box.ctl.demote() used not to do anything with election_mode='off'
if the synchro queue didn't belong to the caller in the same term
as the election state.

The reason could be that if the synchro queue term is "outdated",
there is no guarantee that some other instance doesn't own it in
the latest term right now.

The "problem" is that this could be workarounded easily by just
calling promote + demote together.

There isn't much sense in fixing it for the off-mode because the
only reasons off-mode exists are 1) for people who don't use
synchro at all, 2) who did use it and want to stop. Hence they
need demote just to disown the queue.

The patch "legalizes" the mentioned workaround by allowing to
perform demote in off-mode even if the synchro queue term is old.

Closes #6860

NO_DOC=bugfix

(cherry picked from commit 1afe2274)

`key_part::offset_slot_cache` and `key_part::format_epoch` are used for
speeding up tuple field lookup in `tuple_field_raw_by_part()`. These
structure members are accessed and updated without any locks, assuming
this code is executed exclusively in the tx thread. However, this isn't
necessarily true because we also perform tuple field lookups in vinyl
read threads. Apparently, this can result in unexpected races and bugs,
for example:

```
  #1  0x590be9f7eb6d in crash_collect+256
  #2  0x590be9f7f5a9 in crash_signal_cb+100
  #3  0x72b111642520 in __sigaction+80
  #4  0x590bea385e3c in load_u32+35
  #5  0x590bea231eba in field_map_get_offset+46
  #6  0x590bea23242a in tuple_field_raw_by_path+417
  #7  0x590bea23282b in tuple_field_raw_by_part+203
  #8  0x590bea23288c in tuple_field_by_part+91
  #9  0x590bea24cd2d in unsigned long tuple_hint<(field_type)5, false, false>(tuple*, key_def*)+103
  #10 0x590be9d4fba3 in tuple_hint+40
  #11 0x590be9d50acf in vy_stmt_hint+178
  #12 0x590be9d53531 in vy_page_stmt+168
  #13 0x590be9d535ea in vy_page_find_key+142
  #14 0x590be9d545e6 in vy_page_read_cb+210
  #15 0x590be9f94ef0 in cbus_call_perform+44
  #16 0x590be9f94eae in cmsg_deliver+52
  #17 0x590be9f9583e in cbus_process+100
  #18 0x590be9f958a5 in cbus_loop+28
  #19 0x590be9d512da in vy_run_reader_f+381
  #20 0x590be9cb4147 in fiber_cxx_invoke(int (*)(__va_list_tag*), __va_list_tag*)+34
  #21 0x590be9f8b697 in fiber_loop+219
  #22 0x590bea374bb6 in coro_init+120
```

Fix this by skipping this optimization for threads other than tx.

No test is added because reproducing this race is tricky. Ideally, bugs
like this one should be caught by fuzzing tests or thread sanitizers.

Closes #10123

NO_DOC=bug fix
NO_TEST=tested manually with fuzzer

(cherry picked from commit 19d1f1cc)

The tuple cache doesn't store older tuple versions so if a reader is
in a read view, it must skip tuples that are newer than the read view,
see `vy_cache_iterator_stmt_is_visible()`. A reader must also ignore
cached intervals if any of the tuples used as a boundary is invisible
from the read view, see `vy_cache_iterator_skip_to_read_view()`.
There's a bug in `vy_cache_iterator_restore()` because of which such
an interval may be returned to the reader: when we step backwards
from the last returned tuple we consider only one of the boundaries.
As a result, if the other boundary is invisible from the read view,
the reader will assume there's nothing in the index between the
boundaries and skip reading older sources (memory, disk). Fix this by
always checking if the other boundary is visible.

Closes #10109

NO_DOC=bug fix

(cherry picked from commit 7b72080d)

If a run iterator is positioned at a non-terminal statement (UPSERT or
UPDATE), `vy_run_iterator_next()` will iterate over older statements
with the same key using `vy_run_iterator_next_lsn()` to build the key
history. While doing so, it may reach the end of the run file (if the
current key is the last in the run). This would stop iteration
permanently, which is apparently wrong for reverse iterators (LE or LT):
if this happens the run iterator won't return any keys preceding the
last one in the run file. Fix this by removing `vy_run_iterator_stop()`
from `vy_run_iterator_next_lsn()`.

Part of #10109

NO_DOC=bug fix
NO_CHANGELOG=next commit

(cherry picked from commit 72763f94)

We're using LXD containers as aarch64 runners. For some reason, OOM
killer just kills the compilation process while package building when
`make -j $(nproc)`. The issue happens only with builds where LTO is
enabled. It's found, that `-j6` works fine. The bigger value causes
problems.

NO_DOC=ci
NO_TEST=ci
NO_CHANGELOG=ci

Bump test-run to new version with the following improvements:

- Calculate parallel jobs based on available CPUs [1]
- Bump luatest to 1.0.1-15 (--list-test-cases) [2]
- luatest: detox test searching code [3]
- luatest: allow to run test cases in parallel [4]

[1] tarantool/test-run@182aa77
[2] tarantool/test-run@1fbbf9a
[3] tarantool/test-run@3b0ccd0
[4] tarantool/test-run@dd00063

NO_DOC=test
NO_TEST=test
NO_CHANGELOG=test

(cherry picked from commit 32bcea7d)

Disable workaround for LuaJIT profiling tests on aarch64 runners due to
the following error:

    mount: /tmp/luajit-test-vardir: mount failed: Operation not permitted

Looks like it happens because our aarch64 runners are LXD containers.

NO_DOC=ci
NO_TEST=ci
NO_CHANGELOG=ci

(cherry picked from commit e64457d9)

`vy_apply_result_does_cross_pk()` must be called after the new tuple
format is validated, otherwise it may crash in case the new tuple has
fields conflicting with the primary key definition.

While we are at it, fix the operation cursor (`ups_ops`) not advanced
on this kind of error. This resulted in skipped `upsert` statements
following an invalid `upsert` statement in a transaction.

Closes #10099

NO_DOC=bug fix

(cherry picked from commit dd0ac814)

If a secondary index is altered in such a way that its key parts are
extended with the primary key parts, rebuild isn't required because
`cmp_def` doesn't change, see `vinyl_index_def_change_requires_rebuild`.
In this case `vinyl_index_update_def` will try to update `key_def` and
`cmp_def` in-place with `key_def_copy`. This will lead to a crash
because the number of parts in the new `key_def` is greater.

We can't use `key_def_dup` instead of `key_def_copy` there because
there may be read iterators using the old `key_def` by pointer so
there's no other option but to force rebuild in this case.

The bug was introduced in commit 64817066 ("vinyl: use update_def
index method to update vy_lsm on ddl").

Closes #10095

NO_DOC=bug fix

(cherry picked from commit 9b817848)

A DML request (insert, replace, update) can yield while reading from
the disk in order to check unique constraints. In the meantime the index
can be dropped. The DML request can't crash in this case thanks to
commit d3e12369 ("vinyl: abort affected transactions when space is
removed from cache"), but the DDL operation can because:
 - It unreferences the index in `alter_space_commit`, which may result
   in dropping the LSM tree with `vy_lsm_delete`.
 - `vy_lsm_delete` may yield in `vy_range_tree_free_cb` while waiting
   for disk readers to complete.
 - Yielding in commit triggers isn't allowed (crashes).

We already fixed a similar issue when `index.get` crashed if raced
with index drop, see commit 75f03a50 ("vinyl: fix crash if space is
dropped while space.get is reading from it"). Let's fix this issue in
the same way - by taking a reference to the LSM tree while checking
unique constraints. To do that it's enough to move `vy_lsm_ref` from
`vinyl_index_get` to `vy_get`.

Also, let's replace `vy_slice_wait_pinned` with an assertion checking
that the slice pin count is 0 in `vy_range_tree_free_cb` because
`vy_lsm_delete` must not yield.

Closes #10094

NO_DOC=bug fix

(cherry picked from commit bde28f0f)

`tuple_hash_field()` doesn't advance the MsgPack cursor after hashing
a tuple field with the type `double`, which can result in crashes both
in memtx (while inserting a tuple into a hash index) and in vinyl
(while writing a bloom filter on dump or compaction).

The bug was introduced by commit 51af059c ("box: compare and hash
msgpack value of double key field as double").

Closes #10090

NO_DOC=bug fix

(cherry picked from commit bc0daf99)

Bump test-run to new version with the following improvements:

- Bump luatest to 1.0.1-14-gdfee2f3 [1]
- Adjust test result report width to terminal size [2]
- dispatcher: lift pipe buffer size restriction [3]
- flake8: fix E721 do not compare types [4]

[1] tarantool/test-run@84ebae5
[2] tarantool/test-run@1724211
[3] tarantool/test-run@81259c4
[4] tarantool/test-run@1037299

We also have to fix several tests that check that script with luatest
assertions have empty stderr output. test-run brings Luatest which
logs assertions at 'info' level.

Note that gh_8433_raft_is_candidate_test is different. Original
assertion involves logging huge tables that have somewhere closed
sockets inside. And 'socket.__tostring' currently raises error for
closed sockets.

NO_DOC=submodule bump
NO_TEST=submodule bump
NO_CHANGELOG=submodule bump

(cherry picked from commit 97a801e1)

Bump test-run to new version with the following improvements:

- Bump luatest to 1.0.1-5-g105c69d [1]
- tap13: fix worker fail on failed TAP13 parsing [2]

[1] tarantool/test-run@ed5b623
[2] tarantool/test-run@7c1a0a7

NO_DOC=test
NO_TEST=test
NO_CHANGELOG=test

(cherry picked from commit 4466deaf)

- Fix the `test-sdk` job according to recent changes in SDK
- Fix sending VK Teams message on failure

NO_DOC=ci
NO_TEST=ci
NO_CHANGELOG=ci

Logically, we call triggers after running statements. These triggers can
make significant changes (for instance, DDL triggers), so, for consistency,
we should call the statement's `on_rollback` triggers before rolling back
the statement. This also adheres to the logic that transaction
`on_rollback` triggers are called before rolling back individual
transaction statements.

One particular bug that this patch fixes is rolling back of DDL on the
`_space` space. DDL is essentially a replace operation on the `_space`
space, which also invokes the `on_replace_dd_space` trigger. In this
trigger, among other things, we swap the indexes of the original space,
`alter->old_space`, which is equal to the corresponding transaction
`stmt->space`, with the indexes of the newly created space,
`alter->new_space`:
https://github.com/tarantool/tarantool/blob/de80e0264f7deb58ea86ef85b37b92653a803430/src/box/alter.cc#L1036-L1047

If then a rollback happens, we first rollback the replace operation, using
`stmt->space`, and only after that do we swap back the indexes in
`alter_space_rollback`:
https://github.com/tarantool/tarantool/blob/de80e0264f7deb58ea86ef85b37b92653a803430/src/box/memtx_engine.cc#L659-L669
https://github.com/tarantool/tarantool/blob/de80e0264f7deb58ea86ef85b37b92653a803430/src/box/alter.cc#L916-L925

For DDL on the _space space, the replace operation and DDL occur on the
same space. This means that during rollback of the replace, we will try to
do a replace in the empty indexes that were created for `alter->new_space`.
Not only does this break the replace operation, but also the newly inserted
tuple, which remains in the index, gets deleted, and access to it causes
undefined behavior (heap-use-after-free).

As part of the work on this patch, tests of rollback of DDL on system
spaces which use `on_rollback` triggers were enumerated:
* `_sequence` — box/sequence.test.lua;
* `_sequence_data` — box/sequence.test.lua;
* `_space_sequence` — box/sequence.test.lua;
* `_trigger` — sql/ddl.test.lua, sql/errinj.test.lua;
* `_collation` — engine-luatest/gh_4544_collation_drop_test.lua,
                 box/ddl_collation.test.lua;
* `_space` — box/transaction.test.lua, sql/ddl.test.lua;
* `_index` — box/transaction.test.lua, sql/ddl.test.lua;
* `_cluster` — box/transaction.test.lua;
* `_func` — box/transaction.test.lua, box/function1.test.lua;
* `_priv` — box/errinj.test.lua,
            box-luatest/rollback_ddl_on__priv_space_test.lua;
* `_user` — box/transaction.test.lua,
            box-luatest/gh_4348_transactional_ddl_test.lua.

Closes #9893

NO_DOC=<bugfix>

(cherry picked from commit d529082f)

In scope of #9893 we are going to run statement `on_rollback` triggers
before rolling back the corresponding statement. During rollback of DDL in
the `_priv` space, the database is accessed from `user_reload_privs` to
reload user privileges, so we need it to account for the current statement
being rolled back: i.e., the new tuple that was introduced (if any) must
not be used, while the old tuple (if any) must be used.

Needed for #9893

NO_CHANGELOG=<refactoring>
NO_DOC=<refactoring>

(cherry picked from commit 797c04ff)

Currently on_rollback triggers are called on rollback of the whole
transaction. To make it possible to invoke them on rollback to a
savepoint, we need to pass a statement at which the savepoint was
created.

Needed for #9340

NO_DOC=refactoring
NO_TEST=refactroring
NO_CHANGELOG=refactoring

(cherry picked from commit a1d85827)

It could fail in ASAN build. Can't tell why just there.

The main reason was that in a topology server1 + server2->server3
one of the cases
- did a txn on server1,
- then enabled server2->server3 replication,
- then waited for server2->server3 sync,
- and instantly assumed the txn reached server3.

Surely it not always did. At the server2->server3 sync the txn
might not had reached server2 itself yet.

The fix is as simple as explicitly ensure the txn is on server2
before waiting server2->server3 sync.

Another potential for flakiness was that the default timeout in
luatest.helpers.retrying is super low, just 5 seconds. The patch
manually bumps it to 60 seconds to be sure any future failures
wouldn't be related to too small timeout.

Closes #10031

NO_DOC=test
NO_CHANGELOG=test

(cherry picked from commit d4ea121b)

Bump the OpenSSL library version to 3.2.1 and remove OpenSSL patches which
are already present in the updated library version.

Disable modules in OpenSSL configuration to make sure the OpenSSL 3.0
legacy provider is compiled into the library.

Closes #7502

NO_DOC=<dependency bump>
NO_TEST=<dependency bump>

Co-authored-by: Sergey Bronnikov <sergeyb@tarantool.org>
(cherry picked from commit 8de22969)

```
/home/shiny/dev/tarantool/src/lib/core/coio_task.c:114:58:
	error: ‘calloc’ sizes specified with ‘sizeof’ in the earlier argument
	and not in the later argument [-Werror=calloc-transposed-args]
  114 |         struct cord *cord = (struct cord *)calloc(sizeof(struct cord), 1);
```

NO_TEST=build fix
NO_CHANGELOG=build fix
NO_DOC=build fix

(cherry picked from commit fb6b6c60)

When upgrading a space, attribute `has_optional_parts` of indexes can be
changed. So in order to correctly index both old and new tuples we should
set new min_field_count value to the minimal min_field_count of old and
new formats. Actual value will be set when space upgrade completes.

Part of tarantool/tarantool-ee#698
Part of tarantool/tarantool-ee#750

NO_TEST=in ee
NO_CHANGELOG=in ee
NO_DOC=bugfix

(cherry picked from commit c449ada4)

wal_queue_max_size took effect only after the initial box.cfg call,
meaning that users with non-zero `replication_sync_timeout` still synced
using the default 16 Mb queue size. In some cases the default was too
big and the same issues described in #5536 arose.

Fix this.

Closes #10013

NO_DOC=bugfix

(cherry picked from commit ab0f7913)

The code setting ER_TUPLE_FOUND uses index_name_by_id() to find
the index name, but it passes an index in the dense index map to
it while the function expects an index in the sparse index map.
Apparently, this doesn't work as expected after an index is removed
from the middle of the index map. This bug was introduced by
commit fc3834c0 ("vinyl: check key uniqueness before modifying
tx write set").

Instead of just fixing the index passed to index_name_by_id(), we do
a bit of refactoring. We stop passing index_name and space_name to
vy_check_is_unique_*() functions and instead get them right before
raising ER_TUPLE_FOUND. Note, to get the space name, we need to call
space_by_id() but it should be fine because (a) the space is very likely
to be cached as the last accessed one and (b) this is an error path so
it isn't performance critical. We also drop index_name_by_id() and
extract the index name from the LSM tree object.

Closes #5975

NO_DOC=bug fix

(cherry picked from commit 2cfba5eb)

Like UPDATE, UPSERT must not modify primary key parts. Unlike UPDATE,
such an invalid UPSERT statement doesn't fail (raise an error) - we
just log the error and ignore the statement. The problem is, we don't
clear txn_stmt. As a result, if we're currently building a new index,
the on_replace trigger installed by the build procedure will try to
process this statement, triggering the assertion in the transaction
manager that doesn't expect any statements in a secondary index without
the corresponding statement in the primary index:

  ./src/box/vy_tx.c:728: vy_tx_prepare:
    Assertion `lsm->space_id == current_space_id' failed.

Let's fix this by clearing the txn_stmt corresponding to a skipped
UPSERT.

Note, this also means that on_replace triggers installed by the user
won't run on invalid UPSERT (hence test/vinyl/on_replace.result update),
but this is consistent with the memtx engine, which doesn't run them
in this case, either.

Closes #10026

NO_DOC=bug fix

(cherry picked from commit 5ac0d26a)

Not only for own txns, but also on the txns authored by other
instances.

Note that the lag isn't updated when the replica got new txns from
another master. The lag still only reflects the replication
between this relay and its specific applier.

The motivation is that otherwise the lag sometimes shows
irrelevant things, like that the replica is very outdated, while
it keeps replicating just fine. Only not txns of this specific
master, who might even turned into a replica itself already.

Closes #9748

NO_DOC=bugfix

(cherry picked from commit 39af9fbe)

From the code it isn't obvious, but relay->status_msg.vclock and
relay->last_recv_ack.vclock are both coming from the applier.
Status_msg is the previous ack, last_recv_ack is the latest ack.

They can never go down. And are not affected anyhow by the master
committing its own transactions. I.e. master can commit something,
relay->r->vclock (recovery cursor) will go up, and recovery vclock
might become incomparable with the last ACK vclock. But the prev
and last ACK vclocks are always comparable and always go up.

This invariant was broken though, because relay on restart didn't
nullify the current applier status (status_msg). It could break
if the replica would loose its xlog files or its ID would be
taken by another instance - then its vclock would go down, making
last_recv_ack.vclock < status_msg.vclock. But that is not right
and is fixed in this patch.

In scope of #9748

NO_DOC=bugfix
NO_TEST=test 5158 already covers it
NO_CHANGELOG=bugfix

(cherry picked from commit 71dbb47c)

To reduce the insane indentation level. And to isolate the further
changes in next commits more.

Part of #9748

NO_DOC=refactoring
NO_TEST=refactoring
NO_CHANGELOG=refactoring

(cherry picked from commit d6f15a10)

Before the patch if the applier was reconnected, the master would
see downstream lag equal to the time since it replicated the last
txn to this applier.

This happened because applier between reconnects kept the txn
timestamp used for acks. On the master's side the relay was
recreated, received the ack, thought the applier just applied this
txn, and displayed this as a lag.

The test makes a master restart because this is the easiest way to
reproduce it. Most importantly, the applier shouldn't be
re-created, and relay should restart.

Part of #9748

NO_DOC=bugfix
NO_CHANGELOG=later

(cherry picked from commit dda42035)

It was stored in struct replica, now is in struct applier. The
motivation is that applier-specific data must be inside the
applier.

Also it makes the next commits look more logical. They are going
to change this timestamp when applier progresses through its state
machine. It looks strange when the applier is changing the replica
object. Replica is on an upper level in the hierarchy. It owns the
applier and the applier ideally mustn't know about struct replica
(hardly possible to achieve), or at least not change it (this is
feasible).

In scope of #9748

NO_DOC=internal
NO_TEST=refactoring
NO_CHANGELOG=refactoring

(cherry picked from commit 8e5d9f2a)

A unique nullable key definition extended with primary key parts
(cmp_def) assumes that two tuples are equal *without* comparing
primary key fields if all secondary key fields are equal and not
nulls, see tuple_compare_slowpath(). This is a hack required to
ignore the uniqueness constraint for nulls in memtx. The memtx
engine can't use the secondary key definition as is (key_def) for
comparing tuples in the index tree, as it does for a non-nullable
unique index, because this wouldn't allow insertion of any
duplicates, including nulls. It couldn't use cmp_def without this
hack, either, because then conflicting tuples with the same
secondary key fields would always compare as not equal due to
different primary key parts.

For Vinyl, this hack isn't required because it explicitly skips
the uniqueness check if any of the indexed fields are nulls, see
vy_check_is_unique_secondary(). Furthermore, this hack is harmful
because Vinyl relies on the fact that two tuples compare as equal by
cmp_def if and only if *all* key fields (both secondary and primary)
are equal. For example, this is used in the transaction manager,
which overwrites statements equal by cmp_def, see vy_tx_set_entry().

Let's disable this hack by resetting unique_part_count in cmp_def.

Closes #9769

NO_DOC=bug fix

(cherry picked from commit 2e689063)

Between picking an LSM tree from a heap and taking a reference to it in
vy_task_new() there are a few places where the scheduler may yield:
 - in vy_worker_pool_get() to start a worker pool;
 - in vy_task_dump_new() to wait for a memory tree to be unpinned;
 - in vy_task_compaction_new() to commit an entry to the metadata log
   after splitting or coalescing a range.

If a concurrent fiber drops and deletes the LSM tree in the meanwhile,
the scheduler will crash. To avoid that, let's take a reference to
the LSM tree.

It's quite difficult to write a functional test for it without a bunch
of ugly error injections so we rely on fuzzing tests.

Closes #9995

NO_DOC=bug fix
NO_TEST=fuzzing

(cherry picked from commit 1c4605bb)

It encapsulates all the needed actions to connect to a remote console
using a Unix socket.

Part of #9985

NO_DOC=testing helper change
NO_CHANGELOG=see NO_DOC

(cherry picked from commit bb430c55)

See #7169 for details about the hide/show prompt feature. In short, it
hides readline's prompt before `print()` or `log.<level>()` calls and
restores the prompt afterwards.

This feature sometimes badly interferes with
`test.interactive_tarantool` heuristics about readline's command
echoing.

This commit disables the feature in `test.interactive_tarantool` by
default and enables it explicitly where needed.

Part of #9985

NO_DOC=testing helper change
NO_CHANGELOG=see NO_DOC

(cherry picked from commit 23094b6f)

Before this patch the `:roundtrip()` method in the
`test.interative_tarantool` instance considered the following calls as
equivalent:

```lua
g.it = it.new()

-- Doesn't check the response.
g.it:roundtrip('x')

-- Before the patch it was the same as above.
--
-- Now it checks that the response is nil.
local expected = nil
g.it:roundtrip('x', expected)

-- It is the same as previous.
g.it:roundtrip('x', nil)
```

Now the response is checked against the provided expected value if the
value is passed to arguments, even if it is `nil`.

Also, a command's response is now returned from the method. It may be
useful if the response returns some dynamic information (such as a TCP
port number or a file descriptor) that is used later in the test or if
the response should be verified in some non-trivial way, not just a deep
compare.

The `:roundtrip()` method is just `:execute_command()` plus
`:read_response()` plus `luatest.assert_equals()`. However, I like using
`:roundtrip()` even when the assertion is not needed, because it is
shorter and because using the same method brings less context to a
reader.

For example,

```lua
g.it = it.new()
g.it:roundtrip('x = 2')
g.it:roundtrip('y = 3')
g.it:roundtrip('x + y', 6)
```

Part of #9985

NO_DOC=testing helper change
NO_CHANGELOG=see NO_DOC
NO_TEST=see NO_DOC

(cherry picked from commit 7d9e8569)

When a replica subscribes, it might in the beginning try to
position its reader cursor to the end of a large xlog file.

Positioning inside of this file can take significant time during
which the WAL reader yielded and tried to send heartbeats, but
couldn't, because the relay thread wasn't communicating with the
TX thread.

When there are no messages from TX for too long time, the
heartbeats to the replica are not being sent
(commit 56571d83 ("raft: make
followers notice leader hang")).

The relay must communicate with the TX thread even when subscribe
is just being started and opens a large xlog file.

This isn't the first time when the missing heartbeats result into
timeouts. See more here:

- commit 30ad4a55 ("relay: yield
    explicitly every N sent rows").

- commit 17289440 ("recovery: make
    it yield when positioning in a WAL").

- commit ee6de025 ("relay: send
    heartbeats while reading a WAL").

Given that this is fixed fourth time, it might suggest that the
relay has not the best architecture having some slight drawbacks.
See more in #9968.

Closes #9094

NO_DOC=bugfix

(cherry picked from commit f7e6686a)

By default, each job in a workflow can run for up to 6 hours of the
execution time. If a job reaches this limit, the job is terminated by
GitHub automatically and fails to complete. This patch sets job timeouts
to 60 minutes to avoid waiting for jobs to complete for 6 hours.

NO_DOC=ci
NO_TEST=ci
NO_CHANGELOG=ci

Currently, we close the transport from transport from
`luaT_netbox_transport_stop`, and we do not wait for the worker fiber to
stop. This causes several problems.

Firstly, the worker can switch context by yielding (`coio_wait`) or
entering the Lua VM (`netbox_on_state_change`). During a context switch,
the connection can get closed. When the connection is closed, its receive
buffer is reset. If there was some pending response that was partially
retrieved (e.g., a large select), then after resetting the buffer we will
read some inconsistent data. We must not allow this to happen, so let's
check for this case after returning from places where the worker can switch
context. In between closing the connection and cancelling the connection's
worker, an `on_disconnect` trigger can be called, which, in turn, can
also yield, returning control to the worker before it gets cancelled.

Secondly, when the worker enters the Lua VM, garbage collection can be
triggered and the connection owning the worker could get closed
unexpectedly to the worker.

The fundamental source of these problems is that we close the transport
before the worker's loop stops. Instead, we should close it after the
worker's loop stops. In `luaT_netbox_transport_stop`, we should only cancel
the worker, and either wait for the worker to stop, if we are not executing
on it, or otherwise throw an exception (`luaL_testcancel`) to stop the
worker's loop. The user will still have the opportunity to catch this
exception and prevent stoppage of the worker at his own risk. To safeguard
from this scenario, we will now keep the `is_closing` flag enabled once
`luaT_netbox_transport_stop` is called and never disable it.

There also still remains a special case of the connection getting garbage
collected, when it is impossible to stop the worker's loop, since we cannot
join the worker (yielding is forbidden from finalizers), and an exception
will not go past the finalizer. However, this case is safe, since the
connection is not going to be used by this point, so the worker can simply
stop on its own at some point. The only thing we need to account for is
that we cannot wait for the worker to stop: we can reuse the `wait` option
of `luaT_netbox_transport_stop` for this.

Closes #9621
Closes #9826

NO_DOC=<bugfix>

Co-authored-by: Vladimir Davydov <vdavydov@tarantool.org>
(cherry picked from commit fcf7f5c4)

Cherry pick note: Dropped gh_9621_netbox_worker_crash_test because
box.iproto.encode helpers aren't available on 2.11.

Also, remove unreleased/ entries.

NO_DOC=changelog
NO_TEST=changelog
NO_CHANGELOG=changelog

Before this patch MVCC engine expected that if index_replace
sets `result` to NULL then index_replace sets `successor` to
something (NULL or existing tuple, depending on index type).
That looked fine because by contract `successor` is set when
true insertion was happened.

Unfortunately it was not considered that in case of part with
`exclude_null` option in index the insertion can be silently
skipped and thus `successor` can be not set. The latter access
of it was actually an UB.

Fix it by explicit check of tuple_key_is_excluded and work on
this case correctly.

Note that logically `index_replace` should return a flag whether
the new tuple was filtered (excluded) by key_def. But on the other
hand this flag is required only for mvcc while the function is
already has lots of arguments and it's very cheap to determine
this flag right from memtx_tx, so I decided to make the most
simple patch.

NO_DOC=bugfix

(cherry picked from commit 14e21297)