Define MDB_FDATASYNC() instead of redefining fdatasync().

If the database is opened with MDB_RDONLY or MDB_NOSYNC,
instead use the same file descriptor for me_mfd and me_fd.

Also factor out Windows/Unix error handling after open.

Get txn->mt_txnid from the meta page instead.  Otherwise it
and txn->mt_toggle could become inconsistent, since read-only
txns fetched both from MDB_txbody without mutex proteciton.

Make DPRINTF usable in if-else.  Silence signed vs unsigned warning.

Also add mdb_debug/mdb_debug_start to toggle debug output at runtime

Was being fooled before because the newly malloc'd block had
the same address as the previously freed block.

Should probably compare nsize to pg_size/4 instead...

Sub-DBs were correct before

when replacing an existing item

Also allow read access to freelist in mdb_cursor_open

Allow pages from free list to be used when growing the free list.
(Yes, this is self-referential...)

don't skip the cursor adjust checks

Also line up the code logic a bit.

Delay (MDB_txninfo*) cast to after comparing mmap() with MAP_FAILED.
Otherwise, if MAP_FAILED = (void*)-1 but MDB_txninfo requires stricter
alignment, the compiler could assume the result is never MAP_FAILED.

Also store NULL in env->(me_map, me_txns) after mmap failure.

Revealed when gcc optimization is enabled

Track allocations and invalid accesses to freed tmp pages.

Fix for failure when committing a txn that is a child of
some parent txn.

Don't compare cursors that aren't deep enough to match the
current cursor.