#include "private.h"
+#include <ccan/hash/hash.h>
+#include <assert.h>
-/* all contexts, to ensure no double-opens (fcntl locks don't nest!) */
-static struct tdb_context *tdbs = NULL;
+/* all lock info, to detect double-opens (fcntl file don't nest!) */
+static struct tdb_file *files = NULL;
-static bool tdb_already_open(dev_t device, ino_t ino)
+static struct tdb_file *find_file(dev_t device, ino_t ino)
{
- struct tdb_context *i;
+ struct tdb_file *i;
- for (i = tdbs; i; i = i->next) {
+ for (i = files; i; i = i->next) {
if (i->device == device && i->inode == ino) {
- return true;
+ i->refcnt++;
+ break;
}
}
-
- return false;
+ return i;
}
static bool read_all(int fd, void *buf, size_t len)
else
newdb.hdr.hash_seed = random_number(tdb);
newdb.hdr.hash_test = TDB_HASH_MAGIC;
- newdb.hdr.hash_test = tdb->khash(&newdb.hdr.hash_test,
- sizeof(newdb.hdr.hash_test),
- newdb.hdr.hash_seed,
- tdb->hash_priv);
+ newdb.hdr.hash_test = tdb->hash_fn(&newdb.hdr.hash_test,
+ sizeof(newdb.hdr.hash_test),
+ newdb.hdr.hash_seed,
+ tdb->hash_data);
newdb.hdr.recovery = 0;
newdb.hdr.features_used = newdb.hdr.features_offered = TDB_FEATURE_MASK;
+ newdb.hdr.seqnum = 0;
memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
/* Initial hashes are empty. */
memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
*hdr = newdb.hdr;
if (tdb->flags & TDB_INTERNAL) {
- tdb->map_size = sizeof(newdb);
- tdb->map_ptr = malloc(tdb->map_size);
- if (!tdb->map_ptr) {
+ tdb->file->map_size = sizeof(newdb);
+ tdb->file->map_ptr = malloc(tdb->file->map_size);
+ if (!tdb->file->map_ptr) {
return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
"tdb_new_database:"
" failed to allocate");
}
- memcpy(tdb->map_ptr, &newdb, tdb->map_size);
+ memcpy(tdb->file->map_ptr, &newdb, tdb->file->map_size);
return TDB_SUCCESS;
}
- if (lseek(tdb->fd, 0, SEEK_SET) == -1) {
+ if (lseek(tdb->file->fd, 0, SEEK_SET) == -1) {
return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
"tdb_new_database:"
" failed to seek: %s", strerror(errno));
}
- if (ftruncate(tdb->fd, 0) == -1) {
+ if (ftruncate(tdb->file->fd, 0) == -1) {
return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
"tdb_new_database:"
" failed to truncate: %s", strerror(errno));
}
- rlen = write(tdb->fd, &newdb, sizeof(newdb));
+ rlen = write(tdb->file->fd, &newdb, sizeof(newdb));
if (rlen != sizeof(newdb)) {
if (rlen >= 0)
errno = ENOSPC;
return TDB_SUCCESS;
}
+static enum TDB_ERROR tdb_new_file(struct tdb_context *tdb)
+{
+ tdb->file = malloc(sizeof(*tdb->file));
+ if (!tdb->file)
+ return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
+ "tdb_open: cannot alloc tdb_file structure");
+ tdb->file->num_lockrecs = 0;
+ tdb->file->lockrecs = NULL;
+ tdb->file->allrecord_lock.count = 0;
+ tdb->file->refcnt = 1;
+ return TDB_SUCCESS;
+}
+
+enum TDB_ERROR tdb_set_attribute(struct tdb_context *tdb,
+ const union tdb_attribute *attr)
+{
+ switch (attr->base.attr) {
+ case TDB_ATTRIBUTE_LOG:
+ tdb->log_fn = attr->log.fn;
+ tdb->log_data = attr->log.data;
+ break;
+ case TDB_ATTRIBUTE_HASH:
+ case TDB_ATTRIBUTE_SEED:
+ case TDB_ATTRIBUTE_OPENHOOK:
+ return tdb->last_error
+ = tdb_logerr(tdb, TDB_ERR_EINVAL,
+ TDB_LOG_USE_ERROR,
+ "tdb_set_attribute:"
+ " cannot set %s after opening",
+ attr->base.attr == TDB_ATTRIBUTE_HASH
+ ? "TDB_ATTRIBUTE_HASH"
+ : attr->base.attr == TDB_ATTRIBUTE_SEED
+ ? "TDB_ATTRIBUTE_SEED"
+ : "TDB_ATTRIBUTE_OPENHOOK");
+ case TDB_ATTRIBUTE_STATS:
+ return tdb->last_error
+ = tdb_logerr(tdb, TDB_ERR_EINVAL,
+ TDB_LOG_USE_ERROR,
+ "tdb_set_attribute:"
+ " cannot set TDB_ATTRIBUTE_STATS");
+ case TDB_ATTRIBUTE_FLOCK:
+ tdb->lock_fn = attr->flock.lock;
+ tdb->unlock_fn = attr->flock.unlock;
+ tdb->lock_data = attr->flock.data;
+ break;
+ default:
+ return tdb->last_error
+ = tdb_logerr(tdb, TDB_ERR_EINVAL,
+ TDB_LOG_USE_ERROR,
+ "tdb_set_attribute:"
+ " unknown attribute type %u",
+ attr->base.attr);
+ }
+ return TDB_SUCCESS;
+}
+
+static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
+ void *unused)
+{
+ uint64_t ret;
+ /* hash64_stable assumes lower bits are more important; they are a
+ * slightly better hash. We use the upper bits first, so swap them. */
+ ret = hash64_stable((const unsigned char *)key, length, seed);
+ return (ret >> 32) | (ret << 32);
+}
+
+enum TDB_ERROR tdb_get_attribute(struct tdb_context *tdb,
+ union tdb_attribute *attr)
+{
+ switch (attr->base.attr) {
+ case TDB_ATTRIBUTE_LOG:
+ if (!tdb->log_fn)
+ return tdb->last_error = TDB_ERR_NOEXIST;
+ attr->log.fn = tdb->log_fn;
+ attr->log.data = tdb->log_data;
+ break;
+ case TDB_ATTRIBUTE_HASH:
+ attr->hash.fn = tdb->hash_fn;
+ attr->hash.data = tdb->hash_data;
+ break;
+ case TDB_ATTRIBUTE_SEED:
+ attr->seed.seed = tdb->hash_seed;
+ break;
+ case TDB_ATTRIBUTE_OPENHOOK:
+ return tdb->last_error
+ = tdb_logerr(tdb, TDB_ERR_EINVAL,
+ TDB_LOG_USE_ERROR,
+ "tdb_get_attribute:"
+ " cannot get TDB_ATTRIBUTE_OPENHOOK");
+ case TDB_ATTRIBUTE_STATS: {
+ size_t size = attr->stats.size;
+ if (size > tdb->stats.size)
+ size = tdb->stats.size;
+ memcpy(&attr->stats, &tdb->stats, size);
+ break;
+ }
+ case TDB_ATTRIBUTE_FLOCK:
+ attr->flock.lock = tdb->lock_fn;
+ attr->flock.unlock = tdb->unlock_fn;
+ attr->flock.data = tdb->lock_data;
+ break;
+ default:
+ return tdb->last_error
+ = tdb_logerr(tdb, TDB_ERR_EINVAL,
+ TDB_LOG_USE_ERROR,
+ "tdb_get_attribute:"
+ " unknown attribute type %u",
+ attr->base.attr);
+ }
+ attr->base.next = NULL;
+ return TDB_SUCCESS;
+}
+
+void tdb_unset_attribute(struct tdb_context *tdb,
+ enum tdb_attribute_type type)
+{
+ switch (type) {
+ case TDB_ATTRIBUTE_LOG:
+ tdb->log_fn = NULL;
+ break;
+ case TDB_ATTRIBUTE_HASH:
+ case TDB_ATTRIBUTE_SEED:
+ case TDB_ATTRIBUTE_OPENHOOK:
+ tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
+ "tdb_unset_attribute: cannot unset %s after opening",
+ type == TDB_ATTRIBUTE_HASH
+ ? "TDB_ATTRIBUTE_HASH"
+ : type == TDB_ATTRIBUTE_SEED
+ ? "TDB_ATTRIBUTE_SEED"
+ : "TDB_ATTRIBUTE_OPENHOOK");
+ break;
+ case TDB_ATTRIBUTE_STATS:
+ tdb_logerr(tdb, TDB_ERR_EINVAL,
+ TDB_LOG_USE_ERROR,
+ "tdb_unset_attribute:"
+ "cannot unset TDB_ATTRIBUTE_STATS");
+ break;
+ case TDB_ATTRIBUTE_FLOCK:
+ tdb->lock_fn = tdb_fcntl_lock;
+ tdb->unlock_fn = tdb_fcntl_unlock;
+ break;
+ default:
+ tdb_logerr(tdb, TDB_ERR_EINVAL,
+ TDB_LOG_USE_ERROR,
+ "tdb_unset_attribute: unknown attribute type %u",
+ type);
+ }
+}
+
struct tdb_context *tdb_open(const char *name, int tdb_flags,
int open_flags, mode_t mode,
union tdb_attribute *attr)
ssize_t rlen;
struct tdb_header hdr;
struct tdb_attribute_seed *seed = NULL;
+ struct tdb_attribute_openhook *openhook = NULL;
tdb_bool_err berr;
enum TDB_ERROR ecode;
+ int openlock;
- tdb = malloc(sizeof(*tdb));
+ tdb = malloc(sizeof(*tdb) + (name ? strlen(name) + 1 : 0));
if (!tdb) {
/* Can't log this */
errno = ENOMEM;
return NULL;
}
- tdb->name = NULL;
- tdb->map_ptr = NULL;
+ /* Set name immediately for logging functions. */
+ if (name) {
+ tdb->name = strcpy((char *)(tdb + 1), name);
+ } else {
+ tdb->name = NULL;
+ }
tdb->direct_access = 0;
- tdb->fd = -1;
- tdb->map_size = sizeof(struct tdb_header);
tdb->flags = tdb_flags;
- tdb->logfn = NULL;
+ tdb->log_fn = NULL;
tdb->transaction = NULL;
- tdb->stats = NULL;
tdb->access = NULL;
- tdb_hash_init(tdb);
+ tdb->last_error = TDB_SUCCESS;
+ tdb->file = NULL;
+ tdb->lock_fn = tdb_fcntl_lock;
+ tdb->unlock_fn = tdb_fcntl_unlock;
+ tdb->hash_fn = jenkins_hash;
+ memset(&tdb->stats, 0, sizeof(tdb->stats));
+ tdb->stats.base.attr = TDB_ATTRIBUTE_STATS;
+ tdb->stats.size = sizeof(tdb->stats);
tdb_io_init(tdb);
- tdb_lock_init(tdb);
while (attr) {
switch (attr->base.attr) {
- case TDB_ATTRIBUTE_LOG:
- tdb->logfn = attr->log.log_fn;
- tdb->log_private = attr->log.log_private;
- break;
case TDB_ATTRIBUTE_HASH:
- tdb->khash = attr->hash.hash_fn;
- tdb->hash_priv = attr->hash.hash_private;
+ tdb->hash_fn = attr->hash.fn;
+ tdb->hash_data = attr->hash.data;
break;
case TDB_ATTRIBUTE_SEED:
seed = &attr->seed;
break;
- case TDB_ATTRIBUTE_STATS:
- tdb->stats = &attr->stats;
- /* They have stats we don't know about? Tell them. */
- if (tdb->stats->size > sizeof(attr->stats))
- tdb->stats->size = sizeof(attr->stats);
+ case TDB_ATTRIBUTE_OPENHOOK:
+ openhook = &attr->openhook;
break;
default:
- ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
- TDB_LOG_USE_ERROR,
- "tdb_open:"
- " unknown attribute type %u",
- attr->base.attr);
- goto fail;
+ /* These are set as normal. */
+ ecode = tdb_set_attribute(tdb, attr);
+ if (ecode != TDB_SUCCESS)
+ goto fail;
}
attr = attr->base.next;
}
if (tdb_flags & ~(TDB_INTERNAL | TDB_NOLOCK | TDB_NOMMAP | TDB_CONVERT
- | TDB_NOSYNC)) {
+ | TDB_NOSYNC | TDB_SEQNUM | TDB_ALLOW_NESTING)) {
ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
"tdb_open: unknown flags %u", tdb_flags);
goto fail;
if ((open_flags & O_ACCMODE) == O_RDONLY) {
tdb->read_only = true;
tdb->mmap_flags = PROT_READ;
+ openlock = F_RDLCK;
} else {
tdb->read_only = false;
tdb->mmap_flags = PROT_READ | PROT_WRITE;
+ openlock = F_WRLCK;
}
/* internal databases don't need any of the rest. */
if (tdb->flags & TDB_INTERNAL) {
tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
+ ecode = tdb_new_file(tdb);
+ if (ecode != TDB_SUCCESS) {
+ goto fail;
+ }
+ tdb->file->fd = -1;
ecode = tdb_new_database(tdb, seed, &hdr);
if (ecode != TDB_SUCCESS) {
goto fail;
return tdb;
}
- if ((tdb->fd = open(name, open_flags, mode)) == -1) {
- /* errno set by open(2) */
- saved_errno = errno;
- ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
+ if (stat(name, &st) != -1)
+ tdb->file = find_file(st.st_dev, st.st_ino);
+
+ if (!tdb->file) {
+ int fd;
+
+ if ((fd = open(name, open_flags, mode)) == -1) {
+ /* errno set by open(2) */
+ saved_errno = errno;
+ tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
"tdb_open: could not open file %s: %s",
name, strerror(errno));
- goto fail;
- }
+ goto fail_errno;
+ }
- /* on exec, don't inherit the fd */
- v = fcntl(tdb->fd, F_GETFD, 0);
- fcntl(tdb->fd, F_SETFD, v | FD_CLOEXEC);
+ /* on exec, don't inherit the fd */
+ v = fcntl(fd, F_GETFD, 0);
+ fcntl(fd, F_SETFD, v | FD_CLOEXEC);
+
+ if (fstat(fd, &st) == -1) {
+ saved_errno = errno;
+ tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
+ "tdb_open: could not stat open %s: %s",
+ name, strerror(errno));
+ close(fd);
+ goto fail_errno;
+ }
+
+ ecode = tdb_new_file(tdb);
+ if (ecode != TDB_SUCCESS) {
+ close(fd);
+ goto fail;
+ }
+
+ tdb->file->next = files;
+ tdb->file->fd = fd;
+ tdb->file->device = st.st_dev;
+ tdb->file->inode = st.st_ino;
+ tdb->file->map_ptr = NULL;
+ tdb->file->map_size = sizeof(struct tdb_header);
+ }
/* ensure there is only one process initialising at once */
- ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
+ ecode = tdb_lock_open(tdb, openlock, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
if (ecode != TDB_SUCCESS) {
- goto fail;
+ saved_errno = errno;
+ goto fail_errno;
+ }
+
+ /* call their open hook if they gave us one. */
+ if (openhook) {
+ ecode = openhook->fn(tdb->file->fd, openhook->data);
+ if (ecode != TDB_SUCCESS) {
+ tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
+ "tdb_open: open hook failed");
+ goto fail;
+ }
+ open_flags |= O_CREAT;
}
/* If they used O_TRUNC, read will return 0. */
- rlen = read(tdb->fd, &hdr, sizeof(hdr));
+ rlen = pread(tdb->file->fd, &hdr, sizeof(hdr), 0);
if (rlen == 0 && (open_flags & O_CREAT)) {
ecode = tdb_new_database(tdb, seed, &hdr);
if (ecode != TDB_SUCCESS) {
goto fail;
}
- if (fstat(tdb->fd, &st) == -1) {
- saved_errno = errno;
- ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
- "tdb_open: could not stat open %s: %s",
- name, strerror(errno));
- goto fail;
- }
-
- /* Is it already in the open list? If so, fail. */
- if (tdb_already_open(st.st_dev, st.st_ino)) {
- /* FIXME */
- ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_USE_ERROR,
- "tdb_open: %s (%d,%d) is already open"
- " in this process",
- name, (int)st.st_dev, (int)st.st_ino);
- goto fail;
- }
-
- tdb->name = strdup(name);
- if (!tdb->name) {
- ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
- "tdb_open: failed to allocate name");
- goto fail;
- }
-
/* Clear any features we don't understand. */
if ((open_flags & O_ACCMODE) != O_RDONLY) {
hdr.features_used &= TDB_FEATURE_MASK;
goto fail;
}
- tdb->device = st.st_dev;
- tdb->inode = st.st_ino;
- tdb_unlock_open(tdb);
+ tdb_unlock_open(tdb, openlock);
/* This make sure we have current map_size and mmap. */
- tdb->methods->oob(tdb, tdb->map_size + 1, true);
+ tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
/* Now it's fully formed, recover if necessary. */
berr = tdb_needs_recovery(tdb);
goto fail;
}
- tdb->next = tdbs;
- tdbs = tdb;
+ /* Add to linked list if we're new. */
+ if (tdb->file->refcnt == 1)
+ files = tdb->file;
return tdb;
fail:
/* Map ecode to some logical errno. */
- if (!saved_errno) {
- switch (ecode) {
- case TDB_ERR_CORRUPT:
- case TDB_ERR_IO:
- saved_errno = EIO;
- break;
- case TDB_ERR_LOCK:
- saved_errno = EWOULDBLOCK;
- break;
- case TDB_ERR_OOM:
- saved_errno = ENOMEM;
- break;
- case TDB_ERR_EINVAL:
- saved_errno = EINVAL;
- break;
- default:
- saved_errno = EINVAL;
- break;
- }
- }
-
+ switch (ecode) {
+ case TDB_ERR_CORRUPT:
+ case TDB_ERR_IO:
+ saved_errno = EIO;
+ break;
+ case TDB_ERR_LOCK:
+ saved_errno = EWOULDBLOCK;
+ break;
+ case TDB_ERR_OOM:
+ saved_errno = ENOMEM;
+ break;
+ case TDB_ERR_EINVAL:
+ saved_errno = EINVAL;
+ break;
+ default:
+ saved_errno = EINVAL;
+ break;
+ }
+
+fail_errno:
#ifdef TDB_TRACE
close(tdb->tracefd);
#endif
- if (tdb->map_ptr) {
- if (tdb->flags & TDB_INTERNAL) {
- free(tdb->map_ptr);
- } else
- tdb_munmap(tdb);
- }
- free(tdb->lockrecs);
- free((char *)tdb->name);
- if (tdb->fd != -1)
- if (close(tdb->fd) != 0)
- tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
- "tdb_open: failed to close tdb->fd"
- " on error: %s", strerror(errno));
+ if (tdb->file) {
+ tdb_lock_cleanup(tdb);
+ if (--tdb->file->refcnt == 0) {
+ assert(tdb->file->num_lockrecs == 0);
+ if (tdb->file->map_ptr) {
+ if (tdb->flags & TDB_INTERNAL) {
+ free(tdb->file->map_ptr);
+ } else
+ tdb_munmap(tdb->file);
+ }
+ if (close(tdb->file->fd) != 0)
+ tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
+ "tdb_open: failed to close tdb fd"
+ " on error: %s", strerror(errno));
+ free(tdb->file->lockrecs);
+ free(tdb->file);
+ }
+ }
+
free(tdb);
errno = saved_errno;
return NULL;
int tdb_close(struct tdb_context *tdb)
{
- struct tdb_context **i;
int ret = 0;
tdb_trace(tdb, "tdb_close");
tdb_transaction_cancel(tdb);
}
- if (tdb->map_ptr) {
+ if (tdb->file->map_ptr) {
if (tdb->flags & TDB_INTERNAL)
- free(tdb->map_ptr);
+ free(tdb->file->map_ptr);
else
- tdb_munmap(tdb);
+ tdb_munmap(tdb->file);
}
- free((char *)tdb->name);
- if (tdb->fd != -1) {
- ret = close(tdb->fd);
- tdb->fd = -1;
- }
- free(tdb->lockrecs);
+ if (tdb->file) {
+ struct tdb_file **i;
- /* Remove from contexts list */
- for (i = &tdbs; *i; i = &(*i)->next) {
- if (*i == tdb) {
- *i = tdb->next;
- break;
+ tdb_lock_cleanup(tdb);
+ if (--tdb->file->refcnt == 0) {
+ ret = close(tdb->file->fd);
+
+ /* Remove from files list */
+ for (i = &files; *i; i = &(*i)->next) {
+ if (*i == tdb->file) {
+ *i = tdb->file->next;
+ break;
+ }
+ }
+ free(tdb->file->lockrecs);
+ free(tdb->file);
}
}
projects / ccan / blobdiff