Re: mysterious setting of B_DIRECT?

On Thu, Apr 25, 2024 at 8:51 PM Rick Macklem <rick.macklem@gmail.com> wrote:
>
> On Thu, Apr 25, 2024 at 8:09 PM Konstantin Belousov <kib@freebsd.org> wrote:
> >
> > On Thu, Apr 25, 2024 at 07:49:23PM -0700, Rick Macklem wrote:
> > > Hi,
> > >
> > > This week I have been doing active testing as a part of an IETF
> > > bakeathon for NFSv4. During the week I had a NFSv4 client
> > > crash. On the surface, it is straightforward, in that it called
> > > ncl_doio_directwrite() and the field called b_caller1 was NULL.
> > >
> > > Now, here's the weird part...
> > > ncl_doio_directwrite() should never be called because B_DIRECT
> > > should never be set. (The only place B_DIRECT gets set in the code
> > > is never currently executed.)
> > Do you mean the place in nfs_directio_write()?  And the fact that
> > IO_SYNC is always set.
> Yes.
>
> >
> > >
> > > I have a patch that clears out the "never to be executed" code and
> > > this seems to avoid the patch, since with the patch, ncl_doio_directwrite()
> > > no longer exists.
> > >
> > > What I cannot figure out is how B_DIRECT got set?
> > > I can note that UFS was under heavy load when the client crashed,
> > > but I cannot see how a UFS "struct buf" would become a NFS "struct buf"
> > > without b_flags being set to 0.
> >
> > There are also vfs_bio_brelse()/vfs_bio_setflags() functions which can
> > set B_DIRECT.  On the other hand, they are not used by nfs client.
> Yes, again.
>
> >
> > What was the overall state of the buffer with the B_DIRECT flag?  Which
> > vnode it was assigned to?
> Unfortunately I was in a hurry and didn't get more info.
> And, since I have never seen this crash before, I doubt I'll be able
> to reproduce it.
Oh, and I will put the cleanup patch on phabricator. I didn't see the
crash again
during a few days of testing with the patch. This makes sense, since it gets
rid of ncl_doio_directwrite().

>
> Thanks, rick