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 struct eeh_rmv_data {
  38         struct list_head removed_vf_list;
  39         int removed_dev_count;
  40 };
  41
  42 static int eeh_result_priority(enum pci_ers_result result)
  43 {
  44         switch (result) {
  45         case PCI_ERS_RESULT_NONE:
  46                 return 1;
  47         case PCI_ERS_RESULT_NO_AER_DRIVER:
  48                 return 2;
  49         case PCI_ERS_RESULT_RECOVERED:
  50                 return 3;
  51         case PCI_ERS_RESULT_CAN_RECOVER:
  52                 return 4;
  53         case PCI_ERS_RESULT_DISCONNECT:
  54                 return 5;
  55         case PCI_ERS_RESULT_NEED_RESET:
  56                 return 6;
  57         default:
  58                 WARN_ONCE(1, "Unknown pci_ers_result value: %d\n", (int)result);
  59                 return 0;
  60         }
  61 };
  62
  63 static const char *pci_ers_result_name(enum pci_ers_result result)
  64 {
  65         switch (result) {
  66         case PCI_ERS_RESULT_NONE:
  67                 return "none";
  68         case PCI_ERS_RESULT_CAN_RECOVER:
  69                 return "can recover";
  70         case PCI_ERS_RESULT_NEED_RESET:
  71                 return "need reset";
  72         case PCI_ERS_RESULT_DISCONNECT:
  73                 return "disconnect";
  74         case PCI_ERS_RESULT_RECOVERED:
  75                 return "recovered";
  76         case PCI_ERS_RESULT_NO_AER_DRIVER:
  77                 return "no AER driver";
  78         default:
  79                 WARN_ONCE(1, "Unknown result type: %d\n", (int)result);
  80                 return "unknown";
  81         }
  82 };
  83
  84 static __printf(2, 3) void eeh_edev_info(const struct eeh_dev *edev,
  85                                          const char *fmt, ...)
  86 {
  87         struct va_format vaf;
  88         va_list args;
  89
  90         va_start(args, fmt);
  91
  92         vaf.fmt = fmt;
  93         vaf.va = &args;
  94
  95         printk(KERN_INFO "EEH: PE#%x (PCI %s): %pV\n", edev->pe_config_addr,
  96                edev->pdev ? dev_name(&edev->pdev->dev) : "none", &vaf);
  97
  98         va_end(args);
  99 }
 100
 101 static enum pci_ers_result pci_ers_merge_result(enum pci_ers_result old,
 102                                                 enum pci_ers_result new)
 103 {
 104         if (eeh_result_priority(new) > eeh_result_priority(old))
 105                 return new;
 106         return old;
 107 }
 108
 109 static bool eeh_dev_removed(struct eeh_dev *edev)
 110 {
 111         return !edev || (edev->mode & EEH_DEV_REMOVED);
 112 }
 113
 114 static bool eeh_edev_actionable(struct eeh_dev *edev)
 115 {
 116         return (edev->pdev && !eeh_dev_removed(edev) &&
 117                 !eeh_pe_passed(edev->pe));
 118 }
 119
 120 /**
 121  * eeh_pcid_get - Get the PCI device driver
 122  * @pdev: PCI device
 123  *
 124  * The function is used to retrieve the PCI device driver for
 125  * the indicated PCI device. Besides, we will increase the reference
 126  * of the PCI device driver to prevent that being unloaded on
 127  * the fly. Otherwise, kernel crash would be seen.
 128  */
 129 static inline struct pci_driver *eeh_pcid_get(struct pci_dev *pdev)
 130 {
 131         if (!pdev || !pdev->driver)
 132                 return NULL;
 133
 134         if (!try_module_get(pdev->driver->driver.owner))
 135                 return NULL;
 136
 137         return pdev->driver;
 138 }
 139
 140 /**
 141  * eeh_pcid_put - Dereference on the PCI device driver
 142  * @pdev: PCI device
 143  *
 144  * The function is called to do dereference on the PCI device
 145  * driver of the indicated PCI device.
 146  */
 147 static inline void eeh_pcid_put(struct pci_dev *pdev)
 148 {
 149         if (!pdev || !pdev->driver)
 150                 return;
 151
 152         module_put(pdev->driver->driver.owner);
 153 }
 154
 155 /**
 156  * eeh_disable_irq - Disable interrupt for the recovering device
 157  * @dev: PCI device
 158  *
 159  * This routine must be called when reporting temporary or permanent
 160  * error to the particular PCI device to disable interrupt of that
 161  * device. If the device has enabled MSI or MSI-X interrupt, we needn't
 162  * do real work because EEH should freeze DMA transfers for those PCI
 163  * devices encountering EEH errors, which includes MSI or MSI-X.
 164  */
 165 static void eeh_disable_irq(struct eeh_dev *edev)
 166 {
 167         /* Don't disable MSI and MSI-X interrupts. They are
 168          * effectively disabled by the DMA Stopped state
 169          * when an EEH error occurs.
 170          */
 171         if (edev->pdev->msi_enabled || edev->pdev->msix_enabled)
 172                 return;
 173
 174         if (!irq_has_action(edev->pdev->irq))
 175                 return;
 176
 177         edev->mode |= EEH_DEV_IRQ_DISABLED;
 178         disable_irq_nosync(edev->pdev->irq);
 179 }
 180
 181 /**
 182  * eeh_enable_irq - Enable interrupt for the recovering device
 183  * @dev: PCI device
 184  *
 185  * This routine must be called to enable interrupt while failed
 186  * device could be resumed.
 187  */
 188 static void eeh_enable_irq(struct eeh_dev *edev)
 189 {
 190         if ((edev->mode) & EEH_DEV_IRQ_DISABLED) {
 191                 edev->mode &= ~EEH_DEV_IRQ_DISABLED;
 192                 /*
 193                  * FIXME !!!!!
 194                  *
 195                  * This is just ass backwards. This maze has
 196                  * unbalanced irq_enable/disable calls. So instead of
 197                  * finding the root cause it works around the warning
 198                  * in the irq_enable code by conditionally calling
 199                  * into it.
 200                  *
 201                  * That's just wrong.The warning in the core code is
 202                  * there to tell people to fix their asymmetries in
 203                  * their own code, not by abusing the core information
 204                  * to avoid it.
 205                  *
 206                  * I so wish that the assymetry would be the other way
 207                  * round and a few more irq_disable calls render that
 208                  * shit unusable forever.
 209                  *
 210                  *      tglx
 211                  */
 212                 if (irqd_irq_disabled(irq_get_irq_data(edev->pdev->irq)))
 213                         enable_irq(edev->pdev->irq);
 214         }
 215 }
 216
 217 static void *eeh_dev_save_state(struct eeh_dev *edev, void *userdata)
 218 {
 219         struct pci_dev *pdev;
 220
 221         if (!edev)
 222                 return NULL;
 223
 224         /*
 225          * We cannot access the config space on some adapters.
 226          * Otherwise, it will cause fenced PHB. We don't save
 227          * the content in their config space and will restore
 228          * from the initial config space saved when the EEH
 229          * device is created.
 230          */
 231         if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED))
 232                 return NULL;
 233
 234         pdev = eeh_dev_to_pci_dev(edev);
 235         if (!pdev)
 236                 return NULL;
 237
 238         pci_save_state(pdev);
 239         return NULL;
 240 }
 241
 242 static void eeh_set_channel_state(struct eeh_pe *root, enum pci_channel_state s)
 243 {
 244         struct eeh_pe *pe;
 245         struct eeh_dev *edev, *tmp;
 246
 247         eeh_for_each_pe(root, pe)
 248                 eeh_pe_for_each_dev(pe, edev, tmp)
 249                         if (eeh_edev_actionable(edev))
 250                                 edev->pdev->error_state = s;
 251 }
 252
 253 static void eeh_set_irq_state(struct eeh_pe *root, bool enable)
 254 {
 255         struct eeh_pe *pe;
 256         struct eeh_dev *edev, *tmp;
 257
 258         eeh_for_each_pe(root, pe) {
 259                 eeh_pe_for_each_dev(pe, edev, tmp) {
 260                         if (!eeh_edev_actionable(edev))
 261                                 continue;
 262
 263                         if (!eeh_pcid_get(edev->pdev))
 264                                 continue;
 265
 266                         if (enable)
 267                                 eeh_enable_irq(edev);
 268                         else
 269                                 eeh_disable_irq(edev);
 270
 271                         eeh_pcid_put(edev->pdev);
 272                 }
 273         }
 274 }
 275
 276 typedef enum pci_ers_result (*eeh_report_fn)(struct eeh_dev *,
 277                                              struct pci_driver *);
 278 static void eeh_pe_report_edev(struct eeh_dev *edev, eeh_report_fn fn,
 279                                enum pci_ers_result *result)
 280 {
 281         struct pci_driver *driver;
 282         enum pci_ers_result new_result;
 283
 284         if (!edev->pdev) {
 285                 eeh_edev_info(edev, "no device");
 286                 return;
 287         }
 288         device_lock(&edev->pdev->dev);
 289         if (eeh_edev_actionable(edev)) {
 290                 driver = eeh_pcid_get(edev->pdev);
 291
 292                 if (!driver)
 293                         eeh_edev_info(edev, "no driver");
 294                 else if (!driver->err_handler)
 295                         eeh_edev_info(edev, "driver not EEH aware");
 296                 else if (edev->mode & EEH_DEV_NO_HANDLER)
 297                         eeh_edev_info(edev, "driver bound too late");
 298                 else {
 299                         new_result = fn(edev, driver);
 300                         eeh_edev_info(edev, "%s driver reports: '%s'",
 301                                       driver->name,
 302                                       pci_ers_result_name(new_result));
 303                         if (result)
 304                                 *result = pci_ers_merge_result(*result,
 305                                                                new_result);
 306                 }
 307                 if (driver)
 308                         eeh_pcid_put(edev->pdev);
 309         } else {
 310                 eeh_edev_info(edev, "not actionable (%d,%d,%d)", !!edev->pdev,
 311                               !eeh_dev_removed(edev), !eeh_pe_passed(edev->pe));
 312         }
 313         device_unlock(&edev->pdev->dev);
 314 }
 315
 316 static void eeh_pe_report(const char *name, struct eeh_pe *root,
 317                           eeh_report_fn fn, enum pci_ers_result *result)
 318 {
 319         struct eeh_pe *pe;
 320         struct eeh_dev *edev, *tmp;
 321
 322         pr_info("EEH: Beginning: '%s'\n", name);
 323         eeh_for_each_pe(root, pe) eeh_pe_for_each_dev(pe, edev, tmp)
 324                 eeh_pe_report_edev(edev, fn, result);
 325         if (result)
 326                 pr_info("EEH: Finished:'%s' with aggregate recovery state:'%s'\n",
 327                         name, pci_ers_result_name(*result));
 328         else
 329                 pr_info("EEH: Finished:'%s'", name);
 330 }
 331
 332 /**
 333  * eeh_report_error - Report pci error to each device driver
 334  * @edev: eeh device
 335  * @driver: device's PCI driver
 336  *
 337  * Report an EEH error to each device driver.
 338  */
 339 static enum pci_ers_result eeh_report_error(struct eeh_dev *edev,
 340                                             struct pci_driver *driver)
 341 {
 342         enum pci_ers_result rc;
 343         struct pci_dev *dev = edev->pdev;
 344
 345         if (!driver->err_handler->error_detected)
 346                 return PCI_ERS_RESULT_NONE;
 347
 348         eeh_edev_info(edev, "Invoking %s->error_detected(IO frozen)",
 349                       driver->name);
 350         rc = driver->err_handler->error_detected(dev, pci_channel_io_frozen);
 351
 352         edev->in_error = true;
 353         pci_uevent_ers(dev, PCI_ERS_RESULT_NONE);
 354         return rc;
 355 }
 356
 357 /**
 358  * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
 359  * @edev: eeh device
 360  * @driver: device's PCI driver
 361  *
 362  * Tells each device driver that IO ports, MMIO and config space I/O
 363  * are now enabled.
 364  */
 365 static enum pci_ers_result eeh_report_mmio_enabled(struct eeh_dev *edev,
 366                                                    struct pci_driver *driver)
 367 {
 368         if (!driver->err_handler->mmio_enabled)
 369                 return PCI_ERS_RESULT_NONE;
 370         eeh_edev_info(edev, "Invoking %s->mmio_enabled()", driver->name);
 371         return driver->err_handler->mmio_enabled(edev->pdev);
 372 }
 373
 374 /**
 375  * eeh_report_reset - Tell device that slot has been reset
 376  * @edev: eeh device
 377  * @driver: device's PCI driver
 378  *
 379  * This routine must be called while EEH tries to reset particular
 380  * PCI device so that the associated PCI device driver could take
 381  * some actions, usually to save data the driver needs so that the
 382  * driver can work again while the device is recovered.
 383  */
 384 static enum pci_ers_result eeh_report_reset(struct eeh_dev *edev,
 385                                             struct pci_driver *driver)
 386 {
 387         if (!driver->err_handler->slot_reset || !edev->in_error)
 388                 return PCI_ERS_RESULT_NONE;
 389         eeh_edev_info(edev, "Invoking %s->slot_reset()", driver->name);
 390         return driver->err_handler->slot_reset(edev->pdev);
 391 }
 392
 393 static void *eeh_dev_restore_state(struct eeh_dev *edev, void *userdata)
 394 {
 395         struct pci_dev *pdev;
 396
 397         if (!edev)
 398                 return NULL;
 399
 400         /*
 401          * The content in the config space isn't saved because
 402          * the blocked config space on some adapters. We have
 403          * to restore the initial saved config space when the
 404          * EEH device is created.
 405          */
 406         if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED)) {
 407                 if (list_is_last(&edev->entry, &edev->pe->edevs))
 408                         eeh_pe_restore_bars(edev->pe);
 409
 410                 return NULL;
 411         }
 412
 413         pdev = eeh_dev_to_pci_dev(edev);
 414         if (!pdev)
 415                 return NULL;
 416
 417         pci_restore_state(pdev);
 418         return NULL;
 419 }
 420
 421 /**
 422  * eeh_report_resume - Tell device to resume normal operations
 423  * @edev: eeh device
 424  * @driver: device's PCI driver
 425  *
 426  * This routine must be called to notify the device driver that it
 427  * could resume so that the device driver can do some initialization
 428  * to make the recovered device work again.
 429  */
 430 static enum pci_ers_result eeh_report_resume(struct eeh_dev *edev,
 431                                              struct pci_driver *driver)
 432 {
 433         if (!driver->err_handler->resume || !edev->in_error)
 434                 return PCI_ERS_RESULT_NONE;
 435
 436         eeh_edev_info(edev, "Invoking %s->resume()", driver->name);
 437         driver->err_handler->resume(edev->pdev);
 438
 439         pci_uevent_ers(edev->pdev, PCI_ERS_RESULT_RECOVERED);
 440 #ifdef CONFIG_PCI_IOV
 441         if (eeh_ops->notify_resume && eeh_dev_to_pdn(edev))
 442                 eeh_ops->notify_resume(eeh_dev_to_pdn(edev));
 443 #endif
 444         return PCI_ERS_RESULT_NONE;
 445 }
 446
 447 /**
 448  * eeh_report_failure - Tell device driver that device is dead.
 449  * @edev: eeh device
 450  * @driver: device's PCI driver
 451  *
 452  * This informs the device driver that the device is permanently
 453  * dead, and that no further recovery attempts will be made on it.
 454  */
 455 static enum pci_ers_result eeh_report_failure(struct eeh_dev *edev,
 456                                               struct pci_driver *driver)
 457 {
 458         enum pci_ers_result rc;
 459
 460         if (!driver->err_handler->error_detected)
 461                 return PCI_ERS_RESULT_NONE;
 462
 463         eeh_edev_info(edev, "Invoking %s->error_detected(permanent failure)",
 464                       driver->name);
 465         rc = driver->err_handler->error_detected(edev->pdev,
 466                                                  pci_channel_io_perm_failure);
 467
 468         pci_uevent_ers(edev->pdev, PCI_ERS_RESULT_DISCONNECT);
 469         return rc;
 470 }
 471
 472 static void *eeh_add_virt_device(struct eeh_dev *edev)
 473 {
 474         struct pci_driver *driver;
 475         struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 476         struct pci_dn *pdn = eeh_dev_to_pdn(edev);
 477
 478         if (!(edev->physfn)) {
 479                 pr_warn("%s: EEH dev %04x:%02x:%02x.%01x not for VF\n",
 480                         __func__, pdn->phb->global_number, pdn->busno,
 481                         PCI_SLOT(pdn->devfn), PCI_FUNC(pdn->devfn));
 482                 return NULL;
 483         }
 484
 485         driver = eeh_pcid_get(dev);
 486         if (driver) {
 487                 if (driver->err_handler) {
 488                         eeh_pcid_put(dev);
 489                         return NULL;
 490                 }
 491                 eeh_pcid_put(dev);
 492         }
 493
 494 #ifdef CONFIG_PCI_IOV
 495         pci_iov_add_virtfn(edev->physfn, pdn->vf_index);
 496 #endif
 497         return NULL;
 498 }
 499
 500 static void *eeh_rmv_device(struct eeh_dev *edev, void *userdata)
 501 {
 502         struct pci_driver *driver;
 503         struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 504         struct eeh_rmv_data *rmv_data = (struct eeh_rmv_data *)userdata;
 505
 506         /*
 507          * Actually, we should remove the PCI bridges as well.
 508          * However, that's lots of complexity to do that,
 509          * particularly some of devices under the bridge might
 510          * support EEH. So we just care about PCI devices for
 511          * simplicity here.
 512          */
 513         if (!eeh_edev_actionable(edev) ||
 514             (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE))
 515                 return NULL;
 516
 517         if (rmv_data) {
 518                 driver = eeh_pcid_get(dev);
 519                 if (driver) {
 520                         if (driver->err_handler &&
 521                             driver->err_handler->error_detected &&
 522                             driver->err_handler->slot_reset) {
 523                                 eeh_pcid_put(dev);
 524                                 return NULL;
 525                         }
 526                         eeh_pcid_put(dev);
 527                 }
 528         }
 529
 530         /* Remove it from PCI subsystem */
 531         pr_info("EEH: Removing %s without EEH sensitive driver\n",
 532                 pci_name(dev));
 533         edev->mode |= EEH_DEV_DISCONNECTED;
 534         if (rmv_data)
 535                 rmv_data->removed_dev_count++;
 536
 537         if (edev->physfn) {
 538 #ifdef CONFIG_PCI_IOV
 539                 struct pci_dn *pdn = eeh_dev_to_pdn(edev);
 540
 541                 pci_iov_remove_virtfn(edev->physfn, pdn->vf_index);
 542                 edev->pdev = NULL;
 543
 544                 /*
 545                  * We have to set the VF PE number to invalid one, which is
 546                  * required to plug the VF successfully.
 547                  */
 548                 pdn->pe_number = IODA_INVALID_PE;
 549 #endif
 550                 if (rmv_data)
 551                         list_add(&edev->rmv_entry, &rmv_data->removed_vf_list);
 552         } else {
 553                 pci_lock_rescan_remove();
 554                 pci_stop_and_remove_bus_device(dev);
 555                 pci_unlock_rescan_remove();
 556         }
 557
 558         return NULL;
 559 }
 560
 561 static void *eeh_pe_detach_dev(struct eeh_pe *pe, void *userdata)
 562 {
 563         struct eeh_dev *edev, *tmp;
 564
 565         eeh_pe_for_each_dev(pe, edev, tmp) {
 566                 if (!(edev->mode & EEH_DEV_DISCONNECTED))
 567                         continue;
 568
 569                 edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
 570                 eeh_rmv_from_parent_pe(edev);
 571         }
 572
 573         return NULL;
 574 }
 575
 576 /*
 577  * Explicitly clear PE's frozen state for PowerNV where
 578  * we have frozen PE until BAR restore is completed. It's
 579  * harmless to clear it for pSeries. To be consistent with
 580  * PE reset (for 3 times), we try to clear the frozen state
 581  * for 3 times as well.
 582  */
 583 static int eeh_clear_pe_frozen_state(struct eeh_pe *root, bool include_passed)
 584 {
 585         struct eeh_pe *pe;
 586         int i;
 587
 588         eeh_for_each_pe(root, pe) {
 589                 if (include_passed || !eeh_pe_passed(pe)) {
 590                         for (i = 0; i < 3; i++)
 591                                 if (!eeh_unfreeze_pe(pe))
 592                                         break;
 593                         if (i >= 3)
 594                                 return -EIO;
 595                 }
 596         }
 597         eeh_pe_state_clear(root, EEH_PE_ISOLATED, include_passed);
 598         return 0;
 599 }
 600
 601 int eeh_pe_reset_and_recover(struct eeh_pe *pe)
 602 {
 603         int ret;
 604
 605         /* Bail if the PE is being recovered */
 606         if (pe->state & EEH_PE_RECOVERING)
 607                 return 0;
 608
 609         /* Put the PE into recovery mode */
 610         eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
 611
 612         /* Save states */
 613         eeh_pe_dev_traverse(pe, eeh_dev_save_state, NULL);
 614
 615         /* Issue reset */
 616         ret = eeh_pe_reset_full(pe, true);
 617         if (ret) {
 618                 eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 619                 return ret;
 620         }
 621
 622         /* Unfreeze the PE */
 623         ret = eeh_clear_pe_frozen_state(pe, true);
 624         if (ret) {
 625                 eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 626                 return ret;
 627         }
 628
 629         /* Restore device state */
 630         eeh_pe_dev_traverse(pe, eeh_dev_restore_state, NULL);
 631
 632         /* Clear recovery mode */
 633         eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 634
 635         return 0;
 636 }
 637
 638 /**
 639  * eeh_reset_device - Perform actual reset of a pci slot
 640  * @driver_eeh_aware: Does the device's driver provide EEH support?
 641  * @pe: EEH PE
 642  * @bus: PCI bus corresponding to the isolcated slot
 643  * @rmv_data: Optional, list to record removed devices
 644  *
 645  * This routine must be called to do reset on the indicated PE.
 646  * During the reset, udev might be invoked because those affected
 647  * PCI devices will be removed and then added.
 648  */
 649 static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus,
 650                             struct eeh_rmv_data *rmv_data,
 651                             bool driver_eeh_aware)
 652 {
 653         time64_t tstamp;
 654         int cnt, rc;
 655         struct eeh_dev *edev;
 656         struct eeh_pe *tmp_pe;
 657         bool any_passed = false;
 658
 659         eeh_for_each_pe(pe, tmp_pe)
 660                 any_passed |= eeh_pe_passed(tmp_pe);
 661
 662         /* pcibios will clear the counter; save the value */
 663         cnt = pe->freeze_count;
 664         tstamp = pe->tstamp;
 665
 666         /*
 667          * We don't remove the corresponding PE instances because
 668          * we need the information afterwords. The attached EEH
 669          * devices are expected to be attached soon when calling
 670          * into pci_hp_add_devices().
 671          */
 672         eeh_pe_state_mark(pe, EEH_PE_KEEP);
 673         if (any_passed || driver_eeh_aware || (pe->type & EEH_PE_VF)) {
 674                 eeh_pe_dev_traverse(pe, eeh_rmv_device, rmv_data);
 675         } else {
 676                 pci_lock_rescan_remove();
 677                 pci_hp_remove_devices(bus);
 678                 pci_unlock_rescan_remove();
 679         }
 680
 681         /*
 682          * Reset the pci controller. (Asserts RST#; resets config space).
 683          * Reconfigure bridges and devices. Don't try to bring the system
 684          * up if the reset failed for some reason.
 685          *
 686          * During the reset, it's very dangerous to have uncontrolled PCI
 687          * config accesses. So we prefer to block them. However, controlled
 688          * PCI config accesses initiated from EEH itself are allowed.
 689          */
 690         rc = eeh_pe_reset_full(pe, false);
 691         if (rc)
 692                 return rc;
 693
 694         pci_lock_rescan_remove();
 695
 696         /* Restore PE */
 697         eeh_ops->configure_bridge(pe);
 698         eeh_pe_restore_bars(pe);
 699
 700         /* Clear frozen state */
 701         rc = eeh_clear_pe_frozen_state(pe, false);
 702         if (rc) {
 703                 pci_unlock_rescan_remove();
 704                 return rc;
 705         }
 706
 707         /* Give the system 5 seconds to finish running the user-space
 708          * hotplug shutdown scripts, e.g. ifdown for ethernet.  Yes,
 709          * this is a hack, but if we don't do this, and try to bring
 710          * the device up before the scripts have taken it down,
 711          * potentially weird things happen.
 712          */
 713         if (!driver_eeh_aware || rmv_data->removed_dev_count) {
 714                 pr_info("EEH: Sleep 5s ahead of %s hotplug\n",
 715                         (driver_eeh_aware ? "partial" : "complete"));
 716                 ssleep(5);
 717
 718                 /*
 719                  * The EEH device is still connected with its parent
 720                  * PE. We should disconnect it so the binding can be
 721                  * rebuilt when adding PCI devices.
 722                  */
 723                 edev = list_first_entry(&pe->edevs, struct eeh_dev, entry);
 724                 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
 725                 if (pe->type & EEH_PE_VF) {
 726                         eeh_add_virt_device(edev);
 727                 } else {
 728                         if (!driver_eeh_aware)
 729                                 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
 730                         pci_hp_add_devices(bus);
 731                 }
 732         }
 733         eeh_pe_state_clear(pe, EEH_PE_KEEP, true);
 734
 735         pe->tstamp = tstamp;
 736         pe->freeze_count = cnt;
 737
 738         pci_unlock_rescan_remove();
 739         return 0;
 740 }
 741
 742 /* The longest amount of time to wait for a pci device
 743  * to come back on line, in seconds.
 744  */
 745 #define MAX_WAIT_FOR_RECOVERY 300
 746
 747
 748 /* Walks the PE tree after processing an event to remove any stale PEs.
 749  *
 750  * NB: This needs to be recursive to ensure the leaf PEs get removed
 751  * before their parents do. Although this is possible to do recursively
 752  * we don't since this is easier to read and we need to garantee
 753  * the leaf nodes will be handled first.
 754  */
 755 static void eeh_pe_cleanup(struct eeh_pe *pe)
 756 {
 757         struct eeh_pe *child_pe, *tmp;
 758
 759         list_for_each_entry_safe(child_pe, tmp, &pe->child_list, child)
 760                 eeh_pe_cleanup(child_pe);
 761
 762         if (pe->state & EEH_PE_KEEP)
 763                 return;
 764
 765         if (!(pe->state & EEH_PE_INVALID))
 766                 return;
 767
 768         if (list_empty(&pe->edevs) && list_empty(&pe->child_list)) {
 769                 list_del(&pe->child);
 770                 kfree(pe);
 771         }
 772 }
 773
 774 /**
 775  * eeh_handle_normal_event - Handle EEH events on a specific PE
 776  * @pe: EEH PE - which should not be used after we return, as it may
 777  * have been invalidated.
 778  *
 779  * Attempts to recover the given PE.  If recovery fails or the PE has failed
 780  * too many times, remove the PE.
 781  *
 782  * While PHB detects address or data parity errors on particular PCI
 783  * slot, the associated PE will be frozen. Besides, DMA's occurring
 784  * to wild addresses (which usually happen due to bugs in device
 785  * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
 786  * #PERR or other misc PCI-related errors also can trigger EEH errors.
 787  *
 788  * Recovery process consists of unplugging the device driver (which
 789  * generated hotplug events to userspace), then issuing a PCI #RST to
 790  * the device, then reconfiguring the PCI config space for all bridges
 791  * & devices under this slot, and then finally restarting the device
 792  * drivers (which cause a second set of hotplug events to go out to
 793  * userspace).
 794  */
 795 void eeh_handle_normal_event(struct eeh_pe *pe)
 796 {
 797         struct pci_bus *bus;
 798         struct eeh_dev *edev, *tmp;
 799         struct eeh_pe *tmp_pe;
 800         int rc = 0;
 801         enum pci_ers_result result = PCI_ERS_RESULT_NONE;
 802         struct eeh_rmv_data rmv_data =
 803                 {LIST_HEAD_INIT(rmv_data.removed_vf_list), 0};
 804
 805         bus = eeh_pe_bus_get(pe);
 806         if (!bus) {
 807                 pr_err("%s: Cannot find PCI bus for PHB#%x-PE#%x\n",
 808                         __func__, pe->phb->global_number, pe->addr);
 809                 return;
 810         }
 811
 812         eeh_pe_update_time_stamp(pe);
 813         pe->freeze_count++;
 814         if (pe->freeze_count > eeh_max_freezes) {
 815                 pr_err("EEH: PHB#%x-PE#%x has failed %d times in the last hour and has been permanently disabled.\n",
 816                        pe->phb->global_number, pe->addr,
 817                        pe->freeze_count);
 818                 result = PCI_ERS_RESULT_DISCONNECT;
 819         }
 820
 821         eeh_for_each_pe(pe, tmp_pe)
 822                 eeh_pe_for_each_dev(tmp_pe, edev, tmp)
 823                         edev->mode &= ~EEH_DEV_NO_HANDLER;
 824
 825         /* Walk the various device drivers attached to this slot through
 826          * a reset sequence, giving each an opportunity to do what it needs
 827          * to accomplish the reset.  Each child gets a report of the
 828          * status ... if any child can't handle the reset, then the entire
 829          * slot is dlpar removed and added.
 830          *
 831          * When the PHB is fenced, we have to issue a reset to recover from
 832          * the error. Override the result if necessary to have partially
 833          * hotplug for this case.
 834          */
 835         if (result != PCI_ERS_RESULT_DISCONNECT) {
 836                 pr_warn("EEH: This PCI device has failed %d times in the last hour and will be permanently disabled after %d failures.\n",
 837                         pe->freeze_count, eeh_max_freezes);
 838                 pr_info("EEH: Notify device drivers to shutdown\n");
 839                 eeh_set_channel_state(pe, pci_channel_io_frozen);
 840                 eeh_set_irq_state(pe, false);
 841                 eeh_pe_report("error_detected(IO frozen)", pe,
 842                               eeh_report_error, &result);
 843                 if ((pe->type & EEH_PE_PHB) &&
 844                     result != PCI_ERS_RESULT_NONE &&
 845                     result != PCI_ERS_RESULT_NEED_RESET)
 846                         result = PCI_ERS_RESULT_NEED_RESET;
 847         }
 848
 849         /* Get the current PCI slot state. This can take a long time,
 850          * sometimes over 300 seconds for certain systems.
 851          */
 852         if (result != PCI_ERS_RESULT_DISCONNECT) {
 853                 rc = eeh_wait_state(pe, MAX_WAIT_FOR_RECOVERY*1000);
 854                 if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
 855                         pr_warn("EEH: Permanent failure\n");
 856                         result = PCI_ERS_RESULT_DISCONNECT;
 857                 }
 858         }
 859
 860         /* Since rtas may enable MMIO when posting the error log,
 861          * don't post the error log until after all dev drivers
 862          * have been informed.
 863          */
 864         if (result != PCI_ERS_RESULT_DISCONNECT) {
 865                 pr_info("EEH: Collect temporary log\n");
 866                 eeh_slot_error_detail(pe, EEH_LOG_TEMP);
 867         }
 868
 869         /* If all device drivers were EEH-unaware, then shut
 870          * down all of the device drivers, and hope they
 871          * go down willingly, without panicing the system.
 872          */
 873         if (result == PCI_ERS_RESULT_NONE) {
 874                 pr_info("EEH: Reset with hotplug activity\n");
 875                 rc = eeh_reset_device(pe, bus, NULL, false);
 876                 if (rc) {
 877                         pr_warn("%s: Unable to reset, err=%d\n",
 878                                 __func__, rc);
 879                         result = PCI_ERS_RESULT_DISCONNECT;
 880                 }
 881         }
 882
 883         /* If all devices reported they can proceed, then re-enable MMIO */
 884         if (result == PCI_ERS_RESULT_CAN_RECOVER) {
 885                 pr_info("EEH: Enable I/O for affected devices\n");
 886                 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
 887
 888                 if (rc < 0) {
 889                         result = PCI_ERS_RESULT_DISCONNECT;
 890                 } else if (rc) {
 891                         result = PCI_ERS_RESULT_NEED_RESET;
 892                 } else {
 893                         pr_info("EEH: Notify device drivers to resume I/O\n");
 894                         eeh_pe_report("mmio_enabled", pe,
 895                                       eeh_report_mmio_enabled, &result);
 896                 }
 897         }
 898
 899         /* If all devices reported they can proceed, then re-enable DMA */
 900         if (result == PCI_ERS_RESULT_CAN_RECOVER) {
 901                 pr_info("EEH: Enabled DMA for affected devices\n");
 902                 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
 903
 904                 if (rc < 0) {
 905                         result = PCI_ERS_RESULT_DISCONNECT;
 906                 } else if (rc) {
 907                         result = PCI_ERS_RESULT_NEED_RESET;
 908                 } else {
 909                         /*
 910                          * We didn't do PE reset for the case. The PE
 911                          * is still in frozen state. Clear it before
 912                          * resuming the PE.
 913                          */
 914                         eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
 915                         result = PCI_ERS_RESULT_RECOVERED;
 916                 }
 917         }
 918
 919         /* If any device called out for a reset, then reset the slot */
 920         if (result == PCI_ERS_RESULT_NEED_RESET) {
 921                 pr_info("EEH: Reset without hotplug activity\n");
 922                 rc = eeh_reset_device(pe, bus, &rmv_data, true);
 923                 if (rc) {
 924                         pr_warn("%s: Cannot reset, err=%d\n",
 925                                 __func__, rc);
 926                         result = PCI_ERS_RESULT_DISCONNECT;
 927                 } else {
 928                         result = PCI_ERS_RESULT_NONE;
 929                         eeh_set_channel_state(pe, pci_channel_io_normal);
 930                         eeh_set_irq_state(pe, true);
 931                         eeh_pe_report("slot_reset", pe, eeh_report_reset,
 932                                       &result);
 933                 }
 934         }
 935
 936         if ((result == PCI_ERS_RESULT_RECOVERED) ||
 937             (result == PCI_ERS_RESULT_NONE)) {
 938                 /*
 939                  * For those hot removed VFs, we should add back them after PF
 940                  * get recovered properly.
 941                  */
 942                 list_for_each_entry_safe(edev, tmp, &rmv_data.removed_vf_list,
 943                                          rmv_entry) {
 944                         eeh_add_virt_device(edev);
 945                         list_del(&edev->rmv_entry);
 946                 }
 947
 948                 /* Tell all device drivers that they can resume operations */
 949                 pr_info("EEH: Notify device driver to resume\n");
 950                 eeh_set_channel_state(pe, pci_channel_io_normal);
 951                 eeh_set_irq_state(pe, true);
 952                 eeh_pe_report("resume", pe, eeh_report_resume, NULL);
 953                 eeh_for_each_pe(pe, tmp_pe) {
 954                         eeh_pe_for_each_dev(tmp_pe, edev, tmp) {
 955                                 edev->mode &= ~EEH_DEV_NO_HANDLER;
 956                                 edev->in_error = false;
 957                         }
 958                 }
 959
 960                 pr_info("EEH: Recovery successful.\n");
 961         } else  {
 962                 /*
 963                  * About 90% of all real-life EEH failures in the field
 964                  * are due to poorly seated PCI cards. Only 10% or so are
 965                  * due to actual, failed cards.
 966                  */
 967                 pr_err("EEH: Unable to recover from failure from PHB#%x-PE#%x.\n"
 968                        "Please try reseating or replacing it\n",
 969                         pe->phb->global_number, pe->addr);
 970
 971                 eeh_slot_error_detail(pe, EEH_LOG_PERM);
 972
 973                 /* Notify all devices that they're about to go down. */
 974                 eeh_set_channel_state(pe, pci_channel_io_perm_failure);
 975                 eeh_set_irq_state(pe, false);
 976                 eeh_pe_report("error_detected(permanent failure)", pe,
 977                               eeh_report_failure, NULL);
 978
 979                 /* Mark the PE to be removed permanently */
 980                 eeh_pe_state_mark(pe, EEH_PE_REMOVED);
 981
 982                 /*
 983                  * Shut down the device drivers for good. We mark
 984                  * all removed devices correctly to avoid access
 985                  * the their PCI config any more.
 986                  */
 987                 if (pe->type & EEH_PE_VF) {
 988                         eeh_pe_dev_traverse(pe, eeh_rmv_device, NULL);
 989                         eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
 990                 } else {
 991                         eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
 992                         eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
 993
 994                         pci_lock_rescan_remove();
 995                         pci_hp_remove_devices(bus);
 996                         pci_unlock_rescan_remove();
 997                         /* The passed PE should no longer be used */
 998                         return;
 999                 }
1000         }
1001
1002         /*
1003          * Clean up any PEs without devices. While marked as EEH_PE_RECOVERYING
1004          * we don't want to modify the PE tree structure so we do it here.
1005          */
1006         eeh_pe_cleanup(pe);
1007         eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
1008 }
1009
1010 /**
1011  * eeh_handle_special_event - Handle EEH events without a specific failing PE
1012  *
1013  * Called when an EEH event is detected but can't be narrowed down to a
1014  * specific PE.  Iterates through possible failures and handles them as
1015  * necessary.
1016  */
1017 void eeh_handle_special_event(void)
1018 {
1019         struct eeh_pe *pe, *phb_pe, *tmp_pe;
1020         struct eeh_dev *edev, *tmp_edev;
1021         struct pci_bus *bus;
1022         struct pci_controller *hose;
1023         unsigned long flags;
1024         int rc;
1025
1026
1027         do {
1028                 rc = eeh_ops->next_error(&pe);
1029
1030                 switch (rc) {
1031                 case EEH_NEXT_ERR_DEAD_IOC:
1032                         /* Mark all PHBs in dead state */
1033                         eeh_serialize_lock(&flags);
1034
1035                         /* Purge all events */
1036                         eeh_remove_event(NULL, true);
1037
1038                         list_for_each_entry(hose, &hose_list, list_node) {
1039                                 phb_pe = eeh_phb_pe_get(hose);
1040                                 if (!phb_pe) continue;
1041
1042                                 eeh_pe_mark_isolated(phb_pe);
1043                         }
1044
1045                         eeh_serialize_unlock(flags);
1046
1047                         break;
1048                 case EEH_NEXT_ERR_FROZEN_PE:
1049                 case EEH_NEXT_ERR_FENCED_PHB:
1050                 case EEH_NEXT_ERR_DEAD_PHB:
1051                         /* Mark the PE in fenced state */
1052                         eeh_serialize_lock(&flags);
1053
1054                         /* Purge all events of the PHB */
1055                         eeh_remove_event(pe, true);
1056
1057                         if (rc != EEH_NEXT_ERR_DEAD_PHB)
1058                                 eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
1059                         eeh_pe_mark_isolated(pe);
1060
1061                         eeh_serialize_unlock(flags);
1062
1063                         break;
1064                 case EEH_NEXT_ERR_NONE:
1065                         return;
1066                 default:
1067                         pr_warn("%s: Invalid value %d from next_error()\n",
1068                                 __func__, rc);
1069                         return;
1070                 }
1071
1072                 /*
1073                  * For fenced PHB and frozen PE, it's handled as normal
1074                  * event. We have to remove the affected PHBs for dead
1075                  * PHB and IOC
1076                  */
1077                 if (rc == EEH_NEXT_ERR_FROZEN_PE ||
1078                     rc == EEH_NEXT_ERR_FENCED_PHB) {
1079                         eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
1080                         eeh_handle_normal_event(pe);
1081                 } else {
1082                         pci_lock_rescan_remove();
1083                         list_for_each_entry(hose, &hose_list, list_node) {
1084                                 phb_pe = eeh_phb_pe_get(hose);
1085                                 if (!phb_pe ||
1086                                     !(phb_pe->state & EEH_PE_ISOLATED) ||
1087                                     (phb_pe->state & EEH_PE_RECOVERING))
1088                                         continue;
1089
1090                                 eeh_for_each_pe(pe, tmp_pe)
1091                                         eeh_pe_for_each_dev(tmp_pe, edev, tmp_edev)
1092                                                 edev->mode &= ~EEH_DEV_NO_HANDLER;
1093
1094                                 /* Notify all devices to be down */
1095                                 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
1096                                 eeh_set_channel_state(pe, pci_channel_io_perm_failure);
1097                                 eeh_pe_report(
1098                                         "error_detected(permanent failure)", pe,
1099                                         eeh_report_failure, NULL);
1100                                 bus = eeh_pe_bus_get(phb_pe);
1101                                 if (!bus) {
1102                                         pr_err("%s: Cannot find PCI bus for "
1103                                                "PHB#%x-PE#%x\n",
1104                                                __func__,
1105                                                pe->phb->global_number,
1106                                                pe->addr);
1107                                         break;
1108                                 }
1109                                 pci_hp_remove_devices(bus);
1110                         }
1111                         pci_unlock_rescan_remove();
1112                 }
1113
1114                 /*
1115                  * If we have detected dead IOC, we needn't proceed
1116                  * any more since all PHBs would have been removed
1117                  */
1118                 if (rc == EEH_NEXT_ERR_DEAD_IOC)
1119                         break;
1120         } while (rc != EEH_NEXT_ERR_NONE);
1121 }