2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
65 #if defined(__APPLE__) || defined (BSD)
66 # define MDB_USE_POSIX_SEM 1
67 # define MDB_FDATASYNC fsync
68 #elif defined(ANDROID)
69 # define MDB_FDATASYNC fsync
74 #ifdef MDB_USE_POSIX_SEM
75 #include <semaphore.h>
80 #include <valgrind/memcheck.h>
81 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
82 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
83 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
84 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
85 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
87 #define VGMEMP_CREATE(h,r,z)
88 #define VGMEMP_ALLOC(h,a,s)
89 #define VGMEMP_FREE(h,a)
90 #define VGMEMP_DESTROY(h)
91 #define VGMEMP_DEFINED(a,s)
95 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
96 /* Solaris just defines one or the other */
97 # define LITTLE_ENDIAN 1234
98 # define BIG_ENDIAN 4321
99 # ifdef _LITTLE_ENDIAN
100 # define BYTE_ORDER LITTLE_ENDIAN
102 # define BYTE_ORDER BIG_ENDIAN
105 # define BYTE_ORDER __BYTE_ORDER
109 #ifndef LITTLE_ENDIAN
110 #define LITTLE_ENDIAN __LITTLE_ENDIAN
113 #define BIG_ENDIAN __BIG_ENDIAN
116 #if defined(__i386) || defined(__x86_64)
117 #define MISALIGNED_OK 1
123 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
124 # error "Unknown or unsupported endianness (BYTE_ORDER)"
125 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
126 # error "Two's complement, reasonably sized integer types, please"
129 /** @defgroup internal MDB Internals
132 /** @defgroup compat Windows Compatibility Macros
133 * A bunch of macros to minimize the amount of platform-specific ifdefs
134 * needed throughout the rest of the code. When the features this library
135 * needs are similar enough to POSIX to be hidden in a one-or-two line
136 * replacement, this macro approach is used.
140 #define pthread_t DWORD
141 #define pthread_mutex_t HANDLE
142 #define pthread_key_t DWORD
143 #define pthread_self() GetCurrentThreadId()
144 #define pthread_key_create(x,y) \
145 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
146 #define pthread_key_delete(x) TlsFree(x)
147 #define pthread_getspecific(x) TlsGetValue(x)
148 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
149 #define pthread_mutex_unlock(x) ReleaseMutex(x)
150 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
151 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
152 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
153 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
154 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
155 #define getpid() GetCurrentProcessId()
156 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
157 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
158 #define ErrCode() GetLastError()
159 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
160 #define close(fd) CloseHandle(fd)
161 #define munmap(ptr,len) UnmapViewOfFile(ptr)
164 #ifdef MDB_USE_POSIX_SEM
166 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
167 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
168 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
169 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
172 mdb_sem_wait(sem_t *sem)
175 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
180 /** Lock the reader mutex.
182 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
183 /** Unlock the reader mutex.
185 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
187 /** Lock the writer mutex.
188 * Only a single write transaction is allowed at a time. Other writers
189 * will block waiting for this mutex.
191 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
192 /** Unlock the writer mutex.
194 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
195 #endif /* MDB_USE_POSIX_SEM */
197 /** Get the error code for the last failed system function.
199 #define ErrCode() errno
201 /** An abstraction for a file handle.
202 * On POSIX systems file handles are small integers. On Windows
203 * they're opaque pointers.
207 /** A value for an invalid file handle.
208 * Mainly used to initialize file variables and signify that they are
211 #define INVALID_HANDLE_VALUE (-1)
213 /** Get the size of a memory page for the system.
214 * This is the basic size that the platform's memory manager uses, and is
215 * fundamental to the use of memory-mapped files.
217 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
220 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
223 #define MNAME_LEN (sizeof(pthread_mutex_t))
229 /** A flag for opening a file and requesting synchronous data writes.
230 * This is only used when writing a meta page. It's not strictly needed;
231 * we could just do a normal write and then immediately perform a flush.
232 * But if this flag is available it saves us an extra system call.
234 * @note If O_DSYNC is undefined but exists in /usr/include,
235 * preferably set some compiler flag to get the definition.
236 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
239 # define MDB_DSYNC O_DSYNC
243 /** Function for flushing the data of a file. Define this to fsync
244 * if fdatasync() is not supported.
246 #ifndef MDB_FDATASYNC
247 # define MDB_FDATASYNC fdatasync
251 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
262 /** A page number in the database.
263 * Note that 64 bit page numbers are overkill, since pages themselves
264 * already represent 12-13 bits of addressable memory, and the OS will
265 * always limit applications to a maximum of 63 bits of address space.
267 * @note In the #MDB_node structure, we only store 48 bits of this value,
268 * which thus limits us to only 60 bits of addressable data.
270 typedef MDB_ID pgno_t;
272 /** A transaction ID.
273 * See struct MDB_txn.mt_txnid for details.
275 typedef MDB_ID txnid_t;
277 /** @defgroup debug Debug Macros
281 /** Enable debug output.
282 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
283 * read from and written to the database (used for free space management).
288 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
289 # define DPRINTF (void) /* Vararg macros may be unsupported */
291 static int mdb_debug;
292 static txnid_t mdb_debug_start;
294 /** Print a debug message with printf formatting. */
295 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
296 ((void) ((mdb_debug) && \
297 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
299 # define DPRINTF(fmt, ...) ((void) 0)
300 # define MDB_DEBUG_SKIP
302 /** Print a debug string.
303 * The string is printed literally, with no format processing.
305 #define DPUTS(arg) DPRINTF("%s", arg)
308 /** A default memory page size.
309 * The actual size is platform-dependent, but we use this for
310 * boot-strapping. We probably should not be using this any more.
311 * The #GET_PAGESIZE() macro is used to get the actual size.
313 * Note that we don't currently support Huge pages. On Linux,
314 * regular data files cannot use Huge pages, and in general
315 * Huge pages aren't actually pageable. We rely on the OS
316 * demand-pager to read our data and page it out when memory
317 * pressure from other processes is high. So until OSs have
318 * actual paging support for Huge pages, they're not viable.
320 #define MDB_PAGESIZE 4096
322 /** The minimum number of keys required in a database page.
323 * Setting this to a larger value will place a smaller bound on the
324 * maximum size of a data item. Data items larger than this size will
325 * be pushed into overflow pages instead of being stored directly in
326 * the B-tree node. This value used to default to 4. With a page size
327 * of 4096 bytes that meant that any item larger than 1024 bytes would
328 * go into an overflow page. That also meant that on average 2-3KB of
329 * each overflow page was wasted space. The value cannot be lower than
330 * 2 because then there would no longer be a tree structure. With this
331 * value, items larger than 2KB will go into overflow pages, and on
332 * average only 1KB will be wasted.
334 #define MDB_MINKEYS 2
336 /** A stamp that identifies a file as an MDB file.
337 * There's nothing special about this value other than that it is easily
338 * recognizable, and it will reflect any byte order mismatches.
340 #define MDB_MAGIC 0xBEEFC0DE
342 /** The version number for a database's file format. */
343 #define MDB_VERSION 1
345 /** @brief The maximum size of a key in the database.
347 * We require that keys all fit onto a regular page. This limit
348 * could be raised a bit further if needed; to something just
349 * under #MDB_PAGESIZE / #MDB_MINKEYS.
351 * Note that data items in an #MDB_DUPSORT database are actually keys
352 * of a subDB, so they're also limited to this size.
354 #ifndef MDB_MAXKEYSIZE
355 #define MDB_MAXKEYSIZE 511
358 /** @brief The maximum size of a data item.
360 * We only store a 32 bit value for node sizes.
362 #define MAXDATASIZE 0xffffffffUL
367 * This is used for printing a hex dump of a key's contents.
369 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
370 /** Display a key in hex.
372 * Invoke a function to display a key in hex.
374 #define DKEY(x) mdb_dkey(x, kbuf)
376 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
380 /** An invalid page number.
381 * Mainly used to denote an empty tree.
383 #define P_INVALID (~(pgno_t)0)
385 /** Test if the flags \b f are set in a flag word \b w. */
386 #define F_ISSET(w, f) (((w) & (f)) == (f))
388 /** Used for offsets within a single page.
389 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
392 typedef uint16_t indx_t;
394 /** Default size of memory map.
395 * This is certainly too small for any actual applications. Apps should always set
396 * the size explicitly using #mdb_env_set_mapsize().
398 #define DEFAULT_MAPSIZE 1048576
400 /** @defgroup readers Reader Lock Table
401 * Readers don't acquire any locks for their data access. Instead, they
402 * simply record their transaction ID in the reader table. The reader
403 * mutex is needed just to find an empty slot in the reader table. The
404 * slot's address is saved in thread-specific data so that subsequent read
405 * transactions started by the same thread need no further locking to proceed.
407 * No reader table is used if the database is on a read-only filesystem.
409 * Since the database uses multi-version concurrency control, readers don't
410 * actually need any locking. This table is used to keep track of which
411 * readers are using data from which old transactions, so that we'll know
412 * when a particular old transaction is no longer in use. Old transactions
413 * that have discarded any data pages can then have those pages reclaimed
414 * for use by a later write transaction.
416 * The lock table is constructed such that reader slots are aligned with the
417 * processor's cache line size. Any slot is only ever used by one thread.
418 * This alignment guarantees that there will be no contention or cache
419 * thrashing as threads update their own slot info, and also eliminates
420 * any need for locking when accessing a slot.
422 * A writer thread will scan every slot in the table to determine the oldest
423 * outstanding reader transaction. Any freed pages older than this will be
424 * reclaimed by the writer. The writer doesn't use any locks when scanning
425 * this table. This means that there's no guarantee that the writer will
426 * see the most up-to-date reader info, but that's not required for correct
427 * operation - all we need is to know the upper bound on the oldest reader,
428 * we don't care at all about the newest reader. So the only consequence of
429 * reading stale information here is that old pages might hang around a
430 * while longer before being reclaimed. That's actually good anyway, because
431 * the longer we delay reclaiming old pages, the more likely it is that a
432 * string of contiguous pages can be found after coalescing old pages from
433 * many old transactions together.
435 * @todo We don't actually do such coalescing yet, we grab pages from one
436 * old transaction at a time.
439 /** Number of slots in the reader table.
440 * This value was chosen somewhat arbitrarily. 126 readers plus a
441 * couple mutexes fit exactly into 8KB on my development machine.
442 * Applications should set the table size using #mdb_env_set_maxreaders().
444 #define DEFAULT_READERS 126
446 /** The size of a CPU cache line in bytes. We want our lock structures
447 * aligned to this size to avoid false cache line sharing in the
449 * This value works for most CPUs. For Itanium this should be 128.
455 /** The information we store in a single slot of the reader table.
456 * In addition to a transaction ID, we also record the process and
457 * thread ID that owns a slot, so that we can detect stale information,
458 * e.g. threads or processes that went away without cleaning up.
459 * @note We currently don't check for stale records. We simply re-init
460 * the table when we know that we're the only process opening the
463 typedef struct MDB_rxbody {
464 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
465 * Multiple readers that start at the same time will probably have the
466 * same ID here. Again, it's not important to exclude them from
467 * anything; all we need to know is which version of the DB they
468 * started from so we can avoid overwriting any data used in that
469 * particular version.
472 /** The process ID of the process owning this reader txn. */
474 /** The thread ID of the thread owning this txn. */
478 /** The actual reader record, with cacheline padding. */
479 typedef struct MDB_reader {
482 /** shorthand for mrb_txnid */
483 #define mr_txnid mru.mrx.mrb_txnid
484 #define mr_pid mru.mrx.mrb_pid
485 #define mr_tid mru.mrx.mrb_tid
486 /** cache line alignment */
487 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
491 /** The header for the reader table.
492 * The table resides in a memory-mapped file. (This is a different file
493 * than is used for the main database.)
495 * For POSIX the actual mutexes reside in the shared memory of this
496 * mapped file. On Windows, mutexes are named objects allocated by the
497 * kernel; we store the mutex names in this mapped file so that other
498 * processes can grab them. This same approach is also used on
499 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
500 * process-shared POSIX mutexes. For these cases where a named object
501 * is used, the object name is derived from a 64 bit FNV hash of the
502 * environment pathname. As such, naming collisions are extremely
503 * unlikely. If a collision occurs, the results are unpredictable.
505 typedef struct MDB_txbody {
506 /** Stamp identifying this as an MDB file. It must be set
509 /** Version number of this lock file. Must be set to #MDB_VERSION. */
510 uint32_t mtb_version;
511 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
512 char mtb_rmname[MNAME_LEN];
514 /** Mutex protecting access to this table.
515 * This is the reader lock that #LOCK_MUTEX_R acquires.
517 pthread_mutex_t mtb_mutex;
519 /** The ID of the last transaction committed to the database.
520 * This is recorded here only for convenience; the value can always
521 * be determined by reading the main database meta pages.
524 /** The number of slots that have been used in the reader table.
525 * This always records the maximum count, it is not decremented
526 * when readers release their slots.
528 unsigned mtb_numreaders;
531 /** The actual reader table definition. */
532 typedef struct MDB_txninfo {
535 #define mti_magic mt1.mtb.mtb_magic
536 #define mti_version mt1.mtb.mtb_version
537 #define mti_mutex mt1.mtb.mtb_mutex
538 #define mti_rmname mt1.mtb.mtb_rmname
539 #define mti_txnid mt1.mtb.mtb_txnid
540 #define mti_numreaders mt1.mtb.mtb_numreaders
541 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
544 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
545 char mt2_wmname[MNAME_LEN];
546 #define mti_wmname mt2.mt2_wmname
548 pthread_mutex_t mt2_wmutex;
549 #define mti_wmutex mt2.mt2_wmutex
551 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
553 MDB_reader mti_readers[1];
557 /** Common header for all page types.
558 * Overflow records occupy a number of contiguous pages with no
559 * headers on any page after the first.
561 typedef struct MDB_page {
562 #define mp_pgno mp_p.p_pgno
563 #define mp_next mp_p.p_next
565 pgno_t p_pgno; /**< page number */
566 void * p_next; /**< for in-memory list of freed structs */
569 /** @defgroup mdb_page Page Flags
571 * Flags for the page headers.
574 #define P_BRANCH 0x01 /**< branch page */
575 #define P_LEAF 0x02 /**< leaf page */
576 #define P_OVERFLOW 0x04 /**< overflow page */
577 #define P_META 0x08 /**< meta page */
578 #define P_DIRTY 0x10 /**< dirty page */
579 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
580 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
582 uint16_t mp_flags; /**< @ref mdb_page */
583 #define mp_lower mp_pb.pb.pb_lower
584 #define mp_upper mp_pb.pb.pb_upper
585 #define mp_pages mp_pb.pb_pages
588 indx_t pb_lower; /**< lower bound of free space */
589 indx_t pb_upper; /**< upper bound of free space */
591 uint32_t pb_pages; /**< number of overflow pages */
593 indx_t mp_ptrs[1]; /**< dynamic size */
596 /** Size of the page header, excluding dynamic data at the end */
597 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
599 /** Address of first usable data byte in a page, after the header */
600 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
602 /** Number of nodes on a page */
603 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
605 /** The amount of space remaining in the page */
606 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
608 /** The percentage of space used in the page, in tenths of a percent. */
609 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
610 ((env)->me_psize - PAGEHDRSZ))
611 /** The minimum page fill factor, in tenths of a percent.
612 * Pages emptier than this are candidates for merging.
614 #define FILL_THRESHOLD 250
616 /** Test if a page is a leaf page */
617 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
618 /** Test if a page is a LEAF2 page */
619 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
620 /** Test if a page is a branch page */
621 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
622 /** Test if a page is an overflow page */
623 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
624 /** Test if a page is a sub page */
625 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
627 /** The number of overflow pages needed to store the given size. */
628 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
630 /** Header for a single key/data pair within a page.
631 * We guarantee 2-byte alignment for nodes.
633 typedef struct MDB_node {
634 /** lo and hi are used for data size on leaf nodes and for
635 * child pgno on branch nodes. On 64 bit platforms, flags
636 * is also used for pgno. (Branch nodes have no flags).
637 * They are in host byte order in case that lets some
638 * accesses be optimized into a 32-bit word access.
640 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
641 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
642 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
643 /** @defgroup mdb_node Node Flags
645 * Flags for node headers.
648 #define F_BIGDATA 0x01 /**< data put on overflow page */
649 #define F_SUBDATA 0x02 /**< data is a sub-database */
650 #define F_DUPDATA 0x04 /**< data has duplicates */
652 /** valid flags for #mdb_node_add() */
653 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
656 unsigned short mn_flags; /**< @ref mdb_node */
657 unsigned short mn_ksize; /**< key size */
658 char mn_data[1]; /**< key and data are appended here */
661 /** Size of the node header, excluding dynamic data at the end */
662 #define NODESIZE offsetof(MDB_node, mn_data)
664 /** Bit position of top word in page number, for shifting mn_flags */
665 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
667 /** Size of a node in a branch page with a given key.
668 * This is just the node header plus the key, there is no data.
670 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
672 /** Size of a node in a leaf page with a given key and data.
673 * This is node header plus key plus data size.
675 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
677 /** Address of node \b i in page \b p */
678 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
680 /** Address of the key for the node */
681 #define NODEKEY(node) (void *)((node)->mn_data)
683 /** Address of the data for a node */
684 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
686 /** Get the page number pointed to by a branch node */
687 #define NODEPGNO(node) \
688 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
689 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
690 /** Set the page number in a branch node */
691 #define SETPGNO(node,pgno) do { \
692 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
693 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
695 /** Get the size of the data in a leaf node */
696 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
697 /** Set the size of the data for a leaf node */
698 #define SETDSZ(node,size) do { \
699 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
700 /** The size of a key in a node */
701 #define NODEKSZ(node) ((node)->mn_ksize)
703 /** Copy a page number from src to dst */
705 #define COPY_PGNO(dst,src) dst = src
707 #if SIZE_MAX > 4294967295UL
708 #define COPY_PGNO(dst,src) do { \
709 unsigned short *s, *d; \
710 s = (unsigned short *)&(src); \
711 d = (unsigned short *)&(dst); \
718 #define COPY_PGNO(dst,src) do { \
719 unsigned short *s, *d; \
720 s = (unsigned short *)&(src); \
721 d = (unsigned short *)&(dst); \
727 /** The address of a key in a LEAF2 page.
728 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
729 * There are no node headers, keys are stored contiguously.
731 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
733 /** Set the \b node's key into \b key, if requested. */
734 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
735 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
737 /** Information about a single database in the environment. */
738 typedef struct MDB_db {
739 uint32_t md_pad; /**< also ksize for LEAF2 pages */
740 uint16_t md_flags; /**< @ref mdb_dbi_open */
741 uint16_t md_depth; /**< depth of this tree */
742 pgno_t md_branch_pages; /**< number of internal pages */
743 pgno_t md_leaf_pages; /**< number of leaf pages */
744 pgno_t md_overflow_pages; /**< number of overflow pages */
745 size_t md_entries; /**< number of data items */
746 pgno_t md_root; /**< the root page of this tree */
749 /** Handle for the DB used to track free pages. */
751 /** Handle for the default DB. */
754 /** Meta page content. */
755 typedef struct MDB_meta {
756 /** Stamp identifying this as an MDB file. It must be set
759 /** Version number of this lock file. Must be set to #MDB_VERSION. */
761 void *mm_address; /**< address for fixed mapping */
762 size_t mm_mapsize; /**< size of mmap region */
763 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
764 /** The size of pages used in this DB */
765 #define mm_psize mm_dbs[0].md_pad
766 /** Any persistent environment flags. @ref mdb_env */
767 #define mm_flags mm_dbs[0].md_flags
768 pgno_t mm_last_pg; /**< last used page in file */
769 txnid_t mm_txnid; /**< txnid that committed this page */
772 /** Buffer for a stack-allocated dirty page.
773 * The members define size and alignment, and silence type
774 * aliasing warnings. They are not used directly; that could
775 * mean incorrectly using several union members in parallel.
777 typedef union MDB_pagebuf {
778 char mb_raw[MDB_PAGESIZE];
781 char mm_pad[PAGEHDRSZ];
786 /** Auxiliary DB info.
787 * The information here is mostly static/read-only. There is
788 * only a single copy of this record in the environment.
790 typedef struct MDB_dbx {
791 MDB_val md_name; /**< name of the database */
792 MDB_cmp_func *md_cmp; /**< function for comparing keys */
793 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
794 MDB_rel_func *md_rel; /**< user relocate function */
795 void *md_relctx; /**< user-provided context for md_rel */
798 /** A database transaction.
799 * Every operation requires a transaction handle.
802 MDB_txn *mt_parent; /**< parent of a nested txn */
803 MDB_txn *mt_child; /**< nested txn under this txn */
804 pgno_t mt_next_pgno; /**< next unallocated page */
805 /** The ID of this transaction. IDs are integers incrementing from 1.
806 * Only committed write transactions increment the ID. If a transaction
807 * aborts, the ID may be re-used by the next writer.
810 MDB_env *mt_env; /**< the DB environment */
811 /** The list of pages that became unused during this transaction.
815 MDB_ID2L dirty_list; /**< for write txns: modified pages */
816 MDB_reader *reader; /**< this thread's reader table slot or NULL */
818 /** Array of records for each DB known in the environment. */
820 /** Array of MDB_db records for each known DB */
822 /** @defgroup mt_dbflag Transaction DB Flags
826 #define DB_DIRTY 0x01 /**< DB was written in this txn */
827 #define DB_STALE 0x02 /**< DB record is older than txnID */
829 /** In write txns, array of cursors for each DB */
830 MDB_cursor **mt_cursors;
831 /** Array of flags for each DB */
832 unsigned char *mt_dbflags;
833 /** Number of DB records in use. This number only ever increments;
834 * we don't decrement it when individual DB handles are closed.
838 /** @defgroup mdb_txn Transaction Flags
842 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
843 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
844 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
846 unsigned int mt_flags; /**< @ref mdb_txn */
847 /** dirty_list maxsize - #allocated pages including in parent txns */
848 unsigned int mt_dirty_room;
849 /** Tracks which of the two meta pages was used at the start
850 * of this transaction.
852 unsigned int mt_toggle;
855 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
856 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
857 * raise this on a 64 bit machine.
859 #define CURSOR_STACK 32
863 /** Cursors are used for all DB operations */
865 /** Next cursor on this DB in this txn */
867 /** Original cursor if this is a shadow */
869 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
870 struct MDB_xcursor *mc_xcursor;
871 /** The transaction that owns this cursor */
873 /** The database handle this cursor operates on */
875 /** The database record for this cursor */
877 /** The database auxiliary record for this cursor */
879 /** The @ref mt_dbflag for this database */
880 unsigned char *mc_dbflag;
881 unsigned short mc_snum; /**< number of pushed pages */
882 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
883 /** @defgroup mdb_cursor Cursor Flags
885 * Cursor state flags.
888 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
889 #define C_EOF 0x02 /**< No more data */
890 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
891 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
892 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
893 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
895 unsigned int mc_flags; /**< @ref mdb_cursor */
896 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
897 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
900 /** Context for sorted-dup records.
901 * We could have gone to a fully recursive design, with arbitrarily
902 * deep nesting of sub-databases. But for now we only handle these
903 * levels - main DB, optional sub-DB, sorted-duplicate DB.
905 typedef struct MDB_xcursor {
906 /** A sub-cursor for traversing the Dup DB */
907 MDB_cursor mx_cursor;
908 /** The database record for this Dup DB */
910 /** The auxiliary DB record for this Dup DB */
912 /** The @ref mt_dbflag for this Dup DB */
913 unsigned char mx_dbflag;
916 /** State of FreeDB old pages, stored in the MDB_env */
917 typedef struct MDB_pgstate {
918 txnid_t mf_pglast; /**< ID of last old page record we used */
919 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
920 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
923 /** The database environment. */
925 HANDLE me_fd; /**< The main data file */
926 HANDLE me_lfd; /**< The lock file */
927 HANDLE me_mfd; /**< just for writing the meta pages */
928 /** Failed to update the meta page. Probably an I/O error. */
929 #define MDB_FATAL_ERROR 0x80000000U
930 /** Read-only Filesystem. Allow read access, no locking. */
931 #define MDB_ROFS 0x40000000U
932 /** Some fields are initialized. */
933 #define MDB_ENV_ACTIVE 0x20000000U
934 uint32_t me_flags; /**< @ref mdb_env */
935 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
936 unsigned int me_maxreaders; /**< size of the reader table */
937 unsigned int me_numreaders; /**< max numreaders set by this env */
938 MDB_dbi me_numdbs; /**< number of DBs opened */
939 MDB_dbi me_maxdbs; /**< size of the DB table */
940 pid_t me_pid; /**< process ID of this env */
941 char *me_path; /**< path to the DB files */
942 char *me_map; /**< the memory map of the data file */
943 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
944 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
945 MDB_txn *me_txn; /**< current write transaction */
946 size_t me_mapsize; /**< size of the data memory map */
947 off_t me_size; /**< current file size */
948 pgno_t me_maxpg; /**< me_mapsize / me_psize */
949 MDB_dbx *me_dbxs; /**< array of static DB info */
950 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
951 pthread_key_t me_txkey; /**< thread-key for readers */
952 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
953 # define me_pglast me_pgstate.mf_pglast
954 # define me_pghead me_pgstate.mf_pghead
955 # define me_pgfree me_pgstate.mf_pgfree
956 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
957 /** IDL of pages that became unused in a write txn */
959 /** ID2L of pages that were written during a write txn */
960 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
961 /** Max number of freelist items that can fit in a single overflow page */
962 unsigned int me_maxfree_1pg;
964 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
966 #elif defined(MDB_USE_POSIX_SEM)
967 sem_t *me_rmutex; /* Shared mutexes are not supported */
972 /** Nested transaction */
973 typedef struct MDB_ntxn {
974 MDB_txn mnt_txn; /* the transaction */
975 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
978 /** max number of pages to commit in one writev() call */
979 #define MDB_COMMIT_PAGES 64
980 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
981 #undef MDB_COMMIT_PAGES
982 #define MDB_COMMIT_PAGES IOV_MAX
985 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
986 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
987 static int mdb_page_touch(MDB_cursor *mc);
989 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
990 static int mdb_page_search_root(MDB_cursor *mc,
991 MDB_val *key, int modify);
992 #define MDB_PS_MODIFY 1
993 #define MDB_PS_ROOTONLY 2
994 static int mdb_page_search(MDB_cursor *mc,
995 MDB_val *key, int flags);
996 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
998 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
999 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1000 pgno_t newpgno, unsigned int nflags);
1002 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1003 static int mdb_env_pick_meta(const MDB_env *env);
1004 static int mdb_env_write_meta(MDB_txn *txn);
1005 static void mdb_env_close0(MDB_env *env, int excl);
1007 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1008 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1009 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1010 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1011 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1012 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1013 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1014 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1015 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1017 static int mdb_rebalance(MDB_cursor *mc);
1018 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1020 static void mdb_cursor_pop(MDB_cursor *mc);
1021 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1023 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1024 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1025 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1026 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1027 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1029 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1030 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1032 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1033 static void mdb_xcursor_init0(MDB_cursor *mc);
1034 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1036 static int mdb_drop0(MDB_cursor *mc, int subs);
1037 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1040 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1044 static SECURITY_DESCRIPTOR mdb_null_sd;
1045 static SECURITY_ATTRIBUTES mdb_all_sa;
1046 static int mdb_sec_inited;
1049 /** Return the library version info. */
1051 mdb_version(int *major, int *minor, int *patch)
1053 if (major) *major = MDB_VERSION_MAJOR;
1054 if (minor) *minor = MDB_VERSION_MINOR;
1055 if (patch) *patch = MDB_VERSION_PATCH;
1056 return MDB_VERSION_STRING;
1059 /** Table of descriptions for MDB @ref errors */
1060 static char *const mdb_errstr[] = {
1061 "MDB_KEYEXIST: Key/data pair already exists",
1062 "MDB_NOTFOUND: No matching key/data pair found",
1063 "MDB_PAGE_NOTFOUND: Requested page not found",
1064 "MDB_CORRUPTED: Located page was wrong type",
1065 "MDB_PANIC: Update of meta page failed",
1066 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1067 "MDB_INVALID: File is not an MDB file",
1068 "MDB_MAP_FULL: Environment mapsize limit reached",
1069 "MDB_DBS_FULL: Environment maxdbs limit reached",
1070 "MDB_READERS_FULL: Environment maxreaders limit reached",
1071 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1072 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1073 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1074 "MDB_PAGE_FULL: Internal error - page has no more space",
1075 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1076 "MDB_INCOMPATIBLE: Operation is incompatible with database",
1080 mdb_strerror(int err)
1084 return ("Successful return: 0");
1086 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1087 i = err - MDB_KEYEXIST;
1088 return mdb_errstr[i];
1091 return strerror(err);
1095 /** Display a key in hexadecimal and return the address of the result.
1096 * @param[in] key the key to display
1097 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1098 * @return The key in hexadecimal form.
1101 mdb_dkey(MDB_val *key, char *buf)
1104 unsigned char *c = key->mv_data;
1110 if (key->mv_size > MDB_MAXKEYSIZE)
1111 return "MDB_MAXKEYSIZE";
1112 /* may want to make this a dynamic check: if the key is mostly
1113 * printable characters, print it as-is instead of converting to hex.
1117 for (i=0; i<key->mv_size; i++)
1118 ptr += sprintf(ptr, "%02x", *c++);
1120 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1125 /** Display all the keys in the page. */
1127 mdb_page_list(MDB_page *mp)
1130 unsigned int i, nkeys, nsize;
1134 nkeys = NUMKEYS(mp);
1135 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1136 for (i=0; i<nkeys; i++) {
1137 node = NODEPTR(mp, i);
1138 key.mv_size = node->mn_ksize;
1139 key.mv_data = node->mn_data;
1140 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1141 if (IS_BRANCH(mp)) {
1142 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1145 if (F_ISSET(node->mn_flags, F_BIGDATA))
1146 nsize += sizeof(pgno_t);
1148 nsize += NODEDSZ(node);
1149 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1155 mdb_cursor_chk(MDB_cursor *mc)
1161 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1162 for (i=0; i<mc->mc_top; i++) {
1164 node = NODEPTR(mp, mc->mc_ki[i]);
1165 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1168 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1174 /** Count all the pages in each DB and in the freelist
1175 * and make sure it matches the actual number of pages
1178 static void mdb_audit(MDB_txn *txn)
1182 MDB_ID freecount, count;
1187 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1188 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1189 freecount += *(MDB_ID *)data.mv_data;
1192 for (i = 0; i<txn->mt_numdbs; i++) {
1193 MDB_xcursor mx, *mxp;
1194 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1195 mdb_cursor_init(&mc, txn, i, mxp);
1196 if (txn->mt_dbs[i].md_root == P_INVALID)
1198 count += txn->mt_dbs[i].md_branch_pages +
1199 txn->mt_dbs[i].md_leaf_pages +
1200 txn->mt_dbs[i].md_overflow_pages;
1201 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1202 mdb_page_search(&mc, NULL, 0);
1206 mp = mc.mc_pg[mc.mc_top];
1207 for (j=0; j<NUMKEYS(mp); j++) {
1208 MDB_node *leaf = NODEPTR(mp, j);
1209 if (leaf->mn_flags & F_SUBDATA) {
1211 memcpy(&db, NODEDATA(leaf), sizeof(db));
1212 count += db.md_branch_pages + db.md_leaf_pages +
1213 db.md_overflow_pages;
1217 while (mdb_cursor_sibling(&mc, 1) == 0);
1220 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1221 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1222 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1228 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1230 return txn->mt_dbxs[dbi].md_cmp(a, b);
1234 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1236 if (txn->mt_dbxs[dbi].md_dcmp)
1237 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1239 return EINVAL; /* too bad you can't distinguish this from a valid result */
1242 /** Allocate a single page.
1243 * Re-use old malloc'd pages first, otherwise just malloc.
1246 mdb_page_malloc(MDB_cursor *mc) {
1248 size_t sz = mc->mc_txn->mt_env->me_psize;
1249 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1250 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1251 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1252 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1253 } else if ((ret = malloc(sz)) != NULL) {
1254 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1260 mdb_page_free(MDB_env *env, MDB_page *mp)
1262 mp->mp_next = env->me_dpages;
1263 VGMEMP_FREE(env, mp);
1264 env->me_dpages = mp;
1267 /** Allocate pages for writing.
1268 * If there are free pages available from older transactions, they
1269 * will be re-used first. Otherwise a new page will be allocated.
1270 * @param[in] mc cursor A cursor handle identifying the transaction and
1271 * database for which we are allocating.
1272 * @param[in] num the number of pages to allocate.
1273 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1274 * will always be satisfied by a single contiguous chunk of memory.
1275 * @return 0 on success, non-zero on failure.
1278 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1280 MDB_txn *txn = mc->mc_txn;
1282 pgno_t pgno = P_INVALID;
1284 txnid_t oldest = 0, last;
1289 /* If our dirty list is already full, we can't do anything */
1290 if (txn->mt_dirty_room == 0)
1291 return MDB_TXN_FULL;
1293 /* The free list won't have any content at all until txn 2 has
1294 * committed. The pages freed by txn 2 will be unreferenced
1295 * after txn 3 commits, and so will be safe to re-use in txn 4.
1297 if (txn->mt_txnid > 3) {
1298 if (!txn->mt_env->me_pghead &&
1299 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1300 /* See if there's anything in the free DB */
1307 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1308 if (!txn->mt_env->me_pglast) {
1309 mdb_page_search(&m2, NULL, 0);
1310 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1311 kptr = (txnid_t *)NODEKEY(leaf);
1316 last = txn->mt_env->me_pglast + 1;
1318 key.mv_data = &last;
1319 key.mv_size = sizeof(last);
1320 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1323 last = *(txnid_t *)key.mv_data;
1329 oldest = txn->mt_txnid - 1;
1330 nr = txn->mt_env->me_txns->mti_numreaders;
1331 r = txn->mt_env->me_txns->mti_readers;
1332 for (i=0; i<nr; i++) {
1333 if (!r[i].mr_pid) continue;
1340 if (oldest > last) {
1341 /* It's usable, grab it.
1345 if (!txn->mt_env->me_pglast) {
1346 mdb_node_read(txn, leaf, &data);
1348 idl = (MDB_ID *) data.mv_data;
1349 /* We might have a zero-length IDL due to freelist growth
1350 * during a prior commit
1353 txn->mt_env->me_pglast = last;
1356 mop = malloc(MDB_IDL_SIZEOF(idl));
1359 txn->mt_env->me_pglast = last;
1360 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1361 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1366 DPRINTF("IDL read txn %zu root %zu num %zu",
1367 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1368 for (i=0; i<idl[0]; i++) {
1369 DPRINTF("IDL %zu", idl[i+1]);
1376 if (txn->mt_env->me_pghead) {
1377 pgno_t *mop = txn->mt_env->me_pghead;
1380 int retry = 1, readit = 0, n2 = num-1;
1381 unsigned int i, j, k;
1383 /* If current list is too short, must fetch more and coalesce */
1384 if (mop[0] < (unsigned)num)
1387 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1389 /* If on freelist, don't try to read more. If what we have
1390 * right now isn't enough just use new pages.
1391 * TODO: get all of this working. Many circular dependencies...
1393 if (mc->mc_dbi == FREE_DBI) {
1401 last = txn->mt_env->me_pglast + 1;
1403 /* We haven't hit the readers list yet? */
1409 oldest = txn->mt_txnid - 1;
1410 nr = txn->mt_env->me_txns->mti_numreaders;
1411 r = txn->mt_env->me_txns->mti_readers;
1412 for (i=0; i<nr; i++) {
1413 if (!r[i].mr_pid) continue;
1420 /* There's nothing we can use on the freelist */
1421 if (oldest - last < 1)
1424 key.mv_data = &last;
1425 key.mv_size = sizeof(last);
1426 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1428 if (rc == MDB_NOTFOUND)
1432 last = *(txnid_t*)key.mv_data;
1435 idl = (MDB_ID *) data.mv_data;
1436 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1439 /* merge in sorted order */
1440 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1442 while (i>0 || j>0) {
1443 if (i && idl[i] < mop[j])
1444 mop2[k--] = idl[i--];
1446 mop2[k--] = mop[j--];
1448 txn->mt_env->me_pglast = last;
1449 free(txn->mt_env->me_pgfree);
1450 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1452 /* Keep trying to read until we have enough */
1453 if (mop[0] < (unsigned)num) {
1458 /* current list has enough pages, but are they contiguous? */
1459 for (i=mop[0]; i>=(unsigned)num; i--) {
1460 if (mop[i-n2] == mop[i] + n2) {
1463 /* move any stragglers down */
1464 for (j=i+num; j<=mop[0]; j++)
1471 /* Stop if we succeeded, or no retries */
1472 if (!retry || pgno != P_INVALID)
1478 /* peel pages off tail, so we only have to truncate the list */
1479 pgno = MDB_IDL_LAST(mop);
1482 if (MDB_IDL_IS_ZERO(mop)) {
1483 free(txn->mt_env->me_pgfree);
1484 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1489 if (pgno == P_INVALID) {
1490 /* DB size is maxed out */
1491 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1492 DPUTS("DB size maxed out");
1493 return MDB_MAP_FULL;
1496 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1497 if (pgno == P_INVALID) {
1498 pgno = txn->mt_next_pgno;
1499 txn->mt_next_pgno += num;
1501 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1504 if (txn->mt_env->me_dpages && num == 1) {
1505 np = txn->mt_env->me_dpages;
1506 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1507 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1508 txn->mt_env->me_dpages = np->mp_next;
1510 size_t sz = txn->mt_env->me_psize * num;
1511 if ((np = malloc(sz)) == NULL)
1513 VGMEMP_ALLOC(txn->mt_env, np, sz);
1515 if (pgno == P_INVALID) {
1516 np->mp_pgno = txn->mt_next_pgno;
1517 txn->mt_next_pgno += num;
1522 mid.mid = np->mp_pgno;
1524 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1525 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1527 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1529 txn->mt_dirty_room--;
1535 /** Copy a page: avoid copying unused portions of the page.
1536 * @param[in] dst page to copy into
1537 * @param[in] src page to copy from
1540 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1542 dst->mp_flags = src->mp_flags | P_DIRTY;
1543 dst->mp_pages = src->mp_pages;
1545 if (IS_LEAF2(src)) {
1546 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1548 unsigned int i, nkeys = NUMKEYS(src);
1549 for (i=0; i<nkeys; i++)
1550 dst->mp_ptrs[i] = src->mp_ptrs[i];
1551 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1552 psize - src->mp_upper);
1556 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1557 * @param[in] mc cursor pointing to the page to be touched
1558 * @return 0 on success, non-zero on failure.
1561 mdb_page_touch(MDB_cursor *mc)
1563 MDB_page *mp = mc->mc_pg[mc->mc_top];
1567 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1569 if ((rc = mdb_page_alloc(mc, 1, &np)))
1571 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1572 assert(mp->mp_pgno != np->mp_pgno);
1573 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1575 /* If page isn't full, just copy the used portion */
1576 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1579 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1581 np->mp_flags |= P_DIRTY;
1586 /* Adjust other cursors pointing to mp */
1587 if (mc->mc_flags & C_SUB) {
1588 MDB_cursor *m2, *m3;
1589 MDB_dbi dbi = mc->mc_dbi-1;
1591 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1592 if (m2 == mc) continue;
1593 m3 = &m2->mc_xcursor->mx_cursor;
1594 if (m3->mc_snum < mc->mc_snum) continue;
1595 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1596 m3->mc_pg[mc->mc_top] = mp;
1602 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1603 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1604 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1605 m2->mc_pg[mc->mc_top] = mp;
1609 mc->mc_pg[mc->mc_top] = mp;
1610 /** If this page has a parent, update the parent to point to
1614 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1616 mc->mc_db->md_root = mp->mp_pgno;
1617 } else if (mc->mc_txn->mt_parent) {
1620 /* If txn has a parent, make sure the page is in our
1623 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1624 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1625 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1626 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1627 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1628 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1629 mc->mc_pg[mc->mc_top] = mp;
1634 assert(mc->mc_txn->mt_u.dirty_list[0].mid < MDB_IDL_UM_MAX);
1636 np = mdb_page_malloc(mc);
1639 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1640 mid.mid = np->mp_pgno;
1642 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1650 mdb_env_sync(MDB_env *env, int force)
1653 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1654 if (env->me_flags & MDB_WRITEMAP) {
1655 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1656 ? MS_ASYNC : MS_SYNC;
1657 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1660 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1664 if (MDB_FDATASYNC(env->me_fd))
1671 /** Make shadow copies of all of parent txn's cursors */
1673 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1675 MDB_cursor *mc, *m2;
1676 unsigned int i, j, size;
1678 for (i=0;i<src->mt_numdbs; i++) {
1679 if (src->mt_cursors[i]) {
1680 size = sizeof(MDB_cursor);
1681 if (src->mt_cursors[i]->mc_xcursor)
1682 size += sizeof(MDB_xcursor);
1683 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1690 mc->mc_db = &dst->mt_dbs[i];
1691 mc->mc_dbx = m2->mc_dbx;
1692 mc->mc_dbflag = &dst->mt_dbflags[i];
1693 mc->mc_snum = m2->mc_snum;
1694 mc->mc_top = m2->mc_top;
1695 mc->mc_flags = m2->mc_flags | C_SHADOW;
1696 for (j=0; j<mc->mc_snum; j++) {
1697 mc->mc_pg[j] = m2->mc_pg[j];
1698 mc->mc_ki[j] = m2->mc_ki[j];
1700 if (m2->mc_xcursor) {
1701 MDB_xcursor *mx, *mx2;
1702 mx = (MDB_xcursor *)(mc+1);
1703 mc->mc_xcursor = mx;
1704 mx2 = m2->mc_xcursor;
1705 mx->mx_db = mx2->mx_db;
1706 mx->mx_dbx = mx2->mx_dbx;
1707 mx->mx_dbflag = mx2->mx_dbflag;
1708 mx->mx_cursor.mc_txn = dst;
1709 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1710 mx->mx_cursor.mc_db = &mx->mx_db;
1711 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1712 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1713 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1714 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1715 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1716 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1717 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1718 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1721 mc->mc_xcursor = NULL;
1723 mc->mc_next = dst->mt_cursors[i];
1724 dst->mt_cursors[i] = mc;
1731 /** Merge shadow cursors back into parent's */
1733 mdb_cursor_merge(MDB_txn *txn)
1736 for (i=0; i<txn->mt_numdbs; i++) {
1737 if (txn->mt_cursors[i]) {
1739 while ((mc = txn->mt_cursors[i])) {
1740 txn->mt_cursors[i] = mc->mc_next;
1741 if (mc->mc_flags & C_SHADOW) {
1742 MDB_cursor *m2 = mc->mc_orig;
1744 m2->mc_snum = mc->mc_snum;
1745 m2->mc_top = mc->mc_top;
1746 for (j=0; j<mc->mc_snum; j++) {
1747 m2->mc_pg[j] = mc->mc_pg[j];
1748 m2->mc_ki[j] = mc->mc_ki[j];
1751 if (mc->mc_flags & C_ALLOCD)
1759 mdb_txn_reset0(MDB_txn *txn);
1761 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1762 * @param[in] txn the transaction handle to initialize
1763 * @return 0 on success, non-zero on failure. This can only
1764 * fail for read-only transactions, and then only if the
1765 * reader table is full.
1768 mdb_txn_renew0(MDB_txn *txn)
1770 MDB_env *env = txn->mt_env;
1775 txn->mt_numdbs = env->me_numdbs;
1776 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1778 if (txn->mt_flags & MDB_TXN_RDONLY) {
1779 if (env->me_flags & MDB_ROFS) {
1780 i = mdb_env_pick_meta(env);
1781 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1782 txn->mt_u.reader = NULL;
1784 MDB_reader *r = pthread_getspecific(env->me_txkey);
1786 pid_t pid = env->me_pid;
1787 pthread_t tid = pthread_self();
1790 for (i=0; i<env->me_txns->mti_numreaders; i++)
1791 if (env->me_txns->mti_readers[i].mr_pid == 0)
1793 if (i == env->me_maxreaders) {
1794 UNLOCK_MUTEX_R(env);
1795 return MDB_READERS_FULL;
1797 env->me_txns->mti_readers[i].mr_pid = pid;
1798 env->me_txns->mti_readers[i].mr_tid = tid;
1799 if (i >= env->me_txns->mti_numreaders)
1800 env->me_txns->mti_numreaders = i+1;
1801 /* Save numreaders for un-mutexed mdb_env_close() */
1802 env->me_numreaders = env->me_txns->mti_numreaders;
1803 UNLOCK_MUTEX_R(env);
1804 r = &env->me_txns->mti_readers[i];
1805 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1806 env->me_txns->mti_readers[i].mr_pid = 0;
1810 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1811 txn->mt_u.reader = r;
1813 txn->mt_toggle = txn->mt_txnid & 1;
1814 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1818 txn->mt_txnid = env->me_txns->mti_txnid;
1819 txn->mt_toggle = txn->mt_txnid & 1;
1820 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1823 if (txn->mt_txnid == mdb_debug_start)
1826 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1827 txn->mt_u.dirty_list = env->me_dirty_list;
1828 txn->mt_u.dirty_list[0].mid = 0;
1829 txn->mt_free_pgs = env->me_free_pgs;
1830 txn->mt_free_pgs[0] = 0;
1834 /* Copy the DB info and flags */
1835 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1836 for (i=2; i<txn->mt_numdbs; i++)
1837 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1838 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1839 if (txn->mt_numdbs > 2)
1840 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1842 if (env->me_maxpg < txn->mt_next_pgno) {
1843 mdb_txn_reset0(txn);
1844 return MDB_MAP_RESIZED;
1851 mdb_txn_renew(MDB_txn *txn)
1855 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1858 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1859 DPUTS("environment had fatal error, must shutdown!");
1863 rc = mdb_txn_renew0(txn);
1864 if (rc == MDB_SUCCESS) {
1865 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1866 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1867 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1873 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1877 int rc, size, tsize = sizeof(MDB_txn);
1879 if (env->me_flags & MDB_FATAL_ERROR) {
1880 DPUTS("environment had fatal error, must shutdown!");
1883 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1886 /* Nested transactions: Max 1 child, write txns only, no writemap */
1887 if (parent->mt_child ||
1888 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1889 (env->me_flags & MDB_WRITEMAP))
1893 tsize = sizeof(MDB_ntxn);
1895 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1896 if (!(flags & MDB_RDONLY))
1897 size += env->me_maxdbs * sizeof(MDB_cursor *);
1899 if ((txn = calloc(1, size)) == NULL) {
1900 DPRINTF("calloc: %s", strerror(ErrCode()));
1903 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1904 if (flags & MDB_RDONLY) {
1905 txn->mt_flags |= MDB_TXN_RDONLY;
1906 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1908 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1909 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1914 txn->mt_free_pgs = mdb_midl_alloc();
1915 if (!txn->mt_free_pgs) {
1919 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1920 if (!txn->mt_u.dirty_list) {
1921 free(txn->mt_free_pgs);
1925 txn->mt_txnid = parent->mt_txnid;
1926 txn->mt_toggle = parent->mt_toggle;
1927 txn->mt_dirty_room = parent->mt_dirty_room;
1928 txn->mt_u.dirty_list[0].mid = 0;
1929 txn->mt_free_pgs[0] = 0;
1930 txn->mt_next_pgno = parent->mt_next_pgno;
1931 parent->mt_child = txn;
1932 txn->mt_parent = parent;
1933 txn->mt_numdbs = parent->mt_numdbs;
1934 txn->mt_dbxs = parent->mt_dbxs;
1935 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1936 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1938 ntxn = (MDB_ntxn *)txn;
1939 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1940 if (env->me_pghead) {
1941 size = MDB_IDL_SIZEOF(env->me_pghead);
1942 env->me_pghead = malloc(size);
1944 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1948 env->me_pgfree = env->me_pghead;
1950 rc = mdb_cursor_shadow(parent, txn);
1952 mdb_txn_reset0(txn);
1954 rc = mdb_txn_renew0(txn);
1960 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1961 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1962 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1968 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1969 * @param[in] txn the transaction handle to reset
1972 mdb_txn_reset0(MDB_txn *txn)
1974 MDB_env *env = txn->mt_env;
1976 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1977 if (!(env->me_flags & MDB_ROFS))
1978 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1983 /* close(free) all cursors */
1984 for (i=0; i<txn->mt_numdbs; i++) {
1985 if (txn->mt_cursors[i]) {
1987 while ((mc = txn->mt_cursors[i])) {
1988 txn->mt_cursors[i] = mc->mc_next;
1989 if (mc->mc_flags & C_ALLOCD)
1995 if (!(env->me_flags & MDB_WRITEMAP)) {
1996 /* return all dirty pages to dpage list */
1997 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1998 dp = txn->mt_u.dirty_list[i].mptr;
1999 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2000 mdb_page_free(txn->mt_env, dp);
2002 /* large pages just get freed directly */
2003 VGMEMP_FREE(txn->mt_env, dp);
2009 free(env->me_pgfree);
2011 if (txn->mt_parent) {
2012 txn->mt_parent->mt_child = NULL;
2013 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2014 mdb_midl_free(txn->mt_free_pgs);
2015 free(txn->mt_u.dirty_list);
2018 if (mdb_midl_shrink(&txn->mt_free_pgs))
2019 env->me_free_pgs = txn->mt_free_pgs;
2022 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2023 txn->mt_env->me_pglast = 0;
2026 /* The writer mutex was locked in mdb_txn_begin. */
2027 UNLOCK_MUTEX_W(env);
2032 mdb_txn_reset(MDB_txn *txn)
2037 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2038 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2039 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2041 mdb_txn_reset0(txn);
2045 mdb_txn_abort(MDB_txn *txn)
2050 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2051 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2052 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2055 mdb_txn_abort(txn->mt_child);
2057 mdb_txn_reset0(txn);
2062 mdb_txn_commit(MDB_txn *txn)
2070 pgno_t next, freecnt;
2071 txnid_t oldpg_txnid, id;
2074 assert(txn != NULL);
2075 assert(txn->mt_env != NULL);
2077 if (txn->mt_child) {
2078 mdb_txn_commit(txn->mt_child);
2079 txn->mt_child = NULL;
2084 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2085 if (txn->mt_numdbs > env->me_numdbs) {
2086 /* update the DB flags */
2088 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2089 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2096 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2097 DPUTS("error flag is set, can't commit");
2099 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2104 if (txn->mt_parent) {
2105 MDB_txn *parent = txn->mt_parent;
2109 /* Append our free list to parent's */
2110 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2114 mdb_midl_free(txn->mt_free_pgs);
2116 parent->mt_next_pgno = txn->mt_next_pgno;
2117 parent->mt_flags = txn->mt_flags;
2119 /* Merge (and close) our cursors with parent's */
2120 mdb_cursor_merge(txn);
2122 /* Update parent's DB table. */
2123 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2124 memcpy(parent->mt_dbflags, txn->mt_dbflags, txn->mt_numdbs);
2125 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2127 dst = txn->mt_parent->mt_u.dirty_list;
2128 src = txn->mt_u.dirty_list;
2129 /* Find len = length of merging our dirty list with parent's */
2131 dst[0].mid = 0; /* simplify loops */
2132 if (parent->mt_parent) {
2133 len = x + src[0].mid;
2134 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2135 for (i = x; y && i; y--) {
2136 pgno_t yp = src[y].mid;
2137 while (yp < dst[i].mid)
2139 if (yp == dst[i].mid) {
2144 } else { /* Simplify the above for single-ancestor case */
2145 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2147 /* Merge our dirty list with parent's */
2149 for (i = len; y; dst[i--] = src[y--]) {
2150 pgno_t yp = src[y].mid;
2151 while (yp < dst[x].mid)
2152 dst[i--] = dst[x--];
2153 if (yp == dst[x].mid)
2154 free(dst[x--].mptr);
2158 free(txn->mt_u.dirty_list);
2159 parent->mt_dirty_room = txn->mt_dirty_room;
2161 txn->mt_parent->mt_child = NULL;
2162 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2167 if (txn != env->me_txn) {
2168 DPUTS("attempt to commit unknown transaction");
2173 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2176 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2177 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2179 /* Update DB root pointers */
2180 if (txn->mt_numdbs > 2) {
2183 data.mv_size = sizeof(MDB_db);
2185 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2186 for (i = 2; i < txn->mt_numdbs; i++) {
2187 if (txn->mt_dbflags[i] & DB_DIRTY) {
2188 data.mv_data = &txn->mt_dbs[i];
2189 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2196 /* Save the freelist as of this transaction to the freeDB. This
2197 * can change the freelist, so keep trying until it stabilizes.
2199 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2200 * except the code below can decrease env->me_pglast to split pghead.
2201 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2202 * can only appear in env->me_pghead when env->me_pglast increases.
2203 * Until then, the me_pghead pointer won't move but can become NULL.
2206 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2207 oldpg_txnid = id = 0;
2210 /* should only be one record now */
2211 if (env->me_pghead || env->me_pglast) {
2212 /* make sure first page of freeDB is touched and on freelist */
2213 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2214 if (rc && rc != MDB_NOTFOUND) {
2221 /* Delete IDLs we used from the free list */
2222 if (env->me_pglast) {
2227 rc = mdb_cursor_first(&mc, &key, NULL);
2230 oldpg_txnid = *(txnid_t *)key.mv_data;
2232 assert(oldpg_txnid <= env->me_pglast);
2234 rc = mdb_cursor_del(&mc, 0);
2237 } while (oldpg_txnid < env->me_pglast);
2240 /* Save IDL of pages freed by this txn, to freeDB */
2242 if (freecnt != txn->mt_free_pgs[0]) {
2245 /* make sure last page of freeDB is touched and on freelist */
2246 key.mv_size = MDB_MAXKEYSIZE+1;
2248 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2249 if (rc && rc != MDB_NOTFOUND)
2255 MDB_IDL idl = txn->mt_free_pgs;
2256 mdb_midl_sort(txn->mt_free_pgs);
2257 DPRINTF("IDL write txn %zu root %zu num %zu",
2258 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2259 for (i=1; i<=idl[0]; i++) {
2260 DPRINTF("IDL %zu", idl[i]);
2264 /* write to last page of freeDB */
2265 key.mv_size = sizeof(pgno_t);
2266 key.mv_data = &txn->mt_txnid;
2267 /* The free list can still grow during this call,
2268 * despite the pre-emptive touches above. So retry
2269 * until the reserved space remains big enough.
2272 assert(freecnt < txn->mt_free_pgs[0]);
2273 freecnt = txn->mt_free_pgs[0];
2274 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2275 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2278 } while (freecnt != txn->mt_free_pgs[0]);
2279 mdb_midl_sort(txn->mt_free_pgs);
2280 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2281 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2282 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2285 /* Put back page numbers we took from freeDB but did not use */
2286 if (env->me_pghead) {
2291 mop = env->me_pghead;
2292 id = env->me_pglast;
2293 key.mv_size = sizeof(id);
2295 /* These steps may grow the freelist again
2296 * due to freed overflow pages...
2301 if (orig > env->me_maxfree_1pg && id > 4)
2302 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2303 data.mv_size = (orig + 1) * sizeof(pgno_t);
2304 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2307 assert(!env->me_pghead || env->me_pglast);
2308 /* mop could have been used again here */
2309 if (id != env->me_pglast || env->me_pghead == NULL)
2310 goto again; /* was completely used up */
2311 assert(mop == env->me_pghead);
2312 } while (mop[0] < orig && --i);
2313 memcpy(data.mv_data, mop, data.mv_size);
2316 *(pgno_t *)data.mv_data = orig;
2317 mop[orig] = mop[0] - orig;
2318 env->me_pghead = mop += orig;
2319 /* Save more oldpages at the previous txnid. */
2320 assert(env->me_pglast == id && id == oldpg_txnid);
2321 env->me_pglast = --oldpg_txnid;
2325 /* Check for growth of freelist again */
2326 if (freecnt != txn->mt_free_pgs[0])
2329 free(env->me_pgfree);
2330 env->me_pghead = env->me_pgfree = NULL;
2332 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2333 if (mdb_midl_shrink(&txn->mt_free_pgs))
2334 env->me_free_pgs = txn->mt_free_pgs;
2341 if (env->me_flags & MDB_WRITEMAP) {
2342 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2343 dp = txn->mt_u.dirty_list[i].mptr;
2344 /* clear dirty flag */
2345 dp->mp_flags &= ~P_DIRTY;
2346 txn->mt_u.dirty_list[i].mid = 0;
2348 txn->mt_u.dirty_list[0].mid = 0;
2352 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2358 /* Windows actually supports scatter/gather I/O, but only on
2359 * unbuffered file handles. Since we're relying on the OS page
2360 * cache for all our data, that's self-defeating. So we just
2361 * write pages one at a time. We use the ov structure to set
2362 * the write offset, to at least save the overhead of a Seek
2366 memset(&ov, 0, sizeof(ov));
2367 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2369 dp = txn->mt_u.dirty_list[i].mptr;
2370 DPRINTF("committing page %zu", dp->mp_pgno);
2371 size = dp->mp_pgno * env->me_psize;
2372 ov.Offset = size & 0xffffffff;
2373 ov.OffsetHigh = size >> 16;
2374 ov.OffsetHigh >>= 16;
2375 /* clear dirty flag */
2376 dp->mp_flags &= ~P_DIRTY;
2377 wsize = env->me_psize;
2378 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2379 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2382 DPRINTF("WriteFile: %d", n);
2389 struct iovec iov[MDB_COMMIT_PAGES];
2393 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2394 dp = txn->mt_u.dirty_list[i].mptr;
2395 if (dp->mp_pgno != next) {
2397 rc = writev(env->me_fd, iov, n);
2401 DPUTS("short write, filesystem full?");
2403 DPRINTF("writev: %s", strerror(n));
2410 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2413 DPRINTF("committing page %zu", dp->mp_pgno);
2414 iov[n].iov_len = env->me_psize;
2415 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2416 iov[n].iov_base = (char *)dp;
2417 size += iov[n].iov_len;
2418 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2419 /* clear dirty flag */
2420 dp->mp_flags &= ~P_DIRTY;
2421 if (++n >= MDB_COMMIT_PAGES) {
2431 rc = writev(env->me_fd, iov, n);
2435 DPUTS("short write, filesystem full?");
2437 DPRINTF("writev: %s", strerror(n));
2444 /* Drop the dirty pages.
2446 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2447 dp = txn->mt_u.dirty_list[i].mptr;
2448 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2449 mdb_page_free(txn->mt_env, dp);
2451 VGMEMP_FREE(txn->mt_env, dp);
2454 txn->mt_u.dirty_list[i].mid = 0;
2456 txn->mt_u.dirty_list[0].mid = 0;
2459 if ((n = mdb_env_sync(env, 0)) != 0 ||
2460 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2468 if (txn->mt_numdbs > env->me_numdbs) {
2469 /* update the DB flags */
2471 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2472 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2476 UNLOCK_MUTEX_W(env);
2482 /** Read the environment parameters of a DB environment before
2483 * mapping it into memory.
2484 * @param[in] env the environment handle
2485 * @param[out] meta address of where to store the meta information
2486 * @return 0 on success, non-zero on failure.
2489 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2496 /* We don't know the page size yet, so use a minimum value.
2497 * Read both meta pages so we can use the latest one.
2500 for (i=0; i<2; i++) {
2502 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2504 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2509 else if (rc != MDB_PAGESIZE) {
2513 DPRINTF("read: %s", strerror(err));
2517 p = (MDB_page *)&pbuf;
2519 if (!F_ISSET(p->mp_flags, P_META)) {
2520 DPRINTF("page %zu not a meta page", p->mp_pgno);
2525 if (m->mm_magic != MDB_MAGIC) {
2526 DPUTS("meta has invalid magic");
2530 if (m->mm_version != MDB_VERSION) {
2531 DPRINTF("database is version %u, expected version %u",
2532 m->mm_version, MDB_VERSION);
2533 return MDB_VERSION_MISMATCH;
2537 if (m->mm_txnid > meta->mm_txnid)
2538 memcpy(meta, m, sizeof(*m));
2540 memcpy(meta, m, sizeof(*m));
2542 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2544 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2552 /** Write the environment parameters of a freshly created DB environment.
2553 * @param[in] env the environment handle
2554 * @param[out] meta address of where to store the meta information
2555 * @return 0 on success, non-zero on failure.
2558 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2565 DPUTS("writing new meta page");
2567 GET_PAGESIZE(psize);
2569 meta->mm_magic = MDB_MAGIC;
2570 meta->mm_version = MDB_VERSION;
2571 meta->mm_mapsize = env->me_mapsize;
2572 meta->mm_psize = psize;
2573 meta->mm_last_pg = 1;
2574 meta->mm_flags = env->me_flags & 0xffff;
2575 meta->mm_flags |= MDB_INTEGERKEY;
2576 meta->mm_dbs[0].md_root = P_INVALID;
2577 meta->mm_dbs[1].md_root = P_INVALID;
2579 p = calloc(2, psize);
2581 p->mp_flags = P_META;
2584 memcpy(m, meta, sizeof(*meta));
2586 q = (MDB_page *)((char *)p + psize);
2589 q->mp_flags = P_META;
2592 memcpy(m, meta, sizeof(*meta));
2597 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2598 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2599 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2602 lseek(env->me_fd, 0, SEEK_SET);
2603 rc = write(env->me_fd, p, psize * 2);
2604 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2610 /** Update the environment info to commit a transaction.
2611 * @param[in] txn the transaction that's being committed
2612 * @return 0 on success, non-zero on failure.
2615 mdb_env_write_meta(MDB_txn *txn)
2618 MDB_meta meta, metab, *mp;
2620 int rc, len, toggle;
2627 assert(txn != NULL);
2628 assert(txn->mt_env != NULL);
2630 toggle = !txn->mt_toggle;
2631 DPRINTF("writing meta page %d for root page %zu",
2632 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2635 mp = env->me_metas[toggle];
2637 if (env->me_flags & MDB_WRITEMAP) {
2638 /* Persist any increases of mapsize config */
2639 if (env->me_mapsize > mp->mm_mapsize)
2640 mp->mm_mapsize = env->me_mapsize;
2641 mp->mm_dbs[0] = txn->mt_dbs[0];
2642 mp->mm_dbs[1] = txn->mt_dbs[1];
2643 mp->mm_last_pg = txn->mt_next_pgno - 1;
2644 mp->mm_txnid = txn->mt_txnid;
2645 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2646 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2649 ptr += env->me_psize;
2650 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2657 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2658 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2660 ptr = (char *)&meta;
2661 if (env->me_mapsize > mp->mm_mapsize) {
2662 /* Persist any increases of mapsize config */
2663 meta.mm_mapsize = env->me_mapsize;
2664 off = offsetof(MDB_meta, mm_mapsize);
2666 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2668 len = sizeof(MDB_meta) - off;
2671 meta.mm_dbs[0] = txn->mt_dbs[0];
2672 meta.mm_dbs[1] = txn->mt_dbs[1];
2673 meta.mm_last_pg = txn->mt_next_pgno - 1;
2674 meta.mm_txnid = txn->mt_txnid;
2677 off += env->me_psize;
2680 /* Write to the SYNC fd */
2681 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2682 env->me_fd : env->me_mfd;
2685 memset(&ov, 0, sizeof(ov));
2687 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2690 rc = pwrite(mfd, ptr, len, off);
2695 DPUTS("write failed, disk error?");
2696 /* On a failure, the pagecache still contains the new data.
2697 * Write some old data back, to prevent it from being used.
2698 * Use the non-SYNC fd; we know it will fail anyway.
2700 meta.mm_last_pg = metab.mm_last_pg;
2701 meta.mm_txnid = metab.mm_txnid;
2703 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2705 r2 = pwrite(env->me_fd, ptr, len, off);
2708 env->me_flags |= MDB_FATAL_ERROR;
2712 /* Memory ordering issues are irrelevant; since the entire writer
2713 * is wrapped by wmutex, all of these changes will become visible
2714 * after the wmutex is unlocked. Since the DB is multi-version,
2715 * readers will get consistent data regardless of how fresh or
2716 * how stale their view of these values is.
2718 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2723 /** Check both meta pages to see which one is newer.
2724 * @param[in] env the environment handle
2725 * @return meta toggle (0 or 1).
2728 mdb_env_pick_meta(const MDB_env *env)
2730 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2734 mdb_env_create(MDB_env **env)
2738 e = calloc(1, sizeof(MDB_env));
2742 e->me_free_pgs = mdb_midl_alloc();
2743 if (!e->me_free_pgs) {
2747 e->me_maxreaders = DEFAULT_READERS;
2749 e->me_fd = INVALID_HANDLE_VALUE;
2750 e->me_lfd = INVALID_HANDLE_VALUE;
2751 e->me_mfd = INVALID_HANDLE_VALUE;
2752 #ifdef MDB_USE_POSIX_SEM
2753 e->me_rmutex = SEM_FAILED;
2754 e->me_wmutex = SEM_FAILED;
2756 e->me_pid = getpid();
2757 VGMEMP_CREATE(e,0,0);
2763 mdb_env_set_mapsize(MDB_env *env, size_t size)
2767 env->me_mapsize = size;
2769 env->me_maxpg = env->me_mapsize / env->me_psize;
2774 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2778 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2783 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2785 if (env->me_map || readers < 1)
2787 env->me_maxreaders = readers;
2792 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2794 if (!env || !readers)
2796 *readers = env->me_maxreaders;
2800 /** Further setup required for opening an MDB environment
2803 mdb_env_open2(MDB_env *env)
2805 unsigned int flags = env->me_flags;
2806 int i, newenv = 0, prot;
2810 memset(&meta, 0, sizeof(meta));
2812 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2815 DPUTS("new mdbenv");
2819 /* Was a mapsize configured? */
2820 if (!env->me_mapsize) {
2821 /* If this is a new environment, take the default,
2822 * else use the size recorded in the existing env.
2824 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2825 } else if (env->me_mapsize < meta.mm_mapsize) {
2826 /* If the configured size is smaller, make sure it's
2827 * still big enough. Silently round up to minimum if not.
2829 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2830 if (env->me_mapsize < minsize)
2831 env->me_mapsize = minsize;
2837 LONG sizelo, sizehi;
2838 sizelo = env->me_mapsize & 0xffffffff;
2839 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2841 /* Windows won't create mappings for zero length files.
2842 * Just allocate the maxsize right now.
2845 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2846 if (!SetEndOfFile(env->me_fd))
2848 SetFilePointer(env->me_fd, 0, NULL, 0);
2850 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2851 PAGE_READWRITE : PAGE_READONLY,
2852 sizehi, sizelo, NULL);
2855 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2856 FILE_MAP_WRITE : FILE_MAP_READ,
2857 0, 0, env->me_mapsize, meta.mm_address);
2865 if (flags & MDB_WRITEMAP) {
2867 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2870 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2872 if (env->me_map == MAP_FAILED) {
2879 if (flags & MDB_FIXEDMAP)
2880 meta.mm_address = env->me_map;
2881 i = mdb_env_init_meta(env, &meta);
2882 if (i != MDB_SUCCESS) {
2885 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2886 /* Can happen because the address argument to mmap() is just a
2887 * hint. mmap() can pick another, e.g. if the range is in use.
2888 * The MAP_FIXED flag would prevent that, but then mmap could
2889 * instead unmap existing pages to make room for the new map.
2891 return EBUSY; /* TODO: Make a new MDB_* error code? */
2893 env->me_psize = meta.mm_psize;
2894 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2896 env->me_maxpg = env->me_mapsize / env->me_psize;
2898 p = (MDB_page *)env->me_map;
2899 env->me_metas[0] = METADATA(p);
2900 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2904 int toggle = mdb_env_pick_meta(env);
2905 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2907 DPRINTF("opened database version %u, pagesize %u",
2908 env->me_metas[0]->mm_version, env->me_psize);
2909 DPRINTF("using meta page %d", toggle);
2910 DPRINTF("depth: %u", db->md_depth);
2911 DPRINTF("entries: %zu", db->md_entries);
2912 DPRINTF("branch pages: %zu", db->md_branch_pages);
2913 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2914 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2915 DPRINTF("root: %zu", db->md_root);
2923 /** Release a reader thread's slot in the reader lock table.
2924 * This function is called automatically when a thread exits.
2925 * @param[in] ptr This points to the slot in the reader lock table.
2928 mdb_env_reader_dest(void *ptr)
2930 MDB_reader *reader = ptr;
2936 /** Junk for arranging thread-specific callbacks on Windows. This is
2937 * necessarily platform and compiler-specific. Windows supports up
2938 * to 1088 keys. Let's assume nobody opens more than 64 environments
2939 * in a single process, for now. They can override this if needed.
2941 #ifndef MAX_TLS_KEYS
2942 #define MAX_TLS_KEYS 64
2944 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2945 static int mdb_tls_nkeys;
2947 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2951 case DLL_PROCESS_ATTACH: break;
2952 case DLL_THREAD_ATTACH: break;
2953 case DLL_THREAD_DETACH:
2954 for (i=0; i<mdb_tls_nkeys; i++) {
2955 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2956 mdb_env_reader_dest(r);
2959 case DLL_PROCESS_DETACH: break;
2964 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2966 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2970 /* Force some symbol references.
2971 * _tls_used forces the linker to create the TLS directory if not already done
2972 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2974 #pragma comment(linker, "/INCLUDE:_tls_used")
2975 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2976 #pragma const_seg(".CRT$XLB")
2977 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2978 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2981 #pragma comment(linker, "/INCLUDE:__tls_used")
2982 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2983 #pragma data_seg(".CRT$XLB")
2984 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2986 #endif /* WIN 32/64 */
2987 #endif /* !__GNUC__ */
2990 /** Downgrade the exclusive lock on the region back to shared */
2992 mdb_env_share_locks(MDB_env *env, int *excl)
2994 int rc = 0, toggle = mdb_env_pick_meta(env);
2996 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3001 /* First acquire a shared lock. The Unlock will
3002 * then release the existing exclusive lock.
3004 memset(&ov, 0, sizeof(ov));
3005 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3008 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3014 struct flock lock_info;
3015 /* The shared lock replaces the existing lock */
3016 memset((void *)&lock_info, 0, sizeof(lock_info));
3017 lock_info.l_type = F_RDLCK;
3018 lock_info.l_whence = SEEK_SET;
3019 lock_info.l_start = 0;
3020 lock_info.l_len = 1;
3021 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3022 (rc = ErrCode()) == EINTR) ;
3023 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3030 /** Try to get exlusive lock, otherwise shared.
3031 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3034 mdb_env_excl_lock(MDB_env *env, int *excl)
3038 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3042 memset(&ov, 0, sizeof(ov));
3043 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3050 struct flock lock_info;
3051 memset((void *)&lock_info, 0, sizeof(lock_info));
3052 lock_info.l_type = F_WRLCK;
3053 lock_info.l_whence = SEEK_SET;
3054 lock_info.l_start = 0;
3055 lock_info.l_len = 1;
3056 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3057 (rc = ErrCode()) == EINTR) ;
3061 # ifdef MDB_USE_POSIX_SEM
3062 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3065 lock_info.l_type = F_RDLCK;
3066 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3067 (rc = ErrCode()) == EINTR) ;
3075 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3077 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3079 * @(#) $Revision: 5.1 $
3080 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3081 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3083 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3087 * Please do not copyright this code. This code is in the public domain.
3089 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3090 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3091 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3092 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3093 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3094 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3095 * PERFORMANCE OF THIS SOFTWARE.
3098 * chongo <Landon Curt Noll> /\oo/\
3099 * http://www.isthe.com/chongo/
3101 * Share and Enjoy! :-)
3104 typedef unsigned long long mdb_hash_t;
3105 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3107 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3108 * @param[in] str string to hash
3109 * @param[in] hval initial value for hash
3110 * @return 64 bit hash
3112 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3113 * hval arg on the first call.
3116 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3118 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3119 unsigned char *end = s + val->mv_size;
3121 * FNV-1a hash each octet of the string
3124 /* xor the bottom with the current octet */
3125 hval ^= (mdb_hash_t)*s++;
3127 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3128 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3129 (hval << 7) + (hval << 8) + (hval << 40);
3131 /* return our new hash value */
3135 /** Hash the string and output the hash in hex.
3136 * @param[in] str string to hash
3137 * @param[out] hexbuf an array of 17 chars to hold the hash
3140 mdb_hash_hex(MDB_val *val, char *hexbuf)
3143 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3144 for (i=0; i<8; i++) {
3145 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3151 /** Open and/or initialize the lock region for the environment.
3152 * @param[in] env The MDB environment.
3153 * @param[in] lpath The pathname of the file used for the lock region.
3154 * @param[in] mode The Unix permissions for the file, if we create it.
3155 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3156 * @return 0 on success, non-zero on failure.
3159 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3167 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3168 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3169 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3171 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3172 env->me_flags |= MDB_ROFS;
3177 /* Try to get exclusive lock. If we succeed, then
3178 * nobody is using the lock region and we should initialize it.
3180 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3181 size = GetFileSize(env->me_lfd, NULL);
3187 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3189 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3190 env->me_flags |= MDB_ROFS;
3195 /* Lose record locks when exec*() */
3196 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3197 fcntl(env->me_lfd, F_SETFD, fdflags);
3199 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3200 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3202 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3203 env->me_flags |= MDB_ROFS;
3210 /* Try to get exclusive lock. If we succeed, then
3211 * nobody is using the lock region and we should initialize it.
3213 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3215 size = lseek(env->me_lfd, 0, SEEK_END);
3217 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3218 if (size < rsize && *excl > 0) {
3220 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3221 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3223 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3227 size = rsize - sizeof(MDB_txninfo);
3228 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3233 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3235 if (!mh) goto fail_errno;
3236 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3238 if (!env->me_txns) goto fail_errno;
3240 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3242 if (m == MAP_FAILED) goto fail_errno;
3248 BY_HANDLE_FILE_INFORMATION stbuf;
3257 if (!mdb_sec_inited) {
3258 InitializeSecurityDescriptor(&mdb_null_sd,
3259 SECURITY_DESCRIPTOR_REVISION);
3260 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3261 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3262 mdb_all_sa.bInheritHandle = FALSE;
3263 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3266 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3267 idbuf.volume = stbuf.dwVolumeSerialNumber;
3268 idbuf.nhigh = stbuf.nFileIndexHigh;
3269 idbuf.nlow = stbuf.nFileIndexLow;
3270 val.mv_data = &idbuf;
3271 val.mv_size = sizeof(idbuf);
3272 mdb_hash_hex(&val, hexbuf);
3273 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3274 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3275 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3276 if (!env->me_rmutex) goto fail_errno;
3277 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3278 if (!env->me_wmutex) goto fail_errno;
3279 #elif defined(MDB_USE_POSIX_SEM)
3288 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3289 idbuf.dev = stbuf.st_dev;
3290 idbuf.ino = stbuf.st_ino;
3291 val.mv_data = &idbuf;
3292 val.mv_size = sizeof(idbuf);
3293 mdb_hash_hex(&val, hexbuf);
3294 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3295 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3296 /* Clean up after a previous run, if needed: Try to
3297 * remove both semaphores before doing anything else.
3299 sem_unlink(env->me_txns->mti_rmname);
3300 sem_unlink(env->me_txns->mti_wmname);
3301 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3302 O_CREAT|O_EXCL, mode, 1);
3303 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3304 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3305 O_CREAT|O_EXCL, mode, 1);
3306 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3307 #else /* MDB_USE_POSIX_SEM */
3308 pthread_mutexattr_t mattr;
3310 if ((rc = pthread_mutexattr_init(&mattr))
3311 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3312 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3313 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3315 pthread_mutexattr_destroy(&mattr);
3316 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3318 env->me_txns->mti_version = MDB_VERSION;
3319 env->me_txns->mti_magic = MDB_MAGIC;
3320 env->me_txns->mti_txnid = 0;
3321 env->me_txns->mti_numreaders = 0;
3324 if (env->me_txns->mti_magic != MDB_MAGIC) {
3325 DPUTS("lock region has invalid magic");
3329 if (env->me_txns->mti_version != MDB_VERSION) {
3330 DPRINTF("lock region is version %u, expected version %u",
3331 env->me_txns->mti_version, MDB_VERSION);
3332 rc = MDB_VERSION_MISMATCH;
3336 if (rc != EACCES && rc != EAGAIN) {
3340 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3341 if (!env->me_rmutex) goto fail_errno;
3342 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3343 if (!env->me_wmutex) goto fail_errno;
3344 #elif defined(MDB_USE_POSIX_SEM)
3345 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3346 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3347 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3348 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3359 /** The name of the lock file in the DB environment */
3360 #define LOCKNAME "/lock.mdb"
3361 /** The name of the data file in the DB environment */
3362 #define DATANAME "/data.mdb"
3363 /** The suffix of the lock file when no subdir is used */
3364 #define LOCKSUFF "-lock"
3365 /** Only a subset of the @ref mdb_env flags can be changed
3366 * at runtime. Changing other flags requires closing the
3367 * environment and re-opening it with the new flags.
3369 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3370 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3373 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3375 int oflags, rc, len, excl;
3376 char *lpath, *dpath;
3378 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3382 if (flags & MDB_NOSUBDIR) {
3383 rc = len + sizeof(LOCKSUFF) + len + 1;
3385 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3390 if (flags & MDB_NOSUBDIR) {
3391 dpath = lpath + len + sizeof(LOCKSUFF);
3392 sprintf(lpath, "%s" LOCKSUFF, path);
3393 strcpy(dpath, path);
3395 dpath = lpath + len + sizeof(LOCKNAME);
3396 sprintf(lpath, "%s" LOCKNAME, path);
3397 sprintf(dpath, "%s" DATANAME, path);
3400 flags |= env->me_flags;
3401 /* silently ignore WRITEMAP if we're only getting read access */
3402 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3403 flags ^= MDB_WRITEMAP;
3404 env->me_flags = flags |= MDB_ENV_ACTIVE;
3406 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3411 if (F_ISSET(flags, MDB_RDONLY)) {
3412 oflags = GENERIC_READ;
3413 len = OPEN_EXISTING;
3415 oflags = GENERIC_READ|GENERIC_WRITE;
3418 mode = FILE_ATTRIBUTE_NORMAL;
3419 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3420 NULL, len, mode, NULL);
3422 if (F_ISSET(flags, MDB_RDONLY))
3425 oflags = O_RDWR | O_CREAT;
3427 env->me_fd = open(dpath, oflags, mode);
3429 if (env->me_fd == INVALID_HANDLE_VALUE) {
3434 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3435 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3436 env->me_mfd = env->me_fd;
3438 /* Synchronous fd for meta writes. Needed even with
3439 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3442 env->me_mfd = CreateFile(dpath, oflags,
3443 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3444 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3446 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3448 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3453 DPRINTF("opened dbenv %p", (void *) env);
3454 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3457 env->me_numdbs = 2; /* this notes that me_txkey was set */
3459 /* Windows TLS callbacks need help finding their TLS info. */
3460 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3461 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3468 rc = mdb_env_share_locks(env, &excl);
3472 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3473 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3474 env->me_path = strdup(path);
3475 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3481 mdb_env_close0(env, excl);
3487 /** Destroy resources from mdb_env_open() and clear our readers */
3489 mdb_env_close0(MDB_env *env, int excl)
3493 if (!(env->me_flags & MDB_ENV_ACTIVE))
3496 free(env->me_dbflags);
3500 if (env->me_numdbs) {
3501 pthread_key_delete(env->me_txkey);
3503 /* Delete our key from the global list */
3504 for (i=0; i<mdb_tls_nkeys; i++)
3505 if (mdb_tls_keys[i] == env->me_txkey) {
3506 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3514 munmap(env->me_map, env->me_mapsize);
3516 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3518 if (env->me_fd != INVALID_HANDLE_VALUE)
3521 pid_t pid = env->me_pid;
3522 /* Clearing readers is done in this function because
3523 * me_txkey with its destructor must be disabled first.
3525 for (i = env->me_numreaders; --i >= 0; )
3526 if (env->me_txns->mti_readers[i].mr_pid == pid)
3527 env->me_txns->mti_readers[i].mr_pid = 0;
3529 if (env->me_rmutex) {
3530 CloseHandle(env->me_rmutex);
3531 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3533 /* Windows automatically destroys the mutexes when
3534 * the last handle closes.
3536 #elif defined(MDB_USE_POSIX_SEM)
3537 if (env->me_rmutex != SEM_FAILED) {
3538 sem_close(env->me_rmutex);
3539 if (env->me_wmutex != SEM_FAILED)
3540 sem_close(env->me_wmutex);
3541 /* If we have the filelock: If we are the
3542 * only remaining user, clean up semaphores.
3545 mdb_env_excl_lock(env, &excl);
3547 sem_unlink(env->me_txns->mti_rmname);
3548 sem_unlink(env->me_txns->mti_wmname);
3552 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3554 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3557 /* Unlock the lockfile. Windows would have unlocked it
3558 * after closing anyway, but not necessarily at once.
3560 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3566 env->me_flags &= ~MDB_ENV_ACTIVE;
3570 mdb_env_copy(MDB_env *env, const char *path)
3572 MDB_txn *txn = NULL;
3576 HANDLE newfd = INVALID_HANDLE_VALUE;
3578 if (env->me_flags & MDB_NOSUBDIR) {
3579 lpath = (char *)path;
3582 len += sizeof(DATANAME);
3583 lpath = malloc(len);
3586 sprintf(lpath, "%s" DATANAME, path);
3589 /* The destination path must exist, but the destination file must not.
3590 * We don't want the OS to cache the writes, since the source data is
3591 * already in the OS cache.
3594 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3595 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3597 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3603 if (!(env->me_flags & MDB_NOSUBDIR))
3605 if (newfd == INVALID_HANDLE_VALUE) {
3610 #ifdef F_NOCACHE /* __APPLE__ */
3611 rc = fcntl(newfd, F_NOCACHE, 1);
3618 /* Do the lock/unlock of the reader mutex before starting the
3619 * write txn. Otherwise other read txns could block writers.
3621 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3625 if (!(env->me_flags & MDB_ROFS)) {
3626 /* We must start the actual read txn after blocking writers */
3627 mdb_txn_reset0(txn);
3629 /* Temporarily block writers until we snapshot the meta pages */
3632 rc = mdb_txn_renew0(txn);
3634 UNLOCK_MUTEX_W(env);
3639 wsize = env->me_psize * 2;
3643 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3644 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3647 rc = write(newfd, env->me_map, wsize);
3648 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3650 if (! (env->me_flags & MDB_ROFS))
3651 UNLOCK_MUTEX_W(env);
3656 ptr = env->me_map + wsize;
3657 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3658 #define MAX_WRITE 2147483648U
3662 if (wsize > MAX_WRITE)
3666 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3667 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3676 if (wsize > MAX_WRITE)
3680 wres = write(newfd, ptr, w2);
3681 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3690 if (newfd != INVALID_HANDLE_VALUE)
3697 mdb_env_close(MDB_env *env)
3704 VGMEMP_DESTROY(env);
3705 while ((dp = env->me_dpages) != NULL) {
3706 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3707 env->me_dpages = dp->mp_next;
3711 mdb_env_close0(env, 0);
3712 mdb_midl_free(env->me_free_pgs);
3716 /** Compare two items pointing at aligned size_t's */
3718 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3720 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3721 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3724 /** Compare two items pointing at aligned int's */
3726 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3728 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3729 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3732 /** Compare two items pointing at ints of unknown alignment.
3733 * Nodes and keys are guaranteed to be 2-byte aligned.
3736 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3738 #if BYTE_ORDER == LITTLE_ENDIAN
3739 unsigned short *u, *c;
3742 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3743 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3746 } while(!x && u > (unsigned short *)a->mv_data);
3749 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3753 /** Compare two items lexically */
3755 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3762 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3768 diff = memcmp(a->mv_data, b->mv_data, len);
3769 return diff ? diff : len_diff<0 ? -1 : len_diff;
3772 /** Compare two items in reverse byte order */
3774 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3776 const unsigned char *p1, *p2, *p1_lim;
3780 p1_lim = (const unsigned char *)a->mv_data;
3781 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3782 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3784 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3790 while (p1 > p1_lim) {
3791 diff = *--p1 - *--p2;
3795 return len_diff<0 ? -1 : len_diff;
3798 /** Search for key within a page, using binary search.
3799 * Returns the smallest entry larger or equal to the key.
3800 * If exactp is non-null, stores whether the found entry was an exact match
3801 * in *exactp (1 or 0).
3802 * Updates the cursor index with the index of the found entry.
3803 * If no entry larger or equal to the key is found, returns NULL.
3806 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3808 unsigned int i = 0, nkeys;
3811 MDB_page *mp = mc->mc_pg[mc->mc_top];
3812 MDB_node *node = NULL;
3817 nkeys = NUMKEYS(mp);
3822 COPY_PGNO(pgno, mp->mp_pgno);
3823 DPRINTF("searching %u keys in %s %spage %zu",
3824 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3831 low = IS_LEAF(mp) ? 0 : 1;
3833 cmp = mc->mc_dbx->md_cmp;
3835 /* Branch pages have no data, so if using integer keys,
3836 * alignment is guaranteed. Use faster mdb_cmp_int.
3838 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3839 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3846 nodekey.mv_size = mc->mc_db->md_pad;
3847 node = NODEPTR(mp, 0); /* fake */
3848 while (low <= high) {
3849 i = (low + high) >> 1;
3850 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3851 rc = cmp(key, &nodekey);
3852 DPRINTF("found leaf index %u [%s], rc = %i",
3853 i, DKEY(&nodekey), rc);
3862 while (low <= high) {
3863 i = (low + high) >> 1;
3865 node = NODEPTR(mp, i);
3866 nodekey.mv_size = NODEKSZ(node);
3867 nodekey.mv_data = NODEKEY(node);
3869 rc = cmp(key, &nodekey);
3872 DPRINTF("found leaf index %u [%s], rc = %i",
3873 i, DKEY(&nodekey), rc);
3875 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3876 i, DKEY(&nodekey), NODEPGNO(node), rc);
3887 if (rc > 0) { /* Found entry is less than the key. */
3888 i++; /* Skip to get the smallest entry larger than key. */
3890 node = NODEPTR(mp, i);
3893 *exactp = (rc == 0);
3894 /* store the key index */
3895 mc->mc_ki[mc->mc_top] = i;
3897 /* There is no entry larger or equal to the key. */
3900 /* nodeptr is fake for LEAF2 */
3906 mdb_cursor_adjust(MDB_cursor *mc, func)
3910 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3911 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3918 /** Pop a page off the top of the cursor's stack. */
3920 mdb_cursor_pop(MDB_cursor *mc)
3923 #ifndef MDB_DEBUG_SKIP
3924 MDB_page *top = mc->mc_pg[mc->mc_top];
3930 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3931 mc->mc_dbi, (void *) mc);
3935 /** Push a page onto the top of the cursor's stack. */
3937 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3939 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3940 mc->mc_dbi, (void *) mc);
3942 if (mc->mc_snum >= CURSOR_STACK) {
3943 assert(mc->mc_snum < CURSOR_STACK);
3944 return MDB_CURSOR_FULL;
3947 mc->mc_top = mc->mc_snum++;
3948 mc->mc_pg[mc->mc_top] = mp;
3949 mc->mc_ki[mc->mc_top] = 0;
3954 /** Find the address of the page corresponding to a given page number.
3955 * @param[in] txn the transaction for this access.
3956 * @param[in] pgno the page number for the page to retrieve.
3957 * @param[out] ret address of a pointer where the page's address will be stored.
3958 * @return 0 on success, non-zero on failure.
3961 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3965 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3966 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3970 MDB_ID2L dl = tx2->mt_u.dirty_list;
3972 unsigned x = mdb_mid2l_search(dl, pgno);
3973 if (x <= dl[0].mid && dl[x].mid == pgno) {
3978 } while ((tx2 = tx2->mt_parent) != NULL);
3981 if (pgno < txn->mt_next_pgno) {
3982 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3984 DPRINTF("page %zu not found", pgno);
3990 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3993 /** Search for the page a given key should be in.
3994 * Pushes parent pages on the cursor stack. This function continues a
3995 * search on a cursor that has already been initialized. (Usually by
3996 * #mdb_page_search() but also by #mdb_node_move().)
3997 * @param[in,out] mc the cursor for this operation.
3998 * @param[in] key the key to search for. If NULL, search for the lowest
3999 * page. (This is used by #mdb_cursor_first().)
4000 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4001 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4002 * @return 0 on success, non-zero on failure.
4005 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4007 MDB_page *mp = mc->mc_pg[mc->mc_top];
4012 while (IS_BRANCH(mp)) {
4016 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4017 assert(NUMKEYS(mp) > 1);
4018 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4020 if (key == NULL) /* Initialize cursor to first page. */
4022 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4023 /* cursor to last page */
4027 node = mdb_node_search(mc, key, &exact);
4029 i = NUMKEYS(mp) - 1;
4031 i = mc->mc_ki[mc->mc_top];
4040 DPRINTF("following index %u for key [%s]",
4042 assert(i < NUMKEYS(mp));
4043 node = NODEPTR(mp, i);
4045 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
4048 mc->mc_ki[mc->mc_top] = i;
4049 if ((rc = mdb_cursor_push(mc, mp)))
4053 if ((rc = mdb_page_touch(mc)) != 0)
4055 mp = mc->mc_pg[mc->mc_top];
4060 DPRINTF("internal error, index points to a %02X page!?",
4062 return MDB_CORRUPTED;
4065 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4066 key ? DKEY(key) : NULL);
4071 /** Search for the page a given key should be in.
4072 * Pushes parent pages on the cursor stack. This function just sets up
4073 * the search; it finds the root page for \b mc's database and sets this
4074 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4075 * called to complete the search.
4076 * @param[in,out] mc the cursor for this operation.
4077 * @param[in] key the key to search for. If NULL, search for the lowest
4078 * page. (This is used by #mdb_cursor_first().)
4079 * @param[in] modify If true, visited pages are updated with new page numbers.
4080 * @return 0 on success, non-zero on failure.
4083 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4088 /* Make sure the txn is still viable, then find the root from
4089 * the txn's db table.
4091 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4092 DPUTS("transaction has failed, must abort");
4095 /* Make sure we're using an up-to-date root */
4096 if (mc->mc_dbi > MAIN_DBI) {
4097 if ((*mc->mc_dbflag & DB_STALE) ||
4098 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4100 unsigned char dbflag = 0;
4101 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4102 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4105 if (*mc->mc_dbflag & DB_STALE) {
4108 MDB_node *leaf = mdb_node_search(&mc2,
4109 &mc->mc_dbx->md_name, &exact);
4111 return MDB_NOTFOUND;
4112 mdb_node_read(mc->mc_txn, leaf, &data);
4113 /* The txn may not know this DBI, or another process may
4114 * have dropped and recreated the DB with other flags.
4116 if (mc->mc_db->md_flags != *(uint16_t *)
4117 ((char *) data.mv_data + offsetof(MDB_db, md_flags)))
4118 return MDB_INCOMPATIBLE;
4119 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4121 if (flags & MDB_PS_MODIFY)
4123 *mc->mc_dbflag = dbflag;
4126 root = mc->mc_db->md_root;
4128 if (root == P_INVALID) { /* Tree is empty. */
4129 DPUTS("tree is empty");
4130 return MDB_NOTFOUND;
4135 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4136 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4142 DPRINTF("db %u root page %zu has flags 0x%X",
4143 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4145 if (flags & MDB_PS_MODIFY) {
4146 if ((rc = mdb_page_touch(mc)))
4150 if (flags & MDB_PS_ROOTONLY)
4153 return mdb_page_search_root(mc, key, flags);
4156 /** Return the data associated with a given node.
4157 * @param[in] txn The transaction for this operation.
4158 * @param[in] leaf The node being read.
4159 * @param[out] data Updated to point to the node's data.
4160 * @return 0 on success, non-zero on failure.
4163 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4165 MDB_page *omp; /* overflow page */
4169 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4170 data->mv_size = NODEDSZ(leaf);
4171 data->mv_data = NODEDATA(leaf);
4175 /* Read overflow data.
4177 data->mv_size = NODEDSZ(leaf);
4178 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4179 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4180 DPRINTF("read overflow page %zu failed", pgno);
4183 data->mv_data = METADATA(omp);
4189 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4190 MDB_val *key, MDB_val *data)
4199 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4201 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4204 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4208 mdb_cursor_init(&mc, txn, dbi, &mx);
4209 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4212 /** Find a sibling for a page.
4213 * Replaces the page at the top of the cursor's stack with the
4214 * specified sibling, if one exists.
4215 * @param[in] mc The cursor for this operation.
4216 * @param[in] move_right Non-zero if the right sibling is requested,
4217 * otherwise the left sibling.
4218 * @return 0 on success, non-zero on failure.
4221 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4227 if (mc->mc_snum < 2) {
4228 return MDB_NOTFOUND; /* root has no siblings */
4232 DPRINTF("parent page is page %zu, index %u",
4233 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4235 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4236 : (mc->mc_ki[mc->mc_top] == 0)) {
4237 DPRINTF("no more keys left, moving to %s sibling",
4238 move_right ? "right" : "left");
4239 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4240 /* undo cursor_pop before returning */
4247 mc->mc_ki[mc->mc_top]++;
4249 mc->mc_ki[mc->mc_top]--;
4250 DPRINTF("just moving to %s index key %u",
4251 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4253 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4255 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4256 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4259 mdb_cursor_push(mc, mp);
4264 /** Move the cursor to the next data item. */
4266 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4272 if (mc->mc_flags & C_EOF) {
4273 return MDB_NOTFOUND;
4276 assert(mc->mc_flags & C_INITIALIZED);
4278 mp = mc->mc_pg[mc->mc_top];
4280 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4281 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4282 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4283 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4284 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4285 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4289 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4290 if (op == MDB_NEXT_DUP)
4291 return MDB_NOTFOUND;
4295 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4297 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4298 DPUTS("=====> move to next sibling page");
4299 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4300 mc->mc_flags |= C_EOF;
4301 mc->mc_flags &= ~C_INITIALIZED;
4302 return MDB_NOTFOUND;
4304 mp = mc->mc_pg[mc->mc_top];
4305 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4307 mc->mc_ki[mc->mc_top]++;
4309 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4310 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4313 key->mv_size = mc->mc_db->md_pad;
4314 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4318 assert(IS_LEAF(mp));
4319 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4321 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4322 mdb_xcursor_init1(mc, leaf);
4325 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4328 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4329 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4330 if (rc != MDB_SUCCESS)
4335 MDB_GET_KEY(leaf, key);
4339 /** Move the cursor to the previous data item. */
4341 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4347 assert(mc->mc_flags & C_INITIALIZED);
4349 mp = mc->mc_pg[mc->mc_top];
4351 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4352 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4353 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4354 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4355 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4356 if (op != MDB_PREV || rc == MDB_SUCCESS)
4359 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4360 if (op == MDB_PREV_DUP)
4361 return MDB_NOTFOUND;
4366 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4368 if (mc->mc_ki[mc->mc_top] == 0) {
4369 DPUTS("=====> move to prev sibling page");
4370 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4371 mc->mc_flags &= ~C_INITIALIZED;
4372 return MDB_NOTFOUND;
4374 mp = mc->mc_pg[mc->mc_top];
4375 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4376 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4378 mc->mc_ki[mc->mc_top]--;
4380 mc->mc_flags &= ~C_EOF;
4382 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4383 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4386 key->mv_size = mc->mc_db->md_pad;
4387 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4391 assert(IS_LEAF(mp));
4392 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4394 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4395 mdb_xcursor_init1(mc, leaf);
4398 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4401 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4402 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4403 if (rc != MDB_SUCCESS)
4408 MDB_GET_KEY(leaf, key);
4412 /** Set the cursor on a specific data item. */
4414 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4415 MDB_cursor_op op, int *exactp)
4419 MDB_node *leaf = NULL;
4424 assert(key->mv_size > 0);
4426 /* See if we're already on the right page */
4427 if (mc->mc_flags & C_INITIALIZED) {
4430 mp = mc->mc_pg[mc->mc_top];
4432 mc->mc_ki[mc->mc_top] = 0;
4433 return MDB_NOTFOUND;
4435 if (mp->mp_flags & P_LEAF2) {
4436 nodekey.mv_size = mc->mc_db->md_pad;
4437 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4439 leaf = NODEPTR(mp, 0);
4440 MDB_GET_KEY(leaf, &nodekey);
4442 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4444 /* Probably happens rarely, but first node on the page
4445 * was the one we wanted.
4447 mc->mc_ki[mc->mc_top] = 0;
4454 unsigned int nkeys = NUMKEYS(mp);
4456 if (mp->mp_flags & P_LEAF2) {
4457 nodekey.mv_data = LEAF2KEY(mp,
4458 nkeys-1, nodekey.mv_size);
4460 leaf = NODEPTR(mp, nkeys-1);
4461 MDB_GET_KEY(leaf, &nodekey);
4463 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4465 /* last node was the one we wanted */
4466 mc->mc_ki[mc->mc_top] = nkeys-1;
4472 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4473 /* This is definitely the right page, skip search_page */
4474 if (mp->mp_flags & P_LEAF2) {
4475 nodekey.mv_data = LEAF2KEY(mp,
4476 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4478 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4479 MDB_GET_KEY(leaf, &nodekey);
4481 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4483 /* current node was the one we wanted */
4493 /* If any parents have right-sibs, search.
4494 * Otherwise, there's nothing further.
4496 for (i=0; i<mc->mc_top; i++)
4498 NUMKEYS(mc->mc_pg[i])-1)
4500 if (i == mc->mc_top) {
4501 /* There are no other pages */
4502 mc->mc_ki[mc->mc_top] = nkeys;
4503 return MDB_NOTFOUND;
4507 /* There are no other pages */
4508 mc->mc_ki[mc->mc_top] = 0;
4509 return MDB_NOTFOUND;
4513 rc = mdb_page_search(mc, key, 0);
4514 if (rc != MDB_SUCCESS)
4517 mp = mc->mc_pg[mc->mc_top];
4518 assert(IS_LEAF(mp));
4521 leaf = mdb_node_search(mc, key, exactp);
4522 if (exactp != NULL && !*exactp) {
4523 /* MDB_SET specified and not an exact match. */
4524 return MDB_NOTFOUND;
4528 DPUTS("===> inexact leaf not found, goto sibling");
4529 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4530 return rc; /* no entries matched */
4531 mp = mc->mc_pg[mc->mc_top];
4532 assert(IS_LEAF(mp));
4533 leaf = NODEPTR(mp, 0);
4537 mc->mc_flags |= C_INITIALIZED;
4538 mc->mc_flags &= ~C_EOF;
4541 key->mv_size = mc->mc_db->md_pad;
4542 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4546 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4547 mdb_xcursor_init1(mc, leaf);
4550 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4551 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4552 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4555 if (op == MDB_GET_BOTH) {
4561 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4562 if (rc != MDB_SUCCESS)
4565 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4567 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4569 rc = mc->mc_dbx->md_dcmp(data, &d2);
4571 if (op == MDB_GET_BOTH || rc > 0)
4572 return MDB_NOTFOUND;
4577 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4578 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4583 /* The key already matches in all other cases */
4584 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4585 MDB_GET_KEY(leaf, key);
4586 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4591 /** Move the cursor to the first item in the database. */
4593 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4598 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4599 rc = mdb_page_search(mc, NULL, 0);
4600 if (rc != MDB_SUCCESS)
4603 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4605 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4606 mc->mc_flags |= C_INITIALIZED;
4607 mc->mc_flags &= ~C_EOF;
4609 mc->mc_ki[mc->mc_top] = 0;
4611 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4612 key->mv_size = mc->mc_db->md_pad;
4613 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4618 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4619 mdb_xcursor_init1(mc, leaf);
4620 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4625 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4626 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4630 MDB_GET_KEY(leaf, key);
4634 /** Move the cursor to the last item in the database. */
4636 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4641 if (!(mc->mc_flags & C_EOF)) {
4643 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4646 lkey.mv_size = MDB_MAXKEYSIZE+1;
4647 lkey.mv_data = NULL;
4648 rc = mdb_page_search(mc, &lkey, 0);
4649 if (rc != MDB_SUCCESS)
4652 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4654 mc->mc_flags |= C_INITIALIZED|C_EOF;
4655 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4657 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4659 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4660 key->mv_size = mc->mc_db->md_pad;
4661 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4666 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4667 mdb_xcursor_init1(mc, leaf);
4668 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4673 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4674 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4679 MDB_GET_KEY(leaf, key);
4684 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4693 case MDB_GET_CURRENT:
4694 if (!(mc->mc_flags & C_INITIALIZED)) {
4697 MDB_page *mp = mc->mc_pg[mc->mc_top];
4699 mc->mc_ki[mc->mc_top] = 0;
4705 key->mv_size = mc->mc_db->md_pad;
4706 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4708 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4709 MDB_GET_KEY(leaf, key);
4711 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4712 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4714 rc = mdb_node_read(mc->mc_txn, leaf, data);
4721 case MDB_GET_BOTH_RANGE:
4722 if (data == NULL || mc->mc_xcursor == NULL) {
4730 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4732 } else if (op == MDB_SET_RANGE)
4733 rc = mdb_cursor_set(mc, key, data, op, NULL);
4735 rc = mdb_cursor_set(mc, key, data, op, &exact);
4737 case MDB_GET_MULTIPLE:
4739 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4740 !(mc->mc_flags & C_INITIALIZED)) {
4745 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4746 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4749 case MDB_NEXT_MULTIPLE:
4751 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4755 if (!(mc->mc_flags & C_INITIALIZED))
4756 rc = mdb_cursor_first(mc, key, data);
4758 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4759 if (rc == MDB_SUCCESS) {
4760 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4763 mx = &mc->mc_xcursor->mx_cursor;
4764 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4766 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4767 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4775 case MDB_NEXT_NODUP:
4776 if (!(mc->mc_flags & C_INITIALIZED))
4777 rc = mdb_cursor_first(mc, key, data);
4779 rc = mdb_cursor_next(mc, key, data, op);
4783 case MDB_PREV_NODUP:
4784 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4785 rc = mdb_cursor_last(mc, key, data);
4786 mc->mc_flags |= C_INITIALIZED;
4787 mc->mc_ki[mc->mc_top]++;
4789 rc = mdb_cursor_prev(mc, key, data, op);
4792 rc = mdb_cursor_first(mc, key, data);
4796 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4797 !(mc->mc_flags & C_INITIALIZED) ||
4798 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4802 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4805 rc = mdb_cursor_last(mc, key, data);
4809 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4810 !(mc->mc_flags & C_INITIALIZED) ||
4811 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4815 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4818 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4826 /** Touch all the pages in the cursor stack.
4827 * Makes sure all the pages are writable, before attempting a write operation.
4828 * @param[in] mc The cursor to operate on.
4831 mdb_cursor_touch(MDB_cursor *mc)
4835 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4838 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4839 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4840 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4843 *mc->mc_dbflag = DB_DIRTY;
4845 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4846 rc = mdb_page_touch(mc);
4850 mc->mc_top = mc->mc_snum-1;
4855 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4858 MDB_node *leaf = NULL;
4859 MDB_val xdata, *rdata, dkey;
4862 int do_sub = 0, insert = 0;
4863 unsigned int mcount = 0;
4867 char dbuf[MDB_MAXKEYSIZE+1];
4868 unsigned int nflags;
4871 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4874 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4877 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4880 #if SIZE_MAX > MAXDATASIZE
4881 if (data->mv_size > MAXDATASIZE)
4885 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4886 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4890 if (flags == MDB_CURRENT) {
4891 if (!(mc->mc_flags & C_INITIALIZED))
4894 } else if (mc->mc_db->md_root == P_INVALID) {
4896 /* new database, write a root leaf page */
4897 DPUTS("allocating new root leaf page");
4898 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4902 mdb_cursor_push(mc, np);
4903 mc->mc_db->md_root = np->mp_pgno;
4904 mc->mc_db->md_depth++;
4905 *mc->mc_dbflag = DB_DIRTY;
4906 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4908 np->mp_flags |= P_LEAF2;
4909 mc->mc_flags |= C_INITIALIZED;
4915 if (flags & MDB_APPEND) {
4917 rc = mdb_cursor_last(mc, &k2, &d2);
4919 rc = mc->mc_dbx->md_cmp(key, &k2);
4922 mc->mc_ki[mc->mc_top]++;
4928 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4930 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4931 DPRINTF("duplicate key [%s]", DKEY(key));
4933 return MDB_KEYEXIST;
4935 if (rc && rc != MDB_NOTFOUND)
4939 /* Cursor is positioned, now make sure all pages are writable */
4940 rc2 = mdb_cursor_touch(mc);
4945 /* The key already exists */
4946 if (rc == MDB_SUCCESS) {
4947 /* there's only a key anyway, so this is a no-op */
4948 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4949 unsigned int ksize = mc->mc_db->md_pad;
4950 if (key->mv_size != ksize)
4952 if (flags == MDB_CURRENT) {
4953 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4954 memcpy(ptr, key->mv_data, ksize);
4959 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4962 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4963 /* Was a single item before, must convert now */
4965 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4966 /* Just overwrite the current item */
4967 if (flags == MDB_CURRENT)
4970 dkey.mv_size = NODEDSZ(leaf);
4971 dkey.mv_data = NODEDATA(leaf);
4972 #if UINT_MAX < SIZE_MAX
4973 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4974 #ifdef MISALIGNED_OK
4975 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4977 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4980 /* if data matches, ignore it */
4981 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4982 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4984 /* create a fake page for the dup items */
4985 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4986 dkey.mv_data = dbuf;
4987 fp = (MDB_page *)&pbuf;
4988 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4989 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4990 fp->mp_lower = PAGEHDRSZ;
4991 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4992 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4993 fp->mp_flags |= P_LEAF2;
4994 fp->mp_pad = data->mv_size;
4995 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4997 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4998 (dkey.mv_size & 1) + (data->mv_size & 1);
5000 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5003 xdata.mv_size = fp->mp_upper;
5008 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5009 /* See if we need to convert from fake page to subDB */
5011 unsigned int offset;
5014 fp = NODEDATA(leaf);
5015 if (flags == MDB_CURRENT) {
5017 fp->mp_flags |= P_DIRTY;
5018 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5019 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5023 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5024 offset = fp->mp_pad;
5025 if (SIZELEFT(fp) >= offset)
5027 offset *= 4; /* space for 4 more */
5029 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5031 offset += offset & 1;
5032 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5033 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
5035 /* yes, convert it */
5037 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5038 dummy.md_pad = fp->mp_pad;
5039 dummy.md_flags = MDB_DUPFIXED;
5040 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5041 dummy.md_flags |= MDB_INTEGERKEY;
5044 dummy.md_branch_pages = 0;
5045 dummy.md_leaf_pages = 1;
5046 dummy.md_overflow_pages = 0;
5047 dummy.md_entries = NUMKEYS(fp);
5049 xdata.mv_size = sizeof(MDB_db);
5050 xdata.mv_data = &dummy;
5051 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5053 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5054 flags |= F_DUPDATA|F_SUBDATA;
5055 dummy.md_root = mp->mp_pgno;
5057 /* no, just grow it */
5059 xdata.mv_size = NODEDSZ(leaf) + offset;
5060 xdata.mv_data = &pbuf;
5061 mp = (MDB_page *)&pbuf;
5062 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5065 mp->mp_flags = fp->mp_flags | P_DIRTY;
5066 mp->mp_pad = fp->mp_pad;
5067 mp->mp_lower = fp->mp_lower;
5068 mp->mp_upper = fp->mp_upper + offset;
5070 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5072 nsize = NODEDSZ(leaf) - fp->mp_upper;
5073 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5074 for (i=0; i<NUMKEYS(fp); i++)
5075 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5077 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5081 /* data is on sub-DB, just store it */
5082 flags |= F_DUPDATA|F_SUBDATA;
5086 /* overflow page overwrites need special handling */
5087 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5090 int ovpages, dpages;
5092 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5093 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5094 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5095 mdb_page_get(mc->mc_txn, pg, &omp);
5096 /* Is the ov page writable and large enough? */
5097 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5098 /* yes, overwrite it. Note in this case we don't
5099 * bother to try shrinking the node if the new data
5100 * is smaller than the overflow threshold.
5102 if (F_ISSET(flags, MDB_RESERVE))
5103 data->mv_data = METADATA(omp);
5105 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5108 /* no, free ovpages */
5110 mc->mc_db->md_overflow_pages -= ovpages;
5111 for (i=0; i<ovpages; i++) {
5112 DPRINTF("freed ov page %zu", pg);
5113 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5117 } else if (NODEDSZ(leaf) == data->mv_size) {
5118 /* same size, just replace it. Note that we could
5119 * also reuse this node if the new data is smaller,
5120 * but instead we opt to shrink the node in that case.
5122 if (F_ISSET(flags, MDB_RESERVE))
5123 data->mv_data = NODEDATA(leaf);
5124 else if (data->mv_size)
5125 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5127 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5130 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5131 mc->mc_db->md_entries--;
5133 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5140 nflags = flags & NODE_ADD_FLAGS;
5141 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5142 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5143 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5144 nflags &= ~MDB_APPEND;
5146 nflags |= MDB_SPLIT_REPLACE;
5147 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5149 /* There is room already in this leaf page. */
5150 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5151 if (rc == 0 && !do_sub && insert) {
5152 /* Adjust other cursors pointing to mp */
5153 MDB_cursor *m2, *m3;
5154 MDB_dbi dbi = mc->mc_dbi;
5155 unsigned i = mc->mc_top;
5156 MDB_page *mp = mc->mc_pg[i];
5158 if (mc->mc_flags & C_SUB)
5161 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5162 if (mc->mc_flags & C_SUB)
5163 m3 = &m2->mc_xcursor->mx_cursor;
5166 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5167 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5174 if (rc != MDB_SUCCESS)
5175 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5177 /* Now store the actual data in the child DB. Note that we're
5178 * storing the user data in the keys field, so there are strict
5179 * size limits on dupdata. The actual data fields of the child
5180 * DB are all zero size.
5187 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5188 if (flags & MDB_CURRENT) {
5189 xflags = MDB_CURRENT;
5191 mdb_xcursor_init1(mc, leaf);
5192 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5194 /* converted, write the original data first */
5196 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5200 /* Adjust other cursors pointing to mp */
5202 unsigned i = mc->mc_top;
5203 MDB_page *mp = mc->mc_pg[i];
5205 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5206 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5207 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5208 mdb_xcursor_init1(m2, leaf);
5213 if (flags & MDB_APPENDDUP)
5214 xflags |= MDB_APPEND;
5215 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5216 if (flags & F_SUBDATA) {
5217 void *db = NODEDATA(leaf);
5218 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5221 /* sub-writes might have failed so check rc again.
5222 * Don't increment count if we just replaced an existing item.
5224 if (!rc && !(flags & MDB_CURRENT))
5225 mc->mc_db->md_entries++;
5226 if (flags & MDB_MULTIPLE) {
5228 if (mcount < data[1].mv_size) {
5229 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5230 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5236 /* If we succeeded and the key didn't exist before, make sure
5237 * the cursor is marked valid.
5240 mc->mc_flags |= C_INITIALIZED;
5245 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5250 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5253 if (!(mc->mc_flags & C_INITIALIZED))
5256 rc = mdb_cursor_touch(mc);
5260 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5262 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5263 if (flags != MDB_NODUPDATA) {
5264 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5265 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5267 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5268 /* If sub-DB still has entries, we're done */
5269 if (mc->mc_xcursor->mx_db.md_entries) {
5270 if (leaf->mn_flags & F_SUBDATA) {
5271 /* update subDB info */
5272 void *db = NODEDATA(leaf);
5273 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5275 /* shrink fake page */
5276 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5278 mc->mc_db->md_entries--;
5281 /* otherwise fall thru and delete the sub-DB */
5284 if (leaf->mn_flags & F_SUBDATA) {
5285 /* add all the child DB's pages to the free list */
5286 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5287 if (rc == MDB_SUCCESS) {
5288 mc->mc_db->md_entries -=
5289 mc->mc_xcursor->mx_db.md_entries;
5294 return mdb_cursor_del0(mc, leaf);
5297 /** Allocate and initialize new pages for a database.
5298 * @param[in] mc a cursor on the database being added to.
5299 * @param[in] flags flags defining what type of page is being allocated.
5300 * @param[in] num the number of pages to allocate. This is usually 1,
5301 * unless allocating overflow pages for a large record.
5302 * @param[out] mp Address of a page, or NULL on failure.
5303 * @return 0 on success, non-zero on failure.
5306 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5311 if ((rc = mdb_page_alloc(mc, num, &np)))
5313 DPRINTF("allocated new mpage %zu, page size %u",
5314 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5315 np->mp_flags = flags | P_DIRTY;
5316 np->mp_lower = PAGEHDRSZ;
5317 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5320 mc->mc_db->md_branch_pages++;
5321 else if (IS_LEAF(np))
5322 mc->mc_db->md_leaf_pages++;
5323 else if (IS_OVERFLOW(np)) {
5324 mc->mc_db->md_overflow_pages += num;
5332 /** Calculate the size of a leaf node.
5333 * The size depends on the environment's page size; if a data item
5334 * is too large it will be put onto an overflow page and the node
5335 * size will only include the key and not the data. Sizes are always
5336 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5337 * of the #MDB_node headers.
5338 * @param[in] env The environment handle.
5339 * @param[in] key The key for the node.
5340 * @param[in] data The data for the node.
5341 * @return The number of bytes needed to store the node.
5344 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5348 sz = LEAFSIZE(key, data);
5349 if (sz >= env->me_psize / MDB_MINKEYS) {
5350 /* put on overflow page */
5351 sz -= data->mv_size - sizeof(pgno_t);
5355 return sz + sizeof(indx_t);
5358 /** Calculate the size of a branch node.
5359 * The size should depend on the environment's page size but since
5360 * we currently don't support spilling large keys onto overflow
5361 * pages, it's simply the size of the #MDB_node header plus the
5362 * size of the key. Sizes are always rounded up to an even number
5363 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5364 * @param[in] env The environment handle.
5365 * @param[in] key The key for the node.
5366 * @return The number of bytes needed to store the node.
5369 mdb_branch_size(MDB_env *env, MDB_val *key)
5374 if (sz >= env->me_psize / MDB_MINKEYS) {
5375 /* put on overflow page */
5376 /* not implemented */
5377 /* sz -= key->size - sizeof(pgno_t); */
5380 return sz + sizeof(indx_t);
5383 /** Add a node to the page pointed to by the cursor.
5384 * @param[in] mc The cursor for this operation.
5385 * @param[in] indx The index on the page where the new node should be added.
5386 * @param[in] key The key for the new node.
5387 * @param[in] data The data for the new node, if any.
5388 * @param[in] pgno The page number, if adding a branch node.
5389 * @param[in] flags Flags for the node.
5390 * @return 0 on success, non-zero on failure. Possible errors are:
5392 * <li>ENOMEM - failed to allocate overflow pages for the node.
5393 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5394 * should never happen since all callers already calculate the
5395 * page's free space before calling this function.
5399 mdb_node_add(MDB_cursor *mc, indx_t indx,
5400 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5403 size_t node_size = NODESIZE;
5406 MDB_page *mp = mc->mc_pg[mc->mc_top];
5407 MDB_page *ofp = NULL; /* overflow page */
5410 assert(mp->mp_upper >= mp->mp_lower);
5412 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5413 IS_LEAF(mp) ? "leaf" : "branch",
5414 IS_SUBP(mp) ? "sub-" : "",
5415 mp->mp_pgno, indx, data ? data->mv_size : 0,
5416 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5419 /* Move higher keys up one slot. */
5420 int ksize = mc->mc_db->md_pad, dif;
5421 char *ptr = LEAF2KEY(mp, indx, ksize);
5422 dif = NUMKEYS(mp) - indx;
5424 memmove(ptr+ksize, ptr, dif*ksize);
5425 /* insert new key */
5426 memcpy(ptr, key->mv_data, ksize);
5428 /* Just using these for counting */
5429 mp->mp_lower += sizeof(indx_t);
5430 mp->mp_upper -= ksize - sizeof(indx_t);
5435 node_size += key->mv_size;
5439 if (F_ISSET(flags, F_BIGDATA)) {
5440 /* Data already on overflow page. */
5441 node_size += sizeof(pgno_t);
5442 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5443 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5445 /* Put data on overflow page. */
5446 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5447 data->mv_size, node_size+data->mv_size);
5448 node_size += sizeof(pgno_t);
5449 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5451 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5454 node_size += data->mv_size;
5457 node_size += node_size & 1;
5459 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5460 DPRINTF("not enough room in page %zu, got %u ptrs",
5461 mp->mp_pgno, NUMKEYS(mp));
5462 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5463 mp->mp_upper - mp->mp_lower);
5464 DPRINTF("node size = %zu", node_size);
5465 return MDB_PAGE_FULL;
5468 /* Move higher pointers up one slot. */
5469 for (i = NUMKEYS(mp); i > indx; i--)
5470 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5472 /* Adjust free space offsets. */
5473 ofs = mp->mp_upper - node_size;
5474 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5475 mp->mp_ptrs[indx] = ofs;
5477 mp->mp_lower += sizeof(indx_t);
5479 /* Write the node data. */
5480 node = NODEPTR(mp, indx);
5481 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5482 node->mn_flags = flags;
5484 SETDSZ(node,data->mv_size);
5489 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5494 if (F_ISSET(flags, F_BIGDATA))
5495 memcpy(node->mn_data + key->mv_size, data->mv_data,
5497 else if (F_ISSET(flags, MDB_RESERVE))
5498 data->mv_data = node->mn_data + key->mv_size;
5500 memcpy(node->mn_data + key->mv_size, data->mv_data,
5503 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5505 if (F_ISSET(flags, MDB_RESERVE))
5506 data->mv_data = METADATA(ofp);
5508 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5515 /** Delete the specified node from a page.
5516 * @param[in] mp The page to operate on.
5517 * @param[in] indx The index of the node to delete.
5518 * @param[in] ksize The size of a node. Only used if the page is
5519 * part of a #MDB_DUPFIXED database.
5522 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5525 indx_t i, j, numkeys, ptr;
5532 COPY_PGNO(pgno, mp->mp_pgno);
5533 DPRINTF("delete node %u on %s page %zu", indx,
5534 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5537 assert(indx < NUMKEYS(mp));
5540 int x = NUMKEYS(mp) - 1 - indx;
5541 base = LEAF2KEY(mp, indx, ksize);
5543 memmove(base, base + ksize, x * ksize);
5544 mp->mp_lower -= sizeof(indx_t);
5545 mp->mp_upper += ksize - sizeof(indx_t);
5549 node = NODEPTR(mp, indx);
5550 sz = NODESIZE + node->mn_ksize;
5552 if (F_ISSET(node->mn_flags, F_BIGDATA))
5553 sz += sizeof(pgno_t);
5555 sz += NODEDSZ(node);
5559 ptr = mp->mp_ptrs[indx];
5560 numkeys = NUMKEYS(mp);
5561 for (i = j = 0; i < numkeys; i++) {
5563 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5564 if (mp->mp_ptrs[i] < ptr)
5565 mp->mp_ptrs[j] += sz;
5570 base = (char *)mp + mp->mp_upper;
5571 memmove(base + sz, base, ptr - mp->mp_upper);
5573 mp->mp_lower -= sizeof(indx_t);
5577 /** Compact the main page after deleting a node on a subpage.
5578 * @param[in] mp The main page to operate on.
5579 * @param[in] indx The index of the subpage on the main page.
5582 mdb_node_shrink(MDB_page *mp, indx_t indx)
5589 indx_t i, numkeys, ptr;
5591 node = NODEPTR(mp, indx);
5592 sp = (MDB_page *)NODEDATA(node);
5593 osize = NODEDSZ(node);
5595 delta = sp->mp_upper - sp->mp_lower;
5596 SETDSZ(node, osize - delta);
5597 xp = (MDB_page *)((char *)sp + delta);
5599 /* shift subpage upward */
5601 nsize = NUMKEYS(sp) * sp->mp_pad;
5602 memmove(METADATA(xp), METADATA(sp), nsize);
5605 nsize = osize - sp->mp_upper;
5606 numkeys = NUMKEYS(sp);
5607 for (i=numkeys-1; i>=0; i--)
5608 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5610 xp->mp_upper = sp->mp_lower;
5611 xp->mp_lower = sp->mp_lower;
5612 xp->mp_flags = sp->mp_flags;
5613 xp->mp_pad = sp->mp_pad;
5614 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5616 /* shift lower nodes upward */
5617 ptr = mp->mp_ptrs[indx];
5618 numkeys = NUMKEYS(mp);
5619 for (i = 0; i < numkeys; i++) {
5620 if (mp->mp_ptrs[i] <= ptr)
5621 mp->mp_ptrs[i] += delta;
5624 base = (char *)mp + mp->mp_upper;
5625 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5626 mp->mp_upper += delta;
5629 /** Initial setup of a sorted-dups cursor.
5630 * Sorted duplicates are implemented as a sub-database for the given key.
5631 * The duplicate data items are actually keys of the sub-database.
5632 * Operations on the duplicate data items are performed using a sub-cursor
5633 * initialized when the sub-database is first accessed. This function does
5634 * the preliminary setup of the sub-cursor, filling in the fields that
5635 * depend only on the parent DB.
5636 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5639 mdb_xcursor_init0(MDB_cursor *mc)
5641 MDB_xcursor *mx = mc->mc_xcursor;
5643 mx->mx_cursor.mc_xcursor = NULL;
5644 mx->mx_cursor.mc_txn = mc->mc_txn;
5645 mx->mx_cursor.mc_db = &mx->mx_db;
5646 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5647 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5648 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5649 mx->mx_cursor.mc_snum = 0;
5650 mx->mx_cursor.mc_top = 0;
5651 mx->mx_cursor.mc_flags = C_SUB;
5652 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5653 mx->mx_dbx.md_dcmp = NULL;
5654 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5657 /** Final setup of a sorted-dups cursor.
5658 * Sets up the fields that depend on the data from the main cursor.
5659 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5660 * @param[in] node The data containing the #MDB_db record for the
5661 * sorted-dup database.
5664 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5666 MDB_xcursor *mx = mc->mc_xcursor;
5668 if (node->mn_flags & F_SUBDATA) {
5669 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5670 mx->mx_cursor.mc_pg[0] = 0;
5671 mx->mx_cursor.mc_snum = 0;
5672 mx->mx_cursor.mc_flags = C_SUB;
5674 MDB_page *fp = NODEDATA(node);
5675 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5676 mx->mx_db.md_flags = 0;
5677 mx->mx_db.md_depth = 1;
5678 mx->mx_db.md_branch_pages = 0;
5679 mx->mx_db.md_leaf_pages = 1;
5680 mx->mx_db.md_overflow_pages = 0;
5681 mx->mx_db.md_entries = NUMKEYS(fp);
5682 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5683 mx->mx_cursor.mc_snum = 1;
5684 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5685 mx->mx_cursor.mc_top = 0;
5686 mx->mx_cursor.mc_pg[0] = fp;
5687 mx->mx_cursor.mc_ki[0] = 0;
5688 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5689 mx->mx_db.md_flags = MDB_DUPFIXED;
5690 mx->mx_db.md_pad = fp->mp_pad;
5691 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5692 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5695 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5697 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5699 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5700 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5701 #if UINT_MAX < SIZE_MAX
5702 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5703 #ifdef MISALIGNED_OK
5704 mx->mx_dbx.md_cmp = mdb_cmp_long;
5706 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5711 /** Initialize a cursor for a given transaction and database. */
5713 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5718 mc->mc_db = &txn->mt_dbs[dbi];
5719 mc->mc_dbx = &txn->mt_dbxs[dbi];
5720 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5725 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5727 mc->mc_xcursor = mx;
5728 mdb_xcursor_init0(mc);
5730 mc->mc_xcursor = NULL;
5732 if (*mc->mc_dbflag & DB_STALE) {
5733 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5738 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5741 MDB_xcursor *mx = NULL;
5742 size_t size = sizeof(MDB_cursor);
5744 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5747 /* Allow read access to the freelist */
5748 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5751 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5752 size += sizeof(MDB_xcursor);
5754 if ((mc = malloc(size)) != NULL) {
5755 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5756 mx = (MDB_xcursor *)(mc + 1);
5758 mdb_cursor_init(mc, txn, dbi, mx);
5759 if (txn->mt_cursors) {
5760 mc->mc_next = txn->mt_cursors[dbi];
5761 txn->mt_cursors[dbi] = mc;
5763 mc->mc_flags |= C_ALLOCD;
5774 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5776 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5779 if (txn->mt_cursors)
5782 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5786 /* Return the count of duplicate data items for the current key */
5788 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5792 if (mc == NULL || countp == NULL)
5795 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5798 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5799 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5802 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5805 *countp = mc->mc_xcursor->mx_db.md_entries;
5811 mdb_cursor_close(MDB_cursor *mc)
5814 /* remove from txn, if tracked */
5815 if (mc->mc_txn->mt_cursors) {
5816 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5817 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5819 *prev = mc->mc_next;
5821 if (mc->mc_flags & C_ALLOCD)
5827 mdb_cursor_txn(MDB_cursor *mc)
5829 if (!mc) return NULL;
5834 mdb_cursor_dbi(MDB_cursor *mc)
5840 /** Replace the key for a node with a new key.
5841 * @param[in] mp The page containing the node to operate on.
5842 * @param[in] indx The index of the node to operate on.
5843 * @param[in] key The new key to use.
5844 * @return 0 on success, non-zero on failure.
5847 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5854 indx_t ptr, i, numkeys, indx;
5857 indx = mc->mc_ki[mc->mc_top];
5858 mp = mc->mc_pg[mc->mc_top];
5859 node = NODEPTR(mp, indx);
5860 ptr = mp->mp_ptrs[indx];
5864 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5865 k2.mv_data = NODEKEY(node);
5866 k2.mv_size = node->mn_ksize;
5867 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5869 mdb_dkey(&k2, kbuf2),
5875 delta0 = delta = key->mv_size - node->mn_ksize;
5877 /* Must be 2-byte aligned. If new key is
5878 * shorter by 1, the shift will be skipped.
5880 delta += (delta & 1);
5882 if (delta > 0 && SIZELEFT(mp) < delta) {
5884 /* not enough space left, do a delete and split */
5885 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5886 pgno = NODEPGNO(node);
5887 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5888 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5891 numkeys = NUMKEYS(mp);
5892 for (i = 0; i < numkeys; i++) {
5893 if (mp->mp_ptrs[i] <= ptr)
5894 mp->mp_ptrs[i] -= delta;
5897 base = (char *)mp + mp->mp_upper;
5898 len = ptr - mp->mp_upper + NODESIZE;
5899 memmove(base - delta, base, len);
5900 mp->mp_upper -= delta;
5902 node = NODEPTR(mp, indx);
5905 /* But even if no shift was needed, update ksize */
5907 node->mn_ksize = key->mv_size;
5910 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5916 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5918 /** Move a node from csrc to cdst.
5921 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5928 unsigned short flags;
5932 /* Mark src and dst as dirty. */
5933 if ((rc = mdb_page_touch(csrc)) ||
5934 (rc = mdb_page_touch(cdst)))
5937 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5938 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5939 key.mv_size = csrc->mc_db->md_pad;
5940 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5942 data.mv_data = NULL;
5946 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5947 assert(!((long)srcnode&1));
5948 srcpg = NODEPGNO(srcnode);
5949 flags = srcnode->mn_flags;
5950 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5951 unsigned int snum = csrc->mc_snum;
5953 /* must find the lowest key below src */
5954 mdb_page_search_root(csrc, NULL, 0);
5955 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5956 key.mv_size = csrc->mc_db->md_pad;
5957 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5959 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5960 key.mv_size = NODEKSZ(s2);
5961 key.mv_data = NODEKEY(s2);
5963 csrc->mc_snum = snum--;
5964 csrc->mc_top = snum;
5966 key.mv_size = NODEKSZ(srcnode);
5967 key.mv_data = NODEKEY(srcnode);
5969 data.mv_size = NODEDSZ(srcnode);
5970 data.mv_data = NODEDATA(srcnode);
5972 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5973 unsigned int snum = cdst->mc_snum;
5976 /* must find the lowest key below dst */
5977 mdb_page_search_root(cdst, NULL, 0);
5978 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5979 bkey.mv_size = cdst->mc_db->md_pad;
5980 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5982 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5983 bkey.mv_size = NODEKSZ(s2);
5984 bkey.mv_data = NODEKEY(s2);
5986 cdst->mc_snum = snum--;
5987 cdst->mc_top = snum;
5988 mdb_cursor_copy(cdst, &mn);
5990 rc = mdb_update_key(&mn, &bkey);
5995 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5996 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5997 csrc->mc_ki[csrc->mc_top],
5999 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6000 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6002 /* Add the node to the destination page.
6004 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6005 if (rc != MDB_SUCCESS)
6008 /* Delete the node from the source page.
6010 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6013 /* Adjust other cursors pointing to mp */
6014 MDB_cursor *m2, *m3;
6015 MDB_dbi dbi = csrc->mc_dbi;
6016 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6018 if (csrc->mc_flags & C_SUB)
6021 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6022 if (m2 == csrc) continue;
6023 if (csrc->mc_flags & C_SUB)
6024 m3 = &m2->mc_xcursor->mx_cursor;
6027 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6028 csrc->mc_ki[csrc->mc_top]) {
6029 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6030 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6035 /* Update the parent separators.
6037 if (csrc->mc_ki[csrc->mc_top] == 0) {
6038 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6039 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6040 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6042 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6043 key.mv_size = NODEKSZ(srcnode);
6044 key.mv_data = NODEKEY(srcnode);
6046 DPRINTF("update separator for source page %zu to [%s]",
6047 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6048 mdb_cursor_copy(csrc, &mn);
6051 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6054 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6056 indx_t ix = csrc->mc_ki[csrc->mc_top];
6057 nullkey.mv_size = 0;
6058 csrc->mc_ki[csrc->mc_top] = 0;
6059 rc = mdb_update_key(csrc, &nullkey);
6060 csrc->mc_ki[csrc->mc_top] = ix;
6061 assert(rc == MDB_SUCCESS);
6065 if (cdst->mc_ki[cdst->mc_top] == 0) {
6066 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6067 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6068 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6070 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6071 key.mv_size = NODEKSZ(srcnode);
6072 key.mv_data = NODEKEY(srcnode);
6074 DPRINTF("update separator for destination page %zu to [%s]",
6075 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6076 mdb_cursor_copy(cdst, &mn);
6079 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6082 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6084 indx_t ix = cdst->mc_ki[cdst->mc_top];
6085 nullkey.mv_size = 0;
6086 cdst->mc_ki[cdst->mc_top] = 0;
6087 rc = mdb_update_key(cdst, &nullkey);
6088 cdst->mc_ki[cdst->mc_top] = ix;
6089 assert(rc == MDB_SUCCESS);
6096 /** Merge one page into another.
6097 * The nodes from the page pointed to by \b csrc will
6098 * be copied to the page pointed to by \b cdst and then
6099 * the \b csrc page will be freed.
6100 * @param[in] csrc Cursor pointing to the source page.
6101 * @param[in] cdst Cursor pointing to the destination page.
6104 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6112 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6113 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6115 assert(csrc->mc_snum > 1); /* can't merge root page */
6116 assert(cdst->mc_snum > 1);
6118 /* Mark dst as dirty. */
6119 if ((rc = mdb_page_touch(cdst)))
6122 /* Move all nodes from src to dst.
6124 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6125 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6126 key.mv_size = csrc->mc_db->md_pad;
6127 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6128 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6129 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6130 if (rc != MDB_SUCCESS)
6132 key.mv_data = (char *)key.mv_data + key.mv_size;
6135 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6136 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6137 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6138 unsigned int snum = csrc->mc_snum;
6140 /* must find the lowest key below src */
6141 mdb_page_search_root(csrc, NULL, 0);
6142 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6143 key.mv_size = csrc->mc_db->md_pad;
6144 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6146 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6147 key.mv_size = NODEKSZ(s2);
6148 key.mv_data = NODEKEY(s2);
6150 csrc->mc_snum = snum--;
6151 csrc->mc_top = snum;
6153 key.mv_size = srcnode->mn_ksize;
6154 key.mv_data = NODEKEY(srcnode);
6157 data.mv_size = NODEDSZ(srcnode);
6158 data.mv_data = NODEDATA(srcnode);
6159 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6160 if (rc != MDB_SUCCESS)
6165 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6166 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]), (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10);
6168 /* Unlink the src page from parent and add to free list.
6170 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6171 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6174 rc = mdb_update_key(csrc, &key);
6180 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6181 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6182 csrc->mc_db->md_leaf_pages--;
6184 csrc->mc_db->md_branch_pages--;
6186 /* Adjust other cursors pointing to mp */
6187 MDB_cursor *m2, *m3;
6188 MDB_dbi dbi = csrc->mc_dbi;
6189 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6191 if (csrc->mc_flags & C_SUB)
6194 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6195 if (csrc->mc_flags & C_SUB)
6196 m3 = &m2->mc_xcursor->mx_cursor;
6199 if (m3 == csrc) continue;
6200 if (m3->mc_snum < csrc->mc_snum) continue;
6201 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6202 m3->mc_pg[csrc->mc_top] = mp;
6203 m3->mc_ki[csrc->mc_top] += nkeys;
6207 mdb_cursor_pop(csrc);
6209 return mdb_rebalance(csrc);
6212 /** Copy the contents of a cursor.
6213 * @param[in] csrc The cursor to copy from.
6214 * @param[out] cdst The cursor to copy to.
6217 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6221 cdst->mc_txn = csrc->mc_txn;
6222 cdst->mc_dbi = csrc->mc_dbi;
6223 cdst->mc_db = csrc->mc_db;
6224 cdst->mc_dbx = csrc->mc_dbx;
6225 cdst->mc_snum = csrc->mc_snum;
6226 cdst->mc_top = csrc->mc_top;
6227 cdst->mc_flags = csrc->mc_flags;
6229 for (i=0; i<csrc->mc_snum; i++) {
6230 cdst->mc_pg[i] = csrc->mc_pg[i];
6231 cdst->mc_ki[i] = csrc->mc_ki[i];
6235 /** Rebalance the tree after a delete operation.
6236 * @param[in] mc Cursor pointing to the page where rebalancing
6238 * @return 0 on success, non-zero on failure.
6241 mdb_rebalance(MDB_cursor *mc)
6251 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6252 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6253 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6254 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6258 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6261 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6262 DPRINTF("no need to rebalance page %zu, above fill threshold",
6268 if (mc->mc_snum < 2) {
6269 MDB_page *mp = mc->mc_pg[0];
6270 if (NUMKEYS(mp) == 0) {
6271 DPUTS("tree is completely empty");
6272 mc->mc_db->md_root = P_INVALID;
6273 mc->mc_db->md_depth = 0;
6274 mc->mc_db->md_leaf_pages = 0;
6275 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6279 /* Adjust other cursors pointing to mp */
6280 MDB_cursor *m2, *m3;
6281 MDB_dbi dbi = mc->mc_dbi;
6283 if (mc->mc_flags & C_SUB)
6286 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6287 if (m2 == mc) continue;
6288 if (mc->mc_flags & C_SUB)
6289 m3 = &m2->mc_xcursor->mx_cursor;
6292 if (m3->mc_snum < mc->mc_snum) continue;
6293 if (m3->mc_pg[0] == mp) {
6299 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6300 DPUTS("collapsing root page!");
6301 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6302 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6303 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6306 mc->mc_db->md_depth--;
6307 mc->mc_db->md_branch_pages--;
6309 /* Adjust other cursors pointing to mp */
6310 MDB_cursor *m2, *m3;
6311 MDB_dbi dbi = mc->mc_dbi;
6313 if (mc->mc_flags & C_SUB)
6316 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6317 if (m2 == mc) continue;
6318 if (mc->mc_flags & C_SUB)
6319 m3 = &m2->mc_xcursor->mx_cursor;
6322 if (m3->mc_snum < mc->mc_snum) continue;
6323 if (m3->mc_pg[0] == mp) {
6324 m3->mc_pg[0] = mc->mc_pg[0];
6329 DPUTS("root page doesn't need rebalancing");
6333 /* The parent (branch page) must have at least 2 pointers,
6334 * otherwise the tree is invalid.
6336 ptop = mc->mc_top-1;
6337 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6339 /* Leaf page fill factor is below the threshold.
6340 * Try to move keys from left or right neighbor, or
6341 * merge with a neighbor page.
6346 mdb_cursor_copy(mc, &mn);
6347 mn.mc_xcursor = NULL;
6349 if (mc->mc_ki[ptop] == 0) {
6350 /* We're the leftmost leaf in our parent.
6352 DPUTS("reading right neighbor");
6354 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6355 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6357 mn.mc_ki[mn.mc_top] = 0;
6358 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6360 /* There is at least one neighbor to the left.
6362 DPUTS("reading left neighbor");
6364 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6365 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6367 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6368 mc->mc_ki[mc->mc_top] = 0;
6371 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6372 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10);
6374 /* If the neighbor page is above threshold and has at least two
6375 * keys, move one key from it.
6377 * Otherwise we should try to merge them.
6379 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6380 return mdb_node_move(&mn, mc);
6382 if (mc->mc_ki[ptop] == 0)
6383 rc = mdb_page_merge(&mn, mc);
6385 rc = mdb_page_merge(mc, &mn);
6386 mc->mc_flags &= ~C_INITIALIZED;
6391 /** Complete a delete operation started by #mdb_cursor_del(). */
6393 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6397 /* add overflow pages to free list */
6398 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6402 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6403 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6404 mc->mc_db->md_overflow_pages -= ovpages;
6405 for (i=0; i<ovpages; i++) {
6406 DPRINTF("freed ov page %zu", pg);
6407 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6411 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6412 mc->mc_db->md_entries--;
6413 rc = mdb_rebalance(mc);
6414 if (rc != MDB_SUCCESS)
6415 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6421 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6422 MDB_val *key, MDB_val *data)
6427 MDB_val rdata, *xdata;
6431 assert(key != NULL);
6433 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6435 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6438 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6442 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6446 mdb_cursor_init(&mc, txn, dbi, &mx);
6457 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6459 /* let mdb_page_split know about this cursor if needed:
6460 * delete will trigger a rebalance; if it needs to move
6461 * a node from one page to another, it will have to
6462 * update the parent's separator key(s). If the new sepkey
6463 * is larger than the current one, the parent page may
6464 * run out of space, triggering a split. We need this
6465 * cursor to be consistent until the end of the rebalance.
6467 mc.mc_next = txn->mt_cursors[dbi];
6468 txn->mt_cursors[dbi] = &mc;
6469 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6470 txn->mt_cursors[dbi] = mc.mc_next;
6475 /** Split a page and insert a new node.
6476 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6477 * The cursor will be updated to point to the actual page and index where
6478 * the node got inserted after the split.
6479 * @param[in] newkey The key for the newly inserted node.
6480 * @param[in] newdata The data for the newly inserted node.
6481 * @param[in] newpgno The page number, if the new node is a branch node.
6482 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6483 * @return 0 on success, non-zero on failure.
6486 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6487 unsigned int nflags)
6490 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6493 unsigned int i, j, split_indx, nkeys, pmax;
6495 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6497 MDB_page *mp, *rp, *pp;
6502 mp = mc->mc_pg[mc->mc_top];
6503 newindx = mc->mc_ki[mc->mc_top];
6505 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6506 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6507 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6509 /* Create a right sibling. */
6510 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6512 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6514 if (mc->mc_snum < 2) {
6515 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6517 /* shift current top to make room for new parent */
6518 mc->mc_pg[1] = mc->mc_pg[0];
6519 mc->mc_ki[1] = mc->mc_ki[0];
6522 mc->mc_db->md_root = pp->mp_pgno;
6523 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6524 mc->mc_db->md_depth++;
6527 /* Add left (implicit) pointer. */
6528 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6529 /* undo the pre-push */
6530 mc->mc_pg[0] = mc->mc_pg[1];
6531 mc->mc_ki[0] = mc->mc_ki[1];
6532 mc->mc_db->md_root = mp->mp_pgno;
6533 mc->mc_db->md_depth--;
6540 ptop = mc->mc_top-1;
6541 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6544 mc->mc_flags |= C_SPLITTING;
6545 mdb_cursor_copy(mc, &mn);
6546 mn.mc_pg[mn.mc_top] = rp;
6547 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6549 if (nflags & MDB_APPEND) {
6550 mn.mc_ki[mn.mc_top] = 0;
6552 split_indx = newindx;
6557 nkeys = NUMKEYS(mp);
6558 split_indx = nkeys / 2;
6559 if (newindx < split_indx)
6565 unsigned int lsize, rsize, ksize;
6566 /* Move half of the keys to the right sibling */
6568 x = mc->mc_ki[mc->mc_top] - split_indx;
6569 ksize = mc->mc_db->md_pad;
6570 split = LEAF2KEY(mp, split_indx, ksize);
6571 rsize = (nkeys - split_indx) * ksize;
6572 lsize = (nkeys - split_indx) * sizeof(indx_t);
6573 mp->mp_lower -= lsize;
6574 rp->mp_lower += lsize;
6575 mp->mp_upper += rsize - lsize;
6576 rp->mp_upper -= rsize - lsize;
6577 sepkey.mv_size = ksize;
6578 if (newindx == split_indx) {
6579 sepkey.mv_data = newkey->mv_data;
6581 sepkey.mv_data = split;
6584 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6585 memcpy(rp->mp_ptrs, split, rsize);
6586 sepkey.mv_data = rp->mp_ptrs;
6587 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6588 memcpy(ins, newkey->mv_data, ksize);
6589 mp->mp_lower += sizeof(indx_t);
6590 mp->mp_upper -= ksize - sizeof(indx_t);
6593 memcpy(rp->mp_ptrs, split, x * ksize);
6594 ins = LEAF2KEY(rp, x, ksize);
6595 memcpy(ins, newkey->mv_data, ksize);
6596 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6597 rp->mp_lower += sizeof(indx_t);
6598 rp->mp_upper -= ksize - sizeof(indx_t);
6599 mc->mc_ki[mc->mc_top] = x;
6600 mc->mc_pg[mc->mc_top] = rp;
6605 /* For leaf pages, check the split point based on what
6606 * fits where, since otherwise mdb_node_add can fail.
6608 * This check is only needed when the data items are
6609 * relatively large, such that being off by one will
6610 * make the difference between success or failure.
6612 * It's also relevant if a page happens to be laid out
6613 * such that one half of its nodes are all "small" and
6614 * the other half of its nodes are "large." If the new
6615 * item is also "large" and falls on the half with
6616 * "large" nodes, it also may not fit.
6619 unsigned int psize, nsize;
6620 /* Maximum free space in an empty page */
6621 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6622 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6623 if ((nkeys < 20) || (nsize > pmax/16)) {
6624 if (newindx <= split_indx) {
6627 for (i=0; i<split_indx; i++) {
6628 node = NODEPTR(mp, i);
6629 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6630 if (F_ISSET(node->mn_flags, F_BIGDATA))
6631 psize += sizeof(pgno_t);
6633 psize += NODEDSZ(node);
6637 split_indx = newindx;
6648 for (i=nkeys-1; i>=split_indx; i--) {
6649 node = NODEPTR(mp, i);
6650 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6651 if (F_ISSET(node->mn_flags, F_BIGDATA))
6652 psize += sizeof(pgno_t);
6654 psize += NODEDSZ(node);
6658 split_indx = newindx;
6669 /* First find the separating key between the split pages.
6670 * The case where newindx == split_indx is ambiguous; the
6671 * new item could go to the new page or stay on the original
6672 * page. If newpos == 1 it goes to the new page.
6674 if (newindx == split_indx && newpos) {
6675 sepkey.mv_size = newkey->mv_size;
6676 sepkey.mv_data = newkey->mv_data;
6678 node = NODEPTR(mp, split_indx);
6679 sepkey.mv_size = node->mn_ksize;
6680 sepkey.mv_data = NODEKEY(node);
6684 DPRINTF("separator is [%s]", DKEY(&sepkey));
6686 /* Copy separator key to the parent.
6688 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6692 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6695 if (mn.mc_snum == mc->mc_snum) {
6696 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6697 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6698 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6699 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6704 /* Right page might now have changed parent.
6705 * Check if left page also changed parent.
6707 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6708 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6709 for (i=0; i<ptop; i++) {
6710 mc->mc_pg[i] = mn.mc_pg[i];
6711 mc->mc_ki[i] = mn.mc_ki[i];
6713 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6714 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6718 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6721 mc->mc_flags ^= C_SPLITTING;
6722 if (rc != MDB_SUCCESS) {
6725 if (nflags & MDB_APPEND) {
6726 mc->mc_pg[mc->mc_top] = rp;
6727 mc->mc_ki[mc->mc_top] = 0;
6728 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6731 for (i=0; i<mc->mc_top; i++)
6732 mc->mc_ki[i] = mn.mc_ki[i];
6739 /* Move half of the keys to the right sibling. */
6741 /* grab a page to hold a temporary copy */
6742 copy = mdb_page_malloc(mc);
6746 copy->mp_pgno = mp->mp_pgno;
6747 copy->mp_flags = mp->mp_flags;
6748 copy->mp_lower = PAGEHDRSZ;
6749 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6750 mc->mc_pg[mc->mc_top] = copy;
6751 for (i = j = 0; i <= nkeys; j++) {
6752 if (i == split_indx) {
6753 /* Insert in right sibling. */
6754 /* Reset insert index for right sibling. */
6755 if (i != newindx || (newpos ^ ins_new)) {
6757 mc->mc_pg[mc->mc_top] = rp;
6761 if (i == newindx && !ins_new) {
6762 /* Insert the original entry that caused the split. */
6763 rkey.mv_data = newkey->mv_data;
6764 rkey.mv_size = newkey->mv_size;
6773 /* Update index for the new key. */
6774 mc->mc_ki[mc->mc_top] = j;
6775 } else if (i == nkeys) {
6778 node = NODEPTR(mp, i);
6779 rkey.mv_data = NODEKEY(node);
6780 rkey.mv_size = node->mn_ksize;
6782 xdata.mv_data = NODEDATA(node);
6783 xdata.mv_size = NODEDSZ(node);
6786 pgno = NODEPGNO(node);
6787 flags = node->mn_flags;
6792 if (!IS_LEAF(mp) && j == 0) {
6793 /* First branch index doesn't need key data. */
6797 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6801 nkeys = NUMKEYS(copy);
6802 for (i=0; i<nkeys; i++)
6803 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6804 mp->mp_lower = copy->mp_lower;
6805 mp->mp_upper = copy->mp_upper;
6806 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6807 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6809 /* reset back to original page */
6810 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6811 mc->mc_pg[mc->mc_top] = mp;
6812 if (nflags & MDB_RESERVE) {
6813 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6814 if (!(node->mn_flags & F_BIGDATA))
6815 newdata->mv_data = NODEDATA(node);
6821 /* return tmp page to freelist */
6822 mdb_page_free(mc->mc_txn->mt_env, copy);
6825 /* Adjust other cursors pointing to mp */
6826 MDB_cursor *m2, *m3;
6827 MDB_dbi dbi = mc->mc_dbi;
6828 int fixup = NUMKEYS(mp);
6830 if (mc->mc_flags & C_SUB)
6833 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6834 if (m2 == mc) continue;
6835 if (mc->mc_flags & C_SUB)
6836 m3 = &m2->mc_xcursor->mx_cursor;
6839 if (!(m3->mc_flags & C_INITIALIZED))
6841 if (m3->mc_flags & C_SPLITTING)
6846 for (k=m3->mc_top; k>=0; k--) {
6847 m3->mc_ki[k+1] = m3->mc_ki[k];
6848 m3->mc_pg[k+1] = m3->mc_pg[k];
6850 if (m3->mc_ki[0] >= split_indx) {
6855 m3->mc_pg[0] = mc->mc_pg[0];
6859 if (m3->mc_pg[mc->mc_top] == mp) {
6860 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6861 m3->mc_ki[mc->mc_top]++;
6862 if (m3->mc_ki[mc->mc_top] >= fixup) {
6863 m3->mc_pg[mc->mc_top] = rp;
6864 m3->mc_ki[mc->mc_top] -= fixup;
6865 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6867 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6868 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6877 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6878 MDB_val *key, MDB_val *data, unsigned int flags)
6883 assert(key != NULL);
6884 assert(data != NULL);
6886 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6889 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6893 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6897 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6900 mdb_cursor_init(&mc, txn, dbi, &mx);
6901 return mdb_cursor_put(&mc, key, data, flags);
6905 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6907 if ((flag & CHANGEABLE) != flag)
6910 env->me_flags |= flag;
6912 env->me_flags &= ~flag;
6917 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6922 *arg = env->me_flags;
6927 mdb_env_get_path(MDB_env *env, const char **arg)
6932 *arg = env->me_path;
6936 /** Common code for #mdb_stat() and #mdb_env_stat().
6937 * @param[in] env the environment to operate in.
6938 * @param[in] db the #MDB_db record containing the stats to return.
6939 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6940 * @return 0, this function always succeeds.
6943 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6945 arg->ms_psize = env->me_psize;
6946 arg->ms_depth = db->md_depth;
6947 arg->ms_branch_pages = db->md_branch_pages;
6948 arg->ms_leaf_pages = db->md_leaf_pages;
6949 arg->ms_overflow_pages = db->md_overflow_pages;
6950 arg->ms_entries = db->md_entries;
6955 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6959 if (env == NULL || arg == NULL)
6962 toggle = mdb_env_pick_meta(env);
6964 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6968 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6972 if (env == NULL || arg == NULL)
6975 toggle = mdb_env_pick_meta(env);
6976 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6977 arg->me_mapsize = env->me_mapsize;
6978 arg->me_maxreaders = env->me_maxreaders;
6979 arg->me_numreaders = env->me_numreaders;
6980 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6981 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6985 /** Set the default comparison functions for a database.
6986 * Called immediately after a database is opened to set the defaults.
6987 * The user can then override them with #mdb_set_compare() or
6988 * #mdb_set_dupsort().
6989 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6990 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6993 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6995 uint16_t f = txn->mt_dbs[dbi].md_flags;
6997 txn->mt_dbxs[dbi].md_cmp =
6998 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6999 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7001 txn->mt_dbxs[dbi].md_dcmp =
7002 !(f & MDB_DUPSORT) ? 0 :
7003 ((f & MDB_INTEGERDUP)
7004 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7005 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7008 #define PERSISTENT_FLAGS 0xffff
7009 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
7010 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
7011 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7017 int rc, dbflag, exact;
7018 unsigned int unused = 0;
7021 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7022 mdb_default_cmp(txn, FREE_DBI);
7025 if ((flags & VALID_FLAGS) != flags)
7031 if (flags & PERSISTENT_FLAGS) {
7032 uint16_t f2 = flags & PERSISTENT_FLAGS;
7033 /* make sure flag changes get committed */
7034 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7035 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7036 txn->mt_flags |= MDB_TXN_DIRTY;
7039 mdb_default_cmp(txn, MAIN_DBI);
7043 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7044 mdb_default_cmp(txn, MAIN_DBI);
7047 /* Is the DB already open? */
7049 for (i=2; i<txn->mt_numdbs; i++) {
7050 if (!txn->mt_dbxs[i].md_name.mv_size) {
7051 /* Remember this free slot */
7052 if (!unused) unused = i;
7055 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7056 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7062 /* If no free slot and max hit, fail */
7063 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7064 return MDB_DBS_FULL;
7066 /* Find the DB info */
7070 key.mv_data = (void *)name;
7071 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7072 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7073 if (rc == MDB_SUCCESS) {
7074 /* make sure this is actually a DB */
7075 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7076 if (!(node->mn_flags & F_SUBDATA))
7078 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7079 /* Create if requested */
7081 data.mv_size = sizeof(MDB_db);
7082 data.mv_data = &dummy;
7083 memset(&dummy, 0, sizeof(dummy));
7084 dummy.md_root = P_INVALID;
7085 dummy.md_flags = flags & PERSISTENT_FLAGS;
7086 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7090 /* OK, got info, add to table */
7091 if (rc == MDB_SUCCESS) {
7092 unsigned int slot = unused ? unused : txn->mt_numdbs;
7093 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7094 txn->mt_dbxs[slot].md_name.mv_size = len;
7095 txn->mt_dbxs[slot].md_rel = NULL;
7096 txn->mt_dbflags[slot] = dbflag;
7097 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7099 txn->mt_env->me_dbflags[slot] = mdflags = txn->mt_dbs[slot].md_flags;
7100 mdb_default_cmp(txn, slot);
7103 txn->mt_env->me_numdbs++;
7105 /* Open the DB in parent txns as well */
7106 while ((txn = txn->mt_parent) != NULL) {
7107 txn->mt_dbflags[slot] = DB_STALE;
7108 txn->mt_dbs[slot].md_flags = mdflags;
7117 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7119 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7122 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7125 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7128 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7130 ptr = env->me_dbxs[dbi].md_name.mv_data;
7131 env->me_dbxs[dbi].md_name.mv_data = NULL;
7132 env->me_dbxs[dbi].md_name.mv_size = 0;
7136 /** Add all the DB's pages to the free list.
7137 * @param[in] mc Cursor on the DB to free.
7138 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7139 * @return 0 on success, non-zero on failure.
7142 mdb_drop0(MDB_cursor *mc, int subs)
7146 rc = mdb_page_search(mc, NULL, 0);
7147 if (rc == MDB_SUCCESS) {
7152 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7153 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
7156 mdb_cursor_copy(mc, &mx);
7157 while (mc->mc_snum > 0) {
7158 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7159 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7160 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7161 if (ni->mn_flags & F_SUBDATA) {
7162 mdb_xcursor_init1(mc, ni);
7163 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7169 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7171 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7174 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7179 rc = mdb_cursor_sibling(mc, 1);
7181 /* no more siblings, go back to beginning
7182 * of previous level.
7185 for (i=1; i<mc->mc_top; i++)
7186 mc->mc_pg[i] = mx.mc_pg[i];
7190 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7191 mc->mc_db->md_root);
7196 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7201 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7204 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7207 rc = mdb_cursor_open(txn, dbi, &mc);
7211 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7215 /* Can't delete the main DB */
7216 if (del && dbi > MAIN_DBI) {
7217 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7219 txn->mt_dbflags[dbi] = DB_STALE;
7220 mdb_dbi_close(txn->mt_env, dbi);
7223 /* reset the DB record, mark it dirty */
7224 txn->mt_dbflags[dbi] |= DB_DIRTY;
7225 txn->mt_dbs[dbi].md_depth = 0;
7226 txn->mt_dbs[dbi].md_branch_pages = 0;
7227 txn->mt_dbs[dbi].md_leaf_pages = 0;
7228 txn->mt_dbs[dbi].md_overflow_pages = 0;
7229 txn->mt_dbs[dbi].md_entries = 0;
7230 txn->mt_dbs[dbi].md_root = P_INVALID;
7232 if (!txn->mt_u.dirty_list[0].mid) {
7235 /* make sure we have at least one dirty page in this txn
7236 * otherwise these changes will be ignored.
7238 key.mv_size = sizeof(txnid_t);
7239 key.mv_data = &txn->mt_txnid;
7240 data.mv_size = sizeof(MDB_ID);
7241 data.mv_data = txn->mt_free_pgs;
7242 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7243 rc = mdb_cursor_put(&m2, &key, &data, 0);
7247 mdb_cursor_close(mc);
7251 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7253 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7256 txn->mt_dbxs[dbi].md_cmp = cmp;
7260 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7262 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7265 txn->mt_dbxs[dbi].md_dcmp = cmp;
7269 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7271 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7274 txn->mt_dbxs[dbi].md_rel = rel;
7278 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7280 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7283 txn->mt_dbxs[dbi].md_relctx = ctx;