[PATCH] nptl: Add one more barrier to nptl/tst-create1

[PATCH] nptl: Add one more barrier to nptl/tst-create1

[Florian Weimer]

Without the bar_ctor_finish barrier, it was possible that thread2
re-locked user_lock before ctor had a chance to lock it.  ctor then
blocked in its locking operation, xdlopen from the main thread
did not return, and thread2 was stuck waiting in bar_dtor:

thread 1: started.
thread 2: started.
thread 2: locked user_lock.
constructor started: 0.
thread 1: in ctor: started.
thread 3: started.
thread 3: done.
thread 2: unlocked user_lock.
thread 2: locked user_lock.

Fixes test in commit 83b5323261bb72313bffcf37476c1b8f0847c736
("elf: Avoid deadlock between pthread_create and ctors [BZ #28357]").

Tested on x86_64-gnu-linux with lots and lots of iterations in a loop.

---
 sysdeps/pthread/tst-create1.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/sysdeps/pthread/tst-create1.c b/sysdeps/pthread/tst-create1.c
index 932586c309..763ded8d79 100644
--- a/sysdeps/pthread/tst-create1.c
+++ b/sysdeps/pthread/tst-create1.c
@@ -33,6 +33,7 @@ thread 2: lock(user_lock) -> pthread_create
 */
 
 static pthread_barrier_t bar_ctor;
+static pthread_barrier_t bar_ctor_finish;
 static pthread_barrier_t bar_dtor;
 static pthread_mutex_t user_lock = PTHREAD_MUTEX_INITIALIZER;
 
@@ -46,6 +47,7 @@ ctor (void)
   xpthread_mutex_unlock (&user_lock);
   dprintf (1, "thread 1: in ctor: unlocked user_lock.\n");
   dprintf (1, "thread 1: in ctor: done.\n");
+  xpthread_barrier_wait (&bar_ctor_finish);
 }
 
 void
@@ -81,6 +83,7 @@ thread2 (void *a)
   xpthread_mutex_unlock (&user_lock);
   dprintf (1, "thread 2: unlocked user_lock.\n");
   xpthread_join (t3);
+  xpthread_barrier_wait (&bar_ctor_finish);
 
   xpthread_mutex_lock (&user_lock);
   dprintf (1, "thread 2: locked user_lock.\n");
@@ -99,6 +102,7 @@ thread1 (void)
 {
   dprintf (1, "thread 1: started.\n");
   xpthread_barrier_init (&bar_ctor, NULL, 2);
+  xpthread_barrier_init (&bar_ctor_finish, NULL, 2);
   xpthread_barrier_init (&bar_dtor, NULL, 2);
   pthread_t t2 = xpthread_create (0, thread2, 0);
   void *p = xdlopen ("tst-create1mod.so", RTLD_NOW | RTLD_GLOBAL);

Re: [PATCH] nptl: Add one more barrier to nptl/tst-create1

[Szabolcs Nagy]

The 12/10/2021 05:28, Florian Weimer wrote:
> Without the bar_ctor_finish barrier, it was possible that thread2
> re-locked user_lock before ctor had a chance to lock it.  ctor then
> blocked in its locking operation, xdlopen from the main thread
> did not return, and thread2 was stuck waiting in bar_dtor:
> 
> thread 1: started.
> thread 2: started.
> thread 2: locked user_lock.
> constructor started: 0.
> thread 1: in ctor: started.
> thread 3: started.
> thread 3: done.
> thread 2: unlocked user_lock.
> thread 2: locked user_lock.
> 
> Fixes test in commit 83b5323261bb72313bffcf37476c1b8f0847c736
> ("elf: Avoid deadlock between pthread_create and ctors [BZ #28357]").
> 
> Tested on x86_64-gnu-linux with lots and lots of iterations in a loop.

yes this looks right, thanks

Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>