X-Git-Url: https://git.sur5r.net/?a=blobdiff_plain;f=libraries%2Fliblmdb%2Fmdb.c;h=e31867de8207a9f0476c0707185e107f6feee4fa;hb=f6514da36080efb8876d274c1d01d9c9207335f7;hp=4e4385c04c547846dd71d779e96bfce435524a59;hpb=f2ecddbcf76dec209de2e4d15f2a2d7eb41ce20f;p=openldap diff --git a/libraries/liblmdb/mdb.c b/libraries/liblmdb/mdb.c index 4e4385c04c..e31867de82 100644 --- a/libraries/liblmdb/mdb.c +++ b/libraries/liblmdb/mdb.c @@ -5,7 +5,7 @@ * BerkeleyDB API, but much simplified. */ /* - * Copyright 2011-2016 Howard Chu, Symas Corp. + * Copyright 2011-2017 Howard Chu, Symas Corp. * All rights reserved. * * Redistribution and use in source and binary forms, with or without @@ -41,6 +41,8 @@ #ifdef _WIN32 #include #include +#include /* get wcscpy() */ + /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it * as int64 which is wrong. MSVC doesn't define it at all, so just * don't use it. @@ -111,6 +113,10 @@ typedef SSIZE_T ssize_t; /* Most platforms have posix_memalign, older may only have memalign */ #define HAVE_MEMALIGN 1 #include +/* On Solaris, we need the POSIX sigwait function */ +#if defined (__sun) +# define _POSIX_PTHREAD_SEMANTICS 1 +#endif #endif #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER)) @@ -118,7 +124,7 @@ typedef SSIZE_T ssize_t; #include /* defines BYTE_ORDER on HPUX and Solaris */ #endif -#if defined(__APPLE__) || defined (BSD) +#if defined(__APPLE__) || defined (BSD) || defined(__FreeBSD_kernel__) # define MDB_USE_POSIX_SEM 1 # define MDB_FDATASYNC fsync #elif defined(ANDROID) @@ -127,6 +133,7 @@ typedef SSIZE_T ssize_t; #ifndef _WIN32 #include +#include #ifdef MDB_USE_POSIX_SEM # define MDB_USE_HASH 1 #include @@ -338,15 +345,15 @@ mdb_sem_wait(sem_t *sem) } #else /* MDB_USE_POSIX_MUTEX: */ - /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as - * local variables keep it (mdb_mutexref_t). + /** Shared mutex/semaphore as the original is stored. * - * When #mdb_mutexref_t is a pointer declaration and #mdb_mutex_t is - * not, then it is array[size 1] so it can be assigned to a pointer. - * @{ + * Not for copies. Instead it can be assigned to an #mdb_mutexref_t. + * When mdb_mutexref_t is a pointer and mdb_mutex_t is not, then it + * is array[size 1] so it can be assigned to the pointer. */ -typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t; - /* @} */ +typedef pthread_mutex_t mdb_mutex_t[1]; + /** Reference to an #mdb_mutex_t */ +typedef pthread_mutex_t *mdb_mutexref_t; /** Lock the reader or writer mutex. * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX(). */ @@ -871,19 +878,26 @@ typedef struct MDB_page { /** Header for a single key/data pair within a page. * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2. * We guarantee 2-byte alignment for 'MDB_node's. + * + * #mn_lo and #mn_hi are used for data size on leaf nodes, and for child + * pgno on branch nodes. On 64 bit platforms, #mn_flags is also used + * for pgno. (Branch nodes have no flags). Lo and hi are in host byte + * order in case some accesses can be optimized to 32-bit word access. + * + * Leaf node flags describe node contents. #F_BIGDATA says the node's + * data part is the page number of an overflow page with actual data. + * #F_DUPDATA and #F_SUBDATA can be combined giving duplicate data in + * a sub-page/sub-database, and named databases (just #F_SUBDATA). */ typedef struct MDB_node { - /** lo and hi are used for data size on leaf nodes and for - * child pgno on branch nodes. On 64 bit platforms, flags - * is also used for pgno. (Branch nodes have no flags). - * They are in host byte order in case that lets some - * accesses be optimized into a 32-bit word access. - */ + /** part of data size or pgno + * @{ */ #if BYTE_ORDER == LITTLE_ENDIAN - unsigned short mn_lo, mn_hi; /**< part of data size or pgno */ + unsigned short mn_lo, mn_hi; #else unsigned short mn_hi, mn_lo; #endif + /** @} */ /** @defgroup mdb_node Node Flags * @ingroup internal * Flags for node headers. @@ -1212,17 +1226,19 @@ typedef struct MDB_xcursor { unsigned char mx_dbflag; } MDB_xcursor; - /** Check if there is an inited xcursor, so #XCURSOR_REFRESH() is proper */ + /** Check if there is an inited xcursor */ #define XCURSOR_INITED(mc) \ ((mc)->mc_xcursor && ((mc)->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) - /** Update sub-page pointer, if any, in \b mc->mc_xcursor. Needed + /** Update the xcursor's sub-page pointer, if any, in \b mc. Needed * when the node which contains the sub-page may have moved. Called - * with \b mp = mc->mc_pg[mc->mc_top], \b ki = mc->mc_ki[mc->mc_top]. + * with leaf page \b mp = mc->mc_pg[\b top]. */ -#define XCURSOR_REFRESH(mc, mp, ki) do { \ +#define XCURSOR_REFRESH(mc, top, mp) do { \ MDB_page *xr_pg = (mp); \ - MDB_node *xr_node = NODEPTR(xr_pg, ki); \ + MDB_node *xr_node; \ + if (!XCURSOR_INITED(mc) || (mc)->mc_ki[top] >= NUMKEYS(xr_pg)) break; \ + xr_node = NODEPTR(xr_pg, (mc)->mc_ki[top]); \ if ((xr_node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) \ (mc)->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(xr_node); \ } while (0) @@ -1237,7 +1253,7 @@ typedef struct MDB_pgstate { struct MDB_env { HANDLE me_fd; /**< The main data file */ HANDLE me_lfd; /**< The lock file */ - HANDLE me_mfd; /**< just for writing the meta pages */ + HANDLE me_mfd; /**< For writing and syncing the meta pages */ /** Failed to update the meta page. Probably an I/O error. */ #define MDB_FATAL_ERROR 0x80000000U /** Some fields are initialized. */ @@ -1415,7 +1431,8 @@ static SECURITY_DESCRIPTOR mdb_null_sd; static SECURITY_ATTRIBUTES mdb_all_sa; static int mdb_sec_inited; -static int utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize); +struct MDB_name; +static int utf8_to_utf16(const char *src, struct MDB_name *dst, int xtra); #endif /** Return the library version info. */ @@ -1744,6 +1761,7 @@ mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b) /** Allocate memory for a page. * Re-use old malloc'd pages first for singletons, otherwise just malloc. + * Set #MDB_TXN_ERROR on failure. */ static MDB_page * mdb_page_malloc(MDB_txn *txn, unsigned num) @@ -2111,7 +2129,7 @@ mdb_page_dirty(MDB_txn *txn, MDB_page *mp) } /** Allocate page numbers and memory for writing. Maintain me_pglast, - * me_pghead and mt_next_pgno. + * me_pghead and mt_next_pgno. Set #MDB_TXN_ERROR on failure. * * If there are free pages available from older transactions, they * are re-used first. Otherwise allocate a new page at mt_next_pgno. @@ -2234,7 +2252,7 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) np = m2.mc_pg[m2.mc_top]; leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]); if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS) - return rc; + goto fail; idl = (MDB_ID *) data.mv_data; i = idl[0]; @@ -2384,6 +2402,7 @@ mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret) } /** Touch a page: make it dirty and re-insert into tree with updated pgno. + * Set #MDB_TXN_ERROR on failure. * @param[in] mc cursor pointing to the page to be touched * @return 0 on success, non-zero on failure. */ @@ -2472,8 +2491,8 @@ done: if (m2 == mc) continue; if (m2->mc_pg[mc->mc_top] == mp) { m2->mc_pg[mc->mc_top] = np; - if (XCURSOR_INITED(m2) && IS_LEAF(np)) - XCURSOR_REFRESH(m2, np, m2->mc_ki[mc->mc_top]); + if (IS_LEAF(np)) + XCURSOR_REFRESH(m2, mc->mc_top, np); } } } @@ -3824,7 +3843,10 @@ mdb_env_write_meta(MDB_txn *txn) len = sizeof(MDB_meta) - off; off += (char *)mp - env->me_map; - /* Write to the SYNC fd */ + /* Write to the SYNC fd unless MDB_NOSYNC/MDB_NOMETASYNC. + * (me_mfd goes to the same file as me_fd, but writing to it + * also syncs to disk. Avoids a separate fdatasync() call.) + */ mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd; #ifdef _WIN32 { @@ -4079,6 +4101,189 @@ mdb_fsize(HANDLE fd, size_t *size) return MDB_SUCCESS; } + +#ifdef _WIN32 +typedef wchar_t mdb_nchar_t; +# define MDB_NAME(str) L##str +# define mdb_name_cpy wcscpy +#else +/** Character type for file names: char on Unix, wchar_t on Windows */ +typedef char mdb_nchar_t; +# define MDB_NAME(str) str /**< #mdb_nchar_t[] string literal */ +# define mdb_name_cpy strcpy /**< Copy name (#mdb_nchar_t string) */ +#endif + +/** Filename - string of #mdb_nchar_t[] */ +typedef struct MDB_name { + int mn_len; /**< Length */ + int mn_alloced; /**< True if #mn_val was malloced */ + mdb_nchar_t *mn_val; /**< Contents */ +} MDB_name; + +/** Filename suffixes [datafile,lockfile][without,with MDB_NOSUBDIR] */ +static const mdb_nchar_t *const mdb_suffixes[2][2] = { + { MDB_NAME("/data.mdb"), MDB_NAME("") }, + { MDB_NAME("/lock.mdb"), MDB_NAME("-lock") } +}; + +#define MDB_SUFFLEN 9 /**< Max string length in #mdb_suffixes[] */ + +/** Set up filename + scratch area for filename suffix, for opening files. + * It should be freed with #mdb_fname_destroy(). + * On Windows, paths are converted from char *UTF-8 to wchar_t *UTF-16. + * + * @param[in] path Pathname for #mdb_env_open(). + * @param[in] envflags Whether a subdir and/or lockfile will be used. + * @param[out] fname Resulting filename, with room for a suffix if necessary. + */ +static int ESECT +mdb_fname_init(const char *path, unsigned envflags, MDB_name *fname) +{ + int no_suffix = F_ISSET(envflags, MDB_NOSUBDIR|MDB_NOLOCK); + fname->mn_alloced = 0; +#ifdef _WIN32 + return utf8_to_utf16(path, fname, no_suffix ? 0 : MDB_SUFFLEN); +#else + fname->mn_len = strlen(path); + if (no_suffix) + fname->mn_val = (char *) path; + else if ((fname->mn_val = malloc(fname->mn_len + MDB_SUFFLEN+1)) != NULL) { + fname->mn_alloced = 1; + strcpy(fname->mn_val, path); + } + else + return ENOMEM; + return MDB_SUCCESS; +#endif +} + +/** Destroy \b fname from #mdb_fname_init() */ +#define mdb_fname_destroy(fname) \ + do { if ((fname).mn_alloced) free((fname).mn_val); } while (0) + +#ifdef O_CLOEXEC /* POSIX.1-2008: Set FD_CLOEXEC atomically at open() */ +# define MDB_CLOEXEC O_CLOEXEC +#else +# define MDB_CLOEXEC 0 +#endif + +/** File type, access mode etc. for #mdb_fopen() */ +enum mdb_fopen_type { +#ifdef _WIN32 + MDB_O_RDONLY, MDB_O_RDWR, MDB_O_META, MDB_O_COPY, MDB_O_LOCKS +#else + /* A comment in mdb_fopen() explains some O_* flag choices. */ + MDB_O_RDONLY= O_RDONLY, /**< for RDONLY me_fd */ + MDB_O_RDWR = O_RDWR |O_CREAT, /**< for me_fd */ + MDB_O_META = O_WRONLY|MDB_DSYNC |MDB_CLOEXEC, /**< for me_mfd */ + MDB_O_COPY = O_WRONLY|O_CREAT|O_EXCL|MDB_CLOEXEC, /**< for #mdb_env_copy() */ + /** Bitmask for open() flags in enum #mdb_fopen_type. The other bits + * distinguish otherwise-equal MDB_O_* constants from each other. + */ + MDB_O_MASK = MDB_O_RDWR|MDB_CLOEXEC | MDB_O_RDONLY|MDB_O_META|MDB_O_COPY, + MDB_O_LOCKS = MDB_O_RDWR|MDB_CLOEXEC | ((MDB_O_MASK+1) & ~MDB_O_MASK) /**< for me_lfd */ +#endif +}; + +/** Open an LMDB file. + * @param[in] env The LMDB environment. + * @param[in,out] fname Path from from #mdb_fname_init(). A suffix is + * appended if necessary to create the filename, without changing mn_len. + * @param[in] which Determines file type, access mode, etc. + * @param[in] mode The Unix permissions for the file, if we create it. + * @param[out] res Resulting file handle. + * @return 0 on success, non-zero on failure. + */ +static int ESECT +mdb_fopen(const MDB_env *env, MDB_name *fname, + enum mdb_fopen_type which, mdb_mode_t mode, + HANDLE *res) +{ + int rc = MDB_SUCCESS; + HANDLE fd; +#ifdef _WIN32 + DWORD acc, share, disp, attrs; +#else + int flags; +#endif + + if (fname->mn_alloced) /* modifiable copy */ + mdb_name_cpy(fname->mn_val + fname->mn_len, + mdb_suffixes[which==MDB_O_LOCKS][F_ISSET(env->me_flags, MDB_NOSUBDIR)]); + + /* The directory must already exist. Usually the file need not. + * MDB_O_META requires the file because we already created it using + * MDB_O_RDWR. MDB_O_COPY must not overwrite an existing file. + * + * With MDB_O_COPY we do not want the OS to cache the writes, since + * the source data is already in the OS cache. + * + * The lockfile needs FD_CLOEXEC (close file descriptor on exec*()) + * to avoid the flock() issues noted under Caveats in lmdb.h. + * Also set it for other filehandles which the user cannot get at + * and close himself, which he may need after fork(). I.e. all but + * me_fd, which programs do use via mdb_env_get_fd(). + */ + +#ifdef _WIN32 + acc = GENERIC_READ|GENERIC_WRITE; + share = FILE_SHARE_READ|FILE_SHARE_WRITE; + disp = OPEN_ALWAYS; + attrs = FILE_ATTRIBUTE_NORMAL; + switch (which) { + case MDB_O_RDONLY: /* read-only datafile */ + acc = GENERIC_READ; + disp = OPEN_EXISTING; + break; + case MDB_O_META: /* for writing metapages */ + acc = GENERIC_WRITE; + disp = OPEN_EXISTING; + attrs = FILE_ATTRIBUTE_NORMAL|FILE_FLAG_WRITE_THROUGH; + break; + case MDB_O_COPY: /* mdb_env_copy() & co */ + acc = GENERIC_WRITE; + share = 0; + disp = CREATE_NEW; + attrs = FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH; + break; + default: break; /* silence gcc -Wswitch (not all enum values handled) */ + } + fd = CreateFileW(fname->mn_val, acc, share, NULL, disp, attrs, NULL); +#else + fd = open(fname->mn_val, which & MDB_O_MASK, mode); +#endif + + if (fd == INVALID_HANDLE_VALUE) + rc = ErrCode(); +#ifndef _WIN32 + else { + if (which != MDB_O_RDONLY && which != MDB_O_RDWR) { + /* Set CLOEXEC if we could not pass it to open() */ + if (!MDB_CLOEXEC && (flags = fcntl(fd, F_GETFD)) != -1) + (void) fcntl(fd, F_SETFD, flags | FD_CLOEXEC); + } + if (which == MDB_O_COPY && env->me_psize >= env->me_os_psize) { + /* This may require buffer alignment. There is no portable + * way to ask how much, so we require OS pagesize alignment. + */ +# ifdef F_NOCACHE /* __APPLE__ */ + (void) fcntl(fd, F_NOCACHE, 1); +# elif defined O_DIRECT + /* open(...O_DIRECT...) would break on filesystems without + * O_DIRECT support (ITS#7682). Try to set it here instead. + */ + if ((flags = fcntl(fd, F_GETFL)) != -1) + (void) fcntl(fd, F_SETFL, flags | O_DIRECT); +# endif + } + } +#endif /* !_WIN32 */ + + *res = fd; + return rc; +} + + #ifdef BROKEN_FDATASYNC #include #include @@ -4244,7 +4449,11 @@ mdb_env_reader_dest(void *ptr) { MDB_reader *reader = ptr; - reader->mr_pid = 0; +#ifndef _WIN32 + if (reader->mr_pid == getpid()) /* catch pthread_exit() in child process */ +#endif + /* We omit the mutex, so do this atomically (i.e. skip mr_txnid) */ + reader->mr_pid = 0; } #ifdef _WIN32 @@ -4484,52 +4693,30 @@ mdb_hash_enc(MDB_val *val, char *encbuf) /** Open and/or initialize the lock region for the environment. * @param[in] env The LMDB environment. - * @param[in] lpath The pathname of the file used for the lock region. + * @param[in] fname Filename + scratch area, from #mdb_fname_init(). * @param[in] mode The Unix permissions for the file, if we create it. * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive * @return 0 on success, non-zero on failure. */ static int ESECT -mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl) +mdb_env_setup_locks(MDB_env *env, MDB_name *fname, int mode, int *excl) { #ifdef _WIN32 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT #else # define MDB_ERRCODE_ROFS EROFS -#ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */ -# define MDB_CLOEXEC O_CLOEXEC -#else -# define MDB_CLOEXEC 0 -#endif - int fdflags; #endif int rc; off_t size, rsize; -#ifdef _WIN32 - wchar_t *wlpath; - rc = utf8_to_utf16(lpath, -1, &wlpath, NULL); - if (rc) - return rc; - env->me_lfd = CreateFileW(wlpath, GENERIC_READ|GENERIC_WRITE, - FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS, - FILE_ATTRIBUTE_NORMAL, NULL); - free(wlpath); -#else - env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode); -#endif - if (env->me_lfd == INVALID_HANDLE_VALUE) { - rc = ErrCode(); + rc = mdb_fopen(env, fname, MDB_O_LOCKS, mode, &env->me_lfd); + if (rc) { + /* Omit lockfile if read-only env on read-only filesystem */ if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) { return MDB_SUCCESS; } goto fail; } -#ifndef _WIN32 - /* Lose record locks when exec*() */ - if (!(MDB_CLOEXEC) && (fdflags = fcntl(env->me_lfd, F_GETFD)) != -1) - fcntl(env->me_lfd, F_SETFD, fdflags | FD_CLOEXEC); -#endif if (!(env->me_flags & MDB_NOTLS)) { rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest); @@ -4719,12 +4906,6 @@ fail: return rc; } - /** The name of the lock file in the DB environment */ -#define LOCKNAME "/lock.mdb" - /** The name of the data file in the DB environment */ -#define DATANAME "/data.mdb" - /** The suffix of the lock file when no subdir is used */ -#define LOCKSUFF "-lock" /** Only a subset of the @ref mdb_env flags can be changed * at runtime. Changing other flags requires closing the * environment and re-opening it with the new flags. @@ -4740,36 +4921,18 @@ fail: int ESECT mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode) { - int oflags, rc, len, excl = -1; - char *lpath, *dpath; -#ifdef _WIN32 - wchar_t *wpath; -#endif + int rc, excl = -1; + MDB_name fname; if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS))) return EINVAL; - len = strlen(path); - if (flags & MDB_NOSUBDIR) { - rc = len + sizeof(LOCKSUFF) + len + 1; - } else { - rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME); - } - lpath = malloc(rc); - if (!lpath) - return ENOMEM; - if (flags & MDB_NOSUBDIR) { - dpath = lpath + len + sizeof(LOCKSUFF); - sprintf(lpath, "%s" LOCKSUFF, path); - strcpy(dpath, path); - } else { - dpath = lpath + len + sizeof(LOCKNAME); - sprintf(lpath, "%s" LOCKNAME, path); - sprintf(dpath, "%s" DATANAME, path); - } - - rc = MDB_SUCCESS; flags |= env->me_flags; + + rc = mdb_fname_init(path, flags, &fname); + if (rc) + return rc; + if (flags & MDB_RDONLY) { /* silently ignore WRITEMAP when we're only getting read access */ flags &= ~MDB_WRITEMAP; @@ -4794,69 +4957,31 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode /* For RDONLY, get lockfile after we know datafile exists */ if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) { - rc = mdb_env_setup_locks(env, lpath, mode, &excl); + rc = mdb_env_setup_locks(env, &fname, mode, &excl); if (rc) goto leave; } -#ifdef _WIN32 - if (F_ISSET(flags, MDB_RDONLY)) { - oflags = GENERIC_READ; - len = OPEN_EXISTING; - } else { - oflags = GENERIC_READ|GENERIC_WRITE; - len = OPEN_ALWAYS; - } - mode = FILE_ATTRIBUTE_NORMAL; - rc = utf8_to_utf16(dpath, -1, &wpath, NULL); + rc = mdb_fopen(env, &fname, + (flags & MDB_RDONLY) ? MDB_O_RDONLY : MDB_O_RDWR, + mode, &env->me_fd); if (rc) goto leave; - env->me_fd = CreateFileW(wpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE, - NULL, len, mode, NULL); - free(wpath); -#else - if (F_ISSET(flags, MDB_RDONLY)) - oflags = O_RDONLY; - else - oflags = O_RDWR | O_CREAT; - - env->me_fd = open(dpath, oflags, mode); -#endif - if (env->me_fd == INVALID_HANDLE_VALUE) { - rc = ErrCode(); - goto leave; - } if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) { - rc = mdb_env_setup_locks(env, lpath, mode, &excl); + rc = mdb_env_setup_locks(env, &fname, mode, &excl); if (rc) goto leave; } if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) { - if (flags & (MDB_RDONLY|MDB_WRITEMAP)) { - env->me_mfd = env->me_fd; - } else { + if (!(flags & (MDB_RDONLY|MDB_WRITEMAP))) { /* Synchronous fd for meta writes. Needed even with * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset. */ -#ifdef _WIN32 - len = OPEN_EXISTING; - rc = utf8_to_utf16(dpath, -1, &wpath, NULL); + rc = mdb_fopen(env, &fname, MDB_O_META, mode, &env->me_mfd); if (rc) goto leave; - env->me_mfd = CreateFileW(wpath, oflags, - FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len, - mode | FILE_FLAG_WRITE_THROUGH, NULL); - free(wpath); -#else - oflags &= ~O_CREAT; - env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode); -#endif - if (env->me_mfd == INVALID_HANDLE_VALUE) { - rc = ErrCode(); - goto leave; - } } DPRINTF(("opened dbenv %p", (void *) env)); if (excl > 0) { @@ -4889,7 +5014,7 @@ leave: if (rc) { mdb_env_close0(env, excl); } - free(lpath); + mdb_fname_destroy(fname); return rc; } @@ -4933,7 +5058,7 @@ mdb_env_close0(MDB_env *env, int excl) if (env->me_map) { munmap(env->me_map, env->me_mapsize); } - if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE) + if (env->me_mfd != INVALID_HANDLE_VALUE) (void) close(env->me_mfd); if (env->me_fd != INVALID_HANDLE_VALUE) (void) close(env->me_fd); @@ -5233,7 +5358,9 @@ mdb_cursor_pop(MDB_cursor *mc) } } -/** Push a page onto the top of the cursor's stack. */ +/** Push a page onto the top of the cursor's stack. + * Set #MDB_TXN_ERROR on failure. + */ static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp) { @@ -5253,6 +5380,7 @@ mdb_cursor_push(MDB_cursor *mc, MDB_page *mp) } /** Find the address of the page corresponding to a given page number. + * Set #MDB_TXN_ERROR on failure. * @param[in] mc the cursor accessing the page. * @param[in] pgno the page number for the page to retrieve. * @param[out] ret address of a pointer where the page's address will be stored. @@ -5337,8 +5465,17 @@ mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags) if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) { i = 0; - if (flags & MDB_PS_LAST) + if (flags & MDB_PS_LAST) { i = NUMKEYS(mp) - 1; + /* if already init'd, see if we're already in right place */ + if (mc->mc_flags & C_INITIALIZED) { + if (mc->mc_ki[mc->mc_top] == i) { + mc->mc_top = mc->mc_snum++; + mp = mc->mc_pg[mc->mc_top]; + goto ready; + } + } + } } else { int exact; node = mdb_node_search(mc, key, &exact); @@ -5364,6 +5501,7 @@ mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags) if ((rc = mdb_cursor_push(mc, mp))) return rc; +ready: if (flags & MDB_PS_MODIFY) { if ((rc = mdb_page_touch(mc)) != 0) return rc; @@ -5689,15 +5827,20 @@ mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op) MDB_node *leaf; int rc; - if ((mc->mc_flags & C_EOF) || - ((mc->mc_flags & C_DEL) && op == MDB_NEXT_DUP)) { + if ((mc->mc_flags & C_DEL && op == MDB_NEXT_DUP)) return MDB_NOTFOUND; - } + if (!(mc->mc_flags & C_INITIALIZED)) return mdb_cursor_first(mc, key, data); mp = mc->mc_pg[mc->mc_top]; + if (mc->mc_flags & C_EOF) { + if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mp)-1) + return MDB_NOTFOUND; + mc->mc_flags ^= C_EOF; + } + if (mc->mc_db->md_flags & MDB_DUPSORT) { leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]); if (F_ISSET(leaf->mn_flags, F_DUPDATA)) { @@ -6102,16 +6245,13 @@ mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data) if (mc->mc_xcursor) mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF); - if (!(mc->mc_flags & C_EOF)) { - - if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) { - rc = mdb_page_search(mc, NULL, MDB_PS_LAST); - if (rc != MDB_SUCCESS) - return rc; - } - mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); - + if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) { + rc = mdb_page_search(mc, NULL, MDB_PS_LAST); + if (rc != MDB_SUCCESS) + return rc; } + mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top])); + mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1; mc->mc_flags |= C_INITIALIZED|C_EOF; leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); @@ -6288,6 +6428,11 @@ fetchm: rc = MDB_INCOMPATIBLE; break; } + if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) { + mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]); + rc = MDB_NOTFOUND; + break; + } { MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { @@ -6784,8 +6929,7 @@ new_sub: if (m3->mc_ki[i] >= mc->mc_ki[i] && insert_key) { m3->mc_ki[i]++; } - if (XCURSOR_INITED(m3)) - XCURSOR_REFRESH(m3, mp, m3->mc_ki[i]); + XCURSOR_REFRESH(m3, i, mp); } } } @@ -6827,7 +6971,6 @@ put_sub: MDB_xcursor *mx = mc->mc_xcursor; unsigned i = mc->mc_top; MDB_page *mp = mc->mc_pg[i]; - int nkeys = NUMKEYS(mp); for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) { if (m2 == mc || m2->mc_snum < mc->mc_snum) continue; @@ -6835,8 +6978,8 @@ put_sub: if (m2->mc_pg[i] == mp) { if (m2->mc_ki[i] == mc->mc_ki[i]) { mdb_xcursor_init2(m2, mx, new_dupdata); - } else if (!insert_key && m2->mc_ki[i] < nkeys) { - XCURSOR_REFRESH(m2, mp, m2->mc_ki[i]); + } else if (!insert_key) { + XCURSOR_REFRESH(m2, i, mp); } } } @@ -6941,12 +7084,7 @@ mdb_cursor_del(MDB_cursor *mc, unsigned int flags) if (m2 == mc || m2->mc_snum < mc->mc_snum) continue; if (!(m2->mc_flags & C_INITIALIZED)) continue; if (m2->mc_pg[mc->mc_top] == mp) { - MDB_node *n2 = leaf; - if (m2->mc_ki[mc->mc_top] != mc->mc_ki[mc->mc_top]) { - n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]); - if (n2->mn_flags & F_SUBDATA) continue; - } - m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2); + XCURSOR_REFRESH(m2, mc->mc_top, mp); } } } @@ -6991,6 +7129,7 @@ fail: } /** Allocate and initialize new pages for a database. + * Set #MDB_TXN_ERROR on failure. * @param[in] mc a cursor on the database being added to. * @param[in] flags flags defining what type of page is being allocated. * @param[in] num the number of pages to allocate. This is usually 1, @@ -7076,6 +7215,7 @@ mdb_branch_size(MDB_env *env, MDB_val *key) } /** Add a node to the page pointed to by the cursor. + * Set #MDB_TXN_ERROR on failure. * @param[in] mc The cursor for this operation. * @param[in] indx The index on the page where the new node should be added. * @param[in] key The key for the new node. @@ -7520,9 +7660,15 @@ mdb_cursor_count(MDB_cursor *mc, size_t *countp) if (!(mc->mc_flags & C_INITIALIZED)) return EINVAL; - if (!mc->mc_snum || (mc->mc_flags & C_EOF)) + if (!mc->mc_snum) return MDB_NOTFOUND; + if (mc->mc_flags & C_EOF) { + if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top])) + return MDB_NOTFOUND; + mc->mc_flags ^= C_EOF; + } + leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { *countp = 1; @@ -7564,6 +7710,7 @@ mdb_cursor_dbi(MDB_cursor *mc) } /** Replace the key for a branch node with a new key. + * Set #MDB_TXN_ERROR on failure. * @param[in] mc Cursor pointing to the node to operate on. * @param[in] key The new key to use. * @return 0 on success, non-zero on failure. @@ -7783,8 +7930,8 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft) m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top]; m3->mc_ki[csrc->mc_top-1]++; } - if (XCURSOR_INITED(m3) && IS_LEAF(mps)) - XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]); + if (IS_LEAF(mps)) + XCURSOR_REFRESH(m3, csrc->mc_top, m3->mc_pg[csrc->mc_top]); } } else /* Adding on the right, bump others down */ @@ -7805,8 +7952,8 @@ mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst, int fromleft) } else { m3->mc_ki[csrc->mc_top]--; } - if (XCURSOR_INITED(m3) && IS_LEAF(mps)) - XCURSOR_REFRESH(m3, m3->mc_pg[csrc->mc_top], m3->mc_ki[csrc->mc_top]); + if (IS_LEAF(mps)) + XCURSOR_REFRESH(m3, csrc->mc_top, m3->mc_pg[csrc->mc_top]); } } } @@ -8007,8 +8154,8 @@ mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst) m3->mc_ki[top-1] > csrc->mc_ki[top-1]) { m3->mc_ki[top-1]--; } - if (XCURSOR_INITED(m3) && IS_LEAF(psrc)) - XCURSOR_REFRESH(m3, m3->mc_pg[top], m3->mc_ki[top]); + if (IS_LEAF(psrc)) + XCURSOR_REFRESH(m3, top, m3->mc_pg[top]); } } { @@ -8270,8 +8417,7 @@ mdb_cursor_del0(MDB_cursor *mc) } else if (m3->mc_ki[mc->mc_top] > ki) { m3->mc_ki[mc->mc_top]--; } - if (XCURSOR_INITED(m3)) - XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]); + XCURSOR_REFRESH(m3, mc->mc_top, mp); } } } @@ -8308,15 +8454,20 @@ mdb_cursor_del0(MDB_cursor *mc) } if (mc->mc_db->md_flags & MDB_DUPSORT) { MDB_node *node = NODEPTR(m3->mc_pg[m3->mc_top], m3->mc_ki[m3->mc_top]); - /* If this node is a fake page, it needs to be reinited - * because its data has moved. But just reset mc_pg[0] - * if the xcursor is already live. + /* If this node has dupdata, it may need to be reinited + * because its data has moved. + * If the xcursor was not initd it must be reinited. + * Else if node points to a subDB, nothing is needed. + * Else (xcursor was initd, not a subDB) needs mc_pg[0] reset. */ - if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA) { - if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) - m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node); - else + if (node->mn_flags & F_DUPDATA) { + if (m3->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) { + if (!(node->mn_flags & F_SUBDATA)) + m3->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(node); + } else { mdb_xcursor_init1(m3, node); + m3->mc_xcursor->mx_cursor.mc_flags |= C_DEL; + } } } } @@ -8392,6 +8543,7 @@ mdb_del0(MDB_txn *txn, MDB_dbi dbi, } /** Split a page and insert a new node. + * Set #MDB_TXN_ERROR on failure. * @param[in,out] mc Cursor pointing to the page and desired insertion index. * The cursor will be updated to point to the actual page and index where * the node got inserted after the split. @@ -8801,8 +8953,8 @@ mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno m3->mc_ki[ptop] >= mc->mc_ki[ptop]) { m3->mc_ki[ptop]++; } - if (XCURSOR_INITED(m3) && IS_LEAF(mp)) - XCURSOR_REFRESH(m3, m3->mc_pg[mc->mc_top], m3->mc_ki[mc->mc_top]); + if (IS_LEAF(mp)) + XCURSOR_REFRESH(m3, mc->mc_top, m3->mc_pg[mc->mc_top]); } } DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp))); @@ -8878,6 +9030,13 @@ mdb_env_copythr(void *arg) #else int len; #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0) +#ifdef SIGPIPE + sigset_t set; + sigemptyset(&set); + sigaddset(&set, SIGPIPE); + if ((rc = pthread_sigmask(SIG_BLOCK, &set, NULL)) != 0) + my->mc_error = rc; +#endif #endif pthread_mutex_lock(&my->mc_mutex); @@ -8894,6 +9053,15 @@ again: DO_WRITE(rc, my->mc_fd, ptr, wsize, len); if (!rc) { rc = ErrCode(); +#if defined(SIGPIPE) && !defined(_WIN32) + if (rc == EPIPE) { + /* Collect the pending SIGPIPE, otherwise at least OS X + * gives it to the process on thread-exit (ITS#8504). + */ + int tmp; + sigwait(&set, &tmp); + } +#endif break; } else if (len > 0) { rc = MDB_SUCCESS; @@ -9356,62 +9524,20 @@ mdb_env_copyfd(MDB_env *env, HANDLE fd) int ESECT mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags) { - int rc, len; - char *lpath; + int rc; + MDB_name fname; HANDLE newfd = INVALID_HANDLE_VALUE; -#ifdef _WIN32 - wchar_t *wpath; -#endif - - if (env->me_flags & MDB_NOSUBDIR) { - lpath = (char *)path; - } else { - len = strlen(path); - len += sizeof(DATANAME); - lpath = malloc(len); - if (!lpath) - return ENOMEM; - sprintf(lpath, "%s" DATANAME, path); - } - /* The destination path must exist, but the destination file must not. - * We don't want the OS to cache the writes, since the source data is - * already in the OS cache. - */ -#ifdef _WIN32 - rc = utf8_to_utf16(lpath, -1, &wpath, NULL); - if (rc) - goto leave; - newfd = CreateFileW(wpath, GENERIC_WRITE, 0, NULL, CREATE_NEW, - FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL); - free(wpath); -#else - newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666); -#endif - if (newfd == INVALID_HANDLE_VALUE) { - rc = ErrCode(); - goto leave; - } - - if (env->me_psize >= env->me_os_psize) { -#ifdef F_NOCACHE /* __APPLE__ */ - (void) fcntl(newfd, F_NOCACHE, 1); -#elif defined O_DIRECT - /* Set O_DIRECT if the file system supports it */ - if ((rc = fcntl(newfd, F_GETFL)) != -1) - (void) fcntl(newfd, F_SETFL, rc | O_DIRECT); -#endif + rc = mdb_fname_init(path, env->me_flags | MDB_NOLOCK, &fname); + if (rc == MDB_SUCCESS) { + rc = mdb_fopen(env, &fname, MDB_O_COPY, 0666, &newfd); + mdb_fname_destroy(fname); } - - rc = mdb_env_copyfd2(env, newfd, flags); - -leave: - if (!(env->me_flags & MDB_NOSUBDIR)) - free(lpath); - if (newfd != INVALID_HANDLE_VALUE) + if (rc == MDB_SUCCESS) { + rc = mdb_env_copyfd2(env, newfd, flags); if (close(newfd) < 0 && rc == MDB_SUCCESS) rc = ErrCode(); - + } return rc; } @@ -9633,8 +9759,11 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]); if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA) return MDB_INCOMPATIBLE; - } else if (! (rc == MDB_NOTFOUND && (flags & MDB_CREATE))) { - return rc; + } else { + if (rc != MDB_NOTFOUND || !(flags & MDB_CREATE)) + return rc; + if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) + return EACCES; } /* Done here so we cannot fail after creating a new DB */ @@ -9648,7 +9777,8 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db memset(&dummy, 0, sizeof(dummy)); dummy.md_root = P_INVALID; dummy.md_flags = flags & PERSISTENT_FLAGS; - rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA); + WITH_CURSOR_TRACKING(mc, + rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA)); dbflag |= DB_DIRTY; } @@ -10001,7 +10131,7 @@ mdb_reader_check(MDB_env *env, int *dead) return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS; } -/** As #mdb_reader_check(). rlocked = . */ +/** As #mdb_reader_check(). \b rlocked is set if caller locked #me_rmutex. */ static int ESECT mdb_reader_check0(MDB_env *env, int rlocked, int *dead) { @@ -10106,27 +10236,28 @@ mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc) #endif /* MDB_ROBUST_SUPPORTED */ #if defined(_WIN32) +/** Convert \b src to new wchar_t[] string with room for \b xtra extra chars */ static int ESECT -utf8_to_utf16(const char *src, int srcsize, wchar_t **dst, int *dstsize) +utf8_to_utf16(const char *src, MDB_name *dst, int xtra) { int rc, need = 0; wchar_t *result = NULL; for (;;) { /* malloc result, then fill it in */ - need = MultiByteToWideChar(CP_UTF8, 0, src, srcsize, result, need); + need = MultiByteToWideChar(CP_UTF8, 0, src, -1, result, need); if (!need) { rc = ErrCode(); free(result); return rc; } if (!result) { - result = malloc(sizeof(wchar_t) * need); + result = malloc(sizeof(wchar_t) * (need + xtra)); if (!result) return ENOMEM; continue; } - if (dstsize) - *dstsize = need; - *dst = result; + dst->mn_alloced = 1; + dst->mn_len = need - 1; + dst->mn_val = result; return MDB_SUCCESS; } }