project-base/server-skynet-source-3rd-jemalloc
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fast path: allow low-water mark changes.

Browse Source 
This lets us put more allocations on an "almost as fast" path after a flush.
This results in around a 4% reduction in malloc cycles in prod workloads
(corresponding to about a 0.1% reduction in overall cycles).
This commit is contained in:
David Goldblatt 2020-03-04 08:58:42 -08:00 committed by David Goldblatt
parent 397da03865
commit d701a085c2
5 changed files with 86 additions and 72 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

93
include/jemalloc/internal/cache_bin.h
View File

@ -46,6 +46,12 @@ struct cache_bin_s {
* the head points to one element past the owned array.
*/
void **stack_head;
/*
* cur_ptr and stats are both modified frequently. Let's keep them
* close so that they have a higher chance of being on the same
* cacheline, thus less write-backs.
*/
cache_bin_stats_t tstats;
/*
* The low bits of the address of the first item in the stack that
@ -76,12 +82,6 @@ struct cache_bin_s {
*/
uint16_t low_bits_empty;
/*
* cur_ptr and stats are both modified frequently. Let's keep them
* close so that they have a higher chance of being on the same
* cacheline, thus less write-backs.
*/
cache_bin_stats_t tstats;
};
typedef struct cache_bin_array_descriptor_s cache_bin_array_descriptor_t;
@ -201,36 +201,7 @@ cache_bin_array_descriptor_init(cache_bin_array_descriptor_t *descriptor,
}
JEMALLOC_ALWAYS_INLINE void *
cache_bin_alloc_easy_impl(cache_bin_t *bin, bool *success,
const bool adjust_low_water) {
/*
* This may read from the empty position; however the loaded value won't
* be used. It's safe because the stack has one more slot reserved.
*/
void *ret = *bin->stack_head;
uint16_t low_bits = (uint16_t)(uintptr_t)bin->stack_head;
void **new_head = bin->stack_head + 1;
/*
* Note that the low water mark is at most empty; if we pass this check,
* we know we're non-empty.
*/
if (unlikely(low_bits == bin->low_bits_low_water)) {
if (adjust_low_water) {
if (unlikely(low_bits == bin->low_bits_empty)) {
*success = false;
return NULL;
}
/* Overflow should be impossible. */
assert(bin->low_bits_low_water
< (uint16_t)(uintptr_t)new_head);
bin->low_bits_low_water = (uint16_t)(uintptr_t)new_head;
} else {
*success = false;
return NULL;
}
}
bin->stack_head = new_head;
cache_bin_alloc_impl(cache_bin_t *bin, bool *success, bool adjust_low_water) {
/*
* success (instead of ret) should be checked upon the return of this
* function. We avoid checking (ret == NULL) because there is never a
@ -238,22 +209,52 @@ cache_bin_alloc_easy_impl(cache_bin_t *bin, bool *success,
* and eagerly checking ret would cause pipeline stall (waiting for the
* cacheline).
*/
*success = true;
/*
* This may read from the empty position; however the loaded value won't
* be used. It's safe because the stack has one more slot reserved.
*/
void *ret = *bin->stack_head;
uint16_t low_bits = (uint16_t)(uintptr_t)bin->stack_head;
void **new_head = bin->stack_head + 1;
/*
* Note that the low water mark is at most empty; if we pass this check,
* we know we're non-empty.
*/
if (likely(low_bits != bin->low_bits_low_water)) {
bin->stack_head = new_head;
*success = true;
return ret;
}
JEMALLOC_ALWAYS_INLINE void *
cache_bin_alloc_easy_reduced(cache_bin_t *bin, bool *success) {
/* We don't look at info if we're not adjusting low-water. */
return cache_bin_alloc_easy_impl(bin, success, false);
if (!adjust_low_water) {
*success = false;
return NULL;
}
/*
* In the fast-path case where we call alloc_easy and then alloc, the
* previous checking and computation is optimized away -- we didn't
* actually commit any of our operations.
*/
if (likely(low_bits != bin->low_bits_empty)) {
bin->stack_head = new_head;
bin->low_bits_low_water = (uint16_t)(uintptr_t)new_head;
*success = true;
return ret;
}
*success = false;
return NULL;
}
JEMALLOC_ALWAYS_INLINE void *
cache_bin_alloc_easy(cache_bin_t *bin, cache_bin_info_t *info, bool *success) {
/* We don't use info now, but we may want to in the future. */
(void)info;
return cache_bin_alloc_easy_impl(bin, success, true);
cache_bin_alloc_easy(cache_bin_t *bin, bool *success) {
/* We don't look at info if we're not adjusting low-water. */
return cache_bin_alloc_impl(bin, success, false);
}
JEMALLOC_ALWAYS_INLINE void *
cache_bin_alloc(cache_bin_t *bin, bool *success) {
return cache_bin_alloc_impl(bin, success, true);
}
JEMALLOC_ALWAYS_INLINE bool

6
include/jemalloc/internal/tcache_inlines.h
View File

@ -36,8 +36,7 @@ tcache_alloc_small(tsd_t *tsd, arena_t *arena, tcache_t *tcache,
assert(binind < SC_NBINS);
bin = tcache_small_bin_get(tcache, binind);
ret = cache_bin_alloc_easy(bin, &tcache_bin_info[binind],
&tcache_success);
ret = cache_bin_alloc(bin, &tcache_success);
assert(tcache_success == (ret != NULL));
if (unlikely(!tcache_success)) {
bool tcache_hard_success;
@ -80,8 +79,7 @@ tcache_alloc_large(tsd_t *tsd, arena_t *arena, tcache_t *tcache, size_t size,
assert(binind >= SC_NBINS &&binind < nhbins);
bin = tcache_large_bin_get(tcache, binind);
ret = cache_bin_alloc_easy(bin, &tcache_bin_info[binind],
&tcache_success);
ret = cache_bin_alloc(bin, &tcache_success);
assert(tcache_success == (ret != NULL));
if (unlikely(!tcache_success)) {
/*

32
src/jemalloc.c
View File

@ -2377,6 +2377,17 @@ malloc_default(size_t size) {
* Begin malloc(3)-compatible functions.
*/
JEMALLOC_ALWAYS_INLINE void
fastpath_success_finish(tsd_t *tsd, uint64_t allocated_after,
cache_bin_t *bin, void *ret) {
thread_allocated_set(tsd, allocated_after);
if (config_stats) {
bin->tstats.nrequests++;
}
LOG("core.malloc.exit", "result: %p", ret);
}
/*
* malloc() fastpath.
*
@ -2451,17 +2462,22 @@ je_malloc(size_t size) {
tcache_t *tcache = tsd_tcachep_get(tsd);
cache_bin_t *bin = tcache_small_bin_get(tcache, ind);
bool tcache_success;
void *ret = cache_bin_alloc_easy_reduced(bin, &tcache_success);
void *ret;
/*
* We split up the code this way so that redundant low-water
* computation doesn't happen on the (more common) case in which we
* don't touch the low water mark. The compiler won't do this
* duplication on its own.
*/
ret = cache_bin_alloc_easy(bin, &tcache_success);
if (tcache_success) {
thread_allocated_set(tsd, allocated_after);
if (config_stats) {
bin->tstats.nrequests++;
fastpath_success_finish(tsd, allocated_after, bin, ret);
return ret;
}
LOG("core.malloc.exit", "result: %p", ret);
/* Fastpath success */
ret = cache_bin_alloc(bin, &tcache_success);
if (tcache_success) {
fastpath_success_finish(tsd, allocated_after, bin, ret);
return ret;
}

3
src/tcache.c
View File

@ -104,8 +104,7 @@ tcache_alloc_small_hard(tsdn_t *tsdn, arena_t *arena, tcache_t *tcache,
assert(tcache->arena != NULL);
arena_tcache_fill_small(tsdn, arena, tcache, tbin, binind);
ret = cache_bin_alloc_easy(tbin, &tcache_bin_info[binind],
tcache_success);
ret = cache_bin_alloc(tbin, tcache_success);
return ret;
}

26
test/unit/cache_bin.c
View File

@ -17,7 +17,7 @@ do_fill_test(cache_bin_t *bin, cache_bin_info_t *info, void **ptrs,
cache_bin_low_water_set(bin);
for (cache_bin_sz_t i = 0; i < nfill_succeed; i++) {
ptr = cache_bin_alloc_easy(bin, info, &success);
ptr = cache_bin_alloc(bin, &success);
expect_true(success, "");
expect_ptr_eq(ptr, (void *)&ptrs[i],
"Should pop in order filled");
@ -48,7 +48,7 @@ do_flush_test(cache_bin_t *bin, cache_bin_info_t *info, void **ptrs,
expect_true(cache_bin_ncached_get(bin, info) == nfill - nflush, "");
while (cache_bin_ncached_get(bin, info) > 0) {
cache_bin_alloc_easy(bin, info, &success);
cache_bin_alloc(bin, &success);
}
}
@ -78,11 +78,11 @@ TEST_BEGIN(test_cache_bin) {
expect_true(cache_bin_ncached_get(&bin, &info) == 0, "");
expect_true(cache_bin_low_water_get(&bin, &info) == 0, "");
ptr = cache_bin_alloc_easy_reduced(&bin, &success);
ptr = cache_bin_alloc_easy(&bin, &success);
expect_false(success, "Shouldn't successfully allocate when empty");
expect_ptr_null(ptr, "Shouldn't get a non-null pointer on failure");
ptr = cache_bin_alloc_easy(&bin, &info, &success);
ptr = cache_bin_alloc(&bin, &success);
expect_false(success, "Shouldn't successfully allocate when empty");
expect_ptr_null(ptr, "Shouldn't get a non-null pointer on failure");
@ -112,10 +112,10 @@ TEST_BEGIN(test_cache_bin) {
expect_true(cache_bin_ncached_get(&bin, &info)
== ncached_max - i, "");
/*
* This should fail -- the reduced version can't change low
* water.
* This should fail -- the easy variant can't change the low
* water mark.
*/
ptr = cache_bin_alloc_easy_reduced(&bin, &success);
ptr = cache_bin_alloc_easy(&bin, &success);
expect_ptr_null(ptr, "");
expect_false(success, "");
expect_true(cache_bin_low_water_get(&bin, &info)
@ -124,7 +124,7 @@ TEST_BEGIN(test_cache_bin) {
== ncached_max - i, "");
/* This should succeed, though. */
ptr = cache_bin_alloc_easy(&bin, &info, &success);
ptr = cache_bin_alloc(&bin, &success);
expect_true(success, "");
expect_ptr_eq(ptr, &ptrs[ncached_max - i - 1],
"Alloc should pop in stack order");
@ -135,10 +135,10 @@ TEST_BEGIN(test_cache_bin) {
}
/* Now we're empty -- all alloc attempts should fail. */
expect_true(cache_bin_ncached_get(&bin, &info) == 0, "");
ptr = cache_bin_alloc_easy_reduced(&bin, &success);
ptr = cache_bin_alloc_easy(&bin, &success);
expect_ptr_null(ptr, "");
expect_false(success, "");
ptr = cache_bin_alloc_easy(&bin, &info, &success);
ptr = cache_bin_alloc(&bin, &success);
expect_ptr_null(ptr, "");
expect_false(success, "");
@ -156,18 +156,18 @@ TEST_BEGIN(test_cache_bin) {
* Size is bigger than low water -- the reduced version should
* succeed.
*/
ptr = cache_bin_alloc_easy_reduced(&bin, &success);
ptr = cache_bin_alloc_easy(&bin, &success);
expect_true(success, "");
expect_ptr_eq(ptr, &ptrs[i], "");
}
/* But now, we've hit low-water. */
ptr = cache_bin_alloc_easy_reduced(&bin, &success);
ptr = cache_bin_alloc_easy(&bin, &success);
expect_false(success, "");
expect_ptr_null(ptr, "");
/* We're going to test filling -- we must be empty to start. */
while (cache_bin_ncached_get(&bin, &info)) {
cache_bin_alloc_easy(&bin, &info, &success);
cache_bin_alloc(&bin, &success);
expect_true(success, "");
}