We need to call `tx_accept_msg` in `tx_process_override` before we pass
message to the override handler. Unfortunately if handler response with
IPROTO_HANDLER_FALLBACK we call the builtin handler for message that
calls `tx_accept_msg` again which is not expected. Some actions of
this function are idempotent and some are not.

Let's make the function NOP if it called once again.

Closes #9345

NO_DOC=bugfix

This commit allows to bootstrap an anonymous replica from a replicaset,
where all the instances are in read-only mode.

The reason of the change is that there are no technical reasons to
forbid this action. An anonymous replica is not registered in `_cluster`
system space, so it can join a replicaset even if there are no writable
instances.

Fixes #9432

@TarantoolBot document
Title: config: anonymous replica is now supported

`replication.anon: true` option is now working.

There are configuration constraints that are related to anonymous
replicas.

* A replicaset must contain at least one non-anonymous instance.
* An anonymous replica can't be configured as writable instance using
  `database.mode` or `<replicaset>.leader` options.
* An anonymous replica can't be configured with
  `replication.election_mode` equals to `candidate`, `voter` or `manual`
  (only `off` is allowed).

A few more nuances about anonymous replicas:

* Anonymous replicas are filtered out from default upstream list.
* A `replication.failover: election` replicaset can contain anonymous
  replicas, but `replication.election_mode` defaults to `off` for them
  (unlike non-anonymous instances, where the default is `candidate`).
* `replication.failover: supervised` skips anonymous replicas, when
  choosing a bootstrap leader.
* A anonymous replica can joined a replicaset, which has all the
  instances in read-only mode (unlike a non-anonymous instance).

See details in [1] and [2].

[1]: https://github.com/tarantool/tarantool/issues/9432
[2]: https://github.com/tarantool/tarantool/pull/9418

This commit adds several checks that are specific for
`replication.failover` mode.

* `replication.failover: off`: an anonymous replica shouldn't be set to
  read-write mode using `database.mode` option.
* `replication.failover: manual`: an anonymous replica shouldn't be
  configured as a replicaset leader using `<replicaset>.leader` option.
* `replication.failover: election`: an anonymous replica can't be
  configured with `replication.election_mode` other than `off`.

This commit also adjusts default `replication.election_mode` to `off`
for an anonymous replica if it is part of a `replication.failover:
election` replicaset (the default for a non-anonymous instance is
`candidate`).

Part of #9432

NO_DOC=The documentation request is in the last commit of the series.

A replicaset that contains only anonymous replicas can't be
bootstrapped, because all the instances must be in read-only mode.

Part of #9432

NO_DOC=The documentation request is in the last commit of the series.

Filter out anonymous replicas when choosing a bootstrap leader in
`replication.failover: supervised` mode. An anonymous replica can't be
in read-write mode, so it can't be a replicaset bootstrap leader.

Part of #9432

NO_DOC=It is bugfix. However, this detail is mentioned in the
       documentation request is in the last commit of the series just in
       case.

This commit effectively allows to set `replication.anon: true` without
specifying `replication.peers`.

Without filtering out anonymous replicas from the list of upstreams we
get an error regarding attempt to use an anonymous replica as an
upstream for a non-anonymous instance.

Also, anonymous replicas are excluded from autogenerated upstream list
for other anonymous replicas. It makes the list the same on all the
peers.

A user can configure a custom data flow using `replication.peers`
option.

Part of #9432

NO_DOC=The documentation request is in the last commit of the series.

The commit effectively enables support of anonymous replicas in the
declarative configuration. It has several caveats (see the changelog
entry), which will be resolved in the following commits of the patchset.

An attempt to persist an instance name of an anonymous replica can't be
successful, because it has no entry in `_cluster` system space. Such an
attempt leads to ER_INSTANCE_NAME_MISMATCH error.

This commit patches the configuration applying logic to skip attempt to
set `box.cfg({instance_name = <...>})` if the instance is configured as
an anonymous replica using `replication.anon: true` option.

Part of #9432

NO_DOC=replication.anon option is already documented in the scope of
       https://github.com/tarantool/doc/issues/3851. The bugfix
       shouldn't affect the documentation pages much, however related
       constraints are summarized in a documentation request in the last
       commit of the series.

NO_DOC=code owners file change, no code changes
NO_TEST=see NO_DOC
NO_CHANGELOG=see NO_DOC

Some changes are not new features, rather developer tools updates
and improvements. There are also number of tweaks we can introduce
to improve testing and tests backporting across branches, which
also are not considered neither feature nor bug fix.

NO_DOC=changelog
NO_TEST=changelog
NO_CHANGELOG=changelog

Sometimes the test fails with "Peer closed" error, logs says, that
fatal error happened: cfg_get('read_only'). This is caused by the
instance processing its own ballot.

The problem is that we set box.cfg to function in before_all trigger,
so it's impossible to access box.cfg table during the whole time of test
execution.

Let's instead set box.cfg to function at the start of every test and
restore box.cfg at the end. This way we'll decrease the time, in which
such fatal error can happen. Even though it's still possible to get it
in theory, the problem is not reproduced anymore. The alternative
solution of introducing errinj seems to be overhead here.

Closes tarantool/tarantool-qa#329

NO_DOC=testfix
NO_CHANGELOG=testfix

We need to apply instance/replicaset_name as soon as the instance
becomes RW, so currently we try to do so at every box.status broadcast.
Even though broadcast happens pretty often, it's not enough:

The bug is reproduced in config-luatest/set_names_reload, which checks
the following situation:

    1. Cluster is recovered from the xlogs without names set.
    2. User forgets to set UUID for one replica, starts the cluster.
    3. Replica, UUID of which have not been set, fails to start.
    4. User notices that and updates config, reloading it on the
       instances, which succeeded to start, starting failed one.
    5. Master must apply name for a failed replica.

The test worked all right in the majority of runs, because box.status
broadcast happens often: e.g. it's broadcasted, when master's applier
synced with replica. However, under heavy load on CPU, the test failed
sometimes, when master fails to subscribe on replica and broadcast
doesn't happen.

Let's try to set names not only, when box.status is broadcasted, but
immediately after reload, as at this time new names, which must be
set, might appear. Let's also change test so that, it doesn't rely on
broadcast anymore.

Closes tarantool/tarantool-qa#328

NO_DOC=bugfix

See the details in the documentation request below.

Fixes #9431

@TarantoolBot document
Title: config: failover mode and election mode consistency

`replication.failover: election` enables RAFT based leader election
mechanism on a replicaset. The instances can be configured in the
following election modes: `off`, `candidate`, `voter`, `manual`. It is
controlled by the `replication.election_mode` parameter.

However, the election mode parameter has no sense and confusing for
other failover modes (`off`, `manual`, `supervised`). So, it is
forbidden to set the election modes other than `off` in failover
modes != `election`.

Summary:

* `replication.failover: off`
  * `replication.election_mode: off`: OK
  * `replication.election_mode: candidate`: FAIL
  * `replication.election_mode: voter`: FAIL
  * `replication.election_mode: manual`: FAIL
* `replication.failover: manual`
  * `replication.election_mode: off`: OK
  * `replication.election_mode: candidate`: FAIL
  * `replication.election_mode: voter`: FAIL
  * `replication.election_mode: manual`: FAIL
* `replication.failover: election`
  * `replication.election_mode: off`: OK
  * `replication.election_mode: candidate`: OK
  * `replication.election_mode: voter`: OK
  * `replication.election_mode: manual`: OK
* `replication.failover: supervised`
  * `replication.election_mode: off`: OK
  * `replication.election_mode: candidate`: FAIL
  * `replication.election_mode: voter`: FAIL
  * `replication.election_mode: manual`: FAIL

A next commit adds one more check and it becomes too large snippet of
the code to be part of the constructor function.

NO_DOC=refactoring
NO_CHANGELOG=see NO_DOC
NO_TEST=see NO_DOC

It allows to start a replicaset from the given declarative
configuration, perform checks on all or particular instances, add a new
instance into the replicaset (without stopping existing instances),
update the config file and reload it on the instances.

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

The module now has only one function, which allows to reduce a
boilerplace code needed to verify a scenario, when all instances of a
replicaset should fail with the same error message.

The module will be extended with replicaset management functions later.

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

It allows to construct a declarative configuration for a test case using
less boilerplace code/options, especially when a replicaset is to be
tested, not a single instance. See a description at top of the file for
details.

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

Relay sometimes decodes the PROMOTE packets to be sent in order
to conditionally delay their dispatch. It was believed that own WALs
can't be corrupted and hence there is no point in checking that decoding
succeeds.

Let's be more strict here and check the decoding result before
proceeding.

Closes #9265

NO_DOC=bugfix
NO_TEST=hard to test
NO_CHANGELOG=not user-visible

Closes #9435

@TarantoolBot document
Title: `login` and `password` fields in URI table

Now `uri.parse()` can parse a table containing the `login` and
`password` fields. Values from these fields take precedence over values
obtained from the string URI. For example, login and password of
`{uri = 'one:two:localhost:3301, login = 'alpha', password = 'omega'}`
will be `alpha` and `omega` respectively. If the `login` field is set
and the `password` field is not set, the password is set to `nil`.
If the `password` field is set, the `login` field must be present.

Before this patch it could happen that deletion of a function from
the function cache didn't delete it from the funcs_by_name map. The
reason is that the check if the function exists in the map performs
the comparison of the search result with the `end` backet ID of the
wrong hash table.

The situation in which this could happen is the following:
1. Insertion of a new function into the cache triggers resize of the
   funcs_by_value map, but the size of the funcs map remains the same.
2. Then user deletes a function. This removes the function from the
   funcs map. Then we check if the function exists in the
   funcs_by_value map. The function exists there, but it so happens
   that it's bucket ID equals to the funcs map bucket count, so the
   incorrect check if the function exists in the funcs_by_value map
   states that the function does not exist there, so it's not dropped
   from the map.
3. Now we have the following result: the function is referenced in
   the funcs_by_value map, but not in funcs map. This triggers the
   assertion failure on any attempt to insert a new function with
   the same name.

Closes #9426

NO_DOC=bugfix

This patch adds support for automatic master discovery in vshard. There
is no longer a need to reapply the vshard storage configuration every
time an instance becomes a master or ceases to be a master.

Automatic master discovery also solves the problem with the rebalancer.
Previously, the rebalancer could not work correctly if the masters of
some replicasets were unknown, and since the vshard config generated by
the config module did not contain information about all the masters, the
rebalancer was disabled. The rebalancer can now perform master discovery
on its own, which is why it is enabled.

Before this patch, the config module was based on the
"test: use dofile() for configs instead of require" commit in vshard.
At a minimum, commit "storage: fix assertion error in conn_manager" is
now required.

Part of #8862

NO_DOC=will be added along with documentation for the rebalancer role
NO_CHANGELOG=will be added along with changelog for the rebalancer role

This module became unused as a result of LuaJIT bump made in the
commit 88333d13 ("luajit: bump new version"), so it can be
purged safely from the Tarantool sources.

Part of #8700

NO_DOC=internal
NO_TEST=internal
NO_CHANGELOG=added within the aforementioned commit

Current split-brain detector implementation raises an error each time a
CONFIRM or ROLLBACK entry is received from the previous synchronous
transaction queue owner. It is assumed that the new queue owner must
have witnessed all the previous CONFIRMS. Besides, according to Raft,
ROLLBACK should never happen.

Actually there is a case when a CONFIRM from an old term is legal: it's
possible that during leader transition old leader writes a CONFIRM for
the same transaction that is confirmed by the new leader's PROMOTE. If
PROMOTE and CONFIRM lsns match there is nothing bad about such
situation.

Symmetrically, when an old leader issues a ROLLBACK with the lsn right
after the new leader's PROMOTE lsn, it is not a split-brain.

Allow such cases by tracking the last confirmed lsn for each synchronous
transaction queue owner and silently nopifying CONFIRMs with an lsn less
than the one recorded and ROLLBACKs with lsn greater than that.

Closes #9138

NO_DOC=bugfix

Previously the replicas only persisted the confirmed lsn of the current
synchronous transaction queue owner. As soon as the onwer changed, the
info about which lsn was confirmed by the previous owner was lost.

Actually, this info is needed to correctly filter synchro requests
coming from the old term, so start tracking confirmed vclock instead of
the confirmed lsn on replicas.

In-scope of #9138

NO_TEST=covered by the next commit
NO_CHANGELOG=internal change

@TarantoolBot document
Title: Document new IPROTO_RAFT_PROMOTE request field

IPROTO_RAFT_PROMOTE and IPROTO_RAFT_DEMOTE requests receive a new key
value pair:

IPROTO_VCLOCK : MP_MAP

The vclock holds a confirmed vclock of the node sending the request.

Synchronous requests will receive a new field encoding a full vclock
soon. Theoretically a vclock may take up to ~ 300-400 bytes (3 bytes for
a map header + 32 components each taking up 1 byte for replica id and up
to 9 bytes for lsn). So it makes no sense to increase
SYNCHRO_BODY_LEN_MAX from 32 to 400-500. It would become almost the same
as plain BODY_LEN_MAX. Simply reuse the latter everywhere.

In-scope-of #9138

NO_DOC=refactoring
NO_TEST=refactoring
NO_CHANGELOG=refactoring

There was an error in xrow_decode_synchro: it compared the expected type
of the value to the type of the key (MP_UINT) instead of the type of the
actual value. This went unnoticed because all values in synchro requests
were integers.

This is going to change soon, when PROMOTE requests will start holding a
vclock, so fix the wrong type check.

In-scope-of #9138

NO_DOC=bugfix
NO_CHANGELOG=not user-visible

Without checking the return value of lua_pcall()` in
`lua_field_inspect_ucdata()`, the error message itself is returned as a
serialized result. The result status of `lua_pcall()` is not ignored
now.

NO_DOC=bugfix

Closes #9396

Netbox internally watches 'box.shutdown' for the sake of graceful
shutdown. The event subscription is async with connection API.
Additionally we check error count on server using different connection.
As a result we may or may not account error for the netbox internal
watch failure.

Let's account the internal watch failure reliably. Also while we at it
let's get rid of races for error count check.

Close #9423

NO_CHANGELOG=internal
NO_DOC=internal

If `config.etcd` is present and non-empty, `config.etcd.prefix` is
required.

This validation check was not performed due to a mistake in a schema
node wrapper that adds a validator that checks an attempt to use an
Enterprise Edition option on Community Edition.

Part of #8862

NO_DOC=bugfix

Looks like this is typo introduced in the commit 0704ebb7 ("xlog:
rework writer API").

Close #9428

NO_TEST=will be tested when fiber_cxx_invoke suppression will be removed
NO_CHANGELOG=introduced in 3.0.0-alpha3
NO_DOC=bugfix

Starting with commit f1c2127d ("replication: add META stage to JOIN")
replication master appends a special section, called IPROTO_JOIN_META to
the initial snapshot sent to the replica. This section contains the
latest raft term and synchronous transaction queue owner and term.

The section is only sent to nodes, which have a non-zero version_id.
For some reason, version_id encoding for FETCH_SNAPSHOT (analog of JOIN
for anonymous replicas) wasn't added in that commit, so anonymous
replicas do not receive synchronous queue state.

This leads to them raising ER_SPLIT_BRAIN errors later after join, when
the first synchronous row arrives.

In order to fix this, start encoding version_id in FETCH_SNAPSHOT
requests.

Closes #9401

@TarantoolBot document
Title: new field in `IPROTO_FETCH_SNAPSHOT` request

`IPROTO_FETCH_SNAPSHOT` request was bodyless (only contained a header)
until now, but now it receives a body with a single field:
`IPROTO_SERVER_VERSION` : MP_UINT -- an encoded representation of the
server version of a replica issuing the request.

Added a new is_sync parameter to `box.begin()`, `box.commit()`, and
`box.atomic()`. To make the transaction synchronous, set the `is_sync`
option to `true`. If any value other than `true/nil` is set, for example
`is_sync = "some string"`, then an error will be thrown.

Example:

```Lua
-- Sync transactions
box.atomic({is_sync = true}, function() ... end)

box.begin({is_sync = true}) ... box.commit({is_sync = true})

box.begin({is_sync = true}) ... box.commit()

box.begin() ... box.commit({is_sync = true})

-- Async transactions
box.atomic(function() ... end)

box.begin() ... box.commit()
```

Closes #8650

@TarantoolBot document
Title: box.atomic({is_sync = true})

Added the new `is_sync` parameter to `box.atomic()`. To make the
transaction synchronous, set the `is_sync` option to `true`. Setting
`is_sync = false` is prohibited. If to set any value other than true
for example `is_sync = "some string"`, then an error will be thrown.

This patch adds dependencies support for roles.

Part of #9078

@TarantoolBot document
Title: dependencies for roles

Roles can now have dependencies. This means that the verify() and
apply() methods will be executed for these roles, taking into account
the dependencies. Dependencies should be written in the "dependencies"
field of the array type. Note, the roles will be loaded (not applied!)
in the same order in which they were specified, i.e. not taking
dependencies into account.

Example:

Dependencies of role A: B, C
Dependencies of role B: D
No other role has dependencies.

Order in which roles were given: [E, C, A, B, D, G]
They will be loaded in the same order: [E, C, A, B, D, G]
The order, in which functions verify() and apply() will be executed:
[E, C, D, B, A, G].

Closes #9405

@TarantoolBot document
Title: Document then new built-in system event `box.wal_error`

The new event is broadcast whenever Tarantool fails to commit
a transaction to the write-ahead log (WAL), which usually means
there's a problem with the underlying disk storage. The new event's
payload is a table that currently contains the only field `count` that
stores the number of WAL errors happened so far or nil if there hasn't
been any WAL errors.

Test suite run can produce coredumps in case of bugs. Unfortunately
coredumps related to bugs are mixed with coredumps produced related to
special test conditions, like when we test Tarantool response to deadly
signal. Avoid producing coredumps in correct test suite run.

NO_CHANGELOG=internal
NO_DOC=internal

The fix is simple: look up the function in `box.func` by name and, if
found, execute its `call` method. The only tricky part is to avoid the
lookup before `box.cfg` is called because `box.func` is unavailable at
the time. We achieve that by checking `box.ctl.is_recovery_finished`.

Closes #9131

NO_DOC=bug fix

On Tarantool shutdown we destroy all the fibers in some sequence. We
don't require that all the fibers are finished before shutdown. So it
may turn out that we first destroy some alive fiber and then destroy
another alive fiber which joins the first one. Currently we have
use-after-free issue in this case because clearing `link` field of
the second fiber changes `wake` field of the first fiber.

Close #9406

NO_DOC=bugfix

Graceful shutdown is done in a special fiber which is started for
example on SIGTERM. So it can run concurrently with fiber executing
Tarantool init script. On init fiber exit we break event loop to pass
control back to the Tarantool initialization code. But we fail to run
event loop a bit more to finish graceful shutdown.

The test is a bit contrived. A more real world case is when Tarantool is
termintated during lingering box.cfg().

Close #9411

NO_DOC=bugfix

It was suggested by Igor Munkin (@igormunkin) in PR #9288.

Part of #8862
Follows up PR #9288

NO_DOC=the help message is not an API, nothing to document
NO_CHANGELOG=see NO_DOC
NO_TEST=see NO_DOC

According to ANSI, EXISTS is a predicate that tests a given subquery and
returns true if it returns more than 0 rows, false otherwise. However,
after 2a720d11, EXISTS worked correctly only if there were exactly 0
or 1 rows, and in all other cases it gave an error. This patch makes
EXITS work properly.

Closes #8676

NO_DOC=bugfix

Before this commit the space rollback had been treated as a new space
creation, so it caused creation of a new space object in the Lua's
box.space namespace. Since the preceding space drop removed the space
object from the namespace, on the space rollback all the Lua users of
the space loosed the track of its changes: the original space object
is never updated anymore.

This is fixed by detecting the space rollback and restoring the old
space object instead of creating a new one.

Closes #9120

NO_DOC=bugfix