On 64-bit iconv called via ffi returns a uint64_t cdata object, which
equals to -1 in case of error so error handling works fine. On 32-bit it
returns a number equal to (size_t)-1, which is equal to 4294967295 and
not equal to -1 when compared to in Lua. As a result Lua binding to
iconv hangs if the source string is incorrect. Fix this by comparing the
result of iconv to ffi.cast('size_t', -1) instead of -1.

Closes #2466

Follow up #2334

An on_replace trigger on replica can issue local changes
using its own replica_id. Check replica_id for every row
during writing to WAL.

Fixes #2682

Fix minor issues found by coverity scan:
 - Do not try to estimate slice size if the run is empty to avoid div/0.
 - Fix assert() that is supposed to check slice->first_page_no, but
   checks last_page_no instead.
 - If the slice doesn't intersect with the run (i.e. slice->end is less
   than or equal to run->begin), leave its stats zeroed out.

Closes #2698

The bug happens when some TX deletes a statement (writes it to
its own write set) and then reads a range of tuples containing
that deleted statement. DELETE must not come to select result
(in this and only transaction before it is committed) and
just are skipped by read iterator. After such a skipping and
getting next true (non-DELETE) statement the iterator considers
that there is no data between the previous statement and the
current, and add them to the cache as a chain. This is wrong
because the data could be hidden by DELETE statement.

Track such jumps over TXW DELETE statements in read iterator and
do not add statements as a chain in this case.

We use a binary tree for tracking reads (read_set). The tree is ordered
by (index, statement, tx). Insertion of a partial statement into the
tree may result in breaking comparison transitivity, which is mandated
by the tree implementation. However, for range requests we insert the
search key into the read set, which may be partial. This may result in a
crash in the read_set.

Tracking the search key doesn't make much sense anyway: it doesn't bring
us any closer to range request tracking, only obscures the semantics of
select. So let's track only statements actually returned by the iterator
plus gaps found by point requests.

Closes #2675

Fix bad merge of 945a97db and 7a5835f4

A dump task can be scheduled or completed after we saved stats.
Make sure this doesn't happen.

Closes #2696

Do not use thumb specifier for asm code, use unified syntax directive
instead.

Revoke permissions granted by the test.

Needed for #946

Relay reader fiber is jonable now and should be joined by relay main
fiber.
See #2284

Don't stop coio twice.

Update comments.

Replication slave send its current vclock to replication master
after recieving new data from master.
See #2284

There is a check in vy_point_iterator_apply_history that compares
statement's lsn with lsn with vlsn and skips history node if the
statement is invisible.

This check is wrong:
 - if TX was originally in a read view it could not read any
invisible statement.
 - if TX was not in a read view but were sent to read during yield,
the new created read view must see all current database state,
including data that was read by the iterator.
 - if TX is in read view and made some changes (in own TX write
set) it will skip TXW statement since it has INT64_MAX lsn.

Replace check with assert. Concern possibility of TXW statement.

Write appropriate test.

test/small symlink was created only on `make test`,
create this symlink `make all`.

Check current user write access on space truncate.

Fixes #2674

_truncate space write access is required for space truncation.
Add this privilege to bootstrap.snap.

See #2674

Closes #2609
Closes #2633
Closes #2616

Follow up #910

vinyl_cache_add takes pair of statements - current and previous.
By design, there must be no statements in cache between them.

Some time ago a check was made that the previous statement is in
cache and lays right next to current statement; if that was not
true, the function exited without building a cache chain.
The was wrong because there are some cases when the statement
could be removed by TX finalize/rollback after prepare or even by
cache garbage collector.

Later in this check 'return' was replaced by 'assert' and that
caused rare test crash.

Fix this check. Check only two things 1) absence of entries between
current and previous statements 2) presence of previous statement
if current is already in chain.

Add a test case.

See code comments for details.

Fixes #2685

Discard xlogs with large tuples after decreasing memtx_max_tuple_size.

Follow up #2667

Add a special dump-waiting error injection to vy_point_iterator_get
that will wait for a dump completion. Use it in test and make code
coverage complete.

In the point iterator test

Fix a couple of mistakes in point_iterator and enable it back.

Disable 'could be serializable' suggestion in tx_conflic test
since new point iterator adds its key to read set before reading
from disk and that leads to unnecessary conflicts sometimes.

Fix #2662

No semantic changes.

Follow up #2047

* Remove dependency on memtx from tuple_format_default
* Add tuple_new() to create standalone tuples
* Allow box.tuple.new() without box.cfg()

Closes #2047

Follow up the previos commit.

See #2667

- Use slab_get() as a fallback for smalloc() for large size
- Set MemTX slab_size to 16Mb and don't use memtx_max_tuple_size
  for calculation (Vinyl is not changed yet)
- Allow dynamic configuration of memtx_max_tuple_size

Closes #2667

Added options:

* verify_host - CURLOPT_SSL_VERIFYHOST;
* verify_peer - CURLOPT_SSL_VERIFYPEER;
* ssl_key - CURLOPT_SSLKEY;
* ssl_cert - CURLOPT_SSLCERT.

This patch adds ability to reload C procedures on the fly without
downtime. To achive that, Tarantool loads a new copy of shared
library and starts routing all new request to the new version.
The previous version remains active until all started calls
are finished. All shared libraries are loaded with RTLD_LOCAL,
therefore two or more copies can co-exist without any problems.
From now box loads all external modules via an unique symlink to
avoid caching inside dlopen().

If one of some module function is reloaded then all other functions
from this module will be reloaded.

Reviewed and heavily patched by Roman Tsisyk.

Closes #910

rbtree_next() has log(N) complexity to advance to the right subtree.
The new iterators saves the iteration path during descent.

Closes #1668

panic() from non-main thread calls exit(), which fires
at_exit triggers, which try to destroy TX data.

Closes #2686