jemalloc/test/unit/retained.c
Dmitry Ilvokhin 46690c9ec0 Fix test_retained on boxes with a lot of CPUs
We are trying to create `ncpus * 2` threads for this test and place them
into `VARIABLE_ARRAY`, but `VARIABLE_ARRAY` can not be more than
`VARIABLE_ARRAY_SIZE_MAX` bytes. When there are a lot of threads on the
box test always fails.

```
$ nproc
176

$ make -j`nproc` tests_unit && ./test/unit/retained
<jemalloc>: ../test/unit/retained.c:123: Failed assertion:
"sizeof(thd_t) * (nthreads) <= VARIABLE_ARRAY_SIZE_MAX"
Aborted (core dumped)
```

There is no need for high concurrency for this test as we are only
checking stats there and it's behaviour is quite stable regarding number
of allocating threads.

Limited number of threads to 16 to save compute resources (on CI for
example) and reduce tests running time.

Before the change (`nproc` is 80 on this box).

```
$ make -j`nproc` tests_unit && time ./test/unit/retained
<...>
real    0m0.372s
user    0m14.236s
sys     0m12.338s
```

After the change (same box).

```
$ make -j`nproc` tests_unit && time ./test/unit/retained
<...>
real    0m0.018s
user    0m0.108s
sys     0m0.068s
```
2024-12-02 14:12:26 -08:00

193 lines
5.1 KiB
C

#include "test/jemalloc_test.h"
#include "jemalloc/internal/san.h"
#include "jemalloc/internal/spin.h"
static unsigned arena_ind;
static size_t sz;
static size_t esz;
#define NEPOCHS 8
#define PER_THD_NALLOCS 1
static atomic_u_t epoch;
static atomic_u_t nfinished;
static unsigned
do_arena_create(extent_hooks_t *h) {
unsigned new_arena_ind;
size_t ind_sz = sizeof(unsigned);
expect_d_eq(mallctl("arenas.create", (void *)&new_arena_ind, &ind_sz,
(void *)(h != NULL ? &h : NULL), (h != NULL ? sizeof(h) : 0)), 0,
"Unexpected mallctl() failure");
return new_arena_ind;
}
static void
do_arena_destroy(unsigned ind) {
size_t mib[3];
size_t miblen;
miblen = sizeof(mib)/sizeof(size_t);
expect_d_eq(mallctlnametomib("arena.0.destroy", mib, &miblen), 0,
"Unexpected mallctlnametomib() failure");
mib[1] = (size_t)ind;
expect_d_eq(mallctlbymib(mib, miblen, NULL, NULL, NULL, 0), 0,
"Unexpected mallctlbymib() failure");
}
static void
do_refresh(void) {
uint64_t refresh_epoch = 1;
expect_d_eq(mallctl("epoch", NULL, NULL, (void *)&refresh_epoch,
sizeof(refresh_epoch)), 0, "Unexpected mallctl() failure");
}
static size_t
do_get_size_impl(const char *cmd, unsigned ind) {
size_t mib[4];
size_t miblen = sizeof(mib) / sizeof(size_t);
size_t z = sizeof(size_t);
expect_d_eq(mallctlnametomib(cmd, mib, &miblen),
0, "Unexpected mallctlnametomib(\"%s\", ...) failure", cmd);
mib[2] = ind;
size_t size;
expect_d_eq(mallctlbymib(mib, miblen, (void *)&size, &z, NULL, 0),
0, "Unexpected mallctlbymib([\"%s\"], ...) failure", cmd);
return size;
}
static size_t
do_get_active(unsigned ind) {
return do_get_size_impl("stats.arenas.0.pactive", ind) * PAGE;
}
static size_t
do_get_mapped(unsigned ind) {
return do_get_size_impl("stats.arenas.0.mapped", ind);
}
static void *
thd_start(void *arg) {
for (unsigned next_epoch = 1; next_epoch < NEPOCHS; next_epoch++) {
/* Busy-wait for next epoch. */
unsigned cur_epoch;
spin_t spinner = SPIN_INITIALIZER;
while ((cur_epoch = atomic_load_u(&epoch, ATOMIC_ACQUIRE)) !=
next_epoch) {
spin_adaptive(&spinner);
}
expect_u_eq(cur_epoch, next_epoch, "Unexpected epoch");
/*
* Allocate. The main thread will reset the arena, so there's
* no need to deallocate.
*/
for (unsigned i = 0; i < PER_THD_NALLOCS; i++) {
void *p = mallocx(sz, MALLOCX_ARENA(arena_ind) |
MALLOCX_TCACHE_NONE
);
expect_ptr_not_null(p,
"Unexpected mallocx() failure\n");
}
/* Let the main thread know we've finished this iteration. */
atomic_fetch_add_u(&nfinished, 1, ATOMIC_RELEASE);
}
return NULL;
}
TEST_BEGIN(test_retained) {
test_skip_if(!config_stats);
test_skip_if(opt_hpa);
arena_ind = do_arena_create(NULL);
sz = nallocx(HUGEPAGE, 0);
size_t guard_sz = san_guard_enabled() ? SAN_PAGE_GUARDS_SIZE : 0;
esz = sz + sz_large_pad + guard_sz;
atomic_store_u(&epoch, 0, ATOMIC_RELAXED);
unsigned nthreads = ncpus * 2;
if (nthreads > 16) {
/*
* Limit number of threads we are creating for following
* reasons.
* 1. On 32-bit platforms could run out of vaddr.
* 2. On boxes with a lot of CPUs we might have not enough
* memory to fit thd_t into VARIABLE_ARRAY.
*/
nthreads = 16;
}
VARIABLE_ARRAY(thd_t, threads, nthreads);
for (unsigned i = 0; i < nthreads; i++) {
thd_create(&threads[i], thd_start, NULL);
}
for (unsigned e = 1; e < NEPOCHS; e++) {
atomic_store_u(&nfinished, 0, ATOMIC_RELEASE);
atomic_store_u(&epoch, e, ATOMIC_RELEASE);
/* Wait for threads to finish allocating. */
spin_t spinner = SPIN_INITIALIZER;
while (atomic_load_u(&nfinished, ATOMIC_ACQUIRE) < nthreads) {
spin_adaptive(&spinner);
}
/*
* Assert that retained is no more than the sum of size classes
* that should have been used to satisfy the worker threads'
* requests, discounting per growth fragmentation.
*/
do_refresh();
size_t allocated = (esz - guard_sz) * nthreads *
PER_THD_NALLOCS;
size_t active = do_get_active(arena_ind);
expect_zu_le(allocated, active, "Unexpected active memory");
size_t mapped = do_get_mapped(arena_ind);
expect_zu_le(active, mapped, "Unexpected mapped memory");
arena_t *arena = arena_get(tsdn_fetch(), arena_ind, false);
size_t usable = 0;
for (pszind_t pind = sz_psz2ind(HUGEPAGE); pind <
arena->pa_shard.pac.exp_grow.next; pind++) {
size_t psz = sz_pind2sz(pind);
size_t psz_fragmented = psz % esz;
size_t psz_usable = psz - psz_fragmented;
/*
* Only consider size classes that wouldn't be skipped.
*/
if (psz_usable > 0) {
expect_zu_lt(usable, allocated,
"Excessive retained memory "
"(%#zx[+%#zx] > %#zx)", usable, psz_usable,
allocated);
usable += psz_usable;
}
}
/*
* Clean up arena. Destroying and recreating the arena
* is simpler that specifying extent hooks that deallocate
* (rather than retaining) during reset.
*/
do_arena_destroy(arena_ind);
expect_u_eq(do_arena_create(NULL), arena_ind,
"Unexpected arena index");
}
for (unsigned i = 0; i < nthreads; i++) {
thd_join(threads[i], NULL);
}
do_arena_destroy(arena_ind);
}
TEST_END
int
main(void) {
return test(
test_retained);
}