tdb2: update documentation
[ccan] / ccan / tdb2 / tdb.c
index 881a6501523a53a7b2358c3640deebe0867bca15..dfa68451e51a878d4be75f5431121ff1f012b9d4 100644 (file)
@@ -1,9 +1,7 @@
 #include "private.h"
 #include <ccan/tdb2/tdb2.h>
-#include <ccan/hash/hash.h>
-#include <ccan/build_assert/build_assert.h>
-#include <ccan/likely/likely.h>
 #include <assert.h>
+#include <stdarg.h>
 
 /* The null return. */
 struct tdb_data tdb_null = { .dptr = NULL, .dsize = 0 };
@@ -11,58 +9,6 @@ struct tdb_data tdb_null = { .dptr = NULL, .dsize = 0 };
 /* all contexts, to ensure no double-opens (fcntl locks don't nest!) */
 static struct tdb_context *tdbs = NULL;
 
-PRINTF_ATTRIBUTE(4, 5) static void
-null_log_fn(struct tdb_context *tdb,
-           enum tdb_debug_level level, void *priv,
-           const char *fmt, ...)
-{
-}
-
-/* We do a lot of work assuming our copy of the header volatile area
- * is uptodate, and usually it is.  However, once we grab a lock, we have to
- * re-check it. */
-bool header_changed(struct tdb_context *tdb)
-{
-       uint64_t gen;
-
-       if (!(tdb->flags & TDB_NOLOCK) && tdb->header_uptodate) {
-               tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
-                        "warning: header uptodate already\n");
-       }
-
-       /* We could get a partial update if we're not holding any locks. */
-       assert((tdb->flags & TDB_NOLOCK) || tdb_has_locks(tdb));
-
-       tdb->header_uptodate = true;
-       gen = tdb_read_off(tdb, offsetof(struct tdb_header, v.generation));
-       if (unlikely(gen != tdb->header.v.generation)) {
-               tdb_read_convert(tdb, offsetof(struct tdb_header, v),
-                                &tdb->header.v, sizeof(tdb->header.v));
-               return true;
-       }
-       return false;
-}
-
-int write_header(struct tdb_context *tdb)
-{
-       assert(tdb_read_off(tdb, offsetof(struct tdb_header, v.generation))
-              == tdb->header.v.generation);
-       tdb->header.v.generation++;
-       return tdb_write_convert(tdb, offsetof(struct tdb_header, v),
-                                &tdb->header.v, sizeof(tdb->header.v));
-}
-
-static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
-                            void *arg)
-{
-       return hash64_stable((const unsigned char *)key, length, seed);
-}
-
-uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len)
-{
-       return tdb->khash(ptr, len, tdb->header.hash_seed, tdb->hash_priv);
-}
-
 static bool tdb_already_open(dev_t device, ino_t ino)
 {
        struct tdb_context *i;
@@ -85,8 +31,8 @@ static uint64_t random_number(struct tdb_context *tdb)
        fd = open("/dev/urandom", O_RDONLY);
        if (fd >= 0) {
                if (tdb_read_all(fd, &ret, sizeof(ret))) {
-                       tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
-                                "tdb_open: random from /dev/urandom\n");
+                       tdb_logerr(tdb, TDB_SUCCESS, TDB_DEBUG_TRACE,
+                                "tdb_open: random from /dev/urandom");
                        close(fd);
                        return ret;
                }
@@ -101,9 +47,9 @@ static uint64_t random_number(struct tdb_context *tdb)
                        char reply[1 + sizeof(uint64_t)];
                        int r = read(fd, reply, sizeof(reply));
                        if (r > 1) {
-                               tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
-                                        "tdb_open: %u random bytes from"
-                                        " /dev/egd-pool\n", r-1);
+                               tdb_logerr(tdb, TDB_SUCCESS, TDB_DEBUG_TRACE,
+                                          "tdb_open: %u random bytes from"
+                                          " /dev/egd-pool", r-1);
                                /* Copy at least some bytes. */
                                memcpy(&ret, reply+1, r - 1);
                                if (reply[0] == sizeof(uint64_t)
@@ -119,72 +65,65 @@ static uint64_t random_number(struct tdb_context *tdb)
        /* Fallback: pid and time. */
        gettimeofday(&now, NULL);
        ret = getpid() * 100132289ULL + now.tv_sec * 1000000ULL + now.tv_usec;
-       tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
-                "tdb_open: random from getpid and time\n");
+       tdb_logerr(tdb, TDB_SUCCESS, TDB_DEBUG_TRACE,
+                  "tdb_open: random from getpid and time");
        return ret;
 }
 
 struct new_database {
        struct tdb_header hdr;
-       struct tdb_used_record hrec;
-       tdb_off_t hash[1ULL << INITIAL_HASH_BITS];
-       struct tdb_used_record frec;
-       tdb_off_t free[INITIAL_FREE_BUCKETS + 1]; /* One overflow bucket */
+       struct tdb_freetable ftable;
 };
 
 /* initialise a new database */
-static int tdb_new_database(struct tdb_context *tdb)
+static int tdb_new_database(struct tdb_context *tdb,
+                           struct tdb_attribute_seed *seed,
+                           struct tdb_header *hdr)
 {
        /* We make it up in memory, then write it out if not internal */
        struct new_database newdb;
-       unsigned int magic_off = offsetof(struct tdb_header, magic_food);
+       unsigned int magic_len;
 
        /* Fill in the header */
        newdb.hdr.version = TDB_VERSION;
-       newdb.hdr.hash_seed = random_number(tdb);
+       if (seed)
+               newdb.hdr.hash_seed = seed->seed;
+       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.v.generation = 0;
-
-       /* The initial zone must cover the initial database size! */
-       BUILD_ASSERT((1ULL << INITIAL_ZONE_BITS) >= sizeof(newdb));
-
-       /* Free array has 1 zone, 10 buckets.  All buckets empty. */
-       newdb.hdr.v.num_zones = 1;
-       newdb.hdr.v.zone_bits = INITIAL_ZONE_BITS;
-       newdb.hdr.v.free_buckets = INITIAL_FREE_BUCKETS;
-       newdb.hdr.v.free_off = offsetof(struct new_database, free);
-       set_header(tdb, &newdb.frec, 0,
-                  sizeof(newdb.free), sizeof(newdb.free), 0);
-       memset(newdb.free, 0, sizeof(newdb.free));
-
+       newdb.hdr.recovery = 0;
+       memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
        /* Initial hashes are empty. */
-       newdb.hdr.v.hash_bits = INITIAL_HASH_BITS;
-       newdb.hdr.v.hash_off = offsetof(struct new_database, hash);
-       set_header(tdb, &newdb.hrec, 0,
-                  sizeof(newdb.hash), sizeof(newdb.hash), 0);
-       memset(newdb.hash, 0, sizeof(newdb.hash));
+       memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
+
+       /* Free is empty. */
+       newdb.hdr.free_table = offsetof(struct new_database, ftable);
+       memset(&newdb.ftable, 0, sizeof(newdb.ftable));
+       set_header(NULL, &newdb.ftable.hdr, TDB_FTABLE_MAGIC, 0,
+                  sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
+                  sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr), 0);
 
        /* Magic food */
        memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
        strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 
        /* This creates an endian-converted database, as if read from disk */
+       magic_len = sizeof(newdb.hdr.magic_food);
        tdb_convert(tdb,
-                   (char *)&newdb.hdr + magic_off,
-                   sizeof(newdb) - magic_off);
+                   (char *)&newdb.hdr + magic_len, sizeof(newdb) - magic_len);
 
-       tdb->header = newdb.hdr;
+       *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->ecode = TDB_ERR_OOM;
+                       tdb_logerr(tdb, TDB_ERR_OOM, TDB_DEBUG_FATAL,
+                                  "tdb_new_database: failed to allocate");
                        return -1;
                }
                memcpy(tdb->map_ptr, &newdb, tdb->map_size);
@@ -197,7 +136,9 @@ static int tdb_new_database(struct tdb_context *tdb)
                return -1;
 
        if (!tdb_pwrite_all(tdb->fd, &newdb, sizeof(newdb), 0)) {
-               tdb->ecode = TDB_ERR_IO;
+               tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_FATAL,
+                          "tdb_new_database: failed to write: %s",
+                          strerror(errno));
                return -1;
        }
        return 0;
@@ -209,57 +150,64 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 {
        struct tdb_context *tdb;
        struct stat st;
-       int save_errno;
+       int saved_errno = 0;
        uint64_t hash_test;
        unsigned v;
+       struct tdb_header hdr;
+       struct tdb_attribute_seed *seed = NULL;
 
        tdb = malloc(sizeof(*tdb));
        if (!tdb) {
                /* Can't log this */
                errno = ENOMEM;
-               goto fail;
+               return NULL;
        }
        tdb->name = NULL;
        tdb->map_ptr = NULL;
+       tdb->direct_access = 0;
        tdb->fd = -1;
-       /* map_size will be set below. */
+       tdb->map_size = sizeof(struct tdb_header);
        tdb->ecode = TDB_SUCCESS;
-       /* header will be read in below. */
-       tdb->header_uptodate = false;
        tdb->flags = tdb_flags;
-       tdb->log = null_log_fn;
-       tdb->log_priv = NULL;
-       tdb->khash = jenkins_hash;
-       tdb->hash_priv = NULL;
+       tdb->logfn = NULL;
        tdb->transaction = NULL;
-       /* last_zone will be set below. */
+       tdb->stats = NULL;
+       tdb->access = NULL;
+       tdb_hash_init(tdb);
        tdb_io_init(tdb);
        tdb_lock_init(tdb);
 
        while (attr) {
                switch (attr->base.attr) {
                case TDB_ATTRIBUTE_LOG:
-                       tdb->log = attr->log.log_fn;
-                       tdb->log_priv = attr->log.log_private;
+                       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;
                        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);
+                       break;
                default:
-                       tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                                "tdb_open: unknown attribute type %u\n",
-                                attr->base.attr);
-                       errno = EINVAL;
+                       tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_DEBUG_ERROR,
+                                  "tdb_open: unknown attribute type %u",
+                                  attr->base.attr);
                        goto fail;
                }
                attr = attr->base.next;
        }
 
        if ((open_flags & O_ACCMODE) == O_WRONLY) {
-               tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                        "tdb_open: can't open tdb %s write-only\n", name);
-               errno = EINVAL;
+               tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_DEBUG_ERROR,
+                          "tdb_open: can't open tdb %s write-only", name);
                goto fail;
        }
 
@@ -267,29 +215,31 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
                tdb->read_only = true;
                /* read only databases don't do locking */
                tdb->flags |= TDB_NOLOCK;
-       } else
+               tdb->mmap_flags = PROT_READ;
+       } else {
                tdb->read_only = false;
+               tdb->mmap_flags = PROT_READ | PROT_WRITE;
+       }
 
        /* internal databases don't need any of the rest. */
        if (tdb->flags & TDB_INTERNAL) {
                tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
-               if (tdb_new_database(tdb) != 0) {
-                       tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                                "tdb_open: tdb_new_database failed!");
+               if (tdb_new_database(tdb, seed, &hdr) != 0) {
                        goto fail;
                }
-               TEST_IT(tdb->flags & TDB_CONVERT);
-               tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
-               /* Zones don't matter for internal db. */
-               tdb->last_zone = 0;
+               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) {
-               tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
-                        "tdb_open: could not open file %s: %s\n",
-                        name, strerror(errno));
-               goto fail;      /* errno set by open(2) */
+               /* errno set by open(2) */
+               saved_errno = errno;
+               tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
+                          "tdb_open: could not open file %s: %s",
+                          name, strerror(errno));
+               goto fail;
        }
 
        /* on exec, don't inherit the fd */
@@ -297,82 +247,115 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
         fcntl(tdb->fd, F_SETFD, v | FD_CLOEXEC);
 
        /* ensure there is only one process initialising at once */
-       if (tdb_lock_open(tdb) == -1) {
-               tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                        "tdb_open: failed to get open lock on %s: %s\n",
-                        name, strerror(errno));
-               goto fail;      /* errno set by tdb_brlock */
+       if (tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK) == -1) {
+               /* errno set by tdb_brlock */
+               saved_errno = errno;
+               goto fail;
        }
 
-       if (!tdb_pread_all(tdb->fd, &tdb->header, sizeof(tdb->header), 0)
-           || strcmp(tdb->header.magic_food, TDB_MAGIC_FOOD) != 0) {
-               if (!(open_flags & O_CREAT) || tdb_new_database(tdb) == -1) {
-                       if (errno == 0) {
-                               errno = EIO; /* ie bad format or something */
-                       }
+       if (!tdb_pread_all(tdb->fd, &hdr, sizeof(hdr), 0)
+           || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
+               if (!(open_flags & O_CREAT)) {
+                       tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
+                                  "tdb_open: %s is not a tdb file", name);
+                       goto fail;
+               }
+               if (tdb_new_database(tdb, seed, &hdr) == -1) {
                        goto fail;
                }
-       } else if (tdb->header.version != TDB_VERSION) {
-               if (tdb->header.version == bswap_64(TDB_VERSION))
+       } else if (hdr.version != TDB_VERSION) {
+               if (hdr.version == bswap_64(TDB_VERSION))
                        tdb->flags |= TDB_CONVERT;
                else {
                        /* wrong version */
-                       tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                                "tdb_open: %s is unknown version 0x%llx\n",
-                                name, (long long)tdb->header.version);
-                       errno = EIO;
+                       tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
+                                  "tdb_open: %s is unknown version 0x%llx",
+                                  name, (long long)hdr.version);
                        goto fail;
                }
        }
 
-       tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
+       tdb_convert(tdb, &hdr, sizeof(hdr));
+       tdb->hash_seed = hdr.hash_seed;
        hash_test = TDB_HASH_MAGIC;
-       hash_test = tdb->khash(&hash_test, sizeof(hash_test),
-                              tdb->header.hash_seed, tdb->hash_priv);
-       if (tdb->header.hash_test != hash_test) {
+       hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
+       if (hdr.hash_test != hash_test) {
                /* wrong hash variant */
-               tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                        "tdb_open: %s uses a different hash function\n",
-                        name);
-               errno = EIO;
+               tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
+                          "tdb_open: %s uses a different hash function",
+                          name);
                goto fail;
        }
 
-       if (fstat(tdb->fd, &st) == -1)
+       if (fstat(tdb->fd, &st) == -1) {
+               saved_errno = errno;
+               tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_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 */
-               tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                        "tdb_open: %s (%d,%d) is already open in this process\n",
-                        name, (int)st.st_dev, (int)st.st_ino);
-               errno = EBUSY;
+               tdb_logerr(tdb, TDB_ERR_NESTING, TDB_DEBUG_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) {
-               errno = ENOMEM;
+               tdb_logerr(tdb, TDB_ERR_OOM, TDB_DEBUG_ERROR,
+                          "tdb_open: failed to allocate name");
                goto fail;
        }
 
-       tdb->map_size = st.st_size;
        tdb->device = st.st_dev;
        tdb->inode = st.st_ino;
-       tdb_mmap(tdb);
        tdb_unlock_open(tdb);
-       tdb_zone_init(tdb);
+
+       /* This make sure we have current map_size and mmap. */
+       tdb->methods->oob(tdb, tdb->map_size + 1, true);
+
+       /* Now it's fully formed, recover if necessary. */
+       if (tdb_needs_recovery(tdb) && tdb_lock_and_recover(tdb) == -1) {
+               goto fail;
+       }
+
+       if (tdb_ftable_init(tdb) == -1)
+               goto fail;
 
        tdb->next = tdbs;
        tdbs = tdb;
        return tdb;
 
  fail:
-       save_errno = errno;
-
-       if (!tdb)
-               return NULL;
+       /* Map ecode to some logical errno. */
+       if (!saved_errno) {
+               switch (tdb->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;
+               case TDB_ERR_NESTING:
+                       saved_errno = EBUSY;
+                       break;
+               default:
+                       saved_errno = EINVAL;
+                       break;
+               }
+       }
 
 #ifdef TDB_TRACE
        close(tdb->tracefd);
@@ -386,154 +369,14 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
        free((char *)tdb->name);
        if (tdb->fd != -1)
                if (close(tdb->fd) != 0)
-                       tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                                "tdb_open: failed to close tdb->fd"
-                                " on error!\n");
+                       tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
+                                  "tdb_open: failed to close tdb->fd"
+                                  " on error!");
        free(tdb);
-       errno = save_errno;
+       errno = saved_errno;
        return NULL;
 }
 
-static tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
-{
-       return tdb->header.v.hash_off
-               + ((list & ((1ULL << tdb->header.v.hash_bits) - 1))
-                  * sizeof(tdb_off_t));
-}
-
-/* Returns 0 if the entry is a zero (definitely not a match).
- * Returns a valid entry offset if it's a match.  Fills in rec.
- * Otherwise returns TDB_OFF_ERR: keep searching. */
-static tdb_off_t entry_matches(struct tdb_context *tdb,
-                              uint64_t list,
-                              uint64_t hash,
-                              const struct tdb_data *key,
-                              struct tdb_used_record *rec)
-{
-       tdb_off_t off;
-       uint64_t keylen;
-       const unsigned char *rkey;
-
-       list &= ((1ULL << tdb->header.v.hash_bits) - 1);
-
-       off = tdb_read_off(tdb, tdb->header.v.hash_off
-                          + list * sizeof(tdb_off_t));
-       if (off == 0 || off == TDB_OFF_ERR)
-               return off;
-
-#if 0 /* FIXME: Check other bits. */
-       unsigned int bits, bitmask, hoffextra;
-       /* Bottom three bits show how many extra hash bits. */
-       bits = (off & ((1 << TDB_EXTRA_HASHBITS_NUM) - 1)) + 1;
-       bitmask = (1 << bits)-1;
-       hoffextra = ((off >> TDB_EXTRA_HASHBITS_NUM) & bitmask);
-       uint64_t hextra = hash >> tdb->header.v.hash_bits;
-       if ((hextra & bitmask) != hoffextra) 
-               return TDB_OFF_ERR;
-       off &= ~...;
-#endif
-
-       if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1)
-               return TDB_OFF_ERR;
-
-       /* FIXME: check extra bits in header! */
-       keylen = rec_key_length(rec);
-       if (keylen != key->dsize)
-               return TDB_OFF_ERR;
-
-       rkey = tdb_access_read(tdb, off + sizeof(*rec), keylen, false);
-       if (!rkey)
-               return TDB_OFF_ERR;
-       if (memcmp(rkey, key->dptr, keylen) != 0)
-               off = TDB_OFF_ERR;
-       tdb_access_release(tdb, rkey);
-       return off;
-}
-
-/* FIXME: Optimize? */
-static void unlock_lists(struct tdb_context *tdb,
-                        tdb_off_t list, tdb_len_t num,
-                        int ltype)
-{
-       tdb_off_t i;
-
-       for (i = list; i < list + num; i++)
-               tdb_unlock_list(tdb, i, ltype);
-}
-
-/* FIXME: Optimize? */
-static int lock_lists(struct tdb_context *tdb,
-                     tdb_off_t list, tdb_len_t num,
-                     int ltype)
-{
-       tdb_off_t i;
-
-       for (i = list; i < list + num; i++) {
-               if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT)
-                   == TDB_OFF_ERR) {
-                       unlock_lists(tdb, list, i - list, ltype);
-                       return -1;
-               }
-       }
-       return 0;
-}
-
-/* We lock hashes up to the next empty offset.  We already hold the
- * lock on the start bucket, but we may need to release and re-grab
- * it.  If we fail, we hold no locks at all! */
-static tdb_len_t relock_hash_to_zero(struct tdb_context *tdb,
-                                    tdb_off_t start, int ltype)
-{
-       tdb_len_t num, len;
-
-again:
-       num = 1ULL << tdb->header.v.hash_bits;
-       len = tdb_find_zero_off(tdb, hash_off(tdb, start), num - start);
-       if (unlikely(len == num - start)) {
-               /* We hit the end of the hash range.  Drop lock: we have
-                  to lock start of hash first. */
-               tdb_len_t pre_locks;
-
-               tdb_unlock_list(tdb, start, ltype);
-
-               /* Grab something, so header is stable. */
-               if (tdb_lock_list(tdb, 0, ltype, TDB_LOCK_WAIT))
-                       return TDB_OFF_ERR;
-               pre_locks = tdb_find_zero_off(tdb, hash_off(tdb, 0), num);
-               /* We want to lock the zero entry as well. */
-               pre_locks++;
-               if (lock_lists(tdb, 1, pre_locks - 1, ltype) == -1) {
-                       tdb_unlock_list(tdb, 0, ltype);
-                       return TDB_OFF_ERR;
-               }
-
-               /* Now lock later ones. */
-               if (unlikely(lock_lists(tdb, start, len, ltype) == -1)) {
-                       unlock_lists(tdb, 0, pre_locks, ltype);
-                       return TDB_OFF_ERR;
-               }
-               len += pre_locks;
-       } else {
-               /* We want to lock the zero entry as well. */
-               len++;
-               /* But we already have lock on start. */
-               if (unlikely(lock_lists(tdb, start+1, len-1, ltype) == -1)) {
-                       tdb_unlock_list(tdb, start, ltype);
-                       return TDB_OFF_ERR;
-               }
-       }
-
-       /* Now, did we lose the race, and it's not zero any more? */
-       if (unlikely(tdb_read_off(tdb, hash_off(tdb, start + len - 1)) != 0)) {
-               /* Leave the start locked, as expected. */
-               unlock_lists(tdb, start + 1, len - 1, ltype);
-               goto again;
-       }
-
-       return len;
-}
-
-/* FIXME: modify, don't rewrite! */
 static int update_rec_hdr(struct tdb_context *tdb,
                          tdb_off_t off,
                          tdb_len_t keylen,
@@ -543,172 +386,63 @@ static int update_rec_hdr(struct tdb_context *tdb,
 {
        uint64_t dataroom = rec_data_length(rec) + rec_extra_padding(rec);
 
-       if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h))
+       if (set_header(tdb, rec, TDB_USED_MAGIC, keylen, datalen,
+                      keylen + dataroom, h))
                return -1;
 
        return tdb_write_convert(tdb, off, rec, sizeof(*rec));
 }
 
-static int hash_add(struct tdb_context *tdb,
-                   uint64_t hash, tdb_off_t off)
-{
-       tdb_off_t i, hoff, len, num;
-
-       /* Look for next space. */
-       i = (hash & ((1ULL << tdb->header.v.hash_bits) - 1));
-       len = (1ULL << tdb->header.v.hash_bits) - i;
-       num = tdb_find_zero_off(tdb, hash_off(tdb, i), len);
-
-       if (unlikely(num == len)) {
-               /* We wrapped.  Look through start of hash table. */
-               hoff = hash_off(tdb, 0);
-               len = (1ULL << tdb->header.v.hash_bits);
-               num = tdb_find_zero_off(tdb, hoff, len);
-               if (i == len) {
-                       tdb->ecode = TDB_ERR_CORRUPT;
-                       tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-                                "hash_add: full hash table!\n");
-                       return -1;
-               }
-       }
-       /* FIXME: Encode extra hash bits! */
-       return tdb_write_off(tdb, hash_off(tdb, i + num), off);
-}
-
-/* If we fail, others will try after us. */
-static void enlarge_hash(struct tdb_context *tdb)
+/* Returns -1 on error, 0 on OK */
+static int replace_data(struct tdb_context *tdb,
+                       struct hash_info *h,
+                       struct tdb_data key, struct tdb_data dbuf,
+                       tdb_off_t old_off, tdb_len_t old_room,
+                       bool growing)
 {
-       tdb_off_t newoff, oldoff, i;
-       tdb_len_t hlen;
-       uint64_t num = 1ULL << tdb->header.v.hash_bits;
-       struct tdb_used_record pad, *r;
-
-       /* FIXME: We should do this without holding locks throughout. */
-       if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
-               return;
-
-       /* Someone else enlarged for us?  Nothing to do. */
-       if ((1ULL << tdb->header.v.hash_bits) != num)
-               goto unlock;
-
-again:
-       /* Allocate our new array. */
-       hlen = num * sizeof(tdb_off_t) * 2;
-       newoff = alloc(tdb, 0, hlen, 0, false);
-       if (unlikely(newoff == TDB_OFF_ERR))
-               goto unlock;
-       if (unlikely(newoff == 0)) {
-               if (tdb_expand(tdb, 0, hlen, false) == -1)
-                       goto unlock;
-               goto again;
-       }
-       /* Step over record header! */
-       newoff += sizeof(struct tdb_used_record);
-
-       /* Starts all zero. */
-       if (zero_out(tdb, newoff, hlen) == -1)
-               goto unlock;
-
-       /* Update header now so we can use normal routines. */
-       oldoff = tdb->header.v.hash_off;
-
-       tdb->header.v.hash_bits++;
-       tdb->header.v.hash_off = newoff;
-
-       /* FIXME: If the space before is empty, we know this is in its ideal
-        * location.  Or steal a bit from the pointer to avoid rehash. */
-       for (i = 0; i < num; i++) {
-               tdb_off_t off;
-               off = tdb_read_off(tdb, oldoff + i * sizeof(tdb_off_t));
-               if (unlikely(off == TDB_OFF_ERR))
-                       goto oldheader;
-               if (off && hash_add(tdb, hash_record(tdb, off), off) == -1)
-                       goto oldheader;
-       }
+       tdb_off_t new_off;
 
-       /* Free up old hash. */
-       r = tdb_get(tdb, oldoff - sizeof(*r), &pad, sizeof(*r));
-       if (!r)
-               goto oldheader;
-       add_free_record(tdb, oldoff - sizeof(*r),
-                       sizeof(*r)+rec_data_length(r)+rec_extra_padding(r));
-
-       /* Now we write the modified header. */
-       write_header(tdb);
-unlock:
-       tdb_allrecord_unlock(tdb, F_WRLCK);
-       return;
-
-oldheader:
-       tdb->header.v.hash_bits--;
-       tdb->header.v.hash_off = oldoff;
-       goto unlock;
-}
-
-/* This is the core routine which searches the hashtable for an entry.
- * On error, no locks are held and TDB_OFF_ERR is returned.
- * Otherwise, *num_locks locks of type ltype from *start_lock are held.
- * The bucket where the entry is (or would be) is in *bucket.
- * If not found, the return value is 0.
- * If found, the return value is the offset, and *rec is the record. */
-static tdb_off_t find_and_lock(struct tdb_context *tdb,
-                              struct tdb_data key,
-                              uint64_t h,
-                              int ltype,
-                              tdb_off_t *start_lock,
-                              tdb_len_t *num_locks,
-                              tdb_off_t *bucket,
-                              struct tdb_used_record *rec)
-{
-       tdb_off_t off;
+       /* Allocate a new record. */
+       new_off = alloc(tdb, key.dsize, dbuf.dsize, h->h, TDB_USED_MAGIC,
+                       growing);
+       if (unlikely(new_off == TDB_OFF_ERR))
+               return -1;
 
-       /* FIXME: can we avoid locks for some fast paths? */
-       *start_lock = tdb_lock_list(tdb, h, ltype, TDB_LOCK_WAIT);
-       if (*start_lock == TDB_OFF_ERR)
-               return TDB_OFF_ERR;
-
-       /* Fast path. */
-       off = entry_matches(tdb, *start_lock, h, &key, rec);
-       if (likely(off != TDB_OFF_ERR)) {
-               *bucket = *start_lock;
-               *num_locks = 1;
-               return off;
+       /* We didn't like the existing one: remove it. */
+       if (old_off) {
+               add_stat(tdb, frees, 1);
+               add_free_record(tdb, old_off,
+                               sizeof(struct tdb_used_record)
+                               + key.dsize + old_room);
+               if (replace_in_hash(tdb, h, new_off) == -1)
+                       return -1;
+       } else {
+               if (add_to_hash(tdb, h, new_off) == -1)
+                       return -1;
        }
 
-       /* Slow path, need to grab more locks and search. */
-
-       /* Warning: this may drop the lock on *bucket! */
-       *num_locks = relock_hash_to_zero(tdb, *start_lock, ltype);
-       if (*num_locks == TDB_OFF_ERR)
-               return TDB_OFF_ERR;
+       new_off += sizeof(struct tdb_used_record);
+       if (tdb->methods->write(tdb, new_off, key.dptr, key.dsize) == -1)
+               return -1;
 
-       for (*bucket = *start_lock;
-            *bucket < *start_lock + *num_locks;
-            (*bucket)++) {
-               off = entry_matches(tdb, *bucket, h, &key, rec);
-               /* Empty entry or we found it? */
-               if (off == 0 || off != TDB_OFF_ERR)
-                       return off;
-       }
+       new_off += key.dsize;
+       if (tdb->methods->write(tdb, new_off, dbuf.dptr, dbuf.dsize) == -1)
+               return -1;
 
-       /* We didn't find a zero entry?  Something went badly wrong... */
-       unlock_lists(tdb, *start_lock, *start_lock + *num_locks, ltype);
-       tdb->ecode = TDB_ERR_CORRUPT;
-       tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-                "find_and_lock: expected to find an empty hash bucket!\n");
-       return TDB_OFF_ERR;
+       /* FIXME: tdb_increment_seqnum(tdb); */
+       return 0;
 }
 
 int tdb_store(struct tdb_context *tdb,
              struct tdb_data key, struct tdb_data dbuf, int flag)
 {
-       tdb_off_t new_off, off, bucket, start, num;
+       struct hash_info h;
+       tdb_off_t off;
+       tdb_len_t old_room = 0;
        struct tdb_used_record rec;
-       uint64_t h;
-       bool growing = false;
+       int ret;
 
-       h = tdb_hash(tdb, key.dptr, key.dsize);
-       off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
+       off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
        if (unlikely(off == TDB_OFF_ERR))
                return -1;
 
@@ -720,18 +454,22 @@ int tdb_store(struct tdb_context *tdb,
                }
        } else {
                if (off) {
-                       if (rec_data_length(&rec) + rec_extra_padding(&rec)
-                           >= dbuf.dsize) {
-                               new_off = off;
+                       old_room = rec_data_length(&rec)
+                               + rec_extra_padding(&rec);
+                       if (old_room >= dbuf.dsize) {
+                               /* Can modify in-place.  Easy! */
                                if (update_rec_hdr(tdb, off,
                                                   key.dsize, dbuf.dsize,
-                                                  &rec, h))
+                                                  &rec, h.h))
                                        goto fail;
-                               goto write;
+                               if (tdb->methods->write(tdb, off + sizeof(rec)
+                                                       + key.dsize,
+                                                       dbuf.dptr, dbuf.dsize))
+                                       goto fail;
+                               tdb_unlock_hashes(tdb, h.hlock_start,
+                                                 h.hlock_range, F_WRLCK);
+                               return 0;
                        }
-                       /* FIXME: See if right record is free? */
-                       /* Hint to allocator that we've realloced. */
-                       growing = true;
                } else {
                        if (flag == TDB_MODIFY) {
                                /* if the record doesn't exist and we
@@ -743,60 +481,93 @@ int tdb_store(struct tdb_context *tdb,
                }
        }
 
-       /* Allocate a new record. */
-       new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
-       if (new_off == 0) {
-               unlock_lists(tdb, start, num, F_WRLCK);
-               /* Expand, then try again... */
-               if (tdb_expand(tdb, key.dsize, dbuf.dsize, growing) == -1)
-                       return -1;
-               return tdb_store(tdb, key, dbuf, flag);
-       }
+       /* If we didn't use the old record, this implies we're growing. */
+       ret = replace_data(tdb, &h, key, dbuf, off, old_room, off != 0);
+       tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
+       return ret;
+
+fail:
+       tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
+       return -1;
+}
+
+int tdb_append(struct tdb_context *tdb,
+              struct tdb_data key, struct tdb_data dbuf)
+{
+       struct hash_info h;
+       tdb_off_t off;
+       struct tdb_used_record rec;
+       tdb_len_t old_room = 0, old_dlen;
+       unsigned char *newdata;
+       struct tdb_data new_dbuf;
+       int ret;
+
+       off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
+       if (unlikely(off == TDB_OFF_ERR))
+               return -1;
 
-       /* We didn't like the existing one: remove it. */
        if (off) {
-               add_free_record(tdb, off, sizeof(struct tdb_used_record)
-                               + rec_key_length(&rec)
-                               + rec_data_length(&rec)
-                               + rec_extra_padding(&rec));
-       }
+               old_dlen = rec_data_length(&rec);
+               old_room = old_dlen + rec_extra_padding(&rec);
 
-       /* FIXME: Encode extra hash bits! */
-       if (tdb_write_off(tdb, hash_off(tdb, bucket), new_off) == -1)
-               goto fail;
+               /* Fast path: can append in place. */
+               if (rec_extra_padding(&rec) >= dbuf.dsize) {
+                       if (update_rec_hdr(tdb, off, key.dsize,
+                                          old_dlen + dbuf.dsize, &rec, h.h))
+                               goto fail;
 
-write:
-       off = new_off + sizeof(struct tdb_used_record);
-       if (tdb->methods->write(tdb, off, key.dptr, key.dsize) == -1)
-               goto fail;
-       off += key.dsize;
-       if (tdb->methods->write(tdb, off, dbuf.dptr, dbuf.dsize) == -1)
-               goto fail;
+                       off += sizeof(rec) + key.dsize + old_dlen;
+                       if (tdb->methods->write(tdb, off, dbuf.dptr,
+                                               dbuf.dsize) == -1)
+                               goto fail;
 
-       /* FIXME: tdb_increment_seqnum(tdb); */
-       unlock_lists(tdb, start, num, F_WRLCK);
+                       /* FIXME: tdb_increment_seqnum(tdb); */
+                       tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range,
+                                         F_WRLCK);
+                       return 0;
+               }
+
+               /* Slow path. */
+               newdata = malloc(key.dsize + old_dlen + dbuf.dsize);
+               if (!newdata) {
+                       tdb_logerr(tdb, TDB_ERR_OOM, TDB_DEBUG_FATAL,
+                                  "tdb_append: failed to allocate %zu bytes",
+                                  (size_t)(key.dsize+old_dlen+dbuf.dsize));
+                       goto fail;
+               }
+               if (tdb->methods->read(tdb, off + sizeof(rec) + key.dsize,
+                                      newdata, old_dlen) != 0) {
+                       free(newdata);
+                       goto fail;
+               }
+               memcpy(newdata + old_dlen, dbuf.dptr, dbuf.dsize);
+               new_dbuf.dptr = newdata;
+               new_dbuf.dsize = old_dlen + dbuf.dsize;
+       } else {
+               newdata = NULL;
+               new_dbuf = dbuf;
+       }
 
-       /* FIXME: by simple simulation, this approximated 60% full.
-        * Check in real case! */
-       if (unlikely(num > 4 * tdb->header.v.hash_bits - 30))
-               enlarge_hash(tdb);
+       /* If they're using tdb_append(), it implies they're growing record. */
+       ret = replace_data(tdb, &h, key, new_dbuf, off, old_room, true);
+       tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
+       free(newdata);
 
-       return 0;
+       return ret;
 
 fail:
-       unlock_lists(tdb, start, num, F_WRLCK);
+       tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
        return -1;
 }
 
 struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
 {
-       tdb_off_t off, start, num, bucket;
+       tdb_off_t off;
        struct tdb_used_record rec;
-       uint64_t h;
+       struct hash_info h;
        struct tdb_data ret;
 
-       h = tdb_hash(tdb, key.dptr, key.dsize);
-       off = find_and_lock(tdb, key, h, F_RDLCK, &start, &num, &bucket, &rec);
+       off = find_and_lock(tdb, key, F_RDLCK, &h, &rec, NULL);
        if (unlikely(off == TDB_OFF_ERR))
                return tdb_null;
 
@@ -809,62 +580,31 @@ struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
                                          ret.dsize);
        }
 
-       unlock_lists(tdb, start, num, F_RDLCK);
+       tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_RDLCK);
        return ret;
 }
 
 int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
 {
-       tdb_off_t i, bucket, off, start, num;
+       tdb_off_t off;
        struct tdb_used_record rec;
-       uint64_t h;
+       struct hash_info h;
 
-       h = tdb_hash(tdb, key.dptr, key.dsize);
-       off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
+       off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
        if (unlikely(off == TDB_OFF_ERR))
                return -1;
 
        if (!off) {
-               unlock_lists(tdb, start, num, F_WRLCK);
+               tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
                tdb->ecode = TDB_ERR_NOEXIST;
                return -1;
        }
 
-       /* This actually unlinks it. */
-       if (tdb_write_off(tdb, hash_off(tdb, bucket), 0) == -1)
+       if (delete_from_hash(tdb, &h) == -1)
                goto unlock_err;
 
-       /* Rehash anything following. */
-       for (i = hash_off(tdb, bucket+1);
-            i != hash_off(tdb, h + num - 1);
-            i += sizeof(tdb_off_t)) {
-               tdb_off_t off2;
-               uint64_t h2;
-
-               off2 = tdb_read_off(tdb, i);
-               if (unlikely(off2 == TDB_OFF_ERR))
-                       goto unlock_err;
-
-               /* This can happen if we raced. */
-               if (unlikely(off2 == 0))
-                       break;
-
-               /* Maybe use a bit to indicate it is in ideal place? */
-               h2 = hash_record(tdb, off2);
-               /* Is it happy where it is? */
-               if (hash_off(tdb, h2) == i)
-                       continue;
-
-               /* Remove it. */
-               if (tdb_write_off(tdb, i, 0) == -1)
-                       goto unlock_err;
-
-               /* Rehash it. */
-               if (hash_add(tdb, h2, off2) == -1)
-                       goto unlock_err;
-       }
-
        /* Free the deleted entry. */
+       add_stat(tdb, frees, 1);
        if (add_free_record(tdb, off,
                            sizeof(struct tdb_used_record)
                            + rec_key_length(&rec)
@@ -872,11 +612,11 @@ int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
                            + rec_extra_padding(&rec)) != 0)
                goto unlock_err;
 
-       unlock_lists(tdb, start, num, F_WRLCK);
+       tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
        return 0;
 
 unlock_err:
-       unlock_lists(tdb, start, num, F_WRLCK);
+       tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
        return -1;
 }
 
@@ -885,12 +625,11 @@ int tdb_close(struct tdb_context *tdb)
        struct tdb_context **i;
        int ret = 0;
 
-       /* FIXME:
+       tdb_trace(tdb, "tdb_close");
+
        if (tdb->transaction) {
                tdb_transaction_cancel(tdb);
        }
-       */
-       tdb_trace(tdb, "tdb_close");
 
        if (tdb->map_ptr) {
                if (tdb->flags & TDB_INTERNAL)
@@ -921,7 +660,60 @@ int tdb_close(struct tdb_context *tdb)
        return ret;
 }
 
-enum TDB_ERROR tdb_error(struct tdb_context *tdb)
+enum TDB_ERROR tdb_error(const struct tdb_context *tdb)
 {
        return tdb->ecode;
 }
+
+const char *tdb_errorstr(const struct tdb_context *tdb)
+{
+       /* Gcc warns if you miss a case in the switch, so use that. */
+       switch (tdb->ecode) {
+       case TDB_SUCCESS: return "Success";
+       case TDB_ERR_CORRUPT: return "Corrupt database";
+       case TDB_ERR_IO: return "IO Error";
+       case TDB_ERR_LOCK: return "Locking error";
+       case TDB_ERR_OOM: return "Out of memory";
+       case TDB_ERR_EXISTS: return "Record exists";
+       case TDB_ERR_NESTING: return "Transaction already started";
+       case TDB_ERR_EINVAL: return "Invalid parameter";
+       case TDB_ERR_NOEXIST: return "Record does not exist";
+       case TDB_ERR_RDONLY: return "write not permitted";
+       }
+       return "Invalid error code";
+}
+
+void COLD tdb_logerr(struct tdb_context *tdb,
+                    enum TDB_ERROR ecode,
+                    enum tdb_debug_level level,
+                    const char *fmt, ...)
+{
+       char *message;
+       va_list ap;
+       size_t len;
+       /* tdb_open paths care about errno, so save it. */
+       int saved_errno = errno;
+
+       tdb->ecode = ecode;
+
+       if (!tdb->logfn)
+               return;
+
+       /* FIXME: Doesn't assume asprintf. */
+       va_start(ap, fmt);
+       len = vsnprintf(NULL, 0, fmt, ap);
+       va_end(ap);
+
+       message = malloc(len + 1);
+       if (!message) {
+               tdb->logfn(tdb, level, tdb->log_private,
+                          "out of memory formatting message");
+               return;
+       }
+       va_start(ap, fmt);
+       len = vsprintf(message, fmt, ap);
+       va_end(ap);
+       tdb->logfn(tdb, level, tdb->log_private, message);
+       free(message);
+       errno = saved_errno;
+}