arch/powerpc/kernel/eeh_driver.c

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