3 * Simple generational GC.
6 * Paolo Molaro (lupus@ximian.com)
7 * Rodrigo Kumpera (kumpera@gmail.com)
9 * Copyright 2005-2011 Novell, Inc (http://www.novell.com)
10 * Copyright 2011 Xamarin Inc (http://www.xamarin.com)
12 * Thread start/stop adapted from Boehm's GC:
13 * Copyright (c) 1994 by Xerox Corporation. All rights reserved.
14 * Copyright (c) 1996 by Silicon Graphics. All rights reserved.
15 * Copyright (c) 1998 by Fergus Henderson. All rights reserved.
16 * Copyright (c) 2000-2004 by Hewlett-Packard Company. All rights reserved.
17 * Copyright 2001-2003 Ximian, Inc
18 * Copyright 2003-2010 Novell, Inc.
19 * Copyright 2011 Xamarin, Inc.
20 * Copyright (C) 2012 Xamarin Inc
22 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
24 * Important: allocation provides always zeroed memory, having to do
25 * a memset after allocation is deadly for performance.
26 * Memory usage at startup is currently as follows:
28 * 64 KB internal space
30 * We should provide a small memory config with half the sizes
32 * We currently try to make as few mono assumptions as possible:
33 * 1) 2-word header with no GC pointers in it (first vtable, second to store the
35 * 2) gc descriptor is the second word in the vtable (first word in the class)
36 * 3) 8 byte alignment is the minimum and enough (not true for special structures (SIMD), FIXME)
37 * 4) there is a function to get an object's size and the number of
38 * elements in an array.
39 * 5) we know the special way bounds are allocated for complex arrays
40 * 6) we know about proxies and how to treat them when domains are unloaded
42 * Always try to keep stack usage to a minimum: no recursive behaviour
43 * and no large stack allocs.
45 * General description.
46 * Objects are initially allocated in a nursery using a fast bump-pointer technique.
47 * When the nursery is full we start a nursery collection: this is performed with a
49 * When the old generation is full we start a copying GC of the old generation as well:
50 * this will be changed to mark&sweep with copying when fragmentation becomes to severe
51 * in the future. Maybe we'll even do both during the same collection like IMMIX.
53 * The things that complicate this description are:
54 * *) pinned objects: we can't move them so we need to keep track of them
55 * *) no precise info of the thread stacks and registers: we need to be able to
56 * quickly find the objects that may be referenced conservatively and pin them
57 * (this makes the first issues more important)
58 * *) large objects are too expensive to be dealt with using copying GC: we handle them
59 * with mark/sweep during major collections
60 * *) some objects need to not move even if they are small (interned strings, Type handles):
61 * we use mark/sweep for them, too: they are not allocated in the nursery, but inside
62 * PinnedChunks regions
68 *) we could have a function pointer in MonoClass to implement
69 customized write barriers for value types
71 *) investigate the stuff needed to advance a thread to a GC-safe
72 point (single-stepping, read from unmapped memory etc) and implement it.
73 This would enable us to inline allocations and write barriers, for example,
74 or at least parts of them, like the write barrier checks.
75 We may need this also for handling precise info on stacks, even simple things
76 as having uninitialized data on the stack and having to wait for the prolog
77 to zero it. Not an issue for the last frame that we scan conservatively.
78 We could always not trust the value in the slots anyway.
80 *) modify the jit to save info about references in stack locations:
81 this can be done just for locals as a start, so that at least
82 part of the stack is handled precisely.
84 *) test/fix endianess issues
86 *) Implement a card table as the write barrier instead of remembered
87 sets? Card tables are not easy to implement with our current
88 memory layout. We have several different kinds of major heap
89 objects: Small objects in regular blocks, small objects in pinned
90 chunks and LOS objects. If we just have a pointer we have no way
91 to tell which kind of object it points into, therefore we cannot
92 know where its card table is. The least we have to do to make
93 this happen is to get rid of write barriers for indirect stores.
96 *) Get rid of write barriers for indirect stores. We can do this by
97 telling the GC to wbarrier-register an object once we do an ldloca
98 or ldelema on it, and to unregister it once it's not used anymore
99 (it can only travel downwards on the stack). The problem with
100 unregistering is that it needs to happen eventually no matter
101 what, even if exceptions are thrown, the thread aborts, etc.
102 Rodrigo suggested that we could do only the registering part and
103 let the collector find out (pessimistically) when it's safe to
104 unregister, namely when the stack pointer of the thread that
105 registered the object is higher than it was when the registering
106 happened. This might make for a good first implementation to get
107 some data on performance.
109 *) Some sort of blacklist support? Blacklists is a concept from the
110 Boehm GC: if during a conservative scan we find pointers to an
111 area which we might use as heap, we mark that area as unusable, so
112 pointer retention by random pinning pointers is reduced.
114 *) experiment with max small object size (very small right now - 2kb,
115 because it's tied to the max freelist size)
117 *) add an option to mmap the whole heap in one chunk: it makes for many
118 simplifications in the checks (put the nursery at the top and just use a single
119 check for inclusion/exclusion): the issue this has is that on 32 bit systems it's
120 not flexible (too much of the address space may be used by default or we can't
121 increase the heap as needed) and we'd need a race-free mechanism to return memory
122 back to the system (mprotect(PROT_NONE) will still keep the memory allocated if it
123 was written to, munmap is needed, but the following mmap may not find the same segment
126 *) memzero the major fragments after restarting the world and optionally a smaller
129 *) investigate having fragment zeroing threads
131 *) separate locks for finalization and other minor stuff to reduce
134 *) try a different copying order to improve memory locality
136 *) a thread abort after a store but before the write barrier will
137 prevent the write barrier from executing
139 *) specialized dynamically generated markers/copiers
141 *) Dynamically adjust TLAB size to the number of threads. If we have
142 too many threads that do allocation, we might need smaller TLABs,
143 and we might get better performance with larger TLABs if we only
144 have a handful of threads. We could sum up the space left in all
145 assigned TLABs and if that's more than some percentage of the
146 nursery size, reduce the TLAB size.
148 *) Explore placing unreachable objects on unused nursery memory.
149 Instead of memset'ng a region to zero, place an int[] covering it.
150 A good place to start is add_nursery_frag. The tricky thing here is
151 placing those objects atomically outside of a collection.
153 *) Allocation should use asymmetric Dekker synchronization:
154 http://blogs.oracle.com/dave/resource/Asymmetric-Dekker-Synchronization.txt
155 This should help weak consistency archs.
162 #define _XOPEN_SOURCE
163 #define _DARWIN_C_SOURCE
169 #ifdef HAVE_PTHREAD_H
172 #ifdef HAVE_PTHREAD_NP_H
173 #include <pthread_np.h>
182 #include "mono/sgen/sgen-gc.h"
183 #include "mono/sgen/sgen-cardtable.h"
184 #include "mono/sgen/sgen-protocol.h"
185 #include "mono/sgen/sgen-memory-governor.h"
186 #include "mono/sgen/sgen-hash-table.h"
187 #include "mono/sgen/sgen-pinning.h"
188 #include "mono/sgen/sgen-workers.h"
189 #include "mono/sgen/sgen-client.h"
190 #include "mono/sgen/sgen-pointer-queue.h"
191 #include "mono/sgen/gc-internal-agnostic.h"
192 #include "mono/utils/mono-proclib.h"
193 #include "mono/utils/mono-memory-model.h"
194 #include "mono/utils/hazard-pointer.h"
196 #include <mono/utils/memcheck.h>
197 #include <mono/utils/mono-mmap-internals.h>
199 #undef pthread_create
201 #undef pthread_detach
204 * ######################################################################
205 * ######## Types and constants used by the GC.
206 * ######################################################################
209 /* 0 means not initialized, 1 is initialized, -1 means in progress */
210 static int gc_initialized = 0;
211 /* If set, check if we need to do something every X allocations */
212 gboolean has_per_allocation_action;
213 /* If set, do a heap check every X allocation */
214 guint32 verify_before_allocs = 0;
215 /* If set, do a minor collection before every X allocation */
216 guint32 collect_before_allocs = 0;
217 /* If set, do a whole heap check before each collection */
218 static gboolean whole_heap_check_before_collection = FALSE;
219 /* If set, do a remset consistency check at various opportunities */
220 static gboolean remset_consistency_checks = FALSE;
221 /* If set, do a mod union consistency check before each finishing collection pause */
222 static gboolean mod_union_consistency_check = FALSE;
223 /* If set, check whether mark bits are consistent after major collections */
224 static gboolean check_mark_bits_after_major_collection = FALSE;
225 /* If set, check that all nursery objects are pinned/not pinned, depending on context */
226 static gboolean check_nursery_objects_pinned = FALSE;
227 /* If set, do a few checks when the concurrent collector is used */
228 static gboolean do_concurrent_checks = FALSE;
229 /* If set, do a plausibility check on the scan_starts before and after
231 static gboolean do_scan_starts_check = FALSE;
233 static gboolean disable_minor_collections = FALSE;
234 static gboolean disable_major_collections = FALSE;
235 static gboolean do_verify_nursery = FALSE;
236 static gboolean do_dump_nursery_content = FALSE;
237 static gboolean enable_nursery_canaries = FALSE;
239 static gboolean precleaning_enabled = TRUE;
240 static gboolean dynamic_nursery = FALSE;
241 static size_t min_nursery_size = 0;
242 static size_t max_nursery_size = 0;
244 #ifdef HEAVY_STATISTICS
245 guint64 stat_objects_alloced_degraded = 0;
246 guint64 stat_bytes_alloced_degraded = 0;
248 guint64 stat_copy_object_called_nursery = 0;
249 guint64 stat_objects_copied_nursery = 0;
250 guint64 stat_copy_object_called_major = 0;
251 guint64 stat_objects_copied_major = 0;
253 guint64 stat_scan_object_called_nursery = 0;
254 guint64 stat_scan_object_called_major = 0;
256 guint64 stat_slots_allocated_in_vain;
258 guint64 stat_nursery_copy_object_failed_from_space = 0;
259 guint64 stat_nursery_copy_object_failed_forwarded = 0;
260 guint64 stat_nursery_copy_object_failed_pinned = 0;
261 guint64 stat_nursery_copy_object_failed_to_space = 0;
263 static guint64 stat_wbarrier_add_to_global_remset = 0;
264 static guint64 stat_wbarrier_arrayref_copy = 0;
265 static guint64 stat_wbarrier_generic_store = 0;
266 static guint64 stat_wbarrier_generic_store_atomic = 0;
267 static guint64 stat_wbarrier_set_root = 0;
270 static guint64 stat_pinned_objects = 0;
272 static guint64 time_minor_pre_collection_fragment_clear = 0;
273 static guint64 time_minor_pinning = 0;
274 static guint64 time_minor_scan_remsets = 0;
275 static guint64 time_minor_scan_major_blocks = 0;
276 static guint64 time_minor_scan_los = 0;
277 static guint64 time_minor_scan_pinned = 0;
278 static guint64 time_minor_scan_roots = 0;
279 static guint64 time_minor_finish_gray_stack = 0;
280 static guint64 time_minor_fragment_creation = 0;
282 static guint64 time_major_pre_collection_fragment_clear = 0;
283 static guint64 time_major_pinning = 0;
284 static guint64 time_major_scan_pinned = 0;
285 static guint64 time_major_scan_roots = 0;
286 static guint64 time_major_scan_mod_union_blocks = 0;
287 static guint64 time_major_scan_mod_union_los = 0;
288 static guint64 time_major_finish_gray_stack = 0;
289 static guint64 time_major_free_bigobjs = 0;
290 static guint64 time_major_los_sweep = 0;
291 static guint64 time_major_sweep = 0;
292 static guint64 time_major_fragment_creation = 0;
294 static guint64 time_max = 0;
296 static int sgen_max_pause_time = SGEN_DEFAULT_MAX_PAUSE_TIME;
297 static float sgen_max_pause_margin = SGEN_DEFAULT_MAX_PAUSE_MARGIN;
299 static SGEN_TV_DECLARE (time_major_conc_collection_start);
300 static SGEN_TV_DECLARE (time_major_conc_collection_end);
302 int gc_debug_level = 0;
304 static char* gc_params_options;
305 static char* gc_debug_options;
309 mono_gc_flush_info (void)
311 fflush (gc_debug_file);
315 #define TV_DECLARE SGEN_TV_DECLARE
316 #define TV_GETTIME SGEN_TV_GETTIME
317 #define TV_ELAPSED SGEN_TV_ELAPSED
319 static SGEN_TV_DECLARE (sgen_init_timestamp);
321 NurseryClearPolicy nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
323 #define object_is_forwarded SGEN_OBJECT_IS_FORWARDED
324 #define object_is_pinned SGEN_OBJECT_IS_PINNED
325 #define pin_object SGEN_PIN_OBJECT
327 #define ptr_in_nursery sgen_ptr_in_nursery
329 #define LOAD_VTABLE SGEN_LOAD_VTABLE
332 nursery_canaries_enabled (void)
334 return enable_nursery_canaries;
337 #define safe_object_get_size sgen_safe_object_get_size
339 #if defined(HAVE_CONC_GC_AS_DEFAULT)
340 /* Use concurrent major on deskstop platforms */
341 #define DEFAULT_MAJOR SGEN_MAJOR_CONCURRENT
343 #define DEFAULT_MAJOR SGEN_MAJOR_SERIAL
349 SGEN_MAJOR_CONCURRENT,
350 SGEN_MAJOR_CONCURRENT_PARALLEL
356 SGEN_MINOR_SIMPLE_PARALLEL,
363 SGEN_MODE_THROUGHPUT,
368 * ######################################################################
369 * ######## Global data.
370 * ######################################################################
372 MonoCoopMutex gc_mutex;
374 #define SCAN_START_SIZE SGEN_SCAN_START_SIZE
376 size_t degraded_mode = 0;
378 static mword bytes_pinned_from_failed_allocation = 0;
380 GCMemSection *nursery_section = NULL;
381 static volatile mword lowest_heap_address = ~(mword)0;
382 static volatile mword highest_heap_address = 0;
384 MonoCoopMutex sgen_interruption_mutex;
386 int current_collection_generation = -1;
387 volatile gboolean concurrent_collection_in_progress = FALSE;
389 /* objects that are ready to be finalized */
390 static SgenPointerQueue fin_ready_queue = SGEN_POINTER_QUEUE_INIT (INTERNAL_MEM_FINALIZE_READY);
391 static SgenPointerQueue critical_fin_queue = SGEN_POINTER_QUEUE_INIT (INTERNAL_MEM_FINALIZE_READY);
393 /* registered roots: the key to the hash is the root start address */
395 * Different kinds of roots are kept separate to speed up pin_from_roots () for example.
397 SgenHashTable roots_hash [ROOT_TYPE_NUM] = {
398 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL),
399 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL),
400 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL)
402 static mword roots_size = 0; /* amount of memory in the root set */
404 /* The size of a TLAB */
405 /* The bigger the value, the less often we have to go to the slow path to allocate a new
406 * one, but the more space is wasted by threads not allocating much memory.
408 * FIXME: Make this self-tuning for each thread.
410 guint32 tlab_size = (1024 * 4);
412 #define MAX_SMALL_OBJ_SIZE SGEN_MAX_SMALL_OBJ_SIZE
414 #define ALLOC_ALIGN SGEN_ALLOC_ALIGN
416 #define ALIGN_UP SGEN_ALIGN_UP
418 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
419 MonoNativeThreadId main_gc_thread = NULL;
422 /*Object was pinned during the current collection*/
423 static mword objects_pinned;
426 * ######################################################################
427 * ######## Macros and function declarations.
428 * ######################################################################
431 /* forward declarations */
432 static void scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx);
434 static void pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx);
435 static void finish_gray_stack (int generation, ScanCopyContext ctx);
438 SgenMajorCollector major_collector;
439 SgenMinorCollector sgen_minor_collector;
441 static SgenRememberedSet remset;
444 * The gray queue a worker job must use. If we're not parallel or
445 * concurrent, we use the main gray queue.
447 static SgenGrayQueue*
448 sgen_workers_get_job_gray_queue (WorkerData *worker_data, SgenGrayQueue *default_gray_queue)
451 return &worker_data->private_gray_queue;
452 SGEN_ASSERT (0, default_gray_queue, "Why don't we have a default gray queue when we're not running in a worker thread?");
453 return default_gray_queue;
457 gray_queue_redirect (SgenGrayQueue *queue)
459 sgen_workers_take_from_queue (current_collection_generation, queue);
463 sgen_scan_area_with_callback (char *start, char *end, IterateObjectCallbackFunc callback, void *data, gboolean allow_flags, gboolean fail_on_canaries)
465 while (start < end) {
469 if (!*(void**)start) {
470 start += sizeof (void*); /* should be ALLOC_ALIGN, really */
475 if (!(obj = (char *)SGEN_OBJECT_IS_FORWARDED (start)))
481 if (!sgen_client_object_is_array_fill ((GCObject*)obj)) {
482 CHECK_CANARY_FOR_OBJECT ((GCObject*)obj, fail_on_canaries);
483 size = ALIGN_UP (safe_object_get_size ((GCObject*)obj));
484 callback ((GCObject*)obj, size, data);
485 CANARIFY_SIZE (size);
487 size = ALIGN_UP (safe_object_get_size ((GCObject*)obj));
495 * sgen_add_to_global_remset:
497 * The global remset contains locations which point into newspace after
498 * a minor collection. This can happen if the objects they point to are pinned.
500 * LOCKING: If called from a parallel collector, the global remset
501 * lock must be held. For serial collectors that is not necessary.
504 sgen_add_to_global_remset (gpointer ptr, GCObject *obj)
506 SGEN_ASSERT (5, sgen_ptr_in_nursery (obj), "Target pointer of global remset must be in the nursery");
508 HEAVY_STAT (++stat_wbarrier_add_to_global_remset);
510 if (!major_collector.is_concurrent) {
511 SGEN_ASSERT (5, current_collection_generation != -1, "Global remsets can only be added during collections");
513 if (current_collection_generation == -1)
514 SGEN_ASSERT (5, sgen_concurrent_collection_in_progress (), "Global remsets outside of collection pauses can only be added by the concurrent collector");
517 if (!object_is_pinned (obj))
518 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");
519 else if (sgen_cement_lookup_or_register (obj))
522 remset.record_pointer (ptr);
524 sgen_pin_stats_register_global_remset (obj);
526 SGEN_LOG (8, "Adding global remset for %p", ptr);
527 binary_protocol_global_remset (ptr, obj, (gpointer)SGEN_LOAD_VTABLE (obj));
531 * sgen_drain_gray_stack:
533 * Scan objects in the gray stack until the stack is empty. This should be called
534 * frequently after each object is copied, to achieve better locality and cache
539 sgen_drain_gray_stack (ScanCopyContext ctx)
541 SGEN_ASSERT (0, ctx.ops->drain_gray_stack, "Why do we have a scan/copy context with a missing drain gray stack function?");
543 return ctx.ops->drain_gray_stack (ctx.queue);
547 * Addresses in the pin queue are already sorted. This function finds
548 * the object header for each address and pins the object. The
549 * addresses must be inside the nursery section. The (start of the)
550 * address array is overwritten with the addresses of the actually
551 * pinned objects. Return the number of pinned objects.
554 pin_objects_from_nursery_pin_queue (gboolean do_scan_objects, ScanCopyContext ctx)
556 GCMemSection *section = nursery_section;
557 void **start = sgen_pinning_get_entry (section->pin_queue_first_entry);
558 void **end = sgen_pinning_get_entry (section->pin_queue_last_entry);
559 void *start_nursery = section->data;
560 void *end_nursery = section->end_data;
565 void *pinning_front = start_nursery;
567 void **definitely_pinned = start;
568 ScanObjectFunc scan_func = ctx.ops->scan_object;
569 SgenGrayQueue *queue = ctx.queue;
571 sgen_nursery_allocator_prepare_for_pinning ();
573 while (start < end) {
574 GCObject *obj_to_pin = NULL;
575 size_t obj_to_pin_size = 0;
580 SGEN_ASSERT (0, addr >= start_nursery && addr < end_nursery, "Potential pinning address out of range");
581 SGEN_ASSERT (0, addr >= last, "Pin queue not sorted");
588 SGEN_LOG (5, "Considering pinning addr %p", addr);
589 /* We've already processed everything up to pinning_front. */
590 if (addr < pinning_front) {
596 * Find the closest scan start <= addr. We might search backward in the
597 * scan_starts array because entries might be NULL. In the worst case we
598 * start at start_nursery.
600 idx = ((char*)addr - (char*)section->data) / SCAN_START_SIZE;
601 SGEN_ASSERT (0, idx < section->num_scan_start, "Scan start index out of range");
602 search_start = (void*)section->scan_starts [idx];
603 if (!search_start || search_start > addr) {
606 search_start = section->scan_starts [idx];
607 if (search_start && search_start <= addr)
610 if (!search_start || search_start > addr)
611 search_start = start_nursery;
615 * If the pinning front is closer than the scan start we found, start
616 * searching at the front.
618 if (search_start < pinning_front)
619 search_start = pinning_front;
622 * Now addr should be in an object a short distance from search_start.
624 * search_start must point to zeroed mem or point to an object.
627 size_t obj_size, canarified_obj_size;
630 if (!*(void**)search_start) {
631 search_start = (void*)ALIGN_UP ((mword)search_start + sizeof (gpointer));
632 /* The loop condition makes sure we don't overrun addr. */
636 canarified_obj_size = obj_size = ALIGN_UP (safe_object_get_size ((GCObject*)search_start));
639 * Filler arrays are marked by an invalid sync word. We don't
640 * consider them for pinning. They are not delimited by canaries,
643 if (!sgen_client_object_is_array_fill ((GCObject*)search_start)) {
644 CHECK_CANARY_FOR_OBJECT (search_start, TRUE);
645 CANARIFY_SIZE (canarified_obj_size);
647 if (addr >= search_start && (char*)addr < (char*)search_start + obj_size) {
648 /* This is the object we're looking for. */
649 obj_to_pin = (GCObject*)search_start;
650 obj_to_pin_size = canarified_obj_size;
655 /* Skip to the next object */
656 search_start = (void*)((char*)search_start + canarified_obj_size);
657 } while (search_start <= addr);
659 /* We've searched past the address we were looking for. */
661 pinning_front = search_start;
662 goto next_pin_queue_entry;
666 * We've found an object to pin. It might still be a dummy array, but we
667 * can advance the pinning front in any case.
669 pinning_front = (char*)obj_to_pin + obj_to_pin_size;
672 * If this is a dummy array marking the beginning of a nursery
673 * fragment, we don't pin it.
675 if (sgen_client_object_is_array_fill (obj_to_pin))
676 goto next_pin_queue_entry;
679 * Finally - pin the object!
681 desc = sgen_obj_get_descriptor_safe (obj_to_pin);
682 if (do_scan_objects) {
683 scan_func (obj_to_pin, desc, queue);
685 SGEN_LOG (4, "Pinned object %p, vtable %p (%s), count %d\n",
686 obj_to_pin, *(void**)obj_to_pin, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj_to_pin)), count);
687 binary_protocol_pin (obj_to_pin,
688 (gpointer)LOAD_VTABLE (obj_to_pin),
689 safe_object_get_size (obj_to_pin));
691 pin_object (obj_to_pin);
692 GRAY_OBJECT_ENQUEUE_SERIAL (queue, obj_to_pin, desc);
693 sgen_pin_stats_register_object (obj_to_pin, GENERATION_NURSERY);
694 definitely_pinned [count] = obj_to_pin;
697 if (concurrent_collection_in_progress)
698 sgen_pinning_register_pinned_in_nursery (obj_to_pin);
700 next_pin_queue_entry:
704 sgen_client_nursery_objects_pinned (definitely_pinned, count);
705 stat_pinned_objects += count;
710 pin_objects_in_nursery (gboolean do_scan_objects, ScanCopyContext ctx)
714 if (nursery_section->pin_queue_first_entry == nursery_section->pin_queue_last_entry)
717 reduced_to = pin_objects_from_nursery_pin_queue (do_scan_objects, ctx);
718 nursery_section->pin_queue_last_entry = nursery_section->pin_queue_first_entry + reduced_to;
722 * This function is only ever called (via `collector_pin_object()` in `sgen-copy-object.h`)
723 * when we can't promote an object because we're out of memory.
726 sgen_pin_object (GCObject *object, SgenGrayQueue *queue)
728 SGEN_ASSERT (0, sgen_ptr_in_nursery (object), "We're only supposed to use this for pinning nursery objects when out of memory.");
731 * All pinned objects are assumed to have been staged, so we need to stage as well.
732 * Also, the count of staged objects shows that "late pinning" happened.
734 sgen_pin_stage_ptr (object);
736 SGEN_PIN_OBJECT (object);
737 binary_protocol_pin (object, (gpointer)LOAD_VTABLE (object), safe_object_get_size (object));
740 sgen_pin_stats_register_object (object, GENERATION_NURSERY);
742 GRAY_OBJECT_ENQUEUE_SERIAL (queue, object, sgen_obj_get_descriptor_safe (object));
745 /* Sort the addresses in array in increasing order.
746 * Done using a by-the book heap sort. Which has decent and stable performance, is pretty cache efficient.
749 sgen_sort_addresses (void **array, size_t size)
754 for (i = 1; i < size; ++i) {
757 size_t parent = (child - 1) / 2;
759 if (array [parent] >= array [child])
762 tmp = array [parent];
763 array [parent] = array [child];
770 for (i = size - 1; i > 0; --i) {
773 array [i] = array [0];
779 while (root * 2 + 1 <= end) {
780 size_t child = root * 2 + 1;
782 if (child < end && array [child] < array [child + 1])
784 if (array [root] >= array [child])
788 array [root] = array [child];
797 * Scan the memory between start and end and queue values which could be pointers
798 * to the area between start_nursery and end_nursery for later consideration.
799 * Typically used for thread stacks.
802 sgen_conservatively_pin_objects_from (void **start, void **end, void *start_nursery, void *end_nursery, int pin_type)
806 SGEN_ASSERT (0, ((mword)start & (SIZEOF_VOID_P - 1)) == 0, "Why are we scanning for references in unaligned memory ?");
808 #if defined(VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE) && !defined(_WIN64)
809 VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE (start, (char*)end - (char*)start);
812 while (start < end) {
814 * *start can point to the middle of an object
815 * note: should we handle pointing at the end of an object?
816 * pinning in C# code disallows pointing at the end of an object
817 * but there is some small chance that an optimizing C compiler
818 * may keep the only reference to an object by pointing
819 * at the end of it. We ignore this small chance for now.
820 * Pointers to the end of an object are indistinguishable
821 * from pointers to the start of the next object in memory
822 * so if we allow that we'd need to pin two objects...
823 * We queue the pointer in an array, the
824 * array will then be sorted and uniqued. This way
825 * we can coalesce several pinning pointers and it should
826 * be faster since we'd do a memory scan with increasing
827 * addresses. Note: we can align the address to the allocation
828 * alignment, so the unique process is more effective.
830 mword addr = (mword)*start;
831 addr &= ~(ALLOC_ALIGN - 1);
832 if (addr >= (mword)start_nursery && addr < (mword)end_nursery) {
833 SGEN_LOG (6, "Pinning address %p from %p", (void*)addr, start);
834 sgen_pin_stage_ptr ((void*)addr);
835 binary_protocol_pin_stage (start, (void*)addr);
836 sgen_pin_stats_register_address ((char*)addr, pin_type);
842 SGEN_LOG (7, "found %d potential pinned heap pointers", count);
846 * The first thing we do in a collection is to identify pinned objects.
847 * This function considers all the areas of memory that need to be
848 * conservatively scanned.
851 pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx)
855 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);
856 /* objects pinned from the API are inside these roots */
857 SGEN_HASH_TABLE_FOREACH (&roots_hash [ROOT_TYPE_PINNED], void **, start_root, RootRecord *, root) {
858 SGEN_LOG (6, "Pinned roots %p-%p", start_root, root->end_root);
859 sgen_conservatively_pin_objects_from (start_root, (void**)root->end_root, start_nursery, end_nursery, PIN_TYPE_OTHER);
860 } SGEN_HASH_TABLE_FOREACH_END;
861 /* now deal with the thread stacks
862 * in the future we should be able to conservatively scan only:
863 * *) the cpu registers
864 * *) the unmanaged stack frames
865 * *) the _last_ managed stack frame
866 * *) pointers slots in managed frames
868 sgen_client_scan_thread_data (start_nursery, end_nursery, FALSE, ctx);
872 single_arg_user_copy_or_mark (GCObject **obj, void *gc_data)
874 ScanCopyContext *ctx = (ScanCopyContext *)gc_data;
875 ctx->ops->copy_or_mark_object (obj, ctx->queue);
879 * The memory area from start_root to end_root contains pointers to objects.
880 * Their position is precisely described by @desc (this means that the pointer
881 * can be either NULL or the pointer to the start of an object).
882 * This functions copies them to to_space updates them.
884 * This function is not thread-safe!
887 precisely_scan_objects_from (void** start_root, void** end_root, char* n_start, char *n_end, SgenDescriptor desc, ScanCopyContext ctx)
889 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
890 ScanPtrFieldFunc scan_field_func = ctx.ops->scan_ptr_field;
891 SgenGrayQueue *queue = ctx.queue;
893 switch (desc & ROOT_DESC_TYPE_MASK) {
894 case ROOT_DESC_BITMAP:
895 desc >>= ROOT_DESC_TYPE_SHIFT;
897 if ((desc & 1) && *start_root) {
898 copy_func ((GCObject**)start_root, queue);
899 SGEN_LOG (9, "Overwrote root at %p with %p", start_root, *start_root);
905 case ROOT_DESC_COMPLEX: {
906 gsize *bitmap_data = (gsize *)sgen_get_complex_descriptor_bitmap (desc);
907 gsize bwords = (*bitmap_data) - 1;
908 void **start_run = start_root;
910 while (bwords-- > 0) {
911 gsize bmap = *bitmap_data++;
912 void **objptr = start_run;
914 if ((bmap & 1) && *objptr) {
915 copy_func ((GCObject**)objptr, queue);
916 SGEN_LOG (9, "Overwrote root at %p with %p", objptr, *objptr);
921 start_run += GC_BITS_PER_WORD;
925 case ROOT_DESC_VECTOR: {
928 for (p = start_root; p < end_root; p++) {
930 scan_field_func (NULL, (GCObject**)p, queue);
934 case ROOT_DESC_USER: {
935 SgenUserRootMarkFunc marker = sgen_get_user_descriptor_func (desc);
936 marker (start_root, single_arg_user_copy_or_mark, &ctx);
939 case ROOT_DESC_RUN_LEN:
940 g_assert_not_reached ();
942 g_assert_not_reached ();
947 reset_heap_boundaries (void)
949 lowest_heap_address = ~(mword)0;
950 highest_heap_address = 0;
954 sgen_update_heap_boundaries (mword low, mword high)
959 old = lowest_heap_address;
962 } while (SGEN_CAS_PTR ((gpointer*)&lowest_heap_address, (gpointer)low, (gpointer)old) != (gpointer)old);
965 old = highest_heap_address;
968 } while (SGEN_CAS_PTR ((gpointer*)&highest_heap_address, (gpointer)high, (gpointer)old) != (gpointer)old);
972 * Allocate and setup the data structures needed to be able to allocate objects
973 * in the nursery. The nursery is stored in nursery_section.
976 alloc_nursery (gboolean dynamic, size_t min_size, size_t max_size)
983 min_size = SGEN_DEFAULT_NURSERY_MIN_SIZE;
985 max_size = SGEN_DEFAULT_NURSERY_MAX_SIZE;
987 SGEN_ASSERT (0, min_size == max_size, "We can't have nursery ranges for static configuration.");
989 min_size = max_size = SGEN_DEFAULT_NURSERY_SIZE;
992 SGEN_ASSERT (0, !nursery_section, "Why are we allocating the nursery twice?");
993 SGEN_LOG (2, "Allocating nursery size: %zu, initial %zu", max_size, min_size);
995 /* FIXME: handle OOM */
996 nursery_section = (GCMemSection *)sgen_alloc_internal (INTERNAL_MEM_SECTION);
998 /* If there isn't enough space even for the nursery we should simply abort. */
999 g_assert (sgen_memgov_try_alloc_space (max_size, SPACE_NURSERY));
1002 * The nursery section range represents the memory section where objects
1003 * can be found. This is used when iterating for objects in the nursery,
1004 * pinning etc. sgen_nursery_max_size represents the total allocated space
1005 * for the nursery. sgen_nursery_size represents the current size of the
1006 * nursery and it is used for allocation limits, heuristics etc. The
1007 * nursery section is not always identical to the current nursery size
1008 * because it can contain pinned objects from when the nursery was larger.
1010 * sgen_nursery_size <= nursery_section size <= sgen_nursery_max_size
1012 data = (char *)major_collector.alloc_heap (max_size, max_size);
1013 sgen_update_heap_boundaries ((mword)data, (mword)(data + max_size));
1014 nursery_section->data = data;
1015 nursery_section->end_data = data + min_size;
1016 scan_starts = (max_size + SCAN_START_SIZE - 1) / SCAN_START_SIZE;
1017 nursery_section->scan_starts = (char **)sgen_alloc_internal_dynamic (sizeof (char*) * scan_starts, INTERNAL_MEM_SCAN_STARTS, TRUE);
1018 nursery_section->num_scan_start = scan_starts;
1020 sgen_nursery_allocator_set_nursery_bounds (data, min_size, max_size);
1024 mono_gc_get_logfile (void)
1026 return gc_debug_file;
1030 mono_gc_params_set (const char* options)
1032 if (gc_params_options)
1033 g_free (gc_params_options);
1035 gc_params_options = g_strdup (options);
1039 mono_gc_debug_set (const char* options)
1041 if (gc_debug_options)
1042 g_free (gc_debug_options);
1044 gc_debug_options = g_strdup (options);
1048 scan_finalizer_entries (SgenPointerQueue *fin_queue, ScanCopyContext ctx)
1050 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
1051 SgenGrayQueue *queue = ctx.queue;
1054 for (i = 0; i < fin_queue->next_slot; ++i) {
1055 GCObject *obj = (GCObject *)fin_queue->data [i];
1058 SGEN_LOG (5, "Scan of fin ready object: %p (%s)\n", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
1059 copy_func ((GCObject**)&fin_queue->data [i], queue);
1064 generation_name (int generation)
1066 switch (generation) {
1067 case GENERATION_NURSERY: return "nursery";
1068 case GENERATION_OLD: return "old";
1069 default: g_assert_not_reached ();
1074 sgen_generation_name (int generation)
1076 return generation_name (generation);
1080 finish_gray_stack (int generation, ScanCopyContext ctx)
1084 int done_with_ephemerons, ephemeron_rounds = 0;
1085 char *start_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_start () : NULL;
1086 char *end_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_end () : (char*)-1;
1087 SgenGrayQueue *queue = ctx.queue;
1089 binary_protocol_finish_gray_stack_start (sgen_timestamp (), generation);
1091 * We copied all the reachable objects. Now it's the time to copy
1092 * the objects that were not referenced by the roots, but by the copied objects.
1093 * we built a stack of objects pointed to by gray_start: they are
1094 * additional roots and we may add more items as we go.
1095 * We loop until gray_start == gray_objects which means no more objects have
1096 * been added. Note this is iterative: no recursion is involved.
1097 * We need to walk the LO list as well in search of marked big objects
1098 * (use a flag since this is needed only on major collections). We need to loop
1099 * here as well, so keep a counter of marked LO (increasing it in copy_object).
1100 * To achieve better cache locality and cache usage, we drain the gray stack
1101 * frequently, after each object is copied, and just finish the work here.
1103 sgen_drain_gray_stack (ctx);
1105 SGEN_LOG (2, "%s generation done", generation_name (generation));
1108 Reset bridge data, we might have lingering data from a previous collection if this is a major
1109 collection trigged by minor overflow.
1111 We must reset the gathered bridges since their original block might be evacuated due to major
1112 fragmentation in the meanwhile and the bridge code should not have to deal with that.
1114 if (sgen_client_bridge_need_processing ())
1115 sgen_client_bridge_reset_data ();
1118 * Mark all strong toggleref objects. This must be done before we walk ephemerons or finalizers
1119 * to ensure they see the full set of live objects.
1121 sgen_client_mark_togglerefs (start_addr, end_addr, ctx);
1124 * Walk the ephemeron tables marking all values with reachable keys. This must be completely done
1125 * before processing finalizable objects and non-tracking weak links to avoid finalizing/clearing
1126 * objects that are in fact reachable.
1128 done_with_ephemerons = 0;
1130 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1131 sgen_drain_gray_stack (ctx);
1133 } while (!done_with_ephemerons);
1135 if (sgen_client_bridge_need_processing ()) {
1136 /*Make sure the gray stack is empty before we process bridge objects so we get liveness right*/
1137 sgen_drain_gray_stack (ctx);
1138 sgen_collect_bridge_objects (generation, ctx);
1139 if (generation == GENERATION_OLD)
1140 sgen_collect_bridge_objects (GENERATION_NURSERY, ctx);
1143 Do the first bridge step here, as the collector liveness state will become useless after that.
1145 An important optimization is to only proccess the possibly dead part of the object graph and skip
1146 over all live objects as we transitively know everything they point must be alive too.
1148 The above invariant is completely wrong if we let the gray queue be drained and mark/copy everything.
1150 This has the unfortunate side effect of making overflow collections perform the first step twice, but
1151 given we now have heuristics that perform major GC in anticipation of minor overflows this should not
1154 sgen_client_bridge_processing_stw_step ();
1158 Make sure we drain the gray stack before processing disappearing links and finalizers.
1159 If we don't make sure it is empty we might wrongly see a live object as dead.
1161 sgen_drain_gray_stack (ctx);
1164 We must clear weak links that don't track resurrection before processing object ready for
1165 finalization so they can be cleared before that.
1167 sgen_null_link_in_range (generation, ctx, FALSE);
1168 if (generation == GENERATION_OLD)
1169 sgen_null_link_in_range (GENERATION_NURSERY, ctx, FALSE);
1172 /* walk the finalization queue and move also the objects that need to be
1173 * finalized: use the finalized objects as new roots so the objects they depend
1174 * on are also not reclaimed. As with the roots above, only objects in the nursery
1175 * are marked/copied.
1177 sgen_finalize_in_range (generation, ctx);
1178 if (generation == GENERATION_OLD)
1179 sgen_finalize_in_range (GENERATION_NURSERY, ctx);
1180 /* drain the new stack that might have been created */
1181 SGEN_LOG (6, "Precise scan of gray area post fin");
1182 sgen_drain_gray_stack (ctx);
1185 * This must be done again after processing finalizable objects since CWL slots are cleared only after the key is finalized.
1187 done_with_ephemerons = 0;
1189 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1190 sgen_drain_gray_stack (ctx);
1192 } while (!done_with_ephemerons);
1194 sgen_client_clear_unreachable_ephemerons (ctx);
1197 * We clear togglerefs only after all possible chances of revival are done.
1198 * This is semantically more inline with what users expect and it allows for
1199 * user finalizers to correctly interact with TR objects.
1201 sgen_client_clear_togglerefs (start_addr, end_addr, ctx);
1204 SGEN_LOG (2, "Finalize queue handling scan for %s generation: %lld usecs %d ephemeron rounds", generation_name (generation), (long long)TV_ELAPSED (atv, btv), ephemeron_rounds);
1207 * handle disappearing links
1208 * Note we do this after checking the finalization queue because if an object
1209 * survives (at least long enough to be finalized) we don't clear the link.
1210 * This also deals with a possible issue with the monitor reclamation: with the Boehm
1211 * GC a finalized object my lose the monitor because it is cleared before the finalizer is
1214 g_assert (sgen_gray_object_queue_is_empty (queue));
1216 sgen_null_link_in_range (generation, ctx, TRUE);
1217 if (generation == GENERATION_OLD)
1218 sgen_null_link_in_range (GENERATION_NURSERY, ctx, TRUE);
1219 if (sgen_gray_object_queue_is_empty (queue))
1221 sgen_drain_gray_stack (ctx);
1224 g_assert (sgen_gray_object_queue_is_empty (queue));
1226 binary_protocol_finish_gray_stack_end (sgen_timestamp (), generation);
1230 sgen_check_section_scan_starts (GCMemSection *section)
1233 for (i = 0; i < section->num_scan_start; ++i) {
1234 if (section->scan_starts [i]) {
1235 mword size = safe_object_get_size ((GCObject*) section->scan_starts [i]);
1236 SGEN_ASSERT (0, size >= SGEN_CLIENT_MINIMUM_OBJECT_SIZE && size <= MAX_SMALL_OBJ_SIZE, "Weird object size at scan starts.");
1242 check_scan_starts (void)
1244 if (!do_scan_starts_check)
1246 sgen_check_section_scan_starts (nursery_section);
1247 major_collector.check_scan_starts ();
1251 scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx)
1255 SGEN_HASH_TABLE_FOREACH (&roots_hash [root_type], void **, start_root, RootRecord *, root) {
1256 SGEN_LOG (6, "Precise root scan %p-%p (desc: %p)", start_root, root->end_root, (void*)root->root_desc);
1257 precisely_scan_objects_from (start_root, (void**)root->end_root, addr_start, addr_end, root->root_desc, ctx);
1258 } SGEN_HASH_TABLE_FOREACH_END;
1264 static gboolean inited = FALSE;
1269 mono_counters_register ("Collection max time", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME | MONO_COUNTER_MONOTONIC, &time_max);
1271 mono_counters_register ("Minor fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pre_collection_fragment_clear);
1272 mono_counters_register ("Minor pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pinning);
1273 mono_counters_register ("Minor scan remembered set", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_remsets);
1274 mono_counters_register ("Minor scan major blocks", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_major_blocks);
1275 mono_counters_register ("Minor scan los", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_los);
1276 mono_counters_register ("Minor scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_pinned);
1277 mono_counters_register ("Minor scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_roots);
1278 mono_counters_register ("Minor fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_fragment_creation);
1280 mono_counters_register ("Major fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pre_collection_fragment_clear);
1281 mono_counters_register ("Major pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pinning);
1282 mono_counters_register ("Major scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_pinned);
1283 mono_counters_register ("Major scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_roots);
1284 mono_counters_register ("Major scan mod union blocks", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_mod_union_blocks);
1285 mono_counters_register ("Major scan mod union los", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_mod_union_los);
1286 mono_counters_register ("Major finish gray stack", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_finish_gray_stack);
1287 mono_counters_register ("Major free big objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_free_bigobjs);
1288 mono_counters_register ("Major LOS sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_los_sweep);
1289 mono_counters_register ("Major sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_sweep);
1290 mono_counters_register ("Major fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_fragment_creation);
1292 mono_counters_register ("Number of pinned objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_pinned_objects);
1294 #ifdef HEAVY_STATISTICS
1295 mono_counters_register ("WBarrier remember pointer", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_add_to_global_remset);
1296 mono_counters_register ("WBarrier arrayref copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_arrayref_copy);
1297 mono_counters_register ("WBarrier generic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store);
1298 mono_counters_register ("WBarrier generic atomic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store_atomic);
1299 mono_counters_register ("WBarrier set root", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_set_root);
1301 mono_counters_register ("# objects allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_alloced_degraded);
1302 mono_counters_register ("bytes allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced_degraded);
1304 mono_counters_register ("# copy_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_nursery);
1305 mono_counters_register ("# objects copied (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_nursery);
1306 mono_counters_register ("# copy_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_major);
1307 mono_counters_register ("# objects copied (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_major);
1309 mono_counters_register ("# scan_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_nursery);
1310 mono_counters_register ("# scan_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_major);
1312 mono_counters_register ("Slots allocated in vain", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_slots_allocated_in_vain);
1314 mono_counters_register ("# nursery copy_object() failed from space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_from_space);
1315 mono_counters_register ("# nursery copy_object() failed forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_forwarded);
1316 mono_counters_register ("# nursery copy_object() failed pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_pinned);
1317 mono_counters_register ("# nursery copy_object() failed to space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_to_space);
1319 sgen_nursery_allocator_init_heavy_stats ();
1327 reset_pinned_from_failed_allocation (void)
1329 bytes_pinned_from_failed_allocation = 0;
1333 sgen_set_pinned_from_failed_allocation (mword objsize)
1335 bytes_pinned_from_failed_allocation += objsize;
1339 sgen_collection_is_concurrent (void)
1341 switch (current_collection_generation) {
1342 case GENERATION_NURSERY:
1344 case GENERATION_OLD:
1345 return concurrent_collection_in_progress;
1347 g_error ("Invalid current generation %d", current_collection_generation);
1353 sgen_concurrent_collection_in_progress (void)
1355 return concurrent_collection_in_progress;
1359 SgenThreadPoolJob job;
1360 SgenObjectOperations *ops;
1361 SgenGrayQueue *gc_thread_gray_queue;
1366 int job_index, job_split_count;
1370 static ScanCopyContext
1371 scan_copy_context_for_scan_job (void *worker_data_untyped, ScanJob *job)
1373 WorkerData *worker_data = (WorkerData *)worker_data_untyped;
1377 * For jobs enqueued on workers we set the ops at job runtime in order
1378 * to be able to profit from on the fly optimized object ops or other
1379 * object ops changes, like forced concurrent finish.
1381 SGEN_ASSERT (0, sgen_workers_is_worker_thread (mono_native_thread_id_get ()), "We need a context for the scan job");
1382 job->ops = sgen_workers_get_idle_func_object_ops (worker_data);
1385 return CONTEXT_FROM_OBJECT_OPERATIONS (job->ops, sgen_workers_get_job_gray_queue (worker_data, job->gc_thread_gray_queue));
1393 } ScanFromRegisteredRootsJob;
1396 job_scan_from_registered_roots (void *worker_data_untyped, SgenThreadPoolJob *job)
1398 ScanFromRegisteredRootsJob *job_data = (ScanFromRegisteredRootsJob*)job;
1399 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, &job_data->scan_job);
1401 scan_from_registered_roots (job_data->heap_start, job_data->heap_end, job_data->root_type, ctx);
1408 } ScanThreadDataJob;
1411 job_scan_thread_data (void *worker_data_untyped, SgenThreadPoolJob *job)
1413 ScanThreadDataJob *job_data = (ScanThreadDataJob*)job;
1414 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, &job_data->scan_job);
1416 sgen_client_scan_thread_data (job_data->heap_start, job_data->heap_end, TRUE, ctx);
1421 SgenPointerQueue *queue;
1422 } ScanFinalizerEntriesJob;
1425 job_scan_finalizer_entries (void *worker_data_untyped, SgenThreadPoolJob *job)
1427 ScanFinalizerEntriesJob *job_data = (ScanFinalizerEntriesJob*)job;
1428 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, &job_data->scan_job);
1430 scan_finalizer_entries (job_data->queue, ctx);
1434 job_scan_wbroots (void *worker_data_untyped, SgenThreadPoolJob *job)
1436 ScanJob *job_data = (ScanJob*)job;
1437 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, job_data);
1439 sgen_wbroots_scan_card_table (ctx);
1443 job_scan_major_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1445 SGEN_TV_DECLARE (atv);
1446 SGEN_TV_DECLARE (btv);
1447 ParallelScanJob *job_data = (ParallelScanJob*)job;
1448 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
1450 SGEN_TV_GETTIME (atv);
1451 major_collector.scan_card_table (CARDTABLE_SCAN_GLOBAL, ctx, job_data->job_index, job_data->job_split_count, job_data->data);
1452 SGEN_TV_GETTIME (btv);
1453 time_minor_scan_major_blocks += SGEN_TV_ELAPSED (atv, btv);
1455 if (worker_data_untyped)
1456 ((WorkerData*)worker_data_untyped)->major_scan_time += SGEN_TV_ELAPSED (atv, btv);
1460 job_scan_los_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1462 SGEN_TV_DECLARE (atv);
1463 SGEN_TV_DECLARE (btv);
1464 ParallelScanJob *job_data = (ParallelScanJob*)job;
1465 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
1467 SGEN_TV_GETTIME (atv);
1468 sgen_los_scan_card_table (CARDTABLE_SCAN_GLOBAL, ctx, job_data->job_index, job_data->job_split_count);
1469 SGEN_TV_GETTIME (btv);
1470 time_minor_scan_los += SGEN_TV_ELAPSED (atv, btv);
1472 if (worker_data_untyped)
1473 ((WorkerData*)worker_data_untyped)->los_scan_time += SGEN_TV_ELAPSED (atv, btv);
1477 job_scan_major_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1479 SGEN_TV_DECLARE (atv);
1480 SGEN_TV_DECLARE (btv);
1481 ParallelScanJob *job_data = (ParallelScanJob*)job;
1482 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
1484 g_assert (concurrent_collection_in_progress);
1485 SGEN_TV_GETTIME (atv);
1486 major_collector.scan_card_table (CARDTABLE_SCAN_MOD_UNION, ctx, job_data->job_index, job_data->job_split_count, job_data->data);
1487 SGEN_TV_GETTIME (btv);
1488 time_major_scan_mod_union_blocks += SGEN_TV_ELAPSED (atv, btv);
1490 if (worker_data_untyped)
1491 ((WorkerData*)worker_data_untyped)->major_scan_time += SGEN_TV_ELAPSED (atv, btv);
1495 job_scan_los_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1497 SGEN_TV_DECLARE (atv);
1498 SGEN_TV_DECLARE (btv);
1499 ParallelScanJob *job_data = (ParallelScanJob*)job;
1500 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
1502 g_assert (concurrent_collection_in_progress);
1503 SGEN_TV_GETTIME (atv);
1504 sgen_los_scan_card_table (CARDTABLE_SCAN_MOD_UNION, ctx, job_data->job_index, job_data->job_split_count);
1505 SGEN_TV_GETTIME (btv);
1506 time_major_scan_mod_union_los += SGEN_TV_ELAPSED (atv, btv);
1508 if (worker_data_untyped)
1509 ((WorkerData*)worker_data_untyped)->los_scan_time += SGEN_TV_ELAPSED (atv, btv);
1513 job_major_mod_union_preclean (void *worker_data_untyped, SgenThreadPoolJob *job)
1515 SGEN_TV_DECLARE (atv);
1516 SGEN_TV_DECLARE (btv);
1517 ParallelScanJob *job_data = (ParallelScanJob*)job;
1518 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
1520 g_assert (concurrent_collection_in_progress);
1521 SGEN_TV_GETTIME (atv);
1522 major_collector.scan_card_table (CARDTABLE_SCAN_MOD_UNION_PRECLEAN, ctx, job_data->job_index, job_data->job_split_count, job_data->data);
1523 SGEN_TV_GETTIME (btv);
1525 g_assert (worker_data_untyped);
1526 ((WorkerData*)worker_data_untyped)->major_scan_time += SGEN_TV_ELAPSED (atv, btv);
1530 job_los_mod_union_preclean (void *worker_data_untyped, SgenThreadPoolJob *job)
1532 SGEN_TV_DECLARE (atv);
1533 SGEN_TV_DECLARE (btv);
1534 ParallelScanJob *job_data = (ParallelScanJob*)job;
1535 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
1537 g_assert (concurrent_collection_in_progress);
1538 SGEN_TV_GETTIME (atv);
1539 sgen_los_scan_card_table (CARDTABLE_SCAN_MOD_UNION_PRECLEAN, ctx, job_data->job_index, job_data->job_split_count);
1540 SGEN_TV_GETTIME (btv);
1542 g_assert (worker_data_untyped);
1543 ((WorkerData*)worker_data_untyped)->los_scan_time += SGEN_TV_ELAPSED (atv, btv);
1547 job_scan_last_pinned (void *worker_data_untyped, SgenThreadPoolJob *job)
1549 ScanJob *job_data = (ScanJob*)job;
1550 ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, job_data);
1552 g_assert (concurrent_collection_in_progress);
1554 sgen_scan_pin_queue_objects (ctx);
1558 workers_finish_callback (void)
1560 ParallelScanJob *psj;
1562 size_t num_major_sections = major_collector.get_num_major_sections ();
1563 int split_count = sgen_workers_get_job_split_count (GENERATION_OLD);
1565 /* Mod union preclean jobs */
1566 for (i = 0; i < split_count; i++) {
1567 psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("preclean major mod union cardtable", job_major_mod_union_preclean, sizeof (ParallelScanJob));
1568 psj->scan_job.gc_thread_gray_queue = NULL;
1570 psj->job_split_count = split_count;
1571 psj->data = num_major_sections / split_count;
1572 sgen_workers_enqueue_job (GENERATION_OLD, &psj->scan_job.job, TRUE);
1575 for (i = 0; i < split_count; i++) {
1576 psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("preclean los mod union cardtable", job_los_mod_union_preclean, sizeof (ParallelScanJob));
1577 psj->scan_job.gc_thread_gray_queue = NULL;
1579 psj->job_split_count = split_count;
1580 sgen_workers_enqueue_job (GENERATION_OLD, &psj->scan_job.job, TRUE);
1583 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan last pinned", job_scan_last_pinned, sizeof (ScanJob));
1584 sj->gc_thread_gray_queue = NULL;
1585 sgen_workers_enqueue_job (GENERATION_OLD, &sj->job, TRUE);
1589 init_gray_queue (SgenGrayQueue *gc_thread_gray_queue)
1591 sgen_gray_object_queue_init (gc_thread_gray_queue, NULL, TRUE);
1595 enqueue_scan_remembered_set_jobs (SgenGrayQueue *gc_thread_gray_queue, SgenObjectOperations *ops, gboolean enqueue)
1597 int i, split_count = sgen_workers_get_job_split_count (GENERATION_NURSERY);
1598 size_t num_major_sections = major_collector.get_num_major_sections ();
1601 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan wbroots", job_scan_wbroots, sizeof (ScanJob));
1603 sj->gc_thread_gray_queue = gc_thread_gray_queue;
1604 sgen_workers_enqueue_job (GENERATION_NURSERY, &sj->job, enqueue);
1606 for (i = 0; i < split_count; i++) {
1607 ParallelScanJob *psj;
1609 psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("scan major remsets", job_scan_major_card_table, sizeof (ParallelScanJob));
1610 psj->scan_job.ops = ops;
1611 psj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
1613 psj->job_split_count = split_count;
1614 psj->data = num_major_sections / split_count;
1615 sgen_workers_enqueue_job (GENERATION_NURSERY, &psj->scan_job.job, enqueue);
1617 psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("scan LOS remsets", job_scan_los_card_table, sizeof (ParallelScanJob));
1618 psj->scan_job.ops = ops;
1619 psj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
1621 psj->job_split_count = split_count;
1622 sgen_workers_enqueue_job (GENERATION_NURSERY, &psj->scan_job.job, enqueue);
1627 enqueue_scan_from_roots_jobs (SgenGrayQueue *gc_thread_gray_queue, char *heap_start, char *heap_end, SgenObjectOperations *ops, gboolean enqueue)
1629 ScanFromRegisteredRootsJob *scrrj;
1630 ScanThreadDataJob *stdj;
1631 ScanFinalizerEntriesJob *sfej;
1633 /* registered roots, this includes static fields */
1635 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots normal", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1636 scrrj->scan_job.ops = ops;
1637 scrrj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
1638 scrrj->heap_start = heap_start;
1639 scrrj->heap_end = heap_end;
1640 scrrj->root_type = ROOT_TYPE_NORMAL;
1641 sgen_workers_enqueue_job (current_collection_generation, &scrrj->scan_job.job, enqueue);
1643 if (current_collection_generation == GENERATION_OLD) {
1644 /* During minors we scan the cardtable for these roots instead */
1645 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots wbarrier", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1646 scrrj->scan_job.ops = ops;
1647 scrrj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
1648 scrrj->heap_start = heap_start;
1649 scrrj->heap_end = heap_end;
1650 scrrj->root_type = ROOT_TYPE_WBARRIER;
1651 sgen_workers_enqueue_job (current_collection_generation, &scrrj->scan_job.job, enqueue);
1656 stdj = (ScanThreadDataJob*)sgen_thread_pool_job_alloc ("scan thread data", job_scan_thread_data, sizeof (ScanThreadDataJob));
1657 stdj->scan_job.ops = ops;
1658 stdj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
1659 stdj->heap_start = heap_start;
1660 stdj->heap_end = heap_end;
1661 sgen_workers_enqueue_job (current_collection_generation, &stdj->scan_job.job, enqueue);
1663 /* Scan the list of objects ready for finalization. */
1665 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1666 sfej->scan_job.ops = ops;
1667 sfej->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
1668 sfej->queue = &fin_ready_queue;
1669 sgen_workers_enqueue_job (current_collection_generation, &sfej->scan_job.job, enqueue);
1671 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan critical finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1672 sfej->scan_job.ops = ops;
1673 sfej->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
1674 sfej->queue = &critical_fin_queue;
1675 sgen_workers_enqueue_job (current_collection_generation, &sfej->scan_job.job, enqueue);
1679 * Perform a nursery collection.
1681 * Return whether any objects were late-pinned due to being out of memory.
1684 collect_nursery (const char *reason, gboolean is_overflow, SgenGrayQueue *unpin_queue)
1686 gboolean needs_major, is_parallel = FALSE;
1687 mword fragment_total;
1688 SgenGrayQueue gc_thread_gray_queue;
1689 SgenObjectOperations *object_ops_nopar, *object_ops_par = NULL;
1690 ScanCopyContext ctx;
1693 SGEN_TV_DECLARE (last_minor_collection_start_tv);
1694 SGEN_TV_DECLARE (last_minor_collection_end_tv);
1695 guint64 major_scan_start = time_minor_scan_major_blocks;
1696 guint64 los_scan_start = time_minor_scan_los;
1697 guint64 finish_gray_start = time_minor_finish_gray_stack;
1699 if (disable_minor_collections)
1702 TV_GETTIME (last_minor_collection_start_tv);
1703 atv = last_minor_collection_start_tv;
1705 binary_protocol_collection_begin (InterlockedRead (&gc_stats.minor_gc_count), GENERATION_NURSERY);
1707 object_ops_nopar = sgen_concurrent_collection_in_progress ()
1708 ? &sgen_minor_collector.serial_ops_with_concurrent_major
1709 : &sgen_minor_collector.serial_ops;
1710 if (sgen_minor_collector.is_parallel && sgen_nursery_size >= SGEN_PARALLEL_MINOR_MIN_NURSERY_SIZE) {
1711 object_ops_par = sgen_concurrent_collection_in_progress ()
1712 ? &sgen_minor_collector.parallel_ops_with_concurrent_major
1713 : &sgen_minor_collector.parallel_ops;
1717 if (do_verify_nursery || do_dump_nursery_content)
1718 sgen_debug_verify_nursery (do_dump_nursery_content);
1720 current_collection_generation = GENERATION_NURSERY;
1722 SGEN_ASSERT (0, !sgen_collection_is_concurrent (), "Why is the nursery collection concurrent?");
1724 reset_pinned_from_failed_allocation ();
1726 check_scan_starts ();
1728 sgen_nursery_alloc_prepare_for_minor ();
1733 SGEN_LOG (1, "Start nursery collection %" G_GINT32_FORMAT " %p-%p, size: %d", InterlockedRead (&gc_stats.minor_gc_count), nursery_section->data, nursery_section->end_data, (int)(nursery_section->end_data - nursery_section->data));
1735 /* world must be stopped already */
1737 time_minor_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1739 sgen_client_pre_collection_checks ();
1741 major_collector.start_nursery_collection ();
1743 sgen_memgov_minor_collection_start ();
1745 init_gray_queue (&gc_thread_gray_queue);
1746 ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops_nopar, &gc_thread_gray_queue);
1748 InterlockedIncrement (&gc_stats.minor_gc_count);
1750 sgen_process_fin_stage_entries ();
1752 /* pin from pinned handles */
1753 sgen_init_pinning ();
1754 if (concurrent_collection_in_progress)
1755 sgen_init_pinning_for_conc ();
1756 sgen_client_binary_protocol_mark_start (GENERATION_NURSERY);
1757 pin_from_roots (nursery_section->data, nursery_section->end_data, ctx);
1758 /* pin cemented objects */
1759 sgen_pin_cemented_objects ();
1760 /* identify pinned objects */
1761 sgen_optimize_pin_queue ();
1762 sgen_pinning_setup_section (nursery_section);
1764 pin_objects_in_nursery (FALSE, ctx);
1765 sgen_pinning_trim_queue_to_section (nursery_section);
1766 if (concurrent_collection_in_progress)
1767 sgen_finish_pinning_for_conc ();
1769 if (remset_consistency_checks)
1770 sgen_check_remset_consistency ();
1772 if (whole_heap_check_before_collection) {
1773 sgen_clear_nursery_fragments ();
1774 sgen_check_whole_heap (FALSE);
1778 time_minor_pinning += TV_ELAPSED (btv, atv);
1779 SGEN_LOG (2, "Finding pinned pointers: %zd in %lld usecs", sgen_get_pinned_count (), (long long)TV_ELAPSED (btv, atv));
1780 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
1782 remset.start_scan_remsets ();
1784 enqueue_scan_remembered_set_jobs (&gc_thread_gray_queue, is_parallel ? NULL : object_ops_nopar, is_parallel);
1786 /* we don't have complete write barrier yet, so we scan all the old generation sections */
1788 time_minor_scan_remsets += TV_ELAPSED (atv, btv);
1789 SGEN_LOG (2, "Old generation scan: %lld usecs", (long long)TV_ELAPSED (atv, btv));
1791 sgen_pin_stats_report ();
1793 /* FIXME: Why do we do this at this specific, seemingly random, point? */
1794 sgen_client_collecting_minor (&fin_ready_queue, &critical_fin_queue);
1797 time_minor_scan_pinned += TV_ELAPSED (btv, atv);
1799 enqueue_scan_from_roots_jobs (&gc_thread_gray_queue, nursery_section->data, nursery_section->end_data, is_parallel ? NULL : object_ops_nopar, is_parallel);
1802 gray_queue_redirect (&gc_thread_gray_queue);
1803 sgen_workers_start_all_workers (GENERATION_NURSERY, object_ops_nopar, object_ops_par, NULL);
1804 sgen_workers_join (GENERATION_NURSERY);
1808 time_minor_scan_roots += TV_ELAPSED (atv, btv);
1810 finish_gray_stack (GENERATION_NURSERY, ctx);
1813 time_minor_finish_gray_stack += TV_ELAPSED (btv, atv);
1814 sgen_client_binary_protocol_mark_end (GENERATION_NURSERY);
1816 if (objects_pinned) {
1817 sgen_optimize_pin_queue ();
1818 sgen_pinning_setup_section (nursery_section);
1822 * This is the latest point at which we can do this check, because
1823 * sgen_build_nursery_fragments() unpins nursery objects again.
1825 if (remset_consistency_checks)
1826 sgen_check_remset_consistency ();
1829 if (sgen_max_pause_time) {
1833 duration = (int)(TV_ELAPSED (last_minor_collection_start_tv, btv) / 10000);
1834 if (duration > (sgen_max_pause_time * sgen_max_pause_margin))
1835 sgen_resize_nursery (TRUE);
1837 sgen_resize_nursery (FALSE);
1839 sgen_resize_nursery (FALSE);
1842 /* walk the pin_queue, build up the fragment list of free memory, unmark
1843 * pinned objects as we go, memzero() the empty fragments so they are ready for the
1846 sgen_client_binary_protocol_reclaim_start (GENERATION_NURSERY);
1847 fragment_total = sgen_build_nursery_fragments (nursery_section, unpin_queue);
1848 if (!fragment_total)
1851 /* Clear TLABs for all threads */
1852 sgen_clear_tlabs ();
1854 sgen_client_binary_protocol_reclaim_end (GENERATION_NURSERY);
1856 time_minor_fragment_creation += TV_ELAPSED (atv, btv);
1857 SGEN_LOG (2, "Fragment creation: %lld usecs, %lu bytes available", (long long)TV_ELAPSED (atv, btv), (unsigned long)fragment_total);
1859 if (remset_consistency_checks)
1860 sgen_check_major_refs ();
1862 major_collector.finish_nursery_collection ();
1864 TV_GETTIME (last_minor_collection_end_tv);
1865 InterlockedAdd64 (&gc_stats.minor_gc_time, TV_ELAPSED (last_minor_collection_start_tv, last_minor_collection_end_tv));
1867 sgen_debug_dump_heap ("minor", InterlockedRead (&gc_stats.minor_gc_count) - 1, NULL);
1869 /* prepare the pin queue for the next collection */
1870 sgen_finish_pinning ();
1871 if (sgen_have_pending_finalizers ()) {
1872 SGEN_LOG (4, "Finalizer-thread wakeup");
1873 sgen_client_finalize_notify ();
1875 sgen_pin_stats_reset ();
1876 /* clear cemented hash */
1877 sgen_cement_clear_below_threshold ();
1879 sgen_gray_object_queue_dispose (&gc_thread_gray_queue);
1881 check_scan_starts ();
1883 binary_protocol_flush_buffers (FALSE);
1885 sgen_memgov_minor_collection_end (reason, is_overflow);
1887 /*objects are late pinned because of lack of memory, so a major is a good call*/
1888 needs_major = objects_pinned > 0;
1889 current_collection_generation = -1;
1893 binary_protocol_collection_end_stats (0, 0, time_minor_finish_gray_stack - finish_gray_start);
1895 binary_protocol_collection_end_stats (
1896 time_minor_scan_major_blocks - major_scan_start,
1897 time_minor_scan_los - los_scan_start,
1898 time_minor_finish_gray_stack - finish_gray_start);
1900 binary_protocol_collection_end (InterlockedRead (&gc_stats.minor_gc_count) - 1, GENERATION_NURSERY, 0, 0);
1902 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
1903 sgen_check_nursery_objects_pinned (unpin_queue != NULL);
1909 COPY_OR_MARK_FROM_ROOTS_SERIAL,
1910 COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT,
1911 COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT
1912 } CopyOrMarkFromRootsMode;
1915 major_copy_or_mark_from_roots (SgenGrayQueue *gc_thread_gray_queue, size_t *old_next_pin_slot, CopyOrMarkFromRootsMode mode, SgenObjectOperations *object_ops_nopar, SgenObjectOperations *object_ops_par)
1920 /* FIXME: only use these values for the precise scan
1921 * note that to_space pointers should be excluded anyway...
1923 char *heap_start = NULL;
1924 char *heap_end = (char*)-1;
1925 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops_nopar, gc_thread_gray_queue);
1926 gboolean concurrent = mode != COPY_OR_MARK_FROM_ROOTS_SERIAL;
1928 SGEN_ASSERT (0, !!concurrent == !!concurrent_collection_in_progress, "We've been called with the wrong mode.");
1930 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1931 /*This cleans up unused fragments */
1932 sgen_nursery_allocator_prepare_for_pinning ();
1934 if (do_concurrent_checks)
1935 sgen_debug_check_nursery_is_clean ();
1937 /* The concurrent collector doesn't touch the nursery. */
1938 sgen_nursery_alloc_prepare_for_major ();
1943 /* Pinning depends on this */
1944 sgen_clear_nursery_fragments ();
1946 if (whole_heap_check_before_collection)
1947 sgen_check_whole_heap (TRUE);
1950 time_major_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1954 sgen_client_pre_collection_checks ();
1956 if (mode != COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1957 /* Remsets are not useful for a major collection */
1958 remset.clear_cards ();
1961 sgen_process_fin_stage_entries ();
1964 sgen_init_pinning ();
1965 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT)
1966 sgen_init_pinning_for_conc ();
1967 SGEN_LOG (6, "Collecting pinned addresses");
1968 pin_from_roots ((void*)lowest_heap_address, (void*)highest_heap_address, ctx);
1969 if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
1970 /* Pin cemented objects that were forced */
1971 sgen_pin_cemented_objects ();
1973 sgen_optimize_pin_queue ();
1974 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1976 * Cemented objects that are in the pinned list will be marked. When
1977 * marking concurrently we won't mark mod-union cards for these objects.
1978 * Instead they will remain cemented until the next major collection,
1979 * when we will recheck if they are still pinned in the roots.
1981 sgen_cement_force_pinned ();
1984 sgen_client_collecting_major_1 ();
1987 * pin_queue now contains all candidate pointers, sorted and
1988 * uniqued. We must do two passes now to figure out which
1989 * objects are pinned.
1991 * The first is to find within the pin_queue the area for each
1992 * section. This requires that the pin_queue be sorted. We
1993 * also process the LOS objects and pinned chunks here.
1995 * The second, destructive, pass is to reduce the section
1996 * areas to pointers to the actually pinned objects.
1998 SGEN_LOG (6, "Pinning from sections");
1999 /* first pass for the sections */
2000 sgen_find_section_pin_queue_start_end (nursery_section);
2001 /* identify possible pointers to the insize of large objects */
2002 SGEN_LOG (6, "Pinning from large objects");
2003 for (bigobj = los_object_list; bigobj; bigobj = bigobj->next) {
2005 if (sgen_find_optimized_pin_queue_area ((char*)bigobj->data, (char*)bigobj->data + sgen_los_object_size (bigobj), &dummy, &dummy)) {
2006 binary_protocol_pin (bigobj->data, (gpointer)LOAD_VTABLE (bigobj->data), safe_object_get_size (bigobj->data));
2008 if (sgen_los_object_is_pinned (bigobj->data)) {
2009 SGEN_ASSERT (0, mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT, "LOS objects can only be pinned here after concurrent marking.");
2012 sgen_los_pin_object (bigobj->data);
2013 if (SGEN_OBJECT_HAS_REFERENCES (bigobj->data))
2014 GRAY_OBJECT_ENQUEUE_SERIAL (gc_thread_gray_queue, bigobj->data, sgen_obj_get_descriptor ((GCObject*)bigobj->data));
2015 sgen_pin_stats_register_object (bigobj->data, GENERATION_OLD);
2016 SGEN_LOG (6, "Marked large object %p (%s) size: %lu from roots", bigobj->data,
2017 sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (bigobj->data)),
2018 (unsigned long)sgen_los_object_size (bigobj));
2020 sgen_client_pinned_los_object (bigobj->data);
2024 pin_objects_in_nursery (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT, ctx);
2025 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
2026 sgen_check_nursery_objects_pinned (mode != COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT);
2028 major_collector.pin_objects (gc_thread_gray_queue);
2029 if (old_next_pin_slot)
2030 *old_next_pin_slot = sgen_get_pinned_count ();
2033 time_major_pinning += TV_ELAPSED (atv, btv);
2034 SGEN_LOG (2, "Finding pinned pointers: %zd in %lld usecs", sgen_get_pinned_count (), (long long)TV_ELAPSED (atv, btv));
2035 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
2037 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT)
2038 sgen_finish_pinning_for_conc ();
2040 major_collector.init_to_space ();
2042 SGEN_ASSERT (0, sgen_workers_all_done (), "Why are the workers not done when we start or finish a major collection?");
2043 if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
2044 if (object_ops_par != NULL)
2045 sgen_workers_set_num_active_workers (GENERATION_OLD, 0);
2046 if (object_ops_par == NULL && sgen_workers_have_idle_work (GENERATION_OLD)) {
2048 * We force the finish of the worker with the new object ops context
2049 * which can also do copying. We need to have finished pinning. On the
2050 * parallel collector, there is no need to drain the private queues
2051 * here, since we can do it as part of the finishing work, achieving
2052 * better work distribution.
2054 sgen_workers_start_all_workers (GENERATION_OLD, object_ops_nopar, object_ops_par, NULL);
2056 sgen_workers_join (GENERATION_OLD);
2060 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
2061 main_gc_thread = mono_native_thread_self ();
2064 sgen_client_collecting_major_2 ();
2067 time_major_scan_pinned += TV_ELAPSED (btv, atv);
2069 sgen_client_collecting_major_3 (&fin_ready_queue, &critical_fin_queue);
2071 enqueue_scan_from_roots_jobs (gc_thread_gray_queue, heap_start, heap_end, object_ops_nopar, FALSE);
2074 time_major_scan_roots += TV_ELAPSED (atv, btv);
2077 * We start the concurrent worker after pinning and after we scanned the roots
2078 * in order to make sure that the worker does not finish before handling all
2081 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
2082 sgen_workers_set_num_active_workers (GENERATION_OLD, 1);
2083 gray_queue_redirect (gc_thread_gray_queue);
2084 if (precleaning_enabled) {
2085 sgen_workers_start_all_workers (GENERATION_OLD, object_ops_nopar, object_ops_par, workers_finish_callback);
2087 sgen_workers_start_all_workers (GENERATION_OLD, object_ops_nopar, object_ops_par, NULL);
2091 if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
2092 int i, split_count = sgen_workers_get_job_split_count (GENERATION_OLD);
2093 size_t num_major_sections = major_collector.get_num_major_sections ();
2094 gboolean parallel = object_ops_par != NULL;
2096 /* If we're not parallel we finish the collection on the gc thread */
2098 gray_queue_redirect (gc_thread_gray_queue);
2100 /* Mod union card table */
2101 for (i = 0; i < split_count; i++) {
2102 ParallelScanJob *psj;
2104 psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("scan mod union cardtable", job_scan_major_mod_union_card_table, sizeof (ParallelScanJob));
2105 psj->scan_job.ops = parallel ? NULL : object_ops_nopar;
2106 psj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
2108 psj->job_split_count = split_count;
2109 psj->data = num_major_sections / split_count;
2110 sgen_workers_enqueue_job (GENERATION_OLD, &psj->scan_job.job, parallel);
2112 psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("scan LOS mod union cardtable", job_scan_los_mod_union_card_table, sizeof (ParallelScanJob));
2113 psj->scan_job.ops = parallel ? NULL : object_ops_nopar;
2114 psj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
2116 psj->job_split_count = split_count;
2117 sgen_workers_enqueue_job (GENERATION_OLD, &psj->scan_job.job, parallel);
2122 * If we enqueue a job while workers are running we need to sgen_workers_ensure_awake
2123 * in order to make sure that we are running the idle func and draining all worker
2124 * gray queues. The operation of starting workers implies this, so we start them after
2125 * in order to avoid doing this operation twice. The workers will drain the main gray
2126 * stack that contained roots and pinned objects and also scan the mod union card
2129 sgen_workers_start_all_workers (GENERATION_OLD, object_ops_nopar, object_ops_par, NULL);
2130 sgen_workers_join (GENERATION_OLD);
2134 sgen_pin_stats_report ();
2136 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
2137 sgen_finish_pinning ();
2139 sgen_pin_stats_reset ();
2141 if (do_concurrent_checks)
2142 sgen_debug_check_nursery_is_clean ();
2147 major_start_collection (SgenGrayQueue *gc_thread_gray_queue, const char *reason, gboolean concurrent, size_t *old_next_pin_slot)
2149 SgenObjectOperations *object_ops_nopar, *object_ops_par = NULL;
2151 binary_protocol_collection_begin (InterlockedRead (&gc_stats.major_gc_count), GENERATION_OLD);
2153 current_collection_generation = GENERATION_OLD;
2155 sgen_workers_assert_gray_queue_is_empty (GENERATION_OLD);
2158 sgen_cement_reset ();
2161 g_assert (major_collector.is_concurrent);
2162 concurrent_collection_in_progress = TRUE;
2164 object_ops_nopar = &major_collector.major_ops_concurrent_start;
2165 if (major_collector.is_parallel)
2166 object_ops_par = &major_collector.major_ops_conc_par_start;
2169 object_ops_nopar = &major_collector.major_ops_serial;
2172 reset_pinned_from_failed_allocation ();
2174 sgen_memgov_major_collection_start (concurrent, reason);
2176 //count_ref_nonref_objs ();
2177 //consistency_check ();
2179 check_scan_starts ();
2182 SGEN_LOG (1, "Start major collection %" G_GINT32_FORMAT, InterlockedRead (&gc_stats.major_gc_count));
2183 InterlockedIncrement (&gc_stats.major_gc_count);
2185 if (major_collector.start_major_collection)
2186 major_collector.start_major_collection ();
2188 major_copy_or_mark_from_roots (gc_thread_gray_queue, old_next_pin_slot, concurrent ? COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT : COPY_OR_MARK_FROM_ROOTS_SERIAL, object_ops_nopar, object_ops_par);
2192 major_finish_collection (SgenGrayQueue *gc_thread_gray_queue, const char *reason, gboolean is_overflow, size_t old_next_pin_slot, gboolean forced)
2194 ScannedObjectCounts counts;
2195 SgenObjectOperations *object_ops_nopar;
2196 mword fragment_total;
2199 guint64 major_scan_start = time_major_scan_mod_union_blocks;
2200 guint64 los_scan_start = time_major_scan_mod_union_los;
2201 guint64 finish_gray_start = time_major_finish_gray_stack;
2203 if (concurrent_collection_in_progress) {
2204 SgenObjectOperations *object_ops_par = NULL;
2206 object_ops_nopar = &major_collector.major_ops_concurrent_finish;
2207 if (major_collector.is_parallel)
2208 object_ops_par = &major_collector.major_ops_conc_par_finish;
2210 major_copy_or_mark_from_roots (gc_thread_gray_queue, NULL, COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT, object_ops_nopar, object_ops_par);
2212 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
2213 main_gc_thread = NULL;
2216 object_ops_nopar = &major_collector.major_ops_serial;
2219 sgen_workers_assert_gray_queue_is_empty (GENERATION_OLD);
2222 finish_gray_stack (GENERATION_OLD, CONTEXT_FROM_OBJECT_OPERATIONS (object_ops_nopar, gc_thread_gray_queue));
2224 time_major_finish_gray_stack += TV_ELAPSED (btv, atv);
2226 SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after joining");
2228 if (objects_pinned) {
2229 g_assert (!concurrent_collection_in_progress);
2232 * This is slow, but we just OOM'd.
2234 * See comment at `sgen_pin_queue_clear_discarded_entries` for how the pin
2235 * queue is laid out at this point.
2237 sgen_pin_queue_clear_discarded_entries (nursery_section, old_next_pin_slot);
2239 * We need to reestablish all pinned nursery objects in the pin queue
2240 * because they're needed for fragment creation. Unpinning happens by
2241 * walking the whole queue, so it's not necessary to reestablish where major
2242 * heap block pins are - all we care is that they're still in there
2245 sgen_optimize_pin_queue ();
2246 sgen_find_section_pin_queue_start_end (nursery_section);
2250 reset_heap_boundaries ();
2251 sgen_update_heap_boundaries ((mword)sgen_get_nursery_start (), (mword)sgen_get_nursery_end ());
2253 /* walk the pin_queue, build up the fragment list of free memory, unmark
2254 * pinned objects as we go, memzero() the empty fragments so they are ready for the
2257 fragment_total = sgen_build_nursery_fragments (nursery_section, NULL);
2258 if (!fragment_total)
2260 SGEN_LOG (4, "Free space in nursery after major %ld", (long)fragment_total);
2262 if (do_concurrent_checks && concurrent_collection_in_progress)
2263 sgen_debug_check_nursery_is_clean ();
2265 /* prepare the pin queue for the next collection */
2266 sgen_finish_pinning ();
2268 /* Clear TLABs for all threads */
2269 sgen_clear_tlabs ();
2271 sgen_pin_stats_reset ();
2273 sgen_cement_clear_below_threshold ();
2275 if (check_mark_bits_after_major_collection)
2276 sgen_check_heap_marked (concurrent_collection_in_progress);
2279 time_major_fragment_creation += TV_ELAPSED (atv, btv);
2281 binary_protocol_sweep_begin (GENERATION_OLD, !major_collector.sweeps_lazily);
2282 sgen_memgov_major_pre_sweep ();
2285 time_major_free_bigobjs += TV_ELAPSED (btv, atv);
2290 time_major_los_sweep += TV_ELAPSED (atv, btv);
2292 major_collector.sweep ();
2294 binary_protocol_sweep_end (GENERATION_OLD, !major_collector.sweeps_lazily);
2297 time_major_sweep += TV_ELAPSED (btv, atv);
2299 sgen_debug_dump_heap ("major", InterlockedRead (&gc_stats.major_gc_count) - 1, reason);
2301 if (sgen_have_pending_finalizers ()) {
2302 SGEN_LOG (4, "Finalizer-thread wakeup");
2303 sgen_client_finalize_notify ();
2306 sgen_memgov_major_collection_end (forced, concurrent_collection_in_progress, reason, is_overflow);
2307 current_collection_generation = -1;
2309 memset (&counts, 0, sizeof (ScannedObjectCounts));
2310 major_collector.finish_major_collection (&counts);
2312 sgen_workers_assert_gray_queue_is_empty (GENERATION_OLD);
2314 SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after major collection has finished");
2315 if (concurrent_collection_in_progress)
2316 concurrent_collection_in_progress = FALSE;
2318 check_scan_starts ();
2320 binary_protocol_flush_buffers (FALSE);
2322 //consistency_check ();
2323 if (major_collector.is_parallel)
2324 binary_protocol_collection_end_stats (0, 0, time_major_finish_gray_stack - finish_gray_start);
2326 binary_protocol_collection_end_stats (
2327 time_major_scan_mod_union_blocks - major_scan_start,
2328 time_major_scan_mod_union_los - los_scan_start,
2329 time_major_finish_gray_stack - finish_gray_start);
2331 binary_protocol_collection_end (InterlockedRead (&gc_stats.major_gc_count) - 1, GENERATION_OLD, counts.num_scanned_objects, counts.num_unique_scanned_objects);
2335 major_do_collection (const char *reason, gboolean is_overflow, gboolean forced)
2337 TV_DECLARE (time_start);
2338 TV_DECLARE (time_end);
2339 size_t old_next_pin_slot;
2340 SgenGrayQueue gc_thread_gray_queue;
2342 if (disable_major_collections)
2345 if (major_collector.get_and_reset_num_major_objects_marked) {
2346 long long num_marked = major_collector.get_and_reset_num_major_objects_marked ();
2347 g_assert (!num_marked);
2350 /* world must be stopped already */
2351 TV_GETTIME (time_start);
2353 init_gray_queue (&gc_thread_gray_queue);
2354 major_start_collection (&gc_thread_gray_queue, reason, FALSE, &old_next_pin_slot);
2355 major_finish_collection (&gc_thread_gray_queue, reason, is_overflow, old_next_pin_slot, forced);
2356 sgen_gray_object_queue_dispose (&gc_thread_gray_queue);
2358 TV_GETTIME (time_end);
2359 InterlockedAdd64 (&gc_stats.major_gc_time, TV_ELAPSED (time_start, time_end));
2361 /* FIXME: also report this to the user, preferably in gc-end. */
2362 if (major_collector.get_and_reset_num_major_objects_marked)
2363 major_collector.get_and_reset_num_major_objects_marked ();
2365 return bytes_pinned_from_failed_allocation > 0;
2369 major_start_concurrent_collection (const char *reason)
2371 TV_DECLARE (time_start);
2372 TV_DECLARE (time_end);
2373 long long num_objects_marked;
2374 SgenGrayQueue gc_thread_gray_queue;
2376 if (disable_major_collections)
2379 TV_GETTIME (time_start);
2380 SGEN_TV_GETTIME (time_major_conc_collection_start);
2382 num_objects_marked = major_collector.get_and_reset_num_major_objects_marked ();
2383 g_assert (num_objects_marked == 0);
2385 binary_protocol_concurrent_start ();
2387 init_gray_queue (&gc_thread_gray_queue);
2388 // FIXME: store reason and pass it when finishing
2389 major_start_collection (&gc_thread_gray_queue, reason, TRUE, NULL);
2390 sgen_gray_object_queue_dispose (&gc_thread_gray_queue);
2392 num_objects_marked = major_collector.get_and_reset_num_major_objects_marked ();
2394 TV_GETTIME (time_end);
2395 InterlockedAdd64 (&gc_stats.major_gc_time, TV_ELAPSED (time_start, time_end));
2397 current_collection_generation = -1;
2401 * Returns whether the major collection has finished.
2404 major_should_finish_concurrent_collection (void)
2406 return sgen_workers_all_done ();
2410 major_update_concurrent_collection (void)
2412 TV_DECLARE (total_start);
2413 TV_DECLARE (total_end);
2415 TV_GETTIME (total_start);
2417 binary_protocol_concurrent_update ();
2419 major_collector.update_cardtable_mod_union ();
2420 sgen_los_update_cardtable_mod_union ();
2422 TV_GETTIME (total_end);
2423 InterlockedAdd64 (&gc_stats.major_gc_time, TV_ELAPSED (total_start, total_end));
2427 major_finish_concurrent_collection (gboolean forced)
2429 SgenGrayQueue gc_thread_gray_queue;
2430 TV_DECLARE (total_start);
2431 TV_DECLARE (total_end);
2433 TV_GETTIME (total_start);
2435 binary_protocol_concurrent_finish ();
2438 * We need to stop all workers since we're updating the cardtable below.
2439 * The workers will be resumed with a finishing pause context to avoid
2440 * additional cardtable and object scanning.
2442 sgen_workers_stop_all_workers (GENERATION_OLD);
2444 SGEN_TV_GETTIME (time_major_conc_collection_end);
2445 InterlockedAdd64 (&gc_stats.major_gc_time_concurrent, SGEN_TV_ELAPSED (time_major_conc_collection_start, time_major_conc_collection_end));
2447 major_collector.update_cardtable_mod_union ();
2448 sgen_los_update_cardtable_mod_union ();
2450 if (mod_union_consistency_check)
2451 sgen_check_mod_union_consistency ();
2453 current_collection_generation = GENERATION_OLD;
2454 sgen_cement_reset ();
2455 init_gray_queue (&gc_thread_gray_queue);
2456 major_finish_collection (&gc_thread_gray_queue, "finishing", FALSE, -1, forced);
2457 sgen_gray_object_queue_dispose (&gc_thread_gray_queue);
2459 TV_GETTIME (total_end);
2460 InterlockedAdd64 (&gc_stats.major_gc_time, TV_ELAPSED (total_start, total_end));
2462 current_collection_generation = -1;
2466 * Ensure an allocation request for @size will succeed by freeing enough memory.
2468 * LOCKING: The GC lock MUST be held.
2471 sgen_ensure_free_space (size_t size, int generation)
2473 int generation_to_collect = -1;
2474 const char *reason = NULL;
2476 if (generation == GENERATION_OLD) {
2477 if (sgen_need_major_collection (size)) {
2478 reason = "LOS overflow";
2479 generation_to_collect = GENERATION_OLD;
2482 if (degraded_mode) {
2483 if (sgen_need_major_collection (size)) {
2484 reason = "Degraded mode overflow";
2485 generation_to_collect = GENERATION_OLD;
2487 } else if (sgen_need_major_collection (size)) {
2488 reason = concurrent_collection_in_progress ? "Forced finish concurrent collection" : "Minor allowance";
2489 generation_to_collect = GENERATION_OLD;
2491 generation_to_collect = GENERATION_NURSERY;
2492 reason = "Nursery full";
2496 if (generation_to_collect == -1) {
2497 if (concurrent_collection_in_progress && sgen_workers_all_done ()) {
2498 generation_to_collect = GENERATION_OLD;
2499 reason = "Finish concurrent collection";
2503 if (generation_to_collect == -1)
2505 sgen_perform_collection (size, generation_to_collect, reason, FALSE, TRUE);
2509 * LOCKING: Assumes the GC lock is held.
2512 sgen_perform_collection (size_t requested_size, int generation_to_collect, const char *reason, gboolean wait_to_finish, gboolean stw)
2514 TV_DECLARE (gc_total_start);
2515 TV_DECLARE (gc_total_end);
2516 int overflow_generation_to_collect = -1;
2517 int oldest_generation_collected = generation_to_collect;
2518 const char *overflow_reason = NULL;
2519 gboolean finish_concurrent = concurrent_collection_in_progress && (major_should_finish_concurrent_collection () || generation_to_collect == GENERATION_OLD);
2521 binary_protocol_collection_requested (generation_to_collect, requested_size, wait_to_finish ? 1 : 0);
2523 SGEN_ASSERT (0, generation_to_collect == GENERATION_NURSERY || generation_to_collect == GENERATION_OLD, "What generation is this?");
2526 sgen_stop_world (generation_to_collect);
2528 SGEN_ASSERT (0, sgen_is_world_stopped (), "We can only collect if the world is stopped");
2531 TV_GETTIME (gc_total_start);
2533 // FIXME: extract overflow reason
2534 // FIXME: minor overflow for concurrent case
2535 if (generation_to_collect == GENERATION_NURSERY && !finish_concurrent) {
2536 if (concurrent_collection_in_progress)
2537 major_update_concurrent_collection ();
2539 if (collect_nursery (reason, FALSE, NULL) && !concurrent_collection_in_progress) {
2540 overflow_generation_to_collect = GENERATION_OLD;
2541 overflow_reason = "Minor overflow";
2543 } else if (finish_concurrent) {
2544 major_finish_concurrent_collection (wait_to_finish);
2545 oldest_generation_collected = GENERATION_OLD;
2547 SGEN_ASSERT (0, generation_to_collect == GENERATION_OLD, "We should have handled nursery collections above");
2548 if (major_collector.is_concurrent && !wait_to_finish) {
2549 collect_nursery ("Concurrent start", FALSE, NULL);
2550 major_start_concurrent_collection (reason);
2551 oldest_generation_collected = GENERATION_NURSERY;
2552 } else if (major_do_collection (reason, FALSE, wait_to_finish)) {
2553 overflow_generation_to_collect = GENERATION_NURSERY;
2554 overflow_reason = "Excessive pinning";
2558 if (overflow_generation_to_collect != -1) {
2559 SGEN_ASSERT (0, !concurrent_collection_in_progress, "We don't yet support overflow collections with the concurrent collector");
2562 * We need to do an overflow collection, either because we ran out of memory
2563 * or the nursery is fully pinned.
2566 if (overflow_generation_to_collect == GENERATION_NURSERY)
2567 collect_nursery (overflow_reason, TRUE, NULL);
2569 major_do_collection (overflow_reason, TRUE, wait_to_finish);
2571 oldest_generation_collected = MAX (oldest_generation_collected, overflow_generation_to_collect);
2574 SGEN_LOG (2, "Heap size: %lu, LOS size: %lu", (unsigned long)sgen_gc_get_total_heap_allocation (), (unsigned long)los_memory_usage);
2576 /* this also sets the proper pointers for the next allocation */
2577 if (generation_to_collect == GENERATION_NURSERY && !sgen_can_alloc_size (requested_size)) {
2578 /* TypeBuilder and MonoMethod are killing mcs with fragmentation */
2579 SGEN_LOG (1, "nursery collection didn't find enough room for %zd alloc (%zd pinned)", requested_size, sgen_get_pinned_count ());
2580 sgen_dump_pin_queue ();
2584 TV_GETTIME (gc_total_end);
2585 time_max = MAX (time_max, TV_ELAPSED (gc_total_start, gc_total_end));
2588 sgen_restart_world (oldest_generation_collected);
2592 * ######################################################################
2593 * ######## Memory allocation from the OS
2594 * ######################################################################
2595 * This section of code deals with getting memory from the OS and
2596 * allocating memory for GC-internal data structures.
2597 * Internal memory can be handled with a freelist for small objects.
2603 G_GNUC_UNUSED static void
2604 report_internal_mem_usage (void)
2606 printf ("Internal memory usage:\n");
2607 sgen_report_internal_mem_usage ();
2608 printf ("Pinned memory usage:\n");
2609 major_collector.report_pinned_memory_usage ();
2613 * ######################################################################
2614 * ######## Finalization support
2615 * ######################################################################
2619 * If the object has been forwarded it means it's still referenced from a root.
2620 * If it is pinned it's still alive as well.
2621 * A LOS object is only alive if we have pinned it.
2622 * Return TRUE if @obj is ready to be finalized.
2624 static inline gboolean
2625 sgen_is_object_alive (GCObject *object)
2627 if (ptr_in_nursery (object))
2628 return sgen_nursery_is_object_alive (object);
2630 return sgen_major_is_object_alive (object);
2634 * This function returns true if @object is either alive and belongs to the
2635 * current collection - major collections are full heap, so old gen objects
2636 * are never alive during a minor collection.
2639 sgen_is_object_alive_and_on_current_collection (GCObject *object)
2641 if (ptr_in_nursery (object))
2642 return sgen_nursery_is_object_alive (object);
2644 if (current_collection_generation == GENERATION_NURSERY)
2647 return sgen_major_is_object_alive (object);
2652 sgen_gc_is_object_ready_for_finalization (GCObject *object)
2654 return !sgen_is_object_alive (object);
2658 sgen_queue_finalization_entry (GCObject *obj)
2660 gboolean critical = sgen_client_object_has_critical_finalizer (obj);
2662 sgen_pointer_queue_add (critical ? &critical_fin_queue : &fin_ready_queue, obj);
2664 sgen_client_object_queued_for_finalization (obj);
2668 sgen_object_is_live (GCObject *obj)
2670 return sgen_is_object_alive_and_on_current_collection (obj);
2674 * `System.GC.WaitForPendingFinalizers` first checks `sgen_have_pending_finalizers()` to
2675 * determine whether it can exit quickly. The latter must therefore only return FALSE if
2676 * all finalizers have really finished running.
2678 * `sgen_gc_invoke_finalizers()` first dequeues a finalizable object, and then finalizes it.
2679 * This means that just checking whether the queues are empty leaves the possibility that an
2680 * object might have been dequeued but not yet finalized. That's why we need the additional
2681 * flag `pending_unqueued_finalizer`.
2684 static volatile gboolean pending_unqueued_finalizer = FALSE;
2685 volatile gboolean sgen_suspend_finalizers = FALSE;
2688 sgen_set_suspend_finalizers (void)
2690 sgen_suspend_finalizers = TRUE;
2694 sgen_gc_invoke_finalizers (void)
2698 g_assert (!pending_unqueued_finalizer);
2700 /* FIXME: batch to reduce lock contention */
2701 while (sgen_have_pending_finalizers ()) {
2707 * We need to set `pending_unqueued_finalizer` before dequeing the
2708 * finalizable object.
2710 if (!sgen_pointer_queue_is_empty (&fin_ready_queue)) {
2711 pending_unqueued_finalizer = TRUE;
2712 mono_memory_write_barrier ();
2713 obj = (GCObject *)sgen_pointer_queue_pop (&fin_ready_queue);
2714 } else if (!sgen_pointer_queue_is_empty (&critical_fin_queue)) {
2715 pending_unqueued_finalizer = TRUE;
2716 mono_memory_write_barrier ();
2717 obj = (GCObject *)sgen_pointer_queue_pop (&critical_fin_queue);
2723 SGEN_LOG (7, "Finalizing object %p (%s)", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
2731 /* the object is on the stack so it is pinned */
2732 /*g_print ("Calling finalizer for object: %p (%s)\n", obj, sgen_client_object_safe_name (obj));*/
2733 sgen_client_run_finalize (obj);
2736 if (pending_unqueued_finalizer) {
2737 mono_memory_write_barrier ();
2738 pending_unqueued_finalizer = FALSE;
2745 sgen_have_pending_finalizers (void)
2747 if (sgen_suspend_finalizers)
2749 return pending_unqueued_finalizer || !sgen_pointer_queue_is_empty (&fin_ready_queue) || !sgen_pointer_queue_is_empty (&critical_fin_queue);
2753 * ######################################################################
2754 * ######## registered roots support
2755 * ######################################################################
2759 * We do not coalesce roots.
2762 sgen_register_root (char *start, size_t size, SgenDescriptor descr, int root_type, int source, const char *msg)
2764 RootRecord new_root;
2767 for (i = 0; i < ROOT_TYPE_NUM; ++i) {
2768 RootRecord *root = (RootRecord *)sgen_hash_table_lookup (&roots_hash [i], start);
2769 /* we allow changing the size and the descriptor (for thread statics etc) */
2771 size_t old_size = root->end_root - start;
2772 root->end_root = start + size;
2773 SGEN_ASSERT (0, !!root->root_desc == !!descr, "Can't change whether a root is precise or conservative.");
2774 SGEN_ASSERT (0, root->source == source, "Can't change a root's source identifier.");
2775 SGEN_ASSERT (0, !!root->msg == !!msg, "Can't change a root's message.");
2776 root->root_desc = descr;
2778 roots_size -= old_size;
2784 new_root.end_root = start + size;
2785 new_root.root_desc = descr;
2786 new_root.source = source;
2789 sgen_hash_table_replace (&roots_hash [root_type], start, &new_root, NULL);
2792 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);
2799 sgen_deregister_root (char* addr)
2805 for (root_type = 0; root_type < ROOT_TYPE_NUM; ++root_type) {
2806 if (sgen_hash_table_remove (&roots_hash [root_type], addr, &root))
2807 roots_size -= (root.end_root - addr);
2813 sgen_wbroots_iterate_live_block_ranges (sgen_cardtable_block_callback cb)
2817 SGEN_HASH_TABLE_FOREACH (&roots_hash [ROOT_TYPE_WBARRIER], void **, start_root, RootRecord *, root) {
2818 cb ((mword)start_root, (mword)root->end_root - (mword)start_root);
2819 } SGEN_HASH_TABLE_FOREACH_END;
2822 /* Root equivalent of sgen_client_cardtable_scan_object */
2824 sgen_wbroot_scan_card_table (void** start_root, mword size, ScanCopyContext ctx)
2826 ScanPtrFieldFunc scan_field_func = ctx.ops->scan_ptr_field;
2827 guint8 *card_data = sgen_card_table_get_card_scan_address ((mword)start_root);
2828 guint8 *card_base = card_data;
2829 mword card_count = sgen_card_table_number_of_cards_in_range ((mword)start_root, size);
2830 guint8 *card_data_end = card_data + card_count;
2831 mword extra_idx = 0;
2832 char *obj_start = sgen_card_table_align_pointer (start_root);
2833 char *obj_end = (char*)start_root + size;
2834 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2835 guint8 *overflow_scan_end = NULL;
2838 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2839 /*Check for overflow and if so, setup to scan in two steps*/
2840 if (card_data_end >= SGEN_SHADOW_CARDTABLE_END) {
2841 overflow_scan_end = sgen_shadow_cardtable + (card_data_end - SGEN_SHADOW_CARDTABLE_END);
2842 card_data_end = SGEN_SHADOW_CARDTABLE_END;
2848 card_data = sgen_find_next_card (card_data, card_data_end);
2850 for (; card_data < card_data_end; card_data = sgen_find_next_card (card_data + 1, card_data_end)) {
2851 size_t idx = (card_data - card_base) + extra_idx;
2852 char *start = (char*)(obj_start + idx * CARD_SIZE_IN_BYTES);
2853 char *card_end = start + CARD_SIZE_IN_BYTES;
2854 char *elem = start, *first_elem = start;
2857 * Don't clean first and last card on 32bit systems since they
2858 * may also be part from other roots.
2860 if (card_data != card_base && card_data != (card_data_end - 1))
2861 sgen_card_table_prepare_card_for_scanning (card_data);
2863 card_end = MIN (card_end, obj_end);
2865 if (elem < (char*)start_root)
2866 first_elem = elem = (char*)start_root;
2868 for (; elem < card_end; elem += SIZEOF_VOID_P) {
2869 if (*(GCObject**)elem)
2870 scan_field_func (NULL, (GCObject**)elem, ctx.queue);
2873 binary_protocol_card_scan (first_elem, elem - first_elem);
2876 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2877 if (overflow_scan_end) {
2878 extra_idx = card_data - card_base;
2879 card_base = card_data = sgen_shadow_cardtable;
2880 card_data_end = overflow_scan_end;
2881 overflow_scan_end = NULL;
2888 sgen_wbroots_scan_card_table (ScanCopyContext ctx)
2893 SGEN_HASH_TABLE_FOREACH (&roots_hash [ROOT_TYPE_WBARRIER], void **, start_root, RootRecord *, root) {
2894 SGEN_ASSERT (0, (root->root_desc & ROOT_DESC_TYPE_MASK) == ROOT_DESC_VECTOR, "Unsupported root type");
2896 sgen_wbroot_scan_card_table (start_root, (mword)root->end_root - (mword)start_root, ctx);
2897 } SGEN_HASH_TABLE_FOREACH_END;
2901 * ######################################################################
2902 * ######## Thread handling (stop/start code)
2903 * ######################################################################
2907 sgen_get_current_collection_generation (void)
2909 return current_collection_generation;
2913 sgen_thread_attach (SgenThreadInfo* info)
2915 info->tlab_start = info->tlab_next = info->tlab_temp_end = info->tlab_real_end = NULL;
2917 sgen_client_thread_attach (info);
2923 sgen_thread_detach_with_lock (SgenThreadInfo *p)
2925 sgen_client_thread_detach_with_lock (p);
2929 * ######################################################################
2930 * ######## Write barriers
2931 * ######################################################################
2935 * Note: the write barriers first do the needed GC work and then do the actual store:
2936 * this way the value is visible to the conservative GC scan after the write barrier
2937 * itself. If a GC interrupts the barrier in the middle, value will be kept alive by
2938 * the conservative scan, otherwise by the remembered set scan.
2942 * mono_gc_wbarrier_arrayref_copy:
2945 mono_gc_wbarrier_arrayref_copy (gpointer dest_ptr, gpointer src_ptr, int count)
2947 HEAVY_STAT (++stat_wbarrier_arrayref_copy);
2948 /*This check can be done without taking a lock since dest_ptr array is pinned*/
2949 if (ptr_in_nursery (dest_ptr) || count <= 0) {
2950 mono_gc_memmove_aligned (dest_ptr, src_ptr, count * sizeof (gpointer));
2954 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2955 if (binary_protocol_is_heavy_enabled ()) {
2957 for (i = 0; i < count; ++i) {
2958 gpointer dest = (gpointer*)dest_ptr + i;
2959 gpointer obj = *((gpointer*)src_ptr + i);
2961 binary_protocol_wbarrier (dest, obj, (gpointer)LOAD_VTABLE (obj));
2966 remset.wbarrier_arrayref_copy (dest_ptr, src_ptr, count);
2970 * mono_gc_wbarrier_generic_nostore:
2973 mono_gc_wbarrier_generic_nostore (gpointer ptr)
2977 HEAVY_STAT (++stat_wbarrier_generic_store);
2979 sgen_client_wbarrier_generic_nostore_check (ptr);
2981 obj = *(gpointer*)ptr;
2983 binary_protocol_wbarrier (ptr, obj, (gpointer)LOAD_VTABLE (obj));
2986 * We need to record old->old pointer locations for the
2987 * concurrent collector.
2989 if (!ptr_in_nursery (obj) && !concurrent_collection_in_progress) {
2990 SGEN_LOG (8, "Skipping remset at %p", ptr);
2994 SGEN_LOG (8, "Adding remset at %p", ptr);
2996 remset.wbarrier_generic_nostore (ptr);
3000 * mono_gc_wbarrier_generic_store:
3003 mono_gc_wbarrier_generic_store (gpointer ptr, GCObject* value)
3005 SGEN_LOG (8, "Wbarrier store at %p to %p (%s)", ptr, value, value ? sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (value)) : "null");
3006 SGEN_UPDATE_REFERENCE_ALLOW_NULL (ptr, value);
3007 if (ptr_in_nursery (value) || concurrent_collection_in_progress)
3008 mono_gc_wbarrier_generic_nostore (ptr);
3009 sgen_dummy_use (value);
3013 * mono_gc_wbarrier_generic_store_atomic:
3014 * Same as \c mono_gc_wbarrier_generic_store but performs the store
3015 * as an atomic operation with release semantics.
3018 mono_gc_wbarrier_generic_store_atomic (gpointer ptr, GCObject *value)
3020 HEAVY_STAT (++stat_wbarrier_generic_store_atomic);
3022 SGEN_LOG (8, "Wbarrier atomic store at %p to %p (%s)", ptr, value, value ? sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (value)) : "null");
3024 InterlockedWritePointer ((volatile gpointer *)ptr, value);
3026 if (ptr_in_nursery (value) || concurrent_collection_in_progress)
3027 mono_gc_wbarrier_generic_nostore (ptr);
3029 sgen_dummy_use (value);
3033 sgen_wbarrier_range_copy (gpointer _dest, gpointer _src, int size)
3035 remset.wbarrier_range_copy (_dest,_src, size);
3039 * ######################################################################
3040 * ######## Other mono public interface functions.
3041 * ######################################################################
3045 sgen_gc_collect (int generation)
3050 sgen_perform_collection (0, generation, "user request", TRUE, TRUE);
3055 sgen_gc_collection_count (int generation)
3057 return InterlockedRead (generation == GENERATION_NURSERY ? &gc_stats.minor_gc_count : &gc_stats.major_gc_count);
3061 sgen_gc_get_used_size (void)
3065 tot = los_memory_usage;
3066 tot += nursery_section->end_data - nursery_section->data;
3067 tot += major_collector.get_used_size ();
3068 /* FIXME: account for pinned objects */
3074 sgen_env_var_error (const char *env_var, const char *fallback, const char *description_format, ...)
3078 va_start (ap, description_format);
3080 fprintf (stderr, "Warning: In environment variable `%s': ", env_var);
3081 vfprintf (stderr, description_format, ap);
3083 fprintf (stderr, " - %s", fallback);
3084 fprintf (stderr, "\n");
3090 parse_double_in_interval (const char *env_var, const char *opt_name, const char *opt, double min, double max, double *result)
3093 double val = strtod (opt, &endptr);
3094 if (endptr == opt) {
3095 sgen_env_var_error (env_var, "Using default value.", "`%s` must be a number.", opt_name);
3098 else if (val < min || val > max) {
3099 sgen_env_var_error (env_var, "Using default value.", "`%s` must be between %.2f - %.2f.", opt_name, min, max);
3107 parse_sgen_minor (const char *opt)
3110 return SGEN_MINOR_DEFAULT;
3112 if (!strcmp (opt, "simple")) {
3113 return SGEN_MINOR_SIMPLE;
3114 } else if (!strcmp (opt, "simple-par")) {
3115 return SGEN_MINOR_SIMPLE_PARALLEL;
3116 } else if (!strcmp (opt, "split")) {
3117 return SGEN_MINOR_SPLIT;
3119 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default instead.", "Unknown minor collector `%s'.", opt);
3120 return SGEN_MINOR_DEFAULT;
3125 parse_sgen_major (const char *opt)
3128 return SGEN_MAJOR_DEFAULT;
3130 if (!strcmp (opt, "marksweep")) {
3131 return SGEN_MAJOR_SERIAL;
3132 } else if (!strcmp (opt, "marksweep-conc")) {
3133 return SGEN_MAJOR_CONCURRENT;
3134 } else if (!strcmp (opt, "marksweep-conc-par")) {
3135 return SGEN_MAJOR_CONCURRENT_PARALLEL;
3137 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default instead.", "Unknown major collector `%s'.", opt);
3138 return SGEN_MAJOR_DEFAULT;
3144 parse_sgen_mode (const char *opt)
3147 return SGEN_MODE_NONE;
3149 if (!strcmp (opt, "balanced")) {
3150 return SGEN_MODE_BALANCED;
3151 } else if (!strcmp (opt, "throughput")) {
3152 return SGEN_MODE_THROUGHPUT;
3153 } else if (!strcmp (opt, "pause") || g_str_has_prefix (opt, "pause:")) {
3154 return SGEN_MODE_PAUSE;
3156 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default configurations.", "Unknown mode `%s'.", opt);
3157 return SGEN_MODE_NONE;
3162 init_sgen_minor (SgenMinor minor)
3165 case SGEN_MINOR_DEFAULT:
3166 case SGEN_MINOR_SIMPLE:
3167 sgen_simple_nursery_init (&sgen_minor_collector, FALSE);
3169 case SGEN_MINOR_SIMPLE_PARALLEL:
3170 sgen_simple_nursery_init (&sgen_minor_collector, TRUE);
3172 case SGEN_MINOR_SPLIT:
3173 sgen_split_nursery_init (&sgen_minor_collector);
3176 g_assert_not_reached ();
3181 init_sgen_major (SgenMajor major)
3183 if (major == SGEN_MAJOR_DEFAULT)
3184 major = DEFAULT_MAJOR;
3187 case SGEN_MAJOR_SERIAL:
3188 sgen_marksweep_init (&major_collector);
3190 case SGEN_MAJOR_CONCURRENT:
3191 sgen_marksweep_conc_init (&major_collector);
3193 case SGEN_MAJOR_CONCURRENT_PARALLEL:
3194 sgen_marksweep_conc_par_init (&major_collector);
3197 g_assert_not_reached ();
3202 * If sgen mode is set, major/minor configuration is fixed. The other gc_params
3203 * are parsed and processed after major/minor initialization, so it can potentially
3204 * override some knobs set by the sgen mode. We can consider locking out additional
3205 * configurations when gc_modes are used.
3208 init_sgen_mode (SgenMode mode)
3210 SgenMinor minor = SGEN_MINOR_DEFAULT;
3211 SgenMajor major = SGEN_MAJOR_DEFAULT;
3214 case SGEN_MODE_BALANCED:
3216 * Use a dynamic parallel nursery with a major concurrent collector.
3217 * This uses the default values for max pause time and nursery size.
3219 minor = SGEN_MINOR_SIMPLE;
3220 major = SGEN_MAJOR_CONCURRENT;
3221 dynamic_nursery = TRUE;
3223 case SGEN_MODE_THROUGHPUT:
3225 * Use concurrent major to let the mutator do more work. Use a larger
3226 * nursery, without pause time constraints, in order to collect more
3227 * objects in parallel and avoid repetitive collection tasks (pinning,
3228 * root scanning etc)
3230 minor = SGEN_MINOR_SIMPLE_PARALLEL;
3231 major = SGEN_MAJOR_CONCURRENT;
3232 dynamic_nursery = TRUE;
3233 sgen_max_pause_time = 0;
3235 case SGEN_MODE_PAUSE:
3237 * Use concurrent major and dynamic nursery with a more
3238 * aggressive shrinking relative to pause times.
3240 minor = SGEN_MINOR_SIMPLE_PARALLEL;
3241 major = SGEN_MAJOR_CONCURRENT;
3242 dynamic_nursery = TRUE;
3243 sgen_max_pause_margin = SGEN_PAUSE_MODE_MAX_PAUSE_MARGIN;
3246 g_assert_not_reached ();
3249 init_sgen_minor (minor);
3250 init_sgen_major (major);
3258 SgenMajor sgen_major = SGEN_MAJOR_DEFAULT;
3259 SgenMinor sgen_minor = SGEN_MINOR_DEFAULT;
3260 SgenMode sgen_mode = SGEN_MODE_NONE;
3261 char *params_opts = NULL;
3262 char *debug_opts = NULL;
3263 size_t max_heap = 0;
3264 size_t soft_limit = 0;
3266 gboolean debug_print_allowance = FALSE;
3267 double allowance_ratio = 0, save_target = 0;
3268 gboolean cement_enabled = TRUE;
3271 result = InterlockedCompareExchange (&gc_initialized, -1, 0);
3274 /* already inited */
3277 /* being inited by another thread */
3278 mono_thread_info_usleep (1000);
3281 /* we will init it */
3284 g_assert_not_reached ();
3286 } while (result != 0);
3288 SGEN_TV_GETTIME (sgen_init_timestamp);
3290 #ifdef SGEN_WITHOUT_MONO
3291 mono_thread_smr_init ();
3294 mono_coop_mutex_init (&gc_mutex);
3296 gc_debug_file = stderr;
3298 mono_coop_mutex_init (&sgen_interruption_mutex);
3300 if ((env = g_getenv (MONO_GC_PARAMS_NAME)) || gc_params_options) {
3301 params_opts = g_strdup_printf ("%s,%s", gc_params_options ? gc_params_options : "", env ? env : "");
3306 opts = g_strsplit (params_opts, ",", -1);
3307 for (ptr = opts; *ptr; ++ptr) {
3309 if (g_str_has_prefix (opt, "major=")) {
3310 opt = strchr (opt, '=') + 1;
3311 sgen_major = parse_sgen_major (opt);
3312 } else if (g_str_has_prefix (opt, "minor=")) {
3313 opt = strchr (opt, '=') + 1;
3314 sgen_minor = parse_sgen_minor (opt);
3315 } else if (g_str_has_prefix (opt, "mode=")) {
3316 opt = strchr (opt, '=') + 1;
3317 sgen_mode = parse_sgen_mode (opt);
3325 sgen_init_internal_allocator ();
3326 sgen_init_nursery_allocator ();
3327 sgen_init_fin_weak_hash ();
3328 sgen_init_hash_table ();
3329 sgen_init_descriptors ();
3330 sgen_init_gray_queues ();
3331 sgen_init_allocator ();
3332 sgen_init_gchandles ();
3334 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_SECTION, SGEN_SIZEOF_GC_MEM_SECTION);
3335 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_GRAY_QUEUE, sizeof (GrayQueueSection));
3337 sgen_client_init ();
3339 if (sgen_mode != SGEN_MODE_NONE) {
3340 if (sgen_minor != SGEN_MINOR_DEFAULT || sgen_major != SGEN_MAJOR_DEFAULT)
3341 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring major/minor configuration", "Major/minor configurations cannot be used with sgen modes");
3342 init_sgen_mode (sgen_mode);
3344 init_sgen_minor (sgen_minor);
3345 init_sgen_major (sgen_major);
3349 gboolean usage_printed = FALSE;
3351 for (ptr = opts; *ptr; ++ptr) {
3353 if (!strcmp (opt, ""))
3355 if (g_str_has_prefix (opt, "major="))
3357 if (g_str_has_prefix (opt, "minor="))
3359 if (g_str_has_prefix (opt, "mode=")) {
3360 if (g_str_has_prefix (opt, "mode=pause:")) {
3361 char *str_pause = strchr (opt, ':') + 1;
3362 int pause = atoi (str_pause);
3364 sgen_max_pause_time = pause;
3366 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default", "Invalid maximum pause time for `pause` sgen mode");
3370 if (g_str_has_prefix (opt, "max-heap-size=")) {
3371 size_t page_size = mono_pagesize ();
3372 size_t max_heap_candidate = 0;
3373 opt = strchr (opt, '=') + 1;
3374 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &max_heap_candidate)) {
3375 max_heap = (max_heap_candidate + page_size - 1) & ~(size_t)(page_size - 1);
3376 if (max_heap != max_heap_candidate)
3377 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Rounding up.", "`max-heap-size` size must be a multiple of %d.", page_size);
3379 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`max-heap-size` must be an integer.");
3383 if (g_str_has_prefix (opt, "soft-heap-limit=")) {
3384 opt = strchr (opt, '=') + 1;
3385 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &soft_limit)) {
3386 if (soft_limit <= 0) {
3387 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be positive.");
3391 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be an integer.");
3395 if (g_str_has_prefix (opt, "nursery-size=")) {
3397 opt = strchr (opt, '=') + 1;
3398 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &val)) {
3399 if ((val & (val - 1))) {
3400 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be a power of two.");
3404 if (val < SGEN_MAX_NURSERY_WASTE) {
3405 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.",
3406 "`nursery-size` must be at least %d bytes.", SGEN_MAX_NURSERY_WASTE);
3410 min_nursery_size = max_nursery_size = val;
3411 dynamic_nursery = FALSE;
3413 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be an integer.");
3418 if (g_str_has_prefix (opt, "save-target-ratio=")) {
3420 opt = strchr (opt, '=') + 1;
3421 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "save-target-ratio", opt,
3422 SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO, &val)) {
3427 if (g_str_has_prefix (opt, "default-allowance-ratio=")) {
3429 opt = strchr (opt, '=') + 1;
3430 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "default-allowance-ratio", opt,
3431 SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, SGEN_MAX_ALLOWANCE_NURSERY_SIZE_RATIO, &val)) {
3432 allowance_ratio = val;
3437 if (!strcmp (opt, "cementing")) {
3438 cement_enabled = TRUE;
3441 if (!strcmp (opt, "no-cementing")) {
3442 cement_enabled = FALSE;
3446 if (!strcmp (opt, "precleaning")) {
3447 precleaning_enabled = TRUE;
3450 if (!strcmp (opt, "no-precleaning")) {
3451 precleaning_enabled = FALSE;
3455 if (!strcmp (opt, "dynamic-nursery")) {
3456 if (sgen_minor_collector.is_split)
3457 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.",
3458 "dynamic-nursery not supported with split-nursery.");
3460 dynamic_nursery = TRUE;
3463 if (!strcmp (opt, "no-dynamic-nursery")) {
3464 dynamic_nursery = FALSE;
3468 if (major_collector.handle_gc_param && major_collector.handle_gc_param (opt))
3471 if (sgen_minor_collector.handle_gc_param && sgen_minor_collector.handle_gc_param (opt))
3474 if (sgen_client_handle_gc_param (opt))
3477 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring.", "Unknown option `%s`.", opt);
3482 fprintf (stderr, "\n%s must be a comma-delimited list of one or more of the following:\n", MONO_GC_PARAMS_NAME);
3483 fprintf (stderr, " max-heap-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
3484 fprintf (stderr, " soft-heap-limit=n (where N is an integer, possibly with a k, m or a g suffix)\n");
3485 fprintf (stderr, " mode=MODE (where MODE is 'balanced', 'throughput' or 'pause[:N]' and N is maximum pause in milliseconds)\n");
3486 fprintf (stderr, " nursery-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
3487 fprintf (stderr, " major=COLLECTOR (where COLLECTOR is `marksweep', `marksweep-conc', `marksweep-par')\n");
3488 fprintf (stderr, " minor=COLLECTOR (where COLLECTOR is `simple' or `split')\n");
3489 fprintf (stderr, " wbarrier=WBARRIER (where WBARRIER is `remset' or `cardtable')\n");
3490 fprintf (stderr, " [no-]cementing\n");
3491 fprintf (stderr, " [no-]dynamic-nursery\n");
3492 if (major_collector.print_gc_param_usage)
3493 major_collector.print_gc_param_usage ();
3494 if (sgen_minor_collector.print_gc_param_usage)
3495 sgen_minor_collector.print_gc_param_usage ();
3496 sgen_client_print_gc_params_usage ();
3497 fprintf (stderr, " Experimental options:\n");
3498 fprintf (stderr, " save-target-ratio=R (where R must be between %.2f - %.2f).\n", SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO);
3499 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);
3500 fprintf (stderr, "\n");
3502 usage_printed = TRUE;
3508 g_free (params_opts);
3510 alloc_nursery (dynamic_nursery, min_nursery_size, max_nursery_size);
3512 sgen_pinning_init ();
3513 sgen_cement_init (cement_enabled);
3515 if ((env = g_getenv (MONO_GC_DEBUG_NAME)) || gc_debug_options) {
3516 debug_opts = g_strdup_printf ("%s,%s", gc_debug_options ? gc_debug_options : "", env ? env : "");
3521 gboolean usage_printed = FALSE;
3523 opts = g_strsplit (debug_opts, ",", -1);
3524 for (ptr = opts; ptr && *ptr; ptr ++) {
3526 if (!strcmp (opt, ""))
3528 if (opt [0] >= '0' && opt [0] <= '9') {
3529 gc_debug_level = atoi (opt);
3534 char *rf = g_strdup_printf ("%s.%d", opt, mono_process_current_pid ());
3535 gc_debug_file = fopen (rf, "wb");
3537 gc_debug_file = stderr;
3540 } else if (!strcmp (opt, "print-allowance")) {
3541 debug_print_allowance = TRUE;
3542 } else if (!strcmp (opt, "print-pinning")) {
3543 sgen_pin_stats_enable ();
3544 } else if (!strcmp (opt, "verify-before-allocs")) {
3545 verify_before_allocs = 1;
3546 has_per_allocation_action = TRUE;
3547 } else if (g_str_has_prefix (opt, "max-valloc-size=")) {
3548 size_t max_valloc_size;
3549 char *arg = strchr (opt, '=') + 1;
3550 if (*opt && mono_gc_parse_environment_string_extract_number (arg, &max_valloc_size)) {
3551 mono_valloc_set_limit (max_valloc_size);
3553 sgen_env_var_error (MONO_GC_DEBUG_NAME, NULL, "`max-valloc-size` must be an integer.");
3556 } else if (g_str_has_prefix (opt, "verify-before-allocs=")) {
3557 char *arg = strchr (opt, '=') + 1;
3558 verify_before_allocs = atoi (arg);
3559 has_per_allocation_action = TRUE;
3560 } else if (!strcmp (opt, "collect-before-allocs")) {
3561 collect_before_allocs = 1;
3562 has_per_allocation_action = TRUE;
3563 } else if (g_str_has_prefix (opt, "collect-before-allocs=")) {
3564 char *arg = strchr (opt, '=') + 1;
3565 has_per_allocation_action = TRUE;
3566 collect_before_allocs = atoi (arg);
3567 } else if (!strcmp (opt, "verify-before-collections")) {
3568 whole_heap_check_before_collection = TRUE;
3569 } else if (!strcmp (opt, "check-remset-consistency")) {
3570 remset_consistency_checks = TRUE;
3571 nursery_clear_policy = CLEAR_AT_GC;
3572 } else if (!strcmp (opt, "mod-union-consistency-check")) {
3573 if (!major_collector.is_concurrent) {
3574 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`mod-union-consistency-check` only works with concurrent major collector.");
3577 mod_union_consistency_check = TRUE;
3578 } else if (!strcmp (opt, "check-mark-bits")) {
3579 check_mark_bits_after_major_collection = TRUE;
3580 } else if (!strcmp (opt, "check-nursery-pinned")) {
3581 check_nursery_objects_pinned = TRUE;
3582 } else if (!strcmp (opt, "clear-at-gc")) {
3583 nursery_clear_policy = CLEAR_AT_GC;
3584 } else if (!strcmp (opt, "clear-nursery-at-gc")) {
3585 nursery_clear_policy = CLEAR_AT_GC;
3586 } else if (!strcmp (opt, "clear-at-tlab-creation")) {
3587 nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
3588 } else if (!strcmp (opt, "debug-clear-at-tlab-creation")) {
3589 nursery_clear_policy = CLEAR_AT_TLAB_CREATION_DEBUG;
3590 } else if (!strcmp (opt, "check-scan-starts")) {
3591 do_scan_starts_check = TRUE;
3592 } else if (!strcmp (opt, "verify-nursery-at-minor-gc")) {
3593 do_verify_nursery = TRUE;
3594 } else if (!strcmp (opt, "check-concurrent")) {
3595 if (!major_collector.is_concurrent) {
3596 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`check-concurrent` only works with concurrent major collectors.");
3599 nursery_clear_policy = CLEAR_AT_GC;
3600 do_concurrent_checks = TRUE;
3601 } else if (!strcmp (opt, "dump-nursery-at-minor-gc")) {
3602 do_dump_nursery_content = TRUE;
3603 } else if (!strcmp (opt, "disable-minor")) {
3604 disable_minor_collections = TRUE;
3605 } else if (!strcmp (opt, "disable-major")) {
3606 disable_major_collections = TRUE;
3607 } else if (g_str_has_prefix (opt, "heap-dump=")) {
3608 char *filename = strchr (opt, '=') + 1;
3609 nursery_clear_policy = CLEAR_AT_GC;
3610 sgen_debug_enable_heap_dump (filename);
3611 } else if (g_str_has_prefix (opt, "binary-protocol=")) {
3612 char *filename = strchr (opt, '=') + 1;
3613 char *colon = strrchr (filename, ':');
3616 if (!mono_gc_parse_environment_string_extract_number (colon + 1, &limit)) {
3617 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring limit.", "Binary protocol file size limit must be an integer.");
3622 binary_protocol_init (filename, (long long)limit);
3623 } else if (!strcmp (opt, "nursery-canaries")) {
3624 do_verify_nursery = TRUE;
3625 enable_nursery_canaries = TRUE;
3626 } else if (!sgen_client_handle_gc_debug (opt)) {
3627 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "Unknown option `%s`.", opt);
3632 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);
3633 fprintf (stderr, "Valid <option>s are:\n");
3634 fprintf (stderr, " collect-before-allocs[=<n>]\n");
3635 fprintf (stderr, " verify-before-allocs[=<n>]\n");
3636 fprintf (stderr, " max-valloc-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
3637 fprintf (stderr, " check-remset-consistency\n");
3638 fprintf (stderr, " check-mark-bits\n");
3639 fprintf (stderr, " check-nursery-pinned\n");
3640 fprintf (stderr, " verify-before-collections\n");
3641 fprintf (stderr, " verify-nursery-at-minor-gc\n");
3642 fprintf (stderr, " dump-nursery-at-minor-gc\n");
3643 fprintf (stderr, " disable-minor\n");
3644 fprintf (stderr, " disable-major\n");
3645 fprintf (stderr, " check-concurrent\n");
3646 fprintf (stderr, " clear-[nursery-]at-gc\n");
3647 fprintf (stderr, " clear-at-tlab-creation\n");
3648 fprintf (stderr, " debug-clear-at-tlab-creation\n");
3649 fprintf (stderr, " check-scan-starts\n");
3650 fprintf (stderr, " print-allowance\n");
3651 fprintf (stderr, " print-pinning\n");
3652 fprintf (stderr, " heap-dump=<filename>\n");
3653 fprintf (stderr, " binary-protocol=<filename>[:<file-size-limit>]\n");
3654 fprintf (stderr, " nursery-canaries\n");
3655 sgen_client_print_gc_debug_usage ();
3656 fprintf (stderr, "\n");
3658 usage_printed = TRUE;
3665 g_free (debug_opts);
3667 if (check_mark_bits_after_major_collection)
3668 nursery_clear_policy = CLEAR_AT_GC;
3670 if (major_collector.post_param_init)
3671 major_collector.post_param_init (&major_collector);
3673 sgen_thread_pool_start ();
3675 sgen_memgov_init (max_heap, soft_limit, debug_print_allowance, allowance_ratio, save_target);
3677 memset (&remset, 0, sizeof (remset));
3679 sgen_card_table_init (&remset);
3681 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");
3685 sgen_init_bridge ();
3689 sgen_gc_initialized ()
3691 return gc_initialized > 0;
3695 sgen_get_nursery_clear_policy (void)
3697 return nursery_clear_policy;
3703 mono_coop_mutex_lock (&gc_mutex);
3707 sgen_gc_unlock (void)
3709 mono_coop_mutex_unlock (&gc_mutex);
3713 sgen_major_collector_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
3715 major_collector.iterate_live_block_ranges (callback);
3719 sgen_major_collector_iterate_block_ranges (sgen_cardtable_block_callback callback)
3721 major_collector.iterate_block_ranges (callback);
3725 sgen_get_major_collector (void)
3727 return &major_collector;
3731 sgen_get_minor_collector (void)
3733 return &sgen_minor_collector;
3737 sgen_get_remset (void)
3743 count_cards (long long *major_total, long long *major_marked, long long *los_total, long long *los_marked)
3745 sgen_get_major_collector ()->count_cards (major_total, major_marked);
3746 sgen_los_count_cards (los_total, los_marked);
3749 static gboolean world_is_stopped = FALSE;
3751 /* LOCKING: assumes the GC lock is held */
3753 sgen_stop_world (int generation)
3755 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3757 SGEN_ASSERT (0, !world_is_stopped, "Why are we stopping a stopped world?");
3759 binary_protocol_world_stopping (generation, sgen_timestamp (), (gpointer) (gsize) mono_native_thread_id_get ());
3761 sgen_client_stop_world (generation);
3763 world_is_stopped = TRUE;
3765 if (binary_protocol_is_heavy_enabled ())
3766 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3767 binary_protocol_world_stopped (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3770 /* LOCKING: assumes the GC lock is held */
3772 sgen_restart_world (int generation)
3774 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3777 SGEN_ASSERT (0, world_is_stopped, "Why are we restarting a running world?");
3779 if (binary_protocol_is_heavy_enabled ())
3780 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3781 binary_protocol_world_restarting (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3783 world_is_stopped = FALSE;
3785 sgen_client_restart_world (generation, &stw_time);
3787 binary_protocol_world_restarted (generation, sgen_timestamp ());
3789 if (sgen_client_bridge_need_processing ())
3790 sgen_client_bridge_processing_finish (generation);
3792 sgen_memgov_collection_end (generation, stw_time);
3796 sgen_is_world_stopped (void)
3798 return world_is_stopped;
3802 sgen_check_whole_heap_stw (void)
3804 sgen_stop_world (0);
3805 sgen_clear_nursery_fragments ();
3806 sgen_check_whole_heap (TRUE);
3807 sgen_restart_world (0);
3811 sgen_timestamp (void)
3813 SGEN_TV_DECLARE (timestamp);
3814 SGEN_TV_GETTIME (timestamp);
3815 return SGEN_TV_ELAPSED (sgen_init_timestamp, timestamp);
3818 #endif /* HAVE_SGEN_GC */