[kvmtool test PATCH 24/24] kvmtool: arm: Add support for creating VM with PA size

On Wed, Jul 04, 2018 at 03:41:18PM +0100, Marc Zyngier wrote:

On Wed, 04 Jul 2018 15:22:42 +0100,
Will Deacon [off-list ref] wrote:

quoted

On Fri, Jun 29, 2018 at 12:15:44PM +0100, Suzuki K Poulose wrote:

quoted

diff --git a/arm/kvm.c b/arm/kvm.c
index 5701d41..b1969be 100644
--- a/arm/kvm.c
+++ b/arm/kvm.c

@@ -11,6 +11,8 @@
 #include <linux/kvm.h>
 #include <linux/sizes.h>
 
+unsigned long kvm_arm_type;
+
 struct kvm_ext kvm_req_ext[] = {
 	{ DEFINE_KVM_EXT(KVM_CAP_IRQCHIP) },
 	{ DEFINE_KVM_EXT(KVM_CAP_ONE_REG) },

@@ -18,6 +20,26 @@ struct kvm_ext kvm_req_ext[] = {
 	{ 0, 0 },
 };
 
+#ifndef KVM_ARM_GET_MAX_VM_PHYS_SHIFT
+#define KVM_ARM_GET_MAX_VM_PHYS_SHIFT		_IO(KVMIO, 0x0b)
+#endif
+
+void kvm__arch_init_hyp(struct kvm *kvm)
+{
+	int max_ipa;
+
+	max_ipa = ioctl(kvm->sys_fd, KVM_ARM_GET_MAX_VM_PHYS_SHIFT);
+	if (max_ipa < 0)
+		max_ipa = 40;
+	if (!kvm->cfg.arch.phys_shift)
+		kvm->cfg.arch.phys_shift = 40;
+	if (kvm->cfg.arch.phys_shift > max_ipa)
+		die("Requested PA size (%u) is not supported by the host (%ubits)\n",
+		    kvm->cfg.arch.phys_shift, max_ipa);
+	if (kvm->cfg.arch.phys_shift != 40)
+		kvm_arm_type = kvm->cfg.arch.phys_shift;
+}

Seems a bit weird that the "machine type identifier" to KVM_CREATE_VM is
dedicated entirely to holding the physical address shift verbatim. Is this
really the ABI?

Also, couldn't KVM figure it out automatically if you add memslots at high
addresses, making this a niche tunable outside of testing?

Not really. Let's say I want my IPA space split in two: memory covers
the low 47 bit, and I want MMIO spanning the top 47 bit. With your
scheme, you'd end-up with a 47bit IPA space, while you really want 48
bits (MMIO space implemented by userspace isn't registered to the
kernel).

That still sounds quite niche for a VM. Does QEMU do that? In any case,
having KVM automatically increase the IPA bits to cover the memslots it
knows about would make sense to me, and also be sufficient for kvmtool
without us having to add an extra command-line argument.

The MMIO case might be better dealt with by having a way to register MMIO
regions rather than having the PA bits exposed directly.

Will