Re: vfs: Add MS_FLUSHONFSYNC mount flag

From: Eric Sandeen <sandeen@redhat.com>
To: "Fernando Luis Vázquez Cao" <fernando@oss.ntt.co.jp>
Cc: Jan Kara <jack@suse.cz>, Theodore Tso <tytso@MIT.EDU>,
	Alan Cox <alan@lxorguk.ukuu.org.uk>, Pavel Machek <pavel@suse.cz>,
	kernel list <linux-kernel@vger.kernel.org>,
	Jens Axboe <jens.axboe@oracle.com>,
	fernando@kic.ac.jp, Ric Wheeler <rwheeler@redhat.com>
Subject: Re: vfs: Add MS_FLUSHONFSYNC mount flag
Date: Fri, 13 Feb 2009 00:20:17 -0600	[thread overview]
Message-ID: <49951121.80807@redhat.com> (raw)
In-Reply-To: <1234487679.3795.15.camel@sebastian.kern.oss.ntt.co.jp>

Fernando Luis Vázquez Cao wrote:
> On Thu, 2009-02-12 at 11:13 -0600, Eric Sandeen wrote:
>> Fernando Luis Vázquez Cao wrote:
>>> This mount flag will be used to determine whether the block device's write
>>> cache should be flush or not on fsync()/fdatasync().
>>>
>>> Signed-off-by: Fernando Luis Vazquez Cao <fernando@oss.ntt.co.jp>
>>> ---
>> Again, apologies for chiming in late.
>>
>> But wouldn't it be better to make this a block device property rather
>> than a new filesystem mount option?
>>
>> That way the filesystem can always do "the right thing" and call the
>> blkdev flush on fsync.
>>
>> The block device *could* choose to ignore this in hardware if it knows
>> it's built with a nonvolatile write cache or if it has no write cache.
>>
>> Somewhere in the middle, if an administrator knows they have a UPS they
>> trust and hardware that stays connected to it, they could tune the bdev
>> to ignore these flush requests.
>>
>> Also that way if you have 8 partitions on a battery-backed blockdev, you
>> can tune it once, instead of needing to mount all 8 filesystems with the
>> new option.
> 
> The main reason I decided to go for the mount option approach is to be
> consistent with what we do when it comes to write barriers. Treating one
> as a mount option and the other as a (possibly) sysfs tunable property
> seems a bit confusing to me.

well... technically, I think barriers really *should* mean "don't
reorder these writes, I need them this way for consistency" - and that
is really specific to the fs implementation, isn't it? (we just happen
to implement them as cache flushes) and so that is a per-fs setting, I
think.

Maybe there is no good argument for ignoring barriers on one fs, and
implementing them on another, other than playing fast & loose &
dangerous.... hrm.

> Do you suggest using sysfs tunables instead?

For a per-bdev flush setting, yes...

I guess I'll have to try to convince myself one way or another whether
barrier mount options are consistent with this view.  :)

I guess sometimes you do have workloads where you simply want speed, and
on a crash you start over.  In this case you don't care about barriers
(ordering constraints - if you don't care about fs integrity if fsck or
re-mkfs is ok) or flushing (caches - if you don't care about data
integrity, you regenerate your results).  That could vary from fs to fs....

I'm just a little leery of the "dangerous" mount option proliferation, I
guess.

-Eric

> - Fernando
> 