This patch series prevents possible collisions in the chain_key hashing macro iterate_chain_key(key1, key2) that can lead to lockdep not detecting very simple deadlocks such as AA or ABBA.
The problem only affects the first allocated lock classes. That could explain why it was not seen while running lockdep's test suite, since by the time the test suite runs there are already registered lock classes and the indexes allocated for the lock classes under test are high enough to avoid collisions. The patch series also extends the tools/liblockdep test suite with tests covering the offending cases. I came across the problem while testing a simple AA deadlock scenario in userspace using a pthread_mutex and tools/liblockdep. In that context it is fairly easy to have a clean and deterministic initial state where the problem can be reproduced. The proposed solution was tested with the newly introduced tests and also with lockdep's test suite: [ 0.000000] Good, all 253 testcases passed! | Alfredo Alvarez Fernandez (3): tools/liblockdep: add userspace version of READ_ONCE tools/liblockdep: add tests lockdep: prevent chain_key collisions kernel/locking/lockdep.c | 14 ++++------ tools/lib/lockdep/tests/AA.c | 8 +++--- tools/lib/lockdep/tests/ABA.c | 13 +++++++++ tools/lib/lockdep/tests/ABBA_2threads.c | 43 +++++++++++++++++++++++++++++ tools/lib/lockdep/uinclude/linux/compiler.h | 1 + 5 files changed, 67 insertions(+), 12 deletions(-) create mode 100644 tools/lib/lockdep/tests/ABA.c create mode 100644 tools/lib/lockdep/tests/ABBA_2threads.c -- 2.5.0
