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>
29 #include <asm/hypervisor.h>
33 #include "amdgpu_gmc.h"
34 #include "amdgpu_ras.h"
35 #include "amdgpu_xgmi.h"
37 #include <drm/drm_drv.h>
38 #include <drm/ttm/ttm_tt.h>
41 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0
43 * @adev: amdgpu_device pointer
45 * Allocate video memory for pdb0 and map it for CPU access
46 * Returns 0 for success, error for failure.
48 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev)
51 struct amdgpu_bo_param bp;
52 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
53 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21;
54 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) -1) >> pde0_page_shift;
56 memset(&bp, 0, sizeof(bp));
57 bp.size = PAGE_ALIGN((npdes + 1) * 8);
58 bp.byte_align = PAGE_SIZE;
59 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
60 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
61 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
62 bp.type = ttm_bo_type_kernel;
64 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
66 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo);
70 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false);
72 goto bo_reserve_failure;
74 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM);
77 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0);
81 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
85 amdgpu_bo_unpin(adev->gmc.pdb0_bo);
87 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
89 amdgpu_bo_unref(&adev->gmc.pdb0_bo);
94 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO
96 * @bo: the BO to get the PDE for
97 * @level: the level in the PD hirarchy
98 * @addr: resulting addr
99 * @flags: resulting flags
101 * Get the address and flags to be used for a PDE (Page Directory Entry).
103 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
104 uint64_t *addr, uint64_t *flags)
106 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
108 switch (bo->tbo.resource->mem_type) {
110 *addr = bo->tbo.ttm->dma_address[0];
113 *addr = amdgpu_bo_gpu_offset(bo);
119 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource);
120 amdgpu_gmc_get_vm_pde(adev, level, addr, flags);
124 * amdgpu_gmc_pd_addr - return the address of the root directory
126 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo)
128 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
131 /* TODO: move that into ASIC specific code */
132 if (adev->asic_type >= CHIP_VEGA10) {
133 uint64_t flags = AMDGPU_PTE_VALID;
135 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags);
138 pd_addr = amdgpu_bo_gpu_offset(bo);
144 * amdgpu_gmc_set_pte_pde - update the page tables using CPU
146 * @adev: amdgpu_device pointer
147 * @cpu_pt_addr: cpu address of the page table
148 * @gpu_page_idx: entry in the page table to update
149 * @addr: dst addr to write into pte/pde
150 * @flags: access flags
152 * Update the page tables using CPU.
154 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
155 uint32_t gpu_page_idx, uint64_t addr,
158 void __iomem *ptr = (void *)cpu_pt_addr;
162 * The following is for PTE only. GART does not have PDEs.
164 value = addr & 0x0000FFFFFFFFF000ULL;
166 writeq(value, ptr + (gpu_page_idx * 8));
172 * amdgpu_gmc_agp_addr - return the address in the AGP address space
174 * @bo: TTM BO which needs the address, must be in GTT domain
176 * Tries to figure out how to access the BO through the AGP aperture. Returns
177 * AMDGPU_BO_INVALID_OFFSET if that is not possible.
179 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo)
181 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
183 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached)
184 return AMDGPU_BO_INVALID_OFFSET;
186 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size)
187 return AMDGPU_BO_INVALID_OFFSET;
189 return adev->gmc.agp_start + bo->ttm->dma_address[0];
193 * amdgpu_gmc_vram_location - try to find VRAM location
195 * @adev: amdgpu device structure holding all necessary information
196 * @mc: memory controller structure holding memory information
197 * @base: base address at which to put VRAM
199 * Function will try to place VRAM at base address provided
202 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
205 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
207 mc->vram_start = base;
208 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
209 if (limit && limit < mc->real_vram_size)
210 mc->real_vram_size = limit;
212 if (mc->xgmi.num_physical_nodes == 0) {
213 mc->fb_start = mc->vram_start;
214 mc->fb_end = mc->vram_end;
216 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
217 mc->mc_vram_size >> 20, mc->vram_start,
218 mc->vram_end, mc->real_vram_size >> 20);
221 /** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture
223 * @adev: amdgpu device structure holding all necessary information
224 * @mc: memory controller structure holding memory information
226 * This function is only used if use GART for FB translation. In such
227 * case, we use sysvm aperture (vmid0 page tables) for both vram
228 * and gart (aka system memory) access.
230 * GPUVM (and our organization of vmid0 page tables) require sysvm
231 * aperture to be placed at a location aligned with 8 times of native
232 * page size. For example, if vm_context0_cntl.page_table_block_size
233 * is 12, then native page size is 8G (2M*2^12), sysvm should start
234 * with a 64G aligned address. For simplicity, we just put sysvm at
235 * address 0. So vram start at address 0 and gart is right after vram.
237 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
239 u64 hive_vram_start = 0;
240 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1;
241 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
242 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
243 mc->gart_start = hive_vram_end + 1;
244 mc->gart_end = mc->gart_start + mc->gart_size - 1;
245 mc->fb_start = hive_vram_start;
246 mc->fb_end = hive_vram_end;
247 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
248 mc->mc_vram_size >> 20, mc->vram_start,
249 mc->vram_end, mc->real_vram_size >> 20);
250 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
251 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
255 * amdgpu_gmc_gart_location - try to find GART location
257 * @adev: amdgpu device structure holding all necessary information
258 * @mc: memory controller structure holding memory information
260 * Function will place try to place GART before or after VRAM.
261 * If GART size is bigger than space left then we ajust GART size.
262 * Thus function will never fails.
264 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
266 const uint64_t four_gb = 0x100000000ULL;
267 u64 size_af, size_bf;
268 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/
269 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1);
271 /* VCE doesn't like it when BOs cross a 4GB segment, so align
272 * the GART base on a 4GB boundary as well.
274 size_bf = mc->fb_start;
275 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb);
277 if (mc->gart_size > max(size_bf, size_af)) {
278 dev_warn(adev->dev, "limiting GART\n");
279 mc->gart_size = max(size_bf, size_af);
282 if ((size_bf >= mc->gart_size && size_bf < size_af) ||
283 (size_af < mc->gart_size))
286 mc->gart_start = max_mc_address - mc->gart_size + 1;
288 mc->gart_start &= ~(four_gb - 1);
289 mc->gart_end = mc->gart_start + mc->gart_size - 1;
290 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
291 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
295 * amdgpu_gmc_agp_location - try to find AGP location
296 * @adev: amdgpu device structure holding all necessary information
297 * @mc: memory controller structure holding memory information
299 * Function will place try to find a place for the AGP BAR in the MC address
302 * AGP BAR will be assigned the largest available hole in the address space.
303 * Should be called after VRAM and GART locations are setup.
305 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
307 const uint64_t sixteen_gb = 1ULL << 34;
308 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1);
309 u64 size_af, size_bf;
311 if (amdgpu_sriov_vf(adev)) {
312 mc->agp_start = 0xffffffffffff;
319 if (mc->fb_start > mc->gart_start) {
320 size_bf = (mc->fb_start & sixteen_gb_mask) -
321 ALIGN(mc->gart_end + 1, sixteen_gb);
322 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb);
324 size_bf = mc->fb_start & sixteen_gb_mask;
325 size_af = (mc->gart_start & sixteen_gb_mask) -
326 ALIGN(mc->fb_end + 1, sixteen_gb);
329 if (size_bf > size_af) {
330 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask;
331 mc->agp_size = size_bf;
333 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb);
334 mc->agp_size = size_af;
337 mc->agp_end = mc->agp_start + mc->agp_size - 1;
338 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n",
339 mc->agp_size >> 20, mc->agp_start, mc->agp_end);
343 * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid
345 * @addr: 48 bit physical address, page aligned (36 significant bits)
346 * @pasid: 16 bit process address space identifier
348 static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid)
350 return addr << 4 | pasid;
354 * amdgpu_gmc_filter_faults - filter VM faults
356 * @adev: amdgpu device structure
357 * @ih: interrupt ring that the fault received from
358 * @addr: address of the VM fault
359 * @pasid: PASID of the process causing the fault
360 * @timestamp: timestamp of the fault
363 * True if the fault was filtered and should not be processed further.
364 * False if the fault is a new one and needs to be handled.
366 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
367 struct amdgpu_ih_ring *ih, uint64_t addr,
368 uint16_t pasid, uint64_t timestamp)
370 struct amdgpu_gmc *gmc = &adev->gmc;
371 uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid);
372 struct amdgpu_gmc_fault *fault;
375 /* Stale retry fault if timestamp goes backward */
376 if (amdgpu_ih_ts_after(timestamp, ih->processed_timestamp))
379 /* If we don't have space left in the ring buffer return immediately */
380 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) -
381 AMDGPU_GMC_FAULT_TIMEOUT;
382 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp)
385 /* Try to find the fault in the hash */
386 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
387 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
388 while (fault->timestamp >= stamp) {
391 if (atomic64_read(&fault->key) == key)
394 tmp = fault->timestamp;
395 fault = &gmc->fault_ring[fault->next];
397 /* Check if the entry was reused */
398 if (fault->timestamp >= tmp)
402 /* Add the fault to the ring */
403 fault = &gmc->fault_ring[gmc->last_fault];
404 atomic64_set(&fault->key, key);
405 fault->timestamp = timestamp;
407 /* And update the hash */
408 fault->next = gmc->fault_hash[hash].idx;
409 gmc->fault_hash[hash].idx = gmc->last_fault++;
414 * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter
416 * @adev: amdgpu device structure
417 * @addr: address of the VM fault
418 * @pasid: PASID of the process causing the fault
420 * Remove the address from fault filter, then future vm fault on this address
421 * will pass to retry fault handler to recover.
423 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
426 struct amdgpu_gmc *gmc = &adev->gmc;
427 uint64_t key = amdgpu_gmc_fault_key(addr, pasid);
428 struct amdgpu_gmc_fault *fault;
432 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
433 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
435 if (atomic64_cmpxchg(&fault->key, key, 0) == key)
438 tmp = fault->timestamp;
439 fault = &gmc->fault_ring[fault->next];
440 } while (fault->timestamp < tmp);
443 int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev)
445 if (!adev->gmc.xgmi.connected_to_cpu) {
446 adev->gmc.xgmi.ras = &xgmi_ras;
447 amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
448 adev->gmc.xgmi.ras_if = &adev->gmc.xgmi.ras->ras_block.ras_comm;
454 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
459 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
465 * The latest engine allocation on gfx9/10 is:
466 * Engine 2, 3: firmware
467 * Engine 0, 1, 4~16: amdgpu ring,
468 * subject to change when ring number changes
469 * Engine 17: Gart flushes
471 #define GFXHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
472 #define MMHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
474 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
476 struct amdgpu_ring *ring;
477 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] =
478 {GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP,
479 GFXHUB_FREE_VM_INV_ENGS_BITMAP};
481 unsigned vmhub, inv_eng;
483 if (adev->enable_mes) {
484 /* reserve engine 5 for firmware */
485 for (vmhub = 0; vmhub < AMDGPU_MAX_VMHUBS; vmhub++)
486 vm_inv_engs[vmhub] &= ~(1 << 5);
489 for (i = 0; i < adev->num_rings; ++i) {
490 ring = adev->rings[i];
491 vmhub = ring->funcs->vmhub;
493 if (ring == &adev->mes.ring)
496 inv_eng = ffs(vm_inv_engs[vmhub]);
498 dev_err(adev->dev, "no VM inv eng for ring %s\n",
503 ring->vm_inv_eng = inv_eng - 1;
504 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
506 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
507 ring->name, ring->vm_inv_eng, ring->funcs->vmhub);
514 * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ
515 * @adev: amdgpu_device pointer
517 * Check and set if an the device @adev supports Trusted Memory
520 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev)
522 switch (adev->ip_versions[GC_HWIP][0]) {
524 case IP_VERSION(9, 2, 2):
525 case IP_VERSION(9, 1, 0):
526 /* RENOIR looks like RAVEN */
527 case IP_VERSION(9, 3, 0):
529 case IP_VERSION(10, 3, 7):
530 if (amdgpu_tmz == 0) {
531 adev->gmc.tmz_enabled = false;
533 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n");
535 adev->gmc.tmz_enabled = true;
537 "Trusted Memory Zone (TMZ) feature enabled\n");
540 case IP_VERSION(10, 1, 10):
541 case IP_VERSION(10, 1, 1):
542 case IP_VERSION(10, 1, 2):
543 case IP_VERSION(10, 1, 3):
544 case IP_VERSION(10, 3, 0):
545 case IP_VERSION(10, 3, 2):
546 case IP_VERSION(10, 3, 4):
547 case IP_VERSION(10, 3, 5):
549 case IP_VERSION(10, 3, 1):
551 case IP_VERSION(10, 3, 3):
552 case IP_VERSION(11, 0, 1):
553 case IP_VERSION(11, 0, 4):
554 /* Don't enable it by default yet.
556 if (amdgpu_tmz < 1) {
557 adev->gmc.tmz_enabled = false;
559 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n");
561 adev->gmc.tmz_enabled = true;
563 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n");
567 adev->gmc.tmz_enabled = false;
569 "Trusted Memory Zone (TMZ) feature not supported\n");
575 * amdgpu_gmc_noretry_set -- set per asic noretry defaults
576 * @adev: amdgpu_device pointer
578 * Set a per asic default for the no-retry parameter.
581 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev)
583 struct amdgpu_gmc *gmc = &adev->gmc;
584 uint32_t gc_ver = adev->ip_versions[GC_HWIP][0];
585 bool noretry_default = (gc_ver == IP_VERSION(9, 0, 1) ||
586 gc_ver == IP_VERSION(9, 3, 0) ||
587 gc_ver == IP_VERSION(9, 4, 0) ||
588 gc_ver == IP_VERSION(9, 4, 1) ||
589 gc_ver == IP_VERSION(9, 4, 2) ||
590 gc_ver >= IP_VERSION(10, 3, 0));
592 gmc->noretry = (amdgpu_noretry == -1) ? noretry_default : amdgpu_noretry;
595 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
598 struct amdgpu_vmhub *hub;
601 hub = &adev->vmhub[hub_type];
602 for (i = 0; i < 16; i++) {
603 reg = hub->vm_context0_cntl + hub->ctx_distance * i;
605 tmp = (hub_type == AMDGPU_GFXHUB_0) ?
606 RREG32_SOC15_IP(GC, reg) :
607 RREG32_SOC15_IP(MMHUB, reg);
610 tmp |= hub->vm_cntx_cntl_vm_fault;
612 tmp &= ~hub->vm_cntx_cntl_vm_fault;
614 (hub_type == AMDGPU_GFXHUB_0) ?
615 WREG32_SOC15_IP(GC, reg, tmp) :
616 WREG32_SOC15_IP(MMHUB, reg, tmp);
620 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev)
625 * Some ASICs need to reserve a region of video memory to avoid access
628 adev->mman.stolen_reserved_offset = 0;
629 adev->mman.stolen_reserved_size = 0;
633 * Currently there is a bug where some memory client outside
634 * of the driver writes to first 8M of VRAM on S3 resume,
635 * this overrides GART which by default gets placed in first 8M and
636 * causes VM_FAULTS once GTT is accessed.
637 * Keep the stolen memory reservation until the while this is not solved.
639 switch (adev->asic_type) {
641 adev->mman.keep_stolen_vga_memory = true;
643 * VEGA10 SRIOV VF with MS_HYPERV host needs some firmware reserved area.
646 if (amdgpu_sriov_vf(adev) && hypervisor_is_type(X86_HYPER_MS_HYPERV)) {
647 adev->mman.stolen_reserved_offset = 0x500000;
648 adev->mman.stolen_reserved_size = 0x200000;
654 adev->mman.keep_stolen_vga_memory = true;
656 case CHIP_YELLOW_CARP:
657 if (amdgpu_discovery == 0) {
658 adev->mman.stolen_reserved_offset = 0x1ffb0000;
659 adev->mman.stolen_reserved_size = 64 * PAGE_SIZE;
663 adev->mman.keep_stolen_vga_memory = false;
667 if (amdgpu_sriov_vf(adev) ||
668 !amdgpu_device_has_display_hardware(adev)) {
671 size = amdgpu_gmc_get_vbios_fb_size(adev);
673 if (adev->mman.keep_stolen_vga_memory)
674 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION);
677 /* set to 0 if the pre-OS buffer uses up most of vram */
678 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
681 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) {
682 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION;
683 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size;
685 adev->mman.stolen_vga_size = size;
686 adev->mman.stolen_extended_size = 0;
691 * amdgpu_gmc_init_pdb0 - initialize PDB0
693 * @adev: amdgpu_device pointer
695 * This function is only used when GART page table is used
696 * for FB address translatioin. In such a case, we construct
697 * a 2-level system VM page table: PDB0->PTB, to cover both
698 * VRAM of the hive and system memory.
700 * PDB0 is static, initialized once on driver initialization.
701 * The first n entries of PDB0 are used as PTE by setting
702 * P bit to 1, pointing to VRAM. The n+1'th entry points
703 * to a big PTB covering system memory.
706 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
709 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW?
710 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M
712 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
713 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21;
714 u64 vram_addr = adev->vm_manager.vram_base_offset -
715 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
716 u64 vram_end = vram_addr + vram_size;
717 u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo);
720 if (!drm_dev_enter(adev_to_drm(adev), &idx))
723 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
724 flags |= AMDGPU_PTE_WRITEABLE;
725 flags |= AMDGPU_PTE_SNOOPED;
726 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1));
727 flags |= AMDGPU_PDE_PTE;
729 /* The first n PDE0 entries are used as PTE,
732 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size)
733 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags);
735 /* The n+1'th PDE0 entry points to a huge
736 * PTB who has more than 512 entries each
737 * pointing to a 4K system page
739 flags = AMDGPU_PTE_VALID;
740 flags |= AMDGPU_PDE_BFS(0) | AMDGPU_PTE_SNOOPED;
741 /* Requires gart_ptb_gpu_pa to be 4K aligned */
742 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags);
747 * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC
750 * @adev: amdgpu_device pointer
751 * @mc_addr: MC address of buffer
753 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr)
755 return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset;
759 * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from
762 * @adev: amdgpu_device pointer
763 * @bo: amdgpu buffer object
765 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
767 return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo));
771 * amdgpu_gmc_vram_cpu_pa - calculate vram buffer object's physical address
774 * @adev: amdgpu_device pointer
775 * @bo: amdgpu buffer object
777 uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
779 return amdgpu_bo_gpu_offset(bo) - adev->gmc.vram_start + adev->gmc.aper_base;
782 int amdgpu_gmc_vram_checking(struct amdgpu_device *adev)
784 struct amdgpu_bo *vram_bo = NULL;
785 uint64_t vram_gpu = 0;
786 void *vram_ptr = NULL;
788 int ret, size = 0x100000;
791 ret = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
792 AMDGPU_GEM_DOMAIN_VRAM,
799 memset(vram_ptr, 0x86, size);
800 memset(cptr, 0x86, 10);
803 * Check the start, the mid, and the end of the memory if the content of
804 * each byte is the pattern "0x86". If yes, we suppose the vram bo is
807 * Note: If check the each byte of whole 1M bo, it will cost too many
808 * seconds, so here, we just pick up three parts for emulation.
810 ret = memcmp(vram_ptr, cptr, 10);
814 ret = memcmp(vram_ptr + (size / 2), cptr, 10);
818 ret = memcmp(vram_ptr + size - 10, cptr, 10);
822 amdgpu_bo_free_kernel(&vram_bo, &vram_gpu,