Message-ID: <CAE=gft7567-2Lq7raJKrOpQ8UAvXTFWwPci=_GCRPET3nS=9SA@mail.gmail.com>
Date:   Tue, 27 Jul 2021 09:31:33 -0700
From:   Evan Green <evgreen@...omium.org>
To:     David Hildenbrand <david@...hat.com>
Cc:     Andrew Morton <akpm@...ux-foundation.org>,
        Michal Hocko <mhocko@...e.com>, Pavel Machek <pavel@....cz>,
        linux-api@...r.kernel.org, Alex Shi <alexs@...nel.org>,
        Alistair Popple <apopple@...dia.com>,
        Johannes Weiner <hannes@...xchg.org>,
        Joonsoo Kim <iamjoonsoo.kim@....com>,
        "Matthew Wilcox (Oracle)" <willy@...radead.org>,
        Miaohe Lin <linmiaohe@...wei.com>,
        Minchan Kim <minchan@...nel.org>,
        Suren Baghdasaryan <surenb@...gle.com>,
        Vlastimil Babka <vbabka@...e.cz>,
        LKML <linux-kernel@...r.kernel.org>, linux-mm@...ck.org
Subject: Re: [PATCH v4] mm: Enable suspend-only swap spaces

On Tue, Jul 27, 2021 at 5:21 AM David Hildenbrand <david@...hat.com> wrote:
>
> On 27.07.21 11:48, David Hildenbrand wrote:
> > On 27.07.21 02:12, Evan Green wrote:
> >> Add a new SWAP_FLAG_HIBERNATE_ONLY that adds a swap region but refuses
> >> to allow generic swapping to it. This region can still be wired up for
> >> use in suspend-to-disk activities, but will never have regular pages
> >> swapped to it. This flag will be passed in by utilities like swapon(8),
> >> usage would probably look something like: swapon -o hibernate /dev/sda2.
> >>
> >> Currently it's not possible to enable hibernation without also enabling
> >> generic swap for a given area. One semi-workaround for this is to delay
> >> the call to swapon() until just before attempting to hibernate, and then
> >> call swapoff() just after hibernate completes. This is somewhat kludgy,
> >> and also doesn't really work to keep swap out of the hibernate region.
> >> When hibernate begins, it starts by allocating a large chunk of memory
> >> for itself. This often ends up forcing a lot of data out into swap. By
> >> this time the hibernate region is eligible for generic swap, so swap
> >> ends up leaking into the hibernate region even with the workaround.
> >>
> >> There are a few reasons why usermode might want to be able to
> >> exclusively steer swap and hibernate. One reason relates to SSD wearing.
> >> Hibernate's endurance and speed requirements are different from swap.
> >> It may for instance be advantageous to keep hibernate in primary
> >> storage, but put swap in an SLC namespace. These namespaces are faster
> >> and have better endurance, but cost 3-4x in terms of capacity.
> >> Exclusively steering hibernate and swap enables system designers to
> >> accurately partition their storage without either wearing out their
> >> primary storage, or overprovisioning their fast swap area.
> >>
> >> Another reason to allow exclusive steering has to do with security.
> >> The requirements for designing systems with resilience against
> >> offline attacks are different between swap and hibernate. Swap
> >> effectively requires a dictionary of hashes, as pages can be added and
> >> removed arbitrarily, whereas hibernate only needs a single hash for the
> >> entire image. If you've set up block-level integrity for swap and
> >> image-level integrity for hibernate, then allowing swap blocks to
> >> possibly leak out to the hibernate region is problematic, since it
> >> creates swap pages not protected by any integrity.
> >>
> >> Swap regions with SWAP_FLAG_HIBERNATE_ONLY set will not appear in
> >> /proc/meminfo under SwapTotal and SwapFree, since they are not usable as
> >> general swap. These regions do still appear in /proc/swaps.
> >
> > Right, and they also don't account towards the memory overcommit
> > calculations.
> >
> > Thanks for extending the patch description!

No problem, thanks for all the brainwaves directed at this.

> >
> > [...]
> >
> >> +    if (swap_flags & SWAP_FLAG_HIBERNATE_ONLY) {
> >> +            if (IS_ENABLED(CONFIG_HIBERNATION)) {
> >> +                    if (swap_flags & ~SWAP_HIBERNATE_ONLY_VALID_FLAGS)
> >> +                            return -EINVAL;
> >> +
> >> +            } else {
> >> +                    return -EINVAL;
> >> +            }
> >> +    }
> >
> > We could do short
> >
> > if ((swap_flags & SWAP_FLAG_HIBERNATE_ONLY) &&
> >        (!IS_ENABLED(CONFIG_HIBERNATION) ||
> >         (swap_flags & ~SWAP_HIBERNATE_ONLY_VALID_FLAGS)))
> >       return -EINVAL;
> >
> > or
> >
> > if (swap_flags & SWAP_FLAG_HIBERNATE_ONLY))
> >       if (!IS_ENABLED(CONFIG_HIBERNATION) ||
> >               (swap_flags & ~SWAP_HIBERNATE_ONLY_VALID_FLAGS))
> >               return -EINVAL;
> >
> >> +
> >>      if (!capable(CAP_SYS_ADMIN))
> >>              return -EPERM;
> >>
> >> @@ -3335,16 +3366,20 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
> >>      if (swap_flags & SWAP_FLAG_PREFER)
> >>              prio =
> >>                (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
> >> +
> >> +    if (swap_flags & SWAP_FLAG_HIBERNATE_ONLY)
> >> +            p->flags |= SWP_HIBERNATE_ONLY;
> >>      enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
> >>
> >> -    pr_info("Adding %uk swap on %s.  Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
> >> +    pr_info("Adding %uk swap on %s.  Priority:%d extents:%d across:%lluk %s%s%s%s%s%s\n",
> >>              p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
> >>              nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
> >>              (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
> >>              (p->flags & SWP_DISCARDABLE) ? "D" : "",
> >>              (p->flags & SWP_AREA_DISCARD) ? "s" : "",
> >>              (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
> >> -            (frontswap_map) ? "FS" : "");
> >> +            (frontswap_map) ? "FS" : "",
> >> +            (p->flags & SWP_HIBERNATE_ONLY) ? "H" : "");
> >>
> >>      mutex_unlock(&swapon_mutex);
> >>      atomic_inc(&proc_poll_event);
> >>
> >
> > Looks like the cleanest alternative to me, as long as we don't want to
> > invent new interfaces.
> >
> > Acked-by: David Hildenbrand <david@...hat.com>
> >
>
> Pavel just mentioned uswsusp, and I wonder if it would be a possible
> alternative to this patch.

I think you're right that it would be possible to isolate the
hibernate image with uswsusp if you avoid using the SNAPSHOT_*SWAP*
ioctls. But I'd expect performance to suffer noticeably, since now
every page is making a round trip out to usermode and back. I'd still
very much use the HIBERNATE_ONLY flag if it were accepted, I think
there's value to it.

-Evan

Powered by blists - more mailing lists