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.
60 #define TLAB_START (__thread_info__->tlab_start)
61 #define TLAB_NEXT (__thread_info__->tlab_next)
62 #define TLAB_TEMP_END (__thread_info__->tlab_temp_end)
63 #define TLAB_REAL_END (__thread_info__->tlab_real_end)
66 alloc_degraded (GCVTable vtable, size_t size, gboolean for_mature)
71 sgen_client_degraded_allocation (size);
72 SGEN_ATOMIC_ADD_P (degraded_mode, size);
73 sgen_ensure_free_space (size, GENERATION_OLD);
75 if (sgen_need_major_collection (size))
76 sgen_perform_collection (size, GENERATION_OLD, "mature allocation failure", !for_mature, TRUE);
80 p = major_collector.alloc_degraded (vtable, size);
83 binary_protocol_alloc_degraded (p, vtable, size, sgen_client_get_provenance ());
89 zero_tlab_if_necessary (void *p, size_t size)
91 if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION || nursery_clear_policy == CLEAR_AT_TLAB_CREATION_DEBUG) {
95 * This function is called for all allocations in
96 * TLABs. TLABs originate from fragments, which are
97 * initialized to be faux arrays. The remainder of
98 * the fragments are zeroed out at initialization for
99 * CLEAR_AT_GC, so here we just need to make sure that
100 * the array header is zeroed. Since we don't know
101 * whether we're called for the start of a fragment or
102 * for somewhere in between, we zero in any case, just
105 sgen_client_zero_array_fill_header (p, size);
110 * Provide a variant that takes just the vtable for small fixed-size objects.
111 * The aligned size is already computed and stored in vt->gc_descr.
112 * Note: every SGEN_SCAN_START_SIZE or so we are given the chance to do some special
113 * processing. We can keep track of where objects start, for example,
114 * so when we scan the thread stacks for pinned objects, we can start
115 * a search for the pinned object in SGEN_SCAN_START_SIZE chunks.
118 sgen_alloc_obj_nolock (GCVTable vtable, size_t size)
120 /* FIXME: handle OOM */
123 size_t real_size = size;
128 HEAVY_STAT (++stat_objects_alloced);
129 if (real_size <= SGEN_MAX_SMALL_OBJ_SIZE)
130 HEAVY_STAT (stat_bytes_alloced += size);
132 HEAVY_STAT (stat_bytes_alloced_los += size);
134 size = ALIGN_UP (size);
136 SGEN_ASSERT (6, sgen_vtable_get_descriptor (vtable), "VTable without descriptor");
138 if (G_UNLIKELY (has_per_allocation_action)) {
139 static int alloc_count;
140 int current_alloc = InterlockedIncrement (&alloc_count);
142 if (collect_before_allocs) {
143 if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
144 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE, TRUE);
145 if (!degraded_mode && sgen_can_alloc_size (size) && real_size <= SGEN_MAX_SMALL_OBJ_SIZE) {
147 g_assert_not_reached ();
150 } else if (verify_before_allocs) {
151 if ((current_alloc % verify_before_allocs) == 0)
152 sgen_check_whole_heap_stw ();
157 * We must already have the lock here instead of after the
158 * fast path because we might be interrupted in the fast path
159 * (after confirming that new_next < TLAB_TEMP_END) by the GC,
160 * and we'll end up allocating an object in a fragment which
161 * no longer belongs to us.
163 * The managed allocator does not do this, but it's treated
164 * specially by the world-stopping code.
167 if (real_size > SGEN_MAX_SMALL_OBJ_SIZE) {
168 p = (void **)sgen_los_alloc_large_inner (vtable, ALIGN_UP (real_size));
170 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
172 p = (void**)TLAB_NEXT;
173 /* FIXME: handle overflow */
174 new_next = (char*)p + size;
175 TLAB_NEXT = new_next;
177 if (G_LIKELY (new_next < TLAB_TEMP_END)) {
180 CANARIFY_ALLOC(p,real_size);
181 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
182 binary_protocol_alloc (p , vtable, size, sgen_client_get_provenance ());
183 g_assert (*p == NULL);
184 mono_atomic_store_seq (p, vtable);
191 /* there are two cases: the object is too big or we run out of space in the TLAB */
192 /* we also reach here when the thread does its first allocation after a minor
193 * collection, since the tlab_ variables are initialized to NULL.
194 * there can be another case (from ORP), if we cooperate with the runtime a bit:
195 * objects that need finalizers can have the high bit set in their size
196 * so the above check fails and we can readily add the object to the queue.
197 * This avoids taking again the GC lock when registering, but this is moot when
198 * doing thread-local allocation, so it may not be a good idea.
200 if (TLAB_NEXT >= TLAB_REAL_END) {
201 int available_in_tlab;
203 * Run out of space in the TLAB. When this happens, some amount of space
204 * remains in the TLAB, but not enough to satisfy the current allocation
205 * request. Currently, we retire the TLAB in all cases, later we could
206 * keep it if the remaining space is above a treshold, and satisfy the
207 * allocation directly from the nursery.
210 /* when running in degraded mode, we continue allocing that way
211 * for a while, to decrease the number of useless nursery collections.
213 if (degraded_mode && degraded_mode < DEFAULT_NURSERY_SIZE)
214 return alloc_degraded (vtable, size, FALSE);
216 available_in_tlab = (int)(TLAB_REAL_END - TLAB_NEXT);//We'll never have tlabs > 2Gb
217 if (size > tlab_size || available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
218 /* Allocate directly from the nursery */
219 p = (void **)sgen_nursery_alloc (size);
222 * We couldn't allocate from the nursery, so we try
223 * collecting. Even after the collection, we might
224 * still not have enough memory to allocate the
225 * object. The reason will most likely be that we've
226 * run out of memory, but there is the theoretical
227 * possibility that other threads might have consumed
228 * the freed up memory ahead of us.
230 * What we do in this case is allocate degraded, i.e.,
231 * from the major heap.
233 * Ideally we'd like to detect the case of other
234 * threads allocating ahead of us and loop (if we
235 * always loop we will loop endlessly in the case of
238 sgen_ensure_free_space (real_size, GENERATION_NURSERY);
240 p = (void **)sgen_nursery_alloc (size);
243 return alloc_degraded (vtable, size, FALSE);
245 zero_tlab_if_necessary (p, size);
247 size_t alloc_size = 0;
249 SGEN_LOG (3, "Retire TLAB: %p-%p [%ld]", TLAB_START, TLAB_REAL_END, (long)(TLAB_REAL_END - TLAB_NEXT - size));
250 sgen_nursery_retire_region (p, available_in_tlab);
252 p = (void **)sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
254 /* See comment above in similar case. */
255 sgen_ensure_free_space (tlab_size, GENERATION_NURSERY);
257 p = (void **)sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
260 return alloc_degraded (vtable, size, FALSE);
262 /* Allocate a new TLAB from the current nursery fragment */
263 TLAB_START = (char*)p;
264 TLAB_NEXT = TLAB_START;
265 TLAB_REAL_END = TLAB_START + alloc_size;
266 TLAB_TEMP_END = TLAB_START + MIN (SGEN_SCAN_START_SIZE, alloc_size);
268 zero_tlab_if_necessary (TLAB_START, alloc_size);
270 /* Allocate from the TLAB */
271 p = (void **)TLAB_NEXT;
273 sgen_set_nursery_scan_start ((char*)p);
276 /* Reached tlab_temp_end */
278 /* record the scan start so we can find pinned objects more easily */
279 sgen_set_nursery_scan_start ((char*)p);
280 /* we just bump tlab_temp_end as well */
281 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
282 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
284 CANARIFY_ALLOC(p,real_size);
288 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
289 binary_protocol_alloc (p, vtable, size, sgen_client_get_provenance ());
290 mono_atomic_store_seq (p, vtable);
297 sgen_try_alloc_obj_nolock (GCVTable vtable, size_t size)
301 size_t real_size = size;
306 size = ALIGN_UP (size);
307 SGEN_ASSERT (9, real_size >= SGEN_CLIENT_MINIMUM_OBJECT_SIZE, "Object too small");
309 SGEN_ASSERT (6, sgen_vtable_get_descriptor (vtable), "VTable without descriptor");
311 if (real_size > SGEN_MAX_SMALL_OBJ_SIZE)
314 if (G_UNLIKELY (size > tlab_size)) {
315 /* Allocate directly from the nursery */
316 p = (void **)sgen_nursery_alloc (size);
319 sgen_set_nursery_scan_start ((char*)p);
321 /*FIXME we should use weak memory ops here. Should help specially on x86. */
322 zero_tlab_if_necessary (p, size);
324 int available_in_tlab;
326 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
328 p = (void**)TLAB_NEXT;
329 /* FIXME: handle overflow */
330 new_next = (char*)p + size;
332 real_end = TLAB_REAL_END;
333 available_in_tlab = (int)(real_end - (char*)p);//We'll never have tlabs > 2Gb
335 if (G_LIKELY (new_next < real_end)) {
336 TLAB_NEXT = new_next;
338 /* Second case, we overflowed temp end */
339 if (G_UNLIKELY (new_next >= TLAB_TEMP_END)) {
340 sgen_set_nursery_scan_start (new_next);
341 /* we just bump tlab_temp_end as well */
342 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
343 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
345 } else if (available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
346 /* Allocate directly from the nursery */
347 p = (void **)sgen_nursery_alloc (size);
351 zero_tlab_if_necessary (p, size);
353 size_t alloc_size = 0;
355 sgen_nursery_retire_region (p, available_in_tlab);
356 new_next = (char *)sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
357 p = (void**)new_next;
361 TLAB_START = (char*)new_next;
362 TLAB_NEXT = new_next + size;
363 TLAB_REAL_END = new_next + alloc_size;
364 TLAB_TEMP_END = new_next + MIN (SGEN_SCAN_START_SIZE, alloc_size);
365 sgen_set_nursery_scan_start ((char*)p);
367 zero_tlab_if_necessary (new_next, alloc_size);
371 HEAVY_STAT (++stat_objects_alloced);
372 HEAVY_STAT (stat_bytes_alloced += size);
374 CANARIFY_ALLOC(p,real_size);
375 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
376 binary_protocol_alloc (p, vtable, size, sgen_client_get_provenance ());
377 g_assert (*p == NULL); /* FIXME disable this in non debug builds */
379 mono_atomic_store_seq (p, vtable);
385 sgen_alloc_obj (GCVTable vtable, size_t size)
390 if (!SGEN_CAN_ALIGN_UP (size))
393 if (G_UNLIKELY (has_per_allocation_action)) {
394 static int alloc_count;
395 int current_alloc = InterlockedIncrement (&alloc_count);
397 if (verify_before_allocs) {
398 if ((current_alloc % verify_before_allocs) == 0) {
400 sgen_check_whole_heap_stw ();
404 if (collect_before_allocs) {
405 if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
407 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE, TRUE);
413 ENTER_CRITICAL_REGION;
414 res = sgen_try_alloc_obj_nolock (vtable, size);
416 EXIT_CRITICAL_REGION;
419 EXIT_CRITICAL_REGION;
422 res = sgen_alloc_obj_nolock (vtable, size);
428 * To be used for interned strings and possibly MonoThread, reflection handles.
429 * We may want to explicitly free these objects.
432 sgen_alloc_obj_pinned (GCVTable vtable, size_t size)
436 if (!SGEN_CAN_ALIGN_UP (size))
438 size = ALIGN_UP (size);
442 if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
443 /* large objects are always pinned anyway */
444 p = (GCObject *)sgen_los_alloc_large_inner (vtable, size);
446 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));
447 p = major_collector.alloc_small_pinned_obj (vtable, size, SGEN_VTABLE_HAS_REFERENCES (vtable));
450 SGEN_LOG (6, "Allocated pinned object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
451 binary_protocol_alloc_pinned (p, vtable, size, sgen_client_get_provenance ());
458 sgen_alloc_obj_mature (GCVTable vtable, size_t size)
462 if (!SGEN_CAN_ALIGN_UP (size))
464 size = ALIGN_UP (size);
467 res = alloc_degraded (vtable, size, TRUE);
474 * Clear the thread local TLAB variables for all threads.
477 sgen_clear_tlabs (void)
479 FOREACH_THREAD (info) {
480 /* A new TLAB will be allocated when the thread does its first allocation */
481 info->tlab_start = NULL;
482 info->tlab_next = NULL;
483 info->tlab_temp_end = NULL;
484 info->tlab_real_end = NULL;
489 sgen_init_allocator (void)
491 #ifdef HEAVY_STATISTICS
492 mono_counters_register ("# objects allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_alloced);
493 mono_counters_register ("bytes allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced);
494 mono_counters_register ("bytes allocated in LOS", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced_los);
498 #endif /*HAVE_SGEN_GC*/