ICR and TSCDEADLINE MSRs write cause the main MSRs write vmexits in our
product observation, multicast IPIs are not as common as unicast IPI like
RESCHEDULE_VECTOR and CALL_FUNCTION_SINGLE_VECTOR etc.

This patch introduce a mechanism to handle certain performance-critical
WRMSRs in a very early stage of KVM VMExit handler.

This mechanism is specifically used for accelerating writes to x2APIC ICR
that attempt to send a virtual IPI with physical destination-mode, fixed
delivery-mode and single target. Which was found as one of the main causes
of VMExits for Linux workloads.

The reason this mechanism significantly reduce the latency of such virtual
IPIs is by sending the physical IPI to the target vCPU in a very early stage
of KVM VMExit handler, before host interrupts are enabled and before expensive
operations such as reacquiring KVM’s SRCU lock.
Latency is reduced even more when KVM is able to use APICv posted-interrupt
mechanism (which allows to deliver the virtual IPI directly to target vCPU
without the need to kick it to host).

Testing on Xeon Skylake server:

The virtual IPI latency from sender send to receiver receive reduces
more than 200+ cpu cycles.
Reviewed-by: NLiran Alon <liran.alon@oracle.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Sean Christopherson <sean.j.christopherson@intel.com>
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: Liran Alon <liran.alon@oracle.com>
Signed-off-by: NWanpeng Li <wanpengli@tencent.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

WARN if root_hpa is invalid when handling a page fault.  The check on
root_hpa exists for historical reasons that no longer apply to the
current KVM code base.

Remove an equivalent debug-only warning in direct_page_fault(), whose
existence more or less confirms that root_hpa should always be valid
when handling a page fault.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

WARN on the existing invalid root_hpa checks in __direct_map() and
FNAME(fetch).  The "legitimate" path that invalidated root_hpa in the
middle of a page fault is long since gone, i.e. it should no longer be
impossible to invalidate in the middle of a page fault[*].

The root_hpa checks were added by two related commits

  989c6b34 ("KVM: MMU: handle invalid root_hpa at __direct_map")
  37f6a4e2 ("KVM: x86: handle invalid root_hpa everywhere")

to fix a bug where nested_vmx_vmexit() could be called *in the middle*
of a page fault.  At the time, vmx_interrupt_allowed(), which was and
still is used by kvm_can_do_async_pf() via ->interrupt_allowed(),
directly invoked nested_vmx_vmexit() to switch from L2 to L1 to emulate
a VM-Exit on a pending interrupt.  Emulating the nested VM-Exit resulted
in root_hpa being invalidated by kvm_mmu_reset_context() without
explicitly terminating the page fault.

Now that root_hpa is checked for validity by kvm_mmu_page_fault(), WARN
on an invalid root_hpa to detect any flows that reset the MMU while
handling a page fault.  The broken vmx_interrupt_allowed() behavior has
long since been fixed and resetting the MMU during a page fault should
not be considered legal behavior.

[*] It's actually technically possible in FNAME(page_fault)() because it
    calls inject_page_fault() when the guest translation is invalid, but
    in that case the page fault handling is immediately terminated.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Add a check on root_hpa at the beginning of the page fault handler to
consolidate several checks on root_hpa that are scattered throughout the
page fault code.  This is a preparatory step towards eventually removing
such checks altogether, or at the very least WARNing if an invalid root
is encountered.  Remove only the checks that can be easily audited to
confirm that root_hpa cannot be invalidated between their current
location and the new check in kvm_mmu_page_fault(), and aren't currently
protected by mmu_lock, i.e. keep the checks in __direct_map() and
FNAME(fetch) for the time being.

The root_hpa checks that are consolidate were all added by commit

  37f6a4e2 ("KVM: x86: handle invalid root_hpa everywhere")

which was a follow up to a bug fix for __direct_map(), commit

  989c6b34 ("KVM: MMU: handle invalid root_hpa at __direct_map")

At the time, nested VMX had, in hindsight, crazy handling of nested
interrupts and would trigger a nested VM-Exit in ->interrupt_allowed(),
and thus unexpectedly reset the MMU in flows such as can_do_async_pf().

Now that the wonky nested VM-Exit behavior is gone, the root_hpa checks
are bogus and confusing, e.g. it's not at all obvious what they actually
protect against, and at first glance they appear to be broken since many
of them run without holding mmu_lock.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Move the calls to thp_adjust() down a level from the page fault handlers
to the map/fetch helpers and remove the page count shuffling done in
thp_adjust().

Despite holding a reference to the underlying page while processing a
page fault, the page fault flows don't actually rely on holding a
reference to the page when thp_adjust() is called.  At that point, the
fault handlers hold mmu_lock, which prevents mmu_notifier from completing
any invalidations, and have verified no invalidations from mmu_notifier
have occurred since the page reference was acquired (which is done prior
to taking mmu_lock).

The kvm_release_pfn_clean()/kvm_get_pfn() dance in thp_adjust() is a
quirk that is necessitated because thp_adjust() modifies the pfn that is
consumed by its caller.  Because the page fault handlers call
kvm_release_pfn_clean() on said pfn, thp_adjust() needs to transfer the
reference to the correct pfn purely for correctness when the pfn is
released.

Calling thp_adjust() from __direct_map() and FNAME(fetch) means the pfn
adjustment doesn't change the pfn as seen by the page fault handlers,
i.e. the pfn released by the page fault handlers is the same pfn that
was returned by gfn_to_pfn().
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Move thp_adjust() above __direct_map() in preparation of calling
thp_adjust() from  __direct_map() and FNAME(fetch).

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Consolidate the direct MMU page fault handlers into a common helper,
direct_page_fault().  Except for unique max level conditions, the tdp
and nonpaging fault handlers are functionally identical.

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Rename __direct_map()'s param that controls whether or not a disallowed
NX large page should be accounted to match what it actually does.  The
nonpaging_page_fault() case unconditionally passes %false for the param
even though it locally sets lpage_disallowed.

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Persist the max page level calculated via gfn_lpage_is_disallowed() to
the max level "returned" by mapping_level() so that its naturally taken
into account by the max level check that conditions calling
transparent_hugepage_adjust().

Drop the gfn_lpage_is_disallowed() check in thp_adjust() as it's now
handled by mapping_level() and its callers.

Add a comment to document the behavior of host_mapping_level() and its
interaction with max level and transparent huge pages.

Note, transferring the gfn_lpage_is_disallowed() from thp_adjust() to
mapping_level() superficially affects how changes to a memslot's
disallow_lpage count will be handled due to thp_adjust() being run while
holding mmu_lock.

In the more common case where a different vCPU increments the count via
account_shadowed(), gfn_lpage_is_disallowed() is rechecked by set_spte()
to ensure a writable large page isn't created.

In the less common case where the count is decremented to zero due to
all shadow pages in the memslot being zapped, THP behavior now matches
hugetlbfs behavior in the sense that a small page will be created when a
large page could be used if the count reaches zero in the miniscule
window between mapping_level() and acquiring mmu_lock.

Lastly, the new THP behavior also follows hugetlbfs behavior in the
absurdly unlikely scenario of a memslot being moved such that the
memslot's compatibility with respect to large pages changes, but without
changing the validity of the gpf->pfn walk.  I.e. if a memslot is moved
between mapping_level() and snapshotting mmu_seq, it's theoretically
possible to consume a stale disallow_lpage count.  But, since KVM zaps
all shadow pages when moving a memslot and forces all vCPUs to reload a
new MMU, the inserted spte will always be thrown away prior to
completing the memslot move, i.e. whether or not the spte accurately
reflects disallow_lpage is irrelevant.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Restrict the max level for a shadow page based on the guest's level
instead of capping the level after the fact for host-mapped huge pages,
e.g. hugetlbfs pages.  Explicitly capping the max level using the guest
mapping level also eliminates FNAME(page_fault)'s subtle dependency on
THP only supporting 2mb pages.

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Refactor the page fault handlers and mapping_level() to track the max
allowed page level instead of only tracking if a 4k page is mandatory
due to one restriction or another.  This paves the way for cleanly
consolidating tdp_page_fault() and nonpaging_page_fault(), and for
eliminating a redundant check on mmu_gfn_lpage_is_disallowed().

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Invert the loop which adjusts the allowed page level based on what's
compatible with the associated memslot to use a largest-to-smallest
page size walk.  This paves the way for passing around a "max level"
variable instead of having redundant checks and/or multiple booleans.

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Pre-calculate the max level for a TDP page with respect to MTRR cache
consistency in preparation of replacing force_pt_level with max_level,
and eventually combining the bulk of nonpaging_page_fault() and
tdp_page_fault() into a common helper.

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Move nonpaging_page_fault() below try_async_pf() to eliminate the
forward declaration of try_async_pf() and to prepare for combining the
bulk of nonpaging_page_fault() and tdp_page_fault() into a common
helper.

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Fold nonpaging_map() into its sole caller, nonpaging_page_fault(), in
preparation for combining the bulk of nonpaging_page_fault() and
tdp_page_fault() into a common helper.

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Move make_mmu_pages_available() above its first user to put it closer
to related code and eliminate a forward declaration.

No functional change intended.
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Convert a plethora of parameters and variables in the MMU and page fault
flows from type gva_t to gpa_t to properly handle TDP on 32-bit KVM.

Thanks to PSE and PAE paging, 32-bit kernels can access 64-bit physical
addresses.  When TDP is enabled, the fault address is a guest physical
address and thus can be a 64-bit value, even when both KVM and its guest
are using 32-bit virtual addressing, e.g. VMX's VMCS.GUEST_PHYSICAL is a
64-bit field, not a natural width field.

Using a gva_t for the fault address means KVM will incorrectly drop the
upper 32-bits of the GPA.  Ditto for gva_to_gpa() when it is used to
translate L2 GPAs to L1 GPAs.

Opportunistically rename variables and parameters to better reflect the
dual address modes, e.g. use "cr2_or_gpa" for fault addresses and plain
"addr" instead of "vaddr" when the address may be either a GVA or an L2
GPA.  Similarly, use "gpa" in the nonpaging_page_fault() flows to avoid
a confusing "gpa_t gva" declaration; this also sets the stage for a
future patch to combing nonpaging_page_fault() and tdp_page_fault() with
minimal churn.

Sprinkle in a few comments to document flows where an address is known
to be a GVA and thus can be safely truncated to a 32-bit value.  Add
WARNs in kvm_handle_page_fault() and FNAME(gva_to_gpa_nested)() to help
document such cases and detect bugs.

Cc: stable@vger.kernel.org
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

WARN once in kvm_load_guest_fpu() if TIF_NEED_FPU_LOAD is observed, as
that would mean that KVM is corrupting userspace's FPU by saving
unknown register state into arch.user_fpu.  Add a comment to explain
why KVM WARNs on TIF_NEED_FPU_LOAD instead of implementing logic
similar to fpu__copy().
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Unlike most state managed by XSAVE, MPX is initialized to zero on INIT.
Because INITs are usually recognized in the context of a VCPU_RUN call,
kvm_vcpu_reset() puts the guest's FPU so that the FPU state is resident
in memory, zeros the MPX state, and reloads FPU state to hardware.  But,
in the unlikely event that an INIT is recognized during
kvm_arch_vcpu_ioctl_get_mpstate() via kvm_apic_accept_events(),
kvm_vcpu_reset() will call kvm_put_guest_fpu() without a preceding
kvm_load_guest_fpu() and corrupt the guest's FPU state (and possibly
userspace's FPU state as well).

Given that MPX is being removed from the kernel[*], fix the bug with the
simple-but-ugly approach of loading the guest's FPU during
KVM_GET_MP_STATE.

[*] See commit f240652b ("x86/mpx: Remove MPX APIs").

Fixes: f775b13e ("x86,kvm: move qemu/guest FPU switching out to vcpu_run")
Cc: stable@vger.kernel.org
Signed-off-by: NSean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Use out_err jump label to handle resource release. It's a
good practice to release resource in one place and help
eliminate some duplicated code.
Signed-off-by: NMiaohe Lin <linmiaohe@huawei.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Use out_err jump label to handle resource release. It's a
good practice to release resource in one place and help
eliminate some duplicated code.
Signed-off-by: NMiaohe Lin <linmiaohe@huawei.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

There are two declarations of kvm_vcpu_kick() in kvm_host.h where
one of them is redundant. Remove to keep the git grep a bit cleaner.
Reviewed-by: NCornelia Huck <cohuck@redhat.com>
Signed-off-by: NZenghui Yu <yuzenghui@huawei.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Apply reverse fir tree declaration order, shorten some variable names
to avoid line wrap, reformat a block comment, delete an extra blank
line, and use BIT(10) instead of (1u << 10).
Signed-off-by: NJim Mattson <jmattson@google.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Sean Christopherson <sean.j.christopherson@intel.com>
Reviewed-by: NPeter Shier <pshier@google.com>
Reviewed-by: NOliver Upton <oupton@google.com>
Reviewed-by: NJon Cargille <jcargill@google.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

According to the SDM, VMWRITE checks to see if the secondary source
operand corresponds to an unsupported VMCS field before it checks to
see if the secondary source operand corresponds to a VM-exit
information field and the processor does not support writing to
VM-exit information fields.

Fixes: 49f705c5 ("KVM: nVMX: Implement VMREAD and VMWRITE")
Signed-off-by: NJim Mattson <jmattson@google.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: NPeter Shier <pshier@google.com>
Reviewed-by: NOliver Upton <oupton@google.com>
Reviewed-by: NJon Cargille <jcargill@google.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

According to the SDM, a VMWRITE in VMX non-root operation with an
invalid VMCS-link pointer results in VMfailInvalid before the validity
of the VMCS field in the secondary source operand is checked.

For consistency, modify both handle_vmwrite and handle_vmread, even
though there was no problem with the latter.

Fixes: 6d894f49 ("KVM: nVMX: vmread/vmwrite: Use shadow vmcs12 if running L2")
Signed-off-by: NJim Mattson <jmattson@google.com>
Cc: Liran Alon <liran.alon@oracle.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: NPeter Shier <pshier@google.com>
Reviewed-by: NOliver Upton <oupton@google.com>
Reviewed-by: NJon Cargille <jcargill@google.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The mis-spelling is found by checkpatch.pl, so fix them.
Signed-off-by: NXiaoyao Li <xiaoyao.li@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Rename the NMI-window exiting related definitions to match the latest
Intel SDM. No functional changes.
Signed-off-by: NXiaoyao Li <xiaoyao.li@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Rename interrupt-windown exiting related definitions to match the
latest Intel SDM. No functional changes.
Signed-off-by: NXiaoyao Li <xiaoyao.li@intel.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Fix some typos in comment.
Signed-off-by: NMiaohe Lin <linmiaohe@huawei.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Change the last users of "shorthand = 0" to use APIC_DEST_NOSHORT.
Reviewed-by: NVitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: NPeter Xu <peterx@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Callers of kvm_apic_match_dest() should always pass in APIC_DEST_*
macros for either dest_mode and short_hand parameters.  Fix up all the
callers of kvm_apic_match_dest() that are not following the rule.

Since at it, rename the parameter from short_hand to shorthand in
kvm_apic_match_dest(), as suggested by Vitaly.
Reported-by: NSean Christopherson <sean.j.christopherson@intel.com>
Reported-by: NVitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: NVitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: NPeter Xu <peterx@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

We have both APIC_SHORT_MASK and KVM_APIC_SHORT_MASK defined for the
shorthand mask.  Similarly, we have both APIC_DEST_MASK and
KVM_APIC_DEST_MASK defined for the destination mode mask.

Drop the KVM_APIC_* macros and replace the only user of them to use
the APIC_DEST_* macros instead.  At the meantime, move APIC_SHORT_MASK
and APIC_DEST_MASK from lapic.c to lapic.h.
Reviewed-by: NVitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: NPeter Xu <peterx@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

We were using either APIC_DEST_PHYSICAL|APIC_DEST_LOGICAL or 0|1 to
fill in kvm_lapic_irq.dest_mode.  It's fine only because in most cases
when we check against dest_mode it's against APIC_DEST_PHYSICAL (which
equals to 0).  However, that's not consistent.  We'll have problem
when we want to start checking against APIC_DEST_LOGICAL, which does
not equals to 1.

This patch firstly introduces kvm_lapic_irq_dest_mode() helper to take
any boolean of destination mode and return the APIC_DEST_* macro.
Then, it replaces the 0|1 settings of irq.dest_mode with the helper.
Signed-off-by: NPeter Xu <peterx@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

kvm_apic_match_dest() is declared in both ioapic.h and lapic.h.
Remove the declaration in ioapic.h.

kvm_apic_compare_prio() is declared in ioapic.h but defined in
lapic.c.  Move the declaration to lapic.h.

kvm_irq_delivery_to_apic() is declared in ioapic.h but defined in
irq_comm.c.  Move the declaration to irq.h.

hyperv.c needs to use kvm_irq_delivery_to_apic(). Include irq.h in
hyperv.c.
Reviewed-by: NVitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: NPeter Xu <peterx@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The 3rd parameter of kvm_apic_match_dest() is the irq shorthand,
rather than the irq delivery mode.

Fixes: 7ee30bc1 ("KVM: x86: deliver KVM IOAPIC scan request to target vCPUs")
Reviewed-by: NVitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: NPeter Xu <peterx@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

When we reach here, we have desc->sptes[j] = NULL with j = 0.
So we can replace desc->sptes[0] with 0 to make it more clear.
Signed-off-by: NMiaohe Lin <linmiaohe@huawei.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

We can get rid of unnecessary var page in
kvm_set_pfn_dirty() , thus make code style
similar with kvm_set_pfn_accessed().
Signed-off-by: NMiaohe Lin <linmiaohe@huawei.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

We have no way to reach the final statement, remove it.
Signed-off-by: NMiaohe Lin <linmiaohe@huawei.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The comment in kvm_get_shadow_phys_bits refers to MKTME, but the same is actually
true of SME and SEV.  Just use CPUID[0x8000_0008].EAX[7:0] unconditionally if
available, it is simplest and works even if memory is not encrypted.

Cc: stable@vger.kernel.org
Reported-by: NTom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>