2 * Written by Josh Dybnis and released to the public domain, as explained at
3 * http://creativecommons.org/licenses/publicdomain
5 * Extreamly fast multi-threaded malloc.
7 #ifndef USE_SYSTEM_MALLOC
8 #define _BSD_SOURCE // so we get MAP_ANON on linux
18 #define MAX_SCALE 36 // allocate blocks up to 64GB (arbitrary, could be bigger)
19 #define MIN_SCALE 3 // smallest allocated block is 8 bytes
20 #define MAX_POINTER_BITS 48
21 #define PAGE_SCALE 21 // 2MB pages
24 #define MIN_SCALE 2 // smallest allocated block is 4 bytes
25 #define MAX_POINTER_BITS 32
26 #define PAGE_SCALE 12 // 4KB pages
28 #define PAGE_SIZE (1ULL << PAGE_SCALE)
29 #define HEADERS_SIZE (((size_t)1ULL << (MAX_POINTER_BITS - PAGE_SCALE)) * sizeof(header_t))
31 typedef struct block {
35 // TODO: Break the page header into two parts. The first part is located in the header region. The
36 // second part is located on the page and is only used when there are free items.
37 typedef struct header {
41 block_t *free_list; // list of free blocks
44 uint8_t owner; // thread id of owner
45 uint8_t scale; // log2 of the block size
49 typedef struct size_class {
50 header_t *active_page;
51 header_t *oldest_partial;
52 header_t *newest_partial;
58 block_t *free_list[MAX_SCALE+1];
61 size_class_t size_class[MAX_SCALE+1];
63 block_t *blocks_from[MAX_NUM_THREADS];
64 block_t *blocks_to[MAX_NUM_THREADS];
65 } __attribute__((aligned(CACHE_LINE_SIZE))) tl_t;
67 static header_t *headers_ = NULL;
69 static tl_t tl_[MAX_NUM_THREADS] = {};
71 static inline header_t *get_header (void *r) {
72 ASSERT(((size_t)r >> PAGE_SCALE) < HEADERS_SIZE);
73 return headers_ + ((size_t)r >> PAGE_SCALE);
76 static void *get_new_region (int block_scale) {
77 int thread_index = GET_THREAD_INDEX();
79 tl_t *tl = &tl_[thread_index]; // thread-local data
80 if (block_scale <= PAGE_SCALE && tl->free_pages != NULL) {
81 void *region = tl->free_pages;
82 tl->free_pages = tl->free_pages->next;
83 get_header(region)->scale = block_scale;
87 size_t region_size = (1ULL << block_scale);
88 if (region_size < PAGE_SIZE) {
89 region_size = PAGE_SIZE;
91 void *region = mmap(NULL, region_size, PROT_READ|PROT_WRITE, MAP_NORESERVE|MAP_ANON|MAP_PRIVATE, -1, 0);
92 TRACE("m1", "get_new_region: mmapped new region %p (size %p)", region, region_size);
93 if (region == (void *)-1) {
94 perror("get_new_region: mmap");
97 if ((size_t)region & (region_size - 1)) {
98 TRACE("m0", "get_new_region: region not aligned", 0, 0);
99 munmap(region, region_size);
100 region = mmap(NULL, region_size * 2, PROT_READ|PROT_WRITE, MAP_NORESERVE|MAP_ANON|MAP_PRIVATE, -1, 0);
101 if (region == (void *)-1) {
102 perror("get_new_region: mmap");
105 TRACE("m0", "get_new_region: mmapped new region %p (size %p)", region, region_size * 2);
106 void *aligned = (void *)(((size_t)region + region_size) & ~(region_size - 1));
107 size_t extra = (char *)aligned - (char *)region;
109 munmap(region, extra);
110 TRACE("m0", "get_new_region: unmapped extra memory %p (size %p)", region, extra);
112 extra = ((char *)region + region_size) - (char *)aligned;
114 munmap((char *)aligned + region_size, extra);
115 TRACE("m0", "get_new_region: unmapped extra memory %p (size %p)", (char *)aligned + region_size, extra);
121 header_t *h = get_header(region);
122 TRACE("m1", "get_new_region: header %p (%p)", h, h - headers_);
123 assert(h->scale == 0);
124 h->scale = block_scale;
125 h->owner = thread_index;
130 void mem_init (void) {
131 assert(headers_ == NULL);
132 // Allocate space for the page headers. This could be a big chunk of memory on 64 bit systems,
133 // but it just takes up virtual address space. Physical space used by the headers is still
134 // proportional to the amount of memory the user mallocs.
135 headers_ = mmap(NULL, HEADERS_SIZE, PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, -1, 0);
136 TRACE("m1", "mem_init: header page %p", headers_, 0);
138 // initialize spsc queues
139 for (int i = 0; i < MAX_NUM_THREADS; ++i) {
140 for (int j = 0; j < MAX_NUM_THREADS; ++j) {
142 tl_[i].blocks_to[j] = (block_t *)&(tl_[j].blocks_from[i]);
148 void nbd_free (void *x) {
149 TRACE("m1", "nbd_free: block %p page %p", x, (size_t)x & ~MASK(PAGE_SCALE));
151 block_t *b = (block_t *)x;
152 header_t *h = get_header(x);
153 int b_scale = h->scale;
154 TRACE("m1", "nbd_free: header %p scale %llu", h, b_scale);
155 ASSERT(b_scale && b_scale <= MAX_SCALE);
157 if (b_scale > PAGE_SCALE) {
158 int rc = munmap(x, 1ULL << b_scale);
164 memset(b, 0xcd, (1ULL << b_scale)); // bear trap
166 int thread_index = GET_THREAD_INDEX();
167 tl_t *tl = &tl_[thread_index]; // thread-local data
168 if (h->owner == thread_index) {
169 TRACE("m1", "nbd_free: private block, old free list head %p", tl->free_list[b_scale], 0);
171 #ifndef RECYCLE_PAGES
172 b->next = tl->free_list[b_scale];
173 tl->free_list[b_scale] = b;
174 #else //RECYCLE_PAGES
175 b->next = h->free_list;
178 size_class_t *sc = &tl->size_class[b_scale];
179 if (sc->active_page != h) {
180 if (h->num_in_use == 0) {
181 // remove <h> from the partial-page list
182 if (h->next != NULL) { h->next->prev = h->prev; }
183 if (h->prev != NULL) { h->prev->next = h->next; }
184 // put <h> on the free-page list
185 h->next = tl->free_pages;
188 // move <h> to the top of the partial-page list
189 if (h->next != NULL) {
190 h->next->prev = h->prev;
191 if (h->prev != NULL) { h->prev->next = h->next; }
192 h->prev = sc->newest_partial;
194 sc->newest_partial = h;
198 #endif//RECYCLE_PAGES
200 // push <b> onto it's owner's queue
201 int b_owner = h->owner;
202 TRACE("m1", "nbd_free: owner %llu", b_owner, 0);
204 // The assignment statements are volatile to prevent the compiler from reordering them.
205 VOLATILE_DEREF(b).next = NULL;
206 VOLATILE_DEREF(tl->blocks_to[b_owner]).next = b;
208 tl->blocks_to[b_owner] = b;
212 static inline void process_incoming_blocks (tl_t *tl) {
213 for (int p = 0; p < MAX_NUM_THREADS; ++p) {
214 block_t *b = tl->blocks_from[p];
215 if (EXPECT_FALSE(b == NULL)) continue; // the queue is completely empty
217 // Leave the last block on the queue. Removing the last block on the queue would create a
218 // race with the producer thread putting a new block on the queue.
219 for (block_t *next = b->next; next != NULL; b = next, next = b->next) {
220 // push <b> onto the appropriate free list
221 #ifndef RECYCLE_PAGES
222 int b_scale = get_header(b)->scale;
223 b->next = tl->free_list[b_scale];
224 tl->free_list[b_scale] = b;
225 #else //RECYCLE_PAGES
226 header_t *h = get_header(b);
227 b->next = h->free_list;
229 #endif//RECYCLE_PAGES
231 tl->blocks_from[p] = b;
235 static inline block_t *pop_free_list (tl_t *tl, int scale) {
236 #ifndef RECYCLE_PAGES
237 block_t **free_list = &tl->free_list[scale];
238 #else //RECYCLE_PAGES
239 size_class_t *sc = &tl->size_class[scale];
240 if (EXPECT_FALSE(sc->active_page == NULL))
242 block_t **free_list = &sc->active_page->free_list;
243 #endif//RECYCLE_PAGES
244 block_t *b = *free_list;
245 if (EXPECT_FALSE(b == NULL))
247 ASSERT(get_header(b)->scale == scale);
248 *free_list = b->next;
252 // Allocate a block of memory at least size <n>. Blocks are binned in powers-of-two. Round up <n> to
253 // the nearest power of two.
255 // First check the current thread's free list for an available block. If there are no blocks on the
256 // free list, pull items off of the current thread's incoming block queues and push them onto the
257 // free list. If we didn't get an appropriate size block off of the block queues then allocate a new
258 // page, break it up into blocks and push them onto the free list.
259 void *nbd_malloc (size_t n) {
260 // the scale is the log base 2 of <n>, rounded up
261 int b_scale = (sizeof(void *) * __CHAR_BIT__) - __builtin_clzl((n) - 1);
262 TRACE("m1", "nbd_malloc: size %llu (scale %llu)", n, b_scale);
264 if (EXPECT_FALSE(b_scale < MIN_SCALE)) { b_scale = MIN_SCALE; }
265 if (EXPECT_FALSE(b_scale > MAX_SCALE)) { return NULL; }
267 tl_t *tl = &tl_[GET_THREAD_INDEX()]; // thread-local data
269 block_t *b = pop_free_list(tl, b_scale);
271 TRACE("m1", "nbd_malloc: returning block %p", b, 0);
276 // The free list is empty so process blocks freed from other threads and then check again.
277 process_incoming_blocks(tl);
278 b = pop_free_list(tl, b_scale);
280 TRACE("m1", "nbd_malloc: returning block %p", b, 0);
286 // The current active page is completely allocated. Make the oldest partially allocated page
287 // the new active page.
288 size_class_t *sc = &tl->size_class[b_scale];
289 if (sc->oldest_partial != NULL) {
290 sc->active_page = sc->oldest_partial;
291 sc->oldest_partial = sc->oldest_partial->next;
292 sc->oldest_partial->prev = NULL;
293 b = pop_free_list(tl, b_scale);
295 TRACE("m1", "nbd_malloc: returning block %p", b, 0);
299 // There are no partially allocated pages so get a new page.
301 #endif//RECYCLE_PAGES
304 char *page = get_new_region(b_scale);
305 b = (block_t *)page; // grab the first block on the page
307 // Break up the remainder of the page into blocks and put them on the free list. Start at the
308 // end of the page so that the free list ends up in increasing order, for ease of debugging.
309 if (b_scale < PAGE_SCALE) {
310 size_t block_size = (1ULL << b_scale);
311 block_t *head = NULL;
312 for (int offset = PAGE_SIZE - block_size; offset > 0; offset -= block_size) {
313 block_t *x = (block_t *)(page + offset);
314 x->next = head; head = x;
316 #ifndef RECYCLE_PAGES
317 tl->free_list[b_scale] = head;
318 #else //RECYCLE_PAGES
319 sc->active_page = get_header(page);
320 sc->active_page->free_list = head;
321 #endif//RECYCLE_PAGES
324 TRACE("m1", "nbd_malloc: returning block %p from new region %p", b, (size_t)b & ~MASK(PAGE_SCALE));
328 #else//USE_SYSTEM_MALLOC
334 void mem_init (void) {
338 void nbd_free (void *x) {
339 TRACE("m1", "nbd_free: %p", x, 0);
341 memset(x, 0xcd, sizeof(void *)); // bear trap
347 void *nbd_malloc (size_t n) {
348 TRACE("m1", "nbd_malloc: request size %llu", n, 0);
350 TRACE("m1", "nbd_malloc: returning %p", x, 0);
353 #endif//USE_SYSTEM_MALLOC