mds: update mds_lock handling in Locker contexts

For some contexts, we expect to be called back from the objecter/filer
on an I/O completion, so we must take mds_lock before updating any
MDS metadata.  In others, we expect to be called back from the MDCache
in response to updates to a CInode's state, so we assert that mds_lock
is already held.

Signed-off-by: John Spray <john.spray@redhat.com>

mds/MDLog: enable Journaler finisher

Signed-off-by: John Spray <john.spray@redhat.com>

osdc: Use a finisher from Journaler

Completions from I/O operations (i.e. the objecter) hop
through the finisher twice, because of the three layers of
locking (MDS::mds_lock -> Journaler::lock -> Objecter osd session lock)

Because on the way "right" we take the locks in that order, to avoid
deadlock we can't take the locks in the opposite order on the way
"left", hence the finishers.

Signed-off-by: John Spray <john.spray@redhat.com>

osdc: locking for Journaler

also remove the lock pointers that JournalPointer load/save
used to require in order to use the objecter.

Signed-off-by: John Spray <john.spray@redhat.com>

mds: fix calls to Objecter::wait_for_map

These were wrong in the earlier commit:
"mds: use lock-safe OSDMap accessors; adjust Objecter wait_for_map call"

Rather than checking epoch explicitly and dropping the lock before
calling wait_for_map, just make a single call to wait_for_map and handle
the return code to learn whether we are waiting or not.

Signed-off-by: John Spray <john.spray@redhat.com>

mds: update MDLog._replay_thread for locking

... no longer use mds_lock as a guard for condition variable
while waiting for journaler contexts.

Signed-off-by: John Spray <john.spray@redhat.com>

mds: use finisher for BootStart contexts

Signed-off-by: John Spray <john.spray@redhat.com>

librados: wait for map on create_ioctx failure

Ensure we have a map so we don't simply complain that a pool doesn't
exists.  Only take the lock and wait if we fail to lookup the pool,
though, so we avoid contending the lock in the general case.

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

librados/RadosClient: remove unused get_pool_name()

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

mds: allow Objecter to dispatch its own messages

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

osdc/Objecter: ignore messenger events before or after initalized

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

mon/MonClient: fix shutdown typo

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

librados dispatch fix

client: let Objecter dispatch directly

Add Objecter as a direct dispatcher.  Drop all of the callbacks and
messages we were passing along.  Wrap the IO completions in client_lock
(via C_Lock) and shunt them to the objecter_finisher.

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

client: add a Finisher for the Objecter

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

include/Context: add generic C_Lock

Wrap a callback in a Mutex.

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

mds: push objecter completions to a Finisher

Most/all of the MDS completions need to be reentrant (and potentially
call back into the Objecter).  Shove them all onto a Finisher to make
sure that is safe.

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

mds: mark objecter completions with _IO_, take mds_lock

For any completion we pass directly to Objecter, make sure we take the
mds_lock in finish(), and mark the class with _IO_ in the name.

Note that this doesn't address the use of Journaler.  And this assumes that
we are not holding the mds_lock already when Objecter::handle_osd_op_reply
is called.

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

osdc/Objecter: take over ownership of OSDMap

Instead of taking a pointer to an existing OSDMap in our constructor,
allocate our own, so that we completely own it.

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

osdc/Objecter: make OSDMap private

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

osdc/Filer: use lock-safe OSDMap accessors

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

mds/SnapServer: use lock safe OSDMap accessors

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

mds: use lock-safe OSDMap accessors; adjust Objecter wait_for_map call

We need to handle a race between when we inspect the OSDMap and when we
try to wait for the next map.

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

mds: use lock safe OSDMap accessors for TMAP2OMAP check

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

client/SyntheticClient: use lock safe OSDMap accessors

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

test_mon_workloadgen: use Objecter::wait_for_osd_map

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

osdc/Objecter: make wait_for_osd_map() locking safe

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

client: use Objecter::{get,put}_osdmap_read() interface

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

librados: use {get,put}_osdmap_read() accessors

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

osd/OSDMap: return const string from get_pool_name

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

librados: drop osdmap pool name cache

We now have a rwlock for Objecter; no more need for a separate cache here.

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

osd/OSDMap: make lookup_pg_pool_name const

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

osdc/Objecter: add name-base pool_delete

Avoid a lookup then delete race.

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

osdc/Objecter: protect pool ops with rwlock

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

osdc/Objecter: add OSDMap accessors that are lock-safe

Everybody needs to use these to inspect the OSDMap.

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

librados: let Objecter dispatch its own messages

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

osd: let Objecter dispatch directly

No need for our ObjecterDispatcher wrapper, now!

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

osdc/Objecter: wire up native Dispatcher fast dispatch

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

osdc/Objecter: make Objecter a Dispatcher

Note that it's not actually doing it yet, though!

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

mds: add finisher

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

objecter: take write lock in ms_handle_reset()

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

rwlock: assert on unlock if not locked

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

objecter: fix locking

ms_handle_reset() takes the session lock, needed to be able to call
_kick_requetss() with it held.

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

objecter: split objecter initialization

Separate objecter initialization to non cluster related work (e.g.,
internal data structures, other registrations), and to operations that
can initiate cluster interaction. This is so that we don't hit a rare
race where we can get called indirectly from one of the dispatcher callbacks
e.g., into handle_osd_map() when not yet being initialized.
This requires that objecter->init() should be called before
messenger->add_dispatcher_head(), and objecter->start() after it.

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

objecter: Objecter::Op is refcounted

Cancellation of lingerin ops collided with ops that we sent osdmap on,
needed to add a refcount for correct object delete.

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

objecter: don't serialize responses if there's no object name

This implicitly fixes an issue with the list_objects() being reentrant,
and triggers a lock dependency issue. The better solution would be to
have the callback context specify whether it's reentrant or not, but
will require a much bigger change.

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

objecter: shard completion_lock

Object ops responses are sharded, lock hashed by object name. This
guarantees ordering on the same object. Cross object order is not
guaranteed anymore.

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

objecter: a major refactoring

Fixes: #7619
Removed the client_lock (that used to pass in as a param) and replaced
it with a read-write lock (completely controlled by the objecter). Also
added a per-session read-write lock. Adapt code to use the new locking
scheme, removed locking where not needed. Replaced various counters to
atomics instead of grabbing the lock for updates. Moved ops to live
under the session.

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>

timer: fix RWTimer shutdown

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

librados: cancel operation

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

test_mon_workloadgen: fix for new objecter interface

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

rgw: abort early from logging function

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

osd: adapt to new Objecter interface

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

librados: remove warning

Seems that the current locking scheme is ok. Objecter handles its
locking internally, librados needs to hold its lock.

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

librados: push snap related functions down to objecter

Need to take appropriate lock.

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

librados: call objecter->handle_osd_map() under lock

The lock is not shared with the objecter anymore so we need to call
it under this lock, otherwise we're going to race with
wait_for_osdmap().

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

librados: adapt to new objecter interface

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

librados: don't hold client lock in certain operations

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

librados: don't hold lock in dispatch()

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

mds: adapt to new objecter interface

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

client: adapt to new objecter interface

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

time: create RWTimer

a timer implementation that uses RWLock

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

common/RWLock: add lock context

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

common/RWLock: add explicit unlock() to *Lockers

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

common/RWLock: add get()

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

common/RWLock: check return values

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

common/RWLock: track read/write locks via atomics for assertion checks

Signed-off-by: Yehuda Sadeh <yehuda@inktank.com>

Merge pull request #2138 from xiaoxichen/fix_crush_ruleset

mon/OSDMonitor : Use user provided ruleset for replicated pool

Reviewed-by: Loic Dachary <loic-201408@dachary.org>

Merge branch 'wip-weak-refs'

vstart.sh: debug rgw = 20 on -d

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

Merge pull request #2292 from lpabon/rgw-doc

vstart.sh to enable support for simple Rados

Tested-by: GWabhishek.lekshmanan@gmail.com

common/shared_cache: dump weak refs on shutdown

Before asserting, dump what the refs are along with the counts.

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

Merge pull request #2304 from majianpeng/fix1

small bug fix for journal

Reviewed-by: Sage Weil <sage@redhat.com>

common/shared_cache: take a cct

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

Merge pull request #2291 from ceph/wip-fingerprint

mon: add a unique clsuter fingerprint

Reviewed-by: Joao Eduardo Luis <joao.luis@inktank.com>

Merge pull request #2307 from dmsimard/crushdoc

Doc: Add config option to turn off crush update

Reviewed-by: Loic Dachary <loic-201408@dachary.org>

Merge pull request #2288 from xiaoxichen/fix_add_simple_ruleset

CrushWrapper: pick a ruleset same as rule_id

Reviewed-by: Loic Dachary <loic-201408@dachary.org>

Doc: Add config option to turn off crush update

Signed-off-by: David Moreau Simard <dmsimard@iweb.com>

CrushWrapper: pick a ruleset same as rule_id

Originally in the add_simple_ruleset funtion, the ruleset_id
is not reused but rule_id is reused. So after some add/remove
against rules, the newly created rule likely to have
ruleset!=rule_id.

We dont want this happen because we are trying to hold the constraint
that ruleset == rule_id.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

Merge pull request #2273 from dachary/wip-mailmap

mailmap updates

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

Reviewed-by: Sage Weil <sage@redhat.com>
Tested-by: Mark Kirkwood <mark.kirkwood@catalyst.net.nz>

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

Revert "Merge pull request #2253 from adamcrume/wip-lttng"

This reverts commit 5b9ade5058f72116eb3dc920f442361c2a33b46b, reversing
changes made to b0aa846b3f81225a779de00100e15334fb8156b3.

I merged the wrong version of wip-lttng.

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

Merge pull request #2274 from ceph/wip-9144

filestore: fix force_sync race with mount + umount + mount

Reviewed-by: David Zafman <dzafman@redhat.com>

Merge pull request #2253 from adamcrume/wip-lttng

RBD-replay

Reviewed-by: Sage Weil <sage@redhat.com>

mon: generate cluster_fingerprint if null

This triggers after an upgrade of a legacy cluster that has no fingerprint.

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

mon: add a cluster fingerprint

Generate it on cluster creations with the initial monmap.  Include it in
the report.  Provide no way for this uuid to be fed in to the cluster
(intentionally or not) so that it can be assumed to be a truly unique
identifier for the cluster.

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

rbd_replay: Add unit test for batch_unreachable_from

This requires a fair amount of code reorganization, since the types in
the function signature were previously not in a header file.

Signed-off-by: Adam Crume <adamcrume@gmail.com>

rbd-replay: Add unit test for Ser

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Split --with-lttng and --with-babeltrace options

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Remove tracing-specfic local variables when lttng disabled

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Add informational messages to configure.ac

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Only install rbd-replay-prep when enabled

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Check for babeltrace when autoconfiguring lttng

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Check that the main babeltrace header is present

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Test for sane lttng-gen-tp

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Don't build trace libraries when lttng is disabled

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Fix build error on 32-bit systems

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Remove -lboost_regex from unit test

Signed-off-by: Adam Crume <adamcrume@gmail.com>

lttng: Don't build trace headers when lttng is disabled

Signed-off-by: Adam Crume <adamcrume@gmail.com>