...early enough that txn state is left unchanged.

Move the check to the end of txn startup.  Catches env
breakage which happens while the new txn waits for a lock.

* Add MDB_TXN_FINISHED, MDB_TXN_HAS_CHILD, MDB_TXN_BLOCKED.
* Clear mt_numdbs in writers, for TXN_DBI_EXIST() to catch.
  We already do in readers.

Side effects:
* Clean txn up a bit even before freeing it.
* Tweak DEBUG output at txn end. Add DEBUG after commit(writer).

MDB_DBI_CHANGED(,MAIN_DBI) is never true.

mt_env is always set.
Commit(mt_child) resets mt_child, so parent need not.

So far just to help aligning mdb.master and mdb.RE.
We'll make it an option later.

Leonid Yuriev authored 9 years ago and Hallvard Furuseth committed 9 years ago

The rest of 9a80a8a8
plus commit a937740a.

mdb_xcursor_init1(): md_pad is only used when MDB_DUPFIXED.
mdb_page_split():    Copy mp_pad too. Used by mdb_page_list().

More fallout from 4d02c741.

Taken from mdb_txn_begin(,,MDB_NOMETASYNC, MDB_NOSYNC,,)
without adding those two flags yet, to align with mdb.master:
Part of 54516639 (Renumber...)
+ 8adee946 (Fix per-txn...).

Leonid Yuriev authored 10 years ago and Hallvard Furuseth committed 9 years ago

commit 9a8eb956 (mdb_txn_renew0)
plus b2ab9910 (reader allocation)
from mdb.master.

David Barbour authored 10 years ago and Hallvard Furuseth committed 9 years ago

commit b660491d (mdb_txn_id())
plus c36c167c (fix prev commit)
from mdb.master.

Hallvard Furuseth authored 10 years ago and Hallvard Furuseth committed 9 years ago

Check new mapsizes against mm_last_pg.  Move
mdb_env_init_meta0() so it can set mm_last_pg earlier.

node_del shifts nodes around, cursors pointing at fake pages
need to have their mc_pg[0] corrected.

Some destination fixups need to happen immediately after nodes
are moved, before rebalancing

The tree height can also increase during rebalance, not just shrink.
This can happen if update_key needs to split a parent branch page.