Re: [RESEND PATCH v8 00/11] Fix BUG_ON in vfio_iommu_group_notifier()

[Robin Murphy]

On 2022-05-02 17:42, Jason Gunthorpe wrote:
> On Mon, May 02, 2022 at 12:12:04PM -0400, Qian Cai wrote:
> > On Mon, Apr 18, 2022 at 08:49:49AM +0800, Lu Baolu wrote:
> > > Hi Joerg,
> > >
> > > This is a resend version of v8 posted here:
> > > https://lore.kernel.org/linux-iommu/20220308054421.847385-1-baolu.lu@xxxxxxxxxxxxxxx/
> > > as we discussed in this thread:
> > > https://lore.kernel.org/linux-iommu/Yk%2Fq1BGN8pC5HVZp@xxxxxxxxxx/
> > >
> > > All patches can be applied perfectly except this one:
> > >   - [PATCH v8 02/11] driver core: Add dma_cleanup callback in bus_type
> > > It conflicts with below refactoring commit:
> > >   - 4b775aaf1ea99 "driver core: Refactor sysfs and drv/bus remove hooks"
> > > The conflict has been fixed in this post.
> > >
> > > No functional changes in this series. I suppress cc-ing this series to
> > > all v8 reviewers in order to avoid spam.
> > >
> > > Please consider it for your iommu tree.
> >
> > Reverting this series fixed an user-after-free while doing SR-IOV.
> >
> >   BUG: KASAN: use-after-free in __lock_acquire
> >   Read of size 8 at addr ffff080279825d78 by task qemu-system-aar/22429
> >   CPU: 24 PID: 22429 Comm: qemu-system-aar Not tainted 5.18.0-rc5-next-20220502 #69
> >   Call trace:
> >    dump_backtrace
> >    show_stack
> >    dump_stack_lvl
> >    print_address_description.constprop.0
> >    print_report
> >    kasan_report
> >    __asan_report_load8_noabort
> >    __lock_acquire
> >    lock_acquire.part.0
> >    lock_acquire
> >    _raw_spin_lock_irqsave
> >    arm_smmu_detach_dev
> >    arm_smmu_detach_dev at drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c:2377
> >    arm_smmu_attach_dev
>
> Hum.
>
> So what has happened is that VFIO does this sequence:
>
>   iommu_detach_group()
>   iommu_domain_free()
>   iommu_group_release_dma_owner()
>
> Which, I think should be valid, API wise.
>
>  From what I can see reading the code SMMUv3 blows up above because it
> doesn't have a detach_dev op:
>
> 	.default_domain_ops = &(const struct iommu_domain_ops) {
> 		.attach_dev		= arm_smmu_attach_dev,
> 		.map_pages		= arm_smmu_map_pages,
> 		.unmap_pages		= arm_smmu_unmap_pages,
> 		.flush_iotlb_all	= arm_smmu_flush_iotlb_all,
> 		.iotlb_sync		= arm_smmu_iotlb_sync,
> 		.iova_to_phys		= arm_smmu_iova_to_phys,
> 		.enable_nesting		= arm_smmu_enable_nesting,
> 		.free			= arm_smmu_domain_free,
> 	}
>
> But it is internally tracking the domain inside the master - so when
> the next domain is attached it does this:
>
> static void arm_smmu_detach_dev(struct arm_smmu_master *master)
> {
> 	struct arm_smmu_domain *smmu_domain = master->domain;
>
> 	spin_lock_irqsave(&smmu_domain->devices_lock, flags);
>
> And explodes as the domain has been freed but master->domain was not
> NULL'd.
>
> It worked before because iommu_detach_group() used to attach the
> default group and that was before the domain was freed in the above
> sequence.

Oof, I totally overlooked the significance of that little subtlety in 
review :(

> I'm guessing SMMU3 needs to call it's arm_smmu_detach_dev(master) from
> the detach_dev op and null it's cached copy of the domain, but I don't
> know this driver.. Robin?

The original intent was that .detach_dev is deprecated in favour of 
default domains, and when the latter are in use, a device is always 
attached *somewhere* once probed (i.e. group->domain is never NULL). At 
face value, the neatest fix IMO would probably be for SMMUv3's 
.domain_free to handle smmu_domain->devices being non-empty and detach 
them at that point. However that wouldn't be viable for virtio-iommu or 
anyone else keeping an internal one-way association of devices to their 
current domains.

If we're giving up entirely on that notion of .detach_dev going away 
then all default-domain-supporting drivers probably want checking to 
make sure that path hasn't bitrotted; both Arm SMMU drivers had it 
proactively removed 6 years ago; virtio-iommu never had it at all; newer 
drivers like apple-dart have some code there, but it won't have ever run 
until now.

We *could* stay true to the original paradigm by introducing some real 
usage of IOMMU_DOMAIN_BLOCKED, such that we could keep one or more of 
those around to actively attach to instead of having groups in this 
unattached limbo state, but that's a bigger job involving adding support 
to drivers as well; too much for a quick fix now...

Robin.