The official macro is __ANDROID__; ANDROID may or may not be defined.

Use the same size dirty list for both 64 and 32 bit.

Re-fix 6b1df0e4 from ITS#8406

Attempt to stop liblmdb variants compiled with conflicting options
from using the lockfile at the same time and thus breaking it.

Simpler usage since it's only called once, rename to match new usage,
and drop 3 loop pessimizations which were optimizations 20 years ago.

This limits the namespace which the user can meddle with for
POSIX semaphores and Windows mutexes.  Their names change a
bit, they no longer have fixed lengths.

since it's variable size on Linux/glibc

Don't try to deref cursor page if txn's pagelist is empty

Fully revert the change to GET_MULTIPLE

GET_MULTIPLE was broken

Allow C_EOF flag to be stale

Optimize mdb_page_search_root(PS_LAST) when cursor is already near
last position, ignoring C_EOF flag for now.

Doesn't have POSIX robust mutexes - GNU userland on BSD kernel

- DISTRIBUTE_GROUP_DOC makes doxygen give several fields the
  same doc: mn_hi + mn_lo in MDB_node.
- With mdb_mutex_t + mdb_mutexref_t, instead split them up.
- Don't hide a doxygen #name inside double quotes.

Never clear mc_error, we could lose a failure in the other thread.

Lorenz Bauer authored Oct 20, 2016 and Hallvard Furuseth committed Dec 13, 2016

Return EPIPE instead.

Don't use it as a "cursor is tracked" hint in mdb_pages_xkeep().
It's been harmless so far, but would break after mdb_cursor_copy().
Checking m0 directly short-circuits better anyway.

Drop ~C_EOF, pointless after 8c215aa9

Other functions depend on them to do so.
For mdb_node_read(), instead remove such a dependence.

Init me_rpmutex independently of MDB_NOLOCK.
Plug leaks on mdb_env_open() failure.
Tweak mdb_env_close0() to handle the rearranged mdb_env_open().

xcursor fixup depends on init state

Revert excess cursor fixup