src/alloc-slab.c

   1 /* alloc-slab.c - Slab debugging allocator
   2  * Copyright 2005 srvx Development Team
   3  *
   4  * This file is part of srvx.
   5  *
   6  * x3 is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  */
  16
  17 #include "common.h"
  18 #include "log.h"
  19
  20 #if defined(HAVE_SYS_MMAN_H)
  21 # include <sys/mman.h>
  22 #endif
  23
  24 #if !defined(HAVE_MMAP)
  25 # error The slab allocator requires that your system have the mmap() system call.
  26 #endif
  27
  28 #define SLAB_DEBUG 1
  29 #define SLAB_RESERVE 1024
  30
  31 #if SLAB_DEBUG
  32
  33 #define ALLOC_MAGIC 0x1a
  34 #define FREE_MAGIC  0xcf
  35
  36 struct alloc_header {
  37     unsigned int size : 24;
  38     unsigned int magic : 8;
  39     unsigned int file_id : 8;
  40     unsigned int line : 16;
  41 };
  42
  43 static const char *file_ids[256];
  44 static struct file_id_entry {
  45     const char *name;
  46     unsigned int id : 8;
  47 } file_id_map[256];
  48 unsigned int file_ids_used;
  49
  50 static int
  51 file_id_cmp(const void *a_, const void *b_)
  52 {
  53     return strcmp(*(const char**)a_, *(const char**)b_);
  54 }
  55
  56 static unsigned int
  57 get_file_id(const char *fname)
  58 {
  59     struct file_id_entry *entry;
  60
  61     entry = bsearch(&fname, file_id_map, file_ids_used, sizeof(file_id_map[0]), file_id_cmp);
  62     if (entry)
  63         return entry->id;
  64     entry = file_id_map + file_ids_used;
  65     file_ids[file_ids_used] = fname;
  66     entry->name = fname;
  67     entry->id = file_ids_used;
  68     qsort(file_id_map, ++file_ids_used, sizeof(file_id_map[0]), file_id_cmp);
  69     return file_ids_used - 1;
  70 }
  71
  72 typedef struct alloc_header alloc_header_t;
  73
  74 #else
  75
  76 typedef size_t alloc_header_t;
  77
  78 #endif
  79
  80 struct slab {
  81     struct slabset *parent;
  82     struct slab *prev;
  83     struct slab *next;
  84     void *base;
  85     void **free;
  86     unsigned int used;
  87 };
  88
  89 struct slabset {
  90     struct slab *child;
  91     size_t nslabs;
  92     size_t nallocs;
  93     size_t size;
  94     size_t items_per_slab;
  95 };
  96
  97 #define SLAB_MIN     (2 * sizeof(void*))
  98 #define SLAB_GRAIN   sizeof(void*)
  99 #define SLAB_ALIGN   SLAB_GRAIN
 100 #define SMALL_CUTOFF 576
 101 /* Element size < SMALL_CUTOFF -> use small slabs.
 102  * Larger elements are allocated directly using mmap().  The largest
 103  * regularly allocated struct in srvx 1.x is smaller than
 104  * SMALL_CUTOFF, so there is not much point in coding support for
 105  * larger slabs.
 106  */
 107
 108 static struct slabset little_slabs[SMALL_CUTOFF / SLAB_GRAIN];
 109 static struct slab *free_slab_head;
 110 static struct slab *free_slab_tail;
 111 unsigned long free_slab_count;
 112 unsigned long big_alloc_count;
 113 unsigned long big_alloc_size;
 114 unsigned long slab_count;
 115 unsigned long slab_alloc_count;
 116 unsigned long slab_alloc_size;
 117
 118 #if defined(MAP_ANON)
 119 #elif defined(MAP_ANONYMOUS)
 120 # define MAP_ANON MAP_ANONYMOUS
 121 #else
 122 # define MAP_ANON 0
 123 #endif
 124
 125 static size_t
 126 slab_pagesize(void)
 127 {
 128     static size_t pagesize;
 129     if (pagesize
 130 #if defined(HAVE_GETPAGESIZE)
 131         || (pagesize  = getpagesize())
 132 #endif
 133 #if defined(HAVE_SYSCONF) && defined(_SC_PAGESIZE)
 134         || (pagesize = sysconf(_SC_PAGESIZE))
 135 #endif
 136 #if defined(HAVE_SYSCONF) && defined(_SC_PAGE_SIZE)
 137         || (pagesize = sysconf(_SC_PAGE_SIZE))
 138 #endif
 139         ) return pagesize;
 140     assert(0 && "unable to find system page size");
 141     return pagesize = 4096;
 142 }
 143
 144 static size_t
 145 slab_round_up(size_t size)
 146 {
 147     return (size + slab_pagesize() - 1) & ~(slab_pagesize() - 1);
 148 }
 149
 150 static void *
 151 slab_map(size_t length)
 152 {
 153     static int mmap_fd = -1;
 154     void *res;
 155
 156 #if ! MAP_ANON
 157     if (mmap_fd < 0) {
 158         mmap_fd = open("/dev/zero", 0);
 159         if (mmap_fd < 0)
 160             log_module(MAIN_LOG, LOG_FATAL, "Unable to open /dev/zero for mmap: %s", strerror(errno()));
 161     }
 162 #endif
 163     res = mmap(0, length, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, mmap_fd, 0);
 164     if (res == MAP_FAILED)
 165         log_module(MAIN_LOG, LOG_FATAL, "Unable to mmap %lu bytes (%s).", (unsigned long)length, strerror(errno));
 166     return res;
 167 }
 168
 169 static void *slab_alloc(struct slabset *sset);
 170 static void slab_unalloc(void *ptr, size_t size);
 171
 172 static struct slabset *
 173 slabset_create(size_t size)
 174 {
 175     unsigned int idx;
 176
 177     size = (size < SLAB_MIN) ? SLAB_MIN : (size + SLAB_GRAIN - 1) & ~(SLAB_GRAIN - 1);
 178     idx = size / SLAB_GRAIN;
 179     assert(idx < ArrayLength(little_slabs));
 180     if (!little_slabs[idx].size) {
 181         little_slabs[idx].size = size;
 182         little_slabs[idx].items_per_slab = (slab_pagesize() - sizeof(struct slab)) / ((size + SLAB_ALIGN - 1) & ~(SLAB_ALIGN - 1));
 183     }
 184     return &little_slabs[idx];
 185 }
 186
 187 static void *
 188 slab_alloc(struct slabset *sset)
 189 {
 190     struct slab *slab;
 191     void **item;
 192
 193     if (!sset->child || !sset->child->free) {
 194         unsigned int ii, step;
 195
 196         /* Allocate new slab. */
 197         if (free_slab_head) {
 198             slab = free_slab_head;
 199             if (!(free_slab_head = slab->next))
 200                 free_slab_tail = NULL;
 201         } else {
 202             item = slab_map(slab_pagesize());
 203             slab = (struct slab*)((char*)item + slab_pagesize() - sizeof(*slab));
 204             slab->base = item;
 205             slab_count++;
 206         }
 207
 208         /* Populate free list. */
 209         step = (sset->size + SLAB_ALIGN - 1) & ~(SLAB_ALIGN - 1);
 210         for (ii = 1, item = slab->free = slab->base;
 211              ii < sset->items_per_slab;
 212              ++ii, item = (*item = (char*)item + step));
 213         *item = NULL;
 214
 215         /* Link to parent slabset. */
 216         slab->parent = sset;
 217         slab->prev = sset->child;
 218         if (slab->prev) {
 219             slab->next = slab->prev->next;
 220             slab->prev->next = slab;
 221             if (slab->next)
 222                 slab->next->prev = slab;
 223         } else
 224             slab->next = NULL;
 225         assert(!slab->next || slab == slab->next->prev);
 226         assert(!slab->prev || slab == slab->prev->next);
 227         sset->child = slab;
 228         sset->nslabs++;
 229     }
 230
 231     slab = sset->child;
 232     item = slab->free;
 233     assert(((unsigned long)item & (slab_pagesize() - 1))
 234            <= (slab_pagesize() - sizeof(*slab) - sset->size));
 235     slab->free = *item;
 236     if (++slab->used == sset->items_per_slab) {
 237         if (sset->child != slab) {
 238             /* Unlink slab and reinsert before sset->child. */
 239             if (slab->prev)
 240                 slab->prev->next = slab->next;
 241             if (slab->next)
 242                 slab->next->prev = slab->prev;
 243             if ((slab->prev = sset->child->prev))
 244                 slab->prev->next = slab;
 245             if ((slab->next = sset->child))
 246                 slab->next->prev = slab;
 247             assert(!slab->next || slab == slab->next->prev);
 248             assert(!slab->prev || slab == slab->prev->next);
 249         } else if (slab->next) {
 250             /* Advance sset->child to next pointer. */
 251             sset->child = slab->next;
 252         }
 253     }
 254     sset->nallocs++;
 255     memset(item, 0, sset->size);
 256     return item;
 257 }
 258
 259 static void
 260 slab_unalloc(void *ptr, size_t size)
 261 {
 262     struct slab *slab, *new_next;
 263
 264     assert(size < SMALL_CUTOFF);
 265     slab = (struct slab*)((((unsigned long)ptr | (slab_pagesize() - 1)) + 1) - sizeof(*slab));
 266     *(void**)ptr = slab->free;
 267     slab->free = ptr;
 268     slab->parent->nallocs--;
 269
 270     if (slab->used-- == slab->parent->items_per_slab
 271         && slab->parent->child != slab) {
 272         /* Unlink from current position, relink as parent's first child. */
 273         new_next = slab->parent->child;
 274         assert(new_next != NULL);
 275         if (slab->prev)
 276             slab->prev->next = slab->next;
 277         if (slab->next)
 278             slab->next->prev = slab->prev;
 279         if ((slab->prev = new_next->prev))
 280             slab->prev->next = slab;
 281         slab->next = new_next;
 282         new_next->prev = slab;
 283         slab->parent->child = slab;
 284         assert(!slab->next || slab == slab->next->prev);
 285         assert(!slab->prev || slab == slab->prev->next);
 286     } else if (!slab->used) {
 287         /* Unlink slab from its parent. */
 288         slab->parent->nslabs--;
 289         if (slab->prev)
 290             slab->prev->next = slab->next;
 291         if (slab->next)
 292             slab->next->prev = slab->prev;
 293         new_next = slab->next ? slab->next : slab->prev;
 294         if (slab == slab->parent->child)
 295             slab->parent->child = new_next;
 296         if (new_next) {
 297             assert(!new_next->next || new_next == new_next->next->prev);
 298             assert(!new_next->prev || new_next == new_next->prev->next);
 299         }
 300
 301 #if SLAB_RESERVE
 302         if (!free_slab_count) {
 303             /* Make sure we have enough free slab pages. */
 304             while (free_slab_count < SLAB_RESERVE) {
 305                 struct slab *tslab;
 306                 void *item;
 307
 308                 item = slab_map(slab_pagesize());
 309                 tslab = (struct slab*)((char*)item + slab_pagesize() - sizeof(*slab));
 310                 tslab->base = item;
 311                 tslab->prev = free_slab_tail;
 312                 free_slab_tail = tslab;
 313                 if (!free_slab_head)
 314                     free_slab_head = tslab;
 315                 free_slab_count++;
 316                 slab_count++;
 317             }
 318         }
 319
 320         /* Unmap old slab, so accesses to stale pointers will fault. */
 321         munmap(slab->base, slab_pagesize());
 322         slab_count--;
 323 #else
 324         /* Link to list of free slabs. */
 325         slab->parent = NULL;
 326         slab->prev = free_slab_tail;
 327         slab->next = NULL;
 328         free_slab_tail = slab;
 329         if (!free_slab_head)
 330             free_slab_head = slab;
 331         free_slab_count++;
 332 #endif
 333     }
 334     (void)size;
 335 }
 336
 337 void *
 338 slab_malloc(const char *file, unsigned int line, size_t size)
 339 {
 340     alloc_header_t *res;
 341     size_t real;
 342
 343     assert(size < 1 << 24);
 344     real = (size + sizeof(*res) + SLAB_GRAIN - 1) & ~(SLAB_GRAIN - 1);
 345     if (real < SMALL_CUTOFF) {
 346         res = slab_alloc(slabset_create(real));
 347         slab_alloc_count++;
 348         slab_alloc_size += size;
 349     } else {
 350         res = slab_map(slab_round_up(real));
 351         big_alloc_count++;
 352         big_alloc_size += size;
 353     }
 354 #if SLAB_DEBUG
 355     res->file_id = get_file_id(file);
 356     res->size = size;
 357     res->line = line;
 358     res->magic = ALLOC_MAGIC;
 359 #else
 360     *res = size;
 361     (void)file; (void)line;
 362 #endif
 363     return res + 1;
 364 }
 365
 366 void *
 367 slab_realloc(const char *file, unsigned int line, void *ptr, size_t size)
 368 {
 369     alloc_header_t *orig;
 370     void *newblock;
 371     size_t osize;
 372
 373     if (!ptr)
 374         return slab_malloc(file, line, size);
 375
 376     verify(ptr);
 377     orig = (alloc_header_t*)ptr - 1;
 378 #if SLAB_DEBUG
 379     osize = orig->size;
 380 #else
 381     osize = *orig;
 382 #endif
 383     if (osize >= size)
 384         return ptr;
 385     newblock = slab_malloc(file, line, size);
 386     memcpy(newblock, ptr, osize);
 387     slab_free(file, line, ptr);
 388     return newblock;
 389 }
 390
 391 char *
 392 slab_strdup(const char *file, unsigned int line, const char *src)
 393 {
 394     char *target;
 395     size_t len;
 396
 397     len = strlen(src) + 1;
 398     target = slab_malloc(file, line, len);
 399     memcpy(target, src, len);
 400     return target;
 401 }
 402
 403 void
 404 slab_free(const char *file, unsigned int line, void *ptr)
 405 {
 406     alloc_header_t *hdr;
 407     size_t user, real;
 408
 409     if (!ptr)
 410         return;
 411     verify(ptr);
 412     hdr = (alloc_header_t*)ptr - 1;
 413 #if SLAB_DEBUG
 414     hdr->file_id = get_file_id(file);
 415     hdr->line = line;
 416     hdr->magic = FREE_MAGIC;
 417     user = hdr->size;
 418 #else
 419     user = *hdr;
 420     (void)file; (void)line;
 421 #endif
 422     real = (user + sizeof(*hdr) + SLAB_GRAIN - 1) & ~(SLAB_GRAIN - 1);
 423     if (real < SMALL_CUTOFF) {
 424         memset(hdr + 1, 0xde, real - sizeof(*hdr));
 425         slab_unalloc(hdr, real);
 426         slab_alloc_count--;
 427         slab_alloc_size -= user;
 428     } else {
 429         munmap(hdr, slab_round_up(real));
 430         big_alloc_count--;
 431         big_alloc_size -= user;
 432     }
 433 }
 434
 435 /* Undefine the verify macro in case we're not debugging. */
 436 #undef verify
 437 void
 438 verify(const void *ptr)
 439 {
 440     alloc_header_t *hdr;
 441     size_t real;
 442
 443     if (!ptr)
 444         return;
 445
 446     hdr = (alloc_header_t*)ptr - 1;
 447 #if SLAB_DEBUG
 448     real = hdr->size + sizeof(*hdr);
 449     assert(hdr->file_id < file_ids_used);
 450     assert(hdr->magic == ALLOC_MAGIC);
 451 #else
 452     real = *hdr + sizeof(*hdr);
 453 #endif
 454     real = (real + SLAB_GRAIN - 1) & ~(SLAB_GRAIN - 1);
 455
 456     if (real >= SMALL_CUTOFF)
 457         assert(((unsigned long)ptr & (slab_pagesize() - 1)) == sizeof(*hdr));
 458     else {
 459         struct slab *slab;
 460         size_t expected;
 461
 462         expected = (real + SLAB_GRAIN - 1) & ~(SLAB_GRAIN - 1);
 463         slab = (struct slab*)((((unsigned long)ptr | (slab_pagesize() - 1)) + 1) - sizeof(*slab));
 464         assert(slab->parent->size == expected);
 465     }
 466 }