2 * sgen-gc.c: Simple generational GC.
5 * Paolo Molaro (lupus@ximian.com)
6 * Rodrigo Kumpera (kumpera@gmail.com)
8 * Copyright 2005-2011 Novell, Inc (http://www.novell.com)
9 * Copyright 2011 Xamarin Inc (http://www.xamarin.com)
11 * Thread start/stop adapted from Boehm's GC:
12 * Copyright (c) 1994 by Xerox Corporation. All rights reserved.
13 * Copyright (c) 1996 by Silicon Graphics. All rights reserved.
14 * Copyright (c) 1998 by Fergus Henderson. All rights reserved.
15 * Copyright (c) 2000-2004 by Hewlett-Packard Company. All rights reserved.
16 * Copyright 2001-2003 Ximian, Inc
17 * Copyright 2003-2010 Novell, Inc.
18 * Copyright 2011 Xamarin, Inc.
19 * Copyright (C) 2012 Xamarin Inc
21 * This library is free software; you can redistribute it and/or
22 * modify it under the terms of the GNU Library General Public
23 * License 2.0 as published by the Free Software Foundation;
25 * This library is distributed in the hope that it will be useful,
26 * but WITHOUT ANY WARRANTY; without even the implied warranty of
27 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
28 * Library General Public License for more details.
30 * You should have received a copy of the GNU Library General Public
31 * License 2.0 along with this library; if not, write to the Free
32 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 * Important: allocation provides always zeroed memory, having to do
35 * a memset after allocation is deadly for performance.
36 * Memory usage at startup is currently as follows:
38 * 64 KB internal space
40 * We should provide a small memory config with half the sizes
42 * We currently try to make as few mono assumptions as possible:
43 * 1) 2-word header with no GC pointers in it (first vtable, second to store the
45 * 2) gc descriptor is the second word in the vtable (first word in the class)
46 * 3) 8 byte alignment is the minimum and enough (not true for special structures (SIMD), FIXME)
47 * 4) there is a function to get an object's size and the number of
48 * elements in an array.
49 * 5) we know the special way bounds are allocated for complex arrays
50 * 6) we know about proxies and how to treat them when domains are unloaded
52 * Always try to keep stack usage to a minimum: no recursive behaviour
53 * and no large stack allocs.
55 * General description.
56 * Objects are initially allocated in a nursery using a fast bump-pointer technique.
57 * When the nursery is full we start a nursery collection: this is performed with a
59 * When the old generation is full we start a copying GC of the old generation as well:
60 * this will be changed to mark&sweep with copying when fragmentation becomes to severe
61 * in the future. Maybe we'll even do both during the same collection like IMMIX.
63 * The things that complicate this description are:
64 * *) pinned objects: we can't move them so we need to keep track of them
65 * *) no precise info of the thread stacks and registers: we need to be able to
66 * quickly find the objects that may be referenced conservatively and pin them
67 * (this makes the first issues more important)
68 * *) large objects are too expensive to be dealt with using copying GC: we handle them
69 * with mark/sweep during major collections
70 * *) some objects need to not move even if they are small (interned strings, Type handles):
71 * we use mark/sweep for them, too: they are not allocated in the nursery, but inside
72 * PinnedChunks regions
78 *) we could have a function pointer in MonoClass to implement
79 customized write barriers for value types
81 *) investigate the stuff needed to advance a thread to a GC-safe
82 point (single-stepping, read from unmapped memory etc) and implement it.
83 This would enable us to inline allocations and write barriers, for example,
84 or at least parts of them, like the write barrier checks.
85 We may need this also for handling precise info on stacks, even simple things
86 as having uninitialized data on the stack and having to wait for the prolog
87 to zero it. Not an issue for the last frame that we scan conservatively.
88 We could always not trust the value in the slots anyway.
90 *) modify the jit to save info about references in stack locations:
91 this can be done just for locals as a start, so that at least
92 part of the stack is handled precisely.
94 *) test/fix endianess issues
96 *) Implement a card table as the write barrier instead of remembered
97 sets? Card tables are not easy to implement with our current
98 memory layout. We have several different kinds of major heap
99 objects: Small objects in regular blocks, small objects in pinned
100 chunks and LOS objects. If we just have a pointer we have no way
101 to tell which kind of object it points into, therefore we cannot
102 know where its card table is. The least we have to do to make
103 this happen is to get rid of write barriers for indirect stores.
106 *) Get rid of write barriers for indirect stores. We can do this by
107 telling the GC to wbarrier-register an object once we do an ldloca
108 or ldelema on it, and to unregister it once it's not used anymore
109 (it can only travel downwards on the stack). The problem with
110 unregistering is that it needs to happen eventually no matter
111 what, even if exceptions are thrown, the thread aborts, etc.
112 Rodrigo suggested that we could do only the registering part and
113 let the collector find out (pessimistically) when it's safe to
114 unregister, namely when the stack pointer of the thread that
115 registered the object is higher than it was when the registering
116 happened. This might make for a good first implementation to get
117 some data on performance.
119 *) Some sort of blacklist support? Blacklists is a concept from the
120 Boehm GC: if during a conservative scan we find pointers to an
121 area which we might use as heap, we mark that area as unusable, so
122 pointer retention by random pinning pointers is reduced.
124 *) experiment with max small object size (very small right now - 2kb,
125 because it's tied to the max freelist size)
127 *) add an option to mmap the whole heap in one chunk: it makes for many
128 simplifications in the checks (put the nursery at the top and just use a single
129 check for inclusion/exclusion): the issue this has is that on 32 bit systems it's
130 not flexible (too much of the address space may be used by default or we can't
131 increase the heap as needed) and we'd need a race-free mechanism to return memory
132 back to the system (mprotect(PROT_NONE) will still keep the memory allocated if it
133 was written to, munmap is needed, but the following mmap may not find the same segment
136 *) memzero the major fragments after restarting the world and optionally a smaller
139 *) investigate having fragment zeroing threads
141 *) separate locks for finalization and other minor stuff to reduce
144 *) try a different copying order to improve memory locality
146 *) a thread abort after a store but before the write barrier will
147 prevent the write barrier from executing
149 *) specialized dynamically generated markers/copiers
151 *) Dynamically adjust TLAB size to the number of threads. If we have
152 too many threads that do allocation, we might need smaller TLABs,
153 and we might get better performance with larger TLABs if we only
154 have a handful of threads. We could sum up the space left in all
155 assigned TLABs and if that's more than some percentage of the
156 nursery size, reduce the TLAB size.
158 *) Explore placing unreachable objects on unused nursery memory.
159 Instead of memset'ng a region to zero, place an int[] covering it.
160 A good place to start is add_nursery_frag. The tricky thing here is
161 placing those objects atomically outside of a collection.
163 *) Allocation should use asymmetric Dekker synchronization:
164 http://blogs.oracle.com/dave/resource/Asymmetric-Dekker-Synchronization.txt
165 This should help weak consistency archs.
172 #define _XOPEN_SOURCE
173 #define _DARWIN_C_SOURCE
179 #ifdef HAVE_PTHREAD_H
182 #ifdef HAVE_PTHREAD_NP_H
183 #include <pthread_np.h>
191 #include "mono/sgen/sgen-gc.h"
192 #include "mono/sgen/sgen-cardtable.h"
193 #include "mono/sgen/sgen-protocol.h"
194 #include "mono/sgen/sgen-memory-governor.h"
195 #include "mono/sgen/sgen-hash-table.h"
196 #include "mono/sgen/sgen-cardtable.h"
197 #include "mono/sgen/sgen-pinning.h"
198 #include "mono/sgen/sgen-workers.h"
199 #include "mono/sgen/sgen-client.h"
200 #include "mono/sgen/sgen-pointer-queue.h"
201 #include "mono/sgen/gc-internal-agnostic.h"
202 #include "mono/utils/mono-proclib.h"
203 #include "mono/utils/mono-memory-model.h"
204 #include "mono/utils/hazard-pointer.h"
206 #include <mono/utils/memcheck.h>
208 #undef pthread_create
210 #undef pthread_detach
213 * ######################################################################
214 * ######## Types and constants used by the GC.
215 * ######################################################################
218 /* 0 means not initialized, 1 is initialized, -1 means in progress */
219 static int gc_initialized = 0;
220 /* If set, check if we need to do something every X allocations */
221 gboolean has_per_allocation_action;
222 /* If set, do a heap check every X allocation */
223 guint32 verify_before_allocs = 0;
224 /* If set, do a minor collection before every X allocation */
225 guint32 collect_before_allocs = 0;
226 /* If set, do a whole heap check before each collection */
227 static gboolean whole_heap_check_before_collection = FALSE;
228 /* If set, do a heap consistency check before each minor collection */
229 static gboolean consistency_check_at_minor_collection = FALSE;
230 /* If set, do a mod union consistency check before each finishing collection pause */
231 static gboolean mod_union_consistency_check = FALSE;
232 /* If set, check whether mark bits are consistent after major collections */
233 static gboolean check_mark_bits_after_major_collection = FALSE;
234 /* If set, check that all nursery objects are pinned/not pinned, depending on context */
235 static gboolean check_nursery_objects_pinned = FALSE;
236 /* If set, do a few checks when the concurrent collector is used */
237 static gboolean do_concurrent_checks = FALSE;
238 /* If set, do a plausibility check on the scan_starts before and after
240 static gboolean do_scan_starts_check = FALSE;
243 * If the major collector is concurrent and this is FALSE, we will
244 * never initiate a synchronous major collection, unless requested via
247 static gboolean allow_synchronous_major = TRUE;
248 static gboolean disable_minor_collections = FALSE;
249 static gboolean disable_major_collections = FALSE;
250 static gboolean do_verify_nursery = FALSE;
251 static gboolean do_dump_nursery_content = FALSE;
252 static gboolean enable_nursery_canaries = FALSE;
254 #ifdef HEAVY_STATISTICS
255 guint64 stat_objects_alloced_degraded = 0;
256 guint64 stat_bytes_alloced_degraded = 0;
258 guint64 stat_copy_object_called_nursery = 0;
259 guint64 stat_objects_copied_nursery = 0;
260 guint64 stat_copy_object_called_major = 0;
261 guint64 stat_objects_copied_major = 0;
263 guint64 stat_scan_object_called_nursery = 0;
264 guint64 stat_scan_object_called_major = 0;
266 guint64 stat_slots_allocated_in_vain;
268 guint64 stat_nursery_copy_object_failed_from_space = 0;
269 guint64 stat_nursery_copy_object_failed_forwarded = 0;
270 guint64 stat_nursery_copy_object_failed_pinned = 0;
271 guint64 stat_nursery_copy_object_failed_to_space = 0;
273 static guint64 stat_wbarrier_add_to_global_remset = 0;
274 static guint64 stat_wbarrier_arrayref_copy = 0;
275 static guint64 stat_wbarrier_generic_store = 0;
276 static guint64 stat_wbarrier_generic_store_atomic = 0;
277 static guint64 stat_wbarrier_set_root = 0;
280 static guint64 stat_pinned_objects = 0;
282 static guint64 time_minor_pre_collection_fragment_clear = 0;
283 static guint64 time_minor_pinning = 0;
284 static guint64 time_minor_scan_remsets = 0;
285 static guint64 time_minor_scan_pinned = 0;
286 static guint64 time_minor_scan_roots = 0;
287 static guint64 time_minor_finish_gray_stack = 0;
288 static guint64 time_minor_fragment_creation = 0;
290 static guint64 time_major_pre_collection_fragment_clear = 0;
291 static guint64 time_major_pinning = 0;
292 static guint64 time_major_scan_pinned = 0;
293 static guint64 time_major_scan_roots = 0;
294 static guint64 time_major_scan_mod_union = 0;
295 static guint64 time_major_finish_gray_stack = 0;
296 static guint64 time_major_free_bigobjs = 0;
297 static guint64 time_major_los_sweep = 0;
298 static guint64 time_major_sweep = 0;
299 static guint64 time_major_fragment_creation = 0;
301 static guint64 time_max = 0;
303 static SGEN_TV_DECLARE (time_major_conc_collection_start);
304 static SGEN_TV_DECLARE (time_major_conc_collection_end);
306 static SGEN_TV_DECLARE (last_minor_collection_start_tv);
307 static SGEN_TV_DECLARE (last_minor_collection_end_tv);
309 int gc_debug_level = 0;
314 mono_gc_flush_info (void)
316 fflush (gc_debug_file);
320 #define TV_DECLARE SGEN_TV_DECLARE
321 #define TV_GETTIME SGEN_TV_GETTIME
322 #define TV_ELAPSED SGEN_TV_ELAPSED
324 static SGEN_TV_DECLARE (sgen_init_timestamp);
326 NurseryClearPolicy nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
328 #define object_is_forwarded SGEN_OBJECT_IS_FORWARDED
329 #define object_is_pinned SGEN_OBJECT_IS_PINNED
330 #define pin_object SGEN_PIN_OBJECT
332 #define ptr_in_nursery sgen_ptr_in_nursery
334 #define LOAD_VTABLE SGEN_LOAD_VTABLE
337 nursery_canaries_enabled (void)
339 return enable_nursery_canaries;
342 #define safe_object_get_size sgen_safe_object_get_size
345 * ######################################################################
346 * ######## Global data.
347 * ######################################################################
349 MonoCoopMutex gc_mutex;
350 gboolean sgen_try_free_some_memory;
352 #define SCAN_START_SIZE SGEN_SCAN_START_SIZE
354 size_t degraded_mode = 0;
356 static mword bytes_pinned_from_failed_allocation = 0;
358 GCMemSection *nursery_section = NULL;
359 static volatile mword lowest_heap_address = ~(mword)0;
360 static volatile mword highest_heap_address = 0;
362 MonoCoopMutex sgen_interruption_mutex;
364 int current_collection_generation = -1;
365 static volatile gboolean concurrent_collection_in_progress = FALSE;
367 /* objects that are ready to be finalized */
368 static SgenPointerQueue fin_ready_queue = SGEN_POINTER_QUEUE_INIT (INTERNAL_MEM_FINALIZE_READY);
369 static SgenPointerQueue critical_fin_queue = SGEN_POINTER_QUEUE_INIT (INTERNAL_MEM_FINALIZE_READY);
371 /* registered roots: the key to the hash is the root start address */
373 * Different kinds of roots are kept separate to speed up pin_from_roots () for example.
375 SgenHashTable roots_hash [ROOT_TYPE_NUM] = {
376 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL),
377 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL),
378 SGEN_HASH_TABLE_INIT (INTERNAL_MEM_ROOTS_TABLE, INTERNAL_MEM_ROOT_RECORD, sizeof (RootRecord), sgen_aligned_addr_hash, NULL)
380 static mword roots_size = 0; /* amount of memory in the root set */
382 /* The size of a TLAB */
383 /* The bigger the value, the less often we have to go to the slow path to allocate a new
384 * one, but the more space is wasted by threads not allocating much memory.
386 * FIXME: Make this self-tuning for each thread.
388 guint32 tlab_size = (1024 * 4);
390 #define MAX_SMALL_OBJ_SIZE SGEN_MAX_SMALL_OBJ_SIZE
392 #define ALLOC_ALIGN SGEN_ALLOC_ALIGN
394 #define ALIGN_UP SGEN_ALIGN_UP
396 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
397 MonoNativeThreadId main_gc_thread = NULL;
400 /*Object was pinned during the current collection*/
401 static mword objects_pinned;
404 * ######################################################################
405 * ######## Macros and function declarations.
406 * ######################################################################
409 typedef SgenGrayQueue GrayQueue;
411 /* forward declarations */
412 static void scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx);
414 static void pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx);
415 static void finish_gray_stack (int generation, ScanCopyContext ctx);
418 SgenMajorCollector major_collector;
419 SgenMinorCollector sgen_minor_collector;
420 /* FIXME: get rid of this */
421 static GrayQueue gray_queue;
423 static SgenRememberedSet remset;
425 /* The gray queue to use from the main collection thread. */
426 #define WORKERS_DISTRIBUTE_GRAY_QUEUE (&gray_queue)
429 * The gray queue a worker job must use. If we're not parallel or
430 * concurrent, we use the main gray queue.
432 static SgenGrayQueue*
433 sgen_workers_get_job_gray_queue (WorkerData *worker_data)
435 return worker_data ? &worker_data->private_gray_queue : WORKERS_DISTRIBUTE_GRAY_QUEUE;
439 gray_queue_redirect (SgenGrayQueue *queue)
441 gboolean wake = FALSE;
444 GrayQueueSection *section = sgen_gray_object_dequeue_section (queue);
447 sgen_section_gray_queue_enqueue (queue->alloc_prepare_data, section);
452 g_assert (concurrent_collection_in_progress);
453 sgen_workers_ensure_awake ();
458 gray_queue_enable_redirect (SgenGrayQueue *queue)
460 if (!concurrent_collection_in_progress)
463 sgen_gray_queue_set_alloc_prepare (queue, gray_queue_redirect, sgen_workers_get_distribute_section_gray_queue ());
464 gray_queue_redirect (queue);
468 sgen_scan_area_with_callback (char *start, char *end, IterateObjectCallbackFunc callback, void *data, gboolean allow_flags, gboolean fail_on_canaries)
470 while (start < end) {
474 if (!*(void**)start) {
475 start += sizeof (void*); /* should be ALLOC_ALIGN, really */
480 if (!(obj = SGEN_OBJECT_IS_FORWARDED (start)))
486 if (!sgen_client_object_is_array_fill ((GCObject*)obj)) {
487 CHECK_CANARY_FOR_OBJECT ((GCObject*)obj, fail_on_canaries);
488 size = ALIGN_UP (safe_object_get_size ((GCObject*)obj));
489 callback ((GCObject*)obj, size, data);
490 CANARIFY_SIZE (size);
492 size = ALIGN_UP (safe_object_get_size ((GCObject*)obj));
500 * sgen_add_to_global_remset:
502 * The global remset contains locations which point into newspace after
503 * a minor collection. This can happen if the objects they point to are pinned.
505 * LOCKING: If called from a parallel collector, the global remset
506 * lock must be held. For serial collectors that is not necessary.
509 sgen_add_to_global_remset (gpointer ptr, gpointer obj)
511 SGEN_ASSERT (5, sgen_ptr_in_nursery (obj), "Target pointer of global remset must be in the nursery");
513 HEAVY_STAT (++stat_wbarrier_add_to_global_remset);
515 if (!major_collector.is_concurrent) {
516 SGEN_ASSERT (5, current_collection_generation != -1, "Global remsets can only be added during collections");
518 if (current_collection_generation == -1)
519 SGEN_ASSERT (5, sgen_concurrent_collection_in_progress (), "Global remsets outside of collection pauses can only be added by the concurrent collector");
522 if (!object_is_pinned (obj))
523 SGEN_ASSERT (5, sgen_minor_collector.is_split || sgen_concurrent_collection_in_progress (), "Non-pinned objects can only remain in nursery if it is a split nursery");
524 else if (sgen_cement_lookup_or_register (obj))
527 remset.record_pointer (ptr);
529 sgen_pin_stats_register_global_remset (obj);
531 SGEN_LOG (8, "Adding global remset for %p", ptr);
532 binary_protocol_global_remset (ptr, obj, (gpointer)SGEN_LOAD_VTABLE (obj));
536 * sgen_drain_gray_stack:
538 * Scan objects in the gray stack until the stack is empty. This should be called
539 * frequently after each object is copied, to achieve better locality and cache
544 sgen_drain_gray_stack (ScanCopyContext ctx)
546 ScanObjectFunc scan_func = ctx.ops->scan_object;
547 GrayQueue *queue = ctx.queue;
549 if (ctx.ops->drain_gray_stack)
550 return ctx.ops->drain_gray_stack (queue);
555 GRAY_OBJECT_DEQUEUE (queue, &obj, &desc);
558 SGEN_LOG (9, "Precise gray object scan %p (%s)", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
559 scan_func (obj, desc, queue);
565 * Addresses in the pin queue are already sorted. This function finds
566 * the object header for each address and pins the object. The
567 * addresses must be inside the nursery section. The (start of the)
568 * address array is overwritten with the addresses of the actually
569 * pinned objects. Return the number of pinned objects.
572 pin_objects_from_nursery_pin_queue (gboolean do_scan_objects, ScanCopyContext ctx)
574 GCMemSection *section = nursery_section;
575 void **start = sgen_pinning_get_entry (section->pin_queue_first_entry);
576 void **end = sgen_pinning_get_entry (section->pin_queue_last_entry);
577 void *start_nursery = section->data;
578 void *end_nursery = section->next_data;
583 void *pinning_front = start_nursery;
585 void **definitely_pinned = start;
586 ScanObjectFunc scan_func = ctx.ops->scan_object;
587 SgenGrayQueue *queue = ctx.queue;
589 sgen_nursery_allocator_prepare_for_pinning ();
591 while (start < end) {
592 GCObject *obj_to_pin = NULL;
593 size_t obj_to_pin_size = 0;
598 SGEN_ASSERT (0, addr >= start_nursery && addr < end_nursery, "Potential pinning address out of range");
599 SGEN_ASSERT (0, addr >= last, "Pin queue not sorted");
606 SGEN_LOG (5, "Considering pinning addr %p", addr);
607 /* We've already processed everything up to pinning_front. */
608 if (addr < pinning_front) {
614 * Find the closest scan start <= addr. We might search backward in the
615 * scan_starts array because entries might be NULL. In the worst case we
616 * start at start_nursery.
618 idx = ((char*)addr - (char*)section->data) / SCAN_START_SIZE;
619 SGEN_ASSERT (0, idx < section->num_scan_start, "Scan start index out of range");
620 search_start = (void*)section->scan_starts [idx];
621 if (!search_start || search_start > addr) {
624 search_start = section->scan_starts [idx];
625 if (search_start && search_start <= addr)
628 if (!search_start || search_start > addr)
629 search_start = start_nursery;
633 * If the pinning front is closer than the scan start we found, start
634 * searching at the front.
636 if (search_start < pinning_front)
637 search_start = pinning_front;
640 * Now addr should be in an object a short distance from search_start.
642 * search_start must point to zeroed mem or point to an object.
645 size_t obj_size, canarified_obj_size;
648 if (!*(void**)search_start) {
649 search_start = (void*)ALIGN_UP ((mword)search_start + sizeof (gpointer));
650 /* The loop condition makes sure we don't overrun addr. */
654 canarified_obj_size = obj_size = ALIGN_UP (safe_object_get_size ((GCObject*)search_start));
657 * Filler arrays are marked by an invalid sync word. We don't
658 * consider them for pinning. They are not delimited by canaries,
661 if (!sgen_client_object_is_array_fill ((GCObject*)search_start)) {
662 CHECK_CANARY_FOR_OBJECT (search_start, TRUE);
663 CANARIFY_SIZE (canarified_obj_size);
665 if (addr >= search_start && (char*)addr < (char*)search_start + obj_size) {
666 /* This is the object we're looking for. */
667 obj_to_pin = (GCObject*)search_start;
668 obj_to_pin_size = canarified_obj_size;
673 /* Skip to the next object */
674 search_start = (void*)((char*)search_start + canarified_obj_size);
675 } while (search_start <= addr);
677 /* We've searched past the address we were looking for. */
679 pinning_front = search_start;
680 goto next_pin_queue_entry;
684 * We've found an object to pin. It might still be a dummy array, but we
685 * can advance the pinning front in any case.
687 pinning_front = (char*)obj_to_pin + obj_to_pin_size;
690 * If this is a dummy array marking the beginning of a nursery
691 * fragment, we don't pin it.
693 if (sgen_client_object_is_array_fill (obj_to_pin))
694 goto next_pin_queue_entry;
697 * Finally - pin the object!
699 desc = sgen_obj_get_descriptor_safe (obj_to_pin);
700 if (do_scan_objects) {
701 scan_func (obj_to_pin, desc, queue);
703 SGEN_LOG (4, "Pinned object %p, vtable %p (%s), count %d\n",
704 obj_to_pin, *(void**)obj_to_pin, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj_to_pin)), count);
705 binary_protocol_pin (obj_to_pin,
706 (gpointer)LOAD_VTABLE (obj_to_pin),
707 safe_object_get_size (obj_to_pin));
709 pin_object (obj_to_pin);
710 GRAY_OBJECT_ENQUEUE (queue, obj_to_pin, desc);
711 sgen_pin_stats_register_object (obj_to_pin, obj_to_pin_size);
712 definitely_pinned [count] = obj_to_pin;
716 next_pin_queue_entry:
720 sgen_client_nursery_objects_pinned (definitely_pinned, count);
721 stat_pinned_objects += count;
726 pin_objects_in_nursery (gboolean do_scan_objects, ScanCopyContext ctx)
730 if (nursery_section->pin_queue_first_entry == nursery_section->pin_queue_last_entry)
733 reduced_to = pin_objects_from_nursery_pin_queue (do_scan_objects, ctx);
734 nursery_section->pin_queue_last_entry = nursery_section->pin_queue_first_entry + reduced_to;
738 * This function is only ever called (via `collector_pin_object()` in `sgen-copy-object.h`)
739 * when we can't promote an object because we're out of memory.
742 sgen_pin_object (GCObject *object, GrayQueue *queue)
745 * All pinned objects are assumed to have been staged, so we need to stage as well.
746 * Also, the count of staged objects shows that "late pinning" happened.
748 sgen_pin_stage_ptr (object);
750 SGEN_PIN_OBJECT (object);
751 binary_protocol_pin (object, (gpointer)LOAD_VTABLE (object), safe_object_get_size (object));
754 sgen_pin_stats_register_object (object, safe_object_get_size (object));
756 GRAY_OBJECT_ENQUEUE (queue, object, sgen_obj_get_descriptor_safe (object));
759 /* Sort the addresses in array in increasing order.
760 * Done using a by-the book heap sort. Which has decent and stable performance, is pretty cache efficient.
763 sgen_sort_addresses (void **array, size_t size)
768 for (i = 1; i < size; ++i) {
771 size_t parent = (child - 1) / 2;
773 if (array [parent] >= array [child])
776 tmp = array [parent];
777 array [parent] = array [child];
784 for (i = size - 1; i > 0; --i) {
787 array [i] = array [0];
793 while (root * 2 + 1 <= end) {
794 size_t child = root * 2 + 1;
796 if (child < end && array [child] < array [child + 1])
798 if (array [root] >= array [child])
802 array [root] = array [child];
811 * Scan the memory between start and end and queue values which could be pointers
812 * to the area between start_nursery and end_nursery for later consideration.
813 * Typically used for thread stacks.
816 sgen_conservatively_pin_objects_from (void **start, void **end, void *start_nursery, void *end_nursery, int pin_type)
820 SGEN_ASSERT (0, ((mword)start & (SIZEOF_VOID_P - 1)) == 0, "Why are we scanning for references in unaligned memory ?");
822 #if defined(VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE) && !defined(_WIN64)
823 VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE (start, (char*)end - (char*)start);
826 while (start < end) {
828 * *start can point to the middle of an object
829 * note: should we handle pointing at the end of an object?
830 * pinning in C# code disallows pointing at the end of an object
831 * but there is some small chance that an optimizing C compiler
832 * may keep the only reference to an object by pointing
833 * at the end of it. We ignore this small chance for now.
834 * Pointers to the end of an object are indistinguishable
835 * from pointers to the start of the next object in memory
836 * so if we allow that we'd need to pin two objects...
837 * We queue the pointer in an array, the
838 * array will then be sorted and uniqued. This way
839 * we can coalesce several pinning pointers and it should
840 * be faster since we'd do a memory scan with increasing
841 * addresses. Note: we can align the address to the allocation
842 * alignment, so the unique process is more effective.
844 mword addr = (mword)*start;
845 addr &= ~(ALLOC_ALIGN - 1);
846 if (addr >= (mword)start_nursery && addr < (mword)end_nursery) {
847 SGEN_LOG (6, "Pinning address %p from %p", (void*)addr, start);
848 sgen_pin_stage_ptr ((void*)addr);
849 binary_protocol_pin_stage (start, (void*)addr);
850 sgen_pin_stats_register_address ((char*)addr, pin_type);
856 SGEN_LOG (7, "found %d potential pinned heap pointers", count);
860 * The first thing we do in a collection is to identify pinned objects.
861 * This function considers all the areas of memory that need to be
862 * conservatively scanned.
865 pin_from_roots (void *start_nursery, void *end_nursery, ScanCopyContext ctx)
869 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);
870 /* objects pinned from the API are inside these roots */
871 SGEN_HASH_TABLE_FOREACH (&roots_hash [ROOT_TYPE_PINNED], start_root, root) {
872 SGEN_LOG (6, "Pinned roots %p-%p", start_root, root->end_root);
873 sgen_conservatively_pin_objects_from (start_root, (void**)root->end_root, start_nursery, end_nursery, PIN_TYPE_OTHER);
874 } SGEN_HASH_TABLE_FOREACH_END;
875 /* now deal with the thread stacks
876 * in the future we should be able to conservatively scan only:
877 * *) the cpu registers
878 * *) the unmanaged stack frames
879 * *) the _last_ managed stack frame
880 * *) pointers slots in managed frames
882 sgen_client_scan_thread_data (start_nursery, end_nursery, FALSE, ctx);
886 single_arg_user_copy_or_mark (GCObject **obj, void *gc_data)
888 ScanCopyContext *ctx = gc_data;
889 ctx->ops->copy_or_mark_object (obj, ctx->queue);
893 * The memory area from start_root to end_root contains pointers to objects.
894 * Their position is precisely described by @desc (this means that the pointer
895 * can be either NULL or the pointer to the start of an object).
896 * This functions copies them to to_space updates them.
898 * This function is not thread-safe!
901 precisely_scan_objects_from (void** start_root, void** end_root, char* n_start, char *n_end, SgenDescriptor desc, ScanCopyContext ctx)
903 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
904 SgenGrayQueue *queue = ctx.queue;
906 switch (desc & ROOT_DESC_TYPE_MASK) {
907 case ROOT_DESC_BITMAP:
908 desc >>= ROOT_DESC_TYPE_SHIFT;
910 if ((desc & 1) && *start_root) {
911 copy_func ((GCObject**)start_root, queue);
912 SGEN_LOG (9, "Overwrote root at %p with %p", start_root, *start_root);
918 case ROOT_DESC_COMPLEX: {
919 gsize *bitmap_data = sgen_get_complex_descriptor_bitmap (desc);
920 gsize bwords = (*bitmap_data) - 1;
921 void **start_run = start_root;
923 while (bwords-- > 0) {
924 gsize bmap = *bitmap_data++;
925 void **objptr = start_run;
927 if ((bmap & 1) && *objptr) {
928 copy_func ((GCObject**)objptr, queue);
929 SGEN_LOG (9, "Overwrote root at %p with %p", objptr, *objptr);
934 start_run += GC_BITS_PER_WORD;
938 case ROOT_DESC_USER: {
939 SgenUserRootMarkFunc marker = sgen_get_user_descriptor_func (desc);
940 marker (start_root, single_arg_user_copy_or_mark, &ctx);
943 case ROOT_DESC_RUN_LEN:
944 g_assert_not_reached ();
946 g_assert_not_reached ();
951 reset_heap_boundaries (void)
953 lowest_heap_address = ~(mword)0;
954 highest_heap_address = 0;
958 sgen_update_heap_boundaries (mword low, mword high)
963 old = lowest_heap_address;
966 } while (SGEN_CAS_PTR ((gpointer*)&lowest_heap_address, (gpointer)low, (gpointer)old) != (gpointer)old);
969 old = highest_heap_address;
972 } while (SGEN_CAS_PTR ((gpointer*)&highest_heap_address, (gpointer)high, (gpointer)old) != (gpointer)old);
976 * Allocate and setup the data structures needed to be able to allocate objects
977 * in the nursery. The nursery is stored in nursery_section.
982 GCMemSection *section;
989 SGEN_LOG (2, "Allocating nursery size: %zu", (size_t)sgen_nursery_size);
990 /* later we will alloc a larger area for the nursery but only activate
991 * what we need. The rest will be used as expansion if we have too many pinned
992 * objects in the existing nursery.
994 /* FIXME: handle OOM */
995 section = sgen_alloc_internal (INTERNAL_MEM_SECTION);
997 alloc_size = sgen_nursery_size;
999 /* If there isn't enough space even for the nursery we should simply abort. */
1000 g_assert (sgen_memgov_try_alloc_space (alloc_size, SPACE_NURSERY));
1002 data = major_collector.alloc_heap (alloc_size, alloc_size, DEFAULT_NURSERY_BITS);
1003 sgen_update_heap_boundaries ((mword)data, (mword)(data + sgen_nursery_size));
1004 SGEN_LOG (4, "Expanding nursery size (%p-%p): %lu, total: %lu", data, data + alloc_size, (unsigned long)sgen_nursery_size, (unsigned long)sgen_gc_get_total_heap_allocation ());
1005 section->data = section->next_data = data;
1006 section->size = alloc_size;
1007 section->end_data = data + sgen_nursery_size;
1008 scan_starts = (alloc_size + SCAN_START_SIZE - 1) / SCAN_START_SIZE;
1009 section->scan_starts = sgen_alloc_internal_dynamic (sizeof (char*) * scan_starts, INTERNAL_MEM_SCAN_STARTS, TRUE);
1010 section->num_scan_start = scan_starts;
1012 nursery_section = section;
1014 sgen_nursery_allocator_set_nursery_bounds (data, data + sgen_nursery_size);
1018 mono_gc_get_logfile (void)
1020 return gc_debug_file;
1024 scan_finalizer_entries (SgenPointerQueue *fin_queue, ScanCopyContext ctx)
1026 CopyOrMarkObjectFunc copy_func = ctx.ops->copy_or_mark_object;
1027 SgenGrayQueue *queue = ctx.queue;
1030 for (i = 0; i < fin_queue->next_slot; ++i) {
1031 GCObject *obj = fin_queue->data [i];
1034 SGEN_LOG (5, "Scan of fin ready object: %p (%s)\n", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
1035 copy_func ((GCObject**)&fin_queue->data [i], queue);
1040 generation_name (int generation)
1042 switch (generation) {
1043 case GENERATION_NURSERY: return "nursery";
1044 case GENERATION_OLD: return "old";
1045 default: g_assert_not_reached ();
1050 sgen_generation_name (int generation)
1052 return generation_name (generation);
1056 finish_gray_stack (int generation, ScanCopyContext ctx)
1060 int done_with_ephemerons, ephemeron_rounds = 0;
1061 char *start_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_start () : NULL;
1062 char *end_addr = generation == GENERATION_NURSERY ? sgen_get_nursery_end () : (char*)-1;
1063 SgenGrayQueue *queue = ctx.queue;
1066 * We copied all the reachable objects. Now it's the time to copy
1067 * the objects that were not referenced by the roots, but by the copied objects.
1068 * we built a stack of objects pointed to by gray_start: they are
1069 * additional roots and we may add more items as we go.
1070 * We loop until gray_start == gray_objects which means no more objects have
1071 * been added. Note this is iterative: no recursion is involved.
1072 * We need to walk the LO list as well in search of marked big objects
1073 * (use a flag since this is needed only on major collections). We need to loop
1074 * here as well, so keep a counter of marked LO (increasing it in copy_object).
1075 * To achieve better cache locality and cache usage, we drain the gray stack
1076 * frequently, after each object is copied, and just finish the work here.
1078 sgen_drain_gray_stack (ctx);
1080 SGEN_LOG (2, "%s generation done", generation_name (generation));
1083 Reset bridge data, we might have lingering data from a previous collection if this is a major
1084 collection trigged by minor overflow.
1086 We must reset the gathered bridges since their original block might be evacuated due to major
1087 fragmentation in the meanwhile and the bridge code should not have to deal with that.
1089 if (sgen_client_bridge_need_processing ())
1090 sgen_client_bridge_reset_data ();
1093 * Walk the ephemeron tables marking all values with reachable keys. This must be completely done
1094 * before processing finalizable objects and non-tracking weak links to avoid finalizing/clearing
1095 * objects that are in fact reachable.
1097 done_with_ephemerons = 0;
1099 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1100 sgen_drain_gray_stack (ctx);
1102 } while (!done_with_ephemerons);
1104 sgen_client_mark_togglerefs (start_addr, end_addr, ctx);
1106 if (sgen_client_bridge_need_processing ()) {
1107 /*Make sure the gray stack is empty before we process bridge objects so we get liveness right*/
1108 sgen_drain_gray_stack (ctx);
1109 sgen_collect_bridge_objects (generation, ctx);
1110 if (generation == GENERATION_OLD)
1111 sgen_collect_bridge_objects (GENERATION_NURSERY, ctx);
1114 Do the first bridge step here, as the collector liveness state will become useless after that.
1116 An important optimization is to only proccess the possibly dead part of the object graph and skip
1117 over all live objects as we transitively know everything they point must be alive too.
1119 The above invariant is completely wrong if we let the gray queue be drained and mark/copy everything.
1121 This has the unfortunate side effect of making overflow collections perform the first step twice, but
1122 given we now have heuristics that perform major GC in anticipation of minor overflows this should not
1125 sgen_client_bridge_processing_stw_step ();
1129 Make sure we drain the gray stack before processing disappearing links and finalizers.
1130 If we don't make sure it is empty we might wrongly see a live object as dead.
1132 sgen_drain_gray_stack (ctx);
1135 We must clear weak links that don't track resurrection before processing object ready for
1136 finalization so they can be cleared before that.
1138 sgen_null_link_in_range (generation, ctx, FALSE);
1139 if (generation == GENERATION_OLD)
1140 sgen_null_link_in_range (GENERATION_NURSERY, ctx, FALSE);
1143 /* walk the finalization queue and move also the objects that need to be
1144 * finalized: use the finalized objects as new roots so the objects they depend
1145 * on are also not reclaimed. As with the roots above, only objects in the nursery
1146 * are marked/copied.
1148 sgen_finalize_in_range (generation, ctx);
1149 if (generation == GENERATION_OLD)
1150 sgen_finalize_in_range (GENERATION_NURSERY, ctx);
1151 /* drain the new stack that might have been created */
1152 SGEN_LOG (6, "Precise scan of gray area post fin");
1153 sgen_drain_gray_stack (ctx);
1156 * This must be done again after processing finalizable objects since CWL slots are cleared only after the key is finalized.
1158 done_with_ephemerons = 0;
1160 done_with_ephemerons = sgen_client_mark_ephemerons (ctx);
1161 sgen_drain_gray_stack (ctx);
1163 } while (!done_with_ephemerons);
1165 sgen_client_clear_unreachable_ephemerons (ctx);
1168 * We clear togglerefs only after all possible chances of revival are done.
1169 * This is semantically more inline with what users expect and it allows for
1170 * user finalizers to correctly interact with TR objects.
1172 sgen_client_clear_togglerefs (start_addr, end_addr, ctx);
1175 SGEN_LOG (2, "Finalize queue handling scan for %s generation: %ld usecs %d ephemeron rounds", generation_name (generation), TV_ELAPSED (atv, btv), ephemeron_rounds);
1178 * handle disappearing links
1179 * Note we do this after checking the finalization queue because if an object
1180 * survives (at least long enough to be finalized) we don't clear the link.
1181 * This also deals with a possible issue with the monitor reclamation: with the Boehm
1182 * GC a finalized object my lose the monitor because it is cleared before the finalizer is
1185 g_assert (sgen_gray_object_queue_is_empty (queue));
1187 sgen_null_link_in_range (generation, ctx, TRUE);
1188 if (generation == GENERATION_OLD)
1189 sgen_null_link_in_range (GENERATION_NURSERY, ctx, TRUE);
1190 if (sgen_gray_object_queue_is_empty (queue))
1192 sgen_drain_gray_stack (ctx);
1195 g_assert (sgen_gray_object_queue_is_empty (queue));
1197 sgen_gray_object_queue_trim_free_list (queue);
1201 sgen_check_section_scan_starts (GCMemSection *section)
1204 for (i = 0; i < section->num_scan_start; ++i) {
1205 if (section->scan_starts [i]) {
1206 mword size = safe_object_get_size ((GCObject*) section->scan_starts [i]);
1207 SGEN_ASSERT (0, size >= SGEN_CLIENT_MINIMUM_OBJECT_SIZE && size <= MAX_SMALL_OBJ_SIZE, "Weird object size at scan starts.");
1213 check_scan_starts (void)
1215 if (!do_scan_starts_check)
1217 sgen_check_section_scan_starts (nursery_section);
1218 major_collector.check_scan_starts ();
1222 scan_from_registered_roots (char *addr_start, char *addr_end, int root_type, ScanCopyContext ctx)
1226 SGEN_HASH_TABLE_FOREACH (&roots_hash [root_type], start_root, root) {
1227 SGEN_LOG (6, "Precise root scan %p-%p (desc: %p)", start_root, root->end_root, (void*)root->root_desc);
1228 precisely_scan_objects_from (start_root, (void**)root->end_root, addr_start, addr_end, root->root_desc, ctx);
1229 } SGEN_HASH_TABLE_FOREACH_END;
1235 static gboolean inited = FALSE;
1240 mono_counters_register ("Collection max time", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME | MONO_COUNTER_MONOTONIC, &time_max);
1242 mono_counters_register ("Minor fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pre_collection_fragment_clear);
1243 mono_counters_register ("Minor pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_pinning);
1244 mono_counters_register ("Minor scan remembered set", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_remsets);
1245 mono_counters_register ("Minor scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_pinned);
1246 mono_counters_register ("Minor scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_scan_roots);
1247 mono_counters_register ("Minor fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_minor_fragment_creation);
1249 mono_counters_register ("Major fragment clear", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pre_collection_fragment_clear);
1250 mono_counters_register ("Major pinning", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_pinning);
1251 mono_counters_register ("Major scan pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_pinned);
1252 mono_counters_register ("Major scan roots", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_roots);
1253 mono_counters_register ("Major scan mod union", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_scan_mod_union);
1254 mono_counters_register ("Major finish gray stack", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_finish_gray_stack);
1255 mono_counters_register ("Major free big objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_free_bigobjs);
1256 mono_counters_register ("Major LOS sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_los_sweep);
1257 mono_counters_register ("Major sweep", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_sweep);
1258 mono_counters_register ("Major fragment creation", MONO_COUNTER_GC | MONO_COUNTER_ULONG | MONO_COUNTER_TIME, &time_major_fragment_creation);
1260 mono_counters_register ("Number of pinned objects", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_pinned_objects);
1262 #ifdef HEAVY_STATISTICS
1263 mono_counters_register ("WBarrier remember pointer", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_add_to_global_remset);
1264 mono_counters_register ("WBarrier arrayref copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_arrayref_copy);
1265 mono_counters_register ("WBarrier generic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store);
1266 mono_counters_register ("WBarrier generic atomic store called", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_generic_store_atomic);
1267 mono_counters_register ("WBarrier set root", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_wbarrier_set_root);
1269 mono_counters_register ("# objects allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_alloced_degraded);
1270 mono_counters_register ("bytes allocated degraded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced_degraded);
1272 mono_counters_register ("# copy_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_nursery);
1273 mono_counters_register ("# objects copied (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_nursery);
1274 mono_counters_register ("# copy_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_copy_object_called_major);
1275 mono_counters_register ("# objects copied (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_copied_major);
1277 mono_counters_register ("# scan_object() called (nursery)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_nursery);
1278 mono_counters_register ("# scan_object() called (major)", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_scan_object_called_major);
1280 mono_counters_register ("Slots allocated in vain", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_slots_allocated_in_vain);
1282 mono_counters_register ("# nursery copy_object() failed from space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_from_space);
1283 mono_counters_register ("# nursery copy_object() failed forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_forwarded);
1284 mono_counters_register ("# nursery copy_object() failed pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_pinned);
1285 mono_counters_register ("# nursery copy_object() failed to space", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_nursery_copy_object_failed_to_space);
1287 sgen_nursery_allocator_init_heavy_stats ();
1295 reset_pinned_from_failed_allocation (void)
1297 bytes_pinned_from_failed_allocation = 0;
1301 sgen_set_pinned_from_failed_allocation (mword objsize)
1303 bytes_pinned_from_failed_allocation += objsize;
1307 sgen_collection_is_concurrent (void)
1309 switch (current_collection_generation) {
1310 case GENERATION_NURSERY:
1312 case GENERATION_OLD:
1313 return concurrent_collection_in_progress;
1315 g_error ("Invalid current generation %d", current_collection_generation);
1321 sgen_concurrent_collection_in_progress (void)
1323 return concurrent_collection_in_progress;
1327 SgenThreadPoolJob job;
1328 SgenObjectOperations *ops;
1332 job_remembered_set_scan (void *worker_data_untyped, SgenThreadPoolJob *job)
1334 WorkerData *worker_data = worker_data_untyped;
1335 ScanJob *job_data = (ScanJob*)job;
1336 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1337 remset.scan_remsets (ctx);
1341 SgenThreadPoolJob job;
1342 SgenObjectOperations *ops;
1346 } ScanFromRegisteredRootsJob;
1349 job_scan_from_registered_roots (void *worker_data_untyped, SgenThreadPoolJob *job)
1351 WorkerData *worker_data = worker_data_untyped;
1352 ScanFromRegisteredRootsJob *job_data = (ScanFromRegisteredRootsJob*)job;
1353 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1355 scan_from_registered_roots (job_data->heap_start, job_data->heap_end, job_data->root_type, ctx);
1359 SgenThreadPoolJob job;
1360 SgenObjectOperations *ops;
1363 } ScanThreadDataJob;
1366 job_scan_thread_data (void *worker_data_untyped, SgenThreadPoolJob *job)
1368 WorkerData *worker_data = worker_data_untyped;
1369 ScanThreadDataJob *job_data = (ScanThreadDataJob*)job;
1370 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1372 sgen_client_scan_thread_data (job_data->heap_start, job_data->heap_end, TRUE, ctx);
1376 SgenThreadPoolJob job;
1377 SgenObjectOperations *ops;
1378 SgenPointerQueue *queue;
1379 } ScanFinalizerEntriesJob;
1382 job_scan_finalizer_entries (void *worker_data_untyped, SgenThreadPoolJob *job)
1384 WorkerData *worker_data = worker_data_untyped;
1385 ScanFinalizerEntriesJob *job_data = (ScanFinalizerEntriesJob*)job;
1386 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1388 scan_finalizer_entries (job_data->queue, ctx);
1392 job_scan_major_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1394 WorkerData *worker_data = worker_data_untyped;
1395 ScanJob *job_data = (ScanJob*)job;
1396 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1398 g_assert (concurrent_collection_in_progress);
1399 major_collector.scan_card_table (TRUE, ctx);
1403 job_scan_los_mod_union_card_table (void *worker_data_untyped, SgenThreadPoolJob *job)
1405 WorkerData *worker_data = worker_data_untyped;
1406 ScanJob *job_data = (ScanJob*)job;
1407 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (job_data->ops, sgen_workers_get_job_gray_queue (worker_data));
1409 g_assert (concurrent_collection_in_progress);
1410 sgen_los_scan_card_table (TRUE, ctx);
1414 init_gray_queue (gboolean use_workers)
1417 sgen_workers_init_distribute_gray_queue ();
1418 sgen_gray_object_queue_init (&gray_queue, NULL);
1422 enqueue_scan_from_roots_jobs (char *heap_start, char *heap_end, SgenObjectOperations *ops, gboolean enqueue)
1424 ScanFromRegisteredRootsJob *scrrj;
1425 ScanThreadDataJob *stdj;
1426 ScanFinalizerEntriesJob *sfej;
1428 /* registered roots, this includes static fields */
1430 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots normal", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1432 scrrj->heap_start = heap_start;
1433 scrrj->heap_end = heap_end;
1434 scrrj->root_type = ROOT_TYPE_NORMAL;
1435 sgen_workers_enqueue_job (&scrrj->job, enqueue);
1437 scrrj = (ScanFromRegisteredRootsJob*)sgen_thread_pool_job_alloc ("scan from registered roots wbarrier", job_scan_from_registered_roots, sizeof (ScanFromRegisteredRootsJob));
1439 scrrj->heap_start = heap_start;
1440 scrrj->heap_end = heap_end;
1441 scrrj->root_type = ROOT_TYPE_WBARRIER;
1442 sgen_workers_enqueue_job (&scrrj->job, enqueue);
1446 stdj = (ScanThreadDataJob*)sgen_thread_pool_job_alloc ("scan thread data", job_scan_thread_data, sizeof (ScanThreadDataJob));
1447 stdj->heap_start = heap_start;
1448 stdj->heap_end = heap_end;
1449 sgen_workers_enqueue_job (&stdj->job, enqueue);
1451 /* Scan the list of objects ready for finalization. */
1453 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1454 sfej->queue = &fin_ready_queue;
1456 sgen_workers_enqueue_job (&sfej->job, enqueue);
1458 sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan critical finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
1459 sfej->queue = &critical_fin_queue;
1461 sgen_workers_enqueue_job (&sfej->job, enqueue);
1465 * Perform a nursery collection.
1467 * Return whether any objects were late-pinned due to being out of memory.
1470 collect_nursery (SgenGrayQueue *unpin_queue, gboolean finish_up_concurrent_mark)
1472 gboolean needs_major;
1473 size_t max_garbage_amount;
1475 mword fragment_total;
1477 SgenObjectOperations *object_ops = &sgen_minor_collector.serial_ops;
1478 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, &gray_queue);
1482 if (disable_minor_collections)
1485 TV_GETTIME (last_minor_collection_start_tv);
1486 atv = last_minor_collection_start_tv;
1488 binary_protocol_collection_begin (gc_stats.minor_gc_count, GENERATION_NURSERY);
1490 if (do_verify_nursery || do_dump_nursery_content)
1491 sgen_debug_verify_nursery (do_dump_nursery_content);
1493 current_collection_generation = GENERATION_NURSERY;
1495 SGEN_ASSERT (0, !sgen_collection_is_concurrent (), "Why is the nursery collection concurrent?");
1497 reset_pinned_from_failed_allocation ();
1499 check_scan_starts ();
1501 sgen_nursery_alloc_prepare_for_minor ();
1505 nursery_next = sgen_nursery_alloc_get_upper_alloc_bound ();
1506 /* FIXME: optimize later to use the higher address where an object can be present */
1507 nursery_next = MAX (nursery_next, sgen_get_nursery_end ());
1509 SGEN_LOG (1, "Start nursery collection %d %p-%p, size: %d", gc_stats.minor_gc_count, sgen_get_nursery_start (), nursery_next, (int)(nursery_next - sgen_get_nursery_start ()));
1510 max_garbage_amount = nursery_next - sgen_get_nursery_start ();
1511 g_assert (nursery_section->size >= max_garbage_amount);
1513 /* world must be stopped already */
1515 time_minor_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1517 sgen_client_pre_collection_checks ();
1519 nursery_section->next_data = nursery_next;
1521 major_collector.start_nursery_collection ();
1523 sgen_memgov_minor_collection_start ();
1525 init_gray_queue (FALSE);
1527 gc_stats.minor_gc_count ++;
1529 if (whole_heap_check_before_collection) {
1530 sgen_clear_nursery_fragments ();
1531 sgen_check_whole_heap (finish_up_concurrent_mark);
1533 if (consistency_check_at_minor_collection)
1534 sgen_check_consistency ();
1536 sgen_process_fin_stage_entries ();
1538 /* pin from pinned handles */
1539 sgen_init_pinning ();
1540 sgen_client_binary_protocol_mark_start (GENERATION_NURSERY);
1541 pin_from_roots (sgen_get_nursery_start (), nursery_next, ctx);
1542 /* pin cemented objects */
1543 sgen_pin_cemented_objects ();
1544 /* identify pinned objects */
1545 sgen_optimize_pin_queue ();
1546 sgen_pinning_setup_section (nursery_section);
1548 pin_objects_in_nursery (FALSE, ctx);
1549 sgen_pinning_trim_queue_to_section (nursery_section);
1552 time_minor_pinning += TV_ELAPSED (btv, atv);
1553 SGEN_LOG (2, "Finding pinned pointers: %zd in %ld usecs", sgen_get_pinned_count (), TV_ELAPSED (btv, atv));
1554 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
1557 * FIXME: When we finish a concurrent collection we do a nursery collection first,
1558 * as part of which we scan the card table. Then, later, we scan the mod union
1559 * cardtable. We should only have to do one.
1561 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan remset", job_remembered_set_scan, sizeof (ScanJob));
1562 sj->ops = object_ops;
1563 sgen_workers_enqueue_job (&sj->job, FALSE);
1565 /* we don't have complete write barrier yet, so we scan all the old generation sections */
1567 time_minor_scan_remsets += TV_ELAPSED (atv, btv);
1568 SGEN_LOG (2, "Old generation scan: %ld usecs", TV_ELAPSED (atv, btv));
1570 sgen_pin_stats_print_class_stats ();
1572 sgen_drain_gray_stack (ctx);
1574 /* FIXME: Why do we do this at this specific, seemingly random, point? */
1575 sgen_client_collecting_minor (&fin_ready_queue, &critical_fin_queue);
1578 time_minor_scan_pinned += TV_ELAPSED (btv, atv);
1580 enqueue_scan_from_roots_jobs (sgen_get_nursery_start (), nursery_next, object_ops, FALSE);
1583 time_minor_scan_roots += TV_ELAPSED (atv, btv);
1585 finish_gray_stack (GENERATION_NURSERY, ctx);
1588 time_minor_finish_gray_stack += TV_ELAPSED (btv, atv);
1589 sgen_client_binary_protocol_mark_end (GENERATION_NURSERY);
1591 if (objects_pinned) {
1592 sgen_optimize_pin_queue ();
1593 sgen_pinning_setup_section (nursery_section);
1596 /* walk the pin_queue, build up the fragment list of free memory, unmark
1597 * pinned objects as we go, memzero() the empty fragments so they are ready for the
1600 sgen_client_binary_protocol_reclaim_start (GENERATION_NURSERY);
1601 fragment_total = sgen_build_nursery_fragments (nursery_section, unpin_queue);
1602 if (!fragment_total)
1605 /* Clear TLABs for all threads */
1606 sgen_clear_tlabs ();
1608 sgen_client_binary_protocol_reclaim_end (GENERATION_NURSERY);
1610 time_minor_fragment_creation += TV_ELAPSED (atv, btv);
1611 SGEN_LOG (2, "Fragment creation: %ld usecs, %lu bytes available", TV_ELAPSED (atv, btv), (unsigned long)fragment_total);
1613 if (consistency_check_at_minor_collection)
1614 sgen_check_major_refs ();
1616 major_collector.finish_nursery_collection ();
1618 TV_GETTIME (last_minor_collection_end_tv);
1619 gc_stats.minor_gc_time += TV_ELAPSED (last_minor_collection_start_tv, last_minor_collection_end_tv);
1621 sgen_debug_dump_heap ("minor", gc_stats.minor_gc_count - 1, NULL);
1623 /* prepare the pin queue for the next collection */
1624 sgen_finish_pinning ();
1625 if (sgen_have_pending_finalizers ()) {
1626 SGEN_LOG (4, "Finalizer-thread wakeup");
1627 sgen_client_finalize_notify ();
1629 sgen_pin_stats_reset ();
1630 /* clear cemented hash */
1631 sgen_cement_clear_below_threshold ();
1633 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
1635 remset.finish_minor_collection ();
1637 check_scan_starts ();
1639 binary_protocol_flush_buffers (FALSE);
1641 sgen_memgov_minor_collection_end ();
1643 /*objects are late pinned because of lack of memory, so a major is a good call*/
1644 needs_major = objects_pinned > 0;
1645 current_collection_generation = -1;
1648 binary_protocol_collection_end (gc_stats.minor_gc_count - 1, GENERATION_NURSERY, 0, 0);
1650 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
1651 sgen_check_nursery_objects_pinned (unpin_queue != NULL);
1657 COPY_OR_MARK_FROM_ROOTS_SERIAL,
1658 COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT,
1659 COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT
1660 } CopyOrMarkFromRootsMode;
1663 major_copy_or_mark_from_roots (size_t *old_next_pin_slot, CopyOrMarkFromRootsMode mode, SgenObjectOperations *object_ops)
1668 /* FIXME: only use these values for the precise scan
1669 * note that to_space pointers should be excluded anyway...
1671 char *heap_start = NULL;
1672 char *heap_end = (char*)-1;
1673 ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, WORKERS_DISTRIBUTE_GRAY_QUEUE);
1674 gboolean concurrent = mode != COPY_OR_MARK_FROM_ROOTS_SERIAL;
1676 SGEN_ASSERT (0, !!concurrent == !!concurrent_collection_in_progress, "We've been called with the wrong mode.");
1678 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1679 /*This cleans up unused fragments */
1680 sgen_nursery_allocator_prepare_for_pinning ();
1682 if (do_concurrent_checks)
1683 sgen_debug_check_nursery_is_clean ();
1685 /* The concurrent collector doesn't touch the nursery. */
1686 sgen_nursery_alloc_prepare_for_major ();
1689 init_gray_queue (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT);
1693 /* Pinning depends on this */
1694 sgen_clear_nursery_fragments ();
1696 if (whole_heap_check_before_collection)
1697 sgen_check_whole_heap (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT);
1700 time_major_pre_collection_fragment_clear += TV_ELAPSED (atv, btv);
1702 if (!sgen_collection_is_concurrent ())
1703 nursery_section->next_data = sgen_get_nursery_end ();
1704 /* we should also coalesce scanning from sections close to each other
1705 * and deal with pointers outside of the sections later.
1710 sgen_client_pre_collection_checks ();
1713 /* Remsets are not useful for a major collection */
1714 remset.clear_cards ();
1717 sgen_process_fin_stage_entries ();
1720 sgen_init_pinning ();
1721 SGEN_LOG (6, "Collecting pinned addresses");
1722 pin_from_roots ((void*)lowest_heap_address, (void*)highest_heap_address, ctx);
1724 sgen_optimize_pin_queue ();
1726 sgen_client_collecting_major_1 ();
1729 * pin_queue now contains all candidate pointers, sorted and
1730 * uniqued. We must do two passes now to figure out which
1731 * objects are pinned.
1733 * The first is to find within the pin_queue the area for each
1734 * section. This requires that the pin_queue be sorted. We
1735 * also process the LOS objects and pinned chunks here.
1737 * The second, destructive, pass is to reduce the section
1738 * areas to pointers to the actually pinned objects.
1740 SGEN_LOG (6, "Pinning from sections");
1741 /* first pass for the sections */
1742 sgen_find_section_pin_queue_start_end (nursery_section);
1743 /* identify possible pointers to the insize of large objects */
1744 SGEN_LOG (6, "Pinning from large objects");
1745 for (bigobj = los_object_list; bigobj; bigobj = bigobj->next) {
1747 if (sgen_find_optimized_pin_queue_area ((char*)bigobj->data, (char*)bigobj->data + sgen_los_object_size (bigobj), &dummy, &dummy)) {
1748 binary_protocol_pin (bigobj->data, (gpointer)LOAD_VTABLE (bigobj->data), safe_object_get_size (bigobj->data));
1750 if (sgen_los_object_is_pinned (bigobj->data)) {
1751 SGEN_ASSERT (0, mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT, "LOS objects can only be pinned here after concurrent marking.");
1754 sgen_los_pin_object (bigobj->data);
1755 if (SGEN_OBJECT_HAS_REFERENCES (bigobj->data))
1756 GRAY_OBJECT_ENQUEUE (WORKERS_DISTRIBUTE_GRAY_QUEUE, bigobj->data, sgen_obj_get_descriptor ((GCObject*)bigobj->data));
1757 sgen_pin_stats_register_object (bigobj->data, safe_object_get_size (bigobj->data));
1758 SGEN_LOG (6, "Marked large object %p (%s) size: %lu from roots", bigobj->data,
1759 sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (bigobj->data)),
1760 (unsigned long)sgen_los_object_size (bigobj));
1762 sgen_client_pinned_los_object (bigobj->data);
1766 pin_objects_in_nursery (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT, ctx);
1767 if (check_nursery_objects_pinned && !sgen_minor_collector.is_split)
1768 sgen_check_nursery_objects_pinned (mode != COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT);
1770 major_collector.pin_objects (WORKERS_DISTRIBUTE_GRAY_QUEUE);
1771 if (old_next_pin_slot)
1772 *old_next_pin_slot = sgen_get_pinned_count ();
1775 time_major_pinning += TV_ELAPSED (atv, btv);
1776 SGEN_LOG (2, "Finding pinned pointers: %zd in %ld usecs", sgen_get_pinned_count (), TV_ELAPSED (atv, btv));
1777 SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
1779 major_collector.init_to_space ();
1782 * The concurrent collector doesn't move objects, neither on
1783 * the major heap nor in the nursery, so we can mark even
1784 * before pinning has finished. For the non-concurrent
1785 * collector we start the workers after pinning.
1787 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1788 SGEN_ASSERT (0, sgen_workers_all_done (), "Why are the workers not done when we start or finish a major collection?");
1789 sgen_workers_start_all_workers (object_ops);
1790 gray_queue_enable_redirect (WORKERS_DISTRIBUTE_GRAY_QUEUE);
1791 } else if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
1792 if (sgen_workers_have_idle_work ()) {
1793 sgen_workers_start_all_workers (object_ops);
1794 sgen_workers_join ();
1798 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
1799 main_gc_thread = mono_native_thread_self ();
1802 sgen_client_collecting_major_2 ();
1805 time_major_scan_pinned += TV_ELAPSED (btv, atv);
1807 sgen_client_collecting_major_3 (&fin_ready_queue, &critical_fin_queue);
1810 * FIXME: is this the right context? It doesn't seem to contain a copy function
1811 * unless we're concurrent.
1813 enqueue_scan_from_roots_jobs (heap_start, heap_end, object_ops, mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT);
1816 time_major_scan_roots += TV_ELAPSED (atv, btv);
1818 if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
1821 /* Mod union card table */
1822 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan mod union cardtable", job_scan_major_mod_union_card_table, sizeof (ScanJob));
1823 sj->ops = object_ops;
1824 sgen_workers_enqueue_job (&sj->job, FALSE);
1826 sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan LOS mod union cardtable", job_scan_los_mod_union_card_table, sizeof (ScanJob));
1827 sj->ops = object_ops;
1828 sgen_workers_enqueue_job (&sj->job, FALSE);
1831 time_major_scan_mod_union += TV_ELAPSED (btv, atv);
1834 sgen_pin_stats_print_class_stats ();
1838 major_finish_copy_or_mark (CopyOrMarkFromRootsMode mode)
1840 if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
1842 * Prepare the pin queue for the next collection. Since pinning runs on the worker
1843 * threads we must wait for the jobs to finish before we can reset it.
1845 sgen_workers_wait_for_jobs_finished ();
1846 sgen_finish_pinning ();
1848 sgen_pin_stats_reset ();
1850 if (do_concurrent_checks)
1851 sgen_debug_check_nursery_is_clean ();
1856 major_start_collection (gboolean concurrent, size_t *old_next_pin_slot)
1858 SgenObjectOperations *object_ops;
1860 binary_protocol_collection_begin (gc_stats.major_gc_count, GENERATION_OLD);
1862 current_collection_generation = GENERATION_OLD;
1864 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
1867 sgen_cement_reset ();
1870 g_assert (major_collector.is_concurrent);
1871 concurrent_collection_in_progress = TRUE;
1873 object_ops = &major_collector.major_ops_concurrent_start;
1875 object_ops = &major_collector.major_ops_serial;
1878 reset_pinned_from_failed_allocation ();
1880 sgen_memgov_major_collection_start ();
1882 //count_ref_nonref_objs ();
1883 //consistency_check ();
1885 check_scan_starts ();
1888 SGEN_LOG (1, "Start major collection %d", gc_stats.major_gc_count);
1889 gc_stats.major_gc_count ++;
1891 if (major_collector.start_major_collection)
1892 major_collector.start_major_collection ();
1894 major_copy_or_mark_from_roots (old_next_pin_slot, concurrent ? COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT : COPY_OR_MARK_FROM_ROOTS_SERIAL, object_ops);
1895 major_finish_copy_or_mark (concurrent ? COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT : COPY_OR_MARK_FROM_ROOTS_SERIAL);
1899 major_finish_collection (const char *reason, size_t old_next_pin_slot, gboolean forced)
1901 ScannedObjectCounts counts;
1902 SgenObjectOperations *object_ops;
1903 mword fragment_total;
1909 if (concurrent_collection_in_progress) {
1910 object_ops = &major_collector.major_ops_concurrent_finish;
1912 major_copy_or_mark_from_roots (NULL, COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT, object_ops);
1914 major_finish_copy_or_mark (COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT);
1916 #ifdef SGEN_DEBUG_INTERNAL_ALLOC
1917 main_gc_thread = NULL;
1920 object_ops = &major_collector.major_ops_serial;
1923 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
1925 /* all the objects in the heap */
1926 finish_gray_stack (GENERATION_OLD, CONTEXT_FROM_OBJECT_OPERATIONS (object_ops, &gray_queue));
1928 time_major_finish_gray_stack += TV_ELAPSED (btv, atv);
1930 SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after joining");
1932 if (objects_pinned) {
1933 g_assert (!concurrent_collection_in_progress);
1936 * This is slow, but we just OOM'd.
1938 * See comment at `sgen_pin_queue_clear_discarded_entries` for how the pin
1939 * queue is laid out at this point.
1941 sgen_pin_queue_clear_discarded_entries (nursery_section, old_next_pin_slot);
1943 * We need to reestablish all pinned nursery objects in the pin queue
1944 * because they're needed for fragment creation. Unpinning happens by
1945 * walking the whole queue, so it's not necessary to reestablish where major
1946 * heap block pins are - all we care is that they're still in there
1949 sgen_optimize_pin_queue ();
1950 sgen_find_section_pin_queue_start_end (nursery_section);
1954 reset_heap_boundaries ();
1955 sgen_update_heap_boundaries ((mword)sgen_get_nursery_start (), (mword)sgen_get_nursery_end ());
1957 /* walk the pin_queue, build up the fragment list of free memory, unmark
1958 * pinned objects as we go, memzero() the empty fragments so they are ready for the
1961 fragment_total = sgen_build_nursery_fragments (nursery_section, NULL);
1962 if (!fragment_total)
1964 SGEN_LOG (4, "Free space in nursery after major %ld", (long)fragment_total);
1966 if (do_concurrent_checks && concurrent_collection_in_progress)
1967 sgen_debug_check_nursery_is_clean ();
1969 /* prepare the pin queue for the next collection */
1970 sgen_finish_pinning ();
1972 /* Clear TLABs for all threads */
1973 sgen_clear_tlabs ();
1975 sgen_pin_stats_reset ();
1977 sgen_cement_clear_below_threshold ();
1979 if (check_mark_bits_after_major_collection)
1980 sgen_check_heap_marked (concurrent_collection_in_progress);
1983 time_major_fragment_creation += TV_ELAPSED (atv, btv);
1985 binary_protocol_sweep_begin (GENERATION_OLD, !major_collector.sweeps_lazily);
1988 time_major_free_bigobjs += TV_ELAPSED (btv, atv);
1993 time_major_los_sweep += TV_ELAPSED (atv, btv);
1995 major_collector.sweep ();
1997 binary_protocol_sweep_end (GENERATION_OLD, !major_collector.sweeps_lazily);
2000 time_major_sweep += TV_ELAPSED (btv, atv);
2002 sgen_debug_dump_heap ("major", gc_stats.major_gc_count - 1, reason);
2004 if (sgen_have_pending_finalizers ()) {
2005 SGEN_LOG (4, "Finalizer-thread wakeup");
2006 sgen_client_finalize_notify ();
2009 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
2011 sgen_memgov_major_collection_end (forced);
2012 current_collection_generation = -1;
2014 memset (&counts, 0, sizeof (ScannedObjectCounts));
2015 major_collector.finish_major_collection (&counts);
2017 g_assert (sgen_section_gray_queue_is_empty (sgen_workers_get_distribute_section_gray_queue ()));
2019 SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after major collection has finished");
2020 if (concurrent_collection_in_progress)
2021 concurrent_collection_in_progress = FALSE;
2023 check_scan_starts ();
2025 binary_protocol_flush_buffers (FALSE);
2027 //consistency_check ();
2029 binary_protocol_collection_end (gc_stats.major_gc_count - 1, GENERATION_OLD, counts.num_scanned_objects, counts.num_unique_scanned_objects);
2033 major_do_collection (const char *reason, gboolean forced)
2035 TV_DECLARE (time_start);
2036 TV_DECLARE (time_end);
2037 size_t old_next_pin_slot;
2039 if (disable_major_collections)
2042 if (major_collector.get_and_reset_num_major_objects_marked) {
2043 long long num_marked = major_collector.get_and_reset_num_major_objects_marked ();
2044 g_assert (!num_marked);
2047 /* world must be stopped already */
2048 TV_GETTIME (time_start);
2050 major_start_collection (FALSE, &old_next_pin_slot);
2051 major_finish_collection (reason, old_next_pin_slot, forced);
2053 TV_GETTIME (time_end);
2054 gc_stats.major_gc_time += TV_ELAPSED (time_start, time_end);
2056 /* FIXME: also report this to the user, preferably in gc-end. */
2057 if (major_collector.get_and_reset_num_major_objects_marked)
2058 major_collector.get_and_reset_num_major_objects_marked ();
2060 return bytes_pinned_from_failed_allocation > 0;
2064 major_start_concurrent_collection (const char *reason)
2066 TV_DECLARE (time_start);
2067 TV_DECLARE (time_end);
2068 long long num_objects_marked;
2070 if (disable_major_collections)
2073 TV_GETTIME (time_start);
2074 SGEN_TV_GETTIME (time_major_conc_collection_start);
2076 num_objects_marked = major_collector.get_and_reset_num_major_objects_marked ();
2077 g_assert (num_objects_marked == 0);
2079 binary_protocol_concurrent_start ();
2081 // FIXME: store reason and pass it when finishing
2082 major_start_collection (TRUE, NULL);
2084 gray_queue_redirect (&gray_queue);
2086 num_objects_marked = major_collector.get_and_reset_num_major_objects_marked ();
2088 TV_GETTIME (time_end);
2089 gc_stats.major_gc_time += TV_ELAPSED (time_start, time_end);
2091 current_collection_generation = -1;
2095 * Returns whether the major collection has finished.
2098 major_should_finish_concurrent_collection (void)
2100 SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&gray_queue), "Why is the gray queue not empty before we have started doing anything?");
2101 return sgen_workers_all_done ();
2105 major_update_concurrent_collection (void)
2107 TV_DECLARE (total_start);
2108 TV_DECLARE (total_end);
2110 TV_GETTIME (total_start);
2112 binary_protocol_concurrent_update ();
2114 major_collector.update_cardtable_mod_union ();
2115 sgen_los_update_cardtable_mod_union ();
2117 TV_GETTIME (total_end);
2118 gc_stats.major_gc_time += TV_ELAPSED (total_start, total_end);
2122 major_finish_concurrent_collection (gboolean forced)
2124 TV_DECLARE (total_start);
2125 TV_DECLARE (total_end);
2127 TV_GETTIME (total_start);
2129 binary_protocol_concurrent_finish ();
2132 * We need to stop all workers since we're updating the cardtable below.
2133 * The workers will be resumed with a finishing pause context to avoid
2134 * additional cardtable and object scanning.
2136 sgen_workers_stop_all_workers ();
2138 SGEN_TV_GETTIME (time_major_conc_collection_end);
2139 gc_stats.major_gc_time_concurrent += SGEN_TV_ELAPSED (time_major_conc_collection_start, time_major_conc_collection_end);
2141 major_collector.update_cardtable_mod_union ();
2142 sgen_los_update_cardtable_mod_union ();
2144 if (mod_union_consistency_check)
2145 sgen_check_mod_union_consistency ();
2147 current_collection_generation = GENERATION_OLD;
2148 sgen_cement_reset ();
2149 major_finish_collection ("finishing", -1, forced);
2151 if (whole_heap_check_before_collection)
2152 sgen_check_whole_heap (FALSE);
2154 TV_GETTIME (total_end);
2155 gc_stats.major_gc_time += TV_ELAPSED (total_start, total_end) - TV_ELAPSED (last_minor_collection_start_tv, last_minor_collection_end_tv);
2157 current_collection_generation = -1;
2161 * Ensure an allocation request for @size will succeed by freeing enough memory.
2163 * LOCKING: The GC lock MUST be held.
2166 sgen_ensure_free_space (size_t size)
2168 int generation_to_collect = -1;
2169 const char *reason = NULL;
2171 if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
2172 if (sgen_need_major_collection (size)) {
2173 reason = "LOS overflow";
2174 generation_to_collect = GENERATION_OLD;
2177 if (degraded_mode) {
2178 if (sgen_need_major_collection (size)) {
2179 reason = "Degraded mode overflow";
2180 generation_to_collect = GENERATION_OLD;
2182 } else if (sgen_need_major_collection (size)) {
2183 reason = concurrent_collection_in_progress ? "Forced finish concurrent collection" : "Minor allowance";
2184 generation_to_collect = GENERATION_OLD;
2186 generation_to_collect = GENERATION_NURSERY;
2187 reason = "Nursery full";
2191 if (generation_to_collect == -1) {
2192 if (concurrent_collection_in_progress && sgen_workers_all_done ()) {
2193 generation_to_collect = GENERATION_OLD;
2194 reason = "Finish concurrent collection";
2198 if (generation_to_collect == -1)
2200 sgen_perform_collection (size, generation_to_collect, reason, FALSE);
2204 * LOCKING: Assumes the GC lock is held.
2207 sgen_perform_collection (size_t requested_size, int generation_to_collect, const char *reason, gboolean wait_to_finish)
2209 TV_DECLARE (gc_start);
2210 TV_DECLARE (gc_end);
2211 TV_DECLARE (gc_total_start);
2212 TV_DECLARE (gc_total_end);
2213 GGTimingInfo infos [2];
2214 int overflow_generation_to_collect = -1;
2215 int oldest_generation_collected = generation_to_collect;
2216 const char *overflow_reason = NULL;
2218 binary_protocol_collection_requested (generation_to_collect, requested_size, wait_to_finish ? 1 : 0);
2220 SGEN_ASSERT (0, generation_to_collect == GENERATION_NURSERY || generation_to_collect == GENERATION_OLD, "What generation is this?");
2222 TV_GETTIME (gc_start);
2224 sgen_stop_world (generation_to_collect);
2226 TV_GETTIME (gc_total_start);
2228 if (concurrent_collection_in_progress) {
2230 * If the concurrent worker is finished or we are asked to do a major collection
2231 * then we finish the concurrent collection.
2233 gboolean finish = major_should_finish_concurrent_collection () || generation_to_collect == GENERATION_OLD;
2236 major_finish_concurrent_collection (wait_to_finish);
2237 oldest_generation_collected = GENERATION_OLD;
2239 SGEN_ASSERT (0, generation_to_collect == GENERATION_NURSERY, "Why aren't we finishing the concurrent collection?");
2240 major_update_concurrent_collection ();
2241 collect_nursery (NULL, FALSE);
2248 * If we've been asked to do a major collection, and the major collector wants to
2249 * run synchronously (to evacuate), we set the flag to do that.
2251 if (generation_to_collect == GENERATION_OLD &&
2252 allow_synchronous_major &&
2253 major_collector.want_synchronous_collection &&
2254 *major_collector.want_synchronous_collection) {
2255 wait_to_finish = TRUE;
2258 SGEN_ASSERT (0, !concurrent_collection_in_progress, "Why did this not get handled above?");
2261 * There's no concurrent collection in progress. Collect the generation we're asked
2262 * to collect. If the major collector is concurrent and we're not forced to wait,
2263 * start a concurrent collection.
2265 // FIXME: extract overflow reason
2266 if (generation_to_collect == GENERATION_NURSERY) {
2267 if (collect_nursery (NULL, FALSE)) {
2268 overflow_generation_to_collect = GENERATION_OLD;
2269 overflow_reason = "Minor overflow";
2272 if (major_collector.is_concurrent && !wait_to_finish) {
2273 collect_nursery (NULL, FALSE);
2274 major_start_concurrent_collection (reason);
2275 // FIXME: set infos[0] properly
2279 if (major_do_collection (reason, wait_to_finish)) {
2280 overflow_generation_to_collect = GENERATION_NURSERY;
2281 overflow_reason = "Excessive pinning";
2285 TV_GETTIME (gc_end);
2287 memset (infos, 0, sizeof (infos));
2288 infos [0].generation = generation_to_collect;
2289 infos [0].reason = reason;
2290 infos [0].is_overflow = FALSE;
2291 infos [1].generation = -1;
2292 infos [0].total_time = SGEN_TV_ELAPSED (gc_start, gc_end);
2294 SGEN_ASSERT (0, !concurrent_collection_in_progress, "Why did this not get handled above?");
2296 if (overflow_generation_to_collect != -1) {
2298 * We need to do an overflow collection, either because we ran out of memory
2299 * or the nursery is fully pinned.
2302 infos [1].generation = overflow_generation_to_collect;
2303 infos [1].reason = overflow_reason;
2304 infos [1].is_overflow = TRUE;
2307 if (overflow_generation_to_collect == GENERATION_NURSERY)
2308 collect_nursery (NULL, FALSE);
2310 major_do_collection (overflow_reason, wait_to_finish);
2312 TV_GETTIME (gc_end);
2313 infos [1].total_time = SGEN_TV_ELAPSED (gc_start, gc_end);
2315 oldest_generation_collected = MAX (oldest_generation_collected, overflow_generation_to_collect);
2318 SGEN_LOG (2, "Heap size: %lu, LOS size: %lu", (unsigned long)sgen_gc_get_total_heap_allocation (), (unsigned long)los_memory_usage);
2320 /* this also sets the proper pointers for the next allocation */
2321 if (generation_to_collect == GENERATION_NURSERY && !sgen_can_alloc_size (requested_size)) {
2322 /* TypeBuilder and MonoMethod are killing mcs with fragmentation */
2323 SGEN_LOG (1, "nursery collection didn't find enough room for %zd alloc (%zd pinned)", requested_size, sgen_get_pinned_count ());
2324 sgen_dump_pin_queue ();
2329 g_assert (sgen_gray_object_queue_is_empty (&gray_queue));
2331 TV_GETTIME (gc_total_end);
2332 time_max = MAX (time_max, TV_ELAPSED (gc_total_start, gc_total_end));
2334 sgen_restart_world (oldest_generation_collected, infos);
2338 * ######################################################################
2339 * ######## Memory allocation from the OS
2340 * ######################################################################
2341 * This section of code deals with getting memory from the OS and
2342 * allocating memory for GC-internal data structures.
2343 * Internal memory can be handled with a freelist for small objects.
2349 G_GNUC_UNUSED static void
2350 report_internal_mem_usage (void)
2352 printf ("Internal memory usage:\n");
2353 sgen_report_internal_mem_usage ();
2354 printf ("Pinned memory usage:\n");
2355 major_collector.report_pinned_memory_usage ();
2359 * ######################################################################
2360 * ######## Finalization support
2361 * ######################################################################
2365 * If the object has been forwarded it means it's still referenced from a root.
2366 * If it is pinned it's still alive as well.
2367 * A LOS object is only alive if we have pinned it.
2368 * Return TRUE if @obj is ready to be finalized.
2370 static inline gboolean
2371 sgen_is_object_alive (GCObject *object)
2373 if (ptr_in_nursery (object))
2374 return sgen_nursery_is_object_alive (object);
2376 return sgen_major_is_object_alive (object);
2380 * This function returns true if @object is either alive and belongs to the
2381 * current collection - major collections are full heap, so old gen objects
2382 * are never alive during a minor collection.
2385 sgen_is_object_alive_and_on_current_collection (GCObject *object)
2387 if (ptr_in_nursery (object))
2388 return sgen_nursery_is_object_alive (object);
2390 if (current_collection_generation == GENERATION_NURSERY)
2393 return sgen_major_is_object_alive (object);
2398 sgen_gc_is_object_ready_for_finalization (GCObject *object)
2400 return !sgen_is_object_alive (object);
2404 sgen_queue_finalization_entry (GCObject *obj)
2406 gboolean critical = sgen_client_object_has_critical_finalizer (obj);
2408 sgen_pointer_queue_add (critical ? &critical_fin_queue : &fin_ready_queue, obj);
2410 sgen_client_object_queued_for_finalization (obj);
2414 sgen_object_is_live (GCObject *obj)
2416 return sgen_is_object_alive_and_on_current_collection (obj);
2420 * `System.GC.WaitForPendingFinalizers` first checks `sgen_have_pending_finalizers()` to
2421 * determine whether it can exit quickly. The latter must therefore only return FALSE if
2422 * all finalizers have really finished running.
2424 * `sgen_gc_invoke_finalizers()` first dequeues a finalizable object, and then finalizes it.
2425 * This means that just checking whether the queues are empty leaves the possibility that an
2426 * object might have been dequeued but not yet finalized. That's why we need the additional
2427 * flag `pending_unqueued_finalizer`.
2430 static volatile gboolean pending_unqueued_finalizer = FALSE;
2433 sgen_gc_invoke_finalizers (void)
2437 g_assert (!pending_unqueued_finalizer);
2439 /* FIXME: batch to reduce lock contention */
2440 while (sgen_have_pending_finalizers ()) {
2446 * We need to set `pending_unqueued_finalizer` before dequeing the
2447 * finalizable object.
2449 if (!sgen_pointer_queue_is_empty (&fin_ready_queue)) {
2450 pending_unqueued_finalizer = TRUE;
2451 mono_memory_write_barrier ();
2452 obj = sgen_pointer_queue_pop (&fin_ready_queue);
2453 } else if (!sgen_pointer_queue_is_empty (&critical_fin_queue)) {
2454 pending_unqueued_finalizer = TRUE;
2455 mono_memory_write_barrier ();
2456 obj = sgen_pointer_queue_pop (&critical_fin_queue);
2462 SGEN_LOG (7, "Finalizing object %p (%s)", obj, sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (obj)));
2470 /* the object is on the stack so it is pinned */
2471 /*g_print ("Calling finalizer for object: %p (%s)\n", obj, sgen_client_object_safe_name (obj));*/
2472 sgen_client_run_finalize (obj);
2475 if (pending_unqueued_finalizer) {
2476 mono_memory_write_barrier ();
2477 pending_unqueued_finalizer = FALSE;
2484 sgen_have_pending_finalizers (void)
2486 return pending_unqueued_finalizer || !sgen_pointer_queue_is_empty (&fin_ready_queue) || !sgen_pointer_queue_is_empty (&critical_fin_queue);
2490 * ######################################################################
2491 * ######## registered roots support
2492 * ######################################################################
2496 * We do not coalesce roots.
2499 sgen_register_root (char *start, size_t size, SgenDescriptor descr, int root_type, int source, const char *msg)
2501 RootRecord new_root;
2504 for (i = 0; i < ROOT_TYPE_NUM; ++i) {
2505 RootRecord *root = sgen_hash_table_lookup (&roots_hash [i], start);
2506 /* we allow changing the size and the descriptor (for thread statics etc) */
2508 size_t old_size = root->end_root - start;
2509 root->end_root = start + size;
2510 SGEN_ASSERT (0, !!root->root_desc == !!descr, "Can't change whether a root is precise or conservative.");
2511 SGEN_ASSERT (0, root->source == source, "Can't change a root's source identifier.");
2512 SGEN_ASSERT (0, !!root->msg == !!msg, "Can't change a root's message.");
2513 root->root_desc = descr;
2515 roots_size -= old_size;
2521 new_root.end_root = start + size;
2522 new_root.root_desc = descr;
2523 new_root.source = source;
2526 sgen_hash_table_replace (&roots_hash [root_type], start, &new_root, NULL);
2529 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);
2536 sgen_deregister_root (char* addr)
2542 for (root_type = 0; root_type < ROOT_TYPE_NUM; ++root_type) {
2543 if (sgen_hash_table_remove (&roots_hash [root_type], addr, &root))
2544 roots_size -= (root.end_root - addr);
2550 * ######################################################################
2551 * ######## Thread handling (stop/start code)
2552 * ######################################################################
2556 sgen_get_current_collection_generation (void)
2558 return current_collection_generation;
2562 sgen_thread_register (SgenThreadInfo* info, void *stack_bottom_fallback)
2564 #ifndef HAVE_KW_THREAD
2565 info->tlab_start = info->tlab_next = info->tlab_temp_end = info->tlab_real_end = NULL;
2568 sgen_init_tlab_info (info);
2570 sgen_client_thread_register (info, stack_bottom_fallback);
2576 sgen_thread_unregister (SgenThreadInfo *p)
2578 sgen_client_thread_unregister (p);
2582 * ######################################################################
2583 * ######## Write barriers
2584 * ######################################################################
2588 * Note: the write barriers first do the needed GC work and then do the actual store:
2589 * this way the value is visible to the conservative GC scan after the write barrier
2590 * itself. If a GC interrupts the barrier in the middle, value will be kept alive by
2591 * the conservative scan, otherwise by the remembered set scan.
2595 mono_gc_wbarrier_arrayref_copy (gpointer dest_ptr, gpointer src_ptr, int count)
2597 HEAVY_STAT (++stat_wbarrier_arrayref_copy);
2598 /*This check can be done without taking a lock since dest_ptr array is pinned*/
2599 if (ptr_in_nursery (dest_ptr) || count <= 0) {
2600 mono_gc_memmove_aligned (dest_ptr, src_ptr, count * sizeof (gpointer));
2604 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2605 if (binary_protocol_is_heavy_enabled ()) {
2607 for (i = 0; i < count; ++i) {
2608 gpointer dest = (gpointer*)dest_ptr + i;
2609 gpointer obj = *((gpointer*)src_ptr + i);
2611 binary_protocol_wbarrier (dest, obj, (gpointer)LOAD_VTABLE (obj));
2616 remset.wbarrier_arrayref_copy (dest_ptr, src_ptr, count);
2620 mono_gc_wbarrier_generic_nostore (gpointer ptr)
2624 HEAVY_STAT (++stat_wbarrier_generic_store);
2626 sgen_client_wbarrier_generic_nostore_check (ptr);
2628 obj = *(gpointer*)ptr;
2630 binary_protocol_wbarrier (ptr, obj, (gpointer)LOAD_VTABLE (obj));
2633 * We need to record old->old pointer locations for the
2634 * concurrent collector.
2636 if (!ptr_in_nursery (obj) && !concurrent_collection_in_progress) {
2637 SGEN_LOG (8, "Skipping remset at %p", ptr);
2641 SGEN_LOG (8, "Adding remset at %p", ptr);
2643 remset.wbarrier_generic_nostore (ptr);
2647 mono_gc_wbarrier_generic_store (gpointer ptr, GCObject* value)
2649 SGEN_LOG (8, "Wbarrier store at %p to %p (%s)", ptr, value, value ? sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (value)) : "null");
2650 SGEN_UPDATE_REFERENCE_ALLOW_NULL (ptr, value);
2651 if (ptr_in_nursery (value) || concurrent_collection_in_progress)
2652 mono_gc_wbarrier_generic_nostore (ptr);
2653 sgen_dummy_use (value);
2656 /* Same as mono_gc_wbarrier_generic_store () but performs the store
2657 * as an atomic operation with release semantics.
2660 mono_gc_wbarrier_generic_store_atomic (gpointer ptr, GCObject *value)
2662 HEAVY_STAT (++stat_wbarrier_generic_store_atomic);
2664 SGEN_LOG (8, "Wbarrier atomic store at %p to %p (%s)", ptr, value, value ? sgen_client_vtable_get_name (SGEN_LOAD_VTABLE (value)) : "null");
2666 InterlockedWritePointer (ptr, value);
2668 if (ptr_in_nursery (value) || concurrent_collection_in_progress)
2669 mono_gc_wbarrier_generic_nostore (ptr);
2671 sgen_dummy_use (value);
2675 sgen_wbarrier_value_copy_bitmap (gpointer _dest, gpointer _src, int size, unsigned bitmap)
2677 GCObject **dest = _dest;
2678 GCObject **src = _src;
2682 mono_gc_wbarrier_generic_store (dest, *src);
2687 size -= SIZEOF_VOID_P;
2693 * ######################################################################
2694 * ######## Other mono public interface functions.
2695 * ######################################################################
2699 sgen_gc_collect (int generation)
2704 sgen_perform_collection (0, generation, "user request", TRUE);
2709 sgen_gc_collection_count (int generation)
2711 if (generation == 0)
2712 return gc_stats.minor_gc_count;
2713 return gc_stats.major_gc_count;
2717 sgen_gc_get_used_size (void)
2721 tot = los_memory_usage;
2722 tot += nursery_section->next_data - nursery_section->data;
2723 tot += major_collector.get_used_size ();
2724 /* FIXME: account for pinned objects */
2730 sgen_set_allow_synchronous_major (gboolean flag)
2732 if (!major_collector.is_concurrent)
2735 allow_synchronous_major = flag;
2740 sgen_env_var_error (const char *env_var, const char *fallback, const char *description_format, ...)
2744 va_start (ap, description_format);
2746 fprintf (stderr, "Warning: In environment variable `%s': ", env_var);
2747 vfprintf (stderr, description_format, ap);
2749 fprintf (stderr, " - %s", fallback);
2750 fprintf (stderr, "\n");
2756 parse_double_in_interval (const char *env_var, const char *opt_name, const char *opt, double min, double max, double *result)
2759 double val = strtod (opt, &endptr);
2760 if (endptr == opt) {
2761 sgen_env_var_error (env_var, "Using default value.", "`%s` must be a number.", opt_name);
2764 else if (val < min || val > max) {
2765 sgen_env_var_error (env_var, "Using default value.", "`%s` must be between %.2f - %.2f.", opt_name, min, max);
2777 char *major_collector_opt = NULL;
2778 char *minor_collector_opt = NULL;
2779 size_t max_heap = 0;
2780 size_t soft_limit = 0;
2782 gboolean debug_print_allowance = FALSE;
2783 double allowance_ratio = 0, save_target = 0;
2784 gboolean cement_enabled = TRUE;
2787 result = InterlockedCompareExchange (&gc_initialized, -1, 0);
2790 /* already inited */
2793 /* being inited by another thread */
2797 /* we will init it */
2800 g_assert_not_reached ();
2802 } while (result != 0);
2804 SGEN_TV_GETTIME (sgen_init_timestamp);
2806 #ifdef SGEN_WITHOUT_MONO
2807 mono_thread_smr_init ();
2810 mono_coop_mutex_init (&gc_mutex);
2812 gc_debug_file = stderr;
2814 mono_coop_mutex_init (&sgen_interruption_mutex);
2816 if ((env = g_getenv (MONO_GC_PARAMS_NAME))) {
2817 opts = g_strsplit (env, ",", -1);
2818 for (ptr = opts; *ptr; ++ptr) {
2820 if (g_str_has_prefix (opt, "major=")) {
2821 opt = strchr (opt, '=') + 1;
2822 major_collector_opt = g_strdup (opt);
2823 } else if (g_str_has_prefix (opt, "minor=")) {
2824 opt = strchr (opt, '=') + 1;
2825 minor_collector_opt = g_strdup (opt);
2833 sgen_init_internal_allocator ();
2834 sgen_init_nursery_allocator ();
2835 sgen_init_fin_weak_hash ();
2836 sgen_init_hash_table ();
2837 sgen_init_descriptors ();
2838 sgen_init_gray_queues ();
2839 sgen_init_allocator ();
2840 sgen_init_gchandles ();
2842 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_SECTION, SGEN_SIZEOF_GC_MEM_SECTION);
2843 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_GRAY_QUEUE, sizeof (GrayQueueSection));
2845 sgen_client_init ();
2847 if (!minor_collector_opt) {
2848 sgen_simple_nursery_init (&sgen_minor_collector);
2850 if (!strcmp (minor_collector_opt, "simple")) {
2852 sgen_simple_nursery_init (&sgen_minor_collector);
2853 } else if (!strcmp (minor_collector_opt, "split")) {
2854 sgen_split_nursery_init (&sgen_minor_collector);
2856 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using `simple` instead.", "Unknown minor collector `%s'.", minor_collector_opt);
2857 goto use_simple_nursery;
2861 if (!major_collector_opt || !strcmp (major_collector_opt, "marksweep")) {
2862 use_marksweep_major:
2863 sgen_marksweep_init (&major_collector);
2864 } else if (!major_collector_opt || !strcmp (major_collector_opt, "marksweep-conc")) {
2865 sgen_marksweep_conc_init (&major_collector);
2867 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using `marksweep` instead.", "Unknown major collector `%s'.", major_collector_opt);
2868 goto use_marksweep_major;
2871 sgen_nursery_size = DEFAULT_NURSERY_SIZE;
2874 gboolean usage_printed = FALSE;
2876 for (ptr = opts; *ptr; ++ptr) {
2878 if (!strcmp (opt, ""))
2880 if (g_str_has_prefix (opt, "major="))
2882 if (g_str_has_prefix (opt, "minor="))
2884 if (g_str_has_prefix (opt, "max-heap-size=")) {
2885 size_t page_size = mono_pagesize ();
2886 size_t max_heap_candidate = 0;
2887 opt = strchr (opt, '=') + 1;
2888 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &max_heap_candidate)) {
2889 max_heap = (max_heap_candidate + page_size - 1) & ~(size_t)(page_size - 1);
2890 if (max_heap != max_heap_candidate)
2891 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Rounding up.", "`max-heap-size` size must be a multiple of %d.", page_size);
2893 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`max-heap-size` must be an integer.");
2897 if (g_str_has_prefix (opt, "soft-heap-limit=")) {
2898 opt = strchr (opt, '=') + 1;
2899 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &soft_limit)) {
2900 if (soft_limit <= 0) {
2901 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be positive.");
2905 sgen_env_var_error (MONO_GC_PARAMS_NAME, NULL, "`soft-heap-limit` must be an integer.");
2911 if (g_str_has_prefix (opt, "nursery-size=")) {
2913 opt = strchr (opt, '=') + 1;
2914 if (*opt && mono_gc_parse_environment_string_extract_number (opt, &val)) {
2915 if ((val & (val - 1))) {
2916 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be a power of two.");
2920 if (val < SGEN_MAX_NURSERY_WASTE) {
2921 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.",
2922 "`nursery-size` must be at least %d bytes.", SGEN_MAX_NURSERY_WASTE);
2926 sgen_nursery_size = val;
2927 sgen_nursery_bits = 0;
2928 while (ONE_P << (++ sgen_nursery_bits) != sgen_nursery_size)
2931 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`nursery-size` must be an integer.");
2937 if (g_str_has_prefix (opt, "save-target-ratio=")) {
2939 opt = strchr (opt, '=') + 1;
2940 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "save-target-ratio", opt,
2941 SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO, &val)) {
2946 if (g_str_has_prefix (opt, "default-allowance-ratio=")) {
2948 opt = strchr (opt, '=') + 1;
2949 if (parse_double_in_interval (MONO_GC_PARAMS_NAME, "default-allowance-ratio", opt,
2950 SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, SGEN_MIN_ALLOWANCE_NURSERY_SIZE_RATIO, &val)) {
2951 allowance_ratio = val;
2955 if (g_str_has_prefix (opt, "allow-synchronous-major=")) {
2956 if (!major_collector.is_concurrent) {
2957 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring.", "`allow-synchronous-major` is only valid for the concurrent major collector.");
2961 opt = strchr (opt, '=') + 1;
2963 if (!strcmp (opt, "yes")) {
2964 allow_synchronous_major = TRUE;
2965 } else if (!strcmp (opt, "no")) {
2966 allow_synchronous_major = FALSE;
2968 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Using default value.", "`allow-synchronous-major` must be either `yes' or `no'.");
2973 if (!strcmp (opt, "cementing")) {
2974 cement_enabled = TRUE;
2977 if (!strcmp (opt, "no-cementing")) {
2978 cement_enabled = FALSE;
2982 if (major_collector.handle_gc_param && major_collector.handle_gc_param (opt))
2985 if (sgen_minor_collector.handle_gc_param && sgen_minor_collector.handle_gc_param (opt))
2988 if (sgen_client_handle_gc_param (opt))
2991 sgen_env_var_error (MONO_GC_PARAMS_NAME, "Ignoring.", "Unknown option `%s`.", opt);
2996 fprintf (stderr, "\n%s must be a comma-delimited list of one or more of the following:\n", MONO_GC_PARAMS_NAME);
2997 fprintf (stderr, " max-heap-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
2998 fprintf (stderr, " soft-heap-limit=n (where N is an integer, possibly with a k, m or a g suffix)\n");
2999 fprintf (stderr, " nursery-size=N (where N is an integer, possibly with a k, m or a g suffix)\n");
3000 fprintf (stderr, " major=COLLECTOR (where COLLECTOR is `marksweep', `marksweep-conc', `marksweep-par')\n");
3001 fprintf (stderr, " minor=COLLECTOR (where COLLECTOR is `simple' or `split')\n");
3002 fprintf (stderr, " wbarrier=WBARRIER (where WBARRIER is `remset' or `cardtable')\n");
3003 fprintf (stderr, " [no-]cementing\n");
3004 if (major_collector.is_concurrent)
3005 fprintf (stderr, " allow-synchronous-major=FLAG (where FLAG is `yes' or `no')\n");
3006 if (major_collector.print_gc_param_usage)
3007 major_collector.print_gc_param_usage ();
3008 if (sgen_minor_collector.print_gc_param_usage)
3009 sgen_minor_collector.print_gc_param_usage ();
3010 sgen_client_print_gc_params_usage ();
3011 fprintf (stderr, " Experimental options:\n");
3012 fprintf (stderr, " save-target-ratio=R (where R must be between %.2f - %.2f).\n", SGEN_MIN_SAVE_TARGET_RATIO, SGEN_MAX_SAVE_TARGET_RATIO);
3013 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);
3014 fprintf (stderr, "\n");
3016 usage_printed = TRUE;
3021 if (major_collector_opt)
3022 g_free (major_collector_opt);
3024 if (minor_collector_opt)
3025 g_free (minor_collector_opt);
3029 sgen_cement_init (cement_enabled);
3031 if ((env = g_getenv (MONO_GC_DEBUG_NAME))) {
3032 gboolean usage_printed = FALSE;
3034 opts = g_strsplit (env, ",", -1);
3035 for (ptr = opts; ptr && *ptr; ptr ++) {
3037 if (!strcmp (opt, ""))
3039 if (opt [0] >= '0' && opt [0] <= '9') {
3040 gc_debug_level = atoi (opt);
3045 char *rf = g_strdup_printf ("%s.%d", opt, mono_process_current_pid ());
3046 gc_debug_file = fopen (rf, "wb");
3048 gc_debug_file = stderr;
3051 } else if (!strcmp (opt, "print-allowance")) {
3052 debug_print_allowance = TRUE;
3053 } else if (!strcmp (opt, "print-pinning")) {
3054 sgen_pin_stats_enable ();
3055 } else if (!strcmp (opt, "verify-before-allocs")) {
3056 verify_before_allocs = 1;
3057 has_per_allocation_action = TRUE;
3058 } else if (g_str_has_prefix (opt, "verify-before-allocs=")) {
3059 char *arg = strchr (opt, '=') + 1;
3060 verify_before_allocs = atoi (arg);
3061 has_per_allocation_action = TRUE;
3062 } else if (!strcmp (opt, "collect-before-allocs")) {
3063 collect_before_allocs = 1;
3064 has_per_allocation_action = TRUE;
3065 } else if (g_str_has_prefix (opt, "collect-before-allocs=")) {
3066 char *arg = strchr (opt, '=') + 1;
3067 has_per_allocation_action = TRUE;
3068 collect_before_allocs = atoi (arg);
3069 } else if (!strcmp (opt, "verify-before-collections")) {
3070 whole_heap_check_before_collection = TRUE;
3071 } else if (!strcmp (opt, "check-at-minor-collections")) {
3072 consistency_check_at_minor_collection = TRUE;
3073 nursery_clear_policy = CLEAR_AT_GC;
3074 } else if (!strcmp (opt, "mod-union-consistency-check")) {
3075 if (!major_collector.is_concurrent) {
3076 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`mod-union-consistency-check` only works with concurrent major collector.");
3079 mod_union_consistency_check = TRUE;
3080 } else if (!strcmp (opt, "check-mark-bits")) {
3081 check_mark_bits_after_major_collection = TRUE;
3082 } else if (!strcmp (opt, "check-nursery-pinned")) {
3083 check_nursery_objects_pinned = TRUE;
3084 } else if (!strcmp (opt, "clear-at-gc")) {
3085 nursery_clear_policy = CLEAR_AT_GC;
3086 } else if (!strcmp (opt, "clear-nursery-at-gc")) {
3087 nursery_clear_policy = CLEAR_AT_GC;
3088 } else if (!strcmp (opt, "clear-at-tlab-creation")) {
3089 nursery_clear_policy = CLEAR_AT_TLAB_CREATION;
3090 } else if (!strcmp (opt, "debug-clear-at-tlab-creation")) {
3091 nursery_clear_policy = CLEAR_AT_TLAB_CREATION_DEBUG;
3092 } else if (!strcmp (opt, "check-scan-starts")) {
3093 do_scan_starts_check = TRUE;
3094 } else if (!strcmp (opt, "verify-nursery-at-minor-gc")) {
3095 do_verify_nursery = TRUE;
3096 } else if (!strcmp (opt, "check-concurrent")) {
3097 if (!major_collector.is_concurrent) {
3098 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "`check-concurrent` only works with concurrent major collectors.");
3101 nursery_clear_policy = CLEAR_AT_GC;
3102 do_concurrent_checks = TRUE;
3103 } else if (!strcmp (opt, "dump-nursery-at-minor-gc")) {
3104 do_dump_nursery_content = TRUE;
3105 } else if (!strcmp (opt, "disable-minor")) {
3106 disable_minor_collections = TRUE;
3107 } else if (!strcmp (opt, "disable-major")) {
3108 disable_major_collections = TRUE;
3109 } else if (g_str_has_prefix (opt, "heap-dump=")) {
3110 char *filename = strchr (opt, '=') + 1;
3111 nursery_clear_policy = CLEAR_AT_GC;
3112 sgen_debug_enable_heap_dump (filename);
3113 } else if (g_str_has_prefix (opt, "binary-protocol=")) {
3114 char *filename = strchr (opt, '=') + 1;
3115 char *colon = strrchr (filename, ':');
3118 if (!mono_gc_parse_environment_string_extract_number (colon + 1, &limit)) {
3119 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring limit.", "Binary protocol file size limit must be an integer.");
3124 binary_protocol_init (filename, (long long)limit);
3125 } else if (!strcmp (opt, "nursery-canaries")) {
3126 do_verify_nursery = TRUE;
3127 enable_nursery_canaries = TRUE;
3128 } else if (!sgen_client_handle_gc_debug (opt)) {
3129 sgen_env_var_error (MONO_GC_DEBUG_NAME, "Ignoring.", "Unknown option `%s`.", opt);
3134 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);
3135 fprintf (stderr, "Valid <option>s are:\n");
3136 fprintf (stderr, " collect-before-allocs[=<n>]\n");
3137 fprintf (stderr, " verify-before-allocs[=<n>]\n");
3138 fprintf (stderr, " check-at-minor-collections\n");
3139 fprintf (stderr, " check-mark-bits\n");
3140 fprintf (stderr, " check-nursery-pinned\n");
3141 fprintf (stderr, " verify-before-collections\n");
3142 fprintf (stderr, " verify-nursery-at-minor-gc\n");
3143 fprintf (stderr, " dump-nursery-at-minor-gc\n");
3144 fprintf (stderr, " disable-minor\n");
3145 fprintf (stderr, " disable-major\n");
3146 fprintf (stderr, " check-concurrent\n");
3147 fprintf (stderr, " clear-[nursery-]at-gc\n");
3148 fprintf (stderr, " clear-at-tlab-creation\n");
3149 fprintf (stderr, " debug-clear-at-tlab-creation\n");
3150 fprintf (stderr, " check-scan-starts\n");
3151 fprintf (stderr, " print-allowance\n");
3152 fprintf (stderr, " print-pinning\n");
3153 fprintf (stderr, " heap-dump=<filename>\n");
3154 fprintf (stderr, " binary-protocol=<filename>[:<file-size-limit>]\n");
3155 fprintf (stderr, " nursery-canaries\n");
3156 sgen_client_print_gc_debug_usage ();
3157 fprintf (stderr, "\n");
3159 usage_printed = TRUE;
3165 if (check_mark_bits_after_major_collection)
3166 nursery_clear_policy = CLEAR_AT_GC;
3168 if (major_collector.post_param_init)
3169 major_collector.post_param_init (&major_collector);
3171 if (major_collector.needs_thread_pool)
3172 sgen_workers_init (1);
3174 sgen_memgov_init (max_heap, soft_limit, debug_print_allowance, allowance_ratio, save_target);
3176 memset (&remset, 0, sizeof (remset));
3178 sgen_card_table_init (&remset);
3180 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");
3186 sgen_get_nursery_clear_policy (void)
3188 return nursery_clear_policy;
3194 mono_coop_mutex_lock (&gc_mutex);
3198 sgen_gc_unlock (void)
3200 gboolean try_free = sgen_try_free_some_memory;
3201 sgen_try_free_some_memory = FALSE;
3202 mono_coop_mutex_unlock (&gc_mutex);
3204 mono_thread_hazardous_try_free_some ();
3208 sgen_major_collector_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
3210 major_collector.iterate_live_block_ranges (callback);
3214 sgen_get_major_collector (void)
3216 return &major_collector;
3220 sgen_get_remset (void)
3226 count_cards (long long *major_total, long long *major_marked, long long *los_total, long long *los_marked)
3228 sgen_get_major_collector ()->count_cards (major_total, major_marked);
3229 sgen_los_count_cards (los_total, los_marked);
3232 static gboolean world_is_stopped = FALSE;
3234 /* LOCKING: assumes the GC lock is held */
3236 sgen_stop_world (int generation)
3238 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3240 SGEN_ASSERT (0, !world_is_stopped, "Why are we stopping a stopped world?");
3242 binary_protocol_world_stopping (generation, sgen_timestamp (), (gpointer)mono_native_thread_id_get ());
3244 sgen_client_stop_world (generation);
3246 world_is_stopped = TRUE;
3248 if (binary_protocol_is_heavy_enabled ())
3249 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3250 binary_protocol_world_stopped (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3253 /* LOCKING: assumes the GC lock is held */
3255 sgen_restart_world (int generation, GGTimingInfo *timing)
3257 long long major_total = -1, major_marked = -1, los_total = -1, los_marked = -1;
3259 SGEN_ASSERT (0, world_is_stopped, "Why are we restarting a running world?");
3261 if (binary_protocol_is_heavy_enabled ())
3262 count_cards (&major_total, &major_marked, &los_total, &los_marked);
3263 binary_protocol_world_restarting (generation, sgen_timestamp (), major_total, major_marked, los_total, los_marked);
3265 sgen_client_restart_world (generation, timing);
3267 world_is_stopped = FALSE;
3269 binary_protocol_world_restarted (generation, sgen_timestamp ());
3271 sgen_try_free_some_memory = TRUE;
3273 if (sgen_client_bridge_need_processing ())
3274 sgen_client_bridge_processing_finish (generation);
3276 sgen_memgov_collection_end (generation, timing, timing ? 2 : 0);
3280 sgen_is_world_stopped (void)
3282 return world_is_stopped;
3286 sgen_check_whole_heap_stw (void)
3288 sgen_stop_world (0);
3289 sgen_clear_nursery_fragments ();
3290 sgen_check_whole_heap (FALSE);
3291 sgen_restart_world (0, NULL);
3295 sgen_timestamp (void)
3297 SGEN_TV_DECLARE (timestamp);
3298 SGEN_TV_GETTIME (timestamp);
3299 return SGEN_TV_ELAPSED (sgen_init_timestamp, timestamp);
3302 #endif /* HAVE_SGEN_GC */