io_uring/io-wq: fix worker accounting when canceling creation callbacks

commit 297b5cc upstream.

create_worker_cb() reserves an io-wq worker slot only after its
task-work callback runs. If the callback is canceled before then,
io_worker_cancel_cb() still decrements acct->nr_workers. When an
existing worker retires with its creation callback pending, that
worker has already decremented the same account's worker count.

The resulting undercount permits worker creation beyond the account's
configured limit. On an AST2600 OpenBMC system, an unchanged sensor
daemon reached 4,291 threads with the original kernel. With an
equivalent downstream fix, 25 passive samples under its normal
workload showed 6-9 threads.

Decrement nr_workers only when the canceled callback is not
create_worker_cb(). Continuation callbacks still release their reserved
slot, and both callback types retain the existing running-count,
reference-count, and create-state cleanup.

[ Backport: retain the existing wq->lock protecting worker accounting. ]

Fixes: 1d5f5ea ("io-wq: remove worker to owner tw dependency")
Cc: stable@vger.kernel.org
Assisted-by: Codex:gpt-5.6-sol
Reviewed-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://patch.msgid.link/20260811-vrazdan-io-wq-b4-submit-v1-1-719ced16c921@openai.com
Signed-off-by: Vishnu Razdan <vrazdan@openai.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

Author: vrazdan-oai

io_uring/io-wq.c

@@ -192,9 +192,12 @@ static void io_worker_cancel_cb(struct io_worker *worker)
 	struct io_wq *wq = worker->wq;
 
 	atomic_dec(&acct->nr_running);
-	raw_spin_lock(&wq->lock);
-	acct->nr_workers--;
-	raw_spin_unlock(&wq->lock);
+	/* create_worker_cb() has not reserved a worker slot yet. */
+	if (worker->create_work.func != create_worker_cb) {
+		raw_spin_lock(&wq->lock);
+		acct->nr_workers--;
+		raw_spin_unlock(&wq->lock);
+	}
 	io_worker_ref_put(wq);
 	clear_bit_unlock(0, &worker->create_state);
 	io_worker_release(worker);