free even more state for exited threads

Author: JeffBezanson

src/gc-stacks.c

@@ -203,6 +203,8 @@ void sweep_stack_pools(void)
     assert(gc_n_threads);
     for (int i = 0; i < gc_n_threads; i++) {
         jl_ptls_t ptls2 = gc_all_tls_states[i];
+        if (ptls2 == NULL)
+            continue;
 
         // free half of stacks that remain unused since last sweep
         for (int p = 0; p < JL_N_STACK_POOLS; p++) {
@@ -227,6 +229,9 @@ void sweep_stack_pools(void)
                 small_arraylist_free(al);
             }
         }
+        if (jl_atomic_load_relaxed(&ptls2->current_task) == NULL) {
+            small_arraylist_free(ptls2->heap.free_stacks);
+        }
 
         small_arraylist_t *live_tasks = &ptls2->heap.live_tasks;
         size_t n = 0;

src/gc.c

@@ -1746,6 +1746,19 @@ void gc_free_pages(void)
     }
 }
 
+void gc_move_to_global_page_pool(jl_gc_page_stack_t *pgstack)
+{
+    while (1) {
+        jl_gc_pagemeta_t *pg = pop_lf_back(pgstack);
+        if (pg == NULL) {
+            break;
+        }
+        jl_gc_free_page(pg);
+        jl_atomic_fetch_add_relaxed(&gc_heap_stats.heap_size, -GC_PAGE_SZ);
+        push_lf_back(&global_page_pool_freed, pg);
+    }
+}
+
 // setup the data-structures for a sweep over all memory pools
 static void gc_sweep_pool(void)
 {
@@ -1818,6 +1831,7 @@ static void gc_sweep_pool(void)
             }
         }
     }
+    free(tmp);
 
     // merge free lists
     for (int t_i = 0; t_i < n_threads; t_i++) {
@@ -1848,6 +1862,7 @@ static void gc_sweep_pool(void)
             }
         }
     }
+    free(pfl);
 
     // cleanup
     free(pfl);
@@ -3776,7 +3791,8 @@ static int _jl_gc_collect(jl_ptls_t ptls, jl_gc_collection_t collection)
             ptls2->heap.remset->len = 0;
         }
         // free empty GC state for threads that have exited
-        if (ptls2->current_task == NULL) {
+        if (jl_atomic_load_relaxed(&ptls2->current_task) == NULL &&
+            (ptls->tid < gc_first_tid || ptls2->tid >= gc_first_tid + jl_n_gcthreads)) {
             jl_thread_heap_t *heap = &ptls2->heap;
             if (heap->weak_refs.len == 0)
                 small_arraylist_free(&heap->weak_refs);
@@ -3787,9 +3803,10 @@ static int _jl_gc_collect(jl_ptls_t ptls, jl_gc_collection_t collection)
             if (heap->last_remset->len == 0)
                 arraylist_free(heap->last_remset);
             if (ptls2->finalizers.len == 0)
-                arraylist_free(&ptls->finalizers);
+                arraylist_free(&ptls2->finalizers);
             if (ptls2->sweep_objs.len == 0)
-                arraylist_free(&ptls->sweep_objs);
+                arraylist_free(&ptls2->sweep_objs);
+            gc_move_to_global_page_pool(&ptls2->page_metadata_buffered);
         }
     }
 

src/threading.c

@@ -435,6 +435,8 @@ JL_DLLEXPORT jl_gcframe_t **jl_adopt_thread(void)
 void jl_task_frame_noreturn(jl_task_t *ct) JL_NOTSAFEPOINT;
 void scheduler_delete_thread(jl_ptls_t ptls) JL_NOTSAFEPOINT;
 
+void jl_free_thread_gc_state(jl_ptls_t ptls);
+
 static void jl_delete_thread(void *value) JL_NOTSAFEPOINT_ENTER
 {
 #ifndef _OS_WINDOWS_
@@ -509,9 +511,10 @@ static void jl_delete_thread(void *value) JL_NOTSAFEPOINT_ENTER
     pthread_mutex_unlock(&in_signal_lock);
 #endif
     free(ptls->bt_data);
+    small_arraylist_free(&ptls->locks);
+    ptls->previous_exception = NULL;
     // allow the page root_task is on to be freed
     ptls->root_task = NULL;
-    void jl_free_thread_gc_state(jl_ptls_t ptls);
     jl_free_thread_gc_state(ptls);
     // then park in safe-region
     (void)jl_gc_safe_enter(ptls);