[PATCH v3 00/25] kcsan: Support detecting a subset of missing memory barriers

From: Marco Elver
Date: Tue Nov 30 2021 - 06:45:13 EST


Detection of some missing memory barriers has been on the KCSAN feature
wishlist for some time: this series adds support for modeling a subset
of weak memory as defined by the LKMM, which enables detection of a
subset of data races due to missing memory barriers.

KCSAN's approach to detecting missing memory barriers is based on
modeling access reordering. Each memory access for which a watchpoint is
set up, is also selected for simulated reordering within the scope of
its function (at most 1 in-flight access).

We are limited to modeling the effects of "buffering" (delaying the
access), since the runtime cannot "prefetch" accesses. Once an access
has been selected for reordering, it is checked along every other access
until the end of the function scope. If an appropriate memory barrier is
encountered, the access will no longer be considered for reordering.

When the result of a memory operation should be ordered by a barrier,
KCSAN can then detect data races where the conflict only occurs as a
result of a missing barrier due to reordering accesses.

Some more details and an example are captured in the updated
<Documentation/dev-tools/kcsan.rst>.

Some light fuzzing with the feature also resulted in a discussion [1]
around an issue which appears to be allowed, but unlikely in practice.

[1] https://lkml.kernel.org/r/YRo58c+JGOvec7tc@xxxxxxxxxxxxxxxx


The first half of the series are core KCSAN changes, documentation
updates, and test changes. The second half adds instrumentation to
barriers, atomics, bitops, along with enabling barrier instrumentation
for some currently uninstrumented subsystems.

Followed by objtool changes to add the usual entries to the uaccess
whitelist, but also instruct objtool to remove memory barrier
instrumentation from noinstr code (on x86), given not all versions of
Clang currently respect __no_kcsan (== __no_sanitize_thread) for the new
instrumentation.

The last 2 patches (new in v3) fix up __no_kcsan for newer versions of
Clang, so that non-x86 architectures can enable weak memory modeling
with Clang 14.0 or newer.

Changelog
---------

v3:
* Rework to avoid kcsan_noinstr hackery, because it is unclear if
this works on architectures like arm64. A better alternative exists
where we can get __no_kcsan to work for barrier instrumentation, too.
Clang's and GCC's __no_kcsan (== __no_sanitize_thread) behave slightly
differently, which is reflected in KCSAN_WEAK_MEMORY's dependencies
(either STACK_VALIDATION for older Clang, or GCC which works as-is).
* Rework to avoid inserting explicit calls for barrier instrumentation,
and instead repurpose __atomic_signal_fence (see comment at
__tsan_atomic_signal_fence), which is handled by fsanitize=thread
instrumentation and can therefore be removed via __no_kcsan.
* objtool: s/removable_instr/profiling_func/, and more comments per
Josh's suggestion.
* Minimize diff in patch removing zero-initialization of globals.
* Don't define kcsan_weak_memory bool if !KCSAN_WEAK_MEMORY.
* Apply Acks.
* 2 new patches to make it work with Clang >= 14.0 without objtool,
which will be required on non-x86 architectures.

v2: https://lkml.kernel.org/r/20211118081027.3175699-1-elver@xxxxxxxxxx
* Rewrite objtool patch after rebase to v5.16-rc1.
* Note the reason in documentation that address or control dependencies
do not require special handling.
* Rename kcsan_atomic_release() to kcsan_atomic_builtin_memorder() to
avoid confusion.
* Define kcsan_noinstr as noinline if we rely on objtool nop'ing out
calls, to avoid things like LTO inlining it.

v1: https://lore.kernel.org/all/20211005105905.1994700-1-elver@xxxxxxxxxx/
---

Alexander Potapenko (1):
compiler_attributes.h: Add __disable_sanitizer_instrumentation

Marco Elver (24):
kcsan: Refactor reading of instrumented memory
kcsan: Remove redundant zero-initialization of globals
kcsan: Avoid checking scoped accesses from nested contexts
kcsan: Add core support for a subset of weak memory modeling
kcsan: Add core memory barrier instrumentation functions
kcsan, kbuild: Add option for barrier instrumentation only
kcsan: Call scoped accesses reordered in reports
kcsan: Show location access was reordered to
kcsan: Document modeling of weak memory
kcsan: test: Match reordered or normal accesses
kcsan: test: Add test cases for memory barrier instrumentation
kcsan: Ignore GCC 11+ warnings about TSan runtime support
kcsan: selftest: Add test case to check memory barrier instrumentation
locking/barriers, kcsan: Add instrumentation for barriers
locking/barriers, kcsan: Support generic instrumentation
locking/atomics, kcsan: Add instrumentation for barriers
asm-generic/bitops, kcsan: Add instrumentation for barriers
x86/barriers, kcsan: Use generic instrumentation for non-smp barriers
x86/qspinlock, kcsan: Instrument barrier of pv_queued_spin_unlock()
mm, kcsan: Enable barrier instrumentation
sched, kcsan: Enable memory barrier instrumentation
objtool, kcsan: Add memory barrier instrumentation to whitelist
objtool, kcsan: Remove memory barrier instrumentation from noinstr
kcsan: Support WEAK_MEMORY with Clang where no objtool support exists

Documentation/dev-tools/kcsan.rst | 76 +++-
arch/x86/include/asm/barrier.h | 10 +-
arch/x86/include/asm/qspinlock.h | 1 +
include/asm-generic/barrier.h | 54 ++-
.../asm-generic/bitops/instrumented-atomic.h | 3 +
.../asm-generic/bitops/instrumented-lock.h | 3 +
include/linux/atomic/atomic-instrumented.h | 135 +++++-
include/linux/compiler_attributes.h | 18 +
include/linux/compiler_types.h | 13 +-
include/linux/kcsan-checks.h | 81 +++-
include/linux/kcsan.h | 11 +-
include/linux/sched.h | 3 +
include/linux/spinlock.h | 2 +-
init/init_task.c | 5 -
kernel/kcsan/Makefile | 2 +
kernel/kcsan/core.c | 345 ++++++++++++---
kernel/kcsan/kcsan_test.c | 415 ++++++++++++++++--
kernel/kcsan/report.c | 51 ++-
kernel/kcsan/selftest.c | 141 ++++++
kernel/sched/Makefile | 7 +-
lib/Kconfig.kcsan | 20 +
mm/Makefile | 2 +
scripts/Makefile.kcsan | 15 +-
scripts/Makefile.lib | 5 +
scripts/atomic/gen-atomic-instrumented.sh | 41 +-
tools/objtool/check.c | 41 +-
tools/objtool/include/objtool/elf.h | 2 +-
27 files changed, 1330 insertions(+), 172 deletions(-)

--
2.34.0.rc2.393.gf8c9666880-goog