The PLE window tracepoint triggers even if the window is not changed,
and the wording can be a bit confusing too. One example line:
kvm_ple_window: vcpu 0: ple_window 4096 (shrink 4096)
It easily let people think of "the window now is 4096 which is
shrinked", but the truth is the value actually didn't change (4096).
Let's only dump this message if the value really changed, and we make
the message even simpler like:
kvm_ple_window: vcpu 4 old 4096 new 8192 (growed)
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The VMX ple_window is 32 bits wide, so logically it can overflow with
an int. The module parameter is declared as unsigned int which is
good, however the dynamic variable is not. Switching all the
ple_window references to use unsigned int.
The tracepoint changes will also affect SVM, but SVM is using an even
smaller width (16 bits) so it's always fine.
Suggested-by: Sean Christopherson <sean.j.christopherson@intel.com>
Reviewed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We can easily route hardware interrupts directly into VM context when
they target the "Fixed" or "LowPriority" delivery modes.
However, on modes such as "SMI" or "Init", we need to go via KVM code
to actually put the vCPU into a different mode of operation, so we can
not post the interrupt
Add code in the VMX and SVM PI logic to explicitly refuse to establish
posted mappings for advanced IRQ deliver modes. This reflects the logic
in __apic_accept_irq() which also only ever passes Fixed and LowPriority
interrupts as posted interrupts into the guest.
This fixes a bug I have with code which configures real hardware to
inject virtual SMIs into my guest.
Signed-off-by: Alexander Graf <graf@amazon.com>
Reviewed-by: Liran Alon <liran.alon@oracle.com>
Reviewed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Reviewed-by: Wanpeng Li <wanpengli@tencent.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM implementations that wrap struct kvm_vcpu with a vendor specific
struct, e.g. struct vcpu_vmx, must place the vcpu member at offset 0,
otherwise the usercopy region intended to encompass struct kvm_vcpu_arch
will instead overlap random chunks of the vendor specific struct.
E.g. padding a large number of bytes before struct kvm_vcpu triggers
a usercopy warn when running with CONFIG_HARDENED_USERCOPY=y.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When shadow paging is enabled, KVM tracks the allowed access type for
MMIO SPTEs so that it can do a permission check on a MMIO GVA cache hit
without having to walk the guest's page tables. The tracking is done
by retaining the WRITE and USER bits of the access when inserting the
MMIO SPTE (read access is implicitly allowed), which allows the MMIO
page fault handler to retrieve and cache the WRITE/USER bits from the
SPTE.
Unfortunately for EPT, the mask used to retain the WRITE/USER bits is
hardcoded using the x86 paging versions of the bits. This funkiness
happens to work because KVM uses a completely different mask/value for
MMIO SPTEs when EPT is enabled, and the EPT mask/value just happens to
overlap exactly with the x86 WRITE/USER bits[*].
Explicitly define the access mask for MMIO SPTEs to accurately reflect
that EPT does not want to incorporate any access bits into the SPTE, and
so that KVM isn't subtly relying on EPT's WX bits always being set in
MMIO SPTEs, e.g. attempting to use other bits for experimentation breaks
horribly.
Note, vcpu_match_mmio_gva() explicits prevents matching GVA==0, and all
TDP flows explicit set mmio_gva to 0, i.e. zeroing vcpu->arch.access for
EPT has no (known) functional impact.
[*] Using WX to generate EPT misconfigurations (equivalent to reserved
bit page fault) ensures KVM can employ its MMIO page fault tricks
even platforms without reserved address bits.
Fixes: ce88decffd ("KVM: MMU: mmio page fault support")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
On AMD, kvm_x86_ops->skip_emulated_instruction(vcpu) can, in theory,
fail: in !nrips case we call kvm_emulate_instruction(EMULTYPE_SKIP).
Currently, we only do printk(KERN_DEBUG) when this happens and this
is not ideal. Propagate the error up the stack.
On VMX, skip_emulated_instruction() doesn't fail, we have two call
sites calling it explicitly: handle_exception_nmi() and
handle_task_switch(), we can just ignore the result.
On SVM, we also have two explicit call sites:
svm_queue_exception() and it seems we don't need to do anything there as
we check if RIP was advanced or not. In task_switch_interception(),
however, we are better off not proceeding to kvm_task_switch() in case
skip_emulated_instruction() failed.
Suggested-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
After commit d73eb57b80 (KVM: Boost vCPUs that are delivering interrupts), a
five years old bug is exposed. Running ebizzy benchmark in three 80 vCPUs VMs
on one 80 pCPUs Skylake server, a lot of rcu_sched stall warning splatting
in the VMs after stress testing:
INFO: rcu_sched detected stalls on CPUs/tasks: { 4 41 57 62 77} (detected by 15, t=60004 jiffies, g=899, c=898, q=15073)
Call Trace:
flush_tlb_mm_range+0x68/0x140
tlb_flush_mmu.part.75+0x37/0xe0
tlb_finish_mmu+0x55/0x60
zap_page_range+0x142/0x190
SyS_madvise+0x3cd/0x9c0
system_call_fastpath+0x1c/0x21
swait_active() sustains to be true before finish_swait() is called in
kvm_vcpu_block(), voluntarily preempted vCPUs are taken into account
by kvm_vcpu_on_spin() loop greatly increases the probability condition
kvm_arch_vcpu_runnable(vcpu) is checked and can be true, when APICv
is enabled the yield-candidate vCPU's VMCS RVI field leaks(by
vmx_sync_pir_to_irr()) into spinning-on-a-taken-lock vCPU's current
VMCS.
This patch fixes it by checking conservatively a subset of events.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Cc: Marc Zyngier <Marc.Zyngier@arm.com>
Cc: stable@vger.kernel.org
Fixes: 98f4a1467 (KVM: add kvm_arch_vcpu_runnable() test to kvm_vcpu_on_spin() loop)
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
After reverting commit 240c35a378 (kvm: x86: Use task structs fpu field
for user), struct kvm_vcpu is 19456 bytes on my server, PAGE_ALLOC_COSTLY_ORDER(3)
is the order at which allocations are deemed costly to service. In serveless
scenario, one host can service hundreds/thoudands firecracker/kata-container
instances, howerver, new instance will fail to launch after memory is too
fragmented to allocate kvm_vcpu struct on host, this was observed in some
cloud provider product environments.
This patch dynamically allocates user_fpu, kvm_vcpu is 15168 bytes now on my
Skylake server.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Dedicated instances are currently disturbed by unnecessary jitter due
to the emulated lapic timers firing on the same pCPUs where the
vCPUs reside. There is no hardware virtual timer on Intel for guest
like ARM, so both programming timer in guest and the emulated timer fires
incur vmexits. This patch tries to avoid vmexit when the emulated timer
fires, at least in dedicated instance scenario when nohz_full is enabled.
In that case, the emulated timers can be offload to the nearest busy
housekeeping cpus since APICv has been found for several years in server
processors. The guest timer interrupt can then be injected via posted interrupts,
which are delivered by the housekeeping cpu once the emulated timer fires.
The host should tuned so that vCPUs are placed on isolated physical
processors, and with several pCPUs surplus for busy housekeeping.
If disabled mwait/hlt/pause vmexits keep the vCPUs in non-root mode,
~3% redis performance benefit can be observed on Skylake server, and the
number of external interrupt vmexits drops substantially. Without patch
VM-EXIT Samples Samples% Time% Min Time Max Time Avg time
EXTERNAL_INTERRUPT 42916 49.43% 39.30% 0.47us 106.09us 0.71us ( +- 1.09% )
While with patch:
VM-EXIT Samples Samples% Time% Min Time Max Time Avg time
EXTERNAL_INTERRUPT 6871 9.29% 2.96% 0.44us 57.88us 0.72us ( +- 4.02% )
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This fixes the following coccinelle warning:
WARNING: return of 0/1 in function 'vmx_need_emulation_on_page_fault'
with return type bool
Return false instead of 0.
Signed-off-by: Yi Wang <wang.yi59@zte.com.cn>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
- Add support for chained PMU counters in guests
- Improve SError handling
- Handle Neoverse N1 erratum #1349291
- Allow side-channel mitigation status to be migrated
- Standardise most AArch64 system register accesses to msr_s/mrs_s
- Fix host MPIDR corruption on 32bit
-----BEGIN PGP SIGNATURE-----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=yfCs
-----END PGP SIGNATURE-----
Merge tag 'kvm-arm-for-5.3' of git://git.kernel.org/pub/scm/linux/kernel/git/kvmarm/kvmarm into HEAD
KVM/arm updates for 5.3
- Add support for chained PMU counters in guests
- Improve SError handling
- Handle Neoverse N1 erratum #1349291
- Allow side-channel mitigation status to be migrated
- Standardise most AArch64 system register accesses to msr_s/mrs_s
- Fix host MPIDR corruption on 32bit
This allows userspace to know which MSRs are supported by the hypervisor.
Unfortunately userspace must resort to tricks for everything except
MSR_IA32_VMX_VMFUNC (which was just added in the previous patch).
One possibility is to use the feature control MSR, which is tied to nested
VMX as well and is present on all KVM versions that support feature MSRs.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Raise #GP when guest read/write IA32_XSS, but the CPUID bits
say that it shouldn't exist.
Fixes: 203000993d (kvm: vmx: add MSR logic for XSAVES)
Reported-by: Xiaoyao Li <xiaoyao.li@linux.intel.com>
Reported-by: Tao Xu <tao3.xu@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Based on 1 normalized pattern(s):
this work is licensed under the terms of the gnu gpl version 2 see
the copying file in the top level directory
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-only
has been chosen to replace the boilerplate/reference in 35 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Reviewed-by: Enrico Weigelt <info@metux.net>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190604081206.797835076@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
VMWRITEs to the major VMCS controls, pin controls included, are
deceptively expensive. CPUs with VMCS caching (Westmere and later) also
optimize away consistency checks on VM-Entry, i.e. skip consistency
checks if the relevant fields have not changed since the last successful
VM-Entry (of the cached VMCS). Because uops are a precious commodity,
uCode's dirty VMCS field tracking isn't as precise as software would
prefer. Notably, writing any of the major VMCS fields effectively marks
the entire VMCS dirty, i.e. causes the next VM-Entry to perform all
consistency checks, which consumes several hundred cycles.
As it pertains to KVM, toggling PIN_BASED_VMX_PREEMPTION_TIMER more than
doubles the latency of the next VM-Entry (and again when/if the flag is
toggled back). In a non-nested scenario, running a "standard" guest
with the preemption timer enabled, toggling the timer flag is uncommon
but not rare, e.g. roughly 1 in 10 entries. Disabling the preemption
timer can change these numbers due to its use for "immediate exits",
even when explicitly disabled by userspace.
Nested virtualization in particular is painful, as the timer flag is set
for the majority of VM-Enters, but prepare_vmcs02() initializes vmcs02's
pin controls to *clear* the flag since its the timer's final state isn't
known until vmx_vcpu_run(). I.e. the majority of nested VM-Enters end
up unnecessarily writing pin controls *twice*.
Rather than toggle the timer flag in pin controls, set the timer value
itself to the largest allowed value to put it into a "soft disabled"
state, and ignore any spurious preemption timer exits.
Sadly, the timer is a 32-bit value and so theoretically it can fire
before the head death of the universe, i.e. spurious exits are possible.
But because KVM does *not* save the timer value on VM-Exit and because
the timer runs at a slower rate than the TSC, the maximuma timer value
is still sufficiently large for KVM's purposes. E.g. on a modern CPU
with a timer that runs at 1/32 the frequency of a 2.4ghz constant-rate
TSC, the timer will fire after ~55 seconds of *uninterrupted* guest
execution. In other words, spurious VM-Exits are effectively only
possible if the host is completely tickless on the logical CPU, the
guest is not using the preemption timer, and the guest is not generating
VM-Exits for any other reason.
To be safe from bad/weird hardware, disable the preemption timer if its
maximum delay is less than ten seconds. Ten seconds is mostly arbitrary
and was selected in no small part because it's a nice round number.
For simplicity and paranoia, fall back to __kvm_request_immediate_exit()
if the preemption timer is disabled by KVM or userspace. Previously
KVM continued to use the preemption timer to force immediate exits even
when the timer was disabled by userspace. Now that KVM leaves the timer
running instead of truly disabling it, allow userspace to kill it
entirely in the unlikely event the timer (or KVM) malfunctions.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
... now that it is fully redundant with the pin controls shadow.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Or: Don't re-initialize vmcs02's controls on every nested VM-Entry.
VMWRITEs to the major VMCS controls are deceptively expensive. Intel
CPUs with VMCS caching (Westmere and later) also optimize away
consistency checks on VM-Entry, i.e. skip consistency checks if the
relevant fields have not changed since the last successful VM-Entry (of
the cached VMCS). Because uops are a precious commodity, uCode's dirty
VMCS field tracking isn't as precise as software would prefer. Notably,
writing any of the major VMCS fields effectively marks the entire VMCS
dirty, i.e. causes the next VM-Entry to perform all consistency checks,
which consumes several hundred cycles.
Zero out the controls' shadow copies during VMCS allocation and use the
optimized setter when "initializing" controls. While this technically
affects both non-nested and nested virtualization, nested virtualization
is the primary beneficiary as avoid VMWRITEs when prepare vmcs02 allows
hardware to optimizie away consistency checks.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Prepare to shadow all major control fields on a per-VMCS basis, which
allows KVM to avoid costly VMWRITEs when switching between vmcs01 and
vmcs02.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Prepare to shadow all major control fields on a per-VMCS basis, which
allows KVM to avoid VMREADs when switching between vmcs01 and vmcs02,
and more importantly can eliminate costly VMWRITEs to controls when
preparing vmcs02.
Shadowing exec controls also saves a VMREAD when opening virtual
INTR/NMI windows, yay...
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Prepare to shadow all major control fields on a per-VMCS basis, which
allows KVM to avoid costly VMWRITEs when switching between vmcs01 and
vmcs02.
Shadowing pin controls also allows a future patch to remove the per-VMCS
'hv_timer_armed' flag, as the shadow copy is a superset of said flag.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM provides a module parameter to allow disabling virtual NMI support
to simplify testing (hardware *without* virtual NMI support is hard to
come by but it does have users). When preparing vmcs02, use the accessor
for pin controls to ensure that the module param is respected for nested
guests.
Opportunistically swap the order of applying L0's and L1's pin controls
to better align with other controls and to prepare for a future patche
that will ignore L1's, but not L0's, preemption timer flag.
Fixes: d02fcf5077 ("kvm: vmx: Allow disabling virtual NMI support")
Cc: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Checking for 32-bit PAE is quite common around code that fiddles with
the PDPTRs. Add a function to compress all checks into a single
invocation.
Reviewed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM unconditionally intercepts WRMSR to MSR_IA32_DEBUGCTLMSR. In the
unlikely event that L1 allows L2 to write L1's MSR_IA32_DEBUGCTLMSR, but
but saves L2's value on VM-Exit, update vmcs12 during L2's WRMSR so as
to eliminate the need to VMREAD the value from vmcs02 on nested VM-Exit.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
For L2, KVM always intercepts WRMSR to SYSENTER MSRs. Update vmcs12 in
the WRMSR handler so that they don't need to be (re)read from vmcs02 on
every nested VM-Exit.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
As alluded to by the TODO comment, KVM unconditionally intercepts writes
to the PAT MSR. In the unlikely event that L1 allows L2 to write L1's
PAT directly but saves L2's PAT on VM-Exit, update vmcs12 when L2 writes
the PAT. This eliminates the need to VMREAD the value from vmcs02 on
VM-Exit as vmcs12 is already up to date in all situations.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When switching between vmcs01 and vmcs02, there is no need to update
state tracking for values that aren't tied to any particular VMCS as
the per-vCPU values are already up-to-date (vmx_switch_vmcs() can only
be called when the vCPU is loaded).
Avoiding the update eliminates a RDMSR, and potentially a RDPKRU and
posted-interrupt update (cmpxchg64() and more).
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When switching between vmcs01 and vmcs02, KVM isn't actually switching
between guest and host. If guest state is already loaded (the likely,
if not guaranteed, case), keep the guest state loaded and manually swap
the loaded_cpu_state pointer after propagating saved host state to the
new vmcs0{1,2}.
Avoiding the switch between guest and host reduces the latency of
switching between vmcs01 and vmcs02 by several hundred cycles, and
reduces the roundtrip time of a nested VM by upwards of 1000 cycles.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
vmx->loaded_cpu_state can only be NULL or equal to vmx->loaded_vmcs,
so change it to a bool. Because the direction of the bool is
now the opposite of vmx->guest_msrs_dirty, change the direction of
vmx->guest_msrs_dirty so that they match.
Finally, do not imply that MSRs have to be reloaded when
vmx->guest_state_loaded is false; instead, set vmx->guest_msrs_ready
to false explicitly in vmx_prepare_switch_to_host.
Cc: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The behavior of WRMSR is in no way dependent on whether or not KVM
consumes the value.
Fixes: 4566654bb9 ("KVM: vmx: Inject #GP on invalid PAT CR")
Cc: stable@vger.kernel.org
Cc: Nadav Amit <nadav.amit@gmail.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Nested virtualization involves copying data between many different types
of VMCSes, e.g. vmcs02, vmcs12, shadow VMCS and eVMCS. Rename a variety
of functions and flags to document both the source and destination of
each sync.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Per commit 1b6269db3f ("KVM: VMX: Handle NMIs before enabling
interrupts and preemption"), NMIs are handled directly in vmx_vcpu_run()
to "make sure we handle NMI on the current cpu, and that we don't
service maskable interrupts before non-maskable ones". The other
exceptions handled by complete_atomic_exit(), e.g. async #PF and #MC,
have similar requirements, and are located there to avoid extra VMREADs
since VMX bins hardware exceptions and NMIs into a single exit reason.
Clean up the code and eliminate the vaguely named complete_atomic_exit()
by moving the interrupts-disabled exception and NMI handling into the
existing handle_external_intrs() callback, and rename the callback to
a more appropriate name. Rename VMexit handlers throughout so that the
atomic and non-atomic counterparts have similar names.
In addition to improving code readability, this also ensures the NMI
handler is run with the host's debug registers loaded in the unlikely
event that the user is debugging NMIs. Accuracy of the last_guest_tsc
field is also improved when handling NMIs (and #MCs) as the handler
will run after updating said field.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
[Naming cleanups. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
VMX can conditionally call kvm_{before,after}_interrupt() since KVM
always uses "ack interrupt on exit" and therefore explicitly handles
interrupts as opposed to blindly enabling irqs.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Although the kernel may use multiple IDTs, KVM should only ever see the
"real" IDT, e.g. the early init IDT is long gone by the time KVM runs
and the debug stack IDT is only used for small windows of time in very
specific flows.
Before commit a547c6db4d ("KVM: VMX: Enable acknowledge interupt on
vmexit"), the kernel's IDT base was consumed by KVM only when setting
constant VMCS state, i.e. to set VMCS.HOST_IDTR_BASE. Because constant
host state is done once per vCPU, there was ostensibly no need to cache
the kernel's IDT base.
When support for "ack interrupt on exit" was introduced, KVM added a
second consumer of the IDT base as handling already-acked interrupts
requires directly calling the interrupt handler, i.e. KVM uses the IDT
base to find the address of the handler. Because interrupts are a fast
path, KVM cached the IDT base to avoid having to VMREAD HOST_IDTR_BASE.
Presumably, the IDT base was cached on a per-vCPU basis simply because
the existing code grabbed the IDT base on a per-vCPU (VMCS) basis.
Note, all post-boot IDTs use the same handlers for external interrupts,
i.e. the "ack interrupt on exit" use of the IDT base would be unaffected
even if the cached IDT somehow did not match the current IDT. And as
for the original use case of setting VMCS.HOST_IDTR_BASE, if any of the
above analysis is wrong then KVM has had a bug since the beginning of
time since KVM has effectively been caching the IDT at vCPU creation
since commit a8b732ca01c ("[PATCH] kvm: userspace interface").
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Generic x86 code invokes the kvm_x86_ops external interrupt handler on
all VM-Exits regardless of the actual exit type. Use the already-cached
EXIT_REASON to determine if the VM-Exit was due to an interrupt, thus
avoiding an extra VMREAD (to query VM_EXIT_INTR_INFO) for all other
types of VM-Exit.
In addition to avoiding the extra VMREAD, checking the EXIT_REASON
instead of VM_EXIT_INTR_INFO makes it more obvious that
vmx_handle_external_intr() is called for all VM-Exits, e.g. someone
unfamiliar with the flow might wonder under what condition(s)
VM_EXIT_INTR_INFO does not contain a valid interrupt, which is
simply not possible since KVM always runs with "ack interrupt on exit".
WARN once if VM_EXIT_INTR_INFO doesn't contain a valid interrupt on
an EXTERNAL_INTERRUPT VM-Exit, as such a condition would indicate a
hardware bug.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The reason for skipping handling of NMI and #MC in handle_exception is
the same, namely they are handled earlier by vmx_complete_atomic_exit.
Calling the machine check handler (which just returns 1) is misleading,
don't do it.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
A previous fix to prevent KVM from consuming stale VMCS state after a
failed VM-Entry inadvertantly blocked KVM's handling of machine checks
that occur during VM-Entry.
Per Intel's SDM, a #MC during VM-Entry is handled in one of three ways,
depending on when the #MC is recognoized. As it pertains to this bug
fix, the third case explicitly states EXIT_REASON_MCE_DURING_VMENTRY
is handled like any other VM-Exit during VM-Entry, i.e. sets bit 31 to
indicate the VM-Entry failed.
If a machine-check event occurs during a VM entry, one of the following occurs:
- The machine-check event is handled as if it occurred before the VM entry:
...
- The machine-check event is handled after VM entry completes:
...
- A VM-entry failure occurs as described in Section 26.7. The basic
exit reason is 41, for "VM-entry failure due to machine-check event".
Explicitly handle EXIT_REASON_MCE_DURING_VMENTRY as a one-off case in
vmx_vcpu_run() instead of binning it into vmx_complete_atomic_exit().
Doing so allows vmx_vcpu_run() to handle VMX_EXIT_REASONS_FAILED_VMENTRY
in a sane fashion and also simplifies vmx_complete_atomic_exit() since
VMCS.VM_EXIT_INTR_INFO is guaranteed to be fresh.
Fixes: b060ca3b2e ("kvm: vmx: Handle VMLAUNCH/VMRESUME failure properly")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Make it available to AMD hosts as well, just in case someone is trying
to use an Intel processor's CPUID setup.
Suggested-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
There is an imperfection in get_vmx_mem_address(): access length is ignored
when checking the limit. To fix this, pass access length as a function argument.
The access length is usually obvious since it is used by callers after
get_vmx_mem_address() call, but for vmread/vmwrite it depends on the
state of 64-bit mode.
Signed-off-by: Eugene Korenevsky <ekorenevsky@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Make all code consistent with kvm_deliver_exception_payload() by using
appropriate symbolic constant instead of hard-coded number.
Reviewed-by: Nikita Leshenko <nikita.leshchenko@oracle.com>
Reviewed-by: Krish Sadhukhan <krish.sadhukhan@oracle.com>
Signed-off-by: Liran Alon <liran.alon@oracle.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Allow guest reads CORE cstate when exposing host CPU power management capabilities
to the guest. PKG cstate is restricted to avoid a guest to get the whole package
information in multi-tenant scenario.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Sean Christopherson <sean.j.christopherson@intel.com>
Cc: Liran Alon <liran.alon@oracle.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add a wrapper to invoke kvm_arch_check_processor_compat() so that the
boilerplate ugliness of checking virtualization support on all CPUs is
hidden from the arch specific code. x86's implementation in particular
is quite heinous, as it unnecessarily propagates the out-param pattern
into kvm_x86_ops.
While the x86 specific issue could be resolved solely by changing
kvm_x86_ops, make the change for all architectures as returning a value
directly is prettier and technically more robust, e.g. s390 doesn't set
the out param, which could lead to subtle breakage in the (highly
unlikely) scenario where the out-param was not pre-initialized by the
caller.
Opportunistically annotate svm_check_processor_compat() with __init.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Advance lapic timer tries to hidden the hypervisor overhead between the
host emulated timer fires and the guest awares the timer is fired. However,
it just hidden the time between apic_timer_fn/handle_preemption_timer ->
wait_lapic_expire, instead of the real position of vmentry which is
mentioned in the orignial commit d0659d946b ("KVM: x86: add option to
advance tscdeadline hrtimer expiration"). There is 700+ cpu cycles between
the end of wait_lapic_expire and before world switch on my haswell desktop.
This patch tries to narrow the last gap(wait_lapic_expire -> world switch),
it takes the real overhead time between apic_timer_fn/handle_preemption_timer
and before world switch into consideration when adaptively tuning timer
advancement. The patch can reduce 40% latency (~1600+ cycles to ~1000+ cycles
on a haswell desktop) for kvm-unit-tests/tscdeadline_latency when testing
busy waits.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Sean Christopherson <sean.j.christopherson@intel.com>
Cc: Liran Alon <liran.alon@oracle.com>
Reviewed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Userspace can easily set up invalid processor state in such a way that
dmesg will be filled with VMCS or VMCB dumps. Disable this by default
using a module parameter.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
* POWER: support for direct access to the POWER9 XIVE interrupt controller,
memory and performance optimizations.
* x86: support for accessing memory not backed by struct page, fixes and refactoring
* Generic: dirty page tracking improvements
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.22 (GNU/Linux)
iQEcBAABAgAGBQJc3qV/AAoJEL/70l94x66Dn3QH/jX1Bn0P/RZAIt4w0SySklSg
PqxUKDyBQqB9vN9Qeb9jWXAKPH2CtM3+up/rz7oRnBWp7qA6vXcC/R/QJYAvzdXE
nklsR/oYCsflR1KdlVYuDvvPCPP2fLBU5zfN83OsaBQ8fNRkm3gN+N5XQ2SbXbLy
Mo9tybS4otY201UAC96e8N0ipwwyCRpDneQpLcl+F5nH3RBt63cVbs04O+70MXn7
eT4I+8K3+Go7LATzT8hglD21D/7uvE31qQb6yr5L33IfhU4GB51RZzBXTNaAdY8n
hT1rMrRkAMAFWYZPQDfoMadjWU3i5DIfstKjDxOr9oTfuOEp5Z+GvJwvVnUDg1I=
=D0+p
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
Pull KVM updates from Paolo Bonzini:
"ARM:
- support for SVE and Pointer Authentication in guests
- PMU improvements
POWER:
- support for direct access to the POWER9 XIVE interrupt controller
- memory and performance optimizations
x86:
- support for accessing memory not backed by struct page
- fixes and refactoring
Generic:
- dirty page tracking improvements"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (155 commits)
kvm: fix compilation on aarch64
Revert "KVM: nVMX: Expose RDPMC-exiting only when guest supports PMU"
kvm: x86: Fix L1TF mitigation for shadow MMU
KVM: nVMX: Disable intercept for FS/GS base MSRs in vmcs02 when possible
KVM: PPC: Book3S: Remove useless checks in 'release' method of KVM device
KVM: PPC: Book3S HV: XIVE: Fix spelling mistake "acessing" -> "accessing"
KVM: PPC: Book3S HV: Make sure to load LPID for radix VCPUs
kvm: nVMX: Set nested_run_pending in vmx_set_nested_state after checks complete
tests: kvm: Add tests for KVM_SET_NESTED_STATE
KVM: nVMX: KVM_SET_NESTED_STATE - Tear down old EVMCS state before setting new state
tests: kvm: Add tests for KVM_CAP_MAX_VCPUS and KVM_CAP_MAX_CPU_ID
tests: kvm: Add tests to .gitignore
KVM: Introduce KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
KVM: Fix kvm_clear_dirty_log_protect off-by-(minus-)one
KVM: Fix the bitmap range to copy during clear dirty
KVM: arm64: Fix ptrauth ID register masking logic
KVM: x86: use direct accessors for RIP and RSP
KVM: VMX: Use accessors for GPRs outside of dedicated caching logic
KVM: x86: Omit caching logic for always-available GPRs
kvm, x86: Properly check whether a pfn is an MMIO or not
...
The RDPMC-exiting control is dependent on the existence of the RDPMC
instruction itself, i.e. is not tied to the "Architectural Performance
Monitoring" feature. For all intents and purposes, the control exists
on all CPUs with VMX support since RDPMC also exists on all VCPUs with
VMX supported. Per Intel's SDM:
The RDPMC instruction was introduced into the IA-32 Architecture in
the Pentium Pro processor and the Pentium processor with MMX technology.
The earlier Pentium processors have performance-monitoring counters, but
they must be read with the RDMSR instruction.
Because RDPMC-exiting always exists, KVM requires the control and refuses
to load if it's not available. As a result, hiding the PMU from a guest
breaks nested virtualization if the guest attemts to use KVM.
While it's not explicitly stated in the RDPMC pseudocode, the VM-Exit
check for RDPMC-exiting follows standard fault vs. VM-Exit prioritization
for privileged instructions, e.g. occurs after the CPL/CR0.PE/CR4.PCE
checks, but before the counter referenced in ECX is checked for validity.
In other words, the original KVM behavior of injecting a #GP was correct,
and the KVM unit test needs to be adjusted accordingly, e.g. eat the #GP
when the unit test guest (L3 in this case) executes RDPMC without
RDPMC-exiting set in the unit test host (L2).
This reverts commit e51bfdb687.
Fixes: e51bfdb687 ("KVM: nVMX: Expose RDPMC-exiting only when guest supports PMU")
Reported-by: David Hill <hilld@binarystorm.net>
Cc: Saar Amar <saaramar@microsoft.com>
Cc: Mihai Carabas <mihai.carabas@oracle.com>
Cc: Jim Mattson <jmattson@google.com>
Cc: Liran Alon <liran.alon@oracle.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Pull x86 MDS mitigations from Thomas Gleixner:
"Microarchitectural Data Sampling (MDS) is a hardware vulnerability
which allows unprivileged speculative access to data which is
available in various CPU internal buffers. This new set of misfeatures
has the following CVEs assigned:
CVE-2018-12126 MSBDS Microarchitectural Store Buffer Data Sampling
CVE-2018-12130 MFBDS Microarchitectural Fill Buffer Data Sampling
CVE-2018-12127 MLPDS Microarchitectural Load Port Data Sampling
CVE-2019-11091 MDSUM Microarchitectural Data Sampling Uncacheable Memory
MDS attacks target microarchitectural buffers which speculatively
forward data under certain conditions. Disclosure gadgets can expose
this data via cache side channels.
Contrary to other speculation based vulnerabilities the MDS
vulnerability does not allow the attacker to control the memory target
address. As a consequence the attacks are purely sampling based, but
as demonstrated with the TLBleed attack samples can be postprocessed
successfully.
The mitigation is to flush the microarchitectural buffers on return to
user space and before entering a VM. It's bolted on the VERW
instruction and requires a microcode update. As some of the attacks
exploit data structures shared between hyperthreads, full protection
requires to disable hyperthreading. The kernel does not do that by
default to avoid breaking unattended updates.
The mitigation set comes with documentation for administrators and a
deeper technical view"
* 'x86-mds-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (23 commits)
x86/speculation/mds: Fix documentation typo
Documentation: Correct the possible MDS sysfs values
x86/mds: Add MDSUM variant to the MDS documentation
x86/speculation/mds: Add 'mitigations=' support for MDS
x86/speculation/mds: Print SMT vulnerable on MSBDS with mitigations off
x86/speculation/mds: Fix comment
x86/speculation/mds: Add SMT warning message
x86/speculation: Move arch_smt_update() call to after mitigation decisions
x86/speculation/mds: Add mds=full,nosmt cmdline option
Documentation: Add MDS vulnerability documentation
Documentation: Move L1TF to separate directory
x86/speculation/mds: Add mitigation mode VMWERV
x86/speculation/mds: Add sysfs reporting for MDS
x86/speculation/mds: Add mitigation control for MDS
x86/speculation/mds: Conditionally clear CPU buffers on idle entry
x86/kvm/vmx: Add MDS protection when L1D Flush is not active
x86/speculation/mds: Clear CPU buffers on exit to user
x86/speculation/mds: Add mds_clear_cpu_buffers()
x86/kvm: Expose X86_FEATURE_MD_CLEAR to guests
x86/speculation/mds: Add BUG_MSBDS_ONLY
...
Pull x86 FPU state handling updates from Borislav Petkov:
"This contains work started by Rik van Riel and brought to fruition by
Sebastian Andrzej Siewior with the main goal to optimize when to load
FPU registers: only when returning to userspace and not on every
context switch (while the task remains in the kernel).
In addition, this optimization makes kernel_fpu_begin() cheaper by
requiring registers saving only on the first invocation and skipping
that in following ones.
What is more, this series cleans up and streamlines many aspects of
the already complex FPU code, hopefully making it more palatable for
future improvements and simplifications.
Finally, there's a __user annotations fix from Jann Horn"
* 'x86-fpu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (29 commits)
x86/fpu: Fault-in user stack if copy_fpstate_to_sigframe() fails
x86/pkeys: Add PKRU value to init_fpstate
x86/fpu: Restore regs in copy_fpstate_to_sigframe() in order to use the fastpath
x86/fpu: Add a fastpath to copy_fpstate_to_sigframe()
x86/fpu: Add a fastpath to __fpu__restore_sig()
x86/fpu: Defer FPU state load until return to userspace
x86/fpu: Merge the two code paths in __fpu__restore_sig()
x86/fpu: Restore from kernel memory on the 64-bit path too
x86/fpu: Inline copy_user_to_fpregs_zeroing()
x86/fpu: Update xstate's PKRU value on write_pkru()
x86/fpu: Prepare copy_fpstate_to_sigframe() for TIF_NEED_FPU_LOAD
x86/fpu: Always store the registers in copy_fpstate_to_sigframe()
x86/entry: Add TIF_NEED_FPU_LOAD
x86/fpu: Eager switch PKRU state
x86/pkeys: Don't check if PKRU is zero before writing it
x86/fpu: Only write PKRU if it is different from current
x86/pkeys: Provide *pkru() helpers
x86/fpu: Use a feature number instead of mask in two more helpers
x86/fpu: Make __raw_xsave_addr() use a feature number instead of mask
x86/fpu: Add an __fpregs_load_activate() internal helper
...
... now that there is no overhead when using dedicated accessors.
Opportunistically remove a bogus "FIXME" in handle_rdmsr() regarding
the upper 32 bits of RAX and RDX. Zeroing the upper 32 bits is
architecturally correct as 32-bit writes in 64-bit mode unconditionally
clear the upper 32 bits.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Except for RSP and RIP, which are held in VMX's VMCS, GPRs are always
treated "available and dirtly" on both VMX and SVM, i.e. are
unconditionally loaded/saved immediately before/after VM-Enter/VM-Exit.
Eliminating the unnecessary caching code reduces the size of KVM by a
non-trivial amount, much of which comes from the most common code paths.
E.g. on x86_64, kvm_emulate_cpuid() is reduced from 342 to 182 bytes and
kvm_emulate_hypercall() from 1362 to 1143, with the total size of KVM
dropping by ~1000 bytes. With CONFIG_RETPOLINE=y, the numbers are even
more pronounced, e.g.: 353->182, 1418->1172 and well over 2000 bytes.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
