* Refactor templates in tuple_compare.cc to generate optimized
  versions for various field types
* Add a function pointer in key_def to select optimized
  version at runtime

Fixes #2084

* Implement tuple_extract_key_sequential() for sequential keys
* Add a function pointer to in key_def to select optimized
  version at runtime

Closes #2048

It uses error injection which is compiled out for release builds.
Also, improve the test output by adding s:count() after each space
modification for checking if the statement was actually inserted.

* update comments, messages
* add a test for changing index id

Allow to alter any index opts which don't change key_def of
the index.

Closes #1931
Closes #2109
Closes #2149

Needed for extractiong write iterator to a separate file.
Also move stat account from the function to caller code.

Change-Id: I73af73e6a34f9d7431a3c8056030c9858226b0ad

time_t is long on all major platforms, not int.

Fixes #2443

Change-Id: I7ca029ccfe87124bdc7cdd2ee1db06404f38fe39

Simplify check_param_table() arguments in index.create() and index.alter().

Closes #2148

Change-Id: I10d8d93a12e8bbd91c36f6eefc0f0e8e07636de5

Currently, only keys from vy_log_key_mask[type] may be present in a
record of the given type. Extra keys will result in an error. In
order to rework space truncate, we will have to add a new key for
VY_LOG_CREATE_INDEX record (space_version). To be able to start from
vylog generated by an older version, we need to allow extra keys. So
this patch deletes vy_log_key_mask[] and makes vy_log_record_encode()
encode only those keys whose value is different from the default one
(similarly to request_encode()). Checking for mandatory keys is now up
to vy_recovery_process_record() (currently there's the only that needs
to be checked - VY_LOG_KEY_DEF)

vy_recovery_iterate() doesn't clean vy_log_record before proceeding to
the next log entry. As a result, a run record passed to the callback by
vy_recovery_iterate() contains extra info left from the index this run
is for: index_lsn, index_id and space_id. We use this in gc and backup
callbacks to format file name. The problem is vy_recovery_iterate() is
also used internally for log rotation. Including extra keys in records
doesn't result in writing them to the log file on rotation, because per
each record type we have a mask of keys corresponding to the record
(vy_log_key_mask). In order to allow optional keys in vylog, the
following patch will change the meaning of the mask so that it only
contains mandatory keys, while a key will be written to the log only if
its value differs from the default (similarly to request_encode). Thus,
to avoid writing keys not relevant to a record type we need to clean
vy_log_record within vy_recovery_iterate() before jumping to the next
record. So this patch weans gc and backup from exploiting this feature -
let them save index_id and space_id in the context, as we do on
replication and recovery.

Info_append_double is needed to print vinyl index.info.bloom_fpr later.

It was helpful when a vinyl index could have a custom path. Currently,
it's forbidden, so we can format index path in place.

We can format it in place when needed.

There must be a reasonable timeout on quota wait time, otherwise we risk
throttling a client forever, e.g. in case of a disk error. This patch
introduces quota timeout. The timeout is configured via vinyl_timeout
configuration option and set to 60 seconds by default.

Closes #2014

Use separate function for each kind of callback. The next patch will add
an argument and a return value to the throttle callback, so it isn't apt
anymore to use the only callback to handle all cases.

Currently, we reserve quota after allocating memory, which can result in
exceeding the memory limit, e.g. after the following script

    box.cfg{vinyl_memory = 1024 * 1024}
    s = box.schema.space.create('test', {engine = 'vinyl'})
    s:create_index('pk')
    pad = string.rep('x', 2 * box.cfg.vinyl_memory / 3)
    s:auto_increment{pad}
    s:auto_increment{pad}

is done, box.info.vinyl().memory.used reports that 1447330 bytes are
allocated. Fix this by reserving quota before allocation. A test is
added later in the series.

Vinyl dump is disabled during local recovery from WAL. This is OK,
because we shouldn't exceed the quota provided we don't replay
statements that were dumped to disk before restart. Check that.

Although upsert optimizers (both sync and background) replace the
last UPSERT with a REPLACE and do not insert new statements, they
use vy_index_insert_stmt(), which increments index->stmt_count.
As a result, index->stmt_count is incremented twice.

Closes #2421

See #2429

Force tarantool-common upgrade to support systemd notifications.

+ Add missing "s" suffix to TimeoutStartSec= option.

See #1923

Check that there are no statements between prev_stmt and stmt
in the cache on vy_cache_add() before trying to build a chain.
This workaround makes vy_cache_add() more fool-proof for cases
when vy_read_iterator skips some keys during restoration.

Between prepare and commit of a transaction it is possible
that some prepared statements are read from mem.
Prepared statement has abnomal lsn > MAX_LSN.

If version of the index is changed after return of a
prepared statement and the iterator is placed to read view,
the restart and restore on a such prepared statement could
lead to restore on LSN bigger than vlsn.

Example:

        FIBER 1                        FIBER 2
       box.begin()
 replace {1}, lsn=MAX_LSN+1
 replace {2}, lsn=MAX_LSN+1
    prepare vy_tx ...
                        ->->->      open iterator
                              read {1},lsn=MAX_LSN+1 from mem
                                     yield ...
                        <-<-<-
    commit vy_tx
 replace {1, 1},lsn=100
send iterator to read view
                        ->->->
                                 iterator read view = 100
                          index version changed - restore on last_stmt
                              last_stmt = {1},lsn=MAX_LSN+1
                              last_stmt LSN > iterator VLSN
                                 ??????????????????????

Lets return in such situations the next key.

If vy_mem_iterator was not started, the vy_mem_iterator_restore had been
able to iterate to the same key, as the target of restore, and then call
vy_mem_iterator_next_lsn_impl to find the older lsn.

But the vy_mem_iterator_start_from did not set 'search_started' flag and
next_lsn_impl would restart the mem iterator again, to the first statement,
which could be incorrect result of restore.

Example:

cache statements: {1}, {2}, {3}
mem statements: {1}, {2}, {3}
iterator: GE
key: {1}

- merge_iterator returns all cached data and then tries to restore mem on {3}.
- then mem_iterator during restore calls vy_mem_iterator_start_from and sees tuple {3}.
- then mem_iterator tries to find older lsn and calls vy_mem_iterator_next_lsn_impl,
  which sees 'search_started = false' and restarts iterator to {1}.
- then read_iterator returns {1} after {3} - ERROR!

Lets set 'search_started' flag inside vy_mem_iterator_start_from.

Under heavy load it is possible, that mem_iterator_restore
inside read_iterator_next returns statement with the same key,
as the previous one.

Check keys in vy_read_iterator_merge_next_key.

If the mem_iterator is started, but hasn't any curr_stmt (maybe it
was finished), it could return its first statement regardless of last_stmt.

Lets call start_from() instead of start() in case of not null last_stmt.

Mem iterator restoration in case of LE and LT was completelly wrong.
Additionally I added more asserts and found that restoration is
called too frequently.
Fix'em both.

fix #2207

The code calculating the quota watermark was written long ago, when we
didn't have the common memory level and didn't use lsregion allocator
and hence it was possible to free all memory used by any range. Things
have changed drastically since then: now it's impossible to free memory
occupied by a range or even an index, because statements are shared
between indexes, so the scheduler effectively dumps all memory on
exceeding the quota. Fix the quota calculation accordingly.

We should use fiber_sleep(0) to yield periodically, not
fiber_reschedule() - the latter doesn't advance the event loop.

See 56462bca

Check that we correctly recover in case not all indexes of the same
space were dumped.

We remove the primary index from the scheduler while we are dumping a
secondary index of the same space ('pin' it). On dump task completion we
add it back. By that time, the primary index could have been dropped, in
which case we may get use-after-free in vy_scheduler_unpin_index(). To
avoid it we should increment a reference counter of a pinned index.

I need to set errinj parameter to a negative value (errinj.param < 0 -
errinj is disabled, errinj.param = X >= 0 - errinj is enabled for X).
I don't think it's a good idea to introduce int64_t in addtion to
uint64_t, because int64_t should be enough for error injection. So,
let's change ERRINJ_U64 (and errinj.u64param) to ERRINJ_INT
(errinj.iparam) of type int64_t.