KVM's recent bug fix to update %rip after emulating I/O broke userspace
that relied on the previous behavior of incrementing %rip prior to
exiting to userspace. When running a Windows XP guest on AMD hardware,
Qemu may patch "OUT 0x7E" instructions in reaction to the OUT itself.
Because KVM's old behavior was to increment %rip before exiting to
userspace to handle the I/O, Qemu manually adjusted %rip to account for
the OUT instruction.
Arguably this is a userspace bug as KVM requires userspace to re-enter
the kernel to complete instruction emulation before taking any other
actions. That being said, this is a bit of a grey area and breaking
userspace that has worked for many years is bad.
Pre-increment %rip on OUT to port 0x7e before exiting to userspace to
hack around the issue.
Fixes: 45def77ebf ("KVM: x86: update %rip after emulating IO")
Reported-by: Simon Becherer <simon@becherer.de>
Reported-and-tested-by: Iakov Karpov <srid@rkmail.ru>
Reported-by: Gabriele Balducci <balducci@units.it>
Reported-by: Antti Antinoja <reader@fennosys.fi>
Cc: stable@vger.kernel.org
Cc: Takashi Iwai <tiwai@suse.com>
Cc: Jiri Slaby <jslaby@suse.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The not-so-recent change to move VMX's VM-Exit handing to a dedicated
"function" unintentionally exposed KVM to a speculative attack from the
guest by executing a RET prior to stuffing the RSB. Make RSB stuffing
happen immediately after VM-Exit, before any unpaired returns.
Alternatively, the VM-Exit path could postpone full RSB stuffing until
its current location by stuffing the RSB only as needed, or by avoiding
returns in the VM-Exit path entirely, but both alternatives are beyond
ugly since vmx_vmexit() has multiple indirect callers (by way of
vmx_vmenter()). And putting the RSB stuffing immediately after VM-Exit
makes it much less likely to be re-broken in the future.
Note, the cost of PUSH/POP could be avoided in the normal flow by
pairing the PUSH RAX with the POP RAX in __vmx_vcpu_run() and adding an
a POP to nested_vmx_check_vmentry_hw(), but such a weird/subtle
dependency is likely to cause problems in the long run, and PUSH/POP
will take all of a few cycles, which is peanuts compared to the number
of cycles required to fill the RSB.
Fixes: 453eafbe65 ("KVM: VMX: Move VM-Enter + VM-Exit handling to non-inline sub-routines")
Reported-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Co-developed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
To minimize the latency of timer interrupts as observed by the guest,
KVM adjusts the values it programs into the host timers to account for
the host's overhead of programming and handling the timer event. In
the event that the adjustments are too aggressive, i.e. the timer fires
earlier than the guest expects, KVM busy waits immediately prior to
entering the guest.
Currently, KVM manually converts the delay from nanoseconds to clock
cycles. But, the conversion is done in the guest's time domain, while
the delay occurs in the host's time domain. This is perfectly ok when
the guest and host are using the same TSC ratio, but if the guest is
using a different ratio then the delay may not be accurate and could
wait too little or too long.
When the guest is not using the host's ratio, convert the delay from
guest clock cycles to host nanoseconds and use ndelay() instead of
__delay() to provide more accurate timing. Because converting to
nanoseconds is relatively expensive, e.g. requires division and more
multiplication ops, continue using __delay() directly when guest and
host TSCs are running at the same ratio.
Cc: Liran Alon <liran.alon@oracle.com>
Cc: Wanpeng Li <wanpengli@tencent.com>
Cc: stable@vger.kernel.org
Fixes: 3b8a5df6c4 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The introduction of adaptive tuning of lapic timer advancement did not
allow for the scenario where userspace would want to disable adaptive
tuning but still employ timer advancement, e.g. for testing purposes or
to handle a use case where adaptive tuning is unable to settle on a
suitable time. This is epecially pertinent now that KVM places a hard
threshold on the maximum advancment time.
Rework the timer semantics to accept signed values, with a value of '-1'
being interpreted as "use adaptive tuning with KVM's internal default",
and any other value being used as an explicit advancement time, e.g. a
time of '0' effectively disables advancement.
Note, this does not completely restore the original behavior of
lapic_timer_advance_ns. Prior to tracking the advancement per vCPU,
which is necessary to support autotuning, userspace could adjust
lapic_timer_advance_ns for *running* vCPU. With per-vCPU tracking, the
module params are snapshotted at vCPU creation, i.e. applying a new
advancement effectively requires restarting a VM.
Dynamically updating a running vCPU is possible, e.g. a helper could be
added to retrieve the desired delay, choosing between the global module
param and the per-VCPU value depending on whether or not auto-tuning is
(globally) enabled, but introduces a great deal of complexity. The
wrapper itself is not complex, but understanding and documenting the
effects of dynamically toggling auto-tuning and/or adjusting the timer
advancement is nigh impossible since the behavior would be dependent on
KVM's implementation as well as compiler optimizations. In other words,
providing stable behavior would require extremely careful consideration
now and in the future.
Given that the expected use of a manually-tuned timer advancement is to
"tune once, run many", use the vastly simpler approach of recognizing
changes to the module params only when creating a new vCPU.
Cc: Liran Alon <liran.alon@oracle.com>
Cc: Wanpeng Li <wanpengli@tencent.com>
Reviewed-by: Liran Alon <liran.alon@oracle.com>
Cc: stable@vger.kernel.org
Fixes: 3b8a5df6c4 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Automatically adjusting the globally-shared timer advancement could
corrupt the timer, e.g. if multiple vCPUs are concurrently adjusting
the advancement value. That could be partially fixed by using a local
variable for the arithmetic, but it would still be susceptible to a
race when setting timer_advance_adjust_done.
And because virtual_tsc_khz and tsc_scaling_ratio are per-vCPU, the
correct calibration for a given vCPU may not apply to all vCPUs.
Furthermore, lapic_timer_advance_ns is marked __read_mostly, which is
effectively violated when finding a stable advancement takes an extended
amount of timer.
Opportunistically change the definition of lapic_timer_advance_ns to
a u32 so that it matches the style of struct kvm_timer. Explicitly
pass the param to kvm_create_lapic() so that it doesn't have to be
exposed to lapic.c, thus reducing the probability of unintentionally
using the global value instead of the per-vCPU value.
Cc: Liran Alon <liran.alon@oracle.com>
Cc: Wanpeng Li <wanpengli@tencent.com>
Reviewed-by: Liran Alon <liran.alon@oracle.com>
Cc: stable@vger.kernel.org
Fixes: 3b8a5df6c4 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
To minimize the latency of timer interrupts as observed by the guest,
KVM adjusts the values it programs into the host timers to account for
the host's overhead of programming and handling the timer event. Now
that the timer advancement is automatically tuned during runtime, it's
effectively unbounded by default, e.g. if KVM is running as L1 the
advancement can measure in hundreds of milliseconds.
Disable timer advancement if adaptive tuning yields an advancement of
more than 5000ns, as large advancements can break reasonable assumptions
of the guest, e.g. that a timer configured to fire after 1ms won't
arrive on the next instruction. Although KVM busy waits to mitigate the
case of a timer event arriving too early, complications can arise when
shifting the interrupt too far, e.g. kvm-unit-test's vmx.interrupt test
will fail when its "host" exits on interrupts as KVM may inject the INTR
before the guest executes STI+HLT. Arguably the unit test is "broken"
in the sense that delaying a timer interrupt by 1ms doesn't technically
guarantee the interrupt will arrive after STI+HLT, but it's a reasonable
assumption that KVM should support.
Furthermore, an unbounded advancement also effectively unbounds the time
spent busy waiting, e.g. if the guest programs a timer with a very large
delay.
5000ns is a somewhat arbitrary threshold. When running on bare metal,
which is the intended use case, timer advancement is expected to be in
the general vicinity of 1000ns. 5000ns is high enough that false
positives are unlikely, while not being so high as to negatively affect
the host's performance/stability.
Note, a future patch will enable userspace to disable KVM's adaptive
tuning, which will allow priveleged userspace will to specifying an
advancement value in excess of this arbitrary threshold in order to
satisfy an abnormal use case.
Cc: Liran Alon <liran.alon@oracle.com>
Cc: Wanpeng Li <wanpengli@tencent.com>
Cc: stable@vger.kernel.org
Fixes: 3b8a5df6c4 ("KVM: LAPIC: Tune lapic_timer_advance_ns automatically")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
It was reported that with some special Multi Processor Group configuration,
e.g:
bcdedit.exe /set groupsize 1
bcdedit.exe /set maxgroup on
bcdedit.exe /set groupaware on
for a 16-vCPU guest WS2012 shows BSOD on boot when PV TLB flush mechanism
is in use.
Tracing kvm_hv_flush_tlb immediately reveals the issue:
kvm_hv_flush_tlb: processor_mask 0x0 address_space 0x0 flags 0x2
The only flag set in this request is HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES,
however, processor_mask is 0x0 and no HV_FLUSH_ALL_PROCESSORS is specified.
We don't flush anything and apparently it's not what Windows expects.
TLFS doesn't say anything about such requests and newer Windows versions
seem to be unaffected. This all feels like a WS2012 bug, which is, however,
easy to workaround in KVM: let's flush everything when we see an empty
flush request, over-flushing doesn't hurt.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If guest sets MSR_IA32_TSCDEADLINE to value such that in host
time-domain it's shorter than lapic_timer_advance_ns, we can
reach a case that we call hrtimer_start() with expiration time set at
the past.
Because lapic_timer.timer is init with HRTIMER_MODE_ABS_PINNED, it
is not allowed to run in softirq and therefore will never expire.
To avoid such a scenario, verify that deadline expiration time is set on
host time-domain further than (now + lapic_timer_advance_ns).
A future patch can also consider adding a min_timer_deadline_ns module parameter,
similar to min_timer_period_us to avoid races that amount of ns it takes
to run logic could still call hrtimer_start() with expiration timer set
at the past.
Reviewed-by: Joao Martins <joao.m.martins@oracle.com>
Signed-off-by: Liran Alon <liran.alon@oracle.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In __apic_accept_irq() interface trig_mode is int and actually on some code
paths it is set above u8:
kvm_apic_set_irq() extracts it from 'struct kvm_lapic_irq' where trig_mode
is u16. This is done on purpose as e.g. kvm_set_msi_irq() sets it to
(1 << 15) & e->msi.data
kvm_apic_local_deliver sets it to reg & (1 << 15).
Fix the immediate issue by making 'tm' into u16. We may also want to adjust
__apic_accept_irq() interface and use proper sizes for vector, level,
trig_mode but this is not urgent.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Changed passing argument as "0 to NULL" which resolves below sparse warning
arch/x86/kvm/x86.c:3096:61: warning: Using plain integer as NULL pointer
Signed-off-by: Hariprasad Kelam <hariprasad.kelam@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Invoking the 64-bit variation on a 32-bit kenrel will crash the guest,
trigger a WARN, and/or lead to a buffer overrun in the host, e.g.
rsm_load_state_64() writes r8-r15 unconditionally, but enum kvm_reg and
thus x86_emulate_ctxt._regs only define r8-r15 for CONFIG_X86_64.
KVM allows userspace to report long mode support via CPUID, even though
the guest is all but guaranteed to crash if it actually tries to enable
long mode. But, a pure 32-bit guest that is ignorant of long mode will
happily plod along.
SMM complicates things as 64-bit CPUs use a different SMRAM save state
area. KVM handles this correctly for 64-bit kernels, e.g. uses the
legacy save state map if userspace has hid long mode from the guest,
but doesn't fare well when userspace reports long mode support on a
32-bit host kernel (32-bit KVM doesn't support 64-bit guests).
Since the alternative is to crash the guest, e.g. by not loading state
or explicitly requesting shutdown, unconditionally use the legacy SMRAM
save state map for 32-bit KVM. If a guest has managed to get far enough
to handle SMIs when running under a weird/buggy userspace hypervisor,
then don't deliberately crash the guest since there are no downsides
(from KVM's perspective) to allow it to continue running.
Fixes: 660a5d517a ("KVM: x86: save/load state on SMM switch")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Neither AMD nor Intel CPUs have an EFER field in the legacy SMRAM save
state area, i.e. don't save/restore EFER across SMM transitions. KVM
somewhat models this, e.g. doesn't clear EFER on entry to SMM if the
guest doesn't support long mode. But during RSM, KVM unconditionally
clears EFER so that it can get back to pure 32-bit mode in order to
start loading CRs with their actual non-SMM values.
Clear EFER only when it will be written when loading the non-SMM state
so as to preserve bits that can theoretically be set on 32-bit vCPUs,
e.g. KVM always emulates EFER_SCE.
And because CR4.PAE is cleared only to play nice with EFER, wrap that
code in the long mode check as well. Note, this may result in a
compiler warning about cr4 being consumed uninitialized. Re-read CR4
even though it's technically unnecessary, as doing so allows for more
readable code and RSM emulation is not a performance critical path.
Fixes: 660a5d517a ("KVM: x86: save/load state on SMM switch")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
RSM emulation is currently broken on VMX when the interrupted guest has
CR4.VMXE=1. Stop dancing around the issue of HF_SMM_MASK being set when
loading SMSTATE into architectural state, e.g. by toggling it for
problematic flows, and simply clear HF_SMM_MASK prior to loading
architectural state (from SMRAM save state area).
Reported-by: Jon Doron <arilou@gmail.com>
Cc: Jim Mattson <jmattson@google.com>
Cc: Liran Alon <liran.alon@oracle.com>
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Fixes: 5bea5123cb ("KVM: VMX: check nested state and CR4.VMXE against SMM")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Tested-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Prepare for clearing HF_SMM_MASK prior to loading state from the SMRAM
save state map, i.e. kvm_smm_changed() needs to be called after state
has been loaded and so cannot be done automatically when setting
hflags from RSM.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
RSM emulation is currently broken on VMX when the interrupted guest has
CR4.VMXE=1. Rather than dance around the issue of HF_SMM_MASK being set
when loading SMSTATE into architectural state, ideally RSM emulation
itself would be reworked to clear HF_SMM_MASK prior to loading non-SMM
architectural state.
Ostensibly, the only motivation for having HF_SMM_MASK set throughout
the loading of state from the SMRAM save state area is so that the
memory accesses from GET_SMSTATE() are tagged with role.smm. Load
all of the SMRAM save state area from guest memory at the beginning of
RSM emulation, and load state from the buffer instead of reading guest
memory one-by-one.
This paves the way for clearing HF_SMM_MASK prior to loading state,
and also aligns RSM with the enter_smm() behavior, which fills a
buffer and writes SMRAM save state in a single go.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Issue was discovered when running kvm-unit-tests on KVM running as L1 on
top of Hyper-V.
When vmx_instruction_intercept unit-test attempts to run RDPMC to test
RDPMC-exiting, it is intercepted by L1 KVM which it's EXIT_REASON_RDPMC
handler raise #GP because vCPU exposed by Hyper-V doesn't support PMU.
Instead of unit-test expectation to be reflected with EXIT_REASON_RDPMC.
The reason vmx_instruction_intercept unit-test attempts to run RDPMC
even though Hyper-V doesn't support PMU is because L1 expose to L2
support for RDPMC-exiting. Which is reasonable to assume that is
supported only in case CPU supports PMU to being with.
Above issue can easily be simulated by modifying
vmx_instruction_intercept config in x86/unittests.cfg to run QEMU with
"-cpu host,+vmx,-pmu" and run unit-test.
To handle issue, change KVM to expose RDPMC-exiting only when guest
supports PMU.
Reported-by: Saar Amar <saaramar@microsoft.com>
Reviewed-by: Mihai Carabas <mihai.carabas@oracle.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Liran Alon <liran.alon@oracle.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Before this change, reading a VMware pseduo PMC will succeed even when
PMU is not supported by guest. This can easily be seen by running
kvm-unit-test vmware_backdoors with "-cpu host,-pmu" option.
Reviewed-by: Mihai Carabas <mihai.carabas@oracle.com>
Signed-off-by: Liran Alon <liran.alon@oracle.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
guest xcr0 could leak into host when MCE happens in guest mode. Because
do_machine_check() could schedule out at a few places.
For example:
kvm_load_guest_xcr0
...
kvm_x86_ops->run(vcpu) {
vmx_vcpu_run
vmx_complete_atomic_exit
kvm_machine_check
do_machine_check
do_memory_failure
memory_failure
lock_page
In this case, host_xcr0 is 0x2ff, guest vcpu xcr0 is 0xff. After schedule
out, host cpu has guest xcr0 loaded (0xff).
In __switch_to {
switch_fpu_finish
copy_kernel_to_fpregs
XRSTORS
If any bit i in XSTATE_BV[i] == 1 and xcr0[i] == 0, XRSTORS will
generate #GP (In this case, bit 9). Then ex_handler_fprestore kicks in
and tries to reinitialize fpu by restoring init fpu state. Same story as
last #GP, except we get DOUBLE FAULT this time.
Cc: stable@vger.kernel.org
Signed-off-by: WANG Chao <chao.wang@ucloud.cn>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
I noticed that apic test from kvm-unit-tests always hangs on my EPYC 7401P,
the hanging test nmi-after-sti is trying to deliver 30000 NMIs and tracing
shows that we're sometimes able to deliver a few but never all.
When we're trying to inject an NMI we may fail to do so immediately for
various reasons, however, we still need to inject it so enable_nmi_window()
arms nmi_singlestep mode. #DB occurs as expected, but we're not checking
for pending NMIs before entering the guest and unless there's a different
event to process, the NMI will never get delivered.
Make KVM_REQ_EVENT request on the vCPU from db_interception() to make sure
pending NMIs are checked and possibly injected.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Only clear the valid bit when invalidate logical APIC id entry.
The current logic clear the valid bit, but also set the rest of
the bits (including reserved bits) to 1.
Fixes: 98d90582be ('svm: Fix AVIC DFR and LDR handling')
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This reverts commit bb218fbcfa.
As Oren Twaig pointed out the old discussion:
https://patchwork.kernel.org/patch/8292231/
that the change coud potentially cause an extra IPI to be sent to
the destination vcpu because the AVIC hardware already set the IRR bit
before the incomplete IPI #VMEXIT with id=1 (target vcpu is not running).
Since writting to ICR and ICR2 will also set the IRR. If something triggers
the destination vcpu to get scheduled before the emulation finishes, then
this could result in an additional IPI.
Also, the issue mentioned in the commit bb218fbcfa was misdiagnosed.
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Reported-by: Oren Twaig <oren@scalemp.com>
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM bases its memory usage limits on the total number of guest pages
across all memslots. However, those limits, and the calculations to
produce them, use 32 bit unsigned integers. This can result in overflow
if a VM has more guest pages that can be represented by a u32. As a
result of this overflow, KVM can use a low limit on the number of MMU
pages it will allocate. This makes KVM unable to map all of guest memory
at once, prompting spurious faults.
Tested: Ran all kvm-unit-tests on an Intel Haswell machine. This patch
introduced no new failures.
Signed-off-by: Ben Gardon <bgardon@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The remaining failures of vmx.flat when EPT is disabled are caused by
incorrectly reflecting VMfails to the L1 hypervisor. What happens is
that nested_vmx_restore_host_state corrupts the guest CR3, reloading it
with the host's shadow CR3 instead, because it blindly loads GUEST_CR3
from the vmcs01.
For simplicity let's just always use hardware VMCS checks when EPT is
disabled. This way, nested_vmx_restore_host_state is not reached at
all (or at least shouldn't be reached).
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
As mentioned in the comment, there are some special cases where we can simply
clear the TPR shadow bit from the CPU-based execution controls in the vmcs02.
Handle them so that we can remove some XFAILs from vmx.flat.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
-----BEGIN PGP SIGNATURE-----
iHUEABYIAB0WIQRTLbB6QfY48x44uB6AXGG7T9hjvgUCXKoNnwAKCRCAXGG7T9hj
vqpEAQCMeiLXXp+BMGI1+x1eeE4ri2woGkK1lsZJLOJhGIqTfgD/dDvmhCSQBDAs
IbDDbNJP1IT4jQ98c5obw+qEt9OWcww=
=J7ME
-----END PGP SIGNATURE-----
Merge tag 'for-linus-5.1b-rc4-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/xen/tip
Pull xen fixes from Juergen Gross:
"One minor fix and a small cleanup for the xen privcmd driver"
* tag 'for-linus-5.1b-rc4-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/xen/tip:
xen: Prevent buffer overflow in privcmd ioctl
xen: use struct_size() helper in kzalloc()
Referring to the "VIRTUALIZING MSR-BASED APIC ACCESSES" chapter of the
SDM, when "virtualize x2APIC mode" is 1 and "APIC-register
virtualization" is 0, a RDMSR of 808H should return the VTPR from the
virtual APIC page.
However, for nested, KVM currently fails to disable the read intercept
for this MSR. This means that a RDMSR exit takes precedence over
"virtualize x2APIC mode", and KVM passes through L1's TPR to L2,
instead of sourcing the value from L2's virtual APIC page.
This patch fixes the issue by disabling the read intercept, in VMCS02,
for the VTPR when "APIC-register virtualization" is 0.
The issue described above and fix prescribed here, were verified with
a related patch in kvm-unit-tests titled "Test VMX's virtualize x2APIC
mode w/ nested".
Signed-off-by: Marc Orr <marcorr@google.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Fixes: c992384bde ("KVM: vmx: speed up MSR bitmap merge")
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The nested_vmx_prepare_msr_bitmap() function doesn't directly guard the
x2APIC MSR intercepts with the "virtualize x2APIC mode" MSR. As a
result, we discovered the potential for a buggy or malicious L1 to get
access to L0's x2APIC MSRs, via an L2, as follows.
1. L1 executes WRMSR(IA32_SPEC_CTRL, 1). This causes the spec_ctrl
variable, in nested_vmx_prepare_msr_bitmap() to become true.
2. L1 disables "virtualize x2APIC mode" in VMCS12.
3. L1 enables "APIC-register virtualization" in VMCS12.
Now, KVM will set VMCS02's x2APIC MSR intercepts from VMCS12, and then
set "virtualize x2APIC mode" to 0 in VMCS02. Oops.
This patch closes the leak by explicitly guarding VMCS02's x2APIC MSR
intercepts with VMCS12's "virtualize x2APIC mode" control.
The scenario outlined above and fix prescribed here, were verified with
a related patch in kvm-unit-tests titled "Add leak scenario to
virt_x2apic_mode_test".
Note, it looks like this issue may have been introduced inadvertently
during a merge---see 15303ba5d1.
Signed-off-by: Marc Orr <marcorr@google.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This ensures that the address and length provided to DBG_DECRYPT and
DBG_ENCRYPT do not cause an overflow.
At the same time, pass the actual number of pages pinned in memory to
sev_unpin_memory() as a cleanup.
Reported-by: Cfir Cohen <cfir@google.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
get_num_contig_pages() could potentially overflow int so make its type
consistent with its usage.
Reported-by: Cfir Cohen <cfir@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The "call" variable comes from the user in privcmd_ioctl_hypercall().
It's an offset into the hypercall_page[] which has (PAGE_SIZE / 32)
elements. We need to put an upper bound on it to prevent an out of
bounds access.
Cc: stable@vger.kernel.org
Fixes: 1246ae0bb9 ("xen: add variable hypercall caller")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Signed-off-by: Juergen Gross <jgross@suse.com>
On top of this, a cleanup of kvm_para.h headers, which were exported by
some architectures even though they not support KVM at all. This is
responsible for all the Kbuild changes in the diffstat.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.22 (GNU/Linux)
iQEcBAABAgAGBQJcoM5VAAoJEL/70l94x66DU3EH/A8sYdsfeqALWElm2Sy9TYas
mntz+oTWsl3vDy8s8zp1ET2NpF7oBlBEMmCWhVEJaD+1qW3VpTRAseR3Zr9ML9xD
k+BQM8SKv47o86ZN+y4XALl30Ckb3DXh/X1xsrV5hF6J3ofC+Ce2tF560l8C9ygC
WyHDxwNHMWVA/6TyW3mhunzuVKgZ/JND9+0zlyY1LKmUQ0BQLle23gseIhhI0YDm
B4VGIYU2Mf8jCH5Ir3N/rQ8pLdo8U7f5P/MMfgXQafksvUHJBg6B6vOhLJh94dLh
J2wixYp1zlT0drBBkvJ0jPZ75skooWWj0o3otEA7GNk/hRj6MTllgfL5SajTHZg=
=/A7u
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
Pull KVM fixes from Paolo Bonzini:
"A collection of x86 and ARM bugfixes, and some improvements to
documentation.
On top of this, a cleanup of kvm_para.h headers, which were exported
by some architectures even though they not support KVM at all. This is
responsible for all the Kbuild changes in the diffstat"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (28 commits)
Documentation: kvm: clarify KVM_SET_USER_MEMORY_REGION
KVM: doc: Document the life cycle of a VM and its resources
KVM: selftests: complete IO before migrating guest state
KVM: selftests: disable stack protector for all KVM tests
KVM: selftests: explicitly disable PIE for tests
KVM: selftests: assert on exit reason in CR4/cpuid sync test
KVM: x86: update %rip after emulating IO
x86/kvm/hyper-v: avoid spurious pending stimer on vCPU init
kvm/x86: Move MSR_IA32_ARCH_CAPABILITIES to array emulated_msrs
KVM: x86: Emulate MSR_IA32_ARCH_CAPABILITIES on AMD hosts
kvm: don't redefine flags as something else
kvm: mmu: Used range based flushing in slot_handle_level_range
KVM: export <linux/kvm_para.h> and <asm/kvm_para.h> iif KVM is supported
KVM: x86: remove check on nr_mmu_pages in kvm_arch_commit_memory_region()
kvm: nVMX: Add a vmentry check for HOST_SYSENTER_ESP and HOST_SYSENTER_EIP fields
KVM: SVM: Workaround errata#1096 (insn_len maybe zero on SMAP violation)
KVM: Reject device ioctls from processes other than the VM's creator
KVM: doc: Fix incorrect word ordering regarding supported use of APIs
KVM: x86: fix handling of role.cr4_pae and rename it to 'gpte_size'
KVM: nVMX: Do not inherit quadrant and invalid for the root shadow EPT
...
Pull x86 fixes from Thomas Gleixner:
"A pile of x86 updates:
- Prevent exceeding he valid physical address space in the /dev/mem
limit checks.
- Move all header content inside the header guard to prevent compile
failures.
- Fix the bogus __percpu annotation in this_cpu_has() which makes
sparse very noisy.
- Disable switch jump tables completely when retpolines are enabled.
- Prevent leaking the trampoline address"
* 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/realmode: Make set_real_mode_mem() static inline
x86/cpufeature: Fix __percpu annotation in this_cpu_has()
x86/mm: Don't exceed the valid physical address space
x86/retpolines: Disable switch jump tables when retpolines are enabled
x86/realmode: Don't leak the trampoline kernel address
x86/boot: Fix incorrect ifdeffery scope
x86/resctrl: Remove unused variable
Most (all?) x86 platforms provide a port IO based reset mechanism, e.g.
OUT 92h or CF9h. Userspace may emulate said mechanism, i.e. reset a
vCPU in response to KVM_EXIT_IO, without explicitly announcing to KVM
that it is doing a reset, e.g. Qemu jams vCPU state and resumes running.
To avoid corruping %rip after such a reset, commit 0967b7bf1c ("KVM:
Skip pio instruction when it is emulated, not executed") changed the
behavior of PIO handlers, i.e. today's "fast" PIO handling to skip the
instruction prior to exiting to userspace. Full emulation doesn't need
such tricks becase re-emulating the instruction will naturally handle
%rip being changed to point at the reset vector.
Updating %rip prior to executing to userspace has several drawbacks:
- Userspace sees the wrong %rip on the exit, e.g. if PIO emulation
fails it will likely yell about the wrong address.
- Single step exits to userspace for are effectively dropped as
KVM_EXIT_DEBUG is overwritten with KVM_EXIT_IO.
- Behavior of PIO emulation is different depending on whether it
goes down the fast path or the slow path.
Rather than skip the PIO instruction before exiting to userspace,
snapshot the linear %rip and cancel PIO completion if the current
value does not match the snapshot. For a 64-bit vCPU, i.e. the most
common scenario, the snapshot and comparison has negligible overhead
as VMCS.GUEST_RIP will be cached regardless, i.e. there is no extra
VMREAD in this case.
All other alternatives to snapshotting the linear %rip that don't
rely on an explicit reset announcenment suffer from one corner case
or another. For example, canceling PIO completion on any write to
%rip fails if userspace does a save/restore of %rip, and attempting to
avoid that issue by canceling PIO only if %rip changed then fails if PIO
collides with the reset %rip. Attempting to zero in on the exact reset
vector won't work for APs, which means adding more hooks such as the
vCPU's MP_STATE, and so on and so forth.
Checking for a linear %rip match technically suffers from corner cases,
e.g. userspace could theoretically rewrite the underlying code page and
expect a different instruction to execute, or the guest hardcodes a PIO
reset at 0xfffffff0, but those are far, far outside of what can be
considered normal operation.
Fixes: 432baf60ee ("KVM: VMX: use kvm_fast_pio_in for handling IN I/O")
Cc: <stable@vger.kernel.org>
Reported-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When userspace initializes guest vCPUs it may want to zero all supported
MSRs including Hyper-V related ones including HV_X64_MSR_STIMERn_CONFIG/
HV_X64_MSR_STIMERn_COUNT. With commit f3b138c5d8 ("kvm/x86: Update SynIC
timers on guest entry only") we began doing stimer_mark_pending()
unconditionally on every config change.
The issue I'm observing manifests itself as following:
- Qemu writes 0 to STIMERn_{CONFIG,COUNT} MSRs and marks all stimers as
pending in stimer_pending_bitmap, arms KVM_REQ_HV_STIMER;
- kvm_hv_has_stimer_pending() starts returning true;
- kvm_vcpu_has_events() starts returning true;
- kvm_arch_vcpu_runnable() starts returning true;
- when kvm_arch_vcpu_ioctl_run() gets into
(vcpu->arch.mp_state == KVM_MP_STATE_UNINITIALIZED) case:
- kvm_vcpu_block() gets in 'kvm_vcpu_check_block(vcpu) < 0' and returns
immediately, avoiding normal wait path;
- -EAGAIN is returned from kvm_arch_vcpu_ioctl_run() immediately forcing
userspace to retry.
So instead of normal wait path we get a busy loop on all secondary vCPUs
before they get INIT signal. This seems to be undesirable, especially given
that this happens even when Hyper-V extensions are not used.
Generally, it seems to be pointless to mark an stimer as pending in
stimer_pending_bitmap and arm KVM_REQ_HV_STIMER as the only thing
kvm_hv_process_stimers() will do is clear the corresponding bit. We may
just not mark disabled timers as pending instead.
Fixes: f3b138c5d8 ("kvm/x86: Update SynIC timers on guest entry only")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Since MSR_IA32_ARCH_CAPABILITIES is emualted unconditionally even if
host doesn't suppot it. We should move it to array emulated_msrs from
arry msrs_to_save, to report to userspace that guest support this msr.
Signed-off-by: Xiaoyao Li <xiaoyao.li@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The CPUID flag ARCH_CAPABILITIES is unconditioinally exposed to host
userspace for all x86 hosts, i.e. KVM advertises ARCH_CAPABILITIES
regardless of hardware support under the pretense that KVM fully
emulates MSR_IA32_ARCH_CAPABILITIES. Unfortunately, only VMX hosts
handle accesses to MSR_IA32_ARCH_CAPABILITIES (despite KVM_GET_MSRS
also reporting MSR_IA32_ARCH_CAPABILITIES for all hosts).
Move the MSR_IA32_ARCH_CAPABILITIES handling to common x86 code so
that it's emulated on AMD hosts.
Fixes: 1eaafe91a0 ("kvm: x86: IA32_ARCH_CAPABILITIES is always supported")
Cc: stable@vger.kernel.org
Reported-by: Xiaoyao Li <xiaoyao.li@linux.intel.com>
Cc: Jim Mattson <jmattson@google.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Replace kvm_flush_remote_tlbs with kvm_flush_remote_tlbs_with_address
in slot_handle_level_range. When range based flushes are not enabled
kvm_flush_remote_tlbs_with_address falls back to kvm_flush_remote_tlbs.
This changes the behavior of many functions that indirectly use
slot_handle_level_range, iff the range based flushes are enabled. The
only potential problem I see with this is that kvm->tlbs_dirty will be
cleared less often, however the only caller of slot_handle_level_range that
checks tlbs_dirty is kvm_mmu_notifier_invalidate_range_start which
checks it and does a kvm_flush_remote_tlbs after calling
kvm_unmap_hva_range anyway.
Tested: Ran all kvm-unit-tests on a Intel Haswell machine with and
without this patch. The patch introduced no new failures.
Signed-off-by: Ben Gardon <bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
* nr_mmu_pages would be non-zero only if kvm->arch.n_requested_mmu_pages is
non-zero.
* nr_mmu_pages is always non-zero, since kvm_mmu_calculate_mmu_pages()
never return zero.
Based on these two reasons, we can merge the two *if* clause and use the
return value from kvm_mmu_calculate_mmu_pages() directly. This simplify
the code and also eliminate the possibility for reader to believe
nr_mmu_pages would be zero.
Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
According to section "Checks on VMX Controls" in Intel SDM vol 3C, the
following check is performed on vmentry of L2 guests:
On processors that support Intel 64 architecture, the IA32_SYSENTER_ESP
field and the IA32_SYSENTER_EIP field must each contain a canonical
address.
Signed-off-by: Krish Sadhukhan <krish.sadhukhan@oracle.com>
Reviewed-by: Mihai Carabas <mihai.carabas@oracle.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Errata#1096:
On a nested data page fault when CR.SMAP=1 and the guest data read
generates a SMAP violation, GuestInstrBytes field of the VMCB on a
VMEXIT will incorrectly return 0h instead the correct guest
instruction bytes .
Recommend Workaround:
To determine what instruction the guest was executing the hypervisor
will have to decode the instruction at the instruction pointer.
The recommended workaround can not be implemented for the SEV
guest because guest memory is encrypted with the guest specific key,
and instruction decoder will not be able to decode the instruction
bytes. If we hit this errata in the SEV guest then log the message
and request a guest shutdown.
Reported-by: Venkatesh Srinivas <venkateshs@google.com>
Cc: Jim Mattson <jmattson@google.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Joerg Roedel <joro@8bytes.org>
Cc: "Radim Krčmář" <rkrcmar@redhat.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The cr4_pae flag is a bit of a misnomer, its purpose is really to track
whether the guest PTE that is being shadowed is a 4-byte entry or an
8-byte entry. Prior to supporting nested EPT, the size of the gpte was
reflected purely by CR4.PAE. KVM fudged things a bit for direct sptes,
but it was mostly harmless since the size of the gpte never mattered.
Now that a spte may be tracking an indirect EPT entry, relying on
CR4.PAE is wrong and ill-named.
For direct shadow pages, force the gpte_size to '1' as they are always
8-byte entries; EPT entries can only be 8-bytes and KVM always uses
8-byte entries for NPT and its identity map (when running with EPT but
not unrestricted guest).
Likewise, nested EPT entries are always 8-bytes. Nested EPT presents a
unique scenario as the size of the entries are not dictated by CR4.PAE,
but neither is the shadow page a direct map. To handle this scenario,
set cr0_wp=1 and smap_andnot_wp=1, an otherwise impossible combination,
to denote a nested EPT shadow page. Use the information to avoid
incorrectly zapping an unsync'd indirect page in __kvm_sync_page().
Providing a consistent and accurate gpte_size fixes a bug reported by
Vitaly where fast_cr3_switch() always fails when switching from L2 to
L1 as kvm_mmu_get_page() would force role.cr4_pae=0 for direct pages,
whereas kvm_calc_mmu_role_common() would set it according to CR4.PAE.
Fixes: 7dcd575520 ("x86/kvm/mmu: check if tdp/shadow MMU reconfiguration is needed")
Reported-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Tested-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Explicitly zero out quadrant and invalid instead of inheriting them from
the root_mmu. Functionally, this patch is a nop as we (should) never
set quadrant for a direct mapped (EPT) root_mmu and nested EPT is only
allowed if EPT is used for L1, and the root_mmu will never be invalid at
this point.
Explicitly setting flags sets the stage for repurposing the legacy
paging bits in role, e.g. nxe, cr0_wp, and sm{a,e}p_andnot_wp, at which
point 'smm' would be the only flag to be inherited from root_mmu.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
&cpu_info.x86_capability is __percpu, and the second argument of
x86_this_cpu_test_bit() is expected to be __percpu. Don't cast the
__percpu away and then implicitly add it again. This gets rid of 106
lines of sparse warnings with the kernel config I'm using.
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Masahiro Yamada <yamada.masahiro@socionext.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: x86-ml <x86@kernel.org>
Link: https://lkml.kernel.org/r/20190328154948.152273-1-jannh@google.com
valid_phys_addr_range() is used to sanity check the physical address range
of an operation, e.g., access to /dev/mem. It uses __pa(high_memory)
internally.
If memory is populated at the end of the physical address space, then
__pa(high_memory) is outside of the physical address space because:
high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
For the comparison in valid_phys_addr_range() this is not an issue, but if
CONFIG_DEBUG_VIRTUAL is enabled, __pa() maps to __phys_addr(), which
verifies that the resulting physical address is within the valid physical
address space of the CPU. So in the case that memory is populated at the
end of the physical address space, this is not true and triggers a
VIRTUAL_BUG_ON().
Use __pa(high_memory - 1) to prevent the conversion from going beyond
the end of valid physical addresses.
Fixes: be62a32044 ("x86/mm: Limit mmap() of /dev/mem to valid physical addresses")
Signed-off-by: Ralph Campbell <rcampbell@nvidia.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Craig Bergstrom <craigb@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Fengguang Wu <fengguang.wu@intel.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Hans Verkuil <hans.verkuil@cisco.com>
Cc: Mauro Carvalho Chehab <mchehab@s-opensource.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sander Eikelenboom <linux@eikelenboom.it>
Cc: Sean Young <sean@mess.org>
Link: https://lkml.kernel.org/r/20190326001817.15413-2-rcampbell@nvidia.com
