On Fri, Jan 05, 2018 at 09:15:41AM -0800, Michael Lyle wrote:

Jens & Kent,

On 01/05/2018 08:05 AM, Jens Axboe wrote:

quoted

On 12/30/17 4:09 PM, Michael Lyle wrote:

quoted

+void __closure_sync(struct closure *cl)
+{
+	struct closure_syncer s = { .task = current };
 
+	cl->s = &s;
+	continue_at(cl, closure_sync_fn, NULL);
+
+	while (1) {
+		__set_current_state(TASK_UNINTERRUPTIBLE);
+		smp_mb(); /* Ensure task state set before load of done flag */

That's why we have set_current_state().

I wrote the comment in question-- it seemed like to me set_current_state
and a store w/ barrier, but I was nervous since I didn't write the code
that there might be another dependency/reason.

Kent-- is there any reason to not just set_current_state(...)?

No, set_current_state() is the right way to do it