eeh.c

来自「底层驱动开发」· C语言 代码 · 共 944 行 · 第 1/2 页

/*
 * eeh.c
 * Copyright (C) 2001 Dave Engebretsen & Todd Inglett IBM Corporation
 * 
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 * 
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 * 
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
 */

#include <linux/bootmem.h>
#include <linux/init.h>
#include <linux/list.h>
#include <linux/mm.h>
#include <linux/notifier.h>
#include <linux/pci.h>
#include <linux/proc_fs.h>
#include <linux/rbtree.h>
#include <linux/seq_file.h>
#include <linux/spinlock.h>
#include <asm/eeh.h>
#include <asm/io.h>
#include <asm/machdep.h>
#include <asm/rtas.h>
#include <asm/atomic.h>
#include <asm/systemcfg.h>
#include "pci.h"

#undef DEBUG

/** Overview:
 *  EEH, or "Extended Error Handling" is a PCI bridge technology for
 *  dealing with PCI bus errors that can't be dealt with within the
 *  usual PCI framework, except by check-stopping the CPU.  Systems
 *  that are designed for high-availability/reliability cannot afford
 *  to crash due to a "mere" PCI error, thus the need for EEH.
 *  An EEH-capable bridge operates by converting a detected error
 *  into a "slot freeze", taking the PCI adapter off-line, making
 *  the slot behave, from the OS'es point of view, as if the slot
 *  were "empty": all reads return 0xff's and all writes are silently
 *  ignored.  EEH slot isolation events can be triggered by parity
 *  errors on the address or data busses (e.g. during posted writes),
 *  which in turn might be caused by dust, vibration, humidity,
 *  radioactivity or plain-old failed hardware.
 *
 *  Note, however, that one of the leading causes of EEH slot
 *  freeze events are buggy device drivers, buggy device microcode,
 *  or buggy device hardware.  This is because any attempt by the
 *  device to bus-master data to a memory address that is not
 *  assigned to the device will trigger a slot freeze.   (The idea
 *  is to prevent devices-gone-wild from corrupting system memory).
 *  Buggy hardware/drivers will have a miserable time co-existing
 *  with EEH.
 *
 *  Ideally, a PCI device driver, when suspecting that an isolation
 *  event has occured (e.g. by reading 0xff's), will then ask EEH
 *  whether this is the case, and then take appropriate steps to
 *  reset the PCI slot, the PCI device, and then resume operations.
 *  However, until that day,  the checking is done here, with the
 *  eeh_check_failure() routine embedded in the MMIO macros.  If
 *  the slot is found to be isolated, an "EEH Event" is synthesized
 *  and sent out for processing.
 */

/** Bus Unit ID macros; get low and hi 32-bits of the 64-bit BUID */
#define BUID_HI(buid) ((buid) >> 32)
#define BUID_LO(buid) ((buid) & 0xffffffff)

/* EEH event workqueue setup. */
static DEFINE_SPINLOCK(eeh_eventlist_lock);
LIST_HEAD(eeh_eventlist);
static void eeh_event_handler(void *);
DECLARE_WORK(eeh_event_wq, eeh_event_handler, NULL);

static struct notifier_block *eeh_notifier_chain;

/*
 * If a device driver keeps reading an MMIO register in an interrupt
 * handler after a slot isolation event has occurred, we assume it
 * is broken and panic.  This sets the threshold for how many read
 * attempts we allow before panicking.
 */
#define EEH_MAX_FAILS	1000
static atomic_t eeh_fail_count;

/* RTAS tokens */
static int ibm_set_eeh_option;
static int ibm_set_slot_reset;
static int ibm_read_slot_reset_state;
static int ibm_read_slot_reset_state2;
static int ibm_slot_error_detail;

static int eeh_subsystem_enabled;

/* Buffer for reporting slot-error-detail rtas calls */
static unsigned char slot_errbuf[RTAS_ERROR_LOG_MAX];
static DEFINE_SPINLOCK(slot_errbuf_lock);
static int eeh_error_buf_size;

/* System monitoring statistics */
static DEFINE_PER_CPU(unsigned long, total_mmio_ffs);
static DEFINE_PER_CPU(unsigned long, false_positives);
static DEFINE_PER_CPU(unsigned long, ignored_failures);
static DEFINE_PER_CPU(unsigned long, slot_resets);

/**
 * The pci address cache subsystem.  This subsystem places
 * PCI device address resources into a red-black tree, sorted
 * according to the address range, so that given only an i/o
 * address, the corresponding PCI device can be **quickly**
 * found. It is safe to perform an address lookup in an interrupt
 * context; this ability is an important feature.
 *
 * Currently, the only customer of this code is the EEH subsystem;
 * thus, this code has been somewhat tailored to suit EEH better.
 * In particular, the cache does *not* hold the addresses of devices
 * for which EEH is not enabled.
 *
 * (Implementation Note: The RB tree seems to be better/faster
 * than any hash algo I could think of for this problem, even
 * with the penalty of slow pointer chases for d-cache misses).
 */
struct pci_io_addr_range
{
	struct rb_node rb_node;
	unsigned long addr_lo;
	unsigned long addr_hi;
	struct pci_dev *pcidev;
	unsigned int flags;
};

static struct pci_io_addr_cache
{
	struct rb_root rb_root;
	spinlock_t piar_lock;
} pci_io_addr_cache_root;

static inline struct pci_dev *__pci_get_device_by_addr(unsigned long addr)
{
	struct rb_node *n = pci_io_addr_cache_root.rb_root.rb_node;

	while (n) {
		struct pci_io_addr_range *piar;
		piar = rb_entry(n, struct pci_io_addr_range, rb_node);

		if (addr < piar->addr_lo) {
			n = n->rb_left;
		} else {
			if (addr > piar->addr_hi) {
				n = n->rb_right;
			} else {
				pci_dev_get(piar->pcidev);
				return piar->pcidev;
			}
		}
	}

	return NULL;
}

/**
 * pci_get_device_by_addr - Get device, given only address
 * @addr: mmio (PIO) phys address or i/o port number
 *
 * Given an mmio phys address, or a port number, find a pci device
 * that implements this address.  Be sure to pci_dev_put the device
 * when finished.  I/O port numbers are assumed to be offset
 * from zero (that is, they do *not* have pci_io_addr added in).
 * It is safe to call this function within an interrupt.
 */
static struct pci_dev *pci_get_device_by_addr(unsigned long addr)
{
	struct pci_dev *dev;
	unsigned long flags;

	spin_lock_irqsave(&pci_io_addr_cache_root.piar_lock, flags);
	dev = __pci_get_device_by_addr(addr);
	spin_unlock_irqrestore(&pci_io_addr_cache_root.piar_lock, flags);
	return dev;
}

#ifdef DEBUG
/*
 * Handy-dandy debug print routine, does nothing more
 * than print out the contents of our addr cache.
 */
static void pci_addr_cache_print(struct pci_io_addr_cache *cache)
{
	struct rb_node *n;
	int cnt = 0;

	n = rb_first(&cache->rb_root);
	while (n) {
		struct pci_io_addr_range *piar;
		piar = rb_entry(n, struct pci_io_addr_range, rb_node);
		printk(KERN_DEBUG "PCI: %s addr range %d [%lx-%lx]: %s\n",
		       (piar->flags & IORESOURCE_IO) ? "i/o" : "mem", cnt,
		       piar->addr_lo, piar->addr_hi, pci_name(piar->pcidev));
		cnt++;
		n = rb_next(n);
	}
}
#endif

/* Insert address range into the rb tree. */
static struct pci_io_addr_range *
pci_addr_cache_insert(struct pci_dev *dev, unsigned long alo,
		      unsigned long ahi, unsigned int flags)
{
	struct rb_node **p = &pci_io_addr_cache_root.rb_root.rb_node;
	struct rb_node *parent = NULL;
	struct pci_io_addr_range *piar;

	/* Walk tree, find a place to insert into tree */
	while (*p) {
		parent = *p;
		piar = rb_entry(parent, struct pci_io_addr_range, rb_node);
		if (alo < piar->addr_lo) {
			p = &parent->rb_left;
		} else if (ahi > piar->addr_hi) {
			p = &parent->rb_right;
		} else {
			if (dev != piar->pcidev ||
			    alo != piar->addr_lo || ahi != piar->addr_hi) {
				printk(KERN_WARNING "PIAR: overlapping address range\n");
			}
			return piar;
		}
	}
	piar = (struct pci_io_addr_range *)kmalloc(sizeof(struct pci_io_addr_range), GFP_ATOMIC);
	if (!piar)
		return NULL;

	piar->addr_lo = alo;
	piar->addr_hi = ahi;
	piar->pcidev = dev;
	piar->flags = flags;

	rb_link_node(&piar->rb_node, parent, p);
	rb_insert_color(&piar->rb_node, &pci_io_addr_cache_root.rb_root);

	return piar;
}

static void __pci_addr_cache_insert_device(struct pci_dev *dev)
{
	struct device_node *dn;
	struct pci_dn *pdn;
	int i;
	int inserted = 0;

	dn = pci_device_to_OF_node(dev);
	if (!dn) {
		printk(KERN_WARNING "PCI: no pci dn found for dev=%s\n",
			pci_name(dev));
		return;
	}

	/* Skip any devices for which EEH is not enabled. */
	pdn = dn->data;
	if (!(pdn->eeh_mode & EEH_MODE_SUPPORTED) ||
	    pdn->eeh_mode & EEH_MODE_NOCHECK) {
#ifdef DEBUG
		printk(KERN_INFO "PCI: skip building address cache for=%s\n",
		       pci_name(dev));
#endif
		return;
	}

	/* The cache holds a reference to the device... */
	pci_dev_get(dev);

	/* Walk resources on this device, poke them into the tree */
	for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
		unsigned long start = pci_resource_start(dev,i);
		unsigned long end = pci_resource_end(dev,i);
		unsigned int flags = pci_resource_flags(dev,i);

		/* We are interested only bus addresses, not dma or other stuff */
		if (0 == (flags & (IORESOURCE_IO | IORESOURCE_MEM)))
			continue;
		if (start == 0 || ~start == 0 || end == 0 || ~end == 0)
			 continue;
		pci_addr_cache_insert(dev, start, end, flags);
		inserted = 1;
	}

	/* If there was nothing to add, the cache has no reference... */
	if (!inserted)
		pci_dev_put(dev);
}

/**
 * pci_addr_cache_insert_device - Add a device to the address cache
 * @dev: PCI device whose I/O addresses we are interested in.
 *
 * In order to support the fast lookup of devices based on addresses,
 * we maintain a cache of devices that can be quickly searched.
 * This routine adds a device to that cache.
 */
void pci_addr_cache_insert_device(struct pci_dev *dev)
{
	unsigned long flags;

	spin_lock_irqsave(&pci_io_addr_cache_root.piar_lock, flags);
	__pci_addr_cache_insert_device(dev);
	spin_unlock_irqrestore(&pci_io_addr_cache_root.piar_lock, flags);
}

static inline void __pci_addr_cache_remove_device(struct pci_dev *dev)
{
	struct rb_node *n;
	int removed = 0;

restart:
	n = rb_first(&pci_io_addr_cache_root.rb_root);
	while (n) {
		struct pci_io_addr_range *piar;
		piar = rb_entry(n, struct pci_io_addr_range, rb_node);

		if (piar->pcidev == dev) {
			rb_erase(n, &pci_io_addr_cache_root.rb_root);
			removed = 1;
			kfree(piar);
			goto restart;
		}
		n = rb_next(n);
	}

	/* The cache no longer holds its reference to this device... */
	if (removed)
		pci_dev_put(dev);
}

/**
 * pci_addr_cache_remove_device - remove pci device from addr cache
 * @dev: device to remove
 *
 * Remove a device from the addr-cache tree.
 * This is potentially expensive, since it will walk
 * the tree multiple times (once per resource).
 * But so what; device removal doesn't need to be that fast.
 */
void pci_addr_cache_remove_device(struct pci_dev *dev)
{
	unsigned long flags;

	spin_lock_irqsave(&pci_io_addr_cache_root.piar_lock, flags);
	__pci_addr_cache_remove_device(dev);
	spin_unlock_irqrestore(&pci_io_addr_cache_root.piar_lock, flags);
}

/**
 * pci_addr_cache_build - Build a cache of I/O addresses
 *
 * Build a cache of pci i/o addresses.  This cache will be used to
 * find the pci device that corresponds to a given address.
 * This routine scans all pci busses to build the cache.
 * Must be run late in boot process, after the pci controllers
 * have been scaned for devices (after all device resources are known).
 */
void __init pci_addr_cache_build(void)
{
	struct pci_dev *dev = NULL;

	spin_lock_init(&pci_io_addr_cache_root.piar_lock);

	while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
		/* Ignore PCI bridges ( XXX why ??) */
		if ((dev->class >> 16) == PCI_BASE_CLASS_BRIDGE) {
			continue;
		}
		pci_addr_cache_insert_device(dev);
	}

#ifdef DEBUG
	/* Verify tree built up above, echo back the list of addrs. */
	pci_addr_cache_print(&pci_io_addr_cache_root);
#endif
}

/* --------------------------------------------------------------- */
/* Above lies the PCI Address Cache. Below lies the EEH event infrastructure */

/**
 * eeh_register_notifier - Register to find out about EEH events.
 * @nb: notifier block to callback on events
 */
int eeh_register_notifier(struct notifier_block *nb)
{
	return notifier_chain_register(&eeh_notifier_chain, nb);
}

/**
 * eeh_unregister_notifier - Unregister to an EEH event notifier.
 * @nb: notifier block to callback on events
 */
int eeh_unregister_notifier(struct notifier_block *nb)
{
	return notifier_chain_unregister(&eeh_notifier_chain, nb);
}

/**
 * read_slot_reset_state - Read the reset state of a device node's slot
 * @dn: device node to read
 * @rets: array to return results in
 */
static int read_slot_reset_state(struct device_node *dn, int rets[])
{
	int token, outputs;
	struct pci_dn *pdn = dn->data;

	if (ibm_read_slot_reset_state2 != RTAS_UNKNOWN_SERVICE) {
		token = ibm_read_slot_reset_state2;
		outputs = 4;
	} else {
		token = ibm_read_slot_reset_state;
		outputs = 3;
	}

	return rtas_call(token, 3, outputs, rets, pdn->eeh_config_addr,
			 BUID_HI(pdn->phb->buid), BUID_LO(pdn->phb->buid));
}

/**
 * eeh_panic - call panic() for an eeh event that cannot be handled.
 * The philosophy of this routine is that it is better to panic and
 * halt the OS than it is to risk possible data corruption by
 * oblivious device drivers that don't know better.
 *
 * @dev pci device that had an eeh event
 * @reset_state current reset state of the device slot
 */
static void eeh_panic(struct pci_dev *dev, int reset_state)
{
	/*
	 * XXX We should create a separate sysctl for this.
	 *
	 * Since the panic_on_oops sysctl is used to halt the system
	 * in light of potential corruption, we can use it here.
	 */
	if (panic_on_oops)
		panic("EEH: MMIO failure (%d) on device:%s\n", reset_state,
		      pci_name(dev));
	else {
		__get_cpu_var(ignored_failures)++;
		printk(KERN_INFO "EEH: Ignored MMIO failure (%d) on device:%s\n",
		       reset_state, pci_name(dev));
	}
}

/**
 * eeh_event_handler - dispatch EEH events.  The detection of a frozen
 * slot can occur inside an interrupt, where it can be hard to do
 * anything about it.  The goal of this routine is to pull these
 * detection events out of the context of the interrupt handler, and
 * re-dispatch them for processing at a later time in a normal context.
 *
 * @dummy - unused
 */
static void eeh_event_handler(void *dummy)
{
	unsigned long flags;