Keir Fraser [Mon, 2 Nov 2009 09:38:34 +0000 (09:38 +0000)]
Fixes after addition of dummy_vcpu_info.
- Clean initialisation of new vcpu_info in map_vcpu_info() if the
vcpu was previously using the shared dummy structure.
- Don't allow a vcpu to run with teh shared dummy info structure, as
no good can come of it.
Keir Fraser [Thu, 29 Oct 2009 14:48:28 +0000 (14:48 +0000)]
Extend the max vcpu number for HVM guest.
- Originally the max vcpu number for HVM guest is 32, this patch
extend the number to 128 on x86_64 hypervisor. (For i386 hypervisor,
the max vcpu number is still 32).
- This patch extends the mp-table size to fit more vcpus.
- HVM PV driver should call VCPUOP_register_vcpu_info hypercall to
initialize the vcpu info if the vcpu number is more than 32.
Keir Fraser [Thu, 29 Oct 2009 14:04:45 +0000 (14:04 +0000)]
stubdom: make stubdom-dm exit properly
The built-in bash command wait should be able to take a pid argument
and just wait for the specified process to die, but it currently has a
bug and what actually does is waiting for the death of all the
children. For this reason the stubdom-dm script doesn't exit properly
after stubdom destruction. This patch solves the issue spawning only
one child, removing the sleep subprocess workaround that was used to
create a usable stdin for "xm console" and replacing it with a fifo.
Keir Fraser [Thu, 29 Oct 2009 14:03:56 +0000 (14:03 +0000)]
Extend max vcpu number for HVM guest
Reduce size of Xen-qemu shared ioreq structure to 32 bytes. This
has two advantages:
1. We can support up to 128 VCPUs with a single shared page
2. If/when we want to go beyond 128 VCPUs, a whole number of ioreq_t
structures will pack into a single shared page, so a multi-page
array will have no ioreq_t straddling a page boundary
Also, while modifying qemu, replace a 32-entry vcpu-indexed array
with a dynamically-allocated array.
Keir Fraser [Thu, 29 Oct 2009 08:34:51 +0000 (08:34 +0000)]
minios: xmalloc and realloc fixes
- xmalloc currently faults if xmalloc_new_page fails due to OOM
- realloc treats xmalloc_hdr.size as the size of just the data region
rather than the total size of data region + headers + padding.
From: James Pendergrass <James.Pendergrass@jhuapl.edu> Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Wed, 28 Oct 2009 17:27:09 +0000 (17:27 +0000)]
vtd: Simplify acpi_dmar_init().
No need to check force_iommu, as that is done later in common code.
Also no need to clear iommu_enabled as again this gets checked
later. Furthermore doing it here, from a non-Intel-specific callsite,
breaks other vendors' IOMMU support.
Keir Fraser [Wed, 28 Oct 2009 17:08:26 +0000 (17:08 +0000)]
AMD IOMMU: Use global interrupt remapping table by default
Using a global interrupt remapping table shared by all devices has
better compatibility with certain old BIOSes. Per-device interrupt
remapping table can still be enabled by using a new parameter
"amd-iommu-perdev-intremap".
Keir Fraser [Wed, 28 Oct 2009 10:56:39 +0000 (10:56 +0000)]
scheduler: small csched_cpu_pick() adjustments
When csched_cpu_pick() decides to move a vCPU to a different pCPU, so
far in the vast majority of cases it selected the first core/thread of
the most idle socket/core. When there are many short executing
entities, this will generally lead to them not getting evenly
distributed (since primary cores/threads will be preferred), making
the need for subsequent migration more likely. Instead, candidate
cores/threads should get treated as symmetrically as possible, and
hence this changes the selection logic to cycle through all
candidates.
Further, since csched_cpu_pick() will never move a vCPU between
threads of the same core (and since the weights calculated for
individual threads of the same core are always identical), rather than
removing just the selected pCPU from the mask that still needs looking
at, all siblings of the chosen pCPU can be removed at once without
affecting the outcome.
Keir Fraser [Wed, 28 Oct 2009 10:54:20 +0000 (10:54 +0000)]
irq cleanup
Make IRQ related data const or __read_mostly where possible/reasonable,
use platform_legacy_irq() where feasible, and remove the now unused
definition of vector_to_irq().
Keir Fraser [Tue, 27 Oct 2009 12:52:57 +0000 (12:52 +0000)]
xsm: Add support for Xen device policies
Add support for Xen ocontext records to enable device polices. The
default policy will not be changed and instructions have been added to
enable the new functionality. Examples on how to use the new policy
language have been added but commented out. The newest version of
checkpolicy (>= 2.0.20) and libsepol (>= 2.0.39) is needed in order to
compile it. Devices can be labeled and enforced using the following
new commands; pirqcon, iomemcon, ioportcon and pcidevicecon.
Signed-off-by : George Coker <gscoker@alpha.ncsc.mil>
Signed-off-by : Paul Nuzzi <pjnuzzi@tycho.ncsc.mil>
Keir Fraser [Mon, 26 Oct 2009 13:33:38 +0000 (13:33 +0000)]
x86: IRQ Migration logic enhancement.
To programme MSI's addr/vector safely, delay irq migration
operation before acking next interrupt. In this way, it should
avoid inconsistent interrupts generation due to non-atomic writing
addr and data registers about MSI.
Keir Fraser [Mon, 26 Oct 2009 13:26:43 +0000 (13:26 +0000)]
x86: Small simplification to get_page_from_l1e().
No need for separate top-level check for page owner being NULL: this
can be folded into the case that page owner is not who the caller
expected (caller will never expect NULL owner).
sh_remove_all_mappings() will walk roughly half of the shadow L1
tables for each MFN it's called with; calling it for every MFN in a
guest's framebuffer can be _very_ expensive, especially with the
shadow lock held across the whole operation. Avoid that by just
blowing away all the shadows.
Keir Fraser [Fri, 23 Oct 2009 09:15:17 +0000 (10:15 +0100)]
x86: Enable TSC_RELIABLE for AMD servers
Except for a published BIOS errata on family 11h processors,
all AMD servers that have the Invariant TSC bit set have
a reliable TSC so Xen should not write to the TSC.
Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com> Acked-by: Mark Langsdorf <mark.langsdorf@amd.com>
Keir Fraser [Fri, 23 Oct 2009 09:13:22 +0000 (10:13 +0100)]
vtd: interrupt remapping fix
Fix the error of translation from int remapping table entry(IRTE) to
MSI msg. This error may write wrong IRTE back to the VTd hardware, and
block physical interrupts.
Keir Fraser [Fri, 23 Oct 2009 09:11:52 +0000 (10:11 +0100)]
x86: Fix RevF detection in powernow.c
The PowerNow! driver does not support RevF and earlier parts.
The current code checks for RevF processors in a function that
is not called. Change the code path so that RevF processors
are detected and the driver fails registration.
Also fix cpufreq_add_cpu() to handle unsuccessful registration.
Signed-off-by: Mark Langsdorf <mark.langsdorf@amd.com>
Keir Fraser [Fri, 23 Oct 2009 09:09:37 +0000 (10:09 +0100)]
blktap2: Fix sysfs handling of blktap2
The pause and unpause paths are currently broken due to a missing
slash. I took advantage of the opportunity to remove code repetition,
repeated strings that should point to the proper constants, etc
From: Andres Lagar Cavilla <andreslc@cs.toronto.edu> Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Fri, 23 Oct 2009 09:04:03 +0000 (10:04 +0100)]
passthrough/stubdom: clean up hypercall privilege checking
This patch adds securty checks for pci passthrough related hypercalls
to enforce that the current domain owns the resources that it is about
to remap. It also adds a call to xc_assign_device to xend and removes
the PRIVILEGED_STUBDOMS flags.
Keir Fraser [Fri, 23 Oct 2009 09:02:09 +0000 (10:02 +0100)]
blktap: Fix check_sharing() in blktapctrl
check_sharing() in blktapctrl does not work.
- It accesses to xenstore by using wrong paths.
- It compares image paths including image types.
- It misjudges a return value of strcmp().
Keir Fraser [Wed, 21 Oct 2009 15:08:28 +0000 (16:08 +0100)]
stubdom: mmap on /dev/mem support
This patch adds support for mmap on /dev/mem in a stubdom; it is
secure because it only works for memory areas that have been
explicitly allowed by the toolstack (xc_domain_iomem_permission).
Incidentally this is all that is needed to make MSI-X passthrough work
with stubdoms.
Keir Fraser [Wed, 21 Oct 2009 15:07:37 +0000 (16:07 +0100)]
x86: Initialize the affinity field after assigning the vector.
To avoid strange output from debug-key "i", desc->affinity should
be the subset of the cfg->domain basically, so copy cfg->domain to
desc->affinity after assigning vector for the irq..
Keir Fraser [Wed, 21 Oct 2009 08:23:10 +0000 (09:23 +0100)]
xend: bootable flag of VBD not always of type int
1. Calling VDB.set_bootable(True) results in string 'True' in managed
config file. After xend restart, conversion int(bootable) in
server/blkif.py fails.
2. selection of bootable disks in XendDomainInfo.py requires
type(bootable) == int not str, otherwise all disks are taken as
bootable.
This patch converts the bootable flag always to int.
Keir Fraser [Wed, 21 Oct 2009 08:21:01 +0000 (09:21 +0100)]
xmalloc_tlsf: Fall back to xmalloc_whole_pages() if xmem_pool_alloc() fails.
This was happening for xmalloc request sizes between 3921 and 3951
bytes. The reason being that xmem_pool_alloc() may add extra padding
to the requested size, making the total block size greater than a
page.
Rather than add yet more smarts about TLSF to _xmalloc(), we just
dumbly attempt any request smaller than a page via xmem_pool_alloc()
first, then fall back on xmalloc_whole_pages() if this fails.
Based on bug diagnosis and initial patch by John Byrne <john.l.byrne@hp.com>
Keir Fraser [Wed, 21 Oct 2009 07:51:10 +0000 (08:51 +0100)]
stubdom: implement pci coldplug
This patch fixes the circular dependency problem in the toolstack that
prevented pci coldplug from working with stubdoms: after creating the
stubdom we wait for it to be properly initialized before going
further. We release the domain lock while we wait.
Keir Fraser [Wed, 21 Oct 2009 07:50:23 +0000 (08:50 +0100)]
x86: MSI: Mask/unmask msi irq during the window which programs msi.
When program msi, it has to mask it first, otherwise, it
may generate inconsistent interrupts. According to spec,
if not masked, the interrupt generation behaviour is undefined.
Keir Fraser [Tue, 20 Oct 2009 07:45:12 +0000 (08:45 +0100)]
Add nomigrate config option to disable migration/restore
The new nomigrate option can be set to non-zero in vm.cfg
(for both hvm and pvm) to disallow a guest from being
migrated or restored. (Save is still allowed for the purpose
of checkpointing.) The option persists into a save file
and is also communicated into the hypervisor, the latter
for the purposes of a to-be-added hypercall for communicating
to guests that migration is disallowed (which will be
used initially for userland TSC-related sensing, but may
find other uses).
Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com>
Keir Fraser [Mon, 19 Oct 2009 10:58:36 +0000 (11:58 +0100)]
Allow guests to register secondary vcpu_time_info
Allow a guest to register a second location for the VCPU time info
structure for each vcpu. This is intended to allow the guest kernel
to map this information into a usermode accessible page, so that
usermode can efficiently calculate system time from the TSC without
having to make a syscall.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com> Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Mon, 19 Oct 2009 09:56:58 +0000 (10:56 +0100)]
x86 shadow: Update cr3 in PAE mode when guest walk succeed but shadow walk fails
When running in PAE mode, Windows 7 (apparently) will occasionally
switch cr3 with one of the L3 entries invalid, make it valid, and then
expect the hardware to load the new value. (This behavior is
explicitly not promised in the hardware manuals.) This leads to a
situation where on a shadow fault, the guest walk succeeds but the
shadow walk fails. The code assumes this can only happen when the
domain is dying, and makes an ASSERT() to that effect. So currently,
in debug mode, this will cause the host to crash; in non-debug mode,
this will cause a page-fault loop.
This patch solves the problem by calling update_cr3() in that path
when the guest is in PAE mode, and only ASSERT()ing when the guest is
not in PAE mode. The guest will get one spurious page fault, but
subsequent accesses will succeed.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
Keir Fraser [Mon, 19 Oct 2009 09:54:35 +0000 (10:54 +0100)]
[IOMMU] clean interrupt remapping and queued invalidation
This patch enlarges interrupt remapping table to fix the out-of range
table access when using many multiple-function PCI devices.
Invalidation queue is also expanded.
Signed-Off-By: Zhai Edwin <edwin.zhai@intel.com> Signed-Off-By: Cui Dexuan <dexuan.cui@intel.com>
Keir Fraser [Mon, 19 Oct 2009 09:50:46 +0000 (10:50 +0100)]
x86: vMSI: Fix msi irq affinity issue for hvm guest.
There is a race between guest setting new vector and doing EOI on old
vector. Once guest sets new vector before its doing EOI on vector,
when guest does eoi, hypervisor may fail to find the related pirq, and
hypervisor may miss to EOI real vector and leads to system hang. We
may need to add a timer for each pirq interrupt source to avoid host
hang, but this is another topic, and will be addressed later.
Keir Fraser [Fri, 16 Oct 2009 07:31:39 +0000 (08:31 +0100)]
x86: Update powernow.c to latest cpufreq code
The general cpufreq infrastructure has been improved over the
last year. Update the AMD PowerNow! driver powernow.c to
take advantage of those improvements.
Specifically, addresses Novell bugzilla # 530035.
Signed-of-by: Mark Langsdorf <mark.langsdorf@amd.com>
Keir Fraser [Fri, 16 Oct 2009 07:30:13 +0000 (08:30 +0100)]
xend: passthrough: do not check non-page-aligned MMIO BAR if not strict-check
When the option pci-passthrough-strict-check of
/etc/xen/xend-config.sxp is set to 'no', we don't check the
non-page-aligned MMIO BAR. This could be useful in some cases, e.g.,
when there is only 1 device in the range of the page and we try to
assign the device to pv guest.
Keir Fraser [Fri, 16 Oct 2009 07:28:47 +0000 (08:28 +0100)]
vt-d: Fixpanic in msi_msg_read_remap_rte with acpi=off
Xen panics when "acpi=off noacpi" is set. Problem is caused by
dereferencing NULL pointer in drhd after calling
acpi_find_matched_drhd_unit. As acpi_find_matched_drhd_unit can
return NULL, checks has to be done before returned value is used.
From: Miroslav Rezanina <mrezanin@redhat.com> Signed-off-by: Keir Fraser <keir.fraser@eu.citrix.com>
Keir Fraser [Fri, 16 Oct 2009 07:25:17 +0000 (08:25 +0100)]
x86: Fix ept and vt-d co-existence issue.
For vt-d's mmio address ranges, once ept enables, they should
be added to ept page tables with p2m lock held, and then guest can
access these ranges like conventional ram, but to change the ept
entries, it should take the p2m lock first.
Keir Fraser [Fri, 16 Oct 2009 07:24:47 +0000 (08:24 +0100)]
xend: add a description config item for each guest.
Add a new option "description=" to each VM to increase the
manageability of VM, which could be accessed via "xm list -l
MACHINE".e.g add "description='(name, james),(priority 5), (owner
james.song@company.com)'" to configure file, User can get the VM's
attribute easily by "xm list -l Machine" or some tools.
Keir Fraser [Thu, 15 Oct 2009 07:31:08 +0000 (08:31 +0100)]
xm: Reuse VDI if location is same
In XenAPI mode, when we start a VM by using xm create command, VDIs
are always automatically created for VBDs. Once we shut down the
VM, then start the VM again, then VDIs are newly created. As a
result, a vdi.xml file continues to expand.
This patch reuses VDIs if location of the VDIs is same.
Keir Fraser [Thu, 15 Oct 2009 07:16:42 +0000 (08:16 +0100)]
xend: destroy stubdoms synchronously
This patch makes the destruction of stubdoms a synchronous event,
therefore it is no longer possible to run out of memory when rebooting
a guest because the stubdom of the old guest is always destroyed
before the creation of the new guest.
Keir Fraser [Wed, 14 Oct 2009 08:09:23 +0000 (09:09 +0100)]
xm,xend: Add new option "change_home_server" to xm migrate
This patch adds a new option to xm migrate command. A concept of the
option is inspired from XenServer/XenCenter. The concept is "Change
home server (affinity)." The option name is "change_home_server."
Currently, a config.sxp file of a managed domain is not migrated to a
destination server even if the migration of the managed domain
succeeds. The config.sxp file is kept in a source server.
By the patch, the config.sxp file is migrated with the managed domain.
The config.sxp file is unregistered from the source server, then the
config.sxp file is registered to the destination server.
BTW, should the config.sxp file be always migrated without the option?
If the managed domain is migrated without the option, the managed
domains exist on both the source server and the destination server.
(Of course, the managed domain on the source server is "halted" state,
and the managed domain on the destination server is "running" state.)
Is it good that the managed domains with a same UUID exist on both
servers? (In the patch, I added the option for compatible.)
Keir Fraser [Wed, 14 Oct 2009 07:55:43 +0000 (08:55 +0100)]
minios: fix minios console
MiniOS' console_print tries to expand '\n' into "\n\r" in place,
causing page faults if the string resides in text.
Use a duplicate of the string instead.
Keir Fraser [Wed, 14 Oct 2009 07:31:07 +0000 (08:31 +0100)]
minios pcifront: translate physical into virtual addresses
Qemu understands physical pci addresses while pciback expects virtual
pci addresses: this patch adds a translation function in pcifront to
make the conversion.
Keir Fraser [Wed, 14 Oct 2009 07:29:56 +0000 (08:29 +0100)]
pv-on-hvm: Adjust mkbuildtree to handle pv_ops header placement
Due to the movement of the arch include directories, we need to adjust
where mkbuildtree looks for headers when building the pv drivers.
Also add a check for the location of features.c
Signed-off-by: Charles Arnold <carnold@novell.com>
Keir Fraser [Fri, 9 Oct 2009 07:56:43 +0000 (08:56 +0100)]
xend: Fix VDI-VBD link for XenAPI
I detected problems of VDI-VBD link again.
- In the case of inactive managed domains, VDI->VBD link was lost
by xend restarting (or host OS rebooting).
- In the case of active domains, both VDI->VBD link and VBD->VDI
link were lost by xend restarting.
When xend is restarted, information of VDI instances is restored from
a vdi.xml file. But the vdi.xml file does not have UUID of VBD
because xend does not write the UUID to the vdi.xml file. Therefore,
VDI->VBD link is lost. When xend is restarted, information of VBD
instances is restored from xenstore. But xenstore does not have UUID
of VDI. Therefore, VBD->VDI link is lost.
This patch solves the problems. VDI instances stop having UUID of
VBD. Instead, xend gathers UUID of VBD each time it's required. The
method is the same as Network->VIF link. Information of VBD instances
is restored not only from xenstore but from a config.sxp file. UUID
of VDI is restored from the config.sxp file.
FYI, VBD->VDI link of inactive managed domains is not lost because
information of VBD instances is restored from the config.sxp file.
UUID of VDI is written by xend to the config.sxp file.
projects / xen.git / log