1 /* $NetBSD: mdb.c,v 1.2 2020/08/11 13:15:38 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 2011-2020 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(__APPLE__) || defined (BSD) || defined(__FreeBSD_kernel__) 130 # define MDB_USE_POSIX_SEM 1 131 # define MDB_FDATASYNC fsync 132 #elif defined(ANDROID) 133 # define MDB_FDATASYNC fsync 134 #endif 135 136 #ifndef _WIN32 137 #include <pthread.h> 138 #include <signal.h> 139 #ifdef MDB_USE_POSIX_SEM 140 # define MDB_USE_HASH 1 141 #include <semaphore.h> 142 #else 143 #define MDB_USE_POSIX_MUTEX 1 144 #endif 145 #endif 146 147 #if defined(_WIN32) + defined(MDB_USE_POSIX_SEM) \ 148 + defined(MDB_USE_POSIX_MUTEX) != 1 149 # error "Ambiguous shared-lock implementation" 150 #endif 151 152 #ifdef USE_VALGRIND 153 #include <valgrind/memcheck.h> 154 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z) 155 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s) 156 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a) 157 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h) 158 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s) 159 #else 160 #define VGMEMP_CREATE(h,r,z) 161 #define VGMEMP_ALLOC(h,a,s) 162 #define VGMEMP_FREE(h,a) 163 #define VGMEMP_DESTROY(h) 164 #define VGMEMP_DEFINED(a,s) 165 #endif 166 167 #ifndef BYTE_ORDER 168 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN)) 169 /* Solaris just defines one or the other */ 170 # define LITTLE_ENDIAN 1234 171 # define BIG_ENDIAN 4321 172 # ifdef _LITTLE_ENDIAN 173 # define BYTE_ORDER LITTLE_ENDIAN 174 # else 175 # define BYTE_ORDER BIG_ENDIAN 176 # endif 177 # else 178 # define BYTE_ORDER __BYTE_ORDER 179 # endif 180 #endif 181 182 #ifndef LITTLE_ENDIAN 183 #define LITTLE_ENDIAN __LITTLE_ENDIAN 184 #endif 185 #ifndef BIG_ENDIAN 186 #define BIG_ENDIAN __BIG_ENDIAN 187 #endif 188 189 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86) 190 #define MISALIGNED_OK 1 191 #endif 192 193 #include "lmdb.h" 194 #include "midl.h" 195 196 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN) 197 # error "Unknown or unsupported endianness (BYTE_ORDER)" 198 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF 199 # error "Two's complement, reasonably sized integer types, please" 200 #endif 201 202 #ifdef __GNUC__ 203 /** Put infrequently used env functions in separate section */ 204 # ifdef __APPLE__ 205 # define ESECT __attribute__ ((section("__TEXT,text_env"))) 206 # else 207 # define ESECT __attribute__ ((section("text_env"))) 208 # endif 209 #else 210 #define ESECT 211 #endif 212 213 #ifdef _WIN32 214 #define CALL_CONV WINAPI 215 #else 216 #define CALL_CONV 217 #endif 218 219 /** @defgroup internal LMDB Internals 220 * @{ 221 */ 222 /** @defgroup compat Compatibility Macros 223 * A bunch of macros to minimize the amount of platform-specific ifdefs 224 * needed throughout the rest of the code. When the features this library 225 * needs are similar enough to POSIX to be hidden in a one-or-two line 226 * replacement, this macro approach is used. 227 * @{ 228 */ 229 230 /** Features under development */ 231 #ifndef MDB_DEVEL 232 #define MDB_DEVEL 0 233 #endif 234 235 /** Wrapper around __func__, which is a C99 feature */ 236 #if __STDC_VERSION__ >= 199901L 237 # define mdb_func_ __func__ 238 #elif __GNUC__ >= 2 || _MSC_VER >= 1300 239 # define mdb_func_ __FUNCTION__ 240 #else 241 /* If a debug message says <mdb_unknown>(), update the #if statements above */ 242 # define mdb_func_ "<mdb_unknown>" 243 #endif 244 245 /* Internal error codes, not exposed outside liblmdb */ 246 #define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10) 247 #ifdef _WIN32 248 #define MDB_OWNERDEAD ((int) WAIT_ABANDONED) 249 #elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD) 250 #define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */ 251 #endif 252 253 #ifdef __GLIBC__ 254 #define GLIBC_VER ((__GLIBC__ << 16 )| __GLIBC_MINOR__) 255 #endif 256 /** Some platforms define the EOWNERDEAD error code 257 * even though they don't support Robust Mutexes. 258 * Compile with -DMDB_USE_ROBUST=0, or use some other 259 * mechanism like -DMDB_USE_POSIX_SEM instead of 260 * -DMDB_USE_POSIX_MUTEX. 261 * (Posix semaphores are not robust.) 262 */ 263 #ifndef MDB_USE_ROBUST 264 /* Android currently lacks Robust Mutex support. So does glibc < 2.4. */ 265 # if defined(MDB_USE_POSIX_MUTEX) && (defined(ANDROID) || \ 266 (defined(__GLIBC__) && GLIBC_VER < 0x020004)) 267 # define MDB_USE_ROBUST 0 268 # else 269 # define MDB_USE_ROBUST 1 270 # endif 271 #endif /* !MDB_USE_ROBUST */ 272 273 #if defined(MDB_USE_POSIX_MUTEX) && (MDB_USE_ROBUST) 274 /* glibc < 2.12 only provided _np API */ 275 # if (defined(__GLIBC__) && GLIBC_VER < 0x02000c) || \ 276 (defined(PTHREAD_MUTEX_ROBUST_NP) && !defined(PTHREAD_MUTEX_ROBUST)) 277 # define PTHREAD_MUTEX_ROBUST PTHREAD_MUTEX_ROBUST_NP 278 # define pthread_mutexattr_setrobust(attr, flag) pthread_mutexattr_setrobust_np(attr, flag) 279 # define pthread_mutex_consistent(mutex) pthread_mutex_consistent_np(mutex) 280 # endif 281 #endif /* MDB_USE_POSIX_MUTEX && MDB_USE_ROBUST */ 282 283 #if defined(MDB_OWNERDEAD) && (MDB_USE_ROBUST) 284 #define MDB_ROBUST_SUPPORTED 1 285 #endif 286 287 #ifdef _WIN32 288 #define MDB_USE_HASH 1 289 #define MDB_PIDLOCK 0 290 #define THREAD_RET DWORD 291 #define pthread_t HANDLE 292 #define pthread_mutex_t HANDLE 293 #define pthread_cond_t HANDLE 294 typedef HANDLE mdb_mutex_t, mdb_mutexref_t; 295 #define pthread_key_t DWORD 296 #define pthread_self() GetCurrentThreadId() 297 #define pthread_key_create(x,y) \ 298 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0) 299 #define pthread_key_delete(x) TlsFree(x) 300 #define pthread_getspecific(x) TlsGetValue(x) 301 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode()) 302 #define pthread_mutex_unlock(x) ReleaseMutex(*x) 303 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE) 304 #define pthread_cond_signal(x) SetEvent(*x) 305 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0) 306 #define THREAD_CREATE(thr,start,arg) \ 307 (((thr) = CreateThread(NULL, 0, start, arg, 0, NULL)) ? 0 : ErrCode()) 308 #define THREAD_FINISH(thr) \ 309 (WaitForSingleObject(thr, INFINITE) ? ErrCode() : 0) 310 #define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE) 311 #define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex) 312 #define mdb_mutex_consistent(mutex) 0 313 #define getpid() GetCurrentProcessId() 314 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd)) 315 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len)) 316 #define ErrCode() GetLastError() 317 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;} 318 #define close(fd) (CloseHandle(fd) ? 0 : -1) 319 #define munmap(ptr,len) UnmapViewOfFile(ptr) 320 #ifdef PROCESS_QUERY_LIMITED_INFORMATION 321 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION 322 #else 323 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000 324 #endif 325 #define Z "I" 326 #else 327 #define THREAD_RET void * 328 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg) 329 #define THREAD_FINISH(thr) pthread_join(thr,NULL) 330 #define Z "z" /**< printf format modifier for size_t */ 331 332 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */ 333 #define MDB_PIDLOCK 1 334 335 #ifdef MDB_USE_POSIX_SEM 336 337 typedef sem_t *mdb_mutex_t, *mdb_mutexref_t; 338 #define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex) 339 #define UNLOCK_MUTEX(mutex) sem_post(mutex) 340 341 static int 342 mdb_sem_wait(sem_t *sem) 343 { 344 int rc; 345 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ; 346 return rc; 347 } 348 349 #else /* MDB_USE_POSIX_MUTEX: */ 350 /** Shared mutex/semaphore as the original is stored. 351 * 352 * Not for copies. Instead it can be assigned to an #mdb_mutexref_t. 353 * When mdb_mutexref_t is a pointer and mdb_mutex_t is not, then it 354 * is array[size 1] so it can be assigned to the pointer. 355 */ 356 typedef pthread_mutex_t mdb_mutex_t[1]; 357 /** Reference to an #mdb_mutex_t */ 358 typedef pthread_mutex_t *mdb_mutexref_t; 359 /** Lock the reader or writer mutex. 360 * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX(). 361 */ 362 #define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex) 363 /** Unlock the reader or writer mutex. 364 */ 365 #define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex) 366 /** Mark mutex-protected data as repaired, after death of previous owner. 367 */ 368 #define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex) 369 #endif /* MDB_USE_POSIX_SEM */ 370 371 /** Get the error code for the last failed system function. 372 */ 373 #define ErrCode() errno 374 375 /** An abstraction for a file handle. 376 * On POSIX systems file handles are small integers. On Windows 377 * they're opaque pointers. 378 */ 379 #define HANDLE int 380 381 /** A value for an invalid file handle. 382 * Mainly used to initialize file variables and signify that they are 383 * unused. 384 */ 385 #define INVALID_HANDLE_VALUE (-1) 386 387 /** Get the size of a memory page for the system. 388 * This is the basic size that the platform's memory manager uses, and is 389 * fundamental to the use of memory-mapped files. 390 */ 391 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE)) 392 #endif 393 394 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM) 395 #define MNAME_LEN 32 396 #else 397 #define MNAME_LEN (sizeof(pthread_mutex_t)) 398 #endif 399 400 /** @} */ 401 402 #ifdef MDB_ROBUST_SUPPORTED 403 /** Lock mutex, handle any error, set rc = result. 404 * Return 0 on success, nonzero (not rc) on error. 405 */ 406 #define LOCK_MUTEX(rc, env, mutex) \ 407 (((rc) = LOCK_MUTEX0(mutex)) && \ 408 ((rc) = mdb_mutex_failed(env, mutex, rc))) 409 static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc); 410 #else 411 #define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex)) 412 #define mdb_mutex_failed(env, mutex, rc) (rc) 413 #endif 414 415 #ifndef _WIN32 416 /** A flag for opening a file and requesting synchronous data writes. 417 * This is only used when writing a meta page. It's not strictly needed; 418 * we could just do a normal write and then immediately perform a flush. 419 * But if this flag is available it saves us an extra system call. 420 * 421 * @note If O_DSYNC is undefined but exists in /usr/include, 422 * preferably set some compiler flag to get the definition. 423 */ 424 #ifndef MDB_DSYNC 425 # ifdef O_DSYNC 426 # define MDB_DSYNC O_DSYNC 427 # else 428 # define MDB_DSYNC O_SYNC 429 # endif 430 #endif 431 #endif 432 433 /** Function for flushing the data of a file. Define this to fsync 434 * if fdatasync() is not supported. 435 */ 436 #ifndef MDB_FDATASYNC 437 # define MDB_FDATASYNC fdatasync 438 #endif 439 440 #ifndef MDB_MSYNC 441 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags) 442 #endif 443 444 #ifndef MS_SYNC 445 #define MS_SYNC 1 446 #endif 447 448 #ifndef MS_ASYNC 449 #define MS_ASYNC 0 450 #endif 451 452 /** A page number in the database. 453 * Note that 64 bit page numbers are overkill, since pages themselves 454 * already represent 12-13 bits of addressable memory, and the OS will 455 * always limit applications to a maximum of 63 bits of address space. 456 * 457 * @note In the #MDB_node structure, we only store 48 bits of this value, 458 * which thus limits us to only 60 bits of addressable data. 459 */ 460 typedef MDB_ID pgno_t; 461 462 /** A transaction ID. 463 * See struct MDB_txn.mt_txnid for details. 464 */ 465 typedef MDB_ID txnid_t; 466 467 /** @defgroup debug Debug Macros 468 * @{ 469 */ 470 #ifndef MDB_DEBUG 471 /** Enable debug output. Needs variable argument macros (a C99 feature). 472 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs 473 * read from and written to the database (used for free space management). 474 */ 475 #define MDB_DEBUG 0 476 #endif 477 478 #if MDB_DEBUG 479 static int mdb_debug; 480 static txnid_t mdb_debug_start; 481 482 /** Print a debug message with printf formatting. 483 * Requires double parenthesis around 2 or more args. 484 */ 485 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args)) 486 # define DPRINTF0(fmt, ...) \ 487 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__) 488 #else 489 # define DPRINTF(args) ((void) 0) 490 #endif 491 /** Print a debug string. 492 * The string is printed literally, with no format processing. 493 */ 494 #define DPUTS(arg) DPRINTF(("%s", arg)) 495 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */ 496 #define DDBI(mc) \ 497 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi) 498 /** @} */ 499 500 /** @brief The maximum size of a database page. 501 * 502 * It is 32k or 64k, since value-PAGEBASE must fit in 503 * #MDB_page.%mp_upper. 504 * 505 * LMDB will use database pages < OS pages if needed. 506 * That causes more I/O in write transactions: The OS must 507 * know (read) the whole page before writing a partial page. 508 * 509 * Note that we don't currently support Huge pages. On Linux, 510 * regular data files cannot use Huge pages, and in general 511 * Huge pages aren't actually pageable. We rely on the OS 512 * demand-pager to read our data and page it out when memory 513 * pressure from other processes is high. So until OSs have 514 * actual paging support for Huge pages, they're not viable. 515 */ 516 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000) 517 518 /** The minimum number of keys required in a database page. 519 * Setting this to a larger value will place a smaller bound on the 520 * maximum size of a data item. Data items larger than this size will 521 * be pushed into overflow pages instead of being stored directly in 522 * the B-tree node. This value used to default to 4. With a page size 523 * of 4096 bytes that meant that any item larger than 1024 bytes would 524 * go into an overflow page. That also meant that on average 2-3KB of 525 * each overflow page was wasted space. The value cannot be lower than 526 * 2 because then there would no longer be a tree structure. With this 527 * value, items larger than 2KB will go into overflow pages, and on 528 * average only 1KB will be wasted. 529 */ 530 #define MDB_MINKEYS 2 531 532 /** A stamp that identifies a file as an LMDB file. 533 * There's nothing special about this value other than that it is easily 534 * recognizable, and it will reflect any byte order mismatches. 535 */ 536 #define MDB_MAGIC 0xBEEFC0DE 537 538 /** The version number for a database's datafile format. */ 539 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1) 540 /** The version number for a database's lockfile format. */ 541 #define MDB_LOCK_VERSION 1 542 543 /** @brief The max size of a key we can write, or 0 for computed max. 544 * 545 * This macro should normally be left alone or set to 0. 546 * Note that a database with big keys or dupsort data cannot be 547 * reliably modified by a liblmdb which uses a smaller max. 548 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL. 549 * 550 * Other values are allowed, for backwards compat. However: 551 * A value bigger than the computed max can break if you do not 552 * know what you are doing, and liblmdb <= 0.9.10 can break when 553 * modifying a DB with keys/dupsort data bigger than its max. 554 * 555 * Data items in an #MDB_DUPSORT database are also limited to 556 * this size, since they're actually keys of a sub-DB. Keys and 557 * #MDB_DUPSORT data items must fit on a node in a regular page. 558 */ 559 #ifndef MDB_MAXKEYSIZE 560 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511) 561 #endif 562 563 /** The maximum size of a key we can write to the environment. */ 564 #if MDB_MAXKEYSIZE 565 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE) 566 #else 567 #define ENV_MAXKEY(env) ((env)->me_maxkey) 568 #endif 569 570 /** @brief The maximum size of a data item. 571 * 572 * We only store a 32 bit value for node sizes. 573 */ 574 #define MAXDATASIZE 0xffffffffUL 575 576 #if MDB_DEBUG 577 /** Key size which fits in a #DKBUF. 578 * @ingroup debug 579 */ 580 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511) 581 /** A key buffer. 582 * @ingroup debug 583 * This is used for printing a hex dump of a key's contents. 584 */ 585 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1] 586 /** Display a key in hex. 587 * @ingroup debug 588 * Invoke a function to display a key in hex. 589 */ 590 #define DKEY(x) mdb_dkey(x, kbuf) 591 #else 592 #define DKBUF 593 #define DKEY(x) 0 594 #endif 595 596 /** An invalid page number. 597 * Mainly used to denote an empty tree. 598 */ 599 #define P_INVALID (~(pgno_t)0) 600 601 /** Test if the flags \b f are set in a flag word \b w. */ 602 #define F_ISSET(w, f) (((w) & (f)) == (f)) 603 604 /** Round \b n up to an even number. */ 605 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */ 606 607 /** Used for offsets within a single page. 608 * Since memory pages are typically 4 or 8KB in size, 12-13 bits, 609 * this is plenty. 610 */ 611 typedef uint16_t indx_t; 612 613 /** Default size of memory map. 614 * This is certainly too small for any actual applications. Apps should always set 615 * the size explicitly using #mdb_env_set_mapsize(). 616 */ 617 #define DEFAULT_MAPSIZE 1048576 618 619 /** @defgroup readers Reader Lock Table 620 * Readers don't acquire any locks for their data access. Instead, they 621 * simply record their transaction ID in the reader table. The reader 622 * mutex is needed just to find an empty slot in the reader table. The 623 * slot's address is saved in thread-specific data so that subsequent read 624 * transactions started by the same thread need no further locking to proceed. 625 * 626 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data. 627 * 628 * No reader table is used if the database is on a read-only filesystem, or 629 * if #MDB_NOLOCK is set. 630 * 631 * Since the database uses multi-version concurrency control, readers don't 632 * actually need any locking. This table is used to keep track of which 633 * readers are using data from which old transactions, so that we'll know 634 * when a particular old transaction is no longer in use. Old transactions 635 * that have discarded any data pages can then have those pages reclaimed 636 * for use by a later write transaction. 637 * 638 * The lock table is constructed such that reader slots are aligned with the 639 * processor's cache line size. Any slot is only ever used by one thread. 640 * This alignment guarantees that there will be no contention or cache 641 * thrashing as threads update their own slot info, and also eliminates 642 * any need for locking when accessing a slot. 643 * 644 * A writer thread will scan every slot in the table to determine the oldest 645 * outstanding reader transaction. Any freed pages older than this will be 646 * reclaimed by the writer. The writer doesn't use any locks when scanning 647 * this table. This means that there's no guarantee that the writer will 648 * see the most up-to-date reader info, but that's not required for correct 649 * operation - all we need is to know the upper bound on the oldest reader, 650 * we don't care at all about the newest reader. So the only consequence of 651 * reading stale information here is that old pages might hang around a 652 * while longer before being reclaimed. That's actually good anyway, because 653 * the longer we delay reclaiming old pages, the more likely it is that a 654 * string of contiguous pages can be found after coalescing old pages from 655 * many old transactions together. 656 * @{ 657 */ 658 /** Number of slots in the reader table. 659 * This value was chosen somewhat arbitrarily. 126 readers plus a 660 * couple mutexes fit exactly into 8KB on my development machine. 661 * Applications should set the table size using #mdb_env_set_maxreaders(). 662 */ 663 #define DEFAULT_READERS 126 664 665 /** The size of a CPU cache line in bytes. We want our lock structures 666 * aligned to this size to avoid false cache line sharing in the 667 * lock table. 668 * This value works for most CPUs. For Itanium this should be 128. 669 */ 670 #ifndef CACHELINE 671 #define CACHELINE 64 672 #endif 673 674 /** The information we store in a single slot of the reader table. 675 * In addition to a transaction ID, we also record the process and 676 * thread ID that owns a slot, so that we can detect stale information, 677 * e.g. threads or processes that went away without cleaning up. 678 * @note We currently don't check for stale records. We simply re-init 679 * the table when we know that we're the only process opening the 680 * lock file. 681 */ 682 typedef struct MDB_rxbody { 683 /** Current Transaction ID when this transaction began, or (txnid_t)-1. 684 * Multiple readers that start at the same time will probably have the 685 * same ID here. Again, it's not important to exclude them from 686 * anything; all we need to know is which version of the DB they 687 * started from so we can avoid overwriting any data used in that 688 * particular version. 689 */ 690 volatile txnid_t mrb_txnid; 691 /** The process ID of the process owning this reader txn. */ 692 volatile MDB_PID_T mrb_pid; 693 /** The thread ID of the thread owning this txn. */ 694 volatile MDB_THR_T mrb_tid; 695 } MDB_rxbody; 696 697 /** The actual reader record, with cacheline padding. */ 698 typedef struct MDB_reader { 699 union { 700 MDB_rxbody mrx; 701 /** shorthand for mrb_txnid */ 702 #define mr_txnid mru.mrx.mrb_txnid 703 #define mr_pid mru.mrx.mrb_pid 704 #define mr_tid mru.mrx.mrb_tid 705 /** cache line alignment */ 706 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)]; 707 } mru; 708 } MDB_reader; 709 710 /** The header for the reader table. 711 * The table resides in a memory-mapped file. (This is a different file 712 * than is used for the main database.) 713 * 714 * For POSIX the actual mutexes reside in the shared memory of this 715 * mapped file. On Windows, mutexes are named objects allocated by the 716 * kernel; we store the mutex names in this mapped file so that other 717 * processes can grab them. This same approach is also used on 718 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support 719 * process-shared POSIX mutexes. For these cases where a named object 720 * is used, the object name is derived from a 64 bit FNV hash of the 721 * environment pathname. As such, naming collisions are extremely 722 * unlikely. If a collision occurs, the results are unpredictable. 723 */ 724 typedef struct MDB_txbody { 725 /** Stamp identifying this as an LMDB file. It must be set 726 * to #MDB_MAGIC. */ 727 uint32_t mtb_magic; 728 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */ 729 uint32_t mtb_format; 730 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM) 731 char mtb_rmname[MNAME_LEN]; 732 #else 733 /** Mutex protecting access to this table. 734 * This is the reader table lock used with LOCK_MUTEX(). 735 */ 736 mdb_mutex_t mtb_rmutex; 737 #endif 738 /** The ID of the last transaction committed to the database. 739 * This is recorded here only for convenience; the value can always 740 * be determined by reading the main database meta pages. 741 */ 742 volatile txnid_t mtb_txnid; 743 /** The number of slots that have been used in the reader table. 744 * This always records the maximum count, it is not decremented 745 * when readers release their slots. 746 */ 747 volatile unsigned mtb_numreaders; 748 } MDB_txbody; 749 750 /** The actual reader table definition. */ 751 typedef struct MDB_txninfo { 752 union { 753 MDB_txbody mtb; 754 #define mti_magic mt1.mtb.mtb_magic 755 #define mti_format mt1.mtb.mtb_format 756 #define mti_rmutex mt1.mtb.mtb_rmutex 757 #define mti_rmname mt1.mtb.mtb_rmname 758 #define mti_txnid mt1.mtb.mtb_txnid 759 #define mti_numreaders mt1.mtb.mtb_numreaders 760 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)]; 761 } mt1; 762 union { 763 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM) 764 char mt2_wmname[MNAME_LEN]; 765 #define mti_wmname mt2.mt2_wmname 766 #else 767 mdb_mutex_t mt2_wmutex; 768 #define mti_wmutex mt2.mt2_wmutex 769 #endif 770 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)]; 771 } mt2; 772 MDB_reader mti_readers[1]; 773 } MDB_txninfo; 774 775 /** Lockfile format signature: version, features and field layout */ 776 #define MDB_LOCK_FORMAT \ 777 ((uint32_t) \ 778 ((MDB_LOCK_VERSION) \ 779 /* Flags which describe functionality */ \ 780 + (((MDB_PIDLOCK) != 0) << 16))) 781 /** @} */ 782 783 /** Common header for all page types. The page type depends on #mp_flags. 784 * 785 * #P_BRANCH and #P_LEAF pages have unsorted '#MDB_node's at the end, with 786 * sorted #mp_ptrs[] entries referring to them. Exception: #P_LEAF2 pages 787 * omit mp_ptrs and pack sorted #MDB_DUPFIXED values after the page header. 788 * 789 * #P_OVERFLOW records occupy one or more contiguous pages where only the 790 * first has a page header. They hold the real data of #F_BIGDATA nodes. 791 * 792 * #P_SUBP sub-pages are small leaf "pages" with duplicate data. 793 * A node with flag #F_DUPDATA but not #F_SUBDATA contains a sub-page. 794 * (Duplicate data can also go in sub-databases, which use normal pages.) 795 * 796 * #P_META pages contain #MDB_meta, the start point of an LMDB snapshot. 797 * 798 * Each non-metapage up to #MDB_meta.%mm_last_pg is reachable exactly once 799 * in the snapshot: Either used by a database or listed in a freeDB record. 800 */ 801 typedef struct MDB_page { 802 #define mp_pgno mp_p.p_pgno 803 #define mp_next mp_p.p_next 804 union { 805 pgno_t p_pgno; /**< page number */ 806 struct MDB_page *p_next; /**< for in-memory list of freed pages */ 807 } mp_p; 808 uint16_t mp_pad; /**< key size if this is a LEAF2 page */ 809 /** @defgroup mdb_page Page Flags 810 * @ingroup internal 811 * Flags for the page headers. 812 * @{ 813 */ 814 #define P_BRANCH 0x01 /**< branch page */ 815 #define P_LEAF 0x02 /**< leaf page */ 816 #define P_OVERFLOW 0x04 /**< overflow page */ 817 #define P_META 0x08 /**< meta page */ 818 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */ 819 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */ 820 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */ 821 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */ 822 #define P_KEEP 0x8000 /**< leave this page alone during spill */ 823 /** @} */ 824 uint16_t mp_flags; /**< @ref mdb_page */ 825 #define mp_lower mp_pb.pb.pb_lower 826 #define mp_upper mp_pb.pb.pb_upper 827 #define mp_pages mp_pb.pb_pages 828 union { 829 struct { 830 indx_t pb_lower; /**< lower bound of free space */ 831 indx_t pb_upper; /**< upper bound of free space */ 832 } pb; 833 uint32_t pb_pages; /**< number of overflow pages */ 834 } mp_pb; 835 indx_t mp_ptrs[1]; /**< dynamic size */ 836 } MDB_page; 837 838 /** Size of the page header, excluding dynamic data at the end */ 839 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs)) 840 841 /** Address of first usable data byte in a page, after the header */ 842 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ)) 843 844 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */ 845 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0) 846 847 /** Number of nodes on a page */ 848 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1) 849 850 /** The amount of space remaining in the page */ 851 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower) 852 853 /** The percentage of space used in the page, in tenths of a percent. */ 854 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \ 855 ((env)->me_psize - PAGEHDRSZ)) 856 /** The minimum page fill factor, in tenths of a percent. 857 * Pages emptier than this are candidates for merging. 858 */ 859 #define FILL_THRESHOLD 250 860 861 /** Test if a page is a leaf page */ 862 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF) 863 /** Test if a page is a LEAF2 page */ 864 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2) 865 /** Test if a page is a branch page */ 866 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH) 867 /** Test if a page is an overflow page */ 868 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW) 869 /** Test if a page is a sub page */ 870 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP) 871 872 /** The number of overflow pages needed to store the given size. */ 873 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1) 874 875 /** Link in #MDB_txn.%mt_loose_pgs list. 876 * Kept outside the page header, which is needed when reusing the page. 877 */ 878 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2)) 879 880 /** Header for a single key/data pair within a page. 881 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2. 882 * We guarantee 2-byte alignment for 'MDB_node's. 883 * 884 * #mn_lo and #mn_hi are used for data size on leaf nodes, and for child 885 * pgno on branch nodes. On 64 bit platforms, #mn_flags is also used 886 * for pgno. (Branch nodes have no flags). Lo and hi are in host byte 887 * order in case some accesses can be optimized to 32-bit word access. 888 * 889 * Leaf node flags describe node contents. #F_BIGDATA says the node's 890 * data part is the page number of an overflow page with actual data. 891 * #F_DUPDATA and #F_SUBDATA can be combined giving duplicate data in 892 * a sub-page/sub-database, and named databases (just #F_SUBDATA). 893 */ 894 typedef struct MDB_node { 895 /** part of data size or pgno 896 * @{ */ 897 #if BYTE_ORDER == LITTLE_ENDIAN 898 unsigned short mn_lo, mn_hi; 899 #else 900 unsigned short mn_hi, mn_lo; 901 #endif 902 /** @} */ 903 /** @defgroup mdb_node Node Flags 904 * @ingroup internal 905 * Flags for node headers. 906 * @{ 907 */ 908 #define F_BIGDATA 0x01 /**< data put on overflow page */ 909 #define F_SUBDATA 0x02 /**< data is a sub-database */ 910 #define F_DUPDATA 0x04 /**< data has duplicates */ 911 912 /** valid flags for #mdb_node_add() */ 913 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND) 914 915 /** @} */ 916 unsigned short mn_flags; /**< @ref mdb_node */ 917 unsigned short mn_ksize; /**< key size */ 918 char mn_data[1]; /**< key and data are appended here */ 919 } MDB_node; 920 921 /** Size of the node header, excluding dynamic data at the end */ 922 #define NODESIZE offsetof(MDB_node, mn_data) 923 924 /** Bit position of top word in page number, for shifting mn_flags */ 925 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0) 926 927 /** Size of a node in a branch page with a given key. 928 * This is just the node header plus the key, there is no data. 929 */ 930 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size)) 931 932 /** Size of a node in a leaf page with a given key and data. 933 * This is node header plus key plus data size. 934 */ 935 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size) 936 937 /** Address of node \b i in page \b p */ 938 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE)) 939 940 /** Address of the key for the node */ 941 #define NODEKEY(node) (void *)((node)->mn_data) 942 943 /** Address of the data for a node */ 944 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize) 945 946 /** Get the page number pointed to by a branch node */ 947 #define NODEPGNO(node) \ 948 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \ 949 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0)) 950 /** Set the page number in a branch node */ 951 #define SETPGNO(node,pgno) do { \ 952 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \ 953 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0) 954 955 /** Get the size of the data in a leaf node */ 956 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16)) 957 /** Set the size of the data for a leaf node */ 958 #define SETDSZ(node,size) do { \ 959 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0) 960 /** The size of a key in a node */ 961 #define NODEKSZ(node) ((node)->mn_ksize) 962 963 /** Copy a page number from src to dst */ 964 #ifdef MISALIGNED_OK 965 #define COPY_PGNO(dst,src) dst = src 966 #else 967 #if SIZE_MAX > 4294967295UL 968 #define COPY_PGNO(dst,src) do { \ 969 unsigned short *s, *d; \ 970 s = (unsigned short *)&(src); \ 971 d = (unsigned short *)&(dst); \ 972 *d++ = *s++; \ 973 *d++ = *s++; \ 974 *d++ = *s++; \ 975 *d = *s; \ 976 } while (0) 977 #else 978 #define COPY_PGNO(dst,src) do { \ 979 unsigned short *s, *d; \ 980 s = (unsigned short *)&(src); \ 981 d = (unsigned short *)&(dst); \ 982 *d++ = *s++; \ 983 *d = *s; \ 984 } while (0) 985 #endif 986 #endif 987 /** The address of a key in a LEAF2 page. 988 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs. 989 * There are no node headers, keys are stored contiguously. 990 */ 991 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks))) 992 993 /** Set the \b node's key into \b keyptr, if requested. */ 994 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \ 995 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } } 996 997 /** Set the \b node's key into \b key. */ 998 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); } 999 1000 /** Information about a single database in the environment. */ 1001 typedef struct MDB_db { 1002 uint32_t md_pad; /**< also ksize for LEAF2 pages */ 1003 uint16_t md_flags; /**< @ref mdb_dbi_open */ 1004 uint16_t md_depth; /**< depth of this tree */ 1005 pgno_t md_branch_pages; /**< number of internal pages */ 1006 pgno_t md_leaf_pages; /**< number of leaf pages */ 1007 pgno_t md_overflow_pages; /**< number of overflow pages */ 1008 size_t md_entries; /**< number of data items */ 1009 pgno_t md_root; /**< the root page of this tree */ 1010 } MDB_db; 1011 1012 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */ 1013 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID)) 1014 /** #mdb_dbi_open() flags */ 1015 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\ 1016 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE) 1017 1018 /** Handle for the DB used to track free pages. */ 1019 #define FREE_DBI 0 1020 /** Handle for the default DB. */ 1021 #define MAIN_DBI 1 1022 /** Number of DBs in metapage (free and main) - also hardcoded elsewhere */ 1023 #define CORE_DBS 2 1024 1025 /** Number of meta pages - also hardcoded elsewhere */ 1026 #define NUM_METAS 2 1027 1028 /** Meta page content. 1029 * A meta page is the start point for accessing a database snapshot. 1030 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2). 1031 */ 1032 typedef struct MDB_meta { 1033 /** Stamp identifying this as an LMDB file. It must be set 1034 * to #MDB_MAGIC. */ 1035 uint32_t mm_magic; 1036 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */ 1037 uint32_t mm_version; 1038 void *mm_address; /**< address for fixed mapping */ 1039 size_t mm_mapsize; /**< size of mmap region */ 1040 MDB_db mm_dbs[CORE_DBS]; /**< first is free space, 2nd is main db */ 1041 /** The size of pages used in this DB */ 1042 #define mm_psize mm_dbs[FREE_DBI].md_pad 1043 /** Any persistent environment flags. @ref mdb_env */ 1044 #define mm_flags mm_dbs[FREE_DBI].md_flags 1045 /** Last used page in the datafile. 1046 * Actually the file may be shorter if the freeDB lists the final pages. 1047 */ 1048 pgno_t mm_last_pg; 1049 volatile txnid_t mm_txnid; /**< txnid that committed this page */ 1050 } MDB_meta; 1051 1052 /** Buffer for a stack-allocated meta page. 1053 * The members define size and alignment, and silence type 1054 * aliasing warnings. They are not used directly; that could 1055 * mean incorrectly using several union members in parallel. 1056 */ 1057 typedef union MDB_metabuf { 1058 MDB_page mb_page; 1059 struct { 1060 char mm_pad[PAGEHDRSZ]; 1061 MDB_meta mm_meta; 1062 } mb_metabuf; 1063 } MDB_metabuf; 1064 1065 /** Auxiliary DB info. 1066 * The information here is mostly static/read-only. There is 1067 * only a single copy of this record in the environment. 1068 */ 1069 typedef struct MDB_dbx { 1070 MDB_val md_name; /**< name of the database */ 1071 MDB_cmp_func *md_cmp; /**< function for comparing keys */ 1072 MDB_cmp_func *md_dcmp; /**< function for comparing data items */ 1073 MDB_rel_func *md_rel; /**< user relocate function */ 1074 void *md_relctx; /**< user-provided context for md_rel */ 1075 } MDB_dbx; 1076 1077 /** A database transaction. 1078 * Every operation requires a transaction handle. 1079 */ 1080 struct MDB_txn { 1081 MDB_txn *mt_parent; /**< parent of a nested txn */ 1082 /** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */ 1083 MDB_txn *mt_child; 1084 pgno_t mt_next_pgno; /**< next unallocated page */ 1085 /** The ID of this transaction. IDs are integers incrementing from 1. 1086 * Only committed write transactions increment the ID. If a transaction 1087 * aborts, the ID may be re-used by the next writer. 1088 */ 1089 txnid_t mt_txnid; 1090 MDB_env *mt_env; /**< the DB environment */ 1091 /** The list of pages that became unused during this transaction. 1092 */ 1093 MDB_IDL mt_free_pgs; 1094 /** The list of loose pages that became unused and may be reused 1095 * in this transaction, linked through #NEXT_LOOSE_PAGE(page). 1096 */ 1097 MDB_page *mt_loose_pgs; 1098 /** Number of loose pages (#mt_loose_pgs) */ 1099 int mt_loose_count; 1100 /** The sorted list of dirty pages we temporarily wrote to disk 1101 * because the dirty list was full. page numbers in here are 1102 * shifted left by 1, deleted slots have the LSB set. 1103 */ 1104 MDB_IDL mt_spill_pgs; 1105 union { 1106 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */ 1107 MDB_ID2L dirty_list; 1108 /** For read txns: This thread/txn's reader table slot, or NULL. */ 1109 MDB_reader *reader; 1110 } mt_u; 1111 /** Array of records for each DB known in the environment. */ 1112 MDB_dbx *mt_dbxs; 1113 /** Array of MDB_db records for each known DB */ 1114 MDB_db *mt_dbs; 1115 /** Array of sequence numbers for each DB handle */ 1116 unsigned int *mt_dbiseqs; 1117 /** @defgroup mt_dbflag Transaction DB Flags 1118 * @ingroup internal 1119 * @{ 1120 */ 1121 #define DB_DIRTY 0x01 /**< DB was written in this txn */ 1122 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */ 1123 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */ 1124 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */ 1125 #define DB_USRVALID 0x10 /**< As #DB_VALID, but not set for #FREE_DBI */ 1126 #define DB_DUPDATA 0x20 /**< DB is #MDB_DUPSORT data */ 1127 /** @} */ 1128 /** In write txns, array of cursors for each DB */ 1129 MDB_cursor **mt_cursors; 1130 /** Array of flags for each DB */ 1131 unsigned char *mt_dbflags; 1132 /** Number of DB records in use, or 0 when the txn is finished. 1133 * This number only ever increments until the txn finishes; we 1134 * don't decrement it when individual DB handles are closed. 1135 */ 1136 MDB_dbi mt_numdbs; 1137 1138 /** @defgroup mdb_txn Transaction Flags 1139 * @ingroup internal 1140 * @{ 1141 */ 1142 /** #mdb_txn_begin() flags */ 1143 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY 1144 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */ 1145 /* internal txn flags */ 1146 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */ 1147 #define MDB_TXN_FINISHED 0x01 /**< txn is finished or never began */ 1148 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */ 1149 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */ 1150 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */ 1151 #define MDB_TXN_HAS_CHILD 0x10 /**< txn has an #MDB_txn.%mt_child */ 1152 /** most operations on the txn are currently illegal */ 1153 #define MDB_TXN_BLOCKED (MDB_TXN_FINISHED|MDB_TXN_ERROR|MDB_TXN_HAS_CHILD) 1154 /** @} */ 1155 unsigned int mt_flags; /**< @ref mdb_txn */ 1156 /** #dirty_list room: Array size - \#dirty pages visible to this txn. 1157 * Includes ancestor txns' dirty pages not hidden by other txns' 1158 * dirty/spilled pages. Thus commit(nested txn) has room to merge 1159 * dirty_list into mt_parent after freeing hidden mt_parent pages. 1160 */ 1161 unsigned int mt_dirty_room; 1162 }; 1163 1164 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty. 1165 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to 1166 * raise this on a 64 bit machine. 1167 */ 1168 #define CURSOR_STACK 32 1169 1170 struct MDB_xcursor; 1171 1172 /** Cursors are used for all DB operations. 1173 * A cursor holds a path of (page pointer, key index) from the DB 1174 * root to a position in the DB, plus other state. #MDB_DUPSORT 1175 * cursors include an xcursor to the current data item. Write txns 1176 * track their cursors and keep them up to date when data moves. 1177 * Exception: An xcursor's pointer to a #P_SUBP page can be stale. 1178 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage). 1179 */ 1180 struct MDB_cursor { 1181 /** Next cursor on this DB in this txn */ 1182 MDB_cursor *mc_next; 1183 /** Backup of the original cursor if this cursor is a shadow */ 1184 MDB_cursor *mc_backup; 1185 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */ 1186 struct MDB_xcursor *mc_xcursor; 1187 /** The transaction that owns this cursor */ 1188 MDB_txn *mc_txn; 1189 /** The database handle this cursor operates on */ 1190 MDB_dbi mc_dbi; 1191 /** The database record for this cursor */ 1192 MDB_db *mc_db; 1193 /** The database auxiliary record for this cursor */ 1194 MDB_dbx *mc_dbx; 1195 /** The @ref mt_dbflag for this database */ 1196 unsigned char *mc_dbflag; 1197 unsigned short mc_snum; /**< number of pushed pages */ 1198 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */ 1199 /** @defgroup mdb_cursor Cursor Flags 1200 * @ingroup internal 1201 * Cursor state flags. 1202 * @{ 1203 */ 1204 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */ 1205 #define C_EOF 0x02 /**< No more data */ 1206 #define C_SUB 0x04 /**< Cursor is a sub-cursor */ 1207 #define C_DEL 0x08 /**< last op was a cursor_del */ 1208 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */ 1209 /** @} */ 1210 unsigned int mc_flags; /**< @ref mdb_cursor */ 1211 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */ 1212 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */ 1213 }; 1214 1215 /** Context for sorted-dup records. 1216 * We could have gone to a fully recursive design, with arbitrarily 1217 * deep nesting of sub-databases. But for now we only handle these 1218 * levels - main DB, optional sub-DB, sorted-duplicate DB. 1219 */ 1220 typedef struct MDB_xcursor { 1221 /** A sub-cursor for traversing the Dup DB */ 1222 MDB_cursor mx_cursor; 1223 /** The database record for this Dup DB */ 1224 MDB_db mx_db; 1225 /** The auxiliary DB record for this Dup DB */ 1226 MDB_dbx mx_dbx; 1227 /** The @ref mt_dbflag for this Dup DB */ 1228 unsigned char mx_dbflag; 1229 } MDB_xcursor; 1230 1231 /** Check if there is an inited xcursor */ 1232 #define XCURSOR_INITED(mc) \ 1233 ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) 1234 1235 /** Update the xcursor's sub-page pointer, if any, in \b mc. Needed 1236 * when the node which contains the sub-page may have moved. Called 1237 * with leaf page \b mp = mc->mc_pg[\b top]. 1238 */ 1239 #define XCURSOR_REFRESH(mc, top, mp) do { \ 1240 MDB_page *xr_pg = (mp); \ 1241 MDB_node *xr_node; \ 1242 if (!XCURSOR_INITED(mc) || (mc)->mc_ki[top] >= NUMKEYS(xr_pg)) break; \ 1243 xr_node = NODEPTR(xr_pg, (mc)->mc_ki[top]); \ 1244 if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \ 1245 (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \ 1246 } while (0) 1247 1248 /** State of FreeDB old pages, stored in the MDB_env */ 1249 typedef struct MDB_pgstate { 1250 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */ 1251 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */ 1252 } MDB_pgstate; 1253 1254 /** The database environment. */ 1255 struct MDB_env { 1256 HANDLE me_fd; /**< The main data file */ 1257 HANDLE me_lfd; /**< The lock file */ 1258 HANDLE me_mfd; /**< For writing and syncing the meta pages */ 1259 /** Failed to update the meta page. Probably an I/O error. */ 1260 #define MDB_FATAL_ERROR 0x80000000U 1261 /** Some fields are initialized. */ 1262 #define MDB_ENV_ACTIVE 0x20000000U 1263 /** me_txkey is set */ 1264 #define MDB_ENV_TXKEY 0x10000000U 1265 /** fdatasync is unreliable */ 1266 #define MDB_FSYNCONLY 0x08000000U 1267 uint32_t me_flags; /**< @ref mdb_env */ 1268 unsigned int me_psize; /**< DB page size, inited from me_os_psize */ 1269 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */ 1270 unsigned int me_maxreaders; /**< size of the reader table */ 1271 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */ 1272 volatile int me_close_readers; 1273 MDB_dbi me_numdbs; /**< number of DBs opened */ 1274 MDB_dbi me_maxdbs; /**< size of the DB table */ 1275 MDB_PID_T me_pid; /**< process ID of this env */ 1276 char *me_path; /**< path to the DB files */ 1277 char *me_map; /**< the memory map of the data file */ 1278 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */ 1279 MDB_meta *me_metas[NUM_METAS]; /**< pointers to the two meta pages */ 1280 void *me_pbuf; /**< scratch area for DUPSORT put() */ 1281 MDB_txn *me_txn; /**< current write transaction */ 1282 MDB_txn *me_txn0; /**< prealloc'd write transaction */ 1283 size_t me_mapsize; /**< size of the data memory map */ 1284 off_t me_size; /**< current file size */ 1285 pgno_t me_maxpg; /**< me_mapsize / me_psize */ 1286 MDB_dbx *me_dbxs; /**< array of static DB info */ 1287 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */ 1288 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */ 1289 pthread_key_t me_txkey; /**< thread-key for readers */ 1290 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */ 1291 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */ 1292 # define me_pglast me_pgstate.mf_pglast 1293 # define me_pghead me_pgstate.mf_pghead 1294 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */ 1295 /** IDL of pages that became unused in a write txn */ 1296 MDB_IDL me_free_pgs; 1297 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */ 1298 MDB_ID2L me_dirty_list; 1299 /** Max number of freelist items that can fit in a single overflow page */ 1300 int me_maxfree_1pg; 1301 /** Max size of a node on a page */ 1302 unsigned int me_nodemax; 1303 #if !(MDB_MAXKEYSIZE) 1304 unsigned int me_maxkey; /**< max size of a key */ 1305 #endif 1306 int me_live_reader; /**< have liveness lock in reader table */ 1307 #ifdef _WIN32 1308 int me_pidquery; /**< Used in OpenProcess */ 1309 #endif 1310 #ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */ 1311 # define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */ 1312 # define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */ 1313 #else 1314 mdb_mutex_t me_rmutex; 1315 mdb_mutex_t me_wmutex; 1316 #endif 1317 void *me_userctx; /**< User-settable context */ 1318 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */ 1319 }; 1320 1321 /** Nested transaction */ 1322 typedef struct MDB_ntxn { 1323 MDB_txn mnt_txn; /**< the transaction */ 1324 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */ 1325 } MDB_ntxn; 1326 1327 /** max number of pages to commit in one writev() call */ 1328 #define MDB_COMMIT_PAGES 64 1329 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES 1330 #undef MDB_COMMIT_PAGES 1331 #define MDB_COMMIT_PAGES IOV_MAX 1332 #endif 1333 1334 /** max bytes to write in one call */ 1335 #define MAX_WRITE (0x40000000U >> (sizeof(ssize_t) == 4)) 1336 1337 /** Check \b txn and \b dbi arguments to a function */ 1338 #define TXN_DBI_EXIST(txn, dbi, validity) \ 1339 ((txn) && (dbi)<(txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & (validity))) 1340 1341 /** Check for misused \b dbi handles */ 1342 #define TXN_DBI_CHANGED(txn, dbi) \ 1343 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi]) 1344 1345 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp); 1346 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp); 1347 static int mdb_page_touch(MDB_cursor *mc); 1348 1349 #define MDB_END_NAMES {"committed", "empty-commit", "abort", "reset", \ 1350 "reset-tmp", "fail-begin", "fail-beginchild"} 1351 enum { 1352 /* mdb_txn_end operation number, for logging */ 1353 MDB_END_COMMITTED, MDB_END_EMPTY_COMMIT, MDB_END_ABORT, MDB_END_RESET, 1354 MDB_END_RESET_TMP, MDB_END_FAIL_BEGIN, MDB_END_FAIL_BEGINCHILD 1355 }; 1356 #define MDB_END_OPMASK 0x0F /**< mask for #mdb_txn_end() operation number */ 1357 #define MDB_END_UPDATE 0x10 /**< update env state (DBIs) */ 1358 #define MDB_END_FREE 0x20 /**< free txn unless it is #MDB_env.%me_txn0 */ 1359 #define MDB_END_SLOT MDB_NOTLS /**< release any reader slot if #MDB_NOTLS */ 1360 static void mdb_txn_end(MDB_txn *txn, unsigned mode); 1361 1362 static int mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **mp, int *lvl); 1363 static int mdb_page_search_root(MDB_cursor *mc, 1364 MDB_val *key, int modify); 1365 #define MDB_PS_MODIFY 1 1366 #define MDB_PS_ROOTONLY 2 1367 #define MDB_PS_FIRST 4 1368 #define MDB_PS_LAST 8 1369 static int mdb_page_search(MDB_cursor *mc, 1370 MDB_val *key, int flags); 1371 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst); 1372 1373 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */ 1374 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, 1375 pgno_t newpgno, unsigned int nflags); 1376 1377 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta); 1378 static MDB_meta *mdb_env_pick_meta(const MDB_env *env); 1379 static int mdb_env_write_meta(MDB_txn *txn); 1380 #ifdef MDB_USE_POSIX_MUTEX /* Drop unused excl arg */ 1381 # define mdb_env_close0(env, excl) mdb_env_close1(env) 1382 #endif 1383 static void mdb_env_close0(MDB_env *env, int excl); 1384 1385 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp); 1386 static int mdb_node_add(MDB_cursor *mc, indx_t indx, 1387 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags); 1388 static void mdb_node_del(MDB_cursor *mc, int ksize); 1389 static void mdb_node_shrink(MDB_page *mp, indx_t indx); 1390 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft); 1391 static int mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data); 1392 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data); 1393 static size_t mdb_branch_size(MDB_env *env, MDB_val *key); 1394 1395 static int mdb_rebalance(MDB_cursor *mc); 1396 static int mdb_update_key(MDB_cursor *mc, MDB_val *key); 1397 1398 static void mdb_cursor_pop(MDB_cursor *mc); 1399 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp); 1400 1401 static int mdb_cursor_del0(MDB_cursor *mc); 1402 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags); 1403 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right); 1404 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op); 1405 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op); 1406 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op, 1407 int *exactp); 1408 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data); 1409 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data); 1410 1411 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx); 1412 static void mdb_xcursor_init0(MDB_cursor *mc); 1413 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node); 1414 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force); 1415 1416 static int mdb_drop0(MDB_cursor *mc, int subs); 1417 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi); 1418 static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead); 1419 1420 /** @cond */ 1421 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long; 1422 /** @endcond */ 1423 1424 /** Compare two items pointing at size_t's of unknown alignment. */ 1425 #ifdef MISALIGNED_OK 1426 # define mdb_cmp_clong mdb_cmp_long 1427 #else 1428 # define mdb_cmp_clong mdb_cmp_cint 1429 #endif 1430 1431 #ifdef _WIN32 1432 static SECURITY_DESCRIPTOR mdb_null_sd; 1433 static SECURITY_ATTRIBUTES mdb_all_sa; 1434 static int mdb_sec_inited; 1435 1436 struct MDB_name; 1437 static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra); 1438 #endif 1439 1440 /** Return the library version info. */ 1441 char * ESECT 1442 mdb_version(int *major, int *minor, int *patch) 1443 { 1444 if (major) *major = MDB_VERSION_MAJOR; 1445 if (minor) *minor = MDB_VERSION_MINOR; 1446 if (patch) *patch = MDB_VERSION_PATCH; 1447 return MDB_VERSION_STRING; 1448 } 1449 1450 /** Table of descriptions for LMDB @ref errors */ 1451 static char *const mdb_errstr[] = { 1452 "MDB_KEYEXIST: Key/data pair already exists", 1453 "MDB_NOTFOUND: No matching key/data pair found", 1454 "MDB_PAGE_NOTFOUND: Requested page not found", 1455 "MDB_CORRUPTED: Located page was wrong type", 1456 "MDB_PANIC: Update of meta page failed or environment had fatal error", 1457 "MDB_VERSION_MISMATCH: Database environment version mismatch", 1458 "MDB_INVALID: File is not an LMDB file", 1459 "MDB_MAP_FULL: Environment mapsize limit reached", 1460 "MDB_DBS_FULL: Environment maxdbs limit reached", 1461 "MDB_READERS_FULL: Environment maxreaders limit reached", 1462 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open", 1463 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big", 1464 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached", 1465 "MDB_PAGE_FULL: Internal error - page has no more space", 1466 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize", 1467 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed", 1468 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot", 1469 "MDB_BAD_TXN: Transaction must abort, has a child, or is invalid", 1470 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size", 1471 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly", 1472 }; 1473 1474 char * 1475 mdb_strerror(int err) 1476 { 1477 #ifdef _WIN32 1478 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf. 1479 * This works as long as no function between the call to mdb_strerror 1480 * and the actual use of the message uses more than 4K of stack. 1481 */ 1482 #define MSGSIZE 1024 1483 #define PADSIZE 4096 1484 char buf[MSGSIZE+PADSIZE], *ptr = buf; 1485 #endif 1486 int i; 1487 if (!err) 1488 return ("Successful return: 0"); 1489 1490 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) { 1491 i = err - MDB_KEYEXIST; 1492 return mdb_errstr[i]; 1493 } 1494 1495 #ifdef _WIN32 1496 /* These are the C-runtime error codes we use. The comment indicates 1497 * their numeric value, and the Win32 error they would correspond to 1498 * if the error actually came from a Win32 API. A major mess, we should 1499 * have used LMDB-specific error codes for everything. 1500 */ 1501 switch(err) { 1502 case ENOENT: /* 2, FILE_NOT_FOUND */ 1503 case EIO: /* 5, ACCESS_DENIED */ 1504 case ENOMEM: /* 12, INVALID_ACCESS */ 1505 case EACCES: /* 13, INVALID_DATA */ 1506 case EBUSY: /* 16, CURRENT_DIRECTORY */ 1507 case EINVAL: /* 22, BAD_COMMAND */ 1508 case ENOSPC: /* 28, OUT_OF_PAPER */ 1509 return strerror(err); 1510 default: 1511 ; 1512 } 1513 buf[0] = 0; 1514 FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM | 1515 FORMAT_MESSAGE_IGNORE_INSERTS, 1516 NULL, err, 0, ptr, MSGSIZE, (va_list *)buf+MSGSIZE); 1517 return ptr; 1518 #else 1519 return strerror(err); 1520 #endif 1521 } 1522 1523 /** assert(3) variant in cursor context */ 1524 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr) 1525 /** assert(3) variant in transaction context */ 1526 #define mdb_tassert(txn, expr) mdb_assert0((txn)->mt_env, expr, #expr) 1527 /** assert(3) variant in environment context */ 1528 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr) 1529 1530 #ifndef NDEBUG 1531 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \ 1532 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__)) 1533 1534 static void ESECT 1535 mdb_assert_fail(MDB_env *env, const char *expr_txt, 1536 const char *func, const char *file, int line) 1537 { 1538 char buf[400]; 1539 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()", 1540 file, line, expr_txt, func); 1541 if (env->me_assert_func) 1542 env->me_assert_func(env, buf); 1543 fprintf(stderr, "%s\n", buf); 1544 abort(); 1545 } 1546 #else 1547 # define mdb_assert0(env, expr, expr_txt) ((void) 0) 1548 #endif /* NDEBUG */ 1549 1550 #if MDB_DEBUG 1551 /** Return the page number of \b mp which may be sub-page, for debug output */ 1552 static pgno_t 1553 mdb_dbg_pgno(MDB_page *mp) 1554 { 1555 pgno_t ret; 1556 COPY_PGNO(ret, mp->mp_pgno); 1557 return ret; 1558 } 1559 1560 /** Display a key in hexadecimal and return the address of the result. 1561 * @param[in] key the key to display 1562 * @param[in] buf the buffer to write into. Should always be #DKBUF. 1563 * @return The key in hexadecimal form. 1564 */ 1565 char * 1566 mdb_dkey(MDB_val *key, char *buf) 1567 { 1568 char *ptr = buf; 1569 unsigned char *c = key->mv_data; 1570 unsigned int i; 1571 1572 if (!key) 1573 return ""; 1574 1575 if (key->mv_size > DKBUF_MAXKEYSIZE) 1576 return "MDB_MAXKEYSIZE"; 1577 /* may want to make this a dynamic check: if the key is mostly 1578 * printable characters, print it as-is instead of converting to hex. 1579 */ 1580 #if 1 1581 buf[0] = '\0'; 1582 for (i=0; i<key->mv_size; i++) 1583 ptr += sprintf(ptr, "%02x", *c++); 1584 #else 1585 sprintf(buf, "%.*s", key->mv_size, key->mv_data); 1586 #endif 1587 return buf; 1588 } 1589 1590 static const char * 1591 mdb_leafnode_type(MDB_node *n) 1592 { 1593 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}}; 1594 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" : 1595 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)]; 1596 } 1597 1598 /** Display all the keys in the page. */ 1599 void 1600 mdb_page_list(MDB_page *mp) 1601 { 1602 pgno_t pgno = mdb_dbg_pgno(mp); 1603 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : ""; 1604 MDB_node *node; 1605 unsigned int i, nkeys, nsize, total = 0; 1606 MDB_val key; 1607 DKBUF; 1608 1609 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) { 1610 case P_BRANCH: type = "Branch page"; break; 1611 case P_LEAF: type = "Leaf page"; break; 1612 case P_LEAF|P_SUBP: type = "Sub-page"; break; 1613 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break; 1614 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break; 1615 case P_OVERFLOW: 1616 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n", 1617 pgno, mp->mp_pages, state); 1618 return; 1619 case P_META: 1620 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n", 1621 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid); 1622 return; 1623 default: 1624 fprintf(stderr, "Bad page %"Z"u flags 0x%X\n", pgno, mp->mp_flags); 1625 return; 1626 } 1627 1628 nkeys = NUMKEYS(mp); 1629 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state); 1630 1631 for (i=0; i<nkeys; i++) { 1632 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */ 1633 key.mv_size = nsize = mp->mp_pad; 1634 key.mv_data = LEAF2KEY(mp, i, nsize); 1635 total += nsize; 1636 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key)); 1637 continue; 1638 } 1639 node = NODEPTR(mp, i); 1640 key.mv_size = node->mn_ksize; 1641 key.mv_data = node->mn_data; 1642 nsize = NODESIZE + key.mv_size; 1643 if (IS_BRANCH(mp)) { 1644 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node), 1645 DKEY(&key)); 1646 total += nsize; 1647 } else { 1648 if (F_ISSET(node->mn_flags, F_BIGDATA)) 1649 nsize += sizeof(pgno_t); 1650 else 1651 nsize += NODEDSZ(node); 1652 total += nsize; 1653 nsize += sizeof(indx_t); 1654 fprintf(stderr, "key %d: nsize %d, %s%s\n", 1655 i, nsize, DKEY(&key), mdb_leafnode_type(node)); 1656 } 1657 total = EVEN(total); 1658 } 1659 fprintf(stderr, "Total: header %d + contents %d + unused %d\n", 1660 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp)); 1661 } 1662 1663 void 1664 mdb_cursor_chk(MDB_cursor *mc) 1665 { 1666 unsigned int i; 1667 MDB_node *node; 1668 MDB_page *mp; 1669 1670 if (!mc->mc_snum || !(mc->mc_flags & C_INITIALIZED)) return; 1671 for (i=0; i<mc->mc_top; i++) { 1672 mp = mc->mc_pg[i]; 1673 node = NODEPTR(mp, mc->mc_ki[i]); 1674 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno) 1675 printf("oops!\n"); 1676 } 1677 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i])) 1678 printf("ack!\n"); 1679 if (XCURSOR_INITED(mc)) { 1680 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 1681 if (((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) && 1682 mc->mc_xcursor->mx_cursor.mc_pg[0] != NODEDATA(node)) { 1683 printf("blah!\n"); 1684 } 1685 } 1686 } 1687 #endif 1688 1689 #if (MDB_DEBUG) > 2 1690 /** Count all the pages in each DB and in the freelist 1691 * and make sure it matches the actual number of pages 1692 * being used. 1693 * All named DBs must be open for a correct count. 1694 */ 1695 static void mdb_audit(MDB_txn *txn) 1696 { 1697 MDB_cursor mc; 1698 MDB_val key, data; 1699 MDB_ID freecount, count; 1700 MDB_dbi i; 1701 int rc; 1702 1703 freecount = 0; 1704 mdb_cursor_init(&mc, txn, FREE_DBI, NULL); 1705 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0) 1706 freecount += *(MDB_ID *)data.mv_data; 1707 mdb_tassert(txn, rc == MDB_NOTFOUND); 1708 1709 count = 0; 1710 for (i = 0; i<txn->mt_numdbs; i++) { 1711 MDB_xcursor mx; 1712 if (!(txn->mt_dbflags[i] & DB_VALID)) 1713 continue; 1714 mdb_cursor_init(&mc, txn, i, &mx); 1715 if (txn->mt_dbs[i].md_root == P_INVALID) 1716 continue; 1717 count += txn->mt_dbs[i].md_branch_pages + 1718 txn->mt_dbs[i].md_leaf_pages + 1719 txn->mt_dbs[i].md_overflow_pages; 1720 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) { 1721 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST); 1722 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) { 1723 unsigned j; 1724 MDB_page *mp; 1725 mp = mc.mc_pg[mc.mc_top]; 1726 for (j=0; j<NUMKEYS(mp); j++) { 1727 MDB_node *leaf = NODEPTR(mp, j); 1728 if (leaf->mn_flags & F_SUBDATA) { 1729 MDB_db db; 1730 memcpy(&db, NODEDATA(leaf), sizeof(db)); 1731 count += db.md_branch_pages + db.md_leaf_pages + 1732 db.md_overflow_pages; 1733 } 1734 } 1735 } 1736 mdb_tassert(txn, rc == MDB_NOTFOUND); 1737 } 1738 } 1739 if (freecount + count + NUM_METAS != txn->mt_next_pgno) { 1740 fprintf(stderr, "audit: %"Z"u freecount: %"Z"u count: %"Z"u total: %"Z"u next_pgno: %"Z"u\n", 1741 txn->mt_txnid, freecount, count+NUM_METAS, 1742 freecount+count+NUM_METAS, txn->mt_next_pgno); 1743 } 1744 } 1745 #endif 1746 1747 int 1748 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b) 1749 { 1750 return txn->mt_dbxs[dbi].md_cmp(a, b); 1751 } 1752 1753 int 1754 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b) 1755 { 1756 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp; 1757 #if UINT_MAX < SIZE_MAX 1758 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t)) 1759 dcmp = mdb_cmp_clong; 1760 #endif 1761 return dcmp(a, b); 1762 } 1763 1764 /** Allocate memory for a page. 1765 * Re-use old malloc'd pages first for singletons, otherwise just malloc. 1766 * Set #MDB_TXN_ERROR on failure. 1767 */ 1768 static MDB_page * 1769 mdb_page_malloc(MDB_txn *txn, unsigned num) 1770 { 1771 MDB_env *env = txn->mt_env; 1772 MDB_page *ret = env->me_dpages; 1773 size_t psize = env->me_psize, sz = psize, off; 1774 /* For ! #MDB_NOMEMINIT, psize counts how much to init. 1775 * For a single page alloc, we init everything after the page header. 1776 * For multi-page, we init the final page; if the caller needed that 1777 * many pages they will be filling in at least up to the last page. 1778 */ 1779 if (num == 1) { 1780 if (ret) { 1781 VGMEMP_ALLOC(env, ret, sz); 1782 VGMEMP_DEFINED(ret, sizeof(ret->mp_next)); 1783 env->me_dpages = ret->mp_next; 1784 return ret; 1785 } 1786 psize -= off = PAGEHDRSZ; 1787 } else { 1788 sz *= num; 1789 off = sz - psize; 1790 } 1791 if ((ret = malloc(sz)) != NULL) { 1792 VGMEMP_ALLOC(env, ret, sz); 1793 if (!(env->me_flags & MDB_NOMEMINIT)) { 1794 memset((char *)ret + off, 0, psize); 1795 ret->mp_pad = 0; 1796 } 1797 } else { 1798 txn->mt_flags |= MDB_TXN_ERROR; 1799 } 1800 return ret; 1801 } 1802 /** Free a single page. 1803 * Saves single pages to a list, for future reuse. 1804 * (This is not used for multi-page overflow pages.) 1805 */ 1806 static void 1807 mdb_page_free(MDB_env *env, MDB_page *mp) 1808 { 1809 mp->mp_next = env->me_dpages; 1810 VGMEMP_FREE(env, mp); 1811 env->me_dpages = mp; 1812 } 1813 1814 /** Free a dirty page */ 1815 static void 1816 mdb_dpage_free(MDB_env *env, MDB_page *dp) 1817 { 1818 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) { 1819 mdb_page_free(env, dp); 1820 } else { 1821 /* large pages just get freed directly */ 1822 VGMEMP_FREE(env, dp); 1823 free(dp); 1824 } 1825 } 1826 1827 /** Return all dirty pages to dpage list */ 1828 static void 1829 mdb_dlist_free(MDB_txn *txn) 1830 { 1831 MDB_env *env = txn->mt_env; 1832 MDB_ID2L dl = txn->mt_u.dirty_list; 1833 unsigned i, n = dl[0].mid; 1834 1835 for (i = 1; i <= n; i++) { 1836 mdb_dpage_free(env, dl[i].mptr); 1837 } 1838 dl[0].mid = 0; 1839 } 1840 1841 /** Loosen or free a single page. 1842 * Saves single pages to a list for future reuse 1843 * in this same txn. It has been pulled from the freeDB 1844 * and already resides on the dirty list, but has been 1845 * deleted. Use these pages first before pulling again 1846 * from the freeDB. 1847 * 1848 * If the page wasn't dirtied in this txn, just add it 1849 * to this txn's free list. 1850 */ 1851 static int 1852 mdb_page_loose(MDB_cursor *mc, MDB_page *mp) 1853 { 1854 int loose = 0; 1855 pgno_t pgno = mp->mp_pgno; 1856 MDB_txn *txn = mc->mc_txn; 1857 1858 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) { 1859 if (txn->mt_parent) { 1860 MDB_ID2 *dl = txn->mt_u.dirty_list; 1861 /* If txn has a parent, make sure the page is in our 1862 * dirty list. 1863 */ 1864 if (dl[0].mid) { 1865 unsigned x = mdb_mid2l_search(dl, pgno); 1866 if (x <= dl[0].mid && dl[x].mid == pgno) { 1867 if (mp != dl[x].mptr) { /* bad cursor? */ 1868 mc->mc_flags &= ~(C_INITIALIZED|C_EOF); 1869 txn->mt_flags |= MDB_TXN_ERROR; 1870 return MDB_CORRUPTED; 1871 } 1872 /* ok, it's ours */ 1873 loose = 1; 1874 } 1875 } 1876 } else { 1877 /* no parent txn, so it's just ours */ 1878 loose = 1; 1879 } 1880 } 1881 if (loose) { 1882 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc), 1883 mp->mp_pgno)); 1884 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs; 1885 txn->mt_loose_pgs = mp; 1886 txn->mt_loose_count++; 1887 mp->mp_flags |= P_LOOSE; 1888 } else { 1889 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno); 1890 if (rc) 1891 return rc; 1892 } 1893 1894 return MDB_SUCCESS; 1895 } 1896 1897 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn. 1898 * @param[in] mc A cursor handle for the current operation. 1899 * @param[in] pflags Flags of the pages to update: 1900 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it. 1901 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush(). 1902 * @return 0 on success, non-zero on failure. 1903 */ 1904 static int 1905 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all) 1906 { 1907 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP }; 1908 MDB_txn *txn = mc->mc_txn; 1909 MDB_cursor *m3, *m0 = mc; 1910 MDB_xcursor *mx; 1911 MDB_page *dp, *mp; 1912 MDB_node *leaf; 1913 unsigned i, j; 1914 int rc = MDB_SUCCESS, level; 1915 1916 /* Mark pages seen by cursors */ 1917 if (mc->mc_flags & C_UNTRACK) 1918 mc = NULL; /* will find mc in mt_cursors */ 1919 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) { 1920 for (; mc; mc=mc->mc_next) { 1921 if (!(mc->mc_flags & C_INITIALIZED)) 1922 continue; 1923 for (m3 = mc;; m3 = &mx->mx_cursor) { 1924 mp = NULL; 1925 for (j=0; j<m3->mc_snum; j++) { 1926 mp = m3->mc_pg[j]; 1927 if ((mp->mp_flags & Mask) == pflags) 1928 mp->mp_flags ^= P_KEEP; 1929 } 1930 mx = m3->mc_xcursor; 1931 /* Proceed to mx if it is at a sub-database */ 1932 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED))) 1933 break; 1934 if (! (mp && (mp->mp_flags & P_LEAF))) 1935 break; 1936 leaf = NODEPTR(mp, m3->mc_ki[j-1]); 1937 if (!(leaf->mn_flags & F_SUBDATA)) 1938 break; 1939 } 1940 } 1941 if (i == 0) 1942 break; 1943 } 1944 1945 if (all) { 1946 /* Mark dirty root pages */ 1947 for (i=0; i<txn->mt_numdbs; i++) { 1948 if (txn->mt_dbflags[i] & DB_DIRTY) { 1949 pgno_t pgno = txn->mt_dbs[i].md_root; 1950 if (pgno == P_INVALID) 1951 continue; 1952 if ((rc = mdb_page_get(m0, pgno, &dp, &level)) != MDB_SUCCESS) 1953 break; 1954 if ((dp->mp_flags & Mask) == pflags && level <= 1) 1955 dp->mp_flags ^= P_KEEP; 1956 } 1957 } 1958 } 1959 1960 return rc; 1961 } 1962 1963 static int mdb_page_flush(MDB_txn *txn, int keep); 1964 1965 /** Spill pages from the dirty list back to disk. 1966 * This is intended to prevent running into #MDB_TXN_FULL situations, 1967 * but note that they may still occur in a few cases: 1968 * 1) our estimate of the txn size could be too small. Currently this 1969 * seems unlikely, except with a large number of #MDB_MULTIPLE items. 1970 * 2) child txns may run out of space if their parents dirtied a 1971 * lot of pages and never spilled them. TODO: we probably should do 1972 * a preemptive spill during #mdb_txn_begin() of a child txn, if 1973 * the parent's dirty_room is below a given threshold. 1974 * 1975 * Otherwise, if not using nested txns, it is expected that apps will 1976 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk 1977 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared. 1978 * If the txn never references them again, they can be left alone. 1979 * If the txn only reads them, they can be used without any fuss. 1980 * If the txn writes them again, they can be dirtied immediately without 1981 * going thru all of the work of #mdb_page_touch(). Such references are 1982 * handled by #mdb_page_unspill(). 1983 * 1984 * Also note, we never spill DB root pages, nor pages of active cursors, 1985 * because we'll need these back again soon anyway. And in nested txns, 1986 * we can't spill a page in a child txn if it was already spilled in a 1987 * parent txn. That would alter the parent txns' data even though 1988 * the child hasn't committed yet, and we'd have no way to undo it if 1989 * the child aborted. 1990 * 1991 * @param[in] m0 cursor A cursor handle identifying the transaction and 1992 * database for which we are checking space. 1993 * @param[in] key For a put operation, the key being stored. 1994 * @param[in] data For a put operation, the data being stored. 1995 * @return 0 on success, non-zero on failure. 1996 */ 1997 static int 1998 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data) 1999 { 2000 MDB_txn *txn = m0->mc_txn; 2001 MDB_page *dp; 2002 MDB_ID2L dl = txn->mt_u.dirty_list; 2003 unsigned int i, j, need; 2004 int rc; 2005 2006 if (m0->mc_flags & C_SUB) 2007 return MDB_SUCCESS; 2008 2009 /* Estimate how much space this op will take */ 2010 i = m0->mc_db->md_depth; 2011 /* Named DBs also dirty the main DB */ 2012 if (m0->mc_dbi >= CORE_DBS) 2013 i += txn->mt_dbs[MAIN_DBI].md_depth; 2014 /* For puts, roughly factor in the key+data size */ 2015 if (key) 2016 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize; 2017 i += i; /* double it for good measure */ 2018 need = i; 2019 2020 if (txn->mt_dirty_room > i) 2021 return MDB_SUCCESS; 2022 2023 if (!txn->mt_spill_pgs) { 2024 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX); 2025 if (!txn->mt_spill_pgs) 2026 return ENOMEM; 2027 } else { 2028 /* purge deleted slots */ 2029 MDB_IDL sl = txn->mt_spill_pgs; 2030 unsigned int num = sl[0]; 2031 j=0; 2032 for (i=1; i<=num; i++) { 2033 if (!(sl[i] & 1)) 2034 sl[++j] = sl[i]; 2035 } 2036 sl[0] = j; 2037 } 2038 2039 /* Preserve pages which may soon be dirtied again */ 2040 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS) 2041 goto done; 2042 2043 /* Less aggressive spill - we originally spilled the entire dirty list, 2044 * with a few exceptions for cursor pages and DB root pages. But this 2045 * turns out to be a lot of wasted effort because in a large txn many 2046 * of those pages will need to be used again. So now we spill only 1/8th 2047 * of the dirty pages. Testing revealed this to be a good tradeoff, 2048 * better than 1/2, 1/4, or 1/10. 2049 */ 2050 if (need < MDB_IDL_UM_MAX / 8) 2051 need = MDB_IDL_UM_MAX / 8; 2052 2053 /* Save the page IDs of all the pages we're flushing */ 2054 /* flush from the tail forward, this saves a lot of shifting later on. */ 2055 for (i=dl[0].mid; i && need; i--) { 2056 MDB_ID pn = dl[i].mid << 1; 2057 dp = dl[i].mptr; 2058 if (dp->mp_flags & (P_LOOSE|P_KEEP)) 2059 continue; 2060 /* Can't spill twice, make sure it's not already in a parent's 2061 * spill list. 2062 */ 2063 if (txn->mt_parent) { 2064 MDB_txn *tx2; 2065 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) { 2066 if (tx2->mt_spill_pgs) { 2067 j = mdb_midl_search(tx2->mt_spill_pgs, pn); 2068 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) { 2069 dp->mp_flags |= P_KEEP; 2070 break; 2071 } 2072 } 2073 } 2074 if (tx2) 2075 continue; 2076 } 2077 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn))) 2078 goto done; 2079 need--; 2080 } 2081 mdb_midl_sort(txn->mt_spill_pgs); 2082 2083 /* Flush the spilled part of dirty list */ 2084 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS) 2085 goto done; 2086 2087 /* Reset any dirty pages we kept that page_flush didn't see */ 2088 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i); 2089 2090 done: 2091 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS; 2092 return rc; 2093 } 2094 2095 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */ 2096 static txnid_t 2097 mdb_find_oldest(MDB_txn *txn) 2098 { 2099 int i; 2100 txnid_t mr, oldest = txn->mt_txnid - 1; 2101 if (txn->mt_env->me_txns) { 2102 MDB_reader *r = txn->mt_env->me_txns->mti_readers; 2103 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) { 2104 if (r[i].mr_pid) { 2105 mr = r[i].mr_txnid; 2106 if (oldest > mr) 2107 oldest = mr; 2108 } 2109 } 2110 } 2111 return oldest; 2112 } 2113 2114 /** Add a page to the txn's dirty list */ 2115 static void 2116 mdb_page_dirty(MDB_txn *txn, MDB_page *mp) 2117 { 2118 MDB_ID2 mid; 2119 int rc, (*insert)(MDB_ID2L, MDB_ID2 *); 2120 2121 if (txn->mt_flags & MDB_TXN_WRITEMAP) { 2122 insert = mdb_mid2l_append; 2123 } else { 2124 insert = mdb_mid2l_insert; 2125 } 2126 mid.mid = mp->mp_pgno; 2127 mid.mptr = mp; 2128 rc = insert(txn->mt_u.dirty_list, &mid); 2129 mdb_tassert(txn, rc == 0); 2130 txn->mt_dirty_room--; 2131 } 2132 2133 /** Allocate page numbers and memory for writing. Maintain me_pglast, 2134 * me_pghead and mt_next_pgno. Set #MDB_TXN_ERROR on failure. 2135 * 2136 * If there are free pages available from older transactions, they 2137 * are re-used first. Otherwise allocate a new page at mt_next_pgno. 2138 * Do not modify the freedB, just merge freeDB records into me_pghead[] 2139 * and move me_pglast to say which records were consumed. Only this 2140 * function can create me_pghead and move me_pglast/mt_next_pgno. 2141 * @param[in] mc cursor A cursor handle identifying the transaction and 2142 * database for which we are allocating. 2143 * @param[in] num the number of pages to allocate. 2144 * @param[out] mp Address of the allocated page(s). Requests for multiple pages 2145 * will always be satisfied by a single contiguous chunk of memory. 2146 * @return 0 on success, non-zero on failure. 2147 */ 2148 static int 2149 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) 2150 { 2151 #ifdef MDB_PARANOID /* Seems like we can ignore this now */ 2152 /* Get at most <Max_retries> more freeDB records once me_pghead 2153 * has enough pages. If not enough, use new pages from the map. 2154 * If <Paranoid> and mc is updating the freeDB, only get new 2155 * records if me_pghead is empty. Then the freelist cannot play 2156 * catch-up with itself by growing while trying to save it. 2157 */ 2158 enum { Paranoid = 1, Max_retries = 500 }; 2159 #else 2160 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ }; 2161 #endif 2162 int rc, retry = num * 60; 2163 MDB_txn *txn = mc->mc_txn; 2164 MDB_env *env = txn->mt_env; 2165 pgno_t pgno, *mop = env->me_pghead; 2166 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1; 2167 MDB_page *np; 2168 txnid_t oldest = 0, last; 2169 MDB_cursor_op op; 2170 MDB_cursor m2; 2171 int found_old = 0; 2172 2173 /* If there are any loose pages, just use them */ 2174 if (num == 1 && txn->mt_loose_pgs) { 2175 np = txn->mt_loose_pgs; 2176 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np); 2177 txn->mt_loose_count--; 2178 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc), 2179 np->mp_pgno)); 2180 *mp = np; 2181 return MDB_SUCCESS; 2182 } 2183 2184 *mp = NULL; 2185 2186 /* If our dirty list is already full, we can't do anything */ 2187 if (txn->mt_dirty_room == 0) { 2188 rc = MDB_TXN_FULL; 2189 goto fail; 2190 } 2191 2192 for (op = MDB_FIRST;; op = MDB_NEXT) { 2193 MDB_val key, data; 2194 MDB_node *leaf; 2195 pgno_t *idl; 2196 2197 /* Seek a big enough contiguous page range. Prefer 2198 * pages at the tail, just truncating the list. 2199 */ 2200 if (mop_len > n2) { 2201 i = mop_len; 2202 do { 2203 pgno = mop[i]; 2204 if (mop[i-n2] == pgno+n2) 2205 goto search_done; 2206 } while (--i > n2); 2207 if (--retry < 0) 2208 break; 2209 } 2210 2211 if (op == MDB_FIRST) { /* 1st iteration */ 2212 /* Prepare to fetch more and coalesce */ 2213 last = env->me_pglast; 2214 oldest = env->me_pgoldest; 2215 mdb_cursor_init(&m2, txn, FREE_DBI, NULL); 2216 if (last) { 2217 op = MDB_SET_RANGE; 2218 key.mv_data = &last; /* will look up last+1 */ 2219 key.mv_size = sizeof(last); 2220 } 2221 if (Paranoid && mc->mc_dbi == FREE_DBI) 2222 retry = -1; 2223 } 2224 if (Paranoid && retry < 0 && mop_len) 2225 break; 2226 2227 last++; 2228 /* Do not fetch more if the record will be too recent */ 2229 if (oldest <= last) { 2230 if (!found_old) { 2231 oldest = mdb_find_oldest(txn); 2232 env->me_pgoldest = oldest; 2233 found_old = 1; 2234 } 2235 if (oldest <= last) 2236 break; 2237 } 2238 rc = mdb_cursor_get(&m2, &key, NULL, op); 2239 if (rc) { 2240 if (rc == MDB_NOTFOUND) 2241 break; 2242 goto fail; 2243 } 2244 last = *(txnid_t*)key.mv_data; 2245 if (oldest <= last) { 2246 if (!found_old) { 2247 oldest = mdb_find_oldest(txn); 2248 env->me_pgoldest = oldest; 2249 found_old = 1; 2250 } 2251 if (oldest <= last) 2252 break; 2253 } 2254 np = m2.mc_pg[m2.mc_top]; 2255 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]); 2256 if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS) 2257 goto fail; 2258 2259 idl = (MDB_ID *) data.mv_data; 2260 i = idl[0]; 2261 if (!mop) { 2262 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) { 2263 rc = ENOMEM; 2264 goto fail; 2265 } 2266 } else { 2267 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0) 2268 goto fail; 2269 mop = env->me_pghead; 2270 } 2271 env->me_pglast = last; 2272 #if (MDB_DEBUG) > 1 2273 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u", 2274 last, txn->mt_dbs[FREE_DBI].md_root, i)); 2275 for (j = i; j; j--) 2276 DPRINTF(("IDL %"Z"u", idl[j])); 2277 #endif 2278 /* Merge in descending sorted order */ 2279 mdb_midl_xmerge(mop, idl); 2280 mop_len = mop[0]; 2281 } 2282 2283 /* Use new pages from the map when nothing suitable in the freeDB */ 2284 i = 0; 2285 pgno = txn->mt_next_pgno; 2286 if (pgno + num >= env->me_maxpg) { 2287 DPUTS("DB size maxed out"); 2288 rc = MDB_MAP_FULL; 2289 goto fail; 2290 } 2291 2292 search_done: 2293 if (env->me_flags & MDB_WRITEMAP) { 2294 np = (MDB_page *)(env->me_map + env->me_psize * pgno); 2295 } else { 2296 if (!(np = mdb_page_malloc(txn, num))) { 2297 rc = ENOMEM; 2298 goto fail; 2299 } 2300 } 2301 if (i) { 2302 mop[0] = mop_len -= num; 2303 /* Move any stragglers down */ 2304 for (j = i-num; j < mop_len; ) 2305 mop[++j] = mop[++i]; 2306 } else { 2307 txn->mt_next_pgno = pgno + num; 2308 } 2309 np->mp_pgno = pgno; 2310 mdb_page_dirty(txn, np); 2311 *mp = np; 2312 2313 return MDB_SUCCESS; 2314 2315 fail: 2316 txn->mt_flags |= MDB_TXN_ERROR; 2317 return rc; 2318 } 2319 2320 /** Copy the used portions of a non-overflow page. 2321 * @param[in] dst page to copy into 2322 * @param[in] src page to copy from 2323 * @param[in] psize size of a page 2324 */ 2325 static void 2326 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize) 2327 { 2328 enum { Align = sizeof(pgno_t) }; 2329 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower; 2330 2331 /* If page isn't full, just copy the used portion. Adjust 2332 * alignment so memcpy may copy words instead of bytes. 2333 */ 2334 if ((unused &= -Align) && !IS_LEAF2(src)) { 2335 upper = (upper + PAGEBASE) & -Align; 2336 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align); 2337 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper), 2338 psize - upper); 2339 } else { 2340 memcpy(dst, src, psize - unused); 2341 } 2342 } 2343 2344 /** Pull a page off the txn's spill list, if present. 2345 * If a page being referenced was spilled to disk in this txn, bring 2346 * it back and make it dirty/writable again. 2347 * @param[in] txn the transaction handle. 2348 * @param[in] mp the page being referenced. It must not be dirty. 2349 * @param[out] ret the writable page, if any. ret is unchanged if 2350 * mp wasn't spilled. 2351 */ 2352 static int 2353 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret) 2354 { 2355 MDB_env *env = txn->mt_env; 2356 const MDB_txn *tx2; 2357 unsigned x; 2358 pgno_t pgno = mp->mp_pgno, pn = pgno << 1; 2359 2360 for (tx2 = txn; tx2; tx2=tx2->mt_parent) { 2361 if (!tx2->mt_spill_pgs) 2362 continue; 2363 x = mdb_midl_search(tx2->mt_spill_pgs, pn); 2364 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) { 2365 MDB_page *np; 2366 int num; 2367 if (txn->mt_dirty_room == 0) 2368 return MDB_TXN_FULL; 2369 if (IS_OVERFLOW(mp)) 2370 num = mp->mp_pages; 2371 else 2372 num = 1; 2373 if (env->me_flags & MDB_WRITEMAP) { 2374 np = mp; 2375 } else { 2376 np = mdb_page_malloc(txn, num); 2377 if (!np) 2378 return ENOMEM; 2379 if (num > 1) 2380 memcpy(np, mp, num * env->me_psize); 2381 else 2382 mdb_page_copy(np, mp, env->me_psize); 2383 } 2384 if (tx2 == txn) { 2385 /* If in current txn, this page is no longer spilled. 2386 * If it happens to be the last page, truncate the spill list. 2387 * Otherwise mark it as deleted by setting the LSB. 2388 */ 2389 if (x == txn->mt_spill_pgs[0]) 2390 txn->mt_spill_pgs[0]--; 2391 else 2392 txn->mt_spill_pgs[x] |= 1; 2393 } /* otherwise, if belonging to a parent txn, the 2394 * page remains spilled until child commits 2395 */ 2396 2397 mdb_page_dirty(txn, np); 2398 np->mp_flags |= P_DIRTY; 2399 *ret = np; 2400 break; 2401 } 2402 } 2403 return MDB_SUCCESS; 2404 } 2405 2406 /** Touch a page: make it dirty and re-insert into tree with updated pgno. 2407 * Set #MDB_TXN_ERROR on failure. 2408 * @param[in] mc cursor pointing to the page to be touched 2409 * @return 0 on success, non-zero on failure. 2410 */ 2411 static int 2412 mdb_page_touch(MDB_cursor *mc) 2413 { 2414 MDB_page *mp = mc->mc_pg[mc->mc_top], *np; 2415 MDB_txn *txn = mc->mc_txn; 2416 MDB_cursor *m2, *m3; 2417 pgno_t pgno; 2418 int rc; 2419 2420 if (!F_ISSET(mp->mp_flags, P_DIRTY)) { 2421 if (txn->mt_flags & MDB_TXN_SPILLS) { 2422 np = NULL; 2423 rc = mdb_page_unspill(txn, mp, &np); 2424 if (rc) 2425 goto fail; 2426 if (np) 2427 goto done; 2428 } 2429 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) || 2430 (rc = mdb_page_alloc(mc, 1, &np))) 2431 goto fail; 2432 pgno = np->mp_pgno; 2433 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc), 2434 mp->mp_pgno, pgno)); 2435 mdb_cassert(mc, mp->mp_pgno != pgno); 2436 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno); 2437 /* Update the parent page, if any, to point to the new page */ 2438 if (mc->mc_top) { 2439 MDB_page *parent = mc->mc_pg[mc->mc_top-1]; 2440 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]); 2441 SETPGNO(node, pgno); 2442 } else { 2443 mc->mc_db->md_root = pgno; 2444 } 2445 } else if (txn->mt_parent && !IS_SUBP(mp)) { 2446 MDB_ID2 mid, *dl = txn->mt_u.dirty_list; 2447 pgno = mp->mp_pgno; 2448 /* If txn has a parent, make sure the page is in our 2449 * dirty list. 2450 */ 2451 if (dl[0].mid) { 2452 unsigned x = mdb_mid2l_search(dl, pgno); 2453 if (x <= dl[0].mid && dl[x].mid == pgno) { 2454 if (mp != dl[x].mptr) { /* bad cursor? */ 2455 mc->mc_flags &= ~(C_INITIALIZED|C_EOF); 2456 txn->mt_flags |= MDB_TXN_ERROR; 2457 return MDB_CORRUPTED; 2458 } 2459 return 0; 2460 } 2461 } 2462 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX); 2463 /* No - copy it */ 2464 np = mdb_page_malloc(txn, 1); 2465 if (!np) 2466 return ENOMEM; 2467 mid.mid = pgno; 2468 mid.mptr = np; 2469 rc = mdb_mid2l_insert(dl, &mid); 2470 mdb_cassert(mc, rc == 0); 2471 } else { 2472 return 0; 2473 } 2474 2475 mdb_page_copy(np, mp, txn->mt_env->me_psize); 2476 np->mp_pgno = pgno; 2477 np->mp_flags |= P_DIRTY; 2478 2479 done: 2480 /* Adjust cursors pointing to mp */ 2481 mc->mc_pg[mc->mc_top] = np; 2482 m2 = txn->mt_cursors[mc->mc_dbi]; 2483 if (mc->mc_flags & C_SUB) { 2484 for (; m2; m2=m2->mc_next) { 2485 m3 = &m2->mc_xcursor->mx_cursor; 2486 if (m3->mc_snum < mc->mc_snum) continue; 2487 if (m3->mc_pg[mc->mc_top] == mp) 2488 m3->mc_pg[mc->mc_top] = np; 2489 } 2490 } else { 2491 for (; m2; m2=m2->mc_next) { 2492 if (m2->mc_snum < mc->mc_snum) continue; 2493 if (m2 == mc) continue; 2494 if (m2->mc_pg[mc->mc_top] == mp) { 2495 m2->mc_pg[mc->mc_top] = np; 2496 if (IS_LEAF(np)) 2497 XCURSOR_REFRESH(m2, mc->mc_top, np); 2498 } 2499 } 2500 } 2501 return 0; 2502 2503 fail: 2504 txn->mt_flags |= MDB_TXN_ERROR; 2505 return rc; 2506 } 2507 2508 int 2509 mdb_env_sync(MDB_env *env, int force) 2510 { 2511 int rc = 0; 2512 if (env->me_flags & MDB_RDONLY) 2513 return EACCES; 2514 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) { 2515 if (env->me_flags & MDB_WRITEMAP) { 2516 int flags = ((env->me_flags & MDB_MAPASYNC) && !force) 2517 ? MS_ASYNC : MS_SYNC; 2518 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags)) 2519 rc = ErrCode(); 2520 #ifdef _WIN32 2521 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd)) 2522 rc = ErrCode(); 2523 #endif 2524 } else { 2525 #ifdef BROKEN_FDATASYNC 2526 if (env->me_flags & MDB_FSYNCONLY) { 2527 if (fsync(env->me_fd)) 2528 rc = ErrCode(); 2529 } else 2530 #endif 2531 if (MDB_FDATASYNC(env->me_fd)) 2532 rc = ErrCode(); 2533 } 2534 } 2535 return rc; 2536 } 2537 2538 /** Back up parent txn's cursors, then grab the originals for tracking */ 2539 static int 2540 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst) 2541 { 2542 MDB_cursor *mc, *bk; 2543 MDB_xcursor *mx; 2544 size_t size; 2545 int i; 2546 2547 for (i = src->mt_numdbs; --i >= 0; ) { 2548 if ((mc = src->mt_cursors[i]) != NULL) { 2549 size = sizeof(MDB_cursor); 2550 if (mc->mc_xcursor) 2551 size += sizeof(MDB_xcursor); 2552 for (; mc; mc = bk->mc_next) { 2553 bk = malloc(size); 2554 if (!bk) 2555 return ENOMEM; 2556 *bk = *mc; 2557 mc->mc_backup = bk; 2558 mc->mc_db = &dst->mt_dbs[i]; 2559 /* Kill pointers into src to reduce abuse: The 2560 * user may not use mc until dst ends. But we need a valid 2561 * txn pointer here for cursor fixups to keep working. 2562 */ 2563 mc->mc_txn = dst; 2564 mc->mc_dbflag = &dst->mt_dbflags[i]; 2565 if ((mx = mc->mc_xcursor) != NULL) { 2566 *(MDB_xcursor *)(bk+1) = *mx; 2567 mx->mx_cursor.mc_txn = dst; 2568 } 2569 mc->mc_next = dst->mt_cursors[i]; 2570 dst->mt_cursors[i] = mc; 2571 } 2572 } 2573 } 2574 return MDB_SUCCESS; 2575 } 2576 2577 /** Close this write txn's cursors, give parent txn's cursors back to parent. 2578 * @param[in] txn the transaction handle. 2579 * @param[in] merge true to keep changes to parent cursors, false to revert. 2580 * @return 0 on success, non-zero on failure. 2581 */ 2582 static void 2583 mdb_cursors_close(MDB_txn *txn, unsigned merge) 2584 { 2585 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk; 2586 MDB_xcursor *mx; 2587 int i; 2588 2589 for (i = txn->mt_numdbs; --i >= 0; ) { 2590 for (mc = cursors[i]; mc; mc = next) { 2591 next = mc->mc_next; 2592 if ((bk = mc->mc_backup) != NULL) { 2593 if (merge) { 2594 /* Commit changes to parent txn */ 2595 mc->mc_next = bk->mc_next; 2596 mc->mc_backup = bk->mc_backup; 2597 mc->mc_txn = bk->mc_txn; 2598 mc->mc_db = bk->mc_db; 2599 mc->mc_dbflag = bk->mc_dbflag; 2600 if ((mx = mc->mc_xcursor) != NULL) 2601 mx->mx_cursor.mc_txn = bk->mc_txn; 2602 } else { 2603 /* Abort nested txn */ 2604 *mc = *bk; 2605 if ((mx = mc->mc_xcursor) != NULL) 2606 *mx = *(MDB_xcursor *)(bk+1); 2607 } 2608 mc = bk; 2609 } 2610 /* Only malloced cursors are permanently tracked. */ 2611 free(mc); 2612 } 2613 cursors[i] = NULL; 2614 } 2615 } 2616 2617 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */ 2618 enum Pidlock_op { 2619 Pidset, Pidcheck 2620 }; 2621 #else 2622 enum Pidlock_op { 2623 Pidset = F_SETLK, Pidcheck = F_GETLK 2624 }; 2625 #endif 2626 2627 /** Set or check a pid lock. Set returns 0 on success. 2628 * Check returns 0 if the process is certainly dead, nonzero if it may 2629 * be alive (the lock exists or an error happened so we do not know). 2630 * 2631 * On Windows Pidset is a no-op, we merely check for the existence 2632 * of the process with the given pid. On POSIX we use a single byte 2633 * lock on the lockfile, set at an offset equal to the pid. 2634 */ 2635 static int 2636 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid) 2637 { 2638 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */ 2639 int ret = 0; 2640 HANDLE h; 2641 if (op == Pidcheck) { 2642 h = OpenProcess(env->me_pidquery, FALSE, pid); 2643 /* No documented "no such process" code, but other program use this: */ 2644 if (!h) 2645 return ErrCode() != ERROR_INVALID_PARAMETER; 2646 /* A process exists until all handles to it close. Has it exited? */ 2647 ret = WaitForSingleObject(h, 0) != 0; 2648 CloseHandle(h); 2649 } 2650 return ret; 2651 #else 2652 for (;;) { 2653 int rc; 2654 struct flock lock_info; 2655 memset(&lock_info, 0, sizeof(lock_info)); 2656 lock_info.l_type = F_WRLCK; 2657 lock_info.l_whence = SEEK_SET; 2658 lock_info.l_start = pid; 2659 lock_info.l_len = 1; 2660 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) { 2661 if (op == F_GETLK && lock_info.l_type != F_UNLCK) 2662 rc = -1; 2663 } else if ((rc = ErrCode()) == EINTR) { 2664 continue; 2665 } 2666 return rc; 2667 } 2668 #endif 2669 } 2670 2671 /** Common code for #mdb_txn_begin() and #mdb_txn_renew(). 2672 * @param[in] txn the transaction handle to initialize 2673 * @return 0 on success, non-zero on failure. 2674 */ 2675 static int 2676 mdb_txn_renew0(MDB_txn *txn) 2677 { 2678 MDB_env *env = txn->mt_env; 2679 MDB_txninfo *ti = env->me_txns; 2680 MDB_meta *meta; 2681 unsigned int i, nr, flags = txn->mt_flags; 2682 uint16_t x; 2683 int rc, new_notls = 0; 2684 2685 if ((flags &= MDB_TXN_RDONLY) != 0) { 2686 if (!ti) { 2687 meta = mdb_env_pick_meta(env); 2688 txn->mt_txnid = meta->mm_txnid; 2689 txn->mt_u.reader = NULL; 2690 } else { 2691 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader : 2692 pthread_getspecific(env->me_txkey); 2693 if (r) { 2694 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1) 2695 return MDB_BAD_RSLOT; 2696 } else { 2697 MDB_PID_T pid = env->me_pid; 2698 MDB_THR_T tid = pthread_self(); 2699 mdb_mutexref_t rmutex = env->me_rmutex; 2700 2701 if (!env->me_live_reader) { 2702 rc = mdb_reader_pid(env, Pidset, pid); 2703 if (rc) 2704 return rc; 2705 env->me_live_reader = 1; 2706 } 2707 2708 if (LOCK_MUTEX(rc, env, rmutex)) 2709 return rc; 2710 nr = ti->mti_numreaders; 2711 for (i=0; i<nr; i++) 2712 if (ti->mti_readers[i].mr_pid == 0) 2713 break; 2714 if (i == env->me_maxreaders) { 2715 UNLOCK_MUTEX(rmutex); 2716 return MDB_READERS_FULL; 2717 } 2718 r = &ti->mti_readers[i]; 2719 /* Claim the reader slot, carefully since other code 2720 * uses the reader table un-mutexed: First reset the 2721 * slot, next publish it in mti_numreaders. After 2722 * that, it is safe for mdb_env_close() to touch it. 2723 * When it will be closed, we can finally claim it. 2724 */ 2725 r->mr_pid = 0; 2726 r->mr_txnid = (txnid_t)-1; 2727 r->mr_tid = tid; 2728 if (i == nr) 2729 ti->mti_numreaders = ++nr; 2730 env->me_close_readers = nr; 2731 r->mr_pid = pid; 2732 UNLOCK_MUTEX(rmutex); 2733 2734 new_notls = (env->me_flags & MDB_NOTLS); 2735 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) { 2736 r->mr_pid = 0; 2737 return rc; 2738 } 2739 } 2740 do /* LY: Retry on a race, ITS#7970. */ 2741 r->mr_txnid = ti->mti_txnid; 2742 while(r->mr_txnid != ti->mti_txnid); 2743 txn->mt_txnid = r->mr_txnid; 2744 txn->mt_u.reader = r; 2745 meta = env->me_metas[txn->mt_txnid & 1]; 2746 } 2747 2748 } else { 2749 /* Not yet touching txn == env->me_txn0, it may be active */ 2750 if (ti) { 2751 if (LOCK_MUTEX(rc, env, env->me_wmutex)) 2752 return rc; 2753 txn->mt_txnid = ti->mti_txnid; 2754 meta = env->me_metas[txn->mt_txnid & 1]; 2755 } else { 2756 meta = mdb_env_pick_meta(env); 2757 txn->mt_txnid = meta->mm_txnid; 2758 } 2759 txn->mt_txnid++; 2760 #if MDB_DEBUG 2761 if (txn->mt_txnid == mdb_debug_start) 2762 mdb_debug = 1; 2763 #endif 2764 txn->mt_child = NULL; 2765 txn->mt_loose_pgs = NULL; 2766 txn->mt_loose_count = 0; 2767 txn->mt_dirty_room = MDB_IDL_UM_MAX; 2768 txn->mt_u.dirty_list = env->me_dirty_list; 2769 txn->mt_u.dirty_list[0].mid = 0; 2770 txn->mt_free_pgs = env->me_free_pgs; 2771 txn->mt_free_pgs[0] = 0; 2772 txn->mt_spill_pgs = NULL; 2773 env->me_txn = txn; 2774 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int)); 2775 } 2776 2777 /* Copy the DB info and flags */ 2778 memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db)); 2779 2780 /* Moved to here to avoid a data race in read TXNs */ 2781 txn->mt_next_pgno = meta->mm_last_pg+1; 2782 2783 txn->mt_flags = flags; 2784 2785 /* Setup db info */ 2786 txn->mt_numdbs = env->me_numdbs; 2787 for (i=CORE_DBS; i<txn->mt_numdbs; i++) { 2788 x = env->me_dbflags[i]; 2789 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS; 2790 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_USRVALID|DB_STALE : 0; 2791 } 2792 txn->mt_dbflags[MAIN_DBI] = DB_VALID|DB_USRVALID; 2793 txn->mt_dbflags[FREE_DBI] = DB_VALID; 2794 2795 if (env->me_flags & MDB_FATAL_ERROR) { 2796 DPUTS("environment had fatal error, must shutdown!"); 2797 rc = MDB_PANIC; 2798 } else if (env->me_maxpg < txn->mt_next_pgno) { 2799 rc = MDB_MAP_RESIZED; 2800 } else { 2801 return MDB_SUCCESS; 2802 } 2803 mdb_txn_end(txn, new_notls /*0 or MDB_END_SLOT*/ | MDB_END_FAIL_BEGIN); 2804 return rc; 2805 } 2806 2807 int 2808 mdb_txn_renew(MDB_txn *txn) 2809 { 2810 int rc; 2811 2812 if (!txn || !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED)) 2813 return EINVAL; 2814 2815 rc = mdb_txn_renew0(txn); 2816 if (rc == MDB_SUCCESS) { 2817 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u", 2818 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', 2819 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root)); 2820 } 2821 return rc; 2822 } 2823 2824 int 2825 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret) 2826 { 2827 MDB_txn *txn; 2828 MDB_ntxn *ntxn; 2829 int rc, size, tsize; 2830 2831 flags &= MDB_TXN_BEGIN_FLAGS; 2832 flags |= env->me_flags & MDB_WRITEMAP; 2833 2834 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */ 2835 return EACCES; 2836 2837 if (parent) { 2838 /* Nested transactions: Max 1 child, write txns only, no writemap */ 2839 flags |= parent->mt_flags; 2840 if (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_BLOCKED)) { 2841 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN; 2842 } 2843 /* Child txns save MDB_pgstate and use own copy of cursors */ 2844 size = env->me_maxdbs * (sizeof(MDB_db)+sizeof(MDB_cursor *)+1); 2845 size += tsize = sizeof(MDB_ntxn); 2846 } else if (flags & MDB_RDONLY) { 2847 size = env->me_maxdbs * (sizeof(MDB_db)+1); 2848 size += tsize = sizeof(MDB_txn); 2849 } else { 2850 /* Reuse preallocated write txn. However, do not touch it until 2851 * mdb_txn_renew0() succeeds, since it currently may be active. 2852 */ 2853 txn = env->me_txn0; 2854 goto renew; 2855 } 2856 if ((txn = calloc(1, size)) == NULL) { 2857 DPRINTF(("calloc: %s", strerror(errno))); 2858 return ENOMEM; 2859 } 2860 txn->mt_dbxs = env->me_dbxs; /* static */ 2861 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize); 2862 txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs; 2863 txn->mt_flags = flags; 2864 txn->mt_env = env; 2865 2866 if (parent) { 2867 unsigned int i; 2868 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs); 2869 txn->mt_dbiseqs = parent->mt_dbiseqs; 2870 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE); 2871 if (!txn->mt_u.dirty_list || 2872 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX))) 2873 { 2874 free(txn->mt_u.dirty_list); 2875 free(txn); 2876 return ENOMEM; 2877 } 2878 txn->mt_txnid = parent->mt_txnid; 2879 txn->mt_dirty_room = parent->mt_dirty_room; 2880 txn->mt_u.dirty_list[0].mid = 0; 2881 txn->mt_spill_pgs = NULL; 2882 txn->mt_next_pgno = parent->mt_next_pgno; 2883 parent->mt_flags |= MDB_TXN_HAS_CHILD; 2884 parent->mt_child = txn; 2885 txn->mt_parent = parent; 2886 txn->mt_numdbs = parent->mt_numdbs; 2887 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db)); 2888 /* Copy parent's mt_dbflags, but clear DB_NEW */ 2889 for (i=0; i<txn->mt_numdbs; i++) 2890 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW; 2891 rc = 0; 2892 ntxn = (MDB_ntxn *)txn; 2893 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */ 2894 if (env->me_pghead) { 2895 size = MDB_IDL_SIZEOF(env->me_pghead); 2896 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]); 2897 if (env->me_pghead) 2898 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size); 2899 else 2900 rc = ENOMEM; 2901 } 2902 if (!rc) 2903 rc = mdb_cursor_shadow(parent, txn); 2904 if (rc) 2905 mdb_txn_end(txn, MDB_END_FAIL_BEGINCHILD); 2906 } else { /* MDB_RDONLY */ 2907 txn->mt_dbiseqs = env->me_dbiseqs; 2908 renew: 2909 rc = mdb_txn_renew0(txn); 2910 } 2911 if (rc) { 2912 if (txn != env->me_txn0) 2913 free(txn); 2914 } else { 2915 txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */ 2916 *ret = txn; 2917 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u", 2918 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w', 2919 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root)); 2920 } 2921 2922 return rc; 2923 } 2924 2925 MDB_env * 2926 mdb_txn_env(MDB_txn *txn) 2927 { 2928 if(!txn) return NULL; 2929 return txn->mt_env; 2930 } 2931 2932 size_t 2933 mdb_txn_id(MDB_txn *txn) 2934 { 2935 if(!txn) return 0; 2936 return txn->mt_txnid; 2937 } 2938 2939 /** Export or close DBI handles opened in this txn. */ 2940 static void 2941 mdb_dbis_update(MDB_txn *txn, int keep) 2942 { 2943 int i; 2944 MDB_dbi n = txn->mt_numdbs; 2945 MDB_env *env = txn->mt_env; 2946 unsigned char *tdbflags = txn->mt_dbflags; 2947 2948 for (i = n; --i >= CORE_DBS;) { 2949 if (tdbflags[i] & DB_NEW) { 2950 if (keep) { 2951 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID; 2952 } else { 2953 char *ptr = env->me_dbxs[i].md_name.mv_data; 2954 if (ptr) { 2955 env->me_dbxs[i].md_name.mv_data = NULL; 2956 env->me_dbxs[i].md_name.mv_size = 0; 2957 env->me_dbflags[i] = 0; 2958 env->me_dbiseqs[i]++; 2959 free(ptr); 2960 } 2961 } 2962 } 2963 } 2964 if (keep && env->me_numdbs < n) 2965 env->me_numdbs = n; 2966 } 2967 2968 /** End a transaction, except successful commit of a nested transaction. 2969 * May be called twice for readonly txns: First reset it, then abort. 2970 * @param[in] txn the transaction handle to end 2971 * @param[in] mode why and how to end the transaction 2972 */ 2973 static void 2974 mdb_txn_end(MDB_txn *txn, unsigned mode) 2975 { 2976 MDB_env *env = txn->mt_env; 2977 #if MDB_DEBUG 2978 static const char *const names[] = MDB_END_NAMES; 2979 #endif 2980 2981 /* Export or close DBI handles opened in this txn */ 2982 mdb_dbis_update(txn, mode & MDB_END_UPDATE); 2983 2984 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u", 2985 names[mode & MDB_END_OPMASK], 2986 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', 2987 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root)); 2988 2989 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) { 2990 if (txn->mt_u.reader) { 2991 txn->mt_u.reader->mr_txnid = (txnid_t)-1; 2992 if (!(env->me_flags & MDB_NOTLS)) { 2993 txn->mt_u.reader = NULL; /* txn does not own reader */ 2994 } else if (mode & MDB_END_SLOT) { 2995 txn->mt_u.reader->mr_pid = 0; 2996 txn->mt_u.reader = NULL; 2997 } /* else txn owns the slot until it does MDB_END_SLOT */ 2998 } 2999 txn->mt_numdbs = 0; /* prevent further DBI activity */ 3000 txn->mt_flags |= MDB_TXN_FINISHED; 3001 3002 } else if (!F_ISSET(txn->mt_flags, MDB_TXN_FINISHED)) { 3003 pgno_t *pghead = env->me_pghead; 3004 3005 if (!(mode & MDB_END_UPDATE)) /* !(already closed cursors) */ 3006 mdb_cursors_close(txn, 0); 3007 if (!(env->me_flags & MDB_WRITEMAP)) { 3008 mdb_dlist_free(txn); 3009 } 3010 3011 txn->mt_numdbs = 0; 3012 txn->mt_flags = MDB_TXN_FINISHED; 3013 3014 if (!txn->mt_parent) { 3015 mdb_midl_shrink(&txn->mt_free_pgs); 3016 env->me_free_pgs = txn->mt_free_pgs; 3017 /* me_pgstate: */ 3018 env->me_pghead = NULL; 3019 env->me_pglast = 0; 3020 3021 env->me_txn = NULL; 3022 mode = 0; /* txn == env->me_txn0, do not free() it */ 3023 3024 /* The writer mutex was locked in mdb_txn_begin. */ 3025 if (env->me_txns) 3026 UNLOCK_MUTEX(env->me_wmutex); 3027 } else { 3028 txn->mt_parent->mt_child = NULL; 3029 txn->mt_parent->mt_flags &= ~MDB_TXN_HAS_CHILD; 3030 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate; 3031 mdb_midl_free(txn->mt_free_pgs); 3032 free(txn->mt_u.dirty_list); 3033 } 3034 mdb_midl_free(txn->mt_spill_pgs); 3035 3036 mdb_midl_free(pghead); 3037 } 3038 3039 if (mode & MDB_END_FREE) 3040 free(txn); 3041 } 3042 3043 void 3044 mdb_txn_reset(MDB_txn *txn) 3045 { 3046 if (txn == NULL) 3047 return; 3048 3049 /* This call is only valid for read-only txns */ 3050 if (!(txn->mt_flags & MDB_TXN_RDONLY)) 3051 return; 3052 3053 mdb_txn_end(txn, MDB_END_RESET); 3054 } 3055 3056 void 3057 mdb_txn_abort(MDB_txn *txn) 3058 { 3059 if (txn == NULL) 3060 return; 3061 3062 if (txn->mt_child) 3063 mdb_txn_abort(txn->mt_child); 3064 3065 mdb_txn_end(txn, MDB_END_ABORT|MDB_END_SLOT|MDB_END_FREE); 3066 } 3067 3068 /** Save the freelist as of this transaction to the freeDB. 3069 * This changes the freelist. Keep trying until it stabilizes. 3070 */ 3071 static int 3072 mdb_freelist_save(MDB_txn *txn) 3073 { 3074 /* env->me_pghead[] can grow and shrink during this call. 3075 * env->me_pglast and txn->mt_free_pgs[] can only grow. 3076 * Page numbers cannot disappear from txn->mt_free_pgs[]. 3077 */ 3078 MDB_cursor mc; 3079 MDB_env *env = txn->mt_env; 3080 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1; 3081 txnid_t pglast = 0, head_id = 0; 3082 pgno_t freecnt = 0, *free_pgs, *mop; 3083 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit; 3084 3085 mdb_cursor_init(&mc, txn, FREE_DBI, NULL); 3086 3087 if (env->me_pghead) { 3088 /* Make sure first page of freeDB is touched and on freelist */ 3089 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY); 3090 if (rc && rc != MDB_NOTFOUND) 3091 return rc; 3092 } 3093 3094 if (!env->me_pghead && txn->mt_loose_pgs) { 3095 /* Put loose page numbers in mt_free_pgs, since 3096 * we may be unable to return them to me_pghead. 3097 */ 3098 MDB_page *mp = txn->mt_loose_pgs; 3099 MDB_ID2 *dl = txn->mt_u.dirty_list; 3100 unsigned x; 3101 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0) 3102 return rc; 3103 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) { 3104 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno); 3105 /* must also remove from dirty list */ 3106 if (txn->mt_flags & MDB_TXN_WRITEMAP) { 3107 for (x=1; x<=dl[0].mid; x++) 3108 if (dl[x].mid == mp->mp_pgno) 3109 break; 3110 mdb_tassert(txn, x <= dl[0].mid); 3111 } else { 3112 x = mdb_mid2l_search(dl, mp->mp_pgno); 3113 mdb_tassert(txn, dl[x].mid == mp->mp_pgno); 3114 mdb_dpage_free(env, mp); 3115 } 3116 dl[x].mptr = NULL; 3117 } 3118 { 3119 /* squash freed slots out of the dirty list */ 3120 unsigned y; 3121 for (y=1; dl[y].mptr && y <= dl[0].mid; y++); 3122 if (y <= dl[0].mid) { 3123 for(x=y, y++;;) { 3124 while (!dl[y].mptr && y <= dl[0].mid) y++; 3125 if (y > dl[0].mid) break; 3126 dl[x++] = dl[y++]; 3127 } 3128 dl[0].mid = x-1; 3129 } else { 3130 /* all slots freed */ 3131 dl[0].mid = 0; 3132 } 3133 } 3134 txn->mt_loose_pgs = NULL; 3135 txn->mt_loose_count = 0; 3136 } 3137 3138 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */ 3139 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP)) 3140 ? SSIZE_MAX : maxfree_1pg; 3141 3142 for (;;) { 3143 /* Come back here after each Put() in case freelist changed */ 3144 MDB_val key, data; 3145 pgno_t *pgs; 3146 ssize_t j; 3147 3148 /* If using records from freeDB which we have not yet 3149 * deleted, delete them and any we reserved for me_pghead. 3150 */ 3151 while (pglast < env->me_pglast) { 3152 rc = mdb_cursor_first(&mc, &key, NULL); 3153 if (rc) 3154 return rc; 3155 pglast = head_id = *(txnid_t *)key.mv_data; 3156 total_room = head_room = 0; 3157 mdb_tassert(txn, pglast <= env->me_pglast); 3158 rc = mdb_cursor_del(&mc, 0); 3159 if (rc) 3160 return rc; 3161 } 3162 3163 /* Save the IDL of pages freed by this txn, to a single record */ 3164 if (freecnt < txn->mt_free_pgs[0]) { 3165 if (!freecnt) { 3166 /* Make sure last page of freeDB is touched and on freelist */ 3167 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY); 3168 if (rc && rc != MDB_NOTFOUND) 3169 return rc; 3170 } 3171 free_pgs = txn->mt_free_pgs; 3172 /* Write to last page of freeDB */ 3173 key.mv_size = sizeof(txn->mt_txnid); 3174 key.mv_data = &txn->mt_txnid; 3175 do { 3176 freecnt = free_pgs[0]; 3177 data.mv_size = MDB_IDL_SIZEOF(free_pgs); 3178 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE); 3179 if (rc) 3180 return rc; 3181 /* Retry if mt_free_pgs[] grew during the Put() */ 3182 free_pgs = txn->mt_free_pgs; 3183 } while (freecnt < free_pgs[0]); 3184 mdb_midl_sort(free_pgs); 3185 memcpy(data.mv_data, free_pgs, data.mv_size); 3186 #if (MDB_DEBUG) > 1 3187 { 3188 unsigned int i = free_pgs[0]; 3189 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u", 3190 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i)); 3191 for (; i; i--) 3192 DPRINTF(("IDL %"Z"u", free_pgs[i])); 3193 } 3194 #endif 3195 continue; 3196 } 3197 3198 mop = env->me_pghead; 3199 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count; 3200 3201 /* Reserve records for me_pghead[]. Split it if multi-page, 3202 * to avoid searching freeDB for a page range. Use keys in 3203 * range [1,me_pglast]: Smaller than txnid of oldest reader. 3204 */ 3205 if (total_room >= mop_len) { 3206 if (total_room == mop_len || --more < 0) 3207 break; 3208 } else if (head_room >= maxfree_1pg && head_id > 1) { 3209 /* Keep current record (overflow page), add a new one */ 3210 head_id--; 3211 head_room = 0; 3212 } 3213 /* (Re)write {key = head_id, IDL length = head_room} */ 3214 total_room -= head_room; 3215 head_room = mop_len - total_room; 3216 if (head_room > maxfree_1pg && head_id > 1) { 3217 /* Overflow multi-page for part of me_pghead */ 3218 head_room /= head_id; /* amortize page sizes */ 3219 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1); 3220 } else if (head_room < 0) { 3221 /* Rare case, not bothering to delete this record */ 3222 head_room = 0; 3223 } 3224 key.mv_size = sizeof(head_id); 3225 key.mv_data = &head_id; 3226 data.mv_size = (head_room + 1) * sizeof(pgno_t); 3227 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE); 3228 if (rc) 3229 return rc; 3230 /* IDL is initially empty, zero out at least the length */ 3231 pgs = (pgno_t *)data.mv_data; 3232 j = head_room > clean_limit ? head_room : 0; 3233 do { 3234 pgs[j] = 0; 3235 } while (--j >= 0); 3236 total_room += head_room; 3237 } 3238 3239 /* Return loose page numbers to me_pghead, though usually none are 3240 * left at this point. The pages themselves remain in dirty_list. 3241 */ 3242 if (txn->mt_loose_pgs) { 3243 MDB_page *mp = txn->mt_loose_pgs; 3244 unsigned count = txn->mt_loose_count; 3245 MDB_IDL loose; 3246 /* Room for loose pages + temp IDL with same */ 3247 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0) 3248 return rc; 3249 mop = env->me_pghead; 3250 loose = mop + MDB_IDL_ALLOCLEN(mop) - count; 3251 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp)) 3252 loose[ ++count ] = mp->mp_pgno; 3253 loose[0] = count; 3254 mdb_midl_sort(loose); 3255 mdb_midl_xmerge(mop, loose); 3256 txn->mt_loose_pgs = NULL; 3257 txn->mt_loose_count = 0; 3258 mop_len = mop[0]; 3259 } 3260 3261 /* Fill in the reserved me_pghead records */ 3262 rc = MDB_SUCCESS; 3263 if (mop_len) { 3264 MDB_val key, data; 3265 3266 mop += mop_len; 3267 rc = mdb_cursor_first(&mc, &key, &data); 3268 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) { 3269 txnid_t id = *(txnid_t *)key.mv_data; 3270 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1; 3271 MDB_ID save; 3272 3273 mdb_tassert(txn, len >= 0 && id <= env->me_pglast); 3274 key.mv_data = &id; 3275 if (len > mop_len) { 3276 len = mop_len; 3277 data.mv_size = (len + 1) * sizeof(MDB_ID); 3278 } 3279 data.mv_data = mop -= len; 3280 save = mop[0]; 3281 mop[0] = len; 3282 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT); 3283 mop[0] = save; 3284 if (rc || !(mop_len -= len)) 3285 break; 3286 } 3287 } 3288 return rc; 3289 } 3290 3291 /** Flush (some) dirty pages to the map, after clearing their dirty flag. 3292 * @param[in] txn the transaction that's being committed 3293 * @param[in] keep number of initial pages in dirty_list to keep dirty. 3294 * @return 0 on success, non-zero on failure. 3295 */ 3296 static int 3297 mdb_page_flush(MDB_txn *txn, int keep) 3298 { 3299 MDB_env *env = txn->mt_env; 3300 MDB_ID2L dl = txn->mt_u.dirty_list; 3301 unsigned psize = env->me_psize, j; 3302 int i, pagecount = dl[0].mid, rc; 3303 size_t size = 0, pos = 0; 3304 pgno_t pgno = 0; 3305 MDB_page *dp = NULL; 3306 #ifdef _WIN32 3307 OVERLAPPED ov; 3308 #else 3309 struct iovec iov[MDB_COMMIT_PAGES]; 3310 ssize_t wpos = 0, wsize = 0, wres; 3311 size_t next_pos = 1; /* impossible pos, so pos != next_pos */ 3312 int n = 0; 3313 #endif 3314 3315 j = i = keep; 3316 3317 if (env->me_flags & MDB_WRITEMAP) { 3318 /* Clear dirty flags */ 3319 while (++i <= pagecount) { 3320 dp = dl[i].mptr; 3321 /* Don't flush this page yet */ 3322 if (dp->mp_flags & (P_LOOSE|P_KEEP)) { 3323 dp->mp_flags &= ~P_KEEP; 3324 dl[++j] = dl[i]; 3325 continue; 3326 } 3327 dp->mp_flags &= ~P_DIRTY; 3328 } 3329 goto done; 3330 } 3331 3332 /* Write the pages */ 3333 for (;;) { 3334 if (++i <= pagecount) { 3335 dp = dl[i].mptr; 3336 /* Don't flush this page yet */ 3337 if (dp->mp_flags & (P_LOOSE|P_KEEP)) { 3338 dp->mp_flags &= ~P_KEEP; 3339 dl[i].mid = 0; 3340 continue; 3341 } 3342 pgno = dl[i].mid; 3343 /* clear dirty flag */ 3344 dp->mp_flags &= ~P_DIRTY; 3345 pos = pgno * psize; 3346 size = psize; 3347 if (IS_OVERFLOW(dp)) size *= dp->mp_pages; 3348 } 3349 #ifdef _WIN32 3350 else break; 3351 3352 /* Windows actually supports scatter/gather I/O, but only on 3353 * unbuffered file handles. Since we're relying on the OS page 3354 * cache for all our data, that's self-defeating. So we just 3355 * write pages one at a time. We use the ov structure to set 3356 * the write offset, to at least save the overhead of a Seek 3357 * system call. 3358 */ 3359 DPRINTF(("committing page %"Z"u", pgno)); 3360 memset(&ov, 0, sizeof(ov)); 3361 ov.Offset = pos & 0xffffffff; 3362 ov.OffsetHigh = pos >> 16 >> 16; 3363 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) { 3364 rc = ErrCode(); 3365 DPRINTF(("WriteFile: %d", rc)); 3366 return rc; 3367 } 3368 #else 3369 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */ 3370 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) { 3371 if (n) { 3372 retry_write: 3373 /* Write previous page(s) */ 3374 #ifdef MDB_USE_PWRITEV 3375 wres = pwritev(env->me_fd, iov, n, wpos); 3376 #else 3377 if (n == 1) { 3378 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos); 3379 } else { 3380 retry_seek: 3381 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) { 3382 rc = ErrCode(); 3383 if (rc == EINTR) 3384 goto retry_seek; 3385 DPRINTF(("lseek: %s", strerror(rc))); 3386 return rc; 3387 } 3388 wres = writev(env->me_fd, iov, n); 3389 } 3390 #endif 3391 if (wres != wsize) { 3392 if (wres < 0) { 3393 rc = ErrCode(); 3394 if (rc == EINTR) 3395 goto retry_write; 3396 DPRINTF(("Write error: %s", strerror(rc))); 3397 } else { 3398 rc = EIO; /* TODO: Use which error code? */ 3399 DPUTS("short write, filesystem full?"); 3400 } 3401 return rc; 3402 } 3403 n = 0; 3404 } 3405 if (i > pagecount) 3406 break; 3407 wpos = pos; 3408 wsize = 0; 3409 } 3410 DPRINTF(("committing page %"Z"u", pgno)); 3411 next_pos = pos + size; 3412 iov[n].iov_len = size; 3413 iov[n].iov_base = (char *)dp; 3414 wsize += size; 3415 n++; 3416 #endif /* _WIN32 */ 3417 } 3418 3419 /* MIPS has cache coherency issues, this is a no-op everywhere else 3420 * Note: for any size >= on-chip cache size, entire on-chip cache is 3421 * flushed. 3422 */ 3423 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE); 3424 3425 for (i = keep; ++i <= pagecount; ) { 3426 dp = dl[i].mptr; 3427 /* This is a page we skipped above */ 3428 if (!dl[i].mid) { 3429 dl[++j] = dl[i]; 3430 dl[j].mid = dp->mp_pgno; 3431 continue; 3432 } 3433 mdb_dpage_free(env, dp); 3434 } 3435 3436 done: 3437 i--; 3438 txn->mt_dirty_room += i - j; 3439 dl[0].mid = j; 3440 return MDB_SUCCESS; 3441 } 3442 3443 int 3444 mdb_txn_commit(MDB_txn *txn) 3445 { 3446 int rc; 3447 unsigned int i, end_mode; 3448 MDB_env *env; 3449 3450 if (txn == NULL) 3451 return EINVAL; 3452 3453 /* mdb_txn_end() mode for a commit which writes nothing */ 3454 end_mode = MDB_END_EMPTY_COMMIT|MDB_END_UPDATE|MDB_END_SLOT|MDB_END_FREE; 3455 3456 if (txn->mt_child) { 3457 rc = mdb_txn_commit(txn->mt_child); 3458 if (rc) 3459 goto fail; 3460 } 3461 3462 env = txn->mt_env; 3463 3464 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) { 3465 goto done; 3466 } 3467 3468 if (txn->mt_flags & (MDB_TXN_FINISHED|MDB_TXN_ERROR)) { 3469 DPUTS("txn has failed/finished, can't commit"); 3470 if (txn->mt_parent) 3471 txn->mt_parent->mt_flags |= MDB_TXN_ERROR; 3472 rc = MDB_BAD_TXN; 3473 goto fail; 3474 } 3475 3476 if (txn->mt_parent) { 3477 MDB_txn *parent = txn->mt_parent; 3478 MDB_page **lp; 3479 MDB_ID2L dst, src; 3480 MDB_IDL pspill; 3481 unsigned x, y, len, ps_len; 3482 3483 /* Append our free list to parent's */ 3484 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs); 3485 if (rc) 3486 goto fail; 3487 mdb_midl_free(txn->mt_free_pgs); 3488 /* Failures after this must either undo the changes 3489 * to the parent or set MDB_TXN_ERROR in the parent. 3490 */ 3491 3492 parent->mt_next_pgno = txn->mt_next_pgno; 3493 parent->mt_flags = txn->mt_flags; 3494 3495 /* Merge our cursors into parent's and close them */ 3496 mdb_cursors_close(txn, 1); 3497 3498 /* Update parent's DB table. */ 3499 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db)); 3500 parent->mt_numdbs = txn->mt_numdbs; 3501 parent->mt_dbflags[FREE_DBI] = txn->mt_dbflags[FREE_DBI]; 3502 parent->mt_dbflags[MAIN_DBI] = txn->mt_dbflags[MAIN_DBI]; 3503 for (i=CORE_DBS; i<txn->mt_numdbs; i++) { 3504 /* preserve parent's DB_NEW status */ 3505 x = parent->mt_dbflags[i] & DB_NEW; 3506 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x; 3507 } 3508 3509 dst = parent->mt_u.dirty_list; 3510 src = txn->mt_u.dirty_list; 3511 /* Remove anything in our dirty list from parent's spill list */ 3512 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) { 3513 x = y = ps_len; 3514 pspill[0] = (pgno_t)-1; 3515 /* Mark our dirty pages as deleted in parent spill list */ 3516 for (i=0, len=src[0].mid; ++i <= len; ) { 3517 MDB_ID pn = src[i].mid << 1; 3518 while (pn > pspill[x]) 3519 x--; 3520 if (pn == pspill[x]) { 3521 pspill[x] = 1; 3522 y = --x; 3523 } 3524 } 3525 /* Squash deleted pagenums if we deleted any */ 3526 for (x=y; ++x <= ps_len; ) 3527 if (!(pspill[x] & 1)) 3528 pspill[++y] = pspill[x]; 3529 pspill[0] = y; 3530 } 3531 3532 /* Remove anything in our spill list from parent's dirty list */ 3533 if (txn->mt_spill_pgs && txn->mt_spill_pgs[0]) { 3534 for (i=1; i<=txn->mt_spill_pgs[0]; i++) { 3535 MDB_ID pn = txn->mt_spill_pgs[i]; 3536 if (pn & 1) 3537 continue; /* deleted spillpg */ 3538 pn >>= 1; 3539 y = mdb_mid2l_search(dst, pn); 3540 if (y <= dst[0].mid && dst[y].mid == pn) { 3541 free(dst[y].mptr); 3542 while (y < dst[0].mid) { 3543 dst[y] = dst[y+1]; 3544 y++; 3545 } 3546 dst[0].mid--; 3547 } 3548 } 3549 } 3550 3551 /* Find len = length of merging our dirty list with parent's */ 3552 x = dst[0].mid; 3553 dst[0].mid = 0; /* simplify loops */ 3554 if (parent->mt_parent) { 3555 len = x + src[0].mid; 3556 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1; 3557 for (i = x; y && i; y--) { 3558 pgno_t yp = src[y].mid; 3559 while (yp < dst[i].mid) 3560 i--; 3561 if (yp == dst[i].mid) { 3562 i--; 3563 len--; 3564 } 3565 } 3566 } else { /* Simplify the above for single-ancestor case */ 3567 len = MDB_IDL_UM_MAX - txn->mt_dirty_room; 3568 } 3569 /* Merge our dirty list with parent's */ 3570 y = src[0].mid; 3571 for (i = len; y; dst[i--] = src[y--]) { 3572 pgno_t yp = src[y].mid; 3573 while (yp < dst[x].mid) 3574 dst[i--] = dst[x--]; 3575 if (yp == dst[x].mid) 3576 free(dst[x--].mptr); 3577 } 3578 mdb_tassert(txn, i == x); 3579 dst[0].mid = len; 3580 free(txn->mt_u.dirty_list); 3581 parent->mt_dirty_room = txn->mt_dirty_room; 3582 if (txn->mt_spill_pgs) { 3583 if (parent->mt_spill_pgs) { 3584 /* TODO: Prevent failure here, so parent does not fail */ 3585 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs); 3586 if (rc) 3587 parent->mt_flags |= MDB_TXN_ERROR; 3588 mdb_midl_free(txn->mt_spill_pgs); 3589 mdb_midl_sort(parent->mt_spill_pgs); 3590 } else { 3591 parent->mt_spill_pgs = txn->mt_spill_pgs; 3592 } 3593 } 3594 3595 /* Append our loose page list to parent's */ 3596 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(*lp)) 3597 ; 3598 *lp = txn->mt_loose_pgs; 3599 parent->mt_loose_count += txn->mt_loose_count; 3600 3601 parent->mt_child = NULL; 3602 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead); 3603 free(txn); 3604 return rc; 3605 } 3606 3607 if (txn != env->me_txn) { 3608 DPUTS("attempt to commit unknown transaction"); 3609 rc = EINVAL; 3610 goto fail; 3611 } 3612 3613 mdb_cursors_close(txn, 0); 3614 3615 if (!txn->mt_u.dirty_list[0].mid && 3616 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS))) 3617 goto done; 3618 3619 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u", 3620 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root)); 3621 3622 /* Update DB root pointers */ 3623 if (txn->mt_numdbs > CORE_DBS) { 3624 MDB_cursor mc; 3625 MDB_dbi i; 3626 MDB_val data; 3627 data.mv_size = sizeof(MDB_db); 3628 3629 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL); 3630 for (i = CORE_DBS; i < txn->mt_numdbs; i++) { 3631 if (txn->mt_dbflags[i] & DB_DIRTY) { 3632 if (TXN_DBI_CHANGED(txn, i)) { 3633 rc = MDB_BAD_DBI; 3634 goto fail; 3635 } 3636 data.mv_data = &txn->mt_dbs[i]; 3637 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 3638 F_SUBDATA); 3639 if (rc) 3640 goto fail; 3641 } 3642 } 3643 } 3644 3645 rc = mdb_freelist_save(txn); 3646 if (rc) 3647 goto fail; 3648 3649 mdb_midl_free(env->me_pghead); 3650 env->me_pghead = NULL; 3651 mdb_midl_shrink(&txn->mt_free_pgs); 3652 3653 #if (MDB_DEBUG) > 2 3654 mdb_audit(txn); 3655 #endif 3656 3657 if ((rc = mdb_page_flush(txn, 0)) || 3658 (rc = mdb_env_sync(env, 0)) || 3659 (rc = mdb_env_write_meta(txn))) 3660 goto fail; 3661 end_mode = MDB_END_COMMITTED|MDB_END_UPDATE; 3662 3663 done: 3664 mdb_txn_end(txn, end_mode); 3665 return MDB_SUCCESS; 3666 3667 fail: 3668 mdb_txn_abort(txn); 3669 return rc; 3670 } 3671 3672 /** Read the environment parameters of a DB environment before 3673 * mapping it into memory. 3674 * @param[in] env the environment handle 3675 * @param[out] meta address of where to store the meta information 3676 * @return 0 on success, non-zero on failure. 3677 */ 3678 static int ESECT 3679 mdb_env_read_header(MDB_env *env, MDB_meta *meta) 3680 { 3681 MDB_metabuf pbuf; 3682 MDB_page *p; 3683 MDB_meta *m; 3684 int i, rc, off; 3685 enum { Size = sizeof(pbuf) }; 3686 3687 /* We don't know the page size yet, so use a minimum value. 3688 * Read both meta pages so we can use the latest one. 3689 */ 3690 3691 for (i=off=0; i<NUM_METAS; i++, off += meta->mm_psize) { 3692 #ifdef _WIN32 3693 DWORD len; 3694 OVERLAPPED ov; 3695 memset(&ov, 0, sizeof(ov)); 3696 ov.Offset = off; 3697 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1; 3698 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF) 3699 rc = 0; 3700 #else 3701 rc = pread(env->me_fd, &pbuf, Size, off); 3702 #endif 3703 if (rc != Size) { 3704 if (rc == 0 && off == 0) 3705 return ENOENT; 3706 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID; 3707 DPRINTF(("read: %s", mdb_strerror(rc))); 3708 return rc; 3709 } 3710 3711 p = (MDB_page *)&pbuf; 3712 3713 if (!F_ISSET(p->mp_flags, P_META)) { 3714 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno)); 3715 return MDB_INVALID; 3716 } 3717 3718 m = METADATA(p); 3719 if (m->mm_magic != MDB_MAGIC) { 3720 DPUTS("meta has invalid magic"); 3721 return MDB_INVALID; 3722 } 3723 3724 if (m->mm_version != MDB_DATA_VERSION) { 3725 DPRINTF(("database is version %u, expected version %u", 3726 m->mm_version, MDB_DATA_VERSION)); 3727 return MDB_VERSION_MISMATCH; 3728 } 3729 3730 if (off == 0 || m->mm_txnid > meta->mm_txnid) 3731 *meta = *m; 3732 } 3733 return 0; 3734 } 3735 3736 /** Fill in most of the zeroed #MDB_meta for an empty database environment */ 3737 static void ESECT 3738 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta) 3739 { 3740 meta->mm_magic = MDB_MAGIC; 3741 meta->mm_version = MDB_DATA_VERSION; 3742 meta->mm_mapsize = env->me_mapsize; 3743 meta->mm_psize = env->me_psize; 3744 meta->mm_last_pg = NUM_METAS-1; 3745 meta->mm_flags = env->me_flags & 0xffff; 3746 meta->mm_flags |= MDB_INTEGERKEY; /* this is mm_dbs[FREE_DBI].md_flags */ 3747 meta->mm_dbs[FREE_DBI].md_root = P_INVALID; 3748 meta->mm_dbs[MAIN_DBI].md_root = P_INVALID; 3749 } 3750 3751 /** Write the environment parameters of a freshly created DB environment. 3752 * @param[in] env the environment handle 3753 * @param[in] meta the #MDB_meta to write 3754 * @return 0 on success, non-zero on failure. 3755 */ 3756 static int ESECT 3757 mdb_env_init_meta(MDB_env *env, MDB_meta *meta) 3758 { 3759 MDB_page *p, *q; 3760 int rc; 3761 unsigned int psize; 3762 #ifdef _WIN32 3763 DWORD len; 3764 OVERLAPPED ov; 3765 memset(&ov, 0, sizeof(ov)); 3766 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \ 3767 ov.Offset = pos; \ 3768 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0) 3769 #else 3770 int len; 3771 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \ 3772 len = pwrite(fd, ptr, size, pos); \ 3773 if (len == -1 && ErrCode() == EINTR) continue; \ 3774 rc = (len >= 0); break; } while(1) 3775 #endif 3776 3777 DPUTS("writing new meta page"); 3778 3779 psize = env->me_psize; 3780 3781 p = calloc(NUM_METAS, psize); 3782 if (!p) 3783 return ENOMEM; 3784 3785 p->mp_pgno = 0; 3786 p->mp_flags = P_META; 3787 *(MDB_meta *)METADATA(p) = *meta; 3788 3789 q = (MDB_page *)((char *)p + psize); 3790 q->mp_pgno = 1; 3791 q->mp_flags = P_META; 3792 *(MDB_meta *)METADATA(q) = *meta; 3793 3794 DO_PWRITE(rc, env->me_fd, p, psize * NUM_METAS, len, 0); 3795 if (!rc) 3796 rc = ErrCode(); 3797 else if ((unsigned) len == psize * NUM_METAS) 3798 rc = MDB_SUCCESS; 3799 else 3800 rc = ENOSPC; 3801 free(p); 3802 return rc; 3803 } 3804 3805 /** Update the environment info to commit a transaction. 3806 * @param[in] txn the transaction that's being committed 3807 * @return 0 on success, non-zero on failure. 3808 */ 3809 static int 3810 mdb_env_write_meta(MDB_txn *txn) 3811 { 3812 MDB_env *env; 3813 MDB_meta meta, metab, *mp; 3814 unsigned flags; 3815 size_t mapsize; 3816 off_t off; 3817 int rc, len, toggle; 3818 char *ptr; 3819 HANDLE mfd; 3820 #ifdef _WIN32 3821 OVERLAPPED ov; 3822 #else 3823 int r2; 3824 #endif 3825 3826 toggle = txn->mt_txnid & 1; 3827 DPRINTF(("writing meta page %d for root page %"Z"u", 3828 toggle, txn->mt_dbs[MAIN_DBI].md_root)); 3829 3830 env = txn->mt_env; 3831 flags = env->me_flags; 3832 mp = env->me_metas[toggle]; 3833 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize; 3834 /* Persist any increases of mapsize config */ 3835 if (mapsize < env->me_mapsize) 3836 mapsize = env->me_mapsize; 3837 3838 if (flags & MDB_WRITEMAP) { 3839 mp->mm_mapsize = mapsize; 3840 mp->mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI]; 3841 mp->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI]; 3842 mp->mm_last_pg = txn->mt_next_pgno - 1; 3843 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \ 3844 !(defined(__i386__) || defined(__x86_64__)) 3845 /* LY: issue a memory barrier, if not x86. ITS#7969 */ 3846 __sync_synchronize(); 3847 #endif 3848 mp->mm_txnid = txn->mt_txnid; 3849 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) { 3850 unsigned meta_size = env->me_psize; 3851 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC; 3852 ptr = (char *)mp - PAGEHDRSZ; 3853 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */ 3854 r2 = (ptr - env->me_map) & (env->me_os_psize - 1); 3855 ptr -= r2; 3856 meta_size += r2; 3857 #endif 3858 if (MDB_MSYNC(ptr, meta_size, rc)) { 3859 rc = ErrCode(); 3860 goto fail; 3861 } 3862 } 3863 goto done; 3864 } 3865 metab.mm_txnid = mp->mm_txnid; 3866 metab.mm_last_pg = mp->mm_last_pg; 3867 3868 meta.mm_mapsize = mapsize; 3869 meta.mm_dbs[FREE_DBI] = txn->mt_dbs[FREE_DBI]; 3870 meta.mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI]; 3871 meta.mm_last_pg = txn->mt_next_pgno - 1; 3872 meta.mm_txnid = txn->mt_txnid; 3873 3874 off = offsetof(MDB_meta, mm_mapsize); 3875 ptr = (char *)&meta + off; 3876 len = sizeof(MDB_meta) - off; 3877 off += (char *)mp - env->me_map; 3878 3879 /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC. 3880 * (me_mfd goes to the same file as me_fd, but writing to it 3881 * also syncs to disk. Avoids a separate fdatasync() call.) 3882 */ 3883 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd; 3884 #ifdef _WIN32 3885 { 3886 memset(&ov, 0, sizeof(ov)); 3887 ov.Offset = off; 3888 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov)) 3889 rc = -1; 3890 } 3891 #else 3892 retry_write: 3893 rc = pwrite(mfd, ptr, len, off); 3894 #endif 3895 if (rc != len) { 3896 rc = rc < 0 ? ErrCode() : EIO; 3897 #ifndef _WIN32 3898 if (rc == EINTR) 3899 goto retry_write; 3900 #endif 3901 DPUTS("write failed, disk error?"); 3902 /* On a failure, the pagecache still contains the new data. 3903 * Write some old data back, to prevent it from being used. 3904 * Use the non-SYNC fd; we know it will fail anyway. 3905 */ 3906 meta.mm_last_pg = metab.mm_last_pg; 3907 meta.mm_txnid = metab.mm_txnid; 3908 #ifdef _WIN32 3909 memset(&ov, 0, sizeof(ov)); 3910 ov.Offset = off; 3911 WriteFile(env->me_fd, ptr, len, NULL, &ov); 3912 #else 3913 r2 = pwrite(env->me_fd, ptr, len, off); 3914 (void)r2; /* Silence warnings. We don't care about pwrite's return value */ 3915 #endif 3916 fail: 3917 env->me_flags |= MDB_FATAL_ERROR; 3918 return rc; 3919 } 3920 /* MIPS has cache coherency issues, this is a no-op everywhere else */ 3921 CACHEFLUSH(env->me_map + off, len, DCACHE); 3922 done: 3923 /* Memory ordering issues are irrelevant; since the entire writer 3924 * is wrapped by wmutex, all of these changes will become visible 3925 * after the wmutex is unlocked. Since the DB is multi-version, 3926 * readers will get consistent data regardless of how fresh or 3927 * how stale their view of these values is. 3928 */ 3929 if (env->me_txns) 3930 env->me_txns->mti_txnid = txn->mt_txnid; 3931 3932 return MDB_SUCCESS; 3933 } 3934 3935 /** Check both meta pages to see which one is newer. 3936 * @param[in] env the environment handle 3937 * @return newest #MDB_meta. 3938 */ 3939 static MDB_meta * 3940 mdb_env_pick_meta(const MDB_env *env) 3941 { 3942 MDB_meta *const *metas = env->me_metas; 3943 return metas[ metas[0]->mm_txnid < metas[1]->mm_txnid ]; 3944 } 3945 3946 int ESECT 3947 mdb_env_create(MDB_env **env) 3948 { 3949 MDB_env *e; 3950 3951 e = calloc(1, sizeof(MDB_env)); 3952 if (!e) 3953 return ENOMEM; 3954 3955 e->me_maxreaders = DEFAULT_READERS; 3956 e->me_maxdbs = e->me_numdbs = CORE_DBS; 3957 e->me_fd = INVALID_HANDLE_VALUE; 3958 e->me_lfd = INVALID_HANDLE_VALUE; 3959 e->me_mfd = INVALID_HANDLE_VALUE; 3960 #ifdef MDB_USE_POSIX_SEM 3961 e->me_rmutex = SEM_FAILED; 3962 e->me_wmutex = SEM_FAILED; 3963 #endif 3964 e->me_pid = getpid(); 3965 GET_PAGESIZE(e->me_os_psize); 3966 VGMEMP_CREATE(e,0,0); 3967 *env = e; 3968 return MDB_SUCCESS; 3969 } 3970 3971 static int ESECT 3972 mdb_env_map(MDB_env *env, void *addr) 3973 { 3974 MDB_page *p; 3975 unsigned int flags = env->me_flags; 3976 #ifdef _WIN32 3977 int rc; 3978 HANDLE mh; 3979 LONG sizelo, sizehi; 3980 size_t msize; 3981 3982 if (flags & MDB_RDONLY) { 3983 /* Don't set explicit map size, use whatever exists */ 3984 msize = 0; 3985 sizelo = 0; 3986 sizehi = 0; 3987 } else { 3988 msize = env->me_mapsize; 3989 sizelo = msize & 0xffffffff; 3990 sizehi = msize >> 16 >> 16; /* only needed on Win64 */ 3991 3992 /* Windows won't create mappings for zero length files. 3993 * and won't map more than the file size. 3994 * Just set the maxsize right now. 3995 */ 3996 if (!(flags & MDB_WRITEMAP) && (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo 3997 || !SetEndOfFile(env->me_fd) 3998 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)) 3999 return ErrCode(); 4000 } 4001 4002 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ? 4003 PAGE_READWRITE : PAGE_READONLY, 4004 sizehi, sizelo, NULL); 4005 if (!mh) 4006 return ErrCode(); 4007 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ? 4008 FILE_MAP_WRITE : FILE_MAP_READ, 4009 0, 0, msize, addr); 4010 rc = env->me_map ? 0 : ErrCode(); 4011 CloseHandle(mh); 4012 if (rc) 4013 return rc; 4014 #else 4015 int mmap_flags = MAP_SHARED; 4016 int prot = PROT_READ; 4017 #ifdef MAP_NOSYNC /* Used on FreeBSD */ 4018 if (flags & MDB_NOSYNC) 4019 mmap_flags |= MAP_NOSYNC; 4020 #endif 4021 if (flags & MDB_WRITEMAP) { 4022 prot |= PROT_WRITE; 4023 if (ftruncate(env->me_fd, env->me_mapsize) < 0) 4024 return ErrCode(); 4025 } 4026 env->me_map = mmap(addr, env->me_mapsize, prot, mmap_flags, 4027 env->me_fd, 0); 4028 if (env->me_map == MAP_FAILED) { 4029 env->me_map = NULL; 4030 return ErrCode(); 4031 } 4032 4033 if (flags & MDB_NORDAHEAD) { 4034 /* Turn off readahead. It's harmful when the DB is larger than RAM. */ 4035 #ifdef MADV_RANDOM 4036 madvise(env->me_map, env->me_mapsize, MADV_RANDOM); 4037 #else 4038 #ifdef POSIX_MADV_RANDOM 4039 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM); 4040 #endif /* POSIX_MADV_RANDOM */ 4041 #endif /* MADV_RANDOM */ 4042 } 4043 #endif /* _WIN32 */ 4044 4045 /* Can happen because the address argument to mmap() is just a 4046 * hint. mmap() can pick another, e.g. if the range is in use. 4047 * The MAP_FIXED flag would prevent that, but then mmap could 4048 * instead unmap existing pages to make room for the new map. 4049 */ 4050 if (addr && env->me_map != addr) 4051 return EBUSY; /* TODO: Make a new MDB_* error code? */ 4052 4053 p = (MDB_page *)env->me_map; 4054 env->me_metas[0] = METADATA(p); 4055 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize); 4056 4057 return MDB_SUCCESS; 4058 } 4059 4060 int ESECT 4061 mdb_env_set_mapsize(MDB_env *env, size_t size) 4062 { 4063 /* If env is already open, caller is responsible for making 4064 * sure there are no active txns. 4065 */ 4066 if (env->me_map) { 4067 int rc; 4068 MDB_meta *meta; 4069 void *old; 4070 if (env->me_txn) 4071 return EINVAL; 4072 meta = mdb_env_pick_meta(env); 4073 if (!size) 4074 size = meta->mm_mapsize; 4075 { 4076 /* Silently round up to minimum if the size is too small */ 4077 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize; 4078 if (size < minsize) 4079 size = minsize; 4080 } 4081 munmap(env->me_map, env->me_mapsize); 4082 env->me_mapsize = size; 4083 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL; 4084 rc = mdb_env_map(env, old); 4085 if (rc) 4086 return rc; 4087 } 4088 env->me_mapsize = size; 4089 if (env->me_psize) 4090 env->me_maxpg = env->me_mapsize / env->me_psize; 4091 return MDB_SUCCESS; 4092 } 4093 4094 int ESECT 4095 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs) 4096 { 4097 if (env->me_map) 4098 return EINVAL; 4099 env->me_maxdbs = dbs + CORE_DBS; 4100 return MDB_SUCCESS; 4101 } 4102 4103 int ESECT 4104 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers) 4105 { 4106 if (env->me_map || readers < 1) 4107 return EINVAL; 4108 env->me_maxreaders = readers; 4109 return MDB_SUCCESS; 4110 } 4111 4112 int ESECT 4113 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers) 4114 { 4115 if (!env || !readers) 4116 return EINVAL; 4117 *readers = env->me_maxreaders; 4118 return MDB_SUCCESS; 4119 } 4120 4121 static int ESECT 4122 mdb_fsize(HANDLE fd, size_t *size) 4123 { 4124 #ifdef _WIN32 4125 LARGE_INTEGER fsize; 4126 4127 if (!GetFileSizeEx(fd, &fsize)) 4128 return ErrCode(); 4129 4130 *size = fsize.QuadPart; 4131 #else 4132 struct stat st; 4133 4134 if (fstat(fd, &st)) 4135 return ErrCode(); 4136 4137 *size = st.st_size; 4138 #endif 4139 return MDB_SUCCESS; 4140 } 4141 4142 4143 #ifdef _WIN32 4144 typedef wchar_t mdb_nchar_t; 4145 # define MDB_NAME(str) L##str 4146 # define mdb_name_cpy wcscpy 4147 #else 4148 /** Character type for file names: char on Unix, wchar_t on Windows */ 4149 typedef char mdb_nchar_t; 4150 # define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */ 4151 # define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */ 4152 #endif 4153 4154 /** Filename - string of #mdb_nchar_t[] */ 4155 typedef struct MDB_name { 4156 int mn_len; /**< Length */ 4157 int mn_alloced; /**< True if #mn_val was malloced */ 4158 mdb_nchar_t *mn_val; /**< Contents */ 4159 } MDB_name; 4160 4161 /** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */ 4162 static const mdb_nchar_t *const mdb_suffixes[2][2] = { 4163 { MDB_NAME("/data.mdb"), MDB_NAME("") }, 4164 { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") } 4165 }; 4166 4167 #define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */ 4168 4169 /** Set up filename + scratch area for filename suffix, for opening files. 4170 * It should be freed with #mdb_fname_destroy(). 4171 * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16. 4172 * 4173 * @param[in] path Pathname for #mdb_env_open(). 4174 * @param[in] envflags Whether a subdir and/or lockfile will be used. 4175 * @param[out] fname Resulting filename, with room for a suffix if necessary. 4176 */ 4177 static int ESECT 4178 mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname) 4179 { 4180 int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK); 4181 fname->mn_alloced = 0; 4182 #ifdef _WIN32 4183 return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN); 4184 #else 4185 fname->mn_len = strlen(path); 4186 if (no_suffix) 4187 fname->mn_val = (char *) path; 4188 else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) { 4189 fname->mn_alloced = 1; 4190 strcpy(fname->mn_val, path); 4191 } 4192 else 4193 return ENOMEM; 4194 return MDB_SUCCESS; 4195 #endif 4196 } 4197 4198 /** Destroy \b fname from #mdb_fname_init() */ 4199 #define mdb_fname_destroy(fname) \ 4200 do { if ((fname).mn_alloced) free((fname).mn_val); } while (0) 4201 4202 #ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */ 4203 # define MDB_CLOEXEC O_CLOEXEC 4204 #else 4205 # define MDB_CLOEXEC 0 4206 #endif 4207 4208 /** File type, access mode etc. for #mdb_fopen() */ 4209 enum mdb_fopen_type { 4210 #ifdef _WIN32 4211 MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS 4212 #else 4213 /* A comment in mdb_fopen() explains some O_* flag choices. */ 4214 MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */ 4215 MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */ 4216 MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */ 4217 MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */ 4218 /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits 4219 * distinguish otherwise-equal MDB_O_* constants from each other. 4220 */ 4221 MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY, 4222 MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */ 4223 #endif 4224 }; 4225 4226 /** Open an LMDB file. 4227 * @param[in] env The LMDB environment. 4228 * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is 4229 * appended if necessary to create the filename, without changing mn_len. 4230 * @param[in] which Determines file type, access mode, etc. 4231 * @param[in] mode The Unix permissions for the file, if we create it. 4232 * @param[out] res Resulting file handle. 4233 * @return 0 on success, non-zero on failure. 4234 */ 4235 static int ESECT 4236 mdb_fopen(const MDB_env *env, MDB_name *fname, 4237 enum mdb_fopen_type which, mdb_mode_t mode, 4238 HANDLE *res) 4239 { 4240 int rc = MDB_SUCCESS; 4241 HANDLE fd; 4242 #ifdef _WIN32 4243 DWORD acc, share, disp, attrs; 4244 #else 4245 int flags; 4246 #endif 4247 4248 if (fname->mn_alloced) /* modifiable copy */ 4249 mdb_name_cpy(fname->mn_val + fname->mn_len, 4250 mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]); 4251 4252 /* The directory must already exist. Usually the file need not. 4253 * MDB_O_META requires the file because we already created it using 4254 * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file. 4255 * 4256 * With MDB_O_COPY we do not want the OS to cache the writes, since 4257 * the source data is already in the OS cache. 4258 * 4259 * The lockfile needs FD_CLOEXEC (close file descriptor on exec*()) 4260 * to avoid the flock() issues noted under Caveats in lmdb.h. 4261 * Also set it for other filehandles which the user cannot get at 4262 * and close himself, which he may need after fork(). I.e. all but 4263 * me_fd, which programs do use via mdb_env_get_fd(). 4264 */ 4265 4266 #ifdef _WIN32 4267 acc = GENERIC_READ|GENERIC_WRITE; 4268 share = FILE_SHARE_READ|FILE_SHARE_WRITE; 4269 disp = OPEN_ALWAYS; 4270 attrs = FILE_ATTRIBUTE_NORMAL; 4271 switch (which) { 4272 case MDB_O_RDONLY: /* read-only datafile */ 4273 acc = GENERIC_READ; 4274 disp = OPEN_EXISTING; 4275 break; 4276 case MDB_O_META: /* for writing metapages */ 4277 acc = GENERIC_WRITE; 4278 disp = OPEN_EXISTING; 4279 attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH; 4280 break; 4281 case MDB_O_COPY: /* mdb_env_copy() & co */ 4282 acc = GENERIC_WRITE; 4283 share = 0; 4284 disp = CREATE_NEW; 4285 attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH; 4286 break; 4287 default: break; /* silence gcc -Wswitch (not all enum values handled) */ 4288 } 4289 fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL); 4290 #else 4291 fd = open(fname->mn_val, which & MDB_O_MASK, mode); 4292 #endif 4293 4294 if (fd == INVALID_HANDLE_VALUE) 4295 rc = ErrCode(); 4296 #ifndef _WIN32 4297 else { 4298 if (which != MDB_O_RDONLY && which != MDB_O_RDWR) { 4299 /* Set CLOEXEC if we could not pass it to open() */ 4300 if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1) 4301 (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC); 4302 } 4303 if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) { 4304 /* This may require buffer alignment. There is no portable 4305 * way to ask how much, so we require OS pagesize alignment. 4306 */ 4307 # ifdef F_NOCACHE /* __APPLE__ */ 4308 (void) fcntl(fd, F_NOCACHE, 1); 4309 # elif defined O_DIRECT 4310 /* open(...O_DIRECT...) would break on filesystems without 4311 * O_DIRECT support (ITS#7682). Try to set it here instead. 4312 */ 4313 if ((flags = fcntl(fd, F_GETFL)) != -1) 4314 (void) fcntl(fd, F_SETFL, flags | O_DIRECT); 4315 # endif 4316 } 4317 } 4318 #endif /* !_WIN32 */ 4319 4320 *res = fd; 4321 return rc; 4322 } 4323 4324 4325 #ifdef BROKEN_FDATASYNC 4326 #include <sys/utsname.h> 4327 #include <sys/vfs.h> 4328 #endif 4329 4330 /** Further setup required for opening an LMDB environment 4331 */ 4332 static int ESECT 4333 mdb_env_open2(MDB_env *env) 4334 { 4335 unsigned int flags = env->me_flags; 4336 int i, newenv = 0, rc; 4337 MDB_meta meta; 4338 4339 #ifdef _WIN32 4340 /* See if we should use QueryLimited */ 4341 rc = GetVersion(); 4342 if ((rc & 0xff) > 5) 4343 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION; 4344 else 4345 env->me_pidquery = PROCESS_QUERY_INFORMATION; 4346 #endif /* _WIN32 */ 4347 4348 #ifdef BROKEN_FDATASYNC 4349 /* ext3/ext4 fdatasync is broken on some older Linux kernels. 4350 * https://lkml.org/lkml/2012/9/3/83 4351 * Kernels after 3.6-rc6 are known good. 4352 * https://lkml.org/lkml/2012/9/10/556 4353 * See if the DB is on ext3/ext4, then check for new enough kernel 4354 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known 4355 * to be patched. 4356 */ 4357 { 4358 struct statfs st; 4359 fstatfs(env->me_fd, &st); 4360 while (st.f_type == 0xEF53) { 4361 struct utsname uts; 4362 int i; 4363 uname(&uts); 4364 if (uts.release[0] < '3') { 4365 if (!strncmp(uts.release, "2.6.32.", 7)) { 4366 i = atoi(uts.release+7); 4367 if (i >= 60) 4368 break; /* 2.6.32.60 and newer is OK */ 4369 } else if (!strncmp(uts.release, "2.6.34.", 7)) { 4370 i = atoi(uts.release+7); 4371 if (i >= 15) 4372 break; /* 2.6.34.15 and newer is OK */ 4373 } 4374 } else if (uts.release[0] == '3') { 4375 i = atoi(uts.release+2); 4376 if (i > 5) 4377 break; /* 3.6 and newer is OK */ 4378 if (i == 5) { 4379 i = atoi(uts.release+4); 4380 if (i >= 4) 4381 break; /* 3.5.4 and newer is OK */ 4382 } else if (i == 2) { 4383 i = atoi(uts.release+4); 4384 if (i >= 30) 4385 break; /* 3.2.30 and newer is OK */ 4386 } 4387 } else { /* 4.x and newer is OK */ 4388 break; 4389 } 4390 env->me_flags |= MDB_FSYNCONLY; 4391 break; 4392 } 4393 } 4394 #endif 4395 4396 if ((i = mdb_env_read_header(env, &meta)) != 0) { 4397 if (i != ENOENT) 4398 return i; 4399 DPUTS("new mdbenv"); 4400 newenv = 1; 4401 env->me_psize = env->me_os_psize; 4402 if (env->me_psize > MAX_PAGESIZE) 4403 env->me_psize = MAX_PAGESIZE; 4404 memset(&meta, 0, sizeof(meta)); 4405 mdb_env_init_meta0(env, &meta); 4406 meta.mm_mapsize = DEFAULT_MAPSIZE; 4407 } else { 4408 env->me_psize = meta.mm_psize; 4409 } 4410 4411 /* Was a mapsize configured? */ 4412 if (!env->me_mapsize) { 4413 env->me_mapsize = meta.mm_mapsize; 4414 } 4415 { 4416 /* Make sure mapsize >= committed data size. Even when using 4417 * mm_mapsize, which could be broken in old files (ITS#7789). 4418 */ 4419 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize; 4420 if (env->me_mapsize < minsize) 4421 env->me_mapsize = minsize; 4422 } 4423 meta.mm_mapsize = env->me_mapsize; 4424 4425 if (newenv && !(flags & MDB_FIXEDMAP)) { 4426 /* mdb_env_map() may grow the datafile. Write the metapages 4427 * first, so the file will be valid if initialization fails. 4428 * Except with FIXEDMAP, since we do not yet know mm_address. 4429 * We could fill in mm_address later, but then a different 4430 * program might end up doing that - one with a memory layout 4431 * and map address which does not suit the main program. 4432 */ 4433 rc = mdb_env_init_meta(env, &meta); 4434 if (rc) 4435 return rc; 4436 newenv = 0; 4437 } 4438 4439 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL); 4440 if (rc) 4441 return rc; 4442 4443 if (newenv) { 4444 if (flags & MDB_FIXEDMAP) 4445 meta.mm_address = env->me_map; 4446 i = mdb_env_init_meta(env, &meta); 4447 if (i != MDB_SUCCESS) { 4448 return i; 4449 } 4450 } 4451 4452 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1; 4453 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2) 4454 - sizeof(indx_t); 4455 #if !(MDB_MAXKEYSIZE) 4456 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db)); 4457 #endif 4458 env->me_maxpg = env->me_mapsize / env->me_psize; 4459 4460 #if MDB_DEBUG 4461 { 4462 MDB_meta *meta = mdb_env_pick_meta(env); 4463 MDB_db *db = &meta->mm_dbs[MAIN_DBI]; 4464 4465 DPRINTF(("opened database version %u, pagesize %u", 4466 meta->mm_version, env->me_psize)); 4467 DPRINTF(("using meta page %d", (int) (meta->mm_txnid & 1))); 4468 DPRINTF(("depth: %u", db->md_depth)); 4469 DPRINTF(("entries: %"Z"u", db->md_entries)); 4470 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages)); 4471 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages)); 4472 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages)); 4473 DPRINTF(("root: %"Z"u", db->md_root)); 4474 } 4475 #endif 4476 4477 return MDB_SUCCESS; 4478 } 4479 4480 4481 /** Release a reader thread's slot in the reader lock table. 4482 * This function is called automatically when a thread exits. 4483 * @param[in] ptr This points to the slot in the reader lock table. 4484 */ 4485 static void 4486 mdb_env_reader_dest(void *ptr) 4487 { 4488 MDB_reader *reader = ptr; 4489 4490 #ifndef _WIN32 4491 if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */ 4492 #endif 4493 /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */ 4494 reader->mr_pid = 0; 4495 } 4496 4497 #ifdef _WIN32 4498 /** Junk for arranging thread-specific callbacks on Windows. This is 4499 * necessarily platform and compiler-specific. Windows supports up 4500 * to 1088 keys. Let's assume nobody opens more than 64 environments 4501 * in a single process, for now. They can override this if needed. 4502 */ 4503 #ifndef MAX_TLS_KEYS 4504 #define MAX_TLS_KEYS 64 4505 #endif 4506 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS]; 4507 static int mdb_tls_nkeys; 4508 4509 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr) 4510 { 4511 int i; 4512 switch(reason) { 4513 case DLL_PROCESS_ATTACH: break; 4514 case DLL_THREAD_ATTACH: break; 4515 case DLL_THREAD_DETACH: 4516 for (i=0; i<mdb_tls_nkeys; i++) { 4517 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]); 4518 if (r) { 4519 mdb_env_reader_dest(r); 4520 } 4521 } 4522 break; 4523 case DLL_PROCESS_DETACH: break; 4524 } 4525 } 4526 #ifdef __GNUC__ 4527 #ifdef _WIN64 4528 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback; 4529 #else 4530 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback; 4531 #endif 4532 #else 4533 #ifdef _WIN64 4534 /* Force some symbol references. 4535 * _tls_used forces the linker to create the TLS directory if not already done 4536 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol. 4537 */ 4538 #pragma comment(linker, "/INCLUDE:_tls_used") 4539 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp") 4540 #pragma const_seg(".CRT$XLB") 4541 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp; 4542 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback; 4543 #pragma const_seg() 4544 #else /* _WIN32 */ 4545 #pragma comment(linker, "/INCLUDE:__tls_used") 4546 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp") 4547 #pragma data_seg(".CRT$XLB") 4548 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback; 4549 #pragma data_seg() 4550 #endif /* WIN 32/64 */ 4551 #endif /* !__GNUC__ */ 4552 #endif 4553 4554 /** Downgrade the exclusive lock on the region back to shared */ 4555 static int ESECT 4556 mdb_env_share_locks(MDB_env *env, int *excl) 4557 { 4558 int rc = 0; 4559 MDB_meta *meta = mdb_env_pick_meta(env); 4560 4561 env->me_txns->mti_txnid = meta->mm_txnid; 4562 4563 #ifdef _WIN32 4564 { 4565 OVERLAPPED ov; 4566 /* First acquire a shared lock. The Unlock will 4567 * then release the existing exclusive lock. 4568 */ 4569 memset(&ov, 0, sizeof(ov)); 4570 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) { 4571 rc = ErrCode(); 4572 } else { 4573 UnlockFile(env->me_lfd, 0, 0, 1, 0); 4574 *excl = 0; 4575 } 4576 } 4577 #else 4578 { 4579 struct flock lock_info; 4580 /* The shared lock replaces the existing lock */ 4581 memset((void *)&lock_info, 0, sizeof(lock_info)); 4582 lock_info.l_type = F_RDLCK; 4583 lock_info.l_whence = SEEK_SET; 4584 lock_info.l_start = 0; 4585 lock_info.l_len = 1; 4586 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) && 4587 (rc = ErrCode()) == EINTR) ; 4588 *excl = rc ? -1 : 0; /* error may mean we lost the lock */ 4589 } 4590 #endif 4591 4592 return rc; 4593 } 4594 4595 /** Try to get exclusive lock, otherwise shared. 4596 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive. 4597 */ 4598 static int ESECT 4599 mdb_env_excl_lock(MDB_env *env, int *excl) 4600 { 4601 int rc = 0; 4602 #ifdef _WIN32 4603 if (LockFile(env->me_lfd, 0, 0, 1, 0)) { 4604 *excl = 1; 4605 } else { 4606 OVERLAPPED ov; 4607 memset(&ov, 0, sizeof(ov)); 4608 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) { 4609 *excl = 0; 4610 } else { 4611 rc = ErrCode(); 4612 } 4613 } 4614 #else 4615 struct flock lock_info; 4616 memset((void *)&lock_info, 0, sizeof(lock_info)); 4617 lock_info.l_type = F_WRLCK; 4618 lock_info.l_whence = SEEK_SET; 4619 lock_info.l_start = 0; 4620 lock_info.l_len = 1; 4621 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) && 4622 (rc = ErrCode()) == EINTR) ; 4623 if (!rc) { 4624 *excl = 1; 4625 } else 4626 # ifndef MDB_USE_POSIX_MUTEX 4627 if (*excl < 0) /* always true when MDB_USE_POSIX_MUTEX */ 4628 # endif 4629 { 4630 lock_info.l_type = F_RDLCK; 4631 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) && 4632 (rc = ErrCode()) == EINTR) ; 4633 if (rc == 0) 4634 *excl = 0; 4635 } 4636 #endif 4637 return rc; 4638 } 4639 4640 #ifdef MDB_USE_HASH 4641 /* 4642 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code 4643 * 4644 * @(#) Revision: 5.1 4645 * @(#) Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp 4646 * @(#) Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v 4647 * 4648 * http://www.isthe.com/chongo/tech/comp/fnv/index.html 4649 * 4650 *** 4651 * 4652 * Please do not copyright this code. This code is in the public domain. 4653 * 4654 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, 4655 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO 4656 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR 4657 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF 4658 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR 4659 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR 4660 * PERFORMANCE OF THIS SOFTWARE. 4661 * 4662 * By: 4663 * chongo <Landon Curt Noll> /\oo/\ 4664 * http://www.isthe.com/chongo/ 4665 * 4666 * Share and Enjoy! :-) 4667 */ 4668 4669 typedef unsigned long long mdb_hash_t; 4670 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL) 4671 4672 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer 4673 * @param[in] val value to hash 4674 * @param[in] hval initial value for hash 4675 * @return 64 bit hash 4676 * 4677 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the 4678 * hval arg on the first call. 4679 */ 4680 static mdb_hash_t 4681 mdb_hash_val(MDB_val *val, mdb_hash_t hval) 4682 { 4683 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */ 4684 unsigned char *end = s + val->mv_size; 4685 /* 4686 * FNV-1a hash each octet of the string 4687 */ 4688 while (s < end) { 4689 /* xor the bottom with the current octet */ 4690 hval ^= (mdb_hash_t)*s++; 4691 4692 /* multiply by the 64 bit FNV magic prime mod 2^64 */ 4693 hval += (hval << 1) + (hval << 4) + (hval << 5) + 4694 (hval << 7) + (hval << 8) + (hval << 40); 4695 } 4696 /* return our new hash value */ 4697 return hval; 4698 } 4699 4700 /** Hash the string and output the encoded hash. 4701 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with 4702 * very short name limits. We don't care about the encoding being reversible, 4703 * we just want to preserve as many bits of the input as possible in a 4704 * small printable string. 4705 * @param[in] str string to hash 4706 * @param[out] encbuf an array of 11 chars to hold the hash 4707 */ 4708 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~"; 4709 4710 static void ESECT 4711 mdb_pack85(unsigned long l, char *out) 4712 { 4713 int i; 4714 4715 for (i=0; i<5; i++) { 4716 *out++ = mdb_a85[l % 85]; 4717 l /= 85; 4718 } 4719 } 4720 4721 static void ESECT 4722 mdb_hash_enc(MDB_val *val, char *encbuf) 4723 { 4724 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT); 4725 4726 mdb_pack85(h, encbuf); 4727 mdb_pack85(h>>32, encbuf+5); 4728 encbuf[10] = '\0'; 4729 } 4730 #endif 4731 4732 /** Open and/or initialize the lock region for the environment. 4733 * @param[in] env The LMDB environment. 4734 * @param[in] fname Filename + scratch area, from #mdb_fname_init(). 4735 * @param[in] mode The Unix permissions for the file, if we create it. 4736 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive 4737 * @return 0 on success, non-zero on failure. 4738 */ 4739 static int ESECT 4740 mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl) 4741 { 4742 #ifdef _WIN32 4743 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT 4744 #else 4745 # define MDB_ERRCODE_ROFS EROFS 4746 #endif 4747 int rc; 4748 off_t size, rsize; 4749 4750 rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd); 4751 if (rc) { 4752 /* Omit lockfile if read-only env on read-only filesystem */ 4753 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) { 4754 return MDB_SUCCESS; 4755 } 4756 goto fail; 4757 } 4758 4759 if (!(env->me_flags & MDB_NOTLS)) { 4760 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest); 4761 if (rc) 4762 goto fail; 4763 env->me_flags |= MDB_ENV_TXKEY; 4764 #ifdef _WIN32 4765 /* Windows TLS callbacks need help finding their TLS info. */ 4766 if (mdb_tls_nkeys >= MAX_TLS_KEYS) { 4767 rc = MDB_TLS_FULL; 4768 goto fail; 4769 } 4770 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey; 4771 #endif 4772 } 4773 4774 /* Try to get exclusive lock. If we succeed, then 4775 * nobody is using the lock region and we should initialize it. 4776 */ 4777 if ((rc = mdb_env_excl_lock(env, excl))) goto fail; 4778 4779 #ifdef _WIN32 4780 size = GetFileSize(env->me_lfd, NULL); 4781 #else 4782 size = lseek(env->me_lfd, 0, SEEK_END); 4783 if (size == -1) goto fail_errno; 4784 #endif 4785 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo); 4786 if (size < rsize && *excl > 0) { 4787 #ifdef _WIN32 4788 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize 4789 || !SetEndOfFile(env->me_lfd)) 4790 goto fail_errno; 4791 #else 4792 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno; 4793 #endif 4794 } else { 4795 rsize = size; 4796 size = rsize - sizeof(MDB_txninfo); 4797 env->me_maxreaders = size/sizeof(MDB_reader) + 1; 4798 } 4799 { 4800 #ifdef _WIN32 4801 HANDLE mh; 4802 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE, 4803 0, 0, NULL); 4804 if (!mh) goto fail_errno; 4805 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL); 4806 CloseHandle(mh); 4807 if (!env->me_txns) goto fail_errno; 4808 #else 4809 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED, 4810 env->me_lfd, 0); 4811 if (m == MAP_FAILED) goto fail_errno; 4812 env->me_txns = m; 4813 #endif 4814 } 4815 if (*excl > 0) { 4816 #ifdef _WIN32 4817 BY_HANDLE_FILE_INFORMATION stbuf; 4818 struct { 4819 DWORD volume; 4820 DWORD nhigh; 4821 DWORD nlow; 4822 } idbuf; 4823 MDB_val val; 4824 char encbuf[11]; 4825 4826 if (!mdb_sec_inited) { 4827 InitializeSecurityDescriptor(&mdb_null_sd, 4828 SECURITY_DESCRIPTOR_REVISION); 4829 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE); 4830 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES); 4831 mdb_all_sa.bInheritHandle = FALSE; 4832 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd; 4833 mdb_sec_inited = 1; 4834 } 4835 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno; 4836 idbuf.volume = stbuf.dwVolumeSerialNumber; 4837 idbuf.nhigh = stbuf.nFileIndexHigh; 4838 idbuf.nlow = stbuf.nFileIndexLow; 4839 val.mv_data = &idbuf; 4840 val.mv_size = sizeof(idbuf); 4841 mdb_hash_enc(&val, encbuf); 4842 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf); 4843 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf); 4844 env->me_rmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_rmname); 4845 if (!env->me_rmutex) goto fail_errno; 4846 env->me_wmutex = CreateMutexA(&mdb_all_sa, FALSE, env->me_txns->mti_wmname); 4847 if (!env->me_wmutex) goto fail_errno; 4848 #elif defined(MDB_USE_POSIX_SEM) 4849 struct stat stbuf; 4850 struct { 4851 dev_t dev; 4852 ino_t ino; 4853 } idbuf; 4854 MDB_val val; 4855 char encbuf[11]; 4856 4857 #if defined(__NetBSD__) 4858 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */ 4859 #endif 4860 if (fstat(env->me_lfd, &stbuf)) goto fail_errno; 4861 idbuf.dev = stbuf.st_dev; 4862 idbuf.ino = stbuf.st_ino; 4863 val.mv_data = &idbuf; 4864 val.mv_size = sizeof(idbuf); 4865 mdb_hash_enc(&val, encbuf); 4866 #ifdef MDB_SHORT_SEMNAMES 4867 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */ 4868 #endif 4869 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf); 4870 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf); 4871 /* Clean up after a previous run, if needed: Try to 4872 * remove both semaphores before doing anything else. 4873 */ 4874 sem_unlink(env->me_txns->mti_rmname); 4875 sem_unlink(env->me_txns->mti_wmname); 4876 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 4877 O_CREAT|O_EXCL, mode, 1); 4878 if (env->me_rmutex == SEM_FAILED) goto fail_errno; 4879 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 4880 O_CREAT|O_EXCL, mode, 1); 4881 if (env->me_wmutex == SEM_FAILED) goto fail_errno; 4882 #else /* MDB_USE_POSIX_MUTEX: */ 4883 pthread_mutexattr_t mattr; 4884 4885 /* Solaris needs this before initing a robust mutex. Otherwise 4886 * it may skip the init and return EBUSY "seems someone already 4887 * inited" or EINVAL "it was inited differently". 4888 */ 4889 memset(env->me_txns->mti_rmutex, 0, sizeof(*env->me_txns->mti_rmutex)); 4890 memset(env->me_txns->mti_wmutex, 0, sizeof(*env->me_txns->mti_wmutex)); 4891 4892 if ((rc = pthread_mutexattr_init(&mattr))) 4893 goto fail; 4894 4895 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED); 4896 #ifdef MDB_ROBUST_SUPPORTED 4897 if (!rc) rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST); 4898 #endif 4899 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr); 4900 if (!rc) rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr); 4901 pthread_mutexattr_destroy(&mattr); 4902 if (rc) 4903 goto fail; 4904 #endif /* _WIN32 || MDB_USE_POSIX_SEM */ 4905 4906 env->me_txns->mti_magic = MDB_MAGIC; 4907 env->me_txns->mti_format = MDB_LOCK_FORMAT; 4908 env->me_txns->mti_txnid = 0; 4909 env->me_txns->mti_numreaders = 0; 4910 4911 } else { 4912 if (env->me_txns->mti_magic != MDB_MAGIC) { 4913 DPUTS("lock region has invalid magic"); 4914 rc = MDB_INVALID; 4915 goto fail; 4916 } 4917 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) { 4918 DPRINTF(("lock region has format+version 0x%x, expected 0x%x", 4919 env->me_txns->mti_format, MDB_LOCK_FORMAT)); 4920 rc = MDB_VERSION_MISMATCH; 4921 goto fail; 4922 } 4923 rc = ErrCode(); 4924 if (rc && rc != EACCES && rc != EAGAIN) { 4925 goto fail; 4926 } 4927 #ifdef _WIN32 4928 env->me_rmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname); 4929 if (!env->me_rmutex) goto fail_errno; 4930 env->me_wmutex = OpenMutexA(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname); 4931 if (!env->me_wmutex) goto fail_errno; 4932 #elif defined(MDB_USE_POSIX_SEM) 4933 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0); 4934 if (env->me_rmutex == SEM_FAILED) goto fail_errno; 4935 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0); 4936 if (env->me_wmutex == SEM_FAILED) goto fail_errno; 4937 #endif 4938 } 4939 return MDB_SUCCESS; 4940 4941 fail_errno: 4942 rc = ErrCode(); 4943 fail: 4944 return rc; 4945 } 4946 4947 /** Only a subset of the @ref mdb_env flags can be changed 4948 * at runtime. Changing other flags requires closing the 4949 * environment and re-opening it with the new flags. 4950 */ 4951 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT) 4952 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \ 4953 MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD) 4954 4955 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS) 4956 # error "Persistent DB flags & env flags overlap, but both go in mm_flags" 4957 #endif 4958 4959 int ESECT 4960 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode) 4961 { 4962 int rc, excl = -1; 4963 MDB_name fname; 4964 4965 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS))) 4966 return EINVAL; 4967 4968 flags |= env->me_flags; 4969 4970 rc = mdb_fname_init(path, flags, &fname); 4971 if (rc) 4972 return rc; 4973 4974 if (flags & MDB_RDONLY) { 4975 /* silently ignore WRITEMAP when we're only getting read access */ 4976 flags &= ~MDB_WRITEMAP; 4977 } else { 4978 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) && 4979 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2))))) 4980 rc = ENOMEM; 4981 } 4982 env->me_flags = flags |= MDB_ENV_ACTIVE; 4983 if (rc) 4984 goto leave; 4985 4986 env->me_path = strdup(path); 4987 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx)); 4988 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t)); 4989 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int)); 4990 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) { 4991 rc = ENOMEM; 4992 goto leave; 4993 } 4994 env->me_dbxs[FREE_DBI].md_cmp = mdb_cmp_long; /* aligned MDB_INTEGERKEY */ 4995 4996 /* For RDONLY, get lockfile after we know datafile exists */ 4997 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) { 4998 rc = mdb_env_setup_locks(env, &fname, mode, &excl); 4999 if (rc) 5000 goto leave; 5001 } 5002 5003 rc = mdb_fopen(env, &fname, 5004 (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR, 5005 mode, &env->me_fd); 5006 if (rc) 5007 goto leave; 5008 5009 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) { 5010 rc = mdb_env_setup_locks(env, &fname, mode, &excl); 5011 if (rc) 5012 goto leave; 5013 } 5014 5015 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) { 5016 if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) { 5017 /* Synchronous fd for meta writes. Needed even with 5018 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset. 5019 */ 5020 rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd); 5021 if (rc) 5022 goto leave; 5023 } 5024 DPRINTF(("opened dbenv %p", (void *) env)); 5025 if (excl > 0) { 5026 rc = mdb_env_share_locks(env, &excl); 5027 if (rc) 5028 goto leave; 5029 } 5030 if (!(flags & MDB_RDONLY)) { 5031 MDB_txn *txn; 5032 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs * 5033 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1); 5034 if ((env->me_pbuf = calloc(1, env->me_psize)) && 5035 (txn = calloc(1, size))) 5036 { 5037 txn->mt_dbs = (MDB_db *)((char *)txn + tsize); 5038 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs); 5039 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs); 5040 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs); 5041 txn->mt_env = env; 5042 txn->mt_dbxs = env->me_dbxs; 5043 txn->mt_flags = MDB_TXN_FINISHED; 5044 env->me_txn0 = txn; 5045 } else { 5046 rc = ENOMEM; 5047 } 5048 } 5049 } 5050 5051 leave: 5052 if (rc) { 5053 mdb_env_close0(env, excl); 5054 } 5055 mdb_fname_destroy(fname); 5056 return rc; 5057 } 5058 5059 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */ 5060 static void ESECT 5061 mdb_env_close0(MDB_env *env, int excl) 5062 { 5063 int i; 5064 5065 if (!(env->me_flags & MDB_ENV_ACTIVE)) 5066 return; 5067 5068 /* Doing this here since me_dbxs may not exist during mdb_env_close */ 5069 if (env->me_dbxs) { 5070 for (i = env->me_maxdbs; --i >= CORE_DBS; ) 5071 free(env->me_dbxs[i].md_name.mv_data); 5072 free(env->me_dbxs); 5073 } 5074 5075 free(env->me_pbuf); 5076 free(env->me_dbiseqs); 5077 free(env->me_dbflags); 5078 free(env->me_path); 5079 free(env->me_dirty_list); 5080 free(env->me_txn0); 5081 mdb_midl_free(env->me_free_pgs); 5082 5083 if (env->me_flags & MDB_ENV_TXKEY) { 5084 pthread_key_delete(env->me_txkey); 5085 #ifdef _WIN32 5086 /* Delete our key from the global list */ 5087 for (i=0; i<mdb_tls_nkeys; i++) 5088 if (mdb_tls_keys[i] == env->me_txkey) { 5089 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1]; 5090 mdb_tls_nkeys--; 5091 break; 5092 } 5093 #endif 5094 } 5095 5096 if (env->me_map) { 5097 munmap(env->me_map, env->me_mapsize); 5098 } 5099 if (env->me_mfd != INVALID_HANDLE_VALUE) 5100 (void) close(env->me_mfd); 5101 if (env->me_fd != INVALID_HANDLE_VALUE) 5102 (void) close(env->me_fd); 5103 if (env->me_txns) { 5104 MDB_PID_T pid = getpid(); 5105 /* Clearing readers is done in this function because 5106 * me_txkey with its destructor must be disabled first. 5107 * 5108 * We skip the the reader mutex, so we touch only 5109 * data owned by this process (me_close_readers and 5110 * our readers), and clear each reader atomically. 5111 */ 5112 for (i = env->me_close_readers; --i >= 0; ) 5113 if (env->me_txns->mti_readers[i].mr_pid == pid) 5114 env->me_txns->mti_readers[i].mr_pid = 0; 5115 #ifdef _WIN32 5116 if (env->me_rmutex) { 5117 CloseHandle(env->me_rmutex); 5118 if (env->me_wmutex) CloseHandle(env->me_wmutex); 5119 } 5120 /* Windows automatically destroys the mutexes when 5121 * the last handle closes. 5122 */ 5123 #elif defined(MDB_USE_POSIX_SEM) 5124 if (env->me_rmutex != SEM_FAILED) { 5125 sem_close(env->me_rmutex); 5126 if (env->me_wmutex != SEM_FAILED) 5127 sem_close(env->me_wmutex); 5128 /* If we have the filelock: If we are the 5129 * only remaining user, clean up semaphores. 5130 */ 5131 if (excl == 0) 5132 mdb_env_excl_lock(env, &excl); 5133 if (excl > 0) { 5134 sem_unlink(env->me_txns->mti_rmname); 5135 sem_unlink(env->me_txns->mti_wmname); 5136 } 5137 } 5138 #endif 5139 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo)); 5140 } 5141 if (env->me_lfd != INVALID_HANDLE_VALUE) { 5142 #ifdef _WIN32 5143 if (excl >= 0) { 5144 /* Unlock the lockfile. Windows would have unlocked it 5145 * after closing anyway, but not necessarily at once. 5146 */ 5147 UnlockFile(env->me_lfd, 0, 0, 1, 0); 5148 } 5149 #endif 5150 (void) close(env->me_lfd); 5151 } 5152 5153 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY); 5154 } 5155 5156 void ESECT 5157 mdb_env_close(MDB_env *env) 5158 { 5159 MDB_page *dp; 5160 5161 if (env == NULL) 5162 return; 5163 5164 VGMEMP_DESTROY(env); 5165 while ((dp = env->me_dpages) != NULL) { 5166 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next)); 5167 env->me_dpages = dp->mp_next; 5168 free(dp); 5169 } 5170 5171 mdb_env_close0(env, 0); 5172 free(env); 5173 } 5174 5175 /** Compare two items pointing at aligned size_t's */ 5176 static int 5177 mdb_cmp_long(const MDB_val *a, const MDB_val *b) 5178 { 5179 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 : 5180 *(size_t *)a->mv_data > *(size_t *)b->mv_data; 5181 } 5182 5183 /** Compare two items pointing at aligned unsigned int's. 5184 * 5185 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp, 5186 * but #mdb_cmp_clong() is called instead if the data type is size_t. 5187 */ 5188 static int 5189 mdb_cmp_int(const MDB_val *a, const MDB_val *b) 5190 { 5191 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 : 5192 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data; 5193 } 5194 5195 /** Compare two items pointing at unsigned ints of unknown alignment. 5196 * Nodes and keys are guaranteed to be 2-byte aligned. 5197 */ 5198 static int 5199 mdb_cmp_cint(const MDB_val *a, const MDB_val *b) 5200 { 5201 #if BYTE_ORDER == LITTLE_ENDIAN 5202 unsigned short *u, *c; 5203 int x; 5204 5205 u = (unsigned short *) ((char *) a->mv_data + a->mv_size); 5206 c = (unsigned short *) ((char *) b->mv_data + a->mv_size); 5207 do { 5208 x = *--u - *--c; 5209 } while(!x && u > (unsigned short *)a->mv_data); 5210 return x; 5211 #else 5212 unsigned short *u, *c, *end; 5213 int x; 5214 5215 end = (unsigned short *) ((char *) a->mv_data + a->mv_size); 5216 u = (unsigned short *)a->mv_data; 5217 c = (unsigned short *)b->mv_data; 5218 do { 5219 x = *u++ - *c++; 5220 } while(!x && u < end); 5221 return x; 5222 #endif 5223 } 5224 5225 /** Compare two items lexically */ 5226 static int 5227 mdb_cmp_memn(const MDB_val *a, const MDB_val *b) 5228 { 5229 int diff; 5230 ssize_t len_diff; 5231 unsigned int len; 5232 5233 len = a->mv_size; 5234 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size; 5235 if (len_diff > 0) { 5236 len = b->mv_size; 5237 len_diff = 1; 5238 } 5239 5240 diff = memcmp(a->mv_data, b->mv_data, len); 5241 return diff ? diff : len_diff<0 ? -1 : len_diff; 5242 } 5243 5244 /** Compare two items in reverse byte order */ 5245 static int 5246 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b) 5247 { 5248 const unsigned char *p1, *p2, *p1_lim; 5249 ssize_t len_diff; 5250 int diff; 5251 5252 p1_lim = (const unsigned char *)a->mv_data; 5253 p1 = (const unsigned char *)a->mv_data + a->mv_size; 5254 p2 = (const unsigned char *)b->mv_data + b->mv_size; 5255 5256 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size; 5257 if (len_diff > 0) { 5258 p1_lim += len_diff; 5259 len_diff = 1; 5260 } 5261 5262 while (p1 > p1_lim) { 5263 diff = *--p1 - *--p2; 5264 if (diff) 5265 return diff; 5266 } 5267 return len_diff<0 ? -1 : len_diff; 5268 } 5269 5270 /** Search for key within a page, using binary search. 5271 * Returns the smallest entry larger or equal to the key. 5272 * If exactp is non-null, stores whether the found entry was an exact match 5273 * in *exactp (1 or 0). 5274 * Updates the cursor index with the index of the found entry. 5275 * If no entry larger or equal to the key is found, returns NULL. 5276 */ 5277 static MDB_node * 5278 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp) 5279 { 5280 unsigned int i = 0, nkeys; 5281 int low, high; 5282 int rc = 0; 5283 MDB_page *mp = mc->mc_pg[mc->mc_top]; 5284 MDB_node *node = NULL; 5285 MDB_val nodekey; 5286 MDB_cmp_func *cmp; 5287 DKBUF; 5288 5289 nkeys = NUMKEYS(mp); 5290 5291 DPRINTF(("searching %u keys in %s %spage %"Z"u", 5292 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "", 5293 mdb_dbg_pgno(mp))); 5294 5295 low = IS_LEAF(mp) ? 0 : 1; 5296 high = nkeys - 1; 5297 cmp = mc->mc_dbx->md_cmp; 5298 5299 /* Branch pages have no data, so if using integer keys, 5300 * alignment is guaranteed. Use faster mdb_cmp_int. 5301 */ 5302 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) { 5303 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t)) 5304 cmp = mdb_cmp_long; 5305 else 5306 cmp = mdb_cmp_int; 5307 } 5308 5309 if (IS_LEAF2(mp)) { 5310 nodekey.mv_size = mc->mc_db->md_pad; 5311 node = NODEPTR(mp, 0); /* fake */ 5312 while (low <= high) { 5313 i = (low + high) >> 1; 5314 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size); 5315 rc = cmp(key, &nodekey); 5316 DPRINTF(("found leaf index %u [%s], rc = %i", 5317 i, DKEY(&nodekey), rc)); 5318 if (rc == 0) 5319 break; 5320 if (rc > 0) 5321 low = i + 1; 5322 else 5323 high = i - 1; 5324 } 5325 } else { 5326 while (low <= high) { 5327 i = (low + high) >> 1; 5328 5329 node = NODEPTR(mp, i); 5330 nodekey.mv_size = NODEKSZ(node); 5331 nodekey.mv_data = NODEKEY(node); 5332 5333 rc = cmp(key, &nodekey); 5334 #if MDB_DEBUG 5335 if (IS_LEAF(mp)) 5336 DPRINTF(("found leaf index %u [%s], rc = %i", 5337 i, DKEY(&nodekey), rc)); 5338 else 5339 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i", 5340 i, DKEY(&nodekey), NODEPGNO(node), rc)); 5341 #endif 5342 if (rc == 0) 5343 break; 5344 if (rc > 0) 5345 low = i + 1; 5346 else 5347 high = i - 1; 5348 } 5349 } 5350 5351 if (rc > 0) { /* Found entry is less than the key. */ 5352 i++; /* Skip to get the smallest entry larger than key. */ 5353 if (!IS_LEAF2(mp)) 5354 node = NODEPTR(mp, i); 5355 } 5356 if (exactp) 5357 *exactp = (rc == 0 && nkeys > 0); 5358 /* store the key index */ 5359 mc->mc_ki[mc->mc_top] = i; 5360 if (i >= nkeys) 5361 /* There is no entry larger or equal to the key. */ 5362 return NULL; 5363 5364 /* nodeptr is fake for LEAF2 */ 5365 return node; 5366 } 5367 5368 #if 0 5369 static void 5370 mdb_cursor_adjust(MDB_cursor *mc, func) 5371 { 5372 MDB_cursor *m2; 5373 5374 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) { 5375 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) { 5376 func(mc, m2); 5377 } 5378 } 5379 } 5380 #endif 5381 5382 /** Pop a page off the top of the cursor's stack. */ 5383 static void 5384 mdb_cursor_pop(MDB_cursor *mc) 5385 { 5386 if (mc->mc_snum) { 5387 DPRINTF(("popping page %"Z"u off db %d cursor %p", 5388 mc->mc_pg[mc->mc_top]->mp_pgno, DDBI(mc), (void *) mc)); 5389 5390 mc->mc_snum--; 5391 if (mc->mc_snum) { 5392 mc->mc_top--; 5393 } else { 5394 mc->mc_flags &= ~C_INITIALIZED; 5395 } 5396 } 5397 } 5398 5399 /** Push a page onto the top of the cursor's stack. 5400 * Set #MDB_TXN_ERROR on failure. 5401 */ 5402 static int 5403 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp) 5404 { 5405 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno, 5406 DDBI(mc), (void *) mc)); 5407 5408 if (mc->mc_snum >= CURSOR_STACK) { 5409 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 5410 return MDB_CURSOR_FULL; 5411 } 5412 5413 mc->mc_top = mc->mc_snum++; 5414 mc->mc_pg[mc->mc_top] = mp; 5415 mc->mc_ki[mc->mc_top] = 0; 5416 5417 return MDB_SUCCESS; 5418 } 5419 5420 /** Find the address of the page corresponding to a given page number. 5421 * Set #MDB_TXN_ERROR on failure. 5422 * @param[in] mc the cursor accessing the page. 5423 * @param[in] pgno the page number for the page to retrieve. 5424 * @param[out] ret address of a pointer where the page's address will be stored. 5425 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page. 5426 * @return 0 on success, non-zero on failure. 5427 */ 5428 static int 5429 mdb_page_get(MDB_cursor *mc, pgno_t pgno, MDB_page **ret, int *lvl) 5430 { 5431 MDB_txn *txn = mc->mc_txn; 5432 MDB_env *env = txn->mt_env; 5433 MDB_page *p = NULL; 5434 int level; 5435 5436 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) { 5437 MDB_txn *tx2 = txn; 5438 level = 1; 5439 do { 5440 MDB_ID2L dl = tx2->mt_u.dirty_list; 5441 unsigned x; 5442 /* Spilled pages were dirtied in this txn and flushed 5443 * because the dirty list got full. Bring this page 5444 * back in from the map (but don't unspill it here, 5445 * leave that unless page_touch happens again). 5446 */ 5447 if (tx2->mt_spill_pgs) { 5448 MDB_ID pn = pgno << 1; 5449 x = mdb_midl_search(tx2->mt_spill_pgs, pn); 5450 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) { 5451 p = (MDB_page *)(env->me_map + env->me_psize * pgno); 5452 goto done; 5453 } 5454 } 5455 if (dl[0].mid) { 5456 unsigned x = mdb_mid2l_search(dl, pgno); 5457 if (x <= dl[0].mid && dl[x].mid == pgno) { 5458 p = dl[x].mptr; 5459 goto done; 5460 } 5461 } 5462 level++; 5463 } while ((tx2 = tx2->mt_parent) != NULL); 5464 } 5465 5466 if (pgno < txn->mt_next_pgno) { 5467 level = 0; 5468 p = (MDB_page *)(env->me_map + env->me_psize * pgno); 5469 } else { 5470 DPRINTF(("page %"Z"u not found", pgno)); 5471 txn->mt_flags |= MDB_TXN_ERROR; 5472 return MDB_PAGE_NOTFOUND; 5473 } 5474 5475 done: 5476 *ret = p; 5477 if (lvl) 5478 *lvl = level; 5479 return MDB_SUCCESS; 5480 } 5481 5482 /** Finish #mdb_page_search() / #mdb_page_search_lowest(). 5483 * The cursor is at the root page, set up the rest of it. 5484 */ 5485 static int 5486 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags) 5487 { 5488 MDB_page *mp = mc->mc_pg[mc->mc_top]; 5489 int rc; 5490 DKBUF; 5491 5492 while (IS_BRANCH(mp)) { 5493 MDB_node *node; 5494 indx_t i; 5495 5496 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp))); 5497 /* Don't assert on branch pages in the FreeDB. We can get here 5498 * while in the process of rebalancing a FreeDB branch page; we must 5499 * let that proceed. ITS#8336 5500 */ 5501 mdb_cassert(mc, !mc->mc_dbi || NUMKEYS(mp) > 1); 5502 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0)))); 5503 5504 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) { 5505 i = 0; 5506 if (flags & MDB_PS_LAST) { 5507 i = NUMKEYS(mp) - 1; 5508 /* if already init'd, see if we're already in right place */ 5509 if (mc->mc_flags & C_INITIALIZED) { 5510 if (mc->mc_ki[mc->mc_top] == i) { 5511 mc->mc_top = mc->mc_snum++; 5512 mp = mc->mc_pg[mc->mc_top]; 5513 goto ready; 5514 } 5515 } 5516 } 5517 } else { 5518 int exact; 5519 node = mdb_node_search(mc, key, &exact); 5520 if (node == NULL) 5521 i = NUMKEYS(mp) - 1; 5522 else { 5523 i = mc->mc_ki[mc->mc_top]; 5524 if (!exact) { 5525 mdb_cassert(mc, i > 0); 5526 i--; 5527 } 5528 } 5529 DPRINTF(("following index %u for key [%s]", i, DKEY(key))); 5530 } 5531 5532 mdb_cassert(mc, i < NUMKEYS(mp)); 5533 node = NODEPTR(mp, i); 5534 5535 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0) 5536 return rc; 5537 5538 mc->mc_ki[mc->mc_top] = i; 5539 if ((rc = mdb_cursor_push(mc, mp))) 5540 return rc; 5541 5542 ready: 5543 if (flags & MDB_PS_MODIFY) { 5544 if ((rc = mdb_page_touch(mc)) != 0) 5545 return rc; 5546 mp = mc->mc_pg[mc->mc_top]; 5547 } 5548 } 5549 5550 if (!IS_LEAF(mp)) { 5551 DPRINTF(("internal error, index points to a %02X page!?", 5552 mp->mp_flags)); 5553 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 5554 return MDB_CORRUPTED; 5555 } 5556 5557 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno, 5558 key ? DKEY(key) : "null")); 5559 mc->mc_flags |= C_INITIALIZED; 5560 mc->mc_flags &= ~C_EOF; 5561 5562 return MDB_SUCCESS; 5563 } 5564 5565 /** Search for the lowest key under the current branch page. 5566 * This just bypasses a NUMKEYS check in the current page 5567 * before calling mdb_page_search_root(), because the callers 5568 * are all in situations where the current page is known to 5569 * be underfilled. 5570 */ 5571 static int 5572 mdb_page_search_lowest(MDB_cursor *mc) 5573 { 5574 MDB_page *mp = mc->mc_pg[mc->mc_top]; 5575 MDB_node *node = NODEPTR(mp, 0); 5576 int rc; 5577 5578 if ((rc = mdb_page_get(mc, NODEPGNO(node), &mp, NULL)) != 0) 5579 return rc; 5580 5581 mc->mc_ki[mc->mc_top] = 0; 5582 if ((rc = mdb_cursor_push(mc, mp))) 5583 return rc; 5584 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST); 5585 } 5586 5587 /** Search for the page a given key should be in. 5588 * Push it and its parent pages on the cursor stack. 5589 * @param[in,out] mc the cursor for this operation. 5590 * @param[in] key the key to search for, or NULL for first/last page. 5591 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB 5592 * are touched (updated with new page numbers). 5593 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf. 5594 * This is used by #mdb_cursor_first() and #mdb_cursor_last(). 5595 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups. 5596 * @return 0 on success, non-zero on failure. 5597 */ 5598 static int 5599 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags) 5600 { 5601 int rc; 5602 pgno_t root; 5603 5604 /* Make sure the txn is still viable, then find the root from 5605 * the txn's db table and set it as the root of the cursor's stack. 5606 */ 5607 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) { 5608 DPUTS("transaction may not be used now"); 5609 return MDB_BAD_TXN; 5610 } else { 5611 /* Make sure we're using an up-to-date root */ 5612 if (*mc->mc_dbflag & DB_STALE) { 5613 MDB_cursor mc2; 5614 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi)) 5615 return MDB_BAD_DBI; 5616 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL); 5617 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0); 5618 if (rc) 5619 return rc; 5620 { 5621 MDB_val data; 5622 int exact = 0; 5623 uint16_t flags; 5624 MDB_node *leaf = mdb_node_search(&mc2, 5625 &mc->mc_dbx->md_name, &exact); 5626 if (!exact) 5627 return MDB_NOTFOUND; 5628 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA) 5629 return MDB_INCOMPATIBLE; /* not a named DB */ 5630 rc = mdb_node_read(&mc2, leaf, &data); 5631 if (rc) 5632 return rc; 5633 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)), 5634 sizeof(uint16_t)); 5635 /* The txn may not know this DBI, or another process may 5636 * have dropped and recreated the DB with other flags. 5637 */ 5638 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags) 5639 return MDB_INCOMPATIBLE; 5640 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db)); 5641 } 5642 *mc->mc_dbflag &= ~DB_STALE; 5643 } 5644 root = mc->mc_db->md_root; 5645 5646 if (root == P_INVALID) { /* Tree is empty. */ 5647 DPUTS("tree is empty"); 5648 return MDB_NOTFOUND; 5649 } 5650 } 5651 5652 mdb_cassert(mc, root > 1); 5653 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) 5654 if ((rc = mdb_page_get(mc, root, &mc->mc_pg[0], NULL)) != 0) 5655 return rc; 5656 5657 mc->mc_snum = 1; 5658 mc->mc_top = 0; 5659 5660 DPRINTF(("db %d root page %"Z"u has flags 0x%X", 5661 DDBI(mc), root, mc->mc_pg[0]->mp_flags)); 5662 5663 if (flags & MDB_PS_MODIFY) { 5664 if ((rc = mdb_page_touch(mc))) 5665 return rc; 5666 } 5667 5668 if (flags & MDB_PS_ROOTONLY) 5669 return MDB_SUCCESS; 5670 5671 return mdb_page_search_root(mc, key, flags); 5672 } 5673 5674 static int 5675 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp) 5676 { 5677 MDB_txn *txn = mc->mc_txn; 5678 pgno_t pg = mp->mp_pgno; 5679 unsigned x = 0, ovpages = mp->mp_pages; 5680 MDB_env *env = txn->mt_env; 5681 MDB_IDL sl = txn->mt_spill_pgs; 5682 MDB_ID pn = pg << 1; 5683 int rc; 5684 5685 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages)); 5686 /* If the page is dirty or on the spill list we just acquired it, 5687 * so we should give it back to our current free list, if any. 5688 * Otherwise put it onto the list of pages we freed in this txn. 5689 * 5690 * Won't create me_pghead: me_pglast must be inited along with it. 5691 * Unsupported in nested txns: They would need to hide the page 5692 * range in ancestor txns' dirty and spilled lists. 5693 */ 5694 if (env->me_pghead && 5695 !txn->mt_parent && 5696 ((mp->mp_flags & P_DIRTY) || 5697 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn))) 5698 { 5699 unsigned i, j; 5700 pgno_t *mop; 5701 MDB_ID2 *dl, ix, iy; 5702 rc = mdb_midl_need(&env->me_pghead, ovpages); 5703 if (rc) 5704 return rc; 5705 if (!(mp->mp_flags & P_DIRTY)) { 5706 /* This page is no longer spilled */ 5707 if (x == sl[0]) 5708 sl[0]--; 5709 else 5710 sl[x] |= 1; 5711 goto release; 5712 } 5713 /* Remove from dirty list */ 5714 dl = txn->mt_u.dirty_list; 5715 x = dl[0].mid--; 5716 for (ix = dl[x]; ix.mptr != mp; ix = iy) { 5717 if (x > 1) { 5718 x--; 5719 iy = dl[x]; 5720 dl[x] = ix; 5721 } else { 5722 mdb_cassert(mc, x > 1); 5723 j = ++(dl[0].mid); 5724 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */ 5725 txn->mt_flags |= MDB_TXN_ERROR; 5726 return MDB_CORRUPTED; 5727 } 5728 } 5729 txn->mt_dirty_room++; 5730 if (!(env->me_flags & MDB_WRITEMAP)) 5731 mdb_dpage_free(env, mp); 5732 release: 5733 /* Insert in me_pghead */ 5734 mop = env->me_pghead; 5735 j = mop[0] + ovpages; 5736 for (i = mop[0]; i && mop[i] < pg; i--) 5737 mop[j--] = mop[i]; 5738 while (j>i) 5739 mop[j--] = pg++; 5740 mop[0] += ovpages; 5741 } else { 5742 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages); 5743 if (rc) 5744 return rc; 5745 } 5746 mc->mc_db->md_overflow_pages -= ovpages; 5747 return 0; 5748 } 5749 5750 /** Return the data associated with a given node. 5751 * @param[in] mc The cursor for this operation. 5752 * @param[in] leaf The node being read. 5753 * @param[out] data Updated to point to the node's data. 5754 * @return 0 on success, non-zero on failure. 5755 */ 5756 static int 5757 mdb_node_read(MDB_cursor *mc, MDB_node *leaf, MDB_val *data) 5758 { 5759 MDB_page *omp; /* overflow page */ 5760 pgno_t pgno; 5761 int rc; 5762 5763 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) { 5764 data->mv_size = NODEDSZ(leaf); 5765 data->mv_data = NODEDATA(leaf); 5766 return MDB_SUCCESS; 5767 } 5768 5769 /* Read overflow data. 5770 */ 5771 data->mv_size = NODEDSZ(leaf); 5772 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno)); 5773 if ((rc = mdb_page_get(mc, pgno, &omp, NULL)) != 0) { 5774 DPRINTF(("read overflow page %"Z"u failed", pgno)); 5775 return rc; 5776 } 5777 data->mv_data = METADATA(omp); 5778 5779 return MDB_SUCCESS; 5780 } 5781 5782 int 5783 mdb_get(MDB_txn *txn, MDB_dbi dbi, 5784 MDB_val *key, MDB_val *data) 5785 { 5786 MDB_cursor mc; 5787 MDB_xcursor mx; 5788 int exact = 0; 5789 DKBUF; 5790 5791 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key))); 5792 5793 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 5794 return EINVAL; 5795 5796 if (txn->mt_flags & MDB_TXN_BLOCKED) 5797 return MDB_BAD_TXN; 5798 5799 mdb_cursor_init(&mc, txn, dbi, &mx); 5800 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact); 5801 } 5802 5803 /** Find a sibling for a page. 5804 * Replaces the page at the top of the cursor's stack with the 5805 * specified sibling, if one exists. 5806 * @param[in] mc The cursor for this operation. 5807 * @param[in] move_right Non-zero if the right sibling is requested, 5808 * otherwise the left sibling. 5809 * @return 0 on success, non-zero on failure. 5810 */ 5811 static int 5812 mdb_cursor_sibling(MDB_cursor *mc, int move_right) 5813 { 5814 int rc; 5815 MDB_node *indx; 5816 MDB_page *mp; 5817 5818 if (mc->mc_snum < 2) { 5819 return MDB_NOTFOUND; /* root has no siblings */ 5820 } 5821 5822 mdb_cursor_pop(mc); 5823 DPRINTF(("parent page is page %"Z"u, index %u", 5824 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top])); 5825 5826 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top])) 5827 : (mc->mc_ki[mc->mc_top] == 0)) { 5828 DPRINTF(("no more keys left, moving to %s sibling", 5829 move_right ? "right" : "left")); 5830 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) { 5831 /* undo cursor_pop before returning */ 5832 mc->mc_top++; 5833 mc->mc_snum++; 5834 return rc; 5835 } 5836 } else { 5837 if (move_right) 5838 mc->mc_ki[mc->mc_top]++; 5839 else 5840 mc->mc_ki[mc->mc_top]--; 5841 DPRINTF(("just moving to %s index key %u", 5842 move_right ? "right" : "left", mc->mc_ki[mc->mc_top])); 5843 } 5844 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top])); 5845 5846 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 5847 if ((rc = mdb_page_get(mc, NODEPGNO(indx), &mp, NULL)) != 0) { 5848 /* mc will be inconsistent if caller does mc_snum++ as above */ 5849 mc->mc_flags &= ~(C_INITIALIZED|C_EOF); 5850 return rc; 5851 } 5852 5853 mdb_cursor_push(mc, mp); 5854 if (!move_right) 5855 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1; 5856 5857 return MDB_SUCCESS; 5858 } 5859 5860 /** Move the cursor to the next data item. */ 5861 static int 5862 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op) 5863 { 5864 MDB_page *mp; 5865 MDB_node *leaf; 5866 int rc; 5867 5868 if ((mc->mc_flags & C_DEL && op == MDB_NEXT_DUP)) 5869 return MDB_NOTFOUND; 5870 5871 if (!(mc->mc_flags & C_INITIALIZED)) 5872 return mdb_cursor_first(mc, key, data); 5873 5874 mp = mc->mc_pg[mc->mc_top]; 5875 5876 if (mc->mc_flags & C_EOF) { 5877 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mp)-1) 5878 return MDB_NOTFOUND; 5879 mc->mc_flags ^= C_EOF; 5880 } 5881 5882 if (mc->mc_db->md_flags & MDB_DUPSORT) { 5883 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 5884 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 5885 if (op == MDB_NEXT || op == MDB_NEXT_DUP) { 5886 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT); 5887 if (op != MDB_NEXT || rc != MDB_NOTFOUND) { 5888 if (rc == MDB_SUCCESS) 5889 MDB_GET_KEY(leaf, key); 5890 return rc; 5891 } 5892 } 5893 } else { 5894 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 5895 if (op == MDB_NEXT_DUP) 5896 return MDB_NOTFOUND; 5897 } 5898 } 5899 5900 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", 5901 mdb_dbg_pgno(mp), (void *) mc)); 5902 if (mc->mc_flags & C_DEL) { 5903 mc->mc_flags ^= C_DEL; 5904 goto skip; 5905 } 5906 5907 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) { 5908 DPUTS("=====> move to next sibling page"); 5909 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) { 5910 mc->mc_flags |= C_EOF; 5911 return rc; 5912 } 5913 mp = mc->mc_pg[mc->mc_top]; 5914 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top])); 5915 } else 5916 mc->mc_ki[mc->mc_top]++; 5917 5918 skip: 5919 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u", 5920 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top])); 5921 5922 if (IS_LEAF2(mp)) { 5923 key->mv_size = mc->mc_db->md_pad; 5924 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size); 5925 return MDB_SUCCESS; 5926 } 5927 5928 mdb_cassert(mc, IS_LEAF(mp)); 5929 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 5930 5931 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 5932 mdb_xcursor_init1(mc, leaf); 5933 } 5934 if (data) { 5935 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 5936 return rc; 5937 5938 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 5939 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); 5940 if (rc != MDB_SUCCESS) 5941 return rc; 5942 } 5943 } 5944 5945 MDB_GET_KEY(leaf, key); 5946 return MDB_SUCCESS; 5947 } 5948 5949 /** Move the cursor to the previous data item. */ 5950 static int 5951 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op) 5952 { 5953 MDB_page *mp; 5954 MDB_node *leaf; 5955 int rc; 5956 5957 if (!(mc->mc_flags & C_INITIALIZED)) { 5958 rc = mdb_cursor_last(mc, key, data); 5959 if (rc) 5960 return rc; 5961 mc->mc_ki[mc->mc_top]++; 5962 } 5963 5964 mp = mc->mc_pg[mc->mc_top]; 5965 5966 if (mc->mc_db->md_flags & MDB_DUPSORT) { 5967 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 5968 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 5969 if (op == MDB_PREV || op == MDB_PREV_DUP) { 5970 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV); 5971 if (op != MDB_PREV || rc != MDB_NOTFOUND) { 5972 if (rc == MDB_SUCCESS) { 5973 MDB_GET_KEY(leaf, key); 5974 mc->mc_flags &= ~C_EOF; 5975 } 5976 return rc; 5977 } 5978 } 5979 } else { 5980 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 5981 if (op == MDB_PREV_DUP) 5982 return MDB_NOTFOUND; 5983 } 5984 } 5985 5986 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", 5987 mdb_dbg_pgno(mp), (void *) mc)); 5988 5989 mc->mc_flags &= ~(C_EOF|C_DEL); 5990 5991 if (mc->mc_ki[mc->mc_top] == 0) { 5992 DPUTS("=====> move to prev sibling page"); 5993 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) { 5994 return rc; 5995 } 5996 mp = mc->mc_pg[mc->mc_top]; 5997 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1; 5998 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top])); 5999 } else 6000 mc->mc_ki[mc->mc_top]--; 6001 6002 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u", 6003 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top])); 6004 6005 if (IS_LEAF2(mp)) { 6006 key->mv_size = mc->mc_db->md_pad; 6007 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size); 6008 return MDB_SUCCESS; 6009 } 6010 6011 mdb_cassert(mc, IS_LEAF(mp)); 6012 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 6013 6014 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6015 mdb_xcursor_init1(mc, leaf); 6016 } 6017 if (data) { 6018 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 6019 return rc; 6020 6021 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6022 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL); 6023 if (rc != MDB_SUCCESS) 6024 return rc; 6025 } 6026 } 6027 6028 MDB_GET_KEY(leaf, key); 6029 return MDB_SUCCESS; 6030 } 6031 6032 /** Set the cursor on a specific data item. */ 6033 static int 6034 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, 6035 MDB_cursor_op op, int *exactp) 6036 { 6037 int rc; 6038 MDB_page *mp; 6039 MDB_node *leaf = NULL; 6040 DKBUF; 6041 6042 if (key->mv_size == 0) 6043 return MDB_BAD_VALSIZE; 6044 6045 if (mc->mc_xcursor) 6046 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 6047 6048 /* See if we're already on the right page */ 6049 if (mc->mc_flags & C_INITIALIZED) { 6050 MDB_val nodekey; 6051 6052 mp = mc->mc_pg[mc->mc_top]; 6053 if (!NUMKEYS(mp)) { 6054 mc->mc_ki[mc->mc_top] = 0; 6055 return MDB_NOTFOUND; 6056 } 6057 if (mp->mp_flags & P_LEAF2) { 6058 nodekey.mv_size = mc->mc_db->md_pad; 6059 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size); 6060 } else { 6061 leaf = NODEPTR(mp, 0); 6062 MDB_GET_KEY2(leaf, nodekey); 6063 } 6064 rc = mc->mc_dbx->md_cmp(key, &nodekey); 6065 if (rc == 0) { 6066 /* Probably happens rarely, but first node on the page 6067 * was the one we wanted. 6068 */ 6069 mc->mc_ki[mc->mc_top] = 0; 6070 if (exactp) 6071 *exactp = 1; 6072 goto set1; 6073 } 6074 if (rc > 0) { 6075 unsigned int i; 6076 unsigned int nkeys = NUMKEYS(mp); 6077 if (nkeys > 1) { 6078 if (mp->mp_flags & P_LEAF2) { 6079 nodekey.mv_data = LEAF2KEY(mp, 6080 nkeys-1, nodekey.mv_size); 6081 } else { 6082 leaf = NODEPTR(mp, nkeys-1); 6083 MDB_GET_KEY2(leaf, nodekey); 6084 } 6085 rc = mc->mc_dbx->md_cmp(key, &nodekey); 6086 if (rc == 0) { 6087 /* last node was the one we wanted */ 6088 mc->mc_ki[mc->mc_top] = nkeys-1; 6089 if (exactp) 6090 *exactp = 1; 6091 goto set1; 6092 } 6093 if (rc < 0) { 6094 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) { 6095 /* This is definitely the right page, skip search_page */ 6096 if (mp->mp_flags & P_LEAF2) { 6097 nodekey.mv_data = LEAF2KEY(mp, 6098 mc->mc_ki[mc->mc_top], nodekey.mv_size); 6099 } else { 6100 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 6101 MDB_GET_KEY2(leaf, nodekey); 6102 } 6103 rc = mc->mc_dbx->md_cmp(key, &nodekey); 6104 if (rc == 0) { 6105 /* current node was the one we wanted */ 6106 if (exactp) 6107 *exactp = 1; 6108 goto set1; 6109 } 6110 } 6111 rc = 0; 6112 mc->mc_flags &= ~C_EOF; 6113 goto set2; 6114 } 6115 } 6116 /* If any parents have right-sibs, search. 6117 * Otherwise, there's nothing further. 6118 */ 6119 for (i=0; i<mc->mc_top; i++) 6120 if (mc->mc_ki[i] < 6121 NUMKEYS(mc->mc_pg[i])-1) 6122 break; 6123 if (i == mc->mc_top) { 6124 /* There are no other pages */ 6125 mc->mc_ki[mc->mc_top] = nkeys; 6126 return MDB_NOTFOUND; 6127 } 6128 } 6129 if (!mc->mc_top) { 6130 /* There are no other pages */ 6131 mc->mc_ki[mc->mc_top] = 0; 6132 if (op == MDB_SET_RANGE && !exactp) { 6133 rc = 0; 6134 goto set1; 6135 } else 6136 return MDB_NOTFOUND; 6137 } 6138 } else { 6139 mc->mc_pg[0] = 0; 6140 } 6141 6142 rc = mdb_page_search(mc, key, 0); 6143 if (rc != MDB_SUCCESS) 6144 return rc; 6145 6146 mp = mc->mc_pg[mc->mc_top]; 6147 mdb_cassert(mc, IS_LEAF(mp)); 6148 6149 set2: 6150 leaf = mdb_node_search(mc, key, exactp); 6151 if (exactp != NULL && !*exactp) { 6152 /* MDB_SET specified and not an exact match. */ 6153 return MDB_NOTFOUND; 6154 } 6155 6156 if (leaf == NULL) { 6157 DPUTS("===> inexact leaf not found, goto sibling"); 6158 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) { 6159 mc->mc_flags |= C_EOF; 6160 return rc; /* no entries matched */ 6161 } 6162 mp = mc->mc_pg[mc->mc_top]; 6163 mdb_cassert(mc, IS_LEAF(mp)); 6164 leaf = NODEPTR(mp, 0); 6165 } 6166 6167 set1: 6168 mc->mc_flags |= C_INITIALIZED; 6169 mc->mc_flags &= ~C_EOF; 6170 6171 if (IS_LEAF2(mp)) { 6172 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) { 6173 key->mv_size = mc->mc_db->md_pad; 6174 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size); 6175 } 6176 return MDB_SUCCESS; 6177 } 6178 6179 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6180 mdb_xcursor_init1(mc, leaf); 6181 } 6182 if (data) { 6183 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6184 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) { 6185 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); 6186 } else { 6187 int ex2, *ex2p; 6188 if (op == MDB_GET_BOTH) { 6189 ex2p = &ex2; 6190 ex2 = 0; 6191 } else { 6192 ex2p = NULL; 6193 } 6194 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p); 6195 if (rc != MDB_SUCCESS) 6196 return rc; 6197 } 6198 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) { 6199 MDB_val olddata; 6200 MDB_cmp_func *dcmp; 6201 if ((rc = mdb_node_read(mc, leaf, &olddata)) != MDB_SUCCESS) 6202 return rc; 6203 dcmp = mc->mc_dbx->md_dcmp; 6204 #if UINT_MAX < SIZE_MAX 6205 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t)) 6206 dcmp = mdb_cmp_clong; 6207 #endif 6208 rc = dcmp(data, &olddata); 6209 if (rc) { 6210 if (op == MDB_GET_BOTH || rc > 0) 6211 return MDB_NOTFOUND; 6212 rc = 0; 6213 } 6214 *data = olddata; 6215 6216 } else { 6217 if (mc->mc_xcursor) 6218 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 6219 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 6220 return rc; 6221 } 6222 } 6223 6224 /* The key already matches in all other cases */ 6225 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) 6226 MDB_GET_KEY(leaf, key); 6227 DPRINTF(("==> cursor placed on key [%s]", DKEY(key))); 6228 6229 return rc; 6230 } 6231 6232 /** Move the cursor to the first item in the database. */ 6233 static int 6234 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data) 6235 { 6236 int rc; 6237 MDB_node *leaf; 6238 6239 if (mc->mc_xcursor) 6240 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 6241 6242 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) { 6243 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST); 6244 if (rc != MDB_SUCCESS) 6245 return rc; 6246 } 6247 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); 6248 6249 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0); 6250 mc->mc_flags |= C_INITIALIZED; 6251 mc->mc_flags &= ~C_EOF; 6252 6253 mc->mc_ki[mc->mc_top] = 0; 6254 6255 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { 6256 key->mv_size = mc->mc_db->md_pad; 6257 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size); 6258 return MDB_SUCCESS; 6259 } 6260 6261 if (data) { 6262 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6263 mdb_xcursor_init1(mc, leaf); 6264 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL); 6265 if (rc) 6266 return rc; 6267 } else { 6268 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 6269 return rc; 6270 } 6271 } 6272 MDB_GET_KEY(leaf, key); 6273 return MDB_SUCCESS; 6274 } 6275 6276 /** Move the cursor to the last item in the database. */ 6277 static int 6278 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data) 6279 { 6280 int rc; 6281 MDB_node *leaf; 6282 6283 if (mc->mc_xcursor) 6284 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 6285 6286 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) { 6287 rc = mdb_page_search(mc, NULL, MDB_PS_LAST); 6288 if (rc != MDB_SUCCESS) 6289 return rc; 6290 } 6291 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); 6292 6293 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1; 6294 mc->mc_flags |= C_INITIALIZED|C_EOF; 6295 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 6296 6297 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { 6298 key->mv_size = mc->mc_db->md_pad; 6299 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size); 6300 return MDB_SUCCESS; 6301 } 6302 6303 if (data) { 6304 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6305 mdb_xcursor_init1(mc, leaf); 6306 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL); 6307 if (rc) 6308 return rc; 6309 } else { 6310 if ((rc = mdb_node_read(mc, leaf, data)) != MDB_SUCCESS) 6311 return rc; 6312 } 6313 } 6314 6315 MDB_GET_KEY(leaf, key); 6316 return MDB_SUCCESS; 6317 } 6318 6319 int 6320 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data, 6321 MDB_cursor_op op) 6322 { 6323 int rc; 6324 int exact = 0; 6325 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data); 6326 6327 if (mc == NULL) 6328 return EINVAL; 6329 6330 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) 6331 return MDB_BAD_TXN; 6332 6333 switch (op) { 6334 case MDB_GET_CURRENT: 6335 if (!(mc->mc_flags & C_INITIALIZED)) { 6336 rc = EINVAL; 6337 } else { 6338 MDB_page *mp = mc->mc_pg[mc->mc_top]; 6339 int nkeys = NUMKEYS(mp); 6340 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) { 6341 mc->mc_ki[mc->mc_top] = nkeys; 6342 rc = MDB_NOTFOUND; 6343 break; 6344 } 6345 rc = MDB_SUCCESS; 6346 if (IS_LEAF2(mp)) { 6347 key->mv_size = mc->mc_db->md_pad; 6348 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size); 6349 } else { 6350 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 6351 MDB_GET_KEY(leaf, key); 6352 if (data) { 6353 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6354 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT); 6355 } else { 6356 rc = mdb_node_read(mc, leaf, data); 6357 } 6358 } 6359 } 6360 } 6361 break; 6362 case MDB_GET_BOTH: 6363 case MDB_GET_BOTH_RANGE: 6364 if (data == NULL) { 6365 rc = EINVAL; 6366 break; 6367 } 6368 if (mc->mc_xcursor == NULL) { 6369 rc = MDB_INCOMPATIBLE; 6370 break; 6371 } 6372 /* FALLTHRU */ 6373 case MDB_SET: 6374 case MDB_SET_KEY: 6375 case MDB_SET_RANGE: 6376 if (key == NULL) { 6377 rc = EINVAL; 6378 } else { 6379 rc = mdb_cursor_set(mc, key, data, op, 6380 op == MDB_SET_RANGE ? NULL : &exact); 6381 } 6382 break; 6383 case MDB_GET_MULTIPLE: 6384 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) { 6385 rc = EINVAL; 6386 break; 6387 } 6388 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) { 6389 rc = MDB_INCOMPATIBLE; 6390 break; 6391 } 6392 rc = MDB_SUCCESS; 6393 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) || 6394 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF)) 6395 break; 6396 goto fetchm; 6397 case MDB_NEXT_MULTIPLE: 6398 if (data == NULL) { 6399 rc = EINVAL; 6400 break; 6401 } 6402 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) { 6403 rc = MDB_INCOMPATIBLE; 6404 break; 6405 } 6406 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP); 6407 if (rc == MDB_SUCCESS) { 6408 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) { 6409 MDB_cursor *mx; 6410 fetchm: 6411 mx = &mc->mc_xcursor->mx_cursor; 6412 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) * 6413 mx->mc_db->md_pad; 6414 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]); 6415 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1; 6416 } else { 6417 rc = MDB_NOTFOUND; 6418 } 6419 } 6420 break; 6421 case MDB_PREV_MULTIPLE: 6422 if (data == NULL) { 6423 rc = EINVAL; 6424 break; 6425 } 6426 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) { 6427 rc = MDB_INCOMPATIBLE; 6428 break; 6429 } 6430 if (!(mc->mc_flags & C_INITIALIZED)) 6431 rc = mdb_cursor_last(mc, key, data); 6432 else 6433 rc = MDB_SUCCESS; 6434 if (rc == MDB_SUCCESS) { 6435 MDB_cursor *mx = &mc->mc_xcursor->mx_cursor; 6436 if (mx->mc_flags & C_INITIALIZED) { 6437 rc = mdb_cursor_sibling(mx, 0); 6438 if (rc == MDB_SUCCESS) 6439 goto fetchm; 6440 } else { 6441 rc = MDB_NOTFOUND; 6442 } 6443 } 6444 break; 6445 case MDB_NEXT: 6446 case MDB_NEXT_DUP: 6447 case MDB_NEXT_NODUP: 6448 rc = mdb_cursor_next(mc, key, data, op); 6449 break; 6450 case MDB_PREV: 6451 case MDB_PREV_DUP: 6452 case MDB_PREV_NODUP: 6453 rc = mdb_cursor_prev(mc, key, data, op); 6454 break; 6455 case MDB_FIRST: 6456 rc = mdb_cursor_first(mc, key, data); 6457 break; 6458 case MDB_FIRST_DUP: 6459 mfunc = mdb_cursor_first; 6460 mmove: 6461 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) { 6462 rc = EINVAL; 6463 break; 6464 } 6465 if (mc->mc_xcursor == NULL) { 6466 rc = MDB_INCOMPATIBLE; 6467 break; 6468 } 6469 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) { 6470 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]); 6471 rc = MDB_NOTFOUND; 6472 break; 6473 } 6474 { 6475 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 6476 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6477 MDB_GET_KEY(leaf, key); 6478 rc = mdb_node_read(mc, leaf, data); 6479 break; 6480 } 6481 } 6482 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) { 6483 rc = EINVAL; 6484 break; 6485 } 6486 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL); 6487 break; 6488 case MDB_LAST: 6489 rc = mdb_cursor_last(mc, key, data); 6490 break; 6491 case MDB_LAST_DUP: 6492 mfunc = mdb_cursor_last; 6493 goto mmove; 6494 default: 6495 DPRINTF(("unhandled/unimplemented cursor operation %u", op)); 6496 rc = EINVAL; 6497 break; 6498 } 6499 6500 if (mc->mc_flags & C_DEL) 6501 mc->mc_flags ^= C_DEL; 6502 6503 return rc; 6504 } 6505 6506 /** Touch all the pages in the cursor stack. Set mc_top. 6507 * Makes sure all the pages are writable, before attempting a write operation. 6508 * @param[in] mc The cursor to operate on. 6509 */ 6510 static int 6511 mdb_cursor_touch(MDB_cursor *mc) 6512 { 6513 int rc = MDB_SUCCESS; 6514 6515 if (mc->mc_dbi >= CORE_DBS && !(*mc->mc_dbflag & (DB_DIRTY|DB_DUPDATA))) { 6516 /* Touch DB record of named DB */ 6517 MDB_cursor mc2; 6518 MDB_xcursor mcx; 6519 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi)) 6520 return MDB_BAD_DBI; 6521 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx); 6522 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY); 6523 if (rc) 6524 return rc; 6525 *mc->mc_dbflag |= DB_DIRTY; 6526 } 6527 mc->mc_top = 0; 6528 if (mc->mc_snum) { 6529 do { 6530 rc = mdb_page_touch(mc); 6531 } while (!rc && ++(mc->mc_top) < mc->mc_snum); 6532 mc->mc_top = mc->mc_snum-1; 6533 } 6534 return rc; 6535 } 6536 6537 /** Do not spill pages to disk if txn is getting full, may fail instead */ 6538 #define MDB_NOSPILL 0x8000 6539 6540 int 6541 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data, 6542 unsigned int flags) 6543 { 6544 MDB_env *env; 6545 MDB_node *leaf = NULL; 6546 MDB_page *fp, *mp, *sub_root = NULL; 6547 uint16_t fp_flags; 6548 MDB_val xdata, *rdata, dkey, olddata; 6549 MDB_db dummy; 6550 int do_sub = 0, insert_key, insert_data; 6551 unsigned int mcount = 0, dcount = 0, nospill; 6552 size_t nsize; 6553 int rc, rc2; 6554 unsigned int nflags; 6555 DKBUF; 6556 6557 if (mc == NULL || key == NULL) 6558 return EINVAL; 6559 6560 env = mc->mc_txn->mt_env; 6561 6562 /* Check this first so counter will always be zero on any 6563 * early failures. 6564 */ 6565 if (flags & MDB_MULTIPLE) { 6566 dcount = data[1].mv_size; 6567 data[1].mv_size = 0; 6568 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED)) 6569 return MDB_INCOMPATIBLE; 6570 } 6571 6572 nospill = flags & MDB_NOSPILL; 6573 flags &= ~MDB_NOSPILL; 6574 6575 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED)) 6576 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN; 6577 6578 if (key->mv_size-1 >= ENV_MAXKEY(env)) 6579 return MDB_BAD_VALSIZE; 6580 6581 #if SIZE_MAX > MAXDATASIZE 6582 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE)) 6583 return MDB_BAD_VALSIZE; 6584 #else 6585 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env)) 6586 return MDB_BAD_VALSIZE; 6587 #endif 6588 6589 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u", 6590 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size)); 6591 6592 dkey.mv_size = 0; 6593 6594 if (flags == MDB_CURRENT) { 6595 if (!(mc->mc_flags & C_INITIALIZED)) 6596 return EINVAL; 6597 rc = MDB_SUCCESS; 6598 } else if (mc->mc_db->md_root == P_INVALID) { 6599 /* new database, cursor has nothing to point to */ 6600 mc->mc_snum = 0; 6601 mc->mc_top = 0; 6602 mc->mc_flags &= ~C_INITIALIZED; 6603 rc = MDB_NO_ROOT; 6604 } else { 6605 int exact = 0; 6606 MDB_val d2; 6607 if (flags & MDB_APPEND) { 6608 MDB_val k2; 6609 rc = mdb_cursor_last(mc, &k2, &d2); 6610 if (rc == 0) { 6611 rc = mc->mc_dbx->md_cmp(key, &k2); 6612 if (rc > 0) { 6613 rc = MDB_NOTFOUND; 6614 mc->mc_ki[mc->mc_top]++; 6615 } else { 6616 /* new key is <= last key */ 6617 rc = MDB_KEYEXIST; 6618 } 6619 } 6620 } else { 6621 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact); 6622 } 6623 if ((flags & MDB_NOOVERWRITE) && rc == 0) { 6624 DPRINTF(("duplicate key [%s]", DKEY(key))); 6625 *data = d2; 6626 return MDB_KEYEXIST; 6627 } 6628 if (rc && rc != MDB_NOTFOUND) 6629 return rc; 6630 } 6631 6632 if (mc->mc_flags & C_DEL) 6633 mc->mc_flags ^= C_DEL; 6634 6635 /* Cursor is positioned, check for room in the dirty list */ 6636 if (!nospill) { 6637 if (flags & MDB_MULTIPLE) { 6638 rdata = &xdata; 6639 xdata.mv_size = data->mv_size * dcount; 6640 } else { 6641 rdata = data; 6642 } 6643 if ((rc2 = mdb_page_spill(mc, key, rdata))) 6644 return rc2; 6645 } 6646 6647 if (rc == MDB_NO_ROOT) { 6648 MDB_page *np; 6649 /* new database, write a root leaf page */ 6650 DPUTS("allocating new root leaf page"); 6651 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) { 6652 return rc2; 6653 } 6654 mdb_cursor_push(mc, np); 6655 mc->mc_db->md_root = np->mp_pgno; 6656 mc->mc_db->md_depth++; 6657 *mc->mc_dbflag |= DB_DIRTY; 6658 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED)) 6659 == MDB_DUPFIXED) 6660 np->mp_flags |= P_LEAF2; 6661 mc->mc_flags |= C_INITIALIZED; 6662 } else { 6663 /* make sure all cursor pages are writable */ 6664 rc2 = mdb_cursor_touch(mc); 6665 if (rc2) 6666 return rc2; 6667 } 6668 6669 insert_key = insert_data = rc; 6670 if (insert_key) { 6671 /* The key does not exist */ 6672 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top])); 6673 if ((mc->mc_db->md_flags & MDB_DUPSORT) && 6674 LEAFSIZE(key, data) > env->me_nodemax) 6675 { 6676 /* Too big for a node, insert in sub-DB. Set up an empty 6677 * "old sub-page" for prep_subDB to expand to a full page. 6678 */ 6679 fp_flags = P_LEAF|P_DIRTY; 6680 fp = env->me_pbuf; 6681 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */ 6682 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE); 6683 olddata.mv_size = PAGEHDRSZ; 6684 goto prep_subDB; 6685 } 6686 } else { 6687 /* there's only a key anyway, so this is a no-op */ 6688 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) { 6689 char *ptr; 6690 unsigned int ksize = mc->mc_db->md_pad; 6691 if (key->mv_size != ksize) 6692 return MDB_BAD_VALSIZE; 6693 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize); 6694 memcpy(ptr, key->mv_data, ksize); 6695 fix_parent: 6696 /* if overwriting slot 0 of leaf, need to 6697 * update branch key if there is a parent page 6698 */ 6699 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) { 6700 unsigned short dtop = 1; 6701 mc->mc_top--; 6702 /* slot 0 is always an empty key, find real slot */ 6703 while (mc->mc_top && !mc->mc_ki[mc->mc_top]) { 6704 mc->mc_top--; 6705 dtop++; 6706 } 6707 if (mc->mc_ki[mc->mc_top]) 6708 rc2 = mdb_update_key(mc, key); 6709 else 6710 rc2 = MDB_SUCCESS; 6711 mc->mc_top += dtop; 6712 if (rc2) 6713 return rc2; 6714 } 6715 return MDB_SUCCESS; 6716 } 6717 6718 more: 6719 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 6720 olddata.mv_size = NODEDSZ(leaf); 6721 olddata.mv_data = NODEDATA(leaf); 6722 6723 /* DB has dups? */ 6724 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) { 6725 /* Prepare (sub-)page/sub-DB to accept the new item, 6726 * if needed. fp: old sub-page or a header faking 6727 * it. mp: new (sub-)page. offset: growth in page 6728 * size. xdata: node data with new page or DB. 6729 */ 6730 unsigned i, offset = 0; 6731 mp = fp = xdata.mv_data = env->me_pbuf; 6732 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno; 6733 6734 /* Was a single item before, must convert now */ 6735 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { 6736 MDB_cmp_func *dcmp; 6737 /* Just overwrite the current item */ 6738 if (flags == MDB_CURRENT) 6739 goto current; 6740 dcmp = mc->mc_dbx->md_dcmp; 6741 #if UINT_MAX < SIZE_MAX 6742 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t)) 6743 dcmp = mdb_cmp_clong; 6744 #endif 6745 /* does data match? */ 6746 if (!dcmp(data, &olddata)) { 6747 if (flags & (MDB_NODUPDATA|MDB_APPENDDUP)) 6748 return MDB_KEYEXIST; 6749 /* overwrite it */ 6750 goto current; 6751 } 6752 6753 /* Back up original data item */ 6754 dkey.mv_size = olddata.mv_size; 6755 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size); 6756 6757 /* Make sub-page header for the dup items, with dummy body */ 6758 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP; 6759 fp->mp_lower = (PAGEHDRSZ-PAGEBASE); 6760 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size; 6761 if (mc->mc_db->md_flags & MDB_DUPFIXED) { 6762 fp->mp_flags |= P_LEAF2; 6763 fp->mp_pad = data->mv_size; 6764 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */ 6765 } else { 6766 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) + 6767 (dkey.mv_size & 1) + (data->mv_size & 1); 6768 } 6769 fp->mp_upper = xdata.mv_size - PAGEBASE; 6770 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */ 6771 } else if (leaf->mn_flags & F_SUBDATA) { 6772 /* Data is on sub-DB, just store it */ 6773 flags |= F_DUPDATA|F_SUBDATA; 6774 goto put_sub; 6775 } else { 6776 /* Data is on sub-page */ 6777 fp = olddata.mv_data; 6778 switch (flags) { 6779 default: 6780 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) { 6781 offset = EVEN(NODESIZE + sizeof(indx_t) + 6782 data->mv_size); 6783 break; 6784 } 6785 offset = fp->mp_pad; 6786 if (SIZELEFT(fp) < offset) { 6787 offset *= 4; /* space for 4 more */ 6788 break; 6789 } 6790 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */ 6791 case MDB_CURRENT: 6792 fp->mp_flags |= P_DIRTY; 6793 COPY_PGNO(fp->mp_pgno, mp->mp_pgno); 6794 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp; 6795 flags |= F_DUPDATA; 6796 goto put_sub; 6797 } 6798 xdata.mv_size = olddata.mv_size + offset; 6799 } 6800 6801 fp_flags = fp->mp_flags; 6802 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) { 6803 /* Too big for a sub-page, convert to sub-DB */ 6804 fp_flags &= ~P_SUBP; 6805 prep_subDB: 6806 if (mc->mc_db->md_flags & MDB_DUPFIXED) { 6807 fp_flags |= P_LEAF2; 6808 dummy.md_pad = fp->mp_pad; 6809 dummy.md_flags = MDB_DUPFIXED; 6810 if (mc->mc_db->md_flags & MDB_INTEGERDUP) 6811 dummy.md_flags |= MDB_INTEGERKEY; 6812 } else { 6813 dummy.md_pad = 0; 6814 dummy.md_flags = 0; 6815 } 6816 dummy.md_depth = 1; 6817 dummy.md_branch_pages = 0; 6818 dummy.md_leaf_pages = 1; 6819 dummy.md_overflow_pages = 0; 6820 dummy.md_entries = NUMKEYS(fp); 6821 xdata.mv_size = sizeof(MDB_db); 6822 xdata.mv_data = &dummy; 6823 if ((rc = mdb_page_alloc(mc, 1, &mp))) 6824 return rc; 6825 offset = env->me_psize - olddata.mv_size; 6826 flags |= F_DUPDATA|F_SUBDATA; 6827 dummy.md_root = mp->mp_pgno; 6828 sub_root = mp; 6829 } 6830 if (mp != fp) { 6831 mp->mp_flags = fp_flags | P_DIRTY; 6832 mp->mp_pad = fp->mp_pad; 6833 mp->mp_lower = fp->mp_lower; 6834 mp->mp_upper = fp->mp_upper + offset; 6835 if (fp_flags & P_LEAF2) { 6836 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad); 6837 } else { 6838 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE, 6839 olddata.mv_size - fp->mp_upper - PAGEBASE); 6840 memcpy((char *)(&mp->mp_ptrs), (char *)(&fp->mp_ptrs), NUMKEYS(fp) * sizeof(mp->mp_ptrs[0])); 6841 for (i=0; i<NUMKEYS(fp); i++) 6842 mp->mp_ptrs[i] += offset; 6843 } 6844 } 6845 6846 rdata = &xdata; 6847 flags |= F_DUPDATA; 6848 do_sub = 1; 6849 if (!insert_key) 6850 mdb_node_del(mc, 0); 6851 goto new_sub; 6852 } 6853 current: 6854 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */ 6855 if ((leaf->mn_flags ^ flags) & F_SUBDATA) 6856 return MDB_INCOMPATIBLE; 6857 /* overflow page overwrites need special handling */ 6858 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) { 6859 MDB_page *omp; 6860 pgno_t pg; 6861 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize); 6862 6863 memcpy(&pg, olddata.mv_data, sizeof(pg)); 6864 if ((rc2 = mdb_page_get(mc, pg, &omp, &level)) != 0) 6865 return rc2; 6866 ovpages = omp->mp_pages; 6867 6868 /* Is the ov page large enough? */ 6869 if (ovpages >= dpages) { 6870 if (!(omp->mp_flags & P_DIRTY) && 6871 (level || (env->me_flags & MDB_WRITEMAP))) 6872 { 6873 rc = mdb_page_unspill(mc->mc_txn, omp, &omp); 6874 if (rc) 6875 return rc; 6876 level = 0; /* dirty in this txn or clean */ 6877 } 6878 /* Is it dirty? */ 6879 if (omp->mp_flags & P_DIRTY) { 6880 /* yes, overwrite it. Note in this case we don't 6881 * bother to try shrinking the page if the new data 6882 * is smaller than the overflow threshold. 6883 */ 6884 if (level > 1) { 6885 /* It is writable only in a parent txn */ 6886 size_t sz = (size_t) env->me_psize * ovpages, off; 6887 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages); 6888 MDB_ID2 id2; 6889 if (!np) 6890 return ENOMEM; 6891 id2.mid = pg; 6892 id2.mptr = np; 6893 /* Note - this page is already counted in parent's dirty_room */ 6894 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2); 6895 mdb_cassert(mc, rc2 == 0); 6896 /* Currently we make the page look as with put() in the 6897 * parent txn, in case the user peeks at MDB_RESERVEd 6898 * or unused parts. Some users treat ovpages specially. 6899 */ 6900 if (!(flags & MDB_RESERVE)) { 6901 /* Skip the part where LMDB will put *data. 6902 * Copy end of page, adjusting alignment so 6903 * compiler may copy words instead of bytes. 6904 */ 6905 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t); 6906 memcpy((size_t *)((char *)np + off), 6907 (size_t *)((char *)omp + off), sz - off); 6908 sz = PAGEHDRSZ; 6909 } 6910 memcpy(np, omp, sz); /* Copy beginning of page */ 6911 omp = np; 6912 } 6913 SETDSZ(leaf, data->mv_size); 6914 if (F_ISSET(flags, MDB_RESERVE)) 6915 data->mv_data = METADATA(omp); 6916 else 6917 memcpy(METADATA(omp), data->mv_data, data->mv_size); 6918 return MDB_SUCCESS; 6919 } 6920 } 6921 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS) 6922 return rc2; 6923 } else if (data->mv_size == olddata.mv_size) { 6924 /* same size, just replace it. Note that we could 6925 * also reuse this node if the new data is smaller, 6926 * but instead we opt to shrink the node in that case. 6927 */ 6928 if (F_ISSET(flags, MDB_RESERVE)) 6929 data->mv_data = olddata.mv_data; 6930 else if (!(mc->mc_flags & C_SUB)) 6931 memcpy(olddata.mv_data, data->mv_data, data->mv_size); 6932 else { 6933 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size); 6934 goto fix_parent; 6935 } 6936 return MDB_SUCCESS; 6937 } 6938 mdb_node_del(mc, 0); 6939 } 6940 6941 rdata = data; 6942 6943 new_sub: 6944 nflags = flags & NODE_ADD_FLAGS; 6945 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata); 6946 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) { 6947 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA ) 6948 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */ 6949 if (!insert_key) 6950 nflags |= MDB_SPLIT_REPLACE; 6951 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags); 6952 } else { 6953 /* There is room already in this leaf page. */ 6954 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags); 6955 if (rc == 0) { 6956 /* Adjust other cursors pointing to mp */ 6957 MDB_cursor *m2, *m3; 6958 MDB_dbi dbi = mc->mc_dbi; 6959 unsigned i = mc->mc_top; 6960 MDB_page *mp = mc->mc_pg[i]; 6961 6962 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 6963 if (mc->mc_flags & C_SUB) 6964 m3 = &m2->mc_xcursor->mx_cursor; 6965 else 6966 m3 = m2; 6967 if (m3 == mc || m3->mc_snum < mc->mc_snum || m3->mc_pg[i] != mp) continue; 6968 if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) { 6969 m3->mc_ki[i]++; 6970 } 6971 XCURSOR_REFRESH(m3, i, mp); 6972 } 6973 } 6974 } 6975 6976 if (rc == MDB_SUCCESS) { 6977 /* Now store the actual data in the child DB. Note that we're 6978 * storing the user data in the keys field, so there are strict 6979 * size limits on dupdata. The actual data fields of the child 6980 * DB are all zero size. 6981 */ 6982 if (do_sub) { 6983 int xflags, new_dupdata; 6984 size_t ecount; 6985 put_sub: 6986 xdata.mv_size = 0; 6987 xdata.mv_data = ""; 6988 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 6989 if (flags & MDB_CURRENT) { 6990 xflags = MDB_CURRENT|MDB_NOSPILL; 6991 } else { 6992 mdb_xcursor_init1(mc, leaf); 6993 xflags = (flags & MDB_NODUPDATA) ? 6994 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL; 6995 } 6996 if (sub_root) 6997 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root; 6998 new_dupdata = (int)dkey.mv_size; 6999 /* converted, write the original data first */ 7000 if (dkey.mv_size) { 7001 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags); 7002 if (rc) 7003 goto bad_sub; 7004 /* we've done our job */ 7005 dkey.mv_size = 0; 7006 } 7007 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) { 7008 /* Adjust other cursors pointing to mp */ 7009 MDB_cursor *m2; 7010 MDB_xcursor *mx = mc->mc_xcursor; 7011 unsigned i = mc->mc_top; 7012 MDB_page *mp = mc->mc_pg[i]; 7013 7014 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) { 7015 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue; 7016 if (!(m2->mc_flags & C_INITIALIZED)) continue; 7017 if (m2->mc_pg[i] == mp) { 7018 if (m2->mc_ki[i] == mc->mc_ki[i]) { 7019 mdb_xcursor_init2(m2, mx, new_dupdata); 7020 } else if (!insert_key) { 7021 XCURSOR_REFRESH(m2, i, mp); 7022 } 7023 } 7024 } 7025 } 7026 ecount = mc->mc_xcursor->mx_db.md_entries; 7027 if (flags & MDB_APPENDDUP) 7028 xflags |= MDB_APPEND; 7029 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags); 7030 if (flags & F_SUBDATA) { 7031 void *db = NODEDATA(leaf); 7032 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db)); 7033 } 7034 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount; 7035 } 7036 /* Increment count unless we just replaced an existing item. */ 7037 if (insert_data) 7038 mc->mc_db->md_entries++; 7039 if (insert_key) { 7040 /* Invalidate txn if we created an empty sub-DB */ 7041 if (rc) 7042 goto bad_sub; 7043 /* If we succeeded and the key didn't exist before, 7044 * make sure the cursor is marked valid. 7045 */ 7046 mc->mc_flags |= C_INITIALIZED; 7047 } 7048 if (flags & MDB_MULTIPLE) { 7049 if (!rc) { 7050 mcount++; 7051 /* let caller know how many succeeded, if any */ 7052 data[1].mv_size = mcount; 7053 if (mcount < dcount) { 7054 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size; 7055 insert_key = insert_data = 0; 7056 goto more; 7057 } 7058 } 7059 } 7060 return rc; 7061 bad_sub: 7062 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */ 7063 rc = MDB_CORRUPTED; 7064 } 7065 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 7066 return rc; 7067 } 7068 7069 int 7070 mdb_cursor_del(MDB_cursor *mc, unsigned int flags) 7071 { 7072 MDB_node *leaf; 7073 MDB_page *mp; 7074 int rc; 7075 7076 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED)) 7077 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN; 7078 7079 if (!(mc->mc_flags & C_INITIALIZED)) 7080 return EINVAL; 7081 7082 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) 7083 return MDB_NOTFOUND; 7084 7085 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL))) 7086 return rc; 7087 7088 rc = mdb_cursor_touch(mc); 7089 if (rc) 7090 return rc; 7091 7092 mp = mc->mc_pg[mc->mc_top]; 7093 if (IS_LEAF2(mp)) 7094 goto del_key; 7095 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 7096 7097 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { 7098 if (flags & MDB_NODUPDATA) { 7099 /* mdb_cursor_del0() will subtract the final entry */ 7100 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1; 7101 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED; 7102 } else { 7103 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) { 7104 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf); 7105 } 7106 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL); 7107 if (rc) 7108 return rc; 7109 /* If sub-DB still has entries, we're done */ 7110 if (mc->mc_xcursor->mx_db.md_entries) { 7111 if (leaf->mn_flags & F_SUBDATA) { 7112 /* update subDB info */ 7113 void *db = NODEDATA(leaf); 7114 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db)); 7115 } else { 7116 MDB_cursor *m2; 7117 /* shrink fake page */ 7118 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]); 7119 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); 7120 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf); 7121 /* fix other sub-DB cursors pointed at fake pages on this page */ 7122 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) { 7123 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue; 7124 if (!(m2->mc_flags & C_INITIALIZED)) continue; 7125 if (m2->mc_pg[mc->mc_top] == mp) { 7126 XCURSOR_REFRESH(m2, mc->mc_top, mp); 7127 } 7128 } 7129 } 7130 mc->mc_db->md_entries--; 7131 return rc; 7132 } else { 7133 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED; 7134 } 7135 /* otherwise fall thru and delete the sub-DB */ 7136 } 7137 7138 if (leaf->mn_flags & F_SUBDATA) { 7139 /* add all the child DB's pages to the free list */ 7140 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0); 7141 if (rc) 7142 goto fail; 7143 } 7144 } 7145 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */ 7146 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) { 7147 rc = MDB_INCOMPATIBLE; 7148 goto fail; 7149 } 7150 7151 /* add overflow pages to free list */ 7152 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) { 7153 MDB_page *omp; 7154 pgno_t pg; 7155 7156 memcpy(&pg, NODEDATA(leaf), sizeof(pg)); 7157 if ((rc = mdb_page_get(mc, pg, &omp, NULL)) || 7158 (rc = mdb_ovpage_free(mc, omp))) 7159 goto fail; 7160 } 7161 7162 del_key: 7163 return mdb_cursor_del0(mc); 7164 7165 fail: 7166 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 7167 return rc; 7168 } 7169 7170 /** Allocate and initialize new pages for a database. 7171 * Set #MDB_TXN_ERROR on failure. 7172 * @param[in] mc a cursor on the database being added to. 7173 * @param[in] flags flags defining what type of page is being allocated. 7174 * @param[in] num the number of pages to allocate. This is usually 1, 7175 * unless allocating overflow pages for a large record. 7176 * @param[out] mp Address of a page, or NULL on failure. 7177 * @return 0 on success, non-zero on failure. 7178 */ 7179 static int 7180 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp) 7181 { 7182 MDB_page *np; 7183 int rc; 7184 7185 if ((rc = mdb_page_alloc(mc, num, &np))) 7186 return rc; 7187 DPRINTF(("allocated new mpage %"Z"u, page size %u", 7188 np->mp_pgno, mc->mc_txn->mt_env->me_psize)); 7189 np->mp_flags = flags | P_DIRTY; 7190 np->mp_lower = (PAGEHDRSZ-PAGEBASE); 7191 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE; 7192 7193 if (IS_BRANCH(np)) 7194 mc->mc_db->md_branch_pages++; 7195 else if (IS_LEAF(np)) 7196 mc->mc_db->md_leaf_pages++; 7197 else if (IS_OVERFLOW(np)) { 7198 mc->mc_db->md_overflow_pages += num; 7199 np->mp_pages = num; 7200 } 7201 *mp = np; 7202 7203 return 0; 7204 } 7205 7206 /** Calculate the size of a leaf node. 7207 * The size depends on the environment's page size; if a data item 7208 * is too large it will be put onto an overflow page and the node 7209 * size will only include the key and not the data. Sizes are always 7210 * rounded up to an even number of bytes, to guarantee 2-byte alignment 7211 * of the #MDB_node headers. 7212 * @param[in] env The environment handle. 7213 * @param[in] key The key for the node. 7214 * @param[in] data The data for the node. 7215 * @return The number of bytes needed to store the node. 7216 */ 7217 static size_t 7218 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data) 7219 { 7220 size_t sz; 7221 7222 sz = LEAFSIZE(key, data); 7223 if (sz > env->me_nodemax) { 7224 /* put on overflow page */ 7225 sz -= data->mv_size - sizeof(pgno_t); 7226 } 7227 7228 return EVEN(sz + sizeof(indx_t)); 7229 } 7230 7231 /** Calculate the size of a branch node. 7232 * The size should depend on the environment's page size but since 7233 * we currently don't support spilling large keys onto overflow 7234 * pages, it's simply the size of the #MDB_node header plus the 7235 * size of the key. Sizes are always rounded up to an even number 7236 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers. 7237 * @param[in] env The environment handle. 7238 * @param[in] key The key for the node. 7239 * @return The number of bytes needed to store the node. 7240 */ 7241 static size_t 7242 mdb_branch_size(MDB_env *env, MDB_val *key) 7243 { 7244 size_t sz; 7245 7246 sz = INDXSIZE(key); 7247 if (sz > env->me_nodemax) { 7248 /* put on overflow page */ 7249 /* not implemented */ 7250 /* sz -= key->size - sizeof(pgno_t); */ 7251 } 7252 7253 return sz + sizeof(indx_t); 7254 } 7255 7256 /** Add a node to the page pointed to by the cursor. 7257 * Set #MDB_TXN_ERROR on failure. 7258 * @param[in] mc The cursor for this operation. 7259 * @param[in] indx The index on the page where the new node should be added. 7260 * @param[in] key The key for the new node. 7261 * @param[in] data The data for the new node, if any. 7262 * @param[in] pgno The page number, if adding a branch node. 7263 * @param[in] flags Flags for the node. 7264 * @return 0 on success, non-zero on failure. Possible errors are: 7265 * <ul> 7266 * <li>ENOMEM - failed to allocate overflow pages for the node. 7267 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error 7268 * should never happen since all callers already calculate the 7269 * page's free space before calling this function. 7270 * </ul> 7271 */ 7272 static int 7273 mdb_node_add(MDB_cursor *mc, indx_t indx, 7274 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags) 7275 { 7276 unsigned int i; 7277 size_t node_size = NODESIZE; 7278 ssize_t room; 7279 indx_t ofs; 7280 MDB_node *node; 7281 MDB_page *mp = mc->mc_pg[mc->mc_top]; 7282 MDB_page *ofp = NULL; /* overflow page */ 7283 void *ndata; 7284 DKBUF; 7285 7286 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower); 7287 7288 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]", 7289 IS_LEAF(mp) ? "leaf" : "branch", 7290 IS_SUBP(mp) ? "sub-" : "", 7291 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0, 7292 key ? key->mv_size : 0, key ? DKEY(key) : "null")); 7293 7294 if (IS_LEAF2(mp)) { 7295 /* Move higher keys up one slot. */ 7296 int ksize = mc->mc_db->md_pad, dif; 7297 char *ptr = LEAF2KEY(mp, indx, ksize); 7298 dif = NUMKEYS(mp) - indx; 7299 if (dif > 0) 7300 memmove(ptr+ksize, ptr, dif*ksize); 7301 /* insert new key */ 7302 memcpy(ptr, key->mv_data, ksize); 7303 7304 /* Just using these for counting */ 7305 mp->mp_lower += sizeof(indx_t); 7306 mp->mp_upper -= ksize - sizeof(indx_t); 7307 return MDB_SUCCESS; 7308 } 7309 7310 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t); 7311 if (key != NULL) 7312 node_size += key->mv_size; 7313 if (IS_LEAF(mp)) { 7314 mdb_cassert(mc, key && data); 7315 if (F_ISSET(flags, F_BIGDATA)) { 7316 /* Data already on overflow page. */ 7317 node_size += sizeof(pgno_t); 7318 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) { 7319 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize); 7320 int rc; 7321 /* Put data on overflow page. */ 7322 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page", 7323 data->mv_size, node_size+data->mv_size)); 7324 node_size = EVEN(node_size + sizeof(pgno_t)); 7325 if ((ssize_t)node_size > room) 7326 goto full; 7327 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp))) 7328 return rc; 7329 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno)); 7330 flags |= F_BIGDATA; 7331 goto update; 7332 } else { 7333 node_size += data->mv_size; 7334 } 7335 } 7336 node_size = EVEN(node_size); 7337 if ((ssize_t)node_size > room) 7338 goto full; 7339 7340 update: 7341 /* Move higher pointers up one slot. */ 7342 for (i = NUMKEYS(mp); i > indx; i--) 7343 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1]; 7344 7345 /* Adjust free space offsets. */ 7346 ofs = mp->mp_upper - node_size; 7347 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t)); 7348 mp->mp_ptrs[indx] = ofs; 7349 mp->mp_upper = ofs; 7350 mp->mp_lower += sizeof(indx_t); 7351 7352 /* Write the node data. */ 7353 node = NODEPTR(mp, indx); 7354 node->mn_ksize = (key == NULL) ? 0 : key->mv_size; 7355 node->mn_flags = flags; 7356 if (IS_LEAF(mp)) 7357 SETDSZ(node,data->mv_size); 7358 else 7359 SETPGNO(node,pgno); 7360 7361 if (key) 7362 memcpy(NODEKEY(node), key->mv_data, key->mv_size); 7363 7364 if (IS_LEAF(mp)) { 7365 ndata = NODEDATA(node); 7366 if (ofp == NULL) { 7367 if (F_ISSET(flags, F_BIGDATA)) 7368 memcpy(ndata, data->mv_data, sizeof(pgno_t)); 7369 else if (F_ISSET(flags, MDB_RESERVE)) 7370 data->mv_data = ndata; 7371 else 7372 memcpy(ndata, data->mv_data, data->mv_size); 7373 } else { 7374 memcpy(ndata, &ofp->mp_pgno, sizeof(pgno_t)); 7375 ndata = METADATA(ofp); 7376 if (F_ISSET(flags, MDB_RESERVE)) 7377 data->mv_data = ndata; 7378 else 7379 memcpy(ndata, data->mv_data, data->mv_size); 7380 } 7381 } 7382 7383 return MDB_SUCCESS; 7384 7385 full: 7386 DPRINTF(("not enough room in page %"Z"u, got %u ptrs", 7387 mdb_dbg_pgno(mp), NUMKEYS(mp))); 7388 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room)); 7389 DPRINTF(("node size = %"Z"u", node_size)); 7390 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 7391 return MDB_PAGE_FULL; 7392 } 7393 7394 /** Delete the specified node from a page. 7395 * @param[in] mc Cursor pointing to the node to delete. 7396 * @param[in] ksize The size of a node. Only used if the page is 7397 * part of a #MDB_DUPFIXED database. 7398 */ 7399 static void 7400 mdb_node_del(MDB_cursor *mc, int ksize) 7401 { 7402 MDB_page *mp = mc->mc_pg[mc->mc_top]; 7403 indx_t indx = mc->mc_ki[mc->mc_top]; 7404 unsigned int sz; 7405 indx_t i, j, numkeys, ptr; 7406 MDB_node *node; 7407 char *base; 7408 7409 DPRINTF(("delete node %u on %s page %"Z"u", indx, 7410 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp))); 7411 numkeys = NUMKEYS(mp); 7412 mdb_cassert(mc, indx < numkeys); 7413 7414 if (IS_LEAF2(mp)) { 7415 int x = numkeys - 1 - indx; 7416 base = LEAF2KEY(mp, indx, ksize); 7417 if (x) 7418 memmove(base, base + ksize, x * ksize); 7419 mp->mp_lower -= sizeof(indx_t); 7420 mp->mp_upper += ksize - sizeof(indx_t); 7421 return; 7422 } 7423 7424 node = NODEPTR(mp, indx); 7425 sz = NODESIZE + node->mn_ksize; 7426 if (IS_LEAF(mp)) { 7427 if (F_ISSET(node->mn_flags, F_BIGDATA)) 7428 sz += sizeof(pgno_t); 7429 else 7430 sz += NODEDSZ(node); 7431 } 7432 sz = EVEN(sz); 7433 7434 ptr = mp->mp_ptrs[indx]; 7435 for (i = j = 0; i < numkeys; i++) { 7436 if (i != indx) { 7437 mp->mp_ptrs[j] = mp->mp_ptrs[i]; 7438 if (mp->mp_ptrs[i] < ptr) 7439 mp->mp_ptrs[j] += sz; 7440 j++; 7441 } 7442 } 7443 7444 base = (char *)mp + mp->mp_upper + PAGEBASE; 7445 memmove(base + sz, base, ptr - mp->mp_upper); 7446 7447 mp->mp_lower -= sizeof(indx_t); 7448 mp->mp_upper += sz; 7449 } 7450 7451 /** Compact the main page after deleting a node on a subpage. 7452 * @param[in] mp The main page to operate on. 7453 * @param[in] indx The index of the subpage on the main page. 7454 */ 7455 static void 7456 mdb_node_shrink(MDB_page *mp, indx_t indx) 7457 { 7458 MDB_node *node; 7459 MDB_page *sp, *xp; 7460 char *base; 7461 indx_t delta, nsize, len, ptr; 7462 int i; 7463 7464 node = NODEPTR(mp, indx); 7465 sp = (MDB_page *)NODEDATA(node); 7466 delta = SIZELEFT(sp); 7467 nsize = NODEDSZ(node) - delta; 7468 7469 /* Prepare to shift upward, set len = length(subpage part to shift) */ 7470 if (IS_LEAF2(sp)) { 7471 len = nsize; 7472 if (nsize & 1) 7473 return; /* do not make the node uneven-sized */ 7474 } else { 7475 xp = (MDB_page *)((char *)sp + delta); /* destination subpage */ 7476 for (i = NUMKEYS(sp); --i >= 0; ) 7477 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta; 7478 len = PAGEHDRSZ; 7479 } 7480 sp->mp_upper = sp->mp_lower; 7481 COPY_PGNO(sp->mp_pgno, mp->mp_pgno); 7482 SETDSZ(node, nsize); 7483 7484 /* Shift <lower nodes...initial part of subpage> upward */ 7485 base = (char *)mp + mp->mp_upper + PAGEBASE; 7486 memmove(base + delta, base, (char *)sp + len - base); 7487 7488 ptr = mp->mp_ptrs[indx]; 7489 for (i = NUMKEYS(mp); --i >= 0; ) { 7490 if (mp->mp_ptrs[i] <= ptr) 7491 mp->mp_ptrs[i] += delta; 7492 } 7493 mp->mp_upper += delta; 7494 } 7495 7496 /** Initial setup of a sorted-dups cursor. 7497 * Sorted duplicates are implemented as a sub-database for the given key. 7498 * The duplicate data items are actually keys of the sub-database. 7499 * Operations on the duplicate data items are performed using a sub-cursor 7500 * initialized when the sub-database is first accessed. This function does 7501 * the preliminary setup of the sub-cursor, filling in the fields that 7502 * depend only on the parent DB. 7503 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized. 7504 */ 7505 static void 7506 mdb_xcursor_init0(MDB_cursor *mc) 7507 { 7508 MDB_xcursor *mx = mc->mc_xcursor; 7509 7510 mx->mx_cursor.mc_xcursor = NULL; 7511 mx->mx_cursor.mc_txn = mc->mc_txn; 7512 mx->mx_cursor.mc_db = &mx->mx_db; 7513 mx->mx_cursor.mc_dbx = &mx->mx_dbx; 7514 mx->mx_cursor.mc_dbi = mc->mc_dbi; 7515 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag; 7516 mx->mx_cursor.mc_snum = 0; 7517 mx->mx_cursor.mc_top = 0; 7518 mx->mx_cursor.mc_flags = C_SUB; 7519 mx->mx_dbx.md_name.mv_size = 0; 7520 mx->mx_dbx.md_name.mv_data = NULL; 7521 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp; 7522 mx->mx_dbx.md_dcmp = NULL; 7523 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel; 7524 } 7525 7526 /** Final setup of a sorted-dups cursor. 7527 * Sets up the fields that depend on the data from the main cursor. 7528 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized. 7529 * @param[in] node The data containing the #MDB_db record for the 7530 * sorted-dup database. 7531 */ 7532 static void 7533 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node) 7534 { 7535 MDB_xcursor *mx = mc->mc_xcursor; 7536 7537 if (node->mn_flags & F_SUBDATA) { 7538 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db)); 7539 mx->mx_cursor.mc_pg[0] = 0; 7540 mx->mx_cursor.mc_snum = 0; 7541 mx->mx_cursor.mc_top = 0; 7542 mx->mx_cursor.mc_flags = C_SUB; 7543 } else { 7544 MDB_page *fp = NODEDATA(node); 7545 mx->mx_db.md_pad = 0; 7546 mx->mx_db.md_flags = 0; 7547 mx->mx_db.md_depth = 1; 7548 mx->mx_db.md_branch_pages = 0; 7549 mx->mx_db.md_leaf_pages = 1; 7550 mx->mx_db.md_overflow_pages = 0; 7551 mx->mx_db.md_entries = NUMKEYS(fp); 7552 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno); 7553 mx->mx_cursor.mc_snum = 1; 7554 mx->mx_cursor.mc_top = 0; 7555 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB; 7556 mx->mx_cursor.mc_pg[0] = fp; 7557 mx->mx_cursor.mc_ki[0] = 0; 7558 if (mc->mc_db->md_flags & MDB_DUPFIXED) { 7559 mx->mx_db.md_flags = MDB_DUPFIXED; 7560 mx->mx_db.md_pad = fp->mp_pad; 7561 if (mc->mc_db->md_flags & MDB_INTEGERDUP) 7562 mx->mx_db.md_flags |= MDB_INTEGERKEY; 7563 } 7564 } 7565 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi, 7566 mx->mx_db.md_root)); 7567 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA; 7568 #if UINT_MAX < SIZE_MAX 7569 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t)) 7570 mx->mx_dbx.md_cmp = mdb_cmp_clong; 7571 #endif 7572 } 7573 7574 7575 /** Fixup a sorted-dups cursor due to underlying update. 7576 * Sets up some fields that depend on the data from the main cursor. 7577 * Almost the same as init1, but skips initialization steps if the 7578 * xcursor had already been used. 7579 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up. 7580 * @param[in] src_mx The xcursor of an up-to-date cursor. 7581 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item. 7582 */ 7583 static void 7584 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata) 7585 { 7586 MDB_xcursor *mx = mc->mc_xcursor; 7587 7588 if (new_dupdata) { 7589 mx->mx_cursor.mc_snum = 1; 7590 mx->mx_cursor.mc_top = 0; 7591 mx->mx_cursor.mc_flags |= C_INITIALIZED; 7592 mx->mx_cursor.mc_ki[0] = 0; 7593 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DUPDATA; 7594 #if UINT_MAX < SIZE_MAX 7595 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp; 7596 #endif 7597 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) { 7598 return; 7599 } 7600 mx->mx_db = src_mx->mx_db; 7601 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0]; 7602 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi, 7603 mx->mx_db.md_root)); 7604 } 7605 7606 /** Initialize a cursor for a given transaction and database. */ 7607 static void 7608 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx) 7609 { 7610 mc->mc_next = NULL; 7611 mc->mc_backup = NULL; 7612 mc->mc_dbi = dbi; 7613 mc->mc_txn = txn; 7614 mc->mc_db = &txn->mt_dbs[dbi]; 7615 mc->mc_dbx = &txn->mt_dbxs[dbi]; 7616 mc->mc_dbflag = &txn->mt_dbflags[dbi]; 7617 mc->mc_snum = 0; 7618 mc->mc_top = 0; 7619 mc->mc_pg[0] = 0; 7620 mc->mc_ki[0] = 0; 7621 mc->mc_flags = 0; 7622 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) { 7623 mdb_tassert(txn, mx != NULL); 7624 mc->mc_xcursor = mx; 7625 mdb_xcursor_init0(mc); 7626 } else { 7627 mc->mc_xcursor = NULL; 7628 } 7629 if (*mc->mc_dbflag & DB_STALE) { 7630 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY); 7631 } 7632 } 7633 7634 int 7635 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret) 7636 { 7637 MDB_cursor *mc; 7638 size_t size = sizeof(MDB_cursor); 7639 7640 if (!ret || !TXN_DBI_EXIST(txn, dbi, DB_VALID)) 7641 return EINVAL; 7642 7643 if (txn->mt_flags & MDB_TXN_BLOCKED) 7644 return MDB_BAD_TXN; 7645 7646 if (dbi == FREE_DBI && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) 7647 return EINVAL; 7648 7649 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) 7650 size += sizeof(MDB_xcursor); 7651 7652 if ((mc = malloc(size)) != NULL) { 7653 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1)); 7654 if (txn->mt_cursors) { 7655 mc->mc_next = txn->mt_cursors[dbi]; 7656 txn->mt_cursors[dbi] = mc; 7657 mc->mc_flags |= C_UNTRACK; 7658 } 7659 } else { 7660 return ENOMEM; 7661 } 7662 7663 *ret = mc; 7664 7665 return MDB_SUCCESS; 7666 } 7667 7668 int 7669 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc) 7670 { 7671 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi, DB_VALID)) 7672 return EINVAL; 7673 7674 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors) 7675 return EINVAL; 7676 7677 if (txn->mt_flags & MDB_TXN_BLOCKED) 7678 return MDB_BAD_TXN; 7679 7680 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor); 7681 return MDB_SUCCESS; 7682 } 7683 7684 /* Return the count of duplicate data items for the current key */ 7685 int 7686 mdb_cursor_count(MDB_cursor *mc, size_t *countp) 7687 { 7688 MDB_node *leaf; 7689 7690 if (mc == NULL || countp == NULL) 7691 return EINVAL; 7692 7693 if (mc->mc_xcursor == NULL) 7694 return MDB_INCOMPATIBLE; 7695 7696 if (mc->mc_txn->mt_flags & MDB_TXN_BLOCKED) 7697 return MDB_BAD_TXN; 7698 7699 if (!(mc->mc_flags & C_INITIALIZED)) 7700 return EINVAL; 7701 7702 if (!mc->mc_snum) 7703 return MDB_NOTFOUND; 7704 7705 if (mc->mc_flags & C_EOF) { 7706 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) 7707 return MDB_NOTFOUND; 7708 mc->mc_flags ^= C_EOF; 7709 } 7710 7711 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 7712 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { 7713 *countp = 1; 7714 } else { 7715 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) 7716 return EINVAL; 7717 7718 *countp = mc->mc_xcursor->mx_db.md_entries; 7719 } 7720 return MDB_SUCCESS; 7721 } 7722 7723 void 7724 mdb_cursor_close(MDB_cursor *mc) 7725 { 7726 if (mc && !mc->mc_backup) { 7727 /* remove from txn, if tracked */ 7728 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) { 7729 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi]; 7730 while (*prev && *prev != mc) prev = &(*prev)->mc_next; 7731 if (*prev == mc) 7732 *prev = mc->mc_next; 7733 } 7734 free(mc); 7735 } 7736 } 7737 7738 MDB_txn * 7739 mdb_cursor_txn(MDB_cursor *mc) 7740 { 7741 if (!mc) return NULL; 7742 return mc->mc_txn; 7743 } 7744 7745 MDB_dbi 7746 mdb_cursor_dbi(MDB_cursor *mc) 7747 { 7748 return mc->mc_dbi; 7749 } 7750 7751 /** Replace the key for a branch node with a new key. 7752 * Set #MDB_TXN_ERROR on failure. 7753 * @param[in] mc Cursor pointing to the node to operate on. 7754 * @param[in] key The new key to use. 7755 * @return 0 on success, non-zero on failure. 7756 */ 7757 static int 7758 mdb_update_key(MDB_cursor *mc, MDB_val *key) 7759 { 7760 MDB_page *mp; 7761 MDB_node *node; 7762 char *base; 7763 size_t len; 7764 int delta, ksize, oksize; 7765 indx_t ptr, i, numkeys, indx; 7766 DKBUF; 7767 7768 indx = mc->mc_ki[mc->mc_top]; 7769 mp = mc->mc_pg[mc->mc_top]; 7770 node = NODEPTR(mp, indx); 7771 ptr = mp->mp_ptrs[indx]; 7772 #if MDB_DEBUG 7773 { 7774 MDB_val k2; 7775 char kbuf2[DKBUF_MAXKEYSIZE*2+1]; 7776 k2.mv_data = NODEKEY(node); 7777 k2.mv_size = node->mn_ksize; 7778 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u", 7779 indx, ptr, 7780 mdb_dkey(&k2, kbuf2), 7781 DKEY(key), 7782 mp->mp_pgno)); 7783 } 7784 #endif 7785 7786 /* Sizes must be 2-byte aligned. */ 7787 ksize = EVEN(key->mv_size); 7788 oksize = EVEN(node->mn_ksize); 7789 delta = ksize - oksize; 7790 7791 /* Shift node contents if EVEN(key length) changed. */ 7792 if (delta) { 7793 if (delta > 0 && SIZELEFT(mp) < delta) { 7794 pgno_t pgno; 7795 /* not enough space left, do a delete and split */ 7796 DPRINTF(("Not enough room, delta = %d, splitting...", delta)); 7797 pgno = NODEPGNO(node); 7798 mdb_node_del(mc, 0); 7799 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE); 7800 } 7801 7802 numkeys = NUMKEYS(mp); 7803 for (i = 0; i < numkeys; i++) { 7804 if (mp->mp_ptrs[i] <= ptr) 7805 mp->mp_ptrs[i] -= delta; 7806 } 7807 7808 base = (char *)mp + mp->mp_upper + PAGEBASE; 7809 len = ptr - mp->mp_upper + NODESIZE; 7810 memmove(base - delta, base, len); 7811 mp->mp_upper -= delta; 7812 7813 node = NODEPTR(mp, indx); 7814 } 7815 7816 /* But even if no shift was needed, update ksize */ 7817 if (node->mn_ksize != key->mv_size) 7818 node->mn_ksize = key->mv_size; 7819 7820 if (key->mv_size) 7821 memcpy(NODEKEY(node), key->mv_data, key->mv_size); 7822 7823 return MDB_SUCCESS; 7824 } 7825 7826 static void 7827 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst); 7828 7829 /** Perform \b act while tracking temporary cursor \b mn */ 7830 #define WITH_CURSOR_TRACKING(mn, act) do { \ 7831 MDB_cursor dummy, *tracked, **tp = &(mn).mc_txn->mt_cursors[mn.mc_dbi]; \ 7832 if ((mn).mc_flags & C_SUB) { \ 7833 dummy.mc_flags = C_INITIALIZED; \ 7834 dummy.mc_xcursor = (MDB_xcursor *)&(mn); \ 7835 tracked = &dummy; \ 7836 } else { \ 7837 tracked = &(mn); \ 7838 } \ 7839 tracked->mc_next = *tp; \ 7840 *tp = tracked; \ 7841 { act; } \ 7842 *tp = tracked->mc_next; \ 7843 } while (0) 7844 7845 /** Move a node from csrc to cdst. 7846 */ 7847 static int 7848 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft) 7849 { 7850 MDB_node *srcnode; 7851 MDB_val key, data; 7852 pgno_t srcpg; 7853 MDB_cursor mn; 7854 int rc; 7855 unsigned short flags; 7856 7857 DKBUF; 7858 7859 /* Mark src and dst as dirty. */ 7860 if ((rc = mdb_page_touch(csrc)) || 7861 (rc = mdb_page_touch(cdst))) 7862 return rc; 7863 7864 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) { 7865 key.mv_size = csrc->mc_db->md_pad; 7866 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size); 7867 data.mv_size = 0; 7868 data.mv_data = NULL; 7869 srcpg = 0; 7870 flags = 0; 7871 } else { 7872 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]); 7873 mdb_cassert(csrc, !((size_t)srcnode & 1)); 7874 srcpg = NODEPGNO(srcnode); 7875 flags = srcnode->mn_flags; 7876 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) { 7877 unsigned int snum = csrc->mc_snum; 7878 MDB_node *s2; 7879 /* must find the lowest key below src */ 7880 rc = mdb_page_search_lowest(csrc); 7881 if (rc) 7882 return rc; 7883 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) { 7884 key.mv_size = csrc->mc_db->md_pad; 7885 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size); 7886 } else { 7887 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); 7888 key.mv_size = NODEKSZ(s2); 7889 key.mv_data = NODEKEY(s2); 7890 } 7891 csrc->mc_snum = snum--; 7892 csrc->mc_top = snum; 7893 } else { 7894 key.mv_size = NODEKSZ(srcnode); 7895 key.mv_data = NODEKEY(srcnode); 7896 } 7897 data.mv_size = NODEDSZ(srcnode); 7898 data.mv_data = NODEDATA(srcnode); 7899 } 7900 mn.mc_xcursor = NULL; 7901 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) { 7902 unsigned int snum = cdst->mc_snum; 7903 MDB_node *s2; 7904 MDB_val bkey; 7905 /* must find the lowest key below dst */ 7906 mdb_cursor_copy(cdst, &mn); 7907 rc = mdb_page_search_lowest(&mn); 7908 if (rc) 7909 return rc; 7910 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) { 7911 bkey.mv_size = mn.mc_db->md_pad; 7912 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size); 7913 } else { 7914 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0); 7915 bkey.mv_size = NODEKSZ(s2); 7916 bkey.mv_data = NODEKEY(s2); 7917 } 7918 mn.mc_snum = snum--; 7919 mn.mc_top = snum; 7920 mn.mc_ki[snum] = 0; 7921 rc = mdb_update_key(&mn, &bkey); 7922 if (rc) 7923 return rc; 7924 } 7925 7926 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u", 7927 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch", 7928 csrc->mc_ki[csrc->mc_top], 7929 DKEY(&key), 7930 csrc->mc_pg[csrc->mc_top]->mp_pgno, 7931 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno)); 7932 7933 /* Add the node to the destination page. 7934 */ 7935 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags); 7936 if (rc != MDB_SUCCESS) 7937 return rc; 7938 7939 /* Delete the node from the source page. 7940 */ 7941 mdb_node_del(csrc, key.mv_size); 7942 7943 { 7944 /* Adjust other cursors pointing to mp */ 7945 MDB_cursor *m2, *m3; 7946 MDB_dbi dbi = csrc->mc_dbi; 7947 MDB_page *mpd, *mps; 7948 7949 mps = csrc->mc_pg[csrc->mc_top]; 7950 /* If we're adding on the left, bump others up */ 7951 if (fromleft) { 7952 mpd = cdst->mc_pg[csrc->mc_top]; 7953 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 7954 if (csrc->mc_flags & C_SUB) 7955 m3 = &m2->mc_xcursor->mx_cursor; 7956 else 7957 m3 = m2; 7958 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top) 7959 continue; 7960 if (m3 != cdst && 7961 m3->mc_pg[csrc->mc_top] == mpd && 7962 m3->mc_ki[csrc->mc_top] >= cdst->mc_ki[csrc->mc_top]) { 7963 m3->mc_ki[csrc->mc_top]++; 7964 } 7965 if (m3 !=csrc && 7966 m3->mc_pg[csrc->mc_top] == mps && 7967 m3->mc_ki[csrc->mc_top] == csrc->mc_ki[csrc->mc_top]) { 7968 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top]; 7969 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top]; 7970 m3->mc_ki[csrc->mc_top-1]++; 7971 } 7972 if (IS_LEAF(mps)) 7973 XCURSOR_REFRESH(m3, csrc->mc_top, m3->mc_pg[csrc->mc_top]); 7974 } 7975 } else 7976 /* Adding on the right, bump others down */ 7977 { 7978 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 7979 if (csrc->mc_flags & C_SUB) 7980 m3 = &m2->mc_xcursor->mx_cursor; 7981 else 7982 m3 = m2; 7983 if (m3 == csrc) continue; 7984 if (!(m3->mc_flags & C_INITIALIZED) || m3->mc_top < csrc->mc_top) 7985 continue; 7986 if (m3->mc_pg[csrc->mc_top] == mps) { 7987 if (!m3->mc_ki[csrc->mc_top]) { 7988 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top]; 7989 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top]; 7990 m3->mc_ki[csrc->mc_top-1]--; 7991 } else { 7992 m3->mc_ki[csrc->mc_top]--; 7993 } 7994 if (IS_LEAF(mps)) 7995 XCURSOR_REFRESH(m3, csrc->mc_top, m3->mc_pg[csrc->mc_top]); 7996 } 7997 } 7998 } 7999 } 8000 8001 /* Update the parent separators. 8002 */ 8003 if (csrc->mc_ki[csrc->mc_top] == 0) { 8004 if (csrc->mc_ki[csrc->mc_top-1] != 0) { 8005 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) { 8006 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size); 8007 } else { 8008 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); 8009 key.mv_size = NODEKSZ(srcnode); 8010 key.mv_data = NODEKEY(srcnode); 8011 } 8012 DPRINTF(("update separator for source page %"Z"u to [%s]", 8013 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key))); 8014 mdb_cursor_copy(csrc, &mn); 8015 mn.mc_snum--; 8016 mn.mc_top--; 8017 /* We want mdb_rebalance to find mn when doing fixups */ 8018 WITH_CURSOR_TRACKING(mn, 8019 rc = mdb_update_key(&mn, &key)); 8020 if (rc) 8021 return rc; 8022 } 8023 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) { 8024 MDB_val nullkey; 8025 indx_t ix = csrc->mc_ki[csrc->mc_top]; 8026 nullkey.mv_size = 0; 8027 csrc->mc_ki[csrc->mc_top] = 0; 8028 rc = mdb_update_key(csrc, &nullkey); 8029 csrc->mc_ki[csrc->mc_top] = ix; 8030 mdb_cassert(csrc, rc == MDB_SUCCESS); 8031 } 8032 } 8033 8034 if (cdst->mc_ki[cdst->mc_top] == 0) { 8035 if (cdst->mc_ki[cdst->mc_top-1] != 0) { 8036 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) { 8037 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size); 8038 } else { 8039 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0); 8040 key.mv_size = NODEKSZ(srcnode); 8041 key.mv_data = NODEKEY(srcnode); 8042 } 8043 DPRINTF(("update separator for destination page %"Z"u to [%s]", 8044 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key))); 8045 mdb_cursor_copy(cdst, &mn); 8046 mn.mc_snum--; 8047 mn.mc_top--; 8048 /* We want mdb_rebalance to find mn when doing fixups */ 8049 WITH_CURSOR_TRACKING(mn, 8050 rc = mdb_update_key(&mn, &key)); 8051 if (rc) 8052 return rc; 8053 } 8054 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) { 8055 MDB_val nullkey; 8056 indx_t ix = cdst->mc_ki[cdst->mc_top]; 8057 nullkey.mv_size = 0; 8058 cdst->mc_ki[cdst->mc_top] = 0; 8059 rc = mdb_update_key(cdst, &nullkey); 8060 cdst->mc_ki[cdst->mc_top] = ix; 8061 mdb_cassert(cdst, rc == MDB_SUCCESS); 8062 } 8063 } 8064 8065 return MDB_SUCCESS; 8066 } 8067 8068 /** Merge one page into another. 8069 * The nodes from the page pointed to by \b csrc will 8070 * be copied to the page pointed to by \b cdst and then 8071 * the \b csrc page will be freed. 8072 * @param[in] csrc Cursor pointing to the source page. 8073 * @param[in] cdst Cursor pointing to the destination page. 8074 * @return 0 on success, non-zero on failure. 8075 */ 8076 static int 8077 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst) 8078 { 8079 MDB_page *psrc, *pdst; 8080 MDB_node *srcnode; 8081 MDB_val key, data; 8082 unsigned nkeys; 8083 int rc; 8084 indx_t i, j; 8085 8086 psrc = csrc->mc_pg[csrc->mc_top]; 8087 pdst = cdst->mc_pg[cdst->mc_top]; 8088 8089 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno)); 8090 8091 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */ 8092 mdb_cassert(csrc, cdst->mc_snum > 1); 8093 8094 /* Mark dst as dirty. */ 8095 if ((rc = mdb_page_touch(cdst))) 8096 return rc; 8097 8098 /* get dst page again now that we've touched it. */ 8099 pdst = cdst->mc_pg[cdst->mc_top]; 8100 8101 /* Move all nodes from src to dst. 8102 */ 8103 j = nkeys = NUMKEYS(pdst); 8104 if (IS_LEAF2(psrc)) { 8105 key.mv_size = csrc->mc_db->md_pad; 8106 key.mv_data = METADATA(psrc); 8107 for (i = 0; i < NUMKEYS(psrc); i++, j++) { 8108 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0); 8109 if (rc != MDB_SUCCESS) 8110 return rc; 8111 key.mv_data = (char *)key.mv_data + key.mv_size; 8112 } 8113 } else { 8114 for (i = 0; i < NUMKEYS(psrc); i++, j++) { 8115 srcnode = NODEPTR(psrc, i); 8116 if (i == 0 && IS_BRANCH(psrc)) { 8117 MDB_cursor mn; 8118 MDB_node *s2; 8119 mdb_cursor_copy(csrc, &mn); 8120 mn.mc_xcursor = NULL; 8121 /* must find the lowest key below src */ 8122 rc = mdb_page_search_lowest(&mn); 8123 if (rc) 8124 return rc; 8125 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) { 8126 key.mv_size = mn.mc_db->md_pad; 8127 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size); 8128 } else { 8129 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0); 8130 key.mv_size = NODEKSZ(s2); 8131 key.mv_data = NODEKEY(s2); 8132 } 8133 } else { 8134 key.mv_size = srcnode->mn_ksize; 8135 key.mv_data = NODEKEY(srcnode); 8136 } 8137 8138 data.mv_size = NODEDSZ(srcnode); 8139 data.mv_data = NODEDATA(srcnode); 8140 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags); 8141 if (rc != MDB_SUCCESS) 8142 return rc; 8143 } 8144 } 8145 8146 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)", 8147 pdst->mp_pgno, NUMKEYS(pdst), 8148 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10)); 8149 8150 /* Unlink the src page from parent and add to free list. 8151 */ 8152 csrc->mc_top--; 8153 mdb_node_del(csrc, 0); 8154 if (csrc->mc_ki[csrc->mc_top] == 0) { 8155 key.mv_size = 0; 8156 rc = mdb_update_key(csrc, &key); 8157 if (rc) { 8158 csrc->mc_top++; 8159 return rc; 8160 } 8161 } 8162 csrc->mc_top++; 8163 8164 psrc = csrc->mc_pg[csrc->mc_top]; 8165 /* If not operating on FreeDB, allow this page to be reused 8166 * in this txn. Otherwise just add to free list. 8167 */ 8168 rc = mdb_page_loose(csrc, psrc); 8169 if (rc) 8170 return rc; 8171 if (IS_LEAF(psrc)) 8172 csrc->mc_db->md_leaf_pages--; 8173 else 8174 csrc->mc_db->md_branch_pages--; 8175 { 8176 /* Adjust other cursors pointing to mp */ 8177 MDB_cursor *m2, *m3; 8178 MDB_dbi dbi = csrc->mc_dbi; 8179 unsigned int top = csrc->mc_top; 8180 8181 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8182 if (csrc->mc_flags & C_SUB) 8183 m3 = &m2->mc_xcursor->mx_cursor; 8184 else 8185 m3 = m2; 8186 if (m3 == csrc) continue; 8187 if (m3->mc_snum < csrc->mc_snum) continue; 8188 if (m3->mc_pg[top] == psrc) { 8189 m3->mc_pg[top] = pdst; 8190 m3->mc_ki[top] += nkeys; 8191 m3->mc_ki[top-1] = cdst->mc_ki[top-1]; 8192 } else if (m3->mc_pg[top-1] == csrc->mc_pg[top-1] && 8193 m3->mc_ki[top-1] > csrc->mc_ki[top-1]) { 8194 m3->mc_ki[top-1]--; 8195 } 8196 if (IS_LEAF(psrc)) 8197 XCURSOR_REFRESH(m3, top, m3->mc_pg[top]); 8198 } 8199 } 8200 { 8201 unsigned int snum = cdst->mc_snum; 8202 uint16_t depth = cdst->mc_db->md_depth; 8203 mdb_cursor_pop(cdst); 8204 rc = mdb_rebalance(cdst); 8205 /* Did the tree height change? */ 8206 if (depth != cdst->mc_db->md_depth) 8207 snum += cdst->mc_db->md_depth - depth; 8208 cdst->mc_snum = snum; 8209 cdst->mc_top = snum-1; 8210 } 8211 return rc; 8212 } 8213 8214 /** Copy the contents of a cursor. 8215 * @param[in] csrc The cursor to copy from. 8216 * @param[out] cdst The cursor to copy to. 8217 */ 8218 static void 8219 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst) 8220 { 8221 unsigned int i; 8222 8223 cdst->mc_txn = csrc->mc_txn; 8224 cdst->mc_dbi = csrc->mc_dbi; 8225 cdst->mc_db = csrc->mc_db; 8226 cdst->mc_dbx = csrc->mc_dbx; 8227 cdst->mc_snum = csrc->mc_snum; 8228 cdst->mc_top = csrc->mc_top; 8229 cdst->mc_flags = csrc->mc_flags; 8230 8231 for (i=0; i<csrc->mc_snum; i++) { 8232 cdst->mc_pg[i] = csrc->mc_pg[i]; 8233 cdst->mc_ki[i] = csrc->mc_ki[i]; 8234 } 8235 } 8236 8237 /** Rebalance the tree after a delete operation. 8238 * @param[in] mc Cursor pointing to the page where rebalancing 8239 * should begin. 8240 * @return 0 on success, non-zero on failure. 8241 */ 8242 static int 8243 mdb_rebalance(MDB_cursor *mc) 8244 { 8245 MDB_node *node; 8246 int rc, fromleft; 8247 unsigned int ptop, minkeys, thresh; 8248 MDB_cursor mn; 8249 indx_t oldki; 8250 8251 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) { 8252 minkeys = 2; 8253 thresh = 1; 8254 } else { 8255 minkeys = 1; 8256 thresh = FILL_THRESHOLD; 8257 } 8258 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)", 8259 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch", 8260 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]), 8261 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10)); 8262 8263 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh && 8264 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) { 8265 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold", 8266 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]))); 8267 return MDB_SUCCESS; 8268 } 8269 8270 if (mc->mc_snum < 2) { 8271 MDB_page *mp = mc->mc_pg[0]; 8272 if (IS_SUBP(mp)) { 8273 DPUTS("Can't rebalance a subpage, ignoring"); 8274 return MDB_SUCCESS; 8275 } 8276 if (NUMKEYS(mp) == 0) { 8277 DPUTS("tree is completely empty"); 8278 mc->mc_db->md_root = P_INVALID; 8279 mc->mc_db->md_depth = 0; 8280 mc->mc_db->md_leaf_pages = 0; 8281 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno); 8282 if (rc) 8283 return rc; 8284 /* Adjust cursors pointing to mp */ 8285 mc->mc_snum = 0; 8286 mc->mc_top = 0; 8287 mc->mc_flags &= ~C_INITIALIZED; 8288 { 8289 MDB_cursor *m2, *m3; 8290 MDB_dbi dbi = mc->mc_dbi; 8291 8292 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8293 if (mc->mc_flags & C_SUB) 8294 m3 = &m2->mc_xcursor->mx_cursor; 8295 else 8296 m3 = m2; 8297 if (!(m3->mc_flags & C_INITIALIZED) || (m3->mc_snum < mc->mc_snum)) 8298 continue; 8299 if (m3->mc_pg[0] == mp) { 8300 m3->mc_snum = 0; 8301 m3->mc_top = 0; 8302 m3->mc_flags &= ~C_INITIALIZED; 8303 } 8304 } 8305 } 8306 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) { 8307 int i; 8308 DPUTS("collapsing root page!"); 8309 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno); 8310 if (rc) 8311 return rc; 8312 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0)); 8313 rc = mdb_page_get(mc, mc->mc_db->md_root, &mc->mc_pg[0], NULL); 8314 if (rc) 8315 return rc; 8316 mc->mc_db->md_depth--; 8317 mc->mc_db->md_branch_pages--; 8318 mc->mc_ki[0] = mc->mc_ki[1]; 8319 for (i = 1; i<mc->mc_db->md_depth; i++) { 8320 mc->mc_pg[i] = mc->mc_pg[i+1]; 8321 mc->mc_ki[i] = mc->mc_ki[i+1]; 8322 } 8323 { 8324 /* Adjust other cursors pointing to mp */ 8325 MDB_cursor *m2, *m3; 8326 MDB_dbi dbi = mc->mc_dbi; 8327 8328 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8329 if (mc->mc_flags & C_SUB) 8330 m3 = &m2->mc_xcursor->mx_cursor; 8331 else 8332 m3 = m2; 8333 if (m3 == mc) continue; 8334 if (!(m3->mc_flags & C_INITIALIZED)) 8335 continue; 8336 if (m3->mc_pg[0] == mp) { 8337 for (i=0; i<mc->mc_db->md_depth; i++) { 8338 m3->mc_pg[i] = m3->mc_pg[i+1]; 8339 m3->mc_ki[i] = m3->mc_ki[i+1]; 8340 } 8341 m3->mc_snum--; 8342 m3->mc_top--; 8343 } 8344 } 8345 } 8346 } else 8347 DPUTS("root page doesn't need rebalancing"); 8348 return MDB_SUCCESS; 8349 } 8350 8351 /* The parent (branch page) must have at least 2 pointers, 8352 * otherwise the tree is invalid. 8353 */ 8354 ptop = mc->mc_top-1; 8355 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1); 8356 8357 /* Leaf page fill factor is below the threshold. 8358 * Try to move keys from left or right neighbor, or 8359 * merge with a neighbor page. 8360 */ 8361 8362 /* Find neighbors. 8363 */ 8364 mdb_cursor_copy(mc, &mn); 8365 mn.mc_xcursor = NULL; 8366 8367 oldki = mc->mc_ki[mc->mc_top]; 8368 if (mc->mc_ki[ptop] == 0) { 8369 /* We're the leftmost leaf in our parent. 8370 */ 8371 DPUTS("reading right neighbor"); 8372 mn.mc_ki[ptop]++; 8373 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]); 8374 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL); 8375 if (rc) 8376 return rc; 8377 mn.mc_ki[mn.mc_top] = 0; 8378 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]); 8379 fromleft = 0; 8380 } else { 8381 /* There is at least one neighbor to the left. 8382 */ 8383 DPUTS("reading left neighbor"); 8384 mn.mc_ki[ptop]--; 8385 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]); 8386 rc = mdb_page_get(mc, NODEPGNO(node), &mn.mc_pg[mn.mc_top], NULL); 8387 if (rc) 8388 return rc; 8389 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1; 8390 mc->mc_ki[mc->mc_top] = 0; 8391 fromleft = 1; 8392 } 8393 8394 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)", 8395 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), 8396 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10)); 8397 8398 /* If the neighbor page is above threshold and has enough keys, 8399 * move one key from it. Otherwise we should try to merge them. 8400 * (A branch page must never have less than 2 keys.) 8401 */ 8402 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) { 8403 rc = mdb_node_move(&mn, mc, fromleft); 8404 if (fromleft) { 8405 /* if we inserted on left, bump position up */ 8406 oldki++; 8407 } 8408 } else { 8409 if (!fromleft) { 8410 rc = mdb_page_merge(&mn, mc); 8411 } else { 8412 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]); 8413 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1; 8414 /* We want mdb_rebalance to find mn when doing fixups */ 8415 WITH_CURSOR_TRACKING(mn, 8416 rc = mdb_page_merge(mc, &mn)); 8417 mdb_cursor_copy(&mn, mc); 8418 } 8419 mc->mc_flags &= ~C_EOF; 8420 } 8421 mc->mc_ki[mc->mc_top] = oldki; 8422 return rc; 8423 } 8424 8425 /** Complete a delete operation started by #mdb_cursor_del(). */ 8426 static int 8427 mdb_cursor_del0(MDB_cursor *mc) 8428 { 8429 int rc; 8430 MDB_page *mp; 8431 indx_t ki; 8432 unsigned int nkeys; 8433 MDB_cursor *m2, *m3; 8434 MDB_dbi dbi = mc->mc_dbi; 8435 8436 ki = mc->mc_ki[mc->mc_top]; 8437 mp = mc->mc_pg[mc->mc_top]; 8438 mdb_node_del(mc, mc->mc_db->md_pad); 8439 mc->mc_db->md_entries--; 8440 { 8441 /* Adjust other cursors pointing to mp */ 8442 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8443 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; 8444 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED)) 8445 continue; 8446 if (m3 == mc || m3->mc_snum < mc->mc_snum) 8447 continue; 8448 if (m3->mc_pg[mc->mc_top] == mp) { 8449 if (m3->mc_ki[mc->mc_top] == ki) { 8450 m3->mc_flags |= C_DEL; 8451 if (mc->mc_db->md_flags & MDB_DUPSORT) { 8452 /* Sub-cursor referred into dataset which is gone */ 8453 m3->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); 8454 } 8455 continue; 8456 } else if (m3->mc_ki[mc->mc_top] > ki) { 8457 m3->mc_ki[mc->mc_top]--; 8458 } 8459 XCURSOR_REFRESH(m3, mc->mc_top, mp); 8460 } 8461 } 8462 } 8463 rc = mdb_rebalance(mc); 8464 8465 if (rc == MDB_SUCCESS) { 8466 /* DB is totally empty now, just bail out. 8467 * Other cursors adjustments were already done 8468 * by mdb_rebalance and aren't needed here. 8469 */ 8470 if (!mc->mc_snum) 8471 return rc; 8472 8473 mp = mc->mc_pg[mc->mc_top]; 8474 nkeys = NUMKEYS(mp); 8475 8476 /* Adjust other cursors pointing to mp */ 8477 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) { 8478 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2; 8479 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED)) 8480 continue; 8481 if (m3->mc_snum < mc->mc_snum) 8482 continue; 8483 if (m3->mc_pg[mc->mc_top] == mp) { 8484 /* if m3 points past last node in page, find next sibling */ 8485 if (m3->mc_ki[mc->mc_top] >= mc->mc_ki[mc->mc_top]) { 8486 if (m3->mc_ki[mc->mc_top] >= nkeys) { 8487 rc = mdb_cursor_sibling(m3, 1); 8488 if (rc == MDB_NOTFOUND) { 8489 m3->mc_flags |= C_EOF; 8490 rc = MDB_SUCCESS; 8491 continue; 8492 } 8493 } 8494 if (mc->mc_db->md_flags & MDB_DUPSORT) { 8495 MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]); 8496 /* If this node has dupdata, it may need to be reinited 8497 * because its data has moved. 8498 * If the xcursor was not initd it must be reinited. 8499 * Else if node points to a subDB, nothing is needed. 8500 * Else (xcursor was initd, not a subDB) needs mc_pg[0] reset. 8501 */ 8502 if (node->mn_flags & F_DUPDATA) { 8503 if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) { 8504 if (!(node->mn_flags & F_SUBDATA)) 8505 m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node); 8506 } else { 8507 mdb_xcursor_init1(m3, node); 8508 m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL; 8509 } 8510 } 8511 } 8512 } 8513 } 8514 } 8515 mc->mc_flags |= C_DEL; 8516 } 8517 8518 if (rc) 8519 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 8520 return rc; 8521 } 8522 8523 int 8524 mdb_del(MDB_txn *txn, MDB_dbi dbi, 8525 MDB_val *key, MDB_val *data) 8526 { 8527 if (!key || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 8528 return EINVAL; 8529 8530 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED)) 8531 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN; 8532 8533 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) { 8534 /* must ignore any data */ 8535 data = NULL; 8536 } 8537 8538 return mdb_del0(txn, dbi, key, data, 0); 8539 } 8540 8541 static int 8542 mdb_del0(MDB_txn *txn, MDB_dbi dbi, 8543 MDB_val *key, MDB_val *data, unsigned flags) 8544 { 8545 MDB_cursor mc; 8546 MDB_xcursor mx; 8547 MDB_cursor_op op; 8548 MDB_val rdata, *xdata; 8549 int rc, exact = 0; 8550 DKBUF; 8551 8552 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key))); 8553 8554 mdb_cursor_init(&mc, txn, dbi, &mx); 8555 8556 if (data) { 8557 op = MDB_GET_BOTH; 8558 rdata = *data; 8559 xdata = &rdata; 8560 } else { 8561 op = MDB_SET; 8562 xdata = NULL; 8563 flags |= MDB_NODUPDATA; 8564 } 8565 rc = mdb_cursor_set(&mc, key, xdata, op, &exact); 8566 if (rc == 0) { 8567 /* let mdb_page_split know about this cursor if needed: 8568 * delete will trigger a rebalance; if it needs to move 8569 * a node from one page to another, it will have to 8570 * update the parent's separator key(s). If the new sepkey 8571 * is larger than the current one, the parent page may 8572 * run out of space, triggering a split. We need this 8573 * cursor to be consistent until the end of the rebalance. 8574 */ 8575 mc.mc_flags |= C_UNTRACK; 8576 mc.mc_next = txn->mt_cursors[dbi]; 8577 txn->mt_cursors[dbi] = &mc; 8578 rc = mdb_cursor_del(&mc, flags); 8579 txn->mt_cursors[dbi] = mc.mc_next; 8580 } 8581 return rc; 8582 } 8583 8584 /** Split a page and insert a new node. 8585 * Set #MDB_TXN_ERROR on failure. 8586 * @param[in,out] mc Cursor pointing to the page and desired insertion index. 8587 * The cursor will be updated to point to the actual page and index where 8588 * the node got inserted after the split. 8589 * @param[in] newkey The key for the newly inserted node. 8590 * @param[in] newdata The data for the newly inserted node. 8591 * @param[in] newpgno The page number, if the new node is a branch node. 8592 * @param[in] nflags The #NODE_ADD_FLAGS for the new node. 8593 * @return 0 on success, non-zero on failure. 8594 */ 8595 static int 8596 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno, 8597 unsigned int nflags) 8598 { 8599 unsigned int flags; 8600 int rc = MDB_SUCCESS, new_root = 0, did_split = 0; 8601 indx_t newindx; 8602 pgno_t pgno = 0; 8603 int i, j, split_indx, nkeys, pmax; 8604 MDB_env *env = mc->mc_txn->mt_env; 8605 MDB_node *node; 8606 MDB_val sepkey, rkey, xdata, *rdata = &xdata; 8607 MDB_page *copy = NULL; 8608 MDB_page *mp, *rp, *pp; 8609 int ptop; 8610 MDB_cursor mn; 8611 DKBUF; 8612 8613 mp = mc->mc_pg[mc->mc_top]; 8614 newindx = mc->mc_ki[mc->mc_top]; 8615 nkeys = NUMKEYS(mp); 8616 8617 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i", 8618 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno, 8619 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys)); 8620 8621 /* Create a right sibling. */ 8622 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp))) 8623 return rc; 8624 rp->mp_pad = mp->mp_pad; 8625 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno)); 8626 8627 /* Usually when splitting the root page, the cursor 8628 * height is 1. But when called from mdb_update_key, 8629 * the cursor height may be greater because it walks 8630 * up the stack while finding the branch slot to update. 8631 */ 8632 if (mc->mc_top < 1) { 8633 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp))) 8634 goto done; 8635 /* shift current top to make room for new parent */ 8636 for (i=mc->mc_snum; i>0; i--) { 8637 mc->mc_pg[i] = mc->mc_pg[i-1]; 8638 mc->mc_ki[i] = mc->mc_ki[i-1]; 8639 } 8640 mc->mc_pg[0] = pp; 8641 mc->mc_ki[0] = 0; 8642 mc->mc_db->md_root = pp->mp_pgno; 8643 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno)); 8644 new_root = mc->mc_db->md_depth++; 8645 8646 /* Add left (implicit) pointer. */ 8647 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) { 8648 /* undo the pre-push */ 8649 mc->mc_pg[0] = mc->mc_pg[1]; 8650 mc->mc_ki[0] = mc->mc_ki[1]; 8651 mc->mc_db->md_root = mp->mp_pgno; 8652 mc->mc_db->md_depth--; 8653 goto done; 8654 } 8655 mc->mc_snum++; 8656 mc->mc_top++; 8657 ptop = 0; 8658 } else { 8659 ptop = mc->mc_top-1; 8660 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno)); 8661 } 8662 8663 mdb_cursor_copy(mc, &mn); 8664 mn.mc_xcursor = NULL; 8665 mn.mc_pg[mn.mc_top] = rp; 8666 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1; 8667 8668 if (nflags & MDB_APPEND) { 8669 mn.mc_ki[mn.mc_top] = 0; 8670 sepkey = *newkey; 8671 split_indx = newindx; 8672 nkeys = 0; 8673 } else { 8674 8675 split_indx = (nkeys+1) / 2; 8676 8677 if (IS_LEAF2(rp)) { 8678 char *split, *ins; 8679 int x; 8680 unsigned int lsize, rsize, ksize; 8681 /* Move half of the keys to the right sibling */ 8682 x = mc->mc_ki[mc->mc_top] - split_indx; 8683 ksize = mc->mc_db->md_pad; 8684 split = LEAF2KEY(mp, split_indx, ksize); 8685 rsize = (nkeys - split_indx) * ksize; 8686 lsize = (nkeys - split_indx) * sizeof(indx_t); 8687 mp->mp_lower -= lsize; 8688 rp->mp_lower += lsize; 8689 mp->mp_upper += rsize - lsize; 8690 rp->mp_upper -= rsize - lsize; 8691 sepkey.mv_size = ksize; 8692 if (newindx == split_indx) { 8693 sepkey.mv_data = newkey->mv_data; 8694 } else { 8695 sepkey.mv_data = split; 8696 } 8697 if (x<0) { 8698 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize); 8699 memcpy(rp->mp_ptrs, split, rsize); 8700 sepkey.mv_data = rp->mp_ptrs; 8701 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize); 8702 memcpy(ins, newkey->mv_data, ksize); 8703 mp->mp_lower += sizeof(indx_t); 8704 mp->mp_upper -= ksize - sizeof(indx_t); 8705 } else { 8706 if (x) 8707 memcpy(rp->mp_ptrs, split, x * ksize); 8708 ins = LEAF2KEY(rp, x, ksize); 8709 memcpy(ins, newkey->mv_data, ksize); 8710 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize); 8711 rp->mp_lower += sizeof(indx_t); 8712 rp->mp_upper -= ksize - sizeof(indx_t); 8713 mc->mc_ki[mc->mc_top] = x; 8714 } 8715 } else { 8716 int psize, nsize, k; 8717 /* Maximum free space in an empty page */ 8718 pmax = env->me_psize - PAGEHDRSZ; 8719 if (IS_LEAF(mp)) 8720 nsize = mdb_leaf_size(env, newkey, newdata); 8721 else 8722 nsize = mdb_branch_size(env, newkey); 8723 nsize = EVEN(nsize); 8724 8725 /* grab a page to hold a temporary copy */ 8726 copy = mdb_page_malloc(mc->mc_txn, 1); 8727 if (copy == NULL) { 8728 rc = ENOMEM; 8729 goto done; 8730 } 8731 copy->mp_pgno = mp->mp_pgno; 8732 copy->mp_flags = mp->mp_flags; 8733 copy->mp_lower = (PAGEHDRSZ-PAGEBASE); 8734 copy->mp_upper = env->me_psize - PAGEBASE; 8735 8736 /* prepare to insert */ 8737 for (i=0, j=0; i<nkeys; i++) { 8738 if (i == newindx) { 8739 copy->mp_ptrs[j++] = 0; 8740 } 8741 copy->mp_ptrs[j++] = mp->mp_ptrs[i]; 8742 } 8743 8744 /* When items are relatively large the split point needs 8745 * to be checked, because being off-by-one will make the 8746 * difference between success or failure in mdb_node_add. 8747 * 8748 * It's also relevant if a page happens to be laid out 8749 * such that one half of its nodes are all "small" and 8750 * the other half of its nodes are "large." If the new 8751 * item is also "large" and falls on the half with 8752 * "large" nodes, it also may not fit. 8753 * 8754 * As a final tweak, if the new item goes on the last 8755 * spot on the page (and thus, onto the new page), bias 8756 * the split so the new page is emptier than the old page. 8757 * This yields better packing during sequential inserts. 8758 */ 8759 if (nkeys < 32 || nsize > pmax/16 || newindx >= nkeys) { 8760 /* Find split point */ 8761 psize = 0; 8762 if (newindx <= split_indx || newindx >= nkeys) { 8763 i = 0; j = 1; 8764 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp); 8765 } else { 8766 i = nkeys; j = -1; 8767 k = split_indx-1; 8768 } 8769 for (; i!=k; i+=j) { 8770 if (i == newindx) { 8771 psize += nsize; 8772 node = NULL; 8773 } else { 8774 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE); 8775 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t); 8776 if (IS_LEAF(mp)) { 8777 if (F_ISSET(node->mn_flags, F_BIGDATA)) 8778 psize += sizeof(pgno_t); 8779 else 8780 psize += NODEDSZ(node); 8781 } 8782 psize = EVEN(psize); 8783 } 8784 if (psize > pmax || i == k-j) { 8785 split_indx = i + (j<0); 8786 break; 8787 } 8788 } 8789 } 8790 if (split_indx == newindx) { 8791 sepkey.mv_size = newkey->mv_size; 8792 sepkey.mv_data = newkey->mv_data; 8793 } else { 8794 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE); 8795 sepkey.mv_size = node->mn_ksize; 8796 sepkey.mv_data = NODEKEY(node); 8797 } 8798 } 8799 } 8800 8801 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey))); 8802 8803 /* Copy separator key to the parent. 8804 */ 8805 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) { 8806 int snum = mc->mc_snum; 8807 mn.mc_snum--; 8808 mn.mc_top--; 8809 did_split = 1; 8810 /* We want other splits to find mn when doing fixups */ 8811 WITH_CURSOR_TRACKING(mn, 8812 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0)); 8813 if (rc) 8814 goto done; 8815 8816 /* root split? */ 8817 if (mc->mc_snum > snum) { 8818 ptop++; 8819 } 8820 /* Right page might now have changed parent. 8821 * Check if left page also changed parent. 8822 */ 8823 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && 8824 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) { 8825 for (i=0; i<ptop; i++) { 8826 mc->mc_pg[i] = mn.mc_pg[i]; 8827 mc->mc_ki[i] = mn.mc_ki[i]; 8828 } 8829 mc->mc_pg[ptop] = mn.mc_pg[ptop]; 8830 if (mn.mc_ki[ptop]) { 8831 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1; 8832 } else { 8833 /* find right page's left sibling */ 8834 mc->mc_ki[ptop] = mn.mc_ki[ptop]; 8835 mdb_cursor_sibling(mc, 0); 8836 } 8837 } 8838 } else { 8839 mn.mc_top--; 8840 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0); 8841 mn.mc_top++; 8842 } 8843 if (rc != MDB_SUCCESS) { 8844 goto done; 8845 } 8846 if (nflags & MDB_APPEND) { 8847 mc->mc_pg[mc->mc_top] = rp; 8848 mc->mc_ki[mc->mc_top] = 0; 8849 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags); 8850 if (rc) 8851 goto done; 8852 for (i=0; i<mc->mc_top; i++) 8853 mc->mc_ki[i] = mn.mc_ki[i]; 8854 } else if (!IS_LEAF2(mp)) { 8855 /* Move nodes */ 8856 mc->mc_pg[mc->mc_top] = rp; 8857 i = split_indx; 8858 j = 0; 8859 do { 8860 if (i == newindx) { 8861 rkey.mv_data = newkey->mv_data; 8862 rkey.mv_size = newkey->mv_size; 8863 if (IS_LEAF(mp)) { 8864 rdata = newdata; 8865 } else 8866 pgno = newpgno; 8867 flags = nflags; 8868 /* Update index for the new key. */ 8869 mc->mc_ki[mc->mc_top] = j; 8870 } else { 8871 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE); 8872 rkey.mv_data = NODEKEY(node); 8873 rkey.mv_size = node->mn_ksize; 8874 if (IS_LEAF(mp)) { 8875 xdata.mv_data = NODEDATA(node); 8876 xdata.mv_size = NODEDSZ(node); 8877 rdata = &xdata; 8878 } else 8879 pgno = NODEPGNO(node); 8880 flags = node->mn_flags; 8881 } 8882 8883 if (!IS_LEAF(mp) && j == 0) { 8884 /* First branch index doesn't need key data. */ 8885 rkey.mv_size = 0; 8886 } 8887 8888 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags); 8889 if (rc) 8890 goto done; 8891 if (i == nkeys) { 8892 i = 0; 8893 j = 0; 8894 mc->mc_pg[mc->mc_top] = copy; 8895 } else { 8896 i++; 8897 j++; 8898 } 8899 } while (i != split_indx); 8900 8901 nkeys = NUMKEYS(copy); 8902 for (i=0; i<nkeys; i++) 8903 mp->mp_ptrs[i] = copy->mp_ptrs[i]; 8904 mp->mp_lower = copy->mp_lower; 8905 mp->mp_upper = copy->mp_upper; 8906 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1), 8907 env->me_psize - copy->mp_upper - PAGEBASE); 8908 8909 /* reset back to original page */ 8910 if (newindx < split_indx) { 8911 mc->mc_pg[mc->mc_top] = mp; 8912 } else { 8913 mc->mc_pg[mc->mc_top] = rp; 8914 mc->mc_ki[ptop]++; 8915 /* Make sure mc_ki is still valid. 8916 */ 8917 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && 8918 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) { 8919 for (i=0; i<=ptop; i++) { 8920 mc->mc_pg[i] = mn.mc_pg[i]; 8921 mc->mc_ki[i] = mn.mc_ki[i]; 8922 } 8923 } 8924 } 8925 if (nflags & MDB_RESERVE) { 8926 node = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); 8927 if (!(node->mn_flags & F_BIGDATA)) 8928 newdata->mv_data = NODEDATA(node); 8929 } 8930 } else { 8931 if (newindx >= split_indx) { 8932 mc->mc_pg[mc->mc_top] = rp; 8933 mc->mc_ki[ptop]++; 8934 /* Make sure mc_ki is still valid. 8935 */ 8936 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] && 8937 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) { 8938 for (i=0; i<=ptop; i++) { 8939 mc->mc_pg[i] = mn.mc_pg[i]; 8940 mc->mc_ki[i] = mn.mc_ki[i]; 8941 } 8942 } 8943 } 8944 } 8945 8946 { 8947 /* Adjust other cursors pointing to mp */ 8948 MDB_cursor *m2, *m3; 8949 MDB_dbi dbi = mc->mc_dbi; 8950 nkeys = NUMKEYS(mp); 8951 8952 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) { 8953 if (mc->mc_flags & C_SUB) 8954 m3 = &m2->mc_xcursor->mx_cursor; 8955 else 8956 m3 = m2; 8957 if (m3 == mc) 8958 continue; 8959 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED)) 8960 continue; 8961 if (new_root) { 8962 int k; 8963 /* sub cursors may be on different DB */ 8964 if (m3->mc_pg[0] != mp) 8965 continue; 8966 /* root split */ 8967 for (k=new_root; k>=0; k--) { 8968 m3->mc_ki[k+1] = m3->mc_ki[k]; 8969 m3->mc_pg[k+1] = m3->mc_pg[k]; 8970 } 8971 if (m3->mc_ki[0] >= nkeys) { 8972 m3->mc_ki[0] = 1; 8973 } else { 8974 m3->mc_ki[0] = 0; 8975 } 8976 m3->mc_pg[0] = mc->mc_pg[0]; 8977 m3->mc_snum++; 8978 m3->mc_top++; 8979 } 8980 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) { 8981 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE)) 8982 m3->mc_ki[mc->mc_top]++; 8983 if (m3->mc_ki[mc->mc_top] >= nkeys) { 8984 m3->mc_pg[mc->mc_top] = rp; 8985 m3->mc_ki[mc->mc_top] -= nkeys; 8986 for (i=0; i<mc->mc_top; i++) { 8987 m3->mc_ki[i] = mn.mc_ki[i]; 8988 m3->mc_pg[i] = mn.mc_pg[i]; 8989 } 8990 } 8991 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] && 8992 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) { 8993 m3->mc_ki[ptop]++; 8994 } 8995 if (IS_LEAF(mp)) 8996 XCURSOR_REFRESH(m3, mc->mc_top, m3->mc_pg[mc->mc_top]); 8997 } 8998 } 8999 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp))); 9000 9001 done: 9002 if (copy) /* tmp page */ 9003 mdb_page_free(env, copy); 9004 if (rc) 9005 mc->mc_txn->mt_flags |= MDB_TXN_ERROR; 9006 return rc; 9007 } 9008 9009 int 9010 mdb_put(MDB_txn *txn, MDB_dbi dbi, 9011 MDB_val *key, MDB_val *data, unsigned int flags) 9012 { 9013 MDB_cursor mc; 9014 MDB_xcursor mx; 9015 int rc; 9016 9017 if (!key || !data || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 9018 return EINVAL; 9019 9020 if (flags & ~(MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) 9021 return EINVAL; 9022 9023 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_BLOCKED)) 9024 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN; 9025 9026 mdb_cursor_init(&mc, txn, dbi, &mx); 9027 mc.mc_next = txn->mt_cursors[dbi]; 9028 txn->mt_cursors[dbi] = &mc; 9029 rc = mdb_cursor_put(&mc, key, data, flags); 9030 txn->mt_cursors[dbi] = mc.mc_next; 9031 return rc; 9032 } 9033 9034 #ifndef MDB_WBUF 9035 #define MDB_WBUF (1024*1024) 9036 #endif 9037 #define MDB_EOF 0x10 /**< #mdb_env_copyfd1() is done reading */ 9038 9039 /** State needed for a double-buffering compacting copy. */ 9040 typedef struct mdb_copy { 9041 MDB_env *mc_env; 9042 MDB_txn *mc_txn; 9043 pthread_mutex_t mc_mutex; 9044 pthread_cond_t mc_cond; /**< Condition variable for #mc_new */ 9045 char *mc_wbuf[2]; 9046 char *mc_over[2]; 9047 int mc_wlen[2]; 9048 int mc_olen[2]; 9049 pgno_t mc_next_pgno; 9050 HANDLE mc_fd; 9051 int mc_toggle; /**< Buffer number in provider */ 9052 int mc_new; /**< (0-2 buffers to write) | (#MDB_EOF at end) */ 9053 /** Error code. Never cleared if set. Both threads can set nonzero 9054 * to fail the copy. Not mutex-protected, LMDB expects atomic int. 9055 */ 9056 volatile int mc_error; 9057 } mdb_copy; 9058 9059 /** Dedicated writer thread for compacting copy. */ 9060 static THREAD_RET ESECT CALL_CONV 9061 mdb_env_copythr(void *arg) 9062 { 9063 mdb_copy *my = arg; 9064 char *ptr; 9065 int toggle = 0, wsize, rc; 9066 #ifdef _WIN32 9067 DWORD len; 9068 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL) 9069 #else 9070 int len; 9071 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0) 9072 #ifdef SIGPIPE 9073 sigset_t set; 9074 sigemptyset(&set); 9075 sigaddset(&set, SIGPIPE); 9076 if ((rc = pthread_sigmask(SIG_BLOCK, &set, NULL)) != 0) 9077 my->mc_error = rc; 9078 #endif 9079 #endif 9080 9081 pthread_mutex_lock(&my->mc_mutex); 9082 for(;;) { 9083 while (!my->mc_new) 9084 pthread_cond_wait(&my->mc_cond, &my->mc_mutex); 9085 if (my->mc_new == 0 + MDB_EOF) /* 0 buffers, just EOF */ 9086 break; 9087 wsize = my->mc_wlen[toggle]; 9088 ptr = my->mc_wbuf[toggle]; 9089 again: 9090 rc = MDB_SUCCESS; 9091 while (wsize > 0 && !my->mc_error) { 9092 DO_WRITE(rc, my->mc_fd, ptr, wsize, len); 9093 if (!rc) { 9094 rc = ErrCode(); 9095 #if defined(SIGPIPE) && !defined(_WIN32) 9096 if (rc == EPIPE) { 9097 /* Collect the pending SIGPIPE, otherwise at least OS X 9098 * gives it to the process on thread-exit (ITS#8504). 9099 */ 9100 int tmp; 9101 sigwait(&set, &tmp); 9102 } 9103 #endif 9104 break; 9105 } else if (len > 0) { 9106 rc = MDB_SUCCESS; 9107 ptr += len; 9108 wsize -= len; 9109 continue; 9110 } else { 9111 rc = EIO; 9112 break; 9113 } 9114 } 9115 if (rc) { 9116 my->mc_error = rc; 9117 } 9118 /* If there's an overflow page tail, write it too */ 9119 if (my->mc_olen[toggle]) { 9120 wsize = my->mc_olen[toggle]; 9121 ptr = my->mc_over[toggle]; 9122 my->mc_olen[toggle] = 0; 9123 goto again; 9124 } 9125 my->mc_wlen[toggle] = 0; 9126 toggle ^= 1; 9127 /* Return the empty buffer to provider */ 9128 my->mc_new--; 9129 pthread_cond_signal(&my->mc_cond); 9130 } 9131 pthread_mutex_unlock(&my->mc_mutex); 9132 return (THREAD_RET)0; 9133 #undef DO_WRITE 9134 } 9135 9136 /** Give buffer and/or #MDB_EOF to writer thread, await unused buffer. 9137 * 9138 * @param[in] my control structure. 9139 * @param[in] adjust (1 to hand off 1 buffer) | (MDB_EOF when ending). 9140 */ 9141 static int ESECT 9142 mdb_env_cthr_toggle(mdb_copy *my, int adjust) 9143 { 9144 pthread_mutex_lock(&my->mc_mutex); 9145 my->mc_new += adjust; 9146 pthread_cond_signal(&my->mc_cond); 9147 while (my->mc_new & 2) /* both buffers in use */ 9148 pthread_cond_wait(&my->mc_cond, &my->mc_mutex); 9149 pthread_mutex_unlock(&my->mc_mutex); 9150 9151 my->mc_toggle ^= (adjust & 1); 9152 /* Both threads reset mc_wlen, to be safe from threading errors */ 9153 my->mc_wlen[my->mc_toggle] = 0; 9154 return my->mc_error; 9155 } 9156 9157 /** Depth-first tree traversal for compacting copy. 9158 * @param[in] my control structure. 9159 * @param[in,out] pg database root. 9160 * @param[in] flags includes #F_DUPDATA if it is a sorted-duplicate sub-DB. 9161 */ 9162 static int ESECT 9163 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags) 9164 { 9165 MDB_cursor mc = {0}; 9166 MDB_node *ni; 9167 MDB_page *mo, *mp, *leaf; 9168 char *buf, *ptr; 9169 int rc, toggle; 9170 unsigned int i; 9171 9172 /* Empty DB, nothing to do */ 9173 if (*pg == P_INVALID) 9174 return MDB_SUCCESS; 9175 9176 mc.mc_snum = 1; 9177 mc.mc_txn = my->mc_txn; 9178 9179 rc = mdb_page_get(&mc, *pg, &mc.mc_pg[0], NULL); 9180 if (rc) 9181 return rc; 9182 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST); 9183 if (rc) 9184 return rc; 9185 9186 /* Make cursor pages writable */ 9187 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum); 9188 if (buf == NULL) 9189 return ENOMEM; 9190 9191 for (i=0; i<mc.mc_top; i++) { 9192 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize); 9193 mc.mc_pg[i] = (MDB_page *)ptr; 9194 ptr += my->mc_env->me_psize; 9195 } 9196 9197 /* This is writable space for a leaf page. Usually not needed. */ 9198 leaf = (MDB_page *)ptr; 9199 9200 toggle = my->mc_toggle; 9201 while (mc.mc_snum > 0) { 9202 unsigned n; 9203 mp = mc.mc_pg[mc.mc_top]; 9204 n = NUMKEYS(mp); 9205 9206 if (IS_LEAF(mp)) { 9207 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) { 9208 for (i=0; i<n; i++) { 9209 ni = NODEPTR(mp, i); 9210 if (ni->mn_flags & F_BIGDATA) { 9211 MDB_page *omp; 9212 pgno_t pg; 9213 9214 /* Need writable leaf */ 9215 if (mp != leaf) { 9216 mc.mc_pg[mc.mc_top] = leaf; 9217 mdb_page_copy(leaf, mp, my->mc_env->me_psize); 9218 mp = leaf; 9219 ni = NODEPTR(mp, i); 9220 } 9221 9222 memcpy(&pg, NODEDATA(ni), sizeof(pg)); 9223 memcpy(NODEDATA(ni), &my->mc_next_pgno, sizeof(pgno_t)); 9224 rc = mdb_page_get(&mc, pg, &omp, NULL); 9225 if (rc) 9226 goto done; 9227 if (my->mc_wlen[toggle] >= MDB_WBUF) { 9228 rc = mdb_env_cthr_toggle(my, 1); 9229 if (rc) 9230 goto done; 9231 toggle = my->mc_toggle; 9232 } 9233 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]); 9234 memcpy(mo, omp, my->mc_env->me_psize); 9235 mo->mp_pgno = my->mc_next_pgno; 9236 my->mc_next_pgno += omp->mp_pages; 9237 my->mc_wlen[toggle] += my->mc_env->me_psize; 9238 if (omp->mp_pages > 1) { 9239 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1); 9240 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize; 9241 rc = mdb_env_cthr_toggle(my, 1); 9242 if (rc) 9243 goto done; 9244 toggle = my->mc_toggle; 9245 } 9246 } else if (ni->mn_flags & F_SUBDATA) { 9247 MDB_db db; 9248 9249 /* Need writable leaf */ 9250 if (mp != leaf) { 9251 mc.mc_pg[mc.mc_top] = leaf; 9252 mdb_page_copy(leaf, mp, my->mc_env->me_psize); 9253 mp = leaf; 9254 ni = NODEPTR(mp, i); 9255 } 9256 9257 memcpy(&db, NODEDATA(ni), sizeof(db)); 9258 my->mc_toggle = toggle; 9259 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA); 9260 if (rc) 9261 goto done; 9262 toggle = my->mc_toggle; 9263 memcpy(NODEDATA(ni), &db, sizeof(db)); 9264 } 9265 } 9266 } 9267 } else { 9268 mc.mc_ki[mc.mc_top]++; 9269 if (mc.mc_ki[mc.mc_top] < n) { 9270 pgno_t pg; 9271 again: 9272 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]); 9273 pg = NODEPGNO(ni); 9274 rc = mdb_page_get(&mc, pg, &mp, NULL); 9275 if (rc) 9276 goto done; 9277 mc.mc_top++; 9278 mc.mc_snum++; 9279 mc.mc_ki[mc.mc_top] = 0; 9280 if (IS_BRANCH(mp)) { 9281 /* Whenever we advance to a sibling branch page, 9282 * we must proceed all the way down to its first leaf. 9283 */ 9284 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize); 9285 goto again; 9286 } else 9287 mc.mc_pg[mc.mc_top] = mp; 9288 continue; 9289 } 9290 } 9291 if (my->mc_wlen[toggle] >= MDB_WBUF) { 9292 rc = mdb_env_cthr_toggle(my, 1); 9293 if (rc) 9294 goto done; 9295 toggle = my->mc_toggle; 9296 } 9297 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]); 9298 mdb_page_copy(mo, mp, my->mc_env->me_psize); 9299 mo->mp_pgno = my->mc_next_pgno++; 9300 my->mc_wlen[toggle] += my->mc_env->me_psize; 9301 if (mc.mc_top) { 9302 /* Update parent if there is one */ 9303 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]); 9304 SETPGNO(ni, mo->mp_pgno); 9305 mdb_cursor_pop(&mc); 9306 } else { 9307 /* Otherwise we're done */ 9308 *pg = mo->mp_pgno; 9309 break; 9310 } 9311 } 9312 done: 9313 free(buf); 9314 return rc; 9315 } 9316 9317 /** Copy environment with compaction. */ 9318 static int ESECT 9319 mdb_env_copyfd1(MDB_env *env, HANDLE fd) 9320 { 9321 MDB_meta *mm; 9322 MDB_page *mp; 9323 mdb_copy my = {0}; 9324 MDB_txn *txn = NULL; 9325 pthread_t thr; 9326 pgno_t root, new_root; 9327 int rc = MDB_SUCCESS; 9328 9329 #ifdef _WIN32 9330 if (!(my.mc_mutex = CreateMutex(NULL, FALSE, NULL)) || 9331 !(my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL))) { 9332 rc = ErrCode(); 9333 goto done; 9334 } 9335 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize); 9336 if (my.mc_wbuf[0] == NULL) { 9337 /* _aligned_malloc() sets errno, but we use Windows error codes */ 9338 rc = ERROR_NOT_ENOUGH_MEMORY; 9339 goto done; 9340 } 9341 #else 9342 if ((rc = pthread_mutex_init(&my.mc_mutex, NULL)) != 0) 9343 return rc; 9344 if ((rc = pthread_cond_init(&my.mc_cond, NULL)) != 0) 9345 goto done2; 9346 #ifdef HAVE_MEMALIGN 9347 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2); 9348 if (my.mc_wbuf[0] == NULL) { 9349 rc = errno; 9350 goto done; 9351 } 9352 #else 9353 { 9354 void *p; 9355 if ((rc = posix_memalign(&p, env->me_os_psize, MDB_WBUF*2)) != 0) 9356 goto done; 9357 my.mc_wbuf[0] = p; 9358 } 9359 #endif 9360 #endif 9361 memset(my.mc_wbuf[0], 0, MDB_WBUF*2); 9362 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF; 9363 my.mc_next_pgno = NUM_METAS; 9364 my.mc_env = env; 9365 my.mc_fd = fd; 9366 rc = THREAD_CREATE(thr, mdb_env_copythr, &my); 9367 if (rc) 9368 goto done; 9369 9370 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn); 9371 if (rc) 9372 goto finish; 9373 9374 mp = (MDB_page *)my.mc_wbuf[0]; 9375 memset(mp, 0, NUM_METAS * env->me_psize); 9376 mp->mp_pgno = 0; 9377 mp->mp_flags = P_META; 9378 mm = (MDB_meta *)METADATA(mp); 9379 mdb_env_init_meta0(env, mm); 9380 mm->mm_address = env->me_metas[0]->mm_address; 9381 9382 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize); 9383 mp->mp_pgno = 1; 9384 mp->mp_flags = P_META; 9385 *(MDB_meta *)METADATA(mp) = *mm; 9386 mm = (MDB_meta *)METADATA(mp); 9387 9388 /* Set metapage 1 with current main DB */ 9389 root = new_root = txn->mt_dbs[MAIN_DBI].md_root; 9390 if (root != P_INVALID) { 9391 /* Count free pages + freeDB pages. Subtract from last_pg 9392 * to find the new last_pg, which also becomes the new root. 9393 */ 9394 MDB_ID freecount = 0; 9395 MDB_cursor mc; 9396 MDB_val key, data; 9397 mdb_cursor_init(&mc, txn, FREE_DBI, NULL); 9398 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0) 9399 freecount += *(MDB_ID *)data.mv_data; 9400 if (rc != MDB_NOTFOUND) 9401 goto finish; 9402 freecount += txn->mt_dbs[FREE_DBI].md_branch_pages + 9403 txn->mt_dbs[FREE_DBI].md_leaf_pages + 9404 txn->mt_dbs[FREE_DBI].md_overflow_pages; 9405 9406 new_root = txn->mt_next_pgno - 1 - freecount; 9407 mm->mm_last_pg = new_root; 9408 mm->mm_dbs[MAIN_DBI] = txn->mt_dbs[MAIN_DBI]; 9409 mm->mm_dbs[MAIN_DBI].md_root = new_root; 9410 } else { 9411 /* When the DB is empty, handle it specially to 9412 * fix any breakage like page leaks from ITS#8174. 9413 */ 9414 mm->mm_dbs[MAIN_DBI].md_flags = txn->mt_dbs[MAIN_DBI].md_flags; 9415 } 9416 if (root != P_INVALID || mm->mm_dbs[MAIN_DBI].md_flags) { 9417 mm->mm_txnid = 1; /* use metapage 1 */ 9418 } 9419 9420 my.mc_wlen[0] = env->me_psize * NUM_METAS; 9421 my.mc_txn = txn; 9422 rc = mdb_env_cwalk(&my, &root, 0); 9423 if (rc == MDB_SUCCESS && root != new_root) { 9424 rc = MDB_INCOMPATIBLE; /* page leak or corrupt DB */ 9425 } 9426 9427 finish: 9428 if (rc) 9429 my.mc_error = rc; 9430 mdb_env_cthr_toggle(&my, 1 | MDB_EOF); 9431 rc = THREAD_FINISH(thr); 9432 mdb_txn_abort(txn); 9433 9434 done: 9435 #ifdef _WIN32 9436 if (my.mc_wbuf[0]) _aligned_free(my.mc_wbuf[0]); 9437 if (my.mc_cond) CloseHandle(my.mc_cond); 9438 if (my.mc_mutex) CloseHandle(my.mc_mutex); 9439 #else 9440 free(my.mc_wbuf[0]); 9441 pthread_cond_destroy(&my.mc_cond); 9442 done2: 9443 pthread_mutex_destroy(&my.mc_mutex); 9444 #endif 9445 return rc ? rc : my.mc_error; 9446 } 9447 9448 /** Copy environment as-is. */ 9449 static int ESECT 9450 mdb_env_copyfd0(MDB_env *env, HANDLE fd) 9451 { 9452 MDB_txn *txn = NULL; 9453 mdb_mutexref_t wmutex = NULL; 9454 int rc; 9455 size_t wsize, w3; 9456 char *ptr; 9457 #ifdef _WIN32 9458 DWORD len, w2; 9459 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL) 9460 #else 9461 ssize_t len; 9462 size_t w2; 9463 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0) 9464 #endif 9465 9466 /* Do the lock/unlock of the reader mutex before starting the 9467 * write txn. Otherwise other read txns could block writers. 9468 */ 9469 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn); 9470 if (rc) 9471 return rc; 9472 9473 if (env->me_txns) { 9474 /* We must start the actual read txn after blocking writers */ 9475 mdb_txn_end(txn, MDB_END_RESET_TMP); 9476 9477 /* Temporarily block writers until we snapshot the meta pages */ 9478 wmutex = env->me_wmutex; 9479 if (LOCK_MUTEX(rc, env, wmutex)) 9480 goto leave; 9481 9482 rc = mdb_txn_renew0(txn); 9483 if (rc) { 9484 UNLOCK_MUTEX(wmutex); 9485 goto leave; 9486 } 9487 } 9488 9489 wsize = env->me_psize * NUM_METAS; 9490 ptr = env->me_map; 9491 w2 = wsize; 9492 while (w2 > 0) { 9493 DO_WRITE(rc, fd, ptr, w2, len); 9494 if (!rc) { 9495 rc = ErrCode(); 9496 break; 9497 } else if (len > 0) { 9498 rc = MDB_SUCCESS; 9499 ptr += len; 9500 w2 -= len; 9501 continue; 9502 } else { 9503 /* Non-blocking or async handles are not supported */ 9504 rc = EIO; 9505 break; 9506 } 9507 } 9508 if (wmutex) 9509 UNLOCK_MUTEX(wmutex); 9510 9511 if (rc) 9512 goto leave; 9513 9514 w3 = txn->mt_next_pgno * env->me_psize; 9515 { 9516 size_t fsize = 0; 9517 if ((rc = mdb_fsize(env->me_fd, &fsize))) 9518 goto leave; 9519 if (w3 > fsize) 9520 w3 = fsize; 9521 } 9522 wsize = w3 - wsize; 9523 while (wsize > 0) { 9524 if (wsize > MAX_WRITE) 9525 w2 = MAX_WRITE; 9526 else 9527 w2 = wsize; 9528 DO_WRITE(rc, fd, ptr, w2, len); 9529 if (!rc) { 9530 rc = ErrCode(); 9531 break; 9532 } else if (len > 0) { 9533 rc = MDB_SUCCESS; 9534 ptr += len; 9535 wsize -= len; 9536 continue; 9537 } else { 9538 rc = EIO; 9539 break; 9540 } 9541 } 9542 9543 leave: 9544 mdb_txn_abort(txn); 9545 return rc; 9546 } 9547 9548 int ESECT 9549 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags) 9550 { 9551 if (flags & MDB_CP_COMPACT) 9552 return mdb_env_copyfd1(env, fd); 9553 else 9554 return mdb_env_copyfd0(env, fd); 9555 } 9556 9557 int ESECT 9558 mdb_env_copyfd(MDB_env *env, HANDLE fd) 9559 { 9560 return mdb_env_copyfd2(env, fd, 0); 9561 } 9562 9563 int ESECT 9564 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags) 9565 { 9566 int rc; 9567 MDB_name fname; 9568 HANDLE newfd = INVALID_HANDLE_VALUE; 9569 9570 rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname); 9571 if (rc == MDB_SUCCESS) { 9572 rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd); 9573 mdb_fname_destroy(fname); 9574 } 9575 if (rc == MDB_SUCCESS) { 9576 rc = mdb_env_copyfd2(env, newfd, flags); 9577 if (close(newfd) < 0 && rc == MDB_SUCCESS) 9578 rc = ErrCode(); 9579 } 9580 return rc; 9581 } 9582 9583 int ESECT 9584 mdb_env_copy(MDB_env *env, const char *path) 9585 { 9586 return mdb_env_copy2(env, path, 0); 9587 } 9588 9589 int ESECT 9590 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff) 9591 { 9592 if (flag & ~CHANGEABLE) 9593 return EINVAL; 9594 if (onoff) 9595 env->me_flags |= flag; 9596 else 9597 env->me_flags &= ~flag; 9598 return MDB_SUCCESS; 9599 } 9600 9601 int ESECT 9602 mdb_env_get_flags(MDB_env *env, unsigned int *arg) 9603 { 9604 if (!env || !arg) 9605 return EINVAL; 9606 9607 *arg = env->me_flags & (CHANGEABLE|CHANGELESS); 9608 return MDB_SUCCESS; 9609 } 9610 9611 int ESECT 9612 mdb_env_set_userctx(MDB_env *env, void *ctx) 9613 { 9614 if (!env) 9615 return EINVAL; 9616 env->me_userctx = ctx; 9617 return MDB_SUCCESS; 9618 } 9619 9620 void * ESECT 9621 mdb_env_get_userctx(MDB_env *env) 9622 { 9623 return env ? env->me_userctx : NULL; 9624 } 9625 9626 int ESECT 9627 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func) 9628 { 9629 if (!env) 9630 return EINVAL; 9631 #ifndef NDEBUG 9632 env->me_assert_func = func; 9633 #endif 9634 return MDB_SUCCESS; 9635 } 9636 9637 int ESECT 9638 mdb_env_get_path(MDB_env *env, const char **arg) 9639 { 9640 if (!env || !arg) 9641 return EINVAL; 9642 9643 *arg = env->me_path; 9644 return MDB_SUCCESS; 9645 } 9646 9647 int ESECT 9648 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg) 9649 { 9650 if (!env || !arg) 9651 return EINVAL; 9652 9653 *arg = env->me_fd; 9654 return MDB_SUCCESS; 9655 } 9656 9657 /** Common code for #mdb_stat() and #mdb_env_stat(). 9658 * @param[in] env the environment to operate in. 9659 * @param[in] db the #MDB_db record containing the stats to return. 9660 * @param[out] arg the address of an #MDB_stat structure to receive the stats. 9661 * @return 0, this function always succeeds. 9662 */ 9663 static int ESECT 9664 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg) 9665 { 9666 arg->ms_psize = env->me_psize; 9667 arg->ms_depth = db->md_depth; 9668 arg->ms_branch_pages = db->md_branch_pages; 9669 arg->ms_leaf_pages = db->md_leaf_pages; 9670 arg->ms_overflow_pages = db->md_overflow_pages; 9671 arg->ms_entries = db->md_entries; 9672 9673 return MDB_SUCCESS; 9674 } 9675 9676 int ESECT 9677 mdb_env_stat(MDB_env *env, MDB_stat *arg) 9678 { 9679 MDB_meta *meta; 9680 9681 if (env == NULL || arg == NULL) 9682 return EINVAL; 9683 9684 meta = mdb_env_pick_meta(env); 9685 9686 return mdb_stat0(env, &meta->mm_dbs[MAIN_DBI], arg); 9687 } 9688 9689 int ESECT 9690 mdb_env_info(MDB_env *env, MDB_envinfo *arg) 9691 { 9692 MDB_meta *meta; 9693 9694 if (env == NULL || arg == NULL) 9695 return EINVAL; 9696 9697 meta = mdb_env_pick_meta(env); 9698 arg->me_mapaddr = meta->mm_address; 9699 arg->me_last_pgno = meta->mm_last_pg; 9700 arg->me_last_txnid = meta->mm_txnid; 9701 9702 arg->me_mapsize = env->me_mapsize; 9703 arg->me_maxreaders = env->me_maxreaders; 9704 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0; 9705 return MDB_SUCCESS; 9706 } 9707 9708 /** Set the default comparison functions for a database. 9709 * Called immediately after a database is opened to set the defaults. 9710 * The user can then override them with #mdb_set_compare() or 9711 * #mdb_set_dupsort(). 9712 * @param[in] txn A transaction handle returned by #mdb_txn_begin() 9713 * @param[in] dbi A database handle returned by #mdb_dbi_open() 9714 */ 9715 static void 9716 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi) 9717 { 9718 uint16_t f = txn->mt_dbs[dbi].md_flags; 9719 9720 txn->mt_dbxs[dbi].md_cmp = 9721 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr : 9722 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn; 9723 9724 txn->mt_dbxs[dbi].md_dcmp = 9725 !(f & MDB_DUPSORT) ? 0 : 9726 ((f & MDB_INTEGERDUP) 9727 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint) 9728 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn)); 9729 } 9730 9731 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi) 9732 { 9733 MDB_val key, data; 9734 MDB_dbi i; 9735 MDB_cursor mc; 9736 MDB_db dummy; 9737 int rc, dbflag, exact; 9738 unsigned int unused = 0, seq; 9739 char *namedup; 9740 size_t len; 9741 9742 if (flags & ~VALID_FLAGS) 9743 return EINVAL; 9744 if (txn->mt_flags & MDB_TXN_BLOCKED) 9745 return MDB_BAD_TXN; 9746 9747 /* main DB? */ 9748 if (!name) { 9749 *dbi = MAIN_DBI; 9750 if (flags & PERSISTENT_FLAGS) { 9751 uint16_t f2 = flags & PERSISTENT_FLAGS; 9752 /* make sure flag changes get committed */ 9753 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) { 9754 txn->mt_dbs[MAIN_DBI].md_flags |= f2; 9755 txn->mt_flags |= MDB_TXN_DIRTY; 9756 } 9757 } 9758 mdb_default_cmp(txn, MAIN_DBI); 9759 return MDB_SUCCESS; 9760 } 9761 9762 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) { 9763 mdb_default_cmp(txn, MAIN_DBI); 9764 } 9765 9766 /* Is the DB already open? */ 9767 len = strlen(name); 9768 for (i=CORE_DBS; i<txn->mt_numdbs; i++) { 9769 if (!txn->mt_dbxs[i].md_name.mv_size) { 9770 /* Remember this free slot */ 9771 if (!unused) unused = i; 9772 continue; 9773 } 9774 if (len == txn->mt_dbxs[i].md_name.mv_size && 9775 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) { 9776 *dbi = i; 9777 return MDB_SUCCESS; 9778 } 9779 } 9780 9781 /* If no free slot and max hit, fail */ 9782 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs) 9783 return MDB_DBS_FULL; 9784 9785 /* Cannot mix named databases with some mainDB flags */ 9786 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY)) 9787 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND; 9788 9789 /* Find the DB info */ 9790 dbflag = DB_NEW|DB_VALID|DB_USRVALID; 9791 exact = 0; 9792 key.mv_size = len; 9793 key.mv_data = (void *)name; 9794 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL); 9795 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact); 9796 if (rc == MDB_SUCCESS) { 9797 /* make sure this is actually a DB */ 9798 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]); 9799 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA) 9800 return MDB_INCOMPATIBLE; 9801 } else { 9802 if (rc != MDB_NOTFOUND || !(flags & MDB_CREATE)) 9803 return rc; 9804 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) 9805 return EACCES; 9806 } 9807 9808 /* Done here so we cannot fail after creating a new DB */ 9809 if ((namedup = strdup(name)) == NULL) 9810 return ENOMEM; 9811 9812 if (rc) { 9813 /* MDB_NOTFOUND and MDB_CREATE: Create new DB */ 9814 data.mv_size = sizeof(MDB_db); 9815 data.mv_data = &dummy; 9816 memset(&dummy, 0, sizeof(dummy)); 9817 dummy.md_root = P_INVALID; 9818 dummy.md_flags = flags & PERSISTENT_FLAGS; 9819 WITH_CURSOR_TRACKING(mc, 9820 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA)); 9821 dbflag |= DB_DIRTY; 9822 } 9823 9824 if (rc) { 9825 free(namedup); 9826 } else { 9827 /* Got info, register DBI in this txn */ 9828 unsigned int slot = unused ? unused : txn->mt_numdbs; 9829 txn->mt_dbxs[slot].md_name.mv_data = namedup; 9830 txn->mt_dbxs[slot].md_name.mv_size = len; 9831 txn->mt_dbxs[slot].md_rel = NULL; 9832 txn->mt_dbflags[slot] = dbflag; 9833 /* txn-> and env-> are the same in read txns, use 9834 * tmp variable to avoid undefined assignment 9835 */ 9836 seq = ++txn->mt_env->me_dbiseqs[slot]; 9837 txn->mt_dbiseqs[slot] = seq; 9838 9839 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db)); 9840 *dbi = slot; 9841 mdb_default_cmp(txn, slot); 9842 if (!unused) { 9843 txn->mt_numdbs++; 9844 } 9845 } 9846 9847 return rc; 9848 } 9849 9850 int ESECT 9851 mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg) 9852 { 9853 if (!arg || !TXN_DBI_EXIST(txn, dbi, DB_VALID)) 9854 return EINVAL; 9855 9856 if (txn->mt_flags & MDB_TXN_BLOCKED) 9857 return MDB_BAD_TXN; 9858 9859 if (txn->mt_dbflags[dbi] & DB_STALE) { 9860 MDB_cursor mc; 9861 MDB_xcursor mx; 9862 /* Stale, must read the DB's root. cursor_init does it for us. */ 9863 mdb_cursor_init(&mc, txn, dbi, &mx); 9864 } 9865 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg); 9866 } 9867 9868 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi) 9869 { 9870 char *ptr; 9871 if (dbi < CORE_DBS || dbi >= env->me_maxdbs) 9872 return; 9873 ptr = env->me_dbxs[dbi].md_name.mv_data; 9874 /* If there was no name, this was already closed */ 9875 if (ptr) { 9876 env->me_dbxs[dbi].md_name.mv_data = NULL; 9877 env->me_dbxs[dbi].md_name.mv_size = 0; 9878 env->me_dbflags[dbi] = 0; 9879 env->me_dbiseqs[dbi]++; 9880 free(ptr); 9881 } 9882 } 9883 9884 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags) 9885 { 9886 /* We could return the flags for the FREE_DBI too but what's the point? */ 9887 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 9888 return EINVAL; 9889 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS; 9890 return MDB_SUCCESS; 9891 } 9892 9893 /** Add all the DB's pages to the free list. 9894 * @param[in] mc Cursor on the DB to free. 9895 * @param[in] subs non-Zero to check for sub-DBs in this DB. 9896 * @return 0 on success, non-zero on failure. 9897 */ 9898 static int 9899 mdb_drop0(MDB_cursor *mc, int subs) 9900 { 9901 int rc; 9902 9903 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST); 9904 if (rc == MDB_SUCCESS) { 9905 MDB_txn *txn = mc->mc_txn; 9906 MDB_node *ni; 9907 MDB_cursor mx; 9908 unsigned int i; 9909 9910 /* DUPSORT sub-DBs have no ovpages/DBs. Omit scanning leaves. 9911 * This also avoids any P_LEAF2 pages, which have no nodes. 9912 * Also if the DB doesn't have sub-DBs and has no overflow 9913 * pages, omit scanning leaves. 9914 */ 9915 if ((mc->mc_flags & C_SUB) || 9916 (!subs && !mc->mc_db->md_overflow_pages)) 9917 mdb_cursor_pop(mc); 9918 9919 mdb_cursor_copy(mc, &mx); 9920 while (mc->mc_snum > 0) { 9921 MDB_page *mp = mc->mc_pg[mc->mc_top]; 9922 unsigned n = NUMKEYS(mp); 9923 if (IS_LEAF(mp)) { 9924 for (i=0; i<n; i++) { 9925 ni = NODEPTR(mp, i); 9926 if (ni->mn_flags & F_BIGDATA) { 9927 MDB_page *omp; 9928 pgno_t pg; 9929 memcpy(&pg, NODEDATA(ni), sizeof(pg)); 9930 rc = mdb_page_get(mc, pg, &omp, NULL); 9931 if (rc != 0) 9932 goto done; 9933 mdb_cassert(mc, IS_OVERFLOW(omp)); 9934 rc = mdb_midl_append_range(&txn->mt_free_pgs, 9935 pg, omp->mp_pages); 9936 if (rc) 9937 goto done; 9938 mc->mc_db->md_overflow_pages -= omp->mp_pages; 9939 if (!mc->mc_db->md_overflow_pages && !subs) 9940 break; 9941 } else if (subs && (ni->mn_flags & F_SUBDATA)) { 9942 mdb_xcursor_init1(mc, ni); 9943 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0); 9944 if (rc) 9945 goto done; 9946 } 9947 } 9948 if (!subs && !mc->mc_db->md_overflow_pages) 9949 goto pop; 9950 } else { 9951 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0) 9952 goto done; 9953 for (i=0; i<n; i++) { 9954 pgno_t pg; 9955 ni = NODEPTR(mp, i); 9956 pg = NODEPGNO(ni); 9957 /* free it */ 9958 mdb_midl_xappend(txn->mt_free_pgs, pg); 9959 } 9960 } 9961 if (!mc->mc_top) 9962 break; 9963 mc->mc_ki[mc->mc_top] = i; 9964 rc = mdb_cursor_sibling(mc, 1); 9965 if (rc) { 9966 if (rc != MDB_NOTFOUND) 9967 goto done; 9968 /* no more siblings, go back to beginning 9969 * of previous level. 9970 */ 9971 pop: 9972 mdb_cursor_pop(mc); 9973 mc->mc_ki[0] = 0; 9974 for (i=1; i<mc->mc_snum; i++) { 9975 mc->mc_ki[i] = 0; 9976 mc->mc_pg[i] = mx.mc_pg[i]; 9977 } 9978 } 9979 } 9980 /* free it */ 9981 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root); 9982 done: 9983 if (rc) 9984 txn->mt_flags |= MDB_TXN_ERROR; 9985 } else if (rc == MDB_NOTFOUND) { 9986 rc = MDB_SUCCESS; 9987 } 9988 mc->mc_flags &= ~C_INITIALIZED; 9989 return rc; 9990 } 9991 9992 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del) 9993 { 9994 MDB_cursor *mc, *m2; 9995 int rc; 9996 9997 if ((unsigned)del > 1 || !TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 9998 return EINVAL; 9999 10000 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) 10001 return EACCES; 10002 10003 if (TXN_DBI_CHANGED(txn, dbi)) 10004 return MDB_BAD_DBI; 10005 10006 rc = mdb_cursor_open(txn, dbi, &mc); 10007 if (rc) 10008 return rc; 10009 10010 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT); 10011 /* Invalidate the dropped DB's cursors */ 10012 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next) 10013 m2->mc_flags &= ~(C_INITIALIZED|C_EOF); 10014 if (rc) 10015 goto leave; 10016 10017 /* Can't delete the main DB */ 10018 if (del && dbi >= CORE_DBS) { 10019 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA); 10020 if (!rc) { 10021 txn->mt_dbflags[dbi] = DB_STALE; 10022 mdb_dbi_close(txn->mt_env, dbi); 10023 } else { 10024 txn->mt_flags |= MDB_TXN_ERROR; 10025 } 10026 } else { 10027 /* reset the DB record, mark it dirty */ 10028 txn->mt_dbflags[dbi] |= DB_DIRTY; 10029 txn->mt_dbs[dbi].md_depth = 0; 10030 txn->mt_dbs[dbi].md_branch_pages = 0; 10031 txn->mt_dbs[dbi].md_leaf_pages = 0; 10032 txn->mt_dbs[dbi].md_overflow_pages = 0; 10033 txn->mt_dbs[dbi].md_entries = 0; 10034 txn->mt_dbs[dbi].md_root = P_INVALID; 10035 10036 txn->mt_flags |= MDB_TXN_DIRTY; 10037 } 10038 leave: 10039 mdb_cursor_close(mc); 10040 return rc; 10041 } 10042 10043 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp) 10044 { 10045 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10046 return EINVAL; 10047 10048 txn->mt_dbxs[dbi].md_cmp = cmp; 10049 return MDB_SUCCESS; 10050 } 10051 10052 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp) 10053 { 10054 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10055 return EINVAL; 10056 10057 txn->mt_dbxs[dbi].md_dcmp = cmp; 10058 return MDB_SUCCESS; 10059 } 10060 10061 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel) 10062 { 10063 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10064 return EINVAL; 10065 10066 txn->mt_dbxs[dbi].md_rel = rel; 10067 return MDB_SUCCESS; 10068 } 10069 10070 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx) 10071 { 10072 if (!TXN_DBI_EXIST(txn, dbi, DB_USRVALID)) 10073 return EINVAL; 10074 10075 txn->mt_dbxs[dbi].md_relctx = ctx; 10076 return MDB_SUCCESS; 10077 } 10078 10079 int ESECT 10080 mdb_env_get_maxkeysize(MDB_env *env) 10081 { 10082 return ENV_MAXKEY(env); 10083 } 10084 10085 int ESECT 10086 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx) 10087 { 10088 unsigned int i, rdrs; 10089 MDB_reader *mr; 10090 char buf[64]; 10091 int rc = 0, first = 1; 10092 10093 if (!env || !func) 10094 return -1; 10095 if (!env->me_txns) { 10096 return func("(no reader locks)\n", ctx); 10097 } 10098 rdrs = env->me_txns->mti_numreaders; 10099 mr = env->me_txns->mti_readers; 10100 for (i=0; i<rdrs; i++) { 10101 if (mr[i].mr_pid) { 10102 txnid_t txnid = mr[i].mr_txnid; 10103 sprintf(buf, txnid == (txnid_t)-1 ? 10104 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n", 10105 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid); 10106 if (first) { 10107 first = 0; 10108 rc = func(" pid thread txnid\n", ctx); 10109 if (rc < 0) 10110 break; 10111 } 10112 rc = func(buf, ctx); 10113 if (rc < 0) 10114 break; 10115 } 10116 } 10117 if (first) { 10118 rc = func("(no active readers)\n", ctx); 10119 } 10120 return rc; 10121 } 10122 10123 /** Insert pid into list if not already present. 10124 * return -1 if already present. 10125 */ 10126 static int ESECT 10127 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid) 10128 { 10129 /* binary search of pid in list */ 10130 unsigned base = 0; 10131 unsigned cursor = 1; 10132 int val = 0; 10133 unsigned n = ids[0]; 10134 10135 while( 0 < n ) { 10136 unsigned pivot = n >> 1; 10137 cursor = base + pivot + 1; 10138 val = pid - ids[cursor]; 10139 10140 if( val < 0 ) { 10141 n = pivot; 10142 10143 } else if ( val > 0 ) { 10144 base = cursor; 10145 n -= pivot + 1; 10146 10147 } else { 10148 /* found, so it's a duplicate */ 10149 return -1; 10150 } 10151 } 10152 10153 if( val > 0 ) { 10154 ++cursor; 10155 } 10156 ids[0]++; 10157 for (n = ids[0]; n > cursor; n--) 10158 ids[n] = ids[n-1]; 10159 ids[n] = pid; 10160 return 0; 10161 } 10162 10163 int ESECT 10164 mdb_reader_check(MDB_env *env, int *dead) 10165 { 10166 if (!env) 10167 return EINVAL; 10168 if (dead) 10169 *dead = 0; 10170 return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS; 10171 } 10172 10173 /** As #mdb_reader_check(). \b rlocked is set if caller locked #me_rmutex. */ 10174 static int ESECT 10175 mdb_reader_check0(MDB_env *env, int rlocked, int *dead) 10176 { 10177 mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex; 10178 unsigned int i, j, rdrs; 10179 MDB_reader *mr; 10180 MDB_PID_T *pids, pid; 10181 int rc = MDB_SUCCESS, count = 0; 10182 10183 rdrs = env->me_txns->mti_numreaders; 10184 pids = malloc((rdrs+1) * sizeof(MDB_PID_T)); 10185 if (!pids) 10186 return ENOMEM; 10187 pids[0] = 0; 10188 mr = env->me_txns->mti_readers; 10189 for (i=0; i<rdrs; i++) { 10190 pid = mr[i].mr_pid; 10191 if (pid && pid != env->me_pid) { 10192 if (mdb_pid_insert(pids, pid) == 0) { 10193 if (!mdb_reader_pid(env, Pidcheck, pid)) { 10194 /* Stale reader found */ 10195 j = i; 10196 if (rmutex) { 10197 if ((rc = LOCK_MUTEX0(rmutex)) != 0) { 10198 if ((rc = mdb_mutex_failed(env, rmutex, rc))) 10199 break; 10200 rdrs = 0; /* the above checked all readers */ 10201 } else { 10202 /* Recheck, a new process may have reused pid */ 10203 if (mdb_reader_pid(env, Pidcheck, pid)) 10204 j = rdrs; 10205 } 10206 } 10207 for (; j<rdrs; j++) 10208 if (mr[j].mr_pid == pid) { 10209 DPRINTF(("clear stale reader pid %u txn %"Z"d", 10210 (unsigned) pid, mr[j].mr_txnid)); 10211 mr[j].mr_pid = 0; 10212 count++; 10213 } 10214 if (rmutex) 10215 UNLOCK_MUTEX(rmutex); 10216 } 10217 } 10218 } 10219 } 10220 free(pids); 10221 if (dead) 10222 *dead = count; 10223 return rc; 10224 } 10225 10226 #ifdef MDB_ROBUST_SUPPORTED 10227 /** Handle #LOCK_MUTEX0() failure. 10228 * Try to repair the lock file if the mutex owner died. 10229 * @param[in] env the environment handle 10230 * @param[in] mutex LOCK_MUTEX0() mutex 10231 * @param[in] rc LOCK_MUTEX0() error (nonzero) 10232 * @return 0 on success with the mutex locked, or an error code on failure. 10233 */ 10234 static int ESECT 10235 mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc) 10236 { 10237 int rlocked, rc2; 10238 MDB_meta *meta; 10239 10240 if (rc == MDB_OWNERDEAD) { 10241 /* We own the mutex. Clean up after dead previous owner. */ 10242 rc = MDB_SUCCESS; 10243 rlocked = (mutex == env->me_rmutex); 10244 if (!rlocked) { 10245 /* Keep mti_txnid updated, otherwise next writer can 10246 * overwrite data which latest meta page refers to. 10247 */ 10248 meta = mdb_env_pick_meta(env); 10249 env->me_txns->mti_txnid = meta->mm_txnid; 10250 /* env is hosed if the dead thread was ours */ 10251 if (env->me_txn) { 10252 env->me_flags |= MDB_FATAL_ERROR; 10253 env->me_txn = NULL; 10254 rc = MDB_PANIC; 10255 } 10256 } 10257 DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'), 10258 (rc ? "this process' env is hosed" : "recovering"))); 10259 rc2 = mdb_reader_check0(env, rlocked, NULL); 10260 if (rc2 == 0) 10261 rc2 = mdb_mutex_consistent(mutex); 10262 if (rc || (rc = rc2)) { 10263 DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc))); 10264 UNLOCK_MUTEX(mutex); 10265 } 10266 } else { 10267 #ifdef _WIN32 10268 rc = ErrCode(); 10269 #endif 10270 DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc))); 10271 } 10272 10273 return rc; 10274 } 10275 #endif /* MDB_ROBUST_SUPPORTED */ 10276 10277 #if defined(_WIN32) 10278 /** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */ 10279 static int ESECT 10280 utf8_to_utf16(const char *src, MDB_name *dst, int xtra) 10281 { 10282 int rc, need = 0; 10283 wchar_t *result = NULL; 10284 for (;;) { /* malloc result, then fill it in */ 10285 need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need); 10286 if (!need) { 10287 rc = ErrCode(); 10288 free(result); 10289 return rc; 10290 } 10291 if (!result) { 10292 result = malloc(sizeof(wchar_t) * (need + xtra)); 10293 if (!result) 10294 return ENOMEM; 10295 continue; 10296 } 10297 dst->mn_alloced = 1; 10298 dst->mn_len = need - 1; 10299 dst->mn_val = result; 10300 return MDB_SUCCESS; 10301 } 10302 } 10303 #endif /* defined(_WIN32) */ 10304 /** @} */ 10305