X-Git-Url: https://git.ozlabs.org/?a=blobdiff_plain;f=ccan%2Ftdb2%2Fopen.c;h=c6fd9a090ecf3078bb4135e6b0651403d8379978;hb=703cea0c78a896c658272af06306218fc0bb23cc;hp=21881c350e084dcc5751c395a773ec974662244b;hpb=efdf0f2d8f34b4c01c82c558b350ec36c7329b1e;p=ccan

diff --git a/ccan/tdb2/open.c b/ccan/tdb2/open.c
index 21881c35..c6fd9a09 100644
--- a/ccan/tdb2/open.c
+++ b/ccan/tdb2/open.c
@@ -1,19 +1,21 @@
 #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)
@@ -100,12 +102,13 @@ static enum TDB_ERROR tdb_new_database(struct tdb_context *tdb,
 	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));
@@ -133,29 +136,29 @@ static enum TDB_ERROR tdb_new_database(struct tdb_context *tdb,
 	*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;
@@ -166,6 +169,142 @@ static enum TDB_ERROR tdb_new_database(struct tdb_context *tdb,
 	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: could 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_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:
+		/* FIXME */
+		return TDB_ERR_EINVAL;
+	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:
+		/* FIXME */
+		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)
@@ -178,6 +317,7 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	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;
 
@@ -188,28 +328,24 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 		return NULL;
 	}
 	tdb->name = NULL;
-	tdb->map_ptr = 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;
 	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;
@@ -220,19 +356,20 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 			if (tdb->stats->size > sizeof(attr->stats))
 				tdb->stats->size = sizeof(attr->stats);
 			break;
+		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)) {
 		ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 				   "tdb_open: unknown flags %u", tdb_flags);
 		goto fail;
@@ -256,37 +393,90 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	/* 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;
 		}
+		if (name) {
+			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;
+			}
+		}
 		tdb_convert(tdb, &hdr.hash_seed, sizeof(hdr.hash_seed));
 		tdb->hash_seed = hdr.hash_seed;
 		tdb_ftable_init(tdb);
 		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(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));
+			goto fail_errno;
+		}
+
+		ecode = tdb_new_file(tdb);
+		if (ecode != TDB_SUCCESS)
+			goto fail;
 
-	/* on exec, don't inherit the fd */
-	v = fcntl(tdb->fd, F_GETFD, 0);
-        fcntl(tdb->fd, F_SETFD, v | FD_CLOEXEC);
+		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);
 	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) {
@@ -330,24 +520,6 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 		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,
@@ -365,12 +537,10 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 			goto fail;
 	}
 
-	tdb->device = st.st_dev;
-	tdb->inode = st.st_ino;
 	tdb_unlock_open(tdb);
 
 	/* 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);
@@ -390,49 +560,56 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 		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));
+	free(cast_const(char *, tdb->name));
+	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;
@@ -440,7 +617,6 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 
 int tdb_close(struct tdb_context *tdb)
 {
-	struct tdb_context **i;
 	int ret = 0;
 
 	tdb_trace(tdb, "tdb_close");
@@ -449,24 +625,29 @@ int tdb_close(struct tdb_context *tdb)
 		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);
+	free(cast_const(char *, tdb->name));
+	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);
 		}
 	}