Port nested transaction allow from newer ctdb.
[ccan] / ccan / tdb / transaction.c
index 83a3ed85ab947d4f8c29b4c9ec2d62adbb803488..e55285f94cc6609a2afa720d39fe66d2261a36fe 100644 (file)
     still available, but no transaction recovery area is used and no
     fsync/msync calls are made.
 
-  - if TDB_NO_NESTING is passed to flags in tdb open then transaction
-    nesting is disabled. tdb_transaction_start() will then implicitely
-    cancel any pending transactions and always start a new transaction
-    context instead of nesting.
+  - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
+    tdb_add_flags() transaction is enabled.
+    The default is that transaction nesting is not allowed and an attempt
+    to create a nested transaction will fail with TDB_ERR_NESTING.
 
+    Beware. when transactions are nested a transaction successfully
+    completed with tdb_transaction_commit() can be silently unrolled later.
 */
 
 
@@ -121,8 +123,15 @@ struct tdb_transaction {
           but don't create a new transaction */
        int nesting;
 
+       /* set when a prepare has already occurred */
+       bool prepared;
+       tdb_off_t magic_offset;
+
        /* old file size before transaction */
        tdb_len_t old_map_size;
+
+       /* we should re-pack on commit */
+       bool need_repack;
 };
 
 
@@ -135,6 +144,14 @@ static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
 {
        uint32_t blk;
 
+       /* Only a commit is allowed on a prepared transaction */
+       if (tdb->transaction->prepared) {
+               tdb->ecode = TDB_ERR_EINVAL;
+               TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: transaction already prepared, read not allowed\n"));
+               tdb->transaction->transaction_error = 1;
+               return -1;
+       }
+
        /* break it down into block sized ops */
        while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
                tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
@@ -192,6 +209,14 @@ static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
 {
        uint32_t blk;
 
+       /* Only a commit is allowed on a prepared transaction */
+       if (tdb->transaction->prepared) {
+               tdb->ecode = TDB_ERR_EINVAL;
+               TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
+               tdb->transaction->transaction_error = 1;
+               return -1;
+       }
+
        /* if the write is to a hash head, then update the transaction
           hash heads */
        if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
@@ -362,7 +387,8 @@ static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
        if (len <= tdb->map_size) {
                return 0;
        }
-       return TDB_ERRCODE(TDB_ERR_IO, -1);
+       tdb->ecode = TDB_ERR_IO;
+       return -1;
 }
 
 /*
@@ -377,6 +403,8 @@ static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
                return -1;
        }
 
+       tdb->transaction->need_repack = true;
+
        return 0;
 }
 
@@ -398,9 +426,38 @@ static const struct tdb_methods transaction_methods = {
        transaction_brlock
 };
 
+/*
+  sync to disk
+*/
+static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
+{      
+       if (tdb->flags & TDB_NOSYNC) {
+               return 0;
+       }
+
+       if (fsync(tdb->fd) != 0) {
+               tdb->ecode = TDB_ERR_IO;
+               TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
+               return -1;
+       }
+#ifdef MS_SYNC
+       if (tdb->map_ptr) {
+               tdb_off_t moffset = offset & ~(tdb->page_size-1);
+               if (msync(moffset + (char *)tdb->map_ptr, 
+                         length + (offset - moffset), MS_SYNC) != 0) {
+                       tdb->ecode = TDB_ERR_IO;
+                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
+                                strerror(errno)));
+                       return -1;
+               }
+       }
+#endif
+       return 0;
+}
+
 int tdb_transaction_cancel_internal(struct tdb_context *tdb)
 {
-       int i;
+       int i, ret = 0;
 
        if (tdb->transaction == NULL) {
                TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
@@ -423,6 +480,18 @@ int tdb_transaction_cancel_internal(struct tdb_context *tdb)
        }
        SAFE_FREE(tdb->transaction->blocks);
 
+       if (tdb->transaction->magic_offset) {
+               const struct tdb_methods *methods = tdb->transaction->io_methods;
+               uint32_t zero = 0;
+
+               /* remove the recovery marker */
+               if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
+               transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
+                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
+                       ret = -1;
+               }
+       }
+
        /* remove any global lock created during the transaction */
        if (tdb->global_lock.count != 0) {
                tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
@@ -448,7 +517,7 @@ int tdb_transaction_cancel_internal(struct tdb_context *tdb)
        SAFE_FREE(tdb->transaction->hash_heads);
        SAFE_FREE(tdb->transaction);
        
-       return 0;
+       return ret;
 }
 
 /*
@@ -457,8 +526,6 @@ int tdb_transaction_cancel_internal(struct tdb_context *tdb)
 */
 int tdb_transaction_start(struct tdb_context *tdb)
 {
-       tdb_trace(tdb, "tdb_transaction_start");
-
        /* some sanity checks */
        if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
                TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
@@ -468,15 +535,15 @@ int tdb_transaction_start(struct tdb_context *tdb)
 
        /* cope with nested tdb_transaction_start() calls */
        if (tdb->transaction != NULL) {
-               if (!tdb->flags & TDB_NO_NESTING) {
-                       tdb->transaction->nesting++;
-                       TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
-                                tdb->transaction->nesting));
-                       return 0;
-               } else {
-                       tdb_transaction_cancel_internal(tdb);
-                       TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: cancelling previous transaction\n"));
+               if (!(tdb->flags & TDB_ALLOW_NESTING)) {
+                       tdb->ecode = TDB_ERR_NESTING;
+                       return -1;
                }
+               tdb_trace(tdb, "tdb_transaction_start");
+               tdb->transaction->nesting++;
+               TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
+                        tdb->transaction->nesting));
+               return 0;
        }
 
        if (tdb->num_locks != 0 || tdb->global_lock.count) {
@@ -549,6 +616,8 @@ int tdb_transaction_start(struct tdb_context *tdb)
        tdb->transaction->io_methods = tdb->methods;
        tdb->methods = &transaction_methods;
 
+       /* Trace at the end, so we get sequence number correct. */
+       tdb_trace(tdb, "tdb_transaction_start");
        return 0;
        
 fail:
@@ -569,31 +638,6 @@ int tdb_transaction_cancel(struct tdb_context *tdb)
        tdb_trace(tdb, "tdb_transaction_cancel");
        return tdb_transaction_cancel_internal(tdb);
 }
-/*
-  sync to disk
-*/
-static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
-{      
-       if (fsync(tdb->fd) != 0) {
-               tdb->ecode = TDB_ERR_IO;
-               TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
-               return -1;
-       }
-#ifdef MS_SYNC
-       if (tdb->map_ptr) {
-               tdb_off_t moffset = offset & ~(tdb->page_size-1);
-               if (msync(moffset + (char *)tdb->map_ptr, 
-                         length + (offset - moffset), MS_SYNC) != 0) {
-                       tdb->ecode = TDB_ERR_IO;
-                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
-                                strerror(errno)));
-                       return -1;
-               }
-       }
-#endif
-       return 0;
-}
-
 
 /*
   work out how much space the linearised recovery data will consume
@@ -841,26 +885,26 @@ static int transaction_setup_recovery(struct tdb_context *tdb,
        return 0;
 }
 
-/*
-  commit the current transaction
-*/
-int tdb_transaction_commit(struct tdb_context *tdb)
+static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb)
 {      
        const struct tdb_methods *methods;
-       tdb_off_t magic_offset = 0;
-       uint32_t zero = 0;
-       int i;
 
-       tdb_trace(tdb, "tdb_transaction_commit");
        if (tdb->transaction == NULL) {
-               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
+               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
+               return -1;
+       }
+
+       if (tdb->transaction->prepared) {
+               tdb->ecode = TDB_ERR_EINVAL;
+               tdb_transaction_cancel(tdb);
+               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
                return -1;
        }
 
        if (tdb->transaction->transaction_error) {
                tdb->ecode = TDB_ERR_IO;
                tdb_transaction_cancel_internal(tdb);
-               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
+               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
                return -1;
        }
 
@@ -870,9 +914,13 @@ int tdb_transaction_commit(struct tdb_context *tdb)
                return 0;
        }               
 
+#ifdef TDB_TRACE
+       /* store seqnum now, before reading becomes illegal. */
+       tdb_ofs_read(tdb, TDB_SEQNUM_OFS, &tdb->transaction_prepare_seqnum);
+#endif
+
        /* check for a null transaction */
        if (tdb->transaction->blocks == NULL) {
-               tdb_transaction_cancel_internal(tdb);
                return 0;
        }
 
@@ -882,14 +930,14 @@ int tdb_transaction_commit(struct tdb_context *tdb)
           nested their locks properly, so fail the transaction */
        if (tdb->num_locks || tdb->global_lock.count) {
                tdb->ecode = TDB_ERR_LOCK;
-               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: locks pending on commit\n"));
+               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
                tdb_transaction_cancel_internal(tdb);
                return -1;
        }
 
        /* upgrade the main transaction lock region to a write lock */
        if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
-               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to upgrade hash locks\n"));
+               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
                tdb->ecode = TDB_ERR_LOCK;
                tdb_transaction_cancel_internal(tdb);
                return -1;
@@ -898,7 +946,7 @@ int tdb_transaction_commit(struct tdb_context *tdb)
        /* get the global lock - this prevents new users attaching to the database
           during the commit */
        if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
-               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: failed to get global lock\n"));
+               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
                tdb->ecode = TDB_ERR_LOCK;
                tdb_transaction_cancel_internal(tdb);
                return -1;
@@ -906,21 +954,23 @@ int tdb_transaction_commit(struct tdb_context *tdb)
 
        if (!(tdb->flags & TDB_NOSYNC)) {
                /* write the recovery data to the end of the file */
-               if (transaction_setup_recovery(tdb, &magic_offset) == -1) {
-                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: failed to setup recovery data\n"));
+               if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
+                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
                        tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
                        tdb_transaction_cancel_internal(tdb);
                        return -1;
                }
        }
 
+       tdb->transaction->prepared = true;
+
        /* expand the file to the new size if needed */
        if (tdb->map_size != tdb->transaction->old_map_size) {
                if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
                                             tdb->map_size - 
                                             tdb->transaction->old_map_size) == -1) {
                        tdb->ecode = TDB_ERR_IO;
-                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: expansion failed\n"));
+                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
                        tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
                        tdb_transaction_cancel_internal(tdb);
                        return -1;
@@ -929,6 +979,69 @@ int tdb_transaction_commit(struct tdb_context *tdb)
                methods->tdb_oob(tdb, tdb->map_size + 1, 1);
        }
 
+       /* Keep the global lock until the actual commit */
+
+       return 0;
+}
+
+/*
+   prepare to commit the current transaction
+*/
+int tdb_transaction_prepare_commit(struct tdb_context *tdb)
+{      
+       tdb_trace(tdb, "tdb_transaction_prepare_commit");
+       return tdb_transaction_prepare_commit_internal(tdb);
+}
+
+/*
+  commit the current transaction
+*/
+int tdb_transaction_commit(struct tdb_context *tdb)
+{      
+       const struct tdb_methods *methods;
+       int i;
+       bool need_repack;
+
+       if (tdb->transaction == NULL) {
+               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
+               return -1;
+       }
+
+       /* If we've prepared, can't read seqnum. */
+       if (tdb->transaction->prepared) {
+               tdb_trace_seqnum(tdb, tdb->transaction_prepare_seqnum,
+                                "tdb_transaction_commit");
+       } else {
+               tdb_trace(tdb, "tdb_transaction_commit");
+       }
+
+       if (tdb->transaction->transaction_error) {
+               tdb->ecode = TDB_ERR_IO;
+               tdb_transaction_cancel(tdb);
+               TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
+               return -1;
+       }
+
+
+       if (tdb->transaction->nesting != 0) {
+               tdb->transaction->nesting--;
+               return 0;
+       }
+
+       /* check for a null transaction */
+       if (tdb->transaction->blocks == NULL) {
+               tdb_transaction_cancel_internal(tdb);
+               return 0;
+       }
+
+       if (!tdb->transaction->prepared) {
+               int ret = tdb_transaction_prepare_commit_internal(tdb);
+               if (ret)
+                       return ret;
+       }
+
+       methods = tdb->transaction->io_methods;
+
        /* perform all the writes */
        for (i=0;i<tdb->transaction->num_blocks;i++) {
                tdb_off_t offset;
@@ -965,22 +1078,9 @@ int tdb_transaction_commit(struct tdb_context *tdb)
        SAFE_FREE(tdb->transaction->blocks);
        tdb->transaction->num_blocks = 0;
 
-       if (!(tdb->flags & TDB_NOSYNC)) {
-               /* ensure the new data is on disk */
-               if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
-                       return -1;
-               }
-
-               /* remove the recovery marker */
-               if (methods->tdb_write(tdb, magic_offset, &zero, 4) == -1) {
-                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: failed to remove recovery magic\n"));
-                       return -1;
-               }
-
-               /* ensure the recovery marker has been removed on disk */
-               if (transaction_sync(tdb, magic_offset, 4) == -1) {
-                       return -1;
-               }
+       /* ensure the new data is on disk */
+       if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
+               return -1;
        }
 
        tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
@@ -1000,10 +1100,16 @@ int tdb_transaction_commit(struct tdb_context *tdb)
        utime(tdb->name, NULL);
 #endif
 
+       need_repack = tdb->transaction->need_repack;
+
        /* use a transaction cancel to free memory and remove the
           transaction locks */
        tdb_transaction_cancel_internal(tdb);
 
+       if (need_repack) {
+               return tdb_repack(tdb);
+       }
+
        return 0;
 }