Essentially reverts part of cd8ff376 to
make sure there is always a contextCSN if the server is in charge of its
own serverID.

Add matching rules for all cases where it was missing.  Cleanup
incorrect types for a few attributes as well.  Fix network-timeout
handling in back-ldap/meta/asyncmeta.

And try another approach. Always write contextCSN updates, but
don't set dont_replicate for updates we want propagated.

Let the entryCSN check do all the work. Reloading a server from an old
backup needs this to go thru.

Update syncprov contextCSNs when context entry is added.
Fix accesslog to properly tag Add op when adding context entry.

Keep s_mutex locked until we know we're removed from queue.
Remember qtask cookie so we can retract if ineeded when deleting
the overlay from running slapd.

config_delete is still unsafe, overlay_remove is running with active
threadpool instead of paused pool.

ITS#8527 - Add additional debug logging on consumer/provider state when the consumer has a newer cookie than the provider

behind LDAP_DEVEL

Ondřej Kuzník authored Aug 25, 2017 and Howard Chu committed Aug 25, 2017

When a checkpoint happens, if we remove the CSN from the pending list,
accesslog won't pass it onto the accesslog DB. But in a delta-mmr
scenario, an accesslog entry without a CSN faces a race where it might
be applied twice - that usually fails and causes a full refresh, other
times it can cause a silent desync - both are undesirable.

This reverts commit a5f3a288.

Don't double-queue delete ops

Must setup psearch before snapshotting ctxcsn

Delay mutex init to avoid leaking the mutex

Check for abandon just before recording psearch