2 * PCI Error Recovery Driver for RPA-compliant PPC64 platform.
3 * Copyright IBM Corp. 2004 2005
4 * Copyright Linas Vepstas <linas@linas.org> 2004, 2005
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or (at
11 * your option) any later version.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
16 * NON INFRINGEMENT. See the GNU General Public License for more
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 * Send comments and feedback to Linas Vepstas <linas@austin.ibm.com>
25 #include <linux/delay.h>
26 #include <linux/interrupt.h>
27 #include <linux/irq.h>
28 #include <linux/module.h>
29 #include <linux/pci.h>
31 #include <asm/eeh_event.h>
32 #include <asm/ppc-pci.h>
33 #include <asm/pci-bridge.h>
38 * eeh_pcid_name - Retrieve name of PCI device driver
41 * This routine is used to retrieve the name of PCI device driver
44 static inline const char *eeh_pcid_name(struct pci_dev *pdev)
46 if (pdev && pdev->dev.driver)
47 return pdev->dev.driver->name;
52 * eeh_pcid_get - Get the PCI device driver
55 * The function is used to retrieve the PCI device driver for
56 * the indicated PCI device. Besides, we will increase the reference
57 * of the PCI device driver to prevent that being unloaded on
58 * the fly. Otherwise, kernel crash would be seen.
60 static inline struct pci_driver *eeh_pcid_get(struct pci_dev *pdev)
62 if (!pdev || !pdev->driver)
65 if (!try_module_get(pdev->driver->driver.owner))
72 * eeh_pcid_put - Dereference on the PCI device driver
75 * The function is called to do dereference on the PCI device
76 * driver of the indicated PCI device.
78 static inline void eeh_pcid_put(struct pci_dev *pdev)
80 if (!pdev || !pdev->driver)
83 module_put(pdev->driver->driver.owner);
87 static void print_device_node_tree(struct pci_dn *pdn, int dent)
90 struct device_node *pc;
94 for (i = 0; i < dent; i++)
96 printk("dn=%s mode=%x \tcfg_addr=%x pe_addr=%x \tfull=%s\n",
97 pdn->node->name, pdn->eeh_mode, pdn->eeh_config_addr,
98 pdn->eeh_pe_config_addr, pdn->node->full_name);
100 pc = pdn->node->child;
102 print_device_node_tree(PCI_DN(pc), dent);
109 * eeh_disable_irq - Disable interrupt for the recovering device
112 * This routine must be called when reporting temporary or permanent
113 * error to the particular PCI device to disable interrupt of that
114 * device. If the device has enabled MSI or MSI-X interrupt, we needn't
115 * do real work because EEH should freeze DMA transfers for those PCI
116 * devices encountering EEH errors, which includes MSI or MSI-X.
118 static void eeh_disable_irq(struct pci_dev *dev)
120 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
122 /* Don't disable MSI and MSI-X interrupts. They are
123 * effectively disabled by the DMA Stopped state
124 * when an EEH error occurs.
126 if (dev->msi_enabled || dev->msix_enabled)
129 if (!irq_has_action(dev->irq))
132 edev->mode |= EEH_DEV_IRQ_DISABLED;
133 disable_irq_nosync(dev->irq);
137 * eeh_enable_irq - Enable interrupt for the recovering device
140 * This routine must be called to enable interrupt while failed
141 * device could be resumed.
143 static void eeh_enable_irq(struct pci_dev *dev)
145 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
147 if ((edev->mode) & EEH_DEV_IRQ_DISABLED) {
148 edev->mode &= ~EEH_DEV_IRQ_DISABLED;
152 * This is just ass backwards. This maze has
153 * unbalanced irq_enable/disable calls. So instead of
154 * finding the root cause it works around the warning
155 * in the irq_enable code by conditionally calling
158 * That's just wrong.The warning in the core code is
159 * there to tell people to fix their assymetries in
160 * their own code, not by abusing the core information
163 * I so wish that the assymetry would be the other way
164 * round and a few more irq_disable calls render that
165 * shit unusable forever.
169 if (irqd_irq_disabled(irq_get_irq_data(dev->irq)))
170 enable_irq(dev->irq);
174 static bool eeh_dev_removed(struct eeh_dev *edev)
176 /* EEH device removed ? */
177 if (!edev || (edev->mode & EEH_DEV_REMOVED))
184 * eeh_report_error - Report pci error to each device driver
186 * @userdata: return value
188 * Report an EEH error to each device driver, collect up and
189 * merge the device driver responses. Cumulative response
190 * passed back in "userdata".
192 static void *eeh_report_error(void *data, void *userdata)
194 struct eeh_dev *edev = (struct eeh_dev *)data;
195 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
196 enum pci_ers_result rc, *res = userdata;
197 struct pci_driver *driver;
199 if (!dev || eeh_dev_removed(edev))
201 dev->error_state = pci_channel_io_frozen;
203 driver = eeh_pcid_get(dev);
204 if (!driver) return NULL;
206 eeh_disable_irq(dev);
208 if (!driver->err_handler ||
209 !driver->err_handler->error_detected) {
214 rc = driver->err_handler->error_detected(dev, pci_channel_io_frozen);
216 /* A driver that needs a reset trumps all others */
217 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
218 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
225 * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
227 * @userdata: return value
229 * Tells each device driver that IO ports, MMIO and config space I/O
230 * are now enabled. Collects up and merges the device driver responses.
231 * Cumulative response passed back in "userdata".
233 static void *eeh_report_mmio_enabled(void *data, void *userdata)
235 struct eeh_dev *edev = (struct eeh_dev *)data;
236 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
237 enum pci_ers_result rc, *res = userdata;
238 struct pci_driver *driver;
240 if (!dev || eeh_dev_removed(edev))
243 driver = eeh_pcid_get(dev);
244 if (!driver) return NULL;
246 if (!driver->err_handler ||
247 !driver->err_handler->mmio_enabled ||
248 (edev->mode & EEH_DEV_NO_HANDLER)) {
253 rc = driver->err_handler->mmio_enabled(dev);
255 /* A driver that needs a reset trumps all others */
256 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
257 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
264 * eeh_report_reset - Tell device that slot has been reset
266 * @userdata: return value
268 * This routine must be called while EEH tries to reset particular
269 * PCI device so that the associated PCI device driver could take
270 * some actions, usually to save data the driver needs so that the
271 * driver can work again while the device is recovered.
273 static void *eeh_report_reset(void *data, void *userdata)
275 struct eeh_dev *edev = (struct eeh_dev *)data;
276 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
277 enum pci_ers_result rc, *res = userdata;
278 struct pci_driver *driver;
280 if (!dev || eeh_dev_removed(edev))
282 dev->error_state = pci_channel_io_normal;
284 driver = eeh_pcid_get(dev);
285 if (!driver) return NULL;
289 if (!driver->err_handler ||
290 !driver->err_handler->slot_reset ||
291 (edev->mode & EEH_DEV_NO_HANDLER)) {
296 rc = driver->err_handler->slot_reset(dev);
297 if ((*res == PCI_ERS_RESULT_NONE) ||
298 (*res == PCI_ERS_RESULT_RECOVERED)) *res = rc;
299 if (*res == PCI_ERS_RESULT_DISCONNECT &&
300 rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
307 * eeh_report_resume - Tell device to resume normal operations
309 * @userdata: return value
311 * This routine must be called to notify the device driver that it
312 * could resume so that the device driver can do some initialization
313 * to make the recovered device work again.
315 static void *eeh_report_resume(void *data, void *userdata)
317 struct eeh_dev *edev = (struct eeh_dev *)data;
318 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
319 struct pci_driver *driver;
321 if (!dev || eeh_dev_removed(edev))
323 dev->error_state = pci_channel_io_normal;
325 driver = eeh_pcid_get(dev);
326 if (!driver) return NULL;
330 if (!driver->err_handler ||
331 !driver->err_handler->resume ||
332 (edev->mode & EEH_DEV_NO_HANDLER)) {
333 edev->mode &= ~EEH_DEV_NO_HANDLER;
338 driver->err_handler->resume(dev);
345 * eeh_report_failure - Tell device driver that device is dead.
347 * @userdata: return value
349 * This informs the device driver that the device is permanently
350 * dead, and that no further recovery attempts will be made on it.
352 static void *eeh_report_failure(void *data, void *userdata)
354 struct eeh_dev *edev = (struct eeh_dev *)data;
355 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
356 struct pci_driver *driver;
358 if (!dev || eeh_dev_removed(edev))
360 dev->error_state = pci_channel_io_perm_failure;
362 driver = eeh_pcid_get(dev);
363 if (!driver) return NULL;
365 eeh_disable_irq(dev);
367 if (!driver->err_handler ||
368 !driver->err_handler->error_detected) {
373 driver->err_handler->error_detected(dev, pci_channel_io_perm_failure);
379 static void *eeh_rmv_device(void *data, void *userdata)
381 struct pci_driver *driver;
382 struct eeh_dev *edev = (struct eeh_dev *)data;
383 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
384 int *removed = (int *)userdata;
387 * Actually, we should remove the PCI bridges as well.
388 * However, that's lots of complexity to do that,
389 * particularly some of devices under the bridge might
390 * support EEH. So we just care about PCI devices for
393 if (!dev || (dev->hdr_type & PCI_HEADER_TYPE_BRIDGE))
397 * We rely on count-based pcibios_release_device() to
398 * detach permanently offlined PEs. Unfortunately, that's
399 * not reliable enough. We might have the permanently
400 * offlined PEs attached, but we needn't take care of
401 * them and their child devices.
403 if (eeh_dev_removed(edev))
406 driver = eeh_pcid_get(dev);
409 if (driver->err_handler)
413 /* Remove it from PCI subsystem */
414 pr_debug("EEH: Removing %s without EEH sensitive driver\n",
416 edev->bus = dev->bus;
417 edev->mode |= EEH_DEV_DISCONNECTED;
420 pci_lock_rescan_remove();
421 pci_stop_and_remove_bus_device(dev);
422 pci_unlock_rescan_remove();
427 static void *eeh_pe_detach_dev(void *data, void *userdata)
429 struct eeh_pe *pe = (struct eeh_pe *)data;
430 struct eeh_dev *edev, *tmp;
432 eeh_pe_for_each_dev(pe, edev, tmp) {
433 if (!(edev->mode & EEH_DEV_DISCONNECTED))
436 edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
437 eeh_rmv_from_parent_pe(edev);
444 * Explicitly clear PE's frozen state for PowerNV where
445 * we have frozen PE until BAR restore is completed. It's
446 * harmless to clear it for pSeries. To be consistent with
447 * PE reset (for 3 times), we try to clear the frozen state
448 * for 3 times as well.
450 static int eeh_clear_pe_frozen_state(struct eeh_pe *pe)
454 for (i = 0; i < 3; i++) {
455 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
458 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
463 /* The PE has been isolated, clear it */
465 pr_warn("%s: Can't clear frozen PHB#%x-PE#%x (%d)\n",
466 __func__, pe->phb->global_number, pe->addr, rc);
468 eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
474 * eeh_reset_device - Perform actual reset of a pci slot
476 * @bus: PCI bus corresponding to the isolcated slot
478 * This routine must be called to do reset on the indicated PE.
479 * During the reset, udev might be invoked because those affected
480 * PCI devices will be removed and then added.
482 static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus)
484 struct pci_bus *frozen_bus = eeh_pe_bus_get(pe);
485 struct timeval tstamp;
486 int cnt, rc, removed = 0;
488 /* pcibios will clear the counter; save the value */
489 cnt = pe->freeze_count;
493 * We don't remove the corresponding PE instances because
494 * we need the information afterwords. The attached EEH
495 * devices are expected to be attached soon when calling
496 * into pcibios_add_pci_devices().
498 eeh_pe_state_mark(pe, EEH_PE_KEEP);
500 pci_lock_rescan_remove();
501 pcibios_remove_pci_devices(bus);
502 pci_unlock_rescan_remove();
503 } else if (frozen_bus) {
504 eeh_pe_dev_traverse(pe, eeh_rmv_device, &removed);
508 * Reset the pci controller. (Asserts RST#; resets config space).
509 * Reconfigure bridges and devices. Don't try to bring the system
510 * up if the reset failed for some reason.
512 * During the reset, it's very dangerous to have uncontrolled PCI
513 * config accesses. So we prefer to block them. However, controlled
514 * PCI config accesses initiated from EEH itself are allowed.
516 eeh_pe_state_mark(pe, EEH_PE_RESET);
517 rc = eeh_reset_pe(pe);
519 eeh_pe_state_clear(pe, EEH_PE_RESET);
523 pci_lock_rescan_remove();
526 eeh_ops->configure_bridge(pe);
527 eeh_pe_restore_bars(pe);
528 eeh_pe_state_clear(pe, EEH_PE_RESET);
530 /* Clear frozen state */
531 rc = eeh_clear_pe_frozen_state(pe);
535 /* Give the system 5 seconds to finish running the user-space
536 * hotplug shutdown scripts, e.g. ifdown for ethernet. Yes,
537 * this is a hack, but if we don't do this, and try to bring
538 * the device up before the scripts have taken it down,
539 * potentially weird things happen.
542 pr_info("EEH: Sleep 5s ahead of complete hotplug\n");
546 * The EEH device is still connected with its parent
547 * PE. We should disconnect it so the binding can be
548 * rebuilt when adding PCI devices.
550 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
551 pcibios_add_pci_devices(bus);
552 } else if (frozen_bus && removed) {
553 pr_info("EEH: Sleep 5s ahead of partial hotplug\n");
556 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
557 pcibios_add_pci_devices(frozen_bus);
559 eeh_pe_state_clear(pe, EEH_PE_KEEP);
562 pe->freeze_count = cnt;
564 pci_unlock_rescan_remove();
568 /* The longest amount of time to wait for a pci device
569 * to come back on line, in seconds.
571 #define MAX_WAIT_FOR_RECOVERY 300
573 static void eeh_handle_normal_event(struct eeh_pe *pe)
575 struct pci_bus *frozen_bus;
577 enum pci_ers_result result = PCI_ERS_RESULT_NONE;
579 frozen_bus = eeh_pe_bus_get(pe);
581 pr_err("%s: Cannot find PCI bus for PHB#%d-PE#%x\n",
582 __func__, pe->phb->global_number, pe->addr);
586 eeh_pe_update_time_stamp(pe);
588 if (pe->freeze_count > EEH_MAX_ALLOWED_FREEZES)
589 goto excess_failures;
590 pr_warning("EEH: This PCI device has failed %d times in the last hour\n",
593 /* Walk the various device drivers attached to this slot through
594 * a reset sequence, giving each an opportunity to do what it needs
595 * to accomplish the reset. Each child gets a report of the
596 * status ... if any child can't handle the reset, then the entire
597 * slot is dlpar removed and added.
599 pr_info("EEH: Notify device drivers to shutdown\n");
600 eeh_pe_dev_traverse(pe, eeh_report_error, &result);
602 /* Get the current PCI slot state. This can take a long time,
603 * sometimes over 3 seconds for certain systems.
605 rc = eeh_ops->wait_state(pe, MAX_WAIT_FOR_RECOVERY*1000);
606 if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
607 pr_warning("EEH: Permanent failure\n");
611 /* Since rtas may enable MMIO when posting the error log,
612 * don't post the error log until after all dev drivers
613 * have been informed.
615 pr_info("EEH: Collect temporary log\n");
616 eeh_slot_error_detail(pe, EEH_LOG_TEMP);
618 /* If all device drivers were EEH-unaware, then shut
619 * down all of the device drivers, and hope they
620 * go down willingly, without panicing the system.
622 if (result == PCI_ERS_RESULT_NONE) {
623 pr_info("EEH: Reset with hotplug activity\n");
624 rc = eeh_reset_device(pe, frozen_bus);
626 pr_warning("%s: Unable to reset, err=%d\n",
632 /* If all devices reported they can proceed, then re-enable MMIO */
633 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
634 pr_info("EEH: Enable I/O for affected devices\n");
635 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
640 result = PCI_ERS_RESULT_NEED_RESET;
642 pr_info("EEH: Notify device drivers to resume I/O\n");
643 eeh_pe_dev_traverse(pe, eeh_report_mmio_enabled, &result);
647 /* If all devices reported they can proceed, then re-enable DMA */
648 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
649 pr_info("EEH: Enabled DMA for affected devices\n");
650 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
655 result = PCI_ERS_RESULT_NEED_RESET;
658 * We didn't do PE reset for the case. The PE
659 * is still in frozen state. Clear it before
662 eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
663 result = PCI_ERS_RESULT_RECOVERED;
667 /* If any device has a hard failure, then shut off everything. */
668 if (result == PCI_ERS_RESULT_DISCONNECT) {
669 pr_warning("EEH: Device driver gave up\n");
673 /* If any device called out for a reset, then reset the slot */
674 if (result == PCI_ERS_RESULT_NEED_RESET) {
675 pr_info("EEH: Reset without hotplug activity\n");
676 rc = eeh_reset_device(pe, NULL);
678 pr_warning("%s: Cannot reset, err=%d\n",
683 pr_info("EEH: Notify device drivers "
684 "the completion of reset\n");
685 result = PCI_ERS_RESULT_NONE;
686 eeh_pe_dev_traverse(pe, eeh_report_reset, &result);
689 /* All devices should claim they have recovered by now. */
690 if ((result != PCI_ERS_RESULT_RECOVERED) &&
691 (result != PCI_ERS_RESULT_NONE)) {
692 pr_warning("EEH: Not recovered\n");
696 /* Tell all device drivers that they can resume operations */
697 pr_info("EEH: Notify device driver to resume\n");
698 eeh_pe_dev_traverse(pe, eeh_report_resume, NULL);
704 * About 90% of all real-life EEH failures in the field
705 * are due to poorly seated PCI cards. Only 10% or so are
706 * due to actual, failed cards.
708 pr_err("EEH: PHB#%d-PE#%x has failed %d times in the\n"
709 "last hour and has been permanently disabled.\n"
710 "Please try reseating or replacing it.\n",
711 pe->phb->global_number, pe->addr,
716 pr_err("EEH: Unable to recover from failure from PHB#%d-PE#%x.\n"
717 "Please try reseating or replacing it\n",
718 pe->phb->global_number, pe->addr);
721 eeh_slot_error_detail(pe, EEH_LOG_PERM);
723 /* Notify all devices that they're about to go down. */
724 eeh_pe_dev_traverse(pe, eeh_report_failure, NULL);
726 /* Mark the PE to be removed permanently */
727 pe->freeze_count = EEH_MAX_ALLOWED_FREEZES + 1;
730 * Shut down the device drivers for good. We mark
731 * all removed devices correctly to avoid access
732 * the their PCI config any more.
735 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
737 pci_lock_rescan_remove();
738 pcibios_remove_pci_devices(frozen_bus);
739 pci_unlock_rescan_remove();
743 static void eeh_handle_special_event(void)
745 struct eeh_pe *pe, *phb_pe;
747 struct pci_controller *hose;
753 rc = eeh_ops->next_error(&pe);
756 case EEH_NEXT_ERR_DEAD_IOC:
757 /* Mark all PHBs in dead state */
758 eeh_serialize_lock(&flags);
760 /* Purge all events */
761 eeh_remove_event(NULL);
763 list_for_each_entry(hose, &hose_list, list_node) {
764 phb_pe = eeh_phb_pe_get(hose);
765 if (!phb_pe) continue;
767 eeh_pe_state_mark(phb_pe, EEH_PE_ISOLATED);
770 eeh_serialize_unlock(flags);
773 case EEH_NEXT_ERR_FROZEN_PE:
774 case EEH_NEXT_ERR_FENCED_PHB:
775 case EEH_NEXT_ERR_DEAD_PHB:
776 /* Mark the PE in fenced state */
777 eeh_serialize_lock(&flags);
779 /* Purge all events of the PHB */
780 eeh_remove_event(pe);
782 if (rc == EEH_NEXT_ERR_DEAD_PHB)
783 eeh_pe_state_mark(pe, EEH_PE_ISOLATED);
785 eeh_pe_state_mark(pe,
786 EEH_PE_ISOLATED | EEH_PE_RECOVERING);
788 eeh_serialize_unlock(flags);
791 case EEH_NEXT_ERR_NONE:
794 pr_warn("%s: Invalid value %d from next_error()\n",
800 * For fenced PHB and frozen PE, it's handled as normal
801 * event. We have to remove the affected PHBs for dead
804 if (rc == EEH_NEXT_ERR_FROZEN_PE ||
805 rc == EEH_NEXT_ERR_FENCED_PHB) {
806 eeh_handle_normal_event(pe);
807 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
809 pci_lock_rescan_remove();
810 list_for_each_entry(hose, &hose_list, list_node) {
811 phb_pe = eeh_phb_pe_get(hose);
813 !(phb_pe->state & EEH_PE_ISOLATED) ||
814 (phb_pe->state & EEH_PE_RECOVERING))
817 /* Notify all devices to be down */
818 bus = eeh_pe_bus_get(phb_pe);
819 eeh_pe_dev_traverse(pe,
820 eeh_report_failure, NULL);
821 pcibios_remove_pci_devices(bus);
823 pci_unlock_rescan_remove();
827 * If we have detected dead IOC, we needn't proceed
828 * any more since all PHBs would have been removed
830 if (rc == EEH_NEXT_ERR_DEAD_IOC)
832 } while (rc != EEH_NEXT_ERR_NONE);
836 * eeh_handle_event - Reset a PCI device after hard lockup.
839 * While PHB detects address or data parity errors on particular PCI
840 * slot, the associated PE will be frozen. Besides, DMA's occurring
841 * to wild addresses (which usually happen due to bugs in device
842 * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
843 * #PERR or other misc PCI-related errors also can trigger EEH errors.
845 * Recovery process consists of unplugging the device driver (which
846 * generated hotplug events to userspace), then issuing a PCI #RST to
847 * the device, then reconfiguring the PCI config space for all bridges
848 * & devices under this slot, and then finally restarting the device
849 * drivers (which cause a second set of hotplug events to go out to
852 void eeh_handle_event(struct eeh_pe *pe)
855 eeh_handle_normal_event(pe);
857 eeh_handle_special_event();