[demrom]

Depends on #192; until it lands this diff also shows its commits.

When commands run from several threads, ThreadSanitizer reports three data races, all of them real:

• executeWebAssemblyCommandsInOrder caches the result of dlsym in a static function pointer that command threads write and read without synchronization. It is now _Atomic, and the lookup still falls back to RTLD_DEFAULT.
• The session's global_errno is written by every command thread and read by the caller of ios_system. It is now _Atomic int.
• The session's mainThreadId is written by ios_system when it starts the first command of a session, and read and cleared by cleanup_function on the command thread. It is now _Atomic(pthread_t). The command thread stores its own id in run_function (a new storeMainThread flag), with atomic_compare_exchange_strong only while the field is empty, so a command that ends before ios_system returns is never recorded after it ended. cleanup_function clears it with atomic_compare_exchange_strong only while it still holds its own thread, and the window-resize code loads it once before pthread_kill.

TSan also reports ios_nextAvailablePid and ios_storeThreadId: the first returns with pid_mtx locked and the second unlocks it on the new command thread, which TSan cannot follow, so it reports two threads holding the mutex at once. That locking stays as it is. tests/tsan.supp suppresses the accesses inside that window (ios_nextAvailablePid, storeEnvironment, environmentVariables, resetEnvironment), ios_restoreParentPid, whose compare-and-swap on the release path races with the plain writes in that window, and ios_currentPid. The ios_currentPid entry also hides the unlocked reads of current_pid against the plain writes in ios_nextAvailablePid, which this change leaves alone. tests/README.md explains how to build with ENABLE_THREAD_SANITIZER=YES and run with the file.

Test

tests/concurrent-fork.m, linked with -fsanitize=thread against a TSan build of the framework from a fresh clone of this branch and run with suppressions=tests/tsan.supp on the iOS 27 Simulator:

• master + Hand the current pid back to the parent only when the ending command still owns it #192: 3 to 4 WARNING: ThreadSanitizer reports per run in 10 of 10 runs, on executeWebAssemblyCommandsInOrder (the cached lookup) and run_function (the global_errno write).
• With this change: 0 reports in 10 of 10 runs, and the test prints 8 threads x 400 commands PASS.

tests/main-thread-winch.m covers the stale id: a short command ends before ios_system returns, then a long command installs a SIGWINCH handler and the test calls ios_setWindowSize and checks that the handler runs on the long command's thread. On master + #192, 5 of 5 runs fail with SIGWINCH was handled on another thread; with this change 10 of 10 runs pass.

With the other four changes of the series merged, TSan also reports the mainThreadId write ios_system makes after creating the command (in __ios_system_block_invoke) against the clear in cleanup_function. Without the mainThreadId part of this change, concurrent-fork reports it in 5 of 10 runs and concurrent-getopt in 1 of 10, one report each; with it, 0 reports in 10 of 10 runs of each.

On an iPhone, in a-Shell mini 2.2.3 (592) built from a-shell master, with an ios_system built from all five changes of this series merged, 120 pipelines of cat | sort -r -u -n | uniq -c | head -n 3 ran one after another from dash, each with the expected output. The races are timing-dependent, so the Simulator TSan runs above are what show them fixed.

[holzschu]

Thanks a lot for that one.