Re: UFS snapshot deadlocks

At Fri, 4 Jun 2004 15:06:55 +0000 (UTC),
Frode Nordahl wrote:
> tty1:
> while (1)
> 	ls -la /usr/.snap
> end
> 
> tty2:
> dump -0af /some/where -C 32 -L /dev/ad0s1f

I did diagnose about this.

L1. ls(1) issues lstat(2) during walking inside directory.
L2. lstat(2) calls namei().
L3. namei() tries to lock VREG /usr/.snap/dump_snapshot after locking
    VDIR /usr/.snap.

D1. dump(8) calls mount(2) via mksnap_ffs(8).
D2. mount(2) calls ffs_snapshot().
D3. ffs_snapshot() locks /usr/.snap/dump_snapshot after VOP_CREATE().
D4. ffs_snapshot() searches active but unlinked files from
    mnt_nvnodelist.
D5. when it founds /usr/.snap vnode, tries to lock it.

My question is:

1. should ordering of such locks be parent dir vnode first, and file
   in that dir vnode second?
2. is comparing vnode pointer like this to skip /usr/.snap safe?

Index: ffs_snapshot.c
===================================================================
RCS file: /home/ncvs/src/sys/ufs/ffs/ffs_snapshot.c,v
retrieving revision 1.81
diff -u -r1.81 ffs_snapshot.c
--- ffs_snapshot.c	16 Jun 2004 00:26:30 -0000	1.81
+++ ffs_snapshot.c	16 Jun 2004 03:34:44 -0000
_at__at_ -424,6 +424,11 _at__at_
 			MNT_ILOCK(mp);
 			continue;
 		}
+		if (xvp == nd.ni_dvp) {
+			VI_UNLOCK(xvp);
+			MNT_ILOCK(mp);
+			continue;
+		}
 		if (vn_lock(xvp, LK_EXCLUSIVE | LK_INTERLOCK, td) != 0) {
 			MNT_ILOCK(mp);
 			goto loop;

I did repeated test like as Frode did, but I cannot see any deadlocks
after this patch.


-- 
Jun Kuriyama <kuriyama_at_imgsrc.co.jp> // IMG SRC, Inc.
             <kuriyama_at_FreeBSD.org> // FreeBSD Project