mdb_env_open() with these flags did not open a synchronizing
file descriptor.  Thus disabling them later did not work.

Drop mmap()'s MAP_FIXED flag, so it returns another address instead
of unmapping existing overlapping pages. Verify the returned address.

MDB_NOSUBDIR was == MDB_REVERSEKEY. That affected the freelist:
Env flags are stored in mm_flags alias mm_dbs[FREE_DBI].md_flags.

It stays a persistent flag, in case mdb_env_open someday wants
to pick/verify which lockfile to use with the datafile.

Catch bad flags so they will no longer make it into the data file.

Wasn't pushing the optimal half of the array onto the stack,
thus used more stack than expected -> overrun.

This reverts commit ef2092ea.

If no data was modified in the txn, mdb_open flag changes were dropped.

Could return 1 garbage record before actual data, if starting from EOF

Fix #define _GNU_SOURCE = 1 to match any command line -D_GNU_SOURCE.
.gitignore += mdb_copy.

Add flag MDB_ENV_ACTIVE instead of using the state of me_lfd.
Require MDB_RDONLY for MDB_ROFS.  Swap reader/writer mutex lock.

This can happen even on local filesystems.

It only worked if the freelist was already non-empty

if it was the only operation in a txn, it would be ignored.

More for 48ef27b6
Our definition of "large" data item needs to be smaller to
handle more cases where a bad split is possible

return whatever the cursor is currently pointing at

page_split with newindex > split_indx

Use explicit MDB_RDONLY flag

Do not try shared lock when closing or after error.  Track file lock
state to decide.  Change meaning of *excl to reflect file lock state.

Drop the sem_unlink() error checks, which could prevent the 2nd
unlink.  Instead use O_EXCL in sem_open().

This makes "open+close the database" the API for trying to clean
away the old semaphores, if they were left behind by a previous run.