Re: [PATCH] barrier patch set

[Chris Mason <mason@suse.com>]

On Wed, 2004-03-31 at 16:26, Jeff Garzik wrote:
>  
> > Yes, it gets ugly in a hurry.  Jeff, look at the whole thread about the
> > O_DIRECT read vs buffered write races.  I don't think we can use FUA for
> 
> Yes, I'm aware of the thread...
> 
> 
> > fsync or O_SYNC without using it for every write.
> 
> Why not for O_SYNC?  Is some crazy userspace application flipping this 
> bit on and off rapidly?
> 

For both fsync and O_SYNC, the pages we want to write synchronously are
also available for some other part of the kernel to write async.  Since
we do know the write is going to be O_SYNC when we are marking the pages
dirty, we could mark them dirty_fua or something as well.

Even assuming we can deal with the data=ordered ext3/reiserfs issues, it
makes the writeback for O_SYNC yet another corner case to check, and one
where we have no useful way to make sure the fua bit really got set on
all the writes for a given O_SYNC (unless we pin the page and check each
one after the writes are complete).

Since O_DIRECT is much less complex I think we should start there.

-chris