Re: [PATCH V2 2/2] fs: print a message when freezing/unfreezing filesystems

[Dave Chinner <david@fromorbit.com>]

On Mon, May 19, 2014 at 11:43:13AM +0200, Jan Kara wrote:
> On Fri 16-05-14 10:11:56, Dave Chinner wrote:
> > On Fri, May 16, 2014 at 01:19:09AM +0200, Mateusz Guzik wrote:
> > > On Fri, May 16, 2014 at 08:51:41AM +1000, Dave Chinner wrote:
> > > > On Fri, May 16, 2014 at 12:34:40AM +0200, Mateusz Guzik wrote:
> > > > > On Fri, May 16, 2014 at 08:21:35AM +1000, Dave Chinner wrote:
> > > > > > > IOW, a new column in mountinfo. For frozen filesystems it would contain
> > > > > > > 'frozen_by=[%s]:[%d]' (escaped comm, pid).
> > > > > > 
> > > > > > I really don't see that the process that froze the filesystem is
> > > > > > particularly useful - it many cases that process is long gone (e.g.
> > > > > > fsfreeze is being used to allow a HW array to take a snapshot). Just
> > > > > > the fact it is in the process of freezing (if stuck, stack trace in
> > > > > > sysrq-w should be present) or frozen (freezing process may be long
> > > > > > gone, and is mostly irrelevant because you're now tracking down why
> > > > > > a thaw hasn't happened)...
> > > > > 
> > > > > There are deamons which perform freezing and unfreezing on their own.
> > > > > Thus storing the name along with pid helps to determine whether someone
> > > > > went behind such daemon's back, or maybe it's the daemon which "forgot" to
> > > > > unfreeze after all.
> > > > 
> > > > Such a daemon should be logging the fact that it's freezing and
> > > > thawing the filesystem. The kernel is not the place to track what
> > > > buggy userspace applications are doing wrong.
> > > > 
> > > 
> > > Except there is no log entry if /var got frozen (and this is not an
> > > imaginary example).
> > 
> > Freezing the filesystem that the freezing daemon logs to is, well, a
> > major application architecture fail. Sorry, catering for the lowest
> > common denominator (i.e. stupidity) is not an valid argument for
> > adding stuff to the kernel....
>   Sure it's not a good architecture but it happens either because of a bug
> or a wrong architecture. So you need to debug it and traces from sysrq-w
> don't tell you who froze the filesystem. Currently you have to use
> tracepoints or similar stuff to find that out (e.g. in one case I was
> debugging it was rpm running a post-install script that froze the fs,
> believe me that was really unexpected :)). But tracepoints aren't useful
> after the fact so sometimes it would be useful to be able to find out
> after the fact who froze the fs (PID and command name to help with
> situations when the process isn't running anymore). Since this is mostly
> debug stuff I'd be OK with dumping this information on sysrq request or as
> Ted suggested from some fs-freeze hang check timer... Hmm?

Sure, along with all the memory allocation, copying and freeing
stuff that isn't guaranteed to be around if a hang occurs. And it
has to be per-sb, because you can indepenently freeze multiple
filesystems at the same time, and so on. I just don't see the
complexity as being worthwhile given how rare freeze problems are...

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com