Skip to contentSkip to conversation

memory leak when mounting and unmounting -onolock filesystems

PATCH In review Neil Brown 2 participants 3 messages Dec 11, 2007 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore

Neil Brown

author
#0 lore raw

PATCH 1/1 memory leak when mounting and unmounting -onolock filesystems

Hi Trond,

We found that a machine which made moderately heavy use of 'automount' was leaking some nfs data structures - particularly the 4K allocated by rpc_alloc_iostats.
It turns out that this only happens with filesystems with -onolock set.

The problem is that if NFS_MOUNT_NONLM is set, nfs_start_lockd doesn't set server->destroy, so when the filesystem is unmounted, the ->client_acl is not shutdown, and so several resources are still held. Multiple mount/umount cycles will slowly eat away memory several pages at a time.

The following patch seems the simplest fix, but it may not be what you prefer. Arguably nfs_start_lockd isn't really the right place to set server->destroy as it relates to both lockd and client_acl. Maybe ->destroy could be set at the end of nfs_init_server in the no-error case.

It looks like this bug was introduced by
54ceac4515986030c2502960be620198dd8fe25b
so has been present since 2.6.19 (we found it in 2.6.22).

NeilBrown

fs/nfs/client.c +2 -2
diff .prev/fs/nfs/client.c ./fs/nfs/client.c
@@ -432,9 +432,9 @@ static int nfs_start_lockd(struct nfs_se
 			IPPROTO_TCP : IPPROTO_UDP);
 	if (error < 0)
 		server->flags |= NFS_MOUNT_NONLM;
-	else
-		server->destroy = nfs_destroy_server;
 out:
+	if (! error)
+		server->destroy = nfs_destroy_server;
 	return error;
 }
 

A message that is not in the archive (sent off-list or not yet received): [email protected]

#2 lore raw
Quoted 5 lines · Neil Brown

The problem is that if NFS_MOUNT_NONLM is set, nfs_start_lockd doesn't set server->destroy, so when the filesystem is unmounted, the ->client_acl is not shutdown, and so several resources are still held. Multiple mount/umount cycles will slowly eat away memory several pages at a time.

Hi Neil,

How about the following alternative for 2.6.24, and then we plan on getting rid of ->destroy() altogether in 2.6.25?

Cheers
Trond
---------------------------------------------------
commit 2c29e75b2908f760c45bbf934f81b8a8af867770

Author: Trond Myklebust <[email protected]>
Date:   Tue Dec 11 11:05:19 2007 -0500

NFSv2/v3: Fix a memory leak when using -onolock

Neil Brown said:
> Hi Trond,
>
> We found that a machine which made moderately heavy use of > 'automount' was leaking some nfs data structures - particularly the > 4K allocated by rpc_alloc_iostats.
> It turns out that this only happens with filesystems with -onolock > set.

> The problem is that if NFS_MOUNT_NONLM is set, nfs_start_lockd doesn't > set server->destroy, so when the filesystem is unmounted, the > ->client_acl is not shutdown, and so several resources are still > held. Multiple mount/umount cycles will slowly eat away memory > several pages at a time.

fs/nfs/client.c +3 -3
index 70587f3..a6f6254 100644
@@ -410,9 +410,6 @@ static int nfs_create_rpc_client(struct nfs_client *clp, int proto,
  */
 static void nfs_destroy_server(struct nfs_server *server)
 {
-	if (!IS_ERR(server->client_acl))
-		rpc_shutdown_client(server->client_acl);
-
 	if (!(server->flags & NFS_MOUNT_NONLM))
 		lockd_down();	/* release rpc.lockd */
 }
@@ -755,6 +752,9 @@ void nfs_free_server(struct nfs_server *server)
 
 	if (server->destroy != NULL)
 		server->destroy(server);
+
+	if (!IS_ERR(server->client_acl))
+		rpc_shutdown_client(server->client_acl);
 	if (!IS_ERR(server->client))
 		rpc_shutdown_client(server->client);
 
author Acked-by
#3 lore raw
Quoted 7 lines · On Tuesday December 11, [email protected] wrote:
Quoted 5 lines · Neil Brown

The problem is that if NFS_MOUNT_NONLM is set, nfs_start_lockd doesn't set server->destroy, so when the filesystem is unmounted, the ->client_acl is not shutdown, and so several resources are still held. Multiple mount/umount cycles will slowly eat away memory several pages at a time.

Quoted 5 lines

Hi Neil,

How about the following alternative for 2.6.24, and then we plan on getting rid of ->destroy() altogether in 2.6.25?

Yep, that looks good. Thanks.

NeilBrown

Quoted 51 lines

Cheers
Trond
---------------------------------------------------
commit 2c29e75b2908f760c45bbf934f81b8a8af867770

Author: Trond Myklebust <[email protected]>
Date:   Tue Dec 11 11:05:19 2007 -0500

NFSv2/v3: Fix a memory leak when using -onolock

Neil Brown said:
> Hi Trond,
>
> We found that a machine which made moderately heavy use of > 'automount' was leaking some nfs data structures - particularly the > 4K allocated by rpc_alloc_iostats.
> It turns out that this only happens with filesystems with -onolock > set.

> The problem is that if NFS_MOUNT_NONLM is set, nfs_start_lockd doesn't > set server->destroy, so when the filesystem is unmounted, the > ->client_acl is not shutdown, and so several resources are still > held. Multiple mount/umount cycles will slowly eat away memory > several pages at a time.

fs/nfs/client.c +3 -3
index 70587f3..a6f6254 100644
@@ -410,9 +410,6 @@ static int nfs_create_rpc_client(struct nfs_client *clp, int proto,
  */
 static void nfs_destroy_server(struct nfs_server *server)
 {
-	if (!IS_ERR(server->client_acl))
-		rpc_shutdown_client(server->client_acl);
-
 	if (!(server->flags & NFS_MOUNT_NONLM))
 		lockd_down();	/* release rpc.lockd */
 }
@@ -755,6 +752,9 @@ void nfs_free_server(struct nfs_server *server)
 
 	if (server->destroy != NULL)
 		server->destroy(server);
+
+	if (!IS_ERR(server->client_acl))
+		rpc_shutdown_client(server->client_acl);
 	if (!IS_ERR(server->client))
 		rpc_shutdown_client(server->client);