Re: [PATCH 2/2] ubifs: Allow O_DIRECT

[Chris Mason <clm@fb.com>]

On Tue, Aug 25, 2015 at 10:00:58AM -0400, Jeff Moyer wrote:
> Dave Chinner <david@fromorbit.com> writes:
> 
> > On Mon, Aug 24, 2015 at 01:19:24PM -0400, Jeff Moyer wrote:
> >> Brian Norris <computersforpeace@gmail.com> writes:
> >> 
> >> > On Mon, Aug 24, 2015 at 10:13:25AM +0300, Artem Bityutskiy wrote:
> >> >> Now, some user-space fails when direct I/O is not supported.
> >> >
> >> > I think the whole argument rested on what it means when "some user space
> >> > fails"; apparently that "user space" is just a test suite (which
> >> > can/should be fixed).
> >> 
> >> Even if it wasn't a test suite it should still fail.  Either the fs
> >> supports O_DIRECT or it doesn't.  Right now, the only way an application
> >> can figure this out is to try an open and see if it fails.  Don't break
> >> that.
> >
> > Who cares how a filesystem implements O_DIRECT as long as it does
> > not corrupt data? ext3 fell back to buffered IO in many situations,
> > yet the only complaints about that were performance. IOWs, it's long been
> > true that if the user cares about O_DIRECT *performance* then they
> > have to be careful about their choice of filesystem.
> 
> > But if it's only 5 lines of code per filesystem to support O_DIRECT
> > *correctly* via buffered IO, then exactly why should userspace have
> > to jump through hoops to explicitly handle open(O_DIRECT) failure?
> 
> > Especially when you consider that all they can do is fall back to
> > buffered IO themselves....
> 
> I had written counterpoints for all of this, but I thought better of
> it.  Old versions of the kernel simply ignore O_DIRECT, so clearly
> there's precedent.
> 
> I do think we should at least document what file systems appear to be
> doing.  Here's a man page patch for open (generated with extra context
> for easier reading).  Let me know what you think.

We shouldn't be ignoring it, but instead call it similar to O_DSYNC plus
removing the pages from cache.

-chris