Pull kvm updates from Paolo Bonzini:
"ARM64:
- Enable the per-vcpu dirty-ring tracking mechanism, together with an
option to keep the good old dirty log around for pages that are
dirtied by something other than a vcpu.
- Switch to the relaxed parallel fault handling, using RCU to delay
page table reclaim and giving better performance under load.
- Relax the MTE ABI, allowing a VMM to use the MAP_SHARED mapping
option, which multi-process VMMs such as crosvm rely on (see merge
commit 382b5b87a9: "Fix a number of issues with MTE, such as
races on the tags being initialised vs the PG_mte_tagged flag as
well as the lack of support for VM_SHARED when KVM is involved.
Patches from Catalin Marinas and Peter Collingbourne").
- Merge the pKVM shadow vcpu state tracking that allows the
hypervisor to have its own view of a vcpu, keeping that state
private.
- Add support for the PMUv3p5 architecture revision, bringing support
for 64bit counters on systems that support it, and fix the
no-quite-compliant CHAIN-ed counter support for the machines that
actually exist out there.
- Fix a handful of minor issues around 52bit VA/PA support (64kB
pages only) as a prefix of the oncoming support for 4kB and 16kB
pages.
- Pick a small set of documentation and spelling fixes, because no
good merge window would be complete without those.
s390:
- Second batch of the lazy destroy patches
- First batch of KVM changes for kernel virtual != physical address
support
- Removal of a unused function
x86:
- Allow compiling out SMM support
- Cleanup and documentation of SMM state save area format
- Preserve interrupt shadow in SMM state save area
- Respond to generic signals during slow page faults
- Fixes and optimizations for the non-executable huge page errata
fix.
- Reprogram all performance counters on PMU filter change
- Cleanups to Hyper-V emulation and tests
- Process Hyper-V TLB flushes from a nested guest (i.e. from a L2
guest running on top of a L1 Hyper-V hypervisor)
- Advertise several new Intel features
- x86 Xen-for-KVM:
- Allow the Xen runstate information to cross a page boundary
- Allow XEN_RUNSTATE_UPDATE flag behaviour to be configured
- Add support for 32-bit guests in SCHEDOP_poll
- Notable x86 fixes and cleanups:
- One-off fixes for various emulation flows (SGX, VMXON, NRIPS=0).
- Reinstate IBPB on emulated VM-Exit that was incorrectly dropped
a few years back when eliminating unnecessary barriers when
switching between vmcs01 and vmcs02.
- Clean up vmread_error_trampoline() to make it more obvious that
params must be passed on the stack, even for x86-64.
- Let userspace set all supported bits in MSR_IA32_FEAT_CTL
irrespective of the current guest CPUID.
- Fudge around a race with TSC refinement that results in KVM
incorrectly thinking a guest needs TSC scaling when running on a
CPU with a constant TSC, but no hardware-enumerated TSC
frequency.
- Advertise (on AMD) that the SMM_CTL MSR is not supported
- Remove unnecessary exports
Generic:
- Support for responding to signals during page faults; introduces
new FOLL_INTERRUPTIBLE flag that was reviewed by mm folks
Selftests:
- Fix an inverted check in the access tracking perf test, and restore
support for asserting that there aren't too many idle pages when
running on bare metal.
- Fix build errors that occur in certain setups (unsure exactly what
is unique about the problematic setup) due to glibc overriding
static_assert() to a variant that requires a custom message.
- Introduce actual atomics for clear/set_bit() in selftests
- Add support for pinning vCPUs in dirty_log_perf_test.
- Rename the so called "perf_util" framework to "memstress".
- Add a lightweight psuedo RNG for guest use, and use it to randomize
the access pattern and write vs. read percentage in the memstress
tests.
- Add a common ucall implementation; code dedup and pre-work for
running SEV (and beyond) guests in selftests.
- Provide a common constructor and arch hook, which will eventually
be used by x86 to automatically select the right hypercall (AMD vs.
Intel).
- A bunch of added/enabled/fixed selftests for ARM64, covering
memslots, breakpoints, stage-2 faults and access tracking.
- x86-specific selftest changes:
- Clean up x86's page table management.
- Clean up and enhance the "smaller maxphyaddr" test, and add a
related test to cover generic emulation failure.
- Clean up the nEPT support checks.
- Add X86_PROPERTY_* framework to retrieve multi-bit CPUID values.
- Fix an ordering issue in the AMX test introduced by recent
conversions to use kvm_cpu_has(), and harden the code to guard
against similar bugs in the future. Anything that tiggers
caching of KVM's supported CPUID, kvm_cpu_has() in this case,
effectively hides opt-in XSAVE features if the caching occurs
before the test opts in via prctl().
Documentation:
- Remove deleted ioctls from documentation
- Clean up the docs for the x86 MSR filter.
- Various fixes"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (361 commits)
KVM: x86: Add proper ReST tables for userspace MSR exits/flags
KVM: selftests: Allocate ucall pool from MEM_REGION_DATA
KVM: arm64: selftests: Align VA space allocator with TTBR0
KVM: arm64: Fix benign bug with incorrect use of VA_BITS
KVM: arm64: PMU: Fix period computation for 64bit counters with 32bit overflow
KVM: x86: Advertise that the SMM_CTL MSR is not supported
KVM: x86: remove unnecessary exports
KVM: selftests: Fix spelling mistake "probabalistic" -> "probabilistic"
tools: KVM: selftests: Convert clear/set_bit() to actual atomics
tools: Drop "atomic_" prefix from atomic test_and_set_bit()
tools: Drop conflicting non-atomic test_and_{clear,set}_bit() helpers
KVM: selftests: Use non-atomic clear/set bit helpers in KVM tests
perf tools: Use dedicated non-atomic clear/set bit helpers
tools: Take @bit as an "unsigned long" in {clear,set}_bit() helpers
KVM: arm64: selftests: Enable single-step without a "full" ucall()
KVM: x86: fix APICv/x2AVIC disabled when vm reboot by itself
KVM: Remove stale comment about KVM_REQ_UNHALT
KVM: Add missing arch for KVM_CREATE_DEVICE and KVM_{SET,GET}_DEVICE_ATTR
KVM: Reference to kvm_userspace_memory_region in doc and comments
KVM: Delete all references to removed KVM_SET_MEMORY_ALIAS ioctl
...
Sharing 'kvm_arm_vmid_bits' between EL1 and EL2 allows the host to
modify the variable arbitrarily, potentially leading to all sorts of
shenanians as this is used to configure the VTTBR register for the
guest stage-2.
In preparation for unmapping host sections entirely from EL2, maintain
a copy of 'kvm_arm_vmid_bits' in the pKVM hypervisor and initialise it
from the host value while it is still trusted.
Tested-by: Vincent Donnefort <vdonnefort@google.com>
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20221110190259.26861-23-will@kernel.org
When pKVM is enabled, the hypervisor at EL2 does not trust the host at
EL1 and must therefore prevent it from having unrestricted access to
internal hypervisor state.
The 'kvm_arm_hyp_percpu_base' array holds the offsets for hypervisor
per-cpu allocations, so move this this into the nVHE code where it
cannot be modified by the untrusted host at EL1.
Tested-by: Vincent Donnefort <vdonnefort@google.com>
Signed-off-by: Quentin Perret <qperret@google.com>
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20221110190259.26861-22-will@kernel.org
In preparation for handling cache maintenance of guest pages from within
the pKVM hypervisor at EL2, introduce an EL2 copy of icache_inval_pou()
which will later be plumbed into the stage-2 page-table cache
maintenance callbacks, ensuring that the initial contents of pages
mapped as executable into the guest stage-2 page-table is visible to the
instruction fetcher.
Tested-by: Vincent Donnefort <vdonnefort@google.com>
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20221110190259.26861-17-will@kernel.org
Clone the implementations of strrchr() and memchr() in lib/string.c so
we can use them in the standalone zboot decompressor app. These routines
are used by the FDT handling code.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Split the efi_printk() routine into its own source file, and provide
local implementations of strlen() and strnlen() so that the standalone
zboot app can efi_err and efi_info etc.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
We will no longer be able to call into the kernel image once we merge
the decompressor with the EFI stub, so we need our own implementation of
memcmp(). Let's add the one from lib/string.c and simplify it.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
In preparation for moving the EFI stub functionality into the zboot
decompressor, switch to the stub's implementation of strncmp()
unconditionally.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
The efi_enter_kernel() routine will be shared between the existing EFI
stub and the zboot decompressor, and the version of
dcache_clean_to_poc() that the core kernel exports to the stub will not
be available in the latter case.
So move the handling into the .c file which will remain part of the stub
build that integrates directly with the kernel proper.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Pull EFI updates from Ard Biesheuvel:
"A bit more going on than usual in the EFI subsystem. The main driver
for this has been the introduction of the LoonArch architecture last
cycle, which inspired some cleanup and refactoring of the EFI code.
Another driver for EFI changes this cycle and in the future is
confidential compute.
The LoongArch architecture does not use either struct bootparams or DT
natively [yet], and so passing information between the EFI stub and
the core kernel using either of those is undesirable. And in general,
overloading DT has been a source of issues on arm64, so using DT for
this on new architectures is a to avoid for the time being (even if we
might converge on something DT based for non-x86 architectures in the
future). For this reason, in addition to the patch that enables EFI
boot for LoongArch, there are a number of refactoring patches applied
on top of which separate the DT bits from the generic EFI stub bits.
These changes are on a separate topich branch that has been shared
with the LoongArch maintainers, who will include it in their pull
request as well. This is not ideal, but the best way to manage the
conflicts without stalling LoongArch for another cycle.
Another development inspired by LoongArch is the newly added support
for EFI based decompressors. Instead of adding yet another
arch-specific incarnation of this pattern for LoongArch, we are
introducing an EFI app based on the existing EFI libstub
infrastructure that encapulates the decompression code we use on other
architectures, but in a way that is fully generic. This has been
developed and tested in collaboration with distro and systemd folks,
who are eager to start using this for systemd-boot and also for arm64
secure boot on Fedora. Note that the EFI zimage files this introduces
can also be decompressed by non-EFI bootloaders if needed, as the
image header describes the location of the payload inside the image,
and the type of compression that was used. (Note that Fedora's arm64
GRUB is buggy [0] so you'll need a recent version or switch to
systemd-boot in order to use this.)
Finally, we are adding TPM measurement of the kernel command line
provided by EFI. There is an oversight in the TCG spec which results
in a blind spot for command line arguments passed to loaded images,
which means that either the loader or the stub needs to take the
measurement. Given the combinatorial explosion I am anticipating when
it comes to firmware/bootloader stacks and firmware based attestation
protocols (SEV-SNP, TDX, DICE, DRTM), it is good to set a baseline now
when it comes to EFI measured boot, which is that the kernel measures
the initrd and command line. Intermediate loaders can measure
additional assets if needed, but with the baseline in place, we can
deploy measured boot in a meaningful way even if you boot into Linux
straight from the EFI firmware.
Summary:
- implement EFI boot support for LoongArch
- implement generic EFI compressed boot support for arm64, RISC-V and
LoongArch, none of which implement a decompressor today
- measure the kernel command line into the TPM if measured boot is in
effect
- refactor the EFI stub code in order to isolate DT dependencies for
architectures other than x86
- avoid calling SetVirtualAddressMap() on arm64 if the configured
size of the VA space guarantees that doing so is unnecessary
- move some ARM specific code out of the generic EFI source files
- unmap kernel code from the x86 mixed mode 1:1 page tables"
* tag 'efi-next-for-v6.1' of git://git.kernel.org/pub/scm/linux/kernel/git/efi/efi: (24 commits)
efi/arm64: libstub: avoid SetVirtualAddressMap() when possible
efi: zboot: create MemoryMapped() device path for the parent if needed
efi: libstub: fix up the last remaining open coded boot service call
efi/arm: libstub: move ARM specific code out of generic routines
efi/libstub: measure EFI LoadOptions
efi/libstub: refactor the initrd measuring functions
efi/loongarch: libstub: remove dependency on flattened DT
efi: libstub: install boot-time memory map as config table
efi: libstub: remove DT dependency from generic stub
efi: libstub: unify initrd loading between architectures
efi: libstub: remove pointless goto kludge
efi: libstub: simplify efi_get_memory_map() and struct efi_boot_memmap
efi: libstub: avoid efi_get_memory_map() for allocating the virt map
efi: libstub: drop pointless get_memory_map() call
efi: libstub: fix type confusion for load_options_size
arm64: efi: enable generic EFI compressed boot
loongarch: efi: enable generic EFI compressed boot
riscv: efi: enable generic EFI compressed boot
efi/libstub: implement generic EFI zboot
efi/libstub: move efi_system_table global var into separate object
...
The stub is used in different execution environments, but on arm64,
RISC-V and LoongArch, we still use the core kernel's implementation of
memcpy and memset, as they are just a branch instruction away, and can
generally be reused even from code such as the EFI stub that runs in a
completely different address space.
KAsan complicates this slightly, resulting in the need for some hacks to
expose the uninstrumented, __ prefixed versions as the normal ones, as
the latter are instrumented to include the KAsan checks, which only work
in the core kernel.
Unfortunately, #define'ing memcpy to __memcpy when building C code does
not guarantee that no explicit memcpy() calls will be emitted. And with
the upcoming zboot support, which consists of a separate binary which
therefore needs its own implementation of memcpy/memset anyway, it's
better to provide one explicitly instead of linking to the existing one.
Given that EFI exposes implementations of memmove() and memset() via the
boot services table, let's wire those up in the appropriate way, and
drop the references to the core kernel ones.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
For each instance of an alternative, the compiler outputs a distinct
copy of the alternative instructions into a subsection. As the compiler
doesn't have special knowledge of alternatives, it cannot coalesce these
to save space.
In a defconfig kernel built with GCC 12.1.0, there are approximately
10,000 instances of alternative_has_feature_likely(), where the
replacement instruction is always a NOP. As NOPs are
position-independent, we don't need a unique copy per alternative
sequence.
This patch adds a callback to patch an alternative sequence with NOPs,
and make use of this in alternative_has_feature_likely(). So that this
can be used for other sites in future, this is written to patch multiple
instructions up to the original sequence length.
For NVHE, an alias is added to image-vars.h.
For modules, the callback is exported. Note that as modules are loaded
within 2GiB of the kernel, an alt_instr entry in a module can always
refer directly to the callback, and no special handling is necessary.
When building with GCC 12.1.0, the vmlinux is ~158KiB smaller, though
the resulting Image size is unchanged due to alignment constraints and
padding:
| % ls -al vmlinux-*
| -rwxr-xr-x 1 mark mark 134644592 Sep 1 14:52 vmlinux-after
| -rwxr-xr-x 1 mark mark 134486232 Sep 1 14:50 vmlinux-before
| % ls -al Image-*
| -rw-r--r-- 1 mark mark 37108224 Sep 1 14:52 Image-after
| -rw-r--r-- 1 mark mark 37108224 Sep 1 14:50 Image-before
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Ard Biesheuvel <ardb@kernel.org>
Cc: James Morse <james.morse@arm.com>
Cc: Joey Gouly <joey.gouly@arm.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Will Deacon <will@kernel.org>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20220912162210.3626215-9-mark.rutland@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Currrently we use a mixture of alternative sequences and static branches
to handle features detected at boot time. For ease of maintenance we
generally prefer to use static branches in C code, but this has a few
downsides:
* Each static branch has metadata in the __jump_table section, which is
not discarded after features are finalized. This wastes some space,
and slows down the patching of other static branches.
* The static branches are patched at a different point in time from the
alternatives, so changes are not atomic. This leaves a transient
period where there could be a mismatch between the behaviour of
alternatives and static branches, which could be problematic for some
features (e.g. pseudo-NMI).
* More (instrumentable) kernel code is executed to patch each static
branch, which can be risky when patching certain features (e.g.
irqflags management for pseudo-NMI).
* When CONFIG_JUMP_LABEL=n, static branches are turned into a load of a
flag and a conditional branch. This means it isn't safe to use such
static branches in an alternative address space (e.g. the NVHE/PKVM
hyp code), where the generated address isn't safe to acccess.
To deal with these issues, this patch introduces new
alternative_has_feature_*() helpers, which work like static branches but
are patched using alternatives. This ensures the patching is performed
at the same time as other alternative patching, allows the metadata to
be freed after patching, and is safe for use in alternative address
spaces.
Note that all supported toolchains have asm goto support, and since
commit:
a0a12c3ed0 ("asm goto: eradicate CC_HAS_ASM_GOTO)"
... the CC_HAS_ASM_GOTO Kconfig symbol has been removed, so no feature
check is necessary, and we can always make use of asm goto.
Additionally, note that:
* This has no impact on cpus_have_cap(), which is a dynamic check.
* This has no functional impact on cpus_have_const_cap(). The branches
are patched slightly later than before this patch, but these branches
are not reachable until caps have been finalised.
* It is now invalid to use cpus_have_final_cap() in the window between
feature detection and patching. All existing uses are only expected
after patching anyway, so this should not be a problem.
* The LSE atomics will now be enabled during alternatives patching
rather than immediately before. As the LL/SC an LSE atomics are
functionally equivalent this should not be problematic.
When building defconfig with GCC 12.1.0, the resulting Image is 64KiB
smaller:
| % ls -al Image-*
| -rw-r--r-- 1 mark mark 37108224 Aug 23 09:56 Image-after
| -rw-r--r-- 1 mark mark 37173760 Aug 23 09:54 Image-before
According to bloat-o-meter.pl:
| add/remove: 44/34 grow/shrink: 602/1294 up/down: 39692/-61108 (-21416)
| Function old new delta
| [...]
| Total: Before=16618336, After=16596920, chg -0.13%
| add/remove: 0/2 grow/shrink: 0/0 up/down: 0/-1296 (-1296)
| Data old new delta
| arm64_const_caps_ready 16 - -16
| cpu_hwcap_keys 1280 - -1280
| Total: Before=8987120, After=8985824, chg -0.01%
| add/remove: 0/0 grow/shrink: 0/0 up/down: 0/0 (0)
| RO Data old new delta
| Total: Before=18408, After=18408, chg +0.00%
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Ard Biesheuvel <ardb@kernel.org>
Cc: James Morse <james.morse@arm.com>
Cc: Joey Gouly <joey.gouly@arm.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Will Deacon <will@kernel.org>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20220912162210.3626215-8-mark.rutland@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Currently, a build with CONFIG_EFI=n and CONFIG_KASAN=y will not
complete successfully because of missing symbols. This is due to the
fact that the __pi_ prefixed aliases for __memcpy/__memmove were put
inside a #ifdef CONFIG_EFI block inadvertently, and are therefore
missing from the build in question.
These definitions should only be provided when needed, as they will
otherwise clutter up the symbol table, kallsyms etc for no reason.
Fortunately, instead of using CPP conditionals, we can achieve the same
result by using the linker's PROVIDE() directive, which only defines a
symbol if it is required to complete the link. So let's use that for all
symbols alias definitions.
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20220629083246.3729177-1-ardb@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
Currently, when KASLR is in effect, we set up the kernel virtual address
space twice: the first time, the KASLR seed is looked up in the device
tree, and the kernel virtual mapping is torn down and recreated again,
after which the relocations are applied a second time. The latter step
means that statically initialized global pointer variables will be reset
to their initial values, and to ensure that BSS variables are not set to
values based on the initial translation, they are cleared again as well.
All of this is needed because we need the command line (taken from the
DT) to tell us whether or not to randomize the virtual address space
before entering the kernel proper. However, this code has expanded
little by little and now creates global state unrelated to the virtual
randomization of the kernel before the mapping is torn down and set up
again, and the BSS cleared for a second time. This has created some
issues in the past, and it would be better to avoid this little dance if
possible.
So instead, let's use the temporary mapping of the device tree, and
execute the bare minimum of code to decide whether or not KASLR should
be enabled, and what the seed is. Only then, create the virtual kernel
mapping, clear BSS, etc and proceed as normal. This avoids the issues
around inconsistent global state due to BSS being cleared twice, and is
generally more maintainable, as it permits us to defer all the remaining
DT parsing and KASLR initialization to a later time.
This means the relocation fixup code runs only a single time as well,
allowing us to simplify the RELR handling code too, which is not
idempotent and was therefore required to keep track of the offset that
was applied the first time around.
Note that this means we have to clone a pair of FDT library objects, so
that we can control how they are built - we need the stack protector
and other instrumentation disabled so that the code can tolerate being
called this early. Note that only the kernel page tables and the
temporary stack are mapped read-write at this point, which ensures that
the early code does not modify any global state inadvertently.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20220624150651.1358849-21-ardb@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>
Pull kvm updates from Paolo Bonzini:
"ARM:
- Proper emulation of the OSLock feature of the debug architecture
- Scalibility improvements for the MMU lock when dirty logging is on
- New VMID allocator, which will eventually help with SVA in VMs
- Better support for PMUs in heterogenous systems
- PSCI 1.1 support, enabling support for SYSTEM_RESET2
- Implement CONFIG_DEBUG_LIST at EL2
- Make CONFIG_ARM64_ERRATUM_2077057 default y
- Reduce the overhead of VM exit when no interrupt is pending
- Remove traces of 32bit ARM host support from the documentation
- Updated vgic selftests
- Various cleanups, doc updates and spelling fixes
RISC-V:
- Prevent KVM_COMPAT from being selected
- Optimize __kvm_riscv_switch_to() implementation
- RISC-V SBI v0.3 support
s390:
- memop selftest
- fix SCK locking
- adapter interruptions virtualization for secure guests
- add Claudio Imbrenda as maintainer
- first step to do proper storage key checking
x86:
- Continue switching kvm_x86_ops to static_call(); introduce
static_call_cond() and __static_call_ret0 when applicable.
- Cleanup unused arguments in several functions
- Synthesize AMD 0x80000021 leaf
- Fixes and optimization for Hyper-V sparse-bank hypercalls
- Implement Hyper-V's enlightened MSR bitmap for nested SVM
- Remove MMU auditing
- Eager splitting of page tables (new aka "TDP" MMU only) when dirty
page tracking is enabled
- Cleanup the implementation of the guest PGD cache
- Preparation for the implementation of Intel IPI virtualization
- Fix some segment descriptor checks in the emulator
- Allow AMD AVIC support on systems with physical APIC ID above 255
- Better API to disable virtualization quirks
- Fixes and optimizations for the zapping of page tables:
- Zap roots in two passes, avoiding RCU read-side critical
sections that last too long for very large guests backed by 4
KiB SPTEs.
- Zap invalid and defunct roots asynchronously via
concurrency-managed work queue.
- Allowing yielding when zapping TDP MMU roots in response to the
root's last reference being put.
- Batch more TLB flushes with an RCU trick. Whoever frees the
paging structure now holds RCU as a proxy for all vCPUs running
in the guest, i.e. to prolongs the grace period on their behalf.
It then kicks the the vCPUs out of guest mode before doing
rcu_read_unlock().
Generic:
- Introduce __vcalloc and use it for very large allocations that need
memcg accounting"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (246 commits)
KVM: use kvcalloc for array allocations
KVM: x86: Introduce KVM_CAP_DISABLE_QUIRKS2
kvm: x86: Require const tsc for RT
KVM: x86: synthesize CPUID leaf 0x80000021h if useful
KVM: x86: add support for CPUID leaf 0x80000021
KVM: x86: do not use KVM_X86_OP_OPTIONAL_RET0 for get_mt_mask
Revert "KVM: x86/mmu: Zap only TDP MMU leafs in kvm_zap_gfn_range()"
kvm: x86/mmu: Flush TLB before zap_gfn_range releases RCU
KVM: arm64: fix typos in comments
KVM: arm64: Generalise VM features into a set of flags
KVM: s390: selftests: Add error memop tests
KVM: s390: selftests: Add more copy memop tests
KVM: s390: selftests: Add named stages for memop test
KVM: s390: selftests: Add macro as abstraction for MEM_OP
KVM: s390: selftests: Split memop tests
KVM: s390x: fix SCK locking
RISC-V: KVM: Implement SBI HSM suspend call
RISC-V: KVM: Add common kvm_riscv_vcpu_wfi() function
RISC-V: Add SBI HSM suspend related defines
RISC-V: KVM: Implement SBI v0.3 SRST extension
...
Future CPUs may implement a clearbhb instruction that is sufficient
to mitigate SpectreBHB. CPUs that implement this instruction, but
not CSV2.3 must be affected by Spectre-BHB.
Add support to use this instruction as the BHB mitigation on CPUs
that support it. The instruction is in the hint space, so it will
be treated by a NOP as older CPUs.
Reviewed-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: James Morse <james.morse@arm.com>
Speculation attacks against some high-performance processors can
make use of branch history to influence future speculation.
When taking an exception from user-space, a sequence of branches
or a firmware call overwrites or invalidates the branch history.
The sequence of branches is added to the vectors, and should appear
before the first indirect branch. For systems using KPTI the sequence
is added to the kpti trampoline where it has a free register as the exit
from the trampoline is via a 'ret'. For systems not using KPTI, the same
register tricks are used to free up a register in the vectors.
For the firmware call, arch-workaround-3 clobbers 4 registers, so
there is no choice but to save them to the EL1 stack. This only happens
for entry from EL0, so if we take an exception due to the stack access,
it will not become re-entrant.
For KVM, the existing branch-predictor-hardening vectors are used.
When a spectre version of these vectors is in use, the firmware call
is sufficient to mitigate against Spectre-BHB. For the non-spectre
versions, the sequence of branches is added to the indirect vector.
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: James Morse <james.morse@arm.com>
Move the definition of kvm_arm_pmu_available to pmu-emul.c and, out of
"necessity", hide it behind CONFIG_HW_PERF_EVENTS. Provide a stub for
the key's wrapper, kvm_arm_support_pmu_v3(). Moving the key's definition
out of perf.c will allow a future commit to delete perf.c entirely.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20211111020738.2512932-16-seanjc@google.com
Although naming across the codebase isn't that consistent, it
tends to follow certain patterns. Moreover, the term "flush"
isn't defined in the Arm Architecture reference manual, and might
be interpreted to mean clean, invalidate, or both for a cache.
Rename arm64-internal functions to make the naming internally
consistent, as well as making it consistent with the Arm ARM, by
specifying whether it applies to the instruction, data, or both
caches, whether the operation is a clean, invalidate, or both.
Also specify which point the operation applies to, i.e., to the
point of unification (PoU), coherency (PoC), or persistence
(PoP).
This commit applies the following sed transformation to all files
under arch/arm64:
"s/\b__flush_cache_range\b/caches_clean_inval_pou_macro/g;"\
"s/\b__flush_icache_range\b/caches_clean_inval_pou/g;"\
"s/\binvalidate_icache_range\b/icache_inval_pou/g;"\
"s/\b__flush_dcache_area\b/dcache_clean_inval_poc/g;"\
"s/\b__inval_dcache_area\b/dcache_inval_poc/g;"\
"s/__clean_dcache_area_poc\b/dcache_clean_poc/g;"\
"s/\b__clean_dcache_area_pop\b/dcache_clean_pop/g;"\
"s/\b__clean_dcache_area_pou\b/dcache_clean_pou/g;"\
"s/\b__flush_cache_user_range\b/caches_clean_inval_user_pou/g;"\
"s/\b__flush_icache_all\b/icache_inval_all_pou/g;"
Note that __clean_dcache_area_poc is deliberately missing a word
boundary check at the beginning in order to match the efistub
symbols in image-vars.h.
Also note that, despite its name, __flush_icache_range operates
on both instruction and data caches. The name change here
reflects that.
No functional change intended.
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Fuad Tabba <tabba@google.com>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20210524083001.2586635-19-tabba@google.com
Signed-off-by: Will Deacon <will@kernel.org>
To aid with debugging, add details of the source of a panic from nVHE
hyp. This is done by having nVHE hyp exit to nvhe_hyp_panic_handler()
rather than directly to panic(). The handler will then add the extra
details for debugging before panicking the kernel.
If the panic was due to a BUG(), look up the metadata to log the file
and line, if available, otherwise log an address that can be looked up
in vmlinux. The hyp offset is also logged to allow other hyp VAs to be
converted, similar to how the kernel offset is logged during a panic.
__hyp_panic_string is now inlined since it no longer needs to be
referenced as a symbol and the message is free to diverge between VHE
and nVHE.
The following is an example of the logs generated by a BUG in nVHE hyp.
[ 46.754840] kvm [307]: nVHE hyp BUG at: arch/arm64/kvm/hyp/nvhe/switch.c:242!
[ 46.755357] kvm [307]: Hyp Offset: 0xfffea6c58e1e0000
[ 46.755824] Kernel panic - not syncing: HYP panic:
[ 46.755824] PS:400003c9 PC:0000d93a82c705ac ESR:f2000800
[ 46.755824] FAR:0000000080080000 HPFAR:0000000000800800 PAR:0000000000000000
[ 46.755824] VCPU:0000d93a880d0000
[ 46.756960] CPU: 3 PID: 307 Comm: kvm-vcpu-0 Not tainted 5.12.0-rc3-00005-gc572b99cf65b-dirty #133
[ 46.757459] Hardware name: QEMU QEMU Virtual Machine, BIOS 0.0.0 02/06/2015
[ 46.758366] Call trace:
[ 46.758601] dump_backtrace+0x0/0x1b0
[ 46.758856] show_stack+0x18/0x70
[ 46.759057] dump_stack+0xd0/0x12c
[ 46.759236] panic+0x16c/0x334
[ 46.759426] arm64_kernel_unmapped_at_el0+0x0/0x30
[ 46.759661] kvm_arch_vcpu_ioctl_run+0x134/0x750
[ 46.759936] kvm_vcpu_ioctl+0x2f0/0x970
[ 46.760156] __arm64_sys_ioctl+0xa8/0xec
[ 46.760379] el0_svc_common.constprop.0+0x60/0x120
[ 46.760627] do_el0_svc+0x24/0x90
[ 46.760766] el0_svc+0x2c/0x54
[ 46.760915] el0_sync_handler+0x1a4/0x1b0
[ 46.761146] el0_sync+0x170/0x180
[ 46.761889] SMP: stopping secondary CPUs
[ 46.762786] Kernel Offset: 0x3e1cd2820000 from 0xffff800010000000
[ 46.763142] PHYS_OFFSET: 0xffffa9f680000000
[ 46.763359] CPU features: 0x00240022,61806008
[ 46.763651] Memory Limit: none
[ 46.813867] ---[ end Kernel panic - not syncing: HYP panic:
[ 46.813867] PS:400003c9 PC:0000d93a82c705ac ESR:f2000800
[ 46.813867] FAR:0000000080080000 HPFAR:0000000000800800 PAR:0000000000000000
[ 46.813867] VCPU:0000d93a880d0000 ]---
Signed-off-by: Andrew Scull <ascull@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20210318143311.839894-6-ascull@google.com
In protected mode, late CPUs are not allowed to boot (enforced by
the PSCI relay). We can thus specialise the read_ctr macro to
always return a pre-computed, sanitised value. Special care is
taken to prevent the use of this custome version outside of
the protected mode.
Reviewed-by: Quentin Perret <qperret@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
When KVM runs in protected nVHE mode, make use of a stage 2 page-table
to give the hypervisor some control over the host memory accesses. The
host stage 2 is created lazily using large block mappings if possible,
and will default to page mappings in absence of a better solution.
>From this point on, memory accesses from the host to protected memory
regions (e.g. not 'owned' by the host) are fatal and lead to hyp_panic().
Acked-by: Will Deacon <will@kernel.org>
Signed-off-by: Quentin Perret <qperret@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20210319100146.1149909-36-qperret@google.com
When memory protection is enabled, the EL2 code needs the ability to
create and manage its own page-table. To do so, introduce a new set of
hypercalls to bootstrap a memory management system at EL2.
This leads to the following boot flow in nVHE Protected mode:
1. the host allocates memory for the hypervisor very early on, using
the memblock API;
2. the host creates a set of stage 1 page-table for EL2, installs the
EL2 vectors, and issues the __pkvm_init hypercall;
3. during __pkvm_init, the hypervisor re-creates its stage 1 page-table
and stores it in the memory pool provided by the host;
4. the hypervisor then extends its stage 1 mappings to include a
vmemmap in the EL2 VA space, hence allowing to use the buddy
allocator introduced in a previous patch;
5. the hypervisor jumps back in the idmap page, switches from the
host-provided page-table to the new one, and wraps up its
initialization by enabling the new allocator, before returning to
the host.
6. the host can free the now unused page-table created for EL2, and
will now need to issue hypercalls to make changes to the EL2 stage 1
mappings instead of modifying them directly.
Note that for the sake of simplifying the review, this patch focuses on
the hypervisor side of things. In other words, this only implements the
new hypercalls, but does not make use of them from the host yet. The
host-side changes will follow in a subsequent patch.
Credits to Will for __pkvm_init_switch_pgd.
Acked-by: Will Deacon <will@kernel.org>
Co-authored-by: Will Deacon <will@kernel.org>
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Quentin Perret <qperret@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20210319100146.1149909-18-qperret@google.com
When compiling with __KVM_NVHE_HYPERVISOR__, redefine per_cpu_offset()
to __hyp_per_cpu_offset() which looks up the base of the nVHE per-CPU
region of the given cpu and computes its offset from the
.hyp.data..percpu section.
This enables use of per_cpu_ptr() helpers in nVHE hyp code. Until now
only this_cpu_ptr() was supported by setting TPIDR_EL2.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20201202184122.26046-14-dbrazdil@google.com
MAIR_EL2 and TCR_EL2 are currently initialized from their _EL1 values.
This will not work once KVM starts intercepting PSCI ON/SUSPEND SMCs
and initializing EL2 state before EL1 state.
Obtain the EL1 values during KVM init and store them in the init params
struct. The struct will stay in memory and can be used when booting new
cores.
Take the opportunity to move copying the T0SZ value from idmap_t0sz in
KVM init rather than in .hyp.idmap.text. This avoids the need for the
idmap_t0sz symbol alias.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20201202184122.26046-12-dbrazdil@google.com
Directly using the kimage_voffset variable is fine for now, but
will become more problematic as we start distrusting EL1.
Instead, patch the kimage_voffset into the HYP text, ensuring
we don't have to load an untrusted value later on.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Now that we can use function pointer, use a dispatch table to call
the individual HVC handlers, leading to more maintainable code.
Further improvements include helpers to declare the mapping of
local variables to values passed in the host context.
Reviewed-by: Alexandru Elisei <alexandru.elisei@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
KVM/arm64 fixes for 5.10, take #1
- Force PTE mapping on device pages provided via VFIO
- Fix detection of cacheable mapping at S2
- Fallback to PMD/PTE mappings for composite huge pages
- Fix accounting of Stage-2 PGD allocation
- Fix AArch32 handling of some of the debug registers
- Simplify host HYP entry
- Fix stray pointer conversion on nVHE TLB invalidation
- Fix initialization of the nVHE code
- Simplify handling of capabilities exposed to HYP
- Nuke VCPUs caught using a forbidden AArch32 EL0
We finalize caps before initializing kvm hyp code, and any use of
cpus_have_const_cap() in kvm hyp code generates redundant and
potentially unsound code to read the cpu_hwcaps array.
A number of helper functions used in both hyp context and regular kernel
context use cpus_have_const_cap(), as some regular kernel code runs
before the capabilities are finalized. It's tedious and error-prone to
write separate copies of these for hyp and non-hyp code.
So that we can avoid the redundant code, let's automatically upgrade
cpus_have_const_cap() to cpus_have_final_cap() when used in hyp context.
With this change, there's never a reason to access to cpu_hwcaps array
from hyp code, and we don't need to create an NVHE alias for this.
This should have no effect on non-hyp code.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Acked-by: Will Deacon <will@kernel.org>
Cc: David Brazdil <dbrazdil@google.com>
Cc: Marc Zyngier <maz@kernel.org>
Cc: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20201026134931.28246-4-mark.rutland@arm.com
Pull KVM updates from Paolo Bonzini:
"For x86, there is a new alternative and (in the future) more scalable
implementation of extended page tables that does not need a reverse
map from guest physical addresses to host physical addresses.
For now it is disabled by default because it is still lacking a few of
the existing MMU's bells and whistles. However it is a very solid
piece of work and it is already available for people to hammer on it.
Other updates:
ARM:
- New page table code for both hypervisor and guest stage-2
- Introduction of a new EL2-private host context
- Allow EL2 to have its own private per-CPU variables
- Support of PMU event filtering
- Complete rework of the Spectre mitigation
PPC:
- Fix for running nested guests with in-kernel IRQ chip
- Fix race condition causing occasional host hard lockup
- Minor cleanups and bugfixes
x86:
- allow trapping unknown MSRs to userspace
- allow userspace to force #GP on specific MSRs
- INVPCID support on AMD
- nested AMD cleanup, on demand allocation of nested SVM state
- hide PV MSRs and hypercalls for features not enabled in CPUID
- new test for MSR_IA32_TSC writes from host and guest
- cleanups: MMU, CPUID, shared MSRs
- LAPIC latency optimizations ad bugfixes"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (232 commits)
kvm: x86/mmu: NX largepage recovery for TDP MMU
kvm: x86/mmu: Don't clear write flooding count for direct roots
kvm: x86/mmu: Support MMIO in the TDP MMU
kvm: x86/mmu: Support write protection for nesting in tdp MMU
kvm: x86/mmu: Support disabling dirty logging for the tdp MMU
kvm: x86/mmu: Support dirty logging for the TDP MMU
kvm: x86/mmu: Support changed pte notifier in tdp MMU
kvm: x86/mmu: Add access tracking for tdp_mmu
kvm: x86/mmu: Support invalidate range MMU notifier for TDP MMU
kvm: x86/mmu: Allocate struct kvm_mmu_pages for all pages in TDP MMU
kvm: x86/mmu: Add TDP MMU PF handler
kvm: x86/mmu: Remove disallowed_hugepage_adjust shadow_walk_iterator arg
kvm: x86/mmu: Support zapping SPTEs in the TDP MMU
KVM: Cache as_id in kvm_memory_slot
kvm: x86/mmu: Add functions to handle changed TDP SPTEs
kvm: x86/mmu: Allocate and free TDP MMU roots
kvm: x86/mmu: Init / Uninit the TDP MMU
kvm: x86/mmu: Introduce tdp_iter
KVM: mmu: extract spte.h and spte.c
KVM: mmu: Separate updating a PTE from kvm_set_pte_rmapp
...
Pull irq updates from Thomas Gleixner:
"Updates for the interrupt subsystem:
Core:
- Allow trimming of interrupt hierarchy to support odd hardware
setups where only a subset of the interrupts requires the full
hierarchy.
- Allow the retrigger mechanism to follow a hierarchy to simplify
driver code.
- Provide a mechanism to force enable wakeup interrrupts on suspend.
- More infrastructure to handle IPIs in the core code
Architectures:
- Convert ARM/ARM64 IPI handling to utilize the interrupt core code.
Drivers:
- The usual pile of new interrupt chips (MStar, Actions Owl, TI
PRUSS, Designware ICTL)
- ARM(64) IPI related conversions
- Wakeup support for Qualcom PDC
- Prevent hierarchy corruption in the NVIDIA Tegra driver
- The usual small fixes, improvements and cleanups all over the
place"
* tag 'irq-core-2020-10-12' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (59 commits)
dt-bindings: interrupt-controller: Add MStar interrupt controller
irqchip/irq-mst: Add MStar interrupt controller support
soc/tegra: pmc: Don't create fake interrupt hierarchy levels
soc/tegra: pmc: Allow optional irq parent callbacks
gpio: tegra186: Allow optional irq parent callbacks
genirq/irqdomain: Allow partial trimming of irq_data hierarchy
irqchip/qcom-pdc: Reset PDC interrupts during init
irqchip/qcom-pdc: Set IRQCHIP_ENABLE_WAKEUP_ON_SUSPEND flag
pinctrl: qcom: Set IRQCHIP_ENABLE_WAKEUP_ON_SUSPEND flag
genirq/PM: Introduce IRQCHIP_ENABLE_WAKEUP_ON_SUSPEND flag
pinctrl: qcom: Use return value from irq_set_wake() call
pinctrl: qcom: Set IRQCHIP_SET_TYPE_MASKED and IRQCHIP_MASK_ON_SUSPEND flags
ARM: Handle no IPI being registered in show_ipi_list()
MAINTAINERS: Add entries for Actions Semi Owl SIRQ controller
irqchip: Add Actions Semi Owl SIRQ controller
dt-bindings: interrupt-controller: Add Actions SIRQ controller binding
dt-bindings: dw-apb-ictl: Update binding to describe use as primary interrupt controller
irqchip/dw-apb-ictl: Add primary interrupt controller support
irqchip/dw-apb-ictl: Refactor priot to introducing hierarchical irq domains
genirq: Add stub for set_handle_irq() when !GENERIC_IRQ_MULTI_HANDLER
...
Hyp keeps track of which cores require SSBD callback by accessing a
kernel-proper global variable. Create an nVHE symbol of the same name
and copy the value from kernel proper to nVHE as KVM is being enabled
on a core.
Done in preparation for separating percpu memory owned by kernel
proper and nVHE.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Acked-by: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20200922204910.7265-8-dbrazdil@google.com
Owing to the fact that the host kernel is always mitigated, we can
drastically simplify the WA2 handling by keeping the mitigation
state ON when entering the guest. This means the guest is either
unaffected or not mitigated.
This results in a nice simplification of the mitigation space,
and the removal of a lot of code that was never really used anyway.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Signed-off-by: Will Deacon <will@kernel.org>
The GIC's internal view of the priority mask register and the assigned
interrupt priorities are based on whether GIC security is enabled and
whether firmware routes Group 0 interrupts to EL3. At the moment, we
support priority masking when ICC_PMR_EL1 and interrupt priorities are
either both modified by the GIC, or both left unchanged.
Trusted Firmware-A's default interrupt routing model allows Group 0
interrupts to be delivered to the non-secure world (SCR_EL3.FIQ == 0).
Unfortunately, this is precisely the case that the GIC driver doesn't
support: ICC_PMR_EL1 remains unchanged, but the GIC's view of interrupt
priorities is different from the software programmed values.
Support pseudo-NMIs when SCR_EL3.FIQ == 0 by using a different value to
mask regular interrupts. All the other values remain the same.
Signed-off-by: Alexandru Elisei <alexandru.elisei@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200912153707.667731-3-alexandru.elisei@arm.com
KVM has a one instruction window where it will allow an SError exception
to be consumed by the hypervisor without treating it as a hypervisor bug.
This is used to consume asynchronous external abort that were caused by
the guest.
As we are about to add another location that survives unexpected exceptions,
generalise this code to make it behave like the host's extable.
KVM's version has to be mapped to EL2 to be accessible on nVHE systems.
The SError vaxorcism code is a one instruction window, so has two entries
in the extable. Because the KVM code is copied for VHE and nVHE, we end up
with four entries, half of which correspond with code that isn't mapped.
Signed-off-by: James Morse <james.morse@arm.com>
Reviewed-by: Marc Zyngier <maz@kernel.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
The following files in hyp/ contain only code shared by VHE/nVHE:
vgic-v3-sr.c, aarch32.c, vgic-v2-cpuif-proxy.c, entry.S, fpsimd.S
Compile them under both configurations. Deletions in image-vars.h reflect
eliminated dependencies of nVHE code on the rest of the kernel.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200625131420.71444-14-dbrazdil@google.com
timer-sr.c contains a HVC handler for setting CNTVOFF_EL2 and two helper
functions for controlling access to physical counter. The former is used by
both VHE/nVHE and is duplicated, the latter are used only by nVHE and moved
to nvhe/timer-sr.c.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200625131420.71444-13-dbrazdil@google.com
sysreg-sr.c contains KVM's code for saving/restoring system registers, with
some code shared between VHE/nVHE. These common routines are moved to
a header file, VHE-specific code is moved to vhe/sysreg-sr.c and nVHE-specific
code to nvhe/sysreg-sr.c.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200625131420.71444-12-dbrazdil@google.com
debug-sr.c contains KVM's code for context-switching debug registers, with some
code shared between VHE/nVHE. These common routines are moved to a header file,
VHE-specific code is moved to vhe/debug-sr.c and nVHE-specific code to
nvhe/debug-sr.c.
Functions are slightly refactored to move code hidden behind `has_vhe()` checks
to the corresponding .c files.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200625131420.71444-11-dbrazdil@google.com
