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;
Message-ID: <2af895fcbdaeab04773cbce275ca7a6d59b879cf.camel@redhat.com>
Subject: Re: [PATCH v3 15/32] NFSD: Leave open files out of the filecache LRU
From:   Jeff Layton <jlayton@redhat.com>
To:     Chuck Lever <chuck.lever@oracle.com>, linux-nfs@vger.kernel.org,
        netdev@vger.kernel.org
Cc:     david@fromorbit.com, tgraf@suug.ch
Date:   Fri, 08 Jul 2022 15:29:55 -0400
In-Reply-To: <165730471781.28142.13547044100953437563.stgit@klimt.1015granger.net>
References: <165730437087.28142.6731645688073512500.stgit@klimt.1015granger.net>
         <165730471781.28142.13547044100953437563.stgit@klimt.1015granger.net>
Content-Type: text/plain; charset="ISO-8859-15"
Content-Transfer-Encoding: quoted-printable
User-Agent: Evolution 3.44.3 (3.44.3-1.fc36) 
MIME-Version: 1.0
Precedence: bulk

On Fri, 2022-07-08 at 14:25 -0400, Chuck Lever wrote:
> There have been reports of problems when running fstests generic/531
> against Linux NFS servers with NFSv4. The NFS server that hosts the
> test's SCRATCH_DEV suffers from CPU soft lock-ups during the test.
> Analysis shows that:
>=20
> fs/nfsd/filecache.c
>  482                 ret =3D list_lru_walk(&nfsd_file_lru,
>  483                                 nfsd_file_lru_cb,
>  484                                 &head, LONG_MAX);
>=20
> causes nfsd_file_gc() to walk the entire length of the filecache LRU
> list every time it is called (which is quite frequently). The walk
> holds a spinlock the entire time that prevents other nfsd threads
> from accessing the filecache.
>=20
> What's more, for NFSv4 workloads, none of the items that are visited
> during this walk may be evicted, since they are all files that are
> held OPEN by NFS clients.
>=20
> Address this by ensuring that open files are not kept on the LRU
> list.
>=20
> Reported-by: Frank van der Linden <fllinden@amazon.com>
> Reported-by: Wang Yugui <wangyugui@e16-tech.com>
> Link: https://bugzilla.linux-nfs.org/show_bug.cgi?id=3D386
> Suggested-by: Trond Myklebust <trond.myklebust@hammerspace.com>
> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> ---
>  fs/nfsd/filecache.c |   24 +++++++++++++++++++-----
>  fs/nfsd/trace.h     |    2 ++
>  2 files changed, 21 insertions(+), 5 deletions(-)
>=20
> diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
> index 37373b012276..6e9e186334ab 100644
> --- a/fs/nfsd/filecache.c
> +++ b/fs/nfsd/filecache.c
> @@ -269,6 +269,7 @@ nfsd_file_flush(struct nfsd_file *nf)
> =20
>  static void nfsd_file_lru_add(struct nfsd_file *nf)
>  {
> +	set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
>  	if (list_lru_add(&nfsd_file_lru, &nf->nf_lru))
>  		trace_nfsd_file_lru_add(nf);
>  }
> @@ -298,7 +299,6 @@ nfsd_file_unhash(struct nfsd_file *nf)
>  {
>  	if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
>  		nfsd_file_do_unhash(nf);
> -		nfsd_file_lru_remove(nf);
>  		return true;
>  	}
>  	return false;
> @@ -319,6 +319,7 @@ nfsd_file_unhash_and_release_locked(struct nfsd_file =
*nf, struct list_head *disp
>  	if (refcount_dec_not_one(&nf->nf_ref))
>  		return true;
> =20
> +	nfsd_file_lru_remove(nf);
>  	list_add(&nf->nf_lru, dispose);
>  	return true;
>  }
> @@ -330,6 +331,7 @@ nfsd_file_put_noref(struct nfsd_file *nf)
> =20
>  	if (refcount_dec_and_test(&nf->nf_ref)) {
>  		WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
> +		nfsd_file_lru_remove(nf);
>  		nfsd_file_free(nf);
>  	}
>  }
> @@ -339,7 +341,7 @@ nfsd_file_put(struct nfsd_file *nf)
>  {
>  	might_sleep();
> =20
> -	set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
> +	nfsd_file_lru_add(nf);

Do you really want to add this on every put? I would have thought you'd
only want to do this on a 2->1 nf_ref transition.

>  	if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) =3D=3D 0) {
>  		nfsd_file_flush(nf);
>  		nfsd_file_put_noref(nf);
> @@ -439,8 +441,18 @@ nfsd_file_dispose_list_delayed(struct list_head *dis=
pose)
>  	}
>  }
> =20
> -/*
> +/**
> + * nfsd_file_lru_cb - Examine an entry on the LRU list
> + * @item: LRU entry to examine
> + * @lru: controlling LRU
> + * @lock: LRU list lock (unused)
> + * @arg: dispose list
> + *
>   * Note this can deadlock with nfsd_file_cache_purge.
> + *
> + * Return values:
> + *   %LRU_REMOVED: @item was removed from the LRU
> + *   %LRU_SKIP: @item cannot be evicted
>   */
>  static enum lru_status
>  nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
> @@ -462,8 +474,9 @@ nfsd_file_lru_cb(struct list_head *item, struct list_=
lru_one *lru,
>  	 * That order is deliberate to ensure that we can do this locklessly.
>  	 */
>  	if (refcount_read(&nf->nf_ref) > 1) {
> +		list_lru_isolate(lru, &nf->nf_lru);
>  		trace_nfsd_file_gc_in_use(nf);
> -		return LRU_SKIP;
> +		return LRU_REMOVED;

Interesting. So you wait until the LRU scanner runs to remove these
entries? I expected to see you do this in nfsd_file_get, but this does
seem likely to be more efficient.

>  	}
> =20
>  	/*
> @@ -1020,6 +1033,7 @@ nfsd_do_file_acquire(struct svc_rqst *rqstp, struct=
 svc_fh *fhp,
>  		goto retry;
>  	}
> =20
> +	nfsd_file_lru_remove(nf);
>  	this_cpu_inc(nfsd_file_cache_hits);
> =20
>  	if (!(may_flags & NFSD_MAY_NOT_BREAK_LEASE)) {
> @@ -1055,7 +1069,6 @@ nfsd_do_file_acquire(struct svc_rqst *rqstp, struct=
 svc_fh *fhp,
>  	refcount_inc(&nf->nf_ref);
>  	__set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
>  	__set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
> -	nfsd_file_lru_add(nf);
>  	hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
>  	++nfsd_file_hashtbl[hashval].nfb_count;
>  	nfsd_file_hashtbl[hashval].nfb_maxcount =3D max(nfsd_file_hashtbl[hashv=
al].nfb_maxcount,
> @@ -1080,6 +1093,7 @@ nfsd_do_file_acquire(struct svc_rqst *rqstp, struct=
 svc_fh *fhp,
>  	 */
>  	if (status !=3D nfs_ok || inode->i_nlink =3D=3D 0) {
>  		bool do_free;
> +		nfsd_file_lru_remove(nf);
>  		spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
>  		do_free =3D nfsd_file_unhash(nf);
>  		spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
> diff --git a/fs/nfsd/trace.h b/fs/nfsd/trace.h
> index 1cc1133371eb..54082b868b72 100644
> --- a/fs/nfsd/trace.h
> +++ b/fs/nfsd/trace.h
> @@ -929,7 +929,9 @@ DEFINE_EVENT(nfsd_file_gc_class, name,					\
>  	TP_ARGS(nf))
> =20
>  DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_lru_add);
> +DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_lru_add_disposed);
>  DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_lru_del);
> +DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_lru_del_disposed);
>  DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_gc_in_use);
>  DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_gc_writeback);
>  DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_gc_referenced);
>=20
>=20

--=20
Jeff Layton <jlayton@redhat.com>