Re: [PATCH 0/9] tools: Make {clear,set}_bit() atomic for reals

From: Paolo Bonzini
Date: Fri Dec 02 2022 - 13:24:21 EST


On 11/19/22 02:34, Sean Christopherson wrote:
For obvious reasons I'd like to route the this through Paolo's tree.
In theory, taking just patch 5 through tip would work, but creating a
topic branch seems like the way to go, though maybe I'm being overly
paranoid. The current tip/perf/core doesn't have any conflicts, nor does
it have new set_bit() or clear_bit() users.

Code sitting in kvm/queue for 6.2 adds functionality that relies on
clear_bit() being an atomic operation. Unfortunately, despite being
implemented in atomic.h (among other strong hits that they should be
atomic), clear_bit() and set_bit() aren't actually atomic (and of course
I realized this _just_ after everything got queued up).

Move current tools/ users of clear_bit() and set_bit() to the
double-underscore versions (which tools/ already provides and documents
as being non-atomic), and then implement clear_bit() and set_bit() as
actual atomics to fix the KVM selftests bug.

Perf and KVM are the only affected users. NVDIMM also has test code
in tools/, but that builds against the kernel proper. The KVM code is
well tested and fully audited. The perf code is lightly tested; if I
understand the build system, it's probably not even fully compile tested.

Patches 1 and 2 are completely unrelated and are fixes for patches
sitting in kvm/queue. Paolo, they can be squashed if you want to rewrite
history.

Patch 3 fixes a hilarious collision in a KVM ARM selftest that will arise
when clear_bit() is converted to an atomic.

Patch 4 changes clear_bit() and set_bit() to take an "unsigned long"
instead of an "int" so that patches 5-6 aren't accompanied by functional
changes. I.e. if something in perf is somehow relying on "bit" being a
signed int, failures will bisect to patch 4 and not to the
supposed-to-be-a-nop conversion to __clear_bit() and __set_bit().

Patch 5-9 switch perf+KVM and complete the conversion.

Applies on:
git://git.kernel.org/pub/scm/virt/kvm/kvm.git queue

Queued, thanks Namhyung for the ACK!

Paolo


Sean Christopherson (9):
KVM: selftests: Add rdmsr_from_l2() implementation in Hyper-V eVMCS
test
KVM: selftests: Remove unused "vcpu" param to fix build error
KVM: arm64: selftests: Enable single-step without a "full" ucall()
tools: Take @bit as an "unsigned long" in {clear,set}_bit() helpers
perf tools: Use dedicated non-atomic clear/set bit helpers
KVM: selftests: Use non-atomic clear/set bit helpers in KVM tests
tools: Drop conflicting non-atomic test_and_{clear,set}_bit() helpers
tools: Drop "atomic_" prefix from atomic test_and_set_bit()
tools: KVM: selftests: Convert clear/set_bit() to actual atomics

tools/arch/x86/include/asm/atomic.h | 6 +++-
tools/include/asm-generic/atomic-gcc.h | 13 ++++++-
tools/include/asm-generic/bitops/atomic.h | 15 ++++----
tools/include/linux/bitmap.h | 34 -------------------
tools/perf/bench/find-bit-bench.c | 2 +-
tools/perf/builtin-c2c.c | 6 ++--
tools/perf/builtin-kwork.c | 6 ++--
tools/perf/builtin-record.c | 6 ++--
tools/perf/builtin-sched.c | 2 +-
tools/perf/tests/bitmap.c | 2 +-
tools/perf/tests/mem2node.c | 2 +-
tools/perf/util/affinity.c | 4 +--
tools/perf/util/header.c | 8 ++---
tools/perf/util/mmap.c | 6 ++--
tools/perf/util/pmu.c | 2 +-
.../util/scripting-engines/trace-event-perl.c | 2 +-
.../scripting-engines/trace-event-python.c | 2 +-
tools/perf/util/session.c | 2 +-
tools/perf/util/svghelper.c | 2 +-
.../selftests/kvm/aarch64/arch_timer.c | 2 +-
.../selftests/kvm/aarch64/debug-exceptions.c | 21 ++++++------
tools/testing/selftests/kvm/dirty_log_test.c | 34 +++++++++----------
.../selftests/kvm/include/ucall_common.h | 8 +++++
.../testing/selftests/kvm/lib/ucall_common.c | 2 +-
.../selftests/kvm/x86_64/hyperv_evmcs.c | 13 +++++--
.../selftests/kvm/x86_64/hyperv_svm_test.c | 4 +--
.../selftests/kvm/x86_64/hyperv_tlb_flush.c | 2 +-
27 files changed, 102 insertions(+), 106 deletions(-)


base-commit: 3321eef4acb51c303f0598d8a8493ca58528a054