Re: [RFC][PATCH][EXPERIMENTAL] Make kernel threads nonfreezable by default

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

On Monday, 28 May 2007 12:33, Pavel Machek wrote:
> Hi!
> 
> > Index: linux-2.6.22-rc3/Documentation/power/kernel_threads.txt
> > ===================================================================
> > --- linux-2.6.22-rc3.orig/Documentation/power/kernel_threads.txt
> > +++ linux-2.6.22-rc3/Documentation/power/kernel_threads.txt
> > @@ -1,13 +1,22 @@
> > -KERNEL THREADS
> > +The Freezer and Kernel Threads
> >  
> > -
> > -Freezer
> > -
> > -Upon entering a suspended state the system will freeze all
> > -tasks. This is done by delivering pseudosignals. This affects
> > -kernel threads, too. To successfully freeze a kernel thread
> > -the thread has to check for the pseudosignal and enter the
> > -refrigerator. Code to do this looks like this:
> > +Before entering a system-wide suspend state as well as before creating a
> > +hibernation snapshot image the system will freeze all tasks, which is done
> > +by delivering fake signals.  This affects kernel threads too, but they won't be
> > +frozen unless they declare that they want to.  For this purpose
> 
> Well... and unless thread that does disk writes or DMA _wants_ to, you
> have nice disk corruption... It should be pointed out that it is not
> voluntary for those types of threads.

Well, that really depends.

I think the only thing that must not be written to disk is filesystem stuff.
The other things (eg. writing directly to block devices etc.) doesn't hurt us.

Also, I think that DMA-ing while we're creating the image should be impossible
due to devices being frozen at that point.

That said, I think we need to write some better freezer documentation ASAP.

Greetings,
Rafael