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 static u32 xgmi_v6_4_get_link_status(struct amdgpu_device *adev, int global_link_num) 298 { 299 const u32 smnpcs_xgmi3x16_pcs_state_hist1 = 0x11a00070; 300 const int xgmi_inst = 2; 301 u32 link_inst; 302 u64 addr; 303 304 link_inst = global_link_num % xgmi_inst; 305 306 addr = (smnpcs_xgmi3x16_pcs_state_hist1 | (link_inst << 20)) + 307 adev->asic_funcs->encode_ext_smn_addressing(global_link_num / xgmi_inst); 308 309 return RREG32_PCIE_EXT(addr); 310 } 311 312 int amdgpu_get_xgmi_link_status(struct amdgpu_device *adev, int global_link_num) 313 { 314 u32 xgmi_state_reg_val; 315 316 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 317 case IP_VERSION(6, 4, 0): 318 case IP_VERSION(6, 4, 1): 319 xgmi_state_reg_val = xgmi_v6_4_get_link_status(adev, global_link_num); 320 break; 321 default: 322 return -EOPNOTSUPP; 323 } 324 325 if ((xgmi_state_reg_val & 0xFF) == XGMI_STATE_DISABLE) 326 return -ENOLINK; 327 328 if ((xgmi_state_reg_val & 0xFF) == XGMI_STATE_LS0) 329 return XGMI_LINK_ACTIVE; 330 331 return XGMI_LINK_INACTIVE; 332 } 333 334 /** 335 * DOC: AMDGPU XGMI Support 336 * 337 * XGMI is a high speed interconnect that joins multiple GPU cards 338 * into a homogeneous memory space that is organized by a collective 339 * hive ID and individual node IDs, both of which are 64-bit numbers. 340 * 341 * The file xgmi_device_id contains the unique per GPU device ID and 342 * is stored in the /sys/class/drm/card${cardno}/device/ directory. 343 * 344 * Inside the device directory a sub-directory 'xgmi_hive_info' is 345 * created which contains the hive ID and the list of nodes. 346 * 347 * The hive ID is stored in: 348 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id 349 * 350 * The node information is stored in numbered directories: 351 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id 352 * 353 * Each device has their own xgmi_hive_info direction with a mirror 354 * set of node sub-directories. 355 * 356 * The XGMI memory space is built by contiguously adding the power of 357 * two padded VRAM space from each node to each other. 358 * 359 */ 360 361 static struct attribute amdgpu_xgmi_hive_id = { 362 .name = "xgmi_hive_id", 363 .mode = S_IRUGO 364 }; 365 366 static struct attribute *amdgpu_xgmi_hive_attrs[] = { 367 &amdgpu_xgmi_hive_id, 368 NULL 369 }; 370 ATTRIBUTE_GROUPS(amdgpu_xgmi_hive); 371 372 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj, 373 struct attribute *attr, char *buf) 374 { 375 struct amdgpu_hive_info *hive = container_of( 376 kobj, struct amdgpu_hive_info, kobj); 377 378 if (attr == &amdgpu_xgmi_hive_id) 379 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id); 380 381 return 0; 382 } 383 384 static void amdgpu_xgmi_hive_release(struct kobject *kobj) 385 { 386 struct amdgpu_hive_info *hive = container_of( 387 kobj, struct amdgpu_hive_info, kobj); 388 389 amdgpu_reset_put_reset_domain(hive->reset_domain); 390 hive->reset_domain = NULL; 391 392 mutex_destroy(&hive->hive_lock); 393 kfree(hive); 394 } 395 396 static const struct sysfs_ops amdgpu_xgmi_hive_ops = { 397 .show = amdgpu_xgmi_show_attrs, 398 }; 399 400 static const struct kobj_type amdgpu_xgmi_hive_type = { 401 .release = amdgpu_xgmi_hive_release, 402 .sysfs_ops = &amdgpu_xgmi_hive_ops, 403 .default_groups = amdgpu_xgmi_hive_groups, 404 }; 405 406 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev, 407 struct device_attribute *attr, 408 char *buf) 409 { 410 struct drm_device *ddev = dev_get_drvdata(dev); 411 struct amdgpu_device *adev = drm_to_adev(ddev); 412 413 return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id); 414 415 } 416 417 static ssize_t amdgpu_xgmi_show_physical_id(struct device *dev, 418 struct device_attribute *attr, 419 char *buf) 420 { 421 struct drm_device *ddev = dev_get_drvdata(dev); 422 struct amdgpu_device *adev = drm_to_adev(ddev); 423 424 return sysfs_emit(buf, "%u\n", adev->gmc.xgmi.physical_node_id); 425 426 } 427 428 static ssize_t amdgpu_xgmi_show_num_hops(struct device *dev, 429 struct device_attribute *attr, 430 char *buf) 431 { 432 struct drm_device *ddev = dev_get_drvdata(dev); 433 struct amdgpu_device *adev = drm_to_adev(ddev); 434 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 435 int i; 436 437 for (i = 0; i < top->num_nodes; i++) 438 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_hops); 439 440 return sysfs_emit(buf, "%s\n", buf); 441 } 442 443 static ssize_t amdgpu_xgmi_show_num_links(struct device *dev, 444 struct device_attribute *attr, 445 char *buf) 446 { 447 struct drm_device *ddev = dev_get_drvdata(dev); 448 struct amdgpu_device *adev = drm_to_adev(ddev); 449 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 450 int i; 451 452 for (i = 0; i < top->num_nodes; i++) 453 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_links); 454 455 return sysfs_emit(buf, "%s\n", buf); 456 } 457 458 static ssize_t amdgpu_xgmi_show_connected_port_num(struct device *dev, 459 struct device_attribute *attr, 460 char *buf) 461 { 462 struct drm_device *ddev = dev_get_drvdata(dev); 463 struct amdgpu_device *adev = drm_to_adev(ddev); 464 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 465 int i, j, size = 0; 466 int current_node; 467 /* 468 * get the node id in the sysfs for the current socket and show 469 * it in the port num info output in the sysfs for easy reading. 470 * it is NOT the one retrieved from xgmi ta. 471 */ 472 for (i = 0; i < top->num_nodes; i++) { 473 if (top->nodes[i].node_id == adev->gmc.xgmi.node_id) { 474 current_node = i; 475 break; 476 } 477 } 478 479 if (i == top->num_nodes) 480 return -EINVAL; 481 482 for (i = 0; i < top->num_nodes; i++) { 483 for (j = 0; j < top->nodes[i].num_links; j++) 484 /* node id in sysfs starts from 1 rather than 0 so +1 here */ 485 size += sysfs_emit_at(buf, size, "%02x:%02x -> %02x:%02x\n", current_node + 1, 486 top->nodes[i].port_num[j].src_xgmi_port_num, i + 1, 487 top->nodes[i].port_num[j].dst_xgmi_port_num); 488 } 489 490 return size; 491 } 492 493 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801) 494 static ssize_t amdgpu_xgmi_show_error(struct device *dev, 495 struct device_attribute *attr, 496 char *buf) 497 { 498 struct drm_device *ddev = dev_get_drvdata(dev); 499 struct amdgpu_device *adev = drm_to_adev(ddev); 500 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in; 501 uint64_t fica_out; 502 unsigned int error_count = 0; 503 504 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200); 505 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208); 506 507 if ((!adev->df.funcs) || 508 (!adev->df.funcs->get_fica) || 509 (!adev->df.funcs->set_fica)) 510 return -EINVAL; 511 512 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in); 513 if (fica_out != 0x1f) 514 pr_err("xGMI error counters not enabled!\n"); 515 516 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in); 517 518 if ((fica_out & 0xffff) == 2) 519 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63); 520 521 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0); 522 523 return sysfs_emit(buf, "%u\n", error_count); 524 } 525 526 527 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL); 528 static DEVICE_ATTR(xgmi_physical_id, 0444, amdgpu_xgmi_show_physical_id, NULL); 529 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL); 530 static DEVICE_ATTR(xgmi_num_hops, S_IRUGO, amdgpu_xgmi_show_num_hops, NULL); 531 static DEVICE_ATTR(xgmi_num_links, S_IRUGO, amdgpu_xgmi_show_num_links, NULL); 532 static DEVICE_ATTR(xgmi_port_num, S_IRUGO, amdgpu_xgmi_show_connected_port_num, NULL); 533 534 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev, 535 struct amdgpu_hive_info *hive) 536 { 537 int ret = 0; 538 char node[10] = { 0 }; 539 540 /* Create xgmi device id file */ 541 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id); 542 if (ret) { 543 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n"); 544 return ret; 545 } 546 547 ret = device_create_file(adev->dev, &dev_attr_xgmi_physical_id); 548 if (ret) { 549 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_physical_id\n"); 550 return ret; 551 } 552 553 /* Create xgmi error file */ 554 ret = device_create_file(adev->dev, &dev_attr_xgmi_error); 555 if (ret) 556 pr_err("failed to create xgmi_error\n"); 557 558 /* Create xgmi num hops file */ 559 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_hops); 560 if (ret) 561 pr_err("failed to create xgmi_num_hops\n"); 562 563 /* Create xgmi num links file */ 564 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_links); 565 if (ret) 566 pr_err("failed to create xgmi_num_links\n"); 567 568 /* Create xgmi port num file if supported */ 569 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) { 570 ret = device_create_file(adev->dev, &dev_attr_xgmi_port_num); 571 if (ret) 572 dev_err(adev->dev, "failed to create xgmi_port_num\n"); 573 } 574 575 /* Create sysfs link to hive info folder on the first device */ 576 if (hive->kobj.parent != (&adev->dev->kobj)) { 577 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj, 578 "xgmi_hive_info"); 579 if (ret) { 580 dev_err(adev->dev, "XGMI: Failed to create link to hive info"); 581 goto remove_file; 582 } 583 } 584 585 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 586 /* Create sysfs link form the hive folder to yourself */ 587 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node); 588 if (ret) { 589 dev_err(adev->dev, "XGMI: Failed to create link from hive info"); 590 goto remove_link; 591 } 592 593 goto success; 594 595 596 remove_link: 597 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique); 598 599 remove_file: 600 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 601 device_remove_file(adev->dev, &dev_attr_xgmi_physical_id); 602 device_remove_file(adev->dev, &dev_attr_xgmi_error); 603 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 604 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 605 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) 606 device_remove_file(adev->dev, &dev_attr_xgmi_port_num); 607 608 success: 609 return ret; 610 } 611 612 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev, 613 struct amdgpu_hive_info *hive) 614 { 615 char node[10]; 616 memset(node, 0, sizeof(node)); 617 618 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 619 device_remove_file(adev->dev, &dev_attr_xgmi_physical_id); 620 device_remove_file(adev->dev, &dev_attr_xgmi_error); 621 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 622 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 623 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) 624 device_remove_file(adev->dev, &dev_attr_xgmi_port_num); 625 626 if (hive->kobj.parent != (&adev->dev->kobj)) 627 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info"); 628 629 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 630 sysfs_remove_link(&hive->kobj, node); 631 632 } 633 634 635 636 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev) 637 { 638 struct amdgpu_hive_info *hive = NULL; 639 int ret; 640 641 if (!adev->gmc.xgmi.hive_id) 642 return NULL; 643 644 if (adev->hive) { 645 kobject_get(&adev->hive->kobj); 646 return adev->hive; 647 } 648 649 mutex_lock(&xgmi_mutex); 650 651 list_for_each_entry(hive, &xgmi_hive_list, node) { 652 if (hive->hive_id == adev->gmc.xgmi.hive_id) 653 goto pro_end; 654 } 655 656 hive = kzalloc(sizeof(*hive), GFP_KERNEL); 657 if (!hive) { 658 dev_err(adev->dev, "XGMI: allocation failed\n"); 659 ret = -ENOMEM; 660 hive = NULL; 661 goto pro_end; 662 } 663 664 /* initialize new hive if not exist */ 665 ret = kobject_init_and_add(&hive->kobj, 666 &amdgpu_xgmi_hive_type, 667 &adev->dev->kobj, 668 "%s", "xgmi_hive_info"); 669 if (ret) { 670 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n"); 671 kobject_put(&hive->kobj); 672 hive = NULL; 673 goto pro_end; 674 } 675 676 /** 677 * Only init hive->reset_domain for none SRIOV configuration. For SRIOV, 678 * Host driver decide how to reset the GPU either through FLR or chain reset. 679 * Guest side will get individual notifications from the host for the FLR 680 * if necessary. 681 */ 682 if (!amdgpu_sriov_vf(adev)) { 683 /** 684 * Avoid recreating reset domain when hive is reconstructed for the case 685 * of reset the devices in the XGMI hive during probe for passthrough GPU 686 * See https://www.spinics.net/lists/amd-gfx/msg58836.html 687 */ 688 if (adev->reset_domain->type != XGMI_HIVE) { 689 hive->reset_domain = 690 amdgpu_reset_create_reset_domain(XGMI_HIVE, "amdgpu-reset-hive"); 691 if (!hive->reset_domain) { 692 dev_err(adev->dev, "XGMI: failed initializing reset domain for xgmi hive\n"); 693 ret = -ENOMEM; 694 kobject_put(&hive->kobj); 695 hive = NULL; 696 goto pro_end; 697 } 698 } else { 699 amdgpu_reset_get_reset_domain(adev->reset_domain); 700 hive->reset_domain = adev->reset_domain; 701 } 702 } 703 704 hive->hive_id = adev->gmc.xgmi.hive_id; 705 INIT_LIST_HEAD(&hive->device_list); 706 INIT_LIST_HEAD(&hive->node); 707 mutex_init(&hive->hive_lock); 708 atomic_set(&hive->number_devices, 0); 709 task_barrier_init(&hive->tb); 710 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN; 711 hive->hi_req_gpu = NULL; 712 atomic_set(&hive->requested_nps_mode, UNKNOWN_MEMORY_PARTITION_MODE); 713 714 /* 715 * hive pstate on boot is high in vega20 so we have to go to low 716 * pstate on after boot. 717 */ 718 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE; 719 list_add_tail(&hive->node, &xgmi_hive_list); 720 721 pro_end: 722 if (hive) 723 kobject_get(&hive->kobj); 724 mutex_unlock(&xgmi_mutex); 725 return hive; 726 } 727 728 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive) 729 { 730 if (hive) 731 kobject_put(&hive->kobj); 732 } 733 734 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate) 735 { 736 int ret = 0; 737 struct amdgpu_hive_info *hive; 738 struct amdgpu_device *request_adev; 739 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20; 740 bool init_low; 741 742 hive = amdgpu_get_xgmi_hive(adev); 743 if (!hive) 744 return 0; 745 746 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev; 747 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN; 748 amdgpu_put_xgmi_hive(hive); 749 /* fw bug so temporarily disable pstate switching */ 750 return 0; 751 752 if (!hive || adev->asic_type != CHIP_VEGA20) 753 return 0; 754 755 mutex_lock(&hive->hive_lock); 756 757 if (is_hi_req) 758 hive->hi_req_count++; 759 else 760 hive->hi_req_count--; 761 762 /* 763 * Vega20 only needs single peer to request pstate high for the hive to 764 * go high but all peers must request pstate low for the hive to go low 765 */ 766 if (hive->pstate == pstate || 767 (!is_hi_req && hive->hi_req_count && !init_low)) 768 goto out; 769 770 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate); 771 772 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate); 773 if (ret) { 774 dev_err(request_adev->dev, 775 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d", 776 request_adev->gmc.xgmi.node_id, 777 request_adev->gmc.xgmi.hive_id, ret); 778 goto out; 779 } 780 781 if (init_low) 782 hive->pstate = hive->hi_req_count ? 783 hive->pstate : AMDGPU_XGMI_PSTATE_MIN; 784 else { 785 hive->pstate = pstate; 786 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ? 787 adev : NULL; 788 } 789 out: 790 mutex_unlock(&hive->hive_lock); 791 return ret; 792 } 793 794 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev) 795 { 796 int ret; 797 798 if (amdgpu_sriov_vf(adev)) 799 return 0; 800 801 /* Each psp need to set the latest topology */ 802 ret = psp_xgmi_set_topology_info(&adev->psp, 803 atomic_read(&hive->number_devices), 804 &adev->psp.xgmi_context.top_info); 805 if (ret) 806 dev_err(adev->dev, 807 "XGMI: Set topology failure on device %llx, hive %llx, ret %d", 808 adev->gmc.xgmi.node_id, 809 adev->gmc.xgmi.hive_id, ret); 810 811 return ret; 812 } 813 814 815 /* 816 * NOTE psp_xgmi_node_info.num_hops layout is as follows: 817 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved) 818 * num_hops[5:3] = reserved 819 * num_hops[2:0] = number of hops 820 */ 821 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev, 822 struct amdgpu_device *peer_adev) 823 { 824 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 825 uint8_t num_hops_mask = 0x7; 826 int i; 827 828 if (!adev->gmc.xgmi.supported) 829 return 0; 830 831 for (i = 0 ; i < top->num_nodes; ++i) 832 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 833 return top->nodes[i].num_hops & num_hops_mask; 834 835 dev_err(adev->dev, "Failed to get xgmi hops count for peer %d.\n", 836 peer_adev->gmc.xgmi.physical_node_id); 837 838 return 0; 839 } 840 841 int amdgpu_xgmi_get_bandwidth(struct amdgpu_device *adev, struct amdgpu_device *peer_adev, 842 enum amdgpu_xgmi_bw_mode bw_mode, enum amdgpu_xgmi_bw_unit bw_unit, 843 uint32_t *min_bw, uint32_t *max_bw) 844 { 845 bool peer_mode = bw_mode == AMDGPU_XGMI_BW_MODE_PER_PEER; 846 int unit_scale = bw_unit == AMDGPU_XGMI_BW_UNIT_MBYTES ? 1000 : 1; 847 int num_lanes = adev->gmc.xgmi.max_width; 848 int speed = adev->gmc.xgmi.max_speed; 849 int num_links = !peer_mode ? 1 : -1; 850 851 if (!(min_bw && max_bw)) 852 return -EINVAL; 853 854 *min_bw = 0; 855 *max_bw = 0; 856 857 if (!adev->gmc.xgmi.supported) 858 return -ENODATA; 859 860 if (peer_mode && !peer_adev) 861 return -EINVAL; 862 863 if (peer_mode) { 864 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 865 int i; 866 867 for (i = 0 ; i < top->num_nodes; ++i) { 868 if (top->nodes[i].node_id != peer_adev->gmc.xgmi.node_id) 869 continue; 870 871 num_links = top->nodes[i].num_links; 872 break; 873 } 874 } 875 876 if (num_links == -1) { 877 dev_err(adev->dev, "Failed to get number of xgmi links for peer %d.\n", 878 peer_adev->gmc.xgmi.physical_node_id); 879 } else if (num_links) { 880 int per_link_bw = (speed * num_lanes * unit_scale)/BITS_PER_BYTE; 881 882 *min_bw = per_link_bw; 883 *max_bw = num_links * per_link_bw; 884 } 885 886 return 0; 887 } 888 889 bool amdgpu_xgmi_get_is_sharing_enabled(struct amdgpu_device *adev, 890 struct amdgpu_device *peer_adev) 891 { 892 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 893 int i; 894 895 /* Sharing should always be enabled for non-SRIOV. */ 896 if (!amdgpu_sriov_vf(adev)) 897 return true; 898 899 for (i = 0 ; i < top->num_nodes; ++i) 900 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 901 return !!top->nodes[i].is_sharing_enabled; 902 903 return false; 904 } 905 906 /* 907 * Devices that support extended data require the entire hive to initialize with 908 * the shared memory buffer flag set. 909 * 910 * Hive locks and conditions apply - see amdgpu_xgmi_add_device 911 */ 912 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive, 913 bool set_extended_data) 914 { 915 struct amdgpu_device *tmp_adev; 916 int ret; 917 918 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 919 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false); 920 if (ret) { 921 dev_err(tmp_adev->dev, 922 "XGMI: Failed to initialize xgmi session for data partition %i\n", 923 set_extended_data); 924 return ret; 925 } 926 927 } 928 929 return 0; 930 } 931 932 static void amdgpu_xgmi_fill_topology_info(struct amdgpu_device *adev, 933 struct amdgpu_device *peer_adev) 934 { 935 struct psp_xgmi_topology_info *top_info = &adev->psp.xgmi_context.top_info; 936 struct psp_xgmi_topology_info *peer_info = &peer_adev->psp.xgmi_context.top_info; 937 938 for (int i = 0; i < peer_info->num_nodes; i++) { 939 if (peer_info->nodes[i].node_id == adev->gmc.xgmi.node_id) { 940 for (int j = 0; j < top_info->num_nodes; j++) { 941 if (top_info->nodes[j].node_id == peer_adev->gmc.xgmi.node_id) { 942 peer_info->nodes[i].num_hops = top_info->nodes[j].num_hops; 943 peer_info->nodes[i].is_sharing_enabled = 944 top_info->nodes[j].is_sharing_enabled; 945 peer_info->nodes[i].num_links = 946 top_info->nodes[j].num_links; 947 return; 948 } 949 } 950 } 951 } 952 } 953 954 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 955 { 956 struct psp_xgmi_topology_info *top_info; 957 struct amdgpu_hive_info *hive; 958 struct amdgpu_xgmi *entry; 959 struct amdgpu_device *tmp_adev = NULL; 960 961 int count = 0, ret = 0; 962 963 if (!adev->gmc.xgmi.supported) 964 return 0; 965 966 if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 967 ret = psp_xgmi_initialize(&adev->psp, false, true); 968 if (ret) { 969 dev_err(adev->dev, 970 "XGMI: Failed to initialize xgmi session\n"); 971 return ret; 972 } 973 974 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 975 if (ret) { 976 dev_err(adev->dev, 977 "XGMI: Failed to get hive id\n"); 978 return ret; 979 } 980 981 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 982 if (ret) { 983 dev_err(adev->dev, 984 "XGMI: Failed to get node id\n"); 985 return ret; 986 } 987 } else { 988 adev->gmc.xgmi.hive_id = 16; 989 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 990 } 991 992 hive = amdgpu_get_xgmi_hive(adev); 993 if (!hive) { 994 ret = -EINVAL; 995 dev_err(adev->dev, 996 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 997 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 998 goto exit; 999 } 1000 mutex_lock(&hive->hive_lock); 1001 1002 top_info = &adev->psp.xgmi_context.top_info; 1003 1004 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 1005 list_for_each_entry(entry, &hive->device_list, head) 1006 top_info->nodes[count++].node_id = entry->node_id; 1007 top_info->num_nodes = count; 1008 atomic_set(&hive->number_devices, count); 1009 1010 task_barrier_add_task(&hive->tb); 1011 1012 if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 1013 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1014 /* update node list for other device in the hive */ 1015 if (tmp_adev != adev) { 1016 top_info = &tmp_adev->psp.xgmi_context.top_info; 1017 top_info->nodes[count - 1].node_id = 1018 adev->gmc.xgmi.node_id; 1019 top_info->num_nodes = count; 1020 } 1021 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 1022 if (ret) 1023 goto exit_unlock; 1024 } 1025 1026 if (amdgpu_sriov_vf(adev) && 1027 adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) { 1028 /* only get topology for VF being init if it can support full duplex */ 1029 ret = psp_xgmi_get_topology_info(&adev->psp, count, 1030 &adev->psp.xgmi_context.top_info, false); 1031 if (ret) { 1032 dev_err(adev->dev, 1033 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 1034 adev->gmc.xgmi.node_id, 1035 adev->gmc.xgmi.hive_id, ret); 1036 /* To do: continue with some node failed or disable the whole hive*/ 1037 goto exit_unlock; 1038 } 1039 1040 /* fill the topology info for peers instead of getting from PSP */ 1041 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1042 amdgpu_xgmi_fill_topology_info(adev, tmp_adev); 1043 } 1044 } else { 1045 /* get latest topology info for each device from psp */ 1046 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1047 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 1048 &tmp_adev->psp.xgmi_context.top_info, false); 1049 if (ret) { 1050 dev_err(tmp_adev->dev, 1051 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 1052 tmp_adev->gmc.xgmi.node_id, 1053 tmp_adev->gmc.xgmi.hive_id, ret); 1054 /* To do : continue with some node failed or disable the whole hive */ 1055 goto exit_unlock; 1056 } 1057 } 1058 } 1059 1060 /* get topology again for hives that support extended data */ 1061 if (adev->psp.xgmi_context.supports_extended_data) { 1062 1063 /* initialize the hive to get extended data. */ 1064 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true); 1065 if (ret) 1066 goto exit_unlock; 1067 1068 /* get the extended data. */ 1069 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1070 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 1071 &tmp_adev->psp.xgmi_context.top_info, true); 1072 if (ret) { 1073 dev_err(tmp_adev->dev, 1074 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d", 1075 tmp_adev->gmc.xgmi.node_id, 1076 tmp_adev->gmc.xgmi.hive_id, ret); 1077 goto exit_unlock; 1078 } 1079 } 1080 1081 /* initialize the hive to get non-extended data for the next round. */ 1082 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false); 1083 if (ret) 1084 goto exit_unlock; 1085 1086 } 1087 } 1088 1089 if (!ret) 1090 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 1091 1092 exit_unlock: 1093 mutex_unlock(&hive->hive_lock); 1094 exit: 1095 if (!ret) { 1096 adev->hive = hive; 1097 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 1098 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 1099 } else { 1100 amdgpu_put_xgmi_hive(hive); 1101 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 1102 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 1103 ret); 1104 } 1105 1106 return ret; 1107 } 1108 1109 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 1110 { 1111 struct amdgpu_hive_info *hive = adev->hive; 1112 1113 if (!adev->gmc.xgmi.supported) 1114 return -EINVAL; 1115 1116 if (!hive) 1117 return -EINVAL; 1118 1119 mutex_lock(&hive->hive_lock); 1120 task_barrier_rem_task(&hive->tb); 1121 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 1122 if (hive->hi_req_gpu == adev) 1123 hive->hi_req_gpu = NULL; 1124 list_del(&adev->gmc.xgmi.head); 1125 mutex_unlock(&hive->hive_lock); 1126 1127 amdgpu_put_xgmi_hive(hive); 1128 adev->hive = NULL; 1129 1130 if (atomic_dec_return(&hive->number_devices) == 0) { 1131 /* Remove the hive from global hive list */ 1132 mutex_lock(&xgmi_mutex); 1133 list_del(&hive->node); 1134 mutex_unlock(&xgmi_mutex); 1135 1136 amdgpu_put_xgmi_hive(hive); 1137 } 1138 1139 return 0; 1140 } 1141 1142 static int xgmi_v6_4_0_aca_bank_parser(struct aca_handle *handle, struct aca_bank *bank, 1143 enum aca_smu_type type, void *data) 1144 { 1145 struct amdgpu_device *adev = handle->adev; 1146 struct aca_bank_info info; 1147 const char *error_str; 1148 u64 status, count; 1149 int ret, ext_error_code; 1150 1151 ret = aca_bank_info_decode(bank, &info); 1152 if (ret) 1153 return ret; 1154 1155 status = bank->regs[ACA_REG_IDX_STATUS]; 1156 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); 1157 1158 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1159 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1160 if (error_str) 1161 dev_info(adev->dev, "%s detected\n", error_str); 1162 1163 count = ACA_REG__MISC0__ERRCNT(bank->regs[ACA_REG_IDX_MISC0]); 1164 1165 switch (type) { 1166 case ACA_SMU_TYPE_UE: 1167 if (ext_error_code != 0 && ext_error_code != 9) 1168 count = 0ULL; 1169 1170 bank->aca_err_type = ACA_ERROR_TYPE_UE; 1171 ret = aca_error_cache_log_bank_error(handle, &info, ACA_ERROR_TYPE_UE, count); 1172 break; 1173 case ACA_SMU_TYPE_CE: 1174 count = ext_error_code == 6 ? count : 0ULL; 1175 bank->aca_err_type = ACA_ERROR_TYPE_CE; 1176 ret = aca_error_cache_log_bank_error(handle, &info, bank->aca_err_type, count); 1177 break; 1178 default: 1179 return -EINVAL; 1180 } 1181 1182 return ret; 1183 } 1184 1185 static const struct aca_bank_ops xgmi_v6_4_0_aca_bank_ops = { 1186 .aca_bank_parser = xgmi_v6_4_0_aca_bank_parser, 1187 }; 1188 1189 static const struct aca_info xgmi_v6_4_0_aca_info = { 1190 .hwip = ACA_HWIP_TYPE_PCS_XGMI, 1191 .mask = ACA_ERROR_UE_MASK | ACA_ERROR_CE_MASK, 1192 .bank_ops = &xgmi_v6_4_0_aca_bank_ops, 1193 }; 1194 1195 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 1196 { 1197 int r; 1198 1199 if (!adev->gmc.xgmi.supported || 1200 adev->gmc.xgmi.num_physical_nodes == 0) 1201 return 0; 1202 1203 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1204 1205 r = amdgpu_ras_block_late_init(adev, ras_block); 1206 if (r) 1207 return r; 1208 1209 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1210 case IP_VERSION(6, 4, 0): 1211 case IP_VERSION(6, 4, 1): 1212 r = amdgpu_ras_bind_aca(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL, 1213 &xgmi_v6_4_0_aca_info, NULL); 1214 if (r) 1215 goto late_fini; 1216 break; 1217 default: 1218 break; 1219 } 1220 1221 return 0; 1222 1223 late_fini: 1224 amdgpu_ras_block_late_fini(adev, ras_block); 1225 1226 return r; 1227 } 1228 1229 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 1230 uint64_t addr) 1231 { 1232 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 1233 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 1234 } 1235 1236 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 1237 { 1238 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 1239 WREG32_PCIE(pcs_status_reg, 0); 1240 } 1241 1242 static void amdgpu_xgmi_legacy_reset_ras_error_count(struct amdgpu_device *adev) 1243 { 1244 uint32_t i; 1245 1246 switch (adev->asic_type) { 1247 case CHIP_ARCTURUS: 1248 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 1249 pcs_clear_status(adev, 1250 xgmi_pcs_err_status_reg_arct[i]); 1251 break; 1252 case CHIP_VEGA20: 1253 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 1254 pcs_clear_status(adev, 1255 xgmi_pcs_err_status_reg_vg20[i]); 1256 break; 1257 case CHIP_ALDEBARAN: 1258 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) 1259 pcs_clear_status(adev, 1260 xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1261 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) 1262 pcs_clear_status(adev, 1263 walf_pcs_err_status_reg_aldebaran[i]); 1264 break; 1265 default: 1266 break; 1267 } 1268 1269 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1270 case IP_VERSION(6, 4, 0): 1271 case IP_VERSION(6, 4, 1): 1272 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) 1273 pcs_clear_status(adev, 1274 xgmi3x16_pcs_err_status_reg_v6_4[i]); 1275 break; 1276 default: 1277 break; 1278 } 1279 } 1280 1281 static void __xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst, u64 mca_base) 1282 { 1283 WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL); 1284 } 1285 1286 static void xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst) 1287 { 1288 int i; 1289 1290 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1291 __xgmi_v6_4_0_reset_error_count(adev, xgmi_inst, xgmi_v6_4_0_mca_base_array[i]); 1292 } 1293 1294 static void xgmi_v6_4_0_reset_ras_error_count(struct amdgpu_device *adev) 1295 { 1296 int i; 1297 1298 for_each_inst(i, adev->aid_mask) 1299 xgmi_v6_4_0_reset_error_count(adev, i); 1300 } 1301 1302 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 1303 { 1304 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1305 case IP_VERSION(6, 4, 0): 1306 case IP_VERSION(6, 4, 1): 1307 xgmi_v6_4_0_reset_ras_error_count(adev); 1308 break; 1309 default: 1310 amdgpu_xgmi_legacy_reset_ras_error_count(adev); 1311 break; 1312 } 1313 } 1314 1315 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 1316 uint32_t value, 1317 uint32_t mask_value, 1318 uint32_t *ue_count, 1319 uint32_t *ce_count, 1320 bool is_xgmi_pcs, 1321 bool check_mask) 1322 { 1323 int i; 1324 int ue_cnt = 0; 1325 const struct amdgpu_pcs_ras_field *pcs_ras_fields = NULL; 1326 uint32_t field_array_size = 0; 1327 1328 if (is_xgmi_pcs) { 1329 if (amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1330 IP_VERSION(6, 1, 0) || 1331 amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1332 IP_VERSION(6, 4, 0) || 1333 amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1334 IP_VERSION(6, 4, 1)) { 1335 pcs_ras_fields = &xgmi3x16_pcs_ras_fields[0]; 1336 field_array_size = ARRAY_SIZE(xgmi3x16_pcs_ras_fields); 1337 } else { 1338 pcs_ras_fields = &xgmi_pcs_ras_fields[0]; 1339 field_array_size = ARRAY_SIZE(xgmi_pcs_ras_fields); 1340 } 1341 } else { 1342 pcs_ras_fields = &wafl_pcs_ras_fields[0]; 1343 field_array_size = ARRAY_SIZE(wafl_pcs_ras_fields); 1344 } 1345 1346 if (check_mask) 1347 value = value & ~mask_value; 1348 1349 /* query xgmi/walf pcs error status, 1350 * only ue is supported */ 1351 for (i = 0; value && i < field_array_size; i++) { 1352 ue_cnt = (value & 1353 pcs_ras_fields[i].pcs_err_mask) >> 1354 pcs_ras_fields[i].pcs_err_shift; 1355 if (ue_cnt) { 1356 dev_info(adev->dev, "%s detected\n", 1357 pcs_ras_fields[i].err_name); 1358 *ue_count += ue_cnt; 1359 } 1360 1361 /* reset bit value if the bit is checked */ 1362 value &= ~(pcs_ras_fields[i].pcs_err_mask); 1363 } 1364 1365 return 0; 1366 } 1367 1368 static void amdgpu_xgmi_legacy_query_ras_error_count(struct amdgpu_device *adev, 1369 void *ras_error_status) 1370 { 1371 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1372 int i, supported = 1; 1373 uint32_t data, mask_data = 0; 1374 uint32_t ue_cnt = 0, ce_cnt = 0; 1375 1376 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 1377 return ; 1378 1379 err_data->ue_count = 0; 1380 err_data->ce_count = 0; 1381 1382 switch (adev->asic_type) { 1383 case CHIP_ARCTURUS: 1384 /* check xgmi pcs error */ 1385 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 1386 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 1387 if (data) 1388 amdgpu_xgmi_query_pcs_error_status(adev, data, 1389 mask_data, &ue_cnt, &ce_cnt, true, false); 1390 } 1391 /* check wafl pcs error */ 1392 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 1393 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 1394 if (data) 1395 amdgpu_xgmi_query_pcs_error_status(adev, data, 1396 mask_data, &ue_cnt, &ce_cnt, false, false); 1397 } 1398 break; 1399 case CHIP_VEGA20: 1400 /* check xgmi pcs error */ 1401 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 1402 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 1403 if (data) 1404 amdgpu_xgmi_query_pcs_error_status(adev, data, 1405 mask_data, &ue_cnt, &ce_cnt, true, false); 1406 } 1407 /* check wafl pcs error */ 1408 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 1409 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 1410 if (data) 1411 amdgpu_xgmi_query_pcs_error_status(adev, data, 1412 mask_data, &ue_cnt, &ce_cnt, false, false); 1413 } 1414 break; 1415 case CHIP_ALDEBARAN: 1416 /* check xgmi3x16 pcs error */ 1417 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) { 1418 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1419 mask_data = 1420 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1421 if (data) 1422 amdgpu_xgmi_query_pcs_error_status(adev, data, 1423 mask_data, &ue_cnt, &ce_cnt, true, true); 1424 } 1425 /* check wafl pcs error */ 1426 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) { 1427 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]); 1428 mask_data = 1429 RREG32_PCIE(walf_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1430 if (data) 1431 amdgpu_xgmi_query_pcs_error_status(adev, data, 1432 mask_data, &ue_cnt, &ce_cnt, false, true); 1433 } 1434 break; 1435 default: 1436 supported = 0; 1437 break; 1438 } 1439 1440 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1441 case IP_VERSION(6, 4, 0): 1442 case IP_VERSION(6, 4, 1): 1443 /* check xgmi3x16 pcs error */ 1444 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) { 1445 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_v6_4[i]); 1446 mask_data = 1447 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[i]); 1448 if (data) 1449 amdgpu_xgmi_query_pcs_error_status(adev, data, 1450 mask_data, &ue_cnt, &ce_cnt, true, true); 1451 } 1452 break; 1453 default: 1454 if (!supported) 1455 dev_warn(adev->dev, "XGMI RAS error query not supported"); 1456 break; 1457 } 1458 1459 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1460 1461 err_data->ue_count += ue_cnt; 1462 err_data->ce_count += ce_cnt; 1463 } 1464 1465 static enum aca_error_type xgmi_v6_4_0_pcs_mca_get_error_type(struct amdgpu_device *adev, u64 status) 1466 { 1467 const char *error_str; 1468 int ext_error_code; 1469 1470 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); 1471 1472 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1473 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1474 if (error_str) 1475 dev_info(adev->dev, "%s detected\n", error_str); 1476 1477 switch (ext_error_code) { 1478 case 0: 1479 return ACA_ERROR_TYPE_UE; 1480 case 6: 1481 return ACA_ERROR_TYPE_CE; 1482 default: 1483 return -EINVAL; 1484 } 1485 1486 return -EINVAL; 1487 } 1488 1489 static void __xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, struct amdgpu_smuio_mcm_config_info *mcm_info, 1490 u64 mca_base, struct ras_err_data *err_data) 1491 { 1492 int xgmi_inst = mcm_info->die_id; 1493 u64 status = 0; 1494 1495 status = RREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS); 1496 if (!ACA_REG__STATUS__VAL(status)) 1497 return; 1498 1499 switch (xgmi_v6_4_0_pcs_mca_get_error_type(adev, status)) { 1500 case ACA_ERROR_TYPE_UE: 1501 amdgpu_ras_error_statistic_ue_count(err_data, mcm_info, 1ULL); 1502 break; 1503 case ACA_ERROR_TYPE_CE: 1504 amdgpu_ras_error_statistic_ce_count(err_data, mcm_info, 1ULL); 1505 break; 1506 default: 1507 break; 1508 } 1509 1510 WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL); 1511 } 1512 1513 static void xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, int xgmi_inst, struct ras_err_data *err_data) 1514 { 1515 struct amdgpu_smuio_mcm_config_info mcm_info = { 1516 .socket_id = adev->smuio.funcs->get_socket_id(adev), 1517 .die_id = xgmi_inst, 1518 }; 1519 int i; 1520 1521 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1522 __xgmi_v6_4_0_query_error_count(adev, &mcm_info, xgmi_v6_4_0_mca_base_array[i], err_data); 1523 } 1524 1525 static void xgmi_v6_4_0_query_ras_error_count(struct amdgpu_device *adev, void *ras_error_status) 1526 { 1527 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1528 int i; 1529 1530 for_each_inst(i, adev->aid_mask) 1531 xgmi_v6_4_0_query_error_count(adev, i, err_data); 1532 } 1533 1534 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 1535 void *ras_error_status) 1536 { 1537 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1538 case IP_VERSION(6, 4, 0): 1539 case IP_VERSION(6, 4, 1): 1540 xgmi_v6_4_0_query_ras_error_count(adev, ras_error_status); 1541 break; 1542 default: 1543 amdgpu_xgmi_legacy_query_ras_error_count(adev, ras_error_status); 1544 break; 1545 } 1546 } 1547 1548 /* Trigger XGMI/WAFL error */ 1549 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev, 1550 void *inject_if, uint32_t instance_mask) 1551 { 1552 int ret1, ret2; 1553 struct ta_ras_trigger_error_input *block_info = 1554 (struct ta_ras_trigger_error_input *)inject_if; 1555 1556 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) 1557 dev_warn(adev->dev, "Failed to disallow df cstate"); 1558 1559 ret1 = amdgpu_dpm_set_pm_policy(adev, PP_PM_POLICY_XGMI_PLPD, XGMI_PLPD_DISALLOW); 1560 if (ret1 && ret1 != -EOPNOTSUPP) 1561 dev_warn(adev->dev, "Failed to disallow XGMI power down"); 1562 1563 ret2 = psp_ras_trigger_error(&adev->psp, block_info, instance_mask); 1564 1565 if (amdgpu_ras_intr_triggered()) 1566 return ret2; 1567 1568 ret1 = amdgpu_dpm_set_pm_policy(adev, PP_PM_POLICY_XGMI_PLPD, XGMI_PLPD_DEFAULT); 1569 if (ret1 && ret1 != -EOPNOTSUPP) 1570 dev_warn(adev->dev, "Failed to allow XGMI power down"); 1571 1572 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW)) 1573 dev_warn(adev->dev, "Failed to allow df cstate"); 1574 1575 return ret2; 1576 } 1577 1578 struct amdgpu_ras_block_hw_ops xgmi_ras_hw_ops = { 1579 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 1580 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 1581 .ras_error_inject = amdgpu_ras_error_inject_xgmi, 1582 }; 1583 1584 struct amdgpu_xgmi_ras xgmi_ras = { 1585 .ras_block = { 1586 .hw_ops = &xgmi_ras_hw_ops, 1587 .ras_late_init = amdgpu_xgmi_ras_late_init, 1588 }, 1589 }; 1590 1591 int amdgpu_xgmi_ras_sw_init(struct amdgpu_device *adev) 1592 { 1593 int err; 1594 struct amdgpu_xgmi_ras *ras; 1595 1596 if (!adev->gmc.xgmi.ras) 1597 return 0; 1598 1599 ras = adev->gmc.xgmi.ras; 1600 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 1601 if (err) { 1602 dev_err(adev->dev, "Failed to register xgmi_wafl_pcs ras block!\n"); 1603 return err; 1604 } 1605 1606 strcpy(ras->ras_block.ras_comm.name, "xgmi_wafl"); 1607 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__XGMI_WAFL; 1608 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 1609 adev->gmc.xgmi.ras_if = &ras->ras_block.ras_comm; 1610 1611 return 0; 1612 } 1613 1614 static void amdgpu_xgmi_reset_on_init_work(struct work_struct *work) 1615 { 1616 struct amdgpu_hive_info *hive = 1617 container_of(work, struct amdgpu_hive_info, reset_on_init_work); 1618 struct amdgpu_reset_context reset_context; 1619 struct amdgpu_device *tmp_adev; 1620 struct list_head device_list; 1621 int r; 1622 1623 mutex_lock(&hive->hive_lock); 1624 1625 INIT_LIST_HEAD(&device_list); 1626 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) 1627 list_add_tail(&tmp_adev->reset_list, &device_list); 1628 1629 tmp_adev = list_first_entry(&device_list, struct amdgpu_device, 1630 reset_list); 1631 amdgpu_device_lock_reset_domain(tmp_adev->reset_domain); 1632 1633 reset_context.method = AMD_RESET_METHOD_ON_INIT; 1634 reset_context.reset_req_dev = tmp_adev; 1635 reset_context.hive = hive; 1636 reset_context.reset_device_list = &device_list; 1637 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); 1638 set_bit(AMDGPU_SKIP_COREDUMP, &reset_context.flags); 1639 1640 amdgpu_reset_do_xgmi_reset_on_init(&reset_context); 1641 mutex_unlock(&hive->hive_lock); 1642 amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain); 1643 1644 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1645 r = amdgpu_ras_init_badpage_info(tmp_adev); 1646 if (r && r != -EHWPOISON) 1647 dev_err(tmp_adev->dev, 1648 "error during bad page data initialization"); 1649 } 1650 } 1651 1652 static void amdgpu_xgmi_schedule_reset_on_init(struct amdgpu_hive_info *hive) 1653 { 1654 INIT_WORK(&hive->reset_on_init_work, amdgpu_xgmi_reset_on_init_work); 1655 amdgpu_reset_domain_schedule(hive->reset_domain, 1656 &hive->reset_on_init_work); 1657 } 1658 1659 int amdgpu_xgmi_reset_on_init(struct amdgpu_device *adev) 1660 { 1661 struct amdgpu_hive_info *hive; 1662 bool reset_scheduled; 1663 int num_devs; 1664 1665 hive = amdgpu_get_xgmi_hive(adev); 1666 if (!hive) 1667 return -EINVAL; 1668 1669 mutex_lock(&hive->hive_lock); 1670 num_devs = atomic_read(&hive->number_devices); 1671 reset_scheduled = false; 1672 if (num_devs == adev->gmc.xgmi.num_physical_nodes) { 1673 amdgpu_xgmi_schedule_reset_on_init(hive); 1674 reset_scheduled = true; 1675 } 1676 1677 mutex_unlock(&hive->hive_lock); 1678 amdgpu_put_xgmi_hive(hive); 1679 1680 if (reset_scheduled) 1681 flush_work(&hive->reset_on_init_work); 1682 1683 return 0; 1684 } 1685 1686 int amdgpu_xgmi_request_nps_change(struct amdgpu_device *adev, 1687 struct amdgpu_hive_info *hive, 1688 int req_nps_mode) 1689 { 1690 struct amdgpu_device *tmp_adev; 1691 int cur_nps_mode, r; 1692 1693 /* This is expected to be called only during unload of driver. The 1694 * request needs to be placed only once for all devices in the hive. If 1695 * one of them fail, revert the request for previous successful devices. 1696 * After placing the request, make hive mode as UNKNOWN so that other 1697 * devices don't request anymore. 1698 */ 1699 mutex_lock(&hive->hive_lock); 1700 if (atomic_read(&hive->requested_nps_mode) == 1701 UNKNOWN_MEMORY_PARTITION_MODE) { 1702 dev_dbg(adev->dev, "Unexpected entry for hive NPS change"); 1703 mutex_unlock(&hive->hive_lock); 1704 return 0; 1705 } 1706 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 1707 r = adev->gmc.gmc_funcs->request_mem_partition_mode( 1708 tmp_adev, req_nps_mode); 1709 if (r) 1710 break; 1711 } 1712 if (r) { 1713 /* Request back current mode if one of the requests failed */ 1714 cur_nps_mode = 1715 adev->gmc.gmc_funcs->query_mem_partition_mode(tmp_adev); 1716 list_for_each_entry_continue_reverse( 1717 tmp_adev, &hive->device_list, gmc.xgmi.head) 1718 adev->gmc.gmc_funcs->request_mem_partition_mode( 1719 tmp_adev, cur_nps_mode); 1720 } 1721 /* Set to UNKNOWN so that other devices don't request anymore */ 1722 atomic_set(&hive->requested_nps_mode, UNKNOWN_MEMORY_PARTITION_MODE); 1723 mutex_unlock(&hive->hive_lock); 1724 1725 return r; 1726 } 1727 1728 bool amdgpu_xgmi_same_hive(struct amdgpu_device *adev, 1729 struct amdgpu_device *bo_adev) 1730 { 1731 return (amdgpu_use_xgmi_p2p && adev != bo_adev && 1732 adev->gmc.xgmi.hive_id && 1733 adev->gmc.xgmi.hive_id == bo_adev->gmc.xgmi.hive_id); 1734 } 1735 1736 void amdgpu_xgmi_early_init(struct amdgpu_device *adev) 1737 { 1738 if (!adev->gmc.xgmi.supported) 1739 return; 1740 1741 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1742 case IP_VERSION(9, 4, 0): 1743 case IP_VERSION(9, 4, 1): 1744 case IP_VERSION(9, 4, 2): 1745 adev->gmc.xgmi.max_speed = XGMI_SPEED_25GT; 1746 adev->gmc.xgmi.max_width = 16; 1747 break; 1748 case IP_VERSION(9, 4, 3): 1749 case IP_VERSION(9, 4, 4): 1750 case IP_VERSION(9, 5, 0): 1751 adev->gmc.xgmi.max_speed = XGMI_SPEED_32GT; 1752 adev->gmc.xgmi.max_width = 16; 1753 break; 1754 default: 1755 break; 1756 } 1757 } 1758