drbd: fix schedule in atomic

An administrative detach used to request a state change directly to D_DISKLESS,
first suspending IO to avoid the last put_ldev() occuring from an endio handler,
potentially in irq context.

This is not enough on the receiving side (typically secondary), we may miss
some peer_req on the way to local disk, which then may do the last put_ldev()
from their drbd_peer_request_endio().

This patch makes the detach always go through the intermediate D_FAILED state.
We may consider to rename it D_DETACHING.

Alternative approach would be to create yet an other work item to be scheduled
on the worker, do the destructor work from there, and get the timing right.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

1 files changed, 9 insertions, 4 deletions

diff --git a/drivers/block/drbd/drbd_nl.c b/drivers/block/drbd/drbd_nl.c
index 25468e2be8d0..7320ac00f0fb 100644
--- a/drivers/block/drbd/drbd_nl.c
+++ b/drivers/block/drbd/drbd_nl.c
@@ -1670,12 +1670,17 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
 static int adm_detach(struct drbd_conf *mdev)
 {
 	enum drbd_state_rv retcode;
+	int ret;
 	drbd_suspend_io(mdev); /* so no-one is stuck in drbd_al_begin_io */
-	retcode = drbd_request_state(mdev, NS(disk, D_DISKLESS));
-	wait_event(mdev->misc_wait,
-			mdev->state.disk != D_DISKLESS ||
-			!atomic_read(&mdev->local_cnt));
+	retcode = drbd_request_state(mdev, NS(disk, D_FAILED));
+	/* D_FAILED will transition to DISKLESS. */
+	ret = wait_event_interruptible(mdev->misc_wait,
+			mdev->state.disk != D_FAILED);
 	drbd_resume_io(mdev);
+	if ((int)retcode == (int)SS_IS_DISKLESS)
+		retcode = SS_NOTHING_TO_DO;
+	if (ret)
+		retcode = ERR_INTR;
 	return retcode;
 }