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 static DEFINE_MUTEX(xgmi_mutex); 44 45 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE 4 46 47 static LIST_HEAD(xgmi_hive_list); 48 49 static const int xgmi_pcs_err_status_reg_vg20[] = { 50 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 51 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 52 }; 53 54 static const int wafl_pcs_err_status_reg_vg20[] = { 55 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 56 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 57 }; 58 59 static const int xgmi_pcs_err_status_reg_arct[] = { 60 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 61 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 62 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000, 63 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000, 64 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000, 65 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000, 66 }; 67 68 /* same as vg20*/ 69 static const int wafl_pcs_err_status_reg_arct[] = { 70 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 71 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 72 }; 73 74 static const int xgmi3x16_pcs_err_status_reg_aldebaran[] = { 75 smnPCS_XGMI3X16_PCS_ERROR_STATUS, 76 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000, 77 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x200000, 78 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x300000, 79 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x400000, 80 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x500000, 81 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x600000, 82 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x700000 83 }; 84 85 static const int xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[] = { 86 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK, 87 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000, 88 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x200000, 89 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x300000, 90 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x400000, 91 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x500000, 92 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x600000, 93 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x700000 94 }; 95 96 static const int walf_pcs_err_status_reg_aldebaran[] = { 97 smnPCS_GOPX1_PCS_ERROR_STATUS, 98 smnPCS_GOPX1_PCS_ERROR_STATUS + 0x100000 99 }; 100 101 static const int walf_pcs_err_noncorrectable_mask_reg_aldebaran[] = { 102 smnPCS_GOPX1_PCS_ERROR_NONCORRECTABLE_MASK, 103 smnPCS_GOPX1_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000 104 }; 105 106 static const int xgmi3x16_pcs_err_status_reg_v6_4[] = { 107 smnPCS_XGMI3X16_PCS_ERROR_STATUS, 108 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000 109 }; 110 111 static const int xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[] = { 112 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK, 113 smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000 114 }; 115 116 static const u64 xgmi_v6_4_0_mca_base_array[] = { 117 0x11a09200, 118 0x11b09200, 119 }; 120 121 static const char *xgmi_v6_4_0_ras_error_code_ext[32] = { 122 [0x00] = "XGMI PCS DataLossErr", 123 [0x01] = "XGMI PCS TrainingErr", 124 [0x02] = "XGMI PCS FlowCtrlAckErr", 125 [0x03] = "XGMI PCS RxFifoUnderflowErr", 126 [0x04] = "XGMI PCS RxFifoOverflowErr", 127 [0x05] = "XGMI PCS CRCErr", 128 [0x06] = "XGMI PCS BERExceededErr", 129 [0x07] = "XGMI PCS TxMetaDataErr", 130 [0x08] = "XGMI PCS ReplayBufParityErr", 131 [0x09] = "XGMI PCS DataParityErr", 132 [0x0a] = "XGMI PCS ReplayFifoOverflowErr", 133 [0x0b] = "XGMI PCS ReplayFifoUnderflowErr", 134 [0x0c] = "XGMI PCS ElasticFifoOverflowErr", 135 [0x0d] = "XGMI PCS DeskewErr", 136 [0x0e] = "XGMI PCS FlowCtrlCRCErr", 137 [0x0f] = "XGMI PCS DataStartupLimitErr", 138 [0x10] = "XGMI PCS FCInitTimeoutErr", 139 [0x11] = "XGMI PCS RecoveryTimeoutErr", 140 [0x12] = "XGMI PCS ReadySerialTimeoutErr", 141 [0x13] = "XGMI PCS ReadySerialAttemptErr", 142 [0x14] = "XGMI PCS RecoveryAttemptErr", 143 [0x15] = "XGMI PCS RecoveryRelockAttemptErr", 144 [0x16] = "XGMI PCS ReplayAttemptErr", 145 [0x17] = "XGMI PCS SyncHdrErr", 146 [0x18] = "XGMI PCS TxReplayTimeoutErr", 147 [0x19] = "XGMI PCS RxReplayTimeoutErr", 148 [0x1a] = "XGMI PCS LinkSubTxTimeoutErr", 149 [0x1b] = "XGMI PCS LinkSubRxTimeoutErr", 150 [0x1c] = "XGMI PCS RxCMDPktErr", 151 }; 152 153 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = { 154 {"XGMI PCS DataLossErr", 155 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)}, 156 {"XGMI PCS TrainingErr", 157 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)}, 158 {"XGMI PCS CRCErr", 159 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)}, 160 {"XGMI PCS BERExceededErr", 161 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)}, 162 {"XGMI PCS TxMetaDataErr", 163 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)}, 164 {"XGMI PCS ReplayBufParityErr", 165 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 166 {"XGMI PCS DataParityErr", 167 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)}, 168 {"XGMI PCS ReplayFifoOverflowErr", 169 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 170 {"XGMI PCS ReplayFifoUnderflowErr", 171 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 172 {"XGMI PCS ElasticFifoOverflowErr", 173 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 174 {"XGMI PCS DeskewErr", 175 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)}, 176 {"XGMI PCS DataStartupLimitErr", 177 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 178 {"XGMI PCS FCInitTimeoutErr", 179 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 180 {"XGMI PCS RecoveryTimeoutErr", 181 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 182 {"XGMI PCS ReadySerialTimeoutErr", 183 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 184 {"XGMI PCS ReadySerialAttemptErr", 185 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 186 {"XGMI PCS RecoveryAttemptErr", 187 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 188 {"XGMI PCS RecoveryRelockAttemptErr", 189 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 190 }; 191 192 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = { 193 {"WAFL PCS DataLossErr", 194 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)}, 195 {"WAFL PCS TrainingErr", 196 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)}, 197 {"WAFL PCS CRCErr", 198 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)}, 199 {"WAFL PCS BERExceededErr", 200 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)}, 201 {"WAFL PCS TxMetaDataErr", 202 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)}, 203 {"WAFL PCS ReplayBufParityErr", 204 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)}, 205 {"WAFL PCS DataParityErr", 206 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)}, 207 {"WAFL PCS ReplayFifoOverflowErr", 208 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 209 {"WAFL PCS ReplayFifoUnderflowErr", 210 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 211 {"WAFL PCS ElasticFifoOverflowErr", 212 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 213 {"WAFL PCS DeskewErr", 214 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)}, 215 {"WAFL PCS DataStartupLimitErr", 216 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)}, 217 {"WAFL PCS FCInitTimeoutErr", 218 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 219 {"WAFL PCS RecoveryTimeoutErr", 220 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 221 {"WAFL PCS ReadySerialTimeoutErr", 222 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 223 {"WAFL PCS ReadySerialAttemptErr", 224 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 225 {"WAFL PCS RecoveryAttemptErr", 226 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 227 {"WAFL PCS RecoveryRelockAttemptErr", 228 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 229 }; 230 231 static const struct amdgpu_pcs_ras_field xgmi3x16_pcs_ras_fields[] = { 232 {"XGMI3X16 PCS DataLossErr", 233 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataLossErr)}, 234 {"XGMI3X16 PCS TrainingErr", 235 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TrainingErr)}, 236 {"XGMI3X16 PCS FlowCtrlAckErr", 237 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FlowCtrlAckErr)}, 238 {"XGMI3X16 PCS RxFifoUnderflowErr", 239 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxFifoUnderflowErr)}, 240 {"XGMI3X16 PCS RxFifoOverflowErr", 241 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxFifoOverflowErr)}, 242 {"XGMI3X16 PCS CRCErr", 243 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, CRCErr)}, 244 {"XGMI3X16 PCS BERExceededErr", 245 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, BERExceededErr)}, 246 {"XGMI3X16 PCS TxVcidDataErr", 247 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TxVcidDataErr)}, 248 {"XGMI3X16 PCS ReplayBufParityErr", 249 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 250 {"XGMI3X16 PCS DataParityErr", 251 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataParityErr)}, 252 {"XGMI3X16 PCS ReplayFifoOverflowErr", 253 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 254 {"XGMI3X16 PCS ReplayFifoUnderflowErr", 255 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 256 {"XGMI3X16 PCS ElasticFifoOverflowErr", 257 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 258 {"XGMI3X16 PCS DeskewErr", 259 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DeskewErr)}, 260 {"XGMI3X16 PCS FlowCtrlCRCErr", 261 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FlowCtrlCRCErr)}, 262 {"XGMI3X16 PCS DataStartupLimitErr", 263 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 264 {"XGMI3X16 PCS FCInitTimeoutErr", 265 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 266 {"XGMI3X16 PCS RecoveryTimeoutErr", 267 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 268 {"XGMI3X16 PCS ReadySerialTimeoutErr", 269 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 270 {"XGMI3X16 PCS ReadySerialAttemptErr", 271 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 272 {"XGMI3X16 PCS RecoveryAttemptErr", 273 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 274 {"XGMI3X16 PCS RecoveryRelockAttemptErr", 275 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 276 {"XGMI3X16 PCS ReplayAttemptErr", 277 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayAttemptErr)}, 278 {"XGMI3X16 PCS SyncHdrErr", 279 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, SyncHdrErr)}, 280 {"XGMI3X16 PCS TxReplayTimeoutErr", 281 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TxReplayTimeoutErr)}, 282 {"XGMI3X16 PCS RxReplayTimeoutErr", 283 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxReplayTimeoutErr)}, 284 {"XGMI3X16 PCS LinkSubTxTimeoutErr", 285 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, LinkSubTxTimeoutErr)}, 286 {"XGMI3X16 PCS LinkSubRxTimeoutErr", 287 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, LinkSubRxTimeoutErr)}, 288 {"XGMI3X16 PCS RxCMDPktErr", 289 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxCMDPktErr)}, 290 }; 291 292 /** 293 * DOC: AMDGPU XGMI Support 294 * 295 * XGMI is a high speed interconnect that joins multiple GPU cards 296 * into a homogeneous memory space that is organized by a collective 297 * hive ID and individual node IDs, both of which are 64-bit numbers. 298 * 299 * The file xgmi_device_id contains the unique per GPU device ID and 300 * is stored in the /sys/class/drm/card${cardno}/device/ directory. 301 * 302 * Inside the device directory a sub-directory 'xgmi_hive_info' is 303 * created which contains the hive ID and the list of nodes. 304 * 305 * The hive ID is stored in: 306 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id 307 * 308 * The node information is stored in numbered directories: 309 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id 310 * 311 * Each device has their own xgmi_hive_info direction with a mirror 312 * set of node sub-directories. 313 * 314 * The XGMI memory space is built by contiguously adding the power of 315 * two padded VRAM space from each node to each other. 316 * 317 */ 318 319 static struct attribute amdgpu_xgmi_hive_id = { 320 .name = "xgmi_hive_id", 321 .mode = S_IRUGO 322 }; 323 324 static struct attribute *amdgpu_xgmi_hive_attrs[] = { 325 &amdgpu_xgmi_hive_id, 326 NULL 327 }; 328 ATTRIBUTE_GROUPS(amdgpu_xgmi_hive); 329 330 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj, 331 struct attribute *attr, char *buf) 332 { 333 struct amdgpu_hive_info *hive = container_of( 334 kobj, struct amdgpu_hive_info, kobj); 335 336 if (attr == &amdgpu_xgmi_hive_id) 337 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id); 338 339 return 0; 340 } 341 342 static void amdgpu_xgmi_hive_release(struct kobject *kobj) 343 { 344 struct amdgpu_hive_info *hive = container_of( 345 kobj, struct amdgpu_hive_info, kobj); 346 347 amdgpu_reset_put_reset_domain(hive->reset_domain); 348 hive->reset_domain = NULL; 349 350 mutex_destroy(&hive->hive_lock); 351 kfree(hive); 352 } 353 354 static const struct sysfs_ops amdgpu_xgmi_hive_ops = { 355 .show = amdgpu_xgmi_show_attrs, 356 }; 357 358 static const struct kobj_type amdgpu_xgmi_hive_type = { 359 .release = amdgpu_xgmi_hive_release, 360 .sysfs_ops = &amdgpu_xgmi_hive_ops, 361 .default_groups = amdgpu_xgmi_hive_groups, 362 }; 363 364 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev, 365 struct device_attribute *attr, 366 char *buf) 367 { 368 struct drm_device *ddev = dev_get_drvdata(dev); 369 struct amdgpu_device *adev = drm_to_adev(ddev); 370 371 return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id); 372 373 } 374 375 static ssize_t amdgpu_xgmi_show_physical_id(struct device *dev, 376 struct device_attribute *attr, 377 char *buf) 378 { 379 struct drm_device *ddev = dev_get_drvdata(dev); 380 struct amdgpu_device *adev = drm_to_adev(ddev); 381 382 return sysfs_emit(buf, "%u\n", adev->gmc.xgmi.physical_node_id); 383 384 } 385 386 static ssize_t amdgpu_xgmi_show_num_hops(struct device *dev, 387 struct device_attribute *attr, 388 char *buf) 389 { 390 struct drm_device *ddev = dev_get_drvdata(dev); 391 struct amdgpu_device *adev = drm_to_adev(ddev); 392 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 393 int i; 394 395 for (i = 0; i < top->num_nodes; i++) 396 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_hops); 397 398 return sysfs_emit(buf, "%s\n", buf); 399 } 400 401 static ssize_t amdgpu_xgmi_show_num_links(struct device *dev, 402 struct device_attribute *attr, 403 char *buf) 404 { 405 struct drm_device *ddev = dev_get_drvdata(dev); 406 struct amdgpu_device *adev = drm_to_adev(ddev); 407 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 408 int i; 409 410 for (i = 0; i < top->num_nodes; i++) 411 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_links); 412 413 return sysfs_emit(buf, "%s\n", buf); 414 } 415 416 static ssize_t amdgpu_xgmi_show_connected_port_num(struct device *dev, 417 struct device_attribute *attr, 418 char *buf) 419 { 420 struct drm_device *ddev = dev_get_drvdata(dev); 421 struct amdgpu_device *adev = drm_to_adev(ddev); 422 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 423 int i, j, size = 0; 424 int current_node; 425 /* 426 * get the node id in the sysfs for the current socket and show 427 * it in the port num info output in the sysfs for easy reading. 428 * it is NOT the one retrieved from xgmi ta. 429 */ 430 for (i = 0; i < top->num_nodes; i++) { 431 if (top->nodes[i].node_id == adev->gmc.xgmi.node_id) { 432 current_node = i; 433 break; 434 } 435 } 436 437 for (i = 0; i < top->num_nodes; i++) { 438 for (j = 0; j < top->nodes[i].num_links; j++) 439 /* node id in sysfs starts from 1 rather than 0 so +1 here */ 440 size += sysfs_emit_at(buf, size, "%02x:%02x -> %02x:%02x\n", current_node + 1, 441 top->nodes[i].port_num[j].src_xgmi_port_num, i + 1, 442 top->nodes[i].port_num[j].dst_xgmi_port_num); 443 } 444 445 return size; 446 } 447 448 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801) 449 static ssize_t amdgpu_xgmi_show_error(struct device *dev, 450 struct device_attribute *attr, 451 char *buf) 452 { 453 struct drm_device *ddev = dev_get_drvdata(dev); 454 struct amdgpu_device *adev = drm_to_adev(ddev); 455 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in; 456 uint64_t fica_out; 457 unsigned int error_count = 0; 458 459 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200); 460 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208); 461 462 if ((!adev->df.funcs) || 463 (!adev->df.funcs->get_fica) || 464 (!adev->df.funcs->set_fica)) 465 return -EINVAL; 466 467 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in); 468 if (fica_out != 0x1f) 469 pr_err("xGMI error counters not enabled!\n"); 470 471 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in); 472 473 if ((fica_out & 0xffff) == 2) 474 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63); 475 476 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0); 477 478 return sysfs_emit(buf, "%u\n", error_count); 479 } 480 481 482 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL); 483 static DEVICE_ATTR(xgmi_physical_id, 0444, amdgpu_xgmi_show_physical_id, NULL); 484 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL); 485 static DEVICE_ATTR(xgmi_num_hops, S_IRUGO, amdgpu_xgmi_show_num_hops, NULL); 486 static DEVICE_ATTR(xgmi_num_links, S_IRUGO, amdgpu_xgmi_show_num_links, NULL); 487 static DEVICE_ATTR(xgmi_port_num, S_IRUGO, amdgpu_xgmi_show_connected_port_num, NULL); 488 489 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev, 490 struct amdgpu_hive_info *hive) 491 { 492 int ret = 0; 493 char node[10] = { 0 }; 494 495 /* Create xgmi device id file */ 496 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id); 497 if (ret) { 498 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n"); 499 return ret; 500 } 501 502 ret = device_create_file(adev->dev, &dev_attr_xgmi_physical_id); 503 if (ret) { 504 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_physical_id\n"); 505 return ret; 506 } 507 508 /* Create xgmi error file */ 509 ret = device_create_file(adev->dev, &dev_attr_xgmi_error); 510 if (ret) 511 pr_err("failed to create xgmi_error\n"); 512 513 /* Create xgmi num hops file */ 514 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_hops); 515 if (ret) 516 pr_err("failed to create xgmi_num_hops\n"); 517 518 /* Create xgmi num links file */ 519 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_links); 520 if (ret) 521 pr_err("failed to create xgmi_num_links\n"); 522 523 /* Create xgmi port num file if supported */ 524 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) { 525 ret = device_create_file(adev->dev, &dev_attr_xgmi_port_num); 526 if (ret) 527 dev_err(adev->dev, "failed to create xgmi_port_num\n"); 528 } 529 530 /* Create sysfs link to hive info folder on the first device */ 531 if (hive->kobj.parent != (&adev->dev->kobj)) { 532 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj, 533 "xgmi_hive_info"); 534 if (ret) { 535 dev_err(adev->dev, "XGMI: Failed to create link to hive info"); 536 goto remove_file; 537 } 538 } 539 540 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 541 /* Create sysfs link form the hive folder to yourself */ 542 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node); 543 if (ret) { 544 dev_err(adev->dev, "XGMI: Failed to create link from hive info"); 545 goto remove_link; 546 } 547 548 goto success; 549 550 551 remove_link: 552 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique); 553 554 remove_file: 555 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 556 device_remove_file(adev->dev, &dev_attr_xgmi_physical_id); 557 device_remove_file(adev->dev, &dev_attr_xgmi_error); 558 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 559 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 560 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) 561 device_remove_file(adev->dev, &dev_attr_xgmi_port_num); 562 563 success: 564 return ret; 565 } 566 567 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev, 568 struct amdgpu_hive_info *hive) 569 { 570 char node[10]; 571 memset(node, 0, sizeof(node)); 572 573 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 574 device_remove_file(adev->dev, &dev_attr_xgmi_physical_id); 575 device_remove_file(adev->dev, &dev_attr_xgmi_error); 576 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 577 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 578 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) 579 device_remove_file(adev->dev, &dev_attr_xgmi_port_num); 580 581 if (hive->kobj.parent != (&adev->dev->kobj)) 582 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info"); 583 584 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 585 sysfs_remove_link(&hive->kobj, node); 586 587 } 588 589 590 591 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev) 592 { 593 struct amdgpu_hive_info *hive = NULL; 594 int ret; 595 596 if (!adev->gmc.xgmi.hive_id) 597 return NULL; 598 599 if (adev->hive) { 600 kobject_get(&adev->hive->kobj); 601 return adev->hive; 602 } 603 604 mutex_lock(&xgmi_mutex); 605 606 list_for_each_entry(hive, &xgmi_hive_list, node) { 607 if (hive->hive_id == adev->gmc.xgmi.hive_id) 608 goto pro_end; 609 } 610 611 hive = kzalloc(sizeof(*hive), GFP_KERNEL); 612 if (!hive) { 613 dev_err(adev->dev, "XGMI: allocation failed\n"); 614 ret = -ENOMEM; 615 hive = NULL; 616 goto pro_end; 617 } 618 619 /* initialize new hive if not exist */ 620 ret = kobject_init_and_add(&hive->kobj, 621 &amdgpu_xgmi_hive_type, 622 &adev->dev->kobj, 623 "%s", "xgmi_hive_info"); 624 if (ret) { 625 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n"); 626 kobject_put(&hive->kobj); 627 hive = NULL; 628 goto pro_end; 629 } 630 631 /** 632 * Only init hive->reset_domain for none SRIOV configuration. For SRIOV, 633 * Host driver decide how to reset the GPU either through FLR or chain reset. 634 * Guest side will get individual notifications from the host for the FLR 635 * if necessary. 636 */ 637 if (!amdgpu_sriov_vf(adev)) { 638 /** 639 * Avoid recreating reset domain when hive is reconstructed for the case 640 * of reset the devices in the XGMI hive during probe for passthrough GPU 641 * See https://www.spinics.net/lists/amd-gfx/msg58836.html 642 */ 643 if (adev->reset_domain->type != XGMI_HIVE) { 644 hive->reset_domain = 645 amdgpu_reset_create_reset_domain(XGMI_HIVE, "amdgpu-reset-hive"); 646 if (!hive->reset_domain) { 647 dev_err(adev->dev, "XGMI: failed initializing reset domain for xgmi hive\n"); 648 ret = -ENOMEM; 649 kobject_put(&hive->kobj); 650 hive = NULL; 651 goto pro_end; 652 } 653 } else { 654 amdgpu_reset_get_reset_domain(adev->reset_domain); 655 hive->reset_domain = adev->reset_domain; 656 } 657 } 658 659 hive->hive_id = adev->gmc.xgmi.hive_id; 660 INIT_LIST_HEAD(&hive->device_list); 661 INIT_LIST_HEAD(&hive->node); 662 mutex_init(&hive->hive_lock); 663 atomic_set(&hive->number_devices, 0); 664 task_barrier_init(&hive->tb); 665 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN; 666 hive->hi_req_gpu = NULL; 667 668 /* 669 * hive pstate on boot is high in vega20 so we have to go to low 670 * pstate on after boot. 671 */ 672 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE; 673 list_add_tail(&hive->node, &xgmi_hive_list); 674 675 pro_end: 676 if (hive) 677 kobject_get(&hive->kobj); 678 mutex_unlock(&xgmi_mutex); 679 return hive; 680 } 681 682 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive) 683 { 684 if (hive) 685 kobject_put(&hive->kobj); 686 } 687 688 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate) 689 { 690 int ret = 0; 691 struct amdgpu_hive_info *hive; 692 struct amdgpu_device *request_adev; 693 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20; 694 bool init_low; 695 696 hive = amdgpu_get_xgmi_hive(adev); 697 if (!hive) 698 return 0; 699 700 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev; 701 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN; 702 amdgpu_put_xgmi_hive(hive); 703 /* fw bug so temporarily disable pstate switching */ 704 return 0; 705 706 if (!hive || adev->asic_type != CHIP_VEGA20) 707 return 0; 708 709 mutex_lock(&hive->hive_lock); 710 711 if (is_hi_req) 712 hive->hi_req_count++; 713 else 714 hive->hi_req_count--; 715 716 /* 717 * Vega20 only needs single peer to request pstate high for the hive to 718 * go high but all peers must request pstate low for the hive to go low 719 */ 720 if (hive->pstate == pstate || 721 (!is_hi_req && hive->hi_req_count && !init_low)) 722 goto out; 723 724 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate); 725 726 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate); 727 if (ret) { 728 dev_err(request_adev->dev, 729 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d", 730 request_adev->gmc.xgmi.node_id, 731 request_adev->gmc.xgmi.hive_id, ret); 732 goto out; 733 } 734 735 if (init_low) 736 hive->pstate = hive->hi_req_count ? 737 hive->pstate : AMDGPU_XGMI_PSTATE_MIN; 738 else { 739 hive->pstate = pstate; 740 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ? 741 adev : NULL; 742 } 743 out: 744 mutex_unlock(&hive->hive_lock); 745 return ret; 746 } 747 748 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev) 749 { 750 int ret; 751 752 if (amdgpu_sriov_vf(adev)) 753 return 0; 754 755 /* Each psp need to set the latest topology */ 756 ret = psp_xgmi_set_topology_info(&adev->psp, 757 atomic_read(&hive->number_devices), 758 &adev->psp.xgmi_context.top_info); 759 if (ret) 760 dev_err(adev->dev, 761 "XGMI: Set topology failure on device %llx, hive %llx, ret %d", 762 adev->gmc.xgmi.node_id, 763 adev->gmc.xgmi.hive_id, ret); 764 765 return ret; 766 } 767 768 769 /* 770 * NOTE psp_xgmi_node_info.num_hops layout is as follows: 771 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved) 772 * num_hops[5:3] = reserved 773 * num_hops[2:0] = number of hops 774 */ 775 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev, 776 struct amdgpu_device *peer_adev) 777 { 778 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 779 uint8_t num_hops_mask = 0x7; 780 int i; 781 782 for (i = 0 ; i < top->num_nodes; ++i) 783 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 784 return top->nodes[i].num_hops & num_hops_mask; 785 return -EINVAL; 786 } 787 788 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev, 789 struct amdgpu_device *peer_adev) 790 { 791 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 792 int i; 793 794 for (i = 0 ; i < top->num_nodes; ++i) 795 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 796 return top->nodes[i].num_links; 797 return -EINVAL; 798 } 799 800 /* 801 * Devices that support extended data require the entire hive to initialize with 802 * the shared memory buffer flag set. 803 * 804 * Hive locks and conditions apply - see amdgpu_xgmi_add_device 805 */ 806 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive, 807 bool set_extended_data) 808 { 809 struct amdgpu_device *tmp_adev; 810 int ret; 811 812 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 813 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false); 814 if (ret) { 815 dev_err(tmp_adev->dev, 816 "XGMI: Failed to initialize xgmi session for data partition %i\n", 817 set_extended_data); 818 return ret; 819 } 820 821 } 822 823 return 0; 824 } 825 826 static void amdgpu_xgmi_fill_topology_info(struct amdgpu_device *adev, 827 struct amdgpu_device *peer_adev) 828 { 829 struct psp_xgmi_topology_info *top_info = &adev->psp.xgmi_context.top_info; 830 struct psp_xgmi_topology_info *peer_info = &peer_adev->psp.xgmi_context.top_info; 831 832 for (int i = 0; i < peer_info->num_nodes; i++) { 833 if (peer_info->nodes[i].node_id == adev->gmc.xgmi.node_id) { 834 for (int j = 0; j < top_info->num_nodes; j++) { 835 if (top_info->nodes[j].node_id == peer_adev->gmc.xgmi.node_id) { 836 peer_info->nodes[i].num_hops = top_info->nodes[j].num_hops; 837 peer_info->nodes[i].is_sharing_enabled = 838 top_info->nodes[j].is_sharing_enabled; 839 peer_info->nodes[i].num_links = 840 top_info->nodes[j].num_links; 841 return; 842 } 843 } 844 } 845 } 846 } 847 848 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 849 { 850 struct psp_xgmi_topology_info *top_info; 851 struct amdgpu_hive_info *hive; 852 struct amdgpu_xgmi *entry; 853 struct amdgpu_device *tmp_adev = NULL; 854 855 int count = 0, ret = 0; 856 857 if (!adev->gmc.xgmi.supported) 858 return 0; 859 860 if (!adev->gmc.xgmi.pending_reset && 861 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 862 ret = psp_xgmi_initialize(&adev->psp, false, true); 863 if (ret) { 864 dev_err(adev->dev, 865 "XGMI: Failed to initialize xgmi session\n"); 866 return ret; 867 } 868 869 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 870 if (ret) { 871 dev_err(adev->dev, 872 "XGMI: Failed to get hive id\n"); 873 return ret; 874 } 875 876 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 877 if (ret) { 878 dev_err(adev->dev, 879 "XGMI: Failed to get node id\n"); 880 return ret; 881 } 882 } else { 883 adev->gmc.xgmi.hive_id = 16; 884 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 885 } 886 887 hive = amdgpu_get_xgmi_hive(adev); 888 if (!hive) { 889 ret = -EINVAL; 890 dev_err(adev->dev, 891 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 892 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 893 goto exit; 894 } 895 mutex_lock(&hive->hive_lock); 896 897 top_info = &adev->psp.xgmi_context.top_info; 898 899 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 900 list_for_each_entry(entry, &hive->device_list, head) 901 top_info->nodes[count++].node_id = entry->node_id; 902 top_info->num_nodes = count; 903 atomic_set(&hive->number_devices, count); 904 905 task_barrier_add_task(&hive->tb); 906 907 if (!adev->gmc.xgmi.pending_reset && 908 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 909 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 910 /* update node list for other device in the hive */ 911 if (tmp_adev != adev) { 912 top_info = &tmp_adev->psp.xgmi_context.top_info; 913 top_info->nodes[count - 1].node_id = 914 adev->gmc.xgmi.node_id; 915 top_info->num_nodes = count; 916 } 917 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 918 if (ret) 919 goto exit_unlock; 920 } 921 922 if (amdgpu_sriov_vf(adev) && 923 adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) { 924 /* only get topology for VF being init if it can support full duplex */ 925 ret = psp_xgmi_get_topology_info(&adev->psp, count, 926 &adev->psp.xgmi_context.top_info, false); 927 if (ret) { 928 dev_err(adev->dev, 929 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 930 adev->gmc.xgmi.node_id, 931 adev->gmc.xgmi.hive_id, ret); 932 /* To do: continue with some node failed or disable the whole hive*/ 933 goto exit_unlock; 934 } 935 936 /* fill the topology info for peers instead of getting from PSP */ 937 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 938 amdgpu_xgmi_fill_topology_info(adev, tmp_adev); 939 } 940 } else { 941 /* get latest topology info for each device from psp */ 942 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 943 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 944 &tmp_adev->psp.xgmi_context.top_info, false); 945 if (ret) { 946 dev_err(tmp_adev->dev, 947 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 948 tmp_adev->gmc.xgmi.node_id, 949 tmp_adev->gmc.xgmi.hive_id, ret); 950 /* To do : continue with some node failed or disable the whole hive */ 951 goto exit_unlock; 952 } 953 } 954 } 955 956 /* get topology again for hives that support extended data */ 957 if (adev->psp.xgmi_context.supports_extended_data) { 958 959 /* initialize the hive to get extended data. */ 960 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true); 961 if (ret) 962 goto exit_unlock; 963 964 /* get the extended data. */ 965 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 966 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 967 &tmp_adev->psp.xgmi_context.top_info, true); 968 if (ret) { 969 dev_err(tmp_adev->dev, 970 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d", 971 tmp_adev->gmc.xgmi.node_id, 972 tmp_adev->gmc.xgmi.hive_id, ret); 973 goto exit_unlock; 974 } 975 } 976 977 /* initialize the hive to get non-extended data for the next round. */ 978 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false); 979 if (ret) 980 goto exit_unlock; 981 982 } 983 } 984 985 if (!ret && !adev->gmc.xgmi.pending_reset) 986 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 987 988 exit_unlock: 989 mutex_unlock(&hive->hive_lock); 990 exit: 991 if (!ret) { 992 adev->hive = hive; 993 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 994 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 995 } else { 996 amdgpu_put_xgmi_hive(hive); 997 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 998 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 999 ret); 1000 } 1001 1002 return ret; 1003 } 1004 1005 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 1006 { 1007 struct amdgpu_hive_info *hive = adev->hive; 1008 1009 if (!adev->gmc.xgmi.supported) 1010 return -EINVAL; 1011 1012 if (!hive) 1013 return -EINVAL; 1014 1015 mutex_lock(&hive->hive_lock); 1016 task_barrier_rem_task(&hive->tb); 1017 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 1018 if (hive->hi_req_gpu == adev) 1019 hive->hi_req_gpu = NULL; 1020 list_del(&adev->gmc.xgmi.head); 1021 mutex_unlock(&hive->hive_lock); 1022 1023 amdgpu_put_xgmi_hive(hive); 1024 adev->hive = NULL; 1025 1026 if (atomic_dec_return(&hive->number_devices) == 0) { 1027 /* Remove the hive from global hive list */ 1028 mutex_lock(&xgmi_mutex); 1029 list_del(&hive->node); 1030 mutex_unlock(&xgmi_mutex); 1031 1032 amdgpu_put_xgmi_hive(hive); 1033 } 1034 1035 return 0; 1036 } 1037 1038 static int xgmi_v6_4_0_aca_bank_generate_report(struct aca_handle *handle, struct aca_bank *bank, enum aca_error_type type, 1039 struct aca_bank_report *report, void *data) 1040 { 1041 struct amdgpu_device *adev = handle->adev; 1042 const char *error_str; 1043 u64 status; 1044 int ret, ext_error_code; 1045 1046 ret = aca_bank_info_decode(bank, &report->info); 1047 if (ret) 1048 return ret; 1049 1050 status = bank->regs[ACA_REG_IDX_STATUS]; 1051 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); 1052 1053 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1054 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1055 if (error_str) 1056 dev_info(adev->dev, "%s detected\n", error_str); 1057 1058 if ((type == ACA_ERROR_TYPE_UE && ext_error_code == 0) || 1059 (type == ACA_ERROR_TYPE_CE && ext_error_code == 6)) 1060 report->count[type] = ACA_REG__MISC0__ERRCNT(bank->regs[ACA_REG_IDX_MISC0]); 1061 1062 return 0; 1063 } 1064 1065 static const struct aca_bank_ops xgmi_v6_4_0_aca_bank_ops = { 1066 .aca_bank_generate_report = xgmi_v6_4_0_aca_bank_generate_report, 1067 }; 1068 1069 static const struct aca_info xgmi_v6_4_0_aca_info = { 1070 .hwip = ACA_HWIP_TYPE_PCS_XGMI, 1071 .mask = ACA_ERROR_UE_MASK | ACA_ERROR_CE_MASK, 1072 .bank_ops = &xgmi_v6_4_0_aca_bank_ops, 1073 }; 1074 1075 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 1076 { 1077 int r; 1078 1079 if (!adev->gmc.xgmi.supported || 1080 adev->gmc.xgmi.num_physical_nodes == 0) 1081 return 0; 1082 1083 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1084 1085 r = amdgpu_ras_block_late_init(adev, ras_block); 1086 if (r) 1087 return r; 1088 1089 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1090 case IP_VERSION(6, 4, 0): 1091 r = amdgpu_ras_bind_aca(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL, 1092 &xgmi_v6_4_0_aca_info, NULL); 1093 if (r) 1094 goto late_fini; 1095 break; 1096 default: 1097 break; 1098 } 1099 1100 return 0; 1101 1102 late_fini: 1103 amdgpu_ras_block_late_fini(adev, ras_block); 1104 1105 return r; 1106 } 1107 1108 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 1109 uint64_t addr) 1110 { 1111 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 1112 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 1113 } 1114 1115 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 1116 { 1117 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 1118 WREG32_PCIE(pcs_status_reg, 0); 1119 } 1120 1121 static void amdgpu_xgmi_legacy_reset_ras_error_count(struct amdgpu_device *adev) 1122 { 1123 uint32_t i; 1124 1125 switch (adev->asic_type) { 1126 case CHIP_ARCTURUS: 1127 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 1128 pcs_clear_status(adev, 1129 xgmi_pcs_err_status_reg_arct[i]); 1130 break; 1131 case CHIP_VEGA20: 1132 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 1133 pcs_clear_status(adev, 1134 xgmi_pcs_err_status_reg_vg20[i]); 1135 break; 1136 case CHIP_ALDEBARAN: 1137 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) 1138 pcs_clear_status(adev, 1139 xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1140 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) 1141 pcs_clear_status(adev, 1142 walf_pcs_err_status_reg_aldebaran[i]); 1143 break; 1144 default: 1145 break; 1146 } 1147 1148 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1149 case IP_VERSION(6, 4, 0): 1150 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) 1151 pcs_clear_status(adev, 1152 xgmi3x16_pcs_err_status_reg_v6_4[i]); 1153 break; 1154 default: 1155 break; 1156 } 1157 } 1158 1159 static void __xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst, u64 mca_base) 1160 { 1161 WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL); 1162 } 1163 1164 static void xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst) 1165 { 1166 int i; 1167 1168 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1169 __xgmi_v6_4_0_reset_error_count(adev, xgmi_inst, xgmi_v6_4_0_mca_base_array[i]); 1170 } 1171 1172 static void xgmi_v6_4_0_reset_ras_error_count(struct amdgpu_device *adev) 1173 { 1174 int i; 1175 1176 for_each_inst(i, adev->aid_mask) 1177 xgmi_v6_4_0_reset_error_count(adev, i); 1178 } 1179 1180 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 1181 { 1182 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1183 case IP_VERSION(6, 4, 0): 1184 xgmi_v6_4_0_reset_ras_error_count(adev); 1185 break; 1186 default: 1187 amdgpu_xgmi_legacy_reset_ras_error_count(adev); 1188 break; 1189 } 1190 } 1191 1192 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 1193 uint32_t value, 1194 uint32_t mask_value, 1195 uint32_t *ue_count, 1196 uint32_t *ce_count, 1197 bool is_xgmi_pcs, 1198 bool check_mask) 1199 { 1200 int i; 1201 int ue_cnt = 0; 1202 const struct amdgpu_pcs_ras_field *pcs_ras_fields = NULL; 1203 uint32_t field_array_size = 0; 1204 1205 if (is_xgmi_pcs) { 1206 if (amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1207 IP_VERSION(6, 1, 0) || 1208 amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1209 IP_VERSION(6, 4, 0)) { 1210 pcs_ras_fields = &xgmi3x16_pcs_ras_fields[0]; 1211 field_array_size = ARRAY_SIZE(xgmi3x16_pcs_ras_fields); 1212 } else { 1213 pcs_ras_fields = &xgmi_pcs_ras_fields[0]; 1214 field_array_size = ARRAY_SIZE(xgmi_pcs_ras_fields); 1215 } 1216 } else { 1217 pcs_ras_fields = &wafl_pcs_ras_fields[0]; 1218 field_array_size = ARRAY_SIZE(wafl_pcs_ras_fields); 1219 } 1220 1221 if (check_mask) 1222 value = value & ~mask_value; 1223 1224 /* query xgmi/walf pcs error status, 1225 * only ue is supported */ 1226 for (i = 0; value && i < field_array_size; i++) { 1227 ue_cnt = (value & 1228 pcs_ras_fields[i].pcs_err_mask) >> 1229 pcs_ras_fields[i].pcs_err_shift; 1230 if (ue_cnt) { 1231 dev_info(adev->dev, "%s detected\n", 1232 pcs_ras_fields[i].err_name); 1233 *ue_count += ue_cnt; 1234 } 1235 1236 /* reset bit value if the bit is checked */ 1237 value &= ~(pcs_ras_fields[i].pcs_err_mask); 1238 } 1239 1240 return 0; 1241 } 1242 1243 static void amdgpu_xgmi_legacy_query_ras_error_count(struct amdgpu_device *adev, 1244 void *ras_error_status) 1245 { 1246 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1247 int i, supported = 1; 1248 uint32_t data, mask_data = 0; 1249 uint32_t ue_cnt = 0, ce_cnt = 0; 1250 1251 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 1252 return ; 1253 1254 err_data->ue_count = 0; 1255 err_data->ce_count = 0; 1256 1257 switch (adev->asic_type) { 1258 case CHIP_ARCTURUS: 1259 /* check xgmi pcs error */ 1260 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 1261 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 1262 if (data) 1263 amdgpu_xgmi_query_pcs_error_status(adev, data, 1264 mask_data, &ue_cnt, &ce_cnt, true, false); 1265 } 1266 /* check wafl pcs error */ 1267 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 1268 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 1269 if (data) 1270 amdgpu_xgmi_query_pcs_error_status(adev, data, 1271 mask_data, &ue_cnt, &ce_cnt, false, false); 1272 } 1273 break; 1274 case CHIP_VEGA20: 1275 /* check xgmi pcs error */ 1276 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 1277 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 1278 if (data) 1279 amdgpu_xgmi_query_pcs_error_status(adev, data, 1280 mask_data, &ue_cnt, &ce_cnt, true, false); 1281 } 1282 /* check wafl pcs error */ 1283 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 1284 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 1285 if (data) 1286 amdgpu_xgmi_query_pcs_error_status(adev, data, 1287 mask_data, &ue_cnt, &ce_cnt, false, false); 1288 } 1289 break; 1290 case CHIP_ALDEBARAN: 1291 /* check xgmi3x16 pcs error */ 1292 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) { 1293 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1294 mask_data = 1295 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1296 if (data) 1297 amdgpu_xgmi_query_pcs_error_status(adev, data, 1298 mask_data, &ue_cnt, &ce_cnt, true, true); 1299 } 1300 /* check wafl pcs error */ 1301 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) { 1302 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]); 1303 mask_data = 1304 RREG32_PCIE(walf_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1305 if (data) 1306 amdgpu_xgmi_query_pcs_error_status(adev, data, 1307 mask_data, &ue_cnt, &ce_cnt, false, true); 1308 } 1309 break; 1310 default: 1311 supported = 0; 1312 break; 1313 } 1314 1315 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1316 case IP_VERSION(6, 4, 0): 1317 /* check xgmi3x16 pcs error */ 1318 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) { 1319 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_v6_4[i]); 1320 mask_data = 1321 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[i]); 1322 if (data) 1323 amdgpu_xgmi_query_pcs_error_status(adev, data, 1324 mask_data, &ue_cnt, &ce_cnt, true, true); 1325 } 1326 break; 1327 default: 1328 if (!supported) 1329 dev_warn(adev->dev, "XGMI RAS error query not supported"); 1330 break; 1331 } 1332 1333 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1334 1335 err_data->ue_count += ue_cnt; 1336 err_data->ce_count += ce_cnt; 1337 } 1338 1339 static enum aca_error_type xgmi_v6_4_0_pcs_mca_get_error_type(struct amdgpu_device *adev, u64 status) 1340 { 1341 const char *error_str; 1342 int ext_error_code; 1343 1344 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); 1345 1346 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1347 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1348 if (error_str) 1349 dev_info(adev->dev, "%s detected\n", error_str); 1350 1351 switch (ext_error_code) { 1352 case 0: 1353 return ACA_ERROR_TYPE_UE; 1354 case 6: 1355 return ACA_ERROR_TYPE_CE; 1356 default: 1357 return -EINVAL; 1358 } 1359 1360 return -EINVAL; 1361 } 1362 1363 static void __xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, struct amdgpu_smuio_mcm_config_info *mcm_info, 1364 u64 mca_base, struct ras_err_data *err_data) 1365 { 1366 int xgmi_inst = mcm_info->die_id; 1367 u64 status = 0; 1368 1369 status = RREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS); 1370 if (!ACA_REG__STATUS__VAL(status)) 1371 return; 1372 1373 switch (xgmi_v6_4_0_pcs_mca_get_error_type(adev, status)) { 1374 case ACA_ERROR_TYPE_UE: 1375 amdgpu_ras_error_statistic_ue_count(err_data, mcm_info, NULL, 1ULL); 1376 break; 1377 case ACA_ERROR_TYPE_CE: 1378 amdgpu_ras_error_statistic_ce_count(err_data, mcm_info, NULL, 1ULL); 1379 break; 1380 default: 1381 break; 1382 } 1383 1384 WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL); 1385 } 1386 1387 static void xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, int xgmi_inst, struct ras_err_data *err_data) 1388 { 1389 struct amdgpu_smuio_mcm_config_info mcm_info = { 1390 .socket_id = adev->smuio.funcs->get_socket_id(adev), 1391 .die_id = xgmi_inst, 1392 }; 1393 int i; 1394 1395 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1396 __xgmi_v6_4_0_query_error_count(adev, &mcm_info, xgmi_v6_4_0_mca_base_array[i], err_data); 1397 } 1398 1399 static void xgmi_v6_4_0_query_ras_error_count(struct amdgpu_device *adev, void *ras_error_status) 1400 { 1401 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1402 int i; 1403 1404 for_each_inst(i, adev->aid_mask) 1405 xgmi_v6_4_0_query_error_count(adev, i, err_data); 1406 } 1407 1408 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 1409 void *ras_error_status) 1410 { 1411 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1412 case IP_VERSION(6, 4, 0): 1413 xgmi_v6_4_0_query_ras_error_count(adev, ras_error_status); 1414 break; 1415 default: 1416 amdgpu_xgmi_legacy_query_ras_error_count(adev, ras_error_status); 1417 break; 1418 } 1419 } 1420 1421 /* Trigger XGMI/WAFL error */ 1422 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev, 1423 void *inject_if, uint32_t instance_mask) 1424 { 1425 int ret1, ret2; 1426 struct ta_ras_trigger_error_input *block_info = 1427 (struct ta_ras_trigger_error_input *)inject_if; 1428 1429 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) 1430 dev_warn(adev->dev, "Failed to disallow df cstate"); 1431 1432 ret1 = amdgpu_dpm_set_xgmi_plpd_mode(adev, XGMI_PLPD_DISALLOW); 1433 if (ret1 && ret1 != -EOPNOTSUPP) 1434 dev_warn(adev->dev, "Failed to disallow XGMI power down"); 1435 1436 ret2 = psp_ras_trigger_error(&adev->psp, block_info, instance_mask); 1437 1438 if (amdgpu_ras_intr_triggered()) 1439 return ret2; 1440 1441 ret1 = amdgpu_dpm_set_xgmi_plpd_mode(adev, XGMI_PLPD_DEFAULT); 1442 if (ret1 && ret1 != -EOPNOTSUPP) 1443 dev_warn(adev->dev, "Failed to allow XGMI power down"); 1444 1445 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW)) 1446 dev_warn(adev->dev, "Failed to allow df cstate"); 1447 1448 return ret2; 1449 } 1450 1451 struct amdgpu_ras_block_hw_ops xgmi_ras_hw_ops = { 1452 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 1453 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 1454 .ras_error_inject = amdgpu_ras_error_inject_xgmi, 1455 }; 1456 1457 struct amdgpu_xgmi_ras xgmi_ras = { 1458 .ras_block = { 1459 .hw_ops = &xgmi_ras_hw_ops, 1460 .ras_late_init = amdgpu_xgmi_ras_late_init, 1461 }, 1462 }; 1463 1464 int amdgpu_xgmi_ras_sw_init(struct amdgpu_device *adev) 1465 { 1466 int err; 1467 struct amdgpu_xgmi_ras *ras; 1468 1469 if (!adev->gmc.xgmi.ras) 1470 return 0; 1471 1472 ras = adev->gmc.xgmi.ras; 1473 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 1474 if (err) { 1475 dev_err(adev->dev, "Failed to register xgmi_wafl_pcs ras block!\n"); 1476 return err; 1477 } 1478 1479 strcpy(ras->ras_block.ras_comm.name, "xgmi_wafl"); 1480 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__XGMI_WAFL; 1481 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 1482 adev->gmc.xgmi.ras_if = &ras->ras_block.ras_comm; 1483 1484 return 0; 1485 } 1486