1 /* GLIB sliced memory - fast concurrent memory chunk allocator
2 * Copyright (C) 2005 Tim Janik
3 *
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 2.1 of the License, or (at your option) any later version.
8 *
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.
13 *
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with this library; if not, see <http://www.gnu.org/licenses/>.
16 */
17 /* MT safe */
18
19 #include "config.h"
20 #include "glibconfig.h"
21
22 #if defined(HAVE_POSIX_MEMALIGN) && !defined(_XOPEN_SOURCE)
23 #define _XOPEN_SOURCE 600 /* posix_memalign() */
24 #endif
25 #include <stdlib.h> /* posix_memalign() */
26 #include <string.h>
27 #include <errno.h>
28
29 #ifdef G_OS_UNIX
30 #include <unistd.h> /* sysconf() */
31 #endif
32 #ifdef G_OS_WIN32
33 #include <windows.h>
34 #include <process.h>
35 #endif
36
37 #include <stdio.h> /* fputs */
38
39 #include "gslice.h"
40
41 #include "gmain.h"
42 #include "gmem.h" /* gslice.h */
43 #include "gstrfuncs.h"
44 #include "gutils.h"
45 #include "gtrashstack.h"
46 #include "gtestutils.h"
47 #include "gthread.h"
48 #include "gthreadprivate.h"
49 #include "glib_trace.h"
50 #include "gprintf.h"
51
52 #include "gvalgrind.h"
53
54 /**
55 * SECTION:memory_slices
56 * @title: Memory Slices
57 * @short_description: efficient way to allocate groups of equal-sized
58 * chunks of memory
59 *
60 * Memory slices provide a space-efficient and multi-processing scalable
61 * way to allocate equal-sized pieces of memory, just like the original
62 * #GMemChunks (from GLib 2.8), while avoiding their excessive
63 * memory-waste, scalability and performance problems.
64 *
65 * To achieve these goals, the slice allocator uses a sophisticated,
66 * layered design that has been inspired by Bonwick's slab allocator
67 * ([Bonwick94](http://citeseer.ist.psu.edu/bonwick94slab.html)
68 * Jeff Bonwick, The slab allocator: An object-caching kernel
69 * memory allocator. USENIX 1994, and
70 * [Bonwick01](http://citeseer.ist.psu.edu/bonwick01magazines.html)
71 * Bonwick and Jonathan Adams, Magazines and vmem: Extending the
72 * slab allocator to many cpu's and arbitrary resources. USENIX 2001)
73 *
74 * It uses posix_memalign() to optimize allocations of many equally-sized
75 * chunks, and has per-thread free lists (the so-called magazine layer)
76 * to quickly satisfy allocation requests of already known structure sizes.
77 * This is accompanied by extra caching logic to keep freed memory around
78 * for some time before returning it to the system. Memory that is unused
79 * due to alignment constraints is used for cache colorization (random
80 * distribution of chunk addresses) to improve CPU cache utilization. The
81 * caching layer of the slice allocator adapts itself to high lock contention
82 * to improve scalability.
83 *
84 * The slice allocator can allocate blocks as small as two pointers, and
85 * unlike malloc(), it does not reserve extra space per block. For large block
86 * sizes, g_slice_new() and g_slice_alloc() will automatically delegate to the
87 * system malloc() implementation. For newly written code it is recommended
88 * to use the new `g_slice` API instead of g_malloc() and
89 * friends, as long as objects are not resized during their lifetime and the
90 * object size used at allocation time is still available when freeing.
91 *
92 * Here is an example for using the slice allocator:
93 * |[<!-- language="C" -->
94 * gchar *mem[10000];
95 * gint i;
96 *
97 * // Allocate 10000 blocks.
98 * for (i = 0; i < 10000; i++)
99 * {
100 * mem[i] = g_slice_alloc (50);
101 *
102 * // Fill in the memory with some junk.
103 * for (j = 0; j < 50; j++)
104 * mem[i][j] = i * j;
105 * }
106 *
107 * // Now free all of the blocks.
108 * for (i = 0; i < 10000; i++)
109 * g_slice_free1 (50, mem[i]);
110 * ]|
111 *
112 * And here is an example for using the using the slice allocator
113 * with data structures:
114 * |[<!-- language="C" -->
115 * GRealArray *array;
116 *
117 * // Allocate one block, using the g_slice_new() macro.
118 * array = g_slice_new (GRealArray);
119 *
120 * // We can now use array just like a normal pointer to a structure.
121 * array->data = NULL;
122 * array->len = 0;
123 * array->alloc = 0;
124 * array->zero_terminated = (zero_terminated ? 1 : 0);
125 * array->clear = (clear ? 1 : 0);
126 * array->elt_size = elt_size;
127 *
128 * // We can free the block, so it can be reused.
129 * g_slice_free (GRealArray, array);
130 * ]|
131 */
132
133 /* the GSlice allocator is split up into 4 layers, roughly modelled after the slab
134 * allocator and magazine extensions as outlined in:
135 * + [Bonwick94] Jeff Bonwick, The slab allocator: An object-caching kernel
136 * memory allocator. USENIX 1994, http://citeseer.ist.psu.edu/bonwick94slab.html
137 * + [Bonwick01] Bonwick and Jonathan Adams, Magazines and vmem: Extending the
138 * slab allocator to many cpu's and arbitrary resources.
139 * USENIX 2001, http://citeseer.ist.psu.edu/bonwick01magazines.html
140 * the layers are:
141 * - the thread magazines. for each (aligned) chunk size, a magazine (a list)
142 * of recently freed and soon to be allocated chunks is maintained per thread.
143 * this way, most alloc/free requests can be quickly satisfied from per-thread
144 * free lists which only require one g_private_get() call to retrive the
145 * thread handle.
146 * - the magazine cache. allocating and freeing chunks to/from threads only
147 * occours at magazine sizes from a global depot of magazines. the depot
148 * maintaines a 15 second working set of allocated magazines, so full
149 * magazines are not allocated and released too often.
150 * the chunk size dependent magazine sizes automatically adapt (within limits,
151 * see [3]) to lock contention to properly scale performance across a variety
152 * of SMP systems.
153 * - the slab allocator. this allocator allocates slabs (blocks of memory) close
154 * to the system page size or multiples thereof which have to be page aligned.
155 * the blocks are divided into smaller chunks which are used to satisfy
156 * allocations from the upper layers. the space provided by the reminder of
157 * the chunk size division is used for cache colorization (random distribution
158 * of chunk addresses) to improve processor cache utilization. multiple slabs
159 * with the same chunk size are kept in a partially sorted ring to allow O(1)
160 * freeing and allocation of chunks (as long as the allocation of an entirely
161 * new slab can be avoided).
162 * - the page allocator. on most modern systems, posix_memalign(3) or
163 * memalign(3) should be available, so this is used to allocate blocks with
164 * system page size based alignments and sizes or multiples thereof.
165 * if no memalign variant is provided, valloc() is used instead and
166 * block sizes are limited to the system page size (no multiples thereof).
167 * as a fallback, on system without even valloc(), a malloc(3)-based page
168 * allocator with alloc-only behaviour is used.
169 *
170 * NOTES:
171 * [1] some systems memalign(3) implementations may rely on boundary tagging for
172 * the handed out memory chunks. to avoid excessive page-wise fragmentation,
173 * we reserve 2 * sizeof (void*) per block size for the systems memalign(3),
174 * specified in NATIVE_MALLOC_PADDING.
175 * [2] using the slab allocator alone already provides for a fast and efficient
176 * allocator, it doesn't properly scale beyond single-threaded uses though.
177 * also, the slab allocator implements eager free(3)-ing, i.e. does not
178 * provide any form of caching or working set maintenance. so if used alone,
179 * it's vulnerable to trashing for sequences of balanced (alloc, free) pairs
180 * at certain thresholds.
181 * [3] magazine sizes are bound by an implementation specific minimum size and
182 * a chunk size specific maximum to limit magazine storage sizes to roughly
183 * 16KB.
184 * [4] allocating ca. 8 chunks per block/page keeps a good balance between
185 * external and internal fragmentation (<= 12.5%). [Bonwick94]
186 */
187
188 /* --- macros and constants --- */
189 #define LARGEALIGNMENT (256)
190 #define P2ALIGNMENT (2 * sizeof (gsize)) /* fits 2 pointers (assumed to be 2 * GLIB_SIZEOF_SIZE_T below) */
191 #define ALIGN(size, base) ((base) * (gsize) (((size) + (base) - 1) / (base)))
192 #define NATIVE_MALLOC_PADDING P2ALIGNMENT /* per-page padding left for native malloc(3) see [1] */
193 #define SLAB_INFO_SIZE P2ALIGN (sizeof (SlabInfo) + NATIVE_MALLOC_PADDING)
194 #define MAX_MAGAZINE_SIZE (256) /* see [3] and allocator_get_magazine_threshold() for this */
195 #define MIN_MAGAZINE_SIZE (4)
196 #define MAX_STAMP_COUNTER (7) /* distributes the load of gettimeofday() */
197 #define MAX_SLAB_CHUNK_SIZE(al) (((al)->max_page_size - SLAB_INFO_SIZE) / 8) /* we want at last 8 chunks per page, see [4] */
198 #define MAX_SLAB_INDEX(al) (SLAB_INDEX (al, MAX_SLAB_CHUNK_SIZE (al)) + 1)
199 #define SLAB_INDEX(al, asize) ((asize) / P2ALIGNMENT - 1) /* asize must be P2ALIGNMENT aligned */
200 #define SLAB_CHUNK_SIZE(al, ix) (((ix) + 1) * P2ALIGNMENT)
201 #define SLAB_BPAGE_SIZE(al,csz) (8 * (csz) + SLAB_INFO_SIZE)
202
203 /* optimized version of ALIGN (size, P2ALIGNMENT) */
204 #if GLIB_SIZEOF_SIZE_T * 2 == 8 /* P2ALIGNMENT */
205 #define P2ALIGN(size) (((size) + 0x7) & ~(gsize) 0x7)
206 #elif GLIB_SIZEOF_SIZE_T * 2 == 16 /* P2ALIGNMENT */
207 #define P2ALIGN(size) (((size) + 0xf) & ~(gsize) 0xf)
208 #else
209 #define P2ALIGN(size) ALIGN (size, P2ALIGNMENT)
210 #endif
211
212 /* special helpers to avoid gmessage.c dependency */
213 static void mem_error (const char *format, ...) G_GNUC_PRINTF (1,2);
214 #define mem_assert(cond) do { if (G_LIKELY (cond)) ; else mem_error ("assertion failed: %s", #cond); } while (0)
215
216 /* --- structures --- */
217 typedef struct _ChunkLink ChunkLink;
218 typedef struct _SlabInfo SlabInfo;
219 typedef struct _CachedMagazine CachedMagazine;
220 struct _ChunkLink {
221 ChunkLink *next;
222 ChunkLink *data;
223 };
224 struct _SlabInfo {
225 ChunkLink *chunks;
226 guint n_allocated;
227 SlabInfo *next, *prev;
228 };
229 typedef struct {
230 ChunkLink *chunks;
231 gsize count; /* approximative chunks list length */
232 } Magazine;
233 typedef struct {
234 Magazine *magazine1; /* array of MAX_SLAB_INDEX (allocator) */
235 Magazine *magazine2; /* array of MAX_SLAB_INDEX (allocator) */
236 } ThreadMemory;
237 typedef struct {
238 gboolean always_malloc;
239 gboolean bypass_magazines;
240 gboolean debug_blocks;
241 gsize working_set_msecs;
242 guint color_increment;
243 } SliceConfig;
244 typedef struct {
245 /* const after initialization */
246 gsize min_page_size, max_page_size;
247 SliceConfig config;
248 gsize max_slab_chunk_size_for_magazine_cache;
249 /* magazine cache */
250 GMutex magazine_mutex;
251 ChunkLink **magazines; /* array of MAX_SLAB_INDEX (allocator) */
252 guint *contention_counters; /* array of MAX_SLAB_INDEX (allocator) */
253 gint mutex_counter;
254 guint stamp_counter;
255 guint last_stamp;
256 /* slab allocator */
257 GMutex slab_mutex;
258 SlabInfo **slab_stack; /* array of MAX_SLAB_INDEX (allocator) */
259 guint color_accu;
260 } Allocator;
261
262 /* --- g-slice prototypes --- */
263 static gpointer slab_allocator_alloc_chunk (gsize chunk_size);
264 static void slab_allocator_free_chunk (gsize chunk_size,
265 gpointer mem);
266 static void private_thread_memory_cleanup (gpointer data);
267 static gpointer allocator_memalign (gsize alignment,
268 gsize memsize);
269 static void allocator_memfree (gsize memsize,
270 gpointer mem);
271 static inline void magazine_cache_update_stamp (void);
272 static inline gsize allocator_get_magazine_threshold (Allocator *allocator,
273 guint ix);
274
275 /* --- g-slice memory checker --- */
276 static void smc_notify_alloc (void *pointer,
277 size_t size);
278 static int smc_notify_free (void *pointer,
279 size_t size);
280
281 /* --- variables --- */
282 static GPrivate private_thread_memory = G_PRIVATE_INIT (private_thread_memory_cleanup);
283 static gsize sys_page_size = 0;
284 static Allocator allocator[1] = { { 0, }, };
285 static SliceConfig slice_config = {
286 FALSE, /* always_malloc */
287 FALSE, /* bypass_magazines */
288 FALSE, /* debug_blocks */
289 15 * 1000, /* working_set_msecs */
290 1, /* color increment, alt: 0x7fffffff */
291 };
292 static GMutex smc_tree_mutex; /* mutex for G_SLICE=debug-blocks */
293
294 /* --- auxiliary funcitons --- */
295 void
g_slice_set_config(GSliceConfig ckey,gint64 value)296 g_slice_set_config (GSliceConfig ckey,
297 gint64 value)
298 {
299 g_return_if_fail (sys_page_size == 0);
300 switch (ckey)
301 {
302 case G_SLICE_CONFIG_ALWAYS_MALLOC:
303 slice_config.always_malloc = value != 0;
304 break;
305 case G_SLICE_CONFIG_BYPASS_MAGAZINES:
306 slice_config.bypass_magazines = value != 0;
307 break;
308 case G_SLICE_CONFIG_WORKING_SET_MSECS:
309 slice_config.working_set_msecs = value;
310 break;
311 case G_SLICE_CONFIG_COLOR_INCREMENT:
312 slice_config.color_increment = value;
313 default: ;
314 }
315 }
316
317 gint64
g_slice_get_config(GSliceConfig ckey)318 g_slice_get_config (GSliceConfig ckey)
319 {
320 switch (ckey)
321 {
322 case G_SLICE_CONFIG_ALWAYS_MALLOC:
323 return slice_config.always_malloc;
324 case G_SLICE_CONFIG_BYPASS_MAGAZINES:
325 return slice_config.bypass_magazines;
326 case G_SLICE_CONFIG_WORKING_SET_MSECS:
327 return slice_config.working_set_msecs;
328 case G_SLICE_CONFIG_CHUNK_SIZES:
329 return MAX_SLAB_INDEX (allocator);
330 case G_SLICE_CONFIG_COLOR_INCREMENT:
331 return slice_config.color_increment;
332 default:
333 return 0;
334 }
335 }
336
337 gint64*
g_slice_get_config_state(GSliceConfig ckey,gint64 address,guint * n_values)338 g_slice_get_config_state (GSliceConfig ckey,
339 gint64 address,
340 guint *n_values)
341 {
342 guint i = 0;
343 g_return_val_if_fail (n_values != NULL, NULL);
344 *n_values = 0;
345 switch (ckey)
346 {
347 gint64 array[64];
348 case G_SLICE_CONFIG_CONTENTION_COUNTER:
349 array[i++] = SLAB_CHUNK_SIZE (allocator, address);
350 array[i++] = allocator->contention_counters[address];
351 array[i++] = allocator_get_magazine_threshold (allocator, address);
352 *n_values = i;
353 return g_memdup (array, sizeof (array[0]) * *n_values);
354 default:
355 return NULL;
356 }
357 }
358
359 static void
slice_config_init(SliceConfig * config)360 slice_config_init (SliceConfig *config)
361 {
362 const gchar *val;
363
364 *config = slice_config;
365
366 val = getenv ("G_SLICE");
367 if (val != NULL)
368 {
369 gint flags;
370 const GDebugKey keys[] = {
371 { "always-malloc", 1 << 0 },
372 { "debug-blocks", 1 << 1 },
373 };
374
375 flags = g_parse_debug_string (val, keys, G_N_ELEMENTS (keys));
376 if (flags & (1 << 0))
377 config->always_malloc = TRUE;
378 if (flags & (1 << 1))
379 config->debug_blocks = TRUE;
380 }
381 else
382 {
383 /* G_SLICE was not specified, so check if valgrind is running and
384 * disable ourselves if it is.
385 *
386 * This way it's possible to force gslice to be enabled under
387 * valgrind just by setting G_SLICE to the empty string.
388 */
389 #ifdef ENABLE_VALGRIND
390 if (RUNNING_ON_VALGRIND)
391 config->always_malloc = TRUE;
392 #endif
393 }
394 }
395
396 static void
g_slice_init_nomessage(void)397 g_slice_init_nomessage (void)
398 {
399 /* we may not use g_error() or friends here */
400 mem_assert (sys_page_size == 0);
401 mem_assert (MIN_MAGAZINE_SIZE >= 4);
402
403 #ifdef G_OS_WIN32
404 {
405 SYSTEM_INFO system_info;
406 GetSystemInfo (&system_info);
407 sys_page_size = system_info.dwPageSize;
408 }
409 #else
410 sys_page_size = sysconf (_SC_PAGESIZE); /* = sysconf (_SC_PAGE_SIZE); = getpagesize(); */
411 #endif
412 mem_assert (sys_page_size >= 2 * LARGEALIGNMENT);
413 mem_assert ((sys_page_size & (sys_page_size - 1)) == 0);
414 slice_config_init (&allocator->config);
415 allocator->min_page_size = sys_page_size;
416 #if HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN
417 /* allow allocation of pages up to 8KB (with 8KB alignment).
418 * this is useful because many medium to large sized structures
419 * fit less than 8 times (see [4]) into 4KB pages.
420 * we allow very small page sizes here, to reduce wastage in
421 * threads if only small allocations are required (this does
422 * bear the risk of increasing allocation times and fragmentation
423 * though).
424 */
425 allocator->min_page_size = MAX (allocator->min_page_size, 4096);
426 allocator->max_page_size = MAX (allocator->min_page_size, 8192);
427 allocator->min_page_size = MIN (allocator->min_page_size, 128);
428 #else
429 /* we can only align to system page size */
430 allocator->max_page_size = sys_page_size;
431 #endif
432 if (allocator->config.always_malloc)
433 {
434 allocator->contention_counters = NULL;
435 allocator->magazines = NULL;
436 allocator->slab_stack = NULL;
437 }
438 else
439 {
440 allocator->contention_counters = g_new0 (guint, MAX_SLAB_INDEX (allocator));
441 allocator->magazines = g_new0 (ChunkLink*, MAX_SLAB_INDEX (allocator));
442 allocator->slab_stack = g_new0 (SlabInfo*, MAX_SLAB_INDEX (allocator));
443 }
444
445 allocator->mutex_counter = 0;
446 allocator->stamp_counter = MAX_STAMP_COUNTER; /* force initial update */
447 allocator->last_stamp = 0;
448 allocator->color_accu = 0;
449 magazine_cache_update_stamp();
450 /* values cached for performance reasons */
451 allocator->max_slab_chunk_size_for_magazine_cache = MAX_SLAB_CHUNK_SIZE (allocator);
452 if (allocator->config.always_malloc || allocator->config.bypass_magazines)
453 allocator->max_slab_chunk_size_for_magazine_cache = 0; /* non-optimized cases */
454 }
455
456 static inline guint
allocator_categorize(gsize aligned_chunk_size)457 allocator_categorize (gsize aligned_chunk_size)
458 {
459 /* speed up the likely path */
460 if (G_LIKELY (aligned_chunk_size && aligned_chunk_size <= allocator->max_slab_chunk_size_for_magazine_cache))
461 return 1; /* use magazine cache */
462
463 if (!allocator->config.always_malloc &&
464 aligned_chunk_size &&
465 aligned_chunk_size <= MAX_SLAB_CHUNK_SIZE (allocator))
466 {
467 if (allocator->config.bypass_magazines)
468 return 2; /* use slab allocator, see [2] */
469 return 1; /* use magazine cache */
470 }
471 return 0; /* use malloc() */
472 }
473
474 static inline void
g_mutex_lock_a(GMutex * mutex,guint * contention_counter)475 g_mutex_lock_a (GMutex *mutex,
476 guint *contention_counter)
477 {
478 gboolean contention = FALSE;
479 if (!g_mutex_trylock (mutex))
480 {
481 g_mutex_lock (mutex);
482 contention = TRUE;
483 }
484 if (contention)
485 {
486 allocator->mutex_counter++;
487 if (allocator->mutex_counter >= 1) /* quickly adapt to contention */
488 {
489 allocator->mutex_counter = 0;
490 *contention_counter = MIN (*contention_counter + 1, MAX_MAGAZINE_SIZE);
491 }
492 }
493 else /* !contention */
494 {
495 allocator->mutex_counter--;
496 if (allocator->mutex_counter < -11) /* moderately recover magazine sizes */
497 {
498 allocator->mutex_counter = 0;
499 *contention_counter = MAX (*contention_counter, 1) - 1;
500 }
501 }
502 }
503
504 static inline ThreadMemory*
thread_memory_from_self(void)505 thread_memory_from_self (void)
506 {
507 ThreadMemory *tmem = g_private_get (&private_thread_memory);
508 if (G_UNLIKELY (!tmem))
509 {
510 static GMutex init_mutex;
511 guint n_magazines;
512
513 g_mutex_lock (&init_mutex);
514 if G_UNLIKELY (sys_page_size == 0)
515 g_slice_init_nomessage ();
516 g_mutex_unlock (&init_mutex);
517
518 n_magazines = MAX_SLAB_INDEX (allocator);
519 tmem = g_private_set_alloc0 (&private_thread_memory, sizeof (ThreadMemory) + sizeof (Magazine) * 2 * n_magazines);
520 tmem->magazine1 = (Magazine*) (tmem + 1);
521 tmem->magazine2 = &tmem->magazine1[n_magazines];
522 }
523 return tmem;
524 }
525
526 static inline ChunkLink*
magazine_chain_pop_head(ChunkLink ** magazine_chunks)527 magazine_chain_pop_head (ChunkLink **magazine_chunks)
528 {
529 /* magazine chains are linked via ChunkLink->next.
530 * each ChunkLink->data of the toplevel chain may point to a subchain,
531 * linked via ChunkLink->next. ChunkLink->data of the subchains just
532 * contains uninitialized junk.
533 */
534 ChunkLink *chunk = (*magazine_chunks)->data;
535 if (G_UNLIKELY (chunk))
536 {
537 /* allocating from freed list */
538 (*magazine_chunks)->data = chunk->next;
539 }
540 else
541 {
542 chunk = *magazine_chunks;
543 *magazine_chunks = chunk->next;
544 }
545 return chunk;
546 }
547
548 #if 0 /* useful for debugging */
549 static guint
550 magazine_count (ChunkLink *head)
551 {
552 guint count = 0;
553 if (!head)
554 return 0;
555 while (head)
556 {
557 ChunkLink *child = head->data;
558 count += 1;
559 for (child = head->data; child; child = child->next)
560 count += 1;
561 head = head->next;
562 }
563 return count;
564 }
565 #endif
566
567 static inline gsize
allocator_get_magazine_threshold(Allocator * allocator,guint ix)568 allocator_get_magazine_threshold (Allocator *allocator,
569 guint ix)
570 {
571 /* the magazine size calculated here has a lower bound of MIN_MAGAZINE_SIZE,
572 * which is required by the implementation. also, for moderately sized chunks
573 * (say >= 64 bytes), magazine sizes shouldn't be much smaller then the number
574 * of chunks available per page/2 to avoid excessive traffic in the magazine
575 * cache for small to medium sized structures.
576 * the upper bound of the magazine size is effectively provided by
577 * MAX_MAGAZINE_SIZE. for larger chunks, this number is scaled down so that
578 * the content of a single magazine doesn't exceed ca. 16KB.
579 */
580 gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
581 guint threshold = MAX (MIN_MAGAZINE_SIZE, allocator->max_page_size / MAX (5 * chunk_size, 5 * 32));
582 guint contention_counter = allocator->contention_counters[ix];
583 if (G_UNLIKELY (contention_counter)) /* single CPU bias */
584 {
585 /* adapt contention counter thresholds to chunk sizes */
586 contention_counter = contention_counter * 64 / chunk_size;
587 threshold = MAX (threshold, contention_counter);
588 }
589 return threshold;
590 }
591
592 /* --- magazine cache --- */
593 static inline void
magazine_cache_update_stamp(void)594 magazine_cache_update_stamp (void)
595 {
596 if (allocator->stamp_counter >= MAX_STAMP_COUNTER)
597 {
598 gint64 now_us = g_get_real_time ();
599 allocator->last_stamp = now_us / 1000; /* milli seconds */
600 allocator->stamp_counter = 0;
601 }
602 else
603 allocator->stamp_counter++;
604 }
605
606 static inline ChunkLink*
magazine_chain_prepare_fields(ChunkLink * magazine_chunks)607 magazine_chain_prepare_fields (ChunkLink *magazine_chunks)
608 {
609 ChunkLink *chunk1;
610 ChunkLink *chunk2;
611 ChunkLink *chunk3;
612 ChunkLink *chunk4;
613 /* checked upon initialization: mem_assert (MIN_MAGAZINE_SIZE >= 4); */
614 /* ensure a magazine with at least 4 unused data pointers */
615 chunk1 = magazine_chain_pop_head (&magazine_chunks);
616 chunk2 = magazine_chain_pop_head (&magazine_chunks);
617 chunk3 = magazine_chain_pop_head (&magazine_chunks);
618 chunk4 = magazine_chain_pop_head (&magazine_chunks);
619 chunk4->next = magazine_chunks;
620 chunk3->next = chunk4;
621 chunk2->next = chunk3;
622 chunk1->next = chunk2;
623 return chunk1;
624 }
625
626 /* access the first 3 fields of a specially prepared magazine chain */
627 #define magazine_chain_prev(mc) ((mc)->data)
628 #define magazine_chain_stamp(mc) ((mc)->next->data)
629 #define magazine_chain_uint_stamp(mc) GPOINTER_TO_UINT ((mc)->next->data)
630 #define magazine_chain_next(mc) ((mc)->next->next->data)
631 #define magazine_chain_count(mc) ((mc)->next->next->next->data)
632
633 static void
magazine_cache_trim(Allocator * allocator,guint ix,guint stamp)634 magazine_cache_trim (Allocator *allocator,
635 guint ix,
636 guint stamp)
637 {
638 /* g_mutex_lock (allocator->mutex); done by caller */
639 /* trim magazine cache from tail */
640 ChunkLink *current = magazine_chain_prev (allocator->magazines[ix]);
641 ChunkLink *trash = NULL;
642 while (!G_APPROX_VALUE(stamp, magazine_chain_uint_stamp (current),
643 allocator->config.working_set_msecs))
644 {
645 /* unlink */
646 ChunkLink *prev = magazine_chain_prev (current);
647 ChunkLink *next = magazine_chain_next (current);
648 magazine_chain_next (prev) = next;
649 magazine_chain_prev (next) = prev;
650 /* clear special fields, put on trash stack */
651 magazine_chain_next (current) = NULL;
652 magazine_chain_count (current) = NULL;
653 magazine_chain_stamp (current) = NULL;
654 magazine_chain_prev (current) = trash;
655 trash = current;
656 /* fixup list head if required */
657 if (current == allocator->magazines[ix])
658 {
659 allocator->magazines[ix] = NULL;
660 break;
661 }
662 current = prev;
663 }
664 g_mutex_unlock (&allocator->magazine_mutex);
665 /* free trash */
666 if (trash)
667 {
668 const gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
669 g_mutex_lock (&allocator->slab_mutex);
670 while (trash)
671 {
672 current = trash;
673 trash = magazine_chain_prev (current);
674 magazine_chain_prev (current) = NULL; /* clear special field */
675 while (current)
676 {
677 ChunkLink *chunk = magazine_chain_pop_head (¤t);
678 slab_allocator_free_chunk (chunk_size, chunk);
679 }
680 }
681 g_mutex_unlock (&allocator->slab_mutex);
682 }
683 }
684
685 static void
magazine_cache_push_magazine(guint ix,ChunkLink * magazine_chunks,gsize count)686 magazine_cache_push_magazine (guint ix,
687 ChunkLink *magazine_chunks,
688 gsize count) /* must be >= MIN_MAGAZINE_SIZE */
689 {
690 ChunkLink *current = magazine_chain_prepare_fields (magazine_chunks);
691 ChunkLink *next, *prev;
692 g_mutex_lock (&allocator->magazine_mutex);
693 /* add magazine at head */
694 next = allocator->magazines[ix];
695 if (next)
696 prev = magazine_chain_prev (next);
697 else
698 next = prev = current;
699 magazine_chain_next (prev) = current;
700 magazine_chain_prev (next) = current;
701 magazine_chain_prev (current) = prev;
702 magazine_chain_next (current) = next;
703 magazine_chain_count (current) = (gpointer) count;
704 /* stamp magazine */
705 magazine_cache_update_stamp();
706 magazine_chain_stamp (current) = GUINT_TO_POINTER (allocator->last_stamp);
707 allocator->magazines[ix] = current;
708 /* free old magazines beyond a certain threshold */
709 magazine_cache_trim (allocator, ix, allocator->last_stamp);
710 /* g_mutex_unlock (allocator->mutex); was done by magazine_cache_trim() */
711 }
712
713 static ChunkLink*
magazine_cache_pop_magazine(guint ix,gsize * countp)714 magazine_cache_pop_magazine (guint ix,
715 gsize *countp)
716 {
717 g_mutex_lock_a (&allocator->magazine_mutex, &allocator->contention_counters[ix]);
718 if (!allocator->magazines[ix])
719 {
720 guint magazine_threshold = allocator_get_magazine_threshold (allocator, ix);
721 gsize i, chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
722 ChunkLink *chunk, *head;
723 g_mutex_unlock (&allocator->magazine_mutex);
724 g_mutex_lock (&allocator->slab_mutex);
725 head = slab_allocator_alloc_chunk (chunk_size);
726 head->data = NULL;
727 chunk = head;
728 for (i = 1; i < magazine_threshold; i++)
729 {
730 chunk->next = slab_allocator_alloc_chunk (chunk_size);
731 chunk = chunk->next;
732 chunk->data = NULL;
733 }
734 chunk->next = NULL;
735 g_mutex_unlock (&allocator->slab_mutex);
736 *countp = i;
737 return head;
738 }
739 else
740 {
741 ChunkLink *current = allocator->magazines[ix];
742 ChunkLink *prev = magazine_chain_prev (current);
743 ChunkLink *next = magazine_chain_next (current);
744 /* unlink */
745 magazine_chain_next (prev) = next;
746 magazine_chain_prev (next) = prev;
747 allocator->magazines[ix] = next == current ? NULL : next;
748 g_mutex_unlock (&allocator->magazine_mutex);
749 /* clear special fields and hand out */
750 *countp = (gsize) magazine_chain_count (current);
751 magazine_chain_prev (current) = NULL;
752 magazine_chain_next (current) = NULL;
753 magazine_chain_count (current) = NULL;
754 magazine_chain_stamp (current) = NULL;
755 return current;
756 }
757 }
758
759 /* --- thread magazines --- */
760 static void
private_thread_memory_cleanup(gpointer data)761 private_thread_memory_cleanup (gpointer data)
762 {
763 ThreadMemory *tmem = data;
764 const guint n_magazines = MAX_SLAB_INDEX (allocator);
765 guint ix;
766 for (ix = 0; ix < n_magazines; ix++)
767 {
768 Magazine *mags[2];
769 guint j;
770 mags[0] = &tmem->magazine1[ix];
771 mags[1] = &tmem->magazine2[ix];
772 for (j = 0; j < 2; j++)
773 {
774 Magazine *mag = mags[j];
775 if (mag->count >= MIN_MAGAZINE_SIZE)
776 magazine_cache_push_magazine (ix, mag->chunks, mag->count);
777 else
778 {
779 const gsize chunk_size = SLAB_CHUNK_SIZE (allocator, ix);
780 g_mutex_lock (&allocator->slab_mutex);
781 while (mag->chunks)
782 {
783 ChunkLink *chunk = magazine_chain_pop_head (&mag->chunks);
784 slab_allocator_free_chunk (chunk_size, chunk);
785 }
786 g_mutex_unlock (&allocator->slab_mutex);
787 }
788 }
789 }
790 g_free (tmem);
791 }
792
793 static void
thread_memory_magazine1_reload(ThreadMemory * tmem,guint ix)794 thread_memory_magazine1_reload (ThreadMemory *tmem,
795 guint ix)
796 {
797 Magazine *mag = &tmem->magazine1[ix];
798 mem_assert (mag->chunks == NULL); /* ensure that we may reset mag->count */
799 mag->count = 0;
800 mag->chunks = magazine_cache_pop_magazine (ix, &mag->count);
801 }
802
803 static void
thread_memory_magazine2_unload(ThreadMemory * tmem,guint ix)804 thread_memory_magazine2_unload (ThreadMemory *tmem,
805 guint ix)
806 {
807 Magazine *mag = &tmem->magazine2[ix];
808 magazine_cache_push_magazine (ix, mag->chunks, mag->count);
809 mag->chunks = NULL;
810 mag->count = 0;
811 }
812
813 static inline void
thread_memory_swap_magazines(ThreadMemory * tmem,guint ix)814 thread_memory_swap_magazines (ThreadMemory *tmem,
815 guint ix)
816 {
817 Magazine xmag = tmem->magazine1[ix];
818 tmem->magazine1[ix] = tmem->magazine2[ix];
819 tmem->magazine2[ix] = xmag;
820 }
821
822 static inline gboolean
thread_memory_magazine1_is_empty(ThreadMemory * tmem,guint ix)823 thread_memory_magazine1_is_empty (ThreadMemory *tmem,
824 guint ix)
825 {
826 return tmem->magazine1[ix].chunks == NULL;
827 }
828
829 static inline gboolean
thread_memory_magazine2_is_full(ThreadMemory * tmem,guint ix)830 thread_memory_magazine2_is_full (ThreadMemory *tmem,
831 guint ix)
832 {
833 return tmem->magazine2[ix].count >= allocator_get_magazine_threshold (allocator, ix);
834 }
835
836 static inline gpointer
thread_memory_magazine1_alloc(ThreadMemory * tmem,guint ix)837 thread_memory_magazine1_alloc (ThreadMemory *tmem,
838 guint ix)
839 {
840 Magazine *mag = &tmem->magazine1[ix];
841 ChunkLink *chunk = magazine_chain_pop_head (&mag->chunks);
842 if (G_LIKELY (mag->count > 0))
843 mag->count--;
844 return chunk;
845 }
846
847 static inline void
thread_memory_magazine2_free(ThreadMemory * tmem,guint ix,gpointer mem)848 thread_memory_magazine2_free (ThreadMemory *tmem,
849 guint ix,
850 gpointer mem)
851 {
852 Magazine *mag = &tmem->magazine2[ix];
853 ChunkLink *chunk = mem;
854 chunk->data = NULL;
855 chunk->next = mag->chunks;
856 mag->chunks = chunk;
857 mag->count++;
858 }
859
860 /* --- API functions --- */
861
862 /**
863 * g_slice_new:
864 * @type: the type to allocate, typically a structure name
865 *
866 * A convenience macro to allocate a block of memory from the
867 * slice allocator.
868 *
869 * It calls g_slice_alloc() with `sizeof (@type)` and casts the
870 * returned pointer to a pointer of the given type, avoiding a type
871 * cast in the source code. Note that the underlying slice allocation
872 * mechanism can be changed with the [`G_SLICE=always-malloc`][G_SLICE]
873 * environment variable.
874 *
875 * This can never return %NULL as the minimum allocation size from
876 * `sizeof (@type)` is 1 byte.
877 *
878 * Returns: (not nullable): a pointer to the allocated block, cast to a pointer
879 * to @type
880 *
881 * Since: 2.10
882 */
883
884 /**
885 * g_slice_new0:
886 * @type: the type to allocate, typically a structure name
887 *
888 * A convenience macro to allocate a block of memory from the
889 * slice allocator and set the memory to 0.
890 *
891 * It calls g_slice_alloc0() with `sizeof (@type)`
892 * and casts the returned pointer to a pointer of the given type,
893 * avoiding a type cast in the source code.
894 * Note that the underlying slice allocation mechanism can
895 * be changed with the [`G_SLICE=always-malloc`][G_SLICE]
896 * environment variable.
897 *
898 * This can never return %NULL as the minimum allocation size from
899 * `sizeof (@type)` is 1 byte.
900 *
901 * Returns: (not nullable): a pointer to the allocated block, cast to a pointer
902 * to @type
903 *
904 * Since: 2.10
905 */
906
907 /**
908 * g_slice_dup:
909 * @type: the type to duplicate, typically a structure name
910 * @mem: (not nullable): the memory to copy into the allocated block
911 *
912 * A convenience macro to duplicate a block of memory using
913 * the slice allocator.
914 *
915 * It calls g_slice_copy() with `sizeof (@type)`
916 * and casts the returned pointer to a pointer of the given type,
917 * avoiding a type cast in the source code.
918 * Note that the underlying slice allocation mechanism can
919 * be changed with the [`G_SLICE=always-malloc`][G_SLICE]
920 * environment variable.
921 *
922 * This can never return %NULL.
923 *
924 * Returns: (not nullable): a pointer to the allocated block, cast to a pointer
925 * to @type
926 *
927 * Since: 2.14
928 */
929
930 /**
931 * g_slice_free:
932 * @type: the type of the block to free, typically a structure name
933 * @mem: a pointer to the block to free
934 *
935 * A convenience macro to free a block of memory that has
936 * been allocated from the slice allocator.
937 *
938 * It calls g_slice_free1() using `sizeof (type)`
939 * as the block size.
940 * Note that the exact release behaviour can be changed with the
941 * [`G_DEBUG=gc-friendly`][G_DEBUG] environment variable, also see
942 * [`G_SLICE`][G_SLICE] for related debugging options.
943 *
944 * If @mem is %NULL, this macro does nothing.
945 *
946 * Since: 2.10
947 */
948
949 /**
950 * g_slice_free_chain:
951 * @type: the type of the @mem_chain blocks
952 * @mem_chain: a pointer to the first block of the chain
953 * @next: the field name of the next pointer in @type
954 *
955 * Frees a linked list of memory blocks of structure type @type.
956 * The memory blocks must be equal-sized, allocated via
957 * g_slice_alloc() or g_slice_alloc0() and linked together by
958 * a @next pointer (similar to #GSList). The name of the
959 * @next field in @type is passed as third argument.
960 * Note that the exact release behaviour can be changed with the
961 * [`G_DEBUG=gc-friendly`][G_DEBUG] environment variable, also see
962 * [`G_SLICE`][G_SLICE] for related debugging options.
963 *
964 * If @mem_chain is %NULL, this function does nothing.
965 *
966 * Since: 2.10
967 */
968
969 /**
970 * g_slice_alloc:
971 * @block_size: the number of bytes to allocate
972 *
973 * Allocates a block of memory from the slice allocator.
974 * The block address handed out can be expected to be aligned
975 * to at least 1 * sizeof (void*),
976 * though in general slices are 2 * sizeof (void*) bytes aligned,
977 * if a malloc() fallback implementation is used instead,
978 * the alignment may be reduced in a libc dependent fashion.
979 * Note that the underlying slice allocation mechanism can
980 * be changed with the [`G_SLICE=always-malloc`][G_SLICE]
981 * environment variable.
982 *
983 * Returns: a pointer to the allocated memory block, which will be %NULL if and
984 * only if @mem_size is 0
985 *
986 * Since: 2.10
987 */
988 gpointer
g_slice_alloc(gsize mem_size)989 g_slice_alloc (gsize mem_size)
990 {
991 ThreadMemory *tmem;
992 gsize chunk_size;
993 gpointer mem;
994 guint acat;
995
996 /* This gets the private structure for this thread. If the private
997 * structure does not yet exist, it is created.
998 *
999 * This has a side effect of causing GSlice to be initialised, so it
1000 * must come first.
1001 */
1002 tmem = thread_memory_from_self ();
1003
1004 chunk_size = P2ALIGN (mem_size);
1005 acat = allocator_categorize (chunk_size);
1006 if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
1007 {
1008 guint ix = SLAB_INDEX (allocator, chunk_size);
1009 if (G_UNLIKELY (thread_memory_magazine1_is_empty (tmem, ix)))
1010 {
1011 thread_memory_swap_magazines (tmem, ix);
1012 if (G_UNLIKELY (thread_memory_magazine1_is_empty (tmem, ix)))
1013 thread_memory_magazine1_reload (tmem, ix);
1014 }
1015 mem = thread_memory_magazine1_alloc (tmem, ix);
1016 }
1017 else if (acat == 2) /* allocate through slab allocator */
1018 {
1019 g_mutex_lock (&allocator->slab_mutex);
1020 mem = slab_allocator_alloc_chunk (chunk_size);
1021 g_mutex_unlock (&allocator->slab_mutex);
1022 }
1023 else /* delegate to system malloc */
1024 mem = g_malloc (mem_size);
1025 if (G_UNLIKELY (allocator->config.debug_blocks))
1026 smc_notify_alloc (mem, mem_size);
1027
1028 TRACE (GLIB_SLICE_ALLOC((void*)mem, mem_size));
1029
1030 return mem;
1031 }
1032
1033 /**
1034 * g_slice_alloc0:
1035 * @block_size: the number of bytes to allocate
1036 *
1037 * Allocates a block of memory via g_slice_alloc() and initializes
1038 * the returned memory to 0. Note that the underlying slice allocation
1039 * mechanism can be changed with the [`G_SLICE=always-malloc`][G_SLICE]
1040 * environment variable.
1041 *
1042 * Returns: a pointer to the allocated block, which will be %NULL if and only
1043 * if @mem_size is 0
1044 *
1045 * Since: 2.10
1046 */
1047 gpointer
g_slice_alloc0(gsize mem_size)1048 g_slice_alloc0 (gsize mem_size)
1049 {
1050 gpointer mem = g_slice_alloc (mem_size);
1051 if (mem)
1052 memset (mem, 0, mem_size);
1053 return mem;
1054 }
1055
1056 /**
1057 * g_slice_copy:
1058 * @block_size: the number of bytes to allocate
1059 * @mem_block: the memory to copy
1060 *
1061 * Allocates a block of memory from the slice allocator
1062 * and copies @block_size bytes into it from @mem_block.
1063 *
1064 * @mem_block must be non-%NULL if @block_size is non-zero.
1065 *
1066 * Returns: a pointer to the allocated memory block, which will be %NULL if and
1067 * only if @mem_size is 0
1068 *
1069 * Since: 2.14
1070 */
1071 gpointer
g_slice_copy(gsize mem_size,gconstpointer mem_block)1072 g_slice_copy (gsize mem_size,
1073 gconstpointer mem_block)
1074 {
1075 gpointer mem = g_slice_alloc (mem_size);
1076 if (mem)
1077 memcpy (mem, mem_block, mem_size);
1078 return mem;
1079 }
1080
1081 /**
1082 * g_slice_free1:
1083 * @block_size: the size of the block
1084 * @mem_block: a pointer to the block to free
1085 *
1086 * Frees a block of memory.
1087 *
1088 * The memory must have been allocated via g_slice_alloc() or
1089 * g_slice_alloc0() and the @block_size has to match the size
1090 * specified upon allocation. Note that the exact release behaviour
1091 * can be changed with the [`G_DEBUG=gc-friendly`][G_DEBUG] environment
1092 * variable, also see [`G_SLICE`][G_SLICE] for related debugging options.
1093 *
1094 * If @mem_block is %NULL, this function does nothing.
1095 *
1096 * Since: 2.10
1097 */
1098 void
g_slice_free1(gsize mem_size,gpointer mem_block)1099 g_slice_free1 (gsize mem_size,
1100 gpointer mem_block)
1101 {
1102 gsize chunk_size = P2ALIGN (mem_size);
1103 guint acat = allocator_categorize (chunk_size);
1104 if (G_UNLIKELY (!mem_block))
1105 return;
1106 if (G_UNLIKELY (allocator->config.debug_blocks) &&
1107 !smc_notify_free (mem_block, mem_size))
1108 abort();
1109 if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
1110 {
1111 ThreadMemory *tmem = thread_memory_from_self();
1112 guint ix = SLAB_INDEX (allocator, chunk_size);
1113 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
1114 {
1115 thread_memory_swap_magazines (tmem, ix);
1116 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
1117 thread_memory_magazine2_unload (tmem, ix);
1118 }
1119 if (G_UNLIKELY (g_mem_gc_friendly))
1120 memset (mem_block, 0, chunk_size);
1121 thread_memory_magazine2_free (tmem, ix, mem_block);
1122 }
1123 else if (acat == 2) /* allocate through slab allocator */
1124 {
1125 if (G_UNLIKELY (g_mem_gc_friendly))
1126 memset (mem_block, 0, chunk_size);
1127 g_mutex_lock (&allocator->slab_mutex);
1128 slab_allocator_free_chunk (chunk_size, mem_block);
1129 g_mutex_unlock (&allocator->slab_mutex);
1130 }
1131 else /* delegate to system malloc */
1132 {
1133 if (G_UNLIKELY (g_mem_gc_friendly))
1134 memset (mem_block, 0, mem_size);
1135 g_free (mem_block);
1136 }
1137 TRACE (GLIB_SLICE_FREE((void*)mem_block, mem_size));
1138 }
1139
1140 /**
1141 * g_slice_free_chain_with_offset:
1142 * @block_size: the size of the blocks
1143 * @mem_chain: a pointer to the first block of the chain
1144 * @next_offset: the offset of the @next field in the blocks
1145 *
1146 * Frees a linked list of memory blocks of structure type @type.
1147 *
1148 * The memory blocks must be equal-sized, allocated via
1149 * g_slice_alloc() or g_slice_alloc0() and linked together by a
1150 * @next pointer (similar to #GSList). The offset of the @next
1151 * field in each block is passed as third argument.
1152 * Note that the exact release behaviour can be changed with the
1153 * [`G_DEBUG=gc-friendly`][G_DEBUG] environment variable, also see
1154 * [`G_SLICE`][G_SLICE] for related debugging options.
1155 *
1156 * If @mem_chain is %NULL, this function does nothing.
1157 *
1158 * Since: 2.10
1159 */
1160 void
g_slice_free_chain_with_offset(gsize mem_size,gpointer mem_chain,gsize next_offset)1161 g_slice_free_chain_with_offset (gsize mem_size,
1162 gpointer mem_chain,
1163 gsize next_offset)
1164 {
1165 gpointer slice = mem_chain;
1166 /* while the thread magazines and the magazine cache are implemented so that
1167 * they can easily be extended to allow for free lists containing more free
1168 * lists for the first level nodes, which would allow O(1) freeing in this
1169 * function, the benefit of such an extension is questionable, because:
1170 * - the magazine size counts will become mere lower bounds which confuses
1171 * the code adapting to lock contention;
1172 * - freeing a single node to the thread magazines is very fast, so this
1173 * O(list_length) operation is multiplied by a fairly small factor;
1174 * - memory usage histograms on larger applications seem to indicate that
1175 * the amount of released multi node lists is negligible in comparison
1176 * to single node releases.
1177 * - the major performance bottle neck, namely g_private_get() or
1178 * g_mutex_lock()/g_mutex_unlock() has already been moved out of the
1179 * inner loop for freeing chained slices.
1180 */
1181 gsize chunk_size = P2ALIGN (mem_size);
1182 guint acat = allocator_categorize (chunk_size);
1183 if (G_LIKELY (acat == 1)) /* allocate through magazine layer */
1184 {
1185 ThreadMemory *tmem = thread_memory_from_self();
1186 guint ix = SLAB_INDEX (allocator, chunk_size);
1187 while (slice)
1188 {
1189 guint8 *current = slice;
1190 slice = *(gpointer*) (current + next_offset);
1191 if (G_UNLIKELY (allocator->config.debug_blocks) &&
1192 !smc_notify_free (current, mem_size))
1193 abort();
1194 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
1195 {
1196 thread_memory_swap_magazines (tmem, ix);
1197 if (G_UNLIKELY (thread_memory_magazine2_is_full (tmem, ix)))
1198 thread_memory_magazine2_unload (tmem, ix);
1199 }
1200 if (G_UNLIKELY (g_mem_gc_friendly))
1201 memset (current, 0, chunk_size);
1202 thread_memory_magazine2_free (tmem, ix, current);
1203 }
1204 }
1205 else if (acat == 2) /* allocate through slab allocator */
1206 {
1207 g_mutex_lock (&allocator->slab_mutex);
1208 while (slice)
1209 {
1210 guint8 *current = slice;
1211 slice = *(gpointer*) (current + next_offset);
1212 if (G_UNLIKELY (allocator->config.debug_blocks) &&
1213 !smc_notify_free (current, mem_size))
1214 abort();
1215 if (G_UNLIKELY (g_mem_gc_friendly))
1216 memset (current, 0, chunk_size);
1217 slab_allocator_free_chunk (chunk_size, current);
1218 }
1219 g_mutex_unlock (&allocator->slab_mutex);
1220 }
1221 else /* delegate to system malloc */
1222 while (slice)
1223 {
1224 guint8 *current = slice;
1225 slice = *(gpointer*) (current + next_offset);
1226 if (G_UNLIKELY (allocator->config.debug_blocks) &&
1227 !smc_notify_free (current, mem_size))
1228 abort();
1229 if (G_UNLIKELY (g_mem_gc_friendly))
1230 memset (current, 0, mem_size);
1231 g_free (current);
1232 }
1233 }
1234
1235 /* --- single page allocator --- */
1236 static void
allocator_slab_stack_push(Allocator * allocator,guint ix,SlabInfo * sinfo)1237 allocator_slab_stack_push (Allocator *allocator,
1238 guint ix,
1239 SlabInfo *sinfo)
1240 {
1241 /* insert slab at slab ring head */
1242 if (!allocator->slab_stack[ix])
1243 {
1244 sinfo->next = sinfo;
1245 sinfo->prev = sinfo;
1246 }
1247 else
1248 {
1249 SlabInfo *next = allocator->slab_stack[ix], *prev = next->prev;
1250 next->prev = sinfo;
1251 prev->next = sinfo;
1252 sinfo->next = next;
1253 sinfo->prev = prev;
1254 }
1255 allocator->slab_stack[ix] = sinfo;
1256 }
1257
1258 static gsize
allocator_aligned_page_size(Allocator * allocator,gsize n_bytes)1259 allocator_aligned_page_size (Allocator *allocator,
1260 gsize n_bytes)
1261 {
1262 gsize val = 1 << g_bit_storage (n_bytes - 1);
1263 val = MAX (val, allocator->min_page_size);
1264 return val;
1265 }
1266
1267 static void
allocator_add_slab(Allocator * allocator,guint ix,gsize chunk_size)1268 allocator_add_slab (Allocator *allocator,
1269 guint ix,
1270 gsize chunk_size)
1271 {
1272 ChunkLink *chunk;
1273 SlabInfo *sinfo;
1274 gsize addr, padding, n_chunks, color = 0;
1275 gsize page_size;
1276 int errsv;
1277 gpointer aligned_memory;
1278 guint8 *mem;
1279 guint i;
1280
1281 page_size = allocator_aligned_page_size (allocator, SLAB_BPAGE_SIZE (allocator, chunk_size));
1282 /* allocate 1 page for the chunks and the slab */
1283 aligned_memory = allocator_memalign (page_size, page_size - NATIVE_MALLOC_PADDING);
1284 errsv = errno;
1285 mem = aligned_memory;
1286
1287 if (!mem)
1288 {
1289 const gchar *syserr = strerror (errsv);
1290 mem_error ("failed to allocate %u bytes (alignment: %u): %s\n",
1291 (guint) (page_size - NATIVE_MALLOC_PADDING), (guint) page_size, syserr);
1292 }
1293 /* mask page address */
1294 addr = ((gsize) mem / page_size) * page_size;
1295 /* assert alignment */
1296 mem_assert (aligned_memory == (gpointer) addr);
1297 /* basic slab info setup */
1298 sinfo = (SlabInfo*) (mem + page_size - SLAB_INFO_SIZE);
1299 sinfo->n_allocated = 0;
1300 sinfo->chunks = NULL;
1301 /* figure cache colorization */
1302 n_chunks = ((guint8*) sinfo - mem) / chunk_size;
1303 padding = ((guint8*) sinfo - mem) - n_chunks * chunk_size;
1304 if (padding)
1305 {
1306 color = (allocator->color_accu * P2ALIGNMENT) % padding;
1307 allocator->color_accu += allocator->config.color_increment;
1308 }
1309 /* add chunks to free list */
1310 chunk = (ChunkLink*) (mem + color);
1311 sinfo->chunks = chunk;
1312 for (i = 0; i < n_chunks - 1; i++)
1313 {
1314 chunk->next = (ChunkLink*) ((guint8*) chunk + chunk_size);
1315 chunk = chunk->next;
1316 }
1317 chunk->next = NULL; /* last chunk */
1318 /* add slab to slab ring */
1319 allocator_slab_stack_push (allocator, ix, sinfo);
1320 }
1321
1322 static gpointer
slab_allocator_alloc_chunk(gsize chunk_size)1323 slab_allocator_alloc_chunk (gsize chunk_size)
1324 {
1325 ChunkLink *chunk;
1326 guint ix = SLAB_INDEX (allocator, chunk_size);
1327 /* ensure non-empty slab */
1328 if (!allocator->slab_stack[ix] || !allocator->slab_stack[ix]->chunks)
1329 allocator_add_slab (allocator, ix, chunk_size);
1330 /* allocate chunk */
1331 chunk = allocator->slab_stack[ix]->chunks;
1332 allocator->slab_stack[ix]->chunks = chunk->next;
1333 allocator->slab_stack[ix]->n_allocated++;
1334 /* rotate empty slabs */
1335 if (!allocator->slab_stack[ix]->chunks)
1336 allocator->slab_stack[ix] = allocator->slab_stack[ix]->next;
1337 return chunk;
1338 }
1339
1340 static void
slab_allocator_free_chunk(gsize chunk_size,gpointer mem)1341 slab_allocator_free_chunk (gsize chunk_size,
1342 gpointer mem)
1343 {
1344 ChunkLink *chunk;
1345 gboolean was_empty;
1346 guint ix = SLAB_INDEX (allocator, chunk_size);
1347 gsize page_size = allocator_aligned_page_size (allocator, SLAB_BPAGE_SIZE (allocator, chunk_size));
1348 gsize addr = ((gsize) mem / page_size) * page_size;
1349 /* mask page address */
1350 guint8 *page = (guint8*) addr;
1351 SlabInfo *sinfo = (SlabInfo*) (page + page_size - SLAB_INFO_SIZE);
1352 /* assert valid chunk count */
1353 mem_assert (sinfo->n_allocated > 0);
1354 /* add chunk to free list */
1355 was_empty = sinfo->chunks == NULL;
1356 chunk = (ChunkLink*) mem;
1357 chunk->next = sinfo->chunks;
1358 sinfo->chunks = chunk;
1359 sinfo->n_allocated--;
1360 /* keep slab ring partially sorted, empty slabs at end */
1361 if (was_empty)
1362 {
1363 /* unlink slab */
1364 SlabInfo *next = sinfo->next, *prev = sinfo->prev;
1365 next->prev = prev;
1366 prev->next = next;
1367 if (allocator->slab_stack[ix] == sinfo)
1368 allocator->slab_stack[ix] = next == sinfo ? NULL : next;
1369 /* insert slab at head */
1370 allocator_slab_stack_push (allocator, ix, sinfo);
1371 }
1372 /* eagerly free complete unused slabs */
1373 if (!sinfo->n_allocated)
1374 {
1375 /* unlink slab */
1376 SlabInfo *next = sinfo->next, *prev = sinfo->prev;
1377 next->prev = prev;
1378 prev->next = next;
1379 if (allocator->slab_stack[ix] == sinfo)
1380 allocator->slab_stack[ix] = next == sinfo ? NULL : next;
1381 /* free slab */
1382 allocator_memfree (page_size, page);
1383 }
1384 }
1385
1386 /* --- memalign implementation --- */
1387 #ifdef HAVE_MALLOC_H
1388 #include <malloc.h> /* memalign() */
1389 #endif
1390
1391 /* from config.h:
1392 * define HAVE_POSIX_MEMALIGN 1 // if free(posix_memalign(3)) works, <stdlib.h>
1393 * define HAVE_MEMALIGN 1 // if free(memalign(3)) works, <malloc.h>
1394 * define HAVE_VALLOC 1 // if free(valloc(3)) works, <stdlib.h> or <malloc.h>
1395 * if none is provided, we implement malloc(3)-based alloc-only page alignment
1396 */
1397
1398 #if !(HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN || HAVE_VALLOC)
1399 G_GNUC_BEGIN_IGNORE_DEPRECATIONS
1400 static GTrashStack *compat_valloc_trash = NULL;
1401 G_GNUC_END_IGNORE_DEPRECATIONS
1402 #endif
1403
1404 static gpointer
allocator_memalign(gsize alignment,gsize memsize)1405 allocator_memalign (gsize alignment,
1406 gsize memsize)
1407 {
1408 gpointer aligned_memory = NULL;
1409 gint err = ENOMEM;
1410 #if HAVE_POSIX_MEMALIGN
1411 err = posix_memalign (&aligned_memory, alignment, memsize);
1412 #elif HAVE_MEMALIGN
1413 errno = 0;
1414 aligned_memory = memalign (alignment, memsize);
1415 err = errno;
1416 #elif HAVE_VALLOC
1417 errno = 0;
1418 aligned_memory = valloc (memsize);
1419 err = errno;
1420 #else
1421 /* simplistic non-freeing page allocator */
1422 mem_assert (alignment == sys_page_size);
1423 mem_assert (memsize <= sys_page_size);
1424 if (!compat_valloc_trash)
1425 {
1426 const guint n_pages = 16;
1427 guint8 *mem = malloc (n_pages * sys_page_size);
1428 err = errno;
1429 if (mem)
1430 {
1431 gint i = n_pages;
1432 guint8 *amem = (guint8*) ALIGN ((gsize) mem, sys_page_size);
1433 if (amem != mem)
1434 i--; /* mem wasn't page aligned */
1435 G_GNUC_BEGIN_IGNORE_DEPRECATIONS
1436 while (--i >= 0)
1437 g_trash_stack_push (&compat_valloc_trash, amem + i * sys_page_size);
1438 G_GNUC_END_IGNORE_DEPRECATIONS
1439 }
1440 }
1441 G_GNUC_BEGIN_IGNORE_DEPRECATIONS
1442 aligned_memory = g_trash_stack_pop (&compat_valloc_trash);
1443 G_GNUC_END_IGNORE_DEPRECATIONS
1444 #endif
1445 if (!aligned_memory)
1446 errno = err;
1447 return aligned_memory;
1448 }
1449
1450 static void
allocator_memfree(gsize memsize,gpointer mem)1451 allocator_memfree (gsize memsize,
1452 gpointer mem)
1453 {
1454 #if HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN || HAVE_VALLOC
1455 free (mem);
1456 #else
1457 mem_assert (memsize <= sys_page_size);
1458 G_GNUC_BEGIN_IGNORE_DEPRECATIONS
1459 g_trash_stack_push (&compat_valloc_trash, mem);
1460 G_GNUC_END_IGNORE_DEPRECATIONS
1461 #endif
1462 }
1463
1464 static void
mem_error(const char * format,...)1465 mem_error (const char *format,
1466 ...)
1467 {
1468 const char *pname;
1469 va_list args;
1470 /* at least, put out "MEMORY-ERROR", in case we segfault during the rest of the function */
1471 fputs ("\n***MEMORY-ERROR***: ", stderr);
1472 pname = g_get_prgname();
1473 g_fprintf (stderr, "%s[%ld]: GSlice: ", pname ? pname : "", (long)getpid());
1474 va_start (args, format);
1475 g_vfprintf (stderr, format, args);
1476 va_end (args);
1477 fputs ("\n", stderr);
1478 abort();
1479 _exit (1);
1480 }
1481
1482 /* --- g-slice memory checker tree --- */
1483 typedef size_t SmcKType; /* key type */
1484 typedef size_t SmcVType; /* value type */
1485 typedef struct {
1486 SmcKType key;
1487 SmcVType value;
1488 } SmcEntry;
1489 static void smc_tree_insert (SmcKType key,
1490 SmcVType value);
1491 static gboolean smc_tree_lookup (SmcKType key,
1492 SmcVType *value_p);
1493 static gboolean smc_tree_remove (SmcKType key);
1494
1495
1496 /* --- g-slice memory checker implementation --- */
1497 static void
smc_notify_alloc(void * pointer,size_t size)1498 smc_notify_alloc (void *pointer,
1499 size_t size)
1500 {
1501 size_t address = (size_t) pointer;
1502 if (pointer)
1503 smc_tree_insert (address, size);
1504 }
1505
1506 #if 0
1507 static void
1508 smc_notify_ignore (void *pointer)
1509 {
1510 size_t address = (size_t) pointer;
1511 if (pointer)
1512 smc_tree_remove (address);
1513 }
1514 #endif
1515
1516 static int
smc_notify_free(void * pointer,size_t size)1517 smc_notify_free (void *pointer,
1518 size_t size)
1519 {
1520 size_t address = (size_t) pointer;
1521 SmcVType real_size;
1522 gboolean found_one;
1523
1524 if (!pointer)
1525 return 1; /* ignore */
1526 found_one = smc_tree_lookup (address, &real_size);
1527 if (!found_one)
1528 {
1529 g_fprintf (stderr, "GSlice: MemChecker: attempt to release non-allocated block: %p size=%" G_GSIZE_FORMAT "\n", pointer, size);
1530 return 0;
1531 }
1532 if (real_size != size && (real_size || size))
1533 {
1534 g_fprintf (stderr, "GSlice: MemChecker: attempt to release block with invalid size: %p size=%" G_GSIZE_FORMAT " invalid-size=%" G_GSIZE_FORMAT "\n", pointer, real_size, size);
1535 return 0;
1536 }
1537 if (!smc_tree_remove (address))
1538 {
1539 g_fprintf (stderr, "GSlice: MemChecker: attempt to release non-allocated block: %p size=%" G_GSIZE_FORMAT "\n", pointer, size);
1540 return 0;
1541 }
1542 return 1; /* all fine */
1543 }
1544
1545 /* --- g-slice memory checker tree implementation --- */
1546 #define SMC_TRUNK_COUNT (4093 /* 16381 */) /* prime, to distribute trunk collisions (big, allocated just once) */
1547 #define SMC_BRANCH_COUNT (511) /* prime, to distribute branch collisions */
1548 #define SMC_TRUNK_EXTENT (SMC_BRANCH_COUNT * 2039) /* key address space per trunk, should distribute uniformly across BRANCH_COUNT */
1549 #define SMC_TRUNK_HASH(k) ((k / SMC_TRUNK_EXTENT) % SMC_TRUNK_COUNT) /* generate new trunk hash per megabyte (roughly) */
1550 #define SMC_BRANCH_HASH(k) (k % SMC_BRANCH_COUNT)
1551
1552 typedef struct {
1553 SmcEntry *entries;
1554 unsigned int n_entries;
1555 } SmcBranch;
1556
1557 static SmcBranch **smc_tree_root = NULL;
1558
1559 static void
smc_tree_abort(int errval)1560 smc_tree_abort (int errval)
1561 {
1562 const char *syserr = strerror (errval);
1563 mem_error ("MemChecker: failure in debugging tree: %s", syserr);
1564 }
1565
1566 static inline SmcEntry*
smc_tree_branch_grow_L(SmcBranch * branch,unsigned int index)1567 smc_tree_branch_grow_L (SmcBranch *branch,
1568 unsigned int index)
1569 {
1570 unsigned int old_size = branch->n_entries * sizeof (branch->entries[0]);
1571 unsigned int new_size = old_size + sizeof (branch->entries[0]);
1572 SmcEntry *entry;
1573 mem_assert (index <= branch->n_entries);
1574 branch->entries = (SmcEntry*) realloc (branch->entries, new_size);
1575 if (!branch->entries)
1576 smc_tree_abort (errno);
1577 entry = branch->entries + index;
1578 memmove (entry + 1, entry, (branch->n_entries - index) * sizeof (entry[0]));
1579 branch->n_entries += 1;
1580 return entry;
1581 }
1582
1583 static inline SmcEntry*
smc_tree_branch_lookup_nearest_L(SmcBranch * branch,SmcKType key)1584 smc_tree_branch_lookup_nearest_L (SmcBranch *branch,
1585 SmcKType key)
1586 {
1587 unsigned int n_nodes = branch->n_entries, offs = 0;
1588 SmcEntry *check = branch->entries;
1589 int cmp = 0;
1590 while (offs < n_nodes)
1591 {
1592 unsigned int i = (offs + n_nodes) >> 1;
1593 check = branch->entries + i;
1594 cmp = key < check->key ? -1 : key != check->key;
1595 if (cmp == 0)
1596 return check; /* return exact match */
1597 else if (cmp < 0)
1598 n_nodes = i;
1599 else /* (cmp > 0) */
1600 offs = i + 1;
1601 }
1602 /* check points at last mismatch, cmp > 0 indicates greater key */
1603 return cmp > 0 ? check + 1 : check; /* return insertion position for inexact match */
1604 }
1605
1606 static void
smc_tree_insert(SmcKType key,SmcVType value)1607 smc_tree_insert (SmcKType key,
1608 SmcVType value)
1609 {
1610 unsigned int ix0, ix1;
1611 SmcEntry *entry;
1612
1613 g_mutex_lock (&smc_tree_mutex);
1614 ix0 = SMC_TRUNK_HASH (key);
1615 ix1 = SMC_BRANCH_HASH (key);
1616 if (!smc_tree_root)
1617 {
1618 smc_tree_root = calloc (SMC_TRUNK_COUNT, sizeof (smc_tree_root[0]));
1619 if (!smc_tree_root)
1620 smc_tree_abort (errno);
1621 }
1622 if (!smc_tree_root[ix0])
1623 {
1624 smc_tree_root[ix0] = calloc (SMC_BRANCH_COUNT, sizeof (smc_tree_root[0][0]));
1625 if (!smc_tree_root[ix0])
1626 smc_tree_abort (errno);
1627 }
1628 entry = smc_tree_branch_lookup_nearest_L (&smc_tree_root[ix0][ix1], key);
1629 if (!entry || /* need create */
1630 entry >= smc_tree_root[ix0][ix1].entries + smc_tree_root[ix0][ix1].n_entries || /* need append */
1631 entry->key != key) /* need insert */
1632 entry = smc_tree_branch_grow_L (&smc_tree_root[ix0][ix1], entry - smc_tree_root[ix0][ix1].entries);
1633 entry->key = key;
1634 entry->value = value;
1635 g_mutex_unlock (&smc_tree_mutex);
1636 }
1637
1638 static gboolean
smc_tree_lookup(SmcKType key,SmcVType * value_p)1639 smc_tree_lookup (SmcKType key,
1640 SmcVType *value_p)
1641 {
1642 SmcEntry *entry = NULL;
1643 unsigned int ix0 = SMC_TRUNK_HASH (key), ix1 = SMC_BRANCH_HASH (key);
1644 gboolean found_one = FALSE;
1645 *value_p = 0;
1646 g_mutex_lock (&smc_tree_mutex);
1647 if (smc_tree_root && smc_tree_root[ix0])
1648 {
1649 entry = smc_tree_branch_lookup_nearest_L (&smc_tree_root[ix0][ix1], key);
1650 if (entry &&
1651 entry < smc_tree_root[ix0][ix1].entries + smc_tree_root[ix0][ix1].n_entries &&
1652 entry->key == key)
1653 {
1654 found_one = TRUE;
1655 *value_p = entry->value;
1656 }
1657 }
1658 g_mutex_unlock (&smc_tree_mutex);
1659 return found_one;
1660 }
1661
1662 static gboolean
smc_tree_remove(SmcKType key)1663 smc_tree_remove (SmcKType key)
1664 {
1665 unsigned int ix0 = SMC_TRUNK_HASH (key), ix1 = SMC_BRANCH_HASH (key);
1666 gboolean found_one = FALSE;
1667 g_mutex_lock (&smc_tree_mutex);
1668 if (smc_tree_root && smc_tree_root[ix0])
1669 {
1670 SmcEntry *entry = smc_tree_branch_lookup_nearest_L (&smc_tree_root[ix0][ix1], key);
1671 if (entry &&
1672 entry < smc_tree_root[ix0][ix1].entries + smc_tree_root[ix0][ix1].n_entries &&
1673 entry->key == key)
1674 {
1675 unsigned int i = entry - smc_tree_root[ix0][ix1].entries;
1676 smc_tree_root[ix0][ix1].n_entries -= 1;
1677 memmove (entry, entry + 1, (smc_tree_root[ix0][ix1].n_entries - i) * sizeof (entry[0]));
1678 if (!smc_tree_root[ix0][ix1].n_entries)
1679 {
1680 /* avoid useless pressure on the memory system */
1681 free (smc_tree_root[ix0][ix1].entries);
1682 smc_tree_root[ix0][ix1].entries = NULL;
1683 }
1684 found_one = TRUE;
1685 }
1686 }
1687 g_mutex_unlock (&smc_tree_mutex);
1688 return found_one;
1689 }
1690
1691 #ifdef G_ENABLE_DEBUG
1692 void
g_slice_debug_tree_statistics(void)1693 g_slice_debug_tree_statistics (void)
1694 {
1695 g_mutex_lock (&smc_tree_mutex);
1696 if (smc_tree_root)
1697 {
1698 unsigned int i, j, t = 0, o = 0, b = 0, su = 0, ex = 0, en = 4294967295u;
1699 double tf, bf;
1700 for (i = 0; i < SMC_TRUNK_COUNT; i++)
1701 if (smc_tree_root[i])
1702 {
1703 t++;
1704 for (j = 0; j < SMC_BRANCH_COUNT; j++)
1705 if (smc_tree_root[i][j].n_entries)
1706 {
1707 b++;
1708 su += smc_tree_root[i][j].n_entries;
1709 en = MIN (en, smc_tree_root[i][j].n_entries);
1710 ex = MAX (ex, smc_tree_root[i][j].n_entries);
1711 }
1712 else if (smc_tree_root[i][j].entries)
1713 o++; /* formerly used, now empty */
1714 }
1715 en = b ? en : 0;
1716 tf = MAX (t, 1.0); /* max(1) to be a valid divisor */
1717 bf = MAX (b, 1.0); /* max(1) to be a valid divisor */
1718 g_fprintf (stderr, "GSlice: MemChecker: %u trunks, %u branches, %u old branches\n", t, b, o);
1719 g_fprintf (stderr, "GSlice: MemChecker: %f branches per trunk, %.2f%% utilization\n",
1720 b / tf,
1721 100.0 - (SMC_BRANCH_COUNT - b / tf) / (0.01 * SMC_BRANCH_COUNT));
1722 g_fprintf (stderr, "GSlice: MemChecker: %f entries per branch, %u minimum, %u maximum\n",
1723 su / bf, en, ex);
1724 }
1725 else
1726 g_fprintf (stderr, "GSlice: MemChecker: root=NULL\n");
1727 g_mutex_unlock (&smc_tree_mutex);
1728
1729 /* sample statistics (beast + GSLice + 24h scripted core & GUI activity):
1730 * PID %CPU %MEM VSZ RSS COMMAND
1731 * 8887 30.3 45.8 456068 414856 beast-0.7.1 empty.bse
1732 * $ cat /proc/8887/statm # total-program-size resident-set-size shared-pages text/code data/stack library dirty-pages
1733 * 114017 103714 2354 344 0 108676 0
1734 * $ cat /proc/8887/status
1735 * Name: beast-0.7.1
1736 * VmSize: 456068 kB
1737 * VmLck: 0 kB
1738 * VmRSS: 414856 kB
1739 * VmData: 434620 kB
1740 * VmStk: 84 kB
1741 * VmExe: 1376 kB
1742 * VmLib: 13036 kB
1743 * VmPTE: 456 kB
1744 * Threads: 3
1745 * (gdb) print g_slice_debug_tree_statistics ()
1746 * GSlice: MemChecker: 422 trunks, 213068 branches, 0 old branches
1747 * GSlice: MemChecker: 504.900474 branches per trunk, 98.81% utilization
1748 * GSlice: MemChecker: 4.965039 entries per branch, 1 minimum, 37 maximum
1749 */
1750 }
1751 #endif /* G_ENABLE_DEBUG */
1752