Skip to content

Commit

Permalink
Simplify sweeping of big values (JuliaLang#54936)
Browse files Browse the repository at this point in the history 
Simplifies the layout of the doubly linked list of big objects to make
it a bit more canonical: let's just store a pointer to the previous
element, instead of storing a "pointer to the next element of the
previous element". This should make the implementation a bit easier to
understand without incurring any memory overhead.

I ran the serial and multithreaded benchmarks from GCBenchmarks and this
seems fairly close to performance neutral on my machine. We also ran our
internal benchmarks on it at RAI and it looks fine from a correctness
and performance point of view.

---------

Co-authored-by: Kiran Pamnany <[email protected]>
  • Loading branch information
@d-netto @kpamnany
d-netto and kpamnany committed Jul 25, 2024
1 parent 9490ee3 commit 1f66e0d
Show file tree
Hide file tree
Showing 4 changed files with 105 additions and 100 deletions.
8 changes: 4 additions & 4 deletions src/gc-debug.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -132,7 +132,7 @@ static void clear_mark(int bits)
}
bigval_t *v;
for (int i = 0; i < gc_n_threads; i++) {
v = gc_all_tls_states[i]->gc_tls.heap.big_objects;
v = gc_all_tls_states[i]->gc_tls.heap.young_generation_of_bigvals;
while (v != NULL) {
void *gcv = &v->header;
if (!gc_verifying)
Expand All @@ -142,7 +142,7 @@ static void clear_mark(int bits)
}
}

v = big_objects_marked;
v = oldest_generation_of_bigvals;
while (v != NULL) {
void *gcv = &v->header;
if (!gc_verifying)
Expand Down Expand Up @@ -965,15 +965,15 @@ void gc_stats_big_obj(void)
size_t nused=0, nbytes=0, nused_old=0, nbytes_old=0;
for (int t_i = 0; t_i < gc_n_threads; t_i++) {
jl_ptls_t ptls2 = gc_all_tls_states[t_i];
bigval_t *v = ptls2->gc_tls.heap.big_objects;
bigval_t *v = ptls2->gc_tls.heap.young_generation_of_bigvals;
while (v != NULL) {
if (gc_marked(v->bits.gc)) {
nused++;
nbytes += v->sz & ~3;
}
v = v->next;
}
v = big_objects_marked;
v = oldest_generation_of_bigvals;
while (v != NULL) {
if (gc_marked(v->bits.gc)) {
nused_old++;
Expand Down
4 changes: 2 additions & 2 deletions src/gc-tls.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -31,8 +31,8 @@ typedef struct {
struct _mallocarray_t *mallocarrays;
struct _mallocarray_t *mafreelist;

// variables for tracking big objects
struct _bigval_t *big_objects;
// variable for tracking young (i.e. not in `GC_OLD_MARKED`/last generation) large objects
struct _bigval_t *young_generation_of_bigvals;

// lower bound of the number of pointers inside remembered values
int remset_nptr;
Expand Down
158 changes: 76 additions & 82 deletions src/gc.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,8 @@ uv_cond_t gc_threads_cond;
uv_sem_t gc_sweep_assists_needed;
// Mutex used to coordinate entry of GC threads in the mark loop
uv_mutex_t gc_queue_observer_lock;
// Tag for sentinel nodes in bigval list
uintptr_t gc_bigval_sentinel_tag;

// Linked list of callback functions

Expand Down Expand Up @@ -150,7 +152,6 @@ JL_DLLEXPORT void jl_gc_set_cb_notify_gc_pressure(jl_gc_cb_notify_gc_pressure_t
// is going to realloc the buffer (of its own list) or accessing the
// list of another thread
static jl_mutex_t finalizers_lock;
static uv_mutex_t gc_cache_lock;

// mutex for gc-heap-snapshot.
jl_mutex_t heapsnapshot_lock;
Expand Down Expand Up @@ -201,8 +202,8 @@ JL_DLLEXPORT uintptr_t jl_get_buff_tag(void) JL_NOTSAFEPOINT
return jl_buff_tag;
}

// List of marked big objects. Not per-thread. Accessed only by master thread.
bigval_t *big_objects_marked = NULL;
// List of big objects in oldest generation (`GC_OLD_MARKED`). Not per-thread. Accessed only by master thread.
bigval_t *oldest_generation_of_bigvals = NULL;

// -- Finalization --
// `ptls->finalizers` and `finalizer_list_marked` might have tagged pointers.
Expand Down Expand Up @@ -759,60 +760,25 @@ static int64_t t_start = 0; // Time GC starts;
static int64_t last_trim_maxrss = 0;
#endif

static void gc_sync_cache_nolock(jl_ptls_t ptls, jl_gc_mark_cache_t *gc_cache) JL_NOTSAFEPOINT
static void gc_sync_cache(jl_ptls_t ptls, jl_gc_mark_cache_t *gc_cache) JL_NOTSAFEPOINT
{
const int nbig = gc_cache->nbig_obj;
for (int i = 0; i < nbig; i++) {
void *ptr = gc_cache->big_obj[i];
bigval_t *hdr = (bigval_t*)gc_ptr_clear_tag(ptr, 1);
gc_big_object_unlink(hdr);
if (gc_ptr_tag(ptr, 1)) {
gc_big_object_link(hdr, &ptls->gc_tls.heap.big_objects);
}
else {
// Move hdr from `big_objects` list to `big_objects_marked list`
gc_big_object_link(hdr, &big_objects_marked);
}
}
gc_cache->nbig_obj = 0;
perm_scanned_bytes += gc_cache->perm_scanned_bytes;
scanned_bytes += gc_cache->scanned_bytes;
gc_cache->perm_scanned_bytes = 0;
gc_cache->scanned_bytes = 0;
}

static void gc_sync_cache(jl_ptls_t ptls) JL_NOTSAFEPOINT
{
uv_mutex_lock(&gc_cache_lock);
gc_sync_cache_nolock(ptls, &ptls->gc_tls.gc_cache);
uv_mutex_unlock(&gc_cache_lock);
}

// No other threads can be running marking at the same time
static void gc_sync_all_caches_nolock(jl_ptls_t ptls)
static void gc_sync_all_caches(jl_ptls_t ptls)
{
assert(gc_n_threads);
for (int t_i = 0; t_i < gc_n_threads; t_i++) {
jl_ptls_t ptls2 = gc_all_tls_states[t_i];
if (ptls2 != NULL)
gc_sync_cache_nolock(ptls, &ptls2->gc_tls.gc_cache);
gc_sync_cache(ptls, &ptls2->gc_tls.gc_cache);
}
}

STATIC_INLINE void gc_queue_big_marked(jl_ptls_t ptls, bigval_t *hdr,
int toyoung) JL_NOTSAFEPOINT
{
const int nentry = sizeof(ptls->gc_tls.gc_cache.big_obj) / sizeof(void*);
size_t nobj = ptls->gc_tls.gc_cache.nbig_obj;
if (__unlikely(nobj >= nentry)) {
gc_sync_cache(ptls);
nobj = 0;
}
uintptr_t v = (uintptr_t)hdr;
ptls->gc_tls.gc_cache.big_obj[nobj] = (void*)(toyoung ? (v | 1) : v);
ptls->gc_tls.gc_cache.nbig_obj = nobj + 1;
}

// Atomically set the mark bit for object and return whether it was previously unmarked
FORCE_INLINE int gc_try_setmark_tag(jl_taggedvalue_t *o, uint8_t mark_mode) JL_NOTSAFEPOINT
{
Expand Down Expand Up @@ -849,16 +815,14 @@ STATIC_INLINE void gc_setmark_big(jl_ptls_t ptls, jl_taggedvalue_t *o,
bigval_t *hdr = bigval_header(o);
if (mark_mode == GC_OLD_MARKED) {
ptls->gc_tls.gc_cache.perm_scanned_bytes += hdr->sz;
gc_queue_big_marked(ptls, hdr, 0);
}
else {
ptls->gc_tls.gc_cache.scanned_bytes += hdr->sz;
// We can't easily tell if the object is old or being promoted
// from the gc bits but if the `age` is `0` then the object
// must be already on a young list.
if (mark_reset_age) {
assert(jl_atomic_load(&gc_n_threads_marking) == 0); // `mark_reset_age` is only used during single-threaded marking
// Reset the object as if it was just allocated
gc_queue_big_marked(ptls, hdr, 1);
gc_big_object_unlink(hdr);
gc_big_object_link(ptls->gc_tls.heap.young_generation_of_bigvals, hdr);
}
}
}
Expand Down Expand Up @@ -1023,7 +987,7 @@ STATIC_INLINE jl_value_t *jl_gc_big_alloc_inner(jl_ptls_t ptls, size_t sz)
memset(v, 0xee, allocsz);
#endif
v->sz = allocsz;
gc_big_object_link(v, &ptls->gc_tls.heap.big_objects);
gc_big_object_link(ptls->gc_tls.heap.young_generation_of_bigvals, v);
return jl_valueof(&v->header);
}

Expand All @@ -1043,60 +1007,85 @@ jl_value_t *jl_gc_big_alloc_noinline(jl_ptls_t ptls, size_t sz) {
return jl_gc_big_alloc_inner(ptls, sz);
}

// Sweep list rooted at *pv, removing and freeing any unmarked objects.
// Return pointer to last `next` field in the culled list.
static bigval_t **sweep_big_list(int sweep_full, bigval_t **pv) JL_NOTSAFEPOINT
FORCE_INLINE void sweep_unlink_and_free(bigval_t *v) JL_NOTSAFEPOINT
{
gc_big_object_unlink(v);
gc_num.freed += v->sz;
#ifdef MEMDEBUG
memset(v, 0xbb, v->sz);
#endif
gc_invoke_callbacks(jl_gc_cb_notify_external_free_t, gc_cblist_notify_external_free, (v));
jl_free_aligned(v);
}

static bigval_t *sweep_list_of_young_bigvals(bigval_t *young) JL_NOTSAFEPOINT
{
bigval_t *v = *pv;
bigval_t *last_node = young;
bigval_t *v = young->next; // skip the sentinel
bigval_t *old = oldest_generation_of_bigvals;
int sweep_full = current_sweep_full; // don't load the global in the hot loop
while (v != NULL) {
bigval_t *nxt = v->next;
int bits = v->bits.gc;
int old_bits = bits;
if (gc_marked(bits)) {
pv = &v->next;
if (sweep_full || bits == GC_MARKED) {
bits = GC_OLD;
last_node = v;
}
else { // `bits == GC_OLD_MARKED`
assert(bits == GC_OLD_MARKED);
// reached oldest generation, move from young list to old list
gc_big_object_unlink(v);
gc_big_object_link(old, v);
}
v->bits.gc = bits;
}
else {
// Remove v from list and free it
*pv = nxt;
if (nxt)
nxt->prev = pv;
gc_num.freed += v->sz;
#ifdef MEMDEBUG
memset(v, 0xbb, v->sz);
#endif
gc_invoke_callbacks(jl_gc_cb_notify_external_free_t,
gc_cblist_notify_external_free, (v));
jl_free_aligned(v);
sweep_unlink_and_free(v);
}
gc_time_count_big(old_bits, bits);
v = nxt;
}
return pv;
return last_node;
}

static void sweep_list_of_oldest_bigvals(bigval_t *young) JL_NOTSAFEPOINT
{
bigval_t *v = oldest_generation_of_bigvals->next; // skip the sentinel
while (v != NULL) {
bigval_t *nxt = v->next;
assert(v->bits.gc == GC_OLD_MARKED);
v->bits.gc = GC_OLD;
gc_time_count_big(GC_OLD_MARKED, GC_OLD);
v = nxt;
}
}

static void sweep_big(jl_ptls_t ptls, int sweep_full) JL_NOTSAFEPOINT
static void sweep_big(jl_ptls_t ptls) JL_NOTSAFEPOINT
{
gc_time_big_start();
assert(gc_n_threads);
bigval_t *last_node_in_my_list = NULL;
for (int i = 0; i < gc_n_threads; i++) {
jl_ptls_t ptls2 = gc_all_tls_states[i];
if (ptls2 != NULL)
sweep_big_list(sweep_full, &ptls2->gc_tls.heap.big_objects);
if (ptls2 != NULL) {
bigval_t *last_node = sweep_list_of_young_bigvals(ptls2->gc_tls.heap.young_generation_of_bigvals);
if (ptls == ptls2) {
last_node_in_my_list = last_node;
}
}
}
if (sweep_full) {
bigval_t **last_next = sweep_big_list(sweep_full, &big_objects_marked);
// Move all survivors from big_objects_marked list to the big_objects list of this thread.
if (ptls->gc_tls.heap.big_objects)
ptls->gc_tls.heap.big_objects->prev = last_next;
*last_next = ptls->gc_tls.heap.big_objects;
ptls->gc_tls.heap.big_objects = big_objects_marked;
if (ptls->gc_tls.heap.big_objects)
ptls->gc_tls.heap.big_objects->prev = &ptls->gc_tls.heap.big_objects;
big_objects_marked = NULL;
if (current_sweep_full) {
sweep_list_of_oldest_bigvals(ptls->gc_tls.heap.young_generation_of_bigvals);
// move all nodes in `oldest_generation_of_bigvals` to my list of bigvals
assert(last_node_in_my_list != NULL);
assert(last_node_in_my_list->next == NULL);
last_node_in_my_list->next = oldest_generation_of_bigvals->next; // skip the sentinel
if (oldest_generation_of_bigvals->next != NULL) {
oldest_generation_of_bigvals->next->prev = last_node_in_my_list;
}
oldest_generation_of_bigvals->next = NULL;
}
gc_time_big_end();
}
Expand Down Expand Up @@ -1536,7 +1525,7 @@ STATIC_INLINE void gc_sweep_pool_page(gc_page_profiler_serializer_t *s, jl_gc_pa
static void gc_sweep_other(jl_ptls_t ptls, int sweep_full) JL_NOTSAFEPOINT
{
sweep_malloced_arrays();
sweep_big(ptls, sweep_full);
sweep_big(ptls);
}

static void gc_pool_sync_nfree(jl_gc_pagemeta_t *pg, jl_taggedvalue_t *last) JL_NOTSAFEPOINT
Expand Down Expand Up @@ -3524,7 +3513,7 @@ static int _jl_gc_collect(jl_ptls_t ptls, jl_gc_collection_t collection)
// marking is over

// Flush everything in mark cache
gc_sync_all_caches_nolock(ptls);
gc_sync_all_caches(ptls);

int64_t live_sz_ub = live_bytes + actual_allocd;
int64_t live_sz_est = scanned_bytes + perm_scanned_bytes;
Expand Down Expand Up @@ -3863,15 +3852,16 @@ void jl_init_thread_heap(jl_ptls_t ptls)
small_arraylist_new(&heap->free_stacks[i], 0);
heap->mallocarrays = NULL;
heap->mafreelist = NULL;
heap->big_objects = NULL;
heap->young_generation_of_bigvals = (bigval_t*)calloc_s(sizeof(bigval_t)); // sentinel
assert(gc_bigval_sentinel_tag != 0); // make sure the sentinel is initialized
heap->young_generation_of_bigvals->header = gc_bigval_sentinel_tag;
arraylist_new(&heap->remset, 0);
arraylist_new(&ptls->finalizers, 0);
arraylist_new(&ptls->gc_tls.sweep_objs, 0);

jl_gc_mark_cache_t *gc_cache = &ptls->gc_tls.gc_cache;
gc_cache->perm_scanned_bytes = 0;
gc_cache->scanned_bytes = 0;
gc_cache->nbig_obj = 0;

// Initialize GC mark-queue
jl_gc_markqueue_t *mq = &ptls->gc_tls.mark_queue;
Expand All @@ -3897,12 +3887,16 @@ void jl_gc_init(void)
JL_MUTEX_INIT(&heapsnapshot_lock, "heapsnapshot_lock");
JL_MUTEX_INIT(&finalizers_lock, "finalizers_lock");
uv_mutex_init(&page_profile_lock);
uv_mutex_init(&gc_cache_lock);
uv_mutex_init(&gc_perm_lock);
uv_mutex_init(&gc_threads_lock);
uv_cond_init(&gc_threads_cond);
uv_sem_init(&gc_sweep_assists_needed, 0);
uv_mutex_init(&gc_queue_observer_lock);
void *_addr = (void*)calloc_s(1); // dummy allocation to get the sentinel tag
uintptr_t addr = (uintptr_t)_addr;
gc_bigval_sentinel_tag = addr;
oldest_generation_of_bigvals = (bigval_t*)calloc_s(sizeof(bigval_t)); // sentinel
oldest_generation_of_bigvals->header = gc_bigval_sentinel_tag;

jl_gc_init_page();
jl_gc_debug_init();
Expand Down
Loading

0 comments on commit 1f66e0d

Please sign in to comment.