test: log_destroy() sometimes segfaults on app-tap/logger.test.lua

[Totktonada]

Tarantool version: 2.3.0-61-g5b9f207df OS version: Gentoo Linux.

How to reproduce:

$ cmake . -DCMAKE_BUILD_TYPE=Debug -DENABLE_BACKTRACE=ON -DENABLE_DIST=ON && make -j
$ TEST_RUN_TESTS="$(yes app-tap/logger.test.lua | head -n 1000)" make test # several times

Backtrace:

(gdb) bt
#0  0x00007fe9392d9f1b in raise () from /lib64/libc.so.6
#1  0x00007fe9392c3535 in abort () from /lib64/libc.so.6
#2  0x0000560611c11ef7 in sig_fatal_cb (signo=11, siginfo=0x7fff252656f0, context=0x7fff252655c0) at /home/alex/p/tarantool-meta/r/t-2/src/main.cc:300
#3  <signal handler called>
#4  0x0000560611d52d13 in fiber_yield () at /home/alex/p/tarantool-meta/r/t-2/src/lib/core/fiber.c:478
#5  0x0000560611d52ed1 in fiber_yield_timeout (delay=3153600000) at /home/alex/p/tarantool-meta/r/t-2/src/lib/core/fiber.c:522
#6  0x0000560611d5b286 in fiber_cond_wait_timeout (c=0x5606122037a0 <log_std+96>, timeout=3153600000) at /home/alex/p/tarantool-meta/r/t-2/src/lib/core/fiber_cond.c:107
#7  0x0000560611d5b354 in fiber_cond_wait (c=0x5606122037a0 <log_std+96>) at /home/alex/p/tarantool-meta/r/t-2/src/lib/core/fiber_cond.c:117
#8  0x0000560611d50af4 in log_destroy (log=0x560612203740 <log_std>) at /home/alex/p/tarantool-meta/r/t-2/src/lib/core/say.c:1178
#9  0x0000560611d4e9d5 in say_logger_free () at /home/alex/p/tarantool-meta/r/t-2/src/lib/core/say.c:714
#10 0x0000560611c12fda in tarantool_free () at /home/alex/p/tarantool-meta/r/t-2/src/main.cc:639
#11 0x0000560611c138cd in main (argc=1, argv=0x560612859f18) at /home/alex/p/tarantool-meta/r/t-2/src/main.cc:864

[ImeevMA]

Way to reproduce:

printf "log = require('log')\nbox.cfg{log = 'temp.log'}\nlog.rotate()\nos.exit()" > repr.lua
./tarantool repr.lua

Result:

Segmentation fault
  code: SEGV_MAPERR
  addr: 0x68
  context: 0x7ffe639ac9c0
  siginfo: 0x7ffe639acaf0
  rax      0x0                0
  rbx      0x0                0
  rcx      0x56024afee480     94567848141952
  rdx      0x47               71
  rsi      0xe                14
  rdi      0x5602490fb2f8     94567815688952
  rsp      0x7ffe639ad070     140730569511024
  rbp      0x7ffe639ad090     140730569511056
  r8       0x7                7
  r9       0x1                1
  r10      0x8                8
  r11      0x202              514
  r12      0x560200000000     94566589923328
  r13      0x0                0
  r14      0x0                0
  r15      0x0                0
  rip      0x560248c0208e     94567810474126
  eflags   0x10202            66050
  cs       0x33               51
  gs       0x0                0
  fs       0x0                0
  cr2      0x68               104
  err      0x4                4
  oldmask  0x0                0
  trapno   0xe                14
Current time: 1591793154
Please file a bug at http://github.com/tarantool/tarantool/issues
Attempting backtrace... Note: since the server has already crashed, 
this may fail as well
#0  0x560248c065a1 in print_backtrace+d
#1  0x560248aaaf82 in _ZL12sig_fatal_cbiP9siginfo_tPv+1eb
#2  0x7fe1817b33c0 in funlockfile+60
#3  0x560248c0208e in fiber_yield+80
#4  0x560248c02243 in fiber_yield_timeout+9e
#5  0x560248c0b018 in fiber_cond_wait_timeout+4f
#6  0x560248c0b0ea in fiber_cond_wait+28
#7  0x560248bff5d9 in log_destroy+46
#8  0x560248bfd47e in say_logger_free+1d
#9  0x560248aac120 in _Z14tarantool_freev+105
#10 0x560248aaca50 in main+735
#11 0x7fe180fa20b3 in __libc_start_main+f3
#12 0x560248aaa5fe in _start+2e
Aborted (core dumped)

[Gerold103]

The problem is that ev_break() is called before modules start destruction. Regardless of how you terminate tarantool - using os.exit(), or the script ended. tarantool_exit() is always called before tarantool_free(), and so ev_break(). It means, inside tarantool_free() it is not ok to use any scheduling related functionality - it relies on the loop being active. Don't have a solution for it yet.

[locker]

Reopened by #7643.