Re: [PATCH] Thaw userspace and kernel space separately.

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

On Monday, 23 October 2006 18:51, Andrew Morton wrote:
> > On Mon, 23 Oct 2006 22:00:11 +1000 Nigel Cunningham <ncunningham@linuxmail.org> wrote:
> > Hi.
> > 
> > On Mon, 2006-10-23 at 12:26 +0200, Rafael J. Wysocki wrote:
> > > On Monday, 23 October 2006 01:48, Nigel Cunningham wrote:
> > > > Modify process thawing so that we can thaw kernel space without thawing
> > > > userspace, and thaw kernelspace first. This will be useful in later
> > > > patches, where I intend to get swsusp thawing kernel threads only before
> > > > seeking to free memory.
> > > 
> > > Please explain why you think it will be necessary/useful.
> > > 
> > > I remember a discussion about it some time ago that didn't indicate
> > > we would need/want to do this.
> > 
> > This is needed to make suspending faster and more reliable when the
> > system is in a low memory situation. Imagine that you have a number of
> > processes trying to allocate memory at the time you're trying to
> > suspend. They want so much memory that when you come to prepare the
> > image, you find that you need to free pages. But your swapfile is on
> > ext3, and you've just frozen all processes, so any attempt to free
> > memory could result in a deadlock while the vm tries to swap out pages
> > using the frozen kjournald. So you need to thaw processes to free the
> > memory. But thawing processes will start the processes allocating memory
> > again, so you'll be fighting an uphill battle.
> > 
> > If you can only thaw the kernel threads, you can free memory without
> > restarting userspace or deadlocking against a frozen kjournald.
> > 
> 
> kjournald will not participate in writing to swapfiles.
> 
> The situation where we would need this feature is where the loop driver is
> involved in the path-to-disk.  But I doubt if that's a thing we'd want to
> support.
> 
> otoh there may be other kernel threads which are a saner thing to have in
> the swapout path and which we do want to support.  md_thread, perhaps?

md_thread needs some consideration I think.  Having a swapfile on RAID
is a legit thing and we should support that.


-- 
You never change things by fighting the existing reality.
		R. Buckminster Fuller