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"
35 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0
37 * @adev: amdgpu_device pointer
39 * Allocate video memory for pdb0 and map it for CPU access
40 * Returns 0 for success, error for failure.
42 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev)
45 struct amdgpu_bo_param bp;
46 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
47 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21;
48 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) -1) >> pde0_page_shift;
50 memset(&bp, 0, sizeof(bp));
51 bp.size = PAGE_ALIGN((npdes + 1) * 8);
52 bp.byte_align = PAGE_SIZE;
53 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
54 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
55 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
56 bp.type = ttm_bo_type_kernel;
58 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo);
62 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false);
64 goto bo_reserve_failure;
66 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM);
69 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0);
73 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
77 amdgpu_bo_unpin(adev->gmc.pdb0_bo);
79 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
81 amdgpu_bo_unref(&adev->gmc.pdb0_bo);
86 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO
88 * @bo: the BO to get the PDE for
89 * @level: the level in the PD hirarchy
90 * @addr: resulting addr
91 * @flags: resulting flags
93 * Get the address and flags to be used for a PDE (Page Directory Entry).
95 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
96 uint64_t *addr, uint64_t *flags)
98 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
100 switch (bo->tbo.mem.mem_type) {
102 *addr = bo->tbo.ttm->dma_address[0];
105 *addr = amdgpu_bo_gpu_offset(bo);
111 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, &bo->tbo.mem);
112 amdgpu_gmc_get_vm_pde(adev, level, addr, flags);
116 * amdgpu_gmc_pd_addr - return the address of the root directory
118 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo)
120 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
123 /* TODO: move that into ASIC specific code */
124 if (adev->asic_type >= CHIP_VEGA10) {
125 uint64_t flags = AMDGPU_PTE_VALID;
127 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags);
130 pd_addr = amdgpu_bo_gpu_offset(bo);
136 * amdgpu_gmc_set_pte_pde - update the page tables using CPU
138 * @adev: amdgpu_device pointer
139 * @cpu_pt_addr: cpu address of the page table
140 * @gpu_page_idx: entry in the page table to update
141 * @addr: dst addr to write into pte/pde
142 * @flags: access flags
144 * Update the page tables using CPU.
146 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
147 uint32_t gpu_page_idx, uint64_t addr,
150 void __iomem *ptr = (void *)cpu_pt_addr;
154 * The following is for PTE only. GART does not have PDEs.
156 value = addr & 0x0000FFFFFFFFF000ULL;
158 writeq(value, ptr + (gpu_page_idx * 8));
163 * amdgpu_gmc_agp_addr - return the address in the AGP address space
165 * @bo: TTM BO which needs the address, must be in GTT domain
167 * Tries to figure out how to access the BO through the AGP aperture. Returns
168 * AMDGPU_BO_INVALID_OFFSET if that is not possible.
170 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo)
172 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
174 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached)
175 return AMDGPU_BO_INVALID_OFFSET;
177 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size)
178 return AMDGPU_BO_INVALID_OFFSET;
180 return adev->gmc.agp_start + bo->ttm->dma_address[0];
184 * amdgpu_gmc_vram_location - try to find VRAM location
186 * @adev: amdgpu device structure holding all necessary information
187 * @mc: memory controller structure holding memory information
188 * @base: base address at which to put VRAM
190 * Function will try to place VRAM at base address provided
193 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
196 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
198 mc->vram_start = base;
199 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
200 if (limit && limit < mc->real_vram_size)
201 mc->real_vram_size = limit;
203 if (mc->xgmi.num_physical_nodes == 0) {
204 mc->fb_start = mc->vram_start;
205 mc->fb_end = mc->vram_end;
207 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
208 mc->mc_vram_size >> 20, mc->vram_start,
209 mc->vram_end, mc->real_vram_size >> 20);
212 /** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture
214 * @adev: amdgpu device structure holding all necessary information
215 * @mc: memory controller structure holding memory information
217 * This function is only used if use GART for FB translation. In such
218 * case, we use sysvm aperture (vmid0 page tables) for both vram
219 * and gart (aka system memory) access.
221 * GPUVM (and our organization of vmid0 page tables) require sysvm
222 * aperture to be placed at a location aligned with 8 times of native
223 * page size. For example, if vm_context0_cntl.page_table_block_size
224 * is 12, then native page size is 8G (2M*2^12), sysvm should start
225 * with a 64G aligned address. For simplicity, we just put sysvm at
226 * address 0. So vram start at address 0 and gart is right after vram.
228 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
230 u64 hive_vram_start = 0;
231 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1;
232 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
233 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
234 mc->gart_start = hive_vram_end + 1;
235 mc->gart_end = mc->gart_start + mc->gart_size - 1;
236 mc->fb_start = hive_vram_start;
237 mc->fb_end = hive_vram_end;
238 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
239 mc->mc_vram_size >> 20, mc->vram_start,
240 mc->vram_end, mc->real_vram_size >> 20);
241 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
242 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
246 * amdgpu_gmc_gart_location - try to find GART location
248 * @adev: amdgpu device structure holding all necessary information
249 * @mc: memory controller structure holding memory information
251 * Function will place try to place GART before or after VRAM.
252 * If GART size is bigger than space left then we ajust GART size.
253 * Thus function will never fails.
255 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
257 const uint64_t four_gb = 0x100000000ULL;
258 u64 size_af, size_bf;
259 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/
260 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1);
262 /* VCE doesn't like it when BOs cross a 4GB segment, so align
263 * the GART base on a 4GB boundary as well.
265 size_bf = mc->fb_start;
266 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb);
268 if (mc->gart_size > max(size_bf, size_af)) {
269 dev_warn(adev->dev, "limiting GART\n");
270 mc->gart_size = max(size_bf, size_af);
273 if ((size_bf >= mc->gart_size && size_bf < size_af) ||
274 (size_af < mc->gart_size))
277 mc->gart_start = max_mc_address - mc->gart_size + 1;
279 mc->gart_start &= ~(four_gb - 1);
280 mc->gart_end = mc->gart_start + mc->gart_size - 1;
281 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
282 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
286 * amdgpu_gmc_agp_location - try to find AGP location
287 * @adev: amdgpu device structure holding all necessary information
288 * @mc: memory controller structure holding memory information
290 * Function will place try to find a place for the AGP BAR in the MC address
293 * AGP BAR will be assigned the largest available hole in the address space.
294 * Should be called after VRAM and GART locations are setup.
296 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
298 const uint64_t sixteen_gb = 1ULL << 34;
299 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1);
300 u64 size_af, size_bf;
302 if (amdgpu_sriov_vf(adev)) {
303 mc->agp_start = 0xffffffffffff;
310 if (mc->fb_start > mc->gart_start) {
311 size_bf = (mc->fb_start & sixteen_gb_mask) -
312 ALIGN(mc->gart_end + 1, sixteen_gb);
313 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb);
315 size_bf = mc->fb_start & sixteen_gb_mask;
316 size_af = (mc->gart_start & sixteen_gb_mask) -
317 ALIGN(mc->fb_end + 1, sixteen_gb);
320 if (size_bf > size_af) {
321 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask;
322 mc->agp_size = size_bf;
324 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb);
325 mc->agp_size = size_af;
328 mc->agp_end = mc->agp_start + mc->agp_size - 1;
329 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n",
330 mc->agp_size >> 20, mc->agp_start, mc->agp_end);
334 * amdgpu_gmc_filter_faults - filter VM faults
336 * @adev: amdgpu device structure
337 * @addr: address of the VM fault
338 * @pasid: PASID of the process causing the fault
339 * @timestamp: timestamp of the fault
342 * True if the fault was filtered and should not be processed further.
343 * False if the fault is a new one and needs to be handled.
345 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, uint64_t addr,
346 uint16_t pasid, uint64_t timestamp)
348 struct amdgpu_gmc *gmc = &adev->gmc;
350 uint64_t stamp, key = addr << 4 | pasid;
351 struct amdgpu_gmc_fault *fault;
354 /* If we don't have space left in the ring buffer return immediately */
355 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) -
356 AMDGPU_GMC_FAULT_TIMEOUT;
357 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp)
360 /* Try to find the fault in the hash */
361 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
362 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
363 while (fault->timestamp >= stamp) {
366 if (fault->key == key)
369 tmp = fault->timestamp;
370 fault = &gmc->fault_ring[fault->next];
372 /* Check if the entry was reused */
373 if (fault->timestamp >= tmp)
377 /* Add the fault to the ring */
378 fault = &gmc->fault_ring[gmc->last_fault];
380 fault->timestamp = timestamp;
382 /* And update the hash */
383 fault->next = gmc->fault_hash[hash].idx;
384 gmc->fault_hash[hash].idx = gmc->last_fault++;
388 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
392 if (adev->umc.funcs && adev->umc.funcs->ras_late_init) {
393 r = adev->umc.funcs->ras_late_init(adev);
398 if (adev->mmhub.funcs && adev->mmhub.funcs->ras_late_init) {
399 r = adev->mmhub.funcs->ras_late_init(adev);
404 return amdgpu_xgmi_ras_late_init(adev);
407 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
409 amdgpu_umc_ras_fini(adev);
410 amdgpu_mmhub_ras_fini(adev);
411 amdgpu_xgmi_ras_fini(adev);
415 * The latest engine allocation on gfx9/10 is:
416 * Engine 2, 3: firmware
417 * Engine 0, 1, 4~16: amdgpu ring,
418 * subject to change when ring number changes
419 * Engine 17: Gart flushes
421 #define GFXHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
422 #define MMHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
424 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
426 struct amdgpu_ring *ring;
427 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] =
428 {GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP,
429 GFXHUB_FREE_VM_INV_ENGS_BITMAP};
431 unsigned vmhub, inv_eng;
433 for (i = 0; i < adev->num_rings; ++i) {
434 ring = adev->rings[i];
435 vmhub = ring->funcs->vmhub;
437 if (ring == &adev->mes.ring)
440 inv_eng = ffs(vm_inv_engs[vmhub]);
442 dev_err(adev->dev, "no VM inv eng for ring %s\n",
447 ring->vm_inv_eng = inv_eng - 1;
448 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
450 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
451 ring->name, ring->vm_inv_eng, ring->funcs->vmhub);
458 * amdgpu_tmz_set -- check and set if a device supports TMZ
459 * @adev: amdgpu_device pointer
461 * Check and set if an the device @adev supports Trusted Memory
464 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev)
466 switch (adev->asic_type) {
468 if (amdgpu_tmz == 0) {
469 adev->gmc.tmz_enabled = false;
471 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n");
473 adev->gmc.tmz_enabled = true;
475 "Trusted Memory Zone (TMZ) feature enabled\n");
483 /* Don't enable it by default yet.
485 if (amdgpu_tmz < 1) {
486 adev->gmc.tmz_enabled = false;
488 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n");
490 adev->gmc.tmz_enabled = true;
492 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n");
496 adev->gmc.tmz_enabled = false;
498 "Trusted Memory Zone (TMZ) feature not supported\n");
504 * amdgpu_noretry_set -- set per asic noretry defaults
505 * @adev: amdgpu_device pointer
507 * Set a per asic default for the no-retry parameter.
510 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev)
512 struct amdgpu_gmc *gmc = &adev->gmc;
514 switch (adev->asic_type) {
519 * noretry = 0 will cause kfd page fault tests fail
520 * for some ASICs, so set default to 1 for these ASICs.
522 if (amdgpu_noretry == -1)
525 gmc->noretry = amdgpu_noretry;
529 /* Raven currently has issues with noretry
530 * regardless of what we decide for other
531 * asics, we should leave raven with
532 * noretry = 0 until we root cause the
535 * default this to 0 for now, but we may want
536 * to change this in the future for certain
537 * GPUs as it can increase performance in
540 if (amdgpu_noretry == -1)
543 gmc->noretry = amdgpu_noretry;
548 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
551 struct amdgpu_vmhub *hub;
554 hub = &adev->vmhub[hub_type];
555 for (i = 0; i < 16; i++) {
556 reg = hub->vm_context0_cntl + hub->ctx_distance * i;
560 tmp |= hub->vm_cntx_cntl_vm_fault;
562 tmp &= ~hub->vm_cntx_cntl_vm_fault;
568 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev)
574 * Currently there is a bug where some memory client outside
575 * of the driver writes to first 8M of VRAM on S3 resume,
576 * this overrides GART which by default gets placed in first 8M and
577 * causes VM_FAULTS once GTT is accessed.
578 * Keep the stolen memory reservation until the while this is not solved.
580 switch (adev->asic_type) {
584 adev->mman.keep_stolen_vga_memory = true;
587 adev->mman.keep_stolen_vga_memory = false;
591 if (amdgpu_sriov_vf(adev) ||
592 !amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_DCE)) {
595 size = amdgpu_gmc_get_vbios_fb_size(adev);
597 if (adev->mman.keep_stolen_vga_memory)
598 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION);
601 /* set to 0 if the pre-OS buffer uses up most of vram */
602 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
605 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) {
606 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION;
607 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size;
609 adev->mman.stolen_vga_size = size;
610 adev->mman.stolen_extended_size = 0;
615 * amdgpu_gmc_init_pdb0 - initialize PDB0
617 * @adev: amdgpu_device pointer
619 * This function is only used when GART page table is used
620 * for FB address translatioin. In such a case, we construct
621 * a 2-level system VM page table: PDB0->PTB, to cover both
622 * VRAM of the hive and system memory.
624 * PDB0 is static, initialized once on driver initialization.
625 * The first n entries of PDB0 are used as PTE by setting
626 * P bit to 1, pointing to VRAM. The n+1'th entry points
627 * to a big PTB covering system memory.
630 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
633 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW?
634 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M
636 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
637 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21;
638 u64 vram_addr = adev->vm_manager.vram_base_offset -
639 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
640 u64 vram_end = vram_addr + vram_size;
641 u64 gart_ptb_gpu_pa = amdgpu_bo_gpu_offset(adev->gart.bo) +
642 adev->vm_manager.vram_base_offset - adev->gmc.vram_start;
644 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
645 flags |= AMDGPU_PTE_WRITEABLE;
646 flags |= AMDGPU_PTE_SNOOPED;
647 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1));
648 flags |= AMDGPU_PDE_PTE;
650 /* The first n PDE0 entries are used as PTE,
653 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size)
654 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags);
656 /* The n+1'th PDE0 entry points to a huge
657 * PTB who has more than 512 entries each
658 * pointing to a 4K system page
660 flags = AMDGPU_PTE_VALID;
661 flags |= AMDGPU_PDE_BFS(0) | AMDGPU_PTE_SNOOPED;
662 /* Requires gart_ptb_gpu_pa to be 4K aligned */
663 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags);