X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Fprivate.h;h=4bce11a1c539ea8e39bb09c6179d495bf669e342;hp=9ae5bed2e6bc6cb8c397d32d1f7f6f6ef8574930;hb=c8f6f8c2dea76042f74c02eff048847e62adcca6;hpb=1ad66fedf81fdaf0f07a2b00d6787614c58a23ef;ds=sidebyside
diff --git a/ccan/tdb2/private.h b/ccan/tdb2/private.h
index 9ae5bed2..4bce11a1 100644
--- a/ccan/tdb2/private.h
+++ b/ccan/tdb2/private.h
@@ -18,26 +18,19 @@
License along with this library; if not, see .
*/
-#include "config.h"
-#include
-#include
+#include
#include
#include
#include
#include
#include
#include
-#include
#include
#include
#include
#include
-#include
#include
-#include
-#if HAVE_BYTESWAP_H
-#include
-#endif
+#include
#ifndef TEST_IT
#define TEST_IT(cond)
@@ -75,7 +68,7 @@ typedef uint64_t tdb_off_t;
/* Packing errors into pointers and v.v. */
#define TDB_PTR_IS_ERR(ptr) \
- unlikely((void *)(ptr) >= (void *)(long)TDB_ERR_LAST)
+ unlikely((unsigned long)(ptr) >= (unsigned long)TDB_ERR_LAST)
#define TDB_PTR_ERR(p) ((enum TDB_ERROR)(long)(p))
#define TDB_ERR_PTR(err) ((void *)(long)(err))
@@ -84,12 +77,12 @@ typedef int tdb_bool_err;
/* Prevent others from opening the file. */
#define TDB_OPEN_LOCK 0
-/* Doing a transaction. */
-#define TDB_TRANSACTION_LOCK 1
/* Expanding file. */
#define TDB_EXPANSION_LOCK 2
+/* Doing a transaction. */
+#define TDB_TRANSACTION_LOCK 8
/* Hash chain locks. */
-#define TDB_HASH_LOCK_START 3
+#define TDB_HASH_LOCK_START 64
/* Range for hash locks. */
#define TDB_HASH_LOCK_RANGE_BITS 30
@@ -134,20 +127,6 @@ typedef int tdb_bool_err;
/* Indicates this entry is not on an flist (can happen during coalescing) */
#define TDB_FTABLE_NONE ((1ULL << TDB_OFF_UPPER_STEAL) - 1)
-#if !HAVE_BSWAP_64
-static inline uint64_t bswap_64(uint64_t x)
-{
- return (((x&0x000000FFULL)<<56)
- | ((x&0x0000FF00ULL)<<48)
- | ((x&0x00FF0000ULL)<<40)
- | ((x&0xFF000000ULL)<<32)
- | ((x>>8)&0xFF000000ULL)
- | ((x>>16)&0x00FF0000ULL)
- | ((x>>24)&0x0000FF00ULL)
- | ((x>>32)&0x000000FFULL));
-}
-#endif
-
struct tdb_used_record {
/* For on-disk compatibility, we avoid bitfields:
magic: 16, (highest)
@@ -219,7 +198,7 @@ static inline unsigned frec_ftable(const struct tdb_free_record *f)
struct tdb_recovery_record {
uint64_t magic;
- /* Length of record. */
+ /* Length of record (add this header to get total length). */
uint64_t max_len;
/* Length used. */
uint64_t len;
@@ -246,7 +225,9 @@ struct tdb_header {
uint64_t features_used; /* Features all writers understand */
uint64_t features_offered; /* Features offered */
- tdb_off_t reserved[24];
+ uint64_t seqnum; /* Sequence number for TDB_SEQNUM */
+
+ tdb_off_t reserved[23];
/* Top level hash table. */
tdb_off_t hashtable[1ULL << TDB_TOPLEVEL_HASH_BITS];
@@ -290,6 +271,9 @@ struct traverse_info {
tdb_off_t prev;
};
+typedef uint32_t tdb1_len_t;
+typedef uint32_t tdb1_off_t;
+
enum tdb_lock_flags {
/* WAIT == F_SETLKW, NOWAIT == F_SETLK */
TDB_LOCK_NOWAIT = 0,
@@ -300,8 +284,9 @@ enum tdb_lock_flags {
TDB_LOCK_NOCHECK = 4,
};
-struct tdb_lock_type {
- uint32_t off;
+struct tdb_lock {
+ struct tdb_context *owner;
+ off_t off;
uint32_t count;
uint32_t ltype;
};
@@ -316,76 +301,27 @@ struct tdb_access_hdr {
};
struct tdb_file {
- /* Single list of all TDBs, to detect multiple opens. */
- struct tdb_file *next;
-
- /* The file descriptor. */
- int fd;
-
- /* Lock information */
- struct tdb_lock_type allrecord_lock;
- size_t num_lockrecs;
- struct tdb_lock_type *lockrecs;
-
- /* Identity of this file. */
- dev_t device;
- ino_t inode;
-};
-
-struct tdb_context {
- /* Filename of the database. */
- const char *name;
+ /* How many are sharing us? */
+ unsigned int refcnt;
/* Mmap (if any), or malloc (for TDB_INTERNAL). */
void *map_ptr;
- /* Are we accessing directly? (debugging check). */
- int direct_access;
-
/* How much space has been mapped (<= current file size) */
tdb_len_t map_size;
- /* Operating read-only? (Opened O_RDONLY, or in traverse_read) */
- bool read_only;
-
- /* mmap read only? */
- int mmap_flags;
-
- /* Error code for last tdb error. */
- enum TDB_ERROR ecode;
-
- /* the flags passed to tdb_open, for tdb_reopen. */
- uint32_t flags;
-
- /* Logging function */
- void (*logfn)(struct tdb_context *tdb,
- enum tdb_log_level level,
- void *log_private,
- const char *message);
- void *log_private;
-
- /* Hash function. */
- uint64_t (*khash)(const void *key, size_t len, uint64_t seed, void *);
- void *hash_priv;
- uint64_t hash_seed;
-
- /* Set if we are in a transaction. */
- struct tdb_transaction *transaction;
-
- /* What free table are we using? */
- tdb_off_t ftable_off;
- unsigned int ftable;
-
- /* IO methods: changes for transactions. */
- const struct tdb_methods *methods;
-
- struct tdb_attribute_stats *stats;
+ /* The file descriptor (-1 for TDB_INTERNAL). */
+ int fd;
- /* Direct access information */
- struct tdb_access_hdr *access;
+ /* Lock information */
+ pid_t locker;
+ struct tdb_lock allrecord_lock;
+ size_t num_lockrecs;
+ struct tdb_lock *lockrecs;
- /* The actual file information */
- struct tdb_file *file;
+ /* Identity of this file. */
+ dev_t device;
+ ino_t inode;
};
struct tdb_methods {
@@ -402,7 +338,8 @@ struct tdb_methods {
internal prototypes
*/
/* hash.c: */
-void tdb_hash_init(struct tdb_context *tdb);
+uint64_t tdb_jenkins_hash(const void *key, size_t length, uint64_t seed,
+ void *unused);
tdb_bool_err first_in_hash(struct tdb_context *tdb,
struct traverse_info *tinfo,
@@ -451,7 +388,9 @@ tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
/* Put this record in a free list. */
enum TDB_ERROR add_free_record(struct tdb_context *tdb,
- tdb_off_t off, tdb_len_t len_with_header);
+ tdb_off_t off, tdb_len_t len_with_header,
+ enum tdb_lock_flags waitflag,
+ bool coalesce_ok);
/* Set up header for a used/ftable/htable/chain record. */
enum TDB_ERROR set_header(struct tdb_context *tdb,
@@ -474,7 +413,7 @@ void tdb_io_init(struct tdb_context *tdb);
void *tdb_convert(const struct tdb_context *tdb, void *buf, tdb_len_t size);
/* Unmap and try to map the tdb. */
-void tdb_munmap(struct tdb_context *tdb);
+void tdb_munmap(struct tdb_file *file);
void tdb_mmap(struct tdb_context *tdb);
/* Either alloc a copy, or give direct access. Release frees or noop. */
@@ -519,13 +458,8 @@ enum TDB_ERROR tdb_write_convert(struct tdb_context *tdb, tdb_off_t off,
enum TDB_ERROR tdb_read_convert(struct tdb_context *tdb, tdb_off_t off,
void *rec, size_t len);
-/* Adds a stat, if it's in range. */
-void add_stat_(struct tdb_context *tdb, uint64_t *stat, size_t val);
-#define add_stat(tdb, statname, val) \
- do { \
- if (unlikely((tdb)->stats)) \
- add_stat_((tdb), &(tdb)->stats->statname, (val)); \
- } while (0)
+/* Bump the seqnum (caller checks for tdb->flags & TDB_SEQNUM) */
+void tdb_inc_seqnum(struct tdb_context *tdb);
/* lock.c: */
/* Lock/unlock a range of hashes. */
@@ -536,6 +470,9 @@ enum TDB_ERROR tdb_unlock_hashes(struct tdb_context *tdb,
tdb_off_t hash_lock,
tdb_len_t hash_range, int ltype);
+/* For closing the file. */
+void tdb_lock_cleanup(struct tdb_context *tdb);
+
/* Lock/unlock a particular free bucket. */
enum TDB_ERROR tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
enum tdb_lock_flags waitflag);
@@ -552,12 +489,12 @@ bool tdb_has_hash_locks(struct tdb_context *tdb);
enum TDB_ERROR tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
enum tdb_lock_flags flags, bool upgradable);
void tdb_allrecord_unlock(struct tdb_context *tdb, int ltype);
-enum TDB_ERROR tdb_allrecord_upgrade(struct tdb_context *tdb);
+enum TDB_ERROR tdb_allrecord_upgrade(struct tdb_context *tdb, off_t start);
/* Serialize db open. */
enum TDB_ERROR tdb_lock_open(struct tdb_context *tdb,
- enum tdb_lock_flags flags);
-void tdb_unlock_open(struct tdb_context *tdb);
+ int ltype, enum tdb_lock_flags flags);
+void tdb_unlock_open(struct tdb_context *tdb, int ltype);
bool tdb_has_open_lock(struct tdb_context *tdb);
/* Serialize db expand. */
@@ -568,10 +505,129 @@ bool tdb_has_expansion_lock(struct tdb_context *tdb);
/* If it needs recovery, grab all the locks and do it. */
enum TDB_ERROR tdb_lock_and_recover(struct tdb_context *tdb);
+/* Byte-range lock wrappers for TDB1 to access. */
+enum TDB_ERROR tdb_brlock(struct tdb_context *tdb,
+ int rw_type, tdb_off_t offset, tdb_off_t len,
+ enum tdb_lock_flags flags);
+
+enum TDB_ERROR tdb_brunlock(struct tdb_context *tdb,
+ int rw_type, tdb_off_t offset, size_t len);
+
+enum TDB_ERROR tdb_nest_lock(struct tdb_context *tdb,
+ tdb_off_t offset, int ltype,
+ enum tdb_lock_flags flags);
+
+enum TDB_ERROR tdb_nest_unlock(struct tdb_context *tdb,
+ tdb_off_t off, int ltype);
+
+enum TDB_ERROR tdb_lock_gradual(struct tdb_context *tdb,
+ int ltype, enum tdb_lock_flags flags,
+ tdb_off_t off, tdb_off_t len);
+
+/* Default lock and unlock functions. */
+int tdb_fcntl_lock(int fd, int rw, off_t off, off_t len, bool waitflag, void *);
+int tdb_fcntl_unlock(int fd, int rw, off_t off, off_t len, void *);
+
/* transaction.c: */
enum TDB_ERROR tdb_transaction_recover(struct tdb_context *tdb);
tdb_bool_err tdb_needs_recovery(struct tdb_context *tdb);
+/* this is stored at the front of every database */
+struct tdb1_header {
+ char magic_food[32]; /* for /etc/magic */
+ uint32_t version; /* version of the code */
+ uint32_t hash_size; /* number of hash entries */
+ tdb1_off_t rwlocks; /* obsolete - kept to detect old formats */
+ tdb1_off_t recovery_start; /* offset of transaction recovery region */
+ tdb1_off_t sequence_number; /* used when TDB1_SEQNUM is set */
+ uint32_t magic1_hash; /* hash of TDB_MAGIC_FOOD. */
+ uint32_t magic2_hash; /* hash of TDB1_MAGIC. */
+ tdb1_off_t reserved[27];
+};
+
+struct tdb1_traverse_lock {
+ struct tdb1_traverse_lock *next;
+ uint32_t off;
+ uint32_t hash;
+ int lock_rw;
+};
+
+struct tdb_context {
+ /* Single list of all TDBs, to detect multiple opens. */
+ struct tdb_context *next;
+
+ /* Filename of the database. */
+ const char *name;
+
+ /* Logging function */
+ void (*log_fn)(struct tdb_context *tdb,
+ enum tdb_log_level level,
+ enum TDB_ERROR ecode,
+ const char *message,
+ void *data);
+ void *log_data;
+
+ /* Open flags passed to tdb_open. */
+ int open_flags;
+
+ /* low level (fnctl) lock functions. */
+ int (*lock_fn)(int fd, int rw, off_t off, off_t len, bool w, void *);
+ int (*unlock_fn)(int fd, int rw, off_t off, off_t len, void *);
+ void *lock_data;
+
+ /* the tdb flags passed to tdb_open. */
+ uint32_t flags;
+
+ /* Our statistics. */
+ struct tdb_attribute_stats stats;
+
+ /* The actual file information */
+ struct tdb_file *file;
+
+ /* Hash function. */
+ uint64_t (*hash_fn)(const void *key, size_t len, uint64_t seed, void *);
+ void *hash_data;
+ uint64_t hash_seed;
+
+ /* Our open hook, if any. */
+ enum TDB_ERROR (*openhook)(int fd, void *data);
+ void *openhook_data;
+
+ /* Last error we returned. */
+ enum TDB_ERROR last_error;
+
+ struct {
+
+ /* Are we accessing directly? (debugging check). */
+ int direct_access;
+
+ /* Set if we are in a transaction. */
+ struct tdb_transaction *transaction;
+
+ /* What free table are we using? */
+ tdb_off_t ftable_off;
+ unsigned int ftable;
+
+ /* IO methods: changes for transactions. */
+ const struct tdb_methods *io;
+
+ /* Direct access information */
+ struct tdb_access_hdr *access;
+ } tdb2;
+
+ struct {
+ int traverse_read; /* read-only traversal */
+ int traverse_write; /* read-write traversal */
+
+ struct tdb1_header header; /* a cached copy of the header */
+ struct tdb1_traverse_lock travlocks; /* current traversal locks */
+ const struct tdb1_methods *io;
+ struct tdb1_transaction *transaction;
+ int page_size;
+ int max_dead_records;
+ } tdb1;
+};
+
/* tdb.c: */
enum TDB_ERROR COLD tdb_logerr(struct tdb_context *tdb,
enum TDB_ERROR ecode,