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