ceph: fix mds sync() race with completing requests
authorSage Weil <sage@newdream.net>
Tue, 16 Mar 2010 22:28:54 +0000 (15:28 -0700)
committerSage Weil <sage@newdream.net>
Tue, 23 Mar 2010 14:46:56 +0000 (07:46 -0700)
The wait_unsafe_requests() helper dropped the mdsc mutex to wait
for each request to complete, and then examined r_node to get the
next request after retaking the lock.  But the request completion
removes the request from the tree, so r_node was always undefined
at this point.  Since it's a small race, it usually led to a
valid request, but not always.  The result was an occasional
crash in rb_next() while dereferencing node->rb_left.

Fix this by clearing the rb_node when removing the request from
the request tree, and not walking off into the weeds when we
are done waiting for a request.  Since the request we waited on
will _always_ be out of the request tree, take a ref on the next
request, in the hopes that it won't be.  But if it is, it's ok:
we can start over from the beginning (and traverse over older read
requests again).

Signed-off-by: Sage Weil <sage@newdream.net>
fs/ceph/mds_client.c

index a2600101ec220d0bd17b657d8a69889c6f0055a2..5ec86419836869da3ee24a400cf21222e25478be 100644 (file)
@@ -529,6 +529,7 @@ static void __unregister_request(struct ceph_mds_client *mdsc,
 {
        dout("__unregister_request %p tid %lld\n", req, req->r_tid);
        rb_erase(&req->r_node, &mdsc->request_tree);
+       RB_CLEAR_NODE(&req->r_node);
        ceph_mdsc_put_request(req);
 
        if (req->r_unsafe_dir) {
@@ -2682,29 +2683,41 @@ void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
  */
 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
 {
-       struct ceph_mds_request *req = NULL;
+       struct ceph_mds_request *req = NULL, *nextreq;
        struct rb_node *n;
 
        mutex_lock(&mdsc->mutex);
        dout("wait_unsafe_requests want %lld\n", want_tid);
+restart:
        req = __get_oldest_req(mdsc);
        while (req && req->r_tid <= want_tid) {
+               /* find next request */
+               n = rb_next(&req->r_node);
+               if (n)
+                       nextreq = rb_entry(n, struct ceph_mds_request, r_node);
+               else
+                       nextreq = NULL;
                if ((req->r_op & CEPH_MDS_OP_WRITE)) {
                        /* write op */
                        ceph_mdsc_get_request(req);
+                       if (nextreq)
+                               ceph_mdsc_get_request(nextreq);
                        mutex_unlock(&mdsc->mutex);
                        dout("wait_unsafe_requests  wait on %llu (want %llu)\n",
                             req->r_tid, want_tid);
                        wait_for_completion(&req->r_safe_completion);
                        mutex_lock(&mdsc->mutex);
-                       n = rb_next(&req->r_node);
                        ceph_mdsc_put_request(req);
-               } else {
-                       n = rb_next(&req->r_node);
+                       if (!nextreq)
+                               break;  /* next dne before, so we're done! */
+                       if (RB_EMPTY_NODE(&nextreq->r_node)) {
+                               /* next request was removed from tree */
+                               ceph_mdsc_put_request(nextreq);
+                               goto restart;
+                       }
+                       ceph_mdsc_put_request(nextreq);  /* won't go away */
                }
-               if (!n)
-                       break;
-               req = rb_entry(n, struct ceph_mds_request, r_node);
+               req = nextreq;
        }
        mutex_unlock(&mdsc->mutex);
        dout("wait_unsafe_requests done\n");