both of the current callers for libxl_domain_resume are calling after
a migration has failed, one is failure to suspend on the sender and
the other is failure to start on the destination, both leading to a
resume attempt on the sender.
However in the first case, failure to suspend, there is no guarantee
that the guest has made it as far as the suspend hypercall and
therefore the fast resume method, which frobs the hypercall return to
indicate a cancelled suspend, cannot safely be used since it will
corrupt %eax/%rax.
For the second case, failure to start on destination, I don't think it
really matters if the resume is fast or slow.
Therefore always use the slow/uncooperative version of xc_domain_resume from
libxl_domain_resume.
This makes a PV domain which failed to suspend (e.g. because the core
Linux PM infrastructure within the guest didn't allow it) recover
gracefully.
Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
If we are watching multiple disks we have to free multiple waiters.
Free the event before continuing the loop.
Destroy the disk device returned by libxl_event_get_disk_eject_info.
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
Otherwise enabling log-dirty mode on a PV guest that already has
a shadow allocation can leave the alloc/free functions pointers NULL,
and later try to dereference them.
Signed-off-by: Tim Deegan <Tim.Deegan@citrix.com>
p2m internals should always gate on whether HAP is enabled for the
domain, not whether a HAP paging mode is currently advertised.
This lets us revert the change to hap_enable() that advertises the
new mode before it's safe to use it.
Signed-off-by: Tim Deegan <Tim.Deegan@citrix.com>
Special case how we validate vhd image files. Without this patch when
tap:aio:vhd prefixed image files are specified in the config file,
disk validation and thus vm creation will fail.
Signed-off-by: Kamala Narasimhan <kamala.narasimhan@citrix.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
Currently we pile all the backend and format information pertaining to
disk option in a single enum. This check-in separates the two and
uses two enums, one for disk format and another for disk backend.
This helps clearly differentiate between disk format and backend
within the implementation and also helps cleanup the code in this area
in preparation for the impending parser revamping to be done post 4.1.
Along with separating format and backend, this check-in also removes
unwanted types and renames variables in the disk interface and fixes
the code affected by the interface changes.
In specific, here are the disk interface changes made - In
libxl_device_disk structure physpath was renamed to pdev_path,
virtpath was renamed to vdev, phystype was removed and replaced with
backend and format enums. Also previously a single enum
libxl_disk_phystype held the values for qcow, qcow2, vhd, aio, file,
phy, empty and that got refactored into two enums, libxl_disk_format
to hold unknown, qcow, qcow2, vhd, raw, empty and libxl_disk_backend
to hold unknown, phy, tap and qdisk.
Signed-off-by: Kamala Narasimhan <kamala.narasimhan@citrix.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
Prior to 22909:6868f7f3ab3f libxl would loop waiting simultaneously
for the domain the acknowledge a PV suspend request (by clearing the
XenStore node) and for the domain to actually suspend. For HVM guests
without PV drivers this same loop was simply waiting for the domain to
suspend.
In 22909:6868f7f3ab3f the original loop was split into two loops
(first waiting for the acknowledgement and then for the actual
suspend). This caused libxl to incorrectly wait for an HVM guest
without PV drivers to acknowledge the XenStore request, which is not
something it would ever do.
Fix this by only waiting for an acknowledgement from a guest which
contains PV drivers.
Previously we were also making the request regardless of whether the
guest had PV drivers, change that to only make the request if the
guest has PV drivers.
Lastly there is no need to sample HVM_PARAM_ACPI_S_STATE twice and not
doing so simplifies the test for PVHVM vs. normal HVM guests.
Tested with:
Windows with GPL PV drivers (event channel suspend mode)
Windows without PV drivers (xc_domain_shutdown mode)
Linux PV (PV with XenBus control node mode)
Linux HVM (PVHVM with XenBus control node mode (*))
Linux HVM (xc_domain_shutdown mode)
(*) In this case the kernel didn't actually suspend, due to:
PM: Device input1 failed to suspend: error -22
xen suspend: dpm_suspend_start -22
which may be a misconfiguration in my setup or may be a kernel
bug, but the libxl side dealt with this as gracefully as it could.
Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
xl vcpu-set currently uses a 32 bit mask for specifying which cpus are to be
set online. This restricts the number of cpus supported by this command.
The patch switches to libxl_cpumap, the interface of libxl_set_vcpuonline()
is changed accordingly.
Signed-off-by: Juergen Gross <juergen.gross@ts.fujitsu.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
xl cpupool-create won't take always extra parameters specified on the command
line, as a 0-byte is missing at the end of the configuration file contents.
Signed-off-by: Juergen Gross <juergen.gross@ts.fujitsu.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
Jan Beulich found that for S3 resume on platforms without ARAT feature
but with MSI capable HPET, request_irq() will be called in
hpet_setup_msi_irq() for irq already setup(no release_irq() called
during S3 suspend), so that always falling back to using
legacy_hpet_event.
Fix it by conditional calling request_irq() for 4.1. Planned to split
the S3 resume path from booting path post 4.1, as Jan suggested.
Signed-off-by: Wei Gang <gang.wei@intel.com>
Acked-by: Jan Beulich <jbeulich@novell.com>
In vif-common.sh, set the shell variable "dev" to the new interface
name when interfaces are renamed, and consistently use this variable
in all the vif scripts.
This fixes hotplug of renamed interfaces.
From: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
From: Patrick Scharrenberg <pittipatti@web.de>
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
Signed-off-by: Patrick Scharrenberg <pittipatti@web.de>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
The PV suspend protocol requires guest co-operating whereby the guest
must respond to a suspend request written to the xenstore control node
by clearing the node and then making a suspend hypercall.
Currently when a guest fails to do this libxl times out and returns
a generic failure code to the caller.
In response to this failure xl attempts to resume the guest. However
if the guest has not responded to the suspend request then the is no
guarantee that the guest has made the suspend hypercall (in fact it is
quite unlikely). Since the resume process attempts to modify the
return value of the hypercall (to indicate a cancelled suspend) this
results in the guest eax/rax register being corrupted!
To fix this change libxl to do the following:
* Wait for the guest to acknowledge the suspend request.
- on timeout cancel the suspend request.
- if cancellation is successful then return a new error code to
indicate that the guest is not responding.
- if the cancel does not succeed then we raced with the guest
which actually did acknowledge at the last minute, so
continue.
* Wait for the guest to suspend.
- on timeout return the standard error code as before
* Guest successfully suspended, return success.
Lastly in xl do not attempt to resume a guest if it has not responded
to the suspend request.
Tested by live migration of PVops kernels which either ignore the
suspend request, have already crashed and those which suspend/resume
correctly. In the first two cases the source domain is left alone (and
continues to function in the first case) and in the third the
migration is successful.
Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
The PV shutdown/reboot/suspend protocol requires that the guest
acknowledge a request by clearing the node therefore it is necessary
to allow the guest to write to the node.
Currently libxl is quite relaxed about this protocol and doesn't
reeally seem to mind that the guest is unable to write the node to
perform the acknowledgement. However in a followup patch libxl needs
to be able to detect that a guest has acknowledged a suspend request.
A side effect of this change is that an empty "control/shutdown" node
is created upon domain creation instead of only being created when a
shutdown/reboot/suspend is requested. This should not (and does not
in my tests) have any negative impact on the guest.
Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
Fix double free due to libxl__file_reference_unmap(&info->kernel) called
multiple times: first at the end of libxl__domain_build and then in
libxl_domain_build_info_destroy.
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
Committed-by: Ian Jackson <ian.jackson@eu.citrix.com>
According to lzma's configure.ac (!) the minimum memory limit to cope
with arbitrary input is 128Mby (!)
This is obviously an unreasonable amount of memory for this kind of
task, but we need to increase the constant limit for it not to
randomly fail. So do so.
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
This follows Linux commit 39fe05e58c5e448601ce46e6b03900d5bf31c4b0,
noticing that all this setup is pointless when ARAT support is there,
and knowing that on SLED11's native kernel it has actually caused S3
resume issues.
A question would be whether HPET legacy interrupts should be forced
off in this case (rather than leaving whatever came from firmware).
Signed-off-by: Jan Beulich <jbeulich@novell.com>
MSRs that control physical CPU aspects generally are pointless (and
possibly dangerous) to be written when the writer isn't sufficiently
aware that it's running virtualized.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
If a vcpu is pinned to multiple physical cpus, the pinning is not
removed if all those physical cpus are removed from the cpupool. When
disabling the scheduler on a cpu, the affinity mask must be checked
against the cpumask of the cpupool.
Signed-off-by: Juergen Gross <juergen.gross@ts.fujitsu.com>
IO Page table growth is triggered by amd_iommu_map_page and grows to
upper level. I have tested it well for different devices (nic and gfx)
and different guests (linux and Win7) with different guest memory
sizes (512M, 1G, 4G and above).
Signed-off-by: Wei Wang <wei.wang2@amd.com>
Firstly, __cpupool_find_by_id() would dereference NULL, at the end of
an exact search if the search loop exited with *q==NULL. Fix this.
Secondly, provide suitable accessor functions so that no caller needs
to use the __-prefixed versions which take a boolean flag.
Signed-off-by: Keir Fraser <keir@xen.org>
Some cpupool actions didn't check the cpupool_id exactly. For some
actions this doesn't make any sense, so refuse those actions if the
specified cpupool doesn't exist.
Signed-off-by: Juergen Gross <juergen.gross@ts.fujitsu.com>
Stefano Stabellini