drbd: fix drbd wire compatibility for empty flushes

DRBD has a concept of request epochs or reorder-domains,
which are separated on the wire by P_BARRIER packets.

Older DRBD is not able to handle zero-sized requests at all,
so we need to map empty flushes to these drbd barriers.

These are the equivalent of empty flushes, and
by default trigger flushes on the receiving side anyways
(unless not supported or explicitly disabled),
so there is no need to handle this differently in newer drbd either.

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

diff --git a/drivers/block/drbd/drbd_req.c b/drivers/block/drbd/drbd_req.c
index 910335c30927f0429a4c4b0fddcfe74033c45e8d..0bb1e41f136f9fcb8e1d4c662d16bff4a1096de3 100644 (file)
--- a/drivers/block/drbd/drbd_req.c
+++ b/drivers/block/drbd/drbd_req.c
@@ -834,7 +834,15 @@ static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio, uns
                req->private_bio = NULL;
        }
        if (rw == WRITE) {
-               remote = 1;
+               /* Need to replicate writes.  Unless it is an empty flush,
+                * which is better mapped to a DRBD P_BARRIER packet,
+                * also for drbd wire protocol compatibility reasons. */
+               if (unlikely(size == 0)) {
+                       /* The only size==0 bios we expect are empty flushes. */
+                       D_ASSERT(bio->bi_rw & REQ_FLUSH);
+                       remote = 0;
+               } else
+                       remote = 1;
        } else {
                /* READ || READA */
                if (local) {
@@ -870,8 +878,11 @@ static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio, uns
         * extent.  This waits for any resync activity in the corresponding
         * resync extent to finish, and, if necessary, pulls in the target
         * extent into the activity log, which involves further disk io because
-        * of transactional on-disk meta data updates. */
-       if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
+        * of transactional on-disk meta data updates.
+        * Empty flushes don't need to go into the activity log, they can only
+        * flush data for pending writes which are already in there. */
+       if (rw == WRITE && local && size
+       && !test_bit(AL_SUSPENDED, &mdev->flags)) {
                req->rq_state |= RQ_IN_ACT_LOG;
                drbd_al_begin_io(mdev, sector);
        }
@@ -994,7 +1005,10 @@ allocate_barrier:
        if (rw == WRITE && _req_conflicts(req))
                goto fail_conflicting;
 
-       list_add_tail(&req->tl_requests, &mdev->newest_tle->requests);
+       /* no point in adding empty flushes to the transfer log,
+        * they are mapped to drbd barriers already. */
+       if (likely(size!=0))
+               list_add_tail(&req->tl_requests, &mdev->newest_tle->requests);
 
        /* NOTE remote first: to get the concurrent write detection right,
         * we must register the request before start of local IO.  */
@@ -1014,6 +1028,14 @@ allocate_barrier:
            mdev->net_conf->on_congestion != OC_BLOCK && mdev->agreed_pro_version >= 96)
                maybe_pull_ahead(mdev);
 
+       /* If this was a flush, queue a drbd barrier/start a new epoch.
+        * Unless the current epoch was empty anyways, or we are not currently
+        * replicating, in which case there is no point. */
+       if (unlikely(bio->bi_rw & REQ_FLUSH)
+               && mdev->newest_tle->n_writes
+               && drbd_should_do_remote(mdev->state))
+               queue_barrier(mdev);
+
        spin_unlock_irq(&mdev->req_lock);
        kfree(b); /* if someone else has beaten us to it... */