* added argparse module. Internal, for now (internal.argparse)
* added cat/play commands for tarantoolctl (with tests)

* add assert instead of useless code
* add huge comment
* fix gh-1899 finally

Adding gh-1904 tests. No fixes since 1.7 has new net.box.

 * :close() didn't stop reconnect activity;
 * cancelling a fiber in the ceratin request stage leads to
   deadlock in subsequent :close().

The test starts 5000 fibers each of which sends 20 requests in two
batches via iproto. After that, it injects a 5 seconds delay to WAL and
then waits until all the fibers have completed.

The test takes about 20 seconds, so it's marked as long.

Intended to catch bugs similar to #1439, #1892.

If a message arrives when there is no output for the connection and the
IPROTO_MSG_MAX limit is hit, the input will never be processed unless
the other end writes more data to the socket.

To make sure a throttled connection will be woken up eventually when the
number of pending requests drops below IPROTO_MSG_MAX, let's link all
throttled connections in a list and wake the oldest element from it
whenever a messages is freed.

The issue was initially attempted to be fixed by commit 52b47951
("Don't stop connection if there is no pending requests") by allowing to
exceed the IPROTO_MSG_MAX limit. That commit, however, introduced the
possibility of a deadlock due to depleting the fiber pool, which was
"resolved" by commit 9492cddb ("iproto: fix IPROTO_MSG_MAX check") while
reintroducing the original issue.

Patch by Vladimir Davydov

We can gather all the stats we need at run time.

 * histogram_discard() to remove previously added observations
 * histogram_snprint() to dump a histogram to a string buffer

Required for maintaining a histogram of number of runs per range in
vinyl.

Fix compilation warnings with disable assertions

Creating only one file per range, as we do now, is inflexible, because
 - we can't easily coalesce two ranges from the tx thread
 - it's unclear how to implement multi-level design without introducing
   holes in range files

So this patch switches vinyl to file-per-run design, where each run has
its own file and a file descriptor to access it. Files are named as
follows: <lsn>.<range_id>.<run_no>.{run,data}. The new name component,
<run_no> is the serial number of the run in the range, starting from 0.
As before, we start recovery from ranges with higher ids. We recover
runs of the same range in the same order they were dumped. Range
boundaries are stored in each run file.

errno returned via errorno member, not in result

Fix #1899

fio_filename replace errno by readlink call and can
be used for error logging

ANY - internal field type used for filling not indexed fields
in struct tuple_format.fields. So the type ANY must be
forbidden for external access and must not be used anywhere except
struct tuple_format.fields.

Closes #1897

Vinyl had its own implementation of comparators functions, because
vinyl statements were different by layout from box tuples.

* Change the `struct vy_stmt` layout to be similar to `struct tuple` layout.

* Create vy_index->format based on vy_index->key_def and remove key_map.

* Use comparators from tuple_compare.h.

See #1572

* Rename INDEX_OFFSET to TUPLE_INDEX_BASE
* Use `fieldno` variables for field numbers
* Update comments and clean up code

In the tuple.h comparators only for 'struct tuple' were declared. This commit
provides comparators that can work with raw data. Thus new comparators can be
used in vinyl.c.

See #1572, stage 3.

We shouldn't allow more than IPROTO_MSG_MAX requests, because this can
result in depleting the tx fiber pool (FIBER_POOL_SIZE) and, as a
result, a deadlock (e.g. WAL needs a fiber to wake up another fiber
waiting for write to complete).

The IPROTO_MSG_MAX check was effectively disabled by commit 52b47951
("Don't stop connection if there is no pending requests"):

> --- a/src/box/iproto.cc
> +++ b/src/box/iproto.cc
> @@ -590,7 +590,10 @@ iproto_connection_on_input(ev_loop *loop, struct ev_io *watcher,
>       try {
>               /* Ensure we have sufficient space for the next round.  */
>               struct iobuf *iobuf;
> -             if (mempool_count(&iproto_msg_pool) > IPROTO_MSG_MAX ||
> +             if ((mempool_count(&iproto_msg_pool) > IPROTO_MSG_MAX &&
> +                     /* Don't stop connection if there is no pending requests */
> +                     (ibuf_used(&con->iobuf[0]->in) > con->parse_size ||
> +                      ibuf_used(&con->iobuf[1]->in))) ||
>                   (iobuf = iproto_connection_input_iobuf(con)) == NULL) {
>
>                       ev_io_stop(loop, &con->input);
>
>         (ibuf_used(&con->iobuf[0]->in) > con->parse_size ||
>          ibuf_used(&con->iobuf[1]->in))

because

  (ibuf_used(&con->iobuf[0]->in) > con->parse_size ||
   ibuf_used(&con->iobuf[1]->in))

is almost always false, since iproto_enqueue_batch() parses and flushes
all available input so that both iobufs are empty and the parse_size is
typically 0.

It looks like the problem the above-mentioned commit tried to address is
different. The thing is that mempool_count(&iproto_msg_pool) accounts
not only active requests, but also disconnect messages allocated for
each connection in advance. As a result, if there were more than
IPROTO_MSG_MAX (768) connections, iproto would get stuck forever.

That said, the true fix would be subtracting the number of connections
from the number of objects allocated from iproto_msg_pool for the check.

Closes #1892

When restoring read iterator, we reopen range iterator with key equal to
the last seen statement. This is incorrect for eq order and partial
initial key. On restore we should position range iterator to the
original range without changing it->key or it->order.

xlog_create() depend on xdir and we need a good name for similar function
without xdir dependency.

Needed #1720

Follow up the previous patch.

Needed for #1720

* Inline vy_range_cmp and vy_range_cmpnode to
  remove double inversion of comparison logic
* Streamline vy_stmt_compare_raw()
* Rename API functions
* Fix comments

See #1572

vy_stmt_compare is split into the set
of functions for comparison
* any type statements
* any type statement with an action statement (REPLACE, DELETE, UPSERT)
* any type statement with a key statement (DELETE, SELECT)
* an action statement with another action
* a key statement with another key

See #1572 stage 3 point 1

Closes #1852

Closes #1755

To be used as an indicator of exception. From now on one should write

  if (fiber_join(...)) // or cord_join(...)
    diag_raise();

instead of

  fiber_join(...);
  diag_raise();

Now we can make diag_raise() assert that diag is not empty.

Closes #1847

vy_checkpoint() expects diag to be set on any dump/compact error,
otherwise it crashes.

Closes #1884

See #1720

 - make them add all runs and mems in a loop
 - remove is_mutable and control_eof arguments