2 * Copyright 2019 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.
23 #include <linux/firmware.h>
24 #include <linux/pci.h>
26 #include <drm/drm_cache.h>
29 #include "amdgpu_atomfirmware.h"
30 #include "gmc_v10_0.h"
33 #include "athub/athub_2_0_0_sh_mask.h"
34 #include "athub/athub_2_0_0_offset.h"
35 #include "dcn/dcn_2_0_0_offset.h"
36 #include "dcn/dcn_2_0_0_sh_mask.h"
37 #include "oss/osssys_5_0_0_offset.h"
38 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
39 #include "navi10_enum.h"
43 #include "soc15_common.h"
45 #include "nbio_v2_3.h"
47 #include "gfxhub_v2_0.h"
48 #include "gfxhub_v2_1.h"
49 #include "mmhub_v2_0.h"
50 #include "mmhub_v2_3.h"
51 #include "athub_v2_0.h"
52 #include "athub_v2_1.h"
54 #include "amdgpu_reset.h"
57 static const struct soc15_reg_golden golden_settings_navi10_hdp[] =
59 /* TODO add golden setting for hdp */
63 static int gmc_v10_0_ecc_interrupt_state(struct amdgpu_device *adev,
64 struct amdgpu_irq_src *src,
66 enum amdgpu_interrupt_state state)
72 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
73 struct amdgpu_irq_src *src, unsigned type,
74 enum amdgpu_interrupt_state state)
77 case AMDGPU_IRQ_STATE_DISABLE:
79 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, false);
81 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, false);
83 case AMDGPU_IRQ_STATE_ENABLE:
85 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, true);
87 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, true);
96 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev,
97 struct amdgpu_irq_src *source,
98 struct amdgpu_iv_entry *entry)
100 bool retry_fault = !!(entry->src_data[1] & 0x80);
101 bool write_fault = !!(entry->src_data[1] & 0x20);
102 struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src];
103 struct amdgpu_task_info task_info;
107 addr = (u64)entry->src_data[0] << 12;
108 addr |= ((u64)entry->src_data[1] & 0xf) << 44;
111 /* Returning 1 here also prevents sending the IV to the KFD */
113 /* Process it onyl if it's the first fault for this address */
114 if (entry->ih != &adev->irq.ih_soft &&
115 amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid,
119 /* Delegate it to a different ring if the hardware hasn't
122 if (entry->ih == &adev->irq.ih) {
123 amdgpu_irq_delegate(adev, entry, 8);
127 /* Try to handle the recoverable page faults by filling page
130 if (amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault))
134 if (!amdgpu_sriov_vf(adev)) {
136 * Issue a dummy read to wait for the status register to
137 * be updated to avoid reading an incorrect value due to
138 * the new fast GRBM interface.
140 if ((entry->vmid_src == AMDGPU_GFXHUB_0) &&
141 (adev->ip_versions[GC_HWIP][0] < IP_VERSION(10, 3, 0)))
142 RREG32(hub->vm_l2_pro_fault_status);
144 status = RREG32(hub->vm_l2_pro_fault_status);
145 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
148 if (!printk_ratelimit())
151 memset(&task_info, 0, sizeof(struct amdgpu_task_info));
152 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
155 "[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, "
156 "for process %s pid %d thread %s pid %d)\n",
157 entry->vmid_src ? "mmhub" : "gfxhub",
158 entry->src_id, entry->ring_id, entry->vmid,
159 entry->pasid, task_info.process_name, task_info.tgid,
160 task_info.task_name, task_info.pid);
161 dev_err(adev->dev, " in page starting at address 0x%016llx from client 0x%x (%s)\n",
162 addr, entry->client_id,
163 soc15_ih_clientid_name[entry->client_id]);
165 if (!amdgpu_sriov_vf(adev))
166 hub->vmhub_funcs->print_l2_protection_fault_status(adev,
172 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = {
173 .set = gmc_v10_0_vm_fault_interrupt_state,
174 .process = gmc_v10_0_process_interrupt,
177 static const struct amdgpu_irq_src_funcs gmc_v10_0_ecc_funcs = {
178 .set = gmc_v10_0_ecc_interrupt_state,
179 .process = amdgpu_umc_process_ecc_irq,
182 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev)
184 adev->gmc.vm_fault.num_types = 1;
185 adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs;
187 if (!amdgpu_sriov_vf(adev)) {
188 adev->gmc.ecc_irq.num_types = 1;
189 adev->gmc.ecc_irq.funcs = &gmc_v10_0_ecc_funcs;
194 * gmc_v10_0_use_invalidate_semaphore - judge whether to use semaphore
196 * @adev: amdgpu_device pointer
200 static bool gmc_v10_0_use_invalidate_semaphore(struct amdgpu_device *adev,
203 return ((vmhub == AMDGPU_MMHUB_0 ||
204 vmhub == AMDGPU_MMHUB_1) &&
205 (!amdgpu_sriov_vf(adev)));
208 static bool gmc_v10_0_get_atc_vmid_pasid_mapping_info(
209 struct amdgpu_device *adev,
210 uint8_t vmid, uint16_t *p_pasid)
214 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
216 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
218 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
223 * VMID 0 is the physical GPU addresses as used by the kernel.
224 * VMIDs 1-15 are used for userspace clients and are handled
225 * by the amdgpu vm/hsa code.
228 static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
229 unsigned int vmhub, uint32_t flush_type)
231 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub);
232 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
233 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
235 /* Use register 17 for GART */
236 const unsigned eng = 17;
238 unsigned char hub_ip = 0;
240 hub_ip = (vmhub == AMDGPU_GFXHUB_0) ?
241 GC_HWIP : MMHUB_HWIP;
243 spin_lock(&adev->gmc.invalidate_lock);
245 * It may lose gpuvm invalidate acknowldege state across power-gating
246 * off cycle, add semaphore acquire before invalidation and semaphore
247 * release after invalidation to avoid entering power gated state
251 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
253 for (i = 0; i < adev->usec_timeout; i++) {
254 /* a read return value of 1 means semaphore acuqire */
255 tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
256 hub->eng_distance * eng, hub_ip);
263 if (i >= adev->usec_timeout)
264 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
267 WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req +
268 hub->eng_distance * eng,
272 * Issue a dummy read to wait for the ACK register to be cleared
273 * to avoid a false ACK due to the new fast GRBM interface.
275 if ((vmhub == AMDGPU_GFXHUB_0) &&
276 (adev->ip_versions[GC_HWIP][0] < IP_VERSION(10, 3, 0)))
277 RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req +
278 hub->eng_distance * eng, hub_ip);
280 /* Wait for ACK with a delay.*/
281 for (i = 0; i < adev->usec_timeout; i++) {
282 tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_ack +
283 hub->eng_distance * eng, hub_ip);
292 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
295 * add semaphore release after invalidation,
296 * write with 0 means semaphore release
298 WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
299 hub->eng_distance * eng, 0, hub_ip);
301 spin_unlock(&adev->gmc.invalidate_lock);
303 if (i < adev->usec_timeout)
306 DRM_ERROR("Timeout waiting for VM flush hub: %d!\n", vmhub);
310 * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback
312 * @adev: amdgpu_device pointer
313 * @vmid: vm instance to flush
315 * @flush_type: the flush type
317 * Flush the TLB for the requested page table.
319 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
320 uint32_t vmhub, uint32_t flush_type)
322 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
323 struct dma_fence *fence;
324 struct amdgpu_job *job;
328 /* flush hdp cache */
329 adev->hdp.funcs->flush_hdp(adev, NULL);
331 /* For SRIOV run time, driver shouldn't access the register through MMIO
332 * Directly use kiq to do the vm invalidation instead
334 if (adev->gfx.kiq.ring.sched.ready && !adev->enable_mes &&
335 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
336 down_read_trylock(&adev->reset_domain->sem)) {
337 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
338 const unsigned eng = 17;
339 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
340 u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
341 u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
343 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
346 up_read(&adev->reset_domain->sem);
350 mutex_lock(&adev->mman.gtt_window_lock);
352 if (vmhub == AMDGPU_MMHUB_0) {
353 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0);
354 mutex_unlock(&adev->mman.gtt_window_lock);
358 BUG_ON(vmhub != AMDGPU_GFXHUB_0);
360 if (!adev->mman.buffer_funcs_enabled ||
361 !adev->ib_pool_ready ||
362 amdgpu_in_reset(adev) ||
363 ring->sched.ready == false) {
364 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_GFXHUB_0, 0);
365 mutex_unlock(&adev->mman.gtt_window_lock);
369 /* The SDMA on Navi has a bug which can theoretically result in memory
370 * corruption if an invalidation happens at the same time as an VA
371 * translation. Avoid this by doing the invalidation from the SDMA
374 r = amdgpu_job_alloc_with_ib(ring->adev, &adev->mman.entity,
375 AMDGPU_FENCE_OWNER_UNDEFINED,
376 16 * 4, AMDGPU_IB_POOL_IMMEDIATE,
381 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
382 job->vm_needs_flush = true;
383 job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
384 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
385 fence = amdgpu_job_submit(job);
387 mutex_unlock(&adev->mman.gtt_window_lock);
389 dma_fence_wait(fence, false);
390 dma_fence_put(fence);
395 mutex_unlock(&adev->mman.gtt_window_lock);
396 DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r);
400 * gmc_v10_0_flush_gpu_tlb_pasid - tlb flush via pasid
402 * @adev: amdgpu_device pointer
403 * @pasid: pasid to be flush
404 * @flush_type: the flush type
405 * @all_hub: Used with PACKET3_INVALIDATE_TLBS_ALL_HUB()
407 * Flush the TLB for the requested pasid.
409 static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
410 uint16_t pasid, uint32_t flush_type,
416 uint16_t queried_pasid;
418 u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT : adev->usec_timeout;
419 struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
420 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
422 if (amdgpu_emu_mode == 0 && ring->sched.ready) {
423 spin_lock(&adev->gfx.kiq.ring_lock);
424 /* 2 dwords flush + 8 dwords fence */
425 amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size + 8);
426 kiq->pmf->kiq_invalidate_tlbs(ring,
427 pasid, flush_type, all_hub);
428 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
430 amdgpu_ring_undo(ring);
431 spin_unlock(&adev->gfx.kiq.ring_lock);
435 amdgpu_ring_commit(ring);
436 spin_unlock(&adev->gfx.kiq.ring_lock);
437 r = amdgpu_fence_wait_polling(ring, seq, usec_timeout);
439 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
446 for (vmid = 1; vmid < AMDGPU_NUM_VMID; vmid++) {
448 ret = gmc_v10_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
450 if (ret && queried_pasid == pasid) {
452 for (i = 0; i < adev->num_vmhubs; i++)
453 gmc_v10_0_flush_gpu_tlb(adev, vmid,
456 gmc_v10_0_flush_gpu_tlb(adev, vmid,
457 AMDGPU_GFXHUB_0, flush_type);
459 if (!adev->enable_mes)
467 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
468 unsigned vmid, uint64_t pd_addr)
470 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub);
471 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub];
472 uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
473 unsigned eng = ring->vm_inv_eng;
476 * It may lose gpuvm invalidate acknowldege state across power-gating
477 * off cycle, add semaphore acquire before invalidation and semaphore
478 * release after invalidation to avoid entering power gated state
482 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
484 /* a read return value of 1 means semaphore acuqire */
485 amdgpu_ring_emit_reg_wait(ring,
486 hub->vm_inv_eng0_sem +
487 hub->eng_distance * eng, 0x1, 0x1);
489 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
490 (hub->ctx_addr_distance * vmid),
491 lower_32_bits(pd_addr));
493 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
494 (hub->ctx_addr_distance * vmid),
495 upper_32_bits(pd_addr));
497 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
498 hub->eng_distance * eng,
499 hub->vm_inv_eng0_ack +
500 hub->eng_distance * eng,
503 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
506 * add semaphore release after invalidation,
507 * write with 0 means semaphore release
509 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
510 hub->eng_distance * eng, 0);
515 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
518 struct amdgpu_device *adev = ring->adev;
521 /* MES fw manages IH_VMID_x_LUT updating */
522 if (ring->is_mes_queue)
525 if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
526 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
528 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
530 amdgpu_ring_emit_wreg(ring, reg, pasid);
534 * PTE format on NAVI 10:
536 * 58 reserved and for sienna_cichlid is used for MALL noalloc
544 * 47:12 4k physical page base address
554 * PDE format on NAVI 10:
555 * 63:59 block fragment size
559 * 47:6 physical base address of PD or PTE
566 static uint64_t gmc_v10_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
569 case AMDGPU_VM_MTYPE_DEFAULT:
570 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
571 case AMDGPU_VM_MTYPE_NC:
572 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
573 case AMDGPU_VM_MTYPE_WC:
574 return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC);
575 case AMDGPU_VM_MTYPE_CC:
576 return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC);
577 case AMDGPU_VM_MTYPE_UC:
578 return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
580 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
584 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level,
585 uint64_t *addr, uint64_t *flags)
587 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
588 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr);
589 BUG_ON(*addr & 0xFFFF00000000003FULL);
591 if (!adev->gmc.translate_further)
594 if (level == AMDGPU_VM_PDB1) {
595 /* Set the block fragment size */
596 if (!(*flags & AMDGPU_PDE_PTE))
597 *flags |= AMDGPU_PDE_BFS(0x9);
599 } else if (level == AMDGPU_VM_PDB0) {
600 if (*flags & AMDGPU_PDE_PTE)
601 *flags &= ~AMDGPU_PDE_PTE;
603 *flags |= AMDGPU_PTE_TF;
607 static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev,
608 struct amdgpu_bo_va_mapping *mapping,
611 struct amdgpu_bo *bo = mapping->bo_va->base.bo;
613 *flags &= ~AMDGPU_PTE_EXECUTABLE;
614 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
616 *flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
617 *flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
619 *flags &= ~AMDGPU_PTE_NOALLOC;
620 *flags |= (mapping->flags & AMDGPU_PTE_NOALLOC);
622 if (mapping->flags & AMDGPU_PTE_PRT) {
623 *flags |= AMDGPU_PTE_PRT;
624 *flags |= AMDGPU_PTE_SNOOPED;
625 *flags |= AMDGPU_PTE_LOG;
626 *flags |= AMDGPU_PTE_SYSTEM;
627 *flags &= ~AMDGPU_PTE_VALID;
630 if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
631 AMDGPU_GEM_CREATE_UNCACHED))
632 *flags = (*flags & ~AMDGPU_PTE_MTYPE_NV10_MASK) |
633 AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
636 static unsigned gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev)
638 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
641 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
642 size = AMDGPU_VBIOS_VGA_ALLOCATION;
647 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
648 pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH);
649 size = (REG_GET_FIELD(viewport,
650 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
651 REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) *
658 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = {
659 .flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb,
660 .flush_gpu_tlb_pasid = gmc_v10_0_flush_gpu_tlb_pasid,
661 .emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb,
662 .emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping,
663 .map_mtype = gmc_v10_0_map_mtype,
664 .get_vm_pde = gmc_v10_0_get_vm_pde,
665 .get_vm_pte = gmc_v10_0_get_vm_pte,
666 .get_vbios_fb_size = gmc_v10_0_get_vbios_fb_size,
669 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev)
671 if (adev->gmc.gmc_funcs == NULL)
672 adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs;
675 static void gmc_v10_0_set_umc_funcs(struct amdgpu_device *adev)
677 switch (adev->ip_versions[UMC_HWIP][0]) {
678 case IP_VERSION(8, 7, 0):
679 adev->umc.max_ras_err_cnt_per_query = UMC_V8_7_TOTAL_CHANNEL_NUM;
680 adev->umc.channel_inst_num = UMC_V8_7_CHANNEL_INSTANCE_NUM;
681 adev->umc.umc_inst_num = UMC_V8_7_UMC_INSTANCE_NUM;
682 adev->umc.channel_offs = UMC_V8_7_PER_CHANNEL_OFFSET_SIENNA;
683 adev->umc.channel_idx_tbl = &umc_v8_7_channel_idx_tbl[0][0];
684 adev->umc.ras = &umc_v8_7_ras;
690 amdgpu_ras_register_ras_block(adev, &adev->umc.ras->ras_block);
692 strcpy(adev->umc.ras->ras_block.ras_comm.name, "umc");
693 adev->umc.ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__UMC;
694 adev->umc.ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
695 adev->umc.ras_if = &adev->umc.ras->ras_block.ras_comm;
697 /* If don't define special ras_late_init function, use default ras_late_init */
698 if (!adev->umc.ras->ras_block.ras_late_init)
699 adev->umc.ras->ras_block.ras_late_init = amdgpu_umc_ras_late_init;
701 /* If not defined special ras_cb function, use default ras_cb */
702 if (!adev->umc.ras->ras_block.ras_cb)
703 adev->umc.ras->ras_block.ras_cb = amdgpu_umc_process_ras_data_cb;
708 static void gmc_v10_0_set_mmhub_funcs(struct amdgpu_device *adev)
710 switch (adev->ip_versions[MMHUB_HWIP][0]) {
711 case IP_VERSION(2, 3, 0):
712 case IP_VERSION(2, 4, 0):
713 case IP_VERSION(2, 4, 1):
714 adev->mmhub.funcs = &mmhub_v2_3_funcs;
717 adev->mmhub.funcs = &mmhub_v2_0_funcs;
722 static void gmc_v10_0_set_gfxhub_funcs(struct amdgpu_device *adev)
724 switch (adev->ip_versions[GC_HWIP][0]) {
725 case IP_VERSION(10, 3, 0):
726 case IP_VERSION(10, 3, 2):
727 case IP_VERSION(10, 3, 1):
728 case IP_VERSION(10, 3, 4):
729 case IP_VERSION(10, 3, 5):
730 case IP_VERSION(10, 3, 6):
731 case IP_VERSION(10, 3, 3):
732 case IP_VERSION(10, 3, 7):
733 adev->gfxhub.funcs = &gfxhub_v2_1_funcs;
736 adev->gfxhub.funcs = &gfxhub_v2_0_funcs;
742 static int gmc_v10_0_early_init(void *handle)
745 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
747 gmc_v10_0_set_mmhub_funcs(adev);
748 gmc_v10_0_set_gfxhub_funcs(adev);
749 gmc_v10_0_set_gmc_funcs(adev);
750 gmc_v10_0_set_irq_funcs(adev);
751 gmc_v10_0_set_umc_funcs(adev);
753 adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
754 adev->gmc.shared_aperture_end =
755 adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
756 adev->gmc.private_aperture_start = 0x1000000000000000ULL;
757 adev->gmc.private_aperture_end =
758 adev->gmc.private_aperture_start + (4ULL << 30) - 1;
760 r = amdgpu_gmc_ras_early_init(adev);
767 static int gmc_v10_0_late_init(void *handle)
769 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
772 r = amdgpu_gmc_allocate_vm_inv_eng(adev);
776 r = amdgpu_gmc_ras_late_init(adev);
780 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
783 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev,
784 struct amdgpu_gmc *mc)
788 base = adev->gfxhub.funcs->get_fb_location(adev);
790 /* add the xgmi offset of the physical node */
791 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
793 amdgpu_gmc_vram_location(adev, &adev->gmc, base);
794 amdgpu_gmc_gart_location(adev, mc);
795 amdgpu_gmc_agp_location(adev, mc);
797 /* base offset of vram pages */
798 adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev);
800 /* add the xgmi offset of the physical node */
801 adev->vm_manager.vram_base_offset +=
802 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
806 * gmc_v10_0_mc_init - initialize the memory controller driver params
808 * @adev: amdgpu_device pointer
810 * Look up the amount of vram, vram width, and decide how to place
811 * vram and gart within the GPU's physical address space.
812 * Returns 0 for success.
814 static int gmc_v10_0_mc_init(struct amdgpu_device *adev)
818 /* size in MB on si */
819 adev->gmc.mc_vram_size =
820 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
821 adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
823 if (!(adev->flags & AMD_IS_APU)) {
824 r = amdgpu_device_resize_fb_bar(adev);
828 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
829 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
832 if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) {
833 adev->gmc.aper_base = adev->gfxhub.funcs->get_mc_fb_offset(adev);
834 adev->gmc.aper_size = adev->gmc.real_vram_size;
838 /* In case the PCI BAR is larger than the actual amount of vram */
839 adev->gmc.visible_vram_size = adev->gmc.aper_size;
840 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
841 adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
843 /* set the gart size */
844 if (amdgpu_gart_size == -1) {
845 switch (adev->ip_versions[GC_HWIP][0]) {
847 adev->gmc.gart_size = 512ULL << 20;
849 case IP_VERSION(10, 3, 1): /* DCE SG support */
850 case IP_VERSION(10, 3, 3): /* DCE SG support */
851 case IP_VERSION(10, 3, 6): /* DCE SG support */
852 case IP_VERSION(10, 3, 7): /* DCE SG support */
853 adev->gmc.gart_size = 1024ULL << 20;
857 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
860 gmc_v10_0_vram_gtt_location(adev, &adev->gmc);
865 static int gmc_v10_0_gart_init(struct amdgpu_device *adev)
870 WARN(1, "NAVI10 PCIE GART already initialized\n");
874 /* Initialize common gart structure */
875 r = amdgpu_gart_init(adev);
879 adev->gart.table_size = adev->gart.num_gpu_pages * 8;
880 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) |
881 AMDGPU_PTE_EXECUTABLE;
883 return amdgpu_gart_table_vram_alloc(adev);
886 static int gmc_v10_0_sw_init(void *handle)
888 int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
889 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
891 adev->gfxhub.funcs->init(adev);
893 adev->mmhub.funcs->init(adev);
895 spin_lock_init(&adev->gmc.invalidate_lock);
897 if ((adev->flags & AMD_IS_APU) && amdgpu_emu_mode == 1) {
898 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_DDR4;
899 adev->gmc.vram_width = 64;
900 } else if (amdgpu_emu_mode == 1) {
901 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_GDDR6;
902 adev->gmc.vram_width = 1 * 128; /* numchan * chansize */
904 r = amdgpu_atomfirmware_get_vram_info(adev,
905 &vram_width, &vram_type, &vram_vendor);
906 adev->gmc.vram_width = vram_width;
908 adev->gmc.vram_type = vram_type;
909 adev->gmc.vram_vendor = vram_vendor;
912 switch (adev->ip_versions[GC_HWIP][0]) {
913 case IP_VERSION(10, 3, 0):
914 adev->gmc.mall_size = 128 * 1024 * 1024;
916 case IP_VERSION(10, 3, 2):
917 adev->gmc.mall_size = 96 * 1024 * 1024;
919 case IP_VERSION(10, 3, 4):
920 adev->gmc.mall_size = 32 * 1024 * 1024;
922 case IP_VERSION(10, 3, 5):
923 adev->gmc.mall_size = 16 * 1024 * 1024;
926 adev->gmc.mall_size = 0;
930 switch (adev->ip_versions[GC_HWIP][0]) {
931 case IP_VERSION(10, 1, 10):
932 case IP_VERSION(10, 1, 1):
933 case IP_VERSION(10, 1, 2):
934 case IP_VERSION(10, 1, 3):
935 case IP_VERSION(10, 1, 4):
936 case IP_VERSION(10, 3, 0):
937 case IP_VERSION(10, 3, 2):
938 case IP_VERSION(10, 3, 1):
939 case IP_VERSION(10, 3, 4):
940 case IP_VERSION(10, 3, 5):
941 case IP_VERSION(10, 3, 6):
942 case IP_VERSION(10, 3, 3):
943 case IP_VERSION(10, 3, 7):
944 adev->num_vmhubs = 2;
946 * To fulfill 4-level page support,
947 * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12,
948 * block size 512 (9bit)
950 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
956 /* This interrupt is VMC page fault.*/
957 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC,
958 VMC_1_0__SRCID__VM_FAULT,
959 &adev->gmc.vm_fault);
964 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2,
965 UTCL2_1_0__SRCID__FAULT,
966 &adev->gmc.vm_fault);
970 if (!amdgpu_sriov_vf(adev)) {
971 /* interrupt sent to DF. */
972 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
979 * Set the internal MC address mask This is the max address of the GPU's
980 * internal address space.
982 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
984 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
986 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
990 adev->need_swiotlb = drm_need_swiotlb(44);
992 r = gmc_v10_0_mc_init(adev);
996 amdgpu_gmc_get_vbios_allocations(adev);
999 r = amdgpu_bo_init(adev);
1003 r = gmc_v10_0_gart_init(adev);
1009 * VMID 0 is reserved for System
1010 * amdgpu graphics/compute will use VMIDs 1-7
1011 * amdkfd will use VMIDs 8-15
1013 adev->vm_manager.first_kfd_vmid = 8;
1015 amdgpu_vm_manager_init(adev);
1021 * gmc_v10_0_gart_fini - vm fini callback
1023 * @adev: amdgpu_device pointer
1025 * Tears down the driver GART/VM setup (CIK).
1027 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev)
1029 amdgpu_gart_table_vram_free(adev);
1032 static int gmc_v10_0_sw_fini(void *handle)
1034 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1036 amdgpu_vm_manager_fini(adev);
1037 gmc_v10_0_gart_fini(adev);
1038 amdgpu_gem_force_release(adev);
1039 amdgpu_bo_fini(adev);
1044 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev)
1049 * gmc_v10_0_gart_enable - gart enable
1051 * @adev: amdgpu_device pointer
1053 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev)
1058 if (adev->gart.bo == NULL) {
1059 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1063 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr);
1064 r = adev->gfxhub.funcs->gart_enable(adev);
1068 r = adev->mmhub.funcs->gart_enable(adev);
1072 adev->hdp.funcs->init_registers(adev);
1074 /* Flush HDP after it is initialized */
1075 adev->hdp.funcs->flush_hdp(adev, NULL);
1077 value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
1080 adev->gfxhub.funcs->set_fault_enable_default(adev, value);
1081 adev->mmhub.funcs->set_fault_enable_default(adev, value);
1082 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0);
1083 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB_0, 0);
1085 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1086 (unsigned)(adev->gmc.gart_size >> 20),
1087 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1092 static int gmc_v10_0_hw_init(void *handle)
1095 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1097 /* The sequence of these two function calls matters.*/
1098 gmc_v10_0_init_golden_registers(adev);
1101 * harvestable groups in gc_utcl2 need to be programmed before any GFX block
1102 * register setup within GMC, or else system hang when harvesting SA.
1104 if (adev->gfxhub.funcs && adev->gfxhub.funcs->utcl2_harvest)
1105 adev->gfxhub.funcs->utcl2_harvest(adev);
1107 r = gmc_v10_0_gart_enable(adev);
1111 if (amdgpu_emu_mode == 1) {
1112 r = amdgpu_gmc_vram_checking(adev);
1117 if (adev->umc.funcs && adev->umc.funcs->init_registers)
1118 adev->umc.funcs->init_registers(adev);
1124 * gmc_v10_0_gart_disable - gart disable
1126 * @adev: amdgpu_device pointer
1128 * This disables all VM page table.
1130 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev)
1132 adev->gfxhub.funcs->gart_disable(adev);
1133 adev->mmhub.funcs->gart_disable(adev);
1136 static int gmc_v10_0_hw_fini(void *handle)
1138 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1140 gmc_v10_0_gart_disable(adev);
1142 if (amdgpu_sriov_vf(adev)) {
1143 /* full access mode, so don't touch any GMC register */
1144 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1148 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1149 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1154 static int gmc_v10_0_suspend(void *handle)
1156 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1158 gmc_v10_0_hw_fini(adev);
1163 static int gmc_v10_0_resume(void *handle)
1166 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1168 r = gmc_v10_0_hw_init(adev);
1172 amdgpu_vmid_reset_all(adev);
1177 static bool gmc_v10_0_is_idle(void *handle)
1179 /* MC is always ready in GMC v10.*/
1183 static int gmc_v10_0_wait_for_idle(void *handle)
1185 /* There is no need to wait for MC idle in GMC v10.*/
1189 static int gmc_v10_0_soft_reset(void *handle)
1194 static int gmc_v10_0_set_clockgating_state(void *handle,
1195 enum amd_clockgating_state state)
1198 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1201 * The issue mmhub can't disconnect from DF with MMHUB clock gating being disabled
1202 * is a new problem observed at DF 3.0.3, however with the same suspend sequence not
1203 * seen any issue on the DF 3.0.2 series platform.
1205 if (adev->in_s0ix && adev->ip_versions[DF_HWIP][0] > IP_VERSION(3, 0, 2)) {
1206 dev_dbg(adev->dev, "keep mmhub clock gating being enabled for s0ix\n");
1210 r = adev->mmhub.funcs->set_clockgating(adev, state);
1214 if (adev->ip_versions[ATHUB_HWIP][0] >= IP_VERSION(2, 1, 0))
1215 return athub_v2_1_set_clockgating(adev, state);
1217 return athub_v2_0_set_clockgating(adev, state);
1220 static void gmc_v10_0_get_clockgating_state(void *handle, u64 *flags)
1222 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1224 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(10, 1, 3) ||
1225 adev->ip_versions[GC_HWIP][0] == IP_VERSION(10, 1, 4))
1228 adev->mmhub.funcs->get_clockgating(adev, flags);
1230 if (adev->ip_versions[ATHUB_HWIP][0] >= IP_VERSION(2, 1, 0))
1231 athub_v2_1_get_clockgating(adev, flags);
1233 athub_v2_0_get_clockgating(adev, flags);
1236 static int gmc_v10_0_set_powergating_state(void *handle,
1237 enum amd_powergating_state state)
1242 const struct amd_ip_funcs gmc_v10_0_ip_funcs = {
1243 .name = "gmc_v10_0",
1244 .early_init = gmc_v10_0_early_init,
1245 .late_init = gmc_v10_0_late_init,
1246 .sw_init = gmc_v10_0_sw_init,
1247 .sw_fini = gmc_v10_0_sw_fini,
1248 .hw_init = gmc_v10_0_hw_init,
1249 .hw_fini = gmc_v10_0_hw_fini,
1250 .suspend = gmc_v10_0_suspend,
1251 .resume = gmc_v10_0_resume,
1252 .is_idle = gmc_v10_0_is_idle,
1253 .wait_for_idle = gmc_v10_0_wait_for_idle,
1254 .soft_reset = gmc_v10_0_soft_reset,
1255 .set_clockgating_state = gmc_v10_0_set_clockgating_state,
1256 .set_powergating_state = gmc_v10_0_set_powergating_state,
1257 .get_clockgating_state = gmc_v10_0_get_clockgating_state,
1260 const struct amdgpu_ip_block_version gmc_v10_0_ip_block =
1262 .type = AMD_IP_BLOCK_TYPE_GMC,
1266 .funcs = &gmc_v10_0_ip_funcs,