Date: Mon, 9 Mar 2009 18:02:13 -0400 From: John Baldwin <jhb@freebsd.org> To: Kostik Belousov <kostikbel@gmail.com> Cc: freebsd-current@freebsd.org Subject: Re: nullfs: panic: vrele: negative ref cnt Message-ID: <200903091802.13623.jhb@freebsd.org> In-Reply-To: <20090309205200.GY41617@deviant.kiev.zoral.com.ua> References: <3a142e750903090615t63a594e8rbadd7f7231049180@mail.gmail.com> <200903091642.25311.jhb@freebsd.org> <20090309205200.GY41617@deviant.kiev.zoral.com.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On Monday 09 March 2009 4:52:00 pm Kostik Belousov wrote:
> On Mon, Mar 09, 2009 at 04:42:25PM -0400, John Baldwin wrote:
> > On Monday 09 March 2009 9:15:22 am Paul B. Mahol wrote:
> > > Hi,
> > > here is part of textdump:
> > >
> > > panic(c061db9e,0,c061cee6,88e,4,...) at panic+0x136
> > > vrele(c4326d9c,0,c433fb5f,82,c060f3c7,...) at vrele+0x111
> > > null_nodeget(c3fd4780,c4326d9c,c3ba8bf4,0,c3ba8be4,...) at
null_nodeget+0xa0
> > > null_bypass(c3ba8be0,c3f35a78,c3ba8c28) at null_bypass+0x141
> > > VOP_VPTOCNP_APV(c4340240,c3ba8be0,c061bf01,387,c3cecc00,...) at
> > > VOP_VPTOCNP_APV+0xb3
> >
> > Try this. null_bypass() can't handle VOP_VPTOCNP because VPTOCNP doesn't
> > return a vnode that is locked and VREF'd, but just a vnode that is
vhold()'d.
> > This patch attempts to give nullfs a VOP_VPTONCP() method which remaps the
> > directory vnode properly on return by locking the directory vnode while
> > invoking null_nodeget().
> >
> > --- //depot/user/jhb/lock/fs/nullfs/null_vnops.c
> > +++ /home/jhb/work/p4/lock/fs/nullfs/null_vnops.c
> > @@ -722,6 +722,34 @@
> > return VOP_VPTOFH(lvp, ap->a_fhp);
> > }
> >
> > +static int
> > +null_vptocnp(struct vop_vptocnp_args *ap)
> > +{
> > + struct vnode *lvp, *dvp;
> > + int error;
> > +
> > + /*
> > + * We can't use null_bypass() because 'dvp' is not returned
> > + * locked. It is merely 'vhold()'ed.
> > + */
> > + lvp = NULLVPTOLOWERVP(ap->a_vp);
> > + error = VOP_VPTOCNP(lvp, &dvp, ap->a_buf, ap->a_buflen);
> > + if (error)
> > + return (error);
> > +
> > + /*
> > + * Map 'dvp' to the corresponding null node. We have to lock
> > + * it before calling null_nodeget().
> > + */
> > + vn_lock(dvp, LK_SHARED | LK_RETRY);
> > + error = null_nodeget(ap->a_vp->v_mount, dvp, ap->a_vpp);
> > + if (error) {
> > + vrele(dvp);
> > + vdrop(dvp);
> > + }
> > + return (error);
> > +}
> > +
> > /*
> > * Global vfs data structures
> > */
> > @@ -743,4 +771,5 @@
> > .vop_strategy = VOP_EOPNOTSUPP,
> > .vop_unlock = null_unlock,
> > .vop_vptofh = null_vptofh,
> > + .vop_vptocnp = null_vptocnp,
> > };
>
> I do not think that nullfs-specific method is needed.
> I already mailed the following patch to the reporters.
>
> diff --git a/sys/fs/nullfs/null_vnops.c b/sys/fs/nullfs/null_vnops.c
> index 5a6823f..38da43f 100644
> --- a/sys/fs/nullfs/null_vnops.c
> +++ b/sys/fs/nullfs/null_vnops.c
> @@ -742,5 +742,6 @@ struct vop_vector null_vnodeops = {
> .vop_setattr = null_setattr,
> .vop_strategy = VOP_EOPNOTSUPP,
> .vop_unlock = null_unlock,
> + .vop_vptocnp = vop_stdvptocnp,
> .vop_vptofh = null_vptofh,
> };
That is a lot more expensive than a bypass method fwiw. For something like
devfs as the lower fs you force it to populate the entire directory rather
than just generating the component.
--
John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200903091802.13623.jhb>