2 * Copyright 2018 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
27 #include <linux/io-64-nonatomic-lo-hi.h>
30 #include "amdgpu_gmc.h"
31 #include "amdgpu_ras.h"
32 #include "amdgpu_xgmi.h"
34 #include <drm/drm_drv.h>
37 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0
39 * @adev: amdgpu_device pointer
41 * Allocate video memory for pdb0 and map it for CPU access
42 * Returns 0 for success, error for failure.
44 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev)
47 struct amdgpu_bo_param bp;
48 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
49 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21;
50 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) -1) >> pde0_page_shift;
52 memset(&bp, 0, sizeof(bp));
53 bp.size = PAGE_ALIGN((npdes + 1) * 8);
54 bp.byte_align = PAGE_SIZE;
55 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
56 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
57 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
58 bp.type = ttm_bo_type_kernel;
60 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
62 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo);
66 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false);
68 goto bo_reserve_failure;
70 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM);
73 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0);
77 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
81 amdgpu_bo_unpin(adev->gmc.pdb0_bo);
83 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
85 amdgpu_bo_unref(&adev->gmc.pdb0_bo);
90 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO
92 * @bo: the BO to get the PDE for
93 * @level: the level in the PD hirarchy
94 * @addr: resulting addr
95 * @flags: resulting flags
97 * Get the address and flags to be used for a PDE (Page Directory Entry).
99 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
100 uint64_t *addr, uint64_t *flags)
102 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
104 switch (bo->tbo.resource->mem_type) {
106 *addr = bo->tbo.ttm->dma_address[0];
109 *addr = amdgpu_bo_gpu_offset(bo);
115 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource);
116 amdgpu_gmc_get_vm_pde(adev, level, addr, flags);
120 * amdgpu_gmc_pd_addr - return the address of the root directory
122 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo)
124 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
127 /* TODO: move that into ASIC specific code */
128 if (adev->asic_type >= CHIP_VEGA10) {
129 uint64_t flags = AMDGPU_PTE_VALID;
131 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags);
134 pd_addr = amdgpu_bo_gpu_offset(bo);
140 * amdgpu_gmc_set_pte_pde - update the page tables using CPU
142 * @adev: amdgpu_device pointer
143 * @cpu_pt_addr: cpu address of the page table
144 * @gpu_page_idx: entry in the page table to update
145 * @addr: dst addr to write into pte/pde
146 * @flags: access flags
148 * Update the page tables using CPU.
150 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
151 uint32_t gpu_page_idx, uint64_t addr,
154 void __iomem *ptr = (void *)cpu_pt_addr;
158 * The following is for PTE only. GART does not have PDEs.
160 value = addr & 0x0000FFFFFFFFF000ULL;
162 writeq(value, ptr + (gpu_page_idx * 8));
168 * amdgpu_gmc_agp_addr - return the address in the AGP address space
170 * @bo: TTM BO which needs the address, must be in GTT domain
172 * Tries to figure out how to access the BO through the AGP aperture. Returns
173 * AMDGPU_BO_INVALID_OFFSET if that is not possible.
175 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo)
177 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
179 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached)
180 return AMDGPU_BO_INVALID_OFFSET;
182 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size)
183 return AMDGPU_BO_INVALID_OFFSET;
185 return adev->gmc.agp_start + bo->ttm->dma_address[0];
189 * amdgpu_gmc_vram_location - try to find VRAM location
191 * @adev: amdgpu device structure holding all necessary information
192 * @mc: memory controller structure holding memory information
193 * @base: base address at which to put VRAM
195 * Function will try to place VRAM at base address provided
198 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
201 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
203 mc->vram_start = base;
204 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
205 if (limit && limit < mc->real_vram_size)
206 mc->real_vram_size = limit;
208 if (mc->xgmi.num_physical_nodes == 0) {
209 mc->fb_start = mc->vram_start;
210 mc->fb_end = mc->vram_end;
212 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
213 mc->mc_vram_size >> 20, mc->vram_start,
214 mc->vram_end, mc->real_vram_size >> 20);
217 /** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture
219 * @adev: amdgpu device structure holding all necessary information
220 * @mc: memory controller structure holding memory information
222 * This function is only used if use GART for FB translation. In such
223 * case, we use sysvm aperture (vmid0 page tables) for both vram
224 * and gart (aka system memory) access.
226 * GPUVM (and our organization of vmid0 page tables) require sysvm
227 * aperture to be placed at a location aligned with 8 times of native
228 * page size. For example, if vm_context0_cntl.page_table_block_size
229 * is 12, then native page size is 8G (2M*2^12), sysvm should start
230 * with a 64G aligned address. For simplicity, we just put sysvm at
231 * address 0. So vram start at address 0 and gart is right after vram.
233 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
235 u64 hive_vram_start = 0;
236 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1;
237 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
238 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
239 mc->gart_start = hive_vram_end + 1;
240 mc->gart_end = mc->gart_start + mc->gart_size - 1;
241 mc->fb_start = hive_vram_start;
242 mc->fb_end = hive_vram_end;
243 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
244 mc->mc_vram_size >> 20, mc->vram_start,
245 mc->vram_end, mc->real_vram_size >> 20);
246 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
247 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
251 * amdgpu_gmc_gart_location - try to find GART location
253 * @adev: amdgpu device structure holding all necessary information
254 * @mc: memory controller structure holding memory information
256 * Function will place try to place GART before or after VRAM.
257 * If GART size is bigger than space left then we ajust GART size.
258 * Thus function will never fails.
260 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
262 const uint64_t four_gb = 0x100000000ULL;
263 u64 size_af, size_bf;
264 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/
265 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1);
267 /* VCE doesn't like it when BOs cross a 4GB segment, so align
268 * the GART base on a 4GB boundary as well.
270 size_bf = mc->fb_start;
271 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb);
273 if (mc->gart_size > max(size_bf, size_af)) {
274 dev_warn(adev->dev, "limiting GART\n");
275 mc->gart_size = max(size_bf, size_af);
278 if ((size_bf >= mc->gart_size && size_bf < size_af) ||
279 (size_af < mc->gart_size))
282 mc->gart_start = max_mc_address - mc->gart_size + 1;
284 mc->gart_start &= ~(four_gb - 1);
285 mc->gart_end = mc->gart_start + mc->gart_size - 1;
286 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
287 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
291 * amdgpu_gmc_agp_location - try to find AGP location
292 * @adev: amdgpu device structure holding all necessary information
293 * @mc: memory controller structure holding memory information
295 * Function will place try to find a place for the AGP BAR in the MC address
298 * AGP BAR will be assigned the largest available hole in the address space.
299 * Should be called after VRAM and GART locations are setup.
301 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
303 const uint64_t sixteen_gb = 1ULL << 34;
304 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1);
305 u64 size_af, size_bf;
307 if (amdgpu_sriov_vf(adev)) {
308 mc->agp_start = 0xffffffffffff;
315 if (mc->fb_start > mc->gart_start) {
316 size_bf = (mc->fb_start & sixteen_gb_mask) -
317 ALIGN(mc->gart_end + 1, sixteen_gb);
318 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb);
320 size_bf = mc->fb_start & sixteen_gb_mask;
321 size_af = (mc->gart_start & sixteen_gb_mask) -
322 ALIGN(mc->fb_end + 1, sixteen_gb);
325 if (size_bf > size_af) {
326 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask;
327 mc->agp_size = size_bf;
329 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb);
330 mc->agp_size = size_af;
333 mc->agp_end = mc->agp_start + mc->agp_size - 1;
334 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n",
335 mc->agp_size >> 20, mc->agp_start, mc->agp_end);
339 * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid
341 * @addr: 48 bit physical address, page aligned (36 significant bits)
342 * @pasid: 16 bit process address space identifier
344 static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid)
346 return addr << 4 | pasid;
350 * amdgpu_gmc_filter_faults - filter VM faults
352 * @adev: amdgpu device structure
353 * @ih: interrupt ring that the fault received from
354 * @addr: address of the VM fault
355 * @pasid: PASID of the process causing the fault
356 * @timestamp: timestamp of the fault
359 * True if the fault was filtered and should not be processed further.
360 * False if the fault is a new one and needs to be handled.
362 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
363 struct amdgpu_ih_ring *ih, uint64_t addr,
364 uint16_t pasid, uint64_t timestamp)
366 struct amdgpu_gmc *gmc = &adev->gmc;
367 uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid);
368 struct amdgpu_gmc_fault *fault;
371 /* Stale retry fault if timestamp goes backward */
372 if (amdgpu_ih_ts_after(timestamp, ih->processed_timestamp))
375 /* If we don't have space left in the ring buffer return immediately */
376 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) -
377 AMDGPU_GMC_FAULT_TIMEOUT;
378 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp)
381 /* Try to find the fault in the hash */
382 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
383 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
384 while (fault->timestamp >= stamp) {
387 if (atomic64_read(&fault->key) == key)
390 tmp = fault->timestamp;
391 fault = &gmc->fault_ring[fault->next];
393 /* Check if the entry was reused */
394 if (fault->timestamp >= tmp)
398 /* Add the fault to the ring */
399 fault = &gmc->fault_ring[gmc->last_fault];
400 atomic64_set(&fault->key, key);
401 fault->timestamp = timestamp;
403 /* And update the hash */
404 fault->next = gmc->fault_hash[hash].idx;
405 gmc->fault_hash[hash].idx = gmc->last_fault++;
410 * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter
412 * @adev: amdgpu device structure
413 * @addr: address of the VM fault
414 * @pasid: PASID of the process causing the fault
416 * Remove the address from fault filter, then future vm fault on this address
417 * will pass to retry fault handler to recover.
419 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
422 struct amdgpu_gmc *gmc = &adev->gmc;
423 uint64_t key = amdgpu_gmc_fault_key(addr, pasid);
424 struct amdgpu_gmc_fault *fault;
428 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
429 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
431 if (atomic64_cmpxchg(&fault->key, key, 0) == key)
434 tmp = fault->timestamp;
435 fault = &gmc->fault_ring[fault->next];
436 } while (fault->timestamp < tmp);
439 int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev)
441 if (!adev->gmc.xgmi.connected_to_cpu) {
442 adev->gmc.xgmi.ras = &xgmi_ras;
443 amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
444 adev->gmc.xgmi.ras_if = &adev->gmc.xgmi.ras->ras_block.ras_comm;
450 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
455 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
461 * The latest engine allocation on gfx9/10 is:
462 * Engine 2, 3: firmware
463 * Engine 0, 1, 4~16: amdgpu ring,
464 * subject to change when ring number changes
465 * Engine 17: Gart flushes
467 #define GFXHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
468 #define MMHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
470 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
472 struct amdgpu_ring *ring;
473 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] =
474 {GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP,
475 GFXHUB_FREE_VM_INV_ENGS_BITMAP};
477 unsigned vmhub, inv_eng;
479 for (i = 0; i < adev->num_rings; ++i) {
480 ring = adev->rings[i];
481 vmhub = ring->funcs->vmhub;
483 if (ring == &adev->mes.ring)
486 inv_eng = ffs(vm_inv_engs[vmhub]);
488 dev_err(adev->dev, "no VM inv eng for ring %s\n",
493 ring->vm_inv_eng = inv_eng - 1;
494 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
496 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
497 ring->name, ring->vm_inv_eng, ring->funcs->vmhub);
504 * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ
505 * @adev: amdgpu_device pointer
507 * Check and set if an the device @adev supports Trusted Memory
510 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev)
512 switch (adev->asic_type) {
515 if (amdgpu_tmz == 0) {
516 adev->gmc.tmz_enabled = false;
518 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n");
520 adev->gmc.tmz_enabled = true;
522 "Trusted Memory Zone (TMZ) feature enabled\n");
529 case CHIP_YELLOW_CARP:
530 case CHIP_IP_DISCOVERY:
531 /* Don't enable it by default yet.
533 if (amdgpu_tmz < 1) {
534 adev->gmc.tmz_enabled = false;
536 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n");
538 adev->gmc.tmz_enabled = true;
540 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n");
544 adev->gmc.tmz_enabled = false;
546 "Trusted Memory Zone (TMZ) feature not supported\n");
552 * amdgpu_gmc_noretry_set -- set per asic noretry defaults
553 * @adev: amdgpu_device pointer
555 * Set a per asic default for the no-retry parameter.
558 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev)
560 struct amdgpu_gmc *gmc = &adev->gmc;
562 switch (adev->ip_versions[GC_HWIP][0]) {
563 case IP_VERSION(9, 0, 1):
564 case IP_VERSION(9, 3, 0):
565 case IP_VERSION(9, 4, 0):
566 case IP_VERSION(9, 4, 1):
567 case IP_VERSION(9, 4, 2):
568 case IP_VERSION(10, 3, 3):
569 case IP_VERSION(10, 3, 4):
570 case IP_VERSION(10, 3, 5):
571 case IP_VERSION(10, 3, 6):
572 case IP_VERSION(10, 3, 7):
574 * noretry = 0 will cause kfd page fault tests fail
575 * for some ASICs, so set default to 1 for these ASICs.
577 if (amdgpu_noretry == -1)
580 gmc->noretry = amdgpu_noretry;
583 /* Raven currently has issues with noretry
584 * regardless of what we decide for other
585 * asics, we should leave raven with
586 * noretry = 0 until we root cause the
589 * default this to 0 for now, but we may want
590 * to change this in the future for certain
591 * GPUs as it can increase performance in
594 if (amdgpu_noretry == -1)
597 gmc->noretry = amdgpu_noretry;
602 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
605 struct amdgpu_vmhub *hub;
608 hub = &adev->vmhub[hub_type];
609 for (i = 0; i < 16; i++) {
610 reg = hub->vm_context0_cntl + hub->ctx_distance * i;
612 tmp = (hub_type == AMDGPU_GFXHUB_0) ?
613 RREG32_SOC15_IP(GC, reg) :
614 RREG32_SOC15_IP(MMHUB, reg);
617 tmp |= hub->vm_cntx_cntl_vm_fault;
619 tmp &= ~hub->vm_cntx_cntl_vm_fault;
621 (hub_type == AMDGPU_GFXHUB_0) ?
622 WREG32_SOC15_IP(GC, reg, tmp) :
623 WREG32_SOC15_IP(MMHUB, reg, tmp);
627 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev)
632 * Some ASICs need to reserve a region of video memory to avoid access
635 adev->mman.stolen_reserved_offset = 0;
636 adev->mman.stolen_reserved_size = 0;
640 * Currently there is a bug where some memory client outside
641 * of the driver writes to first 8M of VRAM on S3 resume,
642 * this overrides GART which by default gets placed in first 8M and
643 * causes VM_FAULTS once GTT is accessed.
644 * Keep the stolen memory reservation until the while this is not solved.
646 switch (adev->asic_type) {
648 adev->mman.keep_stolen_vga_memory = true;
650 * VEGA10 SRIOV VF needs some firmware reserved area.
652 if (amdgpu_sriov_vf(adev)) {
653 adev->mman.stolen_reserved_offset = 0x100000;
654 adev->mman.stolen_reserved_size = 0x600000;
659 adev->mman.keep_stolen_vga_memory = true;
661 case CHIP_YELLOW_CARP:
662 if (amdgpu_discovery == 0) {
663 adev->mman.stolen_reserved_offset = 0x1ffb0000;
664 adev->mman.stolen_reserved_size = 64 * PAGE_SIZE;
668 adev->mman.keep_stolen_vga_memory = false;
672 if (amdgpu_sriov_vf(adev) ||
673 !amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_DCE)) {
676 size = amdgpu_gmc_get_vbios_fb_size(adev);
678 if (adev->mman.keep_stolen_vga_memory)
679 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION);
682 /* set to 0 if the pre-OS buffer uses up most of vram */
683 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
686 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) {
687 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION;
688 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size;
690 adev->mman.stolen_vga_size = size;
691 adev->mman.stolen_extended_size = 0;
696 * amdgpu_gmc_init_pdb0 - initialize PDB0
698 * @adev: amdgpu_device pointer
700 * This function is only used when GART page table is used
701 * for FB address translatioin. In such a case, we construct
702 * a 2-level system VM page table: PDB0->PTB, to cover both
703 * VRAM of the hive and system memory.
705 * PDB0 is static, initialized once on driver initialization.
706 * The first n entries of PDB0 are used as PTE by setting
707 * P bit to 1, pointing to VRAM. The n+1'th entry points
708 * to a big PTB covering system memory.
711 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
714 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW?
715 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M
717 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
718 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21;
719 u64 vram_addr = adev->vm_manager.vram_base_offset -
720 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
721 u64 vram_end = vram_addr + vram_size;
722 u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo);
725 if (!drm_dev_enter(adev_to_drm(adev), &idx))
728 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
729 flags |= AMDGPU_PTE_WRITEABLE;
730 flags |= AMDGPU_PTE_SNOOPED;
731 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1));
732 flags |= AMDGPU_PDE_PTE;
734 /* The first n PDE0 entries are used as PTE,
737 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size)
738 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags);
740 /* The n+1'th PDE0 entry points to a huge
741 * PTB who has more than 512 entries each
742 * pointing to a 4K system page
744 flags = AMDGPU_PTE_VALID;
745 flags |= AMDGPU_PDE_BFS(0) | AMDGPU_PTE_SNOOPED;
746 /* Requires gart_ptb_gpu_pa to be 4K aligned */
747 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags);
752 * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC
755 * @adev: amdgpu_device pointer
756 * @mc_addr: MC address of buffer
758 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr)
760 return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset;
764 * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from
767 * @adev: amdgpu_device pointer
768 * @bo: amdgpu buffer object
770 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
772 return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo));
776 * amdgpu_gmc_vram_cpu_pa - calculate vram buffer object's physical address
779 * @adev: amdgpu_device pointer
780 * @bo: amdgpu buffer object
782 uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
784 return amdgpu_bo_gpu_offset(bo) - adev->gmc.vram_start + adev->gmc.aper_base;
787 int amdgpu_gmc_vram_checking(struct amdgpu_device *adev)
789 struct amdgpu_bo *vram_bo = NULL;
790 uint64_t vram_gpu = 0;
791 void *vram_ptr = NULL;
793 int ret, size = 0x100000;
796 ret = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
797 AMDGPU_GEM_DOMAIN_VRAM,
804 memset(vram_ptr, 0x86, size);
805 memset(cptr, 0x86, 10);
808 * Check the start, the mid, and the end of the memory if the content of
809 * each byte is the pattern "0x86". If yes, we suppose the vram bo is
812 * Note: If check the each byte of whole 1M bo, it will cost too many
813 * seconds, so here, we just pick up three parts for emulation.
815 ret = memcmp(vram_ptr, cptr, 10);
819 ret = memcmp(vram_ptr + (size / 2), cptr, 10);
823 ret = memcmp(vram_ptr + size - 10, cptr, 10);
827 amdgpu_bo_free_kernel(&vram_bo, &vram_gpu,