On Mon, Jul 10, 2023 at 11:02:33PM +0800, Wen Yang wrote:
>
> On 2023/7/10 22:12, Christian Brauner wrote:
> > On Sun, Jul 09, 2023 at 02:54:51PM +0800, wenyang.linux@xxxxxxxxxxx wrote:
> > > From: Wen Yang <wenyang.linux@xxxxxxxxxxx>
> > >
> > > For eventfd with flag EFD_SEMAPHORE, when its ctx->count is 0, calling
> > > eventfd_ctx_do_read will cause ctx->count to overflow to ULLONG_MAX.
> > >
> > > Fixes: cb289d6244a3 ("eventfd - allow atomic read and waitqueue remove")
> > > Signed-off-by: Wen Yang <wenyang.linux@xxxxxxxxxxx>
> > > Cc: Alexander Viro <viro@xxxxxxxxxxxxxxxxxx>
> > > Cc: Jens Axboe <axboe@xxxxxxxxx>
> > > Cc: Christian Brauner <brauner@xxxxxxxxxx>
> > > Cc: Christoph Hellwig <hch@xxxxxx>
> > > Cc: Dylan Yudaken <dylany@xxxxxx>
> > > Cc: David Woodhouse <dwmw@xxxxxxxxxxxx>
> > > Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
> > > Cc: linux-fsdevel@xxxxxxxxxxxxxxx
> > > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > > ---
> > So this looks ok but I would like to see an analysis how the overflow
> > can happen. I'm looking at the callers and it seems that once ctx->count
> > hits 0 eventfd_read() won't call eventfd_ctx_do_read() anymore. So is
> > there a caller that can call directly or indirectly
> > eventfd_ctx_do_read() on a ctx->count == 0?
> eventfd_read() ensures that ctx->count is not 0 before calling
> eventfd_ctx_do_read() and it is correct.
>
> But it is not appropriate for eventfd_ctx_remove_wait_queue() to call
> eventfd_ctx_do_read() unconditionally,
>
> as it may not only causes ctx->count to overflow, but also unnecessarily
> calls wake_up_locked_poll().
Hm, so I think you're right and an underflow can be triggered for at
least three subsystems:
(1) virt/kvm/eventfd.c
(2) drivers/vfio/virqfd.c
(3) drivers/virt/acrn/irqfd.c
where (2) and (3) are just modeled after (1). The eventfd must've been
set to EFD_SEMAPHORE and ctx->count must been or decremented zero. The
only way I can see the _underflow_ happening is if the irqfd is shutdown
through an ioctl() like KVM_IRQFD with KVM_IRQFD_FLAG_DEASSIGN raised
while ctx->count is zero:
kvm_vm_ioctl()
-> kvm_irqfd()
-> kvm_irqfd_deassign()
-> irqfd_deactivate()
-> irqfd_shutdown()
-> eventfd_ctx_remove_wait_queue(&cnt)
which would underflow @cnt and cause a spurious wakeup. Userspace would
still read one because of EFD_SEMAPHORE semantics and wouldn't notice
the underflow.
I think it's probably not that bad because afaict, this really can only
happen when (1)-(3) are shutdown. But we should still fix it ofc.
>
>
> I am sorry for just adding the following string in the patch:
> Fixes: cb289d6244a3 ("eventfd - allow atomic read and waitqueue remove")
>
>
> Looking forward to your suggestions.
What I usually look for is some callchain/analysis that explain under
what circumstance what this is fixing can happen. That makes life for
reviewers a lot easier because they don't have to dig out that work
themselves which takes time.
