io_uring: adjust defer tw counting

[ Upstream commit dc12d1799ce710fd90abbe0ced71e7e1ae0894fc ]

The UINT_MAX work item counting bias in io_req_local_work_add() in case
of !IOU_F_TWQ_LAZY_WAKE works in a sense that we will not miss a wake up,
however it's still eerie. In particular, if we add a lazy work item
after a non-lazy one, we'll increment it and get nr_tw==0, and
subsequent adds may try to unnecessarily wake up the task, which is
though not so likely to happen in real workloads.

Half the bias, it's still large enough to be larger than any valid
->cq_wait_nr, which is limited by IORING_MAX_CQ_ENTRIES, but further
have a good enough of space before it overflows.

Fixes: 8751d15426a31 ("io_uring: reduce scheduling due to tw")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/108b971e958deaf7048342930c341ba90f75d806.1705438669.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Sasha Levin <sashal@kernel.org>

diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 1cc4c7b0594924eeed055e67a2d8a2619bb0b1e9..ea772a02c14055f038339dab58118e6cc59a08b3 100644 (file)
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -1339,7 +1339,7 @@ static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
                nr_tw = nr_tw_prev + 1;
                /* Large enough to fail the nr_wait comparison below */
                if (!(flags & IOU_F_TWQ_LAZY_WAKE))
-                       nr_tw = -1U;
+                       nr_tw = INT_MAX;
 
                req->nr_tw = nr_tw;
                req->io_task_work.node.next = first;