if client will evicted at this time - we start invalidate
thread without reference to import and import can be freed
at same time.

Branch HEAD
b=17802
i=nathan
i=johann

b=17052
i=nathan.rutman
i=eric.mei

b=17461
i=r.read
i=tappro

avoid nesting of obd_uncommitted_replies and rs_lock locks
in ptlrpc_handle_rs.

add debugging level for user space lgss_keyring.

i=andreas
i=nikita
i=nathan
i=manoj

Changelogs, a lightweight mechanism to track filesystem metadata and
namespace changes.

b=17461
i=adilger
i=alex.zhuravlev

Ptlrpc optimizations to minimize COS overhead:
. reply state object use own locking instead of ptlrpc service one
. per-cpu dedicated threads for reply handling
. resource hash table size increase

fix resend requests for ldlm imports.

Branch HEAD
b=17695
i=umka
i=tappro

b=16074
r=oleg,jay

small fixes for liblustre.

i=Nikita
i=umka

	compact FLD feature.

b=16860
i=nathan
i=rread

With AT enabled, the recovery window can be excessively long (6000+
seconds). To address this problem, we no longer use
OBD_RECOVERY_FACTOR when extending the recovery window (the connect
timeout no longer depends on the service time, it is set to
INITIAL_CONNECT_TIMEOUT now) and clients report the old service
time via pb_service_time.

i=umka
i=panda
b=17611

don't override lcm->lcm_name

support gss on mgc-mgs connection.
b=16054
r=umka
r=vitaly

rewrite part of sptlrpc configuration system to be able to interoperate
with 1.8.
b=15363
r=nathan
r=fanyong

Fix broken whitespace in error message.

r=shadow,panda

- fixes long sync bulk unlink in ptlrpcd which stops other rpcs from being handled and also causes asstion in umount time;

- make sure that long unlink wait is done with 1 sec interval to return quickly.

use LASSERT_SPIN_LOCKED() in LASSERT(spin_is_locked()) stead as the latter does not work correctly with the uniprocessor kernels. b=17664; i=robert.read

r=shadow,deen

- fixes writing cookie beyond llcd's boundaries.

b=16129
i=adilger
i=green

- a high priority request list is added into service;
- once a lock is canceled, all the IO requests, including coming
ones, under this lock, are moved into this list;
- PING is also added into this list;
- once a lock cancel timeout occurs, the timeout is prolonged
if there is an IO rpc under this lock;
- another request list is added into the export, used to speed up
the rpc-lock matching.

r=shadow,panda

- fixes race in ptlrpcd caused busy obd and inability to cleanup;
- cleanups and debugs in llcd code.

b=12433
i=Oleg Drokin
i=Yury Umanets

fix imp_inval message

decay qos ost/oss penalties if MDS is not creating objects
i=nathan
i=johann

b=13058
i=johann
i=yury.umanets

reserve some space in mdt_body and mdt_rec_xxx structures and add
MDS_SETATTR64_REC to allow possible expansion of uid/gid to 64 bits
in the future.
b=17601
r=adilger
r=huanghua
r=fanyong

r=adilger,behlendorf1

- avoid div/mod in lustre_hash code

r=shadow,vitaly

- check rq_phase correctly in ptlrpc_expired_set() and couple of other places.

 r=tappro,wangdi

  - fixes ost cleanup issue due to missed llcd_put() in the case ost does not receive disconnect from mds;

  - do not sleep on hanging llcd. Instead assert on it _after_ stopping recov_thread's ptlrpcd which should kill any remeining llcds;

  - fixes and cleanups, comments.

b=17176

fixed a bug in 14774 patch -- compare peer's nid instead of self's nid
in ptlrpc_connection during select failover MDS/OST nodes

i=deen

- fixes manual_umount_client() in conf_sanity.sh to handle --force correctly as it is used by number of tests (this part should be missed in tests porting);

- add assert before decref mgc refcounter;
- fixed possible leaked niduuid in case of error;
- kill @rc in server_put_super() as it is not used;
- use OBD_FAIL_CHECK_ORSET in ptlrpc_check_set() to avoid assert;
- enable test_45 as it is working now.

r=shadow,johann

- make sure that no new inflight rpcs may come after ptlrpcd_deactivate_import() for both
synchronous and asynchronous sending. To do so we make sure that imp_inflight++ is done only when
permission is granted by ptlrpc_import_delay_req() which makes decision should req be sent,
deferred or killed as import is not in the state to send it in observable future. For async
sending, rpc is only counted inflight when its added to sending or delaying list instead of just
adding it to set for processing.

This fixes assert in ptlrpc_invalidate_import() and as number of other issues;

- synchronize imp_inflight and the presence on sending or delaying list for ptlrpc_queue_wait()
case. So that, now it is guaranteed that if im_inflight != 0 we may always find hanging rpc either
in sending or in delaying list;

- make sure that in ptlrcp_queue_wait() we remove rpc from sending or delaying list and dec
inflight only after ptlrpc_unregister_reply() is done. This way we make sure that accounting is
correct. Rpc can't be returned to the pool or counted finished until lnet lets us go with finished
reply unlink;

- check for inflight and rq_list in pinger;

- comments, cleanups;

b=14166

r=tappro,johann

- proper locking in rq pool freeing time

r=johann,shadow

- fixes in ptlrcp to handle possible long reply unlink for HEAD.

b=17357

b=17357

Back the changes.

b=15393
i=alex.zhuravlev@sun.com
i=tappro@sun.com

Commit on sharing. Eliminate inter-client dependencies between
uncommitted transactions by doing transaction commits.
Thereby clients may recovery independently.

b=17357
i=johann, yangsheng

Fixed the endian issues for HEAD.
Author: shadow

  b:16389
  i:rread,nathan

AT: fix reply offset calculate and check.
b=17404
r=h.huang
r=nathan

b=17491
i=nathan
i=rread

Quick fix patch from behlendorf1@llnl.gov.

b=11089
i=nathan.rutman
i=adilger

per-nid client statistics for MDT.