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_get_pde_for_bo - get the PDE for a BO
37 * @bo: the BO to get the PDE for
38 * @level: the level in the PD hirarchy
39 * @addr: resulting addr
40 * @flags: resulting flags
42 * Get the address and flags to be used for a PDE (Page Directory Entry).
44 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
45 uint64_t *addr, uint64_t *flags)
47 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
49 switch (bo->tbo.mem.mem_type) {
51 *addr = bo->tbo.ttm->dma_address[0];
54 *addr = amdgpu_bo_gpu_offset(bo);
60 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, &bo->tbo.mem);
61 amdgpu_gmc_get_vm_pde(adev, level, addr, flags);
65 * amdgpu_gmc_pd_addr - return the address of the root directory
67 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo)
69 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
72 /* TODO: move that into ASIC specific code */
73 if (adev->asic_type >= CHIP_VEGA10) {
74 uint64_t flags = AMDGPU_PTE_VALID;
76 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags);
79 pd_addr = amdgpu_bo_gpu_offset(bo);
85 * amdgpu_gmc_set_pte_pde - update the page tables using CPU
87 * @adev: amdgpu_device pointer
88 * @cpu_pt_addr: cpu address of the page table
89 * @gpu_page_idx: entry in the page table to update
90 * @addr: dst addr to write into pte/pde
91 * @flags: access flags
93 * Update the page tables using CPU.
95 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
96 uint32_t gpu_page_idx, uint64_t addr,
99 void __iomem *ptr = (void *)cpu_pt_addr;
103 * The following is for PTE only. GART does not have PDEs.
105 value = addr & 0x0000FFFFFFFFF000ULL;
107 writeq(value, ptr + (gpu_page_idx * 8));
112 * amdgpu_gmc_agp_addr - return the address in the AGP address space
114 * @bo: TTM BO which needs the address, must be in GTT domain
116 * Tries to figure out how to access the BO through the AGP aperture. Returns
117 * AMDGPU_BO_INVALID_OFFSET if that is not possible.
119 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo)
121 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
123 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached)
124 return AMDGPU_BO_INVALID_OFFSET;
126 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size)
127 return AMDGPU_BO_INVALID_OFFSET;
129 return adev->gmc.agp_start + bo->ttm->dma_address[0];
133 * amdgpu_gmc_vram_location - try to find VRAM location
135 * @adev: amdgpu device structure holding all necessary information
136 * @mc: memory controller structure holding memory information
137 * @base: base address at which to put VRAM
139 * Function will try to place VRAM at base address provided
142 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
145 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
147 mc->vram_start = base;
148 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
149 if (limit && limit < mc->real_vram_size)
150 mc->real_vram_size = limit;
152 if (mc->xgmi.num_physical_nodes == 0) {
153 mc->fb_start = mc->vram_start;
154 mc->fb_end = mc->vram_end;
156 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
157 mc->mc_vram_size >> 20, mc->vram_start,
158 mc->vram_end, mc->real_vram_size >> 20);
162 * amdgpu_gmc_gart_location - try to find GART location
164 * @adev: amdgpu device structure holding all necessary information
165 * @mc: memory controller structure holding memory information
167 * Function will place try to place GART before or after VRAM.
169 * If GART size is bigger than space left then we ajust GART size.
170 * Thus function will never fails.
172 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
174 const uint64_t four_gb = 0x100000000ULL;
175 u64 size_af, size_bf;
176 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/
177 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1);
179 mc->gart_size += adev->pm.smu_prv_buffer_size;
181 /* VCE doesn't like it when BOs cross a 4GB segment, so align
182 * the GART base on a 4GB boundary as well.
184 size_bf = mc->fb_start;
185 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb);
187 if (mc->gart_size > max(size_bf, size_af)) {
188 dev_warn(adev->dev, "limiting GART\n");
189 mc->gart_size = max(size_bf, size_af);
192 if ((size_bf >= mc->gart_size && size_bf < size_af) ||
193 (size_af < mc->gart_size))
196 mc->gart_start = max_mc_address - mc->gart_size + 1;
198 mc->gart_start &= ~(four_gb - 1);
199 mc->gart_end = mc->gart_start + mc->gart_size - 1;
200 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
201 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
205 * amdgpu_gmc_agp_location - try to find AGP location
206 * @adev: amdgpu device structure holding all necessary information
207 * @mc: memory controller structure holding memory information
209 * Function will place try to find a place for the AGP BAR in the MC address
212 * AGP BAR will be assigned the largest available hole in the address space.
213 * Should be called after VRAM and GART locations are setup.
215 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
217 const uint64_t sixteen_gb = 1ULL << 34;
218 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1);
219 u64 size_af, size_bf;
221 if (amdgpu_sriov_vf(adev)) {
222 mc->agp_start = 0xffffffffffff;
229 if (mc->fb_start > mc->gart_start) {
230 size_bf = (mc->fb_start & sixteen_gb_mask) -
231 ALIGN(mc->gart_end + 1, sixteen_gb);
232 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb);
234 size_bf = mc->fb_start & sixteen_gb_mask;
235 size_af = (mc->gart_start & sixteen_gb_mask) -
236 ALIGN(mc->fb_end + 1, sixteen_gb);
239 if (size_bf > size_af) {
240 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask;
241 mc->agp_size = size_bf;
243 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb);
244 mc->agp_size = size_af;
247 mc->agp_end = mc->agp_start + mc->agp_size - 1;
248 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n",
249 mc->agp_size >> 20, mc->agp_start, mc->agp_end);
253 * amdgpu_gmc_filter_faults - filter VM faults
255 * @adev: amdgpu device structure
256 * @addr: address of the VM fault
257 * @pasid: PASID of the process causing the fault
258 * @timestamp: timestamp of the fault
261 * True if the fault was filtered and should not be processed further.
262 * False if the fault is a new one and needs to be handled.
264 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, uint64_t addr,
265 uint16_t pasid, uint64_t timestamp)
267 struct amdgpu_gmc *gmc = &adev->gmc;
269 uint64_t stamp, key = addr << 4 | pasid;
270 struct amdgpu_gmc_fault *fault;
273 /* If we don't have space left in the ring buffer return immediately */
274 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) -
275 AMDGPU_GMC_FAULT_TIMEOUT;
276 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp)
279 /* Try to find the fault in the hash */
280 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
281 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
282 while (fault->timestamp >= stamp) {
285 if (fault->key == key)
288 tmp = fault->timestamp;
289 fault = &gmc->fault_ring[fault->next];
291 /* Check if the entry was reused */
292 if (fault->timestamp >= tmp)
296 /* Add the fault to the ring */
297 fault = &gmc->fault_ring[gmc->last_fault];
299 fault->timestamp = timestamp;
301 /* And update the hash */
302 fault->next = gmc->fault_hash[hash].idx;
303 gmc->fault_hash[hash].idx = gmc->last_fault++;
307 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
311 if (adev->umc.funcs && adev->umc.funcs->ras_late_init) {
312 r = adev->umc.funcs->ras_late_init(adev);
317 if (adev->mmhub.funcs && adev->mmhub.funcs->ras_late_init) {
318 r = adev->mmhub.funcs->ras_late_init(adev);
323 return amdgpu_xgmi_ras_late_init(adev);
326 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
328 amdgpu_umc_ras_fini(adev);
329 amdgpu_mmhub_ras_fini(adev);
330 amdgpu_xgmi_ras_fini(adev);
334 * The latest engine allocation on gfx9/10 is:
335 * Engine 2, 3: firmware
336 * Engine 0, 1, 4~16: amdgpu ring,
337 * subject to change when ring number changes
338 * Engine 17: Gart flushes
340 #define GFXHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
341 #define MMHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
343 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
345 struct amdgpu_ring *ring;
346 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] =
347 {GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP,
348 GFXHUB_FREE_VM_INV_ENGS_BITMAP};
350 unsigned vmhub, inv_eng;
352 for (i = 0; i < adev->num_rings; ++i) {
353 ring = adev->rings[i];
354 vmhub = ring->funcs->vmhub;
356 if (ring == &adev->mes.ring)
359 inv_eng = ffs(vm_inv_engs[vmhub]);
361 dev_err(adev->dev, "no VM inv eng for ring %s\n",
366 ring->vm_inv_eng = inv_eng - 1;
367 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
369 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
370 ring->name, ring->vm_inv_eng, ring->funcs->vmhub);
377 * amdgpu_tmz_set -- check and set if a device supports TMZ
378 * @adev: amdgpu_device pointer
380 * Check and set if an the device @adev supports Trusted Memory
383 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev)
385 switch (adev->asic_type) {
392 /* Don't enable it by default yet.
394 if (amdgpu_tmz < 1) {
395 adev->gmc.tmz_enabled = false;
397 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n");
399 adev->gmc.tmz_enabled = true;
401 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n");
405 adev->gmc.tmz_enabled = false;
407 "Trusted Memory Zone (TMZ) feature not supported\n");
413 * amdgpu_noretry_set -- set per asic noretry defaults
414 * @adev: amdgpu_device pointer
416 * Set a per asic default for the no-retry parameter.
419 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev)
421 struct amdgpu_gmc *gmc = &adev->gmc;
423 switch (adev->asic_type) {
427 * noretry = 0 will cause kfd page fault tests fail
428 * for some ASICs, so set default to 1 for these ASICs.
430 if (amdgpu_noretry == -1)
433 gmc->noretry = amdgpu_noretry;
437 /* Raven currently has issues with noretry
438 * regardless of what we decide for other
439 * asics, we should leave raven with
440 * noretry = 0 until we root cause the
443 * default this to 0 for now, but we may want
444 * to change this in the future for certain
445 * GPUs as it can increase performance in
448 if (amdgpu_noretry == -1)
451 gmc->noretry = amdgpu_noretry;
456 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
459 struct amdgpu_vmhub *hub;
462 hub = &adev->vmhub[hub_type];
463 for (i = 0; i < 16; i++) {
464 reg = hub->vm_context0_cntl + hub->ctx_distance * i;
468 tmp |= hub->vm_cntx_cntl_vm_fault;
470 tmp &= ~hub->vm_cntx_cntl_vm_fault;
476 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev)
482 * Currently there is a bug where some memory client outside
483 * of the driver writes to first 8M of VRAM on S3 resume,
484 * this overrides GART which by default gets placed in first 8M and
485 * causes VM_FAULTS once GTT is accessed.
486 * Keep the stolen memory reservation until the while this is not solved.
488 switch (adev->asic_type) {
492 adev->mman.keep_stolen_vga_memory = true;
495 adev->mman.keep_stolen_vga_memory = false;
499 if (amdgpu_sriov_vf(adev) ||
500 !amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_DCE)) {
503 size = amdgpu_gmc_get_vbios_fb_size(adev);
505 if (adev->mman.keep_stolen_vga_memory)
506 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION);
509 /* set to 0 if the pre-OS buffer uses up most of vram */
510 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
513 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) {
514 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION;
515 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size;
517 adev->mman.stolen_vga_size = size;
518 adev->mman.stolen_extended_size = 0;