Both lock holder vCPU and IPI receiver that has halted are condidate for
boost. However, the PLE handler was originally designed to deal with the
lock holder preemption problem. The Intel PLE occurs when the spinlock
waiter is in kernel mode. This assumption doesn't hold for IPI receiver,
they can be in either kernel or user mode. the vCPU candidate in user mode
will not be boosted even if they should respond to IPIs. Some benchmarks
like pbzip2, swaptions etc do the TLB shootdown in kernel mode and most
of the time they are running in user mode. It can lead to a large number
of continuous PLE events because the IPI sender causes PLE events
repeatedly until the receiver is scheduled while the receiver is not
candidate for a boost.
This patch boosts the vCPU candidiate in user mode which is delivery
interrupt. We can observe the speed of pbzip2 improves 10% in 96 vCPUs
VM in over-subscribe scenario (The host machine is 2 socket, 48 cores,
96 HTs Intel CLX box). There is no performance regression for other
benchmarks like Unixbench spawn (most of the time contend read/write
lock in kernel mode), ebizzy (most of the time contend read/write sem
and TLB shoodtdown in kernel mode).
Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Message-Id: <1618542490-14756-1-git-send-email-wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add a capability for userspace to mirror SEV encryption context from
one vm to another. On our side, this is intended to support a
Migration Helper vCPU, but it can also be used generically to support
other in-guest workloads scheduled by the host. The intention is for
the primary guest and the mirror to have nearly identical memslots.
The primary benefits of this are that:
1) The VMs do not share KVM contexts (think APIC/MSRs/etc), so they
can't accidentally clobber each other.
2) The VMs can have different memory-views, which is necessary for post-copy
migration (the migration vCPUs on the target need to read and write to
pages, when the primary guest would VMEXIT).
This does not change the threat model for AMD SEV. Any memory involved
is still owned by the primary guest and its initial state is still
attested to through the normal SEV_LAUNCH_* flows. If userspace wanted
to circumvent SEV, they could achieve the same effect by simply attaching
a vCPU to the primary VM.
This patch deliberately leaves userspace in charge of the memslots for the
mirror, as it already has the power to mess with them in the primary guest.
This patch does not support SEV-ES (much less SNP), as it does not
handle handing off attested VMSAs to the mirror.
For additional context, we need a Migration Helper because SEV PSP
migration is far too slow for our live migration on its own. Using
an in-guest migrator lets us speed this up significantly.
Signed-off-by: Nathan Tempelman <natet@google.com>
Message-Id: <20210408223214.2582277-1-natet@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Convert a comment above kvm_io_bus_unregister_dev() into an actual
lockdep assertion, and opportunistically add curly braces to a multi-line
for-loop.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210412222050.876100-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Abort the walk of coalesced MMIO zones if kvm_io_bus_unregister_dev()
fails to allocate memory for the new instance of the bus. If it can't
instantiate a new bus, unregister_dev() destroys all devices _except_ the
target device. But, it doesn't tell the caller that it obliterated the
bus and invoked the destructor for all devices that were on the bus. In
the coalesced MMIO case, this can result in a deleted list entry
dereference due to attempting to continue iterating on coalesced_zones
after future entries (in the walk) have been deleted.
Opportunistically add curly braces to the for-loop, which encompasses
many lines but sneaks by without braces due to the guts being a single
if statement.
Fixes: f65886606c ("KVM: fix memory leak in kvm_io_bus_unregister_dev()")
Cc: stable@vger.kernel.org
Reported-by: Hao Sun <sunhao.th@gmail.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210412222050.876100-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If allocating a new instance of an I/O bus fails when unregistering a
device, wait to destroy the device until after all readers are guaranteed
to see the new null bus. Destroying devices before the bus is nullified
could lead to use-after-free since readers expect the devices on their
reference of the bus to remain valid.
Fixes: f65886606c ("KVM: fix memory leak in kvm_io_bus_unregister_dev()")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210412222050.876100-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Defer acquiring mmu_lock in the MMU notifier paths until a "hit" has been
detected in the memslots, i.e. don't take the lock for notifications that
don't affect the guest.
For small VMs, spurious locking is a minor annoyance. And for "volatile"
setups where the majority of notifications _are_ relevant, this barely
qualifies as an optimization.
But, for large VMs (hundreds of threads) with static setups, e.g. no
page migration, no swapping, etc..., the vast majority of MMU notifier
callbacks will be unrelated to the guest, e.g. will often be in response
to the userspace VMM adjusting its own virtual address space. In such
large VMs, acquiring mmu_lock can be painful as it blocks vCPUs from
handling page faults. In some scenarios it can even be "fatal" in the
sense that it causes unacceptable brownouts, e.g. when rebuilding huge
pages after live migration, a significant percentage of vCPUs will be
attempting to handle page faults.
x86's TDP MMU implementation is especially susceptible to spurious
locking due it taking mmu_lock for read when handling page faults.
Because rwlock is fair, a single writer will stall future readers, while
the writer is itself stalled waiting for in-progress readers to complete.
This is exacerbated by the MMU notifiers often firing multiple times in
quick succession, e.g. moving a page will (always?) invoke three separate
notifiers: .invalidate_range_start(), invalidate_range_end(), and
.change_pte(). Unnecessarily taking mmu_lock each time means even a
single spurious sequence can be problematic.
Note, this optimizes only the unpaired callbacks. Optimizing the
.invalidate_range_{start,end}() pairs is more complex and will be done in
a future patch.
Suggested-by: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210402005658.3024832-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Acquire and release mmu_lock in the __kvm_handle_hva_range() helper
instead of requiring the caller to do the same. This paves the way for
future patches to take mmu_lock if and only if an overlapping memslot is
found, without also having to introduce the on_lock() shenanigans used
to manipulate the notifier count and sequence.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210402005658.3024832-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Yank out the hva-based MMU notifier APIs now that all architectures that
use the notifiers have moved to the gfn-based APIs.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210402005658.3024832-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move the hva->gfn lookup for MMU notifiers into common code. Every arch
does a similar lookup, and some arch code is all but identical across
multiple architectures.
In addition to consolidating code, this will allow introducing
optimizations that will benefit all architectures without incurring
multiple walks of the memslots, e.g. by taking mmu_lock if and only if a
relevant range exists in the memslots.
The use of __always_inline to avoid indirect call retpolines, as done by
x86, may also benefit other architectures.
Consolidating the lookups also fixes a wart in x86, where the legacy MMU
and TDP MMU each do their own memslot walks.
Lastly, future enhancements to the memslot implementation, e.g. to add an
interval tree to track host address, will need to touch far less arch
specific code.
MIPS, PPC, and arm64 will be converted one at a time in future patches.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210402005658.3024832-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In KVM's .change_pte() notification callback, replace the notifier
sequence bump with a WARN_ON assertion that the notifier count is
elevated. An elevated count provides stricter protections than bumping
the sequence, and the sequence is guarnateed to be bumped before the
count hits zero.
When .change_pte() was added by commit 828502d300 ("ksm: add
mmu_notifier set_pte_at_notify()"), bumping the sequence was necessary
as .change_pte() would be invoked without any surrounding notifications.
However, since commit 6bdb913f0a ("mm: wrap calls to set_pte_at_notify
with invalidate_range_start and invalidate_range_end"), all calls to
.change_pte() are guaranteed to be surrounded by start() and end(), and
so are guaranteed to run with an elevated notifier count.
Note, wrapping .change_pte() with .invalidate_range_{start,end}() is a
bug of sorts, as invalidating the secondary MMU's (KVM's) PTE defeats
the purpose of .change_pte(). Every arch's kvm_set_spte_hva() assumes
.change_pte() is called when the relevant SPTE is present in KVM's MMU,
as the original goal was to accelerate Kernel Samepage Merging (KSM) by
updating KVM's SPTEs without requiring a VM-Exit (due to invalidating
the SPTE). I.e. it means that .change_pte() is effectively dead code
on _all_ architectures.
x86 and MIPS are clearcut nops if the old SPTE is not-present, and that
is guaranteed due to the prior invalidation. PPC simply unmaps the SPTE,
which again should be a nop due to the invalidation. arm64 is a bit
murky, but it's also likely a nop because kvm_pgtable_stage2_map() is
called without a cache pointer, which means it will map an entry if and
only if an existing PTE was found.
For now, take advantage of the bug to simplify future consolidation of
KVMs's MMU notifier code. Doing so will not greatly complicate fixing
.change_pte(), assuming it's even worth fixing. .change_pte() has been
broken for 8+ years and no one has complained. Even if there are
KSM+KVM users that care deeply about its performance, the benefits of
avoiding VM-Exits via .change_pte() need to be reevaluated to justify
the added complexity and testing burden. Ripping out .change_pte()
entirely would be a lot easier.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use GFP_KERNEL_ACCOUNT when allocating vCPUs to make it more obvious that
that the allocations are accounted, to make it easier to audit KVM's
allocations in the future, and to be consistent with other cache usage in
KVM.
When using SLAB/SLUB, this is a nop as the cache itself is created with
SLAB_ACCOUNT.
When using SLOB, there are caveats within caveats. SLOB doesn't honor
SLAB_ACCOUNT, so passing GFP_KERNEL_ACCOUNT will result in vCPU
allocations now being accounted. But, even that depends on internal
SLOB details as SLOB will only go to the page allocator when its cache is
depleted. That just happens to be extremely likely for vCPUs because the
size of kvm_vcpu is larger than the a page for almost all combinations of
architecture and page size. Whether or not the SLOB behavior is by
design is unknown; it's just as likely that no SLOB users care about
accounding and so no one has bothered to implemented support in SLOB.
Regardless, accounting vCPU allocations will not break SLOB+KVM+cgroup
users, if any exist.
Reviewed-by: Wanpeng Li <kernellwp@gmail.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210406190740.4055679-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move arm64's MMU notifier trace events into common code in preparation
for doing the hva->gfn lookup in common code. The alternative would be
to trace the gfn instead of hva, but that's not obviously better and
could also be done in common code. Tracing the notifiers is also quite
handy for debug regardless of architecture.
Remove a completely redundant tracepoint from PPC e500.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210326021957.1424875-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Track the range being invalidated by mmu_notifier and skip page fault
retries if the fault address is not affected by the in-progress
invalidation. Handle concurrent invalidations by finding the minimal
range which includes all ranges being invalidated. Although the combined
range may include unrelated addresses and cannot be shrunk as individual
invalidation operations complete, it is unlikely the marginal gains of
proper range tracking are worth the additional complexity.
The primary benefit of this change is the reduction in the likelihood of
extreme latency when handing a page fault due to another thread having
been preempted while modifying host virtual addresses.
Signed-off-by: David Stevens <stevensd@chromium.org>
Message-Id: <20210222024522.1751719-3-stevensd@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use kvm_pfn_t, a.k.a. u64, for the local 'pfn' variable when retrieving
a so called "remapped" hva/pfn pair. In theory, the hva could resolve to
a pfn in high memory on a 32-bit kernel.
This bug was inadvertantly exposed by commit bd2fae8da7 ("KVM: do not
assume PTE is writable after follow_pfn"), which added an error PFN value
to the mix, causing gcc to comlain about overflowing the unsigned long.
arch/x86/kvm/../../../virt/kvm/kvm_main.c: In function ‘hva_to_pfn_remapped’:
include/linux/kvm_host.h:89:30: error: conversion from ‘long long unsigned int’
to ‘long unsigned int’ changes value from
‘9218868437227405314’ to ‘2’ [-Werror=overflow]
89 | #define KVM_PFN_ERR_RO_FAULT (KVM_PFN_ERR_MASK + 2)
| ^
virt/kvm/kvm_main.c:1935:9: note: in expansion of macro ‘KVM_PFN_ERR_RO_FAULT’
Cc: stable@vger.kernel.org
Fixes: add6a0cd1c ("KVM: MMU: try to fix up page faults before giving up")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210208201940.1258328-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently, the follow_pfn function is exported for modules but
follow_pte is not. However, follow_pfn is very easy to misuse,
because it does not provide protections (so most of its callers
assume the page is writable!) and because it returns after having
already unlocked the page table lock.
Provide instead a simplified version of follow_pte that does
not have the pmdpp and range arguments. The older version
survives as follow_invalidate_pte() for use by fs/dax.c.
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add a read / write lock to be used in place of the MMU spinlock on x86.
The rwlock will enable the TDP MMU to handle page faults, and other
operations in parallel in future commits.
Reviewed-by: Peter Feiner <pfeiner@google.com>
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210202185734.1680553-19-bgardon@google.com>
[Introduce virt/kvm/mmu_lock.h - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In order to convert an HVA to a PFN, KVM usually tries to use
the get_user_pages family of functinso. This however is not
possible for VM_IO vmas; in that case, KVM instead uses follow_pfn.
In doing this however KVM loses the information on whether the
PFN is writable. That is usually not a problem because the main
use of VM_IO vmas with KVM is for BARs in PCI device assignment,
however it is a bug. To fix it, use follow_pte and check pte_write
while under the protection of the PTE lock. The information can
be used to fail hva_to_pfn_remapped or passed back to the
caller via *writable.
Usage of follow_pfn was introduced in commit add6a0cd1c ("KVM: MMU: try to fix
up page faults before giving up", 2016-07-05); however, even older version
have the same issue, all the way back to commit 2e2e3738af ("KVM:
Handle vma regions with no backing page", 2008-07-20), as they also did
not check whether the PFN was writable.
Fixes: 2e2e3738af ("KVM: Handle vma regions with no backing page")
Reported-by: David Stevens <stevensd@google.com>
Cc: 3pvd@google.com
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM/arm64 fixes for 5.11, take #2
- Don't allow tagged pointers to point to memslots
- Filter out ARMv8.1+ PMU events on v8.0 hardware
- Hide PMU registers from userspace when no PMU is configured
- More PMU cleanups
- Don't try to handle broken PSCI firmware
- More sys_reg() to reg_to_encoding() conversions
The use of a tagged address could be pretty confusing for the
whole memslot infrastructure as well as the MMU notifiers.
Forbid it altogether, as it never quite worked the first place.
Cc: stable@vger.kernel.org
Reported-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Pull kvm fixes from Paolo Bonzini:
"x86:
- Fixes for the new scalable MMU
- Fixes for migration of nested hypervisors on AMD
- Fix for clang integrated assembler
- Fix for left shift by 64 (UBSAN)
- Small cleanups
- Straggler SEV-ES patch
ARM:
- VM init cleanups
- PSCI relay cleanups
- Kill CONFIG_KVM_ARM_PMU
- Fixup __init annotations
- Fixup reg_to_encoding()
- Fix spurious PMCR_EL0 access
Misc:
- selftests cleanups"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (38 commits)
KVM: x86: __kvm_vcpu_halt can be static
KVM: SVM: Add support for booting APs in an SEV-ES guest
KVM: nSVM: cancel KVM_REQ_GET_NESTED_STATE_PAGES on nested vmexit
KVM: nSVM: mark vmcb as dirty when forcingly leaving the guest mode
KVM: nSVM: correctly restore nested_run_pending on migration
KVM: x86/mmu: Clarify TDP MMU page list invariants
KVM: x86/mmu: Ensure TDP MMU roots are freed after yield
kvm: check tlbs_dirty directly
KVM: x86: change in pv_eoi_get_pending() to make code more readable
MAINTAINERS: Really update email address for Sean Christopherson
KVM: x86: fix shift out of bounds reported by UBSAN
KVM: selftests: Implement perf_test_util more conventionally
KVM: selftests: Use vm_create_with_vcpus in create_vm
KVM: selftests: Factor out guest mode code
KVM/SVM: Remove leftover __svm_vcpu_run prototype from svm.c
KVM: SVM: Add register operand to vmsave call in sev_es_vcpu_load
KVM: x86/mmu: Optimize not-present/MMIO SPTE check in get_mmio_spte()
KVM: x86/mmu: Use raw level to index into MMIO walks' sptes array
KVM: x86/mmu: Get root level from walkers when retrieving MMIO SPTE
KVM: x86/mmu: Use -1 to flag an undefined spte in get_mmio_spte()
...
In kvm_mmu_notifier_invalidate_range_start(), tlbs_dirty is used as:
need_tlb_flush |= kvm->tlbs_dirty;
with need_tlb_flush's type being int and tlbs_dirty's type being long.
It means that tlbs_dirty is always used as int and the higher 32 bits
is useless. We need to check tlbs_dirty in a correct way and this
change checks it directly without propagating it to need_tlb_flush.
Note: it's _extremely_ unlikely this neglecting of higher 32 bits can
cause problems in practice. It would require encountering tlbs_dirty
on a 4 billion count boundary, and KVM would need to be using shadow
paging or be running a nested guest.
Cc: stable@vger.kernel.org
Fixes: a4ee1ca4a3 ("KVM: MMU: delay flush all tlbs on sync_page path")
Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Message-Id: <20201217154118.16497-1-jiangshanlai@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Pull KVM updates from Paolo Bonzini:
"Much x86 work was pushed out to 5.12, but ARM more than made up for it.
ARM:
- PSCI relay at EL2 when "protected KVM" is enabled
- New exception injection code
- Simplification of AArch32 system register handling
- Fix PMU accesses when no PMU is enabled
- Expose CSV3 on non-Meltdown hosts
- Cache hierarchy discovery fixes
- PV steal-time cleanups
- Allow function pointers at EL2
- Various host EL2 entry cleanups
- Simplification of the EL2 vector allocation
s390:
- memcg accouting for s390 specific parts of kvm and gmap
- selftest for diag318
- new kvm_stat for when async_pf falls back to sync
x86:
- Tracepoints for the new pagetable code from 5.10
- Catch VFIO and KVM irqfd events before userspace
- Reporting dirty pages to userspace with a ring buffer
- SEV-ES host support
- Nested VMX support for wait-for-SIPI activity state
- New feature flag (AVX512 FP16)
- New system ioctl to report Hyper-V-compatible paravirtualization features
Generic:
- Selftest improvements"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (171 commits)
KVM: SVM: fix 32-bit compilation
KVM: SVM: Add AP_JUMP_TABLE support in prep for AP booting
KVM: SVM: Provide support to launch and run an SEV-ES guest
KVM: SVM: Provide an updated VMRUN invocation for SEV-ES guests
KVM: SVM: Provide support for SEV-ES vCPU loading
KVM: SVM: Provide support for SEV-ES vCPU creation/loading
KVM: SVM: Update ASID allocation to support SEV-ES guests
KVM: SVM: Set the encryption mask for the SVM host save area
KVM: SVM: Add NMI support for an SEV-ES guest
KVM: SVM: Guest FPU state save/restore not needed for SEV-ES guest
KVM: SVM: Do not report support for SMM for an SEV-ES guest
KVM: x86: Update __get_sregs() / __set_sregs() to support SEV-ES
KVM: SVM: Add support for CR8 write traps for an SEV-ES guest
KVM: SVM: Add support for CR4 write traps for an SEV-ES guest
KVM: SVM: Add support for CR0 write traps for an SEV-ES guest
KVM: SVM: Add support for EFER write traps for an SEV-ES guest
KVM: SVM: Support string IO operations for an SEV-ES guest
KVM: SVM: Support MMIO for an SEV-ES guest
KVM: SVM: Create trace events for VMGEXIT MSR protocol processing
KVM: SVM: Create trace events for VMGEXIT processing
...
