io-wq: wait for manager exit on wq destroy
authorJens Axboe <axboe@kernel.dk>
Fri, 26 Feb 2021 16:56:32 +0000 (09:56 -0700)
committerJens Axboe <axboe@kernel.dk>
Thu, 4 Mar 2021 13:33:58 +0000 (06:33 -0700)
The manager waits for the workers, hence the manager is always valid if
workers are running. Now also have wq destroy wait for the manager on
exit, so we now everything is gone.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
fs/io-wq.c

index 1e5b416..7a1d51c 100644 (file)
@@ -119,6 +119,7 @@ struct io_wq {
 
        refcount_t refs;
        struct completion started;
+       struct completion exited;
 
        atomic_t worker_refs;
        struct completion worker_done;
@@ -771,6 +772,7 @@ static int io_wq_manager(void *data)
        if (atomic_read(&wq->worker_refs))
                wait_for_completion(&wq->worker_done);
        wq->manager = NULL;
+       complete(&wq->exited);
        io_wq_put(wq);
        do_exit(0);
 }
@@ -1066,6 +1068,7 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 
        wq->task_pid = current->pid;
        init_completion(&wq->started);
+       init_completion(&wq->exited);
        refcount_set(&wq->refs, 1);
 
        init_completion(&wq->worker_done);
@@ -1095,8 +1098,10 @@ static void io_wq_destroy(struct io_wq *wq)
        cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
 
        set_bit(IO_WQ_BIT_EXIT, &wq->state);
-       if (wq->manager)
+       if (wq->manager) {
                wake_up_process(wq->manager);
+               wait_for_completion(&wq->exited);
+       }
 
        spin_lock_irq(&wq->hash->wait.lock);
        for_each_node(node) {