This series fixes races between PipeFS mount/umount notification calls and
SUNRPC clients creation and destruction.
https://bugzilla.redhat.com/show_bug.cgi?id=924649
v2: Fixed few silly locking bugs.
The following series implements...
---
Stanislav Kinsbursky (4):
SUNRPC: split client creation routine into setup and registration
SUNRPC: fix races on PipeFS MOUNT notifications
SUNRPC: fix races on PipeFS UMOUNT notifications
SUNRPC: PipeFS MOUNT notification optimization for dying clients
net/sunrpc/clnt.c | 71 ++++++++++++++++++++++++++++++-------------------
net/sunrpc/rpc_pipe.c | 5 +++
2 files changed, 48 insertions(+), 28 deletions(-)
This helper moves all "registration" code to the new rpc_client_register()
helper.
This helper will be used later in the series to synchronize against PipeFS
MOUNT/UMOUNT events.
Signed-off-by: Stanislav Kinsbursky <[email protected]>
---
net/sunrpc/clnt.c | 48 ++++++++++++++++++++++++++++++++----------------
1 files changed, 32 insertions(+), 16 deletions(-)
diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
index 5a750b9..4bf2409 100644
--- a/net/sunrpc/clnt.c
+++ b/net/sunrpc/clnt.c
@@ -289,12 +289,39 @@ static void rpc_clnt_set_nodename(struct rpc_clnt *clnt, const char *nodename)
memcpy(clnt->cl_nodename, nodename, clnt->cl_nodelen);
}
+static int rpc_client_register(const struct rpc_create_args *args,
+ struct rpc_clnt *clnt)
+{
+ const struct rpc_program *program = args->program;
+ struct rpc_auth *auth;
+ int err = 0;
+
+ err = rpc_setup_pipedir(clnt, program->pipe_dir_name);
+ if (err < 0)
+ goto out;
+
+ auth = rpcauth_create(args->authflavor, clnt);
+ if (IS_ERR(auth)) {
+ dprintk("RPC: Couldn't create auth handle (flavor %u)\n",
+ args->authflavor);
+ err = PTR_ERR(auth);
+ goto err_auth;
+ }
+
+ rpc_register_client(clnt);
+out:
+ return err;
+
+err_auth:
+ rpc_clnt_remove_pipedir(clnt);
+ goto out;
+}
+
static struct rpc_clnt * rpc_new_client(const struct rpc_create_args *args, struct rpc_xprt *xprt)
{
const struct rpc_program *program = args->program;
const struct rpc_version *version;
struct rpc_clnt *clnt = NULL;
- struct rpc_auth *auth;
int err;
/* sanity check the name before trying to print it */
@@ -354,25 +381,14 @@ static struct rpc_clnt * rpc_new_client(const struct rpc_create_args *args, stru
atomic_set(&clnt->cl_count, 1);
- err = rpc_setup_pipedir(clnt, program->pipe_dir_name);
- if (err < 0)
- goto out_no_path;
-
- auth = rpcauth_create(args->authflavor, clnt);
- if (IS_ERR(auth)) {
- dprintk("RPC: Couldn't create auth handle (flavor %u)\n",
- args->authflavor);
- err = PTR_ERR(auth);
- goto out_no_auth;
- }
-
/* save the nodename */
rpc_clnt_set_nodename(clnt, utsname()->nodename);
- rpc_register_client(clnt);
+
+ err = rpc_client_register(args, clnt);
+ if (err)
+ goto out_no_path;
return clnt;
-out_no_auth:
- rpc_clnt_remove_pipedir(clnt);
out_no_path:
kfree(clnt->cl_principal);
out_no_principal:
Below are races, when RPC client can be created without PiepFS dentries
CPU#0 CPU#1
----------------------------- -----------------------------
rpc_new_client rpc_fill_super
rpc_setup_pipedir
mutex_lock(&sn->pipefs_sb_lock)
rpc_get_sb_net == NULL
(no per-net PipeFS superblock)
sn->pipefs_sb = sb;
notifier_call_chain(MOUNT)
(client is not in the list)
rpc_register_client
(client without pipes dentries)
To fix this patch:
1) makes PipeFS mount notification call with pipefs_sb_lock being held.
2) releases pipefs_sb_lock on new SUNRPC client creation only after
registration.
Signed-off-by: Stanislav Kinsbursky <[email protected]>
Cc: [email protected]
---
net/sunrpc/clnt.c | 24 +++++++++++++-----------
net/sunrpc/rpc_pipe.c | 3 +++
2 files changed, 16 insertions(+), 11 deletions(-)
diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
index 4bf2409..48888ee 100644
--- a/net/sunrpc/clnt.c
+++ b/net/sunrpc/clnt.c
@@ -157,20 +157,15 @@ static struct dentry *rpc_setup_pipedir_sb(struct super_block *sb,
}
static int
-rpc_setup_pipedir(struct rpc_clnt *clnt, const char *dir_name)
+rpc_setup_pipedir(struct rpc_clnt *clnt, const char *dir_name,
+ struct super_block *pipefs_sb)
{
- struct net *net = rpc_net_ns(clnt);
- struct super_block *pipefs_sb;
struct dentry *dentry;
clnt->cl_dentry = NULL;
if (dir_name == NULL)
return 0;
- pipefs_sb = rpc_get_sb_net(net);
- if (!pipefs_sb)
- return 0;
dentry = rpc_setup_pipedir_sb(pipefs_sb, clnt, dir_name);
- rpc_put_sb_net(net);
if (IS_ERR(dentry))
return PTR_ERR(dentry);
clnt->cl_dentry = dentry;
@@ -294,11 +289,16 @@ static int rpc_client_register(const struct rpc_create_args *args,
{
const struct rpc_program *program = args->program;
struct rpc_auth *auth;
+ struct net *net = rpc_net_ns(clnt);
+ struct super_block *pipefs_sb;
int err = 0;
- err = rpc_setup_pipedir(clnt, program->pipe_dir_name);
- if (err < 0)
- goto out;
+ pipefs_sb = rpc_get_sb_net(net);
+ if (pipefs_sb) {
+ err = rpc_setup_pipedir(clnt, program->pipe_dir_name, pipefs_sb);
+ if (err)
+ goto out;
+ }
auth = rpcauth_create(args->authflavor, clnt);
if (IS_ERR(auth)) {
@@ -310,10 +310,12 @@ static int rpc_client_register(const struct rpc_create_args *args,
rpc_register_client(clnt);
out:
+ if (pipefs_sb)
+ rpc_put_sb_net(net);
return err;
err_auth:
- rpc_clnt_remove_pipedir(clnt);
+ __rpc_clnt_remove_pipedir(clnt);
goto out;
}
diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index e7ce4b3..c512448 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -1126,6 +1126,7 @@ rpc_fill_super(struct super_block *sb, void *data, int silent)
return -ENOMEM;
dprintk("RPC: sending pipefs MOUNT notification for net %p%s\n",
net, NET_NAME(net));
+ mutex_lock(&sn->pipefs_sb_lock);
sn->pipefs_sb = sb;
err = blocking_notifier_call_chain(&rpc_pipefs_notifier_list,
RPC_PIPEFS_MOUNT,
@@ -1133,6 +1134,7 @@ rpc_fill_super(struct super_block *sb, void *data, int silent)
if (err)
goto err_depopulate;
sb->s_fs_info = get_net(net);
+ mutex_unlock(&sn->pipefs_sb_lock);
return 0;
err_depopulate:
@@ -1141,6 +1143,7 @@ err_depopulate:
sb);
sn->pipefs_sb = NULL;
__rpc_depopulate(root, files, RPCAUTH_lockd, RPCAUTH_RootEOF);
+ mutex_unlock(&sn->pipefs_sb_lock);
return err;
}
Not need to create pipes for dying client. So just skip them.
Note: we can safely dereference the client structure, because notification
caller is holding sn->pipefs_sb_lock.
Signed-off-by: Stanislav Kinsbursky <[email protected]>
Cc: [email protected]
---
net/sunrpc/clnt.c | 2 ++
1 files changed, 2 insertions(+), 0 deletions(-)
diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
index b4f1711..f0339ae 100644
--- a/net/sunrpc/clnt.c
+++ b/net/sunrpc/clnt.c
@@ -177,6 +177,8 @@ static inline int rpc_clnt_skip_event(struct rpc_clnt *clnt, unsigned long event
if (((event == RPC_PIPEFS_MOUNT) && clnt->cl_dentry) ||
((event == RPC_PIPEFS_UMOUNT) && !clnt->cl_dentry))
return 1;
+ if ((event == RPC_PIPEFS_MOUNT) && atomic_read(&clnt->cl_count) == 0)
+ return 1;
return 0;
}
CPU#0 CPU#1
----------------------------- -----------------------------
rpc_kill_sb
sn->pipefs_sb = NULL rpc_release_client
(UMOUNT_EVENT) rpc_free_auth
rpc_pipefs_event
rpc_get_client_for_event
!atomic_inc_not_zero(cl_count)
<skip the client>
atomic_inc(cl_count)
rpc_free_client
rpc_clnt_remove_pipedir
<skip client dir removing>
To fix this, this patch does the following:
1) Calls RPC_PIPEFS_UMOUNT notification with sn->pipefs_sb_lock being held.
2) Removes SUNRPC client from the list AFTER pipes destroying.
3) Doesn't hold RPC client on notification: if client in the list, then it
can't be destroyed while sn->pipefs_sb_lock in hold by notification caller.
Signed-off-by: Stanislav Kinsbursky <[email protected]>
Cc: [email protected]
---
net/sunrpc/clnt.c | 5 +----
net/sunrpc/rpc_pipe.c | 2 +-
2 files changed, 2 insertions(+), 5 deletions(-)
diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
index 48888ee..b4f1711 100644
--- a/net/sunrpc/clnt.c
+++ b/net/sunrpc/clnt.c
@@ -236,8 +236,6 @@ static struct rpc_clnt *rpc_get_client_for_event(struct net *net, int event)
continue;
if (rpc_clnt_skip_event(clnt, event))
continue;
- if (atomic_inc_not_zero(&clnt->cl_count) == 0)
- continue;
spin_unlock(&sn->rpc_client_lock);
return clnt;
}
@@ -254,7 +252,6 @@ static int rpc_pipefs_event(struct notifier_block *nb, unsigned long event,
while ((clnt = rpc_get_client_for_event(sb->s_fs_info, event))) {
error = __rpc_pipefs_event(clnt, event, sb);
- rpc_release_client(clnt);
if (error)
break;
}
@@ -655,8 +652,8 @@ rpc_free_client(struct rpc_clnt *clnt)
rcu_dereference(clnt->cl_xprt)->servername);
if (clnt->cl_parent != clnt)
rpc_release_client(clnt->cl_parent);
- rpc_unregister_client(clnt);
rpc_clnt_remove_pipedir(clnt);
+ rpc_unregister_client(clnt);
rpc_free_iostats(clnt->cl_metrics);
kfree(clnt->cl_principal);
clnt->cl_metrics = NULL;
diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index c512448..efca2f7 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -1165,7 +1165,6 @@ static void rpc_kill_sb(struct super_block *sb)
goto out;
}
sn->pipefs_sb = NULL;
- mutex_unlock(&sn->pipefs_sb_lock);
dprintk("RPC: sending pipefs UMOUNT notification for net %p%s\n",
net, NET_NAME(net));
blocking_notifier_call_chain(&rpc_pipefs_notifier_list,
@@ -1173,6 +1172,7 @@ static void rpc_kill_sb(struct super_block *sb)
sb);
put_net(net);
out:
+ mutex_unlock(&sn->pipefs_sb_lock);
kill_litter_super(sb);
}
On Tue, 2013-06-11 at 18:39 +0400, Stanislav Kinsbursky wrote:
> Below are races, when RPC client can be created without PiepFS dentries
>
> CPU#0 CPU#1
> ----------------------------- -----------------------------
> rpc_new_client rpc_fill_super
> rpc_setup_pipedir
> mutex_lock(&sn->pipefs_sb_lock)
> rpc_get_sb_net == NULL
> (no per-net PipeFS superblock)
> sn->pipefs_sb = sb;
> notifier_call_chain(MOUNT)
> (client is not in the list)
> rpc_register_client
> (client without pipes dentries)
>
> To fix this patch:
> 1) makes PipeFS mount notification call with pipefs_sb_lock being held.
> 2) releases pipefs_sb_lock on new SUNRPC client creation only after
> registration.
>
> Signed-off-by: Stanislav Kinsbursky <[email protected]>
> Cc: [email protected]
Hi Stanislav,
This isn't going to apply to the stable kernels without the cleanup
patch. Could you please reorganise this patch series so that the cleanup
comes last.
Thanks,
Trond
--
Trond Myklebust
Linux NFS client maintainer
NetApp
[email protected]
http://www.netapp.com
17.06.2013 22:20, Myklebust, Trond пишет:
> On Tue, 2013-06-11 at 18:39 +0400, Stanislav Kinsbursky wrote:
>> Below are races, when RPC client can be created without PiepFS dentries
>>
>> CPU#0 CPU#1
>> ----------------------------- -----------------------------
>> rpc_new_client rpc_fill_super
>> rpc_setup_pipedir
>> mutex_lock(&sn->pipefs_sb_lock)
>> rpc_get_sb_net == NULL
>> (no per-net PipeFS superblock)
>> sn->pipefs_sb = sb;
>> notifier_call_chain(MOUNT)
>> (client is not in the list)
>> rpc_register_client
>> (client without pipes dentries)
>>
>> To fix this patch:
>> 1) makes PipeFS mount notification call with pipefs_sb_lock being held.
>> 2) releases pipefs_sb_lock on new SUNRPC client creation only after
>> registration.
>>
>> Signed-off-by: Stanislav Kinsbursky <[email protected]>
>> Cc: [email protected]
>
> Hi Stanislav,
>
> This isn't going to apply to the stable kernels without the cleanup
> patch. Could you please reorganise this patch series so that the cleanup
> comes last.
>
> Thanks,
> Trond
>
>
Hello, Trond.
Sure, will do.
--
Best regards,
Stanislav Kinsbursky