Decouple incremental marking step from page sizes

Currently, the number of incremental marking steps is calculated based
on the number of pooled pages available. This means that if we make Ruby
heap pages larger, it would run fewer incremental marking steps (which
would mean each incremental marking step takes longer).

This commit changes incremental marking to run after every
INCREMENTAL_MARK_STEP_ALLOCATIONS number of allocations. This means that
the behaviour of incremental marking remains the same regardless of the
Ruby heap page size.

I've benchmarked against discourse benchmarks and did not get a
significant change in response times beyond the margin of error. This is
expected as this new incremental marking algorithm behaves very
similarly to the previous one.
This commit is contained in:
Peter Zhu 2022-03-29 13:57:09 -04:00
Родитель ad808506b3
Коммит dde164e968
2 изменённых файлов: 137 добавлений и 96 удалений

232
gc.c
Просмотреть файл

@ -883,6 +883,10 @@ enum {
#define HEAP_PAGE_ALIGN (1 << HEAP_PAGE_ALIGN_LOG)
#define HEAP_PAGE_SIZE HEAP_PAGE_ALIGN
#if GC_ENABLE_INCREMENTAL_MARK && !defined(INCREMENTAL_MARK_STEP_ALLOCATIONS)
# define INCREMENTAL_MARK_STEP_ALLOCATIONS 500
#endif
#ifdef HAVE_MMAP
# if HAVE_CONST_PAGE_SIZE
/* If we have the HEAP_PAGE and it is a constant, then we can directly use it. */
@ -1182,7 +1186,7 @@ static inline void gc_exit(rb_objspace_t *objspace, enum gc_enter_event event, u
static void gc_marks(rb_objspace_t *objspace, int full_mark);
static void gc_marks_start(rb_objspace_t *objspace, int full);
static int gc_marks_finish(rb_objspace_t *objspace);
static void gc_marks_finish(rb_objspace_t *objspace);
static void gc_marks_rest(rb_objspace_t *objspace);
static void gc_marks_continue(rb_objspace_t *objspace, rb_size_pool_t *size_pool, rb_heap_t *heap);
@ -2398,22 +2402,36 @@ rb_gc_size_allocatable_p(size_t size)
}
static inline VALUE
ractor_cached_free_region(rb_objspace_t *objspace, rb_ractor_t *cr, size_t size_pool_idx)
ractor_cache_allocate_slot(rb_objspace_t *objspace, rb_ractor_newobj_cache_t *cache,
size_t size_pool_idx)
{
rb_ractor_newobj_size_pool_cache_t *cache = &cr->newobj_cache.size_pool_caches[size_pool_idx];
RVALUE *p = cache->freelist;
rb_ractor_newobj_size_pool_cache_t *size_pool_cache = &cache->size_pool_caches[size_pool_idx];
RVALUE *p = size_pool_cache->freelist;
#if GC_ENABLE_INCREMENTAL_MARK
if (is_incremental_marking(objspace)) {
// Not allowed to allocate without running an incremental marking step
if (cache->incremental_mark_step_allocated_slots >= INCREMENTAL_MARK_STEP_ALLOCATIONS) {
return Qfalse;
}
if (p) {
cache->incremental_mark_step_allocated_slots++;
}
}
#endif
if (p) {
VALUE obj = (VALUE)p;
MAYBE_UNUSED(const size_t) stride = size_pool_slot_size(size_pool_idx);
cache->freelist = p->as.free.next;
size_pool_cache->freelist = p->as.free.next;
#if USE_RVARGC
asan_unpoison_memory_region(p, stride, true);
#else
asan_unpoison_object(obj, true);
#endif
#if RGENGC_CHECK_MODE
GC_ASSERT(cache->using_page->slot_size == (short)stride);
GC_ASSERT(rb_gc_obj_slot_size(obj) == stride);
// zero clear
MEMZERO((char *)obj, char, stride);
#endif
@ -2425,14 +2443,14 @@ ractor_cached_free_region(rb_objspace_t *objspace, rb_ractor_t *cr, size_t size_
}
static struct heap_page *
heap_next_freepage(rb_objspace_t *objspace, rb_size_pool_t *size_pool, rb_heap_t *heap)
heap_next_free_page(rb_objspace_t *objspace, rb_size_pool_t *size_pool, rb_heap_t *heap)
{
ASSERT_vm_locking();
struct heap_page *page;
while (heap->free_pages == NULL) {
heap_prepare(objspace, size_pool, heap);
heap_prepare(objspace, size_pool, heap);
}
page = heap->free_pages;
heap->free_pages = page->free_next;
@ -2446,31 +2464,25 @@ heap_next_freepage(rb_objspace_t *objspace, rb_size_pool_t *size_pool, rb_heap_t
}
static inline void
ractor_set_cache(rb_ractor_t *cr, struct heap_page *page, size_t size_pool_idx)
ractor_cache_set_page(rb_ractor_newobj_cache_t *cache, size_t size_pool_idx,
struct heap_page *page)
{
gc_report(3, &rb_objspace, "ractor_set_cache: Using page %p\n", (void *)GET_PAGE_BODY(page->start));
rb_ractor_newobj_size_pool_cache_t *cache = &cr->newobj_cache.size_pool_caches[size_pool_idx];
rb_ractor_newobj_size_pool_cache_t *size_pool_cache = &cache->size_pool_caches[size_pool_idx];
cache->using_page = page;
cache->freelist = page->freelist;
GC_ASSERT(size_pool_cache->freelist == NULL);
GC_ASSERT(page->free_slots != 0);
GC_ASSERT(page->freelist != NULL);
size_pool_cache->using_page = page;
size_pool_cache->freelist = page->freelist;
page->free_slots = 0;
page->freelist = NULL;
asan_unpoison_object((VALUE)cache->freelist, false);
GC_ASSERT(RB_TYPE_P((VALUE)cache->freelist, T_NONE));
asan_poison_object((VALUE)cache->freelist);
}
static inline void
ractor_cache_slots(rb_objspace_t *objspace, rb_ractor_t *cr, size_t size_pool_idx)
{
ASSERT_vm_locking();
rb_size_pool_t *size_pool = &size_pools[size_pool_idx];
struct heap_page *page = heap_next_freepage(objspace, size_pool, SIZE_POOL_EDEN_HEAP(size_pool));
ractor_set_cache(cr, page, size_pool_idx);
asan_unpoison_object((VALUE)size_pool_cache->freelist, false);
GC_ASSERT(RB_TYPE_P((VALUE)size_pool_cache->freelist, T_NONE));
asan_poison_object((VALUE)size_pool_cache->freelist);
}
static inline VALUE
@ -2509,6 +2521,58 @@ size_pool_idx_for_size(size_t size)
#endif
}
static VALUE
newobj_alloc(rb_objspace_t *objspace, rb_ractor_t *cr, size_t size_pool_idx, bool vm_locked)
{
rb_size_pool_t *size_pool = &size_pools[size_pool_idx];
rb_heap_t *heap = SIZE_POOL_EDEN_HEAP(size_pool);
rb_ractor_newobj_cache_t *cache = &cr->newobj_cache;
VALUE obj = ractor_cache_allocate_slot(objspace, cache, size_pool_idx);
if (UNLIKELY(obj == Qfalse)) {
unsigned int lev;
bool unlock_vm = false;
if (!vm_locked) {
RB_VM_LOCK_ENTER_CR_LEV(cr, &lev);
vm_locked = true;
unlock_vm = true;
}
{
ASSERT_vm_locking();
#if GC_ENABLE_INCREMENTAL_MARK
if (is_incremental_marking(objspace)) {
gc_marks_continue(objspace, size_pool, heap);
cache->incremental_mark_step_allocated_slots = 0;
// Retry allocation after resetting incremental_mark_step_allocated_slots
obj = ractor_cache_allocate_slot(objspace, cache, size_pool_idx);
}
#endif
if (obj == Qfalse) {
// Get next free page (possibly running GC)
struct heap_page *page = heap_next_free_page(objspace, size_pool, heap);
ractor_cache_set_page(cache, size_pool_idx, page);
// Retry allocation after moving to new page
obj = ractor_cache_allocate_slot(objspace, cache, size_pool_idx);
GC_ASSERT(obj != Qfalse);
}
}
if (unlock_vm) {
RB_VM_LOCK_LEAVE_CR_LEV(cr, &lev);
}
}
return obj;
}
ALWAYS_INLINE(static VALUE newobj_slowpath(VALUE klass, VALUE flags, rb_objspace_t *objspace, rb_ractor_t *cr, int wb_protected, size_t size_pool_idx));
static inline VALUE
@ -2533,11 +2597,7 @@ newobj_slowpath(VALUE klass, VALUE flags, rb_objspace_t *objspace, rb_ractor_t *
}
}
// allocate new slot
while ((obj = ractor_cached_free_region(objspace, cr, size_pool_idx)) == Qfalse) {
ractor_cache_slots(objspace, cr, size_pool_idx);
}
GC_ASSERT(obj != 0);
obj = newobj_alloc(objspace, cr, size_pool_idx, true);
newobj_init(klass, flags, wb_protected, objspace, obj);
gc_event_hook_prep(objspace, RUBY_INTERNAL_EVENT_NEWOBJ, obj, newobj_fill(obj, 0, 0, 0));
@ -2584,12 +2644,11 @@ newobj_of0(VALUE klass, VALUE flags, int wb_protected, rb_ractor_t *cr, size_t a
size_t size_pool_idx = size_pool_idx_for_size(alloc_size);
if ((!UNLIKELY(during_gc ||
ruby_gc_stressful ||
gc_event_hook_available_p(objspace)) &&
wb_protected &&
(obj = ractor_cached_free_region(objspace, cr, size_pool_idx)) != Qfalse)) {
if (!UNLIKELY(during_gc ||
ruby_gc_stressful ||
gc_event_hook_available_p(objspace)) &&
wb_protected) {
obj = newobj_alloc(objspace, cr, size_pool_idx, false);
newobj_init(klass, flags, wb_protected, objspace, obj);
}
else {
@ -7936,6 +7995,27 @@ gc_verify_transient_heap_internal_consistency(VALUE dmy)
return Qnil;
}
#if GC_ENABLE_INCREMENTAL_MARK
static void
heap_move_pooled_pages_to_free_pages(rb_heap_t *heap)
{
if (heap->pooled_pages) {
if (heap->free_pages) {
struct heap_page *free_pages_tail = heap->free_pages;
while (free_pages_tail->free_next) {
free_pages_tail = free_pages_tail->free_next;
}
free_pages_tail->free_next = heap->pooled_pages;
}
else {
heap->free_pages = heap->pooled_pages;
}
heap->pooled_pages = NULL;
}
}
#endif
/* marks */
static void
@ -7947,7 +8027,8 @@ gc_marks_start(rb_objspace_t *objspace, int full_mark)
if (full_mark) {
#if GC_ENABLE_INCREMENTAL_MARK
objspace->rincgc.step_slots = (objspace->marked_slots * 2) / ((objspace->rincgc.pooled_slots / HEAP_PAGE_OBJ_LIMIT) + 1);
size_t incremental_marking_steps = (objspace->rincgc.pooled_slots / INCREMENTAL_MARK_STEP_ALLOCATIONS) + 1;
objspace->rincgc.step_slots = (objspace->marked_slots * 2) / incremental_marking_steps;
if (0) fprintf(stderr, "objspace->marked_slots: %"PRIdSIZE", "
"objspace->rincgc.pooled_page_num: %"PRIdSIZE", "
@ -7965,7 +8046,10 @@ gc_marks_start(rb_objspace_t *objspace, int full_mark)
objspace->marked_slots = 0;
for (int i = 0; i < SIZE_POOL_COUNT; i++) {
rgengc_mark_and_rememberset_clear(objspace, SIZE_POOL_EDEN_HEAP(&size_pools[i]));
rb_size_pool_t *size_pool = &size_pools[i];
rb_heap_t *heap = SIZE_POOL_EDEN_HEAP(size_pool);
rgengc_mark_and_rememberset_clear(objspace, heap);
heap_move_pooled_pages_to_free_pages(heap);
}
}
else {
@ -8029,48 +8113,21 @@ gc_marks_wb_unprotected_objects(rb_objspace_t *objspace, rb_heap_t *heap)
gc_mark_stacked_objects_all(objspace);
}
static struct heap_page *
heap_move_pooled_pages_to_free_pages(rb_heap_t *heap)
{
struct heap_page *page = heap->pooled_pages;
if (page) {
heap->pooled_pages = page->free_next;
heap_add_freepage(heap, page);
}
return page;
}
#endif
static int
static void
gc_marks_finish(rb_objspace_t *objspace)
{
#if GC_ENABLE_INCREMENTAL_MARK
/* finish incremental GC */
if (is_incremental_marking(objspace)) {
for (int i = 0; i < SIZE_POOL_COUNT; i++) {
rb_heap_t *heap = SIZE_POOL_EDEN_HEAP(&size_pools[i]);
if (heap->pooled_pages) {
heap_move_pooled_pages_to_free_pages(heap);
gc_report(1, objspace, "gc_marks_finish: pooled pages are exists. retry.\n");
return FALSE; /* continue marking phase */
}
}
if (RGENGC_CHECK_MODE && is_mark_stack_empty(&objspace->mark_stack) == 0) {
rb_bug("gc_marks_finish: mark stack is not empty (%"PRIdSIZE").",
mark_stack_size(&objspace->mark_stack));
}
gc_mark_roots(objspace, 0);
if (is_mark_stack_empty(&objspace->mark_stack) == FALSE) {
gc_report(1, objspace, "gc_marks_finish: not empty (%"PRIdSIZE"). retry.\n",
mark_stack_size(&objspace->mark_stack));
return FALSE;
}
while (gc_mark_stacked_objects_incremental(objspace, INT_MAX) == false);
#if RGENGC_CHECK_MODE >= 2
if (gc_verify_heap_pages(objspace) != 0) {
@ -8185,8 +8242,6 @@ gc_marks_finish(rb_objspace_t *objspace)
rb_ractor_finish_marking();
gc_event_hook(objspace, RUBY_INTERNAL_EVENT_GC_END_MARK, 0);
return TRUE;
}
#if GC_ENABLE_INCREMENTAL_MARK
@ -8196,10 +8251,8 @@ gc_marks_step(rb_objspace_t *objspace, size_t slots)
GC_ASSERT(is_marking(objspace));
if (gc_mark_stacked_objects_incremental(objspace, slots)) {
if (gc_marks_finish(objspace)) {
/* finish */
gc_sweep(objspace);
}
gc_marks_finish(objspace);
gc_sweep(objspace);
}
if (0) fprintf(stderr, "objspace->marked_slots: %"PRIdSIZE"\n", objspace->marked_slots);
}
@ -8217,15 +8270,14 @@ gc_marks_rest(rb_objspace_t *objspace)
#endif
if (is_incremental_marking(objspace)) {
do {
while (gc_mark_stacked_objects_incremental(objspace, INT_MAX) == FALSE);
} while (gc_marks_finish(objspace) == FALSE);
while (gc_mark_stacked_objects_incremental(objspace, INT_MAX) == FALSE);
}
else {
gc_mark_stacked_objects_all(objspace);
gc_marks_finish(objspace);
}
gc_marks_finish(objspace);
/* move to sweep */
gc_sweep(objspace);
}
@ -8239,24 +8291,8 @@ gc_marks_continue(rb_objspace_t *objspace, rb_size_pool_t *size_pool, rb_heap_t
unsigned int lock_lev;
gc_enter(objspace, gc_enter_event_mark_continue, &lock_lev);
int slots = 0;
const char *from;
if (heap->pooled_pages) {
while (heap->pooled_pages && slots < HEAP_PAGE_OBJ_LIMIT) {
struct heap_page *page = heap_move_pooled_pages_to_free_pages(heap);
slots += page->free_slots;
}
from = "pooled-pages";
}
else if (heap_increment(objspace, size_pool, heap)) {
slots = heap->free_pages->free_slots;
from = "incremented-pages";
}
if (slots > 0) {
gc_report(2, objspace, "gc_marks_continue: provide %d slots from %s.\n",
slots, from);
if (heap->free_pages) {
gc_report(2, objspace, "gc_marks_continue: has pooled pages");
gc_marks_step(objspace, objspace->rincgc.step_slots);
}
else {
@ -8771,6 +8807,10 @@ rb_obj_gc_flags(VALUE obj, ID* flags, size_t max)
void
rb_gc_ractor_newobj_cache_clear(rb_ractor_newobj_cache_t *newobj_cache)
{
#if GC_ENABLE_INCREMENTAL_MARK
newobj_cache->incremental_mark_step_allocated_slots = 0;
#endif
for (size_t size_pool_idx = 0; size_pool_idx < SIZE_POOL_COUNT; size_pool_idx++) {
rb_ractor_newobj_size_pool_cache_t *cache = &newobj_cache->size_pool_caches[size_pool_idx];

Просмотреть файл

@ -79,6 +79,7 @@ typedef struct ractor_newobj_size_pool_cache {
} rb_ractor_newobj_size_pool_cache_t;
typedef struct ractor_newobj_cache {
size_t incremental_mark_step_allocated_slots;
rb_ractor_newobj_size_pool_cache_t size_pool_caches[SIZE_POOL_COUNT];
} rb_ractor_newobj_cache_t;