Add xlog_create function. xdir-create_xlog will work via xlog_create
Add reuse_fd flag to reuse fd after xlog close.

Follow up 037c4e97

Was:
More than one tuple found by get()

Now:
Get() doesn't support partial keys and non-unique indexes

Fix #1882

Split iterations to xlog transaction iteration and rows from current
transaction iteration. Error handling for failed tx reading or row
decoding can be differ.

No semantic changes.

Functions for tuples creation and validation must not
throw exceptions for being C-compatible - it is need
for #1908 in which this functions will be used in vinyl.c

It's better than open-coded lists.

On compaction we link range->mem of the destination range(s) to the
source range, so that the same ->mem can be accessed from two ranges.
This looks confusing, and this is what prevents us from using lists for
linking mems. Let's instead add dest ranges to range->compact_list of
the source range.

A run may be freed and hence its file descriptor closed by compaction
even if it is currently being read from by a coeio thread. At worst, the
descriptor may be reused for another file or even socket, resulting in
unexpected glitches. To avoid that, take a reference to the run before
issuing coeio_preadn() and only free it and close the file when the
reference drops down to 0.

Since commit 2a2bd84e ("vinyl: split data and index files"), run->size
equals the size of the run header, not the total run data size. So we
should use vy_run_size+vy_run_total instead of it for calculating dump
bandwidth, othe

* remove a separate 'env' module
* add os.environ() and os.putenv() function to Lua built-in 'os' module
* update tests

Issue #1720

1. Change xlog_cursor_load() semantic - now this function ensures that
at least count bytes are available in the read buffer.

2. The write position of the read buffer is now bounded to the 'physical'
read position of xlog file. Now we can just move forward rbuf->rpos and
remove buf_offset from xlog_cursor to simplify the read logic.

3. xlog_cursor_read_tx() now returns the minimal count of data needed to
continue parsing. In this case caller must load this count of bytes into
the read buffer and restart parsing.

4. xlog_cursor_read_tx() now doesn't change the buffer position in case of
error. This change allowed us to so remove search_offset from xlog_cursor.

5. Don't increment parsing position for 1 byte in case of the read errors
anymore, just return a SystemError.

Issue #1720

Fix a bug in mem/run iterator restorations
Fix gh-1914

* Remove exceptions from some key_def functions;
* Add const qualifiers;

Need for #1908

mp_check() checks only the first element of MsgPack and
should be called three times for three elements.

+ Fix possible SIGBUS on ARM while decoding magic (see #1930)
+ Add diag for errors

Fixes #1880

Check magic in xlog_cursor_read_tx.

Issue #1720

* Fix buggy math calculation and simplifies randomization logic
  introduced by gh-732. The primary difference is that the new
  implementation uses fixed periods (=box.cfg.snapshot_period)
  shifted by a random offset instead of floating periods
  (=box.cfg.snapshot_period + random). In addition, snapshot daemon
  now never start snapshoting earlier than box.cfg.snapshot_period time.

* Use fiber.channel() to control daemon

Fixes #1912
See also #732

Fixes #1929

Regions provide faster free of a set of allocated in one region
objects.

Closes #1907

* Split try {} catch() {} block in recover_xlog()
* Remove useless try .. catch wrapper around xlog_cursor_next_xc()
* Add comments

Issues #1720

We can't skip significant data in case of magic corruption,
so incorrect magic should be treated as a parse error.
We add separate function to skip datas until magic found.

Issue #1720

No semantic changes.

* Replace rawmemchr() which is not supported on macOS
* Move hardcoded VCLOCK_LEN_MAX constant to vclock.h
* Add missing `const` modifiers
* Add comments

Issue #1720

Fix side effects of previous tx patch
Fix gh-1864
Fix problem when run is completely squashed during dump

Remove a legacy hard-coded limit on range size.
Affects cloud.mail.ru deployment, which got to 713 000 open
file descriptors on a large database (not good).

No semantic changes.

Change XlogError to SystemError() in xlog_cursor_open().

Follow up the previos commit.

Issue #1720

+ Remove try .. catch block from xlog_cursor_next_row().

Follow up the previous commit