* nscd/cache.c (cache_add): Take additional parameter specifying

	whether this is in response of a cache refill.  Check alignment
	of package data.  Revamp waking of pruning thread.
	(prune_cache): Small optimization.
	* nscd/nscd.h: Adjust cache_add prototypes.
	* nscd/aicache.c: Adjust cache_add calls.
	* nscd/grpcache.c: Likewise.
	* nscd/hstcache.c: Likewise.
	* nscd/initgrcache.c: Likewise.
	* nscd/pwdcache.c: Likewise.
	* nscd/servicescache.c: Likewise.
	* nscd/connections.c (restart): Really disable cache use before
	exec attempt.  If it fails, reenable cache.
	(nscd_run_prune): Initialize wakeup_time.  After wakeup, set wakeup
	time to max to be able to notice concurrent cache additions.  Unlock
	prune_lock while performing gc.  Afterwards compute wakeup time with
	current wakeup_time value in mind.

10 files changed, 114 insertions, 42 deletions

diff --git a/ChangeLog b/ChangeLog
index d053ed080a..da8b1b4a22 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,23 @@
+2008-05-18  Ulrich Drepper  <drepper@redhat.com>
+
+	* nscd/cache.c (cache_add): Take additional parameter specifying
+	whether this is in response of a cache refill.  Check alignment
+	of package data.  Revamp waking of pruning thread.
+	(prune_cache): Small optimization.
+	* nscd/nscd.h: Adjust cache_add prototypes.
+	* nscd/aicache.c: Adjust cache_add calls.
+	* nscd/grpcache.c: Likewise.
+	* nscd/hstcache.c: Likewise.
+	* nscd/initgrcache.c: Likewise.
+	* nscd/pwdcache.c: Likewise.
+	* nscd/servicescache.c: Likewise.
+	* nscd/connections.c (restart): Really disable cache use before
+	exec attempt.  If it fails, reenable cache.
+	(nscd_run_prune): Initialize wakeup_time.  After wakeup, set wakeup
+	time to max to be able to notice concurrent cache additions.  Unlock
+	prune_lock while performing gc.  Afterwards compute wakeup time with
+	current wakeup_time value in mind.
+
 2008-05-17  Ulrich Drepper  <drepper@redhat.com>
 
 	* nscd/mem.c (gc): Avoid stack overflow when allocating move list.
diff --git a/nscd/aicache.c b/nscd/aicache.c
index 7ae5a1645c..5ffab76a04 100644
--- a/nscd/aicache.c
+++ b/nscd/aicache.c
@@ -558,7 +558,7 @@ next_nip:
       pthread_rwlock_rdlock (&db->lock);
 
       (void) cache_add (req->type, key_copy, req->key_len, &dataset->head,
-			true, db, uid);
+			true, db, uid, he == NULL);
 
       pthread_rwlock_unlock (&db->lock);
 
diff --git a/nscd/cache.c b/nscd/cache.c
index f4a9de530a..2faaf3403e 100644
--- a/nscd/cache.c
+++ b/nscd/cache.c
@@ -135,7 +135,7 @@ cache_search (request_type type, void *key, size_t len,
 int
 cache_add (int type, const void *key, size_t len, struct datahead *packet,
 	   bool first, struct database_dyn *table,
-	   uid_t owner)
+	   uid_t owner, bool prune_wakeup)
 {
   if (__builtin_expect (debug_level >= 2, 0))
     {
@@ -180,6 +180,7 @@ cache_add (int type, const void *key, size_t len, struct datahead *packet,
   assert (newp->key + newp->len <= table->head->first_free);
   newp->owner = owner;
   newp->packet = (char *) packet - table->data;
+  assert ((newp->packet & BLOCK_ALIGN_M1) == 0);
 
   /* Put the new entry in the first position.  */
   do
@@ -211,19 +212,27 @@ cache_add (int type, const void *key, size_t len, struct datahead *packet,
 	   (char *) &table->head->array[hash] - (char *) table->head
 	   + sizeof (ref_t), MS_ASYNC);
 
-  /* Perhaps the prune thread for the data is not running in a long
-     time.  Wake it if necessary.  */
-  time_t next_wakeup = table->wakeup_time;
-  while (next_wakeup + CACHE_PRUNE_INTERVAL > packet->timeout)
-    if (atomic_compare_and_exchange_bool_acq (&table->wakeup_time,
-					      packet->timeout,
-					      next_wakeup) == 0)
-      {
+  /* We do not have to worry about the pruning thread if we are
+     re-adding the data since this is done by the pruning thread.  We
+     also do not have to do anything in case this is not the first
+     time the data is entered since different data heads all have the
+     same timeout.  */
+  if (first && prune_wakeup)
+    {
+      /* Perhaps the prune thread for the table is not running in a long
+	 time.  Wake it if necessary.  */
+      pthread_mutex_lock (&table->prune_lock);
+      time_t next_wakeup = table->wakeup_time;
+      bool do_wakeup = false;
+      if (next_wakeup > packet->timeout + CACHE_PRUNE_INTERVAL)
+	{
+	  table->wakeup_time = packet->timeout;
+	  do_wakeup = true;
+	}
+      pthread_mutex_unlock (&table->prune_lock);
+      if (do_wakeup)
 	pthread_cond_signal (&table->prune_cond);
-	break;
-      }
-    else
-      next_wakeup = table->wakeup_time;
+    }
 
   /* Mark the in-flight memory as unused.  */
   for (enum in_flight idx = 0; idx < IDX_last; ++idx)
@@ -436,7 +445,8 @@ prune_cache (struct database_dyn *table, time_t now, int fd)
 	      ref_t *old = &table->head->array[first];
 	      ref_t run = table->head->array[first];
 
-	      while (run != ENDREF)
+	      assert (run != ENDREF);
+	      do
 		{
 		  struct hashentry *runp = (struct hashentry *) (data + run);
 		  struct datahead *dh
@@ -462,6 +472,7 @@ prune_cache (struct database_dyn *table, time_t now, int fd)
 		      run = runp->next;
 		    }
 		}
+	      while (run != ENDREF);
 	    }
 
 	  ++first;
diff --git a/nscd/connections.c b/nscd/connections.c
index 15148bdf3d..e6a2c946ae 100644
--- a/nscd/connections.c
+++ b/nscd/connections.c
@@ -1330,11 +1330,14 @@ cannot change to old working directory: %s; disabling paranoia mode"),
     }
 
   /* Synchronize memory.  */
+  int32_t certainly[lastdb];
   for (int cnt = 0; cnt < lastdb; ++cnt)
     if (dbs[cnt].enabled)
       {
 	/* Make sure nobody keeps using the database.  */
 	dbs[cnt].head->timestamp = 0;
+	certainly[cnt] = dbs[cnt].head->nscd_certainly_running;
+	dbs[cnt].head->nscd_certainly_running = 0;
 
 	if (dbs[cnt].persistent)
 	  // XXX async OK?
@@ -1357,6 +1360,15 @@ cannot change to old working directory: %s; disabling paranoia mode"),
     dbg_log (_("cannot change current working directory to \"/\": %s"),
 	     strerror (errno));
   paranoia = 0;
+
+  /* Reenable the databases.  */
+  time_t now = time (NULL);
+  for (int cnt = 0; cnt < lastdb; ++cnt)
+    if (dbs[cnt].enabled)
+      {
+	dbs[cnt].head->timestamp = now;
+	dbs[cnt].head->nscd_certainly_running = certainly[cnt];
+      }
 }
 
 
@@ -1394,42 +1406,68 @@ nscd_run_prune (void *p)
 
   int dont_need_update = setup_thread (&dbs[my_number]);
 
+  time_t now = time (NULL);
+
   /* We are running.  */
-  dbs[my_number].head->timestamp = time (NULL);
+  dbs[my_number].head->timestamp = now;
 
   struct timespec prune_ts;
-  if (clock_gettime (timeout_clock, &prune_ts) == -1)
+  if (__builtin_expect (clock_gettime (timeout_clock, &prune_ts) == -1, 0))
     /* Should never happen.  */
     abort ();
 
   /* Compute the initial timeout time.  Prevent all the timers to go
      off at the same time by adding a db-based value.  */
   prune_ts.tv_sec += CACHE_PRUNE_INTERVAL + my_number;
+  dbs[my_number].wakeup_time = now + CACHE_PRUNE_INTERVAL + my_number;
+
+  pthread_mutex_t *prune_lock = &dbs[my_number].prune_lock;
+  pthread_cond_t *prune_cond = &dbs[my_number].prune_cond;
 
-  pthread_mutex_lock (&dbs[my_number].prune_lock);
+  pthread_mutex_lock (prune_lock);
   while (1)
     {
       /* Wait, but not forever.  */
-      int e = pthread_cond_timedwait (&dbs[my_number].prune_cond,
-				      &dbs[my_number].prune_lock,
-				      &prune_ts);
-      assert (e == 0 || e == ETIMEDOUT);
+      int e = pthread_cond_timedwait (prune_cond, prune_lock, &prune_ts);
+      assert (__builtin_expect (e == 0 || e == ETIMEDOUT, 1));
 
       time_t next_wait;
-      time_t now = time (NULL);
+      now = time (NULL);
       if (e == ETIMEDOUT || now >= dbs[my_number].wakeup_time)
 	{
+	  /* We will determine the new timout values based on the
+	     cache content.  Should there be concurrent additions to
+	     the cache which are not accounted for in the cache
+	     pruning we want to know about it.  Therefore set the
+	     timeout to the maximum.  It will be descreased when adding
+	     new entries to the cache, if necessary.  */
+	  if (sizeof (time_t) == sizeof (long int))
+	    dbs[my_number].wakeup_time = LONG_MAX;
+	  else
+	    dbs[my_number].wakeup_time = INT_MAX;
+
+	  pthread_mutex_unlock (prune_lock);
+
 	  next_wait = prune_cache (&dbs[my_number], now, -1);
+
 	  next_wait = MAX (next_wait, CACHE_PRUNE_INTERVAL);
 	  /* If clients cannot determine for sure whether nscd is running
 	     we need to wake up occasionally to update the timestamp.
 	     Wait 90% of the update period.  */
 #define UPDATE_MAPPING_TIMEOUT (MAPPING_TIMEOUT * 9 / 10)
 	  if (__builtin_expect (! dont_need_update, 0))
-	    next_wait = MIN (UPDATE_MAPPING_TIMEOUT, next_wait);
+	    {
+	      next_wait = MIN (UPDATE_MAPPING_TIMEOUT, next_wait);
+	      dbs[my_number].head->timestamp = now;
+	    }
+
+	  pthread_mutex_lock (prune_lock);
 
 	  /* Make it known when we will wake up again.  */
-	  dbs[my_number].wakeup_time = now + next_wait;
+	  if (now + next_wait < dbs[my_number].wakeup_time)
+	    dbs[my_number].wakeup_time = now + next_wait;
+	  else
+	    next_wait = dbs[my_number].wakeup_time - now;
 	}
       else
 	/* The cache was just pruned.  Do not do it again now.  Just
diff --git a/nscd/grpcache.c b/nscd/grpcache.c
index 9921ae313e..e391dc3984 100644
--- a/nscd/grpcache.c
+++ b/nscd/grpcache.c
@@ -147,7 +147,7 @@ cache_addgr (struct database_dyn *db, int fd, request_header *req,
 	      pthread_rwlock_rdlock (&db->lock);
 
 	      (void) cache_add (req->type, &dataset->strdata, req->key_len,
-				&dataset->head, true, db, owner);
+				&dataset->head, true, db, owner, he == NULL);
 
 	      pthread_rwlock_unlock (&db->lock);
 
@@ -353,7 +353,7 @@ cache_addgr (struct database_dyn *db, int fd, request_header *req,
 	  if (req->type == GETGRBYGID)
 	    {
 	      if (cache_add (GETGRBYGID, cp, key_offset, &dataset->head, true,
-			     db, owner) < 0)
+			     db, owner, he == NULL) < 0)
 		goto out;
 
 	      first = false;
@@ -362,7 +362,7 @@ cache_addgr (struct database_dyn *db, int fd, request_header *req,
 	  else if (strcmp (key_copy, gr_name) != 0)
 	    {
 	      if (cache_add (GETGRBYNAME, key_copy, key_len + 1,
-			     &dataset->head, true, db, owner) < 0)
+			     &dataset->head, true, db, owner, he == NULL) < 0)
 		goto out;
 
 	      first = false;
@@ -372,12 +372,13 @@ cache_addgr (struct database_dyn *db, int fd, request_header *req,
 	  if ((req->type == GETGRBYNAME || db->propagate)
 	      && __builtin_expect (cache_add (GETGRBYNAME, gr_name,
 					      gr_name_len,
-					      &dataset->head, first, db, owner)
+					      &dataset->head, first, db, owner,
+					      he == NULL)
 				   == 0, 1))
 	    {
 	      if (req->type == GETGRBYNAME && db->propagate)
 		(void) cache_add (GETGRBYGID, cp, key_offset, &dataset->head,
-				  false, db, owner);
+				  false, db, owner, false);
 	    }
 
 	out:
diff --git a/nscd/hstcache.c b/nscd/hstcache.c
index 3ceb6715ce..d4dd51f11a 100644
--- a/nscd/hstcache.c
+++ b/nscd/hstcache.c
@@ -156,7 +156,7 @@ cache_addhst (struct database_dyn *db, int fd, request_header *req,
 	      pthread_rwlock_rdlock (&db->lock);
 
 	      (void) cache_add (req->type, &dataset->strdata, req->key_len,
-				&dataset->head, true, db, owner);
+				&dataset->head, true, db, owner, he == NULL);
 
 	      pthread_rwlock_unlock (&db->lock);
 
@@ -408,7 +408,7 @@ cache_addhst (struct database_dyn *db, int fd, request_header *req,
 		  || req->type == GETHOSTBYADDRv6);
 
 	  (void) cache_add (req->type, key_copy, req->key_len,
-			    &dataset->head, true, db, owner);
+			    &dataset->head, true, db, owner, he == NULL);
 
 	  pthread_rwlock_unlock (&db->lock);
 	}
diff --git a/nscd/initgrcache.c b/nscd/initgrcache.c
index 94e909d4ae..3355df5164 100644
--- a/nscd/initgrcache.c
+++ b/nscd/initgrcache.c
@@ -231,7 +231,7 @@ addinitgroupsX (struct database_dyn *db, int fd, request_header *req,
 	      pthread_rwlock_rdlock (&db->lock);
 
 	      (void) cache_add (req->type, key_copy, req->key_len,
-				&dataset->head, true, db, uid);
+				&dataset->head, true, db, uid, he == NULL);
 
 	      pthread_rwlock_unlock (&db->lock);
 
@@ -398,7 +398,7 @@ addinitgroupsX (struct database_dyn *db, int fd, request_header *req,
 	  pthread_rwlock_rdlock (&db->lock);
 
 	  (void) cache_add (INITGROUPS, cp, req->key_len, &dataset->head, true,
-			    db, uid);
+			    db, uid, he == NULL);
 
 	  pthread_rwlock_unlock (&db->lock);
 	}
diff --git a/nscd/nscd.h b/nscd/nscd.h
index b024017fd4..b07256de80 100644
--- a/nscd/nscd.h
+++ b/nscd/nscd.h
@@ -231,7 +231,8 @@ extern struct datahead *cache_search (request_type, void *key, size_t len,
 				      uid_t owner);
 extern int cache_add (int type, const void *key, size_t len,
 		      struct datahead *packet, bool first,
-		      struct database_dyn *table, uid_t owner);
+		      struct database_dyn *table, uid_t owner,
+		      bool prune_wakeup);
 extern time_t prune_cache (struct database_dyn *table, time_t now, int fd);
 
 /* pwdcache.c */
diff --git a/nscd/pwdcache.c b/nscd/pwdcache.c
index 2d0a26592c..4a0cabd240 100644
--- a/nscd/pwdcache.c
+++ b/nscd/pwdcache.c
@@ -154,7 +154,7 @@ cache_addpw (struct database_dyn *db, int fd, request_header *req,
 	      pthread_rwlock_rdlock (&db->lock);
 
 	      (void) cache_add (req->type, key_copy, req->key_len,
-				&dataset->head, true, db, owner);
+				&dataset->head, true, db, owner, he == NULL);
 
 	      pthread_rwlock_unlock (&db->lock);
 
@@ -348,7 +348,7 @@ cache_addpw (struct database_dyn *db, int fd, request_header *req,
 	  if (req->type == GETPWBYUID)
 	    {
 	      if (cache_add (GETPWBYUID, cp, key_offset, &dataset->head, true,
-			     db, owner) < 0)
+			     db, owner, he == NULL) < 0)
 		goto out;
 
 	      first = false;
@@ -357,7 +357,7 @@ cache_addpw (struct database_dyn *db, int fd, request_header *req,
 	  else if (strcmp (key_copy, dataset->strdata) != 0)
 	    {
 	      if (cache_add (GETPWBYNAME, key_copy, key_len + 1,
-			     &dataset->head, true, db, owner) < 0)
+			     &dataset->head, true, db, owner, he == NULL) < 0)
 		goto out;
 
 	      first = false;
@@ -367,11 +367,12 @@ cache_addpw (struct database_dyn *db, int fd, request_header *req,
 	  if ((req->type == GETPWBYNAME || db->propagate)
 	      && __builtin_expect (cache_add (GETPWBYNAME, dataset->strdata,
 					      pw_name_len, &dataset->head,
-					      first, db, owner) == 0, 1))
+					      first, db, owner, he == NULL)
+				   == 0, 1))
 	    {
 	      if (req->type == GETPWBYNAME && db->propagate)
 		(void) cache_add (GETPWBYUID, cp, key_offset, &dataset->head,
-				  false, db, owner);
+				  false, db, owner, false);
 	    }
 
 	out:
diff --git a/nscd/servicescache.c b/nscd/servicescache.c
index c6f0b47e35..c952fa12c8 100644
--- a/nscd/servicescache.c
+++ b/nscd/servicescache.c
@@ -137,7 +137,7 @@ cache_addserv (struct database_dyn *db, int fd, request_header *req,
 	      pthread_rwlock_rdlock (&db->lock);
 
 	      (void) cache_add (req->type, &dataset->strdata, req->key_len,
-				&dataset->head, true, db, owner);
+				&dataset->head, true, db, owner, he == NULL);
 
 	      pthread_rwlock_unlock (&db->lock);
 
@@ -331,7 +331,7 @@ cache_addserv (struct database_dyn *db, int fd, request_header *req,
 	  pthread_rwlock_rdlock (&db->lock);
 
 	  (void) cache_add (req->type, key_copy, req->key_len,
-			    &dataset->head, true, db, owner);
+			    &dataset->head, true, db, owner, he == NULL);
 
 	  pthread_rwlock_unlock (&db->lock);
 	}