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);
175 * eeh_report_error - Report pci error to each device driver
177 * @userdata: return value
179 * Report an EEH error to each device driver, collect up and
180 * merge the device driver responses. Cumulative response
181 * passed back in "userdata".
183 static void *eeh_report_error(void *data, void *userdata)
185 struct eeh_dev *edev = (struct eeh_dev *)data;
186 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
187 enum pci_ers_result rc, *res = userdata;
188 struct pci_driver *driver;
190 /* We might not have the associated PCI device,
191 * then we should continue for next one.
193 if (!dev) return NULL;
194 dev->error_state = pci_channel_io_frozen;
196 driver = eeh_pcid_get(dev);
197 if (!driver) return NULL;
199 eeh_disable_irq(dev);
201 if (!driver->err_handler ||
202 !driver->err_handler->error_detected) {
207 rc = driver->err_handler->error_detected(dev, pci_channel_io_frozen);
209 /* A driver that needs a reset trumps all others */
210 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
211 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
218 * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
220 * @userdata: return value
222 * Tells each device driver that IO ports, MMIO and config space I/O
223 * are now enabled. Collects up and merges the device driver responses.
224 * Cumulative response passed back in "userdata".
226 static void *eeh_report_mmio_enabled(void *data, void *userdata)
228 struct eeh_dev *edev = (struct eeh_dev *)data;
229 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
230 enum pci_ers_result rc, *res = userdata;
231 struct pci_driver *driver;
233 driver = eeh_pcid_get(dev);
234 if (!driver) return NULL;
236 if (!driver->err_handler ||
237 !driver->err_handler->mmio_enabled ||
238 (edev->mode & EEH_DEV_NO_HANDLER)) {
243 rc = driver->err_handler->mmio_enabled(dev);
245 /* A driver that needs a reset trumps all others */
246 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
247 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
254 * eeh_report_reset - Tell device that slot has been reset
256 * @userdata: return value
258 * This routine must be called while EEH tries to reset particular
259 * PCI device so that the associated PCI device driver could take
260 * some actions, usually to save data the driver needs so that the
261 * driver can work again while the device is recovered.
263 static void *eeh_report_reset(void *data, void *userdata)
265 struct eeh_dev *edev = (struct eeh_dev *)data;
266 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
267 enum pci_ers_result rc, *res = userdata;
268 struct pci_driver *driver;
270 if (!dev) return NULL;
271 dev->error_state = pci_channel_io_normal;
273 driver = eeh_pcid_get(dev);
274 if (!driver) return NULL;
278 if (!driver->err_handler ||
279 !driver->err_handler->slot_reset ||
280 (edev->mode & EEH_DEV_NO_HANDLER)) {
285 rc = driver->err_handler->slot_reset(dev);
286 if ((*res == PCI_ERS_RESULT_NONE) ||
287 (*res == PCI_ERS_RESULT_RECOVERED)) *res = rc;
288 if (*res == PCI_ERS_RESULT_DISCONNECT &&
289 rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
296 * eeh_report_resume - Tell device to resume normal operations
298 * @userdata: return value
300 * This routine must be called to notify the device driver that it
301 * could resume so that the device driver can do some initialization
302 * to make the recovered device work again.
304 static void *eeh_report_resume(void *data, void *userdata)
306 struct eeh_dev *edev = (struct eeh_dev *)data;
307 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
308 struct pci_driver *driver;
310 if (!dev) return NULL;
311 dev->error_state = pci_channel_io_normal;
313 driver = eeh_pcid_get(dev);
314 if (!driver) return NULL;
318 if (!driver->err_handler ||
319 !driver->err_handler->resume ||
320 (edev->mode & EEH_DEV_NO_HANDLER)) {
321 edev->mode &= ~EEH_DEV_NO_HANDLER;
326 driver->err_handler->resume(dev);
333 * eeh_report_failure - Tell device driver that device is dead.
335 * @userdata: return value
337 * This informs the device driver that the device is permanently
338 * dead, and that no further recovery attempts will be made on it.
340 static void *eeh_report_failure(void *data, void *userdata)
342 struct eeh_dev *edev = (struct eeh_dev *)data;
343 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
344 struct pci_driver *driver;
346 if (!dev) return NULL;
347 dev->error_state = pci_channel_io_perm_failure;
349 driver = eeh_pcid_get(dev);
350 if (!driver) return NULL;
352 eeh_disable_irq(dev);
354 if (!driver->err_handler ||
355 !driver->err_handler->error_detected) {
360 driver->err_handler->error_detected(dev, pci_channel_io_perm_failure);
366 static void *eeh_rmv_device(void *data, void *userdata)
368 struct pci_driver *driver;
369 struct eeh_dev *edev = (struct eeh_dev *)data;
370 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
371 int *removed = (int *)userdata;
374 * Actually, we should remove the PCI bridges as well.
375 * However, that's lots of complexity to do that,
376 * particularly some of devices under the bridge might
377 * support EEH. So we just care about PCI devices for
380 if (!dev || (dev->hdr_type & PCI_HEADER_TYPE_BRIDGE))
383 driver = eeh_pcid_get(dev);
386 if (driver->err_handler)
390 /* Remove it from PCI subsystem */
391 pr_debug("EEH: Removing %s without EEH sensitive driver\n",
393 edev->bus = dev->bus;
394 edev->mode |= EEH_DEV_DISCONNECTED;
397 pci_lock_rescan_remove();
398 pci_stop_and_remove_bus_device(dev);
399 pci_unlock_rescan_remove();
404 static void *eeh_pe_detach_dev(void *data, void *userdata)
406 struct eeh_pe *pe = (struct eeh_pe *)data;
407 struct eeh_dev *edev, *tmp;
409 eeh_pe_for_each_dev(pe, edev, tmp) {
410 if (!(edev->mode & EEH_DEV_DISCONNECTED))
413 edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
414 eeh_rmv_from_parent_pe(edev);
421 * Explicitly clear PE's frozen state for PowerNV where
422 * we have frozen PE until BAR restore is completed. It's
423 * harmless to clear it for pSeries. To be consistent with
424 * PE reset (for 3 times), we try to clear the frozen state
425 * for 3 times as well.
427 static int eeh_clear_pe_frozen_state(struct eeh_pe *pe)
431 for (i = 0; i < 3; i++) {
432 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
435 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
440 /* The PE has been isolated, clear it */
442 pr_warn("%s: Can't clear frozen PHB#%x-PE#%x (%d)\n",
443 __func__, pe->phb->global_number, pe->addr, rc);
445 eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
451 * eeh_reset_device - Perform actual reset of a pci slot
453 * @bus: PCI bus corresponding to the isolcated slot
455 * This routine must be called to do reset on the indicated PE.
456 * During the reset, udev might be invoked because those affected
457 * PCI devices will be removed and then added.
459 static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus)
461 struct pci_bus *frozen_bus = eeh_pe_bus_get(pe);
462 struct timeval tstamp;
463 int cnt, rc, removed = 0;
465 /* pcibios will clear the counter; save the value */
466 cnt = pe->freeze_count;
470 * We don't remove the corresponding PE instances because
471 * we need the information afterwords. The attached EEH
472 * devices are expected to be attached soon when calling
473 * into pcibios_add_pci_devices().
475 eeh_pe_state_mark(pe, EEH_PE_KEEP);
477 pci_lock_rescan_remove();
478 pcibios_remove_pci_devices(bus);
479 pci_unlock_rescan_remove();
480 } else if (frozen_bus) {
481 eeh_pe_dev_traverse(pe, eeh_rmv_device, &removed);
485 * Reset the pci controller. (Asserts RST#; resets config space).
486 * Reconfigure bridges and devices. Don't try to bring the system
487 * up if the reset failed for some reason.
489 * During the reset, it's very dangerous to have uncontrolled PCI
490 * config accesses. So we prefer to block them. However, controlled
491 * PCI config accesses initiated from EEH itself are allowed.
493 eeh_pe_state_mark(pe, EEH_PE_RESET);
494 rc = eeh_reset_pe(pe);
496 eeh_pe_state_clear(pe, EEH_PE_RESET);
500 pci_lock_rescan_remove();
503 eeh_ops->configure_bridge(pe);
504 eeh_pe_restore_bars(pe);
505 eeh_pe_state_clear(pe, EEH_PE_RESET);
507 /* Clear frozen state */
508 rc = eeh_clear_pe_frozen_state(pe);
512 /* Give the system 5 seconds to finish running the user-space
513 * hotplug shutdown scripts, e.g. ifdown for ethernet. Yes,
514 * this is a hack, but if we don't do this, and try to bring
515 * the device up before the scripts have taken it down,
516 * potentially weird things happen.
519 pr_info("EEH: Sleep 5s ahead of complete hotplug\n");
523 * The EEH device is still connected with its parent
524 * PE. We should disconnect it so the binding can be
525 * rebuilt when adding PCI devices.
527 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
528 pcibios_add_pci_devices(bus);
529 } else if (frozen_bus && removed) {
530 pr_info("EEH: Sleep 5s ahead of partial hotplug\n");
533 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
534 pcibios_add_pci_devices(frozen_bus);
536 eeh_pe_state_clear(pe, EEH_PE_KEEP);
539 pe->freeze_count = cnt;
541 pci_unlock_rescan_remove();
545 /* The longest amount of time to wait for a pci device
546 * to come back on line, in seconds.
548 #define MAX_WAIT_FOR_RECOVERY 300
550 static void eeh_handle_normal_event(struct eeh_pe *pe)
552 struct pci_bus *frozen_bus;
554 enum pci_ers_result result = PCI_ERS_RESULT_NONE;
556 frozen_bus = eeh_pe_bus_get(pe);
558 pr_err("%s: Cannot find PCI bus for PHB#%d-PE#%x\n",
559 __func__, pe->phb->global_number, pe->addr);
563 eeh_pe_update_time_stamp(pe);
565 if (pe->freeze_count > EEH_MAX_ALLOWED_FREEZES)
566 goto excess_failures;
567 pr_warning("EEH: This PCI device has failed %d times in the last hour\n",
570 /* Walk the various device drivers attached to this slot through
571 * a reset sequence, giving each an opportunity to do what it needs
572 * to accomplish the reset. Each child gets a report of the
573 * status ... if any child can't handle the reset, then the entire
574 * slot is dlpar removed and added.
576 pr_info("EEH: Notify device drivers to shutdown\n");
577 eeh_pe_dev_traverse(pe, eeh_report_error, &result);
579 /* Get the current PCI slot state. This can take a long time,
580 * sometimes over 3 seconds for certain systems.
582 rc = eeh_ops->wait_state(pe, MAX_WAIT_FOR_RECOVERY*1000);
583 if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
584 pr_warning("EEH: Permanent failure\n");
588 /* Since rtas may enable MMIO when posting the error log,
589 * don't post the error log until after all dev drivers
590 * have been informed.
592 pr_info("EEH: Collect temporary log\n");
593 eeh_slot_error_detail(pe, EEH_LOG_TEMP);
595 /* If all device drivers were EEH-unaware, then shut
596 * down all of the device drivers, and hope they
597 * go down willingly, without panicing the system.
599 if (result == PCI_ERS_RESULT_NONE) {
600 pr_info("EEH: Reset with hotplug activity\n");
601 rc = eeh_reset_device(pe, frozen_bus);
603 pr_warning("%s: Unable to reset, err=%d\n",
609 /* If all devices reported they can proceed, then re-enable MMIO */
610 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
611 pr_info("EEH: Enable I/O for affected devices\n");
612 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
617 result = PCI_ERS_RESULT_NEED_RESET;
619 pr_info("EEH: Notify device drivers to resume I/O\n");
620 result = PCI_ERS_RESULT_NONE;
621 eeh_pe_dev_traverse(pe, eeh_report_mmio_enabled, &result);
625 /* If all devices reported they can proceed, then re-enable DMA */
626 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
627 pr_info("EEH: Enabled DMA for affected devices\n");
628 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
633 result = PCI_ERS_RESULT_NEED_RESET;
635 result = PCI_ERS_RESULT_RECOVERED;
638 /* If any device has a hard failure, then shut off everything. */
639 if (result == PCI_ERS_RESULT_DISCONNECT) {
640 pr_warning("EEH: Device driver gave up\n");
644 /* If any device called out for a reset, then reset the slot */
645 if (result == PCI_ERS_RESULT_NEED_RESET) {
646 pr_info("EEH: Reset without hotplug activity\n");
647 rc = eeh_reset_device(pe, NULL);
649 pr_warning("%s: Cannot reset, err=%d\n",
654 pr_info("EEH: Notify device drivers "
655 "the completion of reset\n");
656 result = PCI_ERS_RESULT_NONE;
657 eeh_pe_dev_traverse(pe, eeh_report_reset, &result);
660 /* All devices should claim they have recovered by now. */
661 if ((result != PCI_ERS_RESULT_RECOVERED) &&
662 (result != PCI_ERS_RESULT_NONE)) {
663 pr_warning("EEH: Not recovered\n");
667 /* Tell all device drivers that they can resume operations */
668 pr_info("EEH: Notify device driver to resume\n");
669 eeh_pe_dev_traverse(pe, eeh_report_resume, NULL);
675 * About 90% of all real-life EEH failures in the field
676 * are due to poorly seated PCI cards. Only 10% or so are
677 * due to actual, failed cards.
679 pr_err("EEH: PHB#%d-PE#%x has failed %d times in the\n"
680 "last hour and has been permanently disabled.\n"
681 "Please try reseating or replacing it.\n",
682 pe->phb->global_number, pe->addr,
687 pr_err("EEH: Unable to recover from failure from PHB#%d-PE#%x.\n"
688 "Please try reseating or replacing it\n",
689 pe->phb->global_number, pe->addr);
692 eeh_slot_error_detail(pe, EEH_LOG_PERM);
694 /* Notify all devices that they're about to go down. */
695 eeh_pe_dev_traverse(pe, eeh_report_failure, NULL);
697 /* Shut down the device drivers for good. */
699 pci_lock_rescan_remove();
700 pcibios_remove_pci_devices(frozen_bus);
701 pci_unlock_rescan_remove();
705 static void eeh_handle_special_event(void)
707 struct eeh_pe *pe, *phb_pe;
709 struct pci_controller *hose;
715 rc = eeh_ops->next_error(&pe);
718 case EEH_NEXT_ERR_DEAD_IOC:
719 /* Mark all PHBs in dead state */
720 eeh_serialize_lock(&flags);
722 /* Purge all events */
723 eeh_remove_event(NULL);
725 list_for_each_entry(hose, &hose_list, list_node) {
726 phb_pe = eeh_phb_pe_get(hose);
727 if (!phb_pe) continue;
729 eeh_pe_state_mark(phb_pe, EEH_PE_ISOLATED);
732 eeh_serialize_unlock(flags);
735 case EEH_NEXT_ERR_FROZEN_PE:
736 case EEH_NEXT_ERR_FENCED_PHB:
737 case EEH_NEXT_ERR_DEAD_PHB:
738 /* Mark the PE in fenced state */
739 eeh_serialize_lock(&flags);
741 /* Purge all events of the PHB */
742 eeh_remove_event(pe);
744 if (rc == EEH_NEXT_ERR_DEAD_PHB)
745 eeh_pe_state_mark(pe, EEH_PE_ISOLATED);
747 eeh_pe_state_mark(pe,
748 EEH_PE_ISOLATED | EEH_PE_RECOVERING);
750 eeh_serialize_unlock(flags);
753 case EEH_NEXT_ERR_NONE:
756 pr_warn("%s: Invalid value %d from next_error()\n",
762 * For fenced PHB and frozen PE, it's handled as normal
763 * event. We have to remove the affected PHBs for dead
766 if (rc == EEH_NEXT_ERR_FROZEN_PE ||
767 rc == EEH_NEXT_ERR_FENCED_PHB) {
768 eeh_handle_normal_event(pe);
769 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
771 pci_lock_rescan_remove();
772 list_for_each_entry(hose, &hose_list, list_node) {
773 phb_pe = eeh_phb_pe_get(hose);
775 !(phb_pe->state & EEH_PE_ISOLATED) ||
776 (phb_pe->state & EEH_PE_RECOVERING))
779 /* Notify all devices to be down */
780 bus = eeh_pe_bus_get(phb_pe);
781 eeh_pe_dev_traverse(pe,
782 eeh_report_failure, NULL);
783 pcibios_remove_pci_devices(bus);
785 pci_unlock_rescan_remove();
789 * If we have detected dead IOC, we needn't proceed
790 * any more since all PHBs would have been removed
792 if (rc == EEH_NEXT_ERR_DEAD_IOC)
794 } while (rc != EEH_NEXT_ERR_NONE);
798 * eeh_handle_event - Reset a PCI device after hard lockup.
801 * While PHB detects address or data parity errors on particular PCI
802 * slot, the associated PE will be frozen. Besides, DMA's occurring
803 * to wild addresses (which usually happen due to bugs in device
804 * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
805 * #PERR or other misc PCI-related errors also can trigger EEH errors.
807 * Recovery process consists of unplugging the device driver (which
808 * generated hotplug events to userspace), then issuing a PCI #RST to
809 * the device, then reconfiguring the PCI config space for all bridges
810 * & devices under this slot, and then finally restarting the device
811 * drivers (which cause a second set of hotplug events to go out to
814 void eeh_handle_event(struct eeh_pe *pe)
817 eeh_handle_normal_event(pe);
819 eeh_handle_special_event();