Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
MIME-Version: 1.0
From:   "NeilBrown" <neilb@suse.de>
To:     "Jeff Layton" <jlayton@kernel.org>
Cc:     chuck.lever@oracle.com, linux-nfs@vger.kernel.org
Subject: Re: [PATCH 2/3] nfsd: fix potential race in nfsd_file_close
In-reply-to: <20220930191550.172087-3-jlayton@kernel.org>
References: <20220930191550.172087-1-jlayton@kernel.org>,
 <20220930191550.172087-3-jlayton@kernel.org>
Date:   Sat, 01 Oct 2022 15:03:38 +1000
Message-id: <166460061835.17572.12490851025838613566@noble.neil.brown.name>
Precedence: bulk

On Sat, 01 Oct 2022, Jeff Layton wrote:
> Once we call nfsd_file_put, there is no guarantee that "nf" can still be
> safely accessed. That may have been the last reference.
> 
> Change the code to instead check for whether nf_ref is 2 and then unhash
> it and put the reference if we're successful.
> 
> We might occasionally race with another lookup and end up unhashing it
> when it probably shouldn't have been, but that should hopefully be rare
> and will just result in the competing lookup having to create a new
> nfsd_file.
> 
> Signed-off-by: Jeff Layton <jlayton@kernel.org>
> ---
>  fs/nfsd/filecache.c | 12 +++++++-----
>  1 file changed, 7 insertions(+), 5 deletions(-)
> 
> diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
> index 6237715bd23e..58f4d9267f4a 100644
> --- a/fs/nfsd/filecache.c
> +++ b/fs/nfsd/filecache.c
> @@ -461,12 +461,14 @@ nfsd_file_put(struct nfsd_file *nf)
>   */
>  void nfsd_file_close(struct nfsd_file *nf)
>  {
> -	nfsd_file_put(nf);
> -	if (refcount_dec_if_one(&nf->nf_ref)) {
> -		nfsd_file_unhash(nf);
> -		nfsd_file_lru_remove(nf);
> -		nfsd_file_free(nf);
> +	/* One for the reference being put, and one for the hash */
> +	if (refcount_read(&nf->nf_ref) == 2) {
> +		if (nfsd_file_unhash(nf))
> +			nfsd_file_put_noref(nf);
>  	}
> +	/* put the ref for the stateid */
> +	nfsd_file_put(nf);
> +

This looks racy. What if a get happens after the read and before the unhash?

If we unhash the nfsd_file at last close, why does the hash table hold a
counted reference at all?
When it is hashed, set the NFSD_FILE_HASHED flag.  On last-put, if that
flag is set, unhash it.
If you want to unhash it earlier, test/clear the flag and delete from
rhashtable.

NeilBrown


>  }
>  
>  struct nfsd_file *
> -- 
> 2.37.3
> 
>