2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2014 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.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
93 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
94 #include <netinet/in.h>
95 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
98 #if defined(__APPLE__) || defined (BSD)
99 # define MDB_USE_POSIX_SEM 1
100 # define MDB_FDATASYNC fsync
101 #elif defined(ANDROID)
102 # define MDB_FDATASYNC fsync
107 #ifdef MDB_USE_POSIX_SEM
108 # define MDB_USE_HASH 1
109 #include <semaphore.h>
114 #include <valgrind/memcheck.h>
115 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
116 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
117 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
118 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
119 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
121 #define VGMEMP_CREATE(h,r,z)
122 #define VGMEMP_ALLOC(h,a,s)
123 #define VGMEMP_FREE(h,a)
124 #define VGMEMP_DESTROY(h)
125 #define VGMEMP_DEFINED(a,s)
129 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
130 /* Solaris just defines one or the other */
131 # define LITTLE_ENDIAN 1234
132 # define BIG_ENDIAN 4321
133 # ifdef _LITTLE_ENDIAN
134 # define BYTE_ORDER LITTLE_ENDIAN
136 # define BYTE_ORDER BIG_ENDIAN
139 # define BYTE_ORDER __BYTE_ORDER
143 #ifndef LITTLE_ENDIAN
144 #define LITTLE_ENDIAN __LITTLE_ENDIAN
147 #define BIG_ENDIAN __BIG_ENDIAN
150 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
151 #define MISALIGNED_OK 1
157 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
158 # error "Unknown or unsupported endianness (BYTE_ORDER)"
159 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
160 # error "Two's complement, reasonably sized integer types, please"
164 /** Put infrequently used env functions in separate section */
166 # define ESECT __attribute__ ((section("__TEXT,text_env")))
168 # define ESECT __attribute__ ((section("text_env")))
174 /** @defgroup internal LMDB Internals
177 /** @defgroup compat Compatibility Macros
178 * A bunch of macros to minimize the amount of platform-specific ifdefs
179 * needed throughout the rest of the code. When the features this library
180 * needs are similar enough to POSIX to be hidden in a one-or-two line
181 * replacement, this macro approach is used.
185 /** Features under development */
190 /** Wrapper around __func__, which is a C99 feature */
191 #if __STDC_VERSION__ >= 199901L
192 # define mdb_func_ __func__
193 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
194 # define mdb_func_ __FUNCTION__
196 /* If a debug message says <mdb_unknown>(), update the #if statements above */
197 # define mdb_func_ "<mdb_unknown>"
201 #define MDB_USE_HASH 1
202 #define MDB_PIDLOCK 0
203 #define THREAD_RET DWORD
204 #define pthread_t HANDLE
205 #define pthread_mutex_t HANDLE
206 #define pthread_cond_t HANDLE
207 #define pthread_key_t DWORD
208 #define pthread_self() GetCurrentThreadId()
209 #define pthread_key_create(x,y) \
210 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
211 #define pthread_key_delete(x) TlsFree(x)
212 #define pthread_getspecific(x) TlsGetValue(x)
213 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
214 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
215 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
216 #define pthread_cond_signal(x) SetEvent(*x)
217 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
218 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
219 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
220 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
221 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
222 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
223 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
224 #define getpid() GetCurrentProcessId()
225 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
226 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
227 #define ErrCode() GetLastError()
228 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
229 #define close(fd) (CloseHandle(fd) ? 0 : -1)
230 #define munmap(ptr,len) UnmapViewOfFile(ptr)
231 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
232 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
234 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
238 #define THREAD_RET void *
239 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
240 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
241 #define Z "z" /**< printf format modifier for size_t */
243 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
244 #define MDB_PIDLOCK 1
246 #ifdef MDB_USE_POSIX_SEM
248 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
249 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
250 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
251 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
254 mdb_sem_wait(sem_t *sem)
257 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
262 /** Lock the reader mutex.
264 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
265 /** Unlock the reader mutex.
267 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
269 /** Lock the writer mutex.
270 * Only a single write transaction is allowed at a time. Other writers
271 * will block waiting for this mutex.
273 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
274 /** Unlock the writer mutex.
276 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
277 #endif /* MDB_USE_POSIX_SEM */
279 /** Get the error code for the last failed system function.
281 #define ErrCode() errno
283 /** An abstraction for a file handle.
284 * On POSIX systems file handles are small integers. On Windows
285 * they're opaque pointers.
289 /** A value for an invalid file handle.
290 * Mainly used to initialize file variables and signify that they are
293 #define INVALID_HANDLE_VALUE (-1)
295 /** Get the size of a memory page for the system.
296 * This is the basic size that the platform's memory manager uses, and is
297 * fundamental to the use of memory-mapped files.
299 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
302 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
305 #define MNAME_LEN (sizeof(pthread_mutex_t))
311 /** A flag for opening a file and requesting synchronous data writes.
312 * This is only used when writing a meta page. It's not strictly needed;
313 * we could just do a normal write and then immediately perform a flush.
314 * But if this flag is available it saves us an extra system call.
316 * @note If O_DSYNC is undefined but exists in /usr/include,
317 * preferably set some compiler flag to get the definition.
318 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
321 # define MDB_DSYNC O_DSYNC
325 /** Function for flushing the data of a file. Define this to fsync
326 * if fdatasync() is not supported.
328 #ifndef MDB_FDATASYNC
329 # define MDB_FDATASYNC fdatasync
333 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
344 /** A page number in the database.
345 * Note that 64 bit page numbers are overkill, since pages themselves
346 * already represent 12-13 bits of addressable memory, and the OS will
347 * always limit applications to a maximum of 63 bits of address space.
349 * @note In the #MDB_node structure, we only store 48 bits of this value,
350 * which thus limits us to only 60 bits of addressable data.
352 typedef MDB_ID pgno_t;
354 /** A transaction ID.
355 * See struct MDB_txn.mt_txnid for details.
357 typedef MDB_ID txnid_t;
359 /** @defgroup debug Debug Macros
363 /** Enable debug output. Needs variable argument macros (a C99 feature).
364 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
365 * read from and written to the database (used for free space management).
371 static int mdb_debug;
372 static txnid_t mdb_debug_start;
374 /** Print a debug message with printf formatting.
375 * Requires double parenthesis around 2 or more args.
377 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
378 # define DPRINTF0(fmt, ...) \
379 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
381 # define DPRINTF(args) ((void) 0)
383 /** Print a debug string.
384 * The string is printed literally, with no format processing.
386 #define DPUTS(arg) DPRINTF(("%s", arg))
387 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
389 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
392 /** @brief The maximum size of a database page.
394 * It is 32k or 64k, since value-PAGEBASE must fit in
395 * #MDB_page.%mp_upper.
397 * LMDB will use database pages < OS pages if needed.
398 * That causes more I/O in write transactions: The OS must
399 * know (read) the whole page before writing a partial page.
401 * Note that we don't currently support Huge pages. On Linux,
402 * regular data files cannot use Huge pages, and in general
403 * Huge pages aren't actually pageable. We rely on the OS
404 * demand-pager to read our data and page it out when memory
405 * pressure from other processes is high. So until OSs have
406 * actual paging support for Huge pages, they're not viable.
408 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
410 /** The minimum number of keys required in a database page.
411 * Setting this to a larger value will place a smaller bound on the
412 * maximum size of a data item. Data items larger than this size will
413 * be pushed into overflow pages instead of being stored directly in
414 * the B-tree node. This value used to default to 4. With a page size
415 * of 4096 bytes that meant that any item larger than 1024 bytes would
416 * go into an overflow page. That also meant that on average 2-3KB of
417 * each overflow page was wasted space. The value cannot be lower than
418 * 2 because then there would no longer be a tree structure. With this
419 * value, items larger than 2KB will go into overflow pages, and on
420 * average only 1KB will be wasted.
422 #define MDB_MINKEYS 2
424 /** A stamp that identifies a file as an LMDB file.
425 * There's nothing special about this value other than that it is easily
426 * recognizable, and it will reflect any byte order mismatches.
428 #define MDB_MAGIC 0xBEEFC0DE
430 /** The version number for a database's datafile format. */
431 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
432 /** The version number for a database's lockfile format. */
433 #define MDB_LOCK_VERSION 1
435 /** @brief The max size of a key we can write, or 0 for dynamic max.
437 * Define this as 0 to compute the max from the page size. 511
438 * is default for backwards compat: liblmdb <= 0.9.10 can break
439 * when modifying a DB with keys/dupsort data bigger than its max.
440 * #MDB_DEVEL sets the default to 0.
442 * Data items in an #MDB_DUPSORT database are also limited to
443 * this size, since they're actually keys of a sub-DB. Keys and
444 * #MDB_DUPSORT data items must fit on a node in a regular page.
446 #ifndef MDB_MAXKEYSIZE
447 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
450 /** The maximum size of a key we can write to the environment. */
452 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
454 #define ENV_MAXKEY(env) ((env)->me_maxkey)
457 /** @brief The maximum size of a data item.
459 * We only store a 32 bit value for node sizes.
461 #define MAXDATASIZE 0xffffffffUL
464 /** Key size which fits in a #DKBUF.
467 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
470 * This is used for printing a hex dump of a key's contents.
472 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
473 /** Display a key in hex.
475 * Invoke a function to display a key in hex.
477 #define DKEY(x) mdb_dkey(x, kbuf)
483 /** An invalid page number.
484 * Mainly used to denote an empty tree.
486 #define P_INVALID (~(pgno_t)0)
488 /** Test if the flags \b f are set in a flag word \b w. */
489 #define F_ISSET(w, f) (((w) & (f)) == (f))
491 /** Round \b n up to an even number. */
492 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
494 /** Used for offsets within a single page.
495 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
498 typedef uint16_t indx_t;
500 /** Default size of memory map.
501 * This is certainly too small for any actual applications. Apps should always set
502 * the size explicitly using #mdb_env_set_mapsize().
504 #define DEFAULT_MAPSIZE 1048576
506 /** @defgroup readers Reader Lock Table
507 * Readers don't acquire any locks for their data access. Instead, they
508 * simply record their transaction ID in the reader table. The reader
509 * mutex is needed just to find an empty slot in the reader table. The
510 * slot's address is saved in thread-specific data so that subsequent read
511 * transactions started by the same thread need no further locking to proceed.
513 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
515 * No reader table is used if the database is on a read-only filesystem, or
516 * if #MDB_NOLOCK is set.
518 * Since the database uses multi-version concurrency control, readers don't
519 * actually need any locking. This table is used to keep track of which
520 * readers are using data from which old transactions, so that we'll know
521 * when a particular old transaction is no longer in use. Old transactions
522 * that have discarded any data pages can then have those pages reclaimed
523 * for use by a later write transaction.
525 * The lock table is constructed such that reader slots are aligned with the
526 * processor's cache line size. Any slot is only ever used by one thread.
527 * This alignment guarantees that there will be no contention or cache
528 * thrashing as threads update their own slot info, and also eliminates
529 * any need for locking when accessing a slot.
531 * A writer thread will scan every slot in the table to determine the oldest
532 * outstanding reader transaction. Any freed pages older than this will be
533 * reclaimed by the writer. The writer doesn't use any locks when scanning
534 * this table. This means that there's no guarantee that the writer will
535 * see the most up-to-date reader info, but that's not required for correct
536 * operation - all we need is to know the upper bound on the oldest reader,
537 * we don't care at all about the newest reader. So the only consequence of
538 * reading stale information here is that old pages might hang around a
539 * while longer before being reclaimed. That's actually good anyway, because
540 * the longer we delay reclaiming old pages, the more likely it is that a
541 * string of contiguous pages can be found after coalescing old pages from
542 * many old transactions together.
545 /** Number of slots in the reader table.
546 * This value was chosen somewhat arbitrarily. 126 readers plus a
547 * couple mutexes fit exactly into 8KB on my development machine.
548 * Applications should set the table size using #mdb_env_set_maxreaders().
550 #define DEFAULT_READERS 126
552 /** The size of a CPU cache line in bytes. We want our lock structures
553 * aligned to this size to avoid false cache line sharing in the
555 * This value works for most CPUs. For Itanium this should be 128.
561 /** The information we store in a single slot of the reader table.
562 * In addition to a transaction ID, we also record the process and
563 * thread ID that owns a slot, so that we can detect stale information,
564 * e.g. threads or processes that went away without cleaning up.
565 * @note We currently don't check for stale records. We simply re-init
566 * the table when we know that we're the only process opening the
569 typedef struct MDB_rxbody {
570 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
571 * Multiple readers that start at the same time will probably have the
572 * same ID here. Again, it's not important to exclude them from
573 * anything; all we need to know is which version of the DB they
574 * started from so we can avoid overwriting any data used in that
575 * particular version.
578 /** The process ID of the process owning this reader txn. */
580 /** The thread ID of the thread owning this txn. */
584 /** The actual reader record, with cacheline padding. */
585 typedef struct MDB_reader {
588 /** shorthand for mrb_txnid */
589 #define mr_txnid mru.mrx.mrb_txnid
590 #define mr_pid mru.mrx.mrb_pid
591 #define mr_tid mru.mrx.mrb_tid
592 /** cache line alignment */
593 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
597 /** The header for the reader table.
598 * The table resides in a memory-mapped file. (This is a different file
599 * than is used for the main database.)
601 * For POSIX the actual mutexes reside in the shared memory of this
602 * mapped file. On Windows, mutexes are named objects allocated by the
603 * kernel; we store the mutex names in this mapped file so that other
604 * processes can grab them. This same approach is also used on
605 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
606 * process-shared POSIX mutexes. For these cases where a named object
607 * is used, the object name is derived from a 64 bit FNV hash of the
608 * environment pathname. As such, naming collisions are extremely
609 * unlikely. If a collision occurs, the results are unpredictable.
611 typedef struct MDB_txbody {
612 /** Stamp identifying this as an LMDB file. It must be set
615 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
617 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
618 char mtb_rmname[MNAME_LEN];
620 /** Mutex protecting access to this table.
621 * This is the reader lock that #LOCK_MUTEX_R acquires.
623 pthread_mutex_t mtb_mutex;
625 /** The ID of the last transaction committed to the database.
626 * This is recorded here only for convenience; the value can always
627 * be determined by reading the main database meta pages.
630 /** The number of slots that have been used in the reader table.
631 * This always records the maximum count, it is not decremented
632 * when readers release their slots.
634 unsigned mtb_numreaders;
637 /** The actual reader table definition. */
638 typedef struct MDB_txninfo {
641 #define mti_magic mt1.mtb.mtb_magic
642 #define mti_format mt1.mtb.mtb_format
643 #define mti_mutex mt1.mtb.mtb_mutex
644 #define mti_rmname mt1.mtb.mtb_rmname
645 #define mti_txnid mt1.mtb.mtb_txnid
646 #define mti_numreaders mt1.mtb.mtb_numreaders
647 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
650 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
651 char mt2_wmname[MNAME_LEN];
652 #define mti_wmname mt2.mt2_wmname
654 pthread_mutex_t mt2_wmutex;
655 #define mti_wmutex mt2.mt2_wmutex
657 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
659 MDB_reader mti_readers[1];
662 /** Lockfile format signature: version, features and field layout */
663 #define MDB_LOCK_FORMAT \
665 ((MDB_LOCK_VERSION) \
666 /* Flags which describe functionality */ \
667 + (((MDB_PIDLOCK) != 0) << 16)))
670 /** Common header for all page types.
671 * Overflow records occupy a number of contiguous pages with no
672 * headers on any page after the first.
674 typedef struct MDB_page {
675 #define mp_pgno mp_p.p_pgno
676 #define mp_next mp_p.p_next
678 pgno_t p_pgno; /**< page number */
679 struct MDB_page *p_next; /**< for in-memory list of freed pages */
682 /** @defgroup mdb_page Page Flags
684 * Flags for the page headers.
687 #define P_BRANCH 0x01 /**< branch page */
688 #define P_LEAF 0x02 /**< leaf page */
689 #define P_OVERFLOW 0x04 /**< overflow page */
690 #define P_META 0x08 /**< meta page */
691 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
692 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
693 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
694 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
695 #define P_KEEP 0x8000 /**< leave this page alone during spill */
697 uint16_t mp_flags; /**< @ref mdb_page */
698 #define mp_lower mp_pb.pb.pb_lower
699 #define mp_upper mp_pb.pb.pb_upper
700 #define mp_pages mp_pb.pb_pages
703 indx_t pb_lower; /**< lower bound of free space */
704 indx_t pb_upper; /**< upper bound of free space */
706 uint32_t pb_pages; /**< number of overflow pages */
708 indx_t mp_ptrs[1]; /**< dynamic size */
711 /** Size of the page header, excluding dynamic data at the end */
712 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
714 /** Address of first usable data byte in a page, after the header */
715 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
717 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
718 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
720 /** Number of nodes on a page */
721 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
723 /** The amount of space remaining in the page */
724 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
726 /** The percentage of space used in the page, in tenths of a percent. */
727 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
728 ((env)->me_psize - PAGEHDRSZ))
729 /** The minimum page fill factor, in tenths of a percent.
730 * Pages emptier than this are candidates for merging.
732 #define FILL_THRESHOLD 250
734 /** Test if a page is a leaf page */
735 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
736 /** Test if a page is a LEAF2 page */
737 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
738 /** Test if a page is a branch page */
739 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
740 /** Test if a page is an overflow page */
741 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
742 /** Test if a page is a sub page */
743 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
745 /** The number of overflow pages needed to store the given size. */
746 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
748 /** Link in #MDB_txn.%mt_loose_pgs list */
749 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
751 /** Header for a single key/data pair within a page.
752 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
753 * We guarantee 2-byte alignment for 'MDB_node's.
755 typedef struct MDB_node {
756 /** lo and hi are used for data size on leaf nodes and for
757 * child pgno on branch nodes. On 64 bit platforms, flags
758 * is also used for pgno. (Branch nodes have no flags).
759 * They are in host byte order in case that lets some
760 * accesses be optimized into a 32-bit word access.
762 #if BYTE_ORDER == LITTLE_ENDIAN
763 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
765 unsigned short mn_hi, mn_lo;
767 /** @defgroup mdb_node Node Flags
769 * Flags for node headers.
772 #define F_BIGDATA 0x01 /**< data put on overflow page */
773 #define F_SUBDATA 0x02 /**< data is a sub-database */
774 #define F_DUPDATA 0x04 /**< data has duplicates */
776 /** valid flags for #mdb_node_add() */
777 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
780 unsigned short mn_flags; /**< @ref mdb_node */
781 unsigned short mn_ksize; /**< key size */
782 char mn_data[1]; /**< key and data are appended here */
785 /** Size of the node header, excluding dynamic data at the end */
786 #define NODESIZE offsetof(MDB_node, mn_data)
788 /** Bit position of top word in page number, for shifting mn_flags */
789 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
791 /** Size of a node in a branch page with a given key.
792 * This is just the node header plus the key, there is no data.
794 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
796 /** Size of a node in a leaf page with a given key and data.
797 * This is node header plus key plus data size.
799 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
801 /** Address of node \b i in page \b p */
802 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
804 /** Address of the key for the node */
805 #define NODEKEY(node) (void *)((node)->mn_data)
807 /** Address of the data for a node */
808 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
810 /** Get the page number pointed to by a branch node */
811 #define NODEPGNO(node) \
812 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
813 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
814 /** Set the page number in a branch node */
815 #define SETPGNO(node,pgno) do { \
816 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
817 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
819 /** Get the size of the data in a leaf node */
820 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
821 /** Set the size of the data for a leaf node */
822 #define SETDSZ(node,size) do { \
823 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
824 /** The size of a key in a node */
825 #define NODEKSZ(node) ((node)->mn_ksize)
827 /** Copy a page number from src to dst */
829 #define COPY_PGNO(dst,src) dst = src
831 #if SIZE_MAX > 4294967295UL
832 #define COPY_PGNO(dst,src) do { \
833 unsigned short *s, *d; \
834 s = (unsigned short *)&(src); \
835 d = (unsigned short *)&(dst); \
842 #define COPY_PGNO(dst,src) do { \
843 unsigned short *s, *d; \
844 s = (unsigned short *)&(src); \
845 d = (unsigned short *)&(dst); \
851 /** The address of a key in a LEAF2 page.
852 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
853 * There are no node headers, keys are stored contiguously.
855 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
857 /** Set the \b node's key into \b keyptr, if requested. */
858 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
859 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
861 /** Set the \b node's key into \b key. */
862 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
864 /** Information about a single database in the environment. */
865 typedef struct MDB_db {
866 uint32_t md_pad; /**< also ksize for LEAF2 pages */
867 uint16_t md_flags; /**< @ref mdb_dbi_open */
868 uint16_t md_depth; /**< depth of this tree */
869 pgno_t md_branch_pages; /**< number of internal pages */
870 pgno_t md_leaf_pages; /**< number of leaf pages */
871 pgno_t md_overflow_pages; /**< number of overflow pages */
872 size_t md_entries; /**< number of data items */
873 pgno_t md_root; /**< the root page of this tree */
876 /** mdb_dbi_open flags */
877 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
878 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
879 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
880 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
882 /** Handle for the DB used to track free pages. */
884 /** Handle for the default DB. */
887 /** Meta page content.
888 * A meta page is the start point for accessing a database snapshot.
889 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
891 typedef struct MDB_meta {
892 /** Stamp identifying this as an LMDB file. It must be set
895 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
897 void *mm_address; /**< address for fixed mapping */
898 size_t mm_mapsize; /**< size of mmap region */
899 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
900 /** The size of pages used in this DB */
901 #define mm_psize mm_dbs[0].md_pad
902 /** Any persistent environment flags. @ref mdb_env */
903 #define mm_flags mm_dbs[0].md_flags
904 pgno_t mm_last_pg; /**< last used page in file */
905 txnid_t mm_txnid; /**< txnid that committed this page */
908 /** Buffer for a stack-allocated meta page.
909 * The members define size and alignment, and silence type
910 * aliasing warnings. They are not used directly; that could
911 * mean incorrectly using several union members in parallel.
913 typedef union MDB_metabuf {
916 char mm_pad[PAGEHDRSZ];
921 /** Auxiliary DB info.
922 * The information here is mostly static/read-only. There is
923 * only a single copy of this record in the environment.
925 typedef struct MDB_dbx {
926 MDB_val md_name; /**< name of the database */
927 MDB_cmp_func *md_cmp; /**< function for comparing keys */
928 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
929 MDB_rel_func *md_rel; /**< user relocate function */
930 void *md_relctx; /**< user-provided context for md_rel */
933 /** A database transaction.
934 * Every operation requires a transaction handle.
937 MDB_txn *mt_parent; /**< parent of a nested txn */
938 MDB_txn *mt_child; /**< nested txn under this txn */
939 pgno_t mt_next_pgno; /**< next unallocated page */
940 /** The ID of this transaction. IDs are integers incrementing from 1.
941 * Only committed write transactions increment the ID. If a transaction
942 * aborts, the ID may be re-used by the next writer.
945 MDB_env *mt_env; /**< the DB environment */
946 /** The list of pages that became unused during this transaction.
949 /** The list of loose pages that became unused and may be reused
950 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
952 MDB_page *mt_loose_pgs;
953 /* #Number of loose pages (#mt_loose_pgs) */
955 /** The sorted list of dirty pages we temporarily wrote to disk
956 * because the dirty list was full. page numbers in here are
957 * shifted left by 1, deleted slots have the LSB set.
959 MDB_IDL mt_spill_pgs;
961 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
963 /** For read txns: This thread/txn's reader table slot, or NULL. */
966 /** Array of records for each DB known in the environment. */
968 /** Array of MDB_db records for each known DB */
970 /** Array of sequence numbers for each DB handle */
971 unsigned int *mt_dbiseqs;
972 /** @defgroup mt_dbflag Transaction DB Flags
976 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
977 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
978 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
979 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
981 /** In write txns, array of cursors for each DB */
982 MDB_cursor **mt_cursors;
983 /** Array of flags for each DB */
984 unsigned char *mt_dbflags;
985 /** Number of DB records in use. This number only ever increments;
986 * we don't decrement it when individual DB handles are closed.
990 /** @defgroup mdb_txn Transaction Flags
994 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
995 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
996 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
997 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
999 unsigned int mt_flags; /**< @ref mdb_txn */
1000 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1001 * Includes ancestor txns' dirty pages not hidden by other txns'
1002 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1003 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1005 unsigned int mt_dirty_room;
1008 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1009 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1010 * raise this on a 64 bit machine.
1012 #define CURSOR_STACK 32
1016 /** Cursors are used for all DB operations.
1017 * A cursor holds a path of (page pointer, key index) from the DB
1018 * root to a position in the DB, plus other state. #MDB_DUPSORT
1019 * cursors include an xcursor to the current data item. Write txns
1020 * track their cursors and keep them up to date when data moves.
1021 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1022 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1025 /** Next cursor on this DB in this txn */
1026 MDB_cursor *mc_next;
1027 /** Backup of the original cursor if this cursor is a shadow */
1028 MDB_cursor *mc_backup;
1029 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1030 struct MDB_xcursor *mc_xcursor;
1031 /** The transaction that owns this cursor */
1033 /** The database handle this cursor operates on */
1035 /** The database record for this cursor */
1037 /** The database auxiliary record for this cursor */
1039 /** The @ref mt_dbflag for this database */
1040 unsigned char *mc_dbflag;
1041 unsigned short mc_snum; /**< number of pushed pages */
1042 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1043 /** @defgroup mdb_cursor Cursor Flags
1045 * Cursor state flags.
1048 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1049 #define C_EOF 0x02 /**< No more data */
1050 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1051 #define C_DEL 0x08 /**< last op was a cursor_del */
1052 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1053 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1055 unsigned int mc_flags; /**< @ref mdb_cursor */
1056 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1057 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1060 /** Context for sorted-dup records.
1061 * We could have gone to a fully recursive design, with arbitrarily
1062 * deep nesting of sub-databases. But for now we only handle these
1063 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1065 typedef struct MDB_xcursor {
1066 /** A sub-cursor for traversing the Dup DB */
1067 MDB_cursor mx_cursor;
1068 /** The database record for this Dup DB */
1070 /** The auxiliary DB record for this Dup DB */
1072 /** The @ref mt_dbflag for this Dup DB */
1073 unsigned char mx_dbflag;
1076 /** State of FreeDB old pages, stored in the MDB_env */
1077 typedef struct MDB_pgstate {
1078 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1079 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1082 /** The database environment. */
1084 HANDLE me_fd; /**< The main data file */
1085 HANDLE me_lfd; /**< The lock file */
1086 HANDLE me_mfd; /**< just for writing the meta pages */
1087 /** Failed to update the meta page. Probably an I/O error. */
1088 #define MDB_FATAL_ERROR 0x80000000U
1089 /** Some fields are initialized. */
1090 #define MDB_ENV_ACTIVE 0x20000000U
1091 /** me_txkey is set */
1092 #define MDB_ENV_TXKEY 0x10000000U
1093 uint32_t me_flags; /**< @ref mdb_env */
1094 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1095 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1096 unsigned int me_maxreaders; /**< size of the reader table */
1097 unsigned int me_numreaders; /**< max numreaders set by this env */
1098 MDB_dbi me_numdbs; /**< number of DBs opened */
1099 MDB_dbi me_maxdbs; /**< size of the DB table */
1100 MDB_PID_T me_pid; /**< process ID of this env */
1101 char *me_path; /**< path to the DB files */
1102 char *me_map; /**< the memory map of the data file */
1103 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1104 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1105 void *me_pbuf; /**< scratch area for DUPSORT put() */
1106 MDB_txn *me_txn; /**< current write transaction */
1107 size_t me_mapsize; /**< size of the data memory map */
1108 off_t me_size; /**< current file size */
1109 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1110 MDB_dbx *me_dbxs; /**< array of static DB info */
1111 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1112 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1113 pthread_key_t me_txkey; /**< thread-key for readers */
1114 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1115 # define me_pglast me_pgstate.mf_pglast
1116 # define me_pghead me_pgstate.mf_pghead
1117 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1118 /** IDL of pages that became unused in a write txn */
1119 MDB_IDL me_free_pgs;
1120 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1121 MDB_ID2L me_dirty_list;
1122 /** Max number of freelist items that can fit in a single overflow page */
1124 /** Max size of a node on a page */
1125 unsigned int me_nodemax;
1126 #if !(MDB_MAXKEYSIZE)
1127 unsigned int me_maxkey; /**< max size of a key */
1129 int me_live_reader; /**< have liveness lock in reader table */
1131 int me_pidquery; /**< Used in OpenProcess */
1132 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1134 #elif defined(MDB_USE_POSIX_SEM)
1135 sem_t *me_rmutex; /* Shared mutexes are not supported */
1138 void *me_userctx; /**< User-settable context */
1139 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1142 /** Nested transaction */
1143 typedef struct MDB_ntxn {
1144 MDB_txn mnt_txn; /**< the transaction */
1145 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1148 /** max number of pages to commit in one writev() call */
1149 #define MDB_COMMIT_PAGES 64
1150 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1151 #undef MDB_COMMIT_PAGES
1152 #define MDB_COMMIT_PAGES IOV_MAX
1155 /** max bytes to write in one call */
1156 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1158 /** Check \b txn and \b dbi arguments to a function */
1159 #define TXN_DBI_EXIST(txn, dbi) \
1160 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1162 /** Check for misused \b dbi handles */
1163 #define TXN_DBI_CHANGED(txn, dbi) \
1164 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1166 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1167 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1168 static int mdb_page_touch(MDB_cursor *mc);
1170 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1171 static int mdb_page_search_root(MDB_cursor *mc,
1172 MDB_val *key, int modify);
1173 #define MDB_PS_MODIFY 1
1174 #define MDB_PS_ROOTONLY 2
1175 #define MDB_PS_FIRST 4
1176 #define MDB_PS_LAST 8
1177 static int mdb_page_search(MDB_cursor *mc,
1178 MDB_val *key, int flags);
1179 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1181 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1182 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1183 pgno_t newpgno, unsigned int nflags);
1185 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1186 static int mdb_env_pick_meta(const MDB_env *env);
1187 static int mdb_env_write_meta(MDB_txn *txn);
1188 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1189 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1191 static void mdb_env_close0(MDB_env *env, int excl);
1193 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1194 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1195 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1196 static void mdb_node_del(MDB_cursor *mc, int ksize);
1197 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1198 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1199 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1200 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1201 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1203 static int mdb_rebalance(MDB_cursor *mc);
1204 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1206 static void mdb_cursor_pop(MDB_cursor *mc);
1207 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1209 static int mdb_cursor_del0(MDB_cursor *mc);
1210 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1211 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1212 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1213 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1214 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1216 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1217 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1219 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1220 static void mdb_xcursor_init0(MDB_cursor *mc);
1221 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1223 static int mdb_drop0(MDB_cursor *mc, int subs);
1224 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1227 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1231 static SECURITY_DESCRIPTOR mdb_null_sd;
1232 static SECURITY_ATTRIBUTES mdb_all_sa;
1233 static int mdb_sec_inited;
1236 /** Return the library version info. */
1238 mdb_version(int *major, int *minor, int *patch)
1240 if (major) *major = MDB_VERSION_MAJOR;
1241 if (minor) *minor = MDB_VERSION_MINOR;
1242 if (patch) *patch = MDB_VERSION_PATCH;
1243 return MDB_VERSION_STRING;
1246 /** Table of descriptions for LMDB @ref errors */
1247 static char *const mdb_errstr[] = {
1248 "MDB_KEYEXIST: Key/data pair already exists",
1249 "MDB_NOTFOUND: No matching key/data pair found",
1250 "MDB_PAGE_NOTFOUND: Requested page not found",
1251 "MDB_CORRUPTED: Located page was wrong type",
1252 "MDB_PANIC: Update of meta page failed",
1253 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1254 "MDB_INVALID: File is not an LMDB file",
1255 "MDB_MAP_FULL: Environment mapsize limit reached",
1256 "MDB_DBS_FULL: Environment maxdbs limit reached",
1257 "MDB_READERS_FULL: Environment maxreaders limit reached",
1258 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1259 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1260 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1261 "MDB_PAGE_FULL: Internal error - page has no more space",
1262 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1263 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1264 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1265 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1266 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1267 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1271 mdb_strerror(int err)
1274 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1275 * This works as long as no function between the call to mdb_strerror
1276 * and the actual use of the message uses more than 4K of stack.
1279 char buf[1024], *ptr = buf;
1283 return ("Successful return: 0");
1285 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1286 i = err - MDB_KEYEXIST;
1287 return mdb_errstr[i];
1291 /* These are the C-runtime error codes we use. The comment indicates
1292 * their numeric value, and the Win32 error they would correspond to
1293 * if the error actually came from a Win32 API. A major mess, we should
1294 * have used LMDB-specific error codes for everything.
1297 case ENOENT: /* 2, FILE_NOT_FOUND */
1298 case EIO: /* 5, ACCESS_DENIED */
1299 case ENOMEM: /* 12, INVALID_ACCESS */
1300 case EACCES: /* 13, INVALID_DATA */
1301 case EBUSY: /* 16, CURRENT_DIRECTORY */
1302 case EINVAL: /* 22, BAD_COMMAND */
1303 case ENOSPC: /* 28, OUT_OF_PAPER */
1304 return strerror(err);
1309 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1310 FORMAT_MESSAGE_IGNORE_INSERTS,
1311 NULL, err, 0, ptr, sizeof(buf), pad);
1314 return strerror(err);
1318 /** assert(3) variant in cursor context */
1319 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1320 /** assert(3) variant in transaction context */
1321 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1322 /** assert(3) variant in environment context */
1323 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1326 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1327 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1330 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1331 const char *func, const char *file, int line)
1334 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1335 file, line, expr_txt, func);
1336 if (env->me_assert_func)
1337 env->me_assert_func(env, buf);
1338 fprintf(stderr, "%s\n", buf);
1342 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1346 /** Return the page number of \b mp which may be sub-page, for debug output */
1348 mdb_dbg_pgno(MDB_page *mp)
1351 COPY_PGNO(ret, mp->mp_pgno);
1355 /** Display a key in hexadecimal and return the address of the result.
1356 * @param[in] key the key to display
1357 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1358 * @return The key in hexadecimal form.
1361 mdb_dkey(MDB_val *key, char *buf)
1364 unsigned char *c = key->mv_data;
1370 if (key->mv_size > DKBUF_MAXKEYSIZE)
1371 return "MDB_MAXKEYSIZE";
1372 /* may want to make this a dynamic check: if the key is mostly
1373 * printable characters, print it as-is instead of converting to hex.
1377 for (i=0; i<key->mv_size; i++)
1378 ptr += sprintf(ptr, "%02x", *c++);
1380 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1386 mdb_leafnode_type(MDB_node *n)
1388 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1389 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1390 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1393 /** Display all the keys in the page. */
1395 mdb_page_list(MDB_page *mp)
1397 pgno_t pgno = mdb_dbg_pgno(mp);
1398 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1400 unsigned int i, nkeys, nsize, total = 0;
1404 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1405 case P_BRANCH: type = "Branch page"; break;
1406 case P_LEAF: type = "Leaf page"; break;
1407 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1408 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1409 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1411 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1412 pgno, mp->mp_pages, state);
1415 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1416 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1419 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1423 nkeys = NUMKEYS(mp);
1424 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1426 for (i=0; i<nkeys; i++) {
1427 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1428 key.mv_size = nsize = mp->mp_pad;
1429 key.mv_data = LEAF2KEY(mp, i, nsize);
1431 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1434 node = NODEPTR(mp, i);
1435 key.mv_size = node->mn_ksize;
1436 key.mv_data = node->mn_data;
1437 nsize = NODESIZE + key.mv_size;
1438 if (IS_BRANCH(mp)) {
1439 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1443 if (F_ISSET(node->mn_flags, F_BIGDATA))
1444 nsize += sizeof(pgno_t);
1446 nsize += NODEDSZ(node);
1448 nsize += sizeof(indx_t);
1449 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1450 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1452 total = EVEN(total);
1454 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1455 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1459 mdb_cursor_chk(MDB_cursor *mc)
1465 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1466 for (i=0; i<mc->mc_top; i++) {
1468 node = NODEPTR(mp, mc->mc_ki[i]);
1469 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1472 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1478 /** Count all the pages in each DB and in the freelist
1479 * and make sure it matches the actual number of pages
1481 * All named DBs must be open for a correct count.
1483 static void mdb_audit(MDB_txn *txn)
1487 MDB_ID freecount, count;
1492 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1493 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1494 freecount += *(MDB_ID *)data.mv_data;
1495 mdb_tassert(txn, rc == MDB_NOTFOUND);
1498 for (i = 0; i<txn->mt_numdbs; i++) {
1500 if (!(txn->mt_dbflags[i] & DB_VALID))
1502 mdb_cursor_init(&mc, txn, i, &mx);
1503 if (txn->mt_dbs[i].md_root == P_INVALID)
1505 count += txn->mt_dbs[i].md_branch_pages +
1506 txn->mt_dbs[i].md_leaf_pages +
1507 txn->mt_dbs[i].md_overflow_pages;
1508 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1509 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1510 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1513 mp = mc.mc_pg[mc.mc_top];
1514 for (j=0; j<NUMKEYS(mp); j++) {
1515 MDB_node *leaf = NODEPTR(mp, j);
1516 if (leaf->mn_flags & F_SUBDATA) {
1518 memcpy(&db, NODEDATA(leaf), sizeof(db));
1519 count += db.md_branch_pages + db.md_leaf_pages +
1520 db.md_overflow_pages;
1524 mdb_tassert(txn, rc == MDB_NOTFOUND);
1527 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1528 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1529 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1535 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1537 return txn->mt_dbxs[dbi].md_cmp(a, b);
1541 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1543 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1546 /** Allocate memory for a page.
1547 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1550 mdb_page_malloc(MDB_txn *txn, unsigned num)
1552 MDB_env *env = txn->mt_env;
1553 MDB_page *ret = env->me_dpages;
1554 size_t psize = env->me_psize, sz = psize, off;
1555 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1556 * For a single page alloc, we init everything after the page header.
1557 * For multi-page, we init the final page; if the caller needed that
1558 * many pages they will be filling in at least up to the last page.
1562 VGMEMP_ALLOC(env, ret, sz);
1563 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1564 env->me_dpages = ret->mp_next;
1567 psize -= off = PAGEHDRSZ;
1572 if ((ret = malloc(sz)) != NULL) {
1573 VGMEMP_ALLOC(env, ret, sz);
1574 if (!(env->me_flags & MDB_NOMEMINIT)) {
1575 memset((char *)ret + off, 0, psize);
1579 txn->mt_flags |= MDB_TXN_ERROR;
1583 /** Free a single page.
1584 * Saves single pages to a list, for future reuse.
1585 * (This is not used for multi-page overflow pages.)
1588 mdb_page_free(MDB_env *env, MDB_page *mp)
1590 mp->mp_next = env->me_dpages;
1591 VGMEMP_FREE(env, mp);
1592 env->me_dpages = mp;
1595 /** Free a dirty page */
1597 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1599 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1600 mdb_page_free(env, dp);
1602 /* large pages just get freed directly */
1603 VGMEMP_FREE(env, dp);
1608 /** Return all dirty pages to dpage list */
1610 mdb_dlist_free(MDB_txn *txn)
1612 MDB_env *env = txn->mt_env;
1613 MDB_ID2L dl = txn->mt_u.dirty_list;
1614 unsigned i, n = dl[0].mid;
1616 for (i = 1; i <= n; i++) {
1617 mdb_dpage_free(env, dl[i].mptr);
1622 /** Loosen or free a single page.
1623 * Saves single pages to a list for future reuse
1624 * in this same txn. It has been pulled from the freeDB
1625 * and already resides on the dirty list, but has been
1626 * deleted. Use these pages first before pulling again
1629 * If the page wasn't dirtied in this txn, just add it
1630 * to this txn's free list.
1633 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1636 pgno_t pgno = mp->mp_pgno;
1638 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1639 if (mc->mc_txn->mt_parent) {
1640 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1641 /* If txn has a parent, make sure the page is in our
1645 unsigned x = mdb_mid2l_search(dl, pgno);
1646 if (x <= dl[0].mid && dl[x].mid == pgno) {
1647 if (mp != dl[x].mptr) { /* bad cursor? */
1648 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1649 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1650 return MDB_CORRUPTED;
1657 /* no parent txn, so it's just ours */
1662 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1664 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1665 mc->mc_txn->mt_loose_pgs = mp;
1666 mc->mc_txn->mt_loose_count++;
1667 mp->mp_flags |= P_LOOSE;
1669 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1677 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1678 * @param[in] mc A cursor handle for the current operation.
1679 * @param[in] pflags Flags of the pages to update:
1680 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1681 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1682 * @return 0 on success, non-zero on failure.
1685 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1687 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1688 MDB_txn *txn = mc->mc_txn;
1694 int rc = MDB_SUCCESS, level;
1696 /* Mark pages seen by cursors */
1697 if (mc->mc_flags & C_UNTRACK)
1698 mc = NULL; /* will find mc in mt_cursors */
1699 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1700 for (; mc; mc=mc->mc_next) {
1701 if (!(mc->mc_flags & C_INITIALIZED))
1703 for (m3 = mc;; m3 = &mx->mx_cursor) {
1705 for (j=0; j<m3->mc_snum; j++) {
1707 if ((mp->mp_flags & Mask) == pflags)
1708 mp->mp_flags ^= P_KEEP;
1710 mx = m3->mc_xcursor;
1711 /* Proceed to mx if it is at a sub-database */
1712 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1714 if (! (mp && (mp->mp_flags & P_LEAF)))
1716 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1717 if (!(leaf->mn_flags & F_SUBDATA))
1726 /* Mark dirty root pages */
1727 for (i=0; i<txn->mt_numdbs; i++) {
1728 if (txn->mt_dbflags[i] & DB_DIRTY) {
1729 pgno_t pgno = txn->mt_dbs[i].md_root;
1730 if (pgno == P_INVALID)
1732 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1734 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1735 dp->mp_flags ^= P_KEEP;
1743 static int mdb_page_flush(MDB_txn *txn, int keep);
1745 /** Spill pages from the dirty list back to disk.
1746 * This is intended to prevent running into #MDB_TXN_FULL situations,
1747 * but note that they may still occur in a few cases:
1748 * 1) our estimate of the txn size could be too small. Currently this
1749 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1750 * 2) child txns may run out of space if their parents dirtied a
1751 * lot of pages and never spilled them. TODO: we probably should do
1752 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1753 * the parent's dirty_room is below a given threshold.
1755 * Otherwise, if not using nested txns, it is expected that apps will
1756 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1757 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1758 * If the txn never references them again, they can be left alone.
1759 * If the txn only reads them, they can be used without any fuss.
1760 * If the txn writes them again, they can be dirtied immediately without
1761 * going thru all of the work of #mdb_page_touch(). Such references are
1762 * handled by #mdb_page_unspill().
1764 * Also note, we never spill DB root pages, nor pages of active cursors,
1765 * because we'll need these back again soon anyway. And in nested txns,
1766 * we can't spill a page in a child txn if it was already spilled in a
1767 * parent txn. That would alter the parent txns' data even though
1768 * the child hasn't committed yet, and we'd have no way to undo it if
1769 * the child aborted.
1771 * @param[in] m0 cursor A cursor handle identifying the transaction and
1772 * database for which we are checking space.
1773 * @param[in] key For a put operation, the key being stored.
1774 * @param[in] data For a put operation, the data being stored.
1775 * @return 0 on success, non-zero on failure.
1778 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1780 MDB_txn *txn = m0->mc_txn;
1782 MDB_ID2L dl = txn->mt_u.dirty_list;
1783 unsigned int i, j, need;
1786 if (m0->mc_flags & C_SUB)
1789 /* Estimate how much space this op will take */
1790 i = m0->mc_db->md_depth;
1791 /* Named DBs also dirty the main DB */
1792 if (m0->mc_dbi > MAIN_DBI)
1793 i += txn->mt_dbs[MAIN_DBI].md_depth;
1794 /* For puts, roughly factor in the key+data size */
1796 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1797 i += i; /* double it for good measure */
1800 if (txn->mt_dirty_room > i)
1803 if (!txn->mt_spill_pgs) {
1804 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1805 if (!txn->mt_spill_pgs)
1808 /* purge deleted slots */
1809 MDB_IDL sl = txn->mt_spill_pgs;
1810 unsigned int num = sl[0];
1812 for (i=1; i<=num; i++) {
1819 /* Preserve pages which may soon be dirtied again */
1820 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1823 /* Less aggressive spill - we originally spilled the entire dirty list,
1824 * with a few exceptions for cursor pages and DB root pages. But this
1825 * turns out to be a lot of wasted effort because in a large txn many
1826 * of those pages will need to be used again. So now we spill only 1/8th
1827 * of the dirty pages. Testing revealed this to be a good tradeoff,
1828 * better than 1/2, 1/4, or 1/10.
1830 if (need < MDB_IDL_UM_MAX / 8)
1831 need = MDB_IDL_UM_MAX / 8;
1833 /* Save the page IDs of all the pages we're flushing */
1834 /* flush from the tail forward, this saves a lot of shifting later on. */
1835 for (i=dl[0].mid; i && need; i--) {
1836 MDB_ID pn = dl[i].mid << 1;
1838 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1840 /* Can't spill twice, make sure it's not already in a parent's
1843 if (txn->mt_parent) {
1845 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1846 if (tx2->mt_spill_pgs) {
1847 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1848 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1849 dp->mp_flags |= P_KEEP;
1857 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1861 mdb_midl_sort(txn->mt_spill_pgs);
1863 /* Flush the spilled part of dirty list */
1864 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1867 /* Reset any dirty pages we kept that page_flush didn't see */
1868 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1871 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1875 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1877 mdb_find_oldest(MDB_txn *txn)
1880 txnid_t mr, oldest = txn->mt_txnid - 1;
1881 if (txn->mt_env->me_txns) {
1882 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1883 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1894 /** Add a page to the txn's dirty list */
1896 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1899 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1901 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1902 insert = mdb_mid2l_append;
1904 insert = mdb_mid2l_insert;
1906 mid.mid = mp->mp_pgno;
1908 rc = insert(txn->mt_u.dirty_list, &mid);
1909 mdb_tassert(txn, rc == 0);
1910 txn->mt_dirty_room--;
1913 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1914 * me_pghead and mt_next_pgno.
1916 * If there are free pages available from older transactions, they
1917 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1918 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1919 * and move me_pglast to say which records were consumed. Only this
1920 * function can create me_pghead and move me_pglast/mt_next_pgno.
1921 * @param[in] mc cursor A cursor handle identifying the transaction and
1922 * database for which we are allocating.
1923 * @param[in] num the number of pages to allocate.
1924 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1925 * will always be satisfied by a single contiguous chunk of memory.
1926 * @return 0 on success, non-zero on failure.
1929 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1931 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1932 /* Get at most <Max_retries> more freeDB records once me_pghead
1933 * has enough pages. If not enough, use new pages from the map.
1934 * If <Paranoid> and mc is updating the freeDB, only get new
1935 * records if me_pghead is empty. Then the freelist cannot play
1936 * catch-up with itself by growing while trying to save it.
1938 enum { Paranoid = 1, Max_retries = 500 };
1940 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1942 int rc, retry = num * 20;
1943 MDB_txn *txn = mc->mc_txn;
1944 MDB_env *env = txn->mt_env;
1945 pgno_t pgno, *mop = env->me_pghead;
1946 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1948 txnid_t oldest = 0, last;
1952 /* If there are any loose pages, just use them */
1953 if (num == 1 && txn->mt_loose_pgs) {
1954 np = txn->mt_loose_pgs;
1955 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1956 txn->mt_loose_count--;
1957 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1965 /* If our dirty list is already full, we can't do anything */
1966 if (txn->mt_dirty_room == 0) {
1971 for (op = MDB_FIRST;; op = MDB_NEXT) {
1976 /* Seek a big enough contiguous page range. Prefer
1977 * pages at the tail, just truncating the list.
1983 if (mop[i-n2] == pgno+n2)
1990 if (op == MDB_FIRST) { /* 1st iteration */
1991 /* Prepare to fetch more and coalesce */
1992 oldest = mdb_find_oldest(txn);
1993 last = env->me_pglast;
1994 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1997 key.mv_data = &last; /* will look up last+1 */
1998 key.mv_size = sizeof(last);
2000 if (Paranoid && mc->mc_dbi == FREE_DBI)
2003 if (Paranoid && retry < 0 && mop_len)
2007 /* Do not fetch more if the record will be too recent */
2010 rc = mdb_cursor_get(&m2, &key, NULL, op);
2012 if (rc == MDB_NOTFOUND)
2016 last = *(txnid_t*)key.mv_data;
2019 np = m2.mc_pg[m2.mc_top];
2020 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2021 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2024 idl = (MDB_ID *) data.mv_data;
2027 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2032 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2034 mop = env->me_pghead;
2036 env->me_pglast = last;
2038 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2039 last, txn->mt_dbs[FREE_DBI].md_root, i));
2041 DPRINTF(("IDL %"Z"u", idl[j]));
2043 /* Merge in descending sorted order */
2044 mdb_midl_xmerge(mop, idl);
2048 /* Use new pages from the map when nothing suitable in the freeDB */
2050 pgno = txn->mt_next_pgno;
2051 if (pgno + num >= env->me_maxpg) {
2052 DPUTS("DB size maxed out");
2058 if (env->me_flags & MDB_WRITEMAP) {
2059 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2061 if (!(np = mdb_page_malloc(txn, num))) {
2067 mop[0] = mop_len -= num;
2068 /* Move any stragglers down */
2069 for (j = i-num; j < mop_len; )
2070 mop[++j] = mop[++i];
2072 txn->mt_next_pgno = pgno + num;
2075 mdb_page_dirty(txn, np);
2081 txn->mt_flags |= MDB_TXN_ERROR;
2085 /** Copy the used portions of a non-overflow page.
2086 * @param[in] dst page to copy into
2087 * @param[in] src page to copy from
2088 * @param[in] psize size of a page
2091 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2093 enum { Align = sizeof(pgno_t) };
2094 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2096 /* If page isn't full, just copy the used portion. Adjust
2097 * alignment so memcpy may copy words instead of bytes.
2099 if ((unused &= -Align) && !IS_LEAF2(src)) {
2100 upper = (upper + PAGEBASE) & -Align;
2101 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2102 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2105 memcpy(dst, src, psize - unused);
2109 /** Pull a page off the txn's spill list, if present.
2110 * If a page being referenced was spilled to disk in this txn, bring
2111 * it back and make it dirty/writable again.
2112 * @param[in] txn the transaction handle.
2113 * @param[in] mp the page being referenced. It must not be dirty.
2114 * @param[out] ret the writable page, if any. ret is unchanged if
2115 * mp wasn't spilled.
2118 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2120 MDB_env *env = txn->mt_env;
2123 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2125 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2126 if (!tx2->mt_spill_pgs)
2128 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2129 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2132 if (txn->mt_dirty_room == 0)
2133 return MDB_TXN_FULL;
2134 if (IS_OVERFLOW(mp))
2138 if (env->me_flags & MDB_WRITEMAP) {
2141 np = mdb_page_malloc(txn, num);
2145 memcpy(np, mp, num * env->me_psize);
2147 mdb_page_copy(np, mp, env->me_psize);
2150 /* If in current txn, this page is no longer spilled.
2151 * If it happens to be the last page, truncate the spill list.
2152 * Otherwise mark it as deleted by setting the LSB.
2154 if (x == txn->mt_spill_pgs[0])
2155 txn->mt_spill_pgs[0]--;
2157 txn->mt_spill_pgs[x] |= 1;
2158 } /* otherwise, if belonging to a parent txn, the
2159 * page remains spilled until child commits
2162 mdb_page_dirty(txn, np);
2163 np->mp_flags |= P_DIRTY;
2171 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2172 * @param[in] mc cursor pointing to the page to be touched
2173 * @return 0 on success, non-zero on failure.
2176 mdb_page_touch(MDB_cursor *mc)
2178 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2179 MDB_txn *txn = mc->mc_txn;
2180 MDB_cursor *m2, *m3;
2184 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2185 if (txn->mt_flags & MDB_TXN_SPILLS) {
2187 rc = mdb_page_unspill(txn, mp, &np);
2193 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2194 (rc = mdb_page_alloc(mc, 1, &np)))
2197 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2198 mp->mp_pgno, pgno));
2199 mdb_cassert(mc, mp->mp_pgno != pgno);
2200 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2201 /* Update the parent page, if any, to point to the new page */
2203 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2204 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2205 SETPGNO(node, pgno);
2207 mc->mc_db->md_root = pgno;
2209 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2210 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2212 /* If txn has a parent, make sure the page is in our
2216 unsigned x = mdb_mid2l_search(dl, pgno);
2217 if (x <= dl[0].mid && dl[x].mid == pgno) {
2218 if (mp != dl[x].mptr) { /* bad cursor? */
2219 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2220 txn->mt_flags |= MDB_TXN_ERROR;
2221 return MDB_CORRUPTED;
2226 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2228 np = mdb_page_malloc(txn, 1);
2233 rc = mdb_mid2l_insert(dl, &mid);
2234 mdb_cassert(mc, rc == 0);
2239 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2241 np->mp_flags |= P_DIRTY;
2244 /* Adjust cursors pointing to mp */
2245 mc->mc_pg[mc->mc_top] = np;
2246 m2 = txn->mt_cursors[mc->mc_dbi];
2247 if (mc->mc_flags & C_SUB) {
2248 for (; m2; m2=m2->mc_next) {
2249 m3 = &m2->mc_xcursor->mx_cursor;
2250 if (m3->mc_snum < mc->mc_snum) continue;
2251 if (m3->mc_pg[mc->mc_top] == mp)
2252 m3->mc_pg[mc->mc_top] = np;
2255 for (; m2; m2=m2->mc_next) {
2256 if (m2->mc_snum < mc->mc_snum) continue;
2257 if (m2->mc_pg[mc->mc_top] == mp) {
2258 m2->mc_pg[mc->mc_top] = np;
2259 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2261 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2263 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2264 if (!(leaf->mn_flags & F_SUBDATA))
2265 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2273 txn->mt_flags |= MDB_TXN_ERROR;
2278 mdb_env_sync(MDB_env *env, int force)
2281 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2282 if (env->me_flags & MDB_WRITEMAP) {
2283 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2284 ? MS_ASYNC : MS_SYNC;
2285 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2288 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2292 if (MDB_FDATASYNC(env->me_fd))
2299 /** Back up parent txn's cursors, then grab the originals for tracking */
2301 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2303 MDB_cursor *mc, *bk;
2308 for (i = src->mt_numdbs; --i >= 0; ) {
2309 if ((mc = src->mt_cursors[i]) != NULL) {
2310 size = sizeof(MDB_cursor);
2312 size += sizeof(MDB_xcursor);
2313 for (; mc; mc = bk->mc_next) {
2319 mc->mc_db = &dst->mt_dbs[i];
2320 /* Kill pointers into src - and dst to reduce abuse: The
2321 * user may not use mc until dst ends. Otherwise we'd...
2323 mc->mc_txn = NULL; /* ...set this to dst */
2324 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2325 if ((mx = mc->mc_xcursor) != NULL) {
2326 *(MDB_xcursor *)(bk+1) = *mx;
2327 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2329 mc->mc_next = dst->mt_cursors[i];
2330 dst->mt_cursors[i] = mc;
2337 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2338 * @param[in] txn the transaction handle.
2339 * @param[in] merge true to keep changes to parent cursors, false to revert.
2340 * @return 0 on success, non-zero on failure.
2343 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2345 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2349 for (i = txn->mt_numdbs; --i >= 0; ) {
2350 for (mc = cursors[i]; mc; mc = next) {
2352 if ((bk = mc->mc_backup) != NULL) {
2354 /* Commit changes to parent txn */
2355 mc->mc_next = bk->mc_next;
2356 mc->mc_backup = bk->mc_backup;
2357 mc->mc_txn = bk->mc_txn;
2358 mc->mc_db = bk->mc_db;
2359 mc->mc_dbflag = bk->mc_dbflag;
2360 if ((mx = mc->mc_xcursor) != NULL)
2361 mx->mx_cursor.mc_txn = bk->mc_txn;
2363 /* Abort nested txn */
2365 if ((mx = mc->mc_xcursor) != NULL)
2366 *mx = *(MDB_xcursor *)(bk+1);
2370 /* Only malloced cursors are permanently tracked. */
2378 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2381 mdb_txn_reset0(MDB_txn *txn, const char *act);
2383 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2389 Pidset = F_SETLK, Pidcheck = F_GETLK
2393 /** Set or check a pid lock. Set returns 0 on success.
2394 * Check returns 0 if the process is certainly dead, nonzero if it may
2395 * be alive (the lock exists or an error happened so we do not know).
2397 * On Windows Pidset is a no-op, we merely check for the existence
2398 * of the process with the given pid. On POSIX we use a single byte
2399 * lock on the lockfile, set at an offset equal to the pid.
2402 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2404 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2407 if (op == Pidcheck) {
2408 h = OpenProcess(env->me_pidquery, FALSE, pid);
2409 /* No documented "no such process" code, but other program use this: */
2411 return ErrCode() != ERROR_INVALID_PARAMETER;
2412 /* A process exists until all handles to it close. Has it exited? */
2413 ret = WaitForSingleObject(h, 0) != 0;
2420 struct flock lock_info;
2421 memset(&lock_info, 0, sizeof(lock_info));
2422 lock_info.l_type = F_WRLCK;
2423 lock_info.l_whence = SEEK_SET;
2424 lock_info.l_start = pid;
2425 lock_info.l_len = 1;
2426 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2427 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2429 } else if ((rc = ErrCode()) == EINTR) {
2437 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2438 * @param[in] txn the transaction handle to initialize
2439 * @return 0 on success, non-zero on failure.
2442 mdb_txn_renew0(MDB_txn *txn)
2444 MDB_env *env = txn->mt_env;
2445 MDB_txninfo *ti = env->me_txns;
2449 int rc, new_notls = 0;
2452 txn->mt_numdbs = env->me_numdbs;
2453 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2455 if (txn->mt_flags & MDB_TXN_RDONLY) {
2457 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2458 txn->mt_txnid = meta->mm_txnid;
2459 txn->mt_u.reader = NULL;
2461 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2462 pthread_getspecific(env->me_txkey);
2464 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2465 return MDB_BAD_RSLOT;
2467 MDB_PID_T pid = env->me_pid;
2468 MDB_THR_T tid = pthread_self();
2470 if (!env->me_live_reader) {
2471 rc = mdb_reader_pid(env, Pidset, pid);
2474 env->me_live_reader = 1;
2478 nr = ti->mti_numreaders;
2479 for (i=0; i<nr; i++)
2480 if (ti->mti_readers[i].mr_pid == 0)
2482 if (i == env->me_maxreaders) {
2483 UNLOCK_MUTEX_R(env);
2484 return MDB_READERS_FULL;
2486 ti->mti_readers[i].mr_pid = pid;
2487 ti->mti_readers[i].mr_tid = tid;
2489 ti->mti_numreaders = ++nr;
2490 /* Save numreaders for un-mutexed mdb_env_close() */
2491 env->me_numreaders = nr;
2492 UNLOCK_MUTEX_R(env);
2494 r = &ti->mti_readers[i];
2495 new_notls = (env->me_flags & MDB_NOTLS);
2496 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2501 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2502 txn->mt_u.reader = r;
2503 meta = env->me_metas[txn->mt_txnid & 1];
2509 txn->mt_txnid = ti->mti_txnid;
2510 meta = env->me_metas[txn->mt_txnid & 1];
2512 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2513 txn->mt_txnid = meta->mm_txnid;
2517 if (txn->mt_txnid == mdb_debug_start)
2520 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2521 txn->mt_u.dirty_list = env->me_dirty_list;
2522 txn->mt_u.dirty_list[0].mid = 0;
2523 txn->mt_free_pgs = env->me_free_pgs;
2524 txn->mt_free_pgs[0] = 0;
2525 txn->mt_spill_pgs = NULL;
2527 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2530 /* Copy the DB info and flags */
2531 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2533 /* Moved to here to avoid a data race in read TXNs */
2534 txn->mt_next_pgno = meta->mm_last_pg+1;
2536 for (i=2; i<txn->mt_numdbs; i++) {
2537 x = env->me_dbflags[i];
2538 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2539 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2541 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2543 if (env->me_maxpg < txn->mt_next_pgno) {
2544 mdb_txn_reset0(txn, "renew0-mapfail");
2546 txn->mt_u.reader->mr_pid = 0;
2547 txn->mt_u.reader = NULL;
2549 return MDB_MAP_RESIZED;
2556 mdb_txn_renew(MDB_txn *txn)
2560 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2563 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2564 DPUTS("environment had fatal error, must shutdown!");
2568 rc = mdb_txn_renew0(txn);
2569 if (rc == MDB_SUCCESS) {
2570 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2571 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2572 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2578 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2582 int rc, size, tsize = sizeof(MDB_txn);
2584 if (env->me_flags & MDB_FATAL_ERROR) {
2585 DPUTS("environment had fatal error, must shutdown!");
2588 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2591 /* Nested transactions: Max 1 child, write txns only, no writemap */
2592 if (parent->mt_child ||
2593 (flags & MDB_RDONLY) ||
2594 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2595 (env->me_flags & MDB_WRITEMAP))
2597 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2599 tsize = sizeof(MDB_ntxn);
2601 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2602 if (!(flags & MDB_RDONLY)) {
2603 size += env->me_maxdbs * sizeof(MDB_cursor *);
2604 /* child txns use parent's dbiseqs */
2606 size += env->me_maxdbs * sizeof(unsigned int);
2609 if ((txn = calloc(1, size)) == NULL) {
2610 DPRINTF(("calloc: %s", strerror(errno)));
2613 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2614 if (flags & MDB_RDONLY) {
2615 txn->mt_flags |= MDB_TXN_RDONLY;
2616 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2617 txn->mt_dbiseqs = env->me_dbiseqs;
2619 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2621 txn->mt_dbiseqs = parent->mt_dbiseqs;
2622 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2624 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2625 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2632 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2633 if (!txn->mt_u.dirty_list ||
2634 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2636 free(txn->mt_u.dirty_list);
2640 txn->mt_txnid = parent->mt_txnid;
2641 txn->mt_dirty_room = parent->mt_dirty_room;
2642 txn->mt_u.dirty_list[0].mid = 0;
2643 txn->mt_spill_pgs = NULL;
2644 txn->mt_next_pgno = parent->mt_next_pgno;
2645 parent->mt_child = txn;
2646 txn->mt_parent = parent;
2647 txn->mt_numdbs = parent->mt_numdbs;
2648 txn->mt_flags = parent->mt_flags;
2649 txn->mt_dbxs = parent->mt_dbxs;
2650 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2651 /* Copy parent's mt_dbflags, but clear DB_NEW */
2652 for (i=0; i<txn->mt_numdbs; i++)
2653 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2655 ntxn = (MDB_ntxn *)txn;
2656 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2657 if (env->me_pghead) {
2658 size = MDB_IDL_SIZEOF(env->me_pghead);
2659 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2661 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2666 rc = mdb_cursor_shadow(parent, txn);
2668 mdb_txn_reset0(txn, "beginchild-fail");
2670 rc = mdb_txn_renew0(txn);
2676 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2677 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2678 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2685 mdb_txn_env(MDB_txn *txn)
2687 if(!txn) return NULL;
2691 /** Export or close DBI handles opened in this txn. */
2693 mdb_dbis_update(MDB_txn *txn, int keep)
2696 MDB_dbi n = txn->mt_numdbs;
2697 MDB_env *env = txn->mt_env;
2698 unsigned char *tdbflags = txn->mt_dbflags;
2700 for (i = n; --i >= 2;) {
2701 if (tdbflags[i] & DB_NEW) {
2703 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2705 char *ptr = env->me_dbxs[i].md_name.mv_data;
2707 env->me_dbxs[i].md_name.mv_data = NULL;
2708 env->me_dbxs[i].md_name.mv_size = 0;
2709 env->me_dbflags[i] = 0;
2710 env->me_dbiseqs[i]++;
2716 if (keep && env->me_numdbs < n)
2720 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2721 * May be called twice for readonly txns: First reset it, then abort.
2722 * @param[in] txn the transaction handle to reset
2723 * @param[in] act why the transaction is being reset
2726 mdb_txn_reset0(MDB_txn *txn, const char *act)
2728 MDB_env *env = txn->mt_env;
2730 /* Close any DBI handles opened in this txn */
2731 mdb_dbis_update(txn, 0);
2733 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2734 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2735 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2737 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2738 if (txn->mt_u.reader) {
2739 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2740 if (!(env->me_flags & MDB_NOTLS))
2741 txn->mt_u.reader = NULL; /* txn does not own reader */
2743 txn->mt_numdbs = 0; /* close nothing if called again */
2744 txn->mt_dbxs = NULL; /* mark txn as reset */
2746 mdb_cursors_close(txn, 0);
2748 if (!(env->me_flags & MDB_WRITEMAP)) {
2749 mdb_dlist_free(txn);
2751 mdb_midl_free(env->me_pghead);
2753 if (txn->mt_parent) {
2754 txn->mt_parent->mt_child = NULL;
2755 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2756 mdb_midl_free(txn->mt_free_pgs);
2757 mdb_midl_free(txn->mt_spill_pgs);
2758 free(txn->mt_u.dirty_list);
2762 if (mdb_midl_shrink(&txn->mt_free_pgs))
2763 env->me_free_pgs = txn->mt_free_pgs;
2764 env->me_pghead = NULL;
2768 /* The writer mutex was locked in mdb_txn_begin. */
2770 UNLOCK_MUTEX_W(env);
2775 mdb_txn_reset(MDB_txn *txn)
2780 /* This call is only valid for read-only txns */
2781 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2784 mdb_txn_reset0(txn, "reset");
2788 mdb_txn_abort(MDB_txn *txn)
2794 mdb_txn_abort(txn->mt_child);
2796 mdb_txn_reset0(txn, "abort");
2797 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2798 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2799 txn->mt_u.reader->mr_pid = 0;
2804 /** Save the freelist as of this transaction to the freeDB.
2805 * This changes the freelist. Keep trying until it stabilizes.
2808 mdb_freelist_save(MDB_txn *txn)
2810 /* env->me_pghead[] can grow and shrink during this call.
2811 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2812 * Page numbers cannot disappear from txn->mt_free_pgs[].
2815 MDB_env *env = txn->mt_env;
2816 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2817 txnid_t pglast = 0, head_id = 0;
2818 pgno_t freecnt = 0, *free_pgs, *mop;
2819 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2821 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2823 if (env->me_pghead) {
2824 /* Make sure first page of freeDB is touched and on freelist */
2825 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2826 if (rc && rc != MDB_NOTFOUND)
2830 if (!env->me_pghead) {
2831 /* Put loose page numbers in mt_free_pgs, since
2832 * we may be unable to return them to me_pghead.
2834 MDB_page *mp = txn->mt_loose_pgs;
2835 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2837 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2838 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2839 txn->mt_loose_pgs = NULL;
2840 txn->mt_loose_count = 0;
2843 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2844 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2845 ? SSIZE_MAX : maxfree_1pg;
2848 /* Come back here after each Put() in case freelist changed */
2853 /* If using records from freeDB which we have not yet
2854 * deleted, delete them and any we reserved for me_pghead.
2856 while (pglast < env->me_pglast) {
2857 rc = mdb_cursor_first(&mc, &key, NULL);
2860 pglast = head_id = *(txnid_t *)key.mv_data;
2861 total_room = head_room = 0;
2862 mdb_tassert(txn, pglast <= env->me_pglast);
2863 rc = mdb_cursor_del(&mc, 0);
2868 /* Save the IDL of pages freed by this txn, to a single record */
2869 if (freecnt < txn->mt_free_pgs[0]) {
2871 /* Make sure last page of freeDB is touched and on freelist */
2872 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2873 if (rc && rc != MDB_NOTFOUND)
2876 free_pgs = txn->mt_free_pgs;
2877 /* Write to last page of freeDB */
2878 key.mv_size = sizeof(txn->mt_txnid);
2879 key.mv_data = &txn->mt_txnid;
2881 freecnt = free_pgs[0];
2882 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2883 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2886 /* Retry if mt_free_pgs[] grew during the Put() */
2887 free_pgs = txn->mt_free_pgs;
2888 } while (freecnt < free_pgs[0]);
2889 mdb_midl_sort(free_pgs);
2890 memcpy(data.mv_data, free_pgs, data.mv_size);
2893 unsigned int i = free_pgs[0];
2894 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2895 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2897 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2903 mop = env->me_pghead;
2904 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2906 /* Reserve records for me_pghead[]. Split it if multi-page,
2907 * to avoid searching freeDB for a page range. Use keys in
2908 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2910 if (total_room >= mop_len) {
2911 if (total_room == mop_len || --more < 0)
2913 } else if (head_room >= maxfree_1pg && head_id > 1) {
2914 /* Keep current record (overflow page), add a new one */
2918 /* (Re)write {key = head_id, IDL length = head_room} */
2919 total_room -= head_room;
2920 head_room = mop_len - total_room;
2921 if (head_room > maxfree_1pg && head_id > 1) {
2922 /* Overflow multi-page for part of me_pghead */
2923 head_room /= head_id; /* amortize page sizes */
2924 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2925 } else if (head_room < 0) {
2926 /* Rare case, not bothering to delete this record */
2929 key.mv_size = sizeof(head_id);
2930 key.mv_data = &head_id;
2931 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2932 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2935 /* IDL is initially empty, zero out at least the length */
2936 pgs = (pgno_t *)data.mv_data;
2937 j = head_room > clean_limit ? head_room : 0;
2941 total_room += head_room;
2944 /* Return loose page numbers to me_pghead, though usually none are
2945 * left at this point. The pages themselves remain in dirty_list.
2947 if (txn->mt_loose_pgs) {
2948 MDB_page *mp = txn->mt_loose_pgs;
2949 unsigned count = txn->mt_loose_count;
2951 /* Room for loose pages + temp IDL with same */
2952 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
2954 mop = env->me_pghead;
2955 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
2956 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
2957 loose[ ++count ] = mp->mp_pgno;
2959 mdb_midl_sort(loose);
2960 mdb_midl_xmerge(mop, loose);
2961 txn->mt_loose_pgs = NULL;
2962 txn->mt_loose_count = 0;
2966 /* Fill in the reserved me_pghead records */
2972 rc = mdb_cursor_first(&mc, &key, &data);
2973 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2974 txnid_t id = *(txnid_t *)key.mv_data;
2975 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2978 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2980 if (len > mop_len) {
2982 data.mv_size = (len + 1) * sizeof(MDB_ID);
2984 data.mv_data = mop -= len;
2987 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2989 if (rc || !(mop_len -= len))
2996 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2997 * @param[in] txn the transaction that's being committed
2998 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2999 * @return 0 on success, non-zero on failure.
3002 mdb_page_flush(MDB_txn *txn, int keep)
3004 MDB_env *env = txn->mt_env;
3005 MDB_ID2L dl = txn->mt_u.dirty_list;
3006 unsigned psize = env->me_psize, j;
3007 int i, pagecount = dl[0].mid, rc;
3008 size_t size = 0, pos = 0;
3010 MDB_page *dp = NULL;
3014 struct iovec iov[MDB_COMMIT_PAGES];
3015 ssize_t wpos = 0, wsize = 0, wres;
3016 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3022 if (env->me_flags & MDB_WRITEMAP) {
3023 /* Clear dirty flags */
3024 while (++i <= pagecount) {
3026 /* Don't flush this page yet */
3027 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3028 dp->mp_flags &= ~P_KEEP;
3032 dp->mp_flags &= ~P_DIRTY;
3037 /* Write the pages */
3039 if (++i <= pagecount) {
3041 /* Don't flush this page yet */
3042 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3043 dp->mp_flags &= ~P_KEEP;
3048 /* clear dirty flag */
3049 dp->mp_flags &= ~P_DIRTY;
3052 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3057 /* Windows actually supports scatter/gather I/O, but only on
3058 * unbuffered file handles. Since we're relying on the OS page
3059 * cache for all our data, that's self-defeating. So we just
3060 * write pages one at a time. We use the ov structure to set
3061 * the write offset, to at least save the overhead of a Seek
3064 DPRINTF(("committing page %"Z"u", pgno));
3065 memset(&ov, 0, sizeof(ov));
3066 ov.Offset = pos & 0xffffffff;
3067 ov.OffsetHigh = pos >> 16 >> 16;
3068 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3070 DPRINTF(("WriteFile: %d", rc));
3074 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3075 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3077 /* Write previous page(s) */
3078 #ifdef MDB_USE_PWRITEV
3079 wres = pwritev(env->me_fd, iov, n, wpos);
3082 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3084 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3086 DPRINTF(("lseek: %s", strerror(rc)));
3089 wres = writev(env->me_fd, iov, n);
3092 if (wres != wsize) {
3095 DPRINTF(("Write error: %s", strerror(rc)));
3097 rc = EIO; /* TODO: Use which error code? */
3098 DPUTS("short write, filesystem full?");
3109 DPRINTF(("committing page %"Z"u", pgno));
3110 next_pos = pos + size;
3111 iov[n].iov_len = size;
3112 iov[n].iov_base = (char *)dp;
3118 /* MIPS has cache coherency issues, this is a no-op everywhere else
3119 * Note: for any size >= on-chip cache size, entire on-chip cache is
3122 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3124 for (i = keep; ++i <= pagecount; ) {
3126 /* This is a page we skipped above */
3129 dl[j].mid = dp->mp_pgno;
3132 mdb_dpage_free(env, dp);
3137 txn->mt_dirty_room += i - j;
3143 mdb_txn_commit(MDB_txn *txn)
3149 if (txn == NULL || txn->mt_env == NULL)
3152 if (txn->mt_child) {
3153 rc = mdb_txn_commit(txn->mt_child);
3154 txn->mt_child = NULL;
3161 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3162 mdb_dbis_update(txn, 1);
3163 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3168 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3169 DPUTS("error flag is set, can't commit");
3171 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3176 if (txn->mt_parent) {
3177 MDB_txn *parent = txn->mt_parent;
3181 unsigned x, y, len, ps_len;
3183 /* Append our free list to parent's */
3184 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3187 mdb_midl_free(txn->mt_free_pgs);
3188 /* Failures after this must either undo the changes
3189 * to the parent or set MDB_TXN_ERROR in the parent.
3192 parent->mt_next_pgno = txn->mt_next_pgno;
3193 parent->mt_flags = txn->mt_flags;
3195 /* Merge our cursors into parent's and close them */
3196 mdb_cursors_close(txn, 1);
3198 /* Update parent's DB table. */
3199 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3200 parent->mt_numdbs = txn->mt_numdbs;
3201 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3202 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3203 for (i=2; i<txn->mt_numdbs; i++) {
3204 /* preserve parent's DB_NEW status */
3205 x = parent->mt_dbflags[i] & DB_NEW;
3206 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3209 dst = parent->mt_u.dirty_list;
3210 src = txn->mt_u.dirty_list;
3211 /* Remove anything in our dirty list from parent's spill list */
3212 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3214 pspill[0] = (pgno_t)-1;
3215 /* Mark our dirty pages as deleted in parent spill list */
3216 for (i=0, len=src[0].mid; ++i <= len; ) {
3217 MDB_ID pn = src[i].mid << 1;
3218 while (pn > pspill[x])
3220 if (pn == pspill[x]) {
3225 /* Squash deleted pagenums if we deleted any */
3226 for (x=y; ++x <= ps_len; )
3227 if (!(pspill[x] & 1))
3228 pspill[++y] = pspill[x];
3232 /* Find len = length of merging our dirty list with parent's */
3234 dst[0].mid = 0; /* simplify loops */
3235 if (parent->mt_parent) {
3236 len = x + src[0].mid;
3237 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3238 for (i = x; y && i; y--) {
3239 pgno_t yp = src[y].mid;
3240 while (yp < dst[i].mid)
3242 if (yp == dst[i].mid) {
3247 } else { /* Simplify the above for single-ancestor case */
3248 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3250 /* Merge our dirty list with parent's */
3252 for (i = len; y; dst[i--] = src[y--]) {
3253 pgno_t yp = src[y].mid;
3254 while (yp < dst[x].mid)
3255 dst[i--] = dst[x--];
3256 if (yp == dst[x].mid)
3257 free(dst[x--].mptr);
3259 mdb_tassert(txn, i == x);
3261 free(txn->mt_u.dirty_list);
3262 parent->mt_dirty_room = txn->mt_dirty_room;
3263 if (txn->mt_spill_pgs) {
3264 if (parent->mt_spill_pgs) {
3265 /* TODO: Prevent failure here, so parent does not fail */
3266 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3268 parent->mt_flags |= MDB_TXN_ERROR;
3269 mdb_midl_free(txn->mt_spill_pgs);
3270 mdb_midl_sort(parent->mt_spill_pgs);
3272 parent->mt_spill_pgs = txn->mt_spill_pgs;
3276 /* Append our loose page list to parent's */
3277 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3279 *lp = txn->mt_loose_pgs;
3280 parent->mt_loose_count += txn->mt_loose_count;
3282 parent->mt_child = NULL;
3283 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3288 if (txn != env->me_txn) {
3289 DPUTS("attempt to commit unknown transaction");
3294 mdb_cursors_close(txn, 0);
3296 if (!txn->mt_u.dirty_list[0].mid &&
3297 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3300 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3301 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3303 /* Update DB root pointers */
3304 if (txn->mt_numdbs > 2) {
3308 data.mv_size = sizeof(MDB_db);
3310 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3311 for (i = 2; i < txn->mt_numdbs; i++) {
3312 if (txn->mt_dbflags[i] & DB_DIRTY) {
3313 if (TXN_DBI_CHANGED(txn, i)) {
3317 data.mv_data = &txn->mt_dbs[i];
3318 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3325 rc = mdb_freelist_save(txn);
3329 mdb_midl_free(env->me_pghead);
3330 env->me_pghead = NULL;
3331 if (mdb_midl_shrink(&txn->mt_free_pgs))
3332 env->me_free_pgs = txn->mt_free_pgs;
3338 if ((rc = mdb_page_flush(txn, 0)) ||
3339 (rc = mdb_env_sync(env, 0)) ||
3340 (rc = mdb_env_write_meta(txn)))
3343 /* Free P_LOOSE pages left behind in dirty_list */
3344 if (!(env->me_flags & MDB_WRITEMAP))
3345 mdb_dlist_free(txn);
3350 mdb_dbis_update(txn, 1);
3353 UNLOCK_MUTEX_W(env);
3363 /** Read the environment parameters of a DB environment before
3364 * mapping it into memory.
3365 * @param[in] env the environment handle
3366 * @param[out] meta address of where to store the meta information
3367 * @return 0 on success, non-zero on failure.
3370 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3376 enum { Size = sizeof(pbuf) };
3378 /* We don't know the page size yet, so use a minimum value.
3379 * Read both meta pages so we can use the latest one.
3382 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3386 memset(&ov, 0, sizeof(ov));
3388 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3389 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3392 rc = pread(env->me_fd, &pbuf, Size, off);
3395 if (rc == 0 && off == 0)
3397 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3398 DPRINTF(("read: %s", mdb_strerror(rc)));
3402 p = (MDB_page *)&pbuf;
3404 if (!F_ISSET(p->mp_flags, P_META)) {
3405 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3410 if (m->mm_magic != MDB_MAGIC) {
3411 DPUTS("meta has invalid magic");
3415 if (m->mm_version != MDB_DATA_VERSION) {
3416 DPRINTF(("database is version %u, expected version %u",
3417 m->mm_version, MDB_DATA_VERSION));
3418 return MDB_VERSION_MISMATCH;
3421 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3428 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3430 meta->mm_magic = MDB_MAGIC;
3431 meta->mm_version = MDB_DATA_VERSION;
3432 meta->mm_mapsize = env->me_mapsize;
3433 meta->mm_psize = env->me_psize;
3434 meta->mm_last_pg = 1;
3435 meta->mm_flags = env->me_flags & 0xffff;
3436 meta->mm_flags |= MDB_INTEGERKEY;
3437 meta->mm_dbs[0].md_root = P_INVALID;
3438 meta->mm_dbs[1].md_root = P_INVALID;
3441 /** Write the environment parameters of a freshly created DB environment.
3442 * @param[in] env the environment handle
3443 * @param[out] meta address of where to store the meta information
3444 * @return 0 on success, non-zero on failure.
3447 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3455 memset(&ov, 0, sizeof(ov));
3456 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3458 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3461 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3462 len = pwrite(fd, ptr, size, pos); \
3463 rc = (len >= 0); } while(0)
3466 DPUTS("writing new meta page");
3468 psize = env->me_psize;
3470 mdb_env_init_meta0(env, meta);
3472 p = calloc(2, psize);
3474 p->mp_flags = P_META;
3475 *(MDB_meta *)METADATA(p) = *meta;
3477 q = (MDB_page *)((char *)p + psize);
3479 q->mp_flags = P_META;
3480 *(MDB_meta *)METADATA(q) = *meta;
3482 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3485 else if ((unsigned) len == psize * 2)
3493 /** Update the environment info to commit a transaction.
3494 * @param[in] txn the transaction that's being committed
3495 * @return 0 on success, non-zero on failure.
3498 mdb_env_write_meta(MDB_txn *txn)
3501 MDB_meta meta, metab, *mp;
3504 int rc, len, toggle;
3513 toggle = txn->mt_txnid & 1;
3514 DPRINTF(("writing meta page %d for root page %"Z"u",
3515 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3518 mp = env->me_metas[toggle];
3519 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3520 /* Persist any increases of mapsize config */
3521 if (mapsize < env->me_mapsize)
3522 mapsize = env->me_mapsize;
3524 if (env->me_flags & MDB_WRITEMAP) {
3525 mp->mm_mapsize = mapsize;
3526 mp->mm_dbs[0] = txn->mt_dbs[0];
3527 mp->mm_dbs[1] = txn->mt_dbs[1];
3528 mp->mm_last_pg = txn->mt_next_pgno - 1;
3529 mp->mm_txnid = txn->mt_txnid;
3530 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3531 unsigned meta_size = env->me_psize;
3532 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3535 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3536 if (meta_size < env->me_os_psize)
3537 meta_size += meta_size;
3542 if (MDB_MSYNC(ptr, meta_size, rc)) {
3549 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3550 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3552 meta.mm_mapsize = mapsize;
3553 meta.mm_dbs[0] = txn->mt_dbs[0];
3554 meta.mm_dbs[1] = txn->mt_dbs[1];
3555 meta.mm_last_pg = txn->mt_next_pgno - 1;
3556 meta.mm_txnid = txn->mt_txnid;
3558 off = offsetof(MDB_meta, mm_mapsize);
3559 ptr = (char *)&meta + off;
3560 len = sizeof(MDB_meta) - off;
3562 off += env->me_psize;
3565 /* Write to the SYNC fd */
3566 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3567 env->me_fd : env->me_mfd;
3570 memset(&ov, 0, sizeof(ov));
3572 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3576 rc = pwrite(mfd, ptr, len, off);
3579 rc = rc < 0 ? ErrCode() : EIO;
3580 DPUTS("write failed, disk error?");
3581 /* On a failure, the pagecache still contains the new data.
3582 * Write some old data back, to prevent it from being used.
3583 * Use the non-SYNC fd; we know it will fail anyway.
3585 meta.mm_last_pg = metab.mm_last_pg;
3586 meta.mm_txnid = metab.mm_txnid;
3588 memset(&ov, 0, sizeof(ov));
3590 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3592 r2 = pwrite(env->me_fd, ptr, len, off);
3593 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3596 env->me_flags |= MDB_FATAL_ERROR;
3600 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3601 if (!(env->me_flags & MDB_WRITEMAP)) {
3602 CACHEFLUSH(env->me_map + off, len, DCACHE);
3604 /* Memory ordering issues are irrelevant; since the entire writer
3605 * is wrapped by wmutex, all of these changes will become visible
3606 * after the wmutex is unlocked. Since the DB is multi-version,
3607 * readers will get consistent data regardless of how fresh or
3608 * how stale their view of these values is.
3611 env->me_txns->mti_txnid = txn->mt_txnid;
3616 /** Check both meta pages to see which one is newer.
3617 * @param[in] env the environment handle
3618 * @return meta toggle (0 or 1).
3621 mdb_env_pick_meta(const MDB_env *env)
3623 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3627 mdb_env_create(MDB_env **env)
3631 e = calloc(1, sizeof(MDB_env));
3635 e->me_maxreaders = DEFAULT_READERS;
3636 e->me_maxdbs = e->me_numdbs = 2;
3637 e->me_fd = INVALID_HANDLE_VALUE;
3638 e->me_lfd = INVALID_HANDLE_VALUE;
3639 e->me_mfd = INVALID_HANDLE_VALUE;
3640 #ifdef MDB_USE_POSIX_SEM
3641 e->me_rmutex = SEM_FAILED;
3642 e->me_wmutex = SEM_FAILED;
3644 e->me_pid = getpid();
3645 GET_PAGESIZE(e->me_os_psize);
3646 VGMEMP_CREATE(e,0,0);
3652 mdb_env_map(MDB_env *env, void *addr)
3655 unsigned int flags = env->me_flags;
3659 LONG sizelo, sizehi;
3662 if (flags & MDB_RDONLY) {
3663 /* Don't set explicit map size, use whatever exists */
3668 msize = env->me_mapsize;
3669 sizelo = msize & 0xffffffff;
3670 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3672 /* Windows won't create mappings for zero length files.
3673 * and won't map more than the file size.
3674 * Just set the maxsize right now.
3676 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3677 || !SetEndOfFile(env->me_fd)
3678 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3682 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3683 PAGE_READWRITE : PAGE_READONLY,
3684 sizehi, sizelo, NULL);
3687 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3688 FILE_MAP_WRITE : FILE_MAP_READ,
3690 rc = env->me_map ? 0 : ErrCode();
3695 int prot = PROT_READ;
3696 if (flags & MDB_WRITEMAP) {
3698 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3701 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3703 if (env->me_map == MAP_FAILED) {
3708 if (flags & MDB_NORDAHEAD) {
3709 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3711 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3713 #ifdef POSIX_MADV_RANDOM
3714 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3715 #endif /* POSIX_MADV_RANDOM */
3716 #endif /* MADV_RANDOM */
3720 /* Can happen because the address argument to mmap() is just a
3721 * hint. mmap() can pick another, e.g. if the range is in use.
3722 * The MAP_FIXED flag would prevent that, but then mmap could
3723 * instead unmap existing pages to make room for the new map.
3725 if (addr && env->me_map != addr)
3726 return EBUSY; /* TODO: Make a new MDB_* error code? */
3728 p = (MDB_page *)env->me_map;
3729 env->me_metas[0] = METADATA(p);
3730 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3736 mdb_env_set_mapsize(MDB_env *env, size_t size)
3738 /* If env is already open, caller is responsible for making
3739 * sure there are no active txns.
3747 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3748 else if (size < env->me_mapsize) {
3749 /* If the configured size is smaller, make sure it's
3750 * still big enough. Silently round up to minimum if not.
3752 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3756 munmap(env->me_map, env->me_mapsize);
3757 env->me_mapsize = size;
3758 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3759 rc = mdb_env_map(env, old);
3763 env->me_mapsize = size;
3765 env->me_maxpg = env->me_mapsize / env->me_psize;
3770 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3774 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3779 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3781 if (env->me_map || readers < 1)
3783 env->me_maxreaders = readers;
3788 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3790 if (!env || !readers)
3792 *readers = env->me_maxreaders;
3796 /** Further setup required for opening an LMDB environment
3799 mdb_env_open2(MDB_env *env)
3801 unsigned int flags = env->me_flags;
3802 int i, newenv = 0, rc;
3806 /* See if we should use QueryLimited */
3808 if ((rc & 0xff) > 5)
3809 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3811 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3814 memset(&meta, 0, sizeof(meta));
3816 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3819 DPUTS("new mdbenv");
3821 env->me_psize = env->me_os_psize;
3822 if (env->me_psize > MAX_PAGESIZE)
3823 env->me_psize = MAX_PAGESIZE;
3825 env->me_psize = meta.mm_psize;
3828 /* Was a mapsize configured? */
3829 if (!env->me_mapsize) {
3830 /* If this is a new environment, take the default,
3831 * else use the size recorded in the existing env.
3833 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3834 } else if (env->me_mapsize < meta.mm_mapsize) {
3835 /* If the configured size is smaller, make sure it's
3836 * still big enough. Silently round up to minimum if not.
3838 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3839 if (env->me_mapsize < minsize)
3840 env->me_mapsize = minsize;
3843 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3848 if (flags & MDB_FIXEDMAP)
3849 meta.mm_address = env->me_map;
3850 i = mdb_env_init_meta(env, &meta);
3851 if (i != MDB_SUCCESS) {
3856 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3857 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3859 #if !(MDB_MAXKEYSIZE)
3860 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3862 env->me_maxpg = env->me_mapsize / env->me_psize;
3866 int toggle = mdb_env_pick_meta(env);
3867 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3869 DPRINTF(("opened database version %u, pagesize %u",
3870 env->me_metas[0]->mm_version, env->me_psize));
3871 DPRINTF(("using meta page %d", toggle));
3872 DPRINTF(("depth: %u", db->md_depth));
3873 DPRINTF(("entries: %"Z"u", db->md_entries));
3874 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3875 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3876 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3877 DPRINTF(("root: %"Z"u", db->md_root));
3885 /** Release a reader thread's slot in the reader lock table.
3886 * This function is called automatically when a thread exits.
3887 * @param[in] ptr This points to the slot in the reader lock table.
3890 mdb_env_reader_dest(void *ptr)
3892 MDB_reader *reader = ptr;
3898 /** Junk for arranging thread-specific callbacks on Windows. This is
3899 * necessarily platform and compiler-specific. Windows supports up
3900 * to 1088 keys. Let's assume nobody opens more than 64 environments
3901 * in a single process, for now. They can override this if needed.
3903 #ifndef MAX_TLS_KEYS
3904 #define MAX_TLS_KEYS 64
3906 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3907 static int mdb_tls_nkeys;
3909 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3913 case DLL_PROCESS_ATTACH: break;
3914 case DLL_THREAD_ATTACH: break;
3915 case DLL_THREAD_DETACH:
3916 for (i=0; i<mdb_tls_nkeys; i++) {
3917 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3919 mdb_env_reader_dest(r);
3923 case DLL_PROCESS_DETACH: break;
3928 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3930 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3934 /* Force some symbol references.
3935 * _tls_used forces the linker to create the TLS directory if not already done
3936 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3938 #pragma comment(linker, "/INCLUDE:_tls_used")
3939 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3940 #pragma const_seg(".CRT$XLB")
3941 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3942 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3945 #pragma comment(linker, "/INCLUDE:__tls_used")
3946 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3947 #pragma data_seg(".CRT$XLB")
3948 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3950 #endif /* WIN 32/64 */
3951 #endif /* !__GNUC__ */
3954 /** Downgrade the exclusive lock on the region back to shared */
3956 mdb_env_share_locks(MDB_env *env, int *excl)
3958 int rc = 0, toggle = mdb_env_pick_meta(env);
3960 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3965 /* First acquire a shared lock. The Unlock will
3966 * then release the existing exclusive lock.
3968 memset(&ov, 0, sizeof(ov));
3969 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3972 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3978 struct flock lock_info;
3979 /* The shared lock replaces the existing lock */
3980 memset((void *)&lock_info, 0, sizeof(lock_info));
3981 lock_info.l_type = F_RDLCK;
3982 lock_info.l_whence = SEEK_SET;
3983 lock_info.l_start = 0;
3984 lock_info.l_len = 1;
3985 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3986 (rc = ErrCode()) == EINTR) ;
3987 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3994 /** Try to get exlusive lock, otherwise shared.
3995 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3998 mdb_env_excl_lock(MDB_env *env, int *excl)
4002 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4006 memset(&ov, 0, sizeof(ov));
4007 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4014 struct flock lock_info;
4015 memset((void *)&lock_info, 0, sizeof(lock_info));
4016 lock_info.l_type = F_WRLCK;
4017 lock_info.l_whence = SEEK_SET;
4018 lock_info.l_start = 0;
4019 lock_info.l_len = 1;
4020 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4021 (rc = ErrCode()) == EINTR) ;
4025 # ifdef MDB_USE_POSIX_SEM
4026 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4029 lock_info.l_type = F_RDLCK;
4030 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4031 (rc = ErrCode()) == EINTR) ;
4041 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4043 * @(#) $Revision: 5.1 $
4044 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4045 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4047 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4051 * Please do not copyright this code. This code is in the public domain.
4053 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4054 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4055 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4056 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4057 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4058 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4059 * PERFORMANCE OF THIS SOFTWARE.
4062 * chongo <Landon Curt Noll> /\oo/\
4063 * http://www.isthe.com/chongo/
4065 * Share and Enjoy! :-)
4068 typedef unsigned long long mdb_hash_t;
4069 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4071 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4072 * @param[in] val value to hash
4073 * @param[in] hval initial value for hash
4074 * @return 64 bit hash
4076 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4077 * hval arg on the first call.
4080 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4082 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4083 unsigned char *end = s + val->mv_size;
4085 * FNV-1a hash each octet of the string
4088 /* xor the bottom with the current octet */
4089 hval ^= (mdb_hash_t)*s++;
4091 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4092 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4093 (hval << 7) + (hval << 8) + (hval << 40);
4095 /* return our new hash value */
4099 /** Hash the string and output the encoded hash.
4100 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4101 * very short name limits. We don't care about the encoding being reversible,
4102 * we just want to preserve as many bits of the input as possible in a
4103 * small printable string.
4104 * @param[in] str string to hash
4105 * @param[out] encbuf an array of 11 chars to hold the hash
4107 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4110 mdb_pack85(unsigned long l, char *out)
4114 for (i=0; i<5; i++) {
4115 *out++ = mdb_a85[l % 85];
4121 mdb_hash_enc(MDB_val *val, char *encbuf)
4123 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4125 mdb_pack85(h, encbuf);
4126 mdb_pack85(h>>32, encbuf+5);
4131 /** Open and/or initialize the lock region for the environment.
4132 * @param[in] env The LMDB environment.
4133 * @param[in] lpath The pathname of the file used for the lock region.
4134 * @param[in] mode The Unix permissions for the file, if we create it.
4135 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4136 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4137 * @return 0 on success, non-zero on failure.
4140 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4143 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4145 # define MDB_ERRCODE_ROFS EROFS
4146 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4147 # define MDB_CLOEXEC O_CLOEXEC
4150 # define MDB_CLOEXEC 0
4157 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4158 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4159 FILE_ATTRIBUTE_NORMAL, NULL);
4161 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4163 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4165 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4170 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4171 /* Lose record locks when exec*() */
4172 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4173 fcntl(env->me_lfd, F_SETFD, fdflags);
4176 if (!(env->me_flags & MDB_NOTLS)) {
4177 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4180 env->me_flags |= MDB_ENV_TXKEY;
4182 /* Windows TLS callbacks need help finding their TLS info. */
4183 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4187 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4191 /* Try to get exclusive lock. If we succeed, then
4192 * nobody is using the lock region and we should initialize it.
4194 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4197 size = GetFileSize(env->me_lfd, NULL);
4199 size = lseek(env->me_lfd, 0, SEEK_END);
4200 if (size == -1) goto fail_errno;
4202 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4203 if (size < rsize && *excl > 0) {
4205 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4206 || !SetEndOfFile(env->me_lfd))
4209 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4213 size = rsize - sizeof(MDB_txninfo);
4214 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4219 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4221 if (!mh) goto fail_errno;
4222 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4224 if (!env->me_txns) goto fail_errno;
4226 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4228 if (m == MAP_FAILED) goto fail_errno;
4234 BY_HANDLE_FILE_INFORMATION stbuf;
4243 if (!mdb_sec_inited) {
4244 InitializeSecurityDescriptor(&mdb_null_sd,
4245 SECURITY_DESCRIPTOR_REVISION);
4246 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4247 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4248 mdb_all_sa.bInheritHandle = FALSE;
4249 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4252 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4253 idbuf.volume = stbuf.dwVolumeSerialNumber;
4254 idbuf.nhigh = stbuf.nFileIndexHigh;
4255 idbuf.nlow = stbuf.nFileIndexLow;
4256 val.mv_data = &idbuf;
4257 val.mv_size = sizeof(idbuf);
4258 mdb_hash_enc(&val, encbuf);
4259 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4260 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4261 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4262 if (!env->me_rmutex) goto fail_errno;
4263 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4264 if (!env->me_wmutex) goto fail_errno;
4265 #elif defined(MDB_USE_POSIX_SEM)
4274 #if defined(__NetBSD__)
4275 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4277 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4278 idbuf.dev = stbuf.st_dev;
4279 idbuf.ino = stbuf.st_ino;
4280 val.mv_data = &idbuf;
4281 val.mv_size = sizeof(idbuf);
4282 mdb_hash_enc(&val, encbuf);
4283 #ifdef MDB_SHORT_SEMNAMES
4284 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4286 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4287 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4288 /* Clean up after a previous run, if needed: Try to
4289 * remove both semaphores before doing anything else.
4291 sem_unlink(env->me_txns->mti_rmname);
4292 sem_unlink(env->me_txns->mti_wmname);
4293 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4294 O_CREAT|O_EXCL, mode, 1);
4295 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4296 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4297 O_CREAT|O_EXCL, mode, 1);
4298 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4299 #else /* MDB_USE_POSIX_SEM */
4300 pthread_mutexattr_t mattr;
4302 if ((rc = pthread_mutexattr_init(&mattr))
4303 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4304 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4305 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4307 pthread_mutexattr_destroy(&mattr);
4308 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4310 env->me_txns->mti_magic = MDB_MAGIC;
4311 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4312 env->me_txns->mti_txnid = 0;
4313 env->me_txns->mti_numreaders = 0;
4316 if (env->me_txns->mti_magic != MDB_MAGIC) {
4317 DPUTS("lock region has invalid magic");
4321 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4322 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4323 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4324 rc = MDB_VERSION_MISMATCH;
4328 if (rc && rc != EACCES && rc != EAGAIN) {
4332 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4333 if (!env->me_rmutex) goto fail_errno;
4334 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4335 if (!env->me_wmutex) goto fail_errno;
4336 #elif defined(MDB_USE_POSIX_SEM)
4337 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4338 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4339 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4340 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4351 /** The name of the lock file in the DB environment */
4352 #define LOCKNAME "/lock.mdb"
4353 /** The name of the data file in the DB environment */
4354 #define DATANAME "/data.mdb"
4355 /** The suffix of the lock file when no subdir is used */
4356 #define LOCKSUFF "-lock"
4357 /** Only a subset of the @ref mdb_env flags can be changed
4358 * at runtime. Changing other flags requires closing the
4359 * environment and re-opening it with the new flags.
4361 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4362 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4363 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4365 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4366 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4370 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4372 int oflags, rc, len, excl = -1;
4373 char *lpath, *dpath;
4375 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4379 if (flags & MDB_NOSUBDIR) {
4380 rc = len + sizeof(LOCKSUFF) + len + 1;
4382 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4387 if (flags & MDB_NOSUBDIR) {
4388 dpath = lpath + len + sizeof(LOCKSUFF);
4389 sprintf(lpath, "%s" LOCKSUFF, path);
4390 strcpy(dpath, path);
4392 dpath = lpath + len + sizeof(LOCKNAME);
4393 sprintf(lpath, "%s" LOCKNAME, path);
4394 sprintf(dpath, "%s" DATANAME, path);
4398 flags |= env->me_flags;
4399 if (flags & MDB_RDONLY) {
4400 /* silently ignore WRITEMAP when we're only getting read access */
4401 flags &= ~MDB_WRITEMAP;
4403 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4404 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4407 env->me_flags = flags |= MDB_ENV_ACTIVE;
4411 env->me_path = strdup(path);
4412 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4413 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4414 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4415 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4420 /* For RDONLY, get lockfile after we know datafile exists */
4421 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4422 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4428 if (F_ISSET(flags, MDB_RDONLY)) {
4429 oflags = GENERIC_READ;
4430 len = OPEN_EXISTING;
4432 oflags = GENERIC_READ|GENERIC_WRITE;
4435 mode = FILE_ATTRIBUTE_NORMAL;
4436 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4437 NULL, len, mode, NULL);
4439 if (F_ISSET(flags, MDB_RDONLY))
4442 oflags = O_RDWR | O_CREAT;
4444 env->me_fd = open(dpath, oflags, mode);
4446 if (env->me_fd == INVALID_HANDLE_VALUE) {
4451 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4452 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4457 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4458 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4459 env->me_mfd = env->me_fd;
4461 /* Synchronous fd for meta writes. Needed even with
4462 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4465 len = OPEN_EXISTING;
4466 env->me_mfd = CreateFile(dpath, oflags,
4467 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4468 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4471 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4473 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4478 DPRINTF(("opened dbenv %p", (void *) env));
4480 rc = mdb_env_share_locks(env, &excl);
4484 if (!((flags & MDB_RDONLY) ||
4485 (env->me_pbuf = calloc(1, env->me_psize))))
4491 mdb_env_close0(env, excl);
4497 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4499 mdb_env_close0(MDB_env *env, int excl)
4503 if (!(env->me_flags & MDB_ENV_ACTIVE))
4506 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4507 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4508 free(env->me_dbxs[i].md_name.mv_data);
4511 free(env->me_dbiseqs);
4512 free(env->me_dbflags);
4515 free(env->me_dirty_list);
4516 mdb_midl_free(env->me_free_pgs);
4518 if (env->me_flags & MDB_ENV_TXKEY) {
4519 pthread_key_delete(env->me_txkey);
4521 /* Delete our key from the global list */
4522 for (i=0; i<mdb_tls_nkeys; i++)
4523 if (mdb_tls_keys[i] == env->me_txkey) {
4524 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4532 munmap(env->me_map, env->me_mapsize);
4534 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4535 (void) close(env->me_mfd);
4536 if (env->me_fd != INVALID_HANDLE_VALUE)
4537 (void) close(env->me_fd);
4539 MDB_PID_T pid = env->me_pid;
4540 /* Clearing readers is done in this function because
4541 * me_txkey with its destructor must be disabled first.
4543 for (i = env->me_numreaders; --i >= 0; )
4544 if (env->me_txns->mti_readers[i].mr_pid == pid)
4545 env->me_txns->mti_readers[i].mr_pid = 0;
4547 if (env->me_rmutex) {
4548 CloseHandle(env->me_rmutex);
4549 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4551 /* Windows automatically destroys the mutexes when
4552 * the last handle closes.
4554 #elif defined(MDB_USE_POSIX_SEM)
4555 if (env->me_rmutex != SEM_FAILED) {
4556 sem_close(env->me_rmutex);
4557 if (env->me_wmutex != SEM_FAILED)
4558 sem_close(env->me_wmutex);
4559 /* If we have the filelock: If we are the
4560 * only remaining user, clean up semaphores.
4563 mdb_env_excl_lock(env, &excl);
4565 sem_unlink(env->me_txns->mti_rmname);
4566 sem_unlink(env->me_txns->mti_wmname);
4570 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4572 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4575 /* Unlock the lockfile. Windows would have unlocked it
4576 * after closing anyway, but not necessarily at once.
4578 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4581 (void) close(env->me_lfd);
4584 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4589 mdb_env_close(MDB_env *env)
4596 VGMEMP_DESTROY(env);
4597 while ((dp = env->me_dpages) != NULL) {
4598 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4599 env->me_dpages = dp->mp_next;
4603 mdb_env_close0(env, 0);
4607 /** Compare two items pointing at aligned size_t's */
4609 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4611 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4612 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4615 /** Compare two items pointing at aligned unsigned int's */
4617 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4619 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4620 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4623 /** Compare two items pointing at unsigned ints of unknown alignment.
4624 * Nodes and keys are guaranteed to be 2-byte aligned.
4627 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4629 #if BYTE_ORDER == LITTLE_ENDIAN
4630 unsigned short *u, *c;
4633 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4634 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4637 } while(!x && u > (unsigned short *)a->mv_data);
4640 unsigned short *u, *c, *end;
4643 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4644 u = (unsigned short *)a->mv_data;
4645 c = (unsigned short *)b->mv_data;
4648 } while(!x && u < end);
4653 /** Compare two items pointing at size_t's of unknown alignment. */
4654 #ifdef MISALIGNED_OK
4655 # define mdb_cmp_clong mdb_cmp_long
4657 # define mdb_cmp_clong mdb_cmp_cint
4660 /** Compare two items lexically */
4662 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4669 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4675 diff = memcmp(a->mv_data, b->mv_data, len);
4676 return diff ? diff : len_diff<0 ? -1 : len_diff;
4679 /** Compare two items in reverse byte order */
4681 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4683 const unsigned char *p1, *p2, *p1_lim;
4687 p1_lim = (const unsigned char *)a->mv_data;
4688 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4689 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4691 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4697 while (p1 > p1_lim) {
4698 diff = *--p1 - *--p2;
4702 return len_diff<0 ? -1 : len_diff;
4705 /** Search for key within a page, using binary search.
4706 * Returns the smallest entry larger or equal to the key.
4707 * If exactp is non-null, stores whether the found entry was an exact match
4708 * in *exactp (1 or 0).
4709 * Updates the cursor index with the index of the found entry.
4710 * If no entry larger or equal to the key is found, returns NULL.
4713 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4715 unsigned int i = 0, nkeys;
4718 MDB_page *mp = mc->mc_pg[mc->mc_top];
4719 MDB_node *node = NULL;
4724 nkeys = NUMKEYS(mp);
4726 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4727 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4730 low = IS_LEAF(mp) ? 0 : 1;
4732 cmp = mc->mc_dbx->md_cmp;
4734 /* Branch pages have no data, so if using integer keys,
4735 * alignment is guaranteed. Use faster mdb_cmp_int.
4737 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4738 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4745 nodekey.mv_size = mc->mc_db->md_pad;
4746 node = NODEPTR(mp, 0); /* fake */
4747 while (low <= high) {
4748 i = (low + high) >> 1;
4749 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4750 rc = cmp(key, &nodekey);
4751 DPRINTF(("found leaf index %u [%s], rc = %i",
4752 i, DKEY(&nodekey), rc));
4761 while (low <= high) {
4762 i = (low + high) >> 1;
4764 node = NODEPTR(mp, i);
4765 nodekey.mv_size = NODEKSZ(node);
4766 nodekey.mv_data = NODEKEY(node);
4768 rc = cmp(key, &nodekey);
4771 DPRINTF(("found leaf index %u [%s], rc = %i",
4772 i, DKEY(&nodekey), rc));
4774 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4775 i, DKEY(&nodekey), NODEPGNO(node), rc));
4786 if (rc > 0) { /* Found entry is less than the key. */
4787 i++; /* Skip to get the smallest entry larger than key. */
4789 node = NODEPTR(mp, i);
4792 *exactp = (rc == 0 && nkeys > 0);
4793 /* store the key index */
4794 mc->mc_ki[mc->mc_top] = i;
4796 /* There is no entry larger or equal to the key. */
4799 /* nodeptr is fake for LEAF2 */
4805 mdb_cursor_adjust(MDB_cursor *mc, func)
4809 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4810 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4817 /** Pop a page off the top of the cursor's stack. */
4819 mdb_cursor_pop(MDB_cursor *mc)
4823 MDB_page *top = mc->mc_pg[mc->mc_top];
4829 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4830 DDBI(mc), (void *) mc));
4834 /** Push a page onto the top of the cursor's stack. */
4836 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4838 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4839 DDBI(mc), (void *) mc));
4841 if (mc->mc_snum >= CURSOR_STACK) {
4842 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4843 return MDB_CURSOR_FULL;
4846 mc->mc_top = mc->mc_snum++;
4847 mc->mc_pg[mc->mc_top] = mp;
4848 mc->mc_ki[mc->mc_top] = 0;
4853 /** Find the address of the page corresponding to a given page number.
4854 * @param[in] txn the transaction for this access.
4855 * @param[in] pgno the page number for the page to retrieve.
4856 * @param[out] ret address of a pointer where the page's address will be stored.
4857 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4858 * @return 0 on success, non-zero on failure.
4861 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4863 MDB_env *env = txn->mt_env;
4867 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4871 MDB_ID2L dl = tx2->mt_u.dirty_list;
4873 /* Spilled pages were dirtied in this txn and flushed
4874 * because the dirty list got full. Bring this page
4875 * back in from the map (but don't unspill it here,
4876 * leave that unless page_touch happens again).
4878 if (tx2->mt_spill_pgs) {
4879 MDB_ID pn = pgno << 1;
4880 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4881 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4882 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4887 unsigned x = mdb_mid2l_search(dl, pgno);
4888 if (x <= dl[0].mid && dl[x].mid == pgno) {
4894 } while ((tx2 = tx2->mt_parent) != NULL);
4897 if (pgno < txn->mt_next_pgno) {
4899 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4901 DPRINTF(("page %"Z"u not found", pgno));
4902 txn->mt_flags |= MDB_TXN_ERROR;
4903 return MDB_PAGE_NOTFOUND;
4913 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4914 * The cursor is at the root page, set up the rest of it.
4917 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4919 MDB_page *mp = mc->mc_pg[mc->mc_top];
4923 while (IS_BRANCH(mp)) {
4927 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4928 mdb_cassert(mc, NUMKEYS(mp) > 1);
4929 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4931 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4933 if (flags & MDB_PS_LAST)
4934 i = NUMKEYS(mp) - 1;
4937 node = mdb_node_search(mc, key, &exact);
4939 i = NUMKEYS(mp) - 1;
4941 i = mc->mc_ki[mc->mc_top];
4943 mdb_cassert(mc, i > 0);
4947 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4950 mdb_cassert(mc, i < NUMKEYS(mp));
4951 node = NODEPTR(mp, i);
4953 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4956 mc->mc_ki[mc->mc_top] = i;
4957 if ((rc = mdb_cursor_push(mc, mp)))
4960 if (flags & MDB_PS_MODIFY) {
4961 if ((rc = mdb_page_touch(mc)) != 0)
4963 mp = mc->mc_pg[mc->mc_top];
4968 DPRINTF(("internal error, index points to a %02X page!?",
4970 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4971 return MDB_CORRUPTED;
4974 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4975 key ? DKEY(key) : "null"));
4976 mc->mc_flags |= C_INITIALIZED;
4977 mc->mc_flags &= ~C_EOF;
4982 /** Search for the lowest key under the current branch page.
4983 * This just bypasses a NUMKEYS check in the current page
4984 * before calling mdb_page_search_root(), because the callers
4985 * are all in situations where the current page is known to
4989 mdb_page_search_lowest(MDB_cursor *mc)
4991 MDB_page *mp = mc->mc_pg[mc->mc_top];
4992 MDB_node *node = NODEPTR(mp, 0);
4995 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4998 mc->mc_ki[mc->mc_top] = 0;
4999 if ((rc = mdb_cursor_push(mc, mp)))
5001 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5004 /** Search for the page a given key should be in.
5005 * Push it and its parent pages on the cursor stack.
5006 * @param[in,out] mc the cursor for this operation.
5007 * @param[in] key the key to search for, or NULL for first/last page.
5008 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5009 * are touched (updated with new page numbers).
5010 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5011 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5012 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5013 * @return 0 on success, non-zero on failure.
5016 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5021 /* Make sure the txn is still viable, then find the root from
5022 * the txn's db table and set it as the root of the cursor's stack.
5024 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5025 DPUTS("transaction has failed, must abort");
5028 /* Make sure we're using an up-to-date root */
5029 if (*mc->mc_dbflag & DB_STALE) {
5031 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5033 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5034 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5041 MDB_node *leaf = mdb_node_search(&mc2,
5042 &mc->mc_dbx->md_name, &exact);
5044 return MDB_NOTFOUND;
5045 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5048 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5050 /* The txn may not know this DBI, or another process may
5051 * have dropped and recreated the DB with other flags.
5053 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5054 return MDB_INCOMPATIBLE;
5055 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5057 *mc->mc_dbflag &= ~DB_STALE;
5059 root = mc->mc_db->md_root;
5061 if (root == P_INVALID) { /* Tree is empty. */
5062 DPUTS("tree is empty");
5063 return MDB_NOTFOUND;
5067 mdb_cassert(mc, root > 1);
5068 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5069 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5075 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5076 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5078 if (flags & MDB_PS_MODIFY) {
5079 if ((rc = mdb_page_touch(mc)))
5083 if (flags & MDB_PS_ROOTONLY)
5086 return mdb_page_search_root(mc, key, flags);
5090 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5092 MDB_txn *txn = mc->mc_txn;
5093 pgno_t pg = mp->mp_pgno;
5094 unsigned x = 0, ovpages = mp->mp_pages;
5095 MDB_env *env = txn->mt_env;
5096 MDB_IDL sl = txn->mt_spill_pgs;
5097 MDB_ID pn = pg << 1;
5100 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5101 /* If the page is dirty or on the spill list we just acquired it,
5102 * so we should give it back to our current free list, if any.
5103 * Otherwise put it onto the list of pages we freed in this txn.
5105 * Won't create me_pghead: me_pglast must be inited along with it.
5106 * Unsupported in nested txns: They would need to hide the page
5107 * range in ancestor txns' dirty and spilled lists.
5109 if (env->me_pghead &&
5111 ((mp->mp_flags & P_DIRTY) ||
5112 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5116 MDB_ID2 *dl, ix, iy;
5117 rc = mdb_midl_need(&env->me_pghead, ovpages);
5120 if (!(mp->mp_flags & P_DIRTY)) {
5121 /* This page is no longer spilled */
5128 /* Remove from dirty list */
5129 dl = txn->mt_u.dirty_list;
5131 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5137 mdb_cassert(mc, x > 1);
5139 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5140 txn->mt_flags |= MDB_TXN_ERROR;
5141 return MDB_CORRUPTED;
5144 if (!(env->me_flags & MDB_WRITEMAP))
5145 mdb_dpage_free(env, mp);
5147 /* Insert in me_pghead */
5148 mop = env->me_pghead;
5149 j = mop[0] + ovpages;
5150 for (i = mop[0]; i && mop[i] < pg; i--)
5156 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5160 mc->mc_db->md_overflow_pages -= ovpages;
5164 /** Return the data associated with a given node.
5165 * @param[in] txn The transaction for this operation.
5166 * @param[in] leaf The node being read.
5167 * @param[out] data Updated to point to the node's data.
5168 * @return 0 on success, non-zero on failure.
5171 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5173 MDB_page *omp; /* overflow page */
5177 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5178 data->mv_size = NODEDSZ(leaf);
5179 data->mv_data = NODEDATA(leaf);
5183 /* Read overflow data.
5185 data->mv_size = NODEDSZ(leaf);
5186 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5187 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5188 DPRINTF(("read overflow page %"Z"u failed", pgno));
5191 data->mv_data = METADATA(omp);
5197 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5198 MDB_val *key, MDB_val *data)
5205 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5207 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5210 if (txn->mt_flags & MDB_TXN_ERROR)
5213 mdb_cursor_init(&mc, txn, dbi, &mx);
5214 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5217 /** Find a sibling for a page.
5218 * Replaces the page at the top of the cursor's stack with the
5219 * specified sibling, if one exists.
5220 * @param[in] mc The cursor for this operation.
5221 * @param[in] move_right Non-zero if the right sibling is requested,
5222 * otherwise the left sibling.
5223 * @return 0 on success, non-zero on failure.
5226 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5232 if (mc->mc_snum < 2) {
5233 return MDB_NOTFOUND; /* root has no siblings */
5237 DPRINTF(("parent page is page %"Z"u, index %u",
5238 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5240 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5241 : (mc->mc_ki[mc->mc_top] == 0)) {
5242 DPRINTF(("no more keys left, moving to %s sibling",
5243 move_right ? "right" : "left"));
5244 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5245 /* undo cursor_pop before returning */
5252 mc->mc_ki[mc->mc_top]++;
5254 mc->mc_ki[mc->mc_top]--;
5255 DPRINTF(("just moving to %s index key %u",
5256 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5258 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5260 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5261 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5262 /* mc will be inconsistent if caller does mc_snum++ as above */
5263 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5267 mdb_cursor_push(mc, mp);
5269 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5274 /** Move the cursor to the next data item. */
5276 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5282 if (mc->mc_flags & C_EOF) {
5283 return MDB_NOTFOUND;
5286 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5288 mp = mc->mc_pg[mc->mc_top];
5290 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5291 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5292 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5293 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5294 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5295 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5296 if (rc == MDB_SUCCESS)
5297 MDB_GET_KEY(leaf, key);
5302 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5303 if (op == MDB_NEXT_DUP)
5304 return MDB_NOTFOUND;
5308 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5309 mdb_dbg_pgno(mp), (void *) mc));
5310 if (mc->mc_flags & C_DEL)
5313 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5314 DPUTS("=====> move to next sibling page");
5315 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5316 mc->mc_flags |= C_EOF;
5319 mp = mc->mc_pg[mc->mc_top];
5320 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5322 mc->mc_ki[mc->mc_top]++;
5325 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5326 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5329 key->mv_size = mc->mc_db->md_pad;
5330 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5334 mdb_cassert(mc, IS_LEAF(mp));
5335 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5337 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5338 mdb_xcursor_init1(mc, leaf);
5341 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5344 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5345 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5346 if (rc != MDB_SUCCESS)
5351 MDB_GET_KEY(leaf, key);
5355 /** Move the cursor to the previous data item. */
5357 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5363 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5365 mp = mc->mc_pg[mc->mc_top];
5367 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5368 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5369 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5370 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5371 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5372 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5373 if (rc == MDB_SUCCESS) {
5374 MDB_GET_KEY(leaf, key);
5375 mc->mc_flags &= ~C_EOF;
5380 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5381 if (op == MDB_PREV_DUP)
5382 return MDB_NOTFOUND;
5387 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5388 mdb_dbg_pgno(mp), (void *) mc));
5390 if (mc->mc_ki[mc->mc_top] == 0) {
5391 DPUTS("=====> move to prev sibling page");
5392 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5395 mp = mc->mc_pg[mc->mc_top];
5396 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5397 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5399 mc->mc_ki[mc->mc_top]--;
5401 mc->mc_flags &= ~C_EOF;
5403 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5404 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5407 key->mv_size = mc->mc_db->md_pad;
5408 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5412 mdb_cassert(mc, IS_LEAF(mp));
5413 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5415 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5416 mdb_xcursor_init1(mc, leaf);
5419 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5422 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5423 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5424 if (rc != MDB_SUCCESS)
5429 MDB_GET_KEY(leaf, key);
5433 /** Set the cursor on a specific data item. */
5435 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5436 MDB_cursor_op op, int *exactp)
5440 MDB_node *leaf = NULL;
5443 if (key->mv_size == 0)
5444 return MDB_BAD_VALSIZE;
5447 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5449 /* See if we're already on the right page */
5450 if (mc->mc_flags & C_INITIALIZED) {
5453 mp = mc->mc_pg[mc->mc_top];
5455 mc->mc_ki[mc->mc_top] = 0;
5456 return MDB_NOTFOUND;
5458 if (mp->mp_flags & P_LEAF2) {
5459 nodekey.mv_size = mc->mc_db->md_pad;
5460 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5462 leaf = NODEPTR(mp, 0);
5463 MDB_GET_KEY2(leaf, nodekey);
5465 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5467 /* Probably happens rarely, but first node on the page
5468 * was the one we wanted.
5470 mc->mc_ki[mc->mc_top] = 0;
5477 unsigned int nkeys = NUMKEYS(mp);
5479 if (mp->mp_flags & P_LEAF2) {
5480 nodekey.mv_data = LEAF2KEY(mp,
5481 nkeys-1, nodekey.mv_size);
5483 leaf = NODEPTR(mp, nkeys-1);
5484 MDB_GET_KEY2(leaf, nodekey);
5486 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5488 /* last node was the one we wanted */
5489 mc->mc_ki[mc->mc_top] = nkeys-1;
5495 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5496 /* This is definitely the right page, skip search_page */
5497 if (mp->mp_flags & P_LEAF2) {
5498 nodekey.mv_data = LEAF2KEY(mp,
5499 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5501 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5502 MDB_GET_KEY2(leaf, nodekey);
5504 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5506 /* current node was the one we wanted */
5516 /* If any parents have right-sibs, search.
5517 * Otherwise, there's nothing further.
5519 for (i=0; i<mc->mc_top; i++)
5521 NUMKEYS(mc->mc_pg[i])-1)
5523 if (i == mc->mc_top) {
5524 /* There are no other pages */
5525 mc->mc_ki[mc->mc_top] = nkeys;
5526 return MDB_NOTFOUND;
5530 /* There are no other pages */
5531 mc->mc_ki[mc->mc_top] = 0;
5532 if (op == MDB_SET_RANGE && !exactp) {
5536 return MDB_NOTFOUND;
5540 rc = mdb_page_search(mc, key, 0);
5541 if (rc != MDB_SUCCESS)
5544 mp = mc->mc_pg[mc->mc_top];
5545 mdb_cassert(mc, IS_LEAF(mp));
5548 leaf = mdb_node_search(mc, key, exactp);
5549 if (exactp != NULL && !*exactp) {
5550 /* MDB_SET specified and not an exact match. */
5551 return MDB_NOTFOUND;
5555 DPUTS("===> inexact leaf not found, goto sibling");
5556 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5557 return rc; /* no entries matched */
5558 mp = mc->mc_pg[mc->mc_top];
5559 mdb_cassert(mc, IS_LEAF(mp));
5560 leaf = NODEPTR(mp, 0);
5564 mc->mc_flags |= C_INITIALIZED;
5565 mc->mc_flags &= ~C_EOF;
5568 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5569 key->mv_size = mc->mc_db->md_pad;
5570 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5575 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5576 mdb_xcursor_init1(mc, leaf);
5579 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5580 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5581 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5584 if (op == MDB_GET_BOTH) {
5590 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5591 if (rc != MDB_SUCCESS)
5594 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5596 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5598 rc = mc->mc_dbx->md_dcmp(data, &d2);
5600 if (op == MDB_GET_BOTH || rc > 0)
5601 return MDB_NOTFOUND;
5608 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5609 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5614 /* The key already matches in all other cases */
5615 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5616 MDB_GET_KEY(leaf, key);
5617 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5622 /** Move the cursor to the first item in the database. */
5624 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5630 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5632 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5633 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5634 if (rc != MDB_SUCCESS)
5637 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5639 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5640 mc->mc_flags |= C_INITIALIZED;
5641 mc->mc_flags &= ~C_EOF;
5643 mc->mc_ki[mc->mc_top] = 0;
5645 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5646 key->mv_size = mc->mc_db->md_pad;
5647 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5652 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5653 mdb_xcursor_init1(mc, leaf);
5654 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5658 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5662 MDB_GET_KEY(leaf, key);
5666 /** Move the cursor to the last item in the database. */
5668 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5674 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5676 if (!(mc->mc_flags & C_EOF)) {
5678 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5679 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5680 if (rc != MDB_SUCCESS)
5683 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5686 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5687 mc->mc_flags |= C_INITIALIZED|C_EOF;
5688 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5690 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5691 key->mv_size = mc->mc_db->md_pad;
5692 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5697 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5698 mdb_xcursor_init1(mc, leaf);
5699 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5703 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5708 MDB_GET_KEY(leaf, key);
5713 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5718 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5723 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5727 case MDB_GET_CURRENT:
5728 if (!(mc->mc_flags & C_INITIALIZED)) {
5731 MDB_page *mp = mc->mc_pg[mc->mc_top];
5732 int nkeys = NUMKEYS(mp);
5733 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5734 mc->mc_ki[mc->mc_top] = nkeys;
5740 key->mv_size = mc->mc_db->md_pad;
5741 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5743 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5744 MDB_GET_KEY(leaf, key);
5746 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5747 if (mc->mc_flags & C_DEL)
5748 mdb_xcursor_init1(mc, leaf);
5749 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5751 rc = mdb_node_read(mc->mc_txn, leaf, data);
5758 case MDB_GET_BOTH_RANGE:
5763 if (mc->mc_xcursor == NULL) {
5764 rc = MDB_INCOMPATIBLE;
5774 rc = mdb_cursor_set(mc, key, data, op,
5775 op == MDB_SET_RANGE ? NULL : &exact);
5778 case MDB_GET_MULTIPLE:
5779 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5783 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5784 rc = MDB_INCOMPATIBLE;
5788 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5789 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5792 case MDB_NEXT_MULTIPLE:
5797 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5798 rc = MDB_INCOMPATIBLE;
5801 if (!(mc->mc_flags & C_INITIALIZED))
5802 rc = mdb_cursor_first(mc, key, data);
5804 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5805 if (rc == MDB_SUCCESS) {
5806 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5809 mx = &mc->mc_xcursor->mx_cursor;
5810 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5812 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5813 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5821 case MDB_NEXT_NODUP:
5822 if (!(mc->mc_flags & C_INITIALIZED))
5823 rc = mdb_cursor_first(mc, key, data);
5825 rc = mdb_cursor_next(mc, key, data, op);
5829 case MDB_PREV_NODUP:
5830 if (!(mc->mc_flags & C_INITIALIZED)) {
5831 rc = mdb_cursor_last(mc, key, data);
5834 mc->mc_flags |= C_INITIALIZED;
5835 mc->mc_ki[mc->mc_top]++;
5837 rc = mdb_cursor_prev(mc, key, data, op);
5840 rc = mdb_cursor_first(mc, key, data);
5843 mfunc = mdb_cursor_first;
5845 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5849 if (mc->mc_xcursor == NULL) {
5850 rc = MDB_INCOMPATIBLE;
5854 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5855 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5856 MDB_GET_KEY(leaf, key);
5858 rc = mdb_node_read(mc->mc_txn, leaf, data);
5862 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5866 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5869 rc = mdb_cursor_last(mc, key, data);
5872 mfunc = mdb_cursor_last;
5875 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5880 if (mc->mc_flags & C_DEL)
5881 mc->mc_flags ^= C_DEL;
5886 /** Touch all the pages in the cursor stack. Set mc_top.
5887 * Makes sure all the pages are writable, before attempting a write operation.
5888 * @param[in] mc The cursor to operate on.
5891 mdb_cursor_touch(MDB_cursor *mc)
5893 int rc = MDB_SUCCESS;
5895 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5898 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5900 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5901 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5904 *mc->mc_dbflag |= DB_DIRTY;
5909 rc = mdb_page_touch(mc);
5910 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5911 mc->mc_top = mc->mc_snum-1;
5916 /** Do not spill pages to disk if txn is getting full, may fail instead */
5917 #define MDB_NOSPILL 0x8000
5920 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5923 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5925 MDB_node *leaf = NULL;
5928 MDB_val xdata, *rdata, dkey, olddata;
5930 int do_sub = 0, insert_key, insert_data;
5931 unsigned int mcount = 0, dcount = 0, nospill;
5934 unsigned int nflags;
5937 if (mc == NULL || key == NULL)
5940 env = mc->mc_txn->mt_env;
5942 /* Check this first so counter will always be zero on any
5945 if (flags & MDB_MULTIPLE) {
5946 dcount = data[1].mv_size;
5947 data[1].mv_size = 0;
5948 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5949 return MDB_INCOMPATIBLE;
5952 nospill = flags & MDB_NOSPILL;
5953 flags &= ~MDB_NOSPILL;
5955 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5956 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5958 if (key->mv_size-1 >= ENV_MAXKEY(env))
5959 return MDB_BAD_VALSIZE;
5961 #if SIZE_MAX > MAXDATASIZE
5962 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5963 return MDB_BAD_VALSIZE;
5965 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5966 return MDB_BAD_VALSIZE;
5969 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5970 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5974 if (flags == MDB_CURRENT) {
5975 if (!(mc->mc_flags & C_INITIALIZED))
5978 } else if (mc->mc_db->md_root == P_INVALID) {
5979 /* new database, cursor has nothing to point to */
5982 mc->mc_flags &= ~C_INITIALIZED;
5987 if (flags & MDB_APPEND) {
5989 rc = mdb_cursor_last(mc, &k2, &d2);
5991 rc = mc->mc_dbx->md_cmp(key, &k2);
5994 mc->mc_ki[mc->mc_top]++;
5996 /* new key is <= last key */
6001 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6003 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6004 DPRINTF(("duplicate key [%s]", DKEY(key)));
6006 return MDB_KEYEXIST;
6008 if (rc && rc != MDB_NOTFOUND)
6012 if (mc->mc_flags & C_DEL)
6013 mc->mc_flags ^= C_DEL;
6015 /* Cursor is positioned, check for room in the dirty list */
6017 if (flags & MDB_MULTIPLE) {
6019 xdata.mv_size = data->mv_size * dcount;
6023 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6027 if (rc == MDB_NO_ROOT) {
6029 /* new database, write a root leaf page */
6030 DPUTS("allocating new root leaf page");
6031 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6034 mdb_cursor_push(mc, np);
6035 mc->mc_db->md_root = np->mp_pgno;
6036 mc->mc_db->md_depth++;
6037 *mc->mc_dbflag |= DB_DIRTY;
6038 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6040 np->mp_flags |= P_LEAF2;
6041 mc->mc_flags |= C_INITIALIZED;
6043 /* make sure all cursor pages are writable */
6044 rc2 = mdb_cursor_touch(mc);
6049 insert_key = insert_data = rc;
6051 /* The key does not exist */
6052 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6053 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6054 LEAFSIZE(key, data) > env->me_nodemax)
6056 /* Too big for a node, insert in sub-DB. Set up an empty
6057 * "old sub-page" for prep_subDB to expand to a full page.
6059 fp_flags = P_LEAF|P_DIRTY;
6061 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6062 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6063 olddata.mv_size = PAGEHDRSZ;
6067 /* there's only a key anyway, so this is a no-op */
6068 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6070 unsigned int ksize = mc->mc_db->md_pad;
6071 if (key->mv_size != ksize)
6072 return MDB_BAD_VALSIZE;
6073 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6074 memcpy(ptr, key->mv_data, ksize);
6076 /* if overwriting slot 0 of leaf, need to
6077 * update branch key if there is a parent page
6079 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6080 unsigned short top = mc->mc_top;
6082 /* slot 0 is always an empty key, find real slot */
6083 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6085 if (mc->mc_ki[mc->mc_top])
6086 rc2 = mdb_update_key(mc, key);
6097 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6098 olddata.mv_size = NODEDSZ(leaf);
6099 olddata.mv_data = NODEDATA(leaf);
6102 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6103 /* Prepare (sub-)page/sub-DB to accept the new item,
6104 * if needed. fp: old sub-page or a header faking
6105 * it. mp: new (sub-)page. offset: growth in page
6106 * size. xdata: node data with new page or DB.
6108 unsigned i, offset = 0;
6109 mp = fp = xdata.mv_data = env->me_pbuf;
6110 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6112 /* Was a single item before, must convert now */
6113 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6114 /* Just overwrite the current item */
6115 if (flags == MDB_CURRENT)
6118 #if UINT_MAX < SIZE_MAX
6119 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6120 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6122 /* does data match? */
6123 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6124 if (flags & MDB_NODUPDATA)
6125 return MDB_KEYEXIST;
6130 /* Back up original data item */
6131 dkey.mv_size = olddata.mv_size;
6132 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6134 /* Make sub-page header for the dup items, with dummy body */
6135 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6136 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6137 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6138 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6139 fp->mp_flags |= P_LEAF2;
6140 fp->mp_pad = data->mv_size;
6141 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6143 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6144 (dkey.mv_size & 1) + (data->mv_size & 1);
6146 fp->mp_upper = xdata.mv_size - PAGEBASE;
6147 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6148 } else if (leaf->mn_flags & F_SUBDATA) {
6149 /* Data is on sub-DB, just store it */
6150 flags |= F_DUPDATA|F_SUBDATA;
6153 /* Data is on sub-page */
6154 fp = olddata.mv_data;
6157 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6158 offset = EVEN(NODESIZE + sizeof(indx_t) +
6162 offset = fp->mp_pad;
6163 if (SIZELEFT(fp) < offset) {
6164 offset *= 4; /* space for 4 more */
6167 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6169 fp->mp_flags |= P_DIRTY;
6170 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6171 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6175 xdata.mv_size = olddata.mv_size + offset;
6178 fp_flags = fp->mp_flags;
6179 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6180 /* Too big for a sub-page, convert to sub-DB */
6181 fp_flags &= ~P_SUBP;
6183 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6184 fp_flags |= P_LEAF2;
6185 dummy.md_pad = fp->mp_pad;
6186 dummy.md_flags = MDB_DUPFIXED;
6187 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6188 dummy.md_flags |= MDB_INTEGERKEY;
6194 dummy.md_branch_pages = 0;
6195 dummy.md_leaf_pages = 1;
6196 dummy.md_overflow_pages = 0;
6197 dummy.md_entries = NUMKEYS(fp);
6198 xdata.mv_size = sizeof(MDB_db);
6199 xdata.mv_data = &dummy;
6200 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6202 offset = env->me_psize - olddata.mv_size;
6203 flags |= F_DUPDATA|F_SUBDATA;
6204 dummy.md_root = mp->mp_pgno;
6207 mp->mp_flags = fp_flags | P_DIRTY;
6208 mp->mp_pad = fp->mp_pad;
6209 mp->mp_lower = fp->mp_lower;
6210 mp->mp_upper = fp->mp_upper + offset;
6211 if (fp_flags & P_LEAF2) {
6212 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6214 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6215 olddata.mv_size - fp->mp_upper - PAGEBASE);
6216 for (i=0; i<NUMKEYS(fp); i++)
6217 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6225 mdb_node_del(mc, 0);
6229 /* overflow page overwrites need special handling */
6230 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6233 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6235 memcpy(&pg, olddata.mv_data, sizeof(pg));
6236 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6238 ovpages = omp->mp_pages;
6240 /* Is the ov page large enough? */
6241 if (ovpages >= dpages) {
6242 if (!(omp->mp_flags & P_DIRTY) &&
6243 (level || (env->me_flags & MDB_WRITEMAP)))
6245 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6248 level = 0; /* dirty in this txn or clean */
6251 if (omp->mp_flags & P_DIRTY) {
6252 /* yes, overwrite it. Note in this case we don't
6253 * bother to try shrinking the page if the new data
6254 * is smaller than the overflow threshold.
6257 /* It is writable only in a parent txn */
6258 size_t sz = (size_t) env->me_psize * ovpages, off;
6259 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6265 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6266 mdb_cassert(mc, rc2 == 0);
6267 if (!(flags & MDB_RESERVE)) {
6268 /* Copy end of page, adjusting alignment so
6269 * compiler may copy words instead of bytes.
6271 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6272 memcpy((size_t *)((char *)np + off),
6273 (size_t *)((char *)omp + off), sz - off);
6276 memcpy(np, omp, sz); /* Copy beginning of page */
6279 SETDSZ(leaf, data->mv_size);
6280 if (F_ISSET(flags, MDB_RESERVE))
6281 data->mv_data = METADATA(omp);
6283 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6287 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6289 } else if (data->mv_size == olddata.mv_size) {
6290 /* same size, just replace it. Note that we could
6291 * also reuse this node if the new data is smaller,
6292 * but instead we opt to shrink the node in that case.
6294 if (F_ISSET(flags, MDB_RESERVE))
6295 data->mv_data = olddata.mv_data;
6296 else if (!(mc->mc_flags & C_SUB))
6297 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6299 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6304 mdb_node_del(mc, 0);
6310 nflags = flags & NODE_ADD_FLAGS;
6311 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6312 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6313 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6314 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6316 nflags |= MDB_SPLIT_REPLACE;
6317 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6319 /* There is room already in this leaf page. */
6320 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6321 if (rc == 0 && insert_key) {
6322 /* Adjust other cursors pointing to mp */
6323 MDB_cursor *m2, *m3;
6324 MDB_dbi dbi = mc->mc_dbi;
6325 unsigned i = mc->mc_top;
6326 MDB_page *mp = mc->mc_pg[i];
6328 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6329 if (mc->mc_flags & C_SUB)
6330 m3 = &m2->mc_xcursor->mx_cursor;
6333 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6334 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6341 if (rc == MDB_SUCCESS) {
6342 /* Now store the actual data in the child DB. Note that we're
6343 * storing the user data in the keys field, so there are strict
6344 * size limits on dupdata. The actual data fields of the child
6345 * DB are all zero size.
6353 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6354 if (flags & MDB_CURRENT) {
6355 xflags = MDB_CURRENT|MDB_NOSPILL;
6357 mdb_xcursor_init1(mc, leaf);
6358 xflags = (flags & MDB_NODUPDATA) ?
6359 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6361 /* converted, write the original data first */
6363 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6367 /* Adjust other cursors pointing to mp */
6369 unsigned i = mc->mc_top;
6370 MDB_page *mp = mc->mc_pg[i];
6372 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6373 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6374 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6375 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6376 mdb_xcursor_init1(m2, leaf);
6380 /* we've done our job */
6383 ecount = mc->mc_xcursor->mx_db.md_entries;
6384 if (flags & MDB_APPENDDUP)
6385 xflags |= MDB_APPEND;
6386 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6387 if (flags & F_SUBDATA) {
6388 void *db = NODEDATA(leaf);
6389 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6391 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6393 /* Increment count unless we just replaced an existing item. */
6395 mc->mc_db->md_entries++;
6397 /* Invalidate txn if we created an empty sub-DB */
6400 /* If we succeeded and the key didn't exist before,
6401 * make sure the cursor is marked valid.
6403 mc->mc_flags |= C_INITIALIZED;
6405 if (flags & MDB_MULTIPLE) {
6408 /* let caller know how many succeeded, if any */
6409 data[1].mv_size = mcount;
6410 if (mcount < dcount) {
6411 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6412 insert_key = insert_data = 0;
6419 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6422 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6427 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6433 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6434 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6436 if (!(mc->mc_flags & C_INITIALIZED))
6439 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6440 return MDB_NOTFOUND;
6442 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6445 rc = mdb_cursor_touch(mc);
6449 mp = mc->mc_pg[mc->mc_top];
6452 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6454 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6455 if (flags & MDB_NODUPDATA) {
6456 /* mdb_cursor_del0() will subtract the final entry */
6457 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6459 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6460 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6462 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6465 /* If sub-DB still has entries, we're done */
6466 if (mc->mc_xcursor->mx_db.md_entries) {
6467 if (leaf->mn_flags & F_SUBDATA) {
6468 /* update subDB info */
6469 void *db = NODEDATA(leaf);
6470 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6473 /* shrink fake page */
6474 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6475 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6476 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6477 /* fix other sub-DB cursors pointed at this fake page */
6478 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6479 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6480 if (m2->mc_pg[mc->mc_top] == mp &&
6481 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6482 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6485 mc->mc_db->md_entries--;
6486 mc->mc_flags |= C_DEL;
6489 /* otherwise fall thru and delete the sub-DB */
6492 if (leaf->mn_flags & F_SUBDATA) {
6493 /* add all the child DB's pages to the free list */
6494 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6500 /* add overflow pages to free list */
6501 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6505 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6506 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6507 (rc = mdb_ovpage_free(mc, omp)))
6512 return mdb_cursor_del0(mc);
6515 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6519 /** Allocate and initialize new pages for a database.
6520 * @param[in] mc a cursor on the database being added to.
6521 * @param[in] flags flags defining what type of page is being allocated.
6522 * @param[in] num the number of pages to allocate. This is usually 1,
6523 * unless allocating overflow pages for a large record.
6524 * @param[out] mp Address of a page, or NULL on failure.
6525 * @return 0 on success, non-zero on failure.
6528 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6533 if ((rc = mdb_page_alloc(mc, num, &np)))
6535 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6536 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6537 np->mp_flags = flags | P_DIRTY;
6538 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6539 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6542 mc->mc_db->md_branch_pages++;
6543 else if (IS_LEAF(np))
6544 mc->mc_db->md_leaf_pages++;
6545 else if (IS_OVERFLOW(np)) {
6546 mc->mc_db->md_overflow_pages += num;
6554 /** Calculate the size of a leaf node.
6555 * The size depends on the environment's page size; if a data item
6556 * is too large it will be put onto an overflow page and the node
6557 * size will only include the key and not the data. Sizes are always
6558 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6559 * of the #MDB_node headers.
6560 * @param[in] env The environment handle.
6561 * @param[in] key The key for the node.
6562 * @param[in] data The data for the node.
6563 * @return The number of bytes needed to store the node.
6566 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6570 sz = LEAFSIZE(key, data);
6571 if (sz > env->me_nodemax) {
6572 /* put on overflow page */
6573 sz -= data->mv_size - sizeof(pgno_t);
6576 return EVEN(sz + sizeof(indx_t));
6579 /** Calculate the size of a branch node.
6580 * The size should depend on the environment's page size but since
6581 * we currently don't support spilling large keys onto overflow
6582 * pages, it's simply the size of the #MDB_node header plus the
6583 * size of the key. Sizes are always rounded up to an even number
6584 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6585 * @param[in] env The environment handle.
6586 * @param[in] key The key for the node.
6587 * @return The number of bytes needed to store the node.
6590 mdb_branch_size(MDB_env *env, MDB_val *key)
6595 if (sz > env->me_nodemax) {
6596 /* put on overflow page */
6597 /* not implemented */
6598 /* sz -= key->size - sizeof(pgno_t); */
6601 return sz + sizeof(indx_t);
6604 /** Add a node to the page pointed to by the cursor.
6605 * @param[in] mc The cursor for this operation.
6606 * @param[in] indx The index on the page where the new node should be added.
6607 * @param[in] key The key for the new node.
6608 * @param[in] data The data for the new node, if any.
6609 * @param[in] pgno The page number, if adding a branch node.
6610 * @param[in] flags Flags for the node.
6611 * @return 0 on success, non-zero on failure. Possible errors are:
6613 * <li>ENOMEM - failed to allocate overflow pages for the node.
6614 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6615 * should never happen since all callers already calculate the
6616 * page's free space before calling this function.
6620 mdb_node_add(MDB_cursor *mc, indx_t indx,
6621 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6624 size_t node_size = NODESIZE;
6628 MDB_page *mp = mc->mc_pg[mc->mc_top];
6629 MDB_page *ofp = NULL; /* overflow page */
6632 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6634 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6635 IS_LEAF(mp) ? "leaf" : "branch",
6636 IS_SUBP(mp) ? "sub-" : "",
6637 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6638 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6641 /* Move higher keys up one slot. */
6642 int ksize = mc->mc_db->md_pad, dif;
6643 char *ptr = LEAF2KEY(mp, indx, ksize);
6644 dif = NUMKEYS(mp) - indx;
6646 memmove(ptr+ksize, ptr, dif*ksize);
6647 /* insert new key */
6648 memcpy(ptr, key->mv_data, ksize);
6650 /* Just using these for counting */
6651 mp->mp_lower += sizeof(indx_t);
6652 mp->mp_upper -= ksize - sizeof(indx_t);
6656 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6658 node_size += key->mv_size;
6660 mdb_cassert(mc, data);
6661 if (F_ISSET(flags, F_BIGDATA)) {
6662 /* Data already on overflow page. */
6663 node_size += sizeof(pgno_t);
6664 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6665 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6667 /* Put data on overflow page. */
6668 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6669 data->mv_size, node_size+data->mv_size));
6670 node_size = EVEN(node_size + sizeof(pgno_t));
6671 if ((ssize_t)node_size > room)
6673 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6675 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6679 node_size += data->mv_size;
6682 node_size = EVEN(node_size);
6683 if ((ssize_t)node_size > room)
6687 /* Move higher pointers up one slot. */
6688 for (i = NUMKEYS(mp); i > indx; i--)
6689 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6691 /* Adjust free space offsets. */
6692 ofs = mp->mp_upper - node_size;
6693 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6694 mp->mp_ptrs[indx] = ofs;
6696 mp->mp_lower += sizeof(indx_t);
6698 /* Write the node data. */
6699 node = NODEPTR(mp, indx);
6700 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6701 node->mn_flags = flags;
6703 SETDSZ(node,data->mv_size);
6708 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6711 mdb_cassert(mc, key);
6713 if (F_ISSET(flags, F_BIGDATA))
6714 memcpy(node->mn_data + key->mv_size, data->mv_data,
6716 else if (F_ISSET(flags, MDB_RESERVE))
6717 data->mv_data = node->mn_data + key->mv_size;
6719 memcpy(node->mn_data + key->mv_size, data->mv_data,
6722 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6724 if (F_ISSET(flags, MDB_RESERVE))
6725 data->mv_data = METADATA(ofp);
6727 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6734 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6735 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6736 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6737 DPRINTF(("node size = %"Z"u", node_size));
6738 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6739 return MDB_PAGE_FULL;
6742 /** Delete the specified node from a page.
6743 * @param[in] mc Cursor pointing to the node to delete.
6744 * @param[in] ksize The size of a node. Only used if the page is
6745 * part of a #MDB_DUPFIXED database.
6748 mdb_node_del(MDB_cursor *mc, int ksize)
6750 MDB_page *mp = mc->mc_pg[mc->mc_top];
6751 indx_t indx = mc->mc_ki[mc->mc_top];
6753 indx_t i, j, numkeys, ptr;
6757 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6758 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6759 numkeys = NUMKEYS(mp);
6760 mdb_cassert(mc, indx < numkeys);
6763 int x = numkeys - 1 - indx;
6764 base = LEAF2KEY(mp, indx, ksize);
6766 memmove(base, base + ksize, x * ksize);
6767 mp->mp_lower -= sizeof(indx_t);
6768 mp->mp_upper += ksize - sizeof(indx_t);
6772 node = NODEPTR(mp, indx);
6773 sz = NODESIZE + node->mn_ksize;
6775 if (F_ISSET(node->mn_flags, F_BIGDATA))
6776 sz += sizeof(pgno_t);
6778 sz += NODEDSZ(node);
6782 ptr = mp->mp_ptrs[indx];
6783 for (i = j = 0; i < numkeys; i++) {
6785 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6786 if (mp->mp_ptrs[i] < ptr)
6787 mp->mp_ptrs[j] += sz;
6792 base = (char *)mp + mp->mp_upper + PAGEBASE;
6793 memmove(base + sz, base, ptr - mp->mp_upper);
6795 mp->mp_lower -= sizeof(indx_t);
6799 /** Compact the main page after deleting a node on a subpage.
6800 * @param[in] mp The main page to operate on.
6801 * @param[in] indx The index of the subpage on the main page.
6804 mdb_node_shrink(MDB_page *mp, indx_t indx)
6810 indx_t i, numkeys, ptr;
6812 node = NODEPTR(mp, indx);
6813 sp = (MDB_page *)NODEDATA(node);
6814 delta = SIZELEFT(sp);
6815 xp = (MDB_page *)((char *)sp + delta);
6817 /* shift subpage upward */
6819 nsize = NUMKEYS(sp) * sp->mp_pad;
6821 return; /* do not make the node uneven-sized */
6822 memmove(METADATA(xp), METADATA(sp), nsize);
6825 numkeys = NUMKEYS(sp);
6826 for (i=numkeys-1; i>=0; i--)
6827 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6829 xp->mp_upper = sp->mp_lower;
6830 xp->mp_lower = sp->mp_lower;
6831 xp->mp_flags = sp->mp_flags;
6832 xp->mp_pad = sp->mp_pad;
6833 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6835 nsize = NODEDSZ(node) - delta;
6836 SETDSZ(node, nsize);
6838 /* shift lower nodes upward */
6839 ptr = mp->mp_ptrs[indx];
6840 numkeys = NUMKEYS(mp);
6841 for (i = 0; i < numkeys; i++) {
6842 if (mp->mp_ptrs[i] <= ptr)
6843 mp->mp_ptrs[i] += delta;
6846 base = (char *)mp + mp->mp_upper + PAGEBASE;
6847 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6848 mp->mp_upper += delta;
6851 /** Initial setup of a sorted-dups cursor.
6852 * Sorted duplicates are implemented as a sub-database for the given key.
6853 * The duplicate data items are actually keys of the sub-database.
6854 * Operations on the duplicate data items are performed using a sub-cursor
6855 * initialized when the sub-database is first accessed. This function does
6856 * the preliminary setup of the sub-cursor, filling in the fields that
6857 * depend only on the parent DB.
6858 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6861 mdb_xcursor_init0(MDB_cursor *mc)
6863 MDB_xcursor *mx = mc->mc_xcursor;
6865 mx->mx_cursor.mc_xcursor = NULL;
6866 mx->mx_cursor.mc_txn = mc->mc_txn;
6867 mx->mx_cursor.mc_db = &mx->mx_db;
6868 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6869 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6870 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6871 mx->mx_cursor.mc_snum = 0;
6872 mx->mx_cursor.mc_top = 0;
6873 mx->mx_cursor.mc_flags = C_SUB;
6874 mx->mx_dbx.md_name.mv_size = 0;
6875 mx->mx_dbx.md_name.mv_data = NULL;
6876 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6877 mx->mx_dbx.md_dcmp = NULL;
6878 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6881 /** Final setup of a sorted-dups cursor.
6882 * Sets up the fields that depend on the data from the main cursor.
6883 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6884 * @param[in] node The data containing the #MDB_db record for the
6885 * sorted-dup database.
6888 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6890 MDB_xcursor *mx = mc->mc_xcursor;
6892 if (node->mn_flags & F_SUBDATA) {
6893 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6894 mx->mx_cursor.mc_pg[0] = 0;
6895 mx->mx_cursor.mc_snum = 0;
6896 mx->mx_cursor.mc_top = 0;
6897 mx->mx_cursor.mc_flags = C_SUB;
6899 MDB_page *fp = NODEDATA(node);
6900 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6901 mx->mx_db.md_flags = 0;
6902 mx->mx_db.md_depth = 1;
6903 mx->mx_db.md_branch_pages = 0;
6904 mx->mx_db.md_leaf_pages = 1;
6905 mx->mx_db.md_overflow_pages = 0;
6906 mx->mx_db.md_entries = NUMKEYS(fp);
6907 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6908 mx->mx_cursor.mc_snum = 1;
6909 mx->mx_cursor.mc_top = 0;
6910 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6911 mx->mx_cursor.mc_pg[0] = fp;
6912 mx->mx_cursor.mc_ki[0] = 0;
6913 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6914 mx->mx_db.md_flags = MDB_DUPFIXED;
6915 mx->mx_db.md_pad = fp->mp_pad;
6916 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6917 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6920 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6921 mx->mx_db.md_root));
6922 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6923 #if UINT_MAX < SIZE_MAX
6924 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6925 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6929 /** Initialize a cursor for a given transaction and database. */
6931 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6934 mc->mc_backup = NULL;
6937 mc->mc_db = &txn->mt_dbs[dbi];
6938 mc->mc_dbx = &txn->mt_dbxs[dbi];
6939 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6944 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6945 mdb_tassert(txn, mx != NULL);
6946 mc->mc_xcursor = mx;
6947 mdb_xcursor_init0(mc);
6949 mc->mc_xcursor = NULL;
6951 if (*mc->mc_dbflag & DB_STALE) {
6952 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6957 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6960 size_t size = sizeof(MDB_cursor);
6962 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6965 if (txn->mt_flags & MDB_TXN_ERROR)
6968 /* Allow read access to the freelist */
6969 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6972 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6973 size += sizeof(MDB_xcursor);
6975 if ((mc = malloc(size)) != NULL) {
6976 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6977 if (txn->mt_cursors) {
6978 mc->mc_next = txn->mt_cursors[dbi];
6979 txn->mt_cursors[dbi] = mc;
6980 mc->mc_flags |= C_UNTRACK;
6992 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6994 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6997 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7000 if (txn->mt_flags & MDB_TXN_ERROR)
7003 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7007 /* Return the count of duplicate data items for the current key */
7009 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7013 if (mc == NULL || countp == NULL)
7016 if (mc->mc_xcursor == NULL)
7017 return MDB_INCOMPATIBLE;
7019 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7022 if (!(mc->mc_flags & C_INITIALIZED))
7025 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7026 return MDB_NOTFOUND;
7028 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7029 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7032 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7035 *countp = mc->mc_xcursor->mx_db.md_entries;
7041 mdb_cursor_close(MDB_cursor *mc)
7043 if (mc && !mc->mc_backup) {
7044 /* remove from txn, if tracked */
7045 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7046 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7047 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7049 *prev = mc->mc_next;
7056 mdb_cursor_txn(MDB_cursor *mc)
7058 if (!mc) return NULL;
7063 mdb_cursor_dbi(MDB_cursor *mc)
7068 /** Replace the key for a branch node with a new key.
7069 * @param[in] mc Cursor pointing to the node to operate on.
7070 * @param[in] key The new key to use.
7071 * @return 0 on success, non-zero on failure.
7074 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7080 int delta, ksize, oksize;
7081 indx_t ptr, i, numkeys, indx;
7084 indx = mc->mc_ki[mc->mc_top];
7085 mp = mc->mc_pg[mc->mc_top];
7086 node = NODEPTR(mp, indx);
7087 ptr = mp->mp_ptrs[indx];
7091 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7092 k2.mv_data = NODEKEY(node);
7093 k2.mv_size = node->mn_ksize;
7094 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7096 mdb_dkey(&k2, kbuf2),
7102 /* Sizes must be 2-byte aligned. */
7103 ksize = EVEN(key->mv_size);
7104 oksize = EVEN(node->mn_ksize);
7105 delta = ksize - oksize;
7107 /* Shift node contents if EVEN(key length) changed. */
7109 if (delta > 0 && SIZELEFT(mp) < delta) {
7111 /* not enough space left, do a delete and split */
7112 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7113 pgno = NODEPGNO(node);
7114 mdb_node_del(mc, 0);
7115 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7118 numkeys = NUMKEYS(mp);
7119 for (i = 0; i < numkeys; i++) {
7120 if (mp->mp_ptrs[i] <= ptr)
7121 mp->mp_ptrs[i] -= delta;
7124 base = (char *)mp + mp->mp_upper + PAGEBASE;
7125 len = ptr - mp->mp_upper + NODESIZE;
7126 memmove(base - delta, base, len);
7127 mp->mp_upper -= delta;
7129 node = NODEPTR(mp, indx);
7132 /* But even if no shift was needed, update ksize */
7133 if (node->mn_ksize != key->mv_size)
7134 node->mn_ksize = key->mv_size;
7137 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7143 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7145 /** Move a node from csrc to cdst.
7148 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7155 unsigned short flags;
7159 /* Mark src and dst as dirty. */
7160 if ((rc = mdb_page_touch(csrc)) ||
7161 (rc = mdb_page_touch(cdst)))
7164 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7165 key.mv_size = csrc->mc_db->md_pad;
7166 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7168 data.mv_data = NULL;
7172 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7173 mdb_cassert(csrc, !((size_t)srcnode & 1));
7174 srcpg = NODEPGNO(srcnode);
7175 flags = srcnode->mn_flags;
7176 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7177 unsigned int snum = csrc->mc_snum;
7179 /* must find the lowest key below src */
7180 rc = mdb_page_search_lowest(csrc);
7183 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7184 key.mv_size = csrc->mc_db->md_pad;
7185 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7187 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7188 key.mv_size = NODEKSZ(s2);
7189 key.mv_data = NODEKEY(s2);
7191 csrc->mc_snum = snum--;
7192 csrc->mc_top = snum;
7194 key.mv_size = NODEKSZ(srcnode);
7195 key.mv_data = NODEKEY(srcnode);
7197 data.mv_size = NODEDSZ(srcnode);
7198 data.mv_data = NODEDATA(srcnode);
7200 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7201 unsigned int snum = cdst->mc_snum;
7204 /* must find the lowest key below dst */
7205 mdb_cursor_copy(cdst, &mn);
7206 rc = mdb_page_search_lowest(&mn);
7209 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7210 bkey.mv_size = mn.mc_db->md_pad;
7211 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7213 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7214 bkey.mv_size = NODEKSZ(s2);
7215 bkey.mv_data = NODEKEY(s2);
7217 mn.mc_snum = snum--;
7220 rc = mdb_update_key(&mn, &bkey);
7225 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7226 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7227 csrc->mc_ki[csrc->mc_top],
7229 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7230 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7232 /* Add the node to the destination page.
7234 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7235 if (rc != MDB_SUCCESS)
7238 /* Delete the node from the source page.
7240 mdb_node_del(csrc, key.mv_size);
7243 /* Adjust other cursors pointing to mp */
7244 MDB_cursor *m2, *m3;
7245 MDB_dbi dbi = csrc->mc_dbi;
7246 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7248 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7249 if (csrc->mc_flags & C_SUB)
7250 m3 = &m2->mc_xcursor->mx_cursor;
7253 if (m3 == csrc) continue;
7254 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7255 csrc->mc_ki[csrc->mc_top]) {
7256 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7257 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7262 /* Update the parent separators.
7264 if (csrc->mc_ki[csrc->mc_top] == 0) {
7265 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7266 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7267 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7269 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7270 key.mv_size = NODEKSZ(srcnode);
7271 key.mv_data = NODEKEY(srcnode);
7273 DPRINTF(("update separator for source page %"Z"u to [%s]",
7274 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7275 mdb_cursor_copy(csrc, &mn);
7278 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7281 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7283 indx_t ix = csrc->mc_ki[csrc->mc_top];
7284 nullkey.mv_size = 0;
7285 csrc->mc_ki[csrc->mc_top] = 0;
7286 rc = mdb_update_key(csrc, &nullkey);
7287 csrc->mc_ki[csrc->mc_top] = ix;
7288 mdb_cassert(csrc, rc == MDB_SUCCESS);
7292 if (cdst->mc_ki[cdst->mc_top] == 0) {
7293 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7294 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7295 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7297 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7298 key.mv_size = NODEKSZ(srcnode);
7299 key.mv_data = NODEKEY(srcnode);
7301 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7302 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7303 mdb_cursor_copy(cdst, &mn);
7306 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7309 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7311 indx_t ix = cdst->mc_ki[cdst->mc_top];
7312 nullkey.mv_size = 0;
7313 cdst->mc_ki[cdst->mc_top] = 0;
7314 rc = mdb_update_key(cdst, &nullkey);
7315 cdst->mc_ki[cdst->mc_top] = ix;
7316 mdb_cassert(csrc, rc == MDB_SUCCESS);
7323 /** Merge one page into another.
7324 * The nodes from the page pointed to by \b csrc will
7325 * be copied to the page pointed to by \b cdst and then
7326 * the \b csrc page will be freed.
7327 * @param[in] csrc Cursor pointing to the source page.
7328 * @param[in] cdst Cursor pointing to the destination page.
7329 * @return 0 on success, non-zero on failure.
7332 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7334 MDB_page *psrc, *pdst;
7341 psrc = csrc->mc_pg[csrc->mc_top];
7342 pdst = cdst->mc_pg[cdst->mc_top];
7344 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7346 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7347 mdb_cassert(csrc, cdst->mc_snum > 1);
7349 /* Mark dst as dirty. */
7350 if ((rc = mdb_page_touch(cdst)))
7353 /* Move all nodes from src to dst.
7355 j = nkeys = NUMKEYS(pdst);
7356 if (IS_LEAF2(psrc)) {
7357 key.mv_size = csrc->mc_db->md_pad;
7358 key.mv_data = METADATA(psrc);
7359 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7360 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7361 if (rc != MDB_SUCCESS)
7363 key.mv_data = (char *)key.mv_data + key.mv_size;
7366 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7367 srcnode = NODEPTR(psrc, i);
7368 if (i == 0 && IS_BRANCH(psrc)) {
7371 mdb_cursor_copy(csrc, &mn);
7372 /* must find the lowest key below src */
7373 rc = mdb_page_search_lowest(&mn);
7376 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7377 key.mv_size = mn.mc_db->md_pad;
7378 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7380 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7381 key.mv_size = NODEKSZ(s2);
7382 key.mv_data = NODEKEY(s2);
7385 key.mv_size = srcnode->mn_ksize;
7386 key.mv_data = NODEKEY(srcnode);
7389 data.mv_size = NODEDSZ(srcnode);
7390 data.mv_data = NODEDATA(srcnode);
7391 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7392 if (rc != MDB_SUCCESS)
7397 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7398 pdst->mp_pgno, NUMKEYS(pdst),
7399 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7401 /* Unlink the src page from parent and add to free list.
7404 mdb_node_del(csrc, 0);
7405 if (csrc->mc_ki[csrc->mc_top] == 0) {
7407 rc = mdb_update_key(csrc, &key);
7415 psrc = csrc->mc_pg[csrc->mc_top];
7416 /* If not operating on FreeDB, allow this page to be reused
7417 * in this txn. Otherwise just add to free list.
7419 rc = mdb_page_loose(csrc, psrc);
7423 csrc->mc_db->md_leaf_pages--;
7425 csrc->mc_db->md_branch_pages--;
7427 /* Adjust other cursors pointing to mp */
7428 MDB_cursor *m2, *m3;
7429 MDB_dbi dbi = csrc->mc_dbi;
7431 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7432 if (csrc->mc_flags & C_SUB)
7433 m3 = &m2->mc_xcursor->mx_cursor;
7436 if (m3 == csrc) continue;
7437 if (m3->mc_snum < csrc->mc_snum) continue;
7438 if (m3->mc_pg[csrc->mc_top] == psrc) {
7439 m3->mc_pg[csrc->mc_top] = pdst;
7440 m3->mc_ki[csrc->mc_top] += nkeys;
7445 unsigned int snum = cdst->mc_snum;
7446 uint16_t depth = cdst->mc_db->md_depth;
7447 mdb_cursor_pop(cdst);
7448 rc = mdb_rebalance(cdst);
7449 /* Did the tree shrink? */
7450 if (depth > cdst->mc_db->md_depth)
7452 cdst->mc_snum = snum;
7453 cdst->mc_top = snum-1;
7458 /** Copy the contents of a cursor.
7459 * @param[in] csrc The cursor to copy from.
7460 * @param[out] cdst The cursor to copy to.
7463 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7467 cdst->mc_txn = csrc->mc_txn;
7468 cdst->mc_dbi = csrc->mc_dbi;
7469 cdst->mc_db = csrc->mc_db;
7470 cdst->mc_dbx = csrc->mc_dbx;
7471 cdst->mc_snum = csrc->mc_snum;
7472 cdst->mc_top = csrc->mc_top;
7473 cdst->mc_flags = csrc->mc_flags;
7475 for (i=0; i<csrc->mc_snum; i++) {
7476 cdst->mc_pg[i] = csrc->mc_pg[i];
7477 cdst->mc_ki[i] = csrc->mc_ki[i];
7481 /** Rebalance the tree after a delete operation.
7482 * @param[in] mc Cursor pointing to the page where rebalancing
7484 * @return 0 on success, non-zero on failure.
7487 mdb_rebalance(MDB_cursor *mc)
7491 unsigned int ptop, minkeys;
7495 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7496 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7497 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7498 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7499 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7501 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7502 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7503 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7504 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7508 if (mc->mc_snum < 2) {
7509 MDB_page *mp = mc->mc_pg[0];
7511 DPUTS("Can't rebalance a subpage, ignoring");
7514 if (NUMKEYS(mp) == 0) {
7515 DPUTS("tree is completely empty");
7516 mc->mc_db->md_root = P_INVALID;
7517 mc->mc_db->md_depth = 0;
7518 mc->mc_db->md_leaf_pages = 0;
7519 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7522 /* Adjust cursors pointing to mp */
7525 mc->mc_flags &= ~C_INITIALIZED;
7527 MDB_cursor *m2, *m3;
7528 MDB_dbi dbi = mc->mc_dbi;
7530 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7531 if (mc->mc_flags & C_SUB)
7532 m3 = &m2->mc_xcursor->mx_cursor;
7535 if (m3->mc_snum < mc->mc_snum) continue;
7536 if (m3->mc_pg[0] == mp) {
7539 m3->mc_flags &= ~C_INITIALIZED;
7543 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7545 DPUTS("collapsing root page!");
7546 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7549 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7550 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7553 mc->mc_db->md_depth--;
7554 mc->mc_db->md_branch_pages--;
7555 mc->mc_ki[0] = mc->mc_ki[1];
7556 for (i = 1; i<mc->mc_db->md_depth; i++) {
7557 mc->mc_pg[i] = mc->mc_pg[i+1];
7558 mc->mc_ki[i] = mc->mc_ki[i+1];
7561 /* Adjust other cursors pointing to mp */
7562 MDB_cursor *m2, *m3;
7563 MDB_dbi dbi = mc->mc_dbi;
7565 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7566 if (mc->mc_flags & C_SUB)
7567 m3 = &m2->mc_xcursor->mx_cursor;
7570 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7571 if (m3->mc_pg[0] == mp) {
7574 for (i=0; i<m3->mc_snum; i++) {
7575 m3->mc_pg[i] = m3->mc_pg[i+1];
7576 m3->mc_ki[i] = m3->mc_ki[i+1];
7582 DPUTS("root page doesn't need rebalancing");
7586 /* The parent (branch page) must have at least 2 pointers,
7587 * otherwise the tree is invalid.
7589 ptop = mc->mc_top-1;
7590 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7592 /* Leaf page fill factor is below the threshold.
7593 * Try to move keys from left or right neighbor, or
7594 * merge with a neighbor page.
7599 mdb_cursor_copy(mc, &mn);
7600 mn.mc_xcursor = NULL;
7602 oldki = mc->mc_ki[mc->mc_top];
7603 if (mc->mc_ki[ptop] == 0) {
7604 /* We're the leftmost leaf in our parent.
7606 DPUTS("reading right neighbor");
7608 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7609 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7612 mn.mc_ki[mn.mc_top] = 0;
7613 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7615 /* There is at least one neighbor to the left.
7617 DPUTS("reading left neighbor");
7619 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7620 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7623 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7624 mc->mc_ki[mc->mc_top] = 0;
7627 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7628 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7629 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7631 /* If the neighbor page is above threshold and has enough keys,
7632 * move one key from it. Otherwise we should try to merge them.
7633 * (A branch page must never have less than 2 keys.)
7635 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7636 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7637 rc = mdb_node_move(&mn, mc);
7638 if (mc->mc_ki[ptop]) {
7642 if (mc->mc_ki[ptop] == 0) {
7643 rc = mdb_page_merge(&mn, mc);
7645 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7646 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7647 rc = mdb_page_merge(mc, &mn);
7648 mdb_cursor_copy(&mn, mc);
7650 mc->mc_flags &= ~C_EOF;
7652 mc->mc_ki[mc->mc_top] = oldki;
7656 /** Complete a delete operation started by #mdb_cursor_del(). */
7658 mdb_cursor_del0(MDB_cursor *mc)
7665 ki = mc->mc_ki[mc->mc_top];
7666 mdb_node_del(mc, mc->mc_db->md_pad);
7667 mc->mc_db->md_entries--;
7668 rc = mdb_rebalance(mc);
7670 if (rc == MDB_SUCCESS) {
7671 MDB_cursor *m2, *m3;
7672 MDB_dbi dbi = mc->mc_dbi;
7674 mp = mc->mc_pg[mc->mc_top];
7675 nkeys = NUMKEYS(mp);
7677 /* if mc points past last node in page, find next sibling */
7678 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7679 rc = mdb_cursor_sibling(mc, 1);
7680 if (rc == MDB_NOTFOUND) {
7681 mc->mc_flags |= C_EOF;
7686 /* Adjust other cursors pointing to mp */
7687 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7688 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7689 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7691 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7693 if (m3->mc_pg[mc->mc_top] == mp) {
7694 if (m3->mc_ki[mc->mc_top] >= ki) {
7695 m3->mc_flags |= C_DEL;
7696 if (m3->mc_ki[mc->mc_top] > ki)
7697 m3->mc_ki[mc->mc_top]--;
7698 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7699 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7701 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7702 rc = mdb_cursor_sibling(m3, 1);
7703 if (rc == MDB_NOTFOUND) {
7704 m3->mc_flags |= C_EOF;
7710 mc->mc_flags |= C_DEL;
7714 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7719 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7720 MDB_val *key, MDB_val *data)
7722 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7725 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7726 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7728 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7729 /* must ignore any data */
7733 return mdb_del0(txn, dbi, key, data, 0);
7737 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7738 MDB_val *key, MDB_val *data, unsigned flags)
7743 MDB_val rdata, *xdata;
7747 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7749 mdb_cursor_init(&mc, txn, dbi, &mx);
7758 flags |= MDB_NODUPDATA;
7760 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7762 /* let mdb_page_split know about this cursor if needed:
7763 * delete will trigger a rebalance; if it needs to move
7764 * a node from one page to another, it will have to
7765 * update the parent's separator key(s). If the new sepkey
7766 * is larger than the current one, the parent page may
7767 * run out of space, triggering a split. We need this
7768 * cursor to be consistent until the end of the rebalance.
7770 mc.mc_flags |= C_UNTRACK;
7771 mc.mc_next = txn->mt_cursors[dbi];
7772 txn->mt_cursors[dbi] = &mc;
7773 rc = mdb_cursor_del(&mc, flags);
7774 txn->mt_cursors[dbi] = mc.mc_next;
7779 /** Split a page and insert a new node.
7780 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7781 * The cursor will be updated to point to the actual page and index where
7782 * the node got inserted after the split.
7783 * @param[in] newkey The key for the newly inserted node.
7784 * @param[in] newdata The data for the newly inserted node.
7785 * @param[in] newpgno The page number, if the new node is a branch node.
7786 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7787 * @return 0 on success, non-zero on failure.
7790 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7791 unsigned int nflags)
7794 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7797 int i, j, split_indx, nkeys, pmax;
7798 MDB_env *env = mc->mc_txn->mt_env;
7800 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7801 MDB_page *copy = NULL;
7802 MDB_page *mp, *rp, *pp;
7807 mp = mc->mc_pg[mc->mc_top];
7808 newindx = mc->mc_ki[mc->mc_top];
7809 nkeys = NUMKEYS(mp);
7811 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7812 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7813 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7815 /* Create a right sibling. */
7816 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7818 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7820 if (mc->mc_snum < 2) {
7821 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7823 /* shift current top to make room for new parent */
7824 mc->mc_pg[1] = mc->mc_pg[0];
7825 mc->mc_ki[1] = mc->mc_ki[0];
7828 mc->mc_db->md_root = pp->mp_pgno;
7829 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7830 mc->mc_db->md_depth++;
7833 /* Add left (implicit) pointer. */
7834 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7835 /* undo the pre-push */
7836 mc->mc_pg[0] = mc->mc_pg[1];
7837 mc->mc_ki[0] = mc->mc_ki[1];
7838 mc->mc_db->md_root = mp->mp_pgno;
7839 mc->mc_db->md_depth--;
7846 ptop = mc->mc_top-1;
7847 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7850 mc->mc_flags |= C_SPLITTING;
7851 mdb_cursor_copy(mc, &mn);
7852 mn.mc_pg[mn.mc_top] = rp;
7853 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7855 if (nflags & MDB_APPEND) {
7856 mn.mc_ki[mn.mc_top] = 0;
7858 split_indx = newindx;
7862 split_indx = (nkeys+1) / 2;
7867 unsigned int lsize, rsize, ksize;
7868 /* Move half of the keys to the right sibling */
7869 x = mc->mc_ki[mc->mc_top] - split_indx;
7870 ksize = mc->mc_db->md_pad;
7871 split = LEAF2KEY(mp, split_indx, ksize);
7872 rsize = (nkeys - split_indx) * ksize;
7873 lsize = (nkeys - split_indx) * sizeof(indx_t);
7874 mp->mp_lower -= lsize;
7875 rp->mp_lower += lsize;
7876 mp->mp_upper += rsize - lsize;
7877 rp->mp_upper -= rsize - lsize;
7878 sepkey.mv_size = ksize;
7879 if (newindx == split_indx) {
7880 sepkey.mv_data = newkey->mv_data;
7882 sepkey.mv_data = split;
7885 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7886 memcpy(rp->mp_ptrs, split, rsize);
7887 sepkey.mv_data = rp->mp_ptrs;
7888 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7889 memcpy(ins, newkey->mv_data, ksize);
7890 mp->mp_lower += sizeof(indx_t);
7891 mp->mp_upper -= ksize - sizeof(indx_t);
7894 memcpy(rp->mp_ptrs, split, x * ksize);
7895 ins = LEAF2KEY(rp, x, ksize);
7896 memcpy(ins, newkey->mv_data, ksize);
7897 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7898 rp->mp_lower += sizeof(indx_t);
7899 rp->mp_upper -= ksize - sizeof(indx_t);
7900 mc->mc_ki[mc->mc_top] = x;
7901 mc->mc_pg[mc->mc_top] = rp;
7904 int psize, nsize, k;
7905 /* Maximum free space in an empty page */
7906 pmax = env->me_psize - PAGEHDRSZ;
7908 nsize = mdb_leaf_size(env, newkey, newdata);
7910 nsize = mdb_branch_size(env, newkey);
7911 nsize = EVEN(nsize);
7913 /* grab a page to hold a temporary copy */
7914 copy = mdb_page_malloc(mc->mc_txn, 1);
7919 copy->mp_pgno = mp->mp_pgno;
7920 copy->mp_flags = mp->mp_flags;
7921 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7922 copy->mp_upper = env->me_psize - PAGEBASE;
7924 /* prepare to insert */
7925 for (i=0, j=0; i<nkeys; i++) {
7927 copy->mp_ptrs[j++] = 0;
7929 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7932 /* When items are relatively large the split point needs
7933 * to be checked, because being off-by-one will make the
7934 * difference between success or failure in mdb_node_add.
7936 * It's also relevant if a page happens to be laid out
7937 * such that one half of its nodes are all "small" and
7938 * the other half of its nodes are "large." If the new
7939 * item is also "large" and falls on the half with
7940 * "large" nodes, it also may not fit.
7942 * As a final tweak, if the new item goes on the last
7943 * spot on the page (and thus, onto the new page), bias
7944 * the split so the new page is emptier than the old page.
7945 * This yields better packing during sequential inserts.
7947 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7948 /* Find split point */
7950 if (newindx <= split_indx || newindx >= nkeys) {
7952 k = newindx >= nkeys ? nkeys : split_indx+2;
7957 for (; i!=k; i+=j) {
7962 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7963 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7965 if (F_ISSET(node->mn_flags, F_BIGDATA))
7966 psize += sizeof(pgno_t);
7968 psize += NODEDSZ(node);
7970 psize = EVEN(psize);
7972 if (psize > pmax || i == k-j) {
7973 split_indx = i + (j<0);
7978 if (split_indx == newindx) {
7979 sepkey.mv_size = newkey->mv_size;
7980 sepkey.mv_data = newkey->mv_data;
7982 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
7983 sepkey.mv_size = node->mn_ksize;
7984 sepkey.mv_data = NODEKEY(node);
7989 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7991 /* Copy separator key to the parent.
7993 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7997 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8002 if (mn.mc_snum == mc->mc_snum) {
8003 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8004 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8005 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8006 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8011 /* Right page might now have changed parent.
8012 * Check if left page also changed parent.
8014 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8015 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8016 for (i=0; i<ptop; i++) {
8017 mc->mc_pg[i] = mn.mc_pg[i];
8018 mc->mc_ki[i] = mn.mc_ki[i];
8020 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8021 if (mn.mc_ki[ptop]) {
8022 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8024 /* find right page's left sibling */
8025 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8026 mdb_cursor_sibling(mc, 0);
8031 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8034 mc->mc_flags ^= C_SPLITTING;
8035 if (rc != MDB_SUCCESS) {
8038 if (nflags & MDB_APPEND) {
8039 mc->mc_pg[mc->mc_top] = rp;
8040 mc->mc_ki[mc->mc_top] = 0;
8041 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8044 for (i=0; i<mc->mc_top; i++)
8045 mc->mc_ki[i] = mn.mc_ki[i];
8046 } else if (!IS_LEAF2(mp)) {
8048 mc->mc_pg[mc->mc_top] = rp;
8053 rkey.mv_data = newkey->mv_data;
8054 rkey.mv_size = newkey->mv_size;
8060 /* Update index for the new key. */
8061 mc->mc_ki[mc->mc_top] = j;
8063 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8064 rkey.mv_data = NODEKEY(node);
8065 rkey.mv_size = node->mn_ksize;
8067 xdata.mv_data = NODEDATA(node);
8068 xdata.mv_size = NODEDSZ(node);
8071 pgno = NODEPGNO(node);
8072 flags = node->mn_flags;
8075 if (!IS_LEAF(mp) && j == 0) {
8076 /* First branch index doesn't need key data. */
8080 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8086 mc->mc_pg[mc->mc_top] = copy;
8091 } while (i != split_indx);
8093 nkeys = NUMKEYS(copy);
8094 for (i=0; i<nkeys; i++)
8095 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8096 mp->mp_lower = copy->mp_lower;
8097 mp->mp_upper = copy->mp_upper;
8098 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8099 env->me_psize - copy->mp_upper - PAGEBASE);
8101 /* reset back to original page */
8102 if (newindx < split_indx) {
8103 mc->mc_pg[mc->mc_top] = mp;
8104 if (nflags & MDB_RESERVE) {
8105 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8106 if (!(node->mn_flags & F_BIGDATA))
8107 newdata->mv_data = NODEDATA(node);
8110 mc->mc_pg[mc->mc_top] = rp;
8112 /* Make sure mc_ki is still valid.
8114 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8115 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8116 for (i=0; i<=ptop; i++) {
8117 mc->mc_pg[i] = mn.mc_pg[i];
8118 mc->mc_ki[i] = mn.mc_ki[i];
8125 /* Adjust other cursors pointing to mp */
8126 MDB_cursor *m2, *m3;
8127 MDB_dbi dbi = mc->mc_dbi;
8128 int fixup = NUMKEYS(mp);
8130 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8131 if (mc->mc_flags & C_SUB)
8132 m3 = &m2->mc_xcursor->mx_cursor;
8137 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8139 if (m3->mc_flags & C_SPLITTING)
8144 for (k=m3->mc_top; k>=0; k--) {
8145 m3->mc_ki[k+1] = m3->mc_ki[k];
8146 m3->mc_pg[k+1] = m3->mc_pg[k];
8148 if (m3->mc_ki[0] >= split_indx) {
8153 m3->mc_pg[0] = mc->mc_pg[0];
8157 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8158 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8159 m3->mc_ki[mc->mc_top]++;
8160 if (m3->mc_ki[mc->mc_top] >= fixup) {
8161 m3->mc_pg[mc->mc_top] = rp;
8162 m3->mc_ki[mc->mc_top] -= fixup;
8163 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8165 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8166 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8171 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8174 if (copy) /* tmp page */
8175 mdb_page_free(env, copy);
8177 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8182 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8183 MDB_val *key, MDB_val *data, unsigned int flags)
8188 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8191 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8194 mdb_cursor_init(&mc, txn, dbi, &mx);
8195 return mdb_cursor_put(&mc, key, data, flags);
8199 #define MDB_WBUF (1024*1024)
8202 /** State needed for a compacting copy. */
8203 typedef struct mdb_copy {
8204 pthread_mutex_t mc_mutex;
8205 pthread_cond_t mc_cond;
8212 pgno_t mc_next_pgno;
8215 volatile int mc_new;
8220 /** Dedicated writer thread for compacting copy. */
8221 static THREAD_RET ESECT
8222 mdb_env_copythr(void *arg)
8226 int toggle = 0, wsize, rc;
8229 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8232 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8235 pthread_mutex_lock(&my->mc_mutex);
8237 pthread_cond_signal(&my->mc_cond);
8240 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8241 if (my->mc_new < 0) {
8246 wsize = my->mc_wlen[toggle];
8247 ptr = my->mc_wbuf[toggle];
8250 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8254 } else if (len > 0) {
8268 /* If there's an overflow page tail, write it too */
8269 if (my->mc_olen[toggle]) {
8270 wsize = my->mc_olen[toggle];
8271 ptr = my->mc_over[toggle];
8272 my->mc_olen[toggle] = 0;
8275 my->mc_wlen[toggle] = 0;
8277 pthread_cond_signal(&my->mc_cond);
8279 pthread_cond_signal(&my->mc_cond);
8280 pthread_mutex_unlock(&my->mc_mutex);
8281 return (THREAD_RET)0;
8285 /** Tell the writer thread there's a buffer ready to write */
8287 mdb_env_cthr_toggle(mdb_copy *my, int st)
8289 int toggle = my->mc_toggle ^ 1;
8290 pthread_mutex_lock(&my->mc_mutex);
8291 if (my->mc_status) {
8292 pthread_mutex_unlock(&my->mc_mutex);
8293 return my->mc_status;
8295 while (my->mc_new == 1)
8296 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8298 my->mc_toggle = toggle;
8299 pthread_cond_signal(&my->mc_cond);
8300 pthread_mutex_unlock(&my->mc_mutex);
8304 /** Depth-first tree traversal for compacting copy. */
8306 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8309 MDB_txn *txn = my->mc_txn;
8311 MDB_page *mo, *mp, *leaf;
8316 /* Empty DB, nothing to do */
8317 if (*pg == P_INVALID)
8324 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8327 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8331 /* Make cursor pages writable */
8332 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8336 for (i=0; i<mc.mc_top; i++) {
8337 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8338 mc.mc_pg[i] = (MDB_page *)ptr;
8339 ptr += my->mc_env->me_psize;
8342 /* This is writable space for a leaf page. Usually not needed. */
8343 leaf = (MDB_page *)ptr;
8345 toggle = my->mc_toggle;
8346 while (mc.mc_snum > 0) {
8348 mp = mc.mc_pg[mc.mc_top];
8352 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8353 for (i=0; i<n; i++) {
8354 ni = NODEPTR(mp, i);
8355 if (ni->mn_flags & F_BIGDATA) {
8359 /* Need writable leaf */
8361 mc.mc_pg[mc.mc_top] = leaf;
8362 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8364 ni = NODEPTR(mp, i);
8367 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8368 rc = mdb_page_get(txn, pg, &omp, NULL);
8371 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8372 rc = mdb_env_cthr_toggle(my, 1);
8375 toggle = my->mc_toggle;
8377 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8378 memcpy(mo, omp, my->mc_env->me_psize);
8379 mo->mp_pgno = my->mc_next_pgno;
8380 my->mc_next_pgno += omp->mp_pages;
8381 my->mc_wlen[toggle] += my->mc_env->me_psize;
8382 if (omp->mp_pages > 1) {
8383 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8384 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8385 rc = mdb_env_cthr_toggle(my, 1);
8388 toggle = my->mc_toggle;
8390 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8391 } else if (ni->mn_flags & F_SUBDATA) {
8394 /* Need writable leaf */
8396 mc.mc_pg[mc.mc_top] = leaf;
8397 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8399 ni = NODEPTR(mp, i);
8402 memcpy(&db, NODEDATA(ni), sizeof(db));
8403 my->mc_toggle = toggle;
8404 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8407 toggle = my->mc_toggle;
8408 memcpy(NODEDATA(ni), &db, sizeof(db));
8413 mc.mc_ki[mc.mc_top]++;
8414 if (mc.mc_ki[mc.mc_top] < n) {
8417 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8419 rc = mdb_page_get(txn, pg, &mp, NULL);
8424 mc.mc_ki[mc.mc_top] = 0;
8425 if (IS_BRANCH(mp)) {
8426 /* Whenever we advance to a sibling branch page,
8427 * we must proceed all the way down to its first leaf.
8429 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8432 mc.mc_pg[mc.mc_top] = mp;
8436 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8437 rc = mdb_env_cthr_toggle(my, 1);
8440 toggle = my->mc_toggle;
8442 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8443 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8444 mo->mp_pgno = my->mc_next_pgno++;
8445 my->mc_wlen[toggle] += my->mc_env->me_psize;
8447 /* Update parent if there is one */
8448 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8449 SETPGNO(ni, mo->mp_pgno);
8450 mdb_cursor_pop(&mc);
8452 /* Otherwise we're done */
8462 /** Copy environment with compaction. */
8464 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8469 MDB_txn *txn = NULL;
8474 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8475 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8476 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8477 if (my.mc_wbuf[0] == NULL)
8480 pthread_mutex_init(&my.mc_mutex, NULL);
8481 pthread_cond_init(&my.mc_cond, NULL);
8482 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8483 if (my.mc_wbuf[0] == NULL)
8486 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8487 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8492 my.mc_next_pgno = 2;
8498 THREAD_CREATE(thr, mdb_env_copythr, &my);
8500 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8504 mp = (MDB_page *)my.mc_wbuf[0];
8505 memset(mp, 0, 2*env->me_psize);
8507 mp->mp_flags = P_META;
8508 mm = (MDB_meta *)METADATA(mp);
8509 mdb_env_init_meta0(env, mm);
8510 mm->mm_address = env->me_metas[0]->mm_address;
8512 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8514 mp->mp_flags = P_META;
8515 *(MDB_meta *)METADATA(mp) = *mm;
8516 mm = (MDB_meta *)METADATA(mp);
8518 /* Count the number of free pages, subtract from lastpg to find
8519 * number of active pages
8522 MDB_ID freecount = 0;
8525 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8526 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8527 freecount += *(MDB_ID *)data.mv_data;
8528 freecount += txn->mt_dbs[0].md_branch_pages +
8529 txn->mt_dbs[0].md_leaf_pages +
8530 txn->mt_dbs[0].md_overflow_pages;
8532 /* Set metapage 1 */
8533 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8534 mm->mm_dbs[1] = txn->mt_dbs[1];
8535 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8538 my.mc_wlen[0] = env->me_psize * 2;
8540 pthread_mutex_lock(&my.mc_mutex);
8542 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8543 pthread_mutex_unlock(&my.mc_mutex);
8544 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8545 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8546 rc = mdb_env_cthr_toggle(&my, 1);
8547 mdb_env_cthr_toggle(&my, -1);
8548 pthread_mutex_lock(&my.mc_mutex);
8550 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8551 pthread_mutex_unlock(&my.mc_mutex);
8556 CloseHandle(my.mc_cond);
8557 CloseHandle(my.mc_mutex);
8558 _aligned_free(my.mc_wbuf[0]);
8560 pthread_cond_destroy(&my.mc_cond);
8561 pthread_mutex_destroy(&my.mc_mutex);
8562 free(my.mc_wbuf[0]);
8567 /** Copy environment as-is. */
8569 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8571 MDB_txn *txn = NULL;
8577 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8581 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8584 /* Do the lock/unlock of the reader mutex before starting the
8585 * write txn. Otherwise other read txns could block writers.
8587 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8592 /* We must start the actual read txn after blocking writers */
8593 mdb_txn_reset0(txn, "reset-stage1");
8595 /* Temporarily block writers until we snapshot the meta pages */
8598 rc = mdb_txn_renew0(txn);
8600 UNLOCK_MUTEX_W(env);
8605 wsize = env->me_psize * 2;
8609 DO_WRITE(rc, fd, ptr, w2, len);
8613 } else if (len > 0) {
8619 /* Non-blocking or async handles are not supported */
8625 UNLOCK_MUTEX_W(env);
8630 w2 = txn->mt_next_pgno * env->me_psize;
8633 LARGE_INTEGER fsize;
8634 GetFileSizeEx(env->me_fd, &fsize);
8635 if (w2 > fsize.QuadPart)
8636 w2 = fsize.QuadPart;
8641 fstat(env->me_fd, &st);
8642 if (w2 > (size_t)st.st_size)
8648 if (wsize > MAX_WRITE)
8652 DO_WRITE(rc, fd, ptr, w2, len);
8656 } else if (len > 0) {
8673 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8675 if (flags & MDB_CP_COMPACT)
8676 return mdb_env_copyfd1(env, fd);
8678 return mdb_env_copyfd0(env, fd);
8682 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8684 return mdb_env_copyfd2(env, fd, 0);
8688 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8692 HANDLE newfd = INVALID_HANDLE_VALUE;
8694 if (env->me_flags & MDB_NOSUBDIR) {
8695 lpath = (char *)path;
8698 len += sizeof(DATANAME);
8699 lpath = malloc(len);
8702 sprintf(lpath, "%s" DATANAME, path);
8705 /* The destination path must exist, but the destination file must not.
8706 * We don't want the OS to cache the writes, since the source data is
8707 * already in the OS cache.
8710 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8711 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8713 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8715 if (newfd == INVALID_HANDLE_VALUE) {
8720 if (env->me_psize >= env->me_os_psize) {
8722 /* Set O_DIRECT if the file system supports it */
8723 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8724 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8726 #ifdef F_NOCACHE /* __APPLE__ */
8727 rc = fcntl(newfd, F_NOCACHE, 1);
8735 rc = mdb_env_copyfd2(env, newfd, flags);
8738 if (!(env->me_flags & MDB_NOSUBDIR))
8740 if (newfd != INVALID_HANDLE_VALUE)
8741 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8748 mdb_env_copy(MDB_env *env, const char *path)
8750 return mdb_env_copy2(env, path, 0);
8754 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8756 if ((flag & CHANGEABLE) != flag)
8759 env->me_flags |= flag;
8761 env->me_flags &= ~flag;
8766 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8771 *arg = env->me_flags;
8776 mdb_env_set_userctx(MDB_env *env, void *ctx)
8780 env->me_userctx = ctx;
8785 mdb_env_get_userctx(MDB_env *env)
8787 return env ? env->me_userctx : NULL;
8791 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8796 env->me_assert_func = func;
8802 mdb_env_get_path(MDB_env *env, const char **arg)
8807 *arg = env->me_path;
8812 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8821 /** Common code for #mdb_stat() and #mdb_env_stat().
8822 * @param[in] env the environment to operate in.
8823 * @param[in] db the #MDB_db record containing the stats to return.
8824 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8825 * @return 0, this function always succeeds.
8828 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8830 arg->ms_psize = env->me_psize;
8831 arg->ms_depth = db->md_depth;
8832 arg->ms_branch_pages = db->md_branch_pages;
8833 arg->ms_leaf_pages = db->md_leaf_pages;
8834 arg->ms_overflow_pages = db->md_overflow_pages;
8835 arg->ms_entries = db->md_entries;
8841 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8845 if (env == NULL || arg == NULL)
8848 toggle = mdb_env_pick_meta(env);
8850 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8854 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8858 if (env == NULL || arg == NULL)
8861 toggle = mdb_env_pick_meta(env);
8862 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8863 arg->me_mapsize = env->me_mapsize;
8864 arg->me_maxreaders = env->me_maxreaders;
8866 /* me_numreaders may be zero if this process never used any readers. Use
8867 * the shared numreader count if it exists.
8869 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8871 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8872 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8876 /** Set the default comparison functions for a database.
8877 * Called immediately after a database is opened to set the defaults.
8878 * The user can then override them with #mdb_set_compare() or
8879 * #mdb_set_dupsort().
8880 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8881 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8884 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8886 uint16_t f = txn->mt_dbs[dbi].md_flags;
8888 txn->mt_dbxs[dbi].md_cmp =
8889 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8890 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8892 txn->mt_dbxs[dbi].md_dcmp =
8893 !(f & MDB_DUPSORT) ? 0 :
8894 ((f & MDB_INTEGERDUP)
8895 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8896 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8899 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8905 int rc, dbflag, exact;
8906 unsigned int unused = 0, seq;
8909 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8910 mdb_default_cmp(txn, FREE_DBI);
8913 if ((flags & VALID_FLAGS) != flags)
8915 if (txn->mt_flags & MDB_TXN_ERROR)
8921 if (flags & PERSISTENT_FLAGS) {
8922 uint16_t f2 = flags & PERSISTENT_FLAGS;
8923 /* make sure flag changes get committed */
8924 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8925 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8926 txn->mt_flags |= MDB_TXN_DIRTY;
8929 mdb_default_cmp(txn, MAIN_DBI);
8933 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8934 mdb_default_cmp(txn, MAIN_DBI);
8937 /* Is the DB already open? */
8939 for (i=2; i<txn->mt_numdbs; i++) {
8940 if (!txn->mt_dbxs[i].md_name.mv_size) {
8941 /* Remember this free slot */
8942 if (!unused) unused = i;
8945 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8946 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8952 /* If no free slot and max hit, fail */
8953 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8954 return MDB_DBS_FULL;
8956 /* Cannot mix named databases with some mainDB flags */
8957 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8958 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8960 /* Find the DB info */
8961 dbflag = DB_NEW|DB_VALID;
8964 key.mv_data = (void *)name;
8965 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8966 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8967 if (rc == MDB_SUCCESS) {
8968 /* make sure this is actually a DB */
8969 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8970 if (!(node->mn_flags & F_SUBDATA))
8971 return MDB_INCOMPATIBLE;
8972 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8973 /* Create if requested */
8974 data.mv_size = sizeof(MDB_db);
8975 data.mv_data = &dummy;
8976 memset(&dummy, 0, sizeof(dummy));
8977 dummy.md_root = P_INVALID;
8978 dummy.md_flags = flags & PERSISTENT_FLAGS;
8979 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8983 /* OK, got info, add to table */
8984 if (rc == MDB_SUCCESS) {
8985 unsigned int slot = unused ? unused : txn->mt_numdbs;
8986 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8987 txn->mt_dbxs[slot].md_name.mv_size = len;
8988 txn->mt_dbxs[slot].md_rel = NULL;
8989 txn->mt_dbflags[slot] = dbflag;
8990 /* txn-> and env-> are the same in read txns, use
8991 * tmp variable to avoid undefined assignment
8993 seq = ++txn->mt_env->me_dbiseqs[slot];
8994 txn->mt_dbiseqs[slot] = seq;
8996 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8998 mdb_default_cmp(txn, slot);
9007 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9009 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9012 if (txn->mt_flags & MDB_TXN_ERROR)
9015 if (txn->mt_dbflags[dbi] & DB_STALE) {
9018 /* Stale, must read the DB's root. cursor_init does it for us. */
9019 mdb_cursor_init(&mc, txn, dbi, &mx);
9021 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9024 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9027 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9029 ptr = env->me_dbxs[dbi].md_name.mv_data;
9030 /* If there was no name, this was already closed */
9032 env->me_dbxs[dbi].md_name.mv_data = NULL;
9033 env->me_dbxs[dbi].md_name.mv_size = 0;
9034 env->me_dbflags[dbi] = 0;
9035 env->me_dbiseqs[dbi]++;
9040 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9042 /* We could return the flags for the FREE_DBI too but what's the point? */
9043 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9045 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9049 /** Add all the DB's pages to the free list.
9050 * @param[in] mc Cursor on the DB to free.
9051 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9052 * @return 0 on success, non-zero on failure.
9055 mdb_drop0(MDB_cursor *mc, int subs)
9059 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9060 if (rc == MDB_SUCCESS) {
9061 MDB_txn *txn = mc->mc_txn;
9066 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9067 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9070 mdb_cursor_copy(mc, &mx);
9071 while (mc->mc_snum > 0) {
9072 MDB_page *mp = mc->mc_pg[mc->mc_top];
9073 unsigned n = NUMKEYS(mp);
9075 for (i=0; i<n; i++) {
9076 ni = NODEPTR(mp, i);
9077 if (ni->mn_flags & F_BIGDATA) {
9080 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9081 rc = mdb_page_get(txn, pg, &omp, NULL);
9084 mdb_cassert(mc, IS_OVERFLOW(omp));
9085 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9089 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9090 mdb_xcursor_init1(mc, ni);
9091 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9097 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9099 for (i=0; i<n; i++) {
9101 ni = NODEPTR(mp, i);
9104 mdb_midl_xappend(txn->mt_free_pgs, pg);
9109 mc->mc_ki[mc->mc_top] = i;
9110 rc = mdb_cursor_sibling(mc, 1);
9112 if (rc != MDB_NOTFOUND)
9114 /* no more siblings, go back to beginning
9115 * of previous level.
9119 for (i=1; i<mc->mc_snum; i++) {
9121 mc->mc_pg[i] = mx.mc_pg[i];
9126 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9129 txn->mt_flags |= MDB_TXN_ERROR;
9130 } else if (rc == MDB_NOTFOUND) {
9136 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9138 MDB_cursor *mc, *m2;
9141 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9144 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9147 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9150 rc = mdb_cursor_open(txn, dbi, &mc);
9154 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9155 /* Invalidate the dropped DB's cursors */
9156 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9157 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9161 /* Can't delete the main DB */
9162 if (del && dbi > MAIN_DBI) {
9163 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9165 txn->mt_dbflags[dbi] = DB_STALE;
9166 mdb_dbi_close(txn->mt_env, dbi);
9168 txn->mt_flags |= MDB_TXN_ERROR;
9171 /* reset the DB record, mark it dirty */
9172 txn->mt_dbflags[dbi] |= DB_DIRTY;
9173 txn->mt_dbs[dbi].md_depth = 0;
9174 txn->mt_dbs[dbi].md_branch_pages = 0;
9175 txn->mt_dbs[dbi].md_leaf_pages = 0;
9176 txn->mt_dbs[dbi].md_overflow_pages = 0;
9177 txn->mt_dbs[dbi].md_entries = 0;
9178 txn->mt_dbs[dbi].md_root = P_INVALID;
9180 txn->mt_flags |= MDB_TXN_DIRTY;
9183 mdb_cursor_close(mc);
9187 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9189 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9192 txn->mt_dbxs[dbi].md_cmp = cmp;
9196 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9198 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9201 txn->mt_dbxs[dbi].md_dcmp = cmp;
9205 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9207 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9210 txn->mt_dbxs[dbi].md_rel = rel;
9214 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9216 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9219 txn->mt_dbxs[dbi].md_relctx = ctx;
9224 mdb_env_get_maxkeysize(MDB_env *env)
9226 return ENV_MAXKEY(env);
9230 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9232 unsigned int i, rdrs;
9235 int rc = 0, first = 1;
9239 if (!env->me_txns) {
9240 return func("(no reader locks)\n", ctx);
9242 rdrs = env->me_txns->mti_numreaders;
9243 mr = env->me_txns->mti_readers;
9244 for (i=0; i<rdrs; i++) {
9246 txnid_t txnid = mr[i].mr_txnid;
9247 sprintf(buf, txnid == (txnid_t)-1 ?
9248 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9249 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9252 rc = func(" pid thread txnid\n", ctx);
9256 rc = func(buf, ctx);
9262 rc = func("(no active readers)\n", ctx);
9267 /** Insert pid into list if not already present.
9268 * return -1 if already present.
9271 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9273 /* binary search of pid in list */
9275 unsigned cursor = 1;
9277 unsigned n = ids[0];
9280 unsigned pivot = n >> 1;
9281 cursor = base + pivot + 1;
9282 val = pid - ids[cursor];
9287 } else if ( val > 0 ) {
9292 /* found, so it's a duplicate */
9301 for (n = ids[0]; n > cursor; n--)
9308 mdb_reader_check(MDB_env *env, int *dead)
9310 unsigned int i, j, rdrs;
9312 MDB_PID_T *pids, pid;
9321 rdrs = env->me_txns->mti_numreaders;
9322 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9326 mr = env->me_txns->mti_readers;
9327 for (i=0; i<rdrs; i++) {
9328 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9330 if (mdb_pid_insert(pids, pid) == 0) {
9331 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9333 /* Recheck, a new process may have reused pid */
9334 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9335 for (j=i; j<rdrs; j++)
9336 if (mr[j].mr_pid == pid) {
9337 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9338 (unsigned) pid, mr[j].mr_txnid));
9343 UNLOCK_MUTEX_R(env);