Added LevelDBStore

Signed-off-by: Samuel Just <samuel.just@dreamhost.com>

Added leveldb submodule

Signed-off-by: Samuel Just <samuel.just@dreamhost.com>

Makefile: make check-local relative to $(srcdir)

Signed-off-by: Samuel Just <samuel.just@dreamhost.com>

rgw: don't check for ECANCELED in the _impl() functions

We already check it in the outer functions.

Signed-off-by: Yehuda Sadeh <yehuda@hq.newdream.net>

rgw: don't retry certain operations if we raced

The atomic get/put scheme was retrying writes in case where it lost
races (head object was rewritten by another client). Instead we can
just back off and return success.

Signed-off-by: Yehuda Sadeh <yehuda@hq.newdream.net>

msgr: fix race in learned_addr()

- two connect() threads
- both hit if (need_addr) check
- one takes lock, sets addr, need_addr = false, unlocks
- continues to ::encode(ms_addr, ...);
- meanwhile, second thread set ms_addr _again_, but copies peer port into
  place before adjusting it.  racing ::encode() sees bad port and sends it
  to the peer.

Fix this two ways:

- don't copy bad port into place; set it first
- re-check need_addr after taking lock

Fixes: #1747
Signed-off-by: Sage Weil <sage@newdream.net>
Reviewed-by: Greg Farnum <gregory.farnum@dreamhost.com>

msgr: print existing->state before failing assert

May help with #1378.

Signed-off-by: Sage Weil <sage@newdream.net>

Merge remote-tracking branch 'gh/wip-2121'

Reviewed-by: Yehuda Sadeh <yehuda.sadeh@dreamhost.com>

osd: unregister signal handlers on shutdown

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

mon: unregister signal handlers on shutdown

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

mds: unregister SIGHUP too

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

radosgw: handle SIGHUP

Fixes: #2121
Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

init-radosgw: add 'reload' command to send SIGHUP

Fixes: #2121
Signed-off-by: Sage Weil <sage@newdream.net>

osd: fix typo is recovery_state query dump

Signed-off-by: Sage Weil <sage@newdream.net>

osd: add missing space to scrub error

[ERR] 18.5 osd.3: soid 8a5e37ad/rb.0.0.000000002b99/headextra attr _, extra attr snapset

Signed-off-by: Sage Weil <sage@newdream.net>

msgr: discard the local_pipe's queue on shutdown.

To facilitate this, we do two things:
1) actually identify the number of special code values we pass around
2) use that to prevent trying to put() those non-pointer values in
Pipe::discard_queue().
Then we just call local_pipe.discard_queue() in wait() like happens
(indirectly, via reaping) with all the normal Pipes in rank_pipe.

But this does make me think that we may be approaching the point
where it's appropriate to create a subclass LocalPipe (against a
RemotePipe like our current Pipe implementation is mostly intended
to be).

Should fix #2086.

Signed-off-by: Greg Farnum <gregory.farnum@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>

osd: remove down OSDs from peer_info on reset

If an OSD goes down, remove it from peer_info. In particular, I saw

2012-02-28 11:04:25.851038 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3599 mlcod 0'0 peering] state<Started/Primary/Peering>: Peering advmap
2012-02-28 11:04:25.851491 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3599 mlcod 0'0 peering]  PriorSet: affected_by_map osd.1 now down
...
2012-02-28 11:04:25.998186 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3602 mlcod 0'0 peering]  PriorSet: build_prior interval(3587-3597 [3,1]/[3,1] maybe_went_rw)
2012-02-28 11:04:25.998636 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3602 mlcod 0'0 peering]  PriorSet: build_prior  prior osd.1 is down
2012-02-28 11:04:25.999106 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3602 mlcod 0'0 peering]  PriorSet: build_prior final: probe 3,5 down 1 blocked_by {}
...
2012-02-28 11:04:26.001723 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3602 mlcod 0'0 peering] enter Started/Primary/Peering/GetLog
2012-02-28 11:04:26.002428 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3602 mlcod 0'0 peering] calc_acting osd.1 1.15( v 10'1 (0'0,10'1] n=1 ec=1 les/c 0/3587 3598/3598/3598)
2012-02-28 11:04:26.003000 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3602 mlcod 0'0 peering] calc_acting osd.3 1.15( v 10'1 (0'0,10'1] n=1 ec=1 les/c 0/3587 3598/3598/3598)
2012-02-28 11:04:26.003528 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3602 mlcod 0'0 peering] calc_acting osd.5 1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598)
2012-02-28 11:04:26.004109 12e53700 osd.5 3602 pg[1.15( empty n=0 ec=1 les/c 0/3587 3598/3598/3598) [5,3] r=0 lpr=3602 mlcod 0'0 peering] calc_acting newest update on osd.1 with 1.15( v 10'1 (0'0,10'1] n=1 ec=1 les/c 0/3587 3598/3598/3598)

Any time an osd goes down we want to ensure we remove it from peer_info.
Handling this in Reset and Started states captures all of the nested
states, which forward the event (or re-post transit to Reset).  We can
also drop the Primary reaction, which is now superfluous.

Signed-off-by: Sage Weil <sage@newdream.net>
Reviewed-by: Samuel Just <samuel.just@dreamhost.com>
Reviewed-by: Josh Durgin <josh.durgin@dreamhost.com>

Merge branch 'next'

mon: report pgs stuck inactive/unclean/stale in health check

Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>
Reviewed-by: Sage Weil <sage.weil@dreamhost.com>

mon: fix slurp_latest to fill in any missing incrementals

Fixes #1789.

Signed-off-by: Greg Farnum <gregory.farnum@dreamhost.com>

test_osd_types: fix unit test for new pg_t::is_split() prototype

Signed-off-by: Sage Weil <sage@newdream.net>

Makefile: drop separate libjson_spirit.la

automake seems to have difficulty with the .la dependency on another .la.
Since libjson_spirit.la is only used by libcommon.la anyway, just build it
directly into that.  Sigh.

...
CXXLD libjson_spirit.la
AR libmds.a
CXXLD libcls_rbd.la
CXXLD libcls_rgw.la
CXXLD cephfs
CCLD test_ioctls
CC libcommon_la-ceph_ver.lo
CXX libcommon_la-version.lo
CXX ceph_dencoder.o
CCLD mount.ceph
CC ceph_ver.o
CXX test_libhadoopcephfs_build-version.o
CXXLD test_libhadoopcephfs_build
CXXLD libcommon.la
libtool: link: cannot find the library `libjson_spirit.la' or unhandled argument `libjson_spirit.la'

Signed-off-by: Sage Weil <sage@newdream.net>

osd: drop useless ENOMEM check

new throws exception; doesn't return NULL.

Signed-off-by: Sage Weil <sage@newdream.net>

ceph-osd: clarify error messages

So we know where the error came from.  And use real error codes in init().

Signed-off-by: Sage Weil <sage@newdream.net>

init: Actually do start the daemons when 'service ceph start <type>' is specified

A bug in my previous patch prevented any daemon with auto_start set to false from starting.

This patch allows:
* /etc/init.d/ceph start osd|mds|mon
* service ceph start osd|mds|mon

It however does not start daemons if auto_start is disabled when you invoke:
* /etc/init.d/ceph start
* service ceph start

Signed-off-by: Wido den Hollander <wido@widodh.nl>

doc: beginnings of documentation of stuck pgs and pg states

Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>

filestore: make less noise on ENOENT

Don't generate high-level log spam on every open error.

Signed-off-by: Sage Weil <sage@newdream.net>
Reviewed-by: Samuel Just <samuel.just@dreamhost.com>

pg: use get_cluster_inst instead of get_inst in activate

This was mistakenly broken in 4b3bb5ab37a05fa001d59f24da7d9c30d650321b

Signed-off-by: Greg Farnum <gregory.farnum@dreamhost.com>
Reviewed-by: Sam Just <sam.just@dreamhost.com>

Merge branch 'wip-split2'

Reviewed-by: Samuel Just <samuel.just@dreamhost.com>

osd: pg_t::is_split(): make children out param a pointer, and optional

Also unit test it.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: bypass split code

Until it is fully implemented.  It's also disabled in the monitor
currently, but just in case it gets into the OSDMap, do nothing for now.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: fix pg locking flags

Two things we need to handle:

 - callers who already hold map_lock (split_pg())
 - callers who already hold another pg->lock, and want to skip the lockdep
   check for this one.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: partially refactor pg split

This partially refactors the OSD split code to do the split synchronously
when processing a new OSDMap.  It is incomplete in that it does not yet
do anything useful for the PG.  The full solution needs to:

- Do the split synchronously when applying the map update.
- Reset the parent pg so that it repeers.  This will cause problems until
  we consistently consider this a new interval when looking backwards in
  time; this needs to be fixed.  Anybody doing generate_past_intervals()
  or similar will need to consider a split/merge event as an interval
  boundary.
- The recovery state machine should trigger appropriately when this
  happens.
- The old PG that was split should probably be handle identically to the
  new children.  That means deleting the old PG instance and creating a new
  PG object for the newly-split child.  Ditto for merge.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: implement pg_t::is_split()

Test to determine if a pg has split between two pool sizes, and if so,
what its children are.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: factor hobject key into child pgid calc during split

When we calculate the object's new pg, take the locator key into
consideration, to avoid a crash like

osd/OSD.cc: In function 'void OSD::split_pg(PG*, std::map<pg_t, PG*>&,ObjectStore::Transaction&)' thread 7fe3df8c4700 time 2012-02-20 18:22:19.900886
osd/OSD.cc: 4066: FAILED assert(child)

Signed-off-by: Sage Weil <sage@newdream.net>

journaler: log on unexpected objecter error

This will help with #2110, #1796, #1640.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: fix recursive map_lock via check_replay_queue()

Also drop activate_pg() helper while we're at it, so it's clear that we
are the only user.

recursive lock of OSD::map_lock (33)
 ceph version 0.42-146-g7ad35ce (commit:7ad35ce489cc5f9169eb838e1196fa2ca4d6e985)
2012-02-24 12:30:16.541416 1: (PG::lock(bool)+0x2a) [0xa09348]
2012-02-24 12:30:16.541424 2: (OSD::_lookup_lock_pg(pg_t)+0xbd) [0x84b8df]
2012-02-24 12:30:16.541431 3: (OSD::activate_pg(pg_t, utime_t)+0x9f) [0x87463b]
2012-02-24 12:30:16.541442 4: (OSD::check_replay_queue()+0x12f) [0x87452d]
2012-02-24 12:30:16.541450 5: (OSD::tick()+0x23c) [0x8535ea]
2012-02-24 12:30:16.541456 6: (OSD::C_Tick::finish(int)+0x1f) [0x881671]
2012-02-24 12:30:16.541462 7: (SafeTimer::timer_thread()+0x2d5) [0x8f8211]
2012-02-24 12:30:16.541468 8: (SafeTimerThread::entry()+0x1c) [0x8f923c]
2012-02-24 12:30:16.541475 9: (Thread::_entry_func(void*)+0x23) [0x9c8109]
2012-02-24 12:30:16.541485 10: (()+0x68ba) [0x7f9dbed838ba]
2012-02-24 12:30:16.541491 11: (clone()+0x6d) [0x7f9dbd66f02d]
2012-02-24 12:30:16.541495 common/lockdep.cc: In function 'int lockdep_will_lock(const char*, int)' thread 7f9db9d98700 time 2012-02-24 12:30:16.541504

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>
Reviewed-by: Sam Just <samuel.just@dreamhost.com>

init-ceph: stick with /var/run for the time being

/run isn't present on older systems.  Stick with the old location until it
is more pervasive, or we add an autoconf option to control it.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

debian: /var/run/ceph -> /run/ceph

/run/ceph should exists for creating UNIX domain sockets
ceph uses UNIX domain sockets for internal communication. Create their
directory on startup as /run is on a virtual filesystem.

Last-Update: <2012-02-26>
Bug-Debian: http://bugs.debian.org/660238
Forwarded: <ceph-devel@vger.kernel.org>
Signed-off-by: Laszlo Boszormenyi (GCS) <gcs@debian.hu>

debian: build-{indep,arch}

Signed-off-by: Laszlo Boszormenyi <gcs@debian.hu>

debian: sdparm|hdparm, new standards version

Signed-off-by: Laszlo Boszormenyi <gcs@debian.hu>

rgw: initialize bucket_id in bucket structure

might make valgrind a little bit less noisy.

Signed-off-by: Yehuda Sadeh <yehuda@hq.newdream.net>

rgw: _exit(0) on SIGTERM

We need to do something a bit smarter to get coverage information, but this
is a start.

Signed-off-by: Sage Weil <sage@newdream.net>

Merge remote branch 'gh/wip-crush-adjust'

Reviewed-by: Tommi Virtanen <tommi.virtanen@dreamhost.com>

Merge remote branch 'gh/wip-mds-resetter'

Reviewed-by: Greg Farnum <gregory.farnum@dreamhost.com>

Merge branch 'wip-pg-query'

Reviewed-by: Samuel Just <samuel.just@dreamhost.com>

Merge branch 'stable'

v0.42.2

Merge remote-tracking branch 'gh/stable' into stable

Merge branch 'stable'

osd: fix array index

Signed-off-by: Sage Weil <sage@newdream.net>

lockdep: don't make noise on startup

Who cares!

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

formatter: fix trailing dump_stream()

Flush a previous dump_stream() if it was the last thing prior to a
close_section().

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

osd: include timestamps in state json dumps

Include the time we entered this state in the dump.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

Merge branch 'wip-2007'

Reviewed-by: Josh Durgin <josh.durgin@dreamhost.com>

osd: use blocks for readability in list_missing

Signed-off-by: Sage Weil <sage@newdream.net>

osd: dump recovery_state states in json

Use a formatter.  Present a vector of states, inner to outer.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

osd: query Peering substates

Signed-off-by: Sage Weil <sage@newdream.net>

osd: query recovery state machine

For now, just append this to the end of the pg <pgid> query json dump.
We definitely want to do something smarter here, but I'm not sure whether
json or plaintext is the way to go.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: add tunable for number of records in osd command replies

e.g., 'pg <pgid> list_missing [offset]'.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: 'pg <pgid> list_missing <json hobject_t offset>'

Dump missing objects in json.  If more key is non-zero, user should ask for
more by passing the last object as the offset for the next request.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

hobject_t: decode json

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

add libjson_spirit.la

This is lightweight and relies on boost spirit, which we already use, so
there are no new dependencies.

There were some other libraries that also looked good, but they weren't
already packages for existing Debian distros like squeeze or even wheezy.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

osd: pass in data to do_command

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

osd: 'tell osd.N mark_unfound_lost revert' -> 'pg <pgid> mark_unfound_lost revert'

More consistent interface.

Fixes: #2030
Signed-off-by: Sage Weil <sage.weil@dreamhost.com>
Reviewed-by: Josh Durgin <josh.durgin@dreamhost.com>

lockdep: warn on stderr (via derr), not stdout

Otherwise we screw up ceph-conf output and the like.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

do_autogen.sh: -T for --without-tcmalloc

Signed-off-by: Sage Weil <sage@newdream.net>

ceph: fix help.t

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

v0.42.1

debian: add ceph-dencoder

Signed-off-by: Sage Weil <sage@newdream.net>

ceph.spec.in: add ceph-dencoder

Signed-off-by: Sage Weil <sage@newdream.net>

ceph-dencoder: man page

Signed-off-by: Sage Weil <sage@newdream.net>

ceph-tool: remove reference to "stop" command

This doesn't exist any more, and I don't think it
ever "cleanly shut down the filesystem" -- certainly not
within my recent lifetime!

Signed-off-by: Greg Farnum <gregory.farnum@dreamhost.com>
Reviewed-by: Dan Mick <dan.mick@dreamhost.com>

mds: remove unused MDBalancer dump_pop_map() function.

Commenting it out is not the right answer. ;)

Signed-off-by: Greg Farnum <gregory.farnum@dreamhost.com>
Reviewed-by: Dan Mick <dan.mick@dreamhost.com>

mds: clean up useless block

Signed-off-by: Sage Weil <sage@newdream.net>

mds: fix Resetter locking

We need to hold the lock for ms_dispatch, esp calls into objecter.  We
should only drop it when blocking; use distinct naming for the on-stack
mutex used for that.

Reported-by: Alexandre Oliva <oliva@lsd.ic.unicamp.br>
Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

Merge remote branch 'origin/wip-mds-old-inodes'

Reviewed-by: Greg Farnum <gregory.farnum@dreamhost.com>

Merge remote branch 'origin/wip-dencoder'

Reviewed-by: Greg Farnum <gregory.farnum@dreamhost.com>

Merge remote branch 'origin/wip-1820'

Reviewed-by: Greg Farnum <gregory.farnum@dreamhost.com>

osd: only set CLEAN when we are not remapped (up == acting)

If we have a temporary mapping for this PG, consider that unclean.  This
makes CLEAN and REMAPPED mutually exclusive.  For example, a 2 node cluster
with 2x replication and one osd marked out will make the pgs all
active+remapped, not active+clean+remapped.

Fixes: #2094
Signed-off-by: Sage Weil <sage@newdream.net>

Merge remote-tracking branch 'gh/wip-pg-query'

Reviewed-by: Josh Durgin <josh.durgin@dreamhost.com>

osd: conditionally encode old pg_pool_t when no CEPH_FEATURE_OSDENC

This fixes OSDMap compatibility between v0.42 and <v0.42.

For MOSDMap, reencode maps if OSDENC feature is missing.  Also rev the
message version.  We don't use COMPAT version here because v3 can't be
understood by v2 (that's why we're checking feature bits).  (It will be
possible to do that later when our constituent types can be decoded by
multiple versions.)

Fixes: #2095
Signed-off-by: Sage Weil <sage@newdream.net>
Reviewed-by: Greg Farnum <gregory.farnum@dreamhost.com>

Merge remote-tracking branch 'gh/wip-dump-ops-in-flight'

Reviewed-by: Sage Weil <sage@newdream.net>

mon: use pending_mdsmap for deactivate

We should always look at the proposed map to avoid weird races.

Signed-off-by: Sage Weil <sage@newdream.net>

doc: 'deactivate mds' instead of 'stop mds'

Signed-off-by: Sage Weil <sage@newdream.net>

mon: mds "stop" -> "deactivate"

See #1820.

Signed-off-by: Sage Weil <sage@newdream.net>

test: add basic test for the OSD's dump_ops_in_flight adminsocket command

Signed-off-by: Greg Farnum <gregory.farnum@dreamhost.com>

osd: add "dump_ops_in_flight" to the AdminSocket.

Signed-off-by: Greg Farnum <gregory.farnum@dreamhost.com>

mon: refuse to stop mds if max_mds will make it rejoin

Otherwise the MDS will leave the cluster and immediately rejoin, which is
useless and confusing to users.  See #1820.

Signed-off-by: Sage Weil <sage@newdream.net>

crushtool: add --reweight-item cli tests

Test list, tree, and straw buckets.

Signed-off-by: Sage Weil <sage@newdream.net>

crush: fix weight adjust for list, tree buckets

Fix the typo.  Code now matches that for straw buckets.

Reported-by: ZhuRongze <zrz4ceph@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>

Merge branch 'wip-2090'

Reviewed-by: Greg Farnum <gregory.farnum@dreamhost.com>

mon: unlock mon before msgr shutdown

The ceph_mon.cc main() will delete mon when the msgr dispatch thread
completes.  Make sure we unlock before we shut down the messenger, and
avoid touching this after messenger->shutdown().

Fixes: #2090
Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

mon: deprecate mon 'stop' command

Send SIGTERM.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

msgr: join dispatch_thread after it completes

This is just for completeness.  No change in behavior, since we don't
get here until the thread has signaled it is done.

Drop the destroy() overload, since we join earlier.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

Merge remote-tracking branch 'gh/wip-stop'

filestore: use IOC_CLONERANGE intead of IOC_CLONE ioctl

This is functionally equivalent, except that valgrind doesn't complain
about a bad pointer passed to an ioctl.

Signed-off-by: Sage Weil <sage.weil@dreamhost.com>
Reviewed-by: Samuel Just <samuel.just@dreamhost.com>

osd: drop "stop" command

Send SIGINT.

Fixes: #1820
Signed-off-by: Sage Weil <sage@newdream.net>

osd: drop unused "stop" check

This is never reached: both callers handle "stop" explicitly.

Signed-off-by: Sage Weil <sage@newdream.net>

osd: don't complete recovery if unfound

Otherwise we fail the !needs_recovery() assert.  Because we aren't
recovered.  For example,

2012-02-21 16:16:13.104665 1685c700 osd.5 1217 pg[0.16( v 1215'337 lc 19'2 (0'0,1215'337] n=25 ec=1 les/c 0/1061 1210/1210/1210) [5,3] r=0 lpr=1210 mlcod 0'0 active m=23 u=23 snaptrimq=[1~99,9b~e,aa~72,11d~3d,15b~e,16a~f,17a~5,180~4,185~1a,1a0~a,1ac~10,1bd~4,1c2~8,1cb~1,1cd~1,1cf~1a,1ea~10,1fb~6,202~2,205~2,209~2,20c~8,215~2,218~5,21e~1,220~1,222~9,22c~4,231~3,235~2,238~3,23e~2,241~4,246~1,248~1,24b~1,24d~9,257~6,25e~1,263~1,265~2,268~3,26e~1,273~1,275~5,27e~1,280~2]] needs_recovery osd.3 has 23 missing
osd/PG.cc: In function 'boost::statechart::result PG::RecoveryState::Active::react(const PG::RecoveryState::RecoveryComplete&)' thread 1685c700 time 2012-02-21 16:16:13.108923
osd/PG.cc: 4070: FAILED assert(!pg->needs_recovery())
 ceph version 0.42-70-g0e4367a (commit:0e4367aaac88b99c36386b6ce5e8d816fdd4ada0)
 1: (PG::RecoveryState::Active::react(PG::RecoveryState::RecoveryComplete const&)+0x1b3) [0x6a1173]
 2: (boost::statechart::simple_state<PG::RecoveryState::Active, PG::RecoveryState::Primary, boost::mpl::list<mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na, mpl_::na>, (boost::statechart::history_mode)0>::react_impl(boost::statechart::event_base const&, void const*)+0x121) [0x6c7301]
 3: (boost::statechart::state_machine<PG::RecoveryState::RecoveryMachine, PG::RecoveryState::Initial, std::allocator<void>, boost::statechart::null_exception_translator>::process_event(boost::statechart::event_base const&)+0x6b) [0x6bfc6b]
 4: (PG::RecoveryState::handle_recovery_complete(PG::RecoveryCtx*)+0x10c) [0x67c03c]
 5: (ReplicatedPG::start_recovery_ops(int, PG::RecoveryCtx*)+0x241) [0x4f83c1]
 6: (OSD::do_recovery(PG*)+0x345) [0x54b3e5]
 7: (ThreadPool::worker()+0xa26) [0x619e66]
 8: (ThreadPool::WorkThread::entry()+0xd) [0x57ad5d]
 9: (()+0x7971) [0x5037971]
 10: (clone()+0x6d) [0x679f92d]

Signed-off-by: Sage Weil <sage@newdream.net>
Reviewed-by: Samuel Just <samuel.just@dreamhost.com>