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 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 827 { 828 struct psp_xgmi_topology_info *top_info; 829 struct amdgpu_hive_info *hive; 830 struct amdgpu_xgmi *entry; 831 struct amdgpu_device *tmp_adev = NULL; 832 833 int count = 0, ret = 0; 834 835 if (!adev->gmc.xgmi.supported) 836 return 0; 837 838 if (!adev->gmc.xgmi.pending_reset && 839 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 840 ret = psp_xgmi_initialize(&adev->psp, false, true); 841 if (ret) { 842 dev_err(adev->dev, 843 "XGMI: Failed to initialize xgmi session\n"); 844 return ret; 845 } 846 847 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 848 if (ret) { 849 dev_err(adev->dev, 850 "XGMI: Failed to get hive id\n"); 851 return ret; 852 } 853 854 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 855 if (ret) { 856 dev_err(adev->dev, 857 "XGMI: Failed to get node id\n"); 858 return ret; 859 } 860 } else { 861 adev->gmc.xgmi.hive_id = 16; 862 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 863 } 864 865 hive = amdgpu_get_xgmi_hive(adev); 866 if (!hive) { 867 ret = -EINVAL; 868 dev_err(adev->dev, 869 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 870 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 871 goto exit; 872 } 873 mutex_lock(&hive->hive_lock); 874 875 top_info = &adev->psp.xgmi_context.top_info; 876 877 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 878 list_for_each_entry(entry, &hive->device_list, head) 879 top_info->nodes[count++].node_id = entry->node_id; 880 top_info->num_nodes = count; 881 atomic_set(&hive->number_devices, count); 882 883 task_barrier_add_task(&hive->tb); 884 885 if (!adev->gmc.xgmi.pending_reset && 886 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 887 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 888 /* update node list for other device in the hive */ 889 if (tmp_adev != adev) { 890 top_info = &tmp_adev->psp.xgmi_context.top_info; 891 top_info->nodes[count - 1].node_id = 892 adev->gmc.xgmi.node_id; 893 top_info->num_nodes = count; 894 } 895 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 896 if (ret) 897 goto exit_unlock; 898 } 899 900 /* get latest topology info for each device from psp */ 901 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 902 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 903 &tmp_adev->psp.xgmi_context.top_info, false); 904 if (ret) { 905 dev_err(tmp_adev->dev, 906 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 907 tmp_adev->gmc.xgmi.node_id, 908 tmp_adev->gmc.xgmi.hive_id, ret); 909 /* To do : continue with some node failed or disable the whole hive */ 910 goto exit_unlock; 911 } 912 } 913 914 /* get topology again for hives that support extended data */ 915 if (adev->psp.xgmi_context.supports_extended_data) { 916 917 /* initialize the hive to get extended data. */ 918 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true); 919 if (ret) 920 goto exit_unlock; 921 922 /* get the extended data. */ 923 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 924 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 925 &tmp_adev->psp.xgmi_context.top_info, true); 926 if (ret) { 927 dev_err(tmp_adev->dev, 928 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d", 929 tmp_adev->gmc.xgmi.node_id, 930 tmp_adev->gmc.xgmi.hive_id, ret); 931 goto exit_unlock; 932 } 933 } 934 935 /* initialize the hive to get non-extended data for the next round. */ 936 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false); 937 if (ret) 938 goto exit_unlock; 939 940 } 941 } 942 943 if (!ret && !adev->gmc.xgmi.pending_reset) 944 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 945 946 exit_unlock: 947 mutex_unlock(&hive->hive_lock); 948 exit: 949 if (!ret) { 950 adev->hive = hive; 951 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 952 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 953 } else { 954 amdgpu_put_xgmi_hive(hive); 955 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 956 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 957 ret); 958 } 959 960 return ret; 961 } 962 963 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 964 { 965 struct amdgpu_hive_info *hive = adev->hive; 966 967 if (!adev->gmc.xgmi.supported) 968 return -EINVAL; 969 970 if (!hive) 971 return -EINVAL; 972 973 mutex_lock(&hive->hive_lock); 974 task_barrier_rem_task(&hive->tb); 975 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 976 if (hive->hi_req_gpu == adev) 977 hive->hi_req_gpu = NULL; 978 list_del(&adev->gmc.xgmi.head); 979 mutex_unlock(&hive->hive_lock); 980 981 amdgpu_put_xgmi_hive(hive); 982 adev->hive = NULL; 983 984 if (atomic_dec_return(&hive->number_devices) == 0) { 985 /* Remove the hive from global hive list */ 986 mutex_lock(&xgmi_mutex); 987 list_del(&hive->node); 988 mutex_unlock(&xgmi_mutex); 989 990 amdgpu_put_xgmi_hive(hive); 991 } 992 993 return 0; 994 } 995 996 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 997 { 998 if (!adev->gmc.xgmi.supported || 999 adev->gmc.xgmi.num_physical_nodes == 0) 1000 return 0; 1001 1002 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1003 1004 return amdgpu_ras_block_late_init(adev, ras_block); 1005 } 1006 1007 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 1008 uint64_t addr) 1009 { 1010 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 1011 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 1012 } 1013 1014 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 1015 { 1016 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 1017 WREG32_PCIE(pcs_status_reg, 0); 1018 } 1019 1020 static void amdgpu_xgmi_legacy_reset_ras_error_count(struct amdgpu_device *adev) 1021 { 1022 uint32_t i; 1023 1024 switch (adev->asic_type) { 1025 case CHIP_ARCTURUS: 1026 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 1027 pcs_clear_status(adev, 1028 xgmi_pcs_err_status_reg_arct[i]); 1029 break; 1030 case CHIP_VEGA20: 1031 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 1032 pcs_clear_status(adev, 1033 xgmi_pcs_err_status_reg_vg20[i]); 1034 break; 1035 case CHIP_ALDEBARAN: 1036 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) 1037 pcs_clear_status(adev, 1038 xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1039 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) 1040 pcs_clear_status(adev, 1041 walf_pcs_err_status_reg_aldebaran[i]); 1042 break; 1043 default: 1044 break; 1045 } 1046 1047 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1048 case IP_VERSION(6, 4, 0): 1049 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) 1050 pcs_clear_status(adev, 1051 xgmi3x16_pcs_err_status_reg_v6_4[i]); 1052 break; 1053 default: 1054 break; 1055 } 1056 } 1057 1058 static void __xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst, u64 mca_base) 1059 { 1060 WREG64_MCA(xgmi_inst, mca_base, MCA_REG_IDX_STATUS, 0ULL); 1061 } 1062 1063 static void xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst) 1064 { 1065 int i; 1066 1067 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1068 __xgmi_v6_4_0_reset_error_count(adev, xgmi_inst, xgmi_v6_4_0_mca_base_array[i]); 1069 } 1070 1071 static void xgmi_v6_4_0_reset_ras_error_count(struct amdgpu_device *adev) 1072 { 1073 int i; 1074 1075 for_each_inst(i, adev->aid_mask) 1076 xgmi_v6_4_0_reset_error_count(adev, i); 1077 } 1078 1079 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 1080 { 1081 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1082 case IP_VERSION(6, 4, 0): 1083 xgmi_v6_4_0_reset_ras_error_count(adev); 1084 break; 1085 default: 1086 amdgpu_xgmi_legacy_reset_ras_error_count(adev); 1087 break; 1088 } 1089 } 1090 1091 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 1092 uint32_t value, 1093 uint32_t mask_value, 1094 uint32_t *ue_count, 1095 uint32_t *ce_count, 1096 bool is_xgmi_pcs, 1097 bool check_mask) 1098 { 1099 int i; 1100 int ue_cnt = 0; 1101 const struct amdgpu_pcs_ras_field *pcs_ras_fields = NULL; 1102 uint32_t field_array_size = 0; 1103 1104 if (is_xgmi_pcs) { 1105 if (amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1106 IP_VERSION(6, 1, 0) || 1107 amdgpu_ip_version(adev, XGMI_HWIP, 0) == 1108 IP_VERSION(6, 4, 0)) { 1109 pcs_ras_fields = &xgmi3x16_pcs_ras_fields[0]; 1110 field_array_size = ARRAY_SIZE(xgmi3x16_pcs_ras_fields); 1111 } else { 1112 pcs_ras_fields = &xgmi_pcs_ras_fields[0]; 1113 field_array_size = ARRAY_SIZE(xgmi_pcs_ras_fields); 1114 } 1115 } else { 1116 pcs_ras_fields = &wafl_pcs_ras_fields[0]; 1117 field_array_size = ARRAY_SIZE(wafl_pcs_ras_fields); 1118 } 1119 1120 if (check_mask) 1121 value = value & ~mask_value; 1122 1123 /* query xgmi/walf pcs error status, 1124 * only ue is supported */ 1125 for (i = 0; value && i < field_array_size; i++) { 1126 ue_cnt = (value & 1127 pcs_ras_fields[i].pcs_err_mask) >> 1128 pcs_ras_fields[i].pcs_err_shift; 1129 if (ue_cnt) { 1130 dev_info(adev->dev, "%s detected\n", 1131 pcs_ras_fields[i].err_name); 1132 *ue_count += ue_cnt; 1133 } 1134 1135 /* reset bit value if the bit is checked */ 1136 value &= ~(pcs_ras_fields[i].pcs_err_mask); 1137 } 1138 1139 return 0; 1140 } 1141 1142 static void amdgpu_xgmi_legacy_query_ras_error_count(struct amdgpu_device *adev, 1143 void *ras_error_status) 1144 { 1145 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1146 int i, supported = 1; 1147 uint32_t data, mask_data = 0; 1148 uint32_t ue_cnt = 0, ce_cnt = 0; 1149 1150 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 1151 return ; 1152 1153 err_data->ue_count = 0; 1154 err_data->ce_count = 0; 1155 1156 switch (adev->asic_type) { 1157 case CHIP_ARCTURUS: 1158 /* check xgmi pcs error */ 1159 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 1160 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 1161 if (data) 1162 amdgpu_xgmi_query_pcs_error_status(adev, data, 1163 mask_data, &ue_cnt, &ce_cnt, true, false); 1164 } 1165 /* check wafl pcs error */ 1166 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 1167 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 1168 if (data) 1169 amdgpu_xgmi_query_pcs_error_status(adev, data, 1170 mask_data, &ue_cnt, &ce_cnt, false, false); 1171 } 1172 break; 1173 case CHIP_VEGA20: 1174 /* check xgmi pcs error */ 1175 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 1176 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 1177 if (data) 1178 amdgpu_xgmi_query_pcs_error_status(adev, data, 1179 mask_data, &ue_cnt, &ce_cnt, true, false); 1180 } 1181 /* check wafl pcs error */ 1182 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 1183 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 1184 if (data) 1185 amdgpu_xgmi_query_pcs_error_status(adev, data, 1186 mask_data, &ue_cnt, &ce_cnt, false, false); 1187 } 1188 break; 1189 case CHIP_ALDEBARAN: 1190 /* check xgmi3x16 pcs error */ 1191 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) { 1192 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1193 mask_data = 1194 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1195 if (data) 1196 amdgpu_xgmi_query_pcs_error_status(adev, data, 1197 mask_data, &ue_cnt, &ce_cnt, true, true); 1198 } 1199 /* check wafl pcs error */ 1200 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) { 1201 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]); 1202 mask_data = 1203 RREG32_PCIE(walf_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1204 if (data) 1205 amdgpu_xgmi_query_pcs_error_status(adev, data, 1206 mask_data, &ue_cnt, &ce_cnt, false, true); 1207 } 1208 break; 1209 default: 1210 supported = 0; 1211 break; 1212 } 1213 1214 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1215 case IP_VERSION(6, 4, 0): 1216 /* check xgmi3x16 pcs error */ 1217 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) { 1218 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_v6_4[i]); 1219 mask_data = 1220 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[i]); 1221 if (data) 1222 amdgpu_xgmi_query_pcs_error_status(adev, data, 1223 mask_data, &ue_cnt, &ce_cnt, true, true); 1224 } 1225 break; 1226 default: 1227 if (!supported) 1228 dev_warn(adev->dev, "XGMI RAS error query not supported"); 1229 break; 1230 } 1231 1232 amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL); 1233 1234 err_data->ue_count += ue_cnt; 1235 err_data->ce_count += ce_cnt; 1236 } 1237 1238 static enum amdgpu_mca_error_type xgmi_v6_4_0_pcs_mca_get_error_type(struct amdgpu_device *adev, u64 status) 1239 { 1240 const char *error_str; 1241 int ext_error_code; 1242 1243 ext_error_code = MCA_REG__STATUS__ERRORCODEEXT(status); 1244 1245 error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ? 1246 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL; 1247 if (error_str) 1248 dev_info(adev->dev, "%s detected\n", error_str); 1249 1250 switch (ext_error_code) { 1251 case 0: 1252 return AMDGPU_MCA_ERROR_TYPE_UE; 1253 case 6: 1254 return AMDGPU_MCA_ERROR_TYPE_CE; 1255 default: 1256 return -EINVAL; 1257 } 1258 1259 return -EINVAL; 1260 } 1261 1262 static void __xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, struct amdgpu_smuio_mcm_config_info *mcm_info, 1263 u64 mca_base, struct ras_err_data *err_data) 1264 { 1265 int xgmi_inst = mcm_info->die_id; 1266 u64 status = 0; 1267 1268 status = RREG64_MCA(xgmi_inst, mca_base, MCA_REG_IDX_STATUS); 1269 if (!MCA_REG__STATUS__VAL(status)) 1270 return; 1271 1272 switch (xgmi_v6_4_0_pcs_mca_get_error_type(adev, status)) { 1273 case AMDGPU_MCA_ERROR_TYPE_UE: 1274 amdgpu_ras_error_statistic_ue_count(err_data, mcm_info, 1ULL); 1275 break; 1276 case AMDGPU_MCA_ERROR_TYPE_CE: 1277 amdgpu_ras_error_statistic_ce_count(err_data, mcm_info, 1ULL); 1278 break; 1279 default: 1280 break; 1281 } 1282 1283 WREG64_MCA(xgmi_inst, mca_base, MCA_REG_IDX_STATUS, 0ULL); 1284 } 1285 1286 static void xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, int xgmi_inst, struct ras_err_data *err_data) 1287 { 1288 struct amdgpu_smuio_mcm_config_info mcm_info = { 1289 .socket_id = adev->smuio.funcs->get_socket_id(adev), 1290 .die_id = xgmi_inst, 1291 }; 1292 int i; 1293 1294 for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++) 1295 __xgmi_v6_4_0_query_error_count(adev, &mcm_info, xgmi_v6_4_0_mca_base_array[i], err_data); 1296 } 1297 1298 static void xgmi_v6_4_0_query_ras_error_count(struct amdgpu_device *adev, void *ras_error_status) 1299 { 1300 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 1301 int i; 1302 1303 for_each_inst(i, adev->aid_mask) 1304 xgmi_v6_4_0_query_error_count(adev, i, err_data); 1305 } 1306 1307 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 1308 void *ras_error_status) 1309 { 1310 switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) { 1311 case IP_VERSION(6, 4, 0): 1312 xgmi_v6_4_0_query_ras_error_count(adev, ras_error_status); 1313 break; 1314 default: 1315 amdgpu_xgmi_legacy_query_ras_error_count(adev, ras_error_status); 1316 break; 1317 } 1318 } 1319 1320 /* Trigger XGMI/WAFL error */ 1321 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev, 1322 void *inject_if, uint32_t instance_mask) 1323 { 1324 int ret1, ret2; 1325 struct ta_ras_trigger_error_input *block_info = 1326 (struct ta_ras_trigger_error_input *)inject_if; 1327 1328 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) 1329 dev_warn(adev->dev, "Failed to disallow df cstate"); 1330 1331 ret1 = amdgpu_dpm_set_xgmi_plpd_mode(adev, XGMI_PLPD_DISALLOW); 1332 if (ret1 && ret1 != -EOPNOTSUPP) 1333 dev_warn(adev->dev, "Failed to disallow XGMI power down"); 1334 1335 ret2 = psp_ras_trigger_error(&adev->psp, block_info, instance_mask); 1336 1337 if (amdgpu_ras_intr_triggered()) 1338 return ret2; 1339 1340 ret1 = amdgpu_dpm_set_xgmi_plpd_mode(adev, XGMI_PLPD_DEFAULT); 1341 if (ret1 && ret1 != -EOPNOTSUPP) 1342 dev_warn(adev->dev, "Failed to allow XGMI power down"); 1343 1344 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW)) 1345 dev_warn(adev->dev, "Failed to allow df cstate"); 1346 1347 return ret2; 1348 } 1349 1350 struct amdgpu_ras_block_hw_ops xgmi_ras_hw_ops = { 1351 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 1352 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 1353 .ras_error_inject = amdgpu_ras_error_inject_xgmi, 1354 }; 1355 1356 struct amdgpu_xgmi_ras xgmi_ras = { 1357 .ras_block = { 1358 .hw_ops = &xgmi_ras_hw_ops, 1359 .ras_late_init = amdgpu_xgmi_ras_late_init, 1360 }, 1361 }; 1362 1363 int amdgpu_xgmi_ras_sw_init(struct amdgpu_device *adev) 1364 { 1365 int err; 1366 struct amdgpu_xgmi_ras *ras; 1367 1368 if (!adev->gmc.xgmi.ras) 1369 return 0; 1370 1371 ras = adev->gmc.xgmi.ras; 1372 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 1373 if (err) { 1374 dev_err(adev->dev, "Failed to register xgmi_wafl_pcs ras block!\n"); 1375 return err; 1376 } 1377 1378 strcpy(ras->ras_block.ras_comm.name, "xgmi_wafl"); 1379 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__XGMI_WAFL; 1380 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 1381 adev->gmc.xgmi.ras_if = &ras->ras_block.ras_comm; 1382 1383 return 0; 1384 } 1385