|
Hi,
We've run into an issue with MariaDB when running Sysbench "oltp.lua" test with 8 threads. The server daemon crashed mostly with an assertion failure at storage/xtradb/fil/fil0fil.c:5288:
fil_node_complete_io(
|
/*=================*/
|
fil_node_t* node, /*!< in: file node */
|
fil_system_t* system, /*!< in: tablespace memory cache */
|
ulint type) /*!< in: OS_FILE_WRITE or OS_FILE_READ; marks
|
the node as modified if
|
type == OS_FILE_WRITE */
|
{
|
ut_ad(node);
|
ut_ad(system);
|
ut_ad(mutex_own(&(system->mutex)));
|
|
ut_a(node->n_pending > 0); <-- failure point
|
|
node->n_pending--;
|
An attached debugger gave the following backtrace:
(gdb) bt full
|
#0 0x0000007fb1d44d18 in __GI_raise (sig=sig@entry=6)
|
at ../nptl/sysdeps/unix/sysv/linux/raise.c:56
|
_sys_result = 0
|
pd = 0x7fa2fff1a0
|
pid = <optimised out>
|
selftid = 5661
|
#1 0x0000007fb1d4818c in __GI_abort () at abort.c:89
|
save_stage = 2
|
act = {__sigaction_handler = {sa_handler = 0x7f00000000,
|
sa_sigaction = 0x7f00000000}, sa_mask = {__val = {548445445976,
|
404, 1, 404, 0, 366924161824, 548195526816, 366921716804,
|
366927454208, 3, 0, 548434850424, 366927869208, 366936283408,
|
548195527344, 548434850424}}, sa_flags = 5288,
|
sa_restorer = 0xa2fff1a0}
|
sigs = {__val = {32, 0 <repeats 15 times>}}
|
#2 0x000000556e3d1448 in fil_node_complete_io (node=<optimised out>,
|
system=<optimised out>, type=<optimised out>)
|
at /build/buildd/mariadb-5.5-5.5.36/storage/xtradb/fil/fil0fil.c:5288
|
No locals.
|
#3 0x000000556e3db800 in fil_aio_wait (segment=segment@entry=3)
|
at /build/buildd/mariadb-5.5-5.5.36/storage/xtradb/fil/fil0fil.c:5705
|
ret = <optimised out>
|
fil_node = 0x7fb14a0e78
|
message = 0x7fa54d4350
|
type = 10
|
space_id = 0
|
#4 0x000000556e3592a4 in io_handler_thread (arg=<optimised out>)
|
at /build/buildd/mariadb-5.5-5.5.36/storage/xtradb/srv/srv0start.c:486
|
segment = 3
|
#5 0x0000007fb220ae2c in start_thread (arg=0x7fa2fff1a0)
|
at pthread_create.c:314
|
pd = 0x7fa2fff1a0
|
unwind_buf = {cancel_jmp_buf = {{jmp_buf = {548195529120,
|
548981639624, 548449456128, 0, 548449452032, 548195529312,
|
548195527344, 548443965168, 8388608, 548449472512,
|
548195527056, 13770210553321828185, 0, 13770210553602140361,
|
0, 0, 0, 0, 0, 0, 0, 0}, mask_was_saved = 0}}, priv = {pad = {
|
0x0, 0x0, 0x7fb220ad7c <start_thread>, 0x7fa2fff1a0}, data = {
|
prev = 0x0, cleanup = 0x0, canceltype = -1306481284}}}
|
not_first_call = 0
|
pagesize_m1 = <optimised out>
|
sp = <optimised out>
|
freesize = <optimised out>
|
__PRETTY_FUNCTION__ = "start_thread"
|
#6 0x0000007fb1dd9c40 in clone ()
|
at ../ports/sysdeps/unix/sysv/linux/aarch64/nptl/../clone.S:96
|
No locals.
|
Once the daemon crashed we've sometimes been unable to start it again without wiping out the database and re-installing it.
Having done some digging it is apparent that there is a problem in the mutex_exit code path; in particular at:
http://bazaar.launchpad.net/~maria-captains/maria/5.5/view/head:/storage/xtradb/include/sync0sync.ic#L106
A load-acquire is used to exit the mutex rather than a store-release. This leads to unpredictable results for architectures with a weak memory model.
We have the following in program order:
- mutex_enter -> load-acquire lock, loop until it is 0, then set to 1 relaxed
- protected work
- mutex_exit -> load-acquire lock, set it to 0 regardless.
However, the following sequence of events can be observed by another core:
- mutex_enter -> load-acquire lock, loop until it is 0, then set to 1 relaxed
- some of the protected work
- mutex_exit -> load-acquire lock, set it to 0 regardless.
- some more of the protected work (not protected).
The above can (and has for our test system) lead to severe data corruption; that prevents the daemon from even re-starting.
I've attached an emergency patch that re-introduces __ sync_lock_release to release the mutex. This fixes the crash and data corruption issues for me, but I understand from comments in the code that there were issues with this function in the past? Could the gcc intrinsics be moved over to the __ atomic_* functions? Ideally:
To acquire the lock:
__atomic_exchange_n(ptr, (byte) new_val, __ATOMIC_ACQUIRE)
|
To release the lock:
__atomic_store_n(ptr, (byte) new_val, __ATOMIC_RELEASE)
|
(which also worked on my test system).
I believe this issue may affect other versions of MariaDB, but I've only tested 5.5.36.
Cheers,
–
Steve Capper
|
store-release-on-mutex-exit.patch