On Mon, 2022-12-12 at 18:16 +0000, Chuck Lever III wrote:
>
> > On Dec 12, 2022, at 12:44 PM, Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> >
> > On Mon, 2022-12-12 at 18:14 +0100, Greg KH wrote:
> > > On Mon, Dec 12, 2022 at 09:31:19AM -0500, Jeff Layton wrote:
> > > > On Mon, 2022-12-12 at 14:59 +0100, Greg KH wrote:
> > > > > On Mon, Dec 12, 2022 at 08:40:31AM -0500, Jeff Layton wrote:
> > > > > > On Mon, 2022-12-12 at 05:34 -0800, dai.ngo@xxxxxxxxxx wrote:
> > > > > > > On 12/12/22 4:22 AM, Jeff Layton wrote:
> > > > > > > > On Sun, 2022-12-11 at 11:22 -0800, Dai Ngo wrote:
> > > > > > > > > Problem caused by source's vfsmount being unmounted but remains
> > > > > > > > > on the delayed unmount list. This happens when nfs42_ssc_open()
> > > > > > > > > return errors.
> > > > > > > > > Fixed by removing nfsd4_interssc_connect(), leave the vfsmount
> > > > > > > > > for the laundromat to unmount when idle time expires.
> > > > > > > > >
> > > > > > > > > Reported-by: Xingyuan Mo <hdthky0@xxxxxxxxx>
> > > > > > > > > Signed-off-by: Dai Ngo <dai.ngo@xxxxxxxxxx>
> > > > > > > > > ---
> > > > > > > > > fs/nfsd/nfs4proc.c | 23 +++++++----------------
> > > > > > > > > 1 file changed, 7 insertions(+), 16 deletions(-)
> > > > > > > > >
> > > > > > > > > diff --git a/fs/nfsd/nfs4proc.c b/fs/nfsd/nfs4proc.c
> > > > > > > > > index 8beb2bc4c328..756e42cf0d01 100644
> > > > > > > > > --- a/fs/nfsd/nfs4proc.c
> > > > > > > > > +++ b/fs/nfsd/nfs4proc.c
> > > > > > > > > @@ -1463,13 +1463,6 @@ nfsd4_interssc_connect(struct nl4_server *nss, struct svc_rqst *rqstp,
> > > > > > > > > return status;
> > > > > > > > > }
> > > > > > > > >
> > > > > > > > > -static void
> > > > > > > > > -nfsd4_interssc_disconnect(struct vfsmount *ss_mnt)
> > > > > > > > > -{
> > > > > > > > > - nfs_do_sb_deactive(ss_mnt->mnt_sb);
> > > > > > > > > - mntput(ss_mnt);
> > > > > > > > > -}
> > > > > > > > > -
> > > > > > > > > /*
> > > > > > > > > * Verify COPY destination stateid.
> > > > > > > > > *
> > > > > > > > > @@ -1572,11 +1565,6 @@ nfsd4_cleanup_inter_ssc(struct vfsmount *ss_mnt, struct file *filp,
> > > > > > > > > {
> > > > > > > > > }
> > > > > > > > >
> > > > > > > > > -static void
> > > > > > > > > -nfsd4_interssc_disconnect(struct vfsmount *ss_mnt)
> > > > > > > > > -{
> > > > > > > > > -}
> > > > > > > > > -
> > > > > > > > > static struct file *nfs42_ssc_open(struct vfsmount *ss_mnt,
> > > > > > > > > struct nfs_fh *src_fh,
> > > > > > > > > nfs4_stateid *stateid)
> > > > > > > > > @@ -1762,7 +1750,8 @@ static int nfsd4_do_async_copy(void *data)
> > > > > > > > > struct file *filp;
> > > > > > > > >
> > > > > > > > > filp = nfs42_ssc_open(copy->ss_mnt, ©->c_fh,
> > > > > > > > > - ©->stateid);
> > > > > > > > > + ©->stateid);
> > > > > > > > > +
> > > > > > > > > if (IS_ERR(filp)) {
> > > > > > > > > switch (PTR_ERR(filp)) {
> > > > > > > > > case -EBADF:
> > > > > > > > > @@ -1771,7 +1760,7 @@ static int nfsd4_do_async_copy(void *data)
> > > > > > > > > default:
> > > > > > > > > nfserr = nfserr_offload_denied;
> > > > > > > > > }
> > > > > > > > > - nfsd4_interssc_disconnect(copy->ss_mnt);
> > > > > > > > > + /* ss_mnt will be unmounted by the laundromat */
> > > > > > > > > goto do_callback;
> > > > > > > > > }
> > > > > > > > > nfserr = nfsd4_do_copy(copy, filp, copy->nf_dst->nf_file,
> > > > > > > > > @@ -1852,8 +1841,10 @@ nfsd4_copy(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
> > > > > > > > > if (async_copy)
> > > > > > > > > cleanup_async_copy(async_copy);
> > > > > > > > > status = nfserrno(-ENOMEM);
> > > > > > > > > - if (nfsd4_ssc_is_inter(copy))
> > > > > > > > > - nfsd4_interssc_disconnect(copy->ss_mnt);
> > > > > > > > > + /*
> > > > > > > > > + * source's vfsmount of inter-copy will be unmounted
> > > > > > > > > + * by the laundromat
> > > > > > > > > + */
> > > > > > > > > goto out;
> > > > > > > > > }
> > > > > > > > >
> > > > > > > > This looks reasonable at first glance, but I have some concerns with the
> > > > > > > > refcounting around ss_mnt elsewhere in this code. nfsd4_ssc_setup_dul
> > > > > > > > looks for an existing connection and bumps the ni->nsui_refcnt if it
> > > > > > > > finds one.
> > > > > > > >
> > > > > > > > But then later, nfsd4_cleanup_inter_ssc has a couple of cases where it
> > > > > > > > just does a bare mntput:
> > > > > > > >
> > > > > > > > if (!nn) {
> > > > > > > > mntput(ss_mnt);
> > > > > > > > return;
> > > > > > > > }
> > > > > > > > ...
> > > > > > > > if (!found) {
> > > > > > > > mntput(ss_mnt);
> > > > > > > > return;
> > > > > > > > }
> > > > > > > >
> > > > > > > > The first one looks bogus. Can net_generic return NULL? If so how, and
> > > > > > > > why is it not a problem elsewhere in the kernel?
> > > > > > >
> > > > > > > it looks like net_generic can not fail, no where else check for NULL
> > > > > > > so I will remove this check.
> > > > > > >
> > > > > > > >
> > > > > > > > For the second case, if the ni is no longer on the list, where did the
> > > > > > > > extra ss_mnt reference come from? Maybe that should be a WARN_ON or
> > > > > > > > BUG_ON?
> > > > > > >
> > > > > > > if ni is not found on the list then it's a bug somewhere so I will add
> > > > > > > a BUG_ON on this.
> > > > > > >
> > > > > >
> > > > > > Probably better to just WARN_ON and let any references leak in that
> > > > > > case. A BUG_ON implies a panic in some environments, and it's best to
> > > > > > avoid that unless there really is no choice.
> > > > >
> > > > > WARN_ON also causes machines to boot that have panic_on_warn enabled.
> > > > > Why not just handle the error and keep going? Why panic at all?
> > > > >
> > > >
> > > > Who the hell sets panic_on_warn (outside of testing environments)?
> > >
> > > All cloud providers and anyone else that wants to "kill the system that
> > > had a problem and have it reboot fast" in order to keep things working
> > > overall.
> > >
> >
> > If that's the case, then this situation would probably be one where a
> > cloud provider would want to crash it and come back. NFS grace periods
> > can suck though.
> >
> > > > I'm
> > > > suggesting a WARN_ON because not finding an entry at this point
> > > > represents a bug that we'd want reported.
> > >
> > > Your call, but we are generally discouraging adding new WARN_ON() for
> > > anything that userspace could ever trigger. And if userspace can't
> > > trigger it, then it's a normal type of error that you need to handle
> > > anyway, right?
> > >
> > > Anyway, your call, just letting you know.
> > >
> >
> > Understood.
> >
> > > > The caller should hold a reference to the object that holds a vfsmount
> > > > reference. It relies on that vfsmount to do a copy. If it's gone at this
> > > > point where we're releasing that reference, then we're looking at a
> > > > refcounting bug of some sort.
> > >
> > > refcounting in the nfsd code, or outside of that?
> > >
> >
> > It'd be in the nfsd code, but might affect the vfsmount refcount. Inter-
> > server copy is quite the tenuous house of cards. ;)
> >
> > > > I would expect anyone who sets panic_on_warn to _desire_ a panic in this
> > > > situation. After all, they asked for it. Presumably they want it to do
> > > > some coredump analysis or something?
> > > >
> > > > It is debatable whether the stack trace at this point would be helpful
> > > > though, so you might consider a pr_warn or something less log-spammy.
> > >
> > > If you can recover from it, then yeah, pr_warn() is usually best.
> > >
> >
> > It does look like Dai went with pr_warn on his v2 patch.
> >
> > We'd "recover" by leaking a vfsmount reference. The immediate crash
> > would be avoided, but it might make for interesting "fun" later when you
> > went to try and unmount the thing.
>
> This is a red flag for me. If the leak prevents the system from
> shutting down reliably, then we need to do something more than
> a pr_warn(), I would think.
>
Sorry, I should correct myself.
We wouldn't (necessarily) leak a vfsmount reference. If the entry was no
longer on the list, then presumably it has already been cleaned up and
the vfsmount reference put.
It's still a bug though since we _should_ still have a reference to the
nfsd4_ssc_umount_item at this point. So this is really just a potential
use-after-free.
FWIW, the object handling here is somewhat weird as the copy operation
holds a reference to the nfsd4_ssc_umount_item but passes around a
pointer to the vfsmount
I have to wonder if it'd be cleaner to have nfsd4_setup_inter_ssc pass
back a pointer to the nfsd4_ssc_umount_item, so you could pass that to
nfsd4_cleanup_inter_ssc and skip searching for it again at cleanup time.
--
Jeff Layton <jlayton@xxxxxxxxxx>
