Instead of palloc allocator use a member of small allocators family,
region_alloc.

The plan is to remove salloc as well, and switch entire tarantool
to libsmall.

Fix a race condition in replication/swap.test, discovered with the
new allocator.

Do not use atexit() handlers, except the global tarantool_free()
function, which does nothing in the snapshot process.

Review fixes:

- New configuration variable 'replication_protocol'
is defined in the configuration file. The original patch
would hand-craft it into generated files.
- Simplified  numeric version id generation.
- Removed timeouts from new sio_* calls. They did't quite work,
especially with blocking sockits. Timeouts should at least use
select(), which updates wait time on exit..
- Removed include cfg/tarantool_box_cfg.h into recovery.cc, instead
passed around all arguments explicitly.
- Removed unused headers.
- reused handshake code in master and replica
- simplified handshake in replica fiber
- removed clutter
- touched here and there.

Remove get_socket_addr(), use existing sockaddr_in
to save in the cookie.

On snapshot, increase snapshot version id (u32),
save current snapshot version within a tuple.

On tuple free, check if tuple version is equal to current
snapshot id, if so do direct free, otherwise append tuple
to the delayed free list.

Fix https://github.com/tarantool/tarantool/issues/44

Fix C++ compile error of libsmall.
Remove bogus assert() redefinition (why would anyone do such a thing?)
Link tarantool_box with libsmall.

Provide open_memstream() implementation for system which
miss it (Mac OS X).

Allocate slab arena in shared memory to reduce pages splits during
snapshotting.
Adaptation of patch made by Yuri Nevinitsin @nevinitsin.

Implement tuple validation when building a new index.
This introduces such checks as for field type, presence of a field,
etc. in all tuples added to the new index.

Add error code ER_INDEX_ARITY, update tests.

Add a check that the primary key is always unique.

Test unique -> non-unique transition.

Fix memory leaks this test has revealed.

 * sys/prctl.h can be found only on Linux, but cmake recognizes Debian/kFreeBSD
 as Linux.
 * extern "C" for TIMEOUT_INFINITY (it is used in plugins)

 * sys/prctl.h can be found only on Linux, but cmake recognizes Debian/kFreeBSD
 as Linux.
 * extern "C" for TIMEOUT_INFINITY (it is used in plugins)

https://blueprints.launchpad.net/tarantool/+spec/auto-init-storage

What happened in production, during upgrade:
- logger option in the config file was pointing to nowhere
- the server would start, unable to check that the child logger
process died prematurely
- the server would fork spawner process to spawn replication relays
- a replica would come in, asking for data
- the spawner would successfully spawn a relay but die on SIGPIPE
when trying to write to the log (a pipe leading nowhere by now)
about the event
-> we have a master with a dead spawner process, no more replicas,
sorry.

Well, someone should check their configuraiton in the first place.
Nevertheless, try to be nice, and:

- don't die on SIGPIPE in the spawner process.
- when starting the logger, make an effort to check that it has started
successfully. There is no reliable way to find this out, so do
it with the best effort.

Remove memcached support in 1.6.
We're about to change the type system.
About to add authentication.
This all makes supporting memcache within the core too cumbersome.

Fix confetti error output. Remove redundant cfg_reload(). Make sure we
produce a Lua error on error.
Remove duplicate code.
Style renames.

lua console. Refactor errinj lua bindings.

for proper yaml output