From: "Benjamin Coddington" <bcodding@redhat.com>
To: "Trond Myklebust" <trond.myklebust@primarydata.com>
Cc: linux-nfs@vger.kernel.org
Subject: Re: Concurrent `ls` takes out the thrash
Date: Thu, 08 Dec 2016 16:48:44 -0500
Message-ID: <41B5791B-E515-4EA3-85AF-CFB65A2A2D32@redhat.com>
In-Reply-To: <3A6424D7-8207-4392-88C5-6CD3E1F65800@redhat.com>
References: <C2A81CDB-ED8F-4D41-919F-3F92CBFF751B@redhat.com>
 <CEE29C1D-BB8F-4504-940D-43DE1FE63385@primarydata.com>
 <7DA8E9BE-7353-44D5-B982-B477CF7B0A57@redhat.com>
 <934FC075-4393-42AD-92A3-3BC3BE580699@redhat.com>
 <6C5DA8AC-9A42-45FA-892C-E9597A7F7AC9@primarydata.com>
 <EC88DF6D-5A78-499C-A6A8-A5D74083876F@redhat.com>
 <3C863AB4-D07A-494A-AD7B-87BF31496777@primarydata.com>
 <C3824A8A-17EF-465F-8267-55F0AD2C4239@redhat.com>
 <790E1295-EBBE-4059-9EEC-D2C9CD36EA20@redhat.com>
 <3A6424D7-8207-4392-88C5-6CD3E1F65800@redhat.com>
MIME-Version: 1.0
Sender: linux-nfs-owner@vger.kernel.org

> Ehh.. I think it's kinda ugly, but this is what I came up with.
>
> It works well, though, and handles everything I throw at it.  Its not as
> simple as Christoph's suggestion to just go back to .iterate, which would
> solve this in a simpler way.

After many testings, I can't find any way this is faster than using the old
.iterate serialized readdir.  Additionally, at around 1M entries, this way
frequently needs to oom-kill things.  Using .iterate, that doesn't happen,
and is much faster at that scale.

Maybe my feeble brain is unable to think of the right workload to make
.iterate_shared preferable for NFS.  Suggestions welcome.

Otherwise, I think serialized is the way to go.

Ben