[src/trunk]: src/lib/libpthread libpthread(3): Fix a marvellous interaction w...

[riastradh <riastradh%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/15a79c3c503c
branches:  trunk
changeset: 366589:15a79c3c503c
user:      riastradh <riastradh%NetBSD.org@localhost>
date:      Tue May 31 14:23:39 2022 +0000

description:
libpthread(3): Fix a marvellous interaction with rtld.

Patch from chs@.  Comment explaining the story by me.  This patch may
not be optimal -- maybe it would be better in pthread__init, or
better for rtld to call _lwp_unpark after _lwp_park in the contened
case -- but we've tested this version and it's annoying to reproduce,
so let's take this version and worry about testing improvements
later.

diffstat:

 lib/libpthread/pthread.c |  32 ++++++++++++++++++++++++++++++--
 1 files changed, 30 insertions(+), 2 deletions(-)

diffs (53 lines):

diff -r 4dce85080baf -r 15a79c3c503c lib/libpthread/pthread.c
--- a/lib/libpthread/pthread.c  Tue May 31 14:21:44 2022 +0000
+++ b/lib/libpthread/pthread.c  Tue May 31 14:23:39 2022 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: pthread.c,v 1.180 2022/02/12 14:59:32 riastradh Exp $  */
+/*     $NetBSD: pthread.c,v 1.181 2022/05/31 14:23:39 riastradh Exp $  */
 
 /*-
  * Copyright (c) 2001, 2002, 2003, 2006, 2007, 2008, 2020
@@ -31,7 +31,7 @@
  */
 
 #include <sys/cdefs.h>
-__RCSID("$NetBSD: pthread.c,v 1.180 2022/02/12 14:59:32 riastradh Exp $");
+__RCSID("$NetBSD: pthread.c,v 1.181 2022/05/31 14:23:39 riastradh Exp $");
 
 #define        __EXPOSE_STACK  1
 
@@ -414,6 +414,34 @@
        else
                return EINVAL;
 
+       if (!pthread__started) {
+               /*
+                * Force the _lwp_park symbol to be resolved before we
+                * begin any activity that might rely on concurrent
+                * wakeups.
+                *
+                * This is necessary because rtld itself uses _lwp_park
+                * and _lwp_unpark internally for its own locking: If
+                * we wait to resolve _lwp_park until there is an
+                * _lwp_unpark from another thread pending in the
+                * current lwp (for example, pthread_mutex_unlock or
+                * pthread_cond_signal), rtld's internal use of
+                * _lwp_park might consume the pending unpark.  The
+                * result is a deadlock where libpthread and rtld have
+                * both correctly used _lwp_park and _lwp_unpark for
+                * themselves, but rtld has consumed the wakeup meant
+                * for libpthread so it is lost to libpthread.
+                *
+                * For the very first thread, before pthread__started
+                * is set to true, pthread__self()->pt_lid should have
+                * been initialized in pthread__init by the time we get
+                * here to the correct lid so we go to sleep and wake
+                * ourselves at the same time as a no-op.
+                */
+               _lwp_park(CLOCK_REALTIME, 0, NULL, pthread__self()->pt_lid,
+                   NULL, NULL);
+       }
+
        pthread__started = 1;
 
        /* Fetch misc. attributes from the attr structure. */