mdb.c 67.3 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
#include <endian.h>
Howard Chu's avatar
Howard Chu committed
25
26

#include "mdb.h"
Howard Chu's avatar
Howard Chu committed
27

Howard Chu's avatar
Howard Chu committed
28
29
#define ULONG		unsigned long
typedef ULONG		pgno_t;
Howard Chu's avatar
Howard Chu committed
30

Howard Chu's avatar
Howard Chu committed
31
#include "idl.h"
Howard Chu's avatar
Howard Chu committed
32

33
34
35
#ifndef DEBUG
#define DEBUG 1
#endif
Howard Chu's avatar
Howard Chu committed
36

Howard Chu's avatar
Howard Chu committed
37
#if DEBUG && defined(__GNUC__)
38
39
# define DPRINTF(fmt, ...) \
	fprintf(stderr, "%s:%d: " fmt "\n", __func__, __LINE__, ##__VA_ARGS__)
Howard Chu's avatar
Howard Chu committed
40
#else
41
# define DPRINTF(...)	((void) 0)
Howard Chu's avatar
Howard Chu committed
42
43
44
45
46
47
48
49
#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
50
#define P_INVALID	 (~0UL)
Howard Chu's avatar
Howard Chu committed
51
52
53
54
55

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

typedef uint16_t	 indx_t;

Howard Chu's avatar
Howard Chu committed
56
57
58
59
60
#define DEFAULT_READERS	126
#define DEFAULT_MAPSIZE	1048576

/* Lock descriptor stuff */
#define RXBODY	\
Howard Chu's avatar
Howard Chu committed
61
	ULONG		mr_txnid; \
Howard Chu's avatar
Howard Chu committed
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
	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
82
	ULONG		mt_txnid;	\
Howard Chu's avatar
Howard Chu committed
83
84
85
86
87
88
89
90
91
92
93
94
95
	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
96
97
98
99
100
/* 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 */
Howard Chu's avatar
Howard Chu committed
101
102
103
104
105
#define	mp_pgno		mp_p.p_pgno
	union padded {
		pgno_t		p_pgno;		/* page number */
		void *		p_pad;
	} mp_p;
Howard Chu's avatar
Howard Chu committed
106
107
108
109
#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
110
#define	P_DIRTY		 0x10		/* dirty page */
Howard Chu's avatar
Howard Chu committed
111
112
113
114
115
116
117
118
119
120
121
122
123
124
	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;

125
#define PAGEHDRSZ	 ((unsigned) offsetof(MDB_page, mp_ptrs))
Howard Chu's avatar
Howard Chu committed
126
127
128

#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
129
130
#define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
				((env)->me_psize - PAGEHDRSZ))
Howard Chu's avatar
Howard Chu committed
131
132
133
134
#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
135
136
#define OVPAGES(size, psize)	(PAGEHDRSZ + size + psize - 1) / psize;

Howard Chu's avatar
Howard Chu committed
137
138
139
140
141
142
143
144
145
146
147
148
149
150
typedef struct MDB_db {
	uint32_t	md_pad;
	uint16_t	md_flags;
	uint16_t	md_depth;
	ULONG		md_branch_pages;
	ULONG		md_leaf_pages;
	ULONG		md_overflow_pages;
	ULONG		md_entries;
	pgno_t		md_root;
} MDB_db;

#define	FREE_DBI	0
#define	MAIN_DBI	1

Howard Chu's avatar
Howard Chu committed
151
typedef struct MDB_meta {			/* meta (footer) page content */
Howard Chu's avatar
Howard Chu committed
152
153
154
155
	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
156
157
158
	MDB_db		mm_dbs[2];			/* first is free space, 2nd is main db */
#define	mm_psize	mm_dbs[0].md_pad
#define	mm_flags	mm_dbs[0].md_flags
Howard Chu's avatar
Howard Chu committed
159
160
	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
161
162
163
} MDB_meta;

typedef struct MDB_dhead {					/* a dirty page */
Howard Chu's avatar
Howard Chu committed
164
	STAILQ_ENTRY(MDB_dpage)	 md_next;	/* queue of dirty pages */
Howard Chu's avatar
Howard Chu committed
165
	MDB_page	*md_parent;
166
	unsigned	md_pi;				/* parent index */
Howard Chu's avatar
Howard Chu committed
167
168
169
170
171
172
173
174
	int			md_num;
} MDB_dhead;

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

Howard Chu's avatar
Howard Chu committed
175
STAILQ_HEAD(dirty_queue, MDB_dpage);
Howard Chu's avatar
Howard Chu committed
176

Howard Chu's avatar
Howard Chu committed
177
178
typedef struct MDB_oldpages {
	struct MDB_oldpages *mo_next;
Howard Chu's avatar
Howard Chu committed
179
	ULONG		mo_txnid;
Howard Chu's avatar
Howard Chu committed
180
181
182
	pgno_t		mo_pages[1];	/* dynamic */
} MDB_oldpages;

Howard Chu's avatar
Howard Chu committed
183
184
185
typedef struct MDB_pageparent {
	MDB_page *mp_page;
	MDB_page *mp_parent;
186
	unsigned mp_pi;
Howard Chu's avatar
Howard Chu committed
187
188
} MDB_pageparent;

Howard Chu's avatar
Howard Chu committed
189
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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
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_txn		*mc_txn;
	struct page_stack	 mc_stack;		/* stack of parent pages */
Howard Chu's avatar
Howard Chu committed
207
	MDB_dbi		mc_dbi;
Howard Chu's avatar
Howard Chu committed
208
209
210
211
212
213
214
215
216
217
218
219
220
221
	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
222
223
	unsigned int	mn_flags:4;
	unsigned int	mn_ksize:12;			/* key size */
Howard Chu's avatar
Howard Chu committed
224
225
226
227
#define F_BIGDATA	 0x01			/* data put on overflow page */
	char		mn_data[1];
} MDB_node;

Howard Chu's avatar
Howard Chu committed
228

Howard Chu's avatar
Howard Chu committed
229
typedef struct MDB_dbx {
230
	MDB_val		md_name;
Howard Chu's avatar
Howard Chu committed
231
232
233
234
235
	MDB_cmp_func	*md_cmp;		/* user compare function */
	MDB_cmp_func	*md_dcmp;		/* user dupsort function */
	MDB_rel_func	*md_rel;		/* user relocate function */
} MDB_dbx;

Howard Chu's avatar
Howard Chu committed
236
237
struct MDB_txn {
	pgno_t		mt_next_pgno;	/* next unallocated page */
Howard Chu's avatar
Howard Chu committed
238
239
	ULONG		mt_txnid;
	ULONG		mt_oldest;
Howard Chu's avatar
Howard Chu committed
240
	MDB_env		*mt_env;	
Howard Chu's avatar
Howard Chu committed
241
	pgno_t		*mt_free_pgs;	/* this is an IDL */
Howard Chu's avatar
Howard Chu committed
242
243
244
245
	union {
		struct dirty_queue	*dirty_queue;	/* modified pages */
		MDB_reader	*reader;
	} mt_u;
Howard Chu's avatar
Howard Chu committed
246
	MDB_dbx		*mt_dbxs;		/* array */
Howard Chu's avatar
Howard Chu committed
247
	MDB_db		*mt_dbs;
Howard Chu's avatar
Howard Chu committed
248
249
	unsigned int	mt_numdbs;

Howard Chu's avatar
Howard Chu committed
250
251
#define MDB_TXN_RDONLY		 0x01		/* read-only transaction */
#define MDB_TXN_ERROR		 0x02		/* an error has occurred */
Howard Chu's avatar
Howard Chu committed
252
#define MDB_TXN_METOGGLE	0x04		/* used meta page 1 */
Howard Chu's avatar
Howard Chu committed
253
254
255
256
257
	unsigned int		 mt_flags;
};

struct MDB_env {
	int			me_fd;
Howard Chu's avatar
Howard Chu committed
258
	int			me_lfd;
Howard Chu's avatar
Howard Chu committed
259
	uint32_t	me_flags;
Howard Chu's avatar
Howard Chu committed
260
261
262
	unsigned int	me_maxreaders;
	unsigned int	me_numdbs;
	unsigned int	me_maxdbs;
Howard Chu's avatar
Howard Chu committed
263
	char		*me_path;
Howard Chu's avatar
Howard Chu committed
264
	char		*me_map;
Howard Chu's avatar
Howard Chu committed
265
	MDB_txninfo	*me_txns;
Howard Chu's avatar
Howard Chu committed
266
267
	MDB_meta	*me_metas[2];
	MDB_meta	*me_meta;
Howard Chu's avatar
Howard Chu committed
268
	MDB_txn		*me_txn;		/* current write transaction */
Howard Chu's avatar
Howard Chu committed
269
270
	size_t		me_mapsize;
	off_t		me_size;		/* current file size */
Howard Chu's avatar
Howard Chu committed
271
272
273
274
	unsigned int	me_psize;
	int			me_db_toggle;
	MDB_dbx		*me_dbxs;		/* array */
	MDB_db		*me_dbs[2];
Howard Chu's avatar
Howard Chu committed
275
	MDB_oldpages *me_pghead;
Howard Chu's avatar
Howard Chu committed
276
	pthread_key_t	me_txkey;	/* thread-key for readers */
Howard Chu's avatar
Howard Chu committed
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
};

#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  */

Howard Chu's avatar
Howard Chu committed
292
293
static int  mdb_search_page_root(MDB_txn *txn,
			    MDB_dbi dbi, MDB_val *key,
Howard Chu's avatar
Howard Chu committed
294
295
			    MDB_cursor *cursor, int modify,
			    MDB_pageparent *mpp);
Howard Chu's avatar
Howard Chu committed
296
297
static int  mdb_search_page(MDB_txn *txn,
			    MDB_dbi dbi, MDB_val *key,
Howard Chu's avatar
Howard Chu committed
298
299
300
			    MDB_cursor *cursor, int modify,
			    MDB_pageparent *mpp);

Howard Chu's avatar
Howard Chu committed
301
static int  mdbenv_read_header(MDB_env *env, MDB_meta *meta);
Howard Chu's avatar
Howard Chu committed
302
static int  mdbenv_read_meta(MDB_env *env, int *which);
303
static int  mdbenv_write_meta(MDB_txn *txn);
Howard Chu's avatar
Howard Chu committed
304
305
static MDB_page *mdbenv_get_page(MDB_env *env, pgno_t pgno);

Howard Chu's avatar
Howard Chu committed
306
static MDB_node *mdb_search_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
Howard Chu's avatar
Howard Chu committed
307
			    MDB_val *key, int *exactp, unsigned int *kip);
Howard Chu's avatar
Howard Chu committed
308
static int  mdb_add_node(MDB_txn *txn, MDB_dbi dbi, MDB_page *mp,
Howard Chu's avatar
Howard Chu committed
309
310
			    indx_t indx, MDB_val *key, MDB_val *data,
			    pgno_t pgno, uint8_t flags);
Howard Chu's avatar
Howard Chu committed
311
static void mdb_del_node(MDB_page *mp, indx_t indx);
Howard Chu's avatar
Howard Chu committed
312
313
static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, unsigned int ki,
    MDB_pageparent *mpp, MDB_node *leaf);
Howard Chu's avatar
Howard Chu committed
314
315
316
317
318
static int  mdb_read_data(MDB_env *env, MDB_node *leaf, MDB_val *data);

static int		 mdb_rebalance(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *mp);
static int		 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
static int		 mdb_move_node(MDB_txn *txn, MDB_dbi dbi, 
Howard Chu's avatar
Howard Chu committed
319
320
				MDB_pageparent *src, indx_t srcindx,
				MDB_pageparent *dst, indx_t dstindx);
Howard Chu's avatar
Howard Chu committed
321
static int		 mdb_merge(MDB_txn *txn, MDB_dbi dbi, MDB_pageparent *src,
Howard Chu's avatar
Howard Chu committed
322
			    MDB_pageparent *dst);
Howard Chu's avatar
Howard Chu committed
323
static int		 mdb_split(MDB_txn *txn, MDB_dbi dbi, MDB_page **mpp,
Howard Chu's avatar
Howard Chu committed
324
325
			    unsigned int *newindxp, MDB_val *newkey,
			    MDB_val *newdata, pgno_t newpgno);
Howard Chu's avatar
Howard Chu committed
326
static MDB_dpage *mdb_new_page(MDB_txn *txn, MDB_dbi dbi, uint32_t flags, int num);
Howard Chu's avatar
Howard Chu committed
327
328
329
330
331

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

Howard Chu's avatar
Howard Chu committed
332
static int		 mdb_set_key(MDB_node *node, MDB_val *key);
Howard Chu's avatar
Howard Chu committed
333
334
335
336
337
338
339
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);
340
341
static int		 mdb_cursor_last(MDB_cursor *cursor,
			    MDB_val *key, MDB_val *data);
Howard Chu's avatar
Howard Chu committed
342

Howard Chu's avatar
Howard Chu committed
343
static size_t		 mdb_leaf_size(MDB_env *env, MDB_val *key,
Howard Chu's avatar
Howard Chu committed
344
			    MDB_val *data);
Howard Chu's avatar
Howard Chu committed
345
static size_t		 mdb_branch_size(MDB_env *env, MDB_val *key);
Howard Chu's avatar
Howard Chu committed
346
347
348
349
350
351
352
353
354

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)
{
355
356
357
358
359
	int diff, len_diff = -1;

	if (n1 >= n2) {
		len_diff = (n1 > n2);
		n1 = n2;
Howard Chu's avatar
Howard Chu committed
360
	}
361
362
	diff = memcmp(s1, s2, n1);
	return diff ? diff : len_diff;
Howard Chu's avatar
Howard Chu committed
363
364
365
366
367
}

static int
memnrcmp(const void *s1, size_t n1, const void *s2, size_t n2)
{
368
	const unsigned char	*p1, *p2, *p1_lim;
Howard Chu's avatar
Howard Chu committed
369
370

	if (n2 == 0)
371
372
373
		return n1 != 0;
	if (n1 == 0)
		return -1;
Howard Chu's avatar
Howard Chu committed
374
375
376
377

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

378
379
380
	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
381
382
383
384
385
	}
	return *p1 - *p2;
}

int
Howard Chu's avatar
Howard Chu committed
386
mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
Howard Chu's avatar
Howard Chu committed
387
{
Howard Chu's avatar
Howard Chu committed
388
	return txn->mt_dbxs[dbi].md_cmp(a, b);
Howard Chu's avatar
Howard Chu committed
389
390
391
}

static int
Howard Chu's avatar
Howard Chu committed
392
_mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *key1, const MDB_val *key2)
Howard Chu's avatar
Howard Chu committed
393
{
Howard Chu's avatar
Howard Chu committed
394
	if (F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_REVERSEKEY))
Howard Chu's avatar
Howard Chu committed
395
396
397
398
399
400
401
		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
402
mdb_alloc_page(MDB_txn *txn, MDB_page *parent, unsigned int parent_idx, int num)
Howard Chu's avatar
Howard Chu committed
403
404
{
	MDB_dpage *dp;
Howard Chu's avatar
Howard Chu committed
405
	pgno_t pgno = P_INVALID;
Howard Chu's avatar
Howard Chu committed
406
	ULONG oldest = txn->mt_txnid - 2;
Howard Chu's avatar
Howard Chu committed
407

Howard Chu's avatar
Howard Chu committed
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
	if (!txn->mt_env->me_pghead && txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
		/* See if there's anything in the free DB */
		MDB_pageparent mpp;
		MDB_node *leaf;
		ULONG *kptr;

		mpp.mp_parent = NULL;
		mpp.mp_pi = 0;
		mdb_search_page(txn, FREE_DBI, NULL, NULL, 0, &mpp);
		leaf = NODEPTR(mpp.mp_page, 0);
		kptr = (ULONG *)NODEKEY(leaf);

		/* It's potentially usable, unless there are still
		 * older readers outstanding. Grab it.
		 */
		if (oldest > *kptr) {
			MDB_oldpages *mop;
			MDB_val data;
			pgno_t *idl;

			mdb_read_data(txn->mt_env, leaf, &data);
			idl = (ULONG *)data.mv_data;
			mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
			mop->mo_next = txn->mt_env->me_pghead;
			mop->mo_txnid = *kptr;
			txn->mt_env->me_pghead = mop;
			memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));

Howard Chu's avatar
Howard Chu committed
436
437
438
439
440
441
442
443
444
445
#if DEBUG > 1
			{
				unsigned int i;
				DPRINTF("IDL read txn %lu root %lu num %lu",
					mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
				for (i=0; i<idl[0]; i++) {
					DPRINTF("IDL %lu", idl[i+1]);
				}
			}
#endif
Howard Chu's avatar
Howard Chu committed
446
447
448
449
450
451
452
453
			/* drop this IDL from the DB */
			mpp.mp_parent = NULL;
			mpp.mp_pi = 0;
			mdb_search_page(txn, FREE_DBI, NULL, NULL, 1, &mpp);
			leaf = NODEPTR(mpp.mp_page, 0);
			mdb_del0(txn, FREE_DBI, 0, &mpp, leaf);
		}
	}
Howard Chu's avatar
Howard Chu committed
454
	if (txn->mt_env->me_pghead) {
Howard Chu's avatar
Howard Chu committed
455
		unsigned int i;
Howard Chu's avatar
Howard Chu committed
456
		for (i=0; i<txn->mt_env->me_txns->mt_numreaders; i++) {
Howard Chu's avatar
Howard Chu committed
457
458
459
			ULONG mr = txn->mt_env->me_txns->mt_readers[i].mr_txnid;
			if (!mr) continue;
			if (mr < oldest)
Howard Chu's avatar
Howard Chu committed
460
461
462
463
464
465
				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) {
466
467
468
469
				/* 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
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
				;
			} 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;
					free(mop);
				}
			}
		}
	}
Howard Chu's avatar
Howard Chu committed
488

Howard Chu's avatar
Howard Chu committed
489
	if ((dp = malloc(txn->mt_env->me_psize * num + sizeof(MDB_dhead))) == NULL)
Howard Chu's avatar
Howard Chu committed
490
491
492
493
		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
494
	STAILQ_INSERT_TAIL(txn->mt_u.dirty_queue, dp, h.md_next);
Howard Chu's avatar
Howard Chu committed
495
496
497
498
499
500
	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
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516

	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
517
		if ((dp = mdb_alloc_page(txn, pp->mp_parent, pp->mp_pi, 1)) == NULL)
Howard Chu's avatar
Howard Chu committed
518
			return ENOMEM;
Howard Chu's avatar
Howard Chu committed
519
520
		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
521
		pgno = dp->p.mp_pgno;
Howard Chu's avatar
Howard Chu committed
522
		memcpy(&dp->p, mp, txn->mt_env->me_psize);
Howard Chu's avatar
Howard Chu committed
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
		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
550
	int rc, toggle;
Howard Chu's avatar
Howard Chu committed
551
552
553
554
555
556
557
558
559

	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
560
561
		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
562
563
564
			free(txn);
			return ENOMEM;
		}
Howard Chu's avatar
Howard Chu committed
565
		STAILQ_INIT(txn->mt_u.dirty_queue);
Howard Chu's avatar
Howard Chu committed
566

Howard Chu's avatar
Howard Chu committed
567
568
		pthread_mutex_lock(&env->me_txns->mt_wmutex);
		env->me_txns->mt_txnid++;
Howard Chu's avatar
Howard Chu committed
569
570
571
572
573
574
575
		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
576
	}
Howard Chu's avatar
Howard Chu committed
577

Howard Chu's avatar
Howard Chu committed
578
579
580
581
	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
582
			unsigned int i;
Howard Chu's avatar
Howard Chu committed
583
584
585
586
587
			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();
Howard Chu's avatar
Howard Chu committed
588
589
					r = &env->me_txns->mt_readers[i];
					pthread_setspecific(env->me_txkey, r);
Howard Chu's avatar
Howard Chu committed
590
591
592
593
594
595
596
597
598
					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
599
		}
Howard Chu's avatar
Howard Chu committed
600
601
602
		r->mr_txnid = txn->mt_txnid;
		txn->mt_u.reader = r;
	} else {
Howard Chu's avatar
Howard Chu committed
603
604
605
606
607
		env->me_txn = txn;
	}

	txn->mt_env = env;

Howard Chu's avatar
Howard Chu committed
608
	if ((rc = mdbenv_read_meta(env, &toggle)) != MDB_SUCCESS) {
Howard Chu's avatar
Howard Chu committed
609
610
611
612
		mdb_txn_abort(txn);
		return rc;
	}

Howard Chu's avatar
Howard Chu committed
613
614
	/* Copy the DB arrays */
	txn->mt_numdbs = env->me_numdbs;
Howard Chu's avatar
Howard Chu committed
615
616
617
618
619
620
	txn->mt_dbxs = env->me_dbxs;	/* mostly static anyway */
	txn->mt_dbs = malloc(env->me_maxdbs * sizeof(MDB_db));
	memcpy(txn->mt_dbs, env->me_meta->mm_dbs, 2 * sizeof(MDB_db));
	if (txn->mt_numdbs > 2)
		memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
			(txn->mt_numdbs - 2) * sizeof(MDB_db));
Howard Chu's avatar
Howard Chu committed
621
622
623
624

	if (!rdonly) {
		if (toggle)
			txn->mt_flags |= MDB_TXN_METOGGLE;
Howard Chu's avatar
Howard Chu committed
625
		txn->mt_next_pgno = env->me_meta->mm_last_pg+1;
Howard Chu's avatar
Howard Chu committed
626
	}
Howard Chu's avatar
Howard Chu committed
627

628
	DPRINTF("begin transaction %lu on mdbenv %p, root page %lu",
Howard Chu's avatar
Howard Chu committed
629
		txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
Howard Chu's avatar
Howard Chu committed
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644

	*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;
645
	DPRINTF("abort transaction %lu on mdbenv %p, root page %lu",
Howard Chu's avatar
Howard Chu committed
646
		txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
Howard Chu's avatar
Howard Chu committed
647

Howard Chu's avatar
Howard Chu committed
648
	free(txn->mt_dbs);
Howard Chu's avatar
Howard Chu committed
649

Howard Chu's avatar
Howard Chu committed
650
651
652
	if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
		txn->mt_u.reader->mr_txnid = 0;
	} else {
Howard Chu's avatar
Howard Chu committed
653
		MDB_oldpages *mop;
Howard Chu's avatar
Howard Chu committed
654
		/* Discard all dirty pages. */
Howard Chu's avatar
Howard Chu committed
655
656
657
		while (!STAILQ_EMPTY(txn->mt_u.dirty_queue)) {
			dp = STAILQ_FIRST(txn->mt_u.dirty_queue);
			STAILQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
Howard Chu's avatar
Howard Chu committed
658
659
			free(dp);
		}
Howard Chu's avatar
Howard Chu committed
660
		free(txn->mt_free_pgs);
Howard Chu's avatar
Howard Chu committed
661
		free(txn->mt_u.dirty_queue);
Howard Chu's avatar
Howard Chu committed
662
663
664
665
666
667

		while ((mop = txn->mt_env->me_pghead)) {
			txn->mt_env->me_pghead = mop->mo_next;
			free(mop);
		}

Howard Chu's avatar
Howard Chu committed
668
		env->me_txn = NULL;
Howard Chu's avatar
Howard Chu committed
669
		env->me_txns->mt_txnid--;
Howard Chu's avatar
Howard Chu committed
670
		pthread_mutex_unlock(&env->me_txns->mt_wmutex);
Howard Chu's avatar
Howard Chu committed
671
672
673
674
675
676
677
678
679
	}

	free(txn);
}

int
mdb_txn_commit(MDB_txn *txn)
{
	int		 n, done;
Howard Chu's avatar
Howard Chu committed
680
	unsigned int i;
Howard Chu's avatar
Howard Chu committed
681
682
683
684
	ssize_t		 rc;
	off_t		 size;
	MDB_dpage	*dp;
	MDB_env	*env;
685
	pgno_t	next;
Howard Chu's avatar
Howard Chu committed
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
	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
711
	if (STAILQ_EMPTY(txn->mt_u.dirty_queue))
Howard Chu's avatar
Howard Chu committed
712
713
		goto done;

Howard Chu's avatar
Howard Chu committed
714
	DPRINTF("committing transaction %lu on mdbenv %p, root page %lu",
Howard Chu's avatar
Howard Chu committed
715
716
	    txn->mt_txnid, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);

Howard Chu's avatar
Howard Chu committed
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
	/* should only be one record now */
	if (env->me_pghead) {
		MDB_val key, data;
		MDB_oldpages *mop;

		mop = env->me_pghead;
		key.mv_size = sizeof(pgno_t);
		key.mv_data = (char *)&mop->mo_txnid;
		data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
		data.mv_data = mop->mo_pages;
		mdb_put(txn, FREE_DBI, &key, &data, 0);
		free(env->me_pghead);
		env->me_pghead = NULL;
	}
	/* save to free list */
	if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
		MDB_val key, data;
		MDB_pageparent mpp;

		/* make sure last page of freeDB is touched and on freelist */
		key.mv_size = MAXKEYSIZE+1;
		key.mv_data = NULL;
		mpp.mp_parent = NULL;
		mpp.mp_pi = 0;
		mdb_search_page(txn, FREE_DBI, &key, NULL, 1, &mpp);

Howard Chu's avatar
Howard Chu committed
743
744
745
746
747
748
749
750
751
752
753
#if DEBUG > 1
		{
			unsigned int i;
			ULONG *idl = txn->mt_free_pgs;
			DPRINTF("IDL write txn %lu root %lu num %lu",
				txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
			for (i=0; i<idl[0]; i++) {
				DPRINTF("IDL %lu", idl[i+1]);
			}
		}
#endif
Howard Chu's avatar
Howard Chu committed
754
755
756
757
758
759
760
761
		/* write to last page of freeDB */
		key.mv_size = sizeof(pgno_t);
		key.mv_data = (char *)&txn->mt_txnid;
		data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
		data.mv_data = txn->mt_free_pgs;
		mdb_put(txn, FREE_DBI, &key, &data, 0);
	}

Howard Chu's avatar
Howard Chu committed
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
	/* Update DB root pointers. Their pages have already been
	 * touched so this is all in-place and cannot fail.
	 */
	{
		MDB_val data;
		data.mv_size = sizeof(MDB_db);

		for (i = 2; i < env->me_numdbs; i++) {
			if (env->me_dbs[env->me_db_toggle][i].md_root != txn->mt_dbs[i].md_root) {
				data.mv_data = &txn->mt_dbs[i];
				mdb_put(txn, i, &txn->mt_dbxs[i].md_name, &data, 0);
			}
		}
		for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
			data.mv_data = &txn->mt_dbs[i];
			mdb_put(txn, i, &txn->mt_dbxs[i].md_name, &data, 0);
		}
	}
Howard Chu's avatar
Howard Chu committed
780
781
782

	/* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
	 */
783
	next = 0;
Howard Chu's avatar
Howard Chu committed
784
785
786
	do {
		n = 0;
		done = 1;
787
		size = 0;
Howard Chu's avatar
Howard Chu committed
788
		STAILQ_FOREACH(dp, txn->mt_u.dirty_queue, h.md_next) {
789
			if (dp->p.mp_pgno != next) {
Howard Chu's avatar
Howard Chu committed
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
				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
805
				lseek(env->me_fd, dp->p.mp_pgno * env->me_psize, SEEK_SET);
806
807
				next = dp->p.mp_pgno;
			}
Howard Chu's avatar
Howard Chu committed
808
			DPRINTF("committing page %lu", dp->p.mp_pgno);
Howard Chu's avatar
Howard Chu committed
809
			iov[n].iov_len = env->me_psize * dp->h.md_num;
Howard Chu's avatar
Howard Chu committed
810
			iov[n].iov_base = &dp->p;
811
812
			size += iov[n].iov_len;
			next = dp->p.mp_pgno + dp->h.md_num;
Howard Chu's avatar
Howard Chu committed
813
814
815
816
817
818
819
820
821
822
823
824
825
			/* 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);
826
		if (rc != size) {
Howard Chu's avatar
Howard Chu committed
827
828
829
830
831
832
833
834
835
836
837
			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
838
839
	/* Drop the dirty pages.
	 */
Howard Chu's avatar
Howard Chu committed
840
841
842
	while (!STAILQ_EMPTY(txn->mt_u.dirty_queue)) {
		dp = STAILQ_FIRST(txn->mt_u.dirty_queue);
		STAILQ_REMOVE_HEAD(txn->mt_u.dirty_queue, h.md_next);
Howard Chu's avatar
Howard Chu committed
843
844
845
		free(dp);
	}

Howard Chu's avatar
Howard Chu committed
846
	if ((n = mdbenv_sync(env)) != 0 ||
847
	    (n = mdbenv_write_meta(txn)) != MDB_SUCCESS ||
Howard Chu's avatar
Howard Chu committed
848
849
850
851
852
	    (n = mdbenv_sync(env)) != 0) {
		mdb_txn_abort(txn);
		return n;
	}
	env->me_txn = NULL;
Howard Chu's avatar
Howard Chu committed
853

Howard Chu's avatar
Howard Chu committed
854
	/* update the DB tables */
Howard Chu's avatar
Howard Chu committed
855
	{
Howard Chu's avatar
Howard Chu committed
856
		int toggle = !env->me_db_toggle;
Howard Chu's avatar
Howard Chu committed
857

Howard Chu's avatar
Howard Chu committed
858
859
860
861
862
863
864
865
866
		for (i = 0; i < env->me_numdbs; i++) {
			if (env->me_dbs[toggle][i].md_root != txn->mt_dbs[i].md_root)
				env->me_dbs[toggle][i] = txn->mt_dbs[i];
		}
		for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
			env->me_dbxs[i] = txn->mt_dbxs[i];
			env->me_dbs[toggle][i] = txn->mt_dbs[i];
		}
		env->me_db_toggle = toggle;
Howard Chu's avatar
Howard Chu committed
867
		env->me_numdbs = txn->mt_numdbs;
Howard Chu's avatar
Howard Chu committed
868

Howard Chu's avatar
Howard Chu committed
869
		free(txn->mt_dbs);
Howard Chu's avatar
Howard Chu committed
870
871
	}

Howard Chu's avatar
Howard Chu committed
872
	pthread_mutex_unlock(&env->me_txns->mt_wmutex);
Howard Chu's avatar
Howard Chu committed
873
	free(txn->mt_free_pgs);
874
875
	free(txn->mt_u.dirty_queue);
	free(txn);
Howard Chu's avatar
Howard Chu committed
876
	txn = NULL;
Howard Chu's avatar
Howard Chu committed
877
878
879
880
881
882
883
884

done:
	mdb_txn_abort(txn);

	return MDB_SUCCESS;
}

static int
Howard Chu's avatar
Howard Chu committed
885
mdbenv_read_header(MDB_env *env, MDB_meta *meta)
Howard Chu's avatar
Howard Chu committed
886
887
888
{
	char		 page[PAGESIZE];
	MDB_page	*p;
Howard Chu's avatar
Howard Chu committed
889
	MDB_meta	*m;
Howard Chu's avatar
Howard Chu committed
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
	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
908
909
	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
910
911
912
		return EINVAL;
	}

Howard Chu's avatar
Howard Chu committed
913
914
915
	m = METADATA(p);
	if (m->mm_magic != MDB_MAGIC) {
		DPRINTF("meta has invalid magic");
Howard Chu's avatar
Howard Chu committed
916
917
918
		return EINVAL;
	}

Howard Chu's avatar
Howard Chu committed
919
	if (m->mm_version != MDB_VERSION) {
Howard Chu's avatar
Howard Chu committed
920
		DPRINTF("database is version %u, expected version %u",
Howard Chu's avatar
Howard Chu committed
921
		    m->mm_version, MDB_VERSION);
Howard Chu's avatar
Howard Chu committed
922
923
924
		return EINVAL;
	}

Howard Chu's avatar
Howard Chu committed
925
	memcpy(meta, m, sizeof(*m));
Howard Chu's avatar
Howard Chu committed
926
927
928
929
	return 0;
}

static int
Howard Chu's avatar
Howard Chu committed
930
mdbenv_init_meta(MDB_env *env, MDB_meta *meta)
Howard Chu's avatar
Howard Chu committed
931
{
932
	MDB_page *p, *q;
Howard Chu's avatar
Howard Chu committed
933
	MDB_meta *m;
934
	int rc;
Howard Chu's avatar
Howard Chu committed
935
	unsigned int	 psize;
Howard Chu's avatar
Howard Chu committed
936

Howard Chu's avatar
Howard Chu committed
937
938
939
	DPRINTF("writing new meta page");
	psize = sysconf(_SC_PAGE_SIZE);

Howard Chu's avatar
Howard Chu committed
940
941
942
943
944
	meta->mm_magic = MDB_MAGIC;
	meta->mm_version = MDB_VERSION;
	meta->mm_psize = psize;
	meta->mm_last_pg = 1;
	meta->mm_flags = env->me_flags & 0xffff;
Howard Chu's avatar
Howard Chu committed
945
946
947
948
#if __BYTE_ORDER == __LITTLE_ENDIAN
	/* freeDB keys are pgno_t's, must compare in int order */
	meta->mm_flags |= MDB_REVERSEKEY;
#endif
Howard Chu's avatar
Howard Chu committed
949
950
	meta->mm_dbs[0].md_root = P_INVALID;
	meta->mm_dbs[1].md_root = P_INVALID;
Howard Chu's avatar
Howard Chu committed
951
952
953

	p = calloc(2, psize);
	p->mp_pgno = 0;
954
	p->mp_flags = P_META;
Howard Chu's avatar
Howard Chu committed
955

Howard Chu's avatar
Howard Chu committed
956
957
	m = METADATA(p);
	memcpy(m, meta, sizeof(*meta));
Howard Chu's avatar
Howard Chu committed
958

Howard Chu's avatar
Howard Chu committed
959
	q = (MDB_page *)((char *)p + psize);
Howard Chu's avatar
Howard Chu committed
960

Howard Chu's avatar
Howard Chu committed
961
	q->mp_pgno = 1;
962
	q->mp_flags = P_META;
Howard Chu's avatar
Howard Chu committed
963

Howard Chu's avatar
Howard Chu committed
964
965
	m = METADATA(q);
	memcpy(m, meta, sizeof(*meta));
Howard Chu's avatar
Howard Chu committed
966

Howard Chu's avatar
Howard Chu committed
967
	rc = write(env->me_fd, p, psize * 2);
968
	free(p);
Howard Chu's avatar
Howard Chu committed
969
	return (rc == (int)psize * 2) ? MDB_SUCCESS : errno;
970
971
972
973
974
975
976
977
}

static int
mdbenv_write_meta(MDB_txn *txn)
{
	MDB_env *env;
	MDB_meta	meta;
	off_t off;
Howard Chu's avatar
Howard Chu committed
978
979
	int rc, len;
	char *ptr;
Howard Chu's avatar
Howard Chu committed
980

981
982
983
	assert(txn != NULL);
	assert(txn->mt_env != NULL);

Howard Chu's avatar
Howard Chu committed
984
	DPRINTF("writing meta page for root page %lu", txn->mt_dbs[MAIN_DBI].md_root);
985
986
987

	env = txn->mt_env;

Howard Chu's avatar
Howard Chu committed
988
	ptr = (char *)&meta;
Howard Chu's avatar
Howard Chu committed
989
	off = offsetof(MDB_meta, mm_dbs[0].md_depth);
Howard Chu's avatar
Howard Chu committed
990
991
992
	len = sizeof(MDB_meta) - off;

	ptr += off;
Howard Chu's avatar
Howard Chu committed
993
994
	meta.mm_dbs[0] = txn->mt_dbs[0];
	meta.mm_dbs[1] = txn->mt_dbs[1];
Howard Chu's avatar
Howard Chu committed
995
996
	meta.mm_last_pg = txn->mt_next_pgno - 1;
	meta.mm_txnid = txn->mt_txnid;
997

Howard Chu's avatar
Howard Chu committed
998
	if (!F_ISSET(txn->mt_flags, MDB_TXN_METOGGLE))
Howard Chu's avatar
Howard Chu committed
999
		off += env->me_psize;
1000
1001
1002
	off += PAGEHDRSZ;

	lseek(env->me_fd, off, SEEK_SET);
Howard Chu's avatar
Howard Chu committed
1003
1004
	rc = write(env->me_fd, ptr, len);
	if (rc != len) {
1005
1006
		DPRINTF("write failed, disk error?");
		return errno;
Howard Chu's avatar
Howard Chu committed
1007
1008
1009
1010
1011
1012
	}

	return MDB_SUCCESS;
}

static int
Howard Chu's avatar
Howard Chu committed
1013
mdbenv_read_meta(MDB_env *env, int *which)
Howard Chu's avatar
Howard Chu committed
1014
{
Howard Chu's avatar
Howard Chu committed
1015
	int toggle = 0;
Howard Chu's avatar
Howard Chu committed
1016
1017
1018

	assert(env != NULL);

Howard Chu's avatar
Howard Chu committed
1019
	if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1020
		toggle = 1;
Howard Chu's avatar
Howard Chu committed
1021

Howard Chu's avatar
Howard Chu committed
1022
1023
1024
1025
	if (env->me_meta != env->me_metas[toggle])
		env->me_meta = env->me_metas[toggle];
	if (which)
		*which = toggle;
Howard Chu's avatar
Howard Chu committed
1026

1027
	DPRINTF("Using meta page %d", toggle);
Howard Chu's avatar
Howard Chu committed
1028

1029
	return MDB_SUCCESS;
Howard Chu's avatar
Howard Chu committed
1030
1031
1032
}

int
1033
mdbenv_create(MDB_env **env)
Howard Chu's avatar
Howard Chu committed
1034
1035
1036
1037
1038
1039
{
	MDB_env *e;

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

Howard Chu's avatar
Howard Chu committed
1040
	e->me_maxreaders = DEFAULT_READERS;
Howard Chu's avatar
Howard Chu committed
1041
	e->me_maxdbs = 2;
Howard Chu's avatar
Howard Chu committed
1042
1043
	e->me_fd = -1;
	e->me_lfd = -1;
Howard Chu's avatar
Howard Chu committed
1044
1045
1046
1047
	*env = e;
	return MDB_SUCCESS;
}

Howard Chu's avatar
Howard Chu committed
1048
1049
1050
1051
1052
int
mdbenv_set_mapsize(MDB_env *env, size_t size)
{
	if (env->me_map)
		return EINVAL;
Howard Chu's avatar
Howard Chu committed
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
	env->me_mapsize = size;
	return MDB_SUCCESS;
}

int
mdbenv_set_maxdbs(MDB_env *env, int dbs)
{
	if (env->me_map)
		return EINVAL;
	env->me_maxdbs = dbs;
Howard Chu's avatar
Howard Chu committed
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
	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
1082
1083
1084
1085
int
mdbenv_open2(MDB_env *env, unsigned int flags)
{
	int i, newenv = 0;
Howard Chu's avatar
Howard Chu committed
1086
1087
	MDB_meta meta;
	MDB_page *p;
Howard Chu's avatar
Howard Chu committed
1088
1089
1090

	env->me_flags = flags;

Howard Chu's avatar
Howard Chu committed
1091
1092
1093
	memset(&meta, 0, sizeof(meta));

	if ((i = mdbenv_read_header(env, &meta)) != 0) {
Howard Chu's avatar
Howard Chu committed
1094
1095
1096
1097
1098
1099
		if (i != ENOENT)
			return i;
		DPRINTF("new mdbenv");
		newenv = 1;
	}

Howard Chu's avatar
Howard Chu committed
1100
1101
1102
	if (!env->me_mapsize) {
		env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
	}
1103

Howard Chu's avatar
Howard Chu committed
1104
	i = MAP_SHARED;
Howard Chu's avatar
Howard Chu committed
1105
	if (meta.mm_address && (flags & MDB_FIXEDMAP))
Howard Chu's avatar
Howard Chu committed
1106
		i |= MAP_FIXED;
Howard Chu's avatar
Howard Chu committed
1107
	env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
Howard Chu's avatar
Howard Chu committed
1108
1109
1110
1111
1112
		env->me_fd, 0);
	if (env->me_map == MAP_FAILED)
		return errno;

	if (newenv) {
Howard Chu's avatar
Howard Chu committed
1113
		meta.mm_mapsize = env->me_mapsize;
Howard Chu's avatar
Howard Chu committed
1114
		if (flags & MDB_FIXEDMAP)
Howard Chu's avatar
Howard Chu committed
1115
1116
			meta.mm_address = env->me_map;
		i = mdbenv_init_meta(env, &meta);
1117
1118
1119
1120
		if (i != MDB_SUCCESS) {
			munmap(env->me_map, env->me_mapsize);
			return i;
		}
Howard Chu's avatar
Howard Chu committed
1121
	}
Howard Chu's avatar
Howard Chu committed
1122
1123
1124
1125
1126
	env->me_psize = meta.mm_psize;

	p = (MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)(MDB_page *)env->me_map;
	env->me_metas[0] = METADATA(p);
	env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
Howard Chu's avatar
Howard Chu committed
1127

Howard Chu's avatar
Howard Chu committed
1128
	if ((i = mdbenv_read_meta(env, NULL)) != 0)
Howard Chu's avatar
Howard Chu committed
1129
1130
1131
		return i;

	DPRINTF("opened database version %u, pagesize %u",
Howard Chu's avatar
Howard Chu committed
1132
1133
1134
1135
1136
1137
1138
	    env->me_meta->mm_version, env->me_psize);
	DPRINTF("depth: %u", env->me_meta->mm_dbs[MAIN_DBI].md_depth);
	DPRINTF("entries: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_entries);
	DPRINTF("branch pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_branch_pages);
	DPRINTF("leaf pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_leaf_pages);
	DPRINTF("overflow pages: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_overflow_pages);
	DPRINTF("root: %lu", env->me_meta->mm_dbs[MAIN_DBI].md_root);
Howard Chu's avatar
Howard Chu committed
1139
1140
1141
1142

	return MDB_SUCCESS;
}

Howard Chu's avatar
Howard Chu committed
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
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
1153
/* downgrade the exclusive lock on the region back to shared */
Howard Chu's avatar
Howard Chu committed
1154
1155
static void
mdbenv_share_locks(MDB_env *env)
Howard Chu's avatar
Howard Chu committed
1156
{
Howard Chu's avatar
Howard Chu committed
1157
	struct flock lock_info;
1158

Howard Chu's avatar
Howard Chu committed
1159
	env->me_txns->mt_txnid = env->me_meta->mm_txnid;
1160

Howard Chu's avatar
Howard Chu committed
1161
1162
1163
1164
1165
1166
1167
	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);
}
1168

Howard Chu's avatar
Howard Chu committed
1169
1170
1171
1172
1173
1174
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
1175

Howard Chu's avatar
Howard Chu committed
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
	*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
1207
1208
		}
	} else {
Howard Chu's avatar
Howard Chu committed
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
		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
1220
1221
1222
1223
1224
1225
		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);
1226
1227
		env->me_txns->mt_version = MDB_VERSION;
		env->me_txns->mt_magic = MDB_MAGIC;
Howard Chu's avatar
Howard Chu committed
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
		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
1245
	}
Howard Chu's avatar
Howard Chu committed
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
1277
1278
	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
1279
1280
1281
1282
1283
1284

	if ((rc = mdbenv_open2(env, flags)) != MDB_SUCCESS) {
		close(env->me_fd);
		env->me_fd = -1;
	} else {
		env->me_path = strdup(path);
1285
		DPRINTF("opened dbenv %p", (void *) env);
Howard Chu's avatar
Howard Chu committed
1286
1287
1288
		pthread_key_create(&env->me_txkey, mdbenv_reader_dest);
		if (excl)
			mdbenv_share_locks(env);
Howard Chu's avatar
Howard Chu committed
1289
1290
1291
1292
		env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
		env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
		env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
		env->me_numdbs = 2;
Howard Chu's avatar
Howard Chu committed
1293
1294
	}

Howard Chu's avatar
Howard Chu committed
1295
1296
leave:
	free(lpath);
Howard Chu's avatar
Howard Chu committed
1297
1298
1299
1300
1301
1302
1303
1304
1305
	return rc;
}

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

Howard Chu's avatar
Howard Chu committed
1306
	free(env->me_dbxs);
Howard Chu's avatar
Howard Chu committed
1307
1308
1309
1310
1311
1312
	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
1313
	if (env->me_txns) {