From f18f43cb20c3236df02472f84f5183e01c457507 Mon Sep 17 00:00:00 2001 From: Sebastian Riemer Date: Thu, 21 Feb 2013 13:28:09 +1100 Subject: [PATCH] md: protect against crash upon fsync on ro array commit bbfa57c0f2243a7c31fd248d22e9861a2802cad5 upstream. If an fsync occurs on a read-only array, we need to send a completion for the IO and may not increment the active IO count. Otherwise, we hit a bug trace and can't stop the MD array anymore. By advice of Christoph Hellwig we return success upon a flush request but we return -EROFS for other writes. We detect flush requests by checking if the bio has zero sectors. Signed-off-by: Sebastian Riemer Cc: Christoph Hellwig Cc: Ben Hutchings Cc: NeilBrown Reported-by: Ben Hutchings Acked-by: Paul Menzel Signed-off-by: Greg Kroah-Hartman --- drivers/md/md.c | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/drivers/md/md.c b/drivers/md/md.c index 98262e5336d0cc..4ef75e9c7c7751 100644 --- a/drivers/md/md.c +++ b/drivers/md/md.c @@ -299,6 +299,10 @@ static int md_make_request(struct request_queue *q, struct bio *bio) bio_io_error(bio); return 0; } + if (mddev->ro == 1 && unlikely(rw == WRITE)) { + bio_endio(bio, bio_sectors(bio) == 0 ? 0 : -EROFS); + return 0; + } smp_rmb(); /* Ensure implications of 'active' are visible */ rcu_read_lock(); if (mddev->suspended) {