From: Eric Sandeen <sandeen@redhat.com>
Subject: Re: [PATCH] ext4: add max_dir_size_kb mount option
Date: Fri, 10 Aug 2012 15:16:57 -0500
Message-ID: <50256C39.4030502@redhat.com>
References: <1344626638-31548-1-git-send-email-tytso@mit.edu> <50256322.9040409@redhat.com> <20120810195907.GA557@thunk.org>
Mime-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit
Cc: Ext4 Developers List <linux-ext4@vger.kernel.org>
To: "Theodore Ts'o" <tytso@mit.edu>
In-Reply-To: <20120810195907.GA557@thunk.org>
Sender: linux-ext4-owner@vger.kernel.org

On 8/10/12 2:59 PM, Theodore Ts'o wrote:
> On Fri, Aug 10, 2012 at 02:38:10PM -0500, Eric Sandeen wrote:
>>
>> Can the commit message also describe more about the problem: how bad
>> it is, the root cause, and why it's so hard to fix properly?
> 
> The use case is going to be fairly userspace specific, but one example
> might be if the log reaper fails to run for whatever reason, and the
> log directory then proceeds to grow without bound.  And then when the
> log repear *does* have a chance to run, if it happens to be in tight
> memory cgroup, it then dies so the directory grows even larger, and
> then when other processes try to access the directory, a readdir will
> cause them to die because of their memory limitation, and hilarity
> ensues.

Oh, I thought this was papering over a scaling problem in ext4.  The intent
is to protect userspace from arbitrarily large readdir results?

If that's the case, this should probably be proposed as a VFS level
option, and see how it's received there...

(Can you tell I'm not a huge fan of the idea?) ;)

-Eric

> You can fix this in other places in the software stack, but belt and
> suspenders is good, and if there is no reason for directories to grow
> larger than some set size, it's better to get a hard failure with an
> ENOSPC rather than an funny failures caused by OOM's or slower and
> slower performance.

>> Please also update Documentation/filesystems/ext4.txt so people
>> know for sure what the use & intent of this new knob is.
> 
> Yes, I'll do that in the next spin of the patches.
> 
>      	     	     	      	      - Ted
>