Stay in orphan (read-only) mode until local vclock is lower than
master's to make sure that datasets are the same across replicaset.
Update replication/catch test to reflect the change.

Suggested by @kostja

Needed for #3210

Closes #3148

EV_USE_REALTIME and EV_USE_MONOTONIC, which force libev to use
clock_gettime, are enabled automatically on Linux, but not on OS X. We
used to forcefully enable them for performance reasons, but this broke
compilation on certain OS X versions and so was disabled by commit
d36ba279 ("Fix gh-1777: clock_gettime detected but unavailable in
macos"). Today we need these features enabled not just because of
performance, but also to avoid crashes when time changes on the host -
see issue #2527 and commit a6c87bf9 ("Use ev_monotonic_now/time
instead of ev_now/time for timeouts"). Fortunately, we have this cmake
defined macro HAVE_CLOCKGETTIME_DECL, which is set if clock_gettime is
available. Let's enable EV_USE_REALTIME and EV_USE_MONOTONIC if this
macro is defined.

Closes #3299

The bug was that logging we passed to function write
number of bytes which may be more than size of buffer.
This may happen because formatting log string we use vsnprintf which
returns number of bytes would be written to buffer, not the actual
number.

Fix this with limiting number of bytes passing to write function.

Close #3248

* Refactor tests.
* Add ev_async and fiber_cond for thread-safe log_rotate usage.

Follow up #3015

Fix race condition in test on log_rotate.
Test opened file that must be created by log_rotate and read from it.
But as log_rotate is executed in separate thread, file may be not
created or log may be not written yet by the time of opening in test.

Fix this with waiting creation and reading the line.

Netbox does not need nullability or collation info, but some
customers do. Lets fill index parts with these fields.

Fixes #3256

* session_run_on_disconnect_triggers is called only if there are
corresponding triggers so move session_storage_cleanup to
session_destroy.
* fix session storage cleanup path: use
"box.session.aggregate_storage[sid]" instead of
"session.aggregate_storage[sid]" (what was wrong)

Fixed #3279

Empty strings should be ignored, rather than throw an error.

Passing only empty strings (or nothing) to `pathjoin` will return '.' which means the current directory

Every path part passed to `pathjoin` is now converted to a string

The `gsub('/+', '/')` call already does what the removed code does, so avoid the unnecessary work
Simply check if the result equals '/' before removing a trailing '/'.

The previous check did extra work for no gain.

It is possible to discard non-sent responses using a special
sequence of requests and yields. In details: if DML requests
yield on commit too long, there are fast read requests, and a
network is saturated, then some non-sent DML responses are
discarded.

Closes #3255

A vinyl space can be altered in such a way, that key definitions
of indexes are not changed, but comparators do. It is because
space format reset can make some indexed fields optional.

To be able update key definitions in place, they must not be used
in a worker thread. So lets copy key_defs for a worker, and
update index key definitions in place.

An alternative is key_defs reference counting, but there is open
questions what to do in key_defs in mems, ranges, iterators,
runs, slices. Now lets do a hotfix of a crash, and then
refactoring.

Closes #3229

Use key_def_delete instead. It is more safe to use one
destructor everywhere for a case, if in a future key_def will not
be deleted by simple free().

Under FreeBSD getline prototype is not provided by
default due to compatibility problems.
Get rid of getline (use fgets instead).
Based on @locker proposal.

Closes #3217.

log_rotate writes informational message only in plain format,
which is inappropriate when logger is configured to log in json format.

Fix it with replacing write with say_info which is safe
because it is ev signal callback not signal handler.

Fix bug say_format in json formatting,
Log message was invalid in several cases.

Closes #2987

Current log rotation is not async signal safe.
In order to make it so refactor signal handling
with ev_signal.

Log rotation for each logger performs in separate
coio_task to provide async and thread-safe execution.

Relates #3015

Remove yielding and waiting task complete in coio_task_post
in case if timeout is zero.
This patch is inspired by the need in log_rotate posting coio
task. This post should not yield there because the implementation
of multiple loggers works with linked list structure of loggers
which is not fiber-safe.

When a space format is updated, a new min field count must be
calculated before a new format construction to check that some
of fields became optional.

Part of #3229

When scanning a secondary index, we actually track each tuple in the
transaction manager twice - as a part of the interval read from the
secondary index and as a point in the primary index when retrieving
the full tuple. This bloats the read set - instead of storing just one
interval for a range request, we also store each tuple returned by it,
which may count to thousands. There's no point in this extra tracking,
because whenever we change a tuple in the primary index, we also update
it in all secondary indexes. So let's remove it to save us some memory
and cpu cycles.

This is an alternative fix for #2534
It should also mitigate #3197

We never use vy_point_lookup directly, instead we open vy_read_iterator,
which automatically falls back on vy_point_lookup if looking for exact
match (EQ + full key). Due to this we can't add a new point lookup
specific argument (we would have to propagate it through the read
iterator, which is ugly). Let's call vy_point_lookup directly when we
know that vy_read_iterator will fall back on it anyway.

This reverts commit a31c2c10.

The commit reverted by this patch forces all autocommit SELECTs to open
a read view immediately, as a result they can't update tuple cache.
Turned out that one of our customers intensively uses such SELECTs, and
disabling cache for them results in performance degradation.

The reason why that commit was introduced in the first place was to
avoid read set bloating for big SELECTs (e.g. space.count()): currently
we track not only read interval boundaries, but also each tuple fetched
from the primary index if it is a secondary index that is being scanned.
However, it doesn't seem that we really need to do that - tracking an
interval read from a secondary index guarantees that if a tuple returned
by the iterator is modified the transaction will be aborted and so
there's no need to track individual tuples read from the primary index.

That said, let's revert this commit and instead remove point lookup
tracking in case it is a secondary index that is being scanned (done
later in the series).

The first reason of the patch is that vinyl iterators are virtual
already, and 'if's about constant index attributes (like index->id)
can be replaced by new next() implementation. Now in next() index->id
is checked to detect necessity of primary index lookup.

Lets split next() in 2 functions: primary_next() and secondary_next()
to remove 'if'.

The second reason, that in #2129 logic of secondary index lookup
complicates a lot. For example, there is raw idea to do not add
statements into a cache before looking up in a primary index, because
after #2129 any tuple, read from a secondary index, can be dirty.

Needed for #2129

In most cases a tarantool yields on test_run connection and current
transactions rollback, but some times a tarantool console already has more
input to execute and select returns different results.

Fixed #3145

We don't write empty run files anymore. Remove the dead code.

Run iterator uses curr_pos (i.e. page number plus offset) as pointer to
the current position. Whenever it needs to get a statement at curr_pos,
it calls vy_run_iterator_read(), which allocates a new statement. It
doesn't try to cache the last allocated statement, which results in
multiple pointless reallocations of the same statement. For instance,
vy_run_iterator_next_key() rereads the current statement, then moves to
the next key, then calls vy_run_iterator_find_lsn(), which rereads the
current statement again. This is just stupid.

To avoid that, let's keep vy_run_iterator->curr_stmt in sync with
curr_pos. This simplifies the code quite a bit and makes it more
efficient.

vy_run_iterator_get() remembers the position of the last statement
returned by the iterator in curr_stmt_pos. It then uses it to skip
a disk read in case it is called again for the same iterator position.
The code is left from the time when iterators had public virtual
method 'get', which could be called several times without advancing
the iterator. Nowadays, vy_run_iterator_get() is never called twice
for the same iterator position (check coverity scan) so we can zap
this logic.

vy_run_iterator_load_page() keeps two most recently read pages. This
makes sense, because we often probe a page for a better match. Keeping
two pages rather than just one makes sure we won't throw out the current
page if probing fails to find a better match. What doesn't make sense
though is cache promotion logic: we keep promoting the page containing
the current key. The comment says:

        /*
         * The cache is at least two pages. Ensure that
         * subsequent read keeps the cur_key in the cache
         * by moving its page to the start of LRU list.
         */
        vy_run_iterator_cache_touch(itr, cur_key_page_no);

The comment is quite misleading. The "cache" contains at most two pages.
Proudly calling this travesty of a cache LRU is downright ridiculous.

Anyway, touching the current page will simply swap the two cached pages
if a key history spans less than two pages, resulting in no performance
gain or loss whatsoever. However, if a key history spans more than two
pages, it will evict a page that is about to be read.

That said, let's get rid of this piece of crap.

vy_run_iterator_next_key() has to special-case LE/LT for the first
and the last page. This is needed, because this function is used by
vy_read_iterator_seek() for starting iteration. Actually, there's no
point for vy_read_iterator_seek() to use vy_read_iterator_next_key() -
vy_read_iterator_next_pos() + vy_read_iterator_find_lsn() would be
enough. Bearing this in mind, simplify vy_run_iterator_next_key().

Inside a docker 'Connection refused' error turns into
'Cannot assign requested address' - because of it netbox test
fails, which searches for 'Connection refused' in logs.

Let search for both.

Follow up #3164

When reconnect fails, it prints error to log. But reconnect can
fail again and again very many times until the connection is
closed.

Lets print the first error message under warn log level, and the
following ones under verbose log level until error will not be
changed.

Closes #3175

If a connection has reconnect_after > 0, then it is never deleted
until it is explicitly closed or reconnect_after is reset. It is
because worker fiber of such connection holds all references
during yield.

Fix it by do not waiting for a next reconnection inside a state
machine - protocol_sm() function must not be infinite in a case
of error.

Closes #3164