KVM: Do not partially reinitialize gfn=>pfn cache during activation

Don't partially reinitialize a gfn=>pfn cache when activating the cache,
and instead assert that the cache is not valid during activation.  Bug
the VM if the assertion fails, as use-after-free and/or data corruption
is all but guaranteed if KVM ends up with a valid-but-inactive cache.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>

KVM: Drop KVM's API to allow temporarily unmapping gfn=>pfn cache

Drop kvm_gpc_unmap() as it has no users and unclear requirements.  The
API was added as part of the original gfn_to_pfn_cache support, but its
sole usage[*] was never merged.  Fold the guts of kvm_gpc_unmap() into
the deactivate path and drop the API.  Omit acquiring refresh_lock as
as concurrent calls to kvm_gpc_deactivate() are not allowed (this is
not enforced, e.g. via lockdep. due to it being called during vCPU
destruction).

If/when temporary unmapping makes a comeback, the desirable behavior is
likely to restrict temporary unmapping to vCPU-exclusive mappings and
require the vcpu->mutex be held to serialize unmap.  Use of the
refresh_lock to protect unmapping was somewhat specuatively added by
commit 93984f19e7bc ("KVM: Fully serialize gfn=>pfn cache refresh via
mutex") to guard against concurrent unmaps, but the primary use case of
the temporary unmap, nested virtualization[*], doesn't actually need or
want concurrent unmaps.

[*] https://lore.kernel.org/all/20211210163625.2886-7-dwmw2@infradead.org

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>

KVM: Use gfn_to_pfn_cache's immutable "kvm" in kvm_gpc_refresh()

Make kvm_gpc_refresh() use kvm instance cached in gfn_to_pfn_cache.

No functional change intended.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Michal Luczaj <mhal@rbox.co>
[sean: leave kvm_gpc_unmap() as-is]
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>

KVM: Clean up hva_to_pfn_retry()

Make hva_to_pfn_retry() use kvm instance cached in gfn_to_pfn_cache.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Michal Luczaj <mhal@rbox.co>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>

KVM: Use gfn_to_pfn_cache's immutable "kvm" in kvm_gpc_check()

Make kvm_gpc_check() use kvm instance cached in gfn_to_pfn_cache.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Michal Luczaj <mhal@rbox.co>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>

KVM: Store immutable gfn_to_pfn_cache properties

Move the assignment of immutable properties @kvm, @vcpu, and @usage to
the initializer.  Make _activate() and _deactivate() use stored values.

Note, @len is also effectively immutable for most cases, but not in the
case of the Xen runstate cache, which may be split across two pages and
the length of the first segment will depend on its address.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Michal Luczaj <mhal@rbox.co>
[sean: handle @len in a separate patch]
Signed-off-by: Sean Christopherson <seanjc@google.com>
[dwmw2: acknowledge that @len can actually change for some use cases]
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>

KVM: x86/xen: add support for 32-bit guests in SCHEDOP_poll

This patch introduces compat version of struct sched_poll for
SCHEDOP_poll sub-operation of sched_op hypercall, reads correct amount
of data (16 bytes in 32-bit case, 24 bytes otherwise) by using new
compat_sched_poll struct, copies it to sched_poll properly, and lets
rest of the code run as is.

Signed-off-by: Metin Kaya <metikaya@amazon.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Paul Durrant <paul@xen.org>

KVM: x86: fix uninitialized variable use on KVM_REQ_TRIPLE_FAULT

If a triple fault was fixed by kvm_x86_ops.nested_ops->triple_fault (by
turning it into a vmexit), there is no need to leave vcpu_enter_guest().
Any vcpu->requests will be caught later before the actual vmentry,
and in fact vcpu_enter_guest() was not initializing the "r" variable.
Depending on the compiler's whims, this could cause the
x86_64/triple_fault_event_test test to fail.

Cc: Maxim Levitsky <mlevitsk@redhat.com>
Fixes: 92e7d5c83aff ("KVM: x86: allow L1 to not intercept triple fault")
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Remove unused argument in gpc_unmap_khva()

Remove the unused @kvm argument from gpc_unmap_khva().

Signed-off-by: Michal Luczaj <mhal@rbox.co>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: Shorten gfn_to_pfn_cache function names

Formalize "gpc" as the acronym and use it in function names.

No functional change intended.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Michal Luczaj <mhal@rbox.co>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/xen: Add runstate tests for 32-bit mode and crossing page boundary

Torture test the cases where the runstate crosses a page boundary, and
and especially the case where it's configured in 32-bit mode and doesn't,
but then switching to 64-bit mode makes it go onto the second page.

To simplify this, make the KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST ioctl
also update the guest runstate area. It already did so if the actual
runstate changed, as a side-effect of kvm_xen_update_runstate(). So
doing it in the plain adjustment case is making it more consistent, as
well as giving us a nice way to trigger the update without actually
running the vCPU again and changing the values.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Paul Durrant <paul@xen.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/xen: Allow XEN_RUNSTATE_UPDATE flag behaviour to be configured

Closer inspection of the Xen code shows that we aren't supposed to be
using the XEN_RUNSTATE_UPDATE flag unconditionally. It should be
explicitly enabled by guests through the HYPERVISOR_vm_assist hypercall.
If we randomly set the top bit of ->state_entry_time for a guest that
hasn't asked for it and doesn't expect it, that could make the runtimes
fail to add up and confuse the guest. Without the flag it's perfectly
safe for a vCPU to read its own vcpu_runstate_info; just not for one
vCPU to read *another's*.

I briefly pondered adding a word for the whole set of VMASST_TYPE_*
flags but the only one we care about for HVM guests is this, so it
seemed a bit pointless.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Message-Id: <20221127122210.248427-3-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/xen: Compatibility fixes for shared runstate area

The guest runstate area can be arbitrarily byte-aligned. In fact, even
when a sane 32-bit guest aligns the overall structure nicely, the 64-bit
fields in the structure end up being unaligned due to the fact that the
32-bit ABI only aligns them to 32 bits.

So setting the ->state_entry_time field to something|XEN_RUNSTATE_UPDATE
is buggy, because if it's unaligned then we can't update the whole field
atomically; the low bytes might be observable before the _UPDATE bit is.
Xen actually updates the *byte* containing that top bit, on its own. KVM
should do the same.

In addition, we cannot assume that the runstate area fits within a single
page. One option might be to make the gfn_to_pfn cache cope with regions
that cross a page — but getting a contiguous virtual kernel mapping of a
discontiguous set of IOMEM pages is a distinctly non-trivial exercise,
and it seems this is the *only* current use case for the GPC which would
benefit from it.

An earlier version of the runstate code did use a gfn_to_hva cache for
this purpose, but it still had the single-page restriction because it
used the uhva directly — because it needs to be able to do so atomically
when the vCPU is being scheduled out, so it used pagefault_disable()
around the accesses and didn't just use kvm_write_guest_cached() which
has a fallback path.

So... use a pair of GPCs for the first and potential second page covering
the runstate area. We can get away with locking both at once because
nothing else takes more than one GPC lock at a time so we can invent
a trivial ordering rule.

The common case where it's all in the same page is kept as a fast path,
but in both cases, the actual guest structure (compat or not) is built
up from the fields in @vx, following preset pointers to the state and
times fields. The only difference is whether those pointers point to
the kernel stack (in the split case) or to guest memory directly via
the GPC.  The fast path is also fixed to use a byte access for the
XEN_RUNSTATE_UPDATE bit, then the only real difference is the dual
memcpy.

Finally, Xen also does write the runstate area immediately when it's
configured. Flip the kvm_xen_update_runstate() and …_guest() functions
and call the latter directly when the runstate area is set. This means
that other ioctls which modify the runstate also write it immediately
to the guest when they do so, which is also intended.

Update the xen_shinfo_test to exercise the pathological case where the
XEN_RUNSTATE_UPDATE flag in the top byte of the state_entry_time is
actually in a different page to the rest of the 64-bit word.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Merge tag 'kvm-s390-next-6.2-1' of https://git./linux/kernel/git/kvms390/linux into HEAD

- Second batch of the lazy destroy patches
- First batch of KVM changes for kernel virtual != physical address support
- Removal of a unused function

KVM: x86: Advertise PREFETCHIT0/1 CPUID to user space

Latest Intel platform Granite Rapids has introduced a new instruction -
PREFETCHIT0/1, which moves code to memory (cache) closer to the
processor depending on specific hints.

The bit definition:
CPUID.(EAX=7,ECX=1):EDX[bit 14]

PREFETCHIT0/1 is on a KVM-only subleaf. Plus an x86_FEATURE definition
for this feature bit to direct it to the KVM entry.

Advertise PREFETCHIT0/1 to KVM userspace. This is safe because there are
no new VMX controls or additional host enabling required for guests to
use this feature.

Signed-off-by: Jiaxi Chen <jiaxi.chen@linux.intel.com>
Message-Id: <20221125125845.1182922-9-jiaxi.chen@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Advertise AVX-NE-CONVERT CPUID to user space

AVX-NE-CONVERT is a new set of instructions which can convert low
precision floating point like BF16/FP16 to high precision floating point
FP32, and can also convert FP32 elements to BF16. This instruction
allows the platform to have improved AI capabilities and better
compatibility.

The bit definition:
CPUID.(EAX=7,ECX=1):EDX[bit 5]

AVX-NE-CONVERT is on a KVM-only subleaf. Plus an x86_FEATURE definition
for this feature bit to direct it to the KVM entry.

Advertise AVX-NE-CONVERT to KVM userspace. This is safe because there
are no new VMX controls or additional host enabling required for guests
to use this feature.

Signed-off-by: Jiaxi Chen <jiaxi.chen@linux.intel.com>
Message-Id: <20221125125845.1182922-8-jiaxi.chen@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Advertise AVX-VNNI-INT8 CPUID to user space

AVX-VNNI-INT8 is a new set of instructions in the latest Intel platform
Sierra Forest, aims for the platform to have superior AI capabilities.
This instruction multiplies the individual bytes of two unsigned or
unsigned source operands, then adds and accumulates the results into the
destination dword element size operand.

The bit definition:
CPUID.(EAX=7,ECX=1):EDX[bit 4]

AVX-VNNI-INT8 is on a new and sparse CPUID leaf and all bits on this
leaf have no truly kernel use case for now. Given that and to save space
for kernel feature bits, move this new leaf to KVM-only subleaf and plus
an x86_FEATURE definition for AVX-VNNI-INT8 to direct it to the KVM
entry.

Advertise AVX-VNNI-INT8 to KVM userspace. This is safe because there are
no new VMX controls or additional host enabling required for guests to
use this feature.

Signed-off-by: Jiaxi Chen <jiaxi.chen@linux.intel.com>
Message-Id: <20221125125845.1182922-7-jiaxi.chen@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

x86: KVM: Advertise AVX-IFMA CPUID to user space

AVX-IFMA is a new instruction in the latest Intel platform Sierra
Forest. This instruction packed multiplies unsigned 52-bit integers and
adds the low/high 52-bit products to Qword Accumulators.

The bit definition:
CPUID.(EAX=7,ECX=1):EAX[bit 23]

AVX-IFMA is on an expected-dense CPUID leaf and some other bits on this
leaf have kernel usages. Given that, define this feature bit like
X86_FEATURE_<name> in kernel. Considering AVX-IFMA itself has no truly
kernel usages and /proc/cpuinfo has too much unreadable flags, hide this
one in /proc/cpuinfo.

Advertise AVX-IFMA to KVM userspace. This is safe because there are no
new VMX controls or additional host enabling required for guests to use
this feature.

Signed-off-by: Jiaxi Chen <jiaxi.chen@linux.intel.com>
Acked-by: Borislav Petkov <bp@suse.de>
Message-Id: <20221125125845.1182922-6-jiaxi.chen@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

x86: KVM: Advertise AMX-FP16 CPUID to user space

Latest Intel platform Granite Rapids has introduced a new instruction -
AMX-FP16, which performs dot-products of two FP16 tiles and accumulates
the results into a packed single precision tile. AMX-FP16 adds FP16
capability and also allows a FP16 GPU trained model to run faster
without loss of accuracy or added SW overhead.

The bit definition:
CPUID.(EAX=7,ECX=1):EAX[bit 21]

AMX-FP16 is on an expected-dense CPUID leaf and some other bits on this
leaf have kernel usages. Given that, define this feature bit like
X86_FEATURE_<name> in kernel. Considering AMX-FP16 itself has no truly
kernel usages and /proc/cpuinfo has too much unreadable flags, hide this
one in /proc/cpuinfo.

Advertise AMX-FP16 to KVM userspace. This is safe because there are no
new VMX controls or additional host enabling required for guests to use
this feature.

Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
Signed-off-by: Jiaxi Chen <jiaxi.chen@linux.intel.com>
Acked-by: Borislav Petkov <bp@suse.de>
Message-Id: <20221125125845.1182922-5-jiaxi.chen@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

x86: KVM: Advertise CMPccXADD CPUID to user space

CMPccXADD is a new set of instructions in the latest Intel platform
Sierra Forest. This new instruction set includes a semaphore operation
that can compare and add the operands if condition is met, which can
improve database performance.

The bit definition:
CPUID.(EAX=7,ECX=1):EAX[bit 7]

CMPccXADD is on an expected-dense CPUID leaf and some other bits on this
leaf have kernel usages. Given that, define this feature bit like
X86_FEATURE_<name> in kernel. Considering CMPccXADD itself has no truly
kernel usages and /proc/cpuinfo has too much unreadable flags, hide this
one in /proc/cpuinfo.

Advertise CMPCCXADD to KVM userspace. This is safe because there are no
new VMX controls or additional host enabling required for guests to use
this feature.

Signed-off-by: Jiaxi Chen <jiaxi.chen@linux.intel.com>
Acked-by: Borislav Petkov <bp@suse.de>
Message-Id: <20221125125845.1182922-4-jiaxi.chen@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Update KVM-only leaf handling to allow for 100% KVM-only leafs

Rename kvm_cpu_cap_init_scattered() to kvm_cpu_cap_init_kvm_defined() in
anticipation of adding KVM-only CPUID leafs that aren't recognized by the
kernel and thus not scattered, i.e. for leafs that are 100% KVM-defined.

Adjust/add comments to kvm_only_cpuid_leafs and KVM_X86_FEATURE to
document how to create new kvm_only_cpuid_leafs entries for scattered
features as well as features that are entirely unknown to the kernel.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20221125125845.1182922-3-jiaxi.chen@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Add BUILD_BUG_ON() to detect bad usage of "scattered" flags

Add a compile-time assert in the SF() macro to detect improper usage,
i.e. to detect passing in an X86_FEATURE_* flag that isn't actually
scattered by the kernel.  Upcoming feature flags will be 100% KVM-only
and will have X86_FEATURE_* macros that point at a kvm_only_cpuid_leafs
word, not a kernel-defined word.  Using SF() and thus boot_cpu_has() for
such feature flags would access memory beyond x86_capability[NCAPINTS]
and at best incorrectly hide a feature, and at worst leak kernel state to
userspace.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20221125125845.1182922-2-jiaxi.chen@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

MAINTAINERS: Add KVM x86/xen maintainer list

Adding Paul as co-maintainer of Xen support to help ensure that things
don't fall through the cracks when I spend three months at a time
travelling...

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Paul Durrant <paul@xen.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/xen: Add CPL to Xen hypercall tracepoint

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: always declare prototype for kvm_arch_irqchip_in_kernel

Architecture code might want to use it even if CONFIG_HAVE_KVM_IRQ_ROUTING
is false; for example PPC XICS has KVM_IRQ_LINE and wants to use
kvm_arch_irqchip_in_kernel from there, but it does not have
KVM_SET_GSI_ROUTING so the prototype was not provided.

Fixes: d663b8a28598 ("KVM: replace direct irq.h inclusion")
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Merge branch 'kvm-dwmw2-fixes' into HEAD

This brings in a few important fixes for Xen emulation.
While nobody should be enabling it, the bug effectively
allows userspace to read arbitrary memory.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: Update gfn_to_pfn_cache khva when it moves within the same page

In the case where a GPC is refreshed to a different location within the
same page, we didn't bother to update it. Mostly we don't need to, but
since the ->khva field also includes the offset within the page, that
does have to be updated.

Fixes: 3ba2c95ea180 ("KVM: Do not incorporate page offset into gfn=>pfn cache user address")
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Paul Durrant <paul@xen.org>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Cc: stable@kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/xen: Only do in-kernel acceleration of hypercalls for guest CPL0

There are almost no hypercalls which are valid from CPL > 0, and definitely
none which are handled by the kernel.

Fixes: 2fd6df2f2b47 ("KVM: x86/xen: intercept EVTCHNOP_send from guests")
Reported-by: Michal Luczaj <mhal@rbox.co>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Cc: stable@kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/xen: Validate port number in SCHEDOP_poll

We shouldn't allow guests to poll on arbitrary port numbers off the end
of the event channel table.

Fixes: 1a65105a5aba ("KVM: x86/xen: handle PV spinlocks slowpath")
[dwmw2: my bug though; the original version did check the validity as a
 side-effect of an idr_find() which I ripped out in refactoring.]
Reported-by: Michal Luczaj <mhal@rbox.co>
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Cc: stable@kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: s390: remove unused gisa_clear_ipm_gisc() function

clang warns about an unused function:
arch/s390/kvm/interrupt.c:317:20:
  error: unused function 'gisa_clear_ipm_gisc' [-Werror,-Wunused-function]
static inline void gisa_clear_ipm_gisc(struct kvm_s390_gisa *gisa, u32 gisc)

Remove gisa_clear_ipm_gisc(), since it is unused and get rid of this
warning.

Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Link: https://lore.kernel.org/r/20221118151133.2974602-1-hca@linux.ibm.com
Signed-off-by: Christian Borntraeger <borntraeger@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>

s390/vfio-ap: GISA: sort out physical vs virtual pointers usage

Fix virtual vs physical address confusion (which currently are the same)
for the GISA when enabling the IRQ.

Signed-off-by: Nico Boehr <nrb@linux.ibm.com>
Reviewed-by: Halil Pasic <pasic@linux.ibm.com>
Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Link: https://lore.kernel.org/r/20221118100429.70453-1-nrb@linux.ibm.com
Message-Id: <20221118100429.70453-1-nrb@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>

KVM: s390: pv: module parameter to fence asynchronous destroy

Add the module parameter "async_destroy", to allow the asynchronous
destroy mechanism to be switched off. This might be useful for
debugging purposes.

The parameter is enabled by default since the feature is opt-in anyway.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Reviewed-by: Nico Boehr <nrb@linux.ibm.com>
Link: https://lore.kernel.org/r/20221111170632.77622-7-imbrenda@linux.ibm.com
Message-Id: <20221111170632.77622-7-imbrenda@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>

KVM: s390: pv: support for Destroy fast UVC

Add support for the Destroy Secure Configuration Fast Ultravisor call,
and take advantage of it for asynchronous destroy.

When supported, the protected guest is destroyed immediately using the
new UVC, leaving only the memory to be cleaned up asynchronously.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: Nico Boehr <nrb@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Link: https://lore.kernel.org/r/20221111170632.77622-6-imbrenda@linux.ibm.com
Message-Id: <20221111170632.77622-6-imbrenda@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>

KVM: s390: pv: avoid export before import if possible

If the appropriate UV feature bit is set, there is no need to perform
an export before import.

The misc feature indicates, among other things, that importing a shared
page from a different protected VM will automatically also transfer its
ownership.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: Nico Boehr <nrb@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Link: https://lore.kernel.org/r/20221111170632.77622-5-imbrenda@linux.ibm.com
Message-Id: <20221111170632.77622-5-imbrenda@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>

KVM: s390: pv: add KVM_CAP_S390_PROTECTED_ASYNC_DISABLE

Add KVM_CAP_S390_PROTECTED_ASYNC_DISABLE to signal that the
KVM_PV_ASYNC_DISABLE and KVM_PV_ASYNC_DISABLE_PREPARE commands for the
KVM_S390_PV_COMMAND ioctl are available.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: Nico Boehr <nrb@linux.ibm.com>
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Link: https://lore.kernel.org/r/20221111170632.77622-4-imbrenda@linux.ibm.com
Message-Id: <20221111170632.77622-4-imbrenda@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>

KVM: s390: pv: api documentation for asynchronous destroy

Add documentation for the new commands added to the KVM_S390_PV_COMMAND
ioctl.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: Nico Boehr <nrb@linux.ibm.com>
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Link: https://lore.kernel.org/r/20221111170632.77622-3-imbrenda@linux.ibm.com
Message-Id: <20221111170632.77622-3-imbrenda@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>

KVM: s390: pv: asynchronous destroy for reboot

Until now, destroying a protected guest was an entirely synchronous
operation that could potentially take a very long time, depending on
the size of the guest, due to the time needed to clean up the address
space from protected pages.

This patch implements an asynchronous destroy mechanism, that allows a
protected guest to reboot significantly faster than previously.

This is achieved by clearing the pages of the old guest in background.
In case of reboot, the new guest will be able to run in the same
address space almost immediately.

The old protected guest is then only destroyed when all of its memory
has been destroyed or otherwise made non protected.

Two new PV commands are added for the KVM_S390_PV_COMMAND ioctl:

KVM_PV_ASYNC_CLEANUP_PREPARE: set aside the current protected VM for
later asynchronous teardown. The current KVM VM will then continue
immediately as non-protected. If a protected VM had already been
set aside for asynchronous teardown, but without starting the teardown
process, this call will fail. There can be at most one VM set aside at
any time. Once it is set aside, the protected VM only exists in the
context of the Ultravisor, it is not associated with the KVM VM
anymore. Its protected CPUs have already been destroyed, but not its
memory. This command can be issued again immediately after starting
KVM_PV_ASYNC_CLEANUP_PERFORM, without having to wait for completion.

KVM_PV_ASYNC_CLEANUP_PERFORM: tears down the protected VM previously
set aside using KVM_PV_ASYNC_CLEANUP_PREPARE. Ideally the
KVM_PV_ASYNC_CLEANUP_PERFORM PV command should be issued by userspace
from a separate thread. If a fatal signal is received (or if the
process terminates naturally), the command will terminate immediately
without completing. All protected VMs whose teardown was interrupted
will be put in the need_cleanup list. The rest of the normal KVM
teardown process will take care of properly cleaning up all remaining
protected VMs, including the ones on the need_cleanup list.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: Nico Boehr <nrb@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Link: https://lore.kernel.org/r/20221111170632.77622-2-imbrenda@linux.ibm.com
Message-Id: <20221111170632.77622-2-imbrenda@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>

KVM: selftests: Rename 'evmcs_test' to 'hyperv_evmcs'

Conform to the rest of Hyper-V emulation selftests which have 'hyperv'
prefix. Get rid of '_test' suffix as well as the purpose of this code
is fairly obvious.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-49-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: hyperv_svm_test: Introduce L2 TLB flush test

Enable Hyper-V L2 TLB flush and check that Hyper-V TLB flush hypercalls
from L2 don't exit to L1 unless 'TlbLockCount' is set in the Partition
assist page.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-48-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: evmcs_test: Introduce L2 TLB flush test

Enable Hyper-V L2 TLB flush and check that Hyper-V TLB flush hypercalls
from L2 don't exit to L1 unless 'TlbLockCount' is set in the
Partition assist page.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-47-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Introduce rdmsr_from_l2() and use it for MSR-Bitmap tests

Hyper-V MSR-Bitmap tests do RDMSR from L2 to exit to L1. While 'evmcs_test'
correctly clobbers all GPRs (which are not preserved), 'hyperv_svm_test'
does not. Introduce a more generic rdmsr_from_l2() to avoid code
duplication and remove hardcoding of MSRs.  Do not put it in common code
because it is really just a selftests bug rather than a processor
feature that requires it.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-46-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Stuff RAX/RCX with 'safe' values in vmmcall()/vmcall()

vmmcall()/vmcall() are used to exit from L2 to L1 and no concrete hypercall
ABI is currenty followed. With the introduction of Hyper-V L2 TLB flush
it becomes (theoretically) possible that L0 will take responsibility for
handling the call and no L1 exit will happen. Prevent this by stuffing RAX
(KVM ABI) and RCX (Hyper-V ABI) with 'safe' values.

While on it, convert vmmcall() to 'static inline', make it setup stack
frame and move to include/x86_64/svm_util.h.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-45-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Allocate Hyper-V partition assist page

In preparation to testing Hyper-V L2 TLB flush hypercalls, allocate
so-called Partition assist page.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-44-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Create a vendor independent helper to allocate Hyper-V specific test pages

There's no need to pollute VMX and SVM code with Hyper-V specific
stuff and allocate Hyper-V specific test pages for all test as only
few really need them. Create a dedicated struct and an allocation
helper.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-43-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Split off load_evmcs() from load_vmcs()

In preparation to putting Hyper-V specific test pages to a dedicated
struct, move eVMCS load logic from load_vmcs(). Tests call load_vmcs()
directly and the only one which needs 'enlightened' version is
evmcs_test so there's not much gain in having this merged.

Temporary pass both GPA and HVA to load_evmcs().

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-42-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Move Hyper-V VP assist page enablement out of evmcs.h

Hyper-V VP assist page is not eVMCS specific, it is also used for
enlightened nSVM. Move the code to vendor neutral place.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-41-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Sync 'struct hv_vp_assist_page' definition with hyperv-tlfs.h

'struct hv_vp_assist_page' definition doesn't match TLFS. Also, define
'struct hv_nested_enlightenments_control' and use it instead of opaque
'__u64'.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-40-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Sync 'struct hv_enlightened_vmcs' definition with hyperv-tlfs.h

'struct hv_enlightened_vmcs' definition in selftests is not '__packed'
and so we rely on the compiler doing the right padding. This is not
obvious so it seems beneficial to use the same definition as in kernel.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-39-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Hyper-V PV TLB flush selftest

Introduce a selftest for Hyper-V PV TLB flush hypercalls
(HvFlushVirtualAddressSpace/HvFlushVirtualAddressSpaceEx,
HvFlushVirtualAddressList/HvFlushVirtualAddressListEx).

The test creates one 'sender' vCPU and two 'worker' vCPU which do busy
loop reading from a certain GVA checking the observed value. Sender
vCPU swaos the data page with another page filled with a different value.
The expectation for workers is also altered. Without TLB flush on worker
vCPUs, they may continue to observe old value. To guard against accidental
TLB flushes for worker vCPUs the test is repeated 100 times.

Hyper-V TLB flush hypercalls are tested in both 'normal' and 'XMM
fast' modes.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-38-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Export vm_vaddr_unused_gap() to make it possible to request unmapped ranges

Currently, tests can only request a new vaddr range by using
vm_vaddr_alloc()/vm_vaddr_alloc_page()/vm_vaddr_alloc_pages() but
these functions allocate and map physical pages too. Make it possible
to request unmapped range too.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-36-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Fill in vm->vpages_mapped bitmap in virt_map() too

Similar to vm_vaddr_alloc(), virt_map() needs to reflect the mapping
in vm->vpages_mapped.

While on it, remove unneeded code wrapping in vm_vaddr_alloc().

Reviewed-by: Andrew Jones <andrew.jones@linux.dev>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-35-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Hyper-V PV IPI selftest

Introduce a selftest for Hyper-V PV IPI hypercalls
(HvCallSendSyntheticClusterIpi, HvCallSendSyntheticClusterIpiEx).

The test creates one 'sender' vCPU and two 'receiver' vCPU and then
issues various combinations of send IPI hypercalls in both 'normal'
and 'fast' (with XMM input where necessary) mode. Later, the test
checks whether IPIs were delivered to the expected destination vCPU[s].

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-34-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Move the function doing Hyper-V hypercall to a common header

All Hyper-V specific tests issuing hypercalls need this.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-33-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Move HYPERV_LINUX_OS_ID definition to a common header

HYPERV_LINUX_OS_ID needs to be written to HV_X64_MSR_GUEST_OS_ID by
each Hyper-V specific selftest.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-32-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Better XMM read/write helpers

set_xmm()/get_xmm() helpers are fairly useless as they only read 64 bits
from 128-bit registers. Moreover, these helpers are not used. Borrow
_kvm_read_sse_reg()/_kvm_write_sse_reg() from KVM limiting them to
XMM0-XMM8 for now.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-31-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Expose Hyper-V L2 TLB flush feature

With both nSVM and nVMX implementations in place, KVM can now expose
Hyper-V L2 TLB flush feature to userspace.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-30-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: nSVM: hyper-v: Enable L2 TLB flush

Implement Hyper-V L2 TLB flush for nSVM. The feature needs to be enabled
both in extended 'nested controls' in VMCB and VP assist page.
According to Hyper-V TLFS, synthetic vmexit to L1 is performed with
- HV_SVM_EXITCODE_ENL exit_code.
- HV_SVM_ENL_EXITCODE_TRAP_AFTER_FLUSH exit_info_1.

Note: VP assist page is cached in 'struct kvm_vcpu_hv' so
recalc_intercepts() doesn't need to read from guest's memory. KVM
needs to update the case upon each VMRUN and after svm_set_nested_state
(svm_get_nested_state_pages()) to handle the case when the guest got
migrated while L2 was running.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-29-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Make kvm_hv_get_assist_page() return 0/-errno

Convert kvm_hv_get_assist_page() to return 'int' and propagate possible
errors from kvm_read_guest_cached().

Suggested-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-28-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: nVMX: hyper-v: Enable L2 TLB flush

Enable L2 TLB flush feature on nVMX when:
- Enlightened VMCS is in use.
- The feature flag is enabled in eVMCS.
- The feature flag is enabled in partition assist page.

Perform synthetic vmexit to L1 after processing TLB flush call upon
request (HV_VMX_SYNTHETIC_EXIT_REASON_TRAP_AFTER_FLUSH).

Note: nested_evmcs_l2_tlb_flush_enabled() uses cached VP assist page copy
which gets updated from nested_vmx_handle_enlightened_vmptrld(). This is
also guaranteed to happen post migration with eVMCS backed L2 running.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-27-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: nVMX: hyper-v: Cache VP assist page in 'struct kvm_vcpu_hv'

In preparation to enabling L2 TLB flush, cache VP assist page in
'struct kvm_vcpu_hv'. While on it, rename nested_enlightened_vmentry()
to nested_get_evmptr() and make it return eVMCS GPA directly.

No functional change intended.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-26-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Introduce fast guest_hv_cpuid_has_l2_tlb_flush() check

Introduce a helper to quickly check if KVM needs to handle VMCALL/VMMCALL
from L2 in L0 to process L2 TLB flush requests.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-25-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: L2 TLB flush

Handle L2 TLB flush requests by going through all vCPUs and checking
whether there are vCPUs running the same VM_ID with a VP_ID specified
in the requests. Perform synthetic exit to L2 upon finish.

Note, while checking VM_ID/VP_ID of running vCPUs seem to be a bit
racy, we count on the fact that KVM flushes the whole L2 VPID upon
transition. Also, KVM_REQ_HV_TLB_FLUSH request needs to be done upon
transition between L1 and L2 to make sure all pending requests are
always processed.

For the reference, Hyper-V TLFS refers to the feature as "Direct
Virtual Flush".

Note, nVMX/nSVM code does not handle VMCALL/VMMCALL from L2 yet.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-24-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Introduce kvm_hv_is_tlb_flush_hcall()

The newly introduced helper checks whether vCPU is performing a
Hyper-V TLB flush hypercall. This is required to filter out L2 TLB
flush hypercalls for processing.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-23-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Introduce .hv_inject_synthetic_vmexit_post_tlb_flush() nested hook

Hyper-V supports injecting synthetic L2->L1 exit after performing
L2 TLB flush operation but the procedure is vendor specific. Introduce
.hv_inject_synthetic_vmexit_post_tlb_flush nested hook for it.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-22-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: nSVM: Keep track of Hyper-V hv_vm_id/hv_vp_id

Similar to nSVM, KVM needs to know L2's VM_ID/VP_ID and Partition
assist page address to handle L2 TLB flush requests.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-21-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: nVMX: Keep track of hv_vm_id/hv_vp_id when eVMCS is in use

To handle L2 TLB flush requests, KVM needs to keep track of L2's VM_ID/
VP_IDs which are set by L1 hypervisor. 'Partition assist page' address is
also needed to handle post-flush exit to L1 upon request.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-20-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Use preallocated buffer in 'struct kvm_vcpu_hv' instead of on-stack 'sparse_banks'

To make kvm_hv_flush_tlb() ready to handle L2 TLB flush requests, KVM needs
to allow for all 64 sparse vCPU banks regardless of KVM_MAX_VCPUs as L1
may use vCPU overcommit for L2. To avoid growing on-stack allocation, make
'sparse_banks' part of per-vCPU 'struct kvm_vcpu_hv' which is allocated
dynamically.

Note: sparse_set_to_vcpu_mask() can't currently be used to handle L2
requests as KVM does not keep L2 VM_ID -> L2 VCPU_ID -> L1 vCPU mappings,
i.e. its vp_bitmap array is still bounded by the number of L1 vCPUs and so
can remain an on-stack allocation.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-19-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Create a separate fifo for L2 TLB flush

To handle L2 TLB flush requests, KVM needs to use a separate fifo from
regular (L1) Hyper-V TLB flush requests: e.g. when a request to flush
something in L2 is made, the target vCPU can transition from L2 to L1,
receive a request to flush a GVA for L1 and then try to enter L2 back.
The first request needs to be processed at this point. Similarly,
requests to flush GVAs in L1 must wait until L2 exits to L1.

No functional change as KVM doesn't handle L2 TLB flush requests from
L2 yet.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-18-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Don't use sparse_set_to_vcpu_mask() in kvm_hv_send_ipi()

Get rid of on-stack allocation of vcpu_mask and optimize kvm_hv_send_ipi()
for a smaller number of vCPUs in the request. When Hyper-V TLB flush
is in  use, HvSendSyntheticClusterIpi{,Ex} calls are not commonly used to
send IPIs to a large number of vCPUs (and are rarely used in general).

Introduce hv_is_vp_in_sparse_set() to directly check if the specified
VP_ID is present in sparse vCPU set.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-17-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Use HV_MAX_SPARSE_VCPU_BANKS/HV_VCPUS_PER_SPARSE_BANK instead of raw '64'

It may not be clear from where the '64' limit for the maximum sparse
bank number comes from, use HV_MAX_SPARSE_VCPU_BANKS define instead.
Use HV_VCPUS_PER_SPARSE_BANK in KVM_HV_MAX_SPARSE_VCPU_SET_BITS's
definition. Opportunistically adjust the comment around BUILD_BUG_ON().

No functional change.

Suggested-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-16-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

x86/hyperv: Introduce HV_MAX_SPARSE_VCPU_BANKS/HV_VCPUS_PER_SPARSE_BANK constants

It may not come clear from where the magical '64' value used in
__cpumask_to_vpset() come from. Moreover, '64' means both the maximum
sparse bank number as well as the number of vCPUs per bank. Add defines
to make things clear. These defines are also going to be used by KVM.

No functional change.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-15-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Prepare kvm_hv_flush_tlb() to handle L2's GPAs

To handle L2 TLB flush requests, KVM needs to translate the specified
L2 GPA to L1 GPA to read hypercall arguments from there.

No functional change as KVM doesn't handle VMCALL/VMMCALL from L2 yet.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-14-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Expose support for extended gva ranges for flush hypercalls

Extended GVA ranges support bit seems to indicate whether lower 12
bits of GVA can be used to specify up to 4095 additional consequent
GVAs to flush. This is somewhat described in TLFS.

Previously, KVM was handling HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST{,EX}
requests by flushing the whole VPID so technically, extended GVA
ranges were already supported. As such requests are handled more
gently now, advertizing support for extended ranges starts making
sense to reduce the size of TLB flush requests.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-13-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Handle HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST{,EX} calls gently

Currently, HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST{,EX} calls are handled
the exact same way as HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE{,EX}: by
flushing the whole VPID and this is sub-optimal. Switch to handling
these requests with 'flush_tlb_gva()' hooks instead. Use the newly
introduced TLB flush fifo to queue the requests.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-12-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Add helper to read hypercall data for array

Move the guts of kvm_get_sparse_vp_set() to a helper so that the code for
reading a guest-provided array can be reused in the future, e.g. for
getting a list of virtual addresses whose TLB entries need to be flushed.

Opportunisticaly swap the order of the data and XMM adjustment so that
the XMM/gpa offsets are bundled together.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-11-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Introduce TLB flush fifo

To allow flushing individual GVAs instead of always flushing the whole
VPID a per-vCPU structure to pass the requests is needed. Use standard
'kfifo' to queue two types of entries: individual GVA (GFN + up to 4095
following GFNs in the lower 12 bits) and 'flush all'.

The size of the fifo is arbitrarily set to '16'.

Note, kvm_hv_flush_tlb() only queues 'flush all' entries for now and
kvm_hv_vcpu_flush_tlb() doesn't actually read the fifo just resets the
queue before returning -EOPNOTSUPP (which triggers full TLB flush) so
the functional change is very small but the infrastructure is prepared
to handle individual GVA flush requests.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-10-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: hyper-v: Resurrect dedicated KVM_REQ_HV_TLB_FLUSH flag

In preparation to implementing fine-grained Hyper-V TLB flush and
L2 TLB flush, resurrect dedicated KVM_REQ_HV_TLB_FLUSH request bit. As
KVM_REQ_TLB_FLUSH_GUEST is a stronger operation, clear KVM_REQ_HV_TLB_FLUSH
request in kvm_vcpu_flush_tlb_guest().

The flush itself is temporary handled by kvm_vcpu_flush_tlb_guest().

No functional change intended.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-9-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Move clearing of TLB_FLUSH_CURRENT to kvm_vcpu_flush_tlb_all()

Clear KVM_REQ_TLB_FLUSH_CURRENT in kvm_vcpu_flush_tlb_all() instead of in
its sole caller that processes KVM_REQ_TLB_FLUSH.  Regardless of why/when
kvm_vcpu_flush_tlb_all() is called, flushing "all" TLB entries also
flushes "current" TLB entries.

Ideally, there will never be another caller of kvm_vcpu_flush_tlb_all(),
and moving the handling "requires" extra work to document the ordering
requirement, but future Hyper-V paravirt TLB flushing support will add
similar logic for flush "guest" (Hyper-V can flush a subset of "guest"
entries).  And in the Hyper-V case, KVM needs to do more than just clear
the request, the queue of GPAs to flush also needs to purged, and doing
all only in the request path is undesirable as kvm_vcpu_flush_tlb_guest()
does have multiple callers (though it's unlikely KVM's paravirt TLB flush
will coincide with Hyper-V's paravirt TLB flush).

Move the logic even though it adds extra "work" so that KVM will be
consistent with how flush requests are processed when the Hyper-V support
lands.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-8-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: VMX: Rename "vmx/evmcs.{ch}" to "vmx/hyperv.{ch}"

To conform with SVM, rename VMX specific Hyper-V files from "evmcs.{ch}"
to "hyperv.{ch}". While Enlightened VMCS is a lion's share of these
files, some stuff (e.g. enlightened MSR bitmap, the upcoming Hyper-V
L2 TLB flush, ...) goes beyond that.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-7-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: Rename 'enable_direct_tlbflush' to 'enable_l2_tlb_flush'

To make terminology between Hyper-V-on-KVM and KVM-on-Hyper-V consistent,
rename 'enable_direct_tlbflush' to 'enable_l2_tlb_flush'. The change
eliminates the use of confusing 'direct' and adds the missing underscore.

No functional change.

Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-6-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

x86/hyperv: KVM: Rename "hv_enlightenments" to "hv_vmcb_enlightenments"

Now that KVM isn't littered with "struct hv_enlightenments" casts, rename
the struct to "hv_vmcb_enlightenments" to highlight the fact that the
struct is specifically for SVM's VMCB.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Michael Kelley <mikelley@microsoft.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-5-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: SVM: Add a proper field for Hyper-V VMCB enlightenments

Add a union to provide hv_enlightenments side-by-side with the sw_reserved
bytes that Hyper-V's enlightenments overlay.  Casting sw_reserved
everywhere is messy, confusing, and unnecessarily unsafe.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-4-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: Move "struct hv_enlightenments" to x86_64/svm.h

Move Hyper-V's VMCB "struct hv_enlightenments" to the svm.h header so
that the struct can be referenced in "struct vmcb_control_area".
Alternatively, a dedicated header for SVM+Hyper-V could be added, a la
x86_64/evmcs.h, but it doesn't appear that Hyper-V will end up needing
a wholesale replacement for the VMCB.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-3-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

x86/hyperv: Move VMCB enlightenment definitions to hyperv-tlfs.h

Move Hyper-V's VMCB enlightenment definitions to the TLFS header; the
definitions come directly from the TLFS[*], not from KVM.

No functional change intended.

[*] https://learn.microsoft.com/en-us/virtualization/hyper-v-on-windows/tlfs/datatypes/hv_svm_enlightened_vmcb_fields

[vitaly: rename VMCB_HV_ -> HV_VMCB_ to match the rest of
hyperv-tlfs.h, keep svm/hyperv.h]

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20221101145426.251680-2-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: avoid memslot check in NX hugepage recovery if it cannot succeed

Since gfn_to_memslot() is relatively expensive, it helps to
skip it if it the memslot cannot possibly have dirty logging
enabled.  In order to do this, add to struct kvm a counter
of the number of log-page memslots.  While the correct value
can only be read with slots_lock taken, the NX recovery thread
is content with using an approximate value.  Therefore, the
counter is an atomic_t.

Based on https://lore.kernel.org/kvm/20221027200316.2221027-2-dmatlack@google.com/
by David Matlack.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Merge branch 'kvm-svm-harden' into HEAD

This fixes three issues in nested SVM:

1) in the shutdown_interception() vmexit handler we call kvm_vcpu_reset().
However, if running nested and L1 doesn't intercept shutdown, the function
resets vcpu->arch.hflags without properly leaving the nested state.
This leaves the vCPU in inconsistent state and later triggers a kernel
panic in SVM code.  The same bug can likely be triggered by sending INIT
via local apic to a vCPU which runs a nested guest.

On VMX we are lucky that the issue can't happen because VMX always
intercepts triple faults, thus triple fault in L2 will always be
redirected to L1.  Plus, handle_triple_fault() doesn't reset the vCPU.
INIT IPI can't happen on VMX either because INIT events are masked while
in VMX mode.

Secondarily, KVM doesn't honour SHUTDOWN intercept bit of L1 on SVM.
A normal hypervisor should always intercept SHUTDOWN, a unit test on
the other hand might want to not do so.

Finally, the guest can trigger a kernel non rate limited printk on SVM
from the guest, which is fixed as well.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: remove exit_int_info warning in svm_handle_exit

It is valid to receive external interrupt and have broken IDT entry,
which will lead to #GP with exit_int_into that will contain the index of
the IDT entry (e.g any value).

Other exceptions can happen as well, like #NP or #SS
(if stack switch fails).

Thus this warning can be user triggred and has very little value.

Cc: stable@vger.kernel.org
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-10-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: add svm part to triple_fault_test

Add a SVM implementation to triple_fault_test to test that
emulated/injected shutdown works.

Since instead of the VMX, the SVM allows the hypervisor to avoid
intercepting shutdown in guest, don't intercept shutdown to test that
KVM suports this correctly.

Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-9-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: allow L1 to not intercept triple fault

This is SVM correctness fix - although a sane L1 would intercept
SHUTDOWN event, it doesn't have to, so we have to honour this.

Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-8-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

kvm: selftests: add svm nested shutdown test

Add test that tests that on SVM if L1 doesn't intercept SHUTDOWN,
then L2 crashes L1 and doesn't crash L2

Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-7-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: selftests: move idt_entry to header

struct idt_entry will be used for a test which will break IDT on purpose.

Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-6-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: forcibly leave nested mode on vCPU reset

While not obivous, kvm_vcpu_reset() leaves the nested mode by clearing
'vcpu->arch.hflags' but it does so without all the required housekeeping.

On SVM, it is possible to have a vCPU reset while in guest mode because
unlike VMX, on SVM, INIT's are not latched in SVM non root mode and in
addition to that L1 doesn't have to intercept triple fault, which should
also trigger L1's reset if happens in L2 while L1 didn't intercept it.

If one of the above conditions happen, KVM will continue to use vmcb02
while not having in the guest mode.

Later the IA32_EFER will be cleared which will lead to freeing of the
nested guest state which will (correctly) free the vmcb02, but since
KVM still uses it (incorrectly) this will lead to a use after free
and kernel crash.

This issue is assigned CVE-2022-3344

Cc: stable@vger.kernel.org
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-5-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: add kvm_leave_nested

add kvm_leave_nested which wraps a call to nested_ops->leave_nested
into a function.

Cc: stable@vger.kernel.org
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-4-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: nSVM: harden svm_free_nested against freeing vmcb02 while still in use

Make sure that KVM uses vmcb01 before freeing nested state, and warn if
that is not the case.

This is a minimal fix for CVE-2022-3344 making the kernel print a warning
instead of a kernel panic.

Cc: stable@vger.kernel.org
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-3-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86: nSVM: leave nested mode on vCPU free

If the VM was terminated while nested, we free the nested state
while the vCPU still is in nested mode.

Soon a warning will be added for this condition.

Cc: stable@vger.kernel.org
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20221103141351.50662-2-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/mmu: Do not recover dirty-tracked NX Huge Pages

Do not recover (i.e. zap) an NX Huge Page that is being dirty tracked,
as it will just be faulted back in at the same 4KiB granularity when
accessed by a vCPU. This may need to be changed if KVM ever supports
2MiB (or larger) dirty tracking granularity, or faulting huge pages
during dirty tracking for reads/executes. However for now, these zaps
are entirely wasteful.

In order to check if this commit increases the CPU usage of the NX
recovery worker thread I used a modified version of execute_perf_test
[1] that supports splitting guest memory into multiple slots and reports
/proc/pid/schedstat:se.sum_exec_runtime for the NX recovery worker just
before tearing down the VM. The goal was to force a large number of NX
Huge Page recoveries and see if the recovery worker used any more CPU.

Test Setup:

  echo 1000 > /sys/module/kvm/parameters/nx_huge_pages_recovery_period_ms
  echo 10 > /sys/module/kvm/parameters/nx_huge_pages_recovery_ratio

Test Command:

  ./execute_perf_test -v64 -s anonymous_hugetlb_1gb -x 16 -o

        | kvm-nx-lpage-re:se.sum_exec_runtime      |
        | ---------------------------------------- |
Run     | Before             | After               |
------- | ------------------ | ------------------- |
1       | 730.084105         | 724.375314          |
2       | 728.751339         | 740.581988          |
3       | 736.264720         | 757.078163          |

Comparing the median results, this commit results in about a 1% increase
CPU usage of the NX recovery worker when testing a VM with 16 slots.
However, the effect is negligible with the default halving time of NX
pages, which is 1 hour rather than 10 seconds given by period_ms = 1000,
ratio = 10.

[1] https://lore.kernel.org/kvm/20221019234050.3919566-2-dmatlack@google.com/

Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20221103204421.1146958-1-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/mmu: simplify kvm_tdp_mmu_map flow when guest has to retry

A removed SPTE is never present, hence the "if" in kvm_tdp_mmu_map
only fails in the exact same conditions that the earlier loop
tested in order to issue a  "break". So, instead of checking twice the
condition (upper level SPTEs could not be created or was frozen), just
exit the loop with a goto---the usual poor-man C replacement for RAII
early returns.

While at it, do not use the "ret" variable for return values of
functions that do not return a RET_PF_* enum.  This is clearer
and also makes it possible to initialize ret to RET_PF_RETRY.

Suggested-by: Robert Hoo <robert.hu@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

KVM: x86/mmu: Split huge pages mapped by the TDP MMU on fault

Now that the TDP MMU has a mechanism to split huge pages, use it in the
fault path when a huge page needs to be replaced with a mapping at a
lower level.

This change reduces the negative performance impact of NX HugePages.
Prior to this change if a vCPU executed from a huge page and NX
HugePages was enabled, the vCPU would take a fault, zap the huge page,
and mapping the faulting address at 4KiB with execute permissions
enabled. The rest of the memory would be left *unmapped* and have to be
faulted back in by the guest upon access (read, write, or execute). If
guest is backed by 1GiB, a single execute instruction can zap an entire
GiB of its physical address space.

For example, it can take a VM longer to execute from its memory than to
populate that memory in the first place:

$ ./execute_perf_test -s anonymous_hugetlb_1gb -v96

Populating memory             : 2.748378795s
Executing from memory         : 2.899670885s

With this change, such faults split the huge page instead of zapping it,
which avoids the non-present faults on the rest of the huge page:

$ ./execute_perf_test -s anonymous_hugetlb_1gb -v96

Populating memory             : 2.729544474s
Executing from memory         : 0.111965688s   <---

This change also reduces the performance impact of dirty logging when
eager_page_split=N. eager_page_split=N (abbreviated "eps=N" below) can
be desirable for read-heavy workloads, as it avoids allocating memory to
split huge pages that are never written and avoids increasing the TLB
miss cost on reads of those pages.

             | Config: ept=Y, tdp_mmu=Y, 5% writes           |
             | Iteration 1 dirty memory time                 |
             | --------------------------------------------- |
vCPU Count   | eps=N (Before) | eps=N (After) | eps=Y        |
------------ | -------------- | ------------- | ------------ |
2            | 0.332305091s   | 0.019615027s  | 0.006108211s |
4            | 0.353096020s   | 0.019452131s  | 0.006214670s |
8            | 0.453938562s   | 0.019748246s  | 0.006610997s |
16           | 0.719095024s   | 0.019972171s  | 0.007757889s |
32           | 1.698727124s   | 0.021361615s  | 0.012274432s |
64           | 2.630673582s   | 0.031122014s  | 0.016994683s |
96           | 3.016535213s   | 0.062608739s  | 0.044760838s |

Eager page splitting remains beneficial for write-heavy workloads, but
the gap is now reduced.

             | Config: ept=Y, tdp_mmu=Y, 100% writes         |
             | Iteration 1 dirty memory time                 |
             | --------------------------------------------- |
vCPU Count   | eps=N (Before) | eps=N (After) | eps=Y        |
------------ | -------------- | ------------- | ------------ |
2            | 0.317710329s   | 0.296204596s  | 0.058689782s |
4            | 0.337102375s   | 0.299841017s  | 0.060343076s |
8            | 0.386025681s   | 0.297274460s  | 0.060399702s |
16           | 0.791462524s   | 0.298942578s  | 0.062508699s |
32           | 1.719646014s   | 0.313101996s  | 0.075984855s |
64           | 2.527973150s   | 0.455779206s  | 0.079789363s |
96           | 2.681123208s   | 0.673778787s  | 0.165386739s |

Further study is needed to determine if the remaining gap is acceptable
for customer workloads or if eager_page_split=N still requires a-priori
knowledge of the VM workload, especially when considering these costs
extrapolated out to large VMs with e.g. 416 vCPUs and 12TB RAM.

Signed-off-by: David Matlack <dmatlack@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Message-Id: <20221109185905.486172-3-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

Merge tag 'kvm-selftests-6.2-1' of https://github.com/kvm-x86/linux into HEAD

KVM selftests updates for 6.2

perf_util:
 - Add support for pinning vCPUs in dirty_log_perf_test.
 - Add a lightweight psuedo RNG for guest use, and use it to randomize
   the access pattern and write vs. read percentage in the so called
   "perf util" tests.
 - Rename the so called "perf_util" framework to "memstress".

ucall:
 - Add a common pool-based ucall implementation (code dedup and pre-work
   for running SEV (and beyond) guests in selftests.
 - Fix an issue in ARM's single-step test when using the new pool-based
   implementation; LDREX/STREX don't play nice with single-step exceptions.

init:
 - Provide a common constructor and arch hook, which will eventually be
   used by x86 to automatically select the right hypercall (AMD vs. Intel).

x86:
 - Clean up x86's page tabe management.
 - Clean up and enhance the "smaller maxphyaddr" test, and add a related
   test to cover generic emulation failure.
 - Clean up the nEPT support checks.
 - Add X86_PROPERTY_* framework to retrieve multi-bit CPUID values.

KVM: selftests: Check for KVM nEPT support using "feature" MSRs

When checking for nEPT support in KVM, use kvm_get_feature_msr() instead
of vcpu_get_msr() to retrieve KVM's default TRUE_PROCBASED_CTLS and
PROCBASED_CTLS2 MSR values, i.e. don't require a VM+vCPU to query nEPT
support.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Link: https://lore.kernel.org/r/20220927165209.930904-1-dmatlack@google.com
[sean: rebase on merged code, write changelog]
Signed-off-by: Sean Christopherson <seanjc@google.com>