Fix potential race if start_mark_threads called from threads in child

There should be no race in setting GC_markers_m1 value even in case of
a client calls GC_start_mark_threads() from multiple threads in a child
process.

* include/gc/gc.h (GC_start_mark_threads): Add comment that the
allocation lock is acquired internally.
* mark.c [PARALLEL_MARK] (GC_wait_for_markers_init): Change assertion
from I_DONT_HOLD_LOCK to I_HOLD_LOCK.
* pthread_support.c [PARALLEL_MARK] (GC_start_mark_threads_inner):
Likewise.
* win32_threads.c [PARALLEL_MARK] (GC_start_mark_threads_inner):
Likewise.
* mark.c [PARALLEL_MARK] (GC_wait_for_markers_init): Remove LOCK/UNLOCK
around GC_add_to_our_memory() call.
* misc.c [PARALLEL_MARK && GC_ASSERTIONS && GC_ALWAYS_MULTITHREADED]
(GC_init): Remove UNLOCK/LOCK around GC_start_mark_threads_inner call.
* misc.c [THREADS && PARALLEL_MARK && CAN_HANDLE_FORK
&& !THREAD_SANITIZER] (GC_start_mark_threads): Wrap
GC_start_mark_threads_inner() call into LOCK/UNLOCK.

1 files changed, 1 insertions, 4 deletions

diff --git a/mark.c b/mark.c
index b005a1d2..53ca3e4a 100644
--- a/mark.c
+++ b/mark.c
@@ -924,7 +924,7 @@ GC_INNER void GC_wait_for_markers_init(void)
 {
   signed_word count;
 
-  GC_ASSERT(I_DONT_HOLD_LOCK());
+  GC_ASSERT(I_HOLD_LOCK());
   if (GC_markers_m1 == 0)
     return;
 
@@ -937,15 +937,12 @@ GC_INNER void GC_wait_for_markers_init(void)
   {
     size_t bytes_to_get =
                 ROUNDUP_PAGESIZE_IF_MMAP(LOCAL_MARK_STACK_SIZE * sizeof(mse));
-    DCL_LOCK_STATE;
 
     GC_ASSERT(GC_page_size != 0);
     GC_main_local_mark_stack = (mse *)GET_MEM(bytes_to_get);
     if (NULL == GC_main_local_mark_stack)
       ABORT("Insufficient memory for main local_mark_stack");
-    LOCK();
     GC_add_to_our_memory((ptr_t)GC_main_local_mark_stack, bytes_to_get);
-    UNLOCK();
   }
 
   /* Reuse marker lock and builders count to synchronize        */