Use a sub-DB for DUPSORT item #1/#2 per key if needed: Not a sub-
page too big for a node, nor an overflow page (which not all DUPSORT
code checks for). Move "insert" code, to avoid non-loop goto upwards.
(This is the commit which needs the change to xdata.mv_size in
commit 9d6e4a9163 "page sizes".)
Handle keys stored by a liblmdb with a bigger MDB_MAXKEYSIZE.
mdb_get/mdb_del(absent key bigger than our MDB_MAXKEYSIZE)
now return MDB_NOTFOUND instead of MDB_BAD_VALSIZE.
Change me_nodemax to not count the mp_ptrs[] entry. That's mostly
how it was used. Compare node sizes ">" me_nodemax instead of ">=".
The ">=" was a workaround for confusing sizes with and without the
mp_ptrs[] entry, but broke for nodes with size (old me_nodemax-1).
Explicitly make me_nodemax even. An odd value could break the
comparisons. It was even anyway because MDB_MINKEYS == 2.
Use DB page size = min(32k, OS pagesize). Previous limit was 8k
(MDB_MINKEYS*MDB_PAGESIZE).
Handle DB pagesize < OS pagesize. That's an I/O pessimization,
but transactions remain atomic: Only writing the MDB_meta must
be atomic, and it fits in one OS page.
Don't truncate desired subpage size: Asssign it to a size_t
(mv_size), not an uint16_t (mp_upper).
Both were unused and md_name was unmaintained -- except
mdb_cursor_touch(xcursor) would abuse md_name as a key to
touch MAIN_DBI if it could somehow get passed ! DB_DIRTY.
xcursor DBIs were parent DBI+1 for debugging. Instead output
-(parent DBI). Fixes a crash in mdb_cursor_del0()'s xcursor
tracking, it forgot to subtract 1 for C_SUB cursors.
MDB_txn.mt_toggle: Use (mt_txnid & 1) instead.
Drop error checks which will be repeated.
mdb_cursor_set(): Turn assert into if/return to match the above.
mdb_cursor_del(): 'flags' are now used as bitflags.
Only named DBs can have DB_STALE, and they do not use MDB_PS_MODIFY.
Replace magic key values with flags. Drop duplicated comments at
mdb_page_search_root() vs. mdb_page_search(), and rephrase.
Hallvard Furuseth