When libcurl is built with --enable-threaded-resolver (which is default)
and the version of the library is 7.60 or above, libcurl calls a timer
callback with exponentially increasing timeout_ms value during DNS
resolving.

This behaviour was introduced in curl-7_59_0-36-g67636222f (see [1],
[2]). During first ten milliseconds the library sets a timer to a passed
time divided by three (see Curl_resolver_getsock()). It is possible that
passed time is zero during at least several thousands of iterations.

Before this commit we didn't set a libev timer in curl_multi_timer_cb()
when a timeout_ms value is zero, but call curl_multi_process()
immediately. Libcurl however can call curl_multi_timer_cb() again and
here we're going into a recursion that stops only when timeous_ms
becomes positive. Often we generate several thousands of stack frames
within this recursion and exceed 512KiB of a fiber stack size.

The fix is easy: set a libev timer to call curl_multi_process() even
when a timeout_ms value is zero.

The reason why we did the call to curl_multi_process() immediately is
the unclear wording in the CURLMOPT_TIMERFUNCTION option documentation.
This documentation page was fixed in curl-7_64_0-88-g47e540df8 (see [3],
[4], [5]).

There is also the related change in curl-7_60_0-121-g3ef67c686 (see [6],
[7]): after this commit libcurl calls a timer callback with zero
timeout_ms during a first three milliseconds of asynchronous DNS
resolving.

Fixes #4179.

[1]: https://github.com/curl/curl/pull/2419
[2]: https://github.com/curl/curl/commit/67636222f42b7db146b963deb577a981b4fcdfa2
[3]: https://github.com/curl/curl/issues/3537
[4]: https://github.com/curl/curl/pull/3601
[5]: https://github.com/curl/curl/commit/47e540df8f32c8f7298ab1bc96b0087b5738c257
[6]: https://github.com/curl/curl/pull/2685
[7]: https://github.com/curl/curl/commit/3ef67c6861c9d6236a4339d3446a444767598a58

It is important to have testing jobs that build the project with both
-Werror and -O2 to keep the code clean. -O2 is needed, because some
compiler warnings are available only after extra analyzing passes that
are disabled with lesser optimization levels.

The first attempt to add -Werror for release testing jobs was made in
da505ee7 ('Add -Werror for CI (1.10
part)'), but it mistakely doesn't enable -O2 for RelWithDebInfoWError
build. It is possible to fix it in this way:

 | --- a/cmake/compiler.cmake
 | +++ b/cmake/compiler.cmake
 | @@ -113,10 +113,14 @@ set (CMAKE_C_FLAGS_DEBUG
 |      "${CMAKE_C_FLAGS_DEBUG} ${CC_DEBUG_OPT} -O0")
 |  set (CMAKE_C_FLAGS_RELWITHDEBINFO
 |      "${CMAKE_C_FLAGS_RELWITHDEBINFO} ${CC_DEBUG_OPT} -O2")
 | +set (CMAKE_C_FLAGS_RELWITHDEBINFOWERROR
 | +    "${CMAKE_C_FLAGS_RELWITHDEBINFOWERROR} ${CC_DEBUG_OPT} -O2")
 |  set (CMAKE_CXX_FLAGS_DEBUG
 |      "${CMAKE_CXX_FLAGS_DEBUG} ${CC_DEBUG_OPT} -O0")
 |  set (CMAKE_CXX_FLAGS_RELWITHDEBINFO
 |      "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} ${CC_DEBUG_OPT} -O2")
 | +set (CMAKE_CXX_FLAGS_RELWITHDEBINFOWERROR
 | +    "${CMAKE_CXX_FLAGS_RELWITHDEBINFOWERROR} ${CC_DEBUG_OPT} -O2")
 |
 |  unset(CC_DEBUG_OPT)

However I think that a build type (and so `tarantool --version`) should
not show whether -Werror was passed or not. So I have added
ENABLE_WERROR CMake option for that. It can be set like so:

 | cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfo -DENABLE_WERROR=ON

Enabled the option in testing Travis-CI jobs with the RelWithDebInfo
build type. Deploy jobs don't include it as before.

Fixed all -Wmaybe-uninitialized and -Wunused-result warnings. A few
notes about the fixes:

* net.box does not validate received data in general, so I don't add a
  check for autoincrement IDs too. Set the ID to INT64_MIN, because this
  value is less probably will appear here in a normal case and so is the
  best one to signal a user that something probably going wrongly.
* xrow_decode_*() functions could read uninitialized data from
  row->body[0].iov_base in xrow_on_decode_err() when printing a hex code
  for a row. It could be possible when the received msgpack was empty
  (row->bodycnt == 0), but there were expected keys (key_map != 0).
* getcwd() is marked with __attribute__((__warn_unused_result__)) in
  glibc, but the buffer filled by this call is not used anywhere and so
  just removed.
* Vinyl -Wmaybe-uninitialized warnings are false positive ones.

Added comments and quotes into .travis.yml to ease reading. Removed
"test" word from the CentOS 6 job name, because we don't run tests on
this distro (disabled in the RPM spec).

Fixes #4178.

eio library provides a portable version of sendfile syscall
which works a way more efficient than explicit copying file
by 4K chunks.

The same problem that occured with struct swim_member, has
happened with struct swim - it contains a huge structure right in
the middle, struct swim_task. It consumes 1.5Kb and obviously
splits the most accessible struct swim attributes into multiple
cache lines.

This patch moves struct swim_task to the bottom as well as other
members, related to dissemination component.

The bindings mechanism was not updated in scope of BOOLEAN static
type patch. Fixed.

Struct swim_member describes attributes of one remote member of
a SWIM cluster. It is relatively often accessed. And it has two
huge structures - struct swim_packet ping/ack_task. Each is 1500
bytes. When these tasks are in the middle of the structure, they
split and spoil cache lines.

This patch moves the whole failure detection attribute family to
the bottom of the structure, and moves these two tasks to the
end of layout.

Suspicion component is a way how SWIM protects from
false-positive failure detections. When the network is slow, or
a SWIM node does not manage to process messages in time because
of being overloaded, other nodes will not receive ACKs in time,
but it is too soon to declare the member dead.

The nodes will mark the member as suspected, and will ping it
indirectly, via other members. It 1) gives the suspected member
more time to respond on ACKs, 2) protects from the case when it
is a network problem on particular channels.

Part of #3234

Before the patch SWIM packets were being sent quite
straightforward from one instance to another with transparent
routing on Internet Level of TCP/IP. But the SWIM paper
describes last yet not implemented component - suspicion
mechanism.

So as not to overload this message with suspicion details it is
enough to say that it makes possible sending a packet through an
intermediate SWIM instance, not directly.

This commit extends the SWIM protocol with a new transport-level
section named 'routing'. It allows to send indirect SWIM messages
transparently via packet forwarding implemented fully inside
transportation component, in swim_io.c.

Part of #3234

Struct swim_task is an asynchronous task generated by the SWIM
core and scheduled to be sent when a next EV_WRITE event appears.

It has a callback 'complete' called when the task finally sent
its packet into the network. In this callback a next SWIM round
step can be scheduled, or set a deadline for a ping. Usually it
requires to know to which member the packet was sent. For this
UUID is required, but swim_task operates by inet addresses only.

At this moment UUID necessity can be bypassed via container_of
or via some queues, but it is not so once suspicion component is
introduced.

The patch adds sender's UUID to struct swim_task.

Part of #3234

SIO provides a function sio_strfaddr() to obtain string
representation of an arbitrary struct sockaddr. Call of this
function usually looks bulky because it requires explicit cast to
const struct sockaddr *, and expects the address size in the
second paremeter.

SWIM uses only AF_INET addresses and always casts them to
const struct sockaddr * + passes sizeof(struct sockaddr_in) to
each invocation of sio_strfaddr(). This patch wraps
sio_strfaddr() with a function making these preparations.

Part of #3234

Sometimes it is wanted to format multiple addresses and call
sio_strfaddr() multiple times, but it uses one static thread
local buffer, and next calls overwrite results of previous ones.

On the contrary, tt_static_buf() is not a single buffer per
thread - it is 4 buffers. Now sio_strfaddr() uses it and it is
possible to call it 4 times without rewriting old results.

Also, this update makes .bss section a bit smaller - -1 static
buffer of size 1025, and +16 bytes for tt_static buffers. Total:
-1009 bytes.

It appeared  that tt_uuid lib provides the same function:
tt_uuid_str().

The filter is going to be used to test the SWIM suspicion
component. The destination filter will break certain network
channels, and the suspicion component shall withstand that.

Part of #3234

Swim test packet filters are supposed to filter out packets
matching certain criteria or with a probability. They were
implemented as a filter-function and userdata passed into the
former on each invocation. Usually it was allocated on heap and
needed deletion. But it appeared that much simpler is to store
the filters inside struct swim_node, pass it as userdata, and get
rid of userdata destructors and dynamic allocations.

The patch is motivated by necessity to add one new filter, which
anyway will need struct swim_node as userdata.

There are two different structures - public struct swim_member
exposed by SWIM API, and struct swim_node defined and used inside
tests. Before this patch swim_cluster_node() was returning struct
swim_member, just historically. But more and more places appear
where it is wanted to safely take struct swim_node, not
swim_member, with an appropriate assertion on an invalid index.

This patch renames swim_cluster_node() to swim_cluster_member(),
and introduces new swim_cluster_node() returning swim_node.

Fixes #4174.

Tarantool SQL used to return 'number' type in request metadata
for arithmetic operations even when only 'integer's were used.

This also fixes query planner optimisation bug:
SELECT a FROM t1 WHERE a+0 IN (SELECT a FROM t1);
used to open a new ephemeral table with OpenTEphemeral when
it is not required (introduced by 2b22b913).

Closes #4103

When access is performed using VIEW, access rights should be
checked against table[s] which it is referencing, not against
VIEW itself. Added a test case to verify this behaviour.

Closes #4104

The tap:is_deeply call used to return inconsistent result
processing an empty tuple and tuple that has no values:
is_deeply({a = box.NULL}, {})  == true
is_deeply({}, {a = box.NULL}) == false
Fixed to return true by default in such case. You may also
set tap.strict = true to change this behaviour.

@TarantoolBot document
Title: tap test new flag 'strict'
In some scenarios it is convenient to distinguish box.NULL and
nil in tap:is(), tap:isnt(), tap:is_deeply() tests.
For example, {result='Success'} and {result='Success', error=None}
are different HTTP responses. You may set t.strict = true to
behave such way.
Example:
t = require('tap').test('123')
t.strict = true
t:is_deeply({a = box.NULL}, {}) -- false

Closes #4125

VDBE object is used in struct sql_txn to add new autoincrement ids in
sequence_next(). List of these ids is returned later as a query
execution result. sql_txn is created once SQL statement is executed
inside transaction and exists till commit or rollback. After its
creation it contains pointer to current VDBE. Each VDBE is freed after
statement is executed. Hence, after first SQL statement within
transaction is executed, sql_txn will point to freed memory (dangling
pointer). This leads to crash in the next processed statement. Fix to
this bug is simple: we must re-assign pointer to VDBE in sql_txn before
VDBE execution.

Closes #4157

<search condition> is a predicate used as a part of WHERE and JOIN
clauses. ANSI SQL states that <search condition> must accept only
boolean arguments. In our SQL it is implemented as bytecode instruction
OP_If which in turn carries out logic of conditional jump. Hence this
patch makes this opcode accept only boolean values.

Closes #3723

According to ANSI, LIKE predicate should return boolean result.
This patch changes type of return value of LIKE predicate.

Part of #3723

This patch make following predicates accept and return only values
of type boolean: IN, EXISTS, OR, AND, NOT, BETWEEN, IS (NULL).
In terms of approach, it is enough to patch opcodes implementing these
predicates.

Part of #3723

According to ANSI SQL result of comparison predicates must be BOOLEAN.
Before introduction of BOOLEAN type they returned 0 and 1. Now we can
change those values to false and true respectively.

Part of #3723

In most cases we don't assign and store type of node of expression AST
(except for constant literals).  To determine type of node  we use
sql_expr_type() function, which implements logic of quite simple
recursive tree traversal. Before this patch we set type of node after
code generation in sqlExprCodeTarget() without any traversal. This
approach is way worse even then sql_expr_type(). So, to improve accuracy
of type determination, let's always call that method and remove type
assignment in sqlExprCodeTarget().

Closes #4126

This patch introduces basic facilities to operate on boolean type:
boolean literals "true" and "false" where true > false; alias to null -
unknown; column type "BOOLEAN" and shortcut "BOOL"; opportunity to
insert and select boolean values from table; OR and AND predicates
accept boolean arguments; CAST operation involving boolean type;
comparison between boolean values (including VDBE sorter routines).

Part of #3648

It is not used (since we dispatch call of sql_bind_*type*() in
sql_bind_column()), so can be removed.

This patch provides straightforward refactoring replacing enum sql_type
with enum mp_type. Note that we use msgpack format instead of field_type
since it is more suitable when dealing with NULLs.

It is obvious that we can't add string with number except the case when
string is a number literal in quotes (aka '123' or '0.5'). Before this
patch values which can't be converted to numbers were just skipped.
Now error is raised. This patch also removes sql_value_numeric_type()
function since it is not used anymore: instead we invoke
mem_apply_numeric_type().

Fix codestyle according to Tarantool guideline; remove unused argument
from signature; make function non-static - we are going to use it in
aggregate function sum() (src/box/sql/func.c) to attempt at converting
string values to numbers.

Redirects can work improperly with libcurl-7.30 and older. CentOS 7
provides libcurl-7.29, so app-tap/http_client.test.lua fails on the
corresponding test case in CI. We need to disable the test case until
#4180 will be resolved.

While we are here fixed httpd.py to write a response at once, not
char-by-char (the typo from daf1ced8). See
gevent/pywsgy.py::process_result().

Related to #4119 and #4180.

While we are here also added forgotten option descriptions to httpc.lua.

@TarantoolBot document
Title: httpc: new 'follow_location' option

When the option is set to `true` (which is default) and a response has
3xx code the http client will automatically issue another request to a
location that a server sends in 'Location' header. If the new response
is 3xx again, the http.client will issue a next request and so on in a
loop until a non-3xx response will be received. This last response will
be returned as a result.

Setting this option to `false` allows to disable this behaviour. In this
case the http client will return a 3xx response itself.

See https://curl.haxx.se/libcurl/c/CURLOPT_FOLLOWLOCATION.html

Previously Tarantool used to raise the confusing error message in
case of invalid usage of the httpc module. Fixed to follow the
current module API.

Closes #4136

The tuple:update() used to work incorrectly in case of empty
tuple produced with box.tuple.new{} because update_create_rope
unconditionally initialized a new rope with [tuple_data,
mp_next(tuple_data)] field that might not exists.

Closes #4041

There was a variable used as a counter from 0 to
uint32_t space_def->field_count. Obviously it can't be < 0, but
there was a check for that. Drop it.

The check was revealed during SQL BOOLEAN review, where one of
patches gets rid of some explicit Expr.type assignments.

According to the ANSI standard, ltrim, rtrim and trim should
be merged into one unified TRIM() function. The specialization of
trimming (left, right or both and trimming characters) determined
in arguments of this function.

Closes #3879

@TarantoolBot document
Title: TRIM() function

Modify signature of SQL function TRIM(). This function removes
characters included in <trim character> (binary) string from
<trim source> (binary) string until encounter a character that doesn't
belong to <trim character>. Removal occurs on the side, specified by
<trim specification>. Now, syntax is following:
TRIM([ [ <trim specification> ] [ <trim character> ] FROM ] <trim source>).

<trim specification> can be one of the following keywords: LEADING,
TRAILING and BOTH.
<trim character> is the set of trimming characters.
<trim source> is the string, that will be trimmed.
If FROM is specified, then:
1) Either <trim specification> or <trim character> or both shall be
specified.
2) If <trim specification> is not specified, then BOTH is implicit.
3) If <trim character> is not specified, then ' ' is implicit.

This patch removes part of unused functions and macros from
main.c. This will facilitate the work of removing SQL error/status
codes.

Needed for #4074

The patch sets format for spaces with sysview engine.
This is done due to following reasons:
  1. Since an SQL view looks into underneath space's format, set
     it for spaces with sysview engine. Before the patch, spaces
     with sysview enginge didn't have its own tuples and hence
     didn't need to have a format.
  2. To use sysview engine to deal with SQL views. This will allow
     to use sysview machinery to query SQL views from Lua land.

Closes #4111

Update_addr and update_payload need to increment member's
incarnation when it is self. For that they used a special
parameter incarnation_inc set in 1 for self and in 0 for others.

It was used to encapsulate incarnation update + event scheduling
on member attribute updates, but on the other hand it broke
another encapsulation level - there should not be exceptions for
'self' in these functions.

This patch makes incarnation increment explicit in the places
where 'self' is updated.

Payload is arbitrary user data disseminated over the cluster
along with other member attributes.

Part of #3234