diff options
author | teor <teor@torproject.org> | 2019-09-20 11:27:05 +1000 |
---|---|---|
committer | teor <teor@torproject.org> | 2019-09-26 12:37:25 +1000 |
commit | d1eab05834566f998721d3a16107767885711c57 (patch) | |
tree | 4224a8adc04fd130a8794f18e3bd8d54c82d87e8 /src/lib/lock | |
parent | 02840169d860384257042bdf6d7601c2bf48b47b (diff) | |
download | tor-d1eab05834566f998721d3a16107767885711c57.tar.gz tor-d1eab05834566f998721d3a16107767885711c57.zip |
lock: Avoid some undefined behaviour when freeing mutexes.
Fixes bug 31736; bugfix on 0.0.7.
Diffstat (limited to 'src/lib/lock')
-rw-r--r-- | src/lib/lock/compat_mutex.c | 10 | ||||
-rw-r--r-- | src/lib/lock/compat_mutex_pthreads.c | 16 |
2 files changed, 24 insertions, 2 deletions
diff --git a/src/lib/lock/compat_mutex.c b/src/lib/lock/compat_mutex.c index 4ad5929715..670bd0174c 100644 --- a/src/lib/lock/compat_mutex.c +++ b/src/lib/lock/compat_mutex.c @@ -29,7 +29,15 @@ tor_mutex_new_nonrecursive(void) tor_mutex_init_nonrecursive(m); return m; } -/** Release all storage and system resources held by <b>m</b>. */ +/** Release all storage and system resources held by <b>m</b>. + * + * Destroying a locked mutex is undefined behaviour. Global mutexes may be + * locked when they are passed to this function, because multiple threads can + * still access them. So we can either: + * - destroy on shutdown, and re-initialise when tor re-initialises, or + * - skip destroying and re-initialisation, using a sentinel variable. + * See #31735 for details. + */ void tor_mutex_free_(tor_mutex_t *m) { diff --git a/src/lib/lock/compat_mutex_pthreads.c b/src/lib/lock/compat_mutex_pthreads.c index ee5f520cd0..f82ad9f0e8 100644 --- a/src/lib/lock/compat_mutex_pthreads.c +++ b/src/lib/lock/compat_mutex_pthreads.c @@ -88,12 +88,26 @@ tor_mutex_release(tor_mutex_t *m) } /** Clean up the mutex <b>m</b> so that it no longer uses any system * resources. Does not free <b>m</b>. This function must only be called on - * mutexes from tor_mutex_init(). */ + * mutexes from tor_mutex_init(). + * + * Destroying a locked mutex is undefined behaviour. Global mutexes may be + * locked when they are passed to this function, because multiple threads can + * still access them. So we can either: + * - destroy on shutdown, and re-initialise when tor re-initialises, or + * - skip destroying and re-initialisation, using a sentinel variable. + * See #31735 for details. + */ void tor_mutex_uninit(tor_mutex_t *m) { int err; raw_assert(m); + /* If the mutex is already locked, wait until after it is unlocked to destroy + * it. Locking and releasing the mutex makes undefined behaviour less likely, + * but does not prevent it. Another thread can lock the mutex between release + * and destroy. */ + tor_mutex_acquire(m); + tor_mutex_release(m); err = pthread_mutex_destroy(&m->mutex); if (PREDICT_UNLIKELY(err)) { // LCOV_EXCL_START |