On 8/20/20 1:33 AM, Pavel Begunkov wrote: > Setting and clearing REQ_F_OVERFLOW in io_uring_cancel_files() and > io_cqring_overflow_flush() are racy, because they might be called > asynchronously. > > REQ_F_OVERFLOW flag in only needed for files cancellation, so if it can > be guaranteed that requests _currently_ marked inflight can't be > overflown, the problem will be solved with removing the flag > altogether. > > That's how the patch works, it removes inflight status of a request > in io_cqring_fill_event() whenever it should be thrown into CQ-overflow > list. That's Ok to do, because no opcode specific handling can be done > after io_cqring_fill_event(), the same assumption as with "struct > io_completion" patches. > And it already have a good place for such cleanups, which is > io_clean_op(). A nice side effect of this is removing this inflight > check from the hot path. > > note on synchronisation: now __io_cqring_fill_event() may be taking two > spinlocks simultaneously, completion_lock and inflight_lock. It's fine, > because we never do that in reverse order, and CQ-overflow of inflight > requests shouldn't happen often. Thakns, this is a nice cleanup too. -- Jens Axboe
