tdb2: allow multiple chain locks.
[ccan] / ccan / tdb2 / lock.c
index 98570d7627e7a38272c4f72ebcf60edece5dadfd..4033d0afe813e59bf4dee4b8bcc9e5d34757c298 100644 (file)
 #include <assert.h>
 #include <ccan/build_assert/build_assert.h>
 
-static int fcntl_lock(struct tdb_context *tdb,
-                     int rw, off_t off, off_t len, bool waitflag)
+/* If we were threaded, we could wait for unlock, but we're not, so fail. */
+static enum TDB_ERROR owner_conflict(struct tdb_context *tdb, const char *call)
 {
-       struct flock fl;
+       return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
+                         "%s: lock owned by another tdb in this process.",
+                         call);
+}
 
-       fl.l_type = rw;
-       fl.l_whence = SEEK_SET;
-       fl.l_start = off;
-       fl.l_len = len;
-       fl.l_pid = 0;
+/* If we fork, we no longer really own locks: preserves errno */
+static bool check_lock_pid(struct tdb_context *tdb,
+                          const char *call, bool log)
+{
+       /* No locks?  No problem! */
+       if (tdb->file->allrecord_lock.count == 0
+           && tdb->file->num_lockrecs == 0) {
+               return true;
+       }
 
-       add_stat(tdb, lock_lowlevel, 1);
-       if (waitflag)
-               return fcntl(tdb->file->fd, F_SETLKW, &fl);
-       else {
-               add_stat(tdb, lock_nonblock, 1);
-               return fcntl(tdb->file->fd, F_SETLK, &fl);
+       /* No fork?  No problem! */
+       if (tdb->file->locker == getpid()) {
+               return true;
        }
+
+       if (log) {
+               tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
+                          "%s: fork() detected after lock acquisition!"
+                          " (%u vs %u)", call, tdb->file->locker, getpid());
+       }
+       return false;
+}
+
+int tdb_fcntl_lock(int fd, int rw, off_t off, off_t len, bool waitflag,
+                  void *unused)
+{
+       struct flock fl;
+       int ret;
+
+       do {
+               fl.l_type = rw;
+               fl.l_whence = SEEK_SET;
+               fl.l_start = off;
+               fl.l_len = len;
+
+               if (waitflag)
+                       ret = fcntl(fd, F_SETLKW, &fl);
+               else
+                       ret = fcntl(fd, F_SETLK, &fl);
+       } while (ret != 0 && errno == EINTR);
+       return ret;
 }
 
-static int fcntl_unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
+int tdb_fcntl_unlock(int fd, int rw, off_t off, off_t len, void *unused)
 {
        struct flock fl;
+       int ret;
+
+       do {
+               fl.l_type = F_UNLCK;
+               fl.l_whence = SEEK_SET;
+               fl.l_start = off;
+               fl.l_len = len;
+
+               ret = fcntl(fd, F_SETLKW, &fl);
+       } while (ret != 0 && errno == EINTR);
+       return ret;
+}
+
+static int lock(struct tdb_context *tdb,
+                     int rw, off_t off, off_t len, bool waitflag)
+{
+       if (tdb->file->allrecord_lock.count == 0
+           && tdb->file->num_lockrecs == 0) {
+               tdb->file->locker = getpid();
+       }
+
+       tdb->stats.lock_lowlevel++;
+       if (!waitflag)
+               tdb->stats.lock_nonblock++;
+       return tdb->lock_fn(tdb->file->fd, rw, off, len, waitflag,
+                           tdb->lock_data);
+}
+
+static int unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
+{
 #if 0 /* Check they matched up locks and unlocks correctly. */
        char line[80];
        FILE *locks;
@@ -110,13 +171,7 @@ static int fcntl_unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
        fclose(locks);
 #endif
 
-       fl.l_type = F_UNLCK;
-       fl.l_whence = SEEK_SET;
-       fl.l_start = off;
-       fl.l_len = len;
-       fl.l_pid = 0;
-
-       return fcntl(tdb->file->fd, F_SETLKW, &fl);
+       return tdb->unlock_fn(tdb->file->fd, rw, off, len, tdb->lock_data);
 }
 
 /* a byte range locking function - return 0 on success
@@ -147,16 +202,13 @@ static enum TDB_ERROR tdb_brlock(struct tdb_context *tdb,
                                  (long long)(offset + len));
        }
 
-       do {
-               ret = fcntl_lock(tdb, rw_type, offset, len,
-                                flags & TDB_LOCK_WAIT);
-       } while (ret == -1 && errno == EINTR);
-
-       if (ret == -1) {
+       ret = lock(tdb, rw_type, offset, len, flags & TDB_LOCK_WAIT);
+       if (ret != 0) {
                /* Generic lock error. errno set by fcntl.
                 * EAGAIN is an expected return from non-blocking
                 * locks. */
-               if (!(flags & TDB_LOCK_PROBE) && errno != EAGAIN) {
+               if (!(flags & TDB_LOCK_PROBE)
+                   && (errno != EAGAIN && errno != EINTR)) {
                        tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
                                   "tdb_brlock failed (fd=%d) at"
                                   " offset %zu rw_type=%d flags=%d len=%zu:"
@@ -178,16 +230,15 @@ static enum TDB_ERROR tdb_brunlock(struct tdb_context *tdb,
                return TDB_SUCCESS;
        }
 
-       do {
-               ret = fcntl_unlock(tdb, rw_type, offset, len);
-       } while (ret == -1 && errno == EINTR);
+       ret = unlock(tdb, rw_type, offset, len);
 
-       if (ret == -1) {
+       /* If we fail, *then* we verify that we owned the lock.  If not, ok. */
+       if (ret == -1 && check_lock_pid(tdb, "tdb_brunlock", false)) {
                return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
                                  "tdb_brunlock failed (fd=%d) at offset %zu"
-                                 " rw_type=%d len=%zu",
+                                 " rw_type=%d len=%zu: %s",
                                  tdb->file->fd, (size_t)offset, rw_type,
-                                 (size_t)len);
+                                 (size_t)len, strerror(errno));
        }
        return TDB_SUCCESS;
 }
@@ -202,6 +253,9 @@ enum TDB_ERROR tdb_allrecord_upgrade(struct tdb_context *tdb)
 {
        int count = 1000;
 
+       if (!check_lock_pid(tdb, "tdb_transaction_prepare_commit", true))
+               return TDB_ERR_LOCK;
+
        if (tdb->file->allrecord_lock.count != 1) {
                return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
                                  "tdb_allrecord_upgrade failed:"
@@ -215,6 +269,10 @@ enum TDB_ERROR tdb_allrecord_upgrade(struct tdb_context *tdb)
                                  " already upgraded?");
        }
 
+       if (tdb->file->allrecord_lock.owner != tdb) {
+               return owner_conflict(tdb, "tdb_allrecord_upgrade");
+       }
+
        while (count--) {
                struct timeval tv;
                if (tdb_brlock(tdb, F_WRLCK,
@@ -232,17 +290,22 @@ enum TDB_ERROR tdb_allrecord_upgrade(struct tdb_context *tdb)
                tv.tv_usec = 1;
                select(0, NULL, NULL, NULL, &tv);
        }
-       return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
-                         "tdb_allrecord_upgrade failed");
+
+       if (errno != EAGAIN && errno != EINTR)
+               tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
+                          "tdb_allrecord_upgrade failed");
+       return TDB_ERR_LOCK;
 }
 
-static struct tdb_lock_type *find_nestlock(struct tdb_context *tdb,
-                                          tdb_off_t offset)
+static struct tdb_lock *find_nestlock(struct tdb_context *tdb, tdb_off_t offset,
+                                     const struct tdb_context *owner)
 {
        unsigned int i;
 
        for (i=0; i<tdb->file->num_lockrecs; i++) {
                if (tdb->file->lockrecs[i].off == offset) {
+                       if (owner && tdb->file->lockrecs[i].owner != owner)
+                               return NULL;
                        return &tdb->file->lockrecs[i];
                }
        }
@@ -253,6 +316,9 @@ enum TDB_ERROR tdb_lock_and_recover(struct tdb_context *tdb)
 {
        enum TDB_ERROR ecode;
 
+       if (!check_lock_pid(tdb, "tdb_transaction_prepare_commit", true))
+               return TDB_ERR_LOCK;
+
        ecode = tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK,
                                   false);
        if (ecode != TDB_SUCCESS) {
@@ -276,10 +342,11 @@ static enum TDB_ERROR tdb_nest_lock(struct tdb_context *tdb,
                                    tdb_off_t offset, int ltype,
                                    enum tdb_lock_flags flags)
 {
-       struct tdb_lock_type *new_lck;
+       struct tdb_lock *new_lck;
        enum TDB_ERROR ecode;
 
-       if (offset > TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE + tdb->map_size / 8) {
+       if (offset > (TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
+                     + tdb->file->map_size / 8)) {
                return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
                                  "tdb_nest_lock: invalid offset %zu ltype=%d",
                                  (size_t)offset, ltype);
@@ -288,10 +355,18 @@ static enum TDB_ERROR tdb_nest_lock(struct tdb_context *tdb,
        if (tdb->flags & TDB_NOLOCK)
                return TDB_SUCCESS;
 
-       add_stat(tdb, locks, 1);
+       if (!check_lock_pid(tdb, "tdb_nest_lock", true)) {
+               return TDB_ERR_LOCK;
+       }
+
+       tdb->stats.locks++;
 
-       new_lck = find_nestlock(tdb, offset);
+       new_lck = find_nestlock(tdb, offset, NULL);
        if (new_lck) {
+               if (new_lck->owner != tdb) {
+                       return owner_conflict(tdb, "tdb_nest_lock");
+               }
+
                if (new_lck->ltype == F_RDLCK && ltype == F_WRLCK) {
                        return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
                                          "tdb_nest_lock:"
@@ -303,14 +378,16 @@ static enum TDB_ERROR tdb_nest_lock(struct tdb_context *tdb,
                return TDB_SUCCESS;
        }
 
+#if 0
        if (tdb->file->num_lockrecs
            && offset >= TDB_HASH_LOCK_START
            && offset < TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE) {
                return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
                                  "tdb_nest_lock: already have a hash lock?");
        }
+#endif
 
-       new_lck = (struct tdb_lock_type *)realloc(
+       new_lck = (struct tdb_lock *)realloc(
                tdb->file->lockrecs,
                sizeof(*tdb->file->lockrecs) * (tdb->file->num_lockrecs+1));
        if (new_lck == NULL) {
@@ -348,6 +425,7 @@ static enum TDB_ERROR tdb_nest_lock(struct tdb_context *tdb,
                }
        }
 
+       tdb->file->lockrecs[tdb->file->num_lockrecs].owner = tdb;
        tdb->file->lockrecs[tdb->file->num_lockrecs].off = offset;
        tdb->file->lockrecs[tdb->file->num_lockrecs].count = 1;
        tdb->file->lockrecs[tdb->file->num_lockrecs].ltype = ltype;
@@ -359,13 +437,13 @@ static enum TDB_ERROR tdb_nest_lock(struct tdb_context *tdb,
 static enum TDB_ERROR tdb_nest_unlock(struct tdb_context *tdb,
                                      tdb_off_t off, int ltype)
 {
-       struct tdb_lock_type *lck;
+       struct tdb_lock *lck;
        enum TDB_ERROR ecode;
 
        if (tdb->flags & TDB_NOLOCK)
                return TDB_SUCCESS;
 
-       lck = find_nestlock(tdb, off);
+       lck = find_nestlock(tdb, off, tdb);
        if ((lck == NULL) || (lck->count == 0)) {
                return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
                                  "tdb_nest_unlock: no lock for %zu",
@@ -452,14 +530,24 @@ enum TDB_ERROR tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
        enum TDB_ERROR ecode;
        tdb_bool_err berr;
 
-       if (tdb->file->allrecord_lock.count
-           && (ltype == F_RDLCK
-               || tdb->file->allrecord_lock.ltype == F_WRLCK)) {
-               tdb->file->allrecord_lock.count++;
+       if (tdb->flags & TDB_NOLOCK)
                return TDB_SUCCESS;
+
+       if (!check_lock_pid(tdb, "tdb_allrecord_lock", true)) {
+               return TDB_ERR_LOCK;
        }
 
        if (tdb->file->allrecord_lock.count) {
+               if (tdb->file->allrecord_lock.owner != tdb) {
+                       return owner_conflict(tdb, "tdb_allrecord_lock");
+               }
+
+               if (ltype == F_RDLCK
+                   || tdb->file->allrecord_lock.ltype == F_WRLCK) {
+                       tdb->file->allrecord_lock.count++;
+                       return TDB_SUCCESS;
+               }
+
                /* a global lock of a different type exists */
                return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
                                  "tdb_allrecord_lock: already have %s lock",
@@ -481,33 +569,25 @@ enum TDB_ERROR tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
                                  " can't upgrade a write lock");
        }
 
-       add_stat(tdb, locks, 1);
+       tdb->stats.locks++;
 again:
        /* Lock hashes, gradually. */
        ecode = tdb_lock_gradual(tdb, ltype, flags, TDB_HASH_LOCK_START,
                                 TDB_HASH_LOCK_RANGE);
-       if (ecode != TDB_SUCCESS) {
-               if (!(flags & TDB_LOCK_PROBE)) {
-                       tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                                  "tdb_allrecord_lock hashes failed");
-               }
+       if (ecode != TDB_SUCCESS)
                return ecode;
-       }
 
        /* Lock free tables: there to end of file. */
        ecode = tdb_brlock(tdb, ltype,
                           TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE,
                           0, flags);
        if (ecode != TDB_SUCCESS) {
-               if (!(flags & TDB_LOCK_PROBE)) {
-                       tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                                "tdb_allrecord_lock freetables failed");
-               }
                tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START,
                             TDB_HASH_LOCK_RANGE);
                return ecode;
        }
 
+       tdb->file->allrecord_lock.owner = tdb;
        tdb->file->allrecord_lock.count = 1;
        /* If it's upgradable, it's actually exclusive so we can treat
         * it as a write lock. */
@@ -545,7 +625,7 @@ void tdb_unlock_open(struct tdb_context *tdb)
 bool tdb_has_open_lock(struct tdb_context *tdb)
 {
        return !(tdb->flags & TDB_NOLOCK)
-               && find_nestlock(tdb, TDB_OPEN_LOCK) != NULL;
+               && find_nestlock(tdb, TDB_OPEN_LOCK, tdb) != NULL;
 }
 
 enum TDB_ERROR tdb_lock_expand(struct tdb_context *tdb, int ltype)
@@ -563,18 +643,27 @@ void tdb_unlock_expand(struct tdb_context *tdb, int ltype)
 /* unlock entire db */
 void tdb_allrecord_unlock(struct tdb_context *tdb, int ltype)
 {
+       if (tdb->flags & TDB_NOLOCK)
+               return;
+
        if (tdb->file->allrecord_lock.count == 0) {
                tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
                           "tdb_allrecord_unlock: not locked!");
                return;
        }
 
+       if (tdb->file->allrecord_lock.owner != tdb) {
+               tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
+                          "tdb_allrecord_unlock: not locked by us!");
+               return;
+       }
+
        /* Upgradable locks are marked as write locks. */
        if (tdb->file->allrecord_lock.ltype != ltype
            && (!tdb->file->allrecord_lock.off || ltype != F_RDLCK)) {
                tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
                           "tdb_allrecord_unlock: have %s lock",
-                          tdb->allrecord_lock.ltype == F_RDLCK
+                          tdb->file->allrecord_lock.ltype == F_RDLCK
                           ? "read" : "write");
                return;
        }
@@ -592,7 +681,7 @@ void tdb_allrecord_unlock(struct tdb_context *tdb, int ltype)
 
 bool tdb_has_expansion_lock(struct tdb_context *tdb)
 {
-       return find_nestlock(tdb, TDB_EXPANSION_LOCK) != NULL;
+       return find_nestlock(tdb, TDB_EXPANSION_LOCK, tdb) != NULL;
 }
 
 bool tdb_has_hash_locks(struct tdb_context *tdb)
@@ -629,16 +718,21 @@ enum TDB_ERROR tdb_lock_hashes(struct tdb_context *tdb,
                               int ltype, enum tdb_lock_flags waitflag)
 {
        /* FIXME: Do this properly, using hlock_range */
-       unsigned lock = TDB_HASH_LOCK_START
+       unsigned l = TDB_HASH_LOCK_START
                + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
 
        /* a allrecord lock allows us to avoid per chain locks */
-       if (tdb->file->allrecord_lock.count &&
-           (ltype == tdb->file->allrecord_lock.ltype || ltype == F_RDLCK)) {
-               return TDB_SUCCESS;
-       }
-
        if (tdb->file->allrecord_lock.count) {
+               if (!check_lock_pid(tdb, "tdb_lock_hashes", true))
+                       return TDB_ERR_LOCK;
+
+               if (tdb->file->allrecord_lock.owner != tdb)
+                       return owner_conflict(tdb, "tdb_lock_hashes");
+               if (ltype == tdb->file->allrecord_lock.ltype
+                   || ltype == F_RDLCK) {
+                       return TDB_SUCCESS;
+               }
+
                return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
                                  "tdb_lock_hashes:"
                                  " already have %s allrecordlock",
@@ -657,14 +751,14 @@ enum TDB_ERROR tdb_lock_hashes(struct tdb_context *tdb,
                                  " already have expansion lock");
        }
 
-       return tdb_nest_lock(tdb, lock, ltype, waitflag);
+       return tdb_nest_lock(tdb, l, ltype, waitflag);
 }
 
 enum TDB_ERROR tdb_unlock_hashes(struct tdb_context *tdb,
                                 tdb_off_t hash_lock,
                                 tdb_len_t hash_range, int ltype)
 {
-       unsigned lock = TDB_HASH_LOCK_START
+       unsigned l = TDB_HASH_LOCK_START
                + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
 
        if (tdb->flags & TDB_NOLOCK)
@@ -680,7 +774,7 @@ enum TDB_ERROR tdb_unlock_hashes(struct tdb_context *tdb,
                return TDB_SUCCESS;
        }
 
-       return tdb_nest_unlock(tdb, lock, ltype);
+       return tdb_nest_unlock(tdb, l, ltype);
 }
 
 /* Hash locks use TDB_HASH_LOCK_START + the next 30 bits.
@@ -704,6 +798,9 @@ enum TDB_ERROR tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
 
        /* a allrecord lock allows us to avoid per chain locks */
        if (tdb->file->allrecord_lock.count) {
+               if (!check_lock_pid(tdb, "tdb_lock_free_bucket", true))
+                       return TDB_ERR_LOCK;
+
                if (tdb->file->allrecord_lock.ltype == F_WRLCK)
                        return 0;
                return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
@@ -729,3 +826,42 @@ void tdb_unlock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off)
 
        tdb_nest_unlock(tdb, free_lock_off(b_off), F_WRLCK);
 }
+
+enum TDB_ERROR tdb_lockall(struct tdb_context *tdb)
+{
+       return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false);
+}
+
+void tdb_unlockall(struct tdb_context *tdb)
+{
+       tdb_allrecord_unlock(tdb, F_WRLCK);
+}
+
+enum TDB_ERROR tdb_lockall_read(struct tdb_context *tdb)
+{
+       return tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false);
+}
+
+void tdb_unlockall_read(struct tdb_context *tdb)
+{
+       tdb_allrecord_unlock(tdb, F_RDLCK);
+}
+
+void tdb_lock_cleanup(struct tdb_context *tdb)
+{
+       unsigned int i;
+
+       while (tdb->file->allrecord_lock.count
+              && tdb->file->allrecord_lock.owner == tdb) {
+               tdb_allrecord_unlock(tdb, tdb->file->allrecord_lock.ltype);
+       }
+
+       for (i=0; i<tdb->file->num_lockrecs; i++) {
+               if (tdb->file->lockrecs[i].owner == tdb) {
+                       tdb_nest_unlock(tdb,
+                                       tdb->file->lockrecs[i].off,
+                                       tdb->file->lockrecs[i].ltype);
+                       i--;
+               }
+       }
+}