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();
2081 if (!(env->me_flags & MDB_LIVE_READER)) {
2082 rc = mdb_reader_pid(env, Pidset, pid);
2084 UNLOCK_MUTEX_R(env);
2087 env->me_flags |= MDB_LIVE_READER;
2091 for (i=0; i<env->me_txns->mti_numreaders; i++)
2092 if (env->me_txns->mti_readers[i].mr_pid == 0)
2094 if (i == env->me_maxreaders) {
2095 UNLOCK_MUTEX_R(env);
2096 return MDB_READERS_FULL;
2098 env->me_txns->mti_readers[i].mr_pid = pid;
2099 env->me_txns->mti_readers[i].mr_tid = tid;
2100 if (i >= env->me_txns->mti_numreaders)
2101 env->me_txns->mti_numreaders = i+1;
2102 /* Save numreaders for un-mutexed mdb_env_close() */
2103 env->me_numreaders = env->me_txns->mti_numreaders;
2104 UNLOCK_MUTEX_R(env);
2105 r = &env->me_txns->mti_readers[i];
2106 new_notls = (env->me_flags & MDB_NOTLS);
2107 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2112 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2113 txn->mt_u.reader = r;
2115 txn->mt_toggle = txn->mt_txnid & 1;
2119 txn->mt_txnid = env->me_txns->mti_txnid;
2120 txn->mt_toggle = txn->mt_txnid & 1;
2123 if (txn->mt_txnid == mdb_debug_start)
2126 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2127 txn->mt_u.dirty_list = env->me_dirty_list;
2128 txn->mt_u.dirty_list[0].mid = 0;
2129 txn->mt_free_pgs = env->me_free_pgs;
2130 txn->mt_free_pgs[0] = 0;
2131 txn->mt_spill_pgs = NULL;
2135 /* Copy the DB info and flags */
2136 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2138 /* Moved to here to avoid a data race in read TXNs */
2139 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2141 for (i=2; i<txn->mt_numdbs; i++) {
2142 x = env->me_dbflags[i];
2143 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2144 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2146 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2148 if (env->me_maxpg < txn->mt_next_pgno) {
2149 mdb_txn_reset0(txn, "renew0-mapfail");
2151 txn->mt_u.reader->mr_pid = 0;
2152 txn->mt_u.reader = NULL;
2154 return MDB_MAP_RESIZED;
2161 mdb_txn_renew(MDB_txn *txn)
2165 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2168 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2169 DPUTS("environment had fatal error, must shutdown!");
2173 rc = mdb_txn_renew0(txn);
2174 if (rc == MDB_SUCCESS) {
2175 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
2176 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2177 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2183 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2187 int rc, size, tsize = sizeof(MDB_txn);
2189 if (env->me_flags & MDB_FATAL_ERROR) {
2190 DPUTS("environment had fatal error, must shutdown!");
2193 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2196 /* Nested transactions: Max 1 child, write txns only, no writemap */
2197 if (parent->mt_child ||
2198 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
2199 (env->me_flags & MDB_WRITEMAP))
2203 tsize = sizeof(MDB_ntxn);
2205 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2206 if (!(flags & MDB_RDONLY))
2207 size += env->me_maxdbs * sizeof(MDB_cursor *);
2209 if ((txn = calloc(1, size)) == NULL) {
2210 DPRINTF("calloc: %s", strerror(ErrCode()));
2213 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2214 if (flags & MDB_RDONLY) {
2215 txn->mt_flags |= MDB_TXN_RDONLY;
2216 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2218 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2219 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2225 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2226 if (!txn->mt_u.dirty_list ||
2227 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2229 free(txn->mt_u.dirty_list);
2233 txn->mt_txnid = parent->mt_txnid;
2234 txn->mt_toggle = parent->mt_toggle;
2235 txn->mt_dirty_room = parent->mt_dirty_room;
2236 txn->mt_u.dirty_list[0].mid = 0;
2237 txn->mt_spill_pgs = NULL;
2238 txn->mt_next_pgno = parent->mt_next_pgno;
2239 parent->mt_child = txn;
2240 txn->mt_parent = parent;
2241 txn->mt_numdbs = parent->mt_numdbs;
2242 txn->mt_flags = parent->mt_flags;
2243 txn->mt_dbxs = parent->mt_dbxs;
2244 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2245 /* Copy parent's mt_dbflags, but clear DB_NEW */
2246 for (i=0; i<txn->mt_numdbs; i++)
2247 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2249 ntxn = (MDB_ntxn *)txn;
2250 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2251 if (env->me_pghead) {
2252 size = MDB_IDL_SIZEOF(env->me_pghead);
2253 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2255 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2260 rc = mdb_cursor_shadow(parent, txn);
2262 mdb_txn_reset0(txn, "beginchild-fail");
2264 rc = mdb_txn_renew0(txn);
2270 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2271 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2272 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2278 /** Export or close DBI handles opened in this txn. */
2280 mdb_dbis_update(MDB_txn *txn, int keep)
2283 MDB_dbi n = txn->mt_numdbs;
2284 MDB_env *env = txn->mt_env;
2285 unsigned char *tdbflags = txn->mt_dbflags;
2287 for (i = n; --i >= 2;) {
2288 if (tdbflags[i] & DB_NEW) {
2290 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2292 char *ptr = env->me_dbxs[i].md_name.mv_data;
2293 env->me_dbxs[i].md_name.mv_data = NULL;
2294 env->me_dbxs[i].md_name.mv_size = 0;
2295 env->me_dbflags[i] = 0;
2300 if (keep && env->me_numdbs < n)
2304 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2305 * May be called twice for readonly txns: First reset it, then abort.
2306 * @param[in] txn the transaction handle to reset
2309 mdb_txn_reset0(MDB_txn *txn, const char *act)
2311 MDB_env *env = txn->mt_env;
2313 /* Close any DBI handles opened in this txn */
2314 mdb_dbis_update(txn, 0);
2316 DPRINTF("%s txn %zu%c %p on mdbenv %p, root page %zu",
2317 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2318 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2320 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2321 if (txn->mt_u.reader) {
2322 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2323 if (!(env->me_flags & MDB_NOTLS))
2324 txn->mt_u.reader = NULL; /* txn does not own reader */
2326 txn->mt_numdbs = 0; /* close nothing if called again */
2327 txn->mt_dbxs = NULL; /* mark txn as reset */
2329 mdb_cursors_close(txn, 0);
2331 if (!(env->me_flags & MDB_WRITEMAP)) {
2332 mdb_dlist_free(txn);
2334 mdb_midl_free(env->me_pghead);
2336 if (txn->mt_parent) {
2337 txn->mt_parent->mt_child = NULL;
2338 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2339 mdb_midl_free(txn->mt_free_pgs);
2340 mdb_midl_free(txn->mt_spill_pgs);
2341 free(txn->mt_u.dirty_list);
2345 if (mdb_midl_shrink(&txn->mt_free_pgs))
2346 env->me_free_pgs = txn->mt_free_pgs;
2347 env->me_pghead = NULL;
2351 /* The writer mutex was locked in mdb_txn_begin. */
2352 UNLOCK_MUTEX_W(env);
2357 mdb_txn_reset(MDB_txn *txn)
2362 /* This call is only valid for read-only txns */
2363 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2366 mdb_txn_reset0(txn, "reset");
2370 mdb_txn_abort(MDB_txn *txn)
2376 mdb_txn_abort(txn->mt_child);
2378 mdb_txn_reset0(txn, "abort");
2379 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2380 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2381 txn->mt_u.reader->mr_pid = 0;
2386 /** Save the freelist as of this transaction to the freeDB.
2387 * This changes the freelist. Keep trying until it stabilizes.
2390 mdb_freelist_save(MDB_txn *txn)
2392 /* env->me_pghead[] can grow and shrink during this call.
2393 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2394 * Page numbers cannot disappear from txn->mt_free_pgs[].
2397 MDB_env *env = txn->mt_env;
2398 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2399 txnid_t pglast = 0, head_id = 0;
2400 pgno_t freecnt = 0, *free_pgs, *mop;
2401 ssize_t head_room = 0, total_room = 0, mop_len;
2403 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2405 if (env->me_pghead) {
2406 /* Make sure first page of freeDB is touched and on freelist */
2407 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2408 if (rc && rc != MDB_NOTFOUND)
2413 /* Come back here after each Put() in case freelist changed */
2416 /* If using records from freeDB which we have not yet
2417 * deleted, delete them and any we reserved for me_pghead.
2419 while (pglast < env->me_pglast) {
2420 rc = mdb_cursor_first(&mc, &key, NULL);
2423 pglast = head_id = *(txnid_t *)key.mv_data;
2424 total_room = head_room = 0;
2425 assert(pglast <= env->me_pglast);
2426 rc = mdb_cursor_del(&mc, 0);
2431 /* Save the IDL of pages freed by this txn, to a single record */
2432 if (freecnt < txn->mt_free_pgs[0]) {
2434 /* Make sure last page of freeDB is touched and on freelist */
2435 key.mv_size = MDB_MAXKEYSIZE+1;
2437 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2438 if (rc && rc != MDB_NOTFOUND)
2441 free_pgs = txn->mt_free_pgs;
2442 /* Write to last page of freeDB */
2443 key.mv_size = sizeof(txn->mt_txnid);
2444 key.mv_data = &txn->mt_txnid;
2446 freecnt = free_pgs[0];
2447 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2448 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2451 /* Retry if mt_free_pgs[] grew during the Put() */
2452 free_pgs = txn->mt_free_pgs;
2453 } while (freecnt < free_pgs[0]);
2454 mdb_midl_sort(free_pgs);
2455 memcpy(data.mv_data, free_pgs, data.mv_size);
2458 unsigned int i = free_pgs[0];
2459 DPRINTF("IDL write txn %zu root %zu num %u",
2460 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2462 DPRINTF("IDL %zu", free_pgs[i]);
2468 mop = env->me_pghead;
2469 mop_len = mop ? mop[0] : 0;
2471 /* Reserve records for me_pghead[]. Split it if multi-page,
2472 * to avoid searching freeDB for a page range. Use keys in
2473 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2475 if (total_room >= mop_len) {
2476 if (total_room == mop_len || --more < 0)
2478 } else if (head_room >= maxfree_1pg && head_id > 1) {
2479 /* Keep current record (overflow page), add a new one */
2483 /* (Re)write {key = head_id, IDL length = head_room} */
2484 total_room -= head_room;
2485 head_room = mop_len - total_room;
2486 if (head_room > maxfree_1pg && head_id > 1) {
2487 /* Overflow multi-page for part of me_pghead */
2488 head_room /= head_id; /* amortize page sizes */
2489 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2490 } else if (head_room < 0) {
2491 /* Rare case, not bothering to delete this record */
2494 key.mv_size = sizeof(head_id);
2495 key.mv_data = &head_id;
2496 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2497 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2500 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2501 total_room += head_room;
2504 /* Fill in the reserved, touched me_pghead records */
2510 rc = mdb_cursor_first(&mc, &key, &data);
2511 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2512 unsigned flags = MDB_CURRENT;
2513 txnid_t id = *(txnid_t *)key.mv_data;
2514 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2517 assert(len >= 0 && id <= env->me_pglast);
2519 if (len > mop_len) {
2521 data.mv_size = (len + 1) * sizeof(MDB_ID);
2524 data.mv_data = mop -= len;
2527 rc = mdb_cursor_put(&mc, &key, &data, flags);
2529 if (rc || !(mop_len -= len))
2536 /** Flush dirty pages to the map, after clearing their dirty flag.
2539 mdb_page_flush(MDB_txn *txn)
2541 MDB_env *env = txn->mt_env;
2542 MDB_ID2L dl = txn->mt_u.dirty_list;
2543 unsigned psize = env->me_psize, j;
2544 int i, pagecount = dl[0].mid, rc;
2545 size_t size = 0, pos = 0;
2547 MDB_page *dp = NULL;
2551 struct iovec iov[MDB_COMMIT_PAGES];
2552 ssize_t wpos = 0, wsize = 0, wres;
2553 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2558 if (env->me_flags & MDB_WRITEMAP) {
2559 /* Clear dirty flags */
2560 for (i = pagecount; i; i--) {
2562 /* Don't flush this page yet */
2563 if (dp->mp_flags & P_KEEP) {
2564 dp->mp_flags ^= P_KEEP;
2568 dp->mp_flags &= ~P_DIRTY;
2574 /* Write the pages */
2576 if (i <= pagecount) {
2578 /* Don't flush this page yet */
2579 if (dp->mp_flags & P_KEEP) {
2580 dp->mp_flags ^= P_KEEP;
2585 /* clear dirty flag */
2586 dp->mp_flags &= ~P_DIRTY;
2589 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2594 /* Windows actually supports scatter/gather I/O, but only on
2595 * unbuffered file handles. Since we're relying on the OS page
2596 * cache for all our data, that's self-defeating. So we just
2597 * write pages one at a time. We use the ov structure to set
2598 * the write offset, to at least save the overhead of a Seek
2601 DPRINTF("committing page %zu", pgno);
2602 memset(&ov, 0, sizeof(ov));
2603 ov.Offset = pos & 0xffffffff;
2604 ov.OffsetHigh = pos >> 16 >> 16;
2605 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2607 DPRINTF("WriteFile: %d", rc);
2611 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2612 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2614 /* Write previous page(s) */
2615 #ifdef MDB_USE_PWRITEV
2616 wres = pwritev(env->me_fd, iov, n, wpos);
2619 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2621 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2623 DPRINTF("lseek: %s", strerror(rc));
2626 wres = writev(env->me_fd, iov, n);
2629 if (wres != wsize) {
2632 DPRINTF("Write error: %s", strerror(rc));
2634 rc = EIO; /* TODO: Use which error code? */
2635 DPUTS("short write, filesystem full?");
2646 DPRINTF("committing page %zu", pgno);
2647 next_pos = pos + size;
2648 iov[n].iov_len = size;
2649 iov[n].iov_base = (char *)dp;
2656 for (i=1; i<=pagecount; i++) {
2658 /* This is a page we skipped above */
2661 dl[j].mid = dp->mp_pgno;
2664 mdb_dpage_free(env, dp);
2672 mdb_txn_commit(MDB_txn *txn)
2678 assert(txn != NULL);
2679 assert(txn->mt_env != NULL);
2681 if (txn->mt_child) {
2682 rc = mdb_txn_commit(txn->mt_child);
2683 txn->mt_child = NULL;
2690 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2691 mdb_dbis_update(txn, 1);
2692 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2697 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2698 DPUTS("error flag is set, can't commit");
2700 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2705 if (txn->mt_parent) {
2706 MDB_txn *parent = txn->mt_parent;
2710 /* Append our free list to parent's */
2711 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2714 mdb_midl_free(txn->mt_free_pgs);
2716 parent->mt_next_pgno = txn->mt_next_pgno;
2717 parent->mt_flags = txn->mt_flags;
2719 /* Merge our cursors into parent's and close them */
2720 mdb_cursors_close(txn, 1);
2722 /* Update parent's DB table. */
2723 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2724 parent->mt_numdbs = txn->mt_numdbs;
2725 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2726 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2727 for (i=2; i<txn->mt_numdbs; i++) {
2728 /* preserve parent's DB_NEW status */
2729 x = parent->mt_dbflags[i] & DB_NEW;
2730 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2733 dst = parent->mt_u.dirty_list;
2734 src = txn->mt_u.dirty_list;
2735 /* Remove anything in our dirty list from parent's spill list */
2736 if (parent->mt_spill_pgs) {
2737 x = parent->mt_spill_pgs[0];
2739 /* zero out our dirty pages in parent spill list */
2740 for (i=1; i<=src[0].mid; i++) {
2741 if (src[i].mid < parent->mt_spill_pgs[x])
2743 if (src[i].mid > parent->mt_spill_pgs[x]) {
2749 parent->mt_spill_pgs[x] = 0;
2752 /* OK, we had a few hits, squash zeros from the spill list */
2753 if (len < parent->mt_spill_pgs[0]) {
2755 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2756 if (parent->mt_spill_pgs[y]) {
2758 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2763 parent->mt_spill_pgs[0] = len;
2766 /* Find len = length of merging our dirty list with parent's */
2768 dst[0].mid = 0; /* simplify loops */
2769 if (parent->mt_parent) {
2770 len = x + src[0].mid;
2771 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2772 for (i = x; y && i; y--) {
2773 pgno_t yp = src[y].mid;
2774 while (yp < dst[i].mid)
2776 if (yp == dst[i].mid) {
2781 } else { /* Simplify the above for single-ancestor case */
2782 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2784 /* Merge our dirty list with parent's */
2786 for (i = len; y; dst[i--] = src[y--]) {
2787 pgno_t yp = src[y].mid;
2788 while (yp < dst[x].mid)
2789 dst[i--] = dst[x--];
2790 if (yp == dst[x].mid)
2791 free(dst[x--].mptr);
2795 free(txn->mt_u.dirty_list);
2796 parent->mt_dirty_room = txn->mt_dirty_room;
2797 if (txn->mt_spill_pgs) {
2798 if (parent->mt_spill_pgs) {
2799 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2800 mdb_midl_free(txn->mt_spill_pgs);
2801 mdb_midl_sort(parent->mt_spill_pgs);
2803 parent->mt_spill_pgs = txn->mt_spill_pgs;
2807 parent->mt_child = NULL;
2808 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2813 if (txn != env->me_txn) {
2814 DPUTS("attempt to commit unknown transaction");
2819 mdb_cursors_close(txn, 0);
2821 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2824 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2825 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2827 /* Update DB root pointers */
2828 if (txn->mt_numdbs > 2) {
2832 data.mv_size = sizeof(MDB_db);
2834 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2835 for (i = 2; i < txn->mt_numdbs; i++) {
2836 if (txn->mt_dbflags[i] & DB_DIRTY) {
2837 data.mv_data = &txn->mt_dbs[i];
2838 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2845 rc = mdb_freelist_save(txn);
2849 mdb_midl_free(env->me_pghead);
2850 env->me_pghead = NULL;
2851 if (mdb_midl_shrink(&txn->mt_free_pgs))
2852 env->me_free_pgs = txn->mt_free_pgs;
2858 if ((rc = mdb_page_flush(txn)) ||
2859 (rc = mdb_env_sync(env, 0)) ||
2860 (rc = mdb_env_write_meta(txn)))
2866 mdb_dbis_update(txn, 1);
2868 UNLOCK_MUTEX_W(env);
2878 /** Read the environment parameters of a DB environment before
2879 * mapping it into memory.
2880 * @param[in] env the environment handle
2881 * @param[out] meta address of where to store the meta information
2882 * @return 0 on success, non-zero on failure.
2885 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2892 /* We don't know the page size yet, so use a minimum value.
2893 * Read both meta pages so we can use the latest one.
2896 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2900 memset(&ov, 0, sizeof(ov));
2902 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2903 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2906 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2908 if (rc != MDB_PAGESIZE) {
2909 if (rc == 0 && off == 0)
2911 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2912 DPRINTF("read: %s", mdb_strerror(rc));
2916 p = (MDB_page *)&pbuf;
2918 if (!F_ISSET(p->mp_flags, P_META)) {
2919 DPRINTF("page %zu not a meta page", p->mp_pgno);
2924 if (m->mm_magic != MDB_MAGIC) {
2925 DPUTS("meta has invalid magic");
2929 if (m->mm_version != MDB_DATA_VERSION) {
2930 DPRINTF("database is version %u, expected version %u",
2931 m->mm_version, MDB_DATA_VERSION);
2932 return MDB_VERSION_MISMATCH;
2935 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2941 /** Write the environment parameters of a freshly created DB environment.
2942 * @param[in] env the environment handle
2943 * @param[out] meta address of where to store the meta information
2944 * @return 0 on success, non-zero on failure.
2947 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2953 DPUTS("writing new meta page");
2955 GET_PAGESIZE(psize);
2957 meta->mm_magic = MDB_MAGIC;
2958 meta->mm_version = MDB_DATA_VERSION;
2959 meta->mm_mapsize = env->me_mapsize;
2960 meta->mm_psize = psize;
2961 meta->mm_last_pg = 1;
2962 meta->mm_flags = env->me_flags & 0xffff;
2963 meta->mm_flags |= MDB_INTEGERKEY;
2964 meta->mm_dbs[0].md_root = P_INVALID;
2965 meta->mm_dbs[1].md_root = P_INVALID;
2967 p = calloc(2, psize);
2969 p->mp_flags = P_META;
2970 *(MDB_meta *)METADATA(p) = *meta;
2972 q = (MDB_page *)((char *)p + psize);
2974 q->mp_flags = P_META;
2975 *(MDB_meta *)METADATA(q) = *meta;
2981 memset(&ov, 0, sizeof(ov));
2982 rc = WriteFile(env->me_fd, p, psize * 2, &len, &ov);
2983 rc = (len == psize * 2 ? MDB_SUCCESS : ErrCode();
2986 rc = pwrite(env->me_fd, p, psize * 2, 0);
2987 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2993 /** Update the environment info to commit a transaction.
2994 * @param[in] txn the transaction that's being committed
2995 * @return 0 on success, non-zero on failure.
2998 mdb_env_write_meta(MDB_txn *txn)
3001 MDB_meta meta, metab, *mp;
3003 int rc, len, toggle;
3012 assert(txn != NULL);
3013 assert(txn->mt_env != NULL);
3015 toggle = !txn->mt_toggle;
3016 DPRINTF("writing meta page %d for root page %zu",
3017 toggle, txn->mt_dbs[MAIN_DBI].md_root);
3020 mp = env->me_metas[toggle];
3022 if (env->me_flags & MDB_WRITEMAP) {
3023 /* Persist any increases of mapsize config */
3024 if (env->me_mapsize > mp->mm_mapsize)
3025 mp->mm_mapsize = env->me_mapsize;
3026 mp->mm_dbs[0] = txn->mt_dbs[0];
3027 mp->mm_dbs[1] = txn->mt_dbs[1];
3028 mp->mm_last_pg = txn->mt_next_pgno - 1;
3029 mp->mm_txnid = txn->mt_txnid;
3030 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3031 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3034 ptr += env->me_psize;
3035 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3042 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3043 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3045 ptr = (char *)&meta;
3046 if (env->me_mapsize > mp->mm_mapsize) {
3047 /* Persist any increases of mapsize config */
3048 meta.mm_mapsize = env->me_mapsize;
3049 off = offsetof(MDB_meta, mm_mapsize);
3051 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3053 len = sizeof(MDB_meta) - off;
3056 meta.mm_dbs[0] = txn->mt_dbs[0];
3057 meta.mm_dbs[1] = txn->mt_dbs[1];
3058 meta.mm_last_pg = txn->mt_next_pgno - 1;
3059 meta.mm_txnid = txn->mt_txnid;
3062 off += env->me_psize;
3065 /* Write to the SYNC fd */
3066 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3067 env->me_fd : env->me_mfd;
3070 memset(&ov, 0, sizeof(ov));
3072 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3076 rc = pwrite(mfd, ptr, len, off);
3079 rc = rc < 0 ? ErrCode() : EIO;
3080 DPUTS("write failed, disk error?");
3081 /* On a failure, the pagecache still contains the new data.
3082 * Write some old data back, to prevent it from being used.
3083 * Use the non-SYNC fd; we know it will fail anyway.
3085 meta.mm_last_pg = metab.mm_last_pg;
3086 meta.mm_txnid = metab.mm_txnid;
3088 memset(&ov, 0, sizeof(ov));
3090 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3092 r2 = pwrite(env->me_fd, ptr, len, off);
3095 env->me_flags |= MDB_FATAL_ERROR;
3099 /* Memory ordering issues are irrelevant; since the entire writer
3100 * is wrapped by wmutex, all of these changes will become visible
3101 * after the wmutex is unlocked. Since the DB is multi-version,
3102 * readers will get consistent data regardless of how fresh or
3103 * how stale their view of these values is.
3105 env->me_txns->mti_txnid = txn->mt_txnid;
3110 /** Check both meta pages to see which one is newer.
3111 * @param[in] env the environment handle
3112 * @return meta toggle (0 or 1).
3115 mdb_env_pick_meta(const MDB_env *env)
3117 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3121 mdb_env_create(MDB_env **env)
3125 e = calloc(1, sizeof(MDB_env));
3129 e->me_maxreaders = DEFAULT_READERS;
3130 e->me_maxdbs = e->me_numdbs = 2;
3131 e->me_fd = INVALID_HANDLE_VALUE;
3132 e->me_lfd = INVALID_HANDLE_VALUE;
3133 e->me_mfd = INVALID_HANDLE_VALUE;
3134 #ifdef MDB_USE_POSIX_SEM
3135 e->me_rmutex = SEM_FAILED;
3136 e->me_wmutex = SEM_FAILED;
3138 e->me_pid = getpid();
3139 VGMEMP_CREATE(e,0,0);
3145 mdb_env_set_mapsize(MDB_env *env, size_t size)
3149 env->me_mapsize = size;
3151 env->me_maxpg = env->me_mapsize / env->me_psize;
3156 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3160 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3165 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3167 if (env->me_map || readers < 1)
3169 env->me_maxreaders = readers;
3174 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3176 if (!env || !readers)
3178 *readers = env->me_maxreaders;
3182 /** Further setup required for opening an MDB environment
3185 mdb_env_open2(MDB_env *env)
3187 unsigned int flags = env->me_flags;
3195 memset(&meta, 0, sizeof(meta));
3197 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3200 DPUTS("new mdbenv");
3204 /* Was a mapsize configured? */
3205 if (!env->me_mapsize) {
3206 /* If this is a new environment, take the default,
3207 * else use the size recorded in the existing env.
3209 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3210 } else if (env->me_mapsize < meta.mm_mapsize) {
3211 /* If the configured size is smaller, make sure it's
3212 * still big enough. Silently round up to minimum if not.
3214 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3215 if (env->me_mapsize < minsize)
3216 env->me_mapsize = minsize;
3223 LONG sizelo, sizehi;
3224 sizelo = env->me_mapsize & 0xffffffff;
3225 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3227 /* See if we should use QueryLimited */
3229 if ((rc & 0xff) > 5)
3230 env->me_pidquery = PROCESS_QUERY_LIMITED_INFORMATION;
3232 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3234 /* Windows won't create mappings for zero length files.
3235 * Just allocate the maxsize right now.
3238 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3239 || !SetEndOfFile(env->me_fd)
3240 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3243 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3244 PAGE_READWRITE : PAGE_READONLY,
3245 sizehi, sizelo, NULL);
3248 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3249 FILE_MAP_WRITE : FILE_MAP_READ,
3250 0, 0, env->me_mapsize, meta.mm_address);
3251 rc = env->me_map ? 0 : ErrCode();
3259 if (flags & MDB_WRITEMAP) {
3261 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3264 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3266 if (env->me_map == MAP_FAILED) {
3270 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3272 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3274 #ifdef POSIX_MADV_RANDOM
3275 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3276 #endif /* POSIX_MADV_RANDOM */
3277 #endif /* MADV_RANDOM */
3281 if (flags & MDB_FIXEDMAP)
3282 meta.mm_address = env->me_map;
3283 i = mdb_env_init_meta(env, &meta);
3284 if (i != MDB_SUCCESS) {
3287 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3288 /* Can happen because the address argument to mmap() is just a
3289 * hint. mmap() can pick another, e.g. if the range is in use.
3290 * The MAP_FIXED flag would prevent that, but then mmap could
3291 * instead unmap existing pages to make room for the new map.
3293 return EBUSY; /* TODO: Make a new MDB_* error code? */
3295 env->me_psize = meta.mm_psize;
3296 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3297 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3299 env->me_maxpg = env->me_mapsize / env->me_psize;
3301 p = (MDB_page *)env->me_map;
3302 env->me_metas[0] = METADATA(p);
3303 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3307 int toggle = mdb_env_pick_meta(env);
3308 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3310 DPRINTF("opened database version %u, pagesize %u",
3311 env->me_metas[0]->mm_version, env->me_psize);
3312 DPRINTF("using meta page %d", toggle);
3313 DPRINTF("depth: %u", db->md_depth);
3314 DPRINTF("entries: %zu", db->md_entries);
3315 DPRINTF("branch pages: %zu", db->md_branch_pages);
3316 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
3317 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
3318 DPRINTF("root: %zu", db->md_root);
3326 /** Release a reader thread's slot in the reader lock table.
3327 * This function is called automatically when a thread exits.
3328 * @param[in] ptr This points to the slot in the reader lock table.
3331 mdb_env_reader_dest(void *ptr)
3333 MDB_reader *reader = ptr;
3339 /** Junk for arranging thread-specific callbacks on Windows. This is
3340 * necessarily platform and compiler-specific. Windows supports up
3341 * to 1088 keys. Let's assume nobody opens more than 64 environments
3342 * in a single process, for now. They can override this if needed.
3344 #ifndef MAX_TLS_KEYS
3345 #define MAX_TLS_KEYS 64
3347 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3348 static int mdb_tls_nkeys;
3350 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3354 case DLL_PROCESS_ATTACH: break;
3355 case DLL_THREAD_ATTACH: break;
3356 case DLL_THREAD_DETACH:
3357 for (i=0; i<mdb_tls_nkeys; i++) {
3358 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3359 mdb_env_reader_dest(r);
3362 case DLL_PROCESS_DETACH: break;
3367 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3369 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3373 /* Force some symbol references.
3374 * _tls_used forces the linker to create the TLS directory if not already done
3375 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3377 #pragma comment(linker, "/INCLUDE:_tls_used")
3378 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3379 #pragma const_seg(".CRT$XLB")
3380 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3381 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3384 #pragma comment(linker, "/INCLUDE:__tls_used")
3385 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3386 #pragma data_seg(".CRT$XLB")
3387 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3389 #endif /* WIN 32/64 */
3390 #endif /* !__GNUC__ */
3393 /** Downgrade the exclusive lock on the region back to shared */
3395 mdb_env_share_locks(MDB_env *env, int *excl)
3397 int rc = 0, toggle = mdb_env_pick_meta(env);
3399 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3404 /* First acquire a shared lock. The Unlock will
3405 * then release the existing exclusive lock.
3407 memset(&ov, 0, sizeof(ov));
3408 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3411 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3417 struct flock lock_info;
3418 /* The shared lock replaces the existing lock */
3419 memset((void *)&lock_info, 0, sizeof(lock_info));
3420 lock_info.l_type = F_RDLCK;
3421 lock_info.l_whence = SEEK_SET;
3422 lock_info.l_start = 0;
3423 lock_info.l_len = 1;
3424 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3425 (rc = ErrCode()) == EINTR) ;
3426 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3433 /** Try to get exlusive lock, otherwise shared.
3434 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3437 mdb_env_excl_lock(MDB_env *env, int *excl)
3441 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3445 memset(&ov, 0, sizeof(ov));
3446 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3453 struct flock lock_info;
3454 memset((void *)&lock_info, 0, sizeof(lock_info));
3455 lock_info.l_type = F_WRLCK;
3456 lock_info.l_whence = SEEK_SET;
3457 lock_info.l_start = 0;
3458 lock_info.l_len = 1;
3459 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3460 (rc = ErrCode()) == EINTR) ;
3464 # ifdef MDB_USE_POSIX_SEM
3465 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3468 lock_info.l_type = F_RDLCK;
3469 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3470 (rc = ErrCode()) == EINTR) ;
3478 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3480 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3482 * @(#) $Revision: 5.1 $
3483 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3484 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3486 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3490 * Please do not copyright this code. This code is in the public domain.
3492 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3493 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3494 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3495 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3496 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3497 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3498 * PERFORMANCE OF THIS SOFTWARE.
3501 * chongo <Landon Curt Noll> /\oo/\
3502 * http://www.isthe.com/chongo/
3504 * Share and Enjoy! :-)
3507 typedef unsigned long long mdb_hash_t;
3508 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3510 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3511 * @param[in] str string to hash
3512 * @param[in] hval initial value for hash
3513 * @return 64 bit hash
3515 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3516 * hval arg on the first call.
3519 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3521 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3522 unsigned char *end = s + val->mv_size;
3524 * FNV-1a hash each octet of the string
3527 /* xor the bottom with the current octet */
3528 hval ^= (mdb_hash_t)*s++;
3530 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3531 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3532 (hval << 7) + (hval << 8) + (hval << 40);
3534 /* return our new hash value */
3538 /** Hash the string and output the encoded hash.
3539 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3540 * very short name limits. We don't care about the encoding being reversible,
3541 * we just want to preserve as many bits of the input as possible in a
3542 * small printable string.
3543 * @param[in] str string to hash
3544 * @param[out] encbuf an array of 11 chars to hold the hash
3546 const static char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3549 mdb_pack85(unsigned long l, char *out)
3553 for (i=0; i<5; i++) {
3554 *out++ = mdb_a85[l % 85];
3560 mdb_hash_enc(MDB_val *val, char *encbuf)
3562 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3563 unsigned long *l = (unsigned long *)&h;
3565 mdb_pack85(l[0], encbuf);
3566 mdb_pack85(l[1], encbuf+5);
3571 /** Open and/or initialize the lock region for the environment.
3572 * @param[in] env The MDB environment.
3573 * @param[in] lpath The pathname of the file used for the lock region.
3574 * @param[in] mode The Unix permissions for the file, if we create it.
3575 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3576 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3577 * @return 0 on success, non-zero on failure.
3580 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3583 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3585 # define MDB_ERRCODE_ROFS EROFS
3586 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3587 # define MDB_CLOEXEC O_CLOEXEC
3590 # define MDB_CLOEXEC 0
3597 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3598 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3599 FILE_ATTRIBUTE_NORMAL, NULL);
3601 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3603 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3605 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3610 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3611 /* Lose record locks when exec*() */
3612 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3613 fcntl(env->me_lfd, F_SETFD, fdflags);
3616 if (!(env->me_flags & MDB_NOTLS)) {
3617 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3620 env->me_flags |= MDB_ENV_TXKEY;
3622 /* Windows TLS callbacks need help finding their TLS info. */
3623 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3627 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3631 /* Try to get exclusive lock. If we succeed, then
3632 * nobody is using the lock region and we should initialize it.
3634 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3637 size = GetFileSize(env->me_lfd, NULL);
3639 size = lseek(env->me_lfd, 0, SEEK_END);
3640 if (size == -1) goto fail_errno;
3642 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3643 if (size < rsize && *excl > 0) {
3645 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3646 || !SetEndOfFile(env->me_lfd))
3649 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3653 size = rsize - sizeof(MDB_txninfo);
3654 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3659 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3661 if (!mh) goto fail_errno;
3662 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3664 if (!env->me_txns) goto fail_errno;
3666 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3668 if (m == MAP_FAILED) goto fail_errno;
3674 BY_HANDLE_FILE_INFORMATION stbuf;
3683 if (!mdb_sec_inited) {
3684 InitializeSecurityDescriptor(&mdb_null_sd,
3685 SECURITY_DESCRIPTOR_REVISION);
3686 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3687 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3688 mdb_all_sa.bInheritHandle = FALSE;
3689 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3692 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3693 idbuf.volume = stbuf.dwVolumeSerialNumber;
3694 idbuf.nhigh = stbuf.nFileIndexHigh;
3695 idbuf.nlow = stbuf.nFileIndexLow;
3696 val.mv_data = &idbuf;
3697 val.mv_size = sizeof(idbuf);
3698 mdb_hash_enc(&val, encbuf);
3699 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3700 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3701 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3702 if (!env->me_rmutex) goto fail_errno;
3703 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3704 if (!env->me_wmutex) goto fail_errno;
3705 #elif defined(MDB_USE_POSIX_SEM)
3714 #if defined(__NetBSD__)
3715 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3717 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3718 idbuf.dev = stbuf.st_dev;
3719 idbuf.ino = stbuf.st_ino;
3720 val.mv_data = &idbuf;
3721 val.mv_size = sizeof(idbuf);
3722 mdb_hash_enc(&val, encbuf);
3723 #ifdef MDB_SHORT_SEMNAMES
3724 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3726 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3727 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3728 /* Clean up after a previous run, if needed: Try to
3729 * remove both semaphores before doing anything else.
3731 sem_unlink(env->me_txns->mti_rmname);
3732 sem_unlink(env->me_txns->mti_wmname);
3733 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3734 O_CREAT|O_EXCL, mode, 1);
3735 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3736 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3737 O_CREAT|O_EXCL, mode, 1);
3738 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3739 #else /* MDB_USE_POSIX_SEM */
3740 pthread_mutexattr_t mattr;
3742 if ((rc = pthread_mutexattr_init(&mattr))
3743 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3744 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3745 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3747 pthread_mutexattr_destroy(&mattr);
3748 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3750 env->me_txns->mti_version = MDB_LOCK_VERSION;
3751 env->me_txns->mti_magic = MDB_MAGIC;
3752 env->me_txns->mti_txnid = 0;
3753 env->me_txns->mti_numreaders = 0;
3756 if (env->me_txns->mti_magic != MDB_MAGIC) {
3757 DPUTS("lock region has invalid magic");
3761 if (env->me_txns->mti_version != MDB_LOCK_VERSION) {
3762 DPRINTF("lock region is version %u, expected version %u",
3763 env->me_txns->mti_version, MDB_LOCK_VERSION);
3764 rc = MDB_VERSION_MISMATCH;
3768 if (rc && rc != EACCES && rc != EAGAIN) {
3772 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3773 if (!env->me_rmutex) goto fail_errno;
3774 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3775 if (!env->me_wmutex) goto fail_errno;
3776 #elif defined(MDB_USE_POSIX_SEM)
3777 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3778 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3779 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3780 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3791 /** The name of the lock file in the DB environment */
3792 #define LOCKNAME "/lock.mdb"
3793 /** The name of the data file in the DB environment */
3794 #define DATANAME "/data.mdb"
3795 /** The suffix of the lock file when no subdir is used */
3796 #define LOCKSUFF "-lock"
3797 /** Only a subset of the @ref mdb_env flags can be changed
3798 * at runtime. Changing other flags requires closing the
3799 * environment and re-opening it with the new flags.
3801 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3802 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3805 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3807 int oflags, rc, len, excl = -1;
3808 char *lpath, *dpath;
3810 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3814 if (flags & MDB_NOSUBDIR) {
3815 rc = len + sizeof(LOCKSUFF) + len + 1;
3817 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3822 if (flags & MDB_NOSUBDIR) {
3823 dpath = lpath + len + sizeof(LOCKSUFF);
3824 sprintf(lpath, "%s" LOCKSUFF, path);
3825 strcpy(dpath, path);
3827 dpath = lpath + len + sizeof(LOCKNAME);
3828 sprintf(lpath, "%s" LOCKNAME, path);
3829 sprintf(dpath, "%s" DATANAME, path);
3833 flags |= env->me_flags;
3834 if (flags & MDB_RDONLY) {
3835 /* silently ignore WRITEMAP when we're only getting read access */
3836 flags &= ~MDB_WRITEMAP;
3838 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3839 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3842 env->me_flags = flags |= MDB_ENV_ACTIVE;
3846 env->me_path = strdup(path);
3847 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3848 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3849 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3854 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3859 if (F_ISSET(flags, MDB_RDONLY)) {
3860 oflags = GENERIC_READ;
3861 len = OPEN_EXISTING;
3863 oflags = GENERIC_READ|GENERIC_WRITE;
3866 mode = FILE_ATTRIBUTE_NORMAL;
3867 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3868 NULL, len, mode, NULL);
3870 if (F_ISSET(flags, MDB_RDONLY))
3873 oflags = O_RDWR | O_CREAT;
3875 env->me_fd = open(dpath, oflags, mode);
3877 if (env->me_fd == INVALID_HANDLE_VALUE) {
3882 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3883 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3884 env->me_mfd = env->me_fd;
3886 /* Synchronous fd for meta writes. Needed even with
3887 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3890 env->me_mfd = CreateFile(dpath, oflags,
3891 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3892 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3894 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3896 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3901 DPRINTF("opened dbenv %p", (void *) env);
3903 rc = mdb_env_share_locks(env, &excl);
3909 mdb_env_close0(env, excl);
3915 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3917 mdb_env_close0(MDB_env *env, int excl)
3921 if (!(env->me_flags & MDB_ENV_ACTIVE))
3924 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3925 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3926 free(env->me_dbxs[i].md_name.mv_data);
3928 free(env->me_dbflags);
3931 free(env->me_dirty_list);
3932 mdb_midl_free(env->me_free_pgs);
3934 if (env->me_flags & MDB_ENV_TXKEY) {
3935 pthread_key_delete(env->me_txkey);
3937 /* Delete our key from the global list */
3938 for (i=0; i<mdb_tls_nkeys; i++)
3939 if (mdb_tls_keys[i] == env->me_txkey) {
3940 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3948 munmap(env->me_map, env->me_mapsize);
3950 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3951 (void) close(env->me_mfd);
3952 if (env->me_fd != INVALID_HANDLE_VALUE)
3953 (void) close(env->me_fd);
3955 pid_t pid = env->me_pid;
3956 /* Clearing readers is done in this function because
3957 * me_txkey with its destructor must be disabled first.
3959 for (i = env->me_numreaders; --i >= 0; )
3960 if (env->me_txns->mti_readers[i].mr_pid == pid)
3961 env->me_txns->mti_readers[i].mr_pid = 0;
3963 if (env->me_rmutex) {
3964 CloseHandle(env->me_rmutex);
3965 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3967 /* Windows automatically destroys the mutexes when
3968 * the last handle closes.
3970 #elif defined(MDB_USE_POSIX_SEM)
3971 if (env->me_rmutex != SEM_FAILED) {
3972 sem_close(env->me_rmutex);
3973 if (env->me_wmutex != SEM_FAILED)
3974 sem_close(env->me_wmutex);
3975 /* If we have the filelock: If we are the
3976 * only remaining user, clean up semaphores.
3979 mdb_env_excl_lock(env, &excl);
3981 sem_unlink(env->me_txns->mti_rmname);
3982 sem_unlink(env->me_txns->mti_wmname);
3986 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3988 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3991 /* Unlock the lockfile. Windows would have unlocked it
3992 * after closing anyway, but not necessarily at once.
3994 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3997 (void) close(env->me_lfd);
4000 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4004 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4006 MDB_txn *txn = NULL;
4012 #define DOWRITE(fd, ptr, w2, len) WriteFile(fd, ptr, w2, &len)
4016 #define DOWRITE(fd, ptr, w2, len) len = write(fd, ptr, w2)
4019 /* Do the lock/unlock of the reader mutex before starting the
4020 * write txn. Otherwise other read txns could block writers.
4022 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4027 /* We must start the actual read txn after blocking writers */
4028 mdb_txn_reset0(txn, "reset-stage1");
4030 /* Temporarily block writers until we snapshot the meta pages */
4033 rc = mdb_txn_renew0(txn);
4035 UNLOCK_MUTEX_W(env);
4040 wsize = env->me_psize * 2;
4045 DOWRITE(fd, ptr, w2, len);
4052 /* Non-blocking or async handles are not supported */
4061 UNLOCK_MUTEX_W(env);
4066 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4068 if (wsize > MAX_WRITE)
4072 DOWRITE(fd, ptr, w2, len);
4092 mdb_env_copy(MDB_env *env, const char *path)
4096 HANDLE newfd = INVALID_HANDLE_VALUE;
4098 if (env->me_flags & MDB_NOSUBDIR) {
4099 lpath = (char *)path;
4102 len += sizeof(DATANAME);
4103 lpath = malloc(len);
4106 sprintf(lpath, "%s" DATANAME, path);
4109 /* The destination path must exist, but the destination file must not.
4110 * We don't want the OS to cache the writes, since the source data is
4111 * already in the OS cache.
4114 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4115 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4117 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4123 if (newfd == INVALID_HANDLE_VALUE) {
4128 #ifdef F_NOCACHE /* __APPLE__ */
4129 rc = fcntl(newfd, F_NOCACHE, 1);
4136 rc = mdb_env_copyfd(env, newfd);
4139 if (!(env->me_flags & MDB_NOSUBDIR))
4141 if (newfd != INVALID_HANDLE_VALUE)
4142 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4149 mdb_env_close(MDB_env *env)
4156 VGMEMP_DESTROY(env);
4157 while ((dp = env->me_dpages) != NULL) {
4158 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4159 env->me_dpages = dp->mp_next;
4163 mdb_env_close0(env, 0);
4167 /** Compare two items pointing at aligned size_t's */
4169 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4171 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4172 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4175 /** Compare two items pointing at aligned int's */
4177 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4179 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4180 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4183 /** Compare two items pointing at ints of unknown alignment.
4184 * Nodes and keys are guaranteed to be 2-byte aligned.
4187 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4189 #if BYTE_ORDER == LITTLE_ENDIAN
4190 unsigned short *u, *c;
4193 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4194 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4197 } while(!x && u > (unsigned short *)a->mv_data);
4200 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4204 /** Compare two items lexically */
4206 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4213 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4219 diff = memcmp(a->mv_data, b->mv_data, len);
4220 return diff ? diff : len_diff<0 ? -1 : len_diff;
4223 /** Compare two items in reverse byte order */
4225 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4227 const unsigned char *p1, *p2, *p1_lim;
4231 p1_lim = (const unsigned char *)a->mv_data;
4232 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4233 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4235 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4241 while (p1 > p1_lim) {
4242 diff = *--p1 - *--p2;
4246 return len_diff<0 ? -1 : len_diff;
4249 /** Search for key within a page, using binary search.
4250 * Returns the smallest entry larger or equal to the key.
4251 * If exactp is non-null, stores whether the found entry was an exact match
4252 * in *exactp (1 or 0).
4253 * Updates the cursor index with the index of the found entry.
4254 * If no entry larger or equal to the key is found, returns NULL.
4257 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4259 unsigned int i = 0, nkeys;
4262 MDB_page *mp = mc->mc_pg[mc->mc_top];
4263 MDB_node *node = NULL;
4268 nkeys = NUMKEYS(mp);
4273 COPY_PGNO(pgno, mp->mp_pgno);
4274 DPRINTF("searching %u keys in %s %spage %zu",
4275 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4282 low = IS_LEAF(mp) ? 0 : 1;
4284 cmp = mc->mc_dbx->md_cmp;
4286 /* Branch pages have no data, so if using integer keys,
4287 * alignment is guaranteed. Use faster mdb_cmp_int.
4289 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4290 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4297 nodekey.mv_size = mc->mc_db->md_pad;
4298 node = NODEPTR(mp, 0); /* fake */
4299 while (low <= high) {
4300 i = (low + high) >> 1;
4301 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4302 rc = cmp(key, &nodekey);
4303 DPRINTF("found leaf index %u [%s], rc = %i",
4304 i, DKEY(&nodekey), rc);
4313 while (low <= high) {
4314 i = (low + high) >> 1;
4316 node = NODEPTR(mp, i);
4317 nodekey.mv_size = NODEKSZ(node);
4318 nodekey.mv_data = NODEKEY(node);
4320 rc = cmp(key, &nodekey);
4323 DPRINTF("found leaf index %u [%s], rc = %i",
4324 i, DKEY(&nodekey), rc);
4326 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
4327 i, DKEY(&nodekey), NODEPGNO(node), rc);
4338 if (rc > 0) { /* Found entry is less than the key. */
4339 i++; /* Skip to get the smallest entry larger than key. */
4341 node = NODEPTR(mp, i);
4344 *exactp = (rc == 0);
4345 /* store the key index */
4346 mc->mc_ki[mc->mc_top] = i;
4348 /* There is no entry larger or equal to the key. */
4351 /* nodeptr is fake for LEAF2 */
4357 mdb_cursor_adjust(MDB_cursor *mc, func)
4361 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4362 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4369 /** Pop a page off the top of the cursor's stack. */
4371 mdb_cursor_pop(MDB_cursor *mc)
4374 #ifndef MDB_DEBUG_SKIP
4375 MDB_page *top = mc->mc_pg[mc->mc_top];
4381 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4382 mc->mc_dbi, (void *) mc);
4386 /** Push a page onto the top of the cursor's stack. */
4388 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4390 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4391 mc->mc_dbi, (void *) mc);
4393 if (mc->mc_snum >= CURSOR_STACK) {
4394 assert(mc->mc_snum < CURSOR_STACK);
4395 return MDB_CURSOR_FULL;
4398 mc->mc_top = mc->mc_snum++;
4399 mc->mc_pg[mc->mc_top] = mp;
4400 mc->mc_ki[mc->mc_top] = 0;
4405 /** Find the address of the page corresponding to a given page number.
4406 * @param[in] txn the transaction for this access.
4407 * @param[in] pgno the page number for the page to retrieve.
4408 * @param[out] ret address of a pointer where the page's address will be stored.
4409 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4410 * @return 0 on success, non-zero on failure.
4413 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4418 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4419 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4424 MDB_ID2L dl = tx2->mt_u.dirty_list;
4426 /* Spilled pages were dirtied in this txn and flushed
4427 * because the dirty list got full. Bring this page
4428 * back in from the map (but don't unspill it here,
4429 * leave that unless page_touch happens again).
4431 if (tx2->mt_spill_pgs) {
4432 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4433 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4434 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4439 unsigned x = mdb_mid2l_search(dl, pgno);
4440 if (x <= dl[0].mid && dl[x].mid == pgno) {
4446 } while ((tx2 = tx2->mt_parent) != NULL);
4449 if (pgno < txn->mt_next_pgno) {
4451 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4453 DPRINTF("page %zu not found", pgno);
4455 return MDB_PAGE_NOTFOUND;
4465 /** Search for the page a given key should be in.
4466 * Pushes parent pages on the cursor stack. This function continues a
4467 * search on a cursor that has already been initialized. (Usually by
4468 * #mdb_page_search() but also by #mdb_node_move().)
4469 * @param[in,out] mc the cursor for this operation.
4470 * @param[in] key the key to search for. If NULL, search for the lowest
4471 * page. (This is used by #mdb_cursor_first().)
4472 * @param[in] modify If true, visited pages are updated with new page numbers.
4473 * @return 0 on success, non-zero on failure.
4476 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4478 MDB_page *mp = mc->mc_pg[mc->mc_top];
4483 while (IS_BRANCH(mp)) {
4487 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4488 assert(NUMKEYS(mp) > 1);
4489 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4491 if (key == NULL) /* Initialize cursor to first page. */
4493 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4494 /* cursor to last page */
4498 node = mdb_node_search(mc, key, &exact);
4500 i = NUMKEYS(mp) - 1;
4502 i = mc->mc_ki[mc->mc_top];
4511 DPRINTF("following index %u for key [%s]",
4513 assert(i < NUMKEYS(mp));
4514 node = NODEPTR(mp, i);
4516 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4519 mc->mc_ki[mc->mc_top] = i;
4520 if ((rc = mdb_cursor_push(mc, mp)))
4524 if ((rc = mdb_page_touch(mc)) != 0)
4526 mp = mc->mc_pg[mc->mc_top];
4531 DPRINTF("internal error, index points to a %02X page!?",
4533 return MDB_CORRUPTED;
4536 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4537 key ? DKEY(key) : NULL);
4538 mc->mc_flags |= C_INITIALIZED;
4539 mc->mc_flags &= ~C_EOF;
4544 /** Search for the lowest key under the current branch page.
4545 * This just bypasses a NUMKEYS check in the current page
4546 * before calling mdb_page_search_root(), because the callers
4547 * are all in situations where the current page is known to
4551 mdb_page_search_lowest(MDB_cursor *mc)
4553 MDB_page *mp = mc->mc_pg[mc->mc_top];
4554 MDB_node *node = NODEPTR(mp, 0);
4557 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4560 mc->mc_ki[mc->mc_top] = 0;
4561 if ((rc = mdb_cursor_push(mc, mp)))
4563 return mdb_page_search_root(mc, NULL, 0);
4566 /** Search for the page a given key should be in.
4567 * Pushes parent pages on the cursor stack. This function just sets up
4568 * the search; it finds the root page for \b mc's database and sets this
4569 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4570 * called to complete the search.
4571 * @param[in,out] mc the cursor for this operation.
4572 * @param[in] key the key to search for. If NULL, search for the lowest
4573 * page. (This is used by #mdb_cursor_first().)
4574 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4575 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4576 * @return 0 on success, non-zero on failure.
4579 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4584 /* Make sure the txn is still viable, then find the root from
4585 * the txn's db table.
4587 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4588 DPUTS("transaction has failed, must abort");
4591 /* Make sure we're using an up-to-date root */
4592 if (mc->mc_dbi > MAIN_DBI) {
4593 if ((*mc->mc_dbflag & DB_STALE) ||
4594 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4596 unsigned char dbflag = 0;
4597 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4598 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4601 if (*mc->mc_dbflag & DB_STALE) {
4605 MDB_node *leaf = mdb_node_search(&mc2,
4606 &mc->mc_dbx->md_name, &exact);
4608 return MDB_NOTFOUND;
4609 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4612 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4614 /* The txn may not know this DBI, or another process may
4615 * have dropped and recreated the DB with other flags.
4617 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4618 return MDB_INCOMPATIBLE;
4619 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4621 if (flags & MDB_PS_MODIFY)
4623 *mc->mc_dbflag &= ~DB_STALE;
4624 *mc->mc_dbflag |= dbflag;
4627 root = mc->mc_db->md_root;
4629 if (root == P_INVALID) { /* Tree is empty. */
4630 DPUTS("tree is empty");
4631 return MDB_NOTFOUND;
4636 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4637 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4643 DPRINTF("db %u root page %zu has flags 0x%X",
4644 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4646 if (flags & MDB_PS_MODIFY) {
4647 if ((rc = mdb_page_touch(mc)))
4651 if (flags & MDB_PS_ROOTONLY)
4654 return mdb_page_search_root(mc, key, flags);
4658 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4660 MDB_txn *txn = mc->mc_txn;
4661 pgno_t pg = mp->mp_pgno;
4662 unsigned i, ovpages = mp->mp_pages;
4663 MDB_env *env = txn->mt_env;
4666 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4667 /* If the page is dirty or on the spill list we just acquired it,
4668 * so we should give it back to our current free list, if any.
4669 * Not currently supported in nested txns.
4670 * Otherwise put it onto the list of pages we freed in this txn.
4672 if (!(mp->mp_flags & P_DIRTY) && txn->mt_spill_pgs) {
4673 unsigned x = mdb_midl_search(txn->mt_spill_pgs, pg);
4674 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pg) {
4675 /* This page is no longer spilled */
4676 for (; x < txn->mt_spill_pgs[0]; x++)
4677 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
4678 txn->mt_spill_pgs[0]--;
4682 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && env->me_pghead) {
4685 MDB_ID2 *dl, ix, iy;
4686 rc = mdb_midl_need(&env->me_pghead, ovpages);
4689 /* Remove from dirty list */
4690 dl = txn->mt_u.dirty_list;
4692 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4700 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4701 txn->mt_flags |= MDB_TXN_ERROR;
4702 return MDB_CORRUPTED;
4705 if (!(env->me_flags & MDB_WRITEMAP))
4706 mdb_dpage_free(env, mp);
4708 /* Insert in me_pghead */
4709 mop = env->me_pghead;
4710 j = mop[0] + ovpages;
4711 for (i = mop[0]; i && mop[i] < pg; i--)
4717 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4721 mc->mc_db->md_overflow_pages -= ovpages;
4725 /** Return the data associated with a given node.
4726 * @param[in] txn The transaction for this operation.
4727 * @param[in] leaf The node being read.
4728 * @param[out] data Updated to point to the node's data.
4729 * @return 0 on success, non-zero on failure.
4732 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4734 MDB_page *omp; /* overflow page */
4738 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4739 data->mv_size = NODEDSZ(leaf);
4740 data->mv_data = NODEDATA(leaf);
4744 /* Read overflow data.
4746 data->mv_size = NODEDSZ(leaf);
4747 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4748 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4749 DPRINTF("read overflow page %zu failed", pgno);
4752 data->mv_data = METADATA(omp);
4758 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4759 MDB_val *key, MDB_val *data)
4768 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4770 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4773 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4777 mdb_cursor_init(&mc, txn, dbi, &mx);
4778 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4781 /** Find a sibling for a page.
4782 * Replaces the page at the top of the cursor's stack with the
4783 * specified sibling, if one exists.
4784 * @param[in] mc The cursor for this operation.
4785 * @param[in] move_right Non-zero if the right sibling is requested,
4786 * otherwise the left sibling.
4787 * @return 0 on success, non-zero on failure.
4790 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4796 if (mc->mc_snum < 2) {
4797 return MDB_NOTFOUND; /* root has no siblings */
4801 DPRINTF("parent page is page %zu, index %u",
4802 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4804 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4805 : (mc->mc_ki[mc->mc_top] == 0)) {
4806 DPRINTF("no more keys left, moving to %s sibling",
4807 move_right ? "right" : "left");
4808 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4809 /* undo cursor_pop before returning */
4816 mc->mc_ki[mc->mc_top]++;
4818 mc->mc_ki[mc->mc_top]--;
4819 DPRINTF("just moving to %s index key %u",
4820 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4822 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4824 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4825 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4828 mdb_cursor_push(mc, mp);
4830 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4835 /** Move the cursor to the next data item. */
4837 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4843 if (mc->mc_flags & C_EOF) {
4844 return MDB_NOTFOUND;
4847 assert(mc->mc_flags & C_INITIALIZED);
4849 mp = mc->mc_pg[mc->mc_top];
4851 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4852 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4853 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4854 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4855 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4856 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4860 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4861 if (op == MDB_NEXT_DUP)
4862 return MDB_NOTFOUND;
4866 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4868 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4869 DPUTS("=====> move to next sibling page");
4870 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4871 mc->mc_flags |= C_EOF;
4874 mp = mc->mc_pg[mc->mc_top];
4875 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4877 mc->mc_ki[mc->mc_top]++;
4879 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4880 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4883 key->mv_size = mc->mc_db->md_pad;
4884 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4888 assert(IS_LEAF(mp));
4889 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4891 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4892 mdb_xcursor_init1(mc, leaf);
4895 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4898 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4899 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4900 if (rc != MDB_SUCCESS)
4905 MDB_GET_KEY(leaf, key);
4909 /** Move the cursor to the previous data item. */
4911 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4917 assert(mc->mc_flags & C_INITIALIZED);
4919 mp = mc->mc_pg[mc->mc_top];
4921 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4922 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4923 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4924 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4925 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4926 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4929 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4930 if (op == MDB_PREV_DUP)
4931 return MDB_NOTFOUND;
4936 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4938 if (mc->mc_ki[mc->mc_top] == 0) {
4939 DPUTS("=====> move to prev sibling page");
4940 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4943 mp = mc->mc_pg[mc->mc_top];
4944 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4945 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4947 mc->mc_ki[mc->mc_top]--;
4949 mc->mc_flags &= ~C_EOF;
4951 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4952 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4955 key->mv_size = mc->mc_db->md_pad;
4956 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4960 assert(IS_LEAF(mp));
4961 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4963 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4964 mdb_xcursor_init1(mc, leaf);
4967 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4970 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4971 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4972 if (rc != MDB_SUCCESS)
4977 MDB_GET_KEY(leaf, key);
4981 /** Set the cursor on a specific data item. */
4983 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4984 MDB_cursor_op op, int *exactp)
4988 MDB_node *leaf = NULL;
4993 assert(key->mv_size > 0);
4996 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4998 /* See if we're already on the right page */
4999 if (mc->mc_flags & C_INITIALIZED) {
5002 mp = mc->mc_pg[mc->mc_top];
5004 mc->mc_ki[mc->mc_top] = 0;
5005 return MDB_NOTFOUND;
5007 if (mp->mp_flags & P_LEAF2) {
5008 nodekey.mv_size = mc->mc_db->md_pad;
5009 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5011 leaf = NODEPTR(mp, 0);
5012 MDB_GET_KEY(leaf, &nodekey);
5014 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5016 /* Probably happens rarely, but first node on the page
5017 * was the one we wanted.
5019 mc->mc_ki[mc->mc_top] = 0;
5026 unsigned int nkeys = NUMKEYS(mp);
5028 if (mp->mp_flags & P_LEAF2) {
5029 nodekey.mv_data = LEAF2KEY(mp,
5030 nkeys-1, nodekey.mv_size);
5032 leaf = NODEPTR(mp, nkeys-1);
5033 MDB_GET_KEY(leaf, &nodekey);
5035 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5037 /* last node was the one we wanted */
5038 mc->mc_ki[mc->mc_top] = nkeys-1;
5044 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5045 /* This is definitely the right page, skip search_page */
5046 if (mp->mp_flags & P_LEAF2) {
5047 nodekey.mv_data = LEAF2KEY(mp,
5048 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5050 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5051 MDB_GET_KEY(leaf, &nodekey);
5053 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5055 /* current node was the one we wanted */
5065 /* If any parents have right-sibs, search.
5066 * Otherwise, there's nothing further.
5068 for (i=0; i<mc->mc_top; i++)
5070 NUMKEYS(mc->mc_pg[i])-1)
5072 if (i == mc->mc_top) {
5073 /* There are no other pages */
5074 mc->mc_ki[mc->mc_top] = nkeys;
5075 return MDB_NOTFOUND;
5079 /* There are no other pages */
5080 mc->mc_ki[mc->mc_top] = 0;
5081 return MDB_NOTFOUND;
5085 rc = mdb_page_search(mc, key, 0);
5086 if (rc != MDB_SUCCESS)
5089 mp = mc->mc_pg[mc->mc_top];
5090 assert(IS_LEAF(mp));
5093 leaf = mdb_node_search(mc, key, exactp);
5094 if (exactp != NULL && !*exactp) {
5095 /* MDB_SET specified and not an exact match. */
5096 return MDB_NOTFOUND;
5100 DPUTS("===> inexact leaf not found, goto sibling");
5101 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5102 return rc; /* no entries matched */
5103 mp = mc->mc_pg[mc->mc_top];
5104 assert(IS_LEAF(mp));
5105 leaf = NODEPTR(mp, 0);
5109 mc->mc_flags |= C_INITIALIZED;
5110 mc->mc_flags &= ~C_EOF;
5113 key->mv_size = mc->mc_db->md_pad;
5114 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5118 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5119 mdb_xcursor_init1(mc, leaf);
5122 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5123 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5124 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5127 if (op == MDB_GET_BOTH) {
5133 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5134 if (rc != MDB_SUCCESS)
5137 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5139 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5141 rc = mc->mc_dbx->md_dcmp(data, &d2);
5143 if (op == MDB_GET_BOTH || rc > 0)
5144 return MDB_NOTFOUND;
5149 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5150 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5155 /* The key already matches in all other cases */
5156 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5157 MDB_GET_KEY(leaf, key);
5158 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
5163 /** Move the cursor to the first item in the database. */
5165 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5171 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5173 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5174 rc = mdb_page_search(mc, NULL, 0);
5175 if (rc != MDB_SUCCESS)
5178 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5180 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5181 mc->mc_flags |= C_INITIALIZED;
5182 mc->mc_flags &= ~C_EOF;
5184 mc->mc_ki[mc->mc_top] = 0;
5186 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5187 key->mv_size = mc->mc_db->md_pad;
5188 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5193 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5194 mdb_xcursor_init1(mc, leaf);
5195 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5199 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5203 MDB_GET_KEY(leaf, key);
5207 /** Move the cursor to the last item in the database. */
5209 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5215 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5217 if (!(mc->mc_flags & C_EOF)) {
5219 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5222 lkey.mv_size = MDB_MAXKEYSIZE+1;
5223 lkey.mv_data = NULL;
5224 rc = mdb_page_search(mc, &lkey, 0);
5225 if (rc != MDB_SUCCESS)
5228 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5231 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5232 mc->mc_flags |= C_INITIALIZED|C_EOF;
5233 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5235 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5236 key->mv_size = mc->mc_db->md_pad;
5237 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5242 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5243 mdb_xcursor_init1(mc, leaf);
5244 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5248 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5253 MDB_GET_KEY(leaf, key);
5258 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5267 case MDB_GET_CURRENT:
5268 if (!(mc->mc_flags & C_INITIALIZED)) {
5271 MDB_page *mp = mc->mc_pg[mc->mc_top];
5273 mc->mc_ki[mc->mc_top] = 0;
5279 key->mv_size = mc->mc_db->md_pad;
5280 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5282 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5283 MDB_GET_KEY(leaf, key);
5285 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5286 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5288 rc = mdb_node_read(mc->mc_txn, leaf, data);
5295 case MDB_GET_BOTH_RANGE:
5296 if (data == NULL || mc->mc_xcursor == NULL) {
5304 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5306 } else if (op == MDB_SET_RANGE)
5307 rc = mdb_cursor_set(mc, key, data, op, NULL);
5309 rc = mdb_cursor_set(mc, key, data, op, &exact);
5311 case MDB_GET_MULTIPLE:
5313 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
5314 !(mc->mc_flags & C_INITIALIZED)) {
5319 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5320 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5323 case MDB_NEXT_MULTIPLE:
5325 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5329 if (!(mc->mc_flags & C_INITIALIZED))
5330 rc = mdb_cursor_first(mc, key, data);
5332 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5333 if (rc == MDB_SUCCESS) {
5334 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5337 mx = &mc->mc_xcursor->mx_cursor;
5338 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5340 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5341 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5349 case MDB_NEXT_NODUP:
5350 if (!(mc->mc_flags & C_INITIALIZED))
5351 rc = mdb_cursor_first(mc, key, data);
5353 rc = mdb_cursor_next(mc, key, data, op);
5357 case MDB_PREV_NODUP:
5358 if (!(mc->mc_flags & C_INITIALIZED)) {
5359 rc = mdb_cursor_last(mc, key, data);
5362 mc->mc_flags |= C_INITIALIZED;
5363 mc->mc_ki[mc->mc_top]++;
5365 rc = mdb_cursor_prev(mc, key, data, op);
5368 rc = mdb_cursor_first(mc, key, data);
5372 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
5373 !(mc->mc_flags & C_INITIALIZED) ||
5374 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5378 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5381 rc = mdb_cursor_last(mc, key, data);
5385 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
5386 !(mc->mc_flags & C_INITIALIZED) ||
5387 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5391 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5394 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5402 /** Touch all the pages in the cursor stack.
5403 * Makes sure all the pages are writable, before attempting a write operation.
5404 * @param[in] mc The cursor to operate on.
5407 mdb_cursor_touch(MDB_cursor *mc)
5411 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5414 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5415 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5418 *mc->mc_dbflag |= DB_DIRTY;
5420 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5421 rc = mdb_page_touch(mc);
5425 mc->mc_top = mc->mc_snum-1;
5429 /** Do not spill pages to disk if txn is getting full, may fail instead */
5430 #define MDB_NOSPILL 0x8000
5433 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5436 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5437 MDB_node *leaf = NULL;
5438 MDB_val xdata, *rdata, dkey;
5441 int do_sub = 0, insert = 0;
5442 unsigned int mcount = 0, dcount = 0, nospill;
5446 char dbuf[MDB_MAXKEYSIZE+1];
5447 unsigned int nflags;
5450 /* Check this first so counter will always be zero on any
5453 if (flags & MDB_MULTIPLE) {
5454 dcount = data[1].mv_size;
5455 data[1].mv_size = 0;
5456 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5460 nospill = flags & MDB_NOSPILL;
5461 flags &= ~MDB_NOSPILL;
5463 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5466 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5469 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5472 #if SIZE_MAX > MAXDATASIZE
5473 if (data->mv_size > MAXDATASIZE)
5477 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5478 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5482 if (flags == MDB_CURRENT) {
5483 if (!(mc->mc_flags & C_INITIALIZED))
5486 } else if (mc->mc_db->md_root == P_INVALID) {
5487 /* new database, cursor has nothing to point to */
5489 mc->mc_flags &= ~C_INITIALIZED;
5494 if (flags & MDB_APPEND) {
5496 rc = mdb_cursor_last(mc, &k2, &d2);
5498 rc = mc->mc_dbx->md_cmp(key, &k2);
5501 mc->mc_ki[mc->mc_top]++;
5503 /* new key is <= last key */
5508 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5510 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5511 DPRINTF("duplicate key [%s]", DKEY(key));
5513 return MDB_KEYEXIST;
5515 if (rc && rc != MDB_NOTFOUND)
5519 /* Cursor is positioned, check for room in the dirty list */
5521 if (flags & MDB_MULTIPLE) {
5523 xdata.mv_size = data->mv_size * dcount;
5527 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5531 if (rc == MDB_NO_ROOT) {
5533 /* new database, write a root leaf page */
5534 DPUTS("allocating new root leaf page");
5535 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5538 mdb_cursor_push(mc, np);
5539 mc->mc_db->md_root = np->mp_pgno;
5540 mc->mc_db->md_depth++;
5541 *mc->mc_dbflag |= DB_DIRTY;
5542 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5544 np->mp_flags |= P_LEAF2;
5545 mc->mc_flags |= C_INITIALIZED;
5547 /* make sure all cursor pages are writable */
5548 rc2 = mdb_cursor_touch(mc);
5553 /* The key already exists */
5554 if (rc == MDB_SUCCESS) {
5555 /* there's only a key anyway, so this is a no-op */
5556 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5557 unsigned int ksize = mc->mc_db->md_pad;
5558 if (key->mv_size != ksize)
5560 if (flags == MDB_CURRENT) {
5561 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5562 memcpy(ptr, key->mv_data, ksize);
5567 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5570 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5571 /* Was a single item before, must convert now */
5573 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5574 /* Just overwrite the current item */
5575 if (flags == MDB_CURRENT)
5578 dkey.mv_size = NODEDSZ(leaf);
5579 dkey.mv_data = NODEDATA(leaf);
5580 #if UINT_MAX < SIZE_MAX
5581 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5582 #ifdef MISALIGNED_OK
5583 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5585 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5588 /* if data matches, ignore it */
5589 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5590 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5592 /* create a fake page for the dup items */
5593 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5594 dkey.mv_data = dbuf;
5595 fp = (MDB_page *)&pbuf;
5596 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5597 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5598 fp->mp_lower = PAGEHDRSZ;
5599 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5600 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5601 fp->mp_flags |= P_LEAF2;
5602 fp->mp_pad = data->mv_size;
5603 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5605 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5606 (dkey.mv_size & 1) + (data->mv_size & 1);
5608 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5611 xdata.mv_size = fp->mp_upper;
5616 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5617 /* See if we need to convert from fake page to subDB */
5619 unsigned int offset;
5623 fp = NODEDATA(leaf);
5624 if (flags == MDB_CURRENT) {
5626 fp->mp_flags |= P_DIRTY;
5627 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5628 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5632 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5633 offset = fp->mp_pad;
5634 if (SIZELEFT(fp) >= offset)
5636 offset *= 4; /* space for 4 more */
5638 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5640 offset += offset & 1;
5641 fp_flags = fp->mp_flags;
5642 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5643 offset >= mc->mc_txn->mt_env->me_nodemax) {
5644 /* yes, convert it */
5646 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5647 dummy.md_pad = fp->mp_pad;
5648 dummy.md_flags = MDB_DUPFIXED;
5649 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5650 dummy.md_flags |= MDB_INTEGERKEY;
5653 dummy.md_branch_pages = 0;
5654 dummy.md_leaf_pages = 1;
5655 dummy.md_overflow_pages = 0;
5656 dummy.md_entries = NUMKEYS(fp);
5658 xdata.mv_size = sizeof(MDB_db);
5659 xdata.mv_data = &dummy;
5660 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5662 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5663 flags |= F_DUPDATA|F_SUBDATA;
5664 dummy.md_root = mp->mp_pgno;
5665 fp_flags &= ~P_SUBP;
5667 /* no, just grow it */
5669 xdata.mv_size = NODEDSZ(leaf) + offset;
5670 xdata.mv_data = &pbuf;
5671 mp = (MDB_page *)&pbuf;
5672 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5675 mp->mp_flags = fp_flags | P_DIRTY;
5676 mp->mp_pad = fp->mp_pad;
5677 mp->mp_lower = fp->mp_lower;
5678 mp->mp_upper = fp->mp_upper + offset;
5680 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5682 nsize = NODEDSZ(leaf) - fp->mp_upper;
5683 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5684 for (i=0; i<NUMKEYS(fp); i++)
5685 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5687 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5691 /* data is on sub-DB, just store it */
5692 flags |= F_DUPDATA|F_SUBDATA;
5696 /* overflow page overwrites need special handling */
5697 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5700 unsigned psize = mc->mc_txn->mt_env->me_psize;
5701 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5703 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5704 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5706 ovpages = omp->mp_pages;
5708 /* Is the ov page large enough? */
5709 if (ovpages >= dpages) {
5710 if (!(omp->mp_flags & P_DIRTY) &&
5711 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5713 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5716 level = 0; /* dirty in this txn or clean */
5719 if (omp->mp_flags & P_DIRTY) {
5720 /* yes, overwrite it. Note in this case we don't
5721 * bother to try shrinking the page if the new data
5722 * is smaller than the overflow threshold.
5725 /* It is writable only in a parent txn */
5726 size_t sz = (size_t) psize * ovpages, off;
5727 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5733 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5734 if (!(flags & MDB_RESERVE)) {
5735 /* Copy end of page, adjusting alignment so
5736 * compiler may copy words instead of bytes.
5738 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5739 memcpy((size_t *)((char *)np + off),
5740 (size_t *)((char *)omp + off), sz - off);
5743 memcpy(np, omp, sz); /* Copy beginning of page */
5746 SETDSZ(leaf, data->mv_size);
5747 if (F_ISSET(flags, MDB_RESERVE))
5748 data->mv_data = METADATA(omp);
5750 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5754 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5756 } else if (NODEDSZ(leaf) == data->mv_size) {
5757 /* same size, just replace it. Note that we could
5758 * also reuse this node if the new data is smaller,
5759 * but instead we opt to shrink the node in that case.
5761 if (F_ISSET(flags, MDB_RESERVE))
5762 data->mv_data = NODEDATA(leaf);
5763 else if (data->mv_size)
5764 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5766 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5769 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5770 mc->mc_db->md_entries--;
5772 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5779 nflags = flags & NODE_ADD_FLAGS;
5780 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5781 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5782 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5783 nflags &= ~MDB_APPEND;
5785 nflags |= MDB_SPLIT_REPLACE;
5786 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5788 /* There is room already in this leaf page. */
5789 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5790 if (rc == 0 && !do_sub && insert) {
5791 /* Adjust other cursors pointing to mp */
5792 MDB_cursor *m2, *m3;
5793 MDB_dbi dbi = mc->mc_dbi;
5794 unsigned i = mc->mc_top;
5795 MDB_page *mp = mc->mc_pg[i];
5797 if (mc->mc_flags & C_SUB)
5800 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5801 if (mc->mc_flags & C_SUB)
5802 m3 = &m2->mc_xcursor->mx_cursor;
5805 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5806 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5813 if (rc != MDB_SUCCESS)
5814 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5816 /* Now store the actual data in the child DB. Note that we're
5817 * storing the user data in the keys field, so there are strict
5818 * size limits on dupdata. The actual data fields of the child
5819 * DB are all zero size.
5826 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5827 if (flags & MDB_CURRENT) {
5828 xflags = MDB_CURRENT|MDB_NOSPILL;
5830 mdb_xcursor_init1(mc, leaf);
5831 xflags = (flags & MDB_NODUPDATA) ?
5832 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5834 /* converted, write the original data first */
5836 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5840 /* Adjust other cursors pointing to mp */
5842 unsigned i = mc->mc_top;
5843 MDB_page *mp = mc->mc_pg[i];
5845 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5846 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5847 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5848 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5849 mdb_xcursor_init1(m2, leaf);
5853 /* we've done our job */
5856 if (flags & MDB_APPENDDUP)
5857 xflags |= MDB_APPEND;
5858 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5859 if (flags & F_SUBDATA) {
5860 void *db = NODEDATA(leaf);
5861 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5864 /* sub-writes might have failed so check rc again.
5865 * Don't increment count if we just replaced an existing item.
5867 if (!rc && !(flags & MDB_CURRENT))
5868 mc->mc_db->md_entries++;
5869 if (flags & MDB_MULTIPLE) {
5872 if (mcount < dcount) {
5873 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5874 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5878 /* let caller know how many succeeded, if any */
5879 data[1].mv_size = mcount;
5883 /* If we succeeded and the key didn't exist before, make sure
5884 * the cursor is marked valid.
5887 mc->mc_flags |= C_INITIALIZED;
5892 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5897 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5900 if (!(mc->mc_flags & C_INITIALIZED))
5903 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
5905 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
5907 rc = mdb_cursor_touch(mc);
5911 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5913 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5914 if (flags != MDB_NODUPDATA) {
5915 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5916 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5918 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
5919 /* If sub-DB still has entries, we're done */
5920 if (mc->mc_xcursor->mx_db.md_entries) {
5921 if (leaf->mn_flags & F_SUBDATA) {
5922 /* update subDB info */
5923 void *db = NODEDATA(leaf);
5924 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5927 /* shrink fake page */
5928 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5929 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5930 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5931 /* fix other sub-DB cursors pointed at this fake page */
5932 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5933 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5934 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5935 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5936 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5939 mc->mc_db->md_entries--;
5942 /* otherwise fall thru and delete the sub-DB */
5945 if (leaf->mn_flags & F_SUBDATA) {
5946 /* add all the child DB's pages to the free list */
5947 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5948 if (rc == MDB_SUCCESS) {
5949 mc->mc_db->md_entries -=
5950 mc->mc_xcursor->mx_db.md_entries;
5955 return mdb_cursor_del0(mc, leaf);
5958 /** Allocate and initialize new pages for a database.
5959 * @param[in] mc a cursor on the database being added to.
5960 * @param[in] flags flags defining what type of page is being allocated.
5961 * @param[in] num the number of pages to allocate. This is usually 1,
5962 * unless allocating overflow pages for a large record.
5963 * @param[out] mp Address of a page, or NULL on failure.
5964 * @return 0 on success, non-zero on failure.
5967 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5972 if ((rc = mdb_page_alloc(mc, num, &np)))
5974 DPRINTF("allocated new mpage %zu, page size %u",
5975 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5976 np->mp_flags = flags | P_DIRTY;
5977 np->mp_lower = PAGEHDRSZ;
5978 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5981 mc->mc_db->md_branch_pages++;
5982 else if (IS_LEAF(np))
5983 mc->mc_db->md_leaf_pages++;
5984 else if (IS_OVERFLOW(np)) {
5985 mc->mc_db->md_overflow_pages += num;
5993 /** Calculate the size of a leaf node.
5994 * The size depends on the environment's page size; if a data item
5995 * is too large it will be put onto an overflow page and the node
5996 * size will only include the key and not the data. Sizes are always
5997 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5998 * of the #MDB_node headers.
5999 * @param[in] env The environment handle.
6000 * @param[in] key The key for the node.
6001 * @param[in] data The data for the node.
6002 * @return The number of bytes needed to store the node.
6005 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6009 sz = LEAFSIZE(key, data);
6010 if (sz >= env->me_nodemax) {
6011 /* put on overflow page */
6012 sz -= data->mv_size - sizeof(pgno_t);
6016 return sz + sizeof(indx_t);
6019 /** Calculate the size of a branch node.
6020 * The size should depend on the environment's page size but since
6021 * we currently don't support spilling large keys onto overflow
6022 * pages, it's simply the size of the #MDB_node header plus the
6023 * size of the key. Sizes are always rounded up to an even number
6024 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6025 * @param[in] env The environment handle.
6026 * @param[in] key The key for the node.
6027 * @return The number of bytes needed to store the node.
6030 mdb_branch_size(MDB_env *env, MDB_val *key)
6035 if (sz >= env->me_nodemax) {
6036 /* put on overflow page */
6037 /* not implemented */
6038 /* sz -= key->size - sizeof(pgno_t); */
6041 return sz + sizeof(indx_t);
6044 /** Add a node to the page pointed to by the cursor.
6045 * @param[in] mc The cursor for this operation.
6046 * @param[in] indx The index on the page where the new node should be added.
6047 * @param[in] key The key for the new node.
6048 * @param[in] data The data for the new node, if any.
6049 * @param[in] pgno The page number, if adding a branch node.
6050 * @param[in] flags Flags for the node.
6051 * @return 0 on success, non-zero on failure. Possible errors are:
6053 * <li>ENOMEM - failed to allocate overflow pages for the node.
6054 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6055 * should never happen since all callers already calculate the
6056 * page's free space before calling this function.
6060 mdb_node_add(MDB_cursor *mc, indx_t indx,
6061 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6064 size_t node_size = NODESIZE;
6067 MDB_page *mp = mc->mc_pg[mc->mc_top];
6068 MDB_page *ofp = NULL; /* overflow page */
6071 assert(mp->mp_upper >= mp->mp_lower);
6073 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
6074 IS_LEAF(mp) ? "leaf" : "branch",
6075 IS_SUBP(mp) ? "sub-" : "",
6076 mp->mp_pgno, indx, data ? data->mv_size : 0,
6077 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
6080 /* Move higher keys up one slot. */
6081 int ksize = mc->mc_db->md_pad, dif;
6082 char *ptr = LEAF2KEY(mp, indx, ksize);
6083 dif = NUMKEYS(mp) - indx;
6085 memmove(ptr+ksize, ptr, dif*ksize);
6086 /* insert new key */
6087 memcpy(ptr, key->mv_data, ksize);
6089 /* Just using these for counting */
6090 mp->mp_lower += sizeof(indx_t);
6091 mp->mp_upper -= ksize - sizeof(indx_t);
6096 node_size += key->mv_size;
6100 if (F_ISSET(flags, F_BIGDATA)) {
6101 /* Data already on overflow page. */
6102 node_size += sizeof(pgno_t);
6103 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6104 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6106 /* Put data on overflow page. */
6107 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
6108 data->mv_size, node_size+data->mv_size);
6109 node_size += sizeof(pgno_t);
6110 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6112 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
6115 node_size += data->mv_size;
6118 node_size += node_size & 1;
6120 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6121 DPRINTF("not enough room in page %zu, got %u ptrs",
6122 mp->mp_pgno, NUMKEYS(mp));
6123 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6124 mp->mp_upper - mp->mp_lower);
6125 DPRINTF("node size = %zu", node_size);
6126 return MDB_PAGE_FULL;
6129 /* Move higher pointers up one slot. */
6130 for (i = NUMKEYS(mp); i > indx; i--)
6131 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6133 /* Adjust free space offsets. */
6134 ofs = mp->mp_upper - node_size;
6135 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6136 mp->mp_ptrs[indx] = ofs;
6138 mp->mp_lower += sizeof(indx_t);
6140 /* Write the node data. */
6141 node = NODEPTR(mp, indx);
6142 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6143 node->mn_flags = flags;
6145 SETDSZ(node,data->mv_size);
6150 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6155 if (F_ISSET(flags, F_BIGDATA))
6156 memcpy(node->mn_data + key->mv_size, data->mv_data,
6158 else if (F_ISSET(flags, MDB_RESERVE))
6159 data->mv_data = node->mn_data + key->mv_size;
6161 memcpy(node->mn_data + key->mv_size, data->mv_data,
6164 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6166 if (F_ISSET(flags, MDB_RESERVE))
6167 data->mv_data = METADATA(ofp);
6169 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6176 /** Delete the specified node from a page.
6177 * @param[in] mp The page to operate on.
6178 * @param[in] indx The index of the node to delete.
6179 * @param[in] ksize The size of a node. Only used if the page is
6180 * part of a #MDB_DUPFIXED database.
6183 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6186 indx_t i, j, numkeys, ptr;
6193 COPY_PGNO(pgno, mp->mp_pgno);
6194 DPRINTF("delete node %u on %s page %zu", indx,
6195 IS_LEAF(mp) ? "leaf" : "branch", pgno);
6198 assert(indx < NUMKEYS(mp));
6201 int x = NUMKEYS(mp) - 1 - indx;
6202 base = LEAF2KEY(mp, indx, ksize);
6204 memmove(base, base + ksize, x * ksize);
6205 mp->mp_lower -= sizeof(indx_t);
6206 mp->mp_upper += ksize - sizeof(indx_t);
6210 node = NODEPTR(mp, indx);
6211 sz = NODESIZE + node->mn_ksize;
6213 if (F_ISSET(node->mn_flags, F_BIGDATA))
6214 sz += sizeof(pgno_t);
6216 sz += NODEDSZ(node);
6220 ptr = mp->mp_ptrs[indx];
6221 numkeys = NUMKEYS(mp);
6222 for (i = j = 0; i < numkeys; i++) {
6224 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6225 if (mp->mp_ptrs[i] < ptr)
6226 mp->mp_ptrs[j] += sz;
6231 base = (char *)mp + mp->mp_upper;
6232 memmove(base + sz, base, ptr - mp->mp_upper);
6234 mp->mp_lower -= sizeof(indx_t);
6238 /** Compact the main page after deleting a node on a subpage.
6239 * @param[in] mp The main page to operate on.
6240 * @param[in] indx The index of the subpage on the main page.
6243 mdb_node_shrink(MDB_page *mp, indx_t indx)
6250 indx_t i, numkeys, ptr;
6252 node = NODEPTR(mp, indx);
6253 sp = (MDB_page *)NODEDATA(node);
6254 osize = NODEDSZ(node);
6256 delta = sp->mp_upper - sp->mp_lower;
6257 SETDSZ(node, osize - delta);
6258 xp = (MDB_page *)((char *)sp + delta);
6260 /* shift subpage upward */
6262 nsize = NUMKEYS(sp) * sp->mp_pad;
6263 memmove(METADATA(xp), METADATA(sp), nsize);
6266 nsize = osize - sp->mp_upper;
6267 numkeys = NUMKEYS(sp);
6268 for (i=numkeys-1; i>=0; i--)
6269 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6271 xp->mp_upper = sp->mp_lower;
6272 xp->mp_lower = sp->mp_lower;
6273 xp->mp_flags = sp->mp_flags;
6274 xp->mp_pad = sp->mp_pad;
6275 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6277 /* shift lower nodes upward */
6278 ptr = mp->mp_ptrs[indx];
6279 numkeys = NUMKEYS(mp);
6280 for (i = 0; i < numkeys; i++) {
6281 if (mp->mp_ptrs[i] <= ptr)
6282 mp->mp_ptrs[i] += delta;
6285 base = (char *)mp + mp->mp_upper;
6286 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6287 mp->mp_upper += delta;
6290 /** Initial setup of a sorted-dups cursor.
6291 * Sorted duplicates are implemented as a sub-database for the given key.
6292 * The duplicate data items are actually keys of the sub-database.
6293 * Operations on the duplicate data items are performed using a sub-cursor
6294 * initialized when the sub-database is first accessed. This function does
6295 * the preliminary setup of the sub-cursor, filling in the fields that
6296 * depend only on the parent DB.
6297 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6300 mdb_xcursor_init0(MDB_cursor *mc)
6302 MDB_xcursor *mx = mc->mc_xcursor;
6304 mx->mx_cursor.mc_xcursor = NULL;
6305 mx->mx_cursor.mc_txn = mc->mc_txn;
6306 mx->mx_cursor.mc_db = &mx->mx_db;
6307 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6308 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6309 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6310 mx->mx_cursor.mc_snum = 0;
6311 mx->mx_cursor.mc_top = 0;
6312 mx->mx_cursor.mc_flags = C_SUB;
6313 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6314 mx->mx_dbx.md_dcmp = NULL;
6315 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6318 /** Final setup of a sorted-dups cursor.
6319 * Sets up the fields that depend on the data from the main cursor.
6320 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6321 * @param[in] node The data containing the #MDB_db record for the
6322 * sorted-dup database.
6325 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6327 MDB_xcursor *mx = mc->mc_xcursor;
6329 if (node->mn_flags & F_SUBDATA) {
6330 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6331 mx->mx_cursor.mc_pg[0] = 0;
6332 mx->mx_cursor.mc_snum = 0;
6333 mx->mx_cursor.mc_flags = C_SUB;
6335 MDB_page *fp = NODEDATA(node);
6336 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6337 mx->mx_db.md_flags = 0;
6338 mx->mx_db.md_depth = 1;
6339 mx->mx_db.md_branch_pages = 0;
6340 mx->mx_db.md_leaf_pages = 1;
6341 mx->mx_db.md_overflow_pages = 0;
6342 mx->mx_db.md_entries = NUMKEYS(fp);
6343 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6344 mx->mx_cursor.mc_snum = 1;
6345 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6346 mx->mx_cursor.mc_top = 0;
6347 mx->mx_cursor.mc_pg[0] = fp;
6348 mx->mx_cursor.mc_ki[0] = 0;
6349 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6350 mx->mx_db.md_flags = MDB_DUPFIXED;
6351 mx->mx_db.md_pad = fp->mp_pad;
6352 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6353 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6356 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6358 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6360 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6361 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6362 #if UINT_MAX < SIZE_MAX
6363 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6364 #ifdef MISALIGNED_OK
6365 mx->mx_dbx.md_cmp = mdb_cmp_long;
6367 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6372 /** Initialize a cursor for a given transaction and database. */
6374 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6377 mc->mc_backup = NULL;
6380 mc->mc_db = &txn->mt_dbs[dbi];
6381 mc->mc_dbx = &txn->mt_dbxs[dbi];
6382 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6387 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6389 mc->mc_xcursor = mx;
6390 mdb_xcursor_init0(mc);
6392 mc->mc_xcursor = NULL;
6394 if (*mc->mc_dbflag & DB_STALE) {
6395 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6400 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6403 size_t size = sizeof(MDB_cursor);
6405 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6408 /* Allow read access to the freelist */
6409 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6412 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6413 size += sizeof(MDB_xcursor);
6415 if ((mc = malloc(size)) != NULL) {
6416 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6417 if (txn->mt_cursors) {
6418 mc->mc_next = txn->mt_cursors[dbi];
6419 txn->mt_cursors[dbi] = mc;
6420 mc->mc_flags |= C_UNTRACK;
6432 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6434 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6437 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6440 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6444 /* Return the count of duplicate data items for the current key */
6446 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6450 if (mc == NULL || countp == NULL)
6453 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
6456 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6457 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6460 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6463 *countp = mc->mc_xcursor->mx_db.md_entries;
6469 mdb_cursor_close(MDB_cursor *mc)
6471 if (mc && !mc->mc_backup) {
6472 /* remove from txn, if tracked */
6473 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6474 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6475 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6477 *prev = mc->mc_next;
6484 mdb_cursor_txn(MDB_cursor *mc)
6486 if (!mc) return NULL;
6491 mdb_cursor_dbi(MDB_cursor *mc)
6497 /** Replace the key for a node with a new key.
6498 * @param[in] mc Cursor pointing to the node to operate on.
6499 * @param[in] key The new key to use.
6500 * @return 0 on success, non-zero on failure.
6503 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6510 indx_t ptr, i, numkeys, indx;
6513 indx = mc->mc_ki[mc->mc_top];
6514 mp = mc->mc_pg[mc->mc_top];
6515 node = NODEPTR(mp, indx);
6516 ptr = mp->mp_ptrs[indx];
6520 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6521 k2.mv_data = NODEKEY(node);
6522 k2.mv_size = node->mn_ksize;
6523 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6525 mdb_dkey(&k2, kbuf2),
6531 delta0 = delta = key->mv_size - node->mn_ksize;
6533 /* Must be 2-byte aligned. If new key is
6534 * shorter by 1, the shift will be skipped.
6536 delta += (delta & 1);
6538 if (delta > 0 && SIZELEFT(mp) < delta) {
6540 /* not enough space left, do a delete and split */
6541 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6542 pgno = NODEPGNO(node);
6543 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6544 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6547 numkeys = NUMKEYS(mp);
6548 for (i = 0; i < numkeys; i++) {
6549 if (mp->mp_ptrs[i] <= ptr)
6550 mp->mp_ptrs[i] -= delta;
6553 base = (char *)mp + mp->mp_upper;
6554 len = ptr - mp->mp_upper + NODESIZE;
6555 memmove(base - delta, base, len);
6556 mp->mp_upper -= delta;
6558 node = NODEPTR(mp, indx);
6561 /* But even if no shift was needed, update ksize */
6563 node->mn_ksize = key->mv_size;
6566 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6572 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6574 /** Move a node from csrc to cdst.
6577 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6584 unsigned short flags;
6588 /* Mark src and dst as dirty. */
6589 if ((rc = mdb_page_touch(csrc)) ||
6590 (rc = mdb_page_touch(cdst)))
6593 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6594 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6595 key.mv_size = csrc->mc_db->md_pad;
6596 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6598 data.mv_data = NULL;
6602 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6603 assert(!((long)srcnode&1));
6604 srcpg = NODEPGNO(srcnode);
6605 flags = srcnode->mn_flags;
6606 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6607 unsigned int snum = csrc->mc_snum;
6609 /* must find the lowest key below src */
6610 mdb_page_search_lowest(csrc);
6611 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6612 key.mv_size = csrc->mc_db->md_pad;
6613 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6615 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6616 key.mv_size = NODEKSZ(s2);
6617 key.mv_data = NODEKEY(s2);
6619 csrc->mc_snum = snum--;
6620 csrc->mc_top = snum;
6622 key.mv_size = NODEKSZ(srcnode);
6623 key.mv_data = NODEKEY(srcnode);
6625 data.mv_size = NODEDSZ(srcnode);
6626 data.mv_data = NODEDATA(srcnode);
6628 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6629 unsigned int snum = cdst->mc_snum;
6632 /* must find the lowest key below dst */
6633 mdb_page_search_lowest(cdst);
6634 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6635 bkey.mv_size = cdst->mc_db->md_pad;
6636 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6638 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6639 bkey.mv_size = NODEKSZ(s2);
6640 bkey.mv_data = NODEKEY(s2);
6642 cdst->mc_snum = snum--;
6643 cdst->mc_top = snum;
6644 mdb_cursor_copy(cdst, &mn);
6646 rc = mdb_update_key(&mn, &bkey);
6651 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6652 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6653 csrc->mc_ki[csrc->mc_top],
6655 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6656 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6658 /* Add the node to the destination page.
6660 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6661 if (rc != MDB_SUCCESS)
6664 /* Delete the node from the source page.
6666 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6669 /* Adjust other cursors pointing to mp */
6670 MDB_cursor *m2, *m3;
6671 MDB_dbi dbi = csrc->mc_dbi;
6672 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6674 if (csrc->mc_flags & C_SUB)
6677 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6678 if (csrc->mc_flags & C_SUB)
6679 m3 = &m2->mc_xcursor->mx_cursor;
6682 if (m3 == csrc) continue;
6683 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6684 csrc->mc_ki[csrc->mc_top]) {
6685 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6686 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6691 /* Update the parent separators.
6693 if (csrc->mc_ki[csrc->mc_top] == 0) {
6694 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6695 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6696 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6698 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6699 key.mv_size = NODEKSZ(srcnode);
6700 key.mv_data = NODEKEY(srcnode);
6702 DPRINTF("update separator for source page %zu to [%s]",
6703 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6704 mdb_cursor_copy(csrc, &mn);
6707 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6710 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6712 indx_t ix = csrc->mc_ki[csrc->mc_top];
6713 nullkey.mv_size = 0;
6714 csrc->mc_ki[csrc->mc_top] = 0;
6715 rc = mdb_update_key(csrc, &nullkey);
6716 csrc->mc_ki[csrc->mc_top] = ix;
6717 assert(rc == MDB_SUCCESS);
6721 if (cdst->mc_ki[cdst->mc_top] == 0) {
6722 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6723 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6724 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6726 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6727 key.mv_size = NODEKSZ(srcnode);
6728 key.mv_data = NODEKEY(srcnode);
6730 DPRINTF("update separator for destination page %zu to [%s]",
6731 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6732 mdb_cursor_copy(cdst, &mn);
6735 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6738 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6740 indx_t ix = cdst->mc_ki[cdst->mc_top];
6741 nullkey.mv_size = 0;
6742 cdst->mc_ki[cdst->mc_top] = 0;
6743 rc = mdb_update_key(cdst, &nullkey);
6744 cdst->mc_ki[cdst->mc_top] = ix;
6745 assert(rc == MDB_SUCCESS);
6752 /** Merge one page into another.
6753 * The nodes from the page pointed to by \b csrc will
6754 * be copied to the page pointed to by \b cdst and then
6755 * the \b csrc page will be freed.
6756 * @param[in] csrc Cursor pointing to the source page.
6757 * @param[in] cdst Cursor pointing to the destination page.
6760 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6768 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6769 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6771 assert(csrc->mc_snum > 1); /* can't merge root page */
6772 assert(cdst->mc_snum > 1);
6774 /* Mark dst as dirty. */
6775 if ((rc = mdb_page_touch(cdst)))
6778 /* Move all nodes from src to dst.
6780 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6781 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6782 key.mv_size = csrc->mc_db->md_pad;
6783 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6784 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6785 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6786 if (rc != MDB_SUCCESS)
6788 key.mv_data = (char *)key.mv_data + key.mv_size;
6791 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6792 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6793 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6794 unsigned int snum = csrc->mc_snum;
6796 /* must find the lowest key below src */
6797 mdb_page_search_lowest(csrc);
6798 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6799 key.mv_size = csrc->mc_db->md_pad;
6800 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6802 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6803 key.mv_size = NODEKSZ(s2);
6804 key.mv_data = NODEKEY(s2);
6806 csrc->mc_snum = snum--;
6807 csrc->mc_top = snum;
6809 key.mv_size = srcnode->mn_ksize;
6810 key.mv_data = NODEKEY(srcnode);
6813 data.mv_size = NODEDSZ(srcnode);
6814 data.mv_data = NODEDATA(srcnode);
6815 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6816 if (rc != MDB_SUCCESS)
6821 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6822 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);
6824 /* Unlink the src page from parent and add to free list.
6826 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6827 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6830 rc = mdb_update_key(csrc, &key);
6836 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6837 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6840 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6841 csrc->mc_db->md_leaf_pages--;
6843 csrc->mc_db->md_branch_pages--;
6845 /* Adjust other cursors pointing to mp */
6846 MDB_cursor *m2, *m3;
6847 MDB_dbi dbi = csrc->mc_dbi;
6848 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6850 if (csrc->mc_flags & C_SUB)
6853 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6854 if (csrc->mc_flags & C_SUB)
6855 m3 = &m2->mc_xcursor->mx_cursor;
6858 if (m3 == csrc) continue;
6859 if (m3->mc_snum < csrc->mc_snum) continue;
6860 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6861 m3->mc_pg[csrc->mc_top] = mp;
6862 m3->mc_ki[csrc->mc_top] += nkeys;
6866 mdb_cursor_pop(csrc);
6868 return mdb_rebalance(csrc);
6871 /** Copy the contents of a cursor.
6872 * @param[in] csrc The cursor to copy from.
6873 * @param[out] cdst The cursor to copy to.
6876 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6880 cdst->mc_txn = csrc->mc_txn;
6881 cdst->mc_dbi = csrc->mc_dbi;
6882 cdst->mc_db = csrc->mc_db;
6883 cdst->mc_dbx = csrc->mc_dbx;
6884 cdst->mc_snum = csrc->mc_snum;
6885 cdst->mc_top = csrc->mc_top;
6886 cdst->mc_flags = csrc->mc_flags;
6888 for (i=0; i<csrc->mc_snum; i++) {
6889 cdst->mc_pg[i] = csrc->mc_pg[i];
6890 cdst->mc_ki[i] = csrc->mc_ki[i];
6894 /** Rebalance the tree after a delete operation.
6895 * @param[in] mc Cursor pointing to the page where rebalancing
6897 * @return 0 on success, non-zero on failure.
6900 mdb_rebalance(MDB_cursor *mc)
6904 unsigned int ptop, minkeys;
6907 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6911 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6912 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6913 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6914 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6918 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6919 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6922 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6923 DPRINTF("no need to rebalance page %zu, above fill threshold",
6929 if (mc->mc_snum < 2) {
6930 MDB_page *mp = mc->mc_pg[0];
6932 DPUTS("Can't rebalance a subpage, ignoring");
6935 if (NUMKEYS(mp) == 0) {
6936 DPUTS("tree is completely empty");
6937 mc->mc_db->md_root = P_INVALID;
6938 mc->mc_db->md_depth = 0;
6939 mc->mc_db->md_leaf_pages = 0;
6940 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6943 /* Adjust cursors pointing to mp */
6947 MDB_cursor *m2, *m3;
6948 MDB_dbi dbi = mc->mc_dbi;
6950 if (mc->mc_flags & C_SUB)
6953 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6954 if (mc->mc_flags & C_SUB)
6955 m3 = &m2->mc_xcursor->mx_cursor;
6958 if (m3->mc_snum < mc->mc_snum) continue;
6959 if (m3->mc_pg[0] == mp) {
6965 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6966 DPUTS("collapsing root page!");
6967 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6970 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6971 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6974 mc->mc_db->md_depth--;
6975 mc->mc_db->md_branch_pages--;
6976 mc->mc_ki[0] = mc->mc_ki[1];
6978 /* Adjust other cursors pointing to mp */
6979 MDB_cursor *m2, *m3;
6980 MDB_dbi dbi = mc->mc_dbi;
6982 if (mc->mc_flags & C_SUB)
6985 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6986 if (mc->mc_flags & C_SUB)
6987 m3 = &m2->mc_xcursor->mx_cursor;
6990 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6991 if (m3->mc_pg[0] == mp) {
6992 m3->mc_pg[0] = mc->mc_pg[0];
6995 m3->mc_ki[0] = m3->mc_ki[1];
7000 DPUTS("root page doesn't need rebalancing");
7004 /* The parent (branch page) must have at least 2 pointers,
7005 * otherwise the tree is invalid.
7007 ptop = mc->mc_top-1;
7008 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7010 /* Leaf page fill factor is below the threshold.
7011 * Try to move keys from left or right neighbor, or
7012 * merge with a neighbor page.
7017 mdb_cursor_copy(mc, &mn);
7018 mn.mc_xcursor = NULL;
7020 if (mc->mc_ki[ptop] == 0) {
7021 /* We're the leftmost leaf in our parent.
7023 DPUTS("reading right neighbor");
7025 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7026 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7029 mn.mc_ki[mn.mc_top] = 0;
7030 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7032 /* There is at least one neighbor to the left.
7034 DPUTS("reading left neighbor");
7036 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7037 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7040 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7041 mc->mc_ki[mc->mc_top] = 0;
7044 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
7045 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);
7047 /* If the neighbor page is above threshold and has enough keys,
7048 * move one key from it. Otherwise we should try to merge them.
7049 * (A branch page must never have less than 2 keys.)
7051 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7052 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7053 return mdb_node_move(&mn, mc);
7055 if (mc->mc_ki[ptop] == 0)
7056 rc = mdb_page_merge(&mn, mc);
7058 rc = mdb_page_merge(mc, &mn);
7059 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7064 /** Complete a delete operation started by #mdb_cursor_del(). */
7066 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7072 mp = mc->mc_pg[mc->mc_top];
7073 ki = mc->mc_ki[mc->mc_top];
7075 /* add overflow pages to free list */
7076 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7080 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7081 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7082 (rc = mdb_ovpage_free(mc, omp)))
7085 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7086 mc->mc_db->md_entries--;
7087 rc = mdb_rebalance(mc);
7088 if (rc != MDB_SUCCESS)
7089 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7090 /* if mc points past last node in page, invalidate */
7091 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7092 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7095 /* Adjust other cursors pointing to mp */
7098 MDB_dbi dbi = mc->mc_dbi;
7100 mp = mc->mc_pg[mc->mc_top];
7101 nkeys = NUMKEYS(mp);
7102 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7105 if (!(m2->mc_flags & C_INITIALIZED))
7107 if (m2->mc_pg[mc->mc_top] == mp) {
7108 if (m2->mc_ki[mc->mc_top] > ki)
7109 m2->mc_ki[mc->mc_top]--;
7110 if (m2->mc_ki[mc->mc_top] >= nkeys)
7111 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7120 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7121 MDB_val *key, MDB_val *data)
7126 MDB_val rdata, *xdata;
7130 assert(key != NULL);
7132 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
7134 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7137 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7141 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7145 mdb_cursor_init(&mc, txn, dbi, &mx);
7156 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7158 /* let mdb_page_split know about this cursor if needed:
7159 * delete will trigger a rebalance; if it needs to move
7160 * a node from one page to another, it will have to
7161 * update the parent's separator key(s). If the new sepkey
7162 * is larger than the current one, the parent page may
7163 * run out of space, triggering a split. We need this
7164 * cursor to be consistent until the end of the rebalance.
7166 mc.mc_flags |= C_UNTRACK;
7167 mc.mc_next = txn->mt_cursors[dbi];
7168 txn->mt_cursors[dbi] = &mc;
7169 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7170 txn->mt_cursors[dbi] = mc.mc_next;
7175 /** Split a page and insert a new node.
7176 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7177 * The cursor will be updated to point to the actual page and index where
7178 * the node got inserted after the split.
7179 * @param[in] newkey The key for the newly inserted node.
7180 * @param[in] newdata The data for the newly inserted node.
7181 * @param[in] newpgno The page number, if the new node is a branch node.
7182 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7183 * @return 0 on success, non-zero on failure.
7186 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7187 unsigned int nflags)
7190 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7193 unsigned int i, j, split_indx, nkeys, pmax;
7195 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7197 MDB_page *mp, *rp, *pp;
7202 mp = mc->mc_pg[mc->mc_top];
7203 newindx = mc->mc_ki[mc->mc_top];
7205 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
7206 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7207 DKEY(newkey), mc->mc_ki[mc->mc_top]);
7209 /* Create a right sibling. */
7210 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7212 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
7214 if (mc->mc_snum < 2) {
7215 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7217 /* shift current top to make room for new parent */
7218 mc->mc_pg[1] = mc->mc_pg[0];
7219 mc->mc_ki[1] = mc->mc_ki[0];
7222 mc->mc_db->md_root = pp->mp_pgno;
7223 DPRINTF("root split! new root = %zu", pp->mp_pgno);
7224 mc->mc_db->md_depth++;
7227 /* Add left (implicit) pointer. */
7228 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7229 /* undo the pre-push */
7230 mc->mc_pg[0] = mc->mc_pg[1];
7231 mc->mc_ki[0] = mc->mc_ki[1];
7232 mc->mc_db->md_root = mp->mp_pgno;
7233 mc->mc_db->md_depth--;
7240 ptop = mc->mc_top-1;
7241 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
7244 mc->mc_flags |= C_SPLITTING;
7245 mdb_cursor_copy(mc, &mn);
7246 mn.mc_pg[mn.mc_top] = rp;
7247 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7249 if (nflags & MDB_APPEND) {
7250 mn.mc_ki[mn.mc_top] = 0;
7252 split_indx = newindx;
7257 nkeys = NUMKEYS(mp);
7258 split_indx = nkeys / 2;
7259 if (newindx < split_indx)
7265 unsigned int lsize, rsize, ksize;
7266 /* Move half of the keys to the right sibling */
7268 x = mc->mc_ki[mc->mc_top] - split_indx;
7269 ksize = mc->mc_db->md_pad;
7270 split = LEAF2KEY(mp, split_indx, ksize);
7271 rsize = (nkeys - split_indx) * ksize;
7272 lsize = (nkeys - split_indx) * sizeof(indx_t);
7273 mp->mp_lower -= lsize;
7274 rp->mp_lower += lsize;
7275 mp->mp_upper += rsize - lsize;
7276 rp->mp_upper -= rsize - lsize;
7277 sepkey.mv_size = ksize;
7278 if (newindx == split_indx) {
7279 sepkey.mv_data = newkey->mv_data;
7281 sepkey.mv_data = split;
7284 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7285 memcpy(rp->mp_ptrs, split, rsize);
7286 sepkey.mv_data = rp->mp_ptrs;
7287 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7288 memcpy(ins, newkey->mv_data, ksize);
7289 mp->mp_lower += sizeof(indx_t);
7290 mp->mp_upper -= ksize - sizeof(indx_t);
7293 memcpy(rp->mp_ptrs, split, x * ksize);
7294 ins = LEAF2KEY(rp, x, ksize);
7295 memcpy(ins, newkey->mv_data, ksize);
7296 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7297 rp->mp_lower += sizeof(indx_t);
7298 rp->mp_upper -= ksize - sizeof(indx_t);
7299 mc->mc_ki[mc->mc_top] = x;
7300 mc->mc_pg[mc->mc_top] = rp;
7305 /* For leaf pages, check the split point based on what
7306 * fits where, since otherwise mdb_node_add can fail.
7308 * This check is only needed when the data items are
7309 * relatively large, such that being off by one will
7310 * make the difference between success or failure.
7312 * It's also relevant if a page happens to be laid out
7313 * such that one half of its nodes are all "small" and
7314 * the other half of its nodes are "large." If the new
7315 * item is also "large" and falls on the half with
7316 * "large" nodes, it also may not fit.
7319 unsigned int psize, nsize;
7320 /* Maximum free space in an empty page */
7321 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7322 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7323 if ((nkeys < 20) || (nsize > pmax/16)) {
7324 if (newindx <= split_indx) {
7327 for (i=0; i<split_indx; i++) {
7328 node = NODEPTR(mp, i);
7329 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7330 if (F_ISSET(node->mn_flags, F_BIGDATA))
7331 psize += sizeof(pgno_t);
7333 psize += NODEDSZ(node);
7337 split_indx = newindx;
7348 for (i=nkeys-1; i>=split_indx; i--) {
7349 node = NODEPTR(mp, i);
7350 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7351 if (F_ISSET(node->mn_flags, F_BIGDATA))
7352 psize += sizeof(pgno_t);
7354 psize += NODEDSZ(node);
7358 split_indx = newindx;
7369 /* First find the separating key between the split pages.
7370 * The case where newindx == split_indx is ambiguous; the
7371 * new item could go to the new page or stay on the original
7372 * page. If newpos == 1 it goes to the new page.
7374 if (newindx == split_indx && newpos) {
7375 sepkey.mv_size = newkey->mv_size;
7376 sepkey.mv_data = newkey->mv_data;
7378 node = NODEPTR(mp, split_indx);
7379 sepkey.mv_size = node->mn_ksize;
7380 sepkey.mv_data = NODEKEY(node);
7384 DPRINTF("separator is [%s]", DKEY(&sepkey));
7386 /* Copy separator key to the parent.
7388 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7392 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7395 if (mn.mc_snum == mc->mc_snum) {
7396 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7397 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7398 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7399 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7404 /* Right page might now have changed parent.
7405 * Check if left page also changed parent.
7407 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7408 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7409 for (i=0; i<ptop; i++) {
7410 mc->mc_pg[i] = mn.mc_pg[i];
7411 mc->mc_ki[i] = mn.mc_ki[i];
7413 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7414 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7418 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7421 mc->mc_flags ^= C_SPLITTING;
7422 if (rc != MDB_SUCCESS) {
7425 if (nflags & MDB_APPEND) {
7426 mc->mc_pg[mc->mc_top] = rp;
7427 mc->mc_ki[mc->mc_top] = 0;
7428 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7431 for (i=0; i<mc->mc_top; i++)
7432 mc->mc_ki[i] = mn.mc_ki[i];
7439 /* Move half of the keys to the right sibling. */
7441 /* grab a page to hold a temporary copy */
7442 copy = mdb_page_malloc(mc->mc_txn, 1);
7446 copy->mp_pgno = mp->mp_pgno;
7447 copy->mp_flags = mp->mp_flags;
7448 copy->mp_lower = PAGEHDRSZ;
7449 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7450 mc->mc_pg[mc->mc_top] = copy;
7451 for (i = j = 0; i <= nkeys; j++) {
7452 if (i == split_indx) {
7453 /* Insert in right sibling. */
7454 /* Reset insert index for right sibling. */
7455 if (i != newindx || (newpos ^ ins_new)) {
7457 mc->mc_pg[mc->mc_top] = rp;
7461 if (i == newindx && !ins_new) {
7462 /* Insert the original entry that caused the split. */
7463 rkey.mv_data = newkey->mv_data;
7464 rkey.mv_size = newkey->mv_size;
7473 /* Update index for the new key. */
7474 mc->mc_ki[mc->mc_top] = j;
7475 } else if (i == nkeys) {
7478 node = NODEPTR(mp, i);
7479 rkey.mv_data = NODEKEY(node);
7480 rkey.mv_size = node->mn_ksize;
7482 xdata.mv_data = NODEDATA(node);
7483 xdata.mv_size = NODEDSZ(node);
7486 pgno = NODEPGNO(node);
7487 flags = node->mn_flags;
7492 if (!IS_LEAF(mp) && j == 0) {
7493 /* First branch index doesn't need key data. */
7497 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7501 nkeys = NUMKEYS(copy);
7502 for (i=0; i<nkeys; i++)
7503 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7504 mp->mp_lower = copy->mp_lower;
7505 mp->mp_upper = copy->mp_upper;
7506 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7507 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7509 /* reset back to original page */
7510 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7511 mc->mc_pg[mc->mc_top] = mp;
7512 if (nflags & MDB_RESERVE) {
7513 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7514 if (!(node->mn_flags & F_BIGDATA))
7515 newdata->mv_data = NODEDATA(node);
7519 /* Make sure mc_ki is still valid.
7521 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7522 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7523 for (i=0; i<ptop; i++) {
7524 mc->mc_pg[i] = mn.mc_pg[i];
7525 mc->mc_ki[i] = mn.mc_ki[i];
7527 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7528 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7532 /* return tmp page to freelist */
7533 mdb_page_free(mc->mc_txn->mt_env, copy);
7536 /* Adjust other cursors pointing to mp */
7537 MDB_cursor *m2, *m3;
7538 MDB_dbi dbi = mc->mc_dbi;
7539 int fixup = NUMKEYS(mp);
7541 if (mc->mc_flags & C_SUB)
7544 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7545 if (mc->mc_flags & C_SUB)
7546 m3 = &m2->mc_xcursor->mx_cursor;
7551 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7553 if (m3->mc_flags & C_SPLITTING)
7558 for (k=m3->mc_top; k>=0; k--) {
7559 m3->mc_ki[k+1] = m3->mc_ki[k];
7560 m3->mc_pg[k+1] = m3->mc_pg[k];
7562 if (m3->mc_ki[0] >= split_indx) {
7567 m3->mc_pg[0] = mc->mc_pg[0];
7571 if (m3->mc_pg[mc->mc_top] == mp) {
7572 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7573 m3->mc_ki[mc->mc_top]++;
7574 if (m3->mc_ki[mc->mc_top] >= fixup) {
7575 m3->mc_pg[mc->mc_top] = rp;
7576 m3->mc_ki[mc->mc_top] -= fixup;
7577 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7579 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7580 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7589 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7590 MDB_val *key, MDB_val *data, unsigned int flags)
7595 assert(key != NULL);
7596 assert(data != NULL);
7598 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7601 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7605 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7609 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7612 mdb_cursor_init(&mc, txn, dbi, &mx);
7613 return mdb_cursor_put(&mc, key, data, flags);
7617 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7619 if ((flag & CHANGEABLE) != flag)
7622 env->me_flags |= flag;
7624 env->me_flags &= ~flag;
7629 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7634 *arg = env->me_flags;
7639 mdb_env_get_path(MDB_env *env, const char **arg)
7644 *arg = env->me_path;
7648 /** Common code for #mdb_stat() and #mdb_env_stat().
7649 * @param[in] env the environment to operate in.
7650 * @param[in] db the #MDB_db record containing the stats to return.
7651 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7652 * @return 0, this function always succeeds.
7655 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7657 arg->ms_psize = env->me_psize;
7658 arg->ms_depth = db->md_depth;
7659 arg->ms_branch_pages = db->md_branch_pages;
7660 arg->ms_leaf_pages = db->md_leaf_pages;
7661 arg->ms_overflow_pages = db->md_overflow_pages;
7662 arg->ms_entries = db->md_entries;
7667 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7671 if (env == NULL || arg == NULL)
7674 toggle = mdb_env_pick_meta(env);
7676 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7680 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7684 if (env == NULL || arg == NULL)
7687 toggle = mdb_env_pick_meta(env);
7688 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7689 arg->me_mapsize = env->me_mapsize;
7690 arg->me_maxreaders = env->me_maxreaders;
7692 /* me_numreaders may be zero if this process never used any readers. Use
7693 * the shared numreader count if it exists.
7695 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7697 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7698 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7702 /** Set the default comparison functions for a database.
7703 * Called immediately after a database is opened to set the defaults.
7704 * The user can then override them with #mdb_set_compare() or
7705 * #mdb_set_dupsort().
7706 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7707 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7710 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7712 uint16_t f = txn->mt_dbs[dbi].md_flags;
7714 txn->mt_dbxs[dbi].md_cmp =
7715 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7716 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7718 txn->mt_dbxs[dbi].md_dcmp =
7719 !(f & MDB_DUPSORT) ? 0 :
7720 ((f & MDB_INTEGERDUP)
7721 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7722 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7725 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7730 int rc, dbflag, exact;
7731 unsigned int unused = 0;
7734 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7735 mdb_default_cmp(txn, FREE_DBI);
7738 if ((flags & VALID_FLAGS) != flags)
7744 if (flags & PERSISTENT_FLAGS) {
7745 uint16_t f2 = flags & PERSISTENT_FLAGS;
7746 /* make sure flag changes get committed */
7747 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7748 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7749 txn->mt_flags |= MDB_TXN_DIRTY;
7752 mdb_default_cmp(txn, MAIN_DBI);
7756 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7757 mdb_default_cmp(txn, MAIN_DBI);
7760 /* Is the DB already open? */
7762 for (i=2; i<txn->mt_numdbs; i++) {
7763 if (!txn->mt_dbxs[i].md_name.mv_size) {
7764 /* Remember this free slot */
7765 if (!unused) unused = i;
7768 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7769 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7775 /* If no free slot and max hit, fail */
7776 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7777 return MDB_DBS_FULL;
7779 /* Cannot mix named databases with some mainDB flags */
7780 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7781 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7783 /* Find the DB info */
7784 dbflag = DB_NEW|DB_VALID;
7787 key.mv_data = (void *)name;
7788 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7789 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7790 if (rc == MDB_SUCCESS) {
7791 /* make sure this is actually a DB */
7792 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7793 if (!(node->mn_flags & F_SUBDATA))
7795 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7796 /* Create if requested */
7798 data.mv_size = sizeof(MDB_db);
7799 data.mv_data = &dummy;
7800 memset(&dummy, 0, sizeof(dummy));
7801 dummy.md_root = P_INVALID;
7802 dummy.md_flags = flags & PERSISTENT_FLAGS;
7803 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7807 /* OK, got info, add to table */
7808 if (rc == MDB_SUCCESS) {
7809 unsigned int slot = unused ? unused : txn->mt_numdbs;
7810 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7811 txn->mt_dbxs[slot].md_name.mv_size = len;
7812 txn->mt_dbxs[slot].md_rel = NULL;
7813 txn->mt_dbflags[slot] = dbflag;
7814 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7816 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7817 mdb_default_cmp(txn, slot);
7826 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7828 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7831 if (txn->mt_dbflags[dbi] & DB_STALE) {
7834 /* Stale, must read the DB's root. cursor_init does it for us. */
7835 mdb_cursor_init(&mc, txn, dbi, &mx);
7837 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7840 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7843 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7845 ptr = env->me_dbxs[dbi].md_name.mv_data;
7846 env->me_dbxs[dbi].md_name.mv_data = NULL;
7847 env->me_dbxs[dbi].md_name.mv_size = 0;
7848 env->me_dbflags[dbi] = 0;
7852 int mdb_dbi_flags(MDB_env *env, MDB_dbi dbi, unsigned int *flags)
7854 /* We could return the flags for the FREE_DBI too but what's the point? */
7855 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7857 *flags = env->me_dbflags[dbi];
7861 /** Add all the DB's pages to the free list.
7862 * @param[in] mc Cursor on the DB to free.
7863 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7864 * @return 0 on success, non-zero on failure.
7867 mdb_drop0(MDB_cursor *mc, int subs)
7871 rc = mdb_page_search(mc, NULL, 0);
7872 if (rc == MDB_SUCCESS) {
7873 MDB_txn *txn = mc->mc_txn;
7878 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7879 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7882 mdb_cursor_copy(mc, &mx);
7883 while (mc->mc_snum > 0) {
7884 MDB_page *mp = mc->mc_pg[mc->mc_top];
7885 unsigned n = NUMKEYS(mp);
7887 for (i=0; i<n; i++) {
7888 ni = NODEPTR(mp, i);
7889 if (ni->mn_flags & F_BIGDATA) {
7892 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7893 rc = mdb_page_get(txn, pg, &omp, NULL);
7896 assert(IS_OVERFLOW(omp));
7897 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7901 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7902 mdb_xcursor_init1(mc, ni);
7903 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7909 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7911 for (i=0; i<n; i++) {
7913 ni = NODEPTR(mp, i);
7916 mdb_midl_xappend(txn->mt_free_pgs, pg);
7921 mc->mc_ki[mc->mc_top] = i;
7922 rc = mdb_cursor_sibling(mc, 1);
7924 /* no more siblings, go back to beginning
7925 * of previous level.
7929 for (i=1; i<mc->mc_snum; i++) {
7931 mc->mc_pg[i] = mx.mc_pg[i];
7936 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7937 } else if (rc == MDB_NOTFOUND) {
7943 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7945 MDB_cursor *mc, *m2;
7948 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7951 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7954 rc = mdb_cursor_open(txn, dbi, &mc);
7958 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7959 /* Invalidate the dropped DB's cursors */
7960 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7961 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7965 /* Can't delete the main DB */
7966 if (del && dbi > MAIN_DBI) {
7967 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7969 txn->mt_dbflags[dbi] = DB_STALE;
7970 mdb_dbi_close(txn->mt_env, dbi);
7973 /* reset the DB record, mark it dirty */
7974 txn->mt_dbflags[dbi] |= DB_DIRTY;
7975 txn->mt_dbs[dbi].md_depth = 0;
7976 txn->mt_dbs[dbi].md_branch_pages = 0;
7977 txn->mt_dbs[dbi].md_leaf_pages = 0;
7978 txn->mt_dbs[dbi].md_overflow_pages = 0;
7979 txn->mt_dbs[dbi].md_entries = 0;
7980 txn->mt_dbs[dbi].md_root = P_INVALID;
7982 txn->mt_flags |= MDB_TXN_DIRTY;
7985 mdb_cursor_close(mc);
7989 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7991 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7994 txn->mt_dbxs[dbi].md_cmp = cmp;
7998 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8000 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8003 txn->mt_dbxs[dbi].md_dcmp = cmp;
8007 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8009 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8012 txn->mt_dbxs[dbi].md_rel = rel;
8016 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8018 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8021 txn->mt_dbxs[dbi].md_relctx = ctx;
8025 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8027 unsigned int i, rdrs;
8034 if (!env->me_txns) {
8035 return func("(no reader locks)\n", ctx);
8037 rdrs = env->me_txns->mti_numreaders;
8038 mr = env->me_txns->mti_readers;
8039 for (i=0; i<rdrs; i++) {
8044 if (mr[i].mr_txnid == (txnid_t)-1) {
8045 sprintf(buf, "%10d %zx -\n", mr[i].mr_pid, tid);
8047 sprintf(buf, "%10d %zx %zu\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8051 func(" pid thread txnid\n", ctx);
8053 rc = func(buf, ctx);
8059 func("(no active readers)\n", ctx);
8064 /* insert pid into list if not already present.
8065 * return -1 if already present.
8067 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8069 /* binary search of pid in list */
8071 unsigned cursor = 1;
8073 unsigned n = ids[0];
8076 unsigned pivot = n >> 1;
8077 cursor = base + pivot + 1;
8078 val = pid - ids[cursor];
8083 } else if ( val > 0 ) {
8088 /* found, so it's a duplicate */
8097 for (n = ids[0]; n > cursor; n--)
8103 int mdb_reader_check(MDB_env *env, int *dead)
8105 unsigned int i, j, rdrs;
8116 rdrs = env->me_txns->mti_numreaders;
8117 pids = malloc((rdrs+1) * sizeof(pid_t));
8121 mr = env->me_txns->mti_readers;
8123 for (i=0; i<rdrs; i++) {
8124 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8126 if (mdb_pid_insert(pids, pid) == 0) {
8127 if (mdb_reader_pid(env, Pidcheck, pid)) {
8129 if (mdb_reader_pid(env, Pidcheck, pid)) {
8130 for (j=i; j<rdrs; j++)
8131 if (mr[j].mr_pid == pid) {
8136 UNLOCK_MUTEX_R(env);