mdb.c 61.5 KB
Newer Older
Howard Chu's avatar
Howard Chu committed
1
2
3
4
5
6
7
8
9
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/queue.h>
#include <sys/param.h>
#include <sys/uio.h>
#include <sys/mman.h>
#ifdef HAVE_SYS_FILE_H
#include <sys/file.h>
#endif
Howard Chu's avatar
Howard Chu committed
10
#include <fcntl.h>
Howard Chu's avatar
Howard Chu committed
11
12
13
14
15
16
17
18
19
20
21
22

#include <assert.h>
#include <err.h>
#include <errno.h>
#include <fcntl.h>
#include <stddef.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <unistd.h>
Howard Chu's avatar
Howard Chu committed
23
#include <pthread.h>
Howard Chu's avatar
Howard Chu committed
24
25

#include "mdb.h"
Howard Chu's avatar
Howard Chu committed
26

Howard Chu's avatar
Howard Chu committed
27
28
#define ULONG		unsigned long
typedef ULONG		pgno_t;
Howard Chu's avatar
Howard Chu committed
29

Howard Chu's avatar
Howard Chu committed
30
#include "idl.h"
Howard Chu's avatar
Howard Chu committed
31

32
33
34
#ifndef DEBUG
#define DEBUG 1
#endif
Howard Chu's avatar
Howard Chu committed
35

36
37
38
#if (DEBUG +0) && defined(__GNUC__)
# define DPRINTF(fmt, ...) \
	fprintf(stderr, "%s:%d: " fmt "\n", __func__, __LINE__, ##__VA_ARGS__)
Howard Chu's avatar
Howard Chu committed
39
#else
40
# define DPRINTF(...)	((void) 0)
Howard Chu's avatar
Howard Chu committed
41
42
43
44
45
46
47
48
#endif

#define PAGESIZE	 4096
#define MDB_MINKEYS	 4
#define MDB_MAGIC	 0xBEEFC0DE
#define MDB_VERSION	 1
#define MAXKEYSIZE	 255

Hallvard Furuseth's avatar
Hallvard Furuseth committed
49
#define P_INVALID	 (~0UL)
Howard Chu's avatar
Howard Chu committed
50
51
52
53
54

#define F_ISSET(w, f)	 (((w) & (f)) == (f))

typedef uint16_t	 indx_t;

Howard Chu's avatar
Howard Chu committed
55
56
57
58
59
#define DEFAULT_READERS	126
#define DEFAULT_MAPSIZE	1048576

/* Lock descriptor stuff */
#define RXBODY	\
Howard Chu's avatar
Howard Chu committed
60
	ULONG		mr_txnid; \
Howard Chu's avatar
Howard Chu committed
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
	pid_t		mr_pid; \
	pthread_t	mr_tid
typedef struct MDB_rxbody {
	RXBODY;
} MDB_rxbody;

#ifndef CACHELINE
#define CACHELINE	64	/* most CPUs. Itanium uses 128 */
#endif

typedef struct MDB_reader {
	RXBODY;
	/* cache line alignment */
	char pad[CACHELINE-sizeof(MDB_rxbody)];
} MDB_reader;

#define	TXBODY \
	uint32_t	mt_magic;	\
	uint32_t	mt_version;	\
	pthread_mutex_t	mt_mutex;	\
Howard Chu's avatar
Howard Chu committed
81
	ULONG		mt_txnid;	\
Howard Chu's avatar
Howard Chu committed
82
83
84
85
86
87
88
89
90
91
92
93
94
	uint32_t	mt_numreaders
typedef struct MDB_txbody {
	TXBODY;
} MDB_txbody;

typedef struct MDB_txninfo {
	TXBODY;
	char pad[CACHELINE-sizeof(MDB_txbody)];
	pthread_mutex_t	mt_wmutex;
	char pad2[CACHELINE-sizeof(pthread_mutex_t)];
	MDB_reader	mt_readers[1];
} MDB_txninfo;

Howard Chu's avatar
Howard Chu committed
95
96
97
98
99
100
101
102
103
104
/* Common header for all page types. Overflow pages
 * occupy a number of contiguous pages with no
 * headers on any page after the first.
 */
typedef struct MDB_page {		/* represents a page of storage */
	pgno_t		mp_pgno;		/* page number */
#define	P_BRANCH	 0x01		/* branch page */
#define	P_LEAF		 0x02		/* leaf page */
#define	P_OVERFLOW	 0x04		/* overflow page */
#define	P_META		 0x08		/* meta page */
Howard Chu's avatar
Howard Chu committed
105
#define	P_DIRTY		 0x10		/* dirty page */
Howard Chu's avatar
Howard Chu committed
106
107
108
109
110
111
112
113
114
115
116
117
118
119
	uint32_t	mp_flags;
#define mp_lower	mp_pb.pb.pb_lower
#define mp_upper	mp_pb.pb.pb_upper
#define mp_pages	mp_pb.pb_pages
	union page_bounds {
		struct {
			indx_t		pb_lower;		/* lower bound of free space */
			indx_t		pb_upper;		/* upper bound of free space */
		} pb;
		uint32_t	pb_pages;	/* number of overflow pages */
	} mp_pb;
	indx_t		mp_ptrs[1];		/* dynamic size */
} MDB_page;

120
#define PAGEHDRSZ	 ((unsigned) offsetof(MDB_page, mp_ptrs))
Howard Chu's avatar
Howard Chu committed
121
122
123

#define NUMKEYS(p)	 (((p)->mp_lower - PAGEHDRSZ) >> 1)
#define SIZELEFT(p)	 (indx_t)((p)->mp_upper - (p)->mp_lower)
Howard Chu's avatar
Howard Chu committed
124
125
#define PAGEFILL(env, p) (1000L * ((env)->me_meta.mm_psize - PAGEHDRSZ - SIZELEFT(p)) / \
				((env)->me_meta.mm_psize - PAGEHDRSZ))
Howard Chu's avatar
Howard Chu committed
126
127
128
129
#define IS_LEAF(p)	 F_ISSET((p)->mp_flags, P_LEAF)
#define IS_BRANCH(p)	 F_ISSET((p)->mp_flags, P_BRANCH)
#define IS_OVERFLOW(p)	 F_ISSET((p)->mp_flags, P_OVERFLOW)

Howard Chu's avatar
Howard Chu committed
130
131
#define OVPAGES(size, psize)	(PAGEHDRSZ + size + psize - 1) / psize;

Howard Chu's avatar
Howard Chu committed
132
typedef struct MDB_meta {			/* meta (footer) page content */
Howard Chu's avatar
Howard Chu committed
133
134
135
136
	uint32_t	mm_magic;
	uint32_t	mm_version;
	void		*mm_address;		/* address for fixed mapping */
	size_t		mm_mapsize;			/* size of mmap region */
Howard Chu's avatar
Howard Chu committed
137
138
	pgno_t		mm_last_pg;			/* last used page in file */
	ULONG		mm_txnid;			/* txnid that committed this page */
Howard Chu's avatar
Howard Chu committed
139
140
141
	uint32_t	mm_psize;
	uint16_t	mm_flags;
	uint16_t	mm_depth;
Howard Chu's avatar
Howard Chu committed
142
143
144
145
	ULONG		mm_branch_pages;
	ULONG		mm_leaf_pages;
	ULONG		mm_overflow_pages;
	ULONG		mm_entries;
Howard Chu's avatar
Howard Chu committed
146
	pgno_t		mm_root;
Howard Chu's avatar
Howard Chu committed
147
148
149
150
151
} MDB_meta;

typedef struct MDB_dhead {					/* a dirty page */
	SIMPLEQ_ENTRY(MDB_dpage)	 md_next;	/* queue of dirty pages */
	MDB_page	*md_parent;
152
	unsigned	md_pi;				/* parent index */
Howard Chu's avatar
Howard Chu committed
153
154
155
156
157
158
159
160
161
162
	int			md_num;
} MDB_dhead;

typedef struct MDB_dpage {
	MDB_dhead	h;
	MDB_page	p;
} MDB_dpage;

SIMPLEQ_HEAD(dirty_queue, MDB_dpage);

Howard Chu's avatar
Howard Chu committed
163
164
typedef struct MDB_oldpages {
	struct MDB_oldpages *mo_next;
Howard Chu's avatar
Howard Chu committed
165
	ULONG		mo_txnid;
Howard Chu's avatar
Howard Chu committed
166
167
168
	pgno_t		mo_pages[1];	/* dynamic */
} MDB_oldpages;

Howard Chu's avatar
Howard Chu committed
169
170
171
typedef struct MDB_pageparent {
	MDB_page *mp_page;
	MDB_page *mp_parent;
172
	unsigned mp_pi;
Howard Chu's avatar
Howard Chu committed
173
174
} MDB_pageparent;

Howard Chu's avatar
Howard Chu committed
175
static MDB_dpage *mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num);
Howard Chu's avatar
Howard Chu committed
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
static int 		mdb_touch(MDB_txn *txn, MDB_pageparent *mp);

typedef struct MDB_ppage {					/* ordered list of pages */
	SLIST_ENTRY(MDB_ppage)	 mp_entry;
	MDB_page		*mp_page;
	unsigned int	mp_ki;		/* cursor index on page */
} MDB_ppage;
SLIST_HEAD(page_stack, MDB_ppage);

#define CURSOR_EMPTY(c)		 SLIST_EMPTY(&(c)->mc_stack)
#define CURSOR_TOP(c)		 SLIST_FIRST(&(c)->mc_stack)
#define CURSOR_POP(c)		 SLIST_REMOVE_HEAD(&(c)->mc_stack, mp_entry)
#define CURSOR_PUSH(c,p)	 SLIST_INSERT_HEAD(&(c)->mc_stack, p, mp_entry)

struct MDB_cursor {
	MDB_db		*mc_db;
	MDB_txn		*mc_txn;
	struct page_stack	 mc_stack;		/* stack of parent pages */
	short		mc_initialized;	/* 1 if initialized */
	short		mc_eof;		/* 1 if end is reached */
};

#define METAHASHLEN	 offsetof(MDB_meta, mm_hash)
#define METADATA(p)	 ((void *)((char *)p + PAGEHDRSZ))

typedef struct MDB_node {
#define mn_pgno		 mn_p.np_pgno
#define mn_dsize	 mn_p.np_dsize
	union {
		pgno_t		 np_pgno;	/* child page number */
		uint32_t	 np_dsize;	/* leaf data size */
	} mn_p;
Howard Chu's avatar
Howard Chu committed
208
209
	unsigned int	mn_flags:4;
	unsigned int	mn_ksize:12;			/* key size */
Howard Chu's avatar
Howard Chu committed
210
211
212
213
214
215
216
#define F_BIGDATA	 0x01			/* data put on overflow page */
	char		mn_data[1];
} MDB_node;

struct MDB_txn {
	pgno_t		mt_root;		/* current / new root page */
	pgno_t		mt_next_pgno;	/* next unallocated page */
Howard Chu's avatar
Howard Chu committed
217
218
	ULONG		mt_txnid;
	ULONG		mt_oldest;
Howard Chu's avatar
Howard Chu committed
219
	MDB_env		*mt_env;	
Howard Chu's avatar
Howard Chu committed
220
	pgno_t		*mt_free_pgs;	/* this is an IDL */
Howard Chu's avatar
Howard Chu committed
221
222
223
224
	union {
		struct dirty_queue	*dirty_queue;	/* modified pages */
		MDB_reader	*reader;
	} mt_u;
Howard Chu's avatar
Howard Chu committed
225
226
#define MDB_TXN_RDONLY		 0x01		/* read-only transaction */
#define MDB_TXN_ERROR		 0x02		/* an error has occurred */
Howard Chu's avatar
Howard Chu committed
227
#define MDB_TXN_METOGGLE	0x04		/* used meta page 1 */
Howard Chu's avatar
Howard Chu committed
228
229
230
231
	unsigned int		 mt_flags;
};

struct MDB_db {
Howard Chu's avatar
Howard Chu committed
232
233
	MDB_db		*md_next;
	char		*md_name;
Howard Chu's avatar
Howard Chu committed
234
235
236
237
	MDB_cmp_func	*md_cmp;		/* user compare function */
	MDB_rel_func	*md_rel;		/* user relocate function */
	MDB_db			*md_parent;		/* parent tree */
	MDB_env 		*md_env;
Howard Chu's avatar
Howard Chu committed
238
	uint32_t	md_pad;
Howard Chu's avatar
Howard Chu committed
239
240
	uint16_t	md_flags;
	uint16_t	md_depth;
Howard Chu's avatar
Howard Chu committed
241
242
243
244
245
	ULONG		md_branch_pages;
	ULONG		md_leaf_pages;
	ULONG		md_overflow_pages;
	ULONG		md_entries;
	pgno_t		md_root;
Howard Chu's avatar
Howard Chu committed
246
247
248
249
};

struct MDB_env {
	int			me_fd;
Howard Chu's avatar
Howard Chu committed
250
	int			me_lfd;
Howard Chu's avatar
Howard Chu committed
251
	uint32_t	me_flags;
Howard Chu's avatar
Howard Chu committed
252
	unsigned int			me_maxreaders;
Howard Chu's avatar
Howard Chu committed
253
	char		*me_path;
Howard Chu's avatar
Howard Chu committed
254
	char		*me_map;
Howard Chu's avatar
Howard Chu committed
255
	MDB_txninfo	*me_txns;
Howard Chu's avatar
Howard Chu committed
256
	MDB_db		me_db;		/* first DB */
Howard Chu's avatar
Howard Chu committed
257
	MDB_meta	me_meta;
Howard Chu's avatar
Howard Chu committed
258
	MDB_txn		*me_txn;		/* current write transaction */
Howard Chu's avatar
Howard Chu committed
259
260
	size_t		me_mapsize;
	off_t		me_size;		/* current file size */
Howard Chu's avatar
Howard Chu committed
261
	pthread_key_t	me_txkey;	/* thread-key for readers */
Howard Chu's avatar
Howard Chu committed
262
263
	MDB_oldpages *me_pghead;
	MDB_oldpages *me_pgtail;
Howard Chu's avatar
Howard Chu committed
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
};

#define NODESIZE	 offsetof(MDB_node, mn_data)

#define INDXSIZE(k)	 (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
#define LEAFSIZE(k, d)	 (NODESIZE + (k)->mv_size + (d)->mv_size)
#define NODEPTR(p, i)	 ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
#define NODEKEY(node)	 (void *)((node)->mn_data)
#define NODEDATA(node)	 (void *)((char *)(node)->mn_data + (node)->mn_ksize)
#define NODEPGNO(node)	 ((node)->mn_pgno)
#define NODEDSZ(node)	 ((node)->mn_dsize)

#define MDB_COMMIT_PAGES	 64	/* max number of pages to write in one commit */
#define MDB_MAXCACHE_DEF	 1024	/* max number of pages to keep in cache  */

static int  mdb_search_page_root(MDB_db *db,
			    MDB_val *key,
			    MDB_cursor *cursor, int modify,
			    MDB_pageparent *mpp);
static int  mdb_search_page(MDB_db *db,
			    MDB_txn *txn, MDB_val *key,
			    MDB_cursor *cursor, int modify,
			    MDB_pageparent *mpp);

static int  mdbenv_read_header(MDB_env *env);
static int  mdb_check_meta_page(MDB_page *p);
Howard Chu's avatar
Howard Chu committed
290
static int  mdbenv_read_meta(MDB_env *env, int *which);
291
static int  mdbenv_write_meta(MDB_txn *txn);
Howard Chu's avatar
Howard Chu committed
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);

static MDB_node *mdb_search_node(MDB_db *db, MDB_page *mp,
			    MDB_val *key, int *exactp, unsigned int *kip);
static int  mdb_add_node(MDB_db *bt, MDB_page *mp,
			    indx_t indx, MDB_val *key, MDB_val *data,
			    pgno_t pgno, uint8_t flags);
static void mdb_del_node(MDB_db *bt, MDB_page *mp,
			    indx_t indx);
static int  mdb_read_data(MDB_db *bt, MDB_page *mp,
			    MDB_node *leaf, MDB_val *data);

static int		 mdb_rebalance(MDB_db *bt, MDB_pageparent *mp);
static int		 mdb_update_key(MDB_db *bt, MDB_page *mp,
			    indx_t indx, MDB_val *key);
static int		 mdb_move_node(MDB_db *bt, 
				MDB_pageparent *src, indx_t srcindx,
				MDB_pageparent *dst, indx_t dstindx);
static int		 mdb_merge(MDB_db *bt, MDB_pageparent *src,
			    MDB_pageparent *dst);
static int		 mdb_split(MDB_db *bt, MDB_page **mpp,
			    unsigned int *newindxp, MDB_val *newkey,
			    MDB_val *newdata, pgno_t newpgno);
Howard Chu's avatar
Howard Chu committed
315
static MDB_dpage *mdb_new_page(MDB_db *db, uint32_t flags, int num);
Howard Chu's avatar
Howard Chu committed
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342

static void		 cursor_pop_page(MDB_cursor *cursor);
static MDB_ppage *cursor_push_page(MDB_cursor *cursor,
			    MDB_page *mp);

static int		 mdb_set_key(MDB_db *bt, MDB_page *mp,
			    MDB_node *node, MDB_val *key);
static int		 mdb_sibling(MDB_cursor *cursor, int move_right);
static int		 mdb_cursor_next(MDB_cursor *cursor,
			    MDB_val *key, MDB_val *data);
static int		 mdb_cursor_set(MDB_cursor *cursor,
			    MDB_val *key, MDB_val *data, int *exactp);
static int		 mdb_cursor_first(MDB_cursor *cursor,
			    MDB_val *key, MDB_val *data);

static size_t		 mdb_leaf_size(MDB_db *bt, MDB_val *key,
			    MDB_val *data);
static size_t		 mdb_branch_size(MDB_db *bt, MDB_val *key);

static int		 memncmp(const void *s1, size_t n1,
				 const void *s2, size_t n2);
static int		 memnrcmp(const void *s1, size_t n1,
				  const void *s2, size_t n2);

static int
memncmp(const void *s1, size_t n1, const void *s2, size_t n2)
{
343
344
345
346
347
	int diff, len_diff = -1;

	if (n1 >= n2) {
		len_diff = (n1 > n2);
		n1 = n2;
Howard Chu's avatar
Howard Chu committed
348
	}
349
350
	diff = memcmp(s1, s2, n1);
	return diff ? diff : len_diff;
Howard Chu's avatar
Howard Chu committed
351
352
353
354
355
}

static int
memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
{
356
	const unsigned char	*p1, *p2, *p1_lim;
Howard Chu's avatar
Howard Chu committed
357
358

	if (n2 == 0)
359
360
361
		return n1 != 0;
	if (n1 == 0)
		return -1;
Howard Chu's avatar
Howard Chu committed
362
363
364
365

	p1 = (const unsigned char *)s1 + n1 - 1;
	p2 = (const unsigned char *)s2 + n2 - 1;

366
367
368
	for (p1_lim = (n1 <= n2 ? s1 : s2);  *p1 == *p2;  p1--, p2--) {
		if (p1 == p1_lim)
			return (p1 != s1) ? (p1 != p2) : (p2 != s2) ? -1 : 0;
Howard Chu's avatar
Howard Chu committed
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
	}
	return *p1 - *p2;
}

int
mdb_cmp(MDB_db *db, const MDB_val *a, const MDB_val *b)
{
	return db->md_cmp(a, b);
}

static int
_mdb_cmp(MDB_db *db, const MDB_val *key1, const MDB_val *key2)
{
	if (F_ISSET(db->md_flags, MDB_REVERSEKEY))
		return memnrcmp(key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
	else
		return memncmp((char *)key1->mv_data, key1->mv_size, key2->mv_data, key2->mv_size);
}

/* Allocate new page(s) for writing */
static MDB_dpage *
Howard Chu's avatar
Howard Chu committed
390
mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
Howard Chu's avatar
Howard Chu committed
391
392
{
	MDB_dpage *dp;
Howard Chu's avatar
Howard Chu committed
393
394
395
	pgno_t pgno = P_INVALID;

	if (txn->mt_env->me_pghead) {
Howard Chu's avatar
Howard Chu committed
396
		ULONG oldest = txn->mt_txnid - 2;
Howard Chu's avatar
Howard Chu committed
397
		unsigned int i;
Howard Chu's avatar
Howard Chu committed
398
399
400
401
402
403
404
405
		for (i=0; i<txn->mt_env->me_txns->mt_numreaders; i++) {
			if (txn->mt_env->me_txns->mt_readers[i].mr_txnid < oldest)
				oldest = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
		}
		if (oldest > txn->mt_env->me_pghead->mo_txnid) {
			MDB_oldpages *mop = txn->mt_env->me_pghead;
			txn->mt_oldest = oldest;
			if (num > 1) {
406
407
408
409
				/* FIXME: For now, always use fresh pages. We
				 * really ought to search the free list for a
				 * contiguous range.
				 */
Howard Chu's avatar
Howard Chu committed
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
				;
			} else {
				/* peel pages off tail, so we only have to truncate the list */
				pgno = MDB_IDL_LAST(mop->mo_pages);
				if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
					mop->mo_pages[2]++;
					if (mop->mo_pages[2] > mop->mo_pages[1])
						mop->mo_pages[0] = 0;
				} else {
					mop->mo_pages[0]--;
				}
				if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
					txn->mt_env->me_pghead = mop->mo_next;
					if (!txn->mt_env->me_pghead)
						txn->mt_env->me_pgtail = NULL;
					free(mop);
				}
			}
		}
	}
Howard Chu's avatar
Howard Chu committed
430

Howard Chu's avatar
Howard Chu committed
431
	if ((dp = malloc(txn->mt_env->me_meta.mm_psize * num + sizeof(MDB_dhead))) == NULL)
Howard Chu's avatar
Howard Chu committed
432
433
434
435
		return NULL;
	dp->h.md_num = num;
	dp->h.md_parent = parent;
	dp->h.md_pi = parent_idx;
Howard Chu's avatar
Howard Chu committed
436
	SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
Howard Chu's avatar
Howard Chu committed
437
438
439
440
441
442
	if (pgno == P_INVALID) {
		dp->p.mp_pgno = txn->mt_next_pgno;
		txn->mt_next_pgno += num;
	} else {
		dp->p.mp_pgno = pgno;
	}
Howard Chu's avatar
Howard Chu committed
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458

	return dp;
}

/* Touch a page: make it dirty and re-insert into tree with updated pgno.
 */
static int
mdb_touch(MDB_txn *txn, MDB_pageparent *pp)
{
	MDB_page *mp = pp->mp_page;
	pgno_t	pgno;
	assert(txn != NULL);
	assert(pp != NULL);

	if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
		MDB_dpage *dp;
Howard Chu's avatar
Howard Chu committed
459
		if ((dp = mdb_alloc_page(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
Howard Chu's avatar
Howard Chu committed
460
			return ENOMEM;
Howard Chu's avatar
Howard Chu committed
461
462
		DPRINTF("touched page %lu -> %lu", mp->mp_pgno, dp->p.mp_pgno);
		mdb_idl_insert(txn->mt_free_pgs, mp->mp_pgno);
Howard Chu's avatar
Howard Chu committed
463
		pgno = dp->p.mp_pgno;
Howard Chu's avatar
Howard Chu committed
464
		bcopy(mp, &dp->p, txn->mt_env->me_meta.mm_psize);
Howard Chu's avatar
Howard Chu committed
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
		mp = &dp->p;
		mp->mp_pgno = pgno;
		mp->mp_flags |= P_DIRTY;

		/* Update the page number to new touched page. */
		if (pp->mp_parent != NULL)
			NODEPGNO(NODEPTR(pp->mp_parent, pp->mp_pi)) = mp->mp_pgno;
		pp->mp_page = mp;
	}
	return 0;
}

int
mdbenv_sync(MDB_env *env)
{
	int rc = 0;
	if (!F_ISSET(env->me_flags, MDB_NOSYNC)) {
		if (fsync(env->me_fd))
			rc = errno;
	}
	return rc;
}

int
mdb_txn_begin(MDB_env *env, int rdonly, MDB_txn **ret)
{
	MDB_txn	*txn;
Howard Chu's avatar
Howard Chu committed
492
	int rc, toggle;
Howard Chu's avatar
Howard Chu committed
493
494
495
496
497
498
499
500
501

	if ((txn = calloc(1, sizeof(*txn))) == NULL) {
		DPRINTF("calloc: %s", strerror(errno));
		return ENOMEM;
	}

	if (rdonly) {
		txn->mt_flags |= MDB_TXN_RDONLY;
	} else {
Howard Chu's avatar
Howard Chu committed
502
503
		txn->mt_u.dirty_queue = calloc(1, sizeof(*txn->mt_u.dirty_queue));
		if (txn->mt_u.dirty_queue == NULL) {
Howard Chu's avatar
Howard Chu committed
504
505
506
			free(txn);
			return ENOMEM;
		}
Howard Chu's avatar
Howard Chu committed
507
		SIMPLEQ_INIT(txn->mt_u.dirty_queue);
Howard Chu's avatar
Howard Chu committed
508

Howard Chu's avatar
Howard Chu committed
509
510
		pthread_mutex_lock(&env->me_txns->mt_wmutex);
		env->me_txns->mt_txnid++;
Howard Chu's avatar
Howard Chu committed
511
512
513
514
515
516
517
		txn->mt_free_pgs = malloc(MDB_IDL_UM_SIZEOF);
		if (txn->mt_free_pgs == NULL) {
			free(txn->mt_u.dirty_queue);
			free(txn);
			return ENOMEM;
		}
		txn->mt_free_pgs[0] = 0;
Howard Chu's avatar
Howard Chu committed
518
519
520
521
522
	}
	txn->mt_txnid = env->me_txns->mt_txnid;
	if (rdonly) {
		MDB_reader *r = pthread_getspecific(env->me_txkey);
		if (!r) {
Howard Chu's avatar
Howard Chu committed
523
			unsigned int i;
Howard Chu's avatar
Howard Chu committed
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
			pthread_mutex_lock(&env->me_txns->mt_mutex);
			for (i=0; i<env->me_maxreaders; i++) {
				if (env->me_txns->mt_readers[i].mr_pid == 0) {
					env->me_txns->mt_readers[i].mr_pid = getpid();
					env->me_txns->mt_readers[i].mr_tid = pthread_self();
					pthread_setspecific(env->me_txkey, &env->me_txns->mt_readers[i]);
					if (i >= env->me_txns->mt_numreaders)
						env->me_txns->mt_numreaders = i+1;
					break;
				}
			}
			pthread_mutex_unlock(&env->me_txns->mt_mutex);
			if (i == env->me_maxreaders) {
				return ENOSPC;
			}
Howard Chu's avatar
Howard Chu committed
539
		}
Howard Chu's avatar
Howard Chu committed
540
541
542
		r->mr_txnid = txn->mt_txnid;
		txn->mt_u.reader = r;
	} else {
Howard Chu's avatar
Howard Chu committed
543
544
545
546
547
		env->me_txn = txn;
	}

	txn->mt_env = env;

Howard Chu's avatar
Howard Chu committed
548
	if ((rc = mdbenv_read_meta(env, &toggle)) != MDB_SUCCESS) {
Howard Chu's avatar
Howard Chu committed
549
550
551
552
		mdb_txn_abort(txn);
		return rc;
	}

Howard Chu's avatar
Howard Chu committed
553
554
555
	if (toggle)
		txn->mt_flags |= MDB_TXN_METOGGLE;

556
	txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
Howard Chu's avatar
Howard Chu committed
557
	txn->mt_root = env->me_meta.mm_root;
558
559
	DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
		txn->mt_txnid, (void *) env, txn->mt_root);
Howard Chu's avatar
Howard Chu committed
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574

	*ret = txn;
	return MDB_SUCCESS;
}

void
mdb_txn_abort(MDB_txn *txn)
{
	MDB_dpage *dp;
	MDB_env	*env;

	if (txn == NULL)
		return;

	env = txn->mt_env;
575
576
	DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
		txn->mt_txnid, (void *) env, txn->mt_root);
Howard Chu's avatar
Howard Chu committed
577

Howard Chu's avatar
Howard Chu committed
578
579
580
	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
		txn->mt_u.reader->mr_txnid = 0;
	} else {
Howard Chu's avatar
Howard Chu committed
581
582
		/* Discard all dirty pages. Return any re-used pages
		 * to the free list.
Howard Chu's avatar
Howard Chu committed
583
		 */
Howard Chu's avatar
Howard Chu committed
584
		MDB_IDL_ZERO(txn->mt_free_pgs);
Howard Chu's avatar
Howard Chu committed
585
586
587
		while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
			dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
			SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
Howard Chu's avatar
Howard Chu committed
588
589
			if (dp->p.mp_pgno <= env->me_meta.mm_last_pg)
				mdb_idl_insert(txn->mt_free_pgs, dp->p.mp_pgno);
Howard Chu's avatar
Howard Chu committed
590
591
			free(dp);
		}
Howard Chu's avatar
Howard Chu committed
592
593
594
595
596
597
598
599
600
601
602
603
		/* put back to head of free list */
		if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
			MDB_oldpages *mop;

			mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(txn->mt_free_pgs) - sizeof(pgno_t));
			mop->mo_next = env->me_pghead;
			mop->mo_txnid = txn->mt_oldest - 1;
			if (!env->me_pghead) {
				env->me_pgtail = mop;
			}
			env->me_pghead = mop;
			bcopy(txn->mt_free_pgs, mop->mo_pages, MDB_IDL_SIZEOF(txn->mt_free_pgs));
Howard Chu's avatar
Howard Chu committed
604
		}
Howard Chu's avatar
Howard Chu committed
605
606

		free(txn->mt_free_pgs);
Howard Chu's avatar
Howard Chu committed
607
608
		free(txn->mt_u.dirty_queue);
		env->me_txn = NULL;
Howard Chu's avatar
Howard Chu committed
609
		env->me_txns->mt_txnid--;
Howard Chu's avatar
Howard Chu committed
610
		pthread_mutex_unlock(&env->me_txns->mt_wmutex);
Howard Chu's avatar
Howard Chu committed
611
612
613
614
615
616
617
618
619
620
621
622
623
	}

	free(txn);
}

int
mdb_txn_commit(MDB_txn *txn)
{
	int		 n, done;
	ssize_t		 rc;
	off_t		 size;
	MDB_dpage	*dp;
	MDB_env	*env;
624
	pgno_t	next;
Howard Chu's avatar
Howard Chu committed
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
	struct iovec	 iov[MDB_COMMIT_PAGES];

	assert(txn != NULL);
	assert(txn->mt_env != NULL);

	env = txn->mt_env;

	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
		DPRINTF("attempt to commit read-only transaction");
		mdb_txn_abort(txn);
		return EPERM;
	}

	if (txn != env->me_txn) {
		DPRINTF("attempt to commit unknown transaction");
		mdb_txn_abort(txn);
		return EINVAL;
	}

	if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
		DPRINTF("error flag is set, can't commit");
		mdb_txn_abort(txn);
		return EINVAL;
	}

Howard Chu's avatar
Howard Chu committed
650
	if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
Howard Chu's avatar
Howard Chu committed
651
652
		goto done;

Howard Chu's avatar
Howard Chu committed
653
	DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
654
	    txn->mt_txnid, (void *) env, txn->mt_root);
Howard Chu's avatar
Howard Chu committed
655
656
657

	/* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
	 */
658
	next = 0;
Howard Chu's avatar
Howard Chu committed
659
660
661
	do {
		n = 0;
		done = 1;
662
		size = 0;
Howard Chu's avatar
Howard Chu committed
663
		SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
664
			if (dp->p.mp_pgno != next) {
Howard Chu's avatar
Howard Chu committed
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
				if (n) {
					DPRINTF("committing %u dirty pages", n);
					rc = writev(env->me_fd, iov, n);
					if (rc != size) {
						n = errno;
						if (rc > 0)
							DPRINTF("short write, filesystem full?");
						else
							DPRINTF("writev: %s", strerror(errno));
						mdb_txn_abort(txn);
						return n;
					}
					n = 0;
					size = 0;
				}
Howard Chu's avatar
Howard Chu committed
680
				lseek(env->me_fd, dp->p.mp_pgno * env->me_meta.mm_psize, SEEK_SET);
681
682
				next = dp->p.mp_pgno;
			}
Howard Chu's avatar
Howard Chu committed
683
			DPRINTF("committing page %lu", dp->p.mp_pgno);
Howard Chu's avatar
Howard Chu committed
684
			iov[n].iov_len = env->me_meta.mm_psize * dp->h.md_num;
Howard Chu's avatar
Howard Chu committed
685
			iov[n].iov_base = &dp->p;
686
687
			size += iov[n].iov_len;
			next = dp->p.mp_pgno + dp->h.md_num;
Howard Chu's avatar
Howard Chu committed
688
689
690
691
692
693
694
695
696
697
698
699
700
			/* clear dirty flag */
			dp->p.mp_flags &= ~P_DIRTY;
			if (++n >= MDB_COMMIT_PAGES) {
				done = 0;
				break;
			}
		}

		if (n == 0)
			break;

		DPRINTF("committing %u dirty pages", n);
		rc = writev(env->me_fd, iov, n);
701
		if (rc != size) {
Howard Chu's avatar
Howard Chu committed
702
703
704
705
706
707
708
709
710
711
712
			n = errno;
			if (rc > 0)
				DPRINTF("short write, filesystem full?");
			else
				DPRINTF("writev: %s", strerror(errno));
			mdb_txn_abort(txn);
			return n;
		}

	} while (!done);

Howard Chu's avatar
Howard Chu committed
713
714
715
716
717
718
719
720
	/* Drop the dirty pages.
	 */
	while (!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
		dp = SIMPLEQ_FIRST(txn->mt_u.dirty_queue);
		SIMPLEQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
		free(dp);
	}

Howard Chu's avatar
Howard Chu committed
721
	if ((n = mdbenv_sync(env)) != 0 ||
722
	    (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
Howard Chu's avatar
Howard Chu committed
723
724
725
726
727
	    (n = mdbenv_sync(env)) != 0) {
		mdb_txn_abort(txn);
		return n;
	}
	env->me_txn = NULL;
Howard Chu's avatar
Howard Chu committed
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744

	/* add to tail of free list */
	if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
		MDB_oldpages *mop;

		mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(txn->mt_free_pgs) - sizeof(pgno_t));
		mop->mo_next = NULL;
		if (env->me_pghead) {
			env->me_pgtail->mo_next = mop;
		} else {
			env->me_pghead = mop;
		}
		env->me_pgtail = mop;
		bcopy(txn->mt_free_pgs, mop->mo_pages, MDB_IDL_SIZEOF(txn->mt_free_pgs));
		mop->mo_txnid = txn->mt_txnid;
	}

Howard Chu's avatar
Howard Chu committed
745
	pthread_mutex_unlock(&env->me_txns->mt_wmutex);
Howard Chu's avatar
Howard Chu committed
746
	free(txn->mt_free_pgs);
747
748
	free(txn->mt_u.dirty_queue);
	free(txn);
Howard Chu's avatar
Howard Chu committed
749
	txn = NULL;
Howard Chu's avatar
Howard Chu committed
750
751
752
753
754
755
756
757
758
759
760
761

done:
	mdb_txn_abort(txn);

	return MDB_SUCCESS;
}

static int
mdbenv_read_header(MDB_env *env)
{
	char		 page[PAGESIZE];
	MDB_page	*p;
Howard Chu's avatar
Howard Chu committed
762
	MDB_meta	*m;
Howard Chu's avatar
Howard Chu committed
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
	int		 rc;

	assert(env != NULL);

	/* We don't know the page size yet, so use a minimum value.
	 */

	if ((rc = pread(env->me_fd, page, PAGESIZE, 0)) == 0) {
		return ENOENT;
	} else if (rc != PAGESIZE) {
		if (rc > 0)
			errno = EINVAL;
		DPRINTF("read: %s", strerror(errno));
		return errno;
	}

	p = (MDB_page *)page;

Howard Chu's avatar
Howard Chu committed
781
782
	if (!F_ISSET(p->mp_flags, P_META)) {
		DPRINTF("page %lu not a meta page", p->mp_pgno);
Howard Chu's avatar
Howard Chu committed
783
784
785
		return EINVAL;
	}

Howard Chu's avatar
Howard Chu committed
786
787
788
	m = METADATA(p);
	if (m->mm_magic != MDB_MAGIC) {
		DPRINTF("meta has invalid magic");
Howard Chu's avatar
Howard Chu committed
789
790
791
		return EINVAL;
	}

Howard Chu's avatar
Howard Chu committed
792
	if (m->mm_version != MDB_VERSION) {
Howard Chu's avatar
Howard Chu committed
793
		DPRINTF("database is version %u, expected version %u",
Howard Chu's avatar
Howard Chu committed
794
		    m->mm_version, MDB_VERSION);
Howard Chu's avatar
Howard Chu committed
795
796
797
		return EINVAL;
	}

Howard Chu's avatar
Howard Chu committed
798
	bcopy(m, &env->me_meta, sizeof(*m));
Howard Chu's avatar
Howard Chu committed
799
800
801
802
	return 0;
}

static int
803
mdbenv_init_meta(MDB_env *env)
Howard Chu's avatar
Howard Chu committed
804
{
805
806
807
	MDB_page *p, *q;
	MDB_meta *meta;
	int rc;
Howard Chu's avatar
Howard Chu committed
808
	unsigned int	 psize;
Howard Chu's avatar
Howard Chu committed
809

Howard Chu's avatar
Howard Chu committed
810
811
812
813
814
	DPRINTF("writing new meta page");
	psize = sysconf(_SC_PAGE_SIZE);

	env->me_meta.mm_magic = MDB_MAGIC;
	env->me_meta.mm_version = MDB_VERSION;
Howard Chu's avatar
Howard Chu committed
815
816
	env->me_meta.mm_psize = psize;
	env->me_meta.mm_flags = env->me_flags & 0xffff;
Howard Chu's avatar
Howard Chu committed
817
818
819
820
821
	env->me_meta.mm_root = P_INVALID;
	env->me_meta.mm_last_pg = 1;

	p = calloc(2, psize);
	p->mp_pgno = 0;
822
	p->mp_flags = P_META;
Howard Chu's avatar
Howard Chu committed
823

824
	meta = METADATA(p);
Howard Chu's avatar
Howard Chu committed
825
	bcopy(&env->me_meta, meta, sizeof(*meta));
Howard Chu's avatar
Howard Chu committed
826

Howard Chu's avatar
Howard Chu committed
827
	q = (MDB_page *)((char *)p + psize);
Howard Chu's avatar
Howard Chu committed
828

Howard Chu's avatar
Howard Chu committed
829
	q->mp_pgno = 1;
830
	q->mp_flags = P_META;
Howard Chu's avatar
Howard Chu committed
831

832
	meta = METADATA(q);
Howard Chu's avatar
Howard Chu committed
833
	bcopy(&env->me_meta, meta, sizeof(*meta));
Howard Chu's avatar
Howard Chu committed
834

Howard Chu's avatar
Howard Chu committed
835
	rc = write(env->me_fd, p, psize * 2);
836
	free(p);
Howard Chu's avatar
Howard Chu committed
837
	return (rc == (int)psize * 2) ? MDB_SUCCESS : errno;
838
839
840
841
842
843
844
845
}

static int
mdbenv_write_meta(MDB_txn *txn)
{
	MDB_env *env;
	MDB_meta	meta;
	off_t off;
Howard Chu's avatar
Howard Chu committed
846
847
	int rc, len;
	char *ptr;
Howard Chu's avatar
Howard Chu committed
848

849
850
851
852
853
854
855
	assert(txn != NULL);
	assert(txn->mt_env != NULL);

	DPRINTF("writing meta page for root page %lu", txn->mt_root);

	env = txn->mt_env;

Howard Chu's avatar
Howard Chu committed
856
857
858
859
860
861
862
863
864
865
	ptr = (char *)&meta;
	off = offsetof(MDB_meta, mm_depth);
	len = sizeof(MDB_meta) - off;

	ptr += off;
	meta.mm_depth = env->me_db.md_depth;
	meta.mm_branch_pages = env->me_db.md_branch_pages;
	meta.mm_leaf_pages = env->me_db.md_leaf_pages;
	meta.mm_overflow_pages = env->me_db.md_overflow_pages;
	meta.mm_entries = env->me_db.md_entries;
866
867
868
869
	meta.mm_root = txn->mt_root;
	meta.mm_last_pg = txn->mt_next_pgno - 1;
	meta.mm_txnid = txn->mt_txnid;

Howard Chu's avatar
Howard Chu committed
870
	if (!F_ISSET(txn->mt_flags, MDB_TXN_METOGGLE))
Howard Chu's avatar
Howard Chu committed
871
		off += env->me_meta.mm_psize;
872
873
874
	off += PAGEHDRSZ;

	lseek(env->me_fd, off, SEEK_SET);
Howard Chu's avatar
Howard Chu committed
875
876
	rc = write(env->me_fd, ptr, len);
	if (rc != len) {
877
878
		DPRINTF("write failed, disk error?");
		return errno;
Howard Chu's avatar
Howard Chu committed
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
	}

	return MDB_SUCCESS;
}

/* Returns true if page p is a valid meta page, false otherwise.
 */
static int
mdb_check_meta_page(MDB_page *p)
{
	if (!F_ISSET(p->mp_flags, P_META)) {
		DPRINTF("page %lu not a meta page", p->mp_pgno);
		return EINVAL;
	}

	return 0;
}

static int
Howard Chu's avatar
Howard Chu committed
898
mdbenv_read_meta(MDB_env *env, int *which)
Howard Chu's avatar
Howard Chu committed
899
{
900
901
902
	MDB_page	*mp0, *mp1;
	MDB_meta	*meta[2];
	int toggle = 0, rc;
Howard Chu's avatar
Howard Chu committed
903
904
905

	assert(env != NULL);

Howard Chu's avatar
Howard Chu committed
906
907
	if ((mp0 = mdbenv_get_page(env, 0)) == NULL ||
		(mp1 = mdbenv_get_page(env, 1)) == NULL)
908
		return EIO;
Howard Chu's avatar
Howard Chu committed
909

910
911
	rc = mdb_check_meta_page(mp0);
	if (rc) return rc;
Howard Chu's avatar
Howard Chu committed
912

913
914
	rc = mdb_check_meta_page(mp1);
	if (rc) return rc;
Howard Chu's avatar
Howard Chu committed
915

916
917
	meta[0] = METADATA(mp0);
	meta[1] = METADATA(mp1);
Howard Chu's avatar
Howard Chu committed
918

919
920
	if (meta[0]->mm_txnid < meta[1]->mm_txnid)
		toggle = 1;
Howard Chu's avatar
Howard Chu committed
921

Howard Chu's avatar
Howard Chu committed
922
923
	if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
		bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
Howard Chu's avatar
Howard Chu committed
924
925
		if (which)
			*which = toggle;
Howard Chu's avatar
Howard Chu committed
926
	}
Howard Chu's avatar
Howard Chu committed
927

928
	DPRINTF("Using meta page %d", toggle);
Howard Chu's avatar
Howard Chu committed
929

930
	return MDB_SUCCESS;
Howard Chu's avatar
Howard Chu committed
931
932
933
}

int
934
mdbenv_create(MDB_env **env)
Howard Chu's avatar
Howard Chu committed
935
936
937
938
939
940
{
	MDB_env *e;

	e = calloc(1, sizeof(*e));
	if (!e) return ENOMEM;

Howard Chu's avatar
Howard Chu committed
941
	e->me_meta.mm_mapsize = DEFAULT_MAPSIZE;
Howard Chu's avatar
Howard Chu committed
942
	e->me_maxreaders = DEFAULT_READERS;
Howard Chu's avatar
Howard Chu committed
943
	e->me_db.md_env = e;
Howard Chu's avatar
Howard Chu committed
944
945
	e->me_fd = -1;
	e->me_lfd = -1;
Howard Chu's avatar
Howard Chu committed
946
947
948
949
	*env = e;
	return MDB_SUCCESS;
}

Howard Chu's avatar
Howard Chu committed
950
951
952
953
954
int
mdbenv_set_mapsize(MDB_env *env, size_t size)
{
	if (env->me_map)
		return EINVAL;
Howard Chu's avatar
Howard Chu committed
955
	env->me_mapsize = env->me_meta.mm_mapsize = size;
Howard Chu's avatar
Howard Chu committed
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
	return MDB_SUCCESS;
}

int
mdbenv_set_maxreaders(MDB_env *env, int readers)
{
	env->me_maxreaders = readers;
	return MDB_SUCCESS;
}

int
mdbenv_get_maxreaders(MDB_env *env, int *readers)
{
	if (!env || !readers)
		return EINVAL;
	*readers = env->me_maxreaders;
	return MDB_SUCCESS;
}

Howard Chu's avatar
Howard Chu committed
975
976
977
978
979
980
981
982
983
984
985
986
987
988
int
mdbenv_open2(MDB_env *env, unsigned int flags)
{
	int i, newenv = 0;

	env->me_flags = flags;

	if ((i = mdbenv_read_header(env)) != 0) {
		if (i != ENOENT)
			return i;
		DPRINTF("new mdbenv");
		newenv = 1;
	}

989
	if (!env->me_mapsize)
Howard Chu's avatar
Howard Chu committed
990
		env->me_mapsize = env->me_meta.mm_mapsize;
991

Howard Chu's avatar
Howard Chu committed
992
	i = MAP_SHARED;
Howard Chu's avatar
Howard Chu committed
993
	if (env->me_meta.mm_address && (flags & MDB_FIXEDMAP))
Howard Chu's avatar
Howard Chu committed
994
		i |= MAP_FIXED;
Howard Chu's avatar
Howard Chu committed
995
	env->me_map = mmap(env->me_meta.mm_address, env->me_mapsize, PROT_READ, i,
Howard Chu's avatar
Howard Chu committed
996
997
998
999
1000
		env->me_fd, 0);
	if (env->me_map == MAP_FAILED)
		return errno;

	if (newenv) {
Howard Chu's avatar
Howard Chu committed
1001
		env->me_meta.mm_mapsize = env->me_mapsize;
Howard Chu's avatar
Howard Chu committed
1002
		if (flags & MDB_FIXEDMAP)
Howard Chu's avatar
Howard Chu committed
1003
			env->me_meta.mm_address = env->me_map;
1004
1005
1006
1007
1008
		i = mdbenv_init_meta(env);
		if (i != MDB_SUCCESS) {
			munmap(env->me_map, env->me_mapsize);
			return i;
		}
Howard Chu's avatar
Howard Chu committed
1009
1010
	}

Howard Chu's avatar
Howard Chu committed
1011
	if ((i = mdbenv_read_meta(env, NULL)) != 0)
Howard Chu's avatar
Howard Chu committed
1012
1013
1014
		return i;

	DPRINTF("opened database version %u, pagesize %u",
Howard Chu's avatar
Howard Chu committed
1015
1016
1017
1018
1019
1020
	    env->me_meta.mm_version, env->me_meta.mm_psize);
	DPRINTF("depth: %u", env->me_meta.mm_depth);
	DPRINTF("entries: %lu", env->me_meta.mm_entries);
	DPRINTF("branch pages: %lu", env->me_meta.mm_branch_pages);
	DPRINTF("leaf pages: %lu", env->me_meta.mm_leaf_pages);
	DPRINTF("overflow pages: %lu", env->me_meta.mm_overflow_pages);
Howard Chu's avatar
Howard Chu committed
1021
1022
1023
1024
1025
	DPRINTF("root: %lu", env->me_meta.mm_root);

	return MDB_SUCCESS;
}

Howard Chu's avatar
Howard Chu committed
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
static void
mdbenv_reader_dest(void *ptr)
{
	MDB_reader *reader = ptr;

	reader->mr_txnid = 0;
	reader->mr_pid = 0;
	reader->mr_tid = 0;
}

Howard Chu's avatar
Howard Chu committed
1036
1037
static void
mdbenv_share_locks(MDB_env *env)
Howard Chu's avatar
Howard Chu committed
1038
{
Howard Chu's avatar
Howard Chu committed
1039
	struct flock lock_info;
1040

Howard Chu's avatar
Howard Chu committed
1041
	env->me_txns->mt_txnid = env->me_meta.mm_txnid;
1042

Howard Chu's avatar
Howard Chu committed
1043
1044
1045
1046
1047
1048
1049
	memset((void *)&lock_info, 0, sizeof(lock_info));
	lock_info.l_type = F_RDLCK;
	lock_info.l_whence = SEEK_SET;
	lock_info.l_start = 0;
	lock_info.l_len = 1;
	fcntl(env->me_lfd, F_SETLK, &lock_info);
}
1050

Howard Chu's avatar
Howard Chu committed
1051
1052
1053
1054
1055
1056
static int
mdbenv_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
{
	int rc;
	off_t size, rsize;
	struct flock lock_info;
Howard Chu's avatar
Howard Chu committed
1057

Howard Chu's avatar
Howard Chu committed
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
	*excl = 0;

	if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
		rc = errno;
		return rc;
	}
	/* Try to get exclusive lock. If we succeed, then
	 * nobody is using the lock region and we should initialize it.
	 */
	memset((void *)&lock_info, 0, sizeof(lock_info));
	lock_info.l_type = F_WRLCK;
	lock_info.l_whence = SEEK_SET;
	lock_info.l_start = 0;
	lock_info.l_len = 1;
	rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
	if (rc == 0) {
		*excl = 1;
	} else {
		lock_info.l_type = F_RDLCK;
		rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
		if (rc) {
			rc = errno;
			goto fail;
		}
	}
	size = lseek(env->me_lfd, 0, SEEK_END);
	rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
	if (size < rsize && *excl) {
		if (ftruncate(env->me_lfd, rsize) != 0) {
			rc = errno;
			goto fail;
Howard Chu's avatar
Howard Chu committed
1089
1090
		}
	} else {
Howard Chu's avatar
Howard Chu committed
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
		rsize = size;
		size = rsize - sizeof(MDB_txninfo);
		env->me_maxreaders = size/sizeof(MDB_reader) + 1;
	}
	env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
		env->me_lfd, 0);
	if (env->me_txns == MAP_FAILED) {
		rc = errno;
		goto fail;
	}
	if (*excl) {
Howard Chu's avatar
Howard Chu committed
1102
1103
1104
1105
1106
1107
		pthread_mutexattr_t mattr;

		pthread_mutexattr_init(&mattr);
		pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
		pthread_mutex_init(&env->me_txns->mt_mutex, &mattr);
		pthread_mutex_init(&env->me_txns->mt_wmutex, &mattr);
1108
1109
		env->me_txns->mt_version = MDB_VERSION;
		env->me_txns->mt_magic = MDB_MAGIC;
Howard Chu's avatar
Howard Chu committed
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
		env->me_txns->mt_txnid = 0;
		env->me_txns->mt_numreaders = 0;

	} else {
		if (env->me_txns->mt_magic != MDB_MAGIC) {
			DPRINTF("lock region has invalid magic");
			errno = EINVAL;
		}
		if (env->me_txns->mt_version != MDB_VERSION) {
			DPRINTF("lock region is version %u, expected version %u",
				env->me_txns->mt_version, MDB_VERSION);
			errno = EINVAL;
		}
		if (errno != EACCES && errno != EAGAIN) {
			rc = errno;
			goto fail;
		}
Howard Chu's avatar
Howard Chu committed
1127
	}
Howard Chu's avatar
Howard Chu committed
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
	return MDB_SUCCESS;

fail:
	close(env->me_lfd);
	return rc;

}

int
mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
{
	int		oflags, rc, len, excl;
	char *lpath, *dpath;

	len = strlen(path);
	lpath = malloc(len + sizeof("/lock.mdb") + len + sizeof("/data.db"));
	if (!lpath)
		return ENOMEM;
	dpath = lpath + len + sizeof("/lock.mdb");
	sprintf(lpath, "%s/lock.mdb", path);
	sprintf(dpath, "%s/data.mdb", path);

	rc = mdbenv_setup_locks(env, lpath, mode, &excl);
	if (rc)
		goto leave;

	if (F_ISSET(flags, MDB_RDONLY))
		oflags = O_RDONLY;
	else
		oflags = O_RDWR | O_CREAT;

	if ((env->me_fd = open(dpath, oflags, mode)) == -1)
		return errno;
Howard Chu's avatar
Howard Chu committed
1161
1162
1163
1164
1165
1166

	if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
		close(env->me_fd);
		env->me_fd = -1;
	} else {
		env->me_path = strdup(path);
1167
		DPRINTF("opened dbenv %p", (void *) env);
Howard Chu's avatar
Howard Chu committed
1168
1169
1170
		pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
		if (excl)
			mdbenv_share_locks(env);
Howard Chu's avatar
Howard Chu committed
1171
1172
	}

Howard Chu's avatar
Howard Chu committed
1173

Howard Chu's avatar
Howard Chu committed
1174
1175
leave:
	free(lpath);
Howard Chu's avatar
Howard Chu committed
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
	return rc;
}

void
mdbenv_close(MDB_env *env)
{
	if (env == NULL)
		return;

	free(env->me_path);

	if (env->me_map) {
		munmap(env->me_map, env->me_mapsize);
	}
	close(env->me_fd);
Howard Chu's avatar
Howard Chu committed
1191
	if (env->me_txns) {
Howard Chu's avatar
Howard Chu committed
1192
1193
		size_t size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
		munmap(env->me_txns, size);
Howard Chu's avatar
Howard Chu committed
1194
	}
Howard Chu's avatar
Howard Chu committed
1195
	close(env->me_lfd);
Howard Chu's avatar
Howard Chu committed
1196
	free(env);
Howard Chu's avatar
Howard Chu committed
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
}

/* Search for key within a leaf page, using binary search.
 * Returns the smallest entry larger or equal to the key.
 * If exactp is non-null, stores whether the found entry was an exact match
 * in *exactp (1 or 0).
 * If kip is non-null, stores the index of the found entry in *kip.
 * If no entry larger or equal to the key is found, returns NULL.
 */
static MDB_node *
mdb_search_node(MDB_db *bt, MDB_page *mp, MDB_val *key,
    int *exactp, unsigned int *kip)
{
	unsigned int	 i = 0;
	int		 low, high;
	int		 rc = 0;
	MDB_node	*node;
	MDB_val	 nodekey;

1216
	DPRINTF("searching %u keys in %s page %lu",
Howard Chu's avatar
Howard Chu committed
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
	    NUMKEYS(mp),
	    IS_LEAF(mp) ? "leaf" : "branch",
	    mp->mp_pgno);

	assert(NUMKEYS(mp) > 0);

	bzero(&nodekey, sizeof(nodekey));

	low = IS_LEAF(mp) ? 0 : 1;
	high = NUMKEYS(mp) - 1;
	while (low <= high) {
		i = (low + high) >> 1;
		node = NODEPTR(mp, i);

		nodekey.mv_size = node->mn_ksize;
		nodekey.mv_data = NODEKEY(node);

		if (bt->md_cmp)
			rc = bt->md_cmp(key, &nodekey);
		else
			rc = _mdb_cmp(bt, key, &nodekey);

		if (IS_LEAF(mp))
			DPRINTF("found leaf index %u [%.*s], rc = %i",
			    i, (int)nodekey.mv_size, (char *)nodekey.mv_data, rc);
		else
			DPRINTF("found branch index %u [%.*s -> %lu], rc = %i",
			    i, (int)node->mn_ksize, (char *)NODEKEY(node),
			    node->mn_pgno, rc);

		if (rc == 0)
			break;
		if (rc > 0)
			low = i + 1;
		else
			high = i - 1;
	}

	if (rc > 0) {	/* Found entry is less than the key. */
		i++;	/* Skip to get the smallest entry larger than key. */
		if (i >= NUMKEYS(mp))
			/* There is no entry larger or equal to the key. */
			return NULL;
	}
	if (exactp)
		*exactp = (rc == 0);
	if (kip)	/* Store the key index if requested. */
		*kip = i;

	return NODEPTR(mp, i);
}

static void
cursor_pop_page(MDB_cursor *cursor)
{
	MDB_ppage	*top;

	top = CURSOR_TOP(cursor);
	CURSOR_POP(cursor);

1277
	DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, (void *) cursor);
Howard Chu's avatar
Howard Chu committed
1278
1279
1280
1281
1282
1283
1284
1285
1286

	free(top);
}

static MDB_ppage *
cursor_push_page(MDB_cursor *cursor, MDB_page *mp)
{
	MDB_ppage	*ppage;

1287
	DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, (void *) cursor);
Howard Chu's avatar
Howard Chu committed
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300

	if ((ppage = calloc(1, sizeof(*ppage))) == NULL)
		return NULL;
	ppage->mp_page = mp;
	CURSOR_PUSH(cursor, ppage);
	return ppage;
}

static MDB_page *
mdbenv_get_page(MDB_env *env, pgno_t pgno)
{
	MDB_page *p = NULL;
	MDB_txn *txn = env->me_txn;
Howard Chu's avatar
Howard Chu committed
1301
	int found = 0;
Howard Chu's avatar
Howard Chu committed
1302

Howard Chu's avatar
Howard Chu committed
1303
	if (txn && !SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
Howard Chu's avatar
Howard Chu committed
1304
		MDB_dpage *dp;
Howard Chu's avatar
Howard Chu committed
1305
		SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
Howard Chu's avatar
Howard Chu committed
1306
1307
			if (dp->p.mp_pgno == pgno) {
				p = &dp->p;
Howard Chu's avatar
Howard Chu committed
1308
				found = 1;
Howard Chu's avatar
Howard Chu committed
1309
1310
1311
				break;
			}
		}
Howard Chu's avatar
Howard Chu committed
1312
1313
	}
	if (!found) {
Howard Chu's avatar
Howard Chu committed
1314
		p = (MDB_page *)(env->me_map + env->me_meta.mm_psize * pgno);
Howard Chu's avatar
Howard Chu committed
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
	}
	return p;
}

static int
mdb_search_page_root(MDB_db *bt, MDB_val *key,
    MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
{
	MDB_page	*mp = mpp->mp_page;
	int rc;

	if (cursor && cursor_push_page(cursor, mp) == NULL)
		return MDB_FAIL;

	while (IS_BRANCH(mp)) {
		unsigned int	 i = 0;
		MDB_node	*node;

1333
		DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
Howard Chu's avatar
Howard Chu committed
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
		assert(NUMKEYS(mp) > 1);
		DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));

		if (key == NULL)	/* Initialize cursor to first page. */
			i = 0;
		else {
			int	 exact;
			node = mdb_search_node(bt, mp, key, &exact, &i);
			if (node == NULL)
				i = NUMKEYS(mp) - 1;
			else if (!exact) {
				assert(i > 0);
				i--;
			}
		}

		if (key)
			DPRINTF("following index %u for key %.*s",
			    i, (int)key->mv_size, (char *)key->mv_data);
1353
		assert(i < NUMKEYS(mp));
Howard Chu's avatar
Howard Chu committed
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
		node = NODEPTR(mp, i);

		if (cursor)
			CURSOR_TOP(cursor)->mp_ki = i;

		mpp->mp_parent = mp;
		if ((mp = mdbenv_get_page(bt->md_env, NODEPGNO(node))) == NULL)
			return MDB_FAIL;
		mpp->mp_pi = i;
		mpp->mp_page = mp;

		if (cursor && cursor_push_page(cursor, mp) == NULL)
			return MDB_FAIL;

1368
1369
		if (modify) {
			MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1370
			if ((rc = mdb_touch(bt->md_env->me_txn, mpp)) != 0)
1371
1372
1373
1374
1375
1376
				return rc;
			dh = ((MDB_dhead *)mpp->mp_page)-1;
			dh->md_parent = mpp->mp_parent;
			dh->md_pi = mpp->mp_pi;
		}

Howard Chu's avatar
Howard Chu committed
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
		mp = mpp->mp_page;
	}

	if (!IS_LEAF(mp)) {
		DPRINTF("internal error, index points to a %02X page!?",
		    mp->mp_flags);
		return MDB_FAIL;
	}

	DPRINTF("found leaf page %lu for key %.*s", mp->mp_pgno,
	    key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL);

	return MDB_SUCCESS;
}

/* Search for the page a given key should be in.
 * Stores a pointer to the found page in *mpp.
 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
 * If cursor is non-null, pushes parent pages on the cursor stack.
 * If modify is true, visited pages are updated with new page numbers.
 */
static int
mdb_search_page(MDB_db *db, MDB_txn *txn, MDB_val *key,
    MDB_cursor *cursor, int modify, MDB_pageparent *mpp)
{
	int		 rc;
	pgno_t		 root;

	/* Can't modify pages outside a transaction. */
	if (txn == NULL && modify) {
		return EINVAL;
	}

	/* Choose which root page to start with. If a transaction is given
	 * use the root page from the transaction, otherwise read the last
	 * committed root page.
	 */
	if (txn == NULL) {
Howard Chu's avatar
Howard Chu committed
1415
		if ((rc = mdbenv_read_meta(db->md_env, NULL)) != MDB_SUCCESS)
Howard Chu's avatar
Howard Chu committed
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
			return rc;
		root = db->md_env->me_meta.mm_root;
	} else if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
		DPRINTF("transaction has failed, must abort");
		return EINVAL;
	} else
		root = txn->mt_root;

	if (root == P_INVALID) {		/* Tree is empty. */
		DPRINTF("tree is empty");
		return ENOENT;
	}

	if ((mpp->mp_page = mdbenv_get_page(db->md_env, root)) == NULL)
		return MDB_FAIL;

	DPRINTF("root page has flags 0x%X", mpp->mp_page->mp_flags);

	if (modify && !F_ISSET(mpp->mp_page->mp_flags, P_DIRTY)) {
		mpp->mp_parent = NULL;
		mpp->mp_pi = 0;
		if ((rc = mdb_touch(txn, mpp)))
			return rc;
		txn->mt_root = mpp->mp_page->mp_pgno;
	}

	return mdb_search_page_root(db, key, cursor, modify, mpp);
}

static int
mdb_read_data(MDB_db *db, MDB_page *mp, MDB_node *leaf,
    MDB_val *data)
{
	MDB_page	*omp;		/* overflow mpage */
	size_t		 max;
	pgno_t		 pgno;

	bzero(data, sizeof(*data));
Howard Chu's avatar
Howard Chu committed
1454
	max = db->md_env->me_meta.mm_psize - PAGEHDRSZ;
Howard Chu's avatar
Howard Chu committed
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545

	if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
		data->mv_size = leaf->mn_dsize;
		data->mv_data = NODEDATA(leaf);
		return MDB_SUCCESS;
	}

	/* Read overflow data.
	 */
	data->mv_size = leaf->mn_dsize;
	bcopy(NODEDATA(leaf), &pgno, sizeof(pgno));
	if ((omp = mdbenv_get_page(db->md_env, pgno)) == NULL) {
		DPRINTF("read overflow page %lu failed", pgno);
		return MDB_FAIL;
	}
	data->mv_data = omp;

	return MDB_SUCCESS;
}

int
mdb_get(MDB_db *db, MDB_txn *txn,
    MDB_val *key, MDB_val *data)
{
	int		 rc, exact;
	MDB_node	*leaf;
	MDB_pageparent mpp;

	assert(key);
	assert(data);
	DPRINTF("===> get key [%.*s]", (int)key->mv_size, (char *)key->mv_data);

	if (db == NULL)
		return EINVAL;

	if (txn != NULL && db->md_env != txn->mt_env) {
		return EINVAL;
	}

	if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
		return EINVAL;
	}

	if ((rc = mdb_search_page(db, txn, key, NULL, 0, &mpp)) != MDB_SUCCESS)
		return rc;

	leaf = mdb_search_node(db, mpp.mp_page, key, &exact, NULL);
	if (leaf && exact)
		rc = mdb_read_data(db, mpp.mp_page, leaf, data);
	else {
		rc = ENOENT;
	}

	return rc;
}

static int
mdb_sibling(MDB_cursor *cursor, int move_right)
{
	int		 rc;
	MDB_node	*indx;
	MDB_ppage	*parent, *top;
	MDB_page	*mp;

	top = CURSOR_TOP(cursor);
	if ((parent = SLIST_NEXT(top, mp_entry)) == NULL) {
		return ENOENT;		/* root has no siblings */
	}

	DPRINTF("parent page is page %lu, index %u",
	    parent->mp_page->mp_pgno, parent->mp_ki);

	cursor_pop_page(cursor);
	if (move_right ? (parent->mp_ki + 1 >= NUMKEYS</