1 /* 2 * Copyright 2018 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 * 23 */ 24 #include <linux/list.h> 25 #include "amdgpu.h" 26 #include "amdgpu_xgmi.h" 27 #include "amdgpu_ras.h" 28 #include "soc15.h" 29 #include "df/df_3_6_offset.h" 30 #include "xgmi/xgmi_4_0_0_smn.h" 31 #include "xgmi/xgmi_4_0_0_sh_mask.h" 32 #include "xgmi/xgmi_6_1_0_sh_mask.h" 33 #include "wafl/wafl2_4_0_0_smn.h" 34 #include "wafl/wafl2_4_0_0_sh_mask.h" 35 36 #include "amdgpu_reset.h" 37 38 #define smnPCS_XGMI3X16_PCS_ERROR_STATUS 0x11a0020c 39 #define smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK 0x11a00218 40 #define smnPCS_GOPX1_PCS_ERROR_STATUS 0x12200210 41 #define smnPCS_GOPX1_PCS_ERROR_NONCORRECTABLE_MASK 0x12200218 42 43 #define XGMI_STATE_DISABLE 0xD1 44 #define XGMI_STATE_LS0 0x81 45 #define XGMI_LINK_ACTIVE 1 46 #define XGMI_LINK_INACTIVE 0 47 48 static DEFINE_MUTEX(xgmi_mutex); 49 50 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE 4 51 52 static LIST_HEAD(xgmi_hive_list); 53 54 static const int xgmi_pcs_err_status_reg_vg20[] = { 55 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 56 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 57 }; 58 59 static const int wafl_pcs_err_status_reg_vg20[] = { 60 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 61 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 62 }; 63 64 static const int xgmi_pcs_err_status_reg_arct[] = { 65 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 66 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 67 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000, 68 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000, 69 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000, 70 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000, 71 }; 72 73 /* same as vg20*/ 74 static const int wafl_pcs_err_status_reg_arct[] = { 75 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 76 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 77 }; 78 79 static const int xgmi3x16_pcs_err_status_reg_aldebaran[] = { 80 smnPCS_XGMI3X16_PCS_ERROR_STATUS, 81 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000, 82 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x200000, 83 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x300000, 84 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x400000, 85 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x500000, 86 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x600000, 87 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x700000 88 }; 89 90 static const int xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[] = { 91 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK, 92 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000, 93 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x200000, 94 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x300000, 95 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x400000, 96 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x500000, 97 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x600000, 98 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x700000 99 }; 100 101 static const int walf_pcs_err_status_reg_aldebaran[] = { 102 smnPCS_GOPX1_PCS_ERROR_STATUS, 103 smnPCS_GOPX1_PCS_ERROR_STATUS + 0x100000 104 }; 105 106 static const int walf_pcs_err_noncorrectable_mask_reg_aldebaran[] = { 107 smnPCS_GOPX1_PCS_ERROR_NONCORRECTABLE_MASK, 108 smnPCS_GOPX1_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000 109 }; 110 111 static const int xgmi3x16_pcs_err_status_reg_v6_4[] = { 112 smnPCS_XGMI3X16_PCS_ERROR_STATUS, 113 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000 114 }; 115 116 static const int xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[] = { 117 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK, 118 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000 119 }; 120 121 static const u64 xgmi_v6_4_0_mca_base_array[] = { 122 0x11a09200, 123 0x11b09200, 124 }; 125 126 static const char *xgmi_v6_4_0_ras_error_code_ext[32] = { 127 [0x00] = "XGMI PCS DataLossErr", 128 [0x01] = "XGMI PCS TrainingErr", 129 [0x02] = "XGMI PCS FlowCtrlAckErr", 130 [0x03] = "XGMI PCS RxFifoUnderflowErr", 131 [0x04] = "XGMI PCS RxFifoOverflowErr", 132 [0x05] = "XGMI PCS CRCErr", 133 [0x06] = "XGMI PCS BERExceededErr", 134 [0x07] = "XGMI PCS TxMetaDataErr", 135 [0x08] = "XGMI PCS ReplayBufParityErr", 136 [0x09] = "XGMI PCS DataParityErr", 137 [0x0a] = "XGMI PCS ReplayFifoOverflowErr", 138 [0x0b] = "XGMI PCS ReplayFifoUnderflowErr", 139 [0x0c] = "XGMI PCS ElasticFifoOverflowErr", 140 [0x0d] = "XGMI PCS DeskewErr", 141 [0x0e] = "XGMI PCS FlowCtrlCRCErr", 142 [0x0f] = "XGMI PCS DataStartupLimitErr", 143 [0x10] = "XGMI PCS FCInitTimeoutErr", 144 [0x11] = "XGMI PCS RecoveryTimeoutErr", 145 [0x12] = "XGMI PCS ReadySerialTimeoutErr", 146 [0x13] = "XGMI PCS ReadySerialAttemptErr", 147 [0x14] = "XGMI PCS RecoveryAttemptErr", 148 [0x15] = "XGMI PCS RecoveryRelockAttemptErr", 149 [0x16] = "XGMI PCS ReplayAttemptErr", 150 [0x17] = "XGMI PCS SyncHdrErr", 151 [0x18] = "XGMI PCS TxReplayTimeoutErr", 152 [0x19] = "XGMI PCS RxReplayTimeoutErr", 153 [0x1a] = "XGMI PCS LinkSubTxTimeoutErr", 154 [0x1b] = "XGMI PCS LinkSubRxTimeoutErr", 155 [0x1c] = "XGMI PCS RxCMDPktErr", 156 }; 157 158 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = { 159 {"XGMI PCS DataLossErr", 160 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)}, 161 {"XGMI PCS TrainingErr", 162 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)}, 163 {"XGMI PCS CRCErr", 164 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)}, 165 {"XGMI PCS BERExceededErr", 166 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)}, 167 {"XGMI PCS TxMetaDataErr", 168 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)}, 169 {"XGMI PCS ReplayBufParityErr", 170 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 171 {"XGMI PCS DataParityErr", 172 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)}, 173 {"XGMI PCS ReplayFifoOverflowErr", 174 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 175 {"XGMI PCS ReplayFifoUnderflowErr", 176 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 177 {"XGMI PCS ElasticFifoOverflowErr", 178 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 179 {"XGMI PCS DeskewErr", 180 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)}, 181 {"XGMI PCS DataStartupLimitErr", 182 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 183 {"XGMI PCS FCInitTimeoutErr", 184 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 185 {"XGMI PCS RecoveryTimeoutErr", 186 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 187 {"XGMI PCS ReadySerialTimeoutErr", 188 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 189 {"XGMI PCS ReadySerialAttemptErr", 190 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 191 {"XGMI PCS RecoveryAttemptErr", 192 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 193 {"XGMI PCS RecoveryRelockAttemptErr", 194 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 195 }; 196 197 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = { 198 {"WAFL PCS DataLossErr", 199 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)}, 200 {"WAFL PCS TrainingErr", 201 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)}, 202 {"WAFL PCS CRCErr", 203 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)}, 204 {"WAFL PCS BERExceededErr", 205 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)}, 206 {"WAFL PCS TxMetaDataErr", 207 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)}, 208 {"WAFL PCS ReplayBufParityErr", 209 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)}, 210 {"WAFL PCS DataParityErr", 211 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)}, 212 {"WAFL PCS ReplayFifoOverflowErr", 213 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 214 {"WAFL PCS ReplayFifoUnderflowErr", 215 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 216 {"WAFL PCS ElasticFifoOverflowErr", 217 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 218 {"WAFL PCS DeskewErr", 219 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)}, 220 {"WAFL PCS DataStartupLimitErr", 221 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)}, 222 {"WAFL PCS FCInitTimeoutErr", 223 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 224 {"WAFL PCS RecoveryTimeoutErr", 225 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 226 {"WAFL PCS ReadySerialTimeoutErr", 227 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 228 {"WAFL PCS ReadySerialAttemptErr", 229 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 230 {"WAFL PCS RecoveryAttemptErr", 231 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 232 {"WAFL PCS RecoveryRelockAttemptErr", 233 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 234 }; 235 236 static const struct amdgpu_pcs_ras_field xgmi3x16_pcs_ras_fields[] = { 237 {"XGMI3X16 PCS DataLossErr", 238 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataLossErr)}, 239 {"XGMI3X16 PCS TrainingErr", 240 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TrainingErr)}, 241 {"XGMI3X16 PCS FlowCtrlAckErr", 242 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FlowCtrlAckErr)}, 243 {"XGMI3X16 PCS RxFifoUnderflowErr", 244 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxFifoUnderflowErr)}, 245 {"XGMI3X16 PCS RxFifoOverflowErr", 246 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxFifoOverflowErr)}, 247 {"XGMI3X16 PCS CRCErr", 248 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, CRCErr)}, 249 {"XGMI3X16 PCS BERExceededErr", 250 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, BERExceededErr)}, 251 {"XGMI3X16 PCS TxVcidDataErr", 252 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TxVcidDataErr)}, 253 {"XGMI3X16 PCS ReplayBufParityErr", 254 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 255 {"XGMI3X16 PCS DataParityErr", 256 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataParityErr)}, 257 {"XGMI3X16 PCS ReplayFifoOverflowErr", 258 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 259 {"XGMI3X16 PCS ReplayFifoUnderflowErr", 260 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 261 {"XGMI3X16 PCS ElasticFifoOverflowErr", 262 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 263 {"XGMI3X16 PCS DeskewErr", 264 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DeskewErr)}, 265 {"XGMI3X16 PCS FlowCtrlCRCErr", 266 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FlowCtrlCRCErr)}, 267 {"XGMI3X16 PCS DataStartupLimitErr", 268 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 269 {"XGMI3X16 PCS FCInitTimeoutErr", 270 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 271 {"XGMI3X16 PCS RecoveryTimeoutErr", 272 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 273 {"XGMI3X16 PCS ReadySerialTimeoutErr", 274 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 275 {"XGMI3X16 PCS ReadySerialAttemptErr", 276 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 277 {"XGMI3X16 PCS RecoveryAttemptErr", 278 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 279 {"XGMI3X16 PCS RecoveryRelockAttemptErr", 280 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 281 {"XGMI3X16 PCS ReplayAttemptErr", 282 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayAttemptErr)}, 283 {"XGMI3X16 PCS SyncHdrErr", 284 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, SyncHdrErr)}, 285 {"XGMI3X16 PCS TxReplayTimeoutErr", 286 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TxReplayTimeoutErr)}, 287 {"XGMI3X16 PCS RxReplayTimeoutErr", 288 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxReplayTimeoutErr)}, 289 {"XGMI3X16 PCS LinkSubTxTimeoutErr", 290 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, LinkSubTxTimeoutErr)}, 291 {"XGMI3X16 PCS LinkSubRxTimeoutErr", 292 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, LinkSubRxTimeoutErr)}, 293 {"XGMI3X16 PCS RxCMDPktErr", 294 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxCMDPktErr)}, 295 }; 296 297 int amdgpu_xgmi_get_ext_link(struct amdgpu_device *adev, int link_num) 298 { 299 int link_map_6_4_x[8] = { 0, 3, 1, 2, 7, 6, 4, 5 }; 300 301 if (adev->gmc.xgmi.num_physical_nodes <= 1) 302 return -EINVAL; 303 304 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 305 case IP_VERSION(6, 4, 0): 306 case IP_VERSION(6, 4, 1): 307 if (link_num < ARRAY_SIZE(link_map_6_4_x)) 308 return link_map_6_4_x[link_num]; 309 break; 310 default: 311 return -EINVAL; 312 } 313 314 return -EINVAL; 315 } 316 317 static u32 xgmi_v6_4_get_link_status(struct amdgpu_device *adev, int global_link_num) 318 { 319 const u32 smn_xgmi_6_4_pcs_state_hist1[2] = { 0x11a00070, 0x11b00070 }; 320 const u32 smn_xgmi_6_4_1_pcs_state_hist1[2] = { 0x12100070, 321 0x11b00070 }; 322 u32 i, n; 323 u64 addr; 324 325 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 326 case IP_VERSION(6, 4, 0): 327 n = ARRAY_SIZE(smn_xgmi_6_4_pcs_state_hist1); 328 addr = smn_xgmi_6_4_pcs_state_hist1[global_link_num % n]; 329 break; 330 case IP_VERSION(6, 4, 1): 331 n = ARRAY_SIZE(smn_xgmi_6_4_1_pcs_state_hist1); 332 addr = smn_xgmi_6_4_1_pcs_state_hist1[global_link_num % n]; 333 break; 334 default: 335 return U32_MAX; 336 } 337 338 i = global_link_num / n; 339 340 if (!(adev->aid_mask & BIT(i))) 341 return U32_MAX; 342 343 addr += adev->asic_funcs->encode_ext_smn_addressing(i); 344 345 return RREG32_PCIE_EXT(addr); 346 } 347 348 int amdgpu_get_xgmi_link_status(struct amdgpu_device *adev, int global_link_num) 349 { 350 u32 xgmi_state_reg_val; 351 352 if (adev->gmc.xgmi.num_physical_nodes <= 1) 353 return -EINVAL; 354 355 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 356 case IP_VERSION(6, 4, 0): 357 case IP_VERSION(6, 4, 1): 358 xgmi_state_reg_val = xgmi_v6_4_get_link_status(adev, global_link_num); 359 break; 360 default: 361 return -EOPNOTSUPP; 362 } 363 364 if ((xgmi_state_reg_val & 0xFF) == XGMI_STATE_DISABLE) 365 return -ENOLINK; 366 367 if ((xgmi_state_reg_val & 0xFF) == XGMI_STATE_LS0) 368 return XGMI_LINK_ACTIVE; 369 370 return XGMI_LINK_INACTIVE; 371 } 372 373 /** 374 * DOC: AMDGPU XGMI Support 375 * 376 * XGMI is a high speed interconnect that joins multiple GPU cards 377 * into a homogeneous memory space that is organized by a collective 378 * hive ID and individual node IDs, both of which are 64-bit numbers. 379 * 380 * The file xgmi_device_id contains the unique per GPU device ID and 381 * is stored in the /sys/class/drm/card${cardno}/device/ directory. 382 * 383 * Inside the device directory a sub-directory 'xgmi_hive_info' is 384 * created which contains the hive ID and the list of nodes. 385 * 386 * The hive ID is stored in: 387 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id 388 * 389 * The node information is stored in numbered directories: 390 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id 391 * 392 * Each device has their own xgmi_hive_info direction with a mirror 393 * set of node sub-directories. 394 * 395 * The XGMI memory space is built by contiguously adding the power of 396 * two padded VRAM space from each node to each other. 397 * 398 */ 399 400 static struct attribute amdgpu_xgmi_hive_id = { 401 .name = "xgmi_hive_id", 402 .mode = S_IRUGO 403 }; 404 405 static struct attribute *amdgpu_xgmi_hive_attrs[] = { 406 &amdgpu_xgmi_hive_id, 407 NULL 408 }; 409 ATTRIBUTE_GROUPS(amdgpu_xgmi_hive); 410 411 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj, 412 struct attribute *attr, char *buf) 413 { 414 struct amdgpu_hive_info *hive = container_of( 415 kobj, struct amdgpu_hive_info, kobj); 416 417 if (attr == &amdgpu_xgmi_hive_id) 418 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id); 419 420 return 0; 421 } 422 423 static void amdgpu_xgmi_hive_release(struct kobject *kobj) 424 { 425 struct amdgpu_hive_info *hive = container_of( 426 kobj, struct amdgpu_hive_info, kobj); 427 428 amdgpu_reset_put_reset_domain(hive->reset_domain); 429 hive->reset_domain = NULL; 430 431 mutex_destroy(&hive->hive_lock); 432 kfree(hive); 433 } 434 435 static const struct sysfs_ops amdgpu_xgmi_hive_ops = { 436 .show = amdgpu_xgmi_show_attrs, 437 }; 438 439 static const struct kobj_type amdgpu_xgmi_hive_type = { 440 .release = amdgpu_xgmi_hive_release, 441 .sysfs_ops = &amdgpu_xgmi_hive_ops, 442 .default_groups = amdgpu_xgmi_hive_groups, 443 }; 444 445 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev, 446 struct device_attribute *attr, 447 char *buf) 448 { 449 struct drm_device *ddev = dev_get_drvdata(dev); 450 struct amdgpu_device *adev = drm_to_adev(ddev); 451 452 return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id); 453 454 } 455 456 static ssize_t amdgpu_xgmi_show_physical_id(struct device *dev, 457 struct device_attribute *attr, 458 char *buf) 459 { 460 struct drm_device *ddev = dev_get_drvdata(dev); 461 struct amdgpu_device *adev = drm_to_adev(ddev); 462 463 return sysfs_emit(buf, "%u\n", adev->gmc.xgmi.physical_node_id); 464 465 } 466 467 static ssize_t amdgpu_xgmi_show_num_hops(struct device *dev, 468 struct device_attribute *attr, 469 char *buf) 470 { 471 struct drm_device *ddev = dev_get_drvdata(dev); 472 struct amdgpu_device *adev = drm_to_adev(ddev); 473 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 474 int i; 475 476 for (i = 0; i < top->num_nodes; i++) 477 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_hops); 478 479 return sysfs_emit(buf, "%s\n", buf); 480 } 481 482 static ssize_t amdgpu_xgmi_show_num_links(struct device *dev, 483 struct device_attribute *attr, 484 char *buf) 485 { 486 struct drm_device *ddev = dev_get_drvdata(dev); 487 struct amdgpu_device *adev = drm_to_adev(ddev); 488 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 489 int i; 490 491 for (i = 0; i < top->num_nodes; i++) 492 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_links); 493 494 return sysfs_emit(buf, "%s\n", buf); 495 } 496 497 static ssize_t amdgpu_xgmi_show_connected_port_num(struct device *dev, 498 struct device_attribute *attr, 499 char *buf) 500 { 501 struct drm_device *ddev = dev_get_drvdata(dev); 502 struct amdgpu_device *adev = drm_to_adev(ddev); 503 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 504 int i, j, size = 0; 505 int current_node; 506 /* 507 * get the node id in the sysfs for the current socket and show 508 * it in the port num info output in the sysfs for easy reading. 509 * it is NOT the one retrieved from xgmi ta. 510 */ 511 for (i = 0; i < top->num_nodes; i++) { 512 if (top->nodes[i].node_id == adev->gmc.xgmi.node_id) { 513 current_node = i; 514 break; 515 } 516 } 517 518 if (i == top->num_nodes) 519 return -EINVAL; 520 521 for (i = 0; i < top->num_nodes; i++) { 522 for (j = 0; j < top->nodes[i].num_links; j++) 523 /* node id in sysfs starts from 1 rather than 0 so +1 here */ 524 size += sysfs_emit_at(buf, size, "%02x:%02x -> %02x:%02x\n", current_node + 1, 525 top->nodes[i].port_num[j].src_xgmi_port_num, i + 1, 526 top->nodes[i].port_num[j].dst_xgmi_port_num); 527 } 528 529 return size; 530 } 531 532 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801) 533 static ssize_t amdgpu_xgmi_show_error(struct device *dev, 534 struct device_attribute *attr, 535 char *buf) 536 { 537 struct drm_device *ddev = dev_get_drvdata(dev); 538 struct amdgpu_device *adev = drm_to_adev(ddev); 539 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in; 540 uint64_t fica_out; 541 unsigned int error_count = 0; 542 543 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200); 544 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208); 545 546 if ((!adev->df.funcs) || 547 (!adev->df.funcs->get_fica) || 548 (!adev->df.funcs->set_fica)) 549 return -EINVAL; 550 551 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in); 552 if (fica_out != 0x1f) 553 pr_err("xGMI error counters not enabled!\n"); 554 555 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in); 556 557 if ((fica_out & 0xffff) == 2) 558 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63); 559 560 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0); 561 562 return sysfs_emit(buf, "%u\n", error_count); 563 } 564 565 566 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL); 567 static DEVICE_ATTR(xgmi_physical_id, 0444, amdgpu_xgmi_show_physical_id, NULL); 568 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL); 569 static DEVICE_ATTR(xgmi_num_hops, S_IRUGO, amdgpu_xgmi_show_num_hops, NULL); 570 static DEVICE_ATTR(xgmi_num_links, S_IRUGO, amdgpu_xgmi_show_num_links, NULL); 571 static DEVICE_ATTR(xgmi_port_num, S_IRUGO, amdgpu_xgmi_show_connected_port_num, NULL); 572 573 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev, 574 struct amdgpu_hive_info *hive) 575 { 576 int ret = 0; 577 char node[10] = { 0 }; 578 579 /* Create xgmi device id file */ 580 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id); 581 if (ret) { 582 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n"); 583 return ret; 584 } 585 586 ret = device_create_file(adev->dev, &dev_attr_xgmi_physical_id); 587 if (ret) { 588 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_physical_id\n"); 589 return ret; 590 } 591 592 /* Create xgmi error file */ 593 ret = device_create_file(adev->dev, &dev_attr_xgmi_error); 594 if (ret) 595 pr_err("failed to create xgmi_error\n"); 596 597 /* Create xgmi num hops file */ 598 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_hops); 599 if (ret) 600 pr_err("failed to create xgmi_num_hops\n"); 601 602 /* Create xgmi num links file */ 603 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_links); 604 if (ret) 605 pr_err("failed to create xgmi_num_links\n"); 606 607 /* Create xgmi port num file if supported */ 608 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) { 609 ret = device_create_file(adev->dev, &dev_attr_xgmi_port_num); 610 if (ret) 611 dev_err(adev->dev, "failed to create xgmi_port_num\n"); 612 } 613 614 /* Create sysfs link to hive info folder on the first device */ 615 if (hive->kobj.parent != (&adev->dev->kobj)) { 616 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj, 617 "xgmi_hive_info"); 618 if (ret) { 619 dev_err(adev->dev, "XGMI: Failed to create link to hive info"); 620 goto remove_file; 621 } 622 } 623 624 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 625 /* Create sysfs link form the hive folder to yourself */ 626 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node); 627 if (ret) { 628 dev_err(adev->dev, "XGMI: Failed to create link from hive info"); 629 goto remove_link; 630 } 631 632 goto success; 633 634 635 remove_link: 636 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique); 637 638 remove_file: 639 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 640 device_remove_file(adev->dev, &dev_attr_xgmi_physical_id); 641 device_remove_file(adev->dev, &dev_attr_xgmi_error); 642 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 643 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 644 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) 645 device_remove_file(adev->dev, &dev_attr_xgmi_port_num); 646 647 success: 648 return ret; 649 } 650 651 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev, 652 struct amdgpu_hive_info *hive) 653 { 654 char node[10]; 655 memset(node, 0, sizeof(node)); 656 657 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 658 device_remove_file(adev->dev, &dev_attr_xgmi_physical_id); 659 device_remove_file(adev->dev, &dev_attr_xgmi_error); 660 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 661 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 662 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) 663 device_remove_file(adev->dev, &dev_attr_xgmi_port_num); 664 665 if (hive->kobj.parent != (&adev->dev->kobj)) 666 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info"); 667 668 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 669 sysfs_remove_link(&hive->kobj, node); 670 671 } 672 673 674 675 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev) 676 { 677 struct amdgpu_hive_info *hive = NULL; 678 int ret; 679 680 if (!adev->gmc.xgmi.hive_id) 681 return NULL; 682 683 if (adev->hive) { 684 kobject_get(&adev->hive->kobj); 685 return adev->hive; 686 } 687 688 mutex_lock(&xgmi_mutex); 689 690 list_for_each_entry(hive, &xgmi_hive_list, node) { 691 if (hive->hive_id == adev->gmc.xgmi.hive_id) 692 goto pro_end; 693 } 694 695 hive = kzalloc(sizeof(*hive), GFP_KERNEL); 696 if (!hive) { 697 dev_err(adev->dev, "XGMI: allocation failed\n"); 698 ret = -ENOMEM; 699 hive = NULL; 700 goto pro_end; 701 } 702 703 /* initialize new hive if not exist */ 704 ret = kobject_init_and_add(&hive->kobj, 705 &amdgpu_xgmi_hive_type, 706 &adev->dev->kobj, 707 "%s", "xgmi_hive_info"); 708 if (ret) { 709 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n"); 710 kobject_put(&hive->kobj); 711 hive = NULL; 712 goto pro_end; 713 } 714 715 /** 716 * Only init hive->reset_domain for none SRIOV configuration. For SRIOV, 717 * Host driver decide how to reset the GPU either through FLR or chain reset. 718 * Guest side will get individual notifications from the host for the FLR 719 * if necessary. 720 */ 721 if (!amdgpu_sriov_vf(adev)) { 722 /** 723 * Avoid recreating reset domain when hive is reconstructed for the case 724 * of reset the devices in the XGMI hive during probe for passthrough GPU 725 * See https://www.spinics.net/lists/amd-gfx/msg58836.html 726 */ 727 if (adev->reset_domain->type != XGMI_HIVE) { 728 hive->reset_domain = 729 amdgpu_reset_create_reset_domain(XGMI_HIVE, "amdgpu-reset-hive"); 730 if (!hive->reset_domain) { 731 dev_err(adev->dev, "XGMI: failed initializing reset domain for xgmi hive\n"); 732 ret = -ENOMEM; 733 kobject_put(&hive->kobj); 734 hive = NULL; 735 goto pro_end; 736 } 737 } else { 738 amdgpu_reset_get_reset_domain(adev->reset_domain); 739 hive->reset_domain = adev->reset_domain; 740 } 741 } 742 743 hive->hive_id = adev->gmc.xgmi.hive_id; 744 INIT_LIST_HEAD(&hive->device_list); 745 INIT_LIST_HEAD(&hive->node); 746 mutex_init(&hive->hive_lock); 747 atomic_set(&hive->number_devices, 0); 748 task_barrier_init(&hive->tb); 749 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN; 750 hive->hi_req_gpu = NULL; 751 atomic_set(&hive->requested_nps_mode, UNKNOWN_MEMORY_PARTITION_MODE); 752 753 /* 754 * hive pstate on boot is high in vega20 so we have to go to low 755 * pstate on after boot. 756 */ 757 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE; 758 list_add_tail(&hive->node, &xgmi_hive_list); 759 760 pro_end: 761 if (hive) 762 kobject_get(&hive->kobj); 763 mutex_unlock(&xgmi_mutex); 764 return hive; 765 } 766 767 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive) 768 { 769 if (hive) 770 kobject_put(&hive->kobj); 771 } 772 773 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate) 774 { 775 int ret = 0; 776 struct amdgpu_hive_info *hive; 777 struct amdgpu_device *request_adev; 778 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20; 779 bool init_low; 780 781 hive = amdgpu_get_xgmi_hive(adev); 782 if (!hive) 783 return 0; 784 785 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev; 786 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN; 787 amdgpu_put_xgmi_hive(hive); 788 /* fw bug so temporarily disable pstate switching */ 789 return 0; 790 791 if (!hive || adev->asic_type != CHIP_VEGA20) 792 return 0; 793 794 mutex_lock(&hive->hive_lock); 795 796 if (is_hi_req) 797 hive->hi_req_count++; 798 else 799 hive->hi_req_count--; 800 801 /* 802 * Vega20 only needs single peer to request pstate high for the hive to 803 * go high but all peers must request pstate low for the hive to go low 804 */ 805 if (hive->pstate == pstate || 806 (!is_hi_req && hive->hi_req_count && !init_low)) 807 goto out; 808 809 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate); 810 811 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate); 812 if (ret) { 813 dev_err(request_adev->dev, 814 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d", 815 request_adev->gmc.xgmi.node_id, 816 request_adev->gmc.xgmi.hive_id, ret); 817 goto out; 818 } 819 820 if (init_low) 821 hive->pstate = hive->hi_req_count ? 822 hive->pstate : AMDGPU_XGMI_PSTATE_MIN; 823 else { 824 hive->pstate = pstate; 825 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ? 826 adev : NULL; 827 } 828 out: 829 mutex_unlock(&hive->hive_lock); 830 return ret; 831 } 832 833 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev) 834 { 835 int ret; 836 837 if (amdgpu_sriov_vf(adev)) 838 return 0; 839 840 /* Each psp need to set the latest topology */ 841 ret = psp_xgmi_set_topology_info(&adev->psp, 842 atomic_read(&hive->number_devices), 843 &adev->psp.xgmi_context.top_info); 844 if (ret) 845 dev_err(adev->dev, 846 "XGMI: Set topology failure on device %llx, hive %llx, ret %d", 847 adev->gmc.xgmi.node_id, 848 adev->gmc.xgmi.hive_id, ret); 849 850 return ret; 851 } 852 853 854 /* 855 * NOTE psp_xgmi_node_info.num_hops layout is as follows: 856 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved) 857 * num_hops[5:3] = reserved 858 * num_hops[2:0] = number of hops 859 */ 860 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev, 861 struct amdgpu_device *peer_adev) 862 { 863 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 864 uint8_t num_hops_mask = 0x7; 865 int i; 866 867 if (!adev->gmc.xgmi.supported) 868 return 0; 869 870 for (i = 0 ; i < top->num_nodes; ++i) 871 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 872 return top->nodes[i].num_hops & num_hops_mask; 873 874 dev_err(adev->dev, "Failed to get xgmi hops count for peer %d.\n", 875 peer_adev->gmc.xgmi.physical_node_id); 876 877 return 0; 878 } 879 880 int amdgpu_xgmi_get_bandwidth(struct amdgpu_device *adev, struct amdgpu_device *peer_adev, 881 enum amdgpu_xgmi_bw_mode bw_mode, enum amdgpu_xgmi_bw_unit bw_unit, 882 uint32_t *min_bw, uint32_t *max_bw) 883 { 884 bool peer_mode = bw_mode == AMDGPU_XGMI_BW_MODE_PER_PEER; 885 int unit_scale = bw_unit == AMDGPU_XGMI_BW_UNIT_MBYTES ? 1000 : 1; 886 int num_lanes = adev->gmc.xgmi.max_width; 887 int speed = adev->gmc.xgmi.max_speed; 888 int num_links = !peer_mode ? 1 : -1; 889 890 if (!(min_bw && max_bw)) 891 return -EINVAL; 892 893 *min_bw = 0; 894 *max_bw = 0; 895 896 if (!adev->gmc.xgmi.supported) 897 return -ENODATA; 898 899 if (peer_mode && !peer_adev) 900 return -EINVAL; 901 902 if (peer_mode) { 903 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 904 int i; 905 906 for (i = 0 ; i < top->num_nodes; ++i) { 907 if (top->nodes[i].node_id != peer_adev->gmc.xgmi.node_id) 908 continue; 909 910 num_links = top->nodes[i].num_links; 911 break; 912 } 913 } 914 915 if (num_links == -1) { 916 dev_err(adev->dev, "Failed to get number of xgmi links for peer %d.\n", 917 peer_adev->gmc.xgmi.physical_node_id); 918 } else if (num_links) { 919 int per_link_bw = (speed * num_lanes * unit_scale)/BITS_PER_BYTE; 920 921 *min_bw = per_link_bw; 922 *max_bw = num_links * per_link_bw; 923 } 924 925 return 0; 926 } 927 928 bool amdgpu_xgmi_get_is_sharing_enabled(struct amdgpu_device *adev, 929 struct amdgpu_device *peer_adev) 930 { 931 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 932 int i; 933 934 /* Sharing should always be enabled for non-SRIOV. */ 935 if (!amdgpu_sriov_vf(adev)) 936 return true; 937 938 for (i = 0 ; i < top->num_nodes; ++i) 939 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 940 return !!top->nodes[i].is_sharing_enabled; 941 942 return false; 943 } 944 945 /* 946 * Devices that support extended data require the entire hive to initialize with 947 * the shared memory buffer flag set. 948 * 949 * Hive locks and conditions apply - see amdgpu_xgmi_add_device 950 */ 951 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive, 952 bool set_extended_data) 953 { 954 struct amdgpu_device *tmp_adev; 955 int ret; 956 957 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 958 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false); 959 if (ret) { 960 dev_err(tmp_adev->dev, 961 "XGMI: Failed to initialize xgmi session for data partition %i\n", 962 set_extended_data); 963 return ret; 964 } 965 966 } 967 968 return 0; 969 } 970 971 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 972 { 973 struct psp_xgmi_topology_info *top_info; 974 struct amdgpu_hive_info *hive; 975 struct amdgpu_xgmi *entry; 976 struct amdgpu_device *tmp_adev = NULL; 977 978 int count = 0, ret = 0; 979 980 if (!adev->gmc.xgmi.supported) 981 return 0; 982 983 if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 984 ret = psp_xgmi_initialize(&adev->psp, false, true); 985 if (ret) { 986 dev_err(adev->dev, 987 "XGMI: Failed to initialize xgmi session\n"); 988 return ret; 989 } 990 991 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 992 if (ret) { 993 dev_err(adev->dev, 994 "XGMI: Failed to get hive id\n"); 995 return ret; 996 } 997 998 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 999 if (ret) { 1000 dev_err(adev->dev, 1001 "XGMI: Failed to get node id\n"); 1002 return ret; 1003 } 1004 } else { 1005 adev->gmc.xgmi.hive_id = 16; 1006 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 1007 } 1008 1009 hive = amdgpu_get_xgmi_hive(adev); 1010 if (!hive) { 1011 ret = -EINVAL; 1012 dev_err(adev->dev, 1013 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 1014 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 1015 goto exit; 1016 } 1017 mutex_lock(&hive->hive_lock); 1018 1019 top_info = &adev->psp.xgmi_context.top_info; 1020 1021 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 1022 list_for_each_entry(entry, &hive->device_list, head) 1023 top_info->nodes[count++].node_id = entry->node_id; 1024 top_info->num_nodes = count; 1025 atomic_set(&hive->number_devices, count); 1026 1027 task_barrier_add_task(&hive->tb); 1028 1029 if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 1030 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1031 /* update node list for other device in the hive */ 1032 if (tmp_adev != adev) { 1033 top_info = &tmp_adev->psp.xgmi_context.top_info; 1034 top_info->nodes[count - 1].node_id = 1035 adev->gmc.xgmi.node_id; 1036 top_info->num_nodes = count; 1037 } 1038 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 1039 if (ret) 1040 goto exit_unlock; 1041 } 1042 1043 if (amdgpu_sriov_vf(adev) && 1044 adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) { 1045 /* only get topology for VF being init if it can support full duplex */ 1046 ret = psp_xgmi_get_topology_info(&adev->psp, count, 1047 &adev->psp.xgmi_context.top_info, false); 1048 if (ret) { 1049 dev_err(adev->dev, 1050 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 1051 adev->gmc.xgmi.node_id, 1052 adev->gmc.xgmi.hive_id, ret); 1053 /* To do: continue with some node failed or disable the whole hive*/ 1054 goto exit_unlock; 1055 } 1056 } else { 1057 /* get latest topology info for each device from psp */ 1058 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1059 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 1060 &tmp_adev->psp.xgmi_context.top_info, false); 1061 if (ret) { 1062 dev_err(tmp_adev->dev, 1063 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 1064 tmp_adev->gmc.xgmi.node_id, 1065 tmp_adev->gmc.xgmi.hive_id, ret); 1066 /* To do : continue with some node failed or disable the whole hive */ 1067 goto exit_unlock; 1068 } 1069 } 1070 } 1071 1072 /* get topology again for hives that support extended data */ 1073 if (adev->psp.xgmi_context.supports_extended_data) { 1074 1075 /* initialize the hive to get extended data. */ 1076 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true); 1077 if (ret) 1078 goto exit_unlock; 1079 1080 /* get the extended data. */ 1081 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1082 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 1083 &tmp_adev->psp.xgmi_context.top_info, true); 1084 if (ret) { 1085 dev_err(tmp_adev->dev, 1086 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d", 1087 tmp_adev->gmc.xgmi.node_id, 1088 tmp_adev->gmc.xgmi.hive_id, ret); 1089 goto exit_unlock; 1090 } 1091 } 1092 1093 /* initialize the hive to get non-extended data for the next round. */ 1094 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false); 1095 if (ret) 1096 goto exit_unlock; 1097 1098 } 1099 } 1100 1101 if (!ret) 1102 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 1103 1104 exit_unlock: 1105 mutex_unlock(&hive->hive_lock); 1106 exit: 1107 if (!ret) { 1108 adev->hive = hive; 1109 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 1110 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 1111 } else { 1112 amdgpu_put_xgmi_hive(hive); 1113 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 1114 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 1115 ret); 1116 } 1117 1118 return ret; 1119 } 1120 1121 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 1122 { 1123 struct amdgpu_hive_info *hive = adev->hive; 1124 1125 if (!adev->gmc.xgmi.supported) 1126 return -EINVAL; 1127 1128 if (!hive) 1129 return -EINVAL; 1130 1131 mutex_lock(&hive->hive_lock); 1132 task_barrier_rem_task(&hive->tb); 1133 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 1134 if (hive->hi_req_gpu == adev) 1135 hive->hi_req_gpu = NULL; 1136 list_del(&adev->gmc.xgmi.head); 1137 mutex_unlock(&hive->hive_lock); 1138 1139 amdgpu_put_xgmi_hive(hive); 1140 adev->hive = NULL; 1141 1142 if (atomic_dec_return(&hive->number_devices) == 0) { 1143 /* Remove the hive from global hive list */ 1144 mutex_lock(&xgmi_mutex); 1145 list_del(&hive->node); 1146 mutex_unlock(&xgmi_mutex); 1147 1148 amdgpu_put_xgmi_hive(hive); 1149 } 1150 1151 return 0; 1152 } 1153 1154 static int xgmi_v6_4_0_aca_bank_parser(struct aca_handle *handle, struct aca_bank *bank, 1155 enum aca_smu_type type, void *data) 1156 { 1157 struct amdgpu_device *adev = handle->adev; 1158 struct aca_bank_info info; 1159 const char *error_str; 1160 u64 status, count; 1161 int ret, ext_error_code; 1162 1163 ret = aca_bank_info_decode(bank, &info); 1164 if (ret) 1165 return ret; 1166 1167 status = bank->regs[ACA_REG_IDX_STATUS]; 1168 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); 1169 1170 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1171 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1172 if (error_str) 1173 dev_info(adev->dev, "%s detected\n", error_str); 1174 1175 count = ACA_REG__MISC0__ERRCNT(bank->regs[ACA_REG_IDX_MISC0]); 1176 1177 switch (type) { 1178 case ACA_SMU_TYPE_UE: 1179 if (ext_error_code != 0 && ext_error_code != 9) 1180 count = 0ULL; 1181 1182 bank->aca_err_type = ACA_ERROR_TYPE_UE; 1183 ret = aca_error_cache_log_bank_error(handle, &info, ACA_ERROR_TYPE_UE, count); 1184 break; 1185 case ACA_SMU_TYPE_CE: 1186 count = ext_error_code == 6 ? count : 0ULL; 1187 bank->aca_err_type = ACA_ERROR_TYPE_CE; 1188 ret = aca_error_cache_log_bank_error(handle, &info, bank->aca_err_type, count); 1189 break; 1190 default: 1191 return -EINVAL; 1192 } 1193 1194 return ret; 1195 } 1196 1197 static const struct aca_bank_ops xgmi_v6_4_0_aca_bank_ops = { 1198 .aca_bank_parser = xgmi_v6_4_0_aca_bank_parser, 1199 }; 1200 1201 static const struct aca_info xgmi_v6_4_0_aca_info = { 1202 .hwip = ACA_HWIP_TYPE_PCS_XGMI, 1203 .mask = ACA_ERROR_UE_MASK | ACA_ERROR_CE_MASK, 1204 .bank_ops = &xgmi_v6_4_0_aca_bank_ops, 1205 }; 1206 1207 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 1208 { 1209 int r; 1210 1211 if (!adev->gmc.xgmi.supported || 1212 adev->gmc.xgmi.num_physical_nodes == 0) 1213 return 0; 1214 1215 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1216 1217 r = amdgpu_ras_block_late_init(adev, ras_block); 1218 if (r) 1219 return r; 1220 1221 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1222 case IP_VERSION(6, 4, 0): 1223 case IP_VERSION(6, 4, 1): 1224 r = amdgpu_ras_bind_aca(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL, 1225 &xgmi_v6_4_0_aca_info, NULL); 1226 if (r) 1227 goto late_fini; 1228 break; 1229 default: 1230 break; 1231 } 1232 1233 return 0; 1234 1235 late_fini: 1236 amdgpu_ras_block_late_fini(adev, ras_block); 1237 1238 return r; 1239 } 1240 1241 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 1242 uint64_t addr) 1243 { 1244 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 1245 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 1246 } 1247 1248 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 1249 { 1250 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 1251 WREG32_PCIE(pcs_status_reg, 0); 1252 } 1253 1254 static void amdgpu_xgmi_legacy_reset_ras_error_count(struct amdgpu_device *adev) 1255 { 1256 uint32_t i; 1257 1258 switch (adev->asic_type) { 1259 case CHIP_ARCTURUS: 1260 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 1261 pcs_clear_status(adev, 1262 xgmi_pcs_err_status_reg_arct[i]); 1263 break; 1264 case CHIP_VEGA20: 1265 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 1266 pcs_clear_status(adev, 1267 xgmi_pcs_err_status_reg_vg20[i]); 1268 break; 1269 case CHIP_ALDEBARAN: 1270 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) 1271 pcs_clear_status(adev, 1272 xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1273 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) 1274 pcs_clear_status(adev, 1275 walf_pcs_err_status_reg_aldebaran[i]); 1276 break; 1277 default: 1278 break; 1279 } 1280 1281 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1282 case IP_VERSION(6, 4, 0): 1283 case IP_VERSION(6, 4, 1): 1284 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) 1285 pcs_clear_status(adev, 1286 xgmi3x16_pcs_err_status_reg_v6_4[i]); 1287 break; 1288 default: 1289 break; 1290 } 1291 } 1292 1293 static void __xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst, u64 mca_base) 1294 { 1295 WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL); 1296 } 1297 1298 static void xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst) 1299 { 1300 int i; 1301 1302 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1303 __xgmi_v6_4_0_reset_error_count(adev, xgmi_inst, xgmi_v6_4_0_mca_base_array[i]); 1304 } 1305 1306 static void xgmi_v6_4_0_reset_ras_error_count(struct amdgpu_device *adev) 1307 { 1308 int i; 1309 1310 for_each_inst(i, adev->aid_mask) 1311 xgmi_v6_4_0_reset_error_count(adev, i); 1312 } 1313 1314 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 1315 { 1316 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1317 case IP_VERSION(6, 4, 0): 1318 case IP_VERSION(6, 4, 1): 1319 xgmi_v6_4_0_reset_ras_error_count(adev); 1320 break; 1321 default: 1322 amdgpu_xgmi_legacy_reset_ras_error_count(adev); 1323 break; 1324 } 1325 } 1326 1327 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 1328 uint32_t value, 1329 uint32_t mask_value, 1330 uint32_t *ue_count, 1331 uint32_t *ce_count, 1332 bool is_xgmi_pcs, 1333 bool check_mask) 1334 { 1335 int i; 1336 int ue_cnt = 0; 1337 const struct amdgpu_pcs_ras_field *pcs_ras_fields = NULL; 1338 uint32_t field_array_size = 0; 1339 1340 if (is_xgmi_pcs) { 1341 if (amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1342 IP_VERSION(6, 1, 0) || 1343 amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1344 IP_VERSION(6, 4, 0) || 1345 amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1346 IP_VERSION(6, 4, 1)) { 1347 pcs_ras_fields = &xgmi3x16_pcs_ras_fields[0]; 1348 field_array_size = ARRAY_SIZE(xgmi3x16_pcs_ras_fields); 1349 } else { 1350 pcs_ras_fields = &xgmi_pcs_ras_fields[0]; 1351 field_array_size = ARRAY_SIZE(xgmi_pcs_ras_fields); 1352 } 1353 } else { 1354 pcs_ras_fields = &wafl_pcs_ras_fields[0]; 1355 field_array_size = ARRAY_SIZE(wafl_pcs_ras_fields); 1356 } 1357 1358 if (check_mask) 1359 value = value & ~mask_value; 1360 1361 /* query xgmi/walf pcs error status, 1362 * only ue is supported */ 1363 for (i = 0; value && i < field_array_size; i++) { 1364 ue_cnt = (value & 1365 pcs_ras_fields[i].pcs_err_mask) >> 1366 pcs_ras_fields[i].pcs_err_shift; 1367 if (ue_cnt) { 1368 dev_info(adev->dev, "%s detected\n", 1369 pcs_ras_fields[i].err_name); 1370 *ue_count += ue_cnt; 1371 } 1372 1373 /* reset bit value if the bit is checked */ 1374 value &= ~(pcs_ras_fields[i].pcs_err_mask); 1375 } 1376 1377 return 0; 1378 } 1379 1380 static void amdgpu_xgmi_legacy_query_ras_error_count(struct amdgpu_device *adev, 1381 void *ras_error_status) 1382 { 1383 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1384 int i, supported = 1; 1385 uint32_t data, mask_data = 0; 1386 uint32_t ue_cnt = 0, ce_cnt = 0; 1387 1388 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 1389 return ; 1390 1391 err_data->ue_count = 0; 1392 err_data->ce_count = 0; 1393 1394 switch (adev->asic_type) { 1395 case CHIP_ARCTURUS: 1396 /* check xgmi pcs error */ 1397 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 1398 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 1399 if (data) 1400 amdgpu_xgmi_query_pcs_error_status(adev, data, 1401 mask_data, &ue_cnt, &ce_cnt, true, false); 1402 } 1403 /* check wafl pcs error */ 1404 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 1405 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 1406 if (data) 1407 amdgpu_xgmi_query_pcs_error_status(adev, data, 1408 mask_data, &ue_cnt, &ce_cnt, false, false); 1409 } 1410 break; 1411 case CHIP_VEGA20: 1412 /* check xgmi pcs error */ 1413 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 1414 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 1415 if (data) 1416 amdgpu_xgmi_query_pcs_error_status(adev, data, 1417 mask_data, &ue_cnt, &ce_cnt, true, false); 1418 } 1419 /* check wafl pcs error */ 1420 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 1421 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 1422 if (data) 1423 amdgpu_xgmi_query_pcs_error_status(adev, data, 1424 mask_data, &ue_cnt, &ce_cnt, false, false); 1425 } 1426 break; 1427 case CHIP_ALDEBARAN: 1428 /* check xgmi3x16 pcs error */ 1429 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) { 1430 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1431 mask_data = 1432 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1433 if (data) 1434 amdgpu_xgmi_query_pcs_error_status(adev, data, 1435 mask_data, &ue_cnt, &ce_cnt, true, true); 1436 } 1437 /* check wafl pcs error */ 1438 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) { 1439 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]); 1440 mask_data = 1441 RREG32_PCIE(walf_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1442 if (data) 1443 amdgpu_xgmi_query_pcs_error_status(adev, data, 1444 mask_data, &ue_cnt, &ce_cnt, false, true); 1445 } 1446 break; 1447 default: 1448 supported = 0; 1449 break; 1450 } 1451 1452 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1453 case IP_VERSION(6, 4, 0): 1454 case IP_VERSION(6, 4, 1): 1455 /* check xgmi3x16 pcs error */ 1456 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) { 1457 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_v6_4[i]); 1458 mask_data = 1459 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[i]); 1460 if (data) 1461 amdgpu_xgmi_query_pcs_error_status(adev, data, 1462 mask_data, &ue_cnt, &ce_cnt, true, true); 1463 } 1464 break; 1465 default: 1466 if (!supported) 1467 dev_warn(adev->dev, "XGMI RAS error query not supported"); 1468 break; 1469 } 1470 1471 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1472 1473 err_data->ue_count += ue_cnt; 1474 err_data->ce_count += ce_cnt; 1475 } 1476 1477 static enum aca_error_type xgmi_v6_4_0_pcs_mca_get_error_type(struct amdgpu_device *adev, u64 status) 1478 { 1479 const char *error_str; 1480 int ext_error_code; 1481 1482 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); 1483 1484 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1485 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1486 if (error_str) 1487 dev_info(adev->dev, "%s detected\n", error_str); 1488 1489 switch (ext_error_code) { 1490 case 0: 1491 return ACA_ERROR_TYPE_UE; 1492 case 6: 1493 return ACA_ERROR_TYPE_CE; 1494 default: 1495 return -EINVAL; 1496 } 1497 1498 return -EINVAL; 1499 } 1500 1501 static void __xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, struct amdgpu_smuio_mcm_config_info *mcm_info, 1502 u64 mca_base, struct ras_err_data *err_data) 1503 { 1504 int xgmi_inst = mcm_info->die_id; 1505 u64 status = 0; 1506 1507 status = RREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS); 1508 if (!ACA_REG__STATUS__VAL(status)) 1509 return; 1510 1511 switch (xgmi_v6_4_0_pcs_mca_get_error_type(adev, status)) { 1512 case ACA_ERROR_TYPE_UE: 1513 amdgpu_ras_error_statistic_ue_count(err_data, mcm_info, 1ULL); 1514 break; 1515 case ACA_ERROR_TYPE_CE: 1516 amdgpu_ras_error_statistic_ce_count(err_data, mcm_info, 1ULL); 1517 break; 1518 default: 1519 break; 1520 } 1521 1522 WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL); 1523 } 1524 1525 static void xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, int xgmi_inst, struct ras_err_data *err_data) 1526 { 1527 struct amdgpu_smuio_mcm_config_info mcm_info = { 1528 .socket_id = adev->smuio.funcs->get_socket_id(adev), 1529 .die_id = xgmi_inst, 1530 }; 1531 int i; 1532 1533 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1534 __xgmi_v6_4_0_query_error_count(adev, &mcm_info, xgmi_v6_4_0_mca_base_array[i], err_data); 1535 } 1536 1537 static void xgmi_v6_4_0_query_ras_error_count(struct amdgpu_device *adev, void *ras_error_status) 1538 { 1539 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1540 int i; 1541 1542 for_each_inst(i, adev->aid_mask) 1543 xgmi_v6_4_0_query_error_count(adev, i, err_data); 1544 } 1545 1546 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 1547 void *ras_error_status) 1548 { 1549 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1550 case IP_VERSION(6, 4, 0): 1551 case IP_VERSION(6, 4, 1): 1552 xgmi_v6_4_0_query_ras_error_count(adev, ras_error_status); 1553 break; 1554 default: 1555 amdgpu_xgmi_legacy_query_ras_error_count(adev, ras_error_status); 1556 break; 1557 } 1558 } 1559 1560 /* Trigger XGMI/WAFL error */ 1561 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev, 1562 void *inject_if, uint32_t instance_mask) 1563 { 1564 int ret1, ret2; 1565 struct ta_ras_trigger_error_input *block_info = 1566 (struct ta_ras_trigger_error_input *)inject_if; 1567 1568 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) 1569 dev_warn(adev->dev, "Failed to disallow df cstate"); 1570 1571 ret1 = amdgpu_dpm_set_pm_policy(adev, PP_PM_POLICY_XGMI_PLPD, XGMI_PLPD_DISALLOW); 1572 if (ret1 && ret1 != -EOPNOTSUPP) 1573 dev_warn(adev->dev, "Failed to disallow XGMI power down"); 1574 1575 ret2 = psp_ras_trigger_error(&adev->psp, block_info, instance_mask); 1576 1577 if (amdgpu_ras_intr_triggered()) 1578 return ret2; 1579 1580 ret1 = amdgpu_dpm_set_pm_policy(adev, PP_PM_POLICY_XGMI_PLPD, XGMI_PLPD_DEFAULT); 1581 if (ret1 && ret1 != -EOPNOTSUPP) 1582 dev_warn(adev->dev, "Failed to allow XGMI power down"); 1583 1584 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW)) 1585 dev_warn(adev->dev, "Failed to allow df cstate"); 1586 1587 return ret2; 1588 } 1589 1590 struct amdgpu_ras_block_hw_ops xgmi_ras_hw_ops = { 1591 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 1592 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 1593 .ras_error_inject = amdgpu_ras_error_inject_xgmi, 1594 }; 1595 1596 struct amdgpu_xgmi_ras xgmi_ras = { 1597 .ras_block = { 1598 .hw_ops = &xgmi_ras_hw_ops, 1599 .ras_late_init = amdgpu_xgmi_ras_late_init, 1600 }, 1601 }; 1602 1603 int amdgpu_xgmi_ras_sw_init(struct amdgpu_device *adev) 1604 { 1605 int err; 1606 struct amdgpu_xgmi_ras *ras; 1607 1608 if (!adev->gmc.xgmi.ras) 1609 return 0; 1610 1611 ras = adev->gmc.xgmi.ras; 1612 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 1613 if (err) { 1614 dev_err(adev->dev, "Failed to register xgmi_wafl_pcs ras block!\n"); 1615 return err; 1616 } 1617 1618 strcpy(ras->ras_block.ras_comm.name, "xgmi_wafl"); 1619 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__XGMI_WAFL; 1620 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 1621 adev->gmc.xgmi.ras_if = &ras->ras_block.ras_comm; 1622 1623 return 0; 1624 } 1625 1626 static void amdgpu_xgmi_reset_on_init_work(struct work_struct *work) 1627 { 1628 struct amdgpu_hive_info *hive = 1629 container_of(work, struct amdgpu_hive_info, reset_on_init_work); 1630 struct amdgpu_reset_context reset_context; 1631 struct amdgpu_device *tmp_adev; 1632 struct list_head device_list; 1633 int r; 1634 1635 mutex_lock(&hive->hive_lock); 1636 1637 INIT_LIST_HEAD(&device_list); 1638 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) 1639 list_add_tail(&tmp_adev->reset_list, &device_list); 1640 1641 tmp_adev = list_first_entry(&device_list, struct amdgpu_device, 1642 reset_list); 1643 amdgpu_device_lock_reset_domain(tmp_adev->reset_domain); 1644 1645 reset_context.method = AMD_RESET_METHOD_ON_INIT; 1646 reset_context.reset_req_dev = tmp_adev; 1647 reset_context.hive = hive; 1648 reset_context.reset_device_list = &device_list; 1649 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); 1650 set_bit(AMDGPU_SKIP_COREDUMP, &reset_context.flags); 1651 1652 amdgpu_reset_do_xgmi_reset_on_init(&reset_context); 1653 mutex_unlock(&hive->hive_lock); 1654 amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain); 1655 1656 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1657 r = amdgpu_ras_init_badpage_info(tmp_adev); 1658 if (r && r != -EHWPOISON) 1659 dev_err(tmp_adev->dev, 1660 "error during bad page data initialization"); 1661 } 1662 } 1663 1664 static void amdgpu_xgmi_schedule_reset_on_init(struct amdgpu_hive_info *hive) 1665 { 1666 INIT_WORK(&hive->reset_on_init_work, amdgpu_xgmi_reset_on_init_work); 1667 amdgpu_reset_domain_schedule(hive->reset_domain, 1668 &hive->reset_on_init_work); 1669 } 1670 1671 int amdgpu_xgmi_reset_on_init(struct amdgpu_device *adev) 1672 { 1673 struct amdgpu_hive_info *hive; 1674 bool reset_scheduled; 1675 int num_devs; 1676 1677 hive = amdgpu_get_xgmi_hive(adev); 1678 if (!hive) 1679 return -EINVAL; 1680 1681 mutex_lock(&hive->hive_lock); 1682 num_devs = atomic_read(&hive->number_devices); 1683 reset_scheduled = false; 1684 if (num_devs == adev->gmc.xgmi.num_physical_nodes) { 1685 amdgpu_xgmi_schedule_reset_on_init(hive); 1686 reset_scheduled = true; 1687 } 1688 1689 mutex_unlock(&hive->hive_lock); 1690 amdgpu_put_xgmi_hive(hive); 1691 1692 if (reset_scheduled) 1693 flush_work(&hive->reset_on_init_work); 1694 1695 return 0; 1696 } 1697 1698 int amdgpu_xgmi_request_nps_change(struct amdgpu_device *adev, 1699 struct amdgpu_hive_info *hive, 1700 int req_nps_mode) 1701 { 1702 struct amdgpu_device *tmp_adev; 1703 int cur_nps_mode, r; 1704 1705 /* This is expected to be called only during unload of driver. The 1706 * request needs to be placed only once for all devices in the hive. If 1707 * one of them fail, revert the request for previous successful devices. 1708 * After placing the request, make hive mode as UNKNOWN so that other 1709 * devices don't request anymore. 1710 */ 1711 mutex_lock(&hive->hive_lock); 1712 if (atomic_read(&hive->requested_nps_mode) == 1713 UNKNOWN_MEMORY_PARTITION_MODE) { 1714 dev_dbg(adev->dev, "Unexpected entry for hive NPS change"); 1715 mutex_unlock(&hive->hive_lock); 1716 return 0; 1717 } 1718 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1719 r = adev->gmc.gmc_funcs->request_mem_partition_mode( 1720 tmp_adev, req_nps_mode); 1721 if (r) 1722 break; 1723 } 1724 if (r) { 1725 /* Request back current mode if one of the requests failed */ 1726 cur_nps_mode = 1727 adev->gmc.gmc_funcs->query_mem_partition_mode(tmp_adev); 1728 list_for_each_entry_continue_reverse( 1729 tmp_adev, &hive->device_list, gmc.xgmi.head) 1730 adev->gmc.gmc_funcs->request_mem_partition_mode( 1731 tmp_adev, cur_nps_mode); 1732 } 1733 /* Set to UNKNOWN so that other devices don't request anymore */ 1734 atomic_set(&hive->requested_nps_mode, UNKNOWN_MEMORY_PARTITION_MODE); 1735 mutex_unlock(&hive->hive_lock); 1736 1737 return r; 1738 } 1739 1740 bool amdgpu_xgmi_same_hive(struct amdgpu_device *adev, 1741 struct amdgpu_device *bo_adev) 1742 { 1743 return (amdgpu_use_xgmi_p2p && adev != bo_adev && 1744 adev->gmc.xgmi.hive_id && 1745 adev->gmc.xgmi.hive_id == bo_adev->gmc.xgmi.hive_id); 1746 } 1747 1748 void amdgpu_xgmi_early_init(struct amdgpu_device *adev) 1749 { 1750 if (!adev->gmc.xgmi.supported) 1751 return; 1752 1753 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1754 case IP_VERSION(9, 4, 0): 1755 case IP_VERSION(9, 4, 1): 1756 case IP_VERSION(9, 4, 2): 1757 /* 25 GT/s */ 1758 adev->gmc.xgmi.max_speed = 25; 1759 adev->gmc.xgmi.max_width = 16; 1760 break; 1761 case IP_VERSION(9, 4, 3): 1762 case IP_VERSION(9, 4, 4): 1763 case IP_VERSION(9, 5, 0): 1764 /* 32 GT/s */ 1765 adev->gmc.xgmi.max_speed = 32; 1766 adev->gmc.xgmi.max_width = 16; 1767 break; 1768 default: 1769 break; 1770 } 1771 } 1772 1773 void amgpu_xgmi_set_max_speed_width(struct amdgpu_device *adev, 1774 uint16_t max_speed, uint8_t max_width) 1775 { 1776 adev->gmc.xgmi.max_speed = max_speed; 1777 adev->gmc.xgmi.max_width = max_width; 1778 } 1779