Re: Getting rid of freezer for suspend [was Re: [fuse-devel] [PATCH] fuse: make fuse daemon frozen along with kernel threads]

From: "Rafael J. Wysocki" <rjw@sisk.pl>
To: Miklos Szeredi <miklos@szeredi.hu>
Cc: Pavel Machek <pavel@ucw.cz>,
	Goswin von Brederlow <goswin-v-b@web.de>,
	Li Fei <fei.li@intel.com>,
	len.brown@intel.com, mingo@redhat.com, peterz@infradead.org,
	biao.wang@intel.com, linux-pm@vger.kernel.org,
	fuse-devel@lists.sourceforge.net, linux-kernel@vger.kernel.org,
	chuansheng.liu@intel.com
Subject: Re: Getting rid of freezer for suspend [was Re: [fuse-devel] [PATCH] fuse: make fuse daemon frozen along with kernel threads]
Date: Mon, 11 Feb 2013 13:08:56 +0100	[thread overview]
Message-ID: <1660768.prnx6s70rD@vostro.rjw.lan> (raw)
In-Reply-To: <CAJfpegtNUr+AJXZcKmLEGBYqu20DVFwzg-OvOk_nS3fBS3wd=Q@mail.gmail.com>

On Monday, February 11, 2013 11:11:40 AM Miklos Szeredi wrote:
> On Mon, Feb 11, 2013 at 12:31 AM, Rafael J. Wysocki <rjw@sisk.pl> wrote:
> > On Sunday, February 10, 2013 07:55:05 PM Pavel Machek wrote:
> 
> >> Well, from freezer you need:
> >>
> >> 1) user process frozen.
> >>
> >> 2) essential locks _not_ held so that block devices are still functional.
> >>
> >> > > > mmap... what is problem with mmap? For suspend, memory is powered, so
> >> > > > you can permit people changing it.
> >> > >
> >> > > Suppose mmap is used to make the registers of some device available to user
> >> > > space.  Yes, that can happen.
> >>
> >> "Don't do it, then". Yes, can happen, but hopefully is not too common
> >> these days. [And... freezer doing 1) but not 2) would be enough to
> >> handle that. Freezer doing 1) but not 2) would also be simpler...]
> >
> > Again, I'm not sure what you mean.
> >
> > Are you trying to say that it would be OK to freeze user space tasks in
> > the D state?
> 
> I think that's what Pavel is saying.   Processes in D state sleeping
> on non-device mutexes _are_ actually OK to freeze.  And that would
> nicely solve the fuse freeze problem.

That's potentially deeadlock-prone, because a task waiting for mutex X may
very well be holding mutex Y, so if there's another task waiting for mutex Y,
it needs to be frozen at the same time.

> The only little detail is how do we implement that...

This means the only way I can see would be to hack the mutex code so that the
try_to_freeze() was called for user space tasks after the
sched_preempt_enable_no_resched() before schedule().

That shouldn't be a big deal performance-wise, because we are in the slow
path anyway then.  I'm not sure if Peter Z will like it, though.

Moreover, a task waiting for a mutex may be holding a semaphore or be
participating in some other mutual-exclusion mechanism, so we'd need to
address them all.  Plus, as noted by Pavel, freezing those things would make
it difficult to save hibernation images to us.

What about having a "freeze me after all of my children" flag that will be
inherited from parents?  Would that help the fuse case?

Rafael

-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.