Re: [linux-pm] [RFC] userland swsusp

[Nigel Cunningham <ncunningham@cyclades.com>]

Hi

On Thu, 2005-11-17 at 08:35, Pavel Machek wrote:
> Hi, all!
> 
> > > > > It's also implemented in the kernel, which is exactly the wrong place
> > > > > for this.  Pavel is doing this properly, why do you doubt him?
> > > > 
> > > > You yourself called it a hack not long ago.
> > > 
> > > I did, in the proud tradition of neat hacks.  It's a very nice
> > > accomplishment that this even works, and I'm impressed.
> > > 
> > > > I'm not sure why you think the userspace is the right place for
> > > > suspending.
> > > 
> > > If he can come up with an implementation that works, and puts stuff like
> > > the pretty spinning wheels and progress bars and encryption in
> > > userspace, that's great.  That stuff doesn't belong in the kerenel if we
> > > can possibly help it.
> > 
> > I can agree with putting splash screens and userspace stuff in
> > userspace. Suspend2 has had that too, since March. But the guts of
> > the
> 
> Well, I'd say that having to resort to netlink is ... not quite
> nice. You get all the complexity of having userspace running during
> suspend, and get very little benefit.

Mmm, but less complexity than with trying to do the whole suspend from
userspace. (I don't need to export pageflags, bio routines etc or work
around it by using /dev/kmem).

> > code is a different thing. Encryption - well, I think we're both using
> > cryptoapi now, so that's more easily done in the kernel.
> 
> Its not only encryption. It is encryption, compression, support for
> suspend over network, support for suspend into file. That's quite a
> lot of stuff.
> 
> > > Then propose a better way to do this, if you can see one.
> > 
> > We've done the user interface in userspace using netlink to
> > communication.
> > 
> > We've done storing a full image of memory by storing the page cache
> > separately to the rest of the image, so that it doesn't need to have an
> > atomic copy made. (Nothing that uses the page cache is running anyway).
> > Having done this, we can use the memory occupied by the page cache for
> > our atomic copy, and just reread the overwritten page cache pages if we
> > need to cancel the suspend. Suspend2 has done this since... beta18 I
> > think.
> 
> ...at expense of complexity, and hooks all over the kernel. Yes, if
> you modify kernel a bit, nothing will use the page cache.

Could you back your "hooks all over the kernel" statement up? I do have
some BUG_ON()s aimed at double checking that nothing bad happens, but
they never get hit and obviously aren't required to stop processes using
the page cache. All that's really required is to freeze processes.

> Anyway, I believe we have solution for that one. See Rafael's recent
> patches -- "only free as much memory as neccessary" should do the
> trick, without excessive complexity.

That's still imposing a 1/2 of memory limit, though.

> > > > I know that Pavel and I have such different ideas about what should be
> > > > done that it's not worth the effort.
> > > 
> > > I'm sorry that you feel this way.  I thought that after our meeting in
> > > July that things were different.
> > 
> > I'm sorry you came away with that impression. I want to work together,
> > but I'm not willing to settle for a minimalist implementation. Pavel, on
> > the other hand, wanted a minimalist implementation at first. He seems to
> > be changing his mind a bit now, but I'm not sure how far that will go.
> 
> Well, I do not want the complexity of two page sets. I think Rafael's
> patches will provide almost equivalent functionality. Other than that,
> all your features should be doable. I'm not saying I'm going to write
> those patches myself, but I'll certainly not reject them just because
> they are too big.

I'm sorry for making you think that having two pagesets is a complex
issues. I know that when I first did it, I put tight restrictions on
memory usage while the first pageset was written and used a separate
memory pool. Since then, I've realised a far simpler way of handling
this, and the code has been greatly simplified. In essence, all you need
to do is make your I/O code generic enough that it can be passed a list
of pages to write and put page cache pages in a separate list when
figuring out what pages need to be saved. Then you save those pages
before doing your atomic copy of the other pages, and reload them after
restoring the atomic copy at resume time.

Regards,

Nigel


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/