... rather than from the legacy path. Update the API to match guest_cpuid(),
and remove its dependence on current.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Paul Durrant <paul.durrant@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Introduce recalculate_cpuid_policy() which clamps a CPUID policy based on the
domains current restrictions.
Each adjustment introduced here mirrors what currently happens in
{pv,hvm}_cpuid(), although some logic is expressed differently.
* The clearing X86_FEATURE_LM for 32bit PV guests, sanitise_featureset()
takes out all 64bit-dependent features in one go.
* The toolstacks choice of X86_FEATURE_ITSC in (by default) clobbered in
domain_cpuid(), but {pv,hvm}_cpuid() needed to account for the host ITSC
value when masking the toolstack value.
This now requires that sanitise_featureset(), lookup_deep_deps() and
associated data needs to be available at runtime, so moves out of __init.
Recalculate the cpuid policy when:
* The domain is first created
* Switching a PV guest to being compat
* Setting disable_migrate or vTSC modes
* The toolstack sets new policy data
The disable_migrate code was previously common. To compensate, move the code
to each archs arch_do_domctl(), as the implementations now differ.
From this point on, domains have full and correct feature-leaf information in
their CPUID policies, allowing for substantial cleanup and improvements.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Stefano Stabellini <sstabellini@kernel.org>
Introduce init_domain_cpuid_policy() to allocate an appropriate cpuid policy
for the domain (currently the domains maximum applicable policy), and call it
during domain construction.
init_guest_cpuid() now needs calling before dom0 is constructed.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Featuresets will eventually live only once in a struct cpuid_policy, but lots
of code currently uses the global featuresets as a linear bitmap. Remove the
existing global *_featureset bitmaps, replacing them with *_policy objects
containing named featureset words and a fs[] linear bitmap.
Two new helpers are introduced to scatter/gather a linear featureset bitmap
to/from the fixed word locations in struct cpuid_policy.
The existing calculate_raw_policy() already obtains the scattered raw
featureset. Gather the raw featureset into raw_policy.fs in
calculate_raw_policy() and drop calculate_raw_featureset() entirely.
Now that host_featureset can't be a straight define of
boot_cpu_data.x86_capability, introduce calculate_host_policy() to suitably
fill in host_policy from boot_cpu_data.x86_capability. (Future changes will
have additional sanitization logic in this function.)
The PV and HVM policy objects and calculation functions have max introduced to
their names, as there will eventually be a distinction between max and default
policies for each domain type. The existing logic works in terms of linear
bitmaps, so scatter the result back into the policy objects.
Leave some compatibility defines providing the old *_featureset API. This
results in no observed change in the *_featureset values, which are still used
at the hypercall and guest_cpuid() interfaces.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
struct cpuid_policy will eventually be a complete replacement for the cpuids[]
array, with a fixed layout and named fields to allow O(1) access to specific
information.
For now, the CPUID content is capped at the 0xd and 0x8000001c leaves, which
matches the maximum policy that the toolstack will generate for a domain. The
xstate leaves extend up to LWP, and the structured features leaf is
implemented with subleaf properties (in anticipation of subleaf 1 appearing
soon), although only subleaf 0 is currently implemented.
Introduce calculate_raw_policy() which fills raw_policy with information,
making use of the new helpers, cpuid_{,count_}leaf().
Finally, rename calculate_featuresets() to init_guest_cpuid(), as it is going
to perform rather more work.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Longterm, pv_cpuid() and hvm_cpuid() will be merged into a single
guest_cpuid(), which is also capable of working outside of current context.
To aid this transtion, introduce guest_cpuid() with the intended API, which
simply defers back to pv_cpuid() or hvm_cpuid() as appropriate.
Introduce struct cpuid_leaf which is used to represent the results of a CPUID
query in a more efficient mannor than passing four pointers through the
calltree.
Update all codepaths which should use the new guest_cpuid() API. These are
the codepaths which have variable inputs, and (other than some specific
x86_emulate() cases) all pertain to servicing a CPUID instruction from a
guest.
The other codepaths using {pv,hvm}_cpuid() with fixed inputs will later be
adjusted to read their data straight from the policy block.
No intended functional change.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Paul Durrant <paul.durrant@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Reviewed-by: Kevin Tian <kevint.tian@intel.com>
The order of destroy function calls in hvm_vcpu_destroy() should be
the reverse of init calls in hvm_vcpu_initialise().
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
[ Fix up tasklet_kill() position ]
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Commit 9e49dcf67f ("xenstore: add per-node generation counter) changed
the TDB layout, which - in order to not break older xenstored running
on the same system - need to be accompanied by a version bump.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Juergen Gross <jgross@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Simpler non-nested brace expansion.
Some entries in the MAINTAINER are not understood by the script, the
ones that contain {,}. This patch fixes it.
This will convert brace expansion style use in MAINTAINER into a regex
that get_maintainer.pl can use to match a path again a maintainer
section.
It is done by using two different regex, the first one will take care of
converting ',' inside '{}' to a '|', one by one, as long as there is at
least two commas. The second regex will do the final convertion of '{,}'
to '(|)'.
With the patch, the right maintainers are displayed, instead of "THE
REST" maintainers, when using the following command for e.g.
$ ./scripts/get_maintainer.pl -f docs/misc/kconfig.txt
The patch also get rid of the warnings, with recent perl:
Unescaped left brace in regex is deprecated, passed through in regex; marked by <-- HERE in m/^docs/misc/kconfig{ <-- HERE ,-language}\.txt/ at ./scripts/get_maintainer.pl line 731.
Signed-off-by: Anthony PERARD <anthony.perard@citrix.com>
Tested-by: Doug Goldstein <cardoe@cardoe.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
tscmode.txt is referenced in xl.cfg(5). Convert it into a pod
formatted man page.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
pci-device-reservations is references in xl.cfg(5), convert it as a man
page in pod format. The name is now prefixed with 'xen-' to avoid
possible name conflicts.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
channel.txt is referenced in xl.cfg(5). Move it to man pages, section 7.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
vtpmmgr.txt is referenced in a man page, convert it to a man page.
The man page is named xen-vtpmmgr to avoid any conflict with other
potential vtpm docs.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
vtpm.txt is referenced in xl.cfg man page. Convert it to pod,
move it to the man folder and update the reference. The man page
is named xen-vtmp to avoid any potential conflict with other
VTPM documentation.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
docs/misc/xl-numa-placement.markdown is referenced by xl.cfg.5 man page,
move it to a man page, section 7.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Make vbd-interface a man page, section7, as this document is
referenced in other man pages (xl-disk-configuration)
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Convert xl-disk-configuration.txt from plain text file to a POD file
to get it as a man page. The references to it in the other man pages
are also updated.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Move docs/misc/xl-network-configuration.markdown to docs/man and
update the references to it in the other man pages.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Some of the docs/misc documents will need to go in man 7 section,
prepare docs/Makefile for it.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Some of the docs/misc documents are written in markdown language.
As an effort to cleanup man pages these documents will be converted into
man pages. To avoid some more conversion, add rules to the docs/Makefile
to generate man pages out of markdown files as well as pod ones.
However, pandoc doesn't know how to convert man pages links. Thus the
man links in markdown pages won't work.
Signed-off-by: Cédric Bosdonnat <cbosdonnat@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Toolstacks (including some out-of-tree ones) use XEN_DOMCTL_set_address_size
at most once per domain, and it ends up having a destructive effect on the
available CPUID policy for a domain.
To avoid ordering issues between altering the policy via domctl, and the
constructive effects which would have to happen from switching back to native,
explicitly reject this case.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
GCCs of at least 4.4 and earlier do not tollerate the initialisiation of the
$VENDOR_cpu_dev structures, because of c_ident becoming an anonymous union.
Instead of using an anonymous union, reintepret c_ident[] in its CPUID form
just in get_cpu_vendor().
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Tested-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Eliminate the mis-naming of 64-bit fields with 32-bit register names
(eflags instead of rflags etc). To ensure no piece of code was missed,
transiently use the underscore prefixed names only for 32-bit register
accesses. This will be cleaned up subsequently.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Commit dc88221c97 ("x86: rename XMM* features to SSE*") pointlessly
added them - these features are always available on 64-bit CPUs. (Let's
not assume this for MMX though in at least the insn emulator.)
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Instead of using an int and providing a define for TRUE and FALSE,
change the code to use stdbool that Xen provides.
Signed-off-by: Doug Goldstein <cardoe@cardoe.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
[Minor style tweaks]
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Currently HVM guests that use upstream qemu do not update xenstore's
availability entry for VCPUs. While it is not strictly necessary for
hotplug to work, xenstore ends up not reflecting actual status of
VCPUs. We should fix this.
Signed-off-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
While doing archeology I found 38ce7ce3, we should make sure
debug_symbols is responsible for adding "-g" to CFLAGS.
Move adding "-g3" from being guarded by debug to being guarded by
debug_symbols.
Signed-off-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
31d41d7b tried to make debug affect tools build only but failed to take
care of debug_symbols (which appends "-g" to CFLAGS).
Move both to tools/Rules.mk at once in this patch.
Signed-off-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
Having them in StdGNU.mk would affect both hypervisor and tools build.
However judging from the commit message of e4cdd74f LTO was only meant
to affect hypvervisor build.
Move the relevant bits to xen/Rules.mk.
Signed-off-by: Wei Liu <wei.liu2@citrix.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
The I/O bitmap doesn't change function depending on mode. 64bit userspace
such as an X server still needs to enter guest_io_okay() to find that the PV
kernel did set up an appropriate virtual I/O bitmap to permit access.
While moving the check, alter its representation to be easier to read.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
ctxt->addr_size is expressed in bits rather than bytes, and has the value 16,
32 or 64. Comparing < 8 made the intended non-64bit paths dead.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Tested-by: Jan Beulich <jbeulich@suse.com>
Since c/s 0888d36b "x86/emul: Correct the decoding of SReg3 operands",
x86_seg_* have followed hardware encodings, meaning that this translation
table is now an identiy transform.
No functional change.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
This is in preparation of eliminating the mis-naming of 64-bit fields
with 32-bit register names (eflags instead of rflags etc). Use the
guaranteed 32-bit underscore prefixed names for now where appropriate.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
When Xen apicv is enabled, wall clock time is faster on Windows7-32
guest with high payload (with 2vCPU, captured from xentrace, in
high payload, the count of IPI interrupt increases rapidly between
these vCPUs).
If IPI intrrupt (vector 0xe1) and periodic timer interrupt (vector 0xd1)
are both pending (index of bit set in vIRR), unfortunately, the IPI
intrrupt is high priority than periodic timer interrupt. Xen updates
IPI interrupt bit set in vIRR to guest interrupt status (RVI) as a high
priority and apicv (Virtual-Interrupt Delivery) delivers IPI interrupt
within VMX non-root operation without a VM-Exit. Within VMX non-root
operation, if periodic timer interrupt index of bit is set in vIRR and
highest, the apicv delivers periodic timer interrupt within VMX non-root
operation as well.
But in current code, if Xen doesn't update periodic timer interrupt bit
set in vIRR to guest interrupt status (RVI) directly, Xen is not aware
of this case to decrease the count (pending_intr_nr) of pending periodic
timer interrupt, then Xen will deliver a periodic timer interrupt again.
And that we update periodic timer interrupt in every VM-entry, there is
a chance that already-injected instance (before EOI-induced exit happens)
will incur another pending IRR setting if there is a VM-exit happens
between virtual interrupt injection (vIRR->0, vISR->1) and EOI-induced
exit (vISR->0), since pt_intr_post hasn't been invoked yet, then the
guest receives more periodic timer interrupt.
So we set eoi_exit_bitmap for intack.vector - give a chance to post
periodic time interrupts when periodic time interrupts become the
highest one.
Signed-off-by: Quan Xu <xuquan8@huawei.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Chao Gao <chao.gao@intel.com>
The use of X86_FEATURES_ONLY was shortlived in Linux for the same problem
encountered here. The following series needs to add extra includes to
asm/cpuid.h, which breaks the build elsewhere given the current hierachy.
Move the feature definitions into a separate header file, which also matches
the solution Linux used.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
Add CPU ID for Atom Z34xx processors. Datasheets indicate support for this,
detailed information about potential quirks or limitations are missing, though.
So we just reuse the definition from official BSP code.
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>
[Linux commit: 5e7ec268fd48d63cfd0e3a9be6c6443f01673bd4]
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
This is in preparation of eliminating the mis-naming of 64-bit fields
with 32-bit register names (eflags instead of rflags etc).
Note that the result is not fully consistent until after at least one
more patch is in place, primarily to limit patch size (by trying to not
touch the same line twice).
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
While this may appear to introduce a truncation issue, the high 32 bits
get zapped already anyway (early in _PRE_EFLAGS() as well as in
_POST_EFLAGS()). Once a subsequent patch switches to use proper 32-bit
EFLAGS operands, we'll in fact end up with more correct code, as that
zeroing of the upper halves will then go away.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
This involves protmode_load_seg() accepting x86_seg_none as input, with
the meaning to
- suppress any exceptions other than #PF,
- not commit any state.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
GIC_INVALID_LR should be 0xff, but actually, defined as ~(uint8_t)0, is
0xffffffff. Fix the problem by placing the ~ operator before the cast.
Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Julien Grall <julien.grall@arm.com>
Comparing 3 integers is more efficient than using strcmp(), and is more useful
to the gcv_guest case than having to fabricate a suitable string to pass. The
gcv_host cases have both options easily to hand, and experimentally, the
resulting code is more efficient.
Update the cpu_dev structure to be more efficient. c_vendor[] only needs to
be 8 bytes long to cover all the CPU drivers Xen has, which avoids storing an
8-byte pointer to 8 bytes of data. Drop c_ident[1] as we have no CPU drivers
with a second ident string, and turn it into an anonymous union to allow
access to the integer values directly.
This avoids all need for the vendor_id union in update_domain_cpuid_info().
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Otherwise booting a cross-vendor guest would cause PCPU hotplug to
malfunction, because of trying to use the wrong CPU driver.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Andrew Cooper
Suravee Suthikulpanit
Jan Beulich
Anthony PERARD
Eric DeVolder
He Chen
Cédric Bosdonnat
Doug Goldstein
Boris Ostrovsky
Wei Liu
Roger Pau Monne
Quan Xu
Piotr Luc
Andy Shevchenko
Stefano Stabellini