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);
146 struct irq_desc *desc;
148 if ((edev->mode) & EEH_DEV_IRQ_DISABLED) {
149 edev->mode &= ~EEH_DEV_IRQ_DISABLED;
151 desc = irq_to_desc(dev->irq);
152 if (desc && desc->depth > 0)
153 enable_irq(dev->irq);
158 * eeh_report_error - Report pci error to each device driver
160 * @userdata: return value
162 * Report an EEH error to each device driver, collect up and
163 * merge the device driver responses. Cumulative response
164 * passed back in "userdata".
166 static void *eeh_report_error(void *data, void *userdata)
168 struct eeh_dev *edev = (struct eeh_dev *)data;
169 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
170 enum pci_ers_result rc, *res = userdata;
171 struct pci_driver *driver;
173 /* We might not have the associated PCI device,
174 * then we should continue for next one.
176 if (!dev) return NULL;
177 dev->error_state = pci_channel_io_frozen;
179 driver = eeh_pcid_get(dev);
180 if (!driver) return NULL;
182 eeh_disable_irq(dev);
184 if (!driver->err_handler ||
185 !driver->err_handler->error_detected) {
190 rc = driver->err_handler->error_detected(dev, pci_channel_io_frozen);
192 /* A driver that needs a reset trumps all others */
193 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
194 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
201 * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
203 * @userdata: return value
205 * Tells each device driver that IO ports, MMIO and config space I/O
206 * are now enabled. Collects up and merges the device driver responses.
207 * Cumulative response passed back in "userdata".
209 static void *eeh_report_mmio_enabled(void *data, void *userdata)
211 struct eeh_dev *edev = (struct eeh_dev *)data;
212 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
213 enum pci_ers_result rc, *res = userdata;
214 struct pci_driver *driver;
216 driver = eeh_pcid_get(dev);
217 if (!driver) return NULL;
219 if (!driver->err_handler ||
220 !driver->err_handler->mmio_enabled) {
225 rc = driver->err_handler->mmio_enabled(dev);
227 /* A driver that needs a reset trumps all others */
228 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
229 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
236 * eeh_report_reset - Tell device that slot has been reset
238 * @userdata: return value
240 * This routine must be called while EEH tries to reset particular
241 * PCI device so that the associated PCI device driver could take
242 * some actions, usually to save data the driver needs so that the
243 * driver can work again while the device is recovered.
245 static void *eeh_report_reset(void *data, void *userdata)
247 struct eeh_dev *edev = (struct eeh_dev *)data;
248 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
249 enum pci_ers_result rc, *res = userdata;
250 struct pci_driver *driver;
252 if (!dev) return NULL;
253 dev->error_state = pci_channel_io_normal;
255 driver = eeh_pcid_get(dev);
256 if (!driver) return NULL;
260 if (!driver->err_handler ||
261 !driver->err_handler->slot_reset) {
266 rc = driver->err_handler->slot_reset(dev);
267 if ((*res == PCI_ERS_RESULT_NONE) ||
268 (*res == PCI_ERS_RESULT_RECOVERED)) *res = rc;
269 if (*res == PCI_ERS_RESULT_DISCONNECT &&
270 rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
277 * eeh_report_resume - Tell device to resume normal operations
279 * @userdata: return value
281 * This routine must be called to notify the device driver that it
282 * could resume so that the device driver can do some initialization
283 * to make the recovered device work again.
285 static void *eeh_report_resume(void *data, void *userdata)
287 struct eeh_dev *edev = (struct eeh_dev *)data;
288 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
289 struct pci_driver *driver;
291 if (!dev) return NULL;
292 dev->error_state = pci_channel_io_normal;
294 driver = eeh_pcid_get(dev);
295 if (!driver) return NULL;
299 if (!driver->err_handler ||
300 !driver->err_handler->resume) {
305 driver->err_handler->resume(dev);
312 * eeh_report_failure - Tell device driver that device is dead.
314 * @userdata: return value
316 * This informs the device driver that the device is permanently
317 * dead, and that no further recovery attempts will be made on it.
319 static void *eeh_report_failure(void *data, void *userdata)
321 struct eeh_dev *edev = (struct eeh_dev *)data;
322 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
323 struct pci_driver *driver;
325 if (!dev) return NULL;
326 dev->error_state = pci_channel_io_perm_failure;
328 driver = eeh_pcid_get(dev);
329 if (!driver) return NULL;
331 eeh_disable_irq(dev);
333 if (!driver->err_handler ||
334 !driver->err_handler->error_detected) {
339 driver->err_handler->error_detected(dev, pci_channel_io_perm_failure);
345 static void *eeh_rmv_device(void *data, void *userdata)
347 struct pci_driver *driver;
348 struct eeh_dev *edev = (struct eeh_dev *)data;
349 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
350 int *removed = (int *)userdata;
353 * Actually, we should remove the PCI bridges as well.
354 * However, that's lots of complexity to do that,
355 * particularly some of devices under the bridge might
356 * support EEH. So we just care about PCI devices for
359 if (!dev || (dev->hdr_type & PCI_HEADER_TYPE_BRIDGE))
361 driver = eeh_pcid_get(dev);
362 if (driver && driver->err_handler)
365 /* Remove it from PCI subsystem */
366 pr_debug("EEH: Removing %s without EEH sensitive driver\n",
368 edev->bus = dev->bus;
369 edev->mode |= EEH_DEV_DISCONNECTED;
372 pci_lock_rescan_remove();
373 pci_stop_and_remove_bus_device(dev);
374 pci_unlock_rescan_remove();
379 static void *eeh_pe_detach_dev(void *data, void *userdata)
381 struct eeh_pe *pe = (struct eeh_pe *)data;
382 struct eeh_dev *edev, *tmp;
384 eeh_pe_for_each_dev(pe, edev, tmp) {
385 if (!(edev->mode & EEH_DEV_DISCONNECTED))
388 edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
389 eeh_rmv_from_parent_pe(edev);
396 * eeh_reset_device - Perform actual reset of a pci slot
398 * @bus: PCI bus corresponding to the isolcated slot
400 * This routine must be called to do reset on the indicated PE.
401 * During the reset, udev might be invoked because those affected
402 * PCI devices will be removed and then added.
404 static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus)
406 struct pci_bus *frozen_bus = eeh_pe_bus_get(pe);
407 struct timeval tstamp;
408 int cnt, rc, removed = 0;
410 /* pcibios will clear the counter; save the value */
411 cnt = pe->freeze_count;
415 * We don't remove the corresponding PE instances because
416 * we need the information afterwords. The attached EEH
417 * devices are expected to be attached soon when calling
418 * into pcibios_add_pci_devices().
420 eeh_pe_state_mark(pe, EEH_PE_KEEP);
422 pci_lock_rescan_remove();
423 pcibios_remove_pci_devices(bus);
424 pci_unlock_rescan_remove();
425 } else if (frozen_bus) {
426 eeh_pe_dev_traverse(pe, eeh_rmv_device, &removed);
429 /* Reset the pci controller. (Asserts RST#; resets config space).
430 * Reconfigure bridges and devices. Don't try to bring the system
431 * up if the reset failed for some reason.
433 rc = eeh_reset_pe(pe);
437 pci_lock_rescan_remove();
440 eeh_ops->configure_bridge(pe);
441 eeh_pe_restore_bars(pe);
443 /* Give the system 5 seconds to finish running the user-space
444 * hotplug shutdown scripts, e.g. ifdown for ethernet. Yes,
445 * this is a hack, but if we don't do this, and try to bring
446 * the device up before the scripts have taken it down,
447 * potentially weird things happen.
450 pr_info("EEH: Sleep 5s ahead of complete hotplug\n");
454 * The EEH device is still connected with its parent
455 * PE. We should disconnect it so the binding can be
456 * rebuilt when adding PCI devices.
458 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
459 pcibios_add_pci_devices(bus);
460 } else if (frozen_bus && removed) {
461 pr_info("EEH: Sleep 5s ahead of partial hotplug\n");
464 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
465 pcibios_add_pci_devices(frozen_bus);
467 eeh_pe_state_clear(pe, EEH_PE_KEEP);
470 pe->freeze_count = cnt;
472 pci_unlock_rescan_remove();
476 /* The longest amount of time to wait for a pci device
477 * to come back on line, in seconds.
479 #define MAX_WAIT_FOR_RECOVERY 150
481 static void eeh_handle_normal_event(struct eeh_pe *pe)
483 struct pci_bus *frozen_bus;
485 enum pci_ers_result result = PCI_ERS_RESULT_NONE;
487 frozen_bus = eeh_pe_bus_get(pe);
489 pr_err("%s: Cannot find PCI bus for PHB#%d-PE#%x\n",
490 __func__, pe->phb->global_number, pe->addr);
494 eeh_pe_update_time_stamp(pe);
496 if (pe->freeze_count > EEH_MAX_ALLOWED_FREEZES)
497 goto excess_failures;
498 pr_warning("EEH: This PCI device has failed %d times in the last hour\n",
501 /* Walk the various device drivers attached to this slot through
502 * a reset sequence, giving each an opportunity to do what it needs
503 * to accomplish the reset. Each child gets a report of the
504 * status ... if any child can't handle the reset, then the entire
505 * slot is dlpar removed and added.
507 pr_info("EEH: Notify device drivers to shutdown\n");
508 eeh_pe_dev_traverse(pe, eeh_report_error, &result);
510 /* Get the current PCI slot state. This can take a long time,
511 * sometimes over 3 seconds for certain systems.
513 rc = eeh_ops->wait_state(pe, MAX_WAIT_FOR_RECOVERY*1000);
514 if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
515 pr_warning("EEH: Permanent failure\n");
519 /* Since rtas may enable MMIO when posting the error log,
520 * don't post the error log until after all dev drivers
521 * have been informed.
523 pr_info("EEH: Collect temporary log\n");
524 eeh_slot_error_detail(pe, EEH_LOG_TEMP);
526 /* If all device drivers were EEH-unaware, then shut
527 * down all of the device drivers, and hope they
528 * go down willingly, without panicing the system.
530 if (result == PCI_ERS_RESULT_NONE) {
531 pr_info("EEH: Reset with hotplug activity\n");
532 rc = eeh_reset_device(pe, frozen_bus);
534 pr_warning("%s: Unable to reset, err=%d\n",
540 /* If all devices reported they can proceed, then re-enable MMIO */
541 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
542 pr_info("EEH: Enable I/O for affected devices\n");
543 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
548 result = PCI_ERS_RESULT_NEED_RESET;
550 pr_info("EEH: Notify device drivers to resume I/O\n");
551 result = PCI_ERS_RESULT_NONE;
552 eeh_pe_dev_traverse(pe, eeh_report_mmio_enabled, &result);
556 /* If all devices reported they can proceed, then re-enable DMA */
557 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
558 pr_info("EEH: Enabled DMA for affected devices\n");
559 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
564 result = PCI_ERS_RESULT_NEED_RESET;
566 result = PCI_ERS_RESULT_RECOVERED;
569 /* If any device has a hard failure, then shut off everything. */
570 if (result == PCI_ERS_RESULT_DISCONNECT) {
571 pr_warning("EEH: Device driver gave up\n");
575 /* If any device called out for a reset, then reset the slot */
576 if (result == PCI_ERS_RESULT_NEED_RESET) {
577 pr_info("EEH: Reset without hotplug activity\n");
578 rc = eeh_reset_device(pe, NULL);
580 pr_warning("%s: Cannot reset, err=%d\n",
585 pr_info("EEH: Notify device drivers "
586 "the completion of reset\n");
587 result = PCI_ERS_RESULT_NONE;
588 eeh_pe_dev_traverse(pe, eeh_report_reset, &result);
591 /* All devices should claim they have recovered by now. */
592 if ((result != PCI_ERS_RESULT_RECOVERED) &&
593 (result != PCI_ERS_RESULT_NONE)) {
594 pr_warning("EEH: Not recovered\n");
598 /* Tell all device drivers that they can resume operations */
599 pr_info("EEH: Notify device driver to resume\n");
600 eeh_pe_dev_traverse(pe, eeh_report_resume, NULL);
606 * About 90% of all real-life EEH failures in the field
607 * are due to poorly seated PCI cards. Only 10% or so are
608 * due to actual, failed cards.
610 pr_err("EEH: PHB#%d-PE#%x has failed %d times in the\n"
611 "last hour and has been permanently disabled.\n"
612 "Please try reseating or replacing it.\n",
613 pe->phb->global_number, pe->addr,
618 pr_err("EEH: Unable to recover from failure from PHB#%d-PE#%x.\n"
619 "Please try reseating or replacing it\n",
620 pe->phb->global_number, pe->addr);
623 eeh_slot_error_detail(pe, EEH_LOG_PERM);
625 /* Notify all devices that they're about to go down. */
626 eeh_pe_dev_traverse(pe, eeh_report_failure, NULL);
628 /* Shut down the device drivers for good. */
630 pci_lock_rescan_remove();
631 pcibios_remove_pci_devices(frozen_bus);
632 pci_unlock_rescan_remove();
636 static void eeh_handle_special_event(void)
638 struct eeh_pe *pe, *phb_pe;
640 struct pci_controller *hose, *tmp;
645 * The return value from next_error() has been classified as follows.
646 * It might be good to enumerate them. However, next_error() is only
647 * supported by PowerNV platform for now. So it would be fine to use
650 * 4 - Dead IOC 3 - Dead PHB
651 * 2 - Fenced PHB 1 - Frozen PE
655 rc = eeh_ops->next_error(&pe);
661 /* Mark all PHBs in dead state */
662 eeh_serialize_lock(&flags);
663 list_for_each_entry_safe(hose, tmp,
664 &hose_list, list_node) {
665 phb_pe = eeh_phb_pe_get(hose);
666 if (!phb_pe) continue;
668 eeh_pe_state_mark(phb_pe,
669 EEH_PE_ISOLATED | EEH_PE_PHB_DEAD);
671 eeh_serialize_unlock(flags);
673 /* Purge all events */
674 eeh_remove_event(NULL);
679 /* Mark the PE in fenced state */
680 eeh_serialize_lock(&flags);
682 eeh_pe_state_mark(pe,
683 EEH_PE_ISOLATED | EEH_PE_PHB_DEAD);
685 eeh_pe_state_mark(pe,
686 EEH_PE_ISOLATED | EEH_PE_RECOVERING);
687 eeh_serialize_unlock(flags);
689 /* Purge all events of the PHB */
690 eeh_remove_event(pe);
693 pr_err("%s: Invalid value %d from next_error()\n",
699 * For fenced PHB and frozen PE, it's handled as normal
700 * event. We have to remove the affected PHBs for dead
703 if (rc == 2 || rc == 1)
704 eeh_handle_normal_event(pe);
706 pci_lock_rescan_remove();
707 list_for_each_entry_safe(hose, tmp,
708 &hose_list, list_node) {
709 phb_pe = eeh_phb_pe_get(hose);
710 if (!phb_pe || !(phb_pe->state & EEH_PE_PHB_DEAD))
713 bus = eeh_pe_bus_get(phb_pe);
714 /* Notify all devices that they're about to go down. */
715 eeh_pe_dev_traverse(pe, eeh_report_failure, NULL);
716 pcibios_remove_pci_devices(bus);
718 pci_unlock_rescan_remove();
723 * eeh_handle_event - Reset a PCI device after hard lockup.
726 * While PHB detects address or data parity errors on particular PCI
727 * slot, the associated PE will be frozen. Besides, DMA's occurring
728 * to wild addresses (which usually happen due to bugs in device
729 * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
730 * #PERR or other misc PCI-related errors also can trigger EEH errors.
732 * Recovery process consists of unplugging the device driver (which
733 * generated hotplug events to userspace), then issuing a PCI #RST to
734 * the device, then reconfiguring the PCI config space for all bridges
735 * & devices under this slot, and then finally restarting the device
736 * drivers (which cause a second set of hotplug events to go out to
739 void eeh_handle_event(struct eeh_pe *pe)
742 eeh_handle_normal_event(pe);
744 eeh_handle_special_event();