Description:
InnoDB atomics implementation miss load and store operation. To simulate those it uses heavyweight read-modify-write instructions.
Especially bad is replacement of atomic load by atomic read-modify-write. It may cause scalability bottlenecks where it's not necessary.
Following examples of code that can be replaced with atomic loads and stores...
dict/dict0dict.cc:
...
/* We use atomics to read index->zip_pad.pad. Here we use zero
as increment as are not changing the value of the 'pad'. */
pad = os_atomic_increment_ulint(&index->zip_pad.pad, 0);
...
The above is atomic load.
include/os0once.h:
...
const bool swapped = os_compare_and_swap_uint32(
state, IN_PROGRESS, DONE);
ut_a(swapped);
...
The above is atomic store.
include/sync0rw.ic:
...
(void) os_compare_and_swap_ulint(&lock->waiters, 0, 1);
...
(void) os_compare_and_swap_ulint(&lock->waiters, 1, 0);
...
The above are atomic stores.
...
/* Prevent Valgrind warnings about writer_thread being
uninitialized. It does not matter if writer_thread is
uninitialized, because we are comparing writer_thread against
itself, and the operation should always succeed. */
UNIV_MEM_VALID(&lock->writer_thread, sizeof lock->writer_thread);
local_thread = lock->writer_thread;
success = os_compare_and_swap_thread_id(
&lock->writer_thread, local_thread, curr_thread);
ut_a(success);
...
The above is my favorite perl. All lines can be replaced with atomic store.
lock/lock0lock.cc:
...
os_thread_id_t thread_id = os_thread_get_curr_id();
cas = os_compare_and_swap_thread_id(&trx->killed_by, 0, thread_id);
ut_a(cas);
...
The above is atomic store.
trx/trx0purge.cc:
...
while (!os_compare_and_swap_ulint(
&purge_sys->n_completed, n_submitted, n_submitted)) {
...
The above is atomic load.
trx/trx0trx.cc:
...
os_thread_id_t thread_id = trx->killed_by;
os_compare_and_swap_thread_id(&trx->killed_by, thread_id, 0);
...
os_thread_id_t thread_id = victim_trx->killed_by;
os_compare_and_swap_thread_id(&victim_trx->killed_by,
thread_id, 0);
...
The above are atomic stores.
How to repeat:
Code analysis.
Suggested fix:
InnoDB badly needs sane atomics implementation. Follow C11 or gcc atomic builtins API.
Description: InnoDB atomics implementation miss load and store operation. To simulate those it uses heavyweight read-modify-write instructions. Especially bad is replacement of atomic load by atomic read-modify-write. It may cause scalability bottlenecks where it's not necessary. Following examples of code that can be replaced with atomic loads and stores... dict/dict0dict.cc: ... /* We use atomics to read index->zip_pad.pad. Here we use zero as increment as are not changing the value of the 'pad'. */ pad = os_atomic_increment_ulint(&index->zip_pad.pad, 0); ... The above is atomic load. include/os0once.h: ... const bool swapped = os_compare_and_swap_uint32( state, IN_PROGRESS, DONE); ut_a(swapped); ... The above is atomic store. include/sync0rw.ic: ... (void) os_compare_and_swap_ulint(&lock->waiters, 0, 1); ... (void) os_compare_and_swap_ulint(&lock->waiters, 1, 0); ... The above are atomic stores. ... /* Prevent Valgrind warnings about writer_thread being uninitialized. It does not matter if writer_thread is uninitialized, because we are comparing writer_thread against itself, and the operation should always succeed. */ UNIV_MEM_VALID(&lock->writer_thread, sizeof lock->writer_thread); local_thread = lock->writer_thread; success = os_compare_and_swap_thread_id( &lock->writer_thread, local_thread, curr_thread); ut_a(success); ... The above is my favorite perl. All lines can be replaced with atomic store. lock/lock0lock.cc: ... os_thread_id_t thread_id = os_thread_get_curr_id(); cas = os_compare_and_swap_thread_id(&trx->killed_by, 0, thread_id); ut_a(cas); ... The above is atomic store. trx/trx0purge.cc: ... while (!os_compare_and_swap_ulint( &purge_sys->n_completed, n_submitted, n_submitted)) { ... The above is atomic load. trx/trx0trx.cc: ... os_thread_id_t thread_id = trx->killed_by; os_compare_and_swap_thread_id(&trx->killed_by, thread_id, 0); ... os_thread_id_t thread_id = victim_trx->killed_by; os_compare_and_swap_thread_id(&victim_trx->killed_by, thread_id, 0); ... The above are atomic stores. How to repeat: Code analysis. Suggested fix: InnoDB badly needs sane atomics implementation. Follow C11 or gcc atomic builtins API.