Re: [PATCH RESEND 2/2] mm: zswap: use SRCU to synchronize with CPU hotunplug

[Yosry Ahmed <yosryahmed@google.com>]

On Tue, Jan 7, 2025 at 10:13 AM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Tue, Jan 7, 2025 at 10:03 AM Johannes Weiner <hannes@cmpxchg.org> wrote:
> >
> > On Tue, Jan 07, 2025 at 07:47:24AM +0000, Yosry Ahmed wrote:
> > > In zswap_compress() and zswap_decompress(), the per-CPU acomp_ctx of the
> > > current CPU at the beginning of the operation is retrieved and used
> > > throughout.  However, since neither preemption nor migration are disabled,
> > > it is possible that the operation continues on a different CPU.
> > >
> > > If the original CPU is hotunplugged while the acomp_ctx is still in use,
> > > we run into a UAF bug as the resources attached to the acomp_ctx are freed
> > > during hotunplug in zswap_cpu_comp_dead().
> > >
> > > The problem was introduced in commit 1ec3b5fe6eec ("mm/zswap: move to use
> > > crypto_acomp API for hardware acceleration") when the switch to the
> > > crypto_acomp API was made.  Prior to that, the per-CPU crypto_comp was
> > > retrieved using get_cpu_ptr() which disables preemption and makes sure the
> > > CPU cannot go away from under us.  Preemption cannot be disabled with the
> > > crypto_acomp API as a sleepable context is needed.
> > >
> > > Commit 8ba2f844f050 ("mm/zswap: change per-cpu mutex and buffer to
> > > per-acomp_ctx") increased the UAF surface area by making the per-CPU
> > > buffers dynamic, adding yet another resource that can be freed from under
> > > zswap compression/decompression by CPU hotunplug.
> > >
> > > There are a few ways to fix this:
> > > (a) Add a refcount for acomp_ctx.
> > > (b) Disable migration while using the per-CPU acomp_ctx.
> > > (c) Use SRCU to wait for other CPUs using the acomp_ctx of the CPU being
> > > hotunplugged. Normal RCU cannot be used as a sleepable context is
> > > required.
> > >
> > > Implement (c) since it's simpler than (a), and (b) involves using
> > > migrate_disable() which is apparently undesired (see huge comment in
> > > include/linux/preempt.h).
> > >
> > > Fixes: 1ec3b5fe6eec ("mm/zswap: move to use crypto_acomp API for hardware acceleration")
> > > Cc: <stable@vger.kernel.org>
> > > Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
> > > Reported-by: Johannes Weiner <hannes@cmpxchg.org>
> > > Closes: https://lore.kernel.org/lkml/20241113213007.GB1564047@cmpxchg.org/
> > > Reported-by: Sam Sun <samsun1006219@gmail.com>
> > > Closes: https://lore.kernel.org/lkml/CAEkJfYMtSdM5HceNsXUDf5haghD5+o2e7Qv4OcuruL4tPg6OaQ@mail.gmail.com/
> > > ---
> > >  mm/zswap.c | 31 ++++++++++++++++++++++++++++---
> > >  1 file changed, 28 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > index f6316b66fb236..add1406d693b8 100644
> > > --- a/mm/zswap.c
> > > +++ b/mm/zswap.c
> > > @@ -864,12 +864,22 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
> > >       return ret;
> > >  }
> > >
> > > +DEFINE_STATIC_SRCU(acomp_srcu);
> > > +
> > >  static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
> > >  {
> > >       struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
> > >       struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
> > >
> > >       if (!IS_ERR_OR_NULL(acomp_ctx)) {
> > > +             /*
> > > +              * Even though the acomp_ctx should not be currently in use on
> > > +              * @cpu, it may still be used by compress/decompress operations
> > > +              * that started on @cpu and migrated to a different CPU. Wait
> > > +              * for such usages to complete, any news usages would be a bug.
> > > +              */
> > > +             synchronize_srcu(&acomp_srcu);
> >
> > The docs suggest you can't solve it like that :(
> >
> > Documentation/RCU/Design/Requirements/Requirements.rst:
> >
> >   Also unlike other RCU flavors, synchronize_srcu() may **not** be
> >   invoked from CPU-hotplug notifiers, due to the fact that SRCU grace
> >   periods make use of timers and the possibility of timers being
> >   temporarily “stranded” on the outgoing CPU. This stranding of timers
> >   means that timers posted to the outgoing CPU will not fire until
> >   late in the CPU-hotplug process. The problem is that if a notifier
> >   is waiting on an SRCU grace period, that grace period is waiting on
> >   a timer, and that timer is stranded on the outgoing CPU, then the
> >   notifier will never be awakened, in other words, deadlock has
> >   occurred. This same situation of course also prohibits
> >   srcu_barrier() from being invoked from CPU-hotplug notifiers.
>
> Thanks for checking, I completely missed this. I guess it only works
> with SRCU if we use call_srcu(), but then we need to copy the pointers
> to a new struct to avoid racing with the CPU getting onlined again.
> Otherwise we can just bite the bullet and add a refcount, or use
> migrate_disable() despite that being undesirable.
>
> Do you have a favorite? :)

I briefly looked into refcounting. The annoying thing is that we need
to handle the race between putting the last refcount in
zswap_compress()/zswap_decompress(), and the CPU getting onlined again
and re-initializing the refcount. One way to do it would be to put all
dynamically allocated resources in a struct with the same struct with
the new refcount, and use RCU + refcounts to allocate and free the
struct as a whole.

I am leaning toward just disabling migration for now tbh unless there
are objections to that, especially this close to the v6.13 release.