Run GC on multiple threads (JuliaLang#48600)

Author: d-netto

NEWS.md

@@ -38,6 +38,7 @@ Compiler/Runtime improvements
 * All uses of the `@pure` macro in `Base` have been replaced with the now-preferred `Base.@assume_effects` ([#44776]).
 * `invoke(f, invokesig, args...)` calls to a less-specific method than would normally be chosen
   for `f(args...)` are no longer spuriously invalidated when loading package precompile files ([#46010]).
+* The mark phase of the Garbage Collector is now multi-threaded ([#48600]).
 
 Command-line option changes
 ---------------------------
@@ -49,6 +50,8 @@ Command-line option changes
   number of interactive threads to create (`auto` currently means 1) ([#42302]).
 * New option `--heap-size-hint=<size>` suggests a size limit to invoke garbage collection more eagerly.
   The size may be specified in bytes, kilobytes (1000k), megabytes (300M), or gigabytes (1.5G) ([#45369]).
+* New option `--gcthreads` to set how many threads will be used by the Garbage Collector ([#48600]).
+  The default is set to `N/2` where `N` is the amount of worker threads (`--threads`) used by Julia.
 
 Multi-threading changes
 -----------------------

base/options.jl

@@ -11,6 +11,7 @@ struct JLOptions
     cpu_target::Ptr{UInt8}
     nthreadpools::Int16
     nthreads::Int16
+    ngcthreads::Int16
     nthreads_per_pool::Ptr{Int16}
     nprocs::Int32
     machine_file::Ptr{UInt8}

base/threadingconstructs.jl

@@ -136,6 +136,13 @@ function threadpooltids(pool::Symbol)
     end
 end
 
+"""
+    Threads.ngcthreads() -> Int
+
+Returns the number of GC threads currently configured.
+"""
+ngcthreads() = Int(unsafe_load(cglobal(:jl_n_gcthreads, Cint))) + 1
+
 function threading_run(fun, static)
     ccall(:jl_enter_threaded_region, Cvoid, ())
     n = threadpoolsize()

doc/man/julia.1

@@ -118,6 +118,11 @@ supported (Linux and Windows). If this is not supported (macOS) or
 process affinity is not configured, it uses the number of CPU
 threads.
 
+.TP
+--gcthreads <n>
+Enable n GC threads; If unspecified is set to half of the
+compute worker threads.
+
 .TP
 -p, --procs {N|auto}
 Integer value N launches N additional local worker processes `auto` launches as many workers

doc/src/base/multi-threading.md

@@ -10,6 +10,7 @@ Base.Threads.nthreads
 Base.Threads.threadpool
 Base.Threads.nthreadpools
 Base.Threads.threadpoolsize
+Base.Threads.ngcthreads
 ```
 
 See also [Multi-Threading](@ref man-multithreading).

doc/src/manual/command-line-interface.md

@@ -106,8 +106,9 @@ The following is a complete list of command-line switches available when launchi
 |`-e`, `--eval <expr>`                  |Evaluate `<expr>`|
 |`-E`, `--print <expr>`                 |Evaluate `<expr>` and display the result|
 |`-L`, `--load <file>`                  |Load `<file>` immediately on all processors|
-|`-t`, `--threads {N\|auto`}            |Enable N threads; `auto` tries to infer a useful default number of threads to use but the exact behavior might change in the future.  Currently, `auto` uses the number of CPUs assigned to this julia process based on the OS-specific affinity assignment interface, if supported (Linux and Windows). If this is not supported (macOS) or process affinity is not configured, it uses the number of CPU threads.|
-|`-p`, `--procs {N\|auto`}              |Integer value N launches N additional local worker processes; `auto` launches as many workers as the number of local CPU threads (logical cores)|
+|`-t`, `--threads {N\|auto}`            |Enable N threads; `auto` tries to infer a useful default number of threads to use but the exact behavior might change in the future.  Currently, `auto` uses the number of CPUs assigned to this julia process based on the OS-specific affinity assignment interface, if supported (Linux and Windows). If this is not supported (macOS) or process affinity is not configured, it uses the number of CPU threads.|
+| `--gcthreads {N}`                     |Enable N GC threads; If unspecified is set to half of the compute worker threads.|
+|`-p`, `--procs {N\|auto}`              |Integer value N launches N additional local worker processes; `auto` launches as many workers as the number of local CPU threads (logical cores)|
 |`--machine-file <file>`                |Run processes on hosts listed in `<file>`|
 |`-i`                                   |Interactive mode; REPL runs and `isinteractive()` is true|
 |`-q`, `--quiet`                        |Quiet startup: no banner, suppress REPL warnings|

doc/src/manual/environment-variables.md

@@ -315,6 +315,14 @@ then spinning threads never sleep. Otherwise, `$JULIA_THREAD_SLEEP_THRESHOLD` is
 interpreted as an unsigned 64-bit integer (`uint64_t`) and gives, in
 nanoseconds, the amount of time after which spinning threads should sleep.
 
+### [`JULIA_NUM_GC_THREADS`](@id env-gc-threads)
+
+Sets the number of threads used by Garbage Collection. If unspecified is set to
+half of the number of worker threads.
+
+!!! compat "Julia 1.10"
+    The environment variable was added in 1.10
+
 ### `JULIA_EXCLUSIVE`
 
 If set to anything besides `0`, then Julia's thread policy is consistent with

doc/src/manual/multi-threading.md

@@ -72,6 +72,15 @@ julia> Threads.threadid()
     three processes have 2 threads enabled. For more fine grained control over worker
     threads use [`addprocs`](@ref) and pass `-t`/`--threads` as `exeflags`.
 
+### Multiple GC Threads
+
+The Garbage Collector (GC) can use multiple threads. The amount used is either half the number
+of compute worker threads or configured by either the `--gcthreads` command line argument or by using the
+[`JULIA_NUM_GC_THREADS`](@ref env-gc-threads) environment variable.
+
+!!! compat "Julia 1.10"
+    The `--gcthreads` command line argument requires at least Julia 1.10.
+
 ## [Threadpools](@id man-threadpools)
 
 When a program's threads are busy with many tasks to run, tasks may experience

src/Makefile

@@ -99,7 +99,7 @@ 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)/,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 $(SRCDIR)/,work-stealing-queue.h julia.h julia_assert.h julia_threads.h julia_fasttls.h julia_locks.h julia_atomics.h jloptions.h)
 ifeq ($(OS),WINNT)
 PUBLIC_HEADERS += $(addprefix $(SRCDIR)/,win32_ucontext.h)
 endif

src/gc-debug.c

@@ -198,21 +198,32 @@ static void restore(void)
 
 static void gc_verify_track(jl_ptls_t ptls)
 {
-    jl_gc_mark_cache_t *gc_cache = &ptls->gc_cache;
+    // `gc_verify_track` is limited to single-threaded GC
+    if (jl_n_gcthreads != 0)
+        return;
     do {
-        jl_gc_mark_sp_t sp;
-        gc_mark_sp_init(gc_cache, &sp);
+        jl_gc_markqueue_t mq;
+        jl_gc_markqueue_t *mq2 = &ptls->mark_queue;
+        ws_queue_t *cq = &mq.chunk_queue;
+        ws_queue_t *q = &mq.ptr_queue;
+        jl_atomic_store_relaxed(&cq->top, 0);
+        jl_atomic_store_relaxed(&cq->bottom, 0);
+        jl_atomic_store_relaxed(&cq->array, jl_atomic_load_relaxed(&mq2->chunk_queue.array));
+        jl_atomic_store_relaxed(&q->top, 0);
+        jl_atomic_store_relaxed(&q->bottom, 0);
+        jl_atomic_store_relaxed(&q->array, jl_atomic_load_relaxed(&mq2->ptr_queue.array));
+        arraylist_new(&mq.reclaim_set, 32);
         arraylist_push(&lostval_parents_done, lostval);
         jl_safe_printf("Now looking for %p =======\n", lostval);
         clear_mark(GC_CLEAN);
-        gc_mark_queue_all_roots(ptls, &sp);
-        gc_mark_queue_finlist(gc_cache, &sp, &to_finalize, 0);
-        for (int i = 0; i < gc_n_threads; i++) {
+        gc_mark_queue_all_roots(ptls, &mq);
+        gc_mark_finlist(&mq, &to_finalize, 0);
+        for (int i = 0; i < gc_n_threads;i++) {
             jl_ptls_t ptls2 = gc_all_tls_states[i];
-            gc_mark_queue_finlist(gc_cache, &sp, &ptls2->finalizers, 0);
+            gc_mark_finlist(&mq, &ptls2->finalizers, 0);
         }
-        gc_mark_queue_finlist(gc_cache, &sp, &finalizer_list_marked, 0);
-        gc_mark_loop(ptls, sp);
+        gc_mark_finlist(&mq, &finalizer_list_marked, 0);
+        gc_mark_loop_serial_(ptls, &mq);
         if (lostval_parents.len == 0) {
             jl_safe_printf("Could not find the missing link. We missed a toplevel root. This is odd.\n");
             break;
@@ -246,22 +257,35 @@ static void gc_verify_track(jl_ptls_t ptls)
 
 void gc_verify(jl_ptls_t ptls)
 {
-    jl_gc_mark_cache_t *gc_cache = &ptls->gc_cache;
-    jl_gc_mark_sp_t sp;
-    gc_mark_sp_init(gc_cache, &sp);
+    // `gc_verify` is limited to single-threaded GC
+    if (jl_n_gcthreads != 0) {
+        jl_safe_printf("Warn. GC verify disabled in multi-threaded GC\n");
+        return;
+    }
+    jl_gc_markqueue_t mq;
+    jl_gc_markqueue_t *mq2 = &ptls->mark_queue;
+    ws_queue_t *cq = &mq.chunk_queue;
+    ws_queue_t *q = &mq.ptr_queue;
+    jl_atomic_store_relaxed(&cq->top, 0);
+    jl_atomic_store_relaxed(&cq->bottom, 0);
+    jl_atomic_store_relaxed(&cq->array, jl_atomic_load_relaxed(&mq2->chunk_queue.array));
+    jl_atomic_store_relaxed(&q->top, 0);
+    jl_atomic_store_relaxed(&q->bottom, 0);
+    jl_atomic_store_relaxed(&q->array, jl_atomic_load_relaxed(&mq2->ptr_queue.array));
+    arraylist_new(&mq.reclaim_set, 32);
     lostval = NULL;
     lostval_parents.len = 0;
     lostval_parents_done.len = 0;
     clear_mark(GC_CLEAN);
     gc_verifying = 1;
-    gc_mark_queue_all_roots(ptls, &sp);
-    gc_mark_queue_finlist(gc_cache, &sp, &to_finalize, 0);
-    for (int i = 0; i < gc_n_threads; i++) {
+    gc_mark_queue_all_roots(ptls, &mq);
+    gc_mark_finlist(&mq, &to_finalize, 0);
+    for (int i = 0; i < gc_n_threads;i++) {
         jl_ptls_t ptls2 = gc_all_tls_states[i];
-        gc_mark_queue_finlist(gc_cache, &sp, &ptls2->finalizers, 0);
+        gc_mark_finlist(&mq, &ptls2->finalizers, 0);
     }
-    gc_mark_queue_finlist(gc_cache, &sp, &finalizer_list_marked, 0);
-    gc_mark_loop(ptls, sp);
+    gc_mark_finlist(&mq, &finalizer_list_marked, 0);
+    gc_mark_loop_serial_(ptls, &mq);
     int clean_len = bits_save[GC_CLEAN].len;
     for(int i = 0; i < clean_len + bits_save[GC_OLD].len; i++) {
         jl_taggedvalue_t *v = (jl_taggedvalue_t*)bits_save[i >= clean_len ? GC_OLD : GC_CLEAN].items[i >= clean_len ? i - clean_len : i];
@@ -500,7 +524,7 @@ int jl_gc_debug_check_other(void)
     return gc_debug_alloc_check(&jl_gc_debug_env.other);
 }
 
-void jl_gc_debug_print_status(void)
+void jl_gc_debug_print_status(void) JL_NOTSAFEPOINT
 {
     uint64_t pool_count = jl_gc_debug_env.pool.num;
     uint64_t other_count = jl_gc_debug_env.other.num;
@@ -509,7 +533,7 @@ void jl_gc_debug_print_status(void)
                    pool_count + other_count, pool_count, other_count, gc_num.pause);
 }
 
-void jl_gc_debug_critical_error(void)
+void jl_gc_debug_critical_error(void) JL_NOTSAFEPOINT
 {
     jl_gc_debug_print_status();
     if (!jl_gc_debug_env.wait_for_debugger)
@@ -1264,139 +1288,6 @@ int gc_slot_to_arrayidx(void *obj, void *_slot) JL_NOTSAFEPOINT
     return (slot - start) / elsize;
 }
 
-// Print a backtrace from the bottom (start) of the mark stack up to `sp`
-// `pc_offset` will be added to `sp` for convenience in the debugger.
-NOINLINE void gc_mark_loop_unwind(jl_ptls_t ptls, jl_gc_mark_sp_t sp, int pc_offset)
-{
-    jl_jmp_buf *old_buf = jl_get_safe_restore();
-    jl_jmp_buf buf;
-    jl_set_safe_restore(&buf);
-    if (jl_setjmp(buf, 0) != 0) {
-        jl_safe_printf("\n!!! ERROR when unwinding gc mark loop -- ABORTING !!!\n");
-        jl_set_safe_restore(old_buf);
-        return;
-    }
-    void **top = sp.pc + pc_offset;
-    jl_gc_mark_data_t *data_top = sp.data;
-    sp.data = ptls->gc_cache.data_stack;
-    sp.pc = ptls->gc_cache.pc_stack;
-    int isroot = 1;
-    while (sp.pc < top) {
-        void *pc = *sp.pc;
-        const char *prefix = isroot ? "r--" : " `-";
-        isroot = 0;
-        if (pc == gc_mark_label_addrs[GC_MARK_L_marked_obj]) {
-            gc_mark_marked_obj_t *data = gc_repush_markdata(&sp, gc_mark_marked_obj_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_safe_printf("%p: Root object: %p :: %p (bits: %d)\n        of type ",
-                           (void*)data, (void*)data->obj, (void*)data->tag, (int)data->bits);
-            jl_((void*)data->tag);
-            isroot = 1;
-        }
-        else if (pc == gc_mark_label_addrs[GC_MARK_L_scan_only]) {
-            gc_mark_marked_obj_t *data = gc_repush_markdata(&sp, gc_mark_marked_obj_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_safe_printf("%p: Queued root: %p :: %p (bits: %d)\n        of type ",
-                           (void*)data, (void*)data->obj, (void*)data->tag, (int)data->bits);
-            jl_((void*)data->tag);
-            isroot = 1;
-        }
-        else if (pc == gc_mark_label_addrs[GC_MARK_L_finlist]) {
-            gc_mark_finlist_t *data = gc_repush_markdata(&sp, gc_mark_finlist_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_safe_printf("%p: Finalizer list from %p to %p\n",
-                           (void*)data, (void*)data->begin, (void*)data->end);
-            isroot = 1;
-        }
-        else if (pc == gc_mark_label_addrs[GC_MARK_L_objarray]) {
-            gc_mark_objarray_t *data = gc_repush_markdata(&sp, gc_mark_objarray_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_safe_printf("%p:  %s Array in object %p :: %p -- [%p, %p)\n        of type ",
-                           (void*)data, prefix, (void*)data->parent, ((void**)data->parent)[-1],
-                           (void*)data->begin, (void*)data->end);
-            jl_(jl_typeof(data->parent));
-        }
-        else if (pc == gc_mark_label_addrs[GC_MARK_L_obj8]) {
-            gc_mark_obj8_t *data = gc_repush_markdata(&sp, gc_mark_obj8_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_datatype_t *vt = (jl_datatype_t*)jl_typeof(data->parent);
-            uint8_t *desc = (uint8_t*)jl_dt_layout_ptrs(vt->layout);
-            jl_safe_printf("%p:  %s Object (8bit) %p :: %p -- [%d, %d)\n        of type ",
-                           (void*)data, prefix, (void*)data->parent, ((void**)data->parent)[-1],
-                           (int)(data->begin - desc), (int)(data->end - desc));
-            jl_(jl_typeof(data->parent));
-        }
-        else if (pc == gc_mark_label_addrs[GC_MARK_L_obj16]) {
-            gc_mark_obj16_t *data = gc_repush_markdata(&sp, gc_mark_obj16_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_datatype_t *vt = (jl_datatype_t*)jl_typeof(data->parent);
-            uint16_t *desc = (uint16_t*)jl_dt_layout_ptrs(vt->layout);
-            jl_safe_printf("%p:  %s Object (16bit) %p :: %p -- [%d, %d)\n        of type ",
-                           (void*)data, prefix, (void*)data->parent, ((void**)data->parent)[-1],
-                           (int)(data->begin - desc), (int)(data->end - desc));
-            jl_(jl_typeof(data->parent));
-        }
-        else if (pc == gc_mark_label_addrs[GC_MARK_L_obj32]) {
-            gc_mark_obj32_t *data = gc_repush_markdata(&sp, gc_mark_obj32_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_datatype_t *vt = (jl_datatype_t*)jl_typeof(data->parent);
-            uint32_t *desc = (uint32_t*)jl_dt_layout_ptrs(vt->layout);
-            jl_safe_printf("%p:  %s Object (32bit) %p :: %p -- [%d, %d)\n        of type ",
-                           (void*)data, prefix, (void*)data->parent, ((void**)data->parent)[-1],
-                           (int)(data->begin - desc), (int)(data->end - desc));
-            jl_(jl_typeof(data->parent));
-        }
-        else if (pc == gc_mark_label_addrs[GC_MARK_L_stack]) {
-            gc_mark_stackframe_t *data = gc_repush_markdata(&sp, gc_mark_stackframe_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_safe_printf("%p:  %s Stack frame %p -- %d of %d (%s)\n",
-                           (void*)data, prefix, (void*)data->s, (int)data->i,
-                           (int)data->nroots >> 1,
-                           (data->nroots & 1) ? "indirect" : "direct");
-        }
-        else if (pc == gc_mark_label_addrs[GC_MARK_L_module_binding]) {
-            // module_binding
-            gc_mark_binding_t *data = gc_repush_markdata(&sp, gc_mark_binding_t);
-            if ((jl_gc_mark_data_t *)data > data_top) {
-                jl_safe_printf("Mark stack unwind overflow -- ABORTING !!!\n");
-                break;
-            }
-            jl_safe_printf("%p:  %s Module (bindings) %p (bits %d) -- [%p, %p)\n",
-                           (void*)data, prefix, (void*)data->parent, (int)data->bits,
-                           (void*)data->begin, (void*)data->end);
-        }
-        else {
-            jl_safe_printf("Unknown pc %p --- ABORTING !!!\n", pc);
-            break;
-        }
-    }
-    jl_set_safe_restore(old_buf);
-}
-
 static int gc_logging_enabled = 0;
 
 JL_DLLEXPORT void jl_enable_gc_logging(int enable) {