Another bug? linux/fs/nfs/write.c

[Trond Myklebust]

Linus,
  The following mail from Matt describes a problem in the NFS
read/write code that can cause a vicious hang. Obvious patch to fix it
is attached...

Cheers,
   Trond

>>>>> " " == Matt Chapman <[e-mail]> writes:

     > Trond, Here's another bug which seems to be causing crashes.

     > nfs_update_request keeps calling nfs_wait_on_request until the
     > request can be locked.  Presumably it's relying on
     > nfs_wait_on_request to schedule, and hence run rpciod to
     > retransmit any lost requests.  However, if the fs is mounted
     > with the intr option and a signal has arrived, then
     > nfs_wait_on_request (ultimately, wait_event_interruptible)
     > returns immediately with -ERESTARTSYS, and the loop spins.

     > Presumably nfs_update_request needs to check the return value
     > of nfs_wait_on_request and return right up the chain on
     > -ERESTARTSYS.  However, since I don't know this code very well,
     > I'd prefer it if you had a look at it and gave some advice.

diff -u --recursive --new-file linux-2.4.5-fixes/fs/nfs/write.c linux-2.4.5-write/fs/nfs/write.c
--- linux-2.4.5-fixes/fs/nfs/write.c	Mon May 21 11:34:51 2001
+++ linux-2.4.5-write/fs/nfs/write.c	Mon May 21 13:18:47 2001
@@ -863,9 +863,12 @@
 		req = _nfs_find_request(inode, page);
 		if (req) {
 			if (!nfs_lock_request(req)) {
+				int error;
 				spin_unlock(&nfs_wreq_lock);
-				nfs_wait_on_request(req);
+				error = nfs_wait_on_request(req);
 				nfs_release_request(req);
+				if (error < 0)
+					return ERR_PTR(error);
 				continue;
 			}
 			spin_unlock(&nfs_wreq_lock);