- Add extN select-inum-for-allocation support, and use it to make recovery-time

  inode creation more deterministic.

- Test program and ioctl for the aforementioned wantedi support.

- Don't send MSG_LAST_REPLAY with the last-replayed lock.  Really.  I mean it
  this time.

- Some transno-reintegration CERROR removal, because the 9600bps conman on dev
  cramps the style of 300-transaction replays something fierce.

- Sync after every integrated transaction, so that we make forward progress
  in case of interrupted recovery.  (This will be replaced by the handle->h_sync
  fix foreshadowed in fsfilt_extN_commit.)

- Reuse reply_md.start instead of the likely-NULL rq_repmsg when resending reqs.

- Requests without transnos should be ERESTARTed _never_, not _always_.

diff --git a/lustre/obdclass/fsfilt_extN.c b/lustre/obdclass/fsfilt_extN.c
index 98404dd..3dfc0f8 100644 (file)
--- a/lustre/obdclass/fsfilt_extN.c
+++ b/lustre/obdclass/fsfilt_extN.c
@@ -216,12 +216,18 @@ static void *fsfilt_extN_brw_start(int objcount, struct fsfilt_objinfo *fso,
         RETURN(handle);
 }
 
-static int fsfilt_extN_commit(struct inode *inode, void *handle)
+static int fsfilt_extN_commit(struct inode *inode, void *h /*, force_sync */)
 {
         int rc;
+        handle_t *handle = h;
+
+#if 0
+        if (force_sync)
+                handle->h_sync = 1; /* recovery likes this */
+#endif
 
         lock_kernel();
-        rc = journal_stop((handle_t *)handle);
+        rc = journal_stop(handle);
         unlock_kernel();
 
         return rc;