platform/kernel/linux-rpi.git
2 years agokvm: add guest_state_{enter,exit}_irqoff()
Mark Rutland [Tue, 1 Feb 2022 13:29:22 +0000 (13:29 +0000)]
kvm: add guest_state_{enter,exit}_irqoff()

When transitioning to/from guest mode, it is necessary to inform
lockdep, tracing, and RCU in a specific order, similar to the
requirements for transitions to/from user mode. Additionally, it is
necessary to perform vtime accounting for a window around running the
guest, with RCU enabled, such that timer interrupts taken from the guest
can be accounted as guest time.

Most architectures don't handle all the necessary pieces, and a have a
number of common bugs, including unsafe usage of RCU during the window
between guest_enter() and guest_exit().

On x86, this was dealt with across commits:

  87fa7f3e98a1310e ("x86/kvm: Move context tracking where it belongs")
  0642391e2139a2c1 ("x86/kvm/vmx: Add hardirq tracing to guest enter/exit")
  9fc975e9efd03e57 ("x86/kvm/svm: Add hardirq tracing on guest enter/exit")
  3ebccdf373c21d86 ("x86/kvm/vmx: Move guest enter/exit into .noinstr.text")
  135961e0a7d555fc ("x86/kvm/svm: Move guest enter/exit into .noinstr.text")
  160457140187c5fb ("KVM: x86: Defer vtime accounting 'til after IRQ handling")
  bc908e091b326467 ("KVM: x86: Consolidate guest enter/exit logic to common helpers")

... but those fixes are specific to x86, and as the resulting logic
(while correct) is split across generic helper functions and
x86-specific helper functions, it is difficult to see that the
entry/exit accounting is balanced.

This patch adds generic helpers which architectures can use to handle
guest entry/exit consistently and correctly. The guest_{enter,exit}()
helpers are split into guest_timing_{enter,exit}() to perform vtime
accounting, and guest_context_{enter,exit}() to perform the necessary
context tracking and RCU management. The existing guest_{enter,exit}()
heleprs are left as wrappers of these.

Atop this, new guest_state_enter_irqoff() and guest_state_exit_irqoff()
helpers are added to handle the ordering of lockdep, tracing, and RCU
manageent. These are inteneded to mirror exit_to_user_mode() and
enter_from_user_mode().

Subsequent patches will migrate architectures over to the new helpers,
following a sequence:

guest_timing_enter_irqoff();

guest_state_enter_irqoff();
< run the vcpu >
guest_state_exit_irqoff();

< take any pending IRQs >

guest_timing_exit_irqoff();

This sequences handles all of the above correctly, and more clearly
balances the entry and exit portions, making it easier to understand.

The existing helpers are marked as deprecated, and will be removed once
all architectures have been converted.

There should be no functional change as a result of this patch.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Nicolas Saenz Julienne <nsaenzju@redhat.com>
Message-Id: <20220201132926.3301912-2-mark.rutland@arm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Move delivery of non-APICv interrupt into vendor code
Sean Christopherson [Fri, 28 Jan 2022 00:51:48 +0000 (00:51 +0000)]
KVM: x86: Move delivery of non-APICv interrupt into vendor code

Handle non-APICv interrupt delivery in vendor code, even though it means
VMX and SVM will temporarily have duplicate code.  SVM's AVIC has a race
condition that requires KVM to fall back to legacy interrupt injection
_after_ the interrupt has been logged in the vIRR, i.e. to fix the race,
SVM will need to open code the full flow anyways[*].  Refactor the code
so that the SVM bug without introducing other issues, e.g. SVM would
return "success" and thus invoke trace_kvm_apicv_accept_irq() even when
delivery through the AVIC failed, and to opportunistically prepare for
using KVM_X86_OP to fill each vendor's kvm_x86_ops struct, which will
rely on the vendor function matching the kvm_x86_op pointer name.

No functional change intended.

[*] https://lore.kernel.org/all/20211213104634.199141-4-mlevitsk@redhat.com

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220128005208.4008533-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agokvm: Move KVM_GET_XSAVE2 IOCTL definition at the end of kvm.h
Janosch Frank [Fri, 28 Jan 2022 15:40:25 +0000 (15:40 +0000)]
kvm: Move KVM_GET_XSAVE2 IOCTL definition at the end of kvm.h

This way we can more easily find the next free IOCTL number when
adding new IOCTLs.

Fixes: be50b2065dfa ("kvm: x86: Add support for getting/setting expanded xstate buffer")
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
Message-Id: <20220128154025.102666-1-frankja@linux.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoMerge tag 'kvmarm-fixes-5.17-1' of git://git.kernel.org/pub/scm/linux/kernel/git...
Paolo Bonzini [Fri, 28 Jan 2022 12:45:15 +0000 (07:45 -0500)]
Merge tag 'kvmarm-fixes-5.17-1' of git://git./linux/kernel/git/kvmarm/kvmarm into HEAD

KVM/arm64 fixes for 5.17, take #1

- Correctly update the shadow register on exception injection when
  running in nVHE mode

- Correctly use the mm_ops indirection when performing cache invalidation
  from the page-table walker

- Restrict the vgic-v3 workaround for SEIS to the two known broken
  implementations

2 years agoKVM: eventfd: Fix false positive RCU usage warning
Hou Wenlong [Thu, 27 Jan 2022 06:54:49 +0000 (14:54 +0800)]
KVM: eventfd: Fix false positive RCU usage warning

Fix the following false positive warning:
 =============================
 WARNING: suspicious RCU usage
 5.16.0-rc4+ #57 Not tainted
 -----------------------------
 arch/x86/kvm/../../../virt/kvm/eventfd.c:484 RCU-list traversed in non-reader section!!

 other info that might help us debug this:

 rcu_scheduler_active = 2, debug_locks = 1
 3 locks held by fc_vcpu 0/330:
  #0: ffff8884835fc0b0 (&vcpu->mutex){+.+.}-{3:3}, at: kvm_vcpu_ioctl+0x88/0x6f0 [kvm]
  #1: ffffc90004c0bb68 (&kvm->srcu){....}-{0:0}, at: vcpu_enter_guest+0x600/0x1860 [kvm]
  #2: ffffc90004c0c1d0 (&kvm->irq_srcu){....}-{0:0}, at: kvm_notify_acked_irq+0x36/0x180 [kvm]

 stack backtrace:
 CPU: 26 PID: 330 Comm: fc_vcpu 0 Not tainted 5.16.0-rc4+
 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
 Call Trace:
  <TASK>
  dump_stack_lvl+0x44/0x57
  kvm_notify_acked_gsi+0x6b/0x70 [kvm]
  kvm_notify_acked_irq+0x8d/0x180 [kvm]
  kvm_ioapic_update_eoi+0x92/0x240 [kvm]
  kvm_apic_set_eoi_accelerated+0x2a/0xe0 [kvm]
  handle_apic_eoi_induced+0x3d/0x60 [kvm_intel]
  vmx_handle_exit+0x19c/0x6a0 [kvm_intel]
  vcpu_enter_guest+0x66e/0x1860 [kvm]
  kvm_arch_vcpu_ioctl_run+0x438/0x7f0 [kvm]
  kvm_vcpu_ioctl+0x38a/0x6f0 [kvm]
  __x64_sys_ioctl+0x89/0xc0
  do_syscall_64+0x3a/0x90
  entry_SYSCALL_64_after_hwframe+0x44/0xae

Since kvm_unregister_irq_ack_notifier() does synchronize_srcu(&kvm->irq_srcu),
kvm->irq_ack_notifier_list is protected by kvm->irq_srcu. In fact,
kvm->irq_srcu SRCU read lock is held in kvm_notify_acked_irq(), making it
a false positive warning. So use hlist_for_each_entry_srcu() instead of
hlist_for_each_entry_rcu().

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Hou Wenlong <houwenlong93@linux.alibaba.com>
Message-Id: <f98bac4f5052bad2c26df9ad50f7019e40434512.1643265976.git.houwenlong.hwl@antgroup.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Allow VMREAD when Enlightened VMCS is in use
Vitaly Kuznetsov [Wed, 12 Jan 2022 17:01:34 +0000 (18:01 +0100)]
KVM: nVMX: Allow VMREAD when Enlightened VMCS is in use

Hyper-V TLFS explicitly forbids VMREAD and VMWRITE instructions when
Enlightened VMCS interface is in use:

"Any VMREAD or VMWRITE instructions while an enlightened VMCS is
active is unsupported and can result in unexpected behavior.""

Windows 11 + WSL2 seems to ignore this, attempts to VMREAD VMCS field
0x4404 ("VM-exit interruption information") are observed. Failing
these attempts with nested_vmx_failInvalid() makes such guests
unbootable.

Microsoft confirms this is a Hyper-V bug and claims that it'll get fixed
eventually but for the time being we need a workaround. (Temporary) allow
VMREAD to get data from the currently loaded Enlightened VMCS.

Note: VMWRITE instructions remain forbidden, it is not clear how to
handle them properly and hopefully won't ever be needed.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220112170134.1904308-6-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Implement evmcs_field_offset() suitable for handle_vmread()
Vitaly Kuznetsov [Wed, 12 Jan 2022 17:01:33 +0000 (18:01 +0100)]
KVM: nVMX: Implement evmcs_field_offset() suitable for handle_vmread()

In preparation to allowing reads from Enlightened VMCS from
handle_vmread(), implement evmcs_field_offset() to get the correct
read offset. get_evmcs_offset(), which is being used by KVM-on-Hyper-V,
is almost what's needed but a few things need to be adjusted. First,
WARN_ON() is unacceptable for handle_vmread() as any field can (in
theory) be supplied by the guest and not all fields are defined in
eVMCS v1. Second, we need to handle 'holes' in eVMCS (missing fields).
It also sounds like a good idea to WARN_ON() if such fields are ever
accessed by KVM-on-Hyper-V.

Implement dedicated evmcs_field_offset() helper.

No functional change intended.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220112170134.1904308-5-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Rename vmcs_to_field_offset{,_table}
Vitaly Kuznetsov [Wed, 12 Jan 2022 17:01:32 +0000 (18:01 +0100)]
KVM: nVMX: Rename vmcs_to_field_offset{,_table}

vmcs_to_field_offset{,_table} may sound misleading as VMCS is an opaque
blob which is not supposed to be accessed directly. In fact,
vmcs_to_field_offset{,_table} are related to KVM defined VMCS12 structure.

Rename vmcs_field_to_offset() to get_vmcs12_field_offset() for clarity.

No functional change intended.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220112170134.1904308-4-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: eVMCS: Filter out VM_EXIT_SAVE_VMX_PREEMPTION_TIMER
Vitaly Kuznetsov [Wed, 12 Jan 2022 17:01:31 +0000 (18:01 +0100)]
KVM: nVMX: eVMCS: Filter out VM_EXIT_SAVE_VMX_PREEMPTION_TIMER

Enlightened VMCS v1 doesn't have VMX_PREEMPTION_TIMER_VALUE field,
PIN_BASED_VMX_PREEMPTION_TIMER is also filtered out already so it makes
sense to filter out VM_EXIT_SAVE_VMX_PREEMPTION_TIMER too.

Note, none of the currently existing Windows/Hyper-V versions are known
to enable 'save VMX-preemption timer value' when eVMCS is in use, the
change is aimed at making the filtering future proof.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220112170134.1904308-3-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Also filter MSR_IA32_VMX_TRUE_PINBASED_CTLS when eVMCS
Vitaly Kuznetsov [Wed, 12 Jan 2022 17:01:30 +0000 (18:01 +0100)]
KVM: nVMX: Also filter MSR_IA32_VMX_TRUE_PINBASED_CTLS when eVMCS

Similar to MSR_IA32_VMX_EXIT_CTLS/MSR_IA32_VMX_TRUE_EXIT_CTLS,
MSR_IA32_VMX_ENTRY_CTLS/MSR_IA32_VMX_TRUE_ENTRY_CTLS pair,
MSR_IA32_VMX_TRUE_PINBASED_CTLS needs to be filtered the same way
MSR_IA32_VMX_PINBASED_CTLS is currently filtered as guests may solely rely
on 'true' MSR data.

Note, none of the currently existing Windows/Hyper-V versions are known
to stumble upon the unfiltered MSR_IA32_VMX_TRUE_PINBASED_CTLS, the change
is aimed at making the filtering future proof.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220112170134.1904308-2-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm: check dynamic bits against KVM_X86_XCOMP_GUEST_SUPP
Paolo Bonzini [Wed, 26 Jan 2022 12:51:00 +0000 (07:51 -0500)]
selftests: kvm: check dynamic bits against KVM_X86_XCOMP_GUEST_SUPP

Provide coverage for the new API.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: add system attribute to retrieve full set of supported xsave states
Paolo Bonzini [Wed, 26 Jan 2022 12:49:45 +0000 (07:49 -0500)]
KVM: x86: add system attribute to retrieve full set of supported xsave states

Because KVM_GET_SUPPORTED_CPUID is meant to be passed (by simple-minded
VMMs) to KVM_SET_CPUID2, it cannot include any dynamic xsave states that
have not been enabled.  Probing those, for example so that they can be
passed to ARCH_REQ_XCOMP_GUEST_PERM, requires a new ioctl or arch_prctl.
The latter is in fact worse, even though that is what the rest of the
API uses, because it would require supported_xcr0 to be moved from the
KVM module to the kernel just for this use.  In addition, the value
would be nonsensical (or an error would have to be returned) until
the KVM module is loaded in.

Therefore, to limit the growth of system ioctls, add a /dev/kvm
variant of KVM_{GET,HAS}_DEVICE_ATTR, and implement it in x86
with just one group (0) and attribute (KVM_X86_XCOMP_GUEST_SUPP).

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Add a helper to retrieve userspace address from kvm_device_attr
Sean Christopherson [Thu, 27 Jan 2022 15:31:53 +0000 (07:31 -0800)]
KVM: x86: Add a helper to retrieve userspace address from kvm_device_attr

Add a helper to handle converting the u64 userspace address embedded in
struct kvm_device_attr into a userspace pointer, it's all too easy to
forget the intermediate "unsigned long" cast as well as the truncation
check.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm: move vm_xsave_req_perm call to amx_test
Paolo Bonzini [Wed, 26 Jan 2022 12:44:34 +0000 (07:44 -0500)]
selftests: kvm: move vm_xsave_req_perm call to amx_test

There is no need for tests other than amx_test to enable dynamic xsave
states.  Remove the call to vm_xsave_req_perm from generic code,
and move it inside the test.  While at it, allow customizing the bit
that is requested, so that future tests can use it differently.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Sync the states size with the XCR0/IA32_XSS at, any time
Like Xu [Wed, 26 Jan 2022 17:22:26 +0000 (17:22 +0000)]
KVM: x86: Sync the states size with the XCR0/IA32_XSS at, any time

XCR0 is reset to 1 by RESET but not INIT and IA32_XSS is zeroed by
both RESET and INIT. The kvm_set_msr_common()'s handling of MSR_IA32_XSS
also needs to update kvm_update_cpuid_runtime(). In the above cases, the
size in bytes of the XSAVE area containing all states enabled by XCR0 or
(XCRO | IA32_XSS) needs to be updated.

For simplicity and consistency, existing helpers are used to write values
and call kvm_update_cpuid_runtime(), and it's not exactly a fast path.

Fixes: a554d207dc46 ("KVM: X86: Processor States following Reset or INIT")
Cc: stable@vger.kernel.org
Signed-off-by: Like Xu <likexu@tencent.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220126172226.2298529-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Update vCPU's runtime CPUID on write to MSR_IA32_XSS
Like Xu [Wed, 26 Jan 2022 17:22:25 +0000 (17:22 +0000)]
KVM: x86: Update vCPU's runtime CPUID on write to MSR_IA32_XSS

Do a runtime CPUID update for a vCPU if MSR_IA32_XSS is written, as the
size in bytes of the XSAVE area is affected by the states enabled in XSS.

Fixes: 203000993de5 ("kvm: vmx: add MSR logic for XSAVES")
Cc: stable@vger.kernel.org
Signed-off-by: Like Xu <likexu@tencent.com>
[sean: split out as a separate patch, adjust Fixes tag]
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220126172226.2298529-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Keep MSR_IA32_XSS unchanged for INIT
Xiaoyao Li [Wed, 26 Jan 2022 17:22:24 +0000 (17:22 +0000)]
KVM: x86: Keep MSR_IA32_XSS unchanged for INIT

It has been corrected from SDM version 075 that MSR_IA32_XSS is reset to
zero on Power up and Reset but keeps unchanged on INIT.

Fixes: a554d207dc46 ("KVM: X86: Processor States following Reset or INIT")
Cc: stable@vger.kernel.org
Signed-off-by: Xiaoyao Li <xiaoyao.li@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220126172226.2298529-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Free kvm_cpuid_entry2 array on post-KVM_RUN KVM_SET_CPUID{,2}
Sean Christopherson [Tue, 25 Jan 2022 21:04:45 +0000 (21:04 +0000)]
KVM: x86: Free kvm_cpuid_entry2 array on post-KVM_RUN KVM_SET_CPUID{,2}

Free the "struct kvm_cpuid_entry2" array on successful post-KVM_RUN
KVM_SET_CPUID{,2} to fix a memory leak, the callers of kvm_set_cpuid()
free the array only on failure.

 BUG: memory leak
 unreferenced object 0xffff88810963a800 (size 2048):
  comm "syz-executor025", pid 3610, jiffies 4294944928 (age 8.080s)
  hex dump (first 32 bytes):
    00 00 00 00 00 00 00 00 00 00 00 00 0d 00 00 00  ................
    47 65 6e 75 6e 74 65 6c 69 6e 65 49 00 00 00 00  GenuntelineI....
  backtrace:
    [<ffffffff814948ee>] kmalloc_node include/linux/slab.h:604 [inline]
    [<ffffffff814948ee>] kvmalloc_node+0x3e/0x100 mm/util.c:580
    [<ffffffff814950f2>] kvmalloc include/linux/slab.h:732 [inline]
    [<ffffffff814950f2>] vmemdup_user+0x22/0x100 mm/util.c:199
    [<ffffffff8109f5ff>] kvm_vcpu_ioctl_set_cpuid2+0x8f/0xf0 arch/x86/kvm/cpuid.c:423
    [<ffffffff810711b9>] kvm_arch_vcpu_ioctl+0xb99/0x1e60 arch/x86/kvm/x86.c:5251
    [<ffffffff8103e92d>] kvm_vcpu_ioctl+0x4ad/0x950 arch/x86/kvm/../../../virt/kvm/kvm_main.c:4066
    [<ffffffff815afacc>] vfs_ioctl fs/ioctl.c:51 [inline]
    [<ffffffff815afacc>] __do_sys_ioctl fs/ioctl.c:874 [inline]
    [<ffffffff815afacc>] __se_sys_ioctl fs/ioctl.c:860 [inline]
    [<ffffffff815afacc>] __x64_sys_ioctl+0xfc/0x140 fs/ioctl.c:860
    [<ffffffff844a3335>] do_syscall_x64 arch/x86/entry/common.c:50 [inline]
    [<ffffffff844a3335>] do_syscall_64+0x35/0xb0 arch/x86/entry/common.c:80
    [<ffffffff84600068>] entry_SYSCALL_64_after_hwframe+0x44/0xae

Fixes: c6617c61e8fe ("KVM: x86: Partially allow KVM_SET_CPUID{,2} after KVM_RUN")
Cc: stable@vger.kernel.org
Reported-by: syzbot+be576ad7655690586eec@syzkaller.appspotmail.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220125210445.2053429-1-seanjc@google.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: WARN on any attempt to allocate shadow VMCS for vmcs02
Sean Christopherson [Tue, 25 Jan 2022 22:05:27 +0000 (22:05 +0000)]
KVM: nVMX: WARN on any attempt to allocate shadow VMCS for vmcs02

WARN if KVM attempts to allocate a shadow VMCS for vmcs02.  KVM emulates
VMCS shadowing but doesn't virtualize it, i.e. KVM should never allocate
a "real" shadow VMCS for L2.

The previous code WARNed but continued anyway with the allocation,
presumably in an attempt to avoid NULL pointer dereference.
However, alloc_vmcs (and hence alloc_shadow_vmcs) can fail, and
indeed the sole caller does:

if (enable_shadow_vmcs && !alloc_shadow_vmcs(vcpu))
goto out_shadow_vmcs;

which makes it not a useful attempt.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220125220527.2093146-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Don't skip L2's VMCALL in SMM test for SVM guest
Sean Christopherson [Tue, 25 Jan 2022 22:17:25 +0000 (22:17 +0000)]
KVM: selftests: Don't skip L2's VMCALL in SMM test for SVM guest

Don't skip the vmcall() in l2_guest_code() prior to re-entering L2, doing
so will result in L2 running to completion, popping '0' off the stack for
RET, jumping to address '0', and ultimately dying with a triple fault
shutdown.

It's not at all obvious why the test re-enters L2 and re-executes VMCALL,
but presumably it serves a purpose.  The VMX path doesn't skip vmcall(),
and the test can't possibly have passed on SVM, so just do what VMX does.

Fixes: d951b2210c1a ("KVM: selftests: smm_test: Test SMM enter from L2")
Cc: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220125221725.2101126-1-seanjc@google.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Tested-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Check .flags in kvm_cpuid_check_equal() too
Vitaly Kuznetsov [Wed, 26 Jan 2022 13:18:04 +0000 (14:18 +0100)]
KVM: x86: Check .flags in kvm_cpuid_check_equal() too

kvm_cpuid_check_equal() checks for the (full) equality of the supplied
CPUID data so .flags need to be checked too.

Reported-by: Sean Christopherson <seanjc@google.com>
Fixes: c6617c61e8fe ("KVM: x86: Partially allow KVM_SET_CPUID{,2} after KVM_RUN")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220126131804.2839410-1-vkuznets@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Forcibly leave nested virt when SMM state is toggled
Sean Christopherson [Tue, 25 Jan 2022 22:03:58 +0000 (22:03 +0000)]
KVM: x86: Forcibly leave nested virt when SMM state is toggled

Forcibly leave nested virtualization operation if userspace toggles SMM
state via KVM_SET_VCPU_EVENTS or KVM_SYNC_X86_EVENTS.  If userspace
forces the vCPU out of SMM while it's post-VMXON and then injects an SMI,
vmx_enter_smm() will overwrite vmx->nested.smm.vmxon and end up with both
vmxon=false and smm.vmxon=false, but all other nVMX state allocated.

Don't attempt to gracefully handle the transition as (a) most transitions
are nonsencial, e.g. forcing SMM while L2 is running, (b) there isn't
sufficient information to handle all transitions, e.g. SVM wants access
to the SMRAM save state, and (c) KVM_SET_VCPU_EVENTS must precede
KVM_SET_NESTED_STATE during state restore as the latter disallows putting
the vCPU into L2 if SMM is active, and disallows tagging the vCPU as
being post-VMXON in SMM if SMM is not active.

Abuse of KVM_SET_VCPU_EVENTS manifests as a WARN and memory leak in nVMX
due to failure to free vmcs01's shadow VMCS, but the bug goes far beyond
just a memory leak, e.g. toggling SMM on while L2 is active puts the vCPU
in an architecturally impossible state.

  WARNING: CPU: 0 PID: 3606 at free_loaded_vmcs arch/x86/kvm/vmx/vmx.c:2665 [inline]
  WARNING: CPU: 0 PID: 3606 at free_loaded_vmcs+0x158/0x1a0 arch/x86/kvm/vmx/vmx.c:2656
  Modules linked in:
  CPU: 1 PID: 3606 Comm: syz-executor725 Not tainted 5.17.0-rc1-syzkaller #0
  Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
  RIP: 0010:free_loaded_vmcs arch/x86/kvm/vmx/vmx.c:2665 [inline]
  RIP: 0010:free_loaded_vmcs+0x158/0x1a0 arch/x86/kvm/vmx/vmx.c:2656
  Code: <0f> 0b eb b3 e8 8f 4d 9f 00 e9 f7 fe ff ff 48 89 df e8 92 4d 9f 00
  Call Trace:
   <TASK>
   kvm_arch_vcpu_destroy+0x72/0x2f0 arch/x86/kvm/x86.c:11123
   kvm_vcpu_destroy arch/x86/kvm/../../../virt/kvm/kvm_main.c:441 [inline]
   kvm_destroy_vcpus+0x11f/0x290 arch/x86/kvm/../../../virt/kvm/kvm_main.c:460
   kvm_free_vcpus arch/x86/kvm/x86.c:11564 [inline]
   kvm_arch_destroy_vm+0x2e8/0x470 arch/x86/kvm/x86.c:11676
   kvm_destroy_vm arch/x86/kvm/../../../virt/kvm/kvm_main.c:1217 [inline]
   kvm_put_kvm+0x4fa/0xb00 arch/x86/kvm/../../../virt/kvm/kvm_main.c:1250
   kvm_vm_release+0x3f/0x50 arch/x86/kvm/../../../virt/kvm/kvm_main.c:1273
   __fput+0x286/0x9f0 fs/file_table.c:311
   task_work_run+0xdd/0x1a0 kernel/task_work.c:164
   exit_task_work include/linux/task_work.h:32 [inline]
   do_exit+0xb29/0x2a30 kernel/exit.c:806
   do_group_exit+0xd2/0x2f0 kernel/exit.c:935
   get_signal+0x4b0/0x28c0 kernel/signal.c:2862
   arch_do_signal_or_restart+0x2a9/0x1c40 arch/x86/kernel/signal.c:868
   handle_signal_work kernel/entry/common.c:148 [inline]
   exit_to_user_mode_loop kernel/entry/common.c:172 [inline]
   exit_to_user_mode_prepare+0x17d/0x290 kernel/entry/common.c:207
   __syscall_exit_to_user_mode_work kernel/entry/common.c:289 [inline]
   syscall_exit_to_user_mode+0x19/0x60 kernel/entry/common.c:300
   do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86
   entry_SYSCALL_64_after_hwframe+0x44/0xae
   </TASK>

Cc: stable@vger.kernel.org
Reported-by: syzbot+8112db3ab20e70d50c31@syzkaller.appspotmail.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220125220358.2091737-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: drop unnecessary code in svm_hv_vmcb_dirty_nested_enlightenments()
Vitaly Kuznetsov [Mon, 20 Dec 2021 15:21:35 +0000 (16:21 +0100)]
KVM: SVM: drop unnecessary code in svm_hv_vmcb_dirty_nested_enlightenments()

Commit 3fa5e8fd0a0e4 ("KVM: SVM: delay svm_vcpu_init_msrpm after
svm->vmcb is initialized") re-arranged svm_vcpu_init_msrpm() call in
svm_create_vcpu(), thus making the comment about vmcb being NULL
obsolete. Drop it.

While on it, drop superfluous vmcb_is_clean() check: vmcb_mark_dirty()
is a bit flip, an extra check is unlikely to bring any performance gain.
Drop now-unneeded vmcb_is_clean() helper as well.

Fixes: 3fa5e8fd0a0e4 ("KVM: SVM: delay svm_vcpu_init_msrpm after svm->vmcb is initialized")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211220152139.418372-2-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: hyper-v: Enable Enlightened MSR-Bitmap support for real
Vitaly Kuznetsov [Mon, 20 Dec 2021 15:21:36 +0000 (16:21 +0100)]
KVM: SVM: hyper-v: Enable Enlightened MSR-Bitmap support for real

Commit c4327f15dfc7 ("KVM: SVM: hyper-v: Enlightened MSR-Bitmap support")
introduced enlightened MSR-Bitmap support for KVM-on-Hyper-V but it didn't
actually enable the support. Similar to enlightened NPT TLB flush and
direct TLB flush features, the guest (KVM) has to tell L0 (Hyper-V) that
it's using the feature by setting the appropriate feature fit in VMCB
control area (sw reserved fields).

Fixes: c4327f15dfc7 ("KVM: SVM: hyper-v: Enlightened MSR-Bitmap support")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211220152139.418372-3-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Don't kill SEV guest if SMAP erratum triggers in usermode
Sean Christopherson [Thu, 20 Jan 2022 01:07:19 +0000 (01:07 +0000)]
KVM: SVM: Don't kill SEV guest if SMAP erratum triggers in usermode

Inject a #GP instead of synthesizing triple fault to try to avoid killing
the guest if emulation of an SEV guest fails due to encountering the SMAP
erratum.  The injected #GP may still be fatal to the guest, e.g. if the
userspace process is providing critical functionality, but KVM should
make every attempt to keep the guest alive.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Don't apply SEV+SMAP workaround on code fetch or PT access
Sean Christopherson [Thu, 20 Jan 2022 01:07:18 +0000 (01:07 +0000)]
KVM: SVM: Don't apply SEV+SMAP workaround on code fetch or PT access

Resume the guest instead of synthesizing a triple fault shutdown if the
instruction bytes buffer is empty due to the #NPF being on the code fetch
itself or on a page table access.  The SMAP errata applies if and only if
the code fetch was successful and ucode's subsequent data read from the
code page encountered a SMAP violation.  In practice, the guest is likely
hosed either way, but crashing the guest on a code fetch to emulated MMIO
is technically wrong according to the behavior described in the APM.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Inject #UD on attempted emulation for SEV guest w/o insn buffer
Sean Christopherson [Thu, 20 Jan 2022 01:07:17 +0000 (01:07 +0000)]
KVM: SVM: Inject #UD on attempted emulation for SEV guest w/o insn buffer

Inject #UD if KVM attempts emulation for an SEV guests without an insn
buffer and instruction decoding is required.  The previous behavior of
allowing emulation if there is no insn buffer is undesirable as doing so
means KVM is reading guest private memory and thus decoding cyphertext,
i.e. is emulating garbage.  The check was previously necessary as the
emulation type was not provided, i.e. SVM needed to allow emulation to
handle completion of emulation after exiting to userspace to handle I/O.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: WARN if KVM attempts emulation on #UD or #GP for SEV guests
Sean Christopherson [Thu, 20 Jan 2022 01:07:16 +0000 (01:07 +0000)]
KVM: SVM: WARN if KVM attempts emulation on #UD or #GP for SEV guests

WARN if KVM attempts to emulate in response to #UD or #GP for SEV guests,
i.e. if KVM intercepts #UD or #GP, as emulation on any fault except #NPF
is impossible since KVM cannot read guest private memory to get the code
stream, and the CPU's DecodeAssists feature only provides the instruction
bytes on #NPF.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-7-seanjc@google.com>
[Warn on EMULTYPE_TRAP_UD_FORCED according to Liam Merwick's review. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Pass emulation type to can_emulate_instruction()
Sean Christopherson [Thu, 20 Jan 2022 01:07:15 +0000 (01:07 +0000)]
KVM: x86: Pass emulation type to can_emulate_instruction()

Pass the emulation type to kvm_x86_ops.can_emulate_insutrction() so that
a future commit can harden KVM's SEV support to WARN on emulation
scenarios that should never happen.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Explicitly require DECODEASSISTS to enable SEV support
Sean Christopherson [Thu, 20 Jan 2022 01:07:14 +0000 (01:07 +0000)]
KVM: SVM: Explicitly require DECODEASSISTS to enable SEV support

Add a sanity check on DECODEASSIST being support if SEV is supported, as
KVM cannot read guest private memory and thus relies on the CPU to
provide the instruction byte stream on #NPF for emulation.  The intent of
the check is to document the dependency, it should never fail in practice
as producing hardware that supports SEV but not DECODEASSISTS would be
non-sensical.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Don't intercept #GP for SEV guests
Sean Christopherson [Thu, 20 Jan 2022 01:07:13 +0000 (01:07 +0000)]
KVM: SVM: Don't intercept #GP for SEV guests

Never intercept #GP for SEV guests as reading SEV guest private memory
will return cyphertext, i.e. emulating on #GP can't work as intended.

Cc: stable@vger.kernel.org
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoRevert "KVM: SVM: avoid infinite loop on NPF from bad address"
Sean Christopherson [Thu, 20 Jan 2022 01:07:12 +0000 (01:07 +0000)]
Revert "KVM: SVM: avoid infinite loop on NPF from bad address"

Revert a completely broken check on an "invalid" RIP in SVM's workaround
for the DecodeAssists SMAP errata.  kvm_vcpu_gfn_to_memslot() obviously
expects a gfn, i.e. operates in the guest physical address space, whereas
RIP is a virtual (not even linear) address.  The "fix" worked for the
problematic KVM selftest because the test identity mapped RIP.

Fully revert the hack instead of trying to translate RIP to a GPA, as the
non-SEV case is now handled earlier, and KVM cannot access guest page
tables to translate RIP.

This reverts commit e72436bc3a5206f95bb384e741154166ddb3202e.

Fixes: e72436bc3a52 ("KVM: SVM: avoid infinite loop on NPF from bad address")
Reported-by: Liam Merwick <liam.merwick@oracle.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Never reject emulation due to SMAP errata for !SEV guests
Sean Christopherson [Thu, 20 Jan 2022 01:07:11 +0000 (01:07 +0000)]
KVM: SVM: Never reject emulation due to SMAP errata for !SEV guests

Always signal that emulation is possible for !SEV guests regardless of
whether or not the CPU provided a valid instruction byte stream.  KVM can
read all guest state (memory and registers) for !SEV guests, i.e. can
fetch the code stream from memory even if the CPU failed to do so because
of the SMAP errata.

Fixes: 05d5a4863525 ("KVM: SVM: Workaround errata#1096 (insn_len maybe zero on SMAP violation)")
Cc: stable@vger.kernel.org
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Liam Merwick <liam.merwick@oracle.com>
Message-Id: <20220120010719.711476-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: nSVM: skip eax alignment check for non-SVM instructions
Denis Valeev [Sat, 22 Jan 2022 20:13:57 +0000 (23:13 +0300)]
KVM: x86: nSVM: skip eax alignment check for non-SVM instructions

The bug occurs on #GP triggered by VMware backdoor when eax value is
unaligned. eax alignment check should not be applied to non-SVM
instructions because it leads to incorrect omission of the instructions
emulation.
Apply the alignment check only to SVM instructions to fix.

Fixes: d1cba6c92237 ("KVM: x86: nSVM: test eax for 4K alignment for GP errata workaround")
Signed-off-by: Denis Valeev <lemniscattaden@gmail.com>
Message-Id: <Yexlhaoe1Fscm59u@q>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/cpuid: Exclude unpermitted xfeatures sizes at KVM_GET_SUPPORTED_CPUID
Like Xu [Tue, 25 Jan 2022 11:52:23 +0000 (19:52 +0800)]
KVM: x86/cpuid: Exclude unpermitted xfeatures sizes at KVM_GET_SUPPORTED_CPUID

With the help of xstate_get_guest_group_perm(), KVM can exclude unpermitted
xfeatures in cpuid.0xd.0.eax, in which case the corresponding xfeatures
sizes should also be matched to the permitted xfeatures.

To fix this inconsistency, the permitted_xcr0 and permitted_xss are defined
consistently, which implies 'supported' plus certain permissions for this
task, and it also fixes cpuid.0xd.1.ebx and later leaf-by-leaf queries.

Fixes: 445ecdf79be0 ("kvm: x86: Exclude unpermitted xfeatures at KVM_GET_SUPPORTED_CPUID")
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220125115223.33707-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: LAPIC: Also cancel preemption timer during SET_LAPIC
Wanpeng Li [Tue, 25 Jan 2022 09:17:00 +0000 (01:17 -0800)]
KVM: LAPIC: Also cancel preemption timer during SET_LAPIC

The below warning is splatting during guest reboot.

  ------------[ cut here ]------------
  WARNING: CPU: 0 PID: 1931 at arch/x86/kvm/x86.c:10322 kvm_arch_vcpu_ioctl_run+0x874/0x880 [kvm]
  CPU: 0 PID: 1931 Comm: qemu-system-x86 Tainted: G          I       5.17.0-rc1+ #5
  RIP: 0010:kvm_arch_vcpu_ioctl_run+0x874/0x880 [kvm]
  Call Trace:
   <TASK>
   kvm_vcpu_ioctl+0x279/0x710 [kvm]
   __x64_sys_ioctl+0x83/0xb0
   do_syscall_64+0x3b/0xc0
   entry_SYSCALL_64_after_hwframe+0x44/0xae
  RIP: 0033:0x7fd39797350b

This can be triggered by not exposing tsc-deadline mode and doing a reboot in
the guest. The lapic_shutdown() function which is called in sys_reboot path
will not disarm the flying timer, it just masks LVTT. lapic_shutdown() clears
APIC state w/ LVT_MASKED and timer-mode bit is 0, this can trigger timer-mode
switch between tsc-deadline and oneshot/periodic, which can result in preemption
timer be cancelled in apic_update_lvtt(). However, We can't depend on this when
not exposing tsc-deadline mode and oneshot/periodic modes emulated by preemption
timer. Qemu will synchronise states around reset, let's cancel preemption timer
under KVM_SET_LAPIC.

Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Message-Id: <1643102220-35667-1-git-send-email-wanpengli@tencent.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Remove vmcs_config.order
Jim Mattson [Tue, 25 Jan 2022 00:43:59 +0000 (16:43 -0800)]
KVM: VMX: Remove vmcs_config.order

The maximum size of a VMCS (or VMXON region) is 4096. By definition,
these are order 0 allocations.

Signed-off-by: Jim Mattson <jmattson@google.com>
Message-Id: <20220125004359.147600-1-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM/X86: Make kvm_vcpu_reload_apic_access_page() static
Quanfa Fu [Sun, 19 Dec 2021 09:14:46 +0000 (17:14 +0800)]
KVM/X86: Make kvm_vcpu_reload_apic_access_page() static

Make kvm_vcpu_reload_apic_access_page() static
as it is no longer invoked directly by vmx
and it is also no longer exported.

No functional change intended.

Signed-off-by: Quanfa Fu <quanfafu@gmail.com>
Message-Id: <20211219091446.174584-1-quanfafu@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Re-enable access_tracking_perf_test
David Matlack [Thu, 20 Jan 2022 00:38:26 +0000 (00:38 +0000)]
KVM: selftests: Re-enable access_tracking_perf_test

This selftest was accidentally removed by commit 6a58150859fd
("selftest: KVM: Add intra host migration tests"). Add it back.

Fixes: 6a58150859fd ("selftest: KVM: Add intra host migration tests")
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220120003826.2805036-1-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Set vmcs.PENDING_DBG.BS on #DB in STI/MOVSS blocking shadow
Sean Christopherson [Thu, 20 Jan 2022 00:06:24 +0000 (00:06 +0000)]
KVM: VMX: Set vmcs.PENDING_DBG.BS on #DB in STI/MOVSS blocking shadow

Set vmcs.GUEST_PENDING_DBG_EXCEPTIONS.BS, a.k.a. the pending single-step
breakpoint flag, when re-injecting a #DB with RFLAGS.TF=1, and STI or
MOVSS blocking is active.  Setting the flag is necessary to make VM-Entry
consistency checks happy, as VMX has an invariant that if RFLAGS.TF is
set and STI/MOVSS blocking is true, then the previous instruction must
have been STI or MOV/POP, and therefore a single-step #DB must be pending
since the RFLAGS.TF cannot have been set by the previous instruction,
i.e. the one instruction delay after setting RFLAGS.TF must have already
expired.

Normally, the CPU sets vmcs.GUEST_PENDING_DBG_EXCEPTIONS.BS appropriately
when recording guest state as part of a VM-Exit, but #DB VM-Exits
intentionally do not treat the #DB as "guest state" as interception of
the #DB effectively makes the #DB host-owned, thus KVM needs to manually
set PENDING_DBG.BS when forwarding/re-injecting the #DB to the guest.

Note, although this bug can be triggered by guest userspace, doing so
requires IOPL=3, and guest userspace running with IOPL=3 has full access
to all I/O ports (from the guest's perspective) and can crash/reboot the
guest any number of ways.  IOPL=3 is required because STI blocking kicks
in if and only if RFLAGS.IF is toggled 0=>1, and if CPL>IOPL, STI either
takes a #GP or modifies RFLAGS.VIF, not RFLAGS.IF.

MOVSS blocking can be initiated by userspace, but can be coincident with
a #DB if and only if DR7.GD=1 (General Detect enabled) and a MOV DR is
executed in the MOVSS shadow.  MOV DR #GPs at CPL>0, thus MOVSS blocking
is problematic only for CPL0 (and only if the guest is crazy enough to
access a DR in a MOVSS shadow).  All other sources of #DBs are either
suppressed by MOVSS blocking (single-step, code fetch, data, and I/O),
are mutually exclusive with MOVSS blocking (T-bit task switch), or are
already handled by KVM (ICEBP, a.k.a. INT1).

This bug was originally found by running tests[1] created for XSA-308[2].
Note that Xen's userspace test emits ICEBP in the MOVSS shadow, which is
presumably why the Xen bug was deemed to be an exploitable DOS from guest
userspace.  KVM already handles ICEBP by skipping the ICEBP instruction
and thus clears MOVSS blocking as a side effect of its "emulation".

[1] http://xenbits.xenproject.org/docs/xtf/xsa-308_2main_8c_source.html
[2] https://xenbits.xen.org/xsa/advisory-308.html

Reported-by: David Woodhouse <dwmw2@infradead.org>
Reported-by: Alexander Graf <graf@amazon.de>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220120000624.655815-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Move CPUID.(EAX=0x12,ECX=1) mangling to __kvm_update_cpuid_runtime()
Vitaly Kuznetsov [Mon, 24 Jan 2022 10:36:05 +0000 (11:36 +0100)]
KVM: x86: Move CPUID.(EAX=0x12,ECX=1) mangling to __kvm_update_cpuid_runtime()

Full equality check of CPUID data on update (kvm_cpuid_check_equal()) may
fail for SGX enabled CPUs as CPUID.(EAX=0x12,ECX=1) is currently being
mangled in kvm_vcpu_after_set_cpuid(). Move it to
__kvm_update_cpuid_runtime() and split off cpuid_get_supported_xcr0()
helper  as 'vcpu->arch.guest_supported_xcr0' update needs (logically)
to stay in kvm_vcpu_after_set_cpuid().

Cc: stable@vger.kernel.org
Fixes: feb627e8d6f6 ("KVM: x86: Forbid KVM_SET_CPUID{,2} after KVM_RUN")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220124103606.2630588-2-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: remove async parameter of hva_to_pfn_remapped()
Xianting Tian [Mon, 24 Jan 2022 02:04:56 +0000 (10:04 +0800)]
KVM: remove async parameter of hva_to_pfn_remapped()

The async parameter of hva_to_pfn_remapped() is not used, so remove it.

Signed-off-by: Xianting Tian <xianting.tian@linux.alibaba.com>
Message-Id: <20220124020456.156386-1-xianting.tian@linux.alibaba.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agox86,kvm/xen: Remove superfluous .fixup usage
Peter Zijlstra [Sun, 23 Jan 2022 12:42:19 +0000 (13:42 +0100)]
x86,kvm/xen: Remove superfluous .fixup usage

Commit 14243b387137 ("KVM: x86/xen: Add KVM_IRQ_ROUTING_XEN_EVTCHN and
event channel delivery") adds superfluous .fixup usage after the whole
.fixup section was removed in commit e5eefda5aa51 ("x86: Remove .fixup
section").

Fixes: 14243b387137 ("KVM: x86/xen: Add KVM_IRQ_ROUTING_XEN_EVTCHN and event channel delivery")
Reported-by: Borislav Petkov <bp@alien8.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Message-Id: <20220123124219.GH20638@worktop.programming.kicks-ass.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Zero host's SYSENTER_ESP iff SYSENTER is NOT used
Sean Christopherson [Sat, 22 Jan 2022 01:52:11 +0000 (01:52 +0000)]
KVM: VMX: Zero host's SYSENTER_ESP iff SYSENTER is NOT used

Zero vmcs.HOST_IA32_SYSENTER_ESP when initializing *constant* host state
if and only if SYSENTER cannot be used, i.e. the kernel is a 64-bit
kernel and is not emulating 32-bit syscalls.  As the name suggests,
vmx_set_constant_host_state() is intended for state that is *constant*.
When SYSENTER is used, SYSENTER_ESP isn't constant because stacks are
per-CPU, and the VMCS must be updated whenever the vCPU is migrated to a
new CPU.  The logic in vmx_vcpu_load_vmcs() doesn't differentiate between
"never loaded" and "loaded on a different CPU", i.e. setting SYSENTER_ESP
on VMCS load also handles setting correct host state when the VMCS is
first loaded.

Because a VMCS must be loaded before it is initialized during vCPU RESET,
zeroing the field in vmx_set_constant_host_state() obliterates the value
that was written when the VMCS was loaded.  If the vCPU is run before it
is migrated, the subsequent VM-Exit will zero out MSR_IA32_SYSENTER_ESP,
leading to a #DF on the next 32-bit syscall.

  double fault: 0000 [#1] SMP
  CPU: 0 PID: 990 Comm: stable Not tainted 5.16.0+ #97
  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
  EIP: entry_SYSENTER_32+0x0/0xe7
  Code: <9c> 50 eb 17 0f 20 d8 a9 00 10 00 00 74 0d 25 ff ef ff ff 0f 22 d8
  EAX: 000000a2 EBX: a8d1300c ECX: a8d13014 EDX: 00000000
  ESI: a8f87000 EDI: a8d13014 EBP: a8d12fc0 ESP: 00000000
  DS: 007b ES: 007b FS: 0000 GS: 0000 SS: 0068 EFLAGS: 00210093
  CR0: 80050033 CR2: fffffffc CR3: 02c3b000 CR4: 00152e90

Fixes: 6ab8a4053f71 ("KVM: VMX: Avoid to rdmsrl(MSR_IA32_SYSENTER_ESP)")
Cc: Lai Jiangshan <laijs@linux.alibaba.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220122015211.1468758-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: arm64: Use shadow SPSR_EL1 when injecting exceptions on !VHE
Marc Zyngier [Fri, 21 Jan 2022 18:42:07 +0000 (18:42 +0000)]
KVM: arm64: Use shadow SPSR_EL1 when injecting exceptions on !VHE

Injecting an exception into a guest with non-VHE is risky business.
Instead of writing in the shadow register for the switch code to
restore it, we override the CPU register instead. Which gets
overriden a few instructions later by said restore code.

The result is that although the guest correctly gets the exception,
it will return to the original context in some random state,
depending on what was there the first place... Boo.

Fix the issue by writing to the shadow register. The original code
is absolutely fine on VHE, as the state is already loaded, and writing
to the shadow register in that case would actually be a bug.

Fixes: bb666c472ca2 ("KVM: arm64: Inject AArch64 exceptions from HYP")
Cc: stable@vger.kernel.org
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Fuad Tabba <tabba@google.com>
Link: https://lore.kernel.org/r/20220121184207.423426-1-maz@kernel.org
2 years agoKVM: arm64: vgic-v3: Restrict SEIS workaround to known broken systems
Marc Zyngier [Fri, 21 Jan 2022 21:07:47 +0000 (21:07 +0000)]
KVM: arm64: vgic-v3: Restrict SEIS workaround to known broken systems

Contrary to what df652bcf1136 ("KVM: arm64: vgic-v3: Work around GICv3
locally generated SErrors") was asserting, there is at least one other
system out there (Cavium ThunderX2) implementing SEIS, and not in
an obviously broken way.

So instead of imposing the M1 workaround on an innocent bystander,
let's limit it to the two known broken Apple implementations.

Fixes: df652bcf1136 ("KVM: arm64: vgic-v3: Work around GICv3 locally generated SErrors")
Reported-by: Ard Biesheuvel <ardb@kernel.org>
Tested-by: Ard Biesheuvel <ardb@kernel.org>
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20220122103912.795026-1-maz@kernel.org
2 years agodocs: kvm: fix WARNINGs from api.rst
Wei Wang [Thu, 20 Jan 2022 04:50:03 +0000 (23:50 -0500)]
docs: kvm: fix WARNINGs from api.rst

Use the api number 134 for KVM_GET_XSAVE2, instead of 42, which has been
used by KVM_GET_XSAVE.
Also, fix the WARNINGs of the underlines being too short.

Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Wei Wang <wei.w.wang@intel.com>
Tested-by: Stephen Rothwell <sfr@canb.auug.org.au>
Message-Id: <20220120045003.315177-1-wei.w.wang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm/x86: Fix the warning in lib/x86_64/processor.c
Jinrong Liang [Wed, 19 Jan 2022 14:03:25 +0000 (22:03 +0800)]
selftests: kvm/x86: Fix the warning in lib/x86_64/processor.c

The following warning appears when executing
make -C tools/testing/selftests/kvm

include/x86_64/processor.h:290:2: warning: 'ecx' may be used uninitialized in this
function [-Wmaybe-uninitialized]
  asm volatile("cpuid"
  ^~~
lib/x86_64/processor.c:1523:21: note: 'ecx' was declared here
  uint32_t eax, ebx, ecx, edx, max_ext_leaf;

Just initialize ecx to remove this warning.

Fixes: c8cc43c1eae2 ("selftests: KVM: avoid failures due to reserved HyperTransport region")
Signed-off-by: Jinrong Liang <cloudliang@tencent.com>
Message-Id: <20220119140325.59369-1-cloudliang@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm/x86: Fix the warning in pmu_event_filter_test.c
Jinrong Liang [Wed, 19 Jan 2022 13:39:10 +0000 (21:39 +0800)]
selftests: kvm/x86: Fix the warning in pmu_event_filter_test.c

The following warning appears when executing
make -C tools/testing/selftests/kvm

x86_64/pmu_event_filter_test.c: In function 'vcpu_supports_intel_br_retired':
x86_64/pmu_event_filter_test.c:241:28: warning: variable 'cpuid' set but not used [-Wunused-but-set-variable]
  241 |         struct kvm_cpuid2 *cpuid;
      |                            ^~~~~
x86_64/pmu_event_filter_test.c: In function 'vcpu_supports_amd_zen_br_retired':
x86_64/pmu_event_filter_test.c:258:28: warning: variable 'cpuid' set but not used [-Wunused-but-set-variable]
  258 |         struct kvm_cpuid2 *cpuid;
      |                            ^~~~~

Just delete the unused variables to stay away from warnings.

Fixes: dc7e75b3b3ee ("selftests: kvm/x86: Add test for KVM_SET_PMU_EVENT_FILTER")
Signed-off-by: Jinrong Liang <cloudliang@tencent.com>
Message-Id: <20220119133910.56285-1-cloudliang@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agokvm: selftests: Do not indent with spaces
Paolo Bonzini [Wed, 19 Jan 2022 18:34:30 +0000 (13:34 -0500)]
kvm: selftests: Do not indent with spaces

Some indentation with spaces crept in, likely due to terminal-based
cut and paste.  Clean it up.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agokvm: selftests: sync uapi/linux/kvm.h with Linux header
Paolo Bonzini [Wed, 19 Jan 2022 18:30:23 +0000 (13:30 -0500)]
kvm: selftests: sync uapi/linux/kvm.h with Linux header

KVM_CAP_XSAVE2 is out of sync due to a conflict.  Copy the whole
file while at it.

Reported-by: Yang Zhong <yang.zhong@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm: add amx_test to .gitignore
Muhammad Usama Anjum [Tue, 18 Jan 2022 12:20:52 +0000 (17:20 +0500)]
selftests: kvm: add amx_test to .gitignore

amx_test's binary should be present in the .gitignore file for the git
to ignore it.

Fixes: bf70636d9443 ("selftest: kvm: Add amx selftest")
Signed-off-by: Muhammad Usama Anjum <usama.anjum@collabora.com>
Message-Id: <20220118122053.1941915-1-usama.anjum@collabora.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Nullify vcpu_(un)blocking() hooks if AVIC is disabled
Sean Christopherson [Wed, 8 Dec 2021 01:52:33 +0000 (01:52 +0000)]
KVM: SVM: Nullify vcpu_(un)blocking() hooks if AVIC is disabled

Nullify svm_x86_ops.vcpu_(un)blocking if AVIC/APICv is disabled as the
hooks are necessary only to clear the vCPU's IsRunning entry in the
Physical APIC and to update IRTE entries if the VM has a pass-through
device attached.

Opportunistically rename the helpers to clarify their AVIC relationship.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-24-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Move svm_hardware_setup() and its helpers below svm_x86_ops
Sean Christopherson [Wed, 8 Dec 2021 01:52:32 +0000 (01:52 +0000)]
KVM: SVM: Move svm_hardware_setup() and its helpers below svm_x86_ops

Move svm_hardware_setup() below svm_x86_ops so that KVM can modify ops
during setup, e.g. the vcpu_(un)blocking hooks can be nullified if AVIC
is disabled or unsupported.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-23-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Drop AVIC's intermediate avic_set_running() helper
Sean Christopherson [Wed, 8 Dec 2021 01:52:31 +0000 (01:52 +0000)]
KVM: SVM: Drop AVIC's intermediate avic_set_running() helper

Drop avic_set_running() in favor of calling avic_vcpu_{load,put}()
directly, and modify the block+put path to use preempt_disable/enable()
instead of get/put_cpu(), as it doesn't actually care about the current
pCPU associated with the vCPU.  Opportunistically add lockdep assertions
as being preempted in avic_vcpu_put() would lead to consuming stale data,
even though doing so _in the current code base_ would not be fatal.

Add a much needed comment explaining why svm_vcpu_blocking() needs to
unload the AVIC and update the IRTE _before_ the vCPU starts blocking.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-22-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Don't do full kick when handling posted interrupt wakeup
Sean Christopherson [Wed, 8 Dec 2021 01:52:30 +0000 (01:52 +0000)]
KVM: VMX: Don't do full kick when handling posted interrupt wakeup

When waking vCPUs in the posted interrupt wakeup handling, do exactly
that and no more.  There is no need to kick the vCPU as the wakeup
handler just needs to get the vCPU task running, and if it's in the guest
then it's definitely running.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-21-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Fold fallback path into triggering posted IRQ helper
Sean Christopherson [Wed, 8 Dec 2021 01:52:29 +0000 (01:52 +0000)]
KVM: VMX: Fold fallback path into triggering posted IRQ helper

Move the fallback "wake_up" path into the helper to trigger posted
interrupt helper now that the nested and non-nested paths are identical.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-20-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Pass desired vector instead of bool for triggering posted IRQ
Sean Christopherson [Wed, 8 Dec 2021 01:52:28 +0000 (01:52 +0000)]
KVM: VMX: Pass desired vector instead of bool for triggering posted IRQ

Refactor the posted interrupt helper to take the desired notification
vector instead of a bool so that the callers are self-documenting.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-19-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Don't do full kick when triggering posted interrupt "fails"
Sean Christopherson [Wed, 8 Dec 2021 01:52:26 +0000 (01:52 +0000)]
KVM: VMX: Don't do full kick when triggering posted interrupt "fails"

Replace the full "kick" with just the "wake" in the fallback path when
triggering a virtual interrupt via a posted interrupt fails because the
guest is not IN_GUEST_MODE.  If the guest transitions into guest mode
between the check and the kick, then it's guaranteed to see the pending
interrupt as KVM syncs the PIR to IRR (and onto GUEST_RVI) after setting
IN_GUEST_MODE.  Kicking the guest in this case is nothing more than an
unnecessary VM-Exit (and host IRQ).

Opportunistically update comments to explain the various ordering rules
and barriers at play.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-17-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Skip AVIC and IRTE updates when loading blocking vCPU
Sean Christopherson [Wed, 8 Dec 2021 01:52:24 +0000 (01:52 +0000)]
KVM: SVM: Skip AVIC and IRTE updates when loading blocking vCPU

Don't bother updating the Physical APIC table or IRTE when loading a vCPU
that is blocking, i.e. won't be marked IsRun{ning}=1, as the pCPU is
queried if and only if IsRunning is '1'.  If the vCPU was migrated, the
new pCPU will be picked up when avic_vcpu_load() is called by
svm_vcpu_unblocking().

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-15-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Use kvm_vcpu_is_blocking() in AVIC load to handle preemption
Sean Christopherson [Wed, 8 Dec 2021 01:52:23 +0000 (01:52 +0000)]
KVM: SVM: Use kvm_vcpu_is_blocking() in AVIC load to handle preemption

Use kvm_vcpu_is_blocking() to determine whether or not the vCPU should be
marked running during avic_vcpu_load().  Drop avic_is_running, which
really should have been named "vcpu_is_not_blocking", as it tracked if
the vCPU was blocking, not if it was actually running, e.g. it was set
during svm_create_vcpu() when the vCPU was obviously not running.

This is technically a teeny tiny functional change, as the vCPU will be
marked IsRunning=1 on being reloaded if the vCPU is preempted between
svm_vcpu_blocking() and prepare_to_rcuwait().  But that's a benign change
as the vCPU will be marked IsRunning=0 when KVM voluntarily schedules out
the vCPU.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-14-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Remove unnecessary APICv/AVIC update in vCPU unblocking path
Sean Christopherson [Wed, 8 Dec 2021 01:52:22 +0000 (01:52 +0000)]
KVM: SVM: Remove unnecessary APICv/AVIC update in vCPU unblocking path

Remove handling of KVM_REQ_APICV_UPDATE from svm_vcpu_unblocking(), it's
no longer needed as it was made obsolete by commit df7e4827c549 ("KVM:
SVM: call avic_vcpu_load/avic_vcpu_put when enabling/disabling AVIC").
Prior to that commit, the manual check was necessary to ensure the AVIC
stuff was updated by avic_set_running() when a request to enable APICv
became pending while the vCPU was blocking, as the request handling
itself would not do the update.  But, as evidenced by the commit, that
logic was flawed and subject to various races.

Now that svm_refresh_apicv_exec_ctrl() does avic_vcpu_load/put() in
response to an APICv status change, drop the manual check in the
unblocking path.

Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-13-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Don't bother checking for "running" AVIC when kicking for IPIs
Sean Christopherson [Wed, 8 Dec 2021 01:52:21 +0000 (01:52 +0000)]
KVM: SVM: Don't bother checking for "running" AVIC when kicking for IPIs

Drop the avic_vcpu_is_running() check when waking vCPUs in response to a
VM-Exit due to incomplete IPI delivery.  The check isn't wrong per se, but
it's not 100% accurate in the sense that it doesn't guarantee that the vCPU
was one of the vCPUs that didn't receive the IPI.

The check isn't required for correctness as blocking == !running in this
context.

From a performance perspective, waking a live task is not expensive as the
only moderately costly operation is a locked operation to temporarily
disable preemption.  And if that is indeed a performance issue,
kvm_vcpu_is_blocking() would be a better check than poking into the AVIC.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-12-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Signal AVIC doorbell iff vCPU is in guest mode
Sean Christopherson [Wed, 8 Dec 2021 01:52:20 +0000 (01:52 +0000)]
KVM: SVM: Signal AVIC doorbell iff vCPU is in guest mode

Signal the AVIC doorbell iff the vCPU is running in the guest.  If the vCPU
is not IN_GUEST_MODE, it's guaranteed to pick up any pending IRQs on the
next VMRUN, which unconditionally processes the vIRR.

Add comments to document the logic.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-11-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Remove defunct pre_block/post_block kvm_x86_ops hooks
Sean Christopherson [Wed, 8 Dec 2021 01:52:19 +0000 (01:52 +0000)]
KVM: x86: Remove defunct pre_block/post_block kvm_x86_ops hooks

Drop kvm_x86_ops' pre/post_block() now that all implementations are nops.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Unexport LAPIC's switch_to_{hv,sw}_timer() helpers
Sean Christopherson [Wed, 8 Dec 2021 01:52:18 +0000 (01:52 +0000)]
KVM: x86: Unexport LAPIC's switch_to_{hv,sw}_timer() helpers

Unexport switch_to_{hv,sw}_timer() now that common x86 handles the
transitions.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Move preemption timer <=> hrtimer dance to common x86
Sean Christopherson [Wed, 8 Dec 2021 01:52:17 +0000 (01:52 +0000)]
KVM: VMX: Move preemption timer <=> hrtimer dance to common x86

Handle the switch to/from the hypervisor/software timer when a vCPU is
blocking in common x86 instead of in VMX.  Even though VMX is the only
user of a hypervisor timer, the logic and all functions involved are
generic x86 (unless future CPUs do something completely different and
implement a hypervisor timer that runs regardless of mode).

Handling the switch in common x86 will allow for the elimination of the
pre/post_blocks hooks, and also lets KVM switch back to the hypervisor
timer if and only if it was in use (without additional params).  Add a
comment explaining why the switch cannot be deferred to kvm_sched_out()
or kvm_vcpu_block().

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Move x86 VMX's posted interrupt list_head to vcpu_vmx
Sean Christopherson [Wed, 8 Dec 2021 01:52:16 +0000 (01:52 +0000)]
KVM: Move x86 VMX's posted interrupt list_head to vcpu_vmx

Move the seemingly generic block_vcpu_list from kvm_vcpu to vcpu_vmx, and
rename the list and all associated variables to clarify that it tracks
the set of vCPU that need to be poked on a posted interrupt to the wakeup
vector.  The list is not used to track _all_ vCPUs that are blocking, and
the term "blocked" can be misleading as it may refer to a blocking
condition in the host or the guest, where as the PI wakeup case is
specifically for the vCPUs that are actively blocking from within the
guest.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Drop unused kvm_vcpu.pre_pcpu field
Sean Christopherson [Wed, 8 Dec 2021 01:52:15 +0000 (01:52 +0000)]
KVM: Drop unused kvm_vcpu.pre_pcpu field

Remove kvm_vcpu.pre_pcpu as it no longer has any users.  No functional
change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Handle PI descriptor updates during vcpu_put/load
Sean Christopherson [Wed, 8 Dec 2021 01:52:14 +0000 (01:52 +0000)]
KVM: VMX: Handle PI descriptor updates during vcpu_put/load

Move the posted interrupt pre/post_block logic into vcpu_put/load
respectively, using the kvm_vcpu_is_blocking() to determining whether or
not the wakeup handler needs to be set (and unset).  This avoids updating
the PI descriptor if halt-polling is successful, reduces the number of
touchpoints for updating the descriptor, and eliminates the confusing
behavior of intentionally leaving a "stale" PI.NDST when a blocking vCPU
is scheduled back in after preemption.

The downside is that KVM will do the PID update twice if the vCPU is
preempted after prepare_to_rcuwait() but before schedule(), but that's a
rare case (and non-existent on !PREEMPT kernels).

The notable wart is the need to send a self-IPI on the wakeup vector if
an outstanding notification is pending after configuring the wakeup
vector.  Ideally, KVM would just do a kvm_vcpu_wake_up() in this case,
but the scheduler doesn't support waking a task from its preemption
notifier callback, i.e. while the task is right in the middle of
being scheduled out.

Note, setting the wakeup vector before halt-polling is not necessary:
once the pending IRQ will be recorded in the PIR, kvm_vcpu_has_events()
will detect this (via kvm_cpu_get_interrupt(), kvm_apic_get_interrupt(),
apic_has_interrupt_for_ppr() and finally vmx_sync_pir_to_irr()) and
terminate the polling.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoMerge branch 'kvm-pi-raw-spinlock' into HEAD
Paolo Bonzini [Tue, 18 Jan 2022 09:45:32 +0000 (04:45 -0500)]
Merge branch 'kvm-pi-raw-spinlock' into HEAD

Bring in fix for VT-d posted interrupts before further changing the code in 5.17.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: avoid warning on s390 in mark_page_dirty
Christian Borntraeger [Thu, 13 Jan 2022 12:29:24 +0000 (13:29 +0100)]
KVM: avoid warning on s390 in mark_page_dirty

Avoid warnings on s390 like
[ 1801.980931] CPU: 12 PID: 117600 Comm: kworker/12:0 Tainted: G            E     5.17.0-20220113.rc0.git0.32ce2abb03cf.300.fc35.s390x+next #1
[ 1801.980938] Workqueue: events irqfd_inject [kvm]
[...]
[ 1801.981057] Call Trace:
[ 1801.981060]  [<000003ff805f0f5c>] mark_page_dirty_in_slot+0xa4/0xb0 [kvm]
[ 1801.981083]  [<000003ff8060e9fe>] adapter_indicators_set+0xde/0x268 [kvm]
[ 1801.981104]  [<000003ff80613c24>] set_adapter_int+0x64/0xd8 [kvm]
[ 1801.981124]  [<000003ff805fb9aa>] kvm_set_irq+0xc2/0x130 [kvm]
[ 1801.981144]  [<000003ff805f8d86>] irqfd_inject+0x76/0xa0 [kvm]
[ 1801.981164]  [<0000000175e56906>] process_one_work+0x1fe/0x470
[ 1801.981173]  [<0000000175e570a4>] worker_thread+0x64/0x498
[ 1801.981176]  [<0000000175e5ef2c>] kthread+0x10c/0x110
[ 1801.981180]  [<0000000175de73c8>] __ret_from_fork+0x40/0x58
[ 1801.981185]  [<000000017698440a>] ret_from_fork+0xa/0x40

when writing to a guest from an irqfd worker as long as we do not have
the dirty ring.

Signed-off-by: Christian Borntraeger <borntraeger@linux.ibm.com>
Reluctantly-acked-by: David Woodhouse <dwmw@amazon.co.uk>
Message-Id: <20220113122924.740496-1-borntraeger@linux.ibm.com>
Fixes: 2efd61a608b0 ("KVM: Warn if mark_page_dirty() is called without an active vCPU")
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Add a test to force emulation with a pending exception
Sean Christopherson [Tue, 28 Dec 2021 23:24:37 +0000 (23:24 +0000)]
KVM: selftests: Add a test to force emulation with a pending exception

Add a VMX specific test to verify that KVM doesn't explode if userspace
attempts KVM_RUN when emulation is required with a pending exception.
KVM VMX's emulation support for !unrestricted_guest punts exceptions to
userspace instead of attempting to synthesize the exception with all the
correct state (and stack switching, etc...).

Punting is acceptable as there's never been a request to support
injecting exceptions when emulating due to invalid state, but KVM has
historically assumed that userspace will do the right thing and either
clear the exception or kill the guest.  Deliberately do the opposite and
attempt to re-enter the guest with a pending exception and emulation
required to verify KVM continues to punt the combination to userspace,
e.g. doesn't explode, WARN, etc...

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211228232437.1875318-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Reject KVM_RUN if emulation is required with pending exception
Sean Christopherson [Tue, 28 Dec 2021 23:24:36 +0000 (23:24 +0000)]
KVM: VMX: Reject KVM_RUN if emulation is required with pending exception

Reject KVM_RUN if emulation is required (because VMX is running without
unrestricted guest) and an exception is pending, as KVM doesn't support
emulating exceptions except when emulating real mode via vm86.  The vCPU
is hosed either way, but letting KVM_RUN proceed triggers a WARN due to
the impossible condition.  Alternatively, the WARN could be removed, but
then userspace and/or KVM bugs would result in the vCPU silently running
in a bad state, which isn't very friendly to users.

Originally, the bug was hit by syzkaller with a nested guest as that
doesn't require kvm_intel.unrestricted_guest=0.  That particular flavor
is likely fixed by commit cd0e615c49e5 ("KVM: nVMX: Synthesize
TRIPLE_FAULT for L2 if emulation is required"), but it's trivial to
trigger the WARN with a non-nested guest, and userspace can likely force
bad state via ioctls() for a nested guest as well.

Checking for the impossible condition needs to be deferred until KVM_RUN
because KVM can't force specific ordering between ioctls.  E.g. clearing
exception.pending in KVM_SET_SREGS doesn't prevent userspace from setting
it in KVM_SET_VCPU_EVENTS, and disallowing KVM_SET_VCPU_EVENTS with
emulation_required would prevent userspace from queuing an exception and
then stuffing sregs.  Note, if KVM were to try and detect/prevent the
condition prior to KVM_RUN, handle_invalid_guest_state() and/or
handle_emulation_failure() would need to be modified to clear the pending
exception prior to exiting to userspace.

 ------------[ cut here ]------------
 WARNING: CPU: 6 PID: 137812 at arch/x86/kvm/vmx/vmx.c:1623 vmx_queue_exception+0x14f/0x160 [kvm_intel]
 CPU: 6 PID: 137812 Comm: vmx_invalid_nes Not tainted 5.15.2-7cc36c3e14ae-pop #279
 Hardware name: ASUS Q87M-E/Q87M-E, BIOS 1102 03/03/2014
 RIP: 0010:vmx_queue_exception+0x14f/0x160 [kvm_intel]
 Code: <0f> 0b e9 fd fe ff ff 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00
 RSP: 0018:ffffa45c83577d38 EFLAGS: 00010202
 RAX: 0000000000000003 RBX: 0000000080000006 RCX: 0000000000000006
 RDX: 0000000000000000 RSI: 0000000000010002 RDI: ffff9916af734000
 RBP: ffff9916af734000 R08: 0000000000000000 R09: 0000000000000000
 R10: 0000000000000000 R11: 0000000000000001 R12: 0000000000000006
 R13: 0000000000000000 R14: ffff9916af734038 R15: 0000000000000000
 FS:  00007f1e1a47c740(0000) GS:ffff99188fb80000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 00007f1e1a6a8008 CR3: 000000026f83b005 CR4: 00000000001726e0
 Call Trace:
  kvm_arch_vcpu_ioctl_run+0x13a2/0x1f20 [kvm]
  kvm_vcpu_ioctl+0x279/0x690 [kvm]
  __x64_sys_ioctl+0x83/0xb0
  do_syscall_64+0x3b/0xc0
  entry_SYSCALL_64_after_hwframe+0x44/0xae

Reported-by: syzbot+82112403ace4cbd780d8@syzkaller.appspotmail.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211228232437.1875318-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm/x86: Add test for KVM_SET_PMU_EVENT_FILTER
Jim Mattson [Sat, 15 Jan 2022 05:24:31 +0000 (21:24 -0800)]
selftests: kvm/x86: Add test for KVM_SET_PMU_EVENT_FILTER

Verify that the PMU event filter works as expected.

Note that the virtual PMU doesn't work as expected on AMD Zen CPUs (an
intercepted rdmsr is counted as a retired branch instruction), but the
PMU event filter does work.

Signed-off-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20220115052431.447232-7-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm/x86: Introduce x86_model()
Jim Mattson [Sat, 15 Jan 2022 05:24:30 +0000 (21:24 -0800)]
selftests: kvm/x86: Introduce x86_model()

Extract the x86 model number from CPUID.01H:EAX.

Signed-off-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20220115052431.447232-6-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm/x86: Export x86_family() for use outside of processor.c
Jim Mattson [Sat, 15 Jan 2022 05:24:29 +0000 (21:24 -0800)]
selftests: kvm/x86: Export x86_family() for use outside of processor.c

Move this static inline function to processor.h, so that it can be
used in individual tests, as needed.

Opportunistically replace the bare 'unsigned' with 'unsigned int.'

Signed-off-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20220115052431.447232-5-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm/x86: Introduce is_amd_cpu()
Jim Mattson [Sat, 15 Jan 2022 05:24:28 +0000 (21:24 -0800)]
selftests: kvm/x86: Introduce is_amd_cpu()

Replace the one ad hoc "AuthenticAMD" CPUID vendor string comparison
with a new function, is_amd_cpu().

Signed-off-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20220115052431.447232-4-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: kvm/x86: Parameterize the CPUID vendor string check
Jim Mattson [Sat, 15 Jan 2022 05:24:27 +0000 (21:24 -0800)]
selftests: kvm/x86: Parameterize the CPUID vendor string check

Refactor is_intel_cpu() to make it easier to reuse the bulk of the
code for other vendors in the future.

Signed-off-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20220115052431.447232-3-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/pmu: Use binary search to check filtered events
Jim Mattson [Sat, 15 Jan 2022 05:24:26 +0000 (21:24 -0800)]
KVM: x86/pmu: Use binary search to check filtered events

The PMU event filter may contain up to 300 events. Replace the linear
search in reprogram_gp_counter() with a binary search.

Signed-off-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20220115052431.447232-2-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agokvm: selftests: conditionally build vm_xsave_req_perm()
Wei Wang [Tue, 18 Jan 2022 01:48:17 +0000 (20:48 -0500)]
kvm: selftests: conditionally build vm_xsave_req_perm()

vm_xsave_req_perm() is currently defined and used by x86_64 only.
Make it compiled into vm_create_with_vcpus() only when on x86_64
machines. Otherwise, it would cause linkage errors, e.g. on s390x.

Fixes: 415a3c33e8 ("kvm: selftests: Add support for KVM_CAP_XSAVE2")
Reported-by: Janis Schoetterl-Glausch <scgl@linux.ibm.com>
Signed-off-by: Wei Wang <wei.w.wang@intel.com>
Tested-by: Janis Schoetterl-Glausch <scgl@linux.ibm.com>
Message-Id: <20220118014817.30910-1-wei.w.wang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/cpuid: Clear XFD for component i if the base feature is missing
Like Xu [Mon, 17 Jan 2022 07:45:31 +0000 (15:45 +0800)]
KVM: x86/cpuid: Clear XFD for component i if the base feature is missing

According to Intel extended feature disable (XFD) spec, the sub-function i
(i > 1) of CPUID function 0DH enumerates "details for state component i.
ECX[2] enumerates support for XFD support for this state component."

If KVM does not report F(XFD) feature (e.g. due to CONFIG_X86_64),
then the corresponding XFD support for any state component i
should also be removed. Translate this dependency into KVM terms.

Fixes: 690a757d610e ("kvm: x86: Add CPUID support for Intel AMX")
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220117074531.76925-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Improve TLB flush comment in kvm_mmu_slot_remove_write_access()
David Matlack [Thu, 13 Jan 2022 23:30:20 +0000 (23:30 +0000)]
KVM: x86/mmu: Improve TLB flush comment in kvm_mmu_slot_remove_write_access()

Rewrite the comment in kvm_mmu_slot_remove_write_access() that explains
why it is safe to flush TLBs outside of the MMU lock after
write-protecting SPTEs for dirty logging. The current comment is a long
run-on sentence that was difficult to understand. In addition it was
specific to the shadow MMU (mentioning mmu_spte_update()) when the TDP
MMU has to handle this as well.

The new comment explains:
 - Why the TLB flush is necessary at all.
 - Why it is desirable to do the TLB flush outside of the MMU lock.
 - Why it is safe to do the TLB flush outside of the MMU lock.

No functional change intended.

Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220113233020.3986005-5-dmatlack@google.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Document and enforce MMU-writable and Host-writable invariants
David Matlack [Thu, 13 Jan 2022 23:30:19 +0000 (23:30 +0000)]
KVM: x86/mmu: Document and enforce MMU-writable and Host-writable invariants

SPTEs are tagged with software-only bits to indicate if it is
"MMU-writable" and "Host-writable". These bits are used to determine why
KVM has marked an SPTE as read-only.

Document these bits and their invariants, and enforce the invariants
with new WARNs in spte_can_locklessly_be_made_writable() to ensure they
are not accidentally violated in the future.

Opportunistically move DEFAULT_SPTE_{MMU,HOST}_WRITABLE next to
EPT_SPTE_{MMU,HOST}_WRITABLE since the new documentation applies to
both.

No functional change intended.

Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220113233020.3986005-4-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Clear MMU-writable during changed_pte notifier
David Matlack [Thu, 13 Jan 2022 23:30:18 +0000 (23:30 +0000)]
KVM: x86/mmu: Clear MMU-writable during changed_pte notifier

When handling the changed_pte notifier and the new PTE is read-only,
clear both the Host-writable and MMU-writable bits in the SPTE. This
preserves the invariant that MMU-writable is set if-and-only-if
Host-writable is set.

No functional change intended. Nothing currently relies on the
aforementioned invariant and technically the changed_pte notifier is
dead code.

Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220113233020.3986005-3-dmatlack@google.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Fix write-protection of PTs mapped by the TDP MMU
David Matlack [Thu, 13 Jan 2022 23:30:17 +0000 (23:30 +0000)]
KVM: x86/mmu: Fix write-protection of PTs mapped by the TDP MMU

When the TDP MMU is write-protection GFNs for page table protection (as
opposed to for dirty logging, or due to the HVA not being writable), it
checks if the SPTE is already write-protected and if so skips modifying
the SPTE and the TLB flush.

This behavior is incorrect because it fails to check if the SPTE
is write-protected for page table protection, i.e. fails to check
that MMU-writable is '0'.  If the SPTE was write-protected for dirty
logging but not page table protection, the SPTE could locklessly be made
writable, and vCPUs could still be running with writable mappings cached
in their TLB.

Fix this by only skipping setting the SPTE if the SPTE is already
write-protected *and* MMU-writable is already clear.  Technically,
checking only MMU-writable would suffice; a SPTE cannot be writable
without MMU-writable being set.  But check both to be paranoid and
because it arguably yields more readable code.

Fixes: 46044f72c382 ("kvm: x86/mmu: Support write protection for nesting in tdp MMU")
Cc: stable@vger.kernel.org
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220113233020.3986005-2-dmatlack@google.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: switch blocked_vcpu_on_cpu_lock to raw spinlock
Marcelo Tosatti [Tue, 18 Jan 2022 09:34:43 +0000 (04:34 -0500)]
KVM: VMX: switch blocked_vcpu_on_cpu_lock to raw spinlock

blocked_vcpu_on_cpu_lock is taken from hard interrupt context
(pi_wakeup_handler), therefore it cannot sleep.

Switch it to a raw spinlock.

Fixes:

[41297.066254] BUG: scheduling while atomic: CPU 0/KVM/635218/0x00010001
[41297.066323] Preemption disabled at:
[41297.066324] [<ffffffff902ee47f>] irq_enter_rcu+0xf/0x60
[41297.066339] Call Trace:
[41297.066342]  <IRQ>
[41297.066346]  dump_stack_lvl+0x34/0x44
[41297.066353]  ? irq_enter_rcu+0xf/0x60
[41297.066356]  __schedule_bug.cold+0x7d/0x8b
[41297.066361]  __schedule+0x439/0x5b0
[41297.066365]  ? task_blocks_on_rt_mutex.constprop.0.isra.0+0x1b0/0x440
[41297.066369]  schedule_rtlock+0x1e/0x40
[41297.066371]  rtlock_slowlock_locked+0xf1/0x260
[41297.066374]  rt_spin_lock+0x3b/0x60
[41297.066378]  pi_wakeup_handler+0x31/0x90 [kvm_intel]
[41297.066388]  sysvec_kvm_posted_intr_wakeup_ipi+0x9d/0xd0
[41297.066392]  </IRQ>
[41297.066392]  asm_sysvec_kvm_posted_intr_wakeup_ipi+0x12/0x20
...

Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Making the module parameter of vPMU more common
Like Xu [Tue, 11 Jan 2022 07:38:23 +0000 (15:38 +0800)]
KVM: x86: Making the module parameter of vPMU more common

The new module parameter to control PMU virtualization should apply
to Intel as well as AMD, for situations where userspace is not trusted.
If the module parameter allows PMU virtualization, there could be a
new KVM_CAP or guest CPUID bits whereby userspace can enable/disable
PMU virtualization on a per-VM basis.

If the module parameter does not allow PMU virtualization, there
should be no userspace override, since we have no precedent for
authorizing that kind of override. If it's false, other counter-based
profiling features (such as LBR including the associated CPUID bits
if any) will not be exposed.

Change its name from "pmu" to "enable_pmu" as we have temporary
variables with the same name in our code like "struct kvm_pmu *pmu".

Fixes: b1d66dad65dc ("KVM: x86/svm: Add module param to control PMU virtualization")
Suggested-by : Jim Mattson <jmattson@google.com>
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220111073823.21885-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Test KVM_SET_CPUID2 after KVM_RUN
Vitaly Kuznetsov [Mon, 17 Jan 2022 15:05:42 +0000 (16:05 +0100)]
KVM: selftests: Test KVM_SET_CPUID2 after KVM_RUN

KVM forbids KVM_SET_CPUID2 after KVM_RUN was performed on a vCPU unless
the supplied CPUID data is equal to what was previously set. Test this.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220117150542.2176196-5-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Rename 'get_cpuid_test' to 'cpuid_test'
Vitaly Kuznetsov [Mon, 17 Jan 2022 15:05:41 +0000 (16:05 +0100)]
KVM: selftests: Rename 'get_cpuid_test' to 'cpuid_test'

In preparation to reusing the existing 'get_cpuid_test' for testing
"KVM_SET_CPUID{,2} after KVM_RUN" rename it to 'cpuid_test' to avoid
the confusion.

No functional change intended.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220117150542.2176196-4-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Partially allow KVM_SET_CPUID{,2} after KVM_RUN
Vitaly Kuznetsov [Mon, 17 Jan 2022 15:05:40 +0000 (16:05 +0100)]
KVM: x86: Partially allow KVM_SET_CPUID{,2} after KVM_RUN

Commit feb627e8d6f6 ("KVM: x86: Forbid KVM_SET_CPUID{,2} after KVM_RUN")
forbade changing CPUID altogether but unfortunately this is not fully
compatible with existing VMMs. In particular, QEMU reuses vCPU fds for
CPU hotplug after unplug and it calls KVM_SET_CPUID2. Instead of full ban,
check whether the supplied CPUID data is equal to what was previously set.

Reported-by: Igor Mammedov <imammedo@redhat.com>
Fixes: feb627e8d6f6 ("KVM: x86: Forbid KVM_SET_CPUID{,2} after KVM_RUN")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220117150542.2176196-3-vkuznets@redhat.com>
Cc: stable@vger.kernel.org
[Do not call kvm_find_cpuid_entry repeatedly. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Do runtime CPUID update before updating vcpu->arch.cpuid_entries
Vitaly Kuznetsov [Mon, 17 Jan 2022 15:05:39 +0000 (16:05 +0100)]
KVM: x86: Do runtime CPUID update before updating vcpu->arch.cpuid_entries

kvm_update_cpuid_runtime() mangles CPUID data coming from userspace
VMM after updating 'vcpu->arch.cpuid_entries', this makes it
impossible to compare an update with what was previously
supplied. Introduce __kvm_update_cpuid_runtime() version which can be
used to tweak the input before it goes to 'vcpu->arch.cpuid_entries'
so the upcoming update check can compare tweaked data.

No functional change intended.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220117150542.2176196-2-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/pmu: Fix available_event_types check for REF_CPU_CYCLES event
Like Xu [Wed, 5 Jan 2022 05:15:09 +0000 (13:15 +0800)]
KVM: x86/pmu: Fix available_event_types check for REF_CPU_CYCLES event

According to CPUID 0x0A.EBX bit vector, the event [7] should be the
unrealized event "Topdown Slots" instead of the *kernel* generalized
common hardware event "REF_CPU_CYCLES", so we need to skip the cpuid
unavaliblity check in the intel_pmc_perf_hw_id() for the last
REF_CPU_CYCLES event and update the confusing comment.

If the event is marked as unavailable in the Intel guest CPUID
0AH.EBX leaf, we need to avoid any perf_event creation, whether
it's a gp or fixed counter. To distinguish whether it is a rejected
event or an event that needs to be programmed with PERF_TYPE_RAW type,
a new special returned value of "PERF_COUNT_HW_MAX + 1" is introduced.

Fixes: 62079d8a43128 ("KVM: PMU: add proper support for fixed counter 2")
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220105051509.69437-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agox86/fpu: Fix inline prefix warnings
Yang Zhong [Thu, 13 Jan 2022 18:08:25 +0000 (13:08 -0500)]
x86/fpu: Fix inline prefix warnings

Fix sparse warnings in xstate and remove inline prefix.

Fixes: 980fe2fddcff ("x86/fpu: Extend fpu_xstate_prctl() with guest permissions")
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
Reported-by: kernel test robot <lkp@intel.com>
Message-Id: <20220113180825.322333-1-yang.zhong@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftest: kvm: Add amx selftest
Yang Zhong [Thu, 23 Dec 2021 14:53:22 +0000 (09:53 -0500)]
selftest: kvm: Add amx selftest

This selftest covers two aspects of AMX.  The first is triggering #NM
exception and checking the MSR XFD_ERR value.  The second case is
loading tile config and tile data into guest registers and trapping to
the host side for a complete save/load of the guest state.  TMM0
is also checked against memory data after save/restore.

Signed-off-by: Yang Zhong <yang.zhong@intel.com>
Message-Id: <20211223145322.2914028-4-yang.zhong@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftest: kvm: Move struct kvm_x86_state to header
Yang Zhong [Thu, 23 Dec 2021 14:53:21 +0000 (09:53 -0500)]
selftest: kvm: Move struct kvm_x86_state to header

Those changes can avoid dereferencing pointer compile issue
when amx_test.c reference state->xsave.

Move struct kvm_x86_state definition to processor.h.

Signed-off-by: Yang Zhong <yang.zhong@intel.com>
Message-Id: <20211223145322.2914028-3-yang.zhong@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftest: kvm: Reorder vcpu_load_state steps for AMX
Paolo Bonzini [Thu, 23 Dec 2021 14:53:20 +0000 (09:53 -0500)]
selftest: kvm: Reorder vcpu_load_state steps for AMX

For AMX support it is recommended to load XCR0 after XFD, so
that KVM does not see XFD=0, XCR=1 for a save state that will
eventually be disabled (which would lead to premature allocation
of the space required for that save state).

It is also required to load XSAVE data after XCR0 and XFD, so
that KVM can trigger allocation of the extra space required to
store AMX state.

Adjust vcpu_load_state to obey these new requirements.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
Message-Id: <20211223145322.2914028-2-yang.zhong@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agokvm: x86: Disable interception for IA32_XFD on demand
Kevin Tian [Wed, 5 Jan 2022 12:35:32 +0000 (04:35 -0800)]
kvm: x86: Disable interception for IA32_XFD on demand

Always intercepting IA32_XFD causes non-negligible overhead when this
register is updated frequently in the guest.

Disable r/w emulation after intercepting the first WRMSR(IA32_XFD)
with a non-zero value.

Disable WRMSR emulation implies that IA32_XFD becomes out-of-sync
with the software states in fpstate and the per-cpu xfd cache. This
leads to two additional changes accordingly:

  - Call fpu_sync_guest_vmexit_xfd_state() after vm-exit to bring
    software states back in-sync with the MSR, before handle_exit_irqoff()
    is called.

  - Always trap #NM once write interception is disabled for IA32_XFD.
    The #NM exception is rare if the guest doesn't use dynamic
    features. Otherwise, there is at most one exception per guest
    task given a dynamic feature.

p.s. We have confirmed that SDM is being revised to say that
when setting IA32_XFD[18] the AMX register state is not guaranteed
to be preserved. This clarification avoids adding mess for a creative
guest which sets IA32_XFD[18]=1 before saving active AMX state to
its own storage.

Signed-off-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Jing Liu <jing2.liu@intel.com>
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
Message-Id: <20220105123532.12586-22-yang.zhong@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agox86/fpu: Provide fpu_sync_guest_vmexit_xfd_state()
Thomas Gleixner [Wed, 5 Jan 2022 12:35:31 +0000 (04:35 -0800)]
x86/fpu: Provide fpu_sync_guest_vmexit_xfd_state()

KVM can disable the write emulation for the XFD MSR when the vCPU's fpstate
is already correctly sized to reduce the overhead.

When write emulation is disabled the XFD MSR state after a VMEXIT is
unknown and therefore not in sync with the software states in fpstate and
the per CPU XFD cache.

Provide fpu_sync_guest_vmexit_xfd_state() which has to be invoked after a
VMEXIT before enabling interrupts when write emulation is disabled for the
XFD MSR.

It could be invoked unconditionally even when write emulation is enabled
for the price of a pointless MSR read.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Jing Liu <jing2.liu@intel.com>
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
Message-Id: <20220105123532.12586-21-yang.zhong@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agokvm: selftests: Add support for KVM_CAP_XSAVE2
Wei Wang [Wed, 5 Jan 2022 12:35:30 +0000 (04:35 -0800)]
kvm: selftests: Add support for KVM_CAP_XSAVE2

When KVM_CAP_XSAVE2 is supported, userspace is expected to allocate
buffer for KVM_GET_XSAVE2 and KVM_SET_XSAVE using the size returned
by KVM_CHECK_EXTENSION(KVM_CAP_XSAVE2).

Signed-off-by: Wei Wang <wei.w.wang@intel.com>
Signed-off-by: Guang Zeng <guang.zeng@intel.com>
Signed-off-by: Jing Liu <jing2.liu@intel.com>
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
Message-Id: <20220105123532.12586-20-yang.zhong@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>