On Wed, 2013-12-04 at 11:05 +0000, Stefano Stabellini wrote:

On Wed, 4 Dec 2013, Ian Campbell wrote:

quoted

On Wed, 2013-12-04 at 10:51 +0000, Stefano Stabellini wrote:

quoted

On Wed, 4 Dec 2013, Ian Campbell wrote:

quoted

quoted

+bool xen_has_pv_devices(void)
+{
+	if (!xen_domain())
+		return false;
+
+	if (xen_hvm_domain()) {
+		/* User requested no unplug, so no PV drivers. */
+		if (xen_emul_unplug & XEN_UNPLUG_NEVER)
+			return false;

I think you need
		if (xen_emul_unpug & XEN_UNPLUG_UNNECESSARY)
			return true;
don't you?

XEN_UNPLUG_UNNECESSARY was introduced to enable the platform PCI device
even if it didn't respond properly to the unplug protocol.
The corresponding parameter is called "unnecessary" because if you pass
it to the kernel you mean that it is unnecessary to unplug the emulated
devices but you can use the pv devices anyway.

So no, we shouldn't check for XEN_UNPLUG_UNNECESSARY here.

Oh, we will eventually fall through to the return true, so it does
actually work out OK.

I'd still be in favour of handling each option explicitly, for clarity.
Which means checking for XEN_UNPLUG_UNNECESSARY.

I think is wrong to check for any xen_emul_unpug options in this function.
The xen_emul_unpug options should be used to set the right value of
xen_platform_pci_unplug. (See my other reply.)

Whichever one we check we should still be checking explicitly for the
"unnecessary" case, for clarity if nothing else.

TBH I think the split between xen_emul_unplug and
xen_platform_pci_unplug is a bit artificial. There should be one value
which is static to platform-pci-unplug.c and accessor functions should
be provided for other code to use. Open coding all those accesses to
xen_platform_pci_unplug in every driver is just too error prone.

Ian.