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 #ifdef HAVE_POSIX_MEMALIGN
21 #define _XOPEN_SOURCE 600 /* posix_memalign() */
23 #include <stdlib.h> /* posix_memalign() */
27 #include "gmem.h" /* gslice.h */
28 #include "gthreadinit.h"
32 #include <unistd.h> /* sysconf() */
39 /* the GSlice allocator is split up into 4 layers, roughly modelled after the slab
40 * allocator and magazine extensions as outlined in:
41 * + [Bonwick94] Jeff Bonwick, The slab allocator: An object-caching kernel
42 * memory allocator. USENIX 1994, http://citeseer.ist.psu.edu/bonwick94slab.html
43 * + [Bonwick01] Bonwick and Jonathan Adams, Magazines and vmem: Extending the
44 * slab allocator to many cpu's and arbitrary resources.
45 * USENIX 2001, http://citeseer.ist.psu.edu/bonwick01magazines.html
47 * - the thread magazines. for each (aligned) chunk size, a magazine (a list)
48 * of recently freed and soon to be allocated chunks is maintained per thread.
49 * this way, most alloc/free requests can be quickly satisfied from per-thread
50 * free lists which only require one g_private_get() call to retrive the
52 * - the magazine cache. allocating and freeing chunks to/from threads only
53 * occours at magazine sizes from a global depot of magazines. the depot
54 * maintaines a 15 second working set of allocated magazines, so full
55 * magazines are not allocated and released too often.
56 * the chunk size dependent magazine sizes automatically adapt (within limits,
57 * see [3]) to lock contention to properly scale performance across a variety
59 * - the slab allocator. this allocator allocates slabs (blocks of memory) close
60 * to the system page size or multiples thereof which have to be page aligned.
61 * the blocks are divided into smaller chunks which are used to satisfy
62 * allocations from the upper layers. the space provided by the reminder of
63 * the chunk size division is used for cache colorization (random distribution
64 * of chunk addresses) to improve processor cache utilization. multiple slabs
65 * with the same chunk size are kept in a partially sorted ring to allow O(1)
66 * freeing and allocation of chunks (as long as the allocation of an entirely
67 * new slab can be avoided).
68 * - the page allocator. on most modern systems, posix_memalign(3) or
69 * memalign(3) should be available, so this is used to allocate blocks with
70 * system page size based alignments and sizes or multiples thereof.
71 * if no memalign variant is provided, valloc() is used instead and
72 * block sizes are limited to the system page size (no multiples thereof).
73 * as a fallback, on system without even valloc(), a malloc(3)-based page
74 * allocator with alloc-only behaviour is used.
77 * [1] some systems memalign(3) implementations may rely on boundary tagging for
78 * the handed out memory chunks. to avoid excessive page-wise fragmentation,
79 * we reserve 2 * sizeof (void*) per block size for the systems memalign(3),
80 * specified in NATIVE_MALLOC_PADDING.
81 * [2] using the slab allocator alone already provides for a fast and efficient
82 * allocator, it doesn't properly scale beyond single-threaded uses though.
83 * also, the slab allocator implements eager free(3)-ing, i.e. does not
84 * provide any form of caching or working set maintenance. so if used alone,
85 * it's vulnerable to trashing for sequences of balanced (alloc, free) pairs
86 * at certain thresholds.
87 * [3] magazine sizes are bound by an implementation specific minimum size and
88 * a chunk size specific maximum to limit magazine storage sizes to roughly
90 * [4] allocating ca. 8 chunks per block/page keeps a good balance between
91 * external and internal fragmentation (<= 12.5%). [Bonwick94]
94 /* --- macros and constants --- */
95 #define LARGEALIGNMENT (256)
96 #define P2ALIGNMENT (2 * sizeof (gsize)) /* fits 2 pointers (assumed to be 2 * GLIB_SIZEOF_SIZE_T below) */
97 #define ALIGN(size, base) ((base) * (gsize) (((size) + (base) - 1) / (base)))
98 #define NATIVE_MALLOC_PADDING P2ALIGNMENT /* per-page padding left for native malloc(3) see [1] */
99 #define SLAB_INFO_SIZE P2ALIGN (sizeof (SlabInfo) + NATIVE_MALLOC_PADDING)
100 #define MAX_MAGAZINE_SIZE (256) /* see [3] and allocator_get_magazine_threshold() for this */
101 #define MIN_MAGAZINE_SIZE (4)
102 #define MAX_STAMP_COUNTER (7) /* distributes the load of gettimeofday() */
103 #define MAX_SLAB_CHUNK_SIZE(al) (((al)->max_page_size - SLAB_INFO_SIZE) / 8) /* we want at last 8 chunks per page, see [4] */
104 #define MAX_SLAB_INDEX(al) (SLAB_INDEX (al, MAX_SLAB_CHUNK_SIZE (al)) + 1)
105 #define SLAB_INDEX(al, asize) ((asize) / P2ALIGNMENT - 1) /* asize must be P2ALIGNMENT aligned */
106 #define SLAB_CHUNK_SIZE(al, ix) (((ix) + 1) * P2ALIGNMENT)
107 #define SLAB_BPAGE_SIZE(al,csz) (8 * (csz) + SLAB_INFO_SIZE)
109 /* optimized version of ALIGN (size, P2ALIGNMENT) */
110 #if GLIB_SIZEOF_SIZE_T * 2 == 8 /* P2ALIGNMENT */
111 #define P2ALIGN(size) (((size) + 0x7) & ~(gsize) 0x7)
112 #elif GLIB_SIZEOF_SIZE_T * 2 == 16 /* P2ALIGNMENT */
113 #define P2ALIGN(size) (((size) + 0xf) & ~(gsize) 0xf)
115 #define P2ALIGN(size) ALIGN (size, P2ALIGNMENT)
118 /* special helpers to avoid gmessage.c dependency */
119 static void mem_error (const char *format, ...) G_GNUC_PRINTF (1,2);
120 #define mem_assert(cond) do { if (G_LIKELY (cond)) ; else mem_error ("assertion failed: %s", #cond); } while (0)
122 /* --- structures --- */
123 typedef struct _ChunkLink ChunkLink;
124 typedef struct _SlabInfo SlabInfo;
125 typedef struct _CachedMagazine CachedMagazine;
133 SlabInfo *next, *prev;
137 gsize count; /* approximative chunks list length */
140 Magazine *magazine1; /* array of MAX_SLAB_INDEX (allocator) */
141 Magazine *magazine2; /* array of MAX_SLAB_INDEX (allocator) */
144 gboolean always_malloc;
145 gboolean bypass_magazines;
146 gsize working_set_msecs;
147 guint color_increment;
150 /* const after initialization */
151 gsize min_page_size, max_page_size;
153 gsize max_slab_chunk_size_for_magazine_cache;
155 GMutex *magazine_mutex;
156 ChunkLink **magazines; /* array of MAX_SLAB_INDEX (allocator) */
157 guint *contention_counters; /* array of MAX_SLAB_INDEX (allocator) */
163 SlabInfo **slab_stack; /* array of MAX_SLAB_INDEX (allocator) */
167 /* --- prototypes --- */
168 static gpointer slab_allocator_alloc_chunk (gsize chunk_size);
169 static void slab_allocator_free_chunk (gsize chunk_size,
171 static void private_thread_memory_cleanup (gpointer data);
172 static gpointer allocator_memalign (gsize alignment,
174 static void allocator_memfree (gsize memsize,
176 static inline void magazine_cache_update_stamp (void);
177 static inline gsize allocator_get_magazine_threshold (Allocator *allocator,
180 /* --- variables --- */
181 static GPrivate *private_thread_memory = NULL;
182 static gsize sys_page_size = 0;
183 static Allocator allocator[1] = { { 0, }, };
184 static SliceConfig slice_config = {
185 FALSE, /* always_malloc */
186 FALSE, /* bypass_magazines */
187 15 * 1000, /* working_set_msecs */
188 1, /* color increment, alt: 0x7fffffff */
191 /* --- auxillary funcitons --- */
193 g_slice_set_config (GSliceConfig ckey,
196 g_return_if_fail (sys_page_size == 0);
199 case G_SLICE_CONFIG_ALWAYS_MALLOC:
200 slice_config.always_malloc = value != 0;
202 case G_SLICE_CONFIG_BYPASS_MAGAZINES:
203 slice_config.bypass_magazines = value != 0;
205 case G_SLICE_CONFIG_WORKING_SET_MSECS:
206 slice_config.working_set_msecs = value;
208 case G_SLICE_CONFIG_COLOR_INCREMENT:
209 slice_config.color_increment = value;
215 g_slice_get_config (GSliceConfig ckey)
219 case G_SLICE_CONFIG_ALWAYS_MALLOC:
220 return slice_config.always_malloc;
221 case G_SLICE_CONFIG_BYPASS_MAGAZINES:
222 return slice_config.bypass_magazines;
223 case G_SLICE_CONFIG_WORKING_SET_MSECS:
224 return slice_config.working_set_msecs;
225 case G_SLICE_CONFIG_CHUNK_SIZES:
226 return MAX_SLAB_INDEX (allocator);
227 case G_SLICE_CONFIG_COLOR_INCREMENT:
228 return slice_config.color_increment;
235 g_slice_get_config_state (GSliceConfig ckey,
240 g_return_val_if_fail (n_values != NULL, NULL);
245 case G_SLICE_CONFIG_CONTENTION_COUNTER:
246 array[i++] = SLAB_CHUNK_SIZE (allocator, address);
247 array[i++] = allocator->contention_counters[address];
248 array[i++] = allocator_get_magazine_threshold (allocator, address);
250 return g_memdup (array, sizeof (array[0]) * *n_values);
257 g_slice_init_nomessage (void)
259 /* we may not use g_error() or friends here */
260 mem_assert (sys_page_size == 0);
261 mem_assert (MIN_MAGAZINE_SIZE >= 4);
265 SYSTEM_INFO system_info;
266 GetSystemInfo (&system_info);
267 sys_page_size = system_info.dwPageSize;
270 sys_page_size = sysconf (_SC_PAGESIZE); /* = sysconf (_SC_PAGE_SIZE); = getpagesize(); */
272 mem_assert (sys_page_size >= 2 * LARGEALIGNMENT);
273 mem_assert ((sys_page_size & (sys_page_size - 1)) == 0);
274 allocator->config = slice_config;
275 allocator->min_page_size = sys_page_size;
276 #if HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN
277 /* allow allocation of pages up to 8KB (with 8KB alignment).
278 * this is useful because many medium to large sized structures
279 * fit less than 8 times (see [4]) into 4KB pages.
280 * we allow very small page sizes here, to reduce wastage in
281 * threads if only small allocations are required (this does
282 * bear the risk of incresing allocation times and fragmentation
285 allocator->min_page_size = MAX (allocator->min_page_size, 4096);
286 allocator->max_page_size = MAX (allocator->min_page_size, 8192);
287 allocator->min_page_size = MIN (allocator->min_page_size, 128);
289 /* we can only align to system page size */
290 allocator->max_page_size = sys_page_size;
292 allocator->magazine_mutex = NULL; /* _g_slice_thread_init_nomessage() */
293 allocator->magazines = g_new0 (ChunkLink*, MAX_SLAB_INDEX (allocator));
294 allocator->contention_counters = g_new0 (guint, MAX_SLAB_INDEX (allocator));
295 allocator->mutex_counter = 0;
296 allocator->stamp_counter = MAX_STAMP_COUNTER; /* force initial update */
297 allocator->last_stamp = 0;
298 allocator->slab_mutex = NULL; /* _g_slice_thread_init_nomessage() */
299 allocator->slab_stack = g_new0 (SlabInfo*, MAX_SLAB_INDEX (allocator));
300 allocator->color_accu = 0;
301 magazine_cache_update_stamp();
302 /* values cached for performance reasons */
303 allocator->max_slab_chunk_size_for_magazine_cache = MAX_SLAB_CHUNK_SIZE (allocator);
304 if (allocator->config.always_malloc || allocator->config.bypass_magazines)
305 allocator->max_slab_chunk_size_for_magazine_cache = 0; /* non-optimized cases */
309 allocator_categorize (gsize aligned_chunk_size)
311 /* speed up the likely path */
312 if (G_LIKELY (aligned_chunk_size && aligned_chunk_size <= allocator->max_slab_chunk_size_for_magazine_cache))
313 return 1; /* use magazine cache */
315 /* the above will fail (max_slab_chunk_size_for_magazine_cache == 0) if the
316 * allocator is still uninitialized, or if we are not configured to use the
320 g_slice_init_nomessage ();
321 if (!allocator->config.always_malloc &&
322 aligned_chunk_size &&
323 aligned_chunk_size <= MAX_SLAB_CHUNK_SIZE (allocator))
325 if (allocator->config.bypass_magazines)
326 return 2; /* use slab allocator, see [2] */
327 return 1; /* use magazine cache */
329 return 0; /* use malloc() */
333 _g_slice_thread_init_nomessage (void)
335 /* we may not use g_error() or friends here */
337 g_slice_init_nomessage();
338 private_thread_memory = g_private_new (private_thread_memory_cleanup);
339 allocator->magazine_mutex = g_mutex_new();
340 allocator->slab_mutex = g_mutex_new();
344 g_mutex_lock_a (GMutex *mutex,
345 guint *contention_counter)
347 gboolean contention = FALSE;
348 if (!g_mutex_trylock (mutex))
350 g_mutex_lock (mutex);
355 allocator->mutex_counter++;
356 if (allocator->mutex_counter >= 1) /* quickly adapt to contention */
358 allocator->mutex_counter = 0;
359 *contention_counter = MIN (*contention_counter + 1, MAX_MAGAZINE_SIZE);
362 else /* !contention */
364 allocator->mutex_counter--;
365 if (allocator->mutex_counter < -11) /* moderately recover magazine sizes */
367 allocator->mutex_counter = 0;
368 *contention_counter = MAX (*contention_counter, 1) - 1;
373 static inline ThreadMemory*
374 thread_memory_from_self (void)
376 ThreadMemory *tmem = g_private_get (private_thread_memory);
377 if (G_UNLIKELY (!tmem))
379 const guint n_magazines = MAX_SLAB_INDEX (allocator);
380 tmem = g_malloc0 (sizeof (ThreadMemory) + sizeof (Magazine) * 2 * n_magazines);
381 tmem->magazine1 = (Magazine*) (tmem + 1);
382 tmem->magazine2 = &tmem->magazine1[n_magazines];
383 g_private_set (private_thread_memory, tmem);
388 static inline ChunkLink*
389 magazine_chain_pop_head (ChunkLink **magazine_chunks)
391 /* magazine chains are linked via ChunkLink->next.
392 * each ChunkLink->data of the toplevel chain may point to a subchain,
393 * linked via ChunkLink->next. ChunkLink->data of the subchains just
394 * contains uninitialized junk.
396 ChunkLink *chunk = (*magazine_chunks)->data;
397 if (G_UNLIKELY (chunk))
399 /* allocating from freed list */
400 (*magazine_chunks)->data = chunk->next;
404 chunk = *magazine_chunks;
405 *magazine_chunks = chunk->next;
410 #if 0 /* useful for debugging */
412 magazine_count (ChunkLink *head)
419 ChunkLink *child = head->data;
421 for (child = head->data; child; child = child->next)
430 allocator_get_magazine_threshold (Allocator *allocator,
433 /* the magazine size calculated here has a lower bound of MIN_MAGAZINE_SIZE,
434 * which is required by the implementation. also, for moderately sized chunks
435 * (say >= 64 bytes), magazine sizes shouldn't be much smaller then the number
436 * of chunks available per page/2 to avoid excessive traffic in the magazine
437 * cache for small to medium sized structures.
438 * the upper bound of the magazine size is effectively provided by
439 * MAX_MAGAZINE_SIZE. for larger chunks, this number is scaled down so that
440 * the content of a single magazine doesn't exceed ca. 16KB.
442 gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
443 guint threshold = MAX (MIN_MAGAZINE_SIZE, allocator->max_page_size / MAX (5 * chunk_size, 5 * 32));
444 guint contention_counter = allocator->contention_counters[ix];
445 if (G_UNLIKELY (contention_counter)) /* single CPU bias */
447 /* adapt contention counter thresholds to chunk sizes */
448 contention_counter = contention_counter * 64 / chunk_size;
449 threshold = MAX (threshold, contention_counter);
454 /* --- magazine cache --- */
456 magazine_cache_update_stamp (void)
458 if (allocator->stamp_counter >= MAX_STAMP_COUNTER)
461 g_get_current_time (&tv);
462 allocator->last_stamp = tv.tv_sec * 1000 + tv.tv_usec / 1000; /* milli seconds */
463 allocator->stamp_counter = 0;
466 allocator->stamp_counter++;
469 static inline ChunkLink*
470 magazine_chain_prepare_fields (ChunkLink *magazine_chunks)
476 /* checked upon initialization: mem_assert (MIN_MAGAZINE_SIZE >= 4); */
477 /* ensure a magazine with at least 4 unused data pointers */
478 chunk1 = magazine_chain_pop_head (&magazine_chunks);
479 chunk2 = magazine_chain_pop_head (&magazine_chunks);
480 chunk3 = magazine_chain_pop_head (&magazine_chunks);
481 chunk4 = magazine_chain_pop_head (&magazine_chunks);
482 chunk4->next = magazine_chunks;
483 chunk3->next = chunk4;
484 chunk2->next = chunk3;
485 chunk1->next = chunk2;
489 /* access the first 3 fields of a specially prepared magazine chain */
490 #define magazine_chain_prev(mc) ((mc)->data)
491 #define magazine_chain_stamp(mc) ((mc)->next->data)
492 #define magazine_chain_uint_stamp(mc) GPOINTER_TO_UINT ((mc)->next->data)
493 #define magazine_chain_next(mc) ((mc)->next->next->data)
494 #define magazine_chain_count(mc) ((mc)->next->next->next->data)
497 magazine_cache_trim (Allocator *allocator,
501 /* g_mutex_lock (allocator->mutex); done by caller */
502 /* trim magazine cache from tail */
503 ChunkLink *current = magazine_chain_prev (allocator->magazines[ix]);
504 ChunkLink *trash = NULL;
505 while (ABS (stamp - magazine_chain_uint_stamp (current)) >= allocator->config.working_set_msecs)
508 ChunkLink *prev = magazine_chain_prev (current);
509 ChunkLink *next = magazine_chain_next (current);
510 magazine_chain_next (prev) = next;
511 magazine_chain_prev (next) = prev;
512 /* clear special fields, put on trash stack */
513 magazine_chain_next (current) = NULL;
514 magazine_chain_count (current) = NULL;
515 magazine_chain_stamp (current) = NULL;
516 magazine_chain_prev (current) = trash;
518 /* fixup list head if required */
519 if (current == allocator->magazines[ix])
521 allocator->magazines[ix] = NULL;
526 g_mutex_unlock (allocator->magazine_mutex);
530 const gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
531 g_mutex_lock (allocator->slab_mutex);
535 trash = magazine_chain_prev (current);
536 magazine_chain_prev (current) = NULL; /* clear special field */
539 ChunkLink *chunk = magazine_chain_pop_head (¤t);
540 slab_allocator_free_chunk (chunk_size, chunk);
543 g_mutex_unlock (allocator->slab_mutex);
548 magazine_cache_push_magazine (guint ix,
549 ChunkLink *magazine_chunks,
550 gsize count) /* must be >= MIN_MAGAZINE_SIZE */
552 ChunkLink *current = magazine_chain_prepare_fields (magazine_chunks);
553 ChunkLink *next, *prev;
554 g_mutex_lock (allocator->magazine_mutex);
555 /* add magazine at head */
556 next = allocator->magazines[ix];
558 prev = magazine_chain_prev (next);
560 next = prev = current;
561 magazine_chain_next (prev) = current;
562 magazine_chain_prev (next) = current;
563 magazine_chain_prev (current) = prev;
564 magazine_chain_next (current) = next;
565 magazine_chain_count (current) = (gpointer) count;
567 magazine_cache_update_stamp();
568 magazine_chain_stamp (current) = GUINT_TO_POINTER (allocator->last_stamp);
569 allocator->magazines[ix] = current;
570 /* free old magazines beyond a certain threshold */
571 magazine_cache_trim (allocator, ix, allocator->last_stamp);
572 /* g_mutex_unlock (allocator->mutex); was done by magazine_cache_trim() */
576 magazine_cache_pop_magazine (guint ix,
579 g_mutex_lock_a (allocator->magazine_mutex, &allocator->contention_counters[ix]);
580 if (!allocator->magazines[ix])
582 guint magazine_threshold = allocator_get_magazine_threshold (allocator, ix);
583 gsize i, chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
584 ChunkLink *chunk, *head;
585 g_mutex_unlock (allocator->magazine_mutex);
586 g_mutex_lock (allocator->slab_mutex);
587 head = slab_allocator_alloc_chunk (chunk_size);
590 for (i = 1; i < magazine_threshold; i++)
592 chunk->next = slab_allocator_alloc_chunk (chunk_size);
597 g_mutex_unlock (allocator->slab_mutex);
603 ChunkLink *current = allocator->magazines[ix];
604 ChunkLink *prev = magazine_chain_prev (current);
605 ChunkLink *next = magazine_chain_next (current);
607 magazine_chain_next (prev) = next;
608 magazine_chain_prev (next) = prev;
609 allocator->magazines[ix] = next == current ? NULL : next;
610 g_mutex_unlock (allocator->magazine_mutex);
611 /* clear special fields and hand out */
612 *countp = (gsize) magazine_chain_count (current);
613 magazine_chain_prev (current) = NULL;
614 magazine_chain_next (current) = NULL;
615 magazine_chain_count (current) = NULL;
616 magazine_chain_stamp (current) = NULL;
621 /* --- thread magazines --- */
623 private_thread_memory_cleanup (gpointer data)
625 ThreadMemory *tmem = data;
626 const guint n_magazines = MAX_SLAB_INDEX (allocator);
628 for (ix = 0; ix < n_magazines; ix++)
632 mags[0] = &tmem->magazine1[ix];
633 mags[1] = &tmem->magazine2[ix];
634 for (j = 0; j < 2; j++)
636 Magazine *mag = mags[j];
637 if (mag->count >= MIN_MAGAZINE_SIZE)
638 magazine_cache_push_magazine (ix, mag->chunks, mag->count);
641 const gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
642 g_mutex_lock (allocator->slab_mutex);
645 ChunkLink *chunk = magazine_chain_pop_head (&mag->chunks);
646 slab_allocator_free_chunk (chunk_size, chunk);
648 g_mutex_unlock (allocator->slab_mutex);
656 thread_memory_magazine1_reload (ThreadMemory *tmem,
659 Magazine *mag = &tmem->magazine1[ix];
660 mem_assert (mag->chunks == NULL); /* ensure that we may reset mag->count */
662 mag->chunks = magazine_cache_pop_magazine (ix, &mag->count);
666 thread_memory_magazine2_unload (ThreadMemory *tmem,
669 Magazine *mag = &tmem->magazine2[ix];
670 magazine_cache_push_magazine (ix, mag->chunks, mag->count);
676 thread_memory_swap_magazines (ThreadMemory *tmem,
679 Magazine xmag = tmem->magazine1[ix];
680 tmem->magazine1[ix] = tmem->magazine2[ix];
681 tmem->magazine2[ix] = xmag;
684 static inline gboolean
685 thread_memory_magazine1_is_empty (ThreadMemory *tmem,
688 return tmem->magazine1[ix].chunks == NULL;
691 static inline gboolean
692 thread_memory_magazine2_is_full (ThreadMemory *tmem,
695 return tmem->magazine2[ix].count >= allocator_get_magazine_threshold (allocator, ix);
698 static inline gpointer
699 thread_memory_magazine1_alloc (ThreadMemory *tmem,
702 Magazine *mag = &tmem->magazine1[ix];
703 ChunkLink *chunk = magazine_chain_pop_head (&mag->chunks);
704 if (G_LIKELY (mag->count > 0))
710 thread_memory_magazine2_free (ThreadMemory *tmem,
714 Magazine *mag = &tmem->magazine2[ix];
715 ChunkLink *chunk = mem;
717 chunk->next = mag->chunks;
722 /* --- API functions --- */
724 g_slice_alloc (gsize mem_size)
729 chunk_size = P2ALIGN (mem_size);
730 acat = allocator_categorize (chunk_size);
731 if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
733 ThreadMemory *tmem = thread_memory_from_self();
734 guint ix = SLAB_INDEX (allocator, chunk_size);
735 if (G_UNLIKELY (thread_memory_magazine1_is_empty (tmem, ix)))
737 thread_memory_swap_magazines (tmem, ix);
738 if (G_UNLIKELY (thread_memory_magazine1_is_empty (tmem, ix)))
739 thread_memory_magazine1_reload (tmem, ix);
741 mem = thread_memory_magazine1_alloc (tmem, ix);
743 else if (acat == 2) /* allocate through slab allocator */
745 g_mutex_lock (allocator->slab_mutex);
746 mem = slab_allocator_alloc_chunk (chunk_size);
747 g_mutex_unlock (allocator->slab_mutex);
749 else /* delegate to system malloc */
750 mem = g_malloc (mem_size);
755 g_slice_alloc0 (gsize mem_size)
757 gpointer mem = g_slice_alloc (mem_size);
759 memset (mem, 0, mem_size);
764 g_slice_free1 (gsize mem_size,
767 gsize chunk_size = P2ALIGN (mem_size);
768 guint acat = allocator_categorize (chunk_size);
769 if (G_UNLIKELY (!mem_block))
771 else if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
773 ThreadMemory *tmem = thread_memory_from_self();
774 guint ix = SLAB_INDEX (allocator, chunk_size);
775 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
777 thread_memory_swap_magazines (tmem, ix);
778 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
779 thread_memory_magazine2_unload (tmem, ix);
781 thread_memory_magazine2_free (tmem, ix, mem_block);
783 else if (acat == 2) /* allocate through slab allocator */
785 g_mutex_lock (allocator->slab_mutex);
786 slab_allocator_free_chunk (chunk_size, mem_block);
787 g_mutex_unlock (allocator->slab_mutex);
789 else /* delegate to system malloc */
794 g_slice_free_chain_with_offset (gsize mem_size,
798 gpointer slice = mem_chain;
799 /* while the thread magazines and the magazine cache are implemented so that
800 * they can easily be extended to allow for free lists containing more free
801 * lists for the first level nodes, which would allow O(1) freeing in this
802 * function, the benefit of such an extension is questionable, because:
803 * - the magazine size counts will become mere lower bounds which confuses
804 * the code adapting to lock contention;
805 * - freeing a single node to the thread magazines is very fast, so this
806 * O(list_length) operation is multiplied by a fairly small factor;
807 * - memory usage histograms on larger applications seem to indicate that
808 * the amount of released multi node lists is negligible in comparison
809 * to single node releases.
810 * - the major performance bottle neck, namely g_private_get() or
811 * g_mutex_lock()/g_mutex_unlock() has already been moved out of the
812 * inner loop for freeing chained slices.
814 gsize chunk_size = P2ALIGN (mem_size);
815 guint acat = allocator_categorize (chunk_size);
816 if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
818 ThreadMemory *tmem = thread_memory_from_self();
819 guint ix = SLAB_INDEX (allocator, chunk_size);
822 guint8 *current = slice;
823 slice = *(gpointer*) (current + next_offset);
824 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
826 thread_memory_swap_magazines (tmem, ix);
827 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
828 thread_memory_magazine2_unload (tmem, ix);
830 thread_memory_magazine2_free (tmem, ix, current);
833 else if (acat == 2) /* allocate through slab allocator */
835 g_mutex_lock (allocator->slab_mutex);
838 guint8 *current = slice;
839 slice = *(gpointer*) (current + next_offset);
840 slab_allocator_free_chunk (chunk_size, current);
842 g_mutex_unlock (allocator->slab_mutex);
844 else /* delegate to system malloc */
847 guint8 *current = slice;
848 slice = *(gpointer*) (current + next_offset);
853 /* --- single page allocator --- */
855 allocator_slab_stack_push (Allocator *allocator,
859 /* insert slab at slab ring head */
860 if (!allocator->slab_stack[ix])
867 SlabInfo *next = allocator->slab_stack[ix], *prev = next->prev;
873 allocator->slab_stack[ix] = sinfo;
877 allocator_aligned_page_size (Allocator *allocator,
880 gsize val = 1 << g_bit_storage (n_bytes - 1);
881 val = MAX (val, allocator->min_page_size);
886 allocator_add_slab (Allocator *allocator,
892 gsize addr, padding, n_chunks, color = 0;
893 gsize page_size = allocator_aligned_page_size (allocator, SLAB_BPAGE_SIZE (allocator, chunk_size));
894 /* allocate 1 page for the chunks and the slab */
895 gpointer aligned_memory = allocator_memalign (page_size, page_size - NATIVE_MALLOC_PADDING);
896 guint8 *mem = aligned_memory;
900 const gchar *syserr = "unknown error";
902 syserr = strerror (errno);
904 mem_error ("failed to allocate %u bytes (alignment: %u): %s\n",
905 (guint) (page_size - NATIVE_MALLOC_PADDING), (guint) page_size, syserr);
907 /* mask page adress */
908 addr = ((gsize) mem / page_size) * page_size;
909 /* assert alignment */
910 mem_assert (aligned_memory == (gpointer) addr);
911 /* basic slab info setup */
912 sinfo = (SlabInfo*) (mem + page_size - SLAB_INFO_SIZE);
913 sinfo->n_allocated = 0;
914 sinfo->chunks = NULL;
915 /* figure cache colorization */
916 n_chunks = ((guint8*) sinfo - mem) / chunk_size;
917 padding = ((guint8*) sinfo - mem) - n_chunks * chunk_size;
920 color = (allocator->color_accu * P2ALIGNMENT) % padding;
921 allocator->color_accu += allocator->config.color_increment;
923 /* add chunks to free list */
924 chunk = (ChunkLink*) (mem + color);
925 sinfo->chunks = chunk;
926 for (i = 0; i < n_chunks - 1; i++)
928 chunk->next = (ChunkLink*) ((guint8*) chunk + chunk_size);
931 chunk->next = NULL; /* last chunk */
932 /* add slab to slab ring */
933 allocator_slab_stack_push (allocator, ix, sinfo);
937 slab_allocator_alloc_chunk (gsize chunk_size)
940 guint ix = SLAB_INDEX (allocator, chunk_size);
941 /* ensure non-empty slab */
942 if (!allocator->slab_stack[ix] || !allocator->slab_stack[ix]->chunks)
943 allocator_add_slab (allocator, ix, chunk_size);
945 chunk = allocator->slab_stack[ix]->chunks;
946 allocator->slab_stack[ix]->chunks = chunk->next;
947 allocator->slab_stack[ix]->n_allocated++;
948 /* rotate empty slabs */
949 if (!allocator->slab_stack[ix]->chunks)
950 allocator->slab_stack[ix] = allocator->slab_stack[ix]->next;
955 slab_allocator_free_chunk (gsize chunk_size,
960 guint ix = SLAB_INDEX (allocator, chunk_size);
961 gsize page_size = allocator_aligned_page_size (allocator, SLAB_BPAGE_SIZE (allocator, chunk_size));
962 gsize addr = ((gsize) mem / page_size) * page_size;
963 /* mask page adress */
964 guint8 *page = (guint8*) addr;
965 SlabInfo *sinfo = (SlabInfo*) (page + page_size - SLAB_INFO_SIZE);
966 /* assert valid chunk count */
967 mem_assert (sinfo->n_allocated > 0);
968 /* add chunk to free list */
969 was_empty = sinfo->chunks == NULL;
970 chunk = (ChunkLink*) mem;
971 chunk->next = sinfo->chunks;
972 sinfo->chunks = chunk;
973 sinfo->n_allocated--;
974 /* keep slab ring partially sorted, empty slabs at end */
978 SlabInfo *next = sinfo->next, *prev = sinfo->prev;
981 if (allocator->slab_stack[ix] == sinfo)
982 allocator->slab_stack[ix] = next == sinfo ? NULL : next;
983 /* insert slab at head */
984 allocator_slab_stack_push (allocator, ix, sinfo);
986 /* eagerly free complete unused slabs */
987 if (!sinfo->n_allocated)
990 SlabInfo *next = sinfo->next, *prev = sinfo->prev;
993 if (allocator->slab_stack[ix] == sinfo)
994 allocator->slab_stack[ix] = next == sinfo ? NULL : next;
996 allocator_memfree (page_size, page);
1000 /* --- memalign implementation --- */
1001 #ifdef HAVE_MALLOC_H
1002 #include <malloc.h> /* memalign() */
1006 * define HAVE_POSIX_MEMALIGN 1 // if free(posix_memalign(3)) works, <stdlib.h>
1007 * define HAVE_MEMALIGN 1 // if free(memalign(3)) works, <malloc.h>
1008 * define HAVE_VALLOC 1 // if free(valloc(3)) works, <stdlib.h> or <malloc.h>
1009 * if none is provided, we implement malloc(3)-based alloc-only page alignment
1012 #if !(HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN || HAVE_VALLOC)
1013 static GTrashStack *compat_valloc_trash = NULL;
1017 allocator_memalign (gsize alignment,
1020 gpointer aligned_memory = NULL;
1022 #if HAVE_POSIX_MEMALIGN
1023 err = posix_memalign (&aligned_memory, alignment, memsize);
1026 aligned_memory = memalign (alignment, memsize);
1030 aligned_memory = valloc (memsize);
1033 /* simplistic non-freeing page allocator */
1034 mem_assert (alignment == sys_page_size);
1035 mem_assert (memsize <= sys_page_size);
1036 if (!compat_valloc_trash)
1038 const guint n_pages = 16;
1039 guint8 *mem = malloc (n_pages * sys_page_size);
1044 guint8 *amem = (guint8*) ALIGN ((gsize) mem, sys_page_size);
1046 i--; /* mem wasn't page aligned */
1048 g_trash_stack_push (&compat_valloc_trash, amem + i * sys_page_size);
1051 aligned_memory = g_trash_stack_pop (&compat_valloc_trash);
1053 if (!aligned_memory)
1055 return aligned_memory;
1059 allocator_memfree (gsize memsize,
1062 #if HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN || HAVE_VALLOC
1065 mem_assert (memsize <= sys_page_size);
1066 g_trash_stack_push (&compat_valloc_trash, mem);
1073 mem_error (const char *format,
1078 /* at least, put out "MEMORY-ERROR", in case we segfault during the rest of the function */
1079 fputs ("\n***MEMORY-ERROR***: ", stderr);
1080 pname = g_get_prgname();
1081 fprintf (stderr, "%s[%u]: GSlice: ", pname ? pname : "", getpid());
1082 va_start (args, format);
1083 vfprintf (stderr, format, args);
1085 fputs ("\n", stderr);
1089 #define __G_SLICE_C__
1090 #include "galiasdef.c"