This all started about 6 month ago with the attempt to move the Posix CPU
timer heavy lifting out of the timer interrupt code and just have lockless
quick checks in that code path. Trivial 5 patches.
This unearthed an inconsistency in the KVM handling of task work and the
review requested to move all of this into generic code so other
architectures can share.
Valid request and solved with another 25 patches but those unearthed
inconsistencies vs. RCU and instrumentation.
Digging into this made it obvious that there are quite some inconsistencies
vs. instrumentation in general. The int3 text poke handling in particular
was completely unprotected and with the batched update of trace events even
more likely to expose to endless int3 recursion.
In parallel the RCU implications of instrumenting fragile entry code came
up in several discussions.
The conclusion of the X86 maintainer team was to go all the way and make
the protection against any form of instrumentation of fragile and dangerous
code pathes enforcable and verifiable by tooling.
A first batch of preparatory work hit mainline with commit d5f744f9a2.
The (almost) full solution introduced a new code section '.noinstr.text'
into which all code which needs to be protected from instrumentation of all
sorts goes into. Any call into instrumentable code out of this section has
to be annotated. objtool has support to validate this. Kprobes now excludes
this section fully which also prevents BPF from fiddling with it and all
'noinstr' annotated functions also keep ftrace off. The section, kprobes
and objtool changes are already merged.
The major changes coming with this are:
- Preparatory cleanups
- Annotating of relevant functions to move them into the noinstr.text
section or enforcing inlining by marking them __always_inline so the
compiler cannot misplace or instrument them.
- Splitting and simplifying the idtentry macro maze so that it is now
clearly separated into simple exception entries and the more
interesting ones which use interrupt stacks and have the paranoid
handling vs. CR3 and GS.
- Move quite some of the low level ASM functionality into C code:
- enter_from and exit to user space handling. The ASM code now calls
into C after doing the really necessary ASM handling and the return
path goes back out without bells and whistels in ASM.
- exception entry/exit got the equivivalent treatment
- move all IRQ tracepoints from ASM to C so they can be placed as
appropriate which is especially important for the int3 recursion
issue.
- Consolidate the declaration and definition of entry points between 32
and 64 bit. They share a common header and macros now.
- Remove the extra device interrupt entry maze and just use the regular
exception entry code.
- All ASM entry points except NMI are now generated from the shared header
file and the corresponding macros in the 32 and 64 bit entry ASM.
- The C code entry points are consolidated as well with the help of
DEFINE_IDTENTRY*() macros. This allows to ensure at one central point
that all corresponding entry points share the same semantics. The
actual function body for most entry points is in an instrumentable
and sane state.
There are special macros for the more sensitive entry points,
e.g. INT3 and of course the nasty paranoid #NMI, #MCE, #DB and #DF.
They allow to put the whole entry instrumentation and RCU handling
into safe places instead of the previous pray that it is correct
approach.
- The INT3 text poke handling is now completely isolated and the
recursion issue banned. Aside of the entry rework this required other
isolation work, e.g. the ability to force inline bsearch.
- Prevent #DB on fragile entry code, entry relevant memory and disable
it on NMI, #MC entry, which allowed to get rid of the nested #DB IST
stack shifting hackery.
- A few other cleanups and enhancements which have been made possible
through this and already merged changes, e.g. consolidating and
further restricting the IDT code so the IDT table becomes RO after
init which removes yet another popular attack vector
- About 680 lines of ASM maze are gone.
There are a few open issues:
- An escape out of the noinstr section in the MCE handler which needs
some more thought but under the aspect that MCE is a complete
trainwreck by design and the propability to survive it is low, this was
not high on the priority list.
- Paravirtualization
When PV is enabled then objtool complains about a bunch of indirect
calls out of the noinstr section. There are a few straight forward
ways to fix this, but the other issues vs. general correctness were
more pressing than parawitz.
- KVM
KVM is inconsistent as well. Patches have been posted, but they have
not yet been commented on or picked up by the KVM folks.
- IDLE
Pretty much the same problems can be found in the low level idle code
especially the parts where RCU stopped watching. This was beyond the
scope of the more obvious and exposable problems and is on the todo
list.
The lesson learned from this brain melting exercise to morph the evolved
code base into something which can be validated and understood is that once
again the violation of the most important engineering principle
"correctness first" has caused quite a few people to spend valuable time on
problems which could have been avoided in the first place. The "features
first" tinkering mindset really has to stop.
With that I want to say thanks to everyone involved in contributing to this
effort. Special thanks go to the following people (alphabetical order):
Alexandre Chartre
Andy Lutomirski
Borislav Petkov
Brian Gerst
Frederic Weisbecker
Josh Poimboeuf
Juergen Gross
Lai Jiangshan
Macro Elver
Paolo Bonzini
Paul McKenney
Peter Zijlstra
Vitaly Kuznetsov
Will Deacon
-----BEGIN PGP SIGNATURE-----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=llba
-----END PGP SIGNATURE-----
Merge tag 'x86-entry-2020-06-12' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull x86 entry updates from Thomas Gleixner:
"The x86 entry, exception and interrupt code rework
This all started about 6 month ago with the attempt to move the Posix
CPU timer heavy lifting out of the timer interrupt code and just have
lockless quick checks in that code path. Trivial 5 patches.
This unearthed an inconsistency in the KVM handling of task work and
the review requested to move all of this into generic code so other
architectures can share.
Valid request and solved with another 25 patches but those unearthed
inconsistencies vs. RCU and instrumentation.
Digging into this made it obvious that there are quite some
inconsistencies vs. instrumentation in general. The int3 text poke
handling in particular was completely unprotected and with the batched
update of trace events even more likely to expose to endless int3
recursion.
In parallel the RCU implications of instrumenting fragile entry code
came up in several discussions.
The conclusion of the x86 maintainer team was to go all the way and
make the protection against any form of instrumentation of fragile and
dangerous code pathes enforcable and verifiable by tooling.
A first batch of preparatory work hit mainline with commit
d5f744f9a2 ("Pull x86 entry code updates from Thomas Gleixner")
That (almost) full solution introduced a new code section
'.noinstr.text' into which all code which needs to be protected from
instrumentation of all sorts goes into. Any call into instrumentable
code out of this section has to be annotated. objtool has support to
validate this.
Kprobes now excludes this section fully which also prevents BPF from
fiddling with it and all 'noinstr' annotated functions also keep
ftrace off. The section, kprobes and objtool changes are already
merged.
The major changes coming with this are:
- Preparatory cleanups
- Annotating of relevant functions to move them into the
noinstr.text section or enforcing inlining by marking them
__always_inline so the compiler cannot misplace or instrument
them.
- Splitting and simplifying the idtentry macro maze so that it is
now clearly separated into simple exception entries and the more
interesting ones which use interrupt stacks and have the paranoid
handling vs. CR3 and GS.
- Move quite some of the low level ASM functionality into C code:
- enter_from and exit to user space handling. The ASM code now
calls into C after doing the really necessary ASM handling and
the return path goes back out without bells and whistels in
ASM.
- exception entry/exit got the equivivalent treatment
- move all IRQ tracepoints from ASM to C so they can be placed as
appropriate which is especially important for the int3
recursion issue.
- Consolidate the declaration and definition of entry points between
32 and 64 bit. They share a common header and macros now.
- Remove the extra device interrupt entry maze and just use the
regular exception entry code.
- All ASM entry points except NMI are now generated from the shared
header file and the corresponding macros in the 32 and 64 bit
entry ASM.
- The C code entry points are consolidated as well with the help of
DEFINE_IDTENTRY*() macros. This allows to ensure at one central
point that all corresponding entry points share the same
semantics. The actual function body for most entry points is in an
instrumentable and sane state.
There are special macros for the more sensitive entry points, e.g.
INT3 and of course the nasty paranoid #NMI, #MCE, #DB and #DF.
They allow to put the whole entry instrumentation and RCU handling
into safe places instead of the previous pray that it is correct
approach.
- The INT3 text poke handling is now completely isolated and the
recursion issue banned. Aside of the entry rework this required
other isolation work, e.g. the ability to force inline bsearch.
- Prevent #DB on fragile entry code, entry relevant memory and
disable it on NMI, #MC entry, which allowed to get rid of the
nested #DB IST stack shifting hackery.
- A few other cleanups and enhancements which have been made
possible through this and already merged changes, e.g.
consolidating and further restricting the IDT code so the IDT
table becomes RO after init which removes yet another popular
attack vector
- About 680 lines of ASM maze are gone.
There are a few open issues:
- An escape out of the noinstr section in the MCE handler which needs
some more thought but under the aspect that MCE is a complete
trainwreck by design and the propability to survive it is low, this
was not high on the priority list.
- Paravirtualization
When PV is enabled then objtool complains about a bunch of indirect
calls out of the noinstr section. There are a few straight forward
ways to fix this, but the other issues vs. general correctness were
more pressing than parawitz.
- KVM
KVM is inconsistent as well. Patches have been posted, but they
have not yet been commented on or picked up by the KVM folks.
- IDLE
Pretty much the same problems can be found in the low level idle
code especially the parts where RCU stopped watching. This was
beyond the scope of the more obvious and exposable problems and is
on the todo list.
The lesson learned from this brain melting exercise to morph the
evolved code base into something which can be validated and understood
is that once again the violation of the most important engineering
principle "correctness first" has caused quite a few people to spend
valuable time on problems which could have been avoided in the first
place. The "features first" tinkering mindset really has to stop.
With that I want to say thanks to everyone involved in contributing to
this effort. Special thanks go to the following people (alphabetical
order): Alexandre Chartre, Andy Lutomirski, Borislav Petkov, Brian
Gerst, Frederic Weisbecker, Josh Poimboeuf, Juergen Gross, Lai
Jiangshan, Macro Elver, Paolo Bonzin,i Paul McKenney, Peter Zijlstra,
Vitaly Kuznetsov, and Will Deacon"
* tag 'x86-entry-2020-06-12' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (142 commits)
x86/entry: Force rcu_irq_enter() when in idle task
x86/entry: Make NMI use IDTENTRY_RAW
x86/entry: Treat BUG/WARN as NMI-like entries
x86/entry: Unbreak __irqentry_text_start/end magic
x86/entry: __always_inline CR2 for noinstr
lockdep: __always_inline more for noinstr
x86/entry: Re-order #DB handler to avoid *SAN instrumentation
x86/entry: __always_inline arch_atomic_* for noinstr
x86/entry: __always_inline irqflags for noinstr
x86/entry: __always_inline debugreg for noinstr
x86/idt: Consolidate idt functionality
x86/idt: Cleanup trap_init()
x86/idt: Use proper constants for table size
x86/idt: Add comments about early #PF handling
x86/idt: Mark init only functions __init
x86/entry: Rename trace_hardirqs_off_prepare()
x86/entry: Clarify irq_{enter,exit}_rcu()
x86/entry: Remove DBn stacks
x86/entry: Remove debug IDT frobbing
x86/entry: Optimize local_db_save() for virt
...
Convert #MC to IDTENTRY_MCE:
- Implement the C entry points with DEFINE_IDTENTRY_MCE
- Emit the ASM stub with DECLARE_IDTENTRY_MCE
- Remove the ASM idtentry in 64bit
- Remove the open coded ASM entry code in 32bit
- Fixup the XEN/PV code
- Remove the old prototypes
- Remove the error code from *machine_check_vector() as
it is always 0 and not used by any of the functions
it can point to. Fixup all the functions as well.
No functional change.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Alexandre Chartre <alexandre.chartre@oracle.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Andy Lutomirski <luto@kernel.org>
Link: https://lkml.kernel.org/r/20200505135314.334980426@linutronix.de
Syzbot reports the following issue:
WARNING: CPU: 0 PID: 6819 at arch/x86/kvm/x86.c:618
kvm_inject_emulated_page_fault+0x210/0x290 arch/x86/kvm/x86.c:618
...
Call Trace:
...
RIP: 0010:kvm_inject_emulated_page_fault+0x210/0x290 arch/x86/kvm/x86.c:618
...
nested_vmx_get_vmptr+0x1f9/0x2a0 arch/x86/kvm/vmx/nested.c:4638
handle_vmon arch/x86/kvm/vmx/nested.c:4767 [inline]
handle_vmon+0x168/0x3a0 arch/x86/kvm/vmx/nested.c:4728
vmx_handle_exit+0x29c/0x1260 arch/x86/kvm/vmx/vmx.c:6067
'exception' we're trying to inject with kvm_inject_emulated_page_fault()
comes from:
nested_vmx_get_vmptr()
kvm_read_guest_virt()
kvm_read_guest_virt_helper()
vcpu->arch.walk_mmu->gva_to_gpa()
but it is only set when GVA to GPA conversion fails. In case it doesn't but
we still fail kvm_vcpu_read_guest_page(), X86EMUL_IO_NEEDED is returned and
nested_vmx_get_vmptr() calls kvm_inject_emulated_page_fault() with zeroed
'exception'. This happen when the argument is MMIO.
Paolo also noticed that nested_vmx_get_vmptr() is not the only place in
KVM code where kvm_read/write_guest_virt*() return result is mishandled.
VMX instructions along with INVPCID have the same issue. This was already
noticed before, e.g. see commit 541ab2aeb2 ("KVM: x86: work around
leak of uninitialized stack contents") but was never fully fixed.
KVM could've handled the request correctly by going to userspace and
performing I/O but there doesn't seem to be a good need for such requests
in the first place.
Introduce vmx_handle_memory_failure() as an interim solution.
Note, nested_vmx_get_vmptr() now has three possible outcomes: OK, PF,
KVM_EXIT_INTERNAL_ERROR and callers need to know if userspace exit is
needed (for KVM_EXIT_INTERNAL_ERROR) in case of failure. We don't seem
to have a good enum describing this tristate, just add "int *ret" to
nested_vmx_get_vmptr() interface to pass the information.
Reported-by: syzbot+2a7156e11dc199bdbd8a@syzkaller.appspotmail.com
Suggested-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20200605115906.532682-1-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Both Intel and AMD support (MPK) Memory Protection Key feature.
Move the feature detection from VMX to the common code. It should
work for both the platforms now.
Signed-off-by: Babu Moger <babu.moger@amd.com>
Message-Id: <158932795627.44260.15144185478040178638.stgit@naples-babu.amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Intel CPUs have a new alternative MSR range (starting from MSR_IA32_PMC0)
for GP counters that allows writing the full counter width. Enable this
range from a new capability bit (IA32_PERF_CAPABILITIES.FW_WRITE[bit 13]).
The guest would query CPUID to get the counter width, and sign extends
the counter values as needed. The traditional MSRs always limit to 32bit,
even though the counter internally is larger (48 or 57 bits).
When the new capability is set, use the alternative range which do not
have these restrictions. This lowers the overhead of perf stat slightly
because it has to do less interrupts to accumulate the counter value.
Signed-off-by: Like Xu <like.xu@linux.intel.com>
Message-Id: <20200529074347.124619-3-like.xu@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently, APF mechanism relies on the #PF abuse where the token is being
passed through CR2. If we switch to using interrupts to deliver page-ready
notifications we need a different way to pass the data. Extent the existing
'struct kvm_vcpu_pv_apf_data' with token information for page-ready
notifications.
While on it, rename 'reason' to 'flags'. This doesn't change the semantics
as we only have reasons '1' and '2' and these can be treated as bit flags
but KVM_PV_REASON_PAGE_READY is going away with interrupt based delivery
making 'reason' name misleading.
The newly introduced apf_put_user_ready() temporary puts both flags and
token information, this will be changed to put token only when we switch
to interrupt based notifications.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20200525144125.143875-3-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
vmx_load_mmu_pgd is delaying the write of GUEST_CR3 to prepare_vmcs02 as
an optimization, but this is only correct before the nested vmentry.
If userspace is modifying CR3 with KVM_SET_SREGS after the VM has
already been put in guest mode, the value of CR3 will not be updated.
Remove the optimization, which almost never triggers anyway.
Fixes: 04f11ef458 ("KVM: nVMX: Always write vmcs02.GUEST_CR3 during nested VM-Enter")
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In case an interrupt arrives after nested.check_events but before the
call to kvm_cpu_has_injectable_intr, we could end up enabling the interrupt
window even if the interrupt is actually going to be a vmexit. This is
useless rather than harmful, but it really complicates reasoning about
SVM's handling of the VINTR intercept. We'd like to never bother with
the VINTR intercept if V_INTR_MASKING=1 && INTERCEPT_INTR=1, because in
that case there is no interrupt window and we can just exit the nested
guest whenever we want.
This patch moves the opening of the interrupt window inside
inject_pending_event. This consolidates the check for pending
interrupt/NMI/SMI in one place, and makes KVM's usage of immediate
exits more consistent, extending it beyond just nested virtualization.
There are two functional changes here. They only affect corner cases,
but overall they simplify the inject_pending_event.
- re-injection of still-pending events will also use req_immediate_exit
instead of using interrupt-window intercepts. This should have no impact
on performance on Intel since it simply replaces an interrupt-window
or NMI-window exit for a preemption-timer exit. On AMD, which has no
equivalent of the preemption time, it may incur some overhead but an
actual effect on performance should only be visible in pathological cases.
- kvm_arch_interrupt_allowed and kvm_vcpu_has_events will return true
if an interrupt, NMI or SMI is blocked by nested_run_pending. This
makes sense because entering the VM will allow it to make progress
and deliver the event.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The second "/* fall through */" in rmode_exception() makes code harder to
read. Replace it with "return" to indicate they are different cases, only
the #DB and #BP check vcpu->guest_debug, while others don't care. And this
also improves the readability.
Suggested-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
Message-Id: <1582080348-20827-1-git-send-email-linmiaohe@huawei.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Take a u32 for the index in has_emulated_msr() to match hardware, which
treats MSR indices as unsigned 32-bit values. Functionally, taking a
signed int doesn't cause problems with the current code base, but could
theoretically cause problems with 32-bit KVM, e.g. if the index were
checked via a less-than statement, which would evaluate incorrectly for
MSR indices with bit 31 set.
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200218234012.7110-3-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We can simply look at bits 52-53 to identify MMIO entries in KVM's page
tables. Therefore, there is no need to pass a mask to kvm_mmu_set_mmio_spte_mask.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Even though we might not allow the guest to use WAITPKG's new
instructions, we should tell KVM that the feature is supported by the
host CPU.
Note that vmx_waitpkg_supported checks that WAITPKG _can_ be set in
secondary execution controls as specified by VMX capability MSR, rather
that we actually enable it for a guest.
Cc: stable@vger.kernel.org
Fixes: e69e72faa3 ("KVM: x86: Add support for user wait instructions")
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20200523161455.3940-2-mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The hrtimer used to emulate the VMX-preemption timer must be pinned to
the same logical processor as the vCPU thread to be interrupted if we
want to have any hope of adhering to the architectural specification
of the VMX-preemption timer. Even with this change, the emulated
VMX-preemption timer VM-exit occasionally arrives too late.
Signed-off-by: Jim Mattson <jmattson@google.com>
Reviewed-by: Peter Shier <pshier@google.com>
Reviewed-by: Oliver Upton <oupton@google.com>
Message-Id: <20200508203643.85477-4-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Remove a 'struct kvm_x86_ops' param that got left behind when the nested
ops were moved to their own struct.
Fixes: 33b2217245 ("KVM: x86: move nested-related kvm_x86_ops to a separate struct")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200506204653.14683-1-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This patch implements a fastpath for the preemption timer vmexit. The vmexit
can be handled quickly so it can be performed with interrupts off and going
back directly to the guest.
Testing on SKX Server.
cyclictest in guest(w/o mwait exposed, adaptive advance lapic timer is default -1):
5540.5ns -> 4602ns 17%
kvm-unit-test/vmexit.flat:
w/o avanced timer:
tscdeadline_immed: 3028.5 -> 2494.75 17.6%
tscdeadline: 5765.7 -> 5285 8.3%
w/ adaptive advance timer default -1:
tscdeadline_immed: 3123.75 -> 2583 17.3%
tscdeadline: 4663.75 -> 4537 2.7%
Tested-by: Haiwei Li <lihaiwei@tencent.com>
Cc: Haiwei Li <lihaiwei@tencent.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Message-Id: <1588055009-12677-8-git-send-email-wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Replace the ad hoc test in vmx_set_hv_timer with a test in the caller,
start_hv_timer. This test is not Intel-specific and would be duplicated
when introducing the fast path for the TSC deadline MSR.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
While optimizing posted-interrupt delivery especially for the timer
fastpath scenario, I measured kvm_x86_ops.deliver_posted_interrupt()
to introduce substantial latency because the processor has to perform
all vmentry tasks, ack the posted interrupt notification vector,
read the posted-interrupt descriptor etc.
This is not only slow, it is also unnecessary when delivering an
interrupt to the current CPU (as is the case for the LAPIC timer) because
PIR->IRR and IRR->RVI synchronization is already performed on vmentry
Therefore skip kvm_vcpu_trigger_posted_interrupt in this case, and
instead do vmx_sync_pir_to_irr() on the EXIT_FASTPATH_REENTER_GUEST
fastpath as well.
Tested-by: Haiwei Li <lihaiwei@tencent.com>
Cc: Haiwei Li <lihaiwei@tencent.com>
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Message-Id: <1588055009-12677-6-git-send-email-wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Adds a fastpath_t typedef since enum lines are a bit long, and replace
EXIT_FASTPATH_SKIP_EMUL_INS with two new exit_fastpath_completion enum values.
- EXIT_FASTPATH_EXIT_HANDLED kvm will still go through it's full run loop,
but it would skip invoking the exit handler.
- EXIT_FASTPATH_REENTER_GUEST complete fastpath, guest can be re-entered
without invoking the exit handler or going
back to vcpu_run
Tested-by: Haiwei Li <lihaiwei@tencent.com>
Cc: Haiwei Li <lihaiwei@tencent.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Message-Id: <1588055009-12677-4-git-send-email-wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Introduce generic fastpath handler to handle MSR fastpath, VMX-preemption
timer fastpath etc; move it after vmx_complete_interrupts() in order to
catch events delivered to the guest, and abort the fast path in later
patches. While at it, move the kvm_exit tracepoint so that it is printed
for fastpath vmexits as well.
There is no observed performance effect for the IPI fastpath after this patch.
Tested-by: Haiwei Li <lihaiwei@tencent.com>
Cc: Haiwei Li <lihaiwei@tencent.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Suggested-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <1588055009-12677-2-git-send-email-wanpengli@tencent.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Explicitly truncate the data written to vmcs.SYSENTER_EIP/ESP on WRMSR
if the virtual CPU doesn't support 64-bit mode. The SYSENTER address
fields in the VMCS are natural width, i.e. bits 63:32 are dropped if the
CPU doesn't support Intel 64 architectures. This behavior is visible to
the guest after a VM-Exit/VM-Exit roundtrip, e.g. if the guest sets bits
63:32 in the actual MSR.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200428231025.12766-2-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Improve handle_external_interrupt_irqoff inline assembly in several ways:
- remove unneeded %c operand modifiers and "$" prefixes
- use %rsp instead of _ASM_SP, since we are in CONFIG_X86_64 part
- use $-16 immediate to align %rsp
- remove unneeded use of __ASM_SIZE macro
- define "ss" named operand only for X86_64
The patch introduces no functional changes.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Message-Id: <20200504155706.2516956-1-ubizjak@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Replace KVM's PT_PAGE_TABLE_LEVEL, PT_DIRECTORY_LEVEL and PT_PDPE_LEVEL
with the kernel's PG_LEVEL_4K, PG_LEVEL_2M and PG_LEVEL_1G. KVM's
enums are borderline impossible to remember and result in code that is
visually difficult to audit, e.g.
if (!enable_ept)
ept_lpage_level = 0;
else if (cpu_has_vmx_ept_1g_page())
ept_lpage_level = PT_PDPE_LEVEL;
else if (cpu_has_vmx_ept_2m_page())
ept_lpage_level = PT_DIRECTORY_LEVEL;
else
ept_lpage_level = PT_PAGE_TABLE_LEVEL;
versus
if (!enable_ept)
ept_lpage_level = 0;
else if (cpu_has_vmx_ept_1g_page())
ept_lpage_level = PG_LEVEL_1G;
else if (cpu_has_vmx_ept_2m_page())
ept_lpage_level = PG_LEVEL_2M;
else
ept_lpage_level = PG_LEVEL_4K;
No functional change intended.
Suggested-by: Barret Rhoden <brho@google.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200428005422.4235-4-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Snapshot the TDP level now that it's invariant (SVM) or dependent only
on host capabilities and guest CPUID (VMX). This avoids having to call
kvm_x86_ops.get_tdp_level() when initializing a TDP MMU and/or
calculating the page role, and thus avoids the associated retpoline.
Drop the WARN in vmx_get_tdp_level() as updating CPUID while L2 is
active is legal, if dodgy.
No functional change intended.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200502043234.12481-11-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Separate the "core" TDP level handling from the nested EPT path to make
it clear that kvm_x86_ops.get_tdp_level() is used if and only if nested
EPT is not in use (kvm_init_shadow_ept_mmu() calculates the level from
the passed in vmcs12->eptp). Add a WARN_ON() to enforce that the
kvm_x86_ops hook is not called for nested EPT.
This sets the stage for snapshotting the non-"nested EPT" TDP page level
during kvm_cpuid_update() to avoid the retpoline associated with
kvm_x86_ops.get_tdp_level() when resetting the MMU, a relatively
frequent operation when running a nested guest.
No functional change intended.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200502043234.12481-10-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move CR0 caching into the standard register caching mechanism in order
to take advantage of the availability checks provided by regs_avail.
This avoids multiple VMREADs in the (uncommon) case where kvm_read_cr0()
is called multiple times in a single VM-Exit, and more importantly
eliminates a kvm_x86_ops hook, saves a retpoline on SVM when reading
CR0, and squashes the confusing naming discrepancy of "cache_reg" vs.
"decache_cr0_guest_bits".
No functional change intended.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200502043234.12481-8-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move CR4 caching into the standard register caching mechanism in order
to take advantage of the availability checks provided by regs_avail.
This avoids multiple VMREADs and retpolines (when configured) during
nested VMX transitions as kvm_read_cr4_bits() is invoked multiple times
on each transition, e.g. when stuffing CR0 and CR3.
As an added bonus, this eliminates a kvm_x86_ops hook, saves a retpoline
on SVM when reading CR4, and squashes the confusing naming discrepancy
of "cache_reg" vs. "decache_cr4_guest_bits".
No functional change intended.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200502043234.12481-7-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Save L1's TSC offset in 'struct kvm_vcpu_arch' and drop the kvm_x86_ops
hook read_l1_tsc_offset(). This avoids a retpoline (when configured)
when reading L1's effective TSC, which is done at least once on every
VM-Exit.
No functional change intended.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200502043234.12481-2-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Skip the Indirect Branch Prediction Barrier that is triggered on a VMCS
switch when temporarily loading vmcs02 to synchronize it to vmcs12, i.e.
give copy_vmcs02_to_vmcs12_rare() the same treatment as
vmx_switch_vmcs().
Make vmx_vcpu_load() static now that it's only referenced within vmx.c.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200506235850.22600-3-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Skip the Indirect Branch Prediction Barrier that is triggered on a VMCS
switch when running with spectre_v2_user=on/auto if the switch is
between two VMCSes in the same guest, i.e. between vmcs01 and vmcs02.
The IBPB is intended to prevent one guest from attacking another, which
is unnecessary in the nested case as it's the same guest from KVM's
perspective.
This all but eliminates the overhead observed for nested VMX transitions
when running with CONFIG_RETPOLINE=y and spectre_v2_user=on/auto, which
can be significant, e.g. roughly 3x on current systems.
Reported-by: Alexander Graf <graf@amazon.com>
Cc: KarimAllah Raslan <karahmed@amazon.de>
Cc: stable@vger.kernel.org
Fixes: 15d4507152 ("KVM/x86: Add IBPB support")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200501163117.4655-1-sean.j.christopherson@intel.com>
[Invert direction of bool argument. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use vmx_get_intr_info() when grabbing the cached vmcs.INTR_INFO in
handle_exception_nmi() to ensure the cache isn't stale. Bypassing the
caching accessor doesn't cause any known issues as the cache is always
refreshed by handle_exception_nmi_irqoff(), but the whole point of
adding the proper caching mechanism was to avoid such dependencies.
Fixes: 8791585837 ("KVM: VMX: Cache vmcs.EXIT_INTR_INFO using arch avail_reg flags")
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200427171837.22613-1-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM is not handling the case where EIP wraps around the 32-bit address
space (that is, outside long mode). This is needed both in vmx.c
and in emulate.c. SVM with NRIPS is okay, but it can still print
an error to dmesg due to integer overflow.
Reported-by: Nick Peterson <everdox@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add an argument to interrupt_allowed and nmi_allowed, to checking if
interrupt injection is blocked. Use the hook to handle the case where
an interrupt arrives between check_nested_events() and the injection
logic. Drop the retry of check_nested_events() that hack-a-fixed the
same condition.
Blocking injection is also a bit of a hack, e.g. KVM should do exiting
and non-exiting interrupt processing in a single pass, but it's a more
precise hack. The old comment is also misleading, e.g. KVM_REQ_EVENT is
purely an optimization, setting it on every run loop (which KVM doesn't
do) should not affect functionality, only performance.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200423022550.15113-13-sean.j.christopherson@intel.com>
[Extend to SVM, add SMI and NMI. Even though NMI and SMI cannot come
asynchronously right now, making the fix generic is easy and removes a
special case. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use vmx_get_rflags() instead of manually reading vmcs.GUEST_RFLAGS when
querying RFLAGS.IF so that multiple checks against interrupt blocking in
a single run loop only require a single VMREAD.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200423022550.15113-14-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use vmx_interrupt_blocked() instead of bouncing through
vmx_interrupt_allowed() when handling edge cases in vmx_handle_exit().
The nested_run_pending check in vmx_interrupt_allowed() should never
evaluate true in the VM-Exit path.
Hoist the WARN in handle_invalid_guest_state() up to vmx_handle_exit()
to enforce the above assumption for the !enable_vnmi case, and to detect
any other potential bugs with nested VM-Enter.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200423022550.15113-12-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move the architectural (non-KVM specific) interrupt/NMI blocking checks
to a separate helper so that they can be used in a future patch by
vmx_check_nested_events().
No functional change intended.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200423022550.15113-8-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Report NMIs as allowed when the vCPU is in L2 and L2 is being run with
Exit-on-NMI enabled, as NMIs are always unblocked from L1's perspective
in this case.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200423022550.15113-7-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Do not hardcode is_smm so that all the architectural conditions for
blocking SMIs are listed in a single place. Well, in two places because
this introduces some code duplication between Intel and AMD.
This ensures that nested SVM obeys GIF in kvm_vcpu_has_events.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Return an actual bool for kvm_x86_ops' {interrupt_nmi}_allowed() hook to
better reflect the return semantics, and to avoid creating an even
bigger mess when the related VMX code is refactored in upcoming patches.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200423022550.15113-5-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Though rdpkru and wrpkru are contingent upon CR4.PKE, the PKRU
resource isn't. It can be read with XSAVE and written with XRSTOR.
So, if we don't set the guest PKRU value here(kvm_load_guest_xsave_state),
the guest can read the host value.
In case of kvm_load_host_xsave_state, guest with CR4.PKE clear could
potentially use XRSTOR to change the host PKRU value.
While at it, move pkru state save/restore to common code and the
host_pkru field to kvm_vcpu_arch. This will let SVM support protection keys.
Cc: stable@vger.kernel.org
Reported-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Babu Moger <babu.moger@amd.com>
Message-Id: <158932794619.44260.14508381096663848853.stgit@naples-babu.amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When KVM_EXIT_DEBUG is raised for the disabled-breakpoints case (DR7.GD),
DR6 was incorrectly copied from the value in the VM. Instead,
DR6.BD should be set in order to catch this case.
On AMD this does not need any special code because the processor triggers
a #DB exception that is intercepted. However, the testcase would fail
without the previous patch because both DR6.BS and DR6.BD would be set.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
There are two issues with KVM_EXIT_DEBUG on AMD, whose root cause is the
different handling of DR6 on intercepted #DB exceptions on Intel and AMD.
On Intel, #DB exceptions transmit the DR6 value via the exit qualification
field of the VMCS, and the exit qualification only contains the description
of the precise event that caused a vmexit.
On AMD, instead the DR6 field of the VMCB is filled in as if the #DB exception
was to be injected into the guest. This has two effects when guest debugging
is in use:
* the guest DR6 is clobbered
* the kvm_run->debug.arch.dr6 field can accumulate more debug events, rather
than just the last one that happened (the testcase in the next patch covers
this issue).
This patch fixes both issues by emulating, so to speak, the Intel behavior
on AMD processors. The important observation is that (after the previous
patches) the VMCB value of DR6 is only ever observable from the guest is
KVM_DEBUGREG_WONT_EXIT is set. Therefore we can actually set vmcb->save.dr6
to any value we want as long as KVM_DEBUGREG_WONT_EXIT is clear, which it
will be if guest debugging is enabled.
Therefore it is possible to enter the guest with an all-zero DR6,
reconstruct the #DB payload from the DR6 we get at exit time, and let
kvm_deliver_exception_payload move the newly set bits into vcpu->arch.dr6.
Some extra bits may be included in the payload if KVM_DEBUGREG_WONT_EXIT
is set, but this is harmless.
This may not be the most optimized way to deal with this, but it is
simple and, being confined within SVM code, it gets rid of the set_dr6
callback and kvm_update_dr6.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
kvm_x86_ops.set_dr6 is only ever called with vcpu->arch.dr6 as the
second argument. Ensure that the VMCB value is synchronized to
vcpu->arch.dr6 on #DB (both "normal" and nested) and nested vmentry, so
that the current value of DR6 is always available in vcpu->arch.dr6.
The get_dr6 callback can just access vcpu->arch.dr6 and becomes redundant.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
RTM should always been set even with KVM_EXIT_DEBUG on #DB.
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-Id: <20200505205000.188252-2-peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Go through kvm_queue_exception_p so that the payload is correctly delivered
through the exit qualification, and add a kvm_update_dr6 call to
kvm_deliver_exception_payload that is needed on AMD.
Reported-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Clean up some of the patching of kvm_x86_ops, by moving kvm_x86_ops related to
nested virtualization into a separate struct.
As a result, these ops will always be non-NULL on VMX. This is not a problem:
* check_nested_events is only called if is_guest_mode(vcpu) returns true
* get_nested_state treats VMXOFF state the same as nested being disabled
* set_nested_state fails if you attempt to set nested state while
nesting is disabled
* nested_enable_evmcs could already be called on a CPU without VMX enabled
in CPUID.
* nested_get_evmcs_version was fixed in the previous patch
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
IPI and Timer cause the main MSRs write vmexits in cloud environment
observation, let's optimize virtual IPI latency more aggressively to
inject target IPI as soon as possible.
Running kvm-unit-tests/vmexit.flat IPI testing on SKX server, disable
adaptive advance lapic timer and adaptive halt-polling to avoid the
interference, this patch can give another 7% improvement.
w/o fastpath -> x86.c fastpath 4238 -> 3543 16.4%
x86.c fastpath -> vmx.c fastpath 3543 -> 3293 7%
w/o fastpath -> vmx.c fastpath 4238 -> 3293 22.3%
Cc: Haiwei Li <lihaiwei@tencent.com>
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200410174703.1138-3-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Mark the VM-Fail, VM-Exit on VM-Enter, and #MC on VM-Enter paths as
'unlikely' so as to improve code generation so that it favors successful
VM-Enter. The performance of successful VM-Enter is for more important,
irrespective of whether or not success is actually likely.
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Message-Id: <20200410174703.1138-2-sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Linus Torvalds