For the same reason why 'uri' was moved to src/lib - SWIM needs
core and uuid, and SWIM will live in src/lib.

This commit follows 'uri' relocation as a separate one because
'uri' relocation required some changes in the files, moved by
this commit.

Needed for #3234

Http_parser in fact does not depend on anything, even on the core.
As a rule, such basic libraries are stored in src/lib.

The patch is not necessary for anything, but it is a right thing
to do, while some activity is happening there.

Http_parser is a standalone library, which in fact does not
require httpc. And because of that it is going to be moved into
lib/http_parser.

URI and core libraries are going to be used by SWIM, stored in
src/lib. But src/lib can not depend on src/. This patch pushes
URI library down to src/lib - the lowest level of source
dependencies.

Needed for #3234

Memtx uses separate mempools for iterators of different types.
Due to the fact that there will be more iterators of different
sizes in a series of upcoming changes, let's always allocate the
iterator of the largest size.
No changes have been made to the rtree iterators pool because the
size of these structures is significantly larger.

Needed for #3961

This can't entail any consequences, because socket fd is set to -1 in
this case, but this just looks a bit messy. Let's clean it up.

Follow-up commit 305dbcf6 ("iproto: close socket explicitly before
wal_dir at exit").

When we replaced SQLite's ephemeral spaces with our ones, one
optimization concerning ORDER BY clause was disabled. It allows to
reduce number of fields in format of ephemeral space or sorter table.
To illustrate how it works, consider example:

CREATE TABLE t (id INT PRIMARY KEY, b INT);
SELECT * FROM t ORDER BY b;

To sort entries from t, ephemeral space with format [b, id, b] is
created. One can see, that such format contains duplicate of b column.
To avoid such situation, SQLite provides optimization which removes
duplicates. Meanwhile, it doesn't change already set format of ephemeral
(or sorter) space (and SQLite tolerates that format difference). That's
why it was turned off. However, such optimization turns out to be not
optional but required: some column values shouldn't be computed twice.
For instance:

SELECT random() AS x FROM t ORDER BY x;

Without filtering fields from ephemeral space format, it would be like:
[random(), random()]. In other words, results would be sorted by first
call to random() function, but resulting set would consist of values
given by second call of random(). So, to enable it, we should reduce
field count in format of ephemeral space by number of matches between
SELECT and ORDER BY column lists.

Also, type of return value for random() function has been fixed.

Closes #3783

Refactored memtx_tree code so that memtx_tree.h, memtx_rtree.h,
memtx_bitset.h, memtx_hash.h contained only the signature of the
tree object constructor while all implementation details were in
corresponding *.c files.

Needed for #3961

Introduce a macro BPS_TREE_IDENTICAL for BPS TREE class. This
makes possible to define custom comparators for stucture-based
leafs.
Previously, a C++ comparison operator "!=" override was used for
this purpose. Due to the fact that we are not going to rework on
C++ C-only components of Tarantool like memtx_tree, we needed a
way to make complex structures comparisons using preprocessor.

Needed for #3961

The bps_tree.h header defines the macro _api_name, but does not
undefine it at the end. Fixed.

tarantool instance didn't close socket explicitly
which could cause hot standby instance to fail to bind
in case it tries to bind before socket is closed by OS.
Now it is fixed by closing socket explicitly before wal_dir.

Closes #3967

Add preprocessing of the request headers. Each header must be 'string' or 'table'
with '__tostring' metamethod.

Closes #3679

If I create a new fiber and the join to himself we get an abort:

 | tarantool> f = require('fiber')
 | ---
 | ...
 |
 | tarantool> f.join(f.self())
 | tarantool: src/fiber.c:407: fiber_join: Assertion `fiber->flags & FIBER_IS_JOINABLE' failed.
 | Aborted (core dumped)

we should better throw an error.

The only goal of reading and writing heap_node.pos was checking
if a node is now in a heap, or not. This commit encapsulates this
logic into a couple of functions.

Replace assert in socket:read() and add a check in socket:sysread()
when the size is negative.

Closes #3979

It is possible to create vinyl tables, if pragma sql_default_engine is
set to corresponding engine. However, such spaces don't inherit global
vinyl-specific options which can be set by box.cfg{}. Lets fix it and
during encoding of index options fetch appropriate configurations from
cfg.

Closes #3912

Since previous commit allows us to raise an error during msgpack decode
inside VDBE, lets do this if decoded integer is out of
[INT64_MIN, INT64_MAX] range and set "integer is overflowed" diagnostic
message.

Closes #3735
Workaround for #3810

Tarantool allows to hold in INTEGER field values in range
[INT64_MAX, UINT64_MAX], which is obviously larger than common int64_t
range. In this regard, if value of integer field in range
[INT64_MAX, UINT64_MAX] is presented in tuple (e.g. after insertion from
Lua interface), then after decoding msgpack (during processing SQL
query) its value won't fit into int64_t (which in turn is basic type
used to hold integers inside VDBE memory). Now if this happens, instead
of raising an overflow error, value is stored as floating point number
(with precise loss, obviously). Such approach is considered to be messy
and we are going to raise "integer overflow" error instead. To make this
happen, lets firstly refactor sqlVdbeMsgpackGet() to make it return
error code to indicate that something went wrong and move length of
decoded part to output parameters. Codestyle fixes are included as well.

Needed for #3735

Before this patch, if integer overflow occurred during casting to
integer, it was simply ignored. As a result, wrong results might take
place. Lets check possible overflows before CAST conversion, and if it
happens, raise an appropriate error.

Part of #3735

Before this patch, if integer was overflowed during math operations
(OP_Add, OP_Subtract, OP_Multiply, OP_Divide), it would be implicitly
converted and stored as floating point number. This is obviously wrong
way to handle integer overflow errors. Instead, let's raise
corresponding error.

Part of #3735

Before this patch, resulting type of concatenation operator always was
TEXT (i.e. type of memory cell containing result - MEM_Str). Lets fix it
and return type depending on type of concatenation arguments: if both
arguments are TEXT, then resulting type is TEXT; BLOB otherwise. Note
that other options of combining types of arguments are illegal.

Closes #3544

Original SQLite operator of concatenation accepts all types of
arguments. If type of parameter is not TEXT, it is implicitly converted
to TEXT (except for NULLs). That contradicts ANSI (it is regulated by
[1]), so lets allow only TEXT and BLOB as argument type for
concatenation.  Moreover, they both must be of the same type at the same
time (i.e. both TEXT or BLOB).

[1] SQL ANSI 2013, 9.5 Result of data type combination

Part of #3544

Accidentally, mask which is used to map type of VDBE memory cell into
outer API types was replaced with MEM_TypeMask. But value of the latter
is larger then possible values of VDBE memory cells types. Hence, if it
is applied to memory cell, not only pure types is taken into
consideration, but some additional flags (such as MEM_Zero) as well.
Overall, it results in wrong type calculation for zeroed blobs, for
instance. Lets return back previous mask.

Follow-up #3698
Needed for #3544

yaml.encode() now wraps a string literal whose content is equal to a
null or a boolean value representation in YAML into single quotes. Those
literals are: 'false', 'no', 'true', 'yes', '~', 'null'.

Reverted the a2d7643c commit to use single quotes instead of multiline
encoding for 'true' and 'false' string literals.

Follows up #3476
Closes #3662
Closes #3583

Use lua_is*() functions instead of explicit lua_gettop() checks in
yaml.encode() and yaml.decode() functions.

Behaviour changes:

* yaml.decode(object, nil) ignores nil (it is consistent with encode
  behaviour).
* yaml.encode() gives an usage error instead of "unsupported Lua type
  'thread'".
* yaml.encode('', {}, {}) ignores 3rd argument (it is consistent with
  decode behaviour).

Previously readahead was set after the instance synced with all existing
masters. This could lead to a case when some client connections created
rather early would start with default readahead value even when
it was explicitly set to some value.
So, start setting readahead before replicaset sync.

Follow-up: #3958

Iproto connections keep old readahead values for input buffers even
after box.cfg.readahead reconfiguration. This means that for the changes
to take place for the old clients, they have to reconnect. Otherwise
tarantool log will be spammed with 'readahead limit is reached' errors.

To fix this, start updating input buffer size for iproto connections
if needed every time the buffer is empty.

Closes: #3958

Checks for fiber_name legth worked with FIBER_NAME_MAX,
while one byte in fact was occupied by terminating zero.

Closes #4011

vy_run_iterator_seek left the return statement uninitialized under
certain conditions. The branch in the code wasn't covered by any test,
because it could only be triggered by an EQ SELECT following range
coalescing, so we only saw it when a customer reported a crash.
Fix it and improve vinyl/select_consistency test to cover this branch.

Reserve() function by definition should ensure that there is
enough space to store something. But heap reserve() always just
increases the capacity twice. Even if there was already enough
memory. Check for not necessary realloc was done out of this
function. This was wrong. The commit makes reserve() be real
reserve().

The commit is motivated not by pursue of justice, but by
forthcoming usage of reserve() in SWIM module code.

Now heap API works with struct heap_node only, which forces a
user to constantly call container_of. Such a code looks really
awful. This commit makes heap taking and returning user defined
structures, and removes container_of clue.

It is worth noting, that the similar API rb-tree and b-tree
have. Even rlist has its rlist_*_entry() wrappers, and mhash
provides macroses to define your own value type.

During DDL routines we pass encoded space/index/trigger formats
into msgpack to VDBE. EXPLAIN query displays arguments of each opcode of
VDBE program in a readable format. So, lets decode arguments of OP_Blob
opcode with subtype = _MSGPACK before displaying them. Also, lets
enlarge static buffers for P4 operand value and opcode comment to fit
decoded msgpack.

What is more, it fixes buffer-overflow since before this patch operands
of OP_Blob were treated as strings and passed to functions like strlen()
(only during EXPLAIN query). On the other hand, generally speaking
msgpack can come without null termination, or contain '\0' symbols in
the middle of encoded array.

Closes #3868

According to ANSI, result of concatenation operation should derive
collation sequence from its operands. Now it is not true: result is
always comes with no ("none") collation.

In a nutshell[1], rules are quite simple:

a) If some data type has an explicit collation EC1, then every data type
that has an explicit collation shall have declared type collation that
is EC1.  The collation derivation is explicit and the collation is EC1.

b) If every data type has an implicit collation, then:

 - If every data type has the same declared type collation IC1, then
   the collation derivation is implicit and the collation is IC1.

 - Otherwise, the collation derivation is none.

c) Otherwise, the collation derivation is none.

[1] Read complete statement at chapter 9.5 Result of data type
combinations, ANSI 2013, Part 2: Foundations.

Closes #3937

Lets make sql_expr_coll() return error code. It is needed since we are
going to use this function to detect operands of concatenation with
incompatible collations. Hence, struct coll* has been moved to output
parameters.  For the same reason lets make sql_binary_compare_coll_seq()
return error code instead of struct coll*. The latter is really required
only in one out of six usages, so lets remove it at all.

Needed for #3947

Part of #3271

Rename struct fkey and struct fkey_def to fk_constraint and
fk_constraint_def respectively.
Consistently use "fk" for foreign key variables, and fk_def for
foreign key definition variables.

Remove dependency of fkey.h on space.h.
Enfore subject-verb-object naming in a few fk-related methods.

Follow up on the patch adding transaction boundaries to xrow stream.
Use tsn as an abbreviation for transaction identifier (transaction
sequence number). It is an important enough concept to use a short and
a convenient tag name for.
Deploy the name across the code - in names and comments.
Clarify comments.
Still use box_txn_id() as API method since box_tsn() and box_txn() would
be too easy to mistype.

SQLite grammar implementing foreign keys parsing is quite compact, but
on the other hand it allows to define ON DELETE and ON UPDATE actions
multiple times. For instance:

... REFERENCES t ON DELETE UPDATE ON DELETE RESTRICT;

It makes no sense and contradicts original ANSI syntax. So, lets rework
it a bit. Firstly, MATCH clause must come first, so we place it in
independent rule. Then we remove ON INSERT clause, since there is no
such opportunity at all. Finally, we have only 4 options to expose
refargs (i.e. grammar rule to parse FK actions): ON UPDATE, ON DELETE,
ON UPDATE ON DELETE, ON DELETE ON UPDATE. That's it.

Closes #3475