2 * Sherlock Library -- Fast Database Management Routines
4 * (c) 1999 Martin Mares <mj@atrey.karlin.mff.cuni.cz>
8 * This library uses the standard algorithm for external hashing (page directory
9 * mapping topmost K bits of hash value to page address, directory splits and
10 * so on). Peculiarities of this implementation (aka design decisions):
12 * o We allow both fixed and variable length keys and values (this includes
13 * zero size values for cases you want to represent only a set of keys).
14 * o We assume that key_size + val_size < page_size.
15 * o We never shrink the directory nor free empty pages. (The reason is that
16 * if the database was once large, it's likely it will again become large soon.)
17 * o The only pages which can be freed are those of the directory (during
18 * directory split), so we keep only a simple 32-entry free block list
19 * and we assume it's sorted.
29 #include "pagecache.h"
31 #include "db_internal.h"
34 sdbm_open(struct sdbm_options *o)
37 struct sdbm_root root, *r;
38 uns cache_size = o->cache_size ? o->cache_size : 16;
40 d = xmalloc(sizeof(struct sdbm));
43 d->fd = open(o->name, ((d->flags & SDBM_WRITE) ? O_RDWR : O_RDONLY), 0666);
44 if (d->fd >= 0) /* Already exists, let's check it */
46 if (read(d->fd, &root, sizeof(root)) != sizeof(root))
48 if (root.magic != SDBM_MAGIC || root.version != SDBM_VERSION)
50 d->file_size = lseek(d->fd, 0, SEEK_END);
51 d->page_size = 1 << root.page_order;
52 d->cache = pgc_open(d->page_size, cache_size);
53 d->root_page = pgc_read(d->cache, d->fd, 0);
54 d->root = (void *) d->root_page->data;
55 /* FIXME: Should we do some locking? */
57 else if ((d->flags & SDBM_CREAT) && (d->fd = open(o->name, O_RDWR | O_CREAT, 0666)) >= 0)
60 uns page_order = o->page_order;
63 d->page_size = 1 << page_order;
64 d->cache = pgc_open(d->page_size, cache_size);
65 d->root_page = pgc_get_zero(d->cache, d->fd, 0);
66 r = d->root = (void *) d->root_page->data; /* Build root page */
67 r->magic = SDBM_MAGIC;
68 r->version = SDBM_VERSION;
69 r->page_order = page_order;
70 r->key_size = o->key_size;
71 r->val_size = o->val_size;
72 r->dir_start = d->page_size;
74 d->file_size = 3*d->page_size;
75 q = pgc_get_zero(d->cache, d->fd, d->page_size); /* Build page directory */
76 ((u32 *)q->data)[0] = 2*d->page_size;
78 q = pgc_get_zero(d->cache, d->fd, 2*d->page_size); /* Build single data page */
83 d->dir_size = 1 << d->root->dir_order;
84 d->dir_shift = 32 - d->root->dir_order;
85 d->page_order = d->root->page_order;
86 d->page_mask = d->page_size - 1;
87 d->key_size = d->root->key_size;
88 d->val_size = d->root->val_size;
97 sdbm_close(struct sdbm *d)
100 pgc_put(d->cache, d->root_page);
109 sdbm_alloc_pages(struct sdbm *d, uns number)
111 uns where = d->file_size;
112 d->file_size += number << d->page_order;
117 sdbm_alloc_page(struct sdbm *d)
121 if (!d->root->free_pool[0].count)
122 return sdbm_alloc_pages(d, 1);
123 pos = d->root->free_pool[0].first;
124 d->root->free_pool[0].first += d->page_size;
125 if (!--d->root->free_pool[0].count)
127 memmove(d->root->free_pool, d->root->free_pool+1, SDBM_NUM_FREE_PAGE_POOLS * sizeof(d->root->free_pool[0]));
128 d->root->free_pool[SDBM_NUM_FREE_PAGE_POOLS-1].count = 0;
130 pgc_mark_dirty(d->cache, d->root_page);
135 sdbm_free_pages(struct sdbm *d, uns start, uns number)
139 while (d->root->free_pool[i].count)
141 d->root->free_pool[i].first = start;
142 d->root->free_pool[i].count = number;
143 pgc_mark_dirty(d->cache, d->root_page);
147 sdbm_hash(byte *key, uns keylen)
150 * This is the same hash function as GDBM uses.
151 * FIXME: Use a faster one?
156 /* Set the initial value from key. */
157 value = 0x238F13AF * keylen;
158 for (index = 0; index < keylen; index++)
159 value = value + (key[index] << (index*5 % 24));
160 return (1103515243 * value + 12345);
164 sdbm_get_entry(struct sdbm *d, byte *pos, byte **key, uns *keylen, byte **val, uns *vallen)
168 if (d->key_size >= 0)
169 *keylen = d->key_size;
172 *keylen = (p[0] << 8) | p[1];
177 if (d->val_size >= 0)
178 *vallen = d->val_size;
181 *vallen = (p[0] << 8) | p[1];
190 sdbm_entry_len(struct sdbm *d, uns keylen, uns vallen)
192 uns len = keylen + vallen;
201 sdbm_store_entry(struct sdbm *d, byte *pos, byte *key, uns keylen, byte *val, uns vallen)
205 *pos++ = keylen >> 8;
208 memmove(pos, key, keylen);
212 *pos++ = vallen >> 8;
215 memmove(pos, val, vallen);
219 sdbm_page_rank(struct sdbm *d, uns dirpos)
224 uns pm = d->page_mask;
226 b = pgc_read(d->cache, d->fd, d->root->dir_start + (dirpos & ~pm));
227 pg = GET32(b->data, dirpos & pm);
229 while ((l & pm) && GET32(b->data, (l - 4) & pm) == pg)
232 /* We heavily depend on unused directory entries being zero */
233 while ((r & pm) && GET32(b->data, r & pm) == pg)
235 pgc_put(d->cache, b);
237 if (!(l & pm) && !(r & pm))
239 /* Note that if it spans page boundary, it must contain an integer number of pages */
242 b = pgc_read(d->cache, d->fd, d->root->dir_start + ((l - 4) & ~pm));
243 x = GET32(b->data, 0);
244 pgc_put(d->cache, b);
249 while (r < 4*d->dir_size)
251 b = pgc_read(d->cache, d->fd, d->root->dir_start + (r & ~pm));
252 x = GET32(b->data, 0);
253 pgc_put(d->cache, b);
263 sdbm_expand_directory(struct sdbm *d)
269 if (4*d->dir_size < d->page_size)
271 /* It still fits within single page */
272 b = pgc_read(d->cache, d->fd, d->root->dir_start);
273 dir = (u32 *) b->data;
274 for(i=d->dir_size-1; i>=0; i--)
275 dir[2*i] = dir[2*i+1] = dir[i];
276 pgc_mark_dirty(d->cache, b);
277 pgc_put(d->cache, b);
281 uns old_dir = d->root->dir_start;
282 uns old_dir_pages = 1 << (d->root->dir_order + 2 - d->page_order);
283 uns page, ent, new_dir;
284 new_dir = d->root->dir_start = sdbm_alloc_pages(d, 2*old_dir_pages);
285 ent = 1 << (d->page_order - 3);
286 for(page=0; page < old_dir_pages; page++)
288 b = pgc_read(d->cache, d->fd, old_dir + (page << d->page_order));
289 dir = (u32 *) b->data;
290 c = pgc_get(d->cache, d->fd, new_dir + (page << (d->page_order + 1)));
293 t[2*i] = t[2*i+1] = dir[i];
294 pgc_put(d->cache, c);
295 c = pgc_get(d->cache, d->fd, new_dir + (page << (d->page_order + 1)) + d->page_size);
298 t[2*i] = t[2*i+1] = dir[ent+i];
299 pgc_put(d->cache, c);
300 pgc_put(d->cache, b);
302 if (!(d->flags & SDBM_FAST))
305 * Unless in super-fast mode, fill old directory pages with zeroes.
306 * This slows us down a bit, but allows database reconstruction after
307 * the free list is lost.
309 for(page=0; page < old_dir_pages; page++)
311 b = pgc_get_zero(d->cache, d->fd, old_dir + (page << d->page_order));
312 pgc_put(d->cache, b);
315 sdbm_free_pages(d, old_dir, old_dir_pages);
318 d->root->dir_order++;
319 d->dir_size = 1 << d->root->dir_order;
320 d->dir_shift = 32 - d->root->dir_order;
321 pgc_mark_dirty(d->cache, d->root_page);
322 if (!(d->flags & SDBM_FAST))
327 sdbm_split_data(struct sdbm *d, struct sdbm_bucket *s, struct sdbm_bucket *d0, struct sdbm_bucket *d1, uns sigbit)
330 byte *dp[2] = { d0->data, d1->data };
334 while (sp < s->data + s->used)
336 sz = sdbm_get_entry(d, sp, &K, &Kl, &D, &Dl);
338 i = (sdbm_hash(K, Kl) & (1 << sigbit)) ? 1 : 0;
339 sdbm_store_entry(d, dp[i], K, Kl, D, Dl);
342 d0->used = dp[0] - d0->data;
343 d1->used = dp[1] - d1->data;
347 sdbm_split_dir(struct sdbm *d, uns dirpos, uns count, uns pos)
355 b = pgc_read(d->cache, d->fd, d->root->dir_start + (dirpos & ~d->page_mask));
356 i = d->page_size - (dirpos & d->page_mask);
362 GET32(b->data, dirpos & d->page_mask) = pos;
366 pgc_mark_dirty(d->cache, b);
367 pgc_put(d->cache, b);
372 sdbm_split_page(struct sdbm *d, struct page *b, u32 hash, uns dirpos)
375 uns i, rank, sigbit, rank_log;
377 rank = sdbm_page_rank(d, dirpos); /* rank = # of pointers to this page */
380 sdbm_expand_directory(d);
384 rank_log = 1; /* rank_log = log2(rank) */
385 while ((1 << rank_log) < rank)
387 sigbit = d->dir_shift + rank_log - 1; /* sigbit = bit we split on */
389 p[1] = pgc_get_zero(d->cache, d->fd, sdbm_alloc_page(d));
390 sdbm_split_data(d, (void *) b->data, (void *) p[0]->data, (void *) p[1]->data, sigbit);
391 sdbm_split_dir(d, (dirpos & ~(4*rank - 1))+2*rank, rank/2, pgc_page_pos(d->cache, p[1]));
392 pgc_mark_dirty(d->cache, p[0]);
393 i = (hash & (1 << sigbit)) ? 1 : 0;
394 pgc_put(d->cache, p[!i]);
399 sdbm_access(struct sdbm *d, byte *key, uns keylen, byte *val, uns *vallen, uns mode) /* 0=read, 1=store, 2=replace */
402 u32 hash, h, pos, size;
403 struct sdbm_bucket *b;
407 if ((d->key_size >= 0 && keylen != d->key_size) || keylen > 65535)
408 return SDBM_ERROR_BAD_KEY_SIZE;
409 if (val && ((d->val_size >= 0 && *vallen != d->val_size) || *vallen >= 65535))
410 return SDBM_ERROR_BAD_VAL_SIZE;
411 if (!mode && !(d->flags & SDBM_WRITE))
412 return SDBM_ERROR_READ_ONLY;
413 hash = sdbm_hash(key, keylen);
414 if (d->dir_shift != 32) /* avoid shifting by 32 bits */
415 h = (hash >> d->dir_shift) << 2; /* offset in the directory */
418 p = pgc_read(d->cache, d->fd, d->root->dir_start + (h & ~d->page_mask));
419 pos = GET32(p->data, h & d->page_mask);
420 pgc_put(d->cache, p);
421 q = pgc_read(d->cache, d->fd, pos);
422 b = (void *) q->data;
429 s = sdbm_get_entry(d, c, &K, &Kl, &D, &Dl);
430 if (Kl == keylen && !memcmp(K, key, Kl))
435 case 0: /* fetch: found */
445 pgc_put(d->cache, q);
447 case 1: /* store: already present */
448 pgc_put(d->cache, q);
450 default: /* replace: delete the old one */
451 memmove(c, c+s, e-(c+s));
458 if (!mode || !val) /* fetch or delete: no success */
460 pgc_put(d->cache, q);
467 size = sdbm_entry_len(d, keylen, *vallen);
468 while (b->used + size > d->page_size - sizeof(struct sdbm_bucket))
470 /* Page overflow, need to split */
471 q = sdbm_split_page(d, q, hash, h);
472 b = (void *) q->data;
474 sdbm_store_entry(d, b->data + b->used, key, keylen, val, *vallen);
477 pgc_mark_dirty(d->cache, q);
478 pgc_put(d->cache, q);
479 if (d->flags & SDBM_SYNC)
485 sdbm_store(struct sdbm *d, byte *key, uns keylen, byte *val, uns vallen)
487 return sdbm_access(d, key, keylen, val, &vallen, 1);
491 sdbm_replace(struct sdbm *d, byte *key, uns keylen, byte *val, uns vallen)
493 return sdbm_access(d, key, keylen, val, &vallen, 2);
497 sdbm_delete(struct sdbm *d, byte *key, uns keylen)
499 return sdbm_access(d, key, keylen, NULL, NULL, 2);
503 sdbm_fetch(struct sdbm *d, byte *key, uns keylen, byte *val, uns *vallen)
505 return sdbm_access(d, key, keylen, val, vallen, 0);
509 sdbm_rewind(struct sdbm *d)
514 sdbm_get_next(struct sdbm *d, byte *key, uns *keylen, byte *val, uns *vallen)
520 sdbm_sync(struct sdbm *d)
523 if (d->flags & SDBM_FSYNC)
532 struct sdbm_options o = {
534 flags: SDBM_CREAT | SDBM_WRITE,
549 for(i=0; i<1000000; i++)
551 sprintf(buf, "%d", i);
552 k = sdbm_store(d, buf, strlen(buf), (byte *) &i, sizeof(i));
553 // printf("%s:%d\r", buf, k);
559 for(i=0; i<1000000; i++)
561 sprintf(buf, "%d", i);
562 j = sdbm_fetch(d, buf, strlen(buf), NULL, NULL);
563 // printf("%s:%d\r", buf, j);
566 { printf("\nERR: %s %d %x %d\n", buf, j, sdbm_hash(buf, strlen(buf)), d->dir_shift); return 1; }