]> git.ozlabs.org Git - ccan/blobdiff - ccan/tdb2/transaction.c
tdb2: reduce transaction before writing to recovery area.
[ccan] / ccan / tdb2 / transaction.c
index 66565d67c2d762b67cba9169ee3d9f8f661b78d3..4bdc3f32d019b6abe6e9519be298b027a7b804e6 100644 (file)
 
   - 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)
+    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
+    existing transaction record. If the inner transaction is canceled
     then a subsequent commit will fail
 
   - keep a mirrored copy of the tdb hash chain heads to allow for the
@@ -68,7 +68,7 @@
 
   - 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
+    gained until the transaction is committed or canceled
 
   - the commit stategy involves first saving away all modified data
     into a linearised buffer in the transaction recovery area, then
@@ -110,7 +110,7 @@ struct tdb_transaction {
        /* 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;
+       unsigned int nesting;
 
        /* set when a prepare has already occurred */
        bool prepared;
@@ -120,6 +120,8 @@ struct tdb_transaction {
        tdb_len_t old_map_size;
 };
 
+/* This doesn't really need to be pagesize, but we use it for similar reasons. */
+#define PAGESIZE 4096
 
 /*
   read while in a transaction. We need to check first if the data is in our list
@@ -132,8 +134,8 @@ static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
        enum TDB_ERROR ecode;
 
        /* break it down into block sized ops */
-       while (len + (off % getpagesize()) > getpagesize()) {
-               tdb_len_t len2 = getpagesize() - (off % getpagesize());
+       while (len + (off % PAGESIZE) > PAGESIZE) {
+               tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
                ecode = transaction_read(tdb, off, buf, len2);
                if (ecode != TDB_SUCCESS) {
                        return ecode;
@@ -147,7 +149,7 @@ static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
                return TDB_SUCCESS;
        }
 
-       blk = off / getpagesize();
+       blk = off / PAGESIZE;
 
        /* see if we have it in the block list */
        if (tdb->transaction->num_blocks <= blk ||
@@ -169,7 +171,7 @@ static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
        }
 
        /* now copy it out of this block */
-       memcpy(buf, tdb->transaction->blocks[blk] + (off % getpagesize()), len);
+       memcpy(buf, tdb->transaction->blocks[blk] + (off % PAGESIZE), len);
        return TDB_SUCCESS;
 
 fail:
@@ -198,8 +200,8 @@ static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
        }
 
        /* break it up into block sized chunks */
-       while (len + (off % getpagesize()) > getpagesize()) {
-               tdb_len_t len2 = getpagesize() - (off % getpagesize());
+       while (len + (off % PAGESIZE) > PAGESIZE) {
+               tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
                ecode = transaction_write(tdb, off, buf, len2);
                if (ecode != TDB_SUCCESS) {
                        return -1;
@@ -215,8 +217,8 @@ static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
                return TDB_SUCCESS;
        }
 
-       blk = off / getpagesize();
-       off = off % getpagesize();
+       blk = off / PAGESIZE;
+       off = off % PAGESIZE;
 
        if (tdb->transaction->num_blocks <= blk) {
                uint8_t **new_blocks;
@@ -244,20 +246,20 @@ static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
 
        /* allocate and fill a block? */
        if (tdb->transaction->blocks[blk] == NULL) {
-               tdb->transaction->blocks[blk] = (uint8_t *)calloc(getpagesize(), 1);
+               tdb->transaction->blocks[blk] = (uint8_t *)calloc(PAGESIZE, 1);
                if (tdb->transaction->blocks[blk] == NULL) {
                        ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
                                           "transaction_write:"
                                           " failed to allocate");
                        goto fail;
                }
-               if (tdb->transaction->old_map_size > blk * getpagesize()) {
-                       tdb_len_t len2 = getpagesize();
-                       if (len2 + (blk * getpagesize()) > tdb->transaction->old_map_size) {
-                               len2 = tdb->transaction->old_map_size - (blk * getpagesize());
+               if (tdb->transaction->old_map_size > blk * PAGESIZE) {
+                       tdb_len_t len2 = PAGESIZE;
+                       if (len2 + (blk * PAGESIZE) > tdb->transaction->old_map_size) {
+                               len2 = tdb->transaction->old_map_size - (blk * PAGESIZE);
                        }
                        ecode = tdb->transaction->io_methods->tread(tdb,
-                                       blk * getpagesize(),
+                                       blk * PAGESIZE,
                                        tdb->transaction->blocks[blk],
                                        len2);
                        if (ecode != TDB_SUCCESS) {
@@ -297,7 +299,7 @@ fail:
 
 
 /*
-  write while in a transaction - this varient never expands the transaction blocks, it only
+  write while in a transaction - this variant never expands the transaction blocks, it only
   updates existing blocks. This means it cannot change the recovery size
 */
 static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
@@ -306,8 +308,8 @@ static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
        size_t blk;
 
        /* break it up into block sized chunks */
-       while (len + (off % getpagesize()) > getpagesize()) {
-               tdb_len_t len2 = getpagesize() - (off % getpagesize());
+       while (len + (off % PAGESIZE) > PAGESIZE) {
+               tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
                transaction_write_existing(tdb, off, buf, len2);
                len -= len2;
                off += len2;
@@ -320,8 +322,8 @@ static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
                return;
        }
 
-       blk = off / getpagesize();
-       off = off % getpagesize();
+       blk = off / PAGESIZE;
+       off = off % PAGESIZE;
 
        if (tdb->transaction->num_blocks <= blk ||
            tdb->transaction->blocks[blk] == NULL) {
@@ -347,14 +349,14 @@ static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
 static enum TDB_ERROR transaction_oob(struct tdb_context *tdb, tdb_off_t len,
                                      bool probe)
 {
-       if (len <= tdb->map_size) {
+       if (len <= tdb->file->map_size) {
                return TDB_SUCCESS;
        }
        if (!probe) {
                tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
                           "tdb_oob len %lld beyond transaction size %lld",
                           (long long)len,
-                          (long long)tdb->map_size);
+                          (long long)tdb->file->map_size);
        }
        return TDB_ERR_IO;
 }
@@ -369,9 +371,9 @@ static enum TDB_ERROR transaction_expand_file(struct tdb_context *tdb,
 
        /* add a write to the transaction elements, so subsequent
           reads see the zero data */
-       ecode = transaction_write(tdb, tdb->map_size, NULL, addition);
+       ecode = transaction_write(tdb, tdb->file->map_size, NULL, addition);
        if (ecode == TDB_SUCCESS) {
-               tdb->map_size += addition;
+               tdb->file->map_size += addition;
        }
        return ecode;
 }
@@ -379,10 +381,10 @@ static enum TDB_ERROR transaction_expand_file(struct tdb_context *tdb,
 static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
                                size_t len, bool write_mode)
 {
-       size_t blk = off / getpagesize(), end_blk;
+       size_t blk = off / PAGESIZE, end_blk;
 
        /* This is wrong for zero-length blocks, but will fail gracefully */
-       end_blk = (off + len - 1) / getpagesize();
+       end_blk = (off + len - 1) / PAGESIZE;
 
        /* Can only do direct if in single block and we've already copied. */
        if (write_mode) {
@@ -392,17 +394,17 @@ static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
                        return NULL;
                if (tdb->transaction->blocks[blk] == NULL)
                        return NULL;
-               return tdb->transaction->blocks[blk] + off % getpagesize();
+               return tdb->transaction->blocks[blk] + off % PAGESIZE;
        }
 
        /* Single which we have copied? */
        if (blk == end_blk
            && blk < tdb->transaction->num_blocks
            && tdb->transaction->blocks[blk])
-               return tdb->transaction->blocks[blk] + off % getpagesize();
+               return tdb->transaction->blocks[blk] + off % PAGESIZE;
 
        /* Otherwise must be all not copied. */
-       while (blk < end_blk) {
+       while (blk <= end_blk) {
                if (blk >= tdb->transaction->num_blocks)
                        break;
                if (tdb->transaction->blocks[blk])
@@ -430,15 +432,15 @@ static enum TDB_ERROR transaction_sync(struct tdb_context *tdb,
                return TDB_SUCCESS;
        }
 
-       if (fsync(tdb->fd) != 0) {
+       if (fsync(tdb->file->fd) != 0) {
                return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
                                  "tdb_transaction: fsync failed: %s",
                                  strerror(errno));
        }
 #ifdef MS_SYNC
-       if (tdb->map_ptr) {
-               tdb_off_t moffset = offset & ~(getpagesize()-1);
-               if (msync(moffset + (char *)tdb->map_ptr,
+       if (tdb->file->map_ptr) {
+               tdb_off_t moffset = offset & ~(PAGESIZE-1);
+               if (msync(moffset + (char *)tdb->file->map_ptr,
                          length + (offset - moffset), MS_SYNC) != 0) {
                        return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
                                          "tdb_transaction: msync failed: %s",
@@ -467,7 +469,7 @@ static void _tdb_transaction_cancel(struct tdb_context *tdb)
                return;
        }
 
-       tdb->map_size = tdb->transaction->old_map_size;
+       tdb->file->map_size = tdb->transaction->old_map_size;
 
        /* free all the transaction blocks */
        for (i=0;i<tdb->transaction->num_blocks;i++) {
@@ -495,8 +497,8 @@ static void _tdb_transaction_cancel(struct tdb_context *tdb)
                }
        }
 
-       if (tdb->allrecord_lock.count)
-               tdb_allrecord_unlock(tdb, tdb->allrecord_lock.ltype);
+       if (tdb->file->allrecord_lock.count)
+               tdb_allrecord_unlock(tdb, tdb->file->allrecord_lock.ltype);
 
        /* restore the normal io methods */
        tdb->methods = tdb->transaction->io_methods;
@@ -513,42 +515,52 @@ static void _tdb_transaction_cancel(struct tdb_context *tdb)
   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)
+enum TDB_ERROR tdb_transaction_start(struct tdb_context *tdb)
 {
        enum TDB_ERROR ecode;
 
        /* some sanity checks */
        if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
-               tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
-                          "tdb_transaction_start: cannot start a transaction"
-                          " on a read-only or internal db");
-               return -1;
+               return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
+                                                   TDB_LOG_USE_ERROR,
+                                                   "tdb_transaction_start:"
+                                                   " cannot start a"
+                                                   " transaction on a "
+                                                   "read-only or internal db");
        }
 
        /* cope with nested tdb_transaction_start() calls */
        if (tdb->transaction != NULL) {
-               tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_USE_ERROR,
-                          "tdb_transaction_start:"
-                          " already inside transaction");
-               return -1;
+               if (!(tdb->flags & TDB_ALLOW_NESTING)) {
+                       return tdb->last_error
+                               = tdb_logerr(tdb, TDB_ERR_IO,
+                                            TDB_LOG_USE_ERROR,
+                                            "tdb_transaction_start:"
+                                            " already inside transaction");
+               }
+               tdb->transaction->nesting++;
+               return 0;
        }
 
        if (tdb_has_hash_locks(tdb)) {
                /* 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_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
-                          "tdb_transaction_start: cannot start a transaction"
-                          " with locks held");
-               return -1;
+                  of nested locks in POSIX */
+               return tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
+                                                   TDB_LOG_USE_ERROR,
+                                                   "tdb_transaction_start:"
+                                                   " cannot start a"
+                                                   " transaction with locks"
+                                                   " held");
        }
 
        tdb->transaction = (struct tdb_transaction *)
                calloc(sizeof(struct tdb_transaction), 1);
        if (tdb->transaction == NULL) {
-               tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
-                          "tdb_transaction_start: cannot allocate");
-               return -1;
+               return tdb->last_error = tdb_logerr(tdb, TDB_ERR_OOM,
+                                                   TDB_LOG_ERROR,
+                                                   "tdb_transaction_start:"
+                                                   " cannot allocate");
        }
 
        /* get the transaction write lock. This is a blocking lock. As
@@ -556,36 +568,34 @@ int tdb_transaction_start(struct tdb_context *tdb)
           make this async, which we will probably do in the future */
        ecode = tdb_transaction_lock(tdb, F_WRLCK);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
                SAFE_FREE(tdb->transaction->blocks);
                SAFE_FREE(tdb->transaction);
-               return -1;
+               return tdb->last_error = ecode;
        }
 
        /* get a read lock over entire file. This is upgraded to a write
           lock during the commit */
        ecode = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
                goto fail_allrecord_lock;
        }
 
        /* make sure we know about any file expansions already done by
           anyone else */
-       tdb->methods->oob(tdb, tdb->map_size + 1, true);
-       tdb->transaction->old_map_size = tdb->map_size;
+       tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
+       tdb->transaction->old_map_size = tdb->file->map_size;
 
        /* finally hook the io methods, replacing them with
           transaction specific methods */
        tdb->transaction->io_methods = tdb->methods;
        tdb->methods = &transaction_methods;
-       return 0;
+       return tdb->last_error = TDB_SUCCESS;
 
 fail_allrecord_lock:
        tdb_transaction_unlock(tdb, F_WRLCK);
        SAFE_FREE(tdb->transaction->blocks);
        SAFE_FREE(tdb->transaction);
-       return -1;
+       return tdb->last_error = ecode;
 }
 
 
@@ -605,9 +615,9 @@ static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
        tdb_len_t recovery_size = 0;
        int i;
 
-       recovery_size = sizeof(tdb_len_t);
+       recovery_size = 0;
        for (i=0;i<tdb->transaction->num_blocks;i++) {
-               if (i * getpagesize() >= tdb->transaction->old_map_size) {
+               if (i * PAGESIZE >= tdb->transaction->old_map_size) {
                        break;
                }
                if (tdb->transaction->blocks[i] == NULL) {
@@ -617,7 +627,7 @@ static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
                if (i == tdb->transaction->num_blocks-1) {
                        recovery_size += tdb->transaction->last_block_size;
                } else {
-                       recovery_size += getpagesize();
+                       recovery_size += PAGESIZE;
                }
        }
 
@@ -676,9 +686,10 @@ static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
           us an area that is being currently used (as of the start of
           the transaction) */
        if (recovery_head != 0) {
-               add_stat(tdb, frees, 1);
+               tdb->stats.frees++;
                ecode = add_free_record(tdb, recovery_head,
-                                       sizeof(rec) + rec.max_len);
+                                       sizeof(rec) + rec.max_len,
+                                       TDB_LOCK_WAIT, true);
                if (ecode != TDB_SUCCESS) {
                        return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
                                          "tdb_recovery_allocate:"
@@ -690,21 +701,22 @@ static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
        /* the tdb_free() call might have increased the recovery size */
        *recovery_size = tdb_recovery_size(tdb);
 
-       /* round up to a multiple of page size */
+       /* round up to a multiple of page size. Overallocate, since each
+        * such allocation forces us to expand the file. */
        *recovery_max_size
-               = (((sizeof(rec) + *recovery_size) + getpagesize()-1)
-                  & ~(getpagesize()-1))
+               = (((sizeof(rec) + *recovery_size + *recovery_size / 2)
+                   + PAGESIZE-1) & ~(PAGESIZE-1))
                - sizeof(rec);
-       *recovery_offset = tdb->map_size;
+       *recovery_offset = tdb->file->map_size;
        recovery_head = *recovery_offset;
 
        /* Restore ->map_size before calling underlying expand_file.
           Also so that we don't try to expand the file again in the
           transaction commit, which would destroy the recovery
           area */
-       addition = (tdb->map_size - tdb->transaction->old_map_size) +
+       addition = (tdb->file->map_size - tdb->transaction->old_map_size) +
                sizeof(rec) + *recovery_max_size;
-       tdb->map_size = tdb->transaction->old_map_size;
+       tdb->file->map_size = tdb->transaction->old_map_size;
        ecode = methods->expand_file(tdb, addition);
        if (ecode != TDB_SUCCESS) {
                return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
@@ -715,7 +727,7 @@ static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
        /* 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;
+       tdb->transaction->old_map_size = tdb->file->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 */
@@ -745,19 +757,58 @@ static void set_recovery_header(struct tdb_recovery_record *rec,
        rec->eof = oldsize;
 }
 
+static unsigned int same(const unsigned char *new,
+                        const unsigned char *old,
+                        unsigned int length)
+{
+       unsigned int i;
+
+       for (i = 0; i < length; i++) {
+               if (new[i] != old[i])
+                       break;
+       }
+       return i;
+}
+
+static unsigned int different(const unsigned char *new,
+                             const unsigned char *old,
+                             unsigned int length,
+                             unsigned int min_same,
+                             unsigned int *samelen)
+{
+       unsigned int i;
+
+       *samelen = 0;
+       for (i = 0; i < length; i++) {
+               if (new[i] == old[i]) {
+                       (*samelen)++;
+               } else {
+                       if (*samelen >= min_same) {
+                               return i - *samelen;
+                       }
+                       *samelen = 0;
+               }
+       }
+
+       if (*samelen < min_same)
+               *samelen = 0;
+       return length - *samelen;
+}
+
 /*
   setup the recovery data that will be used on a crash during commit
 */
 static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
                                                 tdb_off_t *magic_offset)
 {
-       tdb_len_t recovery_size;
+       /* Initialized for GCC's 4.4.5 overzealous uninitialized warnings. */
+       tdb_len_t recovery_size = 0;
+       tdb_off_t recovery_offset = 0, recovery_max_size = 0;
        unsigned char *data, *p;
        const struct tdb_methods *methods = tdb->transaction->io_methods;
        struct tdb_recovery_record *rec;
-       tdb_off_t recovery_offset, recovery_max_size;
        tdb_off_t old_map_size = tdb->transaction->old_map_size;
-       uint64_t magic, tailer;
+       uint64_t magic;
        int i;
        enum TDB_ERROR ecode;
 
@@ -778,9 +829,6 @@ static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
        }
 
        rec = (struct tdb_recovery_record *)data;
-       set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
-                           recovery_size, recovery_max_size, old_map_size);
-       tdb_convert(tdb, rec, sizeof(*rec));
 
        /* build the recovery data into a single blob to allow us to do a single
           large write, which should be more efficient */
@@ -788,13 +836,15 @@ static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
        for (i=0;i<tdb->transaction->num_blocks;i++) {
                tdb_off_t offset;
                tdb_len_t length;
+               unsigned int off;
+               unsigned char buffer[PAGESIZE];
 
                if (tdb->transaction->blocks[i] == NULL) {
                        continue;
                }
 
-               offset = i * getpagesize();
-               length = getpagesize();
+               offset = i * PAGESIZE;
+               length = PAGESIZE;
                if (i == tdb->transaction->num_blocks-1) {
                        length = tdb->transaction->last_block_size;
                }
@@ -802,52 +852,68 @@ static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
                if (offset >= old_map_size) {
                        continue;
                }
-               if (offset + length > tdb->map_size) {
+
+               if (offset + length > tdb->file->map_size) {
                        free(data);
                        return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
                                          "tdb_transaction_setup_recovery:"
                                          " transaction data over new region"
                                          " boundary");
                }
-               memcpy(p, &offset, sizeof(offset));
-               memcpy(p + sizeof(offset), &length, sizeof(length));
-               tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
-
-               /* the recovery area contains the old data, not the
-                  new data, so we have to call the original tdb_read
-                  method to get it */
-               ecode = methods->tread(tdb, offset,
-                                      p + sizeof(offset) + sizeof(length),
-                                      length);
+               if (offset + length > old_map_size) {
+                       /* Short read at EOF. */
+                       length = old_map_size - offset;
+               }
+               ecode = methods->tread(tdb, offset, buffer, length);
                if (ecode != TDB_SUCCESS) {
                        free(data);
                        return ecode;
                }
-               p += sizeof(offset) + sizeof(length) + length;
+
+               /* Skip over anything the same at the start. */
+               off = same(tdb->transaction->blocks[i], buffer, length);
+               offset += off;
+
+               while (off < length) {
+                       tdb_len_t len;
+                       unsigned int samelen;
+
+                       len = different(tdb->transaction->blocks[i] + off,
+                                       buffer + off, length - off,
+                                       sizeof(offset) + sizeof(len) + 1,
+                                       &samelen);
+
+                       memcpy(p, &offset, sizeof(offset));
+                       memcpy(p + sizeof(offset), &len, sizeof(len));
+                       tdb_convert(tdb, p, sizeof(offset) + sizeof(len));
+                       p += sizeof(offset) + sizeof(len);
+                       memcpy(p, buffer + off, len);
+                       p += len;
+                       off += len + samelen;
+                       offset += len + samelen;
+               }
        }
 
-       /* and the tailer */
-       tailer = sizeof(*rec) + recovery_max_size;
-       memcpy(p, &tailer, sizeof(tailer));
-       tdb_convert(tdb, p, sizeof(tailer));
+       /* Now we know size, set up rec header. */
+       set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
+                           p - data - sizeof(*rec),
+                           recovery_max_size, old_map_size);
+       tdb_convert(tdb, rec, sizeof(*rec));
 
        /* write the recovery data to the recovery area */
-       ecode = methods->twrite(tdb, recovery_offset, data,
-                               sizeof(*rec) + recovery_size);
+       ecode = methods->twrite(tdb, recovery_offset, data, p - data);
        if (ecode != TDB_SUCCESS) {
                free(data);
                return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
                                  "tdb_transaction_setup_recovery:"
                                  " failed to write recovery data");
        }
-       transaction_write_existing(tdb, recovery_offset, data,
-                                  sizeof(*rec) + recovery_size);
+       transaction_write_existing(tdb, recovery_offset, data, p - data);
 
        /* 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 */
-       ecode = transaction_sync(tdb, recovery_offset,
-                                sizeof(*rec) + recovery_size);
+       ecode = transaction_sync(tdb, recovery_offset, p - data);
        if (ecode != TDB_SUCCESS) {
                free(data);
                return ecode;
@@ -900,7 +966,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 
 
        if (tdb->transaction->nesting != 0) {
-               tdb->transaction->nesting--;
                return TDB_SUCCESS;
        }
 
@@ -914,10 +979,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
        /* upgrade the main transaction lock region to a write lock */
        ecode = tdb_allrecord_upgrade(tdb);
        if (ecode != TDB_SUCCESS) {
-               tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                        "tdb_transaction_prepare_commit:"
-                        " failed to upgrade hash locks");
-               _tdb_transaction_cancel(tdb);
                return ecode;
        }
 
@@ -925,10 +986,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
           during the commit */
        ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
        if (ecode != TDB_SUCCESS) {
-               tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                          "tdb_transaction_prepare_commit:"
-                          " failed to get open lock");
-               _tdb_transaction_cancel(tdb);
                return ecode;
        }
 
@@ -939,10 +996,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
                                                   &tdb->transaction
                                                   ->magic_offset);
                if (ecode != TDB_SUCCESS) {
-                       tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                                "tdb_transaction_prepare_commit:"
-                                " failed to setup recovery data");
-                       _tdb_transaction_cancel(tdb);
                        return ecode;
                }
        }
@@ -950,16 +1003,14 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
        tdb->transaction->prepared = true;
 
        /* expand the file to the new size if needed */
-       if (tdb->map_size != tdb->transaction->old_map_size) {
-               tdb_len_t add = tdb->map_size - tdb->transaction->old_map_size;
+       if (tdb->file->map_size != tdb->transaction->old_map_size) {
+               tdb_len_t add;
+
+               add = tdb->file->map_size - tdb->transaction->old_map_size;
                /* Restore original map size for tdb_expand_file */
-               tdb->map_size = tdb->transaction->old_map_size;
+               tdb->file->map_size = tdb->transaction->old_map_size;
                ecode = methods->expand_file(tdb, add);
                if (ecode != TDB_SUCCESS) {
-                       tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                                "tdb_transaction_prepare_commit:"
-                                " expansion failed");
-                       _tdb_transaction_cancel(tdb);
                        return ecode;
                }
        }
@@ -971,46 +1022,46 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 /*
    prepare to commit the current transaction
 */
-int tdb_transaction_prepare_commit(struct tdb_context *tdb)
+enum TDB_ERROR tdb_transaction_prepare_commit(struct tdb_context *tdb)
 {
-       tdb->ecode = _tdb_transaction_prepare_commit(tdb);
-       if (tdb->ecode != TDB_SUCCESS)
-               return -1;
-       return 0;
+       return _tdb_transaction_prepare_commit(tdb);
 }
 
 /*
   commit the current transaction
 */
-int tdb_transaction_commit(struct tdb_context *tdb)
+enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
 {
        const struct tdb_methods *methods;
        int i;
        enum TDB_ERROR ecode;
 
        if (tdb->transaction == NULL) {
-               tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
-                        "tdb_transaction_commit: no transaction");
-               return -1;
+               return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
+                                                   TDB_LOG_USE_ERROR,
+                                                   "tdb_transaction_commit:"
+                                                   " no transaction");
        }
 
        tdb_trace(tdb, "tdb_transaction_commit");
 
        if (tdb->transaction->nesting != 0) {
                tdb->transaction->nesting--;
-               return 0;
+               return tdb->last_error = TDB_SUCCESS;
        }
 
        /* check for a null transaction */
        if (tdb->transaction->blocks == NULL) {
                _tdb_transaction_cancel(tdb);
-               return 0;
+               return tdb->last_error = TDB_SUCCESS;
        }
 
        if (!tdb->transaction->prepared) {
-               tdb->ecode = _tdb_transaction_prepare_commit(tdb);
-               if (tdb->ecode != TDB_SUCCESS)
-                       return -1;
+               ecode = _tdb_transaction_prepare_commit(tdb);
+               if (ecode != TDB_SUCCESS) {
+                       _tdb_transaction_cancel(tdb);
+                       return tdb->last_error = ecode;
+               }
        }
 
        methods = tdb->transaction->io_methods;
@@ -1024,8 +1075,8 @@ int tdb_transaction_commit(struct tdb_context *tdb)
                        continue;
                }
 
-               offset = i * getpagesize();
-               length = getpagesize();
+               offset = i * PAGESIZE;
+               length = PAGESIZE;
                if (i == tdb->transaction->num_blocks-1) {
                        length = tdb->transaction->last_block_size;
                }
@@ -1033,10 +1084,6 @@ int tdb_transaction_commit(struct tdb_context *tdb)
                ecode = methods->twrite(tdb, offset,
                                        tdb->transaction->blocks[i], length);
                if (ecode != TDB_SUCCESS) {
-                       tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                                  "tdb_transaction_commit:"
-                                  " write failed during commit");
-
                        /* we've overwritten part of the data and
                           possibly expanded the file, so we need to
                           run the crash recovery code */
@@ -1045,7 +1092,7 @@ int tdb_transaction_commit(struct tdb_context *tdb)
 
                        _tdb_transaction_cancel(tdb);
 
-                       return -1;
+                       return tdb->last_error = ecode;
                }
                SAFE_FREE(tdb->transaction->blocks[i]);
        }
@@ -1054,10 +1101,9 @@ int tdb_transaction_commit(struct tdb_context *tdb)
        tdb->transaction->num_blocks = 0;
 
        /* ensure the new data is on disk */
-       ecode = transaction_sync(tdb, 0, tdb->map_size);
+       ecode = transaction_sync(tdb, 0, tdb->file->map_size);
        if (ecode != TDB_SUCCESS) {
-               tdb->ecode = ecode;
-               return -1;
+               return tdb->last_error = ecode;
        }
 
        /*
@@ -1076,10 +1122,11 @@ int tdb_transaction_commit(struct tdb_context *tdb)
 #endif
 
        /* use a transaction cancel to free memory and remove the
-          transaction locks */
+          transaction locks: it "restores" map_size, too. */
+       tdb->transaction->old_map_size = tdb->file->map_size;
        _tdb_transaction_cancel(tdb);
 
-       return 0;
+       return tdb->last_error = TDB_SUCCESS;
 }
 
 
@@ -1169,7 +1216,7 @@ enum TDB_ERROR tdb_transaction_recover(struct tdb_context *tdb)
 
        free(data);
 
-       ecode = transaction_sync(tdb, 0, tdb->map_size);
+       ecode = transaction_sync(tdb, 0, tdb->file->map_size);
        if (ecode != TDB_SUCCESS) {
                return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
                                  "tdb_transaction_recover:"