ivmai / bdwgc / 1300

69%

master: 77%

push

travis-ci-com

Fix hang on sem_wait in GC_suspend_thread if thread was resumed recently
(a cherry-pick of commit 0ab51e458 from 'release-8_0')

If GC_resume_thread(t) is followed by GC_suspend_thread(t) immediately,
then there is a risk of the loop (which had been entered before
GC_resume_thread(t) is called) in GC_suspend_self_inner() continue
iterating, thus the suspend signal handler stays pending, and thus
the relevant sem_post() will not be called.

The solution is to replace a boolean flag (suspended_ext) with
a counter (thread_stop_info.ext_suspend_cnt) with the lowest bit
indicating whether the thread is suspended manually.

* include/private/pthread_stop_world.h [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (thread_stop_info.ext_suspend_cnt): Add
field; add comment.
* include/private/pthread_support.h [GC_ENABLE_SUSPEND_THREAD
&& !GC_DARWIN_THREADS && !GC_OPENBSD_UTHREADS && !NACL]
(GC_thread.suspended_ext): Remove field.
* include/private/pthread_support.h [GC_ENABLE_SUSPEND_THREAD
&& !GC_DARWIN_THREADS && !GC_OPENBSD_UTHREADS && !NACL]
(GC_suspend_self_inner): Add suspend_cnt argument.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (GC_suspend_self_inner): Likewise.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (GC_suspend_handler_inner): Declare
suspend_cnt local variable, set it from me->stop_info.ext_suspend_cnt,
pass it GC_suspend_self_inner().
* pthread_support.c [GC_ENABLE_SUSPEND_THREAD && !GC_DARWIN_THREADS
&& !GC_OPENBSD_UTHREADS && !NACL] (GC_do_blocking_inner,
GC_suspend_self_blocked, GC_call_with_gc_active): Likewise.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (GC_suspend_handler_inner): Replace
me->suspended_ext with suspend_cnt&1.
* pthread_stop_world.c [!GC_OPENBSD_UTHREADS && !NACL
&& GC_ENABLE_SUSPEND_THREAD] (GC_suspend_self_inner): Add assertion
that suspend_cnt is odd; replace me->suspended_ext loop condition t... (continued)

32 of 32 new or added lines in 2 files covered. (100.0%)

6107 of 8855 relevant lines covered (68.97%)

1334767.09 hits per line