Al Viro [off-list ref] writes:

On Mon, Aug 28, 2017 at 02:38:37PM +1000, Paul Mackerras wrote:

quoted

On Sun, Aug 27, 2017 at 10:02:20PM +0100, Al Viro wrote:

quoted

On Wed, Aug 23, 2017 at 04:06:24PM +1000, Paul Mackerras wrote:

quoted

It seems to me that it would be better to do the anon_inode_getfd()
call before the kvm_get_kvm() call, and go to the fail label if it
fails.

And what happens if another thread does close() on the (guessed) fd?

Chaos ensues, but mostly because we don't have proper mutual exclusion
on the modifications to the list.  I'll add a mutex_lock/unlock to
kvm_spapr_tce_release() and move the anon_inode_getfd() call inside
the mutex.

It looks like the other possible uses of the fd (mmap, and passing it
as a parameter to the KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE ioctl on a KVM
device fd) are safe.

Frankly, it's a lot saner to have "no failure points past anon_inode_getfd()"
policy...

Actually I thought that was a hard rule. But I don't see it documented
or mentioned anywhere so I'm not sure now why I thought that.

cheers