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 struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = { 107 {"XGMI PCS DataLossErr", 108 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)}, 109 {"XGMI PCS TrainingErr", 110 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)}, 111 {"XGMI PCS CRCErr", 112 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)}, 113 {"XGMI PCS BERExceededErr", 114 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)}, 115 {"XGMI PCS TxMetaDataErr", 116 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)}, 117 {"XGMI PCS ReplayBufParityErr", 118 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 119 {"XGMI PCS DataParityErr", 120 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)}, 121 {"XGMI PCS ReplayFifoOverflowErr", 122 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 123 {"XGMI PCS ReplayFifoUnderflowErr", 124 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 125 {"XGMI PCS ElasticFifoOverflowErr", 126 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 127 {"XGMI PCS DeskewErr", 128 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)}, 129 {"XGMI PCS DataStartupLimitErr", 130 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 131 {"XGMI PCS FCInitTimeoutErr", 132 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 133 {"XGMI PCS RecoveryTimeoutErr", 134 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 135 {"XGMI PCS ReadySerialTimeoutErr", 136 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 137 {"XGMI PCS ReadySerialAttemptErr", 138 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 139 {"XGMI PCS RecoveryAttemptErr", 140 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 141 {"XGMI PCS RecoveryRelockAttemptErr", 142 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 143 }; 144 145 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = { 146 {"WAFL PCS DataLossErr", 147 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)}, 148 {"WAFL PCS TrainingErr", 149 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)}, 150 {"WAFL PCS CRCErr", 151 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)}, 152 {"WAFL PCS BERExceededErr", 153 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)}, 154 {"WAFL PCS TxMetaDataErr", 155 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)}, 156 {"WAFL PCS ReplayBufParityErr", 157 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)}, 158 {"WAFL PCS DataParityErr", 159 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)}, 160 {"WAFL PCS ReplayFifoOverflowErr", 161 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 162 {"WAFL PCS ReplayFifoUnderflowErr", 163 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 164 {"WAFL PCS ElasticFifoOverflowErr", 165 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 166 {"WAFL PCS DeskewErr", 167 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)}, 168 {"WAFL PCS DataStartupLimitErr", 169 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)}, 170 {"WAFL PCS FCInitTimeoutErr", 171 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 172 {"WAFL PCS RecoveryTimeoutErr", 173 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 174 {"WAFL PCS ReadySerialTimeoutErr", 175 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 176 {"WAFL PCS ReadySerialAttemptErr", 177 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 178 {"WAFL PCS RecoveryAttemptErr", 179 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 180 {"WAFL PCS RecoveryRelockAttemptErr", 181 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 182 }; 183 184 static const struct amdgpu_pcs_ras_field xgmi3x16_pcs_ras_fields[] = { 185 {"XGMI3X16 PCS DataLossErr", 186 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataLossErr)}, 187 {"XGMI3X16 PCS TrainingErr", 188 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TrainingErr)}, 189 {"XGMI3X16 PCS FlowCtrlAckErr", 190 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FlowCtrlAckErr)}, 191 {"XGMI3X16 PCS RxFifoUnderflowErr", 192 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxFifoUnderflowErr)}, 193 {"XGMI3X16 PCS RxFifoOverflowErr", 194 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxFifoOverflowErr)}, 195 {"XGMI3X16 PCS CRCErr", 196 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, CRCErr)}, 197 {"XGMI3X16 PCS BERExceededErr", 198 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, BERExceededErr)}, 199 {"XGMI3X16 PCS TxVcidDataErr", 200 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TxVcidDataErr)}, 201 {"XGMI3X16 PCS ReplayBufParityErr", 202 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 203 {"XGMI3X16 PCS DataParityErr", 204 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataParityErr)}, 205 {"XGMI3X16 PCS ReplayFifoOverflowErr", 206 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 207 {"XGMI3X16 PCS ReplayFifoUnderflowErr", 208 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 209 {"XGMI3X16 PCS ElasticFifoOverflowErr", 210 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 211 {"XGMI3X16 PCS DeskewErr", 212 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DeskewErr)}, 213 {"XGMI3X16 PCS FlowCtrlCRCErr", 214 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FlowCtrlCRCErr)}, 215 {"XGMI3X16 PCS DataStartupLimitErr", 216 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 217 {"XGMI3X16 PCS FCInitTimeoutErr", 218 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 219 {"XGMI3X16 PCS RecoveryTimeoutErr", 220 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 221 {"XGMI3X16 PCS ReadySerialTimeoutErr", 222 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 223 {"XGMI3X16 PCS ReadySerialAttemptErr", 224 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 225 {"XGMI3X16 PCS RecoveryAttemptErr", 226 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 227 {"XGMI3X16 PCS RecoveryRelockAttemptErr", 228 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 229 {"XGMI3X16 PCS ReplayAttemptErr", 230 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayAttemptErr)}, 231 {"XGMI3X16 PCS SyncHdrErr", 232 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, SyncHdrErr)}, 233 {"XGMI3X16 PCS TxReplayTimeoutErr", 234 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TxReplayTimeoutErr)}, 235 {"XGMI3X16 PCS RxReplayTimeoutErr", 236 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxReplayTimeoutErr)}, 237 {"XGMI3X16 PCS LinkSubTxTimeoutErr", 238 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, LinkSubTxTimeoutErr)}, 239 {"XGMI3X16 PCS LinkSubRxTimeoutErr", 240 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, LinkSubRxTimeoutErr)}, 241 {"XGMI3X16 PCS RxCMDPktErr", 242 SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxCMDPktErr)}, 243 }; 244 245 /** 246 * DOC: AMDGPU XGMI Support 247 * 248 * XGMI is a high speed interconnect that joins multiple GPU cards 249 * into a homogeneous memory space that is organized by a collective 250 * hive ID and individual node IDs, both of which are 64-bit numbers. 251 * 252 * The file xgmi_device_id contains the unique per GPU device ID and 253 * is stored in the /sys/class/drm/card${cardno}/device/ directory. 254 * 255 * Inside the device directory a sub-directory 'xgmi_hive_info' is 256 * created which contains the hive ID and the list of nodes. 257 * 258 * The hive ID is stored in: 259 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id 260 * 261 * The node information is stored in numbered directories: 262 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id 263 * 264 * Each device has their own xgmi_hive_info direction with a mirror 265 * set of node sub-directories. 266 * 267 * The XGMI memory space is built by contiguously adding the power of 268 * two padded VRAM space from each node to each other. 269 * 270 */ 271 272 static struct attribute amdgpu_xgmi_hive_id = { 273 .name = "xgmi_hive_id", 274 .mode = S_IRUGO 275 }; 276 277 static struct attribute *amdgpu_xgmi_hive_attrs[] = { 278 &amdgpu_xgmi_hive_id, 279 NULL 280 }; 281 ATTRIBUTE_GROUPS(amdgpu_xgmi_hive); 282 283 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj, 284 struct attribute *attr, char *buf) 285 { 286 struct amdgpu_hive_info *hive = container_of( 287 kobj, struct amdgpu_hive_info, kobj); 288 289 if (attr == &amdgpu_xgmi_hive_id) 290 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id); 291 292 return 0; 293 } 294 295 static void amdgpu_xgmi_hive_release(struct kobject *kobj) 296 { 297 struct amdgpu_hive_info *hive = container_of( 298 kobj, struct amdgpu_hive_info, kobj); 299 300 amdgpu_reset_put_reset_domain(hive->reset_domain); 301 hive->reset_domain = NULL; 302 303 mutex_destroy(&hive->hive_lock); 304 kfree(hive); 305 } 306 307 static const struct sysfs_ops amdgpu_xgmi_hive_ops = { 308 .show = amdgpu_xgmi_show_attrs, 309 }; 310 311 static const struct kobj_type amdgpu_xgmi_hive_type = { 312 .release = amdgpu_xgmi_hive_release, 313 .sysfs_ops = &amdgpu_xgmi_hive_ops, 314 .default_groups = amdgpu_xgmi_hive_groups, 315 }; 316 317 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev, 318 struct device_attribute *attr, 319 char *buf) 320 { 321 struct drm_device *ddev = dev_get_drvdata(dev); 322 struct amdgpu_device *adev = drm_to_adev(ddev); 323 324 return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id); 325 326 } 327 328 static ssize_t amdgpu_xgmi_show_num_hops(struct device *dev, 329 struct device_attribute *attr, 330 char *buf) 331 { 332 struct drm_device *ddev = dev_get_drvdata(dev); 333 struct amdgpu_device *adev = drm_to_adev(ddev); 334 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 335 int i; 336 337 for (i = 0; i < top->num_nodes; i++) 338 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_hops); 339 340 return sysfs_emit(buf, "%s\n", buf); 341 } 342 343 static ssize_t amdgpu_xgmi_show_num_links(struct device *dev, 344 struct device_attribute *attr, 345 char *buf) 346 { 347 struct drm_device *ddev = dev_get_drvdata(dev); 348 struct amdgpu_device *adev = drm_to_adev(ddev); 349 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 350 int i; 351 352 for (i = 0; i < top->num_nodes; i++) 353 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_links); 354 355 return sysfs_emit(buf, "%s\n", buf); 356 } 357 358 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801) 359 static ssize_t amdgpu_xgmi_show_error(struct device *dev, 360 struct device_attribute *attr, 361 char *buf) 362 { 363 struct drm_device *ddev = dev_get_drvdata(dev); 364 struct amdgpu_device *adev = drm_to_adev(ddev); 365 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in; 366 uint64_t fica_out; 367 unsigned int error_count = 0; 368 369 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200); 370 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208); 371 372 if ((!adev->df.funcs) || 373 (!adev->df.funcs->get_fica) || 374 (!adev->df.funcs->set_fica)) 375 return -EINVAL; 376 377 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in); 378 if (fica_out != 0x1f) 379 pr_err("xGMI error counters not enabled!\n"); 380 381 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in); 382 383 if ((fica_out & 0xffff) == 2) 384 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63); 385 386 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0); 387 388 return sysfs_emit(buf, "%u\n", error_count); 389 } 390 391 392 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL); 393 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL); 394 static DEVICE_ATTR(xgmi_num_hops, S_IRUGO, amdgpu_xgmi_show_num_hops, NULL); 395 static DEVICE_ATTR(xgmi_num_links, S_IRUGO, amdgpu_xgmi_show_num_links, NULL); 396 397 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev, 398 struct amdgpu_hive_info *hive) 399 { 400 int ret = 0; 401 char node[10] = { 0 }; 402 403 /* Create xgmi device id file */ 404 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id); 405 if (ret) { 406 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n"); 407 return ret; 408 } 409 410 /* Create xgmi error file */ 411 ret = device_create_file(adev->dev, &dev_attr_xgmi_error); 412 if (ret) 413 pr_err("failed to create xgmi_error\n"); 414 415 /* Create xgmi num hops file */ 416 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_hops); 417 if (ret) 418 pr_err("failed to create xgmi_num_hops\n"); 419 420 /* Create xgmi num links file */ 421 ret = device_create_file(adev->dev, &dev_attr_xgmi_num_links); 422 if (ret) 423 pr_err("failed to create xgmi_num_links\n"); 424 425 /* Create sysfs link to hive info folder on the first device */ 426 if (hive->kobj.parent != (&adev->dev->kobj)) { 427 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj, 428 "xgmi_hive_info"); 429 if (ret) { 430 dev_err(adev->dev, "XGMI: Failed to create link to hive info"); 431 goto remove_file; 432 } 433 } 434 435 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 436 /* Create sysfs link form the hive folder to yourself */ 437 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node); 438 if (ret) { 439 dev_err(adev->dev, "XGMI: Failed to create link from hive info"); 440 goto remove_link; 441 } 442 443 goto success; 444 445 446 remove_link: 447 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique); 448 449 remove_file: 450 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 451 device_remove_file(adev->dev, &dev_attr_xgmi_error); 452 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 453 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 454 455 success: 456 return ret; 457 } 458 459 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev, 460 struct amdgpu_hive_info *hive) 461 { 462 char node[10]; 463 memset(node, 0, sizeof(node)); 464 465 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 466 device_remove_file(adev->dev, &dev_attr_xgmi_error); 467 device_remove_file(adev->dev, &dev_attr_xgmi_num_hops); 468 device_remove_file(adev->dev, &dev_attr_xgmi_num_links); 469 470 if (hive->kobj.parent != (&adev->dev->kobj)) 471 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info"); 472 473 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 474 sysfs_remove_link(&hive->kobj, node); 475 476 } 477 478 479 480 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev) 481 { 482 struct amdgpu_hive_info *hive = NULL; 483 int ret; 484 485 if (!adev->gmc.xgmi.hive_id) 486 return NULL; 487 488 if (adev->hive) { 489 kobject_get(&adev->hive->kobj); 490 return adev->hive; 491 } 492 493 mutex_lock(&xgmi_mutex); 494 495 list_for_each_entry(hive, &xgmi_hive_list, node) { 496 if (hive->hive_id == adev->gmc.xgmi.hive_id) 497 goto pro_end; 498 } 499 500 hive = kzalloc(sizeof(*hive), GFP_KERNEL); 501 if (!hive) { 502 dev_err(adev->dev, "XGMI: allocation failed\n"); 503 ret = -ENOMEM; 504 hive = NULL; 505 goto pro_end; 506 } 507 508 /* initialize new hive if not exist */ 509 ret = kobject_init_and_add(&hive->kobj, 510 &amdgpu_xgmi_hive_type, 511 &adev->dev->kobj, 512 "%s", "xgmi_hive_info"); 513 if (ret) { 514 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n"); 515 kobject_put(&hive->kobj); 516 hive = NULL; 517 goto pro_end; 518 } 519 520 /** 521 * Only init hive->reset_domain for none SRIOV configuration. For SRIOV, 522 * Host driver decide how to reset the GPU either through FLR or chain reset. 523 * Guest side will get individual notifications from the host for the FLR 524 * if necessary. 525 */ 526 if (!amdgpu_sriov_vf(adev)) { 527 /** 528 * Avoid recreating reset domain when hive is reconstructed for the case 529 * of reset the devices in the XGMI hive during probe for passthrough GPU 530 * See https://www.spinics.net/lists/amd-gfx/msg58836.html 531 */ 532 if (adev->reset_domain->type != XGMI_HIVE) { 533 hive->reset_domain = 534 amdgpu_reset_create_reset_domain(XGMI_HIVE, "amdgpu-reset-hive"); 535 if (!hive->reset_domain) { 536 dev_err(adev->dev, "XGMI: failed initializing reset domain for xgmi hive\n"); 537 ret = -ENOMEM; 538 kobject_put(&hive->kobj); 539 hive = NULL; 540 goto pro_end; 541 } 542 } else { 543 amdgpu_reset_get_reset_domain(adev->reset_domain); 544 hive->reset_domain = adev->reset_domain; 545 } 546 } 547 548 hive->hive_id = adev->gmc.xgmi.hive_id; 549 INIT_LIST_HEAD(&hive->device_list); 550 INIT_LIST_HEAD(&hive->node); 551 mutex_init(&hive->hive_lock); 552 atomic_set(&hive->number_devices, 0); 553 task_barrier_init(&hive->tb); 554 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN; 555 hive->hi_req_gpu = NULL; 556 557 /* 558 * hive pstate on boot is high in vega20 so we have to go to low 559 * pstate on after boot. 560 */ 561 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE; 562 list_add_tail(&hive->node, &xgmi_hive_list); 563 564 pro_end: 565 if (hive) 566 kobject_get(&hive->kobj); 567 mutex_unlock(&xgmi_mutex); 568 return hive; 569 } 570 571 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive) 572 { 573 if (hive) 574 kobject_put(&hive->kobj); 575 } 576 577 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate) 578 { 579 int ret = 0; 580 struct amdgpu_hive_info *hive; 581 struct amdgpu_device *request_adev; 582 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20; 583 bool init_low; 584 585 hive = amdgpu_get_xgmi_hive(adev); 586 if (!hive) 587 return 0; 588 589 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev; 590 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN; 591 amdgpu_put_xgmi_hive(hive); 592 /* fw bug so temporarily disable pstate switching */ 593 return 0; 594 595 if (!hive || adev->asic_type != CHIP_VEGA20) 596 return 0; 597 598 mutex_lock(&hive->hive_lock); 599 600 if (is_hi_req) 601 hive->hi_req_count++; 602 else 603 hive->hi_req_count--; 604 605 /* 606 * Vega20 only needs single peer to request pstate high for the hive to 607 * go high but all peers must request pstate low for the hive to go low 608 */ 609 if (hive->pstate == pstate || 610 (!is_hi_req && hive->hi_req_count && !init_low)) 611 goto out; 612 613 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate); 614 615 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate); 616 if (ret) { 617 dev_err(request_adev->dev, 618 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d", 619 request_adev->gmc.xgmi.node_id, 620 request_adev->gmc.xgmi.hive_id, ret); 621 goto out; 622 } 623 624 if (init_low) 625 hive->pstate = hive->hi_req_count ? 626 hive->pstate : AMDGPU_XGMI_PSTATE_MIN; 627 else { 628 hive->pstate = pstate; 629 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ? 630 adev : NULL; 631 } 632 out: 633 mutex_unlock(&hive->hive_lock); 634 return ret; 635 } 636 637 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev) 638 { 639 int ret; 640 641 if (amdgpu_sriov_vf(adev)) 642 return 0; 643 644 /* Each psp need to set the latest topology */ 645 ret = psp_xgmi_set_topology_info(&adev->psp, 646 atomic_read(&hive->number_devices), 647 &adev->psp.xgmi_context.top_info); 648 if (ret) 649 dev_err(adev->dev, 650 "XGMI: Set topology failure on device %llx, hive %llx, ret %d", 651 adev->gmc.xgmi.node_id, 652 adev->gmc.xgmi.hive_id, ret); 653 654 return ret; 655 } 656 657 658 /* 659 * NOTE psp_xgmi_node_info.num_hops layout is as follows: 660 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved) 661 * num_hops[5:3] = reserved 662 * num_hops[2:0] = number of hops 663 */ 664 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev, 665 struct amdgpu_device *peer_adev) 666 { 667 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 668 uint8_t num_hops_mask = 0x7; 669 int i; 670 671 for (i = 0 ; i < top->num_nodes; ++i) 672 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 673 return top->nodes[i].num_hops & num_hops_mask; 674 return -EINVAL; 675 } 676 677 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev, 678 struct amdgpu_device *peer_adev) 679 { 680 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 681 int i; 682 683 for (i = 0 ; i < top->num_nodes; ++i) 684 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 685 return top->nodes[i].num_links; 686 return -EINVAL; 687 } 688 689 /* 690 * Devices that support extended data require the entire hive to initialize with 691 * the shared memory buffer flag set. 692 * 693 * Hive locks and conditions apply - see amdgpu_xgmi_add_device 694 */ 695 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive, 696 bool set_extended_data) 697 { 698 struct amdgpu_device *tmp_adev; 699 int ret; 700 701 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 702 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false); 703 if (ret) { 704 dev_err(tmp_adev->dev, 705 "XGMI: Failed to initialize xgmi session for data partition %i\n", 706 set_extended_data); 707 return ret; 708 } 709 710 } 711 712 return 0; 713 } 714 715 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 716 { 717 struct psp_xgmi_topology_info *top_info; 718 struct amdgpu_hive_info *hive; 719 struct amdgpu_xgmi *entry; 720 struct amdgpu_device *tmp_adev = NULL; 721 722 int count = 0, ret = 0; 723 724 if (!adev->gmc.xgmi.supported) 725 return 0; 726 727 if (!adev->gmc.xgmi.pending_reset && 728 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 729 ret = psp_xgmi_initialize(&adev->psp, false, true); 730 if (ret) { 731 dev_err(adev->dev, 732 "XGMI: Failed to initialize xgmi session\n"); 733 return ret; 734 } 735 736 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 737 if (ret) { 738 dev_err(adev->dev, 739 "XGMI: Failed to get hive id\n"); 740 return ret; 741 } 742 743 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 744 if (ret) { 745 dev_err(adev->dev, 746 "XGMI: Failed to get node id\n"); 747 return ret; 748 } 749 } else { 750 adev->gmc.xgmi.hive_id = 16; 751 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 752 } 753 754 hive = amdgpu_get_xgmi_hive(adev); 755 if (!hive) { 756 ret = -EINVAL; 757 dev_err(adev->dev, 758 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 759 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 760 goto exit; 761 } 762 mutex_lock(&hive->hive_lock); 763 764 top_info = &adev->psp.xgmi_context.top_info; 765 766 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 767 list_for_each_entry(entry, &hive->device_list, head) 768 top_info->nodes[count++].node_id = entry->node_id; 769 top_info->num_nodes = count; 770 atomic_set(&hive->number_devices, count); 771 772 task_barrier_add_task(&hive->tb); 773 774 if (!adev->gmc.xgmi.pending_reset && 775 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 776 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 777 /* update node list for other device in the hive */ 778 if (tmp_adev != adev) { 779 top_info = &tmp_adev->psp.xgmi_context.top_info; 780 top_info->nodes[count - 1].node_id = 781 adev->gmc.xgmi.node_id; 782 top_info->num_nodes = count; 783 } 784 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 785 if (ret) 786 goto exit_unlock; 787 } 788 789 /* get latest topology info for each device from psp */ 790 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 791 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 792 &tmp_adev->psp.xgmi_context.top_info, false); 793 if (ret) { 794 dev_err(tmp_adev->dev, 795 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 796 tmp_adev->gmc.xgmi.node_id, 797 tmp_adev->gmc.xgmi.hive_id, ret); 798 /* To do : continue with some node failed or disable the whole hive */ 799 goto exit_unlock; 800 } 801 } 802 803 /* get topology again for hives that support extended data */ 804 if (adev->psp.xgmi_context.supports_extended_data) { 805 806 /* initialize the hive to get extended data. */ 807 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true); 808 if (ret) 809 goto exit_unlock; 810 811 /* get the extended data. */ 812 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 813 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 814 &tmp_adev->psp.xgmi_context.top_info, true); 815 if (ret) { 816 dev_err(tmp_adev->dev, 817 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d", 818 tmp_adev->gmc.xgmi.node_id, 819 tmp_adev->gmc.xgmi.hive_id, ret); 820 goto exit_unlock; 821 } 822 } 823 824 /* initialize the hive to get non-extended data for the next round. */ 825 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false); 826 if (ret) 827 goto exit_unlock; 828 829 } 830 } 831 832 if (!ret && !adev->gmc.xgmi.pending_reset) 833 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 834 835 exit_unlock: 836 mutex_unlock(&hive->hive_lock); 837 exit: 838 if (!ret) { 839 adev->hive = hive; 840 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 841 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 842 } else { 843 amdgpu_put_xgmi_hive(hive); 844 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 845 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 846 ret); 847 } 848 849 return ret; 850 } 851 852 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 853 { 854 struct amdgpu_hive_info *hive = adev->hive; 855 856 if (!adev->gmc.xgmi.supported) 857 return -EINVAL; 858 859 if (!hive) 860 return -EINVAL; 861 862 mutex_lock(&hive->hive_lock); 863 task_barrier_rem_task(&hive->tb); 864 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 865 if (hive->hi_req_gpu == adev) 866 hive->hi_req_gpu = NULL; 867 list_del(&adev->gmc.xgmi.head); 868 mutex_unlock(&hive->hive_lock); 869 870 amdgpu_put_xgmi_hive(hive); 871 adev->hive = NULL; 872 873 if (atomic_dec_return(&hive->number_devices) == 0) { 874 /* Remove the hive from global hive list */ 875 mutex_lock(&xgmi_mutex); 876 list_del(&hive->node); 877 mutex_unlock(&xgmi_mutex); 878 879 amdgpu_put_xgmi_hive(hive); 880 } 881 882 return 0; 883 } 884 885 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 886 { 887 if (!adev->gmc.xgmi.supported || 888 adev->gmc.xgmi.num_physical_nodes == 0) 889 return 0; 890 891 adev->gmc.xgmi.ras->ras_block.hw_ops->reset_ras_error_count(adev); 892 893 return amdgpu_ras_block_late_init(adev, ras_block); 894 } 895 896 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 897 uint64_t addr) 898 { 899 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 900 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 901 } 902 903 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 904 { 905 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 906 WREG32_PCIE(pcs_status_reg, 0); 907 } 908 909 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 910 { 911 uint32_t i; 912 913 switch (adev->asic_type) { 914 case CHIP_ARCTURUS: 915 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 916 pcs_clear_status(adev, 917 xgmi_pcs_err_status_reg_arct[i]); 918 break; 919 case CHIP_VEGA20: 920 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 921 pcs_clear_status(adev, 922 xgmi_pcs_err_status_reg_vg20[i]); 923 break; 924 case CHIP_ALDEBARAN: 925 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) 926 pcs_clear_status(adev, 927 xgmi3x16_pcs_err_status_reg_aldebaran[i]); 928 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) 929 pcs_clear_status(adev, 930 walf_pcs_err_status_reg_aldebaran[i]); 931 break; 932 default: 933 break; 934 } 935 } 936 937 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 938 uint32_t value, 939 uint32_t mask_value, 940 uint32_t *ue_count, 941 uint32_t *ce_count, 942 bool is_xgmi_pcs, 943 bool check_mask) 944 { 945 int i; 946 int ue_cnt = 0; 947 const struct amdgpu_pcs_ras_field *pcs_ras_fields = NULL; 948 uint32_t field_array_size = 0; 949 950 if (is_xgmi_pcs) { 951 if (adev->ip_versions[XGMI_HWIP][0] == IP_VERSION(6, 1, 0)) { 952 pcs_ras_fields = &xgmi3x16_pcs_ras_fields[0]; 953 field_array_size = ARRAY_SIZE(xgmi3x16_pcs_ras_fields); 954 } else { 955 pcs_ras_fields = &xgmi_pcs_ras_fields[0]; 956 field_array_size = ARRAY_SIZE(xgmi_pcs_ras_fields); 957 } 958 } else { 959 pcs_ras_fields = &wafl_pcs_ras_fields[0]; 960 field_array_size = ARRAY_SIZE(wafl_pcs_ras_fields); 961 } 962 963 if (check_mask) 964 value = value & ~mask_value; 965 966 /* query xgmi/walf pcs error status, 967 * only ue is supported */ 968 for (i = 0; value && i < field_array_size; i++) { 969 ue_cnt = (value & 970 pcs_ras_fields[i].pcs_err_mask) >> 971 pcs_ras_fields[i].pcs_err_shift; 972 if (ue_cnt) { 973 dev_info(adev->dev, "%s detected\n", 974 pcs_ras_fields[i].err_name); 975 *ue_count += ue_cnt; 976 } 977 978 /* reset bit value if the bit is checked */ 979 value &= ~(pcs_ras_fields[i].pcs_err_mask); 980 } 981 982 return 0; 983 } 984 985 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 986 void *ras_error_status) 987 { 988 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 989 int i; 990 uint32_t data, mask_data = 0; 991 uint32_t ue_cnt = 0, ce_cnt = 0; 992 993 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 994 return ; 995 996 err_data->ue_count = 0; 997 err_data->ce_count = 0; 998 999 switch (adev->asic_type) { 1000 case CHIP_ARCTURUS: 1001 /* check xgmi pcs error */ 1002 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 1003 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 1004 if (data) 1005 amdgpu_xgmi_query_pcs_error_status(adev, data, 1006 mask_data, &ue_cnt, &ce_cnt, true, false); 1007 } 1008 /* check wafl pcs error */ 1009 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 1010 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 1011 if (data) 1012 amdgpu_xgmi_query_pcs_error_status(adev, data, 1013 mask_data, &ue_cnt, &ce_cnt, false, false); 1014 } 1015 break; 1016 case CHIP_VEGA20: 1017 /* check xgmi pcs error */ 1018 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 1019 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 1020 if (data) 1021 amdgpu_xgmi_query_pcs_error_status(adev, data, 1022 mask_data, &ue_cnt, &ce_cnt, true, false); 1023 } 1024 /* check wafl pcs error */ 1025 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 1026 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 1027 if (data) 1028 amdgpu_xgmi_query_pcs_error_status(adev, data, 1029 mask_data, &ue_cnt, &ce_cnt, false, false); 1030 } 1031 break; 1032 case CHIP_ALDEBARAN: 1033 /* check xgmi3x16 pcs error */ 1034 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) { 1035 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]); 1036 mask_data = 1037 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1038 if (data) 1039 amdgpu_xgmi_query_pcs_error_status(adev, data, 1040 mask_data, &ue_cnt, &ce_cnt, true, true); 1041 } 1042 /* check wafl pcs error */ 1043 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) { 1044 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]); 1045 mask_data = 1046 RREG32_PCIE(walf_pcs_err_noncorrectable_mask_reg_aldebaran[i]); 1047 if (data) 1048 amdgpu_xgmi_query_pcs_error_status(adev, data, 1049 mask_data, &ue_cnt, &ce_cnt, false, true); 1050 } 1051 break; 1052 default: 1053 dev_warn(adev->dev, "XGMI RAS error query not supported"); 1054 break; 1055 } 1056 1057 adev->gmc.xgmi.ras->ras_block.hw_ops->reset_ras_error_count(adev); 1058 1059 err_data->ue_count += ue_cnt; 1060 err_data->ce_count += ce_cnt; 1061 } 1062 1063 /* Trigger XGMI/WAFL error */ 1064 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev, 1065 void *inject_if, uint32_t instance_mask) 1066 { 1067 int ret = 0; 1068 struct ta_ras_trigger_error_input *block_info = 1069 (struct ta_ras_trigger_error_input *)inject_if; 1070 1071 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) 1072 dev_warn(adev->dev, "Failed to disallow df cstate"); 1073 1074 if (amdgpu_dpm_allow_xgmi_power_down(adev, false)) 1075 dev_warn(adev->dev, "Failed to disallow XGMI power down"); 1076 1077 ret = psp_ras_trigger_error(&adev->psp, block_info, instance_mask); 1078 1079 if (amdgpu_ras_intr_triggered()) 1080 return ret; 1081 1082 if (amdgpu_dpm_allow_xgmi_power_down(adev, true)) 1083 dev_warn(adev->dev, "Failed to allow XGMI power down"); 1084 1085 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW)) 1086 dev_warn(adev->dev, "Failed to allow df cstate"); 1087 1088 return ret; 1089 } 1090 1091 struct amdgpu_ras_block_hw_ops xgmi_ras_hw_ops = { 1092 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 1093 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 1094 .ras_error_inject = amdgpu_ras_error_inject_xgmi, 1095 }; 1096 1097 struct amdgpu_xgmi_ras xgmi_ras = { 1098 .ras_block = { 1099 .hw_ops = &xgmi_ras_hw_ops, 1100 .ras_late_init = amdgpu_xgmi_ras_late_init, 1101 }, 1102 }; 1103 1104 int amdgpu_xgmi_ras_sw_init(struct amdgpu_device *adev) 1105 { 1106 int err; 1107 struct amdgpu_xgmi_ras *ras; 1108 1109 if (!adev->gmc.xgmi.ras) 1110 return 0; 1111 1112 ras = adev->gmc.xgmi.ras; 1113 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 1114 if (err) { 1115 dev_err(adev->dev, "Failed to register xgmi_wafl_pcs ras block!\n"); 1116 return err; 1117 } 1118 1119 strcpy(ras->ras_block.ras_comm.name, "xgmi_wafl"); 1120 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__XGMI_WAFL; 1121 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 1122 adev->gmc.xgmi.ras_if = &ras->ras_block.ras_comm; 1123 1124 return 0; 1125 } 1126