create GC TLS (#55086)

Encapsulates all relevant GC thread-local-state into a separate
structure.

Motivation is that MMTk will have its own version of GC
thread-local-state, so doesn't need all of the Julia GC TLS.

In the future, folks who would be using MMTk would be setting a
pre-processor flag which would lead to either the stock Julia GC TLS or
MMTk's GC TLS to be included in `julia_threads.h`.

I.e., we would have something like:


```C
#ifdef MMTK_GC
jl_gc_mmtk_tls_states mmtk_gc_tls;
#else
jl_gc_tls_states gc_tls;
#endif
```

Author: d-netto

src/Makefile

@@ -103,7 +103,7 @@ ifeq ($(USE_SYSTEM_LIBUV),0)
 UV_HEADERS += uv.h
 UV_HEADERS += uv/*.h
 endif
-PUBLIC_HEADERS := $(BUILDDIR)/julia_version.h $(wildcard $(SRCDIR)/support/*.h) $(addprefix $(SRCDIR)/,work-stealing-queue.h julia.h julia_assert.h julia_threads.h julia_fasttls.h julia_locks.h julia_atomics.h jloptions.h)
+PUBLIC_HEADERS := $(BUILDDIR)/julia_version.h $(wildcard $(SRCDIR)/support/*.h) $(addprefix $(SRCDIR)/,work-stealing-queue.h gc-tls.h julia.h julia_assert.h julia_threads.h julia_fasttls.h julia_locks.h julia_atomics.h jloptions.h)
 ifeq ($(OS),WINNT)
 PUBLIC_HEADERS += $(addprefix $(SRCDIR)/,win32_ucontext.h)
 endif

src/array.c

@@ -307,7 +307,7 @@ JL_DLLEXPORT jl_value_t *jl_alloc_string(size_t len)
     const size_t allocsz = sz + sizeof(jl_taggedvalue_t);
     if (sz <= GC_MAX_SZCLASS) {
         int pool_id = jl_gc_szclass_align8(allocsz);
-        jl_gc_pool_t *p = &ptls->heap.norm_pools[pool_id];
+        jl_gc_pool_t *p = &ptls->gc_tls.heap.norm_pools[pool_id];
         int osize = jl_gc_sizeclasses[pool_id];
         // We call `jl_gc_pool_alloc_noinline` instead of `jl_gc_pool_alloc` to avoid double-counting in
         // the Allocations Profiler. (See https://github.com/JuliaLang/julia/pull/43868 for more details.)

src/gc-debug.c

@@ -99,7 +99,7 @@ static arraylist_t bits_save[4];
 static void gc_clear_mark_page(jl_gc_pagemeta_t *pg, int bits)
 {
     jl_ptls_t ptls2 = gc_all_tls_states[pg->thread_n];
-    jl_gc_pool_t *pool = &ptls2->heap.norm_pools[pg->pool_n];
+    jl_gc_pool_t *pool = &ptls2->gc_tls.heap.norm_pools[pg->pool_n];
     jl_taggedvalue_t *pv = (jl_taggedvalue_t*)(pg->data + GC_PAGE_OFFSET);
     char *lim = (char*)pv + GC_PAGE_SZ - GC_PAGE_OFFSET - pool->osize;
     while ((char*)pv <= lim) {
@@ -114,7 +114,7 @@ static void gc_clear_mark_outer(int bits)
 {
     for (int i = 0; i < gc_n_threads; i++) {
         jl_ptls_t ptls2 = gc_all_tls_states[i];
-        jl_gc_pagemeta_t *pg = jl_atomic_load_relaxed(&ptls2->page_metadata_allocd.bottom);
+        jl_gc_pagemeta_t *pg = jl_atomic_load_relaxed(&ptls2->gc_tls.page_metadata_allocd.bottom);
         while (pg != NULL) {
             gc_clear_mark_page(pg, bits);
             pg = pg->next;
@@ -134,7 +134,7 @@ static void clear_mark(int bits)
     }
     bigval_t *v;
     for (int i = 0; i < gc_n_threads; i++) {
-        v = gc_all_tls_states[i]->heap.big_objects;
+        v = gc_all_tls_states[i]->gc_tls.heap.big_objects;
         while (v != NULL) {
             void *gcv = &v->header;
             if (!gc_verifying)
@@ -172,7 +172,7 @@ static void gc_verify_track(jl_ptls_t ptls)
         return;
     do {
         jl_gc_markqueue_t mq;
-        jl_gc_markqueue_t *mq2 = &ptls->mark_queue;
+        jl_gc_markqueue_t *mq2 = &ptls->gc_tls.mark_queue;
         ws_queue_t *cq = &mq.chunk_queue;
         ws_queue_t *q = &mq.ptr_queue;
         jl_atomic_store_relaxed(&cq->top, 0);
@@ -232,7 +232,7 @@ void gc_verify(jl_ptls_t ptls)
         return;
     }
     jl_gc_markqueue_t mq;
-    jl_gc_markqueue_t *mq2 = &ptls->mark_queue;
+    jl_gc_markqueue_t *mq2 = &ptls->gc_tls.mark_queue;
     ws_queue_t *cq = &mq.chunk_queue;
     ws_queue_t *q = &mq.ptr_queue;
     jl_atomic_store_relaxed(&cq->top, 0);
@@ -291,7 +291,7 @@ static void gc_verify_tags_page(jl_gc_pagemeta_t *pg)
     int p_n = pg->pool_n;
     int t_n = pg->thread_n;
     jl_ptls_t ptls2 = gc_all_tls_states[t_n];
-    jl_gc_pool_t *p = &ptls2->heap.norm_pools[p_n];
+    jl_gc_pool_t *p = &ptls2->gc_tls.heap.norm_pools[p_n];
     int osize = pg->osize;
     char *data = pg->data;
     char *page_begin = data + GC_PAGE_OFFSET;
@@ -353,7 +353,7 @@ static void gc_verify_tags_pagestack(void)
 {
     for (int i = 0; i < gc_n_threads; i++) {
         jl_ptls_t ptls2 = gc_all_tls_states[i];
-        jl_gc_page_stack_t *pgstk = &ptls2->page_metadata_allocd;
+        jl_gc_page_stack_t *pgstk = &ptls2->gc_tls.page_metadata_allocd;
         jl_gc_pagemeta_t *pg = jl_atomic_load_relaxed(&pgstk->bottom);
         while (pg != NULL) {
             gc_verify_tags_page(pg);
@@ -369,7 +369,7 @@ void gc_verify_tags(void)
         jl_ptls_t ptls2 = gc_all_tls_states[t_i];
         for (int i = 0; i < JL_GC_N_POOLS; i++) {
             // for all pools, iterate its freelist
-            jl_gc_pool_t *p = &ptls2->heap.norm_pools[i];
+            jl_gc_pool_t *p = &ptls2->gc_tls.heap.norm_pools[i];
             jl_taggedvalue_t *next = p->freelist;
             jl_taggedvalue_t *last = NULL;
             char *allocating = gc_page_data(next);
@@ -811,8 +811,8 @@ void gc_time_mark_pause(int64_t t0, int64_t scanned_bytes,
     int64_t remset_nptr = 0;
     for (int t_i = 0; t_i < gc_n_threads; t_i++) {
         jl_ptls_t ptls2 = gc_all_tls_states[t_i];
-        last_remset_len += ptls2->heap.last_remset->len;
-        remset_nptr = ptls2->heap.remset_nptr;
+        last_remset_len += ptls2->gc_tls.heap.last_remset->len;
+        remset_nptr = ptls2->gc_tls.heap.remset_nptr;
     }
     jl_safe_printf("GC mark pause %.2f ms | "
                    "scanned %" PRId64 " kB = %" PRId64 " + %" PRId64 " | "
@@ -967,13 +967,13 @@ void gc_stats_all_pool(void)
     for (int i = 0; i < JL_GC_N_POOLS; i++) {
         for (int t_i = 0; t_i < gc_n_threads; t_i++) {
             jl_ptls_t ptls2 = gc_all_tls_states[t_i];
-            size_t b = pool_stats(&ptls2->heap.norm_pools[i], &w, &np, &nol);
+            size_t b = pool_stats(&ptls2->gc_tls.heap.norm_pools[i], &w, &np, &nol);
             nb += b;
-            no += (b / ptls2->heap.norm_pools[i].osize);
+            no += (b / ptls2->gc_tls.heap.norm_pools[i].osize);
             tw += w;
             tp += np;
             nold += nol;
-            noldbytes += nol * ptls2->heap.norm_pools[i].osize;
+            noldbytes += nol * ptls2->gc_tls.heap.norm_pools[i].osize;
         }
     }
     jl_safe_printf("%lld objects (%lld%% old), %lld kB (%lld%% old) total allocated, "
@@ -992,7 +992,7 @@ 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->heap.big_objects;
+        bigval_t *v = ptls2->gc_tls.heap.big_objects;
         while (v != NULL) {
             if (gc_marked(v->bits.gc)) {
                 nused++;
@@ -1009,7 +1009,7 @@ void gc_stats_big_obj(void)
             v = v->next;
         }
 
-        mallocarray_t *ma = ptls2->heap.mallocarrays;
+        mallocarray_t *ma = ptls2->gc_tls.heap.mallocarrays;
         while (ma != NULL) {
             if (gc_marked(jl_astaggedvalue(ma->a)->bits.gc)) {
                 nused++;
@@ -1055,7 +1055,7 @@ static void gc_count_pool_pagetable(void)
 {
     for (int i = 0; i < gc_n_threads; i++) {
         jl_ptls_t ptls2 = gc_all_tls_states[i];
-        jl_gc_pagemeta_t *pg = jl_atomic_load_relaxed(&ptls2->page_metadata_allocd.bottom);
+        jl_gc_pagemeta_t *pg = jl_atomic_load_relaxed(&ptls2->gc_tls.page_metadata_allocd.bottom);
         while (pg != NULL) {
             if (gc_alloc_map_is_set(pg->data)) {
                 gc_count_pool_page(pg);

src/gc-stacks.c

@@ -167,7 +167,7 @@ static void _jl_free_stack(jl_ptls_t ptls, void *stkbuf, size_t bufsz) JL_NOTSAF
     if (bufsz <= pool_sizes[JL_N_STACK_POOLS - 1]) {
         unsigned pool_id = select_pool(bufsz);
         if (pool_sizes[pool_id] == bufsz) {
-            small_arraylist_push(&ptls->heap.free_stacks[pool_id], stkbuf);
+            small_arraylist_push(&ptls->gc_tls.heap.free_stacks[pool_id], stkbuf);
             return;
         }
     }
@@ -196,7 +196,7 @@ void jl_release_task_stack(jl_ptls_t ptls, jl_task_t *task)
 #ifdef _COMPILER_ASAN_ENABLED_
             __asan_unpoison_stack_memory((uintptr_t)stkbuf, bufsz);
 #endif
-            small_arraylist_push(&ptls->heap.free_stacks[pool_id], stkbuf);
+            small_arraylist_push(&ptls->gc_tls.heap.free_stacks[pool_id], stkbuf);
         }
     }
 }
@@ -211,7 +211,7 @@ JL_DLLEXPORT void *jl_malloc_stack(size_t *bufsz, jl_task_t *owner) JL_NOTSAFEPO
     if (ssize <= pool_sizes[JL_N_STACK_POOLS - 1]) {
         unsigned pool_id = select_pool(ssize);
         ssize = pool_sizes[pool_id];
-        small_arraylist_t *pool = &ptls->heap.free_stacks[pool_id];
+        small_arraylist_t *pool = &ptls->gc_tls.heap.free_stacks[pool_id];
         if (pool->len > 0) {
             stk = small_arraylist_pop(pool);
         }
@@ -232,7 +232,7 @@ JL_DLLEXPORT void *jl_malloc_stack(size_t *bufsz, jl_task_t *owner) JL_NOTSAFEPO
     }
     *bufsz = ssize;
     if (owner) {
-        small_arraylist_t *live_tasks = &ptls->heap.live_tasks;
+        small_arraylist_t *live_tasks = &ptls->gc_tls.heap.live_tasks;
         mtarraylist_push(live_tasks, owner);
     }
     return stk;
@@ -259,7 +259,7 @@ void sweep_stack_pools(void) JL_NOTSAFEPOINT
 
         // free half of stacks that remain unused since last sweep
         for (int p = 0; p < JL_N_STACK_POOLS; p++) {
-            small_arraylist_t *al = &ptls2->heap.free_stacks[p];
+            small_arraylist_t *al = &ptls2->gc_tls.heap.free_stacks[p];
             size_t n_to_free;
             if (jl_atomic_load_relaxed(&ptls2->current_task) == NULL) {
                 n_to_free = al->len; // not alive yet or dead, so it does not need these anymore
@@ -281,10 +281,10 @@ void sweep_stack_pools(void) JL_NOTSAFEPOINT
             }
         }
         if (jl_atomic_load_relaxed(&ptls2->current_task) == NULL) {
-            small_arraylist_free(ptls2->heap.free_stacks);
+            small_arraylist_free(ptls2->gc_tls.heap.free_stacks);
         }
 
-        small_arraylist_t *live_tasks = &ptls2->heap.live_tasks;
+        small_arraylist_t *live_tasks = &ptls2->gc_tls.heap.live_tasks;
         size_t n = 0;
         size_t ndel = 0;
         size_t l = live_tasks->len;
@@ -339,7 +339,7 @@ JL_DLLEXPORT jl_array_t *jl_live_tasks(void)
         jl_ptls_t ptls2 = allstates[i];
         if (ptls2 == NULL)
             continue;
-        small_arraylist_t *live_tasks = &ptls2->heap.live_tasks;
+        small_arraylist_t *live_tasks = &ptls2->gc_tls.heap.live_tasks;
         size_t n = mtarraylist_length(live_tasks);
         l += n + (ptls2->root_task->stkbuf != NULL);
     }
@@ -362,7 +362,7 @@ JL_DLLEXPORT jl_array_t *jl_live_tasks(void)
                 goto restart;
             jl_array_data(a,void*)[j++] = t;
         }
-        small_arraylist_t *live_tasks = &ptls2->heap.live_tasks;
+        small_arraylist_t *live_tasks = &ptls2->gc_tls.heap.live_tasks;
         size_t n = mtarraylist_length(live_tasks);
         for (size_t i = 0; i < n; i++) {
             jl_task_t *t = (jl_task_t*)mtarraylist_get(live_tasks, i);

src/gc-tls.h

@@ -0,0 +1,103 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+// Meant to be included in "julia_threads.h"
+#ifndef JL_GC_TLS_H
+#define JL_GC_TLS_H
+
+#include "julia_atomics.h"
+#include "work-stealing-queue.h"
+// GC threading ------------------------------------------------------------------
+
+#include "arraylist.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+    struct _jl_taggedvalue_t *freelist; // root of list of free objects
+    struct _jl_taggedvalue_t *newpages; // root of list of chunks of free objects
+    uint16_t osize; // size of objects in this pool
+} jl_gc_pool_t;
+
+typedef struct {
+    // variable for tracking weak references
+    small_arraylist_t weak_refs;
+    // live tasks started on this thread
+    // that are holding onto a stack from the pool
+    small_arraylist_t live_tasks;
+
+    // variables for tracking malloc'd arrays
+    struct _mallocarray_t *mallocarrays;
+    struct _mallocarray_t *mafreelist;
+
+    // variables for tracking big objects
+    struct _bigval_t *big_objects;
+
+    // lower bound of the number of pointers inside remembered values
+    int remset_nptr;
+    // remembered set
+    arraylist_t remset;
+
+    // variables for allocating objects from pools
+#define JL_GC_N_MAX_POOLS 51 // conservative. must be kept in sync with `src/julia_internal.h`
+    jl_gc_pool_t norm_pools[JL_GC_N_MAX_POOLS];
+
+#define JL_N_STACK_POOLS 16
+    small_arraylist_t free_stacks[JL_N_STACK_POOLS];
+} jl_thread_heap_t;
+
+typedef struct {
+    _Atomic(int64_t) allocd;
+    _Atomic(int64_t) pool_live_bytes;
+    _Atomic(uint64_t) malloc;
+    _Atomic(uint64_t) realloc;
+    _Atomic(uint64_t) poolalloc;
+    _Atomic(uint64_t) bigalloc;
+    _Atomic(int64_t) free_acc;
+    _Atomic(uint64_t) alloc_acc;
+} jl_thread_gc_num_t;
+
+typedef struct {
+    ws_queue_t chunk_queue;
+    ws_queue_t ptr_queue;
+    arraylist_t reclaim_set;
+} jl_gc_markqueue_t;
+
+typedef struct {
+    // thread local increment of `perm_scanned_bytes`
+    size_t perm_scanned_bytes;
+    // thread local increment of `scanned_bytes`
+    size_t scanned_bytes;
+    // Number of queued big objects (<= 1024)
+    size_t nbig_obj;
+    // Array of queued big objects to be moved between the young list
+    // and the old list.
+    // A set low bit means that the object should be moved from the old list
+    // to the young list (`mark_reset_age`).
+    // Objects can only be put into this list when the mark bit is flipped to
+    // `1` (atomically). Combining with the sync after marking,
+    // this makes sure that a single objects can only appear once in
+    // the lists (the mark bit cannot be flipped to `0` without sweeping)
+    void *big_obj[1024];
+} jl_gc_mark_cache_t;
+
+typedef struct {
+    _Atomic(struct _jl_gc_pagemeta_t *) bottom;
+} jl_gc_page_stack_t;
+
+typedef struct {
+    jl_thread_heap_t heap;
+    jl_gc_page_stack_t page_metadata_allocd;
+    jl_thread_gc_num_t gc_num;
+    jl_gc_markqueue_t mark_queue;
+    jl_gc_mark_cache_t gc_cache;
+    _Atomic(size_t) gc_sweeps_requested;
+    arraylist_t sweep_objs;
+} jl_gc_tls_states_t;
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // JL_GC_TLS_H