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. 64 bit platforms only!
14 #define GET_SCALE(n) (sizeof(n)*8-__builtin_clzl((n)-1)) // log2 of <n>, rounded up
15 #define MAX_SCALE 31 // allocate blocks up to 4GB in size (arbitrary, could be bigger)
16 #define REGION_SCALE 22 // 4MB regions
17 #define REGION_SIZE (1 << REGION_SCALE)
18 #define HEADER_REGION_SCALE 22 // 4MB is space enough for headers for over 2,000,000 regions
20 typedef struct block {
25 typedef struct header {
26 char owner; // thread id of owner
27 char scale; // log2 of the block size
30 static header_t *region_header_ = NULL;
32 // TODO: experiment with different memory layouts (i.e. separate private and public lists)
33 static block_t free_list_[MAX_NUM_THREADS][MAX_SCALE+1][MAX_NUM_THREADS];
35 static void *get_new_region (int scale) {
36 if (scale < REGION_SCALE) {
39 TRACE("m0", "get_new_region(): mmap new region scale: %llu", scale, 0);
40 void *region = mmap(NULL, (1 << scale), PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, -1, 0);
41 if (region == (void *)-1) {
42 perror("get_new_region: mmap");
49 void mem_init (void) {
50 assert(region_header_ == NULL);
51 region_header_ = (header_t *)get_new_region(HEADER_REGION_SCALE);
52 memset(region_header_, 0, REGION_SIZE);
55 // Put <x> onto its owner's public free list (in the appropriate size bin).
57 // TODO: maybe we want to munmap() larger size blocks to reclaim virtual address space?
58 void nbd_free (void *x) {
59 LOCALIZE_THREAD_LOCAL(tid_, int);
60 block_t *b = (block_t *)x;
61 assert(((size_t)b >> REGION_SCALE) < ((1 << HEADER_REGION_SCALE) / sizeof(header_t)));
62 header_t *h = region_header_ + ((size_t)b >> REGION_SCALE);
63 TRACE("m0", "nbd_free(): block %p scale %llu", x, h->scale);
64 block_t *l = &free_list_[(int)h->owner][(int)h->scale][tid_];
65 TRACE("m0", "nbd_free(): free list %p first block %p", l, l->next);
70 // Allocate a block of memory at least size <n>. Blocks are binned in powers-of-two. Round up
71 // <n> to the nearest power-of-two.
73 // First check the current thread's private free list for an available block. If no blocks are on
74 // the private free list, pull all the available blocks off of the current thread's public free
75 // lists and put them on the private free list. If we didn't find any blocks on the public free
76 // lists, open a new region, break it up into blocks and put them on the private free list.
77 void *nbd_malloc (size_t n) {
78 LOCALIZE_THREAD_LOCAL(tid_, int);
79 if (n < sizeof(block_t)) {
82 int b_scale = GET_SCALE(n);
83 assert(b_scale <= MAX_SCALE);
84 TRACE("m0", "nbd_malloc(): size %llu scale %llu", n, b_scale);
85 block_t *fls = free_list_[tid_][b_scale]; // our free lists
86 block_t *pri = fls + tid_; // our private free list
87 TRACE("m0", "nbd_malloc(): private free list %p first block %p", pri, pri->next);
89 // If our private free list is empty, fill it up with blocks from our public free lists
90 if (EXPECT_FALSE(pri->next == NULL)) {
93 for (int i = 0; i < MAX_NUM_THREADS; ++i) {
94 TRACE("m0", "nbd_malloc(): searching public free lists (%llu)", i, 0);
95 block_t *pub = fls + i; // one of our public free lists
96 TRACE("m0", "nbd_malloc(): public free list %p first block %p", pub, pub->next);
97 if (EXPECT_FALSE(pub == pri))
100 if (pub->next != NULL) {
101 block_t *stolen = SYNC_SWAP(&pub->next, NULL);
102 TRACE("m0", "nbd_malloc(): stole list %p first block %p", pub, stolen);
105 TRACE("m0", "nbd_malloc(): append to last block %p of private free list", last, 0);
108 TRACE("m0", "nbd_malloc(): find last block in list: last %p last->next %p",
115 TRACE("m0", "nbd_malloc(): moved %llu blocks from public to private free lists", cnt, 0);
117 if (b_scale >= REGION_SCALE) {
119 assert(pri->next == NULL);
120 pri->next = (block_t *)get_new_region(b_scale);
121 assert(pri->next->next == NULL);
125 } else if (cnt < (1 << (REGION_SCALE - b_scale - 1))) {
127 // Even if we took a few blocks from our public lists we still break open a new region.
128 // This guarentees that we are amortizing the cost of accessing our public lists accross
129 // many nbd_malloc() calls.
130 char *region = get_new_region(b_scale);
131 size_t b_size = 1 << b_scale;
132 for (int i = REGION_SIZE; i != 0; i -= b_size) {
133 block_t *b = (block_t *)(region + i - b_size);
135 //TRACE("m1", "nbd_malloc(): put new block %p ahead of %p on private list", b, b->next);
144 // Pull a block off of our private free list.
145 block_t *b = pri->next;
146 TRACE("m0", "nbd_malloc(): take block %p off of of private list (new head is %p)", b, pri->next);