APICV_INHIBIT_REASON_HYPERV is currently unconditionally forced upon
SynIC activation as SynIC's AutoEOI is incompatible with APICv/AVIC. It is,
however, possible to track whether the feature was actually used by the
guest and only inhibit APICv/AVIC when needed.
TLFS suggests a dedicated 'HV_DEPRECATING_AEOI_RECOMMENDED' flag to let
Windows know that AutoEOI feature should be avoided. While it's up to
KVM userspace to set the flag, KVM can help a bit by exposing global
APICv/AVIC enablement.
Maxim:
- always set HV_DEPRECATING_AEOI_RECOMMENDED in kvm_get_hv_cpuid,
since this feature can be used regardless of AVIC
Paolo:
- use arch.apicv_update_lock to protect the hv->synic_auto_eoi_used
instead of atomic ops
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210810205251.424103-12-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently on SVM, the kvm_request_apicv_update toggles the APICv
memslot without doing any synchronization.
If there is a mismatch between that memslot state and the AVIC state,
on one of the vCPUs, an APIC mmio access can be lost:
For example:
VCPU0: enable the APIC_ACCESS_PAGE_PRIVATE_MEMSLOT
VCPU1: access an APIC mmio register.
Since AVIC is still disabled on VCPU1, the access will not be intercepted
by it, and neither will it cause MMIO fault, but rather it will just be
read/written from/to the dummy page mapped into the
APIC_ACCESS_PAGE_PRIVATE_MEMSLOT.
Fix that by adding a lock guarding the AVIC state changes, and carefully
order the operations of kvm_request_apicv_update to avoid this race:
1. Take the lock
2. Send KVM_REQ_APICV_UPDATE
3. Update the apic inhibit reason
4. Release the lock
This ensures that at (2) all vCPUs are kicked out of the guest mode,
but don't yet see the new avic state.
Then only after (4) all other vCPUs can update their AVIC state and resume.
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210810205251.424103-10-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
on AMD, APIC virtualization needs to dynamicaly inhibit the AVIC in a
response to some events, and this is problematic and not efficient to do by
enabling/disabling the memslot that covers APIC's mmio range.
Plus due to SRCU locking, it makes it more complex to
request AVIC inhibition.
Instead, the APIC memslot will be always enabled, but be invisible
to the guest, such as the MMU code will not install a SPTE for it,
when it is inhibited and instead jump straight to emulating the access.
When inhibiting the AVIC, this SPTE will be zapped.
This code is based on a suggestion from Sean Christopherson:
https://lkml.org/lkml/2021/7/19/2970
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210810205251.424103-8-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This together with the next patch will fix a future race between
kvm_zap_gfn_range and the page fault handler, which will happen
when AVIC memslot is going to be only partially disabled.
The performance impact is minimal since kvm_zap_gfn_range is only
called by users, update_mtrr() and kvm_post_set_cr0().
Both only use it if the guest has non-coherent DMA, in order to
honor the guest's UC memtype.
MTRR and CD setup only happens at boot, and generally in an area
where the page tables should be small (for CD) or should not
include the affected GFNs at all (for MTRRs).
This is based on a patch suggested by Sean Christopherson:
https://lkml.org/lkml/2021/7/22/1025
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210810205251.424103-2-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use this file to dump rmap statistic information. The statistic is done by
calculating the rmap count and the result is log-2-based.
An example output of this looks like (idle 6GB guest, right after boot linux):
Rmap_Count: 0 1 2-3 4-7 8-15 16-31 32-63 64-127 128-255 256-511 512-1023
Level=4K: 3086676 53045 12330 1272 502 121 76 2 0 0 0
Level=2M: 5947 231 0 0 0 0 0 0 0 0 0
Level=1G: 32 0 0 0 0 0 0 0 0 0 0
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-Id: <20210730220455.26054-5-peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Introduce kvm_mmu_slot_lpages() to calculcate lpage_info and rmap array size.
The other __kvm_mmu_slot_lpages() can take an extra parameter of npages rather
than fetching from the memslot pointer. Start to use the latter one in
kvm_alloc_memslot_metadata().
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-Id: <20210730220455.26054-4-peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If L1 disables VMLOAD/VMSAVE intercepts, and doesn't enable
Virtual VMLOAD/VMSAVE (currently not supported for the nested hypervisor),
then VMLOAD/VMSAVE must operate on the L1 physical memory, which is only
possible by making L0 intercept these instructions.
Failure to do so allowed the nested guest to run VMLOAD/VMSAVE unintercepted,
and thus read/write portions of the host physical memory.
Fixes: 89c8a4984f ("KVM: SVM: Enable Virtual VMLOAD VMSAVE feature")
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
* Invert the mask of bits that we pick from L2 in
nested_vmcb02_prepare_control
* Invert and explicitly use VIRQ related bits bitmask in svm_clear_vintr
This fixes a security issue that allowed a malicious L1 to run L2 with
AVIC enabled, which allowed the L2 to exploit the uninitialized and enabled
AVIC to read/write the host physical memory at some offsets.
Fixes: 3d6368ef58 ("KVM: SVM: Add VMRUN handler")
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Clear nested.pi_pending on nested VM-Enter even if L2 will run without
posted interrupts enabled. If nested.pi_pending is left set from a
previous L2, vmx_complete_nested_posted_interrupt() will pick up the
stale flag and exit to userspace with an "internal emulation error" due
the new L2 not having a valid nested.pi_desc.
Arguably, vmx_complete_nested_posted_interrupt() should first check for
posted interrupts being enabled, but it's also completely reasonable that
KVM wouldn't screw up a fundamental flag. Not to mention that the mere
existence of nested.pi_pending is a long-standing bug as KVM shouldn't
move the posted interrupt out of the IRR until it's actually processed,
e.g. KVM effectively drops an interrupt when it performs a nested VM-Exit
with a "pending" posted interrupt. Fixing the mess is a future problem.
Prior to vmx_complete_nested_posted_interrupt() interpreting a null PI
descriptor as an error, this was a benign bug as the null PI descriptor
effectively served as a check on PI not being enabled. Even then, the
new flow did not become problematic until KVM started checking the result
of kvm_check_nested_events().
Fixes: 705699a139 ("KVM: nVMX: Enable nested posted interrupt processing")
Fixes: 966eefb896 ("KVM: nVMX: Disable vmcs02 posted interrupts if vmcs12 PID isn't mappable")
Fixes: 47d3530f86c0 ("KVM: x86: Exit to userspace when kvm_check_nested_events fails")
Cc: stable@vger.kernel.org
Cc: Jim Mattson <jmattson@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210810144526.2662272-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The ROL16(val, n) macro is repeatedly defined in several vmcs-related
files, and it has never been used outside the KVM context.
Let's move it to vmcs.h without any intended functional changes.
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20210809093410.59304-4-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Remove the __kvm_handle_fault_on_reboot() and __ex() macros now that all
VMX and SVM instructions use asm goto to handle the fault (or in the
case of VMREAD, completely custom logic). Drop kvm_spurious_fault()'s
asmlinkage annotation as __kvm_handle_fault_on_reboot() was the only
flow that invoked it from assembly code.
Cc: Uros Bizjak <ubizjak@gmail.com>
Cc: Like Xu <like.xu.linux@gmail.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210809173955.1710866-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Now that nested VMX pulls KVM's desired VMCS controls from vmcs01 instead
of re-calculating on the fly, bury the helpers that do the calcluations
in vmx.c.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210810171952.2758100-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Remove the secondary execution controls cache now that it's effectively
dead code; it is only read immediately after it is written.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210810171952.2758100-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When preparing controls for vmcs02, grab KVM's desired controls from
vmcs01's shadow state instead of recalculating the controls from scratch,
or in the secondary execution controls, instead of using the dedicated
cache. Calculating secondary exec controls is eye-poppingly expensive
due to the guest CPUID checks, hence the dedicated cache, but the other
calculations aren't exactly free either.
Explicitly clear several bits (x2APIC, DESC exiting, and load EFER on
exit) as appropriate as they may be set in vmcs01, whereas the previous
implementation relied on dynamic bits being cleared in the calculator.
Intentionally propagate VM_{ENTRY,EXIT}_LOAD_IA32_PERF_GLOBAL_CTRL from
vmcs01 to vmcs02. Whether or not PERF_GLOBAL_CTRL is loaded depends on
whether or not perf itself is active, so unless perf stops between the
exit from L1 and entry to L2, vmcs01 will hold the desired value. This
is purely an optimization as atomic_switch_perf_msrs() will set/clear
the control as needed at VM-Enter, i.e. it avoids two extra VMWRITEs in
the case where perf is active (versus starting with the bits clear in
vmcs02, which was the previous behavior).
Cc: Zeng Guang <guang.zeng@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210810171952.2758100-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The commit efdab99281 ("KVM: x86: fix escape of guest dr6 to the host")
fixed a bug by resetting DR6 unconditionally when the vcpu being scheduled out.
But writing to debug registers is slow, and it can be visible in perf results
sometimes, even if neither the host nor the guest activate breakpoints.
Since KVM_DEBUGREG_WONT_EXIT on Intel processors is the only case
where DR6 gets the guest value, and it never happens at all on SVM,
the register can be cleared in vmx.c right after reading it.
Reported-by: Lai Jiangshan <laijs@linux.alibaba.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Commit c77fb5fe6f ("KVM: x86: Allow the guest to run with dirty debug
registers") allows the guest accessing to DRs without exiting when
KVM_DEBUGREG_WONT_EXIT and we need to ensure that they are synchronized
on entry to the guest---including DR6 that was not synced before the commit.
But the commit sets the hardware DR6 not only when KVM_DEBUGREG_WONT_EXIT,
but also when KVM_DEBUGREG_BP_ENABLED. The second case is unnecessary
and just leads to a more case which leaks stale DR6 to the host which has
to be resolved by unconditionally reseting DR6 in kvm_arch_vcpu_put().
Even if KVM_DEBUGREG_WONT_EXIT, however, setting the host DR6 only matters
on VMX because SVM always uses the DR6 value from the VMCB. So move this
line to vmx.c and make it conditional on KVM_DEBUGREG_WONT_EXIT.
Reported-by: Lai Jiangshan <jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Commit ae561edeb4 ("KVM: x86: DR0-DR3 are not clear on reset") added code to
ensure eff_db are updated when they're modified through non-standard paths.
But there is no reason to also update hardware DRs unless hardware breakpoints
are active or DR exiting is disabled, and in those cases updating hardware is
handled by KVM_DEBUGREG_WONT_EXIT and KVM_DEBUGREG_BP_ENABLED.
KVM_DEBUGREG_RELOAD just causes unnecesarry load of hardware DRs and is better
to be removed.
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20210809174307.145263-1-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add yet another spinlock for the TDP MMU and take it when marking indirect
shadow pages unsync. When using the TDP MMU and L1 is running L2(s) with
nested TDP, KVM may encounter shadow pages for the TDP entries managed by
L1 (controlling L2) when handling a TDP MMU page fault. The unsync logic
is not thread safe, e.g. the kvm_mmu_page fields are not atomic, and
misbehaves when a shadow page is marked unsync via a TDP MMU page fault,
which runs with mmu_lock held for read, not write.
Lack of a critical section manifests most visibly as an underflow of
unsync_children in clear_unsync_child_bit() due to unsync_children being
corrupted when multiple CPUs write it without a critical section and
without atomic operations. But underflow is the best case scenario. The
worst case scenario is that unsync_children prematurely hits '0' and
leads to guest memory corruption due to KVM neglecting to properly sync
shadow pages.
Use an entirely new spinlock even though piggybacking tdp_mmu_pages_lock
would functionally be ok. Usurping the lock could degrade performance when
building upper level page tables on different vCPUs, especially since the
unsync flow could hold the lock for a comparatively long time depending on
the number of indirect shadow pages and the depth of the paging tree.
For simplicity, take the lock for all MMUs, even though KVM could fairly
easily know that mmu_lock is held for write. If mmu_lock is held for
write, there cannot be contention for the inner spinlock, and marking
shadow pages unsync across multiple vCPUs will be slow enough that
bouncing the kvm_arch cacheline should be in the noise.
Note, even though L2 could theoretically be given access to its own EPT
entries, a nested MMU must hold mmu_lock for write and thus cannot race
against a TDP MMU page fault. I.e. the additional spinlock only _needs_ to
be taken by the TDP MMU, as opposed to being taken by any MMU for a VM
that is running with the TDP MMU enabled. Holding mmu_lock for read also
prevents the indirect shadow page from being freed. But as above, keep
it simple and always take the lock.
Alternative #1, the TDP MMU could simply pass "false" for can_unsync and
effectively disable unsync behavior for nested TDP. Write protecting leaf
shadow pages is unlikely to noticeably impact traditional L1 VMMs, as such
VMMs typically don't modify TDP entries, but the same may not hold true for
non-standard use cases and/or VMMs that are migrating physical pages (from
L1's perspective).
Alternative #2, the unsync logic could be made thread safe. In theory,
simply converting all relevant kvm_mmu_page fields to atomics and using
atomic bitops for the bitmap would suffice. However, (a) an in-depth audit
would be required, (b) the code churn would be substantial, and (c) legacy
shadow paging would incur additional atomic operations in performance
sensitive paths for no benefit (to legacy shadow paging).
Fixes: a2855afc7e ("KVM: x86/mmu: Allow parallel page faults for the TDP MMU")
Cc: stable@vger.kernel.org
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210812181815.3378104-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Set the min_level for the TDP iterator at the root level when zapping all
SPTEs to optimize the iterator's try_step_down(). Zapping a non-leaf
SPTE will recursively zap all its children, thus there is no need for the
iterator to attempt to step down. This avoids rereading the top-level
SPTEs after they are zapped by causing try_step_down() to short-circuit.
In most cases, optimizing try_step_down() will be in the noise as the cost
of zapping SPTEs completely dominates the overall time. The optimization
is however helpful if the zap occurs with relatively few SPTEs, e.g. if KVM
is zapping in response to multiple memslot updates when userspace is adding
and removing read-only memslots for option ROMs. In that case, the task
doing the zapping likely isn't a vCPU thread, but it still holds mmu_lock
for read and thus can be a noisy neighbor of sorts.
Reviewed-by: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210812181414.3376143-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Pass "all ones" as the end GFN to signal "zap all" for the TDP MMU and
really zap all SPTEs in this case. As is, zap_gfn_range() skips non-leaf
SPTEs whose range exceeds the range to be zapped. If shadow_phys_bits is
not aligned to the range size of top-level SPTEs, e.g. 512gb with 4-level
paging, the "zap all" flows will skip top-level SPTEs whose range extends
beyond shadow_phys_bits and leak their SPs when the VM is destroyed.
Use the current upper bound (based on host.MAXPHYADDR) to detect that the
caller wants to zap all SPTEs, e.g. instead of using the max theoretical
gfn, 1 << (52 - 12). The more precise upper bound allows the TDP iterator
to terminate its walk earlier when running on hosts with MAXPHYADDR < 52.
Add a WARN on kmv->arch.tdp_mmu_pages when the TDP MMU is destroyed to
help future debuggers should KVM decide to leak SPTEs again.
The bug is most easily reproduced by running (and unloading!) KVM in a
VM whose host.MAXPHYADDR < 39, as the SPTE for gfn=0 will be skipped.
=============================================================================
BUG kvm_mmu_page_header (Not tainted): Objects remaining in kvm_mmu_page_header on __kmem_cache_shutdown()
-----------------------------------------------------------------------------
Slab 0x000000004d8f7af1 objects=22 used=2 fp=0x00000000624d29ac flags=0x4000000000000200(slab|zone=1)
CPU: 0 PID: 1582 Comm: rmmod Not tainted 5.14.0-rc2+ #420
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
Call Trace:
dump_stack_lvl+0x45/0x59
slab_err+0x95/0xc9
__kmem_cache_shutdown.cold+0x3c/0x158
kmem_cache_destroy+0x3d/0xf0
kvm_mmu_module_exit+0xa/0x30 [kvm]
kvm_arch_exit+0x5d/0x90 [kvm]
kvm_exit+0x78/0x90 [kvm]
vmx_exit+0x1a/0x50 [kvm_intel]
__x64_sys_delete_module+0x13f/0x220
do_syscall_64+0x3b/0xc0
entry_SYSCALL_64_after_hwframe+0x44/0xae
Fixes: faaf05b00a ("kvm: x86/mmu: Support zapping SPTEs in the TDP MMU")
Cc: stable@vger.kernel.org
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210812181414.3376143-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use vmx_need_pf_intercept() when determining if L0 wants to handle a #PF
in L2 or if the VM-Exit should be forwarded to L1. The current logic fails
to account for the case where #PF is intercepted to handle
guest.MAXPHYADDR < host.MAXPHYADDR and ends up reflecting all #PFs into
L1. At best, L1 will complain and inject the #PF back into L2. At
worst, L1 will eat the unexpected fault and cause L2 to hang on infinite
page faults.
Note, while the bug was technically introduced by the commit that added
support for the MAXPHYADDR madness, the shame is all on commit
a0c134347b ("KVM: VMX: introduce vmx_need_pf_intercept").
Fixes: 1dbf5d68af ("KVM: VMX: Add guest physical address check in EPT violation and misconfig")
Cc: stable@vger.kernel.org
Cc: Peter Shier <pshier@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Jim Mattson <jmattson@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210812045615.3167686-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When a nested EPT violation/misconfig is injected into the guest,
the shadow EPT PTEs associated with that address need to be synced.
This is done by kvm_inject_emulated_page_fault() before it calls
nested_ept_inject_page_fault(). However, that will only sync the
shadow EPT PTE associated with the current L1 EPTP. Since the ASID
is based on EP4TA rather than the full EPTP, so syncing the current
EPTP is not enough. The SPTEs associated with any other L1 EPTPs
in the prev_roots cache with the same EP4TA also need to be synced.
Signed-off-by: Junaid Shahid <junaids@google.com>
Message-Id: <20210806222229.1645356-1-junaids@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
hv_vcpu is initialized again a dozen lines below, and at this
point vcpu->arch.hyperv is not valid. Remove the initializer.
Reported-by: kernel test robot <lkp@intel.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Remove an ancient restriction that disallowed exposing EFER.NX to the
guest if EFER.NX=0 on the host, even if NX is fully supported by the CPU.
The motivation of the check, added by commit 2cc51560ae ("KVM: VMX:
Avoid saving and restoring msr_efer on lightweight vmexit"), was to rule
out the case of host.EFER.NX=0 and guest.EFER.NX=1 so that KVM could run
the guest with the host's EFER.NX and thus avoid context switching EFER
if the only divergence was the NX bit.
Fast forward to today, and KVM has long since stopped running the guest
with the host's EFER.NX. Not only does KVM context switch EFER if
host.EFER.NX=1 && guest.EFER.NX=0, KVM also forces host.EFER.NX=0 &&
guest.EFER.NX=1 when using shadow paging (to emulate SMEP). Furthermore,
the entire motivation for the restriction was made obsolete over a decade
ago when Intel added dedicated host and guest EFER fields in the VMCS
(Nehalem timeframe), which reduced the overhead of context switching EFER
from 400+ cycles (2 * WRMSR + 1 * RDMSR) to a mere ~2 cycles.
In practice, the removed restriction only affects non-PAE 32-bit kernels,
as EFER.NX is set during boot if NX is supported and the kernel will use
PAE paging (32-bit or 64-bit), regardless of whether or not the kernel
will actually use NX itself (mark PTEs non-executable).
Alternatively and/or complementarily, startup_32_smp() in head_32.S could
be modified to set EFER.NX=1 regardless of paging mode, thus eliminating
the scenario where NX is supported but not enabled. However, that runs
the risk of breaking non-KVM non-PAE kernels (though the risk is very,
very low as there are no known EFER.NX errata), and also eliminates an
easy-to-use mechanism for stressing KVM's handling of guest vs. host EFER
across nested virtualization transitions.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210805183804.1221554-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use the secondary_exec_controls_get() accessor in vmx_has_waitpkg() to
effectively get the controls for the current VMCS, as opposed to using
vmx->secondary_exec_controls, which is the cached value of KVM's desired
controls for vmcs01 and truly not reflective of any particular VMCS.
While the waitpkg control is not dynamic, i.e. vmcs01 will always hold
the same waitpkg configuration as vmx->secondary_exec_controls, the same
does not hold true for vmcs02 if the L1 VMM hides the feature from L2.
If L1 hides the feature _and_ does not intercept MSR_IA32_UMWAIT_CONTROL,
L2 could incorrectly read/write L1's virtual MSR instead of taking a #GP.
Fixes: 6e3ba4abce ("KVM: vmx: Emulate MSR IA32_UMWAIT_CONTROL")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210810171952.2758100-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
rmap_add() and rmap_recycle() both run in the context of the vCPU and
thus we can use kvm_vcpu_gfn_to_memslot() to look up the memslot. This
enables rmap_add() and rmap_recycle() to take advantage of
vcpu->last_used_slot and avoid expensive memslot searching.
This change improves the performance of "Populate memory time" in
dirty_log_perf_test with tdp_mmu=N. In addition to improving the
performance, "Populate memory time" no longer scales with the number
of memslots in the VM.
Command | Before | After
------------------------------- | ---------------- | -------------
./dirty_log_perf_test -v64 -x1 | 15.18001570s | 14.99469366s
./dirty_log_perf_test -v64 -x64 | 18.71336392s | 14.98675076s
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20210804222844.1419481-6-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The existing TDP MMU methods to handle dirty logging are vcpu-agnostic
since they can be driven by MMU notifiers and other non-vcpu-specific
events in addition to page faults. However this means that the TDP MMU
is not benefiting from the new vcpu->last_used_slot. Fix that by
introducing a tdp_mmu_map_set_spte_atomic() which is only called during
a TDP page fault and has access to the kvm_vcpu for fast slot lookups.
This improves "Populate memory time" in dirty_log_perf_test by 5%:
Command | Before | After
------------------------------- | ---------------- | -------------
./dirty_log_perf_test -v64 -x64 | 5.472321072s | 5.169832886s
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20210804222844.1419481-5-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Take a signed 'long' instead of an 'unsigned long' for the number of
pages to add/subtract to the total number of pages used by the MMU. This
fixes a zero-extension bug on 32-bit kernels that effectively corrupts
the per-cpu counter used by the shrinker.
Per-cpu counters take a signed 64-bit value on both 32-bit and 64-bit
kernels, whereas kvm_mod_used_mmu_pages() takes an unsigned long and thus
an unsigned 32-bit value on 32-bit kernels. As a result, the value used
to adjust the per-cpu counter is zero-extended (unsigned -> signed), not
sign-extended (signed -> signed), and so KVM's intended -1 gets morphed to
4294967295 and effectively corrupts the counter.
This was found by a staggering amount of sheer dumb luck when running
kvm-unit-tests on a 32-bit KVM build. The shrinker just happened to kick
in while running tests and do_shrink_slab() logged an error about trying
to free a negative number of objects. The truly lucky part is that the
kernel just happened to be a slightly stale build, as the shrinker no
longer yells about negative objects as of commit 18bb473e50 ("mm:
vmscan: shrink deferred objects proportional to priority").
vmscan: shrink_slab: mmu_shrink_scan+0x0/0x210 [kvm] negative objects to delete nr=-858993460
Fixes: bc8a3d8925 ("kvm: mmu: Fix overflow on kvm mmu page limit calculation")
Cc: stable@vger.kernel.org
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210804214609.1096003-1-seanjc@google.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
gfn_to_hva_cache is not thread-safe, so it is usually used only within
a vCPU (whose code is protected by vcpu->mutex). The Xen interface
implementation has such a cache in kvm->arch, but it is not really
used except to store the location of the shared info page. Replace
shinfo_set and shinfo_cache with just the value that is passed via
KVM_XEN_ATTR_TYPE_SHARED_INFO; the only complication is that the
initialization value is not zero anymore and therefore kvm_xen_init_vm
needs to be introduced.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM SEV code uses bitmaps to manage ASID states. ASID 0 was always skipped
because it is never used by VM. Thus, in existing code, ASID value and its
bitmap postion always has an 'offset-by-1' relationship.
Both SEV and SEV-ES shares the ASID space, thus KVM uses a dynamic range
[min_asid, max_asid] to handle SEV and SEV-ES ASIDs separately.
Existing code mixes the usage of ASID value and its bitmap position by
using the same variable called 'min_asid'.
Fix the min_asid usage: ensure that its usage is consistent with its name;
allocate extra size for ASID 0 to ensure that each ASID has the same value
with its bitmap position. Add comments on ASID bitmap allocation to clarify
the size change.
Signed-off-by: Mingwei Zhang <mizhang@google.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Marc Orr <marcorr@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Alper Gun <alpergun@google.com>
Cc: Dionna Glaze <dionnaglaze@google.com>
Cc: Sean Christopherson <seanjc@google.com>
Cc: Vipin Sharma <vipinsh@google.com>
Cc: Peter Gonda <pgonda@google.com>
Cc: Joerg Roedel <joro@8bytes.org>
Message-Id: <20210802180903.159381-1-mizhang@google.com>
[Fix up sev_asid_free to also index by ASID, as suggested by Sean
Christopherson, and use nr_asids in sev_cpu_init. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use the raw ASID, not ASID-1, when nullifying the last used VMCB when
freeing an SEV ASID. The consumer, pre_sev_run(), indexes the array by
the raw ASID, thus KVM could get a false negative when checking for a
different VMCB if KVM manages to reallocate the same ASID+VMCB combo for
a new VM.
Note, this cannot cause a functional issue _in the current code_, as
pre_sev_run() also checks which pCPU last did VMRUN for the vCPU, and
last_vmentry_cpu is initialized to -1 during vCPU creation, i.e. is
guaranteed to mismatch on the first VMRUN. However, prior to commit
8a14fe4f0c ("kvm: x86: Move last_cpu into kvm_vcpu_arch as
last_vmentry_cpu"), SVM tracked pCPU on its own and zero-initialized the
last_cpu variable. Thus it's theoretically possible that older versions
of KVM could miss a TLB flush if the first VMRUN is on pCPU0 and the ASID
and VMCB exactly match those of a prior VM.
Fixes: 70cd94e60c ("KVM: SVM: VMRUN should use associated ASID when SEV is enabled")
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Based on our observations, after any vm-exit associated with vPMU, there
are at least two or more perf interfaces to be called for guest counter
emulation, such as perf_event_{pause, read_value, period}(), and each one
will {lock, unlock} the same perf_event_ctx. The frequency of calls becomes
more severe when guest use counters in a multiplexed manner.
Holding a lock once and completing the KVM request operations in the perf
context would introduce a set of impractical new interfaces. So we can
further optimize the vPMU implementation by avoiding repeated calls to
these interfaces in the KVM context for at least one pattern:
After we call perf_event_pause() once, the event will be disabled and its
internal count will be reset to 0. So there is no need to pause it again
or read its value. Once the event is paused, event period will not be
updated until the next time it's resumed or reprogrammed. And there is
also no need to call perf_event_period twice for a non-running counter,
considering the perf_event for a running counter is never paused.
Based on this implementation, for the following common usage of
sampling 4 events using perf on a 4u8g guest:
echo 0 > /proc/sys/kernel/watchdog
echo 25 > /proc/sys/kernel/perf_cpu_time_max_percent
echo 10000 > /proc/sys/kernel/perf_event_max_sample_rate
echo 0 > /proc/sys/kernel/perf_cpu_time_max_percent
for i in `seq 1 1 10`
do
taskset -c 0 perf record \
-e cpu-cycles -e instructions -e branch-instructions -e cache-misses \
/root/br_instr a
done
the average latency of the guest NMI handler is reduced from
37646.7 ns to 32929.3 ns (~1.14x speed up) on the Intel ICX server.
Also, in addition to collecting more samples, no loss of sampling
accuracy was observed compared to before the optimization.
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20210728120705.6855-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Using rmap_get_first() and rmap_remove() for zapping a huge rmap list could be
slow. The easy way is to travers the rmap list, collecting the a/d bits and
free the slots along the way.
Provide a pte_list_destroy() and do exactly that.
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-Id: <20210730220605.26377-1-peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add a counter field into pte_list_desc, so as to simplify the add/remove/loop
logic. E.g., we don't need to loop over the array any more for most reasons.
This will make more sense after we've switched the array size to be larger
otherwise the counter will be a waste.
Initially I wanted to store a tail pointer at the head of the array list so we
don't need to traverse the list at least for pushing new ones (if without the
counter we traverse both the list and the array). However that'll need
slightly more change without a huge lot benefit, e.g., after we grow entry
numbers per array the list traversing is not so expensive.
So let's be simple but still try to get as much benefit as we can with just
these extra few lines of changes (not to mention the code looks easier too
without looping over arrays).
I used the same a test case to fork 500 child and recycle them ("./rmap_fork
500" [1]), this patch further speeds up the total fork time of about 4%, which
is a total of 33% of vanilla kernel:
Vanilla: 473.90 (+-5.93%)
3->15 slots: 366.10 (+-4.94%)
Add counter: 351.00 (+-3.70%)
[1] 825436f825
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-Id: <20210730220602.26327-1-peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently rmap array element only contains 3 entries. However for EPT=N there
could have a lot of guest pages that got tens of even hundreds of rmap entry.
A normal distribution of a 6G guest (even if idle) shows this with rmap count
statistics:
Rmap_Count: 0 1 2-3 4-7 8-15 16-31 32-63 64-127 128-255 256-511 512-1023
Level=4K: 3089171 49005 14016 1363 235 212 15 7 0 0 0
Level=2M: 5951 227 0 0 0 0 0 0 0 0 0
Level=1G: 32 0 0 0 0 0 0 0 0 0 0
If we do some more fork some pages will grow even larger rmap counts.
This patch makes PTE_LIST_EXT bigger so it'll be more efficient for the general
use case of EPT=N as we do list reference less and the loops over PTE_LIST_EXT
will be slightly more efficient; but still not too large so less waste when
array not full.
It should not affecting EPT=Y since EPT normally only has zero or one rmap
entry for each page, so no array is even allocated.
With a test case to fork 500 child and recycle them ("./rmap_fork 500" [1]),
this patch speeds up fork time of about 29%.
Before: 473.90 (+-5.93%)
After: 366.10 (+-4.94%)
[1] 825436f825
Signed-off-by: Peter Xu <peterx@redhat.com>
Message-Id: <20210730220455.26054-6-peterx@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
