From: Wanpeng Li <wanpengli@xxxxxxxxxxx>
Tim Shearer reported that "There is a guest which is running a packet
forwarding app based on the DPDK (dpdk.org). The packet receive routine
writes to 0xc070 using glibc's "outw_p" function which does an additional
write to I/O port 80. It does this write for every packet that's received,
causing a flood of KVM userspace context switches". He uses mpstat to
observe a CPU performing L2 packet forwarding on a pinned guest vCPU,
the guest time is 95 percent when allowing I/O port 0x80 bypass, however,
it is 65.78 percent when I/O port 0x80 bypss is disabled.
This patch allows I/O port 0x80 bypass when userspace prefer.
Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
Cc: Radim Krčmář <rkrcmar@xxxxxxxxxx>
Cc: Tim Shearer <tshearer@xxxxxxxxxxxxxxx>
Cc: Liran Alon <liran.alon@xxxxxxxxxx>
Signed-off-by: Wanpeng Li <wanpengli@xxxxxxxxxxx>
---
arch/x86/kvm/vmx.c | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index ebf1140..d3e5fef 100644
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -10118,6 +10118,13 @@ static int vmx_vm_init(struct kvm *kvm)
goto out;
memset(kvm_vmx->vmx_io_bitmap[i], 0xff, PAGE_SIZE);
}
+ if (kvm->arch.ioport_disable_intercept) {
+ /*
+ * Allow direct access to the PC debug port (it is often used for I/O
+ * delays, but the vmexits simply slow things down).
+ */
+ clear_bit(0x80, kvm_vmx->vmx_io_bitmap[VMX_IO_BITMAP_A]);
+ }
return 0;
out:
--
2.7.4
