mdb.c 56.7 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
10
11
#include <sys/ipc.h>
#include <sys/shm.h>
Howard Chu's avatar
Howard Chu committed
12
13
14
15
16
17
18
19
20
21
22
23

#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
24
#include <pthread.h>
Howard Chu's avatar
Howard Chu committed
25
26
27

#include "mdb.h"

28
29
30
#ifndef DEBUG
#define DEBUG 1
#endif
Howard Chu's avatar
Howard Chu committed
31

32
33
34
#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
35
#else
36
# define DPRINTF(...)	((void) 0)
Howard Chu's avatar
Howard Chu committed
37
38
39
40
41
42
43
44
#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
45
#define P_INVALID	 (~0UL)
Howard Chu's avatar
Howard Chu committed
46
47
48
49
50
51

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

typedef ulong		 pgno_t;
typedef uint16_t	 indx_t;

Howard Chu's avatar
Howard Chu committed
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
#define DEFAULT_READERS	126
#define DEFAULT_MAPSIZE	1048576

/* Lock descriptor stuff */
#define RXBODY	\
	ulong		mr_txnid; \
	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;	\
	ulong		mt_txnid;	\
	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
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
/* 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 */
#define	P_HEAD		 0x10		/* header page */
#define	P_DIRTY		 0x20		/* dirty page */
	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;

118
#define PAGEHDRSZ	 ((unsigned) offsetof(MDB_page, mp_ptrs))
Howard Chu's avatar
Howard Chu committed
119
120
121

#define NUMKEYS(p)	 (((p)->mp_lower - PAGEHDRSZ) >> 1)
#define SIZELEFT(p)	 (indx_t)((p)->mp_upper - (p)->mp_lower)
122
#define PAGEFILL(env, p) (1000L * ((env)->me_head.mh_psize - PAGEHDRSZ - SIZELEFT(p)) / \
Howard Chu's avatar
Howard Chu committed
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
				((env)->me_head.mh_psize - PAGEHDRSZ))
#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)

typedef struct MDB_head {			/* header page content */
	uint32_t	mh_magic;
	uint32_t	mh_version;
	uint32_t	mh_flags;
	uint32_t	mh_psize;			/* page size */
	void		*mh_address;		/* address for fixed mapping */
	size_t		mh_mapsize;			/* size of mmap region */
} MDB_head;

typedef struct MDB_meta {			/* meta (footer) page content */
	MDB_stat	mm_stat;
139
140
141
	pgno_t		mm_root;			/* page number of root page */
	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
142
143
144
145
146
} MDB_meta;

typedef struct MDB_dhead {					/* a dirty page */
	SIMPLEQ_ENTRY(MDB_dpage)	 md_next;	/* queue of dirty pages */
	MDB_page	*md_parent;
147
	unsigned	md_pi;				/* parent index */
Howard Chu's avatar
Howard Chu committed
148
149
150
151
152
153
154
155
156
157
158
159
160
	int			md_num;
} MDB_dhead;

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

SIMPLEQ_HEAD(dirty_queue, MDB_dpage);

typedef struct MDB_pageparent {
	MDB_page *mp_page;
	MDB_page *mp_parent;
161
	unsigned mp_pi;
Howard Chu's avatar
Howard Chu committed
162
163
} MDB_pageparent;

164
static MDB_dpage *mdb_newpage(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num);
Howard Chu's avatar
Howard Chu committed
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
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
197
198
	unsigned int	mn_flags:4;
	unsigned int	mn_ksize:12;			/* key size */
Howard Chu's avatar
Howard Chu committed
199
200
201
202
203
204
205
206
#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 */
	pgno_t		mt_first_pgno;
Howard Chu's avatar
Howard Chu committed
207
	ulong		mt_txnid;
Howard Chu's avatar
Howard Chu committed
208
	MDB_env		*mt_env;	
Howard Chu's avatar
Howard Chu committed
209
210
211
212
	union {
		struct dirty_queue	*dirty_queue;	/* modified pages */
		MDB_reader	*reader;
	} mt_u;
Howard Chu's avatar
Howard Chu committed
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
#define MDB_TXN_RDONLY		 0x01		/* read-only transaction */
#define MDB_TXN_ERROR		 0x02		/* an error has occurred */
	unsigned int		 mt_flags;
};

/* Must be same as MDB_db, minus md_root/md_stat */
typedef struct MDB_db0 {
	unsigned int	md_flags;
	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;
} MDB_db0;

struct MDB_db {
	unsigned int	md_flags;
	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;
	MDB_stat		md_stat;
234
	pgno_t			md_root;		/* page number of root page */
Howard Chu's avatar
Howard Chu committed
235
236
237
238
};

struct MDB_env {
	int			me_fd;
239
	key_t		me_shmkey;
Howard Chu's avatar
Howard Chu committed
240
	uint32_t	me_flags;
Howard Chu's avatar
Howard Chu committed
241
	int			me_maxreaders;
242
	int			me_metatoggle;
Howard Chu's avatar
Howard Chu committed
243
244
	char		*me_path;
	char *me_map;
Howard Chu's avatar
Howard Chu committed
245
	MDB_txninfo	*me_txns;
Howard Chu's avatar
Howard Chu committed
246
247
248
249
250
251
	MDB_head	me_head;
	MDB_db0		me_db;		/* first DB, overlaps with meta */
	MDB_meta	me_meta;
	MDB_txn	*me_txn;		/* current write transaction */
	size_t		me_mapsize;
	off_t		me_size;		/* current file size */
Howard Chu's avatar
Howard Chu committed
252
	pthread_key_t	me_txkey;	/* thread-key for readers */
Howard Chu's avatar
Howard Chu committed
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
};

#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_write_header(MDB_env *env);
static int  mdbenv_read_header(MDB_env *env);
static int  mdb_check_meta_page(MDB_page *p);
280
281
static int  mdbenv_read_meta(MDB_env *env);
static int  mdbenv_write_meta(MDB_txn *txn);
Howard Chu's avatar
Howard Chu committed
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
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);
static MDB_dpage *mdbenv_new_page(MDB_env *env, uint32_t flags, int num);

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 pgno_t		 mdbenv_compact_tree(MDB_env *env, pgno_t pgno,
			    MDB_env *envc);

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)
{
336
337
338
339
340
	int diff, len_diff = -1;

	if (n1 >= n2) {
		len_diff = (n1 > n2);
		n1 = n2;
Howard Chu's avatar
Howard Chu committed
341
	}
342
343
	diff = memcmp(s1, s2, n1);
	return diff ? diff : len_diff;
Howard Chu's avatar
Howard Chu committed
344
345
346
347
348
}

static int
memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
{
349
	const unsigned char	*p1, *p2, *p1_lim;
Howard Chu's avatar
Howard Chu committed
350
351

	if (n2 == 0)
352
353
354
		return n1 != 0;
	if (n1 == 0)
		return -1;
Howard Chu's avatar
Howard Chu committed
355
356
357
358

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

359
360
361
	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
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
	}
	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 *
383
mdb_newpage(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
Howard Chu's avatar
Howard Chu committed
384
385
386
387
388
389
390
391
{
	MDB_dpage *dp;

	if ((dp = malloc(txn->mt_env->me_head.mh_psize * num + sizeof(MDB_dhead))) == NULL)
		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
392
	SIMPLEQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
Howard Chu's avatar
Howard Chu committed
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
	dp->p.mp_pgno = txn->mt_next_pgno;
	txn->mt_next_pgno += num;

	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;
		DPRINTF("touching page %lu -> %lu", mp->mp_pgno, txn->mt_next_pgno);
		if ((dp = mdb_newpage(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
			return ENOMEM;
		pgno = dp->p.mp_pgno;
		bcopy(mp, &dp->p, txn->mt_env->me_head.mh_psize);
		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;
	int rc;

	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
453
454
		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
455
456
457
			free(txn);
			return ENOMEM;
		}
Howard Chu's avatar
Howard Chu committed
458
		SIMPLEQ_INIT(txn->mt_u.dirty_queue);
Howard Chu's avatar
Howard Chu committed
459

Howard Chu's avatar
Howard Chu committed
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
		pthread_mutex_lock(&env->me_txns->mt_wmutex);
		env->me_txns->mt_txnid++;
	}
	txn->mt_txnid = env->me_txns->mt_txnid;
	if (rdonly) {
		MDB_reader *r = pthread_getspecific(env->me_txkey);
		if (!r) {
			int i;
			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
483
		}
Howard Chu's avatar
Howard Chu committed
484
485
486
		r->mr_txnid = txn->mt_txnid;
		txn->mt_u.reader = r;
	} else {
Howard Chu's avatar
Howard Chu committed
487
488
489
490
491
		env->me_txn = txn;
	}

	txn->mt_env = env;

492
	if ((rc = mdbenv_read_meta(env)) != MDB_SUCCESS) {
Howard Chu's avatar
Howard Chu committed
493
494
495
496
		mdb_txn_abort(txn);
		return rc;
	}

497
	txn->mt_next_pgno = env->me_meta.mm_last_pg+1;
Howard Chu's avatar
Howard Chu committed
498
499
	txn->mt_first_pgno = txn->mt_next_pgno;
	txn->mt_root = env->me_meta.mm_root;
500
501
	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
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516

	*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;
517
518
	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
519

Howard Chu's avatar
Howard Chu committed
520
521
522
	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
		txn->mt_u.reader->mr_txnid = 0;
	} else {
Howard Chu's avatar
Howard Chu committed
523
524
		/* Discard all dirty pages.
		 */
Howard Chu's avatar
Howard Chu committed
525
526
527
		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
528
529
530
531
532
533
534
535
536
537
538
			free(dp);
		}

#if 0
		DPRINTF("releasing write lock on txn %p", txn);
		txn->bt->txn = NULL;
		if (flock(txn->bt->fd, LOCK_UN) != 0) {
			DPRINTF("failed to unlock fd %d: %s",
			    txn->bt->fd, strerror(errno));
		}
#endif
Howard Chu's avatar
Howard Chu committed
539
540
541
		free(txn->mt_u.dirty_queue);
		env->me_txn = NULL;
		pthread_mutex_unlock(&env->me_txns->mt_wmutex);
Howard Chu's avatar
Howard Chu committed
542
543
544
545
546
547
548
549
550
551
552
553
554
	}

	free(txn);
}

int
mdb_txn_commit(MDB_txn *txn)
{
	int		 n, done;
	ssize_t		 rc;
	off_t		 size;
	MDB_dpage	*dp;
	MDB_env	*env;
555
	pgno_t	next;
Howard Chu's avatar
Howard Chu committed
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
	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
581
	if (SIMPLEQ_EMPTY(txn->mt_u.dirty_queue))
Howard Chu's avatar
Howard Chu committed
582
583
		goto done;

Howard Chu's avatar
Howard Chu committed
584
	DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
585
	    txn->mt_txnid, (void *) env, txn->mt_root);
Howard Chu's avatar
Howard Chu committed
586
587
588

	/* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
	 */
589
	next = 0;
Howard Chu's avatar
Howard Chu committed
590
591
592
	do {
		n = 0;
		done = 1;
593
		size = 0;
Howard Chu's avatar
Howard Chu committed
594
		SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
595
596
597
598
599
600
			if (dp->p.mp_pgno != next) {
				lseek(env->me_fd, dp->p.mp_pgno * env->me_head.mh_psize, SEEK_SET);
				next = dp->p.mp_pgno;
				if (n)
					break;
			}
Howard Chu's avatar
Howard Chu committed
601
602
603
			DPRINTF("committing page %lu", dp->p.mp_pgno);
			iov[n].iov_len = env->me_head.mh_psize * dp->h.md_num;
			iov[n].iov_base = &dp->p;
604
605
			size += iov[n].iov_len;
			next = dp->p.mp_pgno + dp->h.md_num;
Howard Chu's avatar
Howard Chu committed
606
607
608
609
610
611
612
613
614
615
616
617
618
			/* 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);
619
		if (rc != size) {
Howard Chu's avatar
Howard Chu committed
620
621
622
623
624
625
626
627
628
629
630
			n = errno;
			if (rc > 0)
				DPRINTF("short write, filesystem full?");
			else
				DPRINTF("writev: %s", strerror(errno));
			mdb_txn_abort(txn);
			return n;
		}

		/* Drop the dirty pages.
		 */
Howard Chu's avatar
Howard Chu committed
631
632
633
		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
634
635
636
637
638
639
640
			free(dp);
			if (--n == 0)
				break;
		}
	} while (!done);

	if ((n = mdbenv_sync(env)) != 0 ||
641
	    (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
Howard Chu's avatar
Howard Chu committed
642
643
644
645
646
	    (n = mdbenv_sync(env)) != 0) {
		mdb_txn_abort(txn);
		return n;
	}
	env->me_txn = NULL;
Howard Chu's avatar
Howard Chu committed
647
	pthread_mutex_unlock(&env->me_txns->mt_wmutex);
648
649
	free(txn->mt_u.dirty_queue);
	free(txn);
Howard Chu's avatar
Howard Chu committed
650
	txn = NULL;
Howard Chu's avatar
Howard Chu committed
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738

done:
	mdb_txn_abort(txn);

	return MDB_SUCCESS;
}

static int
mdbenv_write_header(MDB_env *env)
{
	MDB_head	*h;
	MDB_page	*p;
	ssize_t		 rc;
	unsigned int	 psize;

	DPRINTF("writing header page");
	assert(env != NULL);

	psize = sysconf(_SC_PAGE_SIZE);

	if ((p = calloc(1, psize)) == NULL)
		return ENOMEM;
	p->mp_flags = P_HEAD;

	env->me_head.mh_psize = psize;
	env->me_head.mh_flags = env->me_flags & 0xffff;

	h = METADATA(p);
	bcopy(&env->me_head, h, sizeof(*h));

	rc = write(env->me_fd, p, env->me_head.mh_psize);
	free(p);
	if (rc != (ssize_t)env->me_head.mh_psize) {
		int err = errno;
		if (rc > 0)
			DPRINTF("short write, filesystem full?");
		return err;
	}

	return MDB_SUCCESS;
}

static int
mdbenv_read_header(MDB_env *env)
{
	char		 page[PAGESIZE];
	MDB_page	*p;
	MDB_head	*h;
	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;

	if (!F_ISSET(p->mp_flags, P_HEAD)) {
		DPRINTF("page %lu not a header page", p->mp_pgno);
		return EINVAL;
	}

	h = METADATA(p);
	if (h->mh_magic != MDB_MAGIC) {
		DPRINTF("header has invalid magic");
		return EINVAL;
	}

	if (h->mh_version != MDB_VERSION) {
		DPRINTF("database is version %u, expected version %u",
		    h->mh_version, MDB_VERSION);
		return EINVAL;
	}

	bcopy(h, &env->me_head, sizeof(*h));
	return 0;
}

static int
739
mdbenv_init_meta(MDB_env *env)
Howard Chu's avatar
Howard Chu committed
740
{
741
742
743
	MDB_page *p, *q;
	MDB_meta *meta;
	int rc;
Howard Chu's avatar
Howard Chu committed
744

745
746
747
	p = calloc(2, env->me_head.mh_psize);
	p->mp_pgno = 1;
	p->mp_flags = P_META;
Howard Chu's avatar
Howard Chu committed
748

749
750
751
	meta = METADATA(p);
	meta->mm_root = P_INVALID;
	meta->mm_last_pg = 2;
Howard Chu's avatar
Howard Chu committed
752

753
	q = (MDB_page *)((char *)p + env->me_head.mh_psize);
Howard Chu's avatar
Howard Chu committed
754

755
756
	q->mp_pgno = 2;
	q->mp_flags = P_META;
Howard Chu's avatar
Howard Chu committed
757

758
759
760
	meta = METADATA(q);
	meta->mm_root = P_INVALID;
	meta->mm_last_pg = 2;
Howard Chu's avatar
Howard Chu committed
761

762
763
764
765
766
767
768
769
770
771
772
773
	rc = write(env->me_fd, p, env->me_head.mh_psize * 2);
	free(p);
	return (rc == env->me_head.mh_psize * 2) ? MDB_SUCCESS : errno;
}

static int
mdbenv_write_meta(MDB_txn *txn)
{
	MDB_env *env;
	MDB_meta	meta;
	off_t off;
	int rc;
Howard Chu's avatar
Howard Chu committed
774

775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
	assert(txn != NULL);
	assert(txn->mt_env != NULL);

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

	env = txn->mt_env;

	bcopy(&env->me_meta, &meta, sizeof(meta));
	meta.mm_root = txn->mt_root;
	meta.mm_last_pg = txn->mt_next_pgno - 1;
	meta.mm_txnid = txn->mt_txnid;

	off = env->me_head.mh_psize;
	if (!env->me_metatoggle)
		off *= 2;
	off += PAGEHDRSZ;

	lseek(env->me_fd, off, SEEK_SET);
	rc = write(env->me_fd, &meta, sizeof(meta));
	if (rc != sizeof(meta)) {
		DPRINTF("write failed, disk error?");
		return errno;
Howard Chu's avatar
Howard Chu committed
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
	}

	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
816
mdbenv_read_meta(MDB_env *env)
Howard Chu's avatar
Howard Chu committed
817
{
818
819
820
	MDB_page	*mp0, *mp1;
	MDB_meta	*meta[2];
	int toggle = 0, rc;
Howard Chu's avatar
Howard Chu committed
821
822
823

	assert(env != NULL);

824
825
826
	if ((mp0 = mdbenv_get_page(env, 1)) == NULL ||
		(mp1 = mdbenv_get_page(env, 2)) == NULL)
		return EIO;
Howard Chu's avatar
Howard Chu committed
827

828
829
	rc = mdb_check_meta_page(mp0);
	if (rc) return rc;
Howard Chu's avatar
Howard Chu committed
830

831
832
	rc = mdb_check_meta_page(mp1);
	if (rc) return rc;
Howard Chu's avatar
Howard Chu committed
833

834
835
	meta[0] = METADATA(mp0);
	meta[1] = METADATA(mp1);
Howard Chu's avatar
Howard Chu committed
836

837
838
	if (meta[0]->mm_txnid < meta[1]->mm_txnid)
		toggle = 1;
Howard Chu's avatar
Howard Chu committed
839

Howard Chu's avatar
Howard Chu committed
840
841
842
843
	if (meta[toggle]->mm_txnid > env->me_meta.mm_txnid) {
		bcopy(meta[toggle], &env->me_meta, sizeof(env->me_meta));
		env->me_metatoggle = toggle;
	}
Howard Chu's avatar
Howard Chu committed
844

845
	DPRINTF("Using meta page %d", toggle);
Howard Chu's avatar
Howard Chu committed
846

847
	return MDB_SUCCESS;
Howard Chu's avatar
Howard Chu committed
848
849
850
}

int
851
mdbenv_create(MDB_env **env)
Howard Chu's avatar
Howard Chu committed
852
853
854
855
856
857
858
859
{
	MDB_env *e;

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

	e->me_head.mh_magic = MDB_MAGIC;
	e->me_head.mh_version = MDB_VERSION;
Howard Chu's avatar
Howard Chu committed
860
861
	e->me_head.mh_mapsize = DEFAULT_MAPSIZE;
	e->me_maxreaders = DEFAULT_READERS;
Howard Chu's avatar
Howard Chu committed
862
863
864
865
866
	e->me_db.md_env = e;
	*env = e;
	return MDB_SUCCESS;
}

Howard Chu's avatar
Howard Chu committed
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
int
mdbenv_set_mapsize(MDB_env *env, size_t size)
{
	if (env->me_map)
		return EINVAL;
	env->me_mapsize = env->me_head.mh_mapsize = size;
	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
892
893
894
895
896
897
898
int
mdbenv_open2(MDB_env *env, unsigned int flags)
{
	int i, newenv = 0;

	env->me_flags = flags;
	env->me_meta.mm_root = P_INVALID;
Howard Chu's avatar
Howard Chu committed
899
	env->me_meta.mm_last_pg = 2;
Howard Chu's avatar
Howard Chu committed
900
901
902
903
904
905
906
907

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

908
909
910
	if (!env->me_mapsize)
		env->me_mapsize = env->me_head.mh_mapsize;

Howard Chu's avatar
Howard Chu committed
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
	i = MAP_SHARED;
	if (env->me_head.mh_address && (flags & MDB_FIXEDMAP))
		i |= MAP_FIXED;
	env->me_map = mmap(env->me_head.mh_address, env->me_mapsize, PROT_READ, i,
		env->me_fd, 0);
	if (env->me_map == MAP_FAILED)
		return errno;

	if (newenv) {
		env->me_head.mh_mapsize = env->me_mapsize;
		if (flags & MDB_FIXEDMAP)
			env->me_head.mh_address = env->me_map;
		i = mdbenv_write_header(env);
		if (i != MDB_SUCCESS) {
			munmap(env->me_map, env->me_mapsize);
			return i;
		}
928
929
930
931
932
		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
933
934
	}

935
	if ((i = mdbenv_read_meta(env)) != 0)
Howard Chu's avatar
Howard Chu committed
936
937
938
939
940
941
942
943
944
945
946
947
948
949
		return i;

	DPRINTF("opened database version %u, pagesize %u",
	    env->me_head.mh_version, env->me_head.mh_psize);
	DPRINTF("depth: %u", env->me_meta.mm_stat.ms_depth);
	DPRINTF("entries: %lu", env->me_meta.mm_stat.ms_entries);
	DPRINTF("branch pages: %lu", env->me_meta.mm_stat.ms_branch_pages);
	DPRINTF("leaf pages: %lu", env->me_meta.mm_stat.ms_leaf_pages);
	DPRINTF("overflow pages: %lu", env->me_meta.mm_stat.ms_overflow_pages);
	DPRINTF("root: %lu", env->me_meta.mm_root);

	return MDB_SUCCESS;
}

Howard Chu's avatar
Howard Chu committed
950
951
952
953
954
955
956
957
958
959
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
960
961
962
int
mdbenv_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
{
963
964
965
966
967
968
969
	int		oflags, rc, shmid;
	off_t	size;


	if (F_ISSET(flags, MDB_RDONLY))
		oflags = O_RDONLY;
	else
970
		oflags = O_RDWR | O_CREAT;
971
972

	if ((env->me_fd = open(path, oflags, mode)) == -1)
Howard Chu's avatar
Howard Chu committed
973
974
		return errno;

975
976
977
978
979
980
981
982
983
	env->me_shmkey = ftok(path, 'm');
	size = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
	shmid = shmget(env->me_shmkey, size, IPC_CREAT|IPC_EXCL|mode);
	if (shmid == -1) {
		if (errno == EEXIST) {
			shmid = shmget(env->me_shmkey, size, IPC_CREAT|mode);
			if (shmid == -1)
				return errno;
			env->me_txns = shmat(shmid, NULL, 0);
984
985
986
987
988
989
990
			if (env->me_txns->mt_magic != MDB_MAGIC ||
				env->me_txns->mt_version != MDB_VERSION) {
					DPRINTF("invalid lock region %d", shmid);
					shmdt(env->me_txns);
					env->me_txns = NULL;
					return EIO;
				}
991
992
		} else {
			return errno;
Howard Chu's avatar
Howard Chu committed
993
994
995
996
		}
	} else {
		pthread_mutexattr_t mattr;

997
		env->me_txns = shmat(shmid, NULL, 0);
Howard Chu's avatar
Howard Chu committed
998
999
1000
1001
		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);
1002
1003
		env->me_txns->mt_version = MDB_VERSION;
		env->me_txns->mt_magic = MDB_MAGIC;
Howard Chu's avatar
Howard Chu committed
1004
	}
Howard Chu's avatar
Howard Chu committed
1005
1006
1007
1008
1009
1010

	if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
		close(env->me_fd);
		env->me_fd = -1;
	} else {
		env->me_path = strdup(path);
1011
		DPRINTF("opened dbenv %p", (void *) env);
Howard Chu's avatar
Howard Chu committed
1012
1013
	}

Howard Chu's avatar
Howard Chu committed
1014
1015
	pthread_key_create(&env->me_txkey, mdbenv_reader_dest);

Howard Chu's avatar
Howard Chu committed
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
	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
1031
	if (env->me_txns) {
1032
		shmdt(env->me_txns);
Howard Chu's avatar
Howard Chu committed
1033
1034
	}
	free(env);
Howard Chu's avatar
Howard Chu committed
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
}

/* 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;

1054
	DPRINTF("searching %u keys in %s page %lu",
Howard Chu's avatar
Howard Chu committed
1055
1056
1057
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
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
	    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);

1115
	DPRINTF("popped page %lu off cursor %p", top->mp_page->mp_pgno, (void *) cursor);
Howard Chu's avatar
Howard Chu committed
1116
1117
1118
1119
1120
1121
1122
1123
1124

	free(top);
}

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

1125
	DPRINTF("pushing page %lu on cursor %p", mp->mp_pgno, (void *) cursor);
Howard Chu's avatar
Howard Chu committed
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140

	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;

	if (txn && pgno >= txn->mt_first_pgno &&
Howard Chu's avatar
Howard Chu committed
1141
		!SIMPLEQ_EMPTY(txn->mt_u.dirty_queue)) {
Howard Chu's avatar
Howard Chu committed
1142
		MDB_dpage *dp;
Howard Chu's avatar
Howard Chu committed
1143
		SIMPLEQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
Howard Chu's avatar
Howard Chu committed
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
			if (dp->p.mp_pgno == pgno) {
				p = &dp->p;
				break;
			}
		}
	} else {
		p = (MDB_page *)(env->me_map + env->me_head.mh_psize * pgno);
	}
	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;

1169
		DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
Howard Chu's avatar
Howard Chu committed
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
		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);
1189
		assert(i < NUMKEYS(mp));
Howard Chu's avatar
Howard Chu committed
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
		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;

1204
1205
		if (modify) {
			MDB_dhead *dh = ((MDB_dhead *)mp)-1;
1206
			if ((rc = mdb_touch(bt->md_env->me_txn, mpp)) != 0)
1207
1208
1209
1210
1211
1212
				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
1213
1214
1215
1216
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
		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) {
1251
		if ((rc = mdbenv_read_meta(db->md_env)) != MDB_SUCCESS)
Howard Chu's avatar
Howard Chu committed
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
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
			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));
	max = db->md_env->me_head.mh_psize - PAGEHDRSZ;

	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(parent->mp_page))
		       : (parent->mp_ki == 0)) {
		DPRINTF("no more keys left, moving to %s sibling",
		    move_right ? "right" : "left");
		if ((rc = mdb_sibling(cursor, move_right)) != MDB_SUCCESS)
			return rc;
		parent = CURSOR_TOP(cursor);
	} else {
		if (move_right)
			parent->mp_ki++;
		else
			parent->mp_ki--;
		DPRINTF("just moving to %s index key %u",
		    move_right ? "right" : "left", parent->mp_ki);
	}
	assert(IS_BRANCH(parent->mp_page));

	indx = NODEPTR(parent->mp_page, parent->mp_ki);
1382
	if ((mp = mdbenv_get_page(cursor->mc_db->md_env, indx->mn_pgno)) == NULL)
Howard Chu's avatar
Howard Chu committed
1383
		return MDB_FAIL;
1384
#if 0
Howard Chu's avatar
Howard Chu committed
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
1415
1416
1417
1418
1419
1420
1421
1422
	mp->parent = parent->mp_page;
	mp->parent_index = parent->mp_ki;
#endif

	cursor_push_page(cursor, mp);

	return MDB_SUCCESS;
}

static int
mdb_set_key(MDB_db *bt, MDB_page *mp, MDB_node *node,
    MDB_val *key)
{
	if (key == NULL)
		return 0;

	key->mv_size = node->mn_ksize;
	key->mv_data = NODEKEY(node);

	return 0;
}

static int
mdb_cursor_next(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
{
	MDB_ppage	*top;
	MDB_page	*mp;
	MDB_node	*leaf;

	if (cursor->mc_eof) {
		return ENOENT;
	}

	assert(cursor->mc_initialized);

	top = CURSOR_TOP(cursor);
	mp = top->mp_page;

1423
	DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) cursor);
Howard Chu's avatar
Howard Chu committed
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436

	if (top->mp_ki + 1 >= NUMKEYS(mp)) {
		DPRINTF("=====> move to next sibling page");
		if (mdb_sibling(cursor, 1) != MDB_SUCCESS) {
			cursor->mc_eof = 1;
			return ENOENT;
		}
		top = CURSOR_TOP(cursor);
		mp = top->mp_page;
		DPRINTF("next page is %lu, key index %u", mp->mp_pgno, top->mp_ki);
	} else
		top->mp_ki++;

1437
	DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
Howard Chu's avatar
Howard Chu committed
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
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
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
	    mp->mp_pgno, NUMKEYS(mp), top->mp_ki);

	assert(IS_LEAF(mp));
	leaf = NODEPTR(mp, top->mp_ki);

	if (data && mdb_read_data(cursor->mc_db, mp, leaf, data) != MDB_SUCCESS)
		return MDB_FAIL;

	return mdb_set_key(cursor->mc_db, mp, leaf, key);
}

static int
mdb_cursor_set(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
    int *exactp)
{
	int		 rc;
	MDB_node	*leaf;
	MDB_ppage	*top;
	MDB_pageparent mpp;

	assert(cursor);
	assert(key);
	assert(key->mv_size > 0);

	rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, key, cursor, 0, &mpp);
	if (rc != MDB_SUCCESS)
		return rc;
	assert(IS_LEAF(mpp.mp_page));

	top = CURSOR_TOP(cursor);
	leaf = mdb_search_node(cursor->mc_db, mpp.mp_page, key, exactp, &top->mp_ki);
	if (exactp != NULL && !*exactp) {
		/* MDB_CURSOR_EXACT specified and not an exact match. */
		return ENOENT;
	}

	if (leaf == NULL) {
		DPRINTF("===> inexact leaf not found, goto sibling");
		if ((rc = mdb_sibling(cursor, 1)) != MDB_SUCCESS)
			return rc;		/* no entries matched */
		top = CURSOR_TOP(cursor);
		top->mp_ki = 0;
		mpp.mp_page = top->mp_page;
		assert(IS_LEAF(mpp.mp_page));
		leaf = NODEPTR(mpp.mp_page, 0);
	}

	cursor->mc_initialized = 1;
	cursor->mc_eof = 0;

	if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
		return rc;

	rc = mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
	if (rc == MDB_SUCCESS) {
		DPRINTF("==> cursor placed on key %.*s",
			(int)key->mv_size, (char *)key->mv_data);
		;
	}

	return rc;
}

static int
mdb_cursor_first(MDB_cursor *cursor, MDB_val *key, MDB_val *data)
{
	int		 rc;
	MDB_pageparent	mpp;
	MDB_node	*leaf;

	rc = mdb_search_page(cursor->mc_db, cursor->mc_txn, NULL, cursor, 0, &mpp);
	if (rc != MDB_SUCCESS)
		return rc;
	assert(IS_LEAF(mpp.mp_page));

	leaf = NODEPTR(mpp.mp_page, 0);
	cursor->mc_initialized = 1;
	cursor->mc_eof = 0;

	if (data && (rc = mdb_read_data(cursor->mc_db, mpp.mp_page, leaf, data)) != MDB_SUCCESS)
		return rc;

	return mdb_set_key(cursor->mc_db, mpp.mp_page, leaf, key);
}

int
mdb_cursor_get(MDB_cursor *cursor, MDB_val *key, MDB_val *data,
    MDB_cursor_op op)
{
	int		 rc;
	int		 exact = 0;

	assert(cursor);

	switch (op) {
	case MDB_CURSOR:
	case MDB_CURSOR_EXACT:
		while (CURSOR_TOP(cursor) != NULL)
			cursor_pop_page(cursor);
		if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
			rc = EINVAL;
		} else if (op == MDB_CURSOR_EXACT)
			rc = mdb_cursor_set(cursor, key, data, &exact);
		else
			rc = mdb_cursor_set(cursor, key, data, NULL);
		break;
	case MDB_NEXT:
		if (!cursor->mc_initialized)
			rc = mdb_cursor_first(cursor, key, data);
		else
			rc = mdb_cursor_next(cursor, key, data);
		break;
	case MDB_FIRST:
		while (CURSOR_TOP(cursor) != NULL)
			cursor_pop_page(cursor);
		rc = mdb_cursor_first(cursor, key, data);
		break;
	default:
		DPRINTF("unhandled/unimplemented cursor operation %u", op);
		rc = EINVAL;
		break;
	}

	return rc;
}

/* Allocate a page and initialize it
 */
static MDB_dpage *
mdbenv_new_page(MDB_env *env, uint32_t flags, int num)
{
	MDB_dpage	*dp;

	assert(env != NULL);
	assert(env->me_txn != NULL);

	DPRINTF("allocating new mpage %lu, page size %u",
	    env->me_txn->mt_next_pgno, env->me_head.mh_psize);
	if ((dp = mdb_newpage(env->me_txn, NULL, 0, num)) == NULL)
		return NULL;
1578
	dp->p.mp_flags = flags | P_DIRTY;
Howard Chu's avatar
Howard Chu committed
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
	dp->p.mp_lower = PAGEHDRSZ;
	dp->p.mp_upper = env->me_head.mh_psize;

	if (IS_BRANCH(&dp->p))
		env->me_meta.mm_stat.ms_branch_pages++;
	else if (IS_LEAF(&dp->p))
		env->me_meta.mm_stat.ms_leaf_pages++;
	else if (IS_OVERFLOW(&dp->p)) {
		env->me_meta.mm_stat.ms_overflow_pages += num;
		dp->p.mp_pages = num;
	}

	return dp;
}

static size_t
mdb_leaf_size(MDB_db *db, MDB_val *key, MDB_val *data)
{
	size_t		 sz;

	sz = LEAFSIZE(key, data);
	if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
		/* put on overflow page */
		sz -= data->mv_size - sizeof(pgno_t);
	}

	return sz + sizeof(indx_t);
}

static size_t
mdb_branch_size(MDB_db *db, MDB_val *key)
{
	size_t		 sz;

	sz = INDXSIZE(key);
	if (sz >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
		/* put on overflow page */
		/* not implemented */
		/* sz -= key->size - sizeof(pgno_t); */
	}

	return sz + sizeof(indx_t);
}

static int
mdb_add_node(MDB_db *db, MDB_page *mp, indx_t indx,
    MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
{
	unsigned int	 i;
	size_t		 node_size = NODESIZE;
	indx_t		 ofs;
	MDB_node	*node;
	MDB_dpage	*ofp = NULL;		/* overflow page */

	assert(mp->mp_upper >= mp->mp_lower);

	DPRINTF("add node [%.*s] to %s page %lu at index %i, key size %zu",
	    key ? (int)key->mv_size : 0, key ? (char *)key->mv_data : NULL,
	    IS_LEAF(mp) ? "leaf" : "branch",
	    mp->mp_pgno, indx, key ? key->mv_size : 0);

	if (key != NULL)
		node_size += key->mv_size;

	if (IS_LEAF(mp)) {
		assert(data);
		if (F_ISSET(flags, F_BIGDATA)) {
			/* Data already on overflow page. */
Howard Chu's avatar
Howard Chu committed
1647
			node_size += sizeof(pgno_t);
Howard Chu's avatar
Howard Chu committed
1648
1649
1650
1651
1652
1653
		} else if (data->mv_size >= db->md_env->me_head.mh_psize / MDB_MINKEYS) {
			int ovpages = PAGEHDRSZ + data->mv_size + db->md_env->me_head.mh_psize - 1;
			ovpages /= db->md_env->me_head.mh_psize;
			/* Put data on overflow page. */
			DPRINTF("data size is %zu, put on overflow page",
			    data->mv_size);
Howard Chu's avatar
Howard Chu committed
1654
			node_size += sizeof(pgno_t);
Howard Chu's avatar
Howard Chu committed
1655
1656
1657
1658
			if ((ofp = mdbenv_new_page(db->md_env, P_OVERFLOW, ovpages)) == NULL)
				return MDB_FAIL;
			DPRINTF("allocated overflow page %lu", ofp->p.mp_pgno);
			flags |= F_BIGDATA;
Howard Chu's avatar
Howard Chu committed
1659
1660
		} else {
			node_size += data->mv_size;
Howard Chu's avatar
Howard Chu committed
1661
1662
1663
1664
		}
	}

	if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
1665
		DPRINTF("not enough room in page %lu, got %u ptrs",
Howard Chu's avatar
Howard Chu committed
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
		    mp->mp_pgno, NUMKEYS(mp));
		DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
		    mp->mp_upper - mp->mp_lower);
		DPRINTF("node size = %zu", node_size);
		return ENOSPC;
	}

	/* Move higher pointers up one slot. */
	for (i = NUMKEYS(mp); i > indx; i--)
		mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];

	/* Adjust free space offsets. */
	ofs = mp->mp_upper - node_size;
	assert(ofs >= mp->mp_lower + sizeof(indx_t));
	mp->mp_ptrs[indx] = ofs;
	mp->mp_upper = ofs;
	mp->mp_lower += sizeof(indx_t);

	/* Write the node data. */
	node = NODEPTR(mp, indx);
	node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
	node->mn_flags = flags;
	if (IS_LEAF(mp))
		node->mn_dsize = data->mv_size;
	else
		node->mn_pgno = pgno;

	if (key)
		bcopy(key->mv_data, NODEKEY(node), key->mv_size);

	if (IS_LEAF(mp)) {
		assert(key);
		if (ofp == NULL) {
			if (F_ISSET(flags, F_BIGDATA))
				bcopy(data->mv_data, node->mn_data + key->mv_size,
				    sizeof(pgno_t));
			else
				bcopy(data->mv_data, node->mn_data + key->mv_size,
				    data->mv_size);
		} else {
			bcopy(&ofp->p.mp_pgno, node->mn_data + key->mv_size,
			    sizeof(pgno_t));
			bcopy(data->mv_data, METADATA(&ofp->p), data->mv_size);
		}
	}

	return MDB_SUCCESS;
}

static void
mdb_del_node(MDB_db *db, MDB_page *mp, indx_t indx)
{
	unsigned int	 sz;
	indx_t		 i, j, numkeys, ptr;
	MDB_node	*node;
	char		*base;

	DPRINTF("delete node %u on %s page %lu", indx,
	    IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
	assert(indx < NUMKEYS(mp));

	node = NODEPTR(mp, indx);
	sz = NODESIZE + node->mn_ksize;
	if (IS_LEAF(mp)) {
		if (F_ISSET(node->mn_flags, F_BIGDATA))
			sz += sizeof(pgno_t);
		else
			sz += NODEDSZ(node);
	}

	ptr = mp->mp_ptrs[indx];
	numkeys = NUMKEYS(mp);
	for (i = j = 0; i < numkeys; i++) {
		if (i != indx) {
			mp->mp_ptrs[j] = mp->mp_ptrs[i];
			if (mp->mp_ptrs[i] < ptr)
				mp->mp_ptrs[j] += sz;
			j++;
		}
	}

	base = (char *)mp + mp->mp_upper;
	bcopy(base, base + sz, ptr - mp->mp_upper);

	mp->mp_lower -= sizeof(indx_t);
	mp->mp_upper += sz;
}

int
mdb_cursor_open(MDB_db *db, MDB_txn *txn, MDB_cursor **ret)
{
	MDB_cursor	*cursor;

	if (db == NULL || ret == NULL)
		return EINVAL;

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

	if ((cursor = calloc(1, sizeof(*cursor))) != NULL) {
		SLIST_INIT(&cursor->mc_stack);
		cursor->mc_db = db;
		cursor->mc_txn = txn;
	}

	*ret = cursor;

	return MDB_SUCCESS;
}

void
mdb_cursor_close(MDB_cursor *cursor)
{
	if (cursor != NULL) {
		while (!CURSOR_EMPTY(cursor))
			cursor_pop_page(cursor);

/*		btree_close(cursor->bt); */
		free(cursor);
	}
}

static int
mdb_update_key(MDB_db *db, MDB_page *mp, indx_t indx,
    MDB_val *key)
{
	indx_t			 ptr, i, numkeys;
	int			 delta;
	size_t			 len;
	MDB_node		*node;
	char			*base;

	node = NODEPTR(mp, indx);
	ptr = mp->mp_ptrs[indx];
	DPRINTF("update key %u (ofs %u) [%.*s] to [%.*s] on page %lu",
	    indx, ptr,
	    (int)node->mn_ksize, (char *)NODEKEY(node),
	    (int)key->mv_size, (char *)key->mv_data,
	    mp->mp_pgno);

	delta = key->mv_size - node->mn_ksize;
	if (delta) {
		if (delta > 0 && SIZELEFT(mp) < delta) {
			DPRINTF("OUCH! Not enough room, delta = %d", delta);
			return ENOSPC;
		}

		numkeys = NUMKEYS(mp);
		for (i = 0; i < numkeys; i++) {
			if (mp->mp_ptrs[i] <= ptr)
				mp->mp_ptrs[i] -= delta;
		}

		base = (char *)mp + mp->mp_upper;
		len = ptr - mp->mp_upper + NODESIZE;
		bcopy(base, base - delta, len);
		mp->mp_upper -= delta;

		node = NODEPTR(mp, indx);
		node->mn_ksize = key->mv_size;
	}

	bcopy(key->mv_data, NODEKEY(node), key->mv_size);

	return MDB_SUCCESS;
}

/* Move a node from src to dst.
 */
static int
mdb_move_node(MDB_db *bt, MDB_pageparent *src, indx_t srcindx,
    MDB_pageparent *dst, indx_t dstindx)
{
	int			 rc;
	MDB_node		*srcnode;
	MDB_val		 key, data;

	srcnode = NODEPTR(src->mp_page, srcindx);
	DPRINTF("moving %s node %u [%.*s] on page %lu to node %u on page %lu",
	    IS_LEAF(src->mp_page) ? "leaf" : "branch",
	    srcindx,
	    (int)srcnode->mn_ksize, (char *)NODEKEY(srcnode),
	    src->mp_page->mp_pgno,
	    dstindx, dst->mp_page->mp_pgno);

	/* Mark src and dst as dirty. */
	if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
	    (rc = mdb_touch(bt->md_env->me_txn, dst)))
		return rc;;

	/* Add the node to the destination page.
	 */
	key.mv_size = srcnode->mn_ksize;
	key.mv_data = NODEKEY(srcnode);
	data.mv_size = NODEDSZ(srcnode);
	data.mv_data = NODEDATA(srcnode);
	rc = mdb_add_node(bt, dst->mp_page, dstindx, &key, &data, NODEPGNO(srcnode),
	    srcnode->mn_flags);
	if (rc != MDB_SUCCESS)
		return rc;

	/* Delete the node from the source page.
	 */
	mdb_del_node(bt, src->mp_page, srcindx);

	/* Update the parent separators.
	 */
	if (srcindx == 0 && src->mp_pi != 0) {
		DPRINTF("update separator for source page %lu to [%.*s]",
		    src->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
		if ((rc = mdb_update_key(bt, src->mp_parent, src->mp_pi,
		    &key)) != MDB_SUCCESS)
			return rc;
	}

	if (srcindx == 0 && IS_BRANCH(src->mp_page)) {
		MDB_val	 nullkey;
		nullkey.mv_size = 0;
		assert(mdb_update_key(bt, src->mp_page, 0, &nullkey) == MDB_SUCCESS);
	}

	if (dstindx == 0 && dst->mp_pi != 0) {
		DPRINTF("update separator for destination page %lu to [%.*s]",
		    dst->mp_page->mp_pgno, (int)key.mv_size, (char *)key.mv_data);
		if ((rc = mdb_update_key(bt, dst->mp_parent, dst->mp_pi,
		    &key)) != MDB_SUCCESS)
			return rc;
	}

	if (dstindx == 0 && IS_BRANCH(dst->mp_page)) {
		MDB_val	 nullkey;
		nullkey.mv_size = 0;
		assert(mdb_update_key(bt, dst->mp_page, 0, &nullkey) == MDB_SUCCESS);
	}

	return MDB_SUCCESS;
}

static int
mdb_merge(MDB_db *bt, MDB_pageparent *src, MDB_pageparent *dst)
{
	int			 rc;
	indx_t			 i;
	MDB_node		*srcnode;
	MDB_val		 key, data;
	MDB_pageparent	mpp;
	MDB_dhead *dh;

	DPRINTF("merging page %lu and %lu", src->mp_page->mp_pgno, dst->mp_page->mp_pgno);

	assert(src->mp_parent);	/* can't merge root page */
	assert(dst->mp_parent);
	assert(bt->md_env->me_txn != NULL);

	/* Mark src and dst as dirty. */
	if ((rc = mdb_touch(bt->md_env->me_txn, src)) ||
	    (rc = mdb_touch(bt->md_env->me_txn, dst)))
		return rc;

	/* Move all nodes from src to dst.
	 */
	for (i = 0; i < NUMKEYS(src->mp_page); i++) {
		srcnode = NODEPTR(src->mp_page, i);

		key.mv_size = srcnode->mn_ksize;
		key.mv_data = NODEKEY(srcnode);
		data.mv_size = NODEDSZ(srcnode);
		data.mv_data = NODEDATA(srcnode);
		rc = mdb_add_node(bt, dst->mp_page, NUMKEYS(dst->mp_page), &key,
		    &data, NODEPGNO(srcnode), srcnode->mn_flags);
		if (rc != MDB_SUCCESS)
			return rc;
	}

1941
	DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
Howard Chu's avatar
Howard Chu committed
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
	    dst->mp_page->mp_pgno, NUMKEYS(dst->mp_page), (float)PAGEFILL(bt->md_env, dst->mp_page) / 10);

	/* Unlink the src page from parent.
	 */
	mdb_del_node(bt, src->mp_parent, src->mp_pi);
	if (src->mp_pi == 0) {
		key.mv_size = 0;
		if ((rc = mdb_update_key(bt, src->mp_parent, 0, &key)) != MDB_SUCCESS)
			return rc;
	}

	if (IS_LEAF(src->mp_page))
		bt->md_env->me_meta.mm_stat.ms_leaf_pages--;
	else
		bt->md_env->me_meta.mm_stat.ms_branch_pages--;

	mpp.mp_page = src->mp_parent;
	dh = (MDB_dhead *)src->mp_parent;
	dh--;
	mpp.mp_parent = dh->md_parent;
	mpp.mp_pi = dh->md_pi;

	return mdb_rebalance(bt, &mpp);
}

#define FILL_THRESHOLD	 250

static int
mdb_rebalance(MDB_db *db, MDB_pageparent *mpp)
{
	MDB_node	*node;
	MDB_page	*root;
	MDB_pageparent npp;
	indx_t		 si = 0, di = 0;

	assert(db != NULL);
	assert(db->md_env->me_txn != NULL);
	assert(mpp != NULL);

1981
	DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
Howard Chu's avatar
Howard Chu committed