osdc/Objecter: ignore pings from previous registrations

If we have a ping in flight and then reregister (due to interval change
or whatever), ignore the result of the previous ping.

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

osdc/Objecter: only reconnect watches; skip ack unless needed

Only watches get a RECONNECT op; lingers just get resent.

Only request an ACK in the non-watch reconnect case and when an ack
callback is specified.

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

osdc/Objecter: drop ACK callback for watch

It is semantically useless.

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

librados: move watch/notify into the Objecter

Several things here:
 - we move all of the junk from librados' RadosClient into Objecter
 - we use fast-dispatch to schedule the watch-notify events (this will
   keep them ordered wrt ping)
 - we use the LingerOp * as the librados-exposed handle so that we can
   avoid any lookups for watch-check.

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

rados: break out notify replies by client

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

osdc/Objecter: take Finisher ref from creator

Only needed for watch/notify users, so most can pass NULL.

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

librados: use Objecter's linger_id as the cookie

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

osdc/Objecter: spit out linger register step

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

osdc/Objecter: expose LingerOp instead of linger_id

We want a reference-counted handle, not an id.

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

messages/MWatchNotify: print more nicer

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

librados: watch_check() to check on status of watch

Return either an accumulated error code, or the time since the watch was
last confirmed.

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

ceph_test_rados: use new watch/notify API

Ignore error events for now; we are just stressing the OSD.

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

librados: unwatch doesn't need the oid argument

The handle is like a file handle: there is internal state that is getting
freed and the user needn't provide it again.

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

rados: use new notify API

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

librados: no need for '2' suffix in c++

Compiler disambiguates based on the rest of the signature.

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

librados: mark old rollback call deprecated

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

librados: mark old watch calls deprecated

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

librados: mark old object listing api deprecated

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

librados: mark old rollback method deprecated

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

librados: mark deprecated watch/notify methods with deprecated attribute

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

scratchtoolpp: suppress deprecation warnings

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

ceph_test_rados_api_watch_notify: test watch disconnect, reconnect

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

osd: send watch DISCONNECT to client when a watch is removed

This is an explicit notification to the client that a watch is no longer
valid and needs to be re-registered.

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

ceph_test_rados_api_watch_notify: reorder tests

Group by function, not language/pool type.

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

ceph_test_rados_api_watch_notify: drop 'Test' suffix

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

ceph_test_rados_watch_notify: fix signed/unsigned warnings

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

ceph_test_rados_api_watch_notify: verify notify after missed notify works

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

osdc/Objecter: fix whitespace

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

osdc/Objecter: factor out _send_op_account()

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

osdc/Objecter: skip ping if PAUSERD

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

osdc/Objecter: objecter_inject_no_watch_ping

Tunable to prevent PINGs.  This will let us test paths the client gets
disconnected from the OSD (at least from the client's perspective).

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

osdc/Objecter: stop sending pings once we get an error

If we get ENOTCONN or ETIMEDOUT we can stop pinging.

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

osd/Watch: set timeout event for new watchers

For a new watcher, we set the timeout event on connect, and reset it when
we get a ping.

For an old watcher, we continue to set the timeout on disconnect() and
rely on the msgr Connection reset.

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

make ping verify we are connected

osdc/Objecter: only ping watches (not notifies)

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

osd, librados: distinguish between old and new watchers

Use a new watch op code that is different from the legacy one so we can
tell when the client is doing all the new stuff (like sending pings).

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

librados: tell watcher if we cause a notify timeout

If we are a watcher and we fail to notify in a timely manner, or
circumstances otherwise conspire to prevent out ack from arriving in time,
initiate a callback.

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

osd/Watch: inform watchers when they cause a notify timeout

If a watcher doesn't ack in time, let them know that they triggered a
notify failure.

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

librados: EINVAL on bad args to watch2

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

librados: behave if errcb is NULL

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

osd/Watch: drop redundant in_progress_watchers

This is essentially watchers.size() at all times and obfuscates the code.
In particular this makes the maybe_complete_notify() logic much clearer.

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

osd/Watch: fix comment typo

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

ceph_test_rados_api_watch_notify: verify we get the right # of notifies

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

ceph_test_rados_api_watch_notify: s/handle/cookie/

Use consist nameing for C and C++ tests.

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

clean up CEPH_WATCH_EVENT_* namespace, and make COMPLETE distinct

Use a proper prefix and make a separate event for the notify completion
message.

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

osdc/Objecter: send regular PING ops

Send a full PING op to the object to ensure we are still connected. For
now just use the existing ping interval; we may want to change this in
the future.

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

osdc/Objecter: queue an error if reconnect fails

If we fail to reconnect, trigger the on_error (reusable) context to
let the user know about it.

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

osdc/Objecter: separate WATCH from RECONNECT

Use WATCH op for the initial registration.  This is idempotent in that
it will succeed whether the watch information has been persisted or not.
It is used by the client if it does not know that it is registered.

The RECONNECT op is used for any subsequent session reconnect.  It will
fail if the watch state isn't already persisted on the OSD.

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

rados: improve watch command output

- hexdump the notify payload, and include all of the new metadata
- print error events too

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

librados: add infrastructure to deliver an error notification

Use a reusable context in the WatchNotifyInfo to trigger an error
event, delivered via the existing Finisher thread.  Re-lookup the cookie
in the thread to cope with races with unregister (just as we do with
notify events).

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

librados: define error callback

This is the async notification of a watch error (failure to reconnect to
the object's OSD after a peering event or network disconnect).

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

librados: remove crufty declaration

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

osd/ReplicatedPG: handle PING and RECONNECT watch ops

The ping will essentially assert that a watch is still valid.

A reconnect will reestablish session state *only* if the watch is
still persistent.  If not, it will fail (and the client will know it
may have missed something).

Note that the only difference here is that a PING is a bit lighter
weight; it will not reestablish the session state (which should already
be established).  We could use a single op here but the unique op
code makes the messages easier to understand and simplifies the code
path a bit for PING.

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

librados: define RECONNECT and PING watch ops

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

osdc/Objecter: fix assert_src_version

Don't use the add_watch() helper; do it like assert_version
does it.

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

librados: use new watch op codes; simplify Objecter helpers

- drop the useless add_watch() helper; do it explicitly
- drop the unused var arg everywhere
- make a separate notify member of the union that excludes
  the other unused fields

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

librados: define CEPH_OSD_WATCH_OP_*

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

librados: unconditionally pass WATCH_NOTIFY to legacy callback

This was always WATCH_NOTIFY; just pass that explicitly.  No need to
pay attention to what's in the message.

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

osd: allow multiple watchers/notify replies per client

It is perfectly valid to register multiple watchers via the same
librados client.  Make the notify_replies a multimap to capture all
of the acks.

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

librados: include notifier's gid in notify

The notify replies include the identify of everyone who acked the
notify.  For symmetry, inform the notified client who sent the notify.

Fixes: #9198
Signed-off-by: Sage Weil <sage@redhat.com>

ceph_test_rados_api_watch_notify: test notify2 api

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

librados: update notify2 API to accept reply payloads

Allow the notify2 callers to provide bufferlists or char ** pointers
so that we can pass the reply buffer back to them.

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

osd: implement notify ack payloads

If the notified send back reply payloads, pass them back to the notifier.

Note that we have changed the on-wire behavior of the watch completion
message a bit: instead of sending the original notify payload back to the
notifier, we send the map of notified to replies.  Note that only users of
the new API will know what to do with the notify acknowledgement
information.  At the same time, we stop sending the original payload.
However, the old API users never saw that data; we were uselessly sending
it over the wire.

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

librados: define updated watch/notify interface

- new notify callback with the correct values:
  - notify_id
  - watch handle
  - payload
- new notify_ack call
  - not implicit when the callback returns (for new api only)
  - optional payload
- new watch2 call
  - that provides the new callback
- new notify2 call
  - with the right arguments, and optional timeout

A couple refactors in here:
- IoCtx notify_ack is now called unlocked (Note: this will soon change
  with pending Objecter locking changes)
- Objecter notify_ack takes a buffer

TODO:
- no timeout on the individual watch, yet...

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

librados: drop useless ver arg to _notify_ack

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

Merge pull request #3076 from ceph/wip-mon-mdsdne

MDSMonitor.cc: fix assertion caused by MDSMap::STATE_DNE state

Reviewed-by: Greg Farnum <gfarnum@redhat.com>

Merge pull request #3072 from ceph/wip-10231-gperftools-location

heap_profiler: support new gperftools header locations

gitbuilders are happy

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

Merge pull request #3084 from ceph/wip-rados-include

librados: Fix rados include problem reported by "sponce"

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

librados: Fix rados include problem reported by "sponce"

Caused by a03f85a8e7fab296ea2df70a929a1c5e4aa0f7fb

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

Merge pull request #3059 from dachary/wip-10125-radosgw-init

rgw: run radosgw as apache with systemd

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

rgw: run radosgw as apache with systemd

Same as sysv.

http://tracker.ceph.com/issues/10125 Fixes: #10125

Signed-off-by: Loic Dachary <loic@dachary.org>

Merge pull request #2997 from ceph/wip-snapmapper

osd: require SNAPMAPPER feature; drop tolerate_collection_add_enoent

Reviewed-by: Samuel Just <sjust@redhat.com>

Merge pull request #2979 from ceph/wip-kill-category

osd: kill 'category' from pg_stats_t

Reviewed-by: Samuel Just <sjust@redhat.com>

Merge pull request #2694 from dachary/wip-9728-jerasure-neon

erasure-code: jerasure support for NEON

Reviewed-by: Samuel Just <sjust@redhat.com>

Merge pull request #3082 from scuttlemonkey/wip-doc-10247

Changed os-recommendation table sorting to remove implied bias.

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

Merge pull request #3048 from ceph/wip-10209

osd: tolerate sessionless con in fast dispatch path

Reviewed-by: Greg Farnum <gfarnum@redhat.com>

Merge pull request #3049 from dachary/wip-10211-erasure-code-buffer-alignement

erasure-code: enforce chunk size alignment

Reviewed-by: Samuel Just <sjust@redhat.com>

doc/release-notes: v0.89

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

Merge remote-tracking branch 'gh/next'

Changed os-recommendation table sorting to remove implied bias.

Fixes: #10247
signed-of by: Patrick McGarry <pmcgarry@redhat.com>

Merge pull request #3024 from ceph/wip-librados-symbols

librados: Only public API symbols from the shared library

Reviewed-by: Loic Dachary <ldachary@redhat.com>

MDSMonitor.cc: fix assertion caused by MDSMap::STATE_DNE state

Signed-off-by: Yan, Zheng <zyan@redhat.com>

Merge pull request #3028 from majianpeng/fix4

Add perf reset command.

Reviewed-by: Loic Dachary <ldachary@redhat.com>

Merge pull request #3074 from ceph/wip-10229

osdc/Filer: use finisher to execute C_Probe and C_PurgeRange

osdc/Filer: use finisher to execute C_Probe and C_PurgeRange

Currently contexts C_Probe/C_PurgeRange are executed while holding
OSDSession::completion_lock. C_Probe and C_PurgeRange may call
Objecter::stat() and Objecter::remove() respectively, which acquire
Objecter::rwlock. This can cause deadlock because there is intermediate
dependency between Objecter::rwlock and OSDSession::completion_lock:

 Objecter::rwlock -> OSDSession::lock -> OSDSession::completion_lock

The fix is exexcute C_Probe/C_PurgeRange in finisher thread.

Fixes: #10229
Signed-off-by: Yan, Zheng <zyan@redhat.com>

heap_profiler: support new gperftools header locations

The google/ headers location has been deprecated as of gperftools 2.0.
As of gperftools 2.2rc, the google/ headers will now give deprecation
warnings, and they will probably disappear in a future gperftools
update.

Signed-off-by: Ken Dreyer <kdreyer@redhat.com>

mon: fix signed/unsigned comparison warning

Signed-off-by: Greg Farnum <gfarnum@redhat.com>

0.89

Merge pull request #3065 from ceph/wip-cephtool-exec

qa: fix cephtool tests not getting run

Reviewed-by: Loic Dachary <ldachary@redhat.com>

qa: fix teardown in cephtool's test_mon_mds

This was missing from 17b5fc9a but we didn't notice
because the test wasn't being run by the gitbuilders.

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

Merge pull request #3044 from osynge/wip=suse-systemd

better systemd support

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

Merge pull request #3067 from dachary/wip-cephtool

fix cephtools running only part of the tests + avoid killing wrong daemons

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

common/perf_counters: Remove the duplicated code.

Signed-off-by: Jianpeng Ma <jianpeng.ma@intel.com>

test/perf_counters: Replace perfcounters_dump to perf dump.

The func of command  perfcounters_dump and 'perf dump' are the same .
But from the print 'ceph --admin-daemon help', it only print 'perf
dump'. So replace.
In order to keep consistent, still keep perfcounters_dump in code for
old user.

Signed-off-by: Jianpeng Ma <jianpeng.ma@intel.com>

test/perf_counters: Add test case for cmd 'perf reset'

Signed-off-by: Jianpeng Ma <jianpeng.ma@intel.com>

common/perf_counters: Add 'perf reset all|perfcountername' for 'ceph
--admin-daemon'

Using this command, we can reset all perfcounters or reset on specified
perfcounter.

Signed-off-by: Jianpeng Ma <jianpeng.ma@intel.com>

Merge pull request #3068 from ktdreyer/doc-os-recommendations-cleanup

doc: os recommendations cleanup

Reviewed-by: Loic Dachary <ldachary@redhat.com>

doc: rm Fedora minor versions in os recommendations

Fedora doesn't have point releases, so it doesn't make sense to talk
about "Fedora 18.0" or "Fedora 19.0" - it's just "Fedora 18" or "Fedora
19".

Signed-off-by: Ken Dreyer <kdreyer@redhat.com>

os/ObjectStore: drop tolerate_collection_add_enoent

This is no longer needed and clutters the interface.

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

osd: require SNAPMAPPER feature from peers

This was introduced before cuttlefish.  We require users to upgrade first
to a newer release, so there is no need to support a mixed cluster with
such old code.

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