Received-SPF: pass (google.com: domain of linux-kernel+bounces-135640-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Organization: Red Hat UK Ltd. Registered Address: Red Hat UK Ltd, Amberley
	Place, 107-111 Peascod Street, Windsor, Berkshire, SI4 1TE, United
	Kingdom.
	Registered in England and Wales under Company Registration No. 3798903
From: David Howells <dhowells@redhat.com>
In-Reply-To: <20240328163424.2781320-24-dhowells@redhat.com>
References: <20240328163424.2781320-24-dhowells@redhat.com> <20240328163424.2781320-1-dhowells@redhat.com>
To: Christian Brauner <christian@brauner.io>,
    Matthew Wilcox <willy@infradead.org>
Cc: dhowells@redhat.com, Jeff Layton <jlayton@kernel.org>,
    Gao Xiang <hsiangkao@linux.alibaba.com>,
    Dominique Martinet <asmadeus@codewreck.org>,
    Steve French <smfrench@gmail.com>,
    Marc Dionne <marc.dionne@auristor.com>,
    Paulo Alcantara <pc@manguebit.com>,
    Shyam Prasad N <sprasad@microsoft.com>, Tom Talpey <tom@talpey.com>,
    Eric Van Hensbergen <ericvh@kernel.org>,
    Ilya Dryomov <idryomov@gmail.com>, netfs@lists.linux.dev,
    linux-cachefs@redhat.com, linux-afs@lists.infradead.org,
    linux-cifs@vger.kernel.org, linux-nfs@vger.kernel.org,
    ceph-devel@vger.kernel.org, v9fs@lists.linux.dev,
    linux-erofs@lists.ozlabs.org, linux-fsdevel@vger.kernel.org,
    linux-mm@kvack.org, netdev@vger.kernel.org,
    linux-kernel@vger.kernel.org, Latchesar Ionkov <lucho@ionkov.net>,
    Christian Schoenebeck <linux_oss@crudebyte.com>
Subject: Re: [PATCH 23/26] netfs: Cut over to using new writeback code
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <877901.1712591597.1@warthog.procyon.org.uk>
Content-Transfer-Encoding: quoted-printable
Date: Mon, 08 Apr 2024 16:53:17 +0100
Message-ID: <877902.1712591597@warthog.procyon.org.uk>

David Howells <dhowells@redhat.com> wrote:

> +		/* Wait for writeback to complete.  The writeback engine owns
> +		 * the info in folio->private and may change it until it
> +		 * removes the WB mark.
> +		 */
> +		if (folio_wait_writeback_killable(folio)) {
> +			ret =3D written ? -EINTR : -ERESTARTSYS;
> +			goto error_folio_unlock;
> +		}
> +

It turns out that this really kills performance with fio with as many jobs=
 as
cpus.  It's taking up to around 8x longer to complete a pwrite() on averag=
e
and perf shows a 30% of the CPU cycles are being spent in contention on th=
e
i_rwsem.

The reason this was added here is that writeback cannot take the folio loc=
k in
order to clean up folio->private without risking deadlock vs the truncatio=
n
routines (IIRC).

I can mitigate this by skipping the wait if folio->private is not set and =
if
we're not going to attach anything there (see attached).  Note that if
writeout is ongoing and there is nothing attached to ->private, then we sh=
ould
not be engaging write-streaming mode and attaching a new netfs_folio (and =
if
we did, we'd flush the page and wait for it anyway).

The other possibility is if we have a writeback group to set.  This only
applies to ceph for the moment and is something that will need dealing wit=
h
if/when ceph is made to use this code.

David
---

diff --git a/fs/netfs/buffered_write.c b/fs/netfs/buffered_write.c
index 1eff9413eb1b..279b296f8014 100644
--- a/fs/netfs/buffered_write.c
+++ b/fs/netfs/buffered_write.c
@@ -255,7 +255,8 @@ ssize_t netfs_perform_write(struct kiocb *iocb, struct=
 iov_iter *iter,
 		 * the info in folio->private and may change it until it
 		 * removes the WB mark.
 		 */
-		if (folio_wait_writeback_killable(folio)) {
+		if (folio_get_private(folio) &&
+		    folio_wait_writeback_killable(folio)) {
 			ret =3D written ? -EINTR : -ERESTARTSYS;
 			goto error_folio_unlock;
 		}