1 /* $NetBSD: mdb.c,v 1.3 2021/08/14 16:14:57 christos Exp $ */ 2 3 /** @file mdb.c 4 * @brief Lightning memory-mapped database library 5 * 6 * A Btree-based database management library modeled loosely on the 7 * BerkeleyDB API, but much simplified. 8 */ 9 /* 10 * Copyright 2011-2021 Howard Chu, Symas Corp. 11 * All rights reserved. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted only as authorized by the OpenLDAP 15 * Public License. 16 * 17 * A copy of this license is available in the file LICENSE in the 18 * top-level directory of the distribution or, alternatively, at 19 * <http://www.OpenLDAP.org/license.html>. 20 * 21 * This code is derived from btree.c written by Martin Hedenfalk. 22 * 23 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se> 24 * 25 * Permission to use, copy, modify, and distribute this software for any 26 * purpose with or without fee is hereby granted, provided that the above 27 * copyright notice and this permission notice appear in all copies. 28 * 29 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES 30 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF 31 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR 32 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES 33 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN 34 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF 35 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. 36 */ 37 #ifndef _GNU_SOURCE 38 #define _GNU_SOURCE 1 39 #endif 40 #if defined(__WIN64__) 41 #define _FILE_OFFSET_BITS 64 42 #endif 43 #ifdef _WIN32 44 #include <malloc.h> 45 #include <windows.h> 46 #include <wchar.h> /* get wcscpy() */ 47 48 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it 49 * as int64 which is wrong. MSVC doesn't define it at all, so just 50 * don't use it. 51 */ 52 #define MDB_PID_T int 53 #define MDB_THR_T DWORD 54 #include <sys/types.h> 55 #include <sys/stat.h> 56 #ifdef __GNUC__ 57 # include <sys/param.h> 58 #else 59 # define LITTLE_ENDIAN 1234 60 # define BIG_ENDIAN 4321 61 # define BYTE_ORDER LITTLE_ENDIAN 62 # ifndef SSIZE_MAX 63 # define SSIZE_MAX INT_MAX 64 # endif 65 #endif 66 #else 67 #include <sys/types.h> 68 #include <sys/stat.h> 69 #define MDB_PID_T pid_t 70 #define MDB_THR_T pthread_t 71 #include <sys/param.h> 72 #include <sys/uio.h> 73 #include <sys/mman.h> 74 #ifdef HAVE_SYS_FILE_H 75 #include <sys/file.h> 76 #endif 77 #include <fcntl.h> 78 #endif 79 80 #if defined(__mips) && defined(__linux) 81 /* MIPS has cache coherency issues, requires explicit cache control */ 82 #include <asm/cachectl.h> 83 extern int cacheflush(char *addr, int nbytes, int cache); 84 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache) 85 #else 86 #define CACHEFLUSH(addr, bytes, cache) 87 #endif 88 89 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS) 90 /** fdatasync is broken on ext3/ext4fs on older kernels, see 91 * description in #mdb_env_open2 comments. You can safely 92 * define MDB_FDATASYNC_WORKS if this code will only be run 93 * on kernels 3.6 and newer. 94 */ 95 #define BROKEN_FDATASYNC 96 #endif 97 98 #include <errno.h> 99 #include <limits.h> 100 #include <stddef.h> 101 #include <inttypes.h> 102 #include <stdio.h> 103 #include <stdlib.h> 104 #include <string.h> 105 #include <time.h> 106 107 #ifdef _MSC_VER 108 #include <io.h> 109 typedef SSIZE_T ssize_t; 110 #else 111 #include <unistd.h> 112 #endif 113 114 #if defined(__sun) || defined(ANDROID) 115 /* Most platforms have posix_memalign, older may only have memalign */ 116 #define HAVE_MEMALIGN 1 117 #include <malloc.h> 118 /* On Solaris, we need the POSIX sigwait function */ 119 #if defined (__sun) 120 # define _POSIX_PTHREAD_SEMANTICS 1 121 #endif 122 #endif 123 124 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER)) 125 #include <netinet/in.h> 126 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */ 127 #endif 128 129 #if defined(__FreeBSD__) && defined(__FreeBSD_version) && __FreeBSD_version >= 1100110 130 # define MDB_USE_POSIX_MUTEX 1 131 # define MDB_USE_ROBUST 1 132 #elif defined(__APPLE__) || defined (BSD) || defined(__FreeBSD_kernel__) 133 # define MDB_USE_POSIX_SEM 1 134 # define MDB_FDATASYNC fsync 135 #elif defined(ANDROID) 136 # define MDB_FDATASYNC fsync 137 #endif 138 139 #ifndef _WIN32 140 #include <pthread.h> 141 #include <signal.h> 142 #ifdef MDB_USE_POSIX_SEM 143 # define MDB_USE_HASH 1 144 #include <semaphore.h> 145 #else 146 #define MDB_USE_POSIX_MUTEX 1 147 #endif 148 #endif 149 150 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \ 151 + defined(MDB_USE_POSIX_MUTEX) != 1 152 # error "Ambiguous shared-lock implementation" 153 #endif 154 155 #ifdef USE_VALGRIND 156 #include <valgrind/memcheck.h> 157 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z) 158 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s) 159 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a) 160 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h) 161 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s) 162 #else 163 #define VGMEMP_CREATE(h,r,z) 164 #define VGMEMP_ALLOC(h,a,s) 165 #define VGMEMP_FREE(h,a) 166 #define VGMEMP_DESTROY(h) 167 #define VGMEMP_DEFINED(a,s) 168 #endif 169 170 #ifndef BYTE_ORDER 171 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN)) 172 /* Solaris just defines one or the other */ 173 # define LITTLE_ENDIAN 1234 174 # define BIG_ENDIAN 4321 175 # ifdef _LITTLE_ENDIAN 176 # define BYTE_ORDER LITTLE_ENDIAN 177 # else 178 # define BYTE_ORDER BIG_ENDIAN 179 # endif 180 # else 181 # define BYTE_ORDER __BYTE_ORDER 182 # endif 183 #endif 184 185 #ifndef LITTLE_ENDIAN 186 #define LITTLE_ENDIAN __LITTLE_ENDIAN 187 #endif 188 #ifndef BIG_ENDIAN 189 #define BIG_ENDIAN __BIG_ENDIAN 190 #endif 191 192 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86) 193 #define MISALIGNED_OK 1 194 #endif 195 196 #include "lmdb.h" 197 #include "midl.h" 198 199 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN) 200 # error "Unknown or unsupported endianness (BYTE_ORDER)" 201 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF 202 # error "Two's complement, reasonably sized integer types, please" 203 #endif 204 205 #ifdef __GNUC__ 206 /** Put infrequently used env functions in separate section */ 207 # ifdef __APPLE__ 208 # define ESECT __attribute__ ((section("__TEXT,text_env"))) 209 # else 210 # define ESECT __attribute__ ((section("text_env"))) 211 # endif 212 #else 213 #define ESECT 214 #endif 215 216 #ifdef _WIN32 217 #define CALL_CONV WINAPI 218 #else 219 #define CALL_CONV 220 #endif 221 222 /** @defgroup internal LMDB Internals 223 * @{ 224 */ 225 /** @defgroup compat Compatibility Macros 226 * A bunch of macros to minimize the amount of platform-specific ifdefs 227 * needed throughout the rest of the code. When the features this library 228 * needs are similar enough to POSIX to be hidden in a one-or-two line 229 * replacement, this macro approach is used. 230 * @{ 231 */ 232 233 /** Features under development */ 234 #ifndef MDB_DEVEL 235 #define MDB_DEVEL 0 236 #endif 237 238 /** Wrapper around __func__, which is a C99 feature */ 239 #if __STDC_VERSION__ >= 199901L 240 # define mdb_func_ __func__ 241 #elif __GNUC__ >= 2 || _MSC_VER >= 1300 242 # define mdb_func_ __FUNCTION__ 243 #else 244 /* If a debug message says <mdb_unknown>(), update the #if statements above */ 245 # define mdb_func_ "<mdb_unknown>" 246 #endif 247 248 /* Internal error codes, not exposed outside liblmdb */ 249 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10) 250 #ifdef _WIN32 251 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED) 252 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD) 253 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */ 254 #endif 255 256 #ifdef __GLIBC__ 257 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__) 258 #endif 259 /** Some platforms define the EOWNERDEAD error code 260 * even though they don't support Robust Mutexes. 261 * Compile with -DMDB_USE_ROBUST=0, or use some other 262 * mechanism like -DMDB_USE_POSIX_SEM instead of 263 * -DMDB_USE_POSIX_MUTEX. 264 * (Posix semaphores are not robust.) 265 */ 266 #ifndef MDB_USE_ROBUST 267 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */ 268 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \ 269 (defined(__GLIBC__) && GLIBC_VER < 0x020004)) 270 # define MDB_USE_ROBUST 0 271 # else 272 # define MDB_USE_ROBUST 1 273 # endif 274 #endif /* !MDB_USE_ROBUST */ 275 276 #if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST) 277 /* glibc < 2.12 only provided _np API */ 278 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \ 279 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST)) 280 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP 281 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag) 282 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex) 283 # endif 284 #endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */ 285 286 #if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST) 287 #define MDB_ROBUST_SUPPORTED 1 288 #endif 289 290 #ifdef _WIN32 291 #define MDB_USE_HASH 1 292 #define MDB_PIDLOCK 0 293 #define THREAD_RET DWORD 294 #define pthread_t HANDLE 295 #define pthread_mutex_t HANDLE 296 #define pthread_cond_t HANDLE 297 typedef HANDLE mdb_mutex_t, mdb_mutexref_t; 298 #define pthread_key_t DWORD 299 #define pthread_self() GetCurrentThreadId() 300 #define pthread_key_create(x,y) \ 301 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0) 302 #define pthread_key_delete(x) TlsFree(x) 303 #define pthread_getspecific(x) TlsGetValue(x) 304 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode()) 305 #define pthread_mutex_unlock(x) ReleaseMutex(*x) 306 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE) 307 #define pthread_cond_signal(x) SetEvent(*x) 308 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0) 309 #define THREAD_CREATE(thr,start,arg) \ 310 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode()) 311 #define THREAD_FINISH(thr) \ 312 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0) 313 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE) 314 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex) 315 #define mdb_mutex_consistent(mutex) 0 316 #define getpid() GetCurrentProcessId() 317 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd)) 318 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len)) 319 #define ErrCode() GetLastError() 320 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;} 321 #define close(fd) (CloseHandle(fd) ? 0 : -1) 322 #define munmap(ptr,len) UnmapViewOfFile(ptr) 323 #ifdef PROCESS_QUERY_LIMITED_INFORMATION 324 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION 325 #else 326 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000 327 #endif 328 #define Z "I" 329 #else 330 #define THREAD_RET void * 331 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg) 332 #define THREAD_FINISH(thr) pthread_join(thr,NULL) 333 #define Z "z" /**< printf format modifier for size_t */ 334 335 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */ 336 #define MDB_PIDLOCK 1 337 338 #ifdef MDB_USE_POSIX_SEM 339 340 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t; 341 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex) 342 #define UNLOCK_MUTEX(mutex) sem_post(mutex) 343 344 static int 345 mdb_sem_wait(sem_t *sem) 346 { 347 int rc; 348 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ; 349 return rc; 350 } 351 352 #else /* MDB_USE_POSIX_MUTEX: */ 353 /** Shared mutex/semaphore as the original is stored. 354 * 355 * Not for copies. Instead it can be assigned to an #mdb_mutexref_t. 356 * When mdb_mutexref_t is a pointer and mdb_mutex_t is not, then it 357 * is array[size 1] so it can be assigned to the pointer. 358 */ 359 typedef pthread_mutex_t mdb_mutex_t[1]; 360 /** Reference to an #mdb_mutex_t */ 361 typedef pthread_mutex_t *mdb_mutexref_t; 362 /** Lock the reader or writer mutex. 363 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX(). 364 */ 365 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex) 366 /** Unlock the reader or writer mutex. 367 */ 368 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex) 369 /** Mark mutex-protected data as repaired, after death of previous owner. 370 */ 371 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex) 372 #endif /* MDB_USE_POSIX_SEM */ 373 374 /** Get the error code for the last failed system function. 375 */ 376 #define ErrCode() errno 377 378 /** An abstraction for a file handle. 379 * On POSIX systems file handles are small integers. On Windows 380 * they're opaque pointers. 381 */ 382 #define HANDLE int 383 384 /** A value for an invalid file handle. 385 * Mainly used to initialize file variables and signify that they are 386 * unused. 387 */ 388 #define INVALID_HANDLE_VALUE (-1) 389 390 /** Get the size of a memory page for the system. 391 * This is the basic size that the platform's memory manager uses, and is 392 * fundamental to the use of memory-mapped files. 393 */ 394 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE)) 395 #endif 396 397 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM) 398 #define MNAME_LEN 32 399 #else 400 #define MNAME_LEN (sizeof(pthread_mutex_t)) 401 #endif 402 403 /** @} */ 404 405 #ifdef MDB_ROBUST_SUPPORTED 406 /** Lock mutex, handle any error, set rc = result. 407 * Return 0 on success, nonzero (not rc) on error. 408 */ 409 #define LOCK_MUTEX(rc, env, mutex) \ 410 (((rc) = LOCK_MUTEX0(mutex)) && \ 411 ((rc) = mdb_mutex_failed(env, mutex, rc))) 412 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc); 413 #else 414 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex)) 415 #define mdb_mutex_failed(env, mutex, rc) (rc) 416 #endif 417 418 #ifndef _WIN32 419 /** A flag for opening a file and requesting synchronous data writes. 420 * This is only used when writing a meta page. It's not strictly needed; 421 * we could just do a normal write and then immediately perform a flush. 422 * But if this flag is available it saves us an extra system call. 423 * 424 * @note If O_DSYNC is undefined but exists in /usr/include, 425 * preferably set some compiler flag to get the definition. 426 */ 427 #ifndef MDB_DSYNC 428 # ifdef O_DSYNC 429 # define MDB_DSYNC O_DSYNC 430 # else 431 # define MDB_DSYNC O_SYNC 432 # endif 433 #endif 434 #endif 435 436 /** Function for flushing the data of a file. Define this to fsync 437 * if fdatasync() is not supported. 438 */ 439 #ifndef MDB_FDATASYNC 440 # define MDB_FDATASYNC fdatasync 441 #endif 442 443 #ifndef MDB_MSYNC 444 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags) 445 #endif 446 447 #ifndef MS_SYNC 448 #define MS_SYNC 1 449 #endif 450 451 #ifndef MS_ASYNC 452 #define MS_ASYNC 0 453 #endif 454 455 /** A page number in the database. 456 * Note that 64 bit page numbers are overkill, since pages themselves 457 * already represent 12-13 bits of addressable memory, and the OS will 458 * always limit applications to a maximum of 63 bits of address space. 459 * 460 * @note In the #MDB_node structure, we only store 48 bits of this value, 461 * which thus limits us to only 60 bits of addressable data. 462 */ 463 typedef MDB_ID pgno_t; 464 465 /** A transaction ID. 466 * See struct MDB_txn.mt_txnid for details. 467 */ 468 typedef MDB_ID txnid_t; 469 470 /** @defgroup debug Debug Macros 471 * @{ 472 */ 473 #ifndef MDB_DEBUG 474 /** Enable debug output. Needs variable argument macros (a C99 feature). 475 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs 476 * read from and written to the database (used for free space management). 477 */ 478 #define MDB_DEBUG 0 479 #endif 480 481 #if MDB_DEBUG 482 static int mdb_debug; 483 static txnid_t mdb_debug_start; 484 485 /** Print a debug message with printf formatting. 486 * Requires double parenthesis around 2 or more args. 487 */ 488 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args)) 489 # define DPRINTF0(fmt, ...) \ 490 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__) 491 #else 492 # define DPRINTF(args) ((void) 0) 493 #endif 494 /** Print a debug string. 495 * The string is printed literally, with no format processing. 496 */ 497 #define DPUTS(arg) DPRINTF(("%s", arg)) 498 /** Debugging output value of a cursor DBI: Negative in a sub-cursor. */ 499 #define DDBI(mc) \ 500 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi) 501 /** @} */ 502 503 /** @brief The maximum size of a database page. 504 * 505 * It is 32k or 64k, since value-PAGEBASE must fit in 506 * #MDB_page.%mp_upper. 507 * 508 * LMDB will use database pages < OS pages if needed. 509 * That causes more I/O in write transactions: The OS must 510 * know (read) the whole page before writing a partial page. 511 * 512 * Note that we don't currently support Huge pages. On Linux, 513 * regular data files cannot use Huge pages, and in general 514 * Huge pages aren't actually pageable. We rely on the OS 515 * demand-pager to read our data and page it out when memory 516 * pressure from other processes is high. So until OSs have 517 * actual paging support for Huge pages, they're not viable. 518 */ 519 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000) 520 521 /** The minimum number of keys required in a database page. 522 * Setting this to a larger value will place a smaller bound on the 523 * maximum size of a data item. Data items larger than this size will 524 * be pushed into overflow pages instead of being stored directly in 525 * the B-tree node. This value used to default to 4. With a page size 526 * of 4096 bytes that meant that any item larger than 1024 bytes would 527 * go into an overflow page. That also meant that on average 2-3KB of 528 * each overflow page was wasted space. The value cannot be lower than 529 * 2 because then there would no longer be a tree structure. With this 530 * value, items larger than 2KB will go into overflow pages, and on 531 * average only 1KB will be wasted. 532 */ 533 #define MDB_MINKEYS 2 534 535 /** A stamp that identifies a file as an LMDB file. 536 * There's nothing special about this value other than that it is easily 537 * recognizable, and it will reflect any byte order mismatches. 538 */ 539 #define MDB_MAGIC 0xBEEFC0DE 540 541 /** The version number for a database's datafile format. */ 542 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1) 543 /** The version number for a database's lockfile format. */ 544 #define MDB_LOCK_VERSION 1 545 546 /** @brief The max size of a key we can write, or 0 for computed max. 547 * 548 * This macro should normally be left alone or set to 0. 549 * Note that a database with big keys or dupsort data cannot be 550 * reliably modified by a liblmdb which uses a smaller max. 551 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL. 552 * 553 * Other values are allowed, for backwards compat. However: 554 * A value bigger than the computed max can break if you do not 555 * know what you are doing, and liblmdb <= 0.9.10 can break when 556 * modifying a DB with keys/dupsort data bigger than its max. 557 * 558 * Data items in an #MDB_DUPSORT database are also limited to 559 * this size, since they're actually keys of a sub-DB. Keys and 560 * #MDB_DUPSORT data items must fit on a node in a regular page. 561 */ 562 #ifndef MDB_MAXKEYSIZE 563 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511) 564 #endif 565 566 /** The maximum size of a key we can write to the environment. */ 567 #if MDB_MAXKEYSIZE 568 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE) 569 #else 570 #define ENV_MAXKEY(env) ((env)->me_maxkey) 571 #endif 572 573 /** @brief The maximum size of a data item. 574 * 575 * We only store a 32 bit value for node sizes. 576 */ 577 #define MAXDATASIZE 0xffffffffUL 578 579 #if MDB_DEBUG 580 /** Key size which fits in a #DKBUF. 581 * @ingroup debug 582 */ 583 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511) 584 /** A key buffer. 585 * @ingroup debug 586 * This is used for printing a hex dump of a key's contents. 587 */ 588 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1] 589 /** Display a key in hex. 590 * @ingroup debug 591 * Invoke a function to display a key in hex. 592 */ 593 #define DKEY(x) mdb_dkey(x, kbuf) 594 #else 595 #define DKBUF 596 #define DKEY(x) 0 597 #endif 598 599 /** An invalid page number. 600 * Mainly used to denote an empty tree. 601 */ 602 #define P_INVALID (~(pgno_t)0) 603 604 /** Test if the flags \b f are set in a flag word \b w. */ 605 #define F_ISSET(w, f) (((w) & (f)) == (f)) 606 607 /** Round \b n up to an even number. */ 608 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */ 609 610 /** Used for offsets within a single page. 611 * Since memory pages are typically 4 or 8KB in size, 12-13 bits, 612 * this is plenty. 613 */ 614 typedef uint16_t indx_t; 615 616 /** Default size of memory map. 617 * This is certainly too small for any actual applications. Apps should always set 618 * the size explicitly using #mdb_env_set_mapsize(). 619 */ 620 #define DEFAULT_MAPSIZE 1048576 621 622 /** @defgroup readers Reader Lock Table 623 * Readers don't acquire any locks for their data access. Instead, they 624 * simply record their transaction ID in the reader table. The reader 625 * mutex is needed just to find an empty slot in the reader table. The 626 * slot's address is saved in thread-specific data so that subsequent read 627 * transactions started by the same thread need no further locking to proceed. 628 * 629 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data. 630 * 631 * No reader table is used if the database is on a read-only filesystem, or 632 * if #MDB_NOLOCK is set. 633 * 634 * Since the database uses multi-version concurrency control, readers don't 635 * actually need any locking. This table is used to keep track of which 636 * readers are using data from which old transactions, so that we'll know 637 * when a particular old transaction is no longer in use. Old transactions 638 * that have discarded any data pages can then have those pages reclaimed 639 * for use by a later write transaction. 640 * 641 * The lock table is constructed such that reader slots are aligned with the 642 * processor's cache line size. Any slot is only ever used by one thread. 643 * This alignment guarantees that there will be no contention or cache 644 * thrashing as threads update their own slot info, and also eliminates 645 * any need for locking when accessing a slot. 646 * 647 * A writer thread will scan every slot in the table to determine the oldest 648 * outstanding reader transaction. Any freed pages older than this will be 649 * reclaimed by the writer. The writer doesn't use any locks when scanning 650 * this table. This means that there's no guarantee that the writer will 651 * see the most up-to-date reader info, but that's not required for correct 652 * operation - all we need is to know the upper bound on the oldest reader, 653 * we don't care at all about the newest reader. So the only consequence of 654 * reading stale information here is that old pages might hang around a 655 * while longer before being reclaimed. That's actually good anyway, because 656 * the longer we delay reclaiming old pages, the more likely it is that a 657 * string of contiguous pages can be found after coalescing old pages from 658 * many old transactions together. 659 * @{ 660 */ 661 /** Number of slots in the reader table. 662 * This value was chosen somewhat arbitrarily. 126 readers plus a 663 * couple mutexes fit exactly into 8KB on my development machine. 664 * Applications should set the table size using #mdb_env_set_maxreaders(). 665 */ 666 #define DEFAULT_READERS 126 667 668 /** The size of a CPU cache line in bytes. We want our lock structures 669 * aligned to this size to avoid false cache line sharing in the 670 * lock table. 671 * This value works for most CPUs. For Itanium this should be 128. 672 */ 673 #ifndef CACHELINE 674 #define CACHELINE 64 675 #endif 676 677 /** The information we store in a single slot of the reader table. 678 * In addition to a transaction ID, we also record the process and 679 * thread ID that owns a slot, so that we can detect stale information, 680 * e.g. threads or processes that went away without cleaning up. 681 * @note We currently don't check for stale records. We simply re-init 682 * the table when we know that we're the only process opening the 683 * lock file. 684 */ 685 typedef struct MDB_rxbody { 686 /** Current Transaction ID when this transaction began, or (txnid_t)-1. 687 * Multiple readers that start at the same time will probably have the 688 * same ID here. Again, it's not important to exclude them from 689 * anything; all we need to know is which version of the DB they 690 * started from so we can avoid overwriting any data used in that 691 * particular version. 692 */ 693 volatile txnid_t mrb_txnid; 694 /** The process ID of the process owning this reader txn. */ 695 volatile MDB_PID_T mrb_pid; 696 /** The thread ID of the thread owning this txn. */ 697 volatile MDB_THR_T mrb_tid; 698 } MDB_rxbody; 699 700 /** The actual reader record, with cacheline padding. */ 701 typedef struct MDB_reader { 702 union { 703 MDB_rxbody mrx; 704 /** shorthand for mrb_txnid */ 705 #define mr_txnid mru.mrx.mrb_txnid 706 #define mr_pid mru.mrx.mrb_pid 707 #define mr_tid mru.mrx.mrb_tid 708 /** cache line alignment */ 709 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)]; 710 } mru; 711 } MDB_reader; 712 713 /** The header for the reader table. 714 * The table resides in a memory-mapped file. (This is a different file 715 * than is used for the main database.) 716 * 717 * For POSIX the actual mutexes reside in the shared memory of this 718 * mapped file. On Windows, mutexes are named objects allocated by the 719 * kernel; we store the mutex names in this mapped file so that other 720 * processes can grab them. This same approach is also used on 721 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support 722 * process-shared POSIX mutexes. For these cases where a named object 723 * is used, the object name is derived from a 64 bit FNV hash of the 724 * environment pathname. As such, naming collisions are extremely 725 * unlikely. If a collision occurs, the results are unpredictable. 726 */ 727 typedef struct MDB_txbody { 728 /** Stamp identifying this as an LMDB file. It must be set 729 * to #MDB_MAGIC. */ 730 uint32_t mtb_magic; 731 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */ 732 uint32_t mtb_format; 733 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM) 734 char mtb_rmname[MNAME_LEN]; 735 #else 736 /** Mutex protecting access to this table. 737 * This is the reader table lock used with LOCK_MUTEX(). 738 */ 739 mdb_mutex_t mtb_rmutex; 740 #endif 741 /** The ID of the last transaction committed to the database. 742 * This is recorded here only for convenience; the value can always 743 * be determined by reading the main database meta pages. 744 */ 745 volatile txnid_t mtb_txnid; 746 /** The number of slots that have been used in the reader table. 747 * This always records the maximum count, it is not decremented 748 * when readers release their slots. 749 */ 750 volatile unsigned mtb_numreaders; 751 } MDB_txbody; 752 753 /** The actual reader table definition. */ 754 typedef struct MDB_txninfo { 755 union { 756 MDB_txbody mtb; 757 #define mti_magic mt1.mtb.mtb_magic 758 #define mti_format mt1.mtb.mtb_format 759 #define mti_rmutex mt1.mtb.mtb_rmutex 760 #define mti_rmname mt1.mtb.mtb_rmname 761 #define mti_txnid mt1.mtb.mtb_txnid 762 #define mti_numreaders mt1.mtb.mtb_numreaders 763 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)]; 764 } mt1; 765 union { 766 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM) 767 char mt2_wmname[MNAME_LEN]; 768 #define mti_wmname mt2.mt2_wmname 769 #else 770 mdb_mutex_t mt2_wmutex; 771 #define mti_wmutex mt2.mt2_wmutex 772 #endif 773 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)]; 774 } mt2; 775 MDB_reader mti_readers[1]; 776 } MDB_txninfo; 777 778 /** Lockfile format signature: version, features and field layout */ 779 #define MDB_LOCK_FORMAT \ 780 ((uint32_t) \ 781 ((MDB_LOCK_VERSION) \ 782 /* Flags which describe functionality */ \ 783 + (((MDB_PIDLOCK) != 0) << 16))) 784 /** @} */ 785 786 /** Common header for all page types. The page type depends on #mp_flags. 787 * 788 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with 789 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages 790 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header. 791 * 792 * #P_OVERFLOW records occupy one or more contiguous pages where only the 793 * first has a page header. They hold the real data of #F_BIGDATA nodes. 794 * 795 * #P_SUBP sub-pages are small leaf "pages" with duplicate data. 796 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page. 797 * (Duplicate data can also go in sub-databases, which use normal pages.) 798 * 799 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot. 800 * 801 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once 802 * in the snapshot: Either used by a database or listed in a freeDB record. 803 */ 804 typedef struct MDB_page { 805 #define mp_pgno mp_p.p_pgno 806 #define mp_next mp_p.p_next 807 union { 808 pgno_t p_pgno; /**< page number */ 809 struct MDB_page *p_next; /**< for in-memory list of freed pages */ 810 } mp_p; 811 uint16_t mp_pad; /**< key size if this is a LEAF2 page */ 812 /** @defgroup mdb_page Page Flags 813 * @ingroup internal 814 * Flags for the page headers. 815 * @{ 816 */ 817 #define P_BRANCH 0x01 /**< branch page */ 818 #define P_LEAF 0x02 /**< leaf page */ 819 #define P_OVERFLOW 0x04 /**< overflow page */ 820 #define P_META 0x08 /**< meta page */ 821 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */ 822 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */ 823 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */ 824 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */ 825 #define P_KEEP 0x8000 /**< leave this page alone during spill */ 826 /** @} */ 827 uint16_t mp_flags; /**< @ref mdb_page */ 828 #define mp_lower mp_pb.pb.pb_lower 829 #define mp_upper mp_pb.pb.pb_upper 830 #define mp_pages mp_pb.pb_pages 831 union { 832 struct { 833 indx_t pb_lower; /**< lower bound of free space */ 834 indx_t pb_upper; /**< upper bound of free space */ 835 } pb; 836 uint32_t pb_pages; /**< number of overflow pages */ 837 } mp_pb; 838 indx_t mp_ptrs[1]; /**< dynamic size */ 839 } MDB_page; 840 841 /** Size of the page header, excluding dynamic data at the end */ 842 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs)) 843 844 /** Address of first usable data byte in a page, after the header */ 845 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ)) 846 847 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */ 848 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0) 849 850 /** Number of nodes on a page */ 851 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1) 852 853 /** The amount of space remaining in the page */ 854 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower) 855 856 /** The percentage of space used in the page, in tenths of a percent. */ 857 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \ 858 ((env)->me_psize - PAGEHDRSZ)) 859 /** The minimum page fill factor, in tenths of a percent. 860 * Pages emptier than this are candidates for merging. 861 */ 862 #define FILL_THRESHOLD 250 863 864 /** Test if a page is a leaf page */ 865 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF) 866 /** Test if a page is a LEAF2 page */ 867 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2) 868 /** Test if a page is a branch page */ 869 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH) 870 /** Test if a page is an overflow page */ 871 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW) 872 /** Test if a page is a sub page */ 873 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP) 874 875 /** The number of overflow pages needed to store the given size. */ 876 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1) 877 878 /** Link in #MDB_txn.%mt_loose_pgs list. 879 * Kept outside the page header, which is needed when reusing the page. 880 */ 881 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2)) 882 883 /** Header for a single key/data pair within a page. 884 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2. 885 * We guarantee 2-byte alignment for 'MDB_node's. 886 * 887 * #mn_lo and #mn_hi are used for data size on leaf nodes, and for child 888 * pgno on branch nodes. On 64 bit platforms, #mn_flags is also used 889 * for pgno. (Branch nodes have no flags). Lo and hi are in host byte 890 * order in case some accesses can be optimized to 32-bit word access. 891 * 892 * Leaf node flags describe node contents. #F_BIGDATA says the node's 893 * data part is the page number of an overflow page with actual data. 894 * #F_DUPDATA and #F_SUBDATA can be combined giving duplicate data in 895 * a sub-page/sub-database, and named databases (just #F_SUBDATA). 896 */ 897 typedef struct MDB_node { 898 /** part of data size or pgno 899 * @{ */ 900 #if BYTE_ORDER == LITTLE_ENDIAN 901 unsigned short mn_lo, mn_hi; 902 #else 903 unsigned short mn_hi, mn_lo; 904 #endif 905 /** @} */ 906 /** @defgroup mdb_node Node Flags 907 * @ingroup internal 908 * Flags for node headers. 909 * @{ 910 */ 911 #define F_BIGDATA 0x01 /**< data put on overflow page */ 912 #define F_SUBDATA 0x02 /**< data is a sub-database */ 913 #define F_DUPDATA 0x04 /**< data has duplicates */ 914 915 /** valid flags for #mdb_node_add() */ 916 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND) 917 918 /** @} */ 919 unsigned short mn_flags; /**< @ref mdb_node */ 920 unsigned short mn_ksize; /**< key size */ 921 char mn_data[1]; /**< key and data are appended here */ 922 } MDB_node; 923 924 /** Size of the node header, excluding dynamic data at the end */ 925 #define NODESIZE offsetof(MDB_node, mn_data) 926 927 /** Bit position of top word in page number, for shifting mn_flags */ 928 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0) 929 930 /** Size of a node in a branch page with a given key. 931 * This is just the node header plus the key, there is no data. 932 */ 933 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size)) 934 935 /** Size of a node in a leaf page with a given key and data. 936 * This is node header plus key plus data size. 937 */ 938 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size) 939 940 /** Address of node \b i in page \b p */ 941 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE)) 942 943 /** Address of the key for the node */ 944 #define NODEKEY(node) (void *)((node)->mn_data) 945 946 /** Address of the data for a node */ 947 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize) 948 949 /** Get the page number pointed to by a branch node */ 950 #define NODEPGNO(node) \ 951 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \ 952 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0)) 953 /** Set the page number in a branch node */ 954 #define SETPGNO(node,pgno) do { \ 955 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \ 956 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0) 957 958 /** Get the size of the data in a leaf node */ 959 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16)) 960 /** Set the size of the data for a leaf node */ 961 #define SETDSZ(node,size) do { \ 962 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0) 963 /** The size of a key in a node */ 964 #define NODEKSZ(node) ((node)->mn_ksize) 965 966 /** Copy a page number from src to dst */ 967 #ifdef MISALIGNED_OK 968 #define COPY_PGNO(dst,src) dst = src 969 #else 970 #if SIZE_MAX > 4294967295UL 971 #define COPY_PGNO(dst,src) do { \ 972 unsigned short *s, *d; \ 973 s = (unsigned short *)&(src); \ 974 d = (unsigned short *)&(dst); \ 975 *d++ = *s++; \ 976 *d++ = *s++; \ 977 *d++ = *s++; \ 978 *d = *s; \ 979 } while (0) 980 #else 981 #define COPY_PGNO(dst,src) do { \ 982 unsigned short *s, *d; \ 983 s = (unsigned short *)&(src); \ 984 d = (unsigned short *)&(dst); \ 985 *d++ = *s++; \ 986 *d = *s; \ 987 } while (0) 988 #endif 989 #endif 990 /** The address of a key in a LEAF2 page. 991 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs. 992 * There are no node headers, keys are stored contiguously. 993 */ 994 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks))) 995 996 /** Set the \b node's key into \b keyptr, if requested. */ 997 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \ 998 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } } 999 1000 /** Set the \b node's key into \b key. */ 1001 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); } 1002 1003 /** Information about a single database in the environment. */ 1004 typedef struct MDB_db { 1005 uint32_t md_pad; /**< also ksize for LEAF2 pages */ 1006 uint16_t md_flags; /**< @ref mdb_dbi_open */ 1007 uint16_t md_depth; /**< depth of this tree */ 1008 pgno_t md_branch_pages; /**< number of internal pages */ 1009 pgno_t md_leaf_pages; /**< number of leaf pages */ 1010 pgno_t md_overflow_pages; /**< number of overflow pages */ 1011 size_t md_entries; /**< number of data items */ 1012 pgno_t md_root; /**< the root page of this tree */ 1013 } MDB_db; 1014 1015 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */ 1016 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID)) 1017 /** #mdb_dbi_open() flags */ 1018 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\ 1019 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE) 1020 1021 /** Handle for the DB used to track free pages. */ 1022 #define FREE_DBI 0 1023 /** Handle for the default DB. */ 1024 #define MAIN_DBI 1 1025 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */ 1026 #define CORE_DBS 2 1027 1028 /** Number of meta pages - also hardcoded elsewhere */ 1029 #define NUM_METAS 2 1030 1031 /** Meta page content. 1032 * A meta page is the start point for accessing a database snapshot. 1033 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2). 1034 */ 1035 typedef struct MDB_meta { 1036 /** Stamp identifying this as an LMDB file. It must be set 1037 * to #MDB_MAGIC. */ 1038 uint32_t mm_magic; 1039 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */ 1040 uint32_t mm_version; 1041 void *mm_address; /**< address for fixed mapping */ 1042 size_t mm_mapsize; /**< size of mmap region */ 1043 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */ 1044 /** The size of pages used in this DB */ 1045 #define mm_psize mm_dbs[FREE_DBI].md_pad 1046 /** Any persistent environment flags. @ref mdb_env */ 1047 #define mm_flags mm_dbs[FREE_DBI].md_flags 1048 /** Last used page in the datafile. 1049 * Actually the file may be shorter if the freeDB lists the final pages. 1050 */ 1051 pgno_t mm_last_pg; 1052 volatile txnid_t mm_txnid; /**< txnid that committed this page */ 1053 } MDB_meta; 1054 1055 /** Buffer for a stack-allocated meta page. 1056 * The members define size and alignment, and silence type 1057 * aliasing warnings. They are not used directly; that could 1058 * mean incorrectly using several union members in parallel. 1059 */ 1060 typedef union MDB_metabuf { 1061 MDB_page mb_page; 1062 struct { 1063 char mm_pad[PAGEHDRSZ]; 1064 MDB_meta mm_meta; 1065 } mb_metabuf; 1066 } MDB_metabuf; 1067 1068 /** Auxiliary DB info. 1069 * The information here is mostly static/read-only. There is 1070 * only a single copy of this record in the environment. 1071 */ 1072 typedef struct MDB_dbx { 1073 MDB_val md_name; /**< name of the database */ 1074 MDB_cmp_func *md_cmp; /**< function for comparing keys */ 1075 MDB_cmp_func *md_dcmp; /**< function for comparing data items */ 1076 MDB_rel_func *md_rel; /**< user relocate function */ 1077 void *md_relctx; /**< user-provided context for md_rel */ 1078 } MDB_dbx; 1079 1080 /** A database transaction. 1081 * Every operation requires a transaction handle. 1082 */ 1083 struct MDB_txn { 1084 MDB_txn *mt_parent; /**< parent of a nested txn */ 1085 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */ 1086 MDB_txn *mt_child; 1087 pgno_t mt_next_pgno; /**< next unallocated page */ 1088 /** The ID of this transaction. IDs are integers incrementing from 1. 1089 * Only committed write transactions increment the ID. If a transaction 1090 * aborts, the ID may be re-used by the next writer. 1091 */ 1092 txnid_t mt_txnid; 1093 MDB_env *mt_env; /**< the DB environment */ 1094 /** The list of pages that became unused during this transaction. 1095 */ 1096 MDB_IDL mt_free_pgs; 1097 /** The list of loose pages that became unused and may be reused 1098 * in this transaction, linked through #NEXT_LOOSE_PAGE(page). 1099 */ 1100 MDB_page *mt_loose_pgs; 1101 /** Number of loose pages (#mt_loose_pgs) */ 1102 int mt_loose_count; 1103 /** The sorted list of dirty pages we temporarily wrote to disk 1104 * because the dirty list was full. page numbers in here are 1105 * shifted left by 1, deleted slots have the LSB set. 1106 */ 1107 MDB_IDL mt_spill_pgs; 1108 union { 1109 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */ 1110 MDB_ID2L dirty_list; 1111 /** For read txns: This thread/txn's reader table slot, or NULL. */ 1112 MDB_reader *reader; 1113 } mt_u; 1114 /** Array of records for each DB known in the environment. */ 1115 MDB_dbx *mt_dbxs; 1116 /** Array of MDB_db records for each known DB */ 1117 MDB_db *mt_dbs; 1118 /** Array of sequence numbers for each DB handle */ 1119 unsigned int *mt_dbiseqs; 1120 /** @defgroup mt_dbflag Transaction DB Flags 1121 * @ingroup internal 1122 * @{ 1123 */ 1124 #define DB_DIRTY 0x01 /**< DB was written in this txn */ 1125 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */ 1126 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */ 1127 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */ 1128 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */ 1129 #define DB_DUPDATA 0x20 /**< DB is #MDB_DUPSORT data */ 1130 /** @} */ 1131 /** In write txns, array of cursors for each DB */ 1132 MDB_cursor **mt_cursors; 1133 /** Array of flags for each DB */ 1134 unsigned char *mt_dbflags; 1135 /** Number of DB records in use, or 0 when the txn is finished. 1136 * This number only ever increments until the txn finishes; we 1137 * don't decrement it when individual DB handles are closed. 1138 */ 1139 MDB_dbi mt_numdbs; 1140 1141 /** @defgroup mdb_txn Transaction Flags 1142 * @ingroup internal 1143 * @{ 1144 */ 1145 /** #mdb_txn_begin() flags */ 1146 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY 1147 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */ 1148 /* internal txn flags */ 1149 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */ 1150 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */ 1151 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */ 1152 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */ 1153 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */ 1154 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */ 1155 /** most operations on the txn are currently illegal */ 1156 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD) 1157 /** @} */ 1158 unsigned int mt_flags; /**< @ref mdb_txn */ 1159 /** #dirty_list room: Array size - \#dirty pages visible to this txn. 1160 * Includes ancestor txns' dirty pages not hidden by other txns' 1161 * dirty/spilled pages. Thus commit(nested txn) has room to merge 1162 * dirty_list into mt_parent after freeing hidden mt_parent pages. 1163 */ 1164 unsigned int mt_dirty_room; 1165 }; 1166 1167 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty. 1168 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to 1169 * raise this on a 64 bit machine. 1170 */ 1171 #define CURSOR_STACK 32 1172 1173 struct MDB_xcursor; 1174 1175 /** Cursors are used for all DB operations. 1176 * A cursor holds a path of (page pointer, key index) from the DB 1177 * root to a position in the DB, plus other state. #MDB_DUPSORT 1178 * cursors include an xcursor to the current data item. Write txns 1179 * track their cursors and keep them up to date when data moves. 1180 * Exception: An xcursor's pointer to a #P_SUBP page can be stale. 1181 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage). 1182 */ 1183 struct MDB_cursor { 1184 /** Next cursor on this DB in this txn */ 1185 MDB_cursor *mc_next; 1186 /** Backup of the original cursor if this cursor is a shadow */ 1187 MDB_cursor *mc_backup; 1188 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */ 1189 struct MDB_xcursor *mc_xcursor; 1190 /** The transaction that owns this cursor */ 1191 MDB_txn *mc_txn; 1192 /** The database handle this cursor operates on */ 1193 MDB_dbi mc_dbi; 1194 /** The database record for this cursor */ 1195 MDB_db *mc_db; 1196 /** The database auxiliary record for this cursor */ 1197 MDB_dbx *mc_dbx; 1198 /** The @ref mt_dbflag for this database */ 1199 unsigned char *mc_dbflag; 1200 unsigned short mc_snum; /**< number of pushed pages */ 1201 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */ 1202 /** @defgroup mdb_cursor Cursor Flags 1203 * @ingroup internal 1204 * Cursor state flags. 1205 * @{ 1206 */ 1207 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */ 1208 #define C_EOF 0x02 /**< No more data */ 1209 #define C_SUB 0x04 /**< Cursor is a sub-cursor */ 1210 #define C_DEL 0x08 /**< last op was a cursor_del */ 1211 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */ 1212 /** @} */ 1213 unsigned int mc_flags; /**< @ref mdb_cursor */ 1214 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */ 1215 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */ 1216 }; 1217 1218 /** Context for sorted-dup records. 1219 * We could have gone to a fully recursive design, with arbitrarily 1220 * deep nesting of sub-databases. But for now we only handle these 1221 * levels - main DB, optional sub-DB, sorted-duplicate DB. 1222 */ 1223 typedef struct MDB_xcursor { 1224 /** A sub-cursor for traversing the Dup DB */ 1225 MDB_cursor mx_cursor; 1226 /** The database record for this Dup DB */ 1227 MDB_db mx_db; 1228 /** The auxiliary DB record for this Dup DB */ 1229 MDB_dbx mx_dbx; 1230 /** The @ref mt_dbflag for this Dup DB */ 1231 unsigned char mx_dbflag; 1232 } MDB_xcursor; 1233 1234 /** Check if there is an inited xcursor */ 1235 #define XCURSOR_INITED(mc) \ 1236 ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) 1237 1238 /** Update the xcursor's sub-page pointer, if any, in \b mc. Needed 1239 * when the node which contains the sub-page may have moved. Called 1240 * with leaf page \b mp = mc->mc_pg[\b top]. 1241 */ 1242 #define XCURSOR_REFRESH(mc, top, mp) do { \ 1243 MDB_page *xr_pg = (mp); \ 1244 MDB_node *xr_node; \ 1245 if (!XCURSOR_INITED(mc) || (mc)->mc_ki[top] >= NUMKEYS(xr_pg)) break; \ 1246 xr_node = NODEPTR(xr_pg, (mc)->mc_ki[top]); \ 1247 if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \ 1248 (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \ 1249 } while (0) 1250 1251 /** State of FreeDB old pages, stored in the MDB_env */ 1252 typedef struct MDB_pgstate { 1253 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */ 1254 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */ 1255 } MDB_pgstate; 1256 1257 /** The database environment. */ 1258 struct MDB_env { 1259 HANDLE me_fd; /**< The main data file */ 1260 HANDLE me_lfd; /**< The lock file */ 1261 HANDLE me_mfd; /**< For writing and syncing the meta pages */ 1262 /** Failed to update the meta page. Probably an I/O error. */ 1263 #define MDB_FATAL_ERROR 0x80000000U 1264 /** Some fields are initialized. */ 1265 #define MDB_ENV_ACTIVE 0x20000000U 1266 /** me_txkey is set */ 1267 #define MDB_ENV_TXKEY 0x10000000U 1268 /** fdatasync is unreliable */ 1269 #define MDB_FSYNCONLY 0x08000000U 1270 uint32_t me_flags; /**< @ref mdb_env */ 1271 unsigned int me_psize; /**< DB page size, inited from me_os_psize */ 1272 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */ 1273 unsigned int me_maxreaders; /**< size of the reader table */ 1274 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */ 1275 volatile int me_close_readers; 1276 MDB_dbi me_numdbs; /**< number of DBs opened */ 1277 MDB_dbi me_maxdbs; /**< size of the DB table */ 1278 MDB_PID_T me_pid; /**< process ID of this env */ 1279 char *me_path; /**< path to the DB files */ 1280 char *me_map; /**< the memory map of the data file */ 1281 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */ 1282 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */ 1283 void *me_pbuf; /**< scratch area for DUPSORT put() */ 1284 MDB_txn *me_txn; /**< current write transaction */ 1285 MDB_txn *me_txn0; /**< prealloc'd write transaction */ 1286 size_t me_mapsize; /**< size of the data memory map */ 1287 off_t me_size; /**< current file size */ 1288 pgno_t me_maxpg; /**< me_mapsize / me_psize */ 1289 MDB_dbx *me_dbxs; /**< array of static DB info */ 1290 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */ 1291 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */ 1292 pthread_key_t me_txkey; /**< thread-key for readers */ 1293 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */ 1294 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */ 1295 # define me_pglast me_pgstate.mf_pglast 1296 # define me_pghead me_pgstate.mf_pghead 1297 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */ 1298 /** IDL of pages that became unused in a write txn */ 1299 MDB_IDL me_free_pgs; 1300 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */ 1301 MDB_ID2L me_dirty_list; 1302 /** Max number of freelist items that can fit in a single overflow page */ 1303 int me_maxfree_1pg; 1304 /** Max size of a node on a page */ 1305 unsigned int me_nodemax; 1306 #if !(MDB_MAXKEYSIZE) 1307 unsigned int me_maxkey; /**< max size of a key */ 1308 #endif 1309 int me_live_reader; /**< have liveness lock in reader table */ 1310 #ifdef _WIN32 1311 int me_pidquery; /**< Used in OpenProcess */ 1312 #endif 1313 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */ 1314 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */ 1315 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */ 1316 #else 1317 mdb_mutex_t me_rmutex; 1318 mdb_mutex_t me_wmutex; 1319 #endif 1320 void *me_userctx; /**< User-settable context */ 1321 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */ 1322 }; 1323 1324 /** Nested transaction */ 1325 typedef struct MDB_ntxn { 1326 MDB_txn mnt_txn; /**< the transaction */ 1327 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */ 1328 } MDB_ntxn; 1329 1330 /** max number of pages to commit in one writev() call */ 1331 #define MDB_COMMIT_PAGES 64 1332 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES 1333 #undef MDB_COMMIT_PAGES 1334 #define MDB_COMMIT_PAGES IOV_MAX 1335 #endif 1336 1337 /** max bytes to write in one call */ 1338 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4)) 1339 1340 /** Check \b txn and \b dbi arguments to a function */ 1341 #define TXN_DBI_EXIST(txn, dbi, validity) \ 1342 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity))) 1343 1344 /** Check for misused \b dbi handles */ 1345 #define TXN_DBI_CHANGED(txn, dbi) \ 1346 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi]) 1347 1348 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp); 1349 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp); 1350 static int mdb_page_touch(MDB_cursor *mc); 1351 1352 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \ 1353 "reset-tmp", "fail-begin", "fail-beginchild"} 1354 enum { 1355 /* mdb_txn_end operation number, for logging */ 1356 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET, 1357 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD 1358 }; 1359 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */ 1360 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */ 1361 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */ 1362 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */ 1363 static void mdb_txn_end(MDB_txn *txn, unsigned mode); 1364 1365 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl); 1366 static int mdb_page_search_root(MDB_cursor *mc, 1367 MDB_val *key, int modify); 1368 #define MDB_PS_MODIFY 1 1369 #define MDB_PS_ROOTONLY 2 1370 #define MDB_PS_FIRST 4 1371 #define MDB_PS_LAST 8 1372 static int mdb_page_search(MDB_cursor *mc, 1373 MDB_val *key, int flags); 1374 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst); 1375 1376 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */ 1377 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, 1378 pgno_t newpgno, unsigned int nflags); 1379 1380 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta); 1381 static MDB_meta *mdb_env_pick_meta(const MDB_env *env); 1382 static int mdb_env_write_meta(MDB_txn *txn); 1383 #if defined(MDB_USE_POSIX_MUTEX) && !defined(MDB_ROBUST_SUPPORTED) /* Drop unused excl arg */ 1384 # define mdb_env_close0(env, excl) mdb_env_close1(env) 1385 #endif 1386 static void mdb_env_close0(MDB_env *env, int excl); 1387 1388 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp); 1389 static int mdb_node_add(MDB_cursor *mc, indx_t indx, 1390 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags); 1391 static void mdb_node_del(MDB_cursor *mc, int ksize); 1392 static void mdb_node_shrink(MDB_page *mp, indx_t indx); 1393 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft); 1394 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data); 1395 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data); 1396 static size_t mdb_branch_size(MDB_env *env, MDB_val *key); 1397 1398 static int mdb_rebalance(MDB_cursor *mc); 1399 static int mdb_update_key(MDB_cursor *mc, MDB_val *key); 1400 1401 static void mdb_cursor_pop(MDB_cursor *mc); 1402 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp); 1403 1404 static int mdb_cursor_del0(MDB_cursor *mc); 1405 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags); 1406 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right); 1407 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op); 1408 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op); 1409 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op, 1410 int *exactp); 1411 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data); 1412 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data); 1413 1414 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx); 1415 static void mdb_xcursor_init0(MDB_cursor *mc); 1416 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node); 1417 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force); 1418 1419 static int mdb_drop0(MDB_cursor *mc, int subs); 1420 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi); 1421 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead); 1422 1423 /** @cond */ 1424 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long; 1425 /** @endcond */ 1426 1427 /** Compare two items pointing at size_t's of unknown alignment. */ 1428 #ifdef MISALIGNED_OK 1429 # define mdb_cmp_clong mdb_cmp_long 1430 #else 1431 # define mdb_cmp_clong mdb_cmp_cint 1432 #endif 1433 1434 #ifdef _WIN32 1435 static SECURITY_DESCRIPTOR mdb_null_sd; 1436 static SECURITY_ATTRIBUTES mdb_all_sa; 1437 static int mdb_sec_inited; 1438 1439 struct MDB_name; 1440 static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra); 1441 #endif 1442 1443 /** Return the library version info. */ 1444 char * ESECT 1445 mdb_version(int *major, int *minor, int *patch) 1446 { 1447 if (major) *major = MDB_VERSION_MAJOR; 1448 if (minor) *minor = MDB_VERSION_MINOR; 1449 if (patch) *patch = MDB_VERSION_PATCH; 1450 return MDB_VERSION_STRING; 1451 } 1452 1453 /** Table of descriptions for LMDB @ref errors */ 1454 static char *const mdb_errstr[] = { 1455 "MDB_KEYEXIST: Key/data pair already exists", 1456 "MDB_NOTFOUND: No matching key/data pair found", 1457 "MDB_PAGE_NOTFOUND: Requested page not found", 1458 "MDB_CORRUPTED: Located page was wrong type", 1459 "MDB_PANIC: Update of meta page failed or environment had fatal error", 1460 "MDB_VERSION_MISMATCH: Database environment version mismatch", 1461 "MDB_INVALID: File is not an LMDB file", 1462 "MDB_MAP_FULL: Environment mapsize limit reached", 1463 "MDB_DBS_FULL: Environment maxdbs limit reached", 1464 "MDB_READERS_FULL: Environment maxreaders limit reached", 1465 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open", 1466 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big", 1467 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached", 1468 "MDB_PAGE_FULL: Internal error - page has no more space", 1469 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize", 1470 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed", 1471 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot", 1472 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid", 1473 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size", 1474 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly", 1475 }; 1476 1477 char * 1478 mdb_strerror(int err) 1479 { 1480 #ifdef _WIN32 1481 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf. 1482 * This works as long as no function between the call to mdb_strerror 1483 * and the actual use of the message uses more than 4K of stack. 1484 */ 1485 #define MSGSIZE 1024 1486 #define PADSIZE 4096 1487 char buf[MSGSIZE+PADSIZE], *ptr = buf; 1488 #endif 1489 int i; 1490 if (!err) 1491 return ("Successful return: 0"); 1492 1493 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) { 1494 i = err - MDB_KEYEXIST; 1495 return mdb_errstr[i]; 1496 } 1497 1498 #ifdef _WIN32 1499 /* These are the C-runtime error codes we use. The comment indicates 1500 * their numeric value, and the Win32 error they would correspond to 1501 * if the error actually came from a Win32 API. A major mess, we should 1502 * have used LMDB-specific error codes for everything. 1503 */ 1504 switch(err) { 1505 case ENOENT: /* 2, FILE_NOT_FOUND */ 1506 case EIO: /* 5, ACCESS_DENIED */ 1507 case ENOMEM: /* 12, INVALID_ACCESS */ 1508 case EACCES: /* 13, INVALID_DATA */ 1509 case EBUSY: /* 16, CURRENT_DIRECTORY */ 1510 case EINVAL: /* 22, BAD_COMMAND */ 1511 case ENOSPC: /* 28, OUT_OF_PAPER */ 1512 return strerror(err); 1513 default: 1514 ; 1515 } 1516 buf[0] = 0; 1517 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM | 1518 FORMAT_MESSAGE_IGNORE_INSERTS, 1519 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE); 1520 return ptr; 1521 #else 1522 return strerror(err); 1523 #endif 1524 } 1525 1526 /** assert(3) variant in cursor context */ 1527 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr) 1528 /** assert(3) variant in transaction context */ 1529 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr) 1530 /** assert(3) variant in environment context */ 1531 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr) 1532 1533 #ifndef NDEBUG 1534 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \ 1535 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__)) 1536 1537 static void ESECT 1538 mdb_assert_fail(MDB_env *env, const char *expr_txt, 1539 const char *func, const char *file, int line) 1540 { 1541 char buf[400]; 1542 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()", 1543 file, line, expr_txt, func); 1544 if (env->me_assert_func) 1545 env->me_assert_func(env, buf); 1546 fprintf(stderr, "%s\n", buf); 1547 abort(); 1548 } 1549 #else 1550 # define mdb_assert0(env, expr, expr_txt) ((void) 0) 1551 #endif /* NDEBUG */ 1552 1553 #if MDB_DEBUG 1554 /** Return the page number of \b mp which may be sub-page, for debug output */ 1555 static pgno_t 1556 mdb_dbg_pgno(MDB_page *mp) 1557 { 1558 pgno_t ret; 1559 COPY_PGNO(ret, mp->mp_pgno); 1560 return ret; 1561 } 1562 1563 /** Display a key in hexadecimal and return the address of the result. 1564 * @param[in] key the key to display 1565 * @param[in] buf the buffer to write into. Should always be #DKBUF. 1566 * @return The key in hexadecimal form. 1567 */ 1568 char * 1569 mdb_dkey(MDB_val *key, char *buf) 1570 { 1571 char *ptr = buf; 1572 unsigned char *c = key->mv_data; 1573 unsigned int i; 1574 1575 if (!key) 1576 return ""; 1577 1578 if (key->mv_size > DKBUF_MAXKEYSIZE) 1579 return "MDB_MAXKEYSIZE"; 1580 /* may want to make this a dynamic check: if the key is mostly 1581 * printable characters, print it as-is instead of converting to hex. 1582 */ 1583 #if 1 1584 buf[0] = '\0'; 1585 for (i=0; i<key->mv_size; i++) 1586 ptr += sprintf(ptr, "%02x", *c++); 1587 #else 1588 sprintf(buf, "%.*s", key->mv_size, key->mv_data); 1589 #endif 1590 return buf; 1591 } 1592 1593 static const char * 1594 mdb_leafnode_type(MDB_node *n) 1595 { 1596 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}}; 1597 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" : 1598 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)]; 1599 } 1600 1601 /** Display all the keys in the page. */ 1602 void 1603 mdb_page_list(MDB_page *mp) 1604 { 1605 pgno_t pgno = mdb_dbg_pgno(mp); 1606 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : ""; 1607 MDB_node *node; 1608 unsigned int i, nkeys, nsize, total = 0; 1609 MDB_val key; 1610 DKBUF; 1611 1612 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) { 1613 case P_BRANCH: type = "Branch page"; break; 1614 case P_LEAF: type = "Leaf page"; break; 1615 case P_LEAF|P_SUBP: type = "Sub-page"; break; 1616 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break; 1617 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break; 1618 case P_OVERFLOW: 1619 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n", 1620 pgno, mp->mp_pages, state); 1621 return; 1622 case P_META: 1623 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n", 1624 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid); 1625 return; 1626 default: 1627 fprintf(stderr, "Bad page %"Z"u flags 0x%X\n", pgno, mp->mp_flags); 1628 return; 1629 } 1630 1631 nkeys = NUMKEYS(mp); 1632 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state); 1633 1634 for (i=0; i<nkeys; i++) { 1635 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */ 1636 key.mv_size = nsize = mp->mp_pad; 1637 key.mv_data = LEAF2KEY(mp, i, nsize); 1638 total += nsize; 1639 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key)); 1640 continue; 1641 } 1642 node = NODEPTR(mp, i); 1643 key.mv_size = node->mn_ksize; 1644 key.mv_data = node->mn_data; 1645 nsize = NODESIZE + key.mv_size; 1646 if (IS_BRANCH(mp)) { 1647 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node), 1648 DKEY(&key)); 1649 total += nsize; 1650 } else { 1651 if (F_ISSET(node->mn_flags, F_BIGDATA)) 1652 nsize += sizeof(pgno_t); 1653 else 1654 nsize += NODEDSZ(node); 1655 total += nsize; 1656 nsize += sizeof(indx_t); 1657 fprintf(stderr, "key %d: nsize %d, %s%s\n", 1658 i, nsize, DKEY(&key), mdb_leafnode_type(node)); 1659 } 1660 total = EVEN(total); 1661 } 1662 fprintf(stderr, "Total: header %d + contents %d + unused %d\n", 1663 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp)); 1664 } 1665 1666 void 1667 mdb_cursor_chk(MDB_cursor *mc) 1668 { 1669 unsigned int i; 1670 MDB_node *node; 1671 MDB_page *mp; 1672 1673 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return; 1674 for (i=0; i<mc->mc_top; i++) { 1675 mp = mc->mc_pg[i]; 1676 node = NODEPTR(mp, mc->mc_ki[i]); 1677 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno) 1678 printf("oops!\n"); 1679 } 1680 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i])) 1681 printf("ack!\n"); 1682 if (XCURSOR_INITED(mc)) { 1683 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 1684 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) && 1685 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) { 1686 printf("blah!\n"); 1687 } 1688 } 1689 } 1690 #endif 1691 1692 #if (MDB_DEBUG) > 2 1693 /** Count all the pages in each DB and in the freelist 1694 * and make sure it matches the actual number of pages 1695 * being used. 1696 * All named DBs must be open for a correct count. 1697 */ 1698 static void mdb_audit(MDB_txn *txn) 1699 { 1700 MDB_cursor mc; 1701 MDB_val key, data; 1702 MDB_ID freecount, count; 1703 MDB_dbi i; 1704 int rc; 1705 1706 freecount = 0; 1707 mdb_cursor_init(&mc, txn, FREE_DBI, NULL); 1708 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0) 1709 freecount += *(MDB_ID *)data.mv_data; 1710 mdb_tassert(txn, rc == MDB_NOTFOUND); 1711 1712 count = 0; 1713 for (i = 0; i<txn->mt_numdbs; i++) { 1714 MDB_xcursor mx; 1715 if (!(txn->mt_dbflags[i] & DB_VALID)) 1716 continue; 1717 mdb_cursor_init(&mc, txn, i, &mx); 1718 if (txn->mt_dbs[i].md_root == P_INVALID) 1719 continue; 1720 count += txn->mt_dbs[i].md_branch_pages + 1721 txn->mt_dbs[i].md_leaf_pages + 1722 txn->mt_dbs[i].md_overflow_pages; 1723 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) { 1724 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST); 1725 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) { 1726 unsigned j; 1727 MDB_page *mp; 1728 mp = mc.mc_pg[mc.mc_top]; 1729 for (j=0; j<NUMKEYS(mp); j++) { 1730 MDB_node *leaf = NODEPTR(mp, j); 1731 if (leaf->mn_flags & F_SUBDATA) { 1732 MDB_db db; 1733 memcpy(&db, NODEDATA(leaf), sizeof(db)); 1734 count += db.md_branch_pages + db.md_leaf_pages + 1735 db.md_overflow_pages; 1736 } 1737 } 1738 } 1739 mdb_tassert(txn, rc == MDB_NOTFOUND); 1740 } 1741 } 1742 if (freecount + count + NUM_METAS != txn->mt_next_pgno) { 1743 fprintf(stderr, "audit: %"Z"u freecount: %"Z"u count: %"Z"u total: %"Z"u next_pgno: %"Z"u\n", 1744 txn->mt_txnid, freecount, count+NUM_METAS, 1745 freecount+count+NUM_METAS, txn->mt_next_pgno); 1746 } 1747 } 1748 #endif 1749 1750 int 1751 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b) 1752 { 1753 return txn->mt_dbxs[dbi].md_cmp(a, b); 1754 } 1755 1756 int 1757 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b) 1758 { 1759 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp; 1760 #if UINT_MAX < SIZE_MAX 1761 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t)) 1762 dcmp = mdb_cmp_clong; 1763 #endif 1764 return dcmp(a, b); 1765 } 1766 1767 /** Allocate memory for a page. 1768 * Re-use old malloc'd pages first for singletons, otherwise just malloc. 1769 * Set #MDB_TXN_ERROR on failure. 1770 */ 1771 static MDB_page * 1772 mdb_page_malloc(MDB_txn *txn, unsigned num) 1773 { 1774 MDB_env *env = txn->mt_env; 1775 MDB_page *ret = env->me_dpages; 1776 size_t psize = env->me_psize, sz = psize, off; 1777 /* For ! #MDB_NOMEMINIT, psize counts how much to init. 1778 * For a single page alloc, we init everything after the page header. 1779 * For multi-page, we init the final page; if the caller needed that 1780 * many pages they will be filling in at least up to the last page. 1781 */ 1782 if (num == 1) { 1783 if (ret) { 1784 VGMEMP_ALLOC(env, ret, sz); 1785 VGMEMP_DEFINED(ret, sizeof(ret->mp_next)); 1786 env->me_dpages = ret->mp_next; 1787 return ret; 1788 } 1789 psize -= off = PAGEHDRSZ; 1790 } else { 1791 sz *= num; 1792 off = sz - psize; 1793 } 1794 if ((ret = malloc(sz)) != NULL) { 1795 VGMEMP_ALLOC(env, ret, sz); 1796 if (!(env->me_flags & MDB_NOMEMINIT)) { 1797 memset((char *)ret + off, 0, psize); 1798 ret->mp_pad = 0; 1799 } 1800 } else { 1801 txn->mt_flags |= MDB_TXN_ERROR; 1802 } 1803 return ret; 1804 } 1805 /** Free a single page. 1806 * Saves single pages to a list, for future reuse. 1807 * (This is not used for multi-page overflow pages.) 1808 */ 1809 static void 1810 mdb_page_free(MDB_env *env, MDB_page *mp) 1811 { 1812 mp->mp_next = env->me_dpages; 1813 VGMEMP_FREE(env, mp); 1814 env->me_dpages = mp; 1815 } 1816 1817 /** Free a dirty page */ 1818 static void 1819 mdb_dpage_free(MDB_env *env, MDB_page *dp) 1820 { 1821 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) { 1822 mdb_page_free(env, dp); 1823 } else { 1824 /* large pages just get freed directly */ 1825 VGMEMP_FREE(env, dp); 1826 free(dp); 1827 } 1828 } 1829 1830 /** Return all dirty pages to dpage list */ 1831 static void 1832 mdb_dlist_free(MDB_txn *txn) 1833 { 1834 MDB_env *env = txn->mt_env; 1835 MDB_ID2L dl = txn->mt_u.dirty_list; 1836 unsigned i, n = dl[0].mid; 1837 1838 for (i = 1; i <= n; i++) { 1839 mdb_dpage_free(env, dl[i].mptr); 1840 } 1841 dl[0].mid = 0; 1842 } 1843 1844 /** Loosen or free a single page. 1845 * Saves single pages to a list for future reuse 1846 * in this same txn. It has been pulled from the freeDB 1847 * and already resides on the dirty list, but has been 1848 * deleted. Use these pages first before pulling again 1849 * from the freeDB. 1850 * 1851 * If the page wasn't dirtied in this txn, just add it 1852 * to this txn's free list. 1853 */ 1854 static int 1855 mdb_page_loose(MDB_cursor *mc, MDB_page *mp) 1856 { 1857 int loose = 0; 1858 pgno_t pgno = mp->mp_pgno; 1859 MDB_txn *txn = mc->mc_txn; 1860 1861 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) { 1862 if (txn->mt_parent) { 1863 MDB_ID2 *dl = txn->mt_u.dirty_list; 1864 /* If txn has a parent, make sure the page is in our 1865 * dirty list. 1866 */ 1867 if (dl[0].mid) { 1868 unsigned x = mdb_mid2l_search(dl, pgno); 1869 if (x <= dl[0].mid && dl[x].mid == pgno) { 1870 if (mp != dl[x].mptr) { /* bad cursor? */ 1871 mc->mc_flags &= ~(C_INITIALIZED|C_EOF); 1872 txn->mt_flags |= MDB_TXN_ERROR; 1873 return MDB_CORRUPTED; 1874 } 1875 /* ok, it's ours */ 1876 loose = 1; 1877 } 1878 } 1879 } else { 1880 /* no parent txn, so it's just ours */ 1881 loose = 1; 1882 } 1883 } 1884 if (loose) { 1885 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc), 1886 mp->mp_pgno)); 1887 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs; 1888 txn->mt_loose_pgs = mp; 1889 txn->mt_loose_count++; 1890 mp->mp_flags |= P_LOOSE; 1891 } else { 1892 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno); 1893 if (rc) 1894 return rc; 1895 } 1896 1897 return MDB_SUCCESS; 1898 } 1899 1900 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn. 1901 * @param[in] mc A cursor handle for the current operation. 1902 * @param[in] pflags Flags of the pages to update: 1903 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it. 1904 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush(). 1905 * @return 0 on success, non-zero on failure. 1906 */ 1907 static int 1908 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all) 1909 { 1910 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP }; 1911 MDB_txn *txn = mc->mc_txn; 1912 MDB_cursor *m3, *m0 = mc; 1913 MDB_xcursor *mx; 1914 MDB_page *dp, *mp; 1915 MDB_node *leaf; 1916 unsigned i, j; 1917 int rc = MDB_SUCCESS, level; 1918 1919 /* Mark pages seen by cursors */ 1920 if (mc->mc_flags & C_UNTRACK) 1921 mc = NULL; /* will find mc in mt_cursors */ 1922 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) { 1923 for (; mc; mc=mc->mc_next) { 1924 if (!(mc->mc_flags & C_INITIALIZED)) 1925 continue; 1926 for (m3 = mc;; m3 = &mx->mx_cursor) { 1927 mp = NULL; 1928 for (j=0; j<m3->mc_snum; j++) { 1929 mp = m3->mc_pg[j]; 1930 if ((mp->mp_flags & Mask) == pflags) 1931 mp->mp_flags ^= P_KEEP; 1932 } 1933 mx = m3->mc_xcursor; 1934 /* Proceed to mx if it is at a sub-database */ 1935 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED))) 1936 break; 1937 if (! (mp && (mp->mp_flags & P_LEAF))) 1938 break; 1939 leaf = NODEPTR(mp, m3->mc_ki[j-1]); 1940 if (!(leaf->mn_flags & F_SUBDATA)) 1941 break; 1942 } 1943 } 1944 if (i == 0) 1945 break; 1946 } 1947 1948 if (all) { 1949 /* Mark dirty root pages */ 1950 for (i=0; i<txn->mt_numdbs; i++) { 1951 if (txn->mt_dbflags[i] & DB_DIRTY) { 1952 pgno_t pgno = txn->mt_dbs[i].md_root; 1953 if (pgno == P_INVALID) 1954 continue; 1955 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS) 1956 break; 1957 if ((dp->mp_flags & Mask) == pflags && level <= 1) 1958 dp->mp_flags ^= P_KEEP; 1959 } 1960 } 1961 } 1962 1963 return rc; 1964 } 1965 1966 static int mdb_page_flush(MDB_txn *txn, int keep); 1967 1968 /** Spill pages from the dirty list back to disk. 1969 * This is intended to prevent running into #MDB_TXN_FULL situations, 1970 * but note that they may still occur in a few cases: 1971 * 1) our estimate of the txn size could be too small. Currently this 1972 * seems unlikely, except with a large number of #MDB_MULTIPLE items. 1973 * 2) child txns may run out of space if their parents dirtied a 1974 * lot of pages and never spilled them. TODO: we probably should do 1975 * a preemptive spill during #mdb_txn_begin() of a child txn, if 1976 * the parent's dirty_room is below a given threshold. 1977 * 1978 * Otherwise, if not using nested txns, it is expected that apps will 1979 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk 1980 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared. 1981 * If the txn never references them again, they can be left alone. 1982 * If the txn only reads them, they can be used without any fuss. 1983 * If the txn writes them again, they can be dirtied immediately without 1984 * going thru all of the work of #mdb_page_touch(). Such references are 1985 * handled by #mdb_page_unspill(). 1986 * 1987 * Also note, we never spill DB root pages, nor pages of active cursors, 1988 * because we'll need these back again soon anyway. And in nested txns, 1989 * we can't spill a page in a child txn if it was already spilled in a 1990 * parent txn. That would alter the parent txns' data even though 1991 * the child hasn't committed yet, and we'd have no way to undo it if 1992 * the child aborted. 1993 * 1994 * @param[in] m0 cursor A cursor handle identifying the transaction and 1995 * database for which we are checking space. 1996 * @param[in] key For a put operation, the key being stored. 1997 * @param[in] data For a put operation, the data being stored. 1998 * @return 0 on success, non-zero on failure. 1999 */ 2000 static int 2001 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data) 2002 { 2003 MDB_txn *txn = m0->mc_txn; 2004 MDB_page *dp; 2005 MDB_ID2L dl = txn->mt_u.dirty_list; 2006 unsigned int i, j, need; 2007 int rc; 2008 2009 if (m0->mc_flags & C_SUB) 2010 return MDB_SUCCESS; 2011 2012 /* Estimate how much space this op will take */ 2013 i = m0->mc_db->md_depth; 2014 /* Named DBs also dirty the main DB */ 2015 if (m0->mc_dbi >= CORE_DBS) 2016 i += txn->mt_dbs[MAIN_DBI].md_depth; 2017 /* For puts, roughly factor in the key+data size */ 2018 if (key) 2019 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize; 2020 i += i; /* double it for good measure */ 2021 need = i; 2022 2023 if (txn->mt_dirty_room > i) 2024 return MDB_SUCCESS; 2025 2026 if (!txn->mt_spill_pgs) { 2027 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX); 2028 if (!txn->mt_spill_pgs) 2029 return ENOMEM; 2030 } else { 2031 /* purge deleted slots */ 2032 MDB_IDL sl = txn->mt_spill_pgs; 2033 unsigned int num = sl[0]; 2034 j=0; 2035 for (i=1; i<=num; i++) { 2036 if (!(sl[i] & 1)) 2037 sl[++j] = sl[i]; 2038 } 2039 sl[0] = j; 2040 } 2041 2042 /* Preserve pages which may soon be dirtied again */ 2043 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS) 2044 goto done; 2045 2046 /* Less aggressive spill - we originally spilled the entire dirty list, 2047 * with a few exceptions for cursor pages and DB root pages. But this 2048 * turns out to be a lot of wasted effort because in a large txn many 2049 * of those pages will need to be used again. So now we spill only 1/8th 2050 * of the dirty pages. Testing revealed this to be a good tradeoff, 2051 * better than 1/2, 1/4, or 1/10. 2052 */ 2053 if (need < MDB_IDL_UM_MAX / 8) 2054 need = MDB_IDL_UM_MAX / 8; 2055 2056 /* Save the page IDs of all the pages we're flushing */ 2057 /* flush from the tail forward, this saves a lot of shifting later on. */ 2058 for (i=dl[0].mid; i && need; i--) { 2059 MDB_ID pn = dl[i].mid << 1; 2060 dp = dl[i].mptr; 2061 if (dp->mp_flags & (P_LOOSE|P_KEEP)) 2062 continue; 2063 /* Can't spill twice, make sure it's not already in a parent's 2064 * spill list. 2065 */ 2066 if (txn->mt_parent) { 2067 MDB_txn *tx2; 2068 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) { 2069 if (tx2->mt_spill_pgs) { 2070 j = mdb_midl_search(tx2->mt_spill_pgs, pn); 2071 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) { 2072 dp->mp_flags |= P_KEEP; 2073 break; 2074 } 2075 } 2076 } 2077 if (tx2) 2078 continue; 2079 } 2080 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn))) 2081 goto done; 2082 need--; 2083 } 2084 mdb_midl_sort(txn->mt_spill_pgs); 2085 2086 /* Flush the spilled part of dirty list */ 2087 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS) 2088 goto done; 2089 2090 /* Reset any dirty pages we kept that page_flush didn't see */ 2091 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i); 2092 2093 done: 2094 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS; 2095 return rc; 2096 } 2097 2098 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */ 2099 static txnid_t 2100 mdb_find_oldest(MDB_txn *txn) 2101 { 2102 int i; 2103 txnid_t mr, oldest = txn->mt_txnid - 1; 2104 if (txn->mt_env->me_txns) { 2105 MDB_reader *r = txn->mt_env->me_txns->mti_readers; 2106 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) { 2107 if (r[i].mr_pid) { 2108 mr = r[i].mr_txnid; 2109 if (oldest > mr) 2110 oldest = mr; 2111 } 2112 } 2113 } 2114 return oldest; 2115 } 2116 2117 /** Add a page to the txn's dirty list */ 2118 static void 2119 mdb_page_dirty(MDB_txn *txn, MDB_page *mp) 2120 { 2121 MDB_ID2 mid; 2122 int rc, (*insert)(MDB_ID2L, MDB_ID2 *); 2123 2124 if (txn->mt_flags & MDB_TXN_WRITEMAP) { 2125 insert = mdb_mid2l_append; 2126 } else { 2127 insert = mdb_mid2l_insert; 2128 } 2129 mid.mid = mp->mp_pgno; 2130 mid.mptr = mp; 2131 rc = insert(txn->mt_u.dirty_list, &mid); 2132 mdb_tassert(txn, rc == 0); 2133 txn->mt_dirty_room--; 2134 } 2135 2136 /** Allocate page numbers and memory for writing. Maintain me_pglast, 2137 * me_pghead and mt_next_pgno. Set #MDB_TXN_ERROR on failure. 2138 * 2139 * If there are free pages available from older transactions, they 2140 * are re-used first. Otherwise allocate a new page at mt_next_pgno. 2141 * Do not modify the freedB, just merge freeDB records into me_pghead[] 2142 * and move me_pglast to say which records were consumed. Only this 2143 * function can create me_pghead and move me_pglast/mt_next_pgno. 2144 * @param[in] mc cursor A cursor handle identifying the transaction and 2145 * database for which we are allocating. 2146 * @param[in] num the number of pages to allocate. 2147 * @param[out] mp Address of the allocated page(s). Requests for multiple pages 2148 * will always be satisfied by a single contiguous chunk of memory. 2149 * @return 0 on success, non-zero on failure. 2150 */ 2151 static int 2152 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) 2153 { 2154 #ifdef MDB_PARANOID /* Seems like we can ignore this now */ 2155 /* Get at most <Max_retries> more freeDB records once me_pghead 2156 * has enough pages. If not enough, use new pages from the map. 2157 * If <Paranoid> and mc is updating the freeDB, only get new 2158 * records if me_pghead is empty. Then the freelist cannot play 2159 * catch-up with itself by growing while trying to save it. 2160 */ 2161 enum { Paranoid = 1, Max_retries = 500 }; 2162 #else 2163 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ }; 2164 #endif 2165 int rc, retry = num * 60; 2166 MDB_txn *txn = mc->mc_txn; 2167 MDB_env *env = txn->mt_env; 2168 pgno_t pgno, *mop = env->me_pghead; 2169 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1; 2170 MDB_page *np; 2171 txnid_t oldest = 0, last; 2172 MDB_cursor_op op; 2173 MDB_cursor m2; 2174 int found_old = 0; 2175 2176 /* If there are any loose pages, just use them */ 2177 if (num == 1 && txn->mt_loose_pgs) { 2178 np = txn->mt_loose_pgs; 2179 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np); 2180 txn->mt_loose_count--; 2181 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc), 2182 np->mp_pgno)); 2183 *mp = np; 2184 return MDB_SUCCESS; 2185 } 2186 2187 *mp = NULL; 2188 2189 /* If our dirty list is already full, we can't do anything */ 2190 if (txn->mt_dirty_room == 0) { 2191 rc = MDB_TXN_FULL; 2192 goto fail; 2193 } 2194 2195 for (op = MDB_FIRST;; op = MDB_NEXT) { 2196 MDB_val key, data; 2197 MDB_node *leaf; 2198 pgno_t *idl; 2199 2200 /* Seek a big enough contiguous page range. Prefer 2201 * pages at the tail, just truncating the list. 2202 */ 2203 if (mop_len > n2) { 2204 i = mop_len; 2205 do { 2206 pgno = mop[i]; 2207 if (mop[i-n2] == pgno+n2) 2208 goto search_done; 2209 } while (--i > n2); 2210 if (--retry < 0) 2211 break; 2212 } 2213 2214 if (op == MDB_FIRST) { /* 1st iteration */ 2215 /* Prepare to fetch more and coalesce */ 2216 last = env->me_pglast; 2217 oldest = env->me_pgoldest; 2218 mdb_cursor_init(&m2, txn, FREE_DBI, NULL); 2219 if (last) { 2220 op = MDB_SET_RANGE; 2221 key.mv_data = &last; /* will look up last+1 */ 2222 key.mv_size = sizeof(last); 2223 } 2224 if (Paranoid && mc->mc_dbi == FREE_DBI) 2225 retry = -1; 2226 } 2227 if (Paranoid && retry < 0 && mop_len) 2228 break; 2229 2230 last++; 2231 /* Do not fetch more if the record will be too recent */ 2232 if (oldest <= last) { 2233 if (!found_old) { 2234 oldest = mdb_find_oldest(txn); 2235 env->me_pgoldest = oldest; 2236 found_old = 1; 2237 } 2238 if (oldest <= last) 2239 break; 2240 } 2241 rc = mdb_cursor_get(&m2, &key, NULL, op); 2242 if (rc) { 2243 if (rc == MDB_NOTFOUND) 2244 break; 2245 goto fail; 2246 } 2247 last = *(txnid_t*)key.mv_data; 2248 if (oldest <= last) { 2249 if (!found_old) { 2250 oldest = mdb_find_oldest(txn); 2251 env->me_pgoldest = oldest; 2252 found_old = 1; 2253 } 2254 if (oldest <= last) 2255 break; 2256 } 2257 np = m2.mc_pg[m2.mc_top]; 2258 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]); 2259 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS) 2260 goto fail; 2261 2262 idl = (MDB_ID *) data.mv_data; 2263 i = idl[0]; 2264 if (!mop) { 2265 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) { 2266 rc = ENOMEM; 2267 goto fail; 2268 } 2269 } else { 2270 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0) 2271 goto fail; 2272 mop = env->me_pghead; 2273 } 2274 env->me_pglast = last; 2275 #if (MDB_DEBUG) > 1 2276 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u", 2277 last, txn->mt_dbs[FREE_DBI].md_root, i)); 2278 for (j = i; j; j--) 2279 DPRINTF(("IDL %"Z"u", idl[j])); 2280 #endif 2281 /* Merge in descending sorted order */ 2282 mdb_midl_xmerge(mop, idl); 2283 mop_len = mop[0]; 2284 } 2285 2286 /* Use new pages from the map when nothing suitable in the freeDB */ 2287 i = 0; 2288 pgno = txn->mt_next_pgno; 2289 if (pgno + num >= env->me_maxpg) { 2290 DPUTS("DB size maxed out"); 2291 rc = MDB_MAP_FULL; 2292 goto fail; 2293 } 2294 2295 search_done: 2296 if (env->me_flags & MDB_WRITEMAP) { 2297 np = (MDB_page *)(env->me_map + env->me_psize * pgno); 2298 } else { 2299 if (!(np = mdb_page_malloc(txn, num))) { 2300 rc = ENOMEM; 2301 goto fail; 2302 } 2303 } 2304 if (i) { 2305 mop[0] = mop_len -= num; 2306 /* Move any stragglers down */ 2307 for (j = i-num; j < mop_len; ) 2308 mop[++j] = mop[++i]; 2309 } else { 2310 txn->mt_next_pgno = pgno + num; 2311 } 2312 np->mp_pgno = pgno; 2313 mdb_page_dirty(txn, np); 2314 *mp = np; 2315 2316 return MDB_SUCCESS; 2317 2318 fail: 2319 txn->mt_flags |= MDB_TXN_ERROR; 2320 return rc; 2321 } 2322 2323 /** Copy the used portions of a non-overflow page. 2324 * @param[in] dst page to copy into 2325 * @param[in] src page to copy from 2326 * @param[in] psize size of a page 2327 */ 2328 static void 2329 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize) 2330 { 2331 enum { Align = sizeof(pgno_t) }; 2332 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower; 2333 2334 /* If page isn't full, just copy the used portion. Adjust 2335 * alignment so memcpy may copy words instead of bytes. 2336 */ 2337 if ((unused &= -Align) && !IS_LEAF2(src)) { 2338 upper = (upper + PAGEBASE) & -Align; 2339 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align); 2340 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper), 2341 psize - upper); 2342 } else { 2343 memcpy(dst, src, psize - unused); 2344 } 2345 } 2346 2347 /** Pull a page off the txn's spill list, if present. 2348 * If a page being referenced was spilled to disk in this txn, bring 2349 * it back and make it dirty/writable again. 2350 * @param[in] txn the transaction handle. 2351 * @param[in] mp the page being referenced. It must not be dirty. 2352 * @param[out] ret the writable page, if any. ret is unchanged if 2353 * mp wasn't spilled. 2354 */ 2355 static int 2356 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret) 2357 { 2358 MDB_env *env = txn->mt_env; 2359 const MDB_txn *tx2; 2360 unsigned x; 2361 pgno_t pgno = mp->mp_pgno, pn = pgno << 1; 2362 2363 for (tx2 = txn; tx2; tx2=tx2->mt_parent) { 2364 if (!tx2->mt_spill_pgs) 2365 continue; 2366 x = mdb_midl_search(tx2->mt_spill_pgs, pn); 2367 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) { 2368 MDB_page *np; 2369 int num; 2370 if (txn->mt_dirty_room == 0) 2371 return MDB_TXN_FULL; 2372 if (IS_OVERFLOW(mp)) 2373 num = mp->mp_pages; 2374 else 2375 num = 1; 2376 if (env->me_flags & MDB_WRITEMAP) { 2377 np = mp; 2378 } else { 2379 np = mdb_page_malloc(txn, num); 2380 if (!np) 2381 return ENOMEM; 2382 if (num > 1) 2383 memcpy(np, mp, num * env->me_psize); 2384 else 2385 mdb_page_copy(np, mp, env->me_psize); 2386 } 2387 if (tx2 == txn) { 2388 /* If in current txn, this page is no longer spilled. 2389 * If it happens to be the last page, truncate the spill list. 2390 * Otherwise mark it as deleted by setting the LSB. 2391 */ 2392 if (x == txn->mt_spill_pgs[0]) 2393 txn->mt_spill_pgs[0]--; 2394 else 2395 txn->mt_spill_pgs[x] |= 1; 2396 } /* otherwise, if belonging to a parent txn, the 2397 * page remains spilled until child commits 2398 */ 2399 2400 mdb_page_dirty(txn, np); 2401 np->mp_flags |= P_DIRTY; 2402 *ret = np; 2403 break; 2404 } 2405 } 2406 return MDB_SUCCESS; 2407 } 2408 2409 /** Touch a page: make it dirty and re-insert into tree with updated pgno. 2410 * Set #MDB_TXN_ERROR on failure. 2411 * @param[in] mc cursor pointing to the page to be touched 2412 * @return 0 on success, non-zero on failure. 2413 */ 2414 static int 2415 mdb_page_touch(MDB_cursor *mc) 2416 { 2417 MDB_page *mp = mc->mc_pg[mc->mc_top], *np; 2418 MDB_txn *txn = mc->mc_txn; 2419 MDB_cursor *m2, *m3; 2420 pgno_t pgno; 2421 int rc; 2422 2423 if (!F_ISSET(mp->mp_flags, P_DIRTY)) { 2424 if (txn->mt_flags & MDB_TXN_SPILLS) { 2425 np = NULL; 2426 rc = mdb_page_unspill(txn, mp, &np); 2427 if (rc) 2428 goto fail; 2429 if (np) 2430 goto done; 2431 } 2432 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) || 2433 (rc = mdb_page_alloc(mc, 1, &np))) 2434 goto fail; 2435 pgno = np->mp_pgno; 2436 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc), 2437 mp->mp_pgno, pgno)); 2438 mdb_cassert(mc, mp->mp_pgno != pgno); 2439 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno); 2440 /* Update the parent page, if any, to point to the new page */ 2441 if (mc->mc_top) { 2442 MDB_page *parent = mc->mc_pg[mc->mc_top-1]; 2443 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]); 2444 SETPGNO(node, pgno); 2445 } else { 2446 mc->mc_db->md_root = pgno; 2447 } 2448 } else if (txn->mt_parent && !IS_SUBP(mp)) { 2449 MDB_ID2 mid, *dl = txn->mt_u.dirty_list; 2450 pgno = mp->mp_pgno; 2451 /* If txn has a parent, make sure the page is in our 2452 * dirty list. 2453 */ 2454 if (dl[0].mid) { 2455 unsigned x = mdb_mid2l_search(dl, pgno); 2456 if (x <= dl[0].mid && dl[x].mid == pgno) { 2457 if (mp != dl[x].mptr) { /* bad cursor? */ 2458 mc->mc_flags &= ~(C_INITIALIZED|C_EOF); 2459 txn->mt_flags |= MDB_TXN_ERROR; 2460 return MDB_CORRUPTED; 2461 } 2462 return 0; 2463 } 2464 } 2465 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX); 2466 /* No - copy it */ 2467 np = mdb_page_malloc(txn, 1); 2468 if (!np) 2469 return ENOMEM; 2470 mid.mid = pgno; 2471 mid.mptr = np; 2472 rc = mdb_mid2l_insert(dl, &mid); 2473 mdb_cassert(mc, rc == 0); 2474 } else { 2475 return 0; 2476 } 2477 2478 mdb_page_copy(np, mp, txn->mt_env->me_psize); 2479 np->mp_pgno = pgno; 2480 np->mp_flags |= P_DIRTY; 2481 2482 done: 2483 /* Adjust cursors pointing to mp */ 2484 mc->mc_pg[mc->mc_top] = np; 2485 m2 = txn->mt_cursors[mc->mc_dbi]; 2486 if (mc->mc_flags & C_SUB) { 2487 for (; m2; m2=m2->mc_next) { 2488 m3 = &m2->mc_xcursor->mx_cursor; 2489 if (m3->mc_snum < mc->mc_snum) continue; 2490 if (m3->mc_pg[mc->mc_top] == mp) 2491 m3->mc_pg[mc->mc_top] = np; 2492 } 2493 } else { 2494 for (; m2; m2=m2->mc_next) { 2495 if (m2->mc_snum < mc->mc_snum) continue; 2496 if (m2 == mc) continue; 2497 if (m2->mc_pg[mc->mc_top] == mp) { 2498 m2->mc_pg[mc->mc_top] = np; 2499 if (IS_LEAF(np)) 2500 XCURSOR_REFRESH(m2, mc->mc_top, np); 2501 } 2502 } 2503 } 2504 return 0; 2505 2506 fail: 2507 txn->mt_flags |= MDB_TXN_ERROR; 2508 return rc; 2509 } 2510 2511 int 2512 mdb_env_sync(MDB_env *env, int force) 2513 { 2514 int rc = 0; 2515 if (env->me_flags & MDB_RDONLY) 2516 return EACCES; 2517 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) { 2518 if (env->me_flags & MDB_WRITEMAP) { 2519 int flags = ((env->me_flags & MDB_MAPASYNC) && !force) 2520 ? MS_ASYNC : MS_SYNC; 2521 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags)) 2522 rc = ErrCode(); 2523 #ifdef _WIN32 2524 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd)) 2525 rc = ErrCode(); 2526 #endif 2527 } else { 2528 #ifdef BROKEN_FDATASYNC 2529 if (env->me_flags & MDB_FSYNCONLY) { 2530 if (fsync(env->me_fd)) 2531 rc = ErrCode(); 2532 } else 2533 #endif 2534 if (MDB_FDATASYNC(env->me_fd)) 2535 rc = ErrCode(); 2536 } 2537 } 2538 return rc; 2539 } 2540 2541 /** Back up parent txn's cursors, then grab the originals for tracking */ 2542 static int 2543 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst) 2544 { 2545 MDB_cursor *mc, *bk; 2546 MDB_xcursor *mx; 2547 size_t size; 2548 int i; 2549 2550 for (i = src->mt_numdbs; --i >= 0; ) { 2551 if ((mc = src->mt_cursors[i]) != NULL) { 2552 size = sizeof(MDB_cursor); 2553 if (mc->mc_xcursor) 2554 size += sizeof(MDB_xcursor); 2555 for (; mc; mc = bk->mc_next) { 2556 bk = malloc(size); 2557 if (!bk) 2558 return ENOMEM; 2559 *bk = *mc; 2560 mc->mc_backup = bk; 2561 mc->mc_db = &dst->mt_dbs[i]; 2562 /* Kill pointers into src to reduce abuse: The 2563 * user may not use mc until dst ends. But we need a valid 2564 * txn pointer here for cursor fixups to keep working. 2565 */ 2566 mc->mc_txn = dst; 2567 mc->mc_dbflag = &dst->mt_dbflags[i]; 2568 if ((mx = mc->mc_xcursor) != NULL) { 2569 *(MDB_xcursor *)(bk+1) = *mx; 2570 mx->mx_cursor.mc_txn = dst; 2571 } 2572 mc->mc_next = dst->mt_cursors[i]; 2573 dst->mt_cursors[i] = mc; 2574 } 2575 } 2576 } 2577 return MDB_SUCCESS; 2578 } 2579 2580 /** Close this write txn's cursors, give parent txn's cursors back to parent. 2581 * @param[in] txn the transaction handle. 2582 * @param[in] merge true to keep changes to parent cursors, false to revert. 2583 * @return 0 on success, non-zero on failure. 2584 */ 2585 static void 2586 mdb_cursors_close(MDB_txn *txn, unsigned merge) 2587 { 2588 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk; 2589 MDB_xcursor *mx; 2590 int i; 2591 2592 for (i = txn->mt_numdbs; --i >= 0; ) { 2593 for (mc = cursors[i]; mc; mc = next) { 2594 next = mc->mc_next; 2595 if ((bk = mc->mc_backup) != NULL) { 2596 if (merge) { 2597 /* Commit changes to parent txn */ 2598 mc->mc_next = bk->mc_next; 2599 mc->mc_backup = bk->mc_backup; 2600 mc->mc_txn = bk->mc_txn; 2601 mc->mc_db = bk->mc_db; 2602 mc->mc_dbflag = bk->mc_dbflag; 2603 if ((mx = mc->mc_xcursor) != NULL) 2604 mx->mx_cursor.mc_txn = bk->mc_txn; 2605 } else { 2606 /* Abort nested txn */ 2607 *mc = *bk; 2608 if ((mx = mc->mc_xcursor) != NULL) 2609 *mx = *(MDB_xcursor *)(bk+1); 2610 } 2611 mc = bk; 2612 } 2613 /* Only malloced cursors are permanently tracked. */ 2614 free(mc); 2615 } 2616 cursors[i] = NULL; 2617 } 2618 } 2619 2620 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */ 2621 enum Pidlock_op { 2622 Pidset, Pidcheck 2623 }; 2624 #else 2625 enum Pidlock_op { 2626 Pidset = F_SETLK, Pidcheck = F_GETLK 2627 }; 2628 #endif 2629 2630 /** Set or check a pid lock. Set returns 0 on success. 2631 * Check returns 0 if the process is certainly dead, nonzero if it may 2632 * be alive (the lock exists or an error happened so we do not know). 2633 * 2634 * On Windows Pidset is a no-op, we merely check for the existence 2635 * of the process with the given pid. On POSIX we use a single byte 2636 * lock on the lockfile, set at an offset equal to the pid. 2637 */ 2638 static int 2639 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid) 2640 { 2641 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */ 2642 int ret = 0; 2643 HANDLE h; 2644 if (op == Pidcheck) { 2645 h = OpenProcess(env->me_pidquery, FALSE, pid); 2646 /* No documented "no such process" code, but other program use this: */ 2647 if (!h) 2648 return ErrCode() != ERROR_INVALID_PARAMETER; 2649 /* A process exists until all handles to it close. Has it exited? */ 2650 ret = WaitForSingleObject(h, 0) != 0; 2651 CloseHandle(h); 2652 } 2653 return ret; 2654 #else 2655 for (;;) { 2656 int rc; 2657 struct flock lock_info; 2658 memset(&lock_info, 0, sizeof(lock_info)); 2659 lock_info.l_type = F_WRLCK; 2660 lock_info.l_whence = SEEK_SET; 2661 lock_info.l_start = pid; 2662 lock_info.l_len = 1; 2663 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) { 2664 if (op == F_GETLK && lock_info.l_type != F_UNLCK) 2665 rc = -1; 2666 } else if ((rc = ErrCode()) == EINTR) { 2667 continue; 2668 } 2669 return rc; 2670 } 2671 #endif 2672 } 2673 2674 /** Common code for #mdb_txn_begin() and #mdb_txn_renew(). 2675 * @param[in] txn the transaction handle to initialize 2676 * @return 0 on success, non-zero on failure. 2677 */ 2678 static int 2679 mdb_txn_renew0(MDB_txn *txn) 2680 { 2681 MDB_env *env = txn->mt_env; 2682 MDB_txninfo *ti = env->me_txns; 2683 MDB_meta *meta; 2684 unsigned int i, nr, flags = txn->mt_flags; 2685 uint16_t x; 2686 int rc, new_notls = 0; 2687 2688 if ((flags &= MDB_TXN_RDONLY) != 0) { 2689 if (!ti) { 2690 meta = mdb_env_pick_meta(env); 2691 txn->mt_txnid = meta->mm_txnid; 2692 txn->mt_u.reader = NULL; 2693 } else { 2694 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader : 2695 pthread_getspecific(env->me_txkey); 2696 if (r) { 2697 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1) 2698 return MDB_BAD_RSLOT; 2699 } else { 2700 MDB_PID_T pid = env->me_pid; 2701 MDB_THR_T tid = pthread_self(); 2702 mdb_mutexref_t rmutex = env->me_rmutex; 2703 2704 if (!env->me_live_reader) { 2705 rc = mdb_reader_pid(env, Pidset, pid); 2706 if (rc) 2707 return rc; 2708 env->me_live_reader = 1; 2709 } 2710 2711 if (LOCK_MUTEX(rc, env, rmutex)) 2712 return rc; 2713 nr = ti->mti_numreaders; 2714 for (i=0; i<nr; i++) 2715 if (ti->mti_readers[i].mr_pid == 0) 2716 break; 2717 if (i == env->me_maxreaders) { 2718 UNLOCK_MUTEX(rmutex); 2719 return MDB_READERS_FULL; 2720 } 2721 r = &ti->mti_readers[i]; 2722 /* Claim the reader slot, carefully since other code 2723 * uses the reader table un-mutexed: First reset the 2724 * slot, next publish it in mti_numreaders. After 2725 * that, it is safe for mdb_env_close() to touch it. 2726 * When it will be closed, we can finally claim it. 2727 */ 2728 r->mr_pid = 0; 2729 r->mr_txnid = (txnid_t)-1; 2730 r->mr_tid = tid; 2731 if (i == nr) 2732 ti->mti_numreaders = ++nr; 2733 env->me_close_readers = nr; 2734 r->mr_pid = pid; 2735 UNLOCK_MUTEX(rmutex); 2736 2737 new_notls = (env->me_flags & MDB_NOTLS); 2738 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) { 2739 r->mr_pid = 0; 2740 return rc; 2741 } 2742 } 2743 do /* LY: Retry on a race, ITS#7970. */ 2744 r->mr_txnid = ti->mti_txnid; 2745 while(r->mr_txnid != ti->mti_txnid); 2746 txn->mt_txnid = r->mr_txnid; 2747 txn->mt_u.reader = r; 2748 meta = env->me_metas[txn->mt_txnid & 1]; 2749 } 2750 2751 } else { 2752 /* Not yet touching txn == env->me_txn0, it may be active */ 2753 if (ti) { 2754 if (LOCK_MUTEX(rc, env, env->me_wmutex)) 2755 return rc; 2756 txn->mt_txnid = ti->mti_txnid; 2757 meta = env->me_metas[txn->mt_txnid & 1]; 2758 } else { 2759 meta = mdb_env_pick_meta(env); 2760 txn->mt_txnid = meta->mm_txnid; 2761 } 2762 txn->mt_txnid++; 2763 #if MDB_DEBUG 2764 if (txn->mt_txnid == mdb_debug_start) 2765 mdb_debug = 1; 2766 #endif 2767 txn->mt_child = NULL; 2768 txn->mt_loose_pgs = NULL; 2769 txn->mt_loose_count = 0; 2770 txn->mt_dirty_room = MDB_IDL_UM_MAX; 2771 txn->mt_u.dirty_list = env->me_dirty_list; 2772 txn->mt_u.dirty_list[0].mid = 0; 2773 txn->mt_free_pgs = env->me_free_pgs; 2774 txn->mt_free_pgs[0] = 0; 2775 txn->mt_spill_pgs = NULL; 2776 env->me_txn = txn; 2777 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int)); 2778 } 2779 2780 /* Copy the DB info and flags */ 2781 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db)); 2782 2783 /* Moved to here to avoid a data race in read TXNs */ 2784 txn->mt_next_pgno = meta->mm_last_pg+1; 2785 2786 txn->mt_flags = flags; 2787 2788 /* Setup db info */ 2789 txn->mt_numdbs = env->me_numdbs; 2790 for (i=CORE_DBS; i<txn->mt_numdbs; i++) { 2791 x = env->me_dbflags[i]; 2792 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS; 2793 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0; 2794 } 2795 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID; 2796 txn->mt_dbflags[FREE_DBI] = DB_VALID; 2797 2798 if (env->me_flags & MDB_FATAL_ERROR) { 2799 DPUTS("environment had fatal error, must shutdown!"); 2800 rc = MDB_PANIC; 2801 } else if (env->me_maxpg < txn->mt_next_pgno) { 2802 rc = MDB_MAP_RESIZED; 2803 } else { 2804 return MDB_SUCCESS; 2805 } 2806 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN); 2807 return rc; 2808 } 2809 2810 int 2811 mdb_txn_renew(MDB_txn *txn) 2812 { 2813 int rc; 2814 2815 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED)) 2816 return EINVAL; 2817 2818 rc = mdb_txn_renew0(txn); 2819 if (rc == MDB_SUCCESS) { 2820 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u", 2821 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', 2822 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root)); 2823 } 2824 return rc; 2825 } 2826 2827 int 2828 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret) 2829 { 2830 MDB_txn *txn; 2831 MDB_ntxn *ntxn; 2832 int rc, size, tsize; 2833 2834 flags &= MDB_TXN_BEGIN_FLAGS; 2835 flags |= env->me_flags & MDB_WRITEMAP; 2836 2837 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */ 2838 return EACCES; 2839 2840 if (parent) { 2841 /* Nested transactions: Max 1 child, write txns only, no writemap */ 2842 flags |= parent->mt_flags; 2843 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) { 2844 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN; 2845 } 2846 /* Child txns save MDB_pgstate and use own copy of cursors */ 2847 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1); 2848 size += tsize = sizeof(MDB_ntxn); 2849 } else if (flags & MDB_RDONLY) { 2850 size = env->me_maxdbs * (sizeof(MDB_db)+1); 2851 size += tsize = sizeof(MDB_txn); 2852 } else { 2853 /* Reuse preallocated write txn. However, do not touch it until 2854 * mdb_txn_renew0() succeeds, since it currently may be active. 2855 */ 2856 txn = env->me_txn0; 2857 goto renew; 2858 } 2859 if ((txn = calloc(1, size)) == NULL) { 2860 DPRINTF(("calloc: %s", strerror(errno))); 2861 return ENOMEM; 2862 } 2863 txn->mt_dbxs = env->me_dbxs; /* static */ 2864 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize); 2865 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs; 2866 txn->mt_flags = flags; 2867 txn->mt_env = env; 2868 2869 if (parent) { 2870 unsigned int i; 2871 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs); 2872 txn->mt_dbiseqs = parent->mt_dbiseqs; 2873 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE); 2874 if (!txn->mt_u.dirty_list || 2875 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX))) 2876 { 2877 free(txn->mt_u.dirty_list); 2878 free(txn); 2879 return ENOMEM; 2880 } 2881 txn->mt_txnid = parent->mt_txnid; 2882 txn->mt_dirty_room = parent->mt_dirty_room; 2883 txn->mt_u.dirty_list[0].mid = 0; 2884 txn->mt_spill_pgs = NULL; 2885 txn->mt_next_pgno = parent->mt_next_pgno; 2886 parent->mt_flags |= MDB_TXN_HAS_CHILD; 2887 parent->mt_child = txn; 2888 txn->mt_parent = parent; 2889 txn->mt_numdbs = parent->mt_numdbs; 2890 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db)); 2891 /* Copy parent's mt_dbflags, but clear DB_NEW */ 2892 for (i=0; i<txn->mt_numdbs; i++) 2893 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW; 2894 rc = 0; 2895 ntxn = (MDB_ntxn *)txn; 2896 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */ 2897 if (env->me_pghead) { 2898 size = MDB_IDL_SIZEOF(env->me_pghead); 2899 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]); 2900 if (env->me_pghead) 2901 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size); 2902 else 2903 rc = ENOMEM; 2904 } 2905 if (!rc) 2906 rc = mdb_cursor_shadow(parent, txn); 2907 if (rc) 2908 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD); 2909 } else { /* MDB_RDONLY */ 2910 txn->mt_dbiseqs = env->me_dbiseqs; 2911 renew: 2912 rc = mdb_txn_renew0(txn); 2913 } 2914 if (rc) { 2915 if (txn != env->me_txn0) 2916 free(txn); 2917 } else { 2918 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */ 2919 *ret = txn; 2920 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u", 2921 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w', 2922 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root)); 2923 } 2924 2925 return rc; 2926 } 2927 2928 MDB_env * 2929 mdb_txn_env(MDB_txn *txn) 2930 { 2931 if(!txn) return NULL; 2932 return txn->mt_env; 2933 } 2934 2935 size_t 2936 mdb_txn_id(MDB_txn *txn) 2937 { 2938 if(!txn) return 0; 2939 return txn->mt_txnid; 2940 } 2941 2942 /** Export or close DBI handles opened in this txn. */ 2943 static void 2944 mdb_dbis_update(MDB_txn *txn, int keep) 2945 { 2946 int i; 2947 MDB_dbi n = txn->mt_numdbs; 2948 MDB_env *env = txn->mt_env; 2949 unsigned char *tdbflags = txn->mt_dbflags; 2950 2951 for (i = n; --i >= CORE_DBS;) { 2952 if (tdbflags[i] & DB_NEW) { 2953 if (keep) { 2954 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID; 2955 } else { 2956 char *ptr = env->me_dbxs[i].md_name.mv_data; 2957 if (ptr) { 2958 env->me_dbxs[i].md_name.mv_data = NULL; 2959 env->me_dbxs[i].md_name.mv_size = 0; 2960 env->me_dbflags[i] = 0; 2961 env->me_dbiseqs[i]++; 2962 free(ptr); 2963 } 2964 } 2965 } 2966 } 2967 if (keep && env->me_numdbs < n) 2968 env->me_numdbs = n; 2969 } 2970 2971 /** End a transaction, except successful commit of a nested transaction. 2972 * May be called twice for readonly txns: First reset it, then abort. 2973 * @param[in] txn the transaction handle to end 2974 * @param[in] mode why and how to end the transaction 2975 */ 2976 static void 2977 mdb_txn_end(MDB_txn *txn, unsigned mode) 2978 { 2979 MDB_env *env = txn->mt_env; 2980 #if MDB_DEBUG 2981 static const char *const names[] = MDB_END_NAMES; 2982 #endif 2983 2984 /* Export or close DBI handles opened in this txn */ 2985 mdb_dbis_update(txn, mode & MDB_END_UPDATE); 2986 2987 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u", 2988 names[mode & MDB_END_OPMASK], 2989 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', 2990 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root)); 2991 2992 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) { 2993 if (txn->mt_u.reader) { 2994 txn->mt_u.reader->mr_txnid = (txnid_t)-1; 2995 if (!(env->me_flags & MDB_NOTLS)) { 2996 txn->mt_u.reader = NULL; /* txn does not own reader */ 2997 } else if (mode & MDB_END_SLOT) { 2998 txn->mt_u.reader->mr_pid = 0; 2999 txn->mt_u.reader = NULL; 3000 } /* else txn owns the slot until it does MDB_END_SLOT */ 3001 } 3002 txn->mt_numdbs = 0; /* prevent further DBI activity */ 3003 txn->mt_flags |= MDB_TXN_FINISHED; 3004 3005 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) { 3006 pgno_t *pghead = env->me_pghead; 3007 3008 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */ 3009 mdb_cursors_close(txn, 0); 3010 if (!(env->me_flags & MDB_WRITEMAP)) { 3011 mdb_dlist_free(txn); 3012 } 3013 3014 txn->mt_numdbs = 0; 3015 txn->mt_flags = MDB_TXN_FINISHED; 3016 3017 if (!txn->mt_parent) { 3018 mdb_midl_shrink(&txn->mt_free_pgs); 3019 env->me_free_pgs = txn->mt_free_pgs; 3020 /* me_pgstate: */ 3021 env->me_pghead = NULL; 3022 env->me_pglast = 0; 3023 3024 env->me_txn = NULL; 3025 mode = 0; /* txn == env->me_txn0, do not free() it */ 3026 3027 /* The writer mutex was locked in mdb_txn_begin. */ 3028 if (env->me_txns) 3029 UNLOCK_MUTEX(env->me_wmutex); 3030 } else { 3031 txn->mt_parent->mt_child = NULL; 3032 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD; 3033 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate; 3034 mdb_midl_free(txn->mt_free_pgs); 3035 free(txn->mt_u.dirty_list); 3036 } 3037 mdb_midl_free(txn->mt_spill_pgs); 3038 3039 mdb_midl_free(pghead); 3040 } 3041 3042 if (mode & MDB_END_FREE) 3043 free(txn); 3044 } 3045 3046 void 3047 mdb_txn_reset(MDB_txn *txn) 3048 { 3049 if (txn == NULL) 3050 return; 3051 3052 /* This call is only valid for read-only txns */ 3053 if (!(txn->mt_flags & MDB_TXN_RDONLY)) 3054 return; 3055 3056 mdb_txn_end(txn, MDB_END_RESET); 3057 } 3058 3059 void 3060 mdb_txn_abort(MDB_txn *txn) 3061 { 3062 if (txn == NULL) 3063 return; 3064 3065 if (txn->mt_child) 3066 mdb_txn_abort(txn->mt_child); 3067 3068 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE); 3069 } 3070 3071 /** Save the freelist as of this transaction to the freeDB. 3072 * This changes the freelist. Keep trying until it stabilizes. 3073 */ 3074 static int 3075 mdb_freelist_save(MDB_txn *txn) 3076 { 3077 /* env->me_pghead[] can grow and shrink during this call. 3078 * env->me_pglast and txn->mt_free_pgs[] can only grow. 3079 * Page numbers cannot disappear from txn->mt_free_pgs[]. 3080 */ 3081 MDB_cursor mc; 3082 MDB_env *env = txn->mt_env; 3083 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1; 3084 txnid_t pglast = 0, head_id = 0; 3085 pgno_t freecnt = 0, *free_pgs, *mop; 3086 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit; 3087 3088 mdb_cursor_init(&mc, txn, FREE_DBI, NULL); 3089 3090 if (env->me_pghead) { 3091 /* Make sure first page of freeDB is touched and on freelist */ 3092 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY); 3093 if (rc && rc != MDB_NOTFOUND) 3094 return rc; 3095 } 3096 3097 if (!env->me_pghead && txn->mt_loose_pgs) { 3098 /* Put loose page numbers in mt_free_pgs, since 3099 * we may be unable to return them to me_pghead. 3100 */ 3101 MDB_page *mp = txn->mt_loose_pgs; 3102 MDB_ID2 *dl = txn->mt_u.dirty_list; 3103 unsigned x; 3104 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0) 3105 return rc; 3106 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) { 3107 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno); 3108 /* must also remove from dirty list */ 3109 if (txn->mt_flags & MDB_TXN_WRITEMAP) { 3110 for (x=1; x<=dl[0].mid; x++) 3111 if (dl[x].mid == mp->mp_pgno) 3112 break; 3113 mdb_tassert(txn, x <= dl[0].mid); 3114 } else { 3115 x = mdb_mid2l_search(dl, mp->mp_pgno); 3116 mdb_tassert(txn, dl[x].mid == mp->mp_pgno); 3117 mdb_dpage_free(env, mp); 3118 } 3119 dl[x].mptr = NULL; 3120 } 3121 { 3122 /* squash freed slots out of the dirty list */ 3123 unsigned y; 3124 for (y=1; dl[y].mptr && y <= dl[0].mid; y++); 3125 if (y <= dl[0].mid) { 3126 for(x=y, y++;;) { 3127 while (!dl[y].mptr && y <= dl[0].mid) y++; 3128 if (y > dl[0].mid) break; 3129 dl[x++] = dl[y++]; 3130 } 3131 dl[0].mid = x-1; 3132 } else { 3133 /* all slots freed */ 3134 dl[0].mid = 0; 3135 } 3136 } 3137 txn->mt_loose_pgs = NULL; 3138 txn->mt_loose_count = 0; 3139 } 3140 3141 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */ 3142 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP)) 3143 ? SSIZE_MAX : maxfree_1pg; 3144 3145 for (;;) { 3146 /* Come back here after each Put() in case freelist changed */ 3147 MDB_val key, data; 3148 pgno_t *pgs; 3149 ssize_t j; 3150 3151 /* If using records from freeDB which we have not yet 3152 * deleted, delete them and any we reserved for me_pghead. 3153 */ 3154 while (pglast < env->me_pglast) { 3155 rc = mdb_cursor_first(&mc, &key, NULL); 3156 if (rc) 3157 return rc; 3158 pglast = head_id = *(txnid_t *)key.mv_data; 3159 total_room = head_room = 0; 3160 mdb_tassert(txn, pglast <= env->me_pglast); 3161 rc = mdb_cursor_del(&mc, 0); 3162 if (rc) 3163 return rc; 3164 } 3165 3166 /* Save the IDL of pages freed by this txn, to a single record */ 3167 if (freecnt < txn->mt_free_pgs[0]) { 3168 if (!freecnt) { 3169 /* Make sure last page of freeDB is touched and on freelist */ 3170 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY); 3171 if (rc && rc != MDB_NOTFOUND) 3172 return rc; 3173 } 3174 free_pgs = txn->mt_free_pgs; 3175 /* Write to last page of freeDB */ 3176 key.mv_size = sizeof(txn->mt_txnid); 3177 key.mv_data = &txn->mt_txnid; 3178 do { 3179 freecnt = free_pgs[0]; 3180 data.mv_size = MDB_IDL_SIZEOF(free_pgs); 3181 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE); 3182 if (rc) 3183 return rc; 3184 /* Retry if mt_free_pgs[] grew during the Put() */ 3185 free_pgs = txn->mt_free_pgs; 3186 } while (freecnt < free_pgs[0]); 3187 mdb_midl_sort(free_pgs); 3188 memcpy(data.mv_data, free_pgs, data.mv_size); 3189 #if (MDB_DEBUG) > 1 3190 { 3191 unsigned int i = free_pgs[0]; 3192 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u", 3193 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i)); 3194 for (; i; i--) 3195 DPRINTF(("IDL %"Z"u", free_pgs[i])); 3196 } 3197 #endif 3198 continue; 3199 } 3200 3201 mop = env->me_pghead; 3202 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count; 3203 3204 /* Reserve records for me_pghead[]. Split it if multi-page, 3205 * to avoid searching freeDB for a page range. Use keys in 3206 * range [1,me_pglast]: Smaller than txnid of oldest reader. 3207 */ 3208 if (total_room >= mop_len) { 3209 if (total_room == mop_len || --more < 0) 3210 break; 3211 } else if (head_room >= maxfree_1pg && head_id > 1) { 3212 /* Keep current record (overflow page), add a new one */ 3213 head_id--; 3214 head_room = 0; 3215 } 3216 /* (Re)write {key = head_id, IDL length = head_room} */ 3217 total_room -= head_room; 3218 head_room = mop_len - total_room; 3219 if (head_room > maxfree_1pg && head_id > 1) { 3220 /* Overflow multi-page for part of me_pghead */ 3221 head_room /= head_id; /* amortize page sizes */ 3222 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1); 3223 } else if (head_room < 0) { 3224 /* Rare case, not bothering to delete this record */ 3225 head_room = 0; 3226 } 3227 key.mv_size = sizeof(head_id); 3228 key.mv_data = &head_id; 3229 data.mv_size = (head_room + 1) * sizeof(pgno_t); 3230 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE); 3231 if (rc) 3232 return rc; 3233 /* IDL is initially empty, zero out at least the length */ 3234 pgs = (pgno_t *)data.mv_data; 3235 j = head_room > clean_limit ? head_room : 0; 3236 do { 3237 pgs[j] = 0; 3238 } while (--j >= 0); 3239 total_room += head_room; 3240 } 3241 3242 /* Return loose page numbers to me_pghead, though usually none are 3243 * left at this point. The pages themselves remain in dirty_list. 3244 */ 3245 if (txn->mt_loose_pgs) { 3246 MDB_page *mp = txn->mt_loose_pgs; 3247 unsigned count = txn->mt_loose_count; 3248 MDB_IDL loose; 3249 /* Room for loose pages + temp IDL with same */ 3250 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0) 3251 return rc; 3252 mop = env->me_pghead; 3253 loose = mop + MDB_IDL_ALLOCLEN(mop) - count; 3254 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp)) 3255 loose[ ++count ] = mp->mp_pgno; 3256 loose[0] = count; 3257 mdb_midl_sort(loose); 3258 mdb_midl_xmerge(mop, loose); 3259 txn->mt_loose_pgs = NULL; 3260 txn->mt_loose_count = 0; 3261 mop_len = mop[0]; 3262 } 3263 3264 /* Fill in the reserved me_pghead records */ 3265 rc = MDB_SUCCESS; 3266 if (mop_len) { 3267 MDB_val key, data; 3268 3269 mop += mop_len; 3270 rc = mdb_cursor_first(&mc, &key, &data); 3271 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) { 3272 txnid_t id = *(txnid_t *)key.mv_data; 3273 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1; 3274 MDB_ID save; 3275 3276 mdb_tassert(txn, len >= 0 && id <= env->me_pglast); 3277 key.mv_data = &id; 3278 if (len > mop_len) { 3279 len = mop_len; 3280 data.mv_size = (len + 1) * sizeof(MDB_ID); 3281 } 3282 data.mv_data = mop -= len; 3283 save = mop[0]; 3284 mop[0] = len; 3285 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT); 3286 mop[0] = save; 3287 if (rc || !(mop_len -= len)) 3288 break; 3289 } 3290 } 3291 return rc; 3292 } 3293 3294 /** Flush (some) dirty pages to the map, after clearing their dirty flag. 3295 * @param[in] txn the transaction that's being committed 3296 * @param[in] keep number of initial pages in dirty_list to keep dirty. 3297 * @return 0 on success, non-zero on failure. 3298 */ 3299 static int 3300 mdb_page_flush(MDB_txn *txn, int keep) 3301 { 3302 MDB_env *env = txn->mt_env; 3303 MDB_ID2L dl = txn->mt_u.dirty_list; 3304 unsigned psize = env->me_psize, j; 3305 int i, pagecount = dl[0].mid, rc; 3306 size_t size = 0, pos = 0; 3307 pgno_t pgno = 0; 3308 MDB_page *dp = NULL; 3309 #ifdef _WIN32 3310 OVERLAPPED ov; 3311 #else 3312 struct iovec iov[MDB_COMMIT_PAGES]; 3313 ssize_t wpos = 0, wsize = 0, wres; 3314 size_t next_pos = 1; /* impossible pos, so pos != next_pos */ 3315 int n = 0; 3316 #endif 3317 3318 j = i = keep; 3319 3320 if (env->me_flags & MDB_WRITEMAP) { 3321 /* Clear dirty flags */ 3322 while (++i <= pagecount) { 3323 dp = dl[i].mptr; 3324 /* Don't flush this page yet */ 3325 if (dp->mp_flags & (P_LOOSE|P_KEEP)) { 3326 dp->mp_flags &= ~P_KEEP; 3327 dl[++j] = dl[i]; 3328 continue; 3329 } 3330 dp->mp_flags &= ~P_DIRTY; 3331 } 3332 goto done; 3333 } 3334 3335 /* Write the pages */ 3336 for (;;) { 3337 if (++i <= pagecount) { 3338 dp = dl[i].mptr; 3339 /* Don't flush this page yet */ 3340 if (dp->mp_flags & (P_LOOSE|P_KEEP)) { 3341 dp->mp_flags &= ~P_KEEP; 3342 dl[i].mid = 0; 3343 continue; 3344 } 3345 pgno = dl[i].mid; 3346 /* clear dirty flag */ 3347 dp->mp_flags &= ~P_DIRTY; 3348 pos = pgno * psize; 3349 size = psize; 3350 if (IS_OVERFLOW(dp)) size *= dp->mp_pages; 3351 } 3352 #ifdef _WIN32 3353 else break; 3354 3355 /* Windows actually supports scatter/gather I/O, but only on 3356 * unbuffered file handles. Since we're relying on the OS page 3357 * cache for all our data, that's self-defeating. So we just 3358 * write pages one at a time. We use the ov structure to set 3359 * the write offset, to at least save the overhead of a Seek 3360 * system call. 3361 */ 3362 DPRINTF(("committing page %"Z"u", pgno)); 3363 memset(&ov, 0, sizeof(ov)); 3364 ov.Offset = pos & 0xffffffff; 3365 ov.OffsetHigh = pos >> 16 >> 16; 3366 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) { 3367 rc = ErrCode(); 3368 DPRINTF(("WriteFile: %d", rc)); 3369 return rc; 3370 } 3371 #else 3372 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */ 3373 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) { 3374 if (n) { 3375 retry_write: 3376 /* Write previous page(s) */ 3377 #ifdef MDB_USE_PWRITEV 3378 wres = pwritev(env->me_fd, iov, n, wpos); 3379 #else 3380 if (n == 1) { 3381 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos); 3382 } else { 3383 retry_seek: 3384 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) { 3385 rc = ErrCode(); 3386 if (rc == EINTR) 3387 goto retry_seek; 3388 DPRINTF(("lseek: %s", strerror(rc))); 3389 return rc; 3390 } 3391 wres = writev(env->me_fd, iov, n); 3392 } 3393 #endif 3394 if (wres != wsize) { 3395 if (wres < 0) { 3396 rc = ErrCode(); 3397 if (rc == EINTR) 3398 goto retry_write; 3399 DPRINTF(("Write error: %s", strerror(rc))); 3400 } else { 3401 rc = EIO; /* TODO: Use which error code? */ 3402 DPUTS("short write, filesystem full?"); 3403 } 3404 return rc; 3405 } 3406 n = 0; 3407 } 3408 if (i > pagecount) 3409 break; 3410 wpos = pos; 3411 wsize = 0; 3412 } 3413 DPRINTF(("committing page %"Z"u", pgno)); 3414 next_pos = pos + size; 3415 iov[n].iov_len = size; 3416 iov[n].iov_base = (char *)dp; 3417 wsize += size; 3418 n++; 3419 #endif /* _WIN32 */ 3420 } 3421 3422 /* MIPS has cache coherency issues, this is a no-op everywhere else 3423 * Note: for any size >= on-chip cache size, entire on-chip cache is 3424 * flushed. 3425 */ 3426 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE); 3427 3428 for (i = keep; ++i <= pagecount; ) { 3429 dp = dl[i].mptr; 3430 /* This is a page we skipped above */ 3431 if (!dl[i].mid) { 3432 dl[++j] = dl[i]; 3433 dl[j].mid = dp->mp_pgno; 3434 continue; 3435 } 3436 mdb_dpage_free(env, dp); 3437 } 3438 3439 done: 3440 i--; 3441 txn->mt_dirty_room += i - j; 3442 dl[0].mid = j; 3443 return MDB_SUCCESS; 3444 } 3445 3446 int 3447 mdb_txn_commit(MDB_txn *txn) 3448 { 3449 int rc; 3450 unsigned int i, end_mode; 3451 MDB_env *env; 3452 3453 if (txn == NULL) 3454 return EINVAL; 3455 3456 /* mdb_txn_end() mode for a commit which writes nothing */ 3457 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE; 3458 3459 if (txn->mt_child) { 3460 rc = mdb_txn_commit(txn->mt_child); 3461 if (rc) 3462 goto fail; 3463 } 3464 3465 env = txn->mt_env; 3466 3467 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) { 3468 goto done; 3469 } 3470 3471 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) { 3472 DPUTS("txn has failed/finished, can't commit"); 3473 if (txn->mt_parent) 3474 txn->mt_parent->mt_flags |= MDB_TXN_ERROR; 3475 rc = MDB_BAD_TXN; 3476 goto fail; 3477 } 3478 3479 if (txn->mt_parent) { 3480 MDB_txn *parent = txn->mt_parent; 3481 MDB_page **lp; 3482 MDB_ID2L dst, src; 3483 MDB_IDL pspill; 3484 unsigned x, y, len, ps_len; 3485 3486 /* Append our free list to parent's */ 3487 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs); 3488 if (rc) 3489 goto fail; 3490 mdb_midl_free(txn->mt_free_pgs); 3491 /* Failures after this must either undo the changes 3492 * to the parent or set MDB_TXN_ERROR in the parent. 3493 */ 3494 3495 parent->mt_next_pgno = txn->mt_next_pgno; 3496 parent->mt_flags = txn->mt_flags; 3497 3498 /* Merge our cursors into parent's and close them */ 3499 mdb_cursors_close(txn, 1); 3500 3501 /* Update parent's DB table. */ 3502 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db)); 3503 parent->mt_numdbs = txn->mt_numdbs; 3504 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI]; 3505 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI]; 3506 for (i=CORE_DBS; i<txn->mt_numdbs; i++) { 3507 /* preserve parent's DB_NEW status */ 3508 x = parent->mt_dbflags[i] & DB_NEW; 3509 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x; 3510 } 3511 3512 dst = parent->mt_u.dirty_list; 3513 src = txn->mt_u.dirty_list; 3514 /* Remove anything in our dirty list from parent's spill list */ 3515 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) { 3516 x = y = ps_len; 3517 pspill[0] = (pgno_t)-1; 3518 /* Mark our dirty pages as deleted in parent spill list */ 3519 for (i=0, len=src[0].mid; ++i <= len; ) { 3520 MDB_ID pn = src[i].mid << 1; 3521 while (pn > pspill[x]) 3522 x--; 3523 if (pn == pspill[x]) { 3524 pspill[x] = 1; 3525 y = --x; 3526 } 3527 } 3528 /* Squash deleted pagenums if we deleted any */ 3529 for (x=y; ++x <= ps_len; ) 3530 if (!(pspill[x] & 1)) 3531 pspill[++y] = pspill[x]; 3532 pspill[0] = y; 3533 } 3534 3535 /* Remove anything in our spill list from parent's dirty list */ 3536 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) { 3537 for (i=1; i<=txn->mt_spill_pgs[0]; i++) { 3538 MDB_ID pn = txn->mt_spill_pgs[i]; 3539 if (pn & 1) 3540 continue; /* deleted spillpg */ 3541 pn >>= 1; 3542 y = mdb_mid2l_search(dst, pn); 3543 if (y <= dst[0].mid && dst[y].mid == pn) { 3544 free(dst[y].mptr); 3545 while (y < dst[0].mid) { 3546 dst[y] = dst[y+1]; 3547 y++; 3548 } 3549 dst[0].mid--; 3550 } 3551 } 3552 } 3553 3554 /* Find len = length of merging our dirty list with parent's */ 3555 x = dst[0].mid; 3556 dst[0].mid = 0; /* simplify loops */ 3557 if (parent->mt_parent) { 3558 len = x + src[0].mid; 3559 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1; 3560 for (i = x; y && i; y--) { 3561 pgno_t yp = src[y].mid; 3562 while (yp < dst[i].mid) 3563 i--; 3564 if (yp == dst[i].mid) { 3565 i--; 3566 len--; 3567 } 3568 } 3569 } else { /* Simplify the above for single-ancestor case */ 3570 len = MDB_IDL_UM_MAX - txn->mt_dirty_room; 3571 } 3572 /* Merge our dirty list with parent's */ 3573 y = src[0].mid; 3574 for (i = len; y; dst[i--] = src[y--]) { 3575 pgno_t yp = src[y].mid; 3576 while (yp < dst[x].mid) 3577 dst[i--] = dst[x--]; 3578 if (yp == dst[x].mid) 3579 free(dst[x--].mptr); 3580 } 3581 mdb_tassert(txn, i == x); 3582 dst[0].mid = len; 3583 free(txn->mt_u.dirty_list); 3584 parent->mt_dirty_room = txn->mt_dirty_room; 3585 if (txn->mt_spill_pgs) { 3586 if (parent->mt_spill_pgs) { 3587 /* TODO: Prevent failure here, so parent does not fail */ 3588 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs); 3589 if (rc) 3590 parent->mt_flags |= MDB_TXN_ERROR; 3591 mdb_midl_free(txn->mt_spill_pgs); 3592 mdb_midl_sort(parent->mt_spill_pgs); 3593 } else { 3594 parent->mt_spill_pgs = txn->mt_spill_pgs; 3595 } 3596 } 3597 3598 /* Append our loose page list to parent's */ 3599 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp)) 3600 ; 3601 *lp = txn->mt_loose_pgs; 3602 parent->mt_loose_count += txn->mt_loose_count; 3603 3604 parent->mt_child = NULL; 3605 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead); 3606 free(txn); 3607 return rc; 3608 } 3609 3610 if (txn != env->me_txn) { 3611 DPUTS("attempt to commit unknown transaction"); 3612 rc = EINVAL; 3613 goto fail; 3614 } 3615 3616 mdb_cursors_close(txn, 0); 3617 3618 if (!txn->mt_u.dirty_list[0].mid && 3619 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS))) 3620 goto done; 3621 3622 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u", 3623 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root)); 3624 3625 /* Update DB root pointers */ 3626 if (txn->mt_numdbs > CORE_DBS) { 3627 MDB_cursor mc; 3628 MDB_dbi i; 3629 MDB_val data; 3630 data.mv_size = sizeof(MDB_db); 3631 3632 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL); 3633 for (i = CORE_DBS; i < txn->mt_numdbs; i++) { 3634 if (txn->mt_dbflags[i] & DB_DIRTY) { 3635 if (TXN_DBI_CHANGED(txn, i)) { 3636 rc = MDB_BAD_DBI; 3637 goto fail; 3638 } 3639 data.mv_data = &txn->mt_dbs[i]; 3640 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 3641 F_SUBDATA); 3642 if (rc) 3643 goto fail; 3644 } 3645 } 3646 } 3647 3648 rc = mdb_freelist_save(txn); 3649 if (rc) 3650 goto fail; 3651 3652 mdb_midl_free(env->me_pghead); 3653 env->me_pghead = NULL; 3654 mdb_midl_shrink(&txn->mt_free_pgs); 3655 3656 #if (MDB_DEBUG) > 2 3657 mdb_audit(txn); 3658 #endif 3659 3660 if ((rc = mdb_page_flush(txn, 0)) || 3661 (rc = mdb_env_sync(env, 0)) || 3662 (rc = mdb_env_write_meta(txn))) 3663 goto fail; 3664 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE; 3665 3666 done: 3667 mdb_txn_end(txn, end_mode); 3668 return MDB_SUCCESS; 3669 3670 fail: 3671 mdb_txn_abort(txn); 3672 return rc; 3673 } 3674 3675 /** Read the environment parameters of a DB environment before 3676 * mapping it into memory. 3677 * @param[in] env the environment handle 3678 * @param[out] meta address of where to store the meta information 3679 * @return 0 on success, non-zero on failure. 3680 */ 3681 static int ESECT 3682 mdb_env_read_header(MDB_env *env, MDB_meta *meta) 3683 { 3684 MDB_metabuf pbuf; 3685 MDB_page *p; 3686 MDB_meta *m; 3687 int i, rc, off; 3688 enum { Size = sizeof(pbuf) }; 3689 3690 /* We don't know the page size yet, so use a minimum value. 3691 * Read both meta pages so we can use the latest one. 3692 */ 3693 3694 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) { 3695 #ifdef _WIN32 3696 DWORD len; 3697 OVERLAPPED ov; 3698 memset(&ov, 0, sizeof(ov)); 3699 ov.Offset = off; 3700 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1; 3701 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF) 3702 rc = 0; 3703 #else 3704 rc = pread(env->me_fd, &pbuf, Size, off); 3705 #endif 3706 if (rc != Size) { 3707 if (rc == 0 && off == 0) 3708 return ENOENT; 3709 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID; 3710 DPRINTF(("read: %s", mdb_strerror(rc))); 3711 return rc; 3712 } 3713 3714 p = (MDB_page *)&pbuf; 3715 3716 if (!F_ISSET(p->mp_flags, P_META)) { 3717 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno)); 3718 return MDB_INVALID; 3719 } 3720 3721 m = METADATA(p); 3722 if (m->mm_magic != MDB_MAGIC) { 3723 DPUTS("meta has invalid magic"); 3724 return MDB_INVALID; 3725 } 3726 3727 if (m->mm_version != MDB_DATA_VERSION) { 3728 DPRINTF(("database is version %u, expected version %u", 3729 m->mm_version, MDB_DATA_VERSION)); 3730 return MDB_VERSION_MISMATCH; 3731 } 3732 3733 if (off == 0 || m->mm_txnid > meta->mm_txnid) 3734 *meta = *m; 3735 } 3736 return 0; 3737 } 3738 3739 /** Fill in most of the zeroed #MDB_meta for an empty database environment */ 3740 static void ESECT 3741 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta) 3742 { 3743 meta->mm_magic = MDB_MAGIC; 3744 meta->mm_version = MDB_DATA_VERSION; 3745 meta->mm_mapsize = env->me_mapsize; 3746 meta->mm_psize = env->me_psize; 3747 meta->mm_last_pg = NUM_METAS-1; 3748 meta->mm_flags = env->me_flags & 0xffff; 3749 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */ 3750 meta->mm_dbs[FREE_DBI].md_root = P_INVALID; 3751 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID; 3752 } 3753 3754 /** Write the environment parameters of a freshly created DB environment. 3755 * @param[in] env the environment handle 3756 * @param[in] meta the #MDB_meta to write 3757 * @return 0 on success, non-zero on failure. 3758 */ 3759 static int ESECT 3760 mdb_env_init_meta(MDB_env *env, MDB_meta *meta) 3761 { 3762 MDB_page *p, *q; 3763 int rc; 3764 unsigned int psize; 3765 #ifdef _WIN32 3766 DWORD len; 3767 OVERLAPPED ov; 3768 memset(&ov, 0, sizeof(ov)); 3769 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \ 3770 ov.Offset = pos; \ 3771 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0) 3772 #else 3773 int len; 3774 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \ 3775 len = pwrite(fd, ptr, size, pos); \ 3776 if (len == -1 && ErrCode() == EINTR) continue; \ 3777 rc = (len >= 0); break; } while(1) 3778 #endif 3779 3780 DPUTS("writing new meta page"); 3781 3782 psize = env->me_psize; 3783 3784 p = calloc(NUM_METAS, psize); 3785 if (!p) 3786 return ENOMEM; 3787 3788 p->mp_pgno = 0; 3789 p->mp_flags = P_META; 3790 *(MDB_meta *)METADATA(p) = *meta; 3791 3792 q = (MDB_page *)((char *)p + psize); 3793 q->mp_pgno = 1; 3794 q->mp_flags = P_META; 3795 *(MDB_meta *)METADATA(q) = *meta; 3796 3797 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0); 3798 if (!rc) 3799 rc = ErrCode(); 3800 else if ((unsigned) len == psize * NUM_METAS) 3801 rc = MDB_SUCCESS; 3802 else 3803 rc = ENOSPC; 3804 free(p); 3805 return rc; 3806 } 3807 3808 /** Update the environment info to commit a transaction. 3809 * @param[in] txn the transaction that's being committed 3810 * @return 0 on success, non-zero on failure. 3811 */ 3812 static int 3813 mdb_env_write_meta(MDB_txn *txn) 3814 { 3815 MDB_env *env; 3816 MDB_meta meta, metab, *mp; 3817 unsigned flags; 3818 size_t mapsize; 3819 off_t off; 3820 int rc, len, toggle; 3821 char *ptr; 3822 HANDLE mfd; 3823 #ifdef _WIN32 3824 OVERLAPPED ov; 3825 #else 3826 int r2; 3827 #endif 3828 3829 toggle = txn->mt_txnid & 1; 3830 DPRINTF(("writing meta page %d for root page %"Z"u", 3831 toggle, txn->mt_dbs[MAIN_DBI].md_root)); 3832 3833 env = txn->mt_env; 3834 flags = env->me_flags; 3835 mp = env->me_metas[toggle]; 3836 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize; 3837 /* Persist any increases of mapsize config */ 3838 if (mapsize < env->me_mapsize) 3839 mapsize = env->me_mapsize; 3840 3841 if (flags & MDB_WRITEMAP) { 3842 mp->mm_mapsize = mapsize; 3843 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI]; 3844 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI]; 3845 mp->mm_last_pg = txn->mt_next_pgno - 1; 3846 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \ 3847 !(defined(__i386__) || defined(__x86_64__)) 3848 /* LY: issue a memory barrier, if not x86. ITS#7969 */ 3849 __sync_synchronize(); 3850 #endif 3851 mp->mm_txnid = txn->mt_txnid; 3852 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) { 3853 unsigned meta_size = env->me_psize; 3854 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC; 3855 ptr = (char *)mp - PAGEHDRSZ; 3856 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */ 3857 r2 = (ptr - env->me_map) & (env->me_os_psize - 1); 3858 ptr -= r2; 3859 meta_size += r2; 3860 #endif 3861 if (MDB_MSYNC(ptr, meta_size, rc)) { 3862 rc = ErrCode(); 3863 goto fail; 3864 } 3865 } 3866 goto done; 3867 } 3868 metab.mm_txnid = mp->mm_txnid; 3869 metab.mm_last_pg = mp->mm_last_pg; 3870 3871 meta.mm_mapsize = mapsize; 3872 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI]; 3873 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI]; 3874 meta.mm_last_pg = txn->mt_next_pgno - 1; 3875 meta.mm_txnid = txn->mt_txnid; 3876 3877 off = offsetof(MDB_meta, mm_mapsize); 3878 ptr = (char *)&meta + off; 3879 len = sizeof(MDB_meta) - off; 3880 off += (char *)mp - env->me_map; 3881 3882 /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC. 3883 * (me_mfd goes to the same file as me_fd, but writing to it 3884 * also syncs to disk. Avoids a separate fdatasync() call.) 3885 */ 3886 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd; 3887 #ifdef _WIN32 3888 { 3889 memset(&ov, 0, sizeof(ov)); 3890 ov.Offset = off; 3891 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov)) 3892 rc = -1; 3893 } 3894 #else 3895 retry_write: 3896 rc = pwrite(mfd, ptr, len, off); 3897 #endif 3898 if (rc != len) { 3899 rc = rc < 0 ? ErrCode() : EIO; 3900 #ifndef _WIN32 3901 if (rc == EINTR) 3902 goto retry_write; 3903 #endif 3904 DPUTS("write failed, disk error?"); 3905 /* On a failure, the pagecache still contains the new data. 3906 * Write some old data back, to prevent it from being used. 3907 * Use the non-SYNC fd; we know it will fail anyway. 3908 */ 3909 meta.mm_last_pg = metab.mm_last_pg; 3910 meta.mm_txnid = metab.mm_txnid; 3911 #ifdef _WIN32 3912 memset(&ov, 0, sizeof(ov)); 3913 ov.Offset = off; 3914 WriteFile(env->me_fd, ptr, len, NULL, &ov); 3915 #else 3916 r2 = pwrite(env->me_fd, ptr, len, off); 3917 (void)r2; /* Silence warnings. We don't care about pwrite's return value */ 3918 #endif 3919 fail: 3920 env->me_flags |= MDB_FATAL_ERROR; 3921 return rc; 3922 } 3923 /* MIPS has cache coherency issues, this is a no-op everywhere else */ 3924 CACHEFLUSH(env->me_map + off, len, DCACHE); 3925 done: 3926 /* Memory ordering issues are irrelevant; since the entire writer 3927 * is wrapped by wmutex, all of these changes will become visible 3928 * after the wmutex is unlocked. Since the DB is multi-version, 3929 * readers will get consistent data regardless of how fresh or 3930 * how stale their view of these values is. 3931 */ 3932 if (env->me_txns) 3933 env->me_txns->mti_txnid = txn->mt_txnid; 3934 3935 return MDB_SUCCESS; 3936 } 3937 3938 /** Check both meta pages to see which one is newer. 3939 * @param[in] env the environment handle 3940 * @return newest #MDB_meta. 3941 */ 3942 static MDB_meta * 3943 mdb_env_pick_meta(const MDB_env *env) 3944 { 3945 MDB_meta *const *metas = env->me_metas; 3946 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ]; 3947 } 3948 3949 int ESECT 3950 mdb_env_create(MDB_env **env) 3951 { 3952 MDB_env *e; 3953 3954 e = calloc(1, sizeof(MDB_env)); 3955 if (!e) 3956 return ENOMEM; 3957 3958 e->me_maxreaders = DEFAULT_READERS; 3959 e->me_maxdbs = e->me_numdbs = CORE_DBS; 3960 e->me_fd = INVALID_HANDLE_VALUE; 3961 e->me_lfd = INVALID_HANDLE_VALUE; 3962 e->me_mfd = INVALID_HANDLE_VALUE; 3963 #ifdef MDB_USE_POSIX_SEM 3964 e->me_rmutex = SEM_FAILED; 3965 e->me_wmutex = SEM_FAILED; 3966 #endif 3967 e->me_pid = getpid(); 3968 GET_PAGESIZE(e->me_os_psize); 3969 VGMEMP_CREATE(e,0,0); 3970 *env = e; 3971 return MDB_SUCCESS; 3972 } 3973 3974 static int ESECT 3975 mdb_env_map(MDB_env *env, void *addr) 3976 { 3977 MDB_page *p; 3978 unsigned int flags = env->me_flags; 3979 #ifdef _WIN32 3980 int rc; 3981 HANDLE mh; 3982 LONG sizelo, sizehi; 3983 size_t msize; 3984 3985 if (flags & MDB_RDONLY) { 3986 /* Don't set explicit map size, use whatever exists */ 3987 msize = 0; 3988 sizelo = 0; 3989 sizehi = 0; 3990 } else { 3991 msize = env->me_mapsize; 3992 sizelo = msize & 0xffffffff; 3993 sizehi = msize >> 16 >> 16; /* only needed on Win64 */ 3994 3995 /* Windows won't create mappings for zero length files. 3996 * and won't map more than the file size. 3997 * Just set the maxsize right now. 3998 */ 3999 if (!(flags & MDB_WRITEMAP) && (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo 4000 || !SetEndOfFile(env->me_fd) 4001 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)) 4002 return ErrCode(); 4003 } 4004 4005 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ? 4006 PAGE_READWRITE : PAGE_READONLY, 4007 sizehi, sizelo, NULL); 4008 if (!mh) 4009 return ErrCode(); 4010 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ? 4011 FILE_MAP_WRITE : FILE_MAP_READ, 4012 0, 0, msize, addr); 4013 rc = env->me_map ? 0 : ErrCode(); 4014 CloseHandle(mh); 4015 if (rc) 4016 return rc; 4017 #else 4018 int mmap_flags = MAP_SHARED; 4019 int prot = PROT_READ; 4020 #ifdef MAP_NOSYNC /* Used on FreeBSD */ 4021 if (flags & MDB_NOSYNC) 4022 mmap_flags |= MAP_NOSYNC; 4023 #endif 4024 if (flags & MDB_WRITEMAP) { 4025 prot |= PROT_WRITE; 4026 if (ftruncate(env->me_fd, env->me_mapsize) < 0) 4027 return ErrCode(); 4028 } 4029 env->me_map = mmap(addr, env->me_mapsize, prot, mmap_flags, 4030 env->me_fd, 0); 4031 if (env->me_map == MAP_FAILED) { 4032 env->me_map = NULL; 4033 return ErrCode(); 4034 } 4035 4036 if (flags & MDB_NORDAHEAD) { 4037 /* Turn off readahead. It's harmful when the DB is larger than RAM. */ 4038 #ifdef MADV_RANDOM 4039 madvise(env->me_map, env->me_mapsize, MADV_RANDOM); 4040 #else 4041 #ifdef POSIX_MADV_RANDOM 4042 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM); 4043 #endif /* POSIX_MADV_RANDOM */ 4044 #endif /* MADV_RANDOM */ 4045 } 4046 #endif /* _WIN32 */ 4047 4048 /* Can happen because the address argument to mmap() is just a 4049 * hint. mmap() can pick another, e.g. if the range is in use. 4050 * The MAP_FIXED flag would prevent that, but then mmap could 4051 * instead unmap existing pages to make room for the new map. 4052 */ 4053 if (addr && env->me_map != addr) 4054 return EBUSY; /* TODO: Make a new MDB_* error code? */ 4055 4056 p = (MDB_page *)env->me_map; 4057 env->me_metas[0] = METADATA(p); 4058 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize); 4059 4060 return MDB_SUCCESS; 4061 } 4062 4063 int ESECT 4064 mdb_env_set_mapsize(MDB_env *env, size_t size) 4065 { 4066 /* If env is already open, caller is responsible for making 4067 * sure there are no active txns. 4068 */ 4069 if (env->me_map) { 4070 int rc; 4071 MDB_meta *meta; 4072 void *old; 4073 if (env->me_txn) 4074 return EINVAL; 4075 meta = mdb_env_pick_meta(env); 4076 if (!size) 4077 size = meta->mm_mapsize; 4078 { 4079 /* Silently round up to minimum if the size is too small */ 4080 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize; 4081 if (size < minsize) 4082 size = minsize; 4083 } 4084 munmap(env->me_map, env->me_mapsize); 4085 env->me_mapsize = size; 4086 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL; 4087 rc = mdb_env_map(env, old); 4088 if (rc) 4089 return rc; 4090 } 4091 env->me_mapsize = size; 4092 if (env->me_psize) 4093 env->me_maxpg = env->me_mapsize / env->me_psize; 4094 return MDB_SUCCESS; 4095 } 4096 4097 int ESECT 4098 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs) 4099 { 4100 if (env->me_map) 4101 return EINVAL; 4102 env->me_maxdbs = dbs + CORE_DBS; 4103 return MDB_SUCCESS; 4104 } 4105 4106 int ESECT 4107 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers) 4108 { 4109 if (env->me_map || readers < 1) 4110 return EINVAL; 4111 env->me_maxreaders = readers; 4112 return MDB_SUCCESS; 4113 } 4114 4115 int ESECT 4116 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers) 4117 { 4118 if (!env || !readers) 4119 return EINVAL; 4120 *readers = env->me_maxreaders; 4121 return MDB_SUCCESS; 4122 } 4123 4124 static int ESECT 4125 mdb_fsize(HANDLE fd, size_t *size) 4126 { 4127 #ifdef _WIN32 4128 LARGE_INTEGER fsize; 4129 4130 if (!GetFileSizeEx(fd, &fsize)) 4131 return ErrCode(); 4132 4133 *size = fsize.QuadPart; 4134 #else 4135 struct stat st; 4136 4137 if (fstat(fd, &st)) 4138 return ErrCode(); 4139 4140 *size = st.st_size; 4141 #endif 4142 return MDB_SUCCESS; 4143 } 4144 4145 4146 #ifdef _WIN32 4147 typedef wchar_t mdb_nchar_t; 4148 # define MDB_NAME(str) L##str 4149 # define mdb_name_cpy wcscpy 4150 #else 4151 /** Character type for file names: char on Unix, wchar_t on Windows */ 4152 typedef char mdb_nchar_t; 4153 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */ 4154 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */ 4155 #endif 4156 4157 /** Filename - string of #mdb_nchar_t[] */ 4158 typedef struct MDB_name { 4159 int mn_len; /**< Length */ 4160 int mn_alloced; /**< True if #mn_val was malloced */ 4161 mdb_nchar_t *mn_val; /**< Contents */ 4162 } MDB_name; 4163 4164 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */ 4165 static const mdb_nchar_t *const mdb_suffixes[2][2] = { 4166 { MDB_NAME("/data.mdb"), MDB_NAME("") }, 4167 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") } 4168 }; 4169 4170 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */ 4171 4172 /** Set up filename + scratch area for filename suffix, for opening files. 4173 * It should be freed with #mdb_fname_destroy(). 4174 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16. 4175 * 4176 * @param[in] path Pathname for #mdb_env_open(). 4177 * @param[in] envflags Whether a subdir and/or lockfile will be used. 4178 * @param[out] fname Resulting filename, with room for a suffix if necessary. 4179 */ 4180 static int ESECT 4181 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname) 4182 { 4183 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK); 4184 fname->mn_alloced = 0; 4185 #ifdef _WIN32 4186 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN); 4187 #else 4188 fname->mn_len = strlen(path); 4189 if (no_suffix) 4190 fname->mn_val = (char *) path; 4191 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) { 4192 fname->mn_alloced = 1; 4193 strcpy(fname->mn_val, path); 4194 } 4195 else 4196 return ENOMEM; 4197 return MDB_SUCCESS; 4198 #endif 4199 } 4200 4201 /** Destroy \b fname from #mdb_fname_init() */ 4202 #define mdb_fname_destroy(fname) \ 4203 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0) 4204 4205 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */ 4206 # define MDB_CLOEXEC O_CLOEXEC 4207 #else 4208 # define MDB_CLOEXEC 0 4209 #endif 4210 4211 /** File type, access mode etc. for #mdb_fopen() */ 4212 enum mdb_fopen_type { 4213 #ifdef _WIN32 4214 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS 4215 #else 4216 /* A comment in mdb_fopen() explains some O_* flag choices. */ 4217 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */ 4218 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */ 4219 MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */ 4220 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */ 4221 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits 4222 * distinguish otherwise-equal MDB_O_* constants from each other. 4223 */ 4224 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY, 4225 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */ 4226 #endif 4227 }; 4228 4229 /** Open an LMDB file. 4230 * @param[in] env The LMDB environment. 4231 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is 4232 * appended if necessary to create the filename, without changing mn_len. 4233 * @param[in] which Determines file type, access mode, etc. 4234 * @param[in] mode The Unix permissions for the file, if we create it. 4235 * @param[out] res Resulting file handle. 4236 * @return 0 on success, non-zero on failure. 4237 */ 4238 static int ESECT 4239 mdb_fopen(const MDB_env *env, MDB_name *fname, 4240 enum mdb_fopen_type which, mdb_mode_t mode, 4241 HANDLE *res) 4242 { 4243 int rc = MDB_SUCCESS; 4244 HANDLE fd; 4245 #ifdef _WIN32 4246 DWORD acc, share, disp, attrs; 4247 #else 4248 int flags; 4249 #endif 4250 4251 if (fname->mn_alloced) /* modifiable copy */ 4252 mdb_name_cpy(fname->mn_val + fname->mn_len, 4253 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]); 4254 4255 /* The directory must already exist. Usually the file need not. 4256 * MDB_O_META requires the file because we already created it using 4257 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file. 4258 * 4259 * With MDB_O_COPY we do not want the OS to cache the writes, since 4260 * the source data is already in the OS cache. 4261 * 4262 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*()) 4263 * to avoid the flock() issues noted under Caveats in lmdb.h. 4264 * Also set it for other filehandles which the user cannot get at 4265 * and close himself, which he may need after fork(). I.e. all but 4266 * me_fd, which programs do use via mdb_env_get_fd(). 4267 */ 4268 4269 #ifdef _WIN32 4270 acc = GENERIC_READ|GENERIC_WRITE; 4271 share = FILE_SHARE_READ|FILE_SHARE_WRITE; 4272 disp = OPEN_ALWAYS; 4273 attrs = FILE_ATTRIBUTE_NORMAL; 4274 switch (which) { 4275 case MDB_O_RDONLY: /* read-only datafile */ 4276 acc = GENERIC_READ; 4277 disp = OPEN_EXISTING; 4278 break; 4279 case MDB_O_META: /* for writing metapages */ 4280 acc = GENERIC_WRITE; 4281 disp = OPEN_EXISTING; 4282 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH; 4283 break; 4284 case MDB_O_COPY: /* mdb_env_copy() & co */ 4285 acc = GENERIC_WRITE; 4286 share = 0; 4287 disp = CREATE_NEW; 4288 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH; 4289 break; 4290 default: break; /* silence gcc -Wswitch (not all enum values handled) */ 4291 } 4292 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL); 4293 #else 4294 fd = open(fname->mn_val, which & MDB_O_MASK, mode); 4295 #endif 4296 4297 if (fd == INVALID_HANDLE_VALUE) 4298 rc = ErrCode(); 4299 #ifndef _WIN32 4300 else { 4301 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) { 4302 /* Set CLOEXEC if we could not pass it to open() */ 4303 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1) 4304 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC); 4305 } 4306 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) { 4307 /* This may require buffer alignment. There is no portable 4308 * way to ask how much, so we require OS pagesize alignment. 4309 */ 4310 # ifdef F_NOCACHE /* __APPLE__ */ 4311 (void) fcntl(fd, F_NOCACHE, 1); 4312 # elif defined O_DIRECT 4313 /* open(...O_DIRECT...) would break on filesystems without 4314 * O_DIRECT support (ITS#7682). Try to set it here instead. 4315 */ 4316 if ((flags = fcntl(fd, F_GETFL)) != -1) 4317 (void) fcntl(fd, F_SETFL, flags | O_DIRECT); 4318 # endif 4319 } 4320 } 4321 #endif /* !_WIN32 */ 4322 4323 *res = fd; 4324 return rc; 4325 } 4326 4327 4328 #ifdef BROKEN_FDATASYNC 4329 #include <sys/utsname.h> 4330 #include <sys/vfs.h> 4331 #endif 4332 4333 /** Further setup required for opening an LMDB environment 4334 */ 4335 static int ESECT 4336 mdb_env_open2(MDB_env *env) 4337 { 4338 unsigned int flags = env->me_flags; 4339 int i, newenv = 0, rc; 4340 MDB_meta meta; 4341 4342 #ifdef _WIN32 4343 /* See if we should use QueryLimited */ 4344 rc = GetVersion(); 4345 if ((rc & 0xff) > 5) 4346 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION; 4347 else 4348 env->me_pidquery = PROCESS_QUERY_INFORMATION; 4349 #endif /* _WIN32 */ 4350 4351 #ifdef BROKEN_FDATASYNC 4352 /* ext3/ext4 fdatasync is broken on some older Linux kernels. 4353 * https://lkml.org/lkml/2012/9/3/83 4354 * Kernels after 3.6-rc6 are known good. 4355 * https://lkml.org/lkml/2012/9/10/556 4356 * See if the DB is on ext3/ext4, then check for new enough kernel 4357 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known 4358 * to be patched. 4359 */ 4360 { 4361 struct statfs st; 4362 fstatfs(env->me_fd, &st); 4363 while (st.f_type == 0xEF53) { 4364 struct utsname uts; 4365 int i; 4366 uname(&uts); 4367 if (uts.release[0] < '3') { 4368 if (!strncmp(uts.release, "2.6.32.", 7)) { 4369 i = atoi(uts.release+7); 4370 if (i >= 60) 4371 break; /* 2.6.32.60 and newer is OK */ 4372 } else if (!strncmp(uts.release, "2.6.34.", 7)) { 4373 i = atoi(uts.release+7); 4374 if (i >= 15) 4375 break; /* 2.6.34.15 and newer is OK */ 4376 } 4377 } else if (uts.release[0] == '3') { 4378 i = atoi(uts.release+2); 4379 if (i > 5) 4380 break; /* 3.6 and newer is OK */ 4381 if (i == 5) { 4382 i = atoi(uts.release+4); 4383 if (i >= 4) 4384 break; /* 3.5.4 and newer is OK */ 4385 } else if (i == 2) { 4386 i = atoi(uts.release+4); 4387 if (i >= 30) 4388 break; /* 3.2.30 and newer is OK */ 4389 } 4390 } else { /* 4.x and newer is OK */ 4391 break; 4392 } 4393 env->me_flags |= MDB_FSYNCONLY; 4394 break; 4395 } 4396 } 4397 #endif 4398 4399 if ((i = mdb_env_read_header(env, &meta)) != 0) { 4400 if (i != ENOENT) 4401 return i; 4402 DPUTS("new mdbenv"); 4403 newenv = 1; 4404 env->me_psize = env->me_os_psize; 4405 if (env->me_psize > MAX_PAGESIZE) 4406 env->me_psize = MAX_PAGESIZE; 4407 memset(&meta, 0, sizeof(meta)); 4408 mdb_env_init_meta0(env, &meta); 4409 meta.mm_mapsize = DEFAULT_MAPSIZE; 4410 } else { 4411 env->me_psize = meta.mm_psize; 4412 } 4413 4414 /* Was a mapsize configured? */ 4415 if (!env->me_mapsize) { 4416 env->me_mapsize = meta.mm_mapsize; 4417 } 4418 { 4419 /* Make sure mapsize >= committed data size. Even when using 4420 * mm_mapsize, which could be broken in old files (ITS#7789). 4421 */ 4422 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize; 4423 if (env->me_mapsize < minsize) 4424 env->me_mapsize = minsize; 4425 } 4426 meta.mm_mapsize = env->me_mapsize; 4427 4428 if (newenv && !(flags & MDB_FIXEDMAP)) { 4429 /* mdb_env_map() may grow the datafile. Write the metapages 4430 * first, so the file will be valid if initialization fails. 4431 * Except with FIXEDMAP, since we do not yet know mm_address. 4432 * We could fill in mm_address later, but then a different 4433 * program might end up doing that - one with a memory layout 4434 * and map address which does not suit the main program. 4435 */ 4436 rc = mdb_env_init_meta(env, &meta); 4437 if (rc) 4438 return rc; 4439 newenv = 0; 4440 } 4441 4442 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL); 4443 if (rc) 4444 return rc; 4445 4446 if (newenv) { 4447 if (flags & MDB_FIXEDMAP) 4448 meta.mm_address = env->me_map; 4449 i = mdb_env_init_meta(env, &meta); 4450 if (i != MDB_SUCCESS) { 4451 return i; 4452 } 4453 } 4454 4455 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1; 4456 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2) 4457 - sizeof(indx_t); 4458 #if !(MDB_MAXKEYSIZE) 4459 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db)); 4460 #endif 4461 env->me_maxpg = env->me_mapsize / env->me_psize; 4462 4463 #if MDB_DEBUG 4464 { 4465 MDB_meta *meta = mdb_env_pick_meta(env); 4466 MDB_db *db = &meta->mm_dbs[MAIN_DBI]; 4467 4468 DPRINTF(("opened database version %u, pagesize %u", 4469 meta->mm_version, env->me_psize)); 4470 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1))); 4471 DPRINTF(("depth: %u", db->md_depth)); 4472 DPRINTF(("entries: %"Z"u", db->md_entries)); 4473 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages)); 4474 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages)); 4475 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages)); 4476 DPRINTF(("root: %"Z"u", db->md_root)); 4477 } 4478 #endif 4479 4480 return MDB_SUCCESS; 4481 } 4482 4483 4484 /** Release a reader thread's slot in the reader lock table. 4485 * This function is called automatically when a thread exits. 4486 * @param[in] ptr This points to the slot in the reader lock table. 4487 */ 4488 static void 4489 mdb_env_reader_dest(void *ptr) 4490 { 4491 MDB_reader *reader = ptr; 4492 4493 #ifndef _WIN32 4494 if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */ 4495 #endif 4496 /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */ 4497 reader->mr_pid = 0; 4498 } 4499 4500 #ifdef _WIN32 4501 /** Junk for arranging thread-specific callbacks on Windows. This is 4502 * necessarily platform and compiler-specific. Windows supports up 4503 * to 1088 keys. Let's assume nobody opens more than 64 environments 4504 * in a single process, for now. They can override this if needed. 4505 */ 4506 #ifndef MAX_TLS_KEYS 4507 #define MAX_TLS_KEYS 64 4508 #endif 4509 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS]; 4510 static int mdb_tls_nkeys; 4511 4512 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr) 4513 { 4514 int i; 4515 switch(reason) { 4516 case DLL_PROCESS_ATTACH: break; 4517 case DLL_THREAD_ATTACH: break; 4518 case DLL_THREAD_DETACH: 4519 for (i=0; i<mdb_tls_nkeys; i++) { 4520 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]); 4521 if (r) { 4522 mdb_env_reader_dest(r); 4523 } 4524 } 4525 break; 4526 case DLL_PROCESS_DETACH: break; 4527 } 4528 } 4529 #ifdef __GNUC__ 4530 #ifdef _WIN64 4531 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback; 4532 #else 4533 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback; 4534 #endif 4535 #else 4536 #ifdef _WIN64 4537 /* Force some symbol references. 4538 * _tls_used forces the linker to create the TLS directory if not already done 4539 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol. 4540 */ 4541 #pragma comment(linker, "/INCLUDE:_tls_used") 4542 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp") 4543 #pragma const_seg(".CRT$XLB") 4544 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp; 4545 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback; 4546 #pragma const_seg() 4547 #else /* _WIN32 */ 4548 #pragma comment(linker, "/INCLUDE:__tls_used") 4549 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp") 4550 #pragma data_seg(".CRT$XLB") 4551 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback; 4552 #pragma data_seg() 4553 #endif /* WIN 32/64 */ 4554 #endif /* !__GNUC__ */ 4555 #endif 4556 4557 /** Downgrade the exclusive lock on the region back to shared */ 4558 static int ESECT 4559 mdb_env_share_locks(MDB_env *env, int *excl) 4560 { 4561 int rc = 0; 4562 MDB_meta *meta = mdb_env_pick_meta(env); 4563 4564 env->me_txns->mti_txnid = meta->mm_txnid; 4565 4566 #ifdef _WIN32 4567 { 4568 OVERLAPPED ov; 4569 /* First acquire a shared lock. The Unlock will 4570 * then release the existing exclusive lock. 4571 */ 4572 memset(&ov, 0, sizeof(ov)); 4573 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) { 4574 rc = ErrCode(); 4575 } else { 4576 UnlockFile(env->me_lfd, 0, 0, 1, 0); 4577 *excl = 0; 4578 } 4579 } 4580 #else 4581 { 4582 struct flock lock_info; 4583 /* The shared lock replaces the existing lock */ 4584 memset((void *)&lock_info, 0, sizeof(lock_info)); 4585 lock_info.l_type = F_RDLCK; 4586 lock_info.l_whence = SEEK_SET; 4587 lock_info.l_start = 0; 4588 lock_info.l_len = 1; 4589 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) && 4590 (rc = ErrCode()) == EINTR) ; 4591 *excl = rc ? -1 : 0; /* error may mean we lost the lock */ 4592 } 4593 #endif 4594 4595 return rc; 4596 } 4597 4598 /** Try to get exclusive lock, otherwise shared. 4599 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive. 4600 */ 4601 static int ESECT 4602 mdb_env_excl_lock(MDB_env *env, int *excl) 4603 { 4604 int rc = 0; 4605 #ifdef _WIN32 4606 if (LockFile(env->me_lfd, 0, 0, 1, 0)) { 4607 *excl = 1; 4608 } else { 4609 OVERLAPPED ov; 4610 memset(&ov, 0, sizeof(ov)); 4611 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) { 4612 *excl = 0; 4613 } else { 4614 rc = ErrCode(); 4615 } 4616 } 4617 #else 4618 struct flock lock_info; 4619 memset((void *)&lock_info, 0, sizeof(lock_info)); 4620 lock_info.l_type = F_WRLCK; 4621 lock_info.l_whence = SEEK_SET; 4622 lock_info.l_start = 0; 4623 lock_info.l_len = 1; 4624 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) && 4625 (rc = ErrCode()) == EINTR) ; 4626 if (!rc) { 4627 *excl = 1; 4628 } else 4629 # ifndef MDB_USE_POSIX_MUTEX 4630 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */ 4631 # endif 4632 { 4633 lock_info.l_type = F_RDLCK; 4634 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) && 4635 (rc = ErrCode()) == EINTR) ; 4636 if (rc == 0) 4637 *excl = 0; 4638 } 4639 #endif 4640 return rc; 4641 } 4642 4643 #ifdef MDB_USE_HASH 4644 /* 4645 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code 4646 * 4647 * @(#) Revision: 5.1 4648 * @(#) Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp 4649 * @(#) Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v 4650 * 4651 * http://www.isthe.com/chongo/tech/comp/fnv/index.html 4652 * 4653 *** 4654 * 4655 * Please do not copyright this code. This code is in the public domain. 4656 * 4657 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, 4658 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO 4659 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR 4660 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF 4661 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR 4662 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR 4663 * PERFORMANCE OF THIS SOFTWARE. 4664 * 4665 * By: 4666 * chongo <Landon Curt Noll> /\oo/\ 4667 * http://www.isthe.com/chongo/ 4668 * 4669 * Share and Enjoy! :-) 4670 */ 4671 4672 typedef unsigned long long mdb_hash_t; 4673 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL) 4674 4675 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer 4676 * @param[in] val value to hash 4677 * @param[in] hval initial value for hash 4678 * @return 64 bit hash 4679 * 4680 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the 4681 * hval arg on the first call. 4682 */ 4683 static mdb_hash_t 4684 mdb_hash_val(MDB_val *val, mdb_hash_t hval) 4685 { 4686 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */ 4687 unsigned char *end = s + val->mv_size; 4688 /* 4689 * FNV-1a hash each octet of the string 4690 */ 4691 while (s < end) { 4692 /* xor the bottom with the current octet */ 4693 hval ^= (mdb_hash_t)*s++; 4694 4695 /* multiply by the 64 bit FNV magic prime mod 2^64 */ 4696 hval += (hval << 1) + (hval << 4) + (hval << 5) + 4697 (hval << 7) + (hval << 8) + (hval << 40); 4698 } 4699 /* return our new hash value */ 4700 return hval; 4701 } 4702 4703 /** Hash the string and output the encoded hash. 4704 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with 4705 * very short name limits. We don't care about the encoding being reversible, 4706 * we just want to preserve as many bits of the input as possible in a 4707 * small printable string. 4708 * @param[in] str string to hash 4709 * @param[out] encbuf an array of 11 chars to hold the hash 4710 */ 4711 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~"; 4712 4713 static void ESECT 4714 mdb_pack85(unsigned long l, char *out) 4715 { 4716 int i; 4717 4718 for (i=0; i<5; i++) { 4719 *out++ = mdb_a85[l % 85]; 4720 l /= 85; 4721 } 4722 } 4723 4724 static void ESECT 4725 mdb_hash_enc(MDB_val *val, char *encbuf) 4726 { 4727 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT); 4728 4729 mdb_pack85(h, encbuf); 4730 mdb_pack85(h>>32, encbuf+5); 4731 encbuf[10] = '\0'; 4732 } 4733 #endif 4734 4735 /** Open and/or initialize the lock region for the environment. 4736 * @param[in] env The LMDB environment. 4737 * @param[in] fname Filename + scratch area, from #mdb_fname_init(). 4738 * @param[in] mode The Unix permissions for the file, if we create it. 4739 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive 4740 * @return 0 on success, non-zero on failure. 4741 */ 4742 static int ESECT 4743 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl) 4744 { 4745 #ifdef _WIN32 4746 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT 4747 #else 4748 # define MDB_ERRCODE_ROFS EROFS 4749 #endif 4750 int rc; 4751 off_t size, rsize; 4752 4753 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd); 4754 if (rc) { 4755 /* Omit lockfile if read-only env on read-only filesystem */ 4756 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) { 4757 return MDB_SUCCESS; 4758 } 4759 goto fail; 4760 } 4761 4762 if (!(env->me_flags & MDB_NOTLS)) { 4763 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest); 4764 if (rc) 4765 goto fail; 4766 env->me_flags |= MDB_ENV_TXKEY; 4767 #ifdef _WIN32 4768 /* Windows TLS callbacks need help finding their TLS info. */ 4769 if (mdb_tls_nkeys >= MAX_TLS_KEYS) { 4770 rc = MDB_TLS_FULL; 4771 goto fail; 4772 } 4773 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey; 4774 #endif 4775 } 4776 4777 /* Try to get exclusive lock. If we succeed, then 4778 * nobody is using the lock region and we should initialize it. 4779 */ 4780 if ((rc = mdb_env_excl_lock(env, excl))) goto fail; 4781 4782 #ifdef _WIN32 4783 size = GetFileSize(env->me_lfd, NULL); 4784 #else 4785 size = lseek(env->me_lfd, 0, SEEK_END); 4786 if (size == -1) goto fail_errno; 4787 #endif 4788 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo); 4789 if (size < rsize && *excl > 0) { 4790 #ifdef _WIN32 4791 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize 4792 || !SetEndOfFile(env->me_lfd)) 4793 goto fail_errno; 4794 #else 4795 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno; 4796 #endif 4797 } else { 4798 rsize = size; 4799 size = rsize - sizeof(MDB_txninfo); 4800 env->me_maxreaders = size/sizeof(MDB_reader) + 1; 4801 } 4802 { 4803 #ifdef _WIN32 4804 HANDLE mh; 4805 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE, 4806 0, 0, NULL); 4807 if (!mh) goto fail_errno; 4808 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL); 4809 CloseHandle(mh); 4810 if (!env->me_txns) goto fail_errno; 4811 #else 4812 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED, 4813 env->me_lfd, 0); 4814 if (m == MAP_FAILED) goto fail_errno; 4815 env->me_txns = m; 4816 #endif 4817 } 4818 if (*excl > 0) { 4819 #ifdef _WIN32 4820 BY_HANDLE_FILE_INFORMATION stbuf; 4821 struct { 4822 DWORD volume; 4823 DWORD nhigh; 4824 DWORD nlow; 4825 } idbuf; 4826 MDB_val val; 4827 char encbuf[11]; 4828 4829 if (!mdb_sec_inited) { 4830 InitializeSecurityDescriptor(&mdb_null_sd, 4831 SECURITY_DESCRIPTOR_REVISION); 4832 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE); 4833 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES); 4834 mdb_all_sa.bInheritHandle = FALSE; 4835 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd; 4836 mdb_sec_inited = 1; 4837 } 4838 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno; 4839 idbuf.volume = stbuf.dwVolumeSerialNumber; 4840 idbuf.nhigh = stbuf.nFileIndexHigh; 4841 idbuf.nlow = stbuf.nFileIndexLow; 4842 val.mv_data = &idbuf; 4843 val.mv_size = sizeof(idbuf); 4844 mdb_hash_enc(&val, encbuf); 4845 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf); 4846 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf); 4847 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname); 4848 if (!env->me_rmutex) goto fail_errno; 4849 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname); 4850 if (!env->me_wmutex) goto fail_errno; 4851 #elif defined(MDB_USE_POSIX_SEM) 4852 struct stat stbuf; 4853 struct { 4854 dev_t dev; 4855 ino_t ino; 4856 } idbuf; 4857 MDB_val val; 4858 char encbuf[11]; 4859 4860 #if defined(__NetBSD__) 4861 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */ 4862 #endif 4863 if (fstat(env->me_lfd, &stbuf)) goto fail_errno; 4864 idbuf.dev = stbuf.st_dev; 4865 idbuf.ino = stbuf.st_ino; 4866 val.mv_data = &idbuf; 4867 val.mv_size = sizeof(idbuf); 4868 mdb_hash_enc(&val, encbuf); 4869 #ifdef MDB_SHORT_SEMNAMES 4870 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */ 4871 #endif 4872 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf); 4873 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf); 4874 /* Clean up after a previous run, if needed: Try to 4875 * remove both semaphores before doing anything else. 4876 */ 4877 sem_unlink(env->me_txns->mti_rmname); 4878 sem_unlink(env->me_txns->mti_wmname); 4879 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 4880 O_CREAT|O_EXCL, mode, 1); 4881 if (env->me_rmutex == SEM_FAILED) goto fail_errno; 4882 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 4883 O_CREAT|O_EXCL, mode, 1); 4884 if (env->me_wmutex == SEM_FAILED) goto fail_errno; 4885 #else /* MDB_USE_POSIX_MUTEX: */ 4886 pthread_mutexattr_t mattr; 4887 4888 /* Solaris needs this before initing a robust mutex. Otherwise 4889 * it may skip the init and return EBUSY "seems someone already 4890 * inited" or EINVAL "it was inited differently". 4891 */ 4892 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex)); 4893 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex)); 4894 4895 if ((rc = pthread_mutexattr_init(&mattr))) 4896 goto fail; 4897 4898 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED); 4899 #ifdef MDB_ROBUST_SUPPORTED 4900 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST); 4901 #endif 4902 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr); 4903 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr); 4904 pthread_mutexattr_destroy(&mattr); 4905 if (rc) 4906 goto fail; 4907 #endif /* _WIN32 || MDB_USE_POSIX_SEM */ 4908 4909 env->me_txns->mti_magic = MDB_MAGIC; 4910 env->me_txns->mti_format = MDB_LOCK_FORMAT; 4911 env->me_txns->mti_txnid = 0; 4912 env->me_txns->mti_numreaders = 0; 4913 4914 } else { 4915 if (env->me_txns->mti_magic != MDB_MAGIC) { 4916 DPUTS("lock region has invalid magic"); 4917 rc = MDB_INVALID; 4918 goto fail; 4919 } 4920 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) { 4921 DPRINTF(("lock region has format+version 0x%x, expected 0x%x", 4922 env->me_txns->mti_format, MDB_LOCK_FORMAT)); 4923 rc = MDB_VERSION_MISMATCH; 4924 goto fail; 4925 } 4926 rc = ErrCode(); 4927 if (rc && rc != EACCES && rc != EAGAIN) { 4928 goto fail; 4929 } 4930 #ifdef _WIN32 4931 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname); 4932 if (!env->me_rmutex) goto fail_errno; 4933 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname); 4934 if (!env->me_wmutex) goto fail_errno; 4935 #elif defined(MDB_USE_POSIX_SEM) 4936 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0); 4937 if (env->me_rmutex == SEM_FAILED) goto fail_errno; 4938 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0); 4939 if (env->me_wmutex == SEM_FAILED) goto fail_errno; 4940 #endif 4941 } 4942 return MDB_SUCCESS; 4943 4944 fail_errno: 4945 rc = ErrCode(); 4946 fail: 4947 return rc; 4948 } 4949 4950 /** Only a subset of the @ref mdb_env flags can be changed 4951 * at runtime. Changing other flags requires closing the 4952 * environment and re-opening it with the new flags. 4953 */ 4954 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT) 4955 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \ 4956 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD) 4957 4958 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS) 4959 # error "Persistent DB flags & env flags overlap, but both go in mm_flags" 4960 #endif 4961 4962 int ESECT 4963 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode) 4964 { 4965 int rc, excl = -1; 4966 MDB_name fname; 4967 4968 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS))) 4969 return EINVAL; 4970 4971 flags |= env->me_flags; 4972 4973 rc = mdb_fname_init(path, flags, &fname); 4974 if (rc) 4975 return rc; 4976 4977 if (flags & MDB_RDONLY) { 4978 /* silently ignore WRITEMAP when we're only getting read access */ 4979 flags &= ~MDB_WRITEMAP; 4980 } else { 4981 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) && 4982 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2))))) 4983 rc = ENOMEM; 4984 } 4985 env->me_flags = flags |= MDB_ENV_ACTIVE; 4986 if (rc) 4987 goto leave; 4988 4989 env->me_path = strdup(path); 4990 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx)); 4991 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t)); 4992 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int)); 4993 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) { 4994 rc = ENOMEM; 4995 goto leave; 4996 } 4997 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */ 4998 4999 /* For RDONLY, get lockfile after we know datafile exists */ 5000 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) { 5001 rc = mdb_env_setup_locks(env, &fname, mode, &excl); 5002 if (rc) 5003 goto leave; 5004 } 5005 5006 rc = mdb_fopen(env, &fname, 5007 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR, 5008 mode, &env->me_fd); 5009 if (rc) 5010 goto leave; 5011 5012 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) { 5013 rc = mdb_env_setup_locks(env, &fname, mode, &excl); 5014 if (rc) 5015 goto leave; 5016 } 5017 5018 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) { 5019 if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) { 5020 /* Synchronous fd for meta writes. Needed even with 5021 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset. 5022 */ 5023 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd); 5024 if (rc) 5025 goto leave; 5026 } 5027 DPRINTF(("opened dbenv %p", (void *) env)); 5028 if (excl > 0) { 5029 rc = mdb_env_share_locks(env, &excl); 5030 if (rc) 5031 goto leave; 5032 } 5033 if (!(flags & MDB_RDONLY)) { 5034 MDB_txn *txn; 5035 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs * 5036 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1); 5037 if ((env->me_pbuf = calloc(1, env->me_psize)) && 5038 (txn = calloc(1, size))) 5039 { 5040 txn->mt_dbs = (MDB_db *)((char *)txn + tsize); 5041 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs); 5042 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs); 5043 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs); 5044 txn->mt_env = env; 5045 txn->mt_dbxs = env->me_dbxs; 5046 txn->mt_flags = MDB_TXN_FINISHED; 5047 env->me_txn0 = txn; 5048 } else { 5049 rc = ENOMEM; 5050 } 5051 } 5052 } 5053 5054 leave: 5055 if (rc) { 5056 mdb_env_close0(env, excl); 5057 } 5058 mdb_fname_destroy(fname); 5059 return rc; 5060 } 5061 5062 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */ 5063 static void ESECT 5064 mdb_env_close0(MDB_env *env, int excl) 5065 { 5066 int i; 5067 5068 if (!(env->me_flags & MDB_ENV_ACTIVE)) 5069 return; 5070 5071 /* Doing this here since me_dbxs may not exist during mdb_env_close */ 5072 if (env->me_dbxs) { 5073 for (i = env->me_maxdbs; --i >= CORE_DBS; ) 5074 free(env->me_dbxs[i].md_name.mv_data); 5075 free(env->me_dbxs); 5076 } 5077 5078 free(env->me_pbuf); 5079 free(env->me_dbiseqs); 5080 free(env->me_dbflags); 5081 free(env->me_path); 5082 free(env->me_dirty_list); 5083 free(env->me_txn0); 5084 mdb_midl_free(env->me_free_pgs); 5085 5086 if (env->me_flags & MDB_ENV_TXKEY) { 5087 pthread_key_delete(env->me_txkey); 5088 #ifdef _WIN32 5089 /* Delete our key from the global list */ 5090 for (i=0; i<mdb_tls_nkeys; i++) 5091 if (mdb_tls_keys[i] == env->me_txkey) { 5092 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1]; 5093 mdb_tls_nkeys--; 5094 break; 5095 } 5096 #endif 5097 } 5098 5099 if (env->me_map) { 5100 munmap(env->me_map, env->me_mapsize); 5101 } 5102 if (env->me_mfd != INVALID_HANDLE_VALUE) 5103 (void) close(env->me_mfd); 5104 if (env->me_fd != INVALID_HANDLE_VALUE) 5105 (void) close(env->me_fd); 5106 if (env->me_txns) { 5107 MDB_PID_T pid = getpid(); 5108 /* Clearing readers is done in this function because 5109 * me_txkey with its destructor must be disabled first. 5110 * 5111 * We skip the the reader mutex, so we touch only 5112 * data owned by this process (me_close_readers and 5113 * our readers), and clear each reader atomically. 5114 */ 5115 for (i = env->me_close_readers; --i >= 0; ) 5116 if (env->me_txns->mti_readers[i].mr_pid == pid) 5117 env->me_txns->mti_readers[i].mr_pid = 0; 5118 #ifdef _WIN32 5119 if (env->me_rmutex) { 5120 CloseHandle(env->me_rmutex); 5121 if (env->me_wmutex) CloseHandle(env->me_wmutex); 5122 } 5123 /* Windows automatically destroys the mutexes when 5124 * the last handle closes. 5125 */ 5126 #elif defined(MDB_USE_POSIX_SEM) 5127 if (env->me_rmutex != SEM_FAILED) { 5128 sem_close(env->me_rmutex); 5129 if (env->me_wmutex != SEM_FAILED) 5130 sem_close(env->me_wmutex); 5131 /* If we have the filelock: If we are the 5132 * only remaining user, clean up semaphores. 5133 */ 5134 if (excl == 0) 5135 mdb_env_excl_lock(env, &excl); 5136 if (excl > 0) { 5137 sem_unlink(env->me_txns->mti_rmname); 5138 sem_unlink(env->me_txns->mti_wmname); 5139 } 5140 } 5141 #elif defined(MDB_ROBUST_SUPPORTED) 5142 /* If we have the filelock: If we are the 5143 * only remaining user, clean up robust 5144 * mutexes. 5145 */ 5146 if (excl == 0) 5147 mdb_env_excl_lock(env, &excl); 5148 if (excl > 0) { 5149 pthread_mutex_destroy(env->me_txns->mti_rmutex); 5150 pthread_mutex_destroy(env->me_txns->mti_wmutex); 5151 } 5152 #endif 5153 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo)); 5154 } 5155 if (env->me_lfd != INVALID_HANDLE_VALUE) { 5156 #ifdef _WIN32 5157 if (excl >= 0) { 5158 /* Unlock the lockfile. Windows would have unlocked it 5159 * after closing anyway, but not necessarily at once. 5160 */ 5161 UnlockFile(env->me_lfd, 0, 0, 1, 0); 5162 } 5163 #endif 5164 (void) close(env->me_lfd); 5165 } 5166 5167 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY); 5168 } 5169 5170 void ESECT 5171 mdb_env_close(MDB_env *env) 5172 { 5173 MDB_page *dp; 5174 5175 if (env == NULL) 5176 return; 5177 5178 VGMEMP_DESTROY(env); 5179 while ((dp = env->me_dpages) != NULL) { 5180 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next)); 5181 env->me_dpages = dp->mp_next; 5182 free(dp); 5183 } 5184 5185 mdb_env_close0(env, 0); 5186 free(env); 5187 } 5188 5189 /** Compare two items pointing at aligned size_t's */ 5190 static int 5191 mdb_cmp_long(const MDB_val *a, const MDB_val *b) 5192 { 5193 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 : 5194 *(size_t *)a->mv_data > *(size_t *)b->mv_data; 5195 } 5196 5197 /** Compare two items pointing at aligned unsigned int's. 5198 * 5199 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp, 5200 * but #mdb_cmp_clong() is called instead if the data type is size_t. 5201 */ 5202 static int 5203 mdb_cmp_int(const MDB_val *a, const MDB_val *b) 5204 { 5205 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 : 5206 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data; 5207 } 5208 5209 /** Compare two items pointing at unsigned ints of unknown alignment. 5210 * Nodes and keys are guaranteed to be 2-byte aligned. 5211 */ 5212 static int 5213 mdb_cmp_cint(const MDB_val *a, const MDB_val *b) 5214 { 5215 #if BYTE_ORDER == LITTLE_ENDIAN 5216 unsigned short *u, *c; 5217 int x; 5218 5219 u = (unsigned short *) ((char *) a->mv_data + a->mv_size); 5220 c = (unsigned short *) ((char *) b->mv_data + a->mv_size); 5221 do { 5222 x = *--u - *--c; 5223 } while(!x && u > (unsigned short *)a->mv_data); 5224 return x; 5225 #else 5226 unsigned short *u, *c, *end; 5227 int x; 5228 5229 end = (unsigned short *) ((char *) a->mv_data + a->mv_size); 5230 u = (unsigned short *)a->mv_data; 5231 c = (unsigned short *)b->mv_data; 5232 do { 5233 x = *u++ - *c++; 5234 } while(!x && u < end); 5235 return x; 5236 #endif 5237 } 5238 5239 /** Compare two items lexically */ 5240 static int 5241 mdb_cmp_memn(const MDB_val *a, const MDB_val *b) 5242 { 5243 int diff; 5244 ssize_t len_diff; 5245 unsigned int len; 5246 5247 len = a->mv_size; 5248 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size; 5249 if (len_diff > 0) { 5250 len = b->mv_size; 5251 len_diff = 1; 5252 } 5253 5254 diff = memcmp(a->mv_data, b->mv_data, len); 5255 return diff ? diff : len_diff<0 ? -1 : len_diff; 5256 } 5257 5258 /** Compare two items in reverse byte order */ 5259 static int 5260 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b) 5261 { 5262 const unsigned char *p1, *p2, *p1_lim; 5263 ssize_t len_diff; 5264 int diff; 5265 5266 p1_lim = (const unsigned char *)a->mv_data; 5267 p1 = (const unsigned char *)a->mv_data + a->mv_size; 5268 p2 = (const unsigned char *)b->mv_data + b->mv_size; 5269 5270 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size; 5271 if (len_diff > 0) { 5272 p1_lim += len_diff; 5273 len_diff = 1; 5274 } 5275 5276 while (p1 > p1_lim) { 5277 diff = *--p1 - *--p2; 5278 if (diff) 5279 return diff; 5280 } 5281 return len_diff<0 ? -1 : len_diff; 5282 } 5283 5284 /** Search for key within a page, using binary search. 5285 * Returns the smallest entry larger or equal to the key. 5286 * If exactp is non-null, stores whether the found entry was an exact match 5287 * in *exactp (1 or 0). 5288 * Updates the cursor index with the index of the found entry. 5289 * If no entry larger or equal to the key is found, returns NULL. 5290 */ 5291 static MDB_node * 5292 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp) 5293 { 5294 unsigned int i = 0, nkeys; 5295 int low, high; 5296 int rc = 0; 5297 MDB_page *mp = mc->mc_pg[mc->mc_top]; 5298 MDB_node *node = NULL; 5299 MDB_val nodekey; 5300 MDB_cmp_func *cmp; 5301 DKBUF; 5302 5303 nkeys = NUMKEYS(mp); 5304 5305 DPRINTF(("searching %u keys in %s %spage %"Z"u", 5306 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "", 5307 mdb_dbg_pgno(mp))); 5308 5309 low = IS_LEAF(mp) ? 0 : 1; 5310 high = nkeys - 1; 5311 cmp = mc->mc_dbx->md_cmp; 5312 5313 /* Branch pages have no data, so if using integer keys, 5314 * alignment is guaranteed. Use faster mdb_cmp_int. 5315 */ 5316 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) { 5317 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t)) 5318 cmp = mdb_cmp_long; 5319 else 5320 cmp = mdb_cmp_int; 5321 } 5322 5323 if (IS_LEAF2(mp)) { 5324 nodekey.mv_size = mc->mc_db->md_pad; 5325 node = NODEPTR(mp, 0); /* fake */ 5326 while (low <= high) { 5327 i = (low + high) >> 1; 5328 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size); 5329 rc = cmp(key, &nodekey); 5330 DPRINTF(("found leaf index %u [%s], rc = %i", 5331 i, DKEY(&nodekey), rc)); 5332 if (rc == 0) 5333 break; 5334 if (rc > 0) 5335 low = i + 1; 5336 else 5337 high = i - 1; 5338 } 5339 } else { 5340 while (low <= high) { 5341 i = (low + high) >> 1; 5342 5343 node = NODEPTR(mp, i); 5344 nodekey.mv_size = NODEKSZ(node); 5345 nodekey.mv_data = NODEKEY(node); 5346 5347 rc = cmp(key, &nodekey); 5348 #if MDB_DEBUG 5349 if (IS_LEAF(mp)) 5350 DPRINTF(("found leaf index %u [%s], rc = %i", 5351 i, DKEY(&nodekey), rc)); 5352 else 5353 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i", 5354 i, DKEY(&nodekey), NODEPGNO(node), rc)); 5355 #endif 5356 if (rc == 0) 5357 break; 5358 if (rc > 0) 5359 low = i + 1; 5360 else 5361 high = i - 1; 5362 } 5363 } 5364 5365 if (rc > 0) { /* Found entry is less than the key. */ 5366 i++; /* Skip to get the smallest entry larger than key. */ 5367 if (!IS_LEAF2(mp)) 5368 node = NODEPTR(mp, i); 5369 } 5370 if (exactp) 5371 *exactp = (rc == 0 && nkeys > 0); 5372 /* store the key index */ 5373 mc->mc_ki[mc->mc_top] = i; 5374 if (i >= nkeys) 5375 /* There is no entry larger or equal to the key. */ 5376 return NULL; 5377 5378 /* nodeptr is fake for LEAF2 */ 5379 return node; 5380 } 5381 5382 #if 0 5383 static void 5384 mdb_cursor_adjust(MDB_cursor *mc, func) 5385 { 5386 MDB_cursor *m2; 5387 5388 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) { 5389 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) { 5390 func(mc, m2); 5391 } 5392 } 5393 } 5394 #endif 5395 5396 /** Pop a page off the top of the cursor's stack. */ 5397 static void 5398 mdb_cursor_pop(MDB_cursor *mc) 5399 { 5400 if (mc->mc_snum) { 5401 DPRINTF(("popping page %"Z"u off db %d cursor %p", 5402 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc)); 5403 5404 mc->mc_snum--; 5405 if (mc->mc_snum) { 5406 mc->mc_top--; 5407 } else { 5408 mc->mc_flags &= ~C_INITIALIZED; 5409 } 5410 } 5411 } 5412 5413 /** Push a page onto the top of the cursor's stack. 5414 * Set #MDB_TXN_ERROR on failure. 5415 */ 5416 static int 5417 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp) 5418 { 5419 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno, 5420 DDBI(mc), (void *) mc)); 5421 5422 if (mc->mc_snum >= CURSOR_STACK) { 5423 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 5424 return MDB_CURSOR_FULL; 5425 } 5426 5427 mc->mc_top = mc->mc_snum++; 5428 mc->mc_pg[mc->mc_top] = mp; 5429 mc->mc_ki[mc->mc_top] = 0; 5430 5431 return MDB_SUCCESS; 5432 } 5433 5434 /** Find the address of the page corresponding to a given page number. 5435 * Set #MDB_TXN_ERROR on failure. 5436 * @param[in] mc the cursor accessing the page. 5437 * @param[in] pgno the page number for the page to retrieve. 5438 * @param[out] ret address of a pointer where the page's address will be stored. 5439 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page. 5440 * @return 0 on success, non-zero on failure. 5441 */ 5442 static int 5443 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl) 5444 { 5445 MDB_txn *txn = mc->mc_txn; 5446 MDB_env *env = txn->mt_env; 5447 MDB_page *p = NULL; 5448 int level; 5449 5450 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) { 5451 MDB_txn *tx2 = txn; 5452 level = 1; 5453 do { 5454 MDB_ID2L dl = tx2->mt_u.dirty_list; 5455 unsigned x; 5456 /* Spilled pages were dirtied in this txn and flushed 5457 * because the dirty list got full. Bring this page 5458 * back in from the map (but don't unspill it here, 5459 * leave that unless page_touch happens again). 5460 */ 5461 if (tx2->mt_spill_pgs) { 5462 MDB_ID pn = pgno << 1; 5463 x = mdb_midl_search(tx2->mt_spill_pgs, pn); 5464 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) { 5465 p = (MDB_page *)(env->me_map + env->me_psize * pgno); 5466 goto done; 5467 } 5468 } 5469 if (dl[0].mid) { 5470 unsigned x = mdb_mid2l_search(dl, pgno); 5471 if (x <= dl[0].mid && dl[x].mid == pgno) { 5472 p = dl[x].mptr; 5473 goto done; 5474 } 5475 } 5476 level++; 5477 } while ((tx2 = tx2->mt_parent) != NULL); 5478 } 5479 5480 if (pgno < txn->mt_next_pgno) { 5481 level = 0; 5482 p = (MDB_page *)(env->me_map + env->me_psize * pgno); 5483 } else { 5484 DPRINTF(("page %"Z"u not found", pgno)); 5485 txn->mt_flags |= MDB_TXN_ERROR; 5486 return MDB_PAGE_NOTFOUND; 5487 } 5488 5489 done: 5490 *ret = p; 5491 if (lvl) 5492 *lvl = level; 5493 return MDB_SUCCESS; 5494 } 5495 5496 /** Finish #mdb_page_search() / #mdb_page_search_lowest(). 5497 * The cursor is at the root page, set up the rest of it. 5498 */ 5499 static int 5500 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags) 5501 { 5502 MDB_page *mp = mc->mc_pg[mc->mc_top]; 5503 int rc; 5504 DKBUF; 5505 5506 while (IS_BRANCH(mp)) { 5507 MDB_node *node; 5508 indx_t i; 5509 5510 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp))); 5511 /* Don't assert on branch pages in the FreeDB. We can get here 5512 * while in the process of rebalancing a FreeDB branch page; we must 5513 * let that proceed. ITS#8336 5514 */ 5515 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1); 5516 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0)))); 5517 5518 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) { 5519 i = 0; 5520 if (flags & MDB_PS_LAST) { 5521 i = NUMKEYS(mp) - 1; 5522 /* if already init'd, see if we're already in right place */ 5523 if (mc->mc_flags & C_INITIALIZED) { 5524 if (mc->mc_ki[mc->mc_top] == i) { 5525 mc->mc_top = mc->mc_snum++; 5526 mp = mc->mc_pg[mc->mc_top]; 5527 goto ready; 5528 } 5529 } 5530 } 5531 } else { 5532 int exact; 5533 node = mdb_node_search(mc, key, &exact); 5534 if (node == NULL) 5535 i = NUMKEYS(mp) - 1; 5536 else { 5537 i = mc->mc_ki[mc->mc_top]; 5538 if (!exact) { 5539 mdb_cassert(mc, i > 0); 5540 i--; 5541 } 5542 } 5543 DPRINTF(("following index %u for key [%s]", i, DKEY(key))); 5544 } 5545 5546 mdb_cassert(mc, i < NUMKEYS(mp)); 5547 node = NODEPTR(mp, i); 5548 5549 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0) 5550 return rc; 5551 5552 mc->mc_ki[mc->mc_top] = i; 5553 if ((rc = mdb_cursor_push(mc, mp))) 5554 return rc; 5555 5556 ready: 5557 if (flags & MDB_PS_MODIFY) { 5558 if ((rc = mdb_page_touch(mc)) != 0) 5559 return rc; 5560 mp = mc->mc_pg[mc->mc_top]; 5561 } 5562 } 5563 5564 if (!IS_LEAF(mp)) { 5565 DPRINTF(("internal error, index points to a %02X page!?", 5566 mp->mp_flags)); 5567 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 5568 return MDB_CORRUPTED; 5569 } 5570 5571 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno, 5572 key ? DKEY(key) : "null")); 5573 mc->mc_flags |= C_INITIALIZED; 5574 mc->mc_flags &= ~C_EOF; 5575 5576 return MDB_SUCCESS; 5577 } 5578 5579 /** Search for the lowest key under the current branch page. 5580 * This just bypasses a NUMKEYS check in the current page 5581 * before calling mdb_page_search_root(), because the callers 5582 * are all in situations where the current page is known to 5583 * be underfilled. 5584 */ 5585 static int 5586 mdb_page_search_lowest(MDB_cursor *mc) 5587 { 5588 MDB_page *mp = mc->mc_pg[mc->mc_top]; 5589 MDB_node *node = NODEPTR(mp, 0); 5590 int rc; 5591 5592 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0) 5593 return rc; 5594 5595 mc->mc_ki[mc->mc_top] = 0; 5596 if ((rc = mdb_cursor_push(mc, mp))) 5597 return rc; 5598 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST); 5599 } 5600 5601 /** Search for the page a given key should be in. 5602 * Push it and its parent pages on the cursor stack. 5603 * @param[in,out] mc the cursor for this operation. 5604 * @param[in] key the key to search for, or NULL for first/last page. 5605 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB 5606 * are touched (updated with new page numbers). 5607 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf. 5608 * This is used by #mdb_cursor_first() and #mdb_cursor_last(). 5609 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups. 5610 * @return 0 on success, non-zero on failure. 5611 */ 5612 static int 5613 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags) 5614 { 5615 int rc; 5616 pgno_t root; 5617 5618 /* Make sure the txn is still viable, then find the root from 5619 * the txn's db table and set it as the root of the cursor's stack. 5620 */ 5621 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) { 5622 DPUTS("transaction may not be used now"); 5623 return MDB_BAD_TXN; 5624 } else { 5625 /* Make sure we're using an up-to-date root */ 5626 if (*mc->mc_dbflag & DB_STALE) { 5627 MDB_cursor mc2; 5628 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi)) 5629 return MDB_BAD_DBI; 5630 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL); 5631 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0); 5632 if (rc) 5633 return rc; 5634 { 5635 MDB_val data; 5636 int exact = 0; 5637 uint16_t flags; 5638 MDB_node *leaf = mdb_node_search(&mc2, 5639 &mc->mc_dbx->md_name, &exact); 5640 if (!exact) 5641 return MDB_NOTFOUND; 5642 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA) 5643 return MDB_INCOMPATIBLE; /* not a named DB */ 5644 rc = mdb_node_read(&mc2, leaf, &data); 5645 if (rc) 5646 return rc; 5647 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)), 5648 sizeof(uint16_t)); 5649 /* The txn may not know this DBI, or another process may 5650 * have dropped and recreated the DB with other flags. 5651 */ 5652 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags) 5653 return MDB_INCOMPATIBLE; 5654 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db)); 5655 } 5656 *mc->mc_dbflag &= ~DB_STALE; 5657 } 5658 root = mc->mc_db->md_root; 5659 5660 if (root == P_INVALID) { /* Tree is empty. */ 5661 DPUTS("tree is empty"); 5662 return MDB_NOTFOUND; 5663 } 5664 } 5665 5666 mdb_cassert(mc, root > 1); 5667 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) 5668 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0) 5669 return rc; 5670 5671 mc->mc_snum = 1; 5672 mc->mc_top = 0; 5673 5674 DPRINTF(("db %d root page %"Z"u has flags 0x%X", 5675 DDBI(mc), root, mc->mc_pg[0]->mp_flags)); 5676 5677 if (flags & MDB_PS_MODIFY) { 5678 if ((rc = mdb_page_touch(mc))) 5679 return rc; 5680 } 5681 5682 if (flags & MDB_PS_ROOTONLY) 5683 return MDB_SUCCESS; 5684 5685 return mdb_page_search_root(mc, key, flags); 5686 } 5687 5688 static int 5689 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp) 5690 { 5691 MDB_txn *txn = mc->mc_txn; 5692 pgno_t pg = mp->mp_pgno; 5693 unsigned x = 0, ovpages = mp->mp_pages; 5694 MDB_env *env = txn->mt_env; 5695 MDB_IDL sl = txn->mt_spill_pgs; 5696 MDB_ID pn = pg << 1; 5697 int rc; 5698 5699 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages)); 5700 /* If the page is dirty or on the spill list we just acquired it, 5701 * so we should give it back to our current free list, if any. 5702 * Otherwise put it onto the list of pages we freed in this txn. 5703 * 5704 * Won't create me_pghead: me_pglast must be inited along with it. 5705 * Unsupported in nested txns: They would need to hide the page 5706 * range in ancestor txns' dirty and spilled lists. 5707 */ 5708 if (env->me_pghead && 5709 !txn->mt_parent && 5710 ((mp->mp_flags & P_DIRTY) || 5711 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn))) 5712 { 5713 unsigned i, j; 5714 pgno_t *mop; 5715 MDB_ID2 *dl, ix, iy; 5716 rc = mdb_midl_need(&env->me_pghead, ovpages); 5717 if (rc) 5718 return rc; 5719 if (!(mp->mp_flags & P_DIRTY)) { 5720 /* This page is no longer spilled */ 5721 if (x == sl[0]) 5722 sl[0]--; 5723 else 5724 sl[x] |= 1; 5725 goto release; 5726 } 5727 /* Remove from dirty list */ 5728 dl = txn->mt_u.dirty_list; 5729 x = dl[0].mid--; 5730 for (ix = dl[x]; ix.mptr != mp; ix = iy) { 5731 if (x > 1) { 5732 x--; 5733 iy = dl[x]; 5734 dl[x] = ix; 5735 } else { 5736 mdb_cassert(mc, x > 1); 5737 j = ++(dl[0].mid); 5738 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */ 5739 txn->mt_flags |= MDB_TXN_ERROR; 5740 return MDB_CORRUPTED; 5741 } 5742 } 5743 txn->mt_dirty_room++; 5744 if (!(env->me_flags & MDB_WRITEMAP)) 5745 mdb_dpage_free(env, mp); 5746 release: 5747 /* Insert in me_pghead */ 5748 mop = env->me_pghead; 5749 j = mop[0] + ovpages; 5750 for (i = mop[0]; i && mop[i] < pg; i--) 5751 mop[j--] = mop[i]; 5752 while (j>i) 5753 mop[j--] = pg++; 5754 mop[0] += ovpages; 5755 } else { 5756 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages); 5757 if (rc) 5758 return rc; 5759 } 5760 mc->mc_db->md_overflow_pages -= ovpages; 5761 return 0; 5762 } 5763 5764 /** Return the data associated with a given node. 5765 * @param[in] mc The cursor for this operation. 5766 * @param[in] leaf The node being read. 5767 * @param[out] data Updated to point to the node's data. 5768 * @return 0 on success, non-zero on failure. 5769 */ 5770 static int 5771 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data) 5772 { 5773 MDB_page *omp; /* overflow page */ 5774 pgno_t pgno; 5775 int rc; 5776 5777 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) { 5778 data->mv_size = NODEDSZ(leaf); 5779 data->mv_data = NODEDATA(leaf); 5780 return MDB_SUCCESS; 5781 } 5782 5783 /* Read overflow data. 5784 */ 5785 data->mv_size = NODEDSZ(leaf); 5786 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno)); 5787 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) { 5788 DPRINTF(("read overflow page %"Z"u failed", pgno)); 5789 return rc; 5790 } 5791 data->mv_data = METADATA(omp); 5792 5793 return MDB_SUCCESS; 5794 } 5795 5796 int 5797 mdb_get(MDB_txn *txn, MDB_dbi dbi, 5798 MDB_val *key, MDB_val *data) 5799 { 5800 MDB_cursor mc; 5801 MDB_xcursor mx; 5802 int exact = 0; 5803 DKBUF; 5804 5805 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key))); 5806 5807 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 5808 return EINVAL; 5809 5810 if (txn->mt_flags & MDB_TXN_BLOCKED) 5811 return MDB_BAD_TXN; 5812 5813 mdb_cursor_init(&mc, txn, dbi, &mx); 5814 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact); 5815 } 5816 5817 /** Find a sibling for a page. 5818 * Replaces the page at the top of the cursor's stack with the 5819 * specified sibling, if one exists. 5820 * @param[in] mc The cursor for this operation. 5821 * @param[in] move_right Non-zero if the right sibling is requested, 5822 * otherwise the left sibling. 5823 * @return 0 on success, non-zero on failure. 5824 */ 5825 static int 5826 mdb_cursor_sibling(MDB_cursor *mc, int move_right) 5827 { 5828 int rc; 5829 MDB_node *indx; 5830 MDB_page *mp; 5831 5832 if (mc->mc_snum < 2) { 5833 return MDB_NOTFOUND; /* root has no siblings */ 5834 } 5835 5836 mdb_cursor_pop(mc); 5837 DPRINTF(("parent page is page %"Z"u, index %u", 5838 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top])); 5839 5840 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top])) 5841 : (mc->mc_ki[mc->mc_top] == 0)) { 5842 DPRINTF(("no more keys left, moving to %s sibling", 5843 move_right ? "right" : "left")); 5844 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) { 5845 /* undo cursor_pop before returning */ 5846 mc->mc_top++; 5847 mc->mc_snum++; 5848 return rc; 5849 } 5850 } else { 5851 if (move_right) 5852 mc->mc_ki[mc->mc_top]++; 5853 else 5854 mc->mc_ki[mc->mc_top]--; 5855 DPRINTF(("just moving to %s index key %u", 5856 move_right ? "right" : "left", mc->mc_ki[mc->mc_top])); 5857 } 5858 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top])); 5859 5860 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 5861 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) { 5862 /* mc will be inconsistent if caller does mc_snum++ as above */ 5863 mc->mc_flags &= ~(C_INITIALIZED|C_EOF); 5864 return rc; 5865 } 5866 5867 mdb_cursor_push(mc, mp); 5868 if (!move_right) 5869 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1; 5870 5871 return MDB_SUCCESS; 5872 } 5873 5874 /** Move the cursor to the next data item. */ 5875 static int 5876 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op) 5877 { 5878 MDB_page *mp; 5879 MDB_node *leaf; 5880 int rc; 5881 5882 if ((mc->mc_flags & C_DEL && op == MDB_NEXT_DUP)) 5883 return MDB_NOTFOUND; 5884 5885 if (!(mc->mc_flags & C_INITIALIZED)) 5886 return mdb_cursor_first(mc, key, data); 5887 5888 mp = mc->mc_pg[mc->mc_top]; 5889 5890 if (mc->mc_flags & C_EOF) { 5891 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mp)-1) 5892 return MDB_NOTFOUND; 5893 mc->mc_flags ^= C_EOF; 5894 } 5895 5896 if (mc->mc_db->md_flags & MDB_DUPSORT) { 5897 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 5898 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 5899 if (op == MDB_NEXT || op == MDB_NEXT_DUP) { 5900 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT); 5901 if (op != MDB_NEXT || rc != MDB_NOTFOUND) { 5902 if (rc == MDB_SUCCESS) 5903 MDB_GET_KEY(leaf, key); 5904 return rc; 5905 } 5906 } 5907 } else { 5908 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 5909 if (op == MDB_NEXT_DUP) 5910 return MDB_NOTFOUND; 5911 } 5912 } 5913 5914 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", 5915 mdb_dbg_pgno(mp), (void *) mc)); 5916 if (mc->mc_flags & C_DEL) { 5917 mc->mc_flags ^= C_DEL; 5918 goto skip; 5919 } 5920 5921 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) { 5922 DPUTS("=====> move to next sibling page"); 5923 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) { 5924 mc->mc_flags |= C_EOF; 5925 return rc; 5926 } 5927 mp = mc->mc_pg[mc->mc_top]; 5928 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top])); 5929 } else 5930 mc->mc_ki[mc->mc_top]++; 5931 5932 skip: 5933 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u", 5934 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top])); 5935 5936 if (IS_LEAF2(mp)) { 5937 key->mv_size = mc->mc_db->md_pad; 5938 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size); 5939 return MDB_SUCCESS; 5940 } 5941 5942 mdb_cassert(mc, IS_LEAF(mp)); 5943 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 5944 5945 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 5946 mdb_xcursor_init1(mc, leaf); 5947 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); 5948 if (rc != MDB_SUCCESS) 5949 return rc; 5950 } else if (data) { 5951 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 5952 return rc; 5953 } 5954 5955 MDB_GET_KEY(leaf, key); 5956 return MDB_SUCCESS; 5957 } 5958 5959 /** Move the cursor to the previous data item. */ 5960 static int 5961 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op) 5962 { 5963 MDB_page *mp; 5964 MDB_node *leaf; 5965 int rc; 5966 5967 if (!(mc->mc_flags & C_INITIALIZED)) { 5968 rc = mdb_cursor_last(mc, key, data); 5969 if (rc) 5970 return rc; 5971 mc->mc_ki[mc->mc_top]++; 5972 } 5973 5974 mp = mc->mc_pg[mc->mc_top]; 5975 5976 if ((mc->mc_db->md_flags & MDB_DUPSORT) && 5977 mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) { 5978 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 5979 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 5980 if (op == MDB_PREV || op == MDB_PREV_DUP) { 5981 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV); 5982 if (op != MDB_PREV || rc != MDB_NOTFOUND) { 5983 if (rc == MDB_SUCCESS) { 5984 MDB_GET_KEY(leaf, key); 5985 mc->mc_flags &= ~C_EOF; 5986 } 5987 return rc; 5988 } 5989 } 5990 } else { 5991 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 5992 if (op == MDB_PREV_DUP) 5993 return MDB_NOTFOUND; 5994 } 5995 } 5996 5997 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", 5998 mdb_dbg_pgno(mp), (void *) mc)); 5999 6000 mc->mc_flags &= ~(C_EOF|C_DEL); 6001 6002 if (mc->mc_ki[mc->mc_top] == 0) { 6003 DPUTS("=====> move to prev sibling page"); 6004 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) { 6005 return rc; 6006 } 6007 mp = mc->mc_pg[mc->mc_top]; 6008 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1; 6009 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top])); 6010 } else 6011 mc->mc_ki[mc->mc_top]--; 6012 6013 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u", 6014 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top])); 6015 6016 if (!IS_LEAF(mp)) 6017 return MDB_CORRUPTED; 6018 6019 if (IS_LEAF2(mp)) { 6020 key->mv_size = mc->mc_db->md_pad; 6021 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size); 6022 return MDB_SUCCESS; 6023 } 6024 6025 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 6026 6027 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6028 mdb_xcursor_init1(mc, leaf); 6029 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL); 6030 if (rc != MDB_SUCCESS) 6031 return rc; 6032 } else if (data) { 6033 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 6034 return rc; 6035 } 6036 6037 MDB_GET_KEY(leaf, key); 6038 return MDB_SUCCESS; 6039 } 6040 6041 /** Set the cursor on a specific data item. */ 6042 static int 6043 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, 6044 MDB_cursor_op op, int *exactp) 6045 { 6046 int rc; 6047 MDB_page *mp; 6048 MDB_node *leaf = NULL; 6049 DKBUF; 6050 6051 if (key->mv_size == 0) 6052 return MDB_BAD_VALSIZE; 6053 6054 if (mc->mc_xcursor) 6055 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 6056 6057 /* See if we're already on the right page */ 6058 if (mc->mc_flags & C_INITIALIZED) { 6059 MDB_val nodekey; 6060 6061 mp = mc->mc_pg[mc->mc_top]; 6062 if (!NUMKEYS(mp)) { 6063 mc->mc_ki[mc->mc_top] = 0; 6064 return MDB_NOTFOUND; 6065 } 6066 if (mp->mp_flags & P_LEAF2) { 6067 nodekey.mv_size = mc->mc_db->md_pad; 6068 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size); 6069 } else { 6070 leaf = NODEPTR(mp, 0); 6071 MDB_GET_KEY2(leaf, nodekey); 6072 } 6073 rc = mc->mc_dbx->md_cmp(key, &nodekey); 6074 if (rc == 0) { 6075 /* Probably happens rarely, but first node on the page 6076 * was the one we wanted. 6077 */ 6078 mc->mc_ki[mc->mc_top] = 0; 6079 if (exactp) 6080 *exactp = 1; 6081 goto set1; 6082 } 6083 if (rc > 0) { 6084 unsigned int i; 6085 unsigned int nkeys = NUMKEYS(mp); 6086 if (nkeys > 1) { 6087 if (mp->mp_flags & P_LEAF2) { 6088 nodekey.mv_data = LEAF2KEY(mp, 6089 nkeys-1, nodekey.mv_size); 6090 } else { 6091 leaf = NODEPTR(mp, nkeys-1); 6092 MDB_GET_KEY2(leaf, nodekey); 6093 } 6094 rc = mc->mc_dbx->md_cmp(key, &nodekey); 6095 if (rc == 0) { 6096 /* last node was the one we wanted */ 6097 mc->mc_ki[mc->mc_top] = nkeys-1; 6098 if (exactp) 6099 *exactp = 1; 6100 goto set1; 6101 } 6102 if (rc < 0) { 6103 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) { 6104 /* This is definitely the right page, skip search_page */ 6105 if (mp->mp_flags & P_LEAF2) { 6106 nodekey.mv_data = LEAF2KEY(mp, 6107 mc->mc_ki[mc->mc_top], nodekey.mv_size); 6108 } else { 6109 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 6110 MDB_GET_KEY2(leaf, nodekey); 6111 } 6112 rc = mc->mc_dbx->md_cmp(key, &nodekey); 6113 if (rc == 0) { 6114 /* current node was the one we wanted */ 6115 if (exactp) 6116 *exactp = 1; 6117 goto set1; 6118 } 6119 } 6120 rc = 0; 6121 mc->mc_flags &= ~C_EOF; 6122 goto set2; 6123 } 6124 } 6125 /* If any parents have right-sibs, search. 6126 * Otherwise, there's nothing further. 6127 */ 6128 for (i=0; i<mc->mc_top; i++) 6129 if (mc->mc_ki[i] < 6130 NUMKEYS(mc->mc_pg[i])-1) 6131 break; 6132 if (i == mc->mc_top) { 6133 /* There are no other pages */ 6134 mc->mc_ki[mc->mc_top] = nkeys; 6135 return MDB_NOTFOUND; 6136 } 6137 } 6138 if (!mc->mc_top) { 6139 /* There are no other pages */ 6140 mc->mc_ki[mc->mc_top] = 0; 6141 if (op == MDB_SET_RANGE && !exactp) { 6142 rc = 0; 6143 goto set1; 6144 } else 6145 return MDB_NOTFOUND; 6146 } 6147 } else { 6148 mc->mc_pg[0] = 0; 6149 } 6150 6151 rc = mdb_page_search(mc, key, 0); 6152 if (rc != MDB_SUCCESS) 6153 return rc; 6154 6155 mp = mc->mc_pg[mc->mc_top]; 6156 mdb_cassert(mc, IS_LEAF(mp)); 6157 6158 set2: 6159 leaf = mdb_node_search(mc, key, exactp); 6160 if (exactp != NULL && !*exactp) { 6161 /* MDB_SET specified and not an exact match. */ 6162 return MDB_NOTFOUND; 6163 } 6164 6165 if (leaf == NULL) { 6166 DPUTS("===> inexact leaf not found, goto sibling"); 6167 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) { 6168 mc->mc_flags |= C_EOF; 6169 return rc; /* no entries matched */ 6170 } 6171 mp = mc->mc_pg[mc->mc_top]; 6172 mdb_cassert(mc, IS_LEAF(mp)); 6173 leaf = NODEPTR(mp, 0); 6174 } 6175 6176 set1: 6177 mc->mc_flags |= C_INITIALIZED; 6178 mc->mc_flags &= ~C_EOF; 6179 6180 if (IS_LEAF2(mp)) { 6181 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) { 6182 key->mv_size = mc->mc_db->md_pad; 6183 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size); 6184 } 6185 return MDB_SUCCESS; 6186 } 6187 6188 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6189 mdb_xcursor_init1(mc, leaf); 6190 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) { 6191 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); 6192 } else { 6193 int ex2, *ex2p; 6194 if (op == MDB_GET_BOTH) { 6195 ex2p = &ex2; 6196 ex2 = 0; 6197 } else { 6198 ex2p = NULL; 6199 } 6200 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p); 6201 if (rc != MDB_SUCCESS) 6202 return rc; 6203 } 6204 } else if (data) { 6205 if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) { 6206 MDB_val olddata; 6207 MDB_cmp_func *dcmp; 6208 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS) 6209 return rc; 6210 dcmp = mc->mc_dbx->md_dcmp; 6211 #if UINT_MAX < SIZE_MAX 6212 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t)) 6213 dcmp = mdb_cmp_clong; 6214 #endif 6215 rc = dcmp(data, &olddata); 6216 if (rc) { 6217 if (op == MDB_GET_BOTH || rc > 0) 6218 return MDB_NOTFOUND; 6219 rc = 0; 6220 } 6221 *data = olddata; 6222 6223 } else { 6224 if (mc->mc_xcursor) 6225 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 6226 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 6227 return rc; 6228 } 6229 } 6230 6231 /* The key already matches in all other cases */ 6232 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) 6233 MDB_GET_KEY(leaf, key); 6234 DPRINTF(("==> cursor placed on key [%s]", DKEY(key))); 6235 6236 return rc; 6237 } 6238 6239 /** Move the cursor to the first item in the database. */ 6240 static int 6241 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data) 6242 { 6243 int rc; 6244 MDB_node *leaf; 6245 6246 if (mc->mc_xcursor) 6247 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 6248 6249 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) { 6250 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST); 6251 if (rc != MDB_SUCCESS) 6252 return rc; 6253 } 6254 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); 6255 6256 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0); 6257 mc->mc_flags |= C_INITIALIZED; 6258 mc->mc_flags &= ~C_EOF; 6259 6260 mc->mc_ki[mc->mc_top] = 0; 6261 6262 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { 6263 if ( key ) { 6264 key->mv_size = mc->mc_db->md_pad; 6265 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size); 6266 } 6267 return MDB_SUCCESS; 6268 } 6269 6270 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6271 mdb_xcursor_init1(mc, leaf); 6272 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); 6273 if (rc) 6274 return rc; 6275 } else if (data) { 6276 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 6277 return rc; 6278 } 6279 6280 MDB_GET_KEY(leaf, key); 6281 return MDB_SUCCESS; 6282 } 6283 6284 /** Move the cursor to the last item in the database. */ 6285 static int 6286 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data) 6287 { 6288 int rc; 6289 MDB_node *leaf; 6290 6291 if (mc->mc_xcursor) 6292 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 6293 6294 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) { 6295 rc = mdb_page_search(mc, NULL, MDB_PS_LAST); 6296 if (rc != MDB_SUCCESS) 6297 return rc; 6298 } 6299 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); 6300 6301 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1; 6302 mc->mc_flags |= C_INITIALIZED|C_EOF; 6303 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 6304 6305 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { 6306 if (key) { 6307 key->mv_size = mc->mc_db->md_pad; 6308 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size); 6309 } 6310 return MDB_SUCCESS; 6311 } 6312 6313 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6314 mdb_xcursor_init1(mc, leaf); 6315 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL); 6316 if (rc) 6317 return rc; 6318 } else if (data) { 6319 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 6320 return rc; 6321 } 6322 6323 MDB_GET_KEY(leaf, key); 6324 return MDB_SUCCESS; 6325 } 6326 6327 int 6328 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data, 6329 MDB_cursor_op op) 6330 { 6331 int rc; 6332 int exact = 0; 6333 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data); 6334 6335 if (mc == NULL) 6336 return EINVAL; 6337 6338 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) 6339 return MDB_BAD_TXN; 6340 6341 switch (op) { 6342 case MDB_GET_CURRENT: 6343 if (!(mc->mc_flags & C_INITIALIZED)) { 6344 rc = EINVAL; 6345 } else { 6346 MDB_page *mp = mc->mc_pg[mc->mc_top]; 6347 int nkeys = NUMKEYS(mp); 6348 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) { 6349 mc->mc_ki[mc->mc_top] = nkeys; 6350 rc = MDB_NOTFOUND; 6351 break; 6352 } 6353 rc = MDB_SUCCESS; 6354 if (IS_LEAF2(mp)) { 6355 key->mv_size = mc->mc_db->md_pad; 6356 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size); 6357 } else { 6358 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 6359 MDB_GET_KEY(leaf, key); 6360 if (data) { 6361 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6362 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT); 6363 } else { 6364 rc = mdb_node_read(mc, leaf, data); 6365 } 6366 } 6367 } 6368 } 6369 break; 6370 case MDB_GET_BOTH: 6371 case MDB_GET_BOTH_RANGE: 6372 if (data == NULL) { 6373 rc = EINVAL; 6374 break; 6375 } 6376 if (mc->mc_xcursor == NULL) { 6377 rc = MDB_INCOMPATIBLE; 6378 break; 6379 } 6380 /* FALLTHRU */ 6381 case MDB_SET: 6382 case MDB_SET_KEY: 6383 case MDB_SET_RANGE: 6384 if (key == NULL) { 6385 rc = EINVAL; 6386 } else { 6387 rc = mdb_cursor_set(mc, key, data, op, 6388 op == MDB_SET_RANGE ? NULL : &exact); 6389 } 6390 break; 6391 case MDB_GET_MULTIPLE: 6392 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) { 6393 rc = EINVAL; 6394 break; 6395 } 6396 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) { 6397 rc = MDB_INCOMPATIBLE; 6398 break; 6399 } 6400 rc = MDB_SUCCESS; 6401 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) || 6402 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF)) 6403 break; 6404 goto fetchm; 6405 case MDB_NEXT_MULTIPLE: 6406 if (data == NULL) { 6407 rc = EINVAL; 6408 break; 6409 } 6410 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) { 6411 rc = MDB_INCOMPATIBLE; 6412 break; 6413 } 6414 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP); 6415 if (rc == MDB_SUCCESS) { 6416 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) { 6417 MDB_cursor *mx; 6418 fetchm: 6419 mx = &mc->mc_xcursor->mx_cursor; 6420 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) * 6421 mx->mc_db->md_pad; 6422 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]); 6423 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1; 6424 } else { 6425 rc = MDB_NOTFOUND; 6426 } 6427 } 6428 break; 6429 case MDB_PREV_MULTIPLE: 6430 if (data == NULL) { 6431 rc = EINVAL; 6432 break; 6433 } 6434 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) { 6435 rc = MDB_INCOMPATIBLE; 6436 break; 6437 } 6438 if (!(mc->mc_flags & C_INITIALIZED)) 6439 rc = mdb_cursor_last(mc, key, data); 6440 else 6441 rc = MDB_SUCCESS; 6442 if (rc == MDB_SUCCESS) { 6443 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor; 6444 if (mx->mc_flags & C_INITIALIZED) { 6445 rc = mdb_cursor_sibling(mx, 0); 6446 if (rc == MDB_SUCCESS) 6447 goto fetchm; 6448 } else { 6449 rc = MDB_NOTFOUND; 6450 } 6451 } 6452 break; 6453 case MDB_NEXT: 6454 case MDB_NEXT_DUP: 6455 case MDB_NEXT_NODUP: 6456 rc = mdb_cursor_next(mc, key, data, op); 6457 break; 6458 case MDB_PREV: 6459 case MDB_PREV_DUP: 6460 case MDB_PREV_NODUP: 6461 rc = mdb_cursor_prev(mc, key, data, op); 6462 break; 6463 case MDB_FIRST: 6464 rc = mdb_cursor_first(mc, key, data); 6465 break; 6466 case MDB_FIRST_DUP: 6467 mfunc = mdb_cursor_first; 6468 mmove: 6469 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) { 6470 rc = EINVAL; 6471 break; 6472 } 6473 if (mc->mc_xcursor == NULL) { 6474 rc = MDB_INCOMPATIBLE; 6475 break; 6476 } 6477 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) { 6478 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]); 6479 rc = MDB_NOTFOUND; 6480 break; 6481 } 6482 { 6483 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 6484 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6485 MDB_GET_KEY(leaf, key); 6486 rc = mdb_node_read(mc, leaf, data); 6487 break; 6488 } 6489 } 6490 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) { 6491 rc = EINVAL; 6492 break; 6493 } 6494 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL); 6495 break; 6496 case MDB_LAST: 6497 rc = mdb_cursor_last(mc, key, data); 6498 break; 6499 case MDB_LAST_DUP: 6500 mfunc = mdb_cursor_last; 6501 goto mmove; 6502 default: 6503 DPRINTF(("unhandled/unimplemented cursor operation %u", op)); 6504 rc = EINVAL; 6505 break; 6506 } 6507 6508 if (mc->mc_flags & C_DEL) 6509 mc->mc_flags ^= C_DEL; 6510 6511 return rc; 6512 } 6513 6514 /** Touch all the pages in the cursor stack. Set mc_top. 6515 * Makes sure all the pages are writable, before attempting a write operation. 6516 * @param[in] mc The cursor to operate on. 6517 */ 6518 static int 6519 mdb_cursor_touch(MDB_cursor *mc) 6520 { 6521 int rc = MDB_SUCCESS; 6522 6523 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) { 6524 /* Touch DB record of named DB */ 6525 MDB_cursor mc2; 6526 MDB_xcursor mcx; 6527 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi)) 6528 return MDB_BAD_DBI; 6529 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx); 6530 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY); 6531 if (rc) 6532 return rc; 6533 *mc->mc_dbflag |= DB_DIRTY; 6534 } 6535 mc->mc_top = 0; 6536 if (mc->mc_snum) { 6537 do { 6538 rc = mdb_page_touch(mc); 6539 } while (!rc && ++(mc->mc_top) < mc->mc_snum); 6540 mc->mc_top = mc->mc_snum-1; 6541 } 6542 return rc; 6543 } 6544 6545 /** Do not spill pages to disk if txn is getting full, may fail instead */ 6546 #define MDB_NOSPILL 0x8000 6547 6548 int 6549 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data, 6550 unsigned int flags) 6551 { 6552 MDB_env *env; 6553 MDB_node *leaf = NULL; 6554 MDB_page *fp, *mp, *sub_root = NULL; 6555 uint16_t fp_flags; 6556 MDB_val xdata, *rdata, dkey, olddata; 6557 MDB_db dummy; 6558 int do_sub = 0, insert_key, insert_data; 6559 unsigned int mcount = 0, dcount = 0, nospill; 6560 size_t nsize; 6561 int rc, rc2; 6562 unsigned int nflags; 6563 DKBUF; 6564 6565 if (mc == NULL || key == NULL) 6566 return EINVAL; 6567 6568 env = mc->mc_txn->mt_env; 6569 6570 /* Check this first so counter will always be zero on any 6571 * early failures. 6572 */ 6573 if (flags & MDB_MULTIPLE) { 6574 dcount = data[1].mv_size; 6575 data[1].mv_size = 0; 6576 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED)) 6577 return MDB_INCOMPATIBLE; 6578 } 6579 6580 nospill = flags & MDB_NOSPILL; 6581 flags &= ~MDB_NOSPILL; 6582 6583 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED)) 6584 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN; 6585 6586 if (key->mv_size-1 >= ENV_MAXKEY(env)) 6587 return MDB_BAD_VALSIZE; 6588 6589 #if SIZE_MAX > MAXDATASIZE 6590 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE)) 6591 return MDB_BAD_VALSIZE; 6592 #else 6593 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env)) 6594 return MDB_BAD_VALSIZE; 6595 #endif 6596 6597 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u", 6598 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size)); 6599 6600 dkey.mv_size = 0; 6601 6602 if (flags & MDB_CURRENT) { 6603 if (!(mc->mc_flags & C_INITIALIZED)) 6604 return EINVAL; 6605 rc = MDB_SUCCESS; 6606 } else if (mc->mc_db->md_root == P_INVALID) { 6607 /* new database, cursor has nothing to point to */ 6608 mc->mc_snum = 0; 6609 mc->mc_top = 0; 6610 mc->mc_flags &= ~C_INITIALIZED; 6611 rc = MDB_NO_ROOT; 6612 } else { 6613 int exact = 0; 6614 MDB_val d2; 6615 if (flags & MDB_APPEND) { 6616 MDB_val k2; 6617 rc = mdb_cursor_last(mc, &k2, &d2); 6618 if (rc == 0) { 6619 rc = mc->mc_dbx->md_cmp(key, &k2); 6620 if (rc > 0) { 6621 rc = MDB_NOTFOUND; 6622 mc->mc_ki[mc->mc_top]++; 6623 } else { 6624 /* new key is <= last key */ 6625 rc = MDB_KEYEXIST; 6626 } 6627 } 6628 } else { 6629 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact); 6630 } 6631 if ((flags & MDB_NOOVERWRITE) && rc == 0) { 6632 DPRINTF(("duplicate key [%s]", DKEY(key))); 6633 *data = d2; 6634 return MDB_KEYEXIST; 6635 } 6636 if (rc && rc != MDB_NOTFOUND) 6637 return rc; 6638 } 6639 6640 if (mc->mc_flags & C_DEL) 6641 mc->mc_flags ^= C_DEL; 6642 6643 /* Cursor is positioned, check for room in the dirty list */ 6644 if (!nospill) { 6645 if (flags & MDB_MULTIPLE) { 6646 rdata = &xdata; 6647 xdata.mv_size = data->mv_size * dcount; 6648 } else { 6649 rdata = data; 6650 } 6651 if ((rc2 = mdb_page_spill(mc, key, rdata))) 6652 return rc2; 6653 } 6654 6655 if (rc == MDB_NO_ROOT) { 6656 MDB_page *np; 6657 /* new database, write a root leaf page */ 6658 DPUTS("allocating new root leaf page"); 6659 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) { 6660 return rc2; 6661 } 6662 mdb_cursor_push(mc, np); 6663 mc->mc_db->md_root = np->mp_pgno; 6664 mc->mc_db->md_depth++; 6665 *mc->mc_dbflag |= DB_DIRTY; 6666 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED)) 6667 == MDB_DUPFIXED) 6668 np->mp_flags |= P_LEAF2; 6669 mc->mc_flags |= C_INITIALIZED; 6670 } else { 6671 /* make sure all cursor pages are writable */ 6672 rc2 = mdb_cursor_touch(mc); 6673 if (rc2) 6674 return rc2; 6675 } 6676 6677 insert_key = insert_data = rc; 6678 if (insert_key) { 6679 /* The key does not exist */ 6680 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top])); 6681 if ((mc->mc_db->md_flags & MDB_DUPSORT) && 6682 LEAFSIZE(key, data) > env->me_nodemax) 6683 { 6684 /* Too big for a node, insert in sub-DB. Set up an empty 6685 * "old sub-page" for prep_subDB to expand to a full page. 6686 */ 6687 fp_flags = P_LEAF|P_DIRTY; 6688 fp = env->me_pbuf; 6689 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */ 6690 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE); 6691 olddata.mv_size = PAGEHDRSZ; 6692 goto prep_subDB; 6693 } 6694 } else { 6695 /* there's only a key anyway, so this is a no-op */ 6696 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { 6697 char *ptr; 6698 unsigned int ksize = mc->mc_db->md_pad; 6699 if (key->mv_size != ksize) 6700 return MDB_BAD_VALSIZE; 6701 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize); 6702 memcpy(ptr, key->mv_data, ksize); 6703 fix_parent: 6704 /* if overwriting slot 0 of leaf, need to 6705 * update branch key if there is a parent page 6706 */ 6707 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) { 6708 unsigned short dtop = 1; 6709 mc->mc_top--; 6710 /* slot 0 is always an empty key, find real slot */ 6711 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) { 6712 mc->mc_top--; 6713 dtop++; 6714 } 6715 if (mc->mc_ki[mc->mc_top]) 6716 rc2 = mdb_update_key(mc, key); 6717 else 6718 rc2 = MDB_SUCCESS; 6719 mc->mc_top += dtop; 6720 if (rc2) 6721 return rc2; 6722 } 6723 return MDB_SUCCESS; 6724 } 6725 6726 more: 6727 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 6728 olddata.mv_size = NODEDSZ(leaf); 6729 olddata.mv_data = NODEDATA(leaf); 6730 6731 /* DB has dups? */ 6732 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) { 6733 /* Prepare (sub-)page/sub-DB to accept the new item, 6734 * if needed. fp: old sub-page or a header faking 6735 * it. mp: new (sub-)page. offset: growth in page 6736 * size. xdata: node data with new page or DB. 6737 */ 6738 unsigned i, offset = 0; 6739 mp = fp = xdata.mv_data = env->me_pbuf; 6740 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno; 6741 6742 /* Was a single item before, must convert now */ 6743 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6744 MDB_cmp_func *dcmp; 6745 /* Just overwrite the current item */ 6746 if (flags == MDB_CURRENT) 6747 goto current; 6748 dcmp = mc->mc_dbx->md_dcmp; 6749 #if UINT_MAX < SIZE_MAX 6750 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t)) 6751 dcmp = mdb_cmp_clong; 6752 #endif 6753 /* does data match? */ 6754 if (!dcmp(data, &olddata)) { 6755 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP)) 6756 return MDB_KEYEXIST; 6757 /* overwrite it */ 6758 goto current; 6759 } 6760 6761 /* Back up original data item */ 6762 dkey.mv_size = olddata.mv_size; 6763 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size); 6764 6765 /* Make sub-page header for the dup items, with dummy body */ 6766 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP; 6767 fp->mp_lower = (PAGEHDRSZ-PAGEBASE); 6768 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size; 6769 if (mc->mc_db->md_flags & MDB_DUPFIXED) { 6770 fp->mp_flags |= P_LEAF2; 6771 fp->mp_pad = data->mv_size; 6772 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */ 6773 } else { 6774 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) + 6775 (dkey.mv_size & 1) + (data->mv_size & 1); 6776 } 6777 fp->mp_upper = xdata.mv_size - PAGEBASE; 6778 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */ 6779 } else if (leaf->mn_flags & F_SUBDATA) { 6780 /* Data is on sub-DB, just store it */ 6781 flags |= F_DUPDATA|F_SUBDATA; 6782 goto put_sub; 6783 } else { 6784 /* Data is on sub-page */ 6785 fp = olddata.mv_data; 6786 switch (flags) { 6787 default: 6788 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) { 6789 offset = EVEN(NODESIZE + sizeof(indx_t) + 6790 data->mv_size); 6791 break; 6792 } 6793 offset = fp->mp_pad; 6794 if (SIZELEFT(fp) < offset) { 6795 offset *= 4; /* space for 4 more */ 6796 break; 6797 } 6798 /* FALLTHRU */ /* Big enough MDB_DUPFIXED sub-page */ 6799 case MDB_CURRENT: 6800 fp->mp_flags |= P_DIRTY; 6801 COPY_PGNO(fp->mp_pgno, mp->mp_pgno); 6802 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp; 6803 flags |= F_DUPDATA; 6804 goto put_sub; 6805 } 6806 xdata.mv_size = olddata.mv_size + offset; 6807 } 6808 6809 fp_flags = fp->mp_flags; 6810 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) { 6811 /* Too big for a sub-page, convert to sub-DB */ 6812 fp_flags &= ~P_SUBP; 6813 prep_subDB: 6814 if (mc->mc_db->md_flags & MDB_DUPFIXED) { 6815 fp_flags |= P_LEAF2; 6816 dummy.md_pad = fp->mp_pad; 6817 dummy.md_flags = MDB_DUPFIXED; 6818 if (mc->mc_db->md_flags & MDB_INTEGERDUP) 6819 dummy.md_flags |= MDB_INTEGERKEY; 6820 } else { 6821 dummy.md_pad = 0; 6822 dummy.md_flags = 0; 6823 } 6824 dummy.md_depth = 1; 6825 dummy.md_branch_pages = 0; 6826 dummy.md_leaf_pages = 1; 6827 dummy.md_overflow_pages = 0; 6828 dummy.md_entries = NUMKEYS(fp); 6829 xdata.mv_size = sizeof(MDB_db); 6830 xdata.mv_data = &dummy; 6831 if ((rc = mdb_page_alloc(mc, 1, &mp))) 6832 return rc; 6833 offset = env->me_psize - olddata.mv_size; 6834 flags |= F_DUPDATA|F_SUBDATA; 6835 dummy.md_root = mp->mp_pgno; 6836 sub_root = mp; 6837 } 6838 if (mp != fp) { 6839 mp->mp_flags = fp_flags | P_DIRTY; 6840 mp->mp_pad = fp->mp_pad; 6841 mp->mp_lower = fp->mp_lower; 6842 mp->mp_upper = fp->mp_upper + offset; 6843 if (fp_flags & P_LEAF2) { 6844 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad); 6845 } else { 6846 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE, 6847 olddata.mv_size - fp->mp_upper - PAGEBASE); 6848 memcpy((char *)(&mp->mp_ptrs), (char *)(&fp->mp_ptrs), NUMKEYS(fp) * sizeof(mp->mp_ptrs[0])); 6849 for (i=0; i<NUMKEYS(fp); i++) 6850 mp->mp_ptrs[i] += offset; 6851 } 6852 } 6853 6854 rdata = &xdata; 6855 flags |= F_DUPDATA; 6856 do_sub = 1; 6857 if (!insert_key) 6858 mdb_node_del(mc, 0); 6859 goto new_sub; 6860 } 6861 current: 6862 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */ 6863 if ((leaf->mn_flags ^ flags) & F_SUBDATA) 6864 return MDB_INCOMPATIBLE; 6865 /* overflow page overwrites need special handling */ 6866 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) { 6867 MDB_page *omp; 6868 pgno_t pg; 6869 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize); 6870 6871 memcpy(&pg, olddata.mv_data, sizeof(pg)); 6872 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0) 6873 return rc2; 6874 ovpages = omp->mp_pages; 6875 6876 /* Is the ov page large enough? */ 6877 if (ovpages >= dpages) { 6878 if (!(omp->mp_flags & P_DIRTY) && 6879 (level || (env->me_flags & MDB_WRITEMAP))) 6880 { 6881 rc = mdb_page_unspill(mc->mc_txn, omp, &omp); 6882 if (rc) 6883 return rc; 6884 level = 0; /* dirty in this txn or clean */ 6885 } 6886 /* Is it dirty? */ 6887 if (omp->mp_flags & P_DIRTY) { 6888 /* yes, overwrite it. Note in this case we don't 6889 * bother to try shrinking the page if the new data 6890 * is smaller than the overflow threshold. 6891 */ 6892 if (level > 1) { 6893 /* It is writable only in a parent txn */ 6894 size_t sz = (size_t) env->me_psize * ovpages, off; 6895 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages); 6896 MDB_ID2 id2; 6897 if (!np) 6898 return ENOMEM; 6899 id2.mid = pg; 6900 id2.mptr = np; 6901 /* Note - this page is already counted in parent's dirty_room */ 6902 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2); 6903 mdb_cassert(mc, rc2 == 0); 6904 /* Currently we make the page look as with put() in the 6905 * parent txn, in case the user peeks at MDB_RESERVEd 6906 * or unused parts. Some users treat ovpages specially. 6907 */ 6908 if (!(flags & MDB_RESERVE)) { 6909 /* Skip the part where LMDB will put *data. 6910 * Copy end of page, adjusting alignment so 6911 * compiler may copy words instead of bytes. 6912 */ 6913 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t); 6914 memcpy((size_t *)((char *)np + off), 6915 (size_t *)((char *)omp + off), sz - off); 6916 sz = PAGEHDRSZ; 6917 } 6918 memcpy(np, omp, sz); /* Copy beginning of page */ 6919 omp = np; 6920 } 6921 SETDSZ(leaf, data->mv_size); 6922 if (F_ISSET(flags, MDB_RESERVE)) 6923 data->mv_data = METADATA(omp); 6924 else 6925 memcpy(METADATA(omp), data->mv_data, data->mv_size); 6926 return MDB_SUCCESS; 6927 } 6928 } 6929 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS) 6930 return rc2; 6931 } else if (data->mv_size == olddata.mv_size) { 6932 /* same size, just replace it. Note that we could 6933 * also reuse this node if the new data is smaller, 6934 * but instead we opt to shrink the node in that case. 6935 */ 6936 if (F_ISSET(flags, MDB_RESERVE)) 6937 data->mv_data = olddata.mv_data; 6938 else if (!(mc->mc_flags & C_SUB)) 6939 memcpy(olddata.mv_data, data->mv_data, data->mv_size); 6940 else { 6941 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size); 6942 goto fix_parent; 6943 } 6944 return MDB_SUCCESS; 6945 } 6946 mdb_node_del(mc, 0); 6947 } 6948 6949 rdata = data; 6950 6951 new_sub: 6952 nflags = flags & NODE_ADD_FLAGS; 6953 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata); 6954 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) { 6955 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA ) 6956 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */ 6957 if (!insert_key) 6958 nflags |= MDB_SPLIT_REPLACE; 6959 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags); 6960 } else { 6961 /* There is room already in this leaf page. */ 6962 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags); 6963 if (rc == 0) { 6964 /* Adjust other cursors pointing to mp */ 6965 MDB_cursor *m2, *m3; 6966 MDB_dbi dbi = mc->mc_dbi; 6967 unsigned i = mc->mc_top; 6968 MDB_page *mp = mc->mc_pg[i]; 6969 6970 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 6971 if (mc->mc_flags & C_SUB) 6972 m3 = &m2->mc_xcursor->mx_cursor; 6973 else 6974 m3 = m2; 6975 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue; 6976 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) { 6977 m3->mc_ki[i]++; 6978 } 6979 XCURSOR_REFRESH(m3, i, mp); 6980 } 6981 } 6982 } 6983 6984 if (rc == MDB_SUCCESS) { 6985 /* Now store the actual data in the child DB. Note that we're 6986 * storing the user data in the keys field, so there are strict 6987 * size limits on dupdata. The actual data fields of the child 6988 * DB are all zero size. 6989 */ 6990 if (do_sub) { 6991 int xflags, new_dupdata; 6992 size_t ecount; 6993 put_sub: 6994 xdata.mv_size = 0; 6995 xdata.mv_data = ""; 6996 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 6997 if ((flags & (MDB_CURRENT|MDB_APPENDDUP)) == MDB_CURRENT) { 6998 xflags = MDB_CURRENT|MDB_NOSPILL; 6999 } else { 7000 mdb_xcursor_init1(mc, leaf); 7001 xflags = (flags & MDB_NODUPDATA) ? 7002 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL; 7003 } 7004 if (sub_root) 7005 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root; 7006 new_dupdata = (int)dkey.mv_size; 7007 /* converted, write the original data first */ 7008 if (dkey.mv_size) { 7009 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags); 7010 if (rc) 7011 goto bad_sub; 7012 /* we've done our job */ 7013 dkey.mv_size = 0; 7014 } 7015 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) { 7016 /* Adjust other cursors pointing to mp */ 7017 MDB_cursor *m2; 7018 MDB_xcursor *mx = mc->mc_xcursor; 7019 unsigned i = mc->mc_top; 7020 MDB_page *mp = mc->mc_pg[i]; 7021 7022 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) { 7023 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue; 7024 if (!(m2->mc_flags & C_INITIALIZED)) continue; 7025 if (m2->mc_pg[i] == mp) { 7026 if (m2->mc_ki[i] == mc->mc_ki[i]) { 7027 mdb_xcursor_init2(m2, mx, new_dupdata); 7028 } else if (!insert_key) { 7029 XCURSOR_REFRESH(m2, i, mp); 7030 } 7031 } 7032 } 7033 } 7034 ecount = mc->mc_xcursor->mx_db.md_entries; 7035 if (flags & MDB_APPENDDUP) 7036 xflags |= MDB_APPEND; 7037 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags); 7038 if (flags & F_SUBDATA) { 7039 void *db = NODEDATA(leaf); 7040 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db)); 7041 } 7042 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount; 7043 } 7044 /* Increment count unless we just replaced an existing item. */ 7045 if (insert_data) 7046 mc->mc_db->md_entries++; 7047 if (insert_key) { 7048 /* Invalidate txn if we created an empty sub-DB */ 7049 if (rc) 7050 goto bad_sub; 7051 /* If we succeeded and the key didn't exist before, 7052 * make sure the cursor is marked valid. 7053 */ 7054 mc->mc_flags |= C_INITIALIZED; 7055 } 7056 if (flags & MDB_MULTIPLE) { 7057 if (!rc) { 7058 mcount++; 7059 /* let caller know how many succeeded, if any */ 7060 data[1].mv_size = mcount; 7061 if (mcount < dcount) { 7062 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size; 7063 insert_key = insert_data = 0; 7064 goto more; 7065 } 7066 } 7067 } 7068 return rc; 7069 bad_sub: 7070 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */ 7071 rc = MDB_CORRUPTED; 7072 } 7073 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 7074 return rc; 7075 } 7076 7077 int 7078 mdb_cursor_del(MDB_cursor *mc, unsigned int flags) 7079 { 7080 MDB_node *leaf; 7081 MDB_page *mp; 7082 int rc; 7083 7084 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED)) 7085 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN; 7086 7087 if (!(mc->mc_flags & C_INITIALIZED)) 7088 return EINVAL; 7089 7090 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) 7091 return MDB_NOTFOUND; 7092 7093 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL))) 7094 return rc; 7095 7096 rc = mdb_cursor_touch(mc); 7097 if (rc) 7098 return rc; 7099 7100 mp = mc->mc_pg[mc->mc_top]; 7101 if (!IS_LEAF(mp)) 7102 return MDB_CORRUPTED; 7103 if (IS_LEAF2(mp)) 7104 goto del_key; 7105 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 7106 7107 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 7108 if (flags & MDB_NODUPDATA) { 7109 /* mdb_cursor_del0() will subtract the final entry */ 7110 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1; 7111 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED; 7112 } else { 7113 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) { 7114 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf); 7115 } 7116 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL); 7117 if (rc) 7118 return rc; 7119 /* If sub-DB still has entries, we're done */ 7120 if (mc->mc_xcursor->mx_db.md_entries) { 7121 if (leaf->mn_flags & F_SUBDATA) { 7122 /* update subDB info */ 7123 void *db = NODEDATA(leaf); 7124 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db)); 7125 } else { 7126 MDB_cursor *m2; 7127 /* shrink fake page */ 7128 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]); 7129 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 7130 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf); 7131 /* fix other sub-DB cursors pointed at fake pages on this page */ 7132 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) { 7133 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue; 7134 if (!(m2->mc_flags & C_INITIALIZED)) continue; 7135 if (m2->mc_pg[mc->mc_top] == mp) { 7136 XCURSOR_REFRESH(m2, mc->mc_top, mp); 7137 } 7138 } 7139 } 7140 mc->mc_db->md_entries--; 7141 return rc; 7142 } else { 7143 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED; 7144 } 7145 /* otherwise fall thru and delete the sub-DB */ 7146 } 7147 7148 if (leaf->mn_flags & F_SUBDATA) { 7149 /* add all the child DB's pages to the free list */ 7150 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0); 7151 if (rc) 7152 goto fail; 7153 } 7154 } 7155 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */ 7156 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) { 7157 rc = MDB_INCOMPATIBLE; 7158 goto fail; 7159 } 7160 7161 /* add overflow pages to free list */ 7162 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) { 7163 MDB_page *omp; 7164 pgno_t pg; 7165 7166 memcpy(&pg, NODEDATA(leaf), sizeof(pg)); 7167 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) || 7168 (rc = mdb_ovpage_free(mc, omp))) 7169 goto fail; 7170 } 7171 7172 del_key: 7173 return mdb_cursor_del0(mc); 7174 7175 fail: 7176 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 7177 return rc; 7178 } 7179 7180 /** Allocate and initialize new pages for a database. 7181 * Set #MDB_TXN_ERROR on failure. 7182 * @param[in] mc a cursor on the database being added to. 7183 * @param[in] flags flags defining what type of page is being allocated. 7184 * @param[in] num the number of pages to allocate. This is usually 1, 7185 * unless allocating overflow pages for a large record. 7186 * @param[out] mp Address of a page, or NULL on failure. 7187 * @return 0 on success, non-zero on failure. 7188 */ 7189 static int 7190 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp) 7191 { 7192 MDB_page *np; 7193 int rc; 7194 7195 if ((rc = mdb_page_alloc(mc, num, &np))) 7196 return rc; 7197 DPRINTF(("allocated new mpage %"Z"u, page size %u", 7198 np->mp_pgno, mc->mc_txn->mt_env->me_psize)); 7199 np->mp_flags = flags | P_DIRTY; 7200 np->mp_lower = (PAGEHDRSZ-PAGEBASE); 7201 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE; 7202 7203 if (IS_BRANCH(np)) 7204 mc->mc_db->md_branch_pages++; 7205 else if (IS_LEAF(np)) 7206 mc->mc_db->md_leaf_pages++; 7207 else if (IS_OVERFLOW(np)) { 7208 mc->mc_db->md_overflow_pages += num; 7209 np->mp_pages = num; 7210 } 7211 *mp = np; 7212 7213 return 0; 7214 } 7215 7216 /** Calculate the size of a leaf node. 7217 * The size depends on the environment's page size; if a data item 7218 * is too large it will be put onto an overflow page and the node 7219 * size will only include the key and not the data. Sizes are always 7220 * rounded up to an even number of bytes, to guarantee 2-byte alignment 7221 * of the #MDB_node headers. 7222 * @param[in] env The environment handle. 7223 * @param[in] key The key for the node. 7224 * @param[in] data The data for the node. 7225 * @return The number of bytes needed to store the node. 7226 */ 7227 static size_t 7228 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data) 7229 { 7230 size_t sz; 7231 7232 sz = LEAFSIZE(key, data); 7233 if (sz > env->me_nodemax) { 7234 /* put on overflow page */ 7235 sz -= data->mv_size - sizeof(pgno_t); 7236 } 7237 7238 return EVEN(sz + sizeof(indx_t)); 7239 } 7240 7241 /** Calculate the size of a branch node. 7242 * The size should depend on the environment's page size but since 7243 * we currently don't support spilling large keys onto overflow 7244 * pages, it's simply the size of the #MDB_node header plus the 7245 * size of the key. Sizes are always rounded up to an even number 7246 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers. 7247 * @param[in] env The environment handle. 7248 * @param[in] key The key for the node. 7249 * @return The number of bytes needed to store the node. 7250 */ 7251 static size_t 7252 mdb_branch_size(MDB_env *env, MDB_val *key) 7253 { 7254 size_t sz; 7255 7256 sz = INDXSIZE(key); 7257 if (sz > env->me_nodemax) { 7258 /* put on overflow page */ 7259 /* not implemented */ 7260 /* sz -= key->size - sizeof(pgno_t); */ 7261 } 7262 7263 return sz + sizeof(indx_t); 7264 } 7265 7266 /** Add a node to the page pointed to by the cursor. 7267 * Set #MDB_TXN_ERROR on failure. 7268 * @param[in] mc The cursor for this operation. 7269 * @param[in] indx The index on the page where the new node should be added. 7270 * @param[in] key The key for the new node. 7271 * @param[in] data The data for the new node, if any. 7272 * @param[in] pgno The page number, if adding a branch node. 7273 * @param[in] flags Flags for the node. 7274 * @return 0 on success, non-zero on failure. Possible errors are: 7275 * <ul> 7276 * <li>ENOMEM - failed to allocate overflow pages for the node. 7277 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error 7278 * should never happen since all callers already calculate the 7279 * page's free space before calling this function. 7280 * </ul> 7281 */ 7282 static int 7283 mdb_node_add(MDB_cursor *mc, indx_t indx, 7284 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags) 7285 { 7286 unsigned int i; 7287 size_t node_size = NODESIZE; 7288 ssize_t room; 7289 indx_t ofs; 7290 MDB_node *node; 7291 MDB_page *mp = mc->mc_pg[mc->mc_top]; 7292 MDB_page *ofp = NULL; /* overflow page */ 7293 void *ndata; 7294 DKBUF; 7295 7296 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower); 7297 7298 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]", 7299 IS_LEAF(mp) ? "leaf" : "branch", 7300 IS_SUBP(mp) ? "sub-" : "", 7301 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0, 7302 key ? key->mv_size : 0, key ? DKEY(key) : "null")); 7303 7304 if (IS_LEAF2(mp)) { 7305 /* Move higher keys up one slot. */ 7306 int ksize = mc->mc_db->md_pad, dif; 7307 char *ptr = LEAF2KEY(mp, indx, ksize); 7308 dif = NUMKEYS(mp) - indx; 7309 if (dif > 0) 7310 memmove(ptr+ksize, ptr, dif*ksize); 7311 /* insert new key */ 7312 memcpy(ptr, key->mv_data, ksize); 7313 7314 /* Just using these for counting */ 7315 mp->mp_lower += sizeof(indx_t); 7316 mp->mp_upper -= ksize - sizeof(indx_t); 7317 return MDB_SUCCESS; 7318 } 7319 7320 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t); 7321 if (key != NULL) 7322 node_size += key->mv_size; 7323 if (IS_LEAF(mp)) { 7324 mdb_cassert(mc, key && data); 7325 if (F_ISSET(flags, F_BIGDATA)) { 7326 /* Data already on overflow page. */ 7327 node_size += sizeof(pgno_t); 7328 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) { 7329 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize); 7330 int rc; 7331 /* Put data on overflow page. */ 7332 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page", 7333 data->mv_size, node_size+data->mv_size)); 7334 node_size = EVEN(node_size + sizeof(pgno_t)); 7335 if ((ssize_t)node_size > room) 7336 goto full; 7337 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp))) 7338 return rc; 7339 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno)); 7340 flags |= F_BIGDATA; 7341 goto update; 7342 } else { 7343 node_size += data->mv_size; 7344 } 7345 } 7346 node_size = EVEN(node_size); 7347 if ((ssize_t)node_size > room) 7348 goto full; 7349 7350 update: 7351 /* Move higher pointers up one slot. */ 7352 for (i = NUMKEYS(mp); i > indx; i--) 7353 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1]; 7354 7355 /* Adjust free space offsets. */ 7356 ofs = mp->mp_upper - node_size; 7357 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t)); 7358 mp->mp_ptrs[indx] = ofs; 7359 mp->mp_upper = ofs; 7360 mp->mp_lower += sizeof(indx_t); 7361 7362 /* Write the node data. */ 7363 node = NODEPTR(mp, indx); 7364 node->mn_ksize = (key == NULL) ? 0 : key->mv_size; 7365 node->mn_flags = flags; 7366 if (IS_LEAF(mp)) 7367 SETDSZ(node,data->mv_size); 7368 else 7369 SETPGNO(node,pgno); 7370 7371 if (key) 7372 memcpy(NODEKEY(node), key->mv_data, key->mv_size); 7373 7374 if (IS_LEAF(mp)) { 7375 ndata = NODEDATA(node); 7376 if (ofp == NULL) { 7377 if (F_ISSET(flags, F_BIGDATA)) 7378 memcpy(ndata, data->mv_data, sizeof(pgno_t)); 7379 else if (F_ISSET(flags, MDB_RESERVE)) 7380 data->mv_data = ndata; 7381 else 7382 memcpy(ndata, data->mv_data, data->mv_size); 7383 } else { 7384 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t)); 7385 ndata = METADATA(ofp); 7386 if (F_ISSET(flags, MDB_RESERVE)) 7387 data->mv_data = ndata; 7388 else 7389 memcpy(ndata, data->mv_data, data->mv_size); 7390 } 7391 } 7392 7393 return MDB_SUCCESS; 7394 7395 full: 7396 DPRINTF(("not enough room in page %"Z"u, got %u ptrs", 7397 mdb_dbg_pgno(mp), NUMKEYS(mp))); 7398 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room)); 7399 DPRINTF(("node size = %"Z"u", node_size)); 7400 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 7401 return MDB_PAGE_FULL; 7402 } 7403 7404 /** Delete the specified node from a page. 7405 * @param[in] mc Cursor pointing to the node to delete. 7406 * @param[in] ksize The size of a node. Only used if the page is 7407 * part of a #MDB_DUPFIXED database. 7408 */ 7409 static void 7410 mdb_node_del(MDB_cursor *mc, int ksize) 7411 { 7412 MDB_page *mp = mc->mc_pg[mc->mc_top]; 7413 indx_t indx = mc->mc_ki[mc->mc_top]; 7414 unsigned int sz; 7415 indx_t i, j, numkeys, ptr; 7416 MDB_node *node; 7417 char *base; 7418 7419 DPRINTF(("delete node %u on %s page %"Z"u", indx, 7420 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp))); 7421 numkeys = NUMKEYS(mp); 7422 mdb_cassert(mc, indx < numkeys); 7423 7424 if (IS_LEAF2(mp)) { 7425 int x = numkeys - 1 - indx; 7426 base = LEAF2KEY(mp, indx, ksize); 7427 if (x) 7428 memmove(base, base + ksize, x * ksize); 7429 mp->mp_lower -= sizeof(indx_t); 7430 mp->mp_upper += ksize - sizeof(indx_t); 7431 return; 7432 } 7433 7434 node = NODEPTR(mp, indx); 7435 sz = NODESIZE + node->mn_ksize; 7436 if (IS_LEAF(mp)) { 7437 if (F_ISSET(node->mn_flags, F_BIGDATA)) 7438 sz += sizeof(pgno_t); 7439 else 7440 sz += NODEDSZ(node); 7441 } 7442 sz = EVEN(sz); 7443 7444 ptr = mp->mp_ptrs[indx]; 7445 for (i = j = 0; i < numkeys; i++) { 7446 if (i != indx) { 7447 mp->mp_ptrs[j] = mp->mp_ptrs[i]; 7448 if (mp->mp_ptrs[i] < ptr) 7449 mp->mp_ptrs[j] += sz; 7450 j++; 7451 } 7452 } 7453 7454 base = (char *)mp + mp->mp_upper + PAGEBASE; 7455 memmove(base + sz, base, ptr - mp->mp_upper); 7456 7457 mp->mp_lower -= sizeof(indx_t); 7458 mp->mp_upper += sz; 7459 } 7460 7461 /** Compact the main page after deleting a node on a subpage. 7462 * @param[in] mp The main page to operate on. 7463 * @param[in] indx The index of the subpage on the main page. 7464 */ 7465 static void 7466 mdb_node_shrink(MDB_page *mp, indx_t indx) 7467 { 7468 MDB_node *node; 7469 MDB_page *sp, *xp; 7470 char *base; 7471 indx_t delta, nsize, len, ptr; 7472 int i; 7473 7474 node = NODEPTR(mp, indx); 7475 sp = (MDB_page *)NODEDATA(node); 7476 delta = SIZELEFT(sp); 7477 nsize = NODEDSZ(node) - delta; 7478 7479 /* Prepare to shift upward, set len = length(subpage part to shift) */ 7480 if (IS_LEAF2(sp)) { 7481 len = nsize; 7482 if (nsize & 1) 7483 return; /* do not make the node uneven-sized */ 7484 } else { 7485 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */ 7486 for (i = NUMKEYS(sp); --i >= 0; ) 7487 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta; 7488 len = PAGEHDRSZ; 7489 } 7490 sp->mp_upper = sp->mp_lower; 7491 COPY_PGNO(sp->mp_pgno, mp->mp_pgno); 7492 SETDSZ(node, nsize); 7493 7494 /* Shift <lower nodes...initial part of subpage> upward */ 7495 base = (char *)mp + mp->mp_upper + PAGEBASE; 7496 memmove(base + delta, base, (char *)sp + len - base); 7497 7498 ptr = mp->mp_ptrs[indx]; 7499 for (i = NUMKEYS(mp); --i >= 0; ) { 7500 if (mp->mp_ptrs[i] <= ptr) 7501 mp->mp_ptrs[i] += delta; 7502 } 7503 mp->mp_upper += delta; 7504 } 7505 7506 /** Initial setup of a sorted-dups cursor. 7507 * Sorted duplicates are implemented as a sub-database for the given key. 7508 * The duplicate data items are actually keys of the sub-database. 7509 * Operations on the duplicate data items are performed using a sub-cursor 7510 * initialized when the sub-database is first accessed. This function does 7511 * the preliminary setup of the sub-cursor, filling in the fields that 7512 * depend only on the parent DB. 7513 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized. 7514 */ 7515 static void 7516 mdb_xcursor_init0(MDB_cursor *mc) 7517 { 7518 MDB_xcursor *mx = mc->mc_xcursor; 7519 7520 mx->mx_cursor.mc_xcursor = NULL; 7521 mx->mx_cursor.mc_txn = mc->mc_txn; 7522 mx->mx_cursor.mc_db = &mx->mx_db; 7523 mx->mx_cursor.mc_dbx = &mx->mx_dbx; 7524 mx->mx_cursor.mc_dbi = mc->mc_dbi; 7525 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag; 7526 mx->mx_cursor.mc_snum = 0; 7527 mx->mx_cursor.mc_top = 0; 7528 mx->mx_cursor.mc_flags = C_SUB; 7529 mx->mx_dbx.md_name.mv_size = 0; 7530 mx->mx_dbx.md_name.mv_data = NULL; 7531 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp; 7532 mx->mx_dbx.md_dcmp = NULL; 7533 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel; 7534 } 7535 7536 /** Final setup of a sorted-dups cursor. 7537 * Sets up the fields that depend on the data from the main cursor. 7538 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized. 7539 * @param[in] node The data containing the #MDB_db record for the 7540 * sorted-dup database. 7541 */ 7542 static void 7543 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node) 7544 { 7545 MDB_xcursor *mx = mc->mc_xcursor; 7546 7547 if (node->mn_flags & F_SUBDATA) { 7548 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db)); 7549 mx->mx_cursor.mc_pg[0] = 0; 7550 mx->mx_cursor.mc_snum = 0; 7551 mx->mx_cursor.mc_top = 0; 7552 mx->mx_cursor.mc_flags = C_SUB; 7553 } else { 7554 MDB_page *fp = NODEDATA(node); 7555 mx->mx_db.md_pad = 0; 7556 mx->mx_db.md_flags = 0; 7557 mx->mx_db.md_depth = 1; 7558 mx->mx_db.md_branch_pages = 0; 7559 mx->mx_db.md_leaf_pages = 1; 7560 mx->mx_db.md_overflow_pages = 0; 7561 mx->mx_db.md_entries = NUMKEYS(fp); 7562 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno); 7563 mx->mx_cursor.mc_snum = 1; 7564 mx->mx_cursor.mc_top = 0; 7565 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB; 7566 mx->mx_cursor.mc_pg[0] = fp; 7567 mx->mx_cursor.mc_ki[0] = 0; 7568 if (mc->mc_db->md_flags & MDB_DUPFIXED) { 7569 mx->mx_db.md_flags = MDB_DUPFIXED; 7570 mx->mx_db.md_pad = fp->mp_pad; 7571 if (mc->mc_db->md_flags & MDB_INTEGERDUP) 7572 mx->mx_db.md_flags |= MDB_INTEGERKEY; 7573 } 7574 } 7575 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi, 7576 mx->mx_db.md_root)); 7577 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA; 7578 #if UINT_MAX < SIZE_MAX 7579 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t)) 7580 mx->mx_dbx.md_cmp = mdb_cmp_clong; 7581 #endif 7582 } 7583 7584 7585 /** Fixup a sorted-dups cursor due to underlying update. 7586 * Sets up some fields that depend on the data from the main cursor. 7587 * Almost the same as init1, but skips initialization steps if the 7588 * xcursor had already been used. 7589 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up. 7590 * @param[in] src_mx The xcursor of an up-to-date cursor. 7591 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item. 7592 */ 7593 static void 7594 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata) 7595 { 7596 MDB_xcursor *mx = mc->mc_xcursor; 7597 7598 if (new_dupdata) { 7599 mx->mx_cursor.mc_snum = 1; 7600 mx->mx_cursor.mc_top = 0; 7601 mx->mx_cursor.mc_flags |= C_INITIALIZED; 7602 mx->mx_cursor.mc_ki[0] = 0; 7603 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA; 7604 #if UINT_MAX < SIZE_MAX 7605 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp; 7606 #endif 7607 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) { 7608 return; 7609 } 7610 mx->mx_db = src_mx->mx_db; 7611 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0]; 7612 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi, 7613 mx->mx_db.md_root)); 7614 } 7615 7616 /** Initialize a cursor for a given transaction and database. */ 7617 static void 7618 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx) 7619 { 7620 mc->mc_next = NULL; 7621 mc->mc_backup = NULL; 7622 mc->mc_dbi = dbi; 7623 mc->mc_txn = txn; 7624 mc->mc_db = &txn->mt_dbs[dbi]; 7625 mc->mc_dbx = &txn->mt_dbxs[dbi]; 7626 mc->mc_dbflag = &txn->mt_dbflags[dbi]; 7627 mc->mc_snum = 0; 7628 mc->mc_top = 0; 7629 mc->mc_pg[0] = 0; 7630 mc->mc_ki[0] = 0; 7631 mc->mc_flags = 0; 7632 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) { 7633 mdb_tassert(txn, mx != NULL); 7634 mc->mc_xcursor = mx; 7635 mdb_xcursor_init0(mc); 7636 } else { 7637 mc->mc_xcursor = NULL; 7638 } 7639 if (*mc->mc_dbflag & DB_STALE) { 7640 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY); 7641 } 7642 } 7643 7644 int 7645 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret) 7646 { 7647 MDB_cursor *mc; 7648 size_t size = sizeof(MDB_cursor); 7649 7650 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID)) 7651 return EINVAL; 7652 7653 if (txn->mt_flags & MDB_TXN_BLOCKED) 7654 return MDB_BAD_TXN; 7655 7656 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) 7657 return EINVAL; 7658 7659 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) 7660 size += sizeof(MDB_xcursor); 7661 7662 if ((mc = malloc(size)) != NULL) { 7663 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1)); 7664 if (txn->mt_cursors) { 7665 mc->mc_next = txn->mt_cursors[dbi]; 7666 txn->mt_cursors[dbi] = mc; 7667 mc->mc_flags |= C_UNTRACK; 7668 } 7669 } else { 7670 return ENOMEM; 7671 } 7672 7673 *ret = mc; 7674 7675 return MDB_SUCCESS; 7676 } 7677 7678 int 7679 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc) 7680 { 7681 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID)) 7682 return EINVAL; 7683 7684 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors) 7685 return EINVAL; 7686 7687 if (txn->mt_flags & MDB_TXN_BLOCKED) 7688 return MDB_BAD_TXN; 7689 7690 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor); 7691 return MDB_SUCCESS; 7692 } 7693 7694 /* Return the count of duplicate data items for the current key */ 7695 int 7696 mdb_cursor_count(MDB_cursor *mc, size_t *countp) 7697 { 7698 MDB_node *leaf; 7699 7700 if (mc == NULL || countp == NULL) 7701 return EINVAL; 7702 7703 if (mc->mc_xcursor == NULL) 7704 return MDB_INCOMPATIBLE; 7705 7706 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) 7707 return MDB_BAD_TXN; 7708 7709 if (!(mc->mc_flags & C_INITIALIZED)) 7710 return EINVAL; 7711 7712 if (!mc->mc_snum) 7713 return MDB_NOTFOUND; 7714 7715 if (mc->mc_flags & C_EOF) { 7716 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) 7717 return MDB_NOTFOUND; 7718 mc->mc_flags ^= C_EOF; 7719 } 7720 7721 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 7722 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { 7723 *countp = 1; 7724 } else { 7725 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) 7726 return EINVAL; 7727 7728 *countp = mc->mc_xcursor->mx_db.md_entries; 7729 } 7730 return MDB_SUCCESS; 7731 } 7732 7733 void 7734 mdb_cursor_close(MDB_cursor *mc) 7735 { 7736 if (mc && !mc->mc_backup) { 7737 /* remove from txn, if tracked */ 7738 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) { 7739 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi]; 7740 while (*prev && *prev != mc) prev = &(*prev)->mc_next; 7741 if (*prev == mc) 7742 *prev = mc->mc_next; 7743 } 7744 free(mc); 7745 } 7746 } 7747 7748 MDB_txn * 7749 mdb_cursor_txn(MDB_cursor *mc) 7750 { 7751 if (!mc) return NULL; 7752 return mc->mc_txn; 7753 } 7754 7755 MDB_dbi 7756 mdb_cursor_dbi(MDB_cursor *mc) 7757 { 7758 return mc->mc_dbi; 7759 } 7760 7761 /** Replace the key for a branch node with a new key. 7762 * Set #MDB_TXN_ERROR on failure. 7763 * @param[in] mc Cursor pointing to the node to operate on. 7764 * @param[in] key The new key to use. 7765 * @return 0 on success, non-zero on failure. 7766 */ 7767 static int 7768 mdb_update_key(MDB_cursor *mc, MDB_val *key) 7769 { 7770 MDB_page *mp; 7771 MDB_node *node; 7772 char *base; 7773 size_t len; 7774 int delta, ksize, oksize; 7775 indx_t ptr, i, numkeys, indx; 7776 DKBUF; 7777 7778 indx = mc->mc_ki[mc->mc_top]; 7779 mp = mc->mc_pg[mc->mc_top]; 7780 node = NODEPTR(mp, indx); 7781 ptr = mp->mp_ptrs[indx]; 7782 #if MDB_DEBUG 7783 { 7784 MDB_val k2; 7785 char kbuf2[DKBUF_MAXKEYSIZE*2+1]; 7786 k2.mv_data = NODEKEY(node); 7787 k2.mv_size = node->mn_ksize; 7788 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u", 7789 indx, ptr, 7790 mdb_dkey(&k2, kbuf2), 7791 DKEY(key), 7792 mp->mp_pgno)); 7793 } 7794 #endif 7795 7796 /* Sizes must be 2-byte aligned. */ 7797 ksize = EVEN(key->mv_size); 7798 oksize = EVEN(node->mn_ksize); 7799 delta = ksize - oksize; 7800 7801 /* Shift node contents if EVEN(key length) changed. */ 7802 if (delta) { 7803 if (delta > 0 && SIZELEFT(mp) < delta) { 7804 pgno_t pgno; 7805 /* not enough space left, do a delete and split */ 7806 DPRINTF(("Not enough room, delta = %d, splitting...", delta)); 7807 pgno = NODEPGNO(node); 7808 mdb_node_del(mc, 0); 7809 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE); 7810 } 7811 7812 numkeys = NUMKEYS(mp); 7813 for (i = 0; i < numkeys; i++) { 7814 if (mp->mp_ptrs[i] <= ptr) 7815 mp->mp_ptrs[i] -= delta; 7816 } 7817 7818 base = (char *)mp + mp->mp_upper + PAGEBASE; 7819 len = ptr - mp->mp_upper + NODESIZE; 7820 memmove(base - delta, base, len); 7821 mp->mp_upper -= delta; 7822 7823 node = NODEPTR(mp, indx); 7824 } 7825 7826 /* But even if no shift was needed, update ksize */ 7827 if (node->mn_ksize != key->mv_size) 7828 node->mn_ksize = key->mv_size; 7829 7830 if (key->mv_size) 7831 memcpy(NODEKEY(node), key->mv_data, key->mv_size); 7832 7833 return MDB_SUCCESS; 7834 } 7835 7836 static void 7837 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst); 7838 7839 /** Perform \b act while tracking temporary cursor \b mn */ 7840 #define WITH_CURSOR_TRACKING(mn, act) do { \ 7841 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \ 7842 if ((mn).mc_flags & C_SUB) { \ 7843 dummy.mc_flags = C_INITIALIZED; \ 7844 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \ 7845 tracked = &dummy; \ 7846 } else { \ 7847 tracked = &(mn); \ 7848 } \ 7849 tracked->mc_next = *tp; \ 7850 *tp = tracked; \ 7851 { act; } \ 7852 *tp = tracked->mc_next; \ 7853 } while (0) 7854 7855 /** Move a node from csrc to cdst. 7856 */ 7857 static int 7858 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft) 7859 { 7860 MDB_node *srcnode; 7861 MDB_val key, data; 7862 pgno_t srcpg; 7863 MDB_cursor mn; 7864 int rc; 7865 unsigned short flags; 7866 7867 DKBUF; 7868 7869 /* Mark src and dst as dirty. */ 7870 if ((rc = mdb_page_touch(csrc)) || 7871 (rc = mdb_page_touch(cdst))) 7872 return rc; 7873 7874 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) { 7875 key.mv_size = csrc->mc_db->md_pad; 7876 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size); 7877 data.mv_size = 0; 7878 data.mv_data = NULL; 7879 srcpg = 0; 7880 flags = 0; 7881 } else { 7882 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]); 7883 mdb_cassert(csrc, !((size_t)srcnode & 1)); 7884 srcpg = NODEPGNO(srcnode); 7885 flags = srcnode->mn_flags; 7886 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) { 7887 unsigned int snum = csrc->mc_snum; 7888 MDB_node *s2; 7889 /* must find the lowest key below src */ 7890 rc = mdb_page_search_lowest(csrc); 7891 if (rc) 7892 return rc; 7893 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) { 7894 key.mv_size = csrc->mc_db->md_pad; 7895 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size); 7896 } else { 7897 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); 7898 key.mv_size = NODEKSZ(s2); 7899 key.mv_data = NODEKEY(s2); 7900 } 7901 csrc->mc_snum = snum--; 7902 csrc->mc_top = snum; 7903 } else { 7904 key.mv_size = NODEKSZ(srcnode); 7905 key.mv_data = NODEKEY(srcnode); 7906 } 7907 data.mv_size = NODEDSZ(srcnode); 7908 data.mv_data = NODEDATA(srcnode); 7909 } 7910 mn.mc_xcursor = NULL; 7911 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) { 7912 unsigned int snum = cdst->mc_snum; 7913 MDB_node *s2; 7914 MDB_val bkey; 7915 /* must find the lowest key below dst */ 7916 mdb_cursor_copy(cdst, &mn); 7917 rc = mdb_page_search_lowest(&mn); 7918 if (rc) 7919 return rc; 7920 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) { 7921 bkey.mv_size = mn.mc_db->md_pad; 7922 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size); 7923 } else { 7924 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0); 7925 bkey.mv_size = NODEKSZ(s2); 7926 bkey.mv_data = NODEKEY(s2); 7927 } 7928 mn.mc_snum = snum--; 7929 mn.mc_top = snum; 7930 mn.mc_ki[snum] = 0; 7931 rc = mdb_update_key(&mn, &bkey); 7932 if (rc) 7933 return rc; 7934 } 7935 7936 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u", 7937 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch", 7938 csrc->mc_ki[csrc->mc_top], 7939 DKEY(&key), 7940 csrc->mc_pg[csrc->mc_top]->mp_pgno, 7941 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno)); 7942 7943 /* Add the node to the destination page. 7944 */ 7945 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags); 7946 if (rc != MDB_SUCCESS) 7947 return rc; 7948 7949 /* Delete the node from the source page. 7950 */ 7951 mdb_node_del(csrc, key.mv_size); 7952 7953 { 7954 /* Adjust other cursors pointing to mp */ 7955 MDB_cursor *m2, *m3; 7956 MDB_dbi dbi = csrc->mc_dbi; 7957 MDB_page *mpd, *mps; 7958 7959 mps = csrc->mc_pg[csrc->mc_top]; 7960 /* If we're adding on the left, bump others up */ 7961 if (fromleft) { 7962 mpd = cdst->mc_pg[csrc->mc_top]; 7963 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 7964 if (csrc->mc_flags & C_SUB) 7965 m3 = &m2->mc_xcursor->mx_cursor; 7966 else 7967 m3 = m2; 7968 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top) 7969 continue; 7970 if (m3 != cdst && 7971 m3->mc_pg[csrc->mc_top] == mpd && 7972 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) { 7973 m3->mc_ki[csrc->mc_top]++; 7974 } 7975 if (m3 !=csrc && 7976 m3->mc_pg[csrc->mc_top] == mps && 7977 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) { 7978 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top]; 7979 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top]; 7980 m3->mc_ki[csrc->mc_top-1]++; 7981 } 7982 if (IS_LEAF(mps)) 7983 XCURSOR_REFRESH(m3, csrc->mc_top, m3->mc_pg[csrc->mc_top]); 7984 } 7985 } else 7986 /* Adding on the right, bump others down */ 7987 { 7988 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 7989 if (csrc->mc_flags & C_SUB) 7990 m3 = &m2->mc_xcursor->mx_cursor; 7991 else 7992 m3 = m2; 7993 if (m3 == csrc) continue; 7994 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top) 7995 continue; 7996 if (m3->mc_pg[csrc->mc_top] == mps) { 7997 if (!m3->mc_ki[csrc->mc_top]) { 7998 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top]; 7999 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top]; 8000 m3->mc_ki[csrc->mc_top-1]--; 8001 } else { 8002 m3->mc_ki[csrc->mc_top]--; 8003 } 8004 if (IS_LEAF(mps)) 8005 XCURSOR_REFRESH(m3, csrc->mc_top, m3->mc_pg[csrc->mc_top]); 8006 } 8007 } 8008 } 8009 } 8010 8011 /* Update the parent separators. 8012 */ 8013 if (csrc->mc_ki[csrc->mc_top] == 0) { 8014 if (csrc->mc_ki[csrc->mc_top-1] != 0) { 8015 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) { 8016 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size); 8017 } else { 8018 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); 8019 key.mv_size = NODEKSZ(srcnode); 8020 key.mv_data = NODEKEY(srcnode); 8021 } 8022 DPRINTF(("update separator for source page %"Z"u to [%s]", 8023 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key))); 8024 mdb_cursor_copy(csrc, &mn); 8025 mn.mc_snum--; 8026 mn.mc_top--; 8027 /* We want mdb_rebalance to find mn when doing fixups */ 8028 WITH_CURSOR_TRACKING(mn, 8029 rc = mdb_update_key(&mn, &key)); 8030 if (rc) 8031 return rc; 8032 } 8033 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) { 8034 MDB_val nullkey; 8035 indx_t ix = csrc->mc_ki[csrc->mc_top]; 8036 nullkey.mv_size = 0; 8037 csrc->mc_ki[csrc->mc_top] = 0; 8038 rc = mdb_update_key(csrc, &nullkey); 8039 csrc->mc_ki[csrc->mc_top] = ix; 8040 mdb_cassert(csrc, rc == MDB_SUCCESS); 8041 } 8042 } 8043 8044 if (cdst->mc_ki[cdst->mc_top] == 0) { 8045 if (cdst->mc_ki[cdst->mc_top-1] != 0) { 8046 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) { 8047 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size); 8048 } else { 8049 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0); 8050 key.mv_size = NODEKSZ(srcnode); 8051 key.mv_data = NODEKEY(srcnode); 8052 } 8053 DPRINTF(("update separator for destination page %"Z"u to [%s]", 8054 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key))); 8055 mdb_cursor_copy(cdst, &mn); 8056 mn.mc_snum--; 8057 mn.mc_top--; 8058 /* We want mdb_rebalance to find mn when doing fixups */ 8059 WITH_CURSOR_TRACKING(mn, 8060 rc = mdb_update_key(&mn, &key)); 8061 if (rc) 8062 return rc; 8063 } 8064 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) { 8065 MDB_val nullkey; 8066 indx_t ix = cdst->mc_ki[cdst->mc_top]; 8067 nullkey.mv_size = 0; 8068 cdst->mc_ki[cdst->mc_top] = 0; 8069 rc = mdb_update_key(cdst, &nullkey); 8070 cdst->mc_ki[cdst->mc_top] = ix; 8071 mdb_cassert(cdst, rc == MDB_SUCCESS); 8072 } 8073 } 8074 8075 return MDB_SUCCESS; 8076 } 8077 8078 /** Merge one page into another. 8079 * The nodes from the page pointed to by \b csrc will 8080 * be copied to the page pointed to by \b cdst and then 8081 * the \b csrc page will be freed. 8082 * @param[in] csrc Cursor pointing to the source page. 8083 * @param[in] cdst Cursor pointing to the destination page. 8084 * @return 0 on success, non-zero on failure. 8085 */ 8086 static int 8087 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst) 8088 { 8089 MDB_page *psrc, *pdst; 8090 MDB_node *srcnode; 8091 MDB_val key, data; 8092 unsigned nkeys; 8093 int rc; 8094 indx_t i, j; 8095 8096 psrc = csrc->mc_pg[csrc->mc_top]; 8097 pdst = cdst->mc_pg[cdst->mc_top]; 8098 8099 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno)); 8100 8101 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */ 8102 mdb_cassert(csrc, cdst->mc_snum > 1); 8103 8104 /* Mark dst as dirty. */ 8105 if ((rc = mdb_page_touch(cdst))) 8106 return rc; 8107 8108 /* get dst page again now that we've touched it. */ 8109 pdst = cdst->mc_pg[cdst->mc_top]; 8110 8111 /* Move all nodes from src to dst. 8112 */ 8113 j = nkeys = NUMKEYS(pdst); 8114 if (IS_LEAF2(psrc)) { 8115 key.mv_size = csrc->mc_db->md_pad; 8116 key.mv_data = METADATA(psrc); 8117 for (i = 0; i < NUMKEYS(psrc); i++, j++) { 8118 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0); 8119 if (rc != MDB_SUCCESS) 8120 return rc; 8121 key.mv_data = (char *)key.mv_data + key.mv_size; 8122 } 8123 } else { 8124 for (i = 0; i < NUMKEYS(psrc); i++, j++) { 8125 srcnode = NODEPTR(psrc, i); 8126 if (i == 0 && IS_BRANCH(psrc)) { 8127 MDB_cursor mn; 8128 MDB_node *s2; 8129 mdb_cursor_copy(csrc, &mn); 8130 mn.mc_xcursor = NULL; 8131 /* must find the lowest key below src */ 8132 rc = mdb_page_search_lowest(&mn); 8133 if (rc) 8134 return rc; 8135 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) { 8136 key.mv_size = mn.mc_db->md_pad; 8137 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size); 8138 } else { 8139 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0); 8140 key.mv_size = NODEKSZ(s2); 8141 key.mv_data = NODEKEY(s2); 8142 } 8143 } else { 8144 key.mv_size = srcnode->mn_ksize; 8145 key.mv_data = NODEKEY(srcnode); 8146 } 8147 8148 data.mv_size = NODEDSZ(srcnode); 8149 data.mv_data = NODEDATA(srcnode); 8150 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags); 8151 if (rc != MDB_SUCCESS) 8152 return rc; 8153 } 8154 } 8155 8156 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)", 8157 pdst->mp_pgno, NUMKEYS(pdst), 8158 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10)); 8159 8160 /* Unlink the src page from parent and add to free list. 8161 */ 8162 csrc->mc_top--; 8163 mdb_node_del(csrc, 0); 8164 if (csrc->mc_ki[csrc->mc_top] == 0) { 8165 key.mv_size = 0; 8166 rc = mdb_update_key(csrc, &key); 8167 if (rc) { 8168 csrc->mc_top++; 8169 return rc; 8170 } 8171 } 8172 csrc->mc_top++; 8173 8174 psrc = csrc->mc_pg[csrc->mc_top]; 8175 /* If not operating on FreeDB, allow this page to be reused 8176 * in this txn. Otherwise just add to free list. 8177 */ 8178 rc = mdb_page_loose(csrc, psrc); 8179 if (rc) 8180 return rc; 8181 if (IS_LEAF(psrc)) 8182 csrc->mc_db->md_leaf_pages--; 8183 else 8184 csrc->mc_db->md_branch_pages--; 8185 { 8186 /* Adjust other cursors pointing to mp */ 8187 MDB_cursor *m2, *m3; 8188 MDB_dbi dbi = csrc->mc_dbi; 8189 unsigned int top = csrc->mc_top; 8190 8191 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8192 if (csrc->mc_flags & C_SUB) 8193 m3 = &m2->mc_xcursor->mx_cursor; 8194 else 8195 m3 = m2; 8196 if (m3 == csrc) continue; 8197 if (m3->mc_snum < csrc->mc_snum) continue; 8198 if (m3->mc_pg[top] == psrc) { 8199 m3->mc_pg[top] = pdst; 8200 m3->mc_ki[top] += nkeys; 8201 m3->mc_ki[top-1] = cdst->mc_ki[top-1]; 8202 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] && 8203 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) { 8204 m3->mc_ki[top-1]--; 8205 } 8206 if (IS_LEAF(psrc)) 8207 XCURSOR_REFRESH(m3, top, m3->mc_pg[top]); 8208 } 8209 } 8210 { 8211 unsigned int snum = cdst->mc_snum; 8212 uint16_t depth = cdst->mc_db->md_depth; 8213 mdb_cursor_pop(cdst); 8214 rc = mdb_rebalance(cdst); 8215 /* Did the tree height change? */ 8216 if (depth != cdst->mc_db->md_depth) 8217 snum += cdst->mc_db->md_depth - depth; 8218 cdst->mc_snum = snum; 8219 cdst->mc_top = snum-1; 8220 } 8221 return rc; 8222 } 8223 8224 /** Copy the contents of a cursor. 8225 * @param[in] csrc The cursor to copy from. 8226 * @param[out] cdst The cursor to copy to. 8227 */ 8228 static void 8229 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst) 8230 { 8231 unsigned int i; 8232 8233 cdst->mc_txn = csrc->mc_txn; 8234 cdst->mc_dbi = csrc->mc_dbi; 8235 cdst->mc_db = csrc->mc_db; 8236 cdst->mc_dbx = csrc->mc_dbx; 8237 cdst->mc_snum = csrc->mc_snum; 8238 cdst->mc_top = csrc->mc_top; 8239 cdst->mc_flags = csrc->mc_flags; 8240 8241 for (i=0; i<csrc->mc_snum; i++) { 8242 cdst->mc_pg[i] = csrc->mc_pg[i]; 8243 cdst->mc_ki[i] = csrc->mc_ki[i]; 8244 } 8245 } 8246 8247 /** Rebalance the tree after a delete operation. 8248 * @param[in] mc Cursor pointing to the page where rebalancing 8249 * should begin. 8250 * @return 0 on success, non-zero on failure. 8251 */ 8252 static int 8253 mdb_rebalance(MDB_cursor *mc) 8254 { 8255 MDB_node *node; 8256 int rc, fromleft; 8257 unsigned int ptop, minkeys, thresh; 8258 MDB_cursor mn; 8259 indx_t oldki; 8260 8261 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) { 8262 minkeys = 2; 8263 thresh = 1; 8264 } else { 8265 minkeys = 1; 8266 thresh = FILL_THRESHOLD; 8267 } 8268 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)", 8269 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch", 8270 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]), 8271 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10)); 8272 8273 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh && 8274 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) { 8275 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold", 8276 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]))); 8277 return MDB_SUCCESS; 8278 } 8279 8280 if (mc->mc_snum < 2) { 8281 MDB_page *mp = mc->mc_pg[0]; 8282 if (IS_SUBP(mp)) { 8283 DPUTS("Can't rebalance a subpage, ignoring"); 8284 return MDB_SUCCESS; 8285 } 8286 if (NUMKEYS(mp) == 0) { 8287 DPUTS("tree is completely empty"); 8288 mc->mc_db->md_root = P_INVALID; 8289 mc->mc_db->md_depth = 0; 8290 mc->mc_db->md_leaf_pages = 0; 8291 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno); 8292 if (rc) 8293 return rc; 8294 /* Adjust cursors pointing to mp */ 8295 mc->mc_snum = 0; 8296 mc->mc_top = 0; 8297 mc->mc_flags &= ~C_INITIALIZED; 8298 { 8299 MDB_cursor *m2, *m3; 8300 MDB_dbi dbi = mc->mc_dbi; 8301 8302 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8303 if (mc->mc_flags & C_SUB) 8304 m3 = &m2->mc_xcursor->mx_cursor; 8305 else 8306 m3 = m2; 8307 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum)) 8308 continue; 8309 if (m3->mc_pg[0] == mp) { 8310 m3->mc_snum = 0; 8311 m3->mc_top = 0; 8312 m3->mc_flags &= ~C_INITIALIZED; 8313 } 8314 } 8315 } 8316 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) { 8317 int i; 8318 DPUTS("collapsing root page!"); 8319 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno); 8320 if (rc) 8321 return rc; 8322 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0)); 8323 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL); 8324 if (rc) 8325 return rc; 8326 mc->mc_db->md_depth--; 8327 mc->mc_db->md_branch_pages--; 8328 mc->mc_ki[0] = mc->mc_ki[1]; 8329 for (i = 1; i<mc->mc_db->md_depth; i++) { 8330 mc->mc_pg[i] = mc->mc_pg[i+1]; 8331 mc->mc_ki[i] = mc->mc_ki[i+1]; 8332 } 8333 { 8334 /* Adjust other cursors pointing to mp */ 8335 MDB_cursor *m2, *m3; 8336 MDB_dbi dbi = mc->mc_dbi; 8337 8338 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8339 if (mc->mc_flags & C_SUB) 8340 m3 = &m2->mc_xcursor->mx_cursor; 8341 else 8342 m3 = m2; 8343 if (m3 == mc) continue; 8344 if (!(m3->mc_flags & C_INITIALIZED)) 8345 continue; 8346 if (m3->mc_pg[0] == mp) { 8347 for (i=0; i<mc->mc_db->md_depth; i++) { 8348 m3->mc_pg[i] = m3->mc_pg[i+1]; 8349 m3->mc_ki[i] = m3->mc_ki[i+1]; 8350 } 8351 m3->mc_snum--; 8352 m3->mc_top--; 8353 } 8354 } 8355 } 8356 } else 8357 DPUTS("root page doesn't need rebalancing"); 8358 return MDB_SUCCESS; 8359 } 8360 8361 /* The parent (branch page) must have at least 2 pointers, 8362 * otherwise the tree is invalid. 8363 */ 8364 ptop = mc->mc_top-1; 8365 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1); 8366 8367 /* Leaf page fill factor is below the threshold. 8368 * Try to move keys from left or right neighbor, or 8369 * merge with a neighbor page. 8370 */ 8371 8372 /* Find neighbors. 8373 */ 8374 mdb_cursor_copy(mc, &mn); 8375 mn.mc_xcursor = NULL; 8376 8377 oldki = mc->mc_ki[mc->mc_top]; 8378 if (mc->mc_ki[ptop] == 0) { 8379 /* We're the leftmost leaf in our parent. 8380 */ 8381 DPUTS("reading right neighbor"); 8382 mn.mc_ki[ptop]++; 8383 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]); 8384 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL); 8385 if (rc) 8386 return rc; 8387 mn.mc_ki[mn.mc_top] = 0; 8388 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]); 8389 fromleft = 0; 8390 } else { 8391 /* There is at least one neighbor to the left. 8392 */ 8393 DPUTS("reading left neighbor"); 8394 mn.mc_ki[ptop]--; 8395 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]); 8396 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL); 8397 if (rc) 8398 return rc; 8399 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1; 8400 mc->mc_ki[mc->mc_top] = 0; 8401 fromleft = 1; 8402 } 8403 8404 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)", 8405 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), 8406 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10)); 8407 8408 /* If the neighbor page is above threshold and has enough keys, 8409 * move one key from it. Otherwise we should try to merge them. 8410 * (A branch page must never have less than 2 keys.) 8411 */ 8412 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) { 8413 rc = mdb_node_move(&mn, mc, fromleft); 8414 if (fromleft) { 8415 /* if we inserted on left, bump position up */ 8416 oldki++; 8417 } 8418 } else { 8419 if (!fromleft) { 8420 rc = mdb_page_merge(&mn, mc); 8421 } else { 8422 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]); 8423 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1; 8424 /* We want mdb_rebalance to find mn when doing fixups */ 8425 WITH_CURSOR_TRACKING(mn, 8426 rc = mdb_page_merge(mc, &mn)); 8427 mdb_cursor_copy(&mn, mc); 8428 } 8429 mc->mc_flags &= ~C_EOF; 8430 } 8431 mc->mc_ki[mc->mc_top] = oldki; 8432 return rc; 8433 } 8434 8435 /** Complete a delete operation started by #mdb_cursor_del(). */ 8436 static int 8437 mdb_cursor_del0(MDB_cursor *mc) 8438 { 8439 int rc; 8440 MDB_page *mp; 8441 indx_t ki; 8442 unsigned int nkeys; 8443 MDB_cursor *m2, *m3; 8444 MDB_dbi dbi = mc->mc_dbi; 8445 8446 ki = mc->mc_ki[mc->mc_top]; 8447 mp = mc->mc_pg[mc->mc_top]; 8448 mdb_node_del(mc, mc->mc_db->md_pad); 8449 mc->mc_db->md_entries--; 8450 { 8451 /* Adjust other cursors pointing to mp */ 8452 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8453 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; 8454 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED)) 8455 continue; 8456 if (m3 == mc || m3->mc_snum < mc->mc_snum) 8457 continue; 8458 if (m3->mc_pg[mc->mc_top] == mp) { 8459 if (m3->mc_ki[mc->mc_top] == ki) { 8460 m3->mc_flags |= C_DEL; 8461 if (mc->mc_db->md_flags & MDB_DUPSORT) { 8462 /* Sub-cursor referred into dataset which is gone */ 8463 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 8464 } 8465 continue; 8466 } else if (m3->mc_ki[mc->mc_top] > ki) { 8467 m3->mc_ki[mc->mc_top]--; 8468 } 8469 XCURSOR_REFRESH(m3, mc->mc_top, mp); 8470 } 8471 } 8472 } 8473 rc = mdb_rebalance(mc); 8474 if (rc) 8475 goto fail; 8476 8477 /* DB is totally empty now, just bail out. 8478 * Other cursors adjustments were already done 8479 * by mdb_rebalance and aren't needed here. 8480 */ 8481 if (!mc->mc_snum) { 8482 mc->mc_flags |= C_EOF; 8483 return rc; 8484 } 8485 8486 mp = mc->mc_pg[mc->mc_top]; 8487 nkeys = NUMKEYS(mp); 8488 8489 /* Adjust other cursors pointing to mp */ 8490 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) { 8491 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; 8492 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED)) 8493 continue; 8494 if (m3->mc_snum < mc->mc_snum) 8495 continue; 8496 if (m3->mc_pg[mc->mc_top] == mp) { 8497 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) { 8498 /* if m3 points past last node in page, find next sibling */ 8499 if (m3->mc_ki[mc->mc_top] >= nkeys) { 8500 rc = mdb_cursor_sibling(m3, 1); 8501 if (rc == MDB_NOTFOUND) { 8502 m3->mc_flags |= C_EOF; 8503 rc = MDB_SUCCESS; 8504 continue; 8505 } 8506 if (rc) 8507 goto fail; 8508 } 8509 if (m3->mc_xcursor && !(m3->mc_flags & C_EOF)) { 8510 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]); 8511 /* If this node has dupdata, it may need to be reinited 8512 * because its data has moved. 8513 * If the xcursor was not initd it must be reinited. 8514 * Else if node points to a subDB, nothing is needed. 8515 * Else (xcursor was initd, not a subDB) needs mc_pg[0] reset. 8516 */ 8517 if (node->mn_flags & F_DUPDATA) { 8518 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) { 8519 if (!(node->mn_flags & F_SUBDATA)) 8520 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node); 8521 } else { 8522 mdb_xcursor_init1(m3, node); 8523 rc = mdb_cursor_first(&m3->mc_xcursor->mx_cursor, NULL, NULL); 8524 if (rc) 8525 goto fail; 8526 } 8527 } 8528 m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL; 8529 } 8530 } 8531 } 8532 } 8533 mc->mc_flags |= C_DEL; 8534 8535 fail: 8536 if (rc) 8537 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 8538 return rc; 8539 } 8540 8541 int 8542 mdb_del(MDB_txn *txn, MDB_dbi dbi, 8543 MDB_val *key, MDB_val *data) 8544 { 8545 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 8546 return EINVAL; 8547 8548 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED)) 8549 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN; 8550 8551 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) { 8552 /* must ignore any data */ 8553 data = NULL; 8554 } 8555 8556 return mdb_del0(txn, dbi, key, data, 0); 8557 } 8558 8559 static int 8560 mdb_del0(MDB_txn *txn, MDB_dbi dbi, 8561 MDB_val *key, MDB_val *data, unsigned flags) 8562 { 8563 MDB_cursor mc; 8564 MDB_xcursor mx; 8565 MDB_cursor_op op; 8566 MDB_val rdata, *xdata; 8567 int rc, exact = 0; 8568 DKBUF; 8569 8570 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key))); 8571 8572 mdb_cursor_init(&mc, txn, dbi, &mx); 8573 8574 if (data) { 8575 op = MDB_GET_BOTH; 8576 rdata = *data; 8577 xdata = &rdata; 8578 } else { 8579 op = MDB_SET; 8580 xdata = NULL; 8581 flags |= MDB_NODUPDATA; 8582 } 8583 rc = mdb_cursor_set(&mc, key, xdata, op, &exact); 8584 if (rc == 0) { 8585 /* let mdb_page_split know about this cursor if needed: 8586 * delete will trigger a rebalance; if it needs to move 8587 * a node from one page to another, it will have to 8588 * update the parent's separator key(s). If the new sepkey 8589 * is larger than the current one, the parent page may 8590 * run out of space, triggering a split. We need this 8591 * cursor to be consistent until the end of the rebalance. 8592 */ 8593 mc.mc_flags |= C_UNTRACK; 8594 mc.mc_next = txn->mt_cursors[dbi]; 8595 txn->mt_cursors[dbi] = &mc; 8596 rc = mdb_cursor_del(&mc, flags); 8597 txn->mt_cursors[dbi] = mc.mc_next; 8598 } 8599 return rc; 8600 } 8601 8602 /** Split a page and insert a new node. 8603 * Set #MDB_TXN_ERROR on failure. 8604 * @param[in,out] mc Cursor pointing to the page and desired insertion index. 8605 * The cursor will be updated to point to the actual page and index where 8606 * the node got inserted after the split. 8607 * @param[in] newkey The key for the newly inserted node. 8608 * @param[in] newdata The data for the newly inserted node. 8609 * @param[in] newpgno The page number, if the new node is a branch node. 8610 * @param[in] nflags The #NODE_ADD_FLAGS for the new node. 8611 * @return 0 on success, non-zero on failure. 8612 */ 8613 static int 8614 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno, 8615 unsigned int nflags) 8616 { 8617 unsigned int flags; 8618 int rc = MDB_SUCCESS, new_root = 0, did_split = 0; 8619 indx_t newindx; 8620 pgno_t pgno = 0; 8621 int i, j, split_indx, nkeys, pmax; 8622 MDB_env *env = mc->mc_txn->mt_env; 8623 MDB_node *node; 8624 MDB_val sepkey, rkey, xdata, *rdata = &xdata; 8625 MDB_page *copy = NULL; 8626 MDB_page *mp, *rp, *pp; 8627 int ptop; 8628 MDB_cursor mn; 8629 DKBUF; 8630 8631 mp = mc->mc_pg[mc->mc_top]; 8632 newindx = mc->mc_ki[mc->mc_top]; 8633 nkeys = NUMKEYS(mp); 8634 8635 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i", 8636 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno, 8637 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys)); 8638 8639 /* Create a right sibling. */ 8640 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp))) 8641 return rc; 8642 rp->mp_pad = mp->mp_pad; 8643 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno)); 8644 8645 /* Usually when splitting the root page, the cursor 8646 * height is 1. But when called from mdb_update_key, 8647 * the cursor height may be greater because it walks 8648 * up the stack while finding the branch slot to update. 8649 */ 8650 if (mc->mc_top < 1) { 8651 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp))) 8652 goto done; 8653 /* shift current top to make room for new parent */ 8654 for (i=mc->mc_snum; i>0; i--) { 8655 mc->mc_pg[i] = mc->mc_pg[i-1]; 8656 mc->mc_ki[i] = mc->mc_ki[i-1]; 8657 } 8658 mc->mc_pg[0] = pp; 8659 mc->mc_ki[0] = 0; 8660 mc->mc_db->md_root = pp->mp_pgno; 8661 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno)); 8662 new_root = mc->mc_db->md_depth++; 8663 8664 /* Add left (implicit) pointer. */ 8665 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) { 8666 /* undo the pre-push */ 8667 mc->mc_pg[0] = mc->mc_pg[1]; 8668 mc->mc_ki[0] = mc->mc_ki[1]; 8669 mc->mc_db->md_root = mp->mp_pgno; 8670 mc->mc_db->md_depth--; 8671 goto done; 8672 } 8673 mc->mc_snum++; 8674 mc->mc_top++; 8675 ptop = 0; 8676 } else { 8677 ptop = mc->mc_top-1; 8678 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno)); 8679 } 8680 8681 mdb_cursor_copy(mc, &mn); 8682 mn.mc_xcursor = NULL; 8683 mn.mc_pg[mn.mc_top] = rp; 8684 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1; 8685 8686 if (nflags & MDB_APPEND) { 8687 mn.mc_ki[mn.mc_top] = 0; 8688 sepkey = *newkey; 8689 split_indx = newindx; 8690 nkeys = 0; 8691 } else { 8692 8693 split_indx = (nkeys+1) / 2; 8694 8695 if (IS_LEAF2(rp)) { 8696 char *split, *ins; 8697 int x; 8698 unsigned int lsize, rsize, ksize; 8699 /* Move half of the keys to the right sibling */ 8700 x = mc->mc_ki[mc->mc_top] - split_indx; 8701 ksize = mc->mc_db->md_pad; 8702 split = LEAF2KEY(mp, split_indx, ksize); 8703 rsize = (nkeys - split_indx) * ksize; 8704 lsize = (nkeys - split_indx) * sizeof(indx_t); 8705 mp->mp_lower -= lsize; 8706 rp->mp_lower += lsize; 8707 mp->mp_upper += rsize - lsize; 8708 rp->mp_upper -= rsize - lsize; 8709 sepkey.mv_size = ksize; 8710 if (newindx == split_indx) { 8711 sepkey.mv_data = newkey->mv_data; 8712 } else { 8713 sepkey.mv_data = split; 8714 } 8715 if (x<0) { 8716 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize); 8717 memcpy(rp->mp_ptrs, split, rsize); 8718 sepkey.mv_data = rp->mp_ptrs; 8719 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize); 8720 memcpy(ins, newkey->mv_data, ksize); 8721 mp->mp_lower += sizeof(indx_t); 8722 mp->mp_upper -= ksize - sizeof(indx_t); 8723 } else { 8724 if (x) 8725 memcpy(rp->mp_ptrs, split, x * ksize); 8726 ins = LEAF2KEY(rp, x, ksize); 8727 memcpy(ins, newkey->mv_data, ksize); 8728 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize); 8729 rp->mp_lower += sizeof(indx_t); 8730 rp->mp_upper -= ksize - sizeof(indx_t); 8731 mc->mc_ki[mc->mc_top] = x; 8732 } 8733 } else { 8734 int psize, nsize, k; 8735 /* Maximum free space in an empty page */ 8736 pmax = env->me_psize - PAGEHDRSZ; 8737 if (IS_LEAF(mp)) 8738 nsize = mdb_leaf_size(env, newkey, newdata); 8739 else 8740 nsize = mdb_branch_size(env, newkey); 8741 nsize = EVEN(nsize); 8742 8743 /* grab a page to hold a temporary copy */ 8744 copy = mdb_page_malloc(mc->mc_txn, 1); 8745 if (copy == NULL) { 8746 rc = ENOMEM; 8747 goto done; 8748 } 8749 copy->mp_pgno = mp->mp_pgno; 8750 copy->mp_flags = mp->mp_flags; 8751 copy->mp_lower = (PAGEHDRSZ-PAGEBASE); 8752 copy->mp_upper = env->me_psize - PAGEBASE; 8753 8754 /* prepare to insert */ 8755 for (i=0, j=0; i<nkeys; i++) { 8756 if (i == newindx) { 8757 copy->mp_ptrs[j++] = 0; 8758 } 8759 copy->mp_ptrs[j++] = mp->mp_ptrs[i]; 8760 } 8761 8762 /* When items are relatively large the split point needs 8763 * to be checked, because being off-by-one will make the 8764 * difference between success or failure in mdb_node_add. 8765 * 8766 * It's also relevant if a page happens to be laid out 8767 * such that one half of its nodes are all "small" and 8768 * the other half of its nodes are "large." If the new 8769 * item is also "large" and falls on the half with 8770 * "large" nodes, it also may not fit. 8771 * 8772 * As a final tweak, if the new item goes on the last 8773 * spot on the page (and thus, onto the new page), bias 8774 * the split so the new page is emptier than the old page. 8775 * This yields better packing during sequential inserts. 8776 */ 8777 if (nkeys < 32 || nsize > pmax/16 || newindx >= nkeys) { 8778 /* Find split point */ 8779 psize = 0; 8780 if (newindx <= split_indx || newindx >= nkeys) { 8781 i = 0; j = 1; 8782 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp); 8783 } else { 8784 i = nkeys; j = -1; 8785 k = split_indx-1; 8786 } 8787 for (; i!=k; i+=j) { 8788 if (i == newindx) { 8789 psize += nsize; 8790 node = NULL; 8791 } else { 8792 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE); 8793 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t); 8794 if (IS_LEAF(mp)) { 8795 if (F_ISSET(node->mn_flags, F_BIGDATA)) 8796 psize += sizeof(pgno_t); 8797 else 8798 psize += NODEDSZ(node); 8799 } 8800 psize = EVEN(psize); 8801 } 8802 if (psize > pmax || i == k-j) { 8803 split_indx = i + (j<0); 8804 break; 8805 } 8806 } 8807 } 8808 if (split_indx == newindx) { 8809 sepkey.mv_size = newkey->mv_size; 8810 sepkey.mv_data = newkey->mv_data; 8811 } else { 8812 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE); 8813 sepkey.mv_size = node->mn_ksize; 8814 sepkey.mv_data = NODEKEY(node); 8815 } 8816 } 8817 } 8818 8819 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey))); 8820 8821 /* Copy separator key to the parent. 8822 */ 8823 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) { 8824 int snum = mc->mc_snum; 8825 mn.mc_snum--; 8826 mn.mc_top--; 8827 did_split = 1; 8828 /* We want other splits to find mn when doing fixups */ 8829 WITH_CURSOR_TRACKING(mn, 8830 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0)); 8831 if (rc) 8832 goto done; 8833 8834 /* root split? */ 8835 if (mc->mc_snum > snum) { 8836 ptop++; 8837 } 8838 /* Right page might now have changed parent. 8839 * Check if left page also changed parent. 8840 */ 8841 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && 8842 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) { 8843 for (i=0; i<ptop; i++) { 8844 mc->mc_pg[i] = mn.mc_pg[i]; 8845 mc->mc_ki[i] = mn.mc_ki[i]; 8846 } 8847 mc->mc_pg[ptop] = mn.mc_pg[ptop]; 8848 if (mn.mc_ki[ptop]) { 8849 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1; 8850 } else { 8851 /* find right page's left sibling */ 8852 mc->mc_ki[ptop] = mn.mc_ki[ptop]; 8853 mdb_cursor_sibling(mc, 0); 8854 } 8855 } 8856 } else { 8857 mn.mc_top--; 8858 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0); 8859 mn.mc_top++; 8860 } 8861 if (rc != MDB_SUCCESS) { 8862 goto done; 8863 } 8864 if (nflags & MDB_APPEND) { 8865 mc->mc_pg[mc->mc_top] = rp; 8866 mc->mc_ki[mc->mc_top] = 0; 8867 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags); 8868 if (rc) 8869 goto done; 8870 for (i=0; i<mc->mc_top; i++) 8871 mc->mc_ki[i] = mn.mc_ki[i]; 8872 } else if (!IS_LEAF2(mp)) { 8873 /* Move nodes */ 8874 mc->mc_pg[mc->mc_top] = rp; 8875 i = split_indx; 8876 j = 0; 8877 do { 8878 if (i == newindx) { 8879 rkey.mv_data = newkey->mv_data; 8880 rkey.mv_size = newkey->mv_size; 8881 if (IS_LEAF(mp)) { 8882 rdata = newdata; 8883 } else 8884 pgno = newpgno; 8885 flags = nflags; 8886 /* Update index for the new key. */ 8887 mc->mc_ki[mc->mc_top] = j; 8888 } else { 8889 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE); 8890 rkey.mv_data = NODEKEY(node); 8891 rkey.mv_size = node->mn_ksize; 8892 if (IS_LEAF(mp)) { 8893 xdata.mv_data = NODEDATA(node); 8894 xdata.mv_size = NODEDSZ(node); 8895 rdata = &xdata; 8896 } else 8897 pgno = NODEPGNO(node); 8898 flags = node->mn_flags; 8899 } 8900 8901 if (!IS_LEAF(mp) && j == 0) { 8902 /* First branch index doesn't need key data. */ 8903 rkey.mv_size = 0; 8904 } 8905 8906 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags); 8907 if (rc) 8908 goto done; 8909 if (i == nkeys) { 8910 i = 0; 8911 j = 0; 8912 mc->mc_pg[mc->mc_top] = copy; 8913 } else { 8914 i++; 8915 j++; 8916 } 8917 } while (i != split_indx); 8918 8919 nkeys = NUMKEYS(copy); 8920 for (i=0; i<nkeys; i++) 8921 mp->mp_ptrs[i] = copy->mp_ptrs[i]; 8922 mp->mp_lower = copy->mp_lower; 8923 mp->mp_upper = copy->mp_upper; 8924 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1), 8925 env->me_psize - copy->mp_upper - PAGEBASE); 8926 8927 /* reset back to original page */ 8928 if (newindx < split_indx) { 8929 mc->mc_pg[mc->mc_top] = mp; 8930 } else { 8931 mc->mc_pg[mc->mc_top] = rp; 8932 mc->mc_ki[ptop]++; 8933 /* Make sure mc_ki is still valid. 8934 */ 8935 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && 8936 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) { 8937 for (i=0; i<=ptop; i++) { 8938 mc->mc_pg[i] = mn.mc_pg[i]; 8939 mc->mc_ki[i] = mn.mc_ki[i]; 8940 } 8941 } 8942 } 8943 if (nflags & MDB_RESERVE) { 8944 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 8945 if (!(node->mn_flags & F_BIGDATA)) 8946 newdata->mv_data = NODEDATA(node); 8947 } 8948 } else { 8949 if (newindx >= split_indx) { 8950 mc->mc_pg[mc->mc_top] = rp; 8951 mc->mc_ki[ptop]++; 8952 /* Make sure mc_ki is still valid. 8953 */ 8954 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && 8955 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) { 8956 for (i=0; i<=ptop; i++) { 8957 mc->mc_pg[i] = mn.mc_pg[i]; 8958 mc->mc_ki[i] = mn.mc_ki[i]; 8959 } 8960 } 8961 } 8962 } 8963 8964 { 8965 /* Adjust other cursors pointing to mp */ 8966 MDB_cursor *m2, *m3; 8967 MDB_dbi dbi = mc->mc_dbi; 8968 nkeys = NUMKEYS(mp); 8969 8970 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8971 if (mc->mc_flags & C_SUB) 8972 m3 = &m2->mc_xcursor->mx_cursor; 8973 else 8974 m3 = m2; 8975 if (m3 == mc) 8976 continue; 8977 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED)) 8978 continue; 8979 if (new_root) { 8980 int k; 8981 /* sub cursors may be on different DB */ 8982 if (m3->mc_pg[0] != mp) 8983 continue; 8984 /* root split */ 8985 for (k=new_root; k>=0; k--) { 8986 m3->mc_ki[k+1] = m3->mc_ki[k]; 8987 m3->mc_pg[k+1] = m3->mc_pg[k]; 8988 } 8989 if (m3->mc_ki[0] >= nkeys) { 8990 m3->mc_ki[0] = 1; 8991 } else { 8992 m3->mc_ki[0] = 0; 8993 } 8994 m3->mc_pg[0] = mc->mc_pg[0]; 8995 m3->mc_snum++; 8996 m3->mc_top++; 8997 } 8998 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) { 8999 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE)) 9000 m3->mc_ki[mc->mc_top]++; 9001 if (m3->mc_ki[mc->mc_top] >= nkeys) { 9002 m3->mc_pg[mc->mc_top] = rp; 9003 m3->mc_ki[mc->mc_top] -= nkeys; 9004 for (i=0; i<mc->mc_top; i++) { 9005 m3->mc_ki[i] = mn.mc_ki[i]; 9006 m3->mc_pg[i] = mn.mc_pg[i]; 9007 } 9008 } 9009 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] && 9010 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) { 9011 m3->mc_ki[ptop]++; 9012 } 9013 if (IS_LEAF(mp)) 9014 XCURSOR_REFRESH(m3, mc->mc_top, m3->mc_pg[mc->mc_top]); 9015 } 9016 } 9017 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp))); 9018 9019 done: 9020 if (copy) /* tmp page */ 9021 mdb_page_free(env, copy); 9022 if (rc) 9023 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 9024 return rc; 9025 } 9026 9027 int 9028 mdb_put(MDB_txn *txn, MDB_dbi dbi, 9029 MDB_val *key, MDB_val *data, unsigned int flags) 9030 { 9031 MDB_cursor mc; 9032 MDB_xcursor mx; 9033 int rc; 9034 9035 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 9036 return EINVAL; 9037 9038 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) 9039 return EINVAL; 9040 9041 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED)) 9042 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN; 9043 9044 mdb_cursor_init(&mc, txn, dbi, &mx); 9045 mc.mc_next = txn->mt_cursors[dbi]; 9046 txn->mt_cursors[dbi] = &mc; 9047 rc = mdb_cursor_put(&mc, key, data, flags); 9048 txn->mt_cursors[dbi] = mc.mc_next; 9049 return rc; 9050 } 9051 9052 #ifndef MDB_WBUF 9053 #define MDB_WBUF (1024*1024) 9054 #endif 9055 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */ 9056 9057 /** State needed for a double-buffering compacting copy. */ 9058 typedef struct mdb_copy { 9059 MDB_env *mc_env; 9060 MDB_txn *mc_txn; 9061 pthread_mutex_t mc_mutex; 9062 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */ 9063 char *mc_wbuf[2]; 9064 char *mc_over[2]; 9065 int mc_wlen[2]; 9066 int mc_olen[2]; 9067 pgno_t mc_next_pgno; 9068 HANDLE mc_fd; 9069 int mc_toggle; /**< Buffer number in provider */ 9070 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */ 9071 /** Error code. Never cleared if set. Both threads can set nonzero 9072 * to fail the copy. Not mutex-protected, LMDB expects atomic int. 9073 */ 9074 volatile int mc_error; 9075 } mdb_copy; 9076 9077 /** Dedicated writer thread for compacting copy. */ 9078 static THREAD_RET ESECT CALL_CONV 9079 mdb_env_copythr(void *arg) 9080 { 9081 mdb_copy *my = arg; 9082 char *ptr; 9083 int toggle = 0, wsize, rc; 9084 #ifdef _WIN32 9085 DWORD len; 9086 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL) 9087 #else 9088 int len; 9089 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0) 9090 #ifdef SIGPIPE 9091 sigset_t set; 9092 sigemptyset(&set); 9093 sigaddset(&set, SIGPIPE); 9094 if ((rc = pthread_sigmask(SIG_BLOCK, &set, NULL)) != 0) 9095 my->mc_error = rc; 9096 #endif 9097 #endif 9098 9099 pthread_mutex_lock(&my->mc_mutex); 9100 for(;;) { 9101 while (!my->mc_new) 9102 pthread_cond_wait(&my->mc_cond, &my->mc_mutex); 9103 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */ 9104 break; 9105 wsize = my->mc_wlen[toggle]; 9106 ptr = my->mc_wbuf[toggle]; 9107 again: 9108 rc = MDB_SUCCESS; 9109 while (wsize > 0 && !my->mc_error) { 9110 DO_WRITE(rc, my->mc_fd, ptr, wsize, len); 9111 if (!rc) { 9112 rc = ErrCode(); 9113 #if defined(SIGPIPE) && !defined(_WIN32) 9114 if (rc == EPIPE) { 9115 /* Collect the pending SIGPIPE, otherwise at least OS X 9116 * gives it to the process on thread-exit (ITS#8504). 9117 */ 9118 int tmp; 9119 sigwait(&set, &tmp); 9120 } 9121 #endif 9122 break; 9123 } else if (len > 0) { 9124 rc = MDB_SUCCESS; 9125 ptr += len; 9126 wsize -= len; 9127 continue; 9128 } else { 9129 rc = EIO; 9130 break; 9131 } 9132 } 9133 if (rc) { 9134 my->mc_error = rc; 9135 } 9136 /* If there's an overflow page tail, write it too */ 9137 if (my->mc_olen[toggle]) { 9138 wsize = my->mc_olen[toggle]; 9139 ptr = my->mc_over[toggle]; 9140 my->mc_olen[toggle] = 0; 9141 goto again; 9142 } 9143 my->mc_wlen[toggle] = 0; 9144 toggle ^= 1; 9145 /* Return the empty buffer to provider */ 9146 my->mc_new--; 9147 pthread_cond_signal(&my->mc_cond); 9148 } 9149 pthread_mutex_unlock(&my->mc_mutex); 9150 return (THREAD_RET)0; 9151 #undef DO_WRITE 9152 } 9153 9154 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer. 9155 * 9156 * @param[in] my control structure. 9157 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending). 9158 */ 9159 static int ESECT 9160 mdb_env_cthr_toggle(mdb_copy *my, int adjust) 9161 { 9162 pthread_mutex_lock(&my->mc_mutex); 9163 my->mc_new += adjust; 9164 pthread_cond_signal(&my->mc_cond); 9165 while (my->mc_new & 2) /* both buffers in use */ 9166 pthread_cond_wait(&my->mc_cond, &my->mc_mutex); 9167 pthread_mutex_unlock(&my->mc_mutex); 9168 9169 my->mc_toggle ^= (adjust & 1); 9170 /* Both threads reset mc_wlen, to be safe from threading errors */ 9171 my->mc_wlen[my->mc_toggle] = 0; 9172 return my->mc_error; 9173 } 9174 9175 /** Depth-first tree traversal for compacting copy. 9176 * @param[in] my control structure. 9177 * @param[in,out] pg database root. 9178 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB. 9179 */ 9180 static int ESECT 9181 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags) 9182 { 9183 MDB_cursor mc = {0}; 9184 MDB_node *ni; 9185 MDB_page *mo, *mp, *leaf; 9186 char *buf, *ptr; 9187 int rc, toggle; 9188 unsigned int i; 9189 9190 /* Empty DB, nothing to do */ 9191 if (*pg == P_INVALID) 9192 return MDB_SUCCESS; 9193 9194 mc.mc_snum = 1; 9195 mc.mc_txn = my->mc_txn; 9196 9197 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL); 9198 if (rc) 9199 return rc; 9200 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST); 9201 if (rc) 9202 return rc; 9203 9204 /* Make cursor pages writable */ 9205 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum); 9206 if (buf == NULL) 9207 return ENOMEM; 9208 9209 for (i=0; i<mc.mc_top; i++) { 9210 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize); 9211 mc.mc_pg[i] = (MDB_page *)ptr; 9212 ptr += my->mc_env->me_psize; 9213 } 9214 9215 /* This is writable space for a leaf page. Usually not needed. */ 9216 leaf = (MDB_page *)ptr; 9217 9218 toggle = my->mc_toggle; 9219 while (mc.mc_snum > 0) { 9220 unsigned n; 9221 mp = mc.mc_pg[mc.mc_top]; 9222 n = NUMKEYS(mp); 9223 9224 if (IS_LEAF(mp)) { 9225 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) { 9226 for (i=0; i<n; i++) { 9227 ni = NODEPTR(mp, i); 9228 if (ni->mn_flags & F_BIGDATA) { 9229 MDB_page *omp; 9230 pgno_t pg; 9231 9232 /* Need writable leaf */ 9233 if (mp != leaf) { 9234 mc.mc_pg[mc.mc_top] = leaf; 9235 mdb_page_copy(leaf, mp, my->mc_env->me_psize); 9236 mp = leaf; 9237 ni = NODEPTR(mp, i); 9238 } 9239 9240 memcpy(&pg, NODEDATA(ni), sizeof(pg)); 9241 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t)); 9242 rc = mdb_page_get(&mc, pg, &omp, NULL); 9243 if (rc) 9244 goto done; 9245 if (my->mc_wlen[toggle] >= MDB_WBUF) { 9246 rc = mdb_env_cthr_toggle(my, 1); 9247 if (rc) 9248 goto done; 9249 toggle = my->mc_toggle; 9250 } 9251 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]); 9252 memcpy(mo, omp, my->mc_env->me_psize); 9253 mo->mp_pgno = my->mc_next_pgno; 9254 my->mc_next_pgno += omp->mp_pages; 9255 my->mc_wlen[toggle] += my->mc_env->me_psize; 9256 if (omp->mp_pages > 1) { 9257 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1); 9258 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize; 9259 rc = mdb_env_cthr_toggle(my, 1); 9260 if (rc) 9261 goto done; 9262 toggle = my->mc_toggle; 9263 } 9264 } else if (ni->mn_flags & F_SUBDATA) { 9265 MDB_db db; 9266 9267 /* Need writable leaf */ 9268 if (mp != leaf) { 9269 mc.mc_pg[mc.mc_top] = leaf; 9270 mdb_page_copy(leaf, mp, my->mc_env->me_psize); 9271 mp = leaf; 9272 ni = NODEPTR(mp, i); 9273 } 9274 9275 memcpy(&db, NODEDATA(ni), sizeof(db)); 9276 my->mc_toggle = toggle; 9277 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA); 9278 if (rc) 9279 goto done; 9280 toggle = my->mc_toggle; 9281 memcpy(NODEDATA(ni), &db, sizeof(db)); 9282 } 9283 } 9284 } 9285 } else { 9286 mc.mc_ki[mc.mc_top]++; 9287 if (mc.mc_ki[mc.mc_top] < n) { 9288 pgno_t pg; 9289 again: 9290 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]); 9291 pg = NODEPGNO(ni); 9292 rc = mdb_page_get(&mc, pg, &mp, NULL); 9293 if (rc) 9294 goto done; 9295 mc.mc_top++; 9296 mc.mc_snum++; 9297 mc.mc_ki[mc.mc_top] = 0; 9298 if (IS_BRANCH(mp)) { 9299 /* Whenever we advance to a sibling branch page, 9300 * we must proceed all the way down to its first leaf. 9301 */ 9302 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize); 9303 goto again; 9304 } else 9305 mc.mc_pg[mc.mc_top] = mp; 9306 continue; 9307 } 9308 } 9309 if (my->mc_wlen[toggle] >= MDB_WBUF) { 9310 rc = mdb_env_cthr_toggle(my, 1); 9311 if (rc) 9312 goto done; 9313 toggle = my->mc_toggle; 9314 } 9315 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]); 9316 mdb_page_copy(mo, mp, my->mc_env->me_psize); 9317 mo->mp_pgno = my->mc_next_pgno++; 9318 my->mc_wlen[toggle] += my->mc_env->me_psize; 9319 if (mc.mc_top) { 9320 /* Update parent if there is one */ 9321 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]); 9322 SETPGNO(ni, mo->mp_pgno); 9323 mdb_cursor_pop(&mc); 9324 } else { 9325 /* Otherwise we're done */ 9326 *pg = mo->mp_pgno; 9327 break; 9328 } 9329 } 9330 done: 9331 free(buf); 9332 return rc; 9333 } 9334 9335 /** Copy environment with compaction. */ 9336 static int ESECT 9337 mdb_env_copyfd1(MDB_env *env, HANDLE fd) 9338 { 9339 MDB_meta *mm; 9340 MDB_page *mp; 9341 mdb_copy my = {0}; 9342 MDB_txn *txn = NULL; 9343 pthread_t thr; 9344 pgno_t root, new_root; 9345 int rc = MDB_SUCCESS; 9346 9347 #ifdef _WIN32 9348 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) || 9349 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) { 9350 rc = ErrCode(); 9351 goto done; 9352 } 9353 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize); 9354 if (my.mc_wbuf[0] == NULL) { 9355 /* _aligned_malloc() sets errno, but we use Windows error codes */ 9356 rc = ERROR_NOT_ENOUGH_MEMORY; 9357 goto done; 9358 } 9359 #else 9360 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0) 9361 return rc; 9362 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0) 9363 goto done2; 9364 #ifdef HAVE_MEMALIGN 9365 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2); 9366 if (my.mc_wbuf[0] == NULL) { 9367 rc = errno; 9368 goto done; 9369 } 9370 #else 9371 { 9372 void *p; 9373 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0) 9374 goto done; 9375 my.mc_wbuf[0] = p; 9376 } 9377 #endif 9378 #endif 9379 memset(my.mc_wbuf[0], 0, MDB_WBUF*2); 9380 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF; 9381 my.mc_next_pgno = NUM_METAS; 9382 my.mc_env = env; 9383 my.mc_fd = fd; 9384 rc = THREAD_CREATE(thr, mdb_env_copythr, &my); 9385 if (rc) 9386 goto done; 9387 9388 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn); 9389 if (rc) 9390 goto finish; 9391 9392 mp = (MDB_page *)my.mc_wbuf[0]; 9393 memset(mp, 0, NUM_METAS * env->me_psize); 9394 mp->mp_pgno = 0; 9395 mp->mp_flags = P_META; 9396 mm = (MDB_meta *)METADATA(mp); 9397 mdb_env_init_meta0(env, mm); 9398 mm->mm_address = env->me_metas[0]->mm_address; 9399 9400 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize); 9401 mp->mp_pgno = 1; 9402 mp->mp_flags = P_META; 9403 *(MDB_meta *)METADATA(mp) = *mm; 9404 mm = (MDB_meta *)METADATA(mp); 9405 9406 /* Set metapage 1 with current main DB */ 9407 root = new_root = txn->mt_dbs[MAIN_DBI].md_root; 9408 if (root != P_INVALID) { 9409 /* Count free pages + freeDB pages. Subtract from last_pg 9410 * to find the new last_pg, which also becomes the new root. 9411 */ 9412 MDB_ID freecount = 0; 9413 MDB_cursor mc; 9414 MDB_val key, data; 9415 mdb_cursor_init(&mc, txn, FREE_DBI, NULL); 9416 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0) 9417 freecount += *(MDB_ID *)data.mv_data; 9418 if (rc != MDB_NOTFOUND) 9419 goto finish; 9420 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages + 9421 txn->mt_dbs[FREE_DBI].md_leaf_pages + 9422 txn->mt_dbs[FREE_DBI].md_overflow_pages; 9423 9424 new_root = txn->mt_next_pgno - 1 - freecount; 9425 mm->mm_last_pg = new_root; 9426 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI]; 9427 mm->mm_dbs[MAIN_DBI].md_root = new_root; 9428 } else { 9429 /* When the DB is empty, handle it specially to 9430 * fix any breakage like page leaks from ITS#8174. 9431 */ 9432 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags; 9433 } 9434 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) { 9435 mm->mm_txnid = 1; /* use metapage 1 */ 9436 } 9437 9438 my.mc_wlen[0] = env->me_psize * NUM_METAS; 9439 my.mc_txn = txn; 9440 rc = mdb_env_cwalk(&my, &root, 0); 9441 if (rc == MDB_SUCCESS && root != new_root) { 9442 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */ 9443 } 9444 9445 finish: 9446 if (rc) 9447 my.mc_error = rc; 9448 mdb_env_cthr_toggle(&my, 1 | MDB_EOF); 9449 rc = THREAD_FINISH(thr); 9450 mdb_txn_abort(txn); 9451 9452 done: 9453 #ifdef _WIN32 9454 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]); 9455 if (my.mc_cond) CloseHandle(my.mc_cond); 9456 if (my.mc_mutex) CloseHandle(my.mc_mutex); 9457 #else 9458 free(my.mc_wbuf[0]); 9459 pthread_cond_destroy(&my.mc_cond); 9460 done2: 9461 pthread_mutex_destroy(&my.mc_mutex); 9462 #endif 9463 return rc ? rc : my.mc_error; 9464 } 9465 9466 /** Copy environment as-is. */ 9467 static int ESECT 9468 mdb_env_copyfd0(MDB_env *env, HANDLE fd) 9469 { 9470 MDB_txn *txn = NULL; 9471 mdb_mutexref_t wmutex = NULL; 9472 int rc; 9473 size_t wsize, w3; 9474 char *ptr; 9475 #ifdef _WIN32 9476 DWORD len, w2; 9477 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL) 9478 #else 9479 ssize_t len; 9480 size_t w2; 9481 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0) 9482 #endif 9483 9484 /* Do the lock/unlock of the reader mutex before starting the 9485 * write txn. Otherwise other read txns could block writers. 9486 */ 9487 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn); 9488 if (rc) 9489 return rc; 9490 9491 if (env->me_txns) { 9492 /* We must start the actual read txn after blocking writers */ 9493 mdb_txn_end(txn, MDB_END_RESET_TMP); 9494 9495 /* Temporarily block writers until we snapshot the meta pages */ 9496 wmutex = env->me_wmutex; 9497 if (LOCK_MUTEX(rc, env, wmutex)) 9498 goto leave; 9499 9500 rc = mdb_txn_renew0(txn); 9501 if (rc) { 9502 UNLOCK_MUTEX(wmutex); 9503 goto leave; 9504 } 9505 } 9506 9507 wsize = env->me_psize * NUM_METAS; 9508 ptr = env->me_map; 9509 w2 = wsize; 9510 while (w2 > 0) { 9511 DO_WRITE(rc, fd, ptr, w2, len); 9512 if (!rc) { 9513 rc = ErrCode(); 9514 break; 9515 } else if (len > 0) { 9516 rc = MDB_SUCCESS; 9517 ptr += len; 9518 w2 -= len; 9519 continue; 9520 } else { 9521 /* Non-blocking or async handles are not supported */ 9522 rc = EIO; 9523 break; 9524 } 9525 } 9526 if (wmutex) 9527 UNLOCK_MUTEX(wmutex); 9528 9529 if (rc) 9530 goto leave; 9531 9532 w3 = txn->mt_next_pgno * env->me_psize; 9533 { 9534 size_t fsize = 0; 9535 if ((rc = mdb_fsize(env->me_fd, &fsize))) 9536 goto leave; 9537 if (w3 > fsize) 9538 w3 = fsize; 9539 } 9540 wsize = w3 - wsize; 9541 while (wsize > 0) { 9542 if (wsize > MAX_WRITE) 9543 w2 = MAX_WRITE; 9544 else 9545 w2 = wsize; 9546 DO_WRITE(rc, fd, ptr, w2, len); 9547 if (!rc) { 9548 rc = ErrCode(); 9549 break; 9550 } else if (len > 0) { 9551 rc = MDB_SUCCESS; 9552 ptr += len; 9553 wsize -= len; 9554 continue; 9555 } else { 9556 rc = EIO; 9557 break; 9558 } 9559 } 9560 9561 leave: 9562 mdb_txn_abort(txn); 9563 return rc; 9564 } 9565 9566 int ESECT 9567 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags) 9568 { 9569 if (flags & MDB_CP_COMPACT) 9570 return mdb_env_copyfd1(env, fd); 9571 else 9572 return mdb_env_copyfd0(env, fd); 9573 } 9574 9575 int ESECT 9576 mdb_env_copyfd(MDB_env *env, HANDLE fd) 9577 { 9578 return mdb_env_copyfd2(env, fd, 0); 9579 } 9580 9581 int ESECT 9582 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags) 9583 { 9584 int rc; 9585 MDB_name fname; 9586 HANDLE newfd = INVALID_HANDLE_VALUE; 9587 9588 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname); 9589 if (rc == MDB_SUCCESS) { 9590 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd); 9591 mdb_fname_destroy(fname); 9592 } 9593 if (rc == MDB_SUCCESS) { 9594 rc = mdb_env_copyfd2(env, newfd, flags); 9595 if (close(newfd) < 0 && rc == MDB_SUCCESS) 9596 rc = ErrCode(); 9597 } 9598 return rc; 9599 } 9600 9601 int ESECT 9602 mdb_env_copy(MDB_env *env, const char *path) 9603 { 9604 return mdb_env_copy2(env, path, 0); 9605 } 9606 9607 int ESECT 9608 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff) 9609 { 9610 if (flag & ~CHANGEABLE) 9611 return EINVAL; 9612 if (onoff) 9613 env->me_flags |= flag; 9614 else 9615 env->me_flags &= ~flag; 9616 return MDB_SUCCESS; 9617 } 9618 9619 int ESECT 9620 mdb_env_get_flags(MDB_env *env, unsigned int *arg) 9621 { 9622 if (!env || !arg) 9623 return EINVAL; 9624 9625 *arg = env->me_flags & (CHANGEABLE|CHANGELESS); 9626 return MDB_SUCCESS; 9627 } 9628 9629 int ESECT 9630 mdb_env_set_userctx(MDB_env *env, void *ctx) 9631 { 9632 if (!env) 9633 return EINVAL; 9634 env->me_userctx = ctx; 9635 return MDB_SUCCESS; 9636 } 9637 9638 void * ESECT 9639 mdb_env_get_userctx(MDB_env *env) 9640 { 9641 return env ? env->me_userctx : NULL; 9642 } 9643 9644 int ESECT 9645 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func) 9646 { 9647 if (!env) 9648 return EINVAL; 9649 #ifndef NDEBUG 9650 env->me_assert_func = func; 9651 #endif 9652 return MDB_SUCCESS; 9653 } 9654 9655 int ESECT 9656 mdb_env_get_path(MDB_env *env, const char **arg) 9657 { 9658 if (!env || !arg) 9659 return EINVAL; 9660 9661 *arg = env->me_path; 9662 return MDB_SUCCESS; 9663 } 9664 9665 int ESECT 9666 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg) 9667 { 9668 if (!env || !arg) 9669 return EINVAL; 9670 9671 *arg = env->me_fd; 9672 return MDB_SUCCESS; 9673 } 9674 9675 /** Common code for #mdb_stat() and #mdb_env_stat(). 9676 * @param[in] env the environment to operate in. 9677 * @param[in] db the #MDB_db record containing the stats to return. 9678 * @param[out] arg the address of an #MDB_stat structure to receive the stats. 9679 * @return 0, this function always succeeds. 9680 */ 9681 static int ESECT 9682 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg) 9683 { 9684 arg->ms_psize = env->me_psize; 9685 arg->ms_depth = db->md_depth; 9686 arg->ms_branch_pages = db->md_branch_pages; 9687 arg->ms_leaf_pages = db->md_leaf_pages; 9688 arg->ms_overflow_pages = db->md_overflow_pages; 9689 arg->ms_entries = db->md_entries; 9690 9691 return MDB_SUCCESS; 9692 } 9693 9694 int ESECT 9695 mdb_env_stat(MDB_env *env, MDB_stat *arg) 9696 { 9697 MDB_meta *meta; 9698 9699 if (env == NULL || arg == NULL) 9700 return EINVAL; 9701 9702 meta = mdb_env_pick_meta(env); 9703 9704 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg); 9705 } 9706 9707 int ESECT 9708 mdb_env_info(MDB_env *env, MDB_envinfo *arg) 9709 { 9710 MDB_meta *meta; 9711 9712 if (env == NULL || arg == NULL) 9713 return EINVAL; 9714 9715 meta = mdb_env_pick_meta(env); 9716 arg->me_mapaddr = meta->mm_address; 9717 arg->me_last_pgno = meta->mm_last_pg; 9718 arg->me_last_txnid = meta->mm_txnid; 9719 9720 arg->me_mapsize = env->me_mapsize; 9721 arg->me_maxreaders = env->me_maxreaders; 9722 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0; 9723 return MDB_SUCCESS; 9724 } 9725 9726 /** Set the default comparison functions for a database. 9727 * Called immediately after a database is opened to set the defaults. 9728 * The user can then override them with #mdb_set_compare() or 9729 * #mdb_set_dupsort(). 9730 * @param[in] txn A transaction handle returned by #mdb_txn_begin() 9731 * @param[in] dbi A database handle returned by #mdb_dbi_open() 9732 */ 9733 static void 9734 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi) 9735 { 9736 uint16_t f = txn->mt_dbs[dbi].md_flags; 9737 9738 txn->mt_dbxs[dbi].md_cmp = 9739 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr : 9740 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn; 9741 9742 txn->mt_dbxs[dbi].md_dcmp = 9743 !(f & MDB_DUPSORT) ? 0 : 9744 ((f & MDB_INTEGERDUP) 9745 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint) 9746 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn)); 9747 } 9748 9749 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi) 9750 { 9751 MDB_val key, data; 9752 MDB_dbi i; 9753 MDB_cursor mc; 9754 MDB_db dummy; 9755 int rc, dbflag, exact; 9756 unsigned int unused = 0, seq; 9757 char *namedup; 9758 size_t len; 9759 9760 if (flags & ~VALID_FLAGS) 9761 return EINVAL; 9762 if (txn->mt_flags & MDB_TXN_BLOCKED) 9763 return MDB_BAD_TXN; 9764 9765 /* main DB? */ 9766 if (!name) { 9767 *dbi = MAIN_DBI; 9768 if (flags & PERSISTENT_FLAGS) { 9769 uint16_t f2 = flags & PERSISTENT_FLAGS; 9770 /* make sure flag changes get committed */ 9771 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) { 9772 txn->mt_dbs[MAIN_DBI].md_flags |= f2; 9773 txn->mt_flags |= MDB_TXN_DIRTY; 9774 } 9775 } 9776 mdb_default_cmp(txn, MAIN_DBI); 9777 return MDB_SUCCESS; 9778 } 9779 9780 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) { 9781 mdb_default_cmp(txn, MAIN_DBI); 9782 } 9783 9784 /* Is the DB already open? */ 9785 len = strlen(name); 9786 for (i=CORE_DBS; i<txn->mt_numdbs; i++) { 9787 if (!txn->mt_dbxs[i].md_name.mv_size) { 9788 /* Remember this free slot */ 9789 if (!unused) unused = i; 9790 continue; 9791 } 9792 if (len == txn->mt_dbxs[i].md_name.mv_size && 9793 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) { 9794 *dbi = i; 9795 return MDB_SUCCESS; 9796 } 9797 } 9798 9799 /* If no free slot and max hit, fail */ 9800 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs) 9801 return MDB_DBS_FULL; 9802 9803 /* Cannot mix named databases with some mainDB flags */ 9804 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY)) 9805 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND; 9806 9807 /* Find the DB info */ 9808 dbflag = DB_NEW|DB_VALID|DB_USRVALID; 9809 exact = 0; 9810 key.mv_size = len; 9811 key.mv_data = (void *)name; 9812 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL); 9813 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact); 9814 if (rc == MDB_SUCCESS) { 9815 /* make sure this is actually a DB */ 9816 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]); 9817 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA) 9818 return MDB_INCOMPATIBLE; 9819 } else { 9820 if (rc != MDB_NOTFOUND || !(flags & MDB_CREATE)) 9821 return rc; 9822 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) 9823 return EACCES; 9824 } 9825 9826 /* Done here so we cannot fail after creating a new DB */ 9827 if ((namedup = strdup(name)) == NULL) 9828 return ENOMEM; 9829 9830 if (rc) { 9831 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */ 9832 data.mv_size = sizeof(MDB_db); 9833 data.mv_data = &dummy; 9834 memset(&dummy, 0, sizeof(dummy)); 9835 dummy.md_root = P_INVALID; 9836 dummy.md_flags = flags & PERSISTENT_FLAGS; 9837 WITH_CURSOR_TRACKING(mc, 9838 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA)); 9839 dbflag |= DB_DIRTY; 9840 } 9841 9842 if (rc) { 9843 free(namedup); 9844 } else { 9845 /* Got info, register DBI in this txn */ 9846 unsigned int slot = unused ? unused : txn->mt_numdbs; 9847 txn->mt_dbxs[slot].md_name.mv_data = namedup; 9848 txn->mt_dbxs[slot].md_name.mv_size = len; 9849 txn->mt_dbxs[slot].md_rel = NULL; 9850 txn->mt_dbflags[slot] = dbflag; 9851 /* txn-> and env-> are the same in read txns, use 9852 * tmp variable to avoid undefined assignment 9853 */ 9854 seq = ++txn->mt_env->me_dbiseqs[slot]; 9855 txn->mt_dbiseqs[slot] = seq; 9856 9857 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db)); 9858 *dbi = slot; 9859 mdb_default_cmp(txn, slot); 9860 if (!unused) { 9861 txn->mt_numdbs++; 9862 } 9863 } 9864 9865 return rc; 9866 } 9867 9868 int ESECT 9869 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg) 9870 { 9871 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID)) 9872 return EINVAL; 9873 9874 if (txn->mt_flags & MDB_TXN_BLOCKED) 9875 return MDB_BAD_TXN; 9876 9877 if (txn->mt_dbflags[dbi] & DB_STALE) { 9878 MDB_cursor mc; 9879 MDB_xcursor mx; 9880 /* Stale, must read the DB's root. cursor_init does it for us. */ 9881 mdb_cursor_init(&mc, txn, dbi, &mx); 9882 } 9883 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg); 9884 } 9885 9886 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi) 9887 { 9888 char *ptr; 9889 if (dbi < CORE_DBS || dbi >= env->me_maxdbs) 9890 return; 9891 ptr = env->me_dbxs[dbi].md_name.mv_data; 9892 /* If there was no name, this was already closed */ 9893 if (ptr) { 9894 env->me_dbxs[dbi].md_name.mv_data = NULL; 9895 env->me_dbxs[dbi].md_name.mv_size = 0; 9896 env->me_dbflags[dbi] = 0; 9897 env->me_dbiseqs[dbi]++; 9898 free(ptr); 9899 } 9900 } 9901 9902 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags) 9903 { 9904 /* We could return the flags for the FREE_DBI too but what's the point? */ 9905 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 9906 return EINVAL; 9907 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS; 9908 return MDB_SUCCESS; 9909 } 9910 9911 /** Add all the DB's pages to the free list. 9912 * @param[in] mc Cursor on the DB to free. 9913 * @param[in] subs non-Zero to check for sub-DBs in this DB. 9914 * @return 0 on success, non-zero on failure. 9915 */ 9916 static int 9917 mdb_drop0(MDB_cursor *mc, int subs) 9918 { 9919 int rc; 9920 9921 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST); 9922 if (rc == MDB_SUCCESS) { 9923 MDB_txn *txn = mc->mc_txn; 9924 MDB_node *ni; 9925 MDB_cursor mx; 9926 unsigned int i; 9927 9928 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves. 9929 * This also avoids any P_LEAF2 pages, which have no nodes. 9930 * Also if the DB doesn't have sub-DBs and has no overflow 9931 * pages, omit scanning leaves. 9932 */ 9933 if ((mc->mc_flags & C_SUB) || 9934 (!subs && !mc->mc_db->md_overflow_pages)) 9935 mdb_cursor_pop(mc); 9936 9937 mdb_cursor_copy(mc, &mx); 9938 while (mc->mc_snum > 0) { 9939 MDB_page *mp = mc->mc_pg[mc->mc_top]; 9940 unsigned n = NUMKEYS(mp); 9941 if (IS_LEAF(mp)) { 9942 for (i=0; i<n; i++) { 9943 ni = NODEPTR(mp, i); 9944 if (ni->mn_flags & F_BIGDATA) { 9945 MDB_page *omp; 9946 pgno_t pg; 9947 memcpy(&pg, NODEDATA(ni), sizeof(pg)); 9948 rc = mdb_page_get(mc, pg, &omp, NULL); 9949 if (rc != 0) 9950 goto done; 9951 mdb_cassert(mc, IS_OVERFLOW(omp)); 9952 rc = mdb_midl_append_range(&txn->mt_free_pgs, 9953 pg, omp->mp_pages); 9954 if (rc) 9955 goto done; 9956 mc->mc_db->md_overflow_pages -= omp->mp_pages; 9957 if (!mc->mc_db->md_overflow_pages && !subs) 9958 break; 9959 } else if (subs && (ni->mn_flags & F_SUBDATA)) { 9960 mdb_xcursor_init1(mc, ni); 9961 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0); 9962 if (rc) 9963 goto done; 9964 } 9965 } 9966 if (!subs && !mc->mc_db->md_overflow_pages) 9967 goto pop; 9968 } else { 9969 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0) 9970 goto done; 9971 for (i=0; i<n; i++) { 9972 pgno_t pg; 9973 ni = NODEPTR(mp, i); 9974 pg = NODEPGNO(ni); 9975 /* free it */ 9976 mdb_midl_xappend(txn->mt_free_pgs, pg); 9977 } 9978 } 9979 if (!mc->mc_top) 9980 break; 9981 mc->mc_ki[mc->mc_top] = i; 9982 rc = mdb_cursor_sibling(mc, 1); 9983 if (rc) { 9984 if (rc != MDB_NOTFOUND) 9985 goto done; 9986 /* no more siblings, go back to beginning 9987 * of previous level. 9988 */ 9989 pop: 9990 mdb_cursor_pop(mc); 9991 mc->mc_ki[0] = 0; 9992 for (i=1; i<mc->mc_snum; i++) { 9993 mc->mc_ki[i] = 0; 9994 mc->mc_pg[i] = mx.mc_pg[i]; 9995 } 9996 } 9997 } 9998 /* free it */ 9999 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root); 10000 done: 10001 if (rc) 10002 txn->mt_flags |= MDB_TXN_ERROR; 10003 } else if (rc == MDB_NOTFOUND) { 10004 rc = MDB_SUCCESS; 10005 } 10006 mc->mc_flags &= ~C_INITIALIZED; 10007 return rc; 10008 } 10009 10010 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del) 10011 { 10012 MDB_cursor *mc, *m2; 10013 int rc; 10014 10015 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10016 return EINVAL; 10017 10018 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) 10019 return EACCES; 10020 10021 if (TXN_DBI_CHANGED(txn, dbi)) 10022 return MDB_BAD_DBI; 10023 10024 rc = mdb_cursor_open(txn, dbi, &mc); 10025 if (rc) 10026 return rc; 10027 10028 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT); 10029 /* Invalidate the dropped DB's cursors */ 10030 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next) 10031 m2->mc_flags &= ~(C_INITIALIZED|C_EOF); 10032 if (rc) 10033 goto leave; 10034 10035 /* Can't delete the main DB */ 10036 if (del && dbi >= CORE_DBS) { 10037 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA); 10038 if (!rc) { 10039 txn->mt_dbflags[dbi] = DB_STALE; 10040 mdb_dbi_close(txn->mt_env, dbi); 10041 } else { 10042 txn->mt_flags |= MDB_TXN_ERROR; 10043 } 10044 } else { 10045 /* reset the DB record, mark it dirty */ 10046 txn->mt_dbflags[dbi] |= DB_DIRTY; 10047 txn->mt_dbs[dbi].md_depth = 0; 10048 txn->mt_dbs[dbi].md_branch_pages = 0; 10049 txn->mt_dbs[dbi].md_leaf_pages = 0; 10050 txn->mt_dbs[dbi].md_overflow_pages = 0; 10051 txn->mt_dbs[dbi].md_entries = 0; 10052 txn->mt_dbs[dbi].md_root = P_INVALID; 10053 10054 txn->mt_flags |= MDB_TXN_DIRTY; 10055 } 10056 leave: 10057 mdb_cursor_close(mc); 10058 return rc; 10059 } 10060 10061 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp) 10062 { 10063 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10064 return EINVAL; 10065 10066 txn->mt_dbxs[dbi].md_cmp = cmp; 10067 return MDB_SUCCESS; 10068 } 10069 10070 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp) 10071 { 10072 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10073 return EINVAL; 10074 10075 txn->mt_dbxs[dbi].md_dcmp = cmp; 10076 return MDB_SUCCESS; 10077 } 10078 10079 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel) 10080 { 10081 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10082 return EINVAL; 10083 10084 txn->mt_dbxs[dbi].md_rel = rel; 10085 return MDB_SUCCESS; 10086 } 10087 10088 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx) 10089 { 10090 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10091 return EINVAL; 10092 10093 txn->mt_dbxs[dbi].md_relctx = ctx; 10094 return MDB_SUCCESS; 10095 } 10096 10097 int ESECT 10098 mdb_env_get_maxkeysize(MDB_env *env) 10099 { 10100 return ENV_MAXKEY(env); 10101 } 10102 10103 int ESECT 10104 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx) 10105 { 10106 unsigned int i, rdrs; 10107 MDB_reader *mr; 10108 char buf[64]; 10109 int rc = 0, first = 1; 10110 10111 if (!env || !func) 10112 return -1; 10113 if (!env->me_txns) { 10114 return func("(no reader locks)\n", ctx); 10115 } 10116 rdrs = env->me_txns->mti_numreaders; 10117 mr = env->me_txns->mti_readers; 10118 for (i=0; i<rdrs; i++) { 10119 if (mr[i].mr_pid) { 10120 txnid_t txnid = mr[i].mr_txnid; 10121 sprintf(buf, txnid == (txnid_t)-1 ? 10122 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n", 10123 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid); 10124 if (first) { 10125 first = 0; 10126 rc = func(" pid thread txnid\n", ctx); 10127 if (rc < 0) 10128 break; 10129 } 10130 rc = func(buf, ctx); 10131 if (rc < 0) 10132 break; 10133 } 10134 } 10135 if (first) { 10136 rc = func("(no active readers)\n", ctx); 10137 } 10138 return rc; 10139 } 10140 10141 /** Insert pid into list if not already present. 10142 * return -1 if already present. 10143 */ 10144 static int ESECT 10145 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid) 10146 { 10147 /* binary search of pid in list */ 10148 unsigned base = 0; 10149 unsigned cursor = 1; 10150 int val = 0; 10151 unsigned n = ids[0]; 10152 10153 while( 0 < n ) { 10154 unsigned pivot = n >> 1; 10155 cursor = base + pivot + 1; 10156 val = pid - ids[cursor]; 10157 10158 if( val < 0 ) { 10159 n = pivot; 10160 10161 } else if ( val > 0 ) { 10162 base = cursor; 10163 n -= pivot + 1; 10164 10165 } else { 10166 /* found, so it's a duplicate */ 10167 return -1; 10168 } 10169 } 10170 10171 if( val > 0 ) { 10172 ++cursor; 10173 } 10174 ids[0]++; 10175 for (n = ids[0]; n > cursor; n--) 10176 ids[n] = ids[n-1]; 10177 ids[n] = pid; 10178 return 0; 10179 } 10180 10181 int ESECT 10182 mdb_reader_check(MDB_env *env, int *dead) 10183 { 10184 if (!env) 10185 return EINVAL; 10186 if (dead) 10187 *dead = 0; 10188 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS; 10189 } 10190 10191 /** As #mdb_reader_check(). \b rlocked is set if caller locked #me_rmutex. */ 10192 static int ESECT 10193 mdb_reader_check0(MDB_env *env, int rlocked, int *dead) 10194 { 10195 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex; 10196 unsigned int i, j, rdrs; 10197 MDB_reader *mr; 10198 MDB_PID_T *pids, pid; 10199 int rc = MDB_SUCCESS, count = 0; 10200 10201 rdrs = env->me_txns->mti_numreaders; 10202 pids = malloc((rdrs+1) * sizeof(MDB_PID_T)); 10203 if (!pids) 10204 return ENOMEM; 10205 pids[0] = 0; 10206 mr = env->me_txns->mti_readers; 10207 for (i=0; i<rdrs; i++) { 10208 pid = mr[i].mr_pid; 10209 if (pid && pid != env->me_pid) { 10210 if (mdb_pid_insert(pids, pid) == 0) { 10211 if (!mdb_reader_pid(env, Pidcheck, pid)) { 10212 /* Stale reader found */ 10213 j = i; 10214 if (rmutex) { 10215 if ((rc = LOCK_MUTEX0(rmutex)) != 0) { 10216 if ((rc = mdb_mutex_failed(env, rmutex, rc))) 10217 break; 10218 rdrs = 0; /* the above checked all readers */ 10219 } else { 10220 /* Recheck, a new process may have reused pid */ 10221 if (mdb_reader_pid(env, Pidcheck, pid)) 10222 j = rdrs; 10223 } 10224 } 10225 for (; j<rdrs; j++) 10226 if (mr[j].mr_pid == pid) { 10227 DPRINTF(("clear stale reader pid %u txn %"Z"d", 10228 (unsigned) pid, mr[j].mr_txnid)); 10229 mr[j].mr_pid = 0; 10230 count++; 10231 } 10232 if (rmutex) 10233 UNLOCK_MUTEX(rmutex); 10234 } 10235 } 10236 } 10237 } 10238 free(pids); 10239 if (dead) 10240 *dead = count; 10241 return rc; 10242 } 10243 10244 #ifdef MDB_ROBUST_SUPPORTED 10245 /** Handle #LOCK_MUTEX0() failure. 10246 * Try to repair the lock file if the mutex owner died. 10247 * @param[in] env the environment handle 10248 * @param[in] mutex LOCK_MUTEX0() mutex 10249 * @param[in] rc LOCK_MUTEX0() error (nonzero) 10250 * @return 0 on success with the mutex locked, or an error code on failure. 10251 */ 10252 static int ESECT 10253 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc) 10254 { 10255 int rlocked, rc2; 10256 MDB_meta *meta; 10257 10258 if (rc == MDB_OWNERDEAD) { 10259 /* We own the mutex. Clean up after dead previous owner. */ 10260 rc = MDB_SUCCESS; 10261 rlocked = (mutex == env->me_rmutex); 10262 if (!rlocked) { 10263 /* Keep mti_txnid updated, otherwise next writer can 10264 * overwrite data which latest meta page refers to. 10265 */ 10266 meta = mdb_env_pick_meta(env); 10267 env->me_txns->mti_txnid = meta->mm_txnid; 10268 /* env is hosed if the dead thread was ours */ 10269 if (env->me_txn) { 10270 env->me_flags |= MDB_FATAL_ERROR; 10271 env->me_txn = NULL; 10272 rc = MDB_PANIC; 10273 } 10274 } 10275 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'), 10276 (rc ? "this process' env is hosed" : "recovering"))); 10277 rc2 = mdb_reader_check0(env, rlocked, NULL); 10278 if (rc2 == 0) 10279 rc2 = mdb_mutex_consistent(mutex); 10280 if (rc || (rc = rc2)) { 10281 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc))); 10282 UNLOCK_MUTEX(mutex); 10283 } 10284 } else { 10285 #ifdef _WIN32 10286 rc = ErrCode(); 10287 #endif 10288 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc))); 10289 } 10290 10291 return rc; 10292 } 10293 #endif /* MDB_ROBUST_SUPPORTED */ 10294 10295 #if defined(_WIN32) 10296 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */ 10297 static int ESECT 10298 utf8_to_utf16(const char *src, MDB_name *dst, int xtra) 10299 { 10300 int rc, need = 0; 10301 wchar_t *result = NULL; 10302 for (;;) { /* malloc result, then fill it in */ 10303 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need); 10304 if (!need) { 10305 rc = ErrCode(); 10306 free(result); 10307 return rc; 10308 } 10309 if (!result) { 10310 result = malloc(sizeof(wchar_t) * (need + xtra)); 10311 if (!result) 10312 return ENOMEM; 10313 continue; 10314 } 10315 dst->mn_alloced = 1; 10316 dst->mn_len = need - 1; 10317 dst->mn_val = result; 10318 return MDB_SUCCESS; 10319 } 10320 } 10321 #endif /* defined(_WIN32) */ 10322 /** @} */ 10323