Before this patch it was allowed to rename space which is referenced by
a view. In turn, view contains SELECT statement in a raw form (i.e. as a
string) and it is not modified during renaming routine. Hence, after
renaming space still has referencing counter > 0, but no usage of view
is allowed (since execution of SELECT results in "Space does not
exist"). To avoid such situations, lets ban renaming space if its view
reference counter > 0.

Note that RENAME is ANSI extension, so different DBs behave in this case
in different ways - some of them allow to rename tables referenced by a
view (PostgreSQL), others - don't (Oracle).

Closes #3746

The space_def_destroy_fields routine is used in make_scoped_guard
on alter:space_def_new_from_tuple always pass extern_alloc=true
for sql_expr_delete routine. It shouldn't as the first-time
allocated AST object (field_def_decode) is not external-allocated.
Introduced a new flag 'extern_alloc' for space_def_destroy_fields
routine.

Also fixed def_guard declaration in space_def_new_from_tuple:
it should be right after space_def_new_xc call, otherwise
subsequent tnt_raise/diag_raise call wouldn't fire it.

Closes #3908

Reworked ER_FIELD_TYPE and ER_ACTION_MISMATCH error to pass path
string to field instead of field number.
This patch is required for further JSON patches, to give detailed
information about field on error.

Needed for #1012

This function and the associated table don't depend on key definition,
only on field type. So let's move it from key_def.h to field_def.h.

 - Reworked field types and nullability checks to set error message in
   tuple_init_field_map manually. We will specify full JSON path to the
   field further patches.
 - Introduced field_mp_type_is_compatible routine making field_type and
   mp_type compatibility test taking into account field nullability.
 - Refactored key_part_validate to pass const char *key argument and to
   reuse field_mp_type_is_compatible code.

Needed for #1012

Since the tuple_validate_raw and tuple_init_field_map functions
make the same data checks, we implemented tuple_validate_raw via
tuple_init_field_map called with region memory chunk is passed
as field map. This is required because in subsequent patches
the tuple_init_field_map routine logic will be complicated,
and we want to avoid writing the same checks twice.

Needed for #1012

While building a secondary index for a Vinyl space, we use xm->lsn to
skip statements inserted after build began (as those statements are
propagated by the on_replace trigger callback anyway). The xm->lsn is
advanced every time we insert something into a Vinyl space. The problem
is during recovery we skip statements that have been dumped to disk so
xm->lsn may lag behind the instance vclock once recovery is complete.
In this case if we try to create a new Vinyl index right after recovery
completion, before any DML operation is executed on a Vinyl space, we
will skip statements that would otherwise get into the new index.

Fix this issue by resetting xm->lsn to the instance vclock upon recovery
completion.

Closes #3903

The build time error is confusing for users. The patch makes the error
explicit and moves it to the cmake stage.

Follow up of #3070.

Once a connection is closed, a long-running user
request can not learn about this occasion. Even
box.sesion.push() still works.

This patch makes such disconnected session 'rotten'.
So a user can determine if a connection is closed by
looking at session.fd() == -1, or checking for
errors from box.session.push().

Closes #3859

Before the patch vtabs were stored in a global array
only, called registry and accessed by session.type to
call a virtual method. But it does not allow to change
session vtab without affecting session type. This
feature is necessary to be able to outdate binary
sessions whose socket is closed.

Outdated, closed, sessions will return errors from all
its methods.

Needed for #3859

This reverts commit 022a3c50.

According to FFI documentation [1], ffi.cast() creates a new object for
the given type and initializes it with the given value. So setting a
finalizer for a tuple before casting it to const_tuple_ref_t is totally
wrong as it allows the Lua interpreter to invoke the finalizer as soon
as the cast is complete. That said, we must revert the commit that
swapped ffi.cast and ffi.gc order in tuple_bless() and reopen the issue
it intended to fix - see #3751 - no wonder it improved the garbage
collector performance as it basically made all tuples collectable right
from the moment they are blessed.

To make sure we won't step on the same issue again in future, let's also
add a relevant test case.

[1] http://luajit.org/ext_ffi_api.html

Closes #3902

tuple_format is a complex reference-counted structure so it isn't
always possible to mark it const in a function argument list even if
the function in fact doesn't modify the format. So far we've managed
to retain const qualifier in a bunch of functions that don't involve
reference counting, however with the introduction of indexed JSON
fields we can't do it anymore, because the underlying json_tree
structure doesn't use const qualifier at all. So let's drop const
qualifier everywhere to avoid annoying type conversions.

Needed for #1012

tuple_format_vtab::tuple_new() callback is supposed to return a tuple
with 0 references, but vinyl implementation returns a tuple with the
reference counter set to 1, effectively leaking the new tuple.

How to reproduce: get a tuple from a vinyl space and update it with
box.tuple.update().

The bug was introduced by commit a377332a ("tuple: fix lost format
on update and upsert operations").

The issue has been closed by commit 3eb796a1. To check that
everything is working fine, a test has been added.

Closes #3644

LLVM changes the API in the compiler-rt commit [1]. According to the
date of the commit it is between clang-3.9.0 and clang-3.9.1.

[1]: https://github.com/llvm-mirror/compiler-rt/commit/0b95585616bd28fc0b738289bcc5f7887d7c304e

Fixes #3070.

When we replace txv->stmt in vy_tx_handle_deferred_delete, we must also
fix up tx_manager->write_set_size.

Fixes commit a6edd455 ("vinyl: eliminate disk read on
REPLACE/DELETE").

Closes #3897

The calloc() in parse_headers() leads to lost of 32 bytes on each
http_client request.

The problem was introduced in 1.9.1-38-g3d121dd4a.

Found with ASAN.

Garbage collector should be restarted after it has been stopped.

Follow-up #3224

Add missing closedir() to xdir_collect_inprogress().

Not really critical as this function is only called on recovery.

Fixes commit f41aac61 ("xdir: remove inprogress files after
restart").

If the user does not have permission to read from the “_collation”
space, the “PRAGMA collation_list” command causes a segmentation
fault. Added new check.

Closes #3857

Part of: [https://github.com/tarantool/test-run/issues/113]

* Fix *.reject file directory
* Fixed problem with file descriptors when `use_unix_sockets=True`
* Perform the clean up if param `pre_cleanup = True` in suite.ini.
  It cleans checkpoints in workdir for tap tests. And for app tests
  cleans spaces/users/roles/globals/functions and checks count of
  `box.space._cluster`
* Fix replication-py tests by using `cleanup()` without `full=True`

Fixes #1961

Editorconfig is a standard of defining indentation style and other
settings such as newline types for different file types in a project.

See https://editorconfig.org/

It is achieved by introducing a simple .editorconfig dotfile with
glob expressions and settings per fiile type. Editors then can read
this file and decide which indentation style they should use for a
given project file.

All binary struct *_request are stored in xrow.h/.c
together with their decoders. The only reason why
xrow_decode_sql was implemented in execute.c was a
dependency on struct sql_bind. Now xrow_decode_sql
and struct sql_request operate only by MessagePack and
some iproto constants and are moved to their true
home.

Follow up #3828

Before this patch bind parameters were decoded in
iproto thread in an attempt to save some TX thread
time. But they were stored in an array allocated on
iproto thread memory. Lack of iproto_msg destructor
lead to leak of this array.

This patch makes sql_request store only raw undecoded
MessagePack and moves decoding to TX thread where in
a single function the parameters are decoded, applied
and freed.

There were a couple of alternatives allowing to keep
decoding in iproto thread and to move even more
decoding there, for example, parameters from
IPROTO_UPDATE:

* decode on malloc and add a virtual destructor to
  struct iproto_msg;

* decode on malloc and add a virtual destructor to
  struct cmsg. By the way, it could allow to reuse
  this destructor for struct cbus_call_msg and
  struct relay_gc_msg.

The way how is it solved by this patch has been chosen
thanks to its simplicity.

Closes #3828

This reverts commit bc9e41e9.

This is one step forward to make sql_request store raw
MessagePack.

Part of #3828

sql_request is now decoded entirely in iproto thread
in iproto_msg_decode unlike other similar requests
like CALL or UPDATE which could be decoded in iproto
too.

But iproto thread pays for this optimization with a
leak - sql_request.bind array is allocated on iproto
thread region and never freed.

A fix is to decode, apply and free bind array in tx
thread in one place: tx_process_sql. For this
sql_request should look like other requests: do not
decode arrays, do not store sync, store fields as
raw MsgPack.

First step - remove sync.

Part of #3828

The value of box.cfg.replication_timeout should be the same on all
members of a cluster, otherwise replication may stop on timeout.
Since master.lua has box.cfg.replication_timeout set to 0.1, we should
use the same value in all replica scripts.

This fixes replication/wal_rw_stress test sporadic failure.

Closes #3885

FALLOC_FL_KEEP_SIZE flag has existed since fallocate() was first
introduced, but it was not defined in glibc headers for a while.
Define it manually if necessary.

Follow-up b9db91e1 ("xlog: fix fallocate vs read race").

posix_fallocate(), which is used for preallocating disk space for WAL
files, increases the file size and fills the allocated space with zeros.
The problem is a WAL file may be read by a relay thread at the same time
it is written to. We try to handle the zeroed space in xlog_cursor (see
xlog_cursor_next_tx()), however this turns out to be not enough, because
transactions are written not atomically so it may occur that a writer
writes half a transaction when a reader reads it. Without fallocate, the
reader would stop at EOF until the rest of the transaction is written,
but with fallocate it reads zeroes instead and thinks that the xlog file
is corrupted while actually it is not.

Fix this issue by using fallocate() with FALLOC_FL_KEEP_SIZE flag
instead of posix_fallocate(). With the flag fallocate() won't increase
the file size, it will only allocate disk space beyond EOF.

Closes #3883

VIEW can be created not only in form of AS SELECT but also in form of AS
VALUES (...). In the latter case space's view reference counters were
not incremented. Furthermore, if view came with sub-select, reference
counters of tables within sub-select were not updated as well. These
problems occurred due to simplified traverse over AST: only tables
within FROM clause were accounted. This patch fixes them introducing
complete pass over AST using walker.

Closes #3815

xlog/panic_on_broken_lsn.test.lua fails on
box.schema.space.create('test') after
xlog/checkpoint_threshold.test.lua.

Follow-up db8c7aa3 ("wal: trigger checkpoint if there are
too many WALs").

Add test to check result for indexed char in sub subquery.

Closes #3616

Fix an ability to describe CHAR without specifying the
length as it is allowed by standard. It was
accidentally broken by this commit:
7752cdfd

Length is just an integer, part of a type, not a
whole type.

If VIEW contains constant fields (e.g. CREATE VIEW v AS SELECT 1, 'k';)
it uses string representation of literal as a field name. In the example
above it would be '1' and 'k'. However, if VIEW is created using AS VALUES
syntax, then expressions representing constant literals lack of names
(since span-expression is not assigned in this case).
Lets generate names for all fields which lack names for VIEW.

Closes #3849

Disabled LTO builds, tarballs and packages building on short-term
branches.

Removed 'allow_failures' on coverage / debug build.

Replaced matrix expansion with the list of jobs (because Travis-CI
documentation says it does not support condition jobs with matrix
expansion).

Fixes #3755.