Re: kern/37881: repeatable crash in filesystem when running MP

[Antti Kantee <pooka%cs.hut.fi@localhost>]

The following reply was made to PR kern/37881; it has been noted by GNATS.

From: Antti Kantee <pooka%cs.hut.fi@localhost>
To: dieter.NetBSD%pandora.be@localhost
Cc: gnats-bugs%NetBSD.org@localhost
Subject: Re: kern/37881: repeatable crash in filesystem when running MP
Date: Sun, 27 Jan 2008 15:06:20 +0200

 On Sun Jan 27 2008 at 12:40:00 +0000, dieter.NetBSD%pandora.be@localhost wrote:
 >      Since upgrading my system from 4.99.35 to 4.99.49 and then
 >      4.99.50, I can reproducibly crash it by copying some files
 >      with nautilus from a windows share to my local disk. The
 >      kernel drops in ddb after a uvm_fault with a backtrace like
 >      this:
 > 
 > uvm_fault(0xcd8b1004, 0, 1) -> 0xe
 > knote(cefe5be4,6,0,14c90000,0) at netbsd:knote+0x27
 > ffs_write(cdccbc04,0,c0483fa0,cefe5b20,10002) at netbsd:ffs_write+0x53d
 > VOP_WRITE(cefe5b20,cdccbc7c,10,cdb31540,c0579a84) at netbsd:VOP_WRITE+0x80
 > vn_write(cea71414,cea71440,cdccbc7c,cdb31540,1) at netbsd:vn_write+0xcc
 > dofilewrite(16,cea71414,ba226000,10000,cea71440) at netbsd:dofilewrite+0x75
 > sys_write(cef6eba0,cdccbd00,cdccbd28,ba235000,c037cab5) at 
 > netbsd:sys_write+0x9c
 > syscall(cdccbd48,b3,ba2600ab,ba40001f,ba5f001f) at netbsd:syscall+0xb5
 > 
 >      This does not happen if I disable one of the 2 cores with
 >      cpuctl.
 
 Hmm.  How are you copying the files?  Many at a time?  fd 16 looks
 semi-suspicious.
 
 > #25 0xc02c6727 in knote (list=0xcefe5be4, hint=6)
 >     at /usr/src/sys/kern/kern_event.c:1301
 
 print *list
 print *kn
 
 > #26 0xc025b9ed in ffs_write (v=0xcdccbc04)
 >     at /usr/src/sys/ufs/ufs/ufs_readwrite.c:507
 
 print *uio
 print *vp
 
 > #29 0xc0302cc5 in dofilewrite (fd=22, fp=0xcea71414, buf=0xba226000, 
 >     nbyte=65536, offset=0xcea71440, flags=1, retval=0xcdccbd28)
 >     at /usr/src/sys/kern/sys_generic.c:392
 
 print *offset
 
 print numvnodes
 print desiredvnodes
 
 >      Copy some (large?) files around with nautilus, possibly
 >      from an smb share on a system with more than one CPU active?
 
 Does it happen when you start copying or when you've been copying for
 a while or what?  Be as specific as you can.
 
 
 .. actually, if it's easy to repeat, try this patch, but do the above
 also.  I'm mostly guessing here, as I can't quite wrap my head around
 why this happens now and not before.
 
 Index: vfs_subr.c
 ===================================================================
 RCS file: /cvsroot/src/sys/kern/vfs_subr.c,v
 retrieving revision 1.323
 diff -p -u -r1.323 vfs_subr.c
 --- vfs_subr.c 26 Jan 2008 22:53:10 -0000      1.323
 +++ vfs_subr.c 27 Jan 2008 13:05:54 -0000
 @@ -263,6 +263,7 @@ try_nextlist:
                        spec_node_destroy(vp);
                }
                vp->v_type = VNON;
 +              SLIST_INIT(&vp->v_klist);
        } else {
                /*
                 * Don't return to freelist - the holder of the last
 
 -- 
 Antti Kantee <pooka%iki.fi@localhost>                     Of course he runs 
NetBSD
 http://www.iki.fi/pooka/                          http://www.NetBSD.org/
     "la qualité la plus indispensable du cuisinier est l'exactitude"