I've seen the "server went back in time" message while testing HEAD,
so this will fix that.  Patch already on 1.4.

b=3869,1742

These are the remaining fixes from 1742 that are needed to allow a
client to reconnect during recover.  The request that triggered 3869
on the last run was actually a RESENT request that was sent before
recovery had finished, which this patch will fix.

1) fix some bugs in create_indirect inode.

2)fix bugs in fsfilt_commit in smfs, sometime inode will be null sometime

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

  given ino/generation. we need this for cross-node mkdir() which creates
  inode on remote MDS and adds dir entry then. after that remote MDS may
  fail and primary MDS will have to recreate inode with already gotten
  ino and generation numbers
- export refcounting bug in lmv fixed
- mdc_obj_create() stores just received ino/generation back to request.
  if request replay will happen mdt_obj_create() will use stored numbers
  to recreate lost inode
- mds_filter_recovery_request() accepts OST_CREATE request. need to be
  discussed though
- mdt_obj_create() rewritten to support recovery phase
- needless mds_lmv_disconnect() in mds_fs_cleanup() removed
- possible bug in retrieve_generation_numbers() fixed: it should understand
  cross-node dir entries and put proper mdsnum's

NOTE: please, re-apply extN-wantedi patch

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

b=2939, b=2325, b=2059 (partial), b=2925, b=2399, b=2517

r=zab (partial)

Better handling for OST eviction.  Fixes for clearing the page cache,
imp_invalid handling, and improved import_event interface. Also
includes the tests and fixes added from 2732; one of those tests does
not yet passed so it is excepted for now.

b=2936
r=robert

r=robert

overwrite log when running lconf --write_conf

b=2733, b=2773, b=2529, b=2730, b=2819, b=2814, b=2822, b=1450, b=2676, b=2681
b=2817, b=2706, b=2816, b=1987, b=2884, b=1191, b=2809, b=2765, b=2805, b=1972

b: 2332
r: Shaver

r=shaver

Prevent a race between mds_client_add and mds_client_free by clearing the client
index bit after the data has been zeroed. Also make config messages consistent
(with each other and with the obdfilter messages).

     r: Andreas and Phil
     To increase the mount count of mds.

Fix i_sem/journal inversion in mds_client_add, which was never updated
when we decided to re-order these a few months ago.  This became much
easier to hit after we fixed bug 2306.

b=1600

 - dozens and dozens of fixes for working with 2.6 kernels
 - new 2.4 kernel APIs
 - uncountable bug fixes

- remove extN/ files

- client-side I/O cache
- O_DIRECT fixes for IA64
- liblustre improvements
- various small bug fixes

* bug fixes
 - Fix ldlm_lock_match on the MDS to avoid matching remote locks (592)
 - Fix fsfilt_extN_readpage() to read a full page of directory
   entries, or fake the remainder if PAGE_SIZE != blocksize (500)
 - Avoid extra mdc_getattr() in ll_intent_lock when possible (534, 604)
 - Fix imbalanced LOV object allocation and out-of-bound access (469)
 - Most intent operations were removed, in favour of a new RPC mode
   that does a single RPC to the server and bypasses most of the VFS
 - All LDLM resource ID arrays were removed in favour of ldlm_res_id
 - Aggressively cancel local locks on DLM servers
 - mds_reint_unlink sends EA to the client if it's the last nlink.
   client uses that EA to unlink OST objects.
 - mds_reint_{rename,unlink,link} were rewritten to take ordered locks
 - recursive symlinks were fixed (440)
 - fixed NULL deref in DEBUG_REQ
 - filter_update_lastobjid no longer calls sync, which annoyed extN
 - fixed multi-client small-writes to a single file problem (445)
 - fixed mtime updates during file writes (607)
 - fixed vector writes on obdfilter causing problems when ENOSPC (670)
 - fixed bug in obd_brw_read/write() (under guise of testing 367)
 - fixed Linux OST size reporting problem (444, 656)
 - OST now updates object mtime with writes or setattr (607, 619)
 - client verifies file size before zeroing page past EOF (445)
 - OST now writes last allocated objid to disk with allocation (108)
 - LOV on echo now works (409)
 * protocol changes
 - mds_reint_unlink sends a new buffer, with the EA included.  this
   buffer is only valid if body->valid & OBD_MD_FLEASIZE, which is only
   set if a regular file was being unlinked, and it was the last link
 - use PtlGet from the target for bulk writes (315)
 - OST now updates object mtime with writes or setattr (607, 619)
 - LDLM now has a grant-time callback to revalidate locked items, if
   necessary (604)
 - Many MDS operations were reorganized to combat race conditions
* other changes
 - Merge b_intel branch (updated lprocfs code) - now at /proc/fs/lustre
 - configure check to avoid gcc version 2.96 20000731-2.96-98 (606)

Fixes a _huge_ number of bugs:
  - Fully reactivate OST imports after reconnection (512, others)
  - Make sure client sees our -ENOTCONN from mds_handle (513 - partial)
  - More graceful error handling for truncating on dead OST (515)
  - Don't error out unless we're actually accessing dead stripes (474)
  - Fix garbage sizes when stripes are missing (410)
  - LRU counters were broken, causing constant lock purge (433, 432)
  - garbage on read from stripes with failed OSTs (441)
  - mark OSCs as active before reconnecting during recovery (438)
  - lov_enqueue and lov_cancel need to handle inactive OSTs (403)
  - lfind did not preserve OST order in output (443)
  - symlinks cause hung clients, incorrect data (439)
  - stop dereferencing request after dropping refcount (457)
  - don't LASSERT(spin_is_locked) on non-SMP (455)
  - fixes for many rename() bugs
  - fstat didn't correctly synchronize attributes (399)
  - server must handle lock cancellation during blocking AST prep (487)
  - bulk descriptors were free()d too soon (511)
  - fix paths in lconf, which would load incorrect modules (451, 507)
  - fix confusing lconf 'host not found' error message (386)
  - fix lock order deadlock on OST (O/R i_sem before journal ops, 478)
  - fix race condition in mdc_blocking_ast() for inode access (526)
  - fix lov_unpackmd() unpacking wrong number of stripes (537)
  - fix lov_set_osc_active() marking wrong OSC inactive (440)
  - fix bad lstripe lov_unpackmd() assertion (fix layering too) (527)
  - fix multiple writes of stripe MD to MDS (358, maybe 519)
  - fix lstripe in several ways (kernel side) (527)
  - fix request leak in ldlm_cli_enqueue (262)
  - incorrect OSC was marked inactive after OST failure
  - call mds_fs_cleanup before unmounting filesystem (524)
  - fix races between taking ns_lock and ldlm_lock_change_resource
  - fix races updating LOV export open file list
  - fix lov_enqueue error path, avoid decref-ing bad lock handle
  - fix recovery NULL deref in ldlm_cli_cancel_unused
  - fix some DLM races by using new hash table for lock handles (419)
  - permit the client to specify desired inodes, at replay
  - duplicate requests when we queue them for replay reintegration
  - fix last_rcvd offset calculation
  - sync after each recovered transaction, so we always make progress
  - never, not always, ERESTART requests without transnos
  - store the lov_desc in the MDS, so we don't depend on getlovinfo to set it
  - skip replay if the MDS says that the client is already connected
  - don't check for a recovery-enabled export to match lctl's UUID
  - don't INC_USE_COUNT for phantom exports
  - don't crash when cleaning up phantom exports (567)
  - don't double-finish or set replay data for errored mdc_open requests
  - abort requests when they time out, so we don't get old replies
  - send/receive replies for AST messages again
  - if the client says that it doesn't have the lock, cancel it on the server
  - if we timeout during I/O, don't try to cancel an in-use lock; instead
    mark it as destroyed, it will all work out when decref is called
  - fix module use counts (22, 581)
 * protocol changes
  - ASTs now expect a reply (server cancels lock on error reply)

 - fstat didn't correctly synchronize attributes (399)
 - server must handle lock cancellation during blocking AST prep (487)
 - bulk descriptors were free()d too soon (511)
 - fix paths in lconf, which would load incorrect modules (451, 507)
 - fix confusing lconf 'host not found' error message (386)

 - LRU counters were broken, causing constant lock purge (433, 432)
 - the first part of the fixes for the ldlm_handle2lock race (419)
 - cleanup unused ll_inode_info field (442)
 - keep ioctl semaphore from serializing test_getattr/brw

- fixes of many metadata bugs:
  - multiple client file opens
  - missing/extra intent releases
  - symlink fixes
- OST/MDS server threads clobbering each other
  - dbench 2 works mostly
- many others too numerous to mention here (ChangeLog is a good record)

Highlights:
  - b=324: MDS recovery must replay transactions in strict transno sequence
  - b=325: getattr after OST failure returns -EIO
  - b=326: unlink after OST failure returns -EIO
  - b=400: new client can't join cluster after OST failure
  - b=403: multi-client access failure when OST fails
  - b=410: After an OST failure, lfind incorrectly displays file information
  - b=417: Freeing unreplayable requests twice (aed's fix from b_md)
  - b=402: (partial) give error for lstripe request that exceeds configured OSTs
  - much better support for reconnecting to MDS after network partition
    (still some lock-repeating issues to be resolved for some requests)
  - better support for connecting to multiple MDSes on one host (xid and
    transno and request_list are all per-import now)
  - track disconnecting clients in last_rcvd, for more reliable recovery
  - also, sync last_rcvd after connect/disconnect
  - reduced syslog/CERROR output for recovery (hi, Terry!)
  - server (DLM) timeout is half the system-wide timeout, to avoid cascading
    failure in the face of a dead client
  - don't wait for recovery to finish in order to send disconnect messages
  - removal of c_dying_head
  - don't wait for timeout to trigger recovery after ptl_send_rpc error
  - strict MDS transno ordering via mds_transno_sem (non-optimal, but correct)
  - many !handle -> IS_ERR(handle) fixes around mds_fs_start callers.
  - turn on client-eviction for bulk-timeouts in OST and MDS

Two fixed:
- refcounts on MDS module fixed, and problems with bogus exports
- unsafe list walking on the open file list of the MDS could cause
  random memory problems
- added real handles for open files on the MDS to ensure we don't
  dereference bogus pointers - mike you still need to clean up open
  files on the MDS for disconnect

  the caller to enforce permissions, with a new argument to push_ctxt.

much easier to see correctness, and nano-faster also.

Fixes this bug by remember that sizeof() returns the size in bytes, not bits.

This probably won't fix 600245, but it is related, and would have bitten us
once we grew to more than 32 clients with an ia32 MDS.