Re: BUG: Failure to send REQ_FLUSH on unmount on ext3, ext4, and FS in general

[Alex Bligh <alex@alex.org.uk>]

Jan,

--On 23 May 2011 19:29:06 +0200 Jan Kara <jack@suse.cz> wrote:

>   I wish it was this simple ;) The trouble is that clever filesystems -
> e.g. xfs, ext4 - will send the flush when it's needed (after a transaction
> commit). So sending it after flushing the device (which happens from
> generic sync code) would result in two flushes instead of one - not good
> for performance (although these days when we do merging of flush requests
> the result need not be that bad).
>
> The fs might indicate whether it handles barriers itself or whether it
> wants VFS to handle it but that's where it's gets a bit complicated /
> controversial ;).

Well, to "fix" sync(), one could simply look at whether the file system
had ever sent a REQ_FLUSH or REQ_FUA since that FS was mounted. If there
has been one, assume the FS is taking responsibility for sending them.

I'm presuming that if just umount() were altered to do a REQ_FLUSH,
the potential presence of 2 sync()s would not be too offensive, as
unmount isn't exactly time critical, and as Christoph pointed out in
the other thread, a REQ_FLUSH when the write cache has recently been
emptied isn't going to take long.

>> Would there be any interested in these patches if I cooked them up,
>> or did they die because of opposition before rather than apathy?
>
>   I guess you might come with some proposal and post it to linux-fsdevel
> (include Al Viro and Christoph Hellwig in CC) and see what happens...

Ah, fsdevel not here. OK. Partly I'd like to understand whether
sync() not flushing write caches on barrier-less file systems
is a good thing or a bad thing. I know barriers are better, but if
writing to (e.g.) FAT32, I'm betting there is little prospect of
barrier support.

-- 
Alex Bligh