Keir Fraser [Thu, 12 Jun 2008 15:09:08 +0000 (16:09 +0100)]
x86: machine_halt() forcibly re-enables IRQs as it may be called from
panic() which itself may be called in an IRQ-safe critical section. Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Thu, 12 Jun 2008 15:05:35 +0000 (16:05 +0100)]
x86/32on64: fix physical address restriction
The allocation bit size setting wasn't working anymore after the
recent fix to properly use PAGE_SHIFT instead of PAGE_SIZE. This was
because the bit size implies a power-of-two range that's accessible,
but if all memory is accessible anyway (and its upper boundary is not
a power of two), the domain would either be needlessly restricted or
wouldn't be able to allocate as much memory as was intended for it
(specifically the case for Dom0 without dom0_mem= boot
parameter). Consequently, don't restrict the bit width if all memory
can be accessed.
To avoid needing to adjust this code in two places in the future (it
may need further touching when memory hotplug gets supported), fold
the logic into a function.
Signed-off-by: Jan Beulich <jbeulich@novell.com> Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Thu, 12 Jun 2008 14:42:09 +0000 (15:42 +0100)]
[IA64] compilation fix caused by 17814:9af7a535225f
ia64 compliation fix caused by 17814:9af7a535225f.
- generate_acpi_checksum() is replaced by acpi_tb_checksum()
- move dump_execution_state() definitions from asm-ia64/xenprocessor.h
to asm-ia64/bug.h
Keir Fraser [Wed, 11 Jun 2008 14:19:42 +0000 (15:19 +0100)]
MSI: Fix msi delivery for hvm guest
Some parameters like "dest_mode", "delivery_mode" and "trig_mode"
seem to have been miscalculated before delivering MSI interrupt into
hvm passthru guest. This will cause guest driver to roll back to
legacy interrupt mode.
Keir Fraser [Wed, 11 Jun 2008 10:35:35 +0000 (11:35 +0100)]
x86: cpuid configuration for PV guest
Add pv guest support for the cpuid configuration and checking. That
feature only works for cpuid request which are coming from the guest's
kernel, a process could still call the cpuid directly.
A new policy for pv guest has been created in libxc. dom0 cpuid
emulation is left hardcoded in Xen.
Signed-off-by: Jean Guyader <jean.guyader@eu.citrix.com> Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Wed, 11 Jun 2008 08:36:23 +0000 (09:36 +0100)]
Have `make install' write the Python version number in the #!
The Xen tools contain a number of Python extensions written in C. The
C API to Python, used by these extensions, is not stable from one
version of Python to the next.
Our build system uses whatever version of Python is the default on the
build system at that time to build these extensions. However, the
actual scripts such as `xm' use #!/usr/bin/env python which uses
whichever version of Python is the default on the deployment system at
the time of invocation.
If for any reason these two versions of Python are not the same (eg,
because the system is built on one computer and executed on another,
or because the system's gains a more recent Python installation
alongside the original which changes the previous default), warnings
and failures occur.
In this patch I arrange for Python scripts to be installed via a
special wrapper around `install', which determines the build-time
Python version and path and then writes that into the #! line at the
point of `make install' (or `make dist'). (It can also be overridden
by setting PYTHON_PATH.)
Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
Keir Fraser [Tue, 10 Jun 2008 08:27:01 +0000 (09:27 +0100)]
libxenstat: Fix statistics for blktap disks on linux
Overview: update xenstat vbd statistics parsing from
/sys/devices/xen-backend to process blktap disks
Reason:
-blktap devices (now referenced as tap rather than vbd in
/sys) have statistics counters
(e.g., rd_req, wr_req, oo_req) prepended by tap_
-xenstat behavior did not previously account for this
behavior, which resulted in 0 disks
visible and similarly impacted disk statistics
To reproduce these conditions, make a domain with tap:aio backed disks
on 3.2.x, run xentop, and press B to view VBD stats (nothing will
appear for the domain using tap:aio)
Keir Fraser [Mon, 9 Jun 2008 16:18:27 +0000 (17:18 +0100)]
Allow older PAE Linux guests to access entire compat m2p.
Older PAE guests (prior to xen-unstable.hg 8924:229c602a075a, Feb
2006, this includes some vendor's kernel in the field) use a limit of
0xf6800000 on their code and data segments (i.e. up to the end of the
read-only m2p table).
Newer kernels use a limit of 4G and rely on the hypervisor to clamp to
the actual maximum allowed. 32on64 mode takes advantage of this to
allow a larger m2p than would fit in the PAEonPAE sized hole.
This means that PAE guests with the hardcoded low limit cannot run on
top of a 64 bit hypervisor on a host machine which has more than 16G
of RAM. Fix this by extending any code or data segment which ends
above the start of the hypervisor hole for that guest.
Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
Keir Fraser [Mon, 9 Jun 2008 08:51:14 +0000 (09:51 +0100)]
x86: Fix S3 resume after creating an HVM guest.
If we don't use PIT, we should disable it at boot time and at S3
resuming. In Xen this is done by disable_pit_irq(), but this function
is used by __initcall(). After bootstrap it will be discarded and may be
overwritten. This usually results in S3 resuming failure.
Signed-off-by: Huacai Chen <huacai.chen@intel.com> Signed-off-by: Li Zhang <li.zhang@intel.com>
Keir Fraser [Mon, 9 Jun 2008 08:45:38 +0000 (09:45 +0100)]
xentrace: fix tracing for 64bit guests
Xen tracing some times ago used to put values of type 'long' into the
trace buffer. This has changed to uint32_t. Some trace points log
virtual addresses, which get cropped to 32bit in this case. There were
some inline functions to handle at least PF_XEN and VMEXIT, which
caused a lot of code duplication. The attached patch fixes several
issues:
1. fix and extend tools/xentrace/formats
2. Fix xentrace_format to handle up to 7 parameters
3. create convenience macros to properly log long values
4. remove the inline functions in hvm/trace.h and replace them by macros
5. Change the CPUID trace to work correctly
6. group HVM trace points enable mechanism
I used a similar approach as in PV tracing with bit 8 indicating 64bit
pointers.
Signed-off-by: Andre Przywara <andre.przywara@amd.com>
Keir Fraser [Mon, 9 Jun 2008 08:44:21 +0000 (09:44 +0100)]
AMD IOMMU: add support for PV guest
Assuming iommu_populate_page_table() is only used for PV guest, a new
check is added into assign_device() to avoid duplicating the work done
by amd_iommu_sync_p2m() for hvm guest.
Bring the DisplayState dpy_resize interface back to how it is in qemu
mainstream, thus making the code easier to merge.
In order to support sharing the framebuffer, I am adding a new resize
interface called dpy_resize_shared that also has a depth and a pixels
parameters. As a consequence I could remove the dpy_colourdepth
callback and make the code cleaner and easier to read.
Keir Fraser [Mon, 9 Jun 2008 08:36:27 +0000 (09:36 +0100)]
xm: Set device_model option value default to None.
If device_model is set to '', xend will fail to execute the device
model. If None, xend will detect it and set a sensible qemu-dm path
instead.
Keir Fraser [Thu, 5 Jun 2008 09:36:19 +0000 (10:36 +0100)]
x86 hvm: Improve paging performance for 64b solaris guests
The following patch provides a 'fast-path' for sh_remove_write_access()
for 64 bit Solaris HVM guests. This provides a significant performance
boost for such guests; our testing shows a 200-400% improvement in
microbenchmarks such as fork(), exit(), etc...
From: Gary Pennington <Gary.Pennington@sun.com> Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Tue, 3 Jun 2008 08:41:19 +0000 (09:41 +0100)]
Improve HPET comparator reprog to prevent intr-near-missing case
HPET intr-near-missing means if the current counter value is too close
to the comparator value to be reprogrammed the expected HPET intr may
be missing. Linux kernel uses a mininal 48-hpet-ticks(~3.5us) distance
to workaround this, but personal observation showed there is still
failure case while delta=3D0xba (~13.5us). So choosing 20us as the
MIN_DELTA_NS should be helpful to prevent near-missing from happening.
local_irq_save/restore were used to avoid disturbance. (+ 2) was used
as a final guard to avoid wrong judgement due to the real happened
near-missing case.
Signed-off-by: Wei Gang <gang.wei@intel.com> Signed-off-by: Tian Kevin <kevin.tian@intel.com>
Keir Fraser [Tue, 3 Jun 2008 08:35:02 +0000 (09:35 +0100)]
x86: Enhance Cx stability by adding softirq check before entry cstate
Without checking softirq_pending before entry Cx state, softirq such
as SCHEDULE_SOFTIRQ and TIMER_SOFTIRQ can't be handled timely. It may
cause severe failures such as disk I/O failure.
This patch addresses the issue above, meanwhile move timing points
closer to cstate entry/exit point to make C3 residency more accurate,
and enable irq a little earlier.
Signed-off-by: Wei Gang <gang.wei@intel.com> Signed-off-by: Tian Kevin <kevin.tian@intel.com>
Keir Fraser [Mon, 2 Jun 2008 15:05:07 +0000 (16:05 +0100)]
rombios: S3 resume fixes.
On AMD cpus, HVM S3 resume was hanging in the bios when trying
to load the x_firmware_waking_vector from absolute
address 0xEA01C (that is, a segment offset greater than 64kB).
This change also now uses 20-bits (instead of 16) from the FACS
standard firmware_waking_vector.
Keir Fraser [Mon, 2 Jun 2008 09:55:55 +0000 (10:55 +0100)]
vmx: Better diagnostic messages when VMXON might have failed due to
bad setting of the IA32_FEATURE_CONTROL MSR. Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Fri, 30 May 2008 18:31:18 +0000 (19:31 +0100)]
x86: i8254 device model is shared between PV and HVM guests, so must
take care not to run hvm_get_guest_time() on a PV VCPU. The spinlock
is not initialised and spin_lock() can thus spin forever. Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Keir Fraser [Fri, 30 May 2008 15:04:20 +0000 (16:04 +0100)]
Add xen_phys_start value in the crash info note
This patch makes the vmcore utilities (ex. crash, makedumpfile) be
able to get the relocation address of the xen hypervisor from a
vmcore. It is necessary for the utilities to find the data of the
hypervisor structures.
Note that this patch does not raise any comptibility issue for the
utilities (which I know) nor the other components of xen.
Keir Fraser [Fri, 30 May 2008 14:09:40 +0000 (15:09 +0100)]
[IA64] compilation fix of iommu.h
Trivial compilation fix of iommu.h.
It includes asm/msi.h for struct msi_desc and struct msi_msg
definition.
msi.h doesn't exist yet on ia64 so declare struct msi_desc and
struct msi_msg in iommu.h and don't include asm/msi.h.
Keir Fraser [Fri, 30 May 2008 14:06:08 +0000 (15:06 +0100)]
vtd: Various cleanups and fixes:
* Handle DRHDs with different supported AGAWs. To support this we
create page tables which always have 4 levels, and skip top levels
for units which support only 2 or 3 levels.
* Handle systems with mixed DRHD support for cache snooping. We must
pessimistically CLFLUSH if any DRHD does not support snooping.
Keir Fraser [Thu, 29 May 2008 08:42:59 +0000 (09:42 +0100)]
x86: Fix reboot failure after S3
"warm reset vector" is used to start APs. At booting time, it is set
by BP, and cleared after all APs have started. When do S3, similar work
need to do. However, when enable non-boot CPUs after S3, Xen just sets
this vector, but forgets to clear it after APs startup.
Keir Fraser [Wed, 28 May 2008 15:14:10 +0000 (16:14 +0100)]
vtd: More TLB flush fixes.
- Made the non-present flush testing a bit simpler.
- Removed dma_addr_level_page_maddr(). Use a modified
addr_to_dma_page_maddr() instead.
- Upon mapping new context entry: flush old entry using domid 0 and
always flush iotlb.
Keir Fraser [Wed, 28 May 2008 13:41:23 +0000 (14:41 +0100)]
iommu: Handle sibling device assignment correctly
Domctl interface is extended to allow libxc retrieve device group
information from hypervisor. Vendor-specific iommu_ops is also
extended by adding a new operation "get_device_group_id()", which is
currently a null pointer but could be implemented later for vt-d.
Error will be raised from tools side when user trying to assign PCI
device with a sibling device being driven by dom0. User will keep
being prompted until he has hidden the entire device group (at least,
the sibling devices driven by dom0) in dom0 kernel
parameter. Hopefully this framework could be flexible enough to
support both amd iommu and vt-d.
The following 2 cases are not covered by this patch, but should be
easy to handle.
* Checking for hot-plug devices (maybe we can delay calling
ImageHandler.signalDeviceModel() until all checks are done?)
* Checking for splitted device group between different passthru
domains
Keir Fraser [Wed, 28 May 2008 08:30:48 +0000 (09:30 +0100)]
stubdom: make the build more generic by moving as many stubdom parts
into stubdom/ as possible. That also permits to build all of
ioemu, c and caml stubdoms at the same time.
Signed-off-by: Samuel Thibault <samuel.thibault@eu.citrix.com>
projects / xen.git / log