Re: [PATCH 5/8] xfs_db: check that metadata updates have been committed

[Brian Foster <bfoster@redhat.com>]

On Thu, Feb 20, 2020 at 08:58:40AM -0800, Darrick J. Wong wrote:
> On Thu, Feb 20, 2020 at 09:06:23AM -0500, Brian Foster wrote:
> > On Wed, Feb 19, 2020 at 05:42:13PM -0800, Darrick J. Wong wrote:
> > > From: Darrick J. Wong <darrick.wong@oracle.com>
> > > 
> > > Add a new function that will ensure that everything we scribbled on has
> > > landed on stable media, and report the results.
> > > 
> > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > > ---
> > >  db/init.c |   14 ++++++++++++++
> > >  1 file changed, 14 insertions(+)
> > > 
> > > 
> > > diff --git a/db/init.c b/db/init.c
> > > index 0ac37368..e92de232 100644
> > > --- a/db/init.c
> > > +++ b/db/init.c
> > > @@ -184,6 +184,7 @@ main(
> > >  	char	*input;
> > >  	char	**v;
> > >  	int	start_iocur_sp;
> > > +	int	d, l, r;
> > >  
> > >  	init(argc, argv);
> > >  	start_iocur_sp = iocur_sp;
> > > @@ -216,6 +217,19 @@ main(
> > >  	 */
> > >  	while (iocur_sp > start_iocur_sp)
> > >  		pop_cur();
> > > +
> > > +	libxfs_flush_devices(mp, &d, &l, &r);
> > > +	if (d)
> > > +		fprintf(stderr, _("%s: cannot flush data device (%d).\n"),
> > > +				progname, d);
> > > +	if (l)
> > > +		fprintf(stderr, _("%s: cannot flush log device (%d).\n"),
> > > +				progname, l);
> > > +	if (r)
> > > +		fprintf(stderr, _("%s: cannot flush realtime device (%d).\n"),
> > > +				progname, r);
> > > +
> > > +
> > 
> > Seems like we could reduce some boilerplate by passing progname into
> > libxfs_flush_devices() and letting it dump out of the error messages,
> > unless there's some future code that cares about individual device error
> > state.
> 
> Such a program could call libxfs_flush_devices directly, as we do here.
> 

Right.. but does anything actually care about that level of granularity
right now beyond having a nicer error message?

> Also, progname is defined in libxfs so we don't even need to pass it as
> an argument.
> 

Ok.

> I had originally thought that we should try not to add fprintf calls to
> libxfs because libraries aren't really supposed to be doing things like
> that, but perhaps you're right that all of this should be melded into
> something else.
> 

Yeah, fair point, though I guess it depends on the particular library. 

> > That said, it also seems the semantics of libxfs_flush_devices() are a
> > bit different from convention. Just below we invoke
> > libxfs_device_close() for each device (rather than for all three), and
> > device_close() also happens to call fsync() and platform_flush_device()
> > itself...
> 
> Yeah, the division of responsibilities is a little hazy here -- I would
> think that unmounting a filesystem should flush all the memory caches
> and then the disk cache, but OTOH it's the utility that opens the
> devices and should therefore flush and close them.
> 
> I dunno.  My current thinking is that libxfs_umount should call
> libxfs_flush_devices() and print error messages as necessary, and return
> error codes as appropriate.  xfs_repair can then check the umount return
> value and translate that into exit(1) as required.  The device_close
> functions will fsync a second time, but that shouldn't be a big deal
> because we haven't dirtied anything in the meantime.
> 
> Thoughts?
> 

I was thinking of having a per-device libxfs_device_flush() along the
lines of libxfs_device_close() and separating out that functionality,
but one could argue we're also a bit inconsistent between libxfs_init()
opening the devices and having to close them individually. I think
having libxfs_umount() do a proper purge -> flush and returning any
errors instead is a fair tradeoff for simplicity. Removing the
flush_devices() API also eliminates risk of somebody incorrectly
attempting the flush after the umount frees the buftarg structures
(without reinitializing pointers :P).

Brian

> --D
> 
> > Brian
> > 
> > >  	libxfs_umount(mp);
> > >  	if (x.ddev)
> > >  		libxfs_device_close(x.ddev);
> > > 
> > 
>