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 "wafl/wafl2_4_0_0_smn.h" 33 #include "wafl/wafl2_4_0_0_sh_mask.h" 34 35 #define smnPCS_XGMI23_PCS_ERROR_STATUS 0x11a01210 36 #define smnPCS_XGMI3X16_PCS_ERROR_STATUS 0x11a0020c 37 #define smnPCS_GOPX1_PCS_ERROR_STATUS 0x12200210 38 39 static DEFINE_MUTEX(xgmi_mutex); 40 41 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE 4 42 43 static LIST_HEAD(xgmi_hive_list); 44 45 static const int xgmi_pcs_err_status_reg_vg20[] = { 46 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 47 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 48 }; 49 50 static const int wafl_pcs_err_status_reg_vg20[] = { 51 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 52 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 53 }; 54 55 static const int xgmi_pcs_err_status_reg_arct[] = { 56 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 57 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 58 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000, 59 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000, 60 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000, 61 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000, 62 }; 63 64 /* same as vg20*/ 65 static const int wafl_pcs_err_status_reg_arct[] = { 66 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 67 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 68 }; 69 70 static const int xgmi23_pcs_err_status_reg_aldebaran[] = { 71 smnPCS_XGMI23_PCS_ERROR_STATUS, 72 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x100000, 73 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x200000, 74 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x300000, 75 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x400000, 76 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x500000, 77 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x600000, 78 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x700000 79 }; 80 81 static const int xgmi3x16_pcs_err_status_reg_aldebaran[] = { 82 smnPCS_XGMI3X16_PCS_ERROR_STATUS, 83 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000, 84 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x200000, 85 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x300000, 86 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x400000, 87 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x500000, 88 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x600000, 89 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x700000 90 }; 91 92 static const int walf_pcs_err_status_reg_aldebaran[] = { 93 smnPCS_GOPX1_PCS_ERROR_STATUS, 94 smnPCS_GOPX1_PCS_ERROR_STATUS + 0x100000 95 }; 96 97 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = { 98 {"XGMI PCS DataLossErr", 99 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)}, 100 {"XGMI PCS TrainingErr", 101 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)}, 102 {"XGMI PCS CRCErr", 103 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)}, 104 {"XGMI PCS BERExceededErr", 105 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)}, 106 {"XGMI PCS TxMetaDataErr", 107 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)}, 108 {"XGMI PCS ReplayBufParityErr", 109 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 110 {"XGMI PCS DataParityErr", 111 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)}, 112 {"XGMI PCS ReplayFifoOverflowErr", 113 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 114 {"XGMI PCS ReplayFifoUnderflowErr", 115 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 116 {"XGMI PCS ElasticFifoOverflowErr", 117 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 118 {"XGMI PCS DeskewErr", 119 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)}, 120 {"XGMI PCS DataStartupLimitErr", 121 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 122 {"XGMI PCS FCInitTimeoutErr", 123 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 124 {"XGMI PCS RecoveryTimeoutErr", 125 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 126 {"XGMI PCS ReadySerialTimeoutErr", 127 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 128 {"XGMI PCS ReadySerialAttemptErr", 129 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 130 {"XGMI PCS RecoveryAttemptErr", 131 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 132 {"XGMI PCS RecoveryRelockAttemptErr", 133 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 134 }; 135 136 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = { 137 {"WAFL PCS DataLossErr", 138 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)}, 139 {"WAFL PCS TrainingErr", 140 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)}, 141 {"WAFL PCS CRCErr", 142 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)}, 143 {"WAFL PCS BERExceededErr", 144 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)}, 145 {"WAFL PCS TxMetaDataErr", 146 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)}, 147 {"WAFL PCS ReplayBufParityErr", 148 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)}, 149 {"WAFL PCS DataParityErr", 150 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)}, 151 {"WAFL PCS ReplayFifoOverflowErr", 152 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 153 {"WAFL PCS ReplayFifoUnderflowErr", 154 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 155 {"WAFL PCS ElasticFifoOverflowErr", 156 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 157 {"WAFL PCS DeskewErr", 158 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)}, 159 {"WAFL PCS DataStartupLimitErr", 160 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)}, 161 {"WAFL PCS FCInitTimeoutErr", 162 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 163 {"WAFL PCS RecoveryTimeoutErr", 164 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 165 {"WAFL PCS ReadySerialTimeoutErr", 166 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 167 {"WAFL PCS ReadySerialAttemptErr", 168 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 169 {"WAFL PCS RecoveryAttemptErr", 170 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 171 {"WAFL PCS RecoveryRelockAttemptErr", 172 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 173 }; 174 175 /** 176 * DOC: AMDGPU XGMI Support 177 * 178 * XGMI is a high speed interconnect that joins multiple GPU cards 179 * into a homogeneous memory space that is organized by a collective 180 * hive ID and individual node IDs, both of which are 64-bit numbers. 181 * 182 * The file xgmi_device_id contains the unique per GPU device ID and 183 * is stored in the /sys/class/drm/card${cardno}/device/ directory. 184 * 185 * Inside the device directory a sub-directory 'xgmi_hive_info' is 186 * created which contains the hive ID and the list of nodes. 187 * 188 * The hive ID is stored in: 189 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id 190 * 191 * The node information is stored in numbered directories: 192 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id 193 * 194 * Each device has their own xgmi_hive_info direction with a mirror 195 * set of node sub-directories. 196 * 197 * The XGMI memory space is built by contiguously adding the power of 198 * two padded VRAM space from each node to each other. 199 * 200 */ 201 202 static struct attribute amdgpu_xgmi_hive_id = { 203 .name = "xgmi_hive_id", 204 .mode = S_IRUGO 205 }; 206 207 static struct attribute *amdgpu_xgmi_hive_attrs[] = { 208 &amdgpu_xgmi_hive_id, 209 NULL 210 }; 211 212 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj, 213 struct attribute *attr, char *buf) 214 { 215 struct amdgpu_hive_info *hive = container_of( 216 kobj, struct amdgpu_hive_info, kobj); 217 218 if (attr == &amdgpu_xgmi_hive_id) 219 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id); 220 221 return 0; 222 } 223 224 static void amdgpu_xgmi_hive_release(struct kobject *kobj) 225 { 226 struct amdgpu_hive_info *hive = container_of( 227 kobj, struct amdgpu_hive_info, kobj); 228 229 mutex_destroy(&hive->hive_lock); 230 kfree(hive); 231 } 232 233 static const struct sysfs_ops amdgpu_xgmi_hive_ops = { 234 .show = amdgpu_xgmi_show_attrs, 235 }; 236 237 struct kobj_type amdgpu_xgmi_hive_type = { 238 .release = amdgpu_xgmi_hive_release, 239 .sysfs_ops = &amdgpu_xgmi_hive_ops, 240 .default_attrs = amdgpu_xgmi_hive_attrs, 241 }; 242 243 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev, 244 struct device_attribute *attr, 245 char *buf) 246 { 247 struct drm_device *ddev = dev_get_drvdata(dev); 248 struct amdgpu_device *adev = drm_to_adev(ddev); 249 250 return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id); 251 252 } 253 254 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801) 255 static ssize_t amdgpu_xgmi_show_error(struct device *dev, 256 struct device_attribute *attr, 257 char *buf) 258 { 259 struct drm_device *ddev = dev_get_drvdata(dev); 260 struct amdgpu_device *adev = drm_to_adev(ddev); 261 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in; 262 uint64_t fica_out; 263 unsigned int error_count = 0; 264 265 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200); 266 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208); 267 268 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in); 269 if (fica_out != 0x1f) 270 pr_err("xGMI error counters not enabled!\n"); 271 272 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in); 273 274 if ((fica_out & 0xffff) == 2) 275 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63); 276 277 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0); 278 279 return sysfs_emit(buf, "%u\n", error_count); 280 } 281 282 283 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL); 284 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL); 285 286 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev, 287 struct amdgpu_hive_info *hive) 288 { 289 int ret = 0; 290 char node[10] = { 0 }; 291 292 /* Create xgmi device id file */ 293 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id); 294 if (ret) { 295 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n"); 296 return ret; 297 } 298 299 /* Create xgmi error file */ 300 ret = device_create_file(adev->dev, &dev_attr_xgmi_error); 301 if (ret) 302 pr_err("failed to create xgmi_error\n"); 303 304 305 /* Create sysfs link to hive info folder on the first device */ 306 if (hive->kobj.parent != (&adev->dev->kobj)) { 307 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj, 308 "xgmi_hive_info"); 309 if (ret) { 310 dev_err(adev->dev, "XGMI: Failed to create link to hive info"); 311 goto remove_file; 312 } 313 } 314 315 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 316 /* Create sysfs link form the hive folder to yourself */ 317 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node); 318 if (ret) { 319 dev_err(adev->dev, "XGMI: Failed to create link from hive info"); 320 goto remove_link; 321 } 322 323 goto success; 324 325 326 remove_link: 327 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique); 328 329 remove_file: 330 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 331 332 success: 333 return ret; 334 } 335 336 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev, 337 struct amdgpu_hive_info *hive) 338 { 339 char node[10]; 340 memset(node, 0, sizeof(node)); 341 342 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 343 device_remove_file(adev->dev, &dev_attr_xgmi_error); 344 345 if (hive->kobj.parent != (&adev->dev->kobj)) 346 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info"); 347 348 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 349 sysfs_remove_link(&hive->kobj, node); 350 351 } 352 353 354 355 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev) 356 { 357 struct amdgpu_hive_info *hive = NULL; 358 int ret; 359 360 if (!adev->gmc.xgmi.hive_id) 361 return NULL; 362 363 if (adev->hive) { 364 kobject_get(&adev->hive->kobj); 365 return adev->hive; 366 } 367 368 mutex_lock(&xgmi_mutex); 369 370 list_for_each_entry(hive, &xgmi_hive_list, node) { 371 if (hive->hive_id == adev->gmc.xgmi.hive_id) 372 goto pro_end; 373 } 374 375 hive = kzalloc(sizeof(*hive), GFP_KERNEL); 376 if (!hive) { 377 dev_err(adev->dev, "XGMI: allocation failed\n"); 378 hive = NULL; 379 goto pro_end; 380 } 381 382 /* initialize new hive if not exist */ 383 ret = kobject_init_and_add(&hive->kobj, 384 &amdgpu_xgmi_hive_type, 385 &adev->dev->kobj, 386 "%s", "xgmi_hive_info"); 387 if (ret) { 388 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n"); 389 kobject_put(&hive->kobj); 390 kfree(hive); 391 hive = NULL; 392 goto pro_end; 393 } 394 395 hive->hive_id = adev->gmc.xgmi.hive_id; 396 INIT_LIST_HEAD(&hive->device_list); 397 INIT_LIST_HEAD(&hive->node); 398 mutex_init(&hive->hive_lock); 399 atomic_set(&hive->in_reset, 0); 400 atomic_set(&hive->number_devices, 0); 401 task_barrier_init(&hive->tb); 402 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN; 403 hive->hi_req_gpu = NULL; 404 /* 405 * hive pstate on boot is high in vega20 so we have to go to low 406 * pstate on after boot. 407 */ 408 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE; 409 list_add_tail(&hive->node, &xgmi_hive_list); 410 411 pro_end: 412 if (hive) 413 kobject_get(&hive->kobj); 414 mutex_unlock(&xgmi_mutex); 415 return hive; 416 } 417 418 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive) 419 { 420 if (hive) 421 kobject_put(&hive->kobj); 422 } 423 424 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate) 425 { 426 int ret = 0; 427 struct amdgpu_hive_info *hive; 428 struct amdgpu_device *request_adev; 429 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20; 430 bool init_low; 431 432 hive = amdgpu_get_xgmi_hive(adev); 433 if (!hive) 434 return 0; 435 436 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev; 437 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN; 438 amdgpu_put_xgmi_hive(hive); 439 /* fw bug so temporarily disable pstate switching */ 440 return 0; 441 442 if (!hive || adev->asic_type != CHIP_VEGA20) 443 return 0; 444 445 mutex_lock(&hive->hive_lock); 446 447 if (is_hi_req) 448 hive->hi_req_count++; 449 else 450 hive->hi_req_count--; 451 452 /* 453 * Vega20 only needs single peer to request pstate high for the hive to 454 * go high but all peers must request pstate low for the hive to go low 455 */ 456 if (hive->pstate == pstate || 457 (!is_hi_req && hive->hi_req_count && !init_low)) 458 goto out; 459 460 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate); 461 462 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate); 463 if (ret) { 464 dev_err(request_adev->dev, 465 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d", 466 request_adev->gmc.xgmi.node_id, 467 request_adev->gmc.xgmi.hive_id, ret); 468 goto out; 469 } 470 471 if (init_low) 472 hive->pstate = hive->hi_req_count ? 473 hive->pstate : AMDGPU_XGMI_PSTATE_MIN; 474 else { 475 hive->pstate = pstate; 476 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ? 477 adev : NULL; 478 } 479 out: 480 mutex_unlock(&hive->hive_lock); 481 return ret; 482 } 483 484 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev) 485 { 486 int ret; 487 488 /* Each psp need to set the latest topology */ 489 ret = psp_xgmi_set_topology_info(&adev->psp, 490 atomic_read(&hive->number_devices), 491 &adev->psp.xgmi_context.top_info); 492 if (ret) 493 dev_err(adev->dev, 494 "XGMI: Set topology failure on device %llx, hive %llx, ret %d", 495 adev->gmc.xgmi.node_id, 496 adev->gmc.xgmi.hive_id, ret); 497 498 return ret; 499 } 500 501 502 /* 503 * NOTE psp_xgmi_node_info.num_hops layout is as follows: 504 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved) 505 * num_hops[5:3] = reserved 506 * num_hops[2:0] = number of hops 507 */ 508 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev, 509 struct amdgpu_device *peer_adev) 510 { 511 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 512 uint8_t num_hops_mask = 0x7; 513 int i; 514 515 for (i = 0 ; i < top->num_nodes; ++i) 516 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 517 return top->nodes[i].num_hops & num_hops_mask; 518 return -EINVAL; 519 } 520 521 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev, 522 struct amdgpu_device *peer_adev) 523 { 524 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 525 int i; 526 527 for (i = 0 ; i < top->num_nodes; ++i) 528 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 529 return top->nodes[i].num_links; 530 return -EINVAL; 531 } 532 533 /* 534 * Devices that support extended data require the entire hive to initialize with 535 * the shared memory buffer flag set. 536 * 537 * Hive locks and conditions apply - see amdgpu_xgmi_add_device 538 */ 539 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive, 540 bool set_extended_data) 541 { 542 struct amdgpu_device *tmp_adev; 543 int ret; 544 545 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 546 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false); 547 if (ret) { 548 dev_err(tmp_adev->dev, 549 "XGMI: Failed to initialize xgmi session for data partition %i\n", 550 set_extended_data); 551 return ret; 552 } 553 554 } 555 556 return 0; 557 } 558 559 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 560 { 561 struct psp_xgmi_topology_info *top_info; 562 struct amdgpu_hive_info *hive; 563 struct amdgpu_xgmi *entry; 564 struct amdgpu_device *tmp_adev = NULL; 565 566 int count = 0, ret = 0; 567 568 if (!adev->gmc.xgmi.supported) 569 return 0; 570 571 if (!adev->gmc.xgmi.pending_reset && 572 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 573 ret = psp_xgmi_initialize(&adev->psp, false, true); 574 if (ret) { 575 dev_err(adev->dev, 576 "XGMI: Failed to initialize xgmi session\n"); 577 return ret; 578 } 579 580 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 581 if (ret) { 582 dev_err(adev->dev, 583 "XGMI: Failed to get hive id\n"); 584 return ret; 585 } 586 587 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 588 if (ret) { 589 dev_err(adev->dev, 590 "XGMI: Failed to get node id\n"); 591 return ret; 592 } 593 } else { 594 adev->gmc.xgmi.hive_id = 16; 595 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 596 } 597 598 hive = amdgpu_get_xgmi_hive(adev); 599 if (!hive) { 600 ret = -EINVAL; 601 dev_err(adev->dev, 602 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 603 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 604 goto exit; 605 } 606 mutex_lock(&hive->hive_lock); 607 608 top_info = &adev->psp.xgmi_context.top_info; 609 610 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 611 list_for_each_entry(entry, &hive->device_list, head) 612 top_info->nodes[count++].node_id = entry->node_id; 613 top_info->num_nodes = count; 614 atomic_set(&hive->number_devices, count); 615 616 task_barrier_add_task(&hive->tb); 617 618 if (!adev->gmc.xgmi.pending_reset && 619 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 620 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 621 /* update node list for other device in the hive */ 622 if (tmp_adev != adev) { 623 top_info = &tmp_adev->psp.xgmi_context.top_info; 624 top_info->nodes[count - 1].node_id = 625 adev->gmc.xgmi.node_id; 626 top_info->num_nodes = count; 627 } 628 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 629 if (ret) 630 goto exit_unlock; 631 } 632 633 /* get latest topology info for each device from psp */ 634 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 635 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 636 &tmp_adev->psp.xgmi_context.top_info, false); 637 if (ret) { 638 dev_err(tmp_adev->dev, 639 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 640 tmp_adev->gmc.xgmi.node_id, 641 tmp_adev->gmc.xgmi.hive_id, ret); 642 /* To do : continue with some node failed or disable the whole hive */ 643 goto exit_unlock; 644 } 645 } 646 647 /* get topology again for hives that support extended data */ 648 if (adev->psp.xgmi_context.supports_extended_data) { 649 650 /* initialize the hive to get extended data. */ 651 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true); 652 if (ret) 653 goto exit_unlock; 654 655 /* get the extended data. */ 656 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 657 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 658 &tmp_adev->psp.xgmi_context.top_info, true); 659 if (ret) { 660 dev_err(tmp_adev->dev, 661 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d", 662 tmp_adev->gmc.xgmi.node_id, 663 tmp_adev->gmc.xgmi.hive_id, ret); 664 goto exit_unlock; 665 } 666 } 667 668 /* initialize the hive to get non-extended data for the next round. */ 669 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false); 670 if (ret) 671 goto exit_unlock; 672 673 } 674 } 675 676 if (!ret && !adev->gmc.xgmi.pending_reset) 677 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 678 679 exit_unlock: 680 mutex_unlock(&hive->hive_lock); 681 exit: 682 if (!ret) { 683 adev->hive = hive; 684 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 685 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 686 } else { 687 amdgpu_put_xgmi_hive(hive); 688 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 689 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 690 ret); 691 } 692 693 return ret; 694 } 695 696 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 697 { 698 struct amdgpu_hive_info *hive = adev->hive; 699 700 if (!adev->gmc.xgmi.supported) 701 return -EINVAL; 702 703 if (!hive) 704 return -EINVAL; 705 706 mutex_lock(&hive->hive_lock); 707 task_barrier_rem_task(&hive->tb); 708 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 709 if (hive->hi_req_gpu == adev) 710 hive->hi_req_gpu = NULL; 711 list_del(&adev->gmc.xgmi.head); 712 mutex_unlock(&hive->hive_lock); 713 714 amdgpu_put_xgmi_hive(hive); 715 adev->hive = NULL; 716 717 if (atomic_dec_return(&hive->number_devices) == 0) { 718 /* Remove the hive from global hive list */ 719 mutex_lock(&xgmi_mutex); 720 list_del(&hive->node); 721 mutex_unlock(&xgmi_mutex); 722 723 amdgpu_put_xgmi_hive(hive); 724 } 725 726 return psp_xgmi_terminate(&adev->psp); 727 } 728 729 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev) 730 { 731 int r; 732 struct ras_ih_if ih_info = { 733 .cb = NULL, 734 }; 735 struct ras_fs_if fs_info = { 736 .sysfs_name = "xgmi_wafl_err_count", 737 }; 738 739 if (!adev->gmc.xgmi.supported || 740 adev->gmc.xgmi.num_physical_nodes == 0) 741 return 0; 742 743 adev->gmc.xgmi.ras_funcs->reset_ras_error_count(adev); 744 745 if (!adev->gmc.xgmi.ras_if) { 746 adev->gmc.xgmi.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL); 747 if (!adev->gmc.xgmi.ras_if) 748 return -ENOMEM; 749 adev->gmc.xgmi.ras_if->block = AMDGPU_RAS_BLOCK__XGMI_WAFL; 750 adev->gmc.xgmi.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 751 adev->gmc.xgmi.ras_if->sub_block_index = 0; 752 } 753 ih_info.head = fs_info.head = *adev->gmc.xgmi.ras_if; 754 r = amdgpu_ras_late_init(adev, adev->gmc.xgmi.ras_if, 755 &fs_info, &ih_info); 756 if (r || !amdgpu_ras_is_supported(adev, adev->gmc.xgmi.ras_if->block)) { 757 kfree(adev->gmc.xgmi.ras_if); 758 adev->gmc.xgmi.ras_if = NULL; 759 } 760 761 return r; 762 } 763 764 static void amdgpu_xgmi_ras_fini(struct amdgpu_device *adev) 765 { 766 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL) && 767 adev->gmc.xgmi.ras_if) { 768 struct ras_common_if *ras_if = adev->gmc.xgmi.ras_if; 769 struct ras_ih_if ih_info = { 770 .cb = NULL, 771 }; 772 773 amdgpu_ras_late_fini(adev, ras_if, &ih_info); 774 kfree(ras_if); 775 } 776 } 777 778 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 779 uint64_t addr) 780 { 781 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 782 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 783 } 784 785 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 786 { 787 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 788 WREG32_PCIE(pcs_status_reg, 0); 789 } 790 791 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 792 { 793 uint32_t i; 794 795 switch (adev->asic_type) { 796 case CHIP_ARCTURUS: 797 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 798 pcs_clear_status(adev, 799 xgmi_pcs_err_status_reg_arct[i]); 800 break; 801 case CHIP_VEGA20: 802 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 803 pcs_clear_status(adev, 804 xgmi_pcs_err_status_reg_vg20[i]); 805 break; 806 case CHIP_ALDEBARAN: 807 for (i = 0; i < ARRAY_SIZE(xgmi23_pcs_err_status_reg_aldebaran); i++) 808 pcs_clear_status(adev, 809 xgmi23_pcs_err_status_reg_aldebaran[i]); 810 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) 811 pcs_clear_status(adev, 812 xgmi3x16_pcs_err_status_reg_aldebaran[i]); 813 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) 814 pcs_clear_status(adev, 815 walf_pcs_err_status_reg_aldebaran[i]); 816 break; 817 default: 818 break; 819 } 820 } 821 822 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 823 uint32_t value, 824 uint32_t *ue_count, 825 uint32_t *ce_count, 826 bool is_xgmi_pcs) 827 { 828 int i; 829 int ue_cnt; 830 831 if (is_xgmi_pcs) { 832 /* query xgmi pcs error status, 833 * only ue is supported */ 834 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_ras_fields); i ++) { 835 ue_cnt = (value & 836 xgmi_pcs_ras_fields[i].pcs_err_mask) >> 837 xgmi_pcs_ras_fields[i].pcs_err_shift; 838 if (ue_cnt) { 839 dev_info(adev->dev, "%s detected\n", 840 xgmi_pcs_ras_fields[i].err_name); 841 *ue_count += ue_cnt; 842 } 843 } 844 } else { 845 /* query wafl pcs error status, 846 * only ue is supported */ 847 for (i = 0; i < ARRAY_SIZE(wafl_pcs_ras_fields); i++) { 848 ue_cnt = (value & 849 wafl_pcs_ras_fields[i].pcs_err_mask) >> 850 wafl_pcs_ras_fields[i].pcs_err_shift; 851 if (ue_cnt) { 852 dev_info(adev->dev, "%s detected\n", 853 wafl_pcs_ras_fields[i].err_name); 854 *ue_count += ue_cnt; 855 } 856 } 857 } 858 859 return 0; 860 } 861 862 static int amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 863 void *ras_error_status) 864 { 865 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 866 int i; 867 uint32_t data; 868 uint32_t ue_cnt = 0, ce_cnt = 0; 869 870 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 871 return -EINVAL; 872 873 err_data->ue_count = 0; 874 err_data->ce_count = 0; 875 876 switch (adev->asic_type) { 877 case CHIP_ARCTURUS: 878 /* check xgmi pcs error */ 879 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 880 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 881 if (data) 882 amdgpu_xgmi_query_pcs_error_status(adev, 883 data, &ue_cnt, &ce_cnt, true); 884 } 885 /* check wafl pcs error */ 886 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 887 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 888 if (data) 889 amdgpu_xgmi_query_pcs_error_status(adev, 890 data, &ue_cnt, &ce_cnt, false); 891 } 892 break; 893 case CHIP_VEGA20: 894 /* check xgmi pcs error */ 895 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 896 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 897 if (data) 898 amdgpu_xgmi_query_pcs_error_status(adev, 899 data, &ue_cnt, &ce_cnt, true); 900 } 901 /* check wafl pcs error */ 902 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 903 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 904 if (data) 905 amdgpu_xgmi_query_pcs_error_status(adev, 906 data, &ue_cnt, &ce_cnt, false); 907 } 908 break; 909 case CHIP_ALDEBARAN: 910 /* check xgmi23 pcs error */ 911 for (i = 0; i < ARRAY_SIZE(xgmi23_pcs_err_status_reg_aldebaran); i++) { 912 data = RREG32_PCIE(xgmi23_pcs_err_status_reg_aldebaran[i]); 913 if (data) 914 amdgpu_xgmi_query_pcs_error_status(adev, 915 data, &ue_cnt, &ce_cnt, true); 916 } 917 /* check xgmi3x16 pcs error */ 918 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) { 919 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]); 920 if (data) 921 amdgpu_xgmi_query_pcs_error_status(adev, 922 data, &ue_cnt, &ce_cnt, true); 923 } 924 /* check wafl pcs error */ 925 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) { 926 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]); 927 if (data) 928 amdgpu_xgmi_query_pcs_error_status(adev, 929 data, &ue_cnt, &ce_cnt, false); 930 } 931 break; 932 default: 933 dev_warn(adev->dev, "XGMI RAS error query not supported"); 934 break; 935 } 936 937 adev->gmc.xgmi.ras_funcs->reset_ras_error_count(adev); 938 939 err_data->ue_count += ue_cnt; 940 err_data->ce_count += ce_cnt; 941 942 return 0; 943 } 944 945 const struct amdgpu_xgmi_ras_funcs xgmi_ras_funcs = { 946 .ras_late_init = amdgpu_xgmi_ras_late_init, 947 .ras_fini = amdgpu_xgmi_ras_fini, 948 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 949 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 950 }; 951