]> git.ozlabs.org Git - ccan/blobdiff - ccan/tdb/transaction.c
tdb: no longer need seqnum store hack as we can read after prepare.
[ccan] / ccan / tdb / transaction.c
index 1086c5f037793d9df00269ff78b14ae2ad126e72..34137251bce51f2697184fbd8dc515315b9d64c6 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.
 */
 
 
@@ -127,6 +129,9 @@ struct tdb_transaction {
 
        /* old file size before transaction */
        tdb_len_t old_map_size;
+
+       /* we should re-pack on commit */
+       bool need_repack;
 };
 
 
@@ -139,14 +144,6 @@ 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);
@@ -382,7 +379,8 @@ 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;
 }
 
 /*
@@ -397,14 +395,23 @@ static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
                return -1;
        }
 
+       tdb->transaction->need_repack = true;
+
        return 0;
 }
 
 /*
   brlock during a transaction - ignore them
 */
-static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset, 
-                             int rw_type, int lck_type, int probe, size_t len)
+static int transaction_brlock(struct tdb_context *tdb,
+                             int rw_type, tdb_off_t offset, size_t len,
+                             enum tdb_lock_flags flags)
+{
+       return 0;
+}
+
+static int transaction_brunlock(struct tdb_context *tdb,
+                               int rw_type, tdb_off_t offset, size_t len)
 {
        return 0;
 }
@@ -415,7 +422,8 @@ static const struct tdb_methods transaction_methods = {
        transaction_next_hash_chain,
        transaction_oob,
        transaction_expand_file,
-       transaction_brlock
+       transaction_brlock,
+       transaction_brunlock
 };
 
 /*
@@ -447,7 +455,8 @@ static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t
        return 0;
 }
 
-int tdb_transaction_cancel_internal(struct tdb_context *tdb)
+/* ltype is F_WRLCK after prepare. */
+int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype)
 {
        int i, ret = 0;
 
@@ -486,15 +495,16 @@ int tdb_transaction_cancel_internal(struct tdb_context *tdb)
 
        /* 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);
+               tdb_brunlock(tdb, tdb->global_lock.ltype,
+                            FREELIST_TOP, 4*tdb->header.hash_size);
                tdb->global_lock.count = 0;
        }
 
        /* remove any locks created during the transaction */
        if (tdb->num_locks != 0) {
                for (i=0;i<tdb->num_lockrecs;i++) {
-                       tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
-                                  F_UNLCK,F_SETLKW, 0, 1);
+                       tdb_brunlock(tdb, tdb->lockrecs[i].ltype,
+                                    FREELIST_TOP+4*tdb->lockrecs[i].list, 1);
                }
                tdb->num_locks = 0;
                tdb->num_lockrecs = 0;
@@ -504,8 +514,8 @@ int tdb_transaction_cancel_internal(struct tdb_context *tdb)
        /* restore the normal io methods */
        tdb->methods = tdb->transaction->io_methods;
 
-       tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
-       tdb_transaction_unlock(tdb);
+       tdb_brunlock(tdb, ltype, FREELIST_TOP, 0);
+       tdb_transaction_unlock(tdb, F_WRLCK);
        SAFE_FREE(tdb->transaction->hash_heads);
        SAFE_FREE(tdb->transaction);
        
@@ -527,16 +537,15 @@ int tdb_transaction_start(struct tdb_context *tdb)
 
        /* cope with nested tdb_transaction_start() calls */
        if (tdb->transaction != NULL) {
-               tdb_trace(tdb, "tdb_transaction_start");
-               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) {
@@ -578,7 +587,7 @@ int tdb_transaction_start(struct tdb_context *tdb)
        
        /* get a read lock from the freelist to the end of file. This
           is upgraded to a write lock during the commit */
-       if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
+       if (tdb_brlock(tdb, F_RDLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT) == -1) {
                TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
                tdb->ecode = TDB_ERR_LOCK;
                goto fail;
@@ -614,8 +623,8 @@ int tdb_transaction_start(struct tdb_context *tdb)
        return 0;
        
 fail:
-       tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
-       tdb_transaction_unlock(tdb);
+       tdb_brunlock(tdb, F_RDLCK, FREELIST_TOP, 0);
+       tdb_transaction_unlock(tdb, F_WRLCK);
        SAFE_FREE(tdb->transaction->blocks);
        SAFE_FREE(tdb->transaction->hash_heads);
        SAFE_FREE(tdb->transaction);
@@ -627,9 +636,12 @@ fail:
   cancel the current transaction
 */
 int tdb_transaction_cancel(struct tdb_context *tdb)
-{      
+{
+       int ltype = F_RDLCK;
        tdb_trace(tdb, "tdb_transaction_cancel");
-       return tdb_transaction_cancel_internal(tdb);
+       if (tdb->transaction && tdb->transaction->prepared)
+               ltype = F_WRLCK;
+       return _tdb_transaction_cancel(tdb, ltype);
 }
 
 /*
@@ -878,7 +890,7 @@ static int transaction_setup_recovery(struct tdb_context *tdb,
        return 0;
 }
 
-static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb)
+static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 {      
        const struct tdb_methods *methods;
 
@@ -889,14 +901,14 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb)
 
        if (tdb->transaction->prepared) {
                tdb->ecode = TDB_ERR_EINVAL;
-               tdb_transaction_cancel(tdb);
+               _tdb_transaction_cancel(tdb, F_WRLCK);
                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_transaction_cancel(tdb, F_RDLCK);
                TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
                return -1;
        }
@@ -907,11 +919,6 @@ static int tdb_transaction_prepare_commit_internal(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) {
                return 0;
@@ -924,7 +931,7 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb)
        if (tdb->num_locks || tdb->global_lock.count) {
                tdb->ecode = TDB_ERR_LOCK;
                TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
-               tdb_transaction_cancel_internal(tdb);
+               _tdb_transaction_cancel(tdb, F_RDLCK);
                return -1;
        }
 
@@ -932,16 +939,16 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb)
        if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
                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);
+               _tdb_transaction_cancel(tdb, F_RDLCK);
                return -1;
        }
 
        /* 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) {
+       if (tdb_brlock(tdb, F_WRLCK, GLOBAL_LOCK, 1, TDB_LOCK_WAIT) == -1) {
                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);
+               _tdb_transaction_cancel(tdb, F_WRLCK);
                return -1;
        }
 
@@ -949,8 +956,8 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb)
                /* write the recovery data to the end of the file */
                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);
+                       tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
+                       _tdb_transaction_cancel(tdb, F_WRLCK);
                        return -1;
                }
        }
@@ -964,8 +971,8 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb)
                                             tdb->transaction->old_map_size) == -1) {
                        tdb->ecode = TDB_ERR_IO;
                        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);
+                       tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
+                       _tdb_transaction_cancel(tdb, F_WRLCK);
                        return -1;
                }
                tdb->map_size = tdb->transaction->old_map_size;
@@ -983,7 +990,7 @@ static int tdb_transaction_prepare_commit_internal(struct tdb_context *tdb)
 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
 {      
        tdb_trace(tdb, "tdb_transaction_prepare_commit");
-       return tdb_transaction_prepare_commit_internal(tdb);
+       return _tdb_transaction_prepare_commit(tdb);
 }
 
 /*
@@ -993,19 +1000,14 @@ 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");
-       }
+       tdb_trace(tdb, "tdb_transaction_commit");
 
        if (tdb->transaction->transaction_error) {
                tdb->ecode = TDB_ERR_IO;
@@ -1022,12 +1024,12 @@ int tdb_transaction_commit(struct tdb_context *tdb)
 
        /* check for a null transaction */
        if (tdb->transaction->blocks == NULL) {
-               tdb_transaction_cancel_internal(tdb);
+               _tdb_transaction_cancel(tdb, F_RDLCK);
                return 0;
        }
 
        if (!tdb->transaction->prepared) {
-               int ret = tdb_transaction_prepare_commit_internal(tdb);
+               int ret = _tdb_transaction_prepare_commit(tdb);
                if (ret)
                        return ret;
        }
@@ -1058,8 +1060,8 @@ int tdb_transaction_commit(struct tdb_context *tdb)
                        tdb->methods = methods;
                        tdb_transaction_recover(tdb); 
 
-                       tdb_transaction_cancel_internal(tdb);
-                       tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
+                       _tdb_transaction_cancel(tdb, F_WRLCK);
+                       tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
 
                        TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
                        return -1;
@@ -1075,7 +1077,7 @@ int tdb_transaction_commit(struct tdb_context *tdb)
                return -1;
        }
 
-       tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
+       tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
 
        /*
          TODO: maybe write to some dummy hdr field, or write to magic
@@ -1092,9 +1094,15 @@ 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);
+       _tdb_transaction_cancel(tdb, F_WRLCK);
+
+       if (need_repack) {
+               return tdb_repack(tdb);
+       }
 
        return 0;
 }