Revert the yesterday's patch that installs SIGCHLD
handler to early collect <defunct> server processes
that received SIGSEGV/otherwise somehow terminated
abnormally.

The invinite loop in Tarantool_connection::recvall()
was caused by misuse of the Python socket API,
not by the fact that the sending side was still
"alive".

The correct fix simply checks for EOF received from
the other end and raises an exception if
we expected some data and got a EOF.

Iron out thet sigchld handler.

If Tarantool server segfaults when a test is
running, it used to hand in <defunct> state,
since there was no wait() call in the parent
to collect child process state. All file descriptors,
as well as open sockets, of a <defunct> process were
left open. This lead to the test runner cycling
in an infinite loop in recv() from the server,
and the whole test suite to, ultimately, hanging
in a busy wait.

This looks more like a bug in python pexpect,
but install a SIGCHLD handler to work it around,
and collect child data immediately as soon as
child terminates.

A fix and a test case for
https://bugs.launchpad.net/tarantool/+bug/730593
"Bad data if incomplete tuple"

A partial tuple could get accepted by INSERT
command, and later on show up in all search results
on the index for which this partial tuple was missing
keys.

When inserting into a TREE index Tarantool/Box did not
validate that the tuple which is being inserted actually
has fields used in the index, and would insert instead
an "ASTERISK" (wildcard) value into the index.

Additionally, when such validation was added, it turned
out that box_raise leaves a corrupt tuple in place
since txn_abort() doesn't remove the unless it's marked
as "GHOST", and tuples were marked as "GHOST" after
index insert, not before.

Conflicts:
	CMakeLists.txt
	core/coro.c
	core/fiber.c
	core/palloc.c
	include/fiber.h

A fix and a test case for

https://bugs.launchpad.net/tarantool/+bug/729758
"SELECT fails with a disjunct and small LIMIT"

and

https://bugs.launchpad.net/tarantool/+bug/729879
"Zero limit is treated the same as no limit"

These were simple coding bugs in SELECT main loop.

Conflicts:
	test/lib/test_suite.py

Allow disjunctions in SELECT WHERE clause and LIMIT
clause in SELECT, to be able to cover the problem
reported in Bug#729758.

Fix a bug with prepare_gdb arguments (remove an extra
space in -e option).

Add a simple test for Bug#729758.

The following bugs were reported in the process:

Bug#729789 "Zero limit is treated the same as no limit"
Bug#730593 "Bad data if incomplete tuple"
Bug#730613 "SELECT with a disjunction on multiple keys returns
matching tuples twice"

We don't have to reset struct sigaction
before every call to sigaction().

Use memset() to zero-initialize sigaction,
since there can be platform-specific memebers.

Conflicts:
	core/tarantool.c

A fix and a test casee for
http://bugs.launchpad.net/bugs/727174
"tarantool_silverbox crashes when saving snapshot on SIGUSR1"

Initialize the event loop before registering a child wait
callback.
Make we don't use fiber infrastructure in the signal handler,
which has its own stack.

@todo: Deal with this:

/build/buildd/binutils-2.20.1/builddir-single/bfd/../../bfd/plugin.c:127:
warning: Using 'dlopen' in statically linked applications requires
at runtime the shared libraries from the glibc version used for
linking
libltbox.a(tarantool.c.o): In function `main':
/home/kostja/work/tarantool/core/tarantool.c:453: warning: Using
'getpwnam' in statically linked applications requires at runtime
the shared libraries from the glibc version used for linking

"entering event loop" is printed into tarantool log
before we have successfully bound the TCP ports.
Expect for a correct message when trying to find
out whether or not the server has successfully started.

 - Replace strndrupa with alloca + memcpy.
 - In create_pid function do rewind before read from pid file.

A fix and a test case for Bug#726778 "Gopt broke wal_dir and
snap_dir: they are no longer relative to work_dir".

Ensure that snap_dir and xlog_dir can be relative to
work_dir.
Ensure that chdir happens before we handle --init-storage.
Do not mix declarations and definitions for recovery_state.
Add a test case to args.test.

Implement a number of simple fixes to compile on FreeBSD:
- work correctly with a GCC version that doesn't have
  -Wno-unused-result
- compile even if not GNU libc
- tee doesn't have '--append' option on FreeBSD, use
  the short version ('-a').
- fix a typo in log_io.c, ENOENT, not ENONET.

Implement
https://blueprints.launchpad.net/tarantool/+spec/consistent-header-guards

Audit all header files in tarantool code base for consistent use
of header guards:
- consistent naming scheme, following doc/coding-style.txt
- header guard goes before the copyright
- each header is self-sufficient.
- add copyright headers where they were missing.

Remove __packed__, __unused__, __noinline__ and __regparm2__ from util.h,
since the name __unused__ conflicted with a parameter name in
third_party/valgrind/valrgind.h, which led to a dependency in inclusion
order between util.h and this header. The define names were not standard, I
couldn't come up with some meaningful and standard names, and they were only
saving some typing, while not providing any extra functionality.

Rename 'silverbox' to 'box', across the board.
Delete the remnants of the old build system
(scripts/rules.mk and scripts/config-def.mk).

Do not touch client/perl for now, since it's maintained
elsewhere.

Return a non-zero exit status if there are test
failures, even if run with --force. This necessary to
correctly identify a test failure in buildbot run.

Make 'make test' goal operational on hosts with old
python or python somewhere else than /usr/bin.

Correct a regexp pattern in test suppression filter
to work in a more modern version of python.
In the new python \S matches '"', and the test
output is different compared to the recorded one.

Adjust regexp pattern in one of the tests that
started to fail after an increment in version number.

Start the next development milestone.

Add the simplest CPack configuration for
'make package' and 'make package_source'
targets.
Add an empty snapshot to revision control
to use in 'make package' and thus
spare the user from the need to do
--init-storage.
Modify args.test to accept a bare
(no git version) version number, and thus
ensure the user from a yet another call before
he/she can start work/play around.