shadow mode survived with the domU running lmbench, ltp, kernbench, etc.
by running a script that enables shadow LOGDIRTY mode, CLEAN, and
disables at 5-second intervals in an infinite loop. Thanks Ian for
providing the Python script. Tested x86 domU and VMX domains as well.
Big ones are all new, and are used for 64-bit only. Please apply.
We also verified that shadow_64.c and shadow_public.c could be built for
x86 and that they worked fine there. We can provide a small patch that
does it (once the code is in ;-).
Signed-off-by: Jun Nakajima <jun.nakajima@intel.com>
Signed-off-by: Chengyuan Li <chengyuan.li@intel.com>
Signed-off-by: Yunhong Jiang <yunhong.jiang@intel.com>
Although we don't normally expect a guest to use cmpl against a MMIO region,
this may happen due to bugs. This workaround is needed for 64 bit linux-2.6.
Signed-off-by: Chengyuan Li <chengyuan.li@intel.com>
Signed-off-by: Arun Sharma <arun.sharma@intel.com>
minute, we just pick one vcpu out of the allowed set and allows route
the irq to that one; that's enough for the userspace irq balancer, but
not enough for the kernel-space one.
Whether it's actually worth implementing the full variant is open to
debate.
This also makes IRQ routing across vcpu hotplug events slightly
easier.
Signed-off-by: Steven Smith, sos22@cam.ac.uk
operation allows retrieval of the domain info structures for all domains
in one hypercall.
Using a small test program, on a system with 97 domains with
non-contiguous domain IDs, I found that with this hypercall I could
retrieve the full array of info structures 12840 times per second, an
improvement from 2380 times per second with the DOM0_GETDOMAININFO op.
The patch also adds a libxc wrapper for this new hypercall.
Signed-off-by: Josh Triplett <josht@us.ibm.com>
all bound to vcpu0, rather vcpu0 99% of the time and an entirely
random otherwise. This make sure things work nicely when you start
taking vcpus offline, since vcpu0 can't go offline; the correct
solution is to allow domains to change evtchn<->vcpu bindings.
Signed-off-by: Steven Smith, sos22@cl.cam.ac.uk
which gets rid of the -1 -> # hack and works for cpus > 9.
I ran into some issues with stale vcpu_to_cpu lists when running the
hotplug subprogram. I would take a vcpu offline, and then issue the
command to bring it back and the vcpu_to_cpu list would not have changed
to indicate the the vcpu actually went down. If I injected a xm list -v
(which always showed the correct mapping) then subsequent hotplug
commands would see the state change and fire off the hotplug request. I
don't know that not sending the event when not changing state saves that
much work so I took the state check out and now just send the hotplug
event directly.
> Also the whole hotplug stuff is still missing interrupt re-routing
> when a vcpu is taken down. To do this, we need an evtchn operation to
> change the vcpu affinity of a port by changing notify_vcpu_id.
I don't fully understand all of the mappings that are happening, so this
part of the patch might be way off. In any case, I've added a new
evtchn op to set the notify_vcpu_id field of a channel. I updated the
HOTPLUG_CPU code to use the new routines when bringing cpus up and down.
When taking down a cpu, I route the IPI irq channels to CPU 0, and when
the cpu comes up, it re-routes the channels back to the awakened CPU.
From: Ryan Harper <ryanh@us.ibm.com>
Signed-off-by: ian@xensource.com
