/* Calculate speed of allocation and waste memory of specific space respectively,
* then decide whether to execute a space tuning according to the infomation.*/
static void gc_decide_space_tune(GC* gc)
{
Blocked_Space* mspace = (Blocked_Space*)gc_get_mos((GC_Gen*)gc);
Blocked_Space* fspace = (Blocked_Space*)gc_get_nos((GC_Gen*)gc);
Space* lspace = (Space*)gc_get_los((GC_Gen*)gc);
Space_Tuner* tuner = gc->tuner;
tuner->speed_los = lspace->accumu_alloced_size;
tuner->speed_los = (tuner->speed_los + tuner->last_speed_los) >> 1;
/*The possible survivors from the newly allocated NOS should be counted into the speed of MOS*/
tuner->speed_mos = mspace->accumu_alloced_size + (uint64)((float)fspace->last_alloced_size * fspace->survive_ratio);;
tuner->speed_mos = (tuner->speed_mos + tuner->last_speed_mos) >> 1;
tuner->speed_nos = fspace->accumu_alloced_size;
tuner->speed_nos = (tuner->speed_nos + tuner->last_speed_nos) >> 1;
/*Statistic wasted memory*/
uint64 curr_used_los = lspace->last_surviving_size + lspace->last_alloced_size;
uint64 curr_wast_los = 0;
if(gc->cause != GC_CAUSE_LOS_IS_FULL) curr_wast_los = lspace->committed_heap_size - curr_used_los;
tuner->wast_los += (POINTER_SIZE_INT)curr_wast_los;
uint64 curr_used_mos =
mspace->period_surviving_size + mspace->accumu_alloced_size + (uint64)(fspace->last_alloced_size * fspace->survive_ratio);
float expected_mos_ratio = mspace_get_expected_threshold_ratio((Mspace*)mspace);
uint64 expected_mos = (uint64)((mspace->committed_heap_size + fspace->committed_heap_size) * expected_mos_ratio);
uint64 curr_wast_mos = 0;
if(expected_mos > curr_used_mos) curr_wast_mos = expected_mos - curr_used_mos;
tuner->wast_mos += curr_wast_mos;
tuner->current_dw = ABS_DIFF(tuner->wast_mos, tuner->wast_los);
/*For_statistic ds in heuristic*/
tuner->current_ds = (unsigned int)((float)fspace->committed_heap_size * fspace->survive_ratio);
/*Fixme: Threshold should be computed by heuristic. tslow, total recycled heap size shold be statistic.*/
tuner->threshold_waste = tuner->current_ds;
if(tuner->threshold_waste > 8 * MB) tuner->threshold_waste = 8 * MB;
tuner->min_tuning_size = tuner->current_ds;
if(tuner->min_tuning_size > 4 * MB) tuner->min_tuning_size = 4 * MB;
if(tuner->speed_los == 0) tuner->speed_los = 16;
if(tuner->speed_mos == 0) tuner->speed_mos = 16;
/*Needn't tune if dw does not reach threshold.*/
if(tuner->current_dw > tuner->threshold_waste) tuner->need_tune = 1;
/*If LOS is full, we should tune at lease "tuner->least_tuning_size" size*/
if(gc->cause == GC_CAUSE_LOS_IS_FULL) tuner->force_tune = 1;
return;
}
void verifier_scan_all_objects(Heap_Verifier* heap_verifier)
{
#if !defined(USE_UNIQUE_MARK_SWEEP_GC) && !defined(USE_UNIQUE_MOVE_COMPACT_GC)
GC_Gen* gc = (GC_Gen*)heap_verifier->gc;
Space* fspace = gc_get_nos(gc);
Space* mspace = gc_get_mos(gc);
Space* lspace = gc_get_los(gc);
verifier_scan_nos_mos_objects(fspace, heap_verifier);
verifier_scan_nos_mos_objects(mspace, heap_verifier);
verifier_scan_los_objects(lspace, heap_verifier);
#else
assert(0);
#endif
}
void mutator_initialize(GC* gc, void *unused_gc_information)
{
/* FIXME:: make sure gc_info is cleared */
Mutator *mutator = (Mutator *)STD_MALLOC(sizeof(Mutator));
memset(mutator, 0, sizeof(Mutator));
mutator->alloc_space = gc_get_nos((GC_Gen*)gc);
mutator->gc = gc;
if(gc_is_gen_mode()){
mutator->rem_set = free_set_pool_get_entry(gc->metadata);
assert(vector_block_is_empty(mutator->rem_set));
}
mutator->dirty_set = free_set_pool_get_entry(gc->metadata);
if(!IGNORE_FINREF )
mutator->obj_with_fin = finref_get_free_block(gc);
else
mutator->obj_with_fin = NULL;
#ifdef USE_UNIQUE_MARK_SWEEP_GC
allocator_init_local_chunks((Allocator*)mutator);
#endif
lock(gc->mutator_list_lock); // vvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
mutator->next = (Mutator *)gc->mutator_list;
gc->mutator_list = mutator;
gc->num_mutators++;
/*Begin to measure the mutator thread execution time. */
mutator->time_measurement_start = time_now();
unlock(gc->mutator_list_lock); // ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
gc_set_tls(mutator);
return;
}
/* The tuning size computing before marking is not precise. We only estimate the probable direction of space tuning.
* If this function decide to set TRANS_NOTHING, then we just call the normal marking function.
* Else, we call the marking function for space tuning. */
void gc_compute_space_tune_size_before_marking(GC* gc)
{
if(collect_is_minor()) return;
gc_decide_space_tune(gc);
Space_Tuner* tuner = gc->tuner;
assert((tuner->speed_los != 0) && ( tuner->speed_mos != 0)) ;
if((!tuner->need_tune) && (!tuner->force_tune)) return;
Blocked_Space* mspace = (Blocked_Space*)gc_get_mos((GC_Gen*)gc);
Blocked_Space* fspace = (Blocked_Space*)gc_get_nos((GC_Gen*)gc);
Space* lspace = (Space*)gc_get_los((GC_Gen*)gc);
POINTER_SIZE_INT los_expect_surviving_sz = (POINTER_SIZE_INT)((float)(lspace->last_surviving_size + lspace->last_alloced_size) * lspace->survive_ratio);
POINTER_SIZE_INT los_expect_free_sz = ((lspace->committed_heap_size > los_expect_surviving_sz) ?
(lspace->committed_heap_size - los_expect_surviving_sz) : 0);
POINTER_SIZE_INT mos_expect_survive_sz = (POINTER_SIZE_INT)((float)(mspace->period_surviving_size + mspace->accumu_alloced_size) * mspace->survive_ratio);
float mos_expect_threshold_ratio = mspace_get_expected_threshold_ratio((Mspace*)mspace);
POINTER_SIZE_INT mos_expect_threshold = (POINTER_SIZE_INT)((mspace->committed_heap_size + fspace->committed_heap_size) * mos_expect_threshold_ratio);
POINTER_SIZE_INT mos_expect_free_sz = ((mos_expect_threshold > mos_expect_survive_sz)?
(mos_expect_threshold - mos_expect_survive_sz) : 0);
POINTER_SIZE_INT non_los_expect_surviving_sz = (POINTER_SIZE_INT)(mos_expect_survive_sz + fspace->last_alloced_size * fspace->survive_ratio);
POINTER_SIZE_INT non_los_committed_size = mspace->committed_heap_size + fspace->committed_heap_size;
POINTER_SIZE_INT non_los_expect_free_sz = (non_los_committed_size > non_los_expect_surviving_sz) ? (non_los_committed_size - non_los_expect_surviving_sz):(0) ;
#ifdef SPACE_TUNE_BY_MAJOR_SPEED
/*Fixme: tuner->speed_los here should be computed by sliding compact LOS, to be implemented!*/
POINTER_SIZE_INT total_expect_free_sz = los_expect_free_sz + mos_expect_free_sz;
float new_los_ratio = (float)tuner->speed_los / (float)(tuner->speed_los + tuner->speed_mos);
POINTER_SIZE_INT new_free_los_sz = (POINTER_SIZE_INT)((float)total_expect_free_sz * new_los_ratio);
#else
POINTER_SIZE_INT total_expect_free_sz = los_expect_free_sz + non_los_expect_free_sz;
float new_los_ratio = (float)tuner->speed_los / (float)(tuner->speed_los + tuner->speed_nos);
POINTER_SIZE_INT new_free_los_sz = (POINTER_SIZE_INT)((float)total_expect_free_sz * new_los_ratio);
#endif
/*LOS_Extend:*/
if((new_free_los_sz > los_expect_free_sz) )
{
tuner->kind = TRANS_FROM_MOS_TO_LOS;
tuner->tuning_size = new_free_los_sz - los_expect_free_sz;
}
/*LOS_Shrink:*/
else if(new_free_los_sz < los_expect_free_sz)
{
tuner->kind = TRANS_FROM_LOS_TO_MOS;
tuner->tuning_size = los_expect_free_sz - new_free_los_sz;
}
/*Nothing*/
else
{
tuner->tuning_size = 0;
}
/*If not force tune, and the tuning size is too small, tuner will not take effect.*/
if( (!tuner->force_tune) && (tuner->tuning_size < tuner->min_tuning_size) ){
tuner->kind = TRANS_NOTHING;
tuner->tuning_size = 0;
}
/*If los or non-los is already the smallest size, there is no need to tune anymore.
*But we give "force tune" a chance to extend the whole heap size down there.
*/
if(((lspace->committed_heap_size <= min_los_size_bytes) && (tuner->kind == TRANS_FROM_LOS_TO_MOS)) ||
((fspace->committed_heap_size + mspace->committed_heap_size <= min_none_los_size_bytes) && (tuner->kind == TRANS_FROM_MOS_TO_LOS))){
assert((lspace->committed_heap_size == min_los_size_bytes) || (fspace->committed_heap_size + mspace->committed_heap_size == min_none_los_size_bytes));
tuner->kind = TRANS_NOTHING;
tuner->tuning_size = 0;
}
/*If the strategy upward doesn't decide to extend los, but current GC is caused by los, force an extension here.*/
if(tuner->force_tune){
if(tuner->kind != TRANS_FROM_MOS_TO_LOS){
tuner->kind = TRANS_FROM_MOS_TO_LOS;
tuner->tuning_size = 0;
}
}
return;
}
/* This is the real function that decide tuning_size, because we have know the total size of living objects after "mark_scan_heap_for_space_tune". */
void gc_compute_space_tune_size_after_marking(GC *gc)
{
Blocked_Space* mspace = (Blocked_Space*)gc_get_mos((GC_Gen*)gc);
Blocked_Space* fspace = (Blocked_Space*)gc_get_nos((GC_Gen*)gc);
Lspace *lspace = (Lspace*)gc_get_los((GC_Gen*)gc);
Space_Tuner* tuner = gc->tuner;
POINTER_SIZE_INT max_tuning_size = 0;
POINTER_SIZE_INT non_los_size = mspace->committed_heap_size + fspace->committed_heap_size;
if(LOS_ADJUST_BOUNDARY)
gc_compute_live_object_size_after_marking(gc, non_los_size);
else {
unsigned int collector_num = gc->num_active_collectors;
POINTER_SIZE_INT reserve_size = collector_num <<(GC_BLOCK_SHIFT_COUNT+2);
los_live_obj_size = (POINTER_SIZE_INT) lspace->last_surviving_size + reserve_size;
non_los_live_obj_size = ((POINTER_SIZE_INT)(mspace->free_block_idx-mspace->first_block_idx)<<GC_BLOCK_SHIFT_COUNT)+reserve_size;
non_los_live_obj_size = round_up_to_size(non_los_live_obj_size, SPACE_ALLOC_UNIT);
}
check_tuning_size(gc);
/*We should assure that the non_los area is no less than min_none_los_size_bytes*/
POINTER_SIZE_INT max_tune_for_min_non_los = 0;
if(non_los_size > min_none_los_size_bytes)
max_tune_for_min_non_los = non_los_size - min_none_los_size_bytes;
POINTER_SIZE_INT max_tune_for_min_los = 0;
//debug_adjust
assert(lspace->committed_heap_size >= min_los_size_bytes);
max_tune_for_min_los = lspace->committed_heap_size - min_los_size_bytes;
/*Not force tune, LOS_Extend:*/
if(tuner->kind == TRANS_FROM_MOS_TO_LOS)
{
if (gc->committed_heap_size > lspace->committed_heap_size + non_los_live_obj_size){
max_tuning_size = gc->committed_heap_size - lspace->committed_heap_size - non_los_live_obj_size;
if(max_tuning_size > max_tune_for_min_non_los)
max_tuning_size = max_tune_for_min_non_los;
if( tuner->tuning_size > max_tuning_size)
tuner->tuning_size = max_tuning_size;
/*Round down so as not to break max_tuning_size*/
if(LOS_ADJUST_BOUNDARY)
tuner->tuning_size = round_down_to_size(tuner->tuning_size, GC_BLOCK_SIZE_BYTES);
else
tuner->tuning_size = round_down_to_size(tuner->tuning_size, SPACE_ALLOC_UNIT);
/*If tuning size is zero, we should reset kind to NOTHING, in case that gc_init_block_for_collectors relink the block list.*/
if(tuner->tuning_size == 0) tuner->kind = TRANS_NOTHING;
}else{
tuner->tuning_size = 0;
tuner->kind = TRANS_NOTHING;
}
}
/*Not force tune, LOS Shrink*/
else
{
if(lspace->committed_heap_size > los_live_obj_size){
max_tuning_size = lspace->committed_heap_size - los_live_obj_size;
if(max_tuning_size > max_tune_for_min_los)
max_tuning_size = max_tune_for_min_los;
if(tuner->tuning_size > max_tuning_size)
tuner->tuning_size = max_tuning_size;
/*Round down so as not to break max_tuning_size*/
if (LOS_ADJUST_BOUNDARY)
tuner->tuning_size = round_down_to_size(tuner->tuning_size, GC_BLOCK_SIZE_BYTES);
else
tuner->tuning_size = round_down_to_size(tuner->tuning_size, SPACE_ALLOC_UNIT);
if(tuner->tuning_size == 0) tuner->kind = TRANS_NOTHING;
}else{
/* this is possible because of the reservation in gc_compute_live_object_size_after_marking*/
tuner->tuning_size = 0;
tuner->kind = TRANS_NOTHING;
}
}
/*If the tuning strategy give a bigger tuning_size than failure size, we just follow the strategy and set noforce.*/
Boolean doforce = TRUE;
POINTER_SIZE_INT failure_size = lspace_get_failure_size((Lspace*)lspace);
if( (tuner->kind == TRANS_FROM_MOS_TO_LOS) && (!tuner->reverse) && (tuner->tuning_size > failure_size) )
doforce = FALSE;
if( (tuner->force_tune) && (doforce) )
compute_space_tune_size_for_force_tune(gc, max_tune_for_min_non_los);
return;
}
static void check_tuning_size(GC* gc)
{
Space_Tuner* tuner = gc->tuner;
Lspace *lspace = (Lspace*)gc_get_los((GC_Gen*)gc);
Blocked_Space* mspace = (Blocked_Space*)gc_get_mos((GC_Gen*)gc);
Blocked_Space* fspace = (Blocked_Space*)gc_get_nos((GC_Gen*)gc);
POINTER_SIZE_INT los_free_sz = ((lspace->committed_heap_size > los_live_obj_size) ?
(lspace->committed_heap_size - los_live_obj_size) : 0);
#ifdef SPACE_TUNE_BY_MAJOR_SPEED
float mos_expect_threshold_ratio = mspace_get_expected_threshold_ratio((Mspace*)mspace);
POINTER_SIZE_INT mos_expect_threshold = (POINTER_SIZE_INT)((mspace->committed_heap_size + fspace->committed_heap_size) * mos_expect_threshold_ratio);
POINTER_SIZE_INT mos_free_sz = ((mos_expect_threshold > non_los_live_obj_size)?
(mos_expect_threshold - non_los_live_obj_size) : 0);
POINTER_SIZE_INT total_free_sz = los_free_sz + mos_free_sz;
float new_los_ratio = (float)tuner->speed_los / (float)(tuner->speed_los + tuner->speed_mos);
POINTER_SIZE_INT new_free_los_sz = (POINTER_SIZE_INT)((float)total_free_sz * new_los_ratio);
#else
POINTER_SIZE_INT non_los_committed_size = mspace->committed_heap_size + fspace->committed_heap_size;
POINTER_SIZE_INT non_los_free_sz = ((non_los_committed_size > non_los_live_obj_size)?
(non_los_committed_size - non_los_live_obj_size):0);
POINTER_SIZE_INT total_free_sz = los_free_sz + non_los_free_sz;
float new_los_ratio = (float)tuner->speed_los / (float)(tuner->speed_los + tuner->speed_nos);
POINTER_SIZE_INT new_free_los_sz = (POINTER_SIZE_INT)((float)total_free_sz * new_los_ratio);
#endif
/*LOS_Extend:*/
if((new_free_los_sz > los_free_sz) )
{
tuner->kind = TRANS_FROM_MOS_TO_LOS;
tuner->tuning_size = new_free_los_sz - los_free_sz;
}
/*LOS_Shrink:*/
else if(new_free_los_sz < los_free_sz)
{
tuner->kind = TRANS_FROM_LOS_TO_MOS;
tuner->tuning_size = los_free_sz - new_free_los_sz;
}
/*Nothing*/
else
{
tuner->tuning_size = 0;
/*This is necessary, because the original value of kind might not be NOTHING. */
tuner->kind = TRANS_NOTHING;
}
/*If not force tune, and the tuning size is too small, tuner will not take effect.*/
if( (!tuner->force_tune) && (tuner->tuning_size < tuner->min_tuning_size) ){
tuner->kind = TRANS_NOTHING;
tuner->tuning_size = 0;
}
/*If los or non-los is already the smallest size, there is no need to tune anymore.
*But we give "force tune" a chance to extend the whole heap size down there.
*/
if(((lspace->committed_heap_size <= min_los_size_bytes) && (tuner->kind == TRANS_FROM_LOS_TO_MOS)) ||
((fspace->committed_heap_size + mspace->committed_heap_size <= min_none_los_size_bytes) && (tuner->kind == TRANS_FROM_MOS_TO_LOS))){
assert((lspace->committed_heap_size == min_los_size_bytes) || (fspace->committed_heap_size + mspace->committed_heap_size == min_none_los_size_bytes));
tuner->kind = TRANS_NOTHING;
tuner->tuning_size = 0;
}
if(tuner->force_tune){
if(tuner->kind != TRANS_FROM_MOS_TO_LOS){
tuner->kind = TRANS_FROM_MOS_TO_LOS;
tuner->reverse = 1;
}
}
return;
}
/* If this GC is caused by a LOS allocation failure, we set the "force_tune" flag.
* Attention1: The space tuning strategy will extend or shrink LOS according to the wasted memory size and allocation speed.
* If the strategy decide to shrink or the size extended is not large enough to hold the failed object, we set the "doforce" flag in
* function "gc_compute_space_tune_size_after_marking". And only if "force_tune" and "doforce" are both true, we decide the
* size of extention by this function.
* Attention2: The total heap size might extend in this function. */
static void compute_space_tune_size_for_force_tune(GC *gc, POINTER_SIZE_INT max_tune_for_min_non_los)
{
Space_Tuner* tuner = gc->tuner;
Lspace *lspace = (Lspace*)gc_get_los((GC_Gen*)gc);
Blocked_Space* fspace = (Blocked_Space*)gc_get_nos((GC_Gen*)gc);
POINTER_SIZE_INT max_tuning_size = 0;
POINTER_SIZE_INT failure_size = lspace->failure_size;
POINTER_SIZE_INT lspace_free_size = ( (lspace->committed_heap_size > los_live_obj_size) ? (lspace->committed_heap_size - los_live_obj_size) : (0) );
//debug_adjust
assert(!(lspace_free_size % KB));
assert(!(failure_size % KB));
if(lspace_free_size >= failure_size){
tuner->tuning_size = 0;
tuner->kind = TRANS_NOTHING;
}else{
tuner->tuning_size = failure_size -lspace_free_size;
/*We should assure that the tuning size is no more than the free space of non_los area*/
if( gc->committed_heap_size > lspace->committed_heap_size + non_los_live_obj_size )
max_tuning_size = gc->committed_heap_size - lspace->committed_heap_size - non_los_live_obj_size;
if(max_tuning_size > max_tune_for_min_non_los)
max_tuning_size = max_tune_for_min_non_los;
/*Round up to satisfy LOS alloc demand.*/
if(LOS_ADJUST_BOUNDARY) {
tuner->tuning_size = round_up_to_size(tuner->tuning_size, GC_BLOCK_SIZE_BYTES);
max_tuning_size = round_down_to_size(max_tuning_size, GC_BLOCK_SIZE_BYTES);
}else {
tuner->tuning_size = round_up_to_size(tuner->tuning_size, SPACE_ALLOC_UNIT);
max_tuning_size = round_down_to_size(max_tuning_size, SPACE_ALLOC_UNIT);
}
/*If the tuning size is too large, we did nothing and wait for the OOM of JVM*/
/*Fixme: if the heap size is not mx, we can extend the whole heap size*/
if(tuner->tuning_size > max_tuning_size){
tuner->tuning_size = round_up_to_size(tuner->tuning_size, SPACE_ALLOC_UNIT);
max_tuning_size = round_down_to_size(max_tuning_size, SPACE_ALLOC_UNIT);
//debug_adjust
assert(max_heap_size_bytes >= gc->committed_heap_size);
POINTER_SIZE_INT extend_heap_size = 0;
POINTER_SIZE_INT potential_max_tuning_size = max_tuning_size + max_heap_size_bytes - gc->committed_heap_size;
potential_max_tuning_size -= LOS_HEAD_RESERVE_FOR_HEAP_BASE;
//debug_adjust
assert(!(potential_max_tuning_size % SPACE_ALLOC_UNIT));
if(tuner->tuning_size > potential_max_tuning_size){
tuner->tuning_size = 0;
tuner->kind = TRANS_NOTHING;
}else{
/*We have tuner->tuning_size > max_tuning_size up there.*/
extend_heap_size = tuner->tuning_size - max_tuning_size;
blocked_space_extend(fspace, (unsigned int)extend_heap_size);
gc->committed_heap_size += extend_heap_size;
tuner->kind = TRANS_FROM_MOS_TO_LOS;
}
}
else
{
tuner->kind = TRANS_FROM_MOS_TO_LOS;
}
}
return;
}
void move_compact_mspace(Collector* collector)
{
GC* gc = collector->gc;
Mspace* mspace = (Mspace*)gc_get_mos((GC_Gen*)gc);
Lspace* lspace = (Lspace*)gc_get_los((GC_Gen*)gc);
Blocked_Space* nos = (Blocked_Space*)gc_get_nos((GC_Gen*)gc);
unsigned int num_active_collectors = gc->num_active_collectors;
Boolean is_fallback = collect_is_fallback();
/* Pass 1: **************************************************
mark all live objects in heap, and save all the slots that
have references that are going to be repointed */
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"]: pass1: mark live objects in heap ...");
unsigned int old_num = atomic_cas32( &num_marking_collectors, 0, num_active_collectors+1);
if(!is_fallback)
mark_scan_heap(collector);
else
mark_scan_heap_for_fallback(collector);
old_num = atomic_inc32(&num_marking_collectors);
if( ++old_num == num_active_collectors ){
/* last collector's world here */
/* prepare for next phase */
gc_init_block_for_collectors(gc, mspace);
if(!IGNORE_FINREF )
collector_identify_finref(collector);
#ifndef BUILD_IN_REFERENT
else {
gc_set_weakref_sets(gc);
gc_update_weakref_ignore_finref(gc);
}
#endif
gc_identify_dead_weak_roots(gc);
#ifdef USE_32BITS_HASHCODE
if((!LOS_ADJUST_BOUNDARY) && (is_fallback))
fallback_clear_fwd_obj_oi_init(collector);
#endif
debug_num_compact_blocks = 0;
/* let other collectors go */
num_marking_collectors++;
}
while(num_marking_collectors != num_active_collectors + 1);
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"]: finish pass1");
/* Pass 2: **************************************************
move object and set the forwarding offset table */
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"]: pass2: move object and set the forwarding offset table ...");
atomic_cas32( &num_moving_collectors, 0, num_active_collectors+1);
#ifdef USE_32BITS_HASHCODE
if(is_fallback)
fallback_clear_fwd_obj_oi(collector);
#endif
mspace_move_objects(collector, mspace);
old_num = atomic_inc32(&num_moving_collectors);
if( ++old_num == num_active_collectors ){
/* single thread world */
if(lspace->move_object)
lspace_compute_object_target(collector, lspace);
gc->collect_result = gc_collection_result(gc);
if(!gc->collect_result){
num_moving_collectors++;
return;
}
if(verify_live_heap){
assert( debug_num_compact_blocks == mspace->num_managed_blocks + nos->num_managed_blocks );
debug_num_compact_blocks = 0;
}
gc_reset_block_for_collectors(gc, mspace);
blocked_space_block_iterator_init((Blocked_Space*)mspace);
num_moving_collectors++;
}
while(num_moving_collectors != num_active_collectors + 1);
if(!gc->collect_result) return;
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"]: finish pass2");
/* Pass 3: **************************************************
update all references whose pointed objects were moved */
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"]: pass3: update all references ...");
old_num = atomic_cas32( &num_fixing_collectors, 0, num_active_collectors+1);
mspace_fix_repointed_refs(collector, mspace);
old_num = atomic_inc32(&num_fixing_collectors);
if( ++old_num == num_active_collectors ){
/* last collector's world here */
lspace_fix_repointed_refs(collector, lspace);
gc_fix_rootset(collector, FALSE);
if(lspace->move_object) lspace_sliding_compact(collector, lspace);
num_fixing_collectors++;
}
while(num_fixing_collectors != num_active_collectors + 1);
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"]: finish pass3");
/* Pass 4: **************************************************
restore obj_info . Actually only LOS needs it. Since oi is recorded for new address, so the restoration
doesn't need to to specify space. */
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"]: pass4: restore obj_info ...");
atomic_cas32( &num_restoring_collectors, 0, num_active_collectors);
collector_restore_obj_info(collector);
atomic_inc32(&num_restoring_collectors);
while(num_restoring_collectors != num_active_collectors);
/* Dealing with out of memory in mspace */
if(mspace->free_block_idx > nos->first_block_idx){
atomic_cas32( &num_extending_collectors, 0, num_active_collectors);
mspace_extend_compact(collector);
atomic_inc32(&num_extending_collectors);
while(num_extending_collectors != num_active_collectors);
}
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"]: finish pass4");
/* Leftover: **************************************************
*/
if( (POINTER_SIZE_INT)collector->thread_handle != 0 ){
TRACE2("gc.process", "GC: collector["<<((POINTER_SIZE_INT)collector->thread_handle)<<"] finished");
return;
}
TRACE2("gc.process", "GC: collector[0] finished");
return;
}
void gc_init_block_for_collectors(GC* gc, Mspace* mspace)
{
unsigned int i;
Block_Header* block;
Space_Tuner* tuner = gc->tuner;
Block_Header* nos_last_block;
Block_Header* mos_first_block = (Block_Header*)&mspace->blocks[0];
unsigned int trans_blocks = (unsigned int)(tuner->tuning_size >> GC_BLOCK_SHIFT_COUNT);
/*Needn't change LOS size.*/
if(tuner->kind == TRANS_NOTHING){
for(i=0; i<gc->num_active_collectors; i++){
Collector* collector = gc->collectors[i];
block = (Block_Header*)&mspace->blocks[i];
collector->cur_target_block = block;
collector->cur_compact_block = block;
block->status = BLOCK_TARGET;
}
block = (Block_Header*)&mspace->blocks[i];
next_block_for_target = block;
next_block_for_compact = block;
return;
}
//For_LOS_extend
else if(tuner->kind == TRANS_FROM_MOS_TO_LOS)
{
Blocked_Space* nos = (Blocked_Space*)gc_get_nos((GC_Gen*)gc);
if(nos->num_managed_blocks)
nos_last_block = (Block_Header*)&nos->blocks[nos->num_managed_blocks-1];
else
/*If nos->num_managed_blocks is zero, we take mos_last_block as nos_last_block instead.*/
nos_last_block = (Block_Header*)&mspace->blocks[mspace->num_managed_blocks - 1];
nos_last_block->next = mos_first_block;
((Block_Header*)&(mspace->blocks[trans_blocks - 1]))->next = NULL;
for(i=0; i< gc->num_active_collectors; i++){
Collector* collector = gc->collectors[i];
block = (Block_Header*)&mspace->blocks[i + trans_blocks];
collector->cur_target_block = block;
collector->cur_compact_block = block;
block->status = BLOCK_TARGET;
}
block = (Block_Header*)&mspace->blocks[i+trans_blocks];
next_block_for_target = block;
next_block_for_compact = block;
return;
}else
{
gc_space_tuner_init_fake_blocks_for_los_shrink(gc);
Collector* collector = gc->collectors[0];
collector->cur_target_block = tuner->interim_blocks;
collector->cur_target_block->status = BLOCK_TARGET;
if(trans_blocks >= gc->num_active_collectors)
collector->cur_compact_block = mos_first_block;
else
collector->cur_compact_block = gc->tuner->interim_blocks;
collector->cur_compact_block->status = BLOCK_IN_COMPACT;
for(i=1; i< gc->num_active_collectors; i++){
collector = gc->collectors[i];
collector->cur_target_block = gc->collectors[i - 1]->cur_target_block->next;
collector->cur_target_block->status = BLOCK_TARGET;
collector->cur_compact_block = gc->collectors[i - 1]->cur_compact_block->next;
collector->cur_compact_block->status = BLOCK_IN_COMPACT;
}
next_block_for_target = collector->cur_target_block->next;
next_block_for_compact = collector->cur_compact_block->next;
}
}
Block_Header* mspace_get_next_target_block(Collector* collector, Mspace* mspace)
{
Block_Header* cur_target_block = (Block_Header*)next_block_for_target;
/* firstly, we bump the next_block_for_target global var to the first non BLOCK_TARGET block
This need not atomic op, because the global var is only a hint. */
while(cur_target_block->status == BLOCK_TARGET){
cur_target_block = cur_target_block->next;
}
next_block_for_target = cur_target_block;
/* cur_target_block has to be BLOCK_IN_COMPACT|BLOCK_COMPACTED|BLOCK_TARGET. Reason:
Any block after it must be either BLOCK_TARGET, or:
1. Since cur_target_block < cur_compact_block, we at least can get cur_compact_block as target.
2. For a block that is >=cur_target_block and <cur_compact_block.
Since it is before cur_compact_block, we know it must be a compaction block of some thread.
So it is either BLOCK_IN_COMPACT or BLOCK_COMPACTED.
We care only the BLOCK_COMPACTED block or own BLOCK_IN_COMPACT. But I can't make the assert
as below because of a race condition where the block status is not yet updated by other thread.
assert( cur_target_block->status & (BLOCK_IN_COMPACT|BLOCK_COMPACTED|BLOCK_TARGET));
*/
/* mos may be out of space, so we can use nos blocks for compaction target.
* but we can't use the blocks which are given to los when los extension happens.
* in this case, an out-of-mem should be given to user.
*/
GC* gc = collector->gc;
Blocked_Space* nos = (Blocked_Space*)gc_get_nos((GC_Gen*)gc);
Block_Header *nos_end;
if( nos->num_managed_blocks != 0)
nos_end = ((Block_Header *)&nos->blocks[nos->num_managed_blocks-1])->next;
else
nos_end = ((Block_Header *)&mspace->blocks[mspace->num_managed_blocks-1])->next;
while( cur_target_block != nos_end){
//For_LOS_extend
//assert( cur_target_block <= collector->cur_compact_block);
Block_Header* next_target_block = cur_target_block->next;
volatile unsigned int* p_block_status = &cur_target_block->status;
unsigned int block_status = cur_target_block->status;
//assert( block_status & (BLOCK_IN_COMPACT|BLOCK_COMPACTED|BLOCK_TARGET));
/* if it is not BLOCK_COMPACTED, let's move on to next except it's own cur_compact_block */
if(block_status != BLOCK_COMPACTED){
if(cur_target_block == collector->cur_compact_block){
assert( block_status == BLOCK_IN_COMPACT);
*p_block_status = BLOCK_TARGET;
collector->cur_target_block = cur_target_block;
return cur_target_block;
}
/* it's not my own cur_compact_block, it can be BLOCK_TARGET or other's cur_compact_block */
cur_target_block = next_target_block;
continue;
}
/* else, find a BLOCK_COMPACTED before own cur_compact_block */
unsigned int temp = atomic_cas32(p_block_status, BLOCK_TARGET, BLOCK_COMPACTED);
if(temp == BLOCK_COMPACTED){
collector->cur_target_block = cur_target_block;
return cur_target_block;
}
/* missed it, it must be set by other into BLOCK_TARGET */
assert(temp == BLOCK_TARGET);
cur_target_block = next_target_block;
}
/* mos is run out for major collection */
return NULL;
}
