On 2018/06/13 23:46, Jan Kara wrote:
> On Wed 13-06-18 19:43:47, Tetsuo Handa wrote:
>> Can't we utilize RCU grace period (like shown below) ?
>
> Honestly, the variant 1 looks too ugly to me. However variant 2 looks
> mostly OK. We can also avoid the schedule_timeout_uninterruptible(HZ / 10)
> from your patch by careful handling of the bit waitqueues. Also I'd avoid
> the addition argument to wb_writeback() and split the function instead. The
> patch resulting from your and mine ideas is attached. Thoughts?
>
> Honza
>
+static bool cgwb_start_shutdown(struct bdi_writeback *wb)
+ __releases(cgwb_lock)
+{
+ if (!wb_start_shutdown(wb)) {
+ DEFINE_WAIT(wait);
+ wait_queue_head_t *wqh = bit_waitqueue(&wb->state,
+ WB_shutting_down);
+ bool sleep;
+
+ prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
+ sleep = test_bit(WB_shutting_down, &wb->state);
+ spin_unlock_irq(&cgwb_lock);
+ if (sleep)
+ schedule();
+ return false;
+ }
+ spin_unlock_irq(&cgwb_lock);
+ return true;
+}
Since multiple addresses share bit_wait_table[256], isn't it possible that
cgwb_start_shutdown() prematurely returns false due to wake_up_bit() by
hash-conflicting addresses (i.e. not limited to clear_and_wake_up_bit() from
wb_shutdown())? I think that we cannot be sure without confirming that
test_bit(WB_shutting_down, &wb->state) == false after returning from schedule().
