RE: [PATCH v3] vfio: fix deadlock between group lock and kvm lock

[Tian, Kevin]

> From: Alex Williamson <alex.williamson@xxxxxxxxxx>
> Sent: Friday, February 3, 2023 3:42 AM
> 
> On Thu,  2 Feb 2023 11:24:42 -0500
> Matthew Rosato <mjrosato@xxxxxxxxxxxxx> wrote:
> 
> > After 51cdc8bc120e, we have another deadlock scenario between the
> > kvm->lock and the vfio group_lock with two different codepaths acquiring
> > the locks in different order.  Specifically in vfio_open_device, vfio
> > holds the vfio group_lock when issuing device->ops->open_device but
> some
> > drivers (like vfio-ap) need to acquire kvm->lock during their open_device
> > routine;  Meanwhile, kvm_vfio_release will acquire the kvm->lock first
> > before calling vfio_file_set_kvm which will acquire the vfio group_lock.
> >
> > To resolve this, let's remove the need for the vfio group_lock from the
> > kvm_vfio_release codepath.  This is done by introducing a new spinlock to
> > protect modifications to the vfio group kvm pointer, and acquiring a kvm
> > ref from within vfio while holding this spinlock, with the reference held
> > until the last close for the device in question.
> >
> > Fixes: 51cdc8bc120e ("kvm/vfio: Fix potential deadlock on vfio group_lock")
> > Reported-by: Anthony Krowiak <akrowiak@xxxxxxxxxxxxx>
> > Suggested-by: Jason Gunthorpe <jgg@xxxxxxxxxx>
> > Signed-off-by: Matthew Rosato <mjrosato@xxxxxxxxxxxxx>
> > ---
> > Changes from v2:
> > * Relocate the new functions back to vfio_main and externalize to call
> >   from group (Kevin) since cdev will need this too
> > * s/vfio_kvm_*_kvm/vfio_device_*_kvm/ and only pass device as input.
> >   Handle new kvm_ref_lock directly inside vfio_device_get_kvm (Alex)
> > * Add assert_lockdep_held for dev_set lock (Alex)
> > * Internalize error paths for vfio_device_get_kvm_safe and now return
> >   void - either device->kvm is set with a ref taken or is NULL (Alex)
> > * Other flow suggestions to make the call path cleaner - Thanks! (Alex)
> > * Can't pass group->kvm to vfio_device_open, as it references the value
> >   outside of new lock.  Pass device->kvm to minimize changes in this
> >   fix (Alex, Yi)
> > Changes from v1:
> > * use spin_lock instead of spin_lock_irqsave (Jason)
> > * clear device->kvm_put as part of vfio_kvm_put_kvm (Yi)
> > * Re-arrange code to avoid referencing the group contents from within
> >   vfio_main (Kevin) which meant moving most of the code in this patch
> >   to group.c along with getting/dropping of the dev_set lock
> > ---
> >  drivers/vfio/group.c     | 32 ++++++++++++++----
> >  drivers/vfio/vfio.h      | 14 ++++++++
> >  drivers/vfio/vfio_main.c | 70 ++++++++++++++++++++++++++++++++++++--
> --
> >  include/linux/vfio.h     |  2 +-
> >  4 files changed, 103 insertions(+), 15 deletions(-)
> 
> LGTM.  I'm not sure moving the functions to vfio_main really buys us
> anything since we're making so much use of group fields.  The cdev
> approach will necessarily be different, so the bulk of the get code will
> likely need to move back to group.c anyway.
> 

well my last comment was based on Matthew's v2 where the get code
gets a kvm passed in instead of implicitly retrieving group ref_lock
internally. In that case the get/put helpers only contain device logic
thus fit in vfio_main.c.

with v3 then they have to be in group.c since we don't want to use
group fields in vfio_main.c.

but I still think v2 of the helpers is slightly better. The only difference
between cdev and group when handling this race is using different
ref_lock. the symbol get/put part is exactly same. So even if we
merge v3 like this, very likely Yi has to change it back to v2 style
to share the get/put helpers while just leaving the ref_lock part
handled differently between the two path.

Thanks
Kevin