Run GC on multiple threads

NEWS.md

@@ -17,11 +17,15 @@ Language changes
 
 Compiler/Runtime improvements
 -----------------------------
+
 * The `@pure` macro is now deprecated. Use `Base.@assume_effects :foldable` instead ([#48682]).
+* The mark phase of the Garbage Collector is now multi-threaded ([#48600]).
 
 Command-line option changes
 ---------------------------
 
+* 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

@@ -99,6 +99,13 @@ 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

@@ -107,6 +107,7 @@ The following is a complete list of command-line switches available when launchi
 |`-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.|
+| `--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|

doc/src/manual/environment-variables.md

@@ -316,6 +316,14 @@ 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_IMAGE_THREADS`](@id env-image-threads)
 
 An unsigned 32-bit integer that sets the number of threads used by image

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

@@ -198,12 +198,21 @@ static void restore(void)
 
 static void gc_verify_track(jl_ptls_t ptls)
 {
+    // `gc_verify_track` is limited to single-threaded GC
+    if (jl_n_gcthreads != 0)
+        return;
     do {
         jl_gc_markqueue_t mq;
-        mq.current = mq.start = ptls->mark_queue.start;
-        mq.end = ptls->mark_queue.end;
-        mq.current_chunk = mq.chunk_start = ptls->mark_queue.chunk_start;
-        mq.chunk_end = ptls->mark_queue.chunk_end;
+        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);
@@ -214,7 +223,7 @@ static void gc_verify_track(jl_ptls_t ptls)
             gc_mark_finlist(&mq, &ptls2->finalizers, 0);
         }
         gc_mark_finlist(&mq, &finalizer_list_marked, 0);
-        gc_mark_loop_(ptls, &mq);
+        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;
@@ -248,11 +257,22 @@ static void gc_verify_track(jl_ptls_t ptls)
 
 void gc_verify(jl_ptls_t ptls)
 {
+    // `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;
-    mq.current = mq.start = ptls->mark_queue.start;
-    mq.end = ptls->mark_queue.end;
-    mq.current_chunk = mq.chunk_start = ptls->mark_queue.chunk_start;
-    mq.chunk_end = ptls->mark_queue.chunk_end;
+    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;
@@ -265,7 +285,7 @@ void gc_verify(jl_ptls_t ptls)
         gc_mark_finlist(&mq, &ptls2->finalizers, 0);
     }
     gc_mark_finlist(&mq, &finalizer_list_marked, 0);
-    gc_mark_loop_(ptls, &mq);
+    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];
@@ -1268,30 +1288,6 @@ int gc_slot_to_arrayidx(void *obj, void *_slot) JL_NOTSAFEPOINT
     return (slot - start) / elsize;
 }
 
-// Print a backtrace from the `mq->start` of the mark queue up to `mq->current`
-// `offset` will be added to `mq->current` for convenience in the debugger.
-NOINLINE void gc_mark_loop_unwind(jl_ptls_t ptls, jl_gc_markqueue_t *mq, int 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;
-    }
-    jl_value_t **start = mq->start;
-    jl_value_t **end = mq->current + offset;
-    for (; start < end; start++) {
-        jl_value_t *obj = *start;
-        jl_taggedvalue_t *o = jl_astaggedvalue(obj);
-        jl_safe_printf("Queued object: %p :: (tag: %zu) (bits: %zu)\n", obj,
-                       (uintptr_t)o->header, ((uintptr_t)o->header & 3));
-        jl_((void*)(jl_datatype_t *)(o->header & ~(uintptr_t)0xf));
-    }
-    jl_set_safe_restore(old_buf);
-}
-
 static int gc_logging_enabled = 0;
 
 JL_DLLEXPORT void jl_enable_gc_logging(int enable) {