1 /* 2 * Copyright 2019 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/sort.h> 25 #include "amdgpu.h" 26 #include "umc_v6_7.h" 27 #include "amdgpu_ras_mgr.h" 28 #define MAX_UMC_POISON_POLLING_TIME_SYNC 20 //ms 29 30 #define MAX_UMC_HASH_STRING_SIZE 256 31 32 static int amdgpu_umc_convert_error_address(struct amdgpu_device *adev, 33 struct ras_err_data *err_data, uint64_t err_addr, 34 uint32_t ch_inst, uint32_t umc_inst) 35 { 36 switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) { 37 case IP_VERSION(6, 7, 0): 38 umc_v6_7_convert_error_address(adev, 39 err_data, err_addr, ch_inst, umc_inst); 40 break; 41 default: 42 dev_warn(adev->dev, 43 "UMC address to Physical address translation is not supported\n"); 44 return AMDGPU_RAS_FAIL; 45 } 46 47 return AMDGPU_RAS_SUCCESS; 48 } 49 50 int amdgpu_umc_page_retirement_mca(struct amdgpu_device *adev, 51 uint64_t err_addr, uint32_t ch_inst, uint32_t umc_inst) 52 { 53 struct ras_err_data err_data; 54 int ret; 55 56 ret = amdgpu_ras_error_data_init(&err_data); 57 if (ret) 58 return ret; 59 60 err_data.err_addr = 61 kcalloc(adev->umc.max_ras_err_cnt_per_query, 62 sizeof(struct eeprom_table_record), GFP_KERNEL); 63 if (!err_data.err_addr) { 64 dev_warn(adev->dev, 65 "Failed to alloc memory for umc error record in MCA notifier!\n"); 66 ret = AMDGPU_RAS_FAIL; 67 goto out_fini_err_data; 68 } 69 70 err_data.err_addr_len = adev->umc.max_ras_err_cnt_per_query; 71 72 /* 73 * Translate UMC channel address to Physical address 74 */ 75 ret = amdgpu_umc_convert_error_address(adev, &err_data, err_addr, 76 ch_inst, umc_inst); 77 if (ret) 78 goto out_free_err_addr; 79 80 if (amdgpu_bad_page_threshold != 0) { 81 amdgpu_ras_add_bad_pages(adev, err_data.err_addr, 82 err_data.err_addr_cnt, false); 83 amdgpu_ras_save_bad_pages(adev, NULL); 84 } 85 86 out_free_err_addr: 87 kfree(err_data.err_addr); 88 89 out_fini_err_data: 90 amdgpu_ras_error_data_fini(&err_data); 91 92 return ret; 93 } 94 95 void amdgpu_umc_handle_bad_pages(struct amdgpu_device *adev, 96 void *ras_error_status) 97 { 98 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 99 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 100 struct amdgpu_ras_eeprom_control *control = &con->eeprom_control; 101 unsigned int error_query_mode; 102 int ret = 0; 103 unsigned long err_count; 104 105 amdgpu_ras_get_error_query_mode(adev, &error_query_mode); 106 107 err_data->err_addr = 108 kcalloc(adev->umc.max_ras_err_cnt_per_query, 109 sizeof(struct eeprom_table_record), GFP_KERNEL); 110 111 /* still call query_ras_error_address to clear error status 112 * even NOMEM error is encountered 113 */ 114 if (!err_data->err_addr) 115 dev_warn(adev->dev, 116 "Failed to alloc memory for umc error address record!\n"); 117 else 118 err_data->err_addr_len = adev->umc.max_ras_err_cnt_per_query; 119 120 mutex_lock(&con->page_retirement_lock); 121 if (!amdgpu_ras_smu_eeprom_supported(adev)) { 122 ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(con->umc_ecc)); 123 if (ret == -EOPNOTSUPP && 124 error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY) { 125 if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops && 126 adev->umc.ras->ras_block.hw_ops->query_ras_error_count) 127 adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, 128 ras_error_status); 129 130 if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops && 131 adev->umc.ras->ras_block.hw_ops->query_ras_error_address && 132 adev->umc.max_ras_err_cnt_per_query) { 133 err_data->err_addr = 134 kcalloc(adev->umc.max_ras_err_cnt_per_query, 135 sizeof(struct eeprom_table_record), GFP_KERNEL); 136 137 /* still call query_ras_error_address to clear error status 138 * even NOMEM error is encountered 139 */ 140 if (!err_data->err_addr) 141 dev_warn(adev->dev, 142 "Failed to alloc memory for umc error address record!\n"); 143 else 144 err_data->err_addr_len = 145 adev->umc.max_ras_err_cnt_per_query; 146 147 /* umc query_ras_error_address is also responsible for clearing 148 * error status 149 */ 150 adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, 151 ras_error_status); 152 } 153 } else if (error_query_mode == AMDGPU_RAS_FIRMWARE_ERROR_QUERY || 154 (!ret && error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY)) { 155 if (adev->umc.ras && 156 adev->umc.ras->ecc_info_query_ras_error_count) 157 adev->umc.ras->ecc_info_query_ras_error_count(adev, 158 ras_error_status); 159 160 if (adev->umc.ras && 161 adev->umc.ras->ecc_info_query_ras_error_address && 162 adev->umc.max_ras_err_cnt_per_query) { 163 err_data->err_addr = 164 kcalloc(adev->umc.max_ras_err_cnt_per_query, 165 sizeof(struct eeprom_table_record), GFP_KERNEL); 166 167 /* still call query_ras_error_address to clear error status 168 * even NOMEM error is encountered 169 */ 170 if (!err_data->err_addr) 171 dev_warn(adev->dev, 172 "Failed to alloc memory for umc error address record!\n"); 173 else 174 err_data->err_addr_len = 175 adev->umc.max_ras_err_cnt_per_query; 176 177 /* umc query_ras_error_address is also responsible for clearing 178 * error status 179 */ 180 adev->umc.ras->ecc_info_query_ras_error_address(adev, 181 ras_error_status); 182 } 183 } 184 } else { 185 if (!amdgpu_ras_eeprom_update_record_num(control)) { 186 err_data->err_addr_cnt = err_data->de_count = 187 control->ras_num_recs - control->ras_num_recs_old; 188 amdgpu_ras_eeprom_read_idx(control, err_data->err_addr, 189 control->ras_num_recs_old, err_data->de_count); 190 } 191 } 192 193 /* only uncorrectable error needs gpu reset */ 194 if (err_data->ue_count || err_data->de_count) { 195 err_count = err_data->ue_count + err_data->de_count; 196 if ((amdgpu_bad_page_threshold != 0) && 197 err_data->err_addr_cnt) { 198 amdgpu_ras_add_bad_pages(adev, err_data->err_addr, 199 err_data->err_addr_cnt, amdgpu_ras_smu_eeprom_supported(adev)); 200 amdgpu_ras_save_bad_pages(adev, &err_count); 201 202 amdgpu_dpm_send_hbm_bad_pages_num(adev, 203 con->eeprom_control.ras_num_bad_pages); 204 205 if (con->update_channel_flag == true) { 206 amdgpu_dpm_send_hbm_bad_channel_flag(adev, con->eeprom_control.bad_channel_bitmap); 207 con->update_channel_flag = false; 208 } 209 } 210 } 211 212 kfree(err_data->err_addr); 213 err_data->err_addr = NULL; 214 215 mutex_unlock(&con->page_retirement_lock); 216 } 217 218 static int amdgpu_umc_do_page_retirement(struct amdgpu_device *adev, 219 void *ras_error_status, 220 struct amdgpu_iv_entry *entry, 221 uint32_t reset) 222 { 223 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 224 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 225 226 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev); 227 amdgpu_umc_handle_bad_pages(adev, ras_error_status); 228 229 if ((err_data->ue_count || err_data->de_count) && 230 (reset || amdgpu_ras_is_rma(adev))) { 231 con->gpu_reset_flags |= reset; 232 amdgpu_ras_reset_gpu(adev); 233 } 234 235 return AMDGPU_RAS_SUCCESS; 236 } 237 238 int amdgpu_umc_pasid_poison_handler(struct amdgpu_device *adev, 239 enum amdgpu_ras_block block, uint16_t pasid, 240 pasid_notify pasid_fn, void *data, uint32_t reset) 241 { 242 int ret = AMDGPU_RAS_SUCCESS; 243 244 if (adev->gmc.xgmi.connected_to_cpu || 245 adev->gmc.is_app_apu) { 246 if (reset) { 247 /* MCA poison handler is only responsible for GPU reset, 248 * let MCA notifier do page retirement. 249 */ 250 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev); 251 amdgpu_ras_reset_gpu(adev); 252 } 253 return ret; 254 } 255 256 if (!amdgpu_sriov_vf(adev)) { 257 if (amdgpu_ip_version(adev, UMC_HWIP, 0) < IP_VERSION(12, 0, 0)) { 258 struct ras_err_data err_data; 259 struct ras_common_if head = { 260 .block = AMDGPU_RAS_BLOCK__UMC, 261 }; 262 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head); 263 264 ret = amdgpu_ras_error_data_init(&err_data); 265 if (ret) 266 return ret; 267 268 ret = amdgpu_umc_do_page_retirement(adev, &err_data, NULL, reset); 269 270 if (ret == AMDGPU_RAS_SUCCESS && obj) { 271 obj->err_data.ue_count += err_data.ue_count; 272 obj->err_data.ce_count += err_data.ce_count; 273 obj->err_data.de_count += err_data.de_count; 274 } 275 276 amdgpu_ras_error_data_fini(&err_data); 277 } else if (amdgpu_uniras_enabled(adev)) { 278 struct ras_ih_info ih_info = {0}; 279 280 ih_info.block = block; 281 ih_info.pasid = pasid; 282 ih_info.reset = reset; 283 ih_info.pasid_fn = pasid_fn; 284 ih_info.data = data; 285 amdgpu_ras_mgr_handle_consumer_interrupt(adev, &ih_info); 286 } else { 287 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 288 int ret; 289 290 ret = amdgpu_ras_put_poison_req(adev, 291 block, pasid, pasid_fn, data, reset); 292 if (!ret) { 293 atomic_inc(&con->page_retirement_req_cnt); 294 atomic_inc(&con->poison_consumption_count); 295 wake_up(&con->page_retirement_wq); 296 } 297 } 298 } else { 299 if (adev->virt.ops && adev->virt.ops->ras_poison_handler) 300 adev->virt.ops->ras_poison_handler(adev, block); 301 else 302 dev_warn(adev->dev, 303 "No ras_poison_handler interface in SRIOV!\n"); 304 } 305 306 return ret; 307 } 308 309 int amdgpu_umc_poison_handler(struct amdgpu_device *adev, 310 enum amdgpu_ras_block block, uint32_t reset) 311 { 312 return amdgpu_umc_pasid_poison_handler(adev, 313 block, 0, NULL, NULL, reset); 314 } 315 316 int amdgpu_umc_process_ras_data_cb(struct amdgpu_device *adev, 317 void *ras_error_status, 318 struct amdgpu_iv_entry *entry) 319 { 320 return amdgpu_umc_do_page_retirement(adev, ras_error_status, entry, 321 AMDGPU_RAS_GPU_RESET_MODE1_RESET); 322 } 323 324 int amdgpu_umc_ras_sw_init(struct amdgpu_device *adev) 325 { 326 int err; 327 struct amdgpu_umc_ras *ras; 328 329 if (!adev->umc.ras) 330 return 0; 331 332 ras = adev->umc.ras; 333 334 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 335 if (err) { 336 dev_err(adev->dev, "Failed to register umc ras block!\n"); 337 return err; 338 } 339 340 strcpy(adev->umc.ras->ras_block.ras_comm.name, "umc"); 341 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__UMC; 342 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 343 adev->umc.ras_if = &ras->ras_block.ras_comm; 344 345 if (!ras->ras_block.ras_late_init) 346 ras->ras_block.ras_late_init = amdgpu_umc_ras_late_init; 347 348 if (!ras->ras_block.ras_cb) 349 ras->ras_block.ras_cb = amdgpu_umc_process_ras_data_cb; 350 351 return 0; 352 } 353 354 int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 355 { 356 int r; 357 358 r = amdgpu_ras_block_late_init(adev, ras_block); 359 if (r) 360 return r; 361 362 if (amdgpu_sriov_vf(adev)) 363 return r; 364 365 if (amdgpu_ras_is_supported(adev, ras_block->block)) { 366 r = amdgpu_irq_get(adev, &adev->gmc.ecc_irq, 0); 367 if (r) 368 goto late_fini; 369 } 370 371 /* ras init of specific umc version */ 372 if (adev->umc.ras && 373 adev->umc.ras->err_cnt_init) 374 adev->umc.ras->err_cnt_init(adev); 375 376 return 0; 377 378 late_fini: 379 amdgpu_ras_block_late_fini(adev, ras_block); 380 return r; 381 } 382 383 int amdgpu_umc_process_ecc_irq(struct amdgpu_device *adev, 384 struct amdgpu_irq_src *source, 385 struct amdgpu_iv_entry *entry) 386 { 387 struct ras_common_if *ras_if = adev->umc.ras_if; 388 struct ras_dispatch_if ih_data = { 389 .entry = entry, 390 }; 391 392 if (!ras_if) 393 return 0; 394 395 ih_data.head = *ras_if; 396 397 amdgpu_ras_interrupt_dispatch(adev, &ih_data); 398 return 0; 399 } 400 401 int amdgpu_umc_fill_error_record(struct ras_err_data *err_data, 402 uint64_t err_addr, 403 uint64_t retired_page, 404 uint32_t channel_index, 405 uint32_t umc_inst) 406 { 407 struct eeprom_table_record *err_rec; 408 409 if (!err_data || 410 !err_data->err_addr || 411 (err_data->err_addr_cnt >= err_data->err_addr_len)) 412 return -EINVAL; 413 414 err_rec = &err_data->err_addr[err_data->err_addr_cnt]; 415 416 err_rec->address = err_addr; 417 /* page frame address is saved */ 418 err_rec->retired_page = retired_page >> AMDGPU_GPU_PAGE_SHIFT; 419 err_rec->ts = (uint64_t)ktime_get_real_seconds(); 420 err_rec->err_type = AMDGPU_RAS_EEPROM_ERR_NON_RECOVERABLE; 421 err_rec->cu = 0; 422 err_rec->mem_channel = channel_index; 423 err_rec->mcumc_id = umc_inst; 424 425 err_data->err_addr_cnt++; 426 427 return 0; 428 } 429 430 static int amdgpu_umc_loop_all_aid(struct amdgpu_device *adev, umc_func func, 431 void *data) 432 { 433 uint32_t umc_node_inst; 434 uint32_t node_inst; 435 uint32_t umc_inst; 436 uint32_t ch_inst; 437 int ret; 438 439 /* 440 * This loop is done based on the following - 441 * umc.active mask = mask of active umc instances across all nodes 442 * umc.umc_inst_num = maximum number of umc instancess per node 443 * umc.node_inst_num = maximum number of node instances 444 * Channel instances are not assumed to be harvested. 445 */ 446 dev_dbg(adev->dev, "active umcs :%lx umc_inst per node: %d", 447 adev->umc.active_mask, adev->umc.umc_inst_num); 448 for_each_set_bit(umc_node_inst, &(adev->umc.active_mask), 449 adev->umc.node_inst_num * adev->umc.umc_inst_num) { 450 node_inst = umc_node_inst / adev->umc.umc_inst_num; 451 umc_inst = umc_node_inst % adev->umc.umc_inst_num; 452 LOOP_UMC_CH_INST(ch_inst) { 453 dev_dbg(adev->dev, 454 "node_inst :%d umc_inst: %d ch_inst: %d", 455 node_inst, umc_inst, ch_inst); 456 ret = func(adev, node_inst, umc_inst, ch_inst, data); 457 if (ret) { 458 dev_err(adev->dev, 459 "Node %d umc %d ch %d func returns %d\n", 460 node_inst, umc_inst, ch_inst, ret); 461 return ret; 462 } 463 } 464 } 465 466 return 0; 467 } 468 469 int amdgpu_umc_loop_channels(struct amdgpu_device *adev, 470 umc_func func, void *data) 471 { 472 uint32_t node_inst = 0; 473 uint32_t umc_inst = 0; 474 uint32_t ch_inst = 0; 475 int ret = 0; 476 477 if (adev->aid_mask) 478 return amdgpu_umc_loop_all_aid(adev, func, data); 479 480 if (adev->umc.node_inst_num) { 481 LOOP_UMC_EACH_NODE_INST_AND_CH(node_inst, umc_inst, ch_inst) { 482 ret = func(adev, node_inst, umc_inst, ch_inst, data); 483 if (ret) { 484 dev_err(adev->dev, "Node %d umc %d ch %d func returns %d\n", 485 node_inst, umc_inst, ch_inst, ret); 486 return ret; 487 } 488 } 489 } else { 490 LOOP_UMC_INST_AND_CH(umc_inst, ch_inst) { 491 ret = func(adev, 0, umc_inst, ch_inst, data); 492 if (ret) { 493 dev_err(adev->dev, "Umc %d ch %d func returns %d\n", 494 umc_inst, ch_inst, ret); 495 return ret; 496 } 497 } 498 } 499 500 return 0; 501 } 502 503 int amdgpu_umc_update_ecc_status(struct amdgpu_device *adev, 504 uint64_t status, uint64_t ipid, uint64_t addr) 505 { 506 if (adev->umc.ras->update_ecc_status) 507 return adev->umc.ras->update_ecc_status(adev, 508 status, ipid, addr); 509 return 0; 510 } 511 512 int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev, 513 struct radix_tree_root *ecc_tree, struct ras_ecc_err *ecc_err) 514 { 515 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 516 struct ras_ecc_log_info *ecc_log; 517 int ret; 518 519 ecc_log = &con->umc_ecc_log; 520 521 mutex_lock(&ecc_log->lock); 522 ret = radix_tree_insert(ecc_tree, ecc_err->pa_pfn, ecc_err); 523 if (!ret) 524 radix_tree_tag_set(ecc_tree, 525 ecc_err->pa_pfn, UMC_ECC_NEW_DETECTED_TAG); 526 mutex_unlock(&ecc_log->lock); 527 528 return ret; 529 } 530 531 int amdgpu_umc_pages_in_a_row(struct amdgpu_device *adev, 532 struct ras_err_data *err_data, uint64_t pa_addr) 533 { 534 struct ta_ras_query_address_output addr_out; 535 536 /* reinit err_data */ 537 err_data->err_addr_cnt = 0; 538 err_data->err_addr_len = adev->umc.retire_unit; 539 540 addr_out.pa.pa = pa_addr; 541 if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) 542 return adev->umc.ras->convert_ras_err_addr(adev, err_data, NULL, 543 &addr_out, false); 544 else 545 return -EINVAL; 546 } 547 548 int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, 549 uint64_t pa_addr, uint64_t *pfns, int len) 550 { 551 int i, ret; 552 struct ras_err_data err_data; 553 554 err_data.err_addr = kcalloc(adev->umc.retire_unit, 555 sizeof(struct eeprom_table_record), GFP_KERNEL); 556 if (!err_data.err_addr) { 557 dev_warn(adev->dev, "Failed to alloc memory in bad page lookup!\n"); 558 return 0; 559 } 560 561 ret = amdgpu_umc_pages_in_a_row(adev, &err_data, pa_addr); 562 if (ret) 563 goto out; 564 565 for (i = 0; i < adev->umc.retire_unit; i++) { 566 if (i >= len) 567 goto out; 568 569 pfns[i] = err_data.err_addr[i].retired_page; 570 } 571 ret = i; 572 adev->umc.err_addr_cnt = err_data.err_addr_cnt; 573 574 out: 575 kfree(err_data.err_addr); 576 return ret; 577 } 578 579 int amdgpu_umc_mca_to_addr(struct amdgpu_device *adev, 580 uint64_t err_addr, uint32_t ch, uint32_t umc, 581 uint32_t node, uint32_t socket, 582 struct ta_ras_query_address_output *addr_out, bool dump_addr) 583 { 584 struct ta_ras_query_address_input addr_in; 585 int ret; 586 587 memset(&addr_in, 0, sizeof(addr_in)); 588 addr_in.ma.err_addr = err_addr; 589 addr_in.ma.ch_inst = ch; 590 addr_in.ma.umc_inst = umc; 591 addr_in.ma.node_inst = node; 592 addr_in.ma.socket_id = socket; 593 594 if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) { 595 ret = adev->umc.ras->convert_ras_err_addr(adev, NULL, &addr_in, 596 addr_out, dump_addr); 597 if (ret) 598 return ret; 599 } else { 600 return 0; 601 } 602 603 return 0; 604 } 605 606 int amdgpu_umc_pa2mca(struct amdgpu_device *adev, 607 uint64_t pa, uint64_t *mca, enum amdgpu_memory_partition nps) 608 { 609 struct ta_ras_query_address_input addr_in; 610 struct ta_ras_query_address_output addr_out; 611 int ret; 612 613 /* nps: the pa belongs to */ 614 addr_in.pa.pa = pa | ((uint64_t)nps << 58); 615 addr_in.addr_type = TA_RAS_PA_TO_MCA; 616 ret = psp_ras_query_address(&adev->psp, &addr_in, &addr_out); 617 if (ret) { 618 dev_warn(adev->dev, "Failed to query RAS MCA address for 0x%llx", 619 pa); 620 621 return ret; 622 } 623 624 *mca = addr_out.ma.err_addr; 625 626 return 0; 627 } 628