keep commit_mask live in the cache for better reuse

This commit is contained in:
daan 2020-09-05 21:58:32 -07:00
parent 63a9f45ba6
commit f7dc4847f2
4 changed files with 126 additions and 55 deletions

View file

@ -135,8 +135,8 @@ static bool mi_arena_alloc(mi_arena_t* arena, size_t blocks, mi_bitmap_index_t*
typedef struct mi_cache_slot_s {
void* p;
size_t memid;
mi_commit_mask_t commit_mask;
_Atomic(mi_msecs_t) expire;
bool is_committed; // TODO: use bit from p to reduce size?
} mi_cache_slot_t;
static mi_cache_slot_t cache[MI_CACHE_MAX]; // = 0
@ -147,7 +147,10 @@ static mi_bitmap_field_t cache_available_large[MI_CACHE_FIELDS] = { MI_CACHE_BIT
static mi_bitmap_field_t cache_inuse[MI_CACHE_FIELDS]; // zero bit = free
static void* mi_cache_pop(int numa_node, size_t size, size_t alignment, bool* commit, bool* large, bool* is_zero, size_t* memid, mi_os_tld_t* tld) {
static void* mi_cache_pop(int numa_node, size_t size, size_t alignment, bool commit, mi_commit_mask_t* commit_mask, bool* large, bool* is_zero, size_t* memid, mi_os_tld_t* tld) {
UNUSED(tld);
UNUSED(commit);
// only segment blocks
if (size != MI_SEGMENT_SIZE || alignment > MI_SEGMENT_ALIGN) return NULL;
@ -177,24 +180,55 @@ static void* mi_cache_pop(int numa_node, size_t size, size_t alignment, bool* co
void* p = slot->p;
*memid = slot->memid;
*is_zero = false;
bool committed = slot->is_committed;
mi_commit_mask_t cmask = slot->commit_mask; // copy
slot->p = NULL;
mi_atomic_storei64_release(&slot->expire,(mi_msecs_t)0);
if (*commit && !committed) {
// ignore commit request
/*
if (commit && !mi_commit_mask_is_full(cmask)) {
bool commit_zero;
_mi_os_commit(p, MI_SEGMENT_SIZE, &commit_zero, tld->stats);
*commit = true;
bool ok = _mi_os_commit(p, MI_SEGMENT_SIZE, &commit_zero, tld->stats); // todo: only commit needed parts?
if (!ok) {
*commit_mask = cmask;
}
else {
*commit_mask = mi_commit_mask_full();
}
}
else {
*commit = committed;
}
*/
*commit_mask = cmask;
// mark the slot as free again
mi_assert_internal(mi_bitmap_is_claimed(cache_inuse, MI_CACHE_FIELDS, 1, bitidx));
mi_bitmap_unclaim(cache_inuse, MI_CACHE_FIELDS, 1, bitidx);
return p;
}
static void mi_commit_mask_decommit(mi_commit_mask_t* cmask, void* p, size_t total, mi_stats_t* stats) {
if (mi_commit_mask_is_empty(*cmask)) {
// nothing
}
else if (mi_commit_mask_is_full(*cmask)) {
_mi_os_decommit(p, total, stats);
}
else {
// todo: one call to decommit the whole at once?
mi_assert_internal((total%MI_COMMIT_MASK_BITS)==0);
size_t part = total/MI_COMMIT_MASK_BITS;
uintptr_t idx;
uintptr_t count;
mi_commit_mask_t mask = *cmask;
mi_commit_mask_foreach(mask, idx, count) {
void* start = (uint8_t*)p + (idx*part);
size_t size = count*part;
_mi_os_decommit(start, size, stats);
}
mi_commit_mask_foreach_end()
}
*cmask = mi_commit_mask_empty();
}
static void mi_cache_purge(mi_os_tld_t* tld) {
UNUSED(tld);
mi_msecs_t now = _mi_clock_now();
@ -214,10 +248,11 @@ static void mi_cache_purge(mi_os_tld_t* tld) {
if (expire != 0 && now >= expire) { // safe read
// still expired, decommit it
mi_atomic_storei64_relaxed(&slot->expire,(mi_msecs_t)0);
mi_assert_internal(slot->is_committed && mi_bitmap_is_claimed(cache_available_large, MI_CACHE_FIELDS, 1, bitidx));
mi_assert_internal(!mi_commit_mask_is_empty(slot->commit_mask) && mi_bitmap_is_claimed(cache_available_large, MI_CACHE_FIELDS, 1, bitidx));
_mi_abandoned_await_readers(); // wait until safe to decommit
_mi_os_decommit(slot->p, MI_SEGMENT_SIZE, tld->stats);
slot->is_committed = false;
// decommit committed parts
mi_commit_mask_decommit(&slot->commit_mask, slot->p, MI_SEGMENT_SIZE, tld->stats);
//_mi_os_decommit(slot->p, MI_SEGMENT_SIZE, tld->stats);
}
mi_bitmap_unclaim(cache_available, MI_CACHE_FIELDS, 1, bitidx); // make it available again for a pop
}
@ -226,7 +261,7 @@ static void mi_cache_purge(mi_os_tld_t* tld) {
}
}
static bool mi_cache_push(void* start, size_t size, size_t memid, bool is_committed, bool is_large, mi_os_tld_t* tld)
static bool mi_cache_push(void* start, size_t size, size_t memid, mi_commit_mask_t commit_mask, bool is_large, mi_os_tld_t* tld)
{
// only for segment blocks
if (size != MI_SEGMENT_SIZE || ((uintptr_t)start % MI_SEGMENT_ALIGN) != 0) return false;
@ -255,13 +290,12 @@ static bool mi_cache_push(void* start, size_t size, size_t memid, bool is_commit
slot->p = start;
slot->memid = memid;
mi_atomic_storei64_relaxed(&slot->expire,(mi_msecs_t)0);
slot->is_committed = is_committed;
if (is_committed && !is_large) {
slot->commit_mask = commit_mask;
if (!mi_commit_mask_is_empty(commit_mask) && !is_large) {
long delay = mi_option_get(mi_option_arena_reset_delay);
if (delay == 0) {
_mi_abandoned_await_readers(); // wait until safe to decommit
_mi_os_decommit(start, size, tld->stats);
slot->is_committed = false;
mi_commit_mask_decommit(&slot->commit_mask, start, MI_SEGMENT_SIZE, tld->stats);
}
else {
mi_atomic_storei64_release(&slot->expire, _mi_clock_now() + delay);
@ -311,10 +345,10 @@ static void* mi_arena_alloc_from(mi_arena_t* arena, size_t arena_index, size_t n
}
void* _mi_arena_alloc_aligned(size_t size, size_t alignment,
bool* commit, bool* large, bool* is_zero,
bool commit, mi_commit_mask_t* commit_mask, bool* large, bool* is_zero,
size_t* memid, mi_os_tld_t* tld)
{
mi_assert_internal(commit != NULL && large != NULL && is_zero != NULL && memid != NULL && tld != NULL);
mi_assert_internal(commit_mask != NULL && large != NULL && is_zero != NULL && memid != NULL && tld != NULL);
mi_assert_internal(size > 0);
*memid = MI_MEMID_OS;
*is_zero = false;
@ -339,9 +373,13 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment,
if ((arena->numa_node<0 || arena->numa_node==numa_node) && // numa local?
(*large || !arena->is_large)) // large OS pages allowed, or arena is not large OS pages
{
void* p = mi_arena_alloc_from(arena, i, bcount, commit, large, is_zero, memid, tld);
bool acommit = commit;
void* p = mi_arena_alloc_from(arena, i, bcount, &acommit, large, is_zero, memid, tld);
mi_assert_internal((uintptr_t)p % alignment == 0);
if (p != NULL) return p;
if (p != NULL) {
*commit_mask = (acommit ? mi_commit_mask_full() : mi_commit_mask_empty());
return p;
}
}
}
// try from another numa node instead..
@ -351,43 +389,52 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment,
if ((arena->numa_node>=0 && arena->numa_node!=numa_node) && // not numa local!
(*large || !arena->is_large)) // large OS pages allowed, or arena is not large OS pages
{
void* p = mi_arena_alloc_from(arena, i, bcount, commit, large, is_zero, memid, tld);
bool acommit = commit;
void* p = mi_arena_alloc_from(arena, i, bcount, &acommit, large, is_zero, memid, tld);
mi_assert_internal((uintptr_t)p % alignment == 0);
if (p != NULL) return p;
if (p != NULL) {
*commit_mask = (acommit ? mi_commit_mask_full() : mi_commit_mask_empty());
return p;
}
}
}
}
// try to get from the cache
void* p = mi_cache_pop(numa_node, size, alignment, commit, large, is_zero, memid, tld);
void* p = mi_cache_pop(numa_node, size, alignment, commit, commit_mask, large, is_zero, memid, tld);
if (p != NULL) return p;
// finally, fall back to the OS
*is_zero = true;
*memid = MI_MEMID_OS;
return _mi_os_alloc_aligned(size, alignment, *commit, large, tld);
p = _mi_os_alloc_aligned(size, alignment, commit, large, tld);
*commit_mask = ((p!=NULL && commit) ? mi_commit_mask_full() : mi_commit_mask_empty());
return p;
}
void* _mi_arena_alloc(size_t size, bool* commit, bool* large, bool* is_zero, size_t* memid, mi_os_tld_t* tld)
void* _mi_arena_alloc(size_t size, bool commit, mi_commit_mask_t* commit_mask, bool* large, bool* is_zero, size_t* memid, mi_os_tld_t* tld)
{
return _mi_arena_alloc_aligned(size, MI_ARENA_BLOCK_SIZE, commit, large, is_zero, memid, tld);
return _mi_arena_alloc_aligned(size, MI_ARENA_BLOCK_SIZE, commit, commit_mask, large, is_zero, memid, tld);
}
/* -----------------------------------------------------------
Arena free
----------------------------------------------------------- */
void _mi_arena_free(void* p, size_t size, size_t memid, bool is_committed, bool is_large, mi_os_tld_t* tld) {
void _mi_arena_free(void* p, size_t size, size_t memid, mi_commit_mask_t commit_mask, bool is_large, mi_os_tld_t* tld) {
mi_assert_internal(size > 0 && tld->stats != NULL);
if (p==NULL) return;
if (size==0) return;
if (memid == MI_MEMID_OS) {
// was a direct OS allocation, pass through
if (!mi_cache_push(p, size, memid, is_committed, is_large, tld)) {
_mi_abandoned_await_readers(); // wait unti safe to free
_mi_os_free_ex(p, size, is_committed, tld->stats);
if (!mi_cache_push(p, size, memid, commit_mask, is_large, tld)) {
_mi_abandoned_await_readers(); // wait until safe to free
// TODO: is it safe on all platforms to free even it contains decommitted parts? (eg. macOS)
const size_t csize = mi_commit_mask_committed_size(commit_mask, size);
_mi_stat_decrease(&_mi_stats_main.committed, csize);
_mi_os_free_ex(p, size, false /*pretend decommitted to not double count stats*/, tld->stats);
}
}
else {