tdb2: clarify locking heirarchy; hashes, freelist, then expand lock.
authorRusty Russell <rusty@rustcorp.com.au>
Sun, 17 Oct 2010 04:10:21 +0000 (14:40 +1030)
committerRusty Russell <rusty@rustcorp.com.au>
Sun, 17 Oct 2010 04:10:21 +0000 (14:40 +1030)
ccan/tdb2/io.c
ccan/tdb2/lock.c

index aa79def1e38cc5b49b500d1b513eb9dc446fbbb7..2bcee8a29a17665edf3a08a6b6264bdb7c446322 100644 (file)
@@ -73,7 +73,9 @@ static int tdb_oob(struct tdb_context *tdb, tdb_off_t len, bool probe)
        int ret;
 
        /* We can't hold pointers during this: we could unmap! */
-       assert(!tdb->direct_access || tdb_has_expansion_lock(tdb));
+       assert(!tdb->direct_access
+              || (tdb->flags & TDB_NOLOCK)
+              || tdb_has_expansion_lock(tdb));
 
        if (len <= tdb->map_size)
                return 0;
index 15c97e3b98343b0a02ee910bdac0ec93664e3527..14f9c0ae307b348b1a7401b6001bbd419a8e0c69 100644 (file)
@@ -260,7 +260,7 @@ static int tdb_nest_lock(struct tdb_context *tdb, tdb_off_t offset, int ltype,
        if (offset >= TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE + tdb->map_size / 8) {
                tdb->ecode = TDB_ERR_LOCK;
                tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-                        "tdb_lock: invalid offset %llu ltype=%d\n",
+                        "tdb_nest_lock: invalid offset %llu ltype=%d\n",
                         (long long)offset, ltype);
                return -1;
        }
@@ -284,7 +284,7 @@ static int tdb_nest_lock(struct tdb_context *tdb, tdb_off_t offset, int ltype,
        if (new_lck == NULL) {
                tdb->ecode = TDB_ERR_OOM;
                tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                        "tdb_lock: unable to allocate %llu lock structure",
+                        "tdb_nest_lock: unable to allocate %llu lock struct",
                         (long long)(tdb->num_lockrecs + 1));
                errno = ENOMEM;
                return -1;
@@ -351,7 +351,7 @@ static int tdb_nest_unlock(struct tdb_context *tdb, tdb_off_t off, int ltype)
        if ((lck == NULL) || (lck->count == 0)) {
                tdb->ecode = TDB_ERR_LOCK;
                tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                        "tdb_unlock: no lock for %llu\n", (long long)off);
+                        "tdb_nest_unlock: no lock for %llu\n", (long long)off);
                return -1;
        }
 
@@ -433,9 +433,7 @@ static int tdb_lock_gradual(struct tdb_context *tdb,
 }
 
 /* lock/unlock entire database.  It can only be upgradable if you have some
- * other way of guaranteeing exclusivity (ie. transaction write lock).
- * Note that we don't lock the free chains: currently noone can get those locks
- * without a hash chain lock first. */
+ * other way of guaranteeing exclusivity (ie. transaction write lock). */
 int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
                       enum tdb_lock_flags flags, bool upgradable)
 {
@@ -479,16 +477,30 @@ int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
        }
 
 again:
+       /* Lock hashes, gradually. */
        if (tdb_lock_gradual(tdb, ltype, flags, TDB_HASH_LOCK_START,
                             TDB_HASH_LOCK_RANGE)) {
                if (!(flags & TDB_LOCK_PROBE)) {
                        tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                                "tdb_lockall hashes failed (%s)\n",
+                                "tdb_allrecord_lock hashes failed (%s)\n",
                                 strerror(errno));
                }
                return -1;
        }
 
+       /* Lock free lists: there to end of file. */
+       if (tdb_brlock(tdb, ltype, TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE,
+                      0, flags)) {
+               if (!(flags & TDB_LOCK_PROBE)) {
+                       tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+                                "tdb_allrecord_lock freelist failed (%s)\n",
+                                strerror(errno));
+               }
+               tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START, 
+                            TDB_HASH_LOCK_RANGE);
+               return -1;
+       }
+
        tdb->allrecord_lock.count = 1;
        /* If it's upgradable, it's actually exclusive so we can treat
         * it as a write lock. */
@@ -564,14 +576,12 @@ int tdb_allrecord_unlock(struct tdb_context *tdb, int ltype)
        tdb->allrecord_lock.count = 0;
        tdb->allrecord_lock.ltype = 0;
 
-       return tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START,
-                           TDB_HASH_LOCK_RANGE);
+       return tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START, 0);
 }
 
 bool tdb_has_expansion_lock(struct tdb_context *tdb)
 {
-       return find_nestlock(tdb, TDB_EXPANSION_LOCK) != NULL
-               || (tdb->flags & TDB_NOLOCK);
+       return find_nestlock(tdb, TDB_EXPANSION_LOCK) != NULL;
 }
 
 bool tdb_has_locks(struct tdb_context *tdb)
@@ -625,6 +635,18 @@ int tdb_unlockall_read(struct tdb_context *tdb)
 }
 #endif
 
+static bool tdb_has_free_lock(struct tdb_context *tdb)
+{
+       unsigned int i;
+
+       for (i=0; i<tdb->num_lockrecs; i++) {
+               if (tdb->lockrecs[i].off
+                   > TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE)
+                       return true;
+       }
+       return false;
+}
+
 int tdb_lock_hashes(struct tdb_context *tdb,
                    tdb_off_t hash_lock,
                    tdb_len_t hash_range,
@@ -643,12 +665,26 @@ int tdb_lock_hashes(struct tdb_context *tdb,
        if (tdb->allrecord_lock.count) {
                tdb->ecode = TDB_ERR_LOCK;
                tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-                        "tdb_lock_list: have %s allrecordlock\n",
+                        "tdb_lock_hashes: have %s allrecordlock\n",
                         tdb->allrecord_lock.ltype == F_RDLCK
                         ? "read" : "write");
                return -1;
        }
 
+       if (tdb_has_free_lock(tdb)) {
+               tdb->ecode = TDB_ERR_LOCK;
+               tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+                        "tdb_lock_hashes: have free lock already\n");
+               return -1;
+       }
+
+       if (tdb_has_expansion_lock(tdb)) {
+               tdb->ecode = TDB_ERR_LOCK;
+               tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+                        "tdb_lock_hashes: have expansion lock already\n");
+               return -1;
+       }
+
        return tdb_nest_lock(tdb, lock, ltype, waitflag);
 }
 
@@ -665,7 +701,7 @@ int tdb_unlock_hashes(struct tdb_context *tdb,
                    && ltype == F_WRLCK) {
                        tdb->ecode = TDB_ERR_LOCK;
                        tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-                                "tdb_unlock_list RO allrecord!\n");
+                                "tdb_unlock_hashes RO allrecord!\n");
                        return -1;
                }
                return 0;
@@ -681,7 +717,8 @@ int tdb_unlock_hashes(struct tdb_context *tdb,
  */
 static tdb_off_t free_lock_off(tdb_off_t b_off)
 {
-       return TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE + b_off / sizeof(tdb_off_t);
+       return TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
+               + b_off / sizeof(tdb_off_t);
 }
 
 int tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
@@ -689,24 +726,25 @@ int tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
 {
        assert(b_off >= sizeof(struct tdb_header));
 
-       /* You're supposed to have a hash lock first! */
-       if (!(tdb->flags & TDB_NOLOCK) && !tdb_has_locks(tdb)) {
-               tdb->ecode = TDB_ERR_LOCK;
-               tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-                        "tdb_lock_free_list without lock!\n");
-               return -1;
-       }
-
        /* a allrecord lock allows us to avoid per chain locks */
        if (tdb->allrecord_lock.count) {
                if (tdb->allrecord_lock.ltype == F_WRLCK)
                        return 0;
                tdb->ecode = TDB_ERR_LOCK;
                tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-                        "tdb_lock_free_list with RO allrecordlock!\n");
+                        "tdb_lock_free_bucket with RO allrecordlock!\n");
                return -1;
        }
 
+#if 0 /* FIXME */
+       if (tdb_has_expansion_lock(tdb)) {
+               tdb->ecode = TDB_ERR_LOCK;
+               tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+                        "tdb_lock_free_bucket: have expansion lock already\n");
+               return -1;
+       }
+#endif
+
        return tdb_nest_lock(tdb, free_lock_off(b_off), F_WRLCK, waitflag);
 }