linux_dsm_epyc7002/fs/lockd/host.c

641 lines
16 KiB
C
Raw Normal View History

/*
* linux/fs/lockd/host.c
*
* Management for NLM peer hosts. The nlm_host struct is shared
* between client and server implementation. The only reason to
* do so is to reduce code bloat.
*
* Copyright (C) 1996, Olaf Kirch <okir@monad.swb.de>
*/
#include <linux/types.h>
#include <linux/slab.h>
#include <linux/in.h>
#include <linux/in6.h>
#include <linux/sunrpc/clnt.h>
#include <linux/sunrpc/svc.h>
#include <linux/lockd/lockd.h>
#include <linux/mutex.h>
#include <linux/sunrpc/svc_xprt.h>
#include <net/ipv6.h>
#define NLMDBG_FACILITY NLMDBG_HOSTCACHE
#define NLM_HOST_NRHASH 32
#define NLM_HOST_REBIND (60 * HZ)
#define NLM_HOST_EXPIRE (300 * HZ)
#define NLM_HOST_COLLECT (120 * HZ)
static struct hlist_head nlm_server_hosts[NLM_HOST_NRHASH];
lockd: Create client-side nlm_host cache NFS clients don't need the garbage collection processing that is performed on nlm_host structures. The client picks up an nlm_host at mount time and holds a reference to it until the file system is unmounted. Servers, on the other hand, don't have a precise way to tell when an nlm_host is no longer being used, so zero refcount nlm_host entries are left to expire in the cache after a time. Basically there's nothing holding a reference to an nlm_host between individual server-side NLM requests, but we can't afford the expense of recreating them for every new NLM request from a client. The nlm_host cache adds some lifetime hysteresis to entries in the cache so the next time a particular nlm_host is needed, it's likely to be discovered by a lookup rather than created from whole cloth. With the new implementation, client nlm_host cache items are no longer garbage collected, and are destroyed directly by a new release function specialized for client entries, nlmclnt_release_host(). They are cached in their own data structure, and have their own lookup logic, simplified and specialized for client nlm_host entries. However, the client nlm_host cache still shares reboot recovery logic with the server nlm_host cache. The NSM "peer rebooted" downcall for clients and servers still come through the same RPC call. This is a legacy formal API that would be difficult to alter, and besides, the user space NSM implementation can't tell the difference between peers that are clients or servers. For this reason, the client cache continues to share the nlm_host_mutex (and reboot recovery logic) with the server cache. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2010-12-14 22:05:52 +07:00
static struct hlist_head nlm_client_hosts[NLM_HOST_NRHASH];
#define for_each_host(host, pos, chain, table) \
for ((chain) = (table); \
(chain) < (table) + NLM_HOST_NRHASH; ++(chain)) \
hlist_for_each_entry((host), (pos), (chain), h_hash)
#define for_each_host_safe(host, pos, next, chain, table) \
for ((chain) = (table); \
(chain) < (table) + NLM_HOST_NRHASH; ++(chain)) \
hlist_for_each_entry_safe((host), (pos), (next), \
(chain), h_hash)
static unsigned long next_gc;
static unsigned long nrhosts;
static DEFINE_MUTEX(nlm_host_mutex);
static void nlm_gc_hosts(void);
struct nlm_lookup_host_info {
const int server; /* search for server|client */
const struct sockaddr *sap; /* address to search for */
const size_t salen; /* it's length */
const unsigned short protocol; /* transport to search for*/
const u32 version; /* NLM version to search for */
const char *hostname; /* remote's hostname */
const size_t hostname_len; /* it's length */
const int noresvport; /* use non-priv port */
struct net *net; /* network namespace to bind */
};
/*
* Hash function must work well on big- and little-endian platforms
*/
static unsigned int __nlm_hash32(const __be32 n)
{
unsigned int hash = (__force u32)n ^ ((__force u32)n >> 16);
return hash ^ (hash >> 8);
}
static unsigned int __nlm_hash_addr4(const struct sockaddr *sap)
{
const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
return __nlm_hash32(sin->sin_addr.s_addr);
}
static unsigned int __nlm_hash_addr6(const struct sockaddr *sap)
{
const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
const struct in6_addr addr = sin6->sin6_addr;
return __nlm_hash32(addr.s6_addr32[0]) ^
__nlm_hash32(addr.s6_addr32[1]) ^
__nlm_hash32(addr.s6_addr32[2]) ^
__nlm_hash32(addr.s6_addr32[3]);
}
static unsigned int nlm_hash_address(const struct sockaddr *sap)
{
unsigned int hash;
switch (sap->sa_family) {
case AF_INET:
hash = __nlm_hash_addr4(sap);
break;
case AF_INET6:
hash = __nlm_hash_addr6(sap);
break;
default:
hash = 0;
}
return hash & (NLM_HOST_NRHASH - 1);
}
/*
* Allocate and initialize an nlm_host. Common to both client and server.
*/
static struct nlm_host *nlm_alloc_host(struct nlm_lookup_host_info *ni,
struct nsm_handle *nsm)
{
struct nlm_host *host = NULL;
unsigned long now = jiffies;
if (nsm != NULL)
atomic_inc(&nsm->sm_count);
else {
host = NULL;
nsm = nsm_get_handle(ni->sap, ni->salen,
ni->hostname, ni->hostname_len);
if (unlikely(nsm == NULL)) {
dprintk("lockd: %s failed; no nsm handle\n",
__func__);
goto out;
}
}
host = kmalloc(sizeof(*host), GFP_KERNEL);
if (unlikely(host == NULL)) {
dprintk("lockd: %s failed; no memory\n", __func__);
nsm_release(nsm);
goto out;
}
memcpy(nlm_addr(host), ni->sap, ni->salen);
host->h_addrlen = ni->salen;
rpc_set_port(nlm_addr(host), 0);
host->h_srcaddrlen = 0;
host->h_rpcclnt = NULL;
host->h_name = nsm->sm_name;
host->h_version = ni->version;
host->h_proto = ni->protocol;
host->h_reclaiming = 0;
host->h_server = ni->server;
host->h_noresvport = ni->noresvport;
host->h_inuse = 0;
init_waitqueue_head(&host->h_gracewait);
init_rwsem(&host->h_rwsem);
host->h_state = 0;
host->h_nsmstate = 0;
host->h_pidcount = 0;
atomic_set(&host->h_count, 1);
mutex_init(&host->h_mutex);
host->h_nextrebind = now + NLM_HOST_REBIND;
host->h_expires = now + NLM_HOST_EXPIRE;
INIT_LIST_HEAD(&host->h_lockowners);
spin_lock_init(&host->h_lock);
INIT_LIST_HEAD(&host->h_granted);
INIT_LIST_HEAD(&host->h_reclaim);
host->h_nsmhandle = nsm;
host->h_addrbuf = nsm->sm_addrbuf;
host->net = ni->net;
out:
return host;
}
/*
* Destroy an nlm_host and free associated resources
*
* Caller must hold nlm_host_mutex.
*/
static void nlm_destroy_host_locked(struct nlm_host *host)
{
struct rpc_clnt *clnt;
dprintk("lockd: destroy host %s\n", host->h_name);
BUG_ON(!list_empty(&host->h_lockowners));
BUG_ON(atomic_read(&host->h_count));
hlist_del_init(&host->h_hash);
nsm_unmonitor(host);
nsm_release(host->h_nsmhandle);
clnt = host->h_rpcclnt;
if (clnt != NULL)
rpc_shutdown_client(clnt);
kfree(host);
nrhosts--;
}
/**
* nlmclnt_lookup_host - Find an NLM host handle matching a remote server
* @sap: network address of server
* @salen: length of server address
* @protocol: transport protocol to use
* @version: NLM protocol version
* @hostname: '\0'-terminated hostname of server
* @noresvport: 1 if non-privileged port should be used
*
* Returns an nlm_host structure that matches the passed-in
* [server address, transport protocol, NLM version, server hostname].
* If one doesn't already exist in the host cache, a new handle is
* created and returned.
*/
struct nlm_host *nlmclnt_lookup_host(const struct sockaddr *sap,
const size_t salen,
const unsigned short protocol,
const u32 version,
const char *hostname,
int noresvport,
struct net *net)
{
struct nlm_lookup_host_info ni = {
.server = 0,
.sap = sap,
.salen = salen,
.protocol = protocol,
.version = version,
.hostname = hostname,
.hostname_len = strlen(hostname),
.noresvport = noresvport,
.net = net,
};
lockd: Create client-side nlm_host cache NFS clients don't need the garbage collection processing that is performed on nlm_host structures. The client picks up an nlm_host at mount time and holds a reference to it until the file system is unmounted. Servers, on the other hand, don't have a precise way to tell when an nlm_host is no longer being used, so zero refcount nlm_host entries are left to expire in the cache after a time. Basically there's nothing holding a reference to an nlm_host between individual server-side NLM requests, but we can't afford the expense of recreating them for every new NLM request from a client. The nlm_host cache adds some lifetime hysteresis to entries in the cache so the next time a particular nlm_host is needed, it's likely to be discovered by a lookup rather than created from whole cloth. With the new implementation, client nlm_host cache items are no longer garbage collected, and are destroyed directly by a new release function specialized for client entries, nlmclnt_release_host(). They are cached in their own data structure, and have their own lookup logic, simplified and specialized for client nlm_host entries. However, the client nlm_host cache still shares reboot recovery logic with the server nlm_host cache. The NSM "peer rebooted" downcall for clients and servers still come through the same RPC call. This is a legacy formal API that would be difficult to alter, and besides, the user space NSM implementation can't tell the difference between peers that are clients or servers. For this reason, the client cache continues to share the nlm_host_mutex (and reboot recovery logic) with the server cache. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2010-12-14 22:05:52 +07:00
struct hlist_head *chain;
struct hlist_node *pos;
struct nlm_host *host;
struct nsm_handle *nsm = NULL;
dprintk("lockd: %s(host='%s', vers=%u, proto=%s)\n", __func__,
(hostname ? hostname : "<none>"), version,
(protocol == IPPROTO_UDP ? "udp" : "tcp"));
lockd: Create client-side nlm_host cache NFS clients don't need the garbage collection processing that is performed on nlm_host structures. The client picks up an nlm_host at mount time and holds a reference to it until the file system is unmounted. Servers, on the other hand, don't have a precise way to tell when an nlm_host is no longer being used, so zero refcount nlm_host entries are left to expire in the cache after a time. Basically there's nothing holding a reference to an nlm_host between individual server-side NLM requests, but we can't afford the expense of recreating them for every new NLM request from a client. The nlm_host cache adds some lifetime hysteresis to entries in the cache so the next time a particular nlm_host is needed, it's likely to be discovered by a lookup rather than created from whole cloth. With the new implementation, client nlm_host cache items are no longer garbage collected, and are destroyed directly by a new release function specialized for client entries, nlmclnt_release_host(). They are cached in their own data structure, and have their own lookup logic, simplified and specialized for client nlm_host entries. However, the client nlm_host cache still shares reboot recovery logic with the server nlm_host cache. The NSM "peer rebooted" downcall for clients and servers still come through the same RPC call. This is a legacy formal API that would be difficult to alter, and besides, the user space NSM implementation can't tell the difference between peers that are clients or servers. For this reason, the client cache continues to share the nlm_host_mutex (and reboot recovery logic) with the server cache. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2010-12-14 22:05:52 +07:00
mutex_lock(&nlm_host_mutex);
chain = &nlm_client_hosts[nlm_hash_address(sap)];
hlist_for_each_entry(host, pos, chain, h_hash) {
if (host->net != net)
continue;
lockd: Create client-side nlm_host cache NFS clients don't need the garbage collection processing that is performed on nlm_host structures. The client picks up an nlm_host at mount time and holds a reference to it until the file system is unmounted. Servers, on the other hand, don't have a precise way to tell when an nlm_host is no longer being used, so zero refcount nlm_host entries are left to expire in the cache after a time. Basically there's nothing holding a reference to an nlm_host between individual server-side NLM requests, but we can't afford the expense of recreating them for every new NLM request from a client. The nlm_host cache adds some lifetime hysteresis to entries in the cache so the next time a particular nlm_host is needed, it's likely to be discovered by a lookup rather than created from whole cloth. With the new implementation, client nlm_host cache items are no longer garbage collected, and are destroyed directly by a new release function specialized for client entries, nlmclnt_release_host(). They are cached in their own data structure, and have their own lookup logic, simplified and specialized for client nlm_host entries. However, the client nlm_host cache still shares reboot recovery logic with the server nlm_host cache. The NSM "peer rebooted" downcall for clients and servers still come through the same RPC call. This is a legacy formal API that would be difficult to alter, and besides, the user space NSM implementation can't tell the difference between peers that are clients or servers. For this reason, the client cache continues to share the nlm_host_mutex (and reboot recovery logic) with the server cache. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2010-12-14 22:05:52 +07:00
if (!rpc_cmp_addr(nlm_addr(host), sap))
continue;
/* Same address. Share an NSM handle if we already have one */
if (nsm == NULL)
nsm = host->h_nsmhandle;
if (host->h_proto != protocol)
continue;
if (host->h_version != version)
continue;
nlm_get_host(host);
dprintk("lockd: %s found host %s (%s)\n", __func__,
host->h_name, host->h_addrbuf);
goto out;
}
host = nlm_alloc_host(&ni, nsm);
if (unlikely(host == NULL))
goto out;
hlist_add_head(&host->h_hash, chain);
nrhosts++;
dprintk("lockd: %s created host %s (%s)\n", __func__,
host->h_name, host->h_addrbuf);
out:
mutex_unlock(&nlm_host_mutex);
return host;
}
/**
* nlmclnt_release_host - release client nlm_host
* @host: nlm_host to release
*
*/
void nlmclnt_release_host(struct nlm_host *host)
{
if (host == NULL)
return;
dprintk("lockd: release client host %s\n", host->h_name);
BUG_ON(atomic_read(&host->h_count) < 0);
BUG_ON(host->h_server);
if (atomic_dec_and_test(&host->h_count)) {
BUG_ON(!list_empty(&host->h_lockowners));
BUG_ON(!list_empty(&host->h_granted));
BUG_ON(!list_empty(&host->h_reclaim));
mutex_lock(&nlm_host_mutex);
nlm_destroy_host_locked(host);
mutex_unlock(&nlm_host_mutex);
}
}
/**
* nlmsvc_lookup_host - Find an NLM host handle matching a remote client
* @rqstp: incoming NLM request
* @hostname: name of client host
* @hostname_len: length of client hostname
*
* Returns an nlm_host structure that matches the [client address,
* transport protocol, NLM version, client hostname] of the passed-in
* NLM request. If one doesn't already exist in the host cache, a
* new handle is created and returned.
*
* Before possibly creating a new nlm_host, construct a sockaddr
* for a specific source address in case the local system has
* multiple network addresses. The family of the address in
* rq_daddr is guaranteed to be the same as the family of the
* address in rq_addr, so it's safe to use the same family for
* the source address.
*/
struct nlm_host *nlmsvc_lookup_host(const struct svc_rqst *rqstp,
const char *hostname,
const size_t hostname_len)
{
struct hlist_head *chain;
struct hlist_node *pos;
struct nlm_host *host = NULL;
struct nsm_handle *nsm = NULL;
struct sockaddr *src_sap = svc_daddr(rqstp);
size_t src_len = rqstp->rq_daddrlen;
struct net *net = rqstp->rq_xprt->xpt_net;
struct nlm_lookup_host_info ni = {
.server = 1,
.sap = svc_addr(rqstp),
.salen = rqstp->rq_addrlen,
.protocol = rqstp->rq_prot,
.version = rqstp->rq_vers,
.hostname = hostname,
.hostname_len = hostname_len,
.net = net,
};
dprintk("lockd: %s(host='%*s', vers=%u, proto=%s)\n", __func__,
(int)hostname_len, hostname, rqstp->rq_vers,
(rqstp->rq_prot == IPPROTO_UDP ? "udp" : "tcp"));
mutex_lock(&nlm_host_mutex);
if (time_after_eq(jiffies, next_gc))
nlm_gc_hosts();
chain = &nlm_server_hosts[nlm_hash_address(ni.sap)];
hlist_for_each_entry(host, pos, chain, h_hash) {
if (host->net != net)
continue;
if (!rpc_cmp_addr(nlm_addr(host), ni.sap))
continue;
/* Same address. Share an NSM handle if we already have one */
if (nsm == NULL)
nsm = host->h_nsmhandle;
if (host->h_proto != ni.protocol)
continue;
if (host->h_version != ni.version)
continue;
if (!rpc_cmp_addr(nlm_srcaddr(host), src_sap))
continue;
/* Move to head of hash chain. */
hlist_del(&host->h_hash);
hlist_add_head(&host->h_hash, chain);
nlm_get_host(host);
dprintk("lockd: %s found host %s (%s)\n",
__func__, host->h_name, host->h_addrbuf);
goto out;
}
host = nlm_alloc_host(&ni, nsm);
if (unlikely(host == NULL))
goto out;
memcpy(nlm_srcaddr(host), src_sap, src_len);
host->h_srcaddrlen = src_len;
hlist_add_head(&host->h_hash, chain);
nrhosts++;
dprintk("lockd: %s created host %s (%s)\n",
__func__, host->h_name, host->h_addrbuf);
out:
mutex_unlock(&nlm_host_mutex);
return host;
}
/**
* nlmsvc_release_host - release server nlm_host
* @host: nlm_host to release
*
* Host is destroyed later in nlm_gc_host().
*/
void nlmsvc_release_host(struct nlm_host *host)
{
if (host == NULL)
return;
dprintk("lockd: release server host %s\n", host->h_name);
BUG_ON(atomic_read(&host->h_count) < 0);
BUG_ON(!host->h_server);
atomic_dec(&host->h_count);
}
/*
* Create the NLM RPC client for an NLM peer
*/
struct rpc_clnt *
nlm_bind_host(struct nlm_host *host)
{
struct rpc_clnt *clnt;
dprintk("lockd: nlm_bind_host %s (%s)\n",
host->h_name, host->h_addrbuf);
/* Lock host handle */
mutex_lock(&host->h_mutex);
/* If we've already created an RPC client, check whether
* RPC rebind is required
*/
if ((clnt = host->h_rpcclnt) != NULL) {
if (time_after_eq(jiffies, host->h_nextrebind)) {
rpc_force_rebind(clnt);
host->h_nextrebind = jiffies + NLM_HOST_REBIND;
dprintk("lockd: next rebind in %lu jiffies\n",
host->h_nextrebind - jiffies);
}
} else {
unsigned long increment = nlmsvc_timeout;
struct rpc_timeout timeparms = {
.to_initval = increment,
.to_increment = increment,
.to_maxval = increment * 6UL,
.to_retries = 5U,
};
struct rpc_create_args args = {
.net = host->net,
.protocol = host->h_proto,
.address = nlm_addr(host),
.addrsize = host->h_addrlen,
.timeout = &timeparms,
.servername = host->h_name,
.program = &nlm_program,
.version = host->h_version,
.authflavor = RPC_AUTH_UNIX,
.flags = (RPC_CLNT_CREATE_NOPING |
RPC_CLNT_CREATE_AUTOBIND),
};
/*
* lockd retries server side blocks automatically so we want
* those to be soft RPC calls. Client side calls need to be
* hard RPC tasks.
*/
if (!host->h_server)
args.flags |= RPC_CLNT_CREATE_HARDRTRY;
if (host->h_noresvport)
args.flags |= RPC_CLNT_CREATE_NONPRIVPORT;
if (host->h_srcaddrlen)
args.saddress = nlm_srcaddr(host);
clnt = rpc_create(&args);
if (!IS_ERR(clnt))
host->h_rpcclnt = clnt;
else {
printk("lockd: couldn't create RPC handle for %s\n", host->h_name);
clnt = NULL;
}
}
mutex_unlock(&host->h_mutex);
return clnt;
}
/*
* Force a portmap lookup of the remote lockd port
*/
void
nlm_rebind_host(struct nlm_host *host)
{
dprintk("lockd: rebind host %s\n", host->h_name);
if (host->h_rpcclnt && time_after_eq(jiffies, host->h_nextrebind)) {
rpc_force_rebind(host->h_rpcclnt);
host->h_nextrebind = jiffies + NLM_HOST_REBIND;
}
}
/*
* Increment NLM host count
*/
struct nlm_host * nlm_get_host(struct nlm_host *host)
{
if (host) {
dprintk("lockd: get host %s\n", host->h_name);
atomic_inc(&host->h_count);
host->h_expires = jiffies + NLM_HOST_EXPIRE;
}
return host;
}
static struct nlm_host *next_host_state(struct hlist_head *cache,
struct nsm_handle *nsm,
const struct nlm_reboot *info)
{
NLM: Fix "kernel BUG at fs/lockd/host.c:417!" or ".../host.c:283!" Nick Bowler <nbowler@elliptictech.com> reports: > We were just having some NFS server troubles, and my client machine > running 2.6.38-rc1+ (specifically, commit 2b1caf6ed7b888c95) crashed > hard (syslog output appended to this mail). > > I'm not sure what the exact timeline was or how to reproduce this, > but the server was rebooted during all this. Since I've never seen > this happen before, it is possibly a regression from previous kernel > releases. However, I recently updated my nfs-utils (on the client) to > version 1.2.3, so that might be related as well. [ BUG output redacted ] When done searching, the for_each_host loop in next_host_state() falls through and returns the final host on the host chain without bumping it's reference count. Since the host's ref count is only one at that point, releasing the host in nlm_host_rebooted() attempts to destroy the host prematurely, and therefore hits a BUG(). Likely, the original intent of the for_each_host behavior in next_host_state() was to handle the case when the host chain is empty. Searching the chain and finding no suitable host to return needs to be handled as well. Defensively restructure next_host_state() always to return NULL when the loop falls through. Introduced by commit b10e30f6 "lockd: reorganize nlm_host_rebooted". Cc: J. Bruce Fields <bfields@fieldses.org> Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2011-01-25 03:50:26 +07:00
struct nlm_host *host;
struct hlist_head *chain;
struct hlist_node *pos;
mutex_lock(&nlm_host_mutex);
for_each_host(host, pos, chain, cache) {
if (host->h_nsmhandle == nsm
&& host->h_nsmstate != info->state) {
host->h_nsmstate = info->state;
host->h_state++;
nlm_get_host(host);
NLM: Fix "kernel BUG at fs/lockd/host.c:417!" or ".../host.c:283!" Nick Bowler <nbowler@elliptictech.com> reports: > We were just having some NFS server troubles, and my client machine > running 2.6.38-rc1+ (specifically, commit 2b1caf6ed7b888c95) crashed > hard (syslog output appended to this mail). > > I'm not sure what the exact timeline was or how to reproduce this, > but the server was rebooted during all this. Since I've never seen > this happen before, it is possibly a regression from previous kernel > releases. However, I recently updated my nfs-utils (on the client) to > version 1.2.3, so that might be related as well. [ BUG output redacted ] When done searching, the for_each_host loop in next_host_state() falls through and returns the final host on the host chain without bumping it's reference count. Since the host's ref count is only one at that point, releasing the host in nlm_host_rebooted() attempts to destroy the host prematurely, and therefore hits a BUG(). Likely, the original intent of the for_each_host behavior in next_host_state() was to handle the case when the host chain is empty. Searching the chain and finding no suitable host to return needs to be handled as well. Defensively restructure next_host_state() always to return NULL when the loop falls through. Introduced by commit b10e30f6 "lockd: reorganize nlm_host_rebooted". Cc: J. Bruce Fields <bfields@fieldses.org> Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2011-01-25 03:50:26 +07:00
mutex_unlock(&nlm_host_mutex);
return host;
}
}
NLM: Fix "kernel BUG at fs/lockd/host.c:417!" or ".../host.c:283!" Nick Bowler <nbowler@elliptictech.com> reports: > We were just having some NFS server troubles, and my client machine > running 2.6.38-rc1+ (specifically, commit 2b1caf6ed7b888c95) crashed > hard (syslog output appended to this mail). > > I'm not sure what the exact timeline was or how to reproduce this, > but the server was rebooted during all this. Since I've never seen > this happen before, it is possibly a regression from previous kernel > releases. However, I recently updated my nfs-utils (on the client) to > version 1.2.3, so that might be related as well. [ BUG output redacted ] When done searching, the for_each_host loop in next_host_state() falls through and returns the final host on the host chain without bumping it's reference count. Since the host's ref count is only one at that point, releasing the host in nlm_host_rebooted() attempts to destroy the host prematurely, and therefore hits a BUG(). Likely, the original intent of the for_each_host behavior in next_host_state() was to handle the case when the host chain is empty. Searching the chain and finding no suitable host to return needs to be handled as well. Defensively restructure next_host_state() always to return NULL when the loop falls through. Introduced by commit b10e30f6 "lockd: reorganize nlm_host_rebooted". Cc: J. Bruce Fields <bfields@fieldses.org> Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2011-01-25 03:50:26 +07:00
mutex_unlock(&nlm_host_mutex);
NLM: Fix "kernel BUG at fs/lockd/host.c:417!" or ".../host.c:283!" Nick Bowler <nbowler@elliptictech.com> reports: > We were just having some NFS server troubles, and my client machine > running 2.6.38-rc1+ (specifically, commit 2b1caf6ed7b888c95) crashed > hard (syslog output appended to this mail). > > I'm not sure what the exact timeline was or how to reproduce this, > but the server was rebooted during all this. Since I've never seen > this happen before, it is possibly a regression from previous kernel > releases. However, I recently updated my nfs-utils (on the client) to > version 1.2.3, so that might be related as well. [ BUG output redacted ] When done searching, the for_each_host loop in next_host_state() falls through and returns the final host on the host chain without bumping it's reference count. Since the host's ref count is only one at that point, releasing the host in nlm_host_rebooted() attempts to destroy the host prematurely, and therefore hits a BUG(). Likely, the original intent of the for_each_host behavior in next_host_state() was to handle the case when the host chain is empty. Searching the chain and finding no suitable host to return needs to be handled as well. Defensively restructure next_host_state() always to return NULL when the loop falls through. Introduced by commit b10e30f6 "lockd: reorganize nlm_host_rebooted". Cc: J. Bruce Fields <bfields@fieldses.org> Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2011-01-25 03:50:26 +07:00
return NULL;
}
/**
* nlm_host_rebooted - Release all resources held by rebooted host
* @info: pointer to decoded results of NLM_SM_NOTIFY call
*
* We were notified that the specified host has rebooted. Release
* all resources held by that peer.
*/
void nlm_host_rebooted(const struct nlm_reboot *info)
{
struct nsm_handle *nsm;
struct nlm_host *host;
nsm = nsm_reboot_lookup(info);
if (unlikely(nsm == NULL))
return;
/* Mark all hosts tied to this NSM state as having rebooted.
* We run the loop repeatedly, because we drop the host table
* lock for this.
* To avoid processing a host several times, we match the nsmstate.
*/
while ((host = next_host_state(nlm_server_hosts, nsm, info)) != NULL) {
lockd: Create client-side nlm_host cache NFS clients don't need the garbage collection processing that is performed on nlm_host structures. The client picks up an nlm_host at mount time and holds a reference to it until the file system is unmounted. Servers, on the other hand, don't have a precise way to tell when an nlm_host is no longer being used, so zero refcount nlm_host entries are left to expire in the cache after a time. Basically there's nothing holding a reference to an nlm_host between individual server-side NLM requests, but we can't afford the expense of recreating them for every new NLM request from a client. The nlm_host cache adds some lifetime hysteresis to entries in the cache so the next time a particular nlm_host is needed, it's likely to be discovered by a lookup rather than created from whole cloth. With the new implementation, client nlm_host cache items are no longer garbage collected, and are destroyed directly by a new release function specialized for client entries, nlmclnt_release_host(). They are cached in their own data structure, and have their own lookup logic, simplified and specialized for client nlm_host entries. However, the client nlm_host cache still shares reboot recovery logic with the server nlm_host cache. The NSM "peer rebooted" downcall for clients and servers still come through the same RPC call. This is a legacy formal API that would be difficult to alter, and besides, the user space NSM implementation can't tell the difference between peers that are clients or servers. For this reason, the client cache continues to share the nlm_host_mutex (and reboot recovery logic) with the server cache. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2010-12-14 22:05:52 +07:00
nlmsvc_free_host_resources(host);
nlmsvc_release_host(host);
}
lockd: Create client-side nlm_host cache NFS clients don't need the garbage collection processing that is performed on nlm_host structures. The client picks up an nlm_host at mount time and holds a reference to it until the file system is unmounted. Servers, on the other hand, don't have a precise way to tell when an nlm_host is no longer being used, so zero refcount nlm_host entries are left to expire in the cache after a time. Basically there's nothing holding a reference to an nlm_host between individual server-side NLM requests, but we can't afford the expense of recreating them for every new NLM request from a client. The nlm_host cache adds some lifetime hysteresis to entries in the cache so the next time a particular nlm_host is needed, it's likely to be discovered by a lookup rather than created from whole cloth. With the new implementation, client nlm_host cache items are no longer garbage collected, and are destroyed directly by a new release function specialized for client entries, nlmclnt_release_host(). They are cached in their own data structure, and have their own lookup logic, simplified and specialized for client nlm_host entries. However, the client nlm_host cache still shares reboot recovery logic with the server nlm_host cache. The NSM "peer rebooted" downcall for clients and servers still come through the same RPC call. This is a legacy formal API that would be difficult to alter, and besides, the user space NSM implementation can't tell the difference between peers that are clients or servers. For this reason, the client cache continues to share the nlm_host_mutex (and reboot recovery logic) with the server cache. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
2010-12-14 22:05:52 +07:00
while ((host = next_host_state(nlm_client_hosts, nsm, info)) != NULL) {
nlmclnt_recovery(host);
nlmclnt_release_host(host);
}
nsm_release(nsm);
}
/*
* Shut down the hosts module.
* Note that this routine is called only at server shutdown time.
*/
void
nlm_shutdown_hosts(void)
{
struct hlist_head *chain;
struct hlist_node *pos;
struct nlm_host *host;
dprintk("lockd: shutting down host module\n");
mutex_lock(&nlm_host_mutex);
/* First, make all hosts eligible for gc */
dprintk("lockd: nuking all hosts...\n");
for_each_host(host, pos, chain, nlm_server_hosts) {
host->h_expires = jiffies - 1;
if (host->h_rpcclnt) {
rpc_shutdown_client(host->h_rpcclnt);
host->h_rpcclnt = NULL;
}
}
/* Then, perform a garbage collection pass */
nlm_gc_hosts();
mutex_unlock(&nlm_host_mutex);
/* complain if any hosts are left */
if (nrhosts != 0) {
printk(KERN_WARNING "lockd: couldn't shutdown host module!\n");
dprintk("lockd: %lu hosts left:\n", nrhosts);
for_each_host(host, pos, chain, nlm_server_hosts) {
dprintk(" %s (cnt %d use %d exp %ld)\n",
host->h_name, atomic_read(&host->h_count),
host->h_inuse, host->h_expires);
}
}
}
/*
* Garbage collect any unused NLM hosts.
* This GC combines reference counting for async operations with
* mark & sweep for resources held by remote clients.
*/
static void
nlm_gc_hosts(void)
{
struct hlist_head *chain;
struct hlist_node *pos, *next;
struct nlm_host *host;
dprintk("lockd: host garbage collection\n");
for_each_host(host, pos, chain, nlm_server_hosts)
host->h_inuse = 0;
/* Mark all hosts that hold locks, blocks or shares */
nlmsvc_mark_resources();
for_each_host_safe(host, pos, next, chain, nlm_server_hosts) {
if (atomic_read(&host->h_count) || host->h_inuse
|| time_before(jiffies, host->h_expires)) {
dprintk("nlm_gc_hosts skipping %s "
"(cnt %d use %d exp %ld)\n",
host->h_name, atomic_read(&host->h_count),
host->h_inuse, host->h_expires);
continue;
}
nlm_destroy_host_locked(host);
}
next_gc = jiffies + NLM_HOST_COLLECT;
}