io_uring: fix bug in slow unregistering of nodes
authorDylan Yudaken <dylany@fb.com>
Fri, 21 Jan 2022 12:38:56 +0000 (04:38 -0800)
committerJens Axboe <axboe@kernel.dk>
Sun, 23 Jan 2022 16:12:53 +0000 (09:12 -0700)
commitb36a2050040b2d839bdc044007cdd57101d7f881
tree55dfb795fd070de6f723cec1b4edd56477466fd1
parentdd81e1c7d5fb126e5fbc5c9e334d7b3ec29a16a0
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: 4a38aed2a0a7 ("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>
fs/io_uring.c