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 SGEN_ASSERT (0, ((mword)start & (SIZEOF_VOID_P - 1)) == 0, "Why are we scanning for references in unaligned memory ?");
827 #if defined(VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE) && !defined(_WIN64)
828 VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE (start, (char*)end - (char*)start);
831 while (start < end) {
833 * *start can point to the middle of an object
834 * note: should we handle pointing at the end of an object?
835 * pinning in C# code disallows pointing at the end of an object
836 * but there is some small chance that an optimizing C compiler
837 * may keep the only reference to an object by pointing
838 * at the end of it. We ignore this small chance for now.
839 * Pointers to the end of an object are indistinguishable
840 * from pointers to the start of the next object in memory
841 * so if we allow that we'd need to pin two objects...
842 * We queue the pointer in an array, the
843 * array will then be sorted and uniqued. This way
844 * we can coalesce several pinning pointers and it should
845 * be faster since we'd do a memory scan with increasing
846 * addresses. Note: we can align the address to the allocation
847 * alignment, so the unique process is more effective.
849 mword addr = (mword)*start;
850 addr &= ~(ALLOC_ALIGN - 1);
851 if (addr >= (mword)start_nursery && addr < (mword)end_nursery) {
852 SGEN_LOG (6, "Pinning address %p from %p", (void*)addr, start);
853 sgen_pin_stage_ptr ((void*)addr);
854 binary_protocol_pin_stage (start, (void*)addr);
855 sgen_pin_stats_register_address ((char*)addr, pin_type);
861 SGEN_LOG (7, "found %d potential pinned heap pointers", count);
865 * The first thing we do in a collection is to identify pinned objects.
866 * This function considers all the areas of memory that need to be
867 * conservatively scanned.
870 pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx)
874 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);
875 /* objects pinned from the API are inside these roots */
876 SGEN_HASH_TABLE_FOREACH (&roots_hash [ROOT_TYPE_PINNED], start_root, root) {
877 SGEN_LOG (6, "Pinned roots %p-%p", start_root, root->end_root);
878 sgen_conservatively_pin_objects_from (start_root, (void**)root->end_root, start_nursery, end_nursery, PIN_TYPE_OTHER);
879 } SGEN_HASH_TABLE_FOREACH_END;
880 /* now deal with the thread stacks
881 * in the future we should be able to conservatively scan only:
882 * *) the cpu registers
883 * *) the unmanaged stack frames
884 * *) the _last_ managed stack frame
885 * *) pointers slots in managed frames
887 sgen_client_scan_thread_data (start_nursery, end_nursery, FALSE, ctx);
891 single_arg_user_copy_or_mark (GCObject **obj, void *gc_data)
893 ScanCopyContext *ctx = gc_data;
894 ctx->ops->copy_or_mark_object (obj, ctx->queue);
898 * The memory area from start_root to end_root contains pointers to objects.
899 * Their position is precisely described by @desc (this means that the pointer
900 * can be either NULL or the pointer to the start of an object).
901 * This functions copies them to to_space updates them.
903 * This function is not thread-safe!
906 precisely_scan_objects_from (void** start_root, void** end_root, char* n_start, char *n_end, SgenDescriptor desc, ScanCopyContext ctx)
908 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
909 SgenGrayQueue *queue = ctx.queue;
911 switch (desc & ROOT_DESC_TYPE_MASK) {
912 case ROOT_DESC_BITMAP:
913 desc >>= ROOT_DESC_TYPE_SHIFT;
915 if ((desc & 1) && *start_root) {
916 copy_func ((GCObject**)start_root, queue);
917 SGEN_LOG (9, "Overwrote root at %p with %p", start_root, *start_root);
923 case ROOT_DESC_COMPLEX: {
924 gsize *bitmap_data = sgen_get_complex_descriptor_bitmap (desc);
925 gsize bwords = (*bitmap_data) - 1;
926 void **start_run = start_root;
928 while (bwords-- > 0) {
929 gsize bmap = *bitmap_data++;
930 void **objptr = start_run;
932 if ((bmap & 1) && *objptr) {
933 copy_func ((GCObject**)objptr, queue);
934 SGEN_LOG (9, "Overwrote root at %p with %p", objptr, *objptr);
939 start_run += GC_BITS_PER_WORD;
943 case ROOT_DESC_USER: {
944 SgenUserRootMarkFunc marker = sgen_get_user_descriptor_func (desc);
945 marker (start_root, single_arg_user_copy_or_mark, &ctx);
948 case ROOT_DESC_RUN_LEN:
949 g_assert_not_reached ();
951 g_assert_not_reached ();
956 reset_heap_boundaries (void)
958 lowest_heap_address = ~(mword)0;
959 highest_heap_address = 0;
963 sgen_update_heap_boundaries (mword low, mword high)
968 old = lowest_heap_address;
971 } while (SGEN_CAS_PTR ((gpointer*)&lowest_heap_address, (gpointer)low, (gpointer)old) != (gpointer)old);
974 old = highest_heap_address;
977 } while (SGEN_CAS_PTR ((gpointer*)&highest_heap_address, (gpointer)high, (gpointer)old) != (gpointer)old);
981 * Allocate and setup the data structures needed to be able to allocate objects
982 * in the nursery. The nursery is stored in nursery_section.
987 GCMemSection *section;
994 SGEN_LOG (2, "Allocating nursery size: %zu", (size_t)sgen_nursery_size);
995 /* later we will alloc a larger area for the nursery but only activate
996 * what we need. The rest will be used as expansion if we have too many pinned
997 * objects in the existing nursery.
999 /* FIXME: handle OOM */
1000 section = sgen_alloc_internal (INTERNAL_MEM_SECTION);
1002 alloc_size = sgen_nursery_size;
1004 /* If there isn't enough space even for the nursery we should simply abort. */
1005 g_assert (sgen_memgov_try_alloc_space (alloc_size, SPACE_NURSERY));
1007 data = major_collector.alloc_heap (alloc_size, alloc_size, DEFAULT_NURSERY_BITS);
1008 sgen_update_heap_boundaries ((mword)data, (mword)(data + sgen_nursery_size));
1009 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 ());
1010 section->data = section->next_data = data;
1011 section->size = alloc_size;
1012 section->end_data = data + sgen_nursery_size;
1013 scan_starts = (alloc_size + SCAN_START_SIZE - 1) / SCAN_START_SIZE;
1014 section->scan_starts = sgen_alloc_internal_dynamic (sizeof (char*) * scan_starts, INTERNAL_MEM_SCAN_STARTS, TRUE);
1015 section->num_scan_start = scan_starts;
1017 nursery_section = section;
1019 sgen_nursery_allocator_set_nursery_bounds (data, data + sgen_nursery_size);
1023 mono_gc_get_logfile (void)
1025 return gc_debug_file;
1029 scan_finalizer_entries (SgenPointerQueue *fin_queue, ScanCopyContext ctx)
1031 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
1032 SgenGrayQueue *queue = ctx.queue;
1035 for (i = 0; i < fin_queue->next_slot; ++i) {
1036 GCObject *obj = fin_queue->data [i];
1039 SGEN_LOG (5, "Scan of fin ready object: %p (%s)\n", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
1040 copy_func ((GCObject**)&fin_queue->data [i], queue);
1045 generation_name (int generation)
1047 switch (generation) {
1048 case GENERATION_NURSERY: return "nursery";
1049 case GENERATION_OLD: return "old";
1050 default: g_assert_not_reached ();
1055 sgen_generation_name (int generation)
1057 return generation_name (generation);
1061 finish_gray_stack (int generation, ScanCopyContext ctx)
1065 int done_with_ephemerons, ephemeron_rounds = 0;
1066 char *start_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_start () : NULL;
1067 char *end_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_end () : (char*)-1;
1068 SgenGrayQueue *queue = ctx.queue;
1071 * We copied all the reachable objects. Now it's the time to copy
1072 * the objects that were not referenced by the roots, but by the copied objects.
1073 * we built a stack of objects pointed to by gray_start: they are
1074 * additional roots and we may add more items as we go.
1075 * We loop until gray_start == gray_objects which means no more objects have
1076 * been added. Note this is iterative: no recursion is involved.
1077 * We need to walk the LO list as well in search of marked big objects
1078 * (use a flag since this is needed only on major collections). We need to loop
1079 * here as well, so keep a counter of marked LO (increasing it in copy_object).
1080 * To achieve better cache locality and cache usage, we drain the gray stack
1081 * frequently, after each object is copied, and just finish the work here.
1083 sgen_drain_gray_stack (-1, ctx);
1085 SGEN_LOG (2, "%s generation done", generation_name (generation));
1088 Reset bridge data, we might have lingering data from a previous collection if this is a major
1089 collection trigged by minor overflow.
1091 We must reset the gathered bridges since their original block might be evacuated due to major
1092 fragmentation in the meanwhile and the bridge code should not have to deal with that.
1094 if (sgen_client_bridge_need_processing ())
1095 sgen_client_bridge_reset_data ();
1098 * Walk the ephemeron tables marking all values with reachable keys. This must be completely done
1099 * before processing finalizable objects and non-tracking weak links to avoid finalizing/clearing
1100 * objects that are in fact reachable.
1102 done_with_ephemerons = 0;
1104 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1105 sgen_drain_gray_stack (-1, ctx);
1107 } while (!done_with_ephemerons);
1109 sgen_client_mark_togglerefs (start_addr, end_addr, ctx);
1111 if (sgen_client_bridge_need_processing ()) {
1112 /*Make sure the gray stack is empty before we process bridge objects so we get liveness right*/
1113 sgen_drain_gray_stack (-1, ctx);
1114 sgen_collect_bridge_objects (generation, ctx);
1115 if (generation == GENERATION_OLD)
1116 sgen_collect_bridge_objects (GENERATION_NURSERY, ctx);
1119 Do the first bridge step here, as the collector liveness state will become useless after that.
1121 An important optimization is to only proccess the possibly dead part of the object graph and skip
1122 over all live objects as we transitively know everything they point must be alive too.
1124 The above invariant is completely wrong if we let the gray queue be drained and mark/copy everything.
1126 This has the unfortunate side effect of making overflow collections perform the first step twice, but
1127 given we now have heuristics that perform major GC in anticipation of minor overflows this should not
1130 sgen_client_bridge_processing_stw_step ();
1134 Make sure we drain the gray stack before processing disappearing links and finalizers.
1135 If we don't make sure it is empty we might wrongly see a live object as dead.
1137 sgen_drain_gray_stack (-1, ctx);
1140 We must clear weak links that don't track resurrection before processing object ready for
1141 finalization so they can be cleared before that.
1143 sgen_null_link_in_range (generation, ctx, FALSE);
1144 if (generation == GENERATION_OLD)
1145 sgen_null_link_in_range (GENERATION_NURSERY, ctx, FALSE);
1148 /* walk the finalization queue and move also the objects that need to be
1149 * finalized: use the finalized objects as new roots so the objects they depend
1150 * on are also not reclaimed. As with the roots above, only objects in the nursery
1151 * are marked/copied.
1153 sgen_finalize_in_range (generation, ctx);
1154 if (generation == GENERATION_OLD)
1155 sgen_finalize_in_range (GENERATION_NURSERY, ctx);
1156 /* drain the new stack that might have been created */
1157 SGEN_LOG (6, "Precise scan of gray area post fin");
1158 sgen_drain_gray_stack (-1, ctx);
1161 * This must be done again after processing finalizable objects since CWL slots are cleared only after the key is finalized.
1163 done_with_ephemerons = 0;
1165 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1166 sgen_drain_gray_stack (-1, ctx);
1168 } while (!done_with_ephemerons);
1170 sgen_client_clear_unreachable_ephemerons (ctx);
1173 * We clear togglerefs only after all possible chances of revival are done.
1174 * This is semantically more inline with what users expect and it allows for
1175 * user finalizers to correctly interact with TR objects.
1177 sgen_client_clear_togglerefs (start_addr, end_addr, ctx);
1180 SGEN_LOG (2, "Finalize queue handling scan for %s generation: %ld usecs %d ephemeron rounds", generation_name (generation), TV_ELAPSED (atv, btv), ephemeron_rounds);
1183 * handle disappearing links
1184 * Note we do this after checking the finalization queue because if an object
1185 * survives (at least long enough to be finalized) we don't clear the link.
1186 * This also deals with a possible issue with the monitor reclamation: with the Boehm
1187 * GC a finalized object my lose the monitor because it is cleared before the finalizer is
1190 g_assert (sgen_gray_object_queue_is_empty (queue));
1192 sgen_null_link_in_range (generation, ctx, TRUE);
1193 if (generation == GENERATION_OLD)
1194 sgen_null_link_in_range (GENERATION_NURSERY, ctx, TRUE);
1195 if (sgen_gray_object_queue_is_empty (queue))
1197 sgen_drain_gray_stack (-1, ctx);
1200 g_assert (sgen_gray_object_queue_is_empty (queue));
1202 sgen_gray_object_queue_trim_free_list (queue);
1206 sgen_check_section_scan_starts (GCMemSection *section)
1209 for (i = 0; i < section->num_scan_start; ++i) {
1210 if (section->scan_starts [i]) {
1211 mword size = safe_object_get_size ((GCObject*) section->scan_starts [i]);
1212 SGEN_ASSERT (0, size >= SGEN_CLIENT_MINIMUM_OBJECT_SIZE && size <= MAX_SMALL_OBJ_SIZE, "Weird object size at scan starts.");
1218 check_scan_starts (void)
1220 if (!do_scan_starts_check)
1222 sgen_check_section_scan_starts (nursery_section);
1223 major_collector.check_scan_starts ();
1227 scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx)
1231 SGEN_HASH_TABLE_FOREACH (&roots_hash [root_type], start_root, root) {
1232 SGEN_LOG (6, "Precise root scan %p-%p (desc: %p)", start_root, root->end_root, (void*)root->root_desc);
1233 precisely_scan_objects_from (start_root, (void**)root->end_root, addr_start, addr_end, root->root_desc, ctx);
1234 } SGEN_HASH_TABLE_FOREACH_END;
1240 static gboolean inited = FALSE;
1245 mono_counters_register ("Collection max time", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME | MONO_COUNTER_MONOTONIC, &time_max);
1247 mono_counters_register ("Minor fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pre_collection_fragment_clear);
1248 mono_counters_register ("Minor pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pinning);
1249 mono_counters_register ("Minor scan remembered set", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_remsets);
1250 mono_counters_register ("Minor scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_pinned);
1251 mono_counters_register ("Minor scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_roots);
1252 mono_counters_register ("Minor fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_fragment_creation);
1254 mono_counters_register ("Major fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pre_collection_fragment_clear);
1255 mono_counters_register ("Major pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pinning);
1256 mono_counters_register ("Major scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_pinned);
1257 mono_counters_register ("Major scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_roots);
1258 mono_counters_register ("Major scan mod union", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_mod_union);
1259 mono_counters_register ("Major finish gray stack", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_finish_gray_stack);
1260 mono_counters_register ("Major free big objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_free_bigobjs);
1261 mono_counters_register ("Major LOS sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_los_sweep);
1262 mono_counters_register ("Major sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_sweep);
1263 mono_counters_register ("Major fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_fragment_creation);
1265 mono_counters_register ("Number of pinned objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_pinned_objects);
1267 #ifdef HEAVY_STATISTICS
1268 mono_counters_register ("WBarrier remember pointer", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_add_to_global_remset);
1269 mono_counters_register ("WBarrier arrayref copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_arrayref_copy);
1270 mono_counters_register ("WBarrier generic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store);
1271 mono_counters_register ("WBarrier generic atomic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store_atomic);
1272 mono_counters_register ("WBarrier set root", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_set_root);
1274 mono_counters_register ("# objects allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_alloced_degraded);
1275 mono_counters_register ("bytes allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced_degraded);
1277 mono_counters_register ("# copy_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_nursery);
1278 mono_counters_register ("# objects copied (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_nursery);
1279 mono_counters_register ("# copy_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_major);
1280 mono_counters_register ("# objects copied (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_major);
1282 mono_counters_register ("# scan_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_nursery);
1283 mono_counters_register ("# scan_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_major);
1285 mono_counters_register ("Slots allocated in vain", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_slots_allocated_in_vain);
1287 mono_counters_register ("# nursery copy_object() failed from space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_from_space);
1288 mono_counters_register ("# nursery copy_object() failed forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_forwarded);
1289 mono_counters_register ("# nursery copy_object() failed pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_pinned);
1290 mono_counters_register ("# nursery copy_object() failed to space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_to_space);
1292 sgen_nursery_allocator_init_heavy_stats ();
1300 reset_pinned_from_failed_allocation (void)
1302 bytes_pinned_from_failed_allocation = 0;
1306 sgen_set_pinned_from_failed_allocation (mword objsize)
1308 bytes_pinned_from_failed_allocation += objsize;
1312 sgen_collection_is_concurrent (void)
1314 switch (current_collection_generation) {
1315 case GENERATION_NURSERY:
1317 case GENERATION_OLD:
1318 return concurrent_collection_in_progress;
1320 g_error ("Invalid current generation %d", current_collection_generation);
1326 sgen_concurrent_collection_in_progress (void)
1328 return concurrent_collection_in_progress;
1332 SgenThreadPoolJob job;
1333 SgenObjectOperations *ops;
1337 job_remembered_set_scan (void *worker_data_untyped, SgenThreadPoolJob *job)
1339 WorkerData *worker_data = worker_data_untyped;
1340 ScanJob *job_data = (ScanJob*)job;
1341 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1342 remset.scan_remsets (ctx);
1346 SgenThreadPoolJob job;
1347 SgenObjectOperations *ops;
1351 } ScanFromRegisteredRootsJob;
1354 job_scan_from_registered_roots (void *worker_data_untyped, SgenThreadPoolJob *job)
1356 WorkerData *worker_data = worker_data_untyped;
1357 ScanFromRegisteredRootsJob *job_data = (ScanFromRegisteredRootsJob*)job;
1358 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1360 scan_from_registered_roots (job_data->heap_start, job_data->heap_end, job_data->root_type, ctx);
1364 SgenThreadPoolJob job;
1365 SgenObjectOperations *ops;
1368 } ScanThreadDataJob;
1371 job_scan_thread_data (void *worker_data_untyped, SgenThreadPoolJob *job)
1373 WorkerData *worker_data = worker_data_untyped;
1374 ScanThreadDataJob *job_data = (ScanThreadDataJob*)job;
1375 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1377 sgen_client_scan_thread_data (job_data->heap_start, job_data->heap_end, TRUE, ctx);
1381 SgenThreadPoolJob job;
1382 SgenObjectOperations *ops;
1383 SgenPointerQueue *queue;
1384 } ScanFinalizerEntriesJob;
1387 job_scan_finalizer_entries (void *worker_data_untyped, SgenThreadPoolJob *job)
1389 WorkerData *worker_data = worker_data_untyped;
1390 ScanFinalizerEntriesJob *job_data = (ScanFinalizerEntriesJob*)job;
1391 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1393 scan_finalizer_entries (job_data->queue, ctx);
1397 job_scan_major_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1399 WorkerData *worker_data = worker_data_untyped;
1400 ScanJob *job_data = (ScanJob*)job;
1401 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1403 g_assert (concurrent_collection_in_progress);
1404 major_collector.scan_card_table (TRUE, ctx);
1408 job_scan_los_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1410 WorkerData *worker_data = worker_data_untyped;
1411 ScanJob *job_data = (ScanJob*)job;
1412 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1414 g_assert (concurrent_collection_in_progress);
1415 sgen_los_scan_card_table (TRUE, ctx);
1419 init_gray_queue (void)
1421 if (sgen_collection_is_concurrent ())
1422 sgen_workers_init_distribute_gray_queue ();
1423 sgen_gray_object_queue_init (&gray_queue, NULL);
1427 enqueue_scan_from_roots_jobs (char *heap_start, char *heap_end, SgenObjectOperations *ops)
1429 ScanFromRegisteredRootsJob *scrrj;
1430 ScanThreadDataJob *stdj;
1431 ScanFinalizerEntriesJob *sfej;
1433 /* registered roots, this includes static fields */
1435 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots normal", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1437 scrrj->heap_start = heap_start;
1438 scrrj->heap_end = heap_end;
1439 scrrj->root_type = ROOT_TYPE_NORMAL;
1440 sgen_workers_enqueue_job (&scrrj->job);
1442 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots wbarrier", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1444 scrrj->heap_start = heap_start;
1445 scrrj->heap_end = heap_end;
1446 scrrj->root_type = ROOT_TYPE_WBARRIER;
1447 sgen_workers_enqueue_job (&scrrj->job);
1451 stdj = (ScanThreadDataJob*)sgen_thread_pool_job_alloc ("scan thread data", job_scan_thread_data, sizeof (ScanThreadDataJob));
1452 stdj->heap_start = heap_start;
1453 stdj->heap_end = heap_end;
1454 sgen_workers_enqueue_job (&stdj->job);
1456 /* Scan the list of objects ready for finalization. */
1458 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1459 sfej->queue = &fin_ready_queue;
1461 sgen_workers_enqueue_job (&sfej->job);
1463 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan critical finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1464 sfej->queue = &critical_fin_queue;
1466 sgen_workers_enqueue_job (&sfej->job);
1470 * Perform a nursery collection.
1472 * Return whether any objects were late-pinned due to being out of memory.
1475 collect_nursery (SgenGrayQueue *unpin_queue, gboolean finish_up_concurrent_mark)
1477 gboolean needs_major;
1478 size_t max_garbage_amount;
1480 mword fragment_total;
1482 SgenObjectOperations *object_ops = &sgen_minor_collector.serial_ops;
1483 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, &gray_queue);
1487 if (disable_minor_collections)
1490 TV_GETTIME (last_minor_collection_start_tv);
1491 atv = last_minor_collection_start_tv;
1493 binary_protocol_collection_begin (gc_stats.minor_gc_count, GENERATION_NURSERY);
1495 if (do_verify_nursery || do_dump_nursery_content)
1496 sgen_debug_verify_nursery (do_dump_nursery_content);
1498 current_collection_generation = GENERATION_NURSERY;
1500 SGEN_ASSERT (0, !sgen_collection_is_concurrent (), "Why is the nursery collection concurrent?");
1502 reset_pinned_from_failed_allocation ();
1504 check_scan_starts ();
1506 sgen_nursery_alloc_prepare_for_minor ();
1510 nursery_next = sgen_nursery_alloc_get_upper_alloc_bound ();
1511 /* FIXME: optimize later to use the higher address where an object can be present */
1512 nursery_next = MAX (nursery_next, sgen_get_nursery_end ());
1514 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 ()));
1515 max_garbage_amount = nursery_next - sgen_get_nursery_start ();
1516 g_assert (nursery_section->size >= max_garbage_amount);
1518 /* world must be stopped already */
1520 time_minor_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1522 sgen_client_pre_collection_checks ();
1524 nursery_section->next_data = nursery_next;
1526 major_collector.start_nursery_collection ();
1528 sgen_memgov_minor_collection_start ();
1532 gc_stats.minor_gc_count ++;
1534 if (whole_heap_check_before_collection) {
1535 sgen_clear_nursery_fragments ();
1536 sgen_check_whole_heap (finish_up_concurrent_mark);
1538 if (consistency_check_at_minor_collection)
1539 sgen_check_consistency ();
1541 sgen_process_fin_stage_entries ();
1543 /* pin from pinned handles */
1544 sgen_init_pinning ();
1545 sgen_client_binary_protocol_mark_start (GENERATION_NURSERY);
1546 pin_from_roots (sgen_get_nursery_start (), nursery_next, ctx);
1547 /* pin cemented objects */
1548 sgen_pin_cemented_objects ();
1549 /* identify pinned objects */
1550 sgen_optimize_pin_queue ();
1551 sgen_pinning_setup_section (nursery_section);
1553 pin_objects_in_nursery (FALSE, ctx);
1554 sgen_pinning_trim_queue_to_section (nursery_section);
1557 time_minor_pinning += TV_ELAPSED (btv, atv);
1558 SGEN_LOG (2, "Finding pinned pointers: %zd in %ld usecs", sgen_get_pinned_count (), TV_ELAPSED (btv, atv));
1559 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
1562 * FIXME: When we finish a concurrent collection we do a nursery collection first,
1563 * as part of which we scan the card table. Then, later, we scan the mod union
1564 * cardtable. We should only have to do one.
1566 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan remset", job_remembered_set_scan, sizeof (ScanJob));
1567 sj->ops = object_ops;
1568 sgen_workers_enqueue_job (&sj->job);
1570 /* we don't have complete write barrier yet, so we scan all the old generation sections */
1572 time_minor_scan_remsets += TV_ELAPSED (atv, btv);
1573 SGEN_LOG (2, "Old generation scan: %ld usecs", TV_ELAPSED (atv, btv));
1575 sgen_pin_stats_print_class_stats ();
1577 sgen_drain_gray_stack (-1, ctx);
1579 /* FIXME: Why do we do this at this specific, seemingly random, point? */
1580 sgen_client_collecting_minor (&fin_ready_queue, &critical_fin_queue);
1583 time_minor_scan_pinned += TV_ELAPSED (btv, atv);
1585 enqueue_scan_from_roots_jobs (sgen_get_nursery_start (), nursery_next, object_ops);
1588 time_minor_scan_roots += TV_ELAPSED (atv, btv);
1590 finish_gray_stack (GENERATION_NURSERY, ctx);
1593 time_minor_finish_gray_stack += TV_ELAPSED (btv, atv);
1594 sgen_client_binary_protocol_mark_end (GENERATION_NURSERY);
1596 if (objects_pinned) {
1597 sgen_optimize_pin_queue ();
1598 sgen_pinning_setup_section (nursery_section);
1601 /* walk the pin_queue, build up the fragment list of free memory, unmark
1602 * pinned objects as we go, memzero() the empty fragments so they are ready for the
1605 sgen_client_binary_protocol_reclaim_start (GENERATION_NURSERY);
1606 fragment_total = sgen_build_nursery_fragments (nursery_section, unpin_queue);
1607 if (!fragment_total)
1610 /* Clear TLABs for all threads */
1611 sgen_clear_tlabs ();
1613 sgen_client_binary_protocol_reclaim_end (GENERATION_NURSERY);
1615 time_minor_fragment_creation += TV_ELAPSED (atv, btv);
1616 SGEN_LOG (2, "Fragment creation: %ld usecs, %lu bytes available", TV_ELAPSED (atv, btv), (unsigned long)fragment_total);
1618 if (consistency_check_at_minor_collection)
1619 sgen_check_major_refs ();
1621 major_collector.finish_nursery_collection ();
1623 TV_GETTIME (last_minor_collection_end_tv);
1624 gc_stats.minor_gc_time += TV_ELAPSED (last_minor_collection_start_tv, last_minor_collection_end_tv);
1626 sgen_debug_dump_heap ("minor", gc_stats.minor_gc_count - 1, NULL);
1628 /* prepare the pin queue for the next collection */
1629 sgen_finish_pinning ();
1630 if (sgen_have_pending_finalizers ()) {
1631 SGEN_LOG (4, "Finalizer-thread wakeup");
1632 sgen_client_finalize_notify ();
1634 sgen_pin_stats_reset ();
1635 /* clear cemented hash */
1636 sgen_cement_clear_below_threshold ();
1638 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
1640 remset.finish_minor_collection ();
1642 check_scan_starts ();
1644 binary_protocol_flush_buffers (FALSE);
1646 sgen_memgov_minor_collection_end ();
1648 /*objects are late pinned because of lack of memory, so a major is a good call*/
1649 needs_major = objects_pinned > 0;
1650 current_collection_generation = -1;
1653 binary_protocol_collection_end (gc_stats.minor_gc_count - 1, GENERATION_NURSERY, 0, 0);
1655 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
1656 sgen_check_nursery_objects_pinned (unpin_queue != NULL);
1662 COPY_OR_MARK_FROM_ROOTS_SERIAL,
1663 COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT,
1664 COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT
1665 } CopyOrMarkFromRootsMode;
1668 major_copy_or_mark_from_roots (size_t *old_next_pin_slot, CopyOrMarkFromRootsMode mode, SgenObjectOperations *object_ops)
1673 /* FIXME: only use these values for the precise scan
1674 * note that to_space pointers should be excluded anyway...
1676 char *heap_start = NULL;
1677 char *heap_end = (char*)-1;
1678 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, WORKERS_DISTRIBUTE_GRAY_QUEUE);
1679 gboolean concurrent = mode != COPY_OR_MARK_FROM_ROOTS_SERIAL;
1681 SGEN_ASSERT (0, !!concurrent == !!concurrent_collection_in_progress, "We've been called with the wrong mode.");
1683 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1684 /*This cleans up unused fragments */
1685 sgen_nursery_allocator_prepare_for_pinning ();
1687 if (do_concurrent_checks)
1688 sgen_debug_check_nursery_is_clean ();
1690 /* The concurrent collector doesn't touch the nursery. */
1691 sgen_nursery_alloc_prepare_for_major ();
1698 /* Pinning depends on this */
1699 sgen_clear_nursery_fragments ();
1701 if (whole_heap_check_before_collection)
1702 sgen_check_whole_heap (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT);
1705 time_major_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1707 if (!sgen_collection_is_concurrent ())
1708 nursery_section->next_data = sgen_get_nursery_end ();
1709 /* we should also coalesce scanning from sections close to each other
1710 * and deal with pointers outside of the sections later.
1715 sgen_client_pre_collection_checks ();
1718 /* Remsets are not useful for a major collection */
1719 remset.clear_cards ();
1722 sgen_process_fin_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", (long)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 major_update_concurrent_collection ();
2260 if (generation_to_collect == GENERATION_NURSERY)
2261 collect_nursery (NULL, FALSE);
2268 * If we've been asked to do a major collection, and the major collector wants to
2269 * run synchronously (to evacuate), we set the flag to do that.
2271 if (generation_to_collect == GENERATION_OLD &&
2272 allow_synchronous_major &&
2273 major_collector.want_synchronous_collection &&
2274 *major_collector.want_synchronous_collection) {
2275 wait_to_finish = TRUE;
2278 SGEN_ASSERT (0, !concurrent_collection_in_progress, "Why did this not get handled above?");
2281 * There's no concurrent collection in progress. Collect the generation we're asked
2282 * to collect. If the major collector is concurrent and we're not forced to wait,
2283 * start a concurrent collection.
2285 // FIXME: extract overflow reason
2286 if (generation_to_collect == GENERATION_NURSERY) {
2287 if (collect_nursery (NULL, FALSE)) {
2288 overflow_generation_to_collect = GENERATION_OLD;
2289 overflow_reason = "Minor overflow";
2292 if (major_collector.is_concurrent && !wait_to_finish) {
2293 collect_nursery (NULL, FALSE);
2294 major_start_concurrent_collection (reason);
2295 // FIXME: set infos[0] properly
2299 if (major_do_collection (reason, wait_to_finish)) {
2300 overflow_generation_to_collect = GENERATION_NURSERY;
2301 overflow_reason = "Excessive pinning";
2305 TV_GETTIME (gc_end);
2307 memset (infos, 0, sizeof (infos));
2308 infos [0].generation = generation_to_collect;
2309 infos [0].reason = reason;
2310 infos [0].is_overflow = FALSE;
2311 infos [1].generation = -1;
2312 infos [0].total_time = SGEN_TV_ELAPSED (gc_start, gc_end);
2314 SGEN_ASSERT (0, !concurrent_collection_in_progress, "Why did this not get handled above?");
2316 if (overflow_generation_to_collect != -1) {
2318 * We need to do an overflow collection, either because we ran out of memory
2319 * or the nursery is fully pinned.
2322 infos [1].generation = overflow_generation_to_collect;
2323 infos [1].reason = overflow_reason;
2324 infos [1].is_overflow = TRUE;
2327 if (overflow_generation_to_collect == GENERATION_NURSERY)
2328 collect_nursery (NULL, FALSE);
2330 major_do_collection (overflow_reason, wait_to_finish);
2332 TV_GETTIME (gc_end);
2333 infos [1].total_time = SGEN_TV_ELAPSED (gc_start, gc_end);
2335 oldest_generation_collected = MAX (oldest_generation_collected, overflow_generation_to_collect);
2338 SGEN_LOG (2, "Heap size: %lu, LOS size: %lu", (unsigned long)sgen_gc_get_total_heap_allocation (), (unsigned long)los_memory_usage);
2340 /* this also sets the proper pointers for the next allocation */
2341 if (generation_to_collect == GENERATION_NURSERY && !sgen_can_alloc_size (requested_size)) {
2342 /* TypeBuilder and MonoMethod are killing mcs with fragmentation */
2343 SGEN_LOG (1, "nursery collection didn't find enough room for %zd alloc (%zd pinned)", requested_size, sgen_get_pinned_count ());
2344 sgen_dump_pin_queue ();
2349 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
2351 TV_GETTIME (gc_total_end);
2352 time_max = MAX (time_max, TV_ELAPSED (gc_total_start, gc_total_end));
2354 sgen_restart_world (oldest_generation_collected, infos);
2358 * ######################################################################
2359 * ######## Memory allocation from the OS
2360 * ######################################################################
2361 * This section of code deals with getting memory from the OS and
2362 * allocating memory for GC-internal data structures.
2363 * Internal memory can be handled with a freelist for small objects.
2369 G_GNUC_UNUSED static void
2370 report_internal_mem_usage (void)
2372 printf ("Internal memory usage:\n");
2373 sgen_report_internal_mem_usage ();
2374 printf ("Pinned memory usage:\n");
2375 major_collector.report_pinned_memory_usage ();
2379 * ######################################################################
2380 * ######## Finalization support
2381 * ######################################################################
2385 * If the object has been forwarded it means it's still referenced from a root.
2386 * If it is pinned it's still alive as well.
2387 * A LOS object is only alive if we have pinned it.
2388 * Return TRUE if @obj is ready to be finalized.
2390 static inline gboolean
2391 sgen_is_object_alive (GCObject *object)
2393 if (ptr_in_nursery (object))
2394 return sgen_nursery_is_object_alive (object);
2396 return sgen_major_is_object_alive (object);
2400 * This function returns true if @object is either alive and belongs to the
2401 * current collection - major collections are full heap, so old gen objects
2402 * are never alive during a minor collection.
2405 sgen_is_object_alive_and_on_current_collection (GCObject *object)
2407 if (ptr_in_nursery (object))
2408 return sgen_nursery_is_object_alive (object);
2410 if (current_collection_generation == GENERATION_NURSERY)
2413 return sgen_major_is_object_alive (object);
2418 sgen_gc_is_object_ready_for_finalization (GCObject *object)
2420 return !sgen_is_object_alive (object);
2424 sgen_queue_finalization_entry (GCObject *obj)
2426 gboolean critical = sgen_client_object_has_critical_finalizer (obj);
2428 sgen_pointer_queue_add (critical ? &critical_fin_queue : &fin_ready_queue, obj);
2430 sgen_client_object_queued_for_finalization (obj);
2434 sgen_object_is_live (GCObject *obj)
2436 return sgen_is_object_alive_and_on_current_collection (obj);
2440 * `System.GC.WaitForPendingFinalizers` first checks `sgen_have_pending_finalizers()` to
2441 * determine whether it can exit quickly. The latter must therefore only return FALSE if
2442 * all finalizers have really finished running.
2444 * `sgen_gc_invoke_finalizers()` first dequeues a finalizable object, and then finalizes it.
2445 * This means that just checking whether the queues are empty leaves the possibility that an
2446 * object might have been dequeued but not yet finalized. That's why we need the additional
2447 * flag `pending_unqueued_finalizer`.
2450 static volatile gboolean pending_unqueued_finalizer = FALSE;
2453 sgen_gc_invoke_finalizers (void)
2457 g_assert (!pending_unqueued_finalizer);
2459 /* FIXME: batch to reduce lock contention */
2460 while (sgen_have_pending_finalizers ()) {
2466 * We need to set `pending_unqueued_finalizer` before dequeing the
2467 * finalizable object.
2469 if (!sgen_pointer_queue_is_empty (&fin_ready_queue)) {
2470 pending_unqueued_finalizer = TRUE;
2471 mono_memory_write_barrier ();
2472 obj = sgen_pointer_queue_pop (&fin_ready_queue);
2473 } else if (!sgen_pointer_queue_is_empty (&critical_fin_queue)) {
2474 pending_unqueued_finalizer = TRUE;
2475 mono_memory_write_barrier ();
2476 obj = sgen_pointer_queue_pop (&critical_fin_queue);
2482 SGEN_LOG (7, "Finalizing object %p (%s)", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
2490 /* the object is on the stack so it is pinned */
2491 /*g_print ("Calling finalizer for object: %p (%s)\n", obj, sgen_client_object_safe_name (obj));*/
2492 sgen_client_run_finalize (obj);
2495 if (pending_unqueued_finalizer) {
2496 mono_memory_write_barrier ();
2497 pending_unqueued_finalizer = FALSE;
2504 sgen_have_pending_finalizers (void)
2506 return pending_unqueued_finalizer || !sgen_pointer_queue_is_empty (&fin_ready_queue) || !sgen_pointer_queue_is_empty (&critical_fin_queue);
2510 * ######################################################################
2511 * ######## registered roots support
2512 * ######################################################################
2516 * We do not coalesce roots.
2519 sgen_register_root (char *start, size_t size, SgenDescriptor descr, int root_type, int source, const char *msg)
2521 RootRecord new_root;
2524 for (i = 0; i < ROOT_TYPE_NUM; ++i) {
2525 RootRecord *root = sgen_hash_table_lookup (&roots_hash [i], start);
2526 /* we allow changing the size and the descriptor (for thread statics etc) */
2528 size_t old_size = root->end_root - start;
2529 root->end_root = start + size;
2530 SGEN_ASSERT (0, !!root->root_desc == !!descr, "Can't change whether a root is precise or conservative.");
2531 SGEN_ASSERT (0, root->source == source, "Can't change a root's source identifier.");
2532 SGEN_ASSERT (0, !!root->msg == !!msg, "Can't change a root's message.");
2533 root->root_desc = descr;
2535 roots_size -= old_size;
2541 new_root.end_root = start + size;
2542 new_root.root_desc = descr;
2543 new_root.source = source;
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, (long long)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_set_allow_synchronous_major (gboolean flag)
2752 if (!major_collector.is_concurrent)
2755 allow_synchronous_major = flag;
2760 sgen_env_var_error (const char *env_var, const char *fallback, const char *description_format, ...)
2764 va_start (ap, description_format);
2766 fprintf (stderr, "Warning: In environment variable `%s': ", env_var);
2767 vfprintf (stderr, description_format, ap);
2769 fprintf (stderr, " - %s", fallback);
2770 fprintf (stderr, "\n");
2776 parse_double_in_interval (const char *env_var, const char *opt_name, const char *opt, double min, double max, double *result)
2779 double val = strtod (opt, &endptr);
2780 if (endptr == opt) {
2781 sgen_env_var_error (env_var, "Using default value.", "`%s` must be a number.", opt_name);
2784 else if (val < min || val > max) {
2785 sgen_env_var_error (env_var, "Using default value.", "`%s` must be between %.2f - %.2f.", opt_name, min, max);
2797 char *major_collector_opt = NULL;
2798 char *minor_collector_opt = NULL;
2799 size_t max_heap = 0;
2800 size_t soft_limit = 0;
2802 gboolean debug_print_allowance = FALSE;
2803 double allowance_ratio = 0, save_target = 0;
2804 gboolean cement_enabled = TRUE;
2807 result = InterlockedCompareExchange (&gc_initialized, -1, 0);
2810 /* already inited */
2813 /* being inited by another thread */
2817 /* we will init it */
2820 g_assert_not_reached ();
2822 } while (result != 0);
2824 SGEN_TV_GETTIME (sgen_init_timestamp);
2826 #ifdef SGEN_WITHOUT_MONO
2827 mono_thread_smr_init ();
2830 LOCK_INIT (gc_mutex);
2832 gc_debug_file = stderr;
2834 LOCK_INIT (sgen_interruption_mutex);
2836 if ((env = g_getenv (MONO_GC_PARAMS_NAME))) {
2837 opts = g_strsplit (env, ",", -1);
2838 for (ptr = opts; *ptr; ++ptr) {
2840 if (g_str_has_prefix (opt, "major=")) {
2841 opt = strchr (opt, '=') + 1;
2842 major_collector_opt = g_strdup (opt);
2843 } else if (g_str_has_prefix (opt, "minor=")) {
2844 opt = strchr (opt, '=') + 1;
2845 minor_collector_opt = g_strdup (opt);
2853 sgen_init_internal_allocator ();
2854 sgen_init_nursery_allocator ();
2855 sgen_init_fin_weak_hash ();
2856 sgen_init_hash_table ();
2857 sgen_init_descriptors ();
2858 sgen_init_gray_queues ();
2859 sgen_init_allocator ();
2860 sgen_init_gchandles ();
2862 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_SECTION, SGEN_SIZEOF_GC_MEM_SECTION);
2863 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_GRAY_QUEUE, sizeof (GrayQueueSection));
2865 sgen_client_init ();
2867 if (!minor_collector_opt) {
2868 sgen_simple_nursery_init (&sgen_minor_collector);
2870 if (!strcmp (minor_collector_opt, "simple")) {
2872 sgen_simple_nursery_init (&sgen_minor_collector);
2873 } else if (!strcmp (minor_collector_opt, "split")) {
2874 sgen_split_nursery_init (&sgen_minor_collector);
2876 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using `simple` instead.", "Unknown minor collector `%s'.", minor_collector_opt);
2877 goto use_simple_nursery;
2881 if (!major_collector_opt || !strcmp (major_collector_opt, "marksweep")) {
2882 use_marksweep_major:
2883 sgen_marksweep_init (&major_collector);
2884 } else if (!major_collector_opt || !strcmp (major_collector_opt, "marksweep-conc")) {
2885 sgen_marksweep_conc_init (&major_collector);
2887 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using `marksweep` instead.", "Unknown major collector `%s'.", major_collector_opt);
2888 goto use_marksweep_major;
2891 sgen_nursery_size = DEFAULT_NURSERY_SIZE;
2894 gboolean usage_printed = FALSE;
2896 for (ptr = opts; *ptr; ++ptr) {
2898 if (!strcmp (opt, ""))
2900 if (g_str_has_prefix (opt, "major="))
2902 if (g_str_has_prefix (opt, "minor="))
2904 if (g_str_has_prefix (opt, "max-heap-size=")) {
2905 size_t page_size = mono_pagesize ();
2906 size_t max_heap_candidate = 0;
2907 opt = strchr (opt, '=') + 1;
2908 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &max_heap_candidate)) {
2909 max_heap = (max_heap_candidate + page_size - 1) & ~(size_t)(page_size - 1);
2910 if (max_heap != max_heap_candidate)
2911 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Rounding up.", "`max-heap-size` size must be a multiple of %d.", page_size);
2913 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`max-heap-size` must be an integer.");
2917 if (g_str_has_prefix (opt, "soft-heap-limit=")) {
2918 opt = strchr (opt, '=') + 1;
2919 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &soft_limit)) {
2920 if (soft_limit <= 0) {
2921 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be positive.");
2925 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be an integer.");
2931 if (g_str_has_prefix (opt, "nursery-size=")) {
2933 opt = strchr (opt, '=') + 1;
2934 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &val)) {
2935 if ((val & (val - 1))) {
2936 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be a power of two.");
2940 if (val < SGEN_MAX_NURSERY_WASTE) {
2941 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.",
2942 "`nursery-size` must be at least %d bytes.", SGEN_MAX_NURSERY_WASTE);
2946 sgen_nursery_size = val;
2947 sgen_nursery_bits = 0;
2948 while (ONE_P << (++ sgen_nursery_bits) != sgen_nursery_size)
2951 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be an integer.");
2957 if (g_str_has_prefix (opt, "save-target-ratio=")) {
2959 opt = strchr (opt, '=') + 1;
2960 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "save-target-ratio", opt,
2961 SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO, &val)) {
2966 if (g_str_has_prefix (opt, "default-allowance-ratio=")) {
2968 opt = strchr (opt, '=') + 1;
2969 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "default-allowance-ratio", opt,
2970 SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, &val)) {
2971 allowance_ratio = val;
2975 if (g_str_has_prefix (opt, "allow-synchronous-major=")) {
2976 if (!major_collector.is_concurrent) {
2977 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring.", "`allow-synchronous-major` is only valid for the concurrent major collector.");
2981 opt = strchr (opt, '=') + 1;
2983 if (!strcmp (opt, "yes")) {
2984 allow_synchronous_major = TRUE;
2985 } else if (!strcmp (opt, "no")) {
2986 allow_synchronous_major = FALSE;
2988 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`allow-synchronous-major` must be either `yes' or `no'.");
2993 if (!strcmp (opt, "cementing")) {
2994 cement_enabled = TRUE;
2997 if (!strcmp (opt, "no-cementing")) {
2998 cement_enabled = FALSE;
3002 if (major_collector.handle_gc_param && major_collector.handle_gc_param (opt))
3005 if (sgen_minor_collector.handle_gc_param && sgen_minor_collector.handle_gc_param (opt))
3008 if (sgen_client_handle_gc_param (opt))
3011 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring.", "Unknown option `%s`.", opt);
3016 fprintf (stderr, "\n%s must be a comma-delimited list of one or more of the following:\n", MONO_GC_PARAMS_NAME);
3017 fprintf (stderr, " max-heap-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
3018 fprintf (stderr, " soft-heap-limit=n (where N is an integer, possibly with a k, m or a g suffix)\n");
3019 fprintf (stderr, " nursery-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
3020 fprintf (stderr, " major=COLLECTOR (where COLLECTOR is `marksweep', `marksweep-conc', `marksweep-par')\n");
3021 fprintf (stderr, " minor=COLLECTOR (where COLLECTOR is `simple' or `split')\n");
3022 fprintf (stderr, " wbarrier=WBARRIER (where WBARRIER is `remset' or `cardtable')\n");
3023 fprintf (stderr, " [no-]cementing\n");
3024 if (major_collector.is_concurrent)
3025 fprintf (stderr, " allow-synchronous-major=FLAG (where FLAG is `yes' or `no')\n");
3026 if (major_collector.print_gc_param_usage)
3027 major_collector.print_gc_param_usage ();
3028 if (sgen_minor_collector.print_gc_param_usage)
3029 sgen_minor_collector.print_gc_param_usage ();
3030 sgen_client_print_gc_params_usage ();
3031 fprintf (stderr, " Experimental options:\n");
3032 fprintf (stderr, " save-target-ratio=R (where R must be between %.2f - %.2f).\n", SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO);
3033 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);
3034 fprintf (stderr, "\n");
3036 usage_printed = TRUE;
3041 if (major_collector_opt)
3042 g_free (major_collector_opt);
3044 if (minor_collector_opt)
3045 g_free (minor_collector_opt);
3049 sgen_cement_init (cement_enabled);
3051 if ((env = g_getenv (MONO_GC_DEBUG_NAME))) {
3052 gboolean usage_printed = FALSE;
3054 opts = g_strsplit (env, ",", -1);
3055 for (ptr = opts; ptr && *ptr; ptr ++) {
3057 if (!strcmp (opt, ""))
3059 if (opt [0] >= '0' && opt [0] <= '9') {
3060 gc_debug_level = atoi (opt);
3065 char *rf = g_strdup_printf ("%s.%d", opt, mono_process_current_pid ());
3066 gc_debug_file = fopen (rf, "wb");
3068 gc_debug_file = stderr;
3071 } else if (!strcmp (opt, "print-allowance")) {
3072 debug_print_allowance = TRUE;
3073 } else if (!strcmp (opt, "print-pinning")) {
3074 sgen_pin_stats_enable ();
3075 } else if (!strcmp (opt, "verify-before-allocs")) {
3076 verify_before_allocs = 1;
3077 has_per_allocation_action = TRUE;
3078 } else if (g_str_has_prefix (opt, "verify-before-allocs=")) {
3079 char *arg = strchr (opt, '=') + 1;
3080 verify_before_allocs = atoi (arg);
3081 has_per_allocation_action = TRUE;
3082 } else if (!strcmp (opt, "collect-before-allocs")) {
3083 collect_before_allocs = 1;
3084 has_per_allocation_action = TRUE;
3085 } else if (g_str_has_prefix (opt, "collect-before-allocs=")) {
3086 char *arg = strchr (opt, '=') + 1;
3087 has_per_allocation_action = TRUE;
3088 collect_before_allocs = atoi (arg);
3089 } else if (!strcmp (opt, "verify-before-collections")) {
3090 whole_heap_check_before_collection = TRUE;
3091 } else if (!strcmp (opt, "check-at-minor-collections")) {
3092 consistency_check_at_minor_collection = TRUE;
3093 nursery_clear_policy = CLEAR_AT_GC;
3094 } else if (!strcmp (opt, "mod-union-consistency-check")) {
3095 if (!major_collector.is_concurrent) {
3096 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`mod-union-consistency-check` only works with concurrent major collector.");
3099 mod_union_consistency_check = TRUE;
3100 } else if (!strcmp (opt, "check-mark-bits")) {
3101 check_mark_bits_after_major_collection = TRUE;
3102 } else if (!strcmp (opt, "check-nursery-pinned")) {
3103 check_nursery_objects_pinned = TRUE;
3104 } else if (!strcmp (opt, "clear-at-gc")) {
3105 nursery_clear_policy = CLEAR_AT_GC;
3106 } else if (!strcmp (opt, "clear-nursery-at-gc")) {
3107 nursery_clear_policy = CLEAR_AT_GC;
3108 } else if (!strcmp (opt, "clear-at-tlab-creation")) {
3109 nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
3110 } else if (!strcmp (opt, "debug-clear-at-tlab-creation")) {
3111 nursery_clear_policy = CLEAR_AT_TLAB_CREATION_DEBUG;
3112 } else if (!strcmp (opt, "check-scan-starts")) {
3113 do_scan_starts_check = TRUE;
3114 } else if (!strcmp (opt, "verify-nursery-at-minor-gc")) {
3115 do_verify_nursery = TRUE;
3116 } else if (!strcmp (opt, "check-concurrent")) {
3117 if (!major_collector.is_concurrent) {
3118 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`check-concurrent` only works with concurrent major collectors.");
3121 nursery_clear_policy = CLEAR_AT_GC;
3122 do_concurrent_checks = TRUE;
3123 } else if (!strcmp (opt, "dump-nursery-at-minor-gc")) {
3124 do_dump_nursery_content = TRUE;
3125 } else if (!strcmp (opt, "disable-minor")) {
3126 disable_minor_collections = TRUE;
3127 } else if (!strcmp (opt, "disable-major")) {
3128 disable_major_collections = TRUE;
3129 } else if (g_str_has_prefix (opt, "heap-dump=")) {
3130 char *filename = strchr (opt, '=') + 1;
3131 nursery_clear_policy = CLEAR_AT_GC;
3132 sgen_debug_enable_heap_dump (filename);
3133 } else if (g_str_has_prefix (opt, "binary-protocol=")) {
3134 char *filename = strchr (opt, '=') + 1;
3135 char *colon = strrchr (filename, ':');
3138 if (!mono_gc_parse_environment_string_extract_number (colon + 1, &limit)) {
3139 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring limit.", "Binary protocol file size limit must be an integer.");
3144 binary_protocol_init (filename, (long long)limit);
3145 } else if (!strcmp (opt, "nursery-canaries")) {
3146 do_verify_nursery = TRUE;
3147 enable_nursery_canaries = TRUE;
3148 } else if (!sgen_client_handle_gc_debug (opt)) {
3149 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "Unknown option `%s`.", opt);
3154 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);
3155 fprintf (stderr, "Valid <option>s are:\n");
3156 fprintf (stderr, " collect-before-allocs[=<n>]\n");
3157 fprintf (stderr, " verify-before-allocs[=<n>]\n");
3158 fprintf (stderr, " check-at-minor-collections\n");
3159 fprintf (stderr, " check-mark-bits\n");
3160 fprintf (stderr, " check-nursery-pinned\n");
3161 fprintf (stderr, " verify-before-collections\n");
3162 fprintf (stderr, " verify-nursery-at-minor-gc\n");
3163 fprintf (stderr, " dump-nursery-at-minor-gc\n");
3164 fprintf (stderr, " disable-minor\n");
3165 fprintf (stderr, " disable-major\n");
3166 fprintf (stderr, " check-concurrent\n");
3167 fprintf (stderr, " clear-[nursery-]at-gc\n");
3168 fprintf (stderr, " clear-at-tlab-creation\n");
3169 fprintf (stderr, " debug-clear-at-tlab-creation\n");
3170 fprintf (stderr, " check-scan-starts\n");
3171 fprintf (stderr, " print-allowance\n");
3172 fprintf (stderr, " print-pinning\n");
3173 fprintf (stderr, " heap-dump=<filename>\n");
3174 fprintf (stderr, " binary-protocol=<filename>[:<file-size-limit>]\n");
3175 fprintf (stderr, " nursery-canaries\n");
3176 sgen_client_print_gc_debug_usage ();
3177 fprintf (stderr, "\n");
3179 usage_printed = TRUE;
3185 if (check_mark_bits_after_major_collection)
3186 nursery_clear_policy = CLEAR_AT_GC;
3188 if (major_collector.post_param_init)
3189 major_collector.post_param_init (&major_collector);
3191 if (major_collector.needs_thread_pool)
3192 sgen_workers_init (1);
3194 sgen_memgov_init (max_heap, soft_limit, debug_print_allowance, allowance_ratio, save_target);
3196 memset (&remset, 0, sizeof (remset));
3198 sgen_card_table_init (&remset);
3200 sgen_register_root (NULL, 0, sgen_make_user_root_descriptor (sgen_mark_normal_gc_handles), ROOT_TYPE_NORMAL, MONO_ROOT_SOURCE_GC_HANDLE, "normal gc handles");
3206 sgen_get_nursery_clear_policy (void)
3208 return nursery_clear_policy;
3218 sgen_gc_unlock (void)
3220 gboolean try_free = sgen_try_free_some_memory;
3221 sgen_try_free_some_memory = FALSE;
3222 mono_mutex_unlock (&gc_mutex);
3224 mono_thread_hazardous_try_free_some ();
3228 sgen_major_collector_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
3230 major_collector.iterate_live_block_ranges (callback);
3234 sgen_get_major_collector (void)
3236 return &major_collector;
3240 sgen_get_remset (void)
3246 count_cards (long long *major_total, long long *major_marked, long long *los_total, long long *los_marked)
3248 sgen_get_major_collector ()->count_cards (major_total, major_marked);
3249 sgen_los_count_cards (los_total, los_marked);
3252 static gboolean world_is_stopped = FALSE;
3254 /* LOCKING: assumes the GC lock is held */
3256 sgen_stop_world (int generation)
3258 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3260 SGEN_ASSERT (0, !world_is_stopped, "Why are we stopping a stopped world?");
3262 binary_protocol_world_stopping (generation, sgen_timestamp (), (gpointer)mono_native_thread_id_get ());
3264 sgen_client_stop_world (generation);
3266 world_is_stopped = TRUE;
3268 if (binary_protocol_is_heavy_enabled ())
3269 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3270 binary_protocol_world_stopped (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3273 /* LOCKING: assumes the GC lock is held */
3275 sgen_restart_world (int generation, GGTimingInfo *timing)
3277 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3279 SGEN_ASSERT (0, world_is_stopped, "Why are we restarting a running world?");
3281 if (binary_protocol_is_heavy_enabled ())
3282 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3283 binary_protocol_world_restarting (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3285 sgen_client_restart_world (generation, timing);
3287 world_is_stopped = FALSE;
3289 binary_protocol_world_restarted (generation, sgen_timestamp ());
3291 sgen_try_free_some_memory = TRUE;
3293 if (sgen_client_bridge_need_processing ())
3294 sgen_client_bridge_processing_finish (generation);
3296 sgen_memgov_collection_end (generation, timing, timing ? 2 : 0);
3300 sgen_is_world_stopped (void)
3302 return world_is_stopped;
3306 sgen_check_whole_heap_stw (void)
3308 sgen_stop_world (0);
3309 sgen_clear_nursery_fragments ();
3310 sgen_check_whole_heap (FALSE);
3311 sgen_restart_world (0, NULL);
3315 sgen_timestamp (void)
3317 SGEN_TV_DECLARE (timestamp);
3318 SGEN_TV_GETTIME (timestamp);
3319 return SGEN_TV_ELAPSED (sgen_init_timestamp, timestamp);
3322 #endif /* HAVE_SGEN_GC */