librbd: track block_ofs instead of ofs in aio_sparse_read callback

librbd: fix read_iterate for the case where block ofs is not 0

librbd: sparse read cleanup, fixes

Conflicts:

src/librbd.cc

librbd: fix and cleanup a bit read_iterate

osd: do misdirected request checks before pg active test

We can process+discard a misdirected or invalid request without the pg
being active.  Reorder a few other checks too.

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

filestore: fix fiemap

rbd: fix read_iterate for sparse read

mds: avoid EMetaBlob::fullbit copying with shared_ptr

The fullbit subclass has a pointer, making it unsafe to copy around.  Fix
by adding a undefined copy constructor and operator, and using a shared_ptr
to reference it in the containing STL list<>.

Tested-by: Sam Lang <samlang@gmail.com>
Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

encoding: add list<shared_ptr<T>> macros

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

common: fix descriptinos for -i, -n too

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

update clitests with new usage

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

common: fix generic usage for -D vs -d

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

mds: fix use-after-free on ESession replay

Reported by: Sam Lang <samlang@gmail.com>
Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

v0.29.1

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

librbd: fix block_completion race condition

If block completions finished before all of them were scheduled, the
pending_count could reach 0, and the callback could be called more
than once. This caused the reference counting to be off, so the next
call to the callback passed an invalid pointer, resulting in crashes
like this (or more mysterious ones):

common/Mutex.h: In function 'void Mutex::Lock(bool)', in thread '0x7f5f37e62700'
common/Mutex.h: 118: FAILED assert(r == 0)
 1: (librbd::RBD::AioCompletion::get_return_value()+0x19c) [0x7f5f408835bc]
 2: /usr/bin/kvm() [0x4629fd]
 3: (librbd::AioCompletion::complete_block(librbd::AioBlockCompletion*, long)+0x13c) [0x7f5f4088ab5c]
 4: (librbd::rados_aio_sparse_read_cb(void*, void*)+0x8d) [0x7f5f4088b40d]
 5: (librados::RadosClient::C_aio_sparse_read_Ack::finish(int)+0x12e) [0x7f5f4050beee]
 6: (Objecter::handle_osd_op_reply(MOSDOpReply*)+0x783) [0x7f5f40517863]
 7: (librados::RadosClient::_dispatch(Message*)+0x3c) [0x7f5f404ffe8c]
 8: (librados::RadosClient::ms_dispatch(Message*)+0x33) [0x7f5f404fff53]
 9: (SimpleMessenger::dispatch_entry()+0x69d) [0x7f5f405bf5bd]
 10: (SimpleMessenger::DispatchThread::entry()+0x1c) [0x7f5f4050052c]
 11: (()+0x6d8c) [0x7f5f40c9dd8c]
 12: (clone()+0x6d) [0x7f5f3e0b904d]

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

librbd: add AioCompletion debugging

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

librbd: fix AioCompletion race condition

If block completions finished before all of them where scheduled, the
pending_count would reach 0, and the AioCompletion would be destroyed
since the refcount would also reach 0. To fix this, hold a reference
to the AioCompletion while scheduling more block completions.

This caused a crash like this when running qemu:

1: ceph::__ceph_assert_fail (assertion=<value optimized out>, file=<value optimized out>, line=<value optimized out>, func=<value optimized out>) at common/assert.cc:86
2: Lock (ictx=0x7f83d0001f00, off=5368705024, len=0, buf=0x2fac000 "", c=0x2cebe00) at common/Mutex.h:118
3: add_block_completion (ictx=0x7f83d0001f00, off=5368705024, len=0, buf=0x2fac000 "", c=0x2cebe00) at librbd.cc:159
4: librbd::aio_read (ictx=0x7f83d0001f00, off=5368705024, len=0, buf=0x2fac000 "", c=0x2cebe00) at librbd.cc:1368
5: rbd_aio_rw_vector (bs=<value optimized out>, sector_num=10485752, qiov=<value optimized out>, nb_sectors=<value optimized out>, cb=<value optimized out>, opaque=<value optimized out>, write=0) at block/rbd.c:626
6: qemu_rbd_aio_readv (bs=<value optimized out>, sector_num=<value optimized out>, qiov=<value optimized out>, nb_sectors=<value optimized out>, cb=<value optimized out>, opaque=<value optimized out>) at block/rbd.c:649
7: bdrv_aio_readv (bs=0x2ce92e0, sector_num=10485752, qiov=0x2de08f0, nb_sectors=8, cb=<value optimized out>, opaque=<value optimized out>)

Fixes: #998
Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>

PG: clear scrub_received_maps in scrub_clear_state

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

ReplicatedPG: make_writeable, use correct size for clone_size entry

Previously, we used obs.oi.size for the clone size, but obs refers to
new_obs.  The clone's size should be the old size,
ctx->obc->obs.oi.size.

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

man: update cosd man page to include info on flush-journal option.

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

mkcephfs: fix ceph.conf reference

It's $conf (which is $dir/conf).

Reported-by: biyan chen <riby.chen@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>

Makefile: remove ancient comment

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

mon: weaken pool creation caps check

Allow a pool creation if we specify an auid but are allowed to create
buckets as anybody.

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

moncaps: whitespace

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

Fix typo in usage output for --num-osds

This patch fixes the typo in the usage description to match the actual argument string.

Signed-off-by: Sam Lang <samlang@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>

Fix segfault caused by invalid argument string.

This patchset includes minor fixes to the crushtool utility.  If an invalid bucket type is speicifed on the command line, the code was iterating through bucket_types for the length of the static array, but the last entry in that array has null (0) values, which was causing a segfault.  This patch just checks that bucket_types[i].name is non-null instead.  Also, if the wrong bucket type or algorithm is specified, prints the usage string on exit.

Signed-off-by: Sam Lang <samlang@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>

v0.29

remove dumpjournal

This code has suffered a great deal of bitrot and is superceded
by the cmds --dump-journal functionality anyway, which
is much more full-featured.

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

Merge remote branch 'origin/stable' into next

Merge remote branch 'origin/filestore_debugging' into next

SimpleMessenger: Keep a disposable flag for use in reset

pipes marked disposable must not inherit the lossy policy on reconnect.
Also, in Pipe::writer, when sent.empty() && close_on_empty, mark as
lossy to ensure that fault will close the connection.

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

rados_bencher: re-add written objects constraint to read benchmark.

Somehow, in the last major change, the constraints that kept the
bencher from trying to read non-existent objects got removed. Put
a check back in the main bench loop to fix that.

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

filestore: compare dentry->d_type against d_type constant

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

filestore: stat to test for file type if d_type is unsupported

This only affects list_collections. Previously, when using an FS that
does not support d_type, like xfs, load_pgs would fail to find any
pgs, since list_collections skipped all files.

This made clients hang because all pgs were considered non-existent by
the osd.

Reported-by: Yoshiaki Tamura <yoshi@midokura.jp>
Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>

osd, filestore: debug collection listing

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

Remove unneeded libcrush1 files

mds: journal parents of srci when srcdn is remote

If srcdn is a remote dentry, we will be journaling the src inode to update
the mtime, but we need to ensure the parent dirs are in the metablob.

Fixes: #1132
Signed-off-by: Sage Weil <sage@newdream.net>

journaler: also initialize safe_pos

on reread_head.  Keep consistent across the two methods.

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

journaler: fix trim crash after standby-replay -> active

The reread_head method needs to initialize trimming_pos (like read_head
does) or else we get confused later.

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

mds: cleanup rename_prepare a bit

Use *srci tmp.

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

vstart.sh: turn down debug ms

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

mds: pin inode while one renamed_files list

Make sure the inode is pinned while it is on the LogSegment::renamed_files
list.  Avoids a crash when an inode on that list is wrongly trimmed.

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

filestore: fix fallback/slow do_clone_range

We need to seek to the appropriate offsets on the src and destination
fd's for this to do the right thing.

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

mon: make sure osd paxos is writeable before doing timeouts

The osd paxos machine has to be writeable before we can update it.

Fixes: #1130
Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

OSD: don't keep old connection over new one in update_heartbeat_peers

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

Merge branch 'stable' into next

Allow embedded '\0' in bufferlists when copying to std::string.

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

debian: depend on libboost-dev >= 1.34

for statechart.  Partially fixes #1124.

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

osd: don't leak Connection reference

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

osd: ignore old/stale heartbeat messages

If we get heartbeat messages from old epochs from peers that are not
current, drop them and mark the connection down.  Even if they are peers
we _should_ have (because we haven't gotten a notify yet to learn about
a pg we should have but don't yet) we have a newer map epoch and will learn
about them shortly, reopening the connection.

Fixes: #1107
Signed-off-by: Sage Weil <sage@newdream.net>

osd: fix map sharing due to heartbeats

- share the map with the cluster addr
- use the new {note,get}_peer_epoch helpers to do it sanely
- don't share if we're booting; see 818fa33a661

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

osd: protect recovery_wq ops with the recovery lock

We were calling recovery_item.remove_myself() without holding the
recoveryWQ::lock. Naughty naughty!

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

crushtool: add -v verbose for --test mode

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

Add content to obsync package

v0.28.2

hadoop: track Hadoop API changes

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

SimpleMessenger: allow multiple calls to shutdown

Fixes a case where radostool crashed on an error shutdown.

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

common/Thread.h: const cleanup

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

Merge branch 'wip-obsync'

mkcephfs: pass config to osdmaptool

This lets OSDMap::create_simple() see g_conf.osd_pool_default_size when
creating the initial data, metadata, and rbd pools.

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

drop useless cm.txt

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

osdmap: take default pool size from config

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

crushtool: update help

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

obysnc: rgw target: validate all users

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

Merge branch 'wip-obsync'

mon: remove pg_temp mappings when we delete pools

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

test-obsync: test sync directly from s3->rgw

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

crushtool: fix --add-item weight being zero when parent bucket(s) created

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

obsync: fix bucket creation through rgw target

The rgw: target can now create buckets. Add a test.

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

Merge branch 'stable'

test-obsync: test big objects, user-defined xattr

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

mkcephfs: set rdir for local mon setup

Fixes: #1113
Reported-by: Bernard Grymonpon <bernard@openminds.be>
Signed-off-by: Sage Weil <sage@newdream.net>

init-ceph: ssh

Another bell/whistle.

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

obysnc: fix content-type on RGWStore

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

test-obsync: compare_directory now compares xattrs

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

ceph-pybind-test: test embedded NULLs in data

Test embedded nulls in rados data. Fix a bug in rados.Object.__str__

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

obsync: more fixes for RgwStore

* Fix content-type handling

* add vvprint and use it in Object::equals.

* support RgwStore::prefix

* more tests

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

pybind/rados: correctly return data with NULLs

Correctly handle returning data with embedded NULLs in it.

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

pybind/rados.py: throw NoData on ENODATA

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

mds: fix canceled lock attempt

If client tries to lock a file, has to wait, and then cancels the attempt,
the client will send an unlock request to unwind its state.

 - the unlock now removes the waiting lock attempt from the wait list
 - when the lock request retries and finds it is no longer on the wait
   list it will fail.

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

pybind/rados.py: rados.Object.key should be string

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

obysnc: RgwStore: make sure destination users exist

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

obsync: fix DST_OWNER

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

rgw: return EACCES if acl xattr doesn't exist

obsync: Add boto_retries, remove rgw_store.prefix

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

rados python bindings: handle xattrs with NULL

Handle extended attributes that contain NULL bytes correctly, rather
than treating everything as zero-terminated C strings.

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

PG: fix race in _activate_committed

Previously, _activate_committed would access the osdmap epoch racing
with handle_osd_map's osdmap update.  This would allow a message to be
sent from a replica to the primary tagged with the same epoch as
last_warm_restart, though the event actually occured before
last_warm_restart.  Thus the primary would fail to ignore the event and
transition to crashed.

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

RgwStore: fix some ACL issues

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

test-obysnc.py: support librgw testing

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

Rename RadosStore to RgwStore

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

test-obsync: refactor a little bit

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

Proper ACL support for rados targets

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

mds: do not shift to EXCL or MIX while rdlocked

There was an old change in file_eval() that was allowing us to switch from
SYNC to MIX or EXCL while there were rdlocks, which either caused lots of
lock thrashing or could (I think) hang things up completely.  This was
from ea10a672, an ancient fix for something related that appears to have
taken out the rdlocked check by accident.

In my tests (one writer, one stat-er), this took things from long stalls
(up to 20 seconds) to very responsive stats.  Yay!

Fixes: #791
Signed-off-by: Sage Weil <sage.weil@dreamhost.com>

crushtool: clean up add-item a bit; don't add item to same bucket twice

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

crushtool: fix remove-item

Scan all buckets instead of doing a tree traverse.

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

radosgw_admin: update clitest

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

mkcephfs.in: print out usage if no actions given

If the user didn't specify any actions, print out a usage message rather
than silently exiting.

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

rgw: Fix RGWAccess::init_storage_provider

Signed-off-by: Colin McCabe <colin.mccabe@dreamhost.com>

mkcephfs: error out on bad usage

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

make: fix build for rgw