linux-2.6-microblaze.git
2 years agoMerge tag 'kvm-s390-next-5.17-1' of git://git.kernel.org/pub/scm/linux/kernel/git...
Paolo Bonzini [Tue, 21 Dec 2021 17:59:53 +0000 (12:59 -0500)]
Merge tag 'kvm-s390-next-5.17-1' of git://git./linux/kernel/git/kvms390/linux into HEAD

KVM: s390: Fix and cleanup

- fix sigp sense/start/stop/inconsistency
- cleanups

2 years agoMerge remote-tracking branch 'kvm/master' into HEAD
Paolo Bonzini [Tue, 21 Dec 2021 17:51:09 +0000 (12:51 -0500)]
Merge remote-tracking branch 'kvm/master' into HEAD

Pick commit fdba608f15e2 ("KVM: VMX: Wake vCPU when delivering posted
IRQ even if vCPU == this vCPU").  In addition to fixing a bug, it
also aligns the non-nested and nested usage of triggering posted
interrupts, allowing for additional cleanups.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Wake vCPU when delivering posted IRQ even if vCPU == this vCPU
Sean Christopherson [Tue, 21 Dec 2021 15:37:00 +0000 (10:37 -0500)]
KVM: VMX: Wake vCPU when delivering posted IRQ even if vCPU == this vCPU

Drop a check that guards triggering a posted interrupt on the currently
running vCPU, and more importantly guards waking the target vCPU if
triggering a posted interrupt fails because the vCPU isn't IN_GUEST_MODE.
If a vIRQ is delivered from asynchronous context, the target vCPU can be
the currently running vCPU and can also be blocking, in which case
skipping kvm_vcpu_wake_up() is effectively dropping what is supposed to
be a wake event for the vCPU.

The "do nothing" logic when "vcpu == running_vcpu" mostly works only
because the majority of calls to ->deliver_posted_interrupt(), especially
when using posted interrupts, come from synchronous KVM context.  But if
a device is exposed to the guest using vfio-pci passthrough, the VFIO IRQ
and vCPU are bound to the same pCPU, and the IRQ is _not_ configured to
use posted interrupts, wake events from the device will be delivered to
KVM from IRQ context, e.g.

  vfio_msihandler()
  |
  |-> eventfd_signal()
      |
      |-> ...
          |
          |->  irqfd_wakeup()
               |
               |->kvm_arch_set_irq_inatomic()
                  |
                  |-> kvm_irq_delivery_to_apic_fast()
                      |
                      |-> kvm_apic_set_irq()

This also aligns the non-nested and nested usage of triggering posted
interrupts, and will allow for additional cleanups.

Fixes: 379a3c8ee444 ("KVM: VMX: Optimize posted-interrupt delivery for timer fastpath")
Cc: stable@vger.kernel.org
Reported-by: Longpeng (Mike) <longpeng2@huawei.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211208015236.1616697-18-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Add test to verify TRIPLE_FAULT on invalid L2 guest state
Sean Christopherson [Tue, 7 Dec 2021 19:30:06 +0000 (19:30 +0000)]
KVM: selftests: Add test to verify TRIPLE_FAULT on invalid L2 guest state

Add a selftest to attempt to enter L2 with invalid guests state by
exiting to userspace via I/O from L2, and then using KVM_SET_SREGS to set
invalid guest state (marking TR unusable is arbitrary chosen for its
relative simplicity).

This is a regression test for a bug introduced by commit c8607e4a086f
("KVM: x86: nVMX: don't fail nested VM entry on invalid guest state if
!from_vmentry"), which incorrectly set vmx->fail=true when L2 had invalid
guest state and ultimately triggered a WARN due to nested_vmx_vmexit()
seeing vmx->fail==true while attempting to synthesize a nested VM-Exit.

The is also a functional test to verify that KVM sythesizes TRIPLE_FAULT
for L2, which is somewhat arbitrary behavior, instead of emulating L2.
KVM should never emulate L2 due to invalid guest state, as it's
architecturally impossible for L1 to run an L2 guest with invalid state
as nested VM-Enter should always fail, i.e. L1 needs to do the emulation.
Stuffing state via KVM ioctl() is a non-architctural, out-of-band case,
hence the TRIPLE_FAULT being rather arbitrary.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211207193006.120997-5-seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Fix stale docs for kvm-intel.emulate_invalid_guest_state
Sean Christopherson [Tue, 7 Dec 2021 19:30:05 +0000 (19:30 +0000)]
KVM: VMX: Fix stale docs for kvm-intel.emulate_invalid_guest_state

Update the documentation for kvm-intel's emulate_invalid_guest_state to
rectify the description of KVM's default behavior, and to document that
the behavior and thus parameter only applies to L1.

Fixes: a27685c33acc ("KVM: VMX: Emulate invalid guest state by default")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211207193006.120997-4-seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Synthesize TRIPLE_FAULT for L2 if emulation is required
Sean Christopherson [Tue, 7 Dec 2021 19:30:04 +0000 (19:30 +0000)]
KVM: nVMX: Synthesize TRIPLE_FAULT for L2 if emulation is required

Synthesize a triple fault if L2 guest state is invalid at the time of
VM-Enter, which can happen if L1 modifies SMRAM or if userspace stuffs
guest state via ioctls(), e.g. KVM_SET_SREGS.  KVM should never emulate
invalid guest state, since from L1's perspective, it's architecturally
impossible for L2 to have invalid state while L2 is running in hardware.
E.g. attempts to set CR0 or CR4 to unsupported values will either VM-Exit
or #GP.

Modifying vCPU state via RSM+SMRAM and ioctl() are the only paths that
can trigger this scenario, as nested VM-Enter correctly rejects any
attempt to enter L2 with invalid state.

RSM is a straightforward case as (a) KVM follows AMD's SMRAM layout and
behavior, and (b) Intel's SDM states that loading reserved CR0/CR4 bits
via RSM results in shutdown, i.e. there is precedent for KVM's behavior.
Following AMD's SMRAM layout is important as AMD's layout saves/restores
the descriptor cache information, including CS.RPL and SS.RPL, and also
defines all the fields relevant to invalid guest state as read-only, i.e.
so long as the vCPU had valid state before the SMI, which is guaranteed
for L2, RSM will generate valid state unless SMRAM was modified.  Intel's
layout saves/restores only the selector, which means that scenarios where
the selector and cached RPL don't match, e.g. conforming code segments,
would yield invalid guest state.  Intel CPUs fudge around this issued by
stuffing SS.RPL and CS.RPL on RSM.  Per Intel's SDM on the "Default
Treatment of RSM", paraphrasing for brevity:

  IF internal storage indicates that the [CPU was post-VMXON]
  THEN
     enter VMX operation (root or non-root);
     restore VMX-critical state as defined in Section 34.14.1;
     set to their fixed values any bits in CR0 and CR4 whose values must
     be fixed in VMX operation [unless coming from an unrestricted guest];
     IF RFLAGS.VM = 0 AND (in VMX root operation OR the
        “unrestricted guest” VM-execution control is 0)
     THEN
       CS.RPL := SS.DPL;
       SS.RPL := SS.DPL;
     FI;
     restore current VMCS pointer;
  FI;

Note that Intel CPUs also overwrite the fixed CR0/CR4 bits, whereas KVM
will sythesize TRIPLE_FAULT in this scenario.  KVM's behavior is allowed
as both Intel and AMD define CR0/CR4 SMRAM fields as read-only, i.e. the
only way for CR0 and/or CR4 to have illegal values is if they were
modified by the L1 SMM handler, and Intel's SDM "SMRAM State Save Map"
section states "modifying these registers will result in unpredictable
behavior".

KVM's ioctl() behavior is less straightforward.  Because KVM allows
ioctls() to be executed in any order, rejecting an ioctl() if it would
result in invalid L2 guest state is not an option as KVM cannot know if
a future ioctl() would resolve the invalid state, e.g. KVM_SET_SREGS, or
drop the vCPU out of L2, e.g. KVM_SET_NESTED_STATE.  Ideally, KVM would
reject KVM_RUN if L2 contained invalid guest state, but that carries the
risk of a false positive, e.g. if RSM loaded invalid guest state and KVM
exited to userspace.  Setting a flag/request to detect such a scenario is
undesirable because (a) it's extremely unlikely to add value to KVM as a
whole, and (b) KVM would need to consider ioctl() interactions with such
a flag, e.g. if userspace migrated the vCPU while the flag were set.

Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211207193006.120997-3-seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Always clear vmx->fail on emulation_required
Sean Christopherson [Tue, 7 Dec 2021 19:30:03 +0000 (19:30 +0000)]
KVM: VMX: Always clear vmx->fail on emulation_required

Revert a relatively recent change that set vmx->fail if the vCPU is in L2
and emulation_required is true, as that behavior is completely bogus.
Setting vmx->fail and synthesizing a VM-Exit is contradictory and wrong:

  (a) it's impossible to have both a VM-Fail and VM-Exit
  (b) vmcs.EXIT_REASON is not modified on VM-Fail
  (c) emulation_required refers to guest state and guest state checks are
      always VM-Exits, not VM-Fails.

For KVM specifically, emulation_required is handled before nested exits
in __vmx_handle_exit(), thus setting vmx->fail has no immediate effect,
i.e. KVM calls into handle_invalid_guest_state() and vmx->fail is ignored.
Setting vmx->fail can ultimately result in a WARN in nested_vmx_vmexit()
firing when tearing down the VM as KVM never expects vmx->fail to be set
when L2 is active, KVM always reflects those errors into L1.

  ------------[ cut here ]------------
  WARNING: CPU: 0 PID: 21158 at arch/x86/kvm/vmx/nested.c:4548
                                nested_vmx_vmexit+0x16bd/0x17e0
                                arch/x86/kvm/vmx/nested.c:4547
  Modules linked in:
  CPU: 0 PID: 21158 Comm: syz-executor.1 Not tainted 5.16.0-rc3-syzkaller #0
  Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
  RIP: 0010:nested_vmx_vmexit+0x16bd/0x17e0 arch/x86/kvm/vmx/nested.c:4547
  Code: <0f> 0b e9 2e f8 ff ff e8 57 b3 5d 00 0f 0b e9 00 f1 ff ff 89 e9 80
  Call Trace:
   vmx_leave_nested arch/x86/kvm/vmx/nested.c:6220 [inline]
   nested_vmx_free_vcpu+0x83/0xc0 arch/x86/kvm/vmx/nested.c:330
   vmx_free_vcpu+0x11f/0x2a0 arch/x86/kvm/vmx/vmx.c:6799
   kvm_arch_vcpu_destroy+0x6b/0x240 arch/x86/kvm/x86.c:10989
   kvm_vcpu_destroy+0x29/0x90 arch/x86/kvm/../../../virt/kvm/kvm_main.c:441
   kvm_free_vcpus arch/x86/kvm/x86.c:11426 [inline]
   kvm_arch_destroy_vm+0x3ef/0x6b0 arch/x86/kvm/x86.c:11545
   kvm_destroy_vm arch/x86/kvm/../../../virt/kvm/kvm_main.c:1189 [inline]
   kvm_put_kvm+0x751/0xe40 arch/x86/kvm/../../../virt/kvm/kvm_main.c:1220
   kvm_vcpu_release+0x53/0x60 arch/x86/kvm/../../../virt/kvm/kvm_main.c:3489
   __fput+0x3fc/0x870 fs/file_table.c:280
   task_work_run+0x146/0x1c0 kernel/task_work.c:164
   exit_task_work include/linux/task_work.h:32 [inline]
   do_exit+0x705/0x24f0 kernel/exit.c:832
   do_group_exit+0x168/0x2d0 kernel/exit.c:929
   get_signal+0x1740/0x2120 kernel/signal.c:2852
   arch_do_signal_or_restart+0x9c/0x730 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+0x191/0x220 kernel/entry/common.c:207
   __syscall_exit_to_user_mode_work kernel/entry/common.c:289 [inline]
   syscall_exit_to_user_mode+0x2e/0x70 kernel/entry/common.c:300
   do_syscall_64+0x53/0xd0 arch/x86/entry/common.c:86
   entry_SYSCALL_64_after_hwframe+0x44/0xae

Fixes: c8607e4a086f ("KVM: x86: nVMX: don't fail nested VM entry on invalid guest state if !from_vmentry")
Reported-by: syzbot+f1d2136db9c80d4733e8@syzkaller.appspotmail.com
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211207193006.120997-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: KVM: Fix non-x86 compiling
Andrew Jones [Tue, 14 Dec 2021 15:18:42 +0000 (16:18 +0100)]
selftests: KVM: Fix non-x86 compiling

Attempting to compile on a non-x86 architecture fails with

include/kvm_util.h: In function â€˜vm_compute_max_gfn’:
include/kvm_util.h:79:21: error: dereferencing pointer to incomplete type â€˜struct kvm_vm’
  return ((1ULL << vm->pa_bits) >> vm->page_shift) - 1;
                     ^~

This is because the declaration of struct kvm_vm is in
lib/kvm_util_internal.h as an effort to make it private to
the test lib code. We can still provide arch specific functions,
though, by making the generic function symbols weak. Do that to
fix the compile error.

Fixes: c8cc43c1eae2 ("selftests: KVM: avoid failures due to reserved HyperTransport region")
Cc: stable@vger.kernel.org
Signed-off-by: Andrew Jones <drjones@redhat.com>
Message-Id: <20211214151842.848314-1-drjones@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Always set kvm_run->if_flag
Marc Orr [Thu, 9 Dec 2021 15:52:57 +0000 (07:52 -0800)]
KVM: x86: Always set kvm_run->if_flag

The kvm_run struct's if_flag is a part of the userspace/kernel API. The
SEV-ES patches failed to set this flag because it's no longer needed by
QEMU (according to the comment in the source code). However, other
hypervisors may make use of this flag. Therefore, set the flag for
guests with encrypted registers (i.e., with guest_state_protected set).

Fixes: f1c6366e3043 ("KVM: SVM: Add required changes to support intercepts under SEV-ES")
Signed-off-by: Marc Orr <marcorr@google.com>
Message-Id: <20211209155257.128747-1-marcorr@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
2 years agoKVM: x86/mmu: Don't advance iterator after restart due to yielding
Sean Christopherson [Tue, 14 Dec 2021 03:35:28 +0000 (03:35 +0000)]
KVM: x86/mmu: Don't advance iterator after restart due to yielding

After dropping mmu_lock in the TDP MMU, restart the iterator during
tdp_iter_next() and do not advance the iterator.  Advancing the iterator
results in skipping the top-level SPTE and all its children, which is
fatal if any of the skipped SPTEs were not visited before yielding.

When zapping all SPTEs, i.e. when min_level == root_level, restarting the
iter and then invoking tdp_iter_next() is always fatal if the current gfn
has as a valid SPTE, as advancing the iterator results in try_step_side()
skipping the current gfn, which wasn't visited before yielding.

Sprinkle WARNs on iter->yielded being true in various helpers that are
often used in conjunction with yielding, and tag the helper with
__must_check to reduce the probabily of improper usage.

Failing to zap a top-level SPTE manifests in one of two ways.  If a valid
SPTE is skipped by both kvm_tdp_mmu_zap_all() and kvm_tdp_mmu_put_root(),
the shadow page will be leaked and KVM will WARN accordingly.

  WARNING: CPU: 1 PID: 3509 at arch/x86/kvm/mmu/tdp_mmu.c:46 [kvm]
  RIP: 0010:kvm_mmu_uninit_tdp_mmu+0x3e/0x50 [kvm]
  Call Trace:
   <TASK>
   kvm_arch_destroy_vm+0x130/0x1b0 [kvm]
   kvm_destroy_vm+0x162/0x2a0 [kvm]
   kvm_vcpu_release+0x34/0x60 [kvm]
   __fput+0x82/0x240
   task_work_run+0x5c/0x90
   do_exit+0x364/0xa10
   ? futex_unqueue+0x38/0x60
   do_group_exit+0x33/0xa0
   get_signal+0x155/0x850
   arch_do_signal_or_restart+0xed/0x750
   exit_to_user_mode_prepare+0xc5/0x120
   syscall_exit_to_user_mode+0x1d/0x40
   do_syscall_64+0x48/0xc0
   entry_SYSCALL_64_after_hwframe+0x44/0xae

If kvm_tdp_mmu_zap_all() skips a gfn/SPTE but that SPTE is then zapped by
kvm_tdp_mmu_put_root(), KVM triggers a use-after-free in the form of
marking a struct page as dirty/accessed after it has been put back on the
free list.  This directly triggers a WARN due to encountering a page with
page_count() == 0, but it can also lead to data corruption and additional
errors in the kernel.

  WARNING: CPU: 7 PID: 1995658 at arch/x86/kvm/../../../virt/kvm/kvm_main.c:171
  RIP: 0010:kvm_is_zone_device_pfn.part.0+0x9e/0xd0 [kvm]
  Call Trace:
   <TASK>
   kvm_set_pfn_dirty+0x120/0x1d0 [kvm]
   __handle_changed_spte+0x92e/0xca0 [kvm]
   __handle_changed_spte+0x63c/0xca0 [kvm]
   __handle_changed_spte+0x63c/0xca0 [kvm]
   __handle_changed_spte+0x63c/0xca0 [kvm]
   zap_gfn_range+0x549/0x620 [kvm]
   kvm_tdp_mmu_put_root+0x1b6/0x270 [kvm]
   mmu_free_root_page+0x219/0x2c0 [kvm]
   kvm_mmu_free_roots+0x1b4/0x4e0 [kvm]
   kvm_mmu_unload+0x1c/0xa0 [kvm]
   kvm_arch_destroy_vm+0x1f2/0x5c0 [kvm]
   kvm_put_kvm+0x3b1/0x8b0 [kvm]
   kvm_vcpu_release+0x4e/0x70 [kvm]
   __fput+0x1f7/0x8c0
   task_work_run+0xf8/0x1a0
   do_exit+0x97b/0x2230
   do_group_exit+0xda/0x2a0
   get_signal+0x3be/0x1e50
   arch_do_signal_or_restart+0x244/0x17f0
   exit_to_user_mode_prepare+0xcb/0x120
   syscall_exit_to_user_mode+0x1d/0x40
   do_syscall_64+0x4d/0x90
   entry_SYSCALL_64_after_hwframe+0x44/0xae

Note, the underlying bug existed even before commit 1af4a96025b3 ("KVM:
x86/mmu: Yield in TDU MMU iter even if no SPTES changed") moved calls to
tdp_mmu_iter_cond_resched() to the beginning of loops, as KVM could still
incorrectly advance past a top-level entry when yielding on a lower-level
entry.  But with respect to leaking shadow pages, the bug was introduced
by yielding before processing the current gfn.

Alternatively, tdp_mmu_iter_cond_resched() could simply fall through, or
callers could jump to their "retry" label.  The downside of that approach
is that tdp_mmu_iter_cond_resched() _must_ be called before anything else
in the loop, and there's no easy way to enfornce that requirement.

Ideally, KVM would handling the cond_resched() fully within the iterator
macro (the code is actually quite clean) and avoid this entire class of
bugs, but that is extremely difficult do while also supporting yielding
after tdp_mmu_set_spte_atomic() fails.  Yielding after failing to set a
SPTE is very desirable as the "owner" of the REMOVED_SPTE isn't strictly
bounded, e.g. if it's zapping a high-level shadow page, the REMOVED_SPTE
may block operations on the SPTE for a significant amount of time.

Fixes: faaf05b00aec ("kvm: x86/mmu: Support zapping SPTEs in the TDP MMU")
Fixes: 1af4a96025b3 ("KVM: x86/mmu: Yield in TDU MMU iter even if no SPTES changed")
Reported-by: Ignat Korchagin <ignat@cloudflare.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211214033528.123268-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: remove PMU FIXED_CTR3 from msrs_to_save_all
Wei Wang [Fri, 17 Dec 2021 12:49:34 +0000 (07:49 -0500)]
KVM: x86: remove PMU FIXED_CTR3 from msrs_to_save_all

The fixed counter 3 is used for the Topdown metrics, which hasn't been
enabled for KVM guests. Userspace accessing to it will fail as it's not
included in get_fixed_pmc(). This breaks KVM selftests on ICX+ machines,
which have this counter.

To reproduce it on ICX+ machines, ./state_test reports:
==== Test Assertion Failure ====
lib/x86_64/processor.c:1078: r == nmsrs
pid=4564 tid=4564 - Argument list too long
1  0x000000000040b1b9: vcpu_save_state at processor.c:1077
2  0x0000000000402478: main at state_test.c:209 (discriminator 6)
3  0x00007fbe21ed5f92: ?? ??:0
4  0x000000000040264d: _start at ??:?
 Unexpected result from KVM_GET_MSRS, r: 17 (failed MSR was 0x30c)

With this patch, it works well.

Signed-off-by: Wei Wang <wei.w.wang@intel.com>
Message-Id: <20211217124934.32893-1-wei.w.wang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Retry page fault if MMU reload is pending and root has no sp
Sean Christopherson [Thu, 9 Dec 2021 06:05:46 +0000 (06:05 +0000)]
KVM: x86: Retry page fault if MMU reload is pending and root has no sp

Play nice with a NULL shadow page when checking for an obsolete root in
the page fault handler by flagging the page fault as stale if there's no
shadow page associated with the root and KVM_REQ_MMU_RELOAD is pending.
Invalidating memslots, which is the only case where _all_ roots need to
be reloaded, requests all vCPUs to reload their MMUs while holding
mmu_lock for lock.

The "special" roots, e.g. pae_root when KVM uses PAE paging, are not
backed by a shadow page.  Running with TDP disabled or with nested NPT
explodes spectaculary due to dereferencing a NULL shadow page pointer.

Skip the KVM_REQ_MMU_RELOAD check if there is a valid shadow page for the
root.  Zapping shadow pages in response to guest activity, e.g. when the
guest frees a PGD, can trigger KVM_REQ_MMU_RELOAD even if the current
vCPU isn't using the affected root.  I.e. KVM_REQ_MMU_RELOAD can be seen
with a completely valid root shadow page.  This is a bit of a moot point
as KVM currently unloads all roots on KVM_REQ_MMU_RELOAD, but that will
be cleaned up in the future.

Fixes: a955cad84cda ("KVM: x86/mmu: Retry page fault if root is invalidated by memslot update")
Cc: stable@vger.kernel.org
Cc: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211209060552.2956723-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: vmx_pmu_msrs_test: Drop tests mangling guest visible CPUIDs
Vitaly Kuznetsov [Thu, 16 Dec 2021 16:52:12 +0000 (17:52 +0100)]
KVM: selftests: vmx_pmu_msrs_test: Drop tests mangling guest visible CPUIDs

Host initiated writes to MSR_IA32_PERF_CAPABILITIES should not depend
on guest visible CPUIDs and (incorrect) KVM logic implementing it is
about to change. Also, KVM_SET_CPUID{,2} after KVM_RUN is now forbidden
and causes test to fail.

Reported-by: kernel test robot <oliver.sang@intel.com>
Fixes: feb627e8d6f6 ("KVM: x86: Forbid KVM_SET_CPUID{,2} after KVM_RUN")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211216165213.338923-2-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Drop guest CPUID check for host initiated writes to MSR_IA32_PERF_CAPABILITIES
Vitaly Kuznetsov [Thu, 16 Dec 2021 16:52:13 +0000 (17:52 +0100)]
KVM: x86: Drop guest CPUID check for host initiated writes to MSR_IA32_PERF_CAPABILITIES

The ability to write to MSR_IA32_PERF_CAPABILITIES from the host should
not depend on guest visible CPUID entries, even if just to allow
creating/restoring guest MSRs and CPUIDs in any sequence.

Fixes: 27461da31089 ("KVM: x86/pmu: Support full width counting")
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211216165213.338923-3-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoMerge branch 'topic/ppc-kvm' of https://git.kernel.org/pub/scm/linux/kernel/git/power...
Paolo Bonzini [Sun, 19 Dec 2021 14:27:21 +0000 (15:27 +0100)]
Merge branch 'topic/ppc-kvm' of https://git./linux/kernel/git/powerpc/linux into HEAD

Fix conflicts between memslot overhaul and commit 511d25d6b789f ("KVM:
PPC: Book3S: Suppress warnings when allocating too big memory slots")
from the powerpc tree.

2 years agoKVM: s390: Clarify SIGP orders versus STOP/RESTART
Eric Farman [Mon, 13 Dec 2021 21:05:50 +0000 (22:05 +0100)]
KVM: s390: Clarify SIGP orders versus STOP/RESTART

With KVM_CAP_S390_USER_SIGP, there are only five Signal Processor
orders (CONDITIONAL EMERGENCY SIGNAL, EMERGENCY SIGNAL, EXTERNAL CALL,
SENSE, and SENSE RUNNING STATUS) which are intended for frequent use
and thus are processed in-kernel. The remainder are sent to userspace
with the KVM_CAP_S390_USER_SIGP capability. Of those, three orders
(RESTART, STOP, and STOP AND STORE STATUS) have the potential to
inject work back into the kernel, and thus are asynchronous.

Let's look for those pending IRQs when processing one of the in-kernel
SIGP orders, and return BUSY (CC2) if one is in process. This is in
agreement with the Principles of Operation, which states that only one
order can be "active" on a CPU at a time.

Cc: stable@vger.kernel.org
Suggested-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Eric Farman <farman@linux.ibm.com>
Reviewed-by: Christian Borntraeger <borntraeger@linux.ibm.com>
Acked-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/20211213210550.856213-2-farman@linux.ibm.com
[borntraeger@linux.ibm.com: add stable tag]
Signed-off-by: Christian Borntraeger <borntraeger@linux.ibm.com>
2 years agos390: uv: Add offset comments to UV query struct and fix naming
Janosch Frank [Mon, 7 Jun 2021 08:07:13 +0000 (08:07 +0000)]
s390: uv: Add offset comments to UV query struct and fix naming

Changes to the struct are easier to manage with offset comments so
let's add some. And now that we know that the last struct member has
the wrong name let's also fix this.

Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
2 years agoKVM: s390: gaccess: Cleanup access to guest pages
Janis Schoetterl-Glausch [Fri, 26 Nov 2021 16:45:49 +0000 (17:45 +0100)]
KVM: s390: gaccess: Cleanup access to guest pages

Introduce a helper function for guest frame access.

Signed-off-by: Janis Schoetterl-Glausch <scgl@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Message-Id: <20211126164549.7046-4-scgl@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
2 years agoKVM: s390: gaccess: Refactor access address range check
Janis Schoetterl-Glausch [Fri, 26 Nov 2021 16:45:48 +0000 (17:45 +0100)]
KVM: s390: gaccess: Refactor access address range check

Do not round down the first address to the page boundary, just translate
it normally, which gives the value we care about in the first place.
Given this, translating a single address is just the special case of
translating a range spanning a single page.

Make the output optional, so the function can be used to just check a
range.

Signed-off-by: Janis Schoetterl-Glausch <scgl@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Message-Id: <20211126164549.7046-3-scgl@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
2 years agoKVM: s390: gaccess: Refactor gpa and length calculation
Janis Schoetterl-Glausch [Fri, 26 Nov 2021 16:45:47 +0000 (17:45 +0100)]
KVM: s390: gaccess: Refactor gpa and length calculation

Improve readability by renaming the length variable and
not calculating the offset manually.

Signed-off-by: Janis Schoetterl-Glausch <scgl@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Message-Id: <20211126164549.7046-2-scgl@linux.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
2 years agoKVM: PPC: Book3S HV P9: Use kvm_arch_vcpu_get_wait() to get rcuwait object
Sean Christopherson [Mon, 13 Dec 2021 17:45:56 +0000 (17:45 +0000)]
KVM: PPC: Book3S HV P9: Use kvm_arch_vcpu_get_wait() to get rcuwait object

Use kvm_arch_vcpu_get_wait() to get a vCPU's rcuwait object instead of
using vcpu->wait directly in kvmhv_run_single_vcpu().  Functionally, this
is a nop as vcpu->arch.waitp is guaranteed to point at vcpu->wait.  But
that is not obvious at first glance, and a future change coming in via
the KVM tree, commit 510958e99721 ("KVM: Force PPC to define its own
rcuwait object"), will hide vcpu->wait from architectures that define
__KVM_HAVE_ARCH_WQP to prevent generic KVM from attepting to wake a vCPU
with the wrong rcuwait object.

Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211213174556.3871157-1-seanjc@google.com
2 years agoselftests: KVM: Add test to verify KVM doesn't explode on "bad" I/O
Sean Christopherson [Mon, 25 Oct 2021 20:13:11 +0000 (13:13 -0700)]
selftests: KVM: Add test to verify KVM doesn't explode on "bad" I/O

Add an x86 selftest to verify that KVM doesn't WARN or otherwise explode
if userspace modifies RCX during a userspace exit to handle string I/O.
This is a regression test for a user-triggerable WARN introduced by
commit 3b27de271839 ("KVM: x86: split the two parts of emulator_pio_in").

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211025201311.1881846-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Don't WARN if userspace mucks with RCX during string I/O exit
Sean Christopherson [Mon, 25 Oct 2021 20:13:10 +0000 (13:13 -0700)]
KVM: x86: Don't WARN if userspace mucks with RCX during string I/O exit

Replace a WARN with a comment to call out that userspace can modify RCX
during an exit to userspace to handle string I/O.  KVM doesn't actually
support changing the rep count during an exit, i.e. the scenario can be
ignored, but the WARN needs to go as it's trivial to trigger from
userspace.

Cc: stable@vger.kernel.org
Fixes: 3b27de271839 ("KVM: x86: split the two parts of emulator_pio_in")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211025201311.1881846-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Raise #GP when clearing CR0_PG in 64 bit mode
Lai Jiangshan [Tue, 7 Dec 2021 09:52:30 +0000 (17:52 +0800)]
KVM: X86: Raise #GP when clearing CR0_PG in 64 bit mode

In the SDM:
If the logical processor is in 64-bit mode or if CR4.PCIDE = 1, an
attempt to clear CR0.PG causes a general-protection exception (#GP).
Software should transition to compatibility mode and clear CR4.PCIDE
before attempting to disable paging.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211207095230.53437-1-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoselftests: KVM: avoid failures due to reserved HyperTransport region
Paolo Bonzini [Thu, 5 Aug 2021 10:54:23 +0000 (06:54 -0400)]
selftests: KVM: avoid failures due to reserved HyperTransport region

AMD proceessors define an address range that is reserved by HyperTransport
and causes a failure if used for guest physical addresses.  Avoid
selftests failures by reserving those guest physical addresses; the
rules are:

- On parts with <40 bits, its fully hidden from software.

- Before Fam17h, it was always 12G just below 1T, even if there was more
RAM above this location.  In this case we just not use any RAM above 1T.

- On Fam17h and later, it is variable based on SME, and is either just
below 2^48 (no encryption) or 2^43 (encryption).

Fixes: ef4c9f4f6546 ("KVM: selftests: Fix 32-bit truncation of vm_get_max_gfn()")
Cc: stable@vger.kernel.org
Cc: David Matlack <dmatlack@google.com>
Reported-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20210805105423.412878-1-pbonzini@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Tested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Ignore sparse banks size for an "all CPUs", non-sparse IPI req
Sean Christopherson [Tue, 7 Dec 2021 22:09:19 +0000 (22:09 +0000)]
KVM: x86: Ignore sparse banks size for an "all CPUs", non-sparse IPI req

Do not bail early if there are no bits set in the sparse banks for a
non-sparse, a.k.a. "all CPUs", IPI request.  Per the Hyper-V spec, it is
legal to have a variable length of '0', e.g. VP_SET's BankContents in
this case, if the request can be serviced without the extra info.

  It is possible that for a given invocation of a hypercall that does
  accept variable sized input headers that all the header input fits
  entirely within the fixed size header. In such cases the variable sized
  input header is zero-sized and the corresponding bits in the hypercall
  input should be set to zero.

Bailing early results in KVM failing to send IPIs to all CPUs as expected
by the guest.

Fixes: 214ff83d4473 ("KVM: x86: hyperv: implement PV IPI send hypercalls")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211207220926.718794-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Wait for IPIs to be delivered when handling Hyper-V TLB flush hypercall
Vitaly Kuznetsov [Thu, 9 Dec 2021 10:29:37 +0000 (11:29 +0100)]
KVM: x86: Wait for IPIs to be delivered when handling Hyper-V TLB flush hypercall

Prior to commit 0baedd792713 ("KVM: x86: make Hyper-V PV TLB flush use
tlb_flush_guest()"), kvm_hv_flush_tlb() was using 'KVM_REQ_TLB_FLUSH |
KVM_REQUEST_NO_WAKEUP' when making a request to flush TLBs on other vCPUs
and KVM_REQ_TLB_FLUSH is/was defined as:

 (0 | KVM_REQUEST_WAIT | KVM_REQUEST_NO_WAKEUP)

so KVM_REQUEST_WAIT was lost. Hyper-V TLFS, however, requires that
"This call guarantees that by the time control returns back to the
caller, the observable effects of all flushes on the specified virtual
processors have occurred." and without KVM_REQUEST_WAIT there's a small
chance that the vCPU making the TLB flush will resume running before
all IPIs get delivered to other vCPUs and a stale mapping can get read
there.

Fix the issue by adding KVM_REQUEST_WAIT flag to KVM_REQ_TLB_FLUSH_GUEST:
kvm_hv_flush_tlb() is the sole caller which uses it for
kvm_make_all_cpus_request()/kvm_make_vcpus_request_mask() where
KVM_REQUEST_WAIT makes a difference.

Cc: stable@kernel.org
Fixes: 0baedd792713 ("KVM: x86: make Hyper-V PV TLB flush use tlb_flush_guest()")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211209102937.584397-1-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: arm64: Use Makefile.kvm for common files
David Woodhouse [Sun, 21 Nov 2021 12:54:46 +0000 (12:54 +0000)]
KVM: arm64: Use Makefile.kvm for common files

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Acked-by: Marc Zyngier <maz@kernel.org>
Message-Id: <20211121125451.9489-8-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: powerpc: Use Makefile.kvm for common files
David Woodhouse [Sun, 21 Nov 2021 12:54:45 +0000 (12:54 +0000)]
KVM: powerpc: Use Makefile.kvm for common files

It's all fairly baroque but in the end, I don't think there's any reason
for $(KVM)/irqchip.o to have been handled differently, as they all end
up in $(kvm-y) in the end anyway, regardless of whether they get there
via $(common-objs-y) and the CPU-specific object lists.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Acked-by: Michael Ellerman <mpe@ellerman.id.au> (powerpc)
Message-Id: <20211121125451.9489-7-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: RISC-V: Use Makefile.kvm for common files
David Woodhouse [Sun, 21 Nov 2021 12:54:44 +0000 (12:54 +0000)]
KVM: RISC-V: Use Makefile.kvm for common files

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Message-Id: <20211121125451.9489-6-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: mips: Use Makefile.kvm for common files
David Woodhouse [Sun, 21 Nov 2021 12:54:43 +0000 (12:54 +0000)]
KVM: mips: Use Makefile.kvm for common files

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Message-Id: <20211121125451.9489-5-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: s390: Use Makefile.kvm for common files
David Woodhouse [Sun, 21 Nov 2021 12:54:42 +0000 (12:54 +0000)]
KVM: s390: Use Makefile.kvm for common files

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Message-Id: <20211121125451.9489-4-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Add Makefile.kvm for common files, use it for x86
David Woodhouse [Sun, 21 Nov 2021 12:54:41 +0000 (12:54 +0000)]
KVM: Add Makefile.kvm for common files, use it for x86

Splitting kvm_main.c out into smaller and better-organized files is
slightly non-trivial when it involves editing a bunch of per-arch
KVM makefiles. Provide virt/kvm/Makefile.kvm for them to include.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Acked-by: Marc Zyngier <maz@kernel.org>
Message-Id: <20211121125451.9489-3-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Introduce CONFIG_HAVE_KVM_DIRTY_RING
David Woodhouse [Sun, 21 Nov 2021 12:54:40 +0000 (12:54 +0000)]
KVM: Introduce CONFIG_HAVE_KVM_DIRTY_RING

I'd like to make the build include dirty_ring.c based on whether the
arch wants it or not. That's a whole lot simpler if there's a config
symbol instead of doing it implicitly on KVM_DIRTY_LOG_PAGE_OFFSET
being set to something non-zero.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Message-Id: <20211121125451.9489-2-dwmw2@infradead.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: selftests: svm_int_ctl_test: fix intercept calculation
Maciej S. Szmigiero [Thu, 2 Dec 2021 23:10:13 +0000 (00:10 +0100)]
KVM: x86: selftests: svm_int_ctl_test: fix intercept calculation

INTERCEPT_x are bit positions, but the code was using the raw value of
INTERCEPT_VINTR (4) instead of BIT(INTERCEPT_VINTR).
This resulted in masking of bit 2 - that is, SMI instead of VINTR.

Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Message-Id: <49b9571d25588870db5380b0be1a41df4bbaaf93.1638486479.git.maciej.szmigiero@oracle.com>

2 years agoKVM: VMX: Clean up PI pre/post-block WARNs
Sean Christopherson [Wed, 8 Dec 2021 01:52:13 +0000 (01:52 +0000)]
KVM: VMX: Clean up PI pre/post-block WARNs

Move the WARN sanity checks out of the PI descriptor update loop so as
not to spam the kernel log if the condition is violated and the update
takes multiple attempts due to another writer.  This also eliminates a
few extra uops from the retry path.

Technically not checking every attempt could mean KVM will now fail to
WARN in a scenario that would have failed before, but any such failure
would be inherently racy as some other agent (CPU or device) would have
to concurrent modify the PI descriptor.

Add a helper to handle the actual write and more importantly to document
why the write may need to be retried.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Ensure vCPU honors event request if posting nested IRQ fails
Sean Christopherson [Wed, 8 Dec 2021 01:52:12 +0000 (01:52 +0000)]
KVM: nVMX: Ensure vCPU honors event request if posting nested IRQ fails

Add a memory barrier between writing vcpu->requests and reading
vcpu->guest_mode to ensure the read is ordered after the write when
(potentially) delivering an IRQ to L2 via nested posted interrupt.  If
the request were to be completed after reading vcpu->mode, it would be
possible for the target vCPU to enter the guest without posting the
interrupt and without handling the event request.

Note, the barrier is only for documentation since atomic operations are
serializing on x86.

Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Fixes: 6b6977117f50 ("KVM: nVMX: Fix races when sending nested PI while dest enters/leaves L2")
Fixes: 705699a13994 ("KVM: nVMX: Enable nested posted interrupt processing")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211208015236.1616697-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: add a tracepoint for APICv/AVIC interrupt delivery
Maxim Levitsky [Thu, 9 Dec 2021 11:54:36 +0000 (13:54 +0200)]
KVM: x86: add a tracepoint for APICv/AVIC interrupt delivery

This allows to see how many interrupts were delivered via the
APICv/AVIC from the host.

Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20211209115440.394441-3-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Implement Enlightened MSR Bitmap feature
Vitaly Kuznetsov [Mon, 29 Nov 2021 09:47:04 +0000 (10:47 +0100)]
KVM: nVMX: Implement Enlightened MSR Bitmap feature

Updating MSR bitmap for L2 is not cheap and rearly needed. TLFS for Hyper-V
offers 'Enlightened MSR Bitmap' feature which allows L1 hypervisor to
inform L0 when it changes MSR bitmap, this eliminates the need to examine
L1's MSR bitmap for L2 every time when 'real' MSR bitmap for L2 gets
constructed.

Use 'vmx->nested.msr_bitmap_changed' flag to implement the feature.

Note, KVM already uses 'Enlightened MSR bitmap' feature when it runs as a
nested hypervisor on top of Hyper-V. The newly introduced feature is going
to be used by Hyper-V guests on KVM.

When the feature is enabled for Win10+WSL2, it shaves off around 700 CPU
cycles from a nested vmexit cost (tight cpuid loop test).

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211129094704.326635-5-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Track whether changes in L0 require MSR bitmap for L2 to be rebuilt
Vitaly Kuznetsov [Mon, 29 Nov 2021 09:47:03 +0000 (10:47 +0100)]
KVM: nVMX: Track whether changes in L0 require MSR bitmap for L2 to be rebuilt

Introduce a flag to keep track of whether MSR bitmap for L2 needs to be
rebuilt due to changes in MSR bitmap for L1 or switching to a different
L2. This information will be used for Enlightened MSR Bitmap feature for
Hyper-V guests.

Note, setting msr_bitmap_changed to 'true' from set_current_vmptr() is
not really needed for Enlightened MSR Bitmap as the feature can only
be used in conjunction with Enlightened VMCS but let's keep tracking
information complete, it's cheap and in the future similar PV feature can
easily be implemented for KVM on KVM too.

No functional change intended.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211129094704.326635-4-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Introduce vmx_msr_bitmap_l01_changed() helper
Vitaly Kuznetsov [Mon, 29 Nov 2021 09:47:02 +0000 (10:47 +0100)]
KVM: VMX: Introduce vmx_msr_bitmap_l01_changed() helper

In preparation to enabling 'Enlightened MSR Bitmap' feature for Hyper-V
guests move MSR bitmap update tracking to a dedicated helper.

Note: vmx_msr_bitmap_l01_changed() is called when MSR bitmap might be
updated. KVM doesn't check if the bit we're trying to set is already set
(or the bit it's trying to clear is already cleared). Such situations
should not be common and a few false positives should not be a problem.

No functional change intended.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211129094704.326635-3-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoMerge branch 'kvm-on-hv-msrbm-fix' into HEAD
Paolo Bonzini [Wed, 8 Dec 2021 10:29:16 +0000 (05:29 -0500)]
Merge branch 'kvm-on-hv-msrbm-fix' into HEAD

Merge bugfix for enlightened MSR Bitmap, before adding support
to KVM for exposing the feature to nested guests.

2 years agoKVM: x86: Exit to userspace if emulation prepared a completion callback
Hou Wenlong [Tue, 2 Nov 2021 09:15:32 +0000 (17:15 +0800)]
KVM: x86: Exit to userspace if emulation prepared a completion callback

em_rdmsr() and em_wrmsr() return X86EMUL_IO_NEEDED if MSR accesses
required an exit to userspace. However, x86_emulate_insn() doesn't return
X86EMUL_*, so x86_emulate_instruction() doesn't directly act on
X86EMUL_IO_NEEDED; instead, it looks for other signals to differentiate
between PIO, MMIO, etc. causing RDMSR/WRMSR emulation not to
exit to userspace now.

Nevertheless, if the userspace_msr_exit_test testcase in selftests
is changed to test RDMSR/WRMSR with a forced emulation prefix,
the test passes.  What happens is that first userspace exit
information is filled but the userspace exit does not happen.
Because x86_emulate_instruction() returns 1, the guest retries
the instruction---but this time RIP has already been adjusted
past the forced emulation prefix, so the guest executes RDMSR/WRMSR
and the userspace exit finally happens.

Since the X86EMUL_IO_NEEDED path has provided a complete_userspace_io
callback, x86_emulate_instruction() can just return 0 if the
callback is not NULL. Then RDMSR/WRMSR instruction emulation will
exit to userspace directly, without the RDMSR/WRMSR vmexit.

Fixes: 1ae099540e8c7 ("KVM: x86: Allow deflecting unknown MSR accesses to user space")
Signed-off-by: Hou Wenlong <houwenlong93@linux.alibaba.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <56f9df2ee5c05a81155e2be366c9dc1f7adc8817.1635842679.git.houwenlong93@linux.alibaba.com>

2 years agoKVM: nVMX: Don't use Enlightened MSR Bitmap for L3
Vitaly Kuznetsov [Mon, 29 Nov 2021 09:47:01 +0000 (10:47 +0100)]
KVM: nVMX: Don't use Enlightened MSR Bitmap for L3

When KVM runs as a nested hypervisor on top of Hyper-V it uses Enlightened
VMCS and enables Enlightened MSR Bitmap feature for its L1s and L2s (which
are actually L2s and L3s from Hyper-V's perspective). When MSR bitmap is
updated, KVM has to reset HV_VMX_ENLIGHTENED_CLEAN_FIELD_MSR_BITMAP from
clean fields to make Hyper-V aware of the change. For KVM's L1s, this is
done in vmx_disable_intercept_for_msr()/vmx_enable_intercept_for_msr().
MSR bitmap for L2 is build in nested_vmx_prepare_msr_bitmap() by blending
MSR bitmap for L1 and L1's idea of MSR bitmap for L2. KVM, however, doesn't
check if the resulting bitmap is different and never cleans
HV_VMX_ENLIGHTENED_CLEAN_FIELD_MSR_BITMAP in eVMCS02. This is incorrect and
may result in Hyper-V missing the update.

The issue could've been solved by calling evmcs_touch_msr_bitmap() for
eVMCS02 from nested_vmx_prepare_msr_bitmap() unconditionally but doing so
would not give any performance benefits (compared to not using Enlightened
MSR Bitmap at all). 3-level nesting is also not a very common setup
nowadays.

Don't enable 'Enlightened MSR Bitmap' feature for KVM's L2s (real L3s) for
now.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20211129094704.326635-2-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Use different callback if msr access comes from the emulator
Hou Wenlong [Tue, 2 Nov 2021 09:15:31 +0000 (17:15 +0800)]
KVM: x86: Use different callback if msr access comes from the emulator

If msr access triggers an exit to userspace, the
complete_userspace_io callback would skip instruction by vendor
callback for kvm_skip_emulated_instruction(). However, when msr
access comes from the emulator, e.g. if kvm.force_emulation_prefix
is enabled and the guest uses rdmsr/wrmsr with kvm prefix,
VM_EXIT_INSTRUCTION_LEN in vmcs is invalid and
kvm_emulate_instruction() should be used to skip instruction
instead.

As Sean noted, unlike the previous case, there's no #UD if
unrestricted guest is disabled and the guest accesses an MSR in
Big RM. So the correct way to fix this is to attach a different
callback when the msr access comes from the emulator.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Hou Wenlong <houwenlong93@linux.alibaba.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <34208da8f51580a06e45afefac95afea0e3f96e3.1635842679.git.houwenlong93@linux.alibaba.com>

2 years agoKVM: x86: Add an emulation type to handle completion of user exits
Hou Wenlong [Tue, 2 Nov 2021 09:15:30 +0000 (17:15 +0800)]
KVM: x86: Add an emulation type to handle completion of user exits

The next patch would use kvm_emulate_instruction() with
EMULTYPE_SKIP in complete_userspace_io callback to fix a
problem in msr access emulation. However, EMULTYPE_SKIP
only updates RIP, more things like updating interruptibility
state and injecting single-step #DBs would be done in the
callback. Since the emulator also does those things after
x86_emulate_insn(), add a new emulation type to pair with
EMULTYPE_SKIP to do those things for completion of user exits
within the emulator.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Hou Wenlong <houwenlong93@linux.alibaba.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <8f8c8e268b65f31d55c2881a4b30670946ecfa0d.1635842679.git.houwenlong93@linux.alibaba.com>

2 years agoKVM: x86: Handle 32-bit wrap of EIP for EMULTYPE_SKIP with flat code seg
Sean Christopherson [Tue, 2 Nov 2021 09:15:29 +0000 (17:15 +0800)]
KVM: x86: Handle 32-bit wrap of EIP for EMULTYPE_SKIP with flat code seg

Truncate the new EIP to a 32-bit value when handling EMULTYPE_SKIP as the
decode phase does not truncate _eip.  Wrapping the 32-bit boundary is
legal if and only if CS is a flat code segment, but that check is
implicitly handled in the form of limit checks in the decode phase.

Opportunstically prepare for a future fix by storing the result of any
truncation in "eip" instead of "_eip".

Fixes: 1957aa63be53 ("KVM: VMX: Handle single-step #DB for EMULTYPE_SKIP on EPT misconfig")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <093eabb1eab2965201c9b018373baf26ff256d85.1635842679.git.houwenlong93@linux.alibaba.com>

2 years agoKVM: Clear pv eoi pending bit only when it is set
Li RongQing [Thu, 4 Nov 2021 11:56:14 +0000 (19:56 +0800)]
KVM: Clear pv eoi pending bit only when it is set

merge pv_eoi_get_pending and pv_eoi_clr_pending into a single
function pv_eoi_test_and_clear_pending, which returns and clear
the value of the pending bit.

This makes it possible to clear the pending bit only if the guest
had set it, and otherwise skip the call to pv_eoi_put_user().
This can save up to 300 nsec on AMD EPYC processors.

Suggested-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Message-Id: <1636026974-50555-2-git-send-email-lirongqing@baidu.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: don't print when fail to read/write pv eoi memory
Li RongQing [Thu, 4 Nov 2021 11:56:13 +0000 (19:56 +0800)]
KVM: x86: don't print when fail to read/write pv eoi memory

If guest gives MSR_KVM_PV_EOI_EN a wrong value, this printk() will
be trigged, and kernel log is spammed with the useless message

Fixes: 0d88800d5472 ("kvm: x86: ioapic and apic debug macros cleanup")
Reported-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Cc: stable@kernel.org
Message-Id: <1636026974-50555-1-git-send-email-lirongqing@baidu.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Remove mmu parameter from load_pdptrs()
Lai Jiangshan [Wed, 24 Nov 2021 12:20:52 +0000 (20:20 +0800)]
KVM: X86: Remove mmu parameter from load_pdptrs()

It uses vcpu->arch.walk_mmu always; nested EPT does not have PDPTRs,
and nested NPT treats them like all other non-leaf page table levels
instead of caching them.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211124122055.64424-11-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Rename gpte_is_8_bytes to has_4_byte_gpte and invert the direction
Lai Jiangshan [Wed, 24 Nov 2021 12:20:51 +0000 (20:20 +0800)]
KVM: X86: Rename gpte_is_8_bytes to has_4_byte_gpte and invert the direction

This bit is very close to mean "role.quadrant is not in use", except that
it is false also when the MMU is mapping guest physical addresses
directly.  In that case, role.quadrant is indeed not in use, but there
are no guest PTEs at all.

Changing the name and direction of the bit removes the special case,
since a guest with paging disabled, or not considering guest paging
structures as is the case for two-dimensional paging, does not have
to deal with 4-byte guest PTEs.

Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211124122055.64424-10-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Use ept_caps_to_lpage_level() in hardware_setup()
Lai Jiangshan [Wed, 24 Nov 2021 12:20:50 +0000 (20:20 +0800)]
KVM: VMX: Use ept_caps_to_lpage_level() in hardware_setup()

Using ept_caps_to_lpage_level is simpler.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211124122055.64424-9-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Add parameter huge_page_level to kvm_init_shadow_ept_mmu()
Lai Jiangshan [Wed, 24 Nov 2021 12:20:49 +0000 (20:20 +0800)]
KVM: X86: Add parameter huge_page_level to kvm_init_shadow_ept_mmu()

The level of supported large page on nEPT affects the rsvds_bits_mask.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211124122055.64424-8-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Add huge_page_level to __reset_rsvds_bits_mask_ept()
Lai Jiangshan [Wed, 24 Nov 2021 12:20:48 +0000 (20:20 +0800)]
KVM: X86: Add huge_page_level to __reset_rsvds_bits_mask_ept()

Bit 7 on pte depends on the level of supported large page.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211124122055.64424-7-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Remove mmu->translate_gpa
Lai Jiangshan [Wed, 24 Nov 2021 12:20:45 +0000 (20:20 +0800)]
KVM: X86: Remove mmu->translate_gpa

Reduce an indirect function call (retpoline) and some intialization
code.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211124122055.64424-4-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Add parameter struct kvm_mmu *mmu into mmu->gva_to_gpa()
Lai Jiangshan [Wed, 24 Nov 2021 12:20:44 +0000 (20:20 +0800)]
KVM: X86: Add parameter struct kvm_mmu *mmu into mmu->gva_to_gpa()

The mmu->gva_to_gpa() has no "struct kvm_mmu *mmu", so an extra
FNAME(gva_to_gpa_nested) is needed.

Add the parameter can simplify the code.  And it makes it explicit that
the walk is upon vcpu->arch.walk_mmu for gva and vcpu->arch.mmu for L2
gpa in translate_nested_gpa() via the new parameter.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211124122055.64424-3-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Calculate quadrant when !role.gpte_is_8_bytes
Lai Jiangshan [Thu, 18 Nov 2021 11:08:13 +0000 (19:08 +0800)]
KVM: X86: Calculate quadrant when !role.gpte_is_8_bytes

role.quadrant is only valid when gpte size is 4 bytes and only be
calculated when gpte size is 4 bytes.

Although "vcpu->arch.mmu->root_level <= PT32_ROOT_LEVEL" also means
gpte size is 4 bytes, but using "!role.gpte_is_8_bytes" is clearer

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-15-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Remove useless code to set role.gpte_is_8_bytes when role.direct
Lai Jiangshan [Thu, 18 Nov 2021 11:08:12 +0000 (19:08 +0800)]
KVM: X86: Remove useless code to set role.gpte_is_8_bytes when role.direct

role.gpte_is_8_bytes is unused when role.direct; there is no
point in changing a bit in the role, the value that was set
when the MMU is initialized is just fine.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-14-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Remove unused declaration of __kvm_mmu_free_some_pages()
Lai Jiangshan [Thu, 18 Nov 2021 11:08:11 +0000 (19:08 +0800)]
KVM: X86: Remove unused declaration of __kvm_mmu_free_some_pages()

The body of __kvm_mmu_free_some_pages() has been removed.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-13-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Fix comment in __kvm_mmu_create()
Lai Jiangshan [Thu, 18 Nov 2021 11:08:10 +0000 (19:08 +0800)]
KVM: X86: Fix comment in __kvm_mmu_create()

The allocation of special roots is moved to mmu_alloc_special_roots().

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-12-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Skip allocating pae_root for vcpu->arch.guest_mmu when !tdp_enabled
Lai Jiangshan [Thu, 18 Nov 2021 11:08:09 +0000 (19:08 +0800)]
KVM: X86: Skip allocating pae_root for vcpu->arch.guest_mmu when !tdp_enabled

It is never used.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-11-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Allocate sd->save_area with __GFP_ZERO
Lai Jiangshan [Thu, 18 Nov 2021 11:08:08 +0000 (19:08 +0800)]
KVM: SVM: Allocate sd->save_area with __GFP_ZERO

And remove clear_page() on it.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-10-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Rename get_max_npt_level() to get_npt_level()
Lai Jiangshan [Thu, 18 Nov 2021 11:08:07 +0000 (19:08 +0800)]
KVM: SVM: Rename get_max_npt_level() to get_npt_level()

It returns the only proper NPT level, so the "max" in the name
is not appropriate.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-9-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Change comments about vmx_get_msr()
Lai Jiangshan [Thu, 18 Nov 2021 11:08:06 +0000 (19:08 +0800)]
KVM: VMX: Change comments about vmx_get_msr()

The variable name is changed in the code.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-8-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Use kvm_set_msr_common() for MSR_IA32_TSC_ADJUST in the default way
Lai Jiangshan [Thu, 18 Nov 2021 11:08:05 +0000 (19:08 +0800)]
KVM: VMX: Use kvm_set_msr_common() for MSR_IA32_TSC_ADJUST in the default way

MSR_IA32_TSC_ADJUST can be left to the default way which also uese
kvm_set_msr_common().

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-7-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Save HOST_CR3 in vmx_prepare_switch_to_guest()
Lai Jiangshan [Thu, 18 Nov 2021 11:08:03 +0000 (19:08 +0800)]
KVM: VMX: Save HOST_CR3 in vmx_prepare_switch_to_guest()

The host CR3 in the vcpu thread can only be changed when scheduling.
Moving the code in vmx_prepare_switch_to_guest() makes the code
simpler.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-5-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Update msr value after kvm_set_user_return_msr() succeeds
Lai Jiangshan [Thu, 18 Nov 2021 11:08:02 +0000 (19:08 +0800)]
KVM: VMX: Update msr value after kvm_set_user_return_msr() succeeds

Aoid earlier modification.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-4-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Avoid to rdmsrl(MSR_IA32_SYSENTER_ESP)
Lai Jiangshan [Thu, 18 Nov 2021 11:08:01 +0000 (19:08 +0800)]
KVM: VMX: Avoid to rdmsrl(MSR_IA32_SYSENTER_ESP)

The value of host MSR_IA32_SYSENTER_ESP is known to be constant for
each CPU: (cpu_entry_stack(cpu) + 1) when 32 bit syscall is enabled or
NULL is 32 bit syscall is not enabled.

So rdmsrl() can be avoided for the first case and both rdmsrl() and
vmcs_writel() can be avoided for the second case.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211118110814.2568-3-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Update mmu->pdptrs only when it is changed
Lai Jiangshan [Thu, 11 Nov 2021 14:45:26 +0000 (22:45 +0800)]
KVM: X86: Update mmu->pdptrs only when it is changed

It is unchanged in most cases.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211111144527.88852-1-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Remove kvm_register_clear_available()
Lai Jiangshan [Mon, 8 Nov 2021 12:44:06 +0000 (20:44 +0800)]
KVM: X86: Remove kvm_register_clear_available()

It has no user.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-15-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: vmx, svm: clean up mass updates to regs_avail/regs_dirty bits
Paolo Bonzini [Fri, 26 Nov 2021 12:00:15 +0000 (07:00 -0500)]
KVM: vmx, svm: clean up mass updates to regs_avail/regs_dirty bits

Document the meaning of the three combinations of regs_avail and
regs_dirty.  Update regs_dirty just after writeback instead of
doing it later after vmexit.  After vmexit, instead, we clear the
regs_avail bits corresponding to lazily-loaded registers.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Update vmcs.GUEST_CR3 only when the guest CR3 is dirty
Lai Jiangshan [Mon, 8 Nov 2021 12:44:03 +0000 (20:44 +0800)]
KVM: VMX: Update vmcs.GUEST_CR3 only when the guest CR3 is dirty

When vcpu->arch.cr3 is changed, it is marked dirty, so vmcs.GUEST_CR3
can be updated only when kvm_register_is_dirty(vcpu, VCPU_EXREG_CR3).

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-12-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Mark CR3 dirty when vcpu->arch.cr3 is changed
Lai Jiangshan [Mon, 8 Nov 2021 12:44:02 +0000 (20:44 +0800)]
KVM: X86: Mark CR3 dirty when vcpu->arch.cr3 is changed

When vcpu->arch.cr3 is changed, it should be marked dirty unless it
is being updated to the value of the architecture guest CR3 (i.e.
VMX.GUEST_CR3 or vmcb->save.cr3 when tdp is enabled).

This patch has no functionality changed because
kvm_register_mark_dirty(vcpu, VCPU_EXREG_CR3) is superset of
kvm_register_mark_available(vcpu, VCPU_EXREG_CR3) with additional
change to vcpu->arch.regs_dirty, but no code uses regs_dirty for
VCPU_EXREG_CR3.  (vmx_load_mmu_pgd() uses vcpu->arch.regs_avail instead
to test if VCPU_EXREG_CR3 dirty which means current code (ab)uses
regs_avail for VCPU_EXREG_CR3 dirty information.)

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-11-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Remove references to VCPU_EXREG_CR3
Lai Jiangshan [Mon, 8 Nov 2021 12:44:00 +0000 (20:44 +0800)]
KVM: SVM: Remove references to VCPU_EXREG_CR3

VCPU_EXREG_CR3 is never cleared from vcpu->arch.regs_avail or
vcpu->arch.regs_dirty in SVM; therefore, marking CR3 as available is
merely a NOP, and testing it will likewise always succeed.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-9-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Remove outdated comment in svm_load_mmu_pgd()
Lai Jiangshan [Mon, 8 Nov 2021 12:43:59 +0000 (20:43 +0800)]
KVM: SVM: Remove outdated comment in svm_load_mmu_pgd()

The comment had been added in the commit 689f3bf21628 ("KVM: x86: unify
callbacks to load paging root") and its related code was removed later,
and it has nothing to do with the next line of code.

So the comment should be removed too.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-8-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Move CR0 pdptr_bits into header file as X86_CR0_PDPTR_BITS
Lai Jiangshan [Mon, 8 Nov 2021 12:43:58 +0000 (20:43 +0800)]
KVM: X86: Move CR0 pdptr_bits into header file as X86_CR0_PDPTR_BITS

Not functionality changed.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-7-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Add and use X86_CR4_PDPTR_BITS when !enable_ept
Lai Jiangshan [Mon, 8 Nov 2021 12:43:57 +0000 (20:43 +0800)]
KVM: VMX: Add and use X86_CR4_PDPTR_BITS when !enable_ept

In set_cr4_guest_host_mask(), all cr4 pdptr bits are already set to be
intercepted in an unclear way.

Add X86_CR4_PDPTR_BITS to make it clear and self-documented.

No functionality changed.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-6-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Add and use X86_CR4_TLBFLUSH_BITS when !enable_ept
Lai Jiangshan [Mon, 8 Nov 2021 12:43:56 +0000 (20:43 +0800)]
KVM: VMX: Add and use X86_CR4_TLBFLUSH_BITS when !enable_ept

In set_cr4_guest_host_mask(), X86_CR4_PGE is set to be intercepted when
!enable_ept just because X86_CR4_PGE is the only bit that is
responsible for flushing TLB but listed in KVM_POSSIBLE_CR4_GUEST_BITS.

It is clearer and self-documented to use X86_CR4_TLBFLUSH_BITS instead.

No functionality changed.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-5-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Track dirtiness of PDPTRs even if NPT is disabled
Lai Jiangshan [Mon, 8 Nov 2021 12:43:55 +0000 (20:43 +0800)]
KVM: SVM: Track dirtiness of PDPTRs even if NPT is disabled

Use the same logic to handle the availability of VCPU_EXREG_PDPTR
as VMX, also removing a branch in svm_vcpu_run().

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-4-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Mark VCPU_EXREG_PDPTR available in ept_save_pdptrs()
Lai Jiangshan [Mon, 8 Nov 2021 12:43:54 +0000 (20:43 +0800)]
KVM: VMX: Mark VCPU_EXREG_PDPTR available in ept_save_pdptrs()

mmu->pdptrs[] and vmcs.GUEST_PDPTR[0-3] are synced, so mmu->pdptrs is
available and GUEST_PDPTR[0-3] is not dirty.

Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-3-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: X86: Ensure that dirty PDPTRs are loaded
Lai Jiangshan [Mon, 8 Nov 2021 12:43:53 +0000 (20:43 +0800)]
KVM: X86: Ensure that dirty PDPTRs are loaded

For VMX with EPT, dirty PDPTRs need to be loaded before the next vmentry
via vmx_load_mmu_pgd()

But not all paths that call load_pdptrs() will cause vmx_load_mmu_pgd()
to be invoked.  Normally, kvm_mmu_reset_context() is used to cause
KVM_REQ_LOAD_MMU_PGD, but sometimes it is skipped:

* commit d81135a57aa6("KVM: x86: do not reset mmu if CR0.CD and
CR0.NW are changed") skips kvm_mmu_reset_context() after load_pdptrs()
when changing CR0.CD and CR0.NW.

* commit 21823fbda552("KVM: x86: Invalidate all PGDs for the current
PCID on MOV CR3 w/ flush") skips KVM_REQ_LOAD_MMU_PGD after
load_pdptrs() when rewriting the CR3 with the same value.

* commit a91a7c709600("KVM: X86: Don't reset mmu context when
toggling X86_CR4_PGE") skips kvm_mmu_reset_context() after
load_pdptrs() when changing CR4.PGE.

Fixes: d81135a57aa6 ("KVM: x86: do not reset mmu if CR0.CD and CR0.NW are changed")
Fixes: 21823fbda552 ("KVM: x86: Invalidate all PGDs for the current PCID on MOV CR3 w/ flush")
Fixes: a91a7c709600 ("KVM: X86: Don't reset mmu context when toggling X86_CR4_PGE")
Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20211108124407.12187-2-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/svm: Add module param to control PMU virtualization
Like Xu [Wed, 17 Nov 2021 08:03:04 +0000 (16:03 +0800)]
KVM: x86/svm: Add module param to control PMU virtualization

For Intel, the guest PMU can be disabled via clearing the PMU CPUID.
For AMD, all hw implementations support the base set of four
performance counters, with current mainstream hardware indicating
the presence of two additional counters via X86_FEATURE_PERFCTR_CORE.

In the virtualized world, the AMD guest driver may detect
the presence of at least one counter MSR. Most hypervisor
vendors would introduce a module param (like lbrv for svm)
to disable PMU for all guests.

Another control proposal per-VM is to pass PMU disable information
via MSR_IA32_PERF_CAPABILITIES or one bit in CPUID Fn4000_00[FF:00].
Both of methods require some guest-side changes, so a module
parameter may not be sufficiently granular, but practical enough.

Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20211117080304.38989-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Remove vCPU from PI wakeup list before updating PID.NV
Sean Christopherson [Sat, 9 Oct 2021 02:12:21 +0000 (19:12 -0700)]
KVM: VMX: Remove vCPU from PI wakeup list before updating PID.NV

Remove the vCPU from the wakeup list before updating the notification
vector in the posted interrupt post-block helper.  There is no need to
wake the current vCPU as it is by definition not blocking.  Practically
speaking this is a nop as it only shaves a few meager cycles in the
unlikely case that the vCPU was migrated and the previous pCPU gets a
wakeup IRQ right before PID.NV is updated.  The real motivation is to
allow for more readable code in the future, when post-block is merged
with vmx_vcpu_pi_load(), at which point removal from the list will be
conditional on the old notification vector.

Opportunistically add comments to document why KVM has a per-CPU spinlock
that, at first glance, appears to be taken only on the owning CPU.
Explicitly call out that the spinlock must be taken with IRQs disabled, a
detail that was "lost" when KVM switched from spin_lock_irqsave() to
spin_lock(), with IRQs disabled for the entirety of the relevant path.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-29-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Move Posted Interrupt ndst computation out of write loop
Sean Christopherson [Sat, 9 Oct 2021 02:12:20 +0000 (19:12 -0700)]
KVM: VMX: Move Posted Interrupt ndst computation out of write loop

Hoist the CPU => APIC ID conversion for the Posted Interrupt descriptor
out of the loop to write the descriptor, preemption is disabled so the
CPU won't change, and if the APIC ID changes KVM has bigger problems.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-28-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Read Posted Interrupt "control" exactly once per loop iteration
Sean Christopherson [Sat, 9 Oct 2021 02:12:19 +0000 (19:12 -0700)]
KVM: VMX: Read Posted Interrupt "control" exactly once per loop iteration

Use READ_ONCE() when loading the posted interrupt descriptor control
field to ensure "old" and "new" have the same base value.  If the
compiler emits separate loads, and loads into "new" before "old", KVM
could theoretically drop the ON bit if it were set between the loads.

Fixes: 28b835d60fcc ("KVM: Update Posted-Interrupts Descriptor when vCPU is preempted")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-27-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Save/restore IRQs (instead of CLI/STI) during PI pre/post block
Sean Christopherson [Sat, 9 Oct 2021 02:12:18 +0000 (19:12 -0700)]
KVM: VMX: Save/restore IRQs (instead of CLI/STI) during PI pre/post block

Save/restore IRQs when disabling IRQs in posted interrupt pre/post block
in preparation for moving the code into vcpu_put/load(), where it would be
called with IRQs already disabled.

No functional changed intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-26-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Drop pointless PI.NDST update when blocking
Sean Christopherson [Sat, 9 Oct 2021 02:12:17 +0000 (19:12 -0700)]
KVM: VMX: Drop pointless PI.NDST update when blocking

Don't update Posted Interrupt's NDST, a.k.a. the target pCPU, in the
pre-block path, as NDST is guaranteed to be up-to-date.  The comment
about the vCPU being preempted during the update is simply wrong, as the
update path runs with IRQs disabled (from before snapshotting vcpu->cpu,
until after the update completes).

Since commit 8b306e2f3c41 ("KVM: VMX: avoid double list add with VT-d
posted interrupts", 2017-09-27) The vCPU can get preempted _before_
the update starts, but not during.  And if the vCPU is preempted before,
vmx_vcpu_pi_load() is responsible for updating NDST when the vCPU is
scheduled back in.  In that case, the check against the wakeup vector in
vmx_vcpu_pi_load() cannot be true as that would require the notification
vector to have been set to the wakeup vector _before_ blocking.

Opportunistically switch to using vcpu->cpu for the list/lock lookups,
which do not need pre_pcpu since the same commit.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-25-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Use boolean returns for Posted Interrupt "test" helpers
Sean Christopherson [Sat, 9 Oct 2021 02:12:16 +0000 (19:12 -0700)]
KVM: VMX: Use boolean returns for Posted Interrupt "test" helpers

Return bools instead of ints for the posted interrupt "test" helpers.
The bit position of the flag being test does not matter to the callers,
and is in fact lost by virtue of test_bit() itself returning a bool.

Returning ints is potentially dangerous, e.g. "pi_test_on(pi_desc) == 1"
is safe-ish because ON is bit 0 and thus any sane implementation of
pi_test_on() will work, but for SN (bit 1), checking "== 1" would rely on
pi_test_on() to return 0 or 1, a.k.a. bools, as opposed to 0 or 2 (the
positive bit position).

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-24-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Drop unnecessary PI logic to handle impossible conditions
Sean Christopherson [Sat, 9 Oct 2021 02:12:15 +0000 (19:12 -0700)]
KVM: VMX: Drop unnecessary PI logic to handle impossible conditions

Drop sanity checks on the validity of the previous pCPU when handling
vCPU block/unlock for posted interrupts.  The intention behind the sanity
checks is to avoid memory corruption in case of a race or incorrect locking,
but the code has been stable for a few years now and the checks get in
the way of eliminating kvm_vcpu.pre_cpu.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-23-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Skip Posted Interrupt updates if APICv is hard disabled
Sean Christopherson [Sat, 9 Oct 2021 02:12:13 +0000 (19:12 -0700)]
KVM: VMX: Skip Posted Interrupt updates if APICv is hard disabled

Explicitly skip posted interrupt updates if APICv is disabled in all of
KVM, or if the guest doesn't have an in-kernel APIC.  The PI descriptor
is kept up-to-date if APICv is inhibited, e.g. so that re-enabling APICv
doesn't require a bunch of updates, but neither the module param nor the
APIC type can be changed on-the-fly.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-21-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Add helpers to wake/query blocking vCPU
Sean Christopherson [Sat, 9 Oct 2021 02:12:12 +0000 (19:12 -0700)]
KVM: Add helpers to wake/query blocking vCPU

Add helpers to wake and query a blocking vCPU.  In addition to providing
nice names, the helpers reduce the probability of KVM neglecting to use
kvm_arch_vcpu_get_wait().

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-20-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Invoke kvm_vcpu_block() directly for non-HALTED wait states
Sean Christopherson [Sat, 9 Oct 2021 02:12:11 +0000 (19:12 -0700)]
KVM: x86: Invoke kvm_vcpu_block() directly for non-HALTED wait states

Call kvm_vcpu_block() directly for all wait states except HALTED so that
kvm_vcpu_halt() is no longer a misnomer on x86.

Functionally, this means KVM will never attempt halt-polling or adjust
vcpu->halt_poll_ns for INIT_RECEIVED (a.k.a. Wait-For-SIPI (WFS)) or
AP_RESET_HOLD; UNINITIALIZED is handled in kvm_arch_vcpu_ioctl_run(),
and x86 doesn't use any other "wait" states.

As mentioned above, the motivation of this is purely so that "halt" isn't
overloaded on x86, e.g. in KVM's stats.  Skipping halt-polling for WFS
(and RESET_HOLD) has no meaningful effect on guest performance as there
are typically single-digit numbers of INIT-SIPI sequences per AP vCPU,
per boot, versus thousands of HLTs just to boot to console.

Reviewed-by: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-19-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Directly block (instead of "halting") UNINITIALIZED vCPUs
Sean Christopherson [Sat, 9 Oct 2021 02:12:10 +0000 (19:12 -0700)]
KVM: x86: Directly block (instead of "halting") UNINITIALIZED vCPUs

Go directly to kvm_vcpu_block() when handling the case where userspace
attempts to run an UNINITIALIZED vCPU.  The vCPU is not halted, nor is it
likely that halt-polling will be successful in this case.

Reviewed-by: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-18-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Don't redo ktime_get() when calculating halt-polling stop/deadline
Sean Christopherson [Sat, 9 Oct 2021 02:12:09 +0000 (19:12 -0700)]
KVM: Don't redo ktime_get() when calculating halt-polling stop/deadline

Calculate the halt-polling "stop" time using "start" instead of redoing
ktime_get().  In practice, the numbers involved are in the noise (e.g.,
in the happy case where hardware correctly predicts do_halt_poll and
there are no interrupts, "start" is probably only a few cycles old)
and either approach is perfectly ok.  But it's more precise to count
any extra latency toward the halt-polling time.

Reviewed-by: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-17-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: stats: Add stat to detect if vcpu is currently blocking
Jing Zhang [Sat, 9 Oct 2021 02:12:08 +0000 (19:12 -0700)]
KVM: stats: Add stat to detect if vcpu is currently blocking

Add a "blocking" stat that userspace can use to detect the case where a
vCPU is not being run because of an vCPU/guest action, e.g. HLT or WFS on
x86, WFI on arm64, etc...  Current guest/host/halt stats don't show this
well, e.g. if a guest halts for a long period of time then the vCPU could
could appear pathologically blocked due to a host condition, when in
reality the vCPU has been put into a not-runnable state by the guest.

Originally-by: Cannon Matthews <cannonmatthews@google.com>
Suggested-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: David Matlack <dmatlack@google.com>
Signed-off-by: Jing Zhang <jingzhangos@google.com>
[sean: renamed stat to "blocking", massaged changelog]
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-16-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Split out a kvm_vcpu_block() helper from kvm_vcpu_halt()
Sean Christopherson [Sat, 9 Oct 2021 02:12:07 +0000 (19:12 -0700)]
KVM: Split out a kvm_vcpu_block() helper from kvm_vcpu_halt()

Factor out the "block" part of kvm_vcpu_halt() so that x86 can emulate
non-halt wait/sleep/block conditions that should not be subjected to
halt-polling.

No functional change intended.

Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Reviewed-by: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-15-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Rename kvm_vcpu_block() => kvm_vcpu_halt()
Sean Christopherson [Sat, 9 Oct 2021 02:12:06 +0000 (19:12 -0700)]
KVM: Rename kvm_vcpu_block() => kvm_vcpu_halt()

Rename kvm_vcpu_block() to kvm_vcpu_halt() in preparation for splitting
the actual "block" sequences into a separate helper (to be named
kvm_vcpu_block()).  x86 will use the standalone block-only path to handle
non-halt cases where the vCPU is not runnable.

Rename block_ns to halt_ns to match the new function name.

No functional change intended.

Reviewed-by: David Matlack <dmatlack@google.com>
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-14-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Drop obsolete kvm_arch_vcpu_block_finish()
Sean Christopherson [Sat, 9 Oct 2021 02:12:02 +0000 (19:12 -0700)]
KVM: Drop obsolete kvm_arch_vcpu_block_finish()

Drop kvm_arch_vcpu_block_finish() now that all arch implementations are
nops.

No functional change intended.

Acked-by: Christian Borntraeger <borntraeger@de.ibm.com>
Reviewed-by: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Tweak halt emulation helper names to free up kvm_vcpu_halt()
Sean Christopherson [Sat, 9 Oct 2021 02:12:05 +0000 (19:12 -0700)]
KVM: x86: Tweak halt emulation helper names to free up kvm_vcpu_halt()

Rename a variety of HLT-related helpers to free up the function name
"kvm_vcpu_halt" for future use in generic KVM code, e.g. to differentiate
between "block" and "halt".

No functional change intended.

Reviewed-by: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-13-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Don't block+unblock when halt-polling is successful
Sean Christopherson [Sat, 9 Oct 2021 02:12:04 +0000 (19:12 -0700)]
KVM: Don't block+unblock when halt-polling is successful

Invoke the arch hooks for block+unblock if and only if KVM actually
attempts to block the vCPU.  The only non-nop implementation is on x86,
specifically SVM's AVIC, and there is no need to put the AVIC prior to
halt-polling; KVM x86's kvm_vcpu_has_events() will scour the full vIRR
to find pending IRQs regardless of whether the AVIC is loaded/"running".

The primary motivation is to allow future cleanup to split out "block"
from "halt", but this is also likely a small performance boost on x86 SVM
when halt-polling is successful.

Adjust the post-block path to update "cur" after unblocking, i.e. include
AVIC load time in halt_wait_ns and halt_wait_hist, so that the behavior
is consistent.  Moving just the pre-block arch hook would result in only
the AVIC put latency being included in the halt_wait stats.  There is no
obvious evidence that one way or the other is correct, so just ensure KVM
is consistent.

Note, x86 has two separate paths for handling APICv with respect to vCPU
blocking.  VMX uses hooks in x86's vcpu_block(), while SVM uses the arch
hooks in kvm_vcpu_block().  Prior to this path, the two paths were more
or less functionally identical.  That is very much not the case after
this patch, as the hooks used by VMX _must_ fire before halt-polling.
x86's entire mess will be cleaned up in future patches.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211009021236.4122790-12-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>