1 /* 2 * Copyright 2018 Advanced Micro Devices, Inc. 3 * All Rights Reserved. 4 * 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: 12 * 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. 20 * 21 * The above copyright notice and this permission notice (including the 22 * next paragraph) shall be included in all copies or substantial portions 23 * of the Software. 24 * 25 */ 26 #ifndef __AMDGPU_GMC_H__ 27 #define __AMDGPU_GMC_H__ 28 29 #include <linux/types.h> 30 31 #include "amdgpu_irq.h" 32 33 /* VA hole for 48bit addresses on Vega10 */ 34 #define AMDGPU_GMC_HOLE_START 0x0000800000000000ULL 35 #define AMDGPU_GMC_HOLE_END 0xffff800000000000ULL 36 37 /* 38 * Hardware is programmed as if the hole doesn't exists with start and end 39 * address values. 40 * 41 * This mask is used to remove the upper 16bits of the VA and so come up with 42 * the linear addr value. 43 */ 44 #define AMDGPU_GMC_HOLE_MASK 0x0000ffffffffffffULL 45 46 /* 47 * Ring size as power of two for the log of recent faults. 48 */ 49 #define AMDGPU_GMC_FAULT_RING_ORDER 8 50 #define AMDGPU_GMC_FAULT_RING_SIZE (1 << AMDGPU_GMC_FAULT_RING_ORDER) 51 52 /* 53 * Hash size as power of two for the log of recent faults 54 */ 55 #define AMDGPU_GMC_FAULT_HASH_ORDER 8 56 #define AMDGPU_GMC_FAULT_HASH_SIZE (1 << AMDGPU_GMC_FAULT_HASH_ORDER) 57 58 /* 59 * Number of IH timestamp ticks until a fault is considered handled 60 */ 61 #define AMDGPU_GMC_FAULT_TIMEOUT 5000ULL 62 63 /* 64 * Default stolen memory size, 1024 * 768 * 4 65 */ 66 #define AMDGPU_STOLEN_BIST_TRAINING_DEFAULT_SIZE 0x300000ULL 67 68 struct firmware; 69 70 /* 71 * GMC page fault information 72 */ 73 struct amdgpu_gmc_fault { 74 uint64_t timestamp; 75 uint64_t next:AMDGPU_GMC_FAULT_RING_ORDER; 76 uint64_t key:52; 77 }; 78 79 /* 80 * VMHUB structures, functions & helpers 81 */ 82 struct amdgpu_vmhub { 83 uint32_t ctx0_ptb_addr_lo32; 84 uint32_t ctx0_ptb_addr_hi32; 85 uint32_t vm_inv_eng0_sem; 86 uint32_t vm_inv_eng0_req; 87 uint32_t vm_inv_eng0_ack; 88 uint32_t vm_context0_cntl; 89 uint32_t vm_l2_pro_fault_status; 90 uint32_t vm_l2_pro_fault_cntl; 91 }; 92 93 /* 94 * GPU MC structures, functions & helpers 95 */ 96 struct amdgpu_gmc_funcs { 97 /* flush the vm tlb via mmio */ 98 void (*flush_gpu_tlb)(struct amdgpu_device *adev, uint32_t vmid, 99 uint32_t vmhub, uint32_t flush_type); 100 /* flush the vm tlb via pasid */ 101 int (*flush_gpu_tlb_pasid)(struct amdgpu_device *adev, uint16_t pasid, 102 uint32_t flush_type, bool all_hub); 103 /* flush the vm tlb via ring */ 104 uint64_t (*emit_flush_gpu_tlb)(struct amdgpu_ring *ring, unsigned vmid, 105 uint64_t pd_addr); 106 /* Change the VMID -> PASID mapping */ 107 void (*emit_pasid_mapping)(struct amdgpu_ring *ring, unsigned vmid, 108 unsigned pasid); 109 /* enable/disable PRT support */ 110 void (*set_prt)(struct amdgpu_device *adev, bool enable); 111 /* map mtype to hardware flags */ 112 uint64_t (*map_mtype)(struct amdgpu_device *adev, uint32_t flags); 113 /* get the pde for a given mc addr */ 114 void (*get_vm_pde)(struct amdgpu_device *adev, int level, 115 u64 *dst, u64 *flags); 116 /* get the pte flags to use for a BO VA mapping */ 117 void (*get_vm_pte)(struct amdgpu_device *adev, 118 struct amdgpu_bo_va_mapping *mapping, 119 uint64_t *flags); 120 }; 121 122 struct amdgpu_xgmi { 123 /* from psp */ 124 u64 node_id; 125 u64 hive_id; 126 /* fixed per family */ 127 u64 node_segment_size; 128 /* physical node (0-3) */ 129 unsigned physical_node_id; 130 /* number of nodes (0-4) */ 131 unsigned num_physical_nodes; 132 /* gpu list in the same hive */ 133 struct list_head head; 134 bool supported; 135 struct ras_common_if *ras_if; 136 }; 137 138 struct amdgpu_gmc { 139 /* FB's physical address in MMIO space (for CPU to 140 * map FB). This is different compared to the agp/ 141 * gart/vram_start/end field as the later is from 142 * GPU's view and aper_base is from CPU's view. 143 */ 144 resource_size_t aper_size; 145 resource_size_t aper_base; 146 /* for some chips with <= 32MB we need to lie 147 * about vram size near mc fb location */ 148 u64 mc_vram_size; 149 u64 visible_vram_size; 150 /* AGP aperture start and end in MC address space 151 * Driver find a hole in the MC address space 152 * to place AGP by setting MC_VM_AGP_BOT/TOP registers 153 * Under VMID0, logical address == MC address. AGP 154 * aperture maps to physical bus or IOVA addressed. 155 * AGP aperture is used to simulate FB in ZFB case. 156 * AGP aperture is also used for page table in system 157 * memory (mainly for APU). 158 * 159 */ 160 u64 agp_size; 161 u64 agp_start; 162 u64 agp_end; 163 /* GART aperture start and end in MC address space 164 * Driver find a hole in the MC address space 165 * to place GART by setting VM_CONTEXT0_PAGE_TABLE_START/END_ADDR 166 * registers 167 * Under VMID0, logical address inside GART aperture will 168 * be translated through gpuvm gart page table to access 169 * paged system memory 170 */ 171 u64 gart_size; 172 u64 gart_start; 173 u64 gart_end; 174 /* Frame buffer aperture of this GPU device. Different from 175 * fb_start (see below), this only covers the local GPU device. 176 * Driver get fb_start from MC_VM_FB_LOCATION_BASE (set by vbios) 177 * and calculate vram_start of this local device by adding an 178 * offset inside the XGMI hive. 179 * Under VMID0, logical address == MC address 180 */ 181 u64 vram_start; 182 u64 vram_end; 183 /* FB region , it's same as local vram region in single GPU, in XGMI 184 * configuration, this region covers all GPUs in the same hive , 185 * each GPU in the hive has the same view of this FB region . 186 * GPU0's vram starts at offset (0 * segment size) , 187 * GPU1 starts at offset (1 * segment size), etc. 188 */ 189 u64 fb_start; 190 u64 fb_end; 191 unsigned vram_width; 192 u64 real_vram_size; 193 int vram_mtrr; 194 u64 mc_mask; 195 const struct firmware *fw; /* MC firmware */ 196 uint32_t fw_version; 197 struct amdgpu_irq_src vm_fault; 198 uint32_t vram_type; 199 uint8_t vram_vendor; 200 uint32_t srbm_soft_reset; 201 bool prt_warning; 202 uint64_t stolen_size; 203 /* apertures */ 204 u64 shared_aperture_start; 205 u64 shared_aperture_end; 206 u64 private_aperture_start; 207 u64 private_aperture_end; 208 /* protects concurrent invalidation */ 209 spinlock_t invalidate_lock; 210 bool translate_further; 211 struct kfd_vm_fault_info *vm_fault_info; 212 atomic_t vm_fault_info_updated; 213 214 struct amdgpu_gmc_fault fault_ring[AMDGPU_GMC_FAULT_RING_SIZE]; 215 struct { 216 uint64_t idx:AMDGPU_GMC_FAULT_RING_ORDER; 217 } fault_hash[AMDGPU_GMC_FAULT_HASH_SIZE]; 218 uint64_t last_fault:AMDGPU_GMC_FAULT_RING_ORDER; 219 220 const struct amdgpu_gmc_funcs *gmc_funcs; 221 222 struct amdgpu_xgmi xgmi; 223 struct amdgpu_irq_src ecc_irq; 224 }; 225 226 #define amdgpu_gmc_flush_gpu_tlb(adev, vmid, vmhub, type) ((adev)->gmc.gmc_funcs->flush_gpu_tlb((adev), (vmid), (vmhub), (type))) 227 #define amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, type, allhub) \ 228 ((adev)->gmc.gmc_funcs->flush_gpu_tlb_pasid \ 229 ((adev), (pasid), (type), (allhub))) 230 #define amdgpu_gmc_emit_flush_gpu_tlb(r, vmid, addr) (r)->adev->gmc.gmc_funcs->emit_flush_gpu_tlb((r), (vmid), (addr)) 231 #define amdgpu_gmc_emit_pasid_mapping(r, vmid, pasid) (r)->adev->gmc.gmc_funcs->emit_pasid_mapping((r), (vmid), (pasid)) 232 #define amdgpu_gmc_map_mtype(adev, flags) (adev)->gmc.gmc_funcs->map_mtype((adev),(flags)) 233 #define amdgpu_gmc_get_vm_pde(adev, level, dst, flags) (adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (dst), (flags)) 234 #define amdgpu_gmc_get_vm_pte(adev, mapping, flags) (adev)->gmc.gmc_funcs->get_vm_pte((adev), (mapping), (flags)) 235 236 /** 237 * amdgpu_gmc_vram_full_visible - Check if full VRAM is visible through the BAR 238 * 239 * @adev: amdgpu_device pointer 240 * 241 * Returns: 242 * True if full VRAM is visible through the BAR 243 */ 244 static inline bool amdgpu_gmc_vram_full_visible(struct amdgpu_gmc *gmc) 245 { 246 WARN_ON(gmc->real_vram_size < gmc->visible_vram_size); 247 248 return (gmc->real_vram_size == gmc->visible_vram_size); 249 } 250 251 /** 252 * amdgpu_gmc_sign_extend - sign extend the given gmc address 253 * 254 * @addr: address to extend 255 */ 256 static inline uint64_t amdgpu_gmc_sign_extend(uint64_t addr) 257 { 258 if (addr >= AMDGPU_GMC_HOLE_START) 259 addr |= AMDGPU_GMC_HOLE_END; 260 261 return addr; 262 } 263 264 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level, 265 uint64_t *addr, uint64_t *flags); 266 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr, 267 uint32_t gpu_page_idx, uint64_t addr, 268 uint64_t flags); 269 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo); 270 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo); 271 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc, 272 u64 base); 273 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, 274 struct amdgpu_gmc *mc); 275 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, 276 struct amdgpu_gmc *mc); 277 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, uint64_t addr, 278 uint16_t pasid, uint64_t timestamp); 279 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev); 280 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev); 281 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev); 282 283 #endif 284