On Thu, Nov 17, 2022, Wei Liu wrote:
> On Wed, Nov 16, 2022 at 10:41:36AM -0800, Michael Kelley wrote:
> [...]
> >
> > +static void hv_pci_read_mmio(struct device *dev, phys_addr_t gpa, int size, u32 *val)
> > +{
> > + struct hv_mmio_read_input *in;
> > + struct hv_mmio_read_output *out;
> > + u64 ret;
> > +
> > + /*
> > + * Must be called with interrupts disabled so it is safe
> > + * to use the per-cpu input argument page. Use it for
> > + * both input and output.
> > + */
There's no need to require interrupts to be disabled to safely use a per-cpu
variable, simply disabling preemption also provides the necessary protection.
And this_cpu_ptr() will complain with CONFIG_DEBUG_PREEMPT=y if preemption isn't
disabled.
IIUC, based on the existing code, what is really be guarded against is an IRQ arriving
and initiating a different hypercall from IRQ context, and thus corrupting the page
from this function's perspective.
> Perhaps adding something along this line?
>
> WARN_ON(!irqs_disabled());
Given that every use of hyperv_pcpu_input_arg except hv_common_cpu_init() disables
IRQs, what about adding a helper to retrieve the pointer and assert that IRQs are
disabled? I.e. add the sanity for all usage, not just this one-off case.
And since CPUHP_AP_ONLINE_DYN => hv_common_cpu_init() runs after scheduling is
activated by CPUHP_AP_SCHED_WAIT_EMPTY, I believe that hv_common_cpu_init() is
theoretically broken. Maybe someone can look at that when fixing he KVM vs.
Hyper-V issue?
https://lore.kernel.org/linux-hyperv/878rkqr7ku.fsf@xxxxxxxxxxxxxxxxxxxxxxxxxxx
https://lore.kernel.org/all/87sfikmuop.fsf@xxxxxxxxxx
