2 * Copyright 2016 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
24 #include <linux/firmware.h>
25 #include <linux/pci.h>
27 #include <drm/drm_cache.h>
31 #include "amdgpu_atomfirmware.h"
32 #include "amdgpu_gem.h"
34 #include "gc/gc_9_0_sh_mask.h"
35 #include "dce/dce_12_0_offset.h"
36 #include "dce/dce_12_0_sh_mask.h"
37 #include "vega10_enum.h"
38 #include "mmhub/mmhub_1_0_offset.h"
39 #include "athub/athub_1_0_sh_mask.h"
40 #include "athub/athub_1_0_offset.h"
41 #include "oss/osssys_4_0_offset.h"
45 #include "soc15_common.h"
46 #include "umc/umc_6_0_sh_mask.h"
48 #include "gfxhub_v1_0.h"
49 #include "mmhub_v1_0.h"
50 #include "athub_v1_0.h"
51 #include "gfxhub_v1_1.h"
52 #include "mmhub_v9_4.h"
53 #include "mmhub_v1_7.h"
60 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
62 #include "amdgpu_ras.h"
63 #include "amdgpu_xgmi.h"
65 /* add these here since we already include dce12 headers and these are for DCN */
66 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION 0x055d
67 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX 2
68 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT 0x0
69 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT 0x10
70 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK 0x00003FFFL
71 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK 0x3FFF0000L
72 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0 0x049d
73 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0_BASE_IDX 2
76 static const char *gfxhub_client_ids[] = {
92 static const char *mmhub_client_ids_raven[][2] = {
117 static const char *mmhub_client_ids_renoir[][2] = {
145 static const char *mmhub_client_ids_vega10[][2] = {
158 [32+14][0] = "SDMA0",
171 [32+4][1] = "DCEDWB",
174 [32+14][1] = "SDMA1",
177 static const char *mmhub_client_ids_vega12[][2] = {
190 [32+15][0] = "SDMA0",
200 [32+1][1] = "DCEDWB",
206 [32+15][1] = "SDMA1",
209 static const char *mmhub_client_ids_vega20[][2] = {
223 [32+12][0] = "UTCL2",
224 [32+14][0] = "SDMA1",
242 [32+14][1] = "SDMA1",
245 static const char *mmhub_client_ids_arcturus[][2] = {
286 static const char *mmhub_client_ids_aldebaran[][2] = {
289 [32+1][0] = "DBGU_IO0",
290 [32+2][0] = "DBGU_IO2",
292 [96+11][0] = "JPEG0",
294 [96+13][0] = "VCNU0",
295 [128+11][0] = "JPEG1",
296 [128+12][0] = "VCN1",
297 [128+13][0] = "VCNU1",
300 [256+0][0] = "SDMA0",
301 [256+1][0] = "SDMA1",
302 [256+2][0] = "SDMA2",
303 [256+3][0] = "SDMA3",
304 [256+4][0] = "SDMA4",
308 [32+1][1] = "DBGU_IO0",
309 [32+2][1] = "DBGU_IO2",
311 [96+11][1] = "JPEG0",
313 [96+13][1] = "VCNU0",
314 [128+11][1] = "JPEG1",
315 [128+12][1] = "VCN1",
316 [128+13][1] = "VCNU1",
319 [256+0][1] = "SDMA0",
320 [256+1][1] = "SDMA1",
321 [256+2][1] = "SDMA2",
322 [256+3][1] = "SDMA3",
323 [256+4][1] = "SDMA4",
327 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] =
329 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
330 SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
333 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] =
335 SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
336 SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
339 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
340 (0x000143c0 + 0x00000000),
341 (0x000143c0 + 0x00000800),
342 (0x000143c0 + 0x00001000),
343 (0x000143c0 + 0x00001800),
344 (0x000543c0 + 0x00000000),
345 (0x000543c0 + 0x00000800),
346 (0x000543c0 + 0x00001000),
347 (0x000543c0 + 0x00001800),
348 (0x000943c0 + 0x00000000),
349 (0x000943c0 + 0x00000800),
350 (0x000943c0 + 0x00001000),
351 (0x000943c0 + 0x00001800),
352 (0x000d43c0 + 0x00000000),
353 (0x000d43c0 + 0x00000800),
354 (0x000d43c0 + 0x00001000),
355 (0x000d43c0 + 0x00001800),
356 (0x001143c0 + 0x00000000),
357 (0x001143c0 + 0x00000800),
358 (0x001143c0 + 0x00001000),
359 (0x001143c0 + 0x00001800),
360 (0x001543c0 + 0x00000000),
361 (0x001543c0 + 0x00000800),
362 (0x001543c0 + 0x00001000),
363 (0x001543c0 + 0x00001800),
364 (0x001943c0 + 0x00000000),
365 (0x001943c0 + 0x00000800),
366 (0x001943c0 + 0x00001000),
367 (0x001943c0 + 0x00001800),
368 (0x001d43c0 + 0x00000000),
369 (0x001d43c0 + 0x00000800),
370 (0x001d43c0 + 0x00001000),
371 (0x001d43c0 + 0x00001800),
374 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
375 (0x000143e0 + 0x00000000),
376 (0x000143e0 + 0x00000800),
377 (0x000143e0 + 0x00001000),
378 (0x000143e0 + 0x00001800),
379 (0x000543e0 + 0x00000000),
380 (0x000543e0 + 0x00000800),
381 (0x000543e0 + 0x00001000),
382 (0x000543e0 + 0x00001800),
383 (0x000943e0 + 0x00000000),
384 (0x000943e0 + 0x00000800),
385 (0x000943e0 + 0x00001000),
386 (0x000943e0 + 0x00001800),
387 (0x000d43e0 + 0x00000000),
388 (0x000d43e0 + 0x00000800),
389 (0x000d43e0 + 0x00001000),
390 (0x000d43e0 + 0x00001800),
391 (0x001143e0 + 0x00000000),
392 (0x001143e0 + 0x00000800),
393 (0x001143e0 + 0x00001000),
394 (0x001143e0 + 0x00001800),
395 (0x001543e0 + 0x00000000),
396 (0x001543e0 + 0x00000800),
397 (0x001543e0 + 0x00001000),
398 (0x001543e0 + 0x00001800),
399 (0x001943e0 + 0x00000000),
400 (0x001943e0 + 0x00000800),
401 (0x001943e0 + 0x00001000),
402 (0x001943e0 + 0x00001800),
403 (0x001d43e0 + 0x00000000),
404 (0x001d43e0 + 0x00000800),
405 (0x001d43e0 + 0x00001000),
406 (0x001d43e0 + 0x00001800),
409 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
410 struct amdgpu_irq_src *src,
412 enum amdgpu_interrupt_state state)
414 u32 bits, i, tmp, reg;
416 /* Devices newer then VEGA10/12 shall have these programming
417 sequences performed by PSP BL */
418 if (adev->asic_type >= CHIP_VEGA20)
424 case AMDGPU_IRQ_STATE_DISABLE:
425 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
426 reg = ecc_umc_mcumc_ctrl_addrs[i];
431 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
432 reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
438 case AMDGPU_IRQ_STATE_ENABLE:
439 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
440 reg = ecc_umc_mcumc_ctrl_addrs[i];
445 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
446 reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
459 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
460 struct amdgpu_irq_src *src,
462 enum amdgpu_interrupt_state state)
464 struct amdgpu_vmhub *hub;
465 u32 tmp, reg, bits, i, j;
467 bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
468 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
469 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
470 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
471 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
472 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
473 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
476 case AMDGPU_IRQ_STATE_DISABLE:
477 for (j = 0; j < adev->num_vmhubs; j++) {
478 hub = &adev->vmhub[j];
479 for (i = 0; i < 16; i++) {
480 reg = hub->vm_context0_cntl + i;
487 case AMDGPU_IRQ_STATE_ENABLE:
488 for (j = 0; j < adev->num_vmhubs; j++) {
489 hub = &adev->vmhub[j];
490 for (i = 0; i < 16; i++) {
491 reg = hub->vm_context0_cntl + i;
505 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
506 struct amdgpu_irq_src *source,
507 struct amdgpu_iv_entry *entry)
509 bool retry_fault = !!(entry->src_data[1] & 0x80);
510 uint32_t status = 0, cid = 0, rw = 0;
511 struct amdgpu_task_info task_info;
512 struct amdgpu_vmhub *hub;
513 const char *mmhub_cid;
514 const char *hub_name;
517 addr = (u64)entry->src_data[0] << 12;
518 addr |= ((u64)entry->src_data[1] & 0xf) << 44;
521 /* Returning 1 here also prevents sending the IV to the KFD */
523 /* Process it onyl if it's the first fault for this address */
524 if (entry->ih != &adev->irq.ih_soft &&
525 amdgpu_gmc_filter_faults(adev, addr, entry->pasid,
529 /* Delegate it to a different ring if the hardware hasn't
532 if (entry->ih == &adev->irq.ih) {
533 amdgpu_irq_delegate(adev, entry, 8);
537 /* Try to handle the recoverable page faults by filling page
540 if (amdgpu_vm_handle_fault(adev, entry->pasid, addr))
544 if (!printk_ratelimit())
547 if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
549 hub = &adev->vmhub[AMDGPU_MMHUB_0];
550 } else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
552 hub = &adev->vmhub[AMDGPU_MMHUB_1];
554 hub_name = "gfxhub0";
555 hub = &adev->vmhub[AMDGPU_GFXHUB_0];
558 memset(&task_info, 0, sizeof(struct amdgpu_task_info));
559 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
562 "[%s] %s page fault (src_id:%u ring:%u vmid:%u "
563 "pasid:%u, for process %s pid %d thread %s pid %d)\n",
564 hub_name, retry_fault ? "retry" : "no-retry",
565 entry->src_id, entry->ring_id, entry->vmid,
566 entry->pasid, task_info.process_name, task_info.tgid,
567 task_info.task_name, task_info.pid);
568 dev_err(adev->dev, " in page starting at address 0x%016llx from IH client 0x%x (%s)\n",
569 addr, entry->client_id,
570 soc15_ih_clientid_name[entry->client_id]);
572 if (amdgpu_sriov_vf(adev))
576 * Issue a dummy read to wait for the status register to
577 * be updated to avoid reading an incorrect value due to
578 * the new fast GRBM interface.
580 if ((entry->vmid_src == AMDGPU_GFXHUB_0) &&
581 (adev->asic_type < CHIP_ALDEBARAN))
582 RREG32(hub->vm_l2_pro_fault_status);
584 status = RREG32(hub->vm_l2_pro_fault_status);
585 cid = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, CID);
586 rw = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, RW);
587 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
591 "VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
593 if (hub == &adev->vmhub[AMDGPU_GFXHUB_0]) {
594 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n",
595 cid >= ARRAY_SIZE(gfxhub_client_ids) ? "unknown" :
596 gfxhub_client_ids[cid],
599 switch (adev->asic_type) {
601 mmhub_cid = mmhub_client_ids_vega10[cid][rw];
604 mmhub_cid = mmhub_client_ids_vega12[cid][rw];
607 mmhub_cid = mmhub_client_ids_vega20[cid][rw];
610 mmhub_cid = mmhub_client_ids_arcturus[cid][rw];
613 mmhub_cid = mmhub_client_ids_raven[cid][rw];
616 mmhub_cid = mmhub_client_ids_renoir[cid][rw];
619 mmhub_cid = mmhub_client_ids_aldebaran[cid][rw];
625 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n",
626 mmhub_cid ? mmhub_cid : "unknown", cid);
628 dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
629 REG_GET_FIELD(status,
630 VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
631 dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
632 REG_GET_FIELD(status,
633 VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
634 dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
635 REG_GET_FIELD(status,
636 VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
637 dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
638 REG_GET_FIELD(status,
639 VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
640 dev_err(adev->dev, "\t RW: 0x%x\n", rw);
644 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
645 .set = gmc_v9_0_vm_fault_interrupt_state,
646 .process = gmc_v9_0_process_interrupt,
650 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = {
651 .set = gmc_v9_0_ecc_interrupt_state,
652 .process = amdgpu_umc_process_ecc_irq,
655 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
657 adev->gmc.vm_fault.num_types = 1;
658 adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
660 if (!amdgpu_sriov_vf(adev) &&
661 !adev->gmc.xgmi.connected_to_cpu) {
662 adev->gmc.ecc_irq.num_types = 1;
663 adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs;
667 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
672 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
673 PER_VMID_INVALIDATE_REQ, 1 << vmid);
674 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
675 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
676 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
677 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
678 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
679 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
680 req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
681 CLEAR_PROTECTION_FAULT_STATUS_ADDR, 0);
687 * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore
689 * @adev: amdgpu_device pointer
693 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev,
696 if (adev->asic_type == CHIP_ALDEBARAN)
699 return ((vmhub == AMDGPU_MMHUB_0 ||
700 vmhub == AMDGPU_MMHUB_1) &&
701 (!amdgpu_sriov_vf(adev)) &&
702 (!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) &&
703 (adev->apu_flags & AMD_APU_IS_PICASSO))));
706 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
707 uint8_t vmid, uint16_t *p_pasid)
711 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
713 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
715 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
720 * VMID 0 is the physical GPU addresses as used by the kernel.
721 * VMIDs 1-15 are used for userspace clients and are handled
722 * by the amdgpu vm/hsa code.
726 * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
728 * @adev: amdgpu_device pointer
729 * @vmid: vm instance to flush
730 * @vmhub: which hub to flush
731 * @flush_type: the flush type
733 * Flush the TLB for the requested page table using certain type.
735 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
736 uint32_t vmhub, uint32_t flush_type)
738 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub);
739 const unsigned eng = 17;
740 u32 j, inv_req, inv_req2, tmp;
741 struct amdgpu_vmhub *hub;
743 BUG_ON(vmhub >= adev->num_vmhubs);
745 hub = &adev->vmhub[vmhub];
746 if (adev->gmc.xgmi.num_physical_nodes &&
747 adev->asic_type == CHIP_VEGA20) {
748 /* Vega20+XGMI caches PTEs in TC and TLB. Add a
749 * heavy-weight TLB flush (type 2), which flushes
750 * both. Due to a race condition with concurrent
751 * memory accesses using the same TLB cache line, we
752 * still need a second TLB flush after this.
754 inv_req = gmc_v9_0_get_invalidate_req(vmid, 2);
755 inv_req2 = gmc_v9_0_get_invalidate_req(vmid, flush_type);
757 inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type);
761 /* This is necessary for a HW workaround under SRIOV as well
762 * as GFXOFF under bare metal
764 if (adev->gfx.kiq.ring.sched.ready &&
765 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
766 down_read_trylock(&adev->reset_sem)) {
767 uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
768 uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
770 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
772 up_read(&adev->reset_sem);
776 spin_lock(&adev->gmc.invalidate_lock);
779 * It may lose gpuvm invalidate acknowldege state across power-gating
780 * off cycle, add semaphore acquire before invalidation and semaphore
781 * release after invalidation to avoid entering power gated state
785 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
787 for (j = 0; j < adev->usec_timeout; j++) {
788 /* a read return value of 1 means semaphore acuqire */
789 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem +
790 hub->eng_distance * eng);
796 if (j >= adev->usec_timeout)
797 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
801 WREG32_NO_KIQ(hub->vm_inv_eng0_req +
802 hub->eng_distance * eng, inv_req);
805 * Issue a dummy read to wait for the ACK register to
806 * be cleared to avoid a false ACK due to the new fast
809 if ((vmhub == AMDGPU_GFXHUB_0) &&
810 (adev->asic_type < CHIP_ALDEBARAN))
811 RREG32_NO_KIQ(hub->vm_inv_eng0_req +
812 hub->eng_distance * eng);
814 for (j = 0; j < adev->usec_timeout; j++) {
815 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack +
816 hub->eng_distance * eng);
817 if (tmp & (1 << vmid))
826 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
829 * add semaphore release after invalidation,
830 * write with 0 means semaphore release
832 WREG32_NO_KIQ(hub->vm_inv_eng0_sem +
833 hub->eng_distance * eng, 0);
835 spin_unlock(&adev->gmc.invalidate_lock);
837 if (j < adev->usec_timeout)
840 DRM_ERROR("Timeout waiting for VM flush ACK!\n");
844 * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid
846 * @adev: amdgpu_device pointer
847 * @pasid: pasid to be flush
848 * @flush_type: the flush type
849 * @all_hub: flush all hubs
851 * Flush the TLB for the requested pasid.
853 static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
854 uint16_t pasid, uint32_t flush_type,
860 uint16_t queried_pasid;
862 struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
863 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
865 if (amdgpu_in_reset(adev))
868 if (ring->sched.ready && down_read_trylock(&adev->reset_sem)) {
869 /* Vega20+XGMI caches PTEs in TC and TLB. Add a
870 * heavy-weight TLB flush (type 2), which flushes
871 * both. Due to a race condition with concurrent
872 * memory accesses using the same TLB cache line, we
873 * still need a second TLB flush after this.
875 bool vega20_xgmi_wa = (adev->gmc.xgmi.num_physical_nodes &&
876 adev->asic_type == CHIP_VEGA20);
877 /* 2 dwords flush + 8 dwords fence */
878 unsigned int ndw = kiq->pmf->invalidate_tlbs_size + 8;
881 ndw += kiq->pmf->invalidate_tlbs_size;
883 spin_lock(&adev->gfx.kiq.ring_lock);
884 /* 2 dwords flush + 8 dwords fence */
885 amdgpu_ring_alloc(ring, ndw);
887 kiq->pmf->kiq_invalidate_tlbs(ring,
889 kiq->pmf->kiq_invalidate_tlbs(ring,
890 pasid, flush_type, all_hub);
891 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
893 amdgpu_ring_undo(ring);
894 spin_unlock(&adev->gfx.kiq.ring_lock);
895 up_read(&adev->reset_sem);
899 amdgpu_ring_commit(ring);
900 spin_unlock(&adev->gfx.kiq.ring_lock);
901 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
903 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
904 up_read(&adev->reset_sem);
907 up_read(&adev->reset_sem);
911 for (vmid = 1; vmid < 16; vmid++) {
913 ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
915 if (ret && queried_pasid == pasid) {
917 for (i = 0; i < adev->num_vmhubs; i++)
918 gmc_v9_0_flush_gpu_tlb(adev, vmid,
921 gmc_v9_0_flush_gpu_tlb(adev, vmid,
922 AMDGPU_GFXHUB_0, flush_type);
932 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
933 unsigned vmid, uint64_t pd_addr)
935 bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub);
936 struct amdgpu_device *adev = ring->adev;
937 struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub];
938 uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
939 unsigned eng = ring->vm_inv_eng;
942 * It may lose gpuvm invalidate acknowldege state across power-gating
943 * off cycle, add semaphore acquire before invalidation and semaphore
944 * release after invalidation to avoid entering power gated state
948 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
950 /* a read return value of 1 means semaphore acuqire */
951 amdgpu_ring_emit_reg_wait(ring,
952 hub->vm_inv_eng0_sem +
953 hub->eng_distance * eng, 0x1, 0x1);
955 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
956 (hub->ctx_addr_distance * vmid),
957 lower_32_bits(pd_addr));
959 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
960 (hub->ctx_addr_distance * vmid),
961 upper_32_bits(pd_addr));
963 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
964 hub->eng_distance * eng,
965 hub->vm_inv_eng0_ack +
966 hub->eng_distance * eng,
969 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
972 * add semaphore release after invalidation,
973 * write with 0 means semaphore release
975 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
976 hub->eng_distance * eng, 0);
981 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
984 struct amdgpu_device *adev = ring->adev;
987 /* Do nothing because there's no lut register for mmhub1. */
988 if (ring->funcs->vmhub == AMDGPU_MMHUB_1)
991 if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
992 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
994 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
996 amdgpu_ring_emit_wreg(ring, reg, pasid);
1000 * PTE format on VEGA 10:
1009 * 47:12 4k physical page base address
1019 * PDE format on VEGA 10:
1020 * 63:59 block fragment size
1024 * 47:6 physical base address of PD or PTE
1031 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
1035 case AMDGPU_VM_MTYPE_DEFAULT:
1036 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1037 case AMDGPU_VM_MTYPE_NC:
1038 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1039 case AMDGPU_VM_MTYPE_WC:
1040 return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC);
1041 case AMDGPU_VM_MTYPE_RW:
1042 return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW);
1043 case AMDGPU_VM_MTYPE_CC:
1044 return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC);
1045 case AMDGPU_VM_MTYPE_UC:
1046 return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC);
1048 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1052 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
1053 uint64_t *addr, uint64_t *flags)
1055 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
1056 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr);
1057 BUG_ON(*addr & 0xFFFF00000000003FULL);
1059 if (!adev->gmc.translate_further)
1062 if (level == AMDGPU_VM_PDB1) {
1063 /* Set the block fragment size */
1064 if (!(*flags & AMDGPU_PDE_PTE))
1065 *flags |= AMDGPU_PDE_BFS(0x9);
1067 } else if (level == AMDGPU_VM_PDB0) {
1068 if (*flags & AMDGPU_PDE_PTE)
1069 *flags &= ~AMDGPU_PDE_PTE;
1071 *flags |= AMDGPU_PTE_TF;
1075 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
1076 struct amdgpu_bo_va_mapping *mapping,
1079 *flags &= ~AMDGPU_PTE_EXECUTABLE;
1080 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
1082 *flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
1083 *flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK;
1085 if (mapping->flags & AMDGPU_PTE_PRT) {
1086 *flags |= AMDGPU_PTE_PRT;
1087 *flags &= ~AMDGPU_PTE_VALID;
1090 if ((adev->asic_type == CHIP_ARCTURUS ||
1091 adev->asic_type == CHIP_ALDEBARAN) &&
1092 !(*flags & AMDGPU_PTE_SYSTEM) &&
1093 mapping->bo_va->is_xgmi)
1094 *flags |= AMDGPU_PTE_SNOOPED;
1096 if (adev->asic_type == CHIP_ALDEBARAN)
1097 *flags |= mapping->flags & AMDGPU_PTE_SNOOPED;
1100 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
1102 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
1105 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1106 size = AMDGPU_VBIOS_VGA_ALLOCATION;
1110 switch (adev->asic_type) {
1113 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
1114 size = (REG_GET_FIELD(viewport,
1115 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1116 REG_GET_FIELD(viewport,
1117 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1124 viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
1125 size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1126 REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1135 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
1136 .flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
1137 .flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid,
1138 .emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
1139 .emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
1140 .map_mtype = gmc_v9_0_map_mtype,
1141 .get_vm_pde = gmc_v9_0_get_vm_pde,
1142 .get_vm_pte = gmc_v9_0_get_vm_pte,
1143 .get_vbios_fb_size = gmc_v9_0_get_vbios_fb_size,
1146 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
1148 adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
1151 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
1153 switch (adev->asic_type) {
1155 adev->umc.funcs = &umc_v6_0_funcs;
1158 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
1159 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
1160 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
1161 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20;
1162 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
1163 adev->umc.ras_funcs = &umc_v6_1_ras_funcs;
1166 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
1167 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
1168 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
1169 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT;
1170 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
1171 adev->umc.ras_funcs = &umc_v6_1_ras_funcs;
1173 case CHIP_ALDEBARAN:
1174 adev->umc.max_ras_err_cnt_per_query = UMC_V6_7_TOTAL_CHANNEL_NUM;
1175 adev->umc.channel_inst_num = UMC_V6_7_CHANNEL_INSTANCE_NUM;
1176 adev->umc.umc_inst_num = UMC_V6_7_UMC_INSTANCE_NUM;
1177 adev->umc.channel_offs = UMC_V6_7_PER_CHANNEL_OFFSET;
1178 if (!adev->gmc.xgmi.connected_to_cpu)
1179 adev->umc.ras_funcs = &umc_v6_7_ras_funcs;
1180 if (1 & adev->smuio.funcs->get_die_id(adev))
1181 adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_first[0][0];
1183 adev->umc.channel_idx_tbl = &umc_v6_7_channel_idx_tbl_second[0][0];
1190 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
1192 switch (adev->asic_type) {
1194 adev->mmhub.funcs = &mmhub_v9_4_funcs;
1196 case CHIP_ALDEBARAN:
1197 adev->mmhub.funcs = &mmhub_v1_7_funcs;
1200 adev->mmhub.funcs = &mmhub_v1_0_funcs;
1205 static void gmc_v9_0_set_mmhub_ras_funcs(struct amdgpu_device *adev)
1207 switch (adev->asic_type) {
1209 adev->mmhub.ras_funcs = &mmhub_v1_0_ras_funcs;
1212 adev->mmhub.ras_funcs = &mmhub_v9_4_ras_funcs;
1214 case CHIP_ALDEBARAN:
1215 adev->mmhub.ras_funcs = &mmhub_v1_7_ras_funcs;
1218 /* mmhub ras is not available */
1223 static void gmc_v9_0_set_gfxhub_funcs(struct amdgpu_device *adev)
1225 adev->gfxhub.funcs = &gfxhub_v1_0_funcs;
1228 static void gmc_v9_0_set_hdp_ras_funcs(struct amdgpu_device *adev)
1230 adev->hdp.ras_funcs = &hdp_v4_0_ras_funcs;
1233 static void gmc_v9_0_set_mca_funcs(struct amdgpu_device *adev)
1235 switch (adev->asic_type) {
1236 case CHIP_ALDEBARAN:
1237 if (!adev->gmc.xgmi.connected_to_cpu)
1238 adev->mca.funcs = &mca_v3_0_funcs;
1245 static int gmc_v9_0_early_init(void *handle)
1247 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1249 if (adev->asic_type == CHIP_VEGA20 ||
1250 adev->asic_type == CHIP_ARCTURUS)
1251 adev->gmc.xgmi.supported = true;
1253 if (adev->asic_type == CHIP_ALDEBARAN) {
1254 adev->gmc.xgmi.supported = true;
1255 adev->gmc.xgmi.connected_to_cpu =
1256 adev->smuio.funcs->is_host_gpu_xgmi_supported(adev);
1259 gmc_v9_0_set_gmc_funcs(adev);
1260 gmc_v9_0_set_irq_funcs(adev);
1261 gmc_v9_0_set_umc_funcs(adev);
1262 gmc_v9_0_set_mmhub_funcs(adev);
1263 gmc_v9_0_set_mmhub_ras_funcs(adev);
1264 gmc_v9_0_set_gfxhub_funcs(adev);
1265 gmc_v9_0_set_hdp_ras_funcs(adev);
1266 gmc_v9_0_set_mca_funcs(adev);
1268 adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
1269 adev->gmc.shared_aperture_end =
1270 adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
1271 adev->gmc.private_aperture_start = 0x1000000000000000ULL;
1272 adev->gmc.private_aperture_end =
1273 adev->gmc.private_aperture_start + (4ULL << 30) - 1;
1278 static int gmc_v9_0_late_init(void *handle)
1280 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1283 r = amdgpu_gmc_allocate_vm_inv_eng(adev);
1288 * Workaround performance drop issue with VBIOS enables partial
1289 * writes, while disables HBM ECC for vega10.
1291 if (!amdgpu_sriov_vf(adev) && (adev->asic_type == CHIP_VEGA10)) {
1292 if (!(adev->ras_enabled & (1 << AMDGPU_RAS_BLOCK__UMC))) {
1293 if (adev->df.funcs->enable_ecc_force_par_wr_rmw)
1294 adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false);
1298 if (!amdgpu_persistent_edc_harvesting_supported(adev)) {
1299 if (adev->mmhub.ras_funcs &&
1300 adev->mmhub.ras_funcs->reset_ras_error_count)
1301 adev->mmhub.ras_funcs->reset_ras_error_count(adev);
1303 if (adev->hdp.ras_funcs &&
1304 adev->hdp.ras_funcs->reset_ras_error_count)
1305 adev->hdp.ras_funcs->reset_ras_error_count(adev);
1308 r = amdgpu_gmc_ras_late_init(adev);
1312 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
1315 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
1316 struct amdgpu_gmc *mc)
1318 u64 base = adev->mmhub.funcs->get_fb_location(adev);
1320 /* add the xgmi offset of the physical node */
1321 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1322 if (adev->gmc.xgmi.connected_to_cpu) {
1323 amdgpu_gmc_sysvm_location(adev, mc);
1325 amdgpu_gmc_vram_location(adev, mc, base);
1326 amdgpu_gmc_gart_location(adev, mc);
1327 amdgpu_gmc_agp_location(adev, mc);
1329 /* base offset of vram pages */
1330 adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev);
1332 /* XXX: add the xgmi offset of the physical node? */
1333 adev->vm_manager.vram_base_offset +=
1334 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1338 * gmc_v9_0_mc_init - initialize the memory controller driver params
1340 * @adev: amdgpu_device pointer
1342 * Look up the amount of vram, vram width, and decide how to place
1343 * vram and gart within the GPU's physical address space.
1344 * Returns 0 for success.
1346 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
1350 /* size in MB on si */
1351 adev->gmc.mc_vram_size =
1352 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
1353 adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
1355 if (!(adev->flags & AMD_IS_APU) &&
1356 !adev->gmc.xgmi.connected_to_cpu) {
1357 r = amdgpu_device_resize_fb_bar(adev);
1361 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
1362 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
1364 #ifdef CONFIG_X86_64
1366 * AMD Accelerated Processing Platform (APP) supporting GPU-HOST xgmi
1367 * interface can use VRAM through here as it appears system reserved
1368 * memory in host address space.
1370 * For APUs, VRAM is just the stolen system memory and can be accessed
1373 * Otherwise, use the legacy Host Data Path (HDP) through PCIe BAR.
1376 /* check whether both host-gpu and gpu-gpu xgmi links exist */
1377 if ((adev->flags & AMD_IS_APU) ||
1378 (adev->gmc.xgmi.supported &&
1379 adev->gmc.xgmi.connected_to_cpu)) {
1380 adev->gmc.aper_base =
1381 adev->gfxhub.funcs->get_mc_fb_offset(adev) +
1382 adev->gmc.xgmi.physical_node_id *
1383 adev->gmc.xgmi.node_segment_size;
1384 adev->gmc.aper_size = adev->gmc.real_vram_size;
1388 /* In case the PCI BAR is larger than the actual amount of vram */
1389 adev->gmc.visible_vram_size = adev->gmc.aper_size;
1390 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
1391 adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
1393 /* set the gart size */
1394 if (amdgpu_gart_size == -1) {
1395 switch (adev->asic_type) {
1396 case CHIP_VEGA10: /* all engines support GPUVM */
1397 case CHIP_VEGA12: /* all engines support GPUVM */
1400 case CHIP_ALDEBARAN:
1402 adev->gmc.gart_size = 512ULL << 20;
1404 case CHIP_RAVEN: /* DCE SG support */
1406 adev->gmc.gart_size = 1024ULL << 20;
1410 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
1413 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size;
1415 gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
1420 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
1424 if (adev->gart.bo) {
1425 WARN(1, "VEGA10 PCIE GART already initialized\n");
1429 if (adev->gmc.xgmi.connected_to_cpu) {
1430 adev->gmc.vmid0_page_table_depth = 1;
1431 adev->gmc.vmid0_page_table_block_size = 12;
1433 adev->gmc.vmid0_page_table_depth = 0;
1434 adev->gmc.vmid0_page_table_block_size = 0;
1437 /* Initialize common gart structure */
1438 r = amdgpu_gart_init(adev);
1441 adev->gart.table_size = adev->gart.num_gpu_pages * 8;
1442 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) |
1443 AMDGPU_PTE_EXECUTABLE;
1445 r = amdgpu_gart_table_vram_alloc(adev);
1449 if (adev->gmc.xgmi.connected_to_cpu) {
1450 r = amdgpu_gmc_pdb0_alloc(adev);
1457 * gmc_v9_0_save_registers - saves regs
1459 * @adev: amdgpu_device pointer
1461 * This saves potential register values that should be
1462 * restored upon resume
1464 static void gmc_v9_0_save_registers(struct amdgpu_device *adev)
1466 if (adev->asic_type == CHIP_RAVEN)
1467 adev->gmc.sdpif_register = RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0);
1470 static int gmc_v9_0_sw_init(void *handle)
1472 int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
1473 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1475 adev->gfxhub.funcs->init(adev);
1477 adev->mmhub.funcs->init(adev);
1478 if (adev->mca.funcs)
1479 adev->mca.funcs->init(adev);
1481 spin_lock_init(&adev->gmc.invalidate_lock);
1483 r = amdgpu_atomfirmware_get_vram_info(adev,
1484 &vram_width, &vram_type, &vram_vendor);
1485 if (amdgpu_sriov_vf(adev))
1486 /* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
1487 * and DF related registers is not readable, seems hardcord is the
1488 * only way to set the correct vram_width
1490 adev->gmc.vram_width = 2048;
1491 else if (amdgpu_emu_mode != 1)
1492 adev->gmc.vram_width = vram_width;
1494 if (!adev->gmc.vram_width) {
1495 int chansize, numchan;
1497 /* hbm memory channel size */
1498 if (adev->flags & AMD_IS_APU)
1503 numchan = adev->df.funcs->get_hbm_channel_number(adev);
1504 adev->gmc.vram_width = numchan * chansize;
1507 adev->gmc.vram_type = vram_type;
1508 adev->gmc.vram_vendor = vram_vendor;
1509 switch (adev->asic_type) {
1511 adev->num_vmhubs = 2;
1513 if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
1514 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1516 /* vm_size is 128TB + 512GB for legacy 3-level page support */
1517 amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
1518 adev->gmc.translate_further =
1519 adev->vm_manager.num_level > 1;
1526 case CHIP_ALDEBARAN:
1527 adev->num_vmhubs = 2;
1531 * To fulfill 4-level page support,
1532 * vm size is 256TB (48bit), maximum size of Vega10,
1533 * block size 512 (9bit)
1535 /* sriov restrict max_pfn below AMDGPU_GMC_HOLE */
1536 if (amdgpu_sriov_vf(adev))
1537 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47);
1539 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1542 adev->num_vmhubs = 3;
1544 /* Keep the vm size same with Vega20 */
1545 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1551 /* This interrupt is VMC page fault.*/
1552 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
1553 &adev->gmc.vm_fault);
1557 if (adev->asic_type == CHIP_ARCTURUS) {
1558 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
1559 &adev->gmc.vm_fault);
1564 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
1565 &adev->gmc.vm_fault);
1570 if (!amdgpu_sriov_vf(adev) &&
1571 !adev->gmc.xgmi.connected_to_cpu) {
1572 /* interrupt sent to DF. */
1573 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
1574 &adev->gmc.ecc_irq);
1579 /* Set the internal MC address mask
1580 * This is the max address of the GPU's
1581 * internal address space.
1583 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
1585 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
1587 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
1590 adev->need_swiotlb = drm_need_swiotlb(44);
1592 if (adev->gmc.xgmi.supported) {
1593 r = adev->gfxhub.funcs->get_xgmi_info(adev);
1598 r = gmc_v9_0_mc_init(adev);
1602 amdgpu_gmc_get_vbios_allocations(adev);
1604 /* Memory manager */
1605 r = amdgpu_bo_init(adev);
1609 r = gmc_v9_0_gart_init(adev);
1615 * VMID 0 is reserved for System
1616 * amdgpu graphics/compute will use VMIDs 1..n-1
1617 * amdkfd will use VMIDs n..15
1619 * The first KFD VMID is 8 for GPUs with graphics, 3 for
1620 * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs
1621 * for video processing.
1623 adev->vm_manager.first_kfd_vmid =
1624 (adev->asic_type == CHIP_ARCTURUS ||
1625 adev->asic_type == CHIP_ALDEBARAN) ? 3 : 8;
1627 amdgpu_vm_manager_init(adev);
1629 gmc_v9_0_save_registers(adev);
1634 static int gmc_v9_0_sw_fini(void *handle)
1636 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1638 amdgpu_gmc_ras_fini(adev);
1639 amdgpu_gem_force_release(adev);
1640 amdgpu_vm_manager_fini(adev);
1641 amdgpu_gart_table_vram_free(adev);
1642 amdgpu_bo_unref(&adev->gmc.pdb0_bo);
1643 amdgpu_bo_fini(adev);
1648 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
1651 switch (adev->asic_type) {
1653 if (amdgpu_sriov_vf(adev))
1657 soc15_program_register_sequence(adev,
1658 golden_settings_mmhub_1_0_0,
1659 ARRAY_SIZE(golden_settings_mmhub_1_0_0));
1660 soc15_program_register_sequence(adev,
1661 golden_settings_athub_1_0_0,
1662 ARRAY_SIZE(golden_settings_athub_1_0_0));
1667 /* TODO for renoir */
1668 soc15_program_register_sequence(adev,
1669 golden_settings_athub_1_0_0,
1670 ARRAY_SIZE(golden_settings_athub_1_0_0));
1678 * gmc_v9_0_restore_registers - restores regs
1680 * @adev: amdgpu_device pointer
1682 * This restores register values, saved at suspend.
1684 void gmc_v9_0_restore_registers(struct amdgpu_device *adev)
1686 if (adev->asic_type == CHIP_RAVEN) {
1687 WREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register);
1688 WARN_ON(adev->gmc.sdpif_register !=
1689 RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0));
1694 * gmc_v9_0_gart_enable - gart enable
1696 * @adev: amdgpu_device pointer
1698 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
1702 if (adev->gmc.xgmi.connected_to_cpu)
1703 amdgpu_gmc_init_pdb0(adev);
1705 if (adev->gart.bo == NULL) {
1706 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1710 r = amdgpu_gart_table_vram_pin(adev);
1714 r = adev->gfxhub.funcs->gart_enable(adev);
1718 r = adev->mmhub.funcs->gart_enable(adev);
1722 DRM_INFO("PCIE GART of %uM enabled.\n",
1723 (unsigned)(adev->gmc.gart_size >> 20));
1724 if (adev->gmc.pdb0_bo)
1725 DRM_INFO("PDB0 located at 0x%016llX\n",
1726 (unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo));
1727 DRM_INFO("PTB located at 0x%016llX\n",
1728 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1730 adev->gart.ready = true;
1734 static int gmc_v9_0_hw_init(void *handle)
1736 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1740 /* The sequence of these two function calls matters.*/
1741 gmc_v9_0_init_golden_registers(adev);
1743 if (adev->mode_info.num_crtc) {
1744 /* Lockout access through VGA aperture*/
1745 WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
1746 /* disable VGA render */
1747 WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
1750 if (adev->mmhub.funcs->update_power_gating)
1751 adev->mmhub.funcs->update_power_gating(adev, true);
1753 adev->hdp.funcs->init_registers(adev);
1755 /* After HDP is initialized, flush HDP.*/
1756 adev->hdp.funcs->flush_hdp(adev, NULL);
1758 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
1763 if (!amdgpu_sriov_vf(adev)) {
1764 adev->gfxhub.funcs->set_fault_enable_default(adev, value);
1765 adev->mmhub.funcs->set_fault_enable_default(adev, value);
1767 for (i = 0; i < adev->num_vmhubs; ++i)
1768 gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
1770 if (adev->umc.funcs && adev->umc.funcs->init_registers)
1771 adev->umc.funcs->init_registers(adev);
1773 r = gmc_v9_0_gart_enable(adev);
1779 * gmc_v9_0_gart_disable - gart disable
1781 * @adev: amdgpu_device pointer
1783 * This disables all VM page table.
1785 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
1787 adev->gfxhub.funcs->gart_disable(adev);
1788 adev->mmhub.funcs->gart_disable(adev);
1789 amdgpu_gart_table_vram_unpin(adev);
1792 static int gmc_v9_0_hw_fini(void *handle)
1794 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1796 if (amdgpu_sriov_vf(adev)) {
1797 /* full access mode, so don't touch any GMC register */
1798 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1802 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1803 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1804 gmc_v9_0_gart_disable(adev);
1809 static int gmc_v9_0_suspend(void *handle)
1811 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1813 return gmc_v9_0_hw_fini(adev);
1816 static int gmc_v9_0_resume(void *handle)
1819 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1821 r = gmc_v9_0_hw_init(adev);
1825 amdgpu_vmid_reset_all(adev);
1830 static bool gmc_v9_0_is_idle(void *handle)
1832 /* MC is always ready in GMC v9.*/
1836 static int gmc_v9_0_wait_for_idle(void *handle)
1838 /* There is no need to wait for MC idle in GMC v9.*/
1842 static int gmc_v9_0_soft_reset(void *handle)
1844 /* XXX for emulation.*/
1848 static int gmc_v9_0_set_clockgating_state(void *handle,
1849 enum amd_clockgating_state state)
1851 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1853 adev->mmhub.funcs->set_clockgating(adev, state);
1855 athub_v1_0_set_clockgating(adev, state);
1860 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags)
1862 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1864 adev->mmhub.funcs->get_clockgating(adev, flags);
1866 athub_v1_0_get_clockgating(adev, flags);
1869 static int gmc_v9_0_set_powergating_state(void *handle,
1870 enum amd_powergating_state state)
1875 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
1877 .early_init = gmc_v9_0_early_init,
1878 .late_init = gmc_v9_0_late_init,
1879 .sw_init = gmc_v9_0_sw_init,
1880 .sw_fini = gmc_v9_0_sw_fini,
1881 .hw_init = gmc_v9_0_hw_init,
1882 .hw_fini = gmc_v9_0_hw_fini,
1883 .suspend = gmc_v9_0_suspend,
1884 .resume = gmc_v9_0_resume,
1885 .is_idle = gmc_v9_0_is_idle,
1886 .wait_for_idle = gmc_v9_0_wait_for_idle,
1887 .soft_reset = gmc_v9_0_soft_reset,
1888 .set_clockgating_state = gmc_v9_0_set_clockgating_state,
1889 .set_powergating_state = gmc_v9_0_set_powergating_state,
1890 .get_clockgating_state = gmc_v9_0_get_clockgating_state,
1893 const struct amdgpu_ip_block_version gmc_v9_0_ip_block =
1895 .type = AMD_IP_BLOCK_TYPE_GMC,
1899 .funcs = &gmc_v9_0_ip_funcs,