[v2] Fix data race with the state Field of ThreadPoolElement

Commit Message

Vitalii Mordan Feb. 24, 2025, 4:17 p.m. UTC

TSAN reports a potential data race on the state field of
ThreadPoolElement. This is fixed by using atomic access to the field.

Fixes: d354c7eccf ("aio: add generic thread-pool facility")
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2822
Signed-off-by: Vitalii Mordan <[e-mail]>
---
v2: Addressed the Paolo Bonzini comments.
 util/thread-pool.c | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

Patch

diff --git a/util/thread-pool.c b/util/thread-pool.c
index 27eb777e85..2ef3f1ff84 100644
--- a/util/thread-pool.c
+++ b/util/thread-pool.c
@@ -111,9 +111,8 @@  static void *worker_thread(void *opaque)
         ret = req->func(req->arg);
 
         req->ret = ret;
-        /* Write ret before state.  */
-        smp_wmb();
-        req->state = THREAD_DONE;
+        /* Atomically update state after setting ret.  */
+        qatomic_store_release(&req->state, THREAD_DONE);
 
         qemu_bh_schedule(pool->completion_bh);
         qemu_mutex_lock(&pool->lock);
@@ -180,7 +179,10 @@  static void thread_pool_completion_bh(void *opaque)
 
 restart:
     QLIST_FOREACH_SAFE(elem, &pool->head, all, next) {
-        if (elem->state != THREAD_DONE) {
+        /* Atomically access the state field to synchronize with the
+         * worker_thread and thread_pool_cancel.
+         */
+        if (qatomic_load_acquire(&elem->state) != THREAD_DONE) {
             continue;
         }
 
@@ -189,9 +191,6 @@  restart:
         QLIST_REMOVE(elem, all);
 
         if (elem->common.cb) {
-            /* Read state before ret.  */
-            smp_rmb();
-
             /* Schedule ourselves in case elem->common.cb() calls aio_poll() to
              * wait for another request that completed at the same time.
              */
@@ -223,12 +222,13 @@  static void thread_pool_cancel(BlockAIOCB *acb)
     trace_thread_pool_cancel(elem, elem->common.opaque);
 
     QEMU_LOCK_GUARD(&pool->lock);
-    if (elem->state == THREAD_QUEUED) {
+    if (qatomic_read(&elem->state) == THREAD_QUEUED) {
         QTAILQ_REMOVE(&pool->request_list, elem, reqs);
         qemu_bh_schedule(pool->completion_bh);
 
-        elem->state = THREAD_DONE;
         elem->ret = -ECANCELED;
+        /* Atomically update state after setting ret. */
+        qatomic_store_release(&elem->state, THREAD_DONE);
     }
 
 }