1 /* GLIB sliced memory - fast concurrent memory chunk allocator
2 * Copyright (C) 2005 Tim Janik
4 * This library is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either
7 * version 2 of the License, or (at your option) any later version.
9 * This library is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * Lesser General Public License for more details.
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with this library; if not, write to the
16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17 * Boston, MA 02111-1307, USA.
20 #define _XOPEN_SOURCE 600 /* posix_memalign() */
21 #include <stdlib.h> /* posix_memalign() */
25 #include "gmem.h" /* gslice.h */
26 #include "gthreadinit.h"
30 #include <unistd.h> /* sysconf() */
37 /* the GSlice allocator is split up into 4 layers, roughly modelled after the slab
38 * allocator and magazine extensions as outlined in:
39 * + [Bonwick94] Jeff Bonwick, The slab allocator: An object-caching kernel
40 * memory allocator. USENIX 1994, http://citeseer.ist.psu.edu/bonwick94slab.html
41 * + [Bonwick01] Bonwick and Jonathan Adams, Magazines and vmem: Extending the
42 * slab allocator to many cpu's and arbitrary resources.
43 * USENIX 2001, http://citeseer.ist.psu.edu/bonwick01magazines.html
45 * - the thread magazines. for each (aligned) chunk size, a magazine (a list)
46 * of recently freed and soon to be allocated chunks is maintained per thread.
47 * this way, most alloc/free requests can be quickly satisfied from per-thread
48 * free lists which only require one g_private_get() call to retrive the
50 * - the magazine cache. allocating and freeing chunks to/from threads only
51 * occours at magazine sizes from a global depot of magazines. the depot
52 * maintaines a 15 second working set of allocated magazines, so full
53 * magazines are not allocated and released too often.
54 * the chunk size dependent magazine sizes automatically adapt (within limits,
55 * see [3]) to lock contention to properly scale performance across a variety
57 * - the slab allocator. this allocator allocates slabs (blocks of memory) close
58 * to the system page size or multiples thereof which have to be page aligned.
59 * the blocks are divided into smaller chunks which are used to satisfy
60 * allocations from the upper layers. the space provided by the reminder of
61 * the chunk size division is used for cache colorization (random distribution
62 * of chunk addresses) to improve processor cache utilization. multiple slabs
63 * with the same chunk size are kept in a partially sorted ring to allow O(1)
64 * freeing and allocation of chunks (as long as the allocation of an entirely
65 * new slab can be avoided).
66 * - the page allocator. on most modern systems, posix_memalign(3) or
67 * memalign(3) should be available, so this is used to allocate blocks with
68 * system page size based alignments and sizes or multiples thereof.
69 * if no memalign variant is provided, valloc() is used instead and
70 * block sizes are limited to the system page size (no multiples thereof).
71 * as a fallback, on system without even valloc(), a malloc(3)-based page
72 * allocator with alloc-only behaviour is used.
75 * [1] some systems memalign(3) implementations may rely on boundary tagging for
76 * the handed out memory chunks. to avoid excessive page-wise fragmentation,
77 * we reserve 2 * sizeof (void*) per block size for the systems memalign(3),
78 * specified in NATIVE_MALLOC_PADDING.
79 * [2] using the slab allocator alone already provides for a fast and efficient
80 * allocator, it doesn't properly scale beyond single-threaded uses though.
81 * also, the slab allocator implements eager free(3)-ing, i.e. does not
82 * provide any form of caching or working set maintenance. so if used alone,
83 * it's vulnerable to trashing for sequences of balanced (alloc, free) pairs
84 * at certain thresholds.
85 * [3] magazine sizes are bound by an implementation specific minimum size and
86 * a chunk size specific maximum to limit magazine storage sizes to roughly
88 * [4] allocating ca. 8 chunks per block/page keeps a good balance between
89 * external and internal fragmentation (<= 12.5%). [Bonwick94]
92 /* --- macros and constants --- */
93 #define LARGEALIGNMENT (256)
94 #define P2ALIGNMENT (2 * sizeof (gsize)) /* fits 2 pointers (assumed to be 2 * GLIB_SIZEOF_SIZE_T below) */
95 #define ALIGN(size, base) ((base) * (gsize) (((size) + (base) - 1) / (base)))
96 #define NATIVE_MALLOC_PADDING P2ALIGNMENT /* per-page padding left for native malloc(3) see [1] */
97 #define SLAB_INFO_SIZE P2ALIGN (sizeof (SlabInfo) + NATIVE_MALLOC_PADDING)
98 #define MAX_MAGAZINE_SIZE (256) /* see [3] and allocator_get_magazine_threshold() for this */
99 #define MIN_MAGAZINE_SIZE (4)
100 #define MAX_STAMP_COUNTER (7) /* distributes the load of gettimeofday() */
101 #define MAX_SLAB_CHUNK_SIZE(al) (((al)->max_page_size - SLAB_INFO_SIZE) / 8) /* we want at last 8 chunks per page, see [4] */
102 #define MAX_SLAB_INDEX(al) (SLAB_INDEX (al, MAX_SLAB_CHUNK_SIZE (al)) + 1)
103 #define SLAB_INDEX(al, asize) ((asize) / P2ALIGNMENT - 1) /* asize must be P2ALIGNMENT aligned */
104 #define SLAB_CHUNK_SIZE(al, ix) (((ix) + 1) * P2ALIGNMENT)
105 #define SLAB_BPAGE_SIZE(al,csz) (8 * (csz) + SLAB_INFO_SIZE)
107 /* optimized version of ALIGN (size, P2ALIGNMENT) */
108 #if GLIB_SIZEOF_SIZE_T * 2 == 8 /* P2ALIGNMENT */
109 #define P2ALIGN(size) (((size) + 0x7) & ~(gsize) 0x7)
110 #elif GLIB_SIZEOF_SIZE_T * 2 == 16 /* P2ALIGNMENT */
111 #define P2ALIGN(size) (((size) + 0xf) & ~(gsize) 0xf)
113 #define P2ALIGN(size) ALIGN (size, P2ALIGNMENT)
116 /* special helpers to avoid gmessage.c dependency */
117 static void mem_error (const char *format, ...) G_GNUC_PRINTF (1,2);
118 #define mem_assert(cond) do { if (G_LIKELY (cond)) ; else mem_error ("assertion failed: %s", #cond); } while (0)
120 /* --- structures --- */
121 typedef struct _ChunkLink ChunkLink;
122 typedef struct _SlabInfo SlabInfo;
123 typedef struct _CachedMagazine CachedMagazine;
131 SlabInfo *next, *prev;
135 gsize count; /* approximative chunks list length */
138 Magazine *magazine1; /* array of MAX_SLAB_INDEX (allocator) */
139 Magazine *magazine2; /* array of MAX_SLAB_INDEX (allocator) */
142 gboolean always_malloc;
143 gboolean bypass_magazines;
144 gsize working_set_msecs;
145 guint color_increment;
148 /* const after initialization */
149 gsize min_page_size, max_page_size;
151 gsize max_slab_chunk_size_for_magazine_cache;
153 GMutex *magazine_mutex;
154 ChunkLink **magazines; /* array of MAX_SLAB_INDEX (allocator) */
155 guint *contention_counters; /* array of MAX_SLAB_INDEX (allocator) */
161 SlabInfo **slab_stack; /* array of MAX_SLAB_INDEX (allocator) */
165 /* --- prototypes --- */
166 static gpointer slab_allocator_alloc_chunk (gsize chunk_size);
167 static void slab_allocator_free_chunk (gsize chunk_size,
169 static void private_thread_memory_cleanup (gpointer data);
170 static gpointer allocator_memalign (gsize alignment,
172 static void allocator_memfree (gsize memsize,
174 static inline void magazine_cache_update_stamp (void);
175 static inline gsize allocator_get_magazine_threshold (Allocator *allocator,
178 /* --- variables --- */
179 static GPrivate *private_thread_memory = NULL;
180 static gsize sys_page_size = 0;
181 static Allocator allocator[1] = { { 0, }, };
182 static SliceConfig slice_config = {
183 FALSE, /* always_malloc */
184 FALSE, /* bypass_magazines */
185 15 * 1000, /* working_set_msecs */
186 1, /* color increment, alt: 0x7fffffff */
189 /* --- auxillary funcitons --- */
191 g_slice_set_config (GSliceConfig ckey,
194 g_return_if_fail (sys_page_size == 0);
197 case G_SLICE_CONFIG_ALWAYS_MALLOC:
198 slice_config.always_malloc = value != 0;
200 case G_SLICE_CONFIG_BYPASS_MAGAZINES:
201 slice_config.bypass_magazines = value != 0;
203 case G_SLICE_CONFIG_WORKING_SET_MSECS:
204 slice_config.working_set_msecs = value;
206 case G_SLICE_CONFIG_COLOR_INCREMENT:
207 slice_config.color_increment = value;
213 g_slice_get_config (GSliceConfig ckey)
217 case G_SLICE_CONFIG_ALWAYS_MALLOC:
218 return slice_config.always_malloc;
219 case G_SLICE_CONFIG_BYPASS_MAGAZINES:
220 return slice_config.bypass_magazines;
221 case G_SLICE_CONFIG_WORKING_SET_MSECS:
222 return slice_config.working_set_msecs;
223 case G_SLICE_CONFIG_CHUNK_SIZES:
224 return MAX_SLAB_INDEX (allocator);
225 case G_SLICE_CONFIG_COLOR_INCREMENT:
226 return slice_config.color_increment;
233 g_slice_get_config_state (GSliceConfig ckey,
238 g_return_val_if_fail (n_values != NULL, NULL);
243 case G_SLICE_CONFIG_CONTENTION_COUNTER:
244 array[i++] = SLAB_CHUNK_SIZE (allocator, address);
245 array[i++] = allocator->contention_counters[address];
246 array[i++] = allocator_get_magazine_threshold (allocator, address);
248 return g_memdup (array, sizeof (array[0]) * *n_values);
255 g_slice_init_nomessage (void)
257 /* we may not use g_error() or friends here */
258 mem_assert (sys_page_size == 0);
259 mem_assert (MIN_MAGAZINE_SIZE >= 4);
263 SYSTEM_INFO system_info;
264 GetSystemInfo (&system_info);
265 sys_page_size = system_info.dwPageSize;
268 sys_page_size = sysconf (_SC_PAGESIZE); /* = sysconf (_SC_PAGE_SIZE); = getpagesize(); */
270 mem_assert (sys_page_size >= 2 * LARGEALIGNMENT);
271 mem_assert ((sys_page_size & (sys_page_size - 1)) == 0);
272 allocator->config = slice_config;
273 allocator->min_page_size = sys_page_size;
274 #if HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN
275 /* allow allocation of pages up to 8KB (with 8KB alignment).
276 * this is useful because many medium to large sized structures
277 * fit less than 8 times (see [4]) into 4KB pages.
278 * we allow very small page sizes here, to reduce wastage in
279 * threads if only small allocations are required (this does
280 * bear the risk of incresing allocation times and fragmentation
283 allocator->min_page_size = MAX (allocator->min_page_size, 4096);
284 allocator->max_page_size = MAX (allocator->min_page_size, 8192);
285 allocator->min_page_size = MIN (allocator->min_page_size, 128);
287 /* we can only align to system page size */
288 allocator->max_page_size = sys_page_size;
290 allocator->magazine_mutex = NULL; /* _g_slice_thread_init_nomessage() */
291 allocator->magazines = g_new0 (ChunkLink*, MAX_SLAB_INDEX (allocator));
292 allocator->contention_counters = g_new0 (guint, MAX_SLAB_INDEX (allocator));
293 allocator->mutex_counter = 0;
294 allocator->stamp_counter = MAX_STAMP_COUNTER; /* force initial update */
295 allocator->last_stamp = 0;
296 allocator->slab_mutex = NULL; /* _g_slice_thread_init_nomessage() */
297 allocator->slab_stack = g_new0 (SlabInfo*, MAX_SLAB_INDEX (allocator));
298 allocator->color_accu = 0;
299 magazine_cache_update_stamp();
300 /* values cached for performance reasons */
301 allocator->max_slab_chunk_size_for_magazine_cache = MAX_SLAB_CHUNK_SIZE (allocator);
302 if (allocator->config.always_malloc || allocator->config.bypass_magazines)
303 allocator->max_slab_chunk_size_for_magazine_cache = 0; /* non-optimized cases */
307 allocator_categorize (gsize aligned_chunk_size)
309 /* speed up the likely path */
310 if (G_LIKELY (aligned_chunk_size && aligned_chunk_size <= allocator->max_slab_chunk_size_for_magazine_cache))
311 return 1; /* use magazine cache */
313 /* the above will fail (max_slab_chunk_size_for_magazine_cache == 0) if the
314 * allocator is still uninitialized, or if we are not configured to use the
318 g_slice_init_nomessage ();
319 if (!allocator->config.always_malloc &&
320 aligned_chunk_size &&
321 aligned_chunk_size <= MAX_SLAB_CHUNK_SIZE (allocator))
323 if (allocator->config.bypass_magazines)
324 return 2; /* use slab allocator, see [2] */
325 return 1; /* use magazine cache */
327 return 0; /* use malloc() */
331 _g_slice_thread_init_nomessage (void)
333 /* we may not use g_error() or friends here */
335 g_slice_init_nomessage();
336 private_thread_memory = g_private_new (private_thread_memory_cleanup);
337 allocator->magazine_mutex = g_mutex_new();
338 allocator->slab_mutex = g_mutex_new();
342 g_mutex_lock_a (GMutex *mutex,
343 guint *contention_counter)
345 gboolean contention = FALSE;
346 if (!g_mutex_trylock (mutex))
348 g_mutex_lock (mutex);
353 allocator->mutex_counter++;
354 if (allocator->mutex_counter >= 1) /* quickly adapt to contention */
356 allocator->mutex_counter = 0;
357 *contention_counter = MIN (*contention_counter + 1, MAX_MAGAZINE_SIZE);
360 else /* !contention */
362 allocator->mutex_counter--;
363 if (allocator->mutex_counter < -11) /* moderately recover magazine sizes */
365 allocator->mutex_counter = 0;
366 *contention_counter = MAX (*contention_counter, 1) - 1;
371 static inline ThreadMemory*
372 thread_memory_from_self (void)
374 ThreadMemory *tmem = g_private_get (private_thread_memory);
375 if (G_UNLIKELY (!tmem))
377 const guint n_magazines = MAX_SLAB_INDEX (allocator);
378 tmem = g_malloc0 (sizeof (ThreadMemory) + sizeof (Magazine) * 2 * n_magazines);
379 tmem->magazine1 = (Magazine*) (tmem + 1);
380 tmem->magazine2 = &tmem->magazine1[n_magazines];
381 g_private_set (private_thread_memory, tmem);
386 static inline ChunkLink*
387 magazine_chain_pop_head (ChunkLink **magazine_chunks)
389 /* magazine chains are linked via ChunkLink->next.
390 * each ChunkLink->data of the toplevel chain may point to a subchain,
391 * linked via ChunkLink->next. ChunkLink->data of the subchains just
392 * contains uninitialized junk.
394 ChunkLink *chunk = (*magazine_chunks)->data;
395 if (G_UNLIKELY (chunk))
397 /* allocating from freed list */
398 (*magazine_chunks)->data = chunk->next;
402 chunk = *magazine_chunks;
403 *magazine_chunks = chunk->next;
408 #if 0 /* useful for debugging */
410 magazine_count (ChunkLink *head)
417 ChunkLink *child = head->data;
419 for (child = head->data; child; child = child->next)
428 allocator_get_magazine_threshold (Allocator *allocator,
431 /* the magazine size calculated here has a lower bound of MIN_MAGAZINE_SIZE,
432 * which is required by the implementation. also, for moderately sized chunks
433 * (say >= 64 bytes), magazine sizes shouldn't be much smaller then the number
434 * of chunks available per page/2 to avoid excessive traffic in the magazine
435 * cache for small to medium sized structures.
436 * the upper bound of the magazine size is effectively provided by
437 * MAX_MAGAZINE_SIZE. for larger chunks, this number is scaled down so that
438 * the content of a single magazine doesn't exceed ca. 16KB.
440 gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
441 guint threshold = MAX (MIN_MAGAZINE_SIZE, allocator->max_page_size / MAX (5 * chunk_size, 5 * 32));
442 guint contention_counter = allocator->contention_counters[ix];
443 if (G_UNLIKELY (contention_counter)) /* single CPU bias */
445 /* adapt contention counter thresholds to chunk sizes */
446 contention_counter = contention_counter * 64 / chunk_size;
447 threshold = MAX (threshold, contention_counter);
452 /* --- magazine cache --- */
454 magazine_cache_update_stamp (void)
456 if (allocator->stamp_counter >= MAX_STAMP_COUNTER)
459 g_get_current_time (&tv);
460 allocator->last_stamp = tv.tv_sec * 1000 + tv.tv_usec / 1000; /* milli seconds */
461 allocator->stamp_counter = 0;
464 allocator->stamp_counter++;
467 static inline ChunkLink*
468 magazine_chain_prepare_fields (ChunkLink *magazine_chunks)
474 /* checked upon initialization: mem_assert (MIN_MAGAZINE_SIZE >= 4); */
475 /* ensure a magazine with at least 4 unused data pointers */
476 chunk1 = magazine_chain_pop_head (&magazine_chunks);
477 chunk2 = magazine_chain_pop_head (&magazine_chunks);
478 chunk3 = magazine_chain_pop_head (&magazine_chunks);
479 chunk4 = magazine_chain_pop_head (&magazine_chunks);
480 chunk4->next = magazine_chunks;
481 chunk3->next = chunk4;
482 chunk2->next = chunk3;
483 chunk1->next = chunk2;
487 /* access the first 3 fields of a specially prepared magazine chain */
488 #define magazine_chain_prev(mc) ((mc)->data)
489 #define magazine_chain_stamp(mc) ((mc)->next->data)
490 #define magazine_chain_uint_stamp(mc) GPOINTER_TO_UINT ((mc)->next->data)
491 #define magazine_chain_next(mc) ((mc)->next->next->data)
492 #define magazine_chain_count(mc) ((mc)->next->next->next->data)
495 magazine_cache_trim (Allocator *allocator,
499 /* g_mutex_lock (allocator->mutex); done by caller */
500 /* trim magazine cache from tail */
501 ChunkLink *current = magazine_chain_prev (allocator->magazines[ix]);
502 ChunkLink *trash = NULL;
503 while (ABS (stamp - magazine_chain_uint_stamp (current)) >= allocator->config.working_set_msecs)
506 ChunkLink *prev = magazine_chain_prev (current);
507 ChunkLink *next = magazine_chain_next (current);
508 magazine_chain_next (prev) = next;
509 magazine_chain_prev (next) = prev;
510 /* clear special fields, put on trash stack */
511 magazine_chain_next (current) = NULL;
512 magazine_chain_count (current) = NULL;
513 magazine_chain_stamp (current) = NULL;
514 magazine_chain_prev (current) = trash;
516 /* fixup list head if required */
517 if (current == allocator->magazines[ix])
519 allocator->magazines[ix] = NULL;
524 g_mutex_unlock (allocator->magazine_mutex);
528 const gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
529 g_mutex_lock (allocator->slab_mutex);
533 trash = magazine_chain_prev (current);
534 magazine_chain_prev (current) = NULL; /* clear special field */
537 ChunkLink *chunk = magazine_chain_pop_head (¤t);
538 slab_allocator_free_chunk (chunk_size, chunk);
541 g_mutex_unlock (allocator->slab_mutex);
546 magazine_cache_push_magazine (guint ix,
547 ChunkLink *magazine_chunks,
548 gsize count) /* must be >= MIN_MAGAZINE_SIZE */
550 ChunkLink *current = magazine_chain_prepare_fields (magazine_chunks);
551 ChunkLink *next, *prev;
552 g_mutex_lock (allocator->magazine_mutex);
553 /* add magazine at head */
554 next = allocator->magazines[ix];
556 prev = magazine_chain_prev (next);
558 next = prev = current;
559 magazine_chain_next (prev) = current;
560 magazine_chain_prev (next) = current;
561 magazine_chain_prev (current) = prev;
562 magazine_chain_next (current) = next;
563 magazine_chain_count (current) = (gpointer) count;
565 magazine_cache_update_stamp();
566 magazine_chain_stamp (current) = GUINT_TO_POINTER (allocator->last_stamp);
567 allocator->magazines[ix] = current;
568 /* free old magazines beyond a certain threshold */
569 magazine_cache_trim (allocator, ix, allocator->last_stamp);
570 /* g_mutex_unlock (allocator->mutex); was done by magazine_cache_trim() */
574 magazine_cache_pop_magazine (guint ix,
577 g_mutex_lock_a (allocator->magazine_mutex, &allocator->contention_counters[ix]);
578 if (!allocator->magazines[ix])
580 guint magazine_threshold = allocator_get_magazine_threshold (allocator, ix);
581 gsize i, chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
582 ChunkLink *current = NULL;
583 g_mutex_unlock (allocator->magazine_mutex);
584 g_mutex_lock (allocator->slab_mutex);
585 for (i = 0; i < magazine_threshold; i++)
587 ChunkLink *chunk = slab_allocator_alloc_chunk (chunk_size);
589 chunk->next = current;
592 g_mutex_unlock (allocator->slab_mutex);
598 ChunkLink *current = allocator->magazines[ix];
599 ChunkLink *prev = magazine_chain_prev (current);
600 ChunkLink *next = magazine_chain_next (current);
602 magazine_chain_next (prev) = next;
603 magazine_chain_prev (next) = prev;
604 allocator->magazines[ix] = next == current ? NULL : next;
605 g_mutex_unlock (allocator->magazine_mutex);
606 /* clear special fields and hand out */
607 *countp = (gsize) magazine_chain_count (current);
608 magazine_chain_prev (current) = NULL;
609 magazine_chain_next (current) = NULL;
610 magazine_chain_count (current) = NULL;
611 magazine_chain_stamp (current) = NULL;
616 /* --- thread magazines --- */
618 private_thread_memory_cleanup (gpointer data)
620 ThreadMemory *tmem = data;
621 const guint n_magazines = MAX_SLAB_INDEX (allocator);
623 for (ix = 0; ix < n_magazines; ix++)
627 mags[0] = &tmem->magazine1[ix];
628 mags[1] = &tmem->magazine2[ix];
629 for (j = 0; j < 2; j++)
631 Magazine *mag = mags[j];
632 if (mag->count >= MIN_MAGAZINE_SIZE)
633 magazine_cache_push_magazine (ix, mag->chunks, mag->count);
636 const gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
637 g_mutex_lock (allocator->slab_mutex);
640 ChunkLink *chunk = magazine_chain_pop_head (&mag->chunks);
641 slab_allocator_free_chunk (chunk_size, chunk);
643 g_mutex_unlock (allocator->slab_mutex);
651 thread_memory_magazine1_reload (ThreadMemory *tmem,
654 Magazine *mag = &tmem->magazine1[ix];
655 mem_assert (mag->chunks == NULL); /* ensure that we may reset mag->count */
657 mag->chunks = magazine_cache_pop_magazine (ix, &mag->count);
661 thread_memory_magazine2_unload (ThreadMemory *tmem,
664 Magazine *mag = &tmem->magazine2[ix];
665 magazine_cache_push_magazine (ix, mag->chunks, mag->count);
671 thread_memory_swap_magazines (ThreadMemory *tmem,
674 Magazine xmag = tmem->magazine1[ix];
675 tmem->magazine1[ix] = tmem->magazine2[ix];
676 tmem->magazine2[ix] = xmag;
679 static inline gboolean
680 thread_memory_magazine1_is_empty (ThreadMemory *tmem,
683 return tmem->magazine1[ix].chunks == NULL;
686 static inline gboolean
687 thread_memory_magazine2_is_full (ThreadMemory *tmem,
690 return tmem->magazine2[ix].count >= allocator_get_magazine_threshold (allocator, ix);
693 static inline gpointer
694 thread_memory_magazine1_alloc (ThreadMemory *tmem,
697 Magazine *mag = &tmem->magazine1[ix];
698 ChunkLink *chunk = magazine_chain_pop_head (&mag->chunks);
699 if (G_LIKELY (mag->count > 0))
705 thread_memory_magazine2_free (ThreadMemory *tmem,
709 Magazine *mag = &tmem->magazine2[ix];
710 ChunkLink *chunk = mem;
712 chunk->next = mag->chunks;
717 /* --- API functions --- */
719 g_slice_alloc (gsize mem_size)
724 chunk_size = P2ALIGN (mem_size);
725 acat = allocator_categorize (chunk_size);
726 if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
728 ThreadMemory *tmem = thread_memory_from_self();
729 guint ix = SLAB_INDEX (allocator, chunk_size);
730 if (G_UNLIKELY (thread_memory_magazine1_is_empty (tmem, ix)))
732 thread_memory_swap_magazines (tmem, ix);
733 if (G_UNLIKELY (thread_memory_magazine1_is_empty (tmem, ix)))
734 thread_memory_magazine1_reload (tmem, ix);
736 mem = thread_memory_magazine1_alloc (tmem, ix);
738 else if (acat == 2) /* allocate through slab allocator */
740 g_mutex_lock (allocator->slab_mutex);
741 mem = slab_allocator_alloc_chunk (chunk_size);
742 g_mutex_unlock (allocator->slab_mutex);
744 else /* delegate to system malloc */
745 mem = g_malloc (mem_size);
750 g_slice_alloc0 (gsize mem_size)
752 gpointer mem = g_slice_alloc (mem_size);
754 memset (mem, 0, mem_size);
759 g_slice_free1 (gsize mem_size,
762 gsize chunk_size = P2ALIGN (mem_size);
763 guint acat = allocator_categorize (chunk_size);
764 if (G_UNLIKELY (!mem_block))
766 else if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
768 ThreadMemory *tmem = thread_memory_from_self();
769 guint ix = SLAB_INDEX (allocator, chunk_size);
770 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
772 thread_memory_swap_magazines (tmem, ix);
773 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
774 thread_memory_magazine2_unload (tmem, ix);
776 thread_memory_magazine2_free (tmem, ix, mem_block);
778 else if (acat == 2) /* allocate through slab allocator */
780 g_mutex_lock (allocator->slab_mutex);
781 slab_allocator_free_chunk (chunk_size, mem_block);
782 g_mutex_unlock (allocator->slab_mutex);
784 else /* delegate to system malloc */
789 g_slice_free_chain_with_offset (gsize mem_size,
793 gpointer slice = mem_chain;
794 /* while the thread magazines and the magazine cache are implemented so that
795 * they can easily be extended to allow for free lists containing more free
796 * lists for the first level nodes, which would allow O(1) freeing in this
797 * function, the benefit of such an extension is questionable, because:
798 * - the magazine size counts will become mere lower bounds which confuses
799 * the code adapting to lock contention;
800 * - freeing a single node to the thread magazines is very fast, so this
801 * O(list_length) operation is multiplied by a fairly small factor;
802 * - memory usage histograms on larger applications seem to indicate that
803 * the amount of released multi node lists is negligible in comparison
804 * to single node releases.
805 * - the major performance bottle neck, namely g_private_get() or
806 * g_mutex_lock()/g_mutex_unlock() has already been moved out of the
807 * inner loop for freeing chained slices.
809 gsize chunk_size = P2ALIGN (mem_size);
810 guint acat = allocator_categorize (chunk_size);
811 if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
813 ThreadMemory *tmem = thread_memory_from_self();
814 guint ix = SLAB_INDEX (allocator, chunk_size);
817 guint8 *current = slice;
818 slice = *(gpointer*) (current + next_offset);
819 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
821 thread_memory_swap_magazines (tmem, ix);
822 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
823 thread_memory_magazine2_unload (tmem, ix);
825 thread_memory_magazine2_free (tmem, ix, current);
828 else if (acat == 2) /* allocate through slab allocator */
830 g_mutex_lock (allocator->slab_mutex);
833 guint8 *current = slice;
834 slice = *(gpointer*) (current + next_offset);
835 slab_allocator_free_chunk (chunk_size, current);
837 g_mutex_unlock (allocator->slab_mutex);
839 else /* delegate to system malloc */
842 guint8 *current = slice;
843 slice = *(gpointer*) (current + next_offset);
848 /* --- single page allocator --- */
850 allocator_slab_stack_push (Allocator *allocator,
854 /* insert slab at slab ring head */
855 if (!allocator->slab_stack[ix])
862 SlabInfo *next = allocator->slab_stack[ix], *prev = next->prev;
868 allocator->slab_stack[ix] = sinfo;
872 allocator_aligned_page_size (Allocator *allocator,
875 gsize val = 1 << g_bit_storage (n_bytes - 1);
876 val = MAX (val, allocator->min_page_size);
881 allocator_add_slab (Allocator *allocator,
887 gsize addr, padding, n_chunks, color = 0;
888 gsize page_size = allocator_aligned_page_size (allocator, SLAB_BPAGE_SIZE (allocator, chunk_size));
889 /* allocate 1 page for the chunks and the slab */
890 gpointer aligned_memory = allocator_memalign (page_size, page_size - NATIVE_MALLOC_PADDING);
891 guint8 *mem = aligned_memory;
895 const gchar *syserr = "unknown error";
897 syserr = strerror (errno);
899 mem_error ("failed to allocate %u bytes (alignment: %u): %s\n",
900 (guint) (page_size - NATIVE_MALLOC_PADDING), (guint) page_size, syserr);
902 /* mask page adress */
903 addr = ((gsize) mem / page_size) * page_size;
904 /* assert alignment */
905 mem_assert (aligned_memory == (gpointer) addr);
906 /* basic slab info setup */
907 sinfo = (SlabInfo*) (mem + page_size - SLAB_INFO_SIZE);
908 sinfo->n_allocated = 0;
909 sinfo->chunks = NULL;
910 /* figure cache colorization */
911 n_chunks = ((guint8*) sinfo - mem) / chunk_size;
912 padding = ((guint8*) sinfo - mem) - n_chunks * chunk_size;
915 color = (allocator->color_accu * P2ALIGNMENT) % padding;
916 allocator->color_accu += allocator->config.color_increment;
918 /* add chunks to free list */
919 chunk = (ChunkLink*) (mem + color);
920 sinfo->chunks = chunk;
921 for (i = 0; i < n_chunks - 1; i++)
923 chunk->next = (ChunkLink*) ((guint8*) chunk + chunk_size);
926 chunk->next = NULL; /* last chunk */
927 /* add slab to slab ring */
928 allocator_slab_stack_push (allocator, ix, sinfo);
932 slab_allocator_alloc_chunk (gsize chunk_size)
935 guint ix = SLAB_INDEX (allocator, chunk_size);
936 /* ensure non-empty slab */
937 if (!allocator->slab_stack[ix] || !allocator->slab_stack[ix]->chunks)
938 allocator_add_slab (allocator, ix, chunk_size);
940 chunk = allocator->slab_stack[ix]->chunks;
941 allocator->slab_stack[ix]->chunks = chunk->next;
942 allocator->slab_stack[ix]->n_allocated++;
943 /* rotate empty slabs */
944 if (!allocator->slab_stack[ix]->chunks)
945 allocator->slab_stack[ix] = allocator->slab_stack[ix]->next;
950 slab_allocator_free_chunk (gsize chunk_size,
955 guint ix = SLAB_INDEX (allocator, chunk_size);
956 gsize page_size = allocator_aligned_page_size (allocator, SLAB_BPAGE_SIZE (allocator, chunk_size));
957 gsize addr = ((gsize) mem / page_size) * page_size;
958 /* mask page adress */
959 guint8 *page = (guint8*) addr;
960 SlabInfo *sinfo = (SlabInfo*) (page + page_size - SLAB_INFO_SIZE);
961 /* assert valid chunk count */
962 mem_assert (sinfo->n_allocated > 0);
963 /* add chunk to free list */
964 was_empty = sinfo->chunks == NULL;
965 chunk = (ChunkLink*) mem;
966 chunk->next = sinfo->chunks;
967 sinfo->chunks = chunk;
968 sinfo->n_allocated--;
969 /* keep slab ring partially sorted, empty slabs at end */
973 SlabInfo *next = sinfo->next, *prev = sinfo->prev;
976 if (allocator->slab_stack[ix] == sinfo)
977 allocator->slab_stack[ix] = next == sinfo ? NULL : next;
978 /* insert slab at head */
979 allocator_slab_stack_push (allocator, ix, sinfo);
981 /* eagerly free complete unused slabs */
982 if (!sinfo->n_allocated)
985 SlabInfo *next = sinfo->next, *prev = sinfo->prev;
988 if (allocator->slab_stack[ix] == sinfo)
989 allocator->slab_stack[ix] = next == sinfo ? NULL : next;
991 allocator_memfree (page_size, page);
995 /* --- memalign implementation --- */
996 #include <malloc.h> /* memalign() */
999 * define HAVE_POSIX_MEMALIGN 1 // if free(posix_memalign(3)) works, <stdlib.h>
1000 * define HAVE_MEMALIGN 1 // if free(memalign(3)) works, <malloc.h>
1001 * define HAVE_VALLOC 1 // if free(valloc(3)) works, <stdlib.h> or <malloc.h>
1002 * if none is provided, we implement malloc(3)-based alloc-only page alignment
1005 #if !(HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN || HAVE_VALLOC)
1006 static GTrashStack *compat_valloc_trash = NULL;
1010 allocator_memalign (gsize alignment,
1013 gpointer aligned_memory = NULL;
1015 #if HAVE_POSIX_MEMALIGN
1016 err = posix_memalign (&aligned_memory, alignment, memsize);
1019 aligned_memory = memalign (alignment, memsize);
1023 aligned_memory = valloc (memsize);
1026 /* simplistic non-freeing page allocator */
1027 mem_assert (alignment == sys_page_size);
1028 mem_assert (memsize <= sys_page_size);
1029 if (!compat_valloc_trash)
1031 const guint n_pages = 16;
1032 guint8 *mem = malloc (n_pages * sys_page_size);
1037 guint8 *amem = (guint8*) ALIGN ((gsize) mem, sys_page_size);
1039 i--; /* mem wasn't page aligned */
1041 g_trash_stack_push (&compat_valloc_trash, amem + i * sys_page_size);
1044 aligned_memory = g_trash_stack_pop (&compat_valloc_trash);
1046 if (!aligned_memory)
1048 return aligned_memory;
1052 allocator_memfree (gsize memsize,
1055 #if HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN || HAVE_VALLOC
1058 mem_assert (memsize <= sys_page_size);
1059 g_trash_stack_push (&compat_valloc_trash, mem);
1066 mem_error (const char *format,
1071 /* at least, put out "MEMORY-ERROR", in case we segfault during the rest of the function */
1072 fputs ("\n***MEMORY-ERROR***: ", stderr);
1073 pname = g_get_prgname();
1074 fprintf (stderr, "%s[%u]: GSlice: ", pname ? pname : "", getpid());
1075 va_start (args, format);
1076 vfprintf (stderr, format, args);
1078 fputs ("\n", stderr);
1082 #define __G_SLICE_C__
1083 #include "galiasdef.c"