Change retry to num times 60.  Testing shows that on a known dataset, this has the same growth behavior as 2.4.39, while num times 20 resulted in significant growth.

Leave it in main env struct, it doesn't need to be rolled back
with nested txns.

Don't walk the readers table unless we really need to.

we already checked for data == NULL, drop unnecessary if

Move CACHEFLUSH, avoid unnecessary if()

Of course, MacOSX only has posix_memalign, and only in recent versions.
(Older versions could use valloc instead. That precludes ever changing
the LMDB pagesize to be different from the OS pagesize.) Not all that
crucial since it only affects mdb_env_copyfd1.

Leaving them in dirty_list caused breakage.  Instead merge them
into me_pghead at the end, when no more pages will be allocated.

Don't let dummy go out of scope

Everything has it. Solaris doesn't have posix_memalign.

Do a cacheflush after writing data pages, separate from meta's cacheflush

Paranoia for 36300668 do the
cacheflush before setting mti_txnid.

FIRST_DUP/LAST_DUP should succeed even on non-duplicate values
PREV_DUP should reset cursor's EOF flag.

MIPS chips require manual control of on-chip caches. The cacheflush
syscall being used here only exists on MIPS Linux, other OSs will
require revisiting.

return MDB_NOTFOUND if the cursor is empty

Always write mapsize = max(other metapage:mapsize, env:mapsize).

Revert "ITS#7789 Fix resize vs MDB_NOMETASYNC, and a comment."
Revert "ITS#7789 more"
Revert "Revert "ITS#7789 update RESIZED errmsg text""
Revert "ITS#7789 update RESIZED errmsg text"
Revert "ITS#7789 persist mapsize changes"

This reverts commit 02285aca.
This reverts commit 38e71c68.
This reverts commit a5e4eecb.
This reverts commit d8e18551.
This reverts commit 5926e54b.

Don't pass non-NULL map address unless MDB_FIXEDMAP was used.

Must also set in sub-cursor if deleting entire subDB

Always set the filesize when opening for writes. Otherwise
can't use backups from mdb_copy.

We mainly use Win32 functions, so cannot just use C-runtime strerror
to return error messages. We have to use Win32-specific messages.
Unfortunately, we document the API to return C-runtime error codes,
so we can't just switch to all Win32 error codes.

Off-by-one in end check, would complain on valid input.

persist size changes from env_open() as well. This was the
original behavior.