Linux Audio

Check our new training course

Loading...
v4.17
 
   1/*
   2 * PCI Error Recovery Driver for RPA-compliant PPC64 platform.
   3 * Copyright IBM Corp. 2004 2005
   4 * Copyright Linas Vepstas <linas@linas.org> 2004, 2005
   5 *
   6 * All rights reserved.
   7 *
   8 * This program is free software; you can redistribute it and/or modify
   9 * it under the terms of the GNU General Public License as published by
  10 * the Free Software Foundation; either version 2 of the License, or (at
  11 * your option) any later version.
  12 *
  13 * This program is distributed in the hope that it will be useful, but
  14 * WITHOUT ANY WARRANTY; without even the implied warranty of
  15 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  16 * NON INFRINGEMENT.  See the GNU General Public License for more
  17 * details.
  18 *
  19 * You should have received a copy of the GNU General Public License
  20 * along with this program; if not, write to the Free Software
  21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22 *
  23 * Send comments and feedback to Linas Vepstas <linas@austin.ibm.com>
  24 */
  25#include <linux/delay.h>
  26#include <linux/interrupt.h>
  27#include <linux/irq.h>
  28#include <linux/module.h>
  29#include <linux/pci.h>
 
  30#include <asm/eeh.h>
  31#include <asm/eeh_event.h>
  32#include <asm/ppc-pci.h>
  33#include <asm/pci-bridge.h>
  34#include <asm/prom.h>
  35#include <asm/rtas.h>
  36
  37struct eeh_rmv_data {
  38	struct list_head edev_list;
  39	int removed;
  40};
  41
  42/**
  43 * eeh_pcid_name - Retrieve name of PCI device driver
  44 * @pdev: PCI device
  45 *
  46 * This routine is used to retrieve the name of PCI device driver
  47 * if that's valid.
  48 */
  49static inline const char *eeh_pcid_name(struct pci_dev *pdev)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  50{
  51	if (pdev && pdev->dev.driver)
  52		return pdev->dev.driver->name;
  53	return "";
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  54}
  55
  56/**
  57 * eeh_pcid_get - Get the PCI device driver
  58 * @pdev: PCI device
  59 *
  60 * The function is used to retrieve the PCI device driver for
  61 * the indicated PCI device. Besides, we will increase the reference
  62 * of the PCI device driver to prevent that being unloaded on
  63 * the fly. Otherwise, kernel crash would be seen.
  64 */
  65static inline struct pci_driver *eeh_pcid_get(struct pci_dev *pdev)
  66{
  67	if (!pdev || !pdev->driver)
  68		return NULL;
  69
  70	if (!try_module_get(pdev->driver->driver.owner))
  71		return NULL;
  72
  73	return pdev->driver;
  74}
  75
  76/**
  77 * eeh_pcid_put - Dereference on the PCI device driver
  78 * @pdev: PCI device
  79 *
  80 * The function is called to do dereference on the PCI device
  81 * driver of the indicated PCI device.
  82 */
  83static inline void eeh_pcid_put(struct pci_dev *pdev)
  84{
  85	if (!pdev || !pdev->driver)
  86		return;
  87
  88	module_put(pdev->driver->driver.owner);
  89}
  90
  91/**
  92 * eeh_disable_irq - Disable interrupt for the recovering device
  93 * @dev: PCI device
  94 *
  95 * This routine must be called when reporting temporary or permanent
  96 * error to the particular PCI device to disable interrupt of that
  97 * device. If the device has enabled MSI or MSI-X interrupt, we needn't
  98 * do real work because EEH should freeze DMA transfers for those PCI
  99 * devices encountering EEH errors, which includes MSI or MSI-X.
 100 */
 101static void eeh_disable_irq(struct pci_dev *dev)
 102{
 103	struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
 104
 105	/* Don't disable MSI and MSI-X interrupts. They are
 106	 * effectively disabled by the DMA Stopped state
 107	 * when an EEH error occurs.
 108	 */
 109	if (dev->msi_enabled || dev->msix_enabled)
 110		return;
 111
 112	if (!irq_has_action(dev->irq))
 113		return;
 114
 115	edev->mode |= EEH_DEV_IRQ_DISABLED;
 116	disable_irq_nosync(dev->irq);
 117}
 118
 119/**
 120 * eeh_enable_irq - Enable interrupt for the recovering device
 121 * @dev: PCI device
 122 *
 123 * This routine must be called to enable interrupt while failed
 124 * device could be resumed.
 125 */
 126static void eeh_enable_irq(struct pci_dev *dev)
 127{
 128	struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
 129
 130	if ((edev->mode) & EEH_DEV_IRQ_DISABLED) {
 131		edev->mode &= ~EEH_DEV_IRQ_DISABLED;
 132		/*
 133		 * FIXME !!!!!
 134		 *
 135		 * This is just ass backwards. This maze has
 136		 * unbalanced irq_enable/disable calls. So instead of
 137		 * finding the root cause it works around the warning
 138		 * in the irq_enable code by conditionally calling
 139		 * into it.
 140		 *
 141		 * That's just wrong.The warning in the core code is
 142		 * there to tell people to fix their asymmetries in
 143		 * their own code, not by abusing the core information
 144		 * to avoid it.
 145		 *
 146		 * I so wish that the assymetry would be the other way
 147		 * round and a few more irq_disable calls render that
 148		 * shit unusable forever.
 149		 *
 150		 *	tglx
 151		 */
 152		if (irqd_irq_disabled(irq_get_irq_data(dev->irq)))
 153			enable_irq(dev->irq);
 154	}
 155}
 156
 157static bool eeh_dev_removed(struct eeh_dev *edev)
 158{
 159	/* EEH device removed ? */
 160	if (!edev || (edev->mode & EEH_DEV_REMOVED))
 161		return true;
 162
 163	return false;
 164}
 165
 166static void *eeh_dev_save_state(void *data, void *userdata)
 167{
 168	struct eeh_dev *edev = data;
 169	struct pci_dev *pdev;
 170
 171	if (!edev)
 172		return NULL;
 173
 174	/*
 175	 * We cannot access the config space on some adapters.
 176	 * Otherwise, it will cause fenced PHB. We don't save
 177	 * the content in their config space and will restore
 178	 * from the initial config space saved when the EEH
 179	 * device is created.
 180	 */
 181	if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED))
 182		return NULL;
 183
 184	pdev = eeh_dev_to_pci_dev(edev);
 185	if (!pdev)
 186		return NULL;
 187
 188	pci_save_state(pdev);
 189	return NULL;
 190}
 191
 192/**
 193 * eeh_report_error - Report pci error to each device driver
 194 * @data: eeh device
 195 * @userdata: return value
 196 *
 197 * Report an EEH error to each device driver, collect up and
 198 * merge the device driver responses. Cumulative response
 199 * passed back in "userdata".
 200 */
 201static void *eeh_report_error(void *data, void *userdata)
 202{
 203	struct eeh_dev *edev = (struct eeh_dev *)data;
 204	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 205	enum pci_ers_result rc, *res = userdata;
 206	struct pci_driver *driver;
 207
 208	if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
 209		return NULL;
 210
 211	device_lock(&dev->dev);
 212	dev->error_state = pci_channel_io_frozen;
 213
 214	driver = eeh_pcid_get(dev);
 215	if (!driver) goto out_no_dev;
 216
 217	eeh_disable_irq(dev);
 
 
 
 
 218
 219	if (!driver->err_handler ||
 220	    !driver->err_handler->error_detected)
 221		goto out;
 
 222
 223	rc = driver->err_handler->error_detected(dev, pci_channel_io_frozen);
 
 
 
 224
 225	/* A driver that needs a reset trumps all others */
 226	if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
 227	if (*res == PCI_ERS_RESULT_NONE) *res = rc;
 228
 229	edev->in_error = true;
 230	pci_uevent_ers(dev, PCI_ERS_RESULT_NONE);
 
 
 231
 232out:
 233	eeh_pcid_put(dev);
 234out_no_dev:
 235	device_unlock(&dev->dev);
 236	return NULL;
 237}
 238
 239/**
 240 * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
 241 * @data: eeh device
 242 * @userdata: return value
 243 *
 244 * Tells each device driver that IO ports, MMIO and config space I/O
 245 * are now enabled. Collects up and merges the device driver responses.
 246 * Cumulative response passed back in "userdata".
 247 */
 248static void *eeh_report_mmio_enabled(void *data, void *userdata)
 249{
 250	struct eeh_dev *edev = (struct eeh_dev *)data;
 251	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 252	enum pci_ers_result rc, *res = userdata;
 253	struct pci_driver *driver;
 
 254
 255	if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
 256		return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 257
 258	device_lock(&dev->dev);
 259	driver = eeh_pcid_get(dev);
 260	if (!driver) goto out_no_dev;
 
 
 261
 262	if (!driver->err_handler ||
 263	    !driver->err_handler->mmio_enabled ||
 264	    (edev->mode & EEH_DEV_NO_HANDLER))
 265		goto out;
 
 
 
 
 
 266
 267	rc = driver->err_handler->mmio_enabled(dev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 268
 269	/* A driver that needs a reset trumps all others */
 270	if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
 271	if (*res == PCI_ERS_RESULT_NONE) *res = rc;
 
 272
 273out:
 274	eeh_pcid_put(dev);
 275out_no_dev:
 276	device_unlock(&dev->dev);
 277	return NULL;
 
 
 
 
 
 
 
 
 
 
 
 278}
 279
 280/**
 281 * eeh_report_reset - Tell device that slot has been reset
 282 * @data: eeh device
 283 * @userdata: return value
 284 *
 285 * This routine must be called while EEH tries to reset particular
 286 * PCI device so that the associated PCI device driver could take
 287 * some actions, usually to save data the driver needs so that the
 288 * driver can work again while the device is recovered.
 289 */
 290static void *eeh_report_reset(void *data, void *userdata)
 
 
 291{
 292	struct eeh_dev *edev = (struct eeh_dev *)data;
 293	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 294	enum pci_ers_result rc, *res = userdata;
 295	struct pci_driver *driver;
 296
 297	if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
 298		return NULL;
 299
 300	device_lock(&dev->dev);
 301	dev->error_state = pci_channel_io_normal;
 302
 303	driver = eeh_pcid_get(dev);
 304	if (!driver) goto out_no_dev;
 305
 306	eeh_enable_irq(dev);
 307
 308	if (!driver->err_handler ||
 309	    !driver->err_handler->slot_reset ||
 310	    (edev->mode & EEH_DEV_NO_HANDLER) ||
 311	    (!edev->in_error))
 312		goto out;
 313
 314	rc = driver->err_handler->slot_reset(dev);
 315	if ((*res == PCI_ERS_RESULT_NONE) ||
 316	    (*res == PCI_ERS_RESULT_RECOVERED)) *res = rc;
 317	if (*res == PCI_ERS_RESULT_DISCONNECT &&
 318	     rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
 319
 320out:
 321	eeh_pcid_put(dev);
 322out_no_dev:
 323	device_unlock(&dev->dev);
 324	return NULL;
 325}
 326
 327static void *eeh_dev_restore_state(void *data, void *userdata)
 328{
 329	struct eeh_dev *edev = data;
 330	struct pci_dev *pdev;
 331
 332	if (!edev)
 333		return NULL;
 334
 335	/*
 336	 * The content in the config space isn't saved because
 337	 * the blocked config space on some adapters. We have
 338	 * to restore the initial saved config space when the
 339	 * EEH device is created.
 340	 */
 341	if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED)) {
 342		if (list_is_last(&edev->list, &edev->pe->edevs))
 343			eeh_pe_restore_bars(edev->pe);
 344
 345		return NULL;
 346	}
 347
 348	pdev = eeh_dev_to_pci_dev(edev);
 349	if (!pdev)
 350		return NULL;
 351
 352	pci_restore_state(pdev);
 353	return NULL;
 354}
 355
 356/**
 357 * eeh_report_resume - Tell device to resume normal operations
 358 * @data: eeh device
 359 * @userdata: return value
 360 *
 361 * This routine must be called to notify the device driver that it
 362 * could resume so that the device driver can do some initialization
 363 * to make the recovered device work again.
 364 */
 365static void *eeh_report_resume(void *data, void *userdata)
 
 
 366{
 367	struct eeh_dev *edev = (struct eeh_dev *)data;
 368	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 369	bool was_in_error;
 370	struct pci_driver *driver;
 371
 372	if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
 373		return NULL;
 374
 375	device_lock(&dev->dev);
 376	dev->error_state = pci_channel_io_normal;
 377
 378	driver = eeh_pcid_get(dev);
 379	if (!driver) goto out_no_dev;
 380
 381	was_in_error = edev->in_error;
 382	edev->in_error = false;
 383	eeh_enable_irq(dev);
 384
 385	if (!driver->err_handler ||
 386	    !driver->err_handler->resume ||
 387	    (edev->mode & EEH_DEV_NO_HANDLER) || !was_in_error) {
 388		edev->mode &= ~EEH_DEV_NO_HANDLER;
 389		goto out;
 390	}
 391
 392	driver->err_handler->resume(dev);
 
 393
 394	pci_uevent_ers(dev, PCI_ERS_RESULT_RECOVERED);
 395out:
 396	eeh_pcid_put(dev);
 397#ifdef CONFIG_PCI_IOV
 398	if (eeh_ops->notify_resume && eeh_dev_to_pdn(edev))
 399		eeh_ops->notify_resume(eeh_dev_to_pdn(edev));
 400#endif
 401out_no_dev:
 402	device_unlock(&dev->dev);
 403	return NULL;
 404}
 405
 406/**
 407 * eeh_report_failure - Tell device driver that device is dead.
 408 * @data: eeh device
 409 * @userdata: return value
 410 *
 411 * This informs the device driver that the device is permanently
 412 * dead, and that no further recovery attempts will be made on it.
 413 */
 414static void *eeh_report_failure(void *data, void *userdata)
 
 
 415{
 416	struct eeh_dev *edev = (struct eeh_dev *)data;
 417	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 418	struct pci_driver *driver;
 419
 420	if (!dev || eeh_dev_removed(edev) || eeh_pe_passed(edev->pe))
 421		return NULL;
 422
 423	device_lock(&dev->dev);
 424	dev->error_state = pci_channel_io_perm_failure;
 425
 426	driver = eeh_pcid_get(dev);
 427	if (!driver) goto out_no_dev;
 428
 429	eeh_disable_irq(dev);
 430
 431	if (!driver->err_handler ||
 432	    !driver->err_handler->error_detected)
 433		goto out;
 434
 435	driver->err_handler->error_detected(dev, pci_channel_io_perm_failure);
 
 
 
 436
 437	pci_uevent_ers(dev, PCI_ERS_RESULT_DISCONNECT);
 438out:
 439	eeh_pcid_put(dev);
 440out_no_dev:
 441	device_unlock(&dev->dev);
 442	return NULL;
 443}
 444
 445static void *eeh_add_virt_device(void *data, void *userdata)
 446{
 447	struct pci_driver *driver;
 448	struct eeh_dev *edev = (struct eeh_dev *)data;
 449	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 450	struct pci_dn *pdn = eeh_dev_to_pdn(edev);
 451
 452	if (!(edev->physfn)) {
 453		pr_warn("%s: EEH dev %04x:%02x:%02x.%01x not for VF\n",
 454			__func__, pdn->phb->global_number, pdn->busno,
 455			PCI_SLOT(pdn->devfn), PCI_FUNC(pdn->devfn));
 456		return NULL;
 457	}
 458
 459	driver = eeh_pcid_get(dev);
 460	if (driver) {
 461		eeh_pcid_put(dev);
 462		if (driver->err_handler)
 463			return NULL;
 
 
 464	}
 465
 466#ifdef CONFIG_PCI_IOV
 467	pci_iov_add_virtfn(edev->physfn, pdn->vf_index);
 468#endif
 469	return NULL;
 470}
 471
 472static void *eeh_rmv_device(void *data, void *userdata)
 473{
 474	struct pci_driver *driver;
 475	struct eeh_dev *edev = (struct eeh_dev *)data;
 476	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 477	struct eeh_rmv_data *rmv_data = (struct eeh_rmv_data *)userdata;
 478	int *removed = rmv_data ? &rmv_data->removed : NULL;
 479
 480	/*
 481	 * Actually, we should remove the PCI bridges as well.
 482	 * However, that's lots of complexity to do that,
 483	 * particularly some of devices under the bridge might
 484	 * support EEH. So we just care about PCI devices for
 485	 * simplicity here.
 486	 */
 487	if (!dev || (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE))
 488		return NULL;
 489
 490	/*
 491	 * We rely on count-based pcibios_release_device() to
 492	 * detach permanently offlined PEs. Unfortunately, that's
 493	 * not reliable enough. We might have the permanently
 494	 * offlined PEs attached, but we needn't take care of
 495	 * them and their child devices.
 496	 */
 497	if (eeh_dev_removed(edev))
 498		return NULL;
 499
 500	driver = eeh_pcid_get(dev);
 501	if (driver) {
 502		eeh_pcid_put(dev);
 503		if (removed &&
 504		    eeh_pe_passed(edev->pe))
 505			return NULL;
 506		if (removed &&
 507		    driver->err_handler &&
 508		    driver->err_handler->error_detected &&
 509		    driver->err_handler->slot_reset)
 510			return NULL;
 511	}
 512
 513	/* Remove it from PCI subsystem */
 514	pr_debug("EEH: Removing %s without EEH sensitive driver\n",
 515		 pci_name(dev));
 516	edev->bus = dev->bus;
 517	edev->mode |= EEH_DEV_DISCONNECTED;
 518	if (removed)
 519		(*removed)++;
 520
 521	if (edev->physfn) {
 522#ifdef CONFIG_PCI_IOV
 523		struct pci_dn *pdn = eeh_dev_to_pdn(edev);
 524
 525		pci_iov_remove_virtfn(edev->physfn, pdn->vf_index);
 526		edev->pdev = NULL;
 527
 528		/*
 529		 * We have to set the VF PE number to invalid one, which is
 530		 * required to plug the VF successfully.
 531		 */
 532		pdn->pe_number = IODA_INVALID_PE;
 533#endif
 534		if (rmv_data)
 535			list_add(&edev->rmv_list, &rmv_data->edev_list);
 536	} else {
 537		pci_lock_rescan_remove();
 538		pci_stop_and_remove_bus_device(dev);
 539		pci_unlock_rescan_remove();
 540	}
 541
 542	return NULL;
 543}
 544
 545static void *eeh_pe_detach_dev(void *data, void *userdata)
 546{
 547	struct eeh_pe *pe = (struct eeh_pe *)data;
 548	struct eeh_dev *edev, *tmp;
 549
 550	eeh_pe_for_each_dev(pe, edev, tmp) {
 551		if (!(edev->mode & EEH_DEV_DISCONNECTED))
 552			continue;
 553
 554		edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
 555		eeh_rmv_from_parent_pe(edev);
 556	}
 557
 558	return NULL;
 559}
 560
 561/*
 562 * Explicitly clear PE's frozen state for PowerNV where
 563 * we have frozen PE until BAR restore is completed. It's
 564 * harmless to clear it for pSeries. To be consistent with
 565 * PE reset (for 3 times), we try to clear the frozen state
 566 * for 3 times as well.
 567 */
 568static void *__eeh_clear_pe_frozen_state(void *data, void *flag)
 569{
 570	struct eeh_pe *pe = (struct eeh_pe *)data;
 571	bool clear_sw_state = *(bool *)flag;
 572	int i, rc = 1;
 573
 574	for (i = 0; rc && i < 3; i++)
 575		rc = eeh_unfreeze_pe(pe, clear_sw_state);
 576
 577	/* Stop immediately on any errors */
 578	if (rc) {
 579		pr_warn("%s: Failure %d unfreezing PHB#%x-PE#%x\n",
 580			__func__, rc, pe->phb->global_number, pe->addr);
 581		return (void *)pe;
 582	}
 583
 584	return NULL;
 585}
 586
 587static int eeh_clear_pe_frozen_state(struct eeh_pe *pe,
 588				     bool clear_sw_state)
 589{
 590	void *rc;
 591
 592	rc = eeh_pe_traverse(pe, __eeh_clear_pe_frozen_state, &clear_sw_state);
 593	if (!rc)
 594		eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
 595
 596	return rc ? -EIO : 0;
 597}
 598
 599int eeh_pe_reset_and_recover(struct eeh_pe *pe)
 600{
 601	int ret;
 602
 603	/* Bail if the PE is being recovered */
 604	if (pe->state & EEH_PE_RECOVERING)
 605		return 0;
 606
 607	/* Put the PE into recovery mode */
 608	eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
 609
 610	/* Save states */
 611	eeh_pe_dev_traverse(pe, eeh_dev_save_state, NULL);
 612
 613	/* Issue reset */
 614	ret = eeh_pe_reset_full(pe);
 615	if (ret) {
 616		eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
 617		return ret;
 618	}
 619
 620	/* Unfreeze the PE */
 621	ret = eeh_clear_pe_frozen_state(pe, true);
 622	if (ret) {
 623		eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
 624		return ret;
 625	}
 626
 627	/* Restore device state */
 628	eeh_pe_dev_traverse(pe, eeh_dev_restore_state, NULL);
 629
 630	/* Clear recovery mode */
 631	eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
 632
 633	return 0;
 634}
 635
 636/**
 637 * eeh_reset_device - Perform actual reset of a pci slot
 638 * @driver_eeh_aware: Does the device's driver provide EEH support?
 639 * @pe: EEH PE
 640 * @bus: PCI bus corresponding to the isolcated slot
 641 * @rmv_data: Optional, list to record removed devices
 642 *
 643 * This routine must be called to do reset on the indicated PE.
 644 * During the reset, udev might be invoked because those affected
 645 * PCI devices will be removed and then added.
 646 */
 647static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus,
 648			    struct eeh_rmv_data *rmv_data,
 649			    bool driver_eeh_aware)
 650{
 651	time64_t tstamp;
 652	int cnt, rc;
 653	struct eeh_dev *edev;
 
 
 
 
 
 654
 655	/* pcibios will clear the counter; save the value */
 656	cnt = pe->freeze_count;
 657	tstamp = pe->tstamp;
 658
 659	/*
 660	 * We don't remove the corresponding PE instances because
 661	 * we need the information afterwords. The attached EEH
 662	 * devices are expected to be attached soon when calling
 663	 * into pci_hp_add_devices().
 664	 */
 665	eeh_pe_state_mark(pe, EEH_PE_KEEP);
 666	if (driver_eeh_aware || (pe->type & EEH_PE_VF)) {
 667		eeh_pe_dev_traverse(pe, eeh_rmv_device, rmv_data);
 668	} else {
 669		pci_lock_rescan_remove();
 670		pci_hp_remove_devices(bus);
 671		pci_unlock_rescan_remove();
 672	}
 673
 674	/*
 675	 * Reset the pci controller. (Asserts RST#; resets config space).
 676	 * Reconfigure bridges and devices. Don't try to bring the system
 677	 * up if the reset failed for some reason.
 678	 *
 679	 * During the reset, it's very dangerous to have uncontrolled PCI
 680	 * config accesses. So we prefer to block them. However, controlled
 681	 * PCI config accesses initiated from EEH itself are allowed.
 682	 */
 683	rc = eeh_pe_reset_full(pe);
 684	if (rc)
 685		return rc;
 686
 687	pci_lock_rescan_remove();
 688
 689	/* Restore PE */
 690	eeh_ops->configure_bridge(pe);
 691	eeh_pe_restore_bars(pe);
 692
 693	/* Clear frozen state */
 694	rc = eeh_clear_pe_frozen_state(pe, false);
 695	if (rc) {
 696		pci_unlock_rescan_remove();
 697		return rc;
 698	}
 699
 700	/* Give the system 5 seconds to finish running the user-space
 701	 * hotplug shutdown scripts, e.g. ifdown for ethernet.  Yes,
 702	 * this is a hack, but if we don't do this, and try to bring
 703	 * the device up before the scripts have taken it down,
 704	 * potentially weird things happen.
 705	 */
 706	if (!driver_eeh_aware || rmv_data->removed) {
 707		pr_info("EEH: Sleep 5s ahead of %s hotplug\n",
 708			(driver_eeh_aware ? "partial" : "complete"));
 709		ssleep(5);
 710
 711		/*
 712		 * The EEH device is still connected with its parent
 713		 * PE. We should disconnect it so the binding can be
 714		 * rebuilt when adding PCI devices.
 715		 */
 716		edev = list_first_entry(&pe->edevs, struct eeh_dev, list);
 717		eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
 718		if (pe->type & EEH_PE_VF) {
 719			eeh_add_virt_device(edev, NULL);
 720		} else {
 721			if (!driver_eeh_aware)
 722				eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
 723			pci_hp_add_devices(bus);
 724		}
 725	}
 726	eeh_pe_state_clear(pe, EEH_PE_KEEP);
 727
 728	pe->tstamp = tstamp;
 729	pe->freeze_count = cnt;
 730
 731	pci_unlock_rescan_remove();
 732	return 0;
 733}
 734
 735/* The longest amount of time to wait for a pci device
 736 * to come back on line, in seconds.
 737 */
 738#define MAX_WAIT_FOR_RECOVERY 300
 739
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 740/**
 741 * eeh_handle_normal_event - Handle EEH events on a specific PE
 742 * @pe: EEH PE - which should not be used after we return, as it may
 743 * have been invalidated.
 744 *
 745 * Attempts to recover the given PE.  If recovery fails or the PE has failed
 746 * too many times, remove the PE.
 747 *
 748 * While PHB detects address or data parity errors on particular PCI
 749 * slot, the associated PE will be frozen. Besides, DMA's occurring
 750 * to wild addresses (which usually happen due to bugs in device
 751 * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
 752 * #PERR or other misc PCI-related errors also can trigger EEH errors.
 753 *
 754 * Recovery process consists of unplugging the device driver (which
 755 * generated hotplug events to userspace), then issuing a PCI #RST to
 756 * the device, then reconfiguring the PCI config space for all bridges
 757 * & devices under this slot, and then finally restarting the device
 758 * drivers (which cause a second set of hotplug events to go out to
 759 * userspace).
 760 */
 761void eeh_handle_normal_event(struct eeh_pe *pe)
 762{
 763	struct pci_bus *bus;
 764	struct eeh_dev *edev, *tmp;
 
 765	int rc = 0;
 766	enum pci_ers_result result = PCI_ERS_RESULT_NONE;
 767	struct eeh_rmv_data rmv_data = {LIST_HEAD_INIT(rmv_data.edev_list), 0};
 
 
 768
 769	bus = eeh_pe_bus_get(pe);
 770	if (!bus) {
 771		pr_err("%s: Cannot find PCI bus for PHB#%x-PE#%x\n",
 772			__func__, pe->phb->global_number, pe->addr);
 773		return;
 774	}
 775
 776	eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 777
 778	eeh_pe_update_time_stamp(pe);
 779	pe->freeze_count++;
 780	if (pe->freeze_count > eeh_max_freezes) {
 781		pr_err("EEH: PHB#%x-PE#%x has failed %d times in the\n"
 782		       "last hour and has been permanently disabled.\n",
 783		       pe->phb->global_number, pe->addr,
 784		       pe->freeze_count);
 785		goto hard_fail;
 
 786	}
 787	pr_warn("EEH: This PCI device has failed %d times in the last hour\n",
 788		pe->freeze_count);
 789
 790	/* Walk the various device drivers attached to this slot through
 791	 * a reset sequence, giving each an opportunity to do what it needs
 792	 * to accomplish the reset.  Each child gets a report of the
 793	 * status ... if any child can't handle the reset, then the entire
 794	 * slot is dlpar removed and added.
 795	 *
 796	 * When the PHB is fenced, we have to issue a reset to recover from
 797	 * the error. Override the result if necessary to have partially
 798	 * hotplug for this case.
 799	 */
 
 
 800	pr_info("EEH: Notify device drivers to shutdown\n");
 801	eeh_pe_dev_traverse(pe, eeh_report_error, &result);
 802	if ((pe->type & EEH_PE_PHB) &&
 803	    result != PCI_ERS_RESULT_NONE &&
 804	    result != PCI_ERS_RESULT_NEED_RESET)
 
 
 
 
 
 
 
 
 805		result = PCI_ERS_RESULT_NEED_RESET;
 806
 807	/* Get the current PCI slot state. This can take a long time,
 808	 * sometimes over 300 seconds for certain systems.
 809	 */
 810	rc = eeh_ops->wait_state(pe, MAX_WAIT_FOR_RECOVERY*1000);
 811	if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
 812		pr_warn("EEH: Permanent failure\n");
 813		goto hard_fail;
 814	}
 815
 816	/* Since rtas may enable MMIO when posting the error log,
 817	 * don't post the error log until after all dev drivers
 818	 * have been informed.
 819	 */
 820	pr_info("EEH: Collect temporary log\n");
 821	eeh_slot_error_detail(pe, EEH_LOG_TEMP);
 822
 823	/* If all device drivers were EEH-unaware, then shut
 824	 * down all of the device drivers, and hope they
 825	 * go down willingly, without panicing the system.
 826	 */
 827	if (result == PCI_ERS_RESULT_NONE) {
 828		pr_info("EEH: Reset with hotplug activity\n");
 829		rc = eeh_reset_device(pe, bus, NULL, false);
 830		if (rc) {
 831			pr_warn("%s: Unable to reset, err=%d\n",
 832				__func__, rc);
 833			goto hard_fail;
 834		}
 835	}
 836
 837	/* If all devices reported they can proceed, then re-enable MMIO */
 838	if (result == PCI_ERS_RESULT_CAN_RECOVER) {
 839		pr_info("EEH: Enable I/O for affected devices\n");
 840		rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
 841
 842		if (rc < 0)
 843			goto hard_fail;
 
 844		if (rc) {
 845			result = PCI_ERS_RESULT_NEED_RESET;
 846		} else {
 847			pr_info("EEH: Notify device drivers to resume I/O\n");
 848			eeh_pe_dev_traverse(pe, eeh_report_mmio_enabled, &result);
 
 849		}
 850	}
 851
 852	/* If all devices reported they can proceed, then re-enable DMA */
 853	if (result == PCI_ERS_RESULT_CAN_RECOVER) {
 854		pr_info("EEH: Enabled DMA for affected devices\n");
 855		rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
 856
 857		if (rc < 0)
 858			goto hard_fail;
 
 859		if (rc) {
 860			result = PCI_ERS_RESULT_NEED_RESET;
 861		} else {
 862			/*
 863			 * We didn't do PE reset for the case. The PE
 864			 * is still in frozen state. Clear it before
 865			 * resuming the PE.
 866			 */
 867			eeh_pe_state_clear(pe, EEH_PE_ISOLATED);
 868			result = PCI_ERS_RESULT_RECOVERED;
 869		}
 870	}
 871
 872	/* If any device has a hard failure, then shut off everything. */
 873	if (result == PCI_ERS_RESULT_DISCONNECT) {
 874		pr_warn("EEH: Device driver gave up\n");
 875		goto hard_fail;
 876	}
 877
 878	/* If any device called out for a reset, then reset the slot */
 879	if (result == PCI_ERS_RESULT_NEED_RESET) {
 880		pr_info("EEH: Reset without hotplug activity\n");
 881		rc = eeh_reset_device(pe, bus, &rmv_data, true);
 882		if (rc) {
 883			pr_warn("%s: Cannot reset, err=%d\n",
 884				__func__, rc);
 885			goto hard_fail;
 886		}
 887
 888		pr_info("EEH: Notify device drivers "
 889			"the completion of reset\n");
 890		result = PCI_ERS_RESULT_NONE;
 891		eeh_pe_dev_traverse(pe, eeh_report_reset, &result);
 892	}
 893
 894	/* All devices should claim they have recovered by now. */
 895	if ((result != PCI_ERS_RESULT_RECOVERED) &&
 896	    (result != PCI_ERS_RESULT_NONE)) {
 897		pr_warn("EEH: Not recovered\n");
 898		goto hard_fail;
 899	}
 900
 901	/*
 902	 * For those hot removed VFs, we should add back them after PF get
 903	 * recovered properly.
 904	 */
 905	list_for_each_entry_safe(edev, tmp, &rmv_data.edev_list, rmv_list) {
 906		eeh_add_virt_device(edev, NULL);
 907		list_del(&edev->rmv_list);
 908	}
 
 
 
 909
 910	/* Tell all device drivers that they can resume operations */
 911	pr_info("EEH: Notify device driver to resume\n");
 912	eeh_pe_dev_traverse(pe, eeh_report_resume, NULL);
 
 
 
 
 
 
 
 
 913
 914	goto final;
 
 
 915
 916hard_fail:
 917	/*
 918	 * About 90% of all real-life EEH failures in the field
 919	 * are due to poorly seated PCI cards. Only 10% or so are
 920	 * due to actual, failed cards.
 921	 */
 922	pr_err("EEH: Unable to recover from failure from PHB#%x-PE#%x.\n"
 923	       "Please try reseating or replacing it\n",
 924		pe->phb->global_number, pe->addr);
 925
 926	eeh_slot_error_detail(pe, EEH_LOG_PERM);
 927
 928	/* Notify all devices that they're about to go down. */
 929	eeh_pe_dev_traverse(pe, eeh_report_failure, NULL);
 
 
 
 930
 931	/* Mark the PE to be removed permanently */
 932	eeh_pe_state_mark(pe, EEH_PE_REMOVED);
 933
 934	/*
 935	 * Shut down the device drivers for good. We mark
 936	 * all removed devices correctly to avoid access
 937	 * the their PCI config any more.
 938	 */
 939	if (pe->type & EEH_PE_VF) {
 940		eeh_pe_dev_traverse(pe, eeh_rmv_device, NULL);
 941		eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
 942	} else {
 943		eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
 944		eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
 945
 946		pci_lock_rescan_remove();
 947		pci_hp_remove_devices(bus);
 948		pci_unlock_rescan_remove();
 949		/* The passed PE should no longer be used */
 950		return;
 951	}
 952final:
 953	eeh_pe_state_clear(pe, EEH_PE_RECOVERING);
 
 
 
 
 
 
 
 
 
 
 
 
 954}
 955
 956/**
 957 * eeh_handle_special_event - Handle EEH events without a specific failing PE
 958 *
 959 * Called when an EEH event is detected but can't be narrowed down to a
 960 * specific PE.  Iterates through possible failures and handles them as
 961 * necessary.
 962 */
 963void eeh_handle_special_event(void)
 964{
 965	struct eeh_pe *pe, *phb_pe;
 
 966	struct pci_bus *bus;
 967	struct pci_controller *hose;
 968	unsigned long flags;
 969	int rc;
 970
 971
 972	do {
 973		rc = eeh_ops->next_error(&pe);
 974
 975		switch (rc) {
 976		case EEH_NEXT_ERR_DEAD_IOC:
 977			/* Mark all PHBs in dead state */
 978			eeh_serialize_lock(&flags);
 979
 980			/* Purge all events */
 981			eeh_remove_event(NULL, true);
 982
 983			list_for_each_entry(hose, &hose_list, list_node) {
 984				phb_pe = eeh_phb_pe_get(hose);
 985				if (!phb_pe) continue;
 986
 987				eeh_pe_state_mark(phb_pe, EEH_PE_ISOLATED);
 988			}
 989
 990			eeh_serialize_unlock(flags);
 991
 992			break;
 993		case EEH_NEXT_ERR_FROZEN_PE:
 994		case EEH_NEXT_ERR_FENCED_PHB:
 995		case EEH_NEXT_ERR_DEAD_PHB:
 996			/* Mark the PE in fenced state */
 997			eeh_serialize_lock(&flags);
 998
 999			/* Purge all events of the PHB */
1000			eeh_remove_event(pe, true);
1001
1002			if (rc == EEH_NEXT_ERR_DEAD_PHB)
1003				eeh_pe_state_mark(pe, EEH_PE_ISOLATED);
1004			else
1005				eeh_pe_state_mark(pe,
1006					EEH_PE_ISOLATED | EEH_PE_RECOVERING);
1007
1008			eeh_serialize_unlock(flags);
1009
1010			break;
1011		case EEH_NEXT_ERR_NONE:
1012			return;
1013		default:
1014			pr_warn("%s: Invalid value %d from next_error()\n",
1015				__func__, rc);
1016			return;
1017		}
1018
1019		/*
1020		 * For fenced PHB and frozen PE, it's handled as normal
1021		 * event. We have to remove the affected PHBs for dead
1022		 * PHB and IOC
1023		 */
1024		if (rc == EEH_NEXT_ERR_FROZEN_PE ||
1025		    rc == EEH_NEXT_ERR_FENCED_PHB) {
 
1026			eeh_handle_normal_event(pe);
1027		} else {
 
 
 
 
 
 
 
 
 
 
 
1028			pci_lock_rescan_remove();
1029			list_for_each_entry(hose, &hose_list, list_node) {
1030				phb_pe = eeh_phb_pe_get(hose);
1031				if (!phb_pe ||
1032				    !(phb_pe->state & EEH_PE_ISOLATED) ||
1033				    (phb_pe->state & EEH_PE_RECOVERING))
1034					continue;
1035
1036				/* Notify all devices to be down */
1037				eeh_pe_state_clear(pe, EEH_PE_PRI_BUS);
1038				eeh_pe_dev_traverse(pe,
1039					eeh_report_failure, NULL);
1040				bus = eeh_pe_bus_get(phb_pe);
1041				if (!bus) {
1042					pr_err("%s: Cannot find PCI bus for "
1043					       "PHB#%x-PE#%x\n",
1044					       __func__,
1045					       pe->phb->global_number,
1046					       pe->addr);
1047					break;
1048				}
1049				pci_hp_remove_devices(bus);
1050			}
1051			pci_unlock_rescan_remove();
1052		}
1053
1054		/*
1055		 * If we have detected dead IOC, we needn't proceed
1056		 * any more since all PHBs would have been removed
1057		 */
1058		if (rc == EEH_NEXT_ERR_DEAD_IOC)
1059			break;
1060	} while (rc != EEH_NEXT_ERR_NONE);
1061}
v6.8
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * PCI Error Recovery Driver for RPA-compliant PPC64 platform.
   4 * Copyright IBM Corp. 2004 2005
   5 * Copyright Linas Vepstas <linas@linas.org> 2004, 2005
   6 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   7 * Send comments and feedback to Linas Vepstas <linas@austin.ibm.com>
   8 */
   9#include <linux/delay.h>
  10#include <linux/interrupt.h>
  11#include <linux/irq.h>
  12#include <linux/module.h>
  13#include <linux/pci.h>
  14#include <linux/pci_hotplug.h>
  15#include <asm/eeh.h>
  16#include <asm/eeh_event.h>
  17#include <asm/ppc-pci.h>
  18#include <asm/pci-bridge.h>
 
  19#include <asm/rtas.h>
  20
  21struct eeh_rmv_data {
  22	struct list_head removed_vf_list;
  23	int removed_dev_count;
  24};
  25
  26static int eeh_result_priority(enum pci_ers_result result)
  27{
  28	switch (result) {
  29	case PCI_ERS_RESULT_NONE:
  30		return 1;
  31	case PCI_ERS_RESULT_NO_AER_DRIVER:
  32		return 2;
  33	case PCI_ERS_RESULT_RECOVERED:
  34		return 3;
  35	case PCI_ERS_RESULT_CAN_RECOVER:
  36		return 4;
  37	case PCI_ERS_RESULT_DISCONNECT:
  38		return 5;
  39	case PCI_ERS_RESULT_NEED_RESET:
  40		return 6;
  41	default:
  42		WARN_ONCE(1, "Unknown pci_ers_result value: %d\n", result);
  43		return 0;
  44	}
  45};
  46
  47static const char *pci_ers_result_name(enum pci_ers_result result)
  48{
  49	switch (result) {
  50	case PCI_ERS_RESULT_NONE:
  51		return "none";
  52	case PCI_ERS_RESULT_CAN_RECOVER:
  53		return "can recover";
  54	case PCI_ERS_RESULT_NEED_RESET:
  55		return "need reset";
  56	case PCI_ERS_RESULT_DISCONNECT:
  57		return "disconnect";
  58	case PCI_ERS_RESULT_RECOVERED:
  59		return "recovered";
  60	case PCI_ERS_RESULT_NO_AER_DRIVER:
  61		return "no AER driver";
  62	default:
  63		WARN_ONCE(1, "Unknown result type: %d\n", result);
  64		return "unknown";
  65	}
  66};
  67
  68static enum pci_ers_result pci_ers_merge_result(enum pci_ers_result old,
  69						enum pci_ers_result new)
  70{
  71	if (eeh_result_priority(new) > eeh_result_priority(old))
  72		return new;
  73	return old;
  74}
  75
  76static bool eeh_dev_removed(struct eeh_dev *edev)
  77{
  78	return !edev || (edev->mode & EEH_DEV_REMOVED);
  79}
  80
  81static bool eeh_edev_actionable(struct eeh_dev *edev)
  82{
  83	if (!edev->pdev)
  84		return false;
  85	if (edev->pdev->error_state == pci_channel_io_perm_failure)
  86		return false;
  87	if (eeh_dev_removed(edev))
  88		return false;
  89	if (eeh_pe_passed(edev->pe))
  90		return false;
  91
  92	return true;
  93}
  94
  95/**
  96 * eeh_pcid_get - Get the PCI device driver
  97 * @pdev: PCI device
  98 *
  99 * The function is used to retrieve the PCI device driver for
 100 * the indicated PCI device. Besides, we will increase the reference
 101 * of the PCI device driver to prevent that being unloaded on
 102 * the fly. Otherwise, kernel crash would be seen.
 103 */
 104static inline struct pci_driver *eeh_pcid_get(struct pci_dev *pdev)
 105{
 106	if (!pdev || !pdev->dev.driver)
 107		return NULL;
 108
 109	if (!try_module_get(pdev->dev.driver->owner))
 110		return NULL;
 111
 112	return to_pci_driver(pdev->dev.driver);
 113}
 114
 115/**
 116 * eeh_pcid_put - Dereference on the PCI device driver
 117 * @pdev: PCI device
 118 *
 119 * The function is called to do dereference on the PCI device
 120 * driver of the indicated PCI device.
 121 */
 122static inline void eeh_pcid_put(struct pci_dev *pdev)
 123{
 124	if (!pdev || !pdev->dev.driver)
 125		return;
 126
 127	module_put(pdev->dev.driver->owner);
 128}
 129
 130/**
 131 * eeh_disable_irq - Disable interrupt for the recovering device
 132 * @dev: PCI device
 133 *
 134 * This routine must be called when reporting temporary or permanent
 135 * error to the particular PCI device to disable interrupt of that
 136 * device. If the device has enabled MSI or MSI-X interrupt, we needn't
 137 * do real work because EEH should freeze DMA transfers for those PCI
 138 * devices encountering EEH errors, which includes MSI or MSI-X.
 139 */
 140static void eeh_disable_irq(struct eeh_dev *edev)
 141{
 
 
 142	/* Don't disable MSI and MSI-X interrupts. They are
 143	 * effectively disabled by the DMA Stopped state
 144	 * when an EEH error occurs.
 145	 */
 146	if (edev->pdev->msi_enabled || edev->pdev->msix_enabled)
 147		return;
 148
 149	if (!irq_has_action(edev->pdev->irq))
 150		return;
 151
 152	edev->mode |= EEH_DEV_IRQ_DISABLED;
 153	disable_irq_nosync(edev->pdev->irq);
 154}
 155
 156/**
 157 * eeh_enable_irq - Enable interrupt for the recovering device
 158 * @dev: PCI device
 159 *
 160 * This routine must be called to enable interrupt while failed
 161 * device could be resumed.
 162 */
 163static void eeh_enable_irq(struct eeh_dev *edev)
 164{
 
 
 165	if ((edev->mode) & EEH_DEV_IRQ_DISABLED) {
 166		edev->mode &= ~EEH_DEV_IRQ_DISABLED;
 167		/*
 168		 * FIXME !!!!!
 169		 *
 170		 * This is just ass backwards. This maze has
 171		 * unbalanced irq_enable/disable calls. So instead of
 172		 * finding the root cause it works around the warning
 173		 * in the irq_enable code by conditionally calling
 174		 * into it.
 175		 *
 176		 * That's just wrong.The warning in the core code is
 177		 * there to tell people to fix their asymmetries in
 178		 * their own code, not by abusing the core information
 179		 * to avoid it.
 180		 *
 181		 * I so wish that the assymetry would be the other way
 182		 * round and a few more irq_disable calls render that
 183		 * shit unusable forever.
 184		 *
 185		 *	tglx
 186		 */
 187		if (irqd_irq_disabled(irq_get_irq_data(edev->pdev->irq)))
 188			enable_irq(edev->pdev->irq);
 189	}
 190}
 191
 192static void eeh_dev_save_state(struct eeh_dev *edev, void *userdata)
 
 
 
 
 
 
 
 
 
 193{
 
 194	struct pci_dev *pdev;
 195
 196	if (!edev)
 197		return;
 198
 199	/*
 200	 * We cannot access the config space on some adapters.
 201	 * Otherwise, it will cause fenced PHB. We don't save
 202	 * the content in their config space and will restore
 203	 * from the initial config space saved when the EEH
 204	 * device is created.
 205	 */
 206	if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED))
 207		return;
 208
 209	pdev = eeh_dev_to_pci_dev(edev);
 210	if (!pdev)
 211		return;
 212
 213	pci_save_state(pdev);
 
 214}
 215
 216static void eeh_set_channel_state(struct eeh_pe *root, pci_channel_state_t s)
 
 
 
 
 
 
 
 
 
 217{
 218	struct eeh_pe *pe;
 219	struct eeh_dev *edev, *tmp;
 
 
 
 
 
 
 
 
 
 
 
 220
 221	eeh_for_each_pe(root, pe)
 222		eeh_pe_for_each_dev(pe, edev, tmp)
 223			if (eeh_edev_actionable(edev))
 224				edev->pdev->error_state = s;
 225}
 226
 227static void eeh_set_irq_state(struct eeh_pe *root, bool enable)
 228{
 229	struct eeh_pe *pe;
 230	struct eeh_dev *edev, *tmp;
 231
 232	eeh_for_each_pe(root, pe) {
 233		eeh_pe_for_each_dev(pe, edev, tmp) {
 234			if (!eeh_edev_actionable(edev))
 235				continue;
 236
 237			if (!eeh_pcid_get(edev->pdev))
 238				continue;
 
 239
 240			if (enable)
 241				eeh_enable_irq(edev);
 242			else
 243				eeh_disable_irq(edev);
 244
 245			eeh_pcid_put(edev->pdev);
 246		}
 247	}
 
 
 248}
 249
 250typedef enum pci_ers_result (*eeh_report_fn)(struct eeh_dev *,
 251					     struct pci_dev *,
 252					     struct pci_driver *);
 253static void eeh_pe_report_edev(struct eeh_dev *edev, eeh_report_fn fn,
 254			       enum pci_ers_result *result)
 
 
 
 
 
 255{
 256	struct pci_dev *pdev;
 
 
 257	struct pci_driver *driver;
 258	enum pci_ers_result new_result;
 259
 260	pci_lock_rescan_remove();
 261	pdev = edev->pdev;
 262	if (pdev)
 263		get_device(&pdev->dev);
 264	pci_unlock_rescan_remove();
 265	if (!pdev) {
 266		eeh_edev_info(edev, "no device");
 267		return;
 268	}
 269	device_lock(&pdev->dev);
 270	if (eeh_edev_actionable(edev)) {
 271		driver = eeh_pcid_get(pdev);
 272
 273		if (!driver)
 274			eeh_edev_info(edev, "no driver");
 275		else if (!driver->err_handler)
 276			eeh_edev_info(edev, "driver not EEH aware");
 277		else if (edev->mode & EEH_DEV_NO_HANDLER)
 278			eeh_edev_info(edev, "driver bound too late");
 279		else {
 280			new_result = fn(edev, pdev, driver);
 281			eeh_edev_info(edev, "%s driver reports: '%s'",
 282				      driver->name,
 283				      pci_ers_result_name(new_result));
 284			if (result)
 285				*result = pci_ers_merge_result(*result,
 286							       new_result);
 287		}
 288		if (driver)
 289			eeh_pcid_put(pdev);
 290	} else {
 291		eeh_edev_info(edev, "not actionable (%d,%d,%d)", !!pdev,
 292			      !eeh_dev_removed(edev), !eeh_pe_passed(edev->pe));
 293	}
 294	device_unlock(&pdev->dev);
 295	if (edev->pdev != pdev)
 296		eeh_edev_warn(edev, "Device changed during processing!\n");
 297	put_device(&pdev->dev);
 298}
 299
 300static void eeh_pe_report(const char *name, struct eeh_pe *root,
 301			  eeh_report_fn fn, enum pci_ers_result *result)
 302{
 303	struct eeh_pe *pe;
 304	struct eeh_dev *edev, *tmp;
 305
 306	pr_info("EEH: Beginning: '%s'\n", name);
 307	eeh_for_each_pe(root, pe) eeh_pe_for_each_dev(pe, edev, tmp)
 308		eeh_pe_report_edev(edev, fn, result);
 309	if (result)
 310		pr_info("EEH: Finished:'%s' with aggregate recovery state:'%s'\n",
 311			name, pci_ers_result_name(*result));
 312	else
 313		pr_info("EEH: Finished:'%s'", name);
 314}
 315
 316/**
 317 * eeh_report_error - Report pci error to each device driver
 318 * @edev: eeh device
 319 * @driver: device's PCI driver
 320 *
 321 * Report an EEH error to each device driver.
 322 */
 323static enum pci_ers_result eeh_report_error(struct eeh_dev *edev,
 324					    struct pci_dev *pdev,
 325					    struct pci_driver *driver)
 326{
 327	enum pci_ers_result rc;
 328
 329	if (!driver->err_handler->error_detected)
 330		return PCI_ERS_RESULT_NONE;
 331
 332	eeh_edev_info(edev, "Invoking %s->error_detected(IO frozen)",
 333		      driver->name);
 334	rc = driver->err_handler->error_detected(pdev, pci_channel_io_frozen);
 335
 336	edev->in_error = true;
 337	pci_uevent_ers(pdev, PCI_ERS_RESULT_NONE);
 338	return rc;
 339}
 340
 341/**
 342 * eeh_report_mmio_enabled - Tell drivers that MMIO has been enabled
 343 * @edev: eeh device
 344 * @driver: device's PCI driver
 345 *
 346 * Tells each device driver that IO ports, MMIO and config space I/O
 347 * are now enabled.
 348 */
 349static enum pci_ers_result eeh_report_mmio_enabled(struct eeh_dev *edev,
 350						   struct pci_dev *pdev,
 351						   struct pci_driver *driver)
 352{
 353	if (!driver->err_handler->mmio_enabled)
 354		return PCI_ERS_RESULT_NONE;
 355	eeh_edev_info(edev, "Invoking %s->mmio_enabled()", driver->name);
 356	return driver->err_handler->mmio_enabled(pdev);
 357}
 358
 359/**
 360 * eeh_report_reset - Tell device that slot has been reset
 361 * @edev: eeh device
 362 * @driver: device's PCI driver
 363 *
 364 * This routine must be called while EEH tries to reset particular
 365 * PCI device so that the associated PCI device driver could take
 366 * some actions, usually to save data the driver needs so that the
 367 * driver can work again while the device is recovered.
 368 */
 369static enum pci_ers_result eeh_report_reset(struct eeh_dev *edev,
 370					    struct pci_dev *pdev,
 371					    struct pci_driver *driver)
 372{
 373	if (!driver->err_handler->slot_reset || !edev->in_error)
 374		return PCI_ERS_RESULT_NONE;
 375	eeh_edev_info(edev, "Invoking %s->slot_reset()", driver->name);
 376	return driver->err_handler->slot_reset(pdev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 377}
 378
 379static void eeh_dev_restore_state(struct eeh_dev *edev, void *userdata)
 380{
 
 381	struct pci_dev *pdev;
 382
 383	if (!edev)
 384		return;
 385
 386	/*
 387	 * The content in the config space isn't saved because
 388	 * the blocked config space on some adapters. We have
 389	 * to restore the initial saved config space when the
 390	 * EEH device is created.
 391	 */
 392	if (edev->pe && (edev->pe->state & EEH_PE_CFG_RESTRICTED)) {
 393		if (list_is_last(&edev->entry, &edev->pe->edevs))
 394			eeh_pe_restore_bars(edev->pe);
 395
 396		return;
 397	}
 398
 399	pdev = eeh_dev_to_pci_dev(edev);
 400	if (!pdev)
 401		return;
 402
 403	pci_restore_state(pdev);
 
 404}
 405
 406/**
 407 * eeh_report_resume - Tell device to resume normal operations
 408 * @edev: eeh device
 409 * @driver: device's PCI driver
 410 *
 411 * This routine must be called to notify the device driver that it
 412 * could resume so that the device driver can do some initialization
 413 * to make the recovered device work again.
 414 */
 415static enum pci_ers_result eeh_report_resume(struct eeh_dev *edev,
 416					     struct pci_dev *pdev,
 417					     struct pci_driver *driver)
 418{
 419	if (!driver->err_handler->resume || !edev->in_error)
 420		return PCI_ERS_RESULT_NONE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 421
 422	eeh_edev_info(edev, "Invoking %s->resume()", driver->name);
 423	driver->err_handler->resume(pdev);
 424
 425	pci_uevent_ers(edev->pdev, PCI_ERS_RESULT_RECOVERED);
 
 
 426#ifdef CONFIG_PCI_IOV
 427	if (eeh_ops->notify_resume)
 428		eeh_ops->notify_resume(edev);
 429#endif
 430	return PCI_ERS_RESULT_NONE;
 
 
 431}
 432
 433/**
 434 * eeh_report_failure - Tell device driver that device is dead.
 435 * @edev: eeh device
 436 * @driver: device's PCI driver
 437 *
 438 * This informs the device driver that the device is permanently
 439 * dead, and that no further recovery attempts will be made on it.
 440 */
 441static enum pci_ers_result eeh_report_failure(struct eeh_dev *edev,
 442					      struct pci_dev *pdev,
 443					      struct pci_driver *driver)
 444{
 445	enum pci_ers_result rc;
 
 
 
 
 
 
 
 
 
 
 
 
 
 446
 447	if (!driver->err_handler->error_detected)
 448		return PCI_ERS_RESULT_NONE;
 
 449
 450	eeh_edev_info(edev, "Invoking %s->error_detected(permanent failure)",
 451		      driver->name);
 452	rc = driver->err_handler->error_detected(pdev,
 453						 pci_channel_io_perm_failure);
 454
 455	pci_uevent_ers(pdev, PCI_ERS_RESULT_DISCONNECT);
 456	return rc;
 
 
 
 
 457}
 458
 459static void *eeh_add_virt_device(struct eeh_dev *edev)
 460{
 461	struct pci_driver *driver;
 
 462	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 
 463
 464	if (!(edev->physfn)) {
 465		eeh_edev_warn(edev, "Not for VF\n");
 
 
 466		return NULL;
 467	}
 468
 469	driver = eeh_pcid_get(dev);
 470	if (driver) {
 471		if (driver->err_handler) {
 472			eeh_pcid_put(dev);
 473			return NULL;
 474		}
 475		eeh_pcid_put(dev);
 476	}
 477
 478#ifdef CONFIG_PCI_IOV
 479	pci_iov_add_virtfn(edev->physfn, edev->vf_index);
 480#endif
 481	return NULL;
 482}
 483
 484static void eeh_rmv_device(struct eeh_dev *edev, void *userdata)
 485{
 486	struct pci_driver *driver;
 
 487	struct pci_dev *dev = eeh_dev_to_pci_dev(edev);
 488	struct eeh_rmv_data *rmv_data = (struct eeh_rmv_data *)userdata;
 
 489
 490	/*
 491	 * Actually, we should remove the PCI bridges as well.
 492	 * However, that's lots of complexity to do that,
 493	 * particularly some of devices under the bridge might
 494	 * support EEH. So we just care about PCI devices for
 495	 * simplicity here.
 496	 */
 497	if (!eeh_edev_actionable(edev) ||
 498	    (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE))
 499		return;
 
 
 
 
 
 
 
 
 
 500
 501	if (rmv_data) {
 502		driver = eeh_pcid_get(dev);
 503		if (driver) {
 504			if (driver->err_handler &&
 505			    driver->err_handler->error_detected &&
 506			    driver->err_handler->slot_reset) {
 507				eeh_pcid_put(dev);
 508				return;
 509			}
 510			eeh_pcid_put(dev);
 511		}
 512	}
 513
 514	/* Remove it from PCI subsystem */
 515	pr_info("EEH: Removing %s without EEH sensitive driver\n",
 516		pci_name(dev));
 
 517	edev->mode |= EEH_DEV_DISCONNECTED;
 518	if (rmv_data)
 519		rmv_data->removed_dev_count++;
 520
 521	if (edev->physfn) {
 522#ifdef CONFIG_PCI_IOV
 523		pci_iov_remove_virtfn(edev->physfn, edev->vf_index);
 
 
 524		edev->pdev = NULL;
 
 
 
 
 
 
 525#endif
 526		if (rmv_data)
 527			list_add(&edev->rmv_entry, &rmv_data->removed_vf_list);
 528	} else {
 529		pci_lock_rescan_remove();
 530		pci_stop_and_remove_bus_device(dev);
 531		pci_unlock_rescan_remove();
 532	}
 
 
 533}
 534
 535static void *eeh_pe_detach_dev(struct eeh_pe *pe, void *userdata)
 536{
 
 537	struct eeh_dev *edev, *tmp;
 538
 539	eeh_pe_for_each_dev(pe, edev, tmp) {
 540		if (!(edev->mode & EEH_DEV_DISCONNECTED))
 541			continue;
 542
 543		edev->mode &= ~(EEH_DEV_DISCONNECTED | EEH_DEV_IRQ_DISABLED);
 544		eeh_pe_tree_remove(edev);
 545	}
 546
 547	return NULL;
 548}
 549
 550/*
 551 * Explicitly clear PE's frozen state for PowerNV where
 552 * we have frozen PE until BAR restore is completed. It's
 553 * harmless to clear it for pSeries. To be consistent with
 554 * PE reset (for 3 times), we try to clear the frozen state
 555 * for 3 times as well.
 556 */
 557static int eeh_clear_pe_frozen_state(struct eeh_pe *root, bool include_passed)
 558{
 559	struct eeh_pe *pe;
 560	int i;
 
 561
 562	eeh_for_each_pe(root, pe) {
 563		if (include_passed || !eeh_pe_passed(pe)) {
 564			for (i = 0; i < 3; i++)
 565				if (!eeh_unfreeze_pe(pe))
 566					break;
 567			if (i >= 3)
 568				return -EIO;
 569		}
 570	}
 571	eeh_pe_state_clear(root, EEH_PE_ISOLATED, include_passed);
 572	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 573}
 574
 575int eeh_pe_reset_and_recover(struct eeh_pe *pe)
 576{
 577	int ret;
 578
 579	/* Bail if the PE is being recovered */
 580	if (pe->state & EEH_PE_RECOVERING)
 581		return 0;
 582
 583	/* Put the PE into recovery mode */
 584	eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
 585
 586	/* Save states */
 587	eeh_pe_dev_traverse(pe, eeh_dev_save_state, NULL);
 588
 589	/* Issue reset */
 590	ret = eeh_pe_reset_full(pe, true);
 591	if (ret) {
 592		eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 593		return ret;
 594	}
 595
 596	/* Unfreeze the PE */
 597	ret = eeh_clear_pe_frozen_state(pe, true);
 598	if (ret) {
 599		eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 600		return ret;
 601	}
 602
 603	/* Restore device state */
 604	eeh_pe_dev_traverse(pe, eeh_dev_restore_state, NULL);
 605
 606	/* Clear recovery mode */
 607	eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
 608
 609	return 0;
 610}
 611
 612/**
 613 * eeh_reset_device - Perform actual reset of a pci slot
 614 * @driver_eeh_aware: Does the device's driver provide EEH support?
 615 * @pe: EEH PE
 616 * @bus: PCI bus corresponding to the isolcated slot
 617 * @rmv_data: Optional, list to record removed devices
 618 *
 619 * This routine must be called to do reset on the indicated PE.
 620 * During the reset, udev might be invoked because those affected
 621 * PCI devices will be removed and then added.
 622 */
 623static int eeh_reset_device(struct eeh_pe *pe, struct pci_bus *bus,
 624			    struct eeh_rmv_data *rmv_data,
 625			    bool driver_eeh_aware)
 626{
 627	time64_t tstamp;
 628	int cnt, rc;
 629	struct eeh_dev *edev;
 630	struct eeh_pe *tmp_pe;
 631	bool any_passed = false;
 632
 633	eeh_for_each_pe(pe, tmp_pe)
 634		any_passed |= eeh_pe_passed(tmp_pe);
 635
 636	/* pcibios will clear the counter; save the value */
 637	cnt = pe->freeze_count;
 638	tstamp = pe->tstamp;
 639
 640	/*
 641	 * We don't remove the corresponding PE instances because
 642	 * we need the information afterwords. The attached EEH
 643	 * devices are expected to be attached soon when calling
 644	 * into pci_hp_add_devices().
 645	 */
 646	eeh_pe_state_mark(pe, EEH_PE_KEEP);
 647	if (any_passed || driver_eeh_aware || (pe->type & EEH_PE_VF)) {
 648		eeh_pe_dev_traverse(pe, eeh_rmv_device, rmv_data);
 649	} else {
 650		pci_lock_rescan_remove();
 651		pci_hp_remove_devices(bus);
 652		pci_unlock_rescan_remove();
 653	}
 654
 655	/*
 656	 * Reset the pci controller. (Asserts RST#; resets config space).
 657	 * Reconfigure bridges and devices. Don't try to bring the system
 658	 * up if the reset failed for some reason.
 659	 *
 660	 * During the reset, it's very dangerous to have uncontrolled PCI
 661	 * config accesses. So we prefer to block them. However, controlled
 662	 * PCI config accesses initiated from EEH itself are allowed.
 663	 */
 664	rc = eeh_pe_reset_full(pe, false);
 665	if (rc)
 666		return rc;
 667
 668	pci_lock_rescan_remove();
 669
 670	/* Restore PE */
 671	eeh_ops->configure_bridge(pe);
 672	eeh_pe_restore_bars(pe);
 673
 674	/* Clear frozen state */
 675	rc = eeh_clear_pe_frozen_state(pe, false);
 676	if (rc) {
 677		pci_unlock_rescan_remove();
 678		return rc;
 679	}
 680
 681	/* Give the system 5 seconds to finish running the user-space
 682	 * hotplug shutdown scripts, e.g. ifdown for ethernet.  Yes,
 683	 * this is a hack, but if we don't do this, and try to bring
 684	 * the device up before the scripts have taken it down,
 685	 * potentially weird things happen.
 686	 */
 687	if (!driver_eeh_aware || rmv_data->removed_dev_count) {
 688		pr_info("EEH: Sleep 5s ahead of %s hotplug\n",
 689			(driver_eeh_aware ? "partial" : "complete"));
 690		ssleep(5);
 691
 692		/*
 693		 * The EEH device is still connected with its parent
 694		 * PE. We should disconnect it so the binding can be
 695		 * rebuilt when adding PCI devices.
 696		 */
 697		edev = list_first_entry(&pe->edevs, struct eeh_dev, entry);
 698		eeh_pe_traverse(pe, eeh_pe_detach_dev, NULL);
 699		if (pe->type & EEH_PE_VF) {
 700			eeh_add_virt_device(edev);
 701		} else {
 702			if (!driver_eeh_aware)
 703				eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
 704			pci_hp_add_devices(bus);
 705		}
 706	}
 707	eeh_pe_state_clear(pe, EEH_PE_KEEP, true);
 708
 709	pe->tstamp = tstamp;
 710	pe->freeze_count = cnt;
 711
 712	pci_unlock_rescan_remove();
 713	return 0;
 714}
 715
 716/* The longest amount of time to wait for a pci device
 717 * to come back on line, in seconds.
 718 */
 719#define MAX_WAIT_FOR_RECOVERY 300
 720
 721
 722/* Walks the PE tree after processing an event to remove any stale PEs.
 723 *
 724 * NB: This needs to be recursive to ensure the leaf PEs get removed
 725 * before their parents do. Although this is possible to do recursively
 726 * we don't since this is easier to read and we need to garantee
 727 * the leaf nodes will be handled first.
 728 */
 729static void eeh_pe_cleanup(struct eeh_pe *pe)
 730{
 731	struct eeh_pe *child_pe, *tmp;
 732
 733	list_for_each_entry_safe(child_pe, tmp, &pe->child_list, child)
 734		eeh_pe_cleanup(child_pe);
 735
 736	if (pe->state & EEH_PE_KEEP)
 737		return;
 738
 739	if (!(pe->state & EEH_PE_INVALID))
 740		return;
 741
 742	if (list_empty(&pe->edevs) && list_empty(&pe->child_list)) {
 743		list_del(&pe->child);
 744		kfree(pe);
 745	}
 746}
 747
 748/**
 749 * eeh_check_slot_presence - Check if a device is still present in a slot
 750 * @pdev: pci_dev to check
 751 *
 752 * This function may return a false positive if we can't determine the slot's
 753 * presence state. This might happen for PCIe slots if the PE containing
 754 * the upstream bridge is also frozen, or the bridge is part of the same PE
 755 * as the device.
 756 *
 757 * This shouldn't happen often, but you might see it if you hotplug a PCIe
 758 * switch.
 759 */
 760static bool eeh_slot_presence_check(struct pci_dev *pdev)
 761{
 762	const struct hotplug_slot_ops *ops;
 763	struct pci_slot *slot;
 764	u8 state;
 765	int rc;
 766
 767	if (!pdev)
 768		return false;
 769
 770	if (pdev->error_state == pci_channel_io_perm_failure)
 771		return false;
 772
 773	slot = pdev->slot;
 774	if (!slot || !slot->hotplug)
 775		return true;
 776
 777	ops = slot->hotplug->ops;
 778	if (!ops || !ops->get_adapter_status)
 779		return true;
 780
 781	/* set the attention indicator while we've got the slot ops */
 782	if (ops->set_attention_status)
 783		ops->set_attention_status(slot->hotplug, 1);
 784
 785	rc = ops->get_adapter_status(slot->hotplug, &state);
 786	if (rc)
 787		return true;
 788
 789	return !!state;
 790}
 791
 792static void eeh_clear_slot_attention(struct pci_dev *pdev)
 793{
 794	const struct hotplug_slot_ops *ops;
 795	struct pci_slot *slot;
 796
 797	if (!pdev)
 798		return;
 799
 800	if (pdev->error_state == pci_channel_io_perm_failure)
 801		return;
 802
 803	slot = pdev->slot;
 804	if (!slot || !slot->hotplug)
 805		return;
 806
 807	ops = slot->hotplug->ops;
 808	if (!ops || !ops->set_attention_status)
 809		return;
 810
 811	ops->set_attention_status(slot->hotplug, 0);
 812}
 813
 814/**
 815 * eeh_handle_normal_event - Handle EEH events on a specific PE
 816 * @pe: EEH PE - which should not be used after we return, as it may
 817 * have been invalidated.
 818 *
 819 * Attempts to recover the given PE.  If recovery fails or the PE has failed
 820 * too many times, remove the PE.
 821 *
 822 * While PHB detects address or data parity errors on particular PCI
 823 * slot, the associated PE will be frozen. Besides, DMA's occurring
 824 * to wild addresses (which usually happen due to bugs in device
 825 * drivers or in PCI adapter firmware) can cause EEH error. #SERR,
 826 * #PERR or other misc PCI-related errors also can trigger EEH errors.
 827 *
 828 * Recovery process consists of unplugging the device driver (which
 829 * generated hotplug events to userspace), then issuing a PCI #RST to
 830 * the device, then reconfiguring the PCI config space for all bridges
 831 * & devices under this slot, and then finally restarting the device
 832 * drivers (which cause a second set of hotplug events to go out to
 833 * userspace).
 834 */
 835void eeh_handle_normal_event(struct eeh_pe *pe)
 836{
 837	struct pci_bus *bus;
 838	struct eeh_dev *edev, *tmp;
 839	struct eeh_pe *tmp_pe;
 840	int rc = 0;
 841	enum pci_ers_result result = PCI_ERS_RESULT_NONE;
 842	struct eeh_rmv_data rmv_data =
 843		{LIST_HEAD_INIT(rmv_data.removed_vf_list), 0};
 844	int devices = 0;
 845
 846	bus = eeh_pe_bus_get(pe);
 847	if (!bus) {
 848		pr_err("%s: Cannot find PCI bus for PHB#%x-PE#%x\n",
 849			__func__, pe->phb->global_number, pe->addr);
 850		return;
 851	}
 852
 853	/*
 854	 * When devices are hot-removed we might get an EEH due to
 855	 * a driver attempting to touch the MMIO space of a removed
 856	 * device. In this case we don't have a device to recover
 857	 * so suppress the event if we can't find any present devices.
 858	 *
 859	 * The hotplug driver should take care of tearing down the
 860	 * device itself.
 861	 */
 862	eeh_for_each_pe(pe, tmp_pe)
 863		eeh_pe_for_each_dev(tmp_pe, edev, tmp)
 864			if (eeh_slot_presence_check(edev->pdev))
 865				devices++;
 866
 867	if (!devices) {
 868		pr_debug("EEH: Frozen PHB#%x-PE#%x is empty!\n",
 869			pe->phb->global_number, pe->addr);
 870		goto out; /* nothing to recover */
 871	}
 872
 873	/* Log the event */
 874	if (pe->type & EEH_PE_PHB) {
 875		pr_err("EEH: Recovering PHB#%x, location: %s\n",
 876			pe->phb->global_number, eeh_pe_loc_get(pe));
 877	} else {
 878		struct eeh_pe *phb_pe = eeh_phb_pe_get(pe->phb);
 879
 880		pr_err("EEH: Recovering PHB#%x-PE#%x\n",
 881		       pe->phb->global_number, pe->addr);
 882		pr_err("EEH: PE location: %s, PHB location: %s\n",
 883		       eeh_pe_loc_get(pe), eeh_pe_loc_get(phb_pe));
 884	}
 885
 886#ifdef CONFIG_STACKTRACE
 887	/*
 888	 * Print the saved stack trace now that we've verified there's
 889	 * something to recover.
 890	 */
 891	if (pe->trace_entries) {
 892		void **ptrs = (void **) pe->stack_trace;
 893		int i;
 894
 895		pr_err("EEH: Frozen PHB#%x-PE#%x detected\n",
 896		       pe->phb->global_number, pe->addr);
 897
 898		/* FIXME: Use the same format as dump_stack() */
 899		pr_err("EEH: Call Trace:\n");
 900		for (i = 0; i < pe->trace_entries; i++)
 901			pr_err("EEH: [%pK] %pS\n", ptrs[i], ptrs[i]);
 902
 903		pe->trace_entries = 0;
 904	}
 905#endif /* CONFIG_STACKTRACE */
 906
 907	eeh_for_each_pe(pe, tmp_pe)
 908		eeh_pe_for_each_dev(tmp_pe, edev, tmp)
 909			edev->mode &= ~EEH_DEV_NO_HANDLER;
 910
 911	eeh_pe_update_time_stamp(pe);
 912	pe->freeze_count++;
 913	if (pe->freeze_count > eeh_max_freezes) {
 914		pr_err("EEH: PHB#%x-PE#%x has failed %d times in the last hour and has been permanently disabled.\n",
 
 915		       pe->phb->global_number, pe->addr,
 916		       pe->freeze_count);
 917
 918		goto recover_failed;
 919	}
 
 
 920
 921	/* Walk the various device drivers attached to this slot through
 922	 * a reset sequence, giving each an opportunity to do what it needs
 923	 * to accomplish the reset.  Each child gets a report of the
 924	 * status ... if any child can't handle the reset, then the entire
 925	 * slot is dlpar removed and added.
 926	 *
 927	 * When the PHB is fenced, we have to issue a reset to recover from
 928	 * the error. Override the result if necessary to have partially
 929	 * hotplug for this case.
 930	 */
 931	pr_warn("EEH: This PCI device has failed %d times in the last hour and will be permanently disabled after %d failures.\n",
 932		pe->freeze_count, eeh_max_freezes);
 933	pr_info("EEH: Notify device drivers to shutdown\n");
 934	eeh_set_channel_state(pe, pci_channel_io_frozen);
 935	eeh_set_irq_state(pe, false);
 936	eeh_pe_report("error_detected(IO frozen)", pe,
 937		      eeh_report_error, &result);
 938	if (result == PCI_ERS_RESULT_DISCONNECT)
 939		goto recover_failed;
 940
 941	/*
 942	 * Error logged on a PHB are always fences which need a full
 943	 * PHB reset to clear so force that to happen.
 944	 */
 945	if ((pe->type & EEH_PE_PHB) && result != PCI_ERS_RESULT_NONE)
 946		result = PCI_ERS_RESULT_NEED_RESET;
 947
 948	/* Get the current PCI slot state. This can take a long time,
 949	 * sometimes over 300 seconds for certain systems.
 950	 */
 951	rc = eeh_wait_state(pe, MAX_WAIT_FOR_RECOVERY * 1000);
 952	if (rc < 0 || rc == EEH_STATE_NOT_SUPPORT) {
 953		pr_warn("EEH: Permanent failure\n");
 954		goto recover_failed;
 955	}
 956
 957	/* Since rtas may enable MMIO when posting the error log,
 958	 * don't post the error log until after all dev drivers
 959	 * have been informed.
 960	 */
 961	pr_info("EEH: Collect temporary log\n");
 962	eeh_slot_error_detail(pe, EEH_LOG_TEMP);
 963
 964	/* If all device drivers were EEH-unaware, then shut
 965	 * down all of the device drivers, and hope they
 966	 * go down willingly, without panicing the system.
 967	 */
 968	if (result == PCI_ERS_RESULT_NONE) {
 969		pr_info("EEH: Reset with hotplug activity\n");
 970		rc = eeh_reset_device(pe, bus, NULL, false);
 971		if (rc) {
 972			pr_warn("%s: Unable to reset, err=%d\n", __func__, rc);
 973			goto recover_failed;
 
 974		}
 975	}
 976
 977	/* If all devices reported they can proceed, then re-enable MMIO */
 978	if (result == PCI_ERS_RESULT_CAN_RECOVER) {
 979		pr_info("EEH: Enable I/O for affected devices\n");
 980		rc = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
 
 981		if (rc < 0)
 982			goto recover_failed;
 983
 984		if (rc) {
 985			result = PCI_ERS_RESULT_NEED_RESET;
 986		} else {
 987			pr_info("EEH: Notify device drivers to resume I/O\n");
 988			eeh_pe_report("mmio_enabled", pe,
 989				      eeh_report_mmio_enabled, &result);
 990		}
 991	}
 
 
 992	if (result == PCI_ERS_RESULT_CAN_RECOVER) {
 993		pr_info("EEH: Enabled DMA for affected devices\n");
 994		rc = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
 
 995		if (rc < 0)
 996			goto recover_failed;
 997
 998		if (rc) {
 999			result = PCI_ERS_RESULT_NEED_RESET;
1000		} else {
1001			/*
1002			 * We didn't do PE reset for the case. The PE
1003			 * is still in frozen state. Clear it before
1004			 * resuming the PE.
1005			 */
1006			eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
1007			result = PCI_ERS_RESULT_RECOVERED;
1008		}
1009	}
1010
 
 
 
 
 
 
1011	/* If any device called out for a reset, then reset the slot */
1012	if (result == PCI_ERS_RESULT_NEED_RESET) {
1013		pr_info("EEH: Reset without hotplug activity\n");
1014		rc = eeh_reset_device(pe, bus, &rmv_data, true);
1015		if (rc) {
1016			pr_warn("%s: Cannot reset, err=%d\n", __func__, rc);
1017			goto recover_failed;
 
1018		}
1019
 
 
1020		result = PCI_ERS_RESULT_NONE;
1021		eeh_set_channel_state(pe, pci_channel_io_normal);
1022		eeh_set_irq_state(pe, true);
1023		eeh_pe_report("slot_reset", pe, eeh_report_reset,
1024			      &result);
 
 
 
 
1025	}
1026
1027	if ((result == PCI_ERS_RESULT_RECOVERED) ||
1028	    (result == PCI_ERS_RESULT_NONE)) {
1029		/*
1030		 * For those hot removed VFs, we should add back them after PF
1031		 * get recovered properly.
1032		 */
1033		list_for_each_entry_safe(edev, tmp, &rmv_data.removed_vf_list,
1034					 rmv_entry) {
1035			eeh_add_virt_device(edev);
1036			list_del(&edev->rmv_entry);
1037		}
1038
1039		/* Tell all device drivers that they can resume operations */
1040		pr_info("EEH: Notify device driver to resume\n");
1041		eeh_set_channel_state(pe, pci_channel_io_normal);
1042		eeh_set_irq_state(pe, true);
1043		eeh_pe_report("resume", pe, eeh_report_resume, NULL);
1044		eeh_for_each_pe(pe, tmp_pe) {
1045			eeh_pe_for_each_dev(tmp_pe, edev, tmp) {
1046				edev->mode &= ~EEH_DEV_NO_HANDLER;
1047				edev->in_error = false;
1048			}
1049		}
1050
1051		pr_info("EEH: Recovery successful.\n");
1052		goto out;
1053	}
1054
1055recover_failed:
1056	/*
1057	 * About 90% of all real-life EEH failures in the field
1058	 * are due to poorly seated PCI cards. Only 10% or so are
1059	 * due to actual, failed cards.
1060	 */
1061	pr_err("EEH: Unable to recover from failure from PHB#%x-PE#%x.\n"
1062		"Please try reseating or replacing it\n",
1063		pe->phb->global_number, pe->addr);
1064
1065	eeh_slot_error_detail(pe, EEH_LOG_PERM);
1066
1067	/* Notify all devices that they're about to go down. */
1068	eeh_set_irq_state(pe, false);
1069	eeh_pe_report("error_detected(permanent failure)", pe,
1070		      eeh_report_failure, NULL);
1071	eeh_set_channel_state(pe, pci_channel_io_perm_failure);
1072
1073	/* Mark the PE to be removed permanently */
1074	eeh_pe_state_mark(pe, EEH_PE_REMOVED);
1075
1076	/*
1077	 * Shut down the device drivers for good. We mark
1078	 * all removed devices correctly to avoid access
1079	 * the their PCI config any more.
1080	 */
1081	if (pe->type & EEH_PE_VF) {
1082		eeh_pe_dev_traverse(pe, eeh_rmv_device, NULL);
1083		eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
1084	} else {
1085		eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
1086		eeh_pe_dev_mode_mark(pe, EEH_DEV_REMOVED);
1087
1088		pci_lock_rescan_remove();
1089		pci_hp_remove_devices(bus);
1090		pci_unlock_rescan_remove();
1091		/* The passed PE should no longer be used */
1092		return;
1093	}
1094
1095out:
1096	/*
1097	 * Clean up any PEs without devices. While marked as EEH_PE_RECOVERYING
1098	 * we don't want to modify the PE tree structure so we do it here.
1099	 */
1100	eeh_pe_cleanup(pe);
1101
1102	/* clear the slot attention LED for all recovered devices */
1103	eeh_for_each_pe(pe, tmp_pe)
1104		eeh_pe_for_each_dev(tmp_pe, edev, tmp)
1105			eeh_clear_slot_attention(edev->pdev);
1106
1107	eeh_pe_state_clear(pe, EEH_PE_RECOVERING, true);
1108}
1109
1110/**
1111 * eeh_handle_special_event - Handle EEH events without a specific failing PE
1112 *
1113 * Called when an EEH event is detected but can't be narrowed down to a
1114 * specific PE.  Iterates through possible failures and handles them as
1115 * necessary.
1116 */
1117void eeh_handle_special_event(void)
1118{
1119	struct eeh_pe *pe, *phb_pe, *tmp_pe;
1120	struct eeh_dev *edev, *tmp_edev;
1121	struct pci_bus *bus;
1122	struct pci_controller *hose;
1123	unsigned long flags;
1124	int rc;
1125
1126
1127	do {
1128		rc = eeh_ops->next_error(&pe);
1129
1130		switch (rc) {
1131		case EEH_NEXT_ERR_DEAD_IOC:
1132			/* Mark all PHBs in dead state */
1133			eeh_serialize_lock(&flags);
1134
1135			/* Purge all events */
1136			eeh_remove_event(NULL, true);
1137
1138			list_for_each_entry(hose, &hose_list, list_node) {
1139				phb_pe = eeh_phb_pe_get(hose);
1140				if (!phb_pe) continue;
1141
1142				eeh_pe_mark_isolated(phb_pe);
1143			}
1144
1145			eeh_serialize_unlock(flags);
1146
1147			break;
1148		case EEH_NEXT_ERR_FROZEN_PE:
1149		case EEH_NEXT_ERR_FENCED_PHB:
1150		case EEH_NEXT_ERR_DEAD_PHB:
1151			/* Mark the PE in fenced state */
1152			eeh_serialize_lock(&flags);
1153
1154			/* Purge all events of the PHB */
1155			eeh_remove_event(pe, true);
1156
1157			if (rc != EEH_NEXT_ERR_DEAD_PHB)
1158				eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
1159			eeh_pe_mark_isolated(pe);
 
 
1160
1161			eeh_serialize_unlock(flags);
1162
1163			break;
1164		case EEH_NEXT_ERR_NONE:
1165			return;
1166		default:
1167			pr_warn("%s: Invalid value %d from next_error()\n",
1168				__func__, rc);
1169			return;
1170		}
1171
1172		/*
1173		 * For fenced PHB and frozen PE, it's handled as normal
1174		 * event. We have to remove the affected PHBs for dead
1175		 * PHB and IOC
1176		 */
1177		if (rc == EEH_NEXT_ERR_FROZEN_PE ||
1178		    rc == EEH_NEXT_ERR_FENCED_PHB) {
1179			eeh_pe_state_mark(pe, EEH_PE_RECOVERING);
1180			eeh_handle_normal_event(pe);
1181		} else {
1182			eeh_for_each_pe(pe, tmp_pe)
1183				eeh_pe_for_each_dev(tmp_pe, edev, tmp_edev)
1184					edev->mode &= ~EEH_DEV_NO_HANDLER;
1185
1186			/* Notify all devices to be down */
1187			eeh_pe_state_clear(pe, EEH_PE_PRI_BUS, true);
1188			eeh_pe_report(
1189				"error_detected(permanent failure)", pe,
1190				eeh_report_failure, NULL);
1191			eeh_set_channel_state(pe, pci_channel_io_perm_failure);
1192
1193			pci_lock_rescan_remove();
1194			list_for_each_entry(hose, &hose_list, list_node) {
1195				phb_pe = eeh_phb_pe_get(hose);
1196				if (!phb_pe ||
1197				    !(phb_pe->state & EEH_PE_ISOLATED) ||
1198				    (phb_pe->state & EEH_PE_RECOVERING))
1199					continue;
1200
 
 
 
 
1201				bus = eeh_pe_bus_get(phb_pe);
1202				if (!bus) {
1203					pr_err("%s: Cannot find PCI bus for "
1204					       "PHB#%x-PE#%x\n",
1205					       __func__,
1206					       pe->phb->global_number,
1207					       pe->addr);
1208					break;
1209				}
1210				pci_hp_remove_devices(bus);
1211			}
1212			pci_unlock_rescan_remove();
1213		}
1214
1215		/*
1216		 * If we have detected dead IOC, we needn't proceed
1217		 * any more since all PHBs would have been removed
1218		 */
1219		if (rc == EEH_NEXT_ERR_DEAD_IOC)
1220			break;
1221	} while (rc != EEH_NEXT_ERR_NONE);
1222}