platform/kernel/linux-starfive.git
2 years agoKVM: SVM: Allow AVIC support on system w/ physical APIC ID > 255
Suravee Suthikulpanit [Fri, 11 Feb 2022 00:08:51 +0000 (18:08 -0600)]
KVM: SVM: Allow AVIC support on system w/ physical APIC ID > 255

Expand KVM's mask for the AVIC host physical ID to the full 12 bits defined
by the architecture.  The number of bits consumed by hardware is model
specific, e.g. early CPUs ignored bits 11:8, but there is no way for KVM
to enumerate the "true" size.  So, KVM must allow using all bits, else it
risks rejecting completely legal x2APIC IDs on newer CPUs.

This means KVM relies on hardware to not assign x2APIC IDs that exceed the
"true" width of the field, but presumably hardware is smart enough to tie
the width to the max x2APIC ID.  KVM also relies on hardware to support at
least 8 bits, as the legacy xAPIC ID is writable by software.  But, those
assumptions are unavoidable due to the lack of any way to enumerate the
"true" width.

Cc: stable@vger.kernel.org
Cc: Maxim Levitsky <mlevitsk@redhat.com>
Suggested-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Fixes: 44a95dae1d22 ("KVM: x86: Detect and Initialize AVIC support")
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Message-Id: <20220211000851.185799-1-suravee.suthikulpanit@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Add test to populate a VM with the max possible guest mem
Sean Christopherson [Sat, 26 Feb 2022 00:15:46 +0000 (00:15 +0000)]
KVM: selftests: Add test to populate a VM with the max possible guest mem

Add a selftest that enables populating a VM with the maximum amount of
guest memory allowed by the underlying architecture.  Abuse KVM's
memslots by mapping a single host memory region into multiple memslots so
that the selftest doesn't require a system with terabytes of RAM.

Default to 512gb of guest memory, which isn't all that interesting, but
should work on all MMUs and doesn't take an exorbitant amount of memory
or time.  E.g. testing with ~64tb of guest memory takes the better part
of an hour, and requires 200gb of memory for KVM's page tables when using
4kb pages.

To inflicit maximum abuse on KVM' MMU, default to 4kb pages (or whatever
the not-hugepage size is) in the backing store (memfd).  Use memfd for
the host backing store to ensure that hugepages are guaranteed when
requested, and to give the user explicit control of the size of hugepage
being tested.

By default, spin up as many vCPUs as there are available to the selftest,
and distribute the work of dirtying each 4kb chunk of memory across all
vCPUs.  Dirtying guest memory forces KVM to populate its page tables, and
also forces KVM to write back accessed/dirty information to struct page
when the guest memory is freed.

On x86, perform two passes with a MMU context reset between each pass to
coerce KVM into dropping all references to the MMU root, e.g. to emulate
a vCPU dropping the last reference.  Perform both passes and all
rendezvous on all architectures in the hope that arm64 and s390x can gain
similar shenanigans in the future.

Measure and report the duration of each operation, which is helpful not
only to verify the test is working as intended, but also to easily
evaluate the performance differences different page sizes.

Provide command line options to limit the amount of guest memory, set the
size of each slot (i.e. of the host memory region), set the number of
vCPUs, and to enable usage of hugepages.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-29-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Define cpu_relax() helpers for s390 and x86
Sean Christopherson [Sat, 26 Feb 2022 00:15:45 +0000 (00:15 +0000)]
KVM: selftests: Define cpu_relax() helpers for s390 and x86

Add cpu_relax() for s390 and x86 for use in arch-agnostic tests.  arm64
already defines its own version.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-28-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Split out helper to allocate guest mem via memfd
Sean Christopherson [Sat, 26 Feb 2022 00:15:44 +0000 (00:15 +0000)]
KVM: selftests: Split out helper to allocate guest mem via memfd

Extract the code for allocating guest memory via memfd out of
vm_userspace_mem_region_add() and into a new helper, kvm_memfd_alloc().
A future selftest to populate a guest with the maximum amount of guest
memory will abuse KVM's memslots to alias guest memory regions to a
single memfd-backed host region, i.e. needs to back a guest with memfd
memory without a 1:1 association between a memslot and a memfd instance.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-27-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Move raw KVM_SET_USER_MEMORY_REGION helper to utils
Sean Christopherson [Sat, 26 Feb 2022 00:15:43 +0000 (00:15 +0000)]
KVM: selftests: Move raw KVM_SET_USER_MEMORY_REGION helper to utils

Move set_memory_region_test's KVM_SET_USER_MEMORY_REGION helper to KVM's
utils so that it can be used by other tests.  Provide a raw version as
well as an assert-success version to reduce the amount of boilerplate
code need for basic usage.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-26-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: WARN on any attempt to atomically update REMOVED SPTE
Sean Christopherson [Sat, 26 Feb 2022 00:15:42 +0000 (00:15 +0000)]
KVM: x86/mmu: WARN on any attempt to atomically update REMOVED SPTE

Disallow calling tdp_mmu_set_spte_atomic() with a REMOVED "old" SPTE.
This solves a conundrum introduced by commit 3255530ab191 ("KVM: x86/mmu:
Automatically update iter->old_spte if cmpxchg fails"); if the helper
doesn't update old_spte in the REMOVED case, then theoretically the
caller could get stuck in an infinite loop as it will fail indefinitely
on the REMOVED SPTE.  E.g. until recently, clear_dirty_gfn_range() didn't
check for a present SPTE and would have spun until getting rescheduled.

In practice, only the page fault path should "create" a new SPTE, all
other paths should only operate on existing, a.k.a. shadow present,
SPTEs.  Now that the page fault path pre-checks for a REMOVED SPTE in all
cases, require all other paths to indirectly pre-check by verifying the
target SPTE is a shadow-present SPTE.

Note, this does not guarantee the actual SPTE isn't REMOVED, nor is that
scenario disallowed.  The invariant is only that the caller mustn't
invoke tdp_mmu_set_spte_atomic() if the SPTE was REMOVED when last
observed by the caller.

Cc: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-25-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Check for a REMOVED leaf SPTE before making the SPTE
Sean Christopherson [Sat, 26 Feb 2022 00:15:41 +0000 (00:15 +0000)]
KVM: x86/mmu: Check for a REMOVED leaf SPTE before making the SPTE

Explicitly check for a REMOVED leaf SPTE prior to attempting to map
the final SPTE when handling a TDP MMU fault.  Functionally, this is a
nop as tdp_mmu_set_spte_atomic() will eventually detect the frozen SPTE.
Pre-checking for a REMOVED SPTE is a minor optmization, but the real goal
is to allow tdp_mmu_set_spte_atomic() to have an invariant that the "old"
SPTE is never a REMOVED SPTE.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-24-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Zap defunct roots via asynchronous worker
Paolo Bonzini [Fri, 4 Mar 2022 16:43:13 +0000 (11:43 -0500)]
KVM: x86/mmu: Zap defunct roots via asynchronous worker

Zap defunct roots, a.k.a. roots that have been invalidated after their
last reference was initially dropped, asynchronously via the existing work
queue instead of forcing the work upon the unfortunate task that happened
to drop the last reference.

If a vCPU task drops the last reference, the vCPU is effectively blocked
by the host for the entire duration of the zap.  If the root being zapped
happens be fully populated with 4kb leaf SPTEs, e.g. due to dirty logging
being active, the zap can take several hundred seconds.  Unsurprisingly,
most guests are unhappy if a vCPU disappears for hundreds of seconds.

E.g. running a synthetic selftest that triggers a vCPU root zap with
~64tb of guest memory and 4kb SPTEs blocks the vCPU for 900+ seconds.
Offloading the zap to a worker drops the block time to <100ms.

There is an important nuance to this change.  If the same work item
was queued twice before the work function has run, it would only
execute once and one reference would be leaked.  Therefore, now that
queueing and flushing items is not anymore protected by kvm->slots_lock,
kvm_tdp_mmu_invalidate_all_roots() has to check root->role.invalid and
skip already invalid roots.  On the other hand, kvm_mmu_zap_all_fast()
must return only after those skipped roots have been zapped as well.
These two requirements can be satisfied only if _all_ places that
change invalid to true now schedule the worker before releasing the
mmu_lock.  There are just two, kvm_tdp_mmu_put_root() and
kvm_tdp_mmu_invalidate_all_roots().

Co-developed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-23-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Zap roots in two passes to avoid inducing RCU stalls
Sean Christopherson [Sat, 26 Feb 2022 00:15:39 +0000 (00:15 +0000)]
KVM: x86/mmu: Zap roots in two passes to avoid inducing RCU stalls

When zapping a TDP MMU root, perform the zap in two passes to avoid
zapping an entire top-level SPTE while holding RCU, which can induce RCU
stalls.  In the first pass, zap SPTEs at PG_LEVEL_1G, and then
zap top-level entries in the second pass.

With 4-level paging, zapping a PGD that is fully populated with 4kb leaf
SPTEs take up to ~7 or so seconds (time varies based on kernel config,
number of (v)CPUs, etc...).  With 5-level paging, that time can balloon
well into hundreds of seconds.

Before remote TLB flushes were omitted, the problem was even worse as
waiting for all active vCPUs to respond to the IPI introduced significant
overhead for VMs with large numbers of vCPUs.

By zapping 1gb SPTEs (both shadow pages and hugepages) in the first pass,
the amount of work that is done without dropping RCU protection is
strictly bounded, with the worst case latency for a single operation
being less than 100ms.

Zapping at 1gb in the first pass is not arbitrary.  First and foremost,
KVM relies on being able to zap 1gb shadow pages in a single shot when
when repacing a shadow page with a hugepage.  Zapping a 1gb shadow page
that is fully populated with 4kb dirty SPTEs also triggers the worst case
latency due writing back the struct page accessed/dirty bits for each 4kb
page, i.e. the two-pass approach is guaranteed to work so long as KVM can
cleany zap a 1gb shadow page.

  rcu: INFO: rcu_sched self-detected stall on CPU
  rcu:     52-....: (20999 ticks this GP) idle=7be/1/0x4000000000000000
                                          softirq=15759/15759 fqs=5058
   (t=21016 jiffies g=66453 q=238577)
  NMI backtrace for cpu 52
  Call Trace:
   ...
   mark_page_accessed+0x266/0x2f0
   kvm_set_pfn_accessed+0x31/0x40
   handle_removed_tdp_mmu_page+0x259/0x2e0
   __handle_changed_spte+0x223/0x2c0
   handle_removed_tdp_mmu_page+0x1c1/0x2e0
   __handle_changed_spte+0x223/0x2c0
   handle_removed_tdp_mmu_page+0x1c1/0x2e0
   __handle_changed_spte+0x223/0x2c0
   zap_gfn_range+0x141/0x3b0
   kvm_tdp_mmu_zap_invalidated_roots+0xc8/0x130
   kvm_mmu_zap_all_fast+0x121/0x190
   kvm_mmu_invalidate_zap_pages_in_memslot+0xe/0x10
   kvm_page_track_flush_slot+0x5c/0x80
   kvm_arch_flush_shadow_memslot+0xe/0x10
   kvm_set_memslot+0x172/0x4e0
   __kvm_set_memory_region+0x337/0x590
   kvm_vm_ioctl+0x49c/0xf80

Reported-by: David Matlack <dmatlack@google.com>
Cc: Ben Gardon <bgardon@google.com>
Cc: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-22-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Allow yielding when zapping GFNs for defunct TDP MMU root
Paolo Bonzini [Thu, 3 Mar 2022 06:50:21 +0000 (01:50 -0500)]
KVM: x86/mmu: Allow yielding when zapping GFNs for defunct TDP MMU root

Allow yielding when zapping SPTEs after the last reference to a valid
root is put.  Because KVM must drop all SPTEs in response to relevant
mmu_notifier events, mark defunct roots invalid and reset their refcount
prior to zapping the root.  Keeping the refcount elevated while the zap
is in-progress ensures the root is reachable via mmu_notifier until the
zap completes and the last reference to the invalid, defunct root is put.

Allowing kvm_tdp_mmu_put_root() to yield fixes soft lockup issues if the
root in being put has a massive paging structure, e.g. zapping a root
that is backed entirely by 4kb pages for a guest with 32tb of memory can
take hundreds of seconds to complete.

  watchdog: BUG: soft lockup - CPU#49 stuck for 485s! [max_guest_memor:52368]
  RIP: 0010:kvm_set_pfn_dirty+0x30/0x50 [kvm]
   __handle_changed_spte+0x1b2/0x2f0 [kvm]
   handle_removed_tdp_mmu_page+0x1a7/0x2b8 [kvm]
   __handle_changed_spte+0x1f4/0x2f0 [kvm]
   handle_removed_tdp_mmu_page+0x1a7/0x2b8 [kvm]
   __handle_changed_spte+0x1f4/0x2f0 [kvm]
   tdp_mmu_zap_root+0x307/0x4d0 [kvm]
   kvm_tdp_mmu_put_root+0x7c/0xc0 [kvm]
   kvm_mmu_free_roots+0x22d/0x350 [kvm]
   kvm_mmu_reset_context+0x20/0x60 [kvm]
   kvm_arch_vcpu_ioctl_set_sregs+0x5a/0xc0 [kvm]
   kvm_vcpu_ioctl+0x5bd/0x710 [kvm]
   __se_sys_ioctl+0x77/0xc0
   __x64_sys_ioctl+0x1d/0x20
   do_syscall_64+0x44/0xa0
   entry_SYSCALL_64_after_hwframe+0x44/0xae

KVM currently doesn't put a root from a non-preemptible context, so other
than the mmu_notifier wrinkle, yielding when putting a root is safe.

Yield-unfriendly iteration uses for_each_tdp_mmu_root(), which doesn't
take a reference to each root (it requires mmu_lock be held for the
entire duration of the walk).

tdp_mmu_next_root() is used only by the yield-friendly iterator.

tdp_mmu_zap_root_work() is explicitly yield friendly.

kvm_mmu_free_roots() => mmu_free_root_page() is a much bigger fan-out,
but is still yield-friendly in all call sites, as all callers can be
traced back to some combination of vcpu_run(), kvm_destroy_vm(), and/or
kvm_create_vm().

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-21-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Zap invalidated roots via asynchronous worker
Paolo Bonzini [Wed, 2 Mar 2022 17:02:07 +0000 (12:02 -0500)]
KVM: x86/mmu: Zap invalidated roots via asynchronous worker

Use the system worker threads to zap the roots invalidated
by the TDP MMU's "fast zap" mechanism, implemented by
kvm_tdp_mmu_invalidate_all_roots().

At this point, apart from allowing some parallelism in the zapping of
roots, the workqueue is a glorified linked list: work items are added and
flushed entirely within a single kvm->slots_lock critical section.  However,
the workqueue fixes a latent issue where kvm_mmu_zap_all_invalidated_roots()
assumes that it owns a reference to all invalid roots; therefore, no
one can set the invalid bit outside kvm_mmu_zap_all_fast().  Putting the
invalidated roots on a linked list... erm, on a workqueue ensures that
tdp_mmu_zap_root_work() only puts back those extra references that
kvm_mmu_zap_all_invalidated_roots() had gifted to it.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Defer TLB flush to caller when freeing TDP MMU shadow pages
Sean Christopherson [Sat, 26 Feb 2022 00:15:37 +0000 (00:15 +0000)]
KVM: x86/mmu: Defer TLB flush to caller when freeing TDP MMU shadow pages

Defer TLB flushes to the caller when freeing TDP MMU shadow pages instead
of immediately flushing.  Because the shadow pages are freed in an RCU
callback, so long as at least one CPU holds RCU, all CPUs are protected.
For vCPUs running in the guest, i.e. consuming TLB entries, KVM only
needs to ensure the caller services the pending TLB flush before dropping
its RCU protections.  I.e. use the caller's RCU as a proxy for all vCPUs
running in the guest.

Deferring the flushes allows batching flushes, e.g. when installing a
1gb hugepage and zapping a pile of SPs.  And when zapping an entire root,
deferring flushes allows skipping the flush entirely (because flushes are
not needed in that case).

Avoiding flushes when zapping an entire root is especially important as
synchronizing with other CPUs via IPI after zapping every shadow page can
cause significant performance issues for large VMs.  The issue is
exacerbated by KVM zapping entire top-level entries without dropping
RCU protection, which can lead to RCU stalls even when zapping roots
backing relatively "small" amounts of guest memory, e.g. 2tb.  Removing
the IPI bottleneck largely mitigates the RCU issues, though it's likely
still a problem for 5-level paging.  A future patch will further address
the problem by zapping roots in multiple passes to avoid holding RCU for
an extended duration.

Reviewed-by: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-20-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Do remote TLB flush before dropping RCU in TDP MMU resched
Sean Christopherson [Sat, 26 Feb 2022 00:15:36 +0000 (00:15 +0000)]
KVM: x86/mmu: Do remote TLB flush before dropping RCU in TDP MMU resched

When yielding in the TDP MMU iterator, service any pending TLB flush
before dropping RCU protections in anticipation of using the caller's RCU
"lock" as a proxy for vCPUs in the guest.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-19-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Zap only TDP MMU leafs in kvm_zap_gfn_range()
Sean Christopherson [Sat, 26 Feb 2022 00:15:35 +0000 (00:15 +0000)]
KVM: x86/mmu: Zap only TDP MMU leafs in kvm_zap_gfn_range()

Zap only leaf SPTEs in the TDP MMU's zap_gfn_range(), and rename various
functions accordingly.  When removing mappings for functional correctness
(except for the stupid VFIO GPU passthrough memslots bug), zapping the
leaf SPTEs is sufficient as the paging structures themselves do not point
at guest memory and do not directly impact the final translation (in the
TDP MMU).

Note, this aligns the TDP MMU with the legacy/full MMU, which zaps only
the rmaps, a.k.a. leaf SPTEs, in kvm_zap_gfn_range() and
kvm_unmap_gfn_range().

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-18-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Require mmu_lock be held for write to zap TDP MMU range
Sean Christopherson [Sat, 26 Feb 2022 00:15:34 +0000 (00:15 +0000)]
KVM: x86/mmu: Require mmu_lock be held for write to zap TDP MMU range

Now that all callers of zap_gfn_range() hold mmu_lock for write, drop
support for zapping with mmu_lock held for read.  That all callers hold
mmu_lock for write isn't a random coincidence; now that the paths that
need to zap _everything_ have their own path, the only callers left are
those that need to zap for functional correctness.  And when zapping is
required for functional correctness, mmu_lock must be held for write,
otherwise the caller has no guarantees about the state of the TDP MMU
page tables after it has run, e.g. the SPTE(s) it zapped can be
immediately replaced by a vCPU faulting in a page.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-17-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Add dedicated helper to zap TDP MMU root shadow page
Sean Christopherson [Sat, 26 Feb 2022 00:15:33 +0000 (00:15 +0000)]
KVM: x86/mmu: Add dedicated helper to zap TDP MMU root shadow page

Add a dedicated helper for zapping a TDP MMU root, and use it in the three
flows that do "zap_all" and intentionally do not do a TLB flush if SPTEs
are zapped (zapping an entire root is safe if and only if it cannot be in
use by any vCPU).  Because a TLB flush is never required, unconditionally
pass "false" to tdp_mmu_iter_cond_resched() when potentially yielding.

Opportunistically document why KVM must not yield when zapping roots that
are being zapped by kvm_tdp_mmu_put_root(), i.e. roots whose refcount has
reached zero, and further harden the flow to detect improper KVM behavior
with respect to roots that are supposed to be unreachable.

In addition to hardening zapping of roots, isolating zapping of roots
will allow future simplification of zap_gfn_range() by having it zap only
leaf SPTEs, and by removing its tricky "zap all" heuristic.  By having
all paths that truly need to free _all_ SPs flow through the dedicated
root zapper, the generic zapper can be freed of those concerns.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-16-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Skip remote TLB flush when zapping all of TDP MMU
Sean Christopherson [Sat, 26 Feb 2022 00:15:32 +0000 (00:15 +0000)]
KVM: x86/mmu: Skip remote TLB flush when zapping all of TDP MMU

Don't flush the TLBs when zapping all TDP MMU pages, as the only time KVM
uses the slow version of "zap everything" is when the VM is being
destroyed or the owning mm has exited.  In either case, KVM_RUN is
unreachable for the VM, i.e. the guest TLB entries cannot be consumed.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-15-seanjc@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Zap only the target TDP MMU shadow page in NX recovery
Sean Christopherson [Sat, 26 Feb 2022 00:15:31 +0000 (00:15 +0000)]
KVM: x86/mmu: Zap only the target TDP MMU shadow page in NX recovery

When recovering a potential hugepage that was shattered for the iTLB
multihit workaround, precisely zap only the target page instead of
iterating over the TDP MMU to find the SP that was passed in.  This will
allow future simplification of zap_gfn_range() by having it zap only
leaf SPTEs.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-14-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Refactor low-level TDP MMU set SPTE helper to take raw values
Sean Christopherson [Sat, 26 Feb 2022 00:15:30 +0000 (00:15 +0000)]
KVM: x86/mmu: Refactor low-level TDP MMU set SPTE helper to take raw values

Refactor __tdp_mmu_set_spte() to work with raw values instead of a
tdp_iter objects so that a future patch can modify SPTEs without doing a
walk, and without having to synthesize a tdp_iter.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-13-seanjc@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: WARN if old _or_ new SPTE is REMOVED in non-atomic path
Sean Christopherson [Sat, 26 Feb 2022 00:15:29 +0000 (00:15 +0000)]
KVM: x86/mmu: WARN if old _or_ new SPTE is REMOVED in non-atomic path

WARN if the new_spte being set by __tdp_mmu_set_spte() is a REMOVED_SPTE,
which is called out by the comment as being disallowed but not actually
checked.  Keep the WARN on the old_spte as well, because overwriting a
REMOVED_SPTE in the non-atomic path is also disallowed (as evidence by
lack of splats with the existing WARN).

Fixes: 08f07c800e9d ("KVM: x86/mmu: Flush TLBs after zap in TDP MMU PF handler")
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-12-seanjc@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Add helpers to read/write TDP MMU SPTEs and document RCU
Sean Christopherson [Sat, 26 Feb 2022 00:15:28 +0000 (00:15 +0000)]
KVM: x86/mmu: Add helpers to read/write TDP MMU SPTEs and document RCU

Add helpers to read and write TDP MMU SPTEs instead of open coding
rcu_dereference() all over the place, and to provide a convenient
location to document why KVM doesn't exempt holding mmu_lock for write
from having to hold RCU (and any future changes to the rules).

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-11-seanjc@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Drop RCU after processing each root in MMU notifier hooks
Sean Christopherson [Sat, 26 Feb 2022 00:15:27 +0000 (00:15 +0000)]
KVM: x86/mmu: Drop RCU after processing each root in MMU notifier hooks

Drop RCU protection after processing each root when handling MMU notifier
hooks that aren't the "unmap" path, i.e. aren't zapping.  Temporarily
drop RCU to let RCU do its thing between roots, and to make it clear that
there's no special behavior that relies on holding RCU across all roots.

Currently, the RCU protection is completely superficial, it's necessary
only to make rcu_dereference() of SPTE pointers happy.  A future patch
will rely on holding RCU as a proxy for vCPUs in the guest, e.g. to
ensure shadow pages aren't freed before all vCPUs do a TLB flush (or
rather, acknowledge the need for a flush), but in that case RCU needs to
be held until the flush is complete if and only if the flush is needed
because a shadow page may have been removed.  And except for the "unmap"
path, MMU notifier events cannot remove SPs (don't toggle PRESENT bit,
and can't change the PFN for a SP).

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-10-seanjc@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Batch TLB flushes from TDP MMU for MMU notifier change_spte
Sean Christopherson [Sat, 26 Feb 2022 00:15:26 +0000 (00:15 +0000)]
KVM: x86/mmu: Batch TLB flushes from TDP MMU for MMU notifier change_spte

Batch TLB flushes (with other MMUs) when handling ->change_spte()
notifications in the TDP MMU.  The MMU notifier path in question doesn't
allow yielding and correcty flushes before dropping mmu_lock.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-9-seanjc@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Check for !leaf=>leaf, not PFN change, in TDP MMU SP removal
Sean Christopherson [Sat, 26 Feb 2022 00:15:25 +0000 (00:15 +0000)]
KVM: x86/mmu: Check for !leaf=>leaf, not PFN change, in TDP MMU SP removal

Look for a !leaf=>leaf conversion instead of a PFN change when checking
if a SPTE change removed a TDP MMU shadow page.  Convert the PFN check
into a WARN, as KVM should never change the PFN of a shadow page (except
when its being zapped or replaced).

From a purely theoretical perspective, it's not illegal to replace a SP
with a hugepage pointing at the same PFN.  In practice, it's impossible
as that would require mapping guest memory overtop a kernel-allocated SP.
Either way, the check is odd.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-8-seanjc@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: do not allow readers to acquire references to invalid roots
Paolo Bonzini [Wed, 2 Mar 2022 13:51:05 +0000 (08:51 -0500)]
KVM: x86/mmu: do not allow readers to acquire references to invalid roots

Remove the "shared" argument of for_each_tdp_mmu_root_yield_safe, thus ensuring
that readers do not ever acquire a reference to an invalid root.  After this
patch, all readers except kvm_tdp_mmu_zap_invalidated_roots() treat
refcount=0/valid, refcount=0/invalid and refcount=1/invalid in exactly the
same way.  kvm_tdp_mmu_zap_invalidated_roots() is different but it also
does not acquire a reference to the invalid root, and it cannot see
refcount=0/invalid because it is guaranteed to run after
kvm_tdp_mmu_invalidate_all_roots().

Opportunistically add a lockdep assertion to the yield-safe iterator.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: only perform eager page splitting on valid roots
Paolo Bonzini [Wed, 2 Mar 2022 13:44:22 +0000 (08:44 -0500)]
KVM: x86/mmu: only perform eager page splitting on valid roots

Eager page splitting is an optimization; it does not have to be performed on
invalid roots.  It is also the only case in which a reader might acquire
a reference to an invalid root, so after this change we know that readers
will skip both dying and invalid roots.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Require mmu_lock be held for write in unyielding root iter
Sean Christopherson [Sat, 26 Feb 2022 00:15:24 +0000 (00:15 +0000)]
KVM: x86/mmu: Require mmu_lock be held for write in unyielding root iter

Assert that mmu_lock is held for write by users of the yield-unfriendly
TDP iterator.  The nature of a shared walk means that the caller needs to
play nice with other tasks modifying the page tables, which is more or
less the same thing as playing nice with yielding.  Theoretically, KVM
could gain a flow where it could legitimately take mmu_lock for read in
a non-preemptible context, but that's highly unlikely and any such case
should be viewed with a fair amount of scrutiny.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Document that zapping invalidated roots doesn't need to flush
Sean Christopherson [Sat, 26 Feb 2022 00:15:23 +0000 (00:15 +0000)]
KVM: x86/mmu: Document that zapping invalidated roots doesn't need to flush

Remove the misleading flush "handling" when zapping invalidated TDP MMU
roots, and document that flushing is unnecessary for all flavors of MMUs
when zapping invalid/obsolete roots/pages.  The "handling" in the TDP MMU
is dead code, as zap_gfn_range() is called with shared=true, in which
case it will never return true due to the flushing being handled by
tdp_mmu_zap_spte_atomic().

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Formalize TDP MMU's (unintended?) deferred TLB flush logic
Sean Christopherson [Sat, 26 Feb 2022 00:15:22 +0000 (00:15 +0000)]
KVM: x86/mmu: Formalize TDP MMU's (unintended?) deferred TLB flush logic

Explicitly ignore the result of zap_gfn_range() when putting the last
reference to a TDP MMU root, and add a pile of comments to formalize the
TDP MMU's behavior of deferring TLB flushes to alloc/reuse.  Note, this
only affects the !shared case, as zap_gfn_range() subtly never returns
true for "flush" as the flush is handled by tdp_mmu_zap_spte_atomic().

Putting the root without a flush is ok because even if there are stale
references to the root in the TLB, they are unreachable because KVM will
not run the guest with the same ASID without first flushing (where ASID
in this context refers to both SVM's explicit ASID and Intel's implicit
ASID that is constructed from VPID+PCID+EPT4A+etc...).

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-5-seanjc@google.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Fix wrong/misleading comments in TDP MMU fast zap
Sean Christopherson [Sat, 26 Feb 2022 00:15:21 +0000 (00:15 +0000)]
KVM: x86/mmu: Fix wrong/misleading comments in TDP MMU fast zap

Fix misleading and arguably wrong comments in the TDP MMU's fast zap
flow.  The comments, and the fact that actually zapping invalid roots was
added separately, strongly suggests that zapping invalid roots is an
optimization and not required for correctness.  That is a lie.

KVM _must_ zap invalid roots before returning from kvm_mmu_zap_all_fast(),
because when it's called from kvm_mmu_invalidate_zap_pages_in_memslot(),
KVM is relying on it to fully remove all references to the memslot.  Once
the memslot is gone, KVM's mmu_notifier hooks will be unable to find the
stale references as the hva=>gfn translation is done via the memslots.
If KVM doesn't immediately zap SPTEs and userspace unmaps a range after
deleting a memslot, KVM will fail to zap in response to the mmu_notifier
due to not finding a memslot corresponding to the notifier's range, which
leads to a variation of use-after-free.

The other misleading comment (and code) explicitly states that roots
without a reference should be skipped.  While that's technically true,
it's also extremely misleading as it should be impossible for KVM to
encounter a defunct root on the list while holding mmu_lock for write.
Opportunistically add a WARN to enforce that invariant.

Fixes: b7cccd397f31 ("KVM: x86/mmu: Fast invalidation for TDP MMU")
Fixes: 4c6654bd160d ("KVM: x86/mmu: Tear down roots before kvm_mmu_zap_all_fast returns")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Check for present SPTE when clearing dirty bit in TDP MMU
Sean Christopherson [Sat, 26 Feb 2022 00:15:20 +0000 (00:15 +0000)]
KVM: x86/mmu: Check for present SPTE when clearing dirty bit in TDP MMU

Explicitly check for present SPTEs when clearing dirty bits in the TDP
MMU.  This isn't strictly required for correctness, as setting the dirty
bit in a defunct SPTE will not change the SPTE from !PRESENT to PRESENT.
However, the guarded MMU_WARN_ON() in spte_ad_need_write_protect() would
complain if anyone actually turned on KVM's MMU debugging.

Fixes: a6a0b05da9f3 ("kvm: x86/mmu: Support dirty logging for the TDP MMU")
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-3-seanjc@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: use __vcalloc for very large allocations
Paolo Bonzini [Tue, 8 Mar 2022 09:49:37 +0000 (04:49 -0500)]
KVM: use __vcalloc for very large allocations

Allocations whose size is related to the memslot size can be arbitrarily
large.  Do not use kvzalloc/kvcalloc, as those are limited to "not crazy"
sizes that fit in 32 bits.

Cc: stable@vger.kernel.org
Fixes: 7661809d493b ("mm: don't allow oversized kvmalloc() calls")
Reviewed-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agomm: use vmalloc_array and vcalloc for array allocations
Paolo Bonzini [Tue, 8 Mar 2022 10:02:21 +0000 (05:02 -0500)]
mm: use vmalloc_array and vcalloc for array allocations

Instead of using array_size or just a multiply, use a function that
takes care of both the multiplication and the overflow checks.

Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agomm: vmalloc: introduce array allocation functions
Paolo Bonzini [Tue, 8 Mar 2022 09:47:22 +0000 (04:47 -0500)]
mm: vmalloc: introduce array allocation functions

Linux has dozens of occurrences of vmalloc(array_size()) and
vzalloc(array_size()).  Allow to simplify the code by providing
vmalloc_array and vcalloc, as well as the underscored variants that let
the caller specify the GFP flags.

Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoMerge branch 'kvm-bugfixes' into HEAD
Paolo Bonzini [Fri, 4 Mar 2022 23:39:29 +0000 (18:39 -0500)]
Merge branch 'kvm-bugfixes' into HEAD

Merge bugfixes from 5.17 before merging more tricky work.

2 years agoKVM: x86: pull kvm->srcu read-side to kvm_arch_vcpu_ioctl_run
Paolo Bonzini [Sat, 19 Feb 2022 09:28:20 +0000 (04:28 -0500)]
KVM: x86: pull kvm->srcu read-side to kvm_arch_vcpu_ioctl_run

kvm_arch_vcpu_ioctl_run is already doing srcu_read_lock/unlock in two
places, namely vcpu_run and post_kvm_run_save, and a third is actually
needed around the call to vcpu->arch.complete_userspace_io to avoid
the following splat:

  WARNING: suspicious RCU usage
  arch/x86/kvm/pmu.c:190 suspicious rcu_dereference_check() usage!
  other info that might help us debug this:
  rcu_scheduler_active = 2, debug_locks = 1
  1 lock held by CPU 28/KVM/370841:
  #0: ff11004089f280b8 (&vcpu->mutex){+.+.}-{3:3}, at: kvm_vcpu_ioctl+0x87/0x730 [kvm]
  Call Trace:
   <TASK>
   dump_stack_lvl+0x59/0x73
   reprogram_fixed_counter+0x15d/0x1a0 [kvm]
   kvm_pmu_trigger_event+0x1a3/0x260 [kvm]
   ? free_moved_vector+0x1b4/0x1e0
   complete_fast_pio_in+0x8a/0xd0 [kvm]

This splat is not at all unexpected, since complete_userspace_io callbacks
can execute similar code to vmexits.  For example, SVM with nrips=false
will call into the emulator from svm_skip_emulated_instruction().

While it's tempting to never acquire kvm->srcu for an uninitialized vCPU,
practically speaking there's no penalty to acquiring kvm->srcu "early"
as the KVM_MP_STATE_UNINITIALIZED path is a one-time thing per vCPU.  On
the other hand, seemingly innocuous helpers like kvm_apic_accept_events()
and sync_regs() can theoretically reach code that might access
SRCU-protected data structures, e.g. sync_regs() can trigger forced
existing of nested mode via kvm_vcpu_ioctl_x86_set_vcpu_events().

Reported-by: Like Xu <likexu@tencent.com>
Co-developed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Passing up the error state of mmu_alloc_shadow_roots()
Like Xu [Tue, 1 Mar 2022 12:49:41 +0000 (20:49 +0800)]
KVM: x86/mmu: Passing up the error state of mmu_alloc_shadow_roots()

Just like on the optional mmu_alloc_direct_roots() path, once shadow
path reaches "r = -EIO" somewhere, the caller needs to know the actual
state in order to enter error handling and avoid something worse.

Fixes: 4a38162ee9f1 ("KVM: MMU: load PDPTRs outside mmu_lock")
Signed-off-by: Like Xu <likexu@tencent.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220301124941.48412-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Disable preemption across AVIC load/put during APICv refresh
Sean Christopherson [Tue, 1 Mar 2022 17:05:09 +0000 (09:05 -0800)]
KVM: SVM: Disable preemption across AVIC load/put during APICv refresh

Disable preemption when loading/putting the AVIC during an APICv refresh.
If the vCPU task is preempted and migrated ot a different pCPU, the
unprotected avic_vcpu_load() could set the wrong pCPU in the physical ID
cache/table.

Pull the necessary code out of avic_vcpu_{,un}blocking() and into a new
helper to reduce the probability of introducing this exact bug a third
time.

Fixes: df7e4827c549 ("KVM: SVM: call avic_vcpu_load/avic_vcpu_put when enabling/disabling AVIC")
Cc: stable@vger.kernel.org
Reported-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Exit to userspace on ENOMEM/EFAULT GHCB errors
Sean Christopherson [Fri, 25 Feb 2022 20:52:09 +0000 (20:52 +0000)]
KVM: SVM: Exit to userspace on ENOMEM/EFAULT GHCB errors

Exit to userspace if setup_vmgexit_scratch() fails due to OOM or because
copying data from guest (userspace) memory failed/faulted.  The OOM
scenario is clearcut, it's userspace's decision as to whether it should
terminate the guest, free memory, etc...

As for -EFAULT, arguably, any guest issue is a violation of the guest's
contract with userspace, and thus userspace needs to decide how to
proceed.  E.g. userspace defines what is RAM vs. MMIO and communicates
that directly to the guest, KVM is not involved in deciding what is/isn't
RAM nor in communicating that information to the guest.  If the scratch
GPA doesn't resolve to a memslot, then the guest is not honoring the
memory configuration as defined by userspace.

And if userspace unmaps an hva for whatever reason, then exiting to
userspace with -EFAULT is absolutely the right thing to do.  KVM's ABI
currently sucks and doesn't provide enough information to act on the
-EFAULT, but that will hopefully be remedied in the future as there are
multiple use cases, e.g. uffd and virtiofs truncation, that shouldn't
require any work in KVM beyond returning -EFAULT with a small amount of
metadata.

KVM could define its ABI such that failure to access the scratch area is
reflected into the guest, i.e. establish a contract with userspace, but
that's undesirable as it limits KVM's options in the future, e.g. in the
potential uffd case any failure on a uaccess needs to kick out to
userspace.  KVM does have several cases where it reflects these errors
into the guest, e.g. kvm_pv_clock_pairing() and Hyper-V emulation, but
KVM would preferably "fix" those instead of propagating the falsehood
that any memory failure is the guest's fault.

Lastly, returning a boolean as an "error" for that a helper that isn't
named accordingly never works out well.

Fixes: ad5b353240c8 ("KVM: SVM: Do not terminate SEV-ES guests on GHCB validation failure")
Cc: Alper Gun <alpergun@google.com>
Cc: Peter Gonda <pgonda@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220225205209.3881130-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: WARN if is_unsync_root() is called on a root without a shadow page
Sean Christopherson [Fri, 25 Feb 2022 18:22:48 +0000 (18:22 +0000)]
KVM: WARN if is_unsync_root() is called on a root without a shadow page

WARN and bail if is_unsync_root() is passed a root for which there is no
shadow page, i.e. is passed the physical address of one of the special
roots, which do not have an associated shadow page.  The current usage
squeaks by without bug reports because neither kvm_mmu_sync_roots() nor
kvm_mmu_sync_prev_roots() calls the helper with pae_root or pml4_root,
and 5-level AMD CPUs are not generally available, i.e. no one can coerce
KVM into calling is_unsync_root() on pml5_root.

Note, this doesn't fix the mess with 5-level nNPT, it just (hopefully)
prevents KVM from crashing.

Cc: Lai Jiangshan <jiangshanlai@gmail.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220225182248.3812651-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Drop KVM_REQ_MMU_RELOAD and update vcpu-requests.rst documentation
Sean Christopherson [Fri, 25 Feb 2022 18:22:47 +0000 (18:22 +0000)]
KVM: Drop KVM_REQ_MMU_RELOAD and update vcpu-requests.rst documentation

Remove the now unused KVM_REQ_MMU_RELOAD, shift KVM_REQ_VM_DEAD into the
unoccupied space, and update vcpu-requests.rst, which was missing an
entry for KVM_REQ_VM_DEAD.  Switching KVM_REQ_VM_DEAD to entry '1' also
fixes the stale comment about bits 4-7 being reserved.

Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220225182248.3812651-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: s390: Replace KVM_REQ_MMU_RELOAD usage with arch specific request
Sean Christopherson [Fri, 25 Feb 2022 18:22:46 +0000 (18:22 +0000)]
KVM: s390: Replace KVM_REQ_MMU_RELOAD usage with arch specific request

Add an arch request, KVM_REQ_REFRESH_GUEST_PREFIX, to deal with guest
prefix changes instead of piggybacking KVM_REQ_MMU_RELOAD.  This will
allow for the removal of the generic KVM_REQ_MMU_RELOAD, which isn't
actually used by generic KVM.

No functional change intended.

Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: Janosch Frank <frankja@linux.ibm.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220225182248.3812651-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Zap only obsolete roots if a root shadow page is zapped
Sean Christopherson [Fri, 25 Feb 2022 18:22:45 +0000 (18:22 +0000)]
KVM: x86/mmu: Zap only obsolete roots if a root shadow page is zapped

Zap only obsolete roots when responding to zapping a single root shadow
page.  Because KVM keeps root_count elevated when stuffing a previous
root into its PGD cache, shadowing a 64-bit guest means that zapping any
root causes all vCPUs to reload all roots, even if their current root is
not affected by the zap.

For many kernels, zapping a single root is a frequent operation, e.g. in
Linux it happens whenever an mm is dropped, e.g. process exits, etc...

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220225182248.3812651-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Drop kvm_reload_remote_mmus(), open code request in x86 users
Sean Christopherson [Fri, 25 Feb 2022 18:22:44 +0000 (18:22 +0000)]
KVM: Drop kvm_reload_remote_mmus(), open code request in x86 users

Remove the generic kvm_reload_remote_mmus() and open code its
functionality into the two x86 callers.  x86 is (obviously) the only
architecture that uses the hook, and is also the only architecture that
uses KVM_REQ_MMU_RELOAD in a way that's consistent with the name.  That
will change in a future patch, as x86's usage when zapping a single
shadow page x86 doesn't actually _need_ to reload all vCPUs' MMUs, only
MMUs whose root is being zapped actually need to be reloaded.

s390 also uses KVM_REQ_MMU_RELOAD, but for a slightly different purpose.

Drop the generic code in anticipation of implementing s390 and x86 arch
specific requests, which will allow dropping KVM_REQ_MMU_RELOAD entirely.

Opportunistically reword the x86 TDP MMU comment to avoid making
references to functions (and requests!) when possible, and to remove the
rather ambiguous "this".

No functional change intended.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220225182248.3812651-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Invoke kvm_mmu_unload() directly on CR4.PCIDE change
Sean Christopherson [Fri, 25 Feb 2022 18:22:43 +0000 (18:22 +0000)]
KVM: x86: Invoke kvm_mmu_unload() directly on CR4.PCIDE change

Replace a KVM_REQ_MMU_RELOAD request with a direct kvm_mmu_unload() call
when the guest's CR4.PCIDE changes.  This will allow tweaking the logic
of KVM_REQ_MMU_RELOAD to free only obsolete/invalid roots, which is the
historical intent of KVM_REQ_MMU_RELOAD.  The recent PCIDE behavior is
the only user of KVM_REQ_MMU_RELOAD that doesn't mark affected roots as
obsolete, needs to unconditionally unload the entire MMU, _and_ affects
only the current vCPU.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220225182248.3812651-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/emulator: Move the unhandled outer privilege level logic of far return into...
Hou Wenlong [Tue, 8 Feb 2022 09:34:05 +0000 (17:34 +0800)]
KVM: x86/emulator: Move the unhandled outer privilege level logic of far return into __load_segment_descriptor()

Outer-privilege level return is not implemented in emulator,
move the unhandled logic into __load_segment_descriptor to
make it easier to understand why the checks for RET are
incomplete.

Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Hou Wenlong <houwenlong.hwl@antgroup.com>
Message-Id: <5b7188e6388ac9f4567d14eab32db9adf3e00119.1644292363.git.houwenlong.hwl@antgroup.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/emulator: Fix wrong privilege check for code segment in __load_segment_descr...
Hou Wenlong [Tue, 8 Feb 2022 09:34:04 +0000 (17:34 +0800)]
KVM: x86/emulator: Fix wrong privilege check for code segment in __load_segment_descriptor()

Code segment descriptor can be loaded by jmp/call/ret, iret
and int. The privilege checks are different between those
instructions above realmode. Although, the emulator has
use x86_transfer_type enumerate to differentiate them, but
it is not really used in __load_segment_descriptor(). Note,
far jump/call to call gate, task gate or task state segment
are not implemented in emulator.

As for far jump/call to code segment, if DPL > CPL for conforming
code or (RPL > CPL or DPL != CPL) for non-conforming code, it
should trigger #GP. The current checks are ok.

As for far return, if RPL < CPL or DPL > RPL for conforming
code or DPL != RPL for non-conforming code, it should trigger #GP.
Outer level return is not implemented above virtual-8086 mode in
emulator. So it implies that RPL <= CPL, but the current checks
wouldn't trigger #GP if RPL < CPL.

As for code segment loading in task switch, if DPL > RPL for conforming
code or DPL != RPL for non-conforming code, it should trigger #TS. Since
segment selector is loaded before segment descriptor when load state from
tss, it implies that RPL = CPL, so the current checks are ok.

The only problem in current implementation is missing RPL < CPL check for
far return. However, change code to follow the manual is better.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Hou Wenlong <houwenlong.hwl@antgroup.com>
Message-Id: <e01f5ea70fc1f18f23da1182acdbc5c97c0e5886.1644292363.git.houwenlong.hwl@antgroup.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/emulator: Defer not-present segment check in __load_segment_descriptor()
Hou Wenlong [Tue, 8 Feb 2022 09:34:03 +0000 (17:34 +0800)]
KVM: x86/emulator: Defer not-present segment check in __load_segment_descriptor()

Per Intel's SDM on the "Instruction Set Reference", when
loading segment descriptor, not-present segment check should
be after all type and privilege checks. But the emulator checks
it first, then #NP is triggered instead of #GP if privilege fails
and segment is not present. Put not-present segment check after
type and privilege checks in __load_segment_descriptor().

Fixes: 38ba30ba51a00 (KVM: x86 emulator: Emulate task switch in emulator.c)
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Hou Wenlong <houwenlong.hwl@antgroup.com>
Message-Id: <52573c01d369f506cadcf7233812427cf7db81a7.1644292363.git.houwenlong.hwl@antgroup.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Add test to verify KVM handling of ICR
Sean Christopherson [Fri, 4 Feb 2022 21:42:05 +0000 (21:42 +0000)]
KVM: selftests: Add test to verify KVM handling of ICR

The main thing that the selftest verifies is that KVM copies x2APIC's
ICR[63:32] to/from ICR2 when userspace accesses the vAPIC page via
KVM_{G,S}ET_LAPIC.  KVM previously split x2APIC ICR to ICR+ICR2 at the
time of write (from the guest), and so KVM must preserve that behavior
for backwards compatibility between different versions of KVM.

It will also test other invariants, e.g. that KVM clears the BUSY
flag on ICR writes, that the reserved bits in ICR2 are dropped on writes
from the guest, etc...

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-12-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Make kvm_lapic_set_reg() a "private" xAPIC helper
Sean Christopherson [Fri, 4 Feb 2022 21:42:04 +0000 (21:42 +0000)]
KVM: x86: Make kvm_lapic_set_reg() a "private" xAPIC helper

Hide the lapic's "raw" write helper inside lapic.c to force non-APIC code
to go through proper helpers when modification the vAPIC state.  Keep the
read helper visible to outsiders for now, refactoring KVM to hide it too
is possible, it will just take more work to do so.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-11-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Treat x2APIC's ICR as a 64-bit register, not two 32-bit regs
Sean Christopherson [Fri, 4 Feb 2022 21:42:03 +0000 (21:42 +0000)]
KVM: x86: Treat x2APIC's ICR as a 64-bit register, not two 32-bit regs

Emulate the x2APIC ICR as a single 64-bit register, as opposed to forking
it across ICR and ICR2 as two 32-bit registers.  This mirrors hardware
behavior for Intel's upcoming IPI virtualization support, which does not
split the access.

Previous versions of Intel's SDM and AMD's APM don't explicitly state
exactly how ICR is reflected in the vAPIC page for x2APIC, KVM just
happened to speculate incorrectly.

Handling the upcoming behavior is necessary in order to maintain
backwards compatibility with KVM_{G,S}ET_LAPIC, e.g. failure to shuffle
the 64-bit ICR to ICR+ICR2 and vice versa would break live migration if
IPI virtualization support isn't symmetrical across the source and dest.

Cc: Zeng Guang <guang.zeng@intel.com>
Cc: Chao Gao <chao.gao@intel.com>
Cc: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Add helpers to handle 64-bit APIC MSR read/writes
Sean Christopherson [Fri, 4 Feb 2022 21:42:02 +0000 (21:42 +0000)]
KVM: x86: Add helpers to handle 64-bit APIC MSR read/writes

Add helpers to handle 64-bit APIC read/writes via MSRs to deduplicate the
x2APIC and Hyper-V code needed to service reads/writes to ICR.  Future
support for IPI virtualization will add yet another path where KVM must
handle 64-bit APIC MSR reads/write (to ICR).

Opportunistically fix the comment in the write path; ICR2 holds the
destination (if there's no shorthand), not the vector.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Make kvm_lapic_reg_{read,write}() static
Sean Christopherson [Fri, 4 Feb 2022 21:42:01 +0000 (21:42 +0000)]
KVM: x86: Make kvm_lapic_reg_{read,write}() static

Make the low level read/write lapic helpers static, any accesses to the
local APIC from vendor code or non-APIC code should be routed through
proper helpers.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: WARN if KVM emulates an IPI without clearing the BUSY flag
Sean Christopherson [Fri, 4 Feb 2022 21:42:00 +0000 (21:42 +0000)]
KVM: x86: WARN if KVM emulates an IPI without clearing the BUSY flag

WARN if KVM emulates an IPI without clearing the BUSY flag, failure to do
so could hang the guest if it waits for the IPI be sent.

Opportunistically use APIC_ICR_BUSY macro instead of open coding the
magic number, and add a comment to clarify why kvm_recalculate_apic_map()
is unconditionally invoked (it's really, really confusing for IPIs due to
the existence of fast paths that don't trigger a potential recalc).

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Don't rewrite guest ICR on AVIC IPI virtualization failure
Sean Christopherson [Fri, 4 Feb 2022 21:41:59 +0000 (21:41 +0000)]
KVM: SVM: Don't rewrite guest ICR on AVIC IPI virtualization failure

Don't bother rewriting the ICR value into the vAPIC page on an AVIC IPI
virtualization failure, the access is a trap, i.e. the value has already
been written to the vAPIC page.  The one caveat is if hardware left the
BUSY flag set (which appears to happen somewhat arbitrarily), in which
case go through the "nodecode" APIC-write path in order to clear the BUSY
flag.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: SVM: Use common kvm_apic_write_nodecode() for AVIC write traps
Sean Christopherson [Fri, 4 Feb 2022 21:41:58 +0000 (21:41 +0000)]
KVM: SVM: Use common kvm_apic_write_nodecode() for AVIC write traps

Use the common kvm_apic_write_nodecode() to handle AVIC/APIC-write traps
instead of open coding the same exact code.  This will allow making the
low level lapic helpers inaccessible outside of lapic.c code.

Opportunistically clean up the params to eliminate a bunch of svm=>vcpu
reflection.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Use "raw" APIC register read for handling APIC-write VM-Exit
Sean Christopherson [Fri, 4 Feb 2022 21:41:57 +0000 (21:41 +0000)]
KVM: x86: Use "raw" APIC register read for handling APIC-write VM-Exit

Use the "raw" helper to read the vAPIC register after an APIC-write trap
VM-Exit.  Hardware is responsible for vetting the write, and the caller
is responsible for sanitizing the offset.  This is a functional change,
as it means KVM will consume whatever happens to be in the vAPIC page if
the write was dropped by hardware.  But, unless userspace deliberately
wrote garbage into the vAPIC page via KVM_SET_LAPIC, the value should be
zero since it's not writable by the guest.

This aligns common x86 with SVM's AVIC logic, i.e. paves the way for
using the nodecode path to handle APIC-write traps when AVIC is enabled.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Handle APIC-write offset wrangling in VMX code
Sean Christopherson [Fri, 4 Feb 2022 21:41:56 +0000 (21:41 +0000)]
KVM: VMX: Handle APIC-write offset wrangling in VMX code

Move the vAPIC offset adjustments done in the APIC-write trap path from
common x86 to VMX in anticipation of using the nodecode path for SVM's
AVIC.  The adjustment reflects hardware behavior, i.e. it's technically a
property of VMX, no common x86.  SVM's AVIC behavior is identical, so
it's a bit of a moot point, the goal is purely to make it easier to
understand why the adjustment is ok.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220204214205.3306634-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Do not change ICR on write to APIC_SELF_IPI
Paolo Bonzini [Thu, 24 Feb 2022 14:53:36 +0000 (09:53 -0500)]
KVM: x86: Do not change ICR on write to APIC_SELF_IPI

Emulating writes to SELF_IPI with a write to ICR has an unwanted side effect:
the value of ICR in vAPIC page gets changed.  The lists SELF_IPI as write-only,
with no associated MMIO offset, so any write should have no visible side
effect in the vAPIC page.

Reported-by: Chao Gao <chao.gao@intel.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Fix emulation in writing cr8
Zhenzhong Duan [Thu, 10 Feb 2022 09:45:06 +0000 (17:45 +0800)]
KVM: x86: Fix emulation in writing cr8

In emulation of writing to cr8, one of the lowest four bits in TPR[3:0]
is kept.

According to Intel SDM 10.8.6.1(baremetal scenario):
"APIC.TPR[bits 7:4] = CR8[bits 3:0], APIC.TPR[bits 3:0] = 0";

and SDM 28.3(use TPR shadow):
"MOV to CR8. The instruction stores bits 3:0 of its source operand into
bits 7:4 of VTPR; the remainder of VTPR (bits 3:0 and bits 31:8) are
cleared.";

and AMD's APM 16.6.4:
"Task Priority Sub-class (TPS)-Bits 3 : 0. The TPS field indicates the
current sub-priority to be used when arbitrating lowest-priority messages.
This field is written with zero when TPR is written using the architectural
CR8 register.";

so in KVM emulated scenario, clear TPR[3:0] to make a consistent behavior
as in other scenarios.

This doesn't impact evaluation and delivery of pending virtual interrupts
because processor does not use the processor-priority sub-class to
determine which interrupts to delivery and which to inhibit.

Sub-class is used by hardware to arbitrate lowest priority interrupts,
but KVM just does a round-robin style delivery.

Fixes: b93463aa59d6 ("KVM: Accelerated apic support")
Signed-off-by: Zhenzhong Duan <zhenzhong.duan@intel.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220210094506.20181-1-zhenzhong.duan@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: flush TLB separately from MMU reset
Paolo Bonzini [Mon, 14 Feb 2022 16:50:36 +0000 (11:50 -0500)]
KVM: x86: flush TLB separately from MMU reset

For both CR0 and CR4, disassociate the TLB flush logic from the
MMU role logic.  Instead  of relying on kvm_mmu_reset_context() being
a superset of various TLB flushes (which is not necessarily going to
be the case in the future), always call it if the role changes
but also set the various TLB flush requests according to what is
in the manual.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Yield to IPI target vCPU only if it is busy
Li RongQing [Wed, 9 Feb 2022 04:16:41 +0000 (12:16 +0800)]
KVM: x86: Yield to IPI target vCPU only if it is busy

When sending a call-function IPI-many to vCPUs, yield to the
IPI target vCPU which is marked as preempted.

but when emulating HLT, an idling vCPU will be voluntarily
scheduled out and mark as preempted from the guest kernel
perspective. yielding to idle vCPU is pointless and increase
unnecessary vmexit, maybe miss the true preempted vCPU

so yield to IPI target vCPU only if vCPU is busy and preempted

Signed-off-by: Li RongQing <lirongqing@baidu.com>
Message-Id: <1644380201-29423-1-git-send-email-lirongqing@baidu.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agox86/kvmclock: Fix Hyper-V Isolated VM's boot issue when vCPUs > 64
Dexuan Cui [Fri, 25 Feb 2022 08:46:00 +0000 (00:46 -0800)]
x86/kvmclock: Fix Hyper-V Isolated VM's boot issue when vCPUs > 64

When Linux runs as an Isolated VM on Hyper-V, it supports AMD SEV-SNP
but it's partially enlightened, i.e. cc_platform_has(
CC_ATTR_GUEST_MEM_ENCRYPT) is true but sev_active() is false.

Commit 4d96f9109109 per se is good, but with it now
kvm_setup_vsyscall_timeinfo() -> kvmclock_init_mem() calls
set_memory_decrypted(), and later gets stuck when trying to zere out
the pages pointed by 'hvclock_mem', if Linux runs as an Isolated VM on
Hyper-V. The cause is that here now the Linux VM should no longer access
the original guest physical addrss (GPA); instead the VM should do
memremap() and access the original GPA + ms_hyperv.shared_gpa_boundary:
see the example code in drivers/hv/connection.c: vmbus_connect() or
drivers/hv/ring_buffer.c: hv_ringbuffer_init(). If the VM tries to
access the original GPA, it keepts getting injected a fault by Hyper-V
and gets stuck there.

Here the issue happens only when the VM has >=65 vCPUs, because the
global static array hv_clock_boot[] can hold 64 "struct
pvclock_vsyscall_time_info" (the sizeof of the struct is 64 bytes), so
kvmclock_init_mem() only allocates memory in the case of vCPUs > 64.

Since the 'hvclock_mem' pages are only useful when the kvm clock is
supported by the underlying hypervisor, fix the issue by returning
early when Linux VM runs on Hyper-V, which doesn't support kvm clock.

Fixes: 4d96f9109109 ("x86/sev: Replace occurrences of sev_active() with cc_platform_has()")
Tested-by: Andrea Parri (Microsoft) <parri.andrea@gmail.com>
Signed-off-by: Andrea Parri (Microsoft) <parri.andrea@gmail.com>
Signed-off-by: Dexuan Cui <decui@microsoft.com>
Message-Id: <20220225084600.17817-1-decui@microsoft.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agox86/kvm: Don't waste memory if kvmclock is disabled
Wanpeng Li [Tue, 22 Feb 2022 09:02:03 +0000 (01:02 -0800)]
x86/kvm: Don't waste memory if kvmclock is disabled

Even if "no-kvmclock" is passed in cmdline parameter, the guest kernel
still allocates hvclock_mem which is scaled by the number of vCPUs,
let's check kvmclock enable in advance to avoid this memory waste.

Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Message-Id: <1645520523-30814-1-git-send-email-wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agox86/kvm: Don't use PV TLB/yield when mwait is advertised
Wanpeng Li [Fri, 25 Feb 2022 08:29:40 +0000 (00:29 -0800)]
x86/kvm: Don't use PV TLB/yield when mwait is advertised

MWAIT is advertised in host is not overcommitted scenario, however, PV
TLB/sched yield should be enabled in host overcommitted scenario. Let's
add the MWAIT checking when enabling PV TLB/sched yield.

Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Message-Id: <1645777780-2581-1-git-send-email-wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoMerge tag 'kvmarm-fixes-5.17-4' of git://git.kernel.org/pub/scm/linux/kernel/git...
Paolo Bonzini [Fri, 25 Feb 2022 14:49:30 +0000 (09:49 -0500)]
Merge tag 'kvmarm-fixes-5.17-4' of git://git./linux/kernel/git/kvmarm/kvmarm into HEAD

KVM/arm64 fixes for 5.17, take #4

- Correctly synchronise PMR and co on PSCI CPU_SUSPEND

- Skip tests that depend on GICv3 when the HW isn't available

2 years agoKVM: x86/mmu: clear MMIO cache when unloading the MMU
Paolo Bonzini [Mon, 14 Feb 2022 14:13:48 +0000 (09:13 -0500)]
KVM: x86/mmu: clear MMIO cache when unloading the MMU

For cleanliness, do not leave a stale GVA in the cache after all the roots are
cleared.  In practice, kvm_mmu_load will go through kvm_mmu_sync_roots if
paging is on, and will not use vcpu_match_mmio_gva at all if paging is off.
However, leaving data in the cache might cause bugs in the future.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: Always use current mmu's role when loading new PGD
Paolo Bonzini [Mon, 22 Nov 2021 18:18:23 +0000 (13:18 -0500)]
KVM: x86/mmu: Always use current mmu's role when loading new PGD

Since the guest PGD is now loaded after the MMU has been set up
completely, the desired role for a cache hit is simply the current
mmu_role.  There is no need to compute it again, so __kvm_mmu_new_pgd
can be folded in kvm_mmu_new_pgd.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: load new PGD after the shadow MMU is initialized
Paolo Bonzini [Fri, 4 Feb 2022 09:12:31 +0000 (04:12 -0500)]
KVM: x86/mmu: load new PGD after the shadow MMU is initialized

Now that __kvm_mmu_new_pgd does not look at the MMU's root_level and
shadow_root_level anymore, pull the PGD load after the initialization of
the shadow MMUs.

Besides being more intuitive, this enables future simplifications
and optimizations because it's not necessary anymore to compute the
role outside kvm_init_mmu.  In particular, kvm_mmu_reset_context was not
attempting to use a cached PGD to avoid having to figure out the new role.
With this change, it could follow what nested_{vmx,svm}_load_cr3 are doing,
and avoid unloading all the cached roots.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: look for a cached PGD when going from 32-bit to 64-bit
Paolo Bonzini [Wed, 9 Feb 2022 07:49:47 +0000 (02:49 -0500)]
KVM: x86/mmu: look for a cached PGD when going from 32-bit to 64-bit

Right now, PGD caching avoids placing a PAE root in the cache by using the
old value of mmu->root_level and mmu->shadow_root_level; it does not look
for a cached PGD if the old root is a PAE one, and then frees it using
kvm_mmu_free_roots.

Change the logic instead to free the uncacheable root early.
This way, __kvm_new_mmu_pgd is able to look up the cache when going from
32-bit to 64-bit (if there is a hit, the invalid root becomes the least
recently used).  An example of this is nested virtualization with shadow
paging, when a 64-bit L1 runs a 32-bit L2.

As a side effect (which is actually the reason why this patch was
written), PGD caching does not use the old value of mmu->root_level
and mmu->shadow_root_level anymore.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: do not pass vcpu to root freeing functions
Paolo Bonzini [Mon, 21 Feb 2022 14:31:51 +0000 (09:31 -0500)]
KVM: x86/mmu: do not pass vcpu to root freeing functions

These functions only operate on a given MMU, of which there is more
than one in a vCPU (we care about two, because the third does not have
any roots and is only used to walk guest page tables).  They do need a
struct kvm in order to lock the mmu_lock, but they do not needed anything
else in the struct kvm_vcpu.  So, pass the vcpu->kvm directly to them.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: do not consult levels when freeing roots
Paolo Bonzini [Tue, 8 Feb 2022 22:53:55 +0000 (17:53 -0500)]
KVM: x86/mmu: do not consult levels when freeing roots

Right now, PGD caching requires a complicated dance of first computing
the MMU role and passing it to __kvm_mmu_new_pgd(), and then separately calling
kvm_init_mmu().

Part of this is due to kvm_mmu_free_roots using mmu->root_level and
mmu->shadow_root_level to distinguish whether the page table uses a single
root or 4 PAE roots.  Because kvm_init_mmu() can overwrite mmu->root_level,
kvm_mmu_free_roots() must be called before kvm_init_mmu().

However, even after kvm_init_mmu() there is a way to detect whether the
page table may hold PAE roots, as root.hpa isn't backed by a shadow when
it points at PAE roots.  Using this method results in simpler code, and
is one less obstacle in moving all calls to __kvm_mmu_new_pgd() after the
MMU has been initialized.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: use struct kvm_mmu_root_info for mmu->root
Paolo Bonzini [Mon, 21 Feb 2022 14:28:33 +0000 (09:28 -0500)]
KVM: x86: use struct kvm_mmu_root_info for mmu->root

The root_hpa and root_pgd fields form essentially a struct kvm_mmu_root_info.
Use the struct to have more consistency between mmu->root and
mmu->prev_roots.

The patch is entirely search and replace except for cached_root_available,
which does not need a temporary struct kvm_mmu_root_info anymore.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: avoid NULL-pointer dereference on page freeing bugs
Paolo Bonzini [Wed, 9 Feb 2022 00:08:33 +0000 (19:08 -0500)]
KVM: x86/mmu: avoid NULL-pointer dereference on page freeing bugs

WARN and bail if KVM attempts to free a root that isn't backed by a shadow
page.  KVM allocates a bare page for "special" roots, e.g. when using PAE
paging or shadowing 2/3/4-level page tables with 4/5-level, and so root_hpa
will be valid but won't be backed by a shadow page.  It's all too easy to
blindly call mmu_free_root_page() on root_hpa, be nice and WARN instead of
crashing KVM and possibly the kernel.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: do not deliver asynchronous page faults if CR0.PG=0
Paolo Bonzini [Wed, 9 Feb 2022 10:17:38 +0000 (05:17 -0500)]
KVM: x86: do not deliver asynchronous page faults if CR0.PG=0

Enabling async page faults is nonsensical if paging is disabled, but
it is allowed because CR0.PG=0 does not clear the async page fault
MSR.  Just ignore them and only use the artificial halt state,
similar to what happens in guest mode if async #PF vmexits are disabled.

Given the increasingly complex logic, and the nicer code if the new
"if" is placed last, opportunistically change the "||" into a chain
of "if (...) return false" statements.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Reinitialize context if host userspace toggles EFER.LME
Paolo Bonzini [Wed, 9 Feb 2022 09:56:05 +0000 (04:56 -0500)]
KVM: x86: Reinitialize context if host userspace toggles EFER.LME

While the guest runs, EFER.LME cannot change unless CR0.PG is clear, and
therefore EFER.NX is the only bit that can affect the MMU role.  However,
set_efer accepts a host-initiated change to EFER.LME even with CR0.PG=1.
In that case, the MMU has to be reset.

Fixes: 11988499e62b ("KVM: x86: Skip EFER vs. guest CPUID checks for host-initiated writes")
Cc: stable@vger.kernel.org
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Verify disabling PMU virtualization via KVM_CAP_CONFIG_PMU
David Dunn [Wed, 23 Feb 2022 22:57:43 +0000 (22:57 +0000)]
KVM: selftests: Verify disabling PMU virtualization via KVM_CAP_CONFIG_PMU

On a VM with PMU disabled via KVM_CAP_PMU_CONFIG, the PMU should not be
usable by the guest.

Signed-off-by: David Dunn <daviddunn@google.com>
Message-Id: <20220223225743.2703915-4-daviddunn@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: selftests: Carve out helper to create "default" VM without vCPUs
David Dunn [Wed, 23 Feb 2022 22:57:42 +0000 (22:57 +0000)]
KVM: selftests: Carve out helper to create "default" VM without vCPUs

Carve out portion of vm_create_default so that selftests can modify
a "default" VM prior to creating vcpus.

Signed-off-by: David Dunn <daviddunn@google.com>
Message-Id: <20220223225743.2703915-3-daviddunn@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Provide per VM capability for disabling PMU virtualization
David Dunn [Wed, 23 Feb 2022 22:57:41 +0000 (22:57 +0000)]
KVM: x86: Provide per VM capability for disabling PMU virtualization

Add a new capability, KVM_CAP_PMU_CAPABILITY, that takes a bitmask of
settings/features to allow userspace to configure PMU virtualization on
a per-VM basis.  For now, support a single flag, KVM_PMU_CAP_DISABLE,
to allow disabling PMU virtualization for a VM even when KVM is configured
with enable_pmu=true a module level.

To keep KVM simple, disallow changing VM's PMU configuration after vCPUs
have been created.

Signed-off-by: David Dunn <daviddunn@google.com>
Message-Id: <20220223225743.2703915-2-daviddunn@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: Fix pointer mistmatch warning when patching RET0 static calls
Sean Christopherson [Wed, 23 Feb 2022 16:23:55 +0000 (16:23 +0000)]
KVM: x86: Fix pointer mistmatch warning when patching RET0 static calls

Cast kvm_x86_ops.func to 'void *' when updating KVM static calls that are
conditionally patched to __static_call_return0().  clang complains about
using mismatching pointers in the ternary operator, which breaks the
build when compiling with CONFIG_KVM_WERROR=y.

  >> arch/x86/include/asm/kvm-x86-ops.h:82:1: warning: pointer type mismatch
  ('bool (*)(struct kvm_vcpu *)' and 'void *') [-Wpointer-type-mismatch]

Fixes: 5be2226f417d ("KVM: x86: allow defining return-0 static calls")
Reported-by: Like Xu <like.xu.linux@gmail.com>
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: David Dunn <daviddunn@google.com>
Reviewed-by: Nathan Chancellor <nathan@kernel.org>
Tested-by: Nathan Chancellor <nathan@kernel.org>
Message-Id: <20220223162355.3174907-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: Move VM's worker kthreads back to the original cgroup before exiting.
Vipin Sharma [Tue, 22 Feb 2022 05:48:48 +0000 (05:48 +0000)]
KVM: Move VM's worker kthreads back to the original cgroup before exiting.

VM worker kthreads can linger in the VM process's cgroup for sometime
after KVM terminates the VM process.

KVM terminates the worker kthreads by calling kthread_stop() which waits
on the 'exited' completion, triggered by exit_mm(), via mm_release(), in
do_exit() during the kthread's exit.  However, these kthreads are
removed from the cgroup using the cgroup_exit() which happens after the
exit_mm(). Therefore, A VM process can terminate in between the
exit_mm() and cgroup_exit() calls, leaving only worker kthreads in the
cgroup.

Moving worker kthreads back to the original cgroup (kthreadd_task's
cgroup) makes sure that the cgroup is empty as soon as the main VM
process is terminated.

Signed-off-by: Vipin Sharma <vipinsh@google.com>
Suggested-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220222054848.563321-1-vipinsh@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: VMX: Remove scratch 'cpu' variable that shadows an identical scratch var
Peng Hao [Tue, 22 Feb 2022 10:39:54 +0000 (18:39 +0800)]
KVM: VMX: Remove scratch 'cpu' variable that shadows an identical scratch var

 From: Peng Hao <flyingpeng@tencent.com>

 Remove a redundant 'cpu' declaration from inside an if-statement that
 that shadows an identical declaration at function scope.  Both variables
 are used as scratch variables in for_each_*_cpu() loops, thus there's no
 harm in sharing a variable.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Peng Hao <flyingpeng@tencent.com>
Message-Id: <20220222103954.70062-1-flyingpeng@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agokvm: vmx: Fix typos comment in __loaded_vmcs_clear()
Peng Hao [Tue, 22 Feb 2022 10:40:29 +0000 (18:40 +0800)]
kvm: vmx: Fix typos comment in __loaded_vmcs_clear()

Fix a comment documenting the memory barrier related to clearing a
loaded_vmcs; loaded_vmcs tracks the host CPU the VMCS is loaded on via
the field 'cpu', it doesn't have a 'vcpu' field.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Peng Hao <flyingpeng@tencent.com>
Message-Id: <20220222104029.70129-1-flyingpeng@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: nVMX: Make setup/unsetup under the same conditions
Peng Hao [Tue, 22 Feb 2022 10:40:54 +0000 (18:40 +0800)]
KVM: nVMX: Make setup/unsetup under the same conditions

Make sure nested_vmx_hardware_setup/unsetup() are called in pairs under
the same conditions.  Calling nested_vmx_hardware_unsetup() when nested
is false "works" right now because it only calls free_page() on zero-
initialized pointers, but it's possible that more code will be added to
nested_vmx_hardware_unsetup() in the future.

Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Peng Hao <flyingpeng@tencent.com>
Message-Id: <20220222104054.70286-1-flyingpeng@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoMerge branch 'kvm-hv-xmm-hypercall-fixes' into HEAD
Paolo Bonzini [Fri, 25 Feb 2022 11:28:10 +0000 (06:28 -0500)]
Merge branch 'kvm-hv-xmm-hypercall-fixes' into HEAD

The fixes for 5.17 conflict with cleanups made in the same area
earlier in the 5.18 development cycle.

2 years agoKVM: selftests: aarch64: Skip tests if we can't create a vgic-v3
Mark Brown [Wed, 23 Feb 2022 13:16:24 +0000 (13:16 +0000)]
KVM: selftests: aarch64: Skip tests if we can't create a vgic-v3

The arch_timer and vgic_irq kselftests assume that they can create a
vgic-v3, using the library function vgic_v3_setup() which aborts with a
test failure if it is not possible to do so. Since vgic-v3 can only be
instantiated on systems where the host has GICv3 this leads to false
positives on older systems where that is not the case.

Fix this by changing vgic_v3_setup() to return an error if the vgic can't
be instantiated and have the callers skip if this happens. We could also
exit flagging a skip in vgic_v3_setup() but this would prevent future test
cases conditionally deciding which GIC to use or generally doing more
complex output.

Signed-off-by: Mark Brown <broonie@kernel.org>
Reviewed-by: Andrew Jones <drjones@redhat.com>
Tested-by: Ricardo Koller <ricarkol@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20220223131624.1830351-1-broonie@kernel.org
2 years agoKVM: x86: hyper-v: HVCALL_SEND_IPI_EX is an XMM fast hypercall
Vitaly Kuznetsov [Tue, 22 Feb 2022 15:46:42 +0000 (16:46 +0100)]
KVM: x86: hyper-v: HVCALL_SEND_IPI_EX is an XMM fast hypercall

It has been proven on practice that at least Windows Server 2019 tries
using HVCALL_SEND_IPI_EX in 'XMM fast' mode when it has more than 64 vCPUs
and it needs to send an IPI to a vCPU > 63. Similarly to other XMM Fast
hypercalls (HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE}{,_EX}), this
information is missing in TLFS as of 6.0b. Currently, KVM returns an error
(HV_STATUS_INVALID_HYPERCALL_INPUT) and Windows crashes.

Note, HVCALL_SEND_IPI is a 'standard' fast hypercall (not 'XMM fast') as
all its parameters fit into RDX:R8 and this is handled by KVM correctly.

Cc: stable@vger.kernel.org # 5.14.x: 3244867af8c0: KVM: x86: Ignore sparse banks size for an "all CPUs", non-sparse IPI req
Cc: stable@vger.kernel.org # 5.14.x
Fixes: d8f5537a8816 ("KVM: hyper-v: Advertise support for fast XMM hypercalls")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220222154642.684285-5-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: hyper-v: Fix the maximum number of sparse banks for XMM fast TLB flush...
Vitaly Kuznetsov [Tue, 22 Feb 2022 15:46:41 +0000 (16:46 +0100)]
KVM: x86: hyper-v: Fix the maximum number of sparse banks for XMM fast TLB flush hypercalls

When TLB flush hypercalls (HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE}_EX are
issued in 'XMM fast' mode, the maximum number of allowed sparse_banks is
not 'HV_HYPERCALL_MAX_XMM_REGISTERS - 1' (5) but twice as many (10) as each
XMM register is 128 bit long and can hold two 64 bit long banks.

Cc: stable@vger.kernel.org # 5.14.x
Fixes: 5974565bc26d ("KVM: x86: kvm_hv_flush_tlb use inputs from XMM registers")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220222154642.684285-4-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: hyper-v: Drop redundant 'ex' parameter from kvm_hv_flush_tlb()
Vitaly Kuznetsov [Tue, 22 Feb 2022 15:46:40 +0000 (16:46 +0100)]
KVM: x86: hyper-v: Drop redundant 'ex' parameter from kvm_hv_flush_tlb()

'struct kvm_hv_hcall' has all the required information already,
there's no need to pass 'ex' additionally.

No functional change intended.

Cc: stable@vger.kernel.org # 5.14.x
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220222154642.684285-3-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: hyper-v: Drop redundant 'ex' parameter from kvm_hv_send_ipi()
Vitaly Kuznetsov [Tue, 22 Feb 2022 15:46:39 +0000 (16:46 +0100)]
KVM: x86: hyper-v: Drop redundant 'ex' parameter from kvm_hv_send_ipi()

'struct kvm_hv_hcall' has all the required information already,
there's no need to pass 'ex' additionally.

No functional change intended.

Cc: stable@vger.kernel.org # 5.14.x
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220222154642.684285-2-vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoRevert "KVM: VMX: Save HOST_CR3 in vmx_prepare_switch_to_guest()"
Sean Christopherson [Thu, 24 Feb 2022 19:19:17 +0000 (19:19 +0000)]
Revert "KVM: VMX: Save HOST_CR3 in vmx_prepare_switch_to_guest()"

Revert back to refreshing vmcs.HOST_CR3 immediately prior to VM-Enter.
The PCID (ASID) part of CR3 can be bumped without KVM being scheduled
out, as the kernel will switch CR3 during __text_poke(), e.g. in response
to a static key toggling.  If switch_mm_irqs_off() chooses a new ASID for
the mm associate with KVM, KVM will do VM-Enter => VM-Exit with a stale
vmcs.HOST_CR3.

Add a comment to explain why KVM must wait until VM-Enter is imminent to
refresh vmcs.HOST_CR3.

The following splat was captured by stashing vmcs.HOST_CR3 in kvm_vcpu
and adding a WARN in load_new_mm_cr3() to fire if a new ASID is being
loaded for the KVM-associated mm while KVM has a "running" vCPU:

  static void load_new_mm_cr3(pgd_t *pgdir, u16 new_asid, bool need_flush)
  {
struct kvm_vcpu *vcpu = kvm_get_running_vcpu();

...

WARN(vcpu && (vcpu->cr3 & GENMASK(11, 0)) != (new_mm_cr3 & GENMASK(11, 0)) &&
     (vcpu->cr3 & PHYSICAL_PAGE_MASK) == (new_mm_cr3 & PHYSICAL_PAGE_MASK),
     "KVM is hosed, loading CR3 = %lx, vmcs.HOST_CR3 = %lx", new_mm_cr3, vcpu->cr3);
  }

  ------------[ cut here ]------------
  KVM is hosed, loading CR3 = 8000000105393004, vmcs.HOST_CR3 = 105393003
  WARNING: CPU: 4 PID: 20717 at arch/x86/mm/tlb.c:291 load_new_mm_cr3+0x82/0xe0
  Modules linked in: vhost_net vhost vhost_iotlb tap kvm_intel
  CPU: 4 PID: 20717 Comm: stable Tainted: G        W         5.17.0-rc3+ #747
  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
  RIP: 0010:load_new_mm_cr3+0x82/0xe0
  RSP: 0018:ffffc9000489fa98 EFLAGS: 00010082
  RAX: 0000000000000000 RBX: 8000000105393004 RCX: 0000000000000027
  RDX: 0000000000000027 RSI: 00000000ffffdfff RDI: ffff888277d1b788
  RBP: 0000000000000004 R08: ffff888277d1b780 R09: ffffc9000489f8b8
  R10: 0000000000000001 R11: 0000000000000001 R12: 0000000000000000
  R13: ffff88810678a800 R14: 0000000000000004 R15: 0000000000000c33
  FS:  00007fa9f0e72700(0000) GS:ffff888277d00000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 0000000000000000 CR3: 00000001001b5003 CR4: 0000000000172ea0
  Call Trace:
   <TASK>
   switch_mm_irqs_off+0x1cb/0x460
   __text_poke+0x308/0x3e0
   text_poke_bp_batch+0x168/0x220
   text_poke_finish+0x1b/0x30
   arch_jump_label_transform_apply+0x18/0x30
   static_key_slow_inc_cpuslocked+0x7c/0x90
   static_key_slow_inc+0x16/0x20
   kvm_lapic_set_base+0x116/0x190
   kvm_set_apic_base+0xa5/0xe0
   kvm_set_msr_common+0x2f4/0xf60
   vmx_set_msr+0x355/0xe70 [kvm_intel]
   kvm_set_msr_ignored_check+0x91/0x230
   kvm_emulate_wrmsr+0x36/0x120
   vmx_handle_exit+0x609/0x6c0 [kvm_intel]
   kvm_arch_vcpu_ioctl_run+0x146f/0x1b80
   kvm_vcpu_ioctl+0x279/0x690
   __x64_sys_ioctl+0x83/0xb0
   do_syscall_64+0x3b/0xc0
   entry_SYSCALL_64_after_hwframe+0x44/0xae
   </TASK>
  ---[ end trace 0000000000000000 ]---

This reverts commit 15ad9762d69fd8e40a4a51828c1d6b0c1b8fbea0.

Fixes: 15ad9762d69f ("KVM: VMX: Save HOST_CR3 in vmx_prepare_switch_to_guest()")
Reported-by: Wanpeng Li <kernellwp@gmail.com>
Cc: Lai Jiangshan <laijs@linux.alibaba.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Acked-by: Lai Jiangshan <jiangshanlai@gmail.com>
Message-Id: <20220224191917.3508476-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoRevert "KVM: VMX: Save HOST_CR3 in vmx_set_host_fs_gs()"
Sean Christopherson [Thu, 24 Feb 2022 19:19:16 +0000 (19:19 +0000)]
Revert "KVM: VMX: Save HOST_CR3 in vmx_set_host_fs_gs()"

Undo a nested VMX fix as a step toward reverting the commit it fixed,
15ad9762d69f ("KVM: VMX: Save HOST_CR3 in vmx_prepare_switch_to_guest()"),
as the underlying premise that "host CR3 in the vcpu thread can only be
changed when scheduling" is wrong.

This reverts commit a9f2705ec84449e3b8d70c804766f8e97e23080d.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220224191917.3508476-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86: nSVM: disallow userspace setting of MSR_AMD64_TSC_RATIO to non default...
Maxim Levitsky [Wed, 23 Feb 2022 11:56:49 +0000 (13:56 +0200)]
KVM: x86: nSVM: disallow userspace setting of MSR_AMD64_TSC_RATIO to non default value when tsc scaling disabled

If nested tsc scaling is disabled, MSR_AMD64_TSC_RATIO should
never have non default value.

Due to way nested tsc scaling support was implmented in qemu,
it would set this msr to 0 when nested tsc scaling was disabled.
Ignore that value for now, as it causes no harm.

Fixes: 5228eb96a487 ("KVM: x86: nSVM: implement nested TSC scaling")
Cc: stable@vger.kernel.org
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20220223115649.319134-1-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: x86/mmu: make apf token non-zero to fix bug
Liang Zhang [Tue, 22 Feb 2022 03:12:39 +0000 (11:12 +0800)]
KVM: x86/mmu: make apf token non-zero to fix bug

In current async pagefault logic, when a page is ready, KVM relies on
kvm_arch_can_dequeue_async_page_present() to determine whether to deliver
a READY event to the Guest. This function test token value of struct
kvm_vcpu_pv_apf_data, which must be reset to zero by Guest kernel when a
READY event is finished by Guest. If value is zero meaning that a READY
event is done, so the KVM can deliver another.
But the kvm_arch_setup_async_pf() may produce a valid token with zero
value, which is confused with previous mention and may lead the loss of
this READY event.

This bug may cause task blocked forever in Guest:
 INFO: task stress:7532 blocked for more than 1254 seconds.
       Not tainted 5.10.0 #16
 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
 task:stress          state:D stack:    0 pid: 7532 ppid:  1409
 flags:0x00000080
 Call Trace:
  __schedule+0x1e7/0x650
  schedule+0x46/0xb0
  kvm_async_pf_task_wait_schedule+0xad/0xe0
  ? exit_to_user_mode_prepare+0x60/0x70
  __kvm_handle_async_pf+0x4f/0xb0
  ? asm_exc_page_fault+0x8/0x30
  exc_page_fault+0x6f/0x110
  ? asm_exc_page_fault+0x8/0x30
  asm_exc_page_fault+0x1e/0x30
 RIP: 0033:0x402d00
 RSP: 002b:00007ffd31912500 EFLAGS: 00010206
 RAX: 0000000000071000 RBX: ffffffffffffffff RCX: 00000000021a32b0
 RDX: 000000000007d011 RSI: 000000000007d000 RDI: 00000000021262b0
 RBP: 00000000021262b0 R08: 0000000000000003 R09: 0000000000000086
 R10: 00000000000000eb R11: 00007fefbdf2baa0 R12: 0000000000000000
 R13: 0000000000000002 R14: 000000000007d000 R15: 0000000000001000

Signed-off-by: Liang Zhang <zhangliang5@huawei.com>
Message-Id: <20220222031239.1076682-1-zhangliang5@huawei.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoMerge branch 'kvm-ppc-cap-210' into kvm-next-5.18
Paolo Bonzini [Thu, 24 Feb 2022 13:49:56 +0000 (08:49 -0500)]
Merge branch 'kvm-ppc-cap-210' into kvm-next-5.18

2 years agoMerge tag 'kvm-s390-next-5.18-1' of git://git.kernel.org/pub/scm/linux/kernel/git...
Paolo Bonzini [Tue, 22 Feb 2022 18:07:28 +0000 (13:07 -0500)]
Merge tag 'kvm-s390-next-5.18-1' of git://git./linux/kernel/git/kvms390/linux into HEAD

KVM: s390: Changes for 5.18 part1

- add Claudio as Maintainer
- first step to do proper storage key checking
- testcase for missing memop check

2 years agoMerge branch 'kvm-ppc-cap-210' into kvm-master
Paolo Bonzini [Tue, 22 Feb 2022 14:07:16 +0000 (09:07 -0500)]
Merge branch 'kvm-ppc-cap-210' into kvm-master

By request of Nick Piggin:

> Patch 3 requires a KVM_CAP_PPC number allocated. QEMU maintainers are
> happy with it (link in changelog) just waiting on KVM upstreaming. Do
> you have objections to the series going to ppc/kvm tree first, or
> another option is you could take patch 3 alone first (it's relatively
> independent of the other 2) and ppc/kvm gets it from you?

2 years agoKVM: PPC: reserve capability 210 for KVM_CAP_PPC_AIL_MODE_3
Nicholas Piggin [Tue, 22 Feb 2022 14:06:54 +0000 (09:06 -0500)]
KVM: PPC: reserve capability 210 for KVM_CAP_PPC_AIL_MODE_3

Add KVM_CAP_PPC_AIL_MODE_3 to advertise the capability to set the AIL
resource mode to 3 with the H_SET_MODE hypercall. This capability
differs between processor types and KVM types (PR, HV, Nested HV), and
affects guest-visible behaviour.

QEMU will implement a cap-ail-mode-3 to control this behaviour[1], and
use the KVM CAP if available to determine KVM support[2].

Reviewed-by: Fabiano Rosas <farosas@linux.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2 years agoKVM: s390: Add missing vm MEM_OP size check
Janis Schoetterl-Glausch [Mon, 21 Feb 2022 16:32:37 +0000 (17:32 +0100)]
KVM: s390: Add missing vm MEM_OP size check

Check that size is not zero, preventing the following warning:

WARNING: CPU: 0 PID: 9692 at mm/vmalloc.c:3059 __vmalloc_node_range+0x528/0x648
Modules linked in:
CPU: 0 PID: 9692 Comm: memop Not tainted 5.17.0-rc3-e4+ #80
Hardware name: IBM 8561 T01 701 (LPAR)
Krnl PSW : 0704c00180000000 0000000082dc584c (__vmalloc_node_range+0x52c/0x648)
           R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3
Krnl GPRS: 0000000000000083 ffffffffffffffff 0000000000000000 0000000000000001
           0000038000000000 000003ff80000000 0000000000000cc0 000000008ebb8000
           0000000087a8a700 000000004040aeb1 000003ffd9f7dec8 000000008ebb8000
           000000009d9b8000 000000000102a1b4 00000380035afb68 00000380035afaa8
Krnl Code: 0000000082dc583ed028a7f4ff80        trtr    2036(41,%r10),3968(%r15)
           0000000082dc5844af000000            mc      0,0
          #0000000082dc5848af000000            mc      0,0
          >0000000082dc584ca7d90000            lghi    %r13,0
           0000000082dc5850b904002d            lgr     %r2,%r13
           0000000082dc5854eb6ff1080004        lmg     %r6,%r15,264(%r15)
           0000000082dc585a: 07fe                bcr     15,%r14
           0000000082dc585c47000700            bc      0,1792
Call Trace:
 [<0000000082dc584c>] __vmalloc_node_range+0x52c/0x648
 [<0000000082dc5b62>] vmalloc+0x5a/0x68
 [<000003ff8067f4ca>] kvm_arch_vm_ioctl+0x2da/0x2a30 [kvm]
 [<000003ff806705bc>] kvm_vm_ioctl+0x4ec/0x978 [kvm]
 [<0000000082e562fe>] __s390x_sys_ioctl+0xbe/0x100
 [<000000008360a9bc>] __do_syscall+0x1d4/0x200
 [<0000000083618bd2>] system_call+0x82/0xb0
Last Breaking-Event-Address:
 [<0000000082dc5348>] __vmalloc_node_range+0x28/0x648

Other than the warning, there is no ill effect from the missing check,
the condition is detected by subsequent code and causes a return
with ENOMEM.

Fixes: ef11c9463ae0 (KVM: s390: Add vm IOCTL for key checked guest absolute memory access)
Signed-off-by: Janis Schoetterl-Glausch <scgl@linux.ibm.com>
Link: https://lore.kernel.org/r/20220221163237.4122868-1-scgl@linux.ibm.com
Signed-off-by: Christian Borntraeger <borntraeger@linux.ibm.com>
2 years agoKVM: s390: Clarify key argument for MEM_OP in api docs
Janis Schoetterl-Glausch [Mon, 21 Feb 2022 14:36:57 +0000 (15:36 +0100)]
KVM: s390: Clarify key argument for MEM_OP in api docs

Clarify that the key argument represents the access key, not the whole
storage key.

Signed-off-by: Janis Schoetterl-Glausch <scgl@linux.ibm.com>
Link: https://lore.kernel.org/r/20220221143657.3712481-1-scgl@linux.ibm.com
Fixes: 5e35d0eb472b ("KVM: s390: Update api documentation for memop ioctl")
Signed-off-by: Christian Borntraeger <borntraeger@linux.ibm.com>