Re: [PATCH] hibernate: unlock swap bdev for writing when uswsusp is active

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Sun, Mar 01, 2020 at 10:35:36PM +0100, Rafael J. Wysocki wrote:
> On Sat, Feb 29, 2020 at 9:02 PM Domenico Andreoli
> <domenico.andreoli@linux.com> wrote:
> >
> > On Sat, Feb 29, 2020 at 10:38:20AM -0800, Darrick J. Wong wrote:
> > > On Sat, Feb 29, 2020 at 07:07:16PM +0100, Domenico Andreoli wrote:
> > > > On Sat, Feb 29, 2020 at 09:08:25AM -0800, Darrick J. Wong wrote:
> > > > > From: Darrick J. Wong <darrick.wong@oracle.com>
> > > > >
> > > > > It turns out that there /is/ one use case for programs being able to
> > > > > write to swap devices, and that is the userspace hibernation code.  The
> > > > > uswsusp ioctls allow userspace to lease parts of swap devices, so turn
> > > > > S_SWAPFILE off when invoking suspend.
> > > > >
> > > > > Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices")
> > > > > Reported-by: Domenico Andreoli <domenico.andreoli@linux.com>
> > > > > Reported-by: Marian Klein <mkleinsoft@gmail.com>
> > > >
> > > > I also tested it yesterday but was not satisfied, unfortunately I did
> > > > not come with my comment in time.
> > > >
> > > > Yes, I confirm that the uswsusp works again but also checked that
> > > > swap_relockall() is not triggered at all and therefore after the first
> > > > hibernation cycle the S_SWAPFILE bit remains cleared and the whole
> > > > swap_relockall() is useless.
> > > >
> > > > I'm not sure this patch should be merged in the current form.
> > >
> > > NNGGHHGGHGH /me is rapidly losing his sanity and will soon just revert
> > > the whole security feature because I'm getting fed up with people
> > > yelling at me *while I'm on vacation* trying to *restore* my sanity.  I
> > > really don't want to be QAing userspace-directed hibernation right now.
> >
> > Maybe we could proceed with the first patch to amend the regression and
> > postpone the improved fix to a later patch? Don't loose sanity for this.
> 
> I would concur here.
> 
> > > ...right, the patch is broken because we have to relock the swapfiles in
> > > whatever code executes after we jump back to the restored kernel, not in
> > > the one that's doing the restoring.  Does this help?
> >
> > I made a few unsuccessful attempts in kernel/power/hibernate.c and
> > eventually I'm switching to qemu to speed up the test cycle.
> >
> > > OTOH, maybe we should just leave the swapfiles unlocked after resume.
> > > Userspace has clearly demonstrated the one usecase for writing to the
> > > swapfile, which means anyone could have jumped in while uswsusp was
> > > running and written whatever crap they wanted to the parts of the swap
> > > file that weren't leased for the hibernate image.
> >
> > Essentially, if the hibernation is supported the swapfile is not totally
> > safe.
> 
> But that's only the case with the userspace variant, isn't it?

Yes.

> > Maybe user-space hibernation should be a separate option.
> 
> That actually is not a bad idea at all in my view.

The trouble with kconfig options is that the distros will be pressued
into setting CONFIG_HIBERNATE_USERSPACE=y to avoid regressing their
uswsusp users, which makes the added security code pointless.  As this
has clearly sucked me into a conflict that I don't have the resources to
pursue, I'm going to revert the write patch checks and move on with
life.

--D

> Thanks!