qa/suites/rados/basic/tasks/rados_api_tests: debug ms = 1

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

osd: periodically request newer map from mon if waiting peering events

If we have peering events waiting on a newer map than we have, request it
from the mon.  Do this periodically in tick so that we normally wait to get
it from a peer first.

This avoids a deadlock situation where we are, say, waiting for a newer
map to create a pg or but do not ever get the map to do it (because the
cluster is idle).

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

osd: use rctx transaction for PG removal

In the normal case, queue up the removal work on the rctx transaction.

For the final cleanup, since we need to block, dispatch it ourselves, and
do not do so in OSD.cc.

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

osd: some debug output in identify_split_children

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

osd/PG: do final pg delete transaction on pg sequencer

Simpler, cleaner.  Also, this way we flush before returning.

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

osd: better debug output in identify_splits

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

osd: handle NOUP flag vs boot race

If we digest maps that show a NOUP flag change *and* we also go active,
there is no need to restart the boot process--we can just go/stay active.

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

qa/suites/rados/singleton/all/recovery_preemption: make test more reliable

A 30 second run did only 7000 ops, which means ~50 log entires per pg...
not enough to trigger backfill.

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

qa/suites/rados/singleton/all/mon-seesaw: whitelist PG_AVAILABILITY

The seesaw might delay pg creation by more than 60s.

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

osd/PG: ensure an actual transaction gets queued for recovery finish

Otherwise, this context gets leaked and lost.

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

osd: close split vs query race in consume_map

Consider the race:

- shard 0 consumes epoch E
- shard 1 consumes epoch E
  - shard 1 pg P will split to C
- shard 0 processes query on C, returns DNE
- shard 0 primes slot C

Close race by priming split children before consuming map into each
OSDShard.  That way the query will either (1) arrive before E and before
slot C is primed and wait for E, or find the slot present with
waiting_for_split true.

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

osd: improve documentation for event queue ordering and requeueing rules

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

osd/PG: flush sequencer/collection on shutdown

This should catch any in-flight work we have.

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

osd/PG: move shutdown into PG

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

osd/osd_types: fix pg_t::pool() return type (uint64_t -> int64_t)

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

mon/OSDMonitor: disallow pg_num changes until after pool is created

The pg create handling OSD code does not handle races between a mon create
message and a split message.

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

osd/PG: set send_notify on child

If we are a non-primary, we need to ensure the split children send
notifies.

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

osd: kill broken _process optimization; simplify null pg flow

- drop fast quuee to waiting list optimization: it breaks ordering and is
a useless optimization
- restructure so that we don't drop the lock and revalidate the world if
pg == nullptr

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

osd: fix fast pg create vs limits

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

osd: (pre)publish map before distributing to shards (and pgs)

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

osd: update numpg_* counters when removing a pg

Usually on a pg create we see an OSDMap update; on PG removal completion
we may not.

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

osd: decrement deleting pg count in _delete_some

The exit() method for ToDelete state doesn't run on PG destruction.

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

osd: clear shard osdmaps during shutdown

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

osd: make save osdmap accessor for OSDShard

The advance_pg needs to get the shard osdmap without racing against
consume_map().

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

osd: clean up mutex naming for OSDShard

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

common/tracked_int_ptr: fix operator= return value

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

osd: fix pg removal vs _process race

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

osd: lookup_*pg must return PGRef

Otherwise it is fundamentally unsafe, as the PG might get destroyed out
from under us without a reference.

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

osd: kill pass-through _open_pg

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

osd: remove old min pg epoch tracking

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

osd/PG: remove RecoveryCtx on_applied and on_commit

These were awkward and unnecessary.

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

osd/PG: register delete completion directly on Transaction

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

osd: register split completion directly on Transaction

No need to use wonky RecoveryCtx C_Contexts

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

osd/PG: drop unused context list accessors for RecoveryCtx

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

osd/PG: register recovery finish context directly on Transaction

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

osd/PG: drop unused activate() context list arg

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

osd/PG: register flush completions directly on the Transaction

No need to awkward list passed as an arg; all of these callbacks end up
on the Transaction anyway.

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

osd: wait for pg epochs based on shard tracking

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

osd: index pg (slots) by map epoch within each shard

This will replace the epoch tracking in OSDService shortly.

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

osd/PG: link back to pg slot

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

osd: OSDShard::pg_slot -> OSDShardPGSlot

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

osd: change pg_slots unordered_map to use unique_ptr<>

This avoids moving slots around in memory in the unordered_map... they can
be big!

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

osd: remove some unused methods

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

osd: remove created_pgs tracking in RecoveryCtx

Not needed or used!

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

osd: fix PG::ch init

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

osd: use _attach_pg and _detach_pg helpers; keep PG::osd_shard ptr

Consolidate num_pgs updates (and fix a counting bug along the way).

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

osd: remove old split tracking machinery

This infrastructure is no longer used; simpler split tracking now lives in
the shards pg_slots directly.

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

osd: restructure consume_map in terms of shards

- new split primming machinery
- new primed split cleanup on pg removal
- cover the pg creation path

The old split tracking is now totally unused; will be removed in the next
patch.

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

osd: pass sdata into dequeue_peering_evt (and dequeue_delete)

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

osd: pass data into OpQueueItem::run()

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

osd: kill pg_map

Split doesn't work quite right; num_pgs count is probably off.  But, things
mostly work.

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

osd: rename OSDShard waiting_for_pg_osdmap -> osdmap

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

osd: use _get_pgs() where possible; avoid touching pg_map directly

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

osd: get _get_pgs() and _get_pgids()

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

osd: remove get_mapped_pools command

No in-tree users.

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

osd: move ShardedOpWQ::ShardData -> OSDShard

Soon we will destroy pg_map!

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

osd: kill _open_lock_pg

Move lock call to caller.

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

osd: kill _create_lock_pg

Unused.

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

osd: do not release recovery_ops_reserved on requeue

This doesn't make sense.. although it's the same behavior as
luminous.

The point of the releases here is that if we drop something that is in
the queue we drop the recovery_ops_reserved counter by that much.  However,
if something is in the queue and waiting, and we wake it back up, there
is no net change to _reserved... which is only decremented when we
actually dequeue something.

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

osd: debug recovery_ops_reserved

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

osd: move PG peering waiters into op wq

This resolves problems with a peering event being delivered triggering
advance_pg which triggers a requeue of waiting events that are requeued
*behind* the event we are processing.  It also reduces the number of
wait lists by one, yay!

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

osd: store ec profile with final pool

We need this to reinstantiate semi-deleted ec backends.

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

osd/PG: ignore RecoveryDone in ReplicaActive too

This can be missed on a RepRecovering -> RepNotRecovering ->
RepWaitBackfillReserved transition.  Catch any straggler events in
ReplicaActive.

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

osd/osd_types: include epoch_sent in pg_query_t operator<<

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

osd: restructure pg waiting more

Wait by epoch.  This is less kludgey than before!

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

osd: restructure pg waiting

Rethink the way we wait for PGs.  We need to order peering events relative to
each other; keep them in a separate queue in the pg_slot.

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

osd: normal command uses slow dispatch (it can send messages)

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

osd/OSD,PG: get_osdmap()->get_epoch() -> get_osdmap_epoch()

Avoid wrangling shared_ptr!

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

osd: misc fixes

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

osd: kill disk_tp, recovery_gen_wq

Progress!

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

osd: move recovery contexts to normal wq

We have a specific PGRecoveryContext type/event--even though we are just
calling a GenContext--so that we can distinguish the event type properly.

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

osd: remove _ookup_lock_pg_with_map_lock_held()

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

osd: new MOSDScrub2 message with spg_t, fast dispatch

Send new message to mimic+ OSDs.  Fast dispatch it at the OSD.

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

osd/PG: request scrub via a state machine event

Continuing effort to make PG interactions event based.

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

osd: use peering events for forced recovery

The mgr code is updated to send spg_t's instead of pg_t's (and is slightly
refactored/cleaned).

The PG events are added to the Primary state, unless we're also in the
Clean substate, in which case they are ignored.

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

osd/OSDMap: get_primary_shart() variant that returns primary *and* shard

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

osd: prime pg_slots for to-be-split children

Once we know which PGs are about to be created, we instantiate their
pg_slot and mark them waiting_pg, which blocks all incoming events until
the split completes, the PG is installed, and we call wake_pg_waiters().

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

osd: remove obsolete slow dispatch path for most messages

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

osd: fast dispatch M[Mon]Command

These just get dumped onto a work queue.

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

osd: fast dispatch ping

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

mon/OSDMOnitor: send MOSDPGCreate2 to mimic+ osds

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

osd: handle MOSDPGCreate2 messages (fast dispatch!)

Add a new MOSDPGCreate2 message that sends the spg_t (not just pg_t) and
includes only the info we need.  Fast dispatch it.

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

osd/OSDMapMapping: a getter that returns a spg_t

Note whether a pool is erasure so that we can generate an appropriate
spg_t for a mapping.

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

osd: send pg creations through normal pg queue

Queue a null event tagged with create_info, elimiating the special
legacy path.

These are still not fast dispatch because we need an spg (not pg) to queue
and event, and we need a current osdmap in order to calculate that.  That
isn't possible/a good idea in fast dispatch.  In a subsequent patch we'll
create a new pg create message that includes the correct information and
can be fast dispatched, allowing this path to die off post-nautilus.

Also, improve things so that we ack the pg creation only after the PG has
gone active, meaning it is fully replicated (by at least min_size PGs).

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

osd: fix max pg check for peer events

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

osd: use atomic for pg_map_size

This avoids the need for pg_map_lock in the max pg check.

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

osd/PGPeeringEvent: note mon- vs peer-initiated pg creates

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

osd: fast dispatch peering events (part 2)

This actually puts the remaining peering events into fast dispatch.  The
only remaining event is the pg create from the mon.

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

osd: fast dispatch peering events (part 1)

This is a big commit that lays out the infrastructure changes to fast
dispatch the remaining peering events.  It's hard to separate it all out
so this probably doesn't quite build; it's just easier to review as a
separate patch.

- lock ordering for pg_map has changed:
  before:
    OSD::pg_map_lock
      PG::lock
        ShardData::lock

  after:
    PG::lock
      ShardData::lock
        OSD::pg_map_lock

- queue items are now annotated with whether they can proceed without a
pg at all (e.g., query) or can instantiate a pg (e.g., notify log etc).

- There is some wonkiness around getting the initial Initialize event to
a newly-created PG.  I don't love it but it gets the job done for now.

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

osd: queue null events without PG lock

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

osd: move part of wake_pg_waiters into helper

We'll need this shortly.

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

osd: use MTrim peering event for trimming

This is simpler and cleaner than handling log trimming as a special case.

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

osd: fast dispatch backfill and recovery reservation events

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

osd: move M{Backfill,Recovery}Reserve event logic into message

Better encapsulation!

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

messages/MOSDPeeringOp: add

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

osd/PG: move peering event type out of PG class

We will create these directly from peering Messages shortly.

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

osd/PG: keep epoch, not map ref, of last osdmap for lsat persisted epoch

No need to pin the map in memory!

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

osd/PG: remove old update_store_on_load()

This isn't needed post-luminous.

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

Merge tag 'v13.0.2'

v13.0.2

Merge PR #21180 into master

* refs/pull/21180/head:
vstart_runner: examine check_status before error

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