tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

need help ! stale vnode pointer in ufs_fhtovp

I'm tracking down an issue where ufs_fhtovp() gets back a clean vnode from
ufs_ihashget(), so VTOI(nvp) is NULL and bad things happens.
THis is easily reproduced on a NFS server by running bonnie++ on a NFS
client, and rm'ing bonnie++'s file directly on the server. I think you can
get the same by removing the files from a second NFS client.
I think the problem is that vget() can sleep, and vclean can have
VOP_INACTIVE()/VOP_RECLAIM() the vnode before vget() returns.

I can reproduce this on a Xen DOMU as NFS server, so it's not related to
SMP or preemption. I don't know if this race can affect other users of
vget() but ufs_ihashget().

I tried this patch:

Index: ufs_ihash.c
RCS file: /cvsroot/src/sys/ufs/ufs/ufs_ihash.c,v
retrieving revision 1.26
diff -u -p -u -r1.26 ufs_ihash.c
--- ufs_ihash.c 5 May 2008 17:11:17 -0000       1.26
+++ ufs_ihash.c 8 Sep 2009 13:03:50 -0000
@@ -152,6 +152,14 @@ ufs_ihashget(dev_t dev, ino_t inum, int 
                                if (vget(vp, flags | LK_INTERLOCK))
                                        goto loop;
+                               if (VTOI(vp) != ip ||
+                                   ip->i_number != inum || ip->i_dev != dev) {
+                                       /* lost race against vclean() */
+                                       printf("ufs_ihashget lost race vp 
%p\n", vp);
+                                       vfs_vnode_print(vp, 1, printf);
+                                       vput(vp);
+                                       goto loop;
+                               }
                        return (vp);

But it doens't work. I get:
ufs_ihashget lost race vp 0xffffa0000439cc18
OBJECT 0xffffa0000439cc18: locked=0, pgops=0xffffffff803ea080, npages=0, refs=1
  PAGES <pg,offset>:

mp 0xffffa00003867008 numoutput 0 size 0x0 writesize 0x0
data 0x0 writecount 0 holdcnt 0
tag UNKNOWN(0) type VREG(1) mount 0xffffa00003867008 typedata 0x0
v_lock 0xffffa0000439cd20 v_vnlock 0xffffa0000439cd20
clean bufs:
dirty bufs:
Reader / writer lock error: rw_destroy: assertion failed: (rw->rw_owner & 
~RW_DEBUG) == 0

lock address : 0xffffa0000439cd20 type     :     sleep/adaptive
initialized  : 0xffffffff8032bb26
shared holds :                  0 exclusive:                  1
shares wanted:                  0 exclusive:                  0
current cpu  :                  0 last held:                  0
current lwp  : 0xffffa000045ce7e0 last held: 0xffffa000045ce7e0
last locked  : 0xffffffff803299a6 unlocked : 0xffffffff8032998a
owner/count  : 0xffffa000045ce7e0 flags    : 0x000000000000000c

Turnstile chain at 0xffffffff80612600.
=> No active turnstile for this lock.

fatal breakpoint trap in supervisor mode
trap type 1 code 0 rip ffffffff8037bca5 cs e030 rflags 246 cr2  7f7ffdb9f950 
cpl 0 rsp ffffa00004611460
Stopped in pid 205.3 (nfsd) at  netbsd:breakpoint+0x5:  leave
breakpoint() at netbsd:breakpoint+0x5
panic() at netbsd:panic+0x242
lockdebug_abort1() at netbsd:lockdebug_abort1+0xd3
rw_destroy() at netbsd:rw_destroy+0x46
vnfree() at netbsd:vnfree+0x6e
vrelel() at netbsd:vrelel+0x388
ufs_ihashget() at netbsd:ufs_ihashget+0xd2
ffs_vget() at netbsd:ffs_vget+0xc1
ufs_fhtovp() at netbsd:ufs_fhtovp+0x1f
ffs_fhtovp() at netbsd:ffs_fhtovp+0x58
nfsrv_fhtovp() at netbsd:nfsrv_fhtovp+0xa8
nfsrv_write() at netbsd:nfsrv_write+0x495
nfssvc_nfsd() at netbsd:nfssvc_nfsd+0x41b
sys_nfssvc() at netbsd:sys_nfssvc+0x28f
syscall() at netbsd:syscall+0xb4

When this doesn't panic the nfsd server hangs on tstile, waiting for a lock
whose exclusive owner doens't care any more.
It looks like vput() didn't release the lock ?

Remplacing the vput() with vlockmgr(vp->v_vnlock, LK_RELEASE); vrele(vp);
works (the server doesn't panic, the client gets ESTALE),
but I don't understand why vput() doesn't work.
Also I have no idea if this is the correct fix, or if something needs to be
done in vget. Can anyone help me with this ? Having such a bug affecting
the NFS server is bad ...

Manuel Bouyer <>
     NetBSD: 26 ans d'experience feront toujours la difference

Home | Main Index | Thread Index | Old Index