_device_skip_hw_access(adev))
806 if (adev->mes.ring.sched.ready)
807 return amdgpu_mes_rreg(adev, reg);
809 BUG_ON(!ring->funcs->emit_rreg);
811 spin_lock_irqsave(&kiq->ring_lock, flags);
812 if (amdgpu_device_wb_get(adev, ®_val_offs)) {
813 pr_err("critical bug! too many kiq readers\n");
816 amdgpu_ring_alloc(ring, 32);
817 amdgpu_ring_emit_rreg(ring, reg, reg_val_offs);
818 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
822 amdgpu_ring_commit(ring);
823 spin_unlock_irqrestore(&kiq->ring_lock, flags);
825 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
827 /* don't wait anymore for gpu reset case because this way may
828 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
829 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
830 * never return if we keep waiting in virt_kiq_rreg, which cause
831 * gpu_recover() hang there.
833 * also don't wait anymore for IRQ context
835 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
836 goto failed_kiq_read;
839 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
840 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
841 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
844 if (cnt > MAX_KIQ_REG_TRY)
845 goto failed_kiq_read;
848 value = adev->wb.wb[reg_val_offs];
849 amdgpu_device_wb_free(adev, reg_val_offs);
853 amdgpu_ring_undo(ring);
855 spin_unlock_irqrestore(&kiq->ring_lock, flags);
858 amdgpu_device_wb_free(adev, reg_val_offs);
859 dev_err(adev->dev, "failed to read reg:%x\n", reg);
863 void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
865 signed long r, cnt = 0;
868 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
869 struct amdgpu_ring *ring = &kiq->ring;
871 BUG_ON(!ring->funcs->emit_wreg);
873 if (amdgpu_device_skip_hw_access(adev))
876 if (adev->mes.ring.sched.ready) {
877 amdgpu_mes_wreg(adev, reg, v);
881 spin_lock_irqsave(&kiq->ring_lock, flags);
882 amdgpu_ring_alloc(ring, 32);
883 amdgpu_ring_emit_wreg(ring, reg, v);
884 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
888 amdgpu_ring_commit(ring);
889 spin_unlock_irqrestore(&kiq->ring_lock, flags);
891 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
893 /* don't wait anymore for gpu reset case because this way may
894 * block gpu_recover() routine forever, e.g. this virt_kiq_rreg
895 * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will
896 * never return if we keep waiting in virt_kiq_rreg, which cause
897 * gpu_recover() hang there.
899 * also don't wait anymore for IRQ context
901 if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt()))
902 goto failed_kiq_write;
905 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
907 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
908 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
911 if (cnt > MAX_KIQ_REG_TRY)
912 goto failed_kiq_write;
917 amdgpu_ring_undo(ring);
918 spin_unlock_irqrestore(&kiq->ring_lock, flags);
920 dev_err(adev->dev, "failed to write reg:%x\n", reg);
923 int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev)
925 if (amdgpu_num_kcq == -1) {
927 } else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
928 dev_warn(adev->dev, "set kernel compute queue number to 8 due to invalid parameter provided by user\n");
931 return amdgpu_num_kcq;
934 void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev,
937 const struct gfx_firmware_header_v1_0 *cp_hdr;
938 const struct gfx_firmware_header_v2_0 *cp_hdr_v2_0;
939 struct amdgpu_firmware_info *info = NULL;
940 const struct firmware *ucode_fw;
941 unsigned int fw_size;
944 case AMDGPU_UCODE_ID_CP_PFP:
945 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
946 adev->gfx.pfp_fw->data;
947 adev->gfx.pfp_fw_version =
948 le32_to_cpu(cp_hdr->header.ucode_version);
949 adev->gfx.pfp_feature_version =
950 le32_to_cpu(cp_hdr->ucode_feature_version);
951 ucode_fw = adev->gfx.pfp_fw;
952 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
954 case AMDGPU_UCODE_ID_CP_RS64_PFP:
955 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
956 adev->gfx.pfp_fw->data;
957 adev->gfx.pfp_fw_version =
958 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
959 adev->gfx.pfp_feature_version =
960 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
961 ucode_fw = adev->gfx.pfp_fw;
962 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
964 case AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK:
965 case AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK:
966 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
967 adev->gfx.pfp_fw->data;
968 ucode_fw = adev->gfx.pfp_fw;
969 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
971 case AMDGPU_UCODE_ID_CP_ME:
972 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
973 adev->gfx.me_fw->data;
974 adev->gfx.me_fw_version =
975 le32_to_cpu(cp_hdr->header.ucode_version);
976 adev->gfx.me_feature_version =
977 le32_to_cpu(cp_hdr->ucode_feature_version);
978 ucode_fw = adev->gfx.me_fw;
979 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
981 case AMDGPU_UCODE_ID_CP_RS64_ME:
982 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
983 adev->gfx.me_fw->data;
984 adev->gfx.me_fw_version =
985 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
986 adev->gfx.me_feature_version =
987 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
988 ucode_fw = adev->gfx.me_fw;
989 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
991 case AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK:
992 case AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK:
993 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
994 adev->gfx.me_fw->data;
995 ucode_fw = adev->gfx.me_fw;
996 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
998 case AMDGPU_UCODE_ID_CP_CE:
999 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1000 adev->gfx.ce_fw->data;
1001 adev->gfx.ce_fw_version =
1002 le32_to_cpu(cp_hdr->header.ucode_version);
1003 adev->gfx.ce_feature_version =
1004 le32_to_cpu(cp_hdr->ucode_feature_version);
1005 ucode_fw = adev->gfx.ce_fw;
1006 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes);
1008 case AMDGPU_UCODE_ID_CP_MEC1:
1009 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1010 adev->gfx.mec_fw->data;
1011 adev->gfx.mec_fw_version =
1012 le32_to_cpu(cp_hdr->header.ucode_version);
1013 adev->gfx.mec_feature_version =
1014 le32_to_cpu(cp_hdr->ucode_feature_version);
1015 ucode_fw = adev->gfx.mec_fw;
1016 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1017 le32_to_cpu(cp_hdr->jt_size) * 4;
1019 case AMDGPU_UCODE_ID_CP_MEC1_JT:
1020 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1021 adev->gfx.mec_fw->data;
1022 ucode_fw = adev->gfx.mec_fw;
1023 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1025 case AMDGPU_UCODE_ID_CP_MEC2:
1026 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1027 adev->gfx.mec2_fw->data;
1028 adev->gfx.mec2_fw_version =
1029 le32_to_cpu(cp_hdr->header.ucode_version);
1030 adev->gfx.mec2_feature_version =
1031 le32_to_cpu(cp_hdr->ucode_feature_version);
1032 ucode_fw = adev->gfx.mec2_fw;
1033 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) -
1034 le32_to_cpu(cp_hdr->jt_size) * 4;
1036 case AMDGPU_UCODE_ID_CP_MEC2_JT:
1037 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
1038 adev->gfx.mec2_fw->data;
1039 ucode_fw = adev->gfx.mec2_fw;
1040 fw_size = le32_to_cpu(cp_hdr->jt_size) * 4;
1042 case AMDGPU_UCODE_ID_CP_RS64_MEC:
1043 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1044 adev->gfx.mec_fw->data;
1045 adev->gfx.mec_fw_version =
1046 le32_to_cpu(cp_hdr_v2_0->header.ucode_version);
1047 adev->gfx.mec_feature_version =
1048 le32_to_cpu(cp_hdr_v2_0->ucode_feature_version);
1049 ucode_fw = adev->gfx.mec_fw;
1050 fw_size = le32_to_cpu(cp_hdr_v2_0->ucode_size_bytes);
1052 case AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK:
1053 case AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK:
1054 case AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK:
1055 case AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK:
1056 cp_hdr_v2_0 = (const struct gfx_firmware_header_v2_0 *)
1057 adev->gfx.mec_fw->data;
1058 ucode_fw = adev->gfx.mec_fw;
1059 fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes);
1065 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1066 info = &adev->firmware.ucode[ucode_id];
1067 info->ucode_id = ucode_id;
1068 info->fw = ucode_fw;
1069 adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE);