OpenCloudOS-Kernel/net/sunrpc/svc_xprt.c

1473 lines
39 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
/*
* linux/net/sunrpc/svc_xprt.c
*
* Author: Tom Tucker <tom@opengridcomputing.com>
*/
#include <linux/sched.h>
mm: introduce memalloc_retry_wait() Various places in the kernel - largely in filesystems - respond to a memory allocation failure by looping around and re-trying. Some of these cannot conveniently use __GFP_NOFAIL, for reasons such as: - a GFP_ATOMIC allocation, which __GFP_NOFAIL doesn't work on - a need to check for the process being signalled between failures - the possibility that other recovery actions could be performed - the allocation is quite deep in support code, and passing down an extra flag to say if __GFP_NOFAIL is wanted would be clumsy. Many of these currently use congestion_wait() which (in almost all cases) simply waits the given timeout - congestion isn't tracked for most devices. It isn't clear what the best delay is for loops, but it is clear that the various filesystems shouldn't be responsible for choosing a timeout. This patch introduces memalloc_retry_wait() with takes on that responsibility. Code that wants to retry a memory allocation can call this function passing the GFP flags that were used. It will wait however is appropriate. For now, it only considers __GFP_NORETRY and whatever gfpflags_allow_blocking() tests. If blocking is allowed without __GFP_NORETRY, then alloc_page either made some reclaim progress, or waited for a while, before failing. So there is no need for much further waiting. memalloc_retry_wait() will wait until the current jiffie ends. If this condition is not met, then alloc_page() won't have waited much if at all. In that case memalloc_retry_wait() waits about 200ms. This is the delay that most current loops uses. linux/sched/mm.h needs to be included in some files now, but linux/backing-dev.h does not. Link: https://lkml.kernel.org/r/163754371968.13692.1277530886009912421@noble.neil.brown.name Signed-off-by: NeilBrown <neilb@suse.de> Cc: Dave Chinner <david@fromorbit.com> Cc: Michal Hocko <mhocko@suse.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Jaegeuk Kim <jaegeuk@kernel.org> Cc: Chao Yu <chao@kernel.org> Cc: Darrick J. Wong <djwong@kernel.org> Cc: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:07:14 +08:00
#include <linux/sched/mm.h>
#include <linux/errno.h>
#include <linux/freezer.h>
#include <linux/kthread.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 16:04:11 +08:00
#include <linux/slab.h>
#include <net/sock.h>
#include <linux/sunrpc/addr.h>
#include <linux/sunrpc/stats.h>
#include <linux/sunrpc/svc_xprt.h>
#include <linux/sunrpc/svcsock.h>
#include <linux/sunrpc/xprt.h>
#include <linux/module.h>
#include <linux/netdevice.h>
#include <trace/events/sunrpc.h>
#define RPCDBG_FACILITY RPCDBG_SVCXPRT
static unsigned int svc_rpc_per_connection_limit __read_mostly;
module_param(svc_rpc_per_connection_limit, uint, 0644);
static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
static int svc_deferred_recv(struct svc_rqst *rqstp);
static struct cache_deferred_req *svc_defer(struct cache_req *req);
static void svc_age_temp_xprts(struct timer_list *t);
static void svc_delete_xprt(struct svc_xprt *xprt);
/* apparently the "standard" is that clients close
* idle connections after 5 minutes, servers after
* 6 minutes
* http://nfsv4bat.org/Documents/ConnectAThon/1996/nfstcp.pdf
*/
static int svc_conn_age_period = 6*60;
/* List of registered transport classes */
static DEFINE_SPINLOCK(svc_xprt_class_lock);
static LIST_HEAD(svc_xprt_class_list);
/* SMP locking strategy:
*
* svc_pool->sp_lock protects most of the fields of that pool.
* svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
* when both need to be taken (rare), svc_serv->sv_lock is first.
* The "service mutex" protects svc_serv->sv_nrthread.
* svc_sock->sk_lock protects the svc_sock->sk_deferred list
* and the ->sk_info_authunix cache.
*
* The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
* enqueued multiply. During normal transport processing this bit
* is set by svc_xprt_enqueue and cleared by svc_xprt_received.
* Providers should not manipulate this bit directly.
*
* Some flags can be set to certain values at any time
* providing that certain rules are followed:
*
* XPT_CONN, XPT_DATA:
* - Can be set or cleared at any time.
* - After a set, svc_xprt_enqueue must be called to enqueue
* the transport for processing.
* - After a clear, the transport must be read/accepted.
* If this succeeds, it must be set again.
* XPT_CLOSE:
* - Can set at any time. It is never cleared.
* XPT_DEAD:
* - Can only be set while XPT_BUSY is held which ensures
* that no other thread will be using the transport or will
* try to set XPT_DEAD.
*/
int svc_reg_xprt_class(struct svc_xprt_class *xcl)
{
struct svc_xprt_class *cl;
int res = -EEXIST;
dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
INIT_LIST_HEAD(&xcl->xcl_list);
spin_lock(&svc_xprt_class_lock);
/* Make sure there isn't already a class with the same name */
list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
goto out;
}
list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
res = 0;
out:
spin_unlock(&svc_xprt_class_lock);
return res;
}
EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
{
dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
spin_lock(&svc_xprt_class_lock);
list_del_init(&xcl->xcl_list);
spin_unlock(&svc_xprt_class_lock);
}
EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
/**
* svc_print_xprts - Format the transport list for printing
* @buf: target buffer for formatted address
* @maxlen: length of target buffer
*
* Fills in @buf with a string containing a list of transport names, each name
* terminated with '\n'. If the buffer is too small, some entries may be
* missing, but it is guaranteed that all lines in the output buffer are
* complete.
*
* Returns positive length of the filled-in string.
*/
int svc_print_xprts(char *buf, int maxlen)
{
struct svc_xprt_class *xcl;
char tmpstr[80];
int len = 0;
buf[0] = '\0';
spin_lock(&svc_xprt_class_lock);
list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
int slen;
slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
xcl->xcl_name, xcl->xcl_max_payload);
if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
break;
len += slen;
strcat(buf, tmpstr);
}
spin_unlock(&svc_xprt_class_lock);
return len;
}
/**
* svc_xprt_deferred_close - Close a transport
* @xprt: transport instance
*
* Used in contexts that need to defer the work of shutting down
* the transport to an nfsd thread.
*/
void svc_xprt_deferred_close(struct svc_xprt *xprt)
{
if (!test_and_set_bit(XPT_CLOSE, &xprt->xpt_flags))
svc_xprt_enqueue(xprt);
}
EXPORT_SYMBOL_GPL(svc_xprt_deferred_close);
static void svc_xprt_free(struct kref *kref)
{
struct svc_xprt *xprt =
container_of(kref, struct svc_xprt, xpt_ref);
struct module *owner = xprt->xpt_class->xcl_owner;
if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
svcauth_unix_info_release(xprt);
put_cred(xprt->xpt_cred);
put_net_track(xprt->xpt_net, &xprt->ns_tracker);
/* See comment on corresponding get in xs_setup_bc_tcp(): */
if (xprt->xpt_bc_xprt)
xprt_put(xprt->xpt_bc_xprt);
if (xprt->xpt_bc_xps)
xprt_switch_put(xprt->xpt_bc_xps);
trace_svc_xprt_free(xprt);
xprt->xpt_ops->xpo_free(xprt);
module_put(owner);
}
void svc_xprt_put(struct svc_xprt *xprt)
{
kref_put(&xprt->xpt_ref, svc_xprt_free);
}
EXPORT_SYMBOL_GPL(svc_xprt_put);
/*
* Called by transport drivers to initialize the transport independent
* portion of the transport instance.
*/
void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
struct svc_xprt *xprt, struct svc_serv *serv)
{
memset(xprt, 0, sizeof(*xprt));
xprt->xpt_class = xcl;
xprt->xpt_ops = xcl->xcl_ops;
kref_init(&xprt->xpt_ref);
xprt->xpt_server = serv;
INIT_LIST_HEAD(&xprt->xpt_list);
INIT_LIST_HEAD(&xprt->xpt_ready);
INIT_LIST_HEAD(&xprt->xpt_deferred);
INIT_LIST_HEAD(&xprt->xpt_users);
mutex_init(&xprt->xpt_mutex);
spin_lock_init(&xprt->xpt_lock);
set_bit(XPT_BUSY, &xprt->xpt_flags);
xprt->xpt_net = get_net_track(net, &xprt->ns_tracker, GFP_ATOMIC);
strcpy(xprt->xpt_remotebuf, "uninitialized");
}
EXPORT_SYMBOL_GPL(svc_xprt_init);
static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
struct svc_serv *serv,
struct net *net,
const int family,
const unsigned short port,
int flags)
{
struct sockaddr_in sin = {
.sin_family = AF_INET,
.sin_addr.s_addr = htonl(INADDR_ANY),
.sin_port = htons(port),
};
#if IS_ENABLED(CONFIG_IPV6)
struct sockaddr_in6 sin6 = {
.sin6_family = AF_INET6,
.sin6_addr = IN6ADDR_ANY_INIT,
.sin6_port = htons(port),
};
#endif
struct svc_xprt *xprt;
struct sockaddr *sap;
size_t len;
switch (family) {
case PF_INET:
sap = (struct sockaddr *)&sin;
len = sizeof(sin);
break;
#if IS_ENABLED(CONFIG_IPV6)
case PF_INET6:
sap = (struct sockaddr *)&sin6;
len = sizeof(sin6);
break;
#endif
default:
return ERR_PTR(-EAFNOSUPPORT);
}
xprt = xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
if (IS_ERR(xprt))
trace_svc_xprt_create_err(serv->sv_program->pg_name,
xcl->xcl_name, sap, len, xprt);
return xprt;
}
/**
* svc_xprt_received - start next receiver thread
* @xprt: controlling transport
*
* The caller must hold the XPT_BUSY bit and must
* not thereafter touch transport data.
*
* Note: XPT_DATA only gets cleared when a read-attempt finds no (or
* insufficient) data.
*/
void svc_xprt_received(struct svc_xprt *xprt)
{
if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
return;
}
/* As soon as we clear busy, the xprt could be closed and
* 'put', so we need a reference to call svc_xprt_enqueue with:
*/
svc_xprt_get(xprt);
smp_mb__before_atomic();
clear_bit(XPT_BUSY, &xprt->xpt_flags);
svc_xprt_enqueue(xprt);
svc_xprt_put(xprt);
}
EXPORT_SYMBOL_GPL(svc_xprt_received);
void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
{
clear_bit(XPT_TEMP, &new->xpt_flags);
spin_lock_bh(&serv->sv_lock);
list_add(&new->xpt_list, &serv->sv_permsocks);
spin_unlock_bh(&serv->sv_lock);
svc_xprt_received(new);
}
static int _svc_xprt_create(struct svc_serv *serv, const char *xprt_name,
struct net *net, const int family,
const unsigned short port, int flags,
const struct cred *cred)
{
struct svc_xprt_class *xcl;
spin_lock(&svc_xprt_class_lock);
list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
struct svc_xprt *newxprt;
unsigned short newport;
if (strcmp(xprt_name, xcl->xcl_name))
continue;
if (!try_module_get(xcl->xcl_owner))
goto err;
spin_unlock(&svc_xprt_class_lock);
newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
if (IS_ERR(newxprt)) {
module_put(xcl->xcl_owner);
return PTR_ERR(newxprt);
}
newxprt->xpt_cred = get_cred(cred);
svc_add_new_perm_xprt(serv, newxprt);
newport = svc_xprt_local_port(newxprt);
return newport;
}
err:
spin_unlock(&svc_xprt_class_lock);
/* This errno is exposed to user space. Provide a reasonable
* perror msg for a bad transport. */
return -EPROTONOSUPPORT;
}
/**
* svc_xprt_create - Add a new listener to @serv
* @serv: target RPC service
* @xprt_name: transport class name
* @net: network namespace
* @family: network address family
* @port: listener port
* @flags: SVC_SOCK flags
* @cred: credential to bind to this transport
*
* Return values:
* %0: New listener added successfully
* %-EPROTONOSUPPORT: Requested transport type not supported
*/
int svc_xprt_create(struct svc_serv *serv, const char *xprt_name,
struct net *net, const int family,
const unsigned short port, int flags,
const struct cred *cred)
{
int err;
err = _svc_xprt_create(serv, xprt_name, net, family, port, flags, cred);
if (err == -EPROTONOSUPPORT) {
request_module("svc%s", xprt_name);
err = _svc_xprt_create(serv, xprt_name, net, family, port, flags, cred);
}
return err;
}
EXPORT_SYMBOL_GPL(svc_xprt_create);
svc: Move the sockaddr information to svc_xprt This patch moves the transport sockaddr to the svc_xprt structure. Convenience functions are added to set and get the local and remote addresses of a transport from the transport provider as well as determine the length of a sockaddr. A transport is responsible for setting the xpt_local and xpt_remote addresses in the svc_xprt structure as part of transport creation and xpo_accept processing. This cannot be done in a generic way and in fact varies between TCP, UDP and RDMA. A set of xpo_ functions (e.g. getlocalname, getremotename) could have been added but this would have resulted in additional caching and copying of the addresses around. Note that the xpt_local address should also be set on listening endpoints; for TCP/RDMA this is done as part of endpoint creation. For connected transports like TCP and RDMA, the addresses never change and can be set once and copied into the rqstp structure for each request. For UDP, however, the local and remote addresses may change for each request. In this case, the address information is obtained from the UDP recvmsg info and copied into the rqstp structure from there. A svc_xprt_local_port function was also added that returns the local port given a transport. This is used by svc_create_xprt when returning the port associated with a newly created transport, and later when creating a generic find transport service to check if a service is already listening on a given port. Signed-off-by: Tom Tucker <tom@opengridcomputing.com> Acked-by: Neil Brown <neilb@suse.de> Reviewed-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Greg Banks <gnb@sgi.com> Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
2007-12-31 11:08:12 +08:00
/*
* Copy the local and remote xprt addresses to the rqstp structure
*/
void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
{
memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
rqstp->rq_addrlen = xprt->xpt_remotelen;
/*
* Destination address in request is needed for binding the
* source address in RPC replies/callbacks later.
*/
memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
rqstp->rq_daddrlen = xprt->xpt_locallen;
svc: Move the sockaddr information to svc_xprt This patch moves the transport sockaddr to the svc_xprt structure. Convenience functions are added to set and get the local and remote addresses of a transport from the transport provider as well as determine the length of a sockaddr. A transport is responsible for setting the xpt_local and xpt_remote addresses in the svc_xprt structure as part of transport creation and xpo_accept processing. This cannot be done in a generic way and in fact varies between TCP, UDP and RDMA. A set of xpo_ functions (e.g. getlocalname, getremotename) could have been added but this would have resulted in additional caching and copying of the addresses around. Note that the xpt_local address should also be set on listening endpoints; for TCP/RDMA this is done as part of endpoint creation. For connected transports like TCP and RDMA, the addresses never change and can be set once and copied into the rqstp structure for each request. For UDP, however, the local and remote addresses may change for each request. In this case, the address information is obtained from the UDP recvmsg info and copied into the rqstp structure from there. A svc_xprt_local_port function was also added that returns the local port given a transport. This is used by svc_create_xprt when returning the port associated with a newly created transport, and later when creating a generic find transport service to check if a service is already listening on a given port. Signed-off-by: Tom Tucker <tom@opengridcomputing.com> Acked-by: Neil Brown <neilb@suse.de> Reviewed-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Greg Banks <gnb@sgi.com> Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
2007-12-31 11:08:12 +08:00
}
EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
/**
* svc_print_addr - Format rq_addr field for printing
* @rqstp: svc_rqst struct containing address to print
* @buf: target buffer for formatted address
* @len: length of target buffer
*
*/
char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
{
return __svc_print_addr(svc_addr(rqstp), buf, len);
}
EXPORT_SYMBOL_GPL(svc_print_addr);
static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
{
unsigned int limit = svc_rpc_per_connection_limit;
int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
}
static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
{
if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
if (!svc_xprt_slots_in_range(xprt))
return false;
atomic_inc(&xprt->xpt_nr_rqsts);
set_bit(RQ_DATA, &rqstp->rq_flags);
}
return true;
}
static void svc_xprt_release_slot(struct svc_rqst *rqstp)
{
struct svc_xprt *xprt = rqstp->rq_xprt;
if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
atomic_dec(&xprt->xpt_nr_rqsts);
smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
svc_xprt_enqueue(xprt);
}
}
static bool svc_xprt_ready(struct svc_xprt *xprt)
{
unsigned long xpt_flags;
/*
* If another cpu has recently updated xpt_flags,
* sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
* know about it; otherwise it's possible that both that cpu and
* this one could call svc_xprt_enqueue() without either
* svc_xprt_enqueue() recognizing that the conditions below
* are satisfied, and we could stall indefinitely:
*/
smp_rmb();
xpt_flags = READ_ONCE(xprt->xpt_flags);
if (xpt_flags & BIT(XPT_BUSY))
return false;
if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
return true;
if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
svc_xprt_slots_in_range(xprt))
return true;
trace_svc_xprt_no_write_space(xprt);
return false;
}
return false;
}
/**
* svc_xprt_enqueue - Queue a transport on an idle nfsd thread
* @xprt: transport with data pending
*
*/
void svc_xprt_enqueue(struct svc_xprt *xprt)
{
struct svc_pool *pool;
struct svc_rqst *rqstp = NULL;
if (!svc_xprt_ready(xprt))
return;
/* Mark transport as busy. It will remain in this state until
* the provider calls svc_xprt_received. We update XPT_BUSY
* atomically because it also guards against trying to enqueue
* the transport twice.
*/
if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
return;
pool = svc_pool_for_cpu(xprt->xpt_server);
atomic_long_inc(&pool->sp_stats.packets);
spin_lock_bh(&pool->sp_lock);
list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
pool->sp_stats.sockets_queued++;
spin_unlock_bh(&pool->sp_lock);
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
/* find a thread for this xprt */
rcu_read_lock();
list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
continue;
atomic_long_inc(&pool->sp_stats.threads_woken);
rqstp->rq_qtime = ktime_get();
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
wake_up_process(rqstp->rq_task);
goto out_unlock;
}
set_bit(SP_CONGESTED, &pool->sp_flags);
rqstp = NULL;
out_unlock:
rcu_read_unlock();
trace_svc_xprt_enqueue(xprt, rqstp);
}
EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
/*
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
* Dequeue the first transport, if there is one.
*/
static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
{
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
struct svc_xprt *xprt = NULL;
if (list_empty(&pool->sp_sockets))
goto out;
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
spin_lock_bh(&pool->sp_lock);
if (likely(!list_empty(&pool->sp_sockets))) {
xprt = list_first_entry(&pool->sp_sockets,
struct svc_xprt, xpt_ready);
list_del_init(&xprt->xpt_ready);
svc_xprt_get(xprt);
}
spin_unlock_bh(&pool->sp_lock);
out:
return xprt;
}
/**
* svc_reserve - change the space reserved for the reply to a request.
* @rqstp: The request in question
* @space: new max space to reserve
*
* Each request reserves some space on the output queue of the transport
* to make sure the reply fits. This function reduces that reserved
* space to be the amount of space used already, plus @space.
*
*/
void svc_reserve(struct svc_rqst *rqstp, int space)
{
sunrpc: use-after-free in svc_process_common() if node have NFSv41+ mounts inside several net namespaces it can lead to use-after-free in svc_process_common() svc_process_common() /* Setup reply header */ rqstp->rq_xprt->xpt_ops->xpo_prep_reply_hdr(rqstp); <<< HERE svc_process_common() can use incorrect rqstp->rq_xprt, its caller function bc_svc_process() takes it from serv->sv_bc_xprt. The problem is that serv is global structure but sv_bc_xprt is assigned per-netnamespace. According to Trond, the whole "let's set up rqstp->rq_xprt for the back channel" is nothing but a giant hack in order to work around the fact that svc_process_common() uses it to find the xpt_ops, and perform a couple of (meaningless for the back channel) tests of xpt_flags. All we really need in svc_process_common() is to be able to run rqstp->rq_xprt->xpt_ops->xpo_prep_reply_hdr() Bruce J Fields points that this xpo_prep_reply_hdr() call is an awfully roundabout way just to do "svc_putnl(resv, 0);" in the tcp case. This patch does not initialiuze rqstp->rq_xprt in bc_svc_process(), now it calls svc_process_common() with rqstp->rq_xprt = NULL. To adjust reply header svc_process_common() just check rqstp->rq_prot and calls svc_tcp_prep_reply_hdr() for tcp case. To handle rqstp->rq_xprt = NULL case in functions called from svc_process_common() patch intruduces net namespace pointer svc_rqst->rq_bc_net and adjust SVC_NET() definition. Some other function was also adopted to properly handle described case. Signed-off-by: Vasily Averin <vvs@virtuozzo.com> Cc: stable@vger.kernel.org Fixes: 23c20ecd4475 ("NFS: callback up - users counting cleanup") Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2018-12-24 19:44:52 +08:00
struct svc_xprt *xprt = rqstp->rq_xprt;
space += rqstp->rq_res.head[0].iov_len;
sunrpc: use-after-free in svc_process_common() if node have NFSv41+ mounts inside several net namespaces it can lead to use-after-free in svc_process_common() svc_process_common() /* Setup reply header */ rqstp->rq_xprt->xpt_ops->xpo_prep_reply_hdr(rqstp); <<< HERE svc_process_common() can use incorrect rqstp->rq_xprt, its caller function bc_svc_process() takes it from serv->sv_bc_xprt. The problem is that serv is global structure but sv_bc_xprt is assigned per-netnamespace. According to Trond, the whole "let's set up rqstp->rq_xprt for the back channel" is nothing but a giant hack in order to work around the fact that svc_process_common() uses it to find the xpt_ops, and perform a couple of (meaningless for the back channel) tests of xpt_flags. All we really need in svc_process_common() is to be able to run rqstp->rq_xprt->xpt_ops->xpo_prep_reply_hdr() Bruce J Fields points that this xpo_prep_reply_hdr() call is an awfully roundabout way just to do "svc_putnl(resv, 0);" in the tcp case. This patch does not initialiuze rqstp->rq_xprt in bc_svc_process(), now it calls svc_process_common() with rqstp->rq_xprt = NULL. To adjust reply header svc_process_common() just check rqstp->rq_prot and calls svc_tcp_prep_reply_hdr() for tcp case. To handle rqstp->rq_xprt = NULL case in functions called from svc_process_common() patch intruduces net namespace pointer svc_rqst->rq_bc_net and adjust SVC_NET() definition. Some other function was also adopted to properly handle described case. Signed-off-by: Vasily Averin <vvs@virtuozzo.com> Cc: stable@vger.kernel.org Fixes: 23c20ecd4475 ("NFS: callback up - users counting cleanup") Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2018-12-24 19:44:52 +08:00
if (xprt && space < rqstp->rq_reserved) {
atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
rqstp->rq_reserved = space;
smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
svc_xprt_enqueue(xprt);
}
}
EXPORT_SYMBOL_GPL(svc_reserve);
static void svc_xprt_release(struct svc_rqst *rqstp)
{
struct svc_xprt *xprt = rqstp->rq_xprt;
xprt->xpt_ops->xpo_release_rqst(rqstp);
kfree(rqstp->rq_deferred);
rqstp->rq_deferred = NULL;
pagevec_release(&rqstp->rq_pvec);
svc_free_res_pages(rqstp);
rqstp->rq_res.page_len = 0;
rqstp->rq_res.page_base = 0;
/* Reset response buffer and release
* the reservation.
* But first, check that enough space was reserved
* for the reply, otherwise we have a bug!
*/
if ((rqstp->rq_res.len) > rqstp->rq_reserved)
printk(KERN_ERR "RPC request reserved %d but used %d\n",
rqstp->rq_reserved,
rqstp->rq_res.len);
rqstp->rq_res.head[0].iov_len = 0;
svc_reserve(rqstp, 0);
svc_xprt_release_slot(rqstp);
rqstp->rq_xprt = NULL;
svc_xprt_put(xprt);
}
/*
* Some svc_serv's will have occasional work to do, even when a xprt is not
* waiting to be serviced. This function is there to "kick" a task in one of
* those services so that it can wake up and do that work. Note that we only
* bother with pool 0 as we don't need to wake up more than one thread for
* this purpose.
*/
void svc_wake_up(struct svc_serv *serv)
{
struct svc_rqst *rqstp;
struct svc_pool *pool;
pool = &serv->sv_pools[0];
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
rcu_read_lock();
list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
/* skip any that aren't queued */
if (test_bit(RQ_BUSY, &rqstp->rq_flags))
continue;
rcu_read_unlock();
wake_up_process(rqstp->rq_task);
trace_svc_wake_up(rqstp->rq_task->pid);
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
return;
}
rcu_read_unlock();
/* No free entries available */
set_bit(SP_TASK_PENDING, &pool->sp_flags);
smp_wmb();
trace_svc_wake_up(0);
}
EXPORT_SYMBOL_GPL(svc_wake_up);
int svc_port_is_privileged(struct sockaddr *sin)
{
switch (sin->sa_family) {
case AF_INET:
return ntohs(((struct sockaddr_in *)sin)->sin_port)
< PROT_SOCK;
case AF_INET6:
return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
< PROT_SOCK;
default:
return 0;
}
}
/*
* Make sure that we don't have too many active connections. If we have,
* something must be dropped. It's not clear what will happen if we allow
* "too many" connections, but when dealing with network-facing software,
* we have to code defensively. Here we do that by imposing hard limits.
*
* There's no point in trying to do random drop here for DoS
* prevention. The NFS clients does 1 reconnect in 15 seconds. An
* attacker can easily beat that.
*
* The only somewhat efficient mechanism would be if drop old
* connections from the same IP first. But right now we don't even
* record the client IP in svc_sock.
*
* single-threaded services that expect a lot of clients will probably
* need to set sv_maxconn to override the default value which is based
* on the number of threads
*/
static void svc_check_conn_limits(struct svc_serv *serv)
{
unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
(serv->sv_nrthreads+3) * 20;
if (serv->sv_tmpcnt > limit) {
struct svc_xprt *xprt = NULL;
spin_lock_bh(&serv->sv_lock);
if (!list_empty(&serv->sv_tempsocks)) {
/* Try to help the admin */
net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
serv->sv_name, serv->sv_maxconn ?
"max number of connections" :
"number of threads");
/*
* Always select the oldest connection. It's not fair,
* but so is life
*/
xprt = list_entry(serv->sv_tempsocks.prev,
struct svc_xprt,
xpt_list);
set_bit(XPT_CLOSE, &xprt->xpt_flags);
svc_xprt_get(xprt);
}
spin_unlock_bh(&serv->sv_lock);
if (xprt) {
svc_xprt_enqueue(xprt);
svc_xprt_put(xprt);
}
}
}
static int svc_alloc_arg(struct svc_rqst *rqstp)
{
struct svc_serv *serv = rqstp->rq_server;
SUNRPC: set rq_page_end differently Patch series "SUNRPC consumer for the bulk page allocator" This patch set and the measurements below are based on yesterday's bulk allocator series: git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git mm-bulk-rebase-v5r9 The patches change SUNRPC to invoke the array-based bulk allocator instead of alloc_page(). The micro-benchmark results are promising. I ran a mixture of 256KB reads and writes over NFSv3. The server's kernel is built with KASAN enabled, so the comparison is exaggerated but I believe it is still valid. I instrumented svc_recv() to measure the latency of each call to svc_alloc_arg() and report it via a trace point. The following results are averages across the trace events. Single page: 25.007 us per call over 532,571 calls Bulk list: 6.258 us per call over 517,034 calls Bulk array: 4.590 us per call over 517,442 calls This patch (of 2) Refactor: I'm about to use the loop variable @i for something else. As far as the "i++" is concerned, that is a post-increment. The value of @i is not used subsequently, so the increment operator is unnecessary and can be removed. Also note that nfsd_read_actor() was renamed nfsd_splice_actor() by commit cf8208d0eabd ("sendfile: convert nfsd to splice_direct_to_actor()"). Link: https://lkml.kernel.org/r/20210325114228.27719-7-mgorman@techsingularity.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Alexander Lobakin <alobakin@pm.me> Cc: Alexander Duyck <alexander.duyck@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David Miller <davem@davemloft.net> Cc: Ilias Apalodimas <ilias.apalodimas@linaro.org> Cc: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:01:58 +08:00
struct xdr_buf *arg = &rqstp->rq_arg;
unsigned long pages, filled, ret;
pagevec_init(&rqstp->rq_pvec);
sunrpc: Allocate up to RPCSVC_MAXPAGES per svc_rqst svcrdma needs 259 pages allocated to receive 1MB NFSv4.0 WRITE requests: - 1 page for the transport header and head iovec - 256 pages for the data payload - 1 page for the trailing GETATTR request (since NFSD XDR decoding does not look for a tail iovec, the GETATTR is stuck at the end of the rqstp->rq_arg.pages list) - 1 page for building the reply xdr_buf But RPCSVC_MAXPAGES is already 259 (on x86_64). The problem is that svc_alloc_arg never allocates that many pages. To address this: 1. The final element of rq_pages always points to NULL. To accommodate up to 259 pages in rq_pages, add an extra element to rq_pages for the array termination sentinel. 2. Adjust the calculation of "pages" to match how RPCSVC_MAXPAGES is calculated, so it can go up to 259. Bruce noted that the calculation assumes sv_max_mesg is a multiple of PAGE_SIZE, which might not always be true. I didn't change this assumption. 3. Change the loop boundaries to allow 259 pages to be allocated. Additional clean-up: WARN_ON_ONCE adds an extra conditional branch, which is basically never taken. And there's no need to dump the stack here because svc_alloc_arg has only one caller. Keeping that NULL "array termination sentinel"; there doesn't appear to be any code that depends on it, only code in nfsd_splice_actor() which needs the 259th element to be initialized to *something*. So it's possible we could just keep the array at 259 elements and drop that final NULL, but we're being conservative for now. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2017-07-01 00:03:54 +08:00
pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
if (pages > RPCSVC_MAXPAGES) {
pr_warn_once("svc: warning: pages=%lu > RPCSVC_MAXPAGES=%lu\n",
sunrpc: Allocate up to RPCSVC_MAXPAGES per svc_rqst svcrdma needs 259 pages allocated to receive 1MB NFSv4.0 WRITE requests: - 1 page for the transport header and head iovec - 256 pages for the data payload - 1 page for the trailing GETATTR request (since NFSD XDR decoding does not look for a tail iovec, the GETATTR is stuck at the end of the rqstp->rq_arg.pages list) - 1 page for building the reply xdr_buf But RPCSVC_MAXPAGES is already 259 (on x86_64). The problem is that svc_alloc_arg never allocates that many pages. To address this: 1. The final element of rq_pages always points to NULL. To accommodate up to 259 pages in rq_pages, add an extra element to rq_pages for the array termination sentinel. 2. Adjust the calculation of "pages" to match how RPCSVC_MAXPAGES is calculated, so it can go up to 259. Bruce noted that the calculation assumes sv_max_mesg is a multiple of PAGE_SIZE, which might not always be true. I didn't change this assumption. 3. Change the loop boundaries to allow 259 pages to be allocated. Additional clean-up: WARN_ON_ONCE adds an extra conditional branch, which is basically never taken. And there's no need to dump the stack here because svc_alloc_arg has only one caller. Keeping that NULL "array termination sentinel"; there doesn't appear to be any code that depends on it, only code in nfsd_splice_actor() which needs the 259th element to be initialized to *something*. So it's possible we could just keep the array at 259 elements and drop that final NULL, but we're being conservative for now. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2017-07-01 00:03:54 +08:00
pages, RPCSVC_MAXPAGES);
/* use as many pages as possible */
sunrpc: Allocate up to RPCSVC_MAXPAGES per svc_rqst svcrdma needs 259 pages allocated to receive 1MB NFSv4.0 WRITE requests: - 1 page for the transport header and head iovec - 256 pages for the data payload - 1 page for the trailing GETATTR request (since NFSD XDR decoding does not look for a tail iovec, the GETATTR is stuck at the end of the rqstp->rq_arg.pages list) - 1 page for building the reply xdr_buf But RPCSVC_MAXPAGES is already 259 (on x86_64). The problem is that svc_alloc_arg never allocates that many pages. To address this: 1. The final element of rq_pages always points to NULL. To accommodate up to 259 pages in rq_pages, add an extra element to rq_pages for the array termination sentinel. 2. Adjust the calculation of "pages" to match how RPCSVC_MAXPAGES is calculated, so it can go up to 259. Bruce noted that the calculation assumes sv_max_mesg is a multiple of PAGE_SIZE, which might not always be true. I didn't change this assumption. 3. Change the loop boundaries to allow 259 pages to be allocated. Additional clean-up: WARN_ON_ONCE adds an extra conditional branch, which is basically never taken. And there's no need to dump the stack here because svc_alloc_arg has only one caller. Keeping that NULL "array termination sentinel"; there doesn't appear to be any code that depends on it, only code in nfsd_splice_actor() which needs the 259th element to be initialized to *something*. So it's possible we could just keep the array at 259 elements and drop that final NULL, but we're being conservative for now. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2017-07-01 00:03:54 +08:00
pages = RPCSVC_MAXPAGES;
}
for (filled = 0; filled < pages; filled = ret) {
ret = alloc_pages_bulk_array(GFP_KERNEL, pages,
rqstp->rq_pages);
if (ret > filled)
/* Made progress, don't sleep yet */
continue;
set_current_state(TASK_INTERRUPTIBLE);
if (signalled() || kthread_should_stop()) {
set_current_state(TASK_RUNNING);
return -EINTR;
}
trace_svc_alloc_arg_err(pages);
mm: introduce memalloc_retry_wait() Various places in the kernel - largely in filesystems - respond to a memory allocation failure by looping around and re-trying. Some of these cannot conveniently use __GFP_NOFAIL, for reasons such as: - a GFP_ATOMIC allocation, which __GFP_NOFAIL doesn't work on - a need to check for the process being signalled between failures - the possibility that other recovery actions could be performed - the allocation is quite deep in support code, and passing down an extra flag to say if __GFP_NOFAIL is wanted would be clumsy. Many of these currently use congestion_wait() which (in almost all cases) simply waits the given timeout - congestion isn't tracked for most devices. It isn't clear what the best delay is for loops, but it is clear that the various filesystems shouldn't be responsible for choosing a timeout. This patch introduces memalloc_retry_wait() with takes on that responsibility. Code that wants to retry a memory allocation can call this function passing the GFP flags that were used. It will wait however is appropriate. For now, it only considers __GFP_NORETRY and whatever gfpflags_allow_blocking() tests. If blocking is allowed without __GFP_NORETRY, then alloc_page either made some reclaim progress, or waited for a while, before failing. So there is no need for much further waiting. memalloc_retry_wait() will wait until the current jiffie ends. If this condition is not met, then alloc_page() won't have waited much if at all. In that case memalloc_retry_wait() waits about 200ms. This is the delay that most current loops uses. linux/sched/mm.h needs to be included in some files now, but linux/backing-dev.h does not. Link: https://lkml.kernel.org/r/163754371968.13692.1277530886009912421@noble.neil.brown.name Signed-off-by: NeilBrown <neilb@suse.de> Cc: Dave Chinner <david@fromorbit.com> Cc: Michal Hocko <mhocko@suse.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Jaegeuk Kim <jaegeuk@kernel.org> Cc: Chao Yu <chao@kernel.org> Cc: Darrick J. Wong <djwong@kernel.org> Cc: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-15 06:07:14 +08:00
memalloc_retry_wait(GFP_KERNEL);
}
SUNRPC: set rq_page_end differently Patch series "SUNRPC consumer for the bulk page allocator" This patch set and the measurements below are based on yesterday's bulk allocator series: git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git mm-bulk-rebase-v5r9 The patches change SUNRPC to invoke the array-based bulk allocator instead of alloc_page(). The micro-benchmark results are promising. I ran a mixture of 256KB reads and writes over NFSv3. The server's kernel is built with KASAN enabled, so the comparison is exaggerated but I believe it is still valid. I instrumented svc_recv() to measure the latency of each call to svc_alloc_arg() and report it via a trace point. The following results are averages across the trace events. Single page: 25.007 us per call over 532,571 calls Bulk list: 6.258 us per call over 517,034 calls Bulk array: 4.590 us per call over 517,442 calls This patch (of 2) Refactor: I'm about to use the loop variable @i for something else. As far as the "i++" is concerned, that is a post-increment. The value of @i is not used subsequently, so the increment operator is unnecessary and can be removed. Also note that nfsd_read_actor() was renamed nfsd_splice_actor() by commit cf8208d0eabd ("sendfile: convert nfsd to splice_direct_to_actor()"). Link: https://lkml.kernel.org/r/20210325114228.27719-7-mgorman@techsingularity.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Reviewed-by: Alexander Lobakin <alobakin@pm.me> Cc: Alexander Duyck <alexander.duyck@gmail.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David Miller <davem@davemloft.net> Cc: Ilias Apalodimas <ilias.apalodimas@linaro.org> Cc: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:01:58 +08:00
rqstp->rq_page_end = &rqstp->rq_pages[pages];
rqstp->rq_pages[pages] = NULL; /* this might be seen in nfsd_splice_actor() */
/* Make arg->head point to first page and arg->pages point to rest */
arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
arg->head[0].iov_len = PAGE_SIZE;
arg->pages = rqstp->rq_pages + 1;
arg->page_base = 0;
/* save at least one page for response */
arg->page_len = (pages-2)*PAGE_SIZE;
arg->len = (pages-1)*PAGE_SIZE;
arg->tail[0].iov_len = 0;
return 0;
}
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
static bool
rqst_should_sleep(struct svc_rqst *rqstp)
{
struct svc_pool *pool = rqstp->rq_pool;
/* did someone call svc_wake_up? */
if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
return false;
/* was a socket queued? */
if (!list_empty(&pool->sp_sockets))
return false;
/* are we shutting down? */
if (signalled() || kthread_should_stop())
return false;
/* are we freezing? */
if (freezing(current))
return false;
return true;
}
static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
{
struct svc_pool *pool = rqstp->rq_pool;
long time_left = 0;
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
/* rq_xprt should be clear on entry */
WARN_ON_ONCE(rqstp->rq_xprt);
rqstp->rq_xprt = svc_xprt_dequeue(pool);
if (rqstp->rq_xprt)
goto out_found;
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
/*
* We have to be able to interrupt this wait
* to bring down the daemons ...
*/
set_current_state(TASK_INTERRUPTIBLE);
smp_mb__before_atomic();
clear_bit(SP_CONGESTED, &pool->sp_flags);
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
clear_bit(RQ_BUSY, &rqstp->rq_flags);
smp_mb__after_atomic();
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
if (likely(rqst_should_sleep(rqstp)))
time_left = schedule_timeout(timeout);
else
__set_current_state(TASK_RUNNING);
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
try_to_freeze();
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
set_bit(RQ_BUSY, &rqstp->rq_flags);
smp_mb__after_atomic();
rqstp->rq_xprt = svc_xprt_dequeue(pool);
if (rqstp->rq_xprt)
goto out_found;
sunrpc: convert to lockless lookup of queued server threads Testing has shown that the pool->sp_lock can be a bottleneck on a busy server. Every time data is received on a socket, the server must take that lock in order to dequeue a thread from the sp_threads list. Address this problem by eliminating the sp_threads list (which contains threads that are currently idle) and replacing it with a RQ_BUSY flag in svc_rqst. This allows us to walk the sp_all_threads list under the rcu_read_lock and find a suitable thread for the xprt by doing a test_and_set_bit. Note that we do still have a potential atomicity problem however with this approach. We don't want svc_xprt_do_enqueue to set the rqst->rq_xprt pointer unless a test_and_set_bit of RQ_BUSY returned zero (which indicates that the thread was idle). But, by the time we check that, the bit could be flipped by a waking thread. To address this, we acquire a new per-rqst spinlock (rq_lock) and take that before doing the test_and_set_bit. If that returns false, then we can set rq_xprt and drop the spinlock. Then, when the thread wakes up, it must set the bit under the same spinlock and can trust that if it was already set then the rq_xprt is also properly set. With this scheme, the case where we have an idle thread no longer needs to take the highly contended pool->sp_lock at all, and that removes the bottleneck. That still leaves one issue: What of the case where we walk the whole sp_all_threads list and don't find an idle thread? Because the search is lockess, it's possible for the queueing to race with a thread that is going to sleep. To address that, we queue the xprt and then search again. If we find an idle thread at that point, we can't attach the xprt to it directly since that might race with a different thread waking up and finding it. All we can do is wake the idle thread back up and let it attempt to find the now-queued xprt. Signed-off-by: Jeff Layton <jlayton@primarydata.com> Tested-by: Chris Worley <chris.worley@primarydata.com> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2014-11-22 03:19:30 +08:00
if (!time_left)
atomic_long_inc(&pool->sp_stats.threads_timedout);
if (signalled() || kthread_should_stop())
return ERR_PTR(-EINTR);
return ERR_PTR(-EAGAIN);
out_found:
/* Normally we will wait up to 5 seconds for any required
* cache information to be provided.
*/
if (!test_bit(SP_CONGESTED, &pool->sp_flags))
rqstp->rq_chandle.thread_wait = 5*HZ;
else
rqstp->rq_chandle.thread_wait = 1*HZ;
trace_svc_xprt_dequeue(rqstp);
return rqstp->rq_xprt;
}
static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
{
spin_lock_bh(&serv->sv_lock);
set_bit(XPT_TEMP, &newxpt->xpt_flags);
list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
serv->sv_tmpcnt++;
if (serv->sv_temptimer.function == NULL) {
/* setup timer to age temp transports */
serv->sv_temptimer.function = svc_age_temp_xprts;
mod_timer(&serv->sv_temptimer,
jiffies + svc_conn_age_period * HZ);
}
spin_unlock_bh(&serv->sv_lock);
svc_xprt_received(newxpt);
}
static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
{
struct svc_serv *serv = rqstp->rq_server;
int len = 0;
if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
svc_delete_xprt(xprt);
/* Leave XPT_BUSY set on the dead xprt: */
goto out;
}
if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
struct svc_xprt *newxpt;
/*
* We know this module_get will succeed because the
* listener holds a reference too
*/
__module_get(xprt->xpt_class->xcl_owner);
svc_check_conn_limits(xprt->xpt_server);
newxpt = xprt->xpt_ops->xpo_accept(xprt);
if (newxpt) {
newxpt->xpt_cred = get_cred(xprt->xpt_cred);
svc_add_new_temp_xprt(serv, newxpt);
trace_svc_xprt_accept(newxpt, serv->sv_name);
} else {
module_put(xprt->xpt_class->xcl_owner);
}
svc_xprt_received(xprt);
} else if (svc_xprt_reserve_slot(rqstp, xprt)) {
/* XPT_DATA|XPT_DEFERRED case: */
dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
rqstp, rqstp->rq_pool->sp_id, xprt,
kref_read(&xprt->xpt_ref));
rqstp->rq_deferred = svc_deferred_dequeue(xprt);
if (rqstp->rq_deferred)
len = svc_deferred_recv(rqstp);
else
len = xprt->xpt_ops->xpo_recvfrom(rqstp);
rqstp->rq_stime = ktime_get();
rqstp->rq_reserved = serv->sv_max_mesg;
atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
} else
svc_xprt_received(xprt);
out:
return len;
}
/*
* Receive the next request on any transport. This code is carefully
* organised not to touch any cachelines in the shared svc_serv
* structure, only cachelines in the local svc_pool.
*/
int svc_recv(struct svc_rqst *rqstp, long timeout)
{
struct svc_xprt *xprt = NULL;
struct svc_serv *serv = rqstp->rq_server;
int len, err;
err = svc_alloc_arg(rqstp);
if (err)
goto out;
try_to_freeze();
cond_resched();
err = -EINTR;
if (signalled() || kthread_should_stop())
goto out;
xprt = svc_get_next_xprt(rqstp, timeout);
if (IS_ERR(xprt)) {
err = PTR_ERR(xprt);
goto out;
}
len = svc_handle_xprt(rqstp, xprt);
/* No data, incomplete (TCP) read, or accept() */
err = -EAGAIN;
if (len <= 0)
goto out_release;
trace_svc_xdr_recvfrom(&rqstp->rq_arg);
clear_bit(XPT_OLD, &xprt->xpt_flags);
xprt->xpt_ops->xpo_secure_port(rqstp);
rqstp->rq_chandle.defer = svc_defer;
rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
if (serv->sv_stats)
serv->sv_stats->netcnt++;
return len;
out_release:
rqstp->rq_res.len = 0;
svc_xprt_release(rqstp);
out:
return err;
}
EXPORT_SYMBOL_GPL(svc_recv);
/*
* Drop request
*/
void svc_drop(struct svc_rqst *rqstp)
{
trace_svc_drop(rqstp);
svc_xprt_release(rqstp);
}
EXPORT_SYMBOL_GPL(svc_drop);
/*
* Return reply to client.
*/
int svc_send(struct svc_rqst *rqstp)
{
struct svc_xprt *xprt;
int len = -EFAULT;
struct xdr_buf *xb;
xprt = rqstp->rq_xprt;
if (!xprt)
goto out;
/* calculate over-all length */
xb = &rqstp->rq_res;
xb->len = xb->head[0].iov_len +
xb->page_len +
xb->tail[0].iov_len;
trace_svc_xdr_sendto(rqstp->rq_xid, xb);
trace_svc_stats_latency(rqstp);
len = xprt->xpt_ops->xpo_sendto(rqstp);
trace_svc_send(rqstp, len);
svc_xprt_release(rqstp);
if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
len = 0;
out:
return len;
}
/*
* Timer function to close old temporary transports, using
* a mark-and-sweep algorithm.
*/
static void svc_age_temp_xprts(struct timer_list *t)
{
struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
struct svc_xprt *xprt;
struct list_head *le, *next;
dprintk("svc_age_temp_xprts\n");
if (!spin_trylock_bh(&serv->sv_lock)) {
/* busy, try again 1 sec later */
dprintk("svc_age_temp_xprts: busy\n");
mod_timer(&serv->sv_temptimer, jiffies + HZ);
return;
}
list_for_each_safe(le, next, &serv->sv_tempsocks) {
xprt = list_entry(le, struct svc_xprt, xpt_list);
/* First time through, just mark it OLD. Second time
* through, close it. */
if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
continue;
if (kref_read(&xprt->xpt_ref) > 1 ||
test_bit(XPT_BUSY, &xprt->xpt_flags))
continue;
list_del_init(le);
set_bit(XPT_CLOSE, &xprt->xpt_flags);
dprintk("queuing xprt %p for closing\n", xprt);
/* a thread will dequeue and close it soon */
svc_xprt_enqueue(xprt);
}
spin_unlock_bh(&serv->sv_lock);
mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
}
/* Close temporary transports whose xpt_local matches server_addr immediately
* instead of waiting for them to be picked up by the timer.
*
* This is meant to be called from a notifier_block that runs when an ip
* address is deleted.
*/
void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
{
struct svc_xprt *xprt;
struct list_head *le, *next;
LIST_HEAD(to_be_closed);
spin_lock_bh(&serv->sv_lock);
list_for_each_safe(le, next, &serv->sv_tempsocks) {
xprt = list_entry(le, struct svc_xprt, xpt_list);
if (rpc_cmp_addr(server_addr, (struct sockaddr *)
&xprt->xpt_local)) {
dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
list_move(le, &to_be_closed);
}
}
spin_unlock_bh(&serv->sv_lock);
while (!list_empty(&to_be_closed)) {
le = to_be_closed.next;
list_del_init(le);
xprt = list_entry(le, struct svc_xprt, xpt_list);
set_bit(XPT_CLOSE, &xprt->xpt_flags);
set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
xprt);
svc_xprt_enqueue(xprt);
}
}
EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
static void call_xpt_users(struct svc_xprt *xprt)
{
struct svc_xpt_user *u;
spin_lock(&xprt->xpt_lock);
while (!list_empty(&xprt->xpt_users)) {
u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
list_del_init(&u->list);
u->callback(u);
}
spin_unlock(&xprt->xpt_lock);
}
/*
* Remove a dead transport
*/
static void svc_delete_xprt(struct svc_xprt *xprt)
{
struct svc_serv *serv = xprt->xpt_server;
struct svc_deferred_req *dr;
if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
return;
trace_svc_xprt_detach(xprt);
xprt->xpt_ops->xpo_detach(xprt);
if (xprt->xpt_bc_xprt)
xprt->xpt_bc_xprt->ops->close(xprt->xpt_bc_xprt);
spin_lock_bh(&serv->sv_lock);
list_del_init(&xprt->xpt_list);
WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
if (test_bit(XPT_TEMP, &xprt->xpt_flags))
serv->sv_tmpcnt--;
spin_unlock_bh(&serv->sv_lock);
while ((dr = svc_deferred_dequeue(xprt)) != NULL)
kfree(dr);
call_xpt_users(xprt);
svc_xprt_put(xprt);
}
/**
* svc_xprt_close - Close a client connection
* @xprt: transport to disconnect
*
*/
void svc_xprt_close(struct svc_xprt *xprt)
{
trace_svc_xprt_close(xprt);
set_bit(XPT_CLOSE, &xprt->xpt_flags);
if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
/* someone else will have to effect the close */
return;
/*
* We expect svc_close_xprt() to work even when no threads are
* running (e.g., while configuring the server before starting
* any threads), so if the transport isn't busy, we delete
* it ourself:
*/
svc_delete_xprt(xprt);
}
EXPORT_SYMBOL_GPL(svc_xprt_close);
static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
{
struct svc_xprt *xprt;
int ret = 0;
NFSD: Repair misuse of sv_lock in 5.10.16-rt30. [ This problem is in mainline, but only rt has the chops to be able to detect it. ] Lockdep reports a circular lock dependency between serv->sv_lock and softirq_ctl.lock on system shutdown, when using a kernel built with CONFIG_PREEMPT_RT=y, and a nfs mount exists. This is due to the definition of spin_lock_bh on rt: local_bh_disable(); rt_spin_lock(lock); which forces a softirq_ctl.lock -> serv->sv_lock dependency. This is not a problem as long as _every_ lock of serv->sv_lock is a: spin_lock_bh(&serv->sv_lock); but there is one of the form: spin_lock(&serv->sv_lock); This is what is causing the circular dependency splat. The spin_lock() grabs the lock without first grabbing softirq_ctl.lock via local_bh_disable. If later on in the critical region, someone does a local_bh_disable, we get a serv->sv_lock -> softirq_ctrl.lock dependency established. Deadlock. Fix is to make serv->sv_lock be locked with spin_lock_bh everywhere, no exceptions. [ OK ] Stopped target NFS client services. Stopping Logout off all iSCSI sessions on shutdown... Stopping NFS server and services... [ 109.442380] [ 109.442385] ====================================================== [ 109.442386] WARNING: possible circular locking dependency detected [ 109.442387] 5.10.16-rt30 #1 Not tainted [ 109.442389] ------------------------------------------------------ [ 109.442390] nfsd/1032 is trying to acquire lock: [ 109.442392] ffff994237617f60 ((softirq_ctrl.lock).lock){+.+.}-{2:2}, at: __local_bh_disable_ip+0xd9/0x270 [ 109.442405] [ 109.442405] but task is already holding lock: [ 109.442406] ffff994245cb00b0 (&serv->sv_lock){+.+.}-{0:0}, at: svc_close_list+0x1f/0x90 [ 109.442415] [ 109.442415] which lock already depends on the new lock. [ 109.442415] [ 109.442416] [ 109.442416] the existing dependency chain (in reverse order) is: [ 109.442417] [ 109.442417] -> #1 (&serv->sv_lock){+.+.}-{0:0}: [ 109.442421] rt_spin_lock+0x2b/0xc0 [ 109.442428] svc_add_new_perm_xprt+0x42/0xa0 [ 109.442430] svc_addsock+0x135/0x220 [ 109.442434] write_ports+0x4b3/0x620 [ 109.442438] nfsctl_transaction_write+0x45/0x80 [ 109.442440] vfs_write+0xff/0x420 [ 109.442444] ksys_write+0x4f/0xc0 [ 109.442446] do_syscall_64+0x33/0x40 [ 109.442450] entry_SYSCALL_64_after_hwframe+0x44/0xa9 [ 109.442454] [ 109.442454] -> #0 ((softirq_ctrl.lock).lock){+.+.}-{2:2}: [ 109.442457] __lock_acquire+0x1264/0x20b0 [ 109.442463] lock_acquire+0xc2/0x400 [ 109.442466] rt_spin_lock+0x2b/0xc0 [ 109.442469] __local_bh_disable_ip+0xd9/0x270 [ 109.442471] svc_xprt_do_enqueue+0xc0/0x4d0 [ 109.442474] svc_close_list+0x60/0x90 [ 109.442476] svc_close_net+0x49/0x1a0 [ 109.442478] svc_shutdown_net+0x12/0x40 [ 109.442480] nfsd_destroy+0xc5/0x180 [ 109.442482] nfsd+0x1bc/0x270 [ 109.442483] kthread+0x194/0x1b0 [ 109.442487] ret_from_fork+0x22/0x30 [ 109.442492] [ 109.442492] other info that might help us debug this: [ 109.442492] [ 109.442493] Possible unsafe locking scenario: [ 109.442493] [ 109.442493] CPU0 CPU1 [ 109.442494] ---- ---- [ 109.442495] lock(&serv->sv_lock); [ 109.442496] lock((softirq_ctrl.lock).lock); [ 109.442498] lock(&serv->sv_lock); [ 109.442499] lock((softirq_ctrl.lock).lock); [ 109.442501] [ 109.442501] *** DEADLOCK *** [ 109.442501] [ 109.442501] 3 locks held by nfsd/1032: [ 109.442503] #0: ffffffff93b49258 (nfsd_mutex){+.+.}-{3:3}, at: nfsd+0x19a/0x270 [ 109.442508] #1: ffff994245cb00b0 (&serv->sv_lock){+.+.}-{0:0}, at: svc_close_list+0x1f/0x90 [ 109.442512] #2: ffffffff93a81b20 (rcu_read_lock){....}-{1:2}, at: rt_spin_lock+0x5/0xc0 [ 109.442518] [ 109.442518] stack backtrace: [ 109.442519] CPU: 0 PID: 1032 Comm: nfsd Not tainted 5.10.16-rt30 #1 [ 109.442522] Hardware name: Supermicro X9DRL-3F/iF/X9DRL-3F/iF, BIOS 3.2 09/22/2015 [ 109.442524] Call Trace: [ 109.442527] dump_stack+0x77/0x97 [ 109.442533] check_noncircular+0xdc/0xf0 [ 109.442546] __lock_acquire+0x1264/0x20b0 [ 109.442553] lock_acquire+0xc2/0x400 [ 109.442564] rt_spin_lock+0x2b/0xc0 [ 109.442570] __local_bh_disable_ip+0xd9/0x270 [ 109.442573] svc_xprt_do_enqueue+0xc0/0x4d0 [ 109.442577] svc_close_list+0x60/0x90 [ 109.442581] svc_close_net+0x49/0x1a0 [ 109.442585] svc_shutdown_net+0x12/0x40 [ 109.442588] nfsd_destroy+0xc5/0x180 [ 109.442590] nfsd+0x1bc/0x270 [ 109.442595] kthread+0x194/0x1b0 [ 109.442600] ret_from_fork+0x22/0x30 [ 109.518225] nfsd: last server has exited, flushing export cache [ OK ] Stopped NFSv4 ID-name mapping service. [ OK ] Stopped GSSAPI Proxy Daemon. [ OK ] Stopped NFS Mount Daemon. [ OK ] Stopped NFS status monitor for NFSv2/3 locking.. Fixes: 719f8bcc883e ("svcrpc: fix xpt_list traversal locking on shutdown") Signed-off-by: Joe Korty <joe.korty@concurrent-rt.com> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
2021-02-26 22:38:20 +08:00
spin_lock_bh(&serv->sv_lock);
svcrpc: avoid memory-corruption on pool shutdown Socket callbacks use svc_xprt_enqueue() to add an xprt to a pool->sp_sockets list. In normal operation a server thread will later come along and take the xprt off that list. On shutdown, after all the threads have exited, we instead manually walk the sv_tempsocks and sv_permsocks lists to find all the xprt's and delete them. So the sp_sockets lists don't really matter any more. As a result, we've mostly just ignored them and hoped they would go away. Which has gotten us into trouble; witness for example ebc63e531cc6 "svcrpc: fix list-corrupting race on nfsd shutdown", the result of Ben Greear noticing that a still-running svc_xprt_enqueue() could re-add an xprt to an sp_sockets list just before it was deleted. The fix was to remove it from the list at the end of svc_delete_xprt(). But that only made corruption less likely--I can see nothing that prevents a svc_xprt_enqueue() from adding another xprt to the list at the same moment that we're removing this xprt from the list. In fact, despite the earlier xpo_detach(), I don't even see what guarantees that svc_xprt_enqueue() couldn't still be running on this xprt. So, instead, note that svc_xprt_enqueue() essentially does: lock sp_lock if XPT_BUSY unset add to sp_sockets unlock sp_lock So, if we do: set XPT_BUSY on every xprt. Empty every sp_sockets list, under the sp_socks locks. Then we're left knowing that the sp_sockets lists are all empty and will stay that way, since any svc_xprt_enqueue() will check XPT_BUSY under the sp_lock and see it set. And *then* we can continue deleting the xprt's. (Thanks to Jeff Layton for being correctly suspicious of this code....) Cc: Ben Greear <greearb@candelatech.com> Cc: Jeff Layton <jlayton@redhat.com> Cc: stable@kernel.org Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2011-11-30 06:00:26 +08:00
list_for_each_entry(xprt, xprt_list, xpt_list) {
if (xprt->xpt_net != net)
continue;
ret++;
set_bit(XPT_CLOSE, &xprt->xpt_flags);
svc_xprt_enqueue(xprt);
}
NFSD: Repair misuse of sv_lock in 5.10.16-rt30. [ This problem is in mainline, but only rt has the chops to be able to detect it. ] Lockdep reports a circular lock dependency between serv->sv_lock and softirq_ctl.lock on system shutdown, when using a kernel built with CONFIG_PREEMPT_RT=y, and a nfs mount exists. This is due to the definition of spin_lock_bh on rt: local_bh_disable(); rt_spin_lock(lock); which forces a softirq_ctl.lock -> serv->sv_lock dependency. This is not a problem as long as _every_ lock of serv->sv_lock is a: spin_lock_bh(&serv->sv_lock); but there is one of the form: spin_lock(&serv->sv_lock); This is what is causing the circular dependency splat. The spin_lock() grabs the lock without first grabbing softirq_ctl.lock via local_bh_disable. If later on in the critical region, someone does a local_bh_disable, we get a serv->sv_lock -> softirq_ctrl.lock dependency established. Deadlock. Fix is to make serv->sv_lock be locked with spin_lock_bh everywhere, no exceptions. [ OK ] Stopped target NFS client services. Stopping Logout off all iSCSI sessions on shutdown... Stopping NFS server and services... [ 109.442380] [ 109.442385] ====================================================== [ 109.442386] WARNING: possible circular locking dependency detected [ 109.442387] 5.10.16-rt30 #1 Not tainted [ 109.442389] ------------------------------------------------------ [ 109.442390] nfsd/1032 is trying to acquire lock: [ 109.442392] ffff994237617f60 ((softirq_ctrl.lock).lock){+.+.}-{2:2}, at: __local_bh_disable_ip+0xd9/0x270 [ 109.442405] [ 109.442405] but task is already holding lock: [ 109.442406] ffff994245cb00b0 (&serv->sv_lock){+.+.}-{0:0}, at: svc_close_list+0x1f/0x90 [ 109.442415] [ 109.442415] which lock already depends on the new lock. [ 109.442415] [ 109.442416] [ 109.442416] the existing dependency chain (in reverse order) is: [ 109.442417] [ 109.442417] -> #1 (&serv->sv_lock){+.+.}-{0:0}: [ 109.442421] rt_spin_lock+0x2b/0xc0 [ 109.442428] svc_add_new_perm_xprt+0x42/0xa0 [ 109.442430] svc_addsock+0x135/0x220 [ 109.442434] write_ports+0x4b3/0x620 [ 109.442438] nfsctl_transaction_write+0x45/0x80 [ 109.442440] vfs_write+0xff/0x420 [ 109.442444] ksys_write+0x4f/0xc0 [ 109.442446] do_syscall_64+0x33/0x40 [ 109.442450] entry_SYSCALL_64_after_hwframe+0x44/0xa9 [ 109.442454] [ 109.442454] -> #0 ((softirq_ctrl.lock).lock){+.+.}-{2:2}: [ 109.442457] __lock_acquire+0x1264/0x20b0 [ 109.442463] lock_acquire+0xc2/0x400 [ 109.442466] rt_spin_lock+0x2b/0xc0 [ 109.442469] __local_bh_disable_ip+0xd9/0x270 [ 109.442471] svc_xprt_do_enqueue+0xc0/0x4d0 [ 109.442474] svc_close_list+0x60/0x90 [ 109.442476] svc_close_net+0x49/0x1a0 [ 109.442478] svc_shutdown_net+0x12/0x40 [ 109.442480] nfsd_destroy+0xc5/0x180 [ 109.442482] nfsd+0x1bc/0x270 [ 109.442483] kthread+0x194/0x1b0 [ 109.442487] ret_from_fork+0x22/0x30 [ 109.442492] [ 109.442492] other info that might help us debug this: [ 109.442492] [ 109.442493] Possible unsafe locking scenario: [ 109.442493] [ 109.442493] CPU0 CPU1 [ 109.442494] ---- ---- [ 109.442495] lock(&serv->sv_lock); [ 109.442496] lock((softirq_ctrl.lock).lock); [ 109.442498] lock(&serv->sv_lock); [ 109.442499] lock((softirq_ctrl.lock).lock); [ 109.442501] [ 109.442501] *** DEADLOCK *** [ 109.442501] [ 109.442501] 3 locks held by nfsd/1032: [ 109.442503] #0: ffffffff93b49258 (nfsd_mutex){+.+.}-{3:3}, at: nfsd+0x19a/0x270 [ 109.442508] #1: ffff994245cb00b0 (&serv->sv_lock){+.+.}-{0:0}, at: svc_close_list+0x1f/0x90 [ 109.442512] #2: ffffffff93a81b20 (rcu_read_lock){....}-{1:2}, at: rt_spin_lock+0x5/0xc0 [ 109.442518] [ 109.442518] stack backtrace: [ 109.442519] CPU: 0 PID: 1032 Comm: nfsd Not tainted 5.10.16-rt30 #1 [ 109.442522] Hardware name: Supermicro X9DRL-3F/iF/X9DRL-3F/iF, BIOS 3.2 09/22/2015 [ 109.442524] Call Trace: [ 109.442527] dump_stack+0x77/0x97 [ 109.442533] check_noncircular+0xdc/0xf0 [ 109.442546] __lock_acquire+0x1264/0x20b0 [ 109.442553] lock_acquire+0xc2/0x400 [ 109.442564] rt_spin_lock+0x2b/0xc0 [ 109.442570] __local_bh_disable_ip+0xd9/0x270 [ 109.442573] svc_xprt_do_enqueue+0xc0/0x4d0 [ 109.442577] svc_close_list+0x60/0x90 [ 109.442581] svc_close_net+0x49/0x1a0 [ 109.442585] svc_shutdown_net+0x12/0x40 [ 109.442588] nfsd_destroy+0xc5/0x180 [ 109.442590] nfsd+0x1bc/0x270 [ 109.442595] kthread+0x194/0x1b0 [ 109.442600] ret_from_fork+0x22/0x30 [ 109.518225] nfsd: last server has exited, flushing export cache [ OK ] Stopped NFSv4 ID-name mapping service. [ OK ] Stopped GSSAPI Proxy Daemon. [ OK ] Stopped NFS Mount Daemon. [ OK ] Stopped NFS status monitor for NFSv2/3 locking.. Fixes: 719f8bcc883e ("svcrpc: fix xpt_list traversal locking on shutdown") Signed-off-by: Joe Korty <joe.korty@concurrent-rt.com> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
2021-02-26 22:38:20 +08:00
spin_unlock_bh(&serv->sv_lock);
return ret;
}
static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
{
svcrpc: avoid memory-corruption on pool shutdown Socket callbacks use svc_xprt_enqueue() to add an xprt to a pool->sp_sockets list. In normal operation a server thread will later come along and take the xprt off that list. On shutdown, after all the threads have exited, we instead manually walk the sv_tempsocks and sv_permsocks lists to find all the xprt's and delete them. So the sp_sockets lists don't really matter any more. As a result, we've mostly just ignored them and hoped they would go away. Which has gotten us into trouble; witness for example ebc63e531cc6 "svcrpc: fix list-corrupting race on nfsd shutdown", the result of Ben Greear noticing that a still-running svc_xprt_enqueue() could re-add an xprt to an sp_sockets list just before it was deleted. The fix was to remove it from the list at the end of svc_delete_xprt(). But that only made corruption less likely--I can see nothing that prevents a svc_xprt_enqueue() from adding another xprt to the list at the same moment that we're removing this xprt from the list. In fact, despite the earlier xpo_detach(), I don't even see what guarantees that svc_xprt_enqueue() couldn't still be running on this xprt. So, instead, note that svc_xprt_enqueue() essentially does: lock sp_lock if XPT_BUSY unset add to sp_sockets unlock sp_lock So, if we do: set XPT_BUSY on every xprt. Empty every sp_sockets list, under the sp_socks locks. Then we're left knowing that the sp_sockets lists are all empty and will stay that way, since any svc_xprt_enqueue() will check XPT_BUSY under the sp_lock and see it set. And *then* we can continue deleting the xprt's. (Thanks to Jeff Layton for being correctly suspicious of this code....) Cc: Ben Greear <greearb@candelatech.com> Cc: Jeff Layton <jlayton@redhat.com> Cc: stable@kernel.org Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2011-11-30 06:00:26 +08:00
struct svc_pool *pool;
struct svc_xprt *xprt;
struct svc_xprt *tmp;
svcrpc: avoid memory-corruption on pool shutdown Socket callbacks use svc_xprt_enqueue() to add an xprt to a pool->sp_sockets list. In normal operation a server thread will later come along and take the xprt off that list. On shutdown, after all the threads have exited, we instead manually walk the sv_tempsocks and sv_permsocks lists to find all the xprt's and delete them. So the sp_sockets lists don't really matter any more. As a result, we've mostly just ignored them and hoped they would go away. Which has gotten us into trouble; witness for example ebc63e531cc6 "svcrpc: fix list-corrupting race on nfsd shutdown", the result of Ben Greear noticing that a still-running svc_xprt_enqueue() could re-add an xprt to an sp_sockets list just before it was deleted. The fix was to remove it from the list at the end of svc_delete_xprt(). But that only made corruption less likely--I can see nothing that prevents a svc_xprt_enqueue() from adding another xprt to the list at the same moment that we're removing this xprt from the list. In fact, despite the earlier xpo_detach(), I don't even see what guarantees that svc_xprt_enqueue() couldn't still be running on this xprt. So, instead, note that svc_xprt_enqueue() essentially does: lock sp_lock if XPT_BUSY unset add to sp_sockets unlock sp_lock So, if we do: set XPT_BUSY on every xprt. Empty every sp_sockets list, under the sp_socks locks. Then we're left knowing that the sp_sockets lists are all empty and will stay that way, since any svc_xprt_enqueue() will check XPT_BUSY under the sp_lock and see it set. And *then* we can continue deleting the xprt's. (Thanks to Jeff Layton for being correctly suspicious of this code....) Cc: Ben Greear <greearb@candelatech.com> Cc: Jeff Layton <jlayton@redhat.com> Cc: stable@kernel.org Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2011-11-30 06:00:26 +08:00
int i;
for (i = 0; i < serv->sv_nrpools; i++) {
pool = &serv->sv_pools[i];
spin_lock_bh(&pool->sp_lock);
list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
if (xprt->xpt_net != net)
continue;
svcrpc: avoid memory-corruption on pool shutdown Socket callbacks use svc_xprt_enqueue() to add an xprt to a pool->sp_sockets list. In normal operation a server thread will later come along and take the xprt off that list. On shutdown, after all the threads have exited, we instead manually walk the sv_tempsocks and sv_permsocks lists to find all the xprt's and delete them. So the sp_sockets lists don't really matter any more. As a result, we've mostly just ignored them and hoped they would go away. Which has gotten us into trouble; witness for example ebc63e531cc6 "svcrpc: fix list-corrupting race on nfsd shutdown", the result of Ben Greear noticing that a still-running svc_xprt_enqueue() could re-add an xprt to an sp_sockets list just before it was deleted. The fix was to remove it from the list at the end of svc_delete_xprt(). But that only made corruption less likely--I can see nothing that prevents a svc_xprt_enqueue() from adding another xprt to the list at the same moment that we're removing this xprt from the list. In fact, despite the earlier xpo_detach(), I don't even see what guarantees that svc_xprt_enqueue() couldn't still be running on this xprt. So, instead, note that svc_xprt_enqueue() essentially does: lock sp_lock if XPT_BUSY unset add to sp_sockets unlock sp_lock So, if we do: set XPT_BUSY on every xprt. Empty every sp_sockets list, under the sp_socks locks. Then we're left knowing that the sp_sockets lists are all empty and will stay that way, since any svc_xprt_enqueue() will check XPT_BUSY under the sp_lock and see it set. And *then* we can continue deleting the xprt's. (Thanks to Jeff Layton for being correctly suspicious of this code....) Cc: Ben Greear <greearb@candelatech.com> Cc: Jeff Layton <jlayton@redhat.com> Cc: stable@kernel.org Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2011-11-30 06:00:26 +08:00
list_del_init(&xprt->xpt_ready);
spin_unlock_bh(&pool->sp_lock);
return xprt;
svcrpc: avoid memory-corruption on pool shutdown Socket callbacks use svc_xprt_enqueue() to add an xprt to a pool->sp_sockets list. In normal operation a server thread will later come along and take the xprt off that list. On shutdown, after all the threads have exited, we instead manually walk the sv_tempsocks and sv_permsocks lists to find all the xprt's and delete them. So the sp_sockets lists don't really matter any more. As a result, we've mostly just ignored them and hoped they would go away. Which has gotten us into trouble; witness for example ebc63e531cc6 "svcrpc: fix list-corrupting race on nfsd shutdown", the result of Ben Greear noticing that a still-running svc_xprt_enqueue() could re-add an xprt to an sp_sockets list just before it was deleted. The fix was to remove it from the list at the end of svc_delete_xprt(). But that only made corruption less likely--I can see nothing that prevents a svc_xprt_enqueue() from adding another xprt to the list at the same moment that we're removing this xprt from the list. In fact, despite the earlier xpo_detach(), I don't even see what guarantees that svc_xprt_enqueue() couldn't still be running on this xprt. So, instead, note that svc_xprt_enqueue() essentially does: lock sp_lock if XPT_BUSY unset add to sp_sockets unlock sp_lock So, if we do: set XPT_BUSY on every xprt. Empty every sp_sockets list, under the sp_socks locks. Then we're left knowing that the sp_sockets lists are all empty and will stay that way, since any svc_xprt_enqueue() will check XPT_BUSY under the sp_lock and see it set. And *then* we can continue deleting the xprt's. (Thanks to Jeff Layton for being correctly suspicious of this code....) Cc: Ben Greear <greearb@candelatech.com> Cc: Jeff Layton <jlayton@redhat.com> Cc: stable@kernel.org Signed-off-by: J. Bruce Fields <bfields@redhat.com>
2011-11-30 06:00:26 +08:00
}
spin_unlock_bh(&pool->sp_lock);
}
return NULL;
}
static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
{
struct svc_xprt *xprt;
while ((xprt = svc_dequeue_net(serv, net))) {
set_bit(XPT_CLOSE, &xprt->xpt_flags);
svc_delete_xprt(xprt);
}
}
/**
* svc_xprt_destroy_all - Destroy transports associated with @serv
* @serv: RPC service to be shut down
* @net: target network namespace
*
* Server threads may still be running (especially in the case where the
* service is still running in other network namespaces).
*
* So we shut down sockets the same way we would on a running server, by
* setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
* the close. In the case there are no such other threads,
* threads running, svc_clean_up_xprts() does a simple version of a
* server's main event loop, and in the case where there are other
* threads, we may need to wait a little while and then check again to
* see if they're done.
*/
void svc_xprt_destroy_all(struct svc_serv *serv, struct net *net)
{
int delay = 0;
while (svc_close_list(serv, &serv->sv_permsocks, net) +
svc_close_list(serv, &serv->sv_tempsocks, net)) {
svc_clean_up_xprts(serv, net);
msleep(delay++);
}
}
EXPORT_SYMBOL_GPL(svc_xprt_destroy_all);
/*
* Handle defer and revisit of requests
*/
static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
{
struct svc_deferred_req *dr =
container_of(dreq, struct svc_deferred_req, handle);
struct svc_xprt *xprt = dr->xprt;
spin_lock(&xprt->xpt_lock);
set_bit(XPT_DEFERRED, &xprt->xpt_flags);
if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
spin_unlock(&xprt->xpt_lock);
trace_svc_defer_drop(dr);
svc_xprt_put(xprt);
kfree(dr);
return;
}
dr->xprt = NULL;
list_add(&dr->handle.recent, &xprt->xpt_deferred);
spin_unlock(&xprt->xpt_lock);
trace_svc_defer_queue(dr);
svc_xprt_enqueue(xprt);
svc_xprt_put(xprt);
}
/*
* Save the request off for later processing. The request buffer looks
* like this:
*
* <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
*
* This code can only handle requests that consist of an xprt-header
* and rpc-header.
*/
static struct cache_deferred_req *svc_defer(struct cache_req *req)
{
struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
struct svc_deferred_req *dr;
if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
return NULL; /* if more than a page, give up FIXME */
if (rqstp->rq_deferred) {
dr = rqstp->rq_deferred;
rqstp->rq_deferred = NULL;
} else {
size_t skip;
size_t size;
/* FIXME maybe discard if size too large */
size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
dr = kmalloc(size, GFP_KERNEL);
if (dr == NULL)
return NULL;
dr->handle.owner = rqstp->rq_server;
dr->prot = rqstp->rq_prot;
memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
dr->addrlen = rqstp->rq_addrlen;
dr->daddr = rqstp->rq_daddr;
dr->argslen = rqstp->rq_arg.len >> 2;
dr->xprt_ctxt = rqstp->rq_xprt_ctxt;
rqstp->rq_xprt_ctxt = NULL;
/* back up head to the start of the buffer and copy */
skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
dr->argslen << 2);
}
trace_svc_defer(rqstp);
svc_xprt_get(rqstp->rq_xprt);
dr->xprt = rqstp->rq_xprt;
__set_bit(RQ_DROPME, &rqstp->rq_flags);
dr->handle.revisit = svc_revisit;
return &dr->handle;
}
/*
* recv data from a deferred request into an active one
*/
static noinline int svc_deferred_recv(struct svc_rqst *rqstp)
{
struct svc_deferred_req *dr = rqstp->rq_deferred;
trace_svc_defer_recv(dr);
/* setup iov_base past transport header */
rqstp->rq_arg.head[0].iov_base = dr->args;
/* The iov_len does not include the transport header bytes */
rqstp->rq_arg.head[0].iov_len = dr->argslen << 2;
rqstp->rq_arg.page_len = 0;
/* The rq_arg.len includes the transport header bytes */
rqstp->rq_arg.len = dr->argslen << 2;
rqstp->rq_prot = dr->prot;
memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
rqstp->rq_addrlen = dr->addrlen;
/* Save off transport header len in case we get deferred again */
rqstp->rq_daddr = dr->daddr;
rqstp->rq_respages = rqstp->rq_pages;
rqstp->rq_xprt_ctxt = dr->xprt_ctxt;
svc_xprt_received(rqstp->rq_xprt);
return dr->argslen << 2;
}
static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
{
struct svc_deferred_req *dr = NULL;
if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
return NULL;
spin_lock(&xprt->xpt_lock);
if (!list_empty(&xprt->xpt_deferred)) {
dr = list_entry(xprt->xpt_deferred.next,
struct svc_deferred_req,
handle.recent);
list_del_init(&dr->handle.recent);
} else
clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
spin_unlock(&xprt->xpt_lock);
return dr;
}
/**
* svc_find_xprt - find an RPC transport instance
* @serv: pointer to svc_serv to search
* @xcl_name: C string containing transport's class name
* @net: owner net pointer
* @af: Address family of transport's local address
* @port: transport's IP port number
*
* Return the transport instance pointer for the endpoint accepting
* connections/peer traffic from the specified transport class,
* address family and port.
*
* Specifying 0 for the address family or port is effectively a
* wild-card, and will result in matching the first transport in the
* service's list that has a matching class name.
*/
struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
struct net *net, const sa_family_t af,
const unsigned short port)
{
struct svc_xprt *xprt;
struct svc_xprt *found = NULL;
/* Sanity check the args */
if (serv == NULL || xcl_name == NULL)
return found;
spin_lock_bh(&serv->sv_lock);
list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
if (xprt->xpt_net != net)
continue;
if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
continue;
if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
continue;
if (port != 0 && port != svc_xprt_local_port(xprt))
continue;
found = xprt;
svc_xprt_get(xprt);
break;
}
spin_unlock_bh(&serv->sv_lock);
return found;
}
EXPORT_SYMBOL_GPL(svc_find_xprt);
static int svc_one_xprt_name(const struct svc_xprt *xprt,
char *pos, int remaining)
{
int len;
len = snprintf(pos, remaining, "%s %u\n",
xprt->xpt_class->xcl_name,
svc_xprt_local_port(xprt));
if (len >= remaining)
return -ENAMETOOLONG;
return len;
}
/**
* svc_xprt_names - format a buffer with a list of transport names
* @serv: pointer to an RPC service
* @buf: pointer to a buffer to be filled in
* @buflen: length of buffer to be filled in
*
* Fills in @buf with a string containing a list of transport names,
* each name terminated with '\n'.
*
* Returns positive length of the filled-in string on success; otherwise
* a negative errno value is returned if an error occurs.
*/
int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
{
struct svc_xprt *xprt;
int len, totlen;
char *pos;
/* Sanity check args */
if (!serv)
return 0;
spin_lock_bh(&serv->sv_lock);
pos = buf;
totlen = 0;
list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
len = svc_one_xprt_name(xprt, pos, buflen - totlen);
if (len < 0) {
*buf = '\0';
totlen = len;
}
if (len <= 0)
break;
pos += len;
totlen += len;
}
spin_unlock_bh(&serv->sv_lock);
return totlen;
}
EXPORT_SYMBOL_GPL(svc_xprt_names);
/*----------------------------------------------------------------------------*/
static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
{
unsigned int pidx = (unsigned int)*pos;
struct svc_serv *serv = m->private;
dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
if (!pidx)
return SEQ_START_TOKEN;
return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
}
static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
{
struct svc_pool *pool = p;
struct svc_serv *serv = m->private;
dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
if (p == SEQ_START_TOKEN) {
pool = &serv->sv_pools[0];
} else {
unsigned int pidx = (pool - &serv->sv_pools[0]);
if (pidx < serv->sv_nrpools-1)
pool = &serv->sv_pools[pidx+1];
else
pool = NULL;
}
++*pos;
return pool;
}
static void svc_pool_stats_stop(struct seq_file *m, void *p)
{
}
static int svc_pool_stats_show(struct seq_file *m, void *p)
{
struct svc_pool *pool = p;
if (p == SEQ_START_TOKEN) {
seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
return 0;
}
seq_printf(m, "%u %lu %lu %lu %lu\n",
pool->sp_id,
(unsigned long)atomic_long_read(&pool->sp_stats.packets),
pool->sp_stats.sockets_queued,
(unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
(unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
return 0;
}
static const struct seq_operations svc_pool_stats_seq_ops = {
.start = svc_pool_stats_start,
.next = svc_pool_stats_next,
.stop = svc_pool_stats_stop,
.show = svc_pool_stats_show,
};
int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
{
int err;
err = seq_open(file, &svc_pool_stats_seq_ops);
if (!err)
((struct seq_file *) file->private_data)->private = serv;
return err;
}
EXPORT_SYMBOL(svc_pool_stats_open);
/*----------------------------------------------------------------------------*/