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 if (i == top->num_nodes) 438 return -EINVAL; 439 440 for (i = 0; i < top->num_nodes; i++) { 441 for (j = 0; j < top->nodes[i].num_links; j++) 442 /* node id in sysfs starts from 1 rather than 0 so +1 here */ 443 size += sysfs_emit_at(buf, size, "%02x:%02x -> %02x:%02x\n", current_node + 1, 444 top->nodes[i].port_num[j].src_xgmi_port_num, i + 1, 445 top->nodes[i].port_num[j].dst_xgmi_port_num); 446 } 447 448 return size; 449 } 450 451 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801) 452 static ssize_t amdgpu_xgmi_show_error(struct device *dev, 453 struct device_attribute *attr, 454 char *buf) 455 { 456 struct drm_device *ddev = dev_get_drvdata(dev); 457 struct amdgpu_device *adev = drm_to_adev(ddev); 458 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in; 459 uint64_t fica_out; 460 unsigned int error_count = 0; 461 462 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200); 463 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208); 464 465 if ((!adev->df.funcs) || 466 (!adev->df.funcs->get_fica) || 467 (!adev->df.funcs->set_fica)) 468 return -EINVAL; 469 470 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in); 471 if (fica_out != 0x1f) 472 pr_err("xGMI error counters not enabled!\n"); 473 474 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in); 475 476 if ((fica_out & 0xffff) == 2) 477 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63); 478 479 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0); 480 481 return sysfs_emit(buf, "%u\n", error_count); 482 } 483 484 485 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL); 486 static DEVICE_ATTR(xgmi_physical_id, 0444, amdgpu_xgmi_show_physical_id, NULL); 487 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL); 488 static DEVICE_ATTR(xgmi_num_hops, S_IRUGO, amdgpu_xgmi_show_num_hops, NULL); 489 static DEVICE_ATTR(xgmi_num_links, S_IRUGO, amdgpu_xgmi_show_num_links, NULL); 490 static DEVICE_ATTR(xgmi_port_num, S_IRUGO, amdgpu_xgmi_show_connected_port_num, NULL); 491 492 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev, 493 struct amdgpu_hive_info *hive) 494 { 495 int ret = 0; 496 char node[10] = { 0 }; 497 498 /* Create xgmi device id file */ 499 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id); 500 if (ret) { 501 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n"); 502 return ret; 503 } 504 505 ret = device_create_file(adev->dev, &dev_attr_xgmi_physical_id); 506 if (ret) { 507 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_physical_id\n"); 508 return ret; 509 } 510 511 /* Create xgmi error file */ 512 ret = device_create_file(adev->dev, &dev_attr_xgmi_error); 513 if (ret) 514 pr_err("failed to create xgmi_error\n"); 515 516 /* Create xgmi num hops file */ 517 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_hops); 518 if (ret) 519 pr_err("failed to create xgmi_num_hops\n"); 520 521 /* Create xgmi num links file */ 522 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_links); 523 if (ret) 524 pr_err("failed to create xgmi_num_links\n"); 525 526 /* Create xgmi port num file if supported */ 527 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) { 528 ret = device_create_file(adev->dev, &dev_attr_xgmi_port_num); 529 if (ret) 530 dev_err(adev->dev, "failed to create xgmi_port_num\n"); 531 } 532 533 /* Create sysfs link to hive info folder on the first device */ 534 if (hive->kobj.parent != (&adev->dev->kobj)) { 535 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj, 536 "xgmi_hive_info"); 537 if (ret) { 538 dev_err(adev->dev, "XGMI: Failed to create link to hive info"); 539 goto remove_file; 540 } 541 } 542 543 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 544 /* Create sysfs link form the hive folder to yourself */ 545 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node); 546 if (ret) { 547 dev_err(adev->dev, "XGMI: Failed to create link from hive info"); 548 goto remove_link; 549 } 550 551 goto success; 552 553 554 remove_link: 555 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique); 556 557 remove_file: 558 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 559 device_remove_file(adev->dev, &dev_attr_xgmi_physical_id); 560 device_remove_file(adev->dev, &dev_attr_xgmi_error); 561 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 562 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 563 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) 564 device_remove_file(adev->dev, &dev_attr_xgmi_port_num); 565 566 success: 567 return ret; 568 } 569 570 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev, 571 struct amdgpu_hive_info *hive) 572 { 573 char node[10]; 574 memset(node, 0, sizeof(node)); 575 576 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 577 device_remove_file(adev->dev, &dev_attr_xgmi_physical_id); 578 device_remove_file(adev->dev, &dev_attr_xgmi_error); 579 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 580 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 581 if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) 582 device_remove_file(adev->dev, &dev_attr_xgmi_port_num); 583 584 if (hive->kobj.parent != (&adev->dev->kobj)) 585 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info"); 586 587 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 588 sysfs_remove_link(&hive->kobj, node); 589 590 } 591 592 593 594 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev) 595 { 596 struct amdgpu_hive_info *hive = NULL; 597 int ret; 598 599 if (!adev->gmc.xgmi.hive_id) 600 return NULL; 601 602 if (adev->hive) { 603 kobject_get(&adev->hive->kobj); 604 return adev->hive; 605 } 606 607 mutex_lock(&xgmi_mutex); 608 609 list_for_each_entry(hive, &xgmi_hive_list, node) { 610 if (hive->hive_id == adev->gmc.xgmi.hive_id) 611 goto pro_end; 612 } 613 614 hive = kzalloc(sizeof(*hive), GFP_KERNEL); 615 if (!hive) { 616 dev_err(adev->dev, "XGMI: allocation failed\n"); 617 ret = -ENOMEM; 618 hive = NULL; 619 goto pro_end; 620 } 621 622 /* initialize new hive if not exist */ 623 ret = kobject_init_and_add(&hive->kobj, 624 &amdgpu_xgmi_hive_type, 625 &adev->dev->kobj, 626 "%s", "xgmi_hive_info"); 627 if (ret) { 628 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n"); 629 kobject_put(&hive->kobj); 630 hive = NULL; 631 goto pro_end; 632 } 633 634 /** 635 * Only init hive->reset_domain for none SRIOV configuration. For SRIOV, 636 * Host driver decide how to reset the GPU either through FLR or chain reset. 637 * Guest side will get individual notifications from the host for the FLR 638 * if necessary. 639 */ 640 if (!amdgpu_sriov_vf(adev)) { 641 /** 642 * Avoid recreating reset domain when hive is reconstructed for the case 643 * of reset the devices in the XGMI hive during probe for passthrough GPU 644 * See https://www.spinics.net/lists/amd-gfx/msg58836.html 645 */ 646 if (adev->reset_domain->type != XGMI_HIVE) { 647 hive->reset_domain = 648 amdgpu_reset_create_reset_domain(XGMI_HIVE, "amdgpu-reset-hive"); 649 if (!hive->reset_domain) { 650 dev_err(adev->dev, "XGMI: failed initializing reset domain for xgmi hive\n"); 651 ret = -ENOMEM; 652 kobject_put(&hive->kobj); 653 hive = NULL; 654 goto pro_end; 655 } 656 } else { 657 amdgpu_reset_get_reset_domain(adev->reset_domain); 658 hive->reset_domain = adev->reset_domain; 659 } 660 } 661 662 hive->hive_id = adev->gmc.xgmi.hive_id; 663 INIT_LIST_HEAD(&hive->device_list); 664 INIT_LIST_HEAD(&hive->node); 665 mutex_init(&hive->hive_lock); 666 atomic_set(&hive->number_devices, 0); 667 task_barrier_init(&hive->tb); 668 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN; 669 hive->hi_req_gpu = NULL; 670 671 /* 672 * hive pstate on boot is high in vega20 so we have to go to low 673 * pstate on after boot. 674 */ 675 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE; 676 list_add_tail(&hive->node, &xgmi_hive_list); 677 678 pro_end: 679 if (hive) 680 kobject_get(&hive->kobj); 681 mutex_unlock(&xgmi_mutex); 682 return hive; 683 } 684 685 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive) 686 { 687 if (hive) 688 kobject_put(&hive->kobj); 689 } 690 691 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate) 692 { 693 int ret = 0; 694 struct amdgpu_hive_info *hive; 695 struct amdgpu_device *request_adev; 696 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20; 697 bool init_low; 698 699 hive = amdgpu_get_xgmi_hive(adev); 700 if (!hive) 701 return 0; 702 703 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev; 704 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN; 705 amdgpu_put_xgmi_hive(hive); 706 /* fw bug so temporarily disable pstate switching */ 707 return 0; 708 709 if (!hive || adev->asic_type != CHIP_VEGA20) 710 return 0; 711 712 mutex_lock(&hive->hive_lock); 713 714 if (is_hi_req) 715 hive->hi_req_count++; 716 else 717 hive->hi_req_count--; 718 719 /* 720 * Vega20 only needs single peer to request pstate high for the hive to 721 * go high but all peers must request pstate low for the hive to go low 722 */ 723 if (hive->pstate == pstate || 724 (!is_hi_req && hive->hi_req_count && !init_low)) 725 goto out; 726 727 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate); 728 729 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate); 730 if (ret) { 731 dev_err(request_adev->dev, 732 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d", 733 request_adev->gmc.xgmi.node_id, 734 request_adev->gmc.xgmi.hive_id, ret); 735 goto out; 736 } 737 738 if (init_low) 739 hive->pstate = hive->hi_req_count ? 740 hive->pstate : AMDGPU_XGMI_PSTATE_MIN; 741 else { 742 hive->pstate = pstate; 743 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ? 744 adev : NULL; 745 } 746 out: 747 mutex_unlock(&hive->hive_lock); 748 return ret; 749 } 750 751 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev) 752 { 753 int ret; 754 755 if (amdgpu_sriov_vf(adev)) 756 return 0; 757 758 /* Each psp need to set the latest topology */ 759 ret = psp_xgmi_set_topology_info(&adev->psp, 760 atomic_read(&hive->number_devices), 761 &adev->psp.xgmi_context.top_info); 762 if (ret) 763 dev_err(adev->dev, 764 "XGMI: Set topology failure on device %llx, hive %llx, ret %d", 765 adev->gmc.xgmi.node_id, 766 adev->gmc.xgmi.hive_id, ret); 767 768 return ret; 769 } 770 771 772 /* 773 * NOTE psp_xgmi_node_info.num_hops layout is as follows: 774 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved) 775 * num_hops[5:3] = reserved 776 * num_hops[2:0] = number of hops 777 */ 778 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev, 779 struct amdgpu_device *peer_adev) 780 { 781 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 782 uint8_t num_hops_mask = 0x7; 783 int i; 784 785 for (i = 0 ; i < top->num_nodes; ++i) 786 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 787 return top->nodes[i].num_hops & num_hops_mask; 788 return -EINVAL; 789 } 790 791 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev, 792 struct amdgpu_device *peer_adev) 793 { 794 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 795 int i; 796 797 for (i = 0 ; i < top->num_nodes; ++i) 798 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 799 return top->nodes[i].num_links; 800 return -EINVAL; 801 } 802 803 /* 804 * Devices that support extended data require the entire hive to initialize with 805 * the shared memory buffer flag set. 806 * 807 * Hive locks and conditions apply - see amdgpu_xgmi_add_device 808 */ 809 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive, 810 bool set_extended_data) 811 { 812 struct amdgpu_device *tmp_adev; 813 int ret; 814 815 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 816 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false); 817 if (ret) { 818 dev_err(tmp_adev->dev, 819 "XGMI: Failed to initialize xgmi session for data partition %i\n", 820 set_extended_data); 821 return ret; 822 } 823 824 } 825 826 return 0; 827 } 828 829 static void amdgpu_xgmi_fill_topology_info(struct amdgpu_device *adev, 830 struct amdgpu_device *peer_adev) 831 { 832 struct psp_xgmi_topology_info *top_info = &adev->psp.xgmi_context.top_info; 833 struct psp_xgmi_topology_info *peer_info = &peer_adev->psp.xgmi_context.top_info; 834 835 for (int i = 0; i < peer_info->num_nodes; i++) { 836 if (peer_info->nodes[i].node_id == adev->gmc.xgmi.node_id) { 837 for (int j = 0; j < top_info->num_nodes; j++) { 838 if (top_info->nodes[j].node_id == peer_adev->gmc.xgmi.node_id) { 839 peer_info->nodes[i].num_hops = top_info->nodes[j].num_hops; 840 peer_info->nodes[i].is_sharing_enabled = 841 top_info->nodes[j].is_sharing_enabled; 842 peer_info->nodes[i].num_links = 843 top_info->nodes[j].num_links; 844 return; 845 } 846 } 847 } 848 } 849 } 850 851 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 852 { 853 struct psp_xgmi_topology_info *top_info; 854 struct amdgpu_hive_info *hive; 855 struct amdgpu_xgmi *entry; 856 struct amdgpu_device *tmp_adev = NULL; 857 858 int count = 0, ret = 0; 859 860 if (!adev->gmc.xgmi.supported) 861 return 0; 862 863 if (!adev->gmc.xgmi.pending_reset && 864 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 865 ret = psp_xgmi_initialize(&adev->psp, false, true); 866 if (ret) { 867 dev_err(adev->dev, 868 "XGMI: Failed to initialize xgmi session\n"); 869 return ret; 870 } 871 872 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 873 if (ret) { 874 dev_err(adev->dev, 875 "XGMI: Failed to get hive id\n"); 876 return ret; 877 } 878 879 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 880 if (ret) { 881 dev_err(adev->dev, 882 "XGMI: Failed to get node id\n"); 883 return ret; 884 } 885 } else { 886 adev->gmc.xgmi.hive_id = 16; 887 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 888 } 889 890 hive = amdgpu_get_xgmi_hive(adev); 891 if (!hive) { 892 ret = -EINVAL; 893 dev_err(adev->dev, 894 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 895 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 896 goto exit; 897 } 898 mutex_lock(&hive->hive_lock); 899 900 top_info = &adev->psp.xgmi_context.top_info; 901 902 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 903 list_for_each_entry(entry, &hive->device_list, head) 904 top_info->nodes[count++].node_id = entry->node_id; 905 top_info->num_nodes = count; 906 atomic_set(&hive->number_devices, count); 907 908 task_barrier_add_task(&hive->tb); 909 910 if (!adev->gmc.xgmi.pending_reset && 911 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 912 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 913 /* update node list for other device in the hive */ 914 if (tmp_adev != adev) { 915 top_info = &tmp_adev->psp.xgmi_context.top_info; 916 top_info->nodes[count - 1].node_id = 917 adev->gmc.xgmi.node_id; 918 top_info->num_nodes = count; 919 } 920 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 921 if (ret) 922 goto exit_unlock; 923 } 924 925 if (amdgpu_sriov_vf(adev) && 926 adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) { 927 /* only get topology for VF being init if it can support full duplex */ 928 ret = psp_xgmi_get_topology_info(&adev->psp, count, 929 &adev->psp.xgmi_context.top_info, false); 930 if (ret) { 931 dev_err(adev->dev, 932 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 933 adev->gmc.xgmi.node_id, 934 adev->gmc.xgmi.hive_id, ret); 935 /* To do: continue with some node failed or disable the whole hive*/ 936 goto exit_unlock; 937 } 938 939 /* fill the topology info for peers instead of getting from PSP */ 940 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 941 amdgpu_xgmi_fill_topology_info(adev, tmp_adev); 942 } 943 } else { 944 /* get latest topology info for each device from psp */ 945 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 946 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 947 &tmp_adev->psp.xgmi_context.top_info, false); 948 if (ret) { 949 dev_err(tmp_adev->dev, 950 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 951 tmp_adev->gmc.xgmi.node_id, 952 tmp_adev->gmc.xgmi.hive_id, ret); 953 /* To do : continue with some node failed or disable the whole hive */ 954 goto exit_unlock; 955 } 956 } 957 } 958 959 /* get topology again for hives that support extended data */ 960 if (adev->psp.xgmi_context.supports_extended_data) { 961 962 /* initialize the hive to get extended data. */ 963 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true); 964 if (ret) 965 goto exit_unlock; 966 967 /* get the extended data. */ 968 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 969 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 970 &tmp_adev->psp.xgmi_context.top_info, true); 971 if (ret) { 972 dev_err(tmp_adev->dev, 973 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d", 974 tmp_adev->gmc.xgmi.node_id, 975 tmp_adev->gmc.xgmi.hive_id, ret); 976 goto exit_unlock; 977 } 978 } 979 980 /* initialize the hive to get non-extended data for the next round. */ 981 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false); 982 if (ret) 983 goto exit_unlock; 984 985 } 986 } 987 988 if (!ret && !adev->gmc.xgmi.pending_reset) 989 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 990 991 exit_unlock: 992 mutex_unlock(&hive->hive_lock); 993 exit: 994 if (!ret) { 995 adev->hive = hive; 996 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 997 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 998 } else { 999 amdgpu_put_xgmi_hive(hive); 1000 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 1001 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 1002 ret); 1003 } 1004 1005 return ret; 1006 } 1007 1008 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 1009 { 1010 struct amdgpu_hive_info *hive = adev->hive; 1011 1012 if (!adev->gmc.xgmi.supported) 1013 return -EINVAL; 1014 1015 if (!hive) 1016 return -EINVAL; 1017 1018 mutex_lock(&hive->hive_lock); 1019 task_barrier_rem_task(&hive->tb); 1020 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 1021 if (hive->hi_req_gpu == adev) 1022 hive->hi_req_gpu = NULL; 1023 list_del(&adev->gmc.xgmi.head); 1024 mutex_unlock(&hive->hive_lock); 1025 1026 amdgpu_put_xgmi_hive(hive); 1027 adev->hive = NULL; 1028 1029 if (atomic_dec_return(&hive->number_devices) == 0) { 1030 /* Remove the hive from global hive list */ 1031 mutex_lock(&xgmi_mutex); 1032 list_del(&hive->node); 1033 mutex_unlock(&xgmi_mutex); 1034 1035 amdgpu_put_xgmi_hive(hive); 1036 } 1037 1038 return 0; 1039 } 1040 1041 static int xgmi_v6_4_0_aca_bank_parser(struct aca_handle *handle, struct aca_bank *bank, 1042 enum aca_smu_type type, void *data) 1043 { 1044 struct amdgpu_device *adev = handle->adev; 1045 struct aca_bank_info info; 1046 const char *error_str; 1047 u64 status, count; 1048 int ret, ext_error_code; 1049 1050 ret = aca_bank_info_decode(bank, &info); 1051 if (ret) 1052 return ret; 1053 1054 status = bank->regs[ACA_REG_IDX_STATUS]; 1055 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); 1056 1057 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1058 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1059 if (error_str) 1060 dev_info(adev->dev, "%s detected\n", error_str); 1061 1062 count = ACA_REG__MISC0__ERRCNT(bank->regs[ACA_REG_IDX_MISC0]); 1063 1064 switch (type) { 1065 case ACA_SMU_TYPE_UE: 1066 if (ext_error_code != 0 && ext_error_code != 9) 1067 count = 0ULL; 1068 1069 ret = aca_error_cache_log_bank_error(handle, &info, ACA_ERROR_TYPE_UE, count); 1070 break; 1071 case ACA_SMU_TYPE_CE: 1072 count = ext_error_code == 6 ? count : 0ULL; 1073 ret = aca_error_cache_log_bank_error(handle, &info, ACA_ERROR_TYPE_CE, count); 1074 break; 1075 default: 1076 return -EINVAL; 1077 } 1078 1079 return ret; 1080 } 1081 1082 static const struct aca_bank_ops xgmi_v6_4_0_aca_bank_ops = { 1083 .aca_bank_parser = xgmi_v6_4_0_aca_bank_parser, 1084 }; 1085 1086 static const struct aca_info xgmi_v6_4_0_aca_info = { 1087 .hwip = ACA_HWIP_TYPE_PCS_XGMI, 1088 .mask = ACA_ERROR_UE_MASK | ACA_ERROR_CE_MASK, 1089 .bank_ops = &xgmi_v6_4_0_aca_bank_ops, 1090 }; 1091 1092 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 1093 { 1094 int r; 1095 1096 if (!adev->gmc.xgmi.supported || 1097 adev->gmc.xgmi.num_physical_nodes == 0) 1098 return 0; 1099 1100 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1101 1102 r = amdgpu_ras_block_late_init(adev, ras_block); 1103 if (r) 1104 return r; 1105 1106 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1107 case IP_VERSION(6, 4, 0): 1108 r = amdgpu_ras_bind_aca(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL, 1109 &xgmi_v6_4_0_aca_info, NULL); 1110 if (r) 1111 goto late_fini; 1112 break; 1113 default: 1114 break; 1115 } 1116 1117 return 0; 1118 1119 late_fini: 1120 amdgpu_ras_block_late_fini(adev, ras_block); 1121 1122 return r; 1123 } 1124 1125 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 1126 uint64_t addr) 1127 { 1128 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 1129 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 1130 } 1131 1132 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 1133 { 1134 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 1135 WREG32_PCIE(pcs_status_reg, 0); 1136 } 1137 1138 static void amdgpu_xgmi_legacy_reset_ras_error_count(struct amdgpu_device *adev) 1139 { 1140 uint32_t i; 1141 1142 switch (adev->asic_type) { 1143 case CHIP_ARCTURUS: 1144 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 1145 pcs_clear_status(adev, 1146 xgmi_pcs_err_status_reg_arct[i]); 1147 break; 1148 case CHIP_VEGA20: 1149 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 1150 pcs_clear_status(adev, 1151 xgmi_pcs_err_status_reg_vg20[i]); 1152 break; 1153 case CHIP_ALDEBARAN: 1154 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) 1155 pcs_clear_status(adev, 1156 xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1157 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) 1158 pcs_clear_status(adev, 1159 walf_pcs_err_status_reg_aldebaran[i]); 1160 break; 1161 default: 1162 break; 1163 } 1164 1165 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1166 case IP_VERSION(6, 4, 0): 1167 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) 1168 pcs_clear_status(adev, 1169 xgmi3x16_pcs_err_status_reg_v6_4[i]); 1170 break; 1171 default: 1172 break; 1173 } 1174 } 1175 1176 static void __xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst, u64 mca_base) 1177 { 1178 WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL); 1179 } 1180 1181 static void xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst) 1182 { 1183 int i; 1184 1185 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1186 __xgmi_v6_4_0_reset_error_count(adev, xgmi_inst, xgmi_v6_4_0_mca_base_array[i]); 1187 } 1188 1189 static void xgmi_v6_4_0_reset_ras_error_count(struct amdgpu_device *adev) 1190 { 1191 int i; 1192 1193 for_each_inst(i, adev->aid_mask) 1194 xgmi_v6_4_0_reset_error_count(adev, i); 1195 } 1196 1197 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 1198 { 1199 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1200 case IP_VERSION(6, 4, 0): 1201 xgmi_v6_4_0_reset_ras_error_count(adev); 1202 break; 1203 default: 1204 amdgpu_xgmi_legacy_reset_ras_error_count(adev); 1205 break; 1206 } 1207 } 1208 1209 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 1210 uint32_t value, 1211 uint32_t mask_value, 1212 uint32_t *ue_count, 1213 uint32_t *ce_count, 1214 bool is_xgmi_pcs, 1215 bool check_mask) 1216 { 1217 int i; 1218 int ue_cnt = 0; 1219 const struct amdgpu_pcs_ras_field *pcs_ras_fields = NULL; 1220 uint32_t field_array_size = 0; 1221 1222 if (is_xgmi_pcs) { 1223 if (amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1224 IP_VERSION(6, 1, 0) || 1225 amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1226 IP_VERSION(6, 4, 0)) { 1227 pcs_ras_fields = &xgmi3x16_pcs_ras_fields[0]; 1228 field_array_size = ARRAY_SIZE(xgmi3x16_pcs_ras_fields); 1229 } else { 1230 pcs_ras_fields = &xgmi_pcs_ras_fields[0]; 1231 field_array_size = ARRAY_SIZE(xgmi_pcs_ras_fields); 1232 } 1233 } else { 1234 pcs_ras_fields = &wafl_pcs_ras_fields[0]; 1235 field_array_size = ARRAY_SIZE(wafl_pcs_ras_fields); 1236 } 1237 1238 if (check_mask) 1239 value = value & ~mask_value; 1240 1241 /* query xgmi/walf pcs error status, 1242 * only ue is supported */ 1243 for (i = 0; value && i < field_array_size; i++) { 1244 ue_cnt = (value & 1245 pcs_ras_fields[i].pcs_err_mask) >> 1246 pcs_ras_fields[i].pcs_err_shift; 1247 if (ue_cnt) { 1248 dev_info(adev->dev, "%s detected\n", 1249 pcs_ras_fields[i].err_name); 1250 *ue_count += ue_cnt; 1251 } 1252 1253 /* reset bit value if the bit is checked */ 1254 value &= ~(pcs_ras_fields[i].pcs_err_mask); 1255 } 1256 1257 return 0; 1258 } 1259 1260 static void amdgpu_xgmi_legacy_query_ras_error_count(struct amdgpu_device *adev, 1261 void *ras_error_status) 1262 { 1263 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1264 int i, supported = 1; 1265 uint32_t data, mask_data = 0; 1266 uint32_t ue_cnt = 0, ce_cnt = 0; 1267 1268 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 1269 return ; 1270 1271 err_data->ue_count = 0; 1272 err_data->ce_count = 0; 1273 1274 switch (adev->asic_type) { 1275 case CHIP_ARCTURUS: 1276 /* check xgmi pcs error */ 1277 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 1278 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 1279 if (data) 1280 amdgpu_xgmi_query_pcs_error_status(adev, data, 1281 mask_data, &ue_cnt, &ce_cnt, true, false); 1282 } 1283 /* check wafl pcs error */ 1284 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 1285 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 1286 if (data) 1287 amdgpu_xgmi_query_pcs_error_status(adev, data, 1288 mask_data, &ue_cnt, &ce_cnt, false, false); 1289 } 1290 break; 1291 case CHIP_VEGA20: 1292 /* check xgmi pcs error */ 1293 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 1294 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 1295 if (data) 1296 amdgpu_xgmi_query_pcs_error_status(adev, data, 1297 mask_data, &ue_cnt, &ce_cnt, true, false); 1298 } 1299 /* check wafl pcs error */ 1300 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 1301 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 1302 if (data) 1303 amdgpu_xgmi_query_pcs_error_status(adev, data, 1304 mask_data, &ue_cnt, &ce_cnt, false, false); 1305 } 1306 break; 1307 case CHIP_ALDEBARAN: 1308 /* check xgmi3x16 pcs error */ 1309 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) { 1310 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1311 mask_data = 1312 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1313 if (data) 1314 amdgpu_xgmi_query_pcs_error_status(adev, data, 1315 mask_data, &ue_cnt, &ce_cnt, true, true); 1316 } 1317 /* check wafl pcs error */ 1318 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) { 1319 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]); 1320 mask_data = 1321 RREG32_PCIE(walf_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1322 if (data) 1323 amdgpu_xgmi_query_pcs_error_status(adev, data, 1324 mask_data, &ue_cnt, &ce_cnt, false, true); 1325 } 1326 break; 1327 default: 1328 supported = 0; 1329 break; 1330 } 1331 1332 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1333 case IP_VERSION(6, 4, 0): 1334 /* check xgmi3x16 pcs error */ 1335 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) { 1336 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_v6_4[i]); 1337 mask_data = 1338 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[i]); 1339 if (data) 1340 amdgpu_xgmi_query_pcs_error_status(adev, data, 1341 mask_data, &ue_cnt, &ce_cnt, true, true); 1342 } 1343 break; 1344 default: 1345 if (!supported) 1346 dev_warn(adev->dev, "XGMI RAS error query not supported"); 1347 break; 1348 } 1349 1350 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1351 1352 err_data->ue_count += ue_cnt; 1353 err_data->ce_count += ce_cnt; 1354 } 1355 1356 static enum aca_error_type xgmi_v6_4_0_pcs_mca_get_error_type(struct amdgpu_device *adev, u64 status) 1357 { 1358 const char *error_str; 1359 int ext_error_code; 1360 1361 ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); 1362 1363 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1364 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1365 if (error_str) 1366 dev_info(adev->dev, "%s detected\n", error_str); 1367 1368 switch (ext_error_code) { 1369 case 0: 1370 return ACA_ERROR_TYPE_UE; 1371 case 6: 1372 return ACA_ERROR_TYPE_CE; 1373 default: 1374 return -EINVAL; 1375 } 1376 1377 return -EINVAL; 1378 } 1379 1380 static void __xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, struct amdgpu_smuio_mcm_config_info *mcm_info, 1381 u64 mca_base, struct ras_err_data *err_data) 1382 { 1383 int xgmi_inst = mcm_info->die_id; 1384 u64 status = 0; 1385 1386 status = RREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS); 1387 if (!ACA_REG__STATUS__VAL(status)) 1388 return; 1389 1390 switch (xgmi_v6_4_0_pcs_mca_get_error_type(adev, status)) { 1391 case ACA_ERROR_TYPE_UE: 1392 amdgpu_ras_error_statistic_ue_count(err_data, mcm_info, NULL, 1ULL); 1393 break; 1394 case ACA_ERROR_TYPE_CE: 1395 amdgpu_ras_error_statistic_ce_count(err_data, mcm_info, NULL, 1ULL); 1396 break; 1397 default: 1398 break; 1399 } 1400 1401 WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL); 1402 } 1403 1404 static void xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, int xgmi_inst, struct ras_err_data *err_data) 1405 { 1406 struct amdgpu_smuio_mcm_config_info mcm_info = { 1407 .socket_id = adev->smuio.funcs->get_socket_id(adev), 1408 .die_id = xgmi_inst, 1409 }; 1410 int i; 1411 1412 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1413 __xgmi_v6_4_0_query_error_count(adev, &mcm_info, xgmi_v6_4_0_mca_base_array[i], err_data); 1414 } 1415 1416 static void xgmi_v6_4_0_query_ras_error_count(struct amdgpu_device *adev, void *ras_error_status) 1417 { 1418 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1419 int i; 1420 1421 for_each_inst(i, adev->aid_mask) 1422 xgmi_v6_4_0_query_error_count(adev, i, err_data); 1423 } 1424 1425 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 1426 void *ras_error_status) 1427 { 1428 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1429 case IP_VERSION(6, 4, 0): 1430 xgmi_v6_4_0_query_ras_error_count(adev, ras_error_status); 1431 break; 1432 default: 1433 amdgpu_xgmi_legacy_query_ras_error_count(adev, ras_error_status); 1434 break; 1435 } 1436 } 1437 1438 /* Trigger XGMI/WAFL error */ 1439 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev, 1440 void *inject_if, uint32_t instance_mask) 1441 { 1442 int ret1, ret2; 1443 struct ta_ras_trigger_error_input *block_info = 1444 (struct ta_ras_trigger_error_input *)inject_if; 1445 1446 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) 1447 dev_warn(adev->dev, "Failed to disallow df cstate"); 1448 1449 ret1 = amdgpu_dpm_set_xgmi_plpd_mode(adev, XGMI_PLPD_DISALLOW); 1450 if (ret1 && ret1 != -EOPNOTSUPP) 1451 dev_warn(adev->dev, "Failed to disallow XGMI power down"); 1452 1453 ret2 = psp_ras_trigger_error(&adev->psp, block_info, instance_mask); 1454 1455 if (amdgpu_ras_intr_triggered()) 1456 return ret2; 1457 1458 ret1 = amdgpu_dpm_set_xgmi_plpd_mode(adev, XGMI_PLPD_DEFAULT); 1459 if (ret1 && ret1 != -EOPNOTSUPP) 1460 dev_warn(adev->dev, "Failed to allow XGMI power down"); 1461 1462 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW)) 1463 dev_warn(adev->dev, "Failed to allow df cstate"); 1464 1465 return ret2; 1466 } 1467 1468 struct amdgpu_ras_block_hw_ops xgmi_ras_hw_ops = { 1469 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 1470 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 1471 .ras_error_inject = amdgpu_ras_error_inject_xgmi, 1472 }; 1473 1474 struct amdgpu_xgmi_ras xgmi_ras = { 1475 .ras_block = { 1476 .hw_ops = &xgmi_ras_hw_ops, 1477 .ras_late_init = amdgpu_xgmi_ras_late_init, 1478 }, 1479 }; 1480 1481 int amdgpu_xgmi_ras_sw_init(struct amdgpu_device *adev) 1482 { 1483 int err; 1484 struct amdgpu_xgmi_ras *ras; 1485 1486 if (!adev->gmc.xgmi.ras) 1487 return 0; 1488 1489 ras = adev->gmc.xgmi.ras; 1490 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 1491 if (err) { 1492 dev_err(adev->dev, "Failed to register xgmi_wafl_pcs ras block!\n"); 1493 return err; 1494 } 1495 1496 strcpy(ras->ras_block.ras_comm.name, "xgmi_wafl"); 1497 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__XGMI_WAFL; 1498 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 1499 adev->gmc.xgmi.ras_if = &ras->ras_block.ras_comm; 1500 1501 return 0; 1502 } 1503