Re: [lttng-dev] [PATCH v2 00/12] Add support for TSAN to liburcu
Ondřej Surý via lttng-dev <[email protected]>
| Newsgroups | org.lttng.lists.lttng-dev |
|---|---|
| Message-ID | <[email protected]> |
Olivier, is this somewhere in Gerrit again, please? It’s much easier for us to pull the changes from git than cherry-pick patches from the mailing list. Ondřej -- Ondřej Surý <[email protected]> (He/Him) > On 7. 6. 2023, at 20:54, Olivier Dion <[email protected]> wrote: > > This patch set adds support for TSAN in liburcu. > > * Change since v1 > > ** Adding CMM_SEQ_CST_FENCE memory order to the CMM memory model > > The C11 memory model is incompatible with the memory model used by liburcu, > since the semantic of the C11 memory model is based on a happen-before > (acquire/release) relationship of memory accesses, while liburcu is based on > memory barriers and relaxed memory accesses. > > To circumvent this, a new memory order called CMM_SEQ_CST_FENCE is > introduced. It implies a CMM_SEQ_CST while emitting a thread fence after the > operation. Operations that were documented as emitting memory barriers > before and after the operation are implemented now in term of this new memory > order to keep compatibility. > > However, this model is redundant in some cases and the memory orders were > changed internally in liburcu to simply use the CMM_SEQ_CST. > > ** Adding cmm_emit_legacy_smp_mb() to queue/stack APIs > > The queue/stack APIs document implicit memory barriers before or after > operation. These are now optionally emitted only if > CONFIG_RCU_EMIT_LEGACY_MB is defined in urcu/config.h or manually defined by > the user before including liburcu. That way, users can opt-in even if the > system headers were configured without this feature. > > However, users can not opt-out of that feature if configured by the system. > > * v1 > > ** Here are the major changes > > - Usage of compiler atomic builtins is added to the uatomic API. This is > required for TSAN to understand atomic memory accesses. If the compiler > supports such builtins, they are used by default. User can opt-out and use > the legacy implementation of the uatomic API by using the > `--disable-atomic-builtins' configuration option. > > - The CMM memory model is introduced but yet formalized. It tries to be as > close as possible to the C11 memory model while offering primitives such as > cmm_smp_wmb(), cmm_smp_rmb() and cmm_mb() that can't be expressed in it. > For example, cmm_mb() can be used for ordering memory accesses to MMIO > devices, which is out of the scope of the C11 memory model. > > - The CMM annotation layer is a new public API that is highly experimental and > not guaranteed to be stable at this stage. It serves the dual purpose of > verifying local (intra-thread) relaxed atomic accesses ordering with a > memory barrier and global (inter-thread) relaxed atomic accesses with a > shared state. The second purpose is necessary for TSAN to understand memory > accesses ordering since it does not fully support thread fence yet. > > ** CMM annotation example > > Consider the following pseudo-code of writer side in synchronize_rcu(). An > acquire group is defined on the stack of the writer. Annotations are made > onto the group to ensure ordering of relaxed memory accesses in reader_state() > before the memory barrier at the end of synchronize_rcu(). It also helps TSAN > to understand that the relaxed accesses in reader_state() act like acquire > accesses because of the memory barrier in synchronize_rcu(). > > In other words, the purpose of this annotation is to convert a group of > load-acquire memory operations into load-relaxed memory operations followed by > a single memory barrier. This highly benefits weakly ordered architectures by > having a constant number of memory barriers instead of being linearly > proportional to the number of loads. This does not benefit TSO > architectures. > > > Olivier Dion (12): > configure: Add --disable-atomic-builtins option > urcu/compiler: Use atomic builtins if configured > urcu/arch/generic: Use atomic builtins if configured > urcu/system: Use atomic builtins if configured > urcu/uatomic: Add CMM memory model > urcu-wait: Fix wait state load/store > tests: Use uatomic for accessing global states > benchmark: Use uatomic for accessing global states > tests/unit/test_build: Quiet unused return value > urcu/annotate: Add CMM annotation > Add cmm_emit_legacy_smp_mb() > tests: Add tests for checking race conditions > > README.md | 11 ++ > configure.ac | 39 ++++ > doc/uatomic-api.md | 3 +- > include/Makefile.am | 3 + > include/urcu/annotate.h | 174 ++++++++++++++++++ > include/urcu/arch.h | 6 + > include/urcu/arch/generic.h | 37 ++++ > include/urcu/compiler.h | 22 ++- > include/urcu/config.h.in | 6 + > include/urcu/static/lfstack.h | 25 ++- > include/urcu/static/pointer.h | 40 ++-- > include/urcu/static/rculfqueue.h | 14 +- > include/urcu/static/rculfstack.h | 8 +- > include/urcu/static/urcu-bp.h | 12 +- > include/urcu/static/urcu-common.h | 8 +- > include/urcu/static/urcu-mb.h | 11 +- > include/urcu/static/urcu-memb.h | 26 ++- > include/urcu/static/urcu-qsbr.h | 29 ++- > include/urcu/static/wfcqueue.h | 68 +++---- > include/urcu/static/wfqueue.h | 9 +- > include/urcu/static/wfstack.h | 24 ++- > include/urcu/system.h | 21 +++ > include/urcu/uatomic.h | 63 ++++++- > include/urcu/uatomic/builtins-generic.h | 170 +++++++++++++++++ > include/urcu/uatomic/builtins.h | 79 ++++++++ > include/urcu/uatomic/generic.h | 234 ++++++++++++++++++++++++ > src/rculfhash.c | 92 ++++++---- > src/urcu-bp.c | 17 +- > src/urcu-pointer.c | 9 +- > src/urcu-qsbr.c | 31 +++- > src/urcu-wait.h | 15 +- > src/urcu.c | 24 ++- > tests/benchmark/Makefile.am | 91 ++++----- > tests/benchmark/common-states.c | 1 + > tests/benchmark/common-states.h | 51 ++++++ > tests/benchmark/test_mutex.c | 32 +--- > tests/benchmark/test_perthreadlock.c | 32 +--- > tests/benchmark/test_rwlock.c | 32 +--- > tests/benchmark/test_urcu.c | 33 +--- > tests/benchmark/test_urcu_assign.c | 33 +--- > tests/benchmark/test_urcu_bp.c | 33 +--- > tests/benchmark/test_urcu_defer.c | 33 +--- > tests/benchmark/test_urcu_gc.c | 34 +--- > tests/benchmark/test_urcu_hash.c | 6 +- > tests/benchmark/test_urcu_hash.h | 15 -- > tests/benchmark/test_urcu_hash_rw.c | 10 +- > tests/benchmark/test_urcu_hash_unique.c | 10 +- > tests/benchmark/test_urcu_lfq.c | 20 +- > tests/benchmark/test_urcu_lfs.c | 20 +- > tests/benchmark/test_urcu_lfs_rcu.c | 20 +- > tests/benchmark/test_urcu_qsbr.c | 33 +--- > tests/benchmark/test_urcu_qsbr_gc.c | 34 +--- > tests/benchmark/test_urcu_wfcq.c | 22 +-- > tests/benchmark/test_urcu_wfq.c | 20 +- > tests/benchmark/test_urcu_wfs.c | 22 +-- > tests/common/api.h | 12 +- > tests/regression/rcutorture.h | 106 +++++++---- > tests/unit/test_build.c | 8 +- > tests/unit/test_lfstack.c | 90 +++++++++ > tests/unit/test_wfcqueue.c | 119 ++++++++++++ > tests/unit/test_wfqueue.c | 91 +++++++++ > tests/unit/test_wfstack.c | 90 +++++++++ > 62 files changed, 1799 insertions(+), 684 deletions(-) > create mode 100644 include/urcu/annotate.h > create mode 100644 include/urcu/uatomic/builtins-generic.h > create mode 100644 include/urcu/uatomic/builtins.h > create mode 100644 tests/benchmark/common-states.c > create mode 100644 tests/benchmark/common-states.h > create mode 100644 tests/unit/test_lfstack.c > create mode 100644 tests/unit/test_wfcqueue.c > create mode 100644 tests/unit/test_wfqueue.c > create mode 100644 tests/unit/test_wfstack.c > > -- > 2.40.1 > _______________________________________________ lttng-dev mailing list [email protected] https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev