blob: b36c48583970385c32df6ec2b8e82e89c402e0bd [file] [log] [blame]
/*
URL: svn://svnanon.samba.org/samba/branches/SAMBA_4_0/source/lib/tdb/common
Rev: 23590
Last Changed Date: 2007-06-22 13:36:10 -0400 (Fri, 22 Jun 2007)
*/
/*
trivial database library - standalone version
Copyright (C) Andrew Tridgell 1999-2005
Copyright (C) Jeremy Allison 2000-2006
Copyright (C) Paul `Rusty' Russell 2000
** NOTE! The following LGPL license applies to the tdb
** library. This does NOT imply that all of Samba is released
** under the LGPL
This library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2 of the License, or (at your option) any later version.
This library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with this library; if not, write to the Free Software
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
#ifdef CONFIG_STAND_ALONE
#define HAVE_MMAP
#define HAVE_STRDUP
#define HAVE_SYS_MMAN_H
#define HAVE_UTIME_H
#define HAVE_UTIME
#endif
#ifndef __FreeBSD__
#define _XOPEN_SOURCE 600
#endif
#include "config.h"
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <stdarg.h>
#include <stddef.h>
#include <errno.h>
#include <string.h>
#ifdef HAVE_SYS_SELECT_H
#include <sys/select.h>
#endif
#include <sys/time.h>
#include <sys/types.h>
#include <time.h>
#ifdef HAVE_UTIME_H
#include <utime.h>
#endif
#include <sys/stat.h>
#include <sys/file.h>
#include <fcntl.h>
#ifdef HAVE_SYS_MMAN_H
#include <sys/mman.h>
#endif
#ifndef MAP_FILE
#define MAP_FILE 0
#endif
#ifndef MAP_FAILED
#define MAP_FAILED ((void *)-1)
#endif
#ifndef HAVE_STRDUP
#define strdup rep_strdup
static char *rep_strdup(const char *s)
{
char *ret;
int length;
if (!s)
return NULL;
if (!length)
length = strlen(s);
ret = malloc(length + 1);
if (ret) {
strncpy(ret, s, length);
ret[length] = '\0';
}
return ret;
}
#endif
#ifndef PRINTF_ATTRIBUTE
#if (__GNUC__ >= 3) && (__GNUC_MINOR__ >= 1 )
/** Use gcc attribute to check printf fns. a1 is the 1-based index of
* the parameter containing the format, and a2 the index of the first
* argument. Note that some gcc 2.x versions don't handle this
* properly **/
#define PRINTF_ATTRIBUTE(a1, a2) __attribute__ ((format (__printf__, a1, a2)))
#else
#define PRINTF_ATTRIBUTE(a1, a2)
#endif
#endif
typedef int bool;
#include "tdb.h"
static TDB_DATA tdb_null;
#ifndef u32
#define u32 unsigned
#endif
typedef u32 tdb_len_t;
typedef u32 tdb_off_t;
#ifndef offsetof
#define offsetof(t,f) ((unsigned int)&((t *)0)->f)
#endif
#define TDB_MAGIC_FOOD "TDB file\n"
#define TDB_VERSION (0x26011967 + 6)
#define TDB_MAGIC (0x26011999U)
#define TDB_FREE_MAGIC (~TDB_MAGIC)
#define TDB_DEAD_MAGIC (0xFEE1DEAD)
#define TDB_RECOVERY_MAGIC (0xf53bc0e7U)
#define TDB_ALIGNMENT 4
#define MIN_REC_SIZE (2*sizeof(struct list_struct) + TDB_ALIGNMENT)
#define DEFAULT_HASH_SIZE 131
#define FREELIST_TOP (sizeof(struct tdb_header))
#define TDB_ALIGN(x,a) (((x) + (a)-1) & ~((a)-1))
#define TDB_BYTEREV(x) (((((x)&0xff)<<24)|((x)&0xFF00)<<8)|(((x)>>8)&0xFF00)|((x)>>24))
#define TDB_DEAD(r) ((r)->magic == TDB_DEAD_MAGIC)
#define TDB_BAD_MAGIC(r) ((r)->magic != TDB_MAGIC && !TDB_DEAD(r))
#define TDB_HASH_TOP(hash) (FREELIST_TOP + (BUCKET(hash)+1)*sizeof(tdb_off_t))
#define TDB_HASHTABLE_SIZE(tdb) ((tdb->header.hash_size+1)*sizeof(tdb_off_t))
#define TDB_DATA_START(hash_size) TDB_HASH_TOP(hash_size-1)
#define TDB_RECOVERY_HEAD offsetof(struct tdb_header, recovery_start)
#define TDB_SEQNUM_OFS offsetof(struct tdb_header, sequence_number)
#define TDB_PAD_BYTE 0x42
#define TDB_PAD_U32 0x42424242
/* NB assumes there is a local variable called "tdb" that is the
* current context, also takes doubly-parenthesized print-style
* argument. */
#define TDB_LOG(x) tdb->log.log_fn x
/* lock offsets */
#define GLOBAL_LOCK 0
#define ACTIVE_LOCK 4
#define TRANSACTION_LOCK 8
/* free memory if the pointer is valid and zero the pointer */
#ifndef SAFE_FREE
#define SAFE_FREE(x) do { if ((x) != NULL) {free(x); (x)=NULL;} } while(0)
#endif
#define BUCKET(hash) ((hash) % tdb->header.hash_size)
#define DOCONV() (tdb->flags & TDB_CONVERT)
#define CONVERT(x) (DOCONV() ? tdb_convert(&x, sizeof(x)) : &x)
/* the body of the database is made of one list_struct for the free space
plus a separate data list for each hash value */
struct list_struct {
tdb_off_t next; /* offset of the next record in the list */
tdb_len_t rec_len; /* total byte length of record */
tdb_len_t key_len; /* byte length of key */
tdb_len_t data_len; /* byte length of data */
u32 full_hash; /* the full 32 bit hash of the key */
u32 magic; /* try to catch errors */
/* the following union is implied:
union {
char record[rec_len];
struct {
char key[key_len];
char data[data_len];
}
u32 totalsize; (tailer)
}
*/
};
/* this is stored at the front of every database */
struct tdb_header {
char magic_food[32]; /* for /etc/magic */
u32 version; /* version of the code */
u32 hash_size; /* number of hash entries */
tdb_off_t rwlocks; /* obsolete - kept to detect old formats */
tdb_off_t recovery_start; /* offset of transaction recovery region */
tdb_off_t sequence_number; /* used when TDB_SEQNUM is set */
tdb_off_t reserved[29];
};
struct tdb_lock_type {
int list;
u32 count;
u32 ltype;
};
struct tdb_traverse_lock {
struct tdb_traverse_lock *next;
u32 off;
u32 hash;
int lock_rw;
};
struct tdb_methods {
int (*tdb_read)(struct tdb_context *, tdb_off_t , void *, tdb_len_t , int );
int (*tdb_write)(struct tdb_context *, tdb_off_t, const void *, tdb_len_t);
void (*next_hash_chain)(struct tdb_context *, u32 *);
int (*tdb_oob)(struct tdb_context *, tdb_off_t , int );
int (*tdb_expand_file)(struct tdb_context *, tdb_off_t , tdb_off_t );
int (*tdb_brlock)(struct tdb_context *, tdb_off_t , int, int, int, size_t);
};
struct tdb_context {
char *name; /* the name of the database */
void *map_ptr; /* where it is currently mapped */
int fd; /* open file descriptor for the database */
tdb_len_t map_size; /* how much space has been mapped */
int read_only; /* opened read-only */
int traverse_read; /* read-only traversal */
struct tdb_lock_type global_lock;
int num_lockrecs;
struct tdb_lock_type *lockrecs; /* only real locks, all with count>0 */
enum TDB_ERROR ecode; /* error code for last tdb error */
struct tdb_header header; /* a cached copy of the header */
u32 flags; /* the flags passed to tdb_open */
struct tdb_traverse_lock travlocks; /* current traversal locks */
struct tdb_context *next; /* all tdbs to avoid multiple opens */
dev_t device; /* uniquely identifies this tdb */
ino_t inode; /* uniquely identifies this tdb */
struct tdb_logging_context log;
unsigned int (*hash_fn)(TDB_DATA *key);
int open_flags; /* flags used in the open - needed by reopen */
unsigned int num_locks; /* number of chain locks held */
const struct tdb_methods *methods;
struct tdb_transaction *transaction;
int page_size;
int max_dead_records;
bool have_transaction_lock;
tdb_len_t real_map_size; /* how much space has been mapped */
};
/*
internal prototypes
*/
static int tdb_munmap(struct tdb_context *tdb);
static void tdb_mmap(struct tdb_context *tdb);
static int tdb_lock(struct tdb_context *tdb, int list, int ltype);
static int tdb_unlock(struct tdb_context *tdb, int list, int ltype);
static int tdb_brlock(struct tdb_context *tdb, tdb_off_t offset, int rw_type, int lck_type, int probe, size_t len);
static int tdb_transaction_lock(struct tdb_context *tdb, int ltype);
static int tdb_transaction_unlock(struct tdb_context *tdb);
static int tdb_brlock_upgrade(struct tdb_context *tdb, tdb_off_t offset, size_t len);
static int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off);
static int tdb_write_unlock_record(struct tdb_context *tdb, tdb_off_t off);
static int tdb_ofs_read(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d);
static int tdb_ofs_write(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d);
static void *tdb_convert(void *buf, u32 size);
static int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec);
static tdb_off_t tdb_allocate(struct tdb_context *tdb, tdb_len_t length, struct list_struct *rec);
static int tdb_ofs_read(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d);
static int tdb_ofs_write(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d);
static int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off);
static int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off);
static int tdb_rec_read(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec);
static int tdb_rec_write(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec);
static int tdb_do_delete(struct tdb_context *tdb, tdb_off_t rec_ptr, struct list_struct *rec);
static unsigned char *tdb_alloc_read(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t len);
static int tdb_parse_data(struct tdb_context *tdb, TDB_DATA key,
tdb_off_t offset, tdb_len_t len,
int (*parser)(TDB_DATA key, TDB_DATA data,
void *private_data),
void *private_data);
static tdb_off_t tdb_find_lock_hash(struct tdb_context *tdb, TDB_DATA key, u32 hash, int locktype,
struct list_struct *rec);
static void tdb_io_init(struct tdb_context *tdb);
static int tdb_expand(struct tdb_context *tdb, tdb_off_t size);
static int tdb_rec_free_read(struct tdb_context *tdb, tdb_off_t off,
struct list_struct *rec);
/* file: error.c */
enum TDB_ERROR tdb_error(struct tdb_context *tdb)
{
return tdb->ecode;
}
static struct tdb_errname {
enum TDB_ERROR ecode; const char *estring;
} emap[] = { {TDB_SUCCESS, "Success"},
{TDB_ERR_CORRUPT, "Corrupt database"},
{TDB_ERR_IO, "IO Error"},
{TDB_ERR_LOCK, "Locking error"},
{TDB_ERR_OOM, "Out of memory"},
{TDB_ERR_EXISTS, "Record exists"},
{TDB_ERR_NOLOCK, "Lock exists on other keys"},
{TDB_ERR_EINVAL, "Invalid parameter"},
{TDB_ERR_NOEXIST, "Record does not exist"},
{TDB_ERR_RDONLY, "write not permitted"} };
/* Error string for the last tdb error */
const char *tdb_errorstr(struct tdb_context *tdb)
{
u32 i;
for (i = 0; i < sizeof(emap) / sizeof(struct tdb_errname); i++)
if (tdb->ecode == emap[i].ecode)
return emap[i].estring;
return "Invalid error code";
}
/* file: lock.c */
#define TDB_MARK_LOCK 0x80000000
/* a byte range locking function - return 0 on success
this functions locks/unlocks 1 byte at the specified offset.
On error, errno is also set so that errors are passed back properly
through tdb_open().
note that a len of zero means lock to end of file
*/
int tdb_brlock(struct tdb_context *tdb, tdb_off_t offset,
int rw_type, int lck_type, int probe, size_t len)
{
struct flock fl;
int ret;
if (tdb->flags & TDB_NOLOCK) {
return 0;
}
if ((rw_type == F_WRLCK) && (tdb->read_only || tdb->traverse_read)) {
tdb->ecode = TDB_ERR_RDONLY;
return -1;
}
fl.l_type = rw_type;
fl.l_whence = SEEK_SET;
fl.l_start = offset;
fl.l_len = len;
fl.l_pid = 0;
do {
ret = fcntl(tdb->fd,lck_type,&fl);
} while (ret == -1 && errno == EINTR);
if (ret == -1) {
/* Generic lock error. errno set by fcntl.
* EAGAIN is an expected return from non-blocking
* locks. */
if (!probe && lck_type != F_SETLK) {
/* Ensure error code is set for log fun to examine. */
tdb->ecode = TDB_ERR_LOCK;
TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_brlock failed (fd=%d) at offset %d rw_type=%d lck_type=%d len=%d\n",
tdb->fd, offset, rw_type, lck_type, (int)len));
}
return TDB_ERRCODE(TDB_ERR_LOCK, -1);
}
return 0;
}
/*
upgrade a read lock to a write lock. This needs to be handled in a
special way as some OSes (such as solaris) have too conservative
deadlock detection and claim a deadlock when progress can be
made. For those OSes we may loop for a while.
*/
int tdb_brlock_upgrade(struct tdb_context *tdb, tdb_off_t offset, size_t len)
{
int count = 1000;
while (count--) {
struct timeval tv;
if (tdb_brlock(tdb, offset, F_WRLCK, F_SETLKW, 1, len) == 0) {
return 0;
}
if (errno != EDEADLK) {
break;
}
/* sleep for as short a time as we can - more portable than usleep() */
tv.tv_sec = 0;
tv.tv_usec = 1;
select(0, NULL, NULL, NULL, &tv);
}
TDB_LOG((tdb, TDB_DEBUG_TRACE,"tdb_brlock_upgrade failed at offset %d\n", offset));
return -1;
}
/* lock a list in the database. list -1 is the alloc list */
static int _tdb_lock(struct tdb_context *tdb, int list, int ltype, int op)
{
struct tdb_lock_type *new_lck;
int i;
bool mark_lock = ((ltype & TDB_MARK_LOCK) == TDB_MARK_LOCK);
ltype &= ~TDB_MARK_LOCK;
/* a global lock allows us to avoid per chain locks */
if (tdb->global_lock.count &&
(ltype == tdb->global_lock.ltype || ltype == F_RDLCK)) {
return 0;
}
if (tdb->global_lock.count) {
return TDB_ERRCODE(TDB_ERR_LOCK, -1);
}
if (list < -1 || list >= (int)tdb->header.hash_size) {
TDB_LOG((tdb, TDB_DEBUG_ERROR,"tdb_lock: invalid list %d for ltype=%d\n",
list, ltype));
return -1;
}
if (tdb->flags & TDB_NOLOCK)
return 0;
for (i=0; i<tdb->num_lockrecs; i++) {
if (tdb->lockrecs[i].list == list) {
if (tdb->lockrecs[i].count == 0) {
/*
* Can't happen, see tdb_unlock(). It should
* be an assert.
*/
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lock: "
"lck->count == 0 for list %d", list));
}
/*
* Just increment the in-memory struct, posix locks
* don't stack.
*/
tdb->lockrecs[i].count++;
return 0;
}
}
new_lck = (struct tdb_lock_type *)realloc(
tdb->lockrecs,
sizeof(*tdb->lockrecs) * (tdb->num_lockrecs+1));
if (new_lck == NULL) {
errno = ENOMEM;
return -1;
}
tdb->lockrecs = new_lck;
/* Since fcntl locks don't nest, we do a lock for the first one,
and simply bump the count for future ones */
if (!mark_lock &&
tdb->methods->tdb_brlock(tdb,FREELIST_TOP+4*list, ltype, op,
0, 1)) {
return -1;
}
tdb->num_locks++;
tdb->lockrecs[tdb->num_lockrecs].list = list;
tdb->lockrecs[tdb->num_lockrecs].count = 1;
tdb->lockrecs[tdb->num_lockrecs].ltype = ltype;
tdb->num_lockrecs += 1;
return 0;
}
/* lock a list in the database. list -1 is the alloc list */
int tdb_lock(struct tdb_context *tdb, int list, int ltype)
{
int ret;
ret = _tdb_lock(tdb, list, ltype, F_SETLKW);
if (ret) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lock failed on list %d "
"ltype=%d (%s)\n", list, ltype, strerror(errno)));
}
return ret;
}
/* lock a list in the database. list -1 is the alloc list. non-blocking lock */
int tdb_lock_nonblock(struct tdb_context *tdb, int list, int ltype)
{
return _tdb_lock(tdb, list, ltype, F_SETLK);
}
/* unlock the database: returns void because it's too late for errors. */
/* changed to return int it may be interesting to know there
has been an error --simo */
int tdb_unlock(struct tdb_context *tdb, int list, int ltype)
{
int ret = -1;
int i;
struct tdb_lock_type *lck = NULL;
bool mark_lock = ((ltype & TDB_MARK_LOCK) == TDB_MARK_LOCK);
ltype &= ~TDB_MARK_LOCK;
/* a global lock allows us to avoid per chain locks */
if (tdb->global_lock.count &&
(ltype == tdb->global_lock.ltype || ltype == F_RDLCK)) {
return 0;
}
if (tdb->global_lock.count) {
return TDB_ERRCODE(TDB_ERR_LOCK, -1);
}
if (tdb->flags & TDB_NOLOCK)
return 0;
/* Sanity checks */
if (list < -1 || list >= (int)tdb->header.hash_size) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlock: list %d invalid (%d)\n", list, tdb->header.hash_size));
return ret;
}
for (i=0; i<tdb->num_lockrecs; i++) {
if (tdb->lockrecs[i].list == list) {
lck = &tdb->lockrecs[i];
break;
}
}
if ((lck == NULL) || (lck->count == 0)) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlock: count is 0\n"));
return -1;
}
if (lck->count > 1) {
lck->count--;
return 0;
}
/*
* This lock has count==1 left, so we need to unlock it in the
* kernel. We don't bother with decrementing the in-memory array
* element, we're about to overwrite it with the last array element
* anyway.
*/
if (mark_lock) {
ret = 0;
} else {
ret = tdb->methods->tdb_brlock(tdb, FREELIST_TOP+4*list, F_UNLCK,
F_SETLKW, 0, 1);
}
tdb->num_locks--;
/*
* Shrink the array by overwriting the element just unlocked with the
* last array element.
*/
if (tdb->num_lockrecs > 1) {
*lck = tdb->lockrecs[tdb->num_lockrecs-1];
}
tdb->num_lockrecs -= 1;
/*
* We don't bother with realloc when the array shrinks, but if we have
* a completely idle tdb we should get rid of the locked array.
*/
if (tdb->num_lockrecs == 0) {
SAFE_FREE(tdb->lockrecs);
}
if (ret)
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlock: An error occurred unlocking!\n"));
return ret;
}
/*
get the transaction lock
*/
int tdb_transaction_lock(struct tdb_context *tdb, int ltype)
{
if (tdb->have_transaction_lock || tdb->global_lock.count) {
return 0;
}
if (tdb->methods->tdb_brlock(tdb, TRANSACTION_LOCK, ltype,
F_SETLKW, 0, 1) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_lock: failed to get transaction lock\n"));
tdb->ecode = TDB_ERR_LOCK;
return -1;
}
tdb->have_transaction_lock = 1;
return 0;
}
/*
release the transaction lock
*/
int tdb_transaction_unlock(struct tdb_context *tdb)
{
int ret;
if (!tdb->have_transaction_lock) {
return 0;
}
ret = tdb->methods->tdb_brlock(tdb, TRANSACTION_LOCK, F_UNLCK, F_SETLKW, 0, 1);
if (ret == 0) {
tdb->have_transaction_lock = 0;
}
return ret;
}
/* lock/unlock entire database */
static int _tdb_lockall(struct tdb_context *tdb, int ltype, int op)
{
bool mark_lock = ((ltype & TDB_MARK_LOCK) == TDB_MARK_LOCK);
ltype &= ~TDB_MARK_LOCK;
/* There are no locks on read-only dbs */
if (tdb->read_only || tdb->traverse_read)
return TDB_ERRCODE(TDB_ERR_LOCK, -1);
if (tdb->global_lock.count && tdb->global_lock.ltype == ltype) {
tdb->global_lock.count++;
return 0;
}
if (tdb->global_lock.count) {
/* a global lock of a different type exists */
return TDB_ERRCODE(TDB_ERR_LOCK, -1);
}
if (tdb->num_locks != 0) {
/* can't combine global and chain locks */
return TDB_ERRCODE(TDB_ERR_LOCK, -1);
}
if (!mark_lock &&
tdb->methods->tdb_brlock(tdb, FREELIST_TOP, ltype, op,
0, 4*tdb->header.hash_size)) {
if (op == F_SETLKW) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_lockall failed (%s)\n", strerror(errno)));
}
return -1;
}
tdb->global_lock.count = 1;
tdb->global_lock.ltype = ltype;
return 0;
}
/* unlock entire db */
static int _tdb_unlockall(struct tdb_context *tdb, int ltype)
{
bool mark_lock = ((ltype & TDB_MARK_LOCK) == TDB_MARK_LOCK);
ltype &= ~TDB_MARK_LOCK;
/* There are no locks on read-only dbs */
if (tdb->read_only || tdb->traverse_read) {
return TDB_ERRCODE(TDB_ERR_LOCK, -1);
}
if (tdb->global_lock.ltype != ltype || tdb->global_lock.count == 0) {
return TDB_ERRCODE(TDB_ERR_LOCK, -1);
}
if (tdb->global_lock.count > 1) {
tdb->global_lock.count--;
return 0;
}
if (!mark_lock &&
tdb->methods->tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW,
0, 4*tdb->header.hash_size)) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_unlockall failed (%s)\n", strerror(errno)));
return -1;
}
tdb->global_lock.count = 0;
tdb->global_lock.ltype = 0;
return 0;
}
/* lock entire database with write lock */
int tdb_lockall(struct tdb_context *tdb)
{
return _tdb_lockall(tdb, F_WRLCK, F_SETLKW);
}
/* lock entire database with write lock - mark only */
int tdb_lockall_mark(struct tdb_context *tdb)
{
return _tdb_lockall(tdb, F_WRLCK | TDB_MARK_LOCK, F_SETLKW);
}
/* unlock entire database with write lock - unmark only */
int tdb_lockall_unmark(struct tdb_context *tdb)
{
return _tdb_unlockall(tdb, F_WRLCK | TDB_MARK_LOCK);
}
/* lock entire database with write lock - nonblocking varient */
int tdb_lockall_nonblock(struct tdb_context *tdb)
{
return _tdb_lockall(tdb, F_WRLCK, F_SETLK);
}
/* unlock entire database with write lock */
int tdb_unlockall(struct tdb_context *tdb)
{
return _tdb_unlockall(tdb, F_WRLCK);
}
/* lock entire database with read lock */
int tdb_lockall_read(struct tdb_context *tdb)
{
return _tdb_lockall(tdb, F_RDLCK, F_SETLKW);
}
/* lock entire database with read lock - nonblock varient */
int tdb_lockall_read_nonblock(struct tdb_context *tdb)
{
return _tdb_lockall(tdb, F_RDLCK, F_SETLK);
}
/* unlock entire database with read lock */
int tdb_unlockall_read(struct tdb_context *tdb)
{
return _tdb_unlockall(tdb, F_RDLCK);
}
/* lock/unlock one hash chain. This is meant to be used to reduce
contention - it cannot guarantee how many records will be locked */
int tdb_chainlock(struct tdb_context *tdb, TDB_DATA key)
{
return tdb_lock(tdb, BUCKET(tdb->hash_fn(&key)), F_WRLCK);
}
/* lock/unlock one hash chain, non-blocking. This is meant to be used
to reduce contention - it cannot guarantee how many records will be
locked */
int tdb_chainlock_nonblock(struct tdb_context *tdb, TDB_DATA key)
{
return tdb_lock_nonblock(tdb, BUCKET(tdb->hash_fn(&key)), F_WRLCK);
}
/* mark a chain as locked without actually locking it. Warning! use with great caution! */
int tdb_chainlock_mark(struct tdb_context *tdb, TDB_DATA key)
{
return tdb_lock(tdb, BUCKET(tdb->hash_fn(&key)), F_WRLCK | TDB_MARK_LOCK);
}
/* unmark a chain as locked without actually locking it. Warning! use with great caution! */
int tdb_chainlock_unmark(struct tdb_context *tdb, TDB_DATA key)
{
return tdb_unlock(tdb, BUCKET(tdb->hash_fn(&key)), F_WRLCK | TDB_MARK_LOCK);
}
int tdb_chainunlock(struct tdb_context *tdb, TDB_DATA key)
{
return tdb_unlock(tdb, BUCKET(tdb->hash_fn(&key)), F_WRLCK);
}
int tdb_chainlock_read(struct tdb_context *tdb, TDB_DATA key)
{
return tdb_lock(tdb, BUCKET(tdb->hash_fn(&key)), F_RDLCK);
}
int tdb_chainunlock_read(struct tdb_context *tdb, TDB_DATA key)
{
return tdb_unlock(tdb, BUCKET(tdb->hash_fn(&key)), F_RDLCK);
}
/* record lock stops delete underneath */
int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off)
{
return off ? tdb->methods->tdb_brlock(tdb, off, F_RDLCK, F_SETLKW, 0, 1) : 0;
}
/*
Write locks override our own fcntl readlocks, so check it here.
Note this is meant to be F_SETLK, *not* F_SETLKW, as it's not
an error to fail to get the lock here.
*/
int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off)
{
struct tdb_traverse_lock *i;
for (i = &tdb->travlocks; i; i = i->next)
if (i->off == off)
return -1;
return tdb->methods->tdb_brlock(tdb, off, F_WRLCK, F_SETLK, 1, 1);
}
/*
Note this is meant to be F_SETLK, *not* F_SETLKW, as it's not
an error to fail to get the lock here.
*/
int tdb_write_unlock_record(struct tdb_context *tdb, tdb_off_t off)
{
return tdb->methods->tdb_brlock(tdb, off, F_UNLCK, F_SETLK, 0, 1);
}
/* fcntl locks don't stack: avoid unlocking someone else's */
int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off)
{
struct tdb_traverse_lock *i;
u32 count = 0;
if (off == 0)
return 0;
for (i = &tdb->travlocks; i; i = i->next)
if (i->off == off)
count++;
return (count == 1 ? tdb->methods->tdb_brlock(tdb, off, F_UNLCK, F_SETLKW, 0, 1) : 0);
}
/* file: io.c */
/* check for an out of bounds access - if it is out of bounds then
see if the database has been expanded by someone else and expand
if necessary
note that "len" is the minimum length needed for the db
*/
static int tdb_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
{
struct stat st;
if (len <= tdb->map_size)
return 0;
if (tdb->flags & TDB_INTERNAL) {
if (!probe) {
/* Ensure ecode is set for log fn. */
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_oob len %d beyond internal malloc size %d\n",
(int)len, (int)tdb->map_size));
}
return TDB_ERRCODE(TDB_ERR_IO, -1);
}
if (fstat(tdb->fd, &st) == -1) {
return TDB_ERRCODE(TDB_ERR_IO, -1);
}
if (st.st_size < (size_t)len) {
if (!probe) {
/* Ensure ecode is set for log fn. */
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_oob len %d beyond eof at %d\n",
(int)len, (int)st.st_size));
}
return TDB_ERRCODE(TDB_ERR_IO, -1);
}
/* Unmap, update size, remap */
if (tdb_munmap(tdb) == -1)
return TDB_ERRCODE(TDB_ERR_IO, -1);
tdb->map_size = st.st_size;
tdb_mmap(tdb);
return 0;
}
/* write a lump of data at a specified offset */
static int tdb_write(struct tdb_context *tdb, tdb_off_t off,
const void *buf, tdb_len_t len)
{
if (len == 0) {
return 0;
}
if (tdb->read_only || tdb->traverse_read) {
tdb->ecode = TDB_ERR_RDONLY;
return -1;
}
if (tdb->methods->tdb_oob(tdb, off + len, 0) != 0)
return -1;
if (tdb->map_ptr) {
memcpy(off + (char *)tdb->map_ptr, buf, len);
} else if (pwrite(tdb->fd, buf, len, off) != (ssize_t)len) {
/* Ensure ecode is set for log fn. */
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_write failed at %d len=%d (%s)\n",
off, len, strerror(errno)));
return TDB_ERRCODE(TDB_ERR_IO, -1);
}
return 0;
}
/* Endian conversion: we only ever deal with 4 byte quantities */
void *tdb_convert(void *buf, u32 size)
{
u32 i, *p = (u32 *)buf;
for (i = 0; i < size / 4; i++)
p[i] = TDB_BYTEREV(p[i]);
return buf;
}
/* read a lump of data at a specified offset, maybe convert */
static int tdb_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
tdb_len_t len, int cv)
{
if (tdb->methods->tdb_oob(tdb, off + len, 0) != 0) {
return -1;
}
if (tdb->map_ptr) {
memcpy(buf, off + (char *)tdb->map_ptr, len);
} else {
ssize_t ret = pread(tdb->fd, buf, len, off);
if (ret != (ssize_t)len) {
/* Ensure ecode is set for log fn. */
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_read failed at %d "
"len=%d ret=%d (%s) map_size=%d\n",
(int)off, (int)len, (int)ret, strerror(errno),
(int)tdb->map_size));
return TDB_ERRCODE(TDB_ERR_IO, -1);
}
}
if (cv) {
tdb_convert(buf, len);
}
return 0;
}
/*
do an unlocked scan of the hash table heads to find the next non-zero head. The value
will then be confirmed with the lock held
*/
static void tdb_next_hash_chain(struct tdb_context *tdb, u32 *chain)
{
u32 h = *chain;
if (tdb->map_ptr) {
for (;h < tdb->header.hash_size;h++) {
if (0 != *(u32 *)(TDB_HASH_TOP(h) + (unsigned char *)tdb->map_ptr)) {
break;
}
}
} else {
u32 off=0;
for (;h < tdb->header.hash_size;h++) {
if (tdb_ofs_read(tdb, TDB_HASH_TOP(h), &off) != 0 || off != 0) {
break;
}
}
}
(*chain) = h;
}
int tdb_munmap(struct tdb_context *tdb)
{
if (tdb->flags & TDB_INTERNAL)
return 0;
#ifdef HAVE_MMAP
if (tdb->map_ptr) {
int ret = munmap(tdb->map_ptr, tdb->real_map_size);
if (ret != 0)
return ret;
tdb->real_map_size = 0;
}
#endif
tdb->map_ptr = NULL;
return 0;
}
void tdb_mmap(struct tdb_context *tdb)
{
if (tdb->flags & TDB_INTERNAL)
return;
#ifdef HAVE_MMAP
if (!(tdb->flags & TDB_NOMMAP)) {
tdb->map_ptr = mmap(NULL, tdb->map_size,
PROT_READ|(tdb->read_only? 0:PROT_WRITE),
MAP_SHARED|MAP_FILE, tdb->fd, 0);
/*
* NB. When mmap fails it returns MAP_FAILED *NOT* NULL !!!!
*/
if (tdb->map_ptr == MAP_FAILED) {
tdb->real_map_size = 0;
tdb->map_ptr = NULL;
TDB_LOG((tdb, TDB_DEBUG_WARNING, "tdb_mmap failed for size %d (%s)\n",
tdb->map_size, strerror(errno)));
}
tdb->real_map_size = tdb->map_size;
} else {
tdb->map_ptr = NULL;
}
#else
tdb->map_ptr = NULL;
#endif
}
/* expand a file. we prefer to use ftruncate, as that is what posix
says to use for mmap expansion */
static int tdb_expand_file(struct tdb_context *tdb, tdb_off_t size, tdb_off_t addition)
{
char buf[1024];
if (tdb->read_only || tdb->traverse_read) {
tdb->ecode = TDB_ERR_RDONLY;
return -1;
}
if (ftruncate(tdb->fd, size+addition) == -1) {
char b = 0;
if (pwrite(tdb->fd, &b, 1, (size+addition) - 1) != 1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "expand_file to %d failed (%s)\n",
size+addition, strerror(errno)));
return -1;
}
}
/* now fill the file with something. This ensures that the
file isn't sparse, which would be very bad if we ran out of
disk. This must be done with write, not via mmap */
memset(buf, TDB_PAD_BYTE, sizeof(buf));
while (addition) {
int n = addition>sizeof(buf)?sizeof(buf):addition;
int ret = pwrite(tdb->fd, buf, n, size);
if (ret != n) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "expand_file write of %d failed (%s)\n",
n, strerror(errno)));
return -1;
}
addition -= n;
size += n;
}
return 0;
}
/* expand the database at least size bytes by expanding the underlying
file and doing the mmap again if necessary */
int tdb_expand(struct tdb_context *tdb, tdb_off_t size)
{
struct list_struct rec;
tdb_off_t offset;
if (tdb_lock(tdb, -1, F_WRLCK) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "lock failed in tdb_expand\n"));
return -1;
}
/* must know about any previous expansions by another process */
tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
/* always make room for at least 10 more records, and round
the database up to a multiple of the page size */
size = TDB_ALIGN(tdb->map_size + size*10, tdb->page_size) - tdb->map_size;
if (!(tdb->flags & TDB_INTERNAL))
tdb_munmap(tdb);
/*
* We must ensure the file is unmapped before doing this
* to ensure consistency with systems like OpenBSD where
* writes and mmaps are not consistent.
*/
/* expand the file itself */
if (!(tdb->flags & TDB_INTERNAL)) {
if (tdb->methods->tdb_expand_file(tdb, tdb->map_size, size) != 0)
goto fail;
}
tdb->map_size += size;
if (tdb->flags & TDB_INTERNAL) {
char *new_map_ptr = (char *)realloc(tdb->map_ptr,
tdb->map_size);
if (!new_map_ptr) {
tdb->map_size -= size;
goto fail;
}
tdb->map_ptr = new_map_ptr;
} else {
/*
* We must ensure the file is remapped before adding the space
* to ensure consistency with systems like OpenBSD where
* writes and mmaps are not consistent.
*/
/* We're ok if the mmap fails as we'll fallback to read/write */
tdb_mmap(tdb);
}
/* form a new freelist record */
memset(&rec,'\0',sizeof(rec));
rec.rec_len = size - sizeof(rec);
/* link it into the free list */
offset = tdb->map_size - size;
if (tdb_free(tdb, offset, &rec) == -1)
goto fail;
tdb_unlock(tdb, -1, F_WRLCK);
return 0;
fail:
tdb_unlock(tdb, -1, F_WRLCK);
return -1;
}
/* read/write a tdb_off_t */
int tdb_ofs_read(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d)
{
return tdb->methods->tdb_read(tdb, offset, (char*)d, sizeof(*d), DOCONV());
}
int tdb_ofs_write(struct tdb_context *tdb, tdb_off_t offset, tdb_off_t *d)
{
tdb_off_t off = *d;
return tdb->methods->tdb_write(tdb, offset, CONVERT(off), sizeof(*d));
}
/* read a lump of data, allocating the space for it */
unsigned char *tdb_alloc_read(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t len)
{
unsigned char *buf;
/* some systems don't like zero length malloc */
if (len == 0) {
len = 1;
}
if (!(buf = (unsigned char *)malloc(len))) {
/* Ensure ecode is set for log fn. */
tdb->ecode = TDB_ERR_OOM;
TDB_LOG((tdb, TDB_DEBUG_ERROR,"tdb_alloc_read malloc failed len=%d (%s)\n",
len, strerror(errno)));
return TDB_ERRCODE(TDB_ERR_OOM, buf);
}
if (tdb->methods->tdb_read(tdb, offset, buf, len, 0) == -1) {
SAFE_FREE(buf);
return NULL;
}
return buf;
}
/* Give a piece of tdb data to a parser */
int tdb_parse_data(struct tdb_context *tdb, TDB_DATA key,
tdb_off_t offset, tdb_len_t len,
int (*parser)(TDB_DATA key, TDB_DATA data,
void *private_data),
void *private_data)
{
TDB_DATA data;
int result;
data.dsize = len;
if ((tdb->transaction == NULL) && (tdb->map_ptr != NULL)) {
/*
* Optimize by avoiding the malloc/memcpy/free, point the
* parser directly at the mmap area.
*/
if (tdb->methods->tdb_oob(tdb, offset+len, 0) != 0) {
return -1;
}
data.dptr = offset + (unsigned char *)tdb->map_ptr;
return parser(key, data, private_data);
}
if (!(data.dptr = tdb_alloc_read(tdb, offset, len))) {
return -1;
}
result = parser(key, data, private_data);
free(data.dptr);
return result;
}
/* read/write a record */
int tdb_rec_read(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec)
{
if (tdb->methods->tdb_read(tdb, offset, rec, sizeof(*rec),DOCONV()) == -1)
return -1;
if (TDB_BAD_MAGIC(rec)) {
/* Ensure ecode is set for log fn. */
tdb->ecode = TDB_ERR_CORRUPT;
TDB_LOG((tdb, TDB_DEBUG_FATAL,"tdb_rec_read bad magic 0x%x at offset=%d\n", rec->magic, offset));
return TDB_ERRCODE(TDB_ERR_CORRUPT, -1);
}
return tdb->methods->tdb_oob(tdb, rec->next+sizeof(*rec), 0);
}
int tdb_rec_write(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec)
{
struct list_struct r = *rec;
return tdb->methods->tdb_write(tdb, offset, CONVERT(r), sizeof(r));
}
static const struct tdb_methods io_methods = {
tdb_read,
tdb_write,
tdb_next_hash_chain,
tdb_oob,
tdb_expand_file,
tdb_brlock
};
/*
initialise the default methods table
*/
void tdb_io_init(struct tdb_context *tdb)
{
tdb->methods = &io_methods;
}
/* file: transaction.c */
/*
transaction design:
- only allow a single transaction at a time per database. This makes
using the transaction API simpler, as otherwise the caller would
have to cope with temporary failures in transactions that conflict
with other current transactions
- keep the transaction recovery information in the same file as the
database, using a special 'transaction recovery' record pointed at
by the header. This removes the need for extra journal files as
used by some other databases
- dynamically allocated the transaction recover record, re-using it
for subsequent transactions. If a larger record is needed then
tdb_free() the old record to place it on the normal tdb freelist
before allocating the new record
- during transactions, keep a linked list of writes all that have
been performed by intercepting all tdb_write() calls. The hooked
transaction versions of tdb_read() and tdb_write() check this
linked list and try to use the elements of the list in preference
to the real database.
- don't allow any locks to be held when a transaction starts,
otherwise we can end up with deadlock (plus lack of lock nesting
in posix locks would mean the lock is lost)
- if the caller gains a lock during the transaction but doesn't
release it then fail the commit
- allow for nested calls to tdb_transaction_start(), re-using the
existing transaction record. If the inner transaction is cancelled
then a subsequent commit will fail
- keep a mirrored copy of the tdb hash chain heads to allow for the
fast hash heads scan on traverse, updating the mirrored copy in
the transaction version of tdb_write
- allow callers to mix transaction and non-transaction use of tdb,
although once a transaction is started then an exclusive lock is
gained until the transaction is committed or cancelled
- the commit stategy involves first saving away all modified data
into a linearised buffer in the transaction recovery area, then
marking the transaction recovery area with a magic value to
indicate a valid recovery record. In total 4 fsync/msync calls are
needed per commit to prevent race conditions. It might be possible
to reduce this to 3 or even 2 with some more work.
- check for a valid recovery record on open of the tdb, while the
global lock is held. Automatically recover from the transaction
recovery area if needed, then continue with the open as
usual. This allows for smooth crash recovery with no administrator
intervention.
- if TDB_NOSYNC is passed to flags in tdb_open then transactions are
still available, but no transaction recovery area is used and no
fsync/msync calls are made.
*/
struct tdb_transaction_el {
struct tdb_transaction_el *next, *prev;
tdb_off_t offset;
tdb_len_t length;
unsigned char *data;
};
/*
hold the context of any current transaction
*/
struct tdb_transaction {
/* we keep a mirrored copy of the tdb hash heads here so
tdb_next_hash_chain() can operate efficiently */
u32 *hash_heads;
/* the original io methods - used to do IOs to the real db */
const struct tdb_methods *io_methods;
/* the list of transaction elements. We use a doubly linked
list with a last pointer to allow us to keep the list
ordered, with first element at the front of the list. It
needs to be doubly linked as the read/write traversals need
to be backwards, while the commit needs to be forwards */
struct tdb_transaction_el *elements, *elements_last;
/* non-zero when an internal transaction error has
occurred. All write operations will then fail until the
transaction is ended */
int transaction_error;
/* when inside a transaction we need to keep track of any
nested tdb_transaction_start() calls, as these are allowed,
but don't create a new transaction */
int nesting;
/* old file size before transaction */
tdb_len_t old_map_size;
};
/*
read while in a transaction. We need to check first if the data is in our list
of transaction elements, then if not do a real read
*/
static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
tdb_len_t len, int cv)
{
struct tdb_transaction_el *el;
/* we need to walk the list backwards to get the most recent data */
for (el=tdb->transaction->elements_last;el;el=el->prev) {
tdb_len_t partial;
if (off+len <= el->offset) {
continue;
}
if (off >= el->offset + el->length) {
continue;
}
/* an overlapping read - needs to be split into up to
2 reads and a memcpy */
if (off < el->offset) {
partial = el->offset - off;
if (transaction_read(tdb, off, buf, partial, cv) != 0) {
goto fail;
}
len -= partial;
off += partial;
buf = (void *)(partial + (char *)buf);
}
if (off + len <= el->offset + el->length) {
partial = len;
} else {
partial = el->offset + el->length - off;
}
memcpy(buf, el->data + (off - el->offset), partial);
if (cv) {
tdb_convert(buf, len);
}
len -= partial;
off += partial;
buf = (void *)(partial + (char *)buf);
if (len != 0 && transaction_read(tdb, off, buf, len, cv) != 0) {
goto fail;
}
return 0;
}
/* its not in the transaction elements - do a real read */
return tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv);
fail:
TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
tdb->ecode = TDB_ERR_IO;
tdb->transaction->transaction_error = 1;
return -1;
}
/*
write while in a transaction
*/
static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
const void *buf, tdb_len_t len)
{
struct tdb_transaction_el *el, *best_el=NULL;
if (len == 0) {
return 0;
}
/* if the write is to a hash head, then update the transaction
hash heads */
if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
u32 chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
memcpy(&tdb->transaction->hash_heads[chain], buf, len);
}
/* first see if we can replace an existing entry */
for (el=tdb->transaction->elements_last;el;el=el->prev) {
tdb_len_t partial;
if (best_el == NULL && off == el->offset+el->length) {
best_el = el;
}
if (off+len <= el->offset) {
continue;
}
if (off >= el->offset + el->length) {
continue;
}
/* an overlapping write - needs to be split into up to
2 writes and a memcpy */
if (off < el->offset) {
partial = el->offset - off;
if (transaction_write(tdb, off, buf, partial) != 0) {
goto fail;
}
len -= partial;
off += partial;
buf = (const void *)(partial + (const char *)buf);
}
if (off + len <= el->offset + el->length) {
partial = len;
} else {
partial = el->offset + el->length - off;
}
memcpy(el->data + (off - el->offset), buf, partial);
len -= partial;
off += partial;
buf = (const void *)(partial + (const char *)buf);
if (len != 0 && transaction_write(tdb, off, buf, len) != 0) {
goto fail;
}
return 0;
}
/* see if we can append the new entry to an existing entry */
if (best_el && best_el->offset + best_el->length == off &&
(off+len < tdb->transaction->old_map_size ||
off > tdb->transaction->old_map_size)) {
unsigned char *data = best_el->data;
el = best_el;
el->data = (unsigned char *)realloc(el->data,
el->length + len);
if (el->data == NULL) {
tdb->ecode = TDB_ERR_OOM;
tdb->transaction->transaction_error = 1;
el->data = data;
return -1;
}
if (buf) {
memcpy(el->data + el->length, buf, len);
} else {
memset(el->data + el->length, TDB_PAD_BYTE, len);
}
el->length += len;
return 0;
}
/* add a new entry at the end of the list */
el = (struct tdb_transaction_el *)malloc(sizeof(*el));
if (el == NULL) {
tdb->ecode = TDB_ERR_OOM;
tdb->transaction->transaction_error = 1;
return -1;
}
el->next = NULL;
el->prev = tdb->transaction->elements_last;
el->offset = off;
el->length = len;
el->data = (unsigned char *)malloc(len);
if (el->data == NULL) {
free(el);
tdb->ecode = TDB_ERR_OOM;
tdb->transaction->transaction_error = 1;
return -1;
}
if (buf) {
memcpy(el->data, buf, len);
} else {
memset(el->data, TDB_PAD_BYTE, len);
}
if (el->prev) {
el->prev->next = el;
} else {
tdb->transaction->elements = el;
}
tdb->transaction->elements_last = el;
return 0;
fail:
TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n", off, len));
tdb->ecode = TDB_ERR_IO;
tdb->transaction->transaction_error = 1;
return -1;
}
/*
accelerated hash chain head search, using the cached hash heads
*/
static void transaction_next_hash_chain(struct tdb_context *tdb, u32 *chain)
{
u32 h = *chain;
for (;h < tdb->header.hash_size;h++) {
/* the +1 takes account of the freelist */
if (0 != tdb->transaction->hash_heads[h+1]) {
break;
}
}
(*chain) = h;
}
/*
out of bounds check during a transaction
*/
static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
{
if (len <= tdb->map_size) {
return 0;
}
return TDB_ERRCODE(TDB_ERR_IO, -1);
}
/*
transaction version of tdb_expand().
*/
static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
tdb_off_t addition)
{
/* add a write to the transaction elements, so subsequent
reads see the zero data */
if (transaction_write(tdb, size, NULL, addition) != 0) {
return -1;
}
return 0;
}
/*
brlock during a transaction - ignore them
*/
static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset,
int rw_type, int lck_type, int probe, size_t len)
{
return 0;
}
static const struct tdb_methods transaction_methods = {
transaction_read,
transaction_write,
transaction_next_hash_chain,
transaction_oob,
transaction_expand_file,
transaction_brlock
};
/*
start a tdb transaction. No token is returned, as only a single
transaction is allowed to be pending per tdb_context
*/
int tdb_transaction_start(struct tdb_context *tdb)
{
/* some sanity checks */
if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
tdb->ecode = TDB_ERR_EINVAL;
return -1;
}
/* cope with nested tdb_transaction_start() calls */
if (tdb->transaction != NULL) {
tdb->transaction->nesting++;
TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
tdb->transaction->nesting));
return 0;
}
if (tdb->num_locks != 0 || tdb->global_lock.count) {
/* the caller must not have any locks when starting a
transaction as otherwise we'll be screwed by lack
of nested locks in posix */
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
tdb->ecode = TDB_ERR_LOCK;
return -1;
}
if (tdb->travlocks.next != NULL) {
/* you cannot use transactions inside a traverse (although you can use
traverse inside a transaction) as otherwise you can end up with
deadlock */
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
tdb->ecode = TDB_ERR_LOCK;
return -1;
}
tdb->transaction = (struct tdb_transaction *)
calloc(sizeof(struct tdb_transaction), 1);
if (tdb->transaction == NULL) {
tdb->ecode = TDB_ERR_OOM;
return -1;
}
/* get the transaction write lock. This is a blocking lock. As
discussed with Volker, there are a number of ways we could
make this async, which we will probably do in the future */
if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
SAFE_FREE(tdb->transaction);
return -1;
}
/* get a read lock from the freelist to the end of file. This
is upgraded to a write lock during the commit */
if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
tdb->ecode = TDB_ERR_LOCK;
goto fail;
}
/* setup a copy of the hash table heads so the hash scan in
traverse can be fast */
tdb->transaction->hash_heads = (u32 *)
calloc(tdb->header.hash_size+1, sizeof(u32));
if (tdb->transaction->hash_heads == NULL) {
tdb->ecode = TDB_ERR_OOM;
goto fail;
}
if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
tdb->ecode = TDB_ERR_IO;
goto fail;
}
/* make sure we know about any file expansions already done by
anyone else */
tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
tdb->transaction->old_map_size = tdb->map_size;
/* finally hook the io methods, replacing them with
transaction specific methods */
tdb->transaction->io_methods = tdb->methods;
tdb->methods = &transaction_methods;
/* by calling this transaction write here, we ensure that we don't grow the
transaction linked list due to hash table updates */
if (transaction_write(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
TDB_HASHTABLE_SIZE(tdb)) != 0) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to prime hash table\n"));
tdb->ecode = TDB_ERR_IO;
tdb->methods = tdb->transaction->io_methods;
goto fail;
}
return 0;
fail:
tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
tdb_transaction_unlock(tdb);
SAFE_FREE(tdb->transaction->hash_heads);
SAFE_FREE(tdb->transaction);
return -1;
}
/*
cancel the current transaction
*/
int tdb_transaction_cancel(struct tdb_context *tdb)
{
if (tdb->transaction == NULL) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
return -1;
}
if (tdb->transaction->nesting != 0) {
tdb->transaction->transaction_error = 1;
tdb->transaction->nesting--;
return 0;
}
tdb->map_size = tdb->transaction->old_map_size;
/* free all the transaction elements */
while (tdb->transaction->elements) {
struct tdb_transaction_el *el = tdb->transaction->elements;
tdb->transaction->elements = el->next;
free(el->data);
free(el);
}
/* remove any global lock created during the transaction */
if (tdb->global_lock.count != 0) {
tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
tdb->global_lock.count = 0;
}
/* remove any locks created during the transaction */
if (tdb->num_locks != 0) {
int i;
for (i=0;i<tdb->num_lockrecs;i++) {
tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
F_UNLCK,F_SETLKW, 0, 1);
}
tdb->num_locks = 0;
tdb->num_lockrecs = 0;
SAFE_FREE(tdb->lockrecs);
}
/* restore the normal io methods */
tdb->methods = tdb->transaction->io_methods;
tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
tdb_transaction_unlock(tdb);
SAFE_FREE(tdb->transaction->hash_heads);
SAFE_FREE(tdb->transaction);
return 0;
}
/*
sync to disk
*/
static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
{
if (fsync(tdb->fd) != 0) {
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
return -1;
}
#if defined(HAVE_MSYNC) && defined(MS_SYNC)
if (tdb->map_ptr) {
tdb_off_t moffset = offset & ~(tdb->page_size-1);
if (msync(moffset + (char *)tdb->map_ptr,
length + (offset - moffset), MS_SYNC) != 0) {
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
strerror(errno)));
return -1;
}
}
#endif
return 0;
}
/*
work out how much space the linearised recovery data will consume
*/
static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
{
struct tdb_transaction_el *el;
tdb_len_t recovery_size = 0;
recovery_size = sizeof(u32);
for (el=tdb->transaction->elements;el;el=el->next) {
if (el->offset >= tdb->transaction->old_map_size) {
continue;
}
recovery_size += 2*sizeof(tdb_off_t) + el->length;
}
return recovery_size;
}
/*
allocate the recovery area, or use an existing recovery area if it is
large enough
*/
static int tdb_recovery_allocate(struct tdb_context *tdb,
tdb_len_t *recovery_size,
tdb_off_t *recovery_offset,
tdb_len_t *recovery_max_size)
{
struct list_struct rec;
const struct tdb_methods *methods = tdb->transaction->io_methods;
tdb_off_t recovery_head;
if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
return -1;
}
rec.rec_len = 0;
if (recovery_head != 0 &&
methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
return -1;
}
*recovery_size = tdb_recovery_size(tdb);
if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
/* it fits in the existing area */
*recovery_max_size = rec.rec_len;
*recovery_offset = recovery_head;
return 0;
}
/* we need to free up the old recovery area, then allocate a
new one at the end of the file. Note that we cannot use
tdb_allocate() to allocate the new one as that might return
us an area that is being currently used (as of the start of
the transaction) */
if (recovery_head != 0) {
if (tdb_free(tdb, recovery_head, &rec) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
return -1;
}
}
/* the tdb_free() call might have increased the recovery size */
*recovery_size = tdb_recovery_size(tdb);
/* round up to a multiple of page size */
*recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
*recovery_offset = tdb->map_size;
recovery_head = *recovery_offset;
if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
(tdb->map_size - tdb->transaction->old_map_size) +
sizeof(rec) + *recovery_max_size) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
return -1;
}
/* remap the file (if using mmap) */
methods->tdb_oob(tdb, tdb->map_size + 1, 1);
/* we have to reset the old map size so that we don't try to expand the file
again in the transaction commit, which would destroy the recovery area */
tdb->transaction->old_map_size = tdb->map_size;
/* write the recovery header offset and sync - we can sync without a race here
as the magic ptr in the recovery record has not been set */
CONVERT(recovery_head);
if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
&recovery_head, sizeof(tdb_off_t)) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
return -1;
}
return 0;
}
/*
setup the recovery data that will be used on a crash during commit
*/
static int transaction_setup_recovery(struct tdb_context *tdb,
tdb_off_t *magic_offset)
{
struct tdb_transaction_el *el;
tdb_len_t recovery_size;
unsigned char *data, *p;
const struct tdb_methods *methods = tdb->transaction->io_methods;
struct list_struct *rec;
tdb_off_t recovery_offset, recovery_max_size;
tdb_off_t old_map_size = tdb->transaction->old_map_size;
u32 magic, tailer;
/*
check that the recovery area has enough space
*/
if (tdb_recovery_allocate(tdb, &recovery_size,
&recovery_offset, &recovery_max_size) == -1) {
return -1;
}
data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
if (data == NULL) {
tdb->ecode = TDB_ERR_OOM;
return -1;
}
rec = (struct list_struct *)data;
memset(rec, 0, sizeof(*rec));
rec->magic = 0;
rec->data_len = recovery_size;
rec->rec_len = recovery_max_size;
rec->key_len = old_map_size;
CONVERT(rec);
/* build the recovery data into a single blob to allow us to do a single
large write, which should be more efficient */
p = data + sizeof(*rec);
for (el=tdb->transaction->elements;el;el=el->next) {
if (el->offset >= old_map_size) {
continue;
}
if (el->offset + el->length > tdb->transaction->old_map_size) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
free(data);
tdb->ecode = TDB_ERR_CORRUPT;
return -1;
}
memcpy(p, &el->offset, 4);
memcpy(p+4, &el->length, 4);
if (DOCONV()) {
tdb_convert(p, 8);
}
/* the recovery area contains the old data, not the
new data, so we have to call the original tdb_read
method to get it */
if (methods->tdb_read(tdb, el->offset, p + 8, el->length, 0) != 0) {
free(data);
tdb->ecode = TDB_ERR_IO;
return -1;
}
p += 8 + el->length;
}
/* and the tailer */
tailer = sizeof(*rec) + recovery_max_size;
memcpy(p, &tailer, 4);
CONVERT(p);
/* write the recovery data to the recovery area */
if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
free(data);
tdb->ecode = TDB_ERR_IO;
return -1;
}
/* as we don't have ordered writes, we have to sync the recovery
data before we update the magic to indicate that the recovery
data is present */
if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
free(data);
return -1;
}
free(data);
magic = TDB_RECOVERY_MAGIC;
CONVERT(magic);
*magic_offset = recovery_offset + offsetof(struct list_struct, magic);
if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
/* ensure the recovery magic marker is on disk */
if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
return -1;
}
return 0;
}
/*
commit the current transaction
*/
int tdb_transaction_commit(struct tdb_context *tdb)
{
const struct tdb_methods *methods;
tdb_off_t magic_offset = 0;
u32 zero = 0;
if (tdb->transaction == NULL) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
return -1;
}
if (tdb->transaction->transaction_error) {
tdb->ecode = TDB_ERR_IO;
tdb_transaction_cancel(tdb);
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
return -1;
}
if (tdb->transaction->nesting != 0) {
tdb->transaction->nesting--;
return 0;
}
/* check for a null transaction */
if (tdb->transaction->elements == NULL) {
tdb_transaction_cancel(tdb);
return 0;
}
methods = tdb->transaction->io_methods;
/* if there are any locks pending then the caller has not
nested their locks properly, so fail the transaction */
if (tdb->num_locks || tdb->global_lock.count) {
tdb->ecode = TDB_ERR_LOCK;
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: locks pending on commit\n"));
tdb_transaction_cancel(tdb);
return -1;
}
/* upgrade the main transaction lock region to a write lock */
if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to upgrade hash locks\n"));
tdb->ecode = TDB_ERR_LOCK;
tdb_transaction_cancel(tdb);
return -1;
}
/* get the global lock - this prevents new users attaching to the database
during the commit */
if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: failed to get global lock\n"));
tdb->ecode = TDB_ERR_LOCK;
tdb_transaction_cancel(tdb);
return -1;
}
if (!(tdb->flags & TDB_NOSYNC)) {
/* write the recovery data to the end of the file */
if (transaction_setup_recovery(tdb, &magic_offset) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: failed to setup recovery data\n"));
tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
tdb_transaction_cancel(tdb);
return -1;
}
}
/* expand the file to the new size if needed */
if (tdb->map_size != tdb->transaction->old_map_size) {
if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
tdb->map_size -
tdb->transaction->old_map_size) == -1) {
tdb->ecode = TDB_ERR_IO;
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: expansion failed\n"));
tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
tdb_transaction_cancel(tdb);
return -1;
}
tdb->map_size = tdb->transaction->old_map_size;
methods->tdb_oob(tdb, tdb->map_size + 1, 1);
}
/* perform all the writes */
while (tdb->transaction->elements) {
struct tdb_transaction_el *el = tdb->transaction->elements;
if (methods->tdb_write(tdb, el->offset, el->data, el->length) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
/* we've overwritten part of the data and
possibly expanded the file, so we need to
run the crash recovery code */
tdb->methods = methods;
tdb_transaction_recover(tdb);
tdb_transaction_cancel(tdb);
tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
return -1;
}
tdb->transaction->elements = el->next;
free(el->data);
free(el);
}
if (!(tdb->flags & TDB_NOSYNC)) {
/* ensure the new data is on disk */
if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
return -1;
}
/* remove the recovery marker */
if (methods->tdb_write(tdb, magic_offset, &zero, 4) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: failed to remove recovery magic\n"));
return -1;
}
/* ensure the recovery marker has been removed on disk */
if (transaction_sync(tdb, magic_offset, 4) == -1) {
return -1;
}
}
tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
/*
TODO: maybe write to some dummy hdr field, or write to magic
offset without mmap, before the last sync, instead of the
utime() call
*/
/* on some systems (like Linux 2.6.x) changes via mmap/msync
don't change the mtime of the file, this means the file may
not be backed up (as tdb rounding to block sizes means that
file size changes are quite rare too). The following forces
mtime changes when a transaction completes */
#ifdef HAVE_UTIME
utime(tdb->name, NULL);
#endif
/* use a transaction cancel to free memory and remove the
transaction locks */
tdb_transaction_cancel(tdb);
return 0;
}
/*
recover from an aborted transaction. Must be called with exclusive
database write access already established (including the global
lock to prevent new processes attaching)
*/
int tdb_transaction_recover(struct tdb_context *tdb)
{
tdb_off_t recovery_head, recovery_eof;
unsigned char *data, *p;
u32 zero = 0;
struct list_struct rec;
/* find the recovery area */
if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
if (recovery_head == 0) {
/* we have never allocated a recovery record */
return 0;
}
/* read the recovery record */
if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
sizeof(rec), DOCONV()) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
if (rec.magic != TDB_RECOVERY_MAGIC) {
/* there is no valid recovery data */
return 0;
}
if (tdb->read_only) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
tdb->ecode = TDB_ERR_CORRUPT;
return -1;
}
recovery_eof = rec.key_len;
data = (unsigned char *)malloc(rec.data_len);
if (data == NULL) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
tdb->ecode = TDB_ERR_OOM;
return -1;
}
/* read the full recovery data */
if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
rec.data_len, 0) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
/* recover the file data */
p = data;
while (p+8 < data + rec.data_len) {
u32 ofs, len;
if (DOCONV()) {
tdb_convert(p, 8);
}
memcpy(&ofs, p, 4);
memcpy(&len, p+4, 4);
if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
free(data);
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
tdb->ecode = TDB_ERR_IO;
return -1;
}
p += 8 + len;
}
free(data);
if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
/* if the recovery area is after the recovered eof then remove it */
if (recovery_eof <= recovery_head) {
if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
}
/* remove the recovery magic */
if (tdb_ofs_write(tdb, recovery_head + offsetof(struct list_struct, magic),
&zero) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
/* reduce the file size to the old size */
tdb_munmap(tdb);
if (ftruncate(tdb->fd, recovery_eof) != 0) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
tdb->map_size = recovery_eof;
tdb_mmap(tdb);
if (transaction_sync(tdb, 0, recovery_eof) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
tdb->ecode = TDB_ERR_IO;
return -1;
}
TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",
recovery_eof));
/* all done */
return 0;
}
/* file: freelist.c */
/* read a freelist record and check for simple errors */
static int tdb_rec_free_read(struct tdb_context *tdb, tdb_off_t off, struct list_struct *rec)
{
if (tdb->methods->tdb_read(tdb, off, rec, sizeof(*rec),DOCONV()) == -1)
return -1;
if (rec->magic == TDB_MAGIC) {
/* this happens when a app is showdown while deleting a record - we should
not completely fail when this happens */
TDB_LOG((tdb, TDB_DEBUG_WARNING, "tdb_rec_free_read non-free magic 0x%x at offset=%d - fixing\n",
rec->magic, off));
rec->magic = TDB_FREE_MAGIC;
if (tdb->methods->tdb_write(tdb, off, rec, sizeof(*rec)) == -1)
return -1;
}
if (rec->magic != TDB_FREE_MAGIC) {
/* Ensure ecode is set for log fn. */
tdb->ecode = TDB_ERR_CORRUPT;
TDB_LOG((tdb, TDB_DEBUG_WARNING, "tdb_rec_free_read bad magic 0x%x at offset=%d\n",
rec->magic, off));
return TDB_ERRCODE(TDB_ERR_CORRUPT, -1);
}
if (tdb->methods->tdb_oob(tdb, rec->next+sizeof(*rec), 0) != 0)
return -1;
return 0;
}
/* Remove an element from the freelist. Must have alloc lock. */
static int remove_from_freelist(struct tdb_context *tdb, tdb_off_t off, tdb_off_t next)
{
tdb_off_t last_ptr, i;
/* read in the freelist top */
last_ptr = FREELIST_TOP;
while (tdb_ofs_read(tdb, last_ptr, &i) != -1 && i != 0) {
if (i == off) {
/* We've found it! */
return tdb_ofs_write(tdb, last_ptr, &next);
}
/* Follow chain (next offset is at start of record) */
last_ptr = i;
}
TDB_LOG((tdb, TDB_DEBUG_FATAL,"remove_from_freelist: not on list at off=%d\n", off));
return TDB_ERRCODE(TDB_ERR_CORRUPT, -1);
}
/* update a record tailer (must hold allocation lock) */
static int update_tailer(struct tdb_context *tdb, tdb_off_t offset,
const struct list_struct *rec)
{
tdb_off_t totalsize;
/* Offset of tailer from record header */
totalsize = sizeof(*rec) + rec->rec_len;
return tdb_ofs_write(tdb, offset + totalsize - sizeof(tdb_off_t),
&totalsize);
}
/* Add an element into the freelist. Merge adjacent records if
neccessary. */
int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec)
{
tdb_off_t right, left;
/* Allocation and tailer lock */
if (tdb_lock(tdb, -1, F_WRLCK) != 0)
return -1;
/* set an initial tailer, so if we fail we don't leave a bogus record */
if (update_tailer(tdb, offset, rec) != 0) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed!\n"));
goto fail;
}
/* Look right first (I'm an Australian, dammit) */
right = offset + sizeof(*rec) + rec->rec_len;
if (right + sizeof(*rec) <= tdb->map_size) {
struct list_struct r;
if (tdb->methods->tdb_read(tdb, right, &r, sizeof(r), DOCONV()) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: right read failed at %u\n", right));
goto left;
}
/* If it's free, expand to include it. */
if (r.magic == TDB_FREE_MAGIC) {
if (remove_from_freelist(tdb, right, r.next) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: right free failed at %u\n", right));
goto left;
}
rec->rec_len += sizeof(r) + r.rec_len;
}
}
left:
/* Look left */
left = offset - sizeof(tdb_off_t);
if (left > TDB_DATA_START(tdb->header.hash_size)) {
struct list_struct l;
tdb_off_t leftsize;
/* Read in tailer and jump back to header */
if (tdb_ofs_read(tdb, left, &leftsize) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left offset read failed at %u\n", left));
goto update;
}
/* it could be uninitialised data */
if (leftsize == 0 || leftsize == TDB_PAD_U32) {
goto update;
}
left = offset - leftsize;
/* Now read in record */
if (tdb->methods->tdb_read(tdb, left, &l, sizeof(l), DOCONV()) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left read failed at %u (%u)\n", left, leftsize));
goto update;
}
/* If it's free, expand to include it. */
if (l.magic == TDB_FREE_MAGIC) {
if (remove_from_freelist(tdb, left, l.next) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left free failed at %u\n", left));
goto update;
} else {
offset = left;
rec->rec_len += leftsize;
}
}
}
update:
if (update_tailer(tdb, offset, rec) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
goto fail;
}
/* Now, prepend to free list */
rec->magic = TDB_FREE_MAGIC;
if (tdb_ofs_read(tdb, FREELIST_TOP, &rec->next) == -1 ||
tdb_rec_write(tdb, offset, rec) == -1 ||
tdb_ofs_write(tdb, FREELIST_TOP, &offset) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free record write failed at offset=%d\n", offset));
goto fail;
}
/* And we're done. */
tdb_unlock(tdb, -1, F_WRLCK);
return 0;
fail:
tdb_unlock(tdb, -1, F_WRLCK);
return -1;
}
/*
the core of tdb_allocate - called when we have decided which
free list entry to use
*/
static tdb_off_t tdb_allocate_ofs(struct tdb_context *tdb, tdb_len_t length, tdb_off_t rec_ptr,
struct list_struct *rec, tdb_off_t last_ptr)
{
struct list_struct newrec;
tdb_off_t newrec_ptr;
memset(&newrec, '\0', sizeof(newrec));
/* found it - now possibly split it up */
if (rec->rec_len > length + MIN_REC_SIZE) {
/* Length of left piece */
length = TDB_ALIGN(length, TDB_ALIGNMENT);
/* Right piece to go on free list */
newrec.rec_len = rec->rec_len - (sizeof(*rec) + length);
newrec_ptr = rec_ptr + sizeof(*rec) + length;
/* And left record is shortened */
rec->rec_len = length;
} else {
newrec_ptr = 0;
}
/* Remove allocated record from the free list */
if (tdb_ofs_write(tdb, last_ptr, &rec->next) == -1) {
return 0;
}
/* Update header: do this before we drop alloc
lock, otherwise tdb_free() might try to
merge with us, thinking we're free.
(Thanks Jeremy Allison). */
rec->magic = TDB_MAGIC;
if (tdb_rec_write(tdb, rec_ptr, rec) == -1) {
return 0;
}
/* Did we create new block? */
if (newrec_ptr) {
/* Update allocated record tailer (we
shortened it). */
if (update_tailer(tdb, rec_ptr, rec) == -1) {
return 0;
}
/* Free new record */
if (tdb_free(tdb, newrec_ptr, &newrec) == -1) {
return 0;
}
}
/* all done - return the new record offset */
return rec_ptr;
}
/* allocate some space from the free list. The offset returned points
to a unconnected list_struct within the database with room for at
least length bytes of total data
0 is returned if the space could not be allocated
*/
tdb_off_t tdb_allocate(struct tdb_context *tdb, tdb_len_t length, struct list_struct *rec)
{
tdb_off_t rec_ptr, last_ptr, newrec_ptr;
struct {
tdb_off_t rec_ptr, last_ptr;
tdb_len_t rec_len;
} bestfit;
if (tdb_lock(tdb, -1, F_WRLCK) == -1)
return 0;
/* Extra bytes required for tailer */
length += sizeof(tdb_off_t);
again:
last_ptr = FREELIST_TOP;
/* read in the freelist top */
if (tdb_ofs_read(tdb, FREELIST_TOP, &rec_ptr) == -1)
goto fail;
bestfit.rec_ptr = 0;
bestfit.last_ptr = 0;
bestfit.rec_len = 0;
/*
this is a best fit allocation strategy. Originally we used
a first fit strategy, but it suffered from massive fragmentation
issues when faced with a slowly increasing record size.
*/
while (rec_ptr) {
if (tdb_rec_free_read(tdb, rec_ptr, rec) == -1) {
goto fail;
}
if (rec->rec_len >= length) {
if (bestfit.rec_ptr == 0 ||
rec->rec_len < bestfit.rec_len) {
bestfit.rec_len = rec->rec_len;
bestfit.rec_ptr = rec_ptr;
bestfit.last_ptr = last_ptr;
/* consider a fit to be good enough if
we aren't wasting more than half
the space */
if (bestfit.rec_len < 2*length) {
break;
}
}
}
/* move to the next record */
last_ptr = rec_ptr;
rec_ptr = rec->next;
}
if (bestfit.rec_ptr != 0) {
if (tdb_rec_free_read(tdb, bestfit.rec_ptr, rec) == -1) {
goto fail;
}
newrec_ptr = tdb_allocate_ofs(tdb, length, bestfit.rec_ptr, rec, bestfit.last_ptr);
tdb_unlock(tdb, -1, F_WRLCK);
return newrec_ptr;
}
/* we didn't find enough space. See if we can expand the
database and if we can then try again */
if (tdb_expand(tdb, length + sizeof(*rec)) == 0)
goto again;
fail:
tdb_unlock(tdb, -1, F_WRLCK);
return 0;
}
/* file: freelistcheck.c */
/* Check the freelist is good and contains no loops.
Very memory intensive - only do this as a consistency
checker. Heh heh - uses an in memory tdb as the storage
for the "seen" record list. For some reason this strikes
me as extremely clever as I don't have to write another tree
data structure implementation :-).
*/
static int seen_insert(struct tdb_context *mem_tdb, tdb_off_t rec_ptr)
{
TDB_DATA key, data;
memset(&data, '\0', sizeof(data));
key.dptr = (unsigned char *)&rec_ptr;
key.dsize = sizeof(rec_ptr);
return tdb_store(mem_tdb, key, data, TDB_INSERT);
}
int tdb_validate_freelist(struct tdb_context *tdb, int *pnum_entries)
{
struct tdb_context *mem_tdb = NULL;
struct list_struct rec;
tdb_off_t rec_ptr, last_ptr;
int ret = -1;
*pnum_entries = 0;
mem_tdb = tdb_open("flval", tdb->header.hash_size,
TDB_INTERNAL, O_RDWR, 0600);
if (!mem_tdb) {
return -1;
}
if (tdb_lock(tdb, -1, F_WRLCK) == -1) {
tdb_close(mem_tdb);
return 0;
}
last_ptr = FREELIST_TOP;
/* Store the FREELIST_TOP record. */
if (seen_insert(mem_tdb, last_ptr) == -1) {
ret = TDB_ERRCODE(TDB_ERR_CORRUPT, -1);
goto fail;
}
/* read in the freelist top */
if (tdb_ofs_read(tdb, FREELIST_TOP, &rec_ptr) == -1) {
goto fail;
}
while (rec_ptr) {
/* If we can't store this record (we've seen it
before) then the free list has a loop and must
be corrupt. */
if (seen_insert(mem_tdb, rec_ptr)) {
ret = TDB_ERRCODE(TDB_ERR_CORRUPT, -1);
goto fail;
}
if (tdb_rec_free_read(tdb, rec_ptr, &rec) == -1) {
goto fail;
}
/* move to the next record */
last_ptr = rec_ptr;
rec_ptr = rec.next;
*pnum_entries += 1;
}
ret = 0;
fail:
tdb_close(mem_tdb);
tdb_unlock(tdb, -1, F_WRLCK);
return ret;
}
/* file: traverse.c */
/* Uses traverse lock: 0 = finish, -1 = error, other = record offset */
static int tdb_next_lock(struct tdb_context *tdb, struct tdb_traverse_lock *tlock,
struct list_struct *rec)
{
int want_next = (tlock->off != 0);
/* Lock each chain from the start one. */
for (; tlock->hash < tdb->header.hash_size; tlock->hash++) {
if (!tlock->off && tlock->hash != 0) {
/* this is an optimisation for the common case where
the hash chain is empty, which is particularly
common for the use of tdb with ldb, where large
hashes are used. In that case we spend most of our
time in tdb_brlock(), locking empty hash chains.
To avoid this, we do an unlocked pre-check to see
if the hash chain is empty before starting to look
inside it. If it is empty then we can avoid that
hash chain. If it isn't empty then we can't believe
the value we get back, as we read it without a
lock, so instead we get the lock and re-fetch the
value below.
Notice that not doing this optimisation on the
first hash chain is critical. We must guarantee
that we have done at least one fcntl lock at the
start of a search to guarantee that memory is
coherent on SMP systems. If records are added by
others during the search then thats OK, and we
could possibly miss those with this trick, but we
could miss them anyway without this trick, so the
semantics don't change.