OpenBSD includes DL library in a base system

Part of #4967

Part of #4967

Part of #4967

killpg() on Mac OS may don't deliver a signal to a process: it seems
that there is a race when a process is just forked. It means that
popen_handle:close() may leave a process alive, when `opts.setsid` and
`opts.group_signal` are set.

There is simple reproducer, which does not leave alive `sleep 120`
processes on Linux, but does it on Mac OS (within three-four runs in
rows):

 | #include <signal.h>
 | #include <unistd.h>
 | #include <fcntl.h>
 |
 | int
 | main()
 | {
 | 	char *child_argv[] = {
 | 		"/bin/sh",
 | 		"-c",
 | 		"sleep 120",
 | 		NULL,
 | 	};
 | 	pid_t pid;
 | 	int fd[2];
 | 	pipe(fd);
 | 	fcntl(fd[0], F_SETFD, FD_CLOEXEC);
 | 	fcntl(fd[1], F_SETFD, FD_CLOEXEC);
 |
 | 	if ((pid = fork()) == 0) {
 | 		/* Child. */
 | 		close(fd[0]);
 | 		setpgrp();
 | 		for (int i = 0; i < 10; ++i) {
 | 			/* Proceed with killpg. */
 | 			if (i == 5)
 | 				close(fd[1]);
 | 			if (fork() == 0) {
 | 				/* Child. */
 | 				execve("/bin/sh", child_argv, NULL);
 | 			}
 | 		}
 | 	} else {
 | 		/* Parent. */
 | 		close(fd[1]);
 | 		char c;
 | 		read(fd[0], &c, 1);
 | 		killpg(pid, SIGKILL);
 | 	}
 | 	return 0;
 | }

Compile it (`cc test.c -o test`) and run several times:

$ for i in $(seq 1 1000); do                     \
    echo $i;                                     \
    ./test                                       \
    && ps -o pid,pgid,command -ax | grep [s]leep \
    && break;                                    \
done

This is the reason why `sleep 120` process may be alive even when the
whole test passes.

test-run captures stdout and stderr of a 'core = app' test and waits EOF
on them. If a child process inherit one of them or both, the fd is still
open for writing and so EOF situation will not appear until `sleep 120`
will exit.

This commit doesn't try to overcome the root of the problem, but close
stdout and stderr for the child process that may not be killed / exited
in time.

Aside of this, updated found Mac OS peculiars in API comments of C and
Lua popen modules.

Fixes #4938

@TarantoolBot document
Title: popen: add note re group signaling on Mac OS

Copyed from the popen_handle:signal() updated description:

> Note: Mac OS may don't deliver a signal to a process in a group when
> opts.setsid and opts.group_signal are set. It seems there is a race
> here: when a process is just forked it may be not signaled.

Copyed from the popen_handle:close() updated description:

> Details about signaling:
>
> <...>
> - There are peculiars in group signaling on Mac OS,
>   @see popen_handle:signal() for details.

Follows up https://github.com/tarantool/doc/issues/1248

popen_delete() always frees a handle memory even when it reports a
failure to send SIGKILL, see [1]. We should reflect this contract in
popen_handle:close() and mark the handle as closed despite
popen_delete() return value.

There are cases, when killpg() fails with EPERM on Mac OS, so
popen_delete() reports a failure. See [1] for more information.

[1]: 01657bfb ('popen: always free
resources in popen_delete()')

Fixes #4995

Closes #4689

@TarantoolBot document
Title: box.session.push() 'sync' is deprecated

box.session.push() had two parameters - data to push and 'sync'.
The sync was a request ID with which the out of bound data should
be pushed into a socket.

This was introduced as a workaround for #3450, and is useless
since its resolution.

A user anyway can't push to different sessions, where that
parameter could be useful. And pushing into requests of the same
session, on the contrary, is something not really needed anywhere,
not portable to non-binary session types (console, background),
and is just dangerous since it is easy to add a bug here.

The patch removes the parameter. Now there will be thrown a
'Usage' error at attempt to use 'sync' parameter. In version 2.4
it is deprecated, prints warnings into logs, but still works. In
2.5 it is removed completely.

Revert 03790ac5 commit in small library part.

dynamic-list (exported_symbols_list on Mac) was used to forbid
export of all symbols of the tarantool executable except a given
list. Motivation of that was to avoid hacking the linker with
false usage of symbols needed to be exported. As a consequence,
symbols not listed in these options became invisible.

Before these options, when a symbol was defined, but not used in
the final executable, the linker could throw it away, even though
many symbols were used by Lua FFI, or should be visible for user's
dynamic modules. Where the linker, obviously, can't see if they
are needed.

To make the linker believe the symbols are actually needed there
was a hack with getting pointers at these functions and doing
something with them.

For example, assume we have 'test()' function in 'box' static
library:

    int
    test(void);

It is not used anywhere in the final executable. So to trick the
linker there is a function 'export_syms()' declared, which takes a
pointer at 'test()' and seemingly does something with it (or
actually does - it does not matter):

    void
    export_syms()
    {
        void *syms[] = {test};
        if (time(NULL) == 0) {
            syms[0]();
            syms[1]();
            ...
        }
    }

Some users want to use not documented but visible symbols, so the
patch removes the dynamic-list option, and returns the linker
hack back. But with 0 dependencies in the export file.

Closes #2971

src/cpu_feature.c, src/box/error.cc and src/box/vclock.c were used
twice. As a result it could lead to duplicate symbols.

To make sure ULL constants are not broken.

Part-of #4682

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

I do not drop the serpent module for a while
since I need to make sure that everything work
as expected as we've not break backward compatibility
significantly (though we didn't claim the lua mode
output is stable enough).

Fixes #4682

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

When we print output in console (especially in Lua mode)
it is impossible to find out the internal type an entry
represents, which in turn leads to inability to encode
"ULL" entries properly. Moreover we always introduce new
types (for example decimals, uuids) and the serpent module
we currently use for encodings has no clue about them.

Thus lets implement own lua serializer similarly to the
yaml encoder. This allows us to properly detect every
field and encode it accordingly.

Part-of #4682

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

As we gonna implement lua output serializer lets
rename luaL_yaml_default to serializer_yaml which
will be more general name, for other serializers
we will use same serializer_ prefix.

Part-of #4682

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

This will allow us to implement own formatter for
Lua output mode, so to be precise which exactly formatter
is caller lets rename general "format" to "format_yaml".

Part-of #4682

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

For some reason we're using spaces here to adjust code.

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

For better maintainance.

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

The member is defined as unsigned 32 bit value,
so we have to use an exactly the same type here.

Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>

When a new option is added, app-tap/init_script
outputs big diff. Because all options are printed with
ordinal indexes. Addition of a new option changes
indexes of all options after the new one.

The patch removes indexes from the output making diff
smaller, when a new option is added.

Release branches should be regularly run using gitlab-ci pipeline
schedules:
  https://gitlab.com/tarantool/tarantool/pipeline_schedules
It will help to detect flaky issues. But there is no need to rerun
too long running performance testing, to block it in schedules the
option 'schedules' in 'except:' field was set.

Part of #4974

Set all test suites at OSX testing.

Close #4818

Since the anonymous replica implementation, it was forbidden to
replicate (join/subscribe/register) from anonymous instances.
Actually, only joining and register should be banned, since an anonymous
replica isn't able to register its peer in _cluster anyway.

Let's allow other anonymous replicas, but not the normal ones, to subscribe to
an anonymous replica.
Also remove unnecessary ER_UNSUPPORTED errors from box_process_join()
and box_process_register() for anonymous replicas. These cases are
covered with ER_READONLY checks later on, since anonymous replicas must
be read-only.

Note, this patch doesn't allow normal instances to subscribe to
anonymous ones. Even though it is technically possible, it may bring
more problems than profits. Let's allow it later on if there's an explicit
demand.

Closes #4696

Closes #4900

@TarantoolBot document
Title: add new field to box.info: replication_anon

It is now possible to list all the anonymous replicas following the
instance with a call to `box.info.replication_anon()`
The output is similar to the one produced by `box.info.replication` with
an exception that anonymous replicas are indexed by their uuid strings
rather then server ids, since server ids have no meaning for anonymous
replicas.

Note, that when you issue a plain `box.info.replication_anon`, the only
info returned is the number of anonymous replicas following the current
instance. In order to see the full stats, you have to call
`box.info.replication_anon()`. This is done to not overload the `box.info`
output with excess info, since there may be lots of anonymous replicas.
Example:

```
tarantool> box.info.replication_anon
---
- count: 2
...

tarantool> box.info.replication_anon()
---
- 3a6a2cfb-7e47-42f6-8309-7a25c37feea1:
    id: 0
    uuid: 3a6a2cfb-7e47-42f6-8309-7a25c37feea1
    lsn: 0
    downstream:
      status: follow
      idle: 0.76203499999974
      vclock: {1: 1}
  f58e4cb0-e0a8-42a1-b439-591dd36c8e5e:
    id: 0
    uuid: f58e4cb0-e0a8-42a1-b439-591dd36c8e5e
    lsn: 0
    downstream:
      status: follow
      idle: 0.0041349999992235
      vclock: {1: 1}
...

```

Note, that anonymous replicas hide their lsn from the others, so
anonymous replica lsn will always be reported as zero, even if anonymous
replicas perform some local space operations.
To know the anonymous replica's lsn, you have to issue `box.info.lsn` on
it.

If recovery process fails during range restoration, range itself is
deleted and recovery is assumed to be finished as failed (in case of
casual i.e. not forced recovery). During recovery of particular range,
runs to be restored are refed twice: once when they are created at
vy_run_new() and once when they are attached to slice. This fact is
taken into consideration and after all ranges are recovered: all runs of
lsm tree are unrefed so that slices own run resources (as a result, when
slice is to be deleted its runs unrefed and deleted as well). However, if
range recovery fails, range is dropped alongside with already recovered
slices. This leads to unrefing runs - this is not accounted. To sum up
recovery process below is a brief schema:

foreach range in lsm.ranges {
  vy_lsm_recover_range(range) {
    foreach slice in range.slices {
      // inside recover_slice() each run is refed twice
      if vy_lsm_recover_slice() != 0 {
        // here all already restored slices are deleted and
        // corresponding runs are unrefed, so now they have 1 ref.
        range_delete()
      }
    }
  }
}
foreach run in lsm.runs {
  assert(run->refs > 1)
  vy_run_unref(run)
}

In this case, unrefing such runs one more time would lead to their
destruction. On the other hand, iteration over runs may turn out to
be unsafe, so we should use rlist_foreach_entry_safe(). Moreover, we
should explicitly clean-up these runs calling vy_lsm_remove_run().

Reviewed-by: Vladislav Shpilevoy <vshpilevoi@mail.ru>

Closes #4805

With new macro ERROR_INJECT_COUNTDOWN it is possible to delay error
injection by iparam value: injection will be set only after iparam
times the path is executed. For instance:

void
foo(int i)
{
	/* 2 is delay counter. */
	ERROR_INJECT_COUNTDOWN(ERRINJ_FOO, {
		 printf("Error injection on %d cycle!\n", i);
		});
}

void
boo(void)
{
	for (int i = 0; i < 10; ++i)
		foo(i);
}

box.error.injection.set('ERRINJ_FOO', 2)

The result is "Error injection on 2 cycle!". This type of error
injection can turn out to be useful to set injection in the middle of
query processing. Imagine following scenario:

void
foo(void)
{
	int *fds[10];
	for (int i = 0; i < 10; ++i) {
		fds[i] = malloc(sizeof(int));
		if (fds[i] == NULL)
			goto cleanup;
	}
cleanup:
	free(fds[0]);
}

"cleanup" section obviously contains error and leads to memory leak.
But using means of casual error injection without delay such situation
can't be detected: OOM can be set only for first cycle iteration and in
this particular case no leaks take place.

Reviewed-by: Vladislav Shpilevoy <vshpilevoi@mail.ru>

vy_task_write_run() is executed in auxiliary thread (dump or
compaction). Write iterator is created and used inside this function.
Meanwhile, creating/destroying tuples in these threads does not change
reference counter of corresponding  tuple formats (see vy_tuple_delete()
and vy_stmt_alloc()). Without cleaning up write iterator right in
write_iterator_start() after fail, this procedure takes place in
vy_task_compaction_abort() or vy_task_dump_abort().  These *_abort()
functions in turn are executed in the main thread.  Taking this into
consideration, tuple might be allocated in aux. thread and deleted in
the main thread. As a result, format reference counter might decrease,
whereas it shouldn't change (otherwise tuple format will be destroyed
before all tuples of this format are gone).

Fortunately, clean-up of write iterator in another thread was found only
on 1.10 branch, master branch already contains fix but lacks test
(2f17c929). So let's introduce test with following scenario:

1. run compaction process;
2. add one or more slice sources in vy_write_iterator_start():
corresponding slice_stream structures obtain newly created tuples
in vy_slice_stream_next();
3. the next call of vy_write_iterator_add_src() fails due to OOM,
invalid run file or whatever;
4. if write_iterator_start() didn't provide clean-up of sources, it
would take place in vy_task_dump_abort() which would be executed in
the main thread;
5. now format reference counter would be less than it was before
compaction.

Closes #4864

vy_write_iterator->read_views[i].history objects are allocated on
region (see vy_write_iterator_push_rv()) during building history of the
given key. However, in case of fail of vy_write_iterator_build_history()
region is truncated but pointers to vy_write_history objects are not
nullified. As a result, they may be accessed (for instance while
finalizing write_iterator object in  vy_write_iterator_stop) which in
turn may lead to crash, segfaul or disk formatting. The same may happen
if vy_read_view_merge() fails during processing of read view array.
Let's clean-up those objects in case of error takes place.

Part of #4864

According to dockerfile reference, there are two forms of specifying
entrypoint: exec and shell. Exec form is preferrable and  allows use
this image in scripts.

Close #4960

Added Catalina OSX 10.15 to gitlab-ci testing and removed OSX 10.13,
due to decided to have only 2 last major releases, for now it is
10.14 and 10.15 OSX versions. Also changed the commit job for branches
from 10.14 to 10.15 OSX version.

Additional cleanup for 'box_return_mp' and 'box_session_push',
added API_EXPORT which defines nothrow, compiler warns or errors
depending on the build options.

Part of #4885
Close #4873

Fragiled flaky tests from parallel runs to avoid
of flaky fails in regular testing:

  box-py/snapshot.test.py                ; gh-4514
  replication/misc.test.lua              ; gh-4940
  replication/skip_conflict_row.test.lua ; gh-4958
  replication-py/init_storage.test.py    ; gh-4949
  vinyl/stat.test.lua                    ; gh-4951
  xlog/checkpoint_daemon.test.lua        ; gh-4952

Part of #4953

Gitlab-ci pipeline modified in order to keep
performance results into gitlab-ci artifacts.

Closes #4920

Here is a summary on how and when rollback works in WAL.

Disk write failure can cause rollback. In that case the failed and
all next transactions, sent to WAL, should be rolled back.
Together. Following transactions should be rolled back too,
because they could make their statements based on what they saw in
the failed transaction. Also rollback of the failed transaction
without rollback of the next ones can actually rewrite what they
committed.

So when rollback is started, *all* pending transactions should be
rolled back. However if they would keep coming, the rollback would
be infinite. This means to complete a rollback it is necessary to
stop sending new transactions to WAL, then rollback all already
sent. In the end allow new transactions again.

Step-by-step:

1) stop accepting all new transactions in WAL thread, where
rollback is started. All new transactions don't even try to go to
disk. They added to rollback queue immediately after arriving to
WAL thread.

2) tell TX thread to stop sending new transactions to WAL. So as
the rollback queue would stop growing.

3) rollback all transactions in reverse order.

4) allow transactions again in WAL thread and TX thread.

The algorithm is long, but simple and understandable. However
implementation wasn't so easy. It was done using a 4-hop cbus
route. 2 hops of which were supposed to clear cbus channel from
all other cbus messages. Next two hops implemented steps 3 and 4.
Rollback state of the WAL was signaled by checking internals of a
preallocated cbus message.

The patch makes it simpler and more straightforward. Rollback
state is now signaled by a simple flag, and there is no a hack
about clearing cbus channel, no touching attributes of a cbus
message. The moment when all transactions are stopped and the last
one has returned from WAL is visible explicitly, because the last
sent to WAL journal entry is saved.

Also there is a single route for commit and rollback cbus
messages now, called tx_complete_batch(). This change will come
in hand in scope of synchronous replication, when WAL write won't
be enough for commit. And therefore 'commit' as a concept should
be washed away from WAL's code gradually. Migrate to solely txn
module.

Add check that on_shutdown() triggers were called before exit,
because in case of EOF or Ctrl+D (no signals) they were ignored.

Closes #4703

If vy_key_from_msgpack() fails in vy_lsm_split_range(), clean-up
procedure is called. However, at this moment struct vy_range *parts[2]
is not initialized ergo contains garbage and access to this structure
may result in crash, segfault or disk formatting. Let's move
initialization of mentioned variables before call of
vy_lsm_split_range().

Part of #4864

Found that in commit 'travis-ci/gitlab-ci: add Ubuntu Focal 20.04'
forgot to add Ubuntu Focal to the list of the available Ubuntu
distributions in the script for saving built packages at S3.

Follow up #4863

Closes #4863

API_EXPORT defines nothrow, so compiler warns or errors depending on the
build options.

Closes #4885

Fixed flaky upstream checks at replication/skip_conflict_row test,
also check on lsn set in test-run wait condition routine.

Errors fixed:

[024] @@ -66,11 +66,11 @@
[024]  ...
[024]  box.info.replication[1].upstream.message
[024]  ---
[024] -- null
[024] +- timed out
[024]  ...
[024]  box.info.replication[1].upstream.status
[024]  ---
[024] -- follow
[024] +- disconnected
[024]  ...
[024]  box.space.test:select()
[024]  ---
[024]

[004] @@ -125,11 +125,11 @@
[004]  ...
[004]  box.info.replication[1].upstream.message
[004]  ---
[004] -- Duplicate key exists in unique index 'primary' in space 'test'
[004] -...
[004] -box.info.replication[1].upstream.status
[004] ----
[004] -- stopped
[004] +- null
[004] +...
[004] +box.info.replication[1].upstream.status
[004] +---
[004] +- follow
[004]  ...
[004]  test_run:cmd("switch default")
[004]  ---
[004]

[038] @@ -174,7 +174,7 @@
[038]  ...
[038]  box.info.replication[1].upstream.status
[038]  ---
[038] -- follow
[038] +- disconnected
[038]  ...
[038]  -- write some conflicting records on slave
[038]  for i = 1, 10 do box.space.test:insert({i, 'r'}) end
Line 201 (often):

[039] @@ -201,7 +201,7 @@
[039]  -- lsn should be incremented
[039]  v1 == box.info.vclock[1] - 10
[039]  ---
[039] -- true
[039] +- false
[039]  ...
[039]  -- and state is follow
[039]  box.info.replication[1].upstream.status
[039]

[030] @@ -201,12 +201,12 @@
[030]  -- lsn should be incremented
[030]  v1 == box.info.vclock[1] - 10
[030]  ---
[030] -- true
[030] +- false
[030]  ...
[030]  -- and state is follow
[030]  box.info.replication[1].upstream.status
[030]  ---
[030] -- follow
[030] +- disconnected
[030]  ...
[030]  -- restart server and check replication continues from nop-ed vclock
[030]  test_run:cmd("switch default")
Line 230 (OSX):

[022] --- replication/skip_conflict_row.result	Thu Apr 16 21:54:28 2020
[022] +++ replication/skip_conflict_row.reject	Mon Apr 27 00:52:56 2020
[022] @@ -230,7 +230,7 @@
[022]  ...
[022]  box.info.replication[1].upstream.status
[022]  ---
[022] -- follow
[022] +- disconnected
[022]  ...
[022]  box.space.test:select({11}, {iterator = "GE"})
[022]  ---
[022]

Close #4457

Now we have S3 based infrastructure for RPM / Deb packages and GitLab CI
pipelines, which deploys packages to it.

We don't plan to add 2.5+ repositories on packagecloud.io, so instead of
usual change of target bucket from 2_N to 2_(N+1), the deploy stage is
removed.

Since all distro specific jobs are duplicated in GitLab CI pipelines and
those Travis-CI jobs are needed just for deployment, it worth to remove
them too.

Follows up #3380.
Part of #4947.

A couple of functions were mistakenly declared as 'function'
instead of 'local function' in schema.lua. That led to their
presence in the global namespace.

Closes #4812

When index:alter() was called on a non-functional index with
specified 'func', it led to accessing a not declared variable in
schema.lua.