Re: contemporary -current panic: locking against myself

On Fri, Aug 03, 2007 at 12:20:19PM +0200, Pawel Jakub Dawidek wrote:
> On Thu, Aug 02, 2007 at 03:58:26PM +0400, Dmitry Morozovsky wrote:
> > 
> > Hi there colleagues,
> > 
> > FreeBSD/i386 on Athlon X2, HEAD without WITNESS. 4G of RAM. tmpfs used for 
> > 'make release'.
> > 
> > 
> > panic: lockmgr: locking against myself
> > cpuid = 0
> > KDB: enter: panic
> > [thread pid 19396 tid 100245 ]
> > Stopped at      kdb_enter+0x32: leave
> > 
> > db> tr  
> > Tracing pid 19396 tid 100245 td 0xce194220
> > kdb_enter(c066f664,0,c066dca9,e92799cc,0,...) at kdb_enter+0x32
> > panic(c066dca9,e92799dc,c0559cc7,e9279ac0,ca2f7770,...) at panic+0x124
> > _lockmgr(ca2f77c8,3002,ca2f77f8,ce194220,c0675afc,...) at _lockmgr+0x401
> > vop_stdlock(e9279a5c,ce194220,3002,ca2f7770,e9279a80,...) at vop_stdlock+0x40
> > VOP_LOCK1_APV(d06417e0,e9279a5c,e9279bc0,0,c8d00330,...) at VOP_LOCK1_APV+0x46
> > _vn_lock(ca2f7770,3002,ce194220,c0675afc,7f3,...) at _vn_lock+0x166
> > vget(ca2f7770,1000,ce194220,0,e9279b98,...) at vget+0x114
> > vm_object_reference(d1c70348,e9279b30,c063f81d,c0c71000,e381d000,...) at 
> > vm_object_reference+0x12a
> > kern_execve(ce194220,e9279c5c,0,28204548,282045d8,e381d000,e381d000,e381d015,e381d4dc,e385d000,3fb24,3,20) 
> > at kern_execve+0x31a
> > execve(ce194220,e9279cfc,c,ce194220,e9279d2c,...) at execve+0x4c
> > syscall(e9279d38) at syscall+0x345
> > Xint0x80_syscall() at Xint0x80_syscall+0x20
> > --- syscall (59, FreeBSD ELF32, execve), eip = 0x28146a47, esp = 0xbfbfe4cc, 
> > ebp = 0xbfbfe4e8 ---
> > 
> > db> show lockedvnods
> > Locked vnodes
> > 
> > 0xca2f7770: tag tmpfs, type VREG
> >     usecount 1, writecount 0, refcount 4 mountedhere 0
> >     flags ()
> >     v_object 0xd1c70348 ref 1 pages 19
> >      lock type tmpfs: EXCL (count 1) by thread 0xce194220 (pid 19396) with 1 
> > pending
> > tag VT_TMPFS, tmpfs_node 0xd177f9d4, flags 0x0, links 9
> >         mode 0555, owner 0, group 0, size 76648, status 0x0
> > 
> > It seems there is some locking problem in tmpfs.
> > 
> > What other info should I provide to help resolve the problem?
> 
> Here you can find two patches, which may or may not fix your problem.
> The first one is actually only to improve debug.
> 
> This patch adds all vnode flags to the output, because I believe you
> have VI_OWEINACT set, but not printed:
> 
> 	http://people.freebsd.org/~pjd/patches/vfs_subr.c.4.patch
> 
> The problem here is that vm_object_reference() calls vget() without any
> lock flag and vget() locks vnode exclusively when the VI_OWEINACT flag
> is set. vget() should probably be fixed too, but jeff_at_ opinion is that
> it shouldn't happen in this case, so this may be tmpfs bug.
> 
> The patch below fixes some locking problems in tmpfs:
> 
> 	http://people.freebsd.org/~pjd/patches/tmpfs.patch
> 
> The problems are:
> - tmpfs_root() should honour 'flags' argument, and not always lock the
>   vnode exclusively,
> - tmpfs_lookup() should lock vnode using cnp->cn_lkflags, and not always
>   do it exclusively,
> - in ".." case when we unlock directory vnode to avoid deadlock, we
>   should relock it using the same type of lock it was locked before and
>   not always relock it exclusively,
> 
> Note, that this patch wasn't even compiled tested.

Might be, vget shall check whether the vnode is already locked. On the
other hand, I do not see how this scenario could be realized (note that
usecount is already > 0).

tmpfs may operate on random vnodes due to lack of synchronization between
reclamation and vnode attachment to the tmpfs node. I already discussed
this with delphij_at_.