2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) (CloseHandle(fd) ? 0 : -1)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's datafile format. */
348 #define MDB_DATA_VERSION 1
349 /** The version number for a database's lockfile format. */
350 #define MDB_LOCK_VERSION 1
352 /** @brief The maximum size of a key in the database.
354 * The library rejects bigger keys, and cannot deal with records
355 * with bigger keys stored by a library with bigger max keysize.
357 * We require that keys all fit onto a regular page. This limit
358 * could be raised a bit further if needed; to something just
359 * under #MDB_PAGESIZE / #MDB_MINKEYS.
361 * Note that data items in an #MDB_DUPSORT database are actually keys
362 * of a subDB, so they're also limited to this size.
364 #ifndef MDB_MAXKEYSIZE
365 #define MDB_MAXKEYSIZE 511
368 /** @brief The maximum size of a data item.
370 * We only store a 32 bit value for node sizes.
372 #define MAXDATASIZE 0xffffffffUL
377 * This is used for printing a hex dump of a key's contents.
379 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
380 /** Display a key in hex.
382 * Invoke a function to display a key in hex.
384 #define DKEY(x) mdb_dkey(x, kbuf)
386 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
390 /** An invalid page number.
391 * Mainly used to denote an empty tree.
393 #define P_INVALID (~(pgno_t)0)
395 /** Test if the flags \b f are set in a flag word \b w. */
396 #define F_ISSET(w, f) (((w) & (f)) == (f))
398 /** Used for offsets within a single page.
399 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
402 typedef uint16_t indx_t;
404 /** Default size of memory map.
405 * This is certainly too small for any actual applications. Apps should always set
406 * the size explicitly using #mdb_env_set_mapsize().
408 #define DEFAULT_MAPSIZE 1048576
410 /** @defgroup readers Reader Lock Table
411 * Readers don't acquire any locks for their data access. Instead, they
412 * simply record their transaction ID in the reader table. The reader
413 * mutex is needed just to find an empty slot in the reader table. The
414 * slot's address is saved in thread-specific data so that subsequent read
415 * transactions started by the same thread need no further locking to proceed.
417 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
419 * No reader table is used if the database is on a read-only filesystem.
421 * Since the database uses multi-version concurrency control, readers don't
422 * actually need any locking. This table is used to keep track of which
423 * readers are using data from which old transactions, so that we'll know
424 * when a particular old transaction is no longer in use. Old transactions
425 * that have discarded any data pages can then have those pages reclaimed
426 * for use by a later write transaction.
428 * The lock table is constructed such that reader slots are aligned with the
429 * processor's cache line size. Any slot is only ever used by one thread.
430 * This alignment guarantees that there will be no contention or cache
431 * thrashing as threads update their own slot info, and also eliminates
432 * any need for locking when accessing a slot.
434 * A writer thread will scan every slot in the table to determine the oldest
435 * outstanding reader transaction. Any freed pages older than this will be
436 * reclaimed by the writer. The writer doesn't use any locks when scanning
437 * this table. This means that there's no guarantee that the writer will
438 * see the most up-to-date reader info, but that's not required for correct
439 * operation - all we need is to know the upper bound on the oldest reader,
440 * we don't care at all about the newest reader. So the only consequence of
441 * reading stale information here is that old pages might hang around a
442 * while longer before being reclaimed. That's actually good anyway, because
443 * the longer we delay reclaiming old pages, the more likely it is that a
444 * string of contiguous pages can be found after coalescing old pages from
445 * many old transactions together.
448 /** Number of slots in the reader table.
449 * This value was chosen somewhat arbitrarily. 126 readers plus a
450 * couple mutexes fit exactly into 8KB on my development machine.
451 * Applications should set the table size using #mdb_env_set_maxreaders().
453 #define DEFAULT_READERS 126
455 /** The size of a CPU cache line in bytes. We want our lock structures
456 * aligned to this size to avoid false cache line sharing in the
458 * This value works for most CPUs. For Itanium this should be 128.
464 /** The information we store in a single slot of the reader table.
465 * In addition to a transaction ID, we also record the process and
466 * thread ID that owns a slot, so that we can detect stale information,
467 * e.g. threads or processes that went away without cleaning up.
468 * @note We currently don't check for stale records. We simply re-init
469 * the table when we know that we're the only process opening the
472 typedef struct MDB_rxbody {
473 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
474 * Multiple readers that start at the same time will probably have the
475 * same ID here. Again, it's not important to exclude them from
476 * anything; all we need to know is which version of the DB they
477 * started from so we can avoid overwriting any data used in that
478 * particular version.
481 /** The process ID of the process owning this reader txn. */
483 /** The thread ID of the thread owning this txn. */
487 /** The actual reader record, with cacheline padding. */
488 typedef struct MDB_reader {
491 /** shorthand for mrb_txnid */
492 #define mr_txnid mru.mrx.mrb_txnid
493 #define mr_pid mru.mrx.mrb_pid
494 #define mr_tid mru.mrx.mrb_tid
495 /** cache line alignment */
496 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
500 /** The header for the reader table.
501 * The table resides in a memory-mapped file. (This is a different file
502 * than is used for the main database.)
504 * For POSIX the actual mutexes reside in the shared memory of this
505 * mapped file. On Windows, mutexes are named objects allocated by the
506 * kernel; we store the mutex names in this mapped file so that other
507 * processes can grab them. This same approach is also used on
508 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
509 * process-shared POSIX mutexes. For these cases where a named object
510 * is used, the object name is derived from a 64 bit FNV hash of the
511 * environment pathname. As such, naming collisions are extremely
512 * unlikely. If a collision occurs, the results are unpredictable.
514 typedef struct MDB_txbody {
515 /** Stamp identifying this as an MDB file. It must be set
518 /** Version number of this lock file. Must be set to #MDB_LOCK_VERSION. */
519 uint32_t mtb_version;
520 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
521 char mtb_rmname[MNAME_LEN];
523 /** Mutex protecting access to this table.
524 * This is the reader lock that #LOCK_MUTEX_R acquires.
526 pthread_mutex_t mtb_mutex;
528 /** The ID of the last transaction committed to the database.
529 * This is recorded here only for convenience; the value can always
530 * be determined by reading the main database meta pages.
533 /** The number of slots that have been used in the reader table.
534 * This always records the maximum count, it is not decremented
535 * when readers release their slots.
537 unsigned mtb_numreaders;
540 /** The actual reader table definition. */
541 typedef struct MDB_txninfo {
544 #define mti_magic mt1.mtb.mtb_magic
545 #define mti_version mt1.mtb.mtb_version
546 #define mti_mutex mt1.mtb.mtb_mutex
547 #define mti_rmname mt1.mtb.mtb_rmname
548 #define mti_txnid mt1.mtb.mtb_txnid
549 #define mti_numreaders mt1.mtb.mtb_numreaders
550 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
553 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
554 char mt2_wmname[MNAME_LEN];
555 #define mti_wmname mt2.mt2_wmname
557 pthread_mutex_t mt2_wmutex;
558 #define mti_wmutex mt2.mt2_wmutex
560 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
562 MDB_reader mti_readers[1];
566 /** Common header for all page types.
567 * Overflow records occupy a number of contiguous pages with no
568 * headers on any page after the first.
570 typedef struct MDB_page {
571 #define mp_pgno mp_p.p_pgno
572 #define mp_next mp_p.p_next
574 pgno_t p_pgno; /**< page number */
575 void * p_next; /**< for in-memory list of freed structs */
578 /** @defgroup mdb_page Page Flags
580 * Flags for the page headers.
583 #define P_BRANCH 0x01 /**< branch page */
584 #define P_LEAF 0x02 /**< leaf page */
585 #define P_OVERFLOW 0x04 /**< overflow page */
586 #define P_META 0x08 /**< meta page */
587 #define P_DIRTY 0x10 /**< dirty page */
588 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
589 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
590 #define P_KEEP 0x8000 /**< leave this page alone during spill */
592 uint16_t mp_flags; /**< @ref mdb_page */
593 #define mp_lower mp_pb.pb.pb_lower
594 #define mp_upper mp_pb.pb.pb_upper
595 #define mp_pages mp_pb.pb_pages
598 indx_t pb_lower; /**< lower bound of free space */
599 indx_t pb_upper; /**< upper bound of free space */
601 uint32_t pb_pages; /**< number of overflow pages */
603 indx_t mp_ptrs[1]; /**< dynamic size */
606 /** Size of the page header, excluding dynamic data at the end */
607 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
609 /** Address of first usable data byte in a page, after the header */
610 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
612 /** Number of nodes on a page */
613 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
615 /** The amount of space remaining in the page */
616 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
618 /** The percentage of space used in the page, in tenths of a percent. */
619 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
620 ((env)->me_psize - PAGEHDRSZ))
621 /** The minimum page fill factor, in tenths of a percent.
622 * Pages emptier than this are candidates for merging.
624 #define FILL_THRESHOLD 250
626 /** Test if a page is a leaf page */
627 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
628 /** Test if a page is a LEAF2 page */
629 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
630 /** Test if a page is a branch page */
631 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
632 /** Test if a page is an overflow page */
633 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
634 /** Test if a page is a sub page */
635 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
637 /** The number of overflow pages needed to store the given size. */
638 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
640 /** Header for a single key/data pair within a page.
641 * We guarantee 2-byte alignment for nodes.
643 typedef struct MDB_node {
644 /** lo and hi are used for data size on leaf nodes and for
645 * child pgno on branch nodes. On 64 bit platforms, flags
646 * is also used for pgno. (Branch nodes have no flags).
647 * They are in host byte order in case that lets some
648 * accesses be optimized into a 32-bit word access.
650 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
651 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
652 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
653 /** @defgroup mdb_node Node Flags
655 * Flags for node headers.
658 #define F_BIGDATA 0x01 /**< data put on overflow page */
659 #define F_SUBDATA 0x02 /**< data is a sub-database */
660 #define F_DUPDATA 0x04 /**< data has duplicates */
662 /** valid flags for #mdb_node_add() */
663 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
666 unsigned short mn_flags; /**< @ref mdb_node */
667 unsigned short mn_ksize; /**< key size */
668 char mn_data[1]; /**< key and data are appended here */
671 /** Size of the node header, excluding dynamic data at the end */
672 #define NODESIZE offsetof(MDB_node, mn_data)
674 /** Bit position of top word in page number, for shifting mn_flags */
675 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
677 /** Size of a node in a branch page with a given key.
678 * This is just the node header plus the key, there is no data.
680 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
682 /** Size of a node in a leaf page with a given key and data.
683 * This is node header plus key plus data size.
685 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
687 /** Address of node \b i in page \b p */
688 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
690 /** Address of the key for the node */
691 #define NODEKEY(node) (void *)((node)->mn_data)
693 /** Address of the data for a node */
694 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
696 /** Get the page number pointed to by a branch node */
697 #define NODEPGNO(node) \
698 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
699 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
700 /** Set the page number in a branch node */
701 #define SETPGNO(node,pgno) do { \
702 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
703 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
705 /** Get the size of the data in a leaf node */
706 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
707 /** Set the size of the data for a leaf node */
708 #define SETDSZ(node,size) do { \
709 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
710 /** The size of a key in a node */
711 #define NODEKSZ(node) ((node)->mn_ksize)
713 /** Copy a page number from src to dst */
715 #define COPY_PGNO(dst,src) dst = src
717 #if SIZE_MAX > 4294967295UL
718 #define COPY_PGNO(dst,src) do { \
719 unsigned short *s, *d; \
720 s = (unsigned short *)&(src); \
721 d = (unsigned short *)&(dst); \
728 #define COPY_PGNO(dst,src) do { \
729 unsigned short *s, *d; \
730 s = (unsigned short *)&(src); \
731 d = (unsigned short *)&(dst); \
737 /** The address of a key in a LEAF2 page.
738 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
739 * There are no node headers, keys are stored contiguously.
741 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
743 /** Set the \b node's key into \b key, if requested. */
744 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
745 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
747 /** Information about a single database in the environment. */
748 typedef struct MDB_db {
749 uint32_t md_pad; /**< also ksize for LEAF2 pages */
750 uint16_t md_flags; /**< @ref mdb_dbi_open */
751 uint16_t md_depth; /**< depth of this tree */
752 pgno_t md_branch_pages; /**< number of internal pages */
753 pgno_t md_leaf_pages; /**< number of leaf pages */
754 pgno_t md_overflow_pages; /**< number of overflow pages */
755 size_t md_entries; /**< number of data items */
756 pgno_t md_root; /**< the root page of this tree */
759 /** mdb_dbi_open flags */
760 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
761 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
762 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
763 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
765 /** Handle for the DB used to track free pages. */
767 /** Handle for the default DB. */
770 /** Meta page content. */
771 typedef struct MDB_meta {
772 /** Stamp identifying this as an MDB file. It must be set
775 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
777 void *mm_address; /**< address for fixed mapping */
778 size_t mm_mapsize; /**< size of mmap region */
779 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
780 /** The size of pages used in this DB */
781 #define mm_psize mm_dbs[0].md_pad
782 /** Any persistent environment flags. @ref mdb_env */
783 #define mm_flags mm_dbs[0].md_flags
784 pgno_t mm_last_pg; /**< last used page in file */
785 txnid_t mm_txnid; /**< txnid that committed this page */
788 /** Buffer for a stack-allocated dirty page.
789 * The members define size and alignment, and silence type
790 * aliasing warnings. They are not used directly; that could
791 * mean incorrectly using several union members in parallel.
793 typedef union MDB_pagebuf {
794 char mb_raw[MDB_PAGESIZE];
797 char mm_pad[PAGEHDRSZ];
802 /** Auxiliary DB info.
803 * The information here is mostly static/read-only. There is
804 * only a single copy of this record in the environment.
806 typedef struct MDB_dbx {
807 MDB_val md_name; /**< name of the database */
808 MDB_cmp_func *md_cmp; /**< function for comparing keys */
809 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
810 MDB_rel_func *md_rel; /**< user relocate function */
811 void *md_relctx; /**< user-provided context for md_rel */
814 /** A database transaction.
815 * Every operation requires a transaction handle.
818 MDB_txn *mt_parent; /**< parent of a nested txn */
819 MDB_txn *mt_child; /**< nested txn under this txn */
820 pgno_t mt_next_pgno; /**< next unallocated page */
821 /** The ID of this transaction. IDs are integers incrementing from 1.
822 * Only committed write transactions increment the ID. If a transaction
823 * aborts, the ID may be re-used by the next writer.
826 MDB_env *mt_env; /**< the DB environment */
827 /** The list of pages that became unused during this transaction.
830 /** The list of dirty pages we temporarily wrote to disk
831 * because the dirty list was full.
833 MDB_IDL mt_spill_pgs;
835 MDB_ID2L dirty_list; /**< for write txns: modified pages */
836 MDB_reader *reader; /**< this thread's reader table slot or NULL */
838 /** Array of records for each DB known in the environment. */
840 /** Array of MDB_db records for each known DB */
842 /** @defgroup mt_dbflag Transaction DB Flags
846 #define DB_DIRTY 0x01 /**< DB was written in this txn */
847 #define DB_STALE 0x02 /**< DB record is older than txnID */
848 #define DB_NEW 0x04 /**< DB handle opened in this txn */
849 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
851 /** In write txns, array of cursors for each DB */
852 MDB_cursor **mt_cursors;
853 /** Array of flags for each DB */
854 unsigned char *mt_dbflags;
855 /** Number of DB records in use. This number only ever increments;
856 * we don't decrement it when individual DB handles are closed.
860 /** @defgroup mdb_txn Transaction Flags
864 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
865 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
866 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
867 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
869 unsigned int mt_flags; /**< @ref mdb_txn */
870 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
871 unsigned int mt_dirty_room;
872 /** Tracks which of the two meta pages was used at the start
873 * of this transaction.
875 unsigned int mt_toggle;
878 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
879 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
880 * raise this on a 64 bit machine.
882 #define CURSOR_STACK 32
886 /** Cursors are used for all DB operations */
888 /** Next cursor on this DB in this txn */
890 /** Backup of the original cursor if this cursor is a shadow */
891 MDB_cursor *mc_backup;
892 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
893 struct MDB_xcursor *mc_xcursor;
894 /** The transaction that owns this cursor */
896 /** The database handle this cursor operates on */
898 /** The database record for this cursor */
900 /** The database auxiliary record for this cursor */
902 /** The @ref mt_dbflag for this database */
903 unsigned char *mc_dbflag;
904 unsigned short mc_snum; /**< number of pushed pages */
905 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
906 /** @defgroup mdb_cursor Cursor Flags
908 * Cursor state flags.
911 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
912 #define C_EOF 0x02 /**< No more data */
913 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
914 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
915 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
917 unsigned int mc_flags; /**< @ref mdb_cursor */
918 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
919 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
922 /** Context for sorted-dup records.
923 * We could have gone to a fully recursive design, with arbitrarily
924 * deep nesting of sub-databases. But for now we only handle these
925 * levels - main DB, optional sub-DB, sorted-duplicate DB.
927 typedef struct MDB_xcursor {
928 /** A sub-cursor for traversing the Dup DB */
929 MDB_cursor mx_cursor;
930 /** The database record for this Dup DB */
932 /** The auxiliary DB record for this Dup DB */
934 /** The @ref mt_dbflag for this Dup DB */
935 unsigned char mx_dbflag;
938 /** State of FreeDB old pages, stored in the MDB_env */
939 typedef struct MDB_pgstate {
940 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
941 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
944 /** The database environment. */
946 HANDLE me_fd; /**< The main data file */
947 HANDLE me_lfd; /**< The lock file */
948 HANDLE me_mfd; /**< just for writing the meta pages */
949 /** Failed to update the meta page. Probably an I/O error. */
950 #define MDB_FATAL_ERROR 0x80000000U
951 /** Some fields are initialized. */
952 #define MDB_ENV_ACTIVE 0x20000000U
953 /** me_txkey is set */
954 #define MDB_ENV_TXKEY 0x10000000U
955 /** Have liveness lock in reader table */
956 #define MDB_LIVE_READER 0x08000000U
957 uint32_t me_flags; /**< @ref mdb_env */
958 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
959 unsigned int me_maxreaders; /**< size of the reader table */
960 unsigned int me_numreaders; /**< max numreaders set by this env */
961 MDB_dbi me_numdbs; /**< number of DBs opened */
962 MDB_dbi me_maxdbs; /**< size of the DB table */
963 pid_t me_pid; /**< process ID of this env */
964 char *me_path; /**< path to the DB files */
965 char *me_map; /**< the memory map of the data file */
966 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
967 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
968 MDB_txn *me_txn; /**< current write transaction */
969 size_t me_mapsize; /**< size of the data memory map */
970 off_t me_size; /**< current file size */
971 pgno_t me_maxpg; /**< me_mapsize / me_psize */
972 MDB_dbx *me_dbxs; /**< array of static DB info */
973 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
974 pthread_key_t me_txkey; /**< thread-key for readers */
975 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
976 # define me_pglast me_pgstate.mf_pglast
977 # define me_pghead me_pgstate.mf_pghead
978 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
979 /** IDL of pages that became unused in a write txn */
981 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
982 MDB_ID2L me_dirty_list;
983 /** Max number of freelist items that can fit in a single overflow page */
985 /** Max size of a node on a page */
986 unsigned int me_nodemax;
988 int me_pidquery; /**< Used in OpenProcess */
989 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
991 #elif defined(MDB_USE_POSIX_SEM)
992 sem_t *me_rmutex; /* Shared mutexes are not supported */
997 /** Nested transaction */
998 typedef struct MDB_ntxn {
999 MDB_txn mnt_txn; /* the transaction */
1000 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
1003 /** max number of pages to commit in one writev() call */
1004 #define MDB_COMMIT_PAGES 64
1005 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1006 #undef MDB_COMMIT_PAGES
1007 #define MDB_COMMIT_PAGES IOV_MAX
1010 /* max bytes to write in one call */
1011 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1013 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1014 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1015 static int mdb_page_touch(MDB_cursor *mc);
1017 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1018 static int mdb_page_search_root(MDB_cursor *mc,
1019 MDB_val *key, int modify);
1020 #define MDB_PS_MODIFY 1
1021 #define MDB_PS_ROOTONLY 2
1022 static int mdb_page_search(MDB_cursor *mc,
1023 MDB_val *key, int flags);
1024 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1026 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1027 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1028 pgno_t newpgno, unsigned int nflags);
1030 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1031 static int mdb_env_pick_meta(const MDB_env *env);
1032 static int mdb_env_write_meta(MDB_txn *txn);
1033 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1034 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1036 static void mdb_env_close0(MDB_env *env, int excl);
1038 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1039 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1040 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1041 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1042 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1043 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1044 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1045 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1046 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1048 static int mdb_rebalance(MDB_cursor *mc);
1049 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1051 static void mdb_cursor_pop(MDB_cursor *mc);
1052 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1054 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1055 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1056 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1057 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1058 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1060 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1061 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1063 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1064 static void mdb_xcursor_init0(MDB_cursor *mc);
1065 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1067 static int mdb_drop0(MDB_cursor *mc, int subs);
1068 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1071 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1075 static SECURITY_DESCRIPTOR mdb_null_sd;
1076 static SECURITY_ATTRIBUTES mdb_all_sa;
1077 static int mdb_sec_inited;
1080 /** Return the library version info. */
1082 mdb_version(int *major, int *minor, int *patch)
1084 if (major) *major = MDB_VERSION_MAJOR;
1085 if (minor) *minor = MDB_VERSION_MINOR;
1086 if (patch) *patch = MDB_VERSION_PATCH;
1087 return MDB_VERSION_STRING;
1090 /** Table of descriptions for MDB @ref errors */
1091 static char *const mdb_errstr[] = {
1092 "MDB_KEYEXIST: Key/data pair already exists",
1093 "MDB_NOTFOUND: No matching key/data pair found",
1094 "MDB_PAGE_NOTFOUND: Requested page not found",
1095 "MDB_CORRUPTED: Located page was wrong type",
1096 "MDB_PANIC: Update of meta page failed",
1097 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1098 "MDB_INVALID: File is not an MDB file",
1099 "MDB_MAP_FULL: Environment mapsize limit reached",
1100 "MDB_DBS_FULL: Environment maxdbs limit reached",
1101 "MDB_READERS_FULL: Environment maxreaders limit reached",
1102 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1103 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1104 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1105 "MDB_PAGE_FULL: Internal error - page has no more space",
1106 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1107 "MDB_INCOMPATIBLE: Database flags changed or would change",
1108 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1112 mdb_strerror(int err)
1116 return ("Successful return: 0");
1118 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1119 i = err - MDB_KEYEXIST;
1120 return mdb_errstr[i];
1123 return strerror(err);
1127 /** Display a key in hexadecimal and return the address of the result.
1128 * @param[in] key the key to display
1129 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1130 * @return The key in hexadecimal form.
1133 mdb_dkey(MDB_val *key, char *buf)
1136 unsigned char *c = key->mv_data;
1142 if (key->mv_size > MDB_MAXKEYSIZE)
1143 return "MDB_MAXKEYSIZE";
1144 /* may want to make this a dynamic check: if the key is mostly
1145 * printable characters, print it as-is instead of converting to hex.
1149 for (i=0; i<key->mv_size; i++)
1150 ptr += sprintf(ptr, "%02x", *c++);
1152 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1157 /** Display all the keys in the page. */
1159 mdb_page_list(MDB_page *mp)
1162 unsigned int i, nkeys, nsize;
1166 nkeys = NUMKEYS(mp);
1167 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1168 for (i=0; i<nkeys; i++) {
1169 node = NODEPTR(mp, i);
1170 key.mv_size = node->mn_ksize;
1171 key.mv_data = node->mn_data;
1172 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1173 if (IS_BRANCH(mp)) {
1174 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1177 if (F_ISSET(node->mn_flags, F_BIGDATA))
1178 nsize += sizeof(pgno_t);
1180 nsize += NODEDSZ(node);
1181 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1187 mdb_cursor_chk(MDB_cursor *mc)
1193 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1194 for (i=0; i<mc->mc_top; i++) {
1196 node = NODEPTR(mp, mc->mc_ki[i]);
1197 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1200 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1206 /** Count all the pages in each DB and in the freelist
1207 * and make sure it matches the actual number of pages
1210 static void mdb_audit(MDB_txn *txn)
1214 MDB_ID freecount, count;
1219 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1220 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1221 freecount += *(MDB_ID *)data.mv_data;
1224 for (i = 0; i<txn->mt_numdbs; i++) {
1226 mdb_cursor_init(&mc, txn, i, &mx);
1227 if (txn->mt_dbs[i].md_root == P_INVALID)
1229 count += txn->mt_dbs[i].md_branch_pages +
1230 txn->mt_dbs[i].md_leaf_pages +
1231 txn->mt_dbs[i].md_overflow_pages;
1232 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1233 mdb_page_search(&mc, NULL, 0);
1237 mp = mc.mc_pg[mc.mc_top];
1238 for (j=0; j<NUMKEYS(mp); j++) {
1239 MDB_node *leaf = NODEPTR(mp, j);
1240 if (leaf->mn_flags & F_SUBDATA) {
1242 memcpy(&db, NODEDATA(leaf), sizeof(db));
1243 count += db.md_branch_pages + db.md_leaf_pages +
1244 db.md_overflow_pages;
1248 while (mdb_cursor_sibling(&mc, 1) == 0);
1251 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1252 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1253 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1259 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1261 return txn->mt_dbxs[dbi].md_cmp(a, b);
1265 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1267 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1270 /** Allocate a page.
1271 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1274 mdb_page_malloc(MDB_txn *txn, unsigned num)
1276 MDB_env *env = txn->mt_env;
1277 MDB_page *ret = env->me_dpages;
1278 size_t sz = env->me_psize;
1281 VGMEMP_ALLOC(env, ret, sz);
1282 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1283 env->me_dpages = ret->mp_next;
1289 if ((ret = malloc(sz)) != NULL) {
1290 VGMEMP_ALLOC(env, ret, sz);
1295 /** Free a single page.
1296 * Saves single pages to a list, for future reuse.
1297 * (This is not used for multi-page overflow pages.)
1300 mdb_page_free(MDB_env *env, MDB_page *mp)
1302 mp->mp_next = env->me_dpages;
1303 VGMEMP_FREE(env, mp);
1304 env->me_dpages = mp;
1307 /* Free a dirty page */
1309 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1311 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1312 mdb_page_free(env, dp);
1314 /* large pages just get freed directly */
1315 VGMEMP_FREE(env, dp);
1320 /** Return all dirty pages to dpage list */
1322 mdb_dlist_free(MDB_txn *txn)
1324 MDB_env *env = txn->mt_env;
1325 MDB_ID2L dl = txn->mt_u.dirty_list;
1326 unsigned i, n = dl[0].mid;
1328 for (i = 1; i <= n; i++) {
1329 mdb_dpage_free(env, dl[i].mptr);
1334 /* Set or clear P_KEEP in non-overflow, non-sub pages in known cursors.
1335 * When clearing, only consider backup cursors (from parent txns) since
1336 * other P_KEEP flags have already been cleared.
1337 * @param[in] mc A cursor handle for the current operation.
1338 * @param[in] pflags Flags of the pages to update:
1339 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1342 mdb_cursorpages_mark(MDB_cursor *mc, unsigned pflags)
1344 MDB_txn *txn = mc->mc_txn;
1345 MDB_cursor *m2, *m3;
1349 if (mc->mc_flags & C_UNTRACK)
1350 mc = NULL; /* will find mc in mt_cursors */
1351 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1352 for (; mc; mc=mc->mc_next) {
1353 m2 = pflags == P_DIRTY ? mc : mc->mc_backup;
1354 for (; m2; m2 = m2->mc_backup) {
1355 for (m3=m2; m3->mc_flags & C_INITIALIZED; m3=&mx->mx_cursor) {
1356 for (j=0; j<m3->mc_snum; j++)
1357 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1359 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1360 if (!(m3->mc_db->md_flags & MDB_DUPSORT))
1362 /* Cursor backups have mx malloced at the end of m2 */
1363 mx = (m3 == mc ? m3->mc_xcursor : (MDB_xcursor *)(m3+1));
1372 static int mdb_page_flush(MDB_txn *txn);
1374 /** Spill pages from the dirty list back to disk.
1375 * This is intended to prevent running into #MDB_TXN_FULL situations,
1376 * but note that they may still occur in a few cases:
1377 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1378 * are too many of these dirtied in one txn, the txn may still get
1380 * 2) child txns may run out of space if their parents dirtied a
1381 * lot of pages and never spilled them. TODO: we probably should do
1382 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1383 * the parent's dirty_room is below a given threshold.
1384 * 3) our estimate of the txn size could be too small. At the
1385 * moment this seems unlikely.
1387 * Otherwise, if not using nested txns, it is expected that apps will
1388 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1389 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1390 * If the txn never references them again, they can be left alone.
1391 * If the txn only reads them, they can be used without any fuss.
1392 * If the txn writes them again, they can be dirtied immediately without
1393 * going thru all of the work of #mdb_page_touch(). Such references are
1394 * handled by #mdb_page_unspill().
1396 * Also note, we never spill DB root pages, nor pages of active cursors,
1397 * because we'll need these back again soon anyway. And in nested txns,
1398 * we can't spill a page in a child txn if it was already spilled in a
1399 * parent txn. That would alter the parent txns' data even though
1400 * the child hasn't committed yet, and we'd have no way to undo it if
1401 * the child aborted.
1403 * @param[in] m0 cursor A cursor handle identifying the transaction and
1404 * database for which we are checking space.
1405 * @param[in] key For a put operation, the key being stored.
1406 * @param[in] data For a put operation, the data being stored.
1407 * @return 0 on success, non-zero on failure.
1410 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1412 MDB_txn *txn = m0->mc_txn;
1414 MDB_ID2L dl = txn->mt_u.dirty_list;
1418 if (m0->mc_flags & C_SUB)
1421 /* Estimate how much space this op will take */
1422 i = m0->mc_db->md_depth;
1423 /* Named DBs also dirty the main DB */
1424 if (m0->mc_dbi > MAIN_DBI)
1425 i += txn->mt_dbs[MAIN_DBI].md_depth;
1426 /* For puts, roughly factor in the key+data size */
1428 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1429 i += i; /* double it for good measure */
1431 if (txn->mt_dirty_room > i)
1434 if (!txn->mt_spill_pgs) {
1435 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1436 if (!txn->mt_spill_pgs)
1440 /* Mark all the dirty root pages we want to preserve */
1441 for (i=0; i<txn->mt_numdbs; i++) {
1442 if (txn->mt_dbflags[i] & DB_DIRTY) {
1443 j = mdb_mid2l_search(dl, txn->mt_dbs[i].md_root);
1444 if (j <= dl[0].mid) {
1446 dp->mp_flags |= P_KEEP;
1451 /* Preserve pages used by cursors */
1452 mdb_cursorpages_mark(m0, P_DIRTY);
1454 /* Save the page IDs of all the pages we're flushing */
1455 for (i=1; i<=dl[0].mid; i++) {
1457 if (dp->mp_flags & P_KEEP)
1459 /* Can't spill twice, make sure it's not already in a parent's
1462 if (txn->mt_parent) {
1464 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1465 if (tx2->mt_spill_pgs) {
1466 j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
1467 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
1468 dp->mp_flags |= P_KEEP;
1476 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
1479 mdb_midl_sort(txn->mt_spill_pgs);
1481 rc = mdb_page_flush(txn);
1483 mdb_cursorpages_mark(m0, P_DIRTY|P_KEEP);
1486 if (txn->mt_parent) {
1488 pgno_t pgno = dl[i].mid;
1489 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1490 /* dirty pages that are dirty in an ancestor don't
1491 * count against this txn's dirty_room.
1493 for (i=1; i<=dl[0].mid; i++) {
1494 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1495 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1496 if (j <= tx2->mt_u.dirty_list[0].mid &&
1497 tx2->mt_u.dirty_list[j].mid == pgno) {
1498 txn->mt_dirty_room++;
1504 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1506 txn->mt_flags |= MDB_TXN_SPILLS;
1511 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1513 mdb_find_oldest(MDB_txn *txn)
1516 txnid_t mr, oldest = txn->mt_txnid - 1;
1517 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1518 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1528 /** Add a page to the txn's dirty list */
1530 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1533 int (*insert)(MDB_ID2L, MDB_ID2 *);
1535 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1536 insert = mdb_mid2l_append;
1538 insert = mdb_mid2l_insert;
1540 mid.mid = mp->mp_pgno;
1542 insert(txn->mt_u.dirty_list, &mid);
1543 txn->mt_dirty_room--;
1546 /** Allocate pages for writing.
1547 * If there are free pages available from older transactions, they
1548 * will be re-used first. Otherwise a new page will be allocated.
1549 * @param[in] mc cursor A cursor handle identifying the transaction and
1550 * database for which we are allocating.
1551 * @param[in] num the number of pages to allocate.
1552 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1553 * will always be satisfied by a single contiguous chunk of memory.
1554 * @return 0 on success, non-zero on failure.
1557 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1559 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1560 /* Get at most <Max_retries> more freeDB records once me_pghead
1561 * has enough pages. If not enough, use new pages from the map.
1562 * If <Paranoid> and mc is updating the freeDB, only get new
1563 * records if me_pghead is empty. Then the freelist cannot play
1564 * catch-up with itself by growing while trying to save it.
1566 enum { Paranoid = 1, Max_retries = 500 };
1568 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1570 int rc, n2 = num-1, retry = Max_retries;
1571 MDB_txn *txn = mc->mc_txn;
1572 MDB_env *env = txn->mt_env;
1573 pgno_t pgno, *mop = env->me_pghead;
1574 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1576 txnid_t oldest = 0, last;
1582 /* If our dirty list is already full, we can't do anything */
1583 if (txn->mt_dirty_room == 0)
1584 return MDB_TXN_FULL;
1586 for (op = MDB_FIRST;; op = MDB_NEXT) {
1589 pgno_t *idl, old_id, new_id;
1591 /* Seek a big enough contiguous page range. Prefer
1592 * pages at the tail, just truncating the list.
1594 if (mop_len >= (unsigned)num) {
1598 if (mop[i-n2] == pgno+n2)
1600 } while (--i >= (unsigned)num);
1601 if (Max_retries < INT_MAX && --retry < 0)
1605 if (op == MDB_FIRST) { /* 1st iteration */
1606 /* Prepare to fetch more and coalesce */
1607 oldest = mdb_find_oldest(txn);
1608 last = env->me_pglast;
1609 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1612 key.mv_data = &last; /* will loop up last+1 */
1613 key.mv_size = sizeof(last);
1615 if (Paranoid && mc->mc_dbi == FREE_DBI)
1618 if (Paranoid && retry < 0 && mop_len)
1622 /* Do not fetch more if the record will be too recent */
1625 rc = mdb_cursor_get(&m2, &key, NULL, op);
1627 if (rc == MDB_NOTFOUND)
1631 last = *(txnid_t*)key.mv_data;
1634 np = m2.mc_pg[m2.mc_top];
1635 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1636 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1639 idl = (MDB_ID *) data.mv_data;
1642 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1645 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1647 mop = env->me_pghead;
1649 env->me_pglast = last;
1651 DPRINTF("IDL read txn %zu root %zu num %u",
1652 last, txn->mt_dbs[FREE_DBI].md_root, i);
1654 DPRINTF("IDL %zu", idl[k]);
1656 /* Merge in descending sorted order */
1659 mop[0] = (pgno_t)-1;
1663 for (; old_id < new_id; old_id = mop[--j])
1670 /* Use new pages from the map when nothing suitable in the freeDB */
1672 pgno = txn->mt_next_pgno;
1673 if (pgno + num >= env->me_maxpg) {
1674 DPUTS("DB size maxed out");
1675 return MDB_MAP_FULL;
1679 if (env->me_flags & MDB_WRITEMAP) {
1680 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1682 if (!(np = mdb_page_malloc(txn, num)))
1686 mop[0] = mop_len -= num;
1687 /* Move any stragglers down */
1688 for (j = i-num; j < mop_len; )
1689 mop[++j] = mop[++i];
1691 txn->mt_next_pgno = pgno + num;
1694 mdb_page_dirty(txn, np);
1700 /** Copy the used portions of a non-overflow page.
1701 * @param[in] dst page to copy into
1702 * @param[in] src page to copy from
1703 * @param[in] psize size of a page
1706 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1708 enum { Align = sizeof(pgno_t) };
1709 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1711 /* If page isn't full, just copy the used portion. Adjust
1712 * alignment so memcpy may copy words instead of bytes.
1714 if ((unused &= -Align) && !IS_LEAF2(src)) {
1716 memcpy(dst, src, (lower + (Align-1)) & -Align);
1717 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1720 memcpy(dst, src, psize - unused);
1724 /** Pull a page off the txn's spill list, if present.
1725 * If a page being referenced was spilled to disk in this txn, bring
1726 * it back and make it dirty/writable again.
1727 * @param[in] tx0 the transaction handle.
1728 * @param[in] mp the page being referenced.
1729 * @param[out] ret the writable page, if any. ret is unchanged if
1730 * mp wasn't spilled.
1733 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1735 MDB_env *env = tx0->mt_env;
1738 pgno_t pgno = mp->mp_pgno;
1740 for (txn = tx0; txn; txn=txn->mt_parent) {
1741 if (!txn->mt_spill_pgs)
1743 x = mdb_midl_search(txn->mt_spill_pgs, pgno);
1744 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
1747 if (IS_OVERFLOW(mp))
1751 if (env->me_flags & MDB_WRITEMAP) {
1754 np = mdb_page_malloc(txn, num);
1758 memcpy(np, mp, num * env->me_psize);
1760 mdb_page_copy(np, mp, env->me_psize);
1763 /* If in current txn, this page is no longer spilled */
1764 for (; x < txn->mt_spill_pgs[0]; x++)
1765 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
1766 txn->mt_spill_pgs[0]--;
1767 } /* otherwise, if belonging to a parent txn, the
1768 * page remains spilled until child commits
1771 if (txn->mt_parent) {
1773 /* If this page is also in a parent's dirty list, then
1774 * it's already accounted in dirty_room, and we need to
1775 * cancel out the decrement that mdb_page_dirty does.
1777 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1778 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1779 if (x <= tx2->mt_u.dirty_list[0].mid &&
1780 tx2->mt_u.dirty_list[x].mid == pgno) {
1781 txn->mt_dirty_room++;
1786 mdb_page_dirty(tx0, np);
1787 np->mp_flags |= P_DIRTY;
1795 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1796 * @param[in] mc cursor pointing to the page to be touched
1797 * @return 0 on success, non-zero on failure.
1800 mdb_page_touch(MDB_cursor *mc)
1802 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1803 MDB_txn *txn = mc->mc_txn;
1804 MDB_cursor *m2, *m3;
1809 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1810 if (txn->mt_flags & MDB_TXN_SPILLS) {
1812 rc = mdb_page_unspill(txn, mp, &np);
1818 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1819 (rc = mdb_page_alloc(mc, 1, &np)))
1822 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1823 assert(mp->mp_pgno != pgno);
1824 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1825 /* Update the parent page, if any, to point to the new page */
1827 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1828 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1829 SETPGNO(node, pgno);
1831 mc->mc_db->md_root = pgno;
1833 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1834 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1836 /* If txn has a parent, make sure the page is in our
1840 unsigned x = mdb_mid2l_search(dl, pgno);
1841 if (x <= dl[0].mid && dl[x].mid == pgno) {
1842 if (mp != dl[x].mptr) { /* bad cursor? */
1843 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1844 return MDB_CORRUPTED;
1849 assert(dl[0].mid < MDB_IDL_UM_MAX);
1851 np = mdb_page_malloc(txn, 1);
1856 mdb_mid2l_insert(dl, &mid);
1861 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1863 np->mp_flags |= P_DIRTY;
1866 /* Adjust cursors pointing to mp */
1867 mc->mc_pg[mc->mc_top] = np;
1869 if (mc->mc_flags & C_SUB) {
1871 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1872 m3 = &m2->mc_xcursor->mx_cursor;
1873 if (m3->mc_snum < mc->mc_snum) continue;
1874 if (m3->mc_pg[mc->mc_top] == mp)
1875 m3->mc_pg[mc->mc_top] = np;
1878 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1879 if (m2->mc_snum < mc->mc_snum) continue;
1880 if (m2->mc_pg[mc->mc_top] == mp) {
1881 m2->mc_pg[mc->mc_top] = np;
1882 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1883 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1885 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1886 if (!(leaf->mn_flags & F_SUBDATA))
1887 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1896 mdb_env_sync(MDB_env *env, int force)
1899 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1900 if (env->me_flags & MDB_WRITEMAP) {
1901 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1902 ? MS_ASYNC : MS_SYNC;
1903 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1906 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1910 if (MDB_FDATASYNC(env->me_fd))
1917 /** Back up parent txn's cursors, then grab the originals for tracking */
1919 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1921 MDB_cursor *mc, *bk;
1926 for (i = src->mt_numdbs; --i >= 0; ) {
1927 if ((mc = src->mt_cursors[i]) != NULL) {
1928 size = sizeof(MDB_cursor);
1930 size += sizeof(MDB_xcursor);
1931 for (; mc; mc = bk->mc_next) {
1937 mc->mc_db = &dst->mt_dbs[i];
1938 /* Kill pointers into src - and dst to reduce abuse: The
1939 * user may not use mc until dst ends. Otherwise we'd...
1941 mc->mc_txn = NULL; /* ...set this to dst */
1942 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1943 if ((mx = mc->mc_xcursor) != NULL) {
1944 *(MDB_xcursor *)(bk+1) = *mx;
1945 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
1947 mc->mc_next = dst->mt_cursors[i];
1948 dst->mt_cursors[i] = mc;
1955 /** Close this write txn's cursors, give parent txn's cursors back to parent.
1956 * @param[in] txn the transaction handle.
1957 * @param[in] merge true to keep changes to parent cursors, false to revert.
1958 * @return 0 on success, non-zero on failure.
1961 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1963 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
1967 for (i = txn->mt_numdbs; --i >= 0; ) {
1968 for (mc = cursors[i]; mc; mc = next) {
1970 if ((bk = mc->mc_backup) != NULL) {
1972 /* Commit changes to parent txn */
1973 mc->mc_next = bk->mc_next;
1974 mc->mc_backup = bk->mc_backup;
1975 mc->mc_txn = bk->mc_txn;
1976 mc->mc_db = bk->mc_db;
1977 mc->mc_dbflag = bk->mc_dbflag;
1978 if ((mx = mc->mc_xcursor) != NULL)
1979 mx->mx_cursor.mc_txn = bk->mc_txn;
1981 /* Abort nested txn */
1983 if ((mx = mc->mc_xcursor) != NULL)
1984 *mx = *(MDB_xcursor *)(bk+1);
1994 #ifdef MDB_DEBUG_SKIP
1995 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
1998 mdb_txn_reset0(MDB_txn *txn, const char *act);
2006 Pidset = F_SETLK, Pidcheck = F_GETLK
2010 /** Set or check a pid lock. Set returns 0 on success.
2011 * Check returns 0 if lock exists (meaning the process is alive).
2013 * On Windows Pidset is a no-op, we merely check for the existence
2014 * of the process with the given pid. On POSIX we use a single byte
2015 * lock on the lockfile, set at an offset equal to the pid.
2018 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2027 h = OpenProcess(env->me_pidquery, FALSE, pid);
2029 return GetLastError();
2036 struct flock lock_info;
2037 memset((void *)&lock_info, 0, sizeof(lock_info));
2038 lock_info.l_type = F_WRLCK;
2039 lock_info.l_whence = SEEK_SET;
2040 lock_info.l_start = pid;
2041 lock_info.l_len = 1;
2042 while ((rc = fcntl(env->me_lfd, op, &lock_info)) &&
2043 (rc = ErrCode()) == EINTR) ;
2044 if (op == F_GETLK && rc == 0 && lock_info.l_type == F_UNLCK)
2050 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2051 * @param[in] txn the transaction handle to initialize
2052 * @return 0 on success, non-zero on failure.
2055 mdb_txn_renew0(MDB_txn *txn)
2057 MDB_env *env = txn->mt_env;
2060 int rc, new_notls = 0;
2063 txn->mt_numdbs = env->me_numdbs;
2064 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2066 if (txn->mt_flags & MDB_TXN_RDONLY) {
2067 if (!env->me_txns) {
2068 i = mdb_env_pick_meta(env);
2069 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2070 txn->mt_u.reader = NULL;
2072 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2073 pthread_getspecific(env->me_txkey);
2075 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2076 return MDB_BAD_RSLOT;
2078 pid_t pid = env->me_pid;
2079 pthread_t tid = pthread_self();
2082 for (i=0; i<env->me_txns->mti_numreaders; i++)
2083 if (env->me_txns->mti_readers[i].mr_pid == 0)
2085 if (i == env->me_maxreaders) {
2086 UNLOCK_MUTEX_R(env);
2087 return MDB_READERS_FULL;
2089 if (!(env->me_flags & MDB_LIVE_READER)) {
2090 rc = mdb_reader_pid(env, Pidset, pid);
2092 UNLOCK_MUTEX_R(env);
2095 env->me_flags |= MDB_LIVE_READER;
2097 env->me_txns->mti_readers[i].mr_pid = pid;
2098 env->me_txns->mti_readers[i].mr_tid = tid;
2099 if (i >= env->me_txns->mti_numreaders)
2100 env->me_txns->mti_numreaders = i+1;
2101 /* Save numreaders for un-mutexed mdb_env_close() */
2102 env->me_numreaders = env->me_txns->mti_numreaders;
2103 UNLOCK_MUTEX_R(env);
2104 r = &env->me_txns->mti_readers[i];
2105 new_notls = (env->me_flags & MDB_NOTLS);
2106 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2111 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2112 txn->mt_u.reader = r;
2114 txn->mt_toggle = txn->mt_txnid & 1;
2118 txn->mt_txnid = env->me_txns->mti_txnid;
2119 txn->mt_toggle = txn->mt_txnid & 1;
2122 if (txn->mt_txnid == mdb_debug_start)
2125 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2126 txn->mt_u.dirty_list = env->me_dirty_list;
2127 txn->mt_u.dirty_list[0].mid = 0;
2128 txn->mt_free_pgs = env->me_free_pgs;
2129 txn->mt_free_pgs[0] = 0;
2130 txn->mt_spill_pgs = NULL;
2134 /* Copy the DB info and flags */
2135 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2137 /* Moved to here to avoid a data race in read TXNs */
2138 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2140 for (i=2; i<txn->mt_numdbs; i++) {
2141 x = env->me_dbflags[i];
2142 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2143 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2145 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2147 if (env->me_maxpg < txn->mt_next_pgno) {
2148 mdb_txn_reset0(txn, "renew0-mapfail");
2150 txn->mt_u.reader->mr_pid = 0;
2151 txn->mt_u.reader = NULL;
2153 return MDB_MAP_RESIZED;
2160 mdb_txn_renew(MDB_txn *txn)
2164 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2167 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2168 DPUTS("environment had fatal error, must shutdown!");
2172 rc = mdb_txn_renew0(txn);
2173 if (rc == MDB_SUCCESS) {
2174 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
2175 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2176 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2182 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2186 int rc, size, tsize = sizeof(MDB_txn);
2188 if (env->me_flags & MDB_FATAL_ERROR) {
2189 DPUTS("environment had fatal error, must shutdown!");
2192 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2195 /* Nested transactions: Max 1 child, write txns only, no writemap */
2196 if (parent->mt_child ||
2197 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
2198 (env->me_flags & MDB_WRITEMAP))
2202 tsize = sizeof(MDB_ntxn);
2204 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2205 if (!(flags & MDB_RDONLY))
2206 size += env->me_maxdbs * sizeof(MDB_cursor *);
2208 if ((txn = calloc(1, size)) == NULL) {
2209 DPRINTF("calloc: %s", strerror(ErrCode()));
2212 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2213 if (flags & MDB_RDONLY) {
2214 txn->mt_flags |= MDB_TXN_RDONLY;
2215 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2217 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2218 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2224 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2225 if (!txn->mt_u.dirty_list ||
2226 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2228 free(txn->mt_u.dirty_list);
2232 txn->mt_txnid = parent->mt_txnid;
2233 txn->mt_toggle = parent->mt_toggle;
2234 txn->mt_dirty_room = parent->mt_dirty_room;
2235 txn->mt_u.dirty_list[0].mid = 0;
2236 txn->mt_spill_pgs = NULL;
2237 txn->mt_next_pgno = parent->mt_next_pgno;
2238 parent->mt_child = txn;
2239 txn->mt_parent = parent;
2240 txn->mt_numdbs = parent->mt_numdbs;
2241 txn->mt_flags = parent->mt_flags;
2242 txn->mt_dbxs = parent->mt_dbxs;
2243 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2244 /* Copy parent's mt_dbflags, but clear DB_NEW */
2245 for (i=0; i<txn->mt_numdbs; i++)
2246 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2248 ntxn = (MDB_ntxn *)txn;
2249 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2250 if (env->me_pghead) {
2251 size = MDB_IDL_SIZEOF(env->me_pghead);
2252 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2254 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2259 rc = mdb_cursor_shadow(parent, txn);
2261 mdb_txn_reset0(txn, "beginchild-fail");
2263 rc = mdb_txn_renew0(txn);
2269 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2270 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2271 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2277 /** Export or close DBI handles opened in this txn. */
2279 mdb_dbis_update(MDB_txn *txn, int keep)
2282 MDB_dbi n = txn->mt_numdbs;
2283 MDB_env *env = txn->mt_env;
2284 unsigned char *tdbflags = txn->mt_dbflags;
2286 for (i = n; --i >= 2;) {
2287 if (tdbflags[i] & DB_NEW) {
2289 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2291 char *ptr = env->me_dbxs[i].md_name.mv_data;
2292 env->me_dbxs[i].md_name.mv_data = NULL;
2293 env->me_dbxs[i].md_name.mv_size = 0;
2294 env->me_dbflags[i] = 0;
2299 if (keep && env->me_numdbs < n)
2303 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2304 * May be called twice for readonly txns: First reset it, then abort.
2305 * @param[in] txn the transaction handle to reset
2308 mdb_txn_reset0(MDB_txn *txn, const char *act)
2310 MDB_env *env = txn->mt_env;
2312 /* Close any DBI handles opened in this txn */
2313 mdb_dbis_update(txn, 0);
2315 DPRINTF("%s txn %zu%c %p on mdbenv %p, root page %zu",
2316 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2317 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2319 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2320 if (txn->mt_u.reader) {
2321 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2322 if (!(env->me_flags & MDB_NOTLS))
2323 txn->mt_u.reader = NULL; /* txn does not own reader */
2325 txn->mt_numdbs = 0; /* close nothing if called again */
2326 txn->mt_dbxs = NULL; /* mark txn as reset */
2328 mdb_cursors_close(txn, 0);
2330 if (!(env->me_flags & MDB_WRITEMAP)) {
2331 mdb_dlist_free(txn);
2333 mdb_midl_free(env->me_pghead);
2335 if (txn->mt_parent) {
2336 txn->mt_parent->mt_child = NULL;
2337 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2338 mdb_midl_free(txn->mt_free_pgs);
2339 mdb_midl_free(txn->mt_spill_pgs);
2340 free(txn->mt_u.dirty_list);
2344 if (mdb_midl_shrink(&txn->mt_free_pgs))
2345 env->me_free_pgs = txn->mt_free_pgs;
2346 env->me_pghead = NULL;
2350 /* The writer mutex was locked in mdb_txn_begin. */
2351 UNLOCK_MUTEX_W(env);
2356 mdb_txn_reset(MDB_txn *txn)
2361 /* This call is only valid for read-only txns */
2362 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2365 mdb_txn_reset0(txn, "reset");
2369 mdb_txn_abort(MDB_txn *txn)
2375 mdb_txn_abort(txn->mt_child);
2377 mdb_txn_reset0(txn, "abort");
2378 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2379 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2380 txn->mt_u.reader->mr_pid = 0;
2385 /** Save the freelist as of this transaction to the freeDB.
2386 * This changes the freelist. Keep trying until it stabilizes.
2389 mdb_freelist_save(MDB_txn *txn)
2391 /* env->me_pghead[] can grow and shrink during this call.
2392 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2393 * Page numbers cannot disappear from txn->mt_free_pgs[].
2396 MDB_env *env = txn->mt_env;
2397 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2398 txnid_t pglast = 0, head_id = 0;
2399 pgno_t freecnt = 0, *free_pgs, *mop;
2400 ssize_t head_room = 0, total_room = 0, mop_len;
2402 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2404 if (env->me_pghead) {
2405 /* Make sure first page of freeDB is touched and on freelist */
2406 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2407 if (rc && rc != MDB_NOTFOUND)
2412 /* Come back here after each Put() in case freelist changed */
2415 /* If using records from freeDB which we have not yet
2416 * deleted, delete them and any we reserved for me_pghead.
2418 while (pglast < env->me_pglast) {
2419 rc = mdb_cursor_first(&mc, &key, NULL);
2422 pglast = head_id = *(txnid_t *)key.mv_data;
2423 total_room = head_room = 0;
2424 assert(pglast <= env->me_pglast);
2425 rc = mdb_cursor_del(&mc, 0);
2430 /* Save the IDL of pages freed by this txn, to a single record */
2431 if (freecnt < txn->mt_free_pgs[0]) {
2433 /* Make sure last page of freeDB is touched and on freelist */
2434 key.mv_size = MDB_MAXKEYSIZE+1;
2436 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2437 if (rc && rc != MDB_NOTFOUND)
2440 free_pgs = txn->mt_free_pgs;
2441 /* Write to last page of freeDB */
2442 key.mv_size = sizeof(txn->mt_txnid);
2443 key.mv_data = &txn->mt_txnid;
2445 freecnt = free_pgs[0];
2446 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2447 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2450 /* Retry if mt_free_pgs[] grew during the Put() */
2451 free_pgs = txn->mt_free_pgs;
2452 } while (freecnt < free_pgs[0]);
2453 mdb_midl_sort(free_pgs);
2454 memcpy(data.mv_data, free_pgs, data.mv_size);
2457 unsigned int i = free_pgs[0];
2458 DPRINTF("IDL write txn %zu root %zu num %u",
2459 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2461 DPRINTF("IDL %zu", free_pgs[i]);
2467 mop = env->me_pghead;
2468 mop_len = mop ? mop[0] : 0;
2470 /* Reserve records for me_pghead[]. Split it if multi-page,
2471 * to avoid searching freeDB for a page range. Use keys in
2472 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2474 if (total_room >= mop_len) {
2475 if (total_room == mop_len || --more < 0)
2477 } else if (head_room >= maxfree_1pg && head_id > 1) {
2478 /* Keep current record (overflow page), add a new one */
2482 /* (Re)write {key = head_id, IDL length = head_room} */
2483 total_room -= head_room;
2484 head_room = mop_len - total_room;
2485 if (head_room > maxfree_1pg && head_id > 1) {
2486 /* Overflow multi-page for part of me_pghead */
2487 head_room /= head_id; /* amortize page sizes */
2488 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2489 } else if (head_room < 0) {
2490 /* Rare case, not bothering to delete this record */
2493 key.mv_size = sizeof(head_id);
2494 key.mv_data = &head_id;
2495 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2496 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2499 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2500 total_room += head_room;
2503 /* Fill in the reserved, touched me_pghead records */
2509 rc = mdb_cursor_first(&mc, &key, &data);
2510 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2511 unsigned flags = MDB_CURRENT;
2512 txnid_t id = *(txnid_t *)key.mv_data;
2513 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2516 assert(len >= 0 && id <= env->me_pglast);
2518 if (len > mop_len) {
2520 data.mv_size = (len + 1) * sizeof(MDB_ID);
2523 data.mv_data = mop -= len;
2526 rc = mdb_cursor_put(&mc, &key, &data, flags);
2528 if (rc || !(mop_len -= len))
2535 /** Flush dirty pages to the map, after clearing their dirty flag.
2538 mdb_page_flush(MDB_txn *txn)
2540 MDB_env *env = txn->mt_env;
2541 MDB_ID2L dl = txn->mt_u.dirty_list;
2542 unsigned psize = env->me_psize, j;
2543 int i, pagecount = dl[0].mid, rc;
2544 size_t size = 0, pos = 0;
2546 MDB_page *dp = NULL;
2550 struct iovec iov[MDB_COMMIT_PAGES];
2551 ssize_t wpos = 0, wsize = 0, wres;
2552 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2557 if (env->me_flags & MDB_WRITEMAP) {
2558 /* Clear dirty flags */
2559 for (i = pagecount; i; i--) {
2561 /* Don't flush this page yet */
2562 if (dp->mp_flags & P_KEEP) {
2563 dp->mp_flags ^= P_KEEP;
2567 dp->mp_flags &= ~P_DIRTY;
2573 /* Write the pages */
2575 if (i <= pagecount) {
2577 /* Don't flush this page yet */
2578 if (dp->mp_flags & P_KEEP) {
2579 dp->mp_flags ^= P_KEEP;
2584 /* clear dirty flag */
2585 dp->mp_flags &= ~P_DIRTY;
2588 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2593 /* Windows actually supports scatter/gather I/O, but only on
2594 * unbuffered file handles. Since we're relying on the OS page
2595 * cache for all our data, that's self-defeating. So we just
2596 * write pages one at a time. We use the ov structure to set
2597 * the write offset, to at least save the overhead of a Seek
2600 DPRINTF("committing page %zu", pgno);
2601 memset(&ov, 0, sizeof(ov));
2602 ov.Offset = pos & 0xffffffff;
2603 ov.OffsetHigh = pos >> 16 >> 16;
2604 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2606 DPRINTF("WriteFile: %d", rc);
2610 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2611 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2613 /* Write previous page(s) */
2614 #ifdef MDB_USE_PWRITEV
2615 wres = pwritev(env->me_fd, iov, n, wpos);
2618 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2620 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2622 DPRINTF("lseek: %s", strerror(rc));
2625 wres = writev(env->me_fd, iov, n);
2628 if (wres != wsize) {
2631 DPRINTF("Write error: %s", strerror(rc));
2633 rc = EIO; /* TODO: Use which error code? */
2634 DPUTS("short write, filesystem full?");
2645 DPRINTF("committing page %zu", pgno);
2646 next_pos = pos + size;
2647 iov[n].iov_len = size;
2648 iov[n].iov_base = (char *)dp;
2655 for (i=1; i<=pagecount; i++) {
2657 /* This is a page we skipped above */
2660 dl[j].mid = dp->mp_pgno;
2663 mdb_dpage_free(env, dp);
2671 mdb_txn_commit(MDB_txn *txn)
2677 assert(txn != NULL);
2678 assert(txn->mt_env != NULL);
2680 if (txn->mt_child) {
2681 rc = mdb_txn_commit(txn->mt_child);
2682 txn->mt_child = NULL;
2689 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2690 mdb_dbis_update(txn, 1);
2691 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2696 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2697 DPUTS("error flag is set, can't commit");
2699 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2704 if (txn->mt_parent) {
2705 MDB_txn *parent = txn->mt_parent;
2709 /* Append our free list to parent's */
2710 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2713 mdb_midl_free(txn->mt_free_pgs);
2715 parent->mt_next_pgno = txn->mt_next_pgno;
2716 parent->mt_flags = txn->mt_flags;
2718 /* Merge our cursors into parent's and close them */
2719 mdb_cursors_close(txn, 1);
2721 /* Update parent's DB table. */
2722 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2723 parent->mt_numdbs = txn->mt_numdbs;
2724 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2725 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2726 for (i=2; i<txn->mt_numdbs; i++) {
2727 /* preserve parent's DB_NEW status */
2728 x = parent->mt_dbflags[i] & DB_NEW;
2729 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2732 dst = parent->mt_u.dirty_list;
2733 src = txn->mt_u.dirty_list;
2734 /* Remove anything in our dirty list from parent's spill list */
2735 if (parent->mt_spill_pgs) {
2736 x = parent->mt_spill_pgs[0];
2738 /* zero out our dirty pages in parent spill list */
2739 for (i=1; i<=src[0].mid; i++) {
2740 if (src[i].mid < parent->mt_spill_pgs[x])
2742 if (src[i].mid > parent->mt_spill_pgs[x]) {
2748 parent->mt_spill_pgs[x] = 0;
2751 /* OK, we had a few hits, squash zeros from the spill list */
2752 if (len < parent->mt_spill_pgs[0]) {
2754 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2755 if (parent->mt_spill_pgs[y]) {
2757 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2762 parent->mt_spill_pgs[0] = len;
2765 /* Find len = length of merging our dirty list with parent's */
2767 dst[0].mid = 0; /* simplify loops */
2768 if (parent->mt_parent) {
2769 len = x + src[0].mid;
2770 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2771 for (i = x; y && i; y--) {
2772 pgno_t yp = src[y].mid;
2773 while (yp < dst[i].mid)
2775 if (yp == dst[i].mid) {
2780 } else { /* Simplify the above for single-ancestor case */
2781 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2783 /* Merge our dirty list with parent's */
2785 for (i = len; y; dst[i--] = src[y--]) {
2786 pgno_t yp = src[y].mid;
2787 while (yp < dst[x].mid)
2788 dst[i--] = dst[x--];
2789 if (yp == dst[x].mid)
2790 free(dst[x--].mptr);
2794 free(txn->mt_u.dirty_list);
2795 parent->mt_dirty_room = txn->mt_dirty_room;
2796 if (txn->mt_spill_pgs) {
2797 if (parent->mt_spill_pgs) {
2798 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2799 mdb_midl_free(txn->mt_spill_pgs);
2800 mdb_midl_sort(parent->mt_spill_pgs);
2802 parent->mt_spill_pgs = txn->mt_spill_pgs;
2806 parent->mt_child = NULL;
2807 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2812 if (txn != env->me_txn) {
2813 DPUTS("attempt to commit unknown transaction");
2818 mdb_cursors_close(txn, 0);
2820 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2823 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2824 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2826 /* Update DB root pointers */
2827 if (txn->mt_numdbs > 2) {
2831 data.mv_size = sizeof(MDB_db);
2833 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2834 for (i = 2; i < txn->mt_numdbs; i++) {
2835 if (txn->mt_dbflags[i] & DB_DIRTY) {
2836 data.mv_data = &txn->mt_dbs[i];
2837 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2844 rc = mdb_freelist_save(txn);
2848 mdb_midl_free(env->me_pghead);
2849 env->me_pghead = NULL;
2850 if (mdb_midl_shrink(&txn->mt_free_pgs))
2851 env->me_free_pgs = txn->mt_free_pgs;
2857 if ((rc = mdb_page_flush(txn)) ||
2858 (rc = mdb_env_sync(env, 0)) ||
2859 (rc = mdb_env_write_meta(txn)))
2865 mdb_dbis_update(txn, 1);
2867 UNLOCK_MUTEX_W(env);
2877 /** Read the environment parameters of a DB environment before
2878 * mapping it into memory.
2879 * @param[in] env the environment handle
2880 * @param[out] meta address of where to store the meta information
2881 * @return 0 on success, non-zero on failure.
2884 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2891 /* We don't know the page size yet, so use a minimum value.
2892 * Read both meta pages so we can use the latest one.
2895 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2899 memset(&ov, 0, sizeof(ov));
2901 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2902 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2905 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2907 if (rc != MDB_PAGESIZE) {
2908 if (rc == 0 && off == 0)
2910 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2911 DPRINTF("read: %s", mdb_strerror(rc));
2915 p = (MDB_page *)&pbuf;
2917 if (!F_ISSET(p->mp_flags, P_META)) {
2918 DPRINTF("page %zu not a meta page", p->mp_pgno);
2923 if (m->mm_magic != MDB_MAGIC) {
2924 DPUTS("meta has invalid magic");
2928 if (m->mm_version != MDB_DATA_VERSION) {
2929 DPRINTF("database is version %u, expected version %u",
2930 m->mm_version, MDB_DATA_VERSION);
2931 return MDB_VERSION_MISMATCH;
2934 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2940 /** Write the environment parameters of a freshly created DB environment.
2941 * @param[in] env the environment handle
2942 * @param[out] meta address of where to store the meta information
2943 * @return 0 on success, non-zero on failure.
2946 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2952 DPUTS("writing new meta page");
2954 GET_PAGESIZE(psize);
2956 meta->mm_magic = MDB_MAGIC;
2957 meta->mm_version = MDB_DATA_VERSION;
2958 meta->mm_mapsize = env->me_mapsize;
2959 meta->mm_psize = psize;
2960 meta->mm_last_pg = 1;
2961 meta->mm_flags = env->me_flags & 0xffff;
2962 meta->mm_flags |= MDB_INTEGERKEY;
2963 meta->mm_dbs[0].md_root = P_INVALID;
2964 meta->mm_dbs[1].md_root = P_INVALID;
2966 p = calloc(2, psize);
2968 p->mp_flags = P_META;
2969 *(MDB_meta *)METADATA(p) = *meta;
2971 q = (MDB_page *)((char *)p + psize);
2973 q->mp_flags = P_META;
2974 *(MDB_meta *)METADATA(q) = *meta;
2980 memset(&ov, 0, sizeof(ov));
2981 rc = WriteFile(env->me_fd, p, psize * 2, &len, &ov);
2982 rc = rc ? (len == psize * 2 ? MDB_SUCCESS : EIO) : ErrCode();
2985 rc = pwrite(env->me_fd, p, psize * 2, 0);
2986 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
2992 /** Update the environment info to commit a transaction.
2993 * @param[in] txn the transaction that's being committed
2994 * @return 0 on success, non-zero on failure.
2997 mdb_env_write_meta(MDB_txn *txn)
3000 MDB_meta meta, metab, *mp;
3002 int rc, len, toggle;
3011 assert(txn != NULL);
3012 assert(txn->mt_env != NULL);
3014 toggle = !txn->mt_toggle;
3015 DPRINTF("writing meta page %d for root page %zu",
3016 toggle, txn->mt_dbs[MAIN_DBI].md_root);
3019 mp = env->me_metas[toggle];
3021 if (env->me_flags & MDB_WRITEMAP) {
3022 /* Persist any increases of mapsize config */
3023 if (env->me_mapsize > mp->mm_mapsize)
3024 mp->mm_mapsize = env->me_mapsize;
3025 mp->mm_dbs[0] = txn->mt_dbs[0];
3026 mp->mm_dbs[1] = txn->mt_dbs[1];
3027 mp->mm_last_pg = txn->mt_next_pgno - 1;
3028 mp->mm_txnid = txn->mt_txnid;
3029 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3030 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3033 ptr += env->me_psize;
3034 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3041 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3042 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3044 ptr = (char *)&meta;
3045 if (env->me_mapsize > mp->mm_mapsize) {
3046 /* Persist any increases of mapsize config */
3047 meta.mm_mapsize = env->me_mapsize;
3048 off = offsetof(MDB_meta, mm_mapsize);
3050 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3052 len = sizeof(MDB_meta) - off;
3055 meta.mm_dbs[0] = txn->mt_dbs[0];
3056 meta.mm_dbs[1] = txn->mt_dbs[1];
3057 meta.mm_last_pg = txn->mt_next_pgno - 1;
3058 meta.mm_txnid = txn->mt_txnid;
3061 off += env->me_psize;
3064 /* Write to the SYNC fd */
3065 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3066 env->me_fd : env->me_mfd;
3069 memset(&ov, 0, sizeof(ov));
3071 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3075 rc = pwrite(mfd, ptr, len, off);
3078 rc = rc < 0 ? ErrCode() : EIO;
3079 DPUTS("write failed, disk error?");
3080 /* On a failure, the pagecache still contains the new data.
3081 * Write some old data back, to prevent it from being used.
3082 * Use the non-SYNC fd; we know it will fail anyway.
3084 meta.mm_last_pg = metab.mm_last_pg;
3085 meta.mm_txnid = metab.mm_txnid;
3087 memset(&ov, 0, sizeof(ov));
3089 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3091 r2 = pwrite(env->me_fd, ptr, len, off);
3094 env->me_flags |= MDB_FATAL_ERROR;
3098 /* Memory ordering issues are irrelevant; since the entire writer
3099 * is wrapped by wmutex, all of these changes will become visible
3100 * after the wmutex is unlocked. Since the DB is multi-version,
3101 * readers will get consistent data regardless of how fresh or
3102 * how stale their view of these values is.
3104 env->me_txns->mti_txnid = txn->mt_txnid;
3109 /** Check both meta pages to see which one is newer.
3110 * @param[in] env the environment handle
3111 * @return meta toggle (0 or 1).
3114 mdb_env_pick_meta(const MDB_env *env)
3116 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3120 mdb_env_create(MDB_env **env)
3124 e = calloc(1, sizeof(MDB_env));
3128 e->me_maxreaders = DEFAULT_READERS;
3129 e->me_maxdbs = e->me_numdbs = 2;
3130 e->me_fd = INVALID_HANDLE_VALUE;
3131 e->me_lfd = INVALID_HANDLE_VALUE;
3132 e->me_mfd = INVALID_HANDLE_VALUE;
3133 #ifdef MDB_USE_POSIX_SEM
3134 e->me_rmutex = SEM_FAILED;
3135 e->me_wmutex = SEM_FAILED;
3137 e->me_pid = getpid();
3138 VGMEMP_CREATE(e,0,0);
3144 mdb_env_set_mapsize(MDB_env *env, size_t size)
3148 env->me_mapsize = size;
3150 env->me_maxpg = env->me_mapsize / env->me_psize;
3155 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3159 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3164 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3166 if (env->me_map || readers < 1)
3168 env->me_maxreaders = readers;
3173 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3175 if (!env || !readers)
3177 *readers = env->me_maxreaders;
3181 /** Further setup required for opening an MDB environment
3184 mdb_env_open2(MDB_env *env)
3186 unsigned int flags = env->me_flags;
3194 memset(&meta, 0, sizeof(meta));
3196 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3199 DPUTS("new mdbenv");
3203 /* Was a mapsize configured? */
3204 if (!env->me_mapsize) {
3205 /* If this is a new environment, take the default,
3206 * else use the size recorded in the existing env.
3208 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3209 } else if (env->me_mapsize < meta.mm_mapsize) {
3210 /* If the configured size is smaller, make sure it's
3211 * still big enough. Silently round up to minimum if not.
3213 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3214 if (env->me_mapsize < minsize)
3215 env->me_mapsize = minsize;
3222 LONG sizelo, sizehi;
3223 sizelo = env->me_mapsize & 0xffffffff;
3224 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3226 /* See if we should use QueryLimited */
3228 if ((rc & 0xff) > 5)
3229 env->me_pidquery = PROCESS_QUERY_LIMITED_INFORMATION;
3231 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3233 /* Windows won't create mappings for zero length files.
3234 * Just allocate the maxsize right now.
3237 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3238 || !SetEndOfFile(env->me_fd)
3239 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3242 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3243 PAGE_READWRITE : PAGE_READONLY,
3244 sizehi, sizelo, NULL);
3247 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3248 FILE_MAP_WRITE : FILE_MAP_READ,
3249 0, 0, env->me_mapsize, meta.mm_address);
3250 rc = env->me_map ? 0 : ErrCode();
3258 if (flags & MDB_WRITEMAP) {
3260 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3263 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3265 if (env->me_map == MAP_FAILED) {
3269 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3271 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3273 #ifdef POSIX_MADV_RANDOM
3274 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3275 #endif /* POSIX_MADV_RANDOM */
3276 #endif /* MADV_RANDOM */
3280 if (flags & MDB_FIXEDMAP)
3281 meta.mm_address = env->me_map;
3282 i = mdb_env_init_meta(env, &meta);
3283 if (i != MDB_SUCCESS) {
3286 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3287 /* Can happen because the address argument to mmap() is just a
3288 * hint. mmap() can pick another, e.g. if the range is in use.
3289 * The MAP_FIXED flag would prevent that, but then mmap could
3290 * instead unmap existing pages to make room for the new map.
3292 return EBUSY; /* TODO: Make a new MDB_* error code? */
3294 env->me_psize = meta.mm_psize;
3295 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3296 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3298 env->me_maxpg = env->me_mapsize / env->me_psize;
3300 p = (MDB_page *)env->me_map;
3301 env->me_metas[0] = METADATA(p);
3302 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3306 int toggle = mdb_env_pick_meta(env);
3307 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3309 DPRINTF("opened database version %u, pagesize %u",
3310 env->me_metas[0]->mm_version, env->me_psize);
3311 DPRINTF("using meta page %d", toggle);
3312 DPRINTF("depth: %u", db->md_depth);
3313 DPRINTF("entries: %zu", db->md_entries);
3314 DPRINTF("branch pages: %zu", db->md_branch_pages);
3315 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
3316 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
3317 DPRINTF("root: %zu", db->md_root);
3325 /** Release a reader thread's slot in the reader lock table.
3326 * This function is called automatically when a thread exits.
3327 * @param[in] ptr This points to the slot in the reader lock table.
3330 mdb_env_reader_dest(void *ptr)
3332 MDB_reader *reader = ptr;
3338 /** Junk for arranging thread-specific callbacks on Windows. This is
3339 * necessarily platform and compiler-specific. Windows supports up
3340 * to 1088 keys. Let's assume nobody opens more than 64 environments
3341 * in a single process, for now. They can override this if needed.
3343 #ifndef MAX_TLS_KEYS
3344 #define MAX_TLS_KEYS 64
3346 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3347 static int mdb_tls_nkeys;
3349 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3353 case DLL_PROCESS_ATTACH: break;
3354 case DLL_THREAD_ATTACH: break;
3355 case DLL_THREAD_DETACH:
3356 for (i=0; i<mdb_tls_nkeys; i++) {
3357 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3358 mdb_env_reader_dest(r);
3361 case DLL_PROCESS_DETACH: break;
3366 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3368 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3372 /* Force some symbol references.
3373 * _tls_used forces the linker to create the TLS directory if not already done
3374 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3376 #pragma comment(linker, "/INCLUDE:_tls_used")
3377 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3378 #pragma const_seg(".CRT$XLB")
3379 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3380 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3383 #pragma comment(linker, "/INCLUDE:__tls_used")
3384 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3385 #pragma data_seg(".CRT$XLB")
3386 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3388 #endif /* WIN 32/64 */
3389 #endif /* !__GNUC__ */
3392 /** Downgrade the exclusive lock on the region back to shared */
3394 mdb_env_share_locks(MDB_env *env, int *excl)
3396 int rc = 0, toggle = mdb_env_pick_meta(env);
3398 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3403 /* First acquire a shared lock. The Unlock will
3404 * then release the existing exclusive lock.
3406 memset(&ov, 0, sizeof(ov));
3407 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3410 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3416 struct flock lock_info;
3417 /* The shared lock replaces the existing lock */
3418 memset((void *)&lock_info, 0, sizeof(lock_info));
3419 lock_info.l_type = F_RDLCK;
3420 lock_info.l_whence = SEEK_SET;
3421 lock_info.l_start = 0;
3422 lock_info.l_len = 1;
3423 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3424 (rc = ErrCode()) == EINTR) ;
3425 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3432 /** Try to get exlusive lock, otherwise shared.
3433 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3436 mdb_env_excl_lock(MDB_env *env, int *excl)
3440 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3444 memset(&ov, 0, sizeof(ov));
3445 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3452 struct flock lock_info;
3453 memset((void *)&lock_info, 0, sizeof(lock_info));
3454 lock_info.l_type = F_WRLCK;
3455 lock_info.l_whence = SEEK_SET;
3456 lock_info.l_start = 0;
3457 lock_info.l_len = 1;
3458 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3459 (rc = ErrCode()) == EINTR) ;
3463 # ifdef MDB_USE_POSIX_SEM
3464 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3467 lock_info.l_type = F_RDLCK;
3468 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3469 (rc = ErrCode()) == EINTR) ;
3477 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3479 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3481 * @(#) $Revision: 5.1 $
3482 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3483 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3485 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3489 * Please do not copyright this code. This code is in the public domain.
3491 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3492 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3493 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3494 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3495 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3496 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3497 * PERFORMANCE OF THIS SOFTWARE.
3500 * chongo <Landon Curt Noll> /\oo/\
3501 * http://www.isthe.com/chongo/
3503 * Share and Enjoy! :-)
3506 typedef unsigned long long mdb_hash_t;
3507 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3509 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3510 * @param[in] str string to hash
3511 * @param[in] hval initial value for hash
3512 * @return 64 bit hash
3514 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3515 * hval arg on the first call.
3518 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3520 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3521 unsigned char *end = s + val->mv_size;
3523 * FNV-1a hash each octet of the string
3526 /* xor the bottom with the current octet */
3527 hval ^= (mdb_hash_t)*s++;
3529 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3530 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3531 (hval << 7) + (hval << 8) + (hval << 40);
3533 /* return our new hash value */
3537 /** Hash the string and output the hash in hex.
3538 * @param[in] str string to hash
3539 * @param[out] hexbuf an array of 17 chars to hold the hash
3542 mdb_hash_hex(MDB_val *val, char *hexbuf)
3545 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3546 for (i=0; i<8; i++) {
3547 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3553 /** Open and/or initialize the lock region for the environment.
3554 * @param[in] env The MDB environment.
3555 * @param[in] lpath The pathname of the file used for the lock region.
3556 * @param[in] mode The Unix permissions for the file, if we create it.
3557 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3558 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3559 * @return 0 on success, non-zero on failure.
3562 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3565 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3567 # define MDB_ERRCODE_ROFS EROFS
3568 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3569 # define MDB_CLOEXEC O_CLOEXEC
3572 # define MDB_CLOEXEC 0
3579 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3580 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3581 FILE_ATTRIBUTE_NORMAL, NULL);
3583 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3585 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3587 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3592 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3593 /* Lose record locks when exec*() */
3594 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3595 fcntl(env->me_lfd, F_SETFD, fdflags);
3598 if (!(env->me_flags & MDB_NOTLS)) {
3599 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3602 env->me_flags |= MDB_ENV_TXKEY;
3604 /* Windows TLS callbacks need help finding their TLS info. */
3605 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3609 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3613 /* Try to get exclusive lock. If we succeed, then
3614 * nobody is using the lock region and we should initialize it.
3616 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3619 size = GetFileSize(env->me_lfd, NULL);
3621 size = lseek(env->me_lfd, 0, SEEK_END);
3622 if (size == -1) goto fail_errno;
3624 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3625 if (size < rsize && *excl > 0) {
3627 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3628 || !SetEndOfFile(env->me_lfd))
3631 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3635 size = rsize - sizeof(MDB_txninfo);
3636 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3641 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3643 if (!mh) goto fail_errno;
3644 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3646 if (!env->me_txns) goto fail_errno;
3648 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3650 if (m == MAP_FAILED) goto fail_errno;
3656 BY_HANDLE_FILE_INFORMATION stbuf;
3665 if (!mdb_sec_inited) {
3666 InitializeSecurityDescriptor(&mdb_null_sd,
3667 SECURITY_DESCRIPTOR_REVISION);
3668 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3669 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3670 mdb_all_sa.bInheritHandle = FALSE;
3671 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3674 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3675 idbuf.volume = stbuf.dwVolumeSerialNumber;
3676 idbuf.nhigh = stbuf.nFileIndexHigh;
3677 idbuf.nlow = stbuf.nFileIndexLow;
3678 val.mv_data = &idbuf;
3679 val.mv_size = sizeof(idbuf);
3680 mdb_hash_hex(&val, hexbuf);
3681 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3682 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3683 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3684 if (!env->me_rmutex) goto fail_errno;
3685 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3686 if (!env->me_wmutex) goto fail_errno;
3687 #elif defined(MDB_USE_POSIX_SEM)
3696 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3697 idbuf.dev = stbuf.st_dev;
3698 idbuf.ino = stbuf.st_ino;
3699 val.mv_data = &idbuf;
3700 val.mv_size = sizeof(idbuf);
3701 mdb_hash_hex(&val, hexbuf);
3702 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3703 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3704 /* Clean up after a previous run, if needed: Try to
3705 * remove both semaphores before doing anything else.
3707 sem_unlink(env->me_txns->mti_rmname);
3708 sem_unlink(env->me_txns->mti_wmname);
3709 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3710 O_CREAT|O_EXCL, mode, 1);
3711 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3712 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3713 O_CREAT|O_EXCL, mode, 1);
3714 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3715 #else /* MDB_USE_POSIX_SEM */
3716 pthread_mutexattr_t mattr;
3718 if ((rc = pthread_mutexattr_init(&mattr))
3719 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3720 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3721 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3723 pthread_mutexattr_destroy(&mattr);
3724 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3726 env->me_txns->mti_version = MDB_LOCK_VERSION;
3727 env->me_txns->mti_magic = MDB_MAGIC;
3728 env->me_txns->mti_txnid = 0;
3729 env->me_txns->mti_numreaders = 0;
3732 if (env->me_txns->mti_magic != MDB_MAGIC) {
3733 DPUTS("lock region has invalid magic");
3737 if (env->me_txns->mti_version != MDB_LOCK_VERSION) {
3738 DPRINTF("lock region is version %u, expected version %u",
3739 env->me_txns->mti_version, MDB_LOCK_VERSION);
3740 rc = MDB_VERSION_MISMATCH;
3744 if (rc && rc != EACCES && rc != EAGAIN) {
3748 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3749 if (!env->me_rmutex) goto fail_errno;
3750 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3751 if (!env->me_wmutex) goto fail_errno;
3752 #elif defined(MDB_USE_POSIX_SEM)
3753 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3754 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3755 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3756 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3767 /** The name of the lock file in the DB environment */
3768 #define LOCKNAME "/lock.mdb"
3769 /** The name of the data file in the DB environment */
3770 #define DATANAME "/data.mdb"
3771 /** The suffix of the lock file when no subdir is used */
3772 #define LOCKSUFF "-lock"
3773 /** Only a subset of the @ref mdb_env flags can be changed
3774 * at runtime. Changing other flags requires closing the
3775 * environment and re-opening it with the new flags.
3777 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3778 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3781 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3783 int oflags, rc, len, excl = -1;
3784 char *lpath, *dpath;
3786 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3790 if (flags & MDB_NOSUBDIR) {
3791 rc = len + sizeof(LOCKSUFF) + len + 1;
3793 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3798 if (flags & MDB_NOSUBDIR) {
3799 dpath = lpath + len + sizeof(LOCKSUFF);
3800 sprintf(lpath, "%s" LOCKSUFF, path);
3801 strcpy(dpath, path);
3803 dpath = lpath + len + sizeof(LOCKNAME);
3804 sprintf(lpath, "%s" LOCKNAME, path);
3805 sprintf(dpath, "%s" DATANAME, path);
3809 flags |= env->me_flags;
3810 if (flags & MDB_RDONLY) {
3811 /* silently ignore WRITEMAP when we're only getting read access */
3812 flags &= ~MDB_WRITEMAP;
3814 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3815 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3818 env->me_flags = flags |= MDB_ENV_ACTIVE;
3822 env->me_path = strdup(path);
3823 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3824 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3825 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3830 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3835 if (F_ISSET(flags, MDB_RDONLY)) {
3836 oflags = GENERIC_READ;
3837 len = OPEN_EXISTING;
3839 oflags = GENERIC_READ|GENERIC_WRITE;
3842 mode = FILE_ATTRIBUTE_NORMAL;
3843 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3844 NULL, len, mode, NULL);
3846 if (F_ISSET(flags, MDB_RDONLY))
3849 oflags = O_RDWR | O_CREAT;
3851 env->me_fd = open(dpath, oflags, mode);
3853 if (env->me_fd == INVALID_HANDLE_VALUE) {
3858 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3859 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3860 env->me_mfd = env->me_fd;
3862 /* Synchronous fd for meta writes. Needed even with
3863 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3866 env->me_mfd = CreateFile(dpath, oflags,
3867 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3868 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3870 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3872 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3877 DPRINTF("opened dbenv %p", (void *) env);
3879 rc = mdb_env_share_locks(env, &excl);
3885 mdb_env_close0(env, excl);
3891 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3893 mdb_env_close0(MDB_env *env, int excl)
3897 if (!(env->me_flags & MDB_ENV_ACTIVE))
3900 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3901 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3902 free(env->me_dbxs[i].md_name.mv_data);
3904 free(env->me_dbflags);
3907 free(env->me_dirty_list);
3908 mdb_midl_free(env->me_free_pgs);
3910 if (env->me_flags & MDB_ENV_TXKEY) {
3911 pthread_key_delete(env->me_txkey);
3913 /* Delete our key from the global list */
3914 for (i=0; i<mdb_tls_nkeys; i++)
3915 if (mdb_tls_keys[i] == env->me_txkey) {
3916 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3924 munmap(env->me_map, env->me_mapsize);
3926 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3927 (void) close(env->me_mfd);
3928 if (env->me_fd != INVALID_HANDLE_VALUE)
3929 (void) close(env->me_fd);
3931 pid_t pid = env->me_pid;
3932 /* Clearing readers is done in this function because
3933 * me_txkey with its destructor must be disabled first.
3935 for (i = env->me_numreaders; --i >= 0; )
3936 if (env->me_txns->mti_readers[i].mr_pid == pid)
3937 env->me_txns->mti_readers[i].mr_pid = 0;
3939 if (env->me_rmutex) {
3940 CloseHandle(env->me_rmutex);
3941 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3943 /* Windows automatically destroys the mutexes when
3944 * the last handle closes.
3946 #elif defined(MDB_USE_POSIX_SEM)
3947 if (env->me_rmutex != SEM_FAILED) {
3948 sem_close(env->me_rmutex);
3949 if (env->me_wmutex != SEM_FAILED)
3950 sem_close(env->me_wmutex);
3951 /* If we have the filelock: If we are the
3952 * only remaining user, clean up semaphores.
3955 mdb_env_excl_lock(env, &excl);
3957 sem_unlink(env->me_txns->mti_rmname);
3958 sem_unlink(env->me_txns->mti_wmname);
3962 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3964 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3967 /* Unlock the lockfile. Windows would have unlocked it
3968 * after closing anyway, but not necessarily at once.
3970 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3973 (void) close(env->me_lfd);
3976 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3980 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3982 MDB_txn *txn = NULL;
3987 /* Do the lock/unlock of the reader mutex before starting the
3988 * write txn. Otherwise other read txns could block writers.
3990 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3995 /* We must start the actual read txn after blocking writers */
3996 mdb_txn_reset0(txn, "reset-stage1");
3998 /* Temporarily block writers until we snapshot the meta pages */
4001 rc = mdb_txn_renew0(txn);
4003 UNLOCK_MUTEX_W(env);
4008 wsize = env->me_psize * 2;
4012 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
4013 rc = rc ? (len == wsize ? MDB_SUCCESS : EIO) : ErrCode();
4016 rc = write(fd, env->me_map, wsize);
4017 rc = rc == (int)wsize ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
4020 UNLOCK_MUTEX_W(env);
4025 ptr = env->me_map + wsize;
4026 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4030 if (wsize > MAX_WRITE)
4034 rc = WriteFile(fd, ptr, w2, &len, NULL);
4035 rc = rc ? (len == w2 ? MDB_SUCCESS : EIO) : ErrCode();
4044 if (wsize > MAX_WRITE)
4048 wres = write(fd, ptr, w2);
4049 rc = wres == (ssize_t)w2 ? MDB_SUCCESS : wres < 0 ? ErrCode() : EIO;
4062 mdb_env_copy(MDB_env *env, const char *path)
4066 HANDLE newfd = INVALID_HANDLE_VALUE;
4068 if (env->me_flags & MDB_NOSUBDIR) {
4069 lpath = (char *)path;
4072 len += sizeof(DATANAME);
4073 lpath = malloc(len);
4076 sprintf(lpath, "%s" DATANAME, path);
4079 /* The destination path must exist, but the destination file must not.
4080 * We don't want the OS to cache the writes, since the source data is
4081 * already in the OS cache.
4084 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4085 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4087 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4093 if (newfd == INVALID_HANDLE_VALUE) {
4098 #ifdef F_NOCACHE /* __APPLE__ */
4099 rc = fcntl(newfd, F_NOCACHE, 1);
4106 rc = mdb_env_copyfd(env, newfd);
4109 if (!(env->me_flags & MDB_NOSUBDIR))
4111 if (newfd != INVALID_HANDLE_VALUE)
4112 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4119 mdb_env_close(MDB_env *env)
4126 VGMEMP_DESTROY(env);
4127 while ((dp = env->me_dpages) != NULL) {
4128 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4129 env->me_dpages = dp->mp_next;
4133 mdb_env_close0(env, 0);
4137 /** Compare two items pointing at aligned size_t's */
4139 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4141 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4142 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4145 /** Compare two items pointing at aligned int's */
4147 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4149 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4150 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4153 /** Compare two items pointing at ints of unknown alignment.
4154 * Nodes and keys are guaranteed to be 2-byte aligned.
4157 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4159 #if BYTE_ORDER == LITTLE_ENDIAN
4160 unsigned short *u, *c;
4163 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4164 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4167 } while(!x && u > (unsigned short *)a->mv_data);
4170 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4174 /** Compare two items lexically */
4176 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4183 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4189 diff = memcmp(a->mv_data, b->mv_data, len);
4190 return diff ? diff : len_diff<0 ? -1 : len_diff;
4193 /** Compare two items in reverse byte order */
4195 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4197 const unsigned char *p1, *p2, *p1_lim;
4201 p1_lim = (const unsigned char *)a->mv_data;
4202 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4203 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4205 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4211 while (p1 > p1_lim) {
4212 diff = *--p1 - *--p2;
4216 return len_diff<0 ? -1 : len_diff;
4219 /** Search for key within a page, using binary search.
4220 * Returns the smallest entry larger or equal to the key.
4221 * If exactp is non-null, stores whether the found entry was an exact match
4222 * in *exactp (1 or 0).
4223 * Updates the cursor index with the index of the found entry.
4224 * If no entry larger or equal to the key is found, returns NULL.
4227 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4229 unsigned int i = 0, nkeys;
4232 MDB_page *mp = mc->mc_pg[mc->mc_top];
4233 MDB_node *node = NULL;
4238 nkeys = NUMKEYS(mp);
4243 COPY_PGNO(pgno, mp->mp_pgno);
4244 DPRINTF("searching %u keys in %s %spage %zu",
4245 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4252 low = IS_LEAF(mp) ? 0 : 1;
4254 cmp = mc->mc_dbx->md_cmp;
4256 /* Branch pages have no data, so if using integer keys,
4257 * alignment is guaranteed. Use faster mdb_cmp_int.
4259 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4260 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4267 nodekey.mv_size = mc->mc_db->md_pad;
4268 node = NODEPTR(mp, 0); /* fake */
4269 while (low <= high) {
4270 i = (low + high) >> 1;
4271 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4272 rc = cmp(key, &nodekey);
4273 DPRINTF("found leaf index %u [%s], rc = %i",
4274 i, DKEY(&nodekey), rc);
4283 while (low <= high) {
4284 i = (low + high) >> 1;
4286 node = NODEPTR(mp, i);
4287 nodekey.mv_size = NODEKSZ(node);
4288 nodekey.mv_data = NODEKEY(node);
4290 rc = cmp(key, &nodekey);
4293 DPRINTF("found leaf index %u [%s], rc = %i",
4294 i, DKEY(&nodekey), rc);
4296 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
4297 i, DKEY(&nodekey), NODEPGNO(node), rc);
4308 if (rc > 0) { /* Found entry is less than the key. */
4309 i++; /* Skip to get the smallest entry larger than key. */
4311 node = NODEPTR(mp, i);
4314 *exactp = (rc == 0);
4315 /* store the key index */
4316 mc->mc_ki[mc->mc_top] = i;
4318 /* There is no entry larger or equal to the key. */
4321 /* nodeptr is fake for LEAF2 */
4327 mdb_cursor_adjust(MDB_cursor *mc, func)
4331 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4332 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4339 /** Pop a page off the top of the cursor's stack. */
4341 mdb_cursor_pop(MDB_cursor *mc)
4344 #ifndef MDB_DEBUG_SKIP
4345 MDB_page *top = mc->mc_pg[mc->mc_top];
4351 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4352 mc->mc_dbi, (void *) mc);
4356 /** Push a page onto the top of the cursor's stack. */
4358 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4360 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4361 mc->mc_dbi, (void *) mc);
4363 if (mc->mc_snum >= CURSOR_STACK) {
4364 assert(mc->mc_snum < CURSOR_STACK);
4365 return MDB_CURSOR_FULL;
4368 mc->mc_top = mc->mc_snum++;
4369 mc->mc_pg[mc->mc_top] = mp;
4370 mc->mc_ki[mc->mc_top] = 0;
4375 /** Find the address of the page corresponding to a given page number.
4376 * @param[in] txn the transaction for this access.
4377 * @param[in] pgno the page number for the page to retrieve.
4378 * @param[out] ret address of a pointer where the page's address will be stored.
4379 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4380 * @return 0 on success, non-zero on failure.
4383 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4388 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4389 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4394 MDB_ID2L dl = tx2->mt_u.dirty_list;
4396 /* Spilled pages were dirtied in this txn and flushed
4397 * because the dirty list got full. Bring this page
4398 * back in from the map (but don't unspill it here,
4399 * leave that unless page_touch happens again).
4401 if (tx2->mt_spill_pgs) {
4402 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4403 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4404 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4409 unsigned x = mdb_mid2l_search(dl, pgno);
4410 if (x <= dl[0].mid && dl[x].mid == pgno) {
4416 } while ((tx2 = tx2->mt_parent) != NULL);
4419 if (pgno < txn->mt_next_pgno) {
4421 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4423 DPRINTF("page %zu not found", pgno);
4425 return MDB_PAGE_NOTFOUND;
4435 /** Search for the page a given key should be in.
4436 * Pushes parent pages on the cursor stack. This function continues a
4437 * search on a cursor that has already been initialized. (Usually by
4438 * #mdb_page_search() but also by #mdb_node_move().)
4439 * @param[in,out] mc the cursor for this operation.
4440 * @param[in] key the key to search for. If NULL, search for the lowest
4441 * page. (This is used by #mdb_cursor_first().)
4442 * @param[in] modify If true, visited pages are updated with new page numbers.
4443 * @return 0 on success, non-zero on failure.
4446 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4448 MDB_page *mp = mc->mc_pg[mc->mc_top];
4453 while (IS_BRANCH(mp)) {
4457 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4458 assert(NUMKEYS(mp) > 1);
4459 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4461 if (key == NULL) /* Initialize cursor to first page. */
4463 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4464 /* cursor to last page */
4468 node = mdb_node_search(mc, key, &exact);
4470 i = NUMKEYS(mp) - 1;
4472 i = mc->mc_ki[mc->mc_top];
4481 DPRINTF("following index %u for key [%s]",
4483 assert(i < NUMKEYS(mp));
4484 node = NODEPTR(mp, i);
4486 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4489 mc->mc_ki[mc->mc_top] = i;
4490 if ((rc = mdb_cursor_push(mc, mp)))
4494 if ((rc = mdb_page_touch(mc)) != 0)
4496 mp = mc->mc_pg[mc->mc_top];
4501 DPRINTF("internal error, index points to a %02X page!?",
4503 return MDB_CORRUPTED;
4506 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4507 key ? DKEY(key) : NULL);
4508 mc->mc_flags |= C_INITIALIZED;
4509 mc->mc_flags &= ~C_EOF;
4514 /** Search for the lowest key under the current branch page.
4515 * This just bypasses a NUMKEYS check in the current page
4516 * before calling mdb_page_search_root(), because the callers
4517 * are all in situations where the current page is known to
4521 mdb_page_search_lowest(MDB_cursor *mc)
4523 MDB_page *mp = mc->mc_pg[mc->mc_top];
4524 MDB_node *node = NODEPTR(mp, 0);
4527 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4530 mc->mc_ki[mc->mc_top] = 0;
4531 if ((rc = mdb_cursor_push(mc, mp)))
4533 return mdb_page_search_root(mc, NULL, 0);
4536 /** Search for the page a given key should be in.
4537 * Pushes parent pages on the cursor stack. This function just sets up
4538 * the search; it finds the root page for \b mc's database and sets this
4539 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4540 * called to complete the search.
4541 * @param[in,out] mc the cursor for this operation.
4542 * @param[in] key the key to search for. If NULL, search for the lowest
4543 * page. (This is used by #mdb_cursor_first().)
4544 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4545 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4546 * @return 0 on success, non-zero on failure.
4549 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4554 /* Make sure the txn is still viable, then find the root from
4555 * the txn's db table.
4557 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4558 DPUTS("transaction has failed, must abort");
4561 /* Make sure we're using an up-to-date root */
4562 if (mc->mc_dbi > MAIN_DBI) {
4563 if ((*mc->mc_dbflag & DB_STALE) ||
4564 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4566 unsigned char dbflag = 0;
4567 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4568 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4571 if (*mc->mc_dbflag & DB_STALE) {
4575 MDB_node *leaf = mdb_node_search(&mc2,
4576 &mc->mc_dbx->md_name, &exact);
4578 return MDB_NOTFOUND;
4579 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4582 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4584 /* The txn may not know this DBI, or another process may
4585 * have dropped and recreated the DB with other flags.
4587 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4588 return MDB_INCOMPATIBLE;
4589 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4591 if (flags & MDB_PS_MODIFY)
4593 *mc->mc_dbflag &= ~DB_STALE;
4594 *mc->mc_dbflag |= dbflag;
4597 root = mc->mc_db->md_root;
4599 if (root == P_INVALID) { /* Tree is empty. */
4600 DPUTS("tree is empty");
4601 return MDB_NOTFOUND;
4606 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4607 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4613 DPRINTF("db %u root page %zu has flags 0x%X",
4614 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4616 if (flags & MDB_PS_MODIFY) {
4617 if ((rc = mdb_page_touch(mc)))
4621 if (flags & MDB_PS_ROOTONLY)
4624 return mdb_page_search_root(mc, key, flags);
4628 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4630 MDB_txn *txn = mc->mc_txn;
4631 pgno_t pg = mp->mp_pgno;
4632 unsigned i, ovpages = mp->mp_pages;
4633 MDB_env *env = txn->mt_env;
4636 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4637 /* If the page is dirty or on the spill list we just acquired it,
4638 * so we should give it back to our current free list, if any.
4639 * Not currently supported in nested txns.
4640 * Otherwise put it onto the list of pages we freed in this txn.
4642 if (!(mp->mp_flags & P_DIRTY) && txn->mt_spill_pgs) {
4643 unsigned x = mdb_midl_search(txn->mt_spill_pgs, pg);
4644 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pg) {
4645 /* This page is no longer spilled */
4646 for (; x < txn->mt_spill_pgs[0]; x++)
4647 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
4648 txn->mt_spill_pgs[0]--;
4652 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && env->me_pghead) {
4655 MDB_ID2 *dl, ix, iy;
4656 rc = mdb_midl_need(&env->me_pghead, ovpages);
4659 /* Remove from dirty list */
4660 dl = txn->mt_u.dirty_list;
4662 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4670 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4671 txn->mt_flags |= MDB_TXN_ERROR;
4672 return MDB_CORRUPTED;
4675 if (!(env->me_flags & MDB_WRITEMAP))
4676 mdb_dpage_free(env, mp);
4678 /* Insert in me_pghead */
4679 mop = env->me_pghead;
4680 j = mop[0] + ovpages;
4681 for (i = mop[0]; i && mop[i] < pg; i--)
4687 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4691 mc->mc_db->md_overflow_pages -= ovpages;
4695 /** Return the data associated with a given node.
4696 * @param[in] txn The transaction for this operation.
4697 * @param[in] leaf The node being read.
4698 * @param[out] data Updated to point to the node's data.
4699 * @return 0 on success, non-zero on failure.
4702 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4704 MDB_page *omp; /* overflow page */
4708 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4709 data->mv_size = NODEDSZ(leaf);
4710 data->mv_data = NODEDATA(leaf);
4714 /* Read overflow data.
4716 data->mv_size = NODEDSZ(leaf);
4717 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4718 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4719 DPRINTF("read overflow page %zu failed", pgno);
4722 data->mv_data = METADATA(omp);
4728 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4729 MDB_val *key, MDB_val *data)
4738 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4740 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4743 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4747 mdb_cursor_init(&mc, txn, dbi, &mx);
4748 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4751 /** Find a sibling for a page.
4752 * Replaces the page at the top of the cursor's stack with the
4753 * specified sibling, if one exists.
4754 * @param[in] mc The cursor for this operation.
4755 * @param[in] move_right Non-zero if the right sibling is requested,
4756 * otherwise the left sibling.
4757 * @return 0 on success, non-zero on failure.
4760 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4766 if (mc->mc_snum < 2) {
4767 return MDB_NOTFOUND; /* root has no siblings */
4771 DPRINTF("parent page is page %zu, index %u",
4772 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4774 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4775 : (mc->mc_ki[mc->mc_top] == 0)) {
4776 DPRINTF("no more keys left, moving to %s sibling",
4777 move_right ? "right" : "left");
4778 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4779 /* undo cursor_pop before returning */
4786 mc->mc_ki[mc->mc_top]++;
4788 mc->mc_ki[mc->mc_top]--;
4789 DPRINTF("just moving to %s index key %u",
4790 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4792 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4794 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4795 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4798 mdb_cursor_push(mc, mp);
4800 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4805 /** Move the cursor to the next data item. */
4807 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4813 if (mc->mc_flags & C_EOF) {
4814 return MDB_NOTFOUND;
4817 assert(mc->mc_flags & C_INITIALIZED);
4819 mp = mc->mc_pg[mc->mc_top];
4821 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4822 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4823 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4824 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4825 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4826 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4830 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4831 if (op == MDB_NEXT_DUP)
4832 return MDB_NOTFOUND;
4836 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4838 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4839 DPUTS("=====> move to next sibling page");
4840 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4841 mc->mc_flags |= C_EOF;
4844 mp = mc->mc_pg[mc->mc_top];
4845 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4847 mc->mc_ki[mc->mc_top]++;
4849 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4850 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4853 key->mv_size = mc->mc_db->md_pad;
4854 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4858 assert(IS_LEAF(mp));
4859 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4861 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4862 mdb_xcursor_init1(mc, leaf);
4865 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4868 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4869 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4870 if (rc != MDB_SUCCESS)
4875 MDB_GET_KEY(leaf, key);
4879 /** Move the cursor to the previous data item. */
4881 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4887 assert(mc->mc_flags & C_INITIALIZED);
4889 mp = mc->mc_pg[mc->mc_top];
4891 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4892 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4893 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4894 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4895 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4896 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4899 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4900 if (op == MDB_PREV_DUP)
4901 return MDB_NOTFOUND;
4906 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4908 if (mc->mc_ki[mc->mc_top] == 0) {
4909 DPUTS("=====> move to prev sibling page");
4910 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4913 mp = mc->mc_pg[mc->mc_top];
4914 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4915 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4917 mc->mc_ki[mc->mc_top]--;
4919 mc->mc_flags &= ~C_EOF;
4921 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4922 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4925 key->mv_size = mc->mc_db->md_pad;
4926 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4930 assert(IS_LEAF(mp));
4931 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4933 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4934 mdb_xcursor_init1(mc, leaf);
4937 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4940 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4941 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4942 if (rc != MDB_SUCCESS)
4947 MDB_GET_KEY(leaf, key);
4951 /** Set the cursor on a specific data item. */
4953 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4954 MDB_cursor_op op, int *exactp)
4958 MDB_node *leaf = NULL;
4963 assert(key->mv_size > 0);
4966 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4968 /* See if we're already on the right page */
4969 if (mc->mc_flags & C_INITIALIZED) {
4972 mp = mc->mc_pg[mc->mc_top];
4974 mc->mc_ki[mc->mc_top] = 0;
4975 return MDB_NOTFOUND;
4977 if (mp->mp_flags & P_LEAF2) {
4978 nodekey.mv_size = mc->mc_db->md_pad;
4979 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4981 leaf = NODEPTR(mp, 0);
4982 MDB_GET_KEY(leaf, &nodekey);
4984 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4986 /* Probably happens rarely, but first node on the page
4987 * was the one we wanted.
4989 mc->mc_ki[mc->mc_top] = 0;
4996 unsigned int nkeys = NUMKEYS(mp);
4998 if (mp->mp_flags & P_LEAF2) {
4999 nodekey.mv_data = LEAF2KEY(mp,
5000 nkeys-1, nodekey.mv_size);
5002 leaf = NODEPTR(mp, nkeys-1);
5003 MDB_GET_KEY(leaf, &nodekey);
5005 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5007 /* last node was the one we wanted */
5008 mc->mc_ki[mc->mc_top] = nkeys-1;
5014 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5015 /* This is definitely the right page, skip search_page */
5016 if (mp->mp_flags & P_LEAF2) {
5017 nodekey.mv_data = LEAF2KEY(mp,
5018 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5020 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5021 MDB_GET_KEY(leaf, &nodekey);
5023 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5025 /* current node was the one we wanted */
5035 /* If any parents have right-sibs, search.
5036 * Otherwise, there's nothing further.
5038 for (i=0; i<mc->mc_top; i++)
5040 NUMKEYS(mc->mc_pg[i])-1)
5042 if (i == mc->mc_top) {
5043 /* There are no other pages */
5044 mc->mc_ki[mc->mc_top] = nkeys;
5045 return MDB_NOTFOUND;
5049 /* There are no other pages */
5050 mc->mc_ki[mc->mc_top] = 0;
5051 return MDB_NOTFOUND;
5055 rc = mdb_page_search(mc, key, 0);
5056 if (rc != MDB_SUCCESS)
5059 mp = mc->mc_pg[mc->mc_top];
5060 assert(IS_LEAF(mp));
5063 leaf = mdb_node_search(mc, key, exactp);
5064 if (exactp != NULL && !*exactp) {
5065 /* MDB_SET specified and not an exact match. */
5066 return MDB_NOTFOUND;
5070 DPUTS("===> inexact leaf not found, goto sibling");
5071 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5072 return rc; /* no entries matched */
5073 mp = mc->mc_pg[mc->mc_top];
5074 assert(IS_LEAF(mp));
5075 leaf = NODEPTR(mp, 0);
5079 mc->mc_flags |= C_INITIALIZED;
5080 mc->mc_flags &= ~C_EOF;
5083 key->mv_size = mc->mc_db->md_pad;
5084 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5088 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5089 mdb_xcursor_init1(mc, leaf);
5092 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5093 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5094 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5097 if (op == MDB_GET_BOTH) {
5103 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5104 if (rc != MDB_SUCCESS)
5107 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5109 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5111 rc = mc->mc_dbx->md_dcmp(data, &d2);
5113 if (op == MDB_GET_BOTH || rc > 0)
5114 return MDB_NOTFOUND;
5119 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5120 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5125 /* The key already matches in all other cases */
5126 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5127 MDB_GET_KEY(leaf, key);
5128 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
5133 /** Move the cursor to the first item in the database. */
5135 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5141 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5143 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5144 rc = mdb_page_search(mc, NULL, 0);
5145 if (rc != MDB_SUCCESS)
5148 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5150 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5151 mc->mc_flags |= C_INITIALIZED;
5152 mc->mc_flags &= ~C_EOF;
5154 mc->mc_ki[mc->mc_top] = 0;
5156 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5157 key->mv_size = mc->mc_db->md_pad;
5158 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5163 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5164 mdb_xcursor_init1(mc, leaf);
5165 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5169 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5173 MDB_GET_KEY(leaf, key);
5177 /** Move the cursor to the last item in the database. */
5179 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5185 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5187 if (!(mc->mc_flags & C_EOF)) {
5189 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5192 lkey.mv_size = MDB_MAXKEYSIZE+1;
5193 lkey.mv_data = NULL;
5194 rc = mdb_page_search(mc, &lkey, 0);
5195 if (rc != MDB_SUCCESS)
5198 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5201 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5202 mc->mc_flags |= C_INITIALIZED|C_EOF;
5203 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5205 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5206 key->mv_size = mc->mc_db->md_pad;
5207 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5212 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5213 mdb_xcursor_init1(mc, leaf);
5214 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5218 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5223 MDB_GET_KEY(leaf, key);
5228 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5237 case MDB_GET_CURRENT:
5238 if (!(mc->mc_flags & C_INITIALIZED)) {
5241 MDB_page *mp = mc->mc_pg[mc->mc_top];
5243 mc->mc_ki[mc->mc_top] = 0;
5249 key->mv_size = mc->mc_db->md_pad;
5250 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5252 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5253 MDB_GET_KEY(leaf, key);
5255 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5256 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5258 rc = mdb_node_read(mc->mc_txn, leaf, data);
5265 case MDB_GET_BOTH_RANGE:
5266 if (data == NULL || mc->mc_xcursor == NULL) {
5274 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5276 } else if (op == MDB_SET_RANGE)
5277 rc = mdb_cursor_set(mc, key, data, op, NULL);
5279 rc = mdb_cursor_set(mc, key, data, op, &exact);
5281 case MDB_GET_MULTIPLE:
5283 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
5284 !(mc->mc_flags & C_INITIALIZED)) {
5289 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5290 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5293 case MDB_NEXT_MULTIPLE:
5295 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5299 if (!(mc->mc_flags & C_INITIALIZED))
5300 rc = mdb_cursor_first(mc, key, data);
5302 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5303 if (rc == MDB_SUCCESS) {
5304 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5307 mx = &mc->mc_xcursor->mx_cursor;
5308 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5310 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5311 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5319 case MDB_NEXT_NODUP:
5320 if (!(mc->mc_flags & C_INITIALIZED))
5321 rc = mdb_cursor_first(mc, key, data);
5323 rc = mdb_cursor_next(mc, key, data, op);
5327 case MDB_PREV_NODUP:
5328 if (!(mc->mc_flags & C_INITIALIZED)) {
5329 rc = mdb_cursor_last(mc, key, data);
5332 mc->mc_flags |= C_INITIALIZED;
5333 mc->mc_ki[mc->mc_top]++;
5335 rc = mdb_cursor_prev(mc, key, data, op);
5338 rc = mdb_cursor_first(mc, key, data);
5342 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
5343 !(mc->mc_flags & C_INITIALIZED) ||
5344 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5348 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5351 rc = mdb_cursor_last(mc, key, data);
5355 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
5356 !(mc->mc_flags & C_INITIALIZED) ||
5357 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5361 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5364 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5372 /** Touch all the pages in the cursor stack.
5373 * Makes sure all the pages are writable, before attempting a write operation.
5374 * @param[in] mc The cursor to operate on.
5377 mdb_cursor_touch(MDB_cursor *mc)
5381 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5384 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5385 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5388 *mc->mc_dbflag |= DB_DIRTY;
5390 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5391 rc = mdb_page_touch(mc);
5395 mc->mc_top = mc->mc_snum-1;
5399 /** Do not spill pages to disk if txn is getting full, may fail instead */
5400 #define MDB_NOSPILL 0x8000
5403 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5406 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5407 MDB_node *leaf = NULL;
5408 MDB_val xdata, *rdata, dkey;
5411 int do_sub = 0, insert = 0;
5412 unsigned int mcount = 0, dcount = 0, nospill;
5416 char dbuf[MDB_MAXKEYSIZE+1];
5417 unsigned int nflags;
5420 /* Check this first so counter will always be zero on any
5423 if (flags & MDB_MULTIPLE) {
5424 dcount = data[1].mv_size;
5425 data[1].mv_size = 0;
5426 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5430 nospill = flags & MDB_NOSPILL;
5431 flags &= ~MDB_NOSPILL;
5433 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5436 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5439 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5442 #if SIZE_MAX > MAXDATASIZE
5443 if (data->mv_size > MAXDATASIZE)
5447 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5448 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5452 if (flags == MDB_CURRENT) {
5453 if (!(mc->mc_flags & C_INITIALIZED))
5456 } else if (mc->mc_db->md_root == P_INVALID) {
5457 /* new database, cursor has nothing to point to */
5459 mc->mc_flags &= ~C_INITIALIZED;
5464 if (flags & MDB_APPEND) {
5466 rc = mdb_cursor_last(mc, &k2, &d2);
5468 rc = mc->mc_dbx->md_cmp(key, &k2);
5471 mc->mc_ki[mc->mc_top]++;
5473 /* new key is <= last key */
5478 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5480 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5481 DPRINTF("duplicate key [%s]", DKEY(key));
5483 return MDB_KEYEXIST;
5485 if (rc && rc != MDB_NOTFOUND)
5489 /* Cursor is positioned, check for room in the dirty list */
5491 if (flags & MDB_MULTIPLE) {
5493 xdata.mv_size = data->mv_size * dcount;
5497 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5501 if (rc == MDB_NO_ROOT) {
5503 /* new database, write a root leaf page */
5504 DPUTS("allocating new root leaf page");
5505 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5508 mdb_cursor_push(mc, np);
5509 mc->mc_db->md_root = np->mp_pgno;
5510 mc->mc_db->md_depth++;
5511 *mc->mc_dbflag |= DB_DIRTY;
5512 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5514 np->mp_flags |= P_LEAF2;
5515 mc->mc_flags |= C_INITIALIZED;
5517 /* make sure all cursor pages are writable */
5518 rc2 = mdb_cursor_touch(mc);
5523 /* The key already exists */
5524 if (rc == MDB_SUCCESS) {
5525 /* there's only a key anyway, so this is a no-op */
5526 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5527 unsigned int ksize = mc->mc_db->md_pad;
5528 if (key->mv_size != ksize)
5530 if (flags == MDB_CURRENT) {
5531 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5532 memcpy(ptr, key->mv_data, ksize);
5537 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5540 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5541 /* Was a single item before, must convert now */
5543 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5544 /* Just overwrite the current item */
5545 if (flags == MDB_CURRENT)
5548 dkey.mv_size = NODEDSZ(leaf);
5549 dkey.mv_data = NODEDATA(leaf);
5550 #if UINT_MAX < SIZE_MAX
5551 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5552 #ifdef MISALIGNED_OK
5553 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5555 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5558 /* if data matches, ignore it */
5559 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5560 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5562 /* create a fake page for the dup items */
5563 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5564 dkey.mv_data = dbuf;
5565 fp = (MDB_page *)&pbuf;
5566 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5567 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5568 fp->mp_lower = PAGEHDRSZ;
5569 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5570 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5571 fp->mp_flags |= P_LEAF2;
5572 fp->mp_pad = data->mv_size;
5573 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5575 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5576 (dkey.mv_size & 1) + (data->mv_size & 1);
5578 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5581 xdata.mv_size = fp->mp_upper;
5586 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5587 /* See if we need to convert from fake page to subDB */
5589 unsigned int offset;
5593 fp = NODEDATA(leaf);
5594 if (flags == MDB_CURRENT) {
5596 fp->mp_flags |= P_DIRTY;
5597 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5598 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5602 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5603 offset = fp->mp_pad;
5604 if (SIZELEFT(fp) >= offset)
5606 offset *= 4; /* space for 4 more */
5608 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5610 offset += offset & 1;
5611 fp_flags = fp->mp_flags;
5612 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5613 offset >= mc->mc_txn->mt_env->me_nodemax) {
5614 /* yes, convert it */
5616 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5617 dummy.md_pad = fp->mp_pad;
5618 dummy.md_flags = MDB_DUPFIXED;
5619 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5620 dummy.md_flags |= MDB_INTEGERKEY;
5623 dummy.md_branch_pages = 0;
5624 dummy.md_leaf_pages = 1;
5625 dummy.md_overflow_pages = 0;
5626 dummy.md_entries = NUMKEYS(fp);
5628 xdata.mv_size = sizeof(MDB_db);
5629 xdata.mv_data = &dummy;
5630 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5632 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5633 flags |= F_DUPDATA|F_SUBDATA;
5634 dummy.md_root = mp->mp_pgno;
5635 fp_flags &= ~P_SUBP;
5637 /* no, just grow it */
5639 xdata.mv_size = NODEDSZ(leaf) + offset;
5640 xdata.mv_data = &pbuf;
5641 mp = (MDB_page *)&pbuf;
5642 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5645 mp->mp_flags = fp_flags | P_DIRTY;
5646 mp->mp_pad = fp->mp_pad;
5647 mp->mp_lower = fp->mp_lower;
5648 mp->mp_upper = fp->mp_upper + offset;
5650 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5652 nsize = NODEDSZ(leaf) - fp->mp_upper;
5653 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5654 for (i=0; i<NUMKEYS(fp); i++)
5655 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5657 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5661 /* data is on sub-DB, just store it */
5662 flags |= F_DUPDATA|F_SUBDATA;
5666 /* overflow page overwrites need special handling */
5667 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5670 unsigned psize = mc->mc_txn->mt_env->me_psize;
5671 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5673 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5674 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5676 ovpages = omp->mp_pages;
5678 /* Is the ov page large enough? */
5679 if (ovpages >= dpages) {
5680 if (!(omp->mp_flags & P_DIRTY) &&
5681 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5683 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5686 level = 0; /* dirty in this txn or clean */
5689 if (omp->mp_flags & P_DIRTY) {
5690 /* yes, overwrite it. Note in this case we don't
5691 * bother to try shrinking the page if the new data
5692 * is smaller than the overflow threshold.
5695 /* It is writable only in a parent txn */
5696 size_t sz = (size_t) psize * ovpages, off;
5697 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5703 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5704 if (!(flags & MDB_RESERVE)) {
5705 /* Copy end of page, adjusting alignment so
5706 * compiler may copy words instead of bytes.
5708 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5709 memcpy((size_t *)((char *)np + off),
5710 (size_t *)((char *)omp + off), sz - off);
5713 memcpy(np, omp, sz); /* Copy beginning of page */
5716 SETDSZ(leaf, data->mv_size);
5717 if (F_ISSET(flags, MDB_RESERVE))
5718 data->mv_data = METADATA(omp);
5720 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5724 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5726 } else if (NODEDSZ(leaf) == data->mv_size) {
5727 /* same size, just replace it. Note that we could
5728 * also reuse this node if the new data is smaller,
5729 * but instead we opt to shrink the node in that case.
5731 if (F_ISSET(flags, MDB_RESERVE))
5732 data->mv_data = NODEDATA(leaf);
5733 else if (data->mv_size)
5734 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5736 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5739 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5740 mc->mc_db->md_entries--;
5742 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5749 nflags = flags & NODE_ADD_FLAGS;
5750 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5751 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5752 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5753 nflags &= ~MDB_APPEND;
5755 nflags |= MDB_SPLIT_REPLACE;
5756 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5758 /* There is room already in this leaf page. */
5759 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5760 if (rc == 0 && !do_sub && insert) {
5761 /* Adjust other cursors pointing to mp */
5762 MDB_cursor *m2, *m3;
5763 MDB_dbi dbi = mc->mc_dbi;
5764 unsigned i = mc->mc_top;
5765 MDB_page *mp = mc->mc_pg[i];
5767 if (mc->mc_flags & C_SUB)
5770 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5771 if (mc->mc_flags & C_SUB)
5772 m3 = &m2->mc_xcursor->mx_cursor;
5775 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5776 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5783 if (rc != MDB_SUCCESS)
5784 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5786 /* Now store the actual data in the child DB. Note that we're
5787 * storing the user data in the keys field, so there are strict
5788 * size limits on dupdata. The actual data fields of the child
5789 * DB are all zero size.
5796 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5797 if (flags & MDB_CURRENT) {
5798 xflags = MDB_CURRENT|MDB_NOSPILL;
5800 mdb_xcursor_init1(mc, leaf);
5801 xflags = (flags & MDB_NODUPDATA) ?
5802 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5804 /* converted, write the original data first */
5806 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5810 /* Adjust other cursors pointing to mp */
5812 unsigned i = mc->mc_top;
5813 MDB_page *mp = mc->mc_pg[i];
5815 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5816 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5817 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5818 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5819 mdb_xcursor_init1(m2, leaf);
5823 /* we've done our job */
5826 if (flags & MDB_APPENDDUP)
5827 xflags |= MDB_APPEND;
5828 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5829 if (flags & F_SUBDATA) {
5830 void *db = NODEDATA(leaf);
5831 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5834 /* sub-writes might have failed so check rc again.
5835 * Don't increment count if we just replaced an existing item.
5837 if (!rc && !(flags & MDB_CURRENT))
5838 mc->mc_db->md_entries++;
5839 if (flags & MDB_MULTIPLE) {
5842 if (mcount < dcount) {
5843 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5844 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5848 /* let caller know how many succeeded, if any */
5849 data[1].mv_size = mcount;
5853 /* If we succeeded and the key didn't exist before, make sure
5854 * the cursor is marked valid.
5857 mc->mc_flags |= C_INITIALIZED;
5862 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5867 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5870 if (!(mc->mc_flags & C_INITIALIZED))
5873 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
5875 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
5877 rc = mdb_cursor_touch(mc);
5881 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5883 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5884 if (flags != MDB_NODUPDATA) {
5885 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5886 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5888 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
5889 /* If sub-DB still has entries, we're done */
5890 if (mc->mc_xcursor->mx_db.md_entries) {
5891 if (leaf->mn_flags & F_SUBDATA) {
5892 /* update subDB info */
5893 void *db = NODEDATA(leaf);
5894 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5897 /* shrink fake page */
5898 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5899 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5900 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5901 /* fix other sub-DB cursors pointed at this fake page */
5902 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5903 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5904 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5905 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5906 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5909 mc->mc_db->md_entries--;
5912 /* otherwise fall thru and delete the sub-DB */
5915 if (leaf->mn_flags & F_SUBDATA) {
5916 /* add all the child DB's pages to the free list */
5917 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5918 if (rc == MDB_SUCCESS) {
5919 mc->mc_db->md_entries -=
5920 mc->mc_xcursor->mx_db.md_entries;
5925 return mdb_cursor_del0(mc, leaf);
5928 /** Allocate and initialize new pages for a database.
5929 * @param[in] mc a cursor on the database being added to.
5930 * @param[in] flags flags defining what type of page is being allocated.
5931 * @param[in] num the number of pages to allocate. This is usually 1,
5932 * unless allocating overflow pages for a large record.
5933 * @param[out] mp Address of a page, or NULL on failure.
5934 * @return 0 on success, non-zero on failure.
5937 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5942 if ((rc = mdb_page_alloc(mc, num, &np)))
5944 DPRINTF("allocated new mpage %zu, page size %u",
5945 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5946 np->mp_flags = flags | P_DIRTY;
5947 np->mp_lower = PAGEHDRSZ;
5948 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5951 mc->mc_db->md_branch_pages++;
5952 else if (IS_LEAF(np))
5953 mc->mc_db->md_leaf_pages++;
5954 else if (IS_OVERFLOW(np)) {
5955 mc->mc_db->md_overflow_pages += num;
5963 /** Calculate the size of a leaf node.
5964 * The size depends on the environment's page size; if a data item
5965 * is too large it will be put onto an overflow page and the node
5966 * size will only include the key and not the data. Sizes are always
5967 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5968 * of the #MDB_node headers.
5969 * @param[in] env The environment handle.
5970 * @param[in] key The key for the node.
5971 * @param[in] data The data for the node.
5972 * @return The number of bytes needed to store the node.
5975 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5979 sz = LEAFSIZE(key, data);
5980 if (sz >= env->me_nodemax) {
5981 /* put on overflow page */
5982 sz -= data->mv_size - sizeof(pgno_t);
5986 return sz + sizeof(indx_t);
5989 /** Calculate the size of a branch node.
5990 * The size should depend on the environment's page size but since
5991 * we currently don't support spilling large keys onto overflow
5992 * pages, it's simply the size of the #MDB_node header plus the
5993 * size of the key. Sizes are always rounded up to an even number
5994 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5995 * @param[in] env The environment handle.
5996 * @param[in] key The key for the node.
5997 * @return The number of bytes needed to store the node.
6000 mdb_branch_size(MDB_env *env, MDB_val *key)
6005 if (sz >= env->me_nodemax) {
6006 /* put on overflow page */
6007 /* not implemented */
6008 /* sz -= key->size - sizeof(pgno_t); */
6011 return sz + sizeof(indx_t);
6014 /** Add a node to the page pointed to by the cursor.
6015 * @param[in] mc The cursor for this operation.
6016 * @param[in] indx The index on the page where the new node should be added.
6017 * @param[in] key The key for the new node.
6018 * @param[in] data The data for the new node, if any.
6019 * @param[in] pgno The page number, if adding a branch node.
6020 * @param[in] flags Flags for the node.
6021 * @return 0 on success, non-zero on failure. Possible errors are:
6023 * <li>ENOMEM - failed to allocate overflow pages for the node.
6024 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6025 * should never happen since all callers already calculate the
6026 * page's free space before calling this function.
6030 mdb_node_add(MDB_cursor *mc, indx_t indx,
6031 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6034 size_t node_size = NODESIZE;
6037 MDB_page *mp = mc->mc_pg[mc->mc_top];
6038 MDB_page *ofp = NULL; /* overflow page */
6041 assert(mp->mp_upper >= mp->mp_lower);
6043 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
6044 IS_LEAF(mp) ? "leaf" : "branch",
6045 IS_SUBP(mp) ? "sub-" : "",
6046 mp->mp_pgno, indx, data ? data->mv_size : 0,
6047 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
6050 /* Move higher keys up one slot. */
6051 int ksize = mc->mc_db->md_pad, dif;
6052 char *ptr = LEAF2KEY(mp, indx, ksize);
6053 dif = NUMKEYS(mp) - indx;
6055 memmove(ptr+ksize, ptr, dif*ksize);
6056 /* insert new key */
6057 memcpy(ptr, key->mv_data, ksize);
6059 /* Just using these for counting */
6060 mp->mp_lower += sizeof(indx_t);
6061 mp->mp_upper -= ksize - sizeof(indx_t);
6066 node_size += key->mv_size;
6070 if (F_ISSET(flags, F_BIGDATA)) {
6071 /* Data already on overflow page. */
6072 node_size += sizeof(pgno_t);
6073 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6074 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6076 /* Put data on overflow page. */
6077 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
6078 data->mv_size, node_size+data->mv_size);
6079 node_size += sizeof(pgno_t);
6080 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6082 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
6085 node_size += data->mv_size;
6088 node_size += node_size & 1;
6090 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6091 DPRINTF("not enough room in page %zu, got %u ptrs",
6092 mp->mp_pgno, NUMKEYS(mp));
6093 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6094 mp->mp_upper - mp->mp_lower);
6095 DPRINTF("node size = %zu", node_size);
6096 return MDB_PAGE_FULL;
6099 /* Move higher pointers up one slot. */
6100 for (i = NUMKEYS(mp); i > indx; i--)
6101 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6103 /* Adjust free space offsets. */
6104 ofs = mp->mp_upper - node_size;
6105 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6106 mp->mp_ptrs[indx] = ofs;
6108 mp->mp_lower += sizeof(indx_t);
6110 /* Write the node data. */
6111 node = NODEPTR(mp, indx);
6112 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6113 node->mn_flags = flags;
6115 SETDSZ(node,data->mv_size);
6120 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6125 if (F_ISSET(flags, F_BIGDATA))
6126 memcpy(node->mn_data + key->mv_size, data->mv_data,
6128 else if (F_ISSET(flags, MDB_RESERVE))
6129 data->mv_data = node->mn_data + key->mv_size;
6131 memcpy(node->mn_data + key->mv_size, data->mv_data,
6134 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6136 if (F_ISSET(flags, MDB_RESERVE))
6137 data->mv_data = METADATA(ofp);
6139 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6146 /** Delete the specified node from a page.
6147 * @param[in] mp The page to operate on.
6148 * @param[in] indx The index of the node to delete.
6149 * @param[in] ksize The size of a node. Only used if the page is
6150 * part of a #MDB_DUPFIXED database.
6153 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6156 indx_t i, j, numkeys, ptr;
6163 COPY_PGNO(pgno, mp->mp_pgno);
6164 DPRINTF("delete node %u on %s page %zu", indx,
6165 IS_LEAF(mp) ? "leaf" : "branch", pgno);
6168 assert(indx < NUMKEYS(mp));
6171 int x = NUMKEYS(mp) - 1 - indx;
6172 base = LEAF2KEY(mp, indx, ksize);
6174 memmove(base, base + ksize, x * ksize);
6175 mp->mp_lower -= sizeof(indx_t);
6176 mp->mp_upper += ksize - sizeof(indx_t);
6180 node = NODEPTR(mp, indx);
6181 sz = NODESIZE + node->mn_ksize;
6183 if (F_ISSET(node->mn_flags, F_BIGDATA))
6184 sz += sizeof(pgno_t);
6186 sz += NODEDSZ(node);
6190 ptr = mp->mp_ptrs[indx];
6191 numkeys = NUMKEYS(mp);
6192 for (i = j = 0; i < numkeys; i++) {
6194 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6195 if (mp->mp_ptrs[i] < ptr)
6196 mp->mp_ptrs[j] += sz;
6201 base = (char *)mp + mp->mp_upper;
6202 memmove(base + sz, base, ptr - mp->mp_upper);
6204 mp->mp_lower -= sizeof(indx_t);
6208 /** Compact the main page after deleting a node on a subpage.
6209 * @param[in] mp The main page to operate on.
6210 * @param[in] indx The index of the subpage on the main page.
6213 mdb_node_shrink(MDB_page *mp, indx_t indx)
6220 indx_t i, numkeys, ptr;
6222 node = NODEPTR(mp, indx);
6223 sp = (MDB_page *)NODEDATA(node);
6224 osize = NODEDSZ(node);
6226 delta = sp->mp_upper - sp->mp_lower;
6227 SETDSZ(node, osize - delta);
6228 xp = (MDB_page *)((char *)sp + delta);
6230 /* shift subpage upward */
6232 nsize = NUMKEYS(sp) * sp->mp_pad;
6233 memmove(METADATA(xp), METADATA(sp), nsize);
6236 nsize = osize - sp->mp_upper;
6237 numkeys = NUMKEYS(sp);
6238 for (i=numkeys-1; i>=0; i--)
6239 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6241 xp->mp_upper = sp->mp_lower;
6242 xp->mp_lower = sp->mp_lower;
6243 xp->mp_flags = sp->mp_flags;
6244 xp->mp_pad = sp->mp_pad;
6245 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6247 /* shift lower nodes upward */
6248 ptr = mp->mp_ptrs[indx];
6249 numkeys = NUMKEYS(mp);
6250 for (i = 0; i < numkeys; i++) {
6251 if (mp->mp_ptrs[i] <= ptr)
6252 mp->mp_ptrs[i] += delta;
6255 base = (char *)mp + mp->mp_upper;
6256 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6257 mp->mp_upper += delta;
6260 /** Initial setup of a sorted-dups cursor.
6261 * Sorted duplicates are implemented as a sub-database for the given key.
6262 * The duplicate data items are actually keys of the sub-database.
6263 * Operations on the duplicate data items are performed using a sub-cursor
6264 * initialized when the sub-database is first accessed. This function does
6265 * the preliminary setup of the sub-cursor, filling in the fields that
6266 * depend only on the parent DB.
6267 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6270 mdb_xcursor_init0(MDB_cursor *mc)
6272 MDB_xcursor *mx = mc->mc_xcursor;
6274 mx->mx_cursor.mc_xcursor = NULL;
6275 mx->mx_cursor.mc_txn = mc->mc_txn;
6276 mx->mx_cursor.mc_db = &mx->mx_db;
6277 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6278 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6279 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6280 mx->mx_cursor.mc_snum = 0;
6281 mx->mx_cursor.mc_top = 0;
6282 mx->mx_cursor.mc_flags = C_SUB;
6283 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6284 mx->mx_dbx.md_dcmp = NULL;
6285 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6288 /** Final setup of a sorted-dups cursor.
6289 * Sets up the fields that depend on the data from the main cursor.
6290 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6291 * @param[in] node The data containing the #MDB_db record for the
6292 * sorted-dup database.
6295 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6297 MDB_xcursor *mx = mc->mc_xcursor;
6299 if (node->mn_flags & F_SUBDATA) {
6300 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6301 mx->mx_cursor.mc_pg[0] = 0;
6302 mx->mx_cursor.mc_snum = 0;
6303 mx->mx_cursor.mc_flags = C_SUB;
6305 MDB_page *fp = NODEDATA(node);
6306 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6307 mx->mx_db.md_flags = 0;
6308 mx->mx_db.md_depth = 1;
6309 mx->mx_db.md_branch_pages = 0;
6310 mx->mx_db.md_leaf_pages = 1;
6311 mx->mx_db.md_overflow_pages = 0;
6312 mx->mx_db.md_entries = NUMKEYS(fp);
6313 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6314 mx->mx_cursor.mc_snum = 1;
6315 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6316 mx->mx_cursor.mc_top = 0;
6317 mx->mx_cursor.mc_pg[0] = fp;
6318 mx->mx_cursor.mc_ki[0] = 0;
6319 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6320 mx->mx_db.md_flags = MDB_DUPFIXED;
6321 mx->mx_db.md_pad = fp->mp_pad;
6322 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6323 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6326 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6328 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6330 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6331 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6332 #if UINT_MAX < SIZE_MAX
6333 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6334 #ifdef MISALIGNED_OK
6335 mx->mx_dbx.md_cmp = mdb_cmp_long;
6337 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6342 /** Initialize a cursor for a given transaction and database. */
6344 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6347 mc->mc_backup = NULL;
6350 mc->mc_db = &txn->mt_dbs[dbi];
6351 mc->mc_dbx = &txn->mt_dbxs[dbi];
6352 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6357 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6359 mc->mc_xcursor = mx;
6360 mdb_xcursor_init0(mc);
6362 mc->mc_xcursor = NULL;
6364 if (*mc->mc_dbflag & DB_STALE) {
6365 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6370 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6373 size_t size = sizeof(MDB_cursor);
6375 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6378 /* Allow read access to the freelist */
6379 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6382 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6383 size += sizeof(MDB_xcursor);
6385 if ((mc = malloc(size)) != NULL) {
6386 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6387 if (txn->mt_cursors) {
6388 mc->mc_next = txn->mt_cursors[dbi];
6389 txn->mt_cursors[dbi] = mc;
6390 mc->mc_flags |= C_UNTRACK;
6402 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6404 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6407 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6410 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6414 /* Return the count of duplicate data items for the current key */
6416 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6420 if (mc == NULL || countp == NULL)
6423 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
6426 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6427 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6430 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6433 *countp = mc->mc_xcursor->mx_db.md_entries;
6439 mdb_cursor_close(MDB_cursor *mc)
6441 if (mc && !mc->mc_backup) {
6442 /* remove from txn, if tracked */
6443 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6444 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6445 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6447 *prev = mc->mc_next;
6454 mdb_cursor_txn(MDB_cursor *mc)
6456 if (!mc) return NULL;
6461 mdb_cursor_dbi(MDB_cursor *mc)
6467 /** Replace the key for a node with a new key.
6468 * @param[in] mc Cursor pointing to the node to operate on.
6469 * @param[in] key The new key to use.
6470 * @return 0 on success, non-zero on failure.
6473 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6480 indx_t ptr, i, numkeys, indx;
6483 indx = mc->mc_ki[mc->mc_top];
6484 mp = mc->mc_pg[mc->mc_top];
6485 node = NODEPTR(mp, indx);
6486 ptr = mp->mp_ptrs[indx];
6490 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6491 k2.mv_data = NODEKEY(node);
6492 k2.mv_size = node->mn_ksize;
6493 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6495 mdb_dkey(&k2, kbuf2),
6501 delta0 = delta = key->mv_size - node->mn_ksize;
6503 /* Must be 2-byte aligned. If new key is
6504 * shorter by 1, the shift will be skipped.
6506 delta += (delta & 1);
6508 if (delta > 0 && SIZELEFT(mp) < delta) {
6510 /* not enough space left, do a delete and split */
6511 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6512 pgno = NODEPGNO(node);
6513 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6514 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6517 numkeys = NUMKEYS(mp);
6518 for (i = 0; i < numkeys; i++) {
6519 if (mp->mp_ptrs[i] <= ptr)
6520 mp->mp_ptrs[i] -= delta;
6523 base = (char *)mp + mp->mp_upper;
6524 len = ptr - mp->mp_upper + NODESIZE;
6525 memmove(base - delta, base, len);
6526 mp->mp_upper -= delta;
6528 node = NODEPTR(mp, indx);
6531 /* But even if no shift was needed, update ksize */
6533 node->mn_ksize = key->mv_size;
6536 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6542 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6544 /** Move a node from csrc to cdst.
6547 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6554 unsigned short flags;
6558 /* Mark src and dst as dirty. */
6559 if ((rc = mdb_page_touch(csrc)) ||
6560 (rc = mdb_page_touch(cdst)))
6563 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6564 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6565 key.mv_size = csrc->mc_db->md_pad;
6566 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6568 data.mv_data = NULL;
6572 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6573 assert(!((long)srcnode&1));
6574 srcpg = NODEPGNO(srcnode);
6575 flags = srcnode->mn_flags;
6576 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6577 unsigned int snum = csrc->mc_snum;
6579 /* must find the lowest key below src */
6580 mdb_page_search_lowest(csrc);
6581 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6582 key.mv_size = csrc->mc_db->md_pad;
6583 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6585 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6586 key.mv_size = NODEKSZ(s2);
6587 key.mv_data = NODEKEY(s2);
6589 csrc->mc_snum = snum--;
6590 csrc->mc_top = snum;
6592 key.mv_size = NODEKSZ(srcnode);
6593 key.mv_data = NODEKEY(srcnode);
6595 data.mv_size = NODEDSZ(srcnode);
6596 data.mv_data = NODEDATA(srcnode);
6598 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6599 unsigned int snum = cdst->mc_snum;
6602 /* must find the lowest key below dst */
6603 mdb_page_search_lowest(cdst);
6604 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6605 bkey.mv_size = cdst->mc_db->md_pad;
6606 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6608 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6609 bkey.mv_size = NODEKSZ(s2);
6610 bkey.mv_data = NODEKEY(s2);
6612 cdst->mc_snum = snum--;
6613 cdst->mc_top = snum;
6614 mdb_cursor_copy(cdst, &mn);
6616 rc = mdb_update_key(&mn, &bkey);
6621 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6622 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6623 csrc->mc_ki[csrc->mc_top],
6625 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6626 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6628 /* Add the node to the destination page.
6630 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6631 if (rc != MDB_SUCCESS)
6634 /* Delete the node from the source page.
6636 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6639 /* Adjust other cursors pointing to mp */
6640 MDB_cursor *m2, *m3;
6641 MDB_dbi dbi = csrc->mc_dbi;
6642 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6644 if (csrc->mc_flags & C_SUB)
6647 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6648 if (csrc->mc_flags & C_SUB)
6649 m3 = &m2->mc_xcursor->mx_cursor;
6652 if (m3 == csrc) continue;
6653 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6654 csrc->mc_ki[csrc->mc_top]) {
6655 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6656 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6661 /* Update the parent separators.
6663 if (csrc->mc_ki[csrc->mc_top] == 0) {
6664 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6665 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6666 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6668 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6669 key.mv_size = NODEKSZ(srcnode);
6670 key.mv_data = NODEKEY(srcnode);
6672 DPRINTF("update separator for source page %zu to [%s]",
6673 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6674 mdb_cursor_copy(csrc, &mn);
6677 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6680 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6682 indx_t ix = csrc->mc_ki[csrc->mc_top];
6683 nullkey.mv_size = 0;
6684 csrc->mc_ki[csrc->mc_top] = 0;
6685 rc = mdb_update_key(csrc, &nullkey);
6686 csrc->mc_ki[csrc->mc_top] = ix;
6687 assert(rc == MDB_SUCCESS);
6691 if (cdst->mc_ki[cdst->mc_top] == 0) {
6692 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6693 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6694 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6696 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6697 key.mv_size = NODEKSZ(srcnode);
6698 key.mv_data = NODEKEY(srcnode);
6700 DPRINTF("update separator for destination page %zu to [%s]",
6701 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6702 mdb_cursor_copy(cdst, &mn);
6705 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6708 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6710 indx_t ix = cdst->mc_ki[cdst->mc_top];
6711 nullkey.mv_size = 0;
6712 cdst->mc_ki[cdst->mc_top] = 0;
6713 rc = mdb_update_key(cdst, &nullkey);
6714 cdst->mc_ki[cdst->mc_top] = ix;
6715 assert(rc == MDB_SUCCESS);
6722 /** Merge one page into another.
6723 * The nodes from the page pointed to by \b csrc will
6724 * be copied to the page pointed to by \b cdst and then
6725 * the \b csrc page will be freed.
6726 * @param[in] csrc Cursor pointing to the source page.
6727 * @param[in] cdst Cursor pointing to the destination page.
6730 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6738 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6739 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6741 assert(csrc->mc_snum > 1); /* can't merge root page */
6742 assert(cdst->mc_snum > 1);
6744 /* Mark dst as dirty. */
6745 if ((rc = mdb_page_touch(cdst)))
6748 /* Move all nodes from src to dst.
6750 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6751 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6752 key.mv_size = csrc->mc_db->md_pad;
6753 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6754 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6755 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6756 if (rc != MDB_SUCCESS)
6758 key.mv_data = (char *)key.mv_data + key.mv_size;
6761 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6762 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6763 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6764 unsigned int snum = csrc->mc_snum;
6766 /* must find the lowest key below src */
6767 mdb_page_search_lowest(csrc);
6768 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6769 key.mv_size = csrc->mc_db->md_pad;
6770 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6772 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6773 key.mv_size = NODEKSZ(s2);
6774 key.mv_data = NODEKEY(s2);
6776 csrc->mc_snum = snum--;
6777 csrc->mc_top = snum;
6779 key.mv_size = srcnode->mn_ksize;
6780 key.mv_data = NODEKEY(srcnode);
6783 data.mv_size = NODEDSZ(srcnode);
6784 data.mv_data = NODEDATA(srcnode);
6785 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6786 if (rc != MDB_SUCCESS)
6791 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6792 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]), (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10);
6794 /* Unlink the src page from parent and add to free list.
6796 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6797 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6800 rc = mdb_update_key(csrc, &key);
6806 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6807 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6810 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6811 csrc->mc_db->md_leaf_pages--;
6813 csrc->mc_db->md_branch_pages--;
6815 /* Adjust other cursors pointing to mp */
6816 MDB_cursor *m2, *m3;
6817 MDB_dbi dbi = csrc->mc_dbi;
6818 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6820 if (csrc->mc_flags & C_SUB)
6823 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6824 if (csrc->mc_flags & C_SUB)
6825 m3 = &m2->mc_xcursor->mx_cursor;
6828 if (m3 == csrc) continue;
6829 if (m3->mc_snum < csrc->mc_snum) continue;
6830 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6831 m3->mc_pg[csrc->mc_top] = mp;
6832 m3->mc_ki[csrc->mc_top] += nkeys;
6836 mdb_cursor_pop(csrc);
6838 return mdb_rebalance(csrc);
6841 /** Copy the contents of a cursor.
6842 * @param[in] csrc The cursor to copy from.
6843 * @param[out] cdst The cursor to copy to.
6846 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6850 cdst->mc_txn = csrc->mc_txn;
6851 cdst->mc_dbi = csrc->mc_dbi;
6852 cdst->mc_db = csrc->mc_db;
6853 cdst->mc_dbx = csrc->mc_dbx;
6854 cdst->mc_snum = csrc->mc_snum;
6855 cdst->mc_top = csrc->mc_top;
6856 cdst->mc_flags = csrc->mc_flags;
6858 for (i=0; i<csrc->mc_snum; i++) {
6859 cdst->mc_pg[i] = csrc->mc_pg[i];
6860 cdst->mc_ki[i] = csrc->mc_ki[i];
6864 /** Rebalance the tree after a delete operation.
6865 * @param[in] mc Cursor pointing to the page where rebalancing
6867 * @return 0 on success, non-zero on failure.
6870 mdb_rebalance(MDB_cursor *mc)
6874 unsigned int ptop, minkeys;
6877 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6881 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6882 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6883 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6884 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6888 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6889 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6892 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6893 DPRINTF("no need to rebalance page %zu, above fill threshold",
6899 if (mc->mc_snum < 2) {
6900 MDB_page *mp = mc->mc_pg[0];
6902 DPUTS("Can't rebalance a subpage, ignoring");
6905 if (NUMKEYS(mp) == 0) {
6906 DPUTS("tree is completely empty");
6907 mc->mc_db->md_root = P_INVALID;
6908 mc->mc_db->md_depth = 0;
6909 mc->mc_db->md_leaf_pages = 0;
6910 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6913 /* Adjust cursors pointing to mp */
6917 MDB_cursor *m2, *m3;
6918 MDB_dbi dbi = mc->mc_dbi;
6920 if (mc->mc_flags & C_SUB)
6923 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6924 if (mc->mc_flags & C_SUB)
6925 m3 = &m2->mc_xcursor->mx_cursor;
6928 if (m3->mc_snum < mc->mc_snum) continue;
6929 if (m3->mc_pg[0] == mp) {
6935 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6936 DPUTS("collapsing root page!");
6937 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6940 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6941 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6944 mc->mc_db->md_depth--;
6945 mc->mc_db->md_branch_pages--;
6946 mc->mc_ki[0] = mc->mc_ki[1];
6948 /* Adjust other cursors pointing to mp */
6949 MDB_cursor *m2, *m3;
6950 MDB_dbi dbi = mc->mc_dbi;
6952 if (mc->mc_flags & C_SUB)
6955 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6956 if (mc->mc_flags & C_SUB)
6957 m3 = &m2->mc_xcursor->mx_cursor;
6960 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6961 if (m3->mc_pg[0] == mp) {
6962 m3->mc_pg[0] = mc->mc_pg[0];
6965 m3->mc_ki[0] = m3->mc_ki[1];
6970 DPUTS("root page doesn't need rebalancing");
6974 /* The parent (branch page) must have at least 2 pointers,
6975 * otherwise the tree is invalid.
6977 ptop = mc->mc_top-1;
6978 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6980 /* Leaf page fill factor is below the threshold.
6981 * Try to move keys from left or right neighbor, or
6982 * merge with a neighbor page.
6987 mdb_cursor_copy(mc, &mn);
6988 mn.mc_xcursor = NULL;
6990 if (mc->mc_ki[ptop] == 0) {
6991 /* We're the leftmost leaf in our parent.
6993 DPUTS("reading right neighbor");
6995 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6996 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6999 mn.mc_ki[mn.mc_top] = 0;
7000 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7002 /* There is at least one neighbor to the left.
7004 DPUTS("reading left neighbor");
7006 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7007 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7010 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7011 mc->mc_ki[mc->mc_top] = 0;
7014 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
7015 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10);
7017 /* If the neighbor page is above threshold and has enough keys,
7018 * move one key from it. Otherwise we should try to merge them.
7019 * (A branch page must never have less than 2 keys.)
7021 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7022 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7023 return mdb_node_move(&mn, mc);
7025 if (mc->mc_ki[ptop] == 0)
7026 rc = mdb_page_merge(&mn, mc);
7028 rc = mdb_page_merge(mc, &mn);
7029 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7034 /** Complete a delete operation started by #mdb_cursor_del(). */
7036 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7042 mp = mc->mc_pg[mc->mc_top];
7043 ki = mc->mc_ki[mc->mc_top];
7045 /* add overflow pages to free list */
7046 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7050 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7051 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7052 (rc = mdb_ovpage_free(mc, omp)))
7055 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7056 mc->mc_db->md_entries--;
7057 rc = mdb_rebalance(mc);
7058 if (rc != MDB_SUCCESS)
7059 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7060 /* if mc points past last node in page, invalidate */
7061 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7062 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7065 /* Adjust other cursors pointing to mp */
7068 MDB_dbi dbi = mc->mc_dbi;
7070 mp = mc->mc_pg[mc->mc_top];
7071 nkeys = NUMKEYS(mp);
7072 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7075 if (!(m2->mc_flags & C_INITIALIZED))
7077 if (m2->mc_pg[mc->mc_top] == mp) {
7078 if (m2->mc_ki[mc->mc_top] > ki)
7079 m2->mc_ki[mc->mc_top]--;
7080 if (m2->mc_ki[mc->mc_top] >= nkeys)
7081 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7090 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7091 MDB_val *key, MDB_val *data)
7096 MDB_val rdata, *xdata;
7100 assert(key != NULL);
7102 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
7104 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7107 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7111 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7115 mdb_cursor_init(&mc, txn, dbi, &mx);
7126 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7128 /* let mdb_page_split know about this cursor if needed:
7129 * delete will trigger a rebalance; if it needs to move
7130 * a node from one page to another, it will have to
7131 * update the parent's separator key(s). If the new sepkey
7132 * is larger than the current one, the parent page may
7133 * run out of space, triggering a split. We need this
7134 * cursor to be consistent until the end of the rebalance.
7136 mc.mc_flags |= C_UNTRACK;
7137 mc.mc_next = txn->mt_cursors[dbi];
7138 txn->mt_cursors[dbi] = &mc;
7139 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7140 txn->mt_cursors[dbi] = mc.mc_next;
7145 /** Split a page and insert a new node.
7146 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7147 * The cursor will be updated to point to the actual page and index where
7148 * the node got inserted after the split.
7149 * @param[in] newkey The key for the newly inserted node.
7150 * @param[in] newdata The data for the newly inserted node.
7151 * @param[in] newpgno The page number, if the new node is a branch node.
7152 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7153 * @return 0 on success, non-zero on failure.
7156 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7157 unsigned int nflags)
7160 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7163 unsigned int i, j, split_indx, nkeys, pmax;
7165 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7167 MDB_page *mp, *rp, *pp;
7172 mp = mc->mc_pg[mc->mc_top];
7173 newindx = mc->mc_ki[mc->mc_top];
7175 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
7176 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7177 DKEY(newkey), mc->mc_ki[mc->mc_top]);
7179 /* Create a right sibling. */
7180 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7182 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
7184 if (mc->mc_snum < 2) {
7185 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7187 /* shift current top to make room for new parent */
7188 mc->mc_pg[1] = mc->mc_pg[0];
7189 mc->mc_ki[1] = mc->mc_ki[0];
7192 mc->mc_db->md_root = pp->mp_pgno;
7193 DPRINTF("root split! new root = %zu", pp->mp_pgno);
7194 mc->mc_db->md_depth++;
7197 /* Add left (implicit) pointer. */
7198 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7199 /* undo the pre-push */
7200 mc->mc_pg[0] = mc->mc_pg[1];
7201 mc->mc_ki[0] = mc->mc_ki[1];
7202 mc->mc_db->md_root = mp->mp_pgno;
7203 mc->mc_db->md_depth--;
7210 ptop = mc->mc_top-1;
7211 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
7214 mc->mc_flags |= C_SPLITTING;
7215 mdb_cursor_copy(mc, &mn);
7216 mn.mc_pg[mn.mc_top] = rp;
7217 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7219 if (nflags & MDB_APPEND) {
7220 mn.mc_ki[mn.mc_top] = 0;
7222 split_indx = newindx;
7227 nkeys = NUMKEYS(mp);
7228 split_indx = nkeys / 2;
7229 if (newindx < split_indx)
7235 unsigned int lsize, rsize, ksize;
7236 /* Move half of the keys to the right sibling */
7238 x = mc->mc_ki[mc->mc_top] - split_indx;
7239 ksize = mc->mc_db->md_pad;
7240 split = LEAF2KEY(mp, split_indx, ksize);
7241 rsize = (nkeys - split_indx) * ksize;
7242 lsize = (nkeys - split_indx) * sizeof(indx_t);
7243 mp->mp_lower -= lsize;
7244 rp->mp_lower += lsize;
7245 mp->mp_upper += rsize - lsize;
7246 rp->mp_upper -= rsize - lsize;
7247 sepkey.mv_size = ksize;
7248 if (newindx == split_indx) {
7249 sepkey.mv_data = newkey->mv_data;
7251 sepkey.mv_data = split;
7254 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7255 memcpy(rp->mp_ptrs, split, rsize);
7256 sepkey.mv_data = rp->mp_ptrs;
7257 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7258 memcpy(ins, newkey->mv_data, ksize);
7259 mp->mp_lower += sizeof(indx_t);
7260 mp->mp_upper -= ksize - sizeof(indx_t);
7263 memcpy(rp->mp_ptrs, split, x * ksize);
7264 ins = LEAF2KEY(rp, x, ksize);
7265 memcpy(ins, newkey->mv_data, ksize);
7266 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7267 rp->mp_lower += sizeof(indx_t);
7268 rp->mp_upper -= ksize - sizeof(indx_t);
7269 mc->mc_ki[mc->mc_top] = x;
7270 mc->mc_pg[mc->mc_top] = rp;
7275 /* For leaf pages, check the split point based on what
7276 * fits where, since otherwise mdb_node_add can fail.
7278 * This check is only needed when the data items are
7279 * relatively large, such that being off by one will
7280 * make the difference between success or failure.
7282 * It's also relevant if a page happens to be laid out
7283 * such that one half of its nodes are all "small" and
7284 * the other half of its nodes are "large." If the new
7285 * item is also "large" and falls on the half with
7286 * "large" nodes, it also may not fit.
7289 unsigned int psize, nsize;
7290 /* Maximum free space in an empty page */
7291 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7292 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7293 if ((nkeys < 20) || (nsize > pmax/16)) {
7294 if (newindx <= split_indx) {
7297 for (i=0; i<split_indx; i++) {
7298 node = NODEPTR(mp, i);
7299 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7300 if (F_ISSET(node->mn_flags, F_BIGDATA))
7301 psize += sizeof(pgno_t);
7303 psize += NODEDSZ(node);
7307 split_indx = newindx;
7318 for (i=nkeys-1; i>=split_indx; i--) {
7319 node = NODEPTR(mp, i);
7320 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7321 if (F_ISSET(node->mn_flags, F_BIGDATA))
7322 psize += sizeof(pgno_t);
7324 psize += NODEDSZ(node);
7328 split_indx = newindx;
7339 /* First find the separating key between the split pages.
7340 * The case where newindx == split_indx is ambiguous; the
7341 * new item could go to the new page or stay on the original
7342 * page. If newpos == 1 it goes to the new page.
7344 if (newindx == split_indx && newpos) {
7345 sepkey.mv_size = newkey->mv_size;
7346 sepkey.mv_data = newkey->mv_data;
7348 node = NODEPTR(mp, split_indx);
7349 sepkey.mv_size = node->mn_ksize;
7350 sepkey.mv_data = NODEKEY(node);
7354 DPRINTF("separator is [%s]", DKEY(&sepkey));
7356 /* Copy separator key to the parent.
7358 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7362 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7365 if (mn.mc_snum == mc->mc_snum) {
7366 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7367 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7368 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7369 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7374 /* Right page might now have changed parent.
7375 * Check if left page also changed parent.
7377 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7378 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7379 for (i=0; i<ptop; i++) {
7380 mc->mc_pg[i] = mn.mc_pg[i];
7381 mc->mc_ki[i] = mn.mc_ki[i];
7383 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7384 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7388 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7391 mc->mc_flags ^= C_SPLITTING;
7392 if (rc != MDB_SUCCESS) {
7395 if (nflags & MDB_APPEND) {
7396 mc->mc_pg[mc->mc_top] = rp;
7397 mc->mc_ki[mc->mc_top] = 0;
7398 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7401 for (i=0; i<mc->mc_top; i++)
7402 mc->mc_ki[i] = mn.mc_ki[i];
7409 /* Move half of the keys to the right sibling. */
7411 /* grab a page to hold a temporary copy */
7412 copy = mdb_page_malloc(mc->mc_txn, 1);
7416 copy->mp_pgno = mp->mp_pgno;
7417 copy->mp_flags = mp->mp_flags;
7418 copy->mp_lower = PAGEHDRSZ;
7419 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7420 mc->mc_pg[mc->mc_top] = copy;
7421 for (i = j = 0; i <= nkeys; j++) {
7422 if (i == split_indx) {
7423 /* Insert in right sibling. */
7424 /* Reset insert index for right sibling. */
7425 if (i != newindx || (newpos ^ ins_new)) {
7427 mc->mc_pg[mc->mc_top] = rp;
7431 if (i == newindx && !ins_new) {
7432 /* Insert the original entry that caused the split. */
7433 rkey.mv_data = newkey->mv_data;
7434 rkey.mv_size = newkey->mv_size;
7443 /* Update index for the new key. */
7444 mc->mc_ki[mc->mc_top] = j;
7445 } else if (i == nkeys) {
7448 node = NODEPTR(mp, i);
7449 rkey.mv_data = NODEKEY(node);
7450 rkey.mv_size = node->mn_ksize;
7452 xdata.mv_data = NODEDATA(node);
7453 xdata.mv_size = NODEDSZ(node);
7456 pgno = NODEPGNO(node);
7457 flags = node->mn_flags;
7462 if (!IS_LEAF(mp) && j == 0) {
7463 /* First branch index doesn't need key data. */
7467 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7471 nkeys = NUMKEYS(copy);
7472 for (i=0; i<nkeys; i++)
7473 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7474 mp->mp_lower = copy->mp_lower;
7475 mp->mp_upper = copy->mp_upper;
7476 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7477 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7479 /* reset back to original page */
7480 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7481 mc->mc_pg[mc->mc_top] = mp;
7482 if (nflags & MDB_RESERVE) {
7483 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7484 if (!(node->mn_flags & F_BIGDATA))
7485 newdata->mv_data = NODEDATA(node);
7489 /* Make sure mc_ki is still valid.
7491 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7492 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7493 for (i=0; i<ptop; i++) {
7494 mc->mc_pg[i] = mn.mc_pg[i];
7495 mc->mc_ki[i] = mn.mc_ki[i];
7497 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7498 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7502 /* return tmp page to freelist */
7503 mdb_page_free(mc->mc_txn->mt_env, copy);
7506 /* Adjust other cursors pointing to mp */
7507 MDB_cursor *m2, *m3;
7508 MDB_dbi dbi = mc->mc_dbi;
7509 int fixup = NUMKEYS(mp);
7511 if (mc->mc_flags & C_SUB)
7514 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7515 if (mc->mc_flags & C_SUB)
7516 m3 = &m2->mc_xcursor->mx_cursor;
7521 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7523 if (m3->mc_flags & C_SPLITTING)
7528 for (k=m3->mc_top; k>=0; k--) {
7529 m3->mc_ki[k+1] = m3->mc_ki[k];
7530 m3->mc_pg[k+1] = m3->mc_pg[k];
7532 if (m3->mc_ki[0] >= split_indx) {
7537 m3->mc_pg[0] = mc->mc_pg[0];
7541 if (m3->mc_pg[mc->mc_top] == mp) {
7542 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7543 m3->mc_ki[mc->mc_top]++;
7544 if (m3->mc_ki[mc->mc_top] >= fixup) {
7545 m3->mc_pg[mc->mc_top] = rp;
7546 m3->mc_ki[mc->mc_top] -= fixup;
7547 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7549 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7550 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7559 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7560 MDB_val *key, MDB_val *data, unsigned int flags)
7565 assert(key != NULL);
7566 assert(data != NULL);
7568 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7571 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7575 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7579 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7582 mdb_cursor_init(&mc, txn, dbi, &mx);
7583 return mdb_cursor_put(&mc, key, data, flags);
7587 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7589 if ((flag & CHANGEABLE) != flag)
7592 env->me_flags |= flag;
7594 env->me_flags &= ~flag;
7599 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7604 *arg = env->me_flags;
7609 mdb_env_get_path(MDB_env *env, const char **arg)
7614 *arg = env->me_path;
7618 /** Common code for #mdb_stat() and #mdb_env_stat().
7619 * @param[in] env the environment to operate in.
7620 * @param[in] db the #MDB_db record containing the stats to return.
7621 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7622 * @return 0, this function always succeeds.
7625 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7627 arg->ms_psize = env->me_psize;
7628 arg->ms_depth = db->md_depth;
7629 arg->ms_branch_pages = db->md_branch_pages;
7630 arg->ms_leaf_pages = db->md_leaf_pages;
7631 arg->ms_overflow_pages = db->md_overflow_pages;
7632 arg->ms_entries = db->md_entries;
7637 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7641 if (env == NULL || arg == NULL)
7644 toggle = mdb_env_pick_meta(env);
7646 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7650 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7654 if (env == NULL || arg == NULL)
7657 toggle = mdb_env_pick_meta(env);
7658 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7659 arg->me_mapsize = env->me_mapsize;
7660 arg->me_maxreaders = env->me_maxreaders;
7661 arg->me_numreaders = env->me_numreaders;
7662 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7663 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7667 /** Set the default comparison functions for a database.
7668 * Called immediately after a database is opened to set the defaults.
7669 * The user can then override them with #mdb_set_compare() or
7670 * #mdb_set_dupsort().
7671 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7672 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7675 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7677 uint16_t f = txn->mt_dbs[dbi].md_flags;
7679 txn->mt_dbxs[dbi].md_cmp =
7680 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7681 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7683 txn->mt_dbxs[dbi].md_dcmp =
7684 !(f & MDB_DUPSORT) ? 0 :
7685 ((f & MDB_INTEGERDUP)
7686 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7687 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7690 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7695 int rc, dbflag, exact;
7696 unsigned int unused = 0;
7699 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7700 mdb_default_cmp(txn, FREE_DBI);
7703 if ((flags & VALID_FLAGS) != flags)
7709 if (flags & PERSISTENT_FLAGS) {
7710 uint16_t f2 = flags & PERSISTENT_FLAGS;
7711 /* make sure flag changes get committed */
7712 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7713 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7714 txn->mt_flags |= MDB_TXN_DIRTY;
7717 mdb_default_cmp(txn, MAIN_DBI);
7721 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7722 mdb_default_cmp(txn, MAIN_DBI);
7725 /* Is the DB already open? */
7727 for (i=2; i<txn->mt_numdbs; i++) {
7728 if (!txn->mt_dbxs[i].md_name.mv_size) {
7729 /* Remember this free slot */
7730 if (!unused) unused = i;
7733 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7734 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7740 /* If no free slot and max hit, fail */
7741 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7742 return MDB_DBS_FULL;
7744 /* Cannot mix named databases with some mainDB flags */
7745 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7746 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7748 /* Find the DB info */
7749 dbflag = DB_NEW|DB_VALID;
7752 key.mv_data = (void *)name;
7753 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7754 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7755 if (rc == MDB_SUCCESS) {
7756 /* make sure this is actually a DB */
7757 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7758 if (!(node->mn_flags & F_SUBDATA))
7760 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7761 /* Create if requested */
7763 data.mv_size = sizeof(MDB_db);
7764 data.mv_data = &dummy;
7765 memset(&dummy, 0, sizeof(dummy));
7766 dummy.md_root = P_INVALID;
7767 dummy.md_flags = flags & PERSISTENT_FLAGS;
7768 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7772 /* OK, got info, add to table */
7773 if (rc == MDB_SUCCESS) {
7774 unsigned int slot = unused ? unused : txn->mt_numdbs;
7775 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7776 txn->mt_dbxs[slot].md_name.mv_size = len;
7777 txn->mt_dbxs[slot].md_rel = NULL;
7778 txn->mt_dbflags[slot] = dbflag;
7779 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7781 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7782 mdb_default_cmp(txn, slot);
7791 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7793 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7796 if (txn->mt_dbflags[dbi] & DB_STALE) {
7799 /* Stale, must read the DB's root. cursor_init does it for us. */
7800 mdb_cursor_init(&mc, txn, dbi, &mx);
7802 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7805 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7808 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7810 ptr = env->me_dbxs[dbi].md_name.mv_data;
7811 env->me_dbxs[dbi].md_name.mv_data = NULL;
7812 env->me_dbxs[dbi].md_name.mv_size = 0;
7813 env->me_dbflags[dbi] = 0;
7817 int mdb_dbi_flags(MDB_env *env, MDB_dbi dbi, unsigned int *flags)
7819 /* We could return the flags for the FREE_DBI too but what's the point? */
7820 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7822 *flags = env->me_dbflags[dbi];
7826 /** Add all the DB's pages to the free list.
7827 * @param[in] mc Cursor on the DB to free.
7828 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7829 * @return 0 on success, non-zero on failure.
7832 mdb_drop0(MDB_cursor *mc, int subs)
7836 rc = mdb_page_search(mc, NULL, 0);
7837 if (rc == MDB_SUCCESS) {
7838 MDB_txn *txn = mc->mc_txn;
7843 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7844 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7847 mdb_cursor_copy(mc, &mx);
7848 while (mc->mc_snum > 0) {
7849 MDB_page *mp = mc->mc_pg[mc->mc_top];
7850 unsigned n = NUMKEYS(mp);
7852 for (i=0; i<n; i++) {
7853 ni = NODEPTR(mp, i);
7854 if (ni->mn_flags & F_BIGDATA) {
7857 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7858 rc = mdb_page_get(txn, pg, &omp, NULL);
7861 assert(IS_OVERFLOW(omp));
7862 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7866 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7867 mdb_xcursor_init1(mc, ni);
7868 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7874 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7876 for (i=0; i<n; i++) {
7878 ni = NODEPTR(mp, i);
7881 mdb_midl_xappend(txn->mt_free_pgs, pg);
7886 mc->mc_ki[mc->mc_top] = i;
7887 rc = mdb_cursor_sibling(mc, 1);
7889 /* no more siblings, go back to beginning
7890 * of previous level.
7894 for (i=1; i<mc->mc_snum; i++) {
7896 mc->mc_pg[i] = mx.mc_pg[i];
7901 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7902 } else if (rc == MDB_NOTFOUND) {
7908 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7910 MDB_cursor *mc, *m2;
7913 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7916 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7919 rc = mdb_cursor_open(txn, dbi, &mc);
7923 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7924 /* Invalidate the dropped DB's cursors */
7925 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7926 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7930 /* Can't delete the main DB */
7931 if (del && dbi > MAIN_DBI) {
7932 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7934 txn->mt_dbflags[dbi] = DB_STALE;
7935 mdb_dbi_close(txn->mt_env, dbi);
7938 /* reset the DB record, mark it dirty */
7939 txn->mt_dbflags[dbi] |= DB_DIRTY;
7940 txn->mt_dbs[dbi].md_depth = 0;
7941 txn->mt_dbs[dbi].md_branch_pages = 0;
7942 txn->mt_dbs[dbi].md_leaf_pages = 0;
7943 txn->mt_dbs[dbi].md_overflow_pages = 0;
7944 txn->mt_dbs[dbi].md_entries = 0;
7945 txn->mt_dbs[dbi].md_root = P_INVALID;
7947 txn->mt_flags |= MDB_TXN_DIRTY;
7950 mdb_cursor_close(mc);
7954 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7956 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7959 txn->mt_dbxs[dbi].md_cmp = cmp;
7963 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7965 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7968 txn->mt_dbxs[dbi].md_dcmp = cmp;
7972 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7974 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7977 txn->mt_dbxs[dbi].md_rel = rel;
7981 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7983 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7986 txn->mt_dbxs[dbi].md_relctx = ctx;
7990 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
7992 unsigned int i, rdrs;
7999 if (!env->me_txns) {
8000 return func("(no reader locks)\n", ctx);
8002 rdrs = env->me_txns->mti_numreaders;
8003 mr = env->me_txns->mti_readers;
8004 for (i=0; i<rdrs; i++) {
8009 if (mr[i].mr_txnid == (txnid_t)-1) {
8010 sprintf(buf, "%10d %zx -\n", mr[i].mr_pid, tid);
8012 sprintf(buf, "%10d %zx %zu\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8016 func(" pid thread txnid\n", ctx);
8018 rc = func(buf, ctx);
8024 func("(no active readers)\n", ctx);
8029 /* insert pid into list if not already present.
8030 * return -1 if already present.
8032 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8034 /* binary search of pid in list */
8036 unsigned cursor = 1;
8038 unsigned n = ids[0];
8041 unsigned pivot = n >> 1;
8042 cursor = base + pivot + 1;
8043 val = pid - ids[cursor];
8048 } else if ( val > 0 ) {
8053 /* found, so it's a duplicate */
8062 for (n = ids[0]; n > cursor; n--)
8068 int mdb_reader_check(MDB_env *env, int *dead)
8070 unsigned int i, j, rdrs;
8081 rdrs = env->me_txns->mti_numreaders;
8082 pids = malloc((rdrs+1) * sizeof(pid_t));
8086 mr = env->me_txns->mti_readers;
8088 for (i=0; i<rdrs; i++) {
8089 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8091 if (mdb_pid_insert(pids, pid) == 0) {
8092 if (mdb_reader_pid(env, Pidcheck, pid)) {
8094 for (j=i; j<rdrs; j++)
8095 if (mr[j].mr_pid == pid) {
8099 UNLOCK_MUTEX_R(env);