libceph: requeue only sent requests when kicking

The osd expects incoming requests for a given object from a given
client to arrive in order, with the tid for each request being
greater than the tid for requests that have already arrived.  This
patch fixes two places the osd client might not maintain that
ordering.

For the osd client, the connection fault method is osd_reset().
That function calls __reset_osd() to close and re-open the
connection, then calls __kick_osd_requests() to cause all
outstanding requests for the affected osd to be re-sent after
the connection has been re-established.

When an osd is reset, any in-flight messages will need to be
re-sent.  An osd client maintains distinct lists for unsent and
in-flight messages.  Meanwhile, an osd maintains a single list of
all its requests (both sent and un-sent).  (Each message is linked
into two lists--one for the osd client and one list for the osd.)

To process an osd "kick" operation, the request list for the *osd*
is traversed, and each request is moved off whichever osd *client*
list it was on (unsent or sent) and placed onto the osd client's
unsent list.  (It remains where it is on the osd's request list.)

When that is done, osd_reset() calls __send_queued() to cause each
of the osd client's unsent messages to be sent.

OK, with that background...

As the osd request list is traversed each request is prepended to
the osd client's unsent list in the order they're seen.  The effect
of this is to reverse the order of these requests as they are put
(back) onto the unsent list.

Instead, build up a list of only the requests for an osd that have
already been sent (by checking their r_sent flag values).  Once an
unsent request is found, stop examining requests and prepend the
requests that need re-sending to the osd client's unsent list.

Preserve the original order of requests in the process (previously
re-queued requests were reversed in this process).  Because they
have already been sent, they will have lower tids than any request
already present on the unsent list.

Just below that, traverse the linger list in forward order as
before, but add them to the *tail* of the list rather than the head.
These requests get re-registered, and in the process are give a new
(higher) tid, so the should go at the end.

This partially resolves:
    http://tracker.ceph.com/issues/4392

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-off-by: Sage Weil <sage@inktank.com>

1 files changed, 29 insertions, 4 deletions

diff --git a/net/ceph/osd_client.c b/net/ceph/osd_client.c
index 3723a7f16afd..8b84fb4980ba 100644
--- a/net/ceph/osd_client.c
+++ b/net/ceph/osd_client.c
@@ -570,21 +570,46 @@ static void __kick_osd_requests(struct ceph_osd_client *osdc,
 				struct ceph_osd *osd)
 {
 	struct ceph_osd_request *req, *nreq;
+	LIST_HEAD(resend);
 	int err;
 
 	dout("__kick_osd_requests osd%d\n", osd->o_osd);
 	err = __reset_osd(osdc, osd);
 	if (err)
 		return;
-
+	/*
+	 * Build up a list of requests to resend by traversing the
+	 * osd's list of requests.  Requests for a given object are
+	 * sent in tid order, and that is also the order they're
+	 * kept on this list.  Therefore all requests that are in
+	 * flight will be found first, followed by all requests that
+	 * have not yet been sent.  And to resend requests while
+	 * preserving this order we will want to put any sent
+	 * requests back on the front of the osd client's unsent
+	 * list.
+	 *
+	 * So we build a separate ordered list of already-sent
+	 * requests for the affected osd and splice it onto the
+	 * front of the osd client's unsent list.  Once we've seen a
+	 * request that has not yet been sent we're done.  Those
+	 * requests are already sitting right where they belong.
+	 */
 	list_for_each_entry(req, &osd->o_requests, r_osd_item) {
-		list_move(&req->r_req_lru_item, &osdc->req_unsent);
-		dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
+		if (!req->r_sent)
+			break;
+		list_move_tail(&req->r_req_lru_item, &resend);
+		dout("requeueing %p tid %llu osd%d\n", req, req->r_tid,
 		     osd->o_osd);
 		if (!req->r_linger)
 			req->r_flags |= CEPH_OSD_FLAG_RETRY;
 	}
+	list_splice(&resend, &osdc->req_unsent);
 
+	/*
+	 * Linger requests are re-registered before sending, which
+	 * sets up a new tid for each.  We add them to the unsent
+	 * list at the end to keep things in tid order.
+	 */
 	list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 				 r_linger_osd) {
 		/*
@@ -593,7 +618,7 @@ static void __kick_osd_requests(struct ceph_osd_client *osdc,
 		 */
 		BUG_ON(!list_empty(&req->r_req_lru_item));
 		__register_request(osdc, req);
-		list_add(&req->r_req_lru_item, &osdc->req_unsent);
+		list_add_tail(&req->r_req_lru_item, &osdc->req_unsent);
 		list_add(&req->r_osd_item, &req->r_osd->o_requests);
 		__unregister_linger_request(osdc, req);
 		dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,