2 * sgen-alloc.c: Object allocation routines + managed allocators
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)
10 * Copyright 2011 Xamarin, Inc.
11 * Copyright (C) 2012 Xamarin Inc
13 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
17 * ######################################################################
18 * ######## Object allocation
19 * ######################################################################
20 * This section of code deals with allocating memory for objects.
21 * There are several ways:
22 * *) allocate large objects
23 * *) allocate normal objects
24 * *) fast lock-free allocation
25 * *) allocation of pinned objects
33 #include "mono/sgen/sgen-gc.h"
34 #include "mono/sgen/sgen-protocol.h"
35 #include "mono/sgen/sgen-memory-governor.h"
36 #include "mono/sgen/sgen-client.h"
37 #include "mono/utils/mono-memory-model.h"
39 #define ALIGN_UP SGEN_ALIGN_UP
40 #define ALLOC_ALIGN SGEN_ALLOC_ALIGN
41 #define MAX_SMALL_OBJ_SIZE SGEN_MAX_SMALL_OBJ_SIZE
43 #ifdef HEAVY_STATISTICS
44 static guint64 stat_objects_alloced = 0;
45 static guint64 stat_bytes_alloced = 0;
46 static guint64 stat_bytes_alloced_los = 0;
51 * Allocation is done from a Thread Local Allocation Buffer (TLAB). TLABs are allocated
52 * from nursery fragments.
53 * tlab_next is the pointer to the space inside the TLAB where the next object will
55 * tlab_temp_end is the pointer to the end of the temporary space reserved for
56 * the allocation: it allows us to set the scan starts at reasonable intervals.
57 * tlab_real_end points to the end of the TLAB.
61 * FIXME: What is faster, a TLS variable pointing to a structure, or separate TLS
62 * variables for next+temp_end ?
65 static __thread char *tlab_start;
66 static __thread char *tlab_next;
67 static __thread char *tlab_temp_end;
68 static __thread char *tlab_real_end;
69 /* Used by the managed allocator/wbarrier */
70 static __thread char **tlab_next_addr MONO_ATTR_USED;
74 #define TLAB_START tlab_start
75 #define TLAB_NEXT tlab_next
76 #define TLAB_TEMP_END tlab_temp_end
77 #define TLAB_REAL_END tlab_real_end
79 #define TLAB_START (__thread_info__->tlab_start)
80 #define TLAB_NEXT (__thread_info__->tlab_next)
81 #define TLAB_TEMP_END (__thread_info__->tlab_temp_end)
82 #define TLAB_REAL_END (__thread_info__->tlab_real_end)
86 alloc_degraded (GCVTable vtable, size_t size, gboolean for_mature)
91 sgen_client_degraded_allocation (size);
92 SGEN_ATOMIC_ADD_P (degraded_mode, size);
93 sgen_ensure_free_space (size, GENERATION_OLD);
95 if (sgen_need_major_collection (size))
96 sgen_perform_collection (size, GENERATION_OLD, "mature allocation failure", !for_mature);
100 p = major_collector.alloc_degraded (vtable, size);
103 binary_protocol_alloc_degraded (p, vtable, size, sgen_client_get_provenance ());
109 zero_tlab_if_necessary (void *p, size_t size)
111 if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION || nursery_clear_policy == CLEAR_AT_TLAB_CREATION_DEBUG) {
115 * This function is called for all allocations in
116 * TLABs. TLABs originate from fragments, which are
117 * initialized to be faux arrays. The remainder of
118 * the fragments are zeroed out at initialization for
119 * CLEAR_AT_GC, so here we just need to make sure that
120 * the array header is zeroed. Since we don't know
121 * whether we're called for the start of a fragment or
122 * for somewhere in between, we zero in any case, just
125 sgen_client_zero_array_fill_header (p, size);
130 * Provide a variant that takes just the vtable for small fixed-size objects.
131 * The aligned size is already computed and stored in vt->gc_descr.
132 * Note: every SGEN_SCAN_START_SIZE or so we are given the chance to do some special
133 * processing. We can keep track of where objects start, for example,
134 * so when we scan the thread stacks for pinned objects, we can start
135 * a search for the pinned object in SGEN_SCAN_START_SIZE chunks.
138 sgen_alloc_obj_nolock (GCVTable vtable, size_t size)
140 /* FIXME: handle OOM */
143 size_t real_size = size;
148 HEAVY_STAT (++stat_objects_alloced);
149 if (real_size <= SGEN_MAX_SMALL_OBJ_SIZE)
150 HEAVY_STAT (stat_bytes_alloced += size);
152 HEAVY_STAT (stat_bytes_alloced_los += size);
154 size = ALIGN_UP (size);
156 SGEN_ASSERT (6, sgen_vtable_get_descriptor (vtable), "VTable without descriptor");
158 if (G_UNLIKELY (has_per_allocation_action)) {
159 static int alloc_count;
160 int current_alloc = InterlockedIncrement (&alloc_count);
162 if (collect_before_allocs) {
163 if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
164 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE);
165 if (!degraded_mode && sgen_can_alloc_size (size) && real_size <= SGEN_MAX_SMALL_OBJ_SIZE) {
167 g_assert_not_reached ();
170 } else if (verify_before_allocs) {
171 if ((current_alloc % verify_before_allocs) == 0)
172 sgen_check_whole_heap_stw ();
177 * We must already have the lock here instead of after the
178 * fast path because we might be interrupted in the fast path
179 * (after confirming that new_next < TLAB_TEMP_END) by the GC,
180 * and we'll end up allocating an object in a fragment which
181 * no longer belongs to us.
183 * The managed allocator does not do this, but it's treated
184 * specially by the world-stopping code.
187 if (real_size > SGEN_MAX_SMALL_OBJ_SIZE) {
188 p = (void **)sgen_los_alloc_large_inner (vtable, ALIGN_UP (real_size));
190 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
192 p = (void**)TLAB_NEXT;
193 /* FIXME: handle overflow */
194 new_next = (char*)p + size;
195 TLAB_NEXT = new_next;
197 if (G_LIKELY (new_next < TLAB_TEMP_END)) {
201 * FIXME: We might need a memory barrier here so the change to tlab_next is
202 * visible before the vtable store.
205 CANARIFY_ALLOC(p,real_size);
206 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
207 binary_protocol_alloc (p , vtable, size, sgen_client_get_provenance ());
208 g_assert (*p == NULL);
209 mono_atomic_store_seq (p, vtable);
216 /* there are two cases: the object is too big or we run out of space in the TLAB */
217 /* we also reach here when the thread does its first allocation after a minor
218 * collection, since the tlab_ variables are initialized to NULL.
219 * there can be another case (from ORP), if we cooperate with the runtime a bit:
220 * objects that need finalizers can have the high bit set in their size
221 * so the above check fails and we can readily add the object to the queue.
222 * This avoids taking again the GC lock when registering, but this is moot when
223 * doing thread-local allocation, so it may not be a good idea.
225 if (TLAB_NEXT >= TLAB_REAL_END) {
226 int available_in_tlab;
228 * Run out of space in the TLAB. When this happens, some amount of space
229 * remains in the TLAB, but not enough to satisfy the current allocation
230 * request. Currently, we retire the TLAB in all cases, later we could
231 * keep it if the remaining space is above a treshold, and satisfy the
232 * allocation directly from the nursery.
235 /* when running in degraded mode, we continue allocing that way
236 * for a while, to decrease the number of useless nursery collections.
238 if (degraded_mode && degraded_mode < DEFAULT_NURSERY_SIZE)
239 return alloc_degraded (vtable, size, FALSE);
241 available_in_tlab = (int)(TLAB_REAL_END - TLAB_NEXT);//We'll never have tlabs > 2Gb
242 if (size > tlab_size || available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
243 /* Allocate directly from the nursery */
244 p = (void **)sgen_nursery_alloc (size);
247 * We couldn't allocate from the nursery, so we try
248 * collecting. Even after the collection, we might
249 * still not have enough memory to allocate the
250 * object. The reason will most likely be that we've
251 * run out of memory, but there is the theoretical
252 * possibility that other threads might have consumed
253 * the freed up memory ahead of us.
255 * What we do in this case is allocate degraded, i.e.,
256 * from the major heap.
258 * Ideally we'd like to detect the case of other
259 * threads allocating ahead of us and loop (if we
260 * always loop we will loop endlessly in the case of
263 sgen_ensure_free_space (real_size, GENERATION_NURSERY);
265 p = (void **)sgen_nursery_alloc (size);
268 return alloc_degraded (vtable, size, FALSE);
270 zero_tlab_if_necessary (p, size);
272 size_t alloc_size = 0;
274 SGEN_LOG (3, "Retire TLAB: %p-%p [%ld]", TLAB_START, TLAB_REAL_END, (long)(TLAB_REAL_END - TLAB_NEXT - size));
275 sgen_nursery_retire_region (p, available_in_tlab);
277 p = (void **)sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
279 /* See comment above in similar case. */
280 sgen_ensure_free_space (tlab_size, GENERATION_NURSERY);
282 p = (void **)sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
285 return alloc_degraded (vtable, size, FALSE);
287 /* Allocate a new TLAB from the current nursery fragment */
288 TLAB_START = (char*)p;
289 TLAB_NEXT = TLAB_START;
290 TLAB_REAL_END = TLAB_START + alloc_size;
291 TLAB_TEMP_END = TLAB_START + MIN (SGEN_SCAN_START_SIZE, alloc_size);
293 zero_tlab_if_necessary (TLAB_START, alloc_size);
295 /* Allocate from the TLAB */
296 p = (void **)TLAB_NEXT;
298 sgen_set_nursery_scan_start ((char*)p);
301 /* Reached tlab_temp_end */
303 /* record the scan start so we can find pinned objects more easily */
304 sgen_set_nursery_scan_start ((char*)p);
305 /* we just bump tlab_temp_end as well */
306 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
307 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
309 CANARIFY_ALLOC(p,real_size);
313 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
314 binary_protocol_alloc (p, vtable, size, sgen_client_get_provenance ());
315 mono_atomic_store_seq (p, vtable);
322 sgen_try_alloc_obj_nolock (GCVTable vtable, size_t size)
326 size_t real_size = size;
331 size = ALIGN_UP (size);
332 SGEN_ASSERT (9, real_size >= SGEN_CLIENT_MINIMUM_OBJECT_SIZE, "Object too small");
334 SGEN_ASSERT (6, sgen_vtable_get_descriptor (vtable), "VTable without descriptor");
336 if (real_size > SGEN_MAX_SMALL_OBJ_SIZE)
339 if (G_UNLIKELY (size > tlab_size)) {
340 /* Allocate directly from the nursery */
341 p = (void **)sgen_nursery_alloc (size);
344 sgen_set_nursery_scan_start ((char*)p);
346 /*FIXME we should use weak memory ops here. Should help specially on x86. */
347 zero_tlab_if_necessary (p, size);
349 int available_in_tlab;
351 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
353 p = (void**)TLAB_NEXT;
354 /* FIXME: handle overflow */
355 new_next = (char*)p + size;
357 real_end = TLAB_REAL_END;
358 available_in_tlab = (int)(real_end - (char*)p);//We'll never have tlabs > 2Gb
360 if (G_LIKELY (new_next < real_end)) {
361 TLAB_NEXT = new_next;
363 /* Second case, we overflowed temp end */
364 if (G_UNLIKELY (new_next >= TLAB_TEMP_END)) {
365 sgen_set_nursery_scan_start (new_next);
366 /* we just bump tlab_temp_end as well */
367 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
368 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
370 } else if (available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
371 /* Allocate directly from the nursery */
372 p = (void **)sgen_nursery_alloc (size);
376 zero_tlab_if_necessary (p, size);
378 size_t alloc_size = 0;
380 sgen_nursery_retire_region (p, available_in_tlab);
381 new_next = (char *)sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
382 p = (void**)new_next;
386 TLAB_START = (char*)new_next;
387 TLAB_NEXT = new_next + size;
388 TLAB_REAL_END = new_next + alloc_size;
389 TLAB_TEMP_END = new_next + MIN (SGEN_SCAN_START_SIZE, alloc_size);
390 sgen_set_nursery_scan_start ((char*)p);
392 zero_tlab_if_necessary (new_next, alloc_size);
396 HEAVY_STAT (++stat_objects_alloced);
397 HEAVY_STAT (stat_bytes_alloced += size);
399 CANARIFY_ALLOC(p,real_size);
400 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
401 binary_protocol_alloc (p, vtable, size, sgen_client_get_provenance ());
402 g_assert (*p == NULL); /* FIXME disable this in non debug builds */
404 mono_atomic_store_seq (p, vtable);
410 sgen_alloc_obj (GCVTable vtable, size_t size)
415 if (!SGEN_CAN_ALIGN_UP (size))
418 if (G_UNLIKELY (has_per_allocation_action)) {
419 static int alloc_count;
420 int current_alloc = InterlockedIncrement (&alloc_count);
422 if (verify_before_allocs) {
423 if ((current_alloc % verify_before_allocs) == 0)
424 sgen_check_whole_heap_stw ();
426 if (collect_before_allocs) {
427 if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
429 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE);
435 ENTER_CRITICAL_REGION;
436 res = sgen_try_alloc_obj_nolock (vtable, size);
438 EXIT_CRITICAL_REGION;
441 EXIT_CRITICAL_REGION;
444 res = sgen_alloc_obj_nolock (vtable, size);
450 * To be used for interned strings and possibly MonoThread, reflection handles.
451 * We may want to explicitly free these objects.
454 sgen_alloc_obj_pinned (GCVTable vtable, size_t size)
458 if (!SGEN_CAN_ALIGN_UP (size))
460 size = ALIGN_UP (size);
464 if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
465 /* large objects are always pinned anyway */
466 p = (GCObject *)sgen_los_alloc_large_inner (vtable, size);
468 SGEN_ASSERT (9, sgen_client_vtable_is_inited (vtable), "class %s:%s is not initialized", sgen_client_vtable_get_namespace (vtable), sgen_client_vtable_get_name (vtable));
469 p = major_collector.alloc_small_pinned_obj (vtable, size, SGEN_VTABLE_HAS_REFERENCES (vtable));
472 SGEN_LOG (6, "Allocated pinned object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
473 binary_protocol_alloc_pinned (p, vtable, size, sgen_client_get_provenance ());
480 sgen_alloc_obj_mature (GCVTable vtable, size_t size)
484 if (!SGEN_CAN_ALIGN_UP (size))
486 size = ALIGN_UP (size);
489 res = alloc_degraded (vtable, size, TRUE);
496 sgen_init_tlab_info (SgenThreadInfo* info)
498 #ifndef HAVE_KW_THREAD
499 SgenThreadInfo *__thread_info__ = info;
502 info->tlab_start_addr = &TLAB_START;
503 info->tlab_next_addr = &TLAB_NEXT;
504 info->tlab_temp_end_addr = &TLAB_TEMP_END;
505 info->tlab_real_end_addr = &TLAB_REAL_END;
507 #ifdef HAVE_KW_THREAD
508 tlab_next_addr = &tlab_next;
513 * Clear the thread local TLAB variables for all threads.
516 sgen_clear_tlabs (void)
518 FOREACH_THREAD (info) {
519 /* A new TLAB will be allocated when the thread does its first allocation */
520 *info->tlab_start_addr = NULL;
521 *info->tlab_next_addr = NULL;
522 *info->tlab_temp_end_addr = NULL;
523 *info->tlab_real_end_addr = NULL;
528 sgen_init_allocator (void)
530 #if defined(HAVE_KW_THREAD) && !defined(SGEN_WITHOUT_MONO)
531 int tlab_next_addr_offset = -1;
532 int tlab_temp_end_offset = -1;
535 MONO_THREAD_VAR_OFFSET (tlab_next_addr, tlab_next_addr_offset);
536 MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
538 mono_tls_key_set_offset (TLS_KEY_SGEN_TLAB_NEXT_ADDR, tlab_next_addr_offset);
539 mono_tls_key_set_offset (TLS_KEY_SGEN_TLAB_TEMP_END, tlab_temp_end_offset);
542 #ifdef HEAVY_STATISTICS
543 mono_counters_register ("# objects allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_alloced);
544 mono_counters_register ("bytes allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced);
545 mono_counters_register ("bytes allocated in LOS", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced_los);
549 #endif /*HAVE_SGEN_GC*/