On Sat, 2022-10-01 at 15:33 +0000, Chuck Lever III wrote:
> Hi Jeff-
>
> > On Oct 1, 2022, at 5:59 AM, Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> >
> > nfsd_file is RCU-freed, so it's possible that one could be found that's
> > in the process of being freed and the memory recycled. Ensure we hold
> > the rcu_read_lock while attempting to get a reference on the object.
>
> I'm OK with entertaining clean-up patches in this code, but I
> am skeptical that this patch addresses the concern enumerated
> in bug #394. As you've pointed out to me before, the "UAF on
> DELEGRETURN crashes" appeared well before v5.19, which is the
> kernel release where this bit of code was introduced.
>
Yeah, there may be more than one bug here. In any case, these patches
should close potential races, so I think we ought to take them.
>
> > Cc: NeilBrown <neilb@xxxxxxx>
> > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> > ---
> > fs/nfsd/filecache.c | 14 ++++++++------
> > 1 file changed, 8 insertions(+), 6 deletions(-)
> >
> > diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
> > index d5c57360b418..f4f75ae2e4ea 100644
> > --- a/fs/nfsd/filecache.c
> > +++ b/fs/nfsd/filecache.c
> > @@ -1077,10 +1077,12 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
> >
> > retry:
> > /* Avoid allocation if the item is already in cache */
> > + rcu_read_lock();
> > nf = rhashtable_lookup_fast(&nfsd_file_rhash_tbl, &key,
> > nfsd_file_rhash_params);
> > if (nf)
> > nf = nfsd_file_get(nf);
> > + rcu_read_unlock();
>
> Again:
>
> 657 static inline void *rhashtable_lookup_fast(
> 658 struct rhashtable *ht, const void *key,
> 659 const struct rhashtable_params params)
> 660 {
> 661 void *obj;
> 662
> 663 rcu_read_lock();
> 664 obj = rhashtable_lookup(ht, key, params);
> 665 rcu_read_unlock();
> 666
> 667 return obj;
> 668 }
>
> Since rhashtable_lookup() itself is a public API, please
> just call that in nfsd_file_do_acquire() after explicitly
> taking the RCU read lock.
>
>
Understood. Sorry I missed your point. I'll fix that up.
> > if (nf)
> > goto wait_for_construction;
> >
> > @@ -1090,21 +1092,21 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
> > goto out_status;
> > }
> >
> > + rcu_read_lock();
> > nf = rhashtable_lookup_get_insert_key(&nfsd_file_rhash_tbl,
> > &key, &new->nf_rhash,
> > nfsd_file_rhash_params);
> > + if (!IS_ERR_OR_NULL(nf)) {
> > + nf = nfsd_file_get(nf);
>
> Note that nfsd_file_get() can still return NULL.
>
True, and that would leak. Good catch.
>
> > + nfsd_file_slab_free(&new->nf_rcu);
>
> Why is the slab_free call now inside the RCU critical section?
> Granted this should be a rare case, but this adds unnecessary
> latency while the read lock is held.
>
Fair point.
>
> > + }
> > + rcu_read_unlock();
>
> Is there a problem replacing rhashtable_lookup_get_insert_key()
> with rhashtable_lookup_insert_key() and just retrying the normal
> lookup if insertion returns EEXIST? That way, an nfsd_file_get()
> is necessary only at the rhashtable_lookup() call site above.
>
>
I like this idea, and it allows for a rather nice cleanup of the code.
I'll send a v3 set after I've had a chance to do some testing.
> > if (!nf) {
> > nf = new;
>
> @new was just released above, so won't this set @nf to point
> to freed memory in some cases?
>
>
> > goto open_file;
> > }
> > if (IS_ERR(nf))
> > goto insert_err;
> > - nf = nfsd_file_get(nf);
> > - if (nf == NULL) {
> > - nf = new;
> > - goto open_file;
> > - }
> > - nfsd_file_slab_free(&new->nf_rcu);
> >
> > wait_for_construction:
> > wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
> > --
> > 2.37.3
>
>
> --
> Chuck Lever
>
>
>
--
Jeff Layton <jlayton@xxxxxxxxxx>
