Re: what can we do with frozen processes (or: calling sync with processes frozen)

["Rafael J. Wysocki" <rjw@sisk.pl>]

Hi,

On Sunday, 18 March 2007 00:21, Johannes Berg wrote:
> Hi,
> 
> When on vacation I was doing a bit more stuff w/o battery and noticed
> that I after my patches to implement proper pm_ops for PMU-based
> machines the machine would hang during suspend. Now on the way back I'm
> sitting in a train and it's working again. Couldn't really be the lower
> atmospheric pressure causing it ;) So I looked again and realised that
> when writing those patches I thought doing sys_sync() in the ->prepare
> callback would be a good idea. It is, as it turns out, not such a great
> idea at all, right now anyway.

Well, it's not a good idea indeed.

> When laptop mode is enabled [1], xfs will, upon getting a sync, try to
> wake up one of it's tasks and then wait for something that task does (as
> far as I can tell.) Even though I carry the patch to make xfs's
> workqueues non-freezable this event doesn't seem to happen when
> processes are frozen and so the system hangs. Not sure why that is but
> it isn't really the point right now, read on.

XFS uses more kernel threads that are frozen (see
http://lxr.free-electrons.com/source/fs/xfs/linux-2.6/xfs_buf.c?a=x86_64#1700)

> I have now removed the call to sys_sync() to double-verify and that has
> indeed cured the problem, but the bigger question really is whether it
> should be safe to call it or not [2], especially given that it would
> make sense to call it when processes can no longer submit IO.

We call sys_sync() in freeze_processes() _before_ kernel threads are frozen.
I don't think it's safe to call it any time later before the kernel threads are
thawed.

> The real thing we should be doing is think about a whole range of things
> you can do in the kernel and then think about whether they should be
> allowed to be done in the various stages. A lot of these things are at
> least somewhat documented for early boot stages (and you get into
> trouble when you do it wrong) but it seems that many developers never
> suspend their machines

True, but they should do this.

> so this needs to be documented in a more explicit 
> way rather than just breaking when you try (bootup is part of regular
> testing routine in an obvious way that suspend isn't)
> 
> Yup, this is a lot of hard work, but it could be done gradually as we
> run into these problems.
> 
> Back to the specific sys_sync() issue, what do people feel is
> appropriate? Should XFS get into even more trouble? It seems not to be
> able to guarantee full sync when processes are frozen right now, at
> least not the way sync is currently implemented.

See above.

I think at least some filesystems use freezable kernel threads and will
block if you ask them to sync after these threads have been frozen.

Greetings,
Rafael