svn commit: r340187 - head/sys/geom

[Rodney W. Grimes]

> Reverted, sorry. Turns out that i/o into last_sector+1 is handled
> differently. I'll probably have to use different strategy to properly fail
> out-of-bound ioctl(DIOCGDELETE) or otherwise indicate its result to the
> userland app. To my defense, this patch has been out for 3 weeks on
> freebsd-geom, and I got 0 responses.

Well you have some now, so lets try to sort this out and at least
document what the funny "access one byte past end with zero size
I/O" is about.

A good stratergy when you get no response on a phabricator review
is to poke about it on @current or @arch requesting feedback.
The phabricator notificaiton to potential reviewers is very low.

freebsd-geom is also probably a pretty short list.

> -Max
> 
> On Wed, Nov 7, 2018 at 8:06 AM Rodney W. Grimes <
> freebsd at pdx.rh.cn85.dnsmgr.net> wrote:
> 
> > > On Tue, 2018-11-06 at 16:17 -0800, Rodney W. Grimes wrote:
> > > > >
> > > > > Rodney, this was actually my original intention, however then I
> > noticed in
> > > > > the GEOM code there is at least one case when BIO_FLUSH request is
> > being
> > > > > generated internally with bio_offset == mediasize and bio_lenth ==
> > 0, so I
> > > > > thought there might be some need to allow such requests through. But
> > I'd
> > > > > happily go with the stricter rule if it does no harm. I simply don't
> > know
> > > > > enough about the intended use and the logic behind zero-length
> > transfers to
> > > > > make that call.
> > > > I am not sure enough on if mediasize is 0 based or not,
> > > > if it is then the error case should be fixed, and the
> > > > code you show below should also be fixed as it is
> > > > technically making a request beyond the end of device.
> > > >
> > > > I am also murky on why we are even doing a 0 size
> > > > operation and end of device, is that to validate
> > > > we can access all the media???If so then this wrong
> > > > code and wrong error return should be fixed as it
> > > > is off by 1.
> > > >
> > > > >
> > > > >
> > > > > -Max
> > > > >
> > > > > int
> > > > > g_io_flush(struct g_consumer *cp)
> > > > > {
> > > > > ...
> > > > > ????????bp = g_alloc_bio();
> > > > > ????????bp->bio_cmd = BIO_FLUSH;
> > > > > ...
> > > > > ????????bp->bio_offset = cp->provider->mediasize;
> > > > The above should have a - 1 on it.
> > > >
> > >
> > > Unless offset > mediasize is specifically a signal to downstream code
> > > in some way about how the flush is to be performed.
> >
> > Could very well be, should be documented some place though.
> >
> > > Nearly identical code to create a BIO_FLUSH bio appears in ufs softdeps
> > > and in zfs. Before starting to arbitrarily change code that has worked
> > > since 2006, it might be a good idea to track down why these values are
> > > set the way they are. Unfortunately, there is no clue in the commit
> > > logs, but maybe the author (pjd@, cc'd) can englighten us.
> >
> > I agree with that take on the situation, and it is why I asked
> > for a revert and investigation, rather than trying to solve
> > why we suddenly fail some regression tests.
> >
> > --
> > Rod Grimes
> > rgrimes at freebsd.org
> >
> >

-- 
Rod Grimes                                                 rgrimes at freebsd.org