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 #include "amdgpu_reset.h" 36 37 #define smnPCS_XGMI3X16_PCS_ERROR_STATUS 0x11a0020c 38 #define smnPCS_GOPX1_PCS_ERROR_STATUS 0x12200210 39 40 static DEFINE_MUTEX(xgmi_mutex); 41 42 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE 4 43 44 static LIST_HEAD(xgmi_hive_list); 45 46 static const int xgmi_pcs_err_status_reg_vg20[] = { 47 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 48 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 49 }; 50 51 static const int wafl_pcs_err_status_reg_vg20[] = { 52 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 53 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 54 }; 55 56 static const int xgmi_pcs_err_status_reg_arct[] = { 57 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 58 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 59 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000, 60 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000, 61 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000, 62 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000, 63 }; 64 65 /* same as vg20*/ 66 static const int wafl_pcs_err_status_reg_arct[] = { 67 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 68 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 69 }; 70 71 static const int xgmi3x16_pcs_err_status_reg_aldebaran[] = { 72 smnPCS_XGMI3X16_PCS_ERROR_STATUS, 73 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000, 74 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x200000, 75 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x300000, 76 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x400000, 77 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x500000, 78 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x600000, 79 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x700000 80 }; 81 82 static const int walf_pcs_err_status_reg_aldebaran[] = { 83 smnPCS_GOPX1_PCS_ERROR_STATUS, 84 smnPCS_GOPX1_PCS_ERROR_STATUS + 0x100000 85 }; 86 87 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = { 88 {"XGMI PCS DataLossErr", 89 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)}, 90 {"XGMI PCS TrainingErr", 91 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)}, 92 {"XGMI PCS CRCErr", 93 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)}, 94 {"XGMI PCS BERExceededErr", 95 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)}, 96 {"XGMI PCS TxMetaDataErr", 97 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)}, 98 {"XGMI PCS ReplayBufParityErr", 99 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 100 {"XGMI PCS DataParityErr", 101 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)}, 102 {"XGMI PCS ReplayFifoOverflowErr", 103 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 104 {"XGMI PCS ReplayFifoUnderflowErr", 105 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 106 {"XGMI PCS ElasticFifoOverflowErr", 107 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 108 {"XGMI PCS DeskewErr", 109 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)}, 110 {"XGMI PCS DataStartupLimitErr", 111 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 112 {"XGMI PCS FCInitTimeoutErr", 113 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 114 {"XGMI PCS RecoveryTimeoutErr", 115 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 116 {"XGMI PCS ReadySerialTimeoutErr", 117 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 118 {"XGMI PCS ReadySerialAttemptErr", 119 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 120 {"XGMI PCS RecoveryAttemptErr", 121 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 122 {"XGMI PCS RecoveryRelockAttemptErr", 123 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 124 }; 125 126 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = { 127 {"WAFL PCS DataLossErr", 128 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)}, 129 {"WAFL PCS TrainingErr", 130 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)}, 131 {"WAFL PCS CRCErr", 132 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)}, 133 {"WAFL PCS BERExceededErr", 134 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)}, 135 {"WAFL PCS TxMetaDataErr", 136 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)}, 137 {"WAFL PCS ReplayBufParityErr", 138 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)}, 139 {"WAFL PCS DataParityErr", 140 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)}, 141 {"WAFL PCS ReplayFifoOverflowErr", 142 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 143 {"WAFL PCS ReplayFifoUnderflowErr", 144 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 145 {"WAFL PCS ElasticFifoOverflowErr", 146 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 147 {"WAFL PCS DeskewErr", 148 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)}, 149 {"WAFL PCS DataStartupLimitErr", 150 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)}, 151 {"WAFL PCS FCInitTimeoutErr", 152 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 153 {"WAFL PCS RecoveryTimeoutErr", 154 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 155 {"WAFL PCS ReadySerialTimeoutErr", 156 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 157 {"WAFL PCS ReadySerialAttemptErr", 158 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 159 {"WAFL PCS RecoveryAttemptErr", 160 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 161 {"WAFL PCS RecoveryRelockAttemptErr", 162 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 163 }; 164 165 /** 166 * DOC: AMDGPU XGMI Support 167 * 168 * XGMI is a high speed interconnect that joins multiple GPU cards 169 * into a homogeneous memory space that is organized by a collective 170 * hive ID and individual node IDs, both of which are 64-bit numbers. 171 * 172 * The file xgmi_device_id contains the unique per GPU device ID and 173 * is stored in the /sys/class/drm/card${cardno}/device/ directory. 174 * 175 * Inside the device directory a sub-directory 'xgmi_hive_info' is 176 * created which contains the hive ID and the list of nodes. 177 * 178 * The hive ID is stored in: 179 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id 180 * 181 * The node information is stored in numbered directories: 182 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id 183 * 184 * Each device has their own xgmi_hive_info direction with a mirror 185 * set of node sub-directories. 186 * 187 * The XGMI memory space is built by contiguously adding the power of 188 * two padded VRAM space from each node to each other. 189 * 190 */ 191 192 static struct attribute amdgpu_xgmi_hive_id = { 193 .name = "xgmi_hive_id", 194 .mode = S_IRUGO 195 }; 196 197 static struct attribute *amdgpu_xgmi_hive_attrs[] = { 198 &amdgpu_xgmi_hive_id, 199 NULL 200 }; 201 ATTRIBUTE_GROUPS(amdgpu_xgmi_hive); 202 203 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj, 204 struct attribute *attr, char *buf) 205 { 206 struct amdgpu_hive_info *hive = container_of( 207 kobj, struct amdgpu_hive_info, kobj); 208 209 if (attr == &amdgpu_xgmi_hive_id) 210 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id); 211 212 return 0; 213 } 214 215 static void amdgpu_xgmi_hive_release(struct kobject *kobj) 216 { 217 struct amdgpu_hive_info *hive = container_of( 218 kobj, struct amdgpu_hive_info, kobj); 219 220 amdgpu_reset_put_reset_domain(hive->reset_domain); 221 hive->reset_domain = NULL; 222 223 mutex_destroy(&hive->hive_lock); 224 kfree(hive); 225 } 226 227 static const struct sysfs_ops amdgpu_xgmi_hive_ops = { 228 .show = amdgpu_xgmi_show_attrs, 229 }; 230 231 struct kobj_type amdgpu_xgmi_hive_type = { 232 .release = amdgpu_xgmi_hive_release, 233 .sysfs_ops = &amdgpu_xgmi_hive_ops, 234 .default_groups = amdgpu_xgmi_hive_groups, 235 }; 236 237 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev, 238 struct device_attribute *attr, 239 char *buf) 240 { 241 struct drm_device *ddev = dev_get_drvdata(dev); 242 struct amdgpu_device *adev = drm_to_adev(ddev); 243 244 return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id); 245 246 } 247 248 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801) 249 static ssize_t amdgpu_xgmi_show_error(struct device *dev, 250 struct device_attribute *attr, 251 char *buf) 252 { 253 struct drm_device *ddev = dev_get_drvdata(dev); 254 struct amdgpu_device *adev = drm_to_adev(ddev); 255 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in; 256 uint64_t fica_out; 257 unsigned int error_count = 0; 258 259 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200); 260 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208); 261 262 if ((!adev->df.funcs) || 263 (!adev->df.funcs->get_fica) || 264 (!adev->df.funcs->set_fica)) 265 return -EINVAL; 266 267 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in); 268 if (fica_out != 0x1f) 269 pr_err("xGMI error counters not enabled!\n"); 270 271 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in); 272 273 if ((fica_out & 0xffff) == 2) 274 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63); 275 276 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0); 277 278 return sysfs_emit(buf, "%u\n", error_count); 279 } 280 281 282 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL); 283 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL); 284 285 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev, 286 struct amdgpu_hive_info *hive) 287 { 288 int ret = 0; 289 char node[10] = { 0 }; 290 291 /* Create xgmi device id file */ 292 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id); 293 if (ret) { 294 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n"); 295 return ret; 296 } 297 298 /* Create xgmi error file */ 299 ret = device_create_file(adev->dev, &dev_attr_xgmi_error); 300 if (ret) 301 pr_err("failed to create xgmi_error\n"); 302 303 304 /* Create sysfs link to hive info folder on the first device */ 305 if (hive->kobj.parent != (&adev->dev->kobj)) { 306 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj, 307 "xgmi_hive_info"); 308 if (ret) { 309 dev_err(adev->dev, "XGMI: Failed to create link to hive info"); 310 goto remove_file; 311 } 312 } 313 314 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 315 /* Create sysfs link form the hive folder to yourself */ 316 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node); 317 if (ret) { 318 dev_err(adev->dev, "XGMI: Failed to create link from hive info"); 319 goto remove_link; 320 } 321 322 goto success; 323 324 325 remove_link: 326 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique); 327 328 remove_file: 329 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 330 331 success: 332 return ret; 333 } 334 335 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev, 336 struct amdgpu_hive_info *hive) 337 { 338 char node[10]; 339 memset(node, 0, sizeof(node)); 340 341 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 342 device_remove_file(adev->dev, &dev_attr_xgmi_error); 343 344 if (hive->kobj.parent != (&adev->dev->kobj)) 345 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info"); 346 347 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 348 sysfs_remove_link(&hive->kobj, node); 349 350 } 351 352 353 354 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev) 355 { 356 struct amdgpu_hive_info *hive = NULL; 357 int ret; 358 359 if (!adev->gmc.xgmi.hive_id) 360 return NULL; 361 362 if (adev->hive) { 363 kobject_get(&adev->hive->kobj); 364 return adev->hive; 365 } 366 367 mutex_lock(&xgmi_mutex); 368 369 list_for_each_entry(hive, &xgmi_hive_list, node) { 370 if (hive->hive_id == adev->gmc.xgmi.hive_id) 371 goto pro_end; 372 } 373 374 hive = kzalloc(sizeof(*hive), GFP_KERNEL); 375 if (!hive) { 376 dev_err(adev->dev, "XGMI: allocation failed\n"); 377 hive = NULL; 378 goto pro_end; 379 } 380 381 /* initialize new hive if not exist */ 382 ret = kobject_init_and_add(&hive->kobj, 383 &amdgpu_xgmi_hive_type, 384 &adev->dev->kobj, 385 "%s", "xgmi_hive_info"); 386 if (ret) { 387 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n"); 388 kobject_put(&hive->kobj); 389 hive = NULL; 390 goto pro_end; 391 } 392 393 /** 394 * Only init hive->reset_domain for none SRIOV configuration. For SRIOV, 395 * Host driver decide how to reset the GPU either through FLR or chain reset. 396 * Guest side will get individual notifications from the host for the FLR 397 * if necessary. 398 */ 399 if (!amdgpu_sriov_vf(adev)) { 400 /** 401 * Avoid recreating reset domain when hive is reconstructed for the case 402 * of reset the devices in the XGMI hive during probe for passthrough GPU 403 * See https://www.spinics.net/lists/amd-gfx/msg58836.html 404 */ 405 if (adev->reset_domain->type != XGMI_HIVE) { 406 hive->reset_domain = 407 amdgpu_reset_create_reset_domain(XGMI_HIVE, "amdgpu-reset-hive"); 408 if (!hive->reset_domain) { 409 dev_err(adev->dev, "XGMI: failed initializing reset domain for xgmi hive\n"); 410 ret = -ENOMEM; 411 kobject_put(&hive->kobj); 412 hive = NULL; 413 goto pro_end; 414 } 415 } else { 416 amdgpu_reset_get_reset_domain(adev->reset_domain); 417 hive->reset_domain = adev->reset_domain; 418 } 419 } 420 421 hive->hive_id = adev->gmc.xgmi.hive_id; 422 INIT_LIST_HEAD(&hive->device_list); 423 INIT_LIST_HEAD(&hive->node); 424 mutex_init(&hive->hive_lock); 425 atomic_set(&hive->number_devices, 0); 426 task_barrier_init(&hive->tb); 427 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN; 428 hive->hi_req_gpu = NULL; 429 430 /* 431 * hive pstate on boot is high in vega20 so we have to go to low 432 * pstate on after boot. 433 */ 434 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE; 435 list_add_tail(&hive->node, &xgmi_hive_list); 436 437 pro_end: 438 if (hive) 439 kobject_get(&hive->kobj); 440 mutex_unlock(&xgmi_mutex); 441 return hive; 442 } 443 444 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive) 445 { 446 if (hive) 447 kobject_put(&hive->kobj); 448 } 449 450 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate) 451 { 452 int ret = 0; 453 struct amdgpu_hive_info *hive; 454 struct amdgpu_device *request_adev; 455 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20; 456 bool init_low; 457 458 hive = amdgpu_get_xgmi_hive(adev); 459 if (!hive) 460 return 0; 461 462 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev; 463 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN; 464 amdgpu_put_xgmi_hive(hive); 465 /* fw bug so temporarily disable pstate switching */ 466 return 0; 467 468 if (!hive || adev->asic_type != CHIP_VEGA20) 469 return 0; 470 471 mutex_lock(&hive->hive_lock); 472 473 if (is_hi_req) 474 hive->hi_req_count++; 475 else 476 hive->hi_req_count--; 477 478 /* 479 * Vega20 only needs single peer to request pstate high for the hive to 480 * go high but all peers must request pstate low for the hive to go low 481 */ 482 if (hive->pstate == pstate || 483 (!is_hi_req && hive->hi_req_count && !init_low)) 484 goto out; 485 486 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate); 487 488 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate); 489 if (ret) { 490 dev_err(request_adev->dev, 491 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d", 492 request_adev->gmc.xgmi.node_id, 493 request_adev->gmc.xgmi.hive_id, ret); 494 goto out; 495 } 496 497 if (init_low) 498 hive->pstate = hive->hi_req_count ? 499 hive->pstate : AMDGPU_XGMI_PSTATE_MIN; 500 else { 501 hive->pstate = pstate; 502 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ? 503 adev : NULL; 504 } 505 out: 506 mutex_unlock(&hive->hive_lock); 507 return ret; 508 } 509 510 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev) 511 { 512 int ret; 513 514 if (amdgpu_sriov_vf(adev)) 515 return 0; 516 517 /* Each psp need to set the latest topology */ 518 ret = psp_xgmi_set_topology_info(&adev->psp, 519 atomic_read(&hive->number_devices), 520 &adev->psp.xgmi_context.top_info); 521 if (ret) 522 dev_err(adev->dev, 523 "XGMI: Set topology failure on device %llx, hive %llx, ret %d", 524 adev->gmc.xgmi.node_id, 525 adev->gmc.xgmi.hive_id, ret); 526 527 return ret; 528 } 529 530 531 /* 532 * NOTE psp_xgmi_node_info.num_hops layout is as follows: 533 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved) 534 * num_hops[5:3] = reserved 535 * num_hops[2:0] = number of hops 536 */ 537 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev, 538 struct amdgpu_device *peer_adev) 539 { 540 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 541 uint8_t num_hops_mask = 0x7; 542 int i; 543 544 for (i = 0 ; i < top->num_nodes; ++i) 545 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 546 return top->nodes[i].num_hops & num_hops_mask; 547 return -EINVAL; 548 } 549 550 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev, 551 struct amdgpu_device *peer_adev) 552 { 553 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 554 int i; 555 556 for (i = 0 ; i < top->num_nodes; ++i) 557 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 558 return top->nodes[i].num_links; 559 return -EINVAL; 560 } 561 562 /* 563 * Devices that support extended data require the entire hive to initialize with 564 * the shared memory buffer flag set. 565 * 566 * Hive locks and conditions apply - see amdgpu_xgmi_add_device 567 */ 568 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive, 569 bool set_extended_data) 570 { 571 struct amdgpu_device *tmp_adev; 572 int ret; 573 574 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 575 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false); 576 if (ret) { 577 dev_err(tmp_adev->dev, 578 "XGMI: Failed to initialize xgmi session for data partition %i\n", 579 set_extended_data); 580 return ret; 581 } 582 583 } 584 585 return 0; 586 } 587 588 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 589 { 590 struct psp_xgmi_topology_info *top_info; 591 struct amdgpu_hive_info *hive; 592 struct amdgpu_xgmi *entry; 593 struct amdgpu_device *tmp_adev = NULL; 594 595 int count = 0, ret = 0; 596 597 if (!adev->gmc.xgmi.supported) 598 return 0; 599 600 if (!adev->gmc.xgmi.pending_reset && 601 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 602 ret = psp_xgmi_initialize(&adev->psp, false, true); 603 if (ret) { 604 dev_err(adev->dev, 605 "XGMI: Failed to initialize xgmi session\n"); 606 return ret; 607 } 608 609 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 610 if (ret) { 611 dev_err(adev->dev, 612 "XGMI: Failed to get hive id\n"); 613 return ret; 614 } 615 616 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 617 if (ret) { 618 dev_err(adev->dev, 619 "XGMI: Failed to get node id\n"); 620 return ret; 621 } 622 } else { 623 adev->gmc.xgmi.hive_id = 16; 624 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 625 } 626 627 hive = amdgpu_get_xgmi_hive(adev); 628 if (!hive) { 629 ret = -EINVAL; 630 dev_err(adev->dev, 631 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 632 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 633 goto exit; 634 } 635 mutex_lock(&hive->hive_lock); 636 637 top_info = &adev->psp.xgmi_context.top_info; 638 639 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 640 list_for_each_entry(entry, &hive->device_list, head) 641 top_info->nodes[count++].node_id = entry->node_id; 642 top_info->num_nodes = count; 643 atomic_set(&hive->number_devices, count); 644 645 task_barrier_add_task(&hive->tb); 646 647 if (!adev->gmc.xgmi.pending_reset && 648 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 649 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 650 /* update node list for other device in the hive */ 651 if (tmp_adev != adev) { 652 top_info = &tmp_adev->psp.xgmi_context.top_info; 653 top_info->nodes[count - 1].node_id = 654 adev->gmc.xgmi.node_id; 655 top_info->num_nodes = count; 656 } 657 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 658 if (ret) 659 goto exit_unlock; 660 } 661 662 /* get latest topology info for each device from psp */ 663 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 664 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 665 &tmp_adev->psp.xgmi_context.top_info, false); 666 if (ret) { 667 dev_err(tmp_adev->dev, 668 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 669 tmp_adev->gmc.xgmi.node_id, 670 tmp_adev->gmc.xgmi.hive_id, ret); 671 /* To do : continue with some node failed or disable the whole hive */ 672 goto exit_unlock; 673 } 674 } 675 676 /* get topology again for hives that support extended data */ 677 if (adev->psp.xgmi_context.supports_extended_data) { 678 679 /* initialize the hive to get extended data. */ 680 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true); 681 if (ret) 682 goto exit_unlock; 683 684 /* get the extended data. */ 685 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 686 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 687 &tmp_adev->psp.xgmi_context.top_info, true); 688 if (ret) { 689 dev_err(tmp_adev->dev, 690 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d", 691 tmp_adev->gmc.xgmi.node_id, 692 tmp_adev->gmc.xgmi.hive_id, ret); 693 goto exit_unlock; 694 } 695 } 696 697 /* initialize the hive to get non-extended data for the next round. */ 698 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false); 699 if (ret) 700 goto exit_unlock; 701 702 } 703 } 704 705 if (!ret && !adev->gmc.xgmi.pending_reset) 706 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 707 708 exit_unlock: 709 mutex_unlock(&hive->hive_lock); 710 exit: 711 if (!ret) { 712 adev->hive = hive; 713 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 714 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 715 } else { 716 amdgpu_put_xgmi_hive(hive); 717 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 718 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 719 ret); 720 } 721 722 return ret; 723 } 724 725 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 726 { 727 struct amdgpu_hive_info *hive = adev->hive; 728 729 if (!adev->gmc.xgmi.supported) 730 return -EINVAL; 731 732 if (!hive) 733 return -EINVAL; 734 735 mutex_lock(&hive->hive_lock); 736 task_barrier_rem_task(&hive->tb); 737 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 738 if (hive->hi_req_gpu == adev) 739 hive->hi_req_gpu = NULL; 740 list_del(&adev->gmc.xgmi.head); 741 mutex_unlock(&hive->hive_lock); 742 743 amdgpu_put_xgmi_hive(hive); 744 adev->hive = NULL; 745 746 if (atomic_dec_return(&hive->number_devices) == 0) { 747 /* Remove the hive from global hive list */ 748 mutex_lock(&xgmi_mutex); 749 list_del(&hive->node); 750 mutex_unlock(&xgmi_mutex); 751 752 amdgpu_put_xgmi_hive(hive); 753 } 754 755 return 0; 756 } 757 758 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 759 { 760 if (!adev->gmc.xgmi.supported || 761 adev->gmc.xgmi.num_physical_nodes == 0) 762 return 0; 763 764 adev->gmc.xgmi.ras->ras_block.hw_ops->reset_ras_error_count(adev); 765 766 return amdgpu_ras_block_late_init(adev, ras_block); 767 } 768 769 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 770 uint64_t addr) 771 { 772 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 773 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 774 } 775 776 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 777 { 778 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 779 WREG32_PCIE(pcs_status_reg, 0); 780 } 781 782 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 783 { 784 uint32_t i; 785 786 switch (adev->asic_type) { 787 case CHIP_ARCTURUS: 788 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 789 pcs_clear_status(adev, 790 xgmi_pcs_err_status_reg_arct[i]); 791 break; 792 case CHIP_VEGA20: 793 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 794 pcs_clear_status(adev, 795 xgmi_pcs_err_status_reg_vg20[i]); 796 break; 797 case CHIP_ALDEBARAN: 798 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) 799 pcs_clear_status(adev, 800 xgmi3x16_pcs_err_status_reg_aldebaran[i]); 801 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) 802 pcs_clear_status(adev, 803 walf_pcs_err_status_reg_aldebaran[i]); 804 break; 805 default: 806 break; 807 } 808 } 809 810 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 811 uint32_t value, 812 uint32_t *ue_count, 813 uint32_t *ce_count, 814 bool is_xgmi_pcs) 815 { 816 int i; 817 int ue_cnt; 818 819 if (is_xgmi_pcs) { 820 /* query xgmi pcs error status, 821 * only ue is supported */ 822 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_ras_fields); i ++) { 823 ue_cnt = (value & 824 xgmi_pcs_ras_fields[i].pcs_err_mask) >> 825 xgmi_pcs_ras_fields[i].pcs_err_shift; 826 if (ue_cnt) { 827 dev_info(adev->dev, "%s detected\n", 828 xgmi_pcs_ras_fields[i].err_name); 829 *ue_count += ue_cnt; 830 } 831 } 832 } else { 833 /* query wafl pcs error status, 834 * only ue is supported */ 835 for (i = 0; i < ARRAY_SIZE(wafl_pcs_ras_fields); i++) { 836 ue_cnt = (value & 837 wafl_pcs_ras_fields[i].pcs_err_mask) >> 838 wafl_pcs_ras_fields[i].pcs_err_shift; 839 if (ue_cnt) { 840 dev_info(adev->dev, "%s detected\n", 841 wafl_pcs_ras_fields[i].err_name); 842 *ue_count += ue_cnt; 843 } 844 } 845 } 846 847 return 0; 848 } 849 850 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 851 void *ras_error_status) 852 { 853 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 854 int i; 855 uint32_t data; 856 uint32_t ue_cnt = 0, ce_cnt = 0; 857 858 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 859 return ; 860 861 err_data->ue_count = 0; 862 err_data->ce_count = 0; 863 864 switch (adev->asic_type) { 865 case CHIP_ARCTURUS: 866 /* check xgmi pcs error */ 867 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 868 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 869 if (data) 870 amdgpu_xgmi_query_pcs_error_status(adev, 871 data, &ue_cnt, &ce_cnt, true); 872 } 873 /* check wafl pcs error */ 874 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 875 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 876 if (data) 877 amdgpu_xgmi_query_pcs_error_status(adev, 878 data, &ue_cnt, &ce_cnt, false); 879 } 880 break; 881 case CHIP_VEGA20: 882 /* check xgmi pcs error */ 883 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 884 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 885 if (data) 886 amdgpu_xgmi_query_pcs_error_status(adev, 887 data, &ue_cnt, &ce_cnt, true); 888 } 889 /* check wafl pcs error */ 890 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 891 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 892 if (data) 893 amdgpu_xgmi_query_pcs_error_status(adev, 894 data, &ue_cnt, &ce_cnt, false); 895 } 896 break; 897 case CHIP_ALDEBARAN: 898 /* check xgmi3x16 pcs error */ 899 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) { 900 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]); 901 if (data) 902 amdgpu_xgmi_query_pcs_error_status(adev, 903 data, &ue_cnt, &ce_cnt, true); 904 } 905 /* check wafl pcs error */ 906 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) { 907 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]); 908 if (data) 909 amdgpu_xgmi_query_pcs_error_status(adev, 910 data, &ue_cnt, &ce_cnt, false); 911 } 912 break; 913 default: 914 dev_warn(adev->dev, "XGMI RAS error query not supported"); 915 break; 916 } 917 918 adev->gmc.xgmi.ras->ras_block.hw_ops->reset_ras_error_count(adev); 919 920 err_data->ue_count += ue_cnt; 921 err_data->ce_count += ce_cnt; 922 } 923 924 /* Trigger XGMI/WAFL error */ 925 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev, void *inject_if) 926 { 927 int ret = 0; 928 struct ta_ras_trigger_error_input *block_info = 929 (struct ta_ras_trigger_error_input *)inject_if; 930 931 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) 932 dev_warn(adev->dev, "Failed to disallow df cstate"); 933 934 if (amdgpu_dpm_allow_xgmi_power_down(adev, false)) 935 dev_warn(adev->dev, "Failed to disallow XGMI power down"); 936 937 ret = psp_ras_trigger_error(&adev->psp, block_info); 938 939 if (amdgpu_ras_intr_triggered()) 940 return ret; 941 942 if (amdgpu_dpm_allow_xgmi_power_down(adev, true)) 943 dev_warn(adev->dev, "Failed to allow XGMI power down"); 944 945 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW)) 946 dev_warn(adev->dev, "Failed to allow df cstate"); 947 948 return ret; 949 } 950 951 struct amdgpu_ras_block_hw_ops xgmi_ras_hw_ops = { 952 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 953 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 954 .ras_error_inject = amdgpu_ras_error_inject_xgmi, 955 }; 956 957 struct amdgpu_xgmi_ras xgmi_ras = { 958 .ras_block = { 959 .ras_comm = { 960 .name = "xgmi_wafl", 961 .block = AMDGPU_RAS_BLOCK__XGMI_WAFL, 962 .type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE, 963 }, 964 .hw_ops = &xgmi_ras_hw_ops, 965 .ras_late_init = amdgpu_xgmi_ras_late_init, 966 }, 967 }; 968