Tim Deegan [Tue, 31 Jul 2007 10:37:27 +0000 (11:37 +0100)]
[HVM] Inject #PF when mmio instruction fetch fails
instead of crashing the guest. This can happen if one vcpu pages out
another vcpu's kernel text page while the other is performing an mmio op. Signed-off-by: Tim Deegan <Tim.Deegan@xensource.com>
- don't reload the policy if it has been loaded
- don't always load the policy in the test suite when the policy is
already loaded
- skip tests 07 and 09 when ACM is not enabled and xm is not using the
Xen-API
- fix a problem when trying to remove an invalid label
On some ia64 NUMA machine, we cannot boot dom0.
This issue is caused by different infomation LSAPIC and SRAT.
Xen-ia64 modify LSAPIC IDs of dom0, but it does not modify SRAT.
So we decide disabling SRAT, SLIT of dom0 as first step of NUMA
work.
Xend tries to rebuild the network configuration such as bridges and
vlan interfaces (in those bridges) when starting. Unfortunately this
fails on the first startup since the network script has not run, yet,
and for example created peth0. When restarting xend it works due to
the peth0 being there then. This patch moves the initialization of the
network to an earlier time in the xend initialization process to
mitigate this problem.
Recent changes to the setup of the peth0 interface have resulted in a
change of its MAC address. Previously it seems to have had a MAC
address of 'fe:ff:ff:ff:ff:ff', but now it has the same MAC address as
eth0. As a consequence to this the Xen-API code to create VLANs
(PIF.create_VLAN("peth0",...)) does not work anymore, since peth0 can
not be identified according to this previous criteria of fake MAC
address. This patch fixes this issue by identifying it through the
prefix 'peth'.
A domain is only authorized to run if it has a superset of Simple Type
Enforcement Types in its VM label compared to that of Domain-0, which
itself may not have all STEs available in a policy. This patch adds a
check for this into Xend and the necessary code support into Xen.
xenstored: Fix socket-based connection teardown. Reading zero bytes
*should* in fact cause the connection to be destroyed. Fix this with a
little extra code in the readfd() handler. Signed-off-by: Keir Fraser <keir@xensource.com>
xenstore: Small cleanups and fixes.
1. readfd/writefd account for EINTR/EAGAIN errno returns.
2. Handle zero return from ->read() and ->write() handlers
symmetrically.
3. Fix some indentation issues (use hard tabs). Signed-off-by: Keir Fraser <keir@xensource.com>
Add domain name check and UUID check to 'xm new' command.
Add a domain name check and a UUID check to xm new command. The check
logic is as follows:
- If the UUID is not specified
- If a VM with same name exists
=> Update the config for that existing VM
- Else no vm with same name exists
=> Define a brand new VM with auto-generated UUID
- Else UUID is specified
- If a VM with same UUID exists
- If name is different
=> Error
- Else if name is same
=> Update the config for that existing VM
- Else no VM with same UUID exists
- If name is different
=> Define a branch new VM with that name
- Else if name is same
=> Error
Tim Deegan [Tue, 24 Jul 2007 14:47:36 +0000 (15:47 +0100)]
[HVM] Live-migration fixups after 15639:c585f993385c
Use the same xenstore area for log-dirty commands as for save/continue,
and avoid a race condition by acking the save command after the save. Signed-off-by: Tim Deegan <Tim.Deegan@xensource.com>
Tim Deegan [Tue, 24 Jul 2007 13:52:16 +0000 (14:52 +0100)]
[HVM] Control qemu's state-save via xenstore, instead of SIGUSR1
This lets us verify that qemu has indeed stopped processing before
we start saving guest memory. Also allow qemu to continue processing
after the save has happened, instead of exiting immediately. Signed-off-by: Tim Deegan <Tim.Deegan@xensource.com>
xenstored: Fairly round-robin schedule work across all connections.
Avoids total starvation under some workloads. Signed-off-by: Keir Fraser <keir@xensource.com>
xenstored: Guarantee to fire @releaseDomain watch when a domain is destructed.
Previously this would be missed on some bail paths within xenstored
which would talloc_free() the connection. Signed-off-by: Keir Fraser <keir@xensource.com>
Tim Deegan [Tue, 24 Jul 2007 10:10:08 +0000 (11:10 +0100)]
[HVM] Shadow: release shadow lock during emulation path
and retake it only for the write-back at the end. Signed-off-by: Tim Deegan <Tim.Deegan@xensource.com>
There is a problem in the input of the key in the VNC connection on
the PV domain. When client's keyboard is not the same as the kind of
the keyboard of PVFB and GuestOS, it is not possible to input it correctly.
This patch handled the state of shift from the set keymap. When
client's keyboard is not same as the kind of PVFB/GuestOS, it is
possible to input it correctly. It was confirmed to input it correctly
mutually with this patch between en-us and ja.
There is a problem in the input of the key in the VNC connection on
the HVM domain. When client's keyboard is not the same as the kind of
the keyboard of qemu-dm and GuestOS, it is not possible to input it
correctly.
VNC client qemu-dm & GuestOS
--------------+-----------------------
ja en-us ==> NG
en-us en-us ==> OK
Originally, the same keysym-code between client and qemu-dm is
transmitted. However, even if it is the same character, the state of
shift is different according to the kind of keyboard.
Therefore, it is necessary to handle the state of the shift by setting
qemu-dm and GuestOS. There is information on whether shift is
necessary for each key for the keymap of qemu-dm.
This patch handled the state of shift from the set keymap.
When client's keyboard is not same as the kind of qemu-dm/GuestOS,
it is possible to input it correctly.
It was confirmed to input it correctly mutually with this patch
between en-us and ja.
Fix hypercall migration. schedule_tail() is not required
after set affinity, or else continue stub runs on current
cpu immediately before migrating is checked when switching
out.
Signed-off-by Kevin Tian <kevin.tian@intel.com> Signed-off-by: Keir Fraser <keir@xensource.com>
[xend / libxen] Add support for labeling of virtual network interfaces.
This patch adds labeling of virtual network interfaces to xend and
makes this manageable through the Xen-API. It's a feature that is
only usable if ACM is enabled in Xen and xend is used through the
xen-api. A labeled virtual network interface will be plugged into a
bridge where other domains with the same-labeled network interface are
connected to, so that only same-colored domains can communicate with
each other. The bridge should be connected to the outside world using
VLAN for isolation, extending the isolation beyond the local machine.
If a virtual machine is labeled with a VM label that only has one
Simple Type Enforcement Type then it is not necessary to label the
virtual network interface, but the color of the network interface is
determined from the VM's label. If, however, a virtual machine is
labeled with a VM label that has multiple Simple Type Enforcement
Types, then the explicit labeling of each virtual network interface is
required. To specify the label of a network interface, the vif line in
the VM's configuration file has been extended with parameters similar
use for specifying the label of the VM:
Ian Campbell [Thu, 19 Jul 2007 09:26:00 +0000 (10:26 +0100)]
Build updates for building upstream trees.
* Merge the guts of the various upstream branch buildconfigs into a
single mk.linux-2.6 which they include.
* Add support for the -git snapshot branch.
* Update .hgignore to ignore all these new trees.
* Update buildconfigs/enable-xen-config to make the upstream defconfig
compile without questions.
Signed-off-by: Ian Campbell <ian.campbell@xensource.com>
[xend][xen-api] Save new memory values in domain config when calling
VM.set_memory_[dynamic|static]_[min|max] via Xen API. Signed-off-by: Jim Fehlig <jfehlig@novell.com>
Tim Deegan [Wed, 18 Jul 2007 08:06:41 +0000 (09:06 +0100)]
[HVM] Hide contents of IA32_MCG_STATUS MSR from the guest.
It can't clear the MCE, and it probably can't understand it anyway. Signed-off-by: Tim Deegan <Tim.Deegan@xensource.com>
[VTPM] Make vTPM hotplug scripts use the UUID of the vTPM instance
Make the vTPM hotplug script use the UUID of the vTPM instance for
associating it with its instance number rather than the name of the
domain.
Also change the cleanup in the xm test to use the atexit mechanism
rather than explicitly calling a function to clean up.
Tim Deegan [Mon, 16 Jul 2007 13:27:12 +0000 (14:27 +0100)]
[HVM] Qemu rtl8139: correct ring-buffer wrapping.
The rx DMA should never overrun when it hits the end of a 64k buffer Signed-off-by: Tim Deegan <Tim.Deegan@xensource.com>
Tim Deegan [Mon, 16 Jul 2007 09:36:52 +0000 (10:36 +0100)]
[XEN] HAP: need to intercept CR0 even when using NPT
because we tinker with CR0.TS. Thanks to Wei Huang for pointing this out. Signed-off-by: Tim Deegan <Tim.Deegan@xensource.com>
[POWERPC][XEN] Enable in-guest performance monitoring.
- Lazily save and restore the performance monitor counters when switching
domains.
- Control with the H_PERFMON PAPR hypercall.
- Ignore guest perfmon exceptions that land in Xen. Signed-off-by: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Signed-off-by: Hollis Blanchard <hollisb@us.ibm.com>
[XEN][POWERPC] Fix build after NUMA nodemask changes. Signed-off-by: Jerone Young <jyoung5@us.ibm.com> Signed-off-by: Hollis Blanchard <hollisb@us.ibm.com>
This patch implements DOMCTL_SENDTRIGGER_NMI hypercall for x86 hvm
domains. We can manually inject NMI into a domain by using
'xm trigger [domain] nmi [vcpu]'.
Do not clobber AMD TSC offset on real-mode switch. Signed-off-by: David Lively <dlively@virtualiron.com> Signed-off-by: Ben Guthro <bguthro@virtualiron.com>
[Xend] Fix problem when destroying VMs using the Xen-API
When domains are destroyed/suspend-resume/migration using the Xen-API
things break due to the domid not having been reset. This patch fixes
this. This is the best place I found for fixing this problem. I could
not push this line into _stateSet() for the case of DOM_STATE_HALTED
and left a comment regarding this.
- Allow multiple ChineseWallTypes in a VM labels
- check for surfacing exceptions in the python code
- check for array sizes in the XML DOM in python
- properly display the labels when doing 'xm list --label' in xm's
non-Xen-API mode
- rely on the security checking hooks in xend to check access to the
block interface rather than doing this in xm.
This patch fixes the xc_map_grant_refs() function, which would fail
because, instead of passing the malloc()-ed ioctl argument pointer to
the ioctl, the address of this pointer was being passed instead.
[Docs] Documentation of extension of the Xen-API for managing security policies
Provides an extension to the Xen-API documentation with the classes
and methods implemented for supporting of managing security policies
using the Xen-API.
Need to ensure all the code slice in the wakeup path still
existing. For this purpose, we have to use __devinit instead
of __init, since the former is null for CONFIG_HOTPLUG while
the latter always causes related code to be free-ed after
first boot.
Later when adding __init to some function, be sure to consider
wakeup case and cpu hotplug!
Provide cpu hotplug support to Xen. Note this hotplug
support is specific to PM, instead of for a run-time
single CPU hotplug which can be a separate task. See
embedded comment:
/*
* XXX: One important thing missed here is to migrate vcpus
* from dead cpu to other online ones and then put whole
* system into a stop state. It assures a safe environment
* for a cpu hotplug/remove at normal running state.
*
* However for xen PM case, at this point:
* -> All other domains should be notified with PM event,
* and then in following states:
* * Suspend state, or
* * Paused state, which is a force step to all
* domains if they do nothing to suspend
* -> All vcpus of dom0 (except vcpu0) have already beem
* hot removed
* with the net effect that all other cpus only have idle vcpu
* running. In this special case, we can avoid vcpu migration
* then and system can be considered in a stop state.
*
* So current cpu hotplug is a special version for PM specific
* usage, and need more effort later for full cpu hotplug.
* (ktian1)
*/
Signed-off-by: Kevin Tian <kevin.tian@intel.com> Signed-off-by: Keir Fraser <keir@xensource.com>
Basic infrastructure for Xen S3 support with a common CPU
context save/restore logic for both 32bit and 64bit.
Wakeup code is split into two parts:
- the first locates after trampoline code, to share all the
tricks on the latter, like relocation base and identy mapping
- the 2nd part locates in xen code segment, to do the actual
CPU context restore
Signed-off-by Ke Yu <ke.yu@intel.com>
Signed-off-by Kevin Tian <kevin.tian@intel.com>
Once the presence of LPT1 is baked into a Windows guest, it gets upset
if it is subsequently removed. Hence we add the 'LTP1' device object
back into the ACPI DSDT, and remove IRQ 7 from the list of IRQs usable
by the PCI-ISA links.
Add new domctl hypercall to expose current heap values. This
functionality is needed for probing how much memory is available in a
given node prior to VM creation.
Signed-off-by: Ryan Harper <ryanh@us.ibm.com> Signed-off-by: Keir Fraser <keir@xensource.com>
Replaces an instance of pthread_mutex_lock() that should be
pthreaad_mutex_unlock(). With thanks to our OpenTC partners at the
Technical University of Sofia, whose static analysis found this bug.
Signed-off-by: Derek G. Murray <Derek.Murray@cl.cam.ac.uk>
[xenstore] adds a -w (wide) flag to xenstore-ls to support seeing full
contents of xenstore entries. There is a bit of code cleanup as well
(snprintf vs. sprintf), one formatting fix, and comments).
There is no change to the behavior of xenstore-ls without -w.
Signed-off-by: John Zulauf <john.zulauf@intel.com>
[XM] Tools support for extensions of the Xen-API for managing security policies
This patch adds a couple of new commands for using the Xen-API
extensions for security policies. Older tools are converted to support
going through the Xen-API for their operations rather than doing the
operations directly in their own code.
Make QEMU consistently report write caching support for emulated IDE
drives to fix a hang during SLES 9 HVM guest installation.
Without this, the SLES 9 installer kernels (32 and 64 bit) were
getting inconsistent information from QEMU as to whether the
(emulated) IDE drives support write caching (which they do). So part
of the kernel thought write caching was enabled (and enabled the usage
of barrier writes) and part of it didn't, which triggered a bug in
which the same barrier write is submitted over and over again ...
Fixed by setting another bit in the WIN_IDENTIFY (IDE drive "identify"
command) response to indicate we really, truly support write caching.
Signed-off-by: David Lively <dlively@virtualiron.com> Signed-off-by: Ben Guthro <bguthro@virtualiron.com>
Alex Williamson [Tue, 10 Jul 2007 17:15:54 +0000 (11:15 -0600)]
[IA64] Fixup physinfo
Use max cpus per node to guess at sockets per node. This avoids
averaging problems with offline cpus and nodes without cpus. Also
fill in the cpu_to_node array.
Signed-off-by: Alex Williamson <alex.williamson@hp.com>
xenoprof: Fix initialisation. Much can be done at boot-time, as
indicated by the fact that the low-level functiosn are marked
'__init'. Signed-off-by: Keir Fraser <keir@xensource.com>
projects / xen.git / log