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;
1635 np = mdb_page_malloc(mc);
1638 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1639 mid.mid = np->mp_pgno;
1641 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1649 mdb_env_sync(MDB_env *env, int force)
1652 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1653 if (env->me_flags & MDB_WRITEMAP) {
1654 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1655 ? MS_ASYNC : MS_SYNC;
1656 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1659 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1663 if (MDB_FDATASYNC(env->me_fd))
1670 /** Make shadow copies of all of parent txn's cursors */
1672 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1674 MDB_cursor *mc, *m2;
1675 unsigned int i, j, size;
1677 for (i=0;i<src->mt_numdbs; i++) {
1678 if (src->mt_cursors[i]) {
1679 size = sizeof(MDB_cursor);
1680 if (src->mt_cursors[i]->mc_xcursor)
1681 size += sizeof(MDB_xcursor);
1682 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1689 mc->mc_db = &dst->mt_dbs[i];
1690 mc->mc_dbx = m2->mc_dbx;
1691 mc->mc_dbflag = &dst->mt_dbflags[i];
1692 mc->mc_snum = m2->mc_snum;
1693 mc->mc_top = m2->mc_top;
1694 mc->mc_flags = m2->mc_flags | C_SHADOW;
1695 for (j=0; j<mc->mc_snum; j++) {
1696 mc->mc_pg[j] = m2->mc_pg[j];
1697 mc->mc_ki[j] = m2->mc_ki[j];
1699 if (m2->mc_xcursor) {
1700 MDB_xcursor *mx, *mx2;
1701 mx = (MDB_xcursor *)(mc+1);
1702 mc->mc_xcursor = mx;
1703 mx2 = m2->mc_xcursor;
1704 mx->mx_db = mx2->mx_db;
1705 mx->mx_dbx = mx2->mx_dbx;
1706 mx->mx_dbflag = mx2->mx_dbflag;
1707 mx->mx_cursor.mc_txn = dst;
1708 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1709 mx->mx_cursor.mc_db = &mx->mx_db;
1710 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1711 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1712 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1713 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1714 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1715 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1716 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1717 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1720 mc->mc_xcursor = NULL;
1722 mc->mc_next = dst->mt_cursors[i];
1723 dst->mt_cursors[i] = mc;
1730 /** Merge shadow cursors back into parent's */
1732 mdb_cursor_merge(MDB_txn *txn)
1735 for (i=0; i<txn->mt_numdbs; i++) {
1736 if (txn->mt_cursors[i]) {
1738 while ((mc = txn->mt_cursors[i])) {
1739 txn->mt_cursors[i] = mc->mc_next;
1740 if (mc->mc_flags & C_SHADOW) {
1741 MDB_cursor *m2 = mc->mc_orig;
1743 m2->mc_snum = mc->mc_snum;
1744 m2->mc_top = mc->mc_top;
1745 for (j=0; j<mc->mc_snum; j++) {
1746 m2->mc_pg[j] = mc->mc_pg[j];
1747 m2->mc_ki[j] = mc->mc_ki[j];
1750 if (mc->mc_flags & C_ALLOCD)
1758 mdb_txn_reset0(MDB_txn *txn);
1760 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1761 * @param[in] txn the transaction handle to initialize
1762 * @return 0 on success, non-zero on failure. This can only
1763 * fail for read-only transactions, and then only if the
1764 * reader table is full.
1767 mdb_txn_renew0(MDB_txn *txn)
1769 MDB_env *env = txn->mt_env;
1774 txn->mt_numdbs = env->me_numdbs;
1775 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1777 if (txn->mt_flags & MDB_TXN_RDONLY) {
1778 if (env->me_flags & MDB_ROFS) {
1779 i = mdb_env_pick_meta(env);
1780 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1781 txn->mt_u.reader = NULL;
1783 MDB_reader *r = pthread_getspecific(env->me_txkey);
1785 pid_t pid = env->me_pid;
1786 pthread_t tid = pthread_self();
1789 for (i=0; i<env->me_txns->mti_numreaders; i++)
1790 if (env->me_txns->mti_readers[i].mr_pid == 0)
1792 if (i == env->me_maxreaders) {
1793 UNLOCK_MUTEX_R(env);
1794 return MDB_READERS_FULL;
1796 env->me_txns->mti_readers[i].mr_pid = pid;
1797 env->me_txns->mti_readers[i].mr_tid = tid;
1798 if (i >= env->me_txns->mti_numreaders)
1799 env->me_txns->mti_numreaders = i+1;
1800 /* Save numreaders for un-mutexed mdb_env_close() */
1801 env->me_numreaders = env->me_txns->mti_numreaders;
1802 UNLOCK_MUTEX_R(env);
1803 r = &env->me_txns->mti_readers[i];
1804 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1805 env->me_txns->mti_readers[i].mr_pid = 0;
1809 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1810 txn->mt_u.reader = r;
1812 txn->mt_toggle = txn->mt_txnid & 1;
1813 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1817 txn->mt_txnid = env->me_txns->mti_txnid;
1818 txn->mt_toggle = txn->mt_txnid & 1;
1819 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1822 if (txn->mt_txnid == mdb_debug_start)
1825 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1826 txn->mt_u.dirty_list = env->me_dirty_list;
1827 txn->mt_u.dirty_list[0].mid = 0;
1828 txn->mt_free_pgs = env->me_free_pgs;
1829 txn->mt_free_pgs[0] = 0;
1833 /* Copy the DB info and flags */
1834 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1835 for (i=2; i<txn->mt_numdbs; i++)
1836 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1837 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1838 if (txn->mt_numdbs > 2)
1839 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1841 if (env->me_maxpg < txn->mt_next_pgno) {
1842 mdb_txn_reset0(txn);
1843 return MDB_MAP_RESIZED;
1850 mdb_txn_renew(MDB_txn *txn)
1854 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1857 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1858 DPUTS("environment had fatal error, must shutdown!");
1862 rc = mdb_txn_renew0(txn);
1863 if (rc == MDB_SUCCESS) {
1864 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1865 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1866 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1872 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1876 int rc, size, tsize = sizeof(MDB_txn);
1878 if (env->me_flags & MDB_FATAL_ERROR) {
1879 DPUTS("environment had fatal error, must shutdown!");
1882 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1885 /* Nested transactions: Max 1 child, write txns only, no writemap */
1886 if (parent->mt_child ||
1887 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1888 (env->me_flags & MDB_WRITEMAP))
1892 tsize = sizeof(MDB_ntxn);
1894 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1895 if (!(flags & MDB_RDONLY))
1896 size += env->me_maxdbs * sizeof(MDB_cursor *);
1898 if ((txn = calloc(1, size)) == NULL) {
1899 DPRINTF("calloc: %s", strerror(ErrCode()));
1902 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1903 if (flags & MDB_RDONLY) {
1904 txn->mt_flags |= MDB_TXN_RDONLY;
1905 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1907 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1908 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1913 txn->mt_free_pgs = mdb_midl_alloc();
1914 if (!txn->mt_free_pgs) {
1918 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1919 if (!txn->mt_u.dirty_list) {
1920 free(txn->mt_free_pgs);
1924 txn->mt_txnid = parent->mt_txnid;
1925 txn->mt_toggle = parent->mt_toggle;
1926 txn->mt_dirty_room = parent->mt_dirty_room;
1927 txn->mt_u.dirty_list[0].mid = 0;
1928 txn->mt_free_pgs[0] = 0;
1929 txn->mt_next_pgno = parent->mt_next_pgno;
1930 parent->mt_child = txn;
1931 txn->mt_parent = parent;
1932 txn->mt_numdbs = parent->mt_numdbs;
1933 txn->mt_dbxs = parent->mt_dbxs;
1934 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1935 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1937 ntxn = (MDB_ntxn *)txn;
1938 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1939 if (env->me_pghead) {
1940 size = MDB_IDL_SIZEOF(env->me_pghead);
1941 env->me_pghead = malloc(size);
1943 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1947 env->me_pgfree = env->me_pghead;
1949 rc = mdb_cursor_shadow(parent, txn);
1951 mdb_txn_reset0(txn);
1953 rc = mdb_txn_renew0(txn);
1959 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1960 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1961 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1967 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1968 * @param[in] txn the transaction handle to reset
1971 mdb_txn_reset0(MDB_txn *txn)
1973 MDB_env *env = txn->mt_env;
1975 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1976 if (!(env->me_flags & MDB_ROFS))
1977 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1982 /* close(free) all cursors */
1983 for (i=0; i<txn->mt_numdbs; i++) {
1984 if (txn->mt_cursors[i]) {
1986 while ((mc = txn->mt_cursors[i])) {
1987 txn->mt_cursors[i] = mc->mc_next;
1988 if (mc->mc_flags & C_ALLOCD)
1994 if (!(env->me_flags & MDB_WRITEMAP)) {
1995 /* return all dirty pages to dpage list */
1996 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1997 dp = txn->mt_u.dirty_list[i].mptr;
1998 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1999 mdb_page_free(txn->mt_env, dp);
2001 /* large pages just get freed directly */
2002 VGMEMP_FREE(txn->mt_env, dp);
2008 free(env->me_pgfree);
2010 if (txn->mt_parent) {
2011 txn->mt_parent->mt_child = NULL;
2012 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2013 mdb_midl_free(txn->mt_free_pgs);
2014 free(txn->mt_u.dirty_list);
2017 if (mdb_midl_shrink(&txn->mt_free_pgs))
2018 env->me_free_pgs = txn->mt_free_pgs;
2021 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2022 txn->mt_env->me_pglast = 0;
2025 /* The writer mutex was locked in mdb_txn_begin. */
2026 UNLOCK_MUTEX_W(env);
2031 mdb_txn_reset(MDB_txn *txn)
2036 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2037 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2038 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2040 mdb_txn_reset0(txn);
2044 mdb_txn_abort(MDB_txn *txn)
2049 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2050 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2051 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2054 mdb_txn_abort(txn->mt_child);
2056 mdb_txn_reset0(txn);
2061 mdb_txn_commit(MDB_txn *txn)
2069 pgno_t next, freecnt;
2070 txnid_t oldpg_txnid, id;
2073 assert(txn != NULL);
2074 assert(txn->mt_env != NULL);
2076 if (txn->mt_child) {
2077 mdb_txn_commit(txn->mt_child);
2078 txn->mt_child = NULL;
2083 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2084 if (txn->mt_numdbs > env->me_numdbs) {
2085 /* update the DB flags */
2087 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2088 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2095 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2096 DPUTS("error flag is set, can't commit");
2098 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2103 if (txn->mt_parent) {
2104 MDB_txn *parent = txn->mt_parent;
2108 /* Append our free list to parent's */
2109 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2113 mdb_midl_free(txn->mt_free_pgs);
2115 parent->mt_next_pgno = txn->mt_next_pgno;
2116 parent->mt_flags = txn->mt_flags;
2118 /* Merge (and close) our cursors with parent's */
2119 mdb_cursor_merge(txn);
2121 /* Update parent's DB table. */
2122 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2123 memcpy(parent->mt_dbflags, txn->mt_dbflags, txn->mt_numdbs);
2124 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2126 dst = txn->mt_parent->mt_u.dirty_list;
2127 src = txn->mt_u.dirty_list;
2128 /* Find len = length of merging our dirty list with parent's */
2130 dst[0].mid = 0; /* simplify loops */
2131 if (parent->mt_parent) {
2132 len = x + src[0].mid;
2133 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2134 for (i = x; y && i; y--) {
2135 pgno_t yp = src[y].mid;
2136 while (yp < dst[i].mid)
2138 if (yp == dst[i].mid) {
2143 } else { /* Simplify the above for single-ancestor case */
2144 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2146 /* Merge our dirty list with parent's */
2148 for (i = len; y; dst[i--] = src[y--]) {
2149 pgno_t yp = src[y].mid;
2150 while (yp < dst[x].mid)
2151 dst[i--] = dst[x--];
2152 if (yp == dst[x].mid)
2153 free(dst[x--].mptr);
2157 free(txn->mt_u.dirty_list);
2158 parent->mt_dirty_room = txn->mt_dirty_room;
2160 txn->mt_parent->mt_child = NULL;
2161 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2166 if (txn != env->me_txn) {
2167 DPUTS("attempt to commit unknown transaction");
2172 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2175 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2176 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2178 /* Update DB root pointers */
2179 if (txn->mt_numdbs > 2) {
2182 data.mv_size = sizeof(MDB_db);
2184 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2185 for (i = 2; i < txn->mt_numdbs; i++) {
2186 if (txn->mt_dbflags[i] & DB_DIRTY) {
2187 data.mv_data = &txn->mt_dbs[i];
2188 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2195 /* Save the freelist as of this transaction to the freeDB. This
2196 * can change the freelist, so keep trying until it stabilizes.
2198 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2199 * except the code below can decrease env->me_pglast to split pghead.
2200 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2201 * can only appear in env->me_pghead when env->me_pglast increases.
2202 * Until then, the me_pghead pointer won't move but can become NULL.
2205 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2206 oldpg_txnid = id = 0;
2209 /* should only be one record now */
2210 if (env->me_pghead || env->me_pglast) {
2211 /* make sure first page of freeDB is touched and on freelist */
2212 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2213 if (rc && rc != MDB_NOTFOUND) {
2220 /* Delete IDLs we used from the free list */
2221 if (env->me_pglast) {
2226 rc = mdb_cursor_first(&mc, &key, NULL);
2229 oldpg_txnid = *(txnid_t *)key.mv_data;
2231 assert(oldpg_txnid <= env->me_pglast);
2233 rc = mdb_cursor_del(&mc, 0);
2236 } while (oldpg_txnid < env->me_pglast);
2239 /* Save IDL of pages freed by this txn, to freeDB */
2241 if (freecnt != txn->mt_free_pgs[0]) {
2244 /* make sure last page of freeDB is touched and on freelist */
2245 key.mv_size = MDB_MAXKEYSIZE+1;
2247 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2248 if (rc && rc != MDB_NOTFOUND)
2254 MDB_IDL idl = txn->mt_free_pgs;
2255 mdb_midl_sort(txn->mt_free_pgs);
2256 DPRINTF("IDL write txn %zu root %zu num %zu",
2257 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2258 for (i=1; i<=idl[0]; i++) {
2259 DPRINTF("IDL %zu", idl[i]);
2263 /* write to last page of freeDB */
2264 key.mv_size = sizeof(pgno_t);
2265 key.mv_data = &txn->mt_txnid;
2266 /* The free list can still grow during this call,
2267 * despite the pre-emptive touches above. So retry
2268 * until the reserved space remains big enough.
2271 assert(freecnt < txn->mt_free_pgs[0]);
2272 freecnt = txn->mt_free_pgs[0];
2273 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2274 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2277 } while (freecnt != txn->mt_free_pgs[0]);
2278 mdb_midl_sort(txn->mt_free_pgs);
2279 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2280 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2281 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2284 /* Put back page numbers we took from freeDB but did not use */
2285 if (env->me_pghead) {
2290 mop = env->me_pghead;
2291 id = env->me_pglast;
2292 key.mv_size = sizeof(id);
2294 /* These steps may grow the freelist again
2295 * due to freed overflow pages...
2300 if (orig > env->me_maxfree_1pg && id > 4)
2301 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2302 data.mv_size = (orig + 1) * sizeof(pgno_t);
2303 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2306 assert(!env->me_pghead || env->me_pglast);
2307 /* mop could have been used again here */
2308 if (id != env->me_pglast || env->me_pghead == NULL)
2309 goto again; /* was completely used up */
2310 assert(mop == env->me_pghead);
2311 } while (mop[0] < orig && --i);
2312 memcpy(data.mv_data, mop, data.mv_size);
2315 *(pgno_t *)data.mv_data = orig;
2316 mop[orig] = mop[0] - orig;
2317 env->me_pghead = mop += orig;
2318 /* Save more oldpages at the previous txnid. */
2319 assert(env->me_pglast == id && id == oldpg_txnid);
2320 env->me_pglast = --oldpg_txnid;
2324 /* Check for growth of freelist again */
2325 if (freecnt != txn->mt_free_pgs[0])
2328 free(env->me_pgfree);
2329 env->me_pghead = env->me_pgfree = NULL;
2331 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2332 if (mdb_midl_shrink(&txn->mt_free_pgs))
2333 env->me_free_pgs = txn->mt_free_pgs;
2340 if (env->me_flags & MDB_WRITEMAP) {
2341 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2342 dp = txn->mt_u.dirty_list[i].mptr;
2343 /* clear dirty flag */
2344 dp->mp_flags &= ~P_DIRTY;
2345 txn->mt_u.dirty_list[i].mid = 0;
2347 txn->mt_u.dirty_list[0].mid = 0;
2351 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2357 /* Windows actually supports scatter/gather I/O, but only on
2358 * unbuffered file handles. Since we're relying on the OS page
2359 * cache for all our data, that's self-defeating. So we just
2360 * write pages one at a time. We use the ov structure to set
2361 * the write offset, to at least save the overhead of a Seek
2365 memset(&ov, 0, sizeof(ov));
2366 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2368 dp = txn->mt_u.dirty_list[i].mptr;
2369 DPRINTF("committing page %zu", dp->mp_pgno);
2370 size = dp->mp_pgno * env->me_psize;
2371 ov.Offset = size & 0xffffffff;
2372 ov.OffsetHigh = size >> 16;
2373 ov.OffsetHigh >>= 16;
2374 /* clear dirty flag */
2375 dp->mp_flags &= ~P_DIRTY;
2376 wsize = env->me_psize;
2377 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2378 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2381 DPRINTF("WriteFile: %d", n);
2388 struct iovec iov[MDB_COMMIT_PAGES];
2392 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2393 dp = txn->mt_u.dirty_list[i].mptr;
2394 if (dp->mp_pgno != next) {
2396 rc = writev(env->me_fd, iov, n);
2400 DPUTS("short write, filesystem full?");
2402 DPRINTF("writev: %s", strerror(n));
2409 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2412 DPRINTF("committing page %zu", dp->mp_pgno);
2413 iov[n].iov_len = env->me_psize;
2414 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2415 iov[n].iov_base = (char *)dp;
2416 size += iov[n].iov_len;
2417 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2418 /* clear dirty flag */
2419 dp->mp_flags &= ~P_DIRTY;
2420 if (++n >= MDB_COMMIT_PAGES) {
2430 rc = writev(env->me_fd, iov, n);
2434 DPUTS("short write, filesystem full?");
2436 DPRINTF("writev: %s", strerror(n));
2443 /* Drop the dirty pages.
2445 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2446 dp = txn->mt_u.dirty_list[i].mptr;
2447 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2448 mdb_page_free(txn->mt_env, dp);
2450 VGMEMP_FREE(txn->mt_env, dp);
2453 txn->mt_u.dirty_list[i].mid = 0;
2455 txn->mt_u.dirty_list[0].mid = 0;
2458 if ((n = mdb_env_sync(env, 0)) != 0 ||
2459 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2467 if (txn->mt_numdbs > env->me_numdbs) {
2468 /* update the DB flags */
2470 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2471 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2475 UNLOCK_MUTEX_W(env);
2481 /** Read the environment parameters of a DB environment before
2482 * mapping it into memory.
2483 * @param[in] env the environment handle
2484 * @param[out] meta address of where to store the meta information
2485 * @return 0 on success, non-zero on failure.
2488 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2495 /* We don't know the page size yet, so use a minimum value.
2496 * Read both meta pages so we can use the latest one.
2499 for (i=0; i<2; i++) {
2501 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2503 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2508 else if (rc != MDB_PAGESIZE) {
2512 DPRINTF("read: %s", strerror(err));
2516 p = (MDB_page *)&pbuf;
2518 if (!F_ISSET(p->mp_flags, P_META)) {
2519 DPRINTF("page %zu not a meta page", p->mp_pgno);
2524 if (m->mm_magic != MDB_MAGIC) {
2525 DPUTS("meta has invalid magic");
2529 if (m->mm_version != MDB_VERSION) {
2530 DPRINTF("database is version %u, expected version %u",
2531 m->mm_version, MDB_VERSION);
2532 return MDB_VERSION_MISMATCH;
2536 if (m->mm_txnid > meta->mm_txnid)
2537 memcpy(meta, m, sizeof(*m));
2539 memcpy(meta, m, sizeof(*m));
2541 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2543 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2551 /** Write the environment parameters of a freshly created DB environment.
2552 * @param[in] env the environment handle
2553 * @param[out] meta address of where to store the meta information
2554 * @return 0 on success, non-zero on failure.
2557 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2564 DPUTS("writing new meta page");
2566 GET_PAGESIZE(psize);
2568 meta->mm_magic = MDB_MAGIC;
2569 meta->mm_version = MDB_VERSION;
2570 meta->mm_mapsize = env->me_mapsize;
2571 meta->mm_psize = psize;
2572 meta->mm_last_pg = 1;
2573 meta->mm_flags = env->me_flags & 0xffff;
2574 meta->mm_flags |= MDB_INTEGERKEY;
2575 meta->mm_dbs[0].md_root = P_INVALID;
2576 meta->mm_dbs[1].md_root = P_INVALID;
2578 p = calloc(2, psize);
2580 p->mp_flags = P_META;
2583 memcpy(m, meta, sizeof(*meta));
2585 q = (MDB_page *)((char *)p + psize);
2588 q->mp_flags = P_META;
2591 memcpy(m, meta, sizeof(*meta));
2596 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2597 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2598 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2601 lseek(env->me_fd, 0, SEEK_SET);
2602 rc = write(env->me_fd, p, psize * 2);
2603 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2609 /** Update the environment info to commit a transaction.
2610 * @param[in] txn the transaction that's being committed
2611 * @return 0 on success, non-zero on failure.
2614 mdb_env_write_meta(MDB_txn *txn)
2617 MDB_meta meta, metab, *mp;
2619 int rc, len, toggle;
2626 assert(txn != NULL);
2627 assert(txn->mt_env != NULL);
2629 toggle = !txn->mt_toggle;
2630 DPRINTF("writing meta page %d for root page %zu",
2631 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2634 mp = env->me_metas[toggle];
2636 if (env->me_flags & MDB_WRITEMAP) {
2637 /* Persist any increases of mapsize config */
2638 if (env->me_mapsize > mp->mm_mapsize)
2639 mp->mm_mapsize = env->me_mapsize;
2640 mp->mm_dbs[0] = txn->mt_dbs[0];
2641 mp->mm_dbs[1] = txn->mt_dbs[1];
2642 mp->mm_last_pg = txn->mt_next_pgno - 1;
2643 mp->mm_txnid = txn->mt_txnid;
2644 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2645 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2648 ptr += env->me_psize;
2649 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2656 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2657 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2659 ptr = (char *)&meta;
2660 if (env->me_mapsize > mp->mm_mapsize) {
2661 /* Persist any increases of mapsize config */
2662 meta.mm_mapsize = env->me_mapsize;
2663 off = offsetof(MDB_meta, mm_mapsize);
2665 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2667 len = sizeof(MDB_meta) - off;
2670 meta.mm_dbs[0] = txn->mt_dbs[0];
2671 meta.mm_dbs[1] = txn->mt_dbs[1];
2672 meta.mm_last_pg = txn->mt_next_pgno - 1;
2673 meta.mm_txnid = txn->mt_txnid;
2676 off += env->me_psize;
2679 /* Write to the SYNC fd */
2680 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2681 env->me_fd : env->me_mfd;
2684 memset(&ov, 0, sizeof(ov));
2686 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2689 rc = pwrite(mfd, ptr, len, off);
2694 DPUTS("write failed, disk error?");
2695 /* On a failure, the pagecache still contains the new data.
2696 * Write some old data back, to prevent it from being used.
2697 * Use the non-SYNC fd; we know it will fail anyway.
2699 meta.mm_last_pg = metab.mm_last_pg;
2700 meta.mm_txnid = metab.mm_txnid;
2702 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2704 r2 = pwrite(env->me_fd, ptr, len, off);
2707 env->me_flags |= MDB_FATAL_ERROR;
2711 /* Memory ordering issues are irrelevant; since the entire writer
2712 * is wrapped by wmutex, all of these changes will become visible
2713 * after the wmutex is unlocked. Since the DB is multi-version,
2714 * readers will get consistent data regardless of how fresh or
2715 * how stale their view of these values is.
2717 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2722 /** Check both meta pages to see which one is newer.
2723 * @param[in] env the environment handle
2724 * @return meta toggle (0 or 1).
2727 mdb_env_pick_meta(const MDB_env *env)
2729 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2733 mdb_env_create(MDB_env **env)
2737 e = calloc(1, sizeof(MDB_env));
2741 e->me_free_pgs = mdb_midl_alloc();
2742 if (!e->me_free_pgs) {
2746 e->me_maxreaders = DEFAULT_READERS;
2748 e->me_fd = INVALID_HANDLE_VALUE;
2749 e->me_lfd = INVALID_HANDLE_VALUE;
2750 e->me_mfd = INVALID_HANDLE_VALUE;
2751 #ifdef MDB_USE_POSIX_SEM
2752 e->me_rmutex = SEM_FAILED;
2753 e->me_wmutex = SEM_FAILED;
2755 e->me_pid = getpid();
2756 VGMEMP_CREATE(e,0,0);
2762 mdb_env_set_mapsize(MDB_env *env, size_t size)
2766 env->me_mapsize = size;
2768 env->me_maxpg = env->me_mapsize / env->me_psize;
2773 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2777 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2782 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2784 if (env->me_map || readers < 1)
2786 env->me_maxreaders = readers;
2791 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2793 if (!env || !readers)
2795 *readers = env->me_maxreaders;
2799 /** Further setup required for opening an MDB environment
2802 mdb_env_open2(MDB_env *env)
2804 unsigned int flags = env->me_flags;
2805 int i, newenv = 0, prot;
2809 memset(&meta, 0, sizeof(meta));
2811 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2814 DPUTS("new mdbenv");
2818 /* Was a mapsize configured? */
2819 if (!env->me_mapsize) {
2820 /* If this is a new environment, take the default,
2821 * else use the size recorded in the existing env.
2823 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2824 } else if (env->me_mapsize < meta.mm_mapsize) {
2825 /* If the configured size is smaller, make sure it's
2826 * still big enough. Silently round up to minimum if not.
2828 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2829 if (env->me_mapsize < minsize)
2830 env->me_mapsize = minsize;
2836 LONG sizelo, sizehi;
2837 sizelo = env->me_mapsize & 0xffffffff;
2838 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2840 /* Windows won't create mappings for zero length files.
2841 * Just allocate the maxsize right now.
2844 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2845 if (!SetEndOfFile(env->me_fd))
2847 SetFilePointer(env->me_fd, 0, NULL, 0);
2849 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2850 PAGE_READWRITE : PAGE_READONLY,
2851 sizehi, sizelo, NULL);
2854 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2855 FILE_MAP_WRITE : FILE_MAP_READ,
2856 0, 0, env->me_mapsize, meta.mm_address);
2864 if (flags & MDB_WRITEMAP) {
2866 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2869 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2871 if (env->me_map == MAP_FAILED) {
2878 if (flags & MDB_FIXEDMAP)
2879 meta.mm_address = env->me_map;
2880 i = mdb_env_init_meta(env, &meta);
2881 if (i != MDB_SUCCESS) {
2884 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2885 /* Can happen because the address argument to mmap() is just a
2886 * hint. mmap() can pick another, e.g. if the range is in use.
2887 * The MAP_FIXED flag would prevent that, but then mmap could
2888 * instead unmap existing pages to make room for the new map.
2890 return EBUSY; /* TODO: Make a new MDB_* error code? */
2892 env->me_psize = meta.mm_psize;
2893 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2895 env->me_maxpg = env->me_mapsize / env->me_psize;
2897 p = (MDB_page *)env->me_map;
2898 env->me_metas[0] = METADATA(p);
2899 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2903 int toggle = mdb_env_pick_meta(env);
2904 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2906 DPRINTF("opened database version %u, pagesize %u",
2907 env->me_metas[0]->mm_version, env->me_psize);
2908 DPRINTF("using meta page %d", toggle);
2909 DPRINTF("depth: %u", db->md_depth);
2910 DPRINTF("entries: %zu", db->md_entries);
2911 DPRINTF("branch pages: %zu", db->md_branch_pages);
2912 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2913 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2914 DPRINTF("root: %zu", db->md_root);
2922 /** Release a reader thread's slot in the reader lock table.
2923 * This function is called automatically when a thread exits.
2924 * @param[in] ptr This points to the slot in the reader lock table.
2927 mdb_env_reader_dest(void *ptr)
2929 MDB_reader *reader = ptr;
2935 /** Junk for arranging thread-specific callbacks on Windows. This is
2936 * necessarily platform and compiler-specific. Windows supports up
2937 * to 1088 keys. Let's assume nobody opens more than 64 environments
2938 * in a single process, for now. They can override this if needed.
2940 #ifndef MAX_TLS_KEYS
2941 #define MAX_TLS_KEYS 64
2943 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2944 static int mdb_tls_nkeys;
2946 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2950 case DLL_PROCESS_ATTACH: break;
2951 case DLL_THREAD_ATTACH: break;
2952 case DLL_THREAD_DETACH:
2953 for (i=0; i<mdb_tls_nkeys; i++) {
2954 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2955 mdb_env_reader_dest(r);
2958 case DLL_PROCESS_DETACH: break;
2963 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2965 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2969 /* Force some symbol references.
2970 * _tls_used forces the linker to create the TLS directory if not already done
2971 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2973 #pragma comment(linker, "/INCLUDE:_tls_used")
2974 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2975 #pragma const_seg(".CRT$XLB")
2976 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2977 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2980 #pragma comment(linker, "/INCLUDE:__tls_used")
2981 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2982 #pragma data_seg(".CRT$XLB")
2983 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2985 #endif /* WIN 32/64 */
2986 #endif /* !__GNUC__ */
2989 /** Downgrade the exclusive lock on the region back to shared */
2991 mdb_env_share_locks(MDB_env *env, int *excl)
2993 int rc = 0, toggle = mdb_env_pick_meta(env);
2995 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3000 /* First acquire a shared lock. The Unlock will
3001 * then release the existing exclusive lock.
3003 memset(&ov, 0, sizeof(ov));
3004 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3007 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3013 struct flock lock_info;
3014 /* The shared lock replaces the existing lock */
3015 memset((void *)&lock_info, 0, sizeof(lock_info));
3016 lock_info.l_type = F_RDLCK;
3017 lock_info.l_whence = SEEK_SET;
3018 lock_info.l_start = 0;
3019 lock_info.l_len = 1;
3020 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3021 (rc = ErrCode()) == EINTR) ;
3022 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3029 /** Try to get exlusive lock, otherwise shared.
3030 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3033 mdb_env_excl_lock(MDB_env *env, int *excl)
3037 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3041 memset(&ov, 0, sizeof(ov));
3042 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3049 struct flock lock_info;
3050 memset((void *)&lock_info, 0, sizeof(lock_info));
3051 lock_info.l_type = F_WRLCK;
3052 lock_info.l_whence = SEEK_SET;
3053 lock_info.l_start = 0;
3054 lock_info.l_len = 1;
3055 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3056 (rc = ErrCode()) == EINTR) ;
3060 # ifdef MDB_USE_POSIX_SEM
3061 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3064 lock_info.l_type = F_RDLCK;
3065 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3066 (rc = ErrCode()) == EINTR) ;
3074 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3076 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3078 * @(#) $Revision: 5.1 $
3079 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3080 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3082 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3086 * Please do not copyright this code. This code is in the public domain.
3088 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3089 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3090 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3091 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3092 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3093 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3094 * PERFORMANCE OF THIS SOFTWARE.
3097 * chongo <Landon Curt Noll> /\oo/\
3098 * http://www.isthe.com/chongo/
3100 * Share and Enjoy! :-)
3103 typedef unsigned long long mdb_hash_t;
3104 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3106 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3107 * @param[in] str string to hash
3108 * @param[in] hval initial value for hash
3109 * @return 64 bit hash
3111 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3112 * hval arg on the first call.
3115 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3117 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3118 unsigned char *end = s + val->mv_size;
3120 * FNV-1a hash each octet of the string
3123 /* xor the bottom with the current octet */
3124 hval ^= (mdb_hash_t)*s++;
3126 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3127 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3128 (hval << 7) + (hval << 8) + (hval << 40);
3130 /* return our new hash value */
3134 /** Hash the string and output the hash in hex.
3135 * @param[in] str string to hash
3136 * @param[out] hexbuf an array of 17 chars to hold the hash
3139 mdb_hash_hex(MDB_val *val, char *hexbuf)
3142 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3143 for (i=0; i<8; i++) {
3144 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3150 /** Open and/or initialize the lock region for the environment.
3151 * @param[in] env The MDB environment.
3152 * @param[in] lpath The pathname of the file used for the lock region.
3153 * @param[in] mode The Unix permissions for the file, if we create it.
3154 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3155 * @return 0 on success, non-zero on failure.
3158 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3166 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3167 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3168 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3170 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3171 env->me_flags |= MDB_ROFS;
3176 /* Try to get exclusive lock. If we succeed, then
3177 * nobody is using the lock region and we should initialize it.
3179 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3180 size = GetFileSize(env->me_lfd, NULL);
3186 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3188 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3189 env->me_flags |= MDB_ROFS;
3194 /* Lose record locks when exec*() */
3195 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3196 fcntl(env->me_lfd, F_SETFD, fdflags);
3198 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3199 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3201 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3202 env->me_flags |= MDB_ROFS;
3209 /* Try to get exclusive lock. If we succeed, then
3210 * nobody is using the lock region and we should initialize it.
3212 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3214 size = lseek(env->me_lfd, 0, SEEK_END);
3216 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3217 if (size < rsize && *excl > 0) {
3219 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3220 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3222 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3226 size = rsize - sizeof(MDB_txninfo);
3227 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3232 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3234 if (!mh) goto fail_errno;
3235 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3237 if (!env->me_txns) goto fail_errno;
3239 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3241 if (m == MAP_FAILED) goto fail_errno;
3247 BY_HANDLE_FILE_INFORMATION stbuf;
3256 if (!mdb_sec_inited) {
3257 InitializeSecurityDescriptor(&mdb_null_sd,
3258 SECURITY_DESCRIPTOR_REVISION);
3259 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3260 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3261 mdb_all_sa.bInheritHandle = FALSE;
3262 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3265 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3266 idbuf.volume = stbuf.dwVolumeSerialNumber;
3267 idbuf.nhigh = stbuf.nFileIndexHigh;
3268 idbuf.nlow = stbuf.nFileIndexLow;
3269 val.mv_data = &idbuf;
3270 val.mv_size = sizeof(idbuf);
3271 mdb_hash_hex(&val, hexbuf);
3272 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3273 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3274 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3275 if (!env->me_rmutex) goto fail_errno;
3276 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3277 if (!env->me_wmutex) goto fail_errno;
3278 #elif defined(MDB_USE_POSIX_SEM)
3287 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3288 idbuf.dev = stbuf.st_dev;
3289 idbuf.ino = stbuf.st_ino;
3290 val.mv_data = &idbuf;
3291 val.mv_size = sizeof(idbuf);
3292 mdb_hash_hex(&val, hexbuf);
3293 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3294 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3295 /* Clean up after a previous run, if needed: Try to
3296 * remove both semaphores before doing anything else.
3298 sem_unlink(env->me_txns->mti_rmname);
3299 sem_unlink(env->me_txns->mti_wmname);
3300 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3301 O_CREAT|O_EXCL, mode, 1);
3302 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3303 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3304 O_CREAT|O_EXCL, mode, 1);
3305 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3306 #else /* MDB_USE_POSIX_SEM */
3307 pthread_mutexattr_t mattr;
3309 if ((rc = pthread_mutexattr_init(&mattr))
3310 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3311 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3312 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3314 pthread_mutexattr_destroy(&mattr);
3315 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3317 env->me_txns->mti_version = MDB_VERSION;
3318 env->me_txns->mti_magic = MDB_MAGIC;
3319 env->me_txns->mti_txnid = 0;
3320 env->me_txns->mti_numreaders = 0;
3323 if (env->me_txns->mti_magic != MDB_MAGIC) {
3324 DPUTS("lock region has invalid magic");
3328 if (env->me_txns->mti_version != MDB_VERSION) {
3329 DPRINTF("lock region is version %u, expected version %u",
3330 env->me_txns->mti_version, MDB_VERSION);
3331 rc = MDB_VERSION_MISMATCH;
3335 if (rc != EACCES && rc != EAGAIN) {
3339 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3340 if (!env->me_rmutex) goto fail_errno;
3341 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3342 if (!env->me_wmutex) goto fail_errno;
3343 #elif defined(MDB_USE_POSIX_SEM)
3344 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3345 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3346 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3347 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3358 /** The name of the lock file in the DB environment */
3359 #define LOCKNAME "/lock.mdb"
3360 /** The name of the data file in the DB environment */
3361 #define DATANAME "/data.mdb"
3362 /** The suffix of the lock file when no subdir is used */
3363 #define LOCKSUFF "-lock"
3364 /** Only a subset of the @ref mdb_env flags can be changed
3365 * at runtime. Changing other flags requires closing the
3366 * environment and re-opening it with the new flags.
3368 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3369 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3372 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3374 int oflags, rc, len, excl;
3375 char *lpath, *dpath;
3377 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3381 if (flags & MDB_NOSUBDIR) {
3382 rc = len + sizeof(LOCKSUFF) + len + 1;
3384 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3389 if (flags & MDB_NOSUBDIR) {
3390 dpath = lpath + len + sizeof(LOCKSUFF);
3391 sprintf(lpath, "%s" LOCKSUFF, path);
3392 strcpy(dpath, path);
3394 dpath = lpath + len + sizeof(LOCKNAME);
3395 sprintf(lpath, "%s" LOCKNAME, path);
3396 sprintf(dpath, "%s" DATANAME, path);
3399 flags |= env->me_flags;
3400 /* silently ignore WRITEMAP if we're only getting read access */
3401 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3402 flags ^= MDB_WRITEMAP;
3403 env->me_flags = flags |= MDB_ENV_ACTIVE;
3405 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3410 if (F_ISSET(flags, MDB_RDONLY)) {
3411 oflags = GENERIC_READ;
3412 len = OPEN_EXISTING;
3414 oflags = GENERIC_READ|GENERIC_WRITE;
3417 mode = FILE_ATTRIBUTE_NORMAL;
3418 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3419 NULL, len, mode, NULL);
3421 if (F_ISSET(flags, MDB_RDONLY))
3424 oflags = O_RDWR | O_CREAT;
3426 env->me_fd = open(dpath, oflags, mode);
3428 if (env->me_fd == INVALID_HANDLE_VALUE) {
3433 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3434 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3435 env->me_mfd = env->me_fd;
3437 /* Synchronous fd for meta writes. Needed even with
3438 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3441 env->me_mfd = CreateFile(dpath, oflags,
3442 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3443 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3445 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3447 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3452 DPRINTF("opened dbenv %p", (void *) env);
3453 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3456 env->me_numdbs = 2; /* this notes that me_txkey was set */
3458 /* Windows TLS callbacks need help finding their TLS info. */
3459 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3460 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3467 rc = mdb_env_share_locks(env, &excl);
3471 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3472 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3473 env->me_path = strdup(path);
3474 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3480 mdb_env_close0(env, excl);
3486 /** Destroy resources from mdb_env_open() and clear our readers */
3488 mdb_env_close0(MDB_env *env, int excl)
3492 if (!(env->me_flags & MDB_ENV_ACTIVE))
3495 free(env->me_dbflags);
3499 if (env->me_numdbs) {
3500 pthread_key_delete(env->me_txkey);
3502 /* Delete our key from the global list */
3503 for (i=0; i<mdb_tls_nkeys; i++)
3504 if (mdb_tls_keys[i] == env->me_txkey) {
3505 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3513 munmap(env->me_map, env->me_mapsize);
3515 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3517 if (env->me_fd != INVALID_HANDLE_VALUE)
3520 pid_t pid = env->me_pid;
3521 /* Clearing readers is done in this function because
3522 * me_txkey with its destructor must be disabled first.
3524 for (i = env->me_numreaders; --i >= 0; )
3525 if (env->me_txns->mti_readers[i].mr_pid == pid)
3526 env->me_txns->mti_readers[i].mr_pid = 0;
3528 if (env->me_rmutex) {
3529 CloseHandle(env->me_rmutex);
3530 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3532 /* Windows automatically destroys the mutexes when
3533 * the last handle closes.
3535 #elif defined(MDB_USE_POSIX_SEM)
3536 if (env->me_rmutex != SEM_FAILED) {
3537 sem_close(env->me_rmutex);
3538 if (env->me_wmutex != SEM_FAILED)
3539 sem_close(env->me_wmutex);
3540 /* If we have the filelock: If we are the
3541 * only remaining user, clean up semaphores.
3544 mdb_env_excl_lock(env, &excl);
3546 sem_unlink(env->me_txns->mti_rmname);
3547 sem_unlink(env->me_txns->mti_wmname);
3551 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3553 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3556 /* Unlock the lockfile. Windows would have unlocked it
3557 * after closing anyway, but not necessarily at once.
3559 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3565 env->me_flags &= ~MDB_ENV_ACTIVE;
3569 mdb_env_copy(MDB_env *env, const char *path)
3571 MDB_txn *txn = NULL;
3575 HANDLE newfd = INVALID_HANDLE_VALUE;
3577 if (env->me_flags & MDB_NOSUBDIR) {
3578 lpath = (char *)path;
3581 len += sizeof(DATANAME);
3582 lpath = malloc(len);
3585 sprintf(lpath, "%s" DATANAME, path);
3588 /* The destination path must exist, but the destination file must not.
3589 * We don't want the OS to cache the writes, since the source data is
3590 * already in the OS cache.
3593 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3594 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3596 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3602 if (!(env->me_flags & MDB_NOSUBDIR))
3604 if (newfd == INVALID_HANDLE_VALUE) {
3609 #ifdef F_NOCACHE /* __APPLE__ */
3610 rc = fcntl(newfd, F_NOCACHE, 1);
3617 /* Do the lock/unlock of the reader mutex before starting the
3618 * write txn. Otherwise other read txns could block writers.
3620 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3624 if (!(env->me_flags & MDB_ROFS)) {
3625 /* We must start the actual read txn after blocking writers */
3626 mdb_txn_reset0(txn);
3628 /* Temporarily block writers until we snapshot the meta pages */
3631 rc = mdb_txn_renew0(txn);
3633 UNLOCK_MUTEX_W(env);
3638 wsize = env->me_psize * 2;
3642 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3643 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3646 rc = write(newfd, env->me_map, wsize);
3647 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3649 if (! (env->me_flags & MDB_ROFS))
3650 UNLOCK_MUTEX_W(env);
3655 ptr = env->me_map + wsize;
3656 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3657 #define MAX_WRITE 2147483648U
3661 if (wsize > MAX_WRITE)
3665 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3666 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3675 if (wsize > MAX_WRITE)
3679 wres = write(newfd, ptr, w2);
3680 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3689 if (newfd != INVALID_HANDLE_VALUE)
3696 mdb_env_close(MDB_env *env)
3703 VGMEMP_DESTROY(env);
3704 while ((dp = env->me_dpages) != NULL) {
3705 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3706 env->me_dpages = dp->mp_next;
3710 mdb_env_close0(env, 0);
3711 mdb_midl_free(env->me_free_pgs);
3715 /** Compare two items pointing at aligned size_t's */
3717 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3719 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3720 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3723 /** Compare two items pointing at aligned int's */
3725 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3727 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3728 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3731 /** Compare two items pointing at ints of unknown alignment.
3732 * Nodes and keys are guaranteed to be 2-byte aligned.
3735 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3737 #if BYTE_ORDER == LITTLE_ENDIAN
3738 unsigned short *u, *c;
3741 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3742 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3745 } while(!x && u > (unsigned short *)a->mv_data);
3748 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3752 /** Compare two items lexically */
3754 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3761 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3767 diff = memcmp(a->mv_data, b->mv_data, len);
3768 return diff ? diff : len_diff<0 ? -1 : len_diff;
3771 /** Compare two items in reverse byte order */
3773 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3775 const unsigned char *p1, *p2, *p1_lim;
3779 p1_lim = (const unsigned char *)a->mv_data;
3780 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3781 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3783 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3789 while (p1 > p1_lim) {
3790 diff = *--p1 - *--p2;
3794 return len_diff<0 ? -1 : len_diff;
3797 /** Search for key within a page, using binary search.
3798 * Returns the smallest entry larger or equal to the key.
3799 * If exactp is non-null, stores whether the found entry was an exact match
3800 * in *exactp (1 or 0).
3801 * Updates the cursor index with the index of the found entry.
3802 * If no entry larger or equal to the key is found, returns NULL.
3805 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3807 unsigned int i = 0, nkeys;
3810 MDB_page *mp = mc->mc_pg[mc->mc_top];
3811 MDB_node *node = NULL;
3816 nkeys = NUMKEYS(mp);
3821 COPY_PGNO(pgno, mp->mp_pgno);
3822 DPRINTF("searching %u keys in %s %spage %zu",
3823 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3830 low = IS_LEAF(mp) ? 0 : 1;
3832 cmp = mc->mc_dbx->md_cmp;
3834 /* Branch pages have no data, so if using integer keys,
3835 * alignment is guaranteed. Use faster mdb_cmp_int.
3837 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3838 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3845 nodekey.mv_size = mc->mc_db->md_pad;
3846 node = NODEPTR(mp, 0); /* fake */
3847 while (low <= high) {
3848 i = (low + high) >> 1;
3849 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3850 rc = cmp(key, &nodekey);
3851 DPRINTF("found leaf index %u [%s], rc = %i",
3852 i, DKEY(&nodekey), rc);
3861 while (low <= high) {
3862 i = (low + high) >> 1;
3864 node = NODEPTR(mp, i);
3865 nodekey.mv_size = NODEKSZ(node);
3866 nodekey.mv_data = NODEKEY(node);
3868 rc = cmp(key, &nodekey);
3871 DPRINTF("found leaf index %u [%s], rc = %i",
3872 i, DKEY(&nodekey), rc);
3874 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3875 i, DKEY(&nodekey), NODEPGNO(node), rc);
3886 if (rc > 0) { /* Found entry is less than the key. */
3887 i++; /* Skip to get the smallest entry larger than key. */
3889 node = NODEPTR(mp, i);
3892 *exactp = (rc == 0);
3893 /* store the key index */
3894 mc->mc_ki[mc->mc_top] = i;
3896 /* There is no entry larger or equal to the key. */
3899 /* nodeptr is fake for LEAF2 */
3905 mdb_cursor_adjust(MDB_cursor *mc, func)
3909 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3910 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3917 /** Pop a page off the top of the cursor's stack. */
3919 mdb_cursor_pop(MDB_cursor *mc)
3922 #ifndef MDB_DEBUG_SKIP
3923 MDB_page *top = mc->mc_pg[mc->mc_top];
3929 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3930 mc->mc_dbi, (void *) mc);
3934 /** Push a page onto the top of the cursor's stack. */
3936 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3938 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3939 mc->mc_dbi, (void *) mc);
3941 if (mc->mc_snum >= CURSOR_STACK) {
3942 assert(mc->mc_snum < CURSOR_STACK);
3943 return MDB_CURSOR_FULL;
3946 mc->mc_top = mc->mc_snum++;
3947 mc->mc_pg[mc->mc_top] = mp;
3948 mc->mc_ki[mc->mc_top] = 0;
3953 /** Find the address of the page corresponding to a given page number.
3954 * @param[in] txn the transaction for this access.
3955 * @param[in] pgno the page number for the page to retrieve.
3956 * @param[out] ret address of a pointer where the page's address will be stored.
3957 * @return 0 on success, non-zero on failure.
3960 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3964 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3965 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3969 MDB_ID2L dl = tx2->mt_u.dirty_list;
3971 unsigned x = mdb_mid2l_search(dl, pgno);
3972 if (x <= dl[0].mid && dl[x].mid == pgno) {
3977 } while ((tx2 = tx2->mt_parent) != NULL);
3980 if (pgno < txn->mt_next_pgno) {
3981 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3983 DPRINTF("page %zu not found", pgno);
3989 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3992 /** Search for the page a given key should be in.
3993 * Pushes parent pages on the cursor stack. This function continues a
3994 * search on a cursor that has already been initialized. (Usually by
3995 * #mdb_page_search() but also by #mdb_node_move().)
3996 * @param[in,out] mc the cursor for this operation.
3997 * @param[in] key the key to search for. If NULL, search for the lowest
3998 * page. (This is used by #mdb_cursor_first().)
3999 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4000 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4001 * @return 0 on success, non-zero on failure.
4004 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4006 MDB_page *mp = mc->mc_pg[mc->mc_top];
4011 while (IS_BRANCH(mp)) {
4015 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4016 assert(NUMKEYS(mp) > 1);
4017 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4019 if (key == NULL) /* Initialize cursor to first page. */
4021 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4022 /* cursor to last page */
4026 node = mdb_node_search(mc, key, &exact);
4028 i = NUMKEYS(mp) - 1;
4030 i = mc->mc_ki[mc->mc_top];
4039 DPRINTF("following index %u for key [%s]",
4041 assert(i < NUMKEYS(mp));
4042 node = NODEPTR(mp, i);
4044 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
4047 mc->mc_ki[mc->mc_top] = i;
4048 if ((rc = mdb_cursor_push(mc, mp)))
4052 if ((rc = mdb_page_touch(mc)) != 0)
4054 mp = mc->mc_pg[mc->mc_top];
4059 DPRINTF("internal error, index points to a %02X page!?",
4061 return MDB_CORRUPTED;
4064 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4065 key ? DKEY(key) : NULL);
4070 /** Search for the page a given key should be in.
4071 * Pushes parent pages on the cursor stack. This function just sets up
4072 * the search; it finds the root page for \b mc's database and sets this
4073 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4074 * called to complete the search.
4075 * @param[in,out] mc the cursor for this operation.
4076 * @param[in] key the key to search for. If NULL, search for the lowest
4077 * page. (This is used by #mdb_cursor_first().)
4078 * @param[in] modify If true, visited pages are updated with new page numbers.
4079 * @return 0 on success, non-zero on failure.
4082 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4087 /* Make sure the txn is still viable, then find the root from
4088 * the txn's db table.
4090 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4091 DPUTS("transaction has failed, must abort");
4094 /* Make sure we're using an up-to-date root */
4095 if (mc->mc_dbi > MAIN_DBI) {
4096 if ((*mc->mc_dbflag & DB_STALE) ||
4097 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4099 unsigned char dbflag = 0;
4100 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4101 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4104 if (*mc->mc_dbflag & DB_STALE) {
4107 MDB_node *leaf = mdb_node_search(&mc2,
4108 &mc->mc_dbx->md_name, &exact);
4110 return MDB_NOTFOUND;
4111 mdb_node_read(mc->mc_txn, leaf, &data);
4112 /* The txn may not know this DBI, or another process may
4113 * have dropped and recreated the DB with other flags.
4115 if (mc->mc_db->md_flags != *(uint16_t *)
4116 ((char *) data.mv_data + offsetof(MDB_db, md_flags)))
4117 return MDB_INCOMPATIBLE;
4118 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4120 if (flags & MDB_PS_MODIFY)
4122 *mc->mc_dbflag = dbflag;
4125 root = mc->mc_db->md_root;
4127 if (root == P_INVALID) { /* Tree is empty. */
4128 DPUTS("tree is empty");
4129 return MDB_NOTFOUND;
4134 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4135 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4141 DPRINTF("db %u root page %zu has flags 0x%X",
4142 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4144 if (flags & MDB_PS_MODIFY) {
4145 if ((rc = mdb_page_touch(mc)))
4149 if (flags & MDB_PS_ROOTONLY)
4152 return mdb_page_search_root(mc, key, flags);
4155 /** Return the data associated with a given node.
4156 * @param[in] txn The transaction for this operation.
4157 * @param[in] leaf The node being read.
4158 * @param[out] data Updated to point to the node's data.
4159 * @return 0 on success, non-zero on failure.
4162 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4164 MDB_page *omp; /* overflow page */
4168 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4169 data->mv_size = NODEDSZ(leaf);
4170 data->mv_data = NODEDATA(leaf);
4174 /* Read overflow data.
4176 data->mv_size = NODEDSZ(leaf);
4177 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4178 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4179 DPRINTF("read overflow page %zu failed", pgno);
4182 data->mv_data = METADATA(omp);
4188 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4189 MDB_val *key, MDB_val *data)
4198 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4200 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4203 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4207 mdb_cursor_init(&mc, txn, dbi, &mx);
4208 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4211 /** Find a sibling for a page.
4212 * Replaces the page at the top of the cursor's stack with the
4213 * specified sibling, if one exists.
4214 * @param[in] mc The cursor for this operation.
4215 * @param[in] move_right Non-zero if the right sibling is requested,
4216 * otherwise the left sibling.
4217 * @return 0 on success, non-zero on failure.
4220 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4226 if (mc->mc_snum < 2) {
4227 return MDB_NOTFOUND; /* root has no siblings */
4231 DPRINTF("parent page is page %zu, index %u",
4232 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4234 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4235 : (mc->mc_ki[mc->mc_top] == 0)) {
4236 DPRINTF("no more keys left, moving to %s sibling",
4237 move_right ? "right" : "left");
4238 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4239 /* undo cursor_pop before returning */
4246 mc->mc_ki[mc->mc_top]++;
4248 mc->mc_ki[mc->mc_top]--;
4249 DPRINTF("just moving to %s index key %u",
4250 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4252 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4254 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4255 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4258 mdb_cursor_push(mc, mp);
4263 /** Move the cursor to the next data item. */
4265 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4271 if (mc->mc_flags & C_EOF) {
4272 return MDB_NOTFOUND;
4275 assert(mc->mc_flags & C_INITIALIZED);
4277 mp = mc->mc_pg[mc->mc_top];
4279 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4280 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4281 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4282 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4283 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4284 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4288 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4289 if (op == MDB_NEXT_DUP)
4290 return MDB_NOTFOUND;
4294 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4296 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4297 DPUTS("=====> move to next sibling page");
4298 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4299 mc->mc_flags |= C_EOF;
4300 mc->mc_flags &= ~C_INITIALIZED;
4301 return MDB_NOTFOUND;
4303 mp = mc->mc_pg[mc->mc_top];
4304 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4306 mc->mc_ki[mc->mc_top]++;
4308 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4309 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4312 key->mv_size = mc->mc_db->md_pad;
4313 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4317 assert(IS_LEAF(mp));
4318 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4320 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4321 mdb_xcursor_init1(mc, leaf);
4324 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4327 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4328 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4329 if (rc != MDB_SUCCESS)
4334 MDB_GET_KEY(leaf, key);
4338 /** Move the cursor to the previous data item. */
4340 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4346 assert(mc->mc_flags & C_INITIALIZED);
4348 mp = mc->mc_pg[mc->mc_top];
4350 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4351 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4352 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4353 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4354 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4355 if (op != MDB_PREV || rc == MDB_SUCCESS)
4358 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4359 if (op == MDB_PREV_DUP)
4360 return MDB_NOTFOUND;
4365 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4367 if (mc->mc_ki[mc->mc_top] == 0) {
4368 DPUTS("=====> move to prev sibling page");
4369 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4370 mc->mc_flags &= ~C_INITIALIZED;
4371 return MDB_NOTFOUND;
4373 mp = mc->mc_pg[mc->mc_top];
4374 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4375 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4377 mc->mc_ki[mc->mc_top]--;
4379 mc->mc_flags &= ~C_EOF;
4381 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4382 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4385 key->mv_size = mc->mc_db->md_pad;
4386 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4390 assert(IS_LEAF(mp));
4391 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4393 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4394 mdb_xcursor_init1(mc, leaf);
4397 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4400 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4401 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4402 if (rc != MDB_SUCCESS)
4407 MDB_GET_KEY(leaf, key);
4411 /** Set the cursor on a specific data item. */
4413 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4414 MDB_cursor_op op, int *exactp)
4418 MDB_node *leaf = NULL;
4423 assert(key->mv_size > 0);
4425 /* See if we're already on the right page */
4426 if (mc->mc_flags & C_INITIALIZED) {
4429 mp = mc->mc_pg[mc->mc_top];
4431 mc->mc_ki[mc->mc_top] = 0;
4432 return MDB_NOTFOUND;
4434 if (mp->mp_flags & P_LEAF2) {
4435 nodekey.mv_size = mc->mc_db->md_pad;
4436 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4438 leaf = NODEPTR(mp, 0);
4439 MDB_GET_KEY(leaf, &nodekey);
4441 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4443 /* Probably happens rarely, but first node on the page
4444 * was the one we wanted.
4446 mc->mc_ki[mc->mc_top] = 0;
4453 unsigned int nkeys = NUMKEYS(mp);
4455 if (mp->mp_flags & P_LEAF2) {
4456 nodekey.mv_data = LEAF2KEY(mp,
4457 nkeys-1, nodekey.mv_size);
4459 leaf = NODEPTR(mp, nkeys-1);
4460 MDB_GET_KEY(leaf, &nodekey);
4462 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4464 /* last node was the one we wanted */
4465 mc->mc_ki[mc->mc_top] = nkeys-1;
4471 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4472 /* This is definitely the right page, skip search_page */
4473 if (mp->mp_flags & P_LEAF2) {
4474 nodekey.mv_data = LEAF2KEY(mp,
4475 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4477 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4478 MDB_GET_KEY(leaf, &nodekey);
4480 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4482 /* current node was the one we wanted */
4492 /* If any parents have right-sibs, search.
4493 * Otherwise, there's nothing further.
4495 for (i=0; i<mc->mc_top; i++)
4497 NUMKEYS(mc->mc_pg[i])-1)
4499 if (i == mc->mc_top) {
4500 /* There are no other pages */
4501 mc->mc_ki[mc->mc_top] = nkeys;
4502 return MDB_NOTFOUND;
4506 /* There are no other pages */
4507 mc->mc_ki[mc->mc_top] = 0;
4508 return MDB_NOTFOUND;
4512 rc = mdb_page_search(mc, key, 0);
4513 if (rc != MDB_SUCCESS)
4516 mp = mc->mc_pg[mc->mc_top];
4517 assert(IS_LEAF(mp));
4520 leaf = mdb_node_search(mc, key, exactp);
4521 if (exactp != NULL && !*exactp) {
4522 /* MDB_SET specified and not an exact match. */
4523 return MDB_NOTFOUND;
4527 DPUTS("===> inexact leaf not found, goto sibling");
4528 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4529 return rc; /* no entries matched */
4530 mp = mc->mc_pg[mc->mc_top];
4531 assert(IS_LEAF(mp));
4532 leaf = NODEPTR(mp, 0);
4536 mc->mc_flags |= C_INITIALIZED;
4537 mc->mc_flags &= ~C_EOF;
4540 key->mv_size = mc->mc_db->md_pad;
4541 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4545 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4546 mdb_xcursor_init1(mc, leaf);
4549 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4550 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4551 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4554 if (op == MDB_GET_BOTH) {
4560 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4561 if (rc != MDB_SUCCESS)
4564 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4566 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4568 rc = mc->mc_dbx->md_dcmp(data, &d2);
4570 if (op == MDB_GET_BOTH || rc > 0)
4571 return MDB_NOTFOUND;
4576 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4577 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4582 /* The key already matches in all other cases */
4583 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4584 MDB_GET_KEY(leaf, key);
4585 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4590 /** Move the cursor to the first item in the database. */
4592 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4597 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4598 rc = mdb_page_search(mc, NULL, 0);
4599 if (rc != MDB_SUCCESS)
4602 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4604 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4605 mc->mc_flags |= C_INITIALIZED;
4606 mc->mc_flags &= ~C_EOF;
4608 mc->mc_ki[mc->mc_top] = 0;
4610 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4611 key->mv_size = mc->mc_db->md_pad;
4612 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4617 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4618 mdb_xcursor_init1(mc, leaf);
4619 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4624 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4625 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4629 MDB_GET_KEY(leaf, key);
4633 /** Move the cursor to the last item in the database. */
4635 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4640 if (!(mc->mc_flags & C_EOF)) {
4642 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4645 lkey.mv_size = MDB_MAXKEYSIZE+1;
4646 lkey.mv_data = NULL;
4647 rc = mdb_page_search(mc, &lkey, 0);
4648 if (rc != MDB_SUCCESS)
4651 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4653 mc->mc_flags |= C_INITIALIZED|C_EOF;
4654 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4656 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4658 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4659 key->mv_size = mc->mc_db->md_pad;
4660 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4665 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4666 mdb_xcursor_init1(mc, leaf);
4667 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4672 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4673 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4678 MDB_GET_KEY(leaf, key);
4683 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4692 case MDB_GET_CURRENT:
4693 if (!(mc->mc_flags & C_INITIALIZED)) {
4696 MDB_page *mp = mc->mc_pg[mc->mc_top];
4698 mc->mc_ki[mc->mc_top] = 0;
4704 key->mv_size = mc->mc_db->md_pad;
4705 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4707 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4708 MDB_GET_KEY(leaf, key);
4710 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4711 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4713 rc = mdb_node_read(mc->mc_txn, leaf, data);
4720 case MDB_GET_BOTH_RANGE:
4721 if (data == NULL || mc->mc_xcursor == NULL) {
4729 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4731 } else if (op == MDB_SET_RANGE)
4732 rc = mdb_cursor_set(mc, key, data, op, NULL);
4734 rc = mdb_cursor_set(mc, key, data, op, &exact);
4736 case MDB_GET_MULTIPLE:
4738 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4739 !(mc->mc_flags & C_INITIALIZED)) {
4744 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4745 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4748 case MDB_NEXT_MULTIPLE:
4750 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4754 if (!(mc->mc_flags & C_INITIALIZED))
4755 rc = mdb_cursor_first(mc, key, data);
4757 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4758 if (rc == MDB_SUCCESS) {
4759 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4762 mx = &mc->mc_xcursor->mx_cursor;
4763 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4765 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4766 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4774 case MDB_NEXT_NODUP:
4775 if (!(mc->mc_flags & C_INITIALIZED))
4776 rc = mdb_cursor_first(mc, key, data);
4778 rc = mdb_cursor_next(mc, key, data, op);
4782 case MDB_PREV_NODUP:
4783 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4784 rc = mdb_cursor_last(mc, key, data);
4785 mc->mc_flags |= C_INITIALIZED;
4786 mc->mc_ki[mc->mc_top]++;
4788 rc = mdb_cursor_prev(mc, key, data, op);
4791 rc = mdb_cursor_first(mc, key, data);
4795 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4796 !(mc->mc_flags & C_INITIALIZED) ||
4797 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4801 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4804 rc = mdb_cursor_last(mc, key, data);
4808 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4809 !(mc->mc_flags & C_INITIALIZED) ||
4810 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4814 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4817 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4825 /** Touch all the pages in the cursor stack.
4826 * Makes sure all the pages are writable, before attempting a write operation.
4827 * @param[in] mc The cursor to operate on.
4830 mdb_cursor_touch(MDB_cursor *mc)
4834 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4837 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4838 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4839 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4842 *mc->mc_dbflag = DB_DIRTY;
4844 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4845 rc = mdb_page_touch(mc);
4849 mc->mc_top = mc->mc_snum-1;
4854 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4857 MDB_node *leaf = NULL;
4858 MDB_val xdata, *rdata, dkey;
4861 int do_sub = 0, insert = 0;
4862 unsigned int mcount = 0;
4866 char dbuf[MDB_MAXKEYSIZE+1];
4867 unsigned int nflags;
4870 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4873 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4876 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4879 #if SIZE_MAX > MAXDATASIZE
4880 if (data->mv_size > MAXDATASIZE)
4884 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4885 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4889 if (flags == MDB_CURRENT) {
4890 if (!(mc->mc_flags & C_INITIALIZED))
4893 } else if (mc->mc_db->md_root == P_INVALID) {
4895 /* new database, write a root leaf page */
4896 DPUTS("allocating new root leaf page");
4897 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4901 mdb_cursor_push(mc, np);
4902 mc->mc_db->md_root = np->mp_pgno;
4903 mc->mc_db->md_depth++;
4904 *mc->mc_dbflag = DB_DIRTY;
4905 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4907 np->mp_flags |= P_LEAF2;
4908 mc->mc_flags |= C_INITIALIZED;
4914 if (flags & MDB_APPEND) {
4916 rc = mdb_cursor_last(mc, &k2, &d2);
4918 rc = mc->mc_dbx->md_cmp(key, &k2);
4921 mc->mc_ki[mc->mc_top]++;
4927 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4929 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4930 DPRINTF("duplicate key [%s]", DKEY(key));
4932 return MDB_KEYEXIST;
4934 if (rc && rc != MDB_NOTFOUND)
4938 /* Cursor is positioned, now make sure all pages are writable */
4939 rc2 = mdb_cursor_touch(mc);
4944 /* The key already exists */
4945 if (rc == MDB_SUCCESS) {
4946 /* there's only a key anyway, so this is a no-op */
4947 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4948 unsigned int ksize = mc->mc_db->md_pad;
4949 if (key->mv_size != ksize)
4951 if (flags == MDB_CURRENT) {
4952 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4953 memcpy(ptr, key->mv_data, ksize);
4958 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4961 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4962 /* Was a single item before, must convert now */
4964 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4965 /* Just overwrite the current item */
4966 if (flags == MDB_CURRENT)
4969 dkey.mv_size = NODEDSZ(leaf);
4970 dkey.mv_data = NODEDATA(leaf);
4971 #if UINT_MAX < SIZE_MAX
4972 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4973 #ifdef MISALIGNED_OK
4974 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4976 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4979 /* if data matches, ignore it */
4980 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4981 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4983 /* create a fake page for the dup items */
4984 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4985 dkey.mv_data = dbuf;
4986 fp = (MDB_page *)&pbuf;
4987 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4988 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4989 fp->mp_lower = PAGEHDRSZ;
4990 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4991 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4992 fp->mp_flags |= P_LEAF2;
4993 fp->mp_pad = data->mv_size;
4994 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4996 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4997 (dkey.mv_size & 1) + (data->mv_size & 1);
4999 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5002 xdata.mv_size = fp->mp_upper;
5007 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5008 /* See if we need to convert from fake page to subDB */
5010 unsigned int offset;
5013 fp = NODEDATA(leaf);
5014 if (flags == MDB_CURRENT) {
5016 fp->mp_flags |= P_DIRTY;
5017 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5018 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5022 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5023 offset = fp->mp_pad;
5024 if (SIZELEFT(fp) >= offset)
5026 offset *= 4; /* space for 4 more */
5028 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5030 offset += offset & 1;
5031 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5032 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
5034 /* yes, convert it */
5036 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5037 dummy.md_pad = fp->mp_pad;
5038 dummy.md_flags = MDB_DUPFIXED;
5039 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5040 dummy.md_flags |= MDB_INTEGERKEY;
5043 dummy.md_branch_pages = 0;
5044 dummy.md_leaf_pages = 1;
5045 dummy.md_overflow_pages = 0;
5046 dummy.md_entries = NUMKEYS(fp);
5048 xdata.mv_size = sizeof(MDB_db);
5049 xdata.mv_data = &dummy;
5050 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5052 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5053 flags |= F_DUPDATA|F_SUBDATA;
5054 dummy.md_root = mp->mp_pgno;
5056 /* no, just grow it */
5058 xdata.mv_size = NODEDSZ(leaf) + offset;
5059 xdata.mv_data = &pbuf;
5060 mp = (MDB_page *)&pbuf;
5061 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5064 mp->mp_flags = fp->mp_flags | P_DIRTY;
5065 mp->mp_pad = fp->mp_pad;
5066 mp->mp_lower = fp->mp_lower;
5067 mp->mp_upper = fp->mp_upper + offset;
5069 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5071 nsize = NODEDSZ(leaf) - fp->mp_upper;
5072 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5073 for (i=0; i<NUMKEYS(fp); i++)
5074 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5076 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5080 /* data is on sub-DB, just store it */
5081 flags |= F_DUPDATA|F_SUBDATA;
5085 /* overflow page overwrites need special handling */
5086 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5089 int ovpages, dpages;
5091 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5092 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5093 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5094 mdb_page_get(mc->mc_txn, pg, &omp);
5095 /* Is the ov page writable and large enough? */
5096 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5097 /* yes, overwrite it. Note in this case we don't
5098 * bother to try shrinking the node if the new data
5099 * is smaller than the overflow threshold.
5101 if (F_ISSET(flags, MDB_RESERVE))
5102 data->mv_data = METADATA(omp);
5104 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5107 /* no, free ovpages */
5109 mc->mc_db->md_overflow_pages -= ovpages;
5110 for (i=0; i<ovpages; i++) {
5111 DPRINTF("freed ov page %zu", pg);
5112 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5116 } else if (NODEDSZ(leaf) == data->mv_size) {
5117 /* same size, just replace it. Note that we could
5118 * also reuse this node if the new data is smaller,
5119 * but instead we opt to shrink the node in that case.
5121 if (F_ISSET(flags, MDB_RESERVE))
5122 data->mv_data = NODEDATA(leaf);
5123 else if (data->mv_size)
5124 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5126 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5129 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5130 mc->mc_db->md_entries--;
5132 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5139 nflags = flags & NODE_ADD_FLAGS;
5140 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5141 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5142 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5143 nflags &= ~MDB_APPEND;
5145 nflags |= MDB_SPLIT_REPLACE;
5146 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5148 /* There is room already in this leaf page. */
5149 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5150 if (rc == 0 && !do_sub && insert) {
5151 /* Adjust other cursors pointing to mp */
5152 MDB_cursor *m2, *m3;
5153 MDB_dbi dbi = mc->mc_dbi;
5154 unsigned i = mc->mc_top;
5155 MDB_page *mp = mc->mc_pg[i];
5157 if (mc->mc_flags & C_SUB)
5160 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5161 if (mc->mc_flags & C_SUB)
5162 m3 = &m2->mc_xcursor->mx_cursor;
5165 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5166 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5173 if (rc != MDB_SUCCESS)
5174 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5176 /* Now store the actual data in the child DB. Note that we're
5177 * storing the user data in the keys field, so there are strict
5178 * size limits on dupdata. The actual data fields of the child
5179 * DB are all zero size.
5186 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5187 if (flags & MDB_CURRENT) {
5188 xflags = MDB_CURRENT;
5190 mdb_xcursor_init1(mc, leaf);
5191 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5193 /* converted, write the original data first */
5195 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5199 /* Adjust other cursors pointing to mp */
5201 unsigned i = mc->mc_top;
5202 MDB_page *mp = mc->mc_pg[i];
5204 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5205 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5206 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5207 mdb_xcursor_init1(m2, leaf);
5212 if (flags & MDB_APPENDDUP)
5213 xflags |= MDB_APPEND;
5214 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5215 if (flags & F_SUBDATA) {
5216 void *db = NODEDATA(leaf);
5217 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5220 /* sub-writes might have failed so check rc again.
5221 * Don't increment count if we just replaced an existing item.
5223 if (!rc && !(flags & MDB_CURRENT))
5224 mc->mc_db->md_entries++;
5225 if (flags & MDB_MULTIPLE) {
5227 if (mcount < data[1].mv_size) {
5228 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5229 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5235 /* If we succeeded and the key didn't exist before, make sure
5236 * the cursor is marked valid.
5239 mc->mc_flags |= C_INITIALIZED;
5244 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5249 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5252 if (!(mc->mc_flags & C_INITIALIZED))
5255 rc = mdb_cursor_touch(mc);
5259 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5261 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5262 if (flags != MDB_NODUPDATA) {
5263 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5264 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5266 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5267 /* If sub-DB still has entries, we're done */
5268 if (mc->mc_xcursor->mx_db.md_entries) {
5269 if (leaf->mn_flags & F_SUBDATA) {
5270 /* update subDB info */
5271 void *db = NODEDATA(leaf);
5272 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5274 /* shrink fake page */
5275 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5277 mc->mc_db->md_entries--;
5280 /* otherwise fall thru and delete the sub-DB */
5283 if (leaf->mn_flags & F_SUBDATA) {
5284 /* add all the child DB's pages to the free list */
5285 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5286 if (rc == MDB_SUCCESS) {
5287 mc->mc_db->md_entries -=
5288 mc->mc_xcursor->mx_db.md_entries;
5293 return mdb_cursor_del0(mc, leaf);
5296 /** Allocate and initialize new pages for a database.
5297 * @param[in] mc a cursor on the database being added to.
5298 * @param[in] flags flags defining what type of page is being allocated.
5299 * @param[in] num the number of pages to allocate. This is usually 1,
5300 * unless allocating overflow pages for a large record.
5301 * @param[out] mp Address of a page, or NULL on failure.
5302 * @return 0 on success, non-zero on failure.
5305 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5310 if ((rc = mdb_page_alloc(mc, num, &np)))
5312 DPRINTF("allocated new mpage %zu, page size %u",
5313 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5314 np->mp_flags = flags | P_DIRTY;
5315 np->mp_lower = PAGEHDRSZ;
5316 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5319 mc->mc_db->md_branch_pages++;
5320 else if (IS_LEAF(np))
5321 mc->mc_db->md_leaf_pages++;
5322 else if (IS_OVERFLOW(np)) {
5323 mc->mc_db->md_overflow_pages += num;
5331 /** Calculate the size of a leaf node.
5332 * The size depends on the environment's page size; if a data item
5333 * is too large it will be put onto an overflow page and the node
5334 * size will only include the key and not the data. Sizes are always
5335 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5336 * of the #MDB_node headers.
5337 * @param[in] env The environment handle.
5338 * @param[in] key The key for the node.
5339 * @param[in] data The data for the node.
5340 * @return The number of bytes needed to store the node.
5343 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5347 sz = LEAFSIZE(key, data);
5348 if (sz >= env->me_psize / MDB_MINKEYS) {
5349 /* put on overflow page */
5350 sz -= data->mv_size - sizeof(pgno_t);
5354 return sz + sizeof(indx_t);
5357 /** Calculate the size of a branch node.
5358 * The size should depend on the environment's page size but since
5359 * we currently don't support spilling large keys onto overflow
5360 * pages, it's simply the size of the #MDB_node header plus the
5361 * size of the key. Sizes are always rounded up to an even number
5362 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5363 * @param[in] env The environment handle.
5364 * @param[in] key The key for the node.
5365 * @return The number of bytes needed to store the node.
5368 mdb_branch_size(MDB_env *env, MDB_val *key)
5373 if (sz >= env->me_psize / MDB_MINKEYS) {
5374 /* put on overflow page */
5375 /* not implemented */
5376 /* sz -= key->size - sizeof(pgno_t); */
5379 return sz + sizeof(indx_t);
5382 /** Add a node to the page pointed to by the cursor.
5383 * @param[in] mc The cursor for this operation.
5384 * @param[in] indx The index on the page where the new node should be added.
5385 * @param[in] key The key for the new node.
5386 * @param[in] data The data for the new node, if any.
5387 * @param[in] pgno The page number, if adding a branch node.
5388 * @param[in] flags Flags for the node.
5389 * @return 0 on success, non-zero on failure. Possible errors are:
5391 * <li>ENOMEM - failed to allocate overflow pages for the node.
5392 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5393 * should never happen since all callers already calculate the
5394 * page's free space before calling this function.
5398 mdb_node_add(MDB_cursor *mc, indx_t indx,
5399 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5402 size_t node_size = NODESIZE;
5405 MDB_page *mp = mc->mc_pg[mc->mc_top];
5406 MDB_page *ofp = NULL; /* overflow page */
5409 assert(mp->mp_upper >= mp->mp_lower);
5411 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5412 IS_LEAF(mp) ? "leaf" : "branch",
5413 IS_SUBP(mp) ? "sub-" : "",
5414 mp->mp_pgno, indx, data ? data->mv_size : 0,
5415 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5418 /* Move higher keys up one slot. */
5419 int ksize = mc->mc_db->md_pad, dif;
5420 char *ptr = LEAF2KEY(mp, indx, ksize);
5421 dif = NUMKEYS(mp) - indx;
5423 memmove(ptr+ksize, ptr, dif*ksize);
5424 /* insert new key */
5425 memcpy(ptr, key->mv_data, ksize);
5427 /* Just using these for counting */
5428 mp->mp_lower += sizeof(indx_t);
5429 mp->mp_upper -= ksize - sizeof(indx_t);
5434 node_size += key->mv_size;
5438 if (F_ISSET(flags, F_BIGDATA)) {
5439 /* Data already on overflow page. */
5440 node_size += sizeof(pgno_t);
5441 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5442 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5444 /* Put data on overflow page. */
5445 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5446 data->mv_size, node_size+data->mv_size);
5447 node_size += sizeof(pgno_t);
5448 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5450 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5453 node_size += data->mv_size;
5456 node_size += node_size & 1;
5458 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5459 DPRINTF("not enough room in page %zu, got %u ptrs",
5460 mp->mp_pgno, NUMKEYS(mp));
5461 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5462 mp->mp_upper - mp->mp_lower);
5463 DPRINTF("node size = %zu", node_size);
5464 return MDB_PAGE_FULL;
5467 /* Move higher pointers up one slot. */
5468 for (i = NUMKEYS(mp); i > indx; i--)
5469 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5471 /* Adjust free space offsets. */
5472 ofs = mp->mp_upper - node_size;
5473 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5474 mp->mp_ptrs[indx] = ofs;
5476 mp->mp_lower += sizeof(indx_t);
5478 /* Write the node data. */
5479 node = NODEPTR(mp, indx);
5480 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5481 node->mn_flags = flags;
5483 SETDSZ(node,data->mv_size);
5488 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5493 if (F_ISSET(flags, F_BIGDATA))
5494 memcpy(node->mn_data + key->mv_size, data->mv_data,
5496 else if (F_ISSET(flags, MDB_RESERVE))
5497 data->mv_data = node->mn_data + key->mv_size;
5499 memcpy(node->mn_data + key->mv_size, data->mv_data,
5502 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5504 if (F_ISSET(flags, MDB_RESERVE))
5505 data->mv_data = METADATA(ofp);
5507 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5514 /** Delete the specified node from a page.
5515 * @param[in] mp The page to operate on.
5516 * @param[in] indx The index of the node to delete.
5517 * @param[in] ksize The size of a node. Only used if the page is
5518 * part of a #MDB_DUPFIXED database.
5521 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5524 indx_t i, j, numkeys, ptr;
5531 COPY_PGNO(pgno, mp->mp_pgno);
5532 DPRINTF("delete node %u on %s page %zu", indx,
5533 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5536 assert(indx < NUMKEYS(mp));
5539 int x = NUMKEYS(mp) - 1 - indx;
5540 base = LEAF2KEY(mp, indx, ksize);
5542 memmove(base, base + ksize, x * ksize);
5543 mp->mp_lower -= sizeof(indx_t);
5544 mp->mp_upper += ksize - sizeof(indx_t);
5548 node = NODEPTR(mp, indx);
5549 sz = NODESIZE + node->mn_ksize;
5551 if (F_ISSET(node->mn_flags, F_BIGDATA))
5552 sz += sizeof(pgno_t);
5554 sz += NODEDSZ(node);
5558 ptr = mp->mp_ptrs[indx];
5559 numkeys = NUMKEYS(mp);
5560 for (i = j = 0; i < numkeys; i++) {
5562 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5563 if (mp->mp_ptrs[i] < ptr)
5564 mp->mp_ptrs[j] += sz;
5569 base = (char *)mp + mp->mp_upper;
5570 memmove(base + sz, base, ptr - mp->mp_upper);
5572 mp->mp_lower -= sizeof(indx_t);
5576 /** Compact the main page after deleting a node on a subpage.
5577 * @param[in] mp The main page to operate on.
5578 * @param[in] indx The index of the subpage on the main page.
5581 mdb_node_shrink(MDB_page *mp, indx_t indx)
5588 indx_t i, numkeys, ptr;
5590 node = NODEPTR(mp, indx);
5591 sp = (MDB_page *)NODEDATA(node);
5592 osize = NODEDSZ(node);
5594 delta = sp->mp_upper - sp->mp_lower;
5595 SETDSZ(node, osize - delta);
5596 xp = (MDB_page *)((char *)sp + delta);
5598 /* shift subpage upward */
5600 nsize = NUMKEYS(sp) * sp->mp_pad;
5601 memmove(METADATA(xp), METADATA(sp), nsize);
5604 nsize = osize - sp->mp_upper;
5605 numkeys = NUMKEYS(sp);
5606 for (i=numkeys-1; i>=0; i--)
5607 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5609 xp->mp_upper = sp->mp_lower;
5610 xp->mp_lower = sp->mp_lower;
5611 xp->mp_flags = sp->mp_flags;
5612 xp->mp_pad = sp->mp_pad;
5613 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5615 /* shift lower nodes upward */
5616 ptr = mp->mp_ptrs[indx];
5617 numkeys = NUMKEYS(mp);
5618 for (i = 0; i < numkeys; i++) {
5619 if (mp->mp_ptrs[i] <= ptr)
5620 mp->mp_ptrs[i] += delta;
5623 base = (char *)mp + mp->mp_upper;
5624 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5625 mp->mp_upper += delta;
5628 /** Initial setup of a sorted-dups cursor.
5629 * Sorted duplicates are implemented as a sub-database for the given key.
5630 * The duplicate data items are actually keys of the sub-database.
5631 * Operations on the duplicate data items are performed using a sub-cursor
5632 * initialized when the sub-database is first accessed. This function does
5633 * the preliminary setup of the sub-cursor, filling in the fields that
5634 * depend only on the parent DB.
5635 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5638 mdb_xcursor_init0(MDB_cursor *mc)
5640 MDB_xcursor *mx = mc->mc_xcursor;
5642 mx->mx_cursor.mc_xcursor = NULL;
5643 mx->mx_cursor.mc_txn = mc->mc_txn;
5644 mx->mx_cursor.mc_db = &mx->mx_db;
5645 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5646 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5647 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5648 mx->mx_cursor.mc_snum = 0;
5649 mx->mx_cursor.mc_top = 0;
5650 mx->mx_cursor.mc_flags = C_SUB;
5651 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5652 mx->mx_dbx.md_dcmp = NULL;
5653 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5656 /** Final setup of a sorted-dups cursor.
5657 * Sets up the fields that depend on the data from the main cursor.
5658 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5659 * @param[in] node The data containing the #MDB_db record for the
5660 * sorted-dup database.
5663 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5665 MDB_xcursor *mx = mc->mc_xcursor;
5667 if (node->mn_flags & F_SUBDATA) {
5668 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5669 mx->mx_cursor.mc_pg[0] = 0;
5670 mx->mx_cursor.mc_snum = 0;
5671 mx->mx_cursor.mc_flags = C_SUB;
5673 MDB_page *fp = NODEDATA(node);
5674 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5675 mx->mx_db.md_flags = 0;
5676 mx->mx_db.md_depth = 1;
5677 mx->mx_db.md_branch_pages = 0;
5678 mx->mx_db.md_leaf_pages = 1;
5679 mx->mx_db.md_overflow_pages = 0;
5680 mx->mx_db.md_entries = NUMKEYS(fp);
5681 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5682 mx->mx_cursor.mc_snum = 1;
5683 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5684 mx->mx_cursor.mc_top = 0;
5685 mx->mx_cursor.mc_pg[0] = fp;
5686 mx->mx_cursor.mc_ki[0] = 0;
5687 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5688 mx->mx_db.md_flags = MDB_DUPFIXED;
5689 mx->mx_db.md_pad = fp->mp_pad;
5690 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5691 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5694 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5696 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5698 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5699 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5700 #if UINT_MAX < SIZE_MAX
5701 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5702 #ifdef MISALIGNED_OK
5703 mx->mx_dbx.md_cmp = mdb_cmp_long;
5705 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5710 /** Initialize a cursor for a given transaction and database. */
5712 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5717 mc->mc_db = &txn->mt_dbs[dbi];
5718 mc->mc_dbx = &txn->mt_dbxs[dbi];
5719 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5724 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5726 mc->mc_xcursor = mx;
5727 mdb_xcursor_init0(mc);
5729 mc->mc_xcursor = NULL;
5731 if (*mc->mc_dbflag & DB_STALE) {
5732 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5737 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5740 MDB_xcursor *mx = NULL;
5741 size_t size = sizeof(MDB_cursor);
5743 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5746 /* Allow read access to the freelist */
5747 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5750 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5751 size += sizeof(MDB_xcursor);
5753 if ((mc = malloc(size)) != NULL) {
5754 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5755 mx = (MDB_xcursor *)(mc + 1);
5757 mdb_cursor_init(mc, txn, dbi, mx);
5758 if (txn->mt_cursors) {
5759 mc->mc_next = txn->mt_cursors[dbi];
5760 txn->mt_cursors[dbi] = mc;
5762 mc->mc_flags |= C_ALLOCD;
5773 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5775 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5778 if (txn->mt_cursors)
5781 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5785 /* Return the count of duplicate data items for the current key */
5787 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5791 if (mc == NULL || countp == NULL)
5794 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5797 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5798 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5801 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5804 *countp = mc->mc_xcursor->mx_db.md_entries;
5810 mdb_cursor_close(MDB_cursor *mc)
5813 /* remove from txn, if tracked */
5814 if (mc->mc_txn->mt_cursors) {
5815 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5816 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5818 *prev = mc->mc_next;
5820 if (mc->mc_flags & C_ALLOCD)
5826 mdb_cursor_txn(MDB_cursor *mc)
5828 if (!mc) return NULL;
5833 mdb_cursor_dbi(MDB_cursor *mc)
5839 /** Replace the key for a node with a new key.
5840 * @param[in] mp The page containing the node to operate on.
5841 * @param[in] indx The index of the node to operate on.
5842 * @param[in] key The new key to use.
5843 * @return 0 on success, non-zero on failure.
5846 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5853 indx_t ptr, i, numkeys, indx;
5856 indx = mc->mc_ki[mc->mc_top];
5857 mp = mc->mc_pg[mc->mc_top];
5858 node = NODEPTR(mp, indx);
5859 ptr = mp->mp_ptrs[indx];
5863 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5864 k2.mv_data = NODEKEY(node);
5865 k2.mv_size = node->mn_ksize;
5866 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5868 mdb_dkey(&k2, kbuf2),
5874 delta0 = delta = key->mv_size - node->mn_ksize;
5876 /* Must be 2-byte aligned. If new key is
5877 * shorter by 1, the shift will be skipped.
5879 delta += (delta & 1);
5881 if (delta > 0 && SIZELEFT(mp) < delta) {
5883 /* not enough space left, do a delete and split */
5884 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5885 pgno = NODEPGNO(node);
5886 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5887 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5890 numkeys = NUMKEYS(mp);
5891 for (i = 0; i < numkeys; i++) {
5892 if (mp->mp_ptrs[i] <= ptr)
5893 mp->mp_ptrs[i] -= delta;
5896 base = (char *)mp + mp->mp_upper;
5897 len = ptr - mp->mp_upper + NODESIZE;
5898 memmove(base - delta, base, len);
5899 mp->mp_upper -= delta;
5901 node = NODEPTR(mp, indx);
5904 /* But even if no shift was needed, update ksize */
5906 node->mn_ksize = key->mv_size;
5909 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5915 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5917 /** Move a node from csrc to cdst.
5920 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5927 unsigned short flags;
5931 /* Mark src and dst as dirty. */
5932 if ((rc = mdb_page_touch(csrc)) ||
5933 (rc = mdb_page_touch(cdst)))
5936 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5937 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5938 key.mv_size = csrc->mc_db->md_pad;
5939 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5941 data.mv_data = NULL;
5945 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5946 assert(!((long)srcnode&1));
5947 srcpg = NODEPGNO(srcnode);
5948 flags = srcnode->mn_flags;
5949 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5950 unsigned int snum = csrc->mc_snum;
5952 /* must find the lowest key below src */
5953 mdb_page_search_root(csrc, NULL, 0);
5954 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5955 key.mv_size = csrc->mc_db->md_pad;
5956 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5958 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5959 key.mv_size = NODEKSZ(s2);
5960 key.mv_data = NODEKEY(s2);
5962 csrc->mc_snum = snum--;
5963 csrc->mc_top = snum;
5965 key.mv_size = NODEKSZ(srcnode);
5966 key.mv_data = NODEKEY(srcnode);
5968 data.mv_size = NODEDSZ(srcnode);
5969 data.mv_data = NODEDATA(srcnode);
5971 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5972 unsigned int snum = cdst->mc_snum;
5975 /* must find the lowest key below dst */
5976 mdb_page_search_root(cdst, NULL, 0);
5977 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5978 bkey.mv_size = cdst->mc_db->md_pad;
5979 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5981 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5982 bkey.mv_size = NODEKSZ(s2);
5983 bkey.mv_data = NODEKEY(s2);
5985 cdst->mc_snum = snum--;
5986 cdst->mc_top = snum;
5987 mdb_cursor_copy(cdst, &mn);
5989 rc = mdb_update_key(&mn, &bkey);
5994 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5995 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5996 csrc->mc_ki[csrc->mc_top],
5998 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5999 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6001 /* Add the node to the destination page.
6003 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6004 if (rc != MDB_SUCCESS)
6007 /* Delete the node from the source page.
6009 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6012 /* Adjust other cursors pointing to mp */
6013 MDB_cursor *m2, *m3;
6014 MDB_dbi dbi = csrc->mc_dbi;
6015 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6017 if (csrc->mc_flags & C_SUB)
6020 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6021 if (m2 == csrc) continue;
6022 if (csrc->mc_flags & C_SUB)
6023 m3 = &m2->mc_xcursor->mx_cursor;
6026 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6027 csrc->mc_ki[csrc->mc_top]) {
6028 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6029 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6034 /* Update the parent separators.
6036 if (csrc->mc_ki[csrc->mc_top] == 0) {
6037 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6038 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6039 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6041 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6042 key.mv_size = NODEKSZ(srcnode);
6043 key.mv_data = NODEKEY(srcnode);
6045 DPRINTF("update separator for source page %zu to [%s]",
6046 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6047 mdb_cursor_copy(csrc, &mn);
6050 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6053 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6055 indx_t ix = csrc->mc_ki[csrc->mc_top];
6056 nullkey.mv_size = 0;
6057 csrc->mc_ki[csrc->mc_top] = 0;
6058 rc = mdb_update_key(csrc, &nullkey);
6059 csrc->mc_ki[csrc->mc_top] = ix;
6060 assert(rc == MDB_SUCCESS);
6064 if (cdst->mc_ki[cdst->mc_top] == 0) {
6065 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6066 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6067 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6069 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6070 key.mv_size = NODEKSZ(srcnode);
6071 key.mv_data = NODEKEY(srcnode);
6073 DPRINTF("update separator for destination page %zu to [%s]",
6074 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6075 mdb_cursor_copy(cdst, &mn);
6078 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6081 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6083 indx_t ix = cdst->mc_ki[cdst->mc_top];
6084 nullkey.mv_size = 0;
6085 cdst->mc_ki[cdst->mc_top] = 0;
6086 rc = mdb_update_key(cdst, &nullkey);
6087 cdst->mc_ki[cdst->mc_top] = ix;
6088 assert(rc == MDB_SUCCESS);
6095 /** Merge one page into another.
6096 * The nodes from the page pointed to by \b csrc will
6097 * be copied to the page pointed to by \b cdst and then
6098 * the \b csrc page will be freed.
6099 * @param[in] csrc Cursor pointing to the source page.
6100 * @param[in] cdst Cursor pointing to the destination page.
6103 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6111 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6112 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6114 assert(csrc->mc_snum > 1); /* can't merge root page */
6115 assert(cdst->mc_snum > 1);
6117 /* Mark dst as dirty. */
6118 if ((rc = mdb_page_touch(cdst)))
6121 /* Move all nodes from src to dst.
6123 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6124 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6125 key.mv_size = csrc->mc_db->md_pad;
6126 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6127 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6128 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6129 if (rc != MDB_SUCCESS)
6131 key.mv_data = (char *)key.mv_data + key.mv_size;
6134 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6135 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6136 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6137 unsigned int snum = csrc->mc_snum;
6139 /* must find the lowest key below src */
6140 mdb_page_search_root(csrc, NULL, 0);
6141 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6142 key.mv_size = csrc->mc_db->md_pad;
6143 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6145 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6146 key.mv_size = NODEKSZ(s2);
6147 key.mv_data = NODEKEY(s2);
6149 csrc->mc_snum = snum--;
6150 csrc->mc_top = snum;
6152 key.mv_size = srcnode->mn_ksize;
6153 key.mv_data = NODEKEY(srcnode);
6156 data.mv_size = NODEDSZ(srcnode);
6157 data.mv_data = NODEDATA(srcnode);
6158 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6159 if (rc != MDB_SUCCESS)
6164 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6165 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);
6167 /* Unlink the src page from parent and add to free list.
6169 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6170 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6173 rc = mdb_update_key(csrc, &key);
6179 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6180 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6181 csrc->mc_db->md_leaf_pages--;
6183 csrc->mc_db->md_branch_pages--;
6185 /* Adjust other cursors pointing to mp */
6186 MDB_cursor *m2, *m3;
6187 MDB_dbi dbi = csrc->mc_dbi;
6188 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6190 if (csrc->mc_flags & C_SUB)
6193 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6194 if (csrc->mc_flags & C_SUB)
6195 m3 = &m2->mc_xcursor->mx_cursor;
6198 if (m3 == csrc) continue;
6199 if (m3->mc_snum < csrc->mc_snum) continue;
6200 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6201 m3->mc_pg[csrc->mc_top] = mp;
6202 m3->mc_ki[csrc->mc_top] += nkeys;
6206 mdb_cursor_pop(csrc);
6208 return mdb_rebalance(csrc);
6211 /** Copy the contents of a cursor.
6212 * @param[in] csrc The cursor to copy from.
6213 * @param[out] cdst The cursor to copy to.
6216 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6220 cdst->mc_txn = csrc->mc_txn;
6221 cdst->mc_dbi = csrc->mc_dbi;
6222 cdst->mc_db = csrc->mc_db;
6223 cdst->mc_dbx = csrc->mc_dbx;
6224 cdst->mc_snum = csrc->mc_snum;
6225 cdst->mc_top = csrc->mc_top;
6226 cdst->mc_flags = csrc->mc_flags;
6228 for (i=0; i<csrc->mc_snum; i++) {
6229 cdst->mc_pg[i] = csrc->mc_pg[i];
6230 cdst->mc_ki[i] = csrc->mc_ki[i];
6234 /** Rebalance the tree after a delete operation.
6235 * @param[in] mc Cursor pointing to the page where rebalancing
6237 * @return 0 on success, non-zero on failure.
6240 mdb_rebalance(MDB_cursor *mc)
6250 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6251 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6252 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6253 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6257 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6260 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6261 DPRINTF("no need to rebalance page %zu, above fill threshold",
6267 if (mc->mc_snum < 2) {
6268 MDB_page *mp = mc->mc_pg[0];
6269 if (NUMKEYS(mp) == 0) {
6270 DPUTS("tree is completely empty");
6271 mc->mc_db->md_root = P_INVALID;
6272 mc->mc_db->md_depth = 0;
6273 mc->mc_db->md_leaf_pages = 0;
6274 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6278 /* Adjust other cursors pointing to mp */
6279 MDB_cursor *m2, *m3;
6280 MDB_dbi dbi = mc->mc_dbi;
6282 if (mc->mc_flags & C_SUB)
6285 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6286 if (m2 == mc) continue;
6287 if (mc->mc_flags & C_SUB)
6288 m3 = &m2->mc_xcursor->mx_cursor;
6291 if (m3->mc_snum < mc->mc_snum) continue;
6292 if (m3->mc_pg[0] == mp) {
6298 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6299 DPUTS("collapsing root page!");
6300 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6301 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6302 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6305 mc->mc_db->md_depth--;
6306 mc->mc_db->md_branch_pages--;
6308 /* Adjust other cursors pointing to mp */
6309 MDB_cursor *m2, *m3;
6310 MDB_dbi dbi = mc->mc_dbi;
6312 if (mc->mc_flags & C_SUB)
6315 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6316 if (m2 == mc) continue;
6317 if (mc->mc_flags & C_SUB)
6318 m3 = &m2->mc_xcursor->mx_cursor;
6321 if (m3->mc_snum < mc->mc_snum) continue;
6322 if (m3->mc_pg[0] == mp) {
6323 m3->mc_pg[0] = mc->mc_pg[0];
6328 DPUTS("root page doesn't need rebalancing");
6332 /* The parent (branch page) must have at least 2 pointers,
6333 * otherwise the tree is invalid.
6335 ptop = mc->mc_top-1;
6336 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6338 /* Leaf page fill factor is below the threshold.
6339 * Try to move keys from left or right neighbor, or
6340 * merge with a neighbor page.
6345 mdb_cursor_copy(mc, &mn);
6346 mn.mc_xcursor = NULL;
6348 if (mc->mc_ki[ptop] == 0) {
6349 /* We're the leftmost leaf in our parent.
6351 DPUTS("reading right neighbor");
6353 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6354 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6356 mn.mc_ki[mn.mc_top] = 0;
6357 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6359 /* There is at least one neighbor to the left.
6361 DPUTS("reading left neighbor");
6363 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6364 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6366 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6367 mc->mc_ki[mc->mc_top] = 0;
6370 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6371 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);
6373 /* If the neighbor page is above threshold and has at least two
6374 * keys, move one key from it.
6376 * Otherwise we should try to merge them.
6378 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6379 return mdb_node_move(&mn, mc);
6381 if (mc->mc_ki[ptop] == 0)
6382 rc = mdb_page_merge(&mn, mc);
6384 rc = mdb_page_merge(mc, &mn);
6385 mc->mc_flags &= ~C_INITIALIZED;
6390 /** Complete a delete operation started by #mdb_cursor_del(). */
6392 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6396 /* add overflow pages to free list */
6397 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6401 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6402 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6403 mc->mc_db->md_overflow_pages -= ovpages;
6404 for (i=0; i<ovpages; i++) {
6405 DPRINTF("freed ov page %zu", pg);
6406 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6410 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6411 mc->mc_db->md_entries--;
6412 rc = mdb_rebalance(mc);
6413 if (rc != MDB_SUCCESS)
6414 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6420 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6421 MDB_val *key, MDB_val *data)
6426 MDB_val rdata, *xdata;
6430 assert(key != NULL);
6432 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6434 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6437 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6441 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6445 mdb_cursor_init(&mc, txn, dbi, &mx);
6456 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6458 /* let mdb_page_split know about this cursor if needed:
6459 * delete will trigger a rebalance; if it needs to move
6460 * a node from one page to another, it will have to
6461 * update the parent's separator key(s). If the new sepkey
6462 * is larger than the current one, the parent page may
6463 * run out of space, triggering a split. We need this
6464 * cursor to be consistent until the end of the rebalance.
6466 mc.mc_next = txn->mt_cursors[dbi];
6467 txn->mt_cursors[dbi] = &mc;
6468 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6469 txn->mt_cursors[dbi] = mc.mc_next;
6474 /** Split a page and insert a new node.
6475 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6476 * The cursor will be updated to point to the actual page and index where
6477 * the node got inserted after the split.
6478 * @param[in] newkey The key for the newly inserted node.
6479 * @param[in] newdata The data for the newly inserted node.
6480 * @param[in] newpgno The page number, if the new node is a branch node.
6481 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6482 * @return 0 on success, non-zero on failure.
6485 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6486 unsigned int nflags)
6489 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6492 unsigned int i, j, split_indx, nkeys, pmax;
6494 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6496 MDB_page *mp, *rp, *pp;
6501 mp = mc->mc_pg[mc->mc_top];
6502 newindx = mc->mc_ki[mc->mc_top];
6504 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6505 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6506 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6508 /* Create a right sibling. */
6509 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6511 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6513 if (mc->mc_snum < 2) {
6514 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6516 /* shift current top to make room for new parent */
6517 mc->mc_pg[1] = mc->mc_pg[0];
6518 mc->mc_ki[1] = mc->mc_ki[0];
6521 mc->mc_db->md_root = pp->mp_pgno;
6522 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6523 mc->mc_db->md_depth++;
6526 /* Add left (implicit) pointer. */
6527 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6528 /* undo the pre-push */
6529 mc->mc_pg[0] = mc->mc_pg[1];
6530 mc->mc_ki[0] = mc->mc_ki[1];
6531 mc->mc_db->md_root = mp->mp_pgno;
6532 mc->mc_db->md_depth--;
6539 ptop = mc->mc_top-1;
6540 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6543 mc->mc_flags |= C_SPLITTING;
6544 mdb_cursor_copy(mc, &mn);
6545 mn.mc_pg[mn.mc_top] = rp;
6546 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6548 if (nflags & MDB_APPEND) {
6549 mn.mc_ki[mn.mc_top] = 0;
6551 split_indx = newindx;
6556 nkeys = NUMKEYS(mp);
6557 split_indx = nkeys / 2;
6558 if (newindx < split_indx)
6564 unsigned int lsize, rsize, ksize;
6565 /* Move half of the keys to the right sibling */
6567 x = mc->mc_ki[mc->mc_top] - split_indx;
6568 ksize = mc->mc_db->md_pad;
6569 split = LEAF2KEY(mp, split_indx, ksize);
6570 rsize = (nkeys - split_indx) * ksize;
6571 lsize = (nkeys - split_indx) * sizeof(indx_t);
6572 mp->mp_lower -= lsize;
6573 rp->mp_lower += lsize;
6574 mp->mp_upper += rsize - lsize;
6575 rp->mp_upper -= rsize - lsize;
6576 sepkey.mv_size = ksize;
6577 if (newindx == split_indx) {
6578 sepkey.mv_data = newkey->mv_data;
6580 sepkey.mv_data = split;
6583 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6584 memcpy(rp->mp_ptrs, split, rsize);
6585 sepkey.mv_data = rp->mp_ptrs;
6586 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6587 memcpy(ins, newkey->mv_data, ksize);
6588 mp->mp_lower += sizeof(indx_t);
6589 mp->mp_upper -= ksize - sizeof(indx_t);
6592 memcpy(rp->mp_ptrs, split, x * ksize);
6593 ins = LEAF2KEY(rp, x, ksize);
6594 memcpy(ins, newkey->mv_data, ksize);
6595 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6596 rp->mp_lower += sizeof(indx_t);
6597 rp->mp_upper -= ksize - sizeof(indx_t);
6598 mc->mc_ki[mc->mc_top] = x;
6599 mc->mc_pg[mc->mc_top] = rp;
6604 /* For leaf pages, check the split point based on what
6605 * fits where, since otherwise mdb_node_add can fail.
6607 * This check is only needed when the data items are
6608 * relatively large, such that being off by one will
6609 * make the difference between success or failure.
6611 * It's also relevant if a page happens to be laid out
6612 * such that one half of its nodes are all "small" and
6613 * the other half of its nodes are "large." If the new
6614 * item is also "large" and falls on the half with
6615 * "large" nodes, it also may not fit.
6618 unsigned int psize, nsize;
6619 /* Maximum free space in an empty page */
6620 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6621 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6622 if ((nkeys < 20) || (nsize > pmax/16)) {
6623 if (newindx <= split_indx) {
6626 for (i=0; i<split_indx; i++) {
6627 node = NODEPTR(mp, i);
6628 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6629 if (F_ISSET(node->mn_flags, F_BIGDATA))
6630 psize += sizeof(pgno_t);
6632 psize += NODEDSZ(node);
6636 split_indx = newindx;
6647 for (i=nkeys-1; i>=split_indx; i--) {
6648 node = NODEPTR(mp, i);
6649 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6650 if (F_ISSET(node->mn_flags, F_BIGDATA))
6651 psize += sizeof(pgno_t);
6653 psize += NODEDSZ(node);
6657 split_indx = newindx;
6668 /* First find the separating key between the split pages.
6669 * The case where newindx == split_indx is ambiguous; the
6670 * new item could go to the new page or stay on the original
6671 * page. If newpos == 1 it goes to the new page.
6673 if (newindx == split_indx && newpos) {
6674 sepkey.mv_size = newkey->mv_size;
6675 sepkey.mv_data = newkey->mv_data;
6677 node = NODEPTR(mp, split_indx);
6678 sepkey.mv_size = node->mn_ksize;
6679 sepkey.mv_data = NODEKEY(node);
6683 DPRINTF("separator is [%s]", DKEY(&sepkey));
6685 /* Copy separator key to the parent.
6687 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6691 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6694 if (mn.mc_snum == mc->mc_snum) {
6695 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6696 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6697 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6698 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6703 /* Right page might now have changed parent.
6704 * Check if left page also changed parent.
6706 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6707 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6708 for (i=0; i<ptop; i++) {
6709 mc->mc_pg[i] = mn.mc_pg[i];
6710 mc->mc_ki[i] = mn.mc_ki[i];
6712 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6713 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6717 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6720 mc->mc_flags ^= C_SPLITTING;
6721 if (rc != MDB_SUCCESS) {
6724 if (nflags & MDB_APPEND) {
6725 mc->mc_pg[mc->mc_top] = rp;
6726 mc->mc_ki[mc->mc_top] = 0;
6727 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6730 for (i=0; i<mc->mc_top; i++)
6731 mc->mc_ki[i] = mn.mc_ki[i];
6738 /* Move half of the keys to the right sibling. */
6740 /* grab a page to hold a temporary copy */
6741 copy = mdb_page_malloc(mc);
6745 copy->mp_pgno = mp->mp_pgno;
6746 copy->mp_flags = mp->mp_flags;
6747 copy->mp_lower = PAGEHDRSZ;
6748 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6749 mc->mc_pg[mc->mc_top] = copy;
6750 for (i = j = 0; i <= nkeys; j++) {
6751 if (i == split_indx) {
6752 /* Insert in right sibling. */
6753 /* Reset insert index for right sibling. */
6754 if (i != newindx || (newpos ^ ins_new)) {
6756 mc->mc_pg[mc->mc_top] = rp;
6760 if (i == newindx && !ins_new) {
6761 /* Insert the original entry that caused the split. */
6762 rkey.mv_data = newkey->mv_data;
6763 rkey.mv_size = newkey->mv_size;
6772 /* Update index for the new key. */
6773 mc->mc_ki[mc->mc_top] = j;
6774 } else if (i == nkeys) {
6777 node = NODEPTR(mp, i);
6778 rkey.mv_data = NODEKEY(node);
6779 rkey.mv_size = node->mn_ksize;
6781 xdata.mv_data = NODEDATA(node);
6782 xdata.mv_size = NODEDSZ(node);
6785 pgno = NODEPGNO(node);
6786 flags = node->mn_flags;
6791 if (!IS_LEAF(mp) && j == 0) {
6792 /* First branch index doesn't need key data. */
6796 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6800 nkeys = NUMKEYS(copy);
6801 for (i=0; i<nkeys; i++)
6802 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6803 mp->mp_lower = copy->mp_lower;
6804 mp->mp_upper = copy->mp_upper;
6805 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6806 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6808 /* reset back to original page */
6809 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6810 mc->mc_pg[mc->mc_top] = mp;
6811 if (nflags & MDB_RESERVE) {
6812 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6813 if (!(node->mn_flags & F_BIGDATA))
6814 newdata->mv_data = NODEDATA(node);
6820 /* return tmp page to freelist */
6821 mdb_page_free(mc->mc_txn->mt_env, copy);
6824 /* Adjust other cursors pointing to mp */
6825 MDB_cursor *m2, *m3;
6826 MDB_dbi dbi = mc->mc_dbi;
6827 int fixup = NUMKEYS(mp);
6829 if (mc->mc_flags & C_SUB)
6832 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6833 if (m2 == mc) continue;
6834 if (mc->mc_flags & C_SUB)
6835 m3 = &m2->mc_xcursor->mx_cursor;
6838 if (!(m3->mc_flags & C_INITIALIZED))
6840 if (m3->mc_flags & C_SPLITTING)
6845 for (k=m3->mc_top; k>=0; k--) {
6846 m3->mc_ki[k+1] = m3->mc_ki[k];
6847 m3->mc_pg[k+1] = m3->mc_pg[k];
6849 if (m3->mc_ki[0] >= split_indx) {
6854 m3->mc_pg[0] = mc->mc_pg[0];
6858 if (m3->mc_pg[mc->mc_top] == mp) {
6859 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6860 m3->mc_ki[mc->mc_top]++;
6861 if (m3->mc_ki[mc->mc_top] >= fixup) {
6862 m3->mc_pg[mc->mc_top] = rp;
6863 m3->mc_ki[mc->mc_top] -= fixup;
6864 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6866 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6867 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6876 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6877 MDB_val *key, MDB_val *data, unsigned int flags)
6882 assert(key != NULL);
6883 assert(data != NULL);
6885 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6888 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6892 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6896 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6899 mdb_cursor_init(&mc, txn, dbi, &mx);
6900 return mdb_cursor_put(&mc, key, data, flags);
6904 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6906 if ((flag & CHANGEABLE) != flag)
6909 env->me_flags |= flag;
6911 env->me_flags &= ~flag;
6916 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6921 *arg = env->me_flags;
6926 mdb_env_get_path(MDB_env *env, const char **arg)
6931 *arg = env->me_path;
6935 /** Common code for #mdb_stat() and #mdb_env_stat().
6936 * @param[in] env the environment to operate in.
6937 * @param[in] db the #MDB_db record containing the stats to return.
6938 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6939 * @return 0, this function always succeeds.
6942 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6944 arg->ms_psize = env->me_psize;
6945 arg->ms_depth = db->md_depth;
6946 arg->ms_branch_pages = db->md_branch_pages;
6947 arg->ms_leaf_pages = db->md_leaf_pages;
6948 arg->ms_overflow_pages = db->md_overflow_pages;
6949 arg->ms_entries = db->md_entries;
6954 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6958 if (env == NULL || arg == NULL)
6961 toggle = mdb_env_pick_meta(env);
6963 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6967 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6971 if (env == NULL || arg == NULL)
6974 toggle = mdb_env_pick_meta(env);
6975 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6976 arg->me_mapsize = env->me_mapsize;
6977 arg->me_maxreaders = env->me_maxreaders;
6978 arg->me_numreaders = env->me_numreaders;
6979 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6980 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6984 /** Set the default comparison functions for a database.
6985 * Called immediately after a database is opened to set the defaults.
6986 * The user can then override them with #mdb_set_compare() or
6987 * #mdb_set_dupsort().
6988 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6989 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6992 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6994 uint16_t f = txn->mt_dbs[dbi].md_flags;
6996 txn->mt_dbxs[dbi].md_cmp =
6997 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6998 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7000 txn->mt_dbxs[dbi].md_dcmp =
7001 !(f & MDB_DUPSORT) ? 0 :
7002 ((f & MDB_INTEGERDUP)
7003 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7004 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7007 #define PERSISTENT_FLAGS 0xffff
7008 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
7009 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
7010 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7016 int rc, dbflag, exact;
7017 unsigned int unused = 0;
7020 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7021 mdb_default_cmp(txn, FREE_DBI);
7024 if ((flags & VALID_FLAGS) != flags)
7030 if (flags & PERSISTENT_FLAGS) {
7031 uint16_t f2 = flags & PERSISTENT_FLAGS;
7032 /* make sure flag changes get committed */
7033 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7034 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7035 txn->mt_flags |= MDB_TXN_DIRTY;
7038 mdb_default_cmp(txn, MAIN_DBI);
7042 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7043 mdb_default_cmp(txn, MAIN_DBI);
7046 /* Is the DB already open? */
7048 for (i=2; i<txn->mt_numdbs; i++) {
7049 if (!txn->mt_dbxs[i].md_name.mv_size) {
7050 /* Remember this free slot */
7051 if (!unused) unused = i;
7054 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7055 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7061 /* If no free slot and max hit, fail */
7062 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7063 return MDB_DBS_FULL;
7065 /* Find the DB info */
7069 key.mv_data = (void *)name;
7070 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7071 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7072 if (rc == MDB_SUCCESS) {
7073 /* make sure this is actually a DB */
7074 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7075 if (!(node->mn_flags & F_SUBDATA))
7077 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7078 /* Create if requested */
7080 data.mv_size = sizeof(MDB_db);
7081 data.mv_data = &dummy;
7082 memset(&dummy, 0, sizeof(dummy));
7083 dummy.md_root = P_INVALID;
7084 dummy.md_flags = flags & PERSISTENT_FLAGS;
7085 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7089 /* OK, got info, add to table */
7090 if (rc == MDB_SUCCESS) {
7091 unsigned int slot = unused ? unused : txn->mt_numdbs;
7092 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7093 txn->mt_dbxs[slot].md_name.mv_size = len;
7094 txn->mt_dbxs[slot].md_rel = NULL;
7095 txn->mt_dbflags[slot] = dbflag;
7096 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7098 txn->mt_env->me_dbflags[slot] = mdflags = txn->mt_dbs[slot].md_flags;
7099 mdb_default_cmp(txn, slot);
7102 txn->mt_env->me_numdbs++;
7104 /* Open the DB in parent txns as well */
7105 while ((txn = txn->mt_parent) != NULL) {
7106 txn->mt_dbflags[slot] = DB_STALE;
7107 txn->mt_dbs[slot].md_flags = mdflags;
7116 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7118 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7121 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7124 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7127 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7129 ptr = env->me_dbxs[dbi].md_name.mv_data;
7130 env->me_dbxs[dbi].md_name.mv_data = NULL;
7131 env->me_dbxs[dbi].md_name.mv_size = 0;
7135 /** Add all the DB's pages to the free list.
7136 * @param[in] mc Cursor on the DB to free.
7137 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7138 * @return 0 on success, non-zero on failure.
7141 mdb_drop0(MDB_cursor *mc, int subs)
7145 rc = mdb_page_search(mc, NULL, 0);
7146 if (rc == MDB_SUCCESS) {
7151 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7152 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
7155 mdb_cursor_copy(mc, &mx);
7156 while (mc->mc_snum > 0) {
7157 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7158 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7159 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7160 if (ni->mn_flags & F_SUBDATA) {
7161 mdb_xcursor_init1(mc, ni);
7162 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7168 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7170 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7173 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7178 rc = mdb_cursor_sibling(mc, 1);
7180 /* no more siblings, go back to beginning
7181 * of previous level.
7184 for (i=1; i<mc->mc_top; i++)
7185 mc->mc_pg[i] = mx.mc_pg[i];
7189 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7190 mc->mc_db->md_root);
7195 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7200 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7203 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7206 rc = mdb_cursor_open(txn, dbi, &mc);
7210 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7214 /* Can't delete the main DB */
7215 if (del && dbi > MAIN_DBI) {
7216 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7218 txn->mt_dbflags[dbi] = DB_STALE;
7219 mdb_dbi_close(txn->mt_env, dbi);
7222 /* reset the DB record, mark it dirty */
7223 txn->mt_dbflags[dbi] |= DB_DIRTY;
7224 txn->mt_dbs[dbi].md_depth = 0;
7225 txn->mt_dbs[dbi].md_branch_pages = 0;
7226 txn->mt_dbs[dbi].md_leaf_pages = 0;
7227 txn->mt_dbs[dbi].md_overflow_pages = 0;
7228 txn->mt_dbs[dbi].md_entries = 0;
7229 txn->mt_dbs[dbi].md_root = P_INVALID;
7231 if (!txn->mt_u.dirty_list[0].mid) {
7234 /* make sure we have at least one dirty page in this txn
7235 * otherwise these changes will be ignored.
7237 key.mv_size = sizeof(txnid_t);
7238 key.mv_data = &txn->mt_txnid;
7239 data.mv_size = sizeof(MDB_ID);
7240 data.mv_data = txn->mt_free_pgs;
7241 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7242 rc = mdb_cursor_put(&m2, &key, &data, 0);
7246 mdb_cursor_close(mc);
7250 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7252 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7255 txn->mt_dbxs[dbi].md_cmp = cmp;
7259 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7261 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7264 txn->mt_dbxs[dbi].md_dcmp = cmp;
7268 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7270 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7273 txn->mt_dbxs[dbi].md_rel = rel;
7277 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7279 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7282 txn->mt_dbxs[dbi].md_relctx = ctx;