This is an automated email from the ASF dual-hosted git repository.

xiaoxiang781216 pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/nuttx.git

commit ef20f417f0610a1554109b70762bc2c1a53d5ce2
Author: pangzhen1 <[email protected]>
AuthorDate: Fri Oct 23 22:06:36 2026 +0800

    libc/pthread: Make cond wait release and block atomic.
    
    pthread_cond_wait() and pthread_cond_clockwait() release the mutex and
    only then block on cond->sem.  If the waiter is preempted in between,
    another thread can take the mutex, call pthread_cond_signal() (the post
    finds no blocked waiter and just raises the count) and then wait on the
    same condition itself, consuming that post.  The original waiter blocks
    afterwards and both threads wait forever.
    
    PSE52 pthread_cancel, pthread_sigmask and sigwait stress hang this way
    with RR scheduling: the coredumps show flag == 1, wait_count == 2 and
    semcount == -2 with both threads on the cond waitlist and the mutex
    free.
    
    Hold sched_lock() across the release and the block so no other thread
    on this CPU can run in the window.  Blocking with the scheduler locked
    is already done by nxsem_wait_slow() for priority inheritance.  This
    does not close the window against threads on another CPU under SMP.
    
    Assisted-by: Kiro:auto
    Signed-off-by: pangzhen1 <[email protected]>
---
 libs/libc/pthread/pthread_condclockwait.c | 5 +++++
 libs/libc/pthread/pthread_condwait.c      | 7 +++++++
 2 files changed, 12 insertions(+)

diff --git a/libs/libc/pthread/pthread_condclockwait.c 
b/libs/libc/pthread/pthread_condclockwait.c
index f49ed6659b8..23687307b74 100644
--- a/libs/libc/pthread/pthread_condclockwait.c
+++ b/libs/libc/pthread/pthread_condclockwait.c
@@ -43,6 +43,7 @@
 #include <nuttx/cancelpt.h>
 #include <nuttx/pthread.h>
 #include <nuttx/atomic.h>
+#include <nuttx/sched.h>
 
 #include "pthread.h"
 
@@ -124,6 +125,8 @@ int pthread_cond_clockwait(FAR pthread_cond_t *cond,
 
       sinfo("Give up mutex...\n");
 
+      sched_lock();
+
       atomic_add(COND_WAIT_COUNT(cond), 1);
 
       /* Give up the mutex */
@@ -139,6 +142,8 @@ int pthread_cond_clockwait(FAR pthread_cond_t *cond,
             }
         }
 
+      sched_unlock();
+
       /* Reacquire the mutex (retaining the ret). */
 
       sinfo("Re-locking...\n");
diff --git a/libs/libc/pthread/pthread_condwait.c 
b/libs/libc/pthread/pthread_condwait.c
index 638eb6a8727..be6326c92a3 100644
--- a/libs/libc/pthread/pthread_condwait.c
+++ b/libs/libc/pthread/pthread_condwait.c
@@ -35,6 +35,7 @@
 #include <nuttx/atomic.h>
 #include <nuttx/cancelpt.h>
 #include <nuttx/pthread.h>
+#include <nuttx/sched.h>
 
 #include "pthread.h"
 
@@ -90,10 +91,16 @@ int pthread_cond_wait(FAR pthread_cond_t *cond, FAR 
pthread_mutex_t *mutex)
 
       sinfo("Give up mutex / take cond\n");
 
+      /* Unlock and block atomically, or a later waiter can steal the post */
+
+      sched_lock();
+
       atomic_add(COND_WAIT_COUNT(cond), 1);
       ret = pthread_mutex_breaklock(mutex, &nlocks);
 
       status = -nxsem_wait_uninterruptible(&cond->sem);
+      sched_unlock();
+
       if (ret == OK)
         {
           /* Report the first failure that occurs */

Reply via email to