2 * sgen-gc.c: Simple generational GC.
5 * Paolo Molaro (lupus@ximian.com)
6 * Rodrigo Kumpera (kumpera@gmail.com)
8 * Copyright 2005-2011 Novell, Inc (http://www.novell.com)
9 * Copyright 2011 Xamarin Inc (http://www.xamarin.com)
11 * Thread start/stop adapted from Boehm's GC:
12 * Copyright (c) 1994 by Xerox Corporation. All rights reserved.
13 * Copyright (c) 1996 by Silicon Graphics. All rights reserved.
14 * Copyright (c) 1998 by Fergus Henderson. All rights reserved.
15 * Copyright (c) 2000-2004 by Hewlett-Packard Company. All rights reserved.
16 * Copyright 2001-2003 Ximian, Inc
17 * Copyright 2003-2010 Novell, Inc.
18 * Copyright 2011 Xamarin, Inc.
19 * Copyright (C) 2012 Xamarin Inc
21 * This library is free software; you can redistribute it and/or
22 * modify it under the terms of the GNU Library General Public
23 * License 2.0 as published by the Free Software Foundation;
25 * This library is distributed in the hope that it will be useful,
26 * but WITHOUT ANY WARRANTY; without even the implied warranty of
27 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
28 * Library General Public License for more details.
30 * You should have received a copy of the GNU Library General Public
31 * License 2.0 along with this library; if not, write to the Free
32 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 * Important: allocation provides always zeroed memory, having to do
35 * a memset after allocation is deadly for performance.
36 * Memory usage at startup is currently as follows:
38 * 64 KB internal space
40 * We should provide a small memory config with half the sizes
42 * We currently try to make as few mono assumptions as possible:
43 * 1) 2-word header with no GC pointers in it (first vtable, second to store the
45 * 2) gc descriptor is the second word in the vtable (first word in the class)
46 * 3) 8 byte alignment is the minimum and enough (not true for special structures (SIMD), FIXME)
47 * 4) there is a function to get an object's size and the number of
48 * elements in an array.
49 * 5) we know the special way bounds are allocated for complex arrays
50 * 6) we know about proxies and how to treat them when domains are unloaded
52 * Always try to keep stack usage to a minimum: no recursive behaviour
53 * and no large stack allocs.
55 * General description.
56 * Objects are initially allocated in a nursery using a fast bump-pointer technique.
57 * When the nursery is full we start a nursery collection: this is performed with a
59 * When the old generation is full we start a copying GC of the old generation as well:
60 * this will be changed to mark&sweep with copying when fragmentation becomes to severe
61 * in the future. Maybe we'll even do both during the same collection like IMMIX.
63 * The things that complicate this description are:
64 * *) pinned objects: we can't move them so we need to keep track of them
65 * *) no precise info of the thread stacks and registers: we need to be able to
66 * quickly find the objects that may be referenced conservatively and pin them
67 * (this makes the first issues more important)
68 * *) large objects are too expensive to be dealt with using copying GC: we handle them
69 * with mark/sweep during major collections
70 * *) some objects need to not move even if they are small (interned strings, Type handles):
71 * we use mark/sweep for them, too: they are not allocated in the nursery, but inside
72 * PinnedChunks regions
78 *) we could have a function pointer in MonoClass to implement
79 customized write barriers for value types
81 *) investigate the stuff needed to advance a thread to a GC-safe
82 point (single-stepping, read from unmapped memory etc) and implement it.
83 This would enable us to inline allocations and write barriers, for example,
84 or at least parts of them, like the write barrier checks.
85 We may need this also for handling precise info on stacks, even simple things
86 as having uninitialized data on the stack and having to wait for the prolog
87 to zero it. Not an issue for the last frame that we scan conservatively.
88 We could always not trust the value in the slots anyway.
90 *) modify the jit to save info about references in stack locations:
91 this can be done just for locals as a start, so that at least
92 part of the stack is handled precisely.
94 *) test/fix endianess issues
96 *) Implement a card table as the write barrier instead of remembered
97 sets? Card tables are not easy to implement with our current
98 memory layout. We have several different kinds of major heap
99 objects: Small objects in regular blocks, small objects in pinned
100 chunks and LOS objects. If we just have a pointer we have no way
101 to tell which kind of object it points into, therefore we cannot
102 know where its card table is. The least we have to do to make
103 this happen is to get rid of write barriers for indirect stores.
106 *) Get rid of write barriers for indirect stores. We can do this by
107 telling the GC to wbarrier-register an object once we do an ldloca
108 or ldelema on it, and to unregister it once it's not used anymore
109 (it can only travel downwards on the stack). The problem with
110 unregistering is that it needs to happen eventually no matter
111 what, even if exceptions are thrown, the thread aborts, etc.
112 Rodrigo suggested that we could do only the registering part and
113 let the collector find out (pessimistically) when it's safe to
114 unregister, namely when the stack pointer of the thread that
115 registered the object is higher than it was when the registering
116 happened. This might make for a good first implementation to get
117 some data on performance.
119 *) Some sort of blacklist support? Blacklists is a concept from the
120 Boehm GC: if during a conservative scan we find pointers to an
121 area which we might use as heap, we mark that area as unusable, so
122 pointer retention by random pinning pointers is reduced.
124 *) experiment with max small object size (very small right now - 2kb,
125 because it's tied to the max freelist size)
127 *) add an option to mmap the whole heap in one chunk: it makes for many
128 simplifications in the checks (put the nursery at the top and just use a single
129 check for inclusion/exclusion): the issue this has is that on 32 bit systems it's
130 not flexible (too much of the address space may be used by default or we can't
131 increase the heap as needed) and we'd need a race-free mechanism to return memory
132 back to the system (mprotect(PROT_NONE) will still keep the memory allocated if it
133 was written to, munmap is needed, but the following mmap may not find the same segment
136 *) memzero the major fragments after restarting the world and optionally a smaller
139 *) investigate having fragment zeroing threads
141 *) separate locks for finalization and other minor stuff to reduce
144 *) try a different copying order to improve memory locality
146 *) a thread abort after a store but before the write barrier will
147 prevent the write barrier from executing
149 *) specialized dynamically generated markers/copiers
151 *) Dynamically adjust TLAB size to the number of threads. If we have
152 too many threads that do allocation, we might need smaller TLABs,
153 and we might get better performance with larger TLABs if we only
154 have a handful of threads. We could sum up the space left in all
155 assigned TLABs and if that's more than some percentage of the
156 nursery size, reduce the TLAB size.
158 *) Explore placing unreachable objects on unused nursery memory.
159 Instead of memset'ng a region to zero, place an int[] covering it.
160 A good place to start is add_nursery_frag. The tricky thing here is
161 placing those objects atomically outside of a collection.
163 *) Allocation should use asymmetric Dekker synchronization:
164 http://blogs.oracle.com/dave/resource/Asymmetric-Dekker-Synchronization.txt
165 This should help weak consistency archs.
172 #define _XOPEN_SOURCE
173 #define _DARWIN_C_SOURCE
179 #ifdef HAVE_PTHREAD_H
182 #ifdef HAVE_PTHREAD_NP_H
183 #include <pthread_np.h>
191 #include "mono/sgen/sgen-gc.h"
192 #include "mono/sgen/sgen-cardtable.h"
193 #include "mono/sgen/sgen-protocol.h"
194 #include "mono/sgen/sgen-memory-governor.h"
195 #include "mono/sgen/sgen-hash-table.h"
196 #include "mono/sgen/sgen-cardtable.h"
197 #include "mono/sgen/sgen-pinning.h"
198 #include "mono/sgen/sgen-workers.h"
199 #include "mono/sgen/sgen-client.h"
200 #include "mono/sgen/sgen-pointer-queue.h"
201 #include "mono/sgen/gc-internal-agnostic.h"
202 #include "mono/utils/mono-proclib.h"
203 #include "mono/utils/mono-memory-model.h"
204 #include "mono/utils/hazard-pointer.h"
206 #include <mono/utils/memcheck.h>
208 #undef pthread_create
210 #undef pthread_detach
213 * ######################################################################
214 * ######## Types and constants used by the GC.
215 * ######################################################################
218 /* 0 means not initialized, 1 is initialized, -1 means in progress */
219 static int gc_initialized = 0;
220 /* If set, check if we need to do something every X allocations */
221 gboolean has_per_allocation_action;
222 /* If set, do a heap check every X allocation */
223 guint32 verify_before_allocs = 0;
224 /* If set, do a minor collection before every X allocation */
225 guint32 collect_before_allocs = 0;
226 /* If set, do a whole heap check before each collection */
227 static gboolean whole_heap_check_before_collection = FALSE;
228 /* If set, do a heap consistency check before each minor collection */
229 static gboolean consistency_check_at_minor_collection = FALSE;
230 /* If set, do a mod union consistency check before each finishing collection pause */
231 static gboolean mod_union_consistency_check = FALSE;
232 /* If set, check whether mark bits are consistent after major collections */
233 static gboolean check_mark_bits_after_major_collection = FALSE;
234 /* If set, check that all nursery objects are pinned/not pinned, depending on context */
235 static gboolean check_nursery_objects_pinned = FALSE;
236 /* If set, do a few checks when the concurrent collector is used */
237 static gboolean do_concurrent_checks = FALSE;
238 /* If set, do a plausibility check on the scan_starts before and after
240 static gboolean do_scan_starts_check = FALSE;
243 * If the major collector is concurrent and this is FALSE, we will
244 * never initiate a synchronous major collection, unless requested via
247 static gboolean allow_synchronous_major = TRUE;
248 static gboolean disable_minor_collections = FALSE;
249 static gboolean disable_major_collections = FALSE;
250 static gboolean do_verify_nursery = FALSE;
251 static gboolean do_dump_nursery_content = FALSE;
252 static gboolean enable_nursery_canaries = FALSE;
254 #ifdef HEAVY_STATISTICS
255 guint64 stat_objects_alloced_degraded = 0;
256 guint64 stat_bytes_alloced_degraded = 0;
258 guint64 stat_copy_object_called_nursery = 0;
259 guint64 stat_objects_copied_nursery = 0;
260 guint64 stat_copy_object_called_major = 0;
261 guint64 stat_objects_copied_major = 0;
263 guint64 stat_scan_object_called_nursery = 0;
264 guint64 stat_scan_object_called_major = 0;
266 guint64 stat_slots_allocated_in_vain;
268 guint64 stat_nursery_copy_object_failed_from_space = 0;
269 guint64 stat_nursery_copy_object_failed_forwarded = 0;
270 guint64 stat_nursery_copy_object_failed_pinned = 0;
271 guint64 stat_nursery_copy_object_failed_to_space = 0;
273 static guint64 stat_wbarrier_add_to_global_remset = 0;
274 static guint64 stat_wbarrier_arrayref_copy = 0;
275 static guint64 stat_wbarrier_generic_store = 0;
276 static guint64 stat_wbarrier_generic_store_atomic = 0;
277 static guint64 stat_wbarrier_set_root = 0;
280 static guint64 stat_pinned_objects = 0;
282 static guint64 time_minor_pre_collection_fragment_clear = 0;
283 static guint64 time_minor_pinning = 0;
284 static guint64 time_minor_scan_remsets = 0;
285 static guint64 time_minor_scan_pinned = 0;
286 static guint64 time_minor_scan_roots = 0;
287 static guint64 time_minor_finish_gray_stack = 0;
288 static guint64 time_minor_fragment_creation = 0;
290 static guint64 time_major_pre_collection_fragment_clear = 0;
291 static guint64 time_major_pinning = 0;
292 static guint64 time_major_scan_pinned = 0;
293 static guint64 time_major_scan_roots = 0;
294 static guint64 time_major_scan_mod_union = 0;
295 static guint64 time_major_finish_gray_stack = 0;
296 static guint64 time_major_free_bigobjs = 0;
297 static guint64 time_major_los_sweep = 0;
298 static guint64 time_major_sweep = 0;
299 static guint64 time_major_fragment_creation = 0;
301 static guint64 time_max = 0;
303 static SGEN_TV_DECLARE (time_major_conc_collection_start);
304 static SGEN_TV_DECLARE (time_major_conc_collection_end);
306 static SGEN_TV_DECLARE (last_minor_collection_start_tv);
307 static SGEN_TV_DECLARE (last_minor_collection_end_tv);
309 int gc_debug_level = 0;
314 mono_gc_flush_info (void)
316 fflush (gc_debug_file);
320 #define TV_DECLARE SGEN_TV_DECLARE
321 #define TV_GETTIME SGEN_TV_GETTIME
322 #define TV_ELAPSED SGEN_TV_ELAPSED
324 static SGEN_TV_DECLARE (sgen_init_timestamp);
326 NurseryClearPolicy nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
328 #define object_is_forwarded SGEN_OBJECT_IS_FORWARDED
329 #define object_is_pinned SGEN_OBJECT_IS_PINNED
330 #define pin_object SGEN_PIN_OBJECT
332 #define ptr_in_nursery sgen_ptr_in_nursery
334 #define LOAD_VTABLE SGEN_LOAD_VTABLE
337 nursery_canaries_enabled (void)
339 return enable_nursery_canaries;
342 #define safe_object_get_size sgen_safe_object_get_size
345 * ######################################################################
346 * ######## Global data.
347 * ######################################################################
349 LOCK_DECLARE (gc_mutex);
350 gboolean sgen_try_free_some_memory;
352 #define SCAN_START_SIZE SGEN_SCAN_START_SIZE
354 size_t degraded_mode = 0;
356 static mword bytes_pinned_from_failed_allocation = 0;
358 GCMemSection *nursery_section = NULL;
359 static volatile mword lowest_heap_address = ~(mword)0;
360 static volatile mword highest_heap_address = 0;
362 LOCK_DECLARE (sgen_interruption_mutex);
364 int current_collection_generation = -1;
365 static volatile gboolean concurrent_collection_in_progress = FALSE;
367 /* objects that are ready to be finalized */
368 static SgenPointerQueue fin_ready_queue = SGEN_POINTER_QUEUE_INIT (INTERNAL_MEM_FINALIZE_READY);
369 static SgenPointerQueue critical_fin_queue = SGEN_POINTER_QUEUE_INIT (INTERNAL_MEM_FINALIZE_READY);
371 /* registered roots: the key to the hash is the root start address */
373 * Different kinds of roots are kept separate to speed up pin_from_roots () for example.
375 SgenHashTable roots_hash [ROOT_TYPE_NUM] = {
376 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL),
377 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL),
378 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL)
380 static mword roots_size = 0; /* amount of memory in the root set */
382 /* The size of a TLAB */
383 /* The bigger the value, the less often we have to go to the slow path to allocate a new
384 * one, but the more space is wasted by threads not allocating much memory.
386 * FIXME: Make this self-tuning for each thread.
388 guint32 tlab_size = (1024 * 4);
390 #define MAX_SMALL_OBJ_SIZE SGEN_MAX_SMALL_OBJ_SIZE
392 #define ALLOC_ALIGN SGEN_ALLOC_ALIGN
394 #define ALIGN_UP SGEN_ALIGN_UP
396 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
397 MonoNativeThreadId main_gc_thread = NULL;
400 /*Object was pinned during the current collection*/
401 static mword objects_pinned;
404 * ######################################################################
405 * ######## Macros and function declarations.
406 * ######################################################################
409 typedef SgenGrayQueue GrayQueue;
411 /* forward declarations */
412 static void scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx);
414 static void pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx);
415 static void finish_gray_stack (int generation, ScanCopyContext ctx);
418 SgenMajorCollector major_collector;
419 SgenMinorCollector sgen_minor_collector;
420 /* FIXME: get rid of this */
421 static GrayQueue gray_queue;
423 static SgenRememberedSet remset;
425 /* The gray queue to use from the main collection thread. */
426 #define WORKERS_DISTRIBUTE_GRAY_QUEUE (&gray_queue)
429 * The gray queue a worker job must use. If we're not parallel or
430 * concurrent, we use the main gray queue.
432 static SgenGrayQueue*
433 sgen_workers_get_job_gray_queue (WorkerData *worker_data)
435 return worker_data ? &worker_data->private_gray_queue : WORKERS_DISTRIBUTE_GRAY_QUEUE;
439 gray_queue_redirect (SgenGrayQueue *queue)
441 gboolean wake = FALSE;
444 GrayQueueSection *section = sgen_gray_object_dequeue_section (queue);
447 sgen_section_gray_queue_enqueue (queue->alloc_prepare_data, section);
452 g_assert (concurrent_collection_in_progress);
453 sgen_workers_ensure_awake ();
458 gray_queue_enable_redirect (SgenGrayQueue *queue)
460 if (!concurrent_collection_in_progress)
463 sgen_gray_queue_set_alloc_prepare (queue, gray_queue_redirect, sgen_workers_get_distribute_section_gray_queue ());
464 gray_queue_redirect (queue);
468 sgen_scan_area_with_callback (char *start, char *end, IterateObjectCallbackFunc callback, void *data, gboolean allow_flags, gboolean fail_on_canaries)
470 while (start < end) {
474 if (!*(void**)start) {
475 start += sizeof (void*); /* should be ALLOC_ALIGN, really */
480 if (!(obj = SGEN_OBJECT_IS_FORWARDED (start)))
486 if (!sgen_client_object_is_array_fill ((GCObject*)obj)) {
487 CHECK_CANARY_FOR_OBJECT ((GCObject*)obj, fail_on_canaries);
488 size = ALIGN_UP (safe_object_get_size ((GCObject*)obj));
489 callback ((GCObject*)obj, size, data);
490 CANARIFY_SIZE (size);
492 size = ALIGN_UP (safe_object_get_size ((GCObject*)obj));
500 * sgen_add_to_global_remset:
502 * The global remset contains locations which point into newspace after
503 * a minor collection. This can happen if the objects they point to are pinned.
505 * LOCKING: If called from a parallel collector, the global remset
506 * lock must be held. For serial collectors that is not necessary.
509 sgen_add_to_global_remset (gpointer ptr, gpointer obj)
511 SGEN_ASSERT (5, sgen_ptr_in_nursery (obj), "Target pointer of global remset must be in the nursery");
513 HEAVY_STAT (++stat_wbarrier_add_to_global_remset);
515 if (!major_collector.is_concurrent) {
516 SGEN_ASSERT (5, current_collection_generation != -1, "Global remsets can only be added during collections");
518 if (current_collection_generation == -1)
519 SGEN_ASSERT (5, sgen_concurrent_collection_in_progress (), "Global remsets outside of collection pauses can only be added by the concurrent collector");
522 if (!object_is_pinned (obj))
523 SGEN_ASSERT (5, sgen_minor_collector.is_split || sgen_concurrent_collection_in_progress (), "Non-pinned objects can only remain in nursery if it is a split nursery");
524 else if (sgen_cement_lookup_or_register (obj))
527 remset.record_pointer (ptr);
529 sgen_pin_stats_register_global_remset (obj);
531 SGEN_LOG (8, "Adding global remset for %p", ptr);
532 binary_protocol_global_remset (ptr, obj, (gpointer)SGEN_LOAD_VTABLE (obj));
536 * sgen_drain_gray_stack:
538 * Scan objects in the gray stack until the stack is empty. This should be called
539 * frequently after each object is copied, to achieve better locality and cache
542 * max_objs is the maximum number of objects to scan, or -1 to scan until the stack is
546 sgen_drain_gray_stack (int max_objs, ScanCopyContext ctx)
548 ScanObjectFunc scan_func = ctx.ops->scan_object;
549 GrayQueue *queue = ctx.queue;
551 if (current_collection_generation == GENERATION_OLD && major_collector.drain_gray_stack)
552 return major_collector.drain_gray_stack (ctx);
556 for (i = 0; i != max_objs; ++i) {
559 GRAY_OBJECT_DEQUEUE (queue, &obj, &desc);
562 SGEN_LOG (9, "Precise gray object scan %p (%s)", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
563 scan_func (obj, desc, queue);
565 } while (max_objs < 0);
570 * Addresses in the pin queue are already sorted. This function finds
571 * the object header for each address and pins the object. The
572 * addresses must be inside the nursery section. The (start of the)
573 * address array is overwritten with the addresses of the actually
574 * pinned objects. Return the number of pinned objects.
577 pin_objects_from_nursery_pin_queue (gboolean do_scan_objects, ScanCopyContext ctx)
579 GCMemSection *section = nursery_section;
580 void **start = sgen_pinning_get_entry (section->pin_queue_first_entry);
581 void **end = sgen_pinning_get_entry (section->pin_queue_last_entry);
582 void *start_nursery = section->data;
583 void *end_nursery = section->next_data;
588 void *pinning_front = start_nursery;
590 void **definitely_pinned = start;
591 ScanObjectFunc scan_func = ctx.ops->scan_object;
592 SgenGrayQueue *queue = ctx.queue;
594 sgen_nursery_allocator_prepare_for_pinning ();
596 while (start < end) {
597 GCObject *obj_to_pin = NULL;
598 size_t obj_to_pin_size = 0;
603 SGEN_ASSERT (0, addr >= start_nursery && addr < end_nursery, "Potential pinning address out of range");
604 SGEN_ASSERT (0, addr >= last, "Pin queue not sorted");
611 SGEN_LOG (5, "Considering pinning addr %p", addr);
612 /* We've already processed everything up to pinning_front. */
613 if (addr < pinning_front) {
619 * Find the closest scan start <= addr. We might search backward in the
620 * scan_starts array because entries might be NULL. In the worst case we
621 * start at start_nursery.
623 idx = ((char*)addr - (char*)section->data) / SCAN_START_SIZE;
624 SGEN_ASSERT (0, idx < section->num_scan_start, "Scan start index out of range");
625 search_start = (void*)section->scan_starts [idx];
626 if (!search_start || search_start > addr) {
629 search_start = section->scan_starts [idx];
630 if (search_start && search_start <= addr)
633 if (!search_start || search_start > addr)
634 search_start = start_nursery;
638 * If the pinning front is closer than the scan start we found, start
639 * searching at the front.
641 if (search_start < pinning_front)
642 search_start = pinning_front;
645 * Now addr should be in an object a short distance from search_start.
647 * search_start must point to zeroed mem or point to an object.
650 size_t obj_size, canarified_obj_size;
653 if (!*(void**)search_start) {
654 search_start = (void*)ALIGN_UP ((mword)search_start + sizeof (gpointer));
655 /* The loop condition makes sure we don't overrun addr. */
659 canarified_obj_size = obj_size = ALIGN_UP (safe_object_get_size ((GCObject*)search_start));
662 * Filler arrays are marked by an invalid sync word. We don't
663 * consider them for pinning. They are not delimited by canaries,
666 if (!sgen_client_object_is_array_fill ((GCObject*)search_start)) {
667 CHECK_CANARY_FOR_OBJECT (search_start, TRUE);
668 CANARIFY_SIZE (canarified_obj_size);
670 if (addr >= search_start && (char*)addr < (char*)search_start + obj_size) {
671 /* This is the object we're looking for. */
672 obj_to_pin = (GCObject*)search_start;
673 obj_to_pin_size = canarified_obj_size;
678 /* Skip to the next object */
679 search_start = (void*)((char*)search_start + canarified_obj_size);
680 } while (search_start <= addr);
682 /* We've searched past the address we were looking for. */
684 pinning_front = search_start;
685 goto next_pin_queue_entry;
689 * We've found an object to pin. It might still be a dummy array, but we
690 * can advance the pinning front in any case.
692 pinning_front = (char*)obj_to_pin + obj_to_pin_size;
695 * If this is a dummy array marking the beginning of a nursery
696 * fragment, we don't pin it.
698 if (sgen_client_object_is_array_fill (obj_to_pin))
699 goto next_pin_queue_entry;
702 * Finally - pin the object!
704 desc = sgen_obj_get_descriptor_safe (obj_to_pin);
705 if (do_scan_objects) {
706 scan_func (obj_to_pin, desc, queue);
708 SGEN_LOG (4, "Pinned object %p, vtable %p (%s), count %d\n",
709 obj_to_pin, *(void**)obj_to_pin, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj_to_pin)), count);
710 binary_protocol_pin (obj_to_pin,
711 (gpointer)LOAD_VTABLE (obj_to_pin),
712 safe_object_get_size (obj_to_pin));
714 pin_object (obj_to_pin);
715 GRAY_OBJECT_ENQUEUE (queue, obj_to_pin, desc);
716 sgen_pin_stats_register_object (obj_to_pin, obj_to_pin_size);
717 definitely_pinned [count] = obj_to_pin;
721 next_pin_queue_entry:
725 sgen_client_nursery_objects_pinned (definitely_pinned, count);
726 stat_pinned_objects += count;
731 pin_objects_in_nursery (gboolean do_scan_objects, ScanCopyContext ctx)
735 if (nursery_section->pin_queue_first_entry == nursery_section->pin_queue_last_entry)
738 reduced_to = pin_objects_from_nursery_pin_queue (do_scan_objects, ctx);
739 nursery_section->pin_queue_last_entry = nursery_section->pin_queue_first_entry + reduced_to;
743 * This function is only ever called (via `collector_pin_object()` in `sgen-copy-object.h`)
744 * when we can't promote an object because we're out of memory.
747 sgen_pin_object (GCObject *object, GrayQueue *queue)
750 * All pinned objects are assumed to have been staged, so we need to stage as well.
751 * Also, the count of staged objects shows that "late pinning" happened.
753 sgen_pin_stage_ptr (object);
755 SGEN_PIN_OBJECT (object);
756 binary_protocol_pin (object, (gpointer)LOAD_VTABLE (object), safe_object_get_size (object));
759 sgen_pin_stats_register_object (object, safe_object_get_size (object));
761 GRAY_OBJECT_ENQUEUE (queue, object, sgen_obj_get_descriptor_safe (object));
764 /* Sort the addresses in array in increasing order.
765 * Done using a by-the book heap sort. Which has decent and stable performance, is pretty cache efficient.
768 sgen_sort_addresses (void **array, size_t size)
773 for (i = 1; i < size; ++i) {
776 size_t parent = (child - 1) / 2;
778 if (array [parent] >= array [child])
781 tmp = array [parent];
782 array [parent] = array [child];
789 for (i = size - 1; i > 0; --i) {
792 array [i] = array [0];
798 while (root * 2 + 1 <= end) {
799 size_t child = root * 2 + 1;
801 if (child < end && array [child] < array [child + 1])
803 if (array [root] >= array [child])
807 array [root] = array [child];
816 * Scan the memory between start and end and queue values which could be pointers
817 * to the area between start_nursery and end_nursery for later consideration.
818 * Typically used for thread stacks.
821 sgen_conservatively_pin_objects_from (void **start, void **end, void *start_nursery, void *end_nursery, int pin_type)
825 #if defined(VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE) && !defined(_WIN64)
826 VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE (start, (char*)end - (char*)start);
829 while (start < end) {
831 * *start can point to the middle of an object
832 * note: should we handle pointing at the end of an object?
833 * pinning in C# code disallows pointing at the end of an object
834 * but there is some small chance that an optimizing C compiler
835 * may keep the only reference to an object by pointing
836 * at the end of it. We ignore this small chance for now.
837 * Pointers to the end of an object are indistinguishable
838 * from pointers to the start of the next object in memory
839 * so if we allow that we'd need to pin two objects...
840 * We queue the pointer in an array, the
841 * array will then be sorted and uniqued. This way
842 * we can coalesce several pinning pointers and it should
843 * be faster since we'd do a memory scan with increasing
844 * addresses. Note: we can align the address to the allocation
845 * alignment, so the unique process is more effective.
847 mword addr = (mword)*start;
848 addr &= ~(ALLOC_ALIGN - 1);
849 if (addr >= (mword)start_nursery && addr < (mword)end_nursery) {
850 SGEN_LOG (6, "Pinning address %p from %p", (void*)addr, start);
851 sgen_pin_stage_ptr ((void*)addr);
852 binary_protocol_pin_stage (start, (void*)addr);
853 sgen_pin_stats_register_address ((char*)addr, pin_type);
859 SGEN_LOG (7, "found %d potential pinned heap pointers", count);
863 * The first thing we do in a collection is to identify pinned objects.
864 * This function considers all the areas of memory that need to be
865 * conservatively scanned.
868 pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx)
872 SGEN_LOG (2, "Scanning pinned roots (%d bytes, %d/%d entries)", (int)roots_size, roots_hash [ROOT_TYPE_NORMAL].num_entries, roots_hash [ROOT_TYPE_PINNED].num_entries);
873 /* objects pinned from the API are inside these roots */
874 SGEN_HASH_TABLE_FOREACH (&roots_hash [ROOT_TYPE_PINNED], start_root, root) {
875 SGEN_LOG (6, "Pinned roots %p-%p", start_root, root->end_root);
876 sgen_conservatively_pin_objects_from (start_root, (void**)root->end_root, start_nursery, end_nursery, PIN_TYPE_OTHER);
877 } SGEN_HASH_TABLE_FOREACH_END;
878 /* now deal with the thread stacks
879 * in the future we should be able to conservatively scan only:
880 * *) the cpu registers
881 * *) the unmanaged stack frames
882 * *) the _last_ managed stack frame
883 * *) pointers slots in managed frames
885 sgen_client_scan_thread_data (start_nursery, end_nursery, FALSE, ctx);
889 single_arg_user_copy_or_mark (GCObject **obj, void *gc_data)
891 ScanCopyContext *ctx = gc_data;
892 ctx->ops->copy_or_mark_object (obj, ctx->queue);
896 * The memory area from start_root to end_root contains pointers to objects.
897 * Their position is precisely described by @desc (this means that the pointer
898 * can be either NULL or the pointer to the start of an object).
899 * This functions copies them to to_space updates them.
901 * This function is not thread-safe!
904 precisely_scan_objects_from (void** start_root, void** end_root, char* n_start, char *n_end, SgenDescriptor desc, ScanCopyContext ctx)
906 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
907 SgenGrayQueue *queue = ctx.queue;
909 switch (desc & ROOT_DESC_TYPE_MASK) {
910 case ROOT_DESC_BITMAP:
911 desc >>= ROOT_DESC_TYPE_SHIFT;
913 if ((desc & 1) && *start_root) {
914 copy_func ((GCObject**)start_root, queue);
915 SGEN_LOG (9, "Overwrote root at %p with %p", start_root, *start_root);
921 case ROOT_DESC_COMPLEX: {
922 gsize *bitmap_data = sgen_get_complex_descriptor_bitmap (desc);
923 gsize bwords = (*bitmap_data) - 1;
924 void **start_run = start_root;
926 while (bwords-- > 0) {
927 gsize bmap = *bitmap_data++;
928 void **objptr = start_run;
930 if ((bmap & 1) && *objptr) {
931 copy_func ((GCObject**)objptr, queue);
932 SGEN_LOG (9, "Overwrote root at %p with %p", objptr, *objptr);
937 start_run += GC_BITS_PER_WORD;
941 case ROOT_DESC_USER: {
942 SgenUserRootMarkFunc marker = sgen_get_user_descriptor_func (desc);
943 marker (start_root, single_arg_user_copy_or_mark, &ctx);
946 case ROOT_DESC_RUN_LEN:
947 g_assert_not_reached ();
949 g_assert_not_reached ();
954 reset_heap_boundaries (void)
956 lowest_heap_address = ~(mword)0;
957 highest_heap_address = 0;
961 sgen_update_heap_boundaries (mword low, mword high)
966 old = lowest_heap_address;
969 } while (SGEN_CAS_PTR ((gpointer*)&lowest_heap_address, (gpointer)low, (gpointer)old) != (gpointer)old);
972 old = highest_heap_address;
975 } while (SGEN_CAS_PTR ((gpointer*)&highest_heap_address, (gpointer)high, (gpointer)old) != (gpointer)old);
979 * Allocate and setup the data structures needed to be able to allocate objects
980 * in the nursery. The nursery is stored in nursery_section.
985 GCMemSection *section;
992 SGEN_LOG (2, "Allocating nursery size: %zu", (size_t)sgen_nursery_size);
993 /* later we will alloc a larger area for the nursery but only activate
994 * what we need. The rest will be used as expansion if we have too many pinned
995 * objects in the existing nursery.
997 /* FIXME: handle OOM */
998 section = sgen_alloc_internal (INTERNAL_MEM_SECTION);
1000 alloc_size = sgen_nursery_size;
1002 /* If there isn't enough space even for the nursery we should simply abort. */
1003 g_assert (sgen_memgov_try_alloc_space (alloc_size, SPACE_NURSERY));
1005 data = major_collector.alloc_heap (alloc_size, alloc_size, DEFAULT_NURSERY_BITS);
1006 sgen_update_heap_boundaries ((mword)data, (mword)(data + sgen_nursery_size));
1007 SGEN_LOG (4, "Expanding nursery size (%p-%p): %lu, total: %lu", data, data + alloc_size, (unsigned long)sgen_nursery_size, (unsigned long)sgen_gc_get_total_heap_allocation ());
1008 section->data = section->next_data = data;
1009 section->size = alloc_size;
1010 section->end_data = data + sgen_nursery_size;
1011 scan_starts = (alloc_size + SCAN_START_SIZE - 1) / SCAN_START_SIZE;
1012 section->scan_starts = sgen_alloc_internal_dynamic (sizeof (char*) * scan_starts, INTERNAL_MEM_SCAN_STARTS, TRUE);
1013 section->num_scan_start = scan_starts;
1015 nursery_section = section;
1017 sgen_nursery_allocator_set_nursery_bounds (data, data + sgen_nursery_size);
1021 mono_gc_get_logfile (void)
1023 return gc_debug_file;
1027 scan_finalizer_entries (SgenPointerQueue *fin_queue, ScanCopyContext ctx)
1029 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
1030 SgenGrayQueue *queue = ctx.queue;
1033 for (i = 0; i < fin_queue->next_slot; ++i) {
1034 GCObject *obj = fin_queue->data [i];
1037 SGEN_LOG (5, "Scan of fin ready object: %p (%s)\n", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
1038 copy_func ((GCObject**)&fin_queue->data [i], queue);
1043 generation_name (int generation)
1045 switch (generation) {
1046 case GENERATION_NURSERY: return "nursery";
1047 case GENERATION_OLD: return "old";
1048 default: g_assert_not_reached ();
1053 sgen_generation_name (int generation)
1055 return generation_name (generation);
1059 finish_gray_stack (int generation, ScanCopyContext ctx)
1063 int done_with_ephemerons, ephemeron_rounds = 0;
1064 char *start_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_start () : NULL;
1065 char *end_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_end () : (char*)-1;
1066 SgenGrayQueue *queue = ctx.queue;
1069 * We copied all the reachable objects. Now it's the time to copy
1070 * the objects that were not referenced by the roots, but by the copied objects.
1071 * we built a stack of objects pointed to by gray_start: they are
1072 * additional roots and we may add more items as we go.
1073 * We loop until gray_start == gray_objects which means no more objects have
1074 * been added. Note this is iterative: no recursion is involved.
1075 * We need to walk the LO list as well in search of marked big objects
1076 * (use a flag since this is needed only on major collections). We need to loop
1077 * here as well, so keep a counter of marked LO (increasing it in copy_object).
1078 * To achieve better cache locality and cache usage, we drain the gray stack
1079 * frequently, after each object is copied, and just finish the work here.
1081 sgen_drain_gray_stack (-1, ctx);
1083 SGEN_LOG (2, "%s generation done", generation_name (generation));
1086 Reset bridge data, we might have lingering data from a previous collection if this is a major
1087 collection trigged by minor overflow.
1089 We must reset the gathered bridges since their original block might be evacuated due to major
1090 fragmentation in the meanwhile and the bridge code should not have to deal with that.
1092 if (sgen_client_bridge_need_processing ())
1093 sgen_client_bridge_reset_data ();
1096 * Walk the ephemeron tables marking all values with reachable keys. This must be completely done
1097 * before processing finalizable objects and non-tracking weak links to avoid finalizing/clearing
1098 * objects that are in fact reachable.
1100 done_with_ephemerons = 0;
1102 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1103 sgen_drain_gray_stack (-1, ctx);
1105 } while (!done_with_ephemerons);
1107 sgen_client_mark_togglerefs (start_addr, end_addr, ctx);
1109 if (sgen_client_bridge_need_processing ()) {
1110 /*Make sure the gray stack is empty before we process bridge objects so we get liveness right*/
1111 sgen_drain_gray_stack (-1, ctx);
1112 sgen_collect_bridge_objects (generation, ctx);
1113 if (generation == GENERATION_OLD)
1114 sgen_collect_bridge_objects (GENERATION_NURSERY, ctx);
1117 Do the first bridge step here, as the collector liveness state will become useless after that.
1119 An important optimization is to only proccess the possibly dead part of the object graph and skip
1120 over all live objects as we transitively know everything they point must be alive too.
1122 The above invariant is completely wrong if we let the gray queue be drained and mark/copy everything.
1124 This has the unfortunate side effect of making overflow collections perform the first step twice, but
1125 given we now have heuristics that perform major GC in anticipation of minor overflows this should not
1128 sgen_client_bridge_processing_stw_step ();
1132 Make sure we drain the gray stack before processing disappearing links and finalizers.
1133 If we don't make sure it is empty we might wrongly see a live object as dead.
1135 sgen_drain_gray_stack (-1, ctx);
1138 We must clear weak links that don't track resurrection before processing object ready for
1139 finalization so they can be cleared before that.
1141 sgen_null_link_in_range (generation, TRUE, ctx);
1142 if (generation == GENERATION_OLD)
1143 sgen_null_link_in_range (GENERATION_NURSERY, TRUE, ctx);
1146 /* walk the finalization queue and move also the objects that need to be
1147 * finalized: use the finalized objects as new roots so the objects they depend
1148 * on are also not reclaimed. As with the roots above, only objects in the nursery
1149 * are marked/copied.
1151 sgen_finalize_in_range (generation, ctx);
1152 if (generation == GENERATION_OLD)
1153 sgen_finalize_in_range (GENERATION_NURSERY, ctx);
1154 /* drain the new stack that might have been created */
1155 SGEN_LOG (6, "Precise scan of gray area post fin");
1156 sgen_drain_gray_stack (-1, ctx);
1159 * This must be done again after processing finalizable objects since CWL slots are cleared only after the key is finalized.
1161 done_with_ephemerons = 0;
1163 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1164 sgen_drain_gray_stack (-1, ctx);
1166 } while (!done_with_ephemerons);
1168 sgen_client_clear_unreachable_ephemerons (ctx);
1171 * We clear togglerefs only after all possible chances of revival are done.
1172 * This is semantically more inline with what users expect and it allows for
1173 * user finalizers to correctly interact with TR objects.
1175 sgen_client_clear_togglerefs (start_addr, end_addr, ctx);
1178 SGEN_LOG (2, "Finalize queue handling scan for %s generation: %ld usecs %d ephemeron rounds", generation_name (generation), TV_ELAPSED (atv, btv), ephemeron_rounds);
1181 * handle disappearing links
1182 * Note we do this after checking the finalization queue because if an object
1183 * survives (at least long enough to be finalized) we don't clear the link.
1184 * This also deals with a possible issue with the monitor reclamation: with the Boehm
1185 * GC a finalized object my lose the monitor because it is cleared before the finalizer is
1188 g_assert (sgen_gray_object_queue_is_empty (queue));
1190 sgen_null_link_in_range (generation, FALSE, ctx);
1191 if (generation == GENERATION_OLD)
1192 sgen_null_link_in_range (GENERATION_NURSERY, FALSE, ctx);
1193 if (sgen_gray_object_queue_is_empty (queue))
1195 sgen_drain_gray_stack (-1, ctx);
1198 g_assert (sgen_gray_object_queue_is_empty (queue));
1200 sgen_gray_object_queue_trim_free_list (queue);
1204 sgen_check_section_scan_starts (GCMemSection *section)
1207 for (i = 0; i < section->num_scan_start; ++i) {
1208 if (section->scan_starts [i]) {
1209 mword size = safe_object_get_size ((GCObject*) section->scan_starts [i]);
1210 SGEN_ASSERT (0, size >= SGEN_CLIENT_MINIMUM_OBJECT_SIZE && size <= MAX_SMALL_OBJ_SIZE, "Weird object size at scan starts.");
1216 check_scan_starts (void)
1218 if (!do_scan_starts_check)
1220 sgen_check_section_scan_starts (nursery_section);
1221 major_collector.check_scan_starts ();
1225 scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx)
1229 SGEN_HASH_TABLE_FOREACH (&roots_hash [root_type], start_root, root) {
1230 SGEN_LOG (6, "Precise root scan %p-%p (desc: %p)", start_root, root->end_root, (void*)root->root_desc);
1231 precisely_scan_objects_from (start_root, (void**)root->end_root, addr_start, addr_end, root->root_desc, ctx);
1232 } SGEN_HASH_TABLE_FOREACH_END;
1238 static gboolean inited = FALSE;
1243 mono_counters_register ("Collection max time", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME | MONO_COUNTER_MONOTONIC, &time_max);
1245 mono_counters_register ("Minor fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pre_collection_fragment_clear);
1246 mono_counters_register ("Minor pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pinning);
1247 mono_counters_register ("Minor scan remembered set", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_remsets);
1248 mono_counters_register ("Minor scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_pinned);
1249 mono_counters_register ("Minor scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_roots);
1250 mono_counters_register ("Minor fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_fragment_creation);
1252 mono_counters_register ("Major fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pre_collection_fragment_clear);
1253 mono_counters_register ("Major pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pinning);
1254 mono_counters_register ("Major scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_pinned);
1255 mono_counters_register ("Major scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_roots);
1256 mono_counters_register ("Major scan mod union", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_mod_union);
1257 mono_counters_register ("Major finish gray stack", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_finish_gray_stack);
1258 mono_counters_register ("Major free big objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_free_bigobjs);
1259 mono_counters_register ("Major LOS sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_los_sweep);
1260 mono_counters_register ("Major sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_sweep);
1261 mono_counters_register ("Major fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_fragment_creation);
1263 mono_counters_register ("Number of pinned objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_pinned_objects);
1265 #ifdef HEAVY_STATISTICS
1266 mono_counters_register ("WBarrier remember pointer", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_add_to_global_remset);
1267 mono_counters_register ("WBarrier arrayref copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_arrayref_copy);
1268 mono_counters_register ("WBarrier generic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store);
1269 mono_counters_register ("WBarrier generic atomic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store_atomic);
1270 mono_counters_register ("WBarrier set root", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_set_root);
1272 mono_counters_register ("# objects allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_alloced_degraded);
1273 mono_counters_register ("bytes allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced_degraded);
1275 mono_counters_register ("# copy_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_nursery);
1276 mono_counters_register ("# objects copied (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_nursery);
1277 mono_counters_register ("# copy_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_major);
1278 mono_counters_register ("# objects copied (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_major);
1280 mono_counters_register ("# scan_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_nursery);
1281 mono_counters_register ("# scan_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_major);
1283 mono_counters_register ("Slots allocated in vain", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_slots_allocated_in_vain);
1285 mono_counters_register ("# nursery copy_object() failed from space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_from_space);
1286 mono_counters_register ("# nursery copy_object() failed forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_forwarded);
1287 mono_counters_register ("# nursery copy_object() failed pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_pinned);
1288 mono_counters_register ("# nursery copy_object() failed to space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_to_space);
1290 sgen_nursery_allocator_init_heavy_stats ();
1298 reset_pinned_from_failed_allocation (void)
1300 bytes_pinned_from_failed_allocation = 0;
1304 sgen_set_pinned_from_failed_allocation (mword objsize)
1306 bytes_pinned_from_failed_allocation += objsize;
1310 sgen_collection_is_concurrent (void)
1312 switch (current_collection_generation) {
1313 case GENERATION_NURSERY:
1315 case GENERATION_OLD:
1316 return concurrent_collection_in_progress;
1318 g_error ("Invalid current generation %d", current_collection_generation);
1324 sgen_concurrent_collection_in_progress (void)
1326 return concurrent_collection_in_progress;
1330 SgenThreadPoolJob job;
1331 SgenObjectOperations *ops;
1335 job_remembered_set_scan (void *worker_data_untyped, SgenThreadPoolJob *job)
1337 WorkerData *worker_data = worker_data_untyped;
1338 ScanJob *job_data = (ScanJob*)job;
1339 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1340 remset.scan_remsets (ctx);
1344 SgenThreadPoolJob job;
1345 SgenObjectOperations *ops;
1349 } ScanFromRegisteredRootsJob;
1352 job_scan_from_registered_roots (void *worker_data_untyped, SgenThreadPoolJob *job)
1354 WorkerData *worker_data = worker_data_untyped;
1355 ScanFromRegisteredRootsJob *job_data = (ScanFromRegisteredRootsJob*)job;
1356 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1358 scan_from_registered_roots (job_data->heap_start, job_data->heap_end, job_data->root_type, ctx);
1362 SgenThreadPoolJob job;
1363 SgenObjectOperations *ops;
1366 } ScanThreadDataJob;
1369 job_scan_thread_data (void *worker_data_untyped, SgenThreadPoolJob *job)
1371 WorkerData *worker_data = worker_data_untyped;
1372 ScanThreadDataJob *job_data = (ScanThreadDataJob*)job;
1373 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1375 sgen_client_scan_thread_data (job_data->heap_start, job_data->heap_end, TRUE, ctx);
1379 SgenThreadPoolJob job;
1380 SgenObjectOperations *ops;
1381 SgenPointerQueue *queue;
1382 } ScanFinalizerEntriesJob;
1385 job_scan_finalizer_entries (void *worker_data_untyped, SgenThreadPoolJob *job)
1387 WorkerData *worker_data = worker_data_untyped;
1388 ScanFinalizerEntriesJob *job_data = (ScanFinalizerEntriesJob*)job;
1389 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1391 scan_finalizer_entries (job_data->queue, ctx);
1395 job_scan_major_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1397 WorkerData *worker_data = worker_data_untyped;
1398 ScanJob *job_data = (ScanJob*)job;
1399 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1401 g_assert (concurrent_collection_in_progress);
1402 major_collector.scan_card_table (TRUE, ctx);
1406 job_scan_los_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1408 WorkerData *worker_data = worker_data_untyped;
1409 ScanJob *job_data = (ScanJob*)job;
1410 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1412 g_assert (concurrent_collection_in_progress);
1413 sgen_los_scan_card_table (TRUE, ctx);
1417 init_gray_queue (void)
1419 if (sgen_collection_is_concurrent ())
1420 sgen_workers_init_distribute_gray_queue ();
1421 sgen_gray_object_queue_init (&gray_queue, NULL);
1425 enqueue_scan_from_roots_jobs (char *heap_start, char *heap_end, SgenObjectOperations *ops)
1427 ScanFromRegisteredRootsJob *scrrj;
1428 ScanThreadDataJob *stdj;
1429 ScanFinalizerEntriesJob *sfej;
1431 /* registered roots, this includes static fields */
1433 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots normal", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1435 scrrj->heap_start = heap_start;
1436 scrrj->heap_end = heap_end;
1437 scrrj->root_type = ROOT_TYPE_NORMAL;
1438 sgen_workers_enqueue_job (&scrrj->job);
1440 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots wbarrier", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1442 scrrj->heap_start = heap_start;
1443 scrrj->heap_end = heap_end;
1444 scrrj->root_type = ROOT_TYPE_WBARRIER;
1445 sgen_workers_enqueue_job (&scrrj->job);
1449 stdj = (ScanThreadDataJob*)sgen_thread_pool_job_alloc ("scan thread data", job_scan_thread_data, sizeof (ScanThreadDataJob));
1450 stdj->heap_start = heap_start;
1451 stdj->heap_end = heap_end;
1452 sgen_workers_enqueue_job (&stdj->job);
1454 /* Scan the list of objects ready for finalization. */
1456 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1457 sfej->queue = &fin_ready_queue;
1459 sgen_workers_enqueue_job (&sfej->job);
1461 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan critical finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1462 sfej->queue = &critical_fin_queue;
1464 sgen_workers_enqueue_job (&sfej->job);
1468 * Perform a nursery collection.
1470 * Return whether any objects were late-pinned due to being out of memory.
1473 collect_nursery (SgenGrayQueue *unpin_queue, gboolean finish_up_concurrent_mark)
1475 gboolean needs_major;
1476 size_t max_garbage_amount;
1478 mword fragment_total;
1480 SgenObjectOperations *object_ops = &sgen_minor_collector.serial_ops;
1481 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, &gray_queue);
1485 if (disable_minor_collections)
1488 TV_GETTIME (last_minor_collection_start_tv);
1489 atv = last_minor_collection_start_tv;
1491 binary_protocol_collection_begin (gc_stats.minor_gc_count, GENERATION_NURSERY);
1493 if (do_verify_nursery || do_dump_nursery_content)
1494 sgen_debug_verify_nursery (do_dump_nursery_content);
1496 current_collection_generation = GENERATION_NURSERY;
1498 SGEN_ASSERT (0, !sgen_collection_is_concurrent (), "Why is the nursery collection concurrent?");
1500 reset_pinned_from_failed_allocation ();
1502 check_scan_starts ();
1504 sgen_nursery_alloc_prepare_for_minor ();
1508 nursery_next = sgen_nursery_alloc_get_upper_alloc_bound ();
1509 /* FIXME: optimize later to use the higher address where an object can be present */
1510 nursery_next = MAX (nursery_next, sgen_get_nursery_end ());
1512 SGEN_LOG (1, "Start nursery collection %d %p-%p, size: %d", gc_stats.minor_gc_count, sgen_get_nursery_start (), nursery_next, (int)(nursery_next - sgen_get_nursery_start ()));
1513 max_garbage_amount = nursery_next - sgen_get_nursery_start ();
1514 g_assert (nursery_section->size >= max_garbage_amount);
1516 /* world must be stopped already */
1518 time_minor_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1520 sgen_client_pre_collection_checks ();
1522 nursery_section->next_data = nursery_next;
1524 major_collector.start_nursery_collection ();
1526 sgen_memgov_minor_collection_start ();
1530 gc_stats.minor_gc_count ++;
1532 if (whole_heap_check_before_collection) {
1533 sgen_clear_nursery_fragments ();
1534 sgen_check_whole_heap (finish_up_concurrent_mark);
1536 if (consistency_check_at_minor_collection)
1537 sgen_check_consistency ();
1539 sgen_process_fin_stage_entries ();
1540 sgen_process_dislink_stage_entries ();
1542 /* pin from pinned handles */
1543 sgen_init_pinning ();
1544 sgen_client_binary_protocol_mark_start (GENERATION_NURSERY);
1545 pin_from_roots (sgen_get_nursery_start (), nursery_next, ctx);
1546 /* pin cemented objects */
1547 sgen_pin_cemented_objects ();
1548 /* identify pinned objects */
1549 sgen_optimize_pin_queue ();
1550 sgen_pinning_setup_section (nursery_section);
1552 pin_objects_in_nursery (FALSE, ctx);
1553 sgen_pinning_trim_queue_to_section (nursery_section);
1556 time_minor_pinning += TV_ELAPSED (btv, atv);
1557 SGEN_LOG (2, "Finding pinned pointers: %zd in %ld usecs", sgen_get_pinned_count (), TV_ELAPSED (btv, atv));
1558 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
1561 * FIXME: When we finish a concurrent collection we do a nursery collection first,
1562 * as part of which we scan the card table. Then, later, we scan the mod union
1563 * cardtable. We should only have to do one.
1565 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan remset", job_remembered_set_scan, sizeof (ScanJob));
1566 sj->ops = object_ops;
1567 sgen_workers_enqueue_job (&sj->job);
1569 /* we don't have complete write barrier yet, so we scan all the old generation sections */
1571 time_minor_scan_remsets += TV_ELAPSED (atv, btv);
1572 SGEN_LOG (2, "Old generation scan: %ld usecs", TV_ELAPSED (atv, btv));
1574 sgen_pin_stats_print_class_stats ();
1576 sgen_drain_gray_stack (-1, ctx);
1578 /* FIXME: Why do we do this at this specific, seemingly random, point? */
1579 sgen_client_collecting_minor (&fin_ready_queue, &critical_fin_queue);
1582 time_minor_scan_pinned += TV_ELAPSED (btv, atv);
1584 enqueue_scan_from_roots_jobs (sgen_get_nursery_start (), nursery_next, object_ops);
1587 time_minor_scan_roots += TV_ELAPSED (atv, btv);
1589 finish_gray_stack (GENERATION_NURSERY, ctx);
1592 time_minor_finish_gray_stack += TV_ELAPSED (btv, atv);
1593 sgen_client_binary_protocol_mark_end (GENERATION_NURSERY);
1595 if (objects_pinned) {
1596 sgen_optimize_pin_queue ();
1597 sgen_pinning_setup_section (nursery_section);
1600 /* walk the pin_queue, build up the fragment list of free memory, unmark
1601 * pinned objects as we go, memzero() the empty fragments so they are ready for the
1604 sgen_client_binary_protocol_reclaim_start (GENERATION_NURSERY);
1605 fragment_total = sgen_build_nursery_fragments (nursery_section, unpin_queue);
1606 if (!fragment_total)
1609 /* Clear TLABs for all threads */
1610 sgen_clear_tlabs ();
1612 sgen_client_binary_protocol_reclaim_end (GENERATION_NURSERY);
1614 time_minor_fragment_creation += TV_ELAPSED (atv, btv);
1615 SGEN_LOG (2, "Fragment creation: %ld usecs, %lu bytes available", TV_ELAPSED (atv, btv), (unsigned long)fragment_total);
1617 if (consistency_check_at_minor_collection)
1618 sgen_check_major_refs ();
1620 major_collector.finish_nursery_collection ();
1622 TV_GETTIME (last_minor_collection_end_tv);
1623 gc_stats.minor_gc_time += TV_ELAPSED (last_minor_collection_start_tv, last_minor_collection_end_tv);
1625 sgen_debug_dump_heap ("minor", gc_stats.minor_gc_count - 1, NULL);
1627 /* prepare the pin queue for the next collection */
1628 sgen_finish_pinning ();
1629 if (sgen_have_pending_finalizers ()) {
1630 SGEN_LOG (4, "Finalizer-thread wakeup");
1631 sgen_client_finalize_notify ();
1633 sgen_pin_stats_reset ();
1634 /* clear cemented hash */
1635 sgen_cement_clear_below_threshold ();
1637 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
1639 remset.finish_minor_collection ();
1641 check_scan_starts ();
1643 binary_protocol_flush_buffers (FALSE);
1645 sgen_memgov_minor_collection_end ();
1647 /*objects are late pinned because of lack of memory, so a major is a good call*/
1648 needs_major = objects_pinned > 0;
1649 current_collection_generation = -1;
1652 binary_protocol_collection_end (gc_stats.minor_gc_count - 1, GENERATION_NURSERY, 0, 0);
1654 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
1655 sgen_check_nursery_objects_pinned (unpin_queue != NULL);
1661 COPY_OR_MARK_FROM_ROOTS_SERIAL,
1662 COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT,
1663 COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT
1664 } CopyOrMarkFromRootsMode;
1667 major_copy_or_mark_from_roots (size_t *old_next_pin_slot, CopyOrMarkFromRootsMode mode, SgenObjectOperations *object_ops)
1672 /* FIXME: only use these values for the precise scan
1673 * note that to_space pointers should be excluded anyway...
1675 char *heap_start = NULL;
1676 char *heap_end = (char*)-1;
1677 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, WORKERS_DISTRIBUTE_GRAY_QUEUE);
1678 gboolean concurrent = mode != COPY_OR_MARK_FROM_ROOTS_SERIAL;
1680 SGEN_ASSERT (0, !!concurrent == !!concurrent_collection_in_progress, "We've been called with the wrong mode.");
1682 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1683 /*This cleans up unused fragments */
1684 sgen_nursery_allocator_prepare_for_pinning ();
1686 if (do_concurrent_checks)
1687 sgen_debug_check_nursery_is_clean ();
1689 /* The concurrent collector doesn't touch the nursery. */
1690 sgen_nursery_alloc_prepare_for_major ();
1697 /* Pinning depends on this */
1698 sgen_clear_nursery_fragments ();
1700 if (whole_heap_check_before_collection)
1701 sgen_check_whole_heap (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT);
1704 time_major_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1706 if (!sgen_collection_is_concurrent ())
1707 nursery_section->next_data = sgen_get_nursery_end ();
1708 /* we should also coalesce scanning from sections close to each other
1709 * and deal with pointers outside of the sections later.
1714 sgen_client_pre_collection_checks ();
1717 /* Remsets are not useful for a major collection */
1718 remset.clear_cards ();
1721 sgen_process_fin_stage_entries ();
1722 sgen_process_dislink_stage_entries ();
1725 sgen_init_pinning ();
1726 SGEN_LOG (6, "Collecting pinned addresses");
1727 pin_from_roots ((void*)lowest_heap_address, (void*)highest_heap_address, ctx);
1729 sgen_optimize_pin_queue ();
1731 sgen_client_collecting_major_1 ();
1734 * pin_queue now contains all candidate pointers, sorted and
1735 * uniqued. We must do two passes now to figure out which
1736 * objects are pinned.
1738 * The first is to find within the pin_queue the area for each
1739 * section. This requires that the pin_queue be sorted. We
1740 * also process the LOS objects and pinned chunks here.
1742 * The second, destructive, pass is to reduce the section
1743 * areas to pointers to the actually pinned objects.
1745 SGEN_LOG (6, "Pinning from sections");
1746 /* first pass for the sections */
1747 sgen_find_section_pin_queue_start_end (nursery_section);
1748 /* identify possible pointers to the insize of large objects */
1749 SGEN_LOG (6, "Pinning from large objects");
1750 for (bigobj = los_object_list; bigobj; bigobj = bigobj->next) {
1752 if (sgen_find_optimized_pin_queue_area ((char*)bigobj->data, (char*)bigobj->data + sgen_los_object_size (bigobj), &dummy, &dummy)) {
1753 binary_protocol_pin (bigobj->data, (gpointer)LOAD_VTABLE (bigobj->data), safe_object_get_size (bigobj->data));
1755 if (sgen_los_object_is_pinned (bigobj->data)) {
1756 SGEN_ASSERT (0, mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT, "LOS objects can only be pinned here after concurrent marking.");
1759 sgen_los_pin_object (bigobj->data);
1760 if (SGEN_OBJECT_HAS_REFERENCES (bigobj->data))
1761 GRAY_OBJECT_ENQUEUE (WORKERS_DISTRIBUTE_GRAY_QUEUE, bigobj->data, sgen_obj_get_descriptor ((GCObject*)bigobj->data));
1762 sgen_pin_stats_register_object (bigobj->data, safe_object_get_size (bigobj->data));
1763 SGEN_LOG (6, "Marked large object %p (%s) size: %lu from roots", bigobj->data,
1764 sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (bigobj->data)),
1765 (unsigned long)sgen_los_object_size (bigobj));
1767 sgen_client_pinned_los_object (bigobj->data);
1771 pin_objects_in_nursery (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT, ctx);
1772 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
1773 sgen_check_nursery_objects_pinned (mode != COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT);
1775 major_collector.pin_objects (WORKERS_DISTRIBUTE_GRAY_QUEUE);
1776 if (old_next_pin_slot)
1777 *old_next_pin_slot = sgen_get_pinned_count ();
1780 time_major_pinning += TV_ELAPSED (atv, btv);
1781 SGEN_LOG (2, "Finding pinned pointers: %zd in %ld usecs", sgen_get_pinned_count (), TV_ELAPSED (atv, btv));
1782 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
1784 major_collector.init_to_space ();
1787 * The concurrent collector doesn't move objects, neither on
1788 * the major heap nor in the nursery, so we can mark even
1789 * before pinning has finished. For the non-concurrent
1790 * collector we start the workers after pinning.
1792 if (mode != COPY_OR_MARK_FROM_ROOTS_SERIAL) {
1793 SGEN_ASSERT (0, sgen_workers_all_done (), "Why are the workers not done when we start or finish a major collection?");
1794 sgen_workers_start_all_workers (object_ops);
1795 gray_queue_enable_redirect (WORKERS_DISTRIBUTE_GRAY_QUEUE);
1798 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
1799 main_gc_thread = mono_native_thread_self ();
1802 sgen_client_collecting_major_2 ();
1805 time_major_scan_pinned += TV_ELAPSED (btv, atv);
1807 sgen_client_collecting_major_3 (&fin_ready_queue, &critical_fin_queue);
1810 * FIXME: is this the right context? It doesn't seem to contain a copy function
1811 * unless we're concurrent.
1813 enqueue_scan_from_roots_jobs (heap_start, heap_end, object_ops);
1816 time_major_scan_roots += TV_ELAPSED (atv, btv);
1818 if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
1821 /* Mod union card table */
1822 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan mod union cardtable", job_scan_major_mod_union_card_table, sizeof (ScanJob));
1823 sj->ops = object_ops;
1824 sgen_workers_enqueue_job (&sj->job);
1826 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan LOS mod union cardtable", job_scan_los_mod_union_card_table, sizeof (ScanJob));
1827 sj->ops = object_ops;
1828 sgen_workers_enqueue_job (&sj->job);
1831 time_major_scan_mod_union += TV_ELAPSED (btv, atv);
1834 sgen_pin_stats_print_class_stats ();
1838 major_finish_copy_or_mark (CopyOrMarkFromRootsMode mode)
1841 case COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT:
1843 * Prepare the pin queue for the next collection. Since pinning runs on the worker
1844 * threads we must wait for the jobs to finish before we can reset it.
1846 sgen_workers_wait_for_jobs_finished ();
1847 sgen_finish_pinning ();
1849 sgen_pin_stats_reset ();
1851 if (do_concurrent_checks)
1852 sgen_debug_check_nursery_is_clean ();
1854 case COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT:
1855 sgen_workers_wait_for_jobs_finished ();
1857 case COPY_OR_MARK_FROM_ROOTS_SERIAL:
1860 g_assert_not_reached ();
1865 major_start_collection (gboolean concurrent, size_t *old_next_pin_slot)
1867 SgenObjectOperations *object_ops;
1869 binary_protocol_collection_begin (gc_stats.major_gc_count, GENERATION_OLD);
1871 current_collection_generation = GENERATION_OLD;
1873 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
1876 sgen_cement_reset ();
1879 g_assert (major_collector.is_concurrent);
1880 concurrent_collection_in_progress = TRUE;
1882 object_ops = &major_collector.major_ops_concurrent_start;
1884 object_ops = &major_collector.major_ops_serial;
1887 reset_pinned_from_failed_allocation ();
1889 sgen_memgov_major_collection_start ();
1891 //count_ref_nonref_objs ();
1892 //consistency_check ();
1894 check_scan_starts ();
1897 SGEN_LOG (1, "Start major collection %d", gc_stats.major_gc_count);
1898 gc_stats.major_gc_count ++;
1900 if (major_collector.start_major_collection)
1901 major_collector.start_major_collection ();
1903 major_copy_or_mark_from_roots (old_next_pin_slot, concurrent ? COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT : COPY_OR_MARK_FROM_ROOTS_SERIAL, object_ops);
1904 major_finish_copy_or_mark (concurrent ? COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT : COPY_OR_MARK_FROM_ROOTS_SERIAL);
1908 major_finish_collection (const char *reason, size_t old_next_pin_slot, gboolean forced)
1910 ScannedObjectCounts counts;
1911 SgenObjectOperations *object_ops;
1912 mword fragment_total;
1918 if (concurrent_collection_in_progress) {
1919 object_ops = &major_collector.major_ops_concurrent_finish;
1921 major_copy_or_mark_from_roots (NULL, COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT, object_ops);
1923 major_finish_copy_or_mark (COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT);
1925 sgen_workers_join ();
1927 SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&gray_queue), "Why is the gray queue not empty after workers have finished working?");
1929 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
1930 main_gc_thread = NULL;
1933 object_ops = &major_collector.major_ops_serial;
1937 * The workers have stopped so we need to finish gray queue
1938 * work that might result from finalization in the main GC
1939 * thread. Redirection must therefore be turned off.
1941 sgen_gray_object_queue_disable_alloc_prepare (&gray_queue);
1942 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
1944 /* all the objects in the heap */
1945 finish_gray_stack (GENERATION_OLD, CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, &gray_queue));
1947 time_major_finish_gray_stack += TV_ELAPSED (btv, atv);
1949 SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after joining");
1951 if (objects_pinned) {
1952 g_assert (!concurrent_collection_in_progress);
1955 * This is slow, but we just OOM'd.
1957 * See comment at `sgen_pin_queue_clear_discarded_entries` for how the pin
1958 * queue is laid out at this point.
1960 sgen_pin_queue_clear_discarded_entries (nursery_section, old_next_pin_slot);
1962 * We need to reestablish all pinned nursery objects in the pin queue
1963 * because they're needed for fragment creation. Unpinning happens by
1964 * walking the whole queue, so it's not necessary to reestablish where major
1965 * heap block pins are - all we care is that they're still in there
1968 sgen_optimize_pin_queue ();
1969 sgen_find_section_pin_queue_start_end (nursery_section);
1973 reset_heap_boundaries ();
1974 sgen_update_heap_boundaries ((mword)sgen_get_nursery_start (), (mword)sgen_get_nursery_end ());
1976 /* walk the pin_queue, build up the fragment list of free memory, unmark
1977 * pinned objects as we go, memzero() the empty fragments so they are ready for the
1980 fragment_total = sgen_build_nursery_fragments (nursery_section, NULL);
1981 if (!fragment_total)
1983 SGEN_LOG (4, "Free space in nursery after major %ld", fragment_total);
1985 if (do_concurrent_checks && concurrent_collection_in_progress)
1986 sgen_debug_check_nursery_is_clean ();
1988 /* prepare the pin queue for the next collection */
1989 sgen_finish_pinning ();
1991 /* Clear TLABs for all threads */
1992 sgen_clear_tlabs ();
1994 sgen_pin_stats_reset ();
1996 sgen_cement_clear_below_threshold ();
1998 if (check_mark_bits_after_major_collection)
1999 sgen_check_heap_marked (concurrent_collection_in_progress);
2002 time_major_fragment_creation += TV_ELAPSED (atv, btv);
2004 binary_protocol_sweep_begin (GENERATION_OLD, !major_collector.sweeps_lazily);
2007 time_major_free_bigobjs += TV_ELAPSED (btv, atv);
2012 time_major_los_sweep += TV_ELAPSED (atv, btv);
2014 major_collector.sweep ();
2016 binary_protocol_sweep_end (GENERATION_OLD, !major_collector.sweeps_lazily);
2019 time_major_sweep += TV_ELAPSED (btv, atv);
2021 sgen_debug_dump_heap ("major", gc_stats.major_gc_count - 1, reason);
2023 if (sgen_have_pending_finalizers ()) {
2024 SGEN_LOG (4, "Finalizer-thread wakeup");
2025 sgen_client_finalize_notify ();
2028 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
2030 sgen_memgov_major_collection_end (forced);
2031 current_collection_generation = -1;
2033 memset (&counts, 0, sizeof (ScannedObjectCounts));
2034 major_collector.finish_major_collection (&counts);
2036 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
2038 SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after major collection has finished");
2039 if (concurrent_collection_in_progress)
2040 concurrent_collection_in_progress = FALSE;
2042 check_scan_starts ();
2044 binary_protocol_flush_buffers (FALSE);
2046 //consistency_check ();
2048 binary_protocol_collection_end (gc_stats.major_gc_count - 1, GENERATION_OLD, counts.num_scanned_objects, counts.num_unique_scanned_objects);
2052 major_do_collection (const char *reason, gboolean forced)
2054 TV_DECLARE (time_start);
2055 TV_DECLARE (time_end);
2056 size_t old_next_pin_slot;
2058 if (disable_major_collections)
2061 if (major_collector.get_and_reset_num_major_objects_marked) {
2062 long long num_marked = major_collector.get_and_reset_num_major_objects_marked ();
2063 g_assert (!num_marked);
2066 /* world must be stopped already */
2067 TV_GETTIME (time_start);
2069 major_start_collection (FALSE, &old_next_pin_slot);
2070 major_finish_collection (reason, old_next_pin_slot, forced);
2072 TV_GETTIME (time_end);
2073 gc_stats.major_gc_time += TV_ELAPSED (time_start, time_end);
2075 /* FIXME: also report this to the user, preferably in gc-end. */
2076 if (major_collector.get_and_reset_num_major_objects_marked)
2077 major_collector.get_and_reset_num_major_objects_marked ();
2079 return bytes_pinned_from_failed_allocation > 0;
2083 major_start_concurrent_collection (const char *reason)
2085 TV_DECLARE (time_start);
2086 TV_DECLARE (time_end);
2087 long long num_objects_marked;
2089 if (disable_major_collections)
2092 TV_GETTIME (time_start);
2093 SGEN_TV_GETTIME (time_major_conc_collection_start);
2095 num_objects_marked = major_collector.get_and_reset_num_major_objects_marked ();
2096 g_assert (num_objects_marked == 0);
2098 binary_protocol_concurrent_start ();
2100 // FIXME: store reason and pass it when finishing
2101 major_start_collection (TRUE, NULL);
2103 gray_queue_redirect (&gray_queue);
2105 num_objects_marked = major_collector.get_and_reset_num_major_objects_marked ();
2107 TV_GETTIME (time_end);
2108 gc_stats.major_gc_time += TV_ELAPSED (time_start, time_end);
2110 current_collection_generation = -1;
2114 * Returns whether the major collection has finished.
2117 major_should_finish_concurrent_collection (void)
2119 SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&gray_queue), "Why is the gray queue not empty before we have started doing anything?");
2120 return sgen_workers_all_done ();
2124 major_update_concurrent_collection (void)
2126 TV_DECLARE (total_start);
2127 TV_DECLARE (total_end);
2129 TV_GETTIME (total_start);
2131 binary_protocol_concurrent_update ();
2133 major_collector.update_cardtable_mod_union ();
2134 sgen_los_update_cardtable_mod_union ();
2136 TV_GETTIME (total_end);
2137 gc_stats.major_gc_time += TV_ELAPSED (total_start, total_end);
2141 major_finish_concurrent_collection (gboolean forced)
2143 TV_DECLARE (total_start);
2144 TV_DECLARE (total_end);
2146 TV_GETTIME (total_start);
2148 binary_protocol_concurrent_finish ();
2151 * The major collector can add global remsets which are processed in the finishing
2152 * nursery collection, below. That implies that the workers must have finished
2153 * marking before the nursery collection is allowed to run, otherwise we might miss
2156 sgen_workers_wait ();
2158 SGEN_TV_GETTIME (time_major_conc_collection_end);
2159 gc_stats.major_gc_time_concurrent += SGEN_TV_ELAPSED (time_major_conc_collection_start, time_major_conc_collection_end);
2161 major_collector.update_cardtable_mod_union ();
2162 sgen_los_update_cardtable_mod_union ();
2164 if (mod_union_consistency_check)
2165 sgen_check_mod_union_consistency ();
2167 current_collection_generation = GENERATION_OLD;
2168 sgen_cement_reset ();
2169 major_finish_collection ("finishing", -1, forced);
2171 if (whole_heap_check_before_collection)
2172 sgen_check_whole_heap (FALSE);
2174 TV_GETTIME (total_end);
2175 gc_stats.major_gc_time += TV_ELAPSED (total_start, total_end) - TV_ELAPSED (last_minor_collection_start_tv, last_minor_collection_end_tv);
2177 current_collection_generation = -1;
2181 * Ensure an allocation request for @size will succeed by freeing enough memory.
2183 * LOCKING: The GC lock MUST be held.
2186 sgen_ensure_free_space (size_t size)
2188 int generation_to_collect = -1;
2189 const char *reason = NULL;
2191 if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
2192 if (sgen_need_major_collection (size)) {
2193 reason = "LOS overflow";
2194 generation_to_collect = GENERATION_OLD;
2197 if (degraded_mode) {
2198 if (sgen_need_major_collection (size)) {
2199 reason = "Degraded mode overflow";
2200 generation_to_collect = GENERATION_OLD;
2202 } else if (sgen_need_major_collection (size)) {
2203 reason = "Minor allowance";
2204 generation_to_collect = GENERATION_OLD;
2206 generation_to_collect = GENERATION_NURSERY;
2207 reason = "Nursery full";
2211 if (generation_to_collect == -1) {
2212 if (concurrent_collection_in_progress && sgen_workers_all_done ()) {
2213 generation_to_collect = GENERATION_OLD;
2214 reason = "Finish concurrent collection";
2218 if (generation_to_collect == -1)
2220 sgen_perform_collection (size, generation_to_collect, reason, FALSE);
2224 * LOCKING: Assumes the GC lock is held.
2227 sgen_perform_collection (size_t requested_size, int generation_to_collect, const char *reason, gboolean wait_to_finish)
2229 TV_DECLARE (gc_start);
2230 TV_DECLARE (gc_end);
2231 TV_DECLARE (gc_total_start);
2232 TV_DECLARE (gc_total_end);
2233 GGTimingInfo infos [2];
2234 int overflow_generation_to_collect = -1;
2235 int oldest_generation_collected = generation_to_collect;
2236 const char *overflow_reason = NULL;
2238 binary_protocol_collection_requested (generation_to_collect, requested_size, wait_to_finish ? 1 : 0);
2240 SGEN_ASSERT (0, generation_to_collect == GENERATION_NURSERY || generation_to_collect == GENERATION_OLD, "What generation is this?");
2242 TV_GETTIME (gc_start);
2244 sgen_stop_world (generation_to_collect);
2246 TV_GETTIME (gc_total_start);
2248 if (concurrent_collection_in_progress) {
2250 * We update the concurrent collection. If it finished, we're done. If
2251 * not, and we've been asked to do a nursery collection, we do that.
2253 gboolean finish = major_should_finish_concurrent_collection () || (wait_to_finish && generation_to_collect == GENERATION_OLD);
2256 major_finish_concurrent_collection (wait_to_finish);
2257 oldest_generation_collected = GENERATION_OLD;
2259 sgen_workers_signal_start_nursery_collection_and_wait ();
2261 major_update_concurrent_collection ();
2262 if (generation_to_collect == GENERATION_NURSERY)
2263 collect_nursery (NULL, FALSE);
2265 sgen_workers_signal_finish_nursery_collection ();
2272 * If we've been asked to do a major collection, and the major collector wants to
2273 * run synchronously (to evacuate), we set the flag to do that.
2275 if (generation_to_collect == GENERATION_OLD &&
2276 allow_synchronous_major &&
2277 major_collector.want_synchronous_collection &&
2278 *major_collector.want_synchronous_collection) {
2279 wait_to_finish = TRUE;
2282 SGEN_ASSERT (0, !concurrent_collection_in_progress, "Why did this not get handled above?");
2285 * There's no concurrent collection in progress. Collect the generation we're asked
2286 * to collect. If the major collector is concurrent and we're not forced to wait,
2287 * start a concurrent collection.
2289 // FIXME: extract overflow reason
2290 if (generation_to_collect == GENERATION_NURSERY) {
2291 if (collect_nursery (NULL, FALSE)) {
2292 overflow_generation_to_collect = GENERATION_OLD;
2293 overflow_reason = "Minor overflow";
2296 if (major_collector.is_concurrent && !wait_to_finish) {
2297 collect_nursery (NULL, FALSE);
2298 major_start_concurrent_collection (reason);
2299 // FIXME: set infos[0] properly
2303 if (major_do_collection (reason, wait_to_finish)) {
2304 overflow_generation_to_collect = GENERATION_NURSERY;
2305 overflow_reason = "Excessive pinning";
2309 TV_GETTIME (gc_end);
2311 memset (infos, 0, sizeof (infos));
2312 infos [0].generation = generation_to_collect;
2313 infos [0].reason = reason;
2314 infos [0].is_overflow = FALSE;
2315 infos [1].generation = -1;
2316 infos [0].total_time = SGEN_TV_ELAPSED (gc_start, gc_end);
2318 SGEN_ASSERT (0, !concurrent_collection_in_progress, "Why did this not get handled above?");
2320 if (overflow_generation_to_collect != -1) {
2322 * We need to do an overflow collection, either because we ran out of memory
2323 * or the nursery is fully pinned.
2326 infos [1].generation = overflow_generation_to_collect;
2327 infos [1].reason = overflow_reason;
2328 infos [1].is_overflow = TRUE;
2331 if (overflow_generation_to_collect == GENERATION_NURSERY)
2332 collect_nursery (NULL, FALSE);
2334 major_do_collection (overflow_reason, wait_to_finish);
2336 TV_GETTIME (gc_end);
2337 infos [1].total_time = SGEN_TV_ELAPSED (gc_start, gc_end);
2339 oldest_generation_collected = MAX (oldest_generation_collected, overflow_generation_to_collect);
2342 SGEN_LOG (2, "Heap size: %lu, LOS size: %lu", (unsigned long)sgen_gc_get_total_heap_allocation (), (unsigned long)los_memory_usage);
2344 /* this also sets the proper pointers for the next allocation */
2345 if (generation_to_collect == GENERATION_NURSERY && !sgen_can_alloc_size (requested_size)) {
2346 /* TypeBuilder and MonoMethod are killing mcs with fragmentation */
2347 SGEN_LOG (1, "nursery collection didn't find enough room for %zd alloc (%zd pinned)", requested_size, sgen_get_pinned_count ());
2348 sgen_dump_pin_queue ();
2353 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
2355 TV_GETTIME (gc_total_end);
2356 time_max = MAX (time_max, TV_ELAPSED (gc_total_start, gc_total_end));
2358 sgen_restart_world (oldest_generation_collected, infos);
2362 * ######################################################################
2363 * ######## Memory allocation from the OS
2364 * ######################################################################
2365 * This section of code deals with getting memory from the OS and
2366 * allocating memory for GC-internal data structures.
2367 * Internal memory can be handled with a freelist for small objects.
2373 G_GNUC_UNUSED static void
2374 report_internal_mem_usage (void)
2376 printf ("Internal memory usage:\n");
2377 sgen_report_internal_mem_usage ();
2378 printf ("Pinned memory usage:\n");
2379 major_collector.report_pinned_memory_usage ();
2383 * ######################################################################
2384 * ######## Finalization support
2385 * ######################################################################
2389 * If the object has been forwarded it means it's still referenced from a root.
2390 * If it is pinned it's still alive as well.
2391 * A LOS object is only alive if we have pinned it.
2392 * Return TRUE if @obj is ready to be finalized.
2394 static inline gboolean
2395 sgen_is_object_alive (GCObject *object)
2397 if (ptr_in_nursery (object))
2398 return sgen_nursery_is_object_alive (object);
2400 return sgen_major_is_object_alive (object);
2404 * This function returns true if @object is either alive and belongs to the
2405 * current collection - major collections are full heap, so old gen objects
2406 * are never alive during a minor collection.
2409 sgen_is_object_alive_and_on_current_collection (GCObject *object)
2411 if (ptr_in_nursery (object))
2412 return sgen_nursery_is_object_alive (object);
2414 if (current_collection_generation == GENERATION_NURSERY)
2417 return sgen_major_is_object_alive (object);
2422 sgen_gc_is_object_ready_for_finalization (GCObject *object)
2424 return !sgen_is_object_alive (object);
2428 sgen_queue_finalization_entry (GCObject *obj)
2430 gboolean critical = sgen_client_object_has_critical_finalizer (obj);
2432 sgen_pointer_queue_add (critical ? &critical_fin_queue : &fin_ready_queue, obj);
2434 sgen_client_object_queued_for_finalization (obj);
2438 sgen_object_is_live (GCObject *obj)
2440 return sgen_is_object_alive_and_on_current_collection (obj);
2444 * `System.GC.WaitForPendingFinalizers` first checks `sgen_have_pending_finalizers()` to
2445 * determine whether it can exit quickly. The latter must therefore only return FALSE if
2446 * all finalizers have really finished running.
2448 * `sgen_gc_invoke_finalizers()` first dequeues a finalizable object, and then finalizes it.
2449 * This means that just checking whether the queues are empty leaves the possibility that an
2450 * object might have been dequeued but not yet finalized. That's why we need the additional
2451 * flag `pending_unqueued_finalizer`.
2454 static volatile gboolean pending_unqueued_finalizer = FALSE;
2457 sgen_gc_invoke_finalizers (void)
2461 g_assert (!pending_unqueued_finalizer);
2463 /* FIXME: batch to reduce lock contention */
2464 while (sgen_have_pending_finalizers ()) {
2470 * We need to set `pending_unqueued_finalizer` before dequeing the
2471 * finalizable object.
2473 if (!sgen_pointer_queue_is_empty (&fin_ready_queue)) {
2474 pending_unqueued_finalizer = TRUE;
2475 mono_memory_write_barrier ();
2476 obj = sgen_pointer_queue_pop (&fin_ready_queue);
2477 } else if (!sgen_pointer_queue_is_empty (&critical_fin_queue)) {
2478 pending_unqueued_finalizer = TRUE;
2479 mono_memory_write_barrier ();
2480 obj = sgen_pointer_queue_pop (&critical_fin_queue);
2486 SGEN_LOG (7, "Finalizing object %p (%s)", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
2494 /* the object is on the stack so it is pinned */
2495 /*g_print ("Calling finalizer for object: %p (%s)\n", obj, sgen_client_object_safe_name (obj));*/
2496 sgen_client_run_finalize (obj);
2499 if (pending_unqueued_finalizer) {
2500 mono_memory_write_barrier ();
2501 pending_unqueued_finalizer = FALSE;
2508 sgen_have_pending_finalizers (void)
2510 return pending_unqueued_finalizer || !sgen_pointer_queue_is_empty (&fin_ready_queue) || !sgen_pointer_queue_is_empty (&critical_fin_queue);
2514 * ######################################################################
2515 * ######## registered roots support
2516 * ######################################################################
2520 * We do not coalesce roots.
2523 sgen_register_root (char *start, size_t size, SgenDescriptor descr, int root_type)
2525 RootRecord new_root;
2528 for (i = 0; i < ROOT_TYPE_NUM; ++i) {
2529 RootRecord *root = sgen_hash_table_lookup (&roots_hash [i], start);
2530 /* we allow changing the size and the descriptor (for thread statics etc) */
2532 size_t old_size = root->end_root - start;
2533 root->end_root = start + size;
2534 SGEN_ASSERT (0, !!root->root_desc == !!descr, "Can't change whether a root is precise or conservative.");
2535 root->root_desc = descr;
2537 roots_size -= old_size;
2543 new_root.end_root = start + size;
2544 new_root.root_desc = descr;
2546 sgen_hash_table_replace (&roots_hash [root_type], start, &new_root, NULL);
2549 SGEN_LOG (3, "Added root for range: %p-%p, descr: %llx (%d/%d bytes)", start, new_root.end_root, descr, (int)size, (int)roots_size);
2556 sgen_deregister_root (char* addr)
2562 for (root_type = 0; root_type < ROOT_TYPE_NUM; ++root_type) {
2563 if (sgen_hash_table_remove (&roots_hash [root_type], addr, &root))
2564 roots_size -= (root.end_root - addr);
2570 * ######################################################################
2571 * ######## Thread handling (stop/start code)
2572 * ######################################################################
2576 sgen_get_current_collection_generation (void)
2578 return current_collection_generation;
2582 sgen_thread_register (SgenThreadInfo* info, void *stack_bottom_fallback)
2584 #ifndef HAVE_KW_THREAD
2585 info->tlab_start = info->tlab_next = info->tlab_temp_end = info->tlab_real_end = NULL;
2588 sgen_init_tlab_info (info);
2590 sgen_client_thread_register (info, stack_bottom_fallback);
2596 sgen_thread_unregister (SgenThreadInfo *p)
2598 sgen_client_thread_unregister (p);
2602 * ######################################################################
2603 * ######## Write barriers
2604 * ######################################################################
2608 * Note: the write barriers first do the needed GC work and then do the actual store:
2609 * this way the value is visible to the conservative GC scan after the write barrier
2610 * itself. If a GC interrupts the barrier in the middle, value will be kept alive by
2611 * the conservative scan, otherwise by the remembered set scan.
2615 mono_gc_wbarrier_arrayref_copy (gpointer dest_ptr, gpointer src_ptr, int count)
2617 HEAVY_STAT (++stat_wbarrier_arrayref_copy);
2618 /*This check can be done without taking a lock since dest_ptr array is pinned*/
2619 if (ptr_in_nursery (dest_ptr) || count <= 0) {
2620 mono_gc_memmove_aligned (dest_ptr, src_ptr, count * sizeof (gpointer));
2624 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2625 if (binary_protocol_is_heavy_enabled ()) {
2627 for (i = 0; i < count; ++i) {
2628 gpointer dest = (gpointer*)dest_ptr + i;
2629 gpointer obj = *((gpointer*)src_ptr + i);
2631 binary_protocol_wbarrier (dest, obj, (gpointer)LOAD_VTABLE (obj));
2636 remset.wbarrier_arrayref_copy (dest_ptr, src_ptr, count);
2640 mono_gc_wbarrier_generic_nostore (gpointer ptr)
2644 HEAVY_STAT (++stat_wbarrier_generic_store);
2646 sgen_client_wbarrier_generic_nostore_check (ptr);
2648 obj = *(gpointer*)ptr;
2650 binary_protocol_wbarrier (ptr, obj, (gpointer)LOAD_VTABLE (obj));
2653 * We need to record old->old pointer locations for the
2654 * concurrent collector.
2656 if (!ptr_in_nursery (obj) && !concurrent_collection_in_progress) {
2657 SGEN_LOG (8, "Skipping remset at %p", ptr);
2661 SGEN_LOG (8, "Adding remset at %p", ptr);
2663 remset.wbarrier_generic_nostore (ptr);
2667 mono_gc_wbarrier_generic_store (gpointer ptr, GCObject* value)
2669 SGEN_LOG (8, "Wbarrier store at %p to %p (%s)", ptr, value, value ? sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (value)) : "null");
2670 SGEN_UPDATE_REFERENCE_ALLOW_NULL (ptr, value);
2671 if (ptr_in_nursery (value) || concurrent_collection_in_progress)
2672 mono_gc_wbarrier_generic_nostore (ptr);
2673 sgen_dummy_use (value);
2676 /* Same as mono_gc_wbarrier_generic_store () but performs the store
2677 * as an atomic operation with release semantics.
2680 mono_gc_wbarrier_generic_store_atomic (gpointer ptr, GCObject *value)
2682 HEAVY_STAT (++stat_wbarrier_generic_store_atomic);
2684 SGEN_LOG (8, "Wbarrier atomic store at %p to %p (%s)", ptr, value, value ? sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (value)) : "null");
2686 InterlockedWritePointer (ptr, value);
2688 if (ptr_in_nursery (value) || concurrent_collection_in_progress)
2689 mono_gc_wbarrier_generic_nostore (ptr);
2691 sgen_dummy_use (value);
2695 sgen_wbarrier_value_copy_bitmap (gpointer _dest, gpointer _src, int size, unsigned bitmap)
2697 GCObject **dest = _dest;
2698 GCObject **src = _src;
2702 mono_gc_wbarrier_generic_store (dest, *src);
2707 size -= SIZEOF_VOID_P;
2713 * ######################################################################
2714 * ######## Other mono public interface functions.
2715 * ######################################################################
2719 sgen_gc_collect (int generation)
2724 sgen_perform_collection (0, generation, "user request", TRUE);
2729 sgen_gc_collection_count (int generation)
2731 if (generation == 0)
2732 return gc_stats.minor_gc_count;
2733 return gc_stats.major_gc_count;
2737 sgen_gc_get_used_size (void)
2741 tot = los_memory_usage;
2742 tot += nursery_section->next_data - nursery_section->data;
2743 tot += major_collector.get_used_size ();
2744 /* FIXME: account for pinned objects */
2750 sgen_weak_link_get (void **link_addr)
2752 void * volatile *link_addr_volatile;
2756 link_addr_volatile = link_addr;
2757 ptr = (void*)*link_addr_volatile;
2759 * At this point we have a hidden pointer. If the GC runs
2760 * here, it will not recognize the hidden pointer as a
2761 * reference, and if the object behind it is not referenced
2762 * elsewhere, it will be freed. Once the world is restarted
2763 * we reveal the pointer, giving us a pointer to a freed
2764 * object. To make sure we don't return it, we load the
2765 * hidden pointer again. If it's still the same, we can be
2766 * sure the object reference is valid.
2769 obj = (GCObject*) REVEAL_POINTER (ptr);
2773 mono_memory_barrier ();
2776 * During the second bridge processing step the world is
2777 * running again. That step processes all weak links once
2778 * more to null those that refer to dead objects. Before that
2779 * is completed, those links must not be followed, so we
2780 * conservatively wait for bridge processing when any weak
2781 * link is dereferenced.
2783 sgen_client_bridge_wait_for_processing ();
2785 if ((void*)*link_addr_volatile != ptr)
2792 sgen_set_allow_synchronous_major (gboolean flag)
2794 if (!major_collector.is_concurrent)
2797 allow_synchronous_major = flag;
2802 sgen_env_var_error (const char *env_var, const char *fallback, const char *description_format, ...)
2806 va_start (ap, description_format);
2808 fprintf (stderr, "Warning: In environment variable `%s': ", env_var);
2809 vfprintf (stderr, description_format, ap);
2811 fprintf (stderr, " - %s", fallback);
2812 fprintf (stderr, "\n");
2818 parse_double_in_interval (const char *env_var, const char *opt_name, const char *opt, double min, double max, double *result)
2821 double val = strtod (opt, &endptr);
2822 if (endptr == opt) {
2823 sgen_env_var_error (env_var, "Using default value.", "`%s` must be a number.", opt_name);
2826 else if (val < min || val > max) {
2827 sgen_env_var_error (env_var, "Using default value.", "`%s` must be between %.2f - %.2f.", opt_name, min, max);
2839 char *major_collector_opt = NULL;
2840 char *minor_collector_opt = NULL;
2841 size_t max_heap = 0;
2842 size_t soft_limit = 0;
2844 gboolean debug_print_allowance = FALSE;
2845 double allowance_ratio = 0, save_target = 0;
2846 gboolean cement_enabled = TRUE;
2849 result = InterlockedCompareExchange (&gc_initialized, -1, 0);
2852 /* already inited */
2855 /* being inited by another thread */
2859 /* we will init it */
2862 g_assert_not_reached ();
2864 } while (result != 0);
2866 SGEN_TV_GETTIME (sgen_init_timestamp);
2868 #ifdef SGEN_WITHOUT_MONO
2869 mono_thread_smr_init ();
2872 LOCK_INIT (gc_mutex);
2874 gc_debug_file = stderr;
2876 LOCK_INIT (sgen_interruption_mutex);
2878 if ((env = g_getenv (MONO_GC_PARAMS_NAME))) {
2879 opts = g_strsplit (env, ",", -1);
2880 for (ptr = opts; *ptr; ++ptr) {
2882 if (g_str_has_prefix (opt, "major=")) {
2883 opt = strchr (opt, '=') + 1;
2884 major_collector_opt = g_strdup (opt);
2885 } else if (g_str_has_prefix (opt, "minor=")) {
2886 opt = strchr (opt, '=') + 1;
2887 minor_collector_opt = g_strdup (opt);
2895 sgen_init_internal_allocator ();
2896 sgen_init_nursery_allocator ();
2897 sgen_init_fin_weak_hash ();
2898 sgen_init_hash_table ();
2899 sgen_init_descriptors ();
2900 sgen_init_gray_queues ();
2901 sgen_init_allocator ();
2903 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_SECTION, SGEN_SIZEOF_GC_MEM_SECTION);
2904 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_GRAY_QUEUE, sizeof (GrayQueueSection));
2906 sgen_client_init ();
2908 if (!minor_collector_opt) {
2909 sgen_simple_nursery_init (&sgen_minor_collector);
2911 if (!strcmp (minor_collector_opt, "simple")) {
2913 sgen_simple_nursery_init (&sgen_minor_collector);
2914 } else if (!strcmp (minor_collector_opt, "split")) {
2915 sgen_split_nursery_init (&sgen_minor_collector);
2917 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using `simple` instead.", "Unknown minor collector `%s'.", minor_collector_opt);
2918 goto use_simple_nursery;
2922 if (!major_collector_opt || !strcmp (major_collector_opt, "marksweep")) {
2923 use_marksweep_major:
2924 sgen_marksweep_init (&major_collector);
2925 } else if (!major_collector_opt || !strcmp (major_collector_opt, "marksweep-conc")) {
2926 sgen_marksweep_conc_init (&major_collector);
2928 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using `marksweep` instead.", "Unknown major collector `%s'.", major_collector_opt);
2929 goto use_marksweep_major;
2932 sgen_nursery_size = DEFAULT_NURSERY_SIZE;
2935 gboolean usage_printed = FALSE;
2937 for (ptr = opts; *ptr; ++ptr) {
2939 if (!strcmp (opt, ""))
2941 if (g_str_has_prefix (opt, "major="))
2943 if (g_str_has_prefix (opt, "minor="))
2945 if (g_str_has_prefix (opt, "max-heap-size=")) {
2946 size_t page_size = mono_pagesize ();
2947 size_t max_heap_candidate = 0;
2948 opt = strchr (opt, '=') + 1;
2949 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &max_heap_candidate)) {
2950 max_heap = (max_heap_candidate + page_size - 1) & ~(size_t)(page_size - 1);
2951 if (max_heap != max_heap_candidate)
2952 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Rounding up.", "`max-heap-size` size must be a multiple of %d.", page_size);
2954 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`max-heap-size` must be an integer.");
2958 if (g_str_has_prefix (opt, "soft-heap-limit=")) {
2959 opt = strchr (opt, '=') + 1;
2960 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &soft_limit)) {
2961 if (soft_limit <= 0) {
2962 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be positive.");
2966 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be an integer.");
2972 if (g_str_has_prefix (opt, "nursery-size=")) {
2974 opt = strchr (opt, '=') + 1;
2975 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &val)) {
2976 if ((val & (val - 1))) {
2977 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be a power of two.");
2981 if (val < SGEN_MAX_NURSERY_WASTE) {
2982 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.",
2983 "`nursery-size` must be at least %d bytes.", SGEN_MAX_NURSERY_WASTE);
2987 sgen_nursery_size = val;
2988 sgen_nursery_bits = 0;
2989 while (ONE_P << (++ sgen_nursery_bits) != sgen_nursery_size)
2992 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be an integer.");
2998 if (g_str_has_prefix (opt, "save-target-ratio=")) {
3000 opt = strchr (opt, '=') + 1;
3001 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "save-target-ratio", opt,
3002 SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO, &val)) {
3007 if (g_str_has_prefix (opt, "default-allowance-ratio=")) {
3009 opt = strchr (opt, '=') + 1;
3010 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "default-allowance-ratio", opt,
3011 SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, &val)) {
3012 allowance_ratio = val;
3016 if (g_str_has_prefix (opt, "allow-synchronous-major=")) {
3017 if (!major_collector.is_concurrent) {
3018 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring.", "`allow-synchronous-major` is only valid for the concurrent major collector.");
3022 opt = strchr (opt, '=') + 1;
3024 if (!strcmp (opt, "yes")) {
3025 allow_synchronous_major = TRUE;
3026 } else if (!strcmp (opt, "no")) {
3027 allow_synchronous_major = FALSE;
3029 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`allow-synchronous-major` must be either `yes' or `no'.");
3034 if (!strcmp (opt, "cementing")) {
3035 cement_enabled = TRUE;
3038 if (!strcmp (opt, "no-cementing")) {
3039 cement_enabled = FALSE;
3043 if (major_collector.handle_gc_param && major_collector.handle_gc_param (opt))
3046 if (sgen_minor_collector.handle_gc_param && sgen_minor_collector.handle_gc_param (opt))
3049 if (sgen_client_handle_gc_param (opt))
3052 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring.", "Unknown option `%s`.", opt);
3057 fprintf (stderr, "\n%s must be a comma-delimited list of one or more of the following:\n", MONO_GC_PARAMS_NAME);
3058 fprintf (stderr, " max-heap-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
3059 fprintf (stderr, " soft-heap-limit=n (where N is an integer, possibly with a k, m or a g suffix)\n");
3060 fprintf (stderr, " nursery-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
3061 fprintf (stderr, " major=COLLECTOR (where COLLECTOR is `marksweep', `marksweep-conc', `marksweep-par')\n");
3062 fprintf (stderr, " minor=COLLECTOR (where COLLECTOR is `simple' or `split')\n");
3063 fprintf (stderr, " wbarrier=WBARRIER (where WBARRIER is `remset' or `cardtable')\n");
3064 fprintf (stderr, " [no-]cementing\n");
3065 if (major_collector.is_concurrent)
3066 fprintf (stderr, " allow-synchronous-major=FLAG (where FLAG is `yes' or `no')\n");
3067 if (major_collector.print_gc_param_usage)
3068 major_collector.print_gc_param_usage ();
3069 if (sgen_minor_collector.print_gc_param_usage)
3070 sgen_minor_collector.print_gc_param_usage ();
3071 sgen_client_print_gc_params_usage ();
3072 fprintf (stderr, " Experimental options:\n");
3073 fprintf (stderr, " save-target-ratio=R (where R must be between %.2f - %.2f).\n", SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO);
3074 fprintf (stderr, " default-allowance-ratio=R (where R must be between %.2f - %.2f).\n", SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, SGEN_MAX_ALLOWANCE_NURSERY_SIZE_RATIO);
3075 fprintf (stderr, "\n");
3077 usage_printed = TRUE;
3082 if (major_collector_opt)
3083 g_free (major_collector_opt);
3085 if (minor_collector_opt)
3086 g_free (minor_collector_opt);
3090 sgen_cement_init (cement_enabled);
3092 if ((env = g_getenv (MONO_GC_DEBUG_NAME))) {
3093 gboolean usage_printed = FALSE;
3095 opts = g_strsplit (env, ",", -1);
3096 for (ptr = opts; ptr && *ptr; ptr ++) {
3098 if (!strcmp (opt, ""))
3100 if (opt [0] >= '0' && opt [0] <= '9') {
3101 gc_debug_level = atoi (opt);
3106 char *rf = g_strdup_printf ("%s.%d", opt, mono_process_current_pid ());
3107 gc_debug_file = fopen (rf, "wb");
3109 gc_debug_file = stderr;
3112 } else if (!strcmp (opt, "print-allowance")) {
3113 debug_print_allowance = TRUE;
3114 } else if (!strcmp (opt, "print-pinning")) {
3115 sgen_pin_stats_enable ();
3116 } else if (!strcmp (opt, "verify-before-allocs")) {
3117 verify_before_allocs = 1;
3118 has_per_allocation_action = TRUE;
3119 } else if (g_str_has_prefix (opt, "verify-before-allocs=")) {
3120 char *arg = strchr (opt, '=') + 1;
3121 verify_before_allocs = atoi (arg);
3122 has_per_allocation_action = TRUE;
3123 } else if (!strcmp (opt, "collect-before-allocs")) {
3124 collect_before_allocs = 1;
3125 has_per_allocation_action = TRUE;
3126 } else if (g_str_has_prefix (opt, "collect-before-allocs=")) {
3127 char *arg = strchr (opt, '=') + 1;
3128 has_per_allocation_action = TRUE;
3129 collect_before_allocs = atoi (arg);
3130 } else if (!strcmp (opt, "verify-before-collections")) {
3131 whole_heap_check_before_collection = TRUE;
3132 } else if (!strcmp (opt, "check-at-minor-collections")) {
3133 consistency_check_at_minor_collection = TRUE;
3134 nursery_clear_policy = CLEAR_AT_GC;
3135 } else if (!strcmp (opt, "mod-union-consistency-check")) {
3136 if (!major_collector.is_concurrent) {
3137 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`mod-union-consistency-check` only works with concurrent major collector.");
3140 mod_union_consistency_check = TRUE;
3141 } else if (!strcmp (opt, "check-mark-bits")) {
3142 check_mark_bits_after_major_collection = TRUE;
3143 } else if (!strcmp (opt, "check-nursery-pinned")) {
3144 check_nursery_objects_pinned = TRUE;
3145 } else if (!strcmp (opt, "clear-at-gc")) {
3146 nursery_clear_policy = CLEAR_AT_GC;
3147 } else if (!strcmp (opt, "clear-nursery-at-gc")) {
3148 nursery_clear_policy = CLEAR_AT_GC;
3149 } else if (!strcmp (opt, "clear-at-tlab-creation")) {
3150 nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
3151 } else if (!strcmp (opt, "debug-clear-at-tlab-creation")) {
3152 nursery_clear_policy = CLEAR_AT_TLAB_CREATION_DEBUG;
3153 } else if (!strcmp (opt, "check-scan-starts")) {
3154 do_scan_starts_check = TRUE;
3155 } else if (!strcmp (opt, "verify-nursery-at-minor-gc")) {
3156 do_verify_nursery = TRUE;
3157 } else if (!strcmp (opt, "check-concurrent")) {
3158 if (!major_collector.is_concurrent) {
3159 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`check-concurrent` only works with concurrent major collectors.");
3162 nursery_clear_policy = CLEAR_AT_GC;
3163 do_concurrent_checks = TRUE;
3164 } else if (!strcmp (opt, "dump-nursery-at-minor-gc")) {
3165 do_dump_nursery_content = TRUE;
3166 } else if (!strcmp (opt, "disable-minor")) {
3167 disable_minor_collections = TRUE;
3168 } else if (!strcmp (opt, "disable-major")) {
3169 disable_major_collections = TRUE;
3170 } else if (g_str_has_prefix (opt, "heap-dump=")) {
3171 char *filename = strchr (opt, '=') + 1;
3172 nursery_clear_policy = CLEAR_AT_GC;
3173 sgen_debug_enable_heap_dump (filename);
3174 } else if (g_str_has_prefix (opt, "binary-protocol=")) {
3175 char *filename = strchr (opt, '=') + 1;
3176 char *colon = strrchr (filename, ':');
3179 if (!mono_gc_parse_environment_string_extract_number (colon + 1, &limit)) {
3180 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring limit.", "Binary protocol file size limit must be an integer.");
3185 binary_protocol_init (filename, (long long)limit);
3186 } else if (!strcmp (opt, "nursery-canaries")) {
3187 do_verify_nursery = TRUE;
3188 enable_nursery_canaries = TRUE;
3189 } else if (!sgen_client_handle_gc_debug (opt)) {
3190 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "Unknown option `%s`.", opt);
3195 fprintf (stderr, "\n%s must be of the format [<l>[:<filename>]|<option>]+ where <l> is a debug level 0-9.\n", MONO_GC_DEBUG_NAME);
3196 fprintf (stderr, "Valid <option>s are:\n");
3197 fprintf (stderr, " collect-before-allocs[=<n>]\n");
3198 fprintf (stderr, " verify-before-allocs[=<n>]\n");
3199 fprintf (stderr, " check-at-minor-collections\n");
3200 fprintf (stderr, " check-mark-bits\n");
3201 fprintf (stderr, " check-nursery-pinned\n");
3202 fprintf (stderr, " verify-before-collections\n");
3203 fprintf (stderr, " verify-nursery-at-minor-gc\n");
3204 fprintf (stderr, " dump-nursery-at-minor-gc\n");
3205 fprintf (stderr, " disable-minor\n");
3206 fprintf (stderr, " disable-major\n");
3207 fprintf (stderr, " check-concurrent\n");
3208 fprintf (stderr, " clear-[nursery-]at-gc\n");
3209 fprintf (stderr, " clear-at-tlab-creation\n");
3210 fprintf (stderr, " debug-clear-at-tlab-creation\n");
3211 fprintf (stderr, " check-scan-starts\n");
3212 fprintf (stderr, " print-allowance\n");
3213 fprintf (stderr, " print-pinning\n");
3214 fprintf (stderr, " heap-dump=<filename>\n");
3215 fprintf (stderr, " binary-protocol=<filename>[:<file-size-limit>]\n");
3216 fprintf (stderr, " nursery-canaries\n");
3217 sgen_client_print_gc_debug_usage ();
3218 fprintf (stderr, "\n");
3220 usage_printed = TRUE;
3226 if (check_mark_bits_after_major_collection)
3227 nursery_clear_policy = CLEAR_AT_GC;
3229 if (major_collector.post_param_init)
3230 major_collector.post_param_init (&major_collector);
3232 if (major_collector.needs_thread_pool)
3233 sgen_workers_init (1);
3235 sgen_memgov_init (max_heap, soft_limit, debug_print_allowance, allowance_ratio, save_target);
3237 memset (&remset, 0, sizeof (remset));
3239 sgen_card_table_init (&remset);
3245 sgen_get_nursery_clear_policy (void)
3247 return nursery_clear_policy;
3257 sgen_gc_unlock (void)
3259 gboolean try_free = sgen_try_free_some_memory;
3260 sgen_try_free_some_memory = FALSE;
3261 mono_mutex_unlock (&gc_mutex);
3263 mono_thread_hazardous_try_free_some ();
3267 sgen_major_collector_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
3269 major_collector.iterate_live_block_ranges (callback);
3273 sgen_get_major_collector (void)
3275 return &major_collector;
3279 sgen_get_remset (void)
3285 count_cards (long long *major_total, long long *major_marked, long long *los_total, long long *los_marked)
3287 sgen_get_major_collector ()->count_cards (major_total, major_marked);
3288 sgen_los_count_cards (los_total, los_marked);
3291 static gboolean world_is_stopped = FALSE;
3293 /* LOCKING: assumes the GC lock is held */
3295 sgen_stop_world (int generation)
3297 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3299 SGEN_ASSERT (0, !world_is_stopped, "Why are we stopping a stopped world?");
3301 binary_protocol_world_stopping (generation, sgen_timestamp (), (gpointer)mono_native_thread_id_get ());
3303 sgen_client_stop_world (generation);
3305 world_is_stopped = TRUE;
3307 if (binary_protocol_is_heavy_enabled ())
3308 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3309 binary_protocol_world_stopped (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3312 /* LOCKING: assumes the GC lock is held */
3314 sgen_restart_world (int generation, GGTimingInfo *timing)
3316 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3318 SGEN_ASSERT (0, world_is_stopped, "Why are we restarting a running world?");
3320 if (binary_protocol_is_heavy_enabled ())
3321 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3322 binary_protocol_world_restarting (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3324 sgen_client_restart_world (generation, timing);
3326 world_is_stopped = FALSE;
3328 binary_protocol_world_restarted (generation, sgen_timestamp ());
3330 sgen_try_free_some_memory = TRUE;
3332 if (sgen_client_bridge_need_processing ())
3333 sgen_client_bridge_processing_finish (generation);
3335 sgen_memgov_collection_end (generation, timing, timing ? 2 : 0);
3339 sgen_is_world_stopped (void)
3341 return world_is_stopped;
3345 sgen_check_whole_heap_stw (void)
3347 sgen_stop_world (0);
3348 sgen_clear_nursery_fragments ();
3349 sgen_check_whole_heap (FALSE);
3350 sgen_restart_world (0, NULL);
3354 sgen_timestamp (void)
3356 SGEN_TV_DECLARE (timestamp);
3357 SGEN_TV_GETTIME (timestamp);
3358 return SGEN_TV_ELAPSED (sgen_init_timestamp, timestamp);
3361 #endif /* HAVE_SGEN_GC */