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;
242 static gboolean disable_minor_collections = FALSE;
243 static gboolean disable_major_collections = FALSE;
244 static gboolean do_verify_nursery = FALSE;
245 static gboolean do_dump_nursery_content = FALSE;
246 static gboolean enable_nursery_canaries = FALSE;
248 #ifdef HEAVY_STATISTICS
249 guint64 stat_objects_alloced_degraded = 0;
250 guint64 stat_bytes_alloced_degraded = 0;
252 guint64 stat_copy_object_called_nursery = 0;
253 guint64 stat_objects_copied_nursery = 0;
254 guint64 stat_copy_object_called_major = 0;
255 guint64 stat_objects_copied_major = 0;
257 guint64 stat_scan_object_called_nursery = 0;
258 guint64 stat_scan_object_called_major = 0;
260 guint64 stat_slots_allocated_in_vain;
262 guint64 stat_nursery_copy_object_failed_from_space = 0;
263 guint64 stat_nursery_copy_object_failed_forwarded = 0;
264 guint64 stat_nursery_copy_object_failed_pinned = 0;
265 guint64 stat_nursery_copy_object_failed_to_space = 0;
267 static guint64 stat_wbarrier_add_to_global_remset = 0;
268 static guint64 stat_wbarrier_arrayref_copy = 0;
269 static guint64 stat_wbarrier_generic_store = 0;
270 static guint64 stat_wbarrier_generic_store_atomic = 0;
271 static guint64 stat_wbarrier_set_root = 0;
274 static guint64 stat_pinned_objects = 0;
276 static guint64 time_minor_pre_collection_fragment_clear = 0;
277 static guint64 time_minor_pinning = 0;
278 static guint64 time_minor_scan_remsets = 0;
279 static guint64 time_minor_scan_pinned = 0;
280 static guint64 time_minor_scan_roots = 0;
281 static guint64 time_minor_finish_gray_stack = 0;
282 static guint64 time_minor_fragment_creation = 0;
284 static guint64 time_major_pre_collection_fragment_clear = 0;
285 static guint64 time_major_pinning = 0;
286 static guint64 time_major_scan_pinned = 0;
287 static guint64 time_major_scan_roots = 0;
288 static guint64 time_major_scan_mod_union = 0;
289 static guint64 time_major_finish_gray_stack = 0;
290 static guint64 time_major_free_bigobjs = 0;
291 static guint64 time_major_los_sweep = 0;
292 static guint64 time_major_sweep = 0;
293 static guint64 time_major_fragment_creation = 0;
295 static guint64 time_max = 0;
297 static SGEN_TV_DECLARE (time_major_conc_collection_start);
298 static SGEN_TV_DECLARE (time_major_conc_collection_end);
300 static SGEN_TV_DECLARE (last_minor_collection_start_tv);
301 static SGEN_TV_DECLARE (last_minor_collection_end_tv);
303 int gc_debug_level = 0;
308 mono_gc_flush_info (void)
310 fflush (gc_debug_file);
314 #define TV_DECLARE SGEN_TV_DECLARE
315 #define TV_GETTIME SGEN_TV_GETTIME
316 #define TV_ELAPSED SGEN_TV_ELAPSED
318 static SGEN_TV_DECLARE (sgen_init_timestamp);
320 NurseryClearPolicy nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
322 #define object_is_forwarded SGEN_OBJECT_IS_FORWARDED
323 #define object_is_pinned SGEN_OBJECT_IS_PINNED
324 #define pin_object SGEN_PIN_OBJECT
326 #define ptr_in_nursery sgen_ptr_in_nursery
328 #define LOAD_VTABLE SGEN_LOAD_VTABLE
331 nursery_canaries_enabled (void)
333 return enable_nursery_canaries;
336 #define safe_object_get_size sgen_safe_object_get_size
339 * ######################################################################
340 * ######## Global data.
341 * ######################################################################
343 MonoCoopMutex gc_mutex;
344 gboolean sgen_try_free_some_memory;
346 #define SCAN_START_SIZE SGEN_SCAN_START_SIZE
348 size_t degraded_mode = 0;
350 static mword bytes_pinned_from_failed_allocation = 0;
352 GCMemSection *nursery_section = NULL;
353 static volatile mword lowest_heap_address = ~(mword)0;
354 static volatile mword highest_heap_address = 0;
356 MonoCoopMutex sgen_interruption_mutex;
358 int current_collection_generation = -1;
359 static volatile gboolean concurrent_collection_in_progress = FALSE;
361 /* objects that are ready to be finalized */
362 static SgenPointerQueue fin_ready_queue = SGEN_POINTER_QUEUE_INIT (INTERNAL_MEM_FINALIZE_READY);
363 static SgenPointerQueue critical_fin_queue = SGEN_POINTER_QUEUE_INIT (INTERNAL_MEM_FINALIZE_READY);
365 /* registered roots: the key to the hash is the root start address */
367 * Different kinds of roots are kept separate to speed up pin_from_roots () for example.
369 SgenHashTable roots_hash [ROOT_TYPE_NUM] = {
370 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL),
371 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL),
372 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL)
374 static mword roots_size = 0; /* amount of memory in the root set */
376 /* The size of a TLAB */
377 /* The bigger the value, the less often we have to go to the slow path to allocate a new
378 * one, but the more space is wasted by threads not allocating much memory.
380 * FIXME: Make this self-tuning for each thread.
382 guint32 tlab_size = (1024 * 4);
384 #define MAX_SMALL_OBJ_SIZE SGEN_MAX_SMALL_OBJ_SIZE
386 #define ALLOC_ALIGN SGEN_ALLOC_ALIGN
388 #define ALIGN_UP SGEN_ALIGN_UP
390 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
391 MonoNativeThreadId main_gc_thread = NULL;
394 /*Object was pinned during the current collection*/
395 static mword objects_pinned;
398 * ######################################################################
399 * ######## Macros and function declarations.
400 * ######################################################################
403 typedef SgenGrayQueue GrayQueue;
405 /* forward declarations */
406 static void scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx);
408 static void pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx);
409 static void finish_gray_stack (int generation, ScanCopyContext ctx);
412 SgenMajorCollector major_collector;
413 SgenMinorCollector sgen_minor_collector;
414 /* FIXME: get rid of this */
415 static GrayQueue gray_queue;
417 static SgenRememberedSet remset;
419 /* The gray queue to use from the main collection thread. */
420 #define WORKERS_DISTRIBUTE_GRAY_QUEUE (&gray_queue)
423 * The gray queue a worker job must use. If we're not parallel or
424 * concurrent, we use the main gray queue.
426 static SgenGrayQueue*
427 sgen_workers_get_job_gray_queue (WorkerData *worker_data)
429 return worker_data ? &worker_data->private_gray_queue : WORKERS_DISTRIBUTE_GRAY_QUEUE;
433 gray_queue_redirect (SgenGrayQueue *queue)
435 gboolean wake = FALSE;
438 GrayQueueSection *section = sgen_gray_object_dequeue_section (queue);
441 sgen_section_gray_queue_enqueue ((SgenSectionGrayQueue *)queue->alloc_prepare_data, section);
446 g_assert (concurrent_collection_in_progress);
447 sgen_workers_ensure_awake ();
452 gray_queue_enable_redirect (SgenGrayQueue *queue)
454 if (!concurrent_collection_in_progress)
457 sgen_gray_queue_set_alloc_prepare (queue, gray_queue_redirect, sgen_workers_get_distribute_section_gray_queue ());
458 gray_queue_redirect (queue);
462 sgen_scan_area_with_callback (char *start, char *end, IterateObjectCallbackFunc callback, void *data, gboolean allow_flags, gboolean fail_on_canaries)
464 while (start < end) {
468 if (!*(void**)start) {
469 start += sizeof (void*); /* should be ALLOC_ALIGN, really */
474 if (!(obj = (char *)SGEN_OBJECT_IS_FORWARDED (start)))
480 if (!sgen_client_object_is_array_fill ((GCObject*)obj)) {
481 CHECK_CANARY_FOR_OBJECT ((GCObject*)obj, fail_on_canaries);
482 size = ALIGN_UP (safe_object_get_size ((GCObject*)obj));
483 callback ((GCObject*)obj, size, data);
484 CANARIFY_SIZE (size);
486 size = ALIGN_UP (safe_object_get_size ((GCObject*)obj));
494 * sgen_add_to_global_remset:
496 * The global remset contains locations which point into newspace after
497 * a minor collection. This can happen if the objects they point to are pinned.
499 * LOCKING: If called from a parallel collector, the global remset
500 * lock must be held. For serial collectors that is not necessary.
503 sgen_add_to_global_remset (gpointer ptr, GCObject *obj)
505 SGEN_ASSERT (5, sgen_ptr_in_nursery (obj), "Target pointer of global remset must be in the nursery");
507 HEAVY_STAT (++stat_wbarrier_add_to_global_remset);
509 if (!major_collector.is_concurrent) {
510 SGEN_ASSERT (5, current_collection_generation != -1, "Global remsets can only be added during collections");
512 if (current_collection_generation == -1)
513 SGEN_ASSERT (5, sgen_concurrent_collection_in_progress (), "Global remsets outside of collection pauses can only be added by the concurrent collector");
516 if (!object_is_pinned (obj))
517 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");
518 else if (sgen_cement_lookup_or_register (obj))
521 remset.record_pointer (ptr);
523 sgen_pin_stats_register_global_remset (obj);
525 SGEN_LOG (8, "Adding global remset for %p", ptr);
526 binary_protocol_global_remset (ptr, obj, (gpointer)SGEN_LOAD_VTABLE (obj));
530 * sgen_drain_gray_stack:
532 * Scan objects in the gray stack until the stack is empty. This should be called
533 * frequently after each object is copied, to achieve better locality and cache
538 sgen_drain_gray_stack (ScanCopyContext ctx)
540 ScanObjectFunc scan_func = ctx.ops->scan_object;
541 GrayQueue *queue = ctx.queue;
543 if (ctx.ops->drain_gray_stack)
544 return ctx.ops->drain_gray_stack (queue);
549 GRAY_OBJECT_DEQUEUE (queue, &obj, &desc);
552 SGEN_LOG (9, "Precise gray object scan %p (%s)", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
553 scan_func (obj, desc, queue);
559 * Addresses in the pin queue are already sorted. This function finds
560 * the object header for each address and pins the object. The
561 * addresses must be inside the nursery section. The (start of the)
562 * address array is overwritten with the addresses of the actually
563 * pinned objects. Return the number of pinned objects.
566 pin_objects_from_nursery_pin_queue (gboolean do_scan_objects, ScanCopyContext ctx)
568 GCMemSection *section = nursery_section;
569 void **start = sgen_pinning_get_entry (section->pin_queue_first_entry);
570 void **end = sgen_pinning_get_entry (section->pin_queue_last_entry);
571 void *start_nursery = section->data;
572 void *end_nursery = section->next_data;
577 void *pinning_front = start_nursery;
579 void **definitely_pinned = start;
580 ScanObjectFunc scan_func = ctx.ops->scan_object;
581 SgenGrayQueue *queue = ctx.queue;
583 sgen_nursery_allocator_prepare_for_pinning ();
585 while (start < end) {
586 GCObject *obj_to_pin = NULL;
587 size_t obj_to_pin_size = 0;
592 SGEN_ASSERT (0, addr >= start_nursery && addr < end_nursery, "Potential pinning address out of range");
593 SGEN_ASSERT (0, addr >= last, "Pin queue not sorted");
600 SGEN_LOG (5, "Considering pinning addr %p", addr);
601 /* We've already processed everything up to pinning_front. */
602 if (addr < pinning_front) {
608 * Find the closest scan start <= addr. We might search backward in the
609 * scan_starts array because entries might be NULL. In the worst case we
610 * start at start_nursery.
612 idx = ((char*)addr - (char*)section->data) / SCAN_START_SIZE;
613 SGEN_ASSERT (0, idx < section->num_scan_start, "Scan start index out of range");
614 search_start = (void*)section->scan_starts [idx];
615 if (!search_start || search_start > addr) {
618 search_start = section->scan_starts [idx];
619 if (search_start && search_start <= addr)
622 if (!search_start || search_start > addr)
623 search_start = start_nursery;
627 * If the pinning front is closer than the scan start we found, start
628 * searching at the front.
630 if (search_start < pinning_front)
631 search_start = pinning_front;
634 * Now addr should be in an object a short distance from search_start.
636 * search_start must point to zeroed mem or point to an object.
639 size_t obj_size, canarified_obj_size;
642 if (!*(void**)search_start) {
643 search_start = (void*)ALIGN_UP ((mword)search_start + sizeof (gpointer));
644 /* The loop condition makes sure we don't overrun addr. */
648 canarified_obj_size = obj_size = ALIGN_UP (safe_object_get_size ((GCObject*)search_start));
651 * Filler arrays are marked by an invalid sync word. We don't
652 * consider them for pinning. They are not delimited by canaries,
655 if (!sgen_client_object_is_array_fill ((GCObject*)search_start)) {
656 CHECK_CANARY_FOR_OBJECT (search_start, TRUE);
657 CANARIFY_SIZE (canarified_obj_size);
659 if (addr >= search_start && (char*)addr < (char*)search_start + obj_size) {
660 /* This is the object we're looking for. */
661 obj_to_pin = (GCObject*)search_start;
662 obj_to_pin_size = canarified_obj_size;
667 /* Skip to the next object */
668 search_start = (void*)((char*)search_start + canarified_obj_size);
669 } while (search_start <= addr);
671 /* We've searched past the address we were looking for. */
673 pinning_front = search_start;
674 goto next_pin_queue_entry;
678 * We've found an object to pin. It might still be a dummy array, but we
679 * can advance the pinning front in any case.
681 pinning_front = (char*)obj_to_pin + obj_to_pin_size;
684 * If this is a dummy array marking the beginning of a nursery
685 * fragment, we don't pin it.
687 if (sgen_client_object_is_array_fill (obj_to_pin))
688 goto next_pin_queue_entry;
691 * Finally - pin the object!
693 desc = sgen_obj_get_descriptor_safe (obj_to_pin);
694 if (do_scan_objects) {
695 scan_func (obj_to_pin, desc, queue);
697 SGEN_LOG (4, "Pinned object %p, vtable %p (%s), count %d\n",
698 obj_to_pin, *(void**)obj_to_pin, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj_to_pin)), count);
699 binary_protocol_pin (obj_to_pin,
700 (gpointer)LOAD_VTABLE (obj_to_pin),
701 safe_object_get_size (obj_to_pin));
703 pin_object (obj_to_pin);
704 GRAY_OBJECT_ENQUEUE (queue, obj_to_pin, desc);
705 sgen_pin_stats_register_object (obj_to_pin, obj_to_pin_size);
706 definitely_pinned [count] = obj_to_pin;
710 next_pin_queue_entry:
714 sgen_client_nursery_objects_pinned (definitely_pinned, count);
715 stat_pinned_objects += count;
720 pin_objects_in_nursery (gboolean do_scan_objects, ScanCopyContext ctx)
724 if (nursery_section->pin_queue_first_entry == nursery_section->pin_queue_last_entry)
727 reduced_to = pin_objects_from_nursery_pin_queue (do_scan_objects, ctx);
728 nursery_section->pin_queue_last_entry = nursery_section->pin_queue_first_entry + reduced_to;
732 * This function is only ever called (via `collector_pin_object()` in `sgen-copy-object.h`)
733 * when we can't promote an object because we're out of memory.
736 sgen_pin_object (GCObject *object, GrayQueue *queue)
739 * All pinned objects are assumed to have been staged, so we need to stage as well.
740 * Also, the count of staged objects shows that "late pinning" happened.
742 sgen_pin_stage_ptr (object);
744 SGEN_PIN_OBJECT (object);
745 binary_protocol_pin (object, (gpointer)LOAD_VTABLE (object), safe_object_get_size (object));
748 sgen_pin_stats_register_object (object, safe_object_get_size (object));
750 GRAY_OBJECT_ENQUEUE (queue, object, sgen_obj_get_descriptor_safe (object));
753 /* Sort the addresses in array in increasing order.
754 * Done using a by-the book heap sort. Which has decent and stable performance, is pretty cache efficient.
757 sgen_sort_addresses (void **array, size_t size)
762 for (i = 1; i < size; ++i) {
765 size_t parent = (child - 1) / 2;
767 if (array [parent] >= array [child])
770 tmp = array [parent];
771 array [parent] = array [child];
778 for (i = size - 1; i > 0; --i) {
781 array [i] = array [0];
787 while (root * 2 + 1 <= end) {
788 size_t child = root * 2 + 1;
790 if (child < end && array [child] < array [child + 1])
792 if (array [root] >= array [child])
796 array [root] = array [child];
805 * Scan the memory between start and end and queue values which could be pointers
806 * to the area between start_nursery and end_nursery for later consideration.
807 * Typically used for thread stacks.
810 sgen_conservatively_pin_objects_from (void **start, void **end, void *start_nursery, void *end_nursery, int pin_type)
814 SGEN_ASSERT (0, ((mword)start & (SIZEOF_VOID_P - 1)) == 0, "Why are we scanning for references in unaligned memory ?");
816 #if defined(VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE) && !defined(_WIN64)
817 VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE (start, (char*)end - (char*)start);
820 while (start < end) {
822 * *start can point to the middle of an object
823 * note: should we handle pointing at the end of an object?
824 * pinning in C# code disallows pointing at the end of an object
825 * but there is some small chance that an optimizing C compiler
826 * may keep the only reference to an object by pointing
827 * at the end of it. We ignore this small chance for now.
828 * Pointers to the end of an object are indistinguishable
829 * from pointers to the start of the next object in memory
830 * so if we allow that we'd need to pin two objects...
831 * We queue the pointer in an array, the
832 * array will then be sorted and uniqued. This way
833 * we can coalesce several pinning pointers and it should
834 * be faster since we'd do a memory scan with increasing
835 * addresses. Note: we can align the address to the allocation
836 * alignment, so the unique process is more effective.
838 mword addr = (mword)*start;
839 addr &= ~(ALLOC_ALIGN - 1);
840 if (addr >= (mword)start_nursery && addr < (mword)end_nursery) {
841 SGEN_LOG (6, "Pinning address %p from %p", (void*)addr, start);
842 sgen_pin_stage_ptr ((void*)addr);
843 binary_protocol_pin_stage (start, (void*)addr);
844 sgen_pin_stats_register_address ((char*)addr, pin_type);
850 SGEN_LOG (7, "found %d potential pinned heap pointers", count);
854 * The first thing we do in a collection is to identify pinned objects.
855 * This function considers all the areas of memory that need to be
856 * conservatively scanned.
859 pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx)
863 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);
864 /* objects pinned from the API are inside these roots */
865 SGEN_HASH_TABLE_FOREACH (&roots_hash [ROOT_TYPE_PINNED], void **, start_root, RootRecord *, root) {
866 SGEN_LOG (6, "Pinned roots %p-%p", start_root, root->end_root);
867 sgen_conservatively_pin_objects_from (start_root, (void**)root->end_root, start_nursery, end_nursery, PIN_TYPE_OTHER);
868 } SGEN_HASH_TABLE_FOREACH_END;
869 /* now deal with the thread stacks
870 * in the future we should be able to conservatively scan only:
871 * *) the cpu registers
872 * *) the unmanaged stack frames
873 * *) the _last_ managed stack frame
874 * *) pointers slots in managed frames
876 sgen_client_scan_thread_data (start_nursery, end_nursery, FALSE, ctx);
880 single_arg_user_copy_or_mark (GCObject **obj, void *gc_data)
882 ScanCopyContext *ctx = (ScanCopyContext *)gc_data;
883 ctx->ops->copy_or_mark_object (obj, ctx->queue);
887 * The memory area from start_root to end_root contains pointers to objects.
888 * Their position is precisely described by @desc (this means that the pointer
889 * can be either NULL or the pointer to the start of an object).
890 * This functions copies them to to_space updates them.
892 * This function is not thread-safe!
895 precisely_scan_objects_from (void** start_root, void** end_root, char* n_start, char *n_end, SgenDescriptor desc, ScanCopyContext ctx)
897 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
898 SgenGrayQueue *queue = ctx.queue;
900 switch (desc & ROOT_DESC_TYPE_MASK) {
901 case ROOT_DESC_BITMAP:
902 desc >>= ROOT_DESC_TYPE_SHIFT;
904 if ((desc & 1) && *start_root) {
905 copy_func ((GCObject**)start_root, queue);
906 SGEN_LOG (9, "Overwrote root at %p with %p", start_root, *start_root);
912 case ROOT_DESC_COMPLEX: {
913 gsize *bitmap_data = (gsize *)sgen_get_complex_descriptor_bitmap (desc);
914 gsize bwords = (*bitmap_data) - 1;
915 void **start_run = start_root;
917 while (bwords-- > 0) {
918 gsize bmap = *bitmap_data++;
919 void **objptr = start_run;
921 if ((bmap & 1) && *objptr) {
922 copy_func ((GCObject**)objptr, queue);
923 SGEN_LOG (9, "Overwrote root at %p with %p", objptr, *objptr);
928 start_run += GC_BITS_PER_WORD;
932 case ROOT_DESC_USER: {
933 SgenUserRootMarkFunc marker = sgen_get_user_descriptor_func (desc);
934 marker (start_root, single_arg_user_copy_or_mark, &ctx);
937 case ROOT_DESC_RUN_LEN:
938 g_assert_not_reached ();
940 g_assert_not_reached ();
945 reset_heap_boundaries (void)
947 lowest_heap_address = ~(mword)0;
948 highest_heap_address = 0;
952 sgen_update_heap_boundaries (mword low, mword high)
957 old = lowest_heap_address;
960 } while (SGEN_CAS_PTR ((gpointer*)&lowest_heap_address, (gpointer)low, (gpointer)old) != (gpointer)old);
963 old = highest_heap_address;
966 } while (SGEN_CAS_PTR ((gpointer*)&highest_heap_address, (gpointer)high, (gpointer)old) != (gpointer)old);
970 * Allocate and setup the data structures needed to be able to allocate objects
971 * in the nursery. The nursery is stored in nursery_section.
976 GCMemSection *section;
983 SGEN_LOG (2, "Allocating nursery size: %zu", (size_t)sgen_nursery_size);
984 /* later we will alloc a larger area for the nursery but only activate
985 * what we need. The rest will be used as expansion if we have too many pinned
986 * objects in the existing nursery.
988 /* FIXME: handle OOM */
989 section = (GCMemSection *)sgen_alloc_internal (INTERNAL_MEM_SECTION);
991 alloc_size = sgen_nursery_size;
993 /* If there isn't enough space even for the nursery we should simply abort. */
994 g_assert (sgen_memgov_try_alloc_space (alloc_size, SPACE_NURSERY));
996 data = (char *)major_collector.alloc_heap (alloc_size, alloc_size, DEFAULT_NURSERY_BITS);
997 sgen_update_heap_boundaries ((mword)data, (mword)(data + sgen_nursery_size));
998 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 ());
999 section->data = section->next_data = data;
1000 section->size = alloc_size;
1001 section->end_data = data + sgen_nursery_size;
1002 scan_starts = (alloc_size + SCAN_START_SIZE - 1) / SCAN_START_SIZE;
1003 section->scan_starts = (char **)sgen_alloc_internal_dynamic (sizeof (char*) * scan_starts, INTERNAL_MEM_SCAN_STARTS, TRUE);
1004 section->num_scan_start = scan_starts;
1006 nursery_section = section;
1008 sgen_nursery_allocator_set_nursery_bounds (data, data + sgen_nursery_size);
1012 mono_gc_get_logfile (void)
1014 return gc_debug_file;
1018 scan_finalizer_entries (SgenPointerQueue *fin_queue, ScanCopyContext ctx)
1020 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
1021 SgenGrayQueue *queue = ctx.queue;
1024 for (i = 0; i < fin_queue->next_slot; ++i) {
1025 GCObject *obj = (GCObject *)fin_queue->data [i];
1028 SGEN_LOG (5, "Scan of fin ready object: %p (%s)\n", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
1029 copy_func ((GCObject**)&fin_queue->data [i], queue);
1034 generation_name (int generation)
1036 switch (generation) {
1037 case GENERATION_NURSERY: return "nursery";
1038 case GENERATION_OLD: return "old";
1039 default: g_assert_not_reached ();
1044 sgen_generation_name (int generation)
1046 return generation_name (generation);
1050 finish_gray_stack (int generation, ScanCopyContext ctx)
1054 int done_with_ephemerons, ephemeron_rounds = 0;
1055 char *start_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_start () : NULL;
1056 char *end_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_end () : (char*)-1;
1057 SgenGrayQueue *queue = ctx.queue;
1060 * We copied all the reachable objects. Now it's the time to copy
1061 * the objects that were not referenced by the roots, but by the copied objects.
1062 * we built a stack of objects pointed to by gray_start: they are
1063 * additional roots and we may add more items as we go.
1064 * We loop until gray_start == gray_objects which means no more objects have
1065 * been added. Note this is iterative: no recursion is involved.
1066 * We need to walk the LO list as well in search of marked big objects
1067 * (use a flag since this is needed only on major collections). We need to loop
1068 * here as well, so keep a counter of marked LO (increasing it in copy_object).
1069 * To achieve better cache locality and cache usage, we drain the gray stack
1070 * frequently, after each object is copied, and just finish the work here.
1072 sgen_drain_gray_stack (ctx);
1074 SGEN_LOG (2, "%s generation done", generation_name (generation));
1077 Reset bridge data, we might have lingering data from a previous collection if this is a major
1078 collection trigged by minor overflow.
1080 We must reset the gathered bridges since their original block might be evacuated due to major
1081 fragmentation in the meanwhile and the bridge code should not have to deal with that.
1083 if (sgen_client_bridge_need_processing ())
1084 sgen_client_bridge_reset_data ();
1087 * Walk the ephemeron tables marking all values with reachable keys. This must be completely done
1088 * before processing finalizable objects and non-tracking weak links to avoid finalizing/clearing
1089 * objects that are in fact reachable.
1091 done_with_ephemerons = 0;
1093 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1094 sgen_drain_gray_stack (ctx);
1096 } while (!done_with_ephemerons);
1098 sgen_client_mark_togglerefs (start_addr, end_addr, ctx);
1100 if (sgen_client_bridge_need_processing ()) {
1101 /*Make sure the gray stack is empty before we process bridge objects so we get liveness right*/
1102 sgen_drain_gray_stack (ctx);
1103 sgen_collect_bridge_objects (generation, ctx);
1104 if (generation == GENERATION_OLD)
1105 sgen_collect_bridge_objects (GENERATION_NURSERY, ctx);
1108 Do the first bridge step here, as the collector liveness state will become useless after that.
1110 An important optimization is to only proccess the possibly dead part of the object graph and skip
1111 over all live objects as we transitively know everything they point must be alive too.
1113 The above invariant is completely wrong if we let the gray queue be drained and mark/copy everything.
1115 This has the unfortunate side effect of making overflow collections perform the first step twice, but
1116 given we now have heuristics that perform major GC in anticipation of minor overflows this should not
1119 sgen_client_bridge_processing_stw_step ();
1123 Make sure we drain the gray stack before processing disappearing links and finalizers.
1124 If we don't make sure it is empty we might wrongly see a live object as dead.
1126 sgen_drain_gray_stack (ctx);
1129 We must clear weak links that don't track resurrection before processing object ready for
1130 finalization so they can be cleared before that.
1132 sgen_null_link_in_range (generation, ctx, FALSE);
1133 if (generation == GENERATION_OLD)
1134 sgen_null_link_in_range (GENERATION_NURSERY, ctx, FALSE);
1137 /* walk the finalization queue and move also the objects that need to be
1138 * finalized: use the finalized objects as new roots so the objects they depend
1139 * on are also not reclaimed. As with the roots above, only objects in the nursery
1140 * are marked/copied.
1142 sgen_finalize_in_range (generation, ctx);
1143 if (generation == GENERATION_OLD)
1144 sgen_finalize_in_range (GENERATION_NURSERY, ctx);
1145 /* drain the new stack that might have been created */
1146 SGEN_LOG (6, "Precise scan of gray area post fin");
1147 sgen_drain_gray_stack (ctx);
1150 * This must be done again after processing finalizable objects since CWL slots are cleared only after the key is finalized.
1152 done_with_ephemerons = 0;
1154 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1155 sgen_drain_gray_stack (ctx);
1157 } while (!done_with_ephemerons);
1159 sgen_client_clear_unreachable_ephemerons (ctx);
1162 * We clear togglerefs only after all possible chances of revival are done.
1163 * This is semantically more inline with what users expect and it allows for
1164 * user finalizers to correctly interact with TR objects.
1166 sgen_client_clear_togglerefs (start_addr, end_addr, ctx);
1169 SGEN_LOG (2, "Finalize queue handling scan for %s generation: %ld usecs %d ephemeron rounds", generation_name (generation), TV_ELAPSED (atv, btv), ephemeron_rounds);
1172 * handle disappearing links
1173 * Note we do this after checking the finalization queue because if an object
1174 * survives (at least long enough to be finalized) we don't clear the link.
1175 * This also deals with a possible issue with the monitor reclamation: with the Boehm
1176 * GC a finalized object my lose the monitor because it is cleared before the finalizer is
1179 g_assert (sgen_gray_object_queue_is_empty (queue));
1181 sgen_null_link_in_range (generation, ctx, TRUE);
1182 if (generation == GENERATION_OLD)
1183 sgen_null_link_in_range (GENERATION_NURSERY, ctx, TRUE);
1184 if (sgen_gray_object_queue_is_empty (queue))
1186 sgen_drain_gray_stack (ctx);
1189 g_assert (sgen_gray_object_queue_is_empty (queue));
1191 sgen_gray_object_queue_trim_free_list (queue);
1195 sgen_check_section_scan_starts (GCMemSection *section)
1198 for (i = 0; i < section->num_scan_start; ++i) {
1199 if (section->scan_starts [i]) {
1200 mword size = safe_object_get_size ((GCObject*) section->scan_starts [i]);
1201 SGEN_ASSERT (0, size >= SGEN_CLIENT_MINIMUM_OBJECT_SIZE && size <= MAX_SMALL_OBJ_SIZE, "Weird object size at scan starts.");
1207 check_scan_starts (void)
1209 if (!do_scan_starts_check)
1211 sgen_check_section_scan_starts (nursery_section);
1212 major_collector.check_scan_starts ();
1216 scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx)
1220 SGEN_HASH_TABLE_FOREACH (&roots_hash [root_type], void **, start_root, RootRecord *, root) {
1221 SGEN_LOG (6, "Precise root scan %p-%p (desc: %p)", start_root, root->end_root, (void*)root->root_desc);
1222 precisely_scan_objects_from (start_root, (void**)root->end_root, addr_start, addr_end, root->root_desc, ctx);
1223 } SGEN_HASH_TABLE_FOREACH_END;
1229 static gboolean inited = FALSE;
1234 mono_counters_register ("Collection max time", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME | MONO_COUNTER_MONOTONIC, &time_max);
1236 mono_counters_register ("Minor fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pre_collection_fragment_clear);
1237 mono_counters_register ("Minor pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pinning);
1238 mono_counters_register ("Minor scan remembered set", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_remsets);
1239 mono_counters_register ("Minor scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_pinned);
1240 mono_counters_register ("Minor scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_roots);
1241 mono_counters_register ("Minor fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_fragment_creation);
1243 mono_counters_register ("Major fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pre_collection_fragment_clear);
1244 mono_counters_register ("Major pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pinning);
1245 mono_counters_register ("Major scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_pinned);
1246 mono_counters_register ("Major scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_roots);
1247 mono_counters_register ("Major scan mod union", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_mod_union);
1248 mono_counters_register ("Major finish gray stack", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_finish_gray_stack);
1249 mono_counters_register ("Major free big objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_free_bigobjs);
1250 mono_counters_register ("Major LOS sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_los_sweep);
1251 mono_counters_register ("Major sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_sweep);
1252 mono_counters_register ("Major fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_fragment_creation);
1254 mono_counters_register ("Number of pinned objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_pinned_objects);
1256 #ifdef HEAVY_STATISTICS
1257 mono_counters_register ("WBarrier remember pointer", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_add_to_global_remset);
1258 mono_counters_register ("WBarrier arrayref copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_arrayref_copy);
1259 mono_counters_register ("WBarrier generic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store);
1260 mono_counters_register ("WBarrier generic atomic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store_atomic);
1261 mono_counters_register ("WBarrier set root", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_set_root);
1263 mono_counters_register ("# objects allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_alloced_degraded);
1264 mono_counters_register ("bytes allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced_degraded);
1266 mono_counters_register ("# copy_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_nursery);
1267 mono_counters_register ("# objects copied (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_nursery);
1268 mono_counters_register ("# copy_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_major);
1269 mono_counters_register ("# objects copied (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_major);
1271 mono_counters_register ("# scan_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_nursery);
1272 mono_counters_register ("# scan_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_major);
1274 mono_counters_register ("Slots allocated in vain", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_slots_allocated_in_vain);
1276 mono_counters_register ("# nursery copy_object() failed from space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_from_space);
1277 mono_counters_register ("# nursery copy_object() failed forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_forwarded);
1278 mono_counters_register ("# nursery copy_object() failed pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_pinned);
1279 mono_counters_register ("# nursery copy_object() failed to space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_to_space);
1281 sgen_nursery_allocator_init_heavy_stats ();
1289 reset_pinned_from_failed_allocation (void)
1291 bytes_pinned_from_failed_allocation = 0;
1295 sgen_set_pinned_from_failed_allocation (mword objsize)
1297 bytes_pinned_from_failed_allocation += objsize;
1301 sgen_collection_is_concurrent (void)
1303 switch (current_collection_generation) {
1304 case GENERATION_NURSERY:
1306 case GENERATION_OLD:
1307 return concurrent_collection_in_progress;
1309 g_error ("Invalid current generation %d", current_collection_generation);
1315 sgen_concurrent_collection_in_progress (void)
1317 return concurrent_collection_in_progress;
1321 SgenThreadPoolJob job;
1322 SgenObjectOperations *ops;
1326 job_remembered_set_scan (void *worker_data_untyped, SgenThreadPoolJob *job)
1328 WorkerData *worker_data = (WorkerData *)worker_data_untyped;
1329 ScanJob *job_data = (ScanJob*)job;
1330 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1331 remset.scan_remsets (ctx);
1335 SgenThreadPoolJob job;
1336 SgenObjectOperations *ops;
1340 } ScanFromRegisteredRootsJob;
1343 job_scan_from_registered_roots (void *worker_data_untyped, SgenThreadPoolJob *job)
1345 WorkerData *worker_data = (WorkerData *)worker_data_untyped;
1346 ScanFromRegisteredRootsJob *job_data = (ScanFromRegisteredRootsJob*)job;
1347 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1349 scan_from_registered_roots (job_data->heap_start, job_data->heap_end, job_data->root_type, ctx);
1353 SgenThreadPoolJob job;
1354 SgenObjectOperations *ops;
1357 } ScanThreadDataJob;
1360 job_scan_thread_data (void *worker_data_untyped, SgenThreadPoolJob *job)
1362 WorkerData *worker_data = (WorkerData *)worker_data_untyped;
1363 ScanThreadDataJob *job_data = (ScanThreadDataJob*)job;
1364 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1366 sgen_client_scan_thread_data (job_data->heap_start, job_data->heap_end, TRUE, ctx);
1370 SgenThreadPoolJob job;
1371 SgenObjectOperations *ops;
1372 SgenPointerQueue *queue;
1373 } ScanFinalizerEntriesJob;
1376 job_scan_finalizer_entries (void *worker_data_untyped, SgenThreadPoolJob *job)
1378 WorkerData *worker_data = (WorkerData *)worker_data_untyped;
1379 ScanFinalizerEntriesJob *job_data = (ScanFinalizerEntriesJob*)job;
1380 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1382 scan_finalizer_entries (job_data->queue, ctx);
1386 job_scan_major_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1388 WorkerData *worker_data = (WorkerData *)worker_data_untyped;
1389 ScanJob *job_data = (ScanJob*)job;
1390 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1392 g_assert (concurrent_collection_in_progress);
1393 major_collector.scan_card_table (TRUE, ctx);
1397 job_scan_los_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1399 WorkerData *worker_data = (WorkerData *)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 sgen_los_scan_card_table (TRUE, ctx);
1408 init_gray_queue (gboolean use_workers)
1411 sgen_workers_init_distribute_gray_queue ();
1412 sgen_gray_object_queue_init (&gray_queue, NULL);
1416 enqueue_scan_from_roots_jobs (char *heap_start, char *heap_end, SgenObjectOperations *ops, gboolean enqueue)
1418 ScanFromRegisteredRootsJob *scrrj;
1419 ScanThreadDataJob *stdj;
1420 ScanFinalizerEntriesJob *sfej;
1422 /* registered roots, this includes static fields */
1424 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots normal", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1426 scrrj->heap_start = heap_start;
1427 scrrj->heap_end = heap_end;
1428 scrrj->root_type = ROOT_TYPE_NORMAL;
1429 sgen_workers_enqueue_job (&scrrj->job, enqueue);
1431 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots wbarrier", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1433 scrrj->heap_start = heap_start;
1434 scrrj->heap_end = heap_end;
1435 scrrj->root_type = ROOT_TYPE_WBARRIER;
1436 sgen_workers_enqueue_job (&scrrj->job, enqueue);
1440 stdj = (ScanThreadDataJob*)sgen_thread_pool_job_alloc ("scan thread data", job_scan_thread_data, sizeof (ScanThreadDataJob));
1441 stdj->heap_start = heap_start;
1442 stdj->heap_end = heap_end;
1443 sgen_workers_enqueue_job (&stdj->job, enqueue);
1445 /* Scan the list of objects ready for finalization. */
1447 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1448 sfej->queue = &fin_ready_queue;
1450 sgen_workers_enqueue_job (&sfej->job, enqueue);
1452 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan critical finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1453 sfej->queue = &critical_fin_queue;
1455 sgen_workers_enqueue_job (&sfej->job, enqueue);
1459 * Perform a nursery collection.
1461 * Return whether any objects were late-pinned due to being out of memory.
1464 collect_nursery (SgenGrayQueue *unpin_queue, gboolean finish_up_concurrent_mark)
1466 gboolean needs_major;
1467 size_t max_garbage_amount;
1469 mword fragment_total;
1471 SgenObjectOperations *object_ops = &sgen_minor_collector.serial_ops;
1472 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, &gray_queue);
1476 if (disable_minor_collections)
1479 TV_GETTIME (last_minor_collection_start_tv);
1480 atv = last_minor_collection_start_tv;
1482 binary_protocol_collection_begin (gc_stats.minor_gc_count, GENERATION_NURSERY);
1484 if (do_verify_nursery || do_dump_nursery_content)
1485 sgen_debug_verify_nursery (do_dump_nursery_content);
1487 current_collection_generation = GENERATION_NURSERY;
1489 SGEN_ASSERT (0, !sgen_collection_is_concurrent (), "Why is the nursery collection concurrent?");
1491 reset_pinned_from_failed_allocation ();
1493 check_scan_starts ();
1495 sgen_nursery_alloc_prepare_for_minor ();
1499 nursery_next = sgen_nursery_alloc_get_upper_alloc_bound ();
1500 /* FIXME: optimize later to use the higher address where an object can be present */
1501 nursery_next = MAX (nursery_next, sgen_get_nursery_end ());
1503 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 ()));
1504 max_garbage_amount = nursery_next - sgen_get_nursery_start ();
1505 g_assert (nursery_section->size >= max_garbage_amount);
1507 /* world must be stopped already */
1509 time_minor_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1511 sgen_client_pre_collection_checks ();
1513 nursery_section->next_data = nursery_next;
1515 major_collector.start_nursery_collection ();
1517 sgen_memgov_minor_collection_start ();
1519 init_gray_queue (FALSE);
1521 gc_stats.minor_gc_count ++;
1523 if (whole_heap_check_before_collection) {
1524 sgen_clear_nursery_fragments ();
1525 sgen_check_whole_heap (finish_up_concurrent_mark);
1527 if (consistency_check_at_minor_collection)
1528 sgen_check_consistency ();
1530 sgen_process_fin_stage_entries ();
1532 /* pin from pinned handles */
1533 sgen_init_pinning ();
1534 sgen_client_binary_protocol_mark_start (GENERATION_NURSERY);
1535 pin_from_roots (sgen_get_nursery_start (), nursery_next, ctx);
1536 /* pin cemented objects */
1537 sgen_pin_cemented_objects ();
1538 /* identify pinned objects */
1539 sgen_optimize_pin_queue ();
1540 sgen_pinning_setup_section (nursery_section);
1542 pin_objects_in_nursery (FALSE, ctx);
1543 sgen_pinning_trim_queue_to_section (nursery_section);
1546 time_minor_pinning += TV_ELAPSED (btv, atv);
1547 SGEN_LOG (2, "Finding pinned pointers: %zd in %ld usecs", sgen_get_pinned_count (), TV_ELAPSED (btv, atv));
1548 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
1551 * FIXME: When we finish a concurrent collection we do a nursery collection first,
1552 * as part of which we scan the card table. Then, later, we scan the mod union
1553 * cardtable. We should only have to do one.
1555 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan remset", job_remembered_set_scan, sizeof (ScanJob));
1556 sj->ops = object_ops;
1557 sgen_workers_enqueue_job (&sj->job, FALSE);
1559 /* we don't have complete write barrier yet, so we scan all the old generation sections */
1561 time_minor_scan_remsets += TV_ELAPSED (atv, btv);
1562 SGEN_LOG (2, "Old generation scan: %ld usecs", TV_ELAPSED (atv, btv));
1564 sgen_pin_stats_print_class_stats ();
1566 sgen_drain_gray_stack (ctx);
1568 /* FIXME: Why do we do this at this specific, seemingly random, point? */
1569 sgen_client_collecting_minor (&fin_ready_queue, &critical_fin_queue);
1572 time_minor_scan_pinned += TV_ELAPSED (btv, atv);
1574 enqueue_scan_from_roots_jobs (sgen_get_nursery_start (), nursery_next, object_ops, FALSE);
1577 time_minor_scan_roots += TV_ELAPSED (atv, btv);
1579 finish_gray_stack (GENERATION_NURSERY, ctx);
1582 time_minor_finish_gray_stack += TV_ELAPSED (btv, atv);
1583 sgen_client_binary_protocol_mark_end (GENERATION_NURSERY);
1585 if (objects_pinned) {
1586 sgen_optimize_pin_queue ();
1587 sgen_pinning_setup_section (nursery_section);
1590 /* walk the pin_queue, build up the fragment list of free memory, unmark
1591 * pinned objects as we go, memzero() the empty fragments so they are ready for the
1594 sgen_client_binary_protocol_reclaim_start (GENERATION_NURSERY);
1595 fragment_total = sgen_build_nursery_fragments (nursery_section, unpin_queue);
1596 if (!fragment_total)
1599 /* Clear TLABs for all threads */
1600 sgen_clear_tlabs ();
1602 sgen_client_binary_protocol_reclaim_end (GENERATION_NURSERY);
1604 time_minor_fragment_creation += TV_ELAPSED (atv, btv);
1605 SGEN_LOG (2, "Fragment creation: %ld usecs, %lu bytes available", TV_ELAPSED (atv, btv), (unsigned long)fragment_total);
1607 if (consistency_check_at_minor_collection)
1608 sgen_check_major_refs ();
1610 major_collector.finish_nursery_collection ();
1612 TV_GETTIME (last_minor_collection_end_tv);
1613 gc_stats.minor_gc_time += TV_ELAPSED (last_minor_collection_start_tv, last_minor_collection_end_tv);
1615 sgen_debug_dump_heap ("minor", gc_stats.minor_gc_count - 1, NULL);
1617 /* prepare the pin queue for the next collection */
1618 sgen_finish_pinning ();
1619 if (sgen_have_pending_finalizers ()) {
1620 SGEN_LOG (4, "Finalizer-thread wakeup");
1621 sgen_client_finalize_notify ();
1623 sgen_pin_stats_reset ();
1624 /* clear cemented hash */
1625 sgen_cement_clear_below_threshold ();
1627 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
1629 remset.finish_minor_collection ();
1631 check_scan_starts ();
1633 binary_protocol_flush_buffers (FALSE);
1635 sgen_memgov_minor_collection_end ();
1637 /*objects are late pinned because of lack of memory, so a major is a good call*/
1638 needs_major = objects_pinned > 0;
1639 current_collection_generation = -1;
1642 binary_protocol_collection_end (gc_stats.minor_gc_count - 1, GENERATION_NURSERY, 0, 0);
1644 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
1645 sgen_check_nursery_objects_pinned (unpin_queue != NULL);
1651 COPY_OR_MARK_FROM_ROOTS_SERIAL,
1652 COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT,
1653 COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT
1654 } CopyOrMarkFromRootsMode;
1657 major_copy_or_mark_from_roots (size_t *old_next_pin_slot, CopyOrMarkFromRootsMode mode, SgenObjectOperations *object_ops)
1662 /* FIXME: only use these values for the precise scan
1663 * note that to_space pointers should be excluded anyway...
1665 char *heap_start = NULL;
1666 char *heap_end = (char*)-1;
1667 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, WORKERS_DISTRIBUTE_GRAY_QUEUE);
1668 gboolean concurrent = mode != COPY_OR_MARK_FROM_ROOTS_SERIAL;
1670 SGEN_ASSERT (0, !!concurrent == !!concurrent_collection_in_progress, "We've been called with the wrong mode.");
1672 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1673 /*This cleans up unused fragments */
1674 sgen_nursery_allocator_prepare_for_pinning ();
1676 if (do_concurrent_checks)
1677 sgen_debug_check_nursery_is_clean ();
1679 /* The concurrent collector doesn't touch the nursery. */
1680 sgen_nursery_alloc_prepare_for_major ();
1683 init_gray_queue (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT);
1687 /* Pinning depends on this */
1688 sgen_clear_nursery_fragments ();
1690 if (whole_heap_check_before_collection)
1691 sgen_check_whole_heap (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT);
1694 time_major_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1696 if (!sgen_collection_is_concurrent ())
1697 nursery_section->next_data = sgen_get_nursery_end ();
1698 /* we should also coalesce scanning from sections close to each other
1699 * and deal with pointers outside of the sections later.
1704 sgen_client_pre_collection_checks ();
1707 /* Remsets are not useful for a major collection */
1708 remset.clear_cards ();
1711 sgen_process_fin_stage_entries ();
1714 sgen_init_pinning ();
1715 SGEN_LOG (6, "Collecting pinned addresses");
1716 pin_from_roots ((void*)lowest_heap_address, (void*)highest_heap_address, ctx);
1718 sgen_optimize_pin_queue ();
1720 sgen_client_collecting_major_1 ();
1723 * pin_queue now contains all candidate pointers, sorted and
1724 * uniqued. We must do two passes now to figure out which
1725 * objects are pinned.
1727 * The first is to find within the pin_queue the area for each
1728 * section. This requires that the pin_queue be sorted. We
1729 * also process the LOS objects and pinned chunks here.
1731 * The second, destructive, pass is to reduce the section
1732 * areas to pointers to the actually pinned objects.
1734 SGEN_LOG (6, "Pinning from sections");
1735 /* first pass for the sections */
1736 sgen_find_section_pin_queue_start_end (nursery_section);
1737 /* identify possible pointers to the insize of large objects */
1738 SGEN_LOG (6, "Pinning from large objects");
1739 for (bigobj = los_object_list; bigobj; bigobj = bigobj->next) {
1741 if (sgen_find_optimized_pin_queue_area ((char*)bigobj->data, (char*)bigobj->data + sgen_los_object_size (bigobj), &dummy, &dummy)) {
1742 binary_protocol_pin (bigobj->data, (gpointer)LOAD_VTABLE (bigobj->data), safe_object_get_size (bigobj->data));
1744 if (sgen_los_object_is_pinned (bigobj->data)) {
1745 SGEN_ASSERT (0, mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT, "LOS objects can only be pinned here after concurrent marking.");
1748 sgen_los_pin_object (bigobj->data);
1749 if (SGEN_OBJECT_HAS_REFERENCES (bigobj->data))
1750 GRAY_OBJECT_ENQUEUE (WORKERS_DISTRIBUTE_GRAY_QUEUE, bigobj->data, sgen_obj_get_descriptor ((GCObject*)bigobj->data));
1751 sgen_pin_stats_register_object (bigobj->data, safe_object_get_size (bigobj->data));
1752 SGEN_LOG (6, "Marked large object %p (%s) size: %lu from roots", bigobj->data,
1753 sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (bigobj->data)),
1754 (unsigned long)sgen_los_object_size (bigobj));
1756 sgen_client_pinned_los_object (bigobj->data);
1760 pin_objects_in_nursery (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT, ctx);
1761 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
1762 sgen_check_nursery_objects_pinned (mode != COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT);
1764 major_collector.pin_objects (WORKERS_DISTRIBUTE_GRAY_QUEUE);
1765 if (old_next_pin_slot)
1766 *old_next_pin_slot = sgen_get_pinned_count ();
1769 time_major_pinning += TV_ELAPSED (atv, btv);
1770 SGEN_LOG (2, "Finding pinned pointers: %zd in %ld usecs", sgen_get_pinned_count (), TV_ELAPSED (atv, btv));
1771 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
1773 major_collector.init_to_space ();
1776 * The concurrent collector doesn't move objects, neither on
1777 * the major heap nor in the nursery, so we can mark even
1778 * before pinning has finished. For the non-concurrent
1779 * collector we start the workers after pinning.
1781 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1782 SGEN_ASSERT (0, sgen_workers_all_done (), "Why are the workers not done when we start or finish a major collection?");
1783 sgen_workers_start_all_workers (object_ops);
1784 gray_queue_enable_redirect (WORKERS_DISTRIBUTE_GRAY_QUEUE);
1785 } else if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
1786 if (sgen_workers_have_idle_work ()) {
1787 sgen_workers_start_all_workers (object_ops);
1788 sgen_workers_join ();
1792 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
1793 main_gc_thread = mono_native_thread_self ();
1796 sgen_client_collecting_major_2 ();
1799 time_major_scan_pinned += TV_ELAPSED (btv, atv);
1801 sgen_client_collecting_major_3 (&fin_ready_queue, &critical_fin_queue);
1804 * FIXME: is this the right context? It doesn't seem to contain a copy function
1805 * unless we're concurrent.
1807 enqueue_scan_from_roots_jobs (heap_start, heap_end, object_ops, mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT);
1810 time_major_scan_roots += TV_ELAPSED (atv, btv);
1812 if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
1815 /* Mod union card table */
1816 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan mod union cardtable", job_scan_major_mod_union_card_table, sizeof (ScanJob));
1817 sj->ops = object_ops;
1818 sgen_workers_enqueue_job (&sj->job, FALSE);
1820 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan LOS mod union cardtable", job_scan_los_mod_union_card_table, sizeof (ScanJob));
1821 sj->ops = object_ops;
1822 sgen_workers_enqueue_job (&sj->job, FALSE);
1825 time_major_scan_mod_union += TV_ELAPSED (btv, atv);
1828 sgen_pin_stats_print_class_stats ();
1832 major_finish_copy_or_mark (CopyOrMarkFromRootsMode mode)
1834 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1836 * Prepare the pin queue for the next collection. Since pinning runs on the worker
1837 * threads we must wait for the jobs to finish before we can reset it.
1839 sgen_workers_wait_for_jobs_finished ();
1840 sgen_finish_pinning ();
1842 sgen_pin_stats_reset ();
1844 if (do_concurrent_checks)
1845 sgen_debug_check_nursery_is_clean ();
1850 major_start_collection (gboolean concurrent, size_t *old_next_pin_slot)
1852 SgenObjectOperations *object_ops;
1854 binary_protocol_collection_begin (gc_stats.major_gc_count, GENERATION_OLD);
1856 current_collection_generation = GENERATION_OLD;
1858 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
1861 sgen_cement_reset ();
1864 g_assert (major_collector.is_concurrent);
1865 concurrent_collection_in_progress = TRUE;
1867 object_ops = &major_collector.major_ops_concurrent_start;
1869 object_ops = &major_collector.major_ops_serial;
1872 reset_pinned_from_failed_allocation ();
1874 sgen_memgov_major_collection_start ();
1876 //count_ref_nonref_objs ();
1877 //consistency_check ();
1879 check_scan_starts ();
1882 SGEN_LOG (1, "Start major collection %d", gc_stats.major_gc_count);
1883 gc_stats.major_gc_count ++;
1885 if (major_collector.start_major_collection)
1886 major_collector.start_major_collection ();
1888 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);
1889 major_finish_copy_or_mark (concurrent ? COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT : COPY_OR_MARK_FROM_ROOTS_SERIAL);
1893 major_finish_collection (const char *reason, size_t old_next_pin_slot, gboolean forced)
1895 ScannedObjectCounts counts;
1896 SgenObjectOperations *object_ops;
1897 mword fragment_total;
1903 if (concurrent_collection_in_progress) {
1904 object_ops = &major_collector.major_ops_concurrent_finish;
1906 major_copy_or_mark_from_roots (NULL, COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT, object_ops);
1908 major_finish_copy_or_mark (COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT);
1910 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
1911 main_gc_thread = NULL;
1914 object_ops = &major_collector.major_ops_serial;
1917 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
1919 /* all the objects in the heap */
1920 finish_gray_stack (GENERATION_OLD, CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, &gray_queue));
1922 time_major_finish_gray_stack += TV_ELAPSED (btv, atv);
1924 SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after joining");
1926 if (objects_pinned) {
1927 g_assert (!concurrent_collection_in_progress);
1930 * This is slow, but we just OOM'd.
1932 * See comment at `sgen_pin_queue_clear_discarded_entries` for how the pin
1933 * queue is laid out at this point.
1935 sgen_pin_queue_clear_discarded_entries (nursery_section, old_next_pin_slot);
1937 * We need to reestablish all pinned nursery objects in the pin queue
1938 * because they're needed for fragment creation. Unpinning happens by
1939 * walking the whole queue, so it's not necessary to reestablish where major
1940 * heap block pins are - all we care is that they're still in there
1943 sgen_optimize_pin_queue ();
1944 sgen_find_section_pin_queue_start_end (nursery_section);
1948 reset_heap_boundaries ();
1949 sgen_update_heap_boundaries ((mword)sgen_get_nursery_start (), (mword)sgen_get_nursery_end ());
1951 /* walk the pin_queue, build up the fragment list of free memory, unmark
1952 * pinned objects as we go, memzero() the empty fragments so they are ready for the
1955 fragment_total = sgen_build_nursery_fragments (nursery_section, NULL);
1956 if (!fragment_total)
1958 SGEN_LOG (4, "Free space in nursery after major %ld", (long)fragment_total);
1960 if (do_concurrent_checks && concurrent_collection_in_progress)
1961 sgen_debug_check_nursery_is_clean ();
1963 /* prepare the pin queue for the next collection */
1964 sgen_finish_pinning ();
1966 /* Clear TLABs for all threads */
1967 sgen_clear_tlabs ();
1969 sgen_pin_stats_reset ();
1971 sgen_cement_clear_below_threshold ();
1973 if (check_mark_bits_after_major_collection)
1974 sgen_check_heap_marked (concurrent_collection_in_progress);
1977 time_major_fragment_creation += TV_ELAPSED (atv, btv);
1979 binary_protocol_sweep_begin (GENERATION_OLD, !major_collector.sweeps_lazily);
1982 time_major_free_bigobjs += TV_ELAPSED (btv, atv);
1987 time_major_los_sweep += TV_ELAPSED (atv, btv);
1989 major_collector.sweep ();
1991 binary_protocol_sweep_end (GENERATION_OLD, !major_collector.sweeps_lazily);
1994 time_major_sweep += TV_ELAPSED (btv, atv);
1996 sgen_debug_dump_heap ("major", gc_stats.major_gc_count - 1, reason);
1998 if (sgen_have_pending_finalizers ()) {
1999 SGEN_LOG (4, "Finalizer-thread wakeup");
2000 sgen_client_finalize_notify ();
2003 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
2005 sgen_memgov_major_collection_end (forced);
2006 current_collection_generation = -1;
2008 memset (&counts, 0, sizeof (ScannedObjectCounts));
2009 major_collector.finish_major_collection (&counts);
2011 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
2013 SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after major collection has finished");
2014 if (concurrent_collection_in_progress)
2015 concurrent_collection_in_progress = FALSE;
2017 check_scan_starts ();
2019 binary_protocol_flush_buffers (FALSE);
2021 //consistency_check ();
2023 binary_protocol_collection_end (gc_stats.major_gc_count - 1, GENERATION_OLD, counts.num_scanned_objects, counts.num_unique_scanned_objects);
2027 major_do_collection (const char *reason, gboolean forced)
2029 TV_DECLARE (time_start);
2030 TV_DECLARE (time_end);
2031 size_t old_next_pin_slot;
2033 if (disable_major_collections)
2036 if (major_collector.get_and_reset_num_major_objects_marked) {
2037 long long num_marked = major_collector.get_and_reset_num_major_objects_marked ();
2038 g_assert (!num_marked);
2041 /* world must be stopped already */
2042 TV_GETTIME (time_start);
2044 major_start_collection (FALSE, &old_next_pin_slot);
2045 major_finish_collection (reason, old_next_pin_slot, forced);
2047 TV_GETTIME (time_end);
2048 gc_stats.major_gc_time += TV_ELAPSED (time_start, time_end);
2050 /* FIXME: also report this to the user, preferably in gc-end. */
2051 if (major_collector.get_and_reset_num_major_objects_marked)
2052 major_collector.get_and_reset_num_major_objects_marked ();
2054 return bytes_pinned_from_failed_allocation > 0;
2058 major_start_concurrent_collection (const char *reason)
2060 TV_DECLARE (time_start);
2061 TV_DECLARE (time_end);
2062 long long num_objects_marked;
2064 if (disable_major_collections)
2067 TV_GETTIME (time_start);
2068 SGEN_TV_GETTIME (time_major_conc_collection_start);
2070 num_objects_marked = major_collector.get_and_reset_num_major_objects_marked ();
2071 g_assert (num_objects_marked == 0);
2073 binary_protocol_concurrent_start ();
2075 // FIXME: store reason and pass it when finishing
2076 major_start_collection (TRUE, NULL);
2078 gray_queue_redirect (&gray_queue);
2080 num_objects_marked = major_collector.get_and_reset_num_major_objects_marked ();
2082 TV_GETTIME (time_end);
2083 gc_stats.major_gc_time += TV_ELAPSED (time_start, time_end);
2085 current_collection_generation = -1;
2089 * Returns whether the major collection has finished.
2092 major_should_finish_concurrent_collection (void)
2094 SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&gray_queue), "Why is the gray queue not empty before we have started doing anything?");
2095 return sgen_workers_all_done ();
2099 major_update_concurrent_collection (void)
2101 TV_DECLARE (total_start);
2102 TV_DECLARE (total_end);
2104 TV_GETTIME (total_start);
2106 binary_protocol_concurrent_update ();
2108 major_collector.update_cardtable_mod_union ();
2109 sgen_los_update_cardtable_mod_union ();
2111 TV_GETTIME (total_end);
2112 gc_stats.major_gc_time += TV_ELAPSED (total_start, total_end);
2116 major_finish_concurrent_collection (gboolean forced)
2118 TV_DECLARE (total_start);
2119 TV_DECLARE (total_end);
2121 TV_GETTIME (total_start);
2123 binary_protocol_concurrent_finish ();
2126 * We need to stop all workers since we're updating the cardtable below.
2127 * The workers will be resumed with a finishing pause context to avoid
2128 * additional cardtable and object scanning.
2130 sgen_workers_stop_all_workers ();
2132 SGEN_TV_GETTIME (time_major_conc_collection_end);
2133 gc_stats.major_gc_time_concurrent += SGEN_TV_ELAPSED (time_major_conc_collection_start, time_major_conc_collection_end);
2135 major_collector.update_cardtable_mod_union ();
2136 sgen_los_update_cardtable_mod_union ();
2138 if (mod_union_consistency_check)
2139 sgen_check_mod_union_consistency ();
2141 current_collection_generation = GENERATION_OLD;
2142 sgen_cement_reset ();
2143 major_finish_collection ("finishing", -1, forced);
2145 if (whole_heap_check_before_collection)
2146 sgen_check_whole_heap (FALSE);
2148 TV_GETTIME (total_end);
2149 gc_stats.major_gc_time += TV_ELAPSED (total_start, total_end) - TV_ELAPSED (last_minor_collection_start_tv, last_minor_collection_end_tv);
2151 current_collection_generation = -1;
2155 * Ensure an allocation request for @size will succeed by freeing enough memory.
2157 * LOCKING: The GC lock MUST be held.
2160 sgen_ensure_free_space (size_t size)
2162 int generation_to_collect = -1;
2163 const char *reason = NULL;
2165 if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
2166 if (sgen_need_major_collection (size)) {
2167 reason = "LOS overflow";
2168 generation_to_collect = GENERATION_OLD;
2171 if (degraded_mode) {
2172 if (sgen_need_major_collection (size)) {
2173 reason = "Degraded mode overflow";
2174 generation_to_collect = GENERATION_OLD;
2176 } else if (sgen_need_major_collection (size)) {
2177 reason = concurrent_collection_in_progress ? "Forced finish concurrent collection" : "Minor allowance";
2178 generation_to_collect = GENERATION_OLD;
2180 generation_to_collect = GENERATION_NURSERY;
2181 reason = "Nursery full";
2185 if (generation_to_collect == -1) {
2186 if (concurrent_collection_in_progress && sgen_workers_all_done ()) {
2187 generation_to_collect = GENERATION_OLD;
2188 reason = "Finish concurrent collection";
2192 if (generation_to_collect == -1)
2194 sgen_perform_collection (size, generation_to_collect, reason, FALSE);
2198 * LOCKING: Assumes the GC lock is held.
2201 sgen_perform_collection (size_t requested_size, int generation_to_collect, const char *reason, gboolean wait_to_finish)
2203 TV_DECLARE (gc_start);
2204 TV_DECLARE (gc_end);
2205 TV_DECLARE (gc_total_start);
2206 TV_DECLARE (gc_total_end);
2207 GGTimingInfo infos [2];
2208 int overflow_generation_to_collect = -1;
2209 int oldest_generation_collected = generation_to_collect;
2210 const char *overflow_reason = NULL;
2212 binary_protocol_collection_requested (generation_to_collect, requested_size, wait_to_finish ? 1 : 0);
2214 SGEN_ASSERT (0, generation_to_collect == GENERATION_NURSERY || generation_to_collect == GENERATION_OLD, "What generation is this?");
2216 TV_GETTIME (gc_start);
2218 sgen_stop_world (generation_to_collect);
2220 TV_GETTIME (gc_total_start);
2222 if (concurrent_collection_in_progress) {
2224 * If the concurrent worker is finished or we are asked to do a major collection
2225 * then we finish the concurrent collection.
2227 gboolean finish = major_should_finish_concurrent_collection () || generation_to_collect == GENERATION_OLD;
2230 major_finish_concurrent_collection (wait_to_finish);
2231 oldest_generation_collected = GENERATION_OLD;
2233 SGEN_ASSERT (0, generation_to_collect == GENERATION_NURSERY, "Why aren't we finishing the concurrent collection?");
2234 major_update_concurrent_collection ();
2235 collect_nursery (NULL, FALSE);
2241 SGEN_ASSERT (0, !concurrent_collection_in_progress, "Why did this not get handled above?");
2244 * There's no concurrent collection in progress. Collect the generation we're asked
2245 * to collect. If the major collector is concurrent and we're not forced to wait,
2246 * start a concurrent collection.
2248 // FIXME: extract overflow reason
2249 if (generation_to_collect == GENERATION_NURSERY) {
2250 if (collect_nursery (NULL, FALSE)) {
2251 overflow_generation_to_collect = GENERATION_OLD;
2252 overflow_reason = "Minor overflow";
2255 if (major_collector.is_concurrent && !wait_to_finish) {
2256 collect_nursery (NULL, FALSE);
2257 major_start_concurrent_collection (reason);
2258 // FIXME: set infos[0] properly
2262 if (major_do_collection (reason, wait_to_finish)) {
2263 overflow_generation_to_collect = GENERATION_NURSERY;
2264 overflow_reason = "Excessive pinning";
2268 TV_GETTIME (gc_end);
2270 memset (infos, 0, sizeof (infos));
2271 infos [0].generation = generation_to_collect;
2272 infos [0].reason = reason;
2273 infos [0].is_overflow = FALSE;
2274 infos [1].generation = -1;
2275 infos [0].total_time = SGEN_TV_ELAPSED (gc_start, gc_end);
2277 SGEN_ASSERT (0, !concurrent_collection_in_progress, "Why did this not get handled above?");
2279 if (overflow_generation_to_collect != -1) {
2281 * We need to do an overflow collection, either because we ran out of memory
2282 * or the nursery is fully pinned.
2285 infos [1].generation = overflow_generation_to_collect;
2286 infos [1].reason = overflow_reason;
2287 infos [1].is_overflow = TRUE;
2290 if (overflow_generation_to_collect == GENERATION_NURSERY)
2291 collect_nursery (NULL, FALSE);
2293 major_do_collection (overflow_reason, wait_to_finish);
2295 TV_GETTIME (gc_end);
2296 infos [1].total_time = SGEN_TV_ELAPSED (gc_start, gc_end);
2298 oldest_generation_collected = MAX (oldest_generation_collected, overflow_generation_to_collect);
2301 SGEN_LOG (2, "Heap size: %lu, LOS size: %lu", (unsigned long)sgen_gc_get_total_heap_allocation (), (unsigned long)los_memory_usage);
2303 /* this also sets the proper pointers for the next allocation */
2304 if (generation_to_collect == GENERATION_NURSERY && !sgen_can_alloc_size (requested_size)) {
2305 /* TypeBuilder and MonoMethod are killing mcs with fragmentation */
2306 SGEN_LOG (1, "nursery collection didn't find enough room for %zd alloc (%zd pinned)", requested_size, sgen_get_pinned_count ());
2307 sgen_dump_pin_queue ();
2312 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
2314 TV_GETTIME (gc_total_end);
2315 time_max = MAX (time_max, TV_ELAPSED (gc_total_start, gc_total_end));
2317 sgen_restart_world (oldest_generation_collected, infos);
2321 * ######################################################################
2322 * ######## Memory allocation from the OS
2323 * ######################################################################
2324 * This section of code deals with getting memory from the OS and
2325 * allocating memory for GC-internal data structures.
2326 * Internal memory can be handled with a freelist for small objects.
2332 G_GNUC_UNUSED static void
2333 report_internal_mem_usage (void)
2335 printf ("Internal memory usage:\n");
2336 sgen_report_internal_mem_usage ();
2337 printf ("Pinned memory usage:\n");
2338 major_collector.report_pinned_memory_usage ();
2342 * ######################################################################
2343 * ######## Finalization support
2344 * ######################################################################
2348 * If the object has been forwarded it means it's still referenced from a root.
2349 * If it is pinned it's still alive as well.
2350 * A LOS object is only alive if we have pinned it.
2351 * Return TRUE if @obj is ready to be finalized.
2353 static inline gboolean
2354 sgen_is_object_alive (GCObject *object)
2356 if (ptr_in_nursery (object))
2357 return sgen_nursery_is_object_alive (object);
2359 return sgen_major_is_object_alive (object);
2363 * This function returns true if @object is either alive and belongs to the
2364 * current collection - major collections are full heap, so old gen objects
2365 * are never alive during a minor collection.
2368 sgen_is_object_alive_and_on_current_collection (GCObject *object)
2370 if (ptr_in_nursery (object))
2371 return sgen_nursery_is_object_alive (object);
2373 if (current_collection_generation == GENERATION_NURSERY)
2376 return sgen_major_is_object_alive (object);
2381 sgen_gc_is_object_ready_for_finalization (GCObject *object)
2383 return !sgen_is_object_alive (object);
2387 sgen_queue_finalization_entry (GCObject *obj)
2389 gboolean critical = sgen_client_object_has_critical_finalizer (obj);
2391 sgen_pointer_queue_add (critical ? &critical_fin_queue : &fin_ready_queue, obj);
2393 sgen_client_object_queued_for_finalization (obj);
2397 sgen_object_is_live (GCObject *obj)
2399 return sgen_is_object_alive_and_on_current_collection (obj);
2403 * `System.GC.WaitForPendingFinalizers` first checks `sgen_have_pending_finalizers()` to
2404 * determine whether it can exit quickly. The latter must therefore only return FALSE if
2405 * all finalizers have really finished running.
2407 * `sgen_gc_invoke_finalizers()` first dequeues a finalizable object, and then finalizes it.
2408 * This means that just checking whether the queues are empty leaves the possibility that an
2409 * object might have been dequeued but not yet finalized. That's why we need the additional
2410 * flag `pending_unqueued_finalizer`.
2413 static volatile gboolean pending_unqueued_finalizer = FALSE;
2416 sgen_gc_invoke_finalizers (void)
2420 g_assert (!pending_unqueued_finalizer);
2422 /* FIXME: batch to reduce lock contention */
2423 while (sgen_have_pending_finalizers ()) {
2429 * We need to set `pending_unqueued_finalizer` before dequeing the
2430 * finalizable object.
2432 if (!sgen_pointer_queue_is_empty (&fin_ready_queue)) {
2433 pending_unqueued_finalizer = TRUE;
2434 mono_memory_write_barrier ();
2435 obj = (GCObject *)sgen_pointer_queue_pop (&fin_ready_queue);
2436 } else if (!sgen_pointer_queue_is_empty (&critical_fin_queue)) {
2437 pending_unqueued_finalizer = TRUE;
2438 mono_memory_write_barrier ();
2439 obj = (GCObject *)sgen_pointer_queue_pop (&critical_fin_queue);
2445 SGEN_LOG (7, "Finalizing object %p (%s)", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
2453 /* the object is on the stack so it is pinned */
2454 /*g_print ("Calling finalizer for object: %p (%s)\n", obj, sgen_client_object_safe_name (obj));*/
2455 sgen_client_run_finalize (obj);
2458 if (pending_unqueued_finalizer) {
2459 mono_memory_write_barrier ();
2460 pending_unqueued_finalizer = FALSE;
2467 sgen_have_pending_finalizers (void)
2469 return pending_unqueued_finalizer || !sgen_pointer_queue_is_empty (&fin_ready_queue) || !sgen_pointer_queue_is_empty (&critical_fin_queue);
2473 * ######################################################################
2474 * ######## registered roots support
2475 * ######################################################################
2479 * We do not coalesce roots.
2482 sgen_register_root (char *start, size_t size, SgenDescriptor descr, int root_type, int source, const char *msg)
2484 RootRecord new_root;
2487 for (i = 0; i < ROOT_TYPE_NUM; ++i) {
2488 RootRecord *root = (RootRecord *)sgen_hash_table_lookup (&roots_hash [i], start);
2489 /* we allow changing the size and the descriptor (for thread statics etc) */
2491 size_t old_size = root->end_root - start;
2492 root->end_root = start + size;
2493 SGEN_ASSERT (0, !!root->root_desc == !!descr, "Can't change whether a root is precise or conservative.");
2494 SGEN_ASSERT (0, root->source == source, "Can't change a root's source identifier.");
2495 SGEN_ASSERT (0, !!root->msg == !!msg, "Can't change a root's message.");
2496 root->root_desc = descr;
2498 roots_size -= old_size;
2504 new_root.end_root = start + size;
2505 new_root.root_desc = descr;
2506 new_root.source = source;
2509 sgen_hash_table_replace (&roots_hash [root_type], start, &new_root, NULL);
2512 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);
2519 sgen_deregister_root (char* addr)
2525 for (root_type = 0; root_type < ROOT_TYPE_NUM; ++root_type) {
2526 if (sgen_hash_table_remove (&roots_hash [root_type], addr, &root))
2527 roots_size -= (root.end_root - addr);
2533 * ######################################################################
2534 * ######## Thread handling (stop/start code)
2535 * ######################################################################
2539 sgen_get_current_collection_generation (void)
2541 return current_collection_generation;
2545 sgen_thread_register (SgenThreadInfo* info, void *stack_bottom_fallback)
2547 #ifndef HAVE_KW_THREAD
2548 info->tlab_start = info->tlab_next = info->tlab_temp_end = info->tlab_real_end = NULL;
2551 sgen_init_tlab_info (info);
2553 sgen_client_thread_register (info, stack_bottom_fallback);
2559 sgen_thread_unregister (SgenThreadInfo *p)
2561 sgen_client_thread_unregister (p);
2565 * ######################################################################
2566 * ######## Write barriers
2567 * ######################################################################
2571 * Note: the write barriers first do the needed GC work and then do the actual store:
2572 * this way the value is visible to the conservative GC scan after the write barrier
2573 * itself. If a GC interrupts the barrier in the middle, value will be kept alive by
2574 * the conservative scan, otherwise by the remembered set scan.
2578 mono_gc_wbarrier_arrayref_copy (gpointer dest_ptr, gpointer src_ptr, int count)
2580 HEAVY_STAT (++stat_wbarrier_arrayref_copy);
2581 /*This check can be done without taking a lock since dest_ptr array is pinned*/
2582 if (ptr_in_nursery (dest_ptr) || count <= 0) {
2583 mono_gc_memmove_aligned (dest_ptr, src_ptr, count * sizeof (gpointer));
2587 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2588 if (binary_protocol_is_heavy_enabled ()) {
2590 for (i = 0; i < count; ++i) {
2591 gpointer dest = (gpointer*)dest_ptr + i;
2592 gpointer obj = *((gpointer*)src_ptr + i);
2594 binary_protocol_wbarrier (dest, obj, (gpointer)LOAD_VTABLE (obj));
2599 remset.wbarrier_arrayref_copy (dest_ptr, src_ptr, count);
2603 mono_gc_wbarrier_generic_nostore (gpointer ptr)
2607 HEAVY_STAT (++stat_wbarrier_generic_store);
2609 sgen_client_wbarrier_generic_nostore_check (ptr);
2611 obj = *(gpointer*)ptr;
2613 binary_protocol_wbarrier (ptr, obj, (gpointer)LOAD_VTABLE (obj));
2616 * We need to record old->old pointer locations for the
2617 * concurrent collector.
2619 if (!ptr_in_nursery (obj) && !concurrent_collection_in_progress) {
2620 SGEN_LOG (8, "Skipping remset at %p", ptr);
2624 SGEN_LOG (8, "Adding remset at %p", ptr);
2626 remset.wbarrier_generic_nostore (ptr);
2630 mono_gc_wbarrier_generic_store (gpointer ptr, GCObject* value)
2632 SGEN_LOG (8, "Wbarrier store at %p to %p (%s)", ptr, value, value ? sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (value)) : "null");
2633 SGEN_UPDATE_REFERENCE_ALLOW_NULL (ptr, value);
2634 if (ptr_in_nursery (value) || concurrent_collection_in_progress)
2635 mono_gc_wbarrier_generic_nostore (ptr);
2636 sgen_dummy_use (value);
2639 /* Same as mono_gc_wbarrier_generic_store () but performs the store
2640 * as an atomic operation with release semantics.
2643 mono_gc_wbarrier_generic_store_atomic (gpointer ptr, GCObject *value)
2645 HEAVY_STAT (++stat_wbarrier_generic_store_atomic);
2647 SGEN_LOG (8, "Wbarrier atomic store at %p to %p (%s)", ptr, value, value ? sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (value)) : "null");
2649 InterlockedWritePointer ((volatile gpointer *)ptr, value);
2651 if (ptr_in_nursery (value) || concurrent_collection_in_progress)
2652 mono_gc_wbarrier_generic_nostore (ptr);
2654 sgen_dummy_use (value);
2658 sgen_wbarrier_value_copy_bitmap (gpointer _dest, gpointer _src, int size, unsigned bitmap)
2660 GCObject **dest = (GCObject **)_dest;
2661 GCObject **src = (GCObject **)_src;
2665 mono_gc_wbarrier_generic_store (dest, *src);
2670 size -= SIZEOF_VOID_P;
2676 * ######################################################################
2677 * ######## Other mono public interface functions.
2678 * ######################################################################
2682 sgen_gc_collect (int generation)
2687 sgen_perform_collection (0, generation, "user request", TRUE);
2692 sgen_gc_collection_count (int generation)
2694 if (generation == 0)
2695 return gc_stats.minor_gc_count;
2696 return gc_stats.major_gc_count;
2700 sgen_gc_get_used_size (void)
2704 tot = los_memory_usage;
2705 tot += nursery_section->next_data - nursery_section->data;
2706 tot += major_collector.get_used_size ();
2707 /* FIXME: account for pinned objects */
2713 sgen_env_var_error (const char *env_var, const char *fallback, const char *description_format, ...)
2717 va_start (ap, description_format);
2719 fprintf (stderr, "Warning: In environment variable `%s': ", env_var);
2720 vfprintf (stderr, description_format, ap);
2722 fprintf (stderr, " - %s", fallback);
2723 fprintf (stderr, "\n");
2729 parse_double_in_interval (const char *env_var, const char *opt_name, const char *opt, double min, double max, double *result)
2732 double val = strtod (opt, &endptr);
2733 if (endptr == opt) {
2734 sgen_env_var_error (env_var, "Using default value.", "`%s` must be a number.", opt_name);
2737 else if (val < min || val > max) {
2738 sgen_env_var_error (env_var, "Using default value.", "`%s` must be between %.2f - %.2f.", opt_name, min, max);
2750 char *major_collector_opt = NULL;
2751 char *minor_collector_opt = NULL;
2752 size_t max_heap = 0;
2753 size_t soft_limit = 0;
2755 gboolean debug_print_allowance = FALSE;
2756 double allowance_ratio = 0, save_target = 0;
2757 gboolean cement_enabled = TRUE;
2760 result = InterlockedCompareExchange (&gc_initialized, -1, 0);
2763 /* already inited */
2766 /* being inited by another thread */
2767 mono_thread_info_usleep (1000);
2770 /* we will init it */
2773 g_assert_not_reached ();
2775 } while (result != 0);
2777 SGEN_TV_GETTIME (sgen_init_timestamp);
2779 #ifdef SGEN_WITHOUT_MONO
2780 mono_thread_smr_init ();
2783 mono_coop_mutex_init (&gc_mutex);
2785 gc_debug_file = stderr;
2787 mono_coop_mutex_init (&sgen_interruption_mutex);
2789 if ((env = g_getenv (MONO_GC_PARAMS_NAME))) {
2790 opts = g_strsplit (env, ",", -1);
2791 for (ptr = opts; *ptr; ++ptr) {
2793 if (g_str_has_prefix (opt, "major=")) {
2794 opt = strchr (opt, '=') + 1;
2795 major_collector_opt = g_strdup (opt);
2796 } else if (g_str_has_prefix (opt, "minor=")) {
2797 opt = strchr (opt, '=') + 1;
2798 minor_collector_opt = g_strdup (opt);
2806 sgen_init_internal_allocator ();
2807 sgen_init_nursery_allocator ();
2808 sgen_init_fin_weak_hash ();
2809 sgen_init_hash_table ();
2810 sgen_init_descriptors ();
2811 sgen_init_gray_queues ();
2812 sgen_init_allocator ();
2813 sgen_init_gchandles ();
2815 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_SECTION, SGEN_SIZEOF_GC_MEM_SECTION);
2816 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_GRAY_QUEUE, sizeof (GrayQueueSection));
2818 sgen_client_init ();
2820 if (!minor_collector_opt) {
2821 sgen_simple_nursery_init (&sgen_minor_collector);
2823 if (!strcmp (minor_collector_opt, "simple")) {
2825 sgen_simple_nursery_init (&sgen_minor_collector);
2826 } else if (!strcmp (minor_collector_opt, "split")) {
2827 sgen_split_nursery_init (&sgen_minor_collector);
2829 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using `simple` instead.", "Unknown minor collector `%s'.", minor_collector_opt);
2830 goto use_simple_nursery;
2834 if (!major_collector_opt || !strcmp (major_collector_opt, "marksweep")) {
2835 use_marksweep_major:
2836 sgen_marksweep_init (&major_collector);
2837 } else if (!major_collector_opt || !strcmp (major_collector_opt, "marksweep-conc")) {
2838 sgen_marksweep_conc_init (&major_collector);
2840 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using `marksweep` instead.", "Unknown major collector `%s'.", major_collector_opt);
2841 goto use_marksweep_major;
2844 sgen_nursery_size = DEFAULT_NURSERY_SIZE;
2847 gboolean usage_printed = FALSE;
2849 for (ptr = opts; *ptr; ++ptr) {
2851 if (!strcmp (opt, ""))
2853 if (g_str_has_prefix (opt, "major="))
2855 if (g_str_has_prefix (opt, "minor="))
2857 if (g_str_has_prefix (opt, "max-heap-size=")) {
2858 size_t page_size = mono_pagesize ();
2859 size_t max_heap_candidate = 0;
2860 opt = strchr (opt, '=') + 1;
2861 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &max_heap_candidate)) {
2862 max_heap = (max_heap_candidate + page_size - 1) & ~(size_t)(page_size - 1);
2863 if (max_heap != max_heap_candidate)
2864 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Rounding up.", "`max-heap-size` size must be a multiple of %d.", page_size);
2866 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`max-heap-size` must be an integer.");
2870 if (g_str_has_prefix (opt, "soft-heap-limit=")) {
2871 opt = strchr (opt, '=') + 1;
2872 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &soft_limit)) {
2873 if (soft_limit <= 0) {
2874 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be positive.");
2878 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be an integer.");
2884 if (g_str_has_prefix (opt, "nursery-size=")) {
2886 opt = strchr (opt, '=') + 1;
2887 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &val)) {
2888 if ((val & (val - 1))) {
2889 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be a power of two.");
2893 if (val < SGEN_MAX_NURSERY_WASTE) {
2894 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.",
2895 "`nursery-size` must be at least %d bytes.", SGEN_MAX_NURSERY_WASTE);
2899 sgen_nursery_size = val;
2900 sgen_nursery_bits = 0;
2901 while (ONE_P << (++ sgen_nursery_bits) != sgen_nursery_size)
2904 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be an integer.");
2910 if (g_str_has_prefix (opt, "save-target-ratio=")) {
2912 opt = strchr (opt, '=') + 1;
2913 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "save-target-ratio", opt,
2914 SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO, &val)) {
2919 if (g_str_has_prefix (opt, "default-allowance-ratio=")) {
2921 opt = strchr (opt, '=') + 1;
2922 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "default-allowance-ratio", opt,
2923 SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, &val)) {
2924 allowance_ratio = val;
2929 if (!strcmp (opt, "cementing")) {
2930 cement_enabled = TRUE;
2933 if (!strcmp (opt, "no-cementing")) {
2934 cement_enabled = FALSE;
2938 if (major_collector.handle_gc_param && major_collector.handle_gc_param (opt))
2941 if (sgen_minor_collector.handle_gc_param && sgen_minor_collector.handle_gc_param (opt))
2944 if (sgen_client_handle_gc_param (opt))
2947 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring.", "Unknown option `%s`.", opt);
2952 fprintf (stderr, "\n%s must be a comma-delimited list of one or more of the following:\n", MONO_GC_PARAMS_NAME);
2953 fprintf (stderr, " max-heap-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
2954 fprintf (stderr, " soft-heap-limit=n (where N is an integer, possibly with a k, m or a g suffix)\n");
2955 fprintf (stderr, " nursery-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
2956 fprintf (stderr, " major=COLLECTOR (where COLLECTOR is `marksweep', `marksweep-conc', `marksweep-par')\n");
2957 fprintf (stderr, " minor=COLLECTOR (where COLLECTOR is `simple' or `split')\n");
2958 fprintf (stderr, " wbarrier=WBARRIER (where WBARRIER is `remset' or `cardtable')\n");
2959 fprintf (stderr, " [no-]cementing\n");
2960 if (major_collector.print_gc_param_usage)
2961 major_collector.print_gc_param_usage ();
2962 if (sgen_minor_collector.print_gc_param_usage)
2963 sgen_minor_collector.print_gc_param_usage ();
2964 sgen_client_print_gc_params_usage ();
2965 fprintf (stderr, " Experimental options:\n");
2966 fprintf (stderr, " save-target-ratio=R (where R must be between %.2f - %.2f).\n", SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO);
2967 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);
2968 fprintf (stderr, "\n");
2970 usage_printed = TRUE;
2975 if (major_collector_opt)
2976 g_free (major_collector_opt);
2978 if (minor_collector_opt)
2979 g_free (minor_collector_opt);
2983 sgen_cement_init (cement_enabled);
2985 if ((env = g_getenv (MONO_GC_DEBUG_NAME))) {
2986 gboolean usage_printed = FALSE;
2988 opts = g_strsplit (env, ",", -1);
2989 for (ptr = opts; ptr && *ptr; ptr ++) {
2991 if (!strcmp (opt, ""))
2993 if (opt [0] >= '0' && opt [0] <= '9') {
2994 gc_debug_level = atoi (opt);
2999 char *rf = g_strdup_printf ("%s.%d", opt, mono_process_current_pid ());
3000 gc_debug_file = fopen (rf, "wb");
3002 gc_debug_file = stderr;
3005 } else if (!strcmp (opt, "print-allowance")) {
3006 debug_print_allowance = TRUE;
3007 } else if (!strcmp (opt, "print-pinning")) {
3008 sgen_pin_stats_enable ();
3009 } else if (!strcmp (opt, "verify-before-allocs")) {
3010 verify_before_allocs = 1;
3011 has_per_allocation_action = TRUE;
3012 } else if (g_str_has_prefix (opt, "verify-before-allocs=")) {
3013 char *arg = strchr (opt, '=') + 1;
3014 verify_before_allocs = atoi (arg);
3015 has_per_allocation_action = TRUE;
3016 } else if (!strcmp (opt, "collect-before-allocs")) {
3017 collect_before_allocs = 1;
3018 has_per_allocation_action = TRUE;
3019 } else if (g_str_has_prefix (opt, "collect-before-allocs=")) {
3020 char *arg = strchr (opt, '=') + 1;
3021 has_per_allocation_action = TRUE;
3022 collect_before_allocs = atoi (arg);
3023 } else if (!strcmp (opt, "verify-before-collections")) {
3024 whole_heap_check_before_collection = TRUE;
3025 } else if (!strcmp (opt, "check-at-minor-collections")) {
3026 consistency_check_at_minor_collection = TRUE;
3027 nursery_clear_policy = CLEAR_AT_GC;
3028 } else if (!strcmp (opt, "mod-union-consistency-check")) {
3029 if (!major_collector.is_concurrent) {
3030 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`mod-union-consistency-check` only works with concurrent major collector.");
3033 mod_union_consistency_check = TRUE;
3034 } else if (!strcmp (opt, "check-mark-bits")) {
3035 check_mark_bits_after_major_collection = TRUE;
3036 } else if (!strcmp (opt, "check-nursery-pinned")) {
3037 check_nursery_objects_pinned = TRUE;
3038 } else if (!strcmp (opt, "clear-at-gc")) {
3039 nursery_clear_policy = CLEAR_AT_GC;
3040 } else if (!strcmp (opt, "clear-nursery-at-gc")) {
3041 nursery_clear_policy = CLEAR_AT_GC;
3042 } else if (!strcmp (opt, "clear-at-tlab-creation")) {
3043 nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
3044 } else if (!strcmp (opt, "debug-clear-at-tlab-creation")) {
3045 nursery_clear_policy = CLEAR_AT_TLAB_CREATION_DEBUG;
3046 } else if (!strcmp (opt, "check-scan-starts")) {
3047 do_scan_starts_check = TRUE;
3048 } else if (!strcmp (opt, "verify-nursery-at-minor-gc")) {
3049 do_verify_nursery = TRUE;
3050 } else if (!strcmp (opt, "check-concurrent")) {
3051 if (!major_collector.is_concurrent) {
3052 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`check-concurrent` only works with concurrent major collectors.");
3055 nursery_clear_policy = CLEAR_AT_GC;
3056 do_concurrent_checks = TRUE;
3057 } else if (!strcmp (opt, "dump-nursery-at-minor-gc")) {
3058 do_dump_nursery_content = TRUE;
3059 } else if (!strcmp (opt, "disable-minor")) {
3060 disable_minor_collections = TRUE;
3061 } else if (!strcmp (opt, "disable-major")) {
3062 disable_major_collections = TRUE;
3063 } else if (g_str_has_prefix (opt, "heap-dump=")) {
3064 char *filename = strchr (opt, '=') + 1;
3065 nursery_clear_policy = CLEAR_AT_GC;
3066 sgen_debug_enable_heap_dump (filename);
3067 } else if (g_str_has_prefix (opt, "binary-protocol=")) {
3068 char *filename = strchr (opt, '=') + 1;
3069 char *colon = strrchr (filename, ':');
3072 if (!mono_gc_parse_environment_string_extract_number (colon + 1, &limit)) {
3073 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring limit.", "Binary protocol file size limit must be an integer.");
3078 binary_protocol_init (filename, (long long)limit);
3079 } else if (!strcmp (opt, "nursery-canaries")) {
3080 do_verify_nursery = TRUE;
3081 enable_nursery_canaries = TRUE;
3082 } else if (!sgen_client_handle_gc_debug (opt)) {
3083 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "Unknown option `%s`.", opt);
3088 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);
3089 fprintf (stderr, "Valid <option>s are:\n");
3090 fprintf (stderr, " collect-before-allocs[=<n>]\n");
3091 fprintf (stderr, " verify-before-allocs[=<n>]\n");
3092 fprintf (stderr, " check-at-minor-collections\n");
3093 fprintf (stderr, " check-mark-bits\n");
3094 fprintf (stderr, " check-nursery-pinned\n");
3095 fprintf (stderr, " verify-before-collections\n");
3096 fprintf (stderr, " verify-nursery-at-minor-gc\n");
3097 fprintf (stderr, " dump-nursery-at-minor-gc\n");
3098 fprintf (stderr, " disable-minor\n");
3099 fprintf (stderr, " disable-major\n");
3100 fprintf (stderr, " check-concurrent\n");
3101 fprintf (stderr, " clear-[nursery-]at-gc\n");
3102 fprintf (stderr, " clear-at-tlab-creation\n");
3103 fprintf (stderr, " debug-clear-at-tlab-creation\n");
3104 fprintf (stderr, " check-scan-starts\n");
3105 fprintf (stderr, " print-allowance\n");
3106 fprintf (stderr, " print-pinning\n");
3107 fprintf (stderr, " heap-dump=<filename>\n");
3108 fprintf (stderr, " binary-protocol=<filename>[:<file-size-limit>]\n");
3109 fprintf (stderr, " nursery-canaries\n");
3110 sgen_client_print_gc_debug_usage ();
3111 fprintf (stderr, "\n");
3113 usage_printed = TRUE;
3119 if (check_mark_bits_after_major_collection)
3120 nursery_clear_policy = CLEAR_AT_GC;
3122 if (major_collector.post_param_init)
3123 major_collector.post_param_init (&major_collector);
3125 if (major_collector.needs_thread_pool)
3126 sgen_workers_init (1);
3128 sgen_memgov_init (max_heap, soft_limit, debug_print_allowance, allowance_ratio, save_target);
3130 memset (&remset, 0, sizeof (remset));
3132 sgen_card_table_init (&remset);
3134 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");
3140 sgen_get_nursery_clear_policy (void)
3142 return nursery_clear_policy;
3148 mono_coop_mutex_lock (&gc_mutex);
3152 sgen_gc_unlock (void)
3154 gboolean try_free = sgen_try_free_some_memory;
3155 sgen_try_free_some_memory = FALSE;
3156 mono_coop_mutex_unlock (&gc_mutex);
3158 mono_thread_hazardous_try_free_some ();
3162 sgen_major_collector_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
3164 major_collector.iterate_live_block_ranges (callback);
3168 sgen_get_major_collector (void)
3170 return &major_collector;
3174 sgen_get_remset (void)
3180 count_cards (long long *major_total, long long *major_marked, long long *los_total, long long *los_marked)
3182 sgen_get_major_collector ()->count_cards (major_total, major_marked);
3183 sgen_los_count_cards (los_total, los_marked);
3186 static gboolean world_is_stopped = FALSE;
3188 /* LOCKING: assumes the GC lock is held */
3190 sgen_stop_world (int generation)
3192 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3194 SGEN_ASSERT (0, !world_is_stopped, "Why are we stopping a stopped world?");
3196 binary_protocol_world_stopping (generation, sgen_timestamp (), (gpointer) (gsize) mono_native_thread_id_get ());
3198 sgen_client_stop_world (generation);
3200 world_is_stopped = TRUE;
3202 if (binary_protocol_is_heavy_enabled ())
3203 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3204 binary_protocol_world_stopped (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3207 /* LOCKING: assumes the GC lock is held */
3209 sgen_restart_world (int generation, GGTimingInfo *timing)
3211 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3213 SGEN_ASSERT (0, world_is_stopped, "Why are we restarting a running world?");
3215 if (binary_protocol_is_heavy_enabled ())
3216 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3217 binary_protocol_world_restarting (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3219 sgen_client_restart_world (generation, timing);
3221 world_is_stopped = FALSE;
3223 binary_protocol_world_restarted (generation, sgen_timestamp ());
3225 sgen_try_free_some_memory = TRUE;
3227 if (sgen_client_bridge_need_processing ())
3228 sgen_client_bridge_processing_finish (generation);
3230 sgen_memgov_collection_end (generation, timing, timing ? 2 : 0);
3234 sgen_is_world_stopped (void)
3236 return world_is_stopped;
3240 sgen_check_whole_heap_stw (void)
3242 sgen_stop_world (0);
3243 sgen_clear_nursery_fragments ();
3244 sgen_check_whole_heap (FALSE);
3245 sgen_restart_world (0, NULL);
3249 sgen_timestamp (void)
3251 SGEN_TV_DECLARE (timestamp);
3252 SGEN_TV_GETTIME (timestamp);
3253 return SGEN_TV_ELAPSED (sgen_init_timestamp, timestamp);
3256 #endif /* HAVE_SGEN_GC */