1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * PCI Error Recovery Driver for RPA-compliant PPC64 platform. 4 * Copyright IBM Corp. 2004 2005 5 * Copyright Linas Vepstas <linas@linas.org> 2004, 2005 6 * 7 * Send comments and feedback to Linas Vepstas <linas@austin.ibm.com> 8 */ 9 #include <linux/delay.h> 10 #include <linux/interrupt.h> 11 #include <linux/irq.h> 12 #include <linux/module.h> 13 #include <linux/pci.h> 14 #include <linux/pci_hotplug.h> 15 #include <asm/eeh.h> 16 #include <asm/eeh_event.h> 17 #include <asm/ppc-pci.h> 18 #include <asm/pci-bridge.h> 19 #include <asm/rtas.h> 20 21 struct eeh_rmv_data { 22 struct list_head removed_vf_list; 23 int removed_dev_count; 24 }; 25 26 static int eeh_result_priority(enum pci_ers_result result) 27 { 28 switch (result) { 29 case PCI_ERS_RESULT_NONE: 30 return 1; 31 case PCI_ERS_RESULT_NO_AER_DRIVER: 32 return 2; 33 case PCI_ERS_RESULT_RECOVERED: 34 return 3; 35 case PCI_ERS_RESULT_CAN_RECOVER: 36 return 4; 37 case PCI_ERS_RESULT_DISCONNECT: 38 return 5; 39 case PCI_ERS_RESULT_NEED_RESET: 40 return 6; 41 default: 42 WARN_ONCE(1, "Unknown pci_ers_result value: %d\n", result); 43 return 0; 44 } 45 }; 46 47 static const char *pci_ers_result_name(enum pci_ers_result result) 48 { 49 switch (result) { 50 case PCI_ERS_RESULT_NONE: 51 return "none"; 52 case PCI_ERS_RESULT_CAN_RECOVER: 53 return "can recover"; 54 case PCI_ERS_RESULT_NEED_RESET: 55 return "need reset"; 56 case PCI_ERS_RESULT_DISCONNECT: 57 return "disconnect"; 58 case PCI_ERS_RESULT_RECOVERED: 59 return "recovered"; 60 case PCI_ERS_RESULT_NO_AER_DRIVER: 61 return "no AER driver"; 62 default: 63 WARN_ONCE(1, "Unknown result type: %d\n", result); 64 return "unknown"; 65 } 66 }; 67 68 static enum pci_ers_result pci_ers_merge_result(enum pci_ers_result old, 69 enum pci_ers_result new) 70 { 71 if (eeh_result_priority(new) > eeh_result_priority(old)) 72 return new; 73 return old; 74 } 75 76 static bool eeh_dev_removed(struct eeh_dev *edev) 77 { 78 return !edev || (edev->mode & EEH_DEV_REMOVED); 79 } 80 81 static bool eeh_edev_actionable(struct eeh_dev *edev) 82 { 83 if (!edev->pdev) 84 return false; 85 if (edev->pdev->error_state == pci_channel_io_perm_failure) 86 return false; 87 if (eeh_dev_removed(edev)) 88 return false; 89 if (eeh_pe_passed(edev->pe)) 90 return false; 91 92 return true; 93 } 94 95 /** 96 * eeh_pcid_get - Get the PCI device driver 97 * @pdev: PCI device 98 * 99 * The function is used to retrieve the PCI device driver for 100 * the indicated PCI device. Besides, we will increase the reference 101 * of the PCI device driver to prevent that being unloaded on 102 * the fly. Otherwise, kernel crash would be seen. 103 */ 104 static inline struct pci_driver *eeh_pcid_get(struct pci_dev *pdev) 105 { 106 if (!pdev || !pdev->dev.driver) 107 return NULL; 108 109 if (!try_module_get(pdev->dev.driver->owner)) 110 return NULL; 111 112 return to_pci_driver(pdev->dev.driver); 113 } 114 115 /** 116 * eeh_pcid_put - Dereference on the PCI device driver 117 * @pdev: PCI device 118 * 119 * The function is called to do dereference on the PCI device 120 * driver of the indicated PCI device. 121 */ 122 static inline void eeh_pcid_put(struct pci_dev *pdev) 123 { 124 if (!pdev || !pdev->dev.driver) 125 return; 126 127 module_put(pdev->dev.driver->owner); 128 } 129 130 /** 131 * eeh_disable_irq - Disable interrupt for the recovering device 132 * @dev: PCI device 133 * 134 * This routine must be called when reporting temporary or permanent 135 * error to the particular PCI device to disable interrupt of that 136 * device. If the device has enabled MSI or MSI-X interrupt, we needn't 137 * do real work because EEH should freeze DMA transfers for those PCI 138 * devices encountering EEH errors, which includes MSI or MSI-X. 139 */ 140 static void eeh_disable_irq(struct eeh_dev *edev) 141 { 142 /* Don't disable MSI and MSI-X interrupts. They are 143 * effectively disabled by the DMA Stopped state 144 * when an EEH error occurs. 145 */ 146 if (edev->pdev->msi_enabled || edev->pdev->msix_enabled) 147 return; 148 149 if (!irq_has_action(edev->pdev->irq)) 150 return; 151 152 edev->mode |= EEH_DEV_IRQ_DISABLED; 153 disable_irq_nosync(edev->pdev->irq); 154 } 155 156 /** 157 * eeh_enable_irq - Enable interrupt for the recovering device 158 * @dev: PCI device 159 * 160 * This routine must be called to enable interrupt while failed 161 * device could be resumed. 162 */ 163 static void eeh_enable_irq(struct eeh_dev *edev) 164 { 165 if ((edev->mode) & EEH_DEV_IRQ_DISABLED) { 166 edev->mode &= ~EEH_DEV_IRQ_DISABLED; 167 /* 168 * FIXME !!!!! 169 * 170 * This is just ass backwards. This maze has 171 * unbalanced irq_enable/disable calls. So instead of 172 * finding the root cause it works around the warning 173 * in the irq_enable code by conditionally calling 174 * into it. 175 * 176 * That's just wrong.The warning in the core code is 177 * there to tell people to fix their asymmetries in 178 * their own code, not by abusing the core information 179 * to avoid it. 180 * 181 * I so wish that the assymetry would be the other way 182 * round and a few more irq_disable calls render that 183 * shit unusable forever. 184 * 185 * tglx 186 */ 187 if (irqd_irq_disabled(irq_get_irq_data(edev->pdev->irq))) 188 enable_irq(edev->pdev->irq); 189 } 190 } 191 192 static void eeh_dev_save_state(struct eeh_dev *edev, void *userdata) 193 { 194 struct pci_dev *pdev; 195 196 if (!edev) 197 return; 198 199 /* 200 * We cannot access the config space on some adapters. 201 * Otherwise, it will cause fenced PHB. We don't save 202 * the content in their config space and will restore 203 * from the initial config space saved when the EEH 204 * device is created. 205 */ 206 if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED)) 207 return; 208 209 pdev = eeh_dev_to_pci_dev(edev); 210 if (!pdev) 211 return; 212 213 pci_save_state(pdev); 214 } 215 216 static void eeh_set_channel_state(struct eeh_pe *root, pci_channel_state_t s) 217 { 218 struct eeh_pe *pe; 219 struct eeh_dev *edev, *tmp; 220 221 eeh_for_each_pe(root, pe) 222 eeh_pe_for_each_dev(pe, edev, tmp) 223 if (eeh_edev_actionable(edev)) 224 edev->pdev->error_state = s; 225 } 226 227 static void eeh_set_irq_state(struct eeh_pe *root, bool enable) 228 { 229 struct eeh_pe *pe; 230 struct eeh_dev *edev, *tmp; 231 232 eeh_for_each_pe(root, pe) { 233 eeh_pe_for_each_dev(pe, edev, tmp) { 234 if (!eeh_edev_actionable(edev)) 235 continue; 236 237 if (!eeh_pcid_get(edev->pdev)) 238 continue; 239 240 if (enable) 241 eeh_enable_irq(edev); 242 else 243 eeh_disable_irq(edev); 244 245 eeh_pcid_put(edev->pdev); 246 } 247 } 248 } 249 250 typedef enum pci_ers_result (*eeh_report_fn)(struct eeh_dev *, 251 struct pci_dev *, 252 struct pci_driver *); 253 static void eeh_pe_report_edev(struct eeh_dev *edev, eeh_report_fn fn, 254 enum pci_ers_result *result) 255 { 256 struct pci_dev *pdev; 257 struct pci_driver *driver; 258 enum pci_ers_result new_result; 259 260 pdev = edev->pdev; 261 if (pdev) 262 get_device(&pdev->dev); 263 if (!pdev) { 264 eeh_edev_info(edev, "no device"); 265 *result = PCI_ERS_RESULT_DISCONNECT; 266 return; 267 } 268 device_lock(&pdev->dev); 269 if (eeh_edev_actionable(edev)) { 270 driver = eeh_pcid_get(pdev); 271 272 if (!driver) 273 eeh_edev_info(edev, "no driver"); 274 else if (!driver->err_handler) 275 eeh_edev_info(edev, "driver not EEH aware"); 276 else if (edev->mode & EEH_DEV_NO_HANDLER) 277 eeh_edev_info(edev, "driver bound too late"); 278 else { 279 new_result = fn(edev, pdev, driver); 280 eeh_edev_info(edev, "%s driver reports: '%s'", 281 driver->name, 282 pci_ers_result_name(new_result)); 283 if (result) 284 *result = pci_ers_merge_result(*result, 285 new_result); 286 } 287 if (driver) 288 eeh_pcid_put(pdev); 289 } else { 290 eeh_edev_info(edev, "not actionable (%d,%d,%d)", !!pdev, 291 !eeh_dev_removed(edev), !eeh_pe_passed(edev->pe)); 292 } 293 device_unlock(&pdev->dev); 294 if (edev->pdev != pdev) 295 eeh_edev_warn(edev, "Device changed during processing!\n"); 296 put_device(&pdev->dev); 297 } 298 299 static void eeh_pe_report(const char *name, struct eeh_pe *root, 300 eeh_report_fn fn, enum pci_ers_result *result) 301 { 302 struct eeh_pe *pe; 303 struct eeh_dev *edev, *tmp; 304 305 pr_info("EEH: Beginning: '%s'\n", name); 306 eeh_for_each_pe(root, pe) 307 eeh_pe_for_each_dev(pe, edev, tmp) 308 eeh_pe_report_edev(edev, fn, result); 309 if (result) 310 pr_info("EEH: Finished:'%s' with aggregate recovery state:'%s'\n", 311 name, pci_ers_result_name(*result)); 312 else 313 pr_info("EEH: Finished:'%s'", name); 314 } 315 316 /** 317 * eeh_report_error - Report pci error to each device driver 318 * @edev: eeh device 319 * @driver: device's PCI driver 320 * 321 * Report an EEH error to each device driver. 322 */ 323 static enum pci_ers_result eeh_report_error(struct eeh_dev *edev, 324 struct pci_dev *pdev, 325 struct pci_driver *driver) 326 { 327 enum pci_ers_result rc; 328 329 if (!driver->err_handler->error_detected) 330 return PCI_ERS_RESULT_NONE; 331 332 eeh_edev_info(edev, "Invoking %s->error_detected(IO frozen)", 333 driver->name); 334 rc = driver->err_handler->error_detected(pdev, pci_channel_io_frozen); 335 336 edev->in_error = true; 337 pci_uevent_ers(pdev, rc); 338 return rc; 339 } 340 341 /** 342 * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled 343 * @edev: eeh device 344 * @driver: device's PCI driver 345 * 346 * Tells each device driver that IO ports, MMIO and config space I/O 347 * are now enabled. 348 */ 349 static enum pci_ers_result eeh_report_mmio_enabled(struct eeh_dev *edev, 350 struct pci_dev *pdev, 351 struct pci_driver *driver) 352 { 353 if (!driver->err_handler->mmio_enabled) 354 return PCI_ERS_RESULT_NONE; 355 eeh_edev_info(edev, "Invoking %s->mmio_enabled()", driver->name); 356 return driver->err_handler->mmio_enabled(pdev); 357 } 358 359 /** 360 * eeh_report_reset - Tell device that slot has been reset 361 * @edev: eeh device 362 * @driver: device's PCI driver 363 * 364 * This routine must be called while EEH tries to reset particular 365 * PCI device so that the associated PCI device driver could take 366 * some actions, usually to save data the driver needs so that the 367 * driver can work again while the device is recovered. 368 */ 369 static enum pci_ers_result eeh_report_reset(struct eeh_dev *edev, 370 struct pci_dev *pdev, 371 struct pci_driver *driver) 372 { 373 if (!driver->err_handler->slot_reset || !edev->in_error) 374 return PCI_ERS_RESULT_NONE; 375 eeh_edev_info(edev, "Invoking %s->slot_reset()", driver->name); 376 return driver->err_handler->slot_reset(pdev); 377 } 378 379 static void eeh_dev_restore_state(struct eeh_dev *edev, void *userdata) 380 { 381 struct pci_dev *pdev; 382 383 if (!edev) 384 return; 385 386 pci_lock_rescan_remove(); 387 388 /* 389 * The content in the config space isn't saved because 390 * the blocked config space on some adapters. We have 391 * to restore the initial saved config space when the 392 * EEH device is created. 393 */ 394 if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED)) { 395 if (list_is_last(&edev->entry, &edev->pe->edevs)) 396 eeh_pe_restore_bars(edev->pe); 397 398 pci_unlock_rescan_remove(); 399 return; 400 } 401 402 pdev = eeh_dev_to_pci_dev(edev); 403 if (!pdev) { 404 pci_unlock_rescan_remove(); 405 return; 406 } 407 408 pci_restore_state(pdev); 409 410 pci_unlock_rescan_remove(); 411 } 412 413 /** 414 * eeh_report_resume - Tell device to resume normal operations 415 * @edev: eeh device 416 * @driver: device's PCI driver 417 * 418 * This routine must be called to notify the device driver that it 419 * could resume so that the device driver can do some initialization 420 * to make the recovered device work again. 421 */ 422 static enum pci_ers_result eeh_report_resume(struct eeh_dev *edev, 423 struct pci_dev *pdev, 424 struct pci_driver *driver) 425 { 426 if (!driver->err_handler->resume || !edev->in_error) 427 return PCI_ERS_RESULT_NONE; 428 429 eeh_edev_info(edev, "Invoking %s->resume()", driver->name); 430 driver->err_handler->resume(pdev); 431 432 pci_uevent_ers(edev->pdev, PCI_ERS_RESULT_RECOVERED); 433 #ifdef CONFIG_PCI_IOV 434 if (eeh_ops->notify_resume) 435 eeh_ops->notify_resume(edev); 436 #endif 437 return PCI_ERS_RESULT_NONE; 438 } 439 440 /** 441 * eeh_report_failure - Tell device driver that device is dead. 442 * @edev: eeh device 443 * @driver: device's PCI driver 444 * 445 * This informs the device driver that the device is permanently 446 * dead, and that no further recovery attempts will be made on it. 447 */ 448 static enum pci_ers_result eeh_report_failure(struct eeh_dev *edev, 449 struct pci_dev *pdev, 450 struct pci_driver *driver) 451 { 452 enum pci_ers_result rc; 453 454 if (!driver->err_handler->error_detected) 455 return PCI_ERS_RESULT_NONE; 456 457 eeh_edev_info(edev, "Invoking %s->error_detected(permanent failure)", 458 driver->name); 459 rc = driver->err_handler->error_detected(pdev, 460 pci_channel_io_perm_failure); 461 462 pci_uevent_ers(pdev, PCI_ERS_RESULT_DISCONNECT); 463 return rc; 464 } 465 466 static void *eeh_add_virt_device(struct eeh_dev *edev) 467 { 468 struct pci_driver *driver; 469 struct pci_dev *dev = eeh_dev_to_pci_dev(edev); 470 471 if (!(edev->physfn)) { 472 eeh_edev_warn(edev, "Not for VF\n"); 473 return NULL; 474 } 475 476 driver = eeh_pcid_get(dev); 477 if (driver) { 478 if (driver->err_handler) { 479 eeh_pcid_put(dev); 480 return NULL; 481 } 482 eeh_pcid_put(dev); 483 } 484 485 #ifdef CONFIG_PCI_IOV 486 pci_iov_add_virtfn(edev->physfn, edev->vf_index); 487 #endif 488 return NULL; 489 } 490 491 static void eeh_rmv_device(struct eeh_dev *edev, void *userdata) 492 { 493 struct pci_driver *driver; 494 struct pci_dev *dev = eeh_dev_to_pci_dev(edev); 495 struct eeh_rmv_data *rmv_data = (struct eeh_rmv_data *)userdata; 496 497 /* 498 * Actually, we should remove the PCI bridges as well. 499 * However, that's lots of complexity to do that, 500 * particularly some of devices under the bridge might 501 * support EEH. So we just care about PCI devices for 502 * simplicity here. 503 */ 504 if (!eeh_edev_actionable(edev) || 505 (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE)) 506 return; 507 508 if (rmv_data) { 509 driver = eeh_pcid_get(dev); 510 if (driver) { 511 if (driver->err_handler && 512 driver->err_handler->error_detected && 513 driver->err_handler->slot_reset) { 514 eeh_pcid_put(dev); 515 return; 516 } 517 eeh_pcid_put(dev); 518 } 519 } 520 521 /* Remove it from PCI subsystem */ 522 pr_info("EEH: Removing %s without EEH sensitive driver\n", 523 pci_name(dev)); 524 edev->mode |= EEH_DEV_DISCONNECTED; 525 if (rmv_data) 526 rmv_data->removed_dev_count++; 527 528 if (edev->physfn) { 529 #ifdef CONFIG_PCI_IOV 530 pci_iov_remove_virtfn(edev->physfn, edev->vf_index); 531 edev->pdev = NULL; 532 #endif 533 if (rmv_data) 534 list_add(&edev->rmv_entry, &rmv_data->removed_vf_list); 535 } else { 536 pci_stop_and_remove_bus_device(dev); 537 } 538 } 539 540 static void *eeh_pe_detach_dev(struct eeh_pe *pe, void *userdata) 541 { 542 struct eeh_dev *edev, *tmp; 543 544 eeh_pe_for_each_dev(pe, edev, tmp) { 545 if (!(edev->mode & EEH_DEV_DISCONNECTED)) 546 continue; 547 548 edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED); 549 eeh_pe_tree_remove(edev); 550 } 551 552 return NULL; 553 } 554 555 /* 556 * Explicitly clear PE's frozen state for PowerNV where 557 * we have frozen PE until BAR restore is completed. It's 558 * harmless to clear it for pSeries. To be consistent with 559 * PE reset (for 3 times), we try to clear the frozen state 560 * for 3 times as well. 561 */ 562 static int eeh_clear_pe_frozen_state(struct eeh_pe *root, bool include_passed) 563 { 564 struct eeh_pe *pe; 565 int i; 566 567 eeh_for_each_pe(root, pe) { 568 if (include_passed || !eeh_pe_passed(pe)) { 569 for (i = 0; i < 3; i++) 570 if (!eeh_unfreeze_pe(pe)) 571 break; 572 if (i >= 3) 573 return -EIO; 574 } 575 } 576 eeh_pe_state_clear(root, EEH_PE_ISOLATED, include_passed); 577 return 0; 578 } 579 580 int eeh_pe_reset_and_recover(struct eeh_pe *pe) 581 { 582 int ret; 583 584 /* Bail if the PE is being recovered */ 585 if (pe->state & EEH_PE_RECOVERING) 586 return 0; 587 588 /* Put the PE into recovery mode */ 589 eeh_pe_state_mark(pe, EEH_PE_RECOVERING); 590 591 /* Save states */ 592 eeh_pe_dev_traverse(pe, eeh_dev_save_state, NULL); 593 594 /* Issue reset */ 595 ret = eeh_pe_reset_full(pe, true); 596 if (ret) { 597 eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true); 598 return ret; 599 } 600 601 /* Unfreeze the PE */ 602 ret = eeh_clear_pe_frozen_state(pe, true); 603 if (ret) { 604 eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true); 605 return ret; 606 } 607 608 /* Restore device state */ 609 eeh_pe_dev_traverse(pe, eeh_dev_restore_state, NULL); 610 611 /* Clear recovery mode */ 612 eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true); 613 614 return 0; 615 } 616 617 /** 618 * eeh_reset_device - Perform actual reset of a pci slot 619 * @driver_eeh_aware: Does the device's driver provide EEH support? 620 * @pe: EEH PE 621 * @bus: PCI bus corresponding to the isolcated slot 622 * @rmv_data: Optional, list to record removed devices 623 * 624 * This routine must be called to do reset on the indicated PE. 625 * During the reset, udev might be invoked because those affected 626 * PCI devices will be removed and then added. 627 */ 628 static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus, 629 struct eeh_rmv_data *rmv_data, 630 bool driver_eeh_aware) 631 { 632 time64_t tstamp; 633 int cnt, rc; 634 struct eeh_dev *edev; 635 struct eeh_pe *tmp_pe; 636 bool any_passed = false; 637 638 eeh_for_each_pe(pe, tmp_pe) 639 any_passed |= eeh_pe_passed(tmp_pe); 640 641 /* pcibios will clear the counter; save the value */ 642 cnt = pe->freeze_count; 643 tstamp = pe->tstamp; 644 645 /* 646 * We don't remove the corresponding PE instances because 647 * we need the information afterwords. The attached EEH 648 * devices are expected to be attached soon when calling 649 * into pci_hp_add_devices(). 650 */ 651 eeh_pe_state_mark(pe, EEH_PE_KEEP); 652 if (any_passed || driver_eeh_aware || (pe->type & EEH_PE_VF)) { 653 eeh_pe_dev_traverse(pe, eeh_rmv_device, rmv_data); 654 } else { 655 pci_hp_remove_devices(bus); 656 } 657 658 /* 659 * Reset the pci controller. (Asserts RST#; resets config space). 660 * Reconfigure bridges and devices. Don't try to bring the system 661 * up if the reset failed for some reason. 662 * 663 * During the reset, it's very dangerous to have uncontrolled PCI 664 * config accesses. So we prefer to block them. However, controlled 665 * PCI config accesses initiated from EEH itself are allowed. 666 */ 667 rc = eeh_pe_reset_full(pe, false); 668 if (rc) 669 return rc; 670 671 /* Restore PE */ 672 eeh_ops->configure_bridge(pe); 673 eeh_pe_restore_bars(pe); 674 675 /* Clear frozen state */ 676 rc = eeh_clear_pe_frozen_state(pe, false); 677 if (rc) { 678 return rc; 679 } 680 681 /* Give the system 5 seconds to finish running the user-space 682 * hotplug shutdown scripts, e.g. ifdown for ethernet. Yes, 683 * this is a hack, but if we don't do this, and try to bring 684 * the device up before the scripts have taken it down, 685 * potentially weird things happen. 686 */ 687 if (!driver_eeh_aware || rmv_data->removed_dev_count) { 688 pr_info("EEH: Sleep 5s ahead of %s hotplug\n", 689 (driver_eeh_aware ? "partial" : "complete")); 690 ssleep(5); 691 692 /* 693 * The EEH device is still connected with its parent 694 * PE. We should disconnect it so the binding can be 695 * rebuilt when adding PCI devices. 696 */ 697 edev = list_first_entry(&pe->edevs, struct eeh_dev, entry); 698 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL); 699 if (pe->type & EEH_PE_VF) { 700 eeh_add_virt_device(edev); 701 } else { 702 if (!driver_eeh_aware) 703 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true); 704 pci_hp_add_devices(bus); 705 } 706 } 707 eeh_pe_state_clear(pe, EEH_PE_KEEP, true); 708 709 pe->tstamp = tstamp; 710 pe->freeze_count = cnt; 711 712 return 0; 713 } 714 715 /* The longest amount of time to wait for a pci device 716 * to come back on line, in seconds. 717 */ 718 #define MAX_WAIT_FOR_RECOVERY 300 719 720 721 /* Walks the PE tree after processing an event to remove any stale PEs. 722 * 723 * NB: This needs to be recursive to ensure the leaf PEs get removed 724 * before their parents do. Although this is possible to do recursively 725 * we don't since this is easier to read and we need to garantee 726 * the leaf nodes will be handled first. 727 */ 728 static void eeh_pe_cleanup(struct eeh_pe *pe) 729 { 730 struct eeh_pe *child_pe, *tmp; 731 732 list_for_each_entry_safe(child_pe, tmp, &pe->child_list, child) 733 eeh_pe_cleanup(child_pe); 734 735 if (pe->state & EEH_PE_KEEP) 736 return; 737 738 if (!(pe->state & EEH_PE_INVALID)) 739 return; 740 741 if (list_empty(&pe->edevs) && list_empty(&pe->child_list)) { 742 list_del(&pe->child); 743 kfree(pe); 744 } 745 } 746 747 /** 748 * eeh_check_slot_presence - Check if a device is still present in a slot 749 * @pdev: pci_dev to check 750 * 751 * This function may return a false positive if we can't determine the slot's 752 * presence state. This might happen for PCIe slots if the PE containing 753 * the upstream bridge is also frozen, or the bridge is part of the same PE 754 * as the device. 755 * 756 * This shouldn't happen often, but you might see it if you hotplug a PCIe 757 * switch. 758 */ 759 static bool eeh_slot_presence_check(struct pci_dev *pdev) 760 { 761 const struct hotplug_slot_ops *ops; 762 struct pci_slot *slot; 763 u8 state; 764 int rc; 765 766 if (!pdev) 767 return false; 768 769 if (pdev->error_state == pci_channel_io_perm_failure) 770 return false; 771 772 slot = pdev->slot; 773 if (!slot || !slot->hotplug) 774 return true; 775 776 ops = slot->hotplug->ops; 777 if (!ops || !ops->get_adapter_status) 778 return true; 779 780 /* set the attention indicator while we've got the slot ops */ 781 if (ops->set_attention_status) 782 ops->set_attention_status(slot->hotplug, 1); 783 784 rc = ops->get_adapter_status(slot->hotplug, &state); 785 if (rc) 786 return true; 787 788 return !!state; 789 } 790 791 static void eeh_clear_slot_attention(struct pci_dev *pdev) 792 { 793 const struct hotplug_slot_ops *ops; 794 struct pci_slot *slot; 795 796 if (!pdev) 797 return; 798 799 if (pdev->error_state == pci_channel_io_perm_failure) 800 return; 801 802 slot = pdev->slot; 803 if (!slot || !slot->hotplug) 804 return; 805 806 ops = slot->hotplug->ops; 807 if (!ops || !ops->set_attention_status) 808 return; 809 810 ops->set_attention_status(slot->hotplug, 0); 811 } 812 813 /** 814 * eeh_handle_normal_event - Handle EEH events on a specific PE 815 * @pe: EEH PE - which should not be used after we return, as it may 816 * have been invalidated. 817 * 818 * Attempts to recover the given PE. If recovery fails or the PE has failed 819 * too many times, remove the PE. 820 * 821 * While PHB detects address or data parity errors on particular PCI 822 * slot, the associated PE will be frozen. Besides, DMA's occurring 823 * to wild addresses (which usually happen due to bugs in device 824 * drivers or in PCI adapter firmware) can cause EEH error. #SERR, 825 * #PERR or other misc PCI-related errors also can trigger EEH errors. 826 * 827 * Recovery process consists of unplugging the device driver (which 828 * generated hotplug events to userspace), then issuing a PCI #RST to 829 * the device, then reconfiguring the PCI config space for all bridges 830 * & devices under this slot, and then finally restarting the device 831 * drivers (which cause a second set of hotplug events to go out to 832 * userspace). 833 */ 834 void eeh_handle_normal_event(struct eeh_pe *pe) 835 { 836 struct pci_bus *bus; 837 struct eeh_dev *edev, *tmp; 838 struct eeh_pe *tmp_pe; 839 int rc = 0; 840 enum pci_ers_result result = PCI_ERS_RESULT_NONE; 841 struct eeh_rmv_data rmv_data = 842 {LIST_HEAD_INIT(rmv_data.removed_vf_list), 0}; 843 int devices = 0; 844 845 pci_lock_rescan_remove(); 846 847 bus = eeh_pe_bus_get_nolock(pe); 848 if (!bus) { 849 pr_err("%s: Cannot find PCI bus for PHB#%x-PE#%x\n", 850 __func__, pe->phb->global_number, pe->addr); 851 pci_unlock_rescan_remove(); 852 return; 853 } 854 855 /* 856 * When devices are hot-removed we might get an EEH due to 857 * a driver attempting to touch the MMIO space of a removed 858 * device. In this case we don't have a device to recover 859 * so suppress the event if we can't find any present devices. 860 * 861 * The hotplug driver should take care of tearing down the 862 * device itself. 863 */ 864 eeh_for_each_pe(pe, tmp_pe) 865 eeh_pe_for_each_dev(tmp_pe, edev, tmp) 866 if (eeh_slot_presence_check(edev->pdev)) 867 devices++; 868 869 if (!devices) { 870 pr_warn("EEH: Frozen PHB#%x-PE#%x is empty!\n", 871 pe->phb->global_number, pe->addr); 872 /* 873 * The device is removed, tear down its state, on powernv 874 * hotplug driver would take care of it but not on pseries, 875 * permanently disable the card as it is hot removed. 876 * 877 * In the case of powernv, note that the removal of device 878 * is covered by pci rescan lock, so no problem even if hotplug 879 * driver attempts to remove the device. 880 */ 881 goto recover_failed; 882 } 883 884 /* Log the event */ 885 if (pe->type & EEH_PE_PHB) { 886 pr_err("EEH: Recovering PHB#%x, location: %s\n", 887 pe->phb->global_number, eeh_pe_loc_get_bus(bus)); 888 } else { 889 struct eeh_pe *phb_pe = eeh_phb_pe_get(pe->phb); 890 891 pr_err("EEH: Recovering PHB#%x-PE#%x\n", 892 pe->phb->global_number, pe->addr); 893 pr_err("EEH: PE location: %s, PHB location: %s\n", 894 eeh_pe_loc_get_bus(bus), 895 eeh_pe_loc_get_bus(eeh_pe_bus_get_nolock(phb_pe))); 896 } 897 898 #ifdef CONFIG_STACKTRACE 899 /* 900 * Print the saved stack trace now that we've verified there's 901 * something to recover. 902 */ 903 if (pe->trace_entries) { 904 void **ptrs = (void **) pe->stack_trace; 905 int i; 906 907 pr_err("EEH: Frozen PHB#%x-PE#%x detected\n", 908 pe->phb->global_number, pe->addr); 909 910 /* FIXME: Use the same format as dump_stack() */ 911 pr_err("EEH: Call Trace:\n"); 912 for (i = 0; i < pe->trace_entries; i++) 913 pr_err("EEH: [%p] %pS\n", ptrs[i], ptrs[i]); 914 915 pe->trace_entries = 0; 916 } 917 #endif /* CONFIG_STACKTRACE */ 918 919 eeh_for_each_pe(pe, tmp_pe) 920 eeh_pe_for_each_dev(tmp_pe, edev, tmp) 921 edev->mode &= ~EEH_DEV_NO_HANDLER; 922 923 eeh_pe_update_time_stamp(pe); 924 pe->freeze_count++; 925 if (pe->freeze_count > eeh_max_freezes) { 926 pr_err("EEH: PHB#%x-PE#%x has failed %d times in the last hour and has been permanently disabled.\n", 927 pe->phb->global_number, pe->addr, 928 pe->freeze_count); 929 930 goto recover_failed; 931 } 932 933 /* Walk the various device drivers attached to this slot through 934 * a reset sequence, giving each an opportunity to do what it needs 935 * to accomplish the reset. Each child gets a report of the 936 * status ... if any child can't handle the reset, then the entire 937 * slot is dlpar removed and added. 938 * 939 * When the PHB is fenced, we have to issue a reset to recover from 940 * the error. Override the result if necessary to have partially 941 * hotplug for this case. 942 */ 943 pr_warn("EEH: This PCI device has failed %d times in the last hour and will be permanently disabled after %d failures.\n", 944 pe->freeze_count, eeh_max_freezes); 945 pr_info("EEH: Notify device drivers to shutdown\n"); 946 eeh_set_channel_state(pe, pci_channel_io_frozen); 947 eeh_set_irq_state(pe, false); 948 eeh_pe_report("error_detected(IO frozen)", pe, 949 eeh_report_error, &result); 950 if (result == PCI_ERS_RESULT_DISCONNECT) 951 goto recover_failed; 952 953 /* 954 * Error logged on a PHB are always fences which need a full 955 * PHB reset to clear so force that to happen. 956 */ 957 if ((pe->type & EEH_PE_PHB) && result != PCI_ERS_RESULT_NONE) 958 result = PCI_ERS_RESULT_NEED_RESET; 959 960 /* Get the current PCI slot state. This can take a long time, 961 * sometimes over 300 seconds for certain systems. 962 */ 963 rc = eeh_wait_state(pe, MAX_WAIT_FOR_RECOVERY * 1000); 964 if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) { 965 pr_warn("EEH: Permanent failure\n"); 966 goto recover_failed; 967 } 968 969 /* Since rtas may enable MMIO when posting the error log, 970 * don't post the error log until after all dev drivers 971 * have been informed. 972 */ 973 pr_info("EEH: Collect temporary log\n"); 974 eeh_slot_error_detail(pe, EEH_LOG_TEMP); 975 976 /* If all device drivers were EEH-unaware, then shut 977 * down all of the device drivers, and hope they 978 * go down willingly, without panicing the system. 979 */ 980 if (result == PCI_ERS_RESULT_NONE) { 981 pr_info("EEH: Reset with hotplug activity\n"); 982 rc = eeh_reset_device(pe, bus, NULL, false); 983 if (rc) { 984 pr_warn("%s: Unable to reset, err=%d\n", __func__, rc); 985 goto recover_failed; 986 } 987 } 988 989 /* If all devices reported they can proceed, then re-enable MMIO */ 990 if (result == PCI_ERS_RESULT_CAN_RECOVER) { 991 pr_info("EEH: Enable I/O for affected devices\n"); 992 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO); 993 if (rc < 0) 994 goto recover_failed; 995 996 if (rc) { 997 result = PCI_ERS_RESULT_NEED_RESET; 998 } else { 999 pr_info("EEH: Notify device drivers to resume I/O\n"); 1000 eeh_pe_report("mmio_enabled", pe, 1001 eeh_report_mmio_enabled, &result); 1002 } 1003 } 1004 if (result == PCI_ERS_RESULT_CAN_RECOVER) { 1005 pr_info("EEH: Enabled DMA for affected devices\n"); 1006 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA); 1007 if (rc < 0) 1008 goto recover_failed; 1009 1010 if (rc) { 1011 result = PCI_ERS_RESULT_NEED_RESET; 1012 } else { 1013 /* 1014 * We didn't do PE reset for the case. The PE 1015 * is still in frozen state. Clear it before 1016 * resuming the PE. 1017 */ 1018 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true); 1019 result = PCI_ERS_RESULT_RECOVERED; 1020 } 1021 } 1022 1023 /* If any device called out for a reset, then reset the slot */ 1024 if (result == PCI_ERS_RESULT_NEED_RESET) { 1025 pr_info("EEH: Reset without hotplug activity\n"); 1026 rc = eeh_reset_device(pe, bus, &rmv_data, true); 1027 if (rc) { 1028 pr_warn("%s: Cannot reset, err=%d\n", __func__, rc); 1029 goto recover_failed; 1030 } 1031 1032 result = PCI_ERS_RESULT_NONE; 1033 eeh_set_channel_state(pe, pci_channel_io_normal); 1034 eeh_set_irq_state(pe, true); 1035 eeh_pe_report("slot_reset", pe, eeh_report_reset, 1036 &result); 1037 } 1038 1039 if ((result == PCI_ERS_RESULT_RECOVERED) || 1040 (result == PCI_ERS_RESULT_NONE)) { 1041 /* 1042 * For those hot removed VFs, we should add back them after PF 1043 * get recovered properly. 1044 */ 1045 list_for_each_entry_safe(edev, tmp, &rmv_data.removed_vf_list, 1046 rmv_entry) { 1047 eeh_add_virt_device(edev); 1048 list_del(&edev->rmv_entry); 1049 } 1050 1051 /* Tell all device drivers that they can resume operations */ 1052 pr_info("EEH: Notify device driver to resume\n"); 1053 eeh_set_channel_state(pe, pci_channel_io_normal); 1054 eeh_set_irq_state(pe, true); 1055 eeh_pe_report("resume", pe, eeh_report_resume, NULL); 1056 eeh_for_each_pe(pe, tmp_pe) { 1057 eeh_pe_for_each_dev(tmp_pe, edev, tmp) { 1058 edev->mode &= ~EEH_DEV_NO_HANDLER; 1059 edev->in_error = false; 1060 } 1061 } 1062 1063 pr_info("EEH: Recovery successful.\n"); 1064 goto out; 1065 } 1066 1067 recover_failed: 1068 /* 1069 * About 90% of all real-life EEH failures in the field 1070 * are due to poorly seated PCI cards. Only 10% or so are 1071 * due to actual, failed cards. 1072 */ 1073 pr_err("EEH: Unable to recover from failure from PHB#%x-PE#%x.\n" 1074 "Please try reseating or replacing it\n", 1075 pe->phb->global_number, pe->addr); 1076 1077 eeh_slot_error_detail(pe, EEH_LOG_PERM); 1078 1079 /* Notify all devices that they're about to go down. */ 1080 eeh_set_irq_state(pe, false); 1081 eeh_pe_report("error_detected(permanent failure)", pe, 1082 eeh_report_failure, NULL); 1083 eeh_set_channel_state(pe, pci_channel_io_perm_failure); 1084 1085 /* Mark the PE to be removed permanently */ 1086 eeh_pe_state_mark(pe, EEH_PE_REMOVED); 1087 1088 /* 1089 * Shut down the device drivers for good. We mark 1090 * all removed devices correctly to avoid access 1091 * the their PCI config any more. 1092 */ 1093 if (pe->type & EEH_PE_VF) { 1094 eeh_pe_dev_traverse(pe, eeh_rmv_device, NULL); 1095 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED); 1096 } else { 1097 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true); 1098 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED); 1099 1100 bus = eeh_pe_bus_get_nolock(pe); 1101 if (bus) 1102 pci_hp_remove_devices(bus); 1103 else 1104 pr_err("%s: PCI bus for PHB#%x-PE#%x disappeared\n", 1105 __func__, pe->phb->global_number, pe->addr); 1106 1107 /* The passed PE should no longer be used */ 1108 pci_unlock_rescan_remove(); 1109 return; 1110 } 1111 1112 out: 1113 /* 1114 * Clean up any PEs without devices. While marked as EEH_PE_RECOVERYING 1115 * we don't want to modify the PE tree structure so we do it here. 1116 */ 1117 eeh_pe_cleanup(pe); 1118 1119 /* clear the slot attention LED for all recovered devices */ 1120 eeh_for_each_pe(pe, tmp_pe) 1121 eeh_pe_for_each_dev(tmp_pe, edev, tmp) 1122 eeh_clear_slot_attention(edev->pdev); 1123 1124 eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true); 1125 1126 pci_unlock_rescan_remove(); 1127 } 1128 1129 /** 1130 * eeh_handle_special_event - Handle EEH events without a specific failing PE 1131 * 1132 * Called when an EEH event is detected but can't be narrowed down to a 1133 * specific PE. Iterates through possible failures and handles them as 1134 * necessary. 1135 */ 1136 void eeh_handle_special_event(void) 1137 { 1138 struct eeh_pe *pe, *phb_pe, *tmp_pe; 1139 struct eeh_dev *edev, *tmp_edev; 1140 struct pci_bus *bus; 1141 struct pci_controller *hose; 1142 unsigned long flags; 1143 int rc; 1144 1145 pci_lock_rescan_remove(); 1146 1147 do { 1148 rc = eeh_ops->next_error(&pe); 1149 1150 switch (rc) { 1151 case EEH_NEXT_ERR_DEAD_IOC: 1152 /* Mark all PHBs in dead state */ 1153 eeh_serialize_lock(&flags); 1154 1155 /* Purge all events */ 1156 eeh_remove_event(NULL, true); 1157 1158 list_for_each_entry(hose, &hose_list, list_node) { 1159 phb_pe = eeh_phb_pe_get(hose); 1160 if (!phb_pe) continue; 1161 1162 eeh_pe_mark_isolated(phb_pe); 1163 } 1164 1165 eeh_serialize_unlock(flags); 1166 1167 break; 1168 case EEH_NEXT_ERR_FROZEN_PE: 1169 case EEH_NEXT_ERR_FENCED_PHB: 1170 case EEH_NEXT_ERR_DEAD_PHB: 1171 /* Mark the PE in fenced state */ 1172 eeh_serialize_lock(&flags); 1173 1174 /* Purge all events of the PHB */ 1175 eeh_remove_event(pe, true); 1176 1177 if (rc != EEH_NEXT_ERR_DEAD_PHB) 1178 eeh_pe_state_mark(pe, EEH_PE_RECOVERING); 1179 eeh_pe_mark_isolated(pe); 1180 1181 eeh_serialize_unlock(flags); 1182 1183 break; 1184 case EEH_NEXT_ERR_NONE: 1185 pci_unlock_rescan_remove(); 1186 return; 1187 default: 1188 pr_warn("%s: Invalid value %d from next_error()\n", 1189 __func__, rc); 1190 pci_unlock_rescan_remove(); 1191 return; 1192 } 1193 1194 /* 1195 * For fenced PHB and frozen PE, it's handled as normal 1196 * event. We have to remove the affected PHBs for dead 1197 * PHB and IOC 1198 */ 1199 if (rc == EEH_NEXT_ERR_FROZEN_PE || 1200 rc == EEH_NEXT_ERR_FENCED_PHB) { 1201 eeh_pe_state_mark(pe, EEH_PE_RECOVERING); 1202 pci_unlock_rescan_remove(); 1203 eeh_handle_normal_event(pe); 1204 pci_lock_rescan_remove(); 1205 } else { 1206 eeh_for_each_pe(pe, tmp_pe) 1207 eeh_pe_for_each_dev(tmp_pe, edev, tmp_edev) 1208 edev->mode &= ~EEH_DEV_NO_HANDLER; 1209 1210 /* Notify all devices to be down */ 1211 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true); 1212 eeh_pe_report( 1213 "error_detected(permanent failure)", pe, 1214 eeh_report_failure, NULL); 1215 eeh_set_channel_state(pe, pci_channel_io_perm_failure); 1216 1217 list_for_each_entry(hose, &hose_list, list_node) { 1218 phb_pe = eeh_phb_pe_get(hose); 1219 if (!phb_pe || 1220 !(phb_pe->state & EEH_PE_ISOLATED) || 1221 (phb_pe->state & EEH_PE_RECOVERING)) 1222 continue; 1223 1224 bus = eeh_pe_bus_get_nolock(phb_pe); 1225 if (!bus) { 1226 pr_err("%s: Cannot find PCI bus for " 1227 "PHB#%x-PE#%x\n", 1228 __func__, 1229 pe->phb->global_number, 1230 pe->addr); 1231 break; 1232 } 1233 pci_hp_remove_devices(bus); 1234 } 1235 } 1236 1237 /* 1238 * If we have detected dead IOC, we needn't proceed 1239 * any more since all PHBs would have been removed 1240 */ 1241 if (rc == EEH_NEXT_ERR_DEAD_IOC) 1242 break; 1243 } while (rc != EEH_NEXT_ERR_NONE); 1244 1245 pci_unlock_rescan_remove(); 1246 } 1247