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 <linux/pci_hotplug.h>
  31 #include <asm/eeh.h>
  32 #include <asm/eeh_event.h>
  33 #include <asm/ppc-pci.h>
  34 #include <asm/pci-bridge.h>
  35 #include <asm/prom.h>
  36 #include <asm/rtas.h>
  37
  38 struct eeh_rmv_data {
  39         struct list_head removed_vf_list;
  40         int removed_dev_count;
  41 };
  42
  43 static int eeh_result_priority(enum pci_ers_result result)
  44 {
  45         switch (result) {
  46         case PCI_ERS_RESULT_NONE:
  47                 return 1;
  48         case PCI_ERS_RESULT_NO_AER_DRIVER:
  49                 return 2;
  50         case PCI_ERS_RESULT_RECOVERED:
  51                 return 3;
  52         case PCI_ERS_RESULT_CAN_RECOVER:
  53                 return 4;
  54         case PCI_ERS_RESULT_DISCONNECT:
  55                 return 5;
  56         case PCI_ERS_RESULT_NEED_RESET:
  57                 return 6;
  58         default:
  59                 WARN_ONCE(1, "Unknown pci_ers_result value: %d\n", (int)result);
  60                 return 0;
  61         }
  62 };
  63
  64 static const char *pci_ers_result_name(enum pci_ers_result result)
  65 {
  66         switch (result) {
  67         case PCI_ERS_RESULT_NONE:
  68                 return "none";
  69         case PCI_ERS_RESULT_CAN_RECOVER:
  70                 return "can recover";
  71         case PCI_ERS_RESULT_NEED_RESET:
  72                 return "need reset";
  73         case PCI_ERS_RESULT_DISCONNECT:
  74                 return "disconnect";
  75         case PCI_ERS_RESULT_RECOVERED:
  76                 return "recovered";
  77         case PCI_ERS_RESULT_NO_AER_DRIVER:
  78                 return "no AER driver";
  79         default:
  80                 WARN_ONCE(1, "Unknown result type: %d\n", (int)result);
  81                 return "unknown";
  82         }
  83 };
  84
  85 static enum pci_ers_result pci_ers_merge_result(enum pci_ers_result old,
  86                                                 enum pci_ers_result new)
  87 {
  88         if (eeh_result_priority(new) > eeh_result_priority(old))
  89                 return new;
  90         return old;
  91 }
  92
  93 static bool eeh_dev_removed(struct eeh_dev *edev)
  94 {
  95         return !edev || (edev->mode & EEH_DEV_REMOVED);
  96 }
  97
  98 static bool eeh_edev_actionable(struct eeh_dev *edev)
  99 {
 100         if (!edev->pdev)
 101                 return false;
 102         if (edev->pdev->error_state == pci_channel_io_perm_failure)
 103                 return false;
 104         if (eeh_dev_removed(edev))
 105                 return false;
 106         if (eeh_pe_passed(edev->pe))
 107                 return false;
 108
 109         return true;
 110 }
 111
 112 /**
 113  * eeh_pcid_get - Get the PCI device driver
 114  * @pdev: PCI device
 115  *
 116  * The function is used to retrieve the PCI device driver for
 117  * the indicated PCI device. Besides, we will increase the reference
 118  * of the PCI device driver to prevent that being unloaded on
 119  * the fly. Otherwise, kernel crash would be seen.
 120  */
 121 static inline struct pci_driver *eeh_pcid_get(struct pci_dev *pdev)
 122 {
 123         if (!pdev || !pdev->driver)
 124                 return NULL;
 125
 126         if (!try_module_get(pdev->driver->driver.owner))
 127                 return NULL;
 128
 129         return pdev->driver;
 130 }
 131
 132 /**
 133  * eeh_pcid_put - Dereference on the PCI device driver
 134  * @pdev: PCI device
 135  *
 136  * The function is called to do dereference on the PCI device
 137  * driver of the indicated PCI device.
 138  */
 139 static inline void eeh_pcid_put(struct pci_dev *pdev)
 140 {
 141         if (!pdev || !pdev->driver)
 142                 return;
 143
 144         module_put(pdev->driver->driver.owner);
 145 }
 146
 147 /**
 148  * eeh_disable_irq - Disable interrupt for the recovering device
 149  * @dev: PCI device
 150  *
 151  * This routine must be called when reporting temporary or permanent
 152  * error to the particular PCI device to disable interrupt of that
 153  * device. If the device has enabled MSI or MSI-X interrupt, we needn't
 154  * do real work because EEH should freeze DMA transfers for those PCI
 155  * devices encountering EEH errors, which includes MSI or MSI-X.
 156  */
 157 static void eeh_disable_irq(struct eeh_dev *edev)
 158 {
 159         /* Don't disable MSI and MSI-X interrupts. They are
 160          * effectively disabled by the DMA Stopped state
 161          * when an EEH error occurs.
 162          */
 163         if (edev->pdev->msi_enabled || edev->pdev->msix_enabled)
 164                 return;
 165
 166         if (!irq_has_action(edev->pdev->irq))
 167                 return;
 168
 169         edev->mode |= EEH_DEV_IRQ_DISABLED;
 170         disable_irq_nosync(edev->pdev->irq);
 171 }
 172
 173 /**
 174  * eeh_enable_irq - Enable interrupt for the recovering device
 175  * @dev: PCI device
 176  *
 177  * This routine must be called to enable interrupt while failed
 178  * device could be resumed.
 179  */
 180 static void eeh_enable_irq(struct eeh_dev *edev)
 181 {
 182         if ((edev->mode) & EEH_DEV_IRQ_DISABLED) {
 183                 edev->mode &= ~EEH_DEV_IRQ_DISABLED;
 184                 /*
 185                  * FIXME !!!!!
 186                  *
 187                  * This is just ass backwards. This maze has
 188                  * unbalanced irq_enable/disable calls. So instead of
 189                  * finding the root cause it works around the warning
 190                  * in the irq_enable code by conditionally calling
 191                  * into it.
 192                  *
 193                  * That's just wrong.The warning in the core code is
 194                  * there to tell people to fix their asymmetries in
 195                  * their own code, not by abusing the core information
 196                  * to avoid it.
 197                  *
 198                  * I so wish that the assymetry would be the other way
 199                  * round and a few more irq_disable calls render that
 200                  * shit unusable forever.
 201                  *
 202                  *      tglx
 203                  */
 204                 if (irqd_irq_disabled(irq_get_irq_data(edev->pdev->irq)))
 205                         enable_irq(edev->pdev->irq);
 206         }
 207 }
 208
 209 static void eeh_dev_save_state(struct eeh_dev *edev, void *userdata)
 210 {
 211         struct pci_dev *pdev;
 212
 213         if (!edev)
 214                 return;
 215
 216         /*
 217          * We cannot access the config space on some adapters.
 218          * Otherwise, it will cause fenced PHB. We don't save
 219          * the content in their config space and will restore
 220          * from the initial config space saved when the EEH
 221          * device is created.
 222          */
 223         if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED))
 224                 return;
 225
 226         pdev = eeh_dev_to_pci_dev(edev);
 227         if (!pdev)
 228                 return;
 229
 230         pci_save_state(pdev);
 231 }
 232
 233 static void eeh_set_channel_state(struct eeh_pe *root, enum pci_channel_state s)
 234 {
 235         struct eeh_pe *pe;
 236         struct eeh_dev *edev, *tmp;
 237
 238         eeh_for_each_pe(root, pe)
 239                 eeh_pe_for_each_dev(pe, edev, tmp)
 240                         if (eeh_edev_actionable(edev))
 241                                 edev->pdev->error_state = s;
 242 }
 243
 244 static void eeh_set_irq_state(struct eeh_pe *root, bool enable)
 245 {
 246         struct eeh_pe *pe;
 247         struct eeh_dev *edev, *tmp;
 248
 249         eeh_for_each_pe(root, pe) {
 250                 eeh_pe_for_each_dev(pe, edev, tmp) {
 251                         if (!eeh_edev_actionable(edev))
 252                                 continue;
 253
 254                         if (!eeh_pcid_get(edev->pdev))
 255                                 continue;
 256
 257                         if (enable)
 258                                 eeh_enable_irq(edev);
 259                         else
 260                                 eeh_disable_irq(edev);
 261
 262                         eeh_pcid_put(edev->pdev);
 263                 }
 264         }
 265 }
 266
 267 typedef enum pci_ers_result (*eeh_report_fn)(struct eeh_dev *,
 268                                              struct pci_dev *,
 269                                              struct pci_driver *);
 270 static void eeh_pe_report_edev(struct eeh_dev *edev, eeh_report_fn fn,
 271                                enum pci_ers_result *result)
 272 {
 273         struct pci_dev *pdev;
 274         struct pci_driver *driver;
 275         enum pci_ers_result new_result;
 276
 277         pci_lock_rescan_remove();
 278         pdev = edev->pdev;
 279         if (pdev)
 280                 get_device(&pdev->dev);
 281         pci_unlock_rescan_remove();
 282         if (!pdev) {
 283                 eeh_edev_info(edev, "no device");
 284                 return;
 285         }
 286         device_lock(&pdev->dev);
 287         if (eeh_edev_actionable(edev)) {
 288                 driver = eeh_pcid_get(pdev);
 289
 290                 if (!driver)
 291                         eeh_edev_info(edev, "no driver");
 292                 else if (!driver->err_handler)
 293                         eeh_edev_info(edev, "driver not EEH aware");
 294                 else if (edev->mode & EEH_DEV_NO_HANDLER)
 295                         eeh_edev_info(edev, "driver bound too late");
 296                 else {
 297                         new_result = fn(edev, pdev, driver);
 298                         eeh_edev_info(edev, "%s driver reports: '%s'",
 299                                       driver->name,
 300                                       pci_ers_result_name(new_result));
 301                         if (result)
 302                                 *result = pci_ers_merge_result(*result,
 303                                                                new_result);
 304                 }
 305                 if (driver)
 306                         eeh_pcid_put(pdev);
 307         } else {
 308                 eeh_edev_info(edev, "not actionable (%d,%d,%d)", !!pdev,
 309                               !eeh_dev_removed(edev), !eeh_pe_passed(edev->pe));
 310         }
 311         device_unlock(&pdev->dev);
 312         if (edev->pdev != pdev)
 313                 eeh_edev_warn(edev, "Device changed during processing!\n");
 314         put_device(&pdev->dev);
 315 }
 316
 317 static void eeh_pe_report(const char *name, struct eeh_pe *root,
 318                           eeh_report_fn fn, enum pci_ers_result *result)
 319 {
 320         struct eeh_pe *pe;
 321         struct eeh_dev *edev, *tmp;
 322
 323         pr_info("EEH: Beginning: '%s'\n", name);
 324         eeh_for_each_pe(root, pe) eeh_pe_for_each_dev(pe, edev, tmp)
 325                 eeh_pe_report_edev(edev, fn, result);
 326         if (result)
 327                 pr_info("EEH: Finished:'%s' with aggregate recovery state:'%s'\n",
 328                         name, pci_ers_result_name(*result));
 329         else
 330                 pr_info("EEH: Finished:'%s'", name);
 331 }
 332
 333 /**
 334  * eeh_report_error - Report pci error to each device driver
 335  * @edev: eeh device
 336  * @driver: device's PCI driver
 337  *
 338  * Report an EEH error to each device driver.
 339  */
 340 static enum pci_ers_result eeh_report_error(struct eeh_dev *edev,
 341                                             struct pci_dev *pdev,
 342                                             struct pci_driver *driver)
 343 {
 344         enum pci_ers_result rc;
 345
 346         if (!driver->err_handler->error_detected)
 347                 return PCI_ERS_RESULT_NONE;
 348
 349         eeh_edev_info(edev, "Invoking %s->error_detected(IO frozen)",
 350                       driver->name);
 351         rc = driver->err_handler->error_detected(pdev, pci_channel_io_frozen);
 352
 353         edev->in_error = true;
 354         pci_uevent_ers(pdev, PCI_ERS_RESULT_NONE);
 355         return rc;
 356 }
 357
 358 /**
 359  * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
 360  * @edev: eeh device
 361  * @driver: device's PCI driver
 362  *
 363  * Tells each device driver that IO ports, MMIO and config space I/O
 364  * are now enabled.
 365  */
 366 static enum pci_ers_result eeh_report_mmio_enabled(struct eeh_dev *edev,
 367                                                    struct pci_dev *pdev,
 368                                                    struct pci_driver *driver)
 369 {
 370         if (!driver->err_handler->mmio_enabled)
 371                 return PCI_ERS_RESULT_NONE;
 372         eeh_edev_info(edev, "Invoking %s->mmio_enabled()", driver->name);
 373         return driver->err_handler->mmio_enabled(pdev);
 374 }
 375
 376 /**
 377  * eeh_report_reset - Tell device that slot has been reset
 378  * @edev: eeh device
 379  * @driver: device's PCI driver
 380  *
 381  * This routine must be called while EEH tries to reset particular
 382  * PCI device so that the associated PCI device driver could take
 383  * some actions, usually to save data the driver needs so that the
 384  * driver can work again while the device is recovered.
 385  */
 386 static enum pci_ers_result eeh_report_reset(struct eeh_dev *edev,
 387                                             struct pci_dev *pdev,
 388                                             struct pci_driver *driver)
 389 {
 390         if (!driver->err_handler->slot_reset || !edev->in_error)
 391                 return PCI_ERS_RESULT_NONE;
 392         eeh_edev_info(edev, "Invoking %s->slot_reset()", driver->name);
 393         return driver->err_handler->slot_reset(pdev);
 394 }
 395
 396 static void eeh_dev_restore_state(struct eeh_dev *edev, void *userdata)
 397 {
 398         struct pci_dev *pdev;
 399
 400         if (!edev)
 401                 return;
 402
 403         /*
 404          * The content in the config space isn't saved because
 405          * the blocked config space on some adapters. We have
 406          * to restore the initial saved config space when the
 407          * EEH device is created.
 408          */
 409         if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED)) {
 410                 if (list_is_last(&edev->entry, &edev->pe->edevs))
 411                         eeh_pe_restore_bars(edev->pe);
 412
 413                 return;
 414         }
 415
 416         pdev = eeh_dev_to_pci_dev(edev);
 417         if (!pdev)
 418                 return;
 419
 420         pci_restore_state(pdev);
 421 }
 422
 423 /**
 424  * eeh_report_resume - Tell device to resume normal operations
 425  * @edev: eeh device
 426  * @driver: device's PCI driver
 427  *
 428  * This routine must be called to notify the device driver that it
 429  * could resume so that the device driver can do some initialization
 430  * to make the recovered device work again.
 431  */
 432 static enum pci_ers_result eeh_report_resume(struct eeh_dev *edev,
 433                                              struct pci_dev *pdev,
 434                                              struct pci_driver *driver)
 435 {
 436         if (!driver->err_handler->resume || !edev->in_error)
 437                 return PCI_ERS_RESULT_NONE;
 438
 439         eeh_edev_info(edev, "Invoking %s->resume()", driver->name);
 440         driver->err_handler->resume(pdev);
 441
 442         pci_uevent_ers(edev->pdev, PCI_ERS_RESULT_RECOVERED);
 443 #ifdef CONFIG_PCI_IOV
 444         if (eeh_ops->notify_resume && eeh_dev_to_pdn(edev))
 445                 eeh_ops->notify_resume(eeh_dev_to_pdn(edev));
 446 #endif
 447         return PCI_ERS_RESULT_NONE;
 448 }
 449
 450 /**
 451  * eeh_report_failure - Tell device driver that device is dead.
 452  * @edev: eeh device
 453  * @driver: device's PCI driver
 454  *
 455  * This informs the device driver that the device is permanently
 456  * dead, and that no further recovery attempts will be made on it.
 457  */
 458 static enum pci_ers_result eeh_report_failure(struct eeh_dev *edev,
 459                                               struct pci_dev *pdev,
 460                                               struct pci_driver *driver)
 461 {
 462         enum pci_ers_result rc;
 463
 464         if (!driver->err_handler->error_detected)
 465                 return PCI_ERS_RESULT_NONE;
 466
 467         eeh_edev_info(edev, "Invoking %s->error_detected(permanent failure)",
 468                       driver->name);
 469         rc = driver->err_handler->error_detected(pdev,
 470                                                  pci_channel_io_perm_failure);
 471
 472         pci_uevent_ers(pdev, PCI_ERS_RESULT_DISCONNECT);
 473         return rc;
 474 }
 475
 476 static void *eeh_add_virt_device(struct eeh_dev *edev)
 477 {
 478         struct pci_driver *driver;
 479         struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 480
 481         if (!(edev->physfn)) {
 482                 eeh_edev_warn(edev, "Not for VF\n");
 483                 return NULL;
 484         }
 485
 486         driver = eeh_pcid_get(dev);
 487         if (driver) {
 488                 if (driver->err_handler) {
 489                         eeh_pcid_put(dev);
 490                         return NULL;
 491                 }
 492                 eeh_pcid_put(dev);
 493         }
 494
 495 #ifdef CONFIG_PCI_IOV
 496         pci_iov_add_virtfn(edev->physfn, eeh_dev_to_pdn(edev)->vf_index);
 497 #endif
 498         return NULL;
 499 }
 500
 501 static void eeh_rmv_device(struct eeh_dev *edev, void *userdata)
 502 {
 503         struct pci_driver *driver;
 504         struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 505         struct eeh_rmv_data *rmv_data = (struct eeh_rmv_data *)userdata;
 506
 507         /*
 508          * Actually, we should remove the PCI bridges as well.
 509          * However, that's lots of complexity to do that,
 510          * particularly some of devices under the bridge might
 511          * support EEH. So we just care about PCI devices for
 512          * simplicity here.
 513          */
 514         if (!eeh_edev_actionable(edev) ||
 515             (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE))
 516                 return;
 517
 518         if (rmv_data) {
 519                 driver = eeh_pcid_get(dev);
 520                 if (driver) {
 521                         if (driver->err_handler &&
 522                             driver->err_handler->error_detected &&
 523                             driver->err_handler->slot_reset) {
 524                                 eeh_pcid_put(dev);
 525                                 return;
 526                         }
 527                         eeh_pcid_put(dev);
 528                 }
 529         }
 530
 531         /* Remove it from PCI subsystem */
 532         pr_info("EEH: Removing %s without EEH sensitive driver\n",
 533                 pci_name(dev));
 534         edev->mode |= EEH_DEV_DISCONNECTED;
 535         if (rmv_data)
 536                 rmv_data->removed_dev_count++;
 537
 538         if (edev->physfn) {
 539 #ifdef CONFIG_PCI_IOV
 540                 struct pci_dn *pdn = eeh_dev_to_pdn(edev);
 541
 542                 pci_iov_remove_virtfn(edev->physfn, pdn->vf_index);
 543                 edev->pdev = NULL;
 544 #endif
 545                 if (rmv_data)
 546                         list_add(&edev->rmv_entry, &rmv_data->removed_vf_list);
 547         } else {
 548                 pci_lock_rescan_remove();
 549                 pci_stop_and_remove_bus_device(dev);
 550                 pci_unlock_rescan_remove();
 551         }
 552 }
 553
 554 static void *eeh_pe_detach_dev(struct eeh_pe *pe, void *userdata)
 555 {
 556         struct eeh_dev *edev, *tmp;
 557
 558         eeh_pe_for_each_dev(pe, edev, tmp) {
 559                 if (!(edev->mode & EEH_DEV_DISCONNECTED))
 560                         continue;
 561
 562                 edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
 563                 eeh_rmv_from_parent_pe(edev);
 564         }
 565
 566         return NULL;
 567 }
 568
 569 /*
 570  * Explicitly clear PE's frozen state for PowerNV where
 571  * we have frozen PE until BAR restore is completed. It's
 572  * harmless to clear it for pSeries. To be consistent with
 573  * PE reset (for 3 times), we try to clear the frozen state
 574  * for 3 times as well.
 575  */
 576 static int eeh_clear_pe_frozen_state(struct eeh_pe *root, bool include_passed)
 577 {
 578         struct eeh_pe *pe;
 579         int i;
 580
 581         eeh_for_each_pe(root, pe) {
 582                 if (include_passed || !eeh_pe_passed(pe)) {
 583                         for (i = 0; i < 3; i++)
 584                                 if (!eeh_unfreeze_pe(pe))
 585                                         break;
 586                         if (i >= 3)
 587                                 return -EIO;
 588                 }
 589         }
 590         eeh_pe_state_clear(root, EEH_PE_ISOLATED, include_passed);
 591         return 0;
 592 }
 593
 594 int eeh_pe_reset_and_recover(struct eeh_pe *pe)
 595 {
 596         int ret;
 597
 598         /* Bail if the PE is being recovered */
 599         if (pe->state & EEH_PE_RECOVERING)
 600                 return 0;
 601
 602         /* Put the PE into recovery mode */
 603         eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
 604
 605         /* Save states */
 606         eeh_pe_dev_traverse(pe, eeh_dev_save_state, NULL);
 607
 608         /* Issue reset */
 609         ret = eeh_pe_reset_full(pe, true);
 610         if (ret) {
 611                 eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 612                 return ret;
 613         }
 614
 615         /* Unfreeze the PE */
 616         ret = eeh_clear_pe_frozen_state(pe, true);
 617         if (ret) {
 618                 eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 619                 return ret;
 620         }
 621
 622         /* Restore device state */
 623         eeh_pe_dev_traverse(pe, eeh_dev_restore_state, NULL);
 624
 625         /* Clear recovery mode */
 626         eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 627
 628         return 0;
 629 }
 630
 631 /**
 632  * eeh_reset_device - Perform actual reset of a pci slot
 633  * @driver_eeh_aware: Does the device's driver provide EEH support?
 634  * @pe: EEH PE
 635  * @bus: PCI bus corresponding to the isolcated slot
 636  * @rmv_data: Optional, list to record removed devices
 637  *
 638  * This routine must be called to do reset on the indicated PE.
 639  * During the reset, udev might be invoked because those affected
 640  * PCI devices will be removed and then added.
 641  */
 642 static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus,
 643                             struct eeh_rmv_data *rmv_data,
 644                             bool driver_eeh_aware)
 645 {
 646         time64_t tstamp;
 647         int cnt, rc;
 648         struct eeh_dev *edev;
 649         struct eeh_pe *tmp_pe;
 650         bool any_passed = false;
 651
 652         eeh_for_each_pe(pe, tmp_pe)
 653                 any_passed |= eeh_pe_passed(tmp_pe);
 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 (any_passed || 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, false);
 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_dev_count) {
 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, entry);
 717                 eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
 718                 if (pe->type & EEH_PE_VF) {
 719                         eeh_add_virt_device(edev);
 720                 } else {
 721                         if (!driver_eeh_aware)
 722                                 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
 723                         pci_hp_add_devices(bus);
 724                 }
 725         }
 726         eeh_pe_state_clear(pe, EEH_PE_KEEP, true);
 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 /* Walks the PE tree after processing an event to remove any stale PEs.
 742  *
 743  * NB: This needs to be recursive to ensure the leaf PEs get removed
 744  * before their parents do. Although this is possible to do recursively
 745  * we don't since this is easier to read and we need to garantee
 746  * the leaf nodes will be handled first.
 747  */
 748 static void eeh_pe_cleanup(struct eeh_pe *pe)
 749 {
 750         struct eeh_pe *child_pe, *tmp;
 751
 752         list_for_each_entry_safe(child_pe, tmp, &pe->child_list, child)
 753                 eeh_pe_cleanup(child_pe);
 754
 755         if (pe->state & EEH_PE_KEEP)
 756                 return;
 757
 758         if (!(pe->state & EEH_PE_INVALID))
 759                 return;
 760
 761         if (list_empty(&pe->edevs) && list_empty(&pe->child_list)) {
 762                 list_del(&pe->child);
 763                 kfree(pe);
 764         }
 765 }
 766
 767 /**
 768  * eeh_check_slot_presence - Check if a device is still present in a slot
 769  * @pdev: pci_dev to check
 770  *
 771  * This function may return a false positive if we can't determine the slot's
 772  * presence state. This might happen for for PCIe slots if the PE containing
 773  * the upstream bridge is also frozen, or the bridge is part of the same PE
 774  * as the device.
 775  *
 776  * This shouldn't happen often, but you might see it if you hotplug a PCIe
 777  * switch.
 778  */
 779 static bool eeh_slot_presence_check(struct pci_dev *pdev)
 780 {
 781         const struct hotplug_slot_ops *ops;
 782         struct pci_slot *slot;
 783         u8 state;
 784         int rc;
 785
 786         if (!pdev)
 787                 return false;
 788
 789         if (pdev->error_state == pci_channel_io_perm_failure)
 790                 return false;
 791
 792         slot = pdev->slot;
 793         if (!slot || !slot->hotplug)
 794                 return true;
 795
 796         ops = slot->hotplug->ops;
 797         if (!ops || !ops->get_adapter_status)
 798                 return true;
 799
 800         /* set the attention indicator while we've got the slot ops */
 801         if (ops->set_attention_status)
 802                 ops->set_attention_status(slot->hotplug, 1);
 803
 804         rc = ops->get_adapter_status(slot->hotplug, &state);
 805         if (rc)
 806                 return true;
 807
 808         return !!state;
 809 }
 810
 811 static void eeh_clear_slot_attention(struct pci_dev *pdev)
 812 {
 813         const struct hotplug_slot_ops *ops;
 814         struct pci_slot *slot;
 815
 816         if (!pdev)
 817                 return;
 818
 819         if (pdev->error_state == pci_channel_io_perm_failure)
 820                 return;
 821
 822         slot = pdev->slot;
 823         if (!slot || !slot->hotplug)
 824                 return;
 825
 826         ops = slot->hotplug->ops;
 827         if (!ops || !ops->set_attention_status)
 828                 return;
 829
 830         ops->set_attention_status(slot->hotplug, 0);
 831 }
 832
 833 /**
 834  * eeh_handle_normal_event - Handle EEH events on a specific PE
 835  * @pe: EEH PE - which should not be used after we return, as it may
 836  * have been invalidated.
 837  *
 838  * Attempts to recover the given PE.  If recovery fails or the PE has failed
 839  * too many times, remove the PE.
 840  *
 841  * While PHB detects address or data parity errors on particular PCI
 842  * slot, the associated PE will be frozen. Besides, DMA's occurring
 843  * to wild addresses (which usually happen due to bugs in device
 844  * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
 845  * #PERR or other misc PCI-related errors also can trigger EEH errors.
 846  *
 847  * Recovery process consists of unplugging the device driver (which
 848  * generated hotplug events to userspace), then issuing a PCI #RST to
 849  * the device, then reconfiguring the PCI config space for all bridges
 850  * & devices under this slot, and then finally restarting the device
 851  * drivers (which cause a second set of hotplug events to go out to
 852  * userspace).
 853  */
 854 void eeh_handle_normal_event(struct eeh_pe *pe)
 855 {
 856         struct pci_bus *bus;
 857         struct eeh_dev *edev, *tmp;
 858         struct eeh_pe *tmp_pe;
 859         int rc = 0;
 860         enum pci_ers_result result = PCI_ERS_RESULT_NONE;
 861         struct eeh_rmv_data rmv_data =
 862                 {LIST_HEAD_INIT(rmv_data.removed_vf_list), 0};
 863         int devices = 0;
 864
 865         bus = eeh_pe_bus_get(pe);
 866         if (!bus) {
 867                 pr_err("%s: Cannot find PCI bus for PHB#%x-PE#%x\n",
 868                         __func__, pe->phb->global_number, pe->addr);
 869                 return;
 870         }
 871
 872         /*
 873          * When devices are hot-removed we might get an EEH due to
 874          * a driver attempting to touch the MMIO space of a removed
 875          * device. In this case we don't have a device to recover
 876          * so suppress the event if we can't find any present devices.
 877          *
 878          * The hotplug driver should take care of tearing down the
 879          * device itself.
 880          */
 881         eeh_for_each_pe(pe, tmp_pe)
 882                 eeh_pe_for_each_dev(tmp_pe, edev, tmp)
 883                         if (eeh_slot_presence_check(edev->pdev))
 884                                 devices++;
 885
 886         if (!devices) {
 887                 pr_debug("EEH: Frozen PHB#%x-PE#%x is empty!\n",
 888                         pe->phb->global_number, pe->addr);
 889                 goto out; /* nothing to recover */
 890         }
 891
 892         /* Log the event */
 893         if (pe->type & EEH_PE_PHB) {
 894                 pr_err("EEH: Recovering PHB#%x, location: %s\n",
 895                         pe->phb->global_number, eeh_pe_loc_get(pe));
 896         } else {
 897                 struct eeh_pe *phb_pe = eeh_phb_pe_get(pe->phb);
 898
 899                 pr_err("EEH: Recovering PHB#%x-PE#%x\n",
 900                        pe->phb->global_number, pe->addr);
 901                 pr_err("EEH: PE location: %s, PHB location: %s\n",
 902                        eeh_pe_loc_get(pe), eeh_pe_loc_get(phb_pe));
 903         }
 904
 905 #ifdef CONFIG_STACKTRACE
 906         /*
 907          * Print the saved stack trace now that we've verified there's
 908          * something to recover.
 909          */
 910         if (pe->trace_entries) {
 911                 void **ptrs = (void **) pe->stack_trace;
 912                 int i;
 913
 914                 pr_err("EEH: Frozen PHB#%x-PE#%x detected\n",
 915                        pe->phb->global_number, pe->addr);
 916
 917                 /* FIXME: Use the same format as dump_stack() */
 918                 pr_err("EEH: Call Trace:\n");
 919                 for (i = 0; i < pe->trace_entries; i++)
 920                         pr_err("EEH: [%pK] %pS\n", ptrs[i], ptrs[i]);
 921
 922                 pe->trace_entries = 0;
 923         }
 924 #endif /* CONFIG_STACKTRACE */
 925
 926         eeh_pe_update_time_stamp(pe);
 927         pe->freeze_count++;
 928         if (pe->freeze_count > eeh_max_freezes) {
 929                 pr_err("EEH: PHB#%x-PE#%x has failed %d times in the last hour and has been permanently disabled.\n",
 930                        pe->phb->global_number, pe->addr,
 931                        pe->freeze_count);
 932                 result = PCI_ERS_RESULT_DISCONNECT;
 933         }
 934
 935         eeh_for_each_pe(pe, tmp_pe)
 936                 eeh_pe_for_each_dev(tmp_pe, edev, tmp)
 937                         edev->mode &= ~EEH_DEV_NO_HANDLER;
 938
 939         /* Walk the various device drivers attached to this slot through
 940          * a reset sequence, giving each an opportunity to do what it needs
 941          * to accomplish the reset.  Each child gets a report of the
 942          * status ... if any child can't handle the reset, then the entire
 943          * slot is dlpar removed and added.
 944          *
 945          * When the PHB is fenced, we have to issue a reset to recover from
 946          * the error. Override the result if necessary to have partially
 947          * hotplug for this case.
 948          */
 949         if (result != PCI_ERS_RESULT_DISCONNECT) {
 950                 pr_warn("EEH: This PCI device has failed %d times in the last hour and will be permanently disabled after %d failures.\n",
 951                         pe->freeze_count, eeh_max_freezes);
 952                 pr_info("EEH: Notify device drivers to shutdown\n");
 953                 eeh_set_channel_state(pe, pci_channel_io_frozen);
 954                 eeh_set_irq_state(pe, false);
 955                 eeh_pe_report("error_detected(IO frozen)", pe,
 956                               eeh_report_error, &result);
 957                 if ((pe->type & EEH_PE_PHB) &&
 958                     result != PCI_ERS_RESULT_NONE &&
 959                     result != PCI_ERS_RESULT_NEED_RESET)
 960                         result = PCI_ERS_RESULT_NEED_RESET;
 961         }
 962
 963         /* Get the current PCI slot state. This can take a long time,
 964          * sometimes over 300 seconds for certain systems.
 965          */
 966         if (result != PCI_ERS_RESULT_DISCONNECT) {
 967                 rc = eeh_wait_state(pe, MAX_WAIT_FOR_RECOVERY*1000);
 968                 if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
 969                         pr_warn("EEH: Permanent failure\n");
 970                         result = PCI_ERS_RESULT_DISCONNECT;
 971                 }
 972         }
 973
 974         /* Since rtas may enable MMIO when posting the error log,
 975          * don't post the error log until after all dev drivers
 976          * have been informed.
 977          */
 978         if (result != PCI_ERS_RESULT_DISCONNECT) {
 979                 pr_info("EEH: Collect temporary log\n");
 980                 eeh_slot_error_detail(pe, EEH_LOG_TEMP);
 981         }
 982
 983         /* If all device drivers were EEH-unaware, then shut
 984          * down all of the device drivers, and hope they
 985          * go down willingly, without panicing the system.
 986          */
 987         if (result == PCI_ERS_RESULT_NONE) {
 988                 pr_info("EEH: Reset with hotplug activity\n");
 989                 rc = eeh_reset_device(pe, bus, NULL, false);
 990                 if (rc) {
 991                         pr_warn("%s: Unable to reset, err=%d\n",
 992                                 __func__, rc);
 993                         result = PCI_ERS_RESULT_DISCONNECT;
 994                 }
 995         }
 996
 997         /* If all devices reported they can proceed, then re-enable MMIO */
 998         if (result == PCI_ERS_RESULT_CAN_RECOVER) {
 999                 pr_info("EEH: Enable I/O for affected devices\n");
1000                 rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
1001
1002                 if (rc < 0) {
1003                         result = PCI_ERS_RESULT_DISCONNECT;
1004                 } else if (rc) {
1005                         result = PCI_ERS_RESULT_NEED_RESET;
1006                 } else {
1007                         pr_info("EEH: Notify device drivers to resume I/O\n");
1008                         eeh_pe_report("mmio_enabled", pe,
1009                                       eeh_report_mmio_enabled, &result);
1010                 }
1011         }
1012
1013         /* If all devices reported they can proceed, then re-enable DMA */
1014         if (result == PCI_ERS_RESULT_CAN_RECOVER) {
1015                 pr_info("EEH: Enabled DMA for affected devices\n");
1016                 rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
1017
1018                 if (rc < 0) {
1019                         result = PCI_ERS_RESULT_DISCONNECT;
1020                 } else if (rc) {
1021                         result = PCI_ERS_RESULT_NEED_RESET;
1022                 } else {
1023                         /*
1024                          * We didn't do PE reset for the case. The PE
1025                          * is still in frozen state. Clear it before
1026                          * resuming the PE.
1027                          */
1028                         eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
1029                         result = PCI_ERS_RESULT_RECOVERED;
1030                 }
1031         }
1032
1033         /* If any device called out for a reset, then reset the slot */
1034         if (result == PCI_ERS_RESULT_NEED_RESET) {
1035                 pr_info("EEH: Reset without hotplug activity\n");
1036                 rc = eeh_reset_device(pe, bus, &rmv_data, true);
1037                 if (rc) {
1038                         pr_warn("%s: Cannot reset, err=%d\n",
1039                                 __func__, rc);
1040                         result = PCI_ERS_RESULT_DISCONNECT;
1041                 } else {
1042                         result = PCI_ERS_RESULT_NONE;
1043                         eeh_set_channel_state(pe, pci_channel_io_normal);
1044                         eeh_set_irq_state(pe, true);
1045                         eeh_pe_report("slot_reset", pe, eeh_report_reset,
1046                                       &result);
1047                 }
1048         }
1049
1050         if ((result == PCI_ERS_RESULT_RECOVERED) ||
1051             (result == PCI_ERS_RESULT_NONE)) {
1052                 /*
1053                  * For those hot removed VFs, we should add back them after PF
1054                  * get recovered properly.
1055                  */
1056                 list_for_each_entry_safe(edev, tmp, &rmv_data.removed_vf_list,
1057                                          rmv_entry) {
1058                         eeh_add_virt_device(edev);
1059                         list_del(&edev->rmv_entry);
1060                 }
1061
1062                 /* Tell all device drivers that they can resume operations */
1063                 pr_info("EEH: Notify device driver to resume\n");
1064                 eeh_set_channel_state(pe, pci_channel_io_normal);
1065                 eeh_set_irq_state(pe, true);
1066                 eeh_pe_report("resume", pe, eeh_report_resume, NULL);
1067                 eeh_for_each_pe(pe, tmp_pe) {
1068                         eeh_pe_for_each_dev(tmp_pe, edev, tmp) {
1069                                 edev->mode &= ~EEH_DEV_NO_HANDLER;
1070                                 edev->in_error = false;
1071                         }
1072                 }
1073
1074                 pr_info("EEH: Recovery successful.\n");
1075                 goto out;
1076         }
1077
1078         /*
1079          * About 90% of all real-life EEH failures in the field
1080          * are due to poorly seated PCI cards. Only 10% or so are
1081          * due to actual, failed cards.
1082          */
1083         pr_err("EEH: Unable to recover from failure from PHB#%x-PE#%x.\n"
1084                 "Please try reseating or replacing it\n",
1085                 pe->phb->global_number, pe->addr);
1086
1087         eeh_slot_error_detail(pe, EEH_LOG_PERM);
1088
1089         /* Notify all devices that they're about to go down. */
1090         eeh_set_irq_state(pe, false);
1091         eeh_pe_report("error_detected(permanent failure)", pe,
1092                       eeh_report_failure, NULL);
1093         eeh_set_channel_state(pe, pci_channel_io_perm_failure);
1094
1095         /* Mark the PE to be removed permanently */
1096         eeh_pe_state_mark(pe, EEH_PE_REMOVED);
1097
1098         /*
1099          * Shut down the device drivers for good. We mark
1100          * all removed devices correctly to avoid access
1101          * the their PCI config any more.
1102          */
1103         if (pe->type & EEH_PE_VF) {
1104                 eeh_pe_dev_traverse(pe, eeh_rmv_device, NULL);
1105                 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
1106         } else {
1107                 eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
1108                 eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
1109
1110                 pci_lock_rescan_remove();
1111                 pci_hp_remove_devices(bus);
1112                 pci_unlock_rescan_remove();
1113                 /* The passed PE should no longer be used */
1114                 return;
1115         }
1116
1117 out:
1118         /*
1119          * Clean up any PEs without devices. While marked as EEH_PE_RECOVERYING
1120          * we don't want to modify the PE tree structure so we do it here.
1121          */
1122         eeh_pe_cleanup(pe);
1123
1124         /* clear the slot attention LED for all recovered devices */
1125         eeh_for_each_pe(pe, tmp_pe)
1126                 eeh_pe_for_each_dev(tmp_pe, edev, tmp)
1127                         eeh_clear_slot_attention(edev->pdev);
1128
1129         eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
1130 }
1131
1132 /**
1133  * eeh_handle_special_event - Handle EEH events without a specific failing PE
1134  *
1135  * Called when an EEH event is detected but can't be narrowed down to a
1136  * specific PE.  Iterates through possible failures and handles them as
1137  * necessary.
1138  */
1139 void eeh_handle_special_event(void)
1140 {
1141         struct eeh_pe *pe, *phb_pe, *tmp_pe;
1142         struct eeh_dev *edev, *tmp_edev;
1143         struct pci_bus *bus;
1144         struct pci_controller *hose;
1145         unsigned long flags;
1146         int rc;
1147
1148
1149         do {
1150                 rc = eeh_ops->next_error(&pe);
1151
1152                 switch (rc) {
1153                 case EEH_NEXT_ERR_DEAD_IOC:
1154                         /* Mark all PHBs in dead state */
1155                         eeh_serialize_lock(&flags);
1156
1157                         /* Purge all events */
1158                         eeh_remove_event(NULL, true);
1159
1160                         list_for_each_entry(hose, &hose_list, list_node) {
1161                                 phb_pe = eeh_phb_pe_get(hose);
1162                                 if (!phb_pe) continue;
1163
1164                                 eeh_pe_mark_isolated(phb_pe);
1165                         }
1166
1167                         eeh_serialize_unlock(flags);
1168
1169                         break;
1170                 case EEH_NEXT_ERR_FROZEN_PE:
1171                 case EEH_NEXT_ERR_FENCED_PHB:
1172                 case EEH_NEXT_ERR_DEAD_PHB:
1173                         /* Mark the PE in fenced state */
1174                         eeh_serialize_lock(&flags);
1175
1176                         /* Purge all events of the PHB */
1177                         eeh_remove_event(pe, true);
1178
1179                         if (rc != EEH_NEXT_ERR_DEAD_PHB)
1180                                 eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
1181                         eeh_pe_mark_isolated(pe);
1182
1183                         eeh_serialize_unlock(flags);
1184
1185                         break;
1186                 case EEH_NEXT_ERR_NONE:
1187                         return;
1188                 default:
1189                         pr_warn("%s: Invalid value %d from next_error()\n",
1190                                 __func__, rc);
1191                         return;
1192                 }
1193
1194                 /*
1195                  * For fenced PHB and frozen PE, it's handled as normal
1196                  * event. We have to remove the affected PHBs for dead
1197                  * PHB and IOC
1198                  */
1199                 if (rc == EEH_NEXT_ERR_FROZEN_PE ||
1200                     rc == EEH_NEXT_ERR_FENCED_PHB) {
1201                         eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
1202                         eeh_handle_normal_event(pe);
1203                 } else {
1204                         eeh_for_each_pe(pe, tmp_pe)
1205                                 eeh_pe_for_each_dev(tmp_pe, edev, tmp_edev)
1206                                         edev->mode &= ~EEH_DEV_NO_HANDLER;
1207
1208                         /* Notify all devices to be down */
1209                         eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
1210                         eeh_pe_report(
1211                                 "error_detected(permanent failure)", pe,
1212                                 eeh_report_failure, NULL);
1213                         eeh_set_channel_state(pe, pci_channel_io_perm_failure);
1214
1215                         pci_lock_rescan_remove();
1216                         list_for_each_entry(hose, &hose_list, list_node) {
1217                                 phb_pe = eeh_phb_pe_get(hose);
1218                                 if (!phb_pe ||
1219                                     !(phb_pe->state & EEH_PE_ISOLATED) ||
1220                                     (phb_pe->state & EEH_PE_RECOVERING))
1221                                         continue;
1222
1223                                 bus = eeh_pe_bus_get(phb_pe);
1224                                 if (!bus) {
1225                                         pr_err("%s: Cannot find PCI bus for "
1226                                                "PHB#%x-PE#%x\n",
1227                                                __func__,
1228                                                pe->phb->global_number,
1229                                                pe->addr);
1230                                         break;
1231                                 }
1232                                 pci_hp_remove_devices(bus);
1233                         }
1234                         pci_unlock_rescan_remove();
1235                 }
1236
1237                 /*
1238                  * If we have detected dead IOC, we needn't proceed
1239                  * any more since all PHBs would have been removed
1240                  */
1241                 if (rc == EEH_NEXT_ERR_DEAD_IOC)
1242                         break;
1243         } while (rc != EEH_NEXT_ERR_NONE);
1244 }