From: Chris Mason <clm@fb.com>
Date: Sun, 21 Jan 2018 21:32:50 GMT
Subject: Re: [PATCH] enable core.fsyncObjectFiles by default
Message-ID: <f797983e-1890-d84c-c3c4-87904a0a8135@fb.com>
In-Reply-To: <20180118162721.GA26078@lst.de>

On 01/18/2018 11:27 AM, Christoph Hellwig wrote:
> [adding Chris to the Cc list - this is about the awful ext3 data=ordered
> behavior of syncing the whole file system data and metadata on each
> fsync]
> 
> On Wed, Jan 17, 2018 at 03:57:53PM -0800, Linus Torvalds wrote:
>> On Wed, Jan 17, 2018 at 3:52 PM, Theodore Ts'o <tytso@mit.edu> wrote:
>>>
>>> Well, let's be fair; this is something *ext3* got wrong, and it was
>>> the default file system back them.
>>
>> I'm pretty sure reiserfs and btrfs did too..
> 
> I'm pretty sure btrfs never did, and reiserfs at least looks like
> it currently doesn't but I'd have to dig into history to check if
> it ever did.
> 

Christoph has this right, btrfs only fsyncs the one file that you've 
asked for, and unrelated data/metadata won't be included.

We've seen big fsync stalls on ext4 caused by data=ordered, so it's 
still possible to trigger on ext4, but much better than ext3.

I do share Ted's concern about the perf impact of the fsyncs on tons of 
loose files, but the defaults should be safety first.

-chris