mds: add error handling in PurgeQueue

For decode errors, and for Journaler errors.
Both are considered damage to the MDS rank, as
with other per-rank data structures.

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

mds: expose progress during PurgeQueue drain

We don't track an item count, but we do have
a number of bytes left in the Journaler, so
can use that to give an indication of progress
while the MDS rank shutdown is waiting for
the PurgeQueue to do its thing.

Also lift the ops limit on the PurgeQueue
when it goes into the drain phase.

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

mds: update PurgeQueue for single-ack OSD change

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

mds: create purge queue if it's not found

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

qa/cephfs: add TestStrays.test_purge_on_shutdown

...and change test_migration_on_shutdown to
specifically target non-purgeable strays (i.e.
hardlink-ish things).

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

mds: wait for purgequeue on rank shutdown

Also, move shutdown_pass call from dispatch
to tick, so that it doesn't rely on incoming
messages to make progress.

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

qa: update test_strays for purgequeue

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

mds: implement PurgeQueue throttling

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

mds: add stats to PurgeQueue

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

mds: move dir purge and truncate into purgequeue

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

mds: move throttling code out of StrayManager

This will belong in PurgeQueue from now on.  We assume
that there is no need to throttle the rate of insertions
into purge queue as it is an efficient sequentially-written
journal.

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

mds: move PurgeQueue up to MDSRank

To better reflect its lifecycle: it has a part to play
in create/open and has an init/shutdown, unlike StrayManager.

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

mds: use a persistent queue for purging deleted files

To avoid creating stray directories of unbounded size
and all the associated pain, use a more appropriate
datastructure to store a FIFO of inodes that need
purging.

Fixes: http://tracker.ceph.com/issues/11950
Signed-off-by: John Spray <john.spray@redhat.com>

osdc/Journaler: wrap recover() completion in finisher

Otherwise, the callback will deadlock if it in turn
calls into any Journaler functions.  Don't care
about performance because we do this once at startup.

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

mds: const snaprealm getters on CInode

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

mds: const methods on SnapRealm

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

osdc/Filer: const fix for passed layouts

...so that const references can be passed into
purge calls.

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

common/lockdep: clearer log messages

Previously these were contextless "using id..." messages with
no indication of what subsystem the message came from.

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

osdc/Journaler: add have_waiter()

Allows users of wait_for_readable to conveniently
see if there is already a waiter.  Yes, they could
do this themselves, but I'd rather peek at an existing
variable than add a new one caller-side.

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

osdc/Journaler: remove incorrect assertion

This asserted that flush_pos would be ahead of
safe_pos after calling _flush.  However, this
is not guaranteed to be the case because
prezeroing might prevent us from flushing
right now.

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

osdc/Journaler: assign a name for logging

Now that we have an MDLog journaler and a PurgeQueue journaler,
this is needed to avoid confusion.

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

compact_set: add #includes for dependencies

This was previously working by side effects, I happened
to include it somewhere that its dependencies weren't
already included.

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

Merge pull request #13397 from SUSE/doc-fix-qa-links

doc: update links to point to ceph/qa instead of ceph-qa-suite

Reviewed-by: Kefu Chai <kchai@redhat.com>

Merge pull request #13816 from batrick/i19201

mds: print rank as int

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

Merge pull request #13830 from jcsp/wip-doc-multimds

doc: instructions and guidance for multimds

Reviewed-by: Patrick Donnelly <pdonnell@redhat.com>

doc: mention interactive task in developer guide

Signed-off-by: Nathan Cutler <ncutler@suse.com>

doc: rewrite Deploy a cluster for manual testing section

Signed-off-by: Nathan Cutler <ncutler@suse.com>

doc: update links to point to ceph/qa instead of ceph-qa-suite

Also fix two broken links to install task and two typos.

Signed-off-by: Jan Fajerski <jfajerski@suse.com>

Merge pull request #12506 from SUSE/wip-18259

Revert "dummy: reduce run time, run user.yaml playbook"

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

Merge pull request #13615 from liewegas/wip-osd-full

mon,osd: new mechanism for managing full and nearfull OSDs for luminous

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

Merge pull request #13759 from liewegas/wip-19133

osdc/Objecter: resend RWORDERED ops on full

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

Merge pull request #13734 from liewegas/wip-jewel-x

qa/suite/upgrade/jewel-x: various fixes

Reviewed-by: Yuri Weinstein <yweinstei@redhat.com>

Merge pull request #13846 from rzarzynski/wip-qa-rgw-start-apache-first

qa/tasks/rgw.py: start Apache before RadosGW.

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

qa/tasks/ceph_manager: use new luminous set-full-ratio etc

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

Merge pull request #13410 from yehudasa/wip-tracing-fix

tracing: don't include oid when tracing at dequeue_op()

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

Merge pull request #13839 from theanalyst/release/10.2.6/changelog

doc: add changelog for v10.2.6 Jewel release

doc: add changelog for v10.2.6 Jewel release

Signed-off-by: Abhishek Lekshmanan <abhishek@suse.com>

Merge pull request #13704 from batrick/mds-counter-unify

mds: remove some redundant object counters

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

osdc/Objecter: resend RWORDERED ops on full

Our condition for respecting the FULL flag is complex, and involves
the WRITE | RWORDERED flags vs the FULL_FORCE | FULL_TRY flags.  Previously,
we could block a read bc of RWORDRED but not resend it later.

Fix by capturing the complex condition in a respects_full() bool and using
it both for the blocking-on-send and resending-on-possibly-notfull-later
checks.

Fixes: http://tracker.ceph.com/issues/19133
Signed-off-by: Sage Weil <sage@redhat.com>

qa/tasks/thrashosds: chance_thrash_cluster_full

Induce a momentarily full cluster.

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

Merge pull request #13832 from linuxbox2/wip-rgw-fs_inst

rgw_file:  fix fs_inst progression

Merge pull request #10240 from songbaisen/b2

mon: remove the redudant jugement in paxosservice is_writeable function

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

rgw_file:  fix fs_inst progression

Reported by Gui Hecheng<guimark@126.com>.  This change is a
variation on proposed fix by Dan Gryniewicz<dang@redhat.com>
to take root_fh.state.dev as fs_inst for new handles.

Fixes: http://tracker.ceph.com/issues/19214
Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

qa/tasks/rgw.py: start Apache before RadosGW.

At the end of start_rgw() we wait till establishing HTTP connections
with RadosGW become possible. However, if RadosGW uses the FastCGI,
the condition can't be fulfilled without spawning HTTP server first.

Signed-off-by: Radoslaw Zarzynski <rzarzynski@mirantis.com>

Merge pull request #13323 from yehudasa/wip-18079-2

librados: use cursor for nobjects listing

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

Merge pull request #13262 from batrick/multimds-thrasher

Add multimds:thrash sub-suite and fix bugs in thrasher for multimds

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

doc: instructions and guidance for multimds

Inspired by http://tracker.ceph.com/issues/19135

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

Merge pull request #13560 from wjwithagen/wip-wjw-ceph-disk-tests

ceph-disk/tests: Certain partition types do not work on FreeBSD

Reviewed-by: Kefu Chai <kchai@redhat.com>

Merge pull request #13810 from yuyuyu101/wip-rdma-inflight

msg/async/rdma: destroy QueuePair if needed

Reviewed-by: Adir lev <adirl@mellanox.com>

Merge pull request #13742 from liupan1111/wip-cleanup-journal

os/filestore: use existing variable for same func.

Reviewed-by: Kefu Chai <kchai@redhat.com>

Merge pull request #12177 from kylinstorage/wip-remove-unneeded-loop

os/filestore/FileStore.cc: remove unneeded loop

Reviewed-by: Kefu Chai <kchai@redhat.com>

Merge pull request #13741 from rzarzynski/wip-bs-fastcrc32-in-rocks

os/bluestore: enable SSE-assisted CRC32 calculations in RocksDB

Reviewed-by: Kefu Chai <kchai@redhat.com>

Merge pull request #13768 from tchaikov/wip-clang-fixes

librados, osd: clang fixes

Reviewed-by: Brad Hubbard <bhubbard@redhat.com>

Merge pull request #13794 from liewegas/wip-clog-newlines

common: remove \n on clog messages

Reviewed-by: Brad Hubbard <bhubbard@redhat.com>

Merge pull request #13796 from liewegas/wip-debian-base-dbg

debian/control: add ceph-base-dbg

Reviewed-by: Kefu Chai <kchai@redhat.com>

qa/suites/upgrade/jewel-x/parallel: upgrade mons before osds

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

qa/suites/upgrade/jewel-x/parallel: expand workload matrix

These should run independently against a racing upgrade.

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

Merge pull request #13799 from optimistyzy/36_1

bluestore, NVMEDEVICE: Specify the max io completion in conf

Reviewed-by: Haomai Wang <haomai@xsky.com>
Reviewed-by: Pan Liu <liupan1111@gmail.com>

Merge pull request #13806 from vshankar/rbd-internal-api-move

test/librbd: move tests using non-public api to internal

Reviewed-by: Jason Dillaman <dillaman@redhat.com>
Reviewed-by: Mykola Golub <mgolub@mirantis.com>

test/cli/osdmaptool: fix osdmap output

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

mon/OSDMonitor: generate health warnings for luminous

Note that this tells us how many OSDs are full or nearfull; it
does not include detailed warnings telling you exactly what the
utilization is because we don't have the full osd_stat_t
available.  We leave it to ceph-mgr to generate those health
messages.

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

mon/PGMonitor: stop generating health warnings with luminous

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

mon/OSDMonitor: set cluster flags based on osd flags (luminous)

For luminous, set cluster flags based on osd flags.  Until
require_luminous is set, stick with the old pgmap-based behavior.
Move the new check to encode_pending so that the cluster flag is
set in the same epoch that the osd state(s) change.

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

osd: request a fullness state change during tick if needed

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

osd: require fullness state changes (as needed) before boot

This ensures that we don't have a down osd that is marked full
go up, then realize it's not actually full, and then clear its
full flag.  That would result in a cluster full blip that isn't
needed. This can easily happen if the full_ratio in the osdmap is
increased while the OSD is down.

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

osd: restructure and simplify internal fullness checks

First, eliminate the useless nearfull failsafe--all it did was
generate a log message, which we can do based on the OSDMap
states.

Add some new helpers.

Unify the cluster nearfull/full vs failsafe states so that
failsafe is a "really" full state that is more severe than
full, so we have NONE, NEARFULL, FULL, FAILSAFE.

Pull the full/nearfull ratios out of the OSDMap (remember that
we require luminous mons, so these will be initialized).

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

mon/PGMonitor: disable old 'pg set_[near]full_ratio ...' in luminous

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

qa/workunits/cephtool/test.sh: change [near]full_ratio tests

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

mon/OSDMonitor: implement new 'osd set-[near]full-ratio ...' commands

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

mon/OSDMonitor: initialize osdmap ratios from pgmap on upgrade

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

mon/OSDMonitor: set osdmap ratios on mkfs

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

mon/OSDMonitor: handle MOSDFull messages from OSDs

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

Merge pull request #13755 from liewegas/wip-19131

osd/osd_internal_types: wake snaptrimmer on put_read lock, too

Reviewed-by: Josh Durgin <jdurgin@redhat.com>

mds: print rank as int

If the MDS has no rank then its whoami field would be printed as:

    {"cluster_fsid":"4c1bae66-03fb-4b9a-bd88-108636d29758","whoami":18446744073709551615,"id":54239,"want_state":"up:boot","state":"???","mdsmap_epoch":22,"osdmap_epoch":0,"osdmap_epoch_barrier":0}

Fixes: http://tracker.ceph.com/issues/19201
Signed-off-by: Patrick Donnelly <pdonnell@redhat.com>

Merge pull request #13411 from yehudasa/wip-vstart-rgw-fix

vstart: don't configure rgw_dns_name

Reviewed-by: Casey Bodley <cbodley@redhat.com>

qa/suites/upgrade/jewel-x/stress-split-erasure-code: box thrashosds

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

qa/suites/upgrade/jewel-x/stress-split: finish client.0 upgrade too

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

osd: rename failsafe [near]full getters appropriately

...and make most of these methods private to clarify the public
interface

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

osd/OSDMap: add [near]full_ratio to OSDMap[::Incremental]

This used to live in PGMap; we're moving it here for luminous
(which makes more sense anyway!).

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

osd: add per-osd FULL and NEARFULL state bits

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

Merge pull request #13425 from dzafman/wip-15912

osd: various changes for preventing internal ENOSPC condition

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

Merge pull request #13783 from cbodley/wip-18980

rgw: if user.email is empty, dont try to delete

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

Merge pull request #13379 from jecluis/wip-mon-luminous-features

mon: enable luminous monmap feature on full quorum

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

Merge pull request #13582 from guzhongyan/cleanup_execute_ctx

osd: correct the func name in execute_ctx() log messages

Reviewed-by: Kefu Chai <kchai@redhat.com>

Merge pull request #13699 from yuyuyu101/wipasync-name

msg/async: set thread name for msgr worker

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

Merge pull request #13726 from dachary/wip-crushtool-json

crush: add --dump to crushtool

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

Merge pull request #13730 from idryomov/wip-osd-rm-incremental

osd/OSDMap: don't set weight to IN when OSD is destroyed

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

Merge pull request #13715 from wangzhengyong/master

os/bluestore: remove unused local variable "pos"

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

Merge pull request #13728 from ifed01/wip-bluestore-fix-smallwrite

os/bluestore: fix a bug in small write handling on sharded extents

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

Merge pull request #13736 from rzarzynski/wip-bs-bitmap-exact-type-in-leaf

os/bluestore: leverage the type knowledge in BitMapAreaLeaf.

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

Merge pull request #13743 from rzarzynski/wip-bs-bitmap-zone-direct-vector

os/bluestore: remove unneeded indirection in BitMapZone.

Reviewed-by: Igor Fedotov <ifedotov@mirantis.com>
Reviewed-by: Sage Weil <sage@redhat.com>

msg/async/rdma: destroy QueuePair if needed

Signed-off-by: Haomai Wang <haomai@xsky.com>

test / librbd: move non-public api usage tests to internal

Signed-off-by: Venky Shankar <vshankar@redhat.com>

test / librbd: create_image_full_pp() / clone_image_pp() helpers

Signed-off-by: Venky Shankar <vshankar@redhat.com>

bluestore, NVMEDEVICE: Specify the max io completion in conf

I do not think that 4 is meaninful. If a value needs to be determined,
it should be configured in the configuration file.

Signed-off-by: optimistyzy <optimistyzy@gmail.com>

os/filestore: use existing variable to assign newsize.

Signed-off-by: Pan Liu <liupan1111@gmail.com>

Merge pull request #13228 from trociny/wip-qa-rbd_mirror-opts

qa/workunits/rbd: tweak rbd-mirror config to spead up testing

Reviewed-by: Jason Dillaman <dillaman@redhat.com>

common: remove \n on clog messages

Confirmed these aren't needed, and are inconsistently included.
Remove them where they do appear.

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

Merge pull request #13772 from Liuchang0812/wip-add-override-in-mgr-headers

mgr: add override in headers

Reviewed-by: Kefu Chai <kchai@redhat.com>

os/bluestore: enable SSE-assisted CRC32 calculations in RocksDB.

By default RocksDB extensively employs CRC32. It has two paths
for the checksum calculation:
 * rocksdb::crc32c::Slow_CRC32,
 * rocksdb::crc32c::Fast_CRC32.

The fast path depends on a run-time discovery of CPU capabilities
AND a compile-time define __SSE4_2__. Although my systems really
offer SSE4.2 support, the macro was undefined resulting in poor
RocksDB performance visible especially during WAL transactions.

The patch (awkwardly) adds the -msse4.2 to CXXFLAGS for RocksDB.

Signed-off-by: Radoslaw Zarzynski <rzarzynski@mirantis.com>