xlog for each recovey after new xlog file and pread implementation.
Optimize fake eof (not closed xlog) cause for recovery.

xlog for each recovey after new xlog file and pread implementation.
Optimize fake eof (not closed xlog) cause for recovery.

None of existing task kinds return anything but 0 on completion.
So let's remove the unused ret code to simplify error handling.

Use SystemError and ClientError where appropriate. Note, ER_VINYL is
still there - it is used for reporting about corrupted/incomplete range
files on recovery. However, as soon as we switch to the xlog file
format, we'll probably be able to get rid of it altogether.

We need it to propagate system errors from vinyl.

Closes #1713

Test may fail on OS X due to NOFILE limit.

vy_read_iterator->curr_stmt is supposed to store the last statement
returned by vy_read_iterator_next() to be used for iterator restore.
However, currently it is also updated in vy_read_iterator_next_range().
If the iterator is restored after such an update it will skip the first
statement in the range, which might result in invalid select result. For
instance, if the first statement (the newest one) in a range is a delete
followed by an upsert for the same key (which is older), we will skip
the delete on restore, and return the deleted key, see

  vy_read_iterator_restore:
    vy_merge_iterator_restore:
      vy_run_iterator_restore:

        if (vy_stmt_compare(fnd, last_stmt->data, itr->index->key_def) == 0) {
                position_changed = false;
                if (next_stmt->lsn >= last_stmt->lsn) {
                        /* skip the same stmt to next stmt or older version */

This results in occasional vinyl/options test failures.

Closes #1853

Conflicts:
	test/unit/bps_tree.cc
	test/unit/bps_tree_iterator.cc

Allow to specify allocator context in matras and matras-based data
structures. After this change, the following salad data structures can
be passed a custom allocator context: light, rtree, bps_tree.

Currently, we use a rather dumb way to determine if a statement from
xlog should be committed or not on recovery - we lookup the statement in
all on-disk runs and, if found, compare its lsn. Obviously, this slows
down the recovery procedure significantly. Actually, we can avoid
reading on-disk runs altogether. All we need to do is compare the xlog
statement's lsn with the range's max lsn - we should only commit the
statement if the former is greater.

Closes #1697

Whenever we need VINYL_EQ order in range iterator, we use VINYL_GE.
This behavior exists solely due to historical reasons and needs to
be fixed.

Currently, we never remove delete statements on dump, even if creating
the first run. However, there's no point in preserving deletes in case
there's no on-disk data.

Closes #1823

Fixes #1442

Surround key and lsn in parentheses.

If there are a lot of successive upserts for the same key, select might
take too long to squash them all. So once the number of upserts exceeds
a certain threshold, we schedule a fiber to merge them and substitute
the latest upsert with the resulting replace statement.

Closes #1829

 - Rename vy_index->range_index_version => ->version
          vy_range->range_version => ->version
          vy_merge_iterator->range_index_version => ->index_version
 - Add a comment to vy_range->version
 - Do not increment vy_index->version on dump/compact in case only
   mem/run is added/removed and range index tree is left intact -
   increment vy_range->version instead.

ev_now might be significantly faster on some platforms.

Note, latency info is still reported in nanoseconds.

Closes #1566

 - rename vy_avg to vy_latency
 - remove ->min as it's going to be ~0 anyway - max and avg is enough
   for giving an insight of what actual latency is like
 - remove ->avg as we can always compute it as ->total / ->count
 - zap temporary buffer used for stat reporting
 - report latency stat in a table instead of string

It's good to know not only the total number of events that have happened
since the server start, but also the average over the last few seconds.
So let's use rmean for more stat counters:

 - VY_STAT_GET: number of reads, including from cursor
   (former vy_stat->get)
 - VY_STAT_TX: number of transactions (former vy_stat->tx)
 - VY_STAT_TX_OPS: number of operations (read or write) issued by
   transactions (formerly, we had vy_stat->tx_stmts, which is the
   average number of operations per transaction)
 - VY_STAT_CURSOR: number of cursors used (former vy_stat->cursor)
 - VY_STAT_CURSOR_OPS: number of reads issued by cursors
   (former vy_stat->cursor_ops)

Also, drop get_read_disk and get_read_cache for now, as they are always
0. We might want to reintroduce them once we have read cache.

Instead of creating a temporary storage for vinyl info, we can push it
to lua immediately as we go. This simplifies the code and makes it
possible to use strings allocated on stack as keys and values, not just
literals.

Merge vy_range_iterator_start() to vy_range_iterator_next()

This new function is used to search a single range
that contains specified key.

Closes #1747