Package: sshfs
Version: 2.4-1
Severity: important
Tags: patch
--- Please enter the report below this line. ---
In the event of a network failure, sshfs can become
network-starved, causing programs accessing the sshfs
mount to enter uninterruptible sleep. The solution is
to kill the sshfs process; in contrast to some user
expectations, "fusermount -u /mountpoint" does not
work, since the mountpoint is still busy.
Killing sshfs is annoying at best, and impossible in some
cases (e.g., if $HOME is an sshfs, all new shells will
immediately enter uninterruptible sleep). This scenario
may be related to [1], but that bug appears to also be
associated with an issue on the kernel side.
The attached patch addresses this issue by providing a
new configuration option "-oread_timeout=N" which causes
sshfs to die (or attempt to remount if combined with
"-oremount") if its network connection becomes starved
for N seconds (in reality, the delay may be as long as
2N; eliminating this factor of two would complicate this
patch).
This simple patch protects the critical read() call to the
network socket with a select(). The call must fail only in
the event that sshfs is actually expecting network input,
which is determined by checking if the request tab
"sshfs.reqtab" is nonempty.
I have set a default value of "0" for this option, since
that changes things the least (i.e., it disables the
timeout). It may be best for this default to be nonzero.
Best,
Antonio
[1] http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=571005
P.S. I have attempted to report this bug to upstream,
fuse-ss...@lists.sourceforge.net, but that mailing
list seems to be dead and may have rejected my email.
Hopefully someone else has a better relationship with
upstream and can contact them.
--- System information. ---
Architecture: amd64
Kernel: Linux 3.9-1-amd64
Debian Release: jessie/sid
451 unstable www.deb-multimedia.org
451 unstable ftp.us.debian.org
399 testing ftp.us.debian.org
200 experimental www.deb-multimedia.org
200 experimental ftp.us.debian.org
150 stable dl.google.com
--- Package information. ---
Depends (Version) | Installed
=================================-+-===============
libc6 (>= 2.4) |
libfuse2 (>= 2.8.1) |
libglib2.0-0 (>= 2.12.0) |
fuse (>= 2.7) |
openssh-client (>= 1:4.6p1-1) |
Package's Recommends field is empty.
Package's Suggests field is empty.
--- sshfs.c.old 2012-03-08 01:34:39.000000000 -0800
+++ sshfs.c.new 2013-08-20 21:11:38.274865000 -0700
@@ -219,6 +219,7 @@
int reconnect;
int delay_connect;
int slave;
+ int read_timeout;
char *host;
char *base_path;
GHashTable *reqtab;
@@ -358,6 +359,7 @@
SSHFS_OPT("password_stdin", password_stdin, 1),
SSHFS_OPT("delay_connect", delay_connect, 1),
SSHFS_OPT("slave", slave, 1),
+ SSHFS_OPT("read_timeout=%u", read_timeout, 0),
FUSE_OPT_KEY("-p ", KEY_PORT),
FUSE_OPT_KEY("-C", KEY_COMPRESS),
@@ -1220,7 +1222,34 @@
int res;
uint8_t *p = buf->p;
size_t size = buf->size;
+ unsigned int outstanding_requests = 0;
+
+ struct timeval timeout;
+ fd_set fdset;
+
while (size) {
+ if (sshfs.read_timeout > 0) {
+ FD_ZERO(&fdset);
+ FD_SET(sshfs.rfd, &fdset);
+
+ pthread_mutex_lock(&sshfs.lock);
+ outstanding_requests = (unsigned int)
+ g_hash_table_size(sshfs.reqtab);
+ pthread_mutex_unlock(&sshfs.lock);
+
+ timeout.tv_sec = sshfs.read_timeout;
+ timeout.tv_usec = 0;
+
+ if (select(FD_SETSIZE, &fdset, NULL, NULL, &timeout)<1) {
+ if (outstanding_requests>0) {
+ fprintf(stderr, "wait exceeded read_timeout\n");
+ return -1;
+ } else {
+ continue;
+ }
+ }
+ }
+
res = read(sshfs.rfd, p, size);
if (res == -1) {
perror("read");
@@ -3190,6 +3219,7 @@
" -o sftp_server=SERV path to sftp server or subsystem (default: sftp)\n"
" -o directport=PORT directly connect to PORT bypassing ssh\n"
" -o slave communicate over stdin and stdout bypassing network\n"
+" -o read_timeout=N unmount if a read takes longer than N seconds\n"
" -o transform_symlinks transform absolute symlinks to relative\n"
" -o follow_symlinks follow symlinks on the server\n"
" -o no_check_root don't check for existence of 'dir' on server\n"
