2 * lock-free-alloc.c: Lock free allocator.
4 * (C) Copyright 2011 Novell, Inc
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sublicense, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice shall be
15 * included in all copies or substantial portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
18 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
20 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
21 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
22 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
23 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 * This is a simplified version of the lock-free allocator described in
29 * Scalable Lock-Free Dynamic Memory Allocation
30 * Maged M. Michael, PLDI 2004
32 * I could not get Michael's allocator working bug free under heavy
33 * stress tests. The paper doesn't provide correctness proof and after
34 * failing to formalize the ownership of descriptors I devised this
37 * Allocation within superblocks proceeds exactly like in Michael's
38 * allocator. The simplification is that a thread has to "acquire" a
39 * descriptor before it can allocate from its superblock. While it owns
40 * the descriptor no other thread can acquire and hence allocate from
41 * it. A consequence of this is that the ABA problem cannot occur, so
42 * we don't need the tag field and don't have to use 64 bit CAS.
44 * Descriptors are stored in two locations: The partial queue and the
45 * active field. They can only be in at most one of those at one time.
46 * If a thread wants to allocate, it needs to get a descriptor. It
47 * tries the active descriptor first, CASing it to NULL. If that
48 * doesn't work, it gets a descriptor out of the partial queue. Once it
49 * has the descriptor it owns it because it is not referenced anymore.
50 * It allocates a slot and then gives the descriptor back (unless it is
53 * Note that it is still possible that a slot is freed while an
54 * allocation is in progress from the same superblock. Ownership in
55 * this case is not complicated, though. If the block was FULL and the
56 * free set it to PARTIAL, the free now owns the block (because FULL
57 * blocks are not referenced from partial and active) and has to give it
58 * back. If the block was PARTIAL then the free doesn't own the block
59 * (because it's either still referenced, or an alloc owns it). A
60 * special case of this is that it has changed from PARTIAL to EMPTY and
61 * now needs to be retired. Technically, the free wouldn't have to do
62 * anything in this case because the first thing an alloc does when it
63 * gets ownership of a descriptor is to check whether it is EMPTY and
64 * retire it if that is the case. As an optimization, our free does try
65 * to acquire the descriptor (by CASing the active field, which, if it
66 * is lucky, points to that descriptor) and if it can do so, retire it.
67 * If it can't, it tries to retire other descriptors from the partial
68 * queue, so that we can be sure that even if no more allocations
69 * happen, descriptors are still retired. This is analogous to what
70 * Michael's allocator does.
72 * Another difference to Michael's allocator is not related to
73 * concurrency, however: We don't point from slots to descriptors.
74 * Instead we allocate superblocks aligned and point from the start of
75 * the superblock to the descriptor, so we only need one word of
76 * metadata per superblock.
78 * FIXME: Having more than one allocator per size class is probably
79 * buggy because it was never tested.
85 #include <mono/utils/atomic.h>
86 #ifdef SGEN_WITHOUT_MONO
87 #include <mono/sgen/sgen-gc.h>
88 #include <mono/sgen/sgen-client.h>
90 #include <mono/utils/mono-mmap.h>
92 #include <mono/utils/mono-membar.h>
93 #include <mono/utils/hazard-pointer.h>
94 #include <mono/utils/lock-free-queue.h>
96 #include <mono/utils/lock-free-alloc.h>
98 //#define DESC_AVAIL_DUMMY
115 typedef struct _MonoLockFreeAllocDescriptor Descriptor;
116 struct _MonoLockFreeAllocDescriptor {
117 MonoLockFreeQueueNode node;
118 MonoLockFreeAllocator *heap;
119 volatile Anchor anchor;
120 unsigned int slot_size;
121 unsigned int block_size;
122 unsigned int max_count;
124 #ifndef DESC_AVAIL_DUMMY
125 Descriptor * volatile next;
127 gboolean in_use; /* used for debugging only */
130 #define NUM_DESC_BATCH 64
132 static MONO_ALWAYS_INLINE gpointer
133 sb_header_for_addr (gpointer addr, size_t block_size)
135 return (gpointer)(((size_t)addr) & (~(block_size - 1)));
138 /* Taken from SGen */
141 prot_flags_for_activate (int activate)
143 unsigned long prot_flags = activate? MONO_MMAP_READ|MONO_MMAP_WRITE: MONO_MMAP_NONE;
144 return prot_flags | MONO_MMAP_PRIVATE | MONO_MMAP_ANON;
148 alloc_sb (Descriptor *desc)
150 static int pagesize = -1;
155 pagesize = mono_pagesize ();
157 sb_header = desc->block_size == pagesize ?
158 mono_valloc (NULL, desc->block_size, prot_flags_for_activate (TRUE)) :
159 mono_valloc_aligned (desc->block_size, desc->block_size, prot_flags_for_activate (TRUE));
161 g_assert (sb_header == sb_header_for_addr (sb_header, desc->block_size));
163 *(Descriptor**)sb_header = desc;
164 //g_print ("sb %p for %p\n", sb_header, desc);
166 return (char*)sb_header + LOCK_FREE_ALLOC_SB_HEADER_SIZE;
170 free_sb (gpointer sb, size_t block_size)
172 gpointer sb_header = sb_header_for_addr (sb, block_size);
173 g_assert ((char*)sb_header + LOCK_FREE_ALLOC_SB_HEADER_SIZE == sb);
174 mono_vfree (sb_header, block_size);
175 //g_print ("free sb %p\n", sb_header);
178 #ifndef DESC_AVAIL_DUMMY
179 static Descriptor * volatile desc_avail;
184 MonoThreadHazardPointers *hp = mono_hazard_pointer_get ();
190 desc = (Descriptor *) get_hazardous_pointer ((gpointer * volatile)&desc_avail, hp, 1);
192 Descriptor *next = desc->next;
193 success = (InterlockedCompareExchangePointer ((gpointer * volatile)&desc_avail, next, desc) == desc);
195 size_t desc_size = sizeof (Descriptor);
199 desc = (Descriptor *) mono_valloc (NULL, desc_size * NUM_DESC_BATCH, prot_flags_for_activate (TRUE));
201 /* Organize into linked list. */
203 for (i = 0; i < NUM_DESC_BATCH; ++i) {
204 Descriptor *next = (i == (NUM_DESC_BATCH - 1)) ? NULL : (Descriptor*)((char*)desc + ((i + 1) * desc_size));
206 mono_lock_free_queue_node_init (&d->node, TRUE);
210 mono_memory_write_barrier ();
212 success = (InterlockedCompareExchangePointer ((gpointer * volatile)&desc_avail, desc->next, NULL) == NULL);
215 mono_vfree (desc, desc_size * NUM_DESC_BATCH);
218 mono_hazard_pointer_clear (hp, 1);
224 g_assert (!desc->in_use);
231 desc_enqueue_avail (gpointer _desc)
233 Descriptor *desc = (Descriptor *) _desc;
234 Descriptor *old_head;
236 g_assert (desc->anchor.data.state == STATE_EMPTY);
237 g_assert (!desc->in_use);
240 old_head = desc_avail;
241 desc->next = old_head;
242 mono_memory_write_barrier ();
243 } while (InterlockedCompareExchangePointer ((gpointer * volatile)&desc_avail, desc, old_head) != old_head);
247 desc_retire (Descriptor *desc)
249 g_assert (desc->anchor.data.state == STATE_EMPTY);
250 g_assert (desc->in_use);
251 desc->in_use = FALSE;
252 free_sb (desc->sb, desc->block_size);
253 mono_thread_hazardous_free_or_queue (desc, desc_enqueue_avail, HAZARD_FREE_NO_LOCK, HAZARD_FREE_ASYNC_CTX);
256 MonoLockFreeQueue available_descs;
261 Descriptor *desc = (Descriptor*)mono_lock_free_queue_dequeue (&available_descs);
266 return calloc (1, sizeof (Descriptor));
270 desc_retire (Descriptor *desc)
272 free_sb (desc->sb, desc->block_size);
273 mono_lock_free_queue_enqueue (&available_descs, &desc->node);
278 list_get_partial (MonoLockFreeAllocSizeClass *sc)
281 Descriptor *desc = (Descriptor*) mono_lock_free_queue_dequeue (&sc->partial);
284 if (desc->anchor.data.state != STATE_EMPTY)
291 desc_put_partial (gpointer _desc)
293 Descriptor *desc = (Descriptor *) _desc;
295 g_assert (desc->anchor.data.state != STATE_FULL);
297 mono_lock_free_queue_node_free (&desc->node);
298 mono_lock_free_queue_enqueue (&desc->heap->sc->partial, &desc->node);
302 list_put_partial (Descriptor *desc)
304 g_assert (desc->anchor.data.state != STATE_FULL);
305 mono_thread_hazardous_free_or_queue (desc, desc_put_partial, HAZARD_FREE_NO_LOCK, HAZARD_FREE_ASYNC_CTX);
309 list_remove_empty_desc (MonoLockFreeAllocSizeClass *sc)
311 int num_non_empty = 0;
313 Descriptor *desc = (Descriptor*) mono_lock_free_queue_dequeue (&sc->partial);
317 * We don't need to read atomically because we're the
318 * only thread that references this descriptor.
320 if (desc->anchor.data.state == STATE_EMPTY) {
323 g_assert (desc->heap->sc == sc);
324 mono_thread_hazardous_free_or_queue (desc, desc_put_partial, HAZARD_FREE_NO_LOCK, HAZARD_FREE_ASYNC_CTX);
325 if (++num_non_empty >= 2)
332 heap_get_partial (MonoLockFreeAllocator *heap)
334 return list_get_partial (heap->sc);
338 heap_put_partial (Descriptor *desc)
340 list_put_partial (desc);
344 set_anchor (Descriptor *desc, Anchor old_anchor, Anchor new_anchor)
346 if (old_anchor.data.state == STATE_EMPTY)
347 g_assert (new_anchor.data.state == STATE_EMPTY);
349 return InterlockedCompareExchange (&desc->anchor.value, new_anchor.value, old_anchor.value) == old_anchor.value;
353 alloc_from_active_or_partial (MonoLockFreeAllocator *heap)
356 Anchor old_anchor, new_anchor;
362 if (InterlockedCompareExchangePointer ((gpointer * volatile)&heap->active, NULL, desc) != desc)
365 desc = heap_get_partial (heap);
370 /* Now we own the desc. */
374 new_anchor = old_anchor = *(volatile Anchor*)&desc->anchor.value;
375 if (old_anchor.data.state == STATE_EMPTY) {
376 /* We must free it because we own it. */
380 g_assert (old_anchor.data.state == STATE_PARTIAL);
381 g_assert (old_anchor.data.count > 0);
383 addr = (char*)desc->sb + old_anchor.data.avail * desc->slot_size;
385 mono_memory_read_barrier ();
387 next = *(unsigned int*)addr;
388 g_assert (next < LOCK_FREE_ALLOC_SB_USABLE_SIZE (desc->block_size) / desc->slot_size);
390 new_anchor.data.avail = next;
391 --new_anchor.data.count;
393 if (new_anchor.data.count == 0)
394 new_anchor.data.state = STATE_FULL;
395 } while (!set_anchor (desc, old_anchor, new_anchor));
397 /* If the desc is partial we have to give it back. */
398 if (new_anchor.data.state == STATE_PARTIAL) {
399 if (InterlockedCompareExchangePointer ((gpointer * volatile)&heap->active, desc, NULL) != NULL)
400 heap_put_partial (desc);
407 alloc_from_new_sb (MonoLockFreeAllocator *heap)
409 unsigned int slot_size, block_size, count, i;
410 Descriptor *desc = desc_alloc ();
412 slot_size = desc->slot_size = heap->sc->slot_size;
413 block_size = desc->block_size = heap->sc->block_size;
414 count = LOCK_FREE_ALLOC_SB_USABLE_SIZE (block_size) / slot_size;
418 * Setting avail to 1 because 0 is the block we're allocating
421 desc->anchor.data.avail = 1;
422 desc->slot_size = heap->sc->slot_size;
423 desc->max_count = count;
425 desc->anchor.data.count = desc->max_count - 1;
426 desc->anchor.data.state = STATE_PARTIAL;
428 desc->sb = alloc_sb (desc);
430 /* Organize blocks into linked list. */
431 for (i = 1; i < count - 1; ++i)
432 *(unsigned int*)((char*)desc->sb + i * slot_size) = i + 1;
434 mono_memory_write_barrier ();
436 /* Make it active or free it again. */
437 if (InterlockedCompareExchangePointer ((gpointer * volatile)&heap->active, desc, NULL) == NULL) {
440 desc->anchor.data.state = STATE_EMPTY;
447 mono_lock_free_alloc (MonoLockFreeAllocator *heap)
453 addr = alloc_from_active_or_partial (heap);
457 addr = alloc_from_new_sb (heap);
466 mono_lock_free_free (gpointer ptr, size_t block_size)
468 Anchor old_anchor, new_anchor;
471 MonoLockFreeAllocator *heap = NULL;
473 desc = *(Descriptor**) sb_header_for_addr (ptr, block_size);
474 g_assert (block_size == desc->block_size);
479 new_anchor = old_anchor = *(volatile Anchor*)&desc->anchor.value;
480 *(unsigned int*)ptr = old_anchor.data.avail;
481 new_anchor.data.avail = ((char*)ptr - (char*)sb) / desc->slot_size;
482 g_assert (new_anchor.data.avail < LOCK_FREE_ALLOC_SB_USABLE_SIZE (block_size) / desc->slot_size);
484 if (old_anchor.data.state == STATE_FULL)
485 new_anchor.data.state = STATE_PARTIAL;
487 if (++new_anchor.data.count == desc->max_count) {
489 new_anchor.data.state = STATE_EMPTY;
491 } while (!set_anchor (desc, old_anchor, new_anchor));
493 if (new_anchor.data.state == STATE_EMPTY) {
494 g_assert (old_anchor.data.state != STATE_EMPTY);
496 if (InterlockedCompareExchangePointer ((gpointer * volatile)&heap->active, NULL, desc) == desc) {
497 /* We own it, so we free it. */
501 * Somebody else must free it, so we do some
502 * freeing for others.
504 list_remove_empty_desc (heap->sc);
506 } else if (old_anchor.data.state == STATE_FULL) {
508 * Nobody owned it, now we do, so we need to give it
512 g_assert (new_anchor.data.state == STATE_PARTIAL);
514 if (InterlockedCompareExchangePointer ((gpointer * volatile)&desc->heap->active, desc, NULL) != NULL)
515 heap_put_partial (desc);
519 #define g_assert_OR_PRINT(c, format, ...) do { \
522 g_print ((format), ## __VA_ARGS__); \
529 descriptor_check_consistency (Descriptor *desc, gboolean print)
531 int count = desc->anchor.data.count;
532 int max_count = LOCK_FREE_ALLOC_SB_USABLE_SIZE (desc->block_size) / desc->slot_size;
534 gboolean* linked = alloca(max_count*sizeof(gboolean));
536 gboolean linked [max_count];
541 #ifndef DESC_AVAIL_DUMMY
544 for (avail = desc_avail; avail; avail = avail->next)
545 g_assert_OR_PRINT (desc != avail, "descriptor is in the available list\n");
548 g_assert_OR_PRINT (desc->slot_size == desc->heap->sc->slot_size, "slot size doesn't match size class\n");
551 g_print ("descriptor %p is ", desc);
553 switch (desc->anchor.data.state) {
557 g_assert_OR_PRINT (count == 0, "count is not zero: %d\n", count);
561 g_print ("partial\n");
562 g_assert_OR_PRINT (count < max_count, "count too high: is %d but must be below %d\n", count, max_count);
567 g_assert_OR_PRINT (count == max_count, "count is wrong: is %d but should be %d\n", count, max_count);
570 g_assert_OR_PRINT (FALSE, "invalid state\n");
573 for (i = 0; i < max_count; ++i)
576 index = desc->anchor.data.avail;
578 for (i = 0; i < count; ++i) {
579 gpointer addr = (char*)desc->sb + index * desc->slot_size;
580 g_assert_OR_PRINT (index >= 0 && index < max_count,
581 "index %d for %dth available slot, linked from %d, not in range [0 .. %d)\n",
582 index, i, last, max_count);
583 g_assert_OR_PRINT (!linked [index], "%dth available slot %d linked twice\n", i, index);
586 linked [index] = TRUE;
588 index = *(unsigned int*)addr;
593 mono_lock_free_allocator_check_consistency (MonoLockFreeAllocator *heap)
595 Descriptor *active = heap->active;
598 g_assert (active->anchor.data.state == STATE_PARTIAL);
599 descriptor_check_consistency (active, FALSE);
601 while ((desc = (Descriptor*)mono_lock_free_queue_dequeue (&heap->sc->partial))) {
602 g_assert (desc->anchor.data.state == STATE_PARTIAL || desc->anchor.data.state == STATE_EMPTY);
603 descriptor_check_consistency (desc, FALSE);
609 mono_lock_free_allocator_init_size_class (MonoLockFreeAllocSizeClass *sc, unsigned int slot_size, unsigned int block_size)
611 g_assert (block_size > 0);
612 g_assert ((block_size & (block_size - 1)) == 0); /* check if power of 2 */
613 g_assert (slot_size * 2 <= LOCK_FREE_ALLOC_SB_USABLE_SIZE (block_size));
615 mono_lock_free_queue_init (&sc->partial);
616 sc->slot_size = slot_size;
617 sc->block_size = block_size;
621 mono_lock_free_allocator_init_allocator (MonoLockFreeAllocator *heap, MonoLockFreeAllocSizeClass *sc)