Re: [GIT PULL] Btrfs updates for 6.15

[Josef Bacik <josef@toxicpanda.com>]

On Fri, Mar 28, 2025 at 06:36:44PM +0100, David Sterba wrote:
> On Fri, Mar 28, 2025 at 09:27:51AM -0400, Josef Bacik wrote:
> > On Mon, Mar 24, 2025 at 05:37:51PM +0100, David Sterba wrote:
> > > Hi,
> > > 
> > > please pull the following btrfs updates, thanks.
> > > 
> > > User visible changes:
> > > 
> > > - fall back to buffered write if direct io is done on a file that requires
> > >   checksums
> > 
> > <trimming the everybody linux-btrfs from the cc list>
> > 
> > What?  We use this constantly in a bunch of places to avoid page cache overhead,
> > it's perfectly legitimate to do DIO to a file that requires checksums.  Does the
> > vm case mess this up?  Absolutely, but that's why we say use NOCOW for that
> > case.  We've always had this behavior, we've always been clear that if you break
> > it you buy it.  This is a huge regression for a pretty significant use case.
> 
> The patch has been up for like 2 months and you could have said "don't
> because reasons" any time before the pull request. Now we're left with a
> revert, or other alternatives making the use cases working.

Boris told me about this and I forgot.  I took everybody off the CC list because
I don't want to revert it, in fact generally speaking I'd love to never have
these style of bug reports again.

But it is a pretty significant change.  Are we ok going forward saying you don't
get O_DIRECT unless you want NOCOW? Should we maybe allow for users to indicate
they're not dumb and can be trusted to do O_DIRECT properly? I just think this
opens us up to a lot more uncomfortable conversations than the other behavior.

I personally think this is better, if it's been sitting there for 2 months then
hooray we're in agreement. But I'm also worried it'll come back to bite us.
Thanks,

Josef