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 kzalloc_objs(struct eeprom_table_record, 62 adev->umc.max_ras_err_cnt_per_query, 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 kzalloc_objs(struct eeprom_table_record, 109 adev->umc.max_ras_err_cnt_per_query, 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 kzalloc_objs(struct eeprom_table_record, 135 adev->umc.max_ras_err_cnt_per_query, 136 GFP_KERNEL); 137 138 /* still call query_ras_error_address to clear error status 139 * even NOMEM error is encountered 140 */ 141 if (!err_data->err_addr) 142 dev_warn(adev->dev, 143 "Failed to alloc memory for umc error address record!\n"); 144 else 145 err_data->err_addr_len = 146 adev->umc.max_ras_err_cnt_per_query; 147 148 /* umc query_ras_error_address is also responsible for clearing 149 * error status 150 */ 151 adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, 152 ras_error_status); 153 } 154 } else if (error_query_mode == AMDGPU_RAS_FIRMWARE_ERROR_QUERY || 155 (!ret && error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY)) { 156 if (adev->umc.ras && 157 adev->umc.ras->ecc_info_query_ras_error_count) 158 adev->umc.ras->ecc_info_query_ras_error_count(adev, 159 ras_error_status); 160 161 if (adev->umc.ras && 162 adev->umc.ras->ecc_info_query_ras_error_address && 163 adev->umc.max_ras_err_cnt_per_query) { 164 err_data->err_addr = 165 kzalloc_objs(struct eeprom_table_record, 166 adev->umc.max_ras_err_cnt_per_query, 167 GFP_KERNEL); 168 169 /* still call query_ras_error_address to clear error status 170 * even NOMEM error is encountered 171 */ 172 if (!err_data->err_addr) 173 dev_warn(adev->dev, 174 "Failed to alloc memory for umc error address record!\n"); 175 else 176 err_data->err_addr_len = 177 adev->umc.max_ras_err_cnt_per_query; 178 179 /* umc query_ras_error_address is also responsible for clearing 180 * error status 181 */ 182 adev->umc.ras->ecc_info_query_ras_error_address(adev, 183 ras_error_status); 184 } 185 } 186 } else { 187 if (!amdgpu_ras_eeprom_update_record_num(control)) { 188 err_data->err_addr_cnt = err_data->de_count = 189 control->ras_num_recs - control->ras_num_recs_old; 190 amdgpu_ras_eeprom_read_idx(control, err_data->err_addr, 191 control->ras_num_recs_old, err_data->de_count); 192 } 193 } 194 195 /* only uncorrectable error needs gpu reset */ 196 if (err_data->ue_count || err_data->de_count) { 197 err_count = err_data->ue_count + err_data->de_count; 198 if ((amdgpu_bad_page_threshold != 0) && 199 err_data->err_addr_cnt) { 200 amdgpu_ras_add_bad_pages(adev, err_data->err_addr, 201 err_data->err_addr_cnt, amdgpu_ras_smu_eeprom_supported(adev)); 202 amdgpu_ras_save_bad_pages(adev, &err_count); 203 204 amdgpu_dpm_send_hbm_bad_pages_num(adev, 205 con->eeprom_control.ras_num_bad_pages); 206 207 if (con->update_channel_flag == true) { 208 amdgpu_dpm_send_hbm_bad_channel_flag(adev, con->eeprom_control.bad_channel_bitmap); 209 con->update_channel_flag = false; 210 } 211 } 212 } 213 214 kfree(err_data->err_addr); 215 err_data->err_addr = NULL; 216 217 mutex_unlock(&con->page_retirement_lock); 218 } 219 220 static int amdgpu_umc_do_page_retirement(struct amdgpu_device *adev, 221 void *ras_error_status, 222 struct amdgpu_iv_entry *entry, 223 uint32_t reset) 224 { 225 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 226 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 227 228 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev); 229 amdgpu_umc_handle_bad_pages(adev, ras_error_status); 230 231 if ((err_data->ue_count || err_data->de_count) && 232 (reset || amdgpu_ras_is_rma(adev))) { 233 con->gpu_reset_flags |= reset; 234 amdgpu_ras_reset_gpu(adev); 235 } 236 237 return AMDGPU_RAS_SUCCESS; 238 } 239 240 int amdgpu_umc_pasid_poison_handler(struct amdgpu_device *adev, 241 enum amdgpu_ras_block block, uint16_t pasid, 242 pasid_notify pasid_fn, void *data, uint32_t reset) 243 { 244 int ret = AMDGPU_RAS_SUCCESS; 245 246 if (adev->gmc.xgmi.connected_to_cpu || 247 adev->gmc.is_app_apu) { 248 if (reset) { 249 /* MCA poison handler is only responsible for GPU reset, 250 * let MCA notifier do page retirement. 251 */ 252 kgd2kfd_set_sram_ecc_flag(adev->kfd.dev); 253 amdgpu_ras_reset_gpu(adev); 254 } 255 return ret; 256 } 257 258 if (!amdgpu_sriov_vf(adev)) { 259 if (amdgpu_ip_version(adev, UMC_HWIP, 0) < IP_VERSION(12, 0, 0)) { 260 struct ras_err_data err_data; 261 struct ras_common_if head = { 262 .block = AMDGPU_RAS_BLOCK__UMC, 263 }; 264 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head); 265 266 ret = amdgpu_ras_error_data_init(&err_data); 267 if (ret) 268 return ret; 269 270 ret = amdgpu_umc_do_page_retirement(adev, &err_data, NULL, reset); 271 272 if (ret == AMDGPU_RAS_SUCCESS && obj) { 273 obj->err_data.ue_count += err_data.ue_count; 274 obj->err_data.ce_count += err_data.ce_count; 275 obj->err_data.de_count += err_data.de_count; 276 } 277 278 amdgpu_ras_error_data_fini(&err_data); 279 } else if (amdgpu_uniras_enabled(adev)) { 280 struct ras_ih_info ih_info = {0}; 281 282 ih_info.block = block; 283 ih_info.pasid = pasid; 284 ih_info.reset = reset; 285 ih_info.pasid_fn = pasid_fn; 286 ih_info.data = data; 287 amdgpu_ras_mgr_handle_consumer_interrupt(adev, &ih_info); 288 } else { 289 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 290 int ret; 291 292 ret = amdgpu_ras_put_poison_req(adev, 293 block, pasid, pasid_fn, data, reset); 294 if (!ret) { 295 atomic_inc(&con->page_retirement_req_cnt); 296 atomic_inc(&con->poison_consumption_count); 297 wake_up(&con->page_retirement_wq); 298 } 299 } 300 } else { 301 if (adev->virt.ops && adev->virt.ops->ras_poison_handler) 302 adev->virt.ops->ras_poison_handler(adev, block); 303 else 304 dev_warn(adev->dev, 305 "No ras_poison_handler interface in SRIOV!\n"); 306 } 307 308 return ret; 309 } 310 311 int amdgpu_umc_poison_handler(struct amdgpu_device *adev, 312 enum amdgpu_ras_block block, uint32_t reset) 313 { 314 return amdgpu_umc_pasid_poison_handler(adev, 315 block, 0, NULL, NULL, reset); 316 } 317 318 int amdgpu_umc_process_ras_data_cb(struct amdgpu_device *adev, 319 void *ras_error_status, 320 struct amdgpu_iv_entry *entry) 321 { 322 return amdgpu_umc_do_page_retirement(adev, ras_error_status, entry, 323 AMDGPU_RAS_GPU_RESET_MODE1_RESET); 324 } 325 326 int amdgpu_umc_ras_sw_init(struct amdgpu_device *adev) 327 { 328 int err; 329 struct amdgpu_umc_ras *ras; 330 331 if (!adev->umc.ras) 332 return 0; 333 334 ras = adev->umc.ras; 335 336 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 337 if (err) { 338 dev_err(adev->dev, "Failed to register umc ras block!\n"); 339 return err; 340 } 341 342 strcpy(adev->umc.ras->ras_block.ras_comm.name, "umc"); 343 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__UMC; 344 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 345 adev->umc.ras_if = &ras->ras_block.ras_comm; 346 347 if (!ras->ras_block.ras_late_init) 348 ras->ras_block.ras_late_init = amdgpu_umc_ras_late_init; 349 350 if (!ras->ras_block.ras_cb) 351 ras->ras_block.ras_cb = amdgpu_umc_process_ras_data_cb; 352 353 return 0; 354 } 355 356 int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 357 { 358 int r; 359 360 r = amdgpu_ras_block_late_init(adev, ras_block); 361 if (r) 362 return r; 363 364 if (amdgpu_sriov_vf(adev)) 365 return r; 366 367 if (amdgpu_ras_is_supported(adev, ras_block->block)) { 368 r = amdgpu_irq_get(adev, &adev->gmc.ecc_irq, 0); 369 if (r) 370 goto late_fini; 371 } 372 373 /* ras init of specific umc version */ 374 if (adev->umc.ras && 375 adev->umc.ras->err_cnt_init) 376 adev->umc.ras->err_cnt_init(adev); 377 378 return 0; 379 380 late_fini: 381 amdgpu_ras_block_late_fini(adev, ras_block); 382 return r; 383 } 384 385 int amdgpu_umc_process_ecc_irq(struct amdgpu_device *adev, 386 struct amdgpu_irq_src *source, 387 struct amdgpu_iv_entry *entry) 388 { 389 struct ras_common_if *ras_if = adev->umc.ras_if; 390 struct ras_dispatch_if ih_data = { 391 .entry = entry, 392 }; 393 394 if (!ras_if) 395 return 0; 396 397 ih_data.head = *ras_if; 398 399 amdgpu_ras_interrupt_dispatch(adev, &ih_data); 400 return 0; 401 } 402 403 int amdgpu_umc_fill_error_record(struct ras_err_data *err_data, 404 uint64_t err_addr, 405 uint64_t retired_page, 406 uint32_t channel_index, 407 uint32_t umc_inst) 408 { 409 struct eeprom_table_record *err_rec; 410 411 if (!err_data || 412 !err_data->err_addr || 413 (err_data->err_addr_cnt >= err_data->err_addr_len)) 414 return -EINVAL; 415 416 err_rec = &err_data->err_addr[err_data->err_addr_cnt]; 417 418 err_rec->address = err_addr; 419 /* page frame address is saved */ 420 err_rec->retired_page = retired_page >> AMDGPU_GPU_PAGE_SHIFT; 421 err_rec->ts = (uint64_t)ktime_get_real_seconds(); 422 err_rec->err_type = AMDGPU_RAS_EEPROM_ERR_NON_RECOVERABLE; 423 err_rec->cu = 0; 424 err_rec->mem_channel = channel_index; 425 err_rec->mcumc_id = umc_inst; 426 427 err_data->err_addr_cnt++; 428 429 return 0; 430 } 431 432 static int amdgpu_umc_loop_all_aid(struct amdgpu_device *adev, umc_func func, 433 void *data) 434 { 435 uint32_t umc_node_inst; 436 uint32_t node_inst; 437 uint32_t umc_inst; 438 uint32_t ch_inst; 439 int ret; 440 441 /* 442 * This loop is done based on the following - 443 * umc.active mask = mask of active umc instances across all nodes 444 * umc.umc_inst_num = maximum number of umc instancess per node 445 * umc.node_inst_num = maximum number of node instances 446 * Channel instances are not assumed to be harvested. 447 */ 448 dev_dbg(adev->dev, "active umcs :%lx umc_inst per node: %d", 449 adev->umc.active_mask, adev->umc.umc_inst_num); 450 for_each_set_bit(umc_node_inst, &(adev->umc.active_mask), 451 adev->umc.node_inst_num * adev->umc.umc_inst_num) { 452 node_inst = umc_node_inst / adev->umc.umc_inst_num; 453 umc_inst = umc_node_inst % adev->umc.umc_inst_num; 454 LOOP_UMC_CH_INST(ch_inst) { 455 dev_dbg(adev->dev, 456 "node_inst :%d umc_inst: %d ch_inst: %d", 457 node_inst, umc_inst, ch_inst); 458 ret = func(adev, node_inst, umc_inst, ch_inst, data); 459 if (ret) { 460 dev_err(adev->dev, 461 "Node %d umc %d ch %d func returns %d\n", 462 node_inst, umc_inst, ch_inst, ret); 463 return ret; 464 } 465 } 466 } 467 468 return 0; 469 } 470 471 int amdgpu_umc_loop_channels(struct amdgpu_device *adev, 472 umc_func func, void *data) 473 { 474 uint32_t node_inst = 0; 475 uint32_t umc_inst = 0; 476 uint32_t ch_inst = 0; 477 int ret = 0; 478 479 if (adev->aid_mask) 480 return amdgpu_umc_loop_all_aid(adev, func, data); 481 482 if (adev->umc.node_inst_num) { 483 LOOP_UMC_EACH_NODE_INST_AND_CH(node_inst, umc_inst, ch_inst) { 484 ret = func(adev, node_inst, umc_inst, ch_inst, data); 485 if (ret) { 486 dev_err(adev->dev, "Node %d umc %d ch %d func returns %d\n", 487 node_inst, umc_inst, ch_inst, ret); 488 return ret; 489 } 490 } 491 } else { 492 LOOP_UMC_INST_AND_CH(umc_inst, ch_inst) { 493 ret = func(adev, 0, umc_inst, ch_inst, data); 494 if (ret) { 495 dev_err(adev->dev, "Umc %d ch %d func returns %d\n", 496 umc_inst, ch_inst, ret); 497 return ret; 498 } 499 } 500 } 501 502 return 0; 503 } 504 505 int amdgpu_umc_update_ecc_status(struct amdgpu_device *adev, 506 uint64_t status, uint64_t ipid, uint64_t addr) 507 { 508 if (adev->umc.ras->update_ecc_status) 509 return adev->umc.ras->update_ecc_status(adev, 510 status, ipid, addr); 511 return 0; 512 } 513 514 int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev, 515 struct radix_tree_root *ecc_tree, struct ras_ecc_err *ecc_err) 516 { 517 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 518 struct ras_ecc_log_info *ecc_log; 519 int ret; 520 521 ecc_log = &con->umc_ecc_log; 522 523 mutex_lock(&ecc_log->lock); 524 ret = radix_tree_insert(ecc_tree, ecc_err->pa_pfn, ecc_err); 525 if (!ret) 526 radix_tree_tag_set(ecc_tree, 527 ecc_err->pa_pfn, UMC_ECC_NEW_DETECTED_TAG); 528 mutex_unlock(&ecc_log->lock); 529 530 return ret; 531 } 532 533 int amdgpu_umc_pages_in_a_row(struct amdgpu_device *adev, 534 struct ras_err_data *err_data, uint64_t pa_addr) 535 { 536 struct ta_ras_query_address_output addr_out; 537 538 /* reinit err_data */ 539 err_data->err_addr_cnt = 0; 540 err_data->err_addr_len = adev->umc.retire_unit; 541 542 addr_out.pa.pa = pa_addr; 543 if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) 544 return adev->umc.ras->convert_ras_err_addr(adev, err_data, NULL, 545 &addr_out, false); 546 else 547 return -EINVAL; 548 } 549 550 int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, 551 uint64_t pa_addr, uint64_t *pfns, int len) 552 { 553 int i, ret; 554 struct ras_err_data err_data; 555 556 err_data.err_addr = kzalloc_objs(struct eeprom_table_record, 557 adev->umc.retire_unit, GFP_KERNEL); 558 if (!err_data.err_addr) { 559 dev_warn(adev->dev, "Failed to alloc memory in bad page lookup!\n"); 560 return 0; 561 } 562 563 ret = amdgpu_umc_pages_in_a_row(adev, &err_data, pa_addr); 564 if (ret) 565 goto out; 566 567 for (i = 0; i < adev->umc.retire_unit; i++) { 568 if (i >= len) 569 goto out; 570 571 pfns[i] = err_data.err_addr[i].retired_page; 572 } 573 ret = i; 574 adev->umc.err_addr_cnt = err_data.err_addr_cnt; 575 576 out: 577 kfree(err_data.err_addr); 578 return ret; 579 } 580 581 int amdgpu_umc_mca_to_addr(struct amdgpu_device *adev, 582 uint64_t err_addr, uint32_t ch, uint32_t umc, 583 uint32_t node, uint32_t socket, 584 struct ta_ras_query_address_output *addr_out, bool dump_addr) 585 { 586 struct ta_ras_query_address_input addr_in; 587 int ret; 588 589 memset(&addr_in, 0, sizeof(addr_in)); 590 addr_in.ma.err_addr = err_addr; 591 addr_in.ma.ch_inst = ch; 592 addr_in.ma.umc_inst = umc; 593 addr_in.ma.node_inst = node; 594 addr_in.ma.socket_id = socket; 595 596 if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) { 597 ret = adev->umc.ras->convert_ras_err_addr(adev, NULL, &addr_in, 598 addr_out, dump_addr); 599 if (ret) 600 return ret; 601 } else { 602 return 0; 603 } 604 605 return 0; 606 } 607 608 int amdgpu_umc_pa2mca(struct amdgpu_device *adev, 609 uint64_t pa, uint64_t *mca, enum amdgpu_memory_partition nps) 610 { 611 struct ta_ras_query_address_input addr_in; 612 struct ta_ras_query_address_output addr_out; 613 int ret; 614 615 /* nps: the pa belongs to */ 616 addr_in.pa.pa = pa | ((uint64_t)nps << 58); 617 addr_in.addr_type = TA_RAS_PA_TO_MCA; 618 ret = psp_ras_query_address(&adev->psp, &addr_in, &addr_out); 619 if (ret) { 620 dev_warn(adev->dev, "Failed to query RAS MCA address for 0x%llx", 621 pa); 622 623 return ret; 624 } 625 626 *mca = addr_out.ma.err_addr; 627 628 return 0; 629 } 630