tdb: make _tdb_transaction_cancel static.
[samba.git] / lib / tdb / common / transaction.c
index 501cd62b96ee367626f148f24eebbfd07a5de3b2..07935c9492bca6b7f17141b5095993bd0f7a8cd4 100644 (file)
     fsync/msync calls are made.
 
   - 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.
+    tdb_add_flags() transaction nesting is enabled.
+    It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
+    The default is that transaction nesting is allowed.
+    Note: this default may change in future versions of tdb.
 
     Beware. when transactions are nested a transaction successfully
     completed with tdb_transaction_commit() can be silently unrolled later.
+
+  - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
+    tdb_add_flags() transaction nesting is disabled.
+    It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
+    An attempt create a nested transaction will fail with TDB_ERR_NESTING.
+    The default is that transaction nesting is allowed.
+    Note: this default may change in future versions of tdb.
 */
 
 
@@ -127,6 +135,9 @@ struct tdb_transaction {
        bool prepared;
        tdb_off_t magic_offset;
 
+       /* set when the GLOBAL_LOCK has been taken */
+       bool global_lock_taken;
+
        /* old file size before transaction */
        tdb_len_t old_map_size;
 
@@ -403,8 +414,15 @@ static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
 /*
   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 +433,8 @@ static const struct tdb_methods transaction_methods = {
        transaction_next_hash_chain,
        transaction_oob,
        transaction_expand_file,
-       transaction_brlock
+       transaction_brlock,
+       transaction_brunlock
 };
 
 
@@ -483,7 +502,7 @@ 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;
@@ -519,8 +538,8 @@ 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);
@@ -537,7 +556,7 @@ static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t
                return 0;
        }
 
-       if (fsync(tdb->fd) != 0) {
+       if (fdatasync(tdb->fd) != 0) {
                tdb->ecode = TDB_ERR_IO;
                TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
                return -1;
@@ -558,7 +577,8 @@ static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t
 }
 
 
-int _tdb_transaction_cancel(struct tdb_context *tdb)
+/* ltype is F_WRLCK after prepare. */
+static int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype)
 {      
        int i, ret = 0;
 
@@ -585,27 +605,33 @@ int _tdb_transaction_cancel(struct tdb_context *tdb)
 
        if (tdb->transaction->magic_offset) {
                const struct tdb_methods *methods = tdb->transaction->io_methods;
-               uint32_t zero = 0;
+               const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
 
                /* remove the recovery marker */
-               if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
+               if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 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;
                }
        }
 
+       if (tdb->transaction->global_lock_taken) {
+               tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
+               tdb->transaction->global_lock_taken = false;
+       }
+
        /* 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;
@@ -615,8 +641,8 @@ int _tdb_transaction_cancel(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);
        
@@ -628,8 +654,11 @@ int _tdb_transaction_cancel(struct tdb_context *tdb)
 */
 int tdb_transaction_cancel(struct tdb_context *tdb)
 {
+       int ltype = F_RDLCK;
        tdb_trace(tdb, "tdb_transaction_cancel");
-       return _tdb_transaction_cancel(tdb);
+       if (tdb->transaction && tdb->transaction->prepared)
+               ltype = F_WRLCK;
+       return _tdb_transaction_cancel(tdb, ltype);
 }
 
 /*
@@ -679,10 +708,16 @@ static int tdb_recovery_allocate(struct tdb_context *tdb,
 
        rec.rec_len = 0;
 
-       if (recovery_head != 0 && 
-           methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
-               TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
-               return -1;
+       if (recovery_head != 0) {
+               if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
+                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
+                       return -1;
+               }
+               /* ignore invalid recovery regions: can happen in crash */
+               if (rec.magic != TDB_RECOVERY_MAGIC &&
+                   rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
+                       recovery_head = 0;
+               }
        }
 
        *recovery_size = tdb_recovery_size(tdb);
@@ -777,7 +812,7 @@ static int transaction_setup_recovery(struct tdb_context *tdb,
        rec = (struct tdb_record *)data;
        memset(rec, 0, sizeof(*rec));
 
-       rec->magic    = 0;
+       rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
        rec->data_len = recovery_size;
        rec->rec_len  = recovery_max_size;
        rec->key_len  = old_map_size;
@@ -889,14 +924,14 @@ static int _tdb_transaction_prepare_commit(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(tdb);
+               _tdb_transaction_cancel(tdb, F_RDLCK);
                TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
                return -1;
        }
@@ -918,7 +953,7 @@ static int _tdb_transaction_prepare_commit(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(tdb);
+               _tdb_transaction_cancel(tdb, F_RDLCK);
                return -1;
        }
 
@@ -926,25 +961,26 @@ static int _tdb_transaction_prepare_commit(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(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(tdb);
+               _tdb_transaction_cancel(tdb, F_WRLCK);
                return -1;
        }
 
+       tdb->transaction->global_lock_taken = true;
+
        if (!(tdb->flags & TDB_NOSYNC)) {
                /* 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(tdb);
+                       _tdb_transaction_cancel(tdb, F_WRLCK);
                        return -1;
                }
        }
@@ -958,8 +994,7 @@ static int _tdb_transaction_prepare_commit(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(tdb);
+                       _tdb_transaction_cancel(tdb, F_WRLCK);
                        return -1;
                }
                tdb->map_size = tdb->transaction->old_map_size;
@@ -998,7 +1033,7 @@ int tdb_transaction_commit(struct tdb_context *tdb)
 
        if (tdb->transaction->transaction_error) {
                tdb->ecode = TDB_ERR_IO;
-               _tdb_transaction_cancel(tdb);
+               _tdb_transaction_cancel(tdb, F_RDLCK);
                TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
                return -1;
        }
@@ -1011,7 +1046,7 @@ int tdb_transaction_commit(struct tdb_context *tdb)
 
        /* check for a null transaction */
        if (tdb->transaction->blocks == NULL) {
-               _tdb_transaction_cancel(tdb);
+               _tdb_transaction_cancel(tdb, F_RDLCK);
                return 0;
        }
 
@@ -1047,8 +1082,7 @@ int tdb_transaction_commit(struct tdb_context *tdb)
                        tdb->methods = methods;
                        tdb_transaction_recover(tdb); 
 
-                       _tdb_transaction_cancel(tdb);
-                       tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
+                       _tdb_transaction_cancel(tdb, F_WRLCK);
 
                        TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
                        return -1;
@@ -1064,8 +1098,6 @@ int tdb_transaction_commit(struct tdb_context *tdb)
                return -1;
        }
 
-       tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
-
        /*
          TODO: maybe write to some dummy hdr field, or write to magic
          offset without mmap, before the last sync, instead of the
@@ -1085,7 +1117,7 @@ int tdb_transaction_commit(struct tdb_context *tdb)
 
        /* use a transaction cancel to free memory and remove the
           transaction locks */
-       _tdb_transaction_cancel(tdb);
+       _tdb_transaction_cancel(tdb, F_WRLCK);
 
        if (need_repack) {
                return tdb_repack(tdb);