io_uring: fix bug in slow unregistering of nodes

In some cases io_rsrc_ref_quiesce will call io_rsrc_node_switch_start,
and then immediately flush the delayed work queue &ctx->rsrc_put_work.

However the percpu_ref_put does not immediately destroy the node, it
will be called asynchronously via RCU. That ends up with
io_rsrc_node_ref_zero only being called after rsrc_put_work has been
flushed, and so the process ends up sleeping for 1 second unnecessarily.

This patch executes the put code immediately if we are busy
quiescing.

Fixes: 4a38aed ("io_uring: batch reap of dead file registrations")
Signed-off-by: Dylan Yudaken <dylany@fb.com>
Link: https://lore.kernel.org/r/20220121123856.3557884-1-dylany@fb.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Author: Dylan Yudaken

fs/io_uring.c

@@ -7822,10 +7822,15 @@ static __cold void io_rsrc_node_ref_zero(struct percpu_ref *ref)
 	struct io_ring_ctx *ctx = node->rsrc_data->ctx;
 	unsigned long flags;
 	bool first_add = false;
+	unsigned long delay = HZ;
 
 	spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
 	node->done = true;
 
+	/* if we are mid-quiesce then do not delay */
+	if (node->rsrc_data->quiesce)
+		delay = 0;
+
 	while (!list_empty(&ctx->rsrc_ref_list)) {
 		node = list_first_entry(&ctx->rsrc_ref_list,
 					    struct io_rsrc_node, node);
@@ -7838,7 +7843,7 @@ static __cold void io_rsrc_node_ref_zero(struct percpu_ref *ref)
 	spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
 
 	if (first_add)
-		mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
+		mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
 }
 
 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)