Loading...
1/*
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
5 *
6 * All rights reserved.
7 *
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.
12 *
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
17 * details.
18 *
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.
22 *
23 * Send comments and feedback to Linas Vepstas <linas@austin.ibm.com>
24 */
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>
30#include <asm/eeh.h>
31#include <asm/eeh_event.h>
32#include <asm/ppc-pci.h>
33#include <asm/pci-bridge.h>
34#include <asm/prom.h>
35#include <asm/rtas.h>
36
37struct eeh_rmv_data {
38 struct list_head edev_list;
39 int removed;
40};
41
42/**
43 * eeh_pcid_name - Retrieve name of PCI device driver
44 * @pdev: PCI device
45 *
46 * This routine is used to retrieve the name of PCI device driver
47 * if that's valid.
48 */
49static inline const char *eeh_pcid_name(struct pci_dev *pdev)
50{
51 if (pdev && pdev->dev.driver)
52 return pdev->dev.driver->name;
53 return "";
54}
55
56/**
57 * eeh_pcid_get - Get the PCI device driver
58 * @pdev: PCI device
59 *
60 * The function is used to retrieve the PCI device driver for
61 * the indicated PCI device. Besides, we will increase the reference
62 * of the PCI device driver to prevent that being unloaded on
63 * the fly. Otherwise, kernel crash would be seen.
64 */
65static inline struct pci_driver *eeh_pcid_get(struct pci_dev *pdev)
66{
67 if (!pdev || !pdev->driver)
68 return NULL;
69
70 if (!try_module_get(pdev->driver->driver.owner))
71 return NULL;
72
73 return pdev->driver;
74}
75
76/**
77 * eeh_pcid_put - Dereference on the PCI device driver
78 * @pdev: PCI device
79 *
80 * The function is called to do dereference on the PCI device
81 * driver of the indicated PCI device.
82 */
83static inline void eeh_pcid_put(struct pci_dev *pdev)
84{
85 if (!pdev || !pdev->driver)
86 return;
87
88 module_put(pdev->driver->driver.owner);
89}
90
91/**
92 * eeh_disable_irq - Disable interrupt for the recovering device
93 * @dev: PCI device
94 *
95 * This routine must be called when reporting temporary or permanent
96 * error to the particular PCI device to disable interrupt of that
97 * device. If the device has enabled MSI or MSI-X interrupt, we needn't
98 * do real work because EEH should freeze DMA transfers for those PCI
99 * devices encountering EEH errors, which includes MSI or MSI-X.
100 */
101static void eeh_disable_irq(struct pci_dev *dev)
102{
103 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
104
105 /* Don't disable MSI and MSI-X interrupts. They are
106 * effectively disabled by the DMA Stopped state
107 * when an EEH error occurs.
108 */
109 if (dev->msi_enabled || dev->msix_enabled)
110 return;
111
112 if (!irq_has_action(dev->irq))
113 return;
114
115 edev->mode |= EEH_DEV_IRQ_DISABLED;
116 disable_irq_nosync(dev->irq);
117}
118
119/**
120 * eeh_enable_irq - Enable interrupt for the recovering device
121 * @dev: PCI device
122 *
123 * This routine must be called to enable interrupt while failed
124 * device could be resumed.
125 */
126static void eeh_enable_irq(struct pci_dev *dev)
127{
128 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
129
130 if ((edev->mode) & EEH_DEV_IRQ_DISABLED) {
131 edev->mode &= ~EEH_DEV_IRQ_DISABLED;
132 /*
133 * FIXME !!!!!
134 *
135 * This is just ass backwards. This maze has
136 * unbalanced irq_enable/disable calls. So instead of
137 * finding the root cause it works around the warning
138 * in the irq_enable code by conditionally calling
139 * into it.
140 *
141 * That's just wrong.The warning in the core code is
142 * there to tell people to fix their asymmetries in
143 * their own code, not by abusing the core information
144 * to avoid it.
145 *
146 * I so wish that the assymetry would be the other way
147 * round and a few more irq_disable calls render that
148 * shit unusable forever.
149 *
150 * tglx
151 */
152 if (irqd_irq_disabled(irq_get_irq_data(dev->irq)))
153 enable_irq(dev->irq);
154 }
155}
156
157static bool eeh_dev_removed(struct eeh_dev *edev)
158{
159 /* EEH device removed ? */
160 if (!edev || (edev->mode & EEH_DEV_REMOVED))
161 return true;
162
163 return false;
164}
165
166static void *eeh_dev_save_state(void *data, void *userdata)
167{
168 struct eeh_dev *edev = data;
169 struct pci_dev *pdev;
170
171 if (!edev)
172 return NULL;
173
174 /*
175 * We cannot access the config space on some adapters.
176 * Otherwise, it will cause fenced PHB. We don't save
177 * the content in their config space and will restore
178 * from the initial config space saved when the EEH
179 * device is created.
180 */
181 if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED))
182 return NULL;
183
184 pdev = eeh_dev_to_pci_dev(edev);
185 if (!pdev)
186 return NULL;
187
188 pci_save_state(pdev);
189 return NULL;
190}
191
192/**
193 * eeh_report_error - Report pci error to each device driver
194 * @data: eeh device
195 * @userdata: return value
196 *
197 * Report an EEH error to each device driver, collect up and
198 * merge the device driver responses. Cumulative response
199 * passed back in "userdata".
200 */
201static void *eeh_report_error(void *data, void *userdata)
202{
203 struct eeh_dev *edev = (struct eeh_dev *)data;
204 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
205 enum pci_ers_result rc, *res = userdata;
206 struct pci_driver *driver;
207
208 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
209 return NULL;
210
211 device_lock(&dev->dev);
212 dev->error_state = pci_channel_io_frozen;
213
214 driver = eeh_pcid_get(dev);
215 if (!driver) goto out_no_dev;
216
217 eeh_disable_irq(dev);
218
219 if (!driver->err_handler ||
220 !driver->err_handler->error_detected)
221 goto out;
222
223 rc = driver->err_handler->error_detected(dev, pci_channel_io_frozen);
224
225 /* A driver that needs a reset trumps all others */
226 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
227 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
228
229 edev->in_error = true;
230 pci_uevent_ers(dev, PCI_ERS_RESULT_NONE);
231
232out:
233 eeh_pcid_put(dev);
234out_no_dev:
235 device_unlock(&dev->dev);
236 return NULL;
237}
238
239/**
240 * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
241 * @data: eeh device
242 * @userdata: return value
243 *
244 * Tells each device driver that IO ports, MMIO and config space I/O
245 * are now enabled. Collects up and merges the device driver responses.
246 * Cumulative response passed back in "userdata".
247 */
248static void *eeh_report_mmio_enabled(void *data, void *userdata)
249{
250 struct eeh_dev *edev = (struct eeh_dev *)data;
251 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
252 enum pci_ers_result rc, *res = userdata;
253 struct pci_driver *driver;
254
255 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
256 return NULL;
257
258 device_lock(&dev->dev);
259 driver = eeh_pcid_get(dev);
260 if (!driver) goto out_no_dev;
261
262 if (!driver->err_handler ||
263 !driver->err_handler->mmio_enabled ||
264 (edev->mode & EEH_DEV_NO_HANDLER))
265 goto out;
266
267 rc = driver->err_handler->mmio_enabled(dev);
268
269 /* A driver that needs a reset trumps all others */
270 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
271 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
272
273out:
274 eeh_pcid_put(dev);
275out_no_dev:
276 device_unlock(&dev->dev);
277 return NULL;
278}
279
280/**
281 * eeh_report_reset - Tell device that slot has been reset
282 * @data: eeh device
283 * @userdata: return value
284 *
285 * This routine must be called while EEH tries to reset particular
286 * PCI device so that the associated PCI device driver could take
287 * some actions, usually to save data the driver needs so that the
288 * driver can work again while the device is recovered.
289 */
290static void *eeh_report_reset(void *data, void *userdata)
291{
292 struct eeh_dev *edev = (struct eeh_dev *)data;
293 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
294 enum pci_ers_result rc, *res = userdata;
295 struct pci_driver *driver;
296
297 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
298 return NULL;
299
300 device_lock(&dev->dev);
301 dev->error_state = pci_channel_io_normal;
302
303 driver = eeh_pcid_get(dev);
304 if (!driver) goto out_no_dev;
305
306 eeh_enable_irq(dev);
307
308 if (!driver->err_handler ||
309 !driver->err_handler->slot_reset ||
310 (edev->mode & EEH_DEV_NO_HANDLER) ||
311 (!edev->in_error))
312 goto out;
313
314 rc = driver->err_handler->slot_reset(dev);
315 if ((*res == PCI_ERS_RESULT_NONE) ||
316 (*res == PCI_ERS_RESULT_RECOVERED)) *res = rc;
317 if (*res == PCI_ERS_RESULT_DISCONNECT &&
318 rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
319
320out:
321 eeh_pcid_put(dev);
322out_no_dev:
323 device_unlock(&dev->dev);
324 return NULL;
325}
326
327static void *eeh_dev_restore_state(void *data, void *userdata)
328{
329 struct eeh_dev *edev = data;
330 struct pci_dev *pdev;
331
332 if (!edev)
333 return NULL;
334
335 /*
336 * The content in the config space isn't saved because
337 * the blocked config space on some adapters. We have
338 * to restore the initial saved config space when the
339 * EEH device is created.
340 */
341 if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED)) {
342 if (list_is_last(&edev->list, &edev->pe->edevs))
343 eeh_pe_restore_bars(edev->pe);
344
345 return NULL;
346 }
347
348 pdev = eeh_dev_to_pci_dev(edev);
349 if (!pdev)
350 return NULL;
351
352 pci_restore_state(pdev);
353 return NULL;
354}
355
356/**
357 * eeh_report_resume - Tell device to resume normal operations
358 * @data: eeh device
359 * @userdata: return value
360 *
361 * This routine must be called to notify the device driver that it
362 * could resume so that the device driver can do some initialization
363 * to make the recovered device work again.
364 */
365static void *eeh_report_resume(void *data, void *userdata)
366{
367 struct eeh_dev *edev = (struct eeh_dev *)data;
368 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
369 bool was_in_error;
370 struct pci_driver *driver;
371
372 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
373 return NULL;
374
375 device_lock(&dev->dev);
376 dev->error_state = pci_channel_io_normal;
377
378 driver = eeh_pcid_get(dev);
379 if (!driver) goto out_no_dev;
380
381 was_in_error = edev->in_error;
382 edev->in_error = false;
383 eeh_enable_irq(dev);
384
385 if (!driver->err_handler ||
386 !driver->err_handler->resume ||
387 (edev->mode & EEH_DEV_NO_HANDLER) || !was_in_error) {
388 edev->mode &= ~EEH_DEV_NO_HANDLER;
389 goto out;
390 }
391
392 driver->err_handler->resume(dev);
393
394 pci_uevent_ers(dev, PCI_ERS_RESULT_RECOVERED);
395out:
396 eeh_pcid_put(dev);
397#ifdef CONFIG_PCI_IOV
398 if (eeh_ops->notify_resume && eeh_dev_to_pdn(edev))
399 eeh_ops->notify_resume(eeh_dev_to_pdn(edev));
400#endif
401out_no_dev:
402 device_unlock(&dev->dev);
403 return NULL;
404}
405
406/**
407 * eeh_report_failure - Tell device driver that device is dead.
408 * @data: eeh device
409 * @userdata: return value
410 *
411 * This informs the device driver that the device is permanently
412 * dead, and that no further recovery attempts will be made on it.
413 */
414static void *eeh_report_failure(void *data, void *userdata)
415{
416 struct eeh_dev *edev = (struct eeh_dev *)data;
417 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
418 struct pci_driver *driver;
419
420 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
421 return NULL;
422
423 device_lock(&dev->dev);
424 dev->error_state = pci_channel_io_perm_failure;
425
426 driver = eeh_pcid_get(dev);
427 if (!driver) goto out_no_dev;
428
429 eeh_disable_irq(dev);
430
431 if (!driver->err_handler ||
432 !driver->err_handler->error_detected)
433 goto out;
434
435 driver->err_handler->error_detected(dev, pci_channel_io_perm_failure);
436
437 pci_uevent_ers(dev, PCI_ERS_RESULT_DISCONNECT);
438out:
439 eeh_pcid_put(dev);
440out_no_dev:
441 device_unlock(&dev->dev);
442 return NULL;
443}
444
445static void *eeh_add_virt_device(void *data, void *userdata)
446{
447 struct pci_driver *driver;
448 struct eeh_dev *edev = (struct eeh_dev *)data;
449 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
450 struct pci_dn *pdn = eeh_dev_to_pdn(edev);
451
452 if (!(edev->physfn)) {
453 pr_warn("%s: EEH dev %04x:%02x:%02x.%01x not for VF\n",
454 __func__, pdn->phb->global_number, pdn->busno,
455 PCI_SLOT(pdn->devfn), PCI_FUNC(pdn->devfn));
456 return NULL;
457 }
458
459 driver = eeh_pcid_get(dev);
460 if (driver) {
461 eeh_pcid_put(dev);
462 if (driver->err_handler)
463 return NULL;
464 }
465
466#ifdef CONFIG_PCI_IOV
467 pci_iov_add_virtfn(edev->physfn, pdn->vf_index);
468#endif
469 return NULL;
470}
471
472static void *eeh_rmv_device(void *data, void *userdata)
473{
474 struct pci_driver *driver;
475 struct eeh_dev *edev = (struct eeh_dev *)data;
476 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
477 struct eeh_rmv_data *rmv_data = (struct eeh_rmv_data *)userdata;
478 int *removed = rmv_data ? &rmv_data->removed : NULL;
479
480 /*
481 * Actually, we should remove the PCI bridges as well.
482 * However, that's lots of complexity to do that,
483 * particularly some of devices under the bridge might
484 * support EEH. So we just care about PCI devices for
485 * simplicity here.
486 */
487 if (!dev || (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE))
488 return NULL;
489
490 /*
491 * We rely on count-based pcibios_release_device() to
492 * detach permanently offlined PEs. Unfortunately, that's
493 * not reliable enough. We might have the permanently
494 * offlined PEs attached, but we needn't take care of
495 * them and their child devices.
496 */
497 if (eeh_dev_removed(edev))
498 return NULL;
499
500 driver = eeh_pcid_get(dev);
501 if (driver) {
502 eeh_pcid_put(dev);
503 if (removed &&
504 eeh_pe_passed(edev->pe))
505 return NULL;
506 if (removed &&
507 driver->err_handler &&
508 driver->err_handler->error_detected &&
509 driver->err_handler->slot_reset)
510 return NULL;
511 }
512
513 /* Remove it from PCI subsystem */
514 pr_debug("EEH: Removing %s without EEH sensitive driver\n",
515 pci_name(dev));
516 edev->bus = dev->bus;
517 edev->mode |= EEH_DEV_DISCONNECTED;
518 if (removed)
519 (*removed)++;
520
521 if (edev->physfn) {
522#ifdef CONFIG_PCI_IOV
523 struct pci_dn *pdn = eeh_dev_to_pdn(edev);
524
525 pci_iov_remove_virtfn(edev->physfn, pdn->vf_index);
526 edev->pdev = NULL;
527
528 /*
529 * We have to set the VF PE number to invalid one, which is
530 * required to plug the VF successfully.
531 */
532 pdn->pe_number = IODA_INVALID_PE;
533#endif
534 if (rmv_data)
535 list_add(&edev->rmv_list, &rmv_data->edev_list);
536 } else {
537 pci_lock_rescan_remove();
538 pci_stop_and_remove_bus_device(dev);
539 pci_unlock_rescan_remove();
540 }
541
542 return NULL;
543}
544
545static void *eeh_pe_detach_dev(void *data, void *userdata)
546{
547 struct eeh_pe *pe = (struct eeh_pe *)data;
548 struct eeh_dev *edev, *tmp;
549
550 eeh_pe_for_each_dev(pe, edev, tmp) {
551 if (!(edev->mode & EEH_DEV_DISCONNECTED))
552 continue;
553
554 edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
555 eeh_rmv_from_parent_pe(edev);
556 }
557
558 return NULL;
559}
560
561/*
562 * Explicitly clear PE's frozen state for PowerNV where
563 * we have frozen PE until BAR restore is completed. It's
564 * harmless to clear it for pSeries. To be consistent with
565 * PE reset (for 3 times), we try to clear the frozen state
566 * for 3 times as well.
567 */
568static void *__eeh_clear_pe_frozen_state(void *data, void *flag)
569{
570 struct eeh_pe *pe = (struct eeh_pe *)data;
571 bool clear_sw_state = *(bool *)flag;
572 int i, rc = 1;
573
574 for (i = 0; rc && i < 3; i++)
575 rc = eeh_unfreeze_pe(pe, clear_sw_state);
576
577 /* Stop immediately on any errors */
578 if (rc) {
579 pr_warn("%s: Failure %d unfreezing PHB#%x-PE#%x\n",
580 __func__, rc, pe->phb->global_number, pe->addr);
581 return (void *)pe;
582 }
583
584 return NULL;
585}
586
587static int eeh_clear_pe_frozen_state(struct eeh_pe *pe,
588 bool clear_sw_state)
589{
590 void *rc;
591
592 rc = eeh_pe_traverse(pe, __eeh_clear_pe_frozen_state, &clear_sw_state);
593 if (!rc)
594 eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
595
596 return rc ? -EIO : 0;
597}
598
599int eeh_pe_reset_and_recover(struct eeh_pe *pe)
600{
601 int ret;
602
603 /* Bail if the PE is being recovered */
604 if (pe->state & EEH_PE_RECOVERING)
605 return 0;
606
607 /* Put the PE into recovery mode */
608 eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
609
610 /* Save states */
611 eeh_pe_dev_traverse(pe, eeh_dev_save_state, NULL);
612
613 /* Issue reset */
614 ret = eeh_pe_reset_full(pe);
615 if (ret) {
616 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
617 return ret;
618 }
619
620 /* Unfreeze the PE */
621 ret = eeh_clear_pe_frozen_state(pe, true);
622 if (ret) {
623 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
624 return ret;
625 }
626
627 /* Restore device state */
628 eeh_pe_dev_traverse(pe, eeh_dev_restore_state, NULL);
629
630 /* Clear recovery mode */
631 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
632
633 return 0;
634}
635
636/**
637 * eeh_reset_device - Perform actual reset of a pci slot
638 * @driver_eeh_aware: Does the device's driver provide EEH support?
639 * @pe: EEH PE
640 * @bus: PCI bus corresponding to the isolcated slot
641 * @rmv_data: Optional, list to record removed devices
642 *
643 * This routine must be called to do reset on the indicated PE.
644 * During the reset, udev might be invoked because those affected
645 * PCI devices will be removed and then added.
646 */
647static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus,
648 struct eeh_rmv_data *rmv_data,
649 bool driver_eeh_aware)
650{
651 time64_t tstamp;
652 int cnt, rc;
653 struct eeh_dev *edev;
654
655 /* pcibios will clear the counter; save the value */
656 cnt = pe->freeze_count;
657 tstamp = pe->tstamp;
658
659 /*
660 * We don't remove the corresponding PE instances because
661 * we need the information afterwords. The attached EEH
662 * devices are expected to be attached soon when calling
663 * into pci_hp_add_devices().
664 */
665 eeh_pe_state_mark(pe, EEH_PE_KEEP);
666 if (driver_eeh_aware || (pe->type & EEH_PE_VF)) {
667 eeh_pe_dev_traverse(pe, eeh_rmv_device, rmv_data);
668 } else {
669 pci_lock_rescan_remove();
670 pci_hp_remove_devices(bus);
671 pci_unlock_rescan_remove();
672 }
673
674 /*
675 * Reset the pci controller. (Asserts RST#; resets config space).
676 * Reconfigure bridges and devices. Don't try to bring the system
677 * up if the reset failed for some reason.
678 *
679 * During the reset, it's very dangerous to have uncontrolled PCI
680 * config accesses. So we prefer to block them. However, controlled
681 * PCI config accesses initiated from EEH itself are allowed.
682 */
683 rc = eeh_pe_reset_full(pe);
684 if (rc)
685 return rc;
686
687 pci_lock_rescan_remove();
688
689 /* Restore PE */
690 eeh_ops->configure_bridge(pe);
691 eeh_pe_restore_bars(pe);
692
693 /* Clear frozen state */
694 rc = eeh_clear_pe_frozen_state(pe, false);
695 if (rc) {
696 pci_unlock_rescan_remove();
697 return rc;
698 }
699
700 /* Give the system 5 seconds to finish running the user-space
701 * hotplug shutdown scripts, e.g. ifdown for ethernet. Yes,
702 * this is a hack, but if we don't do this, and try to bring
703 * the device up before the scripts have taken it down,
704 * potentially weird things happen.
705 */
706 if (!driver_eeh_aware || rmv_data->removed) {
707 pr_info("EEH: Sleep 5s ahead of %s hotplug\n",
708 (driver_eeh_aware ? "partial" : "complete"));
709 ssleep(5);
710
711 /*
712 * The EEH device is still connected with its parent
713 * PE. We should disconnect it so the binding can be
714 * rebuilt when adding PCI devices.
715 */
716 edev = list_first_entry(&pe->edevs, struct eeh_dev, list);
717 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
718 if (pe->type & EEH_PE_VF) {
719 eeh_add_virt_device(edev, NULL);
720 } else {
721 if (!driver_eeh_aware)
722 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
723 pci_hp_add_devices(bus);
724 }
725 }
726 eeh_pe_state_clear(pe, EEH_PE_KEEP);
727
728 pe->tstamp = tstamp;
729 pe->freeze_count = cnt;
730
731 pci_unlock_rescan_remove();
732 return 0;
733}
734
735/* The longest amount of time to wait for a pci device
736 * to come back on line, in seconds.
737 */
738#define MAX_WAIT_FOR_RECOVERY 300
739
740/**
741 * eeh_handle_normal_event - Handle EEH events on a specific PE
742 * @pe: EEH PE - which should not be used after we return, as it may
743 * have been invalidated.
744 *
745 * Attempts to recover the given PE. If recovery fails or the PE has failed
746 * too many times, remove the PE.
747 *
748 * While PHB detects address or data parity errors on particular PCI
749 * slot, the associated PE will be frozen. Besides, DMA's occurring
750 * to wild addresses (which usually happen due to bugs in device
751 * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
752 * #PERR or other misc PCI-related errors also can trigger EEH errors.
753 *
754 * Recovery process consists of unplugging the device driver (which
755 * generated hotplug events to userspace), then issuing a PCI #RST to
756 * the device, then reconfiguring the PCI config space for all bridges
757 * & devices under this slot, and then finally restarting the device
758 * drivers (which cause a second set of hotplug events to go out to
759 * userspace).
760 */
761void eeh_handle_normal_event(struct eeh_pe *pe)
762{
763 struct pci_bus *bus;
764 struct eeh_dev *edev, *tmp;
765 int rc = 0;
766 enum pci_ers_result result = PCI_ERS_RESULT_NONE;
767 struct eeh_rmv_data rmv_data = {LIST_HEAD_INIT(rmv_data.edev_list), 0};
768
769 bus = eeh_pe_bus_get(pe);
770 if (!bus) {
771 pr_err("%s: Cannot find PCI bus for PHB#%x-PE#%x\n",
772 __func__, pe->phb->global_number, pe->addr);
773 return;
774 }
775
776 eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
777
778 eeh_pe_update_time_stamp(pe);
779 pe->freeze_count++;
780 if (pe->freeze_count > eeh_max_freezes) {
781 pr_err("EEH: PHB#%x-PE#%x has failed %d times in the\n"
782 "last hour and has been permanently disabled.\n",
783 pe->phb->global_number, pe->addr,
784 pe->freeze_count);
785 goto hard_fail;
786 }
787 pr_warn("EEH: This PCI device has failed %d times in the last hour\n",
788 pe->freeze_count);
789
790 /* Walk the various device drivers attached to this slot through
791 * a reset sequence, giving each an opportunity to do what it needs
792 * to accomplish the reset. Each child gets a report of the
793 * status ... if any child can't handle the reset, then the entire
794 * slot is dlpar removed and added.
795 *
796 * When the PHB is fenced, we have to issue a reset to recover from
797 * the error. Override the result if necessary to have partially
798 * hotplug for this case.
799 */
800 pr_info("EEH: Notify device drivers to shutdown\n");
801 eeh_pe_dev_traverse(pe, eeh_report_error, &result);
802 if ((pe->type & EEH_PE_PHB) &&
803 result != PCI_ERS_RESULT_NONE &&
804 result != PCI_ERS_RESULT_NEED_RESET)
805 result = PCI_ERS_RESULT_NEED_RESET;
806
807 /* Get the current PCI slot state. This can take a long time,
808 * sometimes over 300 seconds for certain systems.
809 */
810 rc = eeh_ops->wait_state(pe, MAX_WAIT_FOR_RECOVERY*1000);
811 if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
812 pr_warn("EEH: Permanent failure\n");
813 goto hard_fail;
814 }
815
816 /* Since rtas may enable MMIO when posting the error log,
817 * don't post the error log until after all dev drivers
818 * have been informed.
819 */
820 pr_info("EEH: Collect temporary log\n");
821 eeh_slot_error_detail(pe, EEH_LOG_TEMP);
822
823 /* If all device drivers were EEH-unaware, then shut
824 * down all of the device drivers, and hope they
825 * go down willingly, without panicing the system.
826 */
827 if (result == PCI_ERS_RESULT_NONE) {
828 pr_info("EEH: Reset with hotplug activity\n");
829 rc = eeh_reset_device(pe, bus, NULL, false);
830 if (rc) {
831 pr_warn("%s: Unable to reset, err=%d\n",
832 __func__, rc);
833 goto hard_fail;
834 }
835 }
836
837 /* If all devices reported they can proceed, then re-enable MMIO */
838 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
839 pr_info("EEH: Enable I/O for affected devices\n");
840 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
841
842 if (rc < 0)
843 goto hard_fail;
844 if (rc) {
845 result = PCI_ERS_RESULT_NEED_RESET;
846 } else {
847 pr_info("EEH: Notify device drivers to resume I/O\n");
848 eeh_pe_dev_traverse(pe, eeh_report_mmio_enabled, &result);
849 }
850 }
851
852 /* If all devices reported they can proceed, then re-enable DMA */
853 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
854 pr_info("EEH: Enabled DMA for affected devices\n");
855 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
856
857 if (rc < 0)
858 goto hard_fail;
859 if (rc) {
860 result = PCI_ERS_RESULT_NEED_RESET;
861 } else {
862 /*
863 * We didn't do PE reset for the case. The PE
864 * is still in frozen state. Clear it before
865 * resuming the PE.
866 */
867 eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
868 result = PCI_ERS_RESULT_RECOVERED;
869 }
870 }
871
872 /* If any device has a hard failure, then shut off everything. */
873 if (result == PCI_ERS_RESULT_DISCONNECT) {
874 pr_warn("EEH: Device driver gave up\n");
875 goto hard_fail;
876 }
877
878 /* If any device called out for a reset, then reset the slot */
879 if (result == PCI_ERS_RESULT_NEED_RESET) {
880 pr_info("EEH: Reset without hotplug activity\n");
881 rc = eeh_reset_device(pe, bus, &rmv_data, true);
882 if (rc) {
883 pr_warn("%s: Cannot reset, err=%d\n",
884 __func__, rc);
885 goto hard_fail;
886 }
887
888 pr_info("EEH: Notify device drivers "
889 "the completion of reset\n");
890 result = PCI_ERS_RESULT_NONE;
891 eeh_pe_dev_traverse(pe, eeh_report_reset, &result);
892 }
893
894 /* All devices should claim they have recovered by now. */
895 if ((result != PCI_ERS_RESULT_RECOVERED) &&
896 (result != PCI_ERS_RESULT_NONE)) {
897 pr_warn("EEH: Not recovered\n");
898 goto hard_fail;
899 }
900
901 /*
902 * For those hot removed VFs, we should add back them after PF get
903 * recovered properly.
904 */
905 list_for_each_entry_safe(edev, tmp, &rmv_data.edev_list, rmv_list) {
906 eeh_add_virt_device(edev, NULL);
907 list_del(&edev->rmv_list);
908 }
909
910 /* Tell all device drivers that they can resume operations */
911 pr_info("EEH: Notify device driver to resume\n");
912 eeh_pe_dev_traverse(pe, eeh_report_resume, NULL);
913
914 goto final;
915
916hard_fail:
917 /*
918 * About 90% of all real-life EEH failures in the field
919 * are due to poorly seated PCI cards. Only 10% or so are
920 * due to actual, failed cards.
921 */
922 pr_err("EEH: Unable to recover from failure from PHB#%x-PE#%x.\n"
923 "Please try reseating or replacing it\n",
924 pe->phb->global_number, pe->addr);
925
926 eeh_slot_error_detail(pe, EEH_LOG_PERM);
927
928 /* Notify all devices that they're about to go down. */
929 eeh_pe_dev_traverse(pe, eeh_report_failure, NULL);
930
931 /* Mark the PE to be removed permanently */
932 eeh_pe_state_mark(pe, EEH_PE_REMOVED);
933
934 /*
935 * Shut down the device drivers for good. We mark
936 * all removed devices correctly to avoid access
937 * the their PCI config any more.
938 */
939 if (pe->type & EEH_PE_VF) {
940 eeh_pe_dev_traverse(pe, eeh_rmv_device, NULL);
941 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
942 } else {
943 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
944 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
945
946 pci_lock_rescan_remove();
947 pci_hp_remove_devices(bus);
948 pci_unlock_rescan_remove();
949 /* The passed PE should no longer be used */
950 return;
951 }
952final:
953 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
954}
955
956/**
957 * eeh_handle_special_event - Handle EEH events without a specific failing PE
958 *
959 * Called when an EEH event is detected but can't be narrowed down to a
960 * specific PE. Iterates through possible failures and handles them as
961 * necessary.
962 */
963void eeh_handle_special_event(void)
964{
965 struct eeh_pe *pe, *phb_pe;
966 struct pci_bus *bus;
967 struct pci_controller *hose;
968 unsigned long flags;
969 int rc;
970
971
972 do {
973 rc = eeh_ops->next_error(&pe);
974
975 switch (rc) {
976 case EEH_NEXT_ERR_DEAD_IOC:
977 /* Mark all PHBs in dead state */
978 eeh_serialize_lock(&flags);
979
980 /* Purge all events */
981 eeh_remove_event(NULL, true);
982
983 list_for_each_entry(hose, &hose_list, list_node) {
984 phb_pe = eeh_phb_pe_get(hose);
985 if (!phb_pe) continue;
986
987 eeh_pe_state_mark(phb_pe, EEH_PE_ISOLATED);
988 }
989
990 eeh_serialize_unlock(flags);
991
992 break;
993 case EEH_NEXT_ERR_FROZEN_PE:
994 case EEH_NEXT_ERR_FENCED_PHB:
995 case EEH_NEXT_ERR_DEAD_PHB:
996 /* Mark the PE in fenced state */
997 eeh_serialize_lock(&flags);
998
999 /* Purge all events of the PHB */
1000 eeh_remove_event(pe, true);
1001
1002 if (rc == EEH_NEXT_ERR_DEAD_PHB)
1003 eeh_pe_state_mark(pe, EEH_PE_ISOLATED);
1004 else
1005 eeh_pe_state_mark(pe,
1006 EEH_PE_ISOLATED | EEH_PE_RECOVERING);
1007
1008 eeh_serialize_unlock(flags);
1009
1010 break;
1011 case EEH_NEXT_ERR_NONE:
1012 return;
1013 default:
1014 pr_warn("%s: Invalid value %d from next_error()\n",
1015 __func__, rc);
1016 return;
1017 }
1018
1019 /*
1020 * For fenced PHB and frozen PE, it's handled as normal
1021 * event. We have to remove the affected PHBs for dead
1022 * PHB and IOC
1023 */
1024 if (rc == EEH_NEXT_ERR_FROZEN_PE ||
1025 rc == EEH_NEXT_ERR_FENCED_PHB) {
1026 eeh_handle_normal_event(pe);
1027 } else {
1028 pci_lock_rescan_remove();
1029 list_for_each_entry(hose, &hose_list, list_node) {
1030 phb_pe = eeh_phb_pe_get(hose);
1031 if (!phb_pe ||
1032 !(phb_pe->state & EEH_PE_ISOLATED) ||
1033 (phb_pe->state & EEH_PE_RECOVERING))
1034 continue;
1035
1036 /* Notify all devices to be down */
1037 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
1038 eeh_pe_dev_traverse(pe,
1039 eeh_report_failure, NULL);
1040 bus = eeh_pe_bus_get(phb_pe);
1041 if (!bus) {
1042 pr_err("%s: Cannot find PCI bus for "
1043 "PHB#%x-PE#%x\n",
1044 __func__,
1045 pe->phb->global_number,
1046 pe->addr);
1047 break;
1048 }
1049 pci_hp_remove_devices(bus);
1050 }
1051 pci_unlock_rescan_remove();
1052 }
1053
1054 /*
1055 * If we have detected dead IOC, we needn't proceed
1056 * any more since all PHBs would have been removed
1057 */
1058 if (rc == EEH_NEXT_ERR_DEAD_IOC)
1059 break;
1060 } while (rc != EEH_NEXT_ERR_NONE);
1061}
1/*
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
5 *
6 * All rights reserved.
7 *
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.
12 *
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
17 * details.
18 *
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.
22 *
23 * Send comments and feedback to Linas Vepstas <linas@austin.ibm.com>
24 */
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>
30#include <asm/eeh.h>
31#include <asm/eeh_event.h>
32#include <asm/ppc-pci.h>
33#include <asm/pci-bridge.h>
34#include <asm/prom.h>
35#include <asm/rtas.h>
36
37struct eeh_rmv_data {
38 struct list_head edev_list;
39 int removed;
40};
41
42/**
43 * eeh_pcid_name - Retrieve name of PCI device driver
44 * @pdev: PCI device
45 *
46 * This routine is used to retrieve the name of PCI device driver
47 * if that's valid.
48 */
49static inline const char *eeh_pcid_name(struct pci_dev *pdev)
50{
51 if (pdev && pdev->dev.driver)
52 return pdev->dev.driver->name;
53 return "";
54}
55
56/**
57 * eeh_pcid_get - Get the PCI device driver
58 * @pdev: PCI device
59 *
60 * The function is used to retrieve the PCI device driver for
61 * the indicated PCI device. Besides, we will increase the reference
62 * of the PCI device driver to prevent that being unloaded on
63 * the fly. Otherwise, kernel crash would be seen.
64 */
65static inline struct pci_driver *eeh_pcid_get(struct pci_dev *pdev)
66{
67 if (!pdev || !pdev->driver)
68 return NULL;
69
70 if (!try_module_get(pdev->driver->driver.owner))
71 return NULL;
72
73 return pdev->driver;
74}
75
76/**
77 * eeh_pcid_put - Dereference on the PCI device driver
78 * @pdev: PCI device
79 *
80 * The function is called to do dereference on the PCI device
81 * driver of the indicated PCI device.
82 */
83static inline void eeh_pcid_put(struct pci_dev *pdev)
84{
85 if (!pdev || !pdev->driver)
86 return;
87
88 module_put(pdev->driver->driver.owner);
89}
90
91/**
92 * eeh_disable_irq - Disable interrupt for the recovering device
93 * @dev: PCI device
94 *
95 * This routine must be called when reporting temporary or permanent
96 * error to the particular PCI device to disable interrupt of that
97 * device. If the device has enabled MSI or MSI-X interrupt, we needn't
98 * do real work because EEH should freeze DMA transfers for those PCI
99 * devices encountering EEH errors, which includes MSI or MSI-X.
100 */
101static void eeh_disable_irq(struct pci_dev *dev)
102{
103 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
104
105 /* Don't disable MSI and MSI-X interrupts. They are
106 * effectively disabled by the DMA Stopped state
107 * when an EEH error occurs.
108 */
109 if (dev->msi_enabled || dev->msix_enabled)
110 return;
111
112 if (!irq_has_action(dev->irq))
113 return;
114
115 edev->mode |= EEH_DEV_IRQ_DISABLED;
116 disable_irq_nosync(dev->irq);
117}
118
119/**
120 * eeh_enable_irq - Enable interrupt for the recovering device
121 * @dev: PCI device
122 *
123 * This routine must be called to enable interrupt while failed
124 * device could be resumed.
125 */
126static void eeh_enable_irq(struct pci_dev *dev)
127{
128 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
129
130 if ((edev->mode) & EEH_DEV_IRQ_DISABLED) {
131 edev->mode &= ~EEH_DEV_IRQ_DISABLED;
132 /*
133 * FIXME !!!!!
134 *
135 * This is just ass backwards. This maze has
136 * unbalanced irq_enable/disable calls. So instead of
137 * finding the root cause it works around the warning
138 * in the irq_enable code by conditionally calling
139 * into it.
140 *
141 * That's just wrong.The warning in the core code is
142 * there to tell people to fix their asymmetries in
143 * their own code, not by abusing the core information
144 * to avoid it.
145 *
146 * I so wish that the assymetry would be the other way
147 * round and a few more irq_disable calls render that
148 * shit unusable forever.
149 *
150 * tglx
151 */
152 if (irqd_irq_disabled(irq_get_irq_data(dev->irq)))
153 enable_irq(dev->irq);
154 }
155}
156
157static bool eeh_dev_removed(struct eeh_dev *edev)
158{
159 /* EEH device removed ? */
160 if (!edev || (edev->mode & EEH_DEV_REMOVED))
161 return true;
162
163 return false;
164}
165
166static void *eeh_dev_save_state(void *data, void *userdata)
167{
168 struct eeh_dev *edev = data;
169 struct pci_dev *pdev;
170
171 if (!edev)
172 return NULL;
173
174 /*
175 * We cannot access the config space on some adapters.
176 * Otherwise, it will cause fenced PHB. We don't save
177 * the content in their config space and will restore
178 * from the initial config space saved when the EEH
179 * device is created.
180 */
181 if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED))
182 return NULL;
183
184 pdev = eeh_dev_to_pci_dev(edev);
185 if (!pdev)
186 return NULL;
187
188 pci_save_state(pdev);
189 return NULL;
190}
191
192/**
193 * eeh_report_error - Report pci error to each device driver
194 * @data: eeh device
195 * @userdata: return value
196 *
197 * Report an EEH error to each device driver, collect up and
198 * merge the device driver responses. Cumulative response
199 * passed back in "userdata".
200 */
201static void *eeh_report_error(void *data, void *userdata)
202{
203 struct eeh_dev *edev = (struct eeh_dev *)data;
204 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
205 enum pci_ers_result rc, *res = userdata;
206 struct pci_driver *driver;
207
208 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
209 return NULL;
210 dev->error_state = pci_channel_io_frozen;
211
212 driver = eeh_pcid_get(dev);
213 if (!driver) return NULL;
214
215 eeh_disable_irq(dev);
216
217 if (!driver->err_handler ||
218 !driver->err_handler->error_detected) {
219 eeh_pcid_put(dev);
220 return NULL;
221 }
222
223 rc = driver->err_handler->error_detected(dev, pci_channel_io_frozen);
224
225 /* A driver that needs a reset trumps all others */
226 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
227 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
228
229 edev->in_error = true;
230 eeh_pcid_put(dev);
231 return NULL;
232}
233
234/**
235 * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
236 * @data: eeh device
237 * @userdata: return value
238 *
239 * Tells each device driver that IO ports, MMIO and config space I/O
240 * are now enabled. Collects up and merges the device driver responses.
241 * Cumulative response passed back in "userdata".
242 */
243static void *eeh_report_mmio_enabled(void *data, void *userdata)
244{
245 struct eeh_dev *edev = (struct eeh_dev *)data;
246 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
247 enum pci_ers_result rc, *res = userdata;
248 struct pci_driver *driver;
249
250 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
251 return NULL;
252
253 driver = eeh_pcid_get(dev);
254 if (!driver) return NULL;
255
256 if (!driver->err_handler ||
257 !driver->err_handler->mmio_enabled ||
258 (edev->mode & EEH_DEV_NO_HANDLER)) {
259 eeh_pcid_put(dev);
260 return NULL;
261 }
262
263 rc = driver->err_handler->mmio_enabled(dev);
264
265 /* A driver that needs a reset trumps all others */
266 if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
267 if (*res == PCI_ERS_RESULT_NONE) *res = rc;
268
269 eeh_pcid_put(dev);
270 return NULL;
271}
272
273/**
274 * eeh_report_reset - Tell device that slot has been reset
275 * @data: eeh device
276 * @userdata: return value
277 *
278 * This routine must be called while EEH tries to reset particular
279 * PCI device so that the associated PCI device driver could take
280 * some actions, usually to save data the driver needs so that the
281 * driver can work again while the device is recovered.
282 */
283static void *eeh_report_reset(void *data, void *userdata)
284{
285 struct eeh_dev *edev = (struct eeh_dev *)data;
286 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
287 enum pci_ers_result rc, *res = userdata;
288 struct pci_driver *driver;
289
290 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
291 return NULL;
292 dev->error_state = pci_channel_io_normal;
293
294 driver = eeh_pcid_get(dev);
295 if (!driver) return NULL;
296
297 eeh_enable_irq(dev);
298
299 if (!driver->err_handler ||
300 !driver->err_handler->slot_reset ||
301 (edev->mode & EEH_DEV_NO_HANDLER) ||
302 (!edev->in_error)) {
303 eeh_pcid_put(dev);
304 return NULL;
305 }
306
307 rc = driver->err_handler->slot_reset(dev);
308 if ((*res == PCI_ERS_RESULT_NONE) ||
309 (*res == PCI_ERS_RESULT_RECOVERED)) *res = rc;
310 if (*res == PCI_ERS_RESULT_DISCONNECT &&
311 rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
312
313 eeh_pcid_put(dev);
314 return NULL;
315}
316
317static void *eeh_dev_restore_state(void *data, void *userdata)
318{
319 struct eeh_dev *edev = data;
320 struct pci_dev *pdev;
321
322 if (!edev)
323 return NULL;
324
325 /*
326 * The content in the config space isn't saved because
327 * the blocked config space on some adapters. We have
328 * to restore the initial saved config space when the
329 * EEH device is created.
330 */
331 if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED)) {
332 if (list_is_last(&edev->list, &edev->pe->edevs))
333 eeh_pe_restore_bars(edev->pe);
334
335 return NULL;
336 }
337
338 pdev = eeh_dev_to_pci_dev(edev);
339 if (!pdev)
340 return NULL;
341
342 pci_restore_state(pdev);
343 return NULL;
344}
345
346/**
347 * eeh_report_resume - Tell device to resume normal operations
348 * @data: eeh device
349 * @userdata: return value
350 *
351 * This routine must be called to notify the device driver that it
352 * could resume so that the device driver can do some initialization
353 * to make the recovered device work again.
354 */
355static void *eeh_report_resume(void *data, void *userdata)
356{
357 struct eeh_dev *edev = (struct eeh_dev *)data;
358 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
359 bool was_in_error;
360 struct pci_driver *driver;
361
362 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
363 return NULL;
364 dev->error_state = pci_channel_io_normal;
365
366 driver = eeh_pcid_get(dev);
367 if (!driver) return NULL;
368
369 was_in_error = edev->in_error;
370 edev->in_error = false;
371 eeh_enable_irq(dev);
372
373 if (!driver->err_handler ||
374 !driver->err_handler->resume ||
375 (edev->mode & EEH_DEV_NO_HANDLER) || !was_in_error) {
376 edev->mode &= ~EEH_DEV_NO_HANDLER;
377 eeh_pcid_put(dev);
378 return NULL;
379 }
380
381 driver->err_handler->resume(dev);
382
383 eeh_pcid_put(dev);
384 return NULL;
385}
386
387/**
388 * eeh_report_failure - Tell device driver that device is dead.
389 * @data: eeh device
390 * @userdata: return value
391 *
392 * This informs the device driver that the device is permanently
393 * dead, and that no further recovery attempts will be made on it.
394 */
395static void *eeh_report_failure(void *data, void *userdata)
396{
397 struct eeh_dev *edev = (struct eeh_dev *)data;
398 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
399 struct pci_driver *driver;
400
401 if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
402 return NULL;
403 dev->error_state = pci_channel_io_perm_failure;
404
405 driver = eeh_pcid_get(dev);
406 if (!driver) return NULL;
407
408 eeh_disable_irq(dev);
409
410 if (!driver->err_handler ||
411 !driver->err_handler->error_detected) {
412 eeh_pcid_put(dev);
413 return NULL;
414 }
415
416 driver->err_handler->error_detected(dev, pci_channel_io_perm_failure);
417
418 eeh_pcid_put(dev);
419 return NULL;
420}
421
422static void *eeh_add_virt_device(void *data, void *userdata)
423{
424 struct pci_driver *driver;
425 struct eeh_dev *edev = (struct eeh_dev *)data;
426 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
427 struct pci_dn *pdn = eeh_dev_to_pdn(edev);
428
429 if (!(edev->physfn)) {
430 pr_warn("%s: EEH dev %04x:%02x:%02x.%01x not for VF\n",
431 __func__, edev->phb->global_number, pdn->busno,
432 PCI_SLOT(pdn->devfn), PCI_FUNC(pdn->devfn));
433 return NULL;
434 }
435
436 driver = eeh_pcid_get(dev);
437 if (driver) {
438 eeh_pcid_put(dev);
439 if (driver->err_handler)
440 return NULL;
441 }
442
443#ifdef CONFIG_PPC_POWERNV
444 pci_iov_add_virtfn(edev->physfn, pdn->vf_index, 0);
445#endif
446 return NULL;
447}
448
449static void *eeh_rmv_device(void *data, void *userdata)
450{
451 struct pci_driver *driver;
452 struct eeh_dev *edev = (struct eeh_dev *)data;
453 struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
454 struct eeh_rmv_data *rmv_data = (struct eeh_rmv_data *)userdata;
455 int *removed = rmv_data ? &rmv_data->removed : NULL;
456
457 /*
458 * Actually, we should remove the PCI bridges as well.
459 * However, that's lots of complexity to do that,
460 * particularly some of devices under the bridge might
461 * support EEH. So we just care about PCI devices for
462 * simplicity here.
463 */
464 if (!dev || (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE))
465 return NULL;
466
467 /*
468 * We rely on count-based pcibios_release_device() to
469 * detach permanently offlined PEs. Unfortunately, that's
470 * not reliable enough. We might have the permanently
471 * offlined PEs attached, but we needn't take care of
472 * them and their child devices.
473 */
474 if (eeh_dev_removed(edev))
475 return NULL;
476
477 driver = eeh_pcid_get(dev);
478 if (driver) {
479 eeh_pcid_put(dev);
480 if (removed &&
481 eeh_pe_passed(edev->pe))
482 return NULL;
483 if (removed &&
484 driver->err_handler &&
485 driver->err_handler->error_detected &&
486 driver->err_handler->slot_reset)
487 return NULL;
488 }
489
490 /* Remove it from PCI subsystem */
491 pr_debug("EEH: Removing %s without EEH sensitive driver\n",
492 pci_name(dev));
493 edev->bus = dev->bus;
494 edev->mode |= EEH_DEV_DISCONNECTED;
495 if (removed)
496 (*removed)++;
497
498 if (edev->physfn) {
499#ifdef CONFIG_PPC_POWERNV
500 struct pci_dn *pdn = eeh_dev_to_pdn(edev);
501
502 pci_iov_remove_virtfn(edev->physfn, pdn->vf_index, 0);
503 edev->pdev = NULL;
504
505 /*
506 * We have to set the VF PE number to invalid one, which is
507 * required to plug the VF successfully.
508 */
509 pdn->pe_number = IODA_INVALID_PE;
510#endif
511 if (rmv_data)
512 list_add(&edev->rmv_list, &rmv_data->edev_list);
513 } else {
514 pci_lock_rescan_remove();
515 pci_stop_and_remove_bus_device(dev);
516 pci_unlock_rescan_remove();
517 }
518
519 return NULL;
520}
521
522static void *eeh_pe_detach_dev(void *data, void *userdata)
523{
524 struct eeh_pe *pe = (struct eeh_pe *)data;
525 struct eeh_dev *edev, *tmp;
526
527 eeh_pe_for_each_dev(pe, edev, tmp) {
528 if (!(edev->mode & EEH_DEV_DISCONNECTED))
529 continue;
530
531 edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
532 eeh_rmv_from_parent_pe(edev);
533 }
534
535 return NULL;
536}
537
538/*
539 * Explicitly clear PE's frozen state for PowerNV where
540 * we have frozen PE until BAR restore is completed. It's
541 * harmless to clear it for pSeries. To be consistent with
542 * PE reset (for 3 times), we try to clear the frozen state
543 * for 3 times as well.
544 */
545static void *__eeh_clear_pe_frozen_state(void *data, void *flag)
546{
547 struct eeh_pe *pe = (struct eeh_pe *)data;
548 bool clear_sw_state = *(bool *)flag;
549 int i, rc = 1;
550
551 for (i = 0; rc && i < 3; i++)
552 rc = eeh_unfreeze_pe(pe, clear_sw_state);
553
554 /* Stop immediately on any errors */
555 if (rc) {
556 pr_warn("%s: Failure %d unfreezing PHB#%x-PE#%x\n",
557 __func__, rc, pe->phb->global_number, pe->addr);
558 return (void *)pe;
559 }
560
561 return NULL;
562}
563
564static int eeh_clear_pe_frozen_state(struct eeh_pe *pe,
565 bool clear_sw_state)
566{
567 void *rc;
568
569 rc = eeh_pe_traverse(pe, __eeh_clear_pe_frozen_state, &clear_sw_state);
570 if (!rc)
571 eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
572
573 return rc ? -EIO : 0;
574}
575
576int eeh_pe_reset_and_recover(struct eeh_pe *pe)
577{
578 int ret;
579
580 /* Bail if the PE is being recovered */
581 if (pe->state & EEH_PE_RECOVERING)
582 return 0;
583
584 /* Put the PE into recovery mode */
585 eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
586
587 /* Save states */
588 eeh_pe_dev_traverse(pe, eeh_dev_save_state, NULL);
589
590 /* Issue reset */
591 ret = eeh_pe_reset_full(pe);
592 if (ret) {
593 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
594 return ret;
595 }
596
597 /* Unfreeze the PE */
598 ret = eeh_clear_pe_frozen_state(pe, true);
599 if (ret) {
600 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
601 return ret;
602 }
603
604 /* Restore device state */
605 eeh_pe_dev_traverse(pe, eeh_dev_restore_state, NULL);
606
607 /* Clear recovery mode */
608 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
609
610 return 0;
611}
612
613/**
614 * eeh_reset_device - Perform actual reset of a pci slot
615 * @pe: EEH PE
616 * @bus: PCI bus corresponding to the isolcated slot
617 *
618 * This routine must be called to do reset on the indicated PE.
619 * During the reset, udev might be invoked because those affected
620 * PCI devices will be removed and then added.
621 */
622static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus,
623 struct eeh_rmv_data *rmv_data)
624{
625 struct pci_bus *frozen_bus = eeh_pe_bus_get(pe);
626 struct timeval tstamp;
627 int cnt, rc;
628 struct eeh_dev *edev;
629
630 /* pcibios will clear the counter; save the value */
631 cnt = pe->freeze_count;
632 tstamp = pe->tstamp;
633
634 /*
635 * We don't remove the corresponding PE instances because
636 * we need the information afterwords. The attached EEH
637 * devices are expected to be attached soon when calling
638 * into pci_hp_add_devices().
639 */
640 eeh_pe_state_mark(pe, EEH_PE_KEEP);
641 if (bus) {
642 if (pe->type & EEH_PE_VF) {
643 eeh_pe_dev_traverse(pe, eeh_rmv_device, NULL);
644 } else {
645 pci_lock_rescan_remove();
646 pci_hp_remove_devices(bus);
647 pci_unlock_rescan_remove();
648 }
649 } else if (frozen_bus) {
650 eeh_pe_dev_traverse(pe, eeh_rmv_device, rmv_data);
651 }
652
653 /*
654 * Reset the pci controller. (Asserts RST#; resets config space).
655 * Reconfigure bridges and devices. Don't try to bring the system
656 * up if the reset failed for some reason.
657 *
658 * During the reset, it's very dangerous to have uncontrolled PCI
659 * config accesses. So we prefer to block them. However, controlled
660 * PCI config accesses initiated from EEH itself are allowed.
661 */
662 rc = eeh_pe_reset_full(pe);
663 if (rc)
664 return rc;
665
666 pci_lock_rescan_remove();
667
668 /* Restore PE */
669 eeh_ops->configure_bridge(pe);
670 eeh_pe_restore_bars(pe);
671
672 /* Clear frozen state */
673 rc = eeh_clear_pe_frozen_state(pe, false);
674 if (rc) {
675 pci_unlock_rescan_remove();
676 return rc;
677 }
678
679 /* Give the system 5 seconds to finish running the user-space
680 * hotplug shutdown scripts, e.g. ifdown for ethernet. Yes,
681 * this is a hack, but if we don't do this, and try to bring
682 * the device up before the scripts have taken it down,
683 * potentially weird things happen.
684 */
685 if (bus) {
686 pr_info("EEH: Sleep 5s ahead of complete hotplug\n");
687 ssleep(5);
688
689 /*
690 * The EEH device is still connected with its parent
691 * PE. We should disconnect it so the binding can be
692 * rebuilt when adding PCI devices.
693 */
694 edev = list_first_entry(&pe->edevs, struct eeh_dev, list);
695 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
696 if (pe->type & EEH_PE_VF) {
697 eeh_add_virt_device(edev, NULL);
698 } else {
699 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
700 pci_hp_add_devices(bus);
701 }
702 } else if (frozen_bus && rmv_data->removed) {
703 pr_info("EEH: Sleep 5s ahead of partial hotplug\n");
704 ssleep(5);
705
706 edev = list_first_entry(&pe->edevs, struct eeh_dev, list);
707 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
708 if (pe->type & EEH_PE_VF)
709 eeh_add_virt_device(edev, NULL);
710 else
711 pci_hp_add_devices(frozen_bus);
712 }
713 eeh_pe_state_clear(pe, EEH_PE_KEEP);
714
715 pe->tstamp = tstamp;
716 pe->freeze_count = cnt;
717
718 pci_unlock_rescan_remove();
719 return 0;
720}
721
722/* The longest amount of time to wait for a pci device
723 * to come back on line, in seconds.
724 */
725#define MAX_WAIT_FOR_RECOVERY 300
726
727static void eeh_handle_normal_event(struct eeh_pe *pe)
728{
729 struct pci_bus *frozen_bus;
730 struct eeh_dev *edev, *tmp;
731 int rc = 0;
732 enum pci_ers_result result = PCI_ERS_RESULT_NONE;
733 struct eeh_rmv_data rmv_data = {LIST_HEAD_INIT(rmv_data.edev_list), 0};
734
735 frozen_bus = eeh_pe_bus_get(pe);
736 if (!frozen_bus) {
737 pr_err("%s: Cannot find PCI bus for PHB#%x-PE#%x\n",
738 __func__, pe->phb->global_number, pe->addr);
739 return;
740 }
741
742 eeh_pe_update_time_stamp(pe);
743 pe->freeze_count++;
744 if (pe->freeze_count > eeh_max_freezes)
745 goto excess_failures;
746 pr_warn("EEH: This PCI device has failed %d times in the last hour\n",
747 pe->freeze_count);
748
749 /* Walk the various device drivers attached to this slot through
750 * a reset sequence, giving each an opportunity to do what it needs
751 * to accomplish the reset. Each child gets a report of the
752 * status ... if any child can't handle the reset, then the entire
753 * slot is dlpar removed and added.
754 *
755 * When the PHB is fenced, we have to issue a reset to recover from
756 * the error. Override the result if necessary to have partially
757 * hotplug for this case.
758 */
759 pr_info("EEH: Notify device drivers to shutdown\n");
760 eeh_pe_dev_traverse(pe, eeh_report_error, &result);
761 if ((pe->type & EEH_PE_PHB) &&
762 result != PCI_ERS_RESULT_NONE &&
763 result != PCI_ERS_RESULT_NEED_RESET)
764 result = PCI_ERS_RESULT_NEED_RESET;
765
766 /* Get the current PCI slot state. This can take a long time,
767 * sometimes over 300 seconds for certain systems.
768 */
769 rc = eeh_ops->wait_state(pe, MAX_WAIT_FOR_RECOVERY*1000);
770 if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
771 pr_warn("EEH: Permanent failure\n");
772 goto hard_fail;
773 }
774
775 /* Since rtas may enable MMIO when posting the error log,
776 * don't post the error log until after all dev drivers
777 * have been informed.
778 */
779 pr_info("EEH: Collect temporary log\n");
780 eeh_slot_error_detail(pe, EEH_LOG_TEMP);
781
782 /* If all device drivers were EEH-unaware, then shut
783 * down all of the device drivers, and hope they
784 * go down willingly, without panicing the system.
785 */
786 if (result == PCI_ERS_RESULT_NONE) {
787 pr_info("EEH: Reset with hotplug activity\n");
788 rc = eeh_reset_device(pe, frozen_bus, NULL);
789 if (rc) {
790 pr_warn("%s: Unable to reset, err=%d\n",
791 __func__, rc);
792 goto hard_fail;
793 }
794 }
795
796 /* If all devices reported they can proceed, then re-enable MMIO */
797 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
798 pr_info("EEH: Enable I/O for affected devices\n");
799 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
800
801 if (rc < 0)
802 goto hard_fail;
803 if (rc) {
804 result = PCI_ERS_RESULT_NEED_RESET;
805 } else {
806 pr_info("EEH: Notify device drivers to resume I/O\n");
807 eeh_pe_dev_traverse(pe, eeh_report_mmio_enabled, &result);
808 }
809 }
810
811 /* If all devices reported they can proceed, then re-enable DMA */
812 if (result == PCI_ERS_RESULT_CAN_RECOVER) {
813 pr_info("EEH: Enabled DMA for affected devices\n");
814 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
815
816 if (rc < 0)
817 goto hard_fail;
818 if (rc) {
819 result = PCI_ERS_RESULT_NEED_RESET;
820 } else {
821 /*
822 * We didn't do PE reset for the case. The PE
823 * is still in frozen state. Clear it before
824 * resuming the PE.
825 */
826 eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
827 result = PCI_ERS_RESULT_RECOVERED;
828 }
829 }
830
831 /* If any device has a hard failure, then shut off everything. */
832 if (result == PCI_ERS_RESULT_DISCONNECT) {
833 pr_warn("EEH: Device driver gave up\n");
834 goto hard_fail;
835 }
836
837 /* If any device called out for a reset, then reset the slot */
838 if (result == PCI_ERS_RESULT_NEED_RESET) {
839 pr_info("EEH: Reset without hotplug activity\n");
840 rc = eeh_reset_device(pe, NULL, &rmv_data);
841 if (rc) {
842 pr_warn("%s: Cannot reset, err=%d\n",
843 __func__, rc);
844 goto hard_fail;
845 }
846
847 pr_info("EEH: Notify device drivers "
848 "the completion of reset\n");
849 result = PCI_ERS_RESULT_NONE;
850 eeh_pe_dev_traverse(pe, eeh_report_reset, &result);
851 }
852
853 /* All devices should claim they have recovered by now. */
854 if ((result != PCI_ERS_RESULT_RECOVERED) &&
855 (result != PCI_ERS_RESULT_NONE)) {
856 pr_warn("EEH: Not recovered\n");
857 goto hard_fail;
858 }
859
860 /*
861 * For those hot removed VFs, we should add back them after PF get
862 * recovered properly.
863 */
864 list_for_each_entry_safe(edev, tmp, &rmv_data.edev_list, rmv_list) {
865 eeh_add_virt_device(edev, NULL);
866 list_del(&edev->rmv_list);
867 }
868
869 /* Tell all device drivers that they can resume operations */
870 pr_info("EEH: Notify device driver to resume\n");
871 eeh_pe_dev_traverse(pe, eeh_report_resume, NULL);
872
873 return;
874
875excess_failures:
876 /*
877 * About 90% of all real-life EEH failures in the field
878 * are due to poorly seated PCI cards. Only 10% or so are
879 * due to actual, failed cards.
880 */
881 pr_err("EEH: PHB#%x-PE#%x has failed %d times in the\n"
882 "last hour and has been permanently disabled.\n"
883 "Please try reseating or replacing it.\n",
884 pe->phb->global_number, pe->addr,
885 pe->freeze_count);
886 goto perm_error;
887
888hard_fail:
889 pr_err("EEH: Unable to recover from failure from PHB#%x-PE#%x.\n"
890 "Please try reseating or replacing it\n",
891 pe->phb->global_number, pe->addr);
892
893perm_error:
894 eeh_slot_error_detail(pe, EEH_LOG_PERM);
895
896 /* Notify all devices that they're about to go down. */
897 eeh_pe_dev_traverse(pe, eeh_report_failure, NULL);
898
899 /* Mark the PE to be removed permanently */
900 eeh_pe_state_mark(pe, EEH_PE_REMOVED);
901
902 /*
903 * Shut down the device drivers for good. We mark
904 * all removed devices correctly to avoid access
905 * the their PCI config any more.
906 */
907 if (frozen_bus) {
908 if (pe->type & EEH_PE_VF) {
909 eeh_pe_dev_traverse(pe, eeh_rmv_device, NULL);
910 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
911 } else {
912 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
913 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
914
915 pci_lock_rescan_remove();
916 pci_hp_remove_devices(frozen_bus);
917 pci_unlock_rescan_remove();
918 }
919 }
920}
921
922static void eeh_handle_special_event(void)
923{
924 struct eeh_pe *pe, *phb_pe;
925 struct pci_bus *bus;
926 struct pci_controller *hose;
927 unsigned long flags;
928 int rc;
929
930
931 do {
932 rc = eeh_ops->next_error(&pe);
933
934 switch (rc) {
935 case EEH_NEXT_ERR_DEAD_IOC:
936 /* Mark all PHBs in dead state */
937 eeh_serialize_lock(&flags);
938
939 /* Purge all events */
940 eeh_remove_event(NULL, true);
941
942 list_for_each_entry(hose, &hose_list, list_node) {
943 phb_pe = eeh_phb_pe_get(hose);
944 if (!phb_pe) continue;
945
946 eeh_pe_state_mark(phb_pe, EEH_PE_ISOLATED);
947 }
948
949 eeh_serialize_unlock(flags);
950
951 break;
952 case EEH_NEXT_ERR_FROZEN_PE:
953 case EEH_NEXT_ERR_FENCED_PHB:
954 case EEH_NEXT_ERR_DEAD_PHB:
955 /* Mark the PE in fenced state */
956 eeh_serialize_lock(&flags);
957
958 /* Purge all events of the PHB */
959 eeh_remove_event(pe, true);
960
961 if (rc == EEH_NEXT_ERR_DEAD_PHB)
962 eeh_pe_state_mark(pe, EEH_PE_ISOLATED);
963 else
964 eeh_pe_state_mark(pe,
965 EEH_PE_ISOLATED | EEH_PE_RECOVERING);
966
967 eeh_serialize_unlock(flags);
968
969 break;
970 case EEH_NEXT_ERR_NONE:
971 return;
972 default:
973 pr_warn("%s: Invalid value %d from next_error()\n",
974 __func__, rc);
975 return;
976 }
977
978 /*
979 * For fenced PHB and frozen PE, it's handled as normal
980 * event. We have to remove the affected PHBs for dead
981 * PHB and IOC
982 */
983 if (rc == EEH_NEXT_ERR_FROZEN_PE ||
984 rc == EEH_NEXT_ERR_FENCED_PHB) {
985 eeh_handle_normal_event(pe);
986 eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
987 } else {
988 pci_lock_rescan_remove();
989 list_for_each_entry(hose, &hose_list, list_node) {
990 phb_pe = eeh_phb_pe_get(hose);
991 if (!phb_pe ||
992 !(phb_pe->state & EEH_PE_ISOLATED) ||
993 (phb_pe->state & EEH_PE_RECOVERING))
994 continue;
995
996 /* Notify all devices to be down */
997 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
998 eeh_pe_dev_traverse(pe,
999 eeh_report_failure, NULL);
1000 bus = eeh_pe_bus_get(phb_pe);
1001 if (!bus) {
1002 pr_err("%s: Cannot find PCI bus for "
1003 "PHB#%x-PE#%x\n",
1004 __func__,
1005 pe->phb->global_number,
1006 pe->addr);
1007 break;
1008 }
1009 pci_hp_remove_devices(bus);
1010 }
1011 pci_unlock_rescan_remove();
1012 }
1013
1014 /*
1015 * If we have detected dead IOC, we needn't proceed
1016 * any more since all PHBs would have been removed
1017 */
1018 if (rc == EEH_NEXT_ERR_DEAD_IOC)
1019 break;
1020 } while (rc != EEH_NEXT_ERR_NONE);
1021}
1022
1023/**
1024 * eeh_handle_event - Reset a PCI device after hard lockup.
1025 * @pe: EEH PE
1026 *
1027 * While PHB detects address or data parity errors on particular PCI
1028 * slot, the associated PE will be frozen. Besides, DMA's occurring
1029 * to wild addresses (which usually happen due to bugs in device
1030 * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
1031 * #PERR or other misc PCI-related errors also can trigger EEH errors.
1032 *
1033 * Recovery process consists of unplugging the device driver (which
1034 * generated hotplug events to userspace), then issuing a PCI #RST to
1035 * the device, then reconfiguring the PCI config space for all bridges
1036 * & devices under this slot, and then finally restarting the device
1037 * drivers (which cause a second set of hotplug events to go out to
1038 * userspace).
1039 */
1040void eeh_handle_event(struct eeh_pe *pe)
1041{
1042 if (pe)
1043 eeh_handle_normal_event(pe);
1044 else
1045 eeh_handle_special_event();
1046}