Re: acpi_pcihp_eject_slot() bug if passed 'slots == 0'

[Michael S. Tsirkin]

On Thu, Mar 26, 2020 at 02:23:17PM +0100, Igor Mammedov wrote:
> On Thu, 26 Mar 2020 11:52:36 +0000
> Peter Maydell <address@hidden> wrote:
> 
> > Hi; Coverity spots that if hw/acpi/pcihp.c:acpi_pcihp_eject_slot()
> > is passed a zero 'slots' argument then ctz32(slots) will return 32,
> > and then the code that does '1U << slot' is C undefined behaviour
> > because it's an oversized shift. (This is CID 1421896.)
> > 
> > Since the pci_write() function in this file can call
> > acpi_pcihp_eject_slot() with an arbitrary value from the guest,
> > I think we need to handle 'slots == 0' safely. But what should
> > the behaviour be?
> 
> it also uncovers a bug, where we are not able to eject slot 0 on bridge,


And that is by design. A standard PCI SHPC register can support up to 31
hotpluggable slots. So we chose slot 0 as non hotpluggable.
It's consistent across SHPC, PCI-E, so I made ACPI match.
You can't hot-add it either.

> can be reproduced with:
> 
>  -enable-kvm -m 4G -device pci-bridge,chassis_nr=1 -global 
> PIIX4_PM.acpi-pci-hotplug-with-bridge-support=on -device 
> virtio-net-pci,bus=pci.1,addr=0,id=netdev12
> 
> (monitor) device_del netdev12
> (monitor) qtree # still shows the device
> 
> reason is that acpi_pcihp_eject_slot()
>    if (PCI_SLOT(dev->devfn) == slot) { # doesn't  match (0 != 32)
> 
> so device is not deleted

We should probably teach QEMU that some slots aren't hotpluggable
even if device in it is hotpluggable in theory. But that is
a separate issue.

> > thanks
> > -- PMM
> >