KVM SVM changes for 6.7:
- Report KVM_EXIT_SHUTDOWN instead of EINVAL if KVM intercepts SHUTDOWN while
running an SEV-ES guest.
- Clean up handling "failures" when KVM detects it can't emulate the "skip"
action for an instruction that has already been partially emulated. Drop a
hack in the SVM code that was fudging around the emulator code not giving
SVM enough information to do the right thing.
KVM PMU change for 6.7:
- Handle NMI/SMI requests after PMU/PMI requests so that a PMI=>NMI doesn't
require redoing the entire run loop due to the NMI not being detected until
the final kvm_vcpu_exit_request() check before entering the guest.
KVM x86 Xen changes for 6.7:
- Omit "struct kvm_vcpu_xen" entirely when CONFIG_KVM_XEN=n.
- Use the fast path directly from the timer callback when delivering Xen timer
events. Avoid the problematic races with using the fast path by ensuring
the hrtimer isn't running when (re)starting the timer or saving the timer
information (for userspace).
- Follow the lead of upstream Xen and ignore the VCPU_SSHOTTMR_future flag.
KVM x86 MMU changes for 6.7:
- Clean up code that deals with honoring guest MTRRs when the VM has
non-coherent DMA and host MTRRs are ignored, i.e. EPT is enabled.
- Zap EPT entries when non-coherent DMA assignment stops/start to prevent
using stale entries with the wrong memtype.
- Don't ignore guest PAT for CR0.CD=1 && KVM_X86_QUIRK_CD_NW_CLEARED=y, as
there's zero reason to ignore guest PAT if the effective MTRR memtype is WB.
This will also allow for future optimizations of handling guest MTRR updates
for VMs with non-coherent DMA and the quirk enabled.
- Harden the fast page fault path to guard against encountering an invalid
root when walking SPTEs.
KVM x86 misc changes for 6.7:
- Add CONFIG_KVM_MAX_NR_VCPUS to allow supporting up to 4096 vCPUs without
forcing more common use cases to eat the extra memory overhead.
- Add IBPB and SBPB virtualization support.
- Fix a bug where restoring a vCPU snapshot that was taken within 1 second of
creating the original vCPU would cause KVM to try to synchronize the vCPU's
TSC and thus clobber the correct TSC being set by userspace.
- Compute guest wall clock using a single TSC read to avoid generating an
inaccurate time, e.g. if the vCPU is preempted between multiple TSC reads.
- "Virtualize" HWCR.TscFreqSel to make Linux guests happy, which complain
about a "Firmware Bug" if the bit isn't set for select F/M/S combos.
- Don't apply side effects to Hyper-V's synthetic timer on writes from
userspace to fix an issue where the auto-enable behavior can trigger
spurious interrupts, i.e. do auto-enabling only for guest writes.
- Remove an unnecessary kick of all vCPUs when synchronizing the dirty log
without PML enabled.
- Advertise "support" for non-serializing FS/GS base MSR writes as appropriate.
- Use octal notation for file permissions through KVM x86.
- Fix a handful of typo fixes and warts.
Service NMI and SMI requests after PMI requests in vcpu_enter_guest() so
that KVM does not need to cancel and redo the VM-Enter if the guest
configures its PMIs to be delivered as NMIs (likely) or SMIs (unlikely).
Because APIC emulation "injects" NMIs via KVM_REQ_NMI, handling PMI
requests after NMI requests (the likely case) means KVM won't detect the
pending NMI request until the final check for outstanding requests.
Detecting requests at the final stage is costly as KVM has already loaded
guest state, potentially queued events for injection, disabled IRQs,
dropped SRCU, etc., most of which needs to be unwound.
Note that changing the order of request processing doesn't change the end
result, as KVM's final check for outstanding requests prevents entering
the guest until all requests are serviced. I.e. KVM will ultimately
coalesce events (or not) regardless of the ordering.
Using SPEC2017 benchmark programs running along with Intel vtune in a VM
demonstrates that the following code change reduces 800~1500 canceled
VM-Enters per second.
Some glory details:
Probe the invocation to vmx_cancel_injection():
$ perf probe -a vmx_cancel_injection
$ perf stat -a -e probe:vmx_cancel_injection -I 10000 # per 10 seconds
Partial results when SPEC2017 with Intel vtune are running in the VM:
On kernel without the change:
10.010018010 14254 probe:vmx_cancel_injection
20.037646388 15207 probe:vmx_cancel_injection
30.078739816 15261 probe:vmx_cancel_injection
40.114033258 15085 probe:vmx_cancel_injection
50.149297460 15112 probe:vmx_cancel_injection
60.185103088 15104 probe:vmx_cancel_injection
On kernel with the change:
10.003595390 40 probe:vmx_cancel_injection
20.017855682 31 probe:vmx_cancel_injection
30.028355883 34 probe:vmx_cancel_injection
40.038686298 31 probe:vmx_cancel_injection
50.048795162 20 probe:vmx_cancel_injection
60.069057747 19 probe:vmx_cancel_injection
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Mingwei Zhang <mizhang@google.com>
Link: https://lore.kernel.org/r/20231002040839.2630027-1-mizhang@google.com
[sean: hoist PMU/PMI above SMI too, massage changelog]
Signed-off-by: Sean Christopherson <seanjc@google.com>
Hyper-V enabled Windows Server 2022 KVM VM cannot be started on Zen1 Ryzen
since it crashes at boot with SYSTEM_THREAD_EXCEPTION_NOT_HANDLED +
STATUS_PRIVILEGED_INSTRUCTION (in other words, because of an unexpected #GP
in the guest kernel).
This is because Windows tries to set bit 8 in MSR_AMD64_TW_CFG and can't
handle receiving a #GP when doing so.
Give this MSR the same treatment that commit 2e32b71906
("x86, kvm: Add MSR_AMD64_BU_CFG2 to the list of ignored MSRs") gave
MSR_AMD64_BU_CFG2 under justification that this MSR is baremetal-relevant
only.
Although apparently it was then needed for Linux guests, not Windows as in
this case.
With this change, the aforementioned guest setup is able to finish booting
successfully.
This issue can be reproduced either on a Summit Ridge Ryzen (with
just "-cpu host") or on a Naples EPYC (with "-cpu host,stepping=1" since
EPYC is ordinarily stepping 2).
Alternatively, userspace could solve the problem by using MSR filters, but
forcing every userspace to define a filter isn't very friendly and doesn't
add much, if any, value. The only potential hiccup is if one of these
"baremetal-only" MSRs ever requires actual emulation and/or has F/M/S
specific behavior. But if that happens, then KVM can still punt *that*
handling to userspace since userspace MSR filters "win" over KVM's default
handling.
Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/1ce85d9c7c9e9632393816cf19c902e0a3f411f1.1697731406.git.maciej.szmigiero@oracle.com
[sean: call out MSR filtering alternative]
Signed-off-by: Sean Christopherson <seanjc@google.com>
Stop kicking vCPUs in kvm_arch_sync_dirty_log() when PML is disabled.
Kicking vCPUs when PML is disabled serves no purpose and could
negatively impact guest performance.
This restores KVM's behavior to prior to 5.12 commit a018eba538 ("KVM:
x86: Move MMU's PML logic to common code"), which replaced a
static_call_cond(kvm_x86_flush_log_dirty) with unconditional calls to
kvm_vcpu_kick().
Fixes: a018eba538 ("KVM: x86: Move MMU's PML logic to common code")
Signed-off-by: David Matlack <dmatlack@google.com>
Link: https://lore.kernel.org/r/20231016221228.1348318-1-dmatlack@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
KVM x86/pmu fixes for 6.6:
- Truncate writes to PMU counters to the counter's width to avoid spurious
overflows when emulating counter events in software.
- Set the LVTPC entry mask bit when handling a PMI (to match Intel-defined
architectural behavior).
- Treat KVM_REQ_PMI as a wake event instead of queueing host IRQ work to
kick the guest out of emulated halt.
Mask off xfeatures that aren't exposed to the guest only when saving guest
state via KVM_GET_XSAVE{2} instead of modifying user_xfeatures directly.
Preserving the maximal set of xfeatures in user_xfeatures restores KVM's
ABI for KVM_SET_XSAVE, which prior to commit ad856280dd ("x86/kvm/fpu:
Limit guest user_xfeatures to supported bits of XCR0") allowed userspace
to load xfeatures that are supported by the host, irrespective of what
xfeatures are exposed to the guest.
There is no known use case where userspace *intentionally* loads xfeatures
that aren't exposed to the guest, but the bug fixed by commit ad856280dd
was specifically that KVM_GET_SAVE{2} would save xfeatures that weren't
exposed to the guest, e.g. would lead to userspace unintentionally loading
guest-unsupported xfeatures when live migrating a VM.
Restricting KVM_SET_XSAVE to guest-supported xfeatures is especially
problematic for QEMU-based setups, as QEMU has a bug where instead of
terminating the VM if KVM_SET_XSAVE fails, QEMU instead simply stops
loading guest state, i.e. resumes the guest after live migration with
incomplete guest state, and ultimately results in guest data corruption.
Note, letting userspace restore all host-supported xfeatures does not fix
setups where a VM is migrated from a host *without* commit ad856280dd,
to a target with a subset of host-supported xfeatures. However there is
no way to safely address that scenario, e.g. KVM could silently drop the
unsupported features, but that would be a clear violation of KVM's ABI and
so would require userspace to opt-in, at which point userspace could
simply be updated to sanitize the to-be-loaded XSAVE state.
Reported-by: Tyler Stachecki <stachecki.tyler@gmail.com>
Closes: https://lore.kernel.org/all/20230914010003.358162-1-tstachecki@bloomberg.net
Fixes: ad856280dd ("x86/kvm/fpu: Limit guest user_xfeatures to supported bits of XCR0")
Cc: stable@vger.kernel.org
Cc: Leonardo Bras <leobras@redhat.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Acked-by: Dave Hansen <dave.hansen@linux.intel.com>
Message-Id: <20230928001956.924301-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Plumb an xfeatures mask into __copy_xstate_to_uabi_buf() so that KVM can
constrain which xfeatures are saved into the userspace buffer without
having to modify the user_xfeatures field in KVM's guest_fpu state.
KVM's ABI for KVM_GET_XSAVE{2} is that features that are not exposed to
guest must not show up in the effective xstate_bv field of the buffer.
Saving only the guest-supported xfeatures allows userspace to load the
saved state on a different host with a fewer xfeatures, so long as the
target host supports the xfeatures that are exposed to the guest.
KVM currently sets user_xfeatures directly to restrict KVM_GET_XSAVE{2} to
the set of guest-supported xfeatures, but doing so broke KVM's historical
ABI for KVM_SET_XSAVE, which allows userspace to load any xfeatures that
are supported by the *host*.
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20230928001956.924301-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Zap KVM TDP when noncoherent DMA assignment starts (noncoherent dma count
transitions from 0 to 1) or stops (noncoherent dma count transitions
from 1 to 0). Before the zap, test if guest MTRR is to be honored after
the assignment starts or was honored before the assignment stops.
When there's no noncoherent DMA device, EPT memory type is
((MTRR_TYPE_WRBACK << VMX_EPT_MT_EPTE_SHIFT) | VMX_EPT_IPAT_BIT)
When there're noncoherent DMA devices, EPT memory type needs to honor
guest CR0.CD and MTRR settings.
So, if noncoherent DMA count transitions between 0 and 1, EPT leaf entries
need to be zapped to clear stale memory type.
This issue might be hidden when the device is statically assigned with
VFIO adding/removing MMIO regions of the noncoherent DMA devices for
several times during guest boot, and current KVM MMU will call
kvm_mmu_zap_all_fast() on the memslot removal.
But if the device is hot-plugged, or if the guest has mmio_always_on for
the device, the MMIO regions of it may only be added for once, then there's
no path to do the EPT entries zapping to clear stale memory type.
Therefore do the EPT zapping when noncoherent assignment starts/stops to
ensure stale entries cleaned away.
Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
Link: https://lore.kernel.org/r/20230714065223.20432-1-yan.y.zhao@intel.com
[sean: fix misspelled words in comment and changelog]
Signed-off-by: Sean Christopherson <seanjc@google.com>
The legacy API for setting the TSC is fundamentally broken, and only
allows userspace to set a TSC "now", without any way to account for
time lost between the calculation of the value, and the kernel eventually
handling the ioctl.
To work around this, KVM has a hack which, if a TSC is set with a value
which is within a second's worth of the last TSC "written" to any vCPU in
the VM, assumes that userspace actually intended the two TSC values to be
in sync and adjusts the newly-written TSC value accordingly.
Thus, when a VMM restores a guest after suspend or migration using the
legacy API, the TSCs aren't necessarily *right*, but at least they're
in sync.
This trick falls down when restoring a guest which genuinely has been
running for less time than the 1 second of imprecision KVM allows for in
in the legacy API. On *creation*, the first vCPU starts its TSC counting
from zero, and the subsequent vCPUs synchronize to that. But then when
the VMM tries to restore a vCPU's intended TSC, because the VM has been
alive for less than 1 second and KVM's default TSC value for new vCPU's is
'0', the intended TSC is within a second of the last "written" TSC and KVM
incorrectly adjusts the intended TSC in an attempt to synchronize.
But further hacks can be piled onto KVM's existing hackish ABI, and
declare that the *first* value written by *userspace* (on any vCPU)
should not be subject to this "correction", i.e. KVM can assume that the
first write from userspace is not an attempt to sync up with TSC values
that only come from the kernel's default vCPU creation.
To that end: Add a flag, kvm->arch.user_set_tsc, protected by
kvm->arch.tsc_write_lock, to record that a TSC for at least one vCPU in
the VM *has* been set by userspace, and make the 1-second slop hack only
trigger if user_set_tsc is already set.
Note that userspace can explicitly request a *synchronization* of the
TSC by writing zero. For the purpose of user_set_tsc, an explicit
synchronization counts as "setting" the TSC, i.e. if userspace then
subsequently writes an explicit non-zero value which happens to be within
1 second of the previous value, the new value will be "corrected". This
behavior is deliberate, as treating explicit synchronization as "setting"
the TSC preserves KVM's existing behaviour inasmuch as possible (KVM
always applied the 1-second "correction" regardless of whether the write
came from userspace vs. the kernel).
Reported-by: Yong He <alexyonghe@tencent.com>
Closes: https://bugzilla.kernel.org/show_bug.cgi?id=217423
Suggested-by: Oliver Upton <oliver.upton@linux.dev>
Original-by: Oliver Upton <oliver.upton@linux.dev>
Original-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Like Xu <likexu@tencent.com>
Tested-by: Yong He <alexyonghe@tencent.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20231008025335.7419-1-likexu@tencent.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
On certain CPUs, Linux guests expect HWCR.TscFreqSel[bit 24] to be
set. If it isn't set, they complain:
[Firmware Bug]: TSC doesn't count with P0 frequency!
Allow userspace (and the guest) to set this bit in the virtual HWCR to
eliminate the above complaint.
Allow the guest to write the bit even though its is R/O on *some* CPUs.
Like many bits in HWRC, TscFreqSel is not architectural at all. On Family
10h[1], it was R/W and powered on as 0. In Family 15h, one of the "changes
relative to Family 10H Revision D processors[2] was:
• MSRC001_0015 [Hardware Configuration (HWCR)]:
• Dropped TscFreqSel; TSC can no longer be selected to run at NB P0-state.
Despite the "Dropped" above, that same document later describes
HWCR[bit 24] as follows:
TscFreqSel: TSC frequency select. Read-only. Reset: 1. 1=The TSC
increments at the P0 frequency
If the guest clears the bit, the worst case scenario is the guest will be
no worse off than it is today, e.g. the whining may return after a guest
clears the bit and kexec()'s into a new kernel.
[1] https://www.amd.com/content/dam/amd/en/documents/archived-tech-docs/programmer-references/31116.pdf
[2] https://www.amd.com/content/dam/amd/en/documents/archived-tech-docs/programmer-references/42301_15h_Mod_00h-0Fh_BKDG.pdf,
Signed-off-by: Jim Mattson <jmattson@google.com>
Link: https://lore.kernel.org/r/20230929230246.1954854-3-jmattson@google.com
[sean: elaborate on why the bit is writable by the guest]
Signed-off-by: Sean Christopherson <seanjc@google.com>
When populating the guest's PV wall clock information, KVM currently does
a simple 'kvm_get_real_ns() - get_kvmclock_ns(kvm)'. This is an antipattern
which should be avoided; when working with the relationship between two
clocks, it's never correct to obtain one of them "now" and then the other
at a slightly different "now" after an unspecified period of preemption
(which might not even be under the control of the kernel, if this is an
L1 hosting an L2 guest under nested virtualization).
Add a kvm_get_wall_clock_epoch() function to return the guest wall clock
epoch in nanoseconds using the same method as __get_kvmclock() — by using
kvm_get_walltime_and_clockread() to calculate both the wall clock and KVM
clock time from a *single* TSC reading.
The condition using get_cpu_tsc_khz() is equivalent to the version in
__get_kvmclock() which separately checks for the CONSTANT_TSC feature or
the per-CPU cpu_tsc_khz. Which is what get_cpu_tsc_khz() does anyway.
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Link: https://lore.kernel.org/r/bfc6d3d7cfb88c47481eabbf5a30a264c58c7789.camel@infradead.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Add support for the AMD Selective Branch Predictor Barrier (SBPB) by
advertising the CPUID bit and handling PRED_CMD writes accordingly.
Note, like SRSO_NO and IBPB_BRTYPE before it, advertise support for SBPB
even if it's not enumerated by in the raw CPUID. Some CPUs that gained
support via a uCode patch don't report SBPB via CPUID (the kernel forces
the flag).
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
Link: https://lore.kernel.org/r/a4ab1e7fe50096d50fde33e739ed2da40b41ea6a.1692919072.git.jpoimboe@kernel.org
Co-developed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Treat EMULTYPE_SKIP failures on SEV guests as unhandleable emulation
instead of simply resuming the guest, and drop the hack-a-fix which
effects that behavior for the INT3/INTO injection path. If KVM can't
skip an instruction for which KVM has already done partial emulation,
resuming the guest is undesirable as doing so may corrupt guest state.
Link: https://lore.kernel.org/r/20230825013621.2845700-5-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Refactor and rename can_emulate_instruction() to allow vendor code to
return more than true/false, e.g. to explicitly differentiate between
"retry", "fault", and "unhandleable". For now, just do the plumbing, a
future patch will expand SVM's implementation to signal outright failure
if KVM attempts EMULTYPE_SKIP on an SEV guest.
No functional change intended (or rather, none that are visible to the
guest or userspace).
Link: https://lore.kernel.org/r/20230825013621.2845700-4-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Userspace can directly modify the content of vCPU's CR0, CR3, and CR4 via
KVM_SYNC_X86_SREGS and KVM_SET_SREGS{,2}. Make sure that KVM flushes guest
TLB entries and paging-structure caches if a (partial) guest TLB flush is
architecturally required based on the CRn changes. To keep things simple,
flush whenever KVM resets the MMU context, i.e. if any bits in CR0, CR3,
CR4, or EFER are modified. This is extreme overkill, but stuffing state
from userspace is not such a hot path that preserving guest TLB state is a
priority.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Michal Luczaj <mhal@rbox.co>
Link: https://lore.kernel.org/r/20230814222358.707877-3-mhal@rbox.co
[sean: call out that the flushing on MMU context resets is for simplicity]
Signed-off-by: Sean Christopherson <seanjc@google.com>
When the irq_work callback, kvm_pmi_trigger_fn(), is invoked during a
VM-exit that also invokes __kvm_perf_overflow() as a result of
instruction emulation, kvm_pmu_deliver_pmi() will be called twice
before the next VM-entry.
Calling kvm_pmu_deliver_pmi() twice is unlikely to be problematic now that
KVM sets the LVTPC mask bit when delivering a PMI. But using IRQ work to
trigger the PMI is still broken, albeit very theoretically.
E.g. if the self-IPI to trigger IRQ work is be delayed long enough for the
vCPU to be migrated to a different pCPU, then it's possible for
kvm_pmi_trigger_fn() to race with the kvm_pmu_deliver_pmi() from
KVM_REQ_PMI and still generate two PMIs.
KVM could set the mask bit using an atomic operation, but that'd just be
piling on unnecessary code to workaround what is effectively a hack. The
*only* reason KVM uses IRQ work is to ensure the PMI is treated as a wake
event, e.g. if the vCPU just executed HLT.
Remove the irq_work callback for synthesizing a PMI, and all of the
logic for invoking it. Instead, to prevent a vcpu from leaving C0 with
a PMI pending, add a check for KVM_REQ_PMI to kvm_vcpu_has_events().
Fixes: 9cd803d496 ("KVM: x86: Update vPMCs when retiring instructions")
Signed-off-by: Jim Mattson <jmattson@google.com>
Tested-by: Mingwei Zhang <mizhang@google.com>
Tested-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Signed-off-by: Mingwei Zhang <mizhang@google.com>
Link: https://lore.kernel.org/r/20230925173448.3518223-2-mizhang@google.com
[sean: massage changelog]
Signed-off-by: Sean Christopherson <seanjc@google.com>
Stop zapping invalidate TDP MMU roots via work queue now that KVM
preserves TDP MMU roots until they are explicitly invalidated. Zapping
roots asynchronously was effectively a workaround to avoid stalling a vCPU
for an extended during if a vCPU unloaded a root, which at the time
happened whenever the guest toggled CR0.WP (a frequent operation for some
guest kernels).
While a clever hack, zapping roots via an unbound worker had subtle,
unintended consequences on host scheduling, especially when zapping
multiple roots, e.g. as part of a memslot. Because the work of zapping a
root is no longer bound to the task that initiated the zap, things like
the CPU affinity and priority of the original task get lost. Losing the
affinity and priority can be especially problematic if unbound workqueues
aren't affined to a small number of CPUs, as zapping multiple roots can
cause KVM to heavily utilize the majority of CPUs in the system, *beyond*
the CPUs KVM is already using to run vCPUs.
When deleting a memslot via KVM_SET_USER_MEMORY_REGION, the async root
zap can result in KVM occupying all logical CPUs for ~8ms, and result in
high priority tasks not being scheduled in in a timely manner. In v5.15,
which doesn't preserve unloaded roots, the issues were even more noticeable
as KVM would zap roots more frequently and could occupy all CPUs for 50ms+.
Consuming all CPUs for an extended duration can lead to significant jitter
throughout the system, e.g. on ChromeOS with virtio-gpu, deleting memslots
is a semi-frequent operation as memslots are deleted and recreated with
different host virtual addresses to react to host GPU drivers allocating
and freeing GPU blobs. On ChromeOS, the jitter manifests as audio blips
during games due to the audio server's tasks not getting scheduled in
promptly, despite the tasks having a high realtime priority.
Deleting memslots isn't exactly a fast path and should be avoided when
possible, and ChromeOS is working towards utilizing MAP_FIXED to avoid the
memslot shenanigans, but KVM is squarely in the wrong. Not to mention
that removing the async zapping eliminates a non-trivial amount of
complexity.
Note, one of the subtle behaviors hidden behind the async zapping is that
KVM would zap invalidated roots only once (ignoring partial zaps from
things like mmu_notifier events). Preserve this behavior by adding a flag
to identify roots that are scheduled to be zapped versus roots that have
already been zapped but not yet freed.
Add a comment calling out why kvm_tdp_mmu_invalidate_all_roots() can
encounter invalid roots, as it's not at all obvious why zapping
invalidated roots shouldn't simply zap all invalid roots.
Reported-by: Pattara Teerapong <pteerapong@google.com>
Cc: David Stevens <stevensd@google.com>
Cc: Yiwei Zhang<zzyiwei@google.com>
Cc: Paul Hsia <paulhsia@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20230916003916.2545000-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Bury the declaration of the page-track helpers that are intended only for
internal KVM use in a "private" header. In addition to guarding against
unwanted usage of the internal-only helpers, dropping their definitions
avoids exposing other structures that should be KVM-internal, e.g. for
memslots. This is a baby step toward making kvm_host.h a KVM-internal
header in the very distant future.
Tested-by: Yongwei Ma <yongwei.ma@intel.com>
Link: https://lore.kernel.org/r/20230729013535.1070024-22-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add a new page-track hook, track_remove_region(), that is called when a
memslot DELETE operation is about to be committed. The "remove" hook
will be used by KVMGT and will effectively replace the existing
track_flush_slot() altogether now that KVM itself doesn't rely on the
"flush" hook either.
The "flush" hook is flawed as it's invoked before the memslot operation
is guaranteed to succeed, i.e. KVM might ultimately keep the existing
memslot without notifying external page track users, a.k.a. KVMGT. In
practice, this can't currently happen on x86, but there are no guarantees
that won't change in the future, not to mention that "flush" does a very
poor job of describing what is happening.
Pass in the gfn+nr_pages instead of the slot itself so external users,
i.e. KVMGT, don't need to exposed to KVM internals (memslots). This will
help set the stage for additional cleanups to the page-track APIs.
Opportunistically align the existing srcu_read_lock_held() usage so that
the new case doesn't stand out like a sore thumb (and not aligning the
new code makes bots unhappy).
Cc: Zhenyu Wang <zhenyuw@linux.intel.com>
Tested-by: Yongwei Ma <yongwei.ma@intel.com>
Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
Co-developed-by: Sean Christopherson <seanjc@google.com>
Link: https://lore.kernel.org/r/20230729013535.1070024-19-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Disallow moving memslots if the VM has external page-track users, i.e. if
KVMGT is being used to expose a virtual GPU to the guest, as KVMGT doesn't
correctly handle moving memory regions.
Note, this is potential ABI breakage! E.g. userspace could move regions
that aren't shadowed by KVMGT without harming the guest. However, the
only known user of KVMGT is QEMU, and QEMU doesn't move generic memory
regions. KVM's own support for moving memory regions was also broken for
multiple years (albeit for an edge case, but arguably moving RAM is
itself an edge case), e.g. see commit edd4fa37ba ("KVM: x86: Allocate
new rmap and large page tracking when moving memslot").
Reviewed-by: Yan Zhao <yan.y.zhao@intel.com>
Tested-by: Yongwei Ma <yongwei.ma@intel.com>
Link: https://lore.kernel.org/r/20230729013535.1070024-17-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move x86's implementation of kvm_arch_flush_shadow_{all,memslot}() into
mmu.c, and make kvm_mmu_zap_all() static as it was globally visible only
for kvm_arch_flush_shadow_all(). This will allow refactoring
kvm_arch_flush_shadow_memslot() to call kvm_mmu_zap_all() directly without
having to expose kvm_mmu_zap_all_fast() outside of mmu.c. Keeping
everything in mmu.c will also likely simplify supporting TDX, which
intends to do zap only relevant SPTEs on memslot updates.
No functional change intended.
Suggested-by: Yan Zhao <yan.y.zhao@intel.com>
Tested-by: Yongwei Ma <yongwei.ma@intel.com>
Link: https://lore.kernel.org/r/20230729013535.1070024-13-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM x86 changes for 6.6:
- Misc cleanups
- Retry APIC optimized recalculation if a vCPU is added/enabled
- Overhaul emergency reboot code to bring SVM up to par with VMX, tie the
"emergency disabling" behavior to KVM actually being loaded, and move all of
the logic within KVM
- Fix user triggerable WARNs in SVM where KVM incorrectly assumes the TSC
ratio MSR can diverge from the default iff TSC scaling is enabled, and clean
up related code
- Add a framework to allow "caching" feature flags so that KVM can check if
the guest can use a feature without needing to search guest CPUID
KVM: x86: Selftests changes for 6.6:
- Add testcases to x86's sync_regs_test for detecting KVM TOCTOU bugs
- Add support for printf() in guest code and covert all guest asserts to use
printf-based reporting
- Clean up the PMU event filter test and add new testcases
- Include x86 selftests in the KVM x86 MAINTAINERS entry
KVM/arm64 updates for Linux 6.6
- Add support for TLB range invalidation of Stage-2 page tables,
avoiding unnecessary invalidations. Systems that do not implement
range invalidation still rely on a full invalidation when dealing
with large ranges.
- Add infrastructure for forwarding traps taken from a L2 guest to
the L1 guest, with L0 acting as the dispatcher, another baby step
towards the full nested support.
- Simplify the way we deal with the (long deprecated) 'CPU target',
resulting in a much needed cleanup.
- Fix another set of PMU bugs, both on the guest and host sides,
as we seem to never have any shortage of those...
- Relax the alignment requirements of EL2 VA allocations for
non-stack allocations, as we were otherwise wasting a lot of that
precious VA space.
- The usual set of non-functional cleanups, although I note the lack
of spelling fixes...
Use the governed feature framework to track if XSAVES is "enabled", i.e.
if XSAVES can be used by the guest. Add a comment in the SVM code to
explain the very unintuitive logic of deliberately NOT checking if XSAVES
is enumerated in the guest CPUID model.
No functional change intended.
Reviewed-by: Yuan Yao <yuan.yao@intel.com>
Link: https://lore.kernel.org/r/20230815203653.519297-7-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Drop the WARN in KVM_RUN that asserts that KVM isn't using the hypervisor
timer, a.k.a. the VMX preemption timer, for a vCPU that is in the
UNINITIALIZIED activity state. The intent of the WARN is to sanity check
that KVM won't drop a timer interrupt due to an unexpected transition to
UNINITIALIZED, but unfortunately userspace can use various ioctl()s to
force the unexpected state.
Drop the sanity check instead of switching from the hypervisor timer to a
software based timer, as the only reason to switch to a software timer
when a vCPU is blocking is to ensure the timer interrupt wakes the vCPU,
but said interrupt isn't a valid wake event for vCPUs in UNINITIALIZED
state *and* the interrupt will be dropped in the end.
Reported-by: Yikebaer Aizezi <yikebaer61@gmail.com>
Closes: https://lore.kernel.org/all/CALcu4rbFrU4go8sBHk3FreP+qjgtZCGcYNpSiEXOLm==qFv7iQ@mail.gmail.com
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Link: https://lore.kernel.org/r/20230808232057.2498287-1-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Move kvm_arch_flush_remote_tlbs_memslot() to common code and drop
"arch_" from the name. kvm_arch_flush_remote_tlbs_memslot() is just a
range-based TLB invalidation where the range is defined by the memslot.
Now that kvm_flush_remote_tlbs_range() can be called from common code we
can just use that and drop a bunch of duplicate code from the arch
directories.
Note this adds a lockdep assertion for slots_lock being held when
calling kvm_flush_remote_tlbs_memslot(), which was previously only
asserted on x86. MIPS has calls to kvm_flush_remote_tlbs_memslot(),
but they all hold the slots_lock, so the lockdep assertion continues to
hold true.
Also drop the CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT ifdef gating
kvm_flush_remote_tlbs_memslot(), since it is no longer necessary.
Signed-off-by: David Matlack <dmatlack@google.com>
Signed-off-by: Raghavendra Rao Ananta <rananta@google.com>
Reviewed-by: Gavin Shan <gshan@redhat.com>
Reviewed-by: Shaoqin Huang <shahuang@redhat.com>
Acked-by: Anup Patel <anup@brainfault.org>
Acked-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20230811045127.3308641-7-rananta@google.com
Pull x86/gds fixes from Dave Hansen:
"Mitigate Gather Data Sampling issue:
- Add Base GDS mitigation
- Support GDS_NO under KVM
- Fix a documentation typo"
* tag 'gds-for-linus-2023-08-01' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
Documentation/x86: Fix backwards on/off logic about YMM support
KVM: Add GDS_NO support to KVM
x86/speculation: Add Kconfig option for GDS
x86/speculation: Add force option to GDS mitigation
x86/speculation: Add Gather Data Sampling mitigation
Drop the @offset and @multiplier params from the kvm_x86_ops hooks for
propagating TSC offsets/multipliers into hardware, and instead have the
vendor implementations pull the information directly from the vCPU
structure. The respective vCPU fields _must_ be written at the same
time in order to maintain consistent state, i.e. it's not random luck
that the value passed in by all callers is grabbed from the vCPU.
Explicitly grabbing the value from the vCPU field in SVM's implementation
in particular will allow for additional cleanup without introducing even
more subtle dependencies. Specifically, SVM can skip the WRMSR if guest
state isn't loaded, i.e. svm_prepare_switch_to_guest() will load the
correct value for the vCPU prior to entering the guest.
This also reconciles KVM's handling of related values that are stored in
the vCPU, as svm_write_tsc_offset() already assumes/requires the caller
to have updated l1_tsc_offset.
Link: https://lore.kernel.org/r/20230729011608.1065019-6-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Snapshot the host's MSR_IA32_ARCH_CAPABILITIES, if it's supported, instead
of reading the MSR every time KVM wants to query the host state, e.g. when
initializing the default value during vCPU creation. The paths that query
ARCH_CAPABILITIES aren't particularly performance sensitive, but creating
vCPUs is a frequent enough operation that burning 8 bytes is a good
trade-off.
Alternatively, KVM could add a field in kvm_caps and thus skip the
on-demand calculations entirely, but a pure snapshot isn't possible due to
the way KVM handles the l1tf_vmx_mitigation module param. And unlike the
other "supported" fields in kvm_caps, KVM doesn't enforce the "supported"
value, i.e. KVM treats ARCH_CAPABILITIES like a CPUID leaf and lets
userspace advertise whatever it wants. Those problems are solvable, but
it's not clear there is real benefit versus snapshotting the host value,
and grabbing the host value will allow additional cleanup of KVM's
FB_CLEAR_CTRL code.
Link: https://lore.kernel.org/all/20230524061634.54141-2-chao.gao@intel.com
Cc: Chao Gao <chao.gao@intel.com>
Cc: Xiaoyao Li <xiaoyao.li@intel.com>
Reviewed-by: Chao Gao <chao.gao@intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20230607004311.1420507-2-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Remove x86_emulate_ops::guest_has_long_mode along with its implementation,
emulator_guest_has_long_mode(). It has been unused since commit
1d0da94cda ("KVM: x86: do not go through ctxt->ops when emulating rsm").
No functional change intended.
Signed-off-by: Michal Luczaj <mhal@rbox.co>
Link: https://lore.kernel.org/r/20230718101809.1249769-1-mhal@rbox.co
Signed-off-by: Sean Christopherson <seanjc@google.com>
In a spirit of using a sledgehammer to crack a nut, make sync_regs() feed
__set_sregs() and kvm_vcpu_ioctl_x86_set_vcpu_events() with kernel's own
copy of data.
Both __set_sregs() and kvm_vcpu_ioctl_x86_set_vcpu_events() assume they
have exclusive rights to structs they operate on. While this is true when
coming from an ioctl handler (caller makes a local copy of user's data),
sync_regs() breaks this contract; a pointer to a user-modifiable memory
(vcpu->run->s.regs) is provided. This can lead to a situation when incoming
data is checked and/or sanitized only to be re-set by a user thread running
in parallel.
Signed-off-by: Michal Luczaj <mhal@rbox.co>
Fixes: 01643c51bf ("KVM: x86: KVM_CAP_SYNC_REGS")
Link: https://lore.kernel.org/r/20230728001606.2275586-2-mhal@rbox.co
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reject KVM_SET_SREGS{2} with -EINVAL if the incoming CR0 is invalid,
e.g. due to setting bits 63:32, illegal combinations, or to a value that
isn't allowed in VMX (non-)root mode. The VMX checks in particular are
"fun" as failure to disallow Real Mode for an L2 that is configured with
unrestricted guest disabled, when KVM itself has unrestricted guest
enabled, will result in KVM forcing VM86 mode to virtual Real Mode for
L2, but then fail to unwind the related metadata when synthesizing a
nested VM-Exit back to L1 (which has unrestricted guest enabled).
Opportunistically fix a benign typo in the prototype for is_valid_cr4().
Cc: stable@vger.kernel.org
Reported-by: syzbot+5feef0b9ee9c8e9e5689@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/000000000000f316b705fdf6e2b4@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20230613203037.1968489-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Temporarily acquire kvm->srcu for read when potentially emulating WRMSR in
the VM-Exit fastpath handler, as several of the common helpers used during
emulation expect the caller to provide SRCU protection. E.g. if the guest
is counting instructions retired, KVM will query the PMU event filter when
stepping over the WRMSR.
dump_stack+0x85/0xdf
lockdep_rcu_suspicious+0x109/0x120
pmc_event_is_allowed+0x165/0x170
kvm_pmu_trigger_event+0xa5/0x190
handle_fastpath_set_msr_irqoff+0xca/0x1e0
svm_vcpu_run+0x5c3/0x7b0 [kvm_amd]
vcpu_enter_guest+0x2108/0x2580
Alternatively, check_pmu_event_filter() could acquire kvm->srcu, but this
isn't the first bug of this nature, e.g. see commit 5c30e8101e ("KVM:
SVM: Skip WRMSR fastpath on VM-Exit if next RIP isn't valid"). Providing
protection for the entirety of WRMSR emulation will allow reverting the
aforementioned commit, and will avoid having to play whack-a-mole when new
uses of SRCU-protected structures are inevitably added in common emulation
helpers.
Fixes: dfdeda67ea ("KVM: x86/pmu: Prevent the PMU from counting disallowed events")
Reported-by: Greg Thelen <gthelen@google.com>
Reported-by: Aaron Lewis <aaronlewis@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20230721224337.2335137-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
As was attempted commit 14717e2031 ("kvm: Conditionally register IRQ
bypass consumer"): "if we don't support a mechanism for bypassing IRQs,
don't register as a consumer. Initially this applied to AMD processors,
but when AVIC support was implemented for assigned devices,
kvm_arch_has_irq_bypass() was always returning true.
We can still skip registering the consumer where enable_apicv
or posted-interrupts capability is unsupported or globally disabled.
This eliminates meaningless dev_info()s when the connect fails
between producer and consumer", such as on Linux hosts where enable_apicv
or posted-interrupts capability is unsupported or globally disabled.
Cc: Alex Williamson <alex.williamson@redhat.com>
Reported-by: Yong He <alexyonghe@tencent.com>
Closes: https://bugzilla.kernel.org/show_bug.cgi?id=217379
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20230724111236.76570-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The code was blindly assuming that kvm_cpu_get_interrupt never returns -1
when there is a pending interrupt.
While this should be true, a bug in KVM can still cause this.
If -1 is returned, the code before this patch was converting it to 0xFF,
and 0xFF interrupt was injected to the guest, which results in an issue
which was hard to debug.
Add WARN_ON_ONCE to catch this case and skip the injection
if this happens again.
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20230726135945.260841-4-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
