Re: [PATCH] nfsd: remove unneeded EEXIST error check in nfsd_do_file_acquire

[Chuck Lever <chuck.lever@oracle.com>]

On Mon, Jul 15, 2024 at 08:25:53AM -0400, Jeff Layton wrote:
> On Mon, 2024-07-15 at 10:27 +1000, NeilBrown wrote:
> > On Fri, 12 Jul 2024, Jeff Layton wrote:
> > > Given that we do the search and insertion while holding the i_lock, I
> > > don't think it's possible for us to get EEXIST here. Remove this case.
> > 
> > I was going to comment that as rhltable_insert() cannot return -EEXIST
> > that is an extra reason to discard the check.  But then I looked at the
> > code an I cannot convince myself that it cannot.
> > If __rhashtable_insert_fast() finds that tbl->future_tbl is not NULL it
> > calls rhashtable_insert_slow(), and that seems to fail if the key
> > already exists.  But it shouldn't for an rhltable, it should just add
> > the new item to the linked list for that key.
> > 
> > It looks like this has always been broken: adding to an rhltable during
> > a resize event can cause EEXIST....
> > 
> > Would anyone like to check my work?  I'm surprise that hasn't been
> > noticed if it is really the case.
> > 
> > 
> 
> I don't know this code well at all, but it looks correct to me:
> 
> static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
>                                    struct rhash_head *obj)
> {
>         struct bucket_table *new_tbl;
>         struct bucket_table *tbl;
>         struct rhash_lock_head __rcu **bkt;
>         unsigned long flags;
>         unsigned int hash;
>         void *data;
> 
>         new_tbl = rcu_dereference(ht->tbl);
> 
>         do {
>                 tbl = new_tbl;
>                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
>                 if (rcu_access_pointer(tbl->future_tbl))
>                         /* Failure is OK */
>                         bkt = rht_bucket_var(tbl, hash);
>                 else
>                         bkt = rht_bucket_insert(ht, tbl, hash);
>                 if (bkt == NULL) {
>                         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
>                         data = ERR_PTR(-EAGAIN);
>                 } else {
>                         flags = rht_lock(tbl, bkt);
>                         data = rhashtable_lookup_one(ht, bkt, tbl,
>                                                      hash, key, obj);
>                         new_tbl = rhashtable_insert_one(ht, bkt, tbl,
>                                                         hash, obj, data);
>                         if (PTR_ERR(new_tbl) != -EEXIST)
>                                 data = ERR_CAST(new_tbl);
> 
>                         rht_unlock(tbl, bkt, flags);
>                 }
>         } while (!IS_ERR_OR_NULL(new_tbl));
> 
>         if (PTR_ERR(data) == -EAGAIN)
>                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
>                                -EAGAIN);
> 
>         return data;
> }
> 
> I'm assuming the part we need to worry about is where
> rhashtable_insert_one returns -EEXIST.
> 
> It holds the rht_lock across the lookup and insert though. So if
> rhashtable_insert_one returns -EEXIST, then "data" must be something
> valid. In that case, "data" won't be overwritten and it will fall
> through and return the pointer to the entry already there.
> 
> That said, this logic is really convoluted, so I may have missed
> something too.

This is the issue I was concerned about after my review: it's
obvious that the rhtable API can return -EEXIST, but it's just
really hard to tell whether the rh/l/table API will ever return
-EEXIST.

As Neil says, the rhtable "hash table full" case should not happen
with rhltable. But can we prove that?

If we are not yet confident, then maybe PATCH 1/3 should replace
the "if (ret == -EEXIST)" with "WARN_ON(ret == -EEXIST)"...? It's
also possible to ask the human(s) who constructed the rhltable
code. :-)


> > > Cc: Youzhong Yang <youzhong@gmail.com>
> > > Fixes: c6593366c0bf ("nfsd: don't kill nfsd_files because of lease break error")
> > > Signed-off-by: Jeff Layton <jlayton@kernel.org>
> > > ---
> > > This is replacement for PATCH 1/3 in the series I sent yesterday. I
> > > think it makes sense to just eliminate this case.
> > > ---
> > >  fs/nfsd/filecache.c | 2 --
> > >  1 file changed, 2 deletions(-)
> > > 
> > > diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
> > > index f84913691b78..b9dc7c22242c 100644
> > > --- a/fs/nfsd/filecache.c
> > > +++ b/fs/nfsd/filecache.c
> > > @@ -1038,8 +1038,6 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
> > >  	if (likely(ret == 0))
> > >  		goto open_file;
> > >  
> > > -	if (ret == -EEXIST)
> > > -		goto retry;
> > >  	trace_nfsd_file_insert_err(rqstp, inode, may_flags, ret);
> > >  	status = nfserr_jukebox;
> > >  	goto construction_err;
> > > 
> > > ---
> > > base-commit: ec1772c39fa8dd85340b1a02040806377ffbff27
> > > change-id: 20240711-nfsd-next-c9d17f66e2bd
> > > 
> > > Best regards,
> > > -- 
> > > Jeff Layton <jlayton@kernel.org>
> > > 
> > > 
> > 
> 
> -- 
> Jeff Layton <jlayton@kernel.org>

-- 
Chuck Lever