i=scjody
i=adilger
clients have imports, not exports

r=adilger,shadow

Check that request is still eligible for replay in mdc_set_replay_data()
before doing anything.

timeout.

b=11203
i=green
i=adilger

    OST_CONNECTs to complete at startup) on HEAD

b=9829
r=adilger

Description: client incorrectly hits assertion in ptlrpc_replay_req()
Details    : for a short time RPCs with bulk IO are in the replay list,
       but replay of bulk IOs is unimplemented.  If the OST filesystem
       is corrupted due to disk cache incoherency and then replay is
       started it is possible to trip an assertion.  Avoid putting
       committed RPCs into the replay list at all to avoid this issue.

 - remove ugly debug bits
 - dump FL_REPLAY reqs once after connect to avoid flooding

 - additional check because a req with FL_REPLAY can be w/o repmsg

 - additional debug mess(age?) to learn 6K depth open req queue on cnbuild

 - the situation looks valid with opencache locks, but it floods consoles
   over. make it silent for a while

  fixes, need new nfsutils patches version 8.
* more precise error notify between kernel and lgssd.
* minor gss api param fix. and various minor gss fixes.

 - workaround for cascading timeouts when failure of 1 server node causes
   failures of unrelated connections. this caused different problems. for
   example, lock cancel timeouts.

 - add a flag in lustre_msg which indicate message type
 - add a security descriptor into most of messages send to MDSs, which
   include the uid, supplementary groups array etc.
 - add group hash
 - MDS retrive local groups database through upcall, and set/merge with
   which sent from client based on policies (here not totally finished)
 - move root_squash test into separate sanity-sec.sh

Move the transno update so it's done after we check for the -ENOTCONN
reply.

A small races exists between receiving a reply and timeout. This diff
prevents ptlrpc_expire_one_request from timing out a req that has
already received the reply.

robert's fixes:
 - add CONNECT_INITAL flag for the initial connect message from a client.  When
   server sees an old client reconnect with that flag, then it needs to set the
   handles as if it was a new client.

 - set the initial imp_conn_cnt from the server's conn_cnt. When client
   connects with INITIAL set, the server sends back the old conn_cnt+1.

 - timeout for inter-MDS requests is set to obd_timeout / 2

- landed on HEAD

- 2nd patch of the bug landed

  after HZ then ptlrpc_connect_interpret() starts reconnection immediately.
  this speeds recovery up a bit
- pinger pings inter-mds connection more ofter then usual ones
- pinger finds the most sooner event and waits for it, not for obd_timeout * 2
- mds_getattr_name() needs not to check whether mds is choosen right if
  attrs are requested by fid

b_cmobd_merge should not be used for any further development or testing!

landed on HEAD

cMD_PRE_CMD_SMASH are the relevant tags.

The pre-smashing contents of HEAD will be committed by yury into b_cmd as a
merge of the two branches.  These will be stabilized on b_cmd and eventually
landed onto HEAD.

b=3284, b=3285, b=3043, b=3236, revert liblustre group changes

b=3043, b=3258, b=3219, b=2410

b=3116, b=3064, b=3107, b=3122, b=1519, 2.6 kernel fixes, group lock fixes

    Superbug 2871:
    Sub-bugs fixed 2869 2870 3034 3036 3037 3035 3038 3135

      ptlrpc_abort_bulk() to always l_wait_event() after doing
      PtlMDUnlink() on the relevant MD.  Previous code assumed that when
      PtlMDUnlink() returned PTL_MD_INVALID, the relevant callback handler
      must have run already, which isn't true for liblustre.  The
      l_wait_event() will be a NOOP for kernel lustre (condition already
      true) but gives the liblustre event handler a chance to call all
      the callbacks.

b=2867