[SSI] openssi/kernel/cluster/ssi/ipc ipc.h, 1.6, 1.7 namesvr_clnt.c, 1.8, 1.9 namesvr_func.c, 1.18, 1.19

Update of /cvsroot/ssic-linux/openssi/kernel/cluster/ssi/ipc
In directory sfp-cvsdas-3.v30.ch3.sourceforge.com:/tmp/cvs-serv8504/kernel/cluster/ssi/ipc

Modified Files:
      Tag: OPENSSI-FC
	ipc.h namesvr_clnt.c namesvr_func.c 
Log Message:
IPC:
- Fix uninitialized iodb_timestamp.
- Fix ssi_growary() can cause array bounds violation when value of iodb_size is
  increased beyond the static size of object table.
- Fix ipcname_failover_data() dereferencing invalid ipc_obj structure; objp
  pointer is uninitialized.
- Fix shmem_svr_lookup() ignored cli_ipcname_getid() error and looked up id 0.
- Fix shmem_svr_lookup() always creates obj with id 0.

IPC (#ifdef SSI_IPC_OBJ_DB_LOCKLESS):
- Remove NSC_IPC_RD/WRLOCK() global object database mutex.
- Implement reference counting for ipc_obj structure.
- Implement lockless iodb_active[] traversal.
- Fix ipcname_failover_data() iodb_seq race with another objsvr_new() thread.
- Remove redundant objsvr_find_key() lookup's in ipcname_getid() path when
  objsvr_new() raced with another.

 cluster/ssi/cfs/cfs_ipcshm.c   |   15 -
 cluster/ssi/ipc/ipc.h          |   26 +
 cluster/ssi/ipc/namesvr_clnt.c |    6 
 cluster/ssi/ipc/namesvr_func.c |  463 ++++++++++++++++++++++++++++++++-
 include/linux/config.h         |    5 
 5 files changed, 495 insertions(+), 20 deletions(-)


Index: ipc.h
===================================================================
RCS file: /cvsroot/ssic-linux/openssi/kernel/cluster/ssi/ipc/ipc.h,v
retrieving revision 1.6
retrieving revision 1.7
diff -u -d -r1.6 -r1.7
--- ipc.h	10 Feb 2005 01:07:20 -0000	1.6
+++ ipc.h	5 Mar 2010 06:25:36 -0000	1.7
@@ -43,6 +43,17 @@
 
 struct ipc_obj
 {
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	atomic_t	io_refcnt;
+	int		io_id;
+	clusternode_t	io_svr_node;
+	int		io_local_view;
+	int		io_size;
+	struct kern_ipc_perm	*io_perm;
+	ipc_obj_db_t	*io_obj_dbp;
+	unsigned short	iodb_active_idx; /* index in iodb_active[] */
+	struct rcu_head	io_rcu;
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 	ipc_obj_db_t	*io_obj_dbp;
 	int		 io_id;
 	unsigned short	 io_seq;
@@ -51,6 +62,7 @@
 	clusternode_t	 svr_node;
 	clusternode_t	 new_node;
 	int		 local_view;
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 };
 
 /*
@@ -63,13 +75,23 @@
 	ipc_obj_t	 *iodb_active[IPCMNI];	/* List of active objects. */
 	struct iodb_ops *iodb_ops;
 	int		 iodb_size;
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	unsigned short	 iodb_seq;
+	unsigned short	 iodb_seq_max;
+	time_t		 iodb_timestamp;	/* timestamp of last update */
+	struct semaphore iodb_sem;		/* database mutex */
+	spinlock_t	 iodb_lock;		/* lock for iodb_active[] */
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 	int		 iodb_max;
 	unsigned short	 iodb_seq;
 	unsigned short	 iodb_seq_max;
 	int		 iodb_objects;		/* number of objects on chain */
 	time_t		 iodb_timestamp;	/* timestamp of last update */
 	RW_LOCK_T	 rwlock;		/* database mutex */
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
+#ifdef SSI_NOTUSED
 	int	 	 iodb_seqnum;		/* Last assigned id */
+#endif
 };
 
 typedef struct kern_ipc_perm ipc_perm_t;
@@ -81,10 +103,12 @@
 	int (*iop_get_check)(ipc_obj_t *, long); /* to check size or nsem. */
 };
 
+#ifndef SSI_IPC_OBJ_DB_LOCKLESS
 #define NSC_IPC_RDLOCK(ipcp)	LOCK_SHR_RW_LOCK(&(ipcp)->rwlock)
 #define NSC_IPC_WRLOCK(ipcp)	LOCK_EXCL_RW_LOCK(&(ipcp)->rwlock)
 #define NSC_IPC_RDUNLOCK(ipcp) 	UNLOCK_SHR_RW_LOCK(&(ipcp)->rwlock)
 #define NSC_IPC_WRUNLOCK(ipcp) 	UNLOCK_EXCL_RW_LOCK(&(ipcp)->rwlock)
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 
 /*
  * Common Distributed Inter-Process Communication routines.
@@ -128,11 +152,13 @@
 	int		view,
 	clusternode_t	node);
 
+#ifdef SSI_NOTYET
 int cli_objsvr_get_check(
 	ipc_obj_db_t	*objdbp,	/* Object type's database. */
 	int		id,
 	long		check_arg,
 	clusternode_t	obj_server,
 	int		flag);
+#endif
 
 #endif /* _NSC_IPC_IPC_H */

Index: namesvr_func.c
===================================================================
RCS file: /cvsroot/ssic-linux/openssi/kernel/cluster/ssi/ipc/namesvr_func.c,v
retrieving revision 1.18
retrieving revision 1.19
diff -u -d -r1.18 -r1.19
--- namesvr_func.c	13 Feb 2010 14:57:54 -0000	1.18
+++ namesvr_func.c	5 Mar 2010 06:25:36 -0000	1.19
@@ -88,6 +88,36 @@
 /*
  * The database for shared memory objects managed by this node.
  */
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+ipc_obj_db_t nsc_name_odb[NAME_SERVICE_MAX] =
+{
+{	"nsc_ipcname (msg)",	/* Service name */
+	NAME_SERVICE_MSG,	/* Service number */
+	{0},			/* Active list */
+	&nsc_name_iodb_ops,
+	MSGMNI,		/* size */
+	1,			/* seq */
+	USHRT_MAX,		/* seq_max */
+},
+{
+	"nsc_ipcname (sem)",	/* Service name */
+	NAME_SERVICE_SEM,	/* Service number */
+	{0},			/* Active list */
+	&nsc_name_iodb_ops,
+	SEMMNI,			/* size */
+	1,			/* seq */
+	USHRT_MAX,		/* seq_max */
+},
+{	"nsc_ipcname (shm)",	/* Service name */
+	NAME_SERVICE_SHM,	/* Service number */
+	{0},			/* Active list */
+	&nsc_name_iodb_ops,
+	SHMMNI,			/* size */
+	1,			/* seq */
+	USHRT_MAX,		/* seq_max */
+}
+};
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 ipc_obj_db_t nsc_name_odb[NAME_SERVICE_MAX] =
 {
 {	"nsc_ipcname (msg)",	/* Service name */
@@ -122,6 +152,7 @@
 	0			/* in_use */
 }
 };
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 
 
 ATOMIC_INT_T local_nameserver_version;
@@ -148,6 +179,14 @@
 void ssi_growary(void *argp, void *datap, int datalen)
 {
 	ipc_obj_db_t	*odbp;
+	int *new_size = (int *)datap;
+
+	/* SSI_XXX: New size is bigger than static object table? */
+	if (*new_size > IPCMNI) {
+		printk(KERN_WARNING "%s: new size %d too big not supported\n",
+				__FUNCTION__, *new_size);
+		*new_size = IPCMNI;
+	}
 
 	/*
 	 * Re-allocate the nodehint array corresponding to the change in the
@@ -179,12 +218,100 @@
 		 * failure handling code here.
 		 */
 		odbp = &nsc_name_odb[NAME_SERVICE_SEM];
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+		down(&odbp->iodb_sem);
+		clreg_update_int(&odbp->iodb_size, datap, datalen);
+		up(&odbp->iodb_sem);
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 		NSC_IPC_WRLOCK(odbp);
 		clreg_update_int(&odbp->iodb_size, datap, datalen);
 		NSC_IPC_WRUNLOCK(odbp);
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 	}
 }
 
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+static int
+objsvr_check_idx(
+	ipc_obj_db_t	*objdbp,	/* Object type's database. */
+	int		idx)
+{
+	ipc_obj_t *objp;
+
+	rcu_read_lock();
+	objp = rcu_dereference(objdbp->iodb_active[idx]);
+	if (objp && !atomic_read(&objp->io_refcnt))
+		objp = NULL;
+	rcu_read_unlock();
+
+	return (objp != NULL);
+}
+
+/*
+ * Scans an object manager database for a valid object with the specified id.
+ * Returns true if found.
+ */
+static int
+objsvr_check_id(
+	ipc_obj_db_t	*objdbp,	/* Object type's database. */
+	int		id)
+{
+	int idx = id % SEQ_MULTIPLIER;
+	if (idx >= objdbp->iodb_size)
+		return 0;
+	return objsvr_check_idx(objdbp, idx);
+}
+
+/*
+ * Scans an object manager database for a valid object with the specified id.
+ * If found, it is held, locked, and returned.
+ */
+ipc_obj_t *
+objsvr_find_id(
+	ipc_obj_db_t	*objdbp,	/* Object type's database. */
+	int		id)
+{
+	ipc_obj_t *objp;
+	int idx;
+
+	idx = id % SEQ_MULTIPLIER;
+	if (idx >= objdbp->iodb_size)
+		return NULL;
+
+	rcu_read_lock();
+	objp = rcu_dereference(objdbp->iodb_active[idx]);
+	if (objp && !atomic_test_and_add(1, &objp->io_refcnt))
+		objp = NULL;
+	rcu_read_unlock();
+
+	return objp;
+}
+
+static void
+ipc_obj_dealloc(
+	struct rcu_head *io_rcu)
+{
+	ipc_obj_t *objp = container_of(io_rcu, ipc_obj_t, io_rcu);
+	ipc_obj_db_t *objdbp = objp->io_obj_dbp;
+
+	(*objdbp->iodb_ops->iop_dealloc)(objp->io_perm);
+	kfree(objp);
+}
+
+static void
+objsvr_put(ipc_obj_t *objp)
+{
+	ipc_obj_db_t *objdbp = objp->io_obj_dbp;
+
+	if (!atomic_dec_and_lock(&objp->io_refcnt, &objdbp->iodb_lock))
+		return;
+	objdbp->iodb_active[objp->iodb_active_idx] = NULL;
+	spin_unlock(&objdbp->iodb_lock);
+
+	set_mb(objdbp->iodb_timestamp, jiffies);
+	call_rcu(&objp->io_rcu, ipc_obj_dealloc);
+}
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 /*
  * Scans an object manager database for a valid object with the specified id.
  * If found, it is held, locked, and returned.
@@ -202,6 +329,7 @@
 	else
 		return (objdbp->iodb_active[idx]);
 }
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 
 /*
  * Called at the potential object server for a new object.
@@ -221,8 +349,117 @@
 	int		view,
 	int		nmsvr_version)
 {
-	int	error, idx=0, newid=id;
 	ipc_perm_t *permp;
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	ipc_obj_t *objp, *old_objp;
+	int objdb_locked, error, idx;
+	extern ATOMIC_INT_T local_nameserver_version;
+
+	*new_objpp = NULL;
+
+	(*objdbp->iodb_ops->iop_alloc)(&permp);
+	if (permp == NULL)
+		return -ENOMEM;
+	permp->mode = (flag & S_IRWXUGO);
+	permp->key = key;
+	permp->cuid = permp->uid = current->euid;
+	permp->cgid = permp->gid = current->egid;
+
+	objp = (typeof(objp)) kmalloc(sizeof(*objp), GFP_KERNEL);
+	if (objp == NULL) {
+		(*objdbp->iodb_ops->iop_dealloc)(permp);
+		return -ENOMEM;
+	}
+
+	objdb_locked = sem_owned(&objdbp->iodb_sem);
+	if (!objdb_locked)
+		down(&objdbp->iodb_sem);
+
+	/*
+	 * If name_server has gone down while this id was in flight to us,
+	 * or if the id has been claimed by another proc after being
+	 * gen'd, we have to try again
+	 */
+	if (id >= 0) {
+		if (nmsvr_version != READ_ATOMIC_INT(&local_nameserver_version)
+		    || objsvr_check_id(objdbp, id)) {
+#ifdef TEST_IPC
+			printk(KERN_WARNING "%s: object pointer is NULL,"
+				" returning -EAGAIN\n", __FUNCTION__);
+#endif
+			error = -EAGAIN;
+			goto out;
+		}
+	}
+
+	/*
+	 *  Check to see whether the object was created while we slept
+	 */
+	old_objp = objsvr_find_key(objdbp, key, view, svr_node);
+	if (old_objp) {
+		*new_objpp = old_objp;
+
+		kfree(objp);
+		(*objdbp->iodb_ops->iop_dealloc)(permp);
+#ifdef TEST_IPC
+		printk("%s\n","find key not null, returning objpp");
+#endif
+		error = 0;
+		goto out;
+	}
+
+	if (id >= 0) {
+		idx = id % SEQ_MULTIPLIER;
+	} else {
+		for (idx=0; idx < objdbp->iodb_size; idx++) {
+			if (!objsvr_check_idx(objdbp, idx))
+				break;
+		}
+	}
+	if (idx >= objdbp->iodb_size) {
+		error = -ENOSPC;
+		goto out;
+	}
+
+	*new_objpp = objp;
+
+	atomic_set(&objp->io_refcnt, 2); /* active + new_objpp */
+	if (id >= 0)
+		objp->io_id = id;
+	if (id < 0) {
+		objdbp->iodb_seq++;
+		if (objdbp->iodb_seq > objdbp->iodb_seq_max)
+			objdbp->iodb_seq = 1;
+		objp->io_id = /* ipc_buildid() */
+			objdbp->iodb_seq * SEQ_MULTIPLIER + idx;
+	}
+	objp->io_svr_node = svr_node;
+	objp->io_local_view = view;
+	/* objp->io_size */
+	objp->io_perm = permp;
+	objp->io_obj_dbp = objdbp;
+	objp->iodb_active_idx = idx;
+	INIT_RCU_HEAD(&objp->io_rcu);
+
+	spin_lock(&objdbp->iodb_lock);
+	rcu_assign_pointer(objdbp->iodb_active[idx], objp);
+	spin_unlock(&objdbp->iodb_lock);
+
+	objdbp->iodb_timestamp = jiffies;
+	error = 0;
+#ifdef TEST_IPC
+	printk("%s\n","new object registered");
+#endif
+out:
+	if (!objdb_locked)
+		up(&objdbp->iodb_sem);
+	if (error) {
+		kfree(objp);
+		(*objdbp->iodb_ops->iop_dealloc)(permp);
+	}
+	return error;
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
+	int error, idx=0, newid=id;
 	extern ATOMIC_INT_T local_nameserver_version;
 
 	error = -ENOSPC;
@@ -307,6 +544,7 @@
 #endif
 out:
 	return error;
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 }
 
 /*
@@ -319,6 +557,31 @@
 	int		view,
 	clusternode_t	node)
 {
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	ipc_obj_t *objp;
+	int idx;
+	char found = 0;
+
+	if (key == IPC_PRIVATE)
+		return NULL;
+
+	rcu_read_lock();
+	for (idx = 0; idx < objdbp->iodb_size; idx++) {
+		objp = rcu_dereference(objdbp->iodb_active[idx]);
+		if (!objp || objp->io_perm->key != key)
+			continue;
+		if (view && objp->io_svr_node != node)
+			continue;
+		if (view ^ objp->io_local_view)
+			continue;
+		if (atomic_test_and_add(1, &objp->io_refcnt))
+			found = 1;
+		break;
+	}
+	rcu_read_unlock();
+
+	return found ? objp : NULL;
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 	ipc_obj_t **objp = objdbp->iodb_active;
 	int idx;
 
@@ -336,9 +599,10 @@
 		}
 	}
 	if (idx < objdbp->iodb_size)
-		return objp[idx];
+		return objp[idx]; /* Caller holds NSC_IPC_RD/WRLOCK */
 	else
 		return NULL;
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 }
 
 /*
@@ -353,8 +617,14 @@
 	printk("%s\n","Name server registered with clms");
 	for (i = 0; i < NAME_SERVICE_MAX; i++) {
 		nsc_name_odb[i].iodb_service = i;
-		INIT_RW_LOCK(&nsc_name_odb[i].rwlock);
 		nsc_name_odb[i].iodb_seq = 1;
+		nsc_name_odb[i].iodb_timestamp = 0;
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+		init_MUTEX(&nsc_name_odb[i].iodb_sem);
+		spin_lock_init(&nsc_name_odb[i].iodb_lock);
+#else
+		INIT_RW_LOCK(&nsc_name_odb[i].rwlock);
+#endif
 	}
 	INIT_ATOMIC_INT(&local_nameserver_version, 1);
 }
@@ -427,6 +697,7 @@
 	return 0;
 }
 
+#ifdef SSI_NOTUSED
 int
 ipcname_genid(ipc_obj_db_t *odbp,
 	      key_t key,
@@ -451,13 +722,18 @@
 	 * lookup if this id exist
 	 */
 	if (id >= 0) {
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+		if (!objsvr_check_id(odbp, id))
+#else
 		if (objsvr_find_id(odbp, id) == NULL)
+#endif
 			*glid = id;
 		else
 			id = -EINVAL;
 	}
 	return id;
 }
+#endif /* SSI_NOUSED */
 
 inline int
 obj_checks(ipc_obj_t *objp, int size)
@@ -495,8 +771,10 @@
 #endif
 	extern ATOMIC_INT_T local_nameserver_version;
 
-
 	odbp = &nsc_name_odb[service];
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	objp = NULL;
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 #ifdef NSC_IPC_RWLOCK_DOWNGRADE
 	NSC_IPC_RDLOCK(odbp);
 #else
@@ -505,6 +783,7 @@
 #ifdef TEST_IPC
 	printk("%s\n","Message nameserver locked in getid");
 #endif
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 	if ((key != IPC_PRIVATE) &&
 	    (objp = objsvr_find_key(odbp, key, view, *server)) != NULL)
 	{
@@ -527,7 +806,11 @@
 
 		*glid    = objp->io_id;
 		*size = 0;
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+		*server = objp->io_svr_node;
+#else
 		*server = objp->svr_node;
+#endif
 		goto out;
 	}
 	if ((key != IPC_PRIVATE) && !(in_flag & IPC_CREAT)) {
@@ -538,12 +821,16 @@
 		goto out;
 	}
 
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	WARN_ON(*glid != -1);
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 #ifdef NSC_IPC_RWLOCK_DOWNGRADE
 	NSC_IPC_RDUNLOCK(odbp);
 	wrlock = 1;
 #else
 	NSC_IPC_WRUNLOCK(odbp);
 #endif
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 
 	error = objsvr_new(odbp, &objp, key, *glid, in_flag, *server, view,
 			   READ_ATOMIC_INT(&local_nameserver_version));
@@ -551,6 +838,14 @@
 	printk("%s\n","returning from nameserver registration");
 #endif
 	if (!error) {
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+		SSI_ASSERT(objp);
+		if (odbp->iodb_service != NAME_SERVICE_MSG)
+			objp->io_size = *size;
+		smp_rmb();
+		*size = 1;
+		*server  = objp->io_svr_node;
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 		if (objp == NULL) {
 			if ((objp = objsvr_find_key(odbp, key, view, *server)) == NULL) {
 				printk("%s\n","objsvr_find_key is NULL");
@@ -566,6 +861,7 @@
 			*size = 1;
 		}
 		*server  = objp->svr_node;
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 		*glid = objp->io_id;
 
 		/*
@@ -579,6 +875,13 @@
 #endif
 	}
 out:
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	if (objp)
+		objsvr_put(objp);
+#ifdef TEST_IPC
+	printk("Returning glid %d\n",*glid);
+#endif
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 #ifdef NSC_IPC_RWLOCK_DOWNGRADE
 	if (!wrlock)
 		NSC_IPC_RDUNLOCK(odbp);
@@ -589,6 +892,7 @@
 	printk("%s\n","name server spin lock dropped in out/getid");
 	printk("Returning glid %d\n",*glid);
 #endif
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 	return error;
 }
 int
@@ -598,7 +902,25 @@
 {
 	ipc_obj_db_t	*odbp;
 	int		idx, count = -1;
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	ipc_obj_t	*objp;
+
+	odbp = &nsc_name_odb[service];
 
+	rcu_read_lock();
+	for (idx = 0; idx < odbp->iodb_size; idx++) {
+		objp = rcu_dereference(odbp->iodb_active[idx]);
+		if (!objp || !atomic_read(&objp->io_refcnt))
+			continue;
+		if (++count >= *size)
+			continue;
+		node_id_pairs[count].node_num = objp->io_svr_node;
+		node_id_pairs[count].ipc_id = objp->io_id;
+	}
+	rcu_read_unlock();
+	smp_rmb();
+	*size = count + 1;
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 	odbp = &nsc_name_odb[service];
 #ifdef NSC_IPC_RWLOCK_DOWNGRADE
 	NSC_IPC_RDLOCK(odbp);
@@ -623,6 +945,7 @@
 	*size = count + 1;
 	NSC_IPC_WRUNLOCK(odbp);
 #endif
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 #ifdef TEST_IPC
 	printk("%s\n","name server spin lock dropped in out/getid");
 #endif
@@ -637,9 +960,25 @@
 int
 ipcname_rmid(int service, global_id_t glid, int flag)
 {
-	ipc_obj_db_t *odbp;
 	int error;
 
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	(void) flag;
+
+	switch(service) {
+		case NAME_SERVICE_SHM:
+		case NAME_SERVICE_MSG:
+		case NAME_SERVICE_SEM:
+			error = nsc_ipcremove(&nsc_name_odb[service], glid);
+			if (error > 0)
+				error = 0;
+			break;
+		default:
+			return -EINVAL;
+	}
+#else
+	ipc_obj_db_t *odbp;
+
 	switch(service) {
 		case NAME_SERVICE_SHM:
 		case NAME_SERVICE_MSG:
@@ -655,6 +994,7 @@
 		default:
 			return -EINVAL;
 	}
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 	return error;
 }
 
@@ -664,6 +1004,23 @@
 {
 	ipc_obj_db_t	*odbp;
 	ipc_obj_t	*objp;
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	odbp = &nsc_name_odb[service];
+
+	objp = objsvr_find_id(odbp, glid);
+	if (objp == NULL)
+		return -EINVAL;
+
+	if (service == NAME_SERVICE_SHM)
+		*sz = objp->io_size;
+	*key = objp->io_perm->key;
+	*flags = objp->io_perm->mode;
+	*svr = objp->io_svr_node;
+	*view = objp->io_local_view;
+
+	objsvr_put(objp);
+	return 0;
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 	int error = 0;
 
 	odbp = &nsc_name_odb[service];
@@ -698,6 +1055,7 @@
 	NSC_IPC_WRUNLOCK(odbp);
 #endif
 	return error;
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 }
 
 int
@@ -715,6 +1073,28 @@
 	int		total, idx, count;
 
 	odbp = &nsc_name_odb[service];
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	total = idx = 0;
+
+	rcu_read_lock();
+	for (count = 0; count < odbp->iodb_size; count++) {
+		objp = rcu_dereference(odbp->iodb_active[count]);
+		if (!objp || !atomic_read(&objp->io_refcnt))
+			continue;
+		if (server && server != objp->io_svr_node)
+			continue;
+		total++;
+		if (start < total && idx < nentries) {
+			*idsp++ = objp->io_id;
+			idx++;
+		}
+	}
+	rcu_read_unlock();
+
+	*num_entriesp = idx;
+	*total_entriesp = total;
+	*timestampp = odbp->iodb_timestamp;
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 	NSC_IPC_RDLOCK(odbp);
 
 	total = idx = 0;
@@ -734,6 +1114,7 @@
 	*timestampp = odbp->iodb_timestamp;
 
 	NSC_IPC_RDUNLOCK(odbp);
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 
 	return 0;
 }
@@ -846,6 +1227,10 @@
 	char *buffer,
 	int data_len)
 {
+	ipc_obj_t *objp;
+	key_t key;
+	clusternode_t nodenum;
+	int id, view, mode, size;
 	int i, j, objects, error;
 	extern ipc_obj_db_t nsc_name_odb[NAME_SERVICE_MAX];
 	extern ATOMIC_INT_T local_nameserver_version;
@@ -853,14 +1238,10 @@
 	for (i = 0; i < NAME_SERVICE_MAX; i++) {
 		objects = *((int *)buffer);
 		buffer += sizeof(int);
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+		down(&nsc_name_odb[i].iodb_sem);
+#endif
 		for (j = 0; j < objects; j++) {
-			ipc_obj_t *objp;
-			int id, view;
-			key_t key;
-			int mode;
-			int size;
-			clusternode_t nodenum;
-
 			id = *((int *)buffer);
 			buffer += sizeof(int);
 			key = *((key_t *)buffer);
@@ -873,6 +1254,8 @@
 			buffer += sizeof(int);
 			nodenum = *((int *)buffer);
 			buffer += sizeof(int);
+
+			objp = NULL;
 			error = objsvr_new(&nsc_name_odb[i],
 					   &objp,
 					   (int)key,
@@ -881,6 +1264,18 @@
 					   nodenum,
 					   view,
 				(int)READ_ATOMIC_INT(&local_nameserver_version));
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+			if (!error && size && i != NAME_SERVICE_MSG) {
+				BUG_ON(!objp);
+				objp->io_size = size;
+			}
+			if (objp)
+				objsvr_put(objp);
+
+			/* Ensure local db seqnum starts at highest in use */
+			if ((id / SEQ_MULTIPLIER) > nsc_name_odb[i].iodb_seq)
+				nsc_name_odb[i].iodb_seq = id/SEQ_MULTIPLIER;
+#else
 			if (!error && size && (i != NAME_SERVICE_MSG))
 				if (!objp)
 					objp = objsvr_find_id(&nsc_name_odb[i],
@@ -892,6 +1287,7 @@
 				nsc_name_odb[i].iodb_seq = id/SEQ_MULTIPLIER;
 
 			NSC_IPC_WRUNLOCK(&nsc_name_odb[i]);
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 			if (error < 0)
 			    printk(KERN_WARNING "ipcname_failover_data: "
 				"error %d attemting to recover id %d, key %d"
@@ -899,8 +1295,12 @@
 				" lost or invalid.\n", error, id, key, i);
 		}
 		nsc_name_odb[i].iodb_seq++;
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+		up(&nsc_name_odb[i].iodb_sem);
+#endif
 	}
 }
+
 /*
  * ipcname_nodedown()
  *
@@ -921,6 +1321,20 @@
 
 	for (service = 0; service < NAME_SERVICE_MAX; service++) {
 		odbp = &nsc_name_odb[service];
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+		rcu_read_lock();
+		for (idx = 0; idx < odbp->iodb_size; idx++) {
+			op = rcu_dereference(odbp->iodb_active[idx]);
+			if (!op || op->io_svr_node != node)
+				continue;
+			if (!atomic_test_and_add(1, &op->io_refcnt))
+				continue;
+			objsvr_put(op); /* active */
+			objsvr_put(op);
+		}
+		rcu_read_unlock();
+	}
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 #ifdef NSC_IPC_RWLOCK_DOWNGRADE
 again:
 #endif
@@ -950,7 +1364,9 @@
 		}
 		NSC_IPC_RDUNLOCK(odbp);
 	}
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 }
+
 /*
  * Called when the service specific creation of a new object
  * fails (after created by nsc_ipcget).  Called with a locked,
@@ -966,7 +1382,29 @@
 	ipc_obj_t *op;
 	clusternode_t svr_node;
 	int idx;
+#ifdef SSI_IPC_OBJ_DB_LOCKLESS
+	char found = 0;
+
+	rcu_read_lock();
+	for (idx = 0; idx < objdbp->iodb_size; idx++) {
+		op = rcu_dereference(objdbp->iodb_active[idx]);
+		if (!op || op->io_id != objid)
+			continue;
+		if (!atomic_test_and_add(1, &op->io_refcnt))
+			continue;
+		found = 1;
+		break;
+	}
+	rcu_read_unlock();
+	if (!found)
+		return -ESRCH;
 
+	svr_node = op->io_svr_node;
+	objsvr_put(op); /* active */
+	objsvr_put(op);
+
+	return (svr_node == name_server_node);
+#else /* SSI_IPC_OBJ_DB_LOCKLESS */
 	for (idx=0; idx<objdbp->iodb_size; idx++) {
 		if((op=objdbp->iodb_active[idx])==NULL)
 			continue;
@@ -988,6 +1426,7 @@
 		}
 	}
 	return -ESRCH;
+#endif /* !SSI_IPC_OBJ_DB_LOCKLESS */
 }
 
 void

Index: namesvr_clnt.c
===================================================================
RCS file: /cvsroot/ssic-linux/openssi/kernel/cluster/ssi/ipc/namesvr_clnt.c,v
retrieving revision 1.8
retrieving revision 1.9
diff -u -d -r1.8 -r1.9
--- namesvr_clnt.c	27 Oct 2009 03:18:29 -0000	1.8
+++ namesvr_clnt.c	5 Mar 2010 06:25:36 -0000	1.9
@@ -71,11 +71,11 @@
 int cli_ipcname_getid(int service, key_t key, int in_flag,
 		     global_id_t *glid, clusternode_t *server, int view, int *sz)
 {
-	int		status;
-	clusternode_t	server_node;
-	int 		rval;
 	u_long		newid;
+	clusternode_t	server_node;
+	int		status, rval;
 	extern clms_key_svc_t ipc_key_service;
+
 again:
 	server_node = name_server_node;
 #ifdef TEST_IPC


------------------------------------------------------------------------------
Download Intel&#174; Parallel Studio Eval
Try the new software tools for yourself. Speed compiling, find bugs
proactively, and fine-tune applications for parallel performance.
See why Intel Parallel Studio got high marks during beta.
http://p.sf.net/sfu/intel-sw-dev