Re: [PATCH] xen: speed up grant-table reclaim

From: Demi Marie Obenour
Date: Tue Jun 13 2023 - 11:12:03 EST

On Tue, Jun 13, 2023 at 08:45:31AM +0200, Juergen Gross wrote:
> On 12.06.23 22:09, Demi Marie Obenour wrote:
> > On Mon, Jun 12, 2023 at 08:27:59AM +0200, Juergen Gross wrote:
> > > On 10.06.23 17:32, Demi Marie Obenour wrote:
> > > > When a grant entry is still in use by the remote domain, Linux must put
> > > > it on a deferred list.
> > >
> > > This lacks quite some context.
> > >
> > > The main problem is related to the grant not having been unmapped after
> > > the end of a request, but the side granting the access is assuming this
> > > should be the case.
> >
> > The GUI agent has relied on deferred grant reclaim for as long as it has
> > existed. One could argue that doing so means that the agent is misusing
> > gntalloc, but this is not documented anywhere. A better fix would be to
> > use IOCTL_GNTDEV_SET_UNMAP_NOTIFY in the GUI daemon.
>
> I don't think this is a gntalloc specific problem. You could create the same
> problem with a kernel implementation.

That is true.

> And yes, using IOCTL_GNTDEV_SET_UNMAP_NOTIFY might be a good idea.
>
> > > In general this means that the two sides implementing the protocol don't
> > > agree how it should work, or that the protocol itself has a flaw.
> >
> > What would a better solution be? This is going to be particularly
> > tricky with Wayland, as the wl_shm protocol makes absolutely no
> > guarantees that compositors will promptly release the mapping and
> > provides no way whatsoever for Wayland clients to know when this has
> > happened. Relying on an LD_PRELOAD hack is not sustainable.
> >
> > > > Normally, this list is very short, because
> > > > the PV network and block protocols expect the backend to unmap the grant
> > > > first.
> > >
> > > Normally the list is just empty. Only in very rare cases like premature
> > > PV frontend module unloading it is expected to see cases of deferred
> > > grant reclaims.
> >
> > In the case of a system using only properly-designed PV protocols
> > implemented in kernel mode I agree. However, both libxenvchan and the
> > Qubes GUI protocol are implemented in user mode and this means that if
> > the frontend process (the one that uses gntalloc) crashes, deferred
> > grant reclaims will occur.
>
> This would be the user space variant of frontend driver unloading.

Yes and no. The main difference is that a frontend driver can prevent
itself from being unloaded while it still has grants outstanding,
whereas a user space process cannot prevent itself from being killed.
The closest a user space process can get to making itself unkillable is
to use mlockall() and mark itself ineligible for the OOM killer. This
would likely require anything that used grants (including vchans!) to be
a privileged daemon that exposes an API over an AF_UNIX socket, which
would be a substantial change in the case of Qubes OS.

> > Worse, it is possible for the domain to use
> > the grant in a PV protocol. If the PV backend driver then maps and
> > unmaps the grant and then tells the frontend driver to reclaim it, but
> > the backend userspace process (the one using gntdev) maps it before the
> > frontend does reclaim it, the frontend will think the backend is trying
> > to exploit XSA-396 and freeze the connection.
>
> Let me sum up how I understand above statement:
>
> 1. Frontend F in DomA is granting DomB access via grant X for PV-device Y.
> 2. DomB backend B for PV-device Y is mapping the page using grant X and uses it.
> 3. DomB backend B unmaps grant X and signals end of usage to DomA frontend F.
> 4. DomB userspace process maps grant X
> 5. DomA frontend F tries to reclaim grant X and fails due to DomB userspace
> mapping

All of these are correct, but there is some context that is missing.

> So why would DomB userspace process map grant X? This would imply a malicious
> process in DomB. This might be possible, but I don't see how such an attack
> would relate to your problem. It could happen with any malicious userspace
> program with root access in a domain running a PV backend.

The complete sequence of events is as follows:

1. DomA userspace process C allocates grant X via gntalloc.
2. DomA userspace process C tells DomB userspace process to map grant X
via e.g. a vchan message.
3. DomB userspace process S receives message telling it to map grant X.
4. (optional) DomB userspace process S checks that DomA userspace
process C is still alive and finds that it is.
5. DomB userspace process S gets preempted.
6. DomA userspace process C exits or is killed.
7. DomA kernel KA deallocates grant X.
8. Frontend F in DomA grants DomB access via grant X for PV-device Y.
9. DomB backend B for PV-device Y is mapping the page using grant X and uses it.
10. DomB backend B unmaps grant X and signals end of usage to DomA frontend F.
11. DomB userspace process S is scheduled again. It finally gets the
chance to use gntdev to map grant X.
12. DomA frontend F tries to reclaim grant X and fails due to DomB userspace
mapping by S.

Given the current kernel API, I’m not aware of any way to avoid this race
unless C (and possibly S) are unkillable. Even with the check in step 4
there is still a TOCTOU problem. Neither kernel has any visibility into
the userspace messaging protocol, so DomA has no way of knowing that S
has been told to map the grant in the future.

There are four other potential solutions I can think of.

1. If a gntalloc handle is closed while it still has mapped grants, leak
the remaining grants. It’s okay to reclaim the pages those grants
refer to, but the grants themselves would not be usable until reboot.

2. Assume that “backend has not unmapped grant” is due to this race, and
fall back to deferred reclaim instead of closing the connection. For
this to work, either the data would need to be discarded by F, or S
would need to follow some sort of protocol where it always checks
that C is alive after mapping the grant but before writing to the
newly mapped page.

3. Implement AF_VSOCK or a similar protocol for kernel-mediated inter-VM
communication, and send grants as control messages. This means that
both kernels are aware of grant lifetimes and can avoid this race.

4. Same as 3, except that grant table entries are replaced by
capabilities that must be sent over Argo. This is obviously the best
long-term solution, but it also requires massive changes in all PV
protocols, so it isn’t practical in the short term.

Additionally, I wonder if grant operations should require CAP_SYS_RAWIO.
Userspace use of grants very much seems like a raw I/O operation,
inasmuch as it is bypassing the protections Linux normally provides.

> > > > However, Qubes OS's GUI protocol is subject to the constraints
> > > > of the X Window System, and as such winds up with the frontend unmapping
> > > > the window first. As a result, the list can grow very large, resulting
> > > > in a massive memory leak and eventual VM freeze.
> > >
> > > I do understand that it is difficult to change the protocol and/or
> > > behavior after the fact, or that performance considerations are in the
> > > way of doing so.
> >
> > Would the correct fix be to use IOCTL_GNTDEV_SET_UNMAP_NOTIFY? That
> > would require that the agent either create a new event channel for each
> > window or maintain a pool of event channels, but that should be doable.
>
> I think this sounds like a promising idea.

Me too.

> > This still does not solve the problem of the frontend exiting
> > unexpectedly, though.
>
> Such cases need to be handled via deferred reclaim. You could add a flag
> to struct deferred_entry when called through gntalloc_vma_close(),
> indicating that this is an expected deferred reclaim not leading to
> error messages.

Deferred reclaim can handle this most of the time, but there is still a
race (see above).

> > > > To partially solve this problem, make the number of entries that the VM
> > > > will attempt to free at each iteration tunable. The default is still
> > > > 10, but it can be overridden at compile-time (via Kconfig), boot-time
> > > > (via a kernel command-line option), or runtime (via sysfs).
> > >
> > > Is there really a need to have another Kconfig option for this? AFAICS
> > > only QubesOS is affected by the problem you are trying to solve. I don't
> > > see why you can't use the command-line option or sysfs node to set the
> > > higher reclaim batch size.
> >
> > Fair. In practice, Qubes OS will need to use the sysfs node, since
> > the other two do not work with in-VM kernels.
> >
> > > > Fixes: 569ca5b3f94c ("xen/gnttab: add deferred freeing logic")
> > >
> > > I don't think this "Fixes:" tag is appropriate. The mentioned commit didn't
> > > have a bug. You are adding new functionality on top of it.
> >
> > I’ll drop the "Fixes:" tag, but I will keep the "Cc: stable@xxxxxxxxxxxxxxx"
> > as I believe this patch meets the following criterion for stable
> > backport (from Documentation/process/stable-kernel-rules.rst):
> >
> > Serious issues as reported by a user of a distribution kernel may also
> > be considered if they fix a notable performance or interactivity issue.
> >
>
> Sure, this seems appropriate.

Thank you.

> > > > Cc: stable@xxxxxxxxxxxxxxx
> > > > Signed-off-by: Demi Marie Obenour <demi@xxxxxxxxxxxxxxxxxxxxxx>
> > > > ---
> > > > drivers/xen/Kconfig | 12 ++++++++++++
> > > > drivers/xen/grant-table.c | 40 ++++++++++++++++++++++++++++-----------
> > > > 2 files changed, 41 insertions(+), 11 deletions(-)
> > > >
> > > > diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig
> > > > index d5d7c402b65112b8592ba10bd3fd1732c26b771e..8f96e1359eb102d6420775b66e7805004a4ce9fe 100644
> > > > --- a/drivers/xen/Kconfig
> > > > +++ b/drivers/xen/Kconfig
> > > > @@ -65,6 +65,18 @@ config XEN_MEMORY_HOTPLUG_LIMIT
> > > > This value is used to allocate enough space in internal
> > > > tables needed for physical memory administration.
> > > > +config XEN_GRANTS_RECLAIM_PER_ITERATION
> > > > + int "Default number of grant entries to reclaim per iteration"
> > > > + default 10
> > > > + range 10 4294967295
> > > > + help
> > > > + This sets the default value for the grant_table.free_per_iteration
> > > > + kernel command line option, which sets the number of grants that
> > > > + Linux will try to reclaim at once. The default is 10, but
> > > > + workloads that make heavy use of gntalloc will likely want to
> > > > + increase this. The current value can be accessed and/or modified
> > > > + via /sys/module/grant_table/parameters/free_per_iteration.
> > > > +
> > >
> > > Apart from the fact that I don't like adding a new Kconfig option, the help
> > > text is not telling the true story. Heavy use of gntalloc isn't to blame, but
> > > the fact that your PV-device implementation is based on the reclaim
> > > functionality. TBH, someone not familiar with the grant reclaim will have no
> > > chance to select a sensible value based on your help text.
> >
> > That’s a good point. Qubes OS will need to use the sysfs value anyway
> > in order to support in-VM kernels, so the Kconfig option is not really
> > useful.
>
> Nice.

Yeah, having one less Kconfig option is nice.

> > > > config XEN_SCRUB_PAGES_DEFAULT
> > > > bool "Scrub pages before returning them to system by default"
> > > > depends on XEN_BALLOON
> > > > diff --git a/drivers/xen/grant-table.c b/drivers/xen/grant-table.c
> > > > index e1ec725c2819d4d5dede063eb00d86a6d52944c0..fa666aa6abc3e786dddc94f895641505ec0b23d8 100644
> > > > --- a/drivers/xen/grant-table.c
> > > > +++ b/drivers/xen/grant-table.c
> > > > @@ -498,14 +498,20 @@ static LIST_HEAD(deferred_list);
> > > > static void gnttab_handle_deferred(struct timer_list *);
> > > > static DEFINE_TIMER(deferred_timer, gnttab_handle_deferred);
> > > > +static atomic64_t deferred_count;
> > > > +static atomic64_t leaked_count;
> > > > +static unsigned int free_per_iteration = CONFIG_XEN_GRANTS_RECLAIM_PER_ITERATION;
> > > > +
> > > > static void gnttab_handle_deferred(struct timer_list *unused)
> > > > {
> > > > - unsigned int nr = 10;
> > > > + unsigned int nr = READ_ONCE(free_per_iteration);
> > > > + const bool ignore_limit = nr == 0;
> > > > struct deferred_entry *first = NULL;
> > > > unsigned long flags;
> > > > + size_t freed = 0;
> > > > spin_lock_irqsave(&gnttab_list_lock, flags);
> > > > - while (nr--) {
> > > > + while ((ignore_limit || nr--) && !list_empty(&deferred_list)) {
> > > > struct deferred_entry *entry
> > > > = list_first_entry(&deferred_list,
> > > > struct deferred_entry, list);
> > > > @@ -515,10 +521,13 @@ static void gnttab_handle_deferred(struct timer_list *unused)
> > > > list_del(&entry->list);
> > > > spin_unlock_irqrestore(&gnttab_list_lock, flags);
> > > > if (_gnttab_end_foreign_access_ref(entry->ref)) {
> > > > + uint64_t ret = atomic64_sub_return(1, &deferred_count);
> > > > put_free_entry(entry->ref);
> > > > - pr_debug("freeing g.e. %#x (pfn %#lx)\n",
> > > > - entry->ref, page_to_pfn(entry->page));
> > > > + pr_debug("freeing g.e. %#x (pfn %#lx), %llu remaining\n",
> > > > + entry->ref, page_to_pfn(entry->page),
> > > > + (unsigned long long)ret);
> > >
> > > I'd prefer not to issue lots of prints (being it debug or info ones) if the
> > > reclaim is expected to happen with a specific PV-device.
> > >
> > > My preferred solution would be a per-device flag, but at least you should
> > > switch to pr_*_ratelimited(). Same applies further down.
> >
> > What do you mean by “per-device flag”? These grants are allocated by
> > userspace using gntalloc, so there is no PV device on which the flag
> > could be set.
>
> In this case the flag should relate to the file pointer used for
> gntalloc_open().

What about adding a warning if a userspace process exits without having
cleaned up all of its grants, and a module parameter that controls
whether the remaining grants are reclaimed or leaked?
--
Sincerely,
Demi Marie Obenour (she/her/hers)
Invisible Things Lab

Attachment: signature.asc
Description: PGP signature