[RFC] new client gssd upcall

[RFC] new client gssd upcall

[J. Bruce Fields]

The client needs a new more estensible text-based upcall to gssd, to
make it easier to support some features that are needed for new krb5
enctypes, for callbacks, etc.

We will have to continue to support the old upcall for backwards
compatibility with older gssd's.  To simplify upgrades, as well as
testing and debugging, it would help if we can upgrade gssd without
having to choose at boot (or module-load) time whether we want the new
or the old upcall.

That means that when gssd opens an rpc upcall pipe, we'd like to just
start using whichever version it chose immediately--ideally without
having to wait the 30 seconds that it would normally take for upcalls
already queued on the wrong upcall pipe to time out.

The following patches do that by queueing the upcalls on whichever of
the two upcall pipes (the new one and the old one) that somebody holds
open.  If nobody's opened anything yet, then upcalls are provisionally
queued on the new pipe.  I add a pipe_open method to allow the gss code
to keep track of which pipe is open, to prevent anyone from attempting
to open both pipes at once, and to cancel any upcalls that got queued to
the wrong pipe.

I did some minimal testing with the old upcall, but haven't tested the
new upcall yet.  (Olga, could you do that?)  So this is just an RFC at
this point.

--b.

[PATCH 1/5] rpc: remove unnecessary assignment

[J. Bruce Fields]

We're just about to kfree() gss_auth, so there's no point to setting any
of its fields.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index cc12d5f..b42ec85 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -665,7 +665,6 @@ static void
 gss_free(struct gss_auth *gss_auth)
 {
 	rpc_unlink(gss_auth->dentry);
-	gss_auth->dentry = NULL;
 	gss_mech_put(gss_auth->mech);
 
 	kfree(gss_auth);
-- 
1.5.5.rc1

[PATCH 2/5] rpc: Use separate spinlock for cred locking in auth_gss.c

[J. Bruce Fields]

The i_lock uses in auth_gss.c are to protect the gc_upcall and gc_ctx
fields of the gss cred; so we can use a separate per-auth spinlock for
those instead of the i_lock.  This simplifies the locking in the case
where we have two different pipes (one the current pipe, one the pipe
for a new more flexible upcall), either of which might be used for gss
upcalls.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |   56 +++++++++++++++++++++------------------
 1 files changed, 30 insertions(+), 26 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index b42ec85..28ca1c9 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -82,6 +82,7 @@ static const struct rpc_credops gss_nullops;
 struct gss_auth {
 	struct kref kref;
 	struct rpc_auth rpc_auth;
+	spinlock_t lock;
 	struct gss_api_mech *mech;
 	enum rpc_gss_svc service;
 	struct rpc_clnt *client;
@@ -108,7 +109,7 @@ gss_put_ctx(struct gss_cl_ctx *ctx)
 /* gss_cred_set_ctx:
  * called by gss_upcall_callback and gss_create_upcall in order
  * to set the gss context. The actual exchange of an old context
- * and a new one is protected by the inode->i_lock.
+ * and a new one is protected by the gss_auth->lock.
  */
 static void
 gss_cred_set_ctx(struct rpc_cred *cred, struct gss_cl_ctx *ctx)
@@ -283,14 +284,14 @@ gss_add_msg(struct gss_auth *gss_auth, struct gss_upcall_msg *gss_msg)
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct gss_upcall_msg *old;
 
-	spin_lock(&inode->i_lock);
+	spin_lock(&gss_auth->lock);
 	old = __gss_find_upcall(rpci, gss_msg->uid);
 	if (old == NULL) {
 		atomic_inc(&gss_msg->count);
 		list_add(&gss_msg->list, &rpci->in_downcall);
 	} else
 		gss_msg = old;
-	spin_unlock(&inode->i_lock);
+	spin_unlock(&gss_auth->lock);
 	return gss_msg;
 }
 
@@ -307,14 +308,13 @@ static void
 gss_unhash_msg(struct gss_upcall_msg *gss_msg)
 {
 	struct gss_auth *gss_auth = gss_msg->auth;
-	struct inode *inode = gss_auth->dentry->d_inode;
 
 	if (list_empty(&gss_msg->list))
 		return;
-	spin_lock(&inode->i_lock);
+	spin_lock(&gss_auth->lock);
 	if (!list_empty(&gss_msg->list))
 		__gss_unhash_msg(gss_msg);
-	spin_unlock(&inode->i_lock);
+	spin_unlock(&gss_auth->lock);
 }
 
 static void
@@ -323,16 +323,16 @@ gss_upcall_callback(struct rpc_task *task)
 	struct gss_cred *gss_cred = container_of(task->tk_msg.rpc_cred,
 			struct gss_cred, gc_base);
 	struct gss_upcall_msg *gss_msg = gss_cred->gc_upcall;
-	struct inode *inode = gss_msg->auth->dentry->d_inode;
+	struct gss_auth *gss_auth = gss_msg->auth;
 
-	spin_lock(&inode->i_lock);
+	spin_lock(&gss_auth->lock);
 	if (gss_msg->ctx)
 		gss_cred_set_ctx(task->tk_msg.rpc_cred, gss_msg->ctx);
 	else
 		task->tk_status = gss_msg->msg.errno;
 	gss_cred->gc_upcall = NULL;
 	rpc_wake_up_status(&gss_msg->rpc_waitqueue, gss_msg->msg.errno);
-	spin_unlock(&inode->i_lock);
+	spin_unlock(&gss_auth->lock);
 	gss_release_msg(gss_msg);
 }
 
@@ -391,7 +391,6 @@ gss_refresh_upcall(struct rpc_task *task)
 	struct gss_cred *gss_cred = container_of(cred,
 			struct gss_cred, gc_base);
 	struct gss_upcall_msg *gss_msg;
-	struct inode *inode = gss_auth->dentry->d_inode;
 	int err = 0;
 
 	dprintk("RPC: %5u gss_refresh_upcall for uid %u\n", task->tk_pid,
@@ -401,7 +400,7 @@ gss_refresh_upcall(struct rpc_task *task)
 		err = PTR_ERR(gss_msg);
 		goto out;
 	}
-	spin_lock(&inode->i_lock);
+	spin_lock(&gss_auth->lock);
 	if (gss_cred->gc_upcall != NULL)
 		rpc_sleep_on(&gss_cred->gc_upcall->rpc_waitqueue, task, NULL);
 	else if (gss_msg->ctx != NULL) {
@@ -416,7 +415,7 @@ gss_refresh_upcall(struct rpc_task *task)
 		rpc_sleep_on(&gss_msg->rpc_waitqueue, task, gss_upcall_callback);
 	} else
 		err = gss_msg->msg.errno;
-	spin_unlock(&inode->i_lock);
+	spin_unlock(&gss_auth->lock);
 	gss_release_msg(gss_msg);
 out:
 	dprintk("RPC: %5u gss_refresh_upcall for uid %u result %d\n",
@@ -427,7 +426,6 @@ out:
 static inline int
 gss_create_upcall(struct gss_auth *gss_auth, struct gss_cred *gss_cred)
 {
-	struct inode *inode = gss_auth->dentry->d_inode;
 	struct rpc_cred *cred = &gss_cred->gc_base;
 	struct gss_upcall_msg *gss_msg;
 	DEFINE_WAIT(wait);
@@ -441,11 +439,11 @@ gss_create_upcall(struct gss_auth *gss_auth, struct gss_cred *gss_cred)
 	}
 	for (;;) {
 		prepare_to_wait(&gss_msg->waitqueue, &wait, TASK_INTERRUPTIBLE);
-		spin_lock(&inode->i_lock);
+		spin_lock(&gss_auth->lock);
 		if (gss_msg->ctx != NULL || gss_msg->msg.errno < 0) {
 			break;
 		}
-		spin_unlock(&inode->i_lock);
+		spin_unlock(&gss_auth->lock);
 		if (signalled()) {
 			err = -ERESTARTSYS;
 			goto out_intr;
@@ -456,7 +454,7 @@ gss_create_upcall(struct gss_auth *gss_auth, struct gss_cred *gss_cred)
 		gss_cred_set_ctx(cred, gss_msg->ctx);
 	else
 		err = gss_msg->msg.errno;
-	spin_unlock(&inode->i_lock);
+	spin_unlock(&gss_auth->lock);
 out_intr:
 	finish_wait(&gss_msg->waitqueue, &wait);
 	gss_release_msg(gss_msg);
@@ -493,9 +491,11 @@ gss_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
 {
 	const void *p, *end;
 	void *buf;
-	struct rpc_clnt *clnt;
 	struct gss_upcall_msg *gss_msg;
 	struct inode *inode = filp->f_path.dentry->d_inode;
+	struct rpc_clnt *clnt = RPC_I(inode)->private;
+	struct gss_auth *gss_auth = container_of(clnt->cl_auth,
+						 struct gss_auth, rpc_auth);
 	struct gss_cl_ctx *ctx;
 	uid_t uid;
 	ssize_t err = -EFBIG;
@@ -526,14 +526,14 @@ gss_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
 
 	err = -ENOENT;
 	/* Find a matching upcall */
-	spin_lock(&inode->i_lock);
+	spin_lock(&gss_auth->lock);
 	gss_msg = __gss_find_upcall(RPC_I(inode), uid);
 	if (gss_msg == NULL) {
-		spin_unlock(&inode->i_lock);
+		spin_unlock(&gss_auth->lock);
 		goto err_put_ctx;
 	}
 	list_del_init(&gss_msg->list);
-	spin_unlock(&inode->i_lock);
+	spin_unlock(&gss_auth->lock);
 
 	p = gss_fill_context(p, end, ctx, gss_msg->auth->mech);
 	if (IS_ERR(p)) {
@@ -545,9 +545,9 @@ gss_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
 	err = mlen;
 
 err_release_msg:
-	spin_lock(&inode->i_lock);
+	spin_lock(&gss_auth->lock);
 	__gss_unhash_msg(gss_msg);
-	spin_unlock(&inode->i_lock);
+	spin_unlock(&gss_auth->lock);
 	gss_release_msg(gss_msg);
 err_put_ctx:
 	gss_put_ctx(ctx);
@@ -562,9 +562,12 @@ static void
 gss_pipe_release(struct inode *inode)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
+	struct rpc_clnt *clnt = rpci->private;
+	struct gss_auth *gss_auth = container_of(clnt->cl_auth,
+						 struct gss_auth, rpc_auth);
 	struct gss_upcall_msg *gss_msg;
 
-	spin_lock(&inode->i_lock);
+	spin_lock(&gss_auth->lock);
 	while (!list_empty(&rpci->in_downcall)) {
 
 		gss_msg = list_entry(rpci->in_downcall.next,
@@ -572,11 +575,11 @@ gss_pipe_release(struct inode *inode)
 		gss_msg->msg.errno = -EPIPE;
 		atomic_inc(&gss_msg->count);
 		__gss_unhash_msg(gss_msg);
-		spin_unlock(&inode->i_lock);
+		spin_unlock(&gss_auth->lock);
 		gss_release_msg(gss_msg);
-		spin_lock(&inode->i_lock);
+		spin_lock(&gss_auth->lock);
 	}
-	spin_unlock(&inode->i_lock);
+	spin_unlock(&gss_auth->lock);
 }
 
 static void
@@ -631,6 +634,7 @@ gss_create(struct rpc_clnt *clnt, rpc_authflavor_t flavor)
 	if (gss_auth->service == 0)
 		goto err_put_mech;
 	auth = &gss_auth->rpc_auth;
+	spin_lock_init(&gss_auth->lock);
 	auth->au_cslack = GSS_CRED_SLACK >> 2;
 	auth->au_rslack = GSS_VERF_SLACK >> 2;
 	auth->au_ops = &authgss_ops;
-- 
1.5.5.rc1

[PATCH 3/5] rpc: move in_downcall list to gss code

[J. Bruce Fields]

This list is only used by the auth_gss code anyway.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 include/linux/sunrpc/rpc_pipe_fs.h |    1 -
 net/sunrpc/auth_gss/auth_gss.c     |   33 +++++++++++++++++++--------------
 net/sunrpc/rpc_pipe.c              |    1 -
 3 files changed, 19 insertions(+), 16 deletions(-)

diff --git a/include/linux/sunrpc/rpc_pipe_fs.h b/include/linux/sunrpc/rpc_pipe_fs.h
index 51b977a..867bea8 100644
--- a/include/linux/sunrpc/rpc_pipe_fs.h
+++ b/include/linux/sunrpc/rpc_pipe_fs.h
@@ -23,7 +23,6 @@ struct rpc_inode {
 	void *private;
 	struct list_head pipe;
 	struct list_head in_upcall;
-	struct list_head in_downcall;
 	int pipelen;
 	int nreaders;
 	int nwriters;
diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 28ca1c9..5e82a69 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -83,6 +83,7 @@ struct gss_auth {
 	struct kref kref;
 	struct rpc_auth rpc_auth;
 	spinlock_t lock;
+	struct list_head in_downcall;
 	struct gss_api_mech *mech;
 	enum rpc_gss_svc service;
 	struct rpc_clnt *client;
@@ -259,10 +260,11 @@ gss_release_msg(struct gss_upcall_msg *gss_msg)
 }
 
 static struct gss_upcall_msg *
-__gss_find_upcall(struct rpc_inode *rpci, uid_t uid)
+__gss_find_upcall(struct gss_auth *gss_auth, uid_t uid)
 {
 	struct gss_upcall_msg *pos;
-	list_for_each_entry(pos, &rpci->in_downcall, list) {
+
+	list_for_each_entry(pos, &gss_auth->in_downcall, list) {
 		if (pos->uid != uid)
 			continue;
 		atomic_inc(&pos->count);
@@ -280,15 +282,13 @@ __gss_find_upcall(struct rpc_inode *rpci, uid_t uid)
 static inline struct gss_upcall_msg *
 gss_add_msg(struct gss_auth *gss_auth, struct gss_upcall_msg *gss_msg)
 {
-	struct inode *inode = gss_auth->dentry->d_inode;
-	struct rpc_inode *rpci = RPC_I(inode);
 	struct gss_upcall_msg *old;
 
 	spin_lock(&gss_auth->lock);
-	old = __gss_find_upcall(rpci, gss_msg->uid);
+	old = __gss_find_upcall(gss_auth, gss_msg->uid);
 	if (old == NULL) {
 		atomic_inc(&gss_msg->count);
-		list_add(&gss_msg->list, &rpci->in_downcall);
+		list_add(&gss_msg->list, &gss_auth->in_downcall);
 	} else
 		gss_msg = old;
 	spin_unlock(&gss_auth->lock);
@@ -484,6 +484,14 @@ gss_pipe_upcall(struct file *filp, struct rpc_pipe_msg *msg,
 	return mlen;
 }
 
+static struct gss_auth *gss_auth_from_dentry(struct dentry *dentry)
+{
+	struct inode *inode = dentry->d_inode;
+	struct rpc_clnt *clnt = RPC_I(inode)->private;
+
+	return container_of(clnt->cl_auth, struct gss_auth, rpc_auth);
+}
+
 #define MSG_BUF_MAXSIZE 1024
 
 static ssize_t
@@ -492,10 +500,7 @@ gss_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
 	const void *p, *end;
 	void *buf;
 	struct gss_upcall_msg *gss_msg;
-	struct inode *inode = filp->f_path.dentry->d_inode;
-	struct rpc_clnt *clnt = RPC_I(inode)->private;
-	struct gss_auth *gss_auth = container_of(clnt->cl_auth,
-						 struct gss_auth, rpc_auth);
+	struct gss_auth *gss_auth = gss_auth_from_dentry(filp->f_path.dentry);
 	struct gss_cl_ctx *ctx;
 	uid_t uid;
 	ssize_t err = -EFBIG;
@@ -507,7 +512,6 @@ gss_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
 	if (!buf)
 		goto out;
 
-	clnt = RPC_I(inode)->private;
 	err = -EFAULT;
 	if (copy_from_user(buf, src, mlen))
 		goto err;
@@ -527,7 +531,7 @@ gss_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
 	err = -ENOENT;
 	/* Find a matching upcall */
 	spin_lock(&gss_auth->lock);
-	gss_msg = __gss_find_upcall(RPC_I(inode), uid);
+	gss_msg = __gss_find_upcall(gss_auth, uid);
 	if (gss_msg == NULL) {
 		spin_unlock(&gss_auth->lock);
 		goto err_put_ctx;
@@ -568,9 +572,9 @@ gss_pipe_release(struct inode *inode)
 	struct gss_upcall_msg *gss_msg;
 
 	spin_lock(&gss_auth->lock);
-	while (!list_empty(&rpci->in_downcall)) {
+	while (!list_empty(&gss_auth->in_downcall)) {
 
-		gss_msg = list_entry(rpci->in_downcall.next,
+		gss_msg = list_entry(gss_auth->in_downcall.next,
 				struct gss_upcall_msg, list);
 		gss_msg->msg.errno = -EPIPE;
 		atomic_inc(&gss_msg->count);
@@ -635,6 +639,7 @@ gss_create(struct rpc_clnt *clnt, rpc_authflavor_t flavor)
 		goto err_put_mech;
 	auth = &gss_auth->rpc_auth;
 	spin_lock_init(&gss_auth->lock);
+	INIT_LIST_HEAD(&gss_auth->in_downcall);
 	auth->au_cslack = GSS_CRED_SLACK >> 2;
 	auth->au_rslack = GSS_VERF_SLACK >> 2;
 	auth->au_ops = &authgss_ops;
diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index 5a9b0e7..41d968f 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -906,7 +906,6 @@ init_once(struct kmem_cache * cachep, void *foo)
 	rpci->nreaders = 0;
 	rpci->nwriters = 0;
 	INIT_LIST_HEAD(&rpci->in_upcall);
-	INIT_LIST_HEAD(&rpci->in_downcall);
 	INIT_LIST_HEAD(&rpci->pipe);
 	rpci->pipelen = 0;
 	init_waitqueue_head(&rpci->waitq);
-- 
1.5.5.rc1

[PATCH 4/5] rpc: add an rpc_pipe_open method

[J. Bruce Fields]

We want to transition to a new gssd upcall which is text-based and more
easily extensible.

To simplify upgrades, as well as testing and debugging, it will help if
we can upgrade gssd (to a version which understands the new upcall)
without having to choose at boot (or module-load) time whether we want
the new or the old upcall.

That means that when gssd opens an rpc upcall pipe, we'd like to just
start using whichever version it chose immediately--ideally without
having to wait the 30 seconds that it would normally take for upcalls
already queued on the wrong upcall pipe to time out.

So we add an rpc_pipe open method, which we will use to let the gssd
code know that a new gssd may have started up.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 include/linux/sunrpc/rpc_pipe_fs.h |    1 +
 net/sunrpc/rpc_pipe.c              |    9 ++++++++-
 2 files changed, 9 insertions(+), 1 deletions(-)

diff --git a/include/linux/sunrpc/rpc_pipe_fs.h b/include/linux/sunrpc/rpc_pipe_fs.h
index 867bea8..439bef3 100644
--- a/include/linux/sunrpc/rpc_pipe_fs.h
+++ b/include/linux/sunrpc/rpc_pipe_fs.h
@@ -15,6 +15,7 @@ struct rpc_pipe_ops {
 	ssize_t (*upcall)(struct file *, struct rpc_pipe_msg *, char __user *, size_t);
 	ssize_t (*downcall)(struct file *, const char __user *, size_t);
 	void (*release_pipe)(struct inode *);
+	int (*open_pipe)(struct inode *);
 	void (*destroy_msg)(struct rpc_pipe_msg *);
 };
 
diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index 41d968f..f7781fc 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -169,9 +169,15 @@ static int
 rpc_pipe_open(struct inode *inode, struct file *filp)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
-	int res = -ENXIO;
+	int res;
 
 	mutex_lock(&inode->i_mutex);
+	if (rpci->ops->open_pipe) {
+		res = rpci->ops->open_pipe(inode);
+		if (res)
+			goto out;
+	}
+	res = -ENXIO;
 	if (rpci->ops != NULL) {
 		if (filp->f_mode & FMODE_READ)
 			rpci->nreaders ++;
@@ -179,6 +185,7 @@ rpc_pipe_open(struct inode *inode, struct file *filp)
 			rpci->nwriters ++;
 		res = 0;
 	}
+out:
 	mutex_unlock(&inode->i_mutex);
 	return res;
 }
-- 
1.5.5.rc1

[PATCH 5/5] rpc: add new gssd upcall pipe

[J. Bruce Fields]

This adds the new text-based gss upcall.  We only allow an open from
either the new pipe or the old one, and use the pipe_open method to
enforce this.

If no pipes are open, we provisionally queue messages for the new pipe;
a subsequent open of the old pipe will cause all those messages to be
purged.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 include/linux/sunrpc/rpc_pipe_fs.h |    1 +
 net/sunrpc/auth_gss/auth_gss.c     |  144 ++++++++++++++++++++++++++++++++---
 net/sunrpc/rpc_pipe.c              |   27 +++++--
 3 files changed, 153 insertions(+), 19 deletions(-)

diff --git a/include/linux/sunrpc/rpc_pipe_fs.h b/include/linux/sunrpc/rpc_pipe_fs.h
index 439bef3..4c0f4a0 100644
--- a/include/linux/sunrpc/rpc_pipe_fs.h
+++ b/include/linux/sunrpc/rpc_pipe_fs.h
@@ -42,6 +42,7 @@ RPC_I(struct inode *inode)
 }
 
 extern int rpc_queue_upcall(struct inode *, struct rpc_pipe_msg *);
+void rpc_pipe_cancel_all(struct inode *inode, int errno);
 
 extern struct dentry *rpc_mkdir(char *, struct rpc_clnt *);
 extern int rpc_rmdir(struct dentry *);
diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 5e82a69..a5a56e8 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -87,7 +87,19 @@ struct gss_auth {
 	struct gss_api_mech *mech;
 	enum rpc_gss_svc service;
 	struct rpc_clnt *client;
-	struct dentry *dentry;
+	/*
+	 * There are two upcall pipes; dentry[1], named "gssd", is used
+	 * for the new text-based upcall; dentry[0] is named after the
+	 * mechanism ("krb5") and exists for backwards-compatibility
+	 * with older gssd's.
+	 */
+	struct dentry *dentry[2];
+	/*
+	 * The upcall_version tracks which of the above upcalls we're
+	 * using, 0 (old) or 1 (new).  While userland has neither pipe
+	 * open, upcall_version is set to -1:
+	 */
+	int upcall_version;
 };
 
 static void gss_free_ctx(struct gss_cl_ctx *);
@@ -235,6 +247,7 @@ err:
 	return p;
 }
 
+#define UPCALL_BUF_LEN 128
 
 struct gss_upcall_msg {
 	atomic_t count;
@@ -245,6 +258,7 @@ struct gss_upcall_msg {
 	struct rpc_wait_queue rpc_waitqueue;
 	wait_queue_head_t waitqueue;
 	struct gss_cl_ctx *ctx;
+	char databuf[UPCALL_BUF_LEN];
 };
 
 static void
@@ -347,14 +361,52 @@ gss_alloc_msg(struct gss_auth *gss_auth, uid_t uid)
 		rpc_init_wait_queue(&gss_msg->rpc_waitqueue, "RPCSEC_GSS upcall waitq");
 		init_waitqueue_head(&gss_msg->waitqueue);
 		atomic_set(&gss_msg->count, 1);
-		gss_msg->msg.data = &gss_msg->uid;
-		gss_msg->msg.len = sizeof(gss_msg->uid);
 		gss_msg->uid = uid;
 		gss_msg->auth = gss_auth;
 	}
 	return gss_msg;
 }
 
+static void gss_encode_v0_msg(struct gss_upcall_msg *gss_msg)
+{
+	gss_msg->msg.data = &gss_msg->uid;
+	gss_msg->msg.len = sizeof(gss_msg->uid);
+}
+
+static void gss_encode_v1_msg(struct gss_upcall_msg *gss_msg)
+{
+	gss_msg->msg.len = sprintf(gss_msg->databuf, "mech=%s uid=%d\n",
+				   gss_msg->auth->mech->gm_name,
+				   gss_msg->uid);
+	gss_msg->msg.data = gss_msg->databuf;
+	BUG_ON(gss_msg->msg.len > UPCALL_BUF_LEN);
+}
+
+static void gss_encode_msg(struct gss_upcall_msg *gss_msg, int vers)
+{
+	if (vers == 0)
+		gss_encode_v0_msg(gss_msg);
+	else
+		gss_encode_v1_msg(gss_msg);
+}
+
+static int gss_queue_upcall(struct gss_auth *gss_auth, struct gss_upcall_msg *gss_msg)
+{
+	int res;
+	int vers;
+
+	spin_lock(&gss_auth->lock);
+	vers = gss_auth->upcall_version;
+	if (vers == -1) {
+		/* Neither pipe opened yet; default to the new one: */
+		vers = 1;
+	}
+	gss_encode_msg(gss_msg, vers);
+	res = rpc_queue_upcall(gss_auth->dentry[vers]->d_inode, &gss_msg->msg);
+	spin_unlock(&gss_auth->lock);
+	return res;
+}
+
 static struct gss_upcall_msg *
 gss_setup_upcall(struct rpc_clnt *clnt, struct gss_auth *gss_auth, struct rpc_cred *cred)
 {
@@ -372,7 +424,7 @@ gss_setup_upcall(struct rpc_clnt *clnt, struct gss_auth *gss_auth, struct rpc_cr
 		return ERR_PTR(-ENOMEM);
 	gss_msg = gss_add_msg(gss_auth, gss_new);
 	if (gss_msg == gss_new) {
-		int res = rpc_queue_upcall(gss_auth->dentry->d_inode, &gss_new->msg);
+		int res = gss_queue_upcall(gss_auth, gss_msg);
 		if (res) {
 			gss_unhash_msg(gss_new);
 			gss_msg = ERR_PTR(res);
@@ -562,6 +614,55 @@ out:
 	return err;
 }
 
+static int gss_pipe_version(struct gss_auth *gss_auth, struct inode *inode)
+{
+	if (gss_auth->dentry[0]->d_inode == inode)
+		return 0;
+	if (gss_auth->dentry[1]->d_inode == inode)
+		return 1;
+	BUG();
+}
+
+static int
+gss_pipe_open(struct inode *inode)
+{
+	struct rpc_inode *rpci = RPC_I(inode);
+	struct rpc_clnt *clnt = rpci->private;
+	struct gss_auth *gss_auth = container_of(clnt->cl_auth,
+						 struct gss_auth, rpc_auth);
+	int vers = gss_pipe_version(gss_auth, inode);
+	int ret;
+
+	spin_lock(&gss_auth->lock);
+	if (gss_auth->upcall_version >= 0) {
+		/*
+		 * If pipe is already opened, permit only opens of the same
+		 * version:
+		 */
+		if (vers == gss_auth->upcall_version)
+			ret = 0;
+		else
+			ret = -EBUSY;
+		spin_unlock(&gss_auth->lock);
+		return ret;
+	}
+
+	/*
+	 * Neither pipe was opened previously, so this open is the first
+	 * hint we've gotten as to which upcall gssd intends to use:
+	 */
+	gss_auth->upcall_version = vers;
+	spin_unlock(&gss_auth->lock);
+	if (vers == 0) {
+		/*
+		 * We default to trying the new upcall, so may have
+		 * already attempted some upcalls there; cancel them:
+		 */
+		rpc_pipe_cancel_all(gss_auth->dentry[1]->d_inode, -EAGAIN);
+	}
+	return 0;
+}
+
 static void
 gss_pipe_release(struct inode *inode)
 {
@@ -572,6 +673,8 @@ gss_pipe_release(struct inode *inode)
 	struct gss_upcall_msg *gss_msg;
 
 	spin_lock(&gss_auth->lock);
+	if (rpci->nreaders == 0)
+		gss_auth->upcall_version = -1;
 	while (!list_empty(&gss_auth->in_downcall)) {
 
 		gss_msg = list_entry(gss_auth->in_downcall.next,
@@ -646,21 +749,34 @@ gss_create(struct rpc_clnt *clnt, rpc_authflavor_t flavor)
 	auth->au_flavor = flavor;
 	atomic_set(&auth->au_count, 1);
 	kref_init(&gss_auth->kref);
-
-	gss_auth->dentry = rpc_mkpipe(clnt->cl_dentry, gss_auth->mech->gm_name,
-			clnt, &gss_upcall_ops, RPC_PIPE_WAIT_FOR_OPEN);
-	if (IS_ERR(gss_auth->dentry)) {
-		err = PTR_ERR(gss_auth->dentry);
+	gss_auth->upcall_version = -1;
+
+	gss_auth->dentry[0] = rpc_mkpipe(clnt->cl_dentry,
+					 gss_auth->mech->gm_name,
+					 clnt, &gss_upcall_ops,
+					 RPC_PIPE_WAIT_FOR_OPEN);
+	if (IS_ERR(gss_auth->dentry[0])) {
+		err = PTR_ERR(gss_auth->dentry[0]);
 		goto err_put_mech;
 	}
+	gss_auth->dentry[1] = rpc_mkpipe(clnt->cl_dentry,
+					 "gssd",
+					 clnt, &gss_upcall_ops,
+					 RPC_PIPE_WAIT_FOR_OPEN);
+	if (IS_ERR(gss_auth->dentry[1])) {
+		err = PTR_ERR(gss_auth->dentry[1]);
+		goto err_unlink_pipe_0;
+	}
 
 	err = rpcauth_init_credcache(auth);
 	if (err)
-		goto err_unlink_pipe;
+		goto err_unlink_pipe_1;
 
 	return auth;
-err_unlink_pipe:
-	rpc_unlink(gss_auth->dentry);
+err_unlink_pipe_1:
+	rpc_unlink(gss_auth->dentry[1]);
+err_unlink_pipe_0:
+	rpc_unlink(gss_auth->dentry[0]);
 err_put_mech:
 	gss_mech_put(gss_auth->mech);
 err_free:
@@ -673,7 +789,8 @@ out_dec:
 static void
 gss_free(struct gss_auth *gss_auth)
 {
-	rpc_unlink(gss_auth->dentry);
+	rpc_unlink(gss_auth->dentry[1]);
+	rpc_unlink(gss_auth->dentry[0]);
 	gss_mech_put(gss_auth->mech);
 
 	kfree(gss_auth);
@@ -1362,6 +1479,7 @@ static struct rpc_pipe_ops gss_upcall_ops = {
 	.downcall	= gss_pipe_downcall,
 	.destroy_msg	= gss_pipe_destroy_msg,
 	.release_pipe	= gss_pipe_release,
+	.open_pipe	= gss_pipe_open,
 };
 
 /*
diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index f7781fc..bbefc3a 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -53,13 +53,18 @@ static void rpc_purge_list(struct rpc_inode *rpci, struct list_head *head,
 	wake_up(&rpci->waitq);
 }
 
-static void
-rpc_timeout_upcall_queue(struct work_struct *work)
+/**
+ * rpc_pipe_cancel_all
+ * @inode: inode of upcall pipe to cancel all upcalls on
+ * @errno: errno (as a negative integer) to return on all pending upcalls
+ *
+ * Call with an inode created by rpc_mkpipe() to remove all queued
+ * upcalls.  Allowed only on pipes not yet opened for read.
+ */
+void rpc_pipe_cancel_all(struct inode *inode, int errno)
 {
 	LIST_HEAD(free_list);
-	struct rpc_inode *rpci =
-		container_of(work, struct rpc_inode, queue_timeout.work);
-	struct inode *inode = &rpci->vfs_inode;
+	struct rpc_inode *rpci = RPC_I(inode);
 	void (*destroy_msg)(struct rpc_pipe_msg *);
 
 	spin_lock(&inode->i_lock);
@@ -73,7 +78,17 @@ rpc_timeout_upcall_queue(struct work_struct *work)
 		rpci->pipelen = 0;
 	}
 	spin_unlock(&inode->i_lock);
-	rpc_purge_list(rpci, &free_list, destroy_msg, -ETIMEDOUT);
+	rpc_purge_list(rpci, &free_list, destroy_msg, errno);
+}
+
+static void
+rpc_timeout_upcall_queue(struct work_struct *work)
+{
+	struct rpc_inode *rpci =
+		container_of(work, struct rpc_inode, queue_timeout.work);
+	struct inode *inode = &rpci->vfs_inode;
+
+	rpc_pipe_cancel_all(inode, -ETIMEDOUT);
 }
 
 /**
-- 
1.5.5.rc1

Re: [PATCH 5/5] rpc: add new gssd upcall pipe

[Trond Myklebust]

On Fri, 2008-06-13 at 18:50 -0400, J. Bruce Fields wrote:
> This adds the new text-based gss upcall.  We only allow an open from
> either the new pipe or the old one, and use the pipe_open method to
> enforce this.
> 
> If no pipes are open, we provisionally queue messages for the new pipe;
> a subsequent open of the old pipe will cause all those messages to be
> purged.
> 
> Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
> ---

I'm not happy about allowing the gss_auth layer to control timeouts and
cancel upcalls either: that caused us way too much grief and bugs in
earlier revisions of rpc_pipefs.

Why not rather disallow queueing (i.e. put the task to sleep, and maybe
print out a reminder on the console every minute or so) until you are
notified by the rpc_pipefs layer that someone has opened the file for
reading?

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 5/5] rpc: add new gssd upcall pipe

[J. Bruce Fields]

On Sat, Jun 14, 2008 at 12:07:18PM -0400, Trond Myklebust wrote:
> On Fri, 2008-06-13 at 18:50 -0400, J. Bruce Fields wrote:
> > This adds the new text-based gss upcall.  We only allow an open from
> > either the new pipe or the old one, and use the pipe_open method to
> > enforce this.
> > 
> > If no pipes are open, we provisionally queue messages for the new pipe;
> > a subsequent open of the old pipe will cause all those messages to be
> > purged.
> > 
> > Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
> > ---
> 
> I'm not happy about allowing the gss_auth layer to control timeouts and
> cancel upcalls either: that caused us way too much grief and bugs in
> earlier revisions of rpc_pipefs.

Understood.  In this case the only cancel occurs before any open has
come--it seems pretty simple.  But perhaps I'm missing something--I'll
think about it some more.

> Why not rather disallow queueing (i.e. put the task to sleep, and maybe
> print out a reminder on the console every minute or so) until you are
> notified by the rpc_pipefs layer that someone has opened the file for
> reading?

I have another version of these patches that did that (in fact, I just
turned off the RPC_PIPE_WAIT_FOR_OPEN stuff and relied entirely on
timeouts in gss code).  I'll take another look at that.

--b.

Re: [PATCH 5/5] rpc: add new gssd upcall pipe

[J. Bruce Fields]

On Sat, Jun 14, 2008 at 12:07:18PM -0400, Trond Myklebust wrote:
> On Fri, 2008-06-13 at 18:50 -0400, J. Bruce Fields wrote:
> > This adds the new text-based gss upcall.  We only allow an open from
> > either the new pipe or the old one, and use the pipe_open method to
> > enforce this.
> > 
> > If no pipes are open, we provisionally queue messages for the new pipe;
> > a subsequent open of the old pipe will cause all those messages to be
> > purged.
> > 
> > Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
> > ---
> 
> I'm not happy about allowing the gss_auth layer to control timeouts and
> cancel upcalls either: that caused us way too much grief and bugs in
> earlier revisions of rpc_pipefs.
> 
> Why not rather disallow queueing (i.e. put the task to sleep, and maybe
> print out a reminder on the console every minute or so) until you are
> notified by the rpc_pipefs layer that someone has opened the file for
> reading?

OK, done.

--b.

Re: [PATCH 3/5] rpc: move in_downcall list to gss code

[Trond Myklebust]

On Fri, 2008-06-13 at 18:50 -0400, J. Bruce Fields wrote:
> This list is only used by the auth_gss code anyway.
> 
> Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>

NACK. This too reintroduces a forced one-to-one relationship between the
inode and the gss_auth.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 2/5] rpc: Use separate spinlock for cred locking in auth_gss.c

[Trond Myklebust]

On Fri, 2008-06-13 at 18:50 -0400, J. Bruce Fields wrote:
> The i_lock uses in auth_gss.c are to protect the gc_upcall and gc_ctx
> fields of the gss cred; so we can use a separate per-auth spinlock for
> those instead of the i_lock.  This simplifies the locking in the case
> where we have two different pipes (one the current pipe, one the pipe
> for a new more flexible upcall), either of which might be used for gss
> upcalls.

NACK. I deliberately ripped out the old struct gss_auth spinlock in
order to allow multiple gss_auth per inode (I believe Kevin was asking
for this).

Trond

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 2/5] rpc: Use separate spinlock for cred locking in auth_gss.c

[J. Bruce Fields]

On Sat, Jun 14, 2008 at 11:58:55AM -0400, Trond Myklebust wrote:
> On Fri, 2008-06-13 at 18:50 -0400, J. Bruce Fields wrote:
> > The i_lock uses in auth_gss.c are to protect the gc_upcall and gc_ctx
> > fields of the gss cred; so we can use a separate per-auth spinlock for
> > those instead of the i_lock.  This simplifies the locking in the case
> > where we have two different pipes (one the current pipe, one the pipe
> > for a new more flexible upcall), either of which might be used for gss
> > upcalls.
> 
> NACK. I deliberately ripped out the old struct gss_auth spinlock in
> order to allow multiple gss_auth per inode (I believe Kevin was asking
> for this).

OK, makes sense.  So what will be the scope of a cred lookup--an rpc
client?

I wish I had a better picture of what the data structures here will
eventually look like.

--b.

Re: [PATCH 2/5] rpc: Use separate spinlock for cred locking in auth_gss.c

[Trond Myklebust]

On Sat, 2008-06-14 at 13:45 -0400, J. Bruce Fields wrote:
> > NACK. I deliberately ripped out the old struct gss_auth spinlock in
> > order to allow multiple gss_auth per inode (I believe Kevin was asking
> > for this).
> 
> OK, makes sense.  So what will be the scope of a cred lookup--an rpc
> client?

That should normally be the case, but why do you need to change the
locking here in the first place? Is there ever going to be a case where
the same rpc_cred has upcalls on several different pipes? I can't see
how that could be justified.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 2/5] rpc: Use separate spinlock for cred locking in auth_gss.c

[J. Bruce Fields]

On Sat, Jun 14, 2008 at 02:16:27PM -0400, Trond Myklebust wrote:
> On Sat, 2008-06-14 at 13:45 -0400, J. Bruce Fields wrote:
> > > NACK. I deliberately ripped out the old struct gss_auth spinlock in
> > > order to allow multiple gss_auth per inode (I believe Kevin was asking
> > > for this).
> > 
> > OK, makes sense.  So what will be the scope of a cred lookup--an rpc
> > client?
> 
> That should normally be the case, but why do you need to change the
> locking here in the first place? Is there ever going to be a case where
> the same rpc_cred has upcalls on several different pipes? I can't see
> how that could be justified.

If you allow changing the upcall version over the life of the kernel,
then it's difficult to avoid.  You can insist that noone have both the
new and old version opened simultaneously, for example, but it's harder
to know when there are no longer messages sitting around that have been
unhashed but are still lying around waiting for a process to wake up and
examine their results.

--b.

Re: [PATCH 2/5] rpc: Use separate spinlock for cred locking in auth_gss.c

[Trond Myklebust]

On Tue, 2008-06-17 at 16:51 -0400, J. Bruce Fields wrote:
> On Sat, Jun 14, 2008 at 02:16:27PM -0400, Trond Myklebust wrote:
> > On Sat, 2008-06-14 at 13:45 -0400, J. Bruce Fields wrote:
> > > > NACK. I deliberately ripped out the old struct gss_auth spinlock in
> > > > order to allow multiple gss_auth per inode (I believe Kevin was asking
> > > > for this).
> > > 
> > > OK, makes sense.  So what will be the scope of a cred lookup--an rpc
> > > client?
> > 
> > That should normally be the case, but why do you need to change the
> > locking here in the first place? Is there ever going to be a case where
> > the same rpc_cred has upcalls on several different pipes? I can't see
> > how that could be justified.
> 
> If you allow changing the upcall version over the life of the kernel,
> then it's difficult to avoid.  You can insist that noone have both the
> new and old version opened simultaneously, for example, but it's harder
> to know when there are no longer messages sitting around that have been
> unhashed but are still lying around waiting for a process to wake up and
> examine their results.

AFAIK, there are two cases when the dying rpc.gssd closes the pipe:

     1. gss_cred->gc_upcall is set. In this case, the gss_cred has a
        full reference to the gss_msg, so it has no trouble locating the
        message and figuring out if it needs to resend (rpc_purge_list()
        will ensure that the error field is set to EAGAIN) or if the
        call completed before gssd died. If you are worried about the
        fact that gss_upcall_callback uses gss_msg->auth to access the
        inode in order to do the locking, then just add a pointer to the
        inode to gss_msg: it is not as if a gss_msg can migrate from one
        upcall queue to another.
     2. gss_cred->gc_upcall is not set. In this case the call to
        rpc_purge_list() in rpc_pipe_release() will ensure that the
        message gets immediately released.

In other words, I can't see that keeping the current behaviour will
cause you to have more than one upcall at a time even if you do restart
rpc.gssd.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 2/5] rpc: Use separate spinlock for cred locking in auth_gss.c

[J. Bruce Fields]

On Tue, Jun 17, 2008 at 05:34:47PM -0400, Trond Myklebust wrote:
> On Tue, 2008-06-17 at 16:51 -0400, J. Bruce Fields wrote:
> > On Sat, Jun 14, 2008 at 02:16:27PM -0400, Trond Myklebust wrote:
> > > On Sat, 2008-06-14 at 13:45 -0400, J. Bruce Fields wrote:
> > > > > NACK. I deliberately ripped out the old struct gss_auth spinlock in
> > > > > order to allow multiple gss_auth per inode (I believe Kevin was asking
> > > > > for this).
> > > > 
> > > > OK, makes sense.  So what will be the scope of a cred lookup--an rpc
> > > > client?
> > > 
> > > That should normally be the case, but why do you need to change the
> > > locking here in the first place? Is there ever going to be a case where
> > > the same rpc_cred has upcalls on several different pipes? I can't see
> > > how that could be justified.
> > 
> > If you allow changing the upcall version over the life of the kernel,
> > then it's difficult to avoid.  You can insist that noone have both the
> > new and old version opened simultaneously, for example, but it's harder
> > to know when there are no longer messages sitting around that have been
> > unhashed but are still lying around waiting for a process to wake up and
> > examine their results.
> 
> AFAIK, there are two cases when the dying rpc.gssd closes the pipe:
> 
>      1. gss_cred->gc_upcall is set. In this case, the gss_cred has a
>         full reference to the gss_msg, so it has no trouble locating the
>         message and figuring out if it needs to resend (rpc_purge_list()
>         will ensure that the error field is set to EAGAIN) or if the
>         call completed before gssd died. If you are worried about the
>         fact that gss_upcall_callback uses gss_msg->auth to access the
>         inode in order to do the locking, then just add a pointer to the
>         inode to gss_msg: it is not as if a gss_msg can migrate from one
>         upcall queue to another.
>      2. gss_cred->gc_upcall is not set. In this case the call to
>         rpc_purge_list() in rpc_pipe_release() will ensure that the
>         message gets immediately released.
> 
> In other words, I can't see that keeping the current behaviour will
> cause you to have more than one upcall at a time even if you do restart
> rpc.gssd.

Yeah, I think that's more or less exactly the sort of case I was worried
about--so let me see if that does the job.  Thanks!--b.

Re: [PATCH 2/5] rpc: Use separate spinlock for cred locking in auth_gss.c

[J. Bruce Fields]

On Tue, Jun 17, 2008 at 05:34:47PM -0400, Trond Myklebust wrote:
> On Tue, 2008-06-17 at 16:51 -0400, J. Bruce Fields wrote:
> > On Sat, Jun 14, 2008 at 02:16:27PM -0400, Trond Myklebust wrote:
> > > On Sat, 2008-06-14 at 13:45 -0400, J. Bruce Fields wrote:
> > > > > NACK. I deliberately ripped out the old struct gss_auth spinlock in
> > > > > order to allow multiple gss_auth per inode (I believe Kevin was asking
> > > > > for this).
> > > > 
> > > > OK, makes sense.  So what will be the scope of a cred lookup--an rpc
> > > > client?
> > > 
> > > That should normally be the case, but why do you need to change the
> > > locking here in the first place? Is there ever going to be a case where
> > > the same rpc_cred has upcalls on several different pipes? I can't see
> > > how that could be justified.
> > 
> > If you allow changing the upcall version over the life of the kernel,
> > then it's difficult to avoid.  You can insist that noone have both the
> > new and old version opened simultaneously, for example, but it's harder
> > to know when there are no longer messages sitting around that have been
> > unhashed but are still lying around waiting for a process to wake up and
> > examine their results.
> 
> AFAIK, there are two cases when the dying rpc.gssd closes the pipe:
> 
>      1. gss_cred->gc_upcall is set. In this case, the gss_cred has a
>         full reference to the gss_msg, so it has no trouble locating the
>         message and figuring out if it needs to resend (rpc_purge_list()
>         will ensure that the error field is set to EAGAIN) or if the
>         call completed before gssd died. If you are worried about the
>         fact that gss_upcall_callback uses gss_msg->auth to access the
>         inode in order to do the locking, then just add a pointer to the
>         inode to gss_msg: it is not as if a gss_msg can migrate from one
>         upcall queue to another.
>      2. gss_cred->gc_upcall is not set. In this case the call to
>         rpc_purge_list() in rpc_pipe_release() will ensure that the
>         message gets immediately released.
> 
> In other words, I can't see that keeping the current behaviour will
> cause you to have more than one upcall at a time even if you do restart
> rpc.gssd.

OK, sorry for the delay.  I believe there is still a race here in the
presence of two pipes that take upcalls, something like:

	task 1				task 2
	------				------

	Create upcall message on
	old pipe using refresh_upcall

		---- gssd closes old pipe, destroys message ----
		---- new gssd opens new pipe ----

	close of old pipe wakes up
	this rpc task; upcall_callback
	called.
					gss_refresh__callback creates new
					upcall on new pipe for the same cred;
					sets gc_upcall.
	gc_upcall := NULL
					woken up later, upcall_callback
					dereferences NULL gc_upcall.

The problem being that access to gc_upcall isn't correctly serialized,
since it's protected here by two different i_lock's (task 1 is using the
old pipe's i_lock, task 2 the new pipe's i_lock).

Anyway, rather than fooling with the basic data structures I figured
it's easiest just to add a separate lock which just controls the change
of version and to forbid changing the version while there are still any
upcalls lying around.  I'll send patches.

--b.

text-based gss upcall

[J. Bruce Fields]

So, this time there aren't any changes to the underlying
gss_auth/rpc_pipe locking, and the gss code doesn't attempt to cancel
rpc calls on its own.  Instead we:

	- keep a single global variable which determines which version (new
	  or old) of the upcall we use;
	- keep a counter which tells us
	  	count of gss pipes open + count of gss upcalls outstanding;
	- permit switching versions only when that counter goes to zero;
	- use pipe_open and pipe_release callbacks to maintain that
	  counter; and
	- allow the gss code to wait on gssd to open a pipe before
	  deciding which version of upcall to make.

And the new locking change is one spinlock to protect the global pipe
version.

I've tested with an unmodified gssd, and checked that things still work
correctly when gssd is killed and restarted.  I haven't tested with a
gssd that uses the new upcall yet.

Most of these patches are just trivial preparatory cleanup, and the
first two (maybe three) patches might be worth applying independently.

--b.

[PATCH 1/9] rpc: remove unnecessary assignment

[J. Bruce Fields]

We're just about to kfree() gss_auth, so there's no point to setting any
of its fields.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 853a414..07bb395 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -650,7 +650,6 @@ static void
 gss_free(struct gss_auth *gss_auth)
 {
 	rpc_unlink(gss_auth->dentry);
-	gss_auth->dentry = NULL;
 	gss_mech_put(gss_auth->mech);
 
 	kfree(gss_auth);
-- 
1.5.5.rc1

[PATCH 2/9] rpc: factor out warning code from gss_pipe_destroy_msg

[J. Bruce Fields]

We'll want to call this from elsewhere soon.  And this is a bit nicer
anyway.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |   23 ++++++++++++++---------
 1 files changed, 14 insertions(+), 9 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 07bb395..1cb4bcc 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -369,6 +369,18 @@ gss_setup_upcall(struct rpc_clnt *clnt, struct gss_auth *gss_auth, struct rpc_cr
 	return gss_msg;
 }
 
+static void warn_gssd(void)
+{
+	static unsigned long ratelimit;
+	unsigned long now = jiffies;
+
+	if (time_after(now, ratelimit)) {
+		printk(KERN_WARNING "RPC: AUTH_GSS upcall timed out.\n"
+				"Please check user daemon is running.\n");
+		ratelimit = now + 15*HZ;
+	}
+}
+
 static inline int
 gss_refresh_upcall(struct rpc_task *task)
 {
@@ -568,21 +580,14 @@ static void
 gss_pipe_destroy_msg(struct rpc_pipe_msg *msg)
 {
 	struct gss_upcall_msg *gss_msg = container_of(msg, struct gss_upcall_msg, msg);
-	static unsigned long ratelimit;
 
 	if (msg->errno < 0) {
 		dprintk("RPC:       gss_pipe_destroy_msg releasing msg %p\n",
 				gss_msg);
 		atomic_inc(&gss_msg->count);
 		gss_unhash_msg(gss_msg);
-		if (msg->errno == -ETIMEDOUT) {
-			unsigned long now = jiffies;
-			if (time_after(now, ratelimit)) {
-				printk(KERN_WARNING "RPC: AUTH_GSS upcall timed out.\n"
-						    "Please check user daemon is running!\n");
-				ratelimit = now + 15*HZ;
-			}
-		}
+		if (msg->errno == -ETIMEDOUT)
+			warn_gssd();
 		gss_release_msg(gss_msg);
 	}
 }
-- 
1.5.5.rc1

[PATCH 3/9] rpc: minor gss_alloc_msg cleanup

[J. Bruce Fields]

I want to add a little more code here, so it'll be convenient to have
this flatter.

Also, I'll want to add another error condition, so it'll be more
convenient to return -ENOMEM than NULL in the error case.  The only
caller is already converting NULL to -ENOMEM anyway.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |   24 ++++++++++++------------
 1 files changed, 12 insertions(+), 12 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 1cb4bcc..7baa432 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -329,16 +329,16 @@ gss_alloc_msg(struct gss_auth *gss_auth, uid_t uid)
 	struct gss_upcall_msg *gss_msg;
 
 	gss_msg = kzalloc(sizeof(*gss_msg), GFP_NOFS);
-	if (gss_msg != NULL) {
-		INIT_LIST_HEAD(&gss_msg->list);
-		rpc_init_wait_queue(&gss_msg->rpc_waitqueue, "RPCSEC_GSS upcall waitq");
-		init_waitqueue_head(&gss_msg->waitqueue);
-		atomic_set(&gss_msg->count, 1);
-		gss_msg->msg.data = &gss_msg->uid;
-		gss_msg->msg.len = sizeof(gss_msg->uid);
-		gss_msg->uid = uid;
-		gss_msg->auth = gss_auth;
-	}
+	if (gss_msg == NULL)
+		return ERR_PTR(-ENOMEM);
+	INIT_LIST_HEAD(&gss_msg->list);
+	rpc_init_wait_queue(&gss_msg->rpc_waitqueue, "RPCSEC_GSS upcall waitq");
+	init_waitqueue_head(&gss_msg->waitqueue);
+	atomic_set(&gss_msg->count, 1);
+	gss_msg->msg.data = &gss_msg->uid;
+	gss_msg->msg.len = sizeof(gss_msg->uid);
+	gss_msg->uid = uid;
+	gss_msg->auth = gss_auth;
 	return gss_msg;
 }
 
@@ -355,8 +355,8 @@ gss_setup_upcall(struct rpc_clnt *clnt, struct gss_auth *gss_auth, struct rpc_cr
 		uid = 0;
 
 	gss_new = gss_alloc_msg(gss_auth, uid);
-	if (gss_new == NULL)
-		return ERR_PTR(-ENOMEM);
+	if (IS_ERR(gss_new))
+		return gss_new;
 	gss_msg = gss_add_msg(gss_auth, gss_new);
 	if (gss_msg == gss_new) {
 		int res = rpc_queue_upcall(gss_auth->dentry->d_inode, &gss_new->msg);
-- 
1.5.5.rc1

[PATCH 4/9] rpc: add an rpc_pipe_open method

[J. Bruce Fields]

We want to transition to a new gssd upcall which is text-based and more
easily extensible.

To simplify upgrades, as well as testing and debugging, it will help if
we can upgrade gssd (to a version which understands the new upcall)
without having to choose at boot (or module-load) time whether we want
the new or the old upcall.

We will do this by providing two different pipes: one named, as
currently, after the mechanism (normally "krb5"), and supporting the
old upcall.  One named "gssd" and supporting the new upcall version.

We allow gssd to indicate which version it supports by its choice of
which pipe to open.

As we have no interest in supporting *simultaneous* use of both
versions, we'll forbid opening both pipes at the same time.

So, add a new pipe_open callback to the rpc_pipefs api, which the gss
code can use to track which pipes have been open, and to refuse opens of
incompatible pipes.

We only need this to be called on the first open of a given pipe.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 include/linux/sunrpc/rpc_pipe_fs.h |    1 +
 net/sunrpc/rpc_pipe.c              |   21 ++++++++++++++-------
 2 files changed, 15 insertions(+), 7 deletions(-)

diff --git a/include/linux/sunrpc/rpc_pipe_fs.h b/include/linux/sunrpc/rpc_pipe_fs.h
index 51b977a..cea764c 100644
--- a/include/linux/sunrpc/rpc_pipe_fs.h
+++ b/include/linux/sunrpc/rpc_pipe_fs.h
@@ -15,6 +15,7 @@ struct rpc_pipe_ops {
 	ssize_t (*upcall)(struct file *, struct rpc_pipe_msg *, char __user *, size_t);
 	ssize_t (*downcall)(struct file *, const char __user *, size_t);
 	void (*release_pipe)(struct inode *);
+	int (*open_pipe)(struct inode *);
 	void (*destroy_msg)(struct rpc_pipe_msg *);
 };
 
diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index 23a2b8f..4171ab7 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -169,16 +169,23 @@ static int
 rpc_pipe_open(struct inode *inode, struct file *filp)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
+	int first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
 	int res = -ENXIO;
 
 	mutex_lock(&inode->i_mutex);
-	if (rpci->ops != NULL) {
-		if (filp->f_mode & FMODE_READ)
-			rpci->nreaders ++;
-		if (filp->f_mode & FMODE_WRITE)
-			rpci->nwriters ++;
-		res = 0;
+	if (rpci->ops == NULL)
+		goto out;
+	if (first_open && rpci->ops->open_pipe) {
+		res = rpci->ops->open_pipe(inode);
+		if (res)
+			goto out;
 	}
+	if (filp->f_mode & FMODE_READ)
+		rpci->nreaders++;
+	if (filp->f_mode & FMODE_WRITE)
+		rpci->nwriters++;
+	res = 0;
+out:
 	mutex_unlock(&inode->i_mutex);
 	return res;
 }
@@ -748,7 +755,7 @@ rpc_rmdir(struct dentry *dentry)
  * @name: name of pipe
  * @private: private data to associate with the pipe, for the caller's use
  * @ops: operations defining the behavior of the pipe: upcall, downcall,
- *	release_pipe, and destroy_msg.
+ *	release_pipe, open_pipe, and destroy_msg.
  * @flags: rpc_inode flags
  *
  * Data is made available for userspace to read by calls to
-- 
1.5.5.rc1

[PATCH 5/9] rpc: call release_pipe only on last close

[J. Bruce Fields]

I can't see any reason we particularly need to call this until the last
gssd closes the pipe.

Also, this allows to guarantee that open_pipe and release_pipe are
called strictly in pairs; open_pipe on the first open, release_pipe on
the last close.  That'll make it very easy for the gss code to keep
track of which pipes gssd is using.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/rpc_pipe.c |    9 ++++++---
 1 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index 4171ab7..712e1ec 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -126,13 +126,14 @@ rpc_close_pipes(struct inode *inode)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct rpc_pipe_ops *ops;
+	int last_close;
 
 	mutex_lock(&inode->i_mutex);
 	ops = rpci->ops;
 	if (ops != NULL) {
 		LIST_HEAD(free_list);
-
 		spin_lock(&inode->i_lock);
+		last_close = rpci->nreaders != 0 || rpci->nwriters != 0;
 		rpci->nreaders = 0;
 		list_splice_init(&rpci->in_upcall, &free_list);
 		list_splice_init(&rpci->pipe, &free_list);
@@ -141,7 +142,7 @@ rpc_close_pipes(struct inode *inode)
 		spin_unlock(&inode->i_lock);
 		rpc_purge_list(rpci, &free_list, ops->destroy_msg, -EPIPE);
 		rpci->nwriters = 0;
-		if (ops->release_pipe)
+		if (last_close && ops->release_pipe)
 			ops->release_pipe(inode);
 		cancel_delayed_work_sync(&rpci->queue_timeout);
 	}
@@ -195,6 +196,7 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct rpc_pipe_msg *msg;
+	int last_close;
 
 	mutex_lock(&inode->i_mutex);
 	if (rpci->ops == NULL)
@@ -221,7 +223,8 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
 					rpci->ops->destroy_msg, -EAGAIN);
 		}
 	}
-	if (rpci->ops->release_pipe)
+	last_close = rpci->nwriters == 0 && rpci->nreaders == 0;
+	if (last_close && rpci->ops->release_pipe)
 		rpci->ops->release_pipe(inode);
 out:
 	mutex_unlock(&inode->i_mutex);
-- 
1.5.5.rc1

[PATCH 6/9] rpc: track number of users of the gss upcall pipe

[J. Bruce Fields]

Keep a count of the number of pipes open plus the number of messages on
a pipe.  This count isn't used yet.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |   14 ++++++++++++++
 1 files changed, 14 insertions(+), 0 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 7baa432..45e6728 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -75,6 +75,8 @@ struct gss_auth {
 	struct dentry *dentry;
 };
 
+static atomic_t pipe_users = ATOMIC_INIT(0);
+
 static void gss_free_ctx(struct gss_cl_ctx *);
 static struct rpc_pipe_ops gss_upcall_ops;
 
@@ -237,6 +239,7 @@ gss_release_msg(struct gss_upcall_msg *gss_msg)
 {
 	if (!atomic_dec_and_test(&gss_msg->count))
 		return;
+	atomic_dec(&pipe_users);
 	BUG_ON(!list_empty(&gss_msg->list));
 	if (gss_msg->ctx != NULL)
 		gss_put_ctx(gss_msg->ctx);
@@ -331,6 +334,7 @@ gss_alloc_msg(struct gss_auth *gss_auth, uid_t uid)
 	gss_msg = kzalloc(sizeof(*gss_msg), GFP_NOFS);
 	if (gss_msg == NULL)
 		return ERR_PTR(-ENOMEM);
+	atomic_inc(&pipe_users);
 	INIT_LIST_HEAD(&gss_msg->list);
 	rpc_init_wait_queue(&gss_msg->rpc_waitqueue, "RPCSEC_GSS upcall waitq");
 	init_waitqueue_head(&gss_msg->waitqueue);
@@ -555,6 +559,13 @@ out:
 	return err;
 }
 
+static int
+gss_pipe_open(struct inode *inode)
+{
+	atomic_inc(&pipe_users);
+	return 0;
+}
+
 static void
 gss_pipe_release(struct inode *inode)
 {
@@ -574,6 +585,8 @@ gss_pipe_release(struct inode *inode)
 		spin_lock(&inode->i_lock);
 	}
 	spin_unlock(&inode->i_lock);
+
+	atomic_dec(&pipe_users);
 }
 
 static void
@@ -1342,6 +1355,7 @@ static struct rpc_pipe_ops gss_upcall_ops = {
 	.upcall		= gss_pipe_upcall,
 	.downcall	= gss_pipe_downcall,
 	.destroy_msg	= gss_pipe_destroy_msg,
+	.open_pipe	= gss_pipe_open,
 	.release_pipe	= gss_pipe_release,
 };
 
-- 
1.5.5.rc1

[PATCH 7/9] rpc: use count of pipe openers to wait for first open

[J. Bruce Fields]

Introduce a global variable pipe_version which will eventually be used
to keep track of which version of the upcall gssd is using.

For now, though, it only keeps track of whether any pipe is open or not;
it is negative if not, zero if one is opened.  We use this to wait for
the first gssd to open a pipe.

(Minor digression: note this waits only for the very first open of any
pipe, not for the first open of a pipe for a given auth; thus we still
need the RPC_PIPE_WAIT_FOR_OPEN behavior to wait for gssd to open new
pipes that pop up on subsequent mounts.)

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |   64 ++++++++++++++++++++++++++++++++++++++--
 1 files changed, 61 insertions(+), 3 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 45e6728..3046f2b 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -75,7 +75,12 @@ struct gss_auth {
 	struct dentry *dentry;
 };
 
+/* pipe_version >= 0 if and only if someone has a pipe open. */
+static int pipe_version = -1;
 static atomic_t pipe_users = ATOMIC_INIT(0);
+static DEFINE_SPINLOCK(pipe_version_lock);
+static struct rpc_wait_queue pipe_version_rpc_waitqueue;
+static DECLARE_WAIT_QUEUE_HEAD(pipe_version_waitqueue);
 
 static void gss_free_ctx(struct gss_cl_ctx *);
 static struct rpc_pipe_ops gss_upcall_ops;
@@ -234,12 +239,34 @@ struct gss_upcall_msg {
 	struct gss_cl_ctx *ctx;
 };
 
+static int get_pipe_version(void)
+{
+	int ret;
+
+	spin_lock(&pipe_version_lock);
+	if (pipe_version >= 0) {
+		atomic_inc(&pipe_users);
+		ret = 0;
+	} else
+		ret = -EAGAIN;
+	spin_unlock(&pipe_version_lock);
+	return ret;
+}
+
+static void put_pipe_version(void)
+{
+	if (atomic_dec_and_lock(&pipe_users, &pipe_version_lock)) {
+		pipe_version = -1;
+		spin_unlock(&pipe_version_lock);
+	}
+}
+
 static void
 gss_release_msg(struct gss_upcall_msg *gss_msg)
 {
 	if (!atomic_dec_and_test(&gss_msg->count))
 		return;
-	atomic_dec(&pipe_users);
+	put_pipe_version();
 	BUG_ON(!list_empty(&gss_msg->list));
 	if (gss_msg->ctx != NULL)
 		gss_put_ctx(gss_msg->ctx);
@@ -330,11 +357,16 @@ static inline struct gss_upcall_msg *
 gss_alloc_msg(struct gss_auth *gss_auth, uid_t uid)
 {
 	struct gss_upcall_msg *gss_msg;
+	int vers;
 
 	gss_msg = kzalloc(sizeof(*gss_msg), GFP_NOFS);
 	if (gss_msg == NULL)
 		return ERR_PTR(-ENOMEM);
-	atomic_inc(&pipe_users);
+	vers = get_pipe_version();
+	if (vers < 0) {
+		kfree(gss_msg);
+		return ERR_PTR(vers);
+	}
 	INIT_LIST_HEAD(&gss_msg->list);
 	rpc_init_wait_queue(&gss_msg->rpc_waitqueue, "RPCSEC_GSS upcall waitq");
 	init_waitqueue_head(&gss_msg->waitqueue);
@@ -400,6 +432,14 @@ gss_refresh_upcall(struct rpc_task *task)
 	dprintk("RPC: %5u gss_refresh_upcall for uid %u\n", task->tk_pid,
 								cred->cr_uid);
 	gss_msg = gss_setup_upcall(task->tk_client, gss_auth, cred);
+	if (IS_ERR(gss_msg) == -EAGAIN) {
+		/* XXX: warning on the first, under the assumption we
+		 * shouldn't normally hit this case on a refresh. */
+		warn_gssd();
+		task->tk_timeout = 15*HZ;
+		rpc_sleep_on(&pipe_version_rpc_waitqueue, task, NULL);
+		return 0;
+	}
 	if (IS_ERR(gss_msg)) {
 		err = PTR_ERR(gss_msg);
 		goto out;
@@ -437,7 +477,17 @@ gss_create_upcall(struct gss_auth *gss_auth, struct gss_cred *gss_cred)
 	int err = 0;
 
 	dprintk("RPC:       gss_upcall for uid %u\n", cred->cr_uid);
+retry:
 	gss_msg = gss_setup_upcall(gss_auth->client, gss_auth, cred);
+	if (PTR_ERR(gss_msg) == -EAGAIN) {
+		err = wait_event_interruptible_timeout(pipe_version_waitqueue,
+				pipe_version >= 0, 15*HZ);
+		if (err)
+			goto out;
+		if (pipe_version < 0)
+			warn_gssd();
+		goto retry;
+	}
 	if (IS_ERR(gss_msg)) {
 		err = PTR_ERR(gss_msg);
 		goto out;
@@ -562,7 +612,14 @@ out:
 static int
 gss_pipe_open(struct inode *inode)
 {
+	spin_lock(&pipe_version_lock);
+	if (pipe_version < 0) {
+		pipe_version = 0;
+		rpc_wake_up(&pipe_version_rpc_waitqueue);
+		wake_up(&pipe_version_waitqueue);
+	}
 	atomic_inc(&pipe_users);
+	spin_unlock(&pipe_version_lock);
 	return 0;
 }
 
@@ -586,7 +643,7 @@ gss_pipe_release(struct inode *inode)
 	}
 	spin_unlock(&inode->i_lock);
 
-	atomic_dec(&pipe_users);
+	put_pipe_version();
 }
 
 static void
@@ -1372,6 +1429,7 @@ static int __init init_rpcsec_gss(void)
 	err = gss_svc_init();
 	if (err)
 		goto out_unregister;
+	rpc_init_wait_queue(&pipe_version_rpc_waitqueue, "gss pipe version");
 	return 0;
 out_unregister:
 	rpcauth_unregister(&authgss_ops);
-- 
1.5.5.rc1

[PATCH 8/9] rpc: store pointer to pipe inode in gss upcall message

[J. Bruce Fields]

Keep a pointer to the inode that the message is queued on in the struct
gss_upcall_msg.  This will be convenient, especially after we have a
choice of two pipes that an upcall could be queued on.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |   20 ++++++++++++--------
 1 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 3046f2b..96c1bab 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -234,6 +234,7 @@ struct gss_upcall_msg {
 	struct rpc_pipe_msg msg;
 	struct list_head list;
 	struct gss_auth *auth;
+	struct rpc_inode *inode;
 	struct rpc_wait_queue rpc_waitqueue;
 	wait_queue_head_t waitqueue;
 	struct gss_cl_ctx *ctx;
@@ -296,8 +297,8 @@ __gss_find_upcall(struct rpc_inode *rpci, uid_t uid)
 static inline struct gss_upcall_msg *
 gss_add_msg(struct gss_auth *gss_auth, struct gss_upcall_msg *gss_msg)
 {
-	struct inode *inode = gss_auth->dentry->d_inode;
-	struct rpc_inode *rpci = RPC_I(inode);
+	struct rpc_inode *rpci = gss_msg->inode;
+	struct inode *inode = &rpci->vfs_inode;
 	struct gss_upcall_msg *old;
 
 	spin_lock(&inode->i_lock);
@@ -323,8 +324,7 @@ __gss_unhash_msg(struct gss_upcall_msg *gss_msg)
 static void
 gss_unhash_msg(struct gss_upcall_msg *gss_msg)
 {
-	struct gss_auth *gss_auth = gss_msg->auth;
-	struct inode *inode = gss_auth->dentry->d_inode;
+	struct inode *inode = &gss_msg->inode->vfs_inode;
 
 	if (list_empty(&gss_msg->list))
 		return;
@@ -340,7 +340,7 @@ gss_upcall_callback(struct rpc_task *task)
 	struct gss_cred *gss_cred = container_of(task->tk_msg.rpc_cred,
 			struct gss_cred, gc_base);
 	struct gss_upcall_msg *gss_msg = gss_cred->gc_upcall;
-	struct inode *inode = gss_msg->auth->dentry->d_inode;
+	struct inode *inode = &gss_msg->inode->vfs_inode;
 
 	spin_lock(&inode->i_lock);
 	if (gss_msg->ctx)
@@ -367,6 +367,7 @@ gss_alloc_msg(struct gss_auth *gss_auth, uid_t uid)
 		kfree(gss_msg);
 		return ERR_PTR(vers);
 	}
+	gss_msg->inode = RPC_I(gss_auth->dentry->d_inode);
 	INIT_LIST_HEAD(&gss_msg->list);
 	rpc_init_wait_queue(&gss_msg->rpc_waitqueue, "RPCSEC_GSS upcall waitq");
 	init_waitqueue_head(&gss_msg->waitqueue);
@@ -395,7 +396,8 @@ gss_setup_upcall(struct rpc_clnt *clnt, struct gss_auth *gss_auth, struct rpc_cr
 		return gss_new;
 	gss_msg = gss_add_msg(gss_auth, gss_new);
 	if (gss_msg == gss_new) {
-		int res = rpc_queue_upcall(gss_auth->dentry->d_inode, &gss_new->msg);
+		struct inode *inode = &gss_new->inode->vfs_inode;
+		int res = rpc_queue_upcall(inode, &gss_new->msg);
 		if (res) {
 			gss_unhash_msg(gss_new);
 			gss_msg = ERR_PTR(res);
@@ -426,7 +428,7 @@ gss_refresh_upcall(struct rpc_task *task)
 	struct gss_cred *gss_cred = container_of(cred,
 			struct gss_cred, gc_base);
 	struct gss_upcall_msg *gss_msg;
-	struct inode *inode = gss_auth->dentry->d_inode;
+	struct inode *inode;
 	int err = 0;
 
 	dprintk("RPC: %5u gss_refresh_upcall for uid %u\n", task->tk_pid,
@@ -444,6 +446,7 @@ gss_refresh_upcall(struct rpc_task *task)
 		err = PTR_ERR(gss_msg);
 		goto out;
 	}
+	inode = &gss_msg->inode->vfs_inode;
 	spin_lock(&inode->i_lock);
 	if (gss_cred->gc_upcall != NULL)
 		rpc_sleep_on(&gss_cred->gc_upcall->rpc_waitqueue, task, NULL);
@@ -470,7 +473,7 @@ out:
 static inline int
 gss_create_upcall(struct gss_auth *gss_auth, struct gss_cred *gss_cred)
 {
-	struct inode *inode = gss_auth->dentry->d_inode;
+	struct inode *inode;
 	struct rpc_cred *cred = &gss_cred->gc_base;
 	struct gss_upcall_msg *gss_msg;
 	DEFINE_WAIT(wait);
@@ -492,6 +495,7 @@ retry:
 		err = PTR_ERR(gss_msg);
 		goto out;
 	}
+	inode = &gss_msg->inode->vfs_inode;
 	for (;;) {
 		prepare_to_wait(&gss_msg->waitqueue, &wait, TASK_INTERRUPTIBLE);
 		spin_lock(&inode->i_lock);
-- 
1.5.5.rc1

[PATCH 9/9] rpc: implement new upcall

[J. Bruce Fields]

Implement the new upcall.  We decide which version of the upcall gssd
will use (new or old), by creating both pipes (the new one named "gssd",
the old one named after the mechanism (e.g., "krb5")), and then waiting
to see which version gssd actually opens.

We don't permit pipes of the two different types to be opened at once.

Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
---
 net/sunrpc/auth_gss/auth_gss.c |   95 +++++++++++++++++++++++++++++++++------
 1 files changed, 80 insertions(+), 15 deletions(-)

diff --git a/net/sunrpc/auth_gss/auth_gss.c b/net/sunrpc/auth_gss/auth_gss.c
index 96c1bab..fbe04cb 100644
--- a/net/sunrpc/auth_gss/auth_gss.c
+++ b/net/sunrpc/auth_gss/auth_gss.c
@@ -72,7 +72,13 @@ struct gss_auth {
 	struct gss_api_mech *mech;
 	enum rpc_gss_svc service;
 	struct rpc_clnt *client;
-	struct dentry *dentry;
+	/*
+	 * There are two upcall pipes; dentry[1], named "gssd", is used
+	 * for the new text-based upcall; dentry[0] is named after the
+	 * mechanism (for example, "krb5") and exists for
+	 * backwards-compatibility with older gssd's.
+	 */
+	struct dentry *dentry[2];
 };
 
 /* pipe_version >= 0 if and only if someone has a pipe open. */
@@ -227,6 +233,7 @@ err:
 	return p;
 }
 
+#define UPCALL_BUF_LEN 128
 
 struct gss_upcall_msg {
 	atomic_t count;
@@ -238,6 +245,7 @@ struct gss_upcall_msg {
 	struct rpc_wait_queue rpc_waitqueue;
 	wait_queue_head_t waitqueue;
 	struct gss_cl_ctx *ctx;
+	char databuf[UPCALL_BUF_LEN];
 };
 
 static int get_pipe_version(void)
@@ -247,7 +255,7 @@ static int get_pipe_version(void)
 	spin_lock(&pipe_version_lock);
 	if (pipe_version >= 0) {
 		atomic_inc(&pipe_users);
-		ret = 0;
+		ret = pipe_version;
 	} else
 		ret = -EAGAIN;
 	spin_unlock(&pipe_version_lock);
@@ -353,6 +361,29 @@ gss_upcall_callback(struct rpc_task *task)
 	gss_release_msg(gss_msg);
 }
 
+static void gss_encode_v0_msg(struct gss_upcall_msg *gss_msg)
+{
+	gss_msg->msg.data = &gss_msg->uid;
+	gss_msg->msg.len = sizeof(gss_msg->uid);
+}
+
+static void gss_encode_v1_msg(struct gss_upcall_msg *gss_msg)
+{
+	gss_msg->msg.len = sprintf(gss_msg->databuf, "mech=%s uid=%d\n",
+				   gss_msg->auth->mech->gm_name,
+				   gss_msg->uid);
+	gss_msg->msg.data = gss_msg->databuf;
+	BUG_ON(gss_msg->msg.len > UPCALL_BUF_LEN);
+}
+
+static void gss_encode_msg(struct gss_upcall_msg *gss_msg)
+{
+	if (pipe_version == 0)
+		gss_encode_v0_msg(gss_msg);
+	else /* pipe_version == 1 */
+		gss_encode_v1_msg(gss_msg);
+}
+
 static inline struct gss_upcall_msg *
 gss_alloc_msg(struct gss_auth *gss_auth, uid_t uid)
 {
@@ -367,13 +398,12 @@ gss_alloc_msg(struct gss_auth *gss_auth, uid_t uid)
 		kfree(gss_msg);
 		return ERR_PTR(vers);
 	}
-	gss_msg->inode = RPC_I(gss_auth->dentry->d_inode);
+	gss_msg->inode = RPC_I(gss_auth->dentry[vers]->d_inode);
 	INIT_LIST_HEAD(&gss_msg->list);
 	rpc_init_wait_queue(&gss_msg->rpc_waitqueue, "RPCSEC_GSS upcall waitq");
 	init_waitqueue_head(&gss_msg->waitqueue);
 	atomic_set(&gss_msg->count, 1);
-	gss_msg->msg.data = &gss_msg->uid;
-	gss_msg->msg.len = sizeof(gss_msg->uid);
+	gss_encode_msg(gss_msg);
 	gss_msg->uid = uid;
 	gss_msg->auth = gss_auth;
 	return gss_msg;
@@ -613,18 +643,40 @@ out:
 	return err;
 }
 
+static int gss_pipe_version(struct gss_auth *gss_auth, struct inode *inode)
+{
+	if (gss_auth->dentry[0]->d_inode == inode)
+		return 0;
+	if (gss_auth->dentry[1]->d_inode == inode)
+		return 1;
+	BUG();
+}
+
 static int
 gss_pipe_open(struct inode *inode)
 {
+	struct rpc_inode *rpci = RPC_I(inode);
+	struct rpc_clnt *clnt = rpci->private;
+	struct gss_auth *gss_auth = container_of(clnt->cl_auth,
+				struct gss_auth, rpc_auth);
+	int new_version = gss_pipe_version(gss_auth, inode);
+	int ret = 0;
+
 	spin_lock(&pipe_version_lock);
 	if (pipe_version < 0) {
-		pipe_version = 0;
+		/* First open of any gss pipe determines the version: */
+		pipe_version = new_version;
 		rpc_wake_up(&pipe_version_rpc_waitqueue);
 		wake_up(&pipe_version_waitqueue);
+	} else if (pipe_version != new_version) {
+		/* Trying to open a pipe of a different version */
+		ret = -EBUSY;
+		goto out;
 	}
 	atomic_inc(&pipe_users);
+out:
 	spin_unlock(&pipe_version_lock);
-	return 0;
+	return ret;
 }
 
 static void
@@ -702,20 +754,32 @@ gss_create(struct rpc_clnt *clnt, rpc_authflavor_t flavor)
 	atomic_set(&auth->au_count, 1);
 	kref_init(&gss_auth->kref);
 
-	gss_auth->dentry = rpc_mkpipe(clnt->cl_dentry, gss_auth->mech->gm_name,
-			clnt, &gss_upcall_ops, RPC_PIPE_WAIT_FOR_OPEN);
-	if (IS_ERR(gss_auth->dentry)) {
-		err = PTR_ERR(gss_auth->dentry);
+	gss_auth->dentry[0] = rpc_mkpipe(clnt->cl_dentry,
+					 gss_auth->mech->gm_name,
+					 clnt, &gss_upcall_ops,
+					 RPC_PIPE_WAIT_FOR_OPEN);
+	if (IS_ERR(gss_auth->dentry[0])) {
+		err = PTR_ERR(gss_auth->dentry[0]);
 		goto err_put_mech;
 	}
 
+	gss_auth->dentry[1] = rpc_mkpipe(clnt->cl_dentry,
+					 "gssd",
+					 clnt, &gss_upcall_ops,
+					 RPC_PIPE_WAIT_FOR_OPEN);
+	if (IS_ERR(gss_auth->dentry[1])) {
+		err = PTR_ERR(gss_auth->dentry[1]);
+		goto err_unlink_pipe_0;
+	}
 	err = rpcauth_init_credcache(auth);
 	if (err)
-		goto err_unlink_pipe;
+		goto err_unlink_pipe_1;
 
 	return auth;
-err_unlink_pipe:
-	rpc_unlink(gss_auth->dentry);
+err_unlink_pipe_1:
+	rpc_unlink(gss_auth->dentry[1]);
+err_unlink_pipe_0:
+	rpc_unlink(gss_auth->dentry[0]);
 err_put_mech:
 	gss_mech_put(gss_auth->mech);
 err_free:
@@ -728,7 +792,8 @@ out_dec:
 static void
 gss_free(struct gss_auth *gss_auth)
 {
-	rpc_unlink(gss_auth->dentry);
+	rpc_unlink(gss_auth->dentry[1]);
+	rpc_unlink(gss_auth->dentry[0]);
 	gss_mech_put(gss_auth->mech);
 
 	kfree(gss_auth);
-- 
1.5.5.rc1

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[Trond Myklebust]

On Sun, 2008-11-09 at 16:04 -0500, J. Bruce Fields wrote:
> I can't see any reason we particularly need to call this until the last
> gssd closes the pipe.

There's a very good reason: if we call rpc_close_pipes() then it is
because the kernel listener is shutting down. At that point, we want to
return EPIPE for all future read() or write() attempts by gssd.

> Also, this allows to guarantee that open_pipe and release_pipe are
> called strictly in pairs; open_pipe on the first open, release_pipe on
> the last close.  That'll make it very easy for the gss code to keep
> track of which pipes gssd is using.

...unless the gss code is no longer running.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 02:11:37PM -0500, Trond Myklebust wrote:
> On Sun, 2008-11-09 at 16:04 -0500, J. Bruce Fields wrote:
> > I can't see any reason we particularly need to call this until the last
> > gssd closes the pipe.
> 
> There's a very good reason: if we call rpc_close_pipes() then it is
> because the kernel listener is shutting down. At that point, we want to
> return EPIPE for all future read() or write() attempts by gssd.

Sure.  This doesn't interfere with that--we're still calling
release_pipe() itself (which is the function that sets rpci->ops to
NULL), it's only the ->release_pipe() op that we're not calling.

> > Also, this allows to guarantee that open_pipe and release_pipe are
> > called strictly in pairs; open_pipe on the first open, release_pipe on
> > the last close.  That'll make it very easy for the gss code to keep
> > track of which pipes gssd is using.
> 
> ...unless the gss code is no longer running.

I'm not following you.

Note the patch calls ->release_pipe() at close_pipe() time if there are
still opens (and then subsequent closes become no-ops).  So it does
guarantee that there's always a matching call.

So my patch description was wrong: "last close" should be something like
"last close, or destruction of an opened pipe."  I'll fix that.

--b.

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[Trond Myklebust]

On Mon, 2008-11-10 at 14:49 -0500, J. Bruce Fields wrote:
> On Mon, Nov 10, 2008 at 02:11:37PM -0500, Trond Myklebust wrote:
> > On Sun, 2008-11-09 at 16:04 -0500, J. Bruce Fields wrote:
> > > I can't see any reason we particularly need to call this until the last
> > > gssd closes the pipe.
> > 
> > There's a very good reason: if we call rpc_close_pipes() then it is
> > because the kernel listener is shutting down. At that point, we want to
> > return EPIPE for all future read() or write() attempts by gssd.
> 
> Sure.  This doesn't interfere with that--we're still calling
> release_pipe() itself (which is the function that sets rpci->ops to
> NULL), it's only the ->release_pipe() op that we're not calling.
>
> > > Also, this allows to guarantee that open_pipe and release_pipe are
> > > called strictly in pairs; open_pipe on the first open, release_pipe on
> > > the last close.  That'll make it very easy for the gss code to keep
> > > track of which pipes gssd is using.
> > 
> > ...unless the gss code is no longer running.
> 
> I'm not following you.
> 
> Note the patch calls ->release_pipe() at close_pipe() time if there are
> still opens (and then subsequent closes become no-ops).  So it does
> guarantee that there's always a matching call.

No it doesn't. Now that rpci->ops has been set to NULL by
rpc_close_pipes(), exactly how are you going to call
rpci->ops->release_pipe() at a later time?

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 03:01:19PM -0500, Trond Myklebust wrote:
> On Mon, 2008-11-10 at 14:49 -0500, J. Bruce Fields wrote:
> > On Mon, Nov 10, 2008 at 02:11:37PM -0500, Trond Myklebust wrote:
> > > On Sun, 2008-11-09 at 16:04 -0500, J. Bruce Fields wrote:
> > > > I can't see any reason we particularly need to call this until the last
> > > > gssd closes the pipe.
> > > 
> > > There's a very good reason: if we call rpc_close_pipes() then it is
> > > because the kernel listener is shutting down. At that point, we want to
> > > return EPIPE for all future read() or write() attempts by gssd.
> > 
> > Sure.  This doesn't interfere with that--we're still calling
> > release_pipe() itself (which is the function that sets rpci->ops to
> > NULL), it's only the ->release_pipe() op that we're not calling.
> >
> > > > Also, this allows to guarantee that open_pipe and release_pipe are
> > > > called strictly in pairs; open_pipe on the first open, release_pipe on
> > > > the last close.  That'll make it very easy for the gss code to keep
> > > > track of which pipes gssd is using.
> > > 
> > > ...unless the gss code is no longer running.
> > 
> > I'm not following you.
> > 
> > Note the patch calls ->release_pipe() at close_pipe() time if there are
> > still opens (and then subsequent closes become no-ops).  So it does
> > guarantee that there's always a matching call.
> 
> No it doesn't. Now that rpci->ops has been set to NULL by
> rpc_close_pipes(), exactly how are you going to call
> rpci->ops->release_pipe() at a later time?

Again, I'm dealing with that case by calling release_pipe() from
rpc_close_pipes(), just as the current code does--the only change being
to do that only when there are still opens:

		last_close = rpci->nreaders != 0 || rpci->nwriters != 0;
 		rpci->nreaders = 0;
		...
 		rpci->nwriters = 0;
		if (last_close && ops->release_pipe)
 			ops->release_pipe(inode);

--b.

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[Trond Myklebust]

On Mon, 2008-11-10 at 15:07 -0500, J. Bruce Fields wrote:

> Again, I'm dealing with that case by calling release_pipe() from
> rpc_close_pipes(), just as the current code does--the only change being
> to do that only when there are still opens:
> 
> 		last_close = rpci->nreaders != 0 || rpci->nwriters != 0;
>  		rpci->nreaders = 0;
> 		...
>  		rpci->nwriters = 0;
> 		if (last_close && ops->release_pipe)
>  			ops->release_pipe(inode);

Which means that if the kernel calls rpc_close_pipes() before gssd has
managed to close, then you _NEVER_ call ops->release_pipe()...



-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 03:11:38PM -0500, Trond Myklebust wrote:
> On Mon, 2008-11-10 at 15:07 -0500, J. Bruce Fields wrote:
> 
> > Again, I'm dealing with that case by calling release_pipe() from
> > rpc_close_pipes(), just as the current code does--the only change being
> > to do that only when there are still opens:
> > 
> > 		last_close = rpci->nreaders != 0 || rpci->nwriters != 0;
> >  		rpci->nreaders = 0;
> > 		...
> >  		rpci->nwriters = 0;
> > 		if (last_close && ops->release_pipe)
> >  			ops->release_pipe(inode);
> 
> Which means that if the kernel calls rpc_close_pipes() before gssd has
> managed to close, then you _NEVER_ call ops->release_pipe()...

So, I take "before gssd has managed to close" to mean that gssd is still
holding the file open.  Thus the statement

	last_close = rpci->nreaders != 0 || rpci->nwriters != 0;

evaluates to true; either nreaders or nwriters must be nonzero.

(And note the open and close code that modifes nreaders and nwriters is
all serialized with this code by the i_mutex.)

--b.

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[Trond Myklebust]

On Mon, 2008-11-10 at 15:17 -0500, J. Bruce Fields wrote:
> On Mon, Nov 10, 2008 at 03:11:38PM -0500, Trond Myklebust wrote:
> > On Mon, 2008-11-10 at 15:07 -0500, J. Bruce Fields wrote:
> > 
> > > Again, I'm dealing with that case by calling release_pipe() from
> > > rpc_close_pipes(), just as the current code does--the only change being
> > > to do that only when there are still opens:
> > > 
> > > 		last_close = rpci->nreaders != 0 || rpci->nwriters != 0;
> > >  		rpci->nreaders = 0;
> > > 		...
> > >  		rpci->nwriters = 0;
> > > 		if (last_close && ops->release_pipe)
> > >  			ops->release_pipe(inode);
> > 
> > Which means that if the kernel calls rpc_close_pipes() before gssd has
> > managed to close, then you _NEVER_ call ops->release_pipe()...
> 
> So, I take "before gssd has managed to close" to mean that gssd is still
> holding the file open.  Thus the statement
> 
> 	last_close = rpci->nreaders != 0 || rpci->nwriters != 0;
> 
> evaluates to true; either nreaders or nwriters must be nonzero.
> 
> (And note the open and close code that modifes nreaders and nwriters is
> all serialized with this code by the i_mutex.)
> 
> --b.

Exactly... That is a very common situation that happens pretty much
every time you unmount. The kernel closes the pipe on its side, and
removes the dentry; it doesn't wait for gssd to close the pipe.

Trond

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 03:21:55PM -0500, Trond Myklebust wrote:
> On Mon, 2008-11-10 at 15:17 -0500, J. Bruce Fields wrote:
> > On Mon, Nov 10, 2008 at 03:11:38PM -0500, Trond Myklebust wrote:
> > > On Mon, 2008-11-10 at 15:07 -0500, J. Bruce Fields wrote:
> > > 
> > > > Again, I'm dealing with that case by calling release_pipe() from
> > > > rpc_close_pipes(), just as the current code does--the only change being
> > > > to do that only when there are still opens:
> > > > 
> > > > 		last_close = rpci->nreaders != 0 || rpci->nwriters != 0;
> > > >  		rpci->nreaders = 0;
> > > > 		...
> > > >  		rpci->nwriters = 0;
> > > > 		if (last_close && ops->release_pipe)
> > > >  			ops->release_pipe(inode);
> > > 
> > > Which means that if the kernel calls rpc_close_pipes() before gssd has
> > > managed to close, then you _NEVER_ call ops->release_pipe()...
> > 
> > So, I take "before gssd has managed to close" to mean that gssd is still
> > holding the file open.  Thus the statement
> > 
> > 	last_close = rpci->nreaders != 0 || rpci->nwriters != 0;
> > 
> > evaluates to true; either nreaders or nwriters must be nonzero.
> > 
> > (And note the open and close code that modifes nreaders and nwriters is
> > all serialized with this code by the i_mutex.)
> > 
> > --b.
> 
> Exactly... That is a very common situation that happens pretty much
> every time you unmount. The kernel closes the pipe on its side, and
> removes the dentry; it doesn't wait for gssd to close the pipe.

Yep.  And when we're in that situation, last_close is true, so

	if (last_close && ops->release_pipe)
		ops->release_pipe(inode)

does ensure that calls to the ->open_pipe() and ->release_pipe() methods
balance, as originally claimed.

Maybe it'd be clearer to call that variable "still_open", or something?

	still_open = rpci->nreaders != 0 || rpci->nwriters != 0;
	...
	if (still_open && ops->release_pipe)
		ops->release_pipe(inode)

??

--b.

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 03:26:40PM -0500, bfields wrote:
> Yep.  And when we're in that situation, last_close is true, so
> 
> 	if (last_close && ops->release_pipe)
> 		ops->release_pipe(inode)
> 
> does ensure that calls to the ->open_pipe() and ->release_pipe() methods
> balance, as originally claimed.

(Trivial to verify, by the way: search the code for every assignment to
nreaders or nwriters, and verify that:

	nreaders+nwriters is always initialized to zero.
	each time nreaders+nwriters transitions from 0 to positive, we
		call ->pipe_open().
	each time nreaders+nwriters transitions from positive to 0, we
		call ->pipe_release().
	nreaders+nwriters always transitions back to 0 before the
		containing inode is freed.

And all of this happens under the i_mutex.)

--b.

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[Trond Myklebust]

On Mon, 2008-11-10 at 15:26 -0500, J. Bruce Fields wrote:
> Yep.  And when we're in that situation, last_close is true, so
> 
> 	if (last_close && ops->release_pipe)
> 		ops->release_pipe(inode)
> 
> does ensure that calls to the ->open_pipe() and ->release_pipe() methods
> balance, as originally claimed.
> 
> Maybe it'd be clearer to call that variable "still_open", or something?
> 
> 	still_open = rpci->nreaders != 0 || rpci->nwriters != 0;
> 	...
> 	if (still_open && ops->release_pipe)
> 		ops->release_pipe(inode)

Call it 'need_release'. It is definitely confusing right now...

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 5/9] rpc: call release_pipe only on last close

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 04:18:03PM -0500, Trond Myklebust wrote:
> On Mon, 2008-11-10 at 15:26 -0500, J. Bruce Fields wrote:
> > Yep.  And when we're in that situation, last_close is true, so
> > 
> > 	if (last_close && ops->release_pipe)
> > 		ops->release_pipe(inode)
> > 
> > does ensure that calls to the ->open_pipe() and ->release_pipe() methods
> > balance, as originally claimed.
> > 
> > Maybe it'd be clearer to call that variable "still_open", or something?
> > 
> > 	still_open = rpci->nreaders != 0 || rpci->nwriters != 0;
> > 	...
> > 	if (still_open && ops->release_pipe)
> > 		ops->release_pipe(inode)
> 
> Call it 'need_release'. It is definitely confusing right now...

Sure.

--b.

commit 572df09a77f3cfdca9d5fba4ca736919294eb177
Author: J. Bruce Fields <bfields@citi.umich.edu>
Date:   Fri Nov 7 18:53:24 2008 -0500

    rpc: call release_pipe only on last close
    
    I can't see any reason we need to call this until either the kernel or
    the last gssd closes the pipe.
    
    Also, this allows to guarantee that open_pipe and release_pipe are
    called strictly in pairs; open_pipe on gssd's first open, release_pipe
    on gssd's last close (or on the close of the kernel side of the pipe, if
    that comes first).
    
    That will make it very easy for the gss code to keep track of which
    pipes gssd is using.
    
    Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>

diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index 4171ab7..a68aaac 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -126,13 +126,14 @@ rpc_close_pipes(struct inode *inode)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct rpc_pipe_ops *ops;
+	int need_release;
 
 	mutex_lock(&inode->i_mutex);
 	ops = rpci->ops;
 	if (ops != NULL) {
 		LIST_HEAD(free_list);
-
 		spin_lock(&inode->i_lock);
+		need_release = rpci->nreaders != 0 || rpci->nwriters != 0;
 		rpci->nreaders = 0;
 		list_splice_init(&rpci->in_upcall, &free_list);
 		list_splice_init(&rpci->pipe, &free_list);
@@ -141,7 +142,7 @@ rpc_close_pipes(struct inode *inode)
 		spin_unlock(&inode->i_lock);
 		rpc_purge_list(rpci, &free_list, ops->destroy_msg, -EPIPE);
 		rpci->nwriters = 0;
-		if (ops->release_pipe)
+		if (need_release && ops->release_pipe)
 			ops->release_pipe(inode);
 		cancel_delayed_work_sync(&rpci->queue_timeout);
 	}
@@ -169,12 +170,13 @@ static int
 rpc_pipe_open(struct inode *inode, struct file *filp)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
-	int first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
+	int first_open;
 	int res = -ENXIO;
 
 	mutex_lock(&inode->i_mutex);
 	if (rpci->ops == NULL)
 		goto out;
+	first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
 	if (first_open && rpci->ops->open_pipe) {
 		res = rpci->ops->open_pipe(inode);
 		if (res)
@@ -195,6 +197,7 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct rpc_pipe_msg *msg;
+	int last_close;
 
 	mutex_lock(&inode->i_mutex);
 	if (rpci->ops == NULL)
@@ -221,7 +224,8 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
 					rpci->ops->destroy_msg, -EAGAIN);
 		}
 	}
-	if (rpci->ops->release_pipe)
+	last_close = rpci->nwriters == 0 && rpci->nreaders == 0;
+	if (last_close && rpci->ops->release_pipe)
 		rpci->ops->release_pipe(inode);
 out:
 	mutex_unlock(&inode->i_mutex);

Re: [PATCH 4/9] rpc: add an rpc_pipe_open method

[Trond Myklebust]

On Sun, 2008-11-09 at 16:04 -0500, J. Bruce Fields wrote:
> We want to transition to a new gssd upcall which is text-based and more
> easily extensible.
> 
> To simplify upgrades, as well as testing and debugging, it will help if
> we can upgrade gssd (to a version which understands the new upcall)
> without having to choose at boot (or module-load) time whether we want
> the new or the old upcall.
> 
> We will do this by providing two different pipes: one named, as
> currently, after the mechanism (normally "krb5"), and supporting the
> old upcall.  One named "gssd" and supporting the new upcall version.
> 
> We allow gssd to indicate which version it supports by its choice of
> which pipe to open.
> 
> As we have no interest in supporting *simultaneous* use of both
> versions, we'll forbid opening both pipes at the same time.
> 
> So, add a new pipe_open callback to the rpc_pipefs api, which the gss
> code can use to track which pipes have been open, and to refuse opens of
> incompatible pipes.
> 
> We only need this to be called on the first open of a given pipe.
> 
> Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
> ---
>  include/linux/sunrpc/rpc_pipe_fs.h |    1 +
>  net/sunrpc/rpc_pipe.c              |   21 ++++++++++++++-------
>  2 files changed, 15 insertions(+), 7 deletions(-)
> 
> diff --git a/include/linux/sunrpc/rpc_pipe_fs.h b/include/linux/sunrpc/rpc_pipe_fs.h
> index 51b977a..cea764c 100644
> --- a/include/linux/sunrpc/rpc_pipe_fs.h
> +++ b/include/linux/sunrpc/rpc_pipe_fs.h
> @@ -15,6 +15,7 @@ struct rpc_pipe_ops {
>  	ssize_t (*upcall)(struct file *, struct rpc_pipe_msg *, char __user *, size_t);
>  	ssize_t (*downcall)(struct file *, const char __user *, size_t);
>  	void (*release_pipe)(struct inode *);
> +	int (*open_pipe)(struct inode *);
>  	void (*destroy_msg)(struct rpc_pipe_msg *);
>  };
>  
> diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
> index 23a2b8f..4171ab7 100644
> --- a/net/sunrpc/rpc_pipe.c
> +++ b/net/sunrpc/rpc_pipe.c
> @@ -169,16 +169,23 @@ static int
>  rpc_pipe_open(struct inode *inode, struct file *filp)
>  {
>  	struct rpc_inode *rpci = RPC_I(inode);
> +	int first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
>  	int res = -ENXIO;
>  
>  	mutex_lock(&inode->i_mutex);
> -	if (rpci->ops != NULL) {
> -		if (filp->f_mode & FMODE_READ)
> -			rpci->nreaders ++;
> -		if (filp->f_mode & FMODE_WRITE)
> -			rpci->nwriters ++;
> -		res = 0;
> +	if (rpci->ops == NULL)
> +		goto out;
> +	if (first_open && rpci->ops->open_pipe) {
> +		res = rpci->ops->open_pipe(inode);
> +		if (res)
> +			goto out;
>  	}
> +	if (filp->f_mode & FMODE_READ)
> +		rpci->nreaders++;
> +	if (filp->f_mode & FMODE_WRITE)
> +		rpci->nwriters++;
> +	res = 0;
> +out:

BTW: This is racy. You have to set first_open _after_ you take the
inode->i_mutex.

>  	mutex_unlock(&inode->i_mutex);
>  	return res;
>  }
> @@ -748,7 +755,7 @@ rpc_rmdir(struct dentry *dentry)
>   * @name: name of pipe
>   * @private: private data to associate with the pipe, for the caller's use
>   * @ops: operations defining the behavior of the pipe: upcall, downcall,
> - *	release_pipe, and destroy_msg.
> + *	release_pipe, open_pipe, and destroy_msg.
>   * @flags: rpc_inode flags
>   *
>   * Data is made available for userspace to read by calls to
-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 4/9] rpc: add an rpc_pipe_open method

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 03:35:22PM -0500, Trond Myklebust wrote:
> On Sun, 2008-11-09 at 16:04 -0500, J. Bruce Fields wrote:
> > We want to transition to a new gssd upcall which is text-based and more
> > easily extensible.
> > 
> > To simplify upgrades, as well as testing and debugging, it will help if
> > we can upgrade gssd (to a version which understands the new upcall)
> > without having to choose at boot (or module-load) time whether we want
> > the new or the old upcall.
> > 
> > We will do this by providing two different pipes: one named, as
> > currently, after the mechanism (normally "krb5"), and supporting the
> > old upcall.  One named "gssd" and supporting the new upcall version.
> > 
> > We allow gssd to indicate which version it supports by its choice of
> > which pipe to open.
> > 
> > As we have no interest in supporting *simultaneous* use of both
> > versions, we'll forbid opening both pipes at the same time.
> > 
> > So, add a new pipe_open callback to the rpc_pipefs api, which the gss
> > code can use to track which pipes have been open, and to refuse opens of
> > incompatible pipes.
> > 
> > We only need this to be called on the first open of a given pipe.
> > 
> > Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
> > ---
> >  include/linux/sunrpc/rpc_pipe_fs.h |    1 +
> >  net/sunrpc/rpc_pipe.c              |   21 ++++++++++++++-------
> >  2 files changed, 15 insertions(+), 7 deletions(-)
> > 
> > diff --git a/include/linux/sunrpc/rpc_pipe_fs.h b/include/linux/sunrpc/rpc_pipe_fs.h
> > index 51b977a..cea764c 100644
> > --- a/include/linux/sunrpc/rpc_pipe_fs.h
> > +++ b/include/linux/sunrpc/rpc_pipe_fs.h
> > @@ -15,6 +15,7 @@ struct rpc_pipe_ops {
> >  	ssize_t (*upcall)(struct file *, struct rpc_pipe_msg *, char __user *, size_t);
> >  	ssize_t (*downcall)(struct file *, const char __user *, size_t);
> >  	void (*release_pipe)(struct inode *);
> > +	int (*open_pipe)(struct inode *);
> >  	void (*destroy_msg)(struct rpc_pipe_msg *);
> >  };
> >  
> > diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
> > index 23a2b8f..4171ab7 100644
> > --- a/net/sunrpc/rpc_pipe.c
> > +++ b/net/sunrpc/rpc_pipe.c
> > @@ -169,16 +169,23 @@ static int
> >  rpc_pipe_open(struct inode *inode, struct file *filp)
> >  {
> >  	struct rpc_inode *rpci = RPC_I(inode);
> > +	int first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
> >  	int res = -ENXIO;
> >  
> >  	mutex_lock(&inode->i_mutex);
> > -	if (rpci->ops != NULL) {
> > -		if (filp->f_mode & FMODE_READ)
> > -			rpci->nreaders ++;
> > -		if (filp->f_mode & FMODE_WRITE)
> > -			rpci->nwriters ++;
> > -		res = 0;
> > +	if (rpci->ops == NULL)
> > +		goto out;
> > +	if (first_open && rpci->ops->open_pipe) {
> > +		res = rpci->ops->open_pipe(inode);
> > +		if (res)
> > +			goto out;
> >  	}
> > +	if (filp->f_mode & FMODE_READ)
> > +		rpci->nreaders++;
> > +	if (filp->f_mode & FMODE_WRITE)
> > +		rpci->nwriters++;
> > +	res = 0;
> > +out:
> 
> BTW: This is racy. You have to set first_open _after_ you take the
> inode->i_mutex.

Whoops!  Good catch, thanks--fixed.

--b.

> 
> >  	mutex_unlock(&inode->i_mutex);
> >  	return res;
> >  }
> > @@ -748,7 +755,7 @@ rpc_rmdir(struct dentry *dentry)
> >   * @name: name of pipe
> >   * @private: private data to associate with the pipe, for the caller's use
> >   * @ops: operations defining the behavior of the pipe: upcall, downcall,
> > - *	release_pipe, and destroy_msg.
> > + *	release_pipe, open_pipe, and destroy_msg.
> >   * @flags: rpc_inode flags
> >   *
> >   * Data is made available for userspace to read by calls to
> -- 
> Trond Myklebust
> Linux NFS client maintainer
> 
> NetApp
> Trond.Myklebust@netapp.com
> www.netapp.com

Re: [PATCH 4/9] rpc: add an rpc_pipe_open method

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 03:37:57PM -0500, bfields wrote:
> On Mon, Nov 10, 2008 at 03:35:22PM -0500, Trond Myklebust wrote:
> > BTW: This is racy. You have to set first_open _after_ you take the
> > inode->i_mutex.
> 
> Whoops!  Good catch, thanks--fixed.

(Updated version follows.)

--b.

commit 7f5c491a32857dc8ba587cc365a2d70024f7b33e
Author: J. Bruce Fields <bfields@citi.umich.edu>
Date:   Fri Nov 7 18:53:24 2008 -0500

    rpc: call release_pipe only on last close
    
    I can't see any reason we need to call this until either the kernel or
    the last gssd closes the pipe.
    
    Also, this allows to guarantee that open_pipe and release_pipe are
    called strictly in pairs; open_pipe on gssd's first open, release_pipe
    on gssd's last close (or on the close of the kernel side of the pipe, if
    that comes first).
    
    That will make it very easy for the gss code to keep track of which
    pipes gssd is using.
    
    Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>

diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index 4171ab7..f734103 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -126,13 +126,14 @@ rpc_close_pipes(struct inode *inode)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct rpc_pipe_ops *ops;
+	int still_open;
 
 	mutex_lock(&inode->i_mutex);
 	ops = rpci->ops;
 	if (ops != NULL) {
 		LIST_HEAD(free_list);
-
 		spin_lock(&inode->i_lock);
+		still_open = rpci->nreaders != 0 || rpci->nwriters != 0;
 		rpci->nreaders = 0;
 		list_splice_init(&rpci->in_upcall, &free_list);
 		list_splice_init(&rpci->pipe, &free_list);
@@ -141,7 +142,7 @@ rpc_close_pipes(struct inode *inode)
 		spin_unlock(&inode->i_lock);
 		rpc_purge_list(rpci, &free_list, ops->destroy_msg, -EPIPE);
 		rpci->nwriters = 0;
-		if (ops->release_pipe)
+		if (still_open && ops->release_pipe)
 			ops->release_pipe(inode);
 		cancel_delayed_work_sync(&rpci->queue_timeout);
 	}
@@ -169,12 +170,13 @@ static int
 rpc_pipe_open(struct inode *inode, struct file *filp)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
-	int first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
+	int first_open;
 	int res = -ENXIO;
 
 	mutex_lock(&inode->i_mutex);
 	if (rpci->ops == NULL)
 		goto out;
+	first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
 	if (first_open && rpci->ops->open_pipe) {
 		res = rpci->ops->open_pipe(inode);
 		if (res)
@@ -195,6 +197,7 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct rpc_pipe_msg *msg;
+	int last_close;
 
 	mutex_lock(&inode->i_mutex);
 	if (rpci->ops == NULL)
@@ -221,7 +224,8 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
 					rpci->ops->destroy_msg, -EAGAIN);
 		}
 	}
-	if (rpci->ops->release_pipe)
+	last_close = rpci->nwriters == 0 && rpci->nreaders == 0;
+	if (last_close && rpci->ops->release_pipe)
 		rpci->ops->release_pipe(inode);
 out:
 	mutex_unlock(&inode->i_mutex);

Re: [PATCH 4/9] rpc: add an rpc_pipe_open method

[Trond Myklebust]

On Mon, 2008-11-10 at 16:18 -0500, J. Bruce Fields wrote:
> On Mon, Nov 10, 2008 at 03:37:57PM -0500, bfields wrote:
> > On Mon, Nov 10, 2008 at 03:35:22PM -0500, Trond Myklebust wrote:
> > > BTW: This is racy. You have to set first_open _after_ you take the
> > > inode->i_mutex.
> > 
> > Whoops!  Good catch, thanks--fixed.
> 
> (Updated version follows.)

Err... This doesn't look right. This would be a mix of 4/9 and 5/9...

> --b.
> 
> commit 7f5c491a32857dc8ba587cc365a2d70024f7b33e
> Author: J. Bruce Fields <bfields@citi.umich.edu>
> Date:   Fri Nov 7 18:53:24 2008 -0500
> 
>     rpc: call release_pipe only on last close
>     
>     I can't see any reason we need to call this until either the kernel or
>     the last gssd closes the pipe.
>     
>     Also, this allows to guarantee that open_pipe and release_pipe are
>     called strictly in pairs; open_pipe on gssd's first open, release_pipe
>     on gssd's last close (or on the close of the kernel side of the pipe, if
>     that comes first).
>     
>     That will make it very easy for the gss code to keep track of which
>     pipes gssd is using.
>     
>     Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
> 
> diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
> index 4171ab7..f734103 100644
> --- a/net/sunrpc/rpc_pipe.c
> +++ b/net/sunrpc/rpc_pipe.c
> @@ -126,13 +126,14 @@ rpc_close_pipes(struct inode *inode)
>  {
>  	struct rpc_inode *rpci = RPC_I(inode);
>  	struct rpc_pipe_ops *ops;
> +	int still_open;
>  
>  	mutex_lock(&inode->i_mutex);
>  	ops = rpci->ops;
>  	if (ops != NULL) {
>  		LIST_HEAD(free_list);
> -
>  		spin_lock(&inode->i_lock);
> +		still_open = rpci->nreaders != 0 || rpci->nwriters != 0;
>  		rpci->nreaders = 0;
>  		list_splice_init(&rpci->in_upcall, &free_list);
>  		list_splice_init(&rpci->pipe, &free_list);
> @@ -141,7 +142,7 @@ rpc_close_pipes(struct inode *inode)
>  		spin_unlock(&inode->i_lock);
>  		rpc_purge_list(rpci, &free_list, ops->destroy_msg, -EPIPE);
>  		rpci->nwriters = 0;
> -		if (ops->release_pipe)
> +		if (still_open && ops->release_pipe)
>  			ops->release_pipe(inode);
>  		cancel_delayed_work_sync(&rpci->queue_timeout);
>  	}
> @@ -169,12 +170,13 @@ static int
>  rpc_pipe_open(struct inode *inode, struct file *filp)
>  {
>  	struct rpc_inode *rpci = RPC_I(inode);
> -	int first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
> +	int first_open;
>  	int res = -ENXIO;
>  
>  	mutex_lock(&inode->i_mutex);
>  	if (rpci->ops == NULL)
>  		goto out;
> +	first_open = rpci->nreaders == 0 && rpci->nwriters == 0;
>  	if (first_open && rpci->ops->open_pipe) {
>  		res = rpci->ops->open_pipe(inode);
>  		if (res)
> @@ -195,6 +197,7 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
>  {
>  	struct rpc_inode *rpci = RPC_I(inode);
>  	struct rpc_pipe_msg *msg;
> +	int last_close;
>  
>  	mutex_lock(&inode->i_mutex);
>  	if (rpci->ops == NULL)
> @@ -221,7 +224,8 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
>  					rpci->ops->destroy_msg, -EAGAIN);
>  		}
>  	}
> -	if (rpci->ops->release_pipe)
> +	last_close = rpci->nwriters == 0 && rpci->nreaders == 0;
> +	if (last_close && rpci->ops->release_pipe)
>  		rpci->ops->release_pipe(inode);
>  out:
>  	mutex_unlock(&inode->i_mutex);
-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH 4/9] rpc: add an rpc_pipe_open method

[J. Bruce Fields]

On Mon, Nov 10, 2008 at 04:48:42PM -0500, Trond Myklebust wrote:
> On Mon, 2008-11-10 at 16:18 -0500, J. Bruce Fields wrote:
> > On Mon, Nov 10, 2008 at 03:37:57PM -0500, bfields wrote:
> > > On Mon, Nov 10, 2008 at 03:35:22PM -0500, Trond Myklebust wrote:
> > > > BTW: This is racy. You have to set first_open _after_ you take the
> > > > inode->i_mutex.
> > > 
> > > Whoops!  Good catch, thanks--fixed.
> > 
> > (Updated version follows.)
> 
> Err... This doesn't look right. This would be a mix of 4/9 and 5/9...

Erp, right.  I folded the fix for the race into the wrong patch.  New
5/9 looks like the below.  (I'm also keeping the whole series in the
for-trond branch at

	git://linux-nfs.org/~bfields/linux.git for-trond

But let me know if you want any patches resent at some point.)

--b.

commit 4e0278688f280ff815505c8cd5287e4e9d553990
Author: J. Bruce Fields <bfields@citi.umich.edu>
Date:   Fri Nov 7 18:53:24 2008 -0500

    rpc: call release_pipe only on last close
    
    I can't see any reason we need to call this until either the kernel or
    the last gssd closes the pipe.
    
    Also, this allows to guarantee that open_pipe and release_pipe are
    called strictly in pairs; open_pipe on gssd's first open, release_pipe
    on gssd's last close (or on the close of the kernel side of the pipe, if
    that comes first).
    
    That will make it very easy for the gss code to keep track of which
    pipes gssd is using.
    
    Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>

diff --git a/net/sunrpc/rpc_pipe.c b/net/sunrpc/rpc_pipe.c
index 57342aa..a68aaac 100644
--- a/net/sunrpc/rpc_pipe.c
+++ b/net/sunrpc/rpc_pipe.c
@@ -126,13 +126,14 @@ rpc_close_pipes(struct inode *inode)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct rpc_pipe_ops *ops;
+	int need_release;
 
 	mutex_lock(&inode->i_mutex);
 	ops = rpci->ops;
 	if (ops != NULL) {
 		LIST_HEAD(free_list);
-
 		spin_lock(&inode->i_lock);
+		need_release = rpci->nreaders != 0 || rpci->nwriters != 0;
 		rpci->nreaders = 0;
 		list_splice_init(&rpci->in_upcall, &free_list);
 		list_splice_init(&rpci->pipe, &free_list);
@@ -141,7 +142,7 @@ rpc_close_pipes(struct inode *inode)
 		spin_unlock(&inode->i_lock);
 		rpc_purge_list(rpci, &free_list, ops->destroy_msg, -EPIPE);
 		rpci->nwriters = 0;
-		if (ops->release_pipe)
+		if (need_release && ops->release_pipe)
 			ops->release_pipe(inode);
 		cancel_delayed_work_sync(&rpci->queue_timeout);
 	}
@@ -196,6 +197,7 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
 {
 	struct rpc_inode *rpci = RPC_I(inode);
 	struct rpc_pipe_msg *msg;
+	int last_close;
 
 	mutex_lock(&inode->i_mutex);
 	if (rpci->ops == NULL)
@@ -222,7 +224,8 @@ rpc_pipe_release(struct inode *inode, struct file *filp)
 					rpci->ops->destroy_msg, -EAGAIN);
 		}
 	}
-	if (rpci->ops->release_pipe)
+	last_close = rpci->nwriters == 0 && rpci->nreaders == 0;
+	if (last_close && rpci->ops->release_pipe)
 		rpci->ops->release_pipe(inode);
 out:
 	mutex_unlock(&inode->i_mutex);

Re: [RFC] new client gssd upcall

[Jeff Layton]

On Fri, 13 Jun 2008 18:50:37 -0400
"J. Bruce Fields" <bfields@citi.umich.edu> wrote:

> The client needs a new more estensible text-based upcall to gssd, to
> make it easier to support some features that are needed for new krb5
> enctypes, for callbacks, etc.
> 
> We will have to continue to support the old upcall for backwards
> compatibility with older gssd's.  To simplify upgrades, as well as
> testing and debugging, it would help if we can upgrade gssd without
> having to choose at boot (or module-load) time whether we want the new
> or the old upcall.
> 
> That means that when gssd opens an rpc upcall pipe, we'd like to just
> start using whichever version it chose immediately--ideally without
> having to wait the 30 seconds that it would normally take for upcalls
> already queued on the wrong upcall pipe to time out.
> 
> The following patches do that by queueing the upcalls on whichever of
> the two upcall pipes (the new one and the old one) that somebody holds
> open.  If nobody's opened anything yet, then upcalls are provisionally
> queued on the new pipe.  I add a pipe_open method to allow the gss code
> to keep track of which pipe is open, to prevent anyone from attempting
> to open both pipes at once, and to cancel any upcalls that got queued to
> the wrong pipe.
> 
> I did some minimal testing with the old upcall, but haven't tested the
> new upcall yet.  (Olga, could you do that?)  So this is just an RFC at
> this point.
> 
> --b.

Has any thought been given to moving all of the rpc_pipefs upcalls to use
the keyctl API that David Howells did? It seems like that would be better
suited to this sort of application than rpc_pipefs...

-- 
Jeff Layton <jlayton@redhat.com>

Re: [RFC] new client gssd upcall

[J. Bruce Fields]

On Mon, Jun 16, 2008 at 10:28:59AM -0400, Jeff Layton wrote:
> On Fri, 13 Jun 2008 18:50:37 -0400
> "J. Bruce Fields" <bfields@citi.umich.edu> wrote:
> 
> > The client needs a new more estensible text-based upcall to gssd, to
> > make it easier to support some features that are needed for new krb5
> > enctypes, for callbacks, etc.
> > 
> > We will have to continue to support the old upcall for backwards
> > compatibility with older gssd's.  To simplify upgrades, as well as
> > testing and debugging, it would help if we can upgrade gssd without
> > having to choose at boot (or module-load) time whether we want the new
> > or the old upcall.
> > 
> > That means that when gssd opens an rpc upcall pipe, we'd like to just
> > start using whichever version it chose immediately--ideally without
> > having to wait the 30 seconds that it would normally take for upcalls
> > already queued on the wrong upcall pipe to time out.
> > 
> > The following patches do that by queueing the upcalls on whichever of
> > the two upcall pipes (the new one and the old one) that somebody holds
> > open.  If nobody's opened anything yet, then upcalls are provisionally
> > queued on the new pipe.  I add a pipe_open method to allow the gss code
> > to keep track of which pipe is open, to prevent anyone from attempting
> > to open both pipes at once, and to cancel any upcalls that got queued to
> > the wrong pipe.
> > 
> > I did some minimal testing with the old upcall, but haven't tested the
> > new upcall yet.  (Olga, could you do that?)  So this is just an RFC at
> > this point.
> > 
> > --b.
> 
> Has any thought been given to moving all of the rpc_pipefs upcalls to use
> the keyctl API that David Howells did? It seems like that would be better
> suited to this sort of application than rpc_pipefs...

I haven't looked at it.  I've just assumed that since Trond and Kevin
have both looked at both API's, then there must be some good reason
we're not using it....

--b.

Re: [RFC] new client gssd upcall

[Trond Myklebust]

On Tue, 2008-06-17 at 17:36 -0400, J. Bruce Fields wrote:
> On Mon, Jun 16, 2008 at 10:28:59AM -0400, Jeff Layton wrote:
> > On Fri, 13 Jun 2008 18:50:37 -0400
> > "J. Bruce Fields" <bfields@citi.umich.edu> wrote:
> > 
> > > The client needs a new more estensible text-based upcall to gssd, to
> > > make it easier to support some features that are needed for new krb5
> > > enctypes, for callbacks, etc.
> > > 
> > > We will have to continue to support the old upcall for backwards
> > > compatibility with older gssd's.  To simplify upgrades, as well as
> > > testing and debugging, it would help if we can upgrade gssd without
> > > having to choose at boot (or module-load) time whether we want the new
> > > or the old upcall.
> > > 
> > > That means that when gssd opens an rpc upcall pipe, we'd like to just
> > > start using whichever version it chose immediately--ideally without
> > > having to wait the 30 seconds that it would normally take for upcalls
> > > already queued on the wrong upcall pipe to time out.
> > > 
> > > The following patches do that by queueing the upcalls on whichever of
> > > the two upcall pipes (the new one and the old one) that somebody holds
> > > open.  If nobody's opened anything yet, then upcalls are provisionally
> > > queued on the new pipe.  I add a pipe_open method to allow the gss code
> > > to keep track of which pipe is open, to prevent anyone from attempting
> > > to open both pipes at once, and to cancel any upcalls that got queued to
> > > the wrong pipe.
> > > 
> > > I did some minimal testing with the old upcall, but haven't tested the
> > > new upcall yet.  (Olga, could you do that?)  So this is just an RFC at
> > > this point.
> > > 
> > > --b.
> > 
> > Has any thought been given to moving all of the rpc_pipefs upcalls to use
> > the keyctl API that David Howells did? It seems like that would be better
> > suited to this sort of application than rpc_pipefs...
> 
> I haven't looked at it.  I've just assumed that since Trond and Kevin
> have both looked at both API's, then there must be some good reason
> we're not using it....

Kevin has spent quite some time working on the keyring support, but as
far as I understand the amount of time he can continue to spend working
for CITI has recently been heavily reduced...

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [RFC] new client gssd upcall

[J. Bruce Fields]

On Tue, Jun 17, 2008 at 05:59:29PM -0400, Trond Myklebust wrote:
> On Tue, 2008-06-17 at 17:36 -0400, J. Bruce Fields wrote:
> > On Mon, Jun 16, 2008 at 10:28:59AM -0400, Jeff Layton wrote:
> > > Has any thought been given to moving all of the rpc_pipefs upcalls to use
> > > the keyctl API that David Howells did? It seems like that would be better
> > > suited to this sort of application than rpc_pipefs...
> > 
> > I haven't looked at it.  I've just assumed that since Trond and Kevin
> > have both looked at both API's, then there must be some good reason
> > we're not using it....
> 
> Kevin has spent quite some time working on the keyring support, but as
> far as I understand the amount of time he can continue to spend working
> for CITI has recently been heavily reduced...

Hm, but I thought that you'd both decided that the gssd upcalls would
end up coexisting with the upcalls in any case?

If there's a chance we might end up replacing the gssd upcalls entirely,
then--while I don't want to stretch out this one task
indefinitely--still it might be worth my time to go take a look at the
keyctl API's Jeff mentions.

--b.

Re: [RFC] new client gssd upcall

[Jeff Layton]

On Tue, 17 Jun 2008 18:09:18 -0400
"J. Bruce Fields" <bfields@fieldses.org> wrote:

> On Tue, Jun 17, 2008 at 05:59:29PM -0400, Trond Myklebust wrote:
> > On Tue, 2008-06-17 at 17:36 -0400, J. Bruce Fields wrote:
> > > On Mon, Jun 16, 2008 at 10:28:59AM -0400, Jeff Layton wrote:
> > > > Has any thought been given to moving all of the rpc_pipefs upcalls to use
> > > > the keyctl API that David Howells did? It seems like that would be better
> > > > suited to this sort of application than rpc_pipefs...
> > > 
> > > I haven't looked at it.  I've just assumed that since Trond and Kevin
> > > have both looked at both API's, then there must be some good reason
> > > we're not using it....
> > 
> > Kevin has spent quite some time working on the keyring support, but as
> > far as I understand the amount of time he can continue to spend working
> > for CITI has recently been heavily reduced...
> 
> Hm, but I thought that you'd both decided that the gssd upcalls would
> end up coexisting with the upcalls in any case?
> 
> If there's a chance we might end up replacing the gssd upcalls entirely,
> then--while I don't want to stretch out this one task
> indefinitely--still it might be worth my time to go take a look at the
> keyctl API's Jeff mentions.
> 

If you're interested, there are several examples of upcalls that use
this API in CIFS. I used it for the SPNEGO/Kerberos stuff that I did
there several months ago. It's pretty easy to use and can handle a
fairly large payload if needed.

When it does the upcall, it runs a particular program, so there's no
need for a running daemon. That might be considered a good thing here
since it may be one less thing that a user has to remember to have
running...

-- 
Jeff Layton <jlayton@redhat.com>

Re: [RFC] new client gssd upcall

[Olga Kornievskaia]

Trond Myklebust wrote:
> On Tue, 2008-06-17 at 17:36 -0400, J. Bruce Fields wrote:
>   
>> On Mon, Jun 16, 2008 at 10:28:59AM -0400, Jeff Layton wrote:
>>     
>>> On Fri, 13 Jun 2008 18:50:37 -0400
>>> "J. Bruce Fields" <bfields@citi.umich.edu> wrote:
>>>
>>>       
>>>> The client needs a new more estensible text-based upcall to gssd, to
>>>> make it easier to support some features that are needed for new krb5
>>>> enctypes, for callbacks, etc.
>>>>
>>>> We will have to continue to support the old upcall for backwards
>>>> compatibility with older gssd's.  To simplify upgrades, as well as
>>>> testing and debugging, it would help if we can upgrade gssd without
>>>> having to choose at boot (or module-load) time whether we want the new
>>>> or the old upcall.
>>>>
>>>> That means that when gssd opens an rpc upcall pipe, we'd like to just
>>>> start using whichever version it chose immediately--ideally without
>>>> having to wait the 30 seconds that it would normally take for upcalls
>>>> already queued on the wrong upcall pipe to time out.
>>>>
>>>> The following patches do that by queueing the upcalls on whichever of
>>>> the two upcall pipes (the new one and the old one) that somebody holds
>>>> open.  If nobody's opened anything yet, then upcalls are provisionally
>>>> queued on the new pipe.  I add a pipe_open method to allow the gss code
>>>> to keep track of which pipe is open, to prevent anyone from attempting
>>>> to open both pipes at once, and to cancel any upcalls that got queued to
>>>> the wrong pipe.
>>>>
>>>> I did some minimal testing with the old upcall, but haven't tested the
>>>> new upcall yet.  (Olga, could you do that?)  So this is just an RFC at
>>>> this point.
>>>>
>>>> --b.
>>>>         
>>> Has any thought been given to moving all of the rpc_pipefs upcalls to use
>>> the keyctl API that David Howells did? It seems like that would be better
>>> suited to this sort of application than rpc_pipefs...
>>>       
>> I haven't looked at it.  I've just assumed that since Trond and Kevin
>> have both looked at both API's, then there must be some good reason
>> we're not using it....
>>     
>
> Kevin has spent quite some time working on the keyring support, but as
> far as I understand the amount of time he can continue to spend working
> for CITI has recently been heavily reduced...
>
>   
I don't think we ever considered replacing NFS's client upcall 
mechanism. Why would we scrap a perfectly good implementation that Trond 
keeps on improving and making better?

Re: [RFC] new client gssd upcall

[Jeff Layton]

On Thu, 19 Jun 2008 11:37:17 -0400
Olga Kornievskaia <aglo@citi.umich.edu> wrote:

> 
> 
> Trond Myklebust wrote:
> > On Tue, 2008-06-17 at 17:36 -0400, J. Bruce Fields wrote:
> >   
> >> On Mon, Jun 16, 2008 at 10:28:59AM -0400, Jeff Layton wrote:
> >>     
> >>> On Fri, 13 Jun 2008 18:50:37 -0400
> >>> "J. Bruce Fields" <bfields@citi.umich.edu> wrote:
> >>>
> >>>       
> >>>> The client needs a new more estensible text-based upcall to gssd, to
> >>>> make it easier to support some features that are needed for new krb5
> >>>> enctypes, for callbacks, etc.
> >>>>
> >>>> We will have to continue to support the old upcall for backwards
> >>>> compatibility with older gssd's.  To simplify upgrades, as well as
> >>>> testing and debugging, it would help if we can upgrade gssd without
> >>>> having to choose at boot (or module-load) time whether we want the new
> >>>> or the old upcall.
> >>>>
> >>>> That means that when gssd opens an rpc upcall pipe, we'd like to just
> >>>> start using whichever version it chose immediately--ideally without
> >>>> having to wait the 30 seconds that it would normally take for upcalls
> >>>> already queued on the wrong upcall pipe to time out.
> >>>>
> >>>> The following patches do that by queueing the upcalls on whichever of
> >>>> the two upcall pipes (the new one and the old one) that somebody holds
> >>>> open.  If nobody's opened anything yet, then upcalls are provisionally
> >>>> queued on the new pipe.  I add a pipe_open method to allow the gss code
> >>>> to keep track of which pipe is open, to prevent anyone from attempting
> >>>> to open both pipes at once, and to cancel any upcalls that got queued to
> >>>> the wrong pipe.
> >>>>
> >>>> I did some minimal testing with the old upcall, but haven't tested the
> >>>> new upcall yet.  (Olga, could you do that?)  So this is just an RFC at
> >>>> this point.
> >>>>
> >>>> --b.
> >>>>         
> >>> Has any thought been given to moving all of the rpc_pipefs upcalls to use
> >>> the keyctl API that David Howells did? It seems like that would be better
> >>> suited to this sort of application than rpc_pipefs...
> >>>       
> >> I haven't looked at it.  I've just assumed that since Trond and Kevin
> >> have both looked at both API's, then there must be some good reason
> >> we're not using it....
> >>     
> >
> > Kevin has spent quite some time working on the keyring support, but as
> > far as I understand the amount of time he can continue to spend working
> > for CITI has recently been heavily reduced...
> >
> >   
> I don't think we ever considered replacing NFS's client upcall 
> mechanism. Why would we scrap a perfectly good implementation that Trond 
> keeps on improving and making better?
> 

Because it's less code that we have to maintain. rpc_pipefs certainly works
(and works fairly well), but we have so many upcall mechanisms in the
kernel already. keyctl also has some nice features (automated cache
timeouts, granular security, etc), and was designed with this sort of use
in mind.

I'm not saying that we absolutely need to scrap rpc_pipefs, but considering
alternatives may mean less work for us all in the long run.

-- 
Jeff Layton <jlayton@redhat.com>

Re: [RFC] new client gssd upcall

[Trond Myklebust]

On Thu, 2008-06-19 at 11:49 -0400, Jeff Layton wrote:
> Because it's less code that we have to maintain. rpc_pipefs certainly works
> (and works fairly well), but we have so many upcall mechanisms in the
> kernel already. keyctl also has some nice features (automated cache
> timeouts, granular security, etc), and was designed with this sort of use
> in mind.
> 
> I'm not saying that we absolutely need to scrap rpc_pipefs, but considering
> alternatives may mean less work for us all in the long run.

Talk about scrapping rpc_pipefs is premature, to say the least: I have
yet to see a plan for replacing the idmapd upcall. I have yet to see
working code that replaces the gss upcall and that works correctly in a
non-blocking environment.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [RFC] new client gssd upcall

[Jeff Layton]

On Thu, 19 Jun 2008 13:06:22 -0400
Trond Myklebust <Trond.Myklebust@netapp.com> wrote:

> On Thu, 2008-06-19 at 11:49 -0400, Jeff Layton wrote:
> > Because it's less code that we have to maintain. rpc_pipefs certainly works
> > (and works fairly well), but we have so many upcall mechanisms in the
> > kernel already. keyctl also has some nice features (automated cache
> > timeouts, granular security, etc), and was designed with this sort of use
> > in mind.
> > 
> > I'm not saying that we absolutely need to scrap rpc_pipefs, but considering
> > alternatives may mean less work for us all in the long run.
> 
> Talk about scrapping rpc_pipefs is premature, to say the least: I have
> yet to see a plan for replacing the idmapd upcall. I have yet to see
> working code that replaces the gss upcall and that works correctly in a
> non-blocking environment.
> 

That was probably too strongly worded. We wouldn't be scrapping
rpc_pipefs for a long time even if we did add new upcall schemes. That
said:

I know that David H. recently added async versions of request_key:

request_key_async()

...and...

request_key_async_with_auxdata()

...assuming they work the way I think they do, they should be OK for
the non-blocking case. There's also no reason we couldn't use keys for
idmap upcalls as well. I'm considering them for a similar idmap scheme
for CIFS.

Of course all of this is handwavy and speculative. I have no working
code. I was just asking whether anyone had considered this API for
this purpose.

-- 
Jeff Layton <jlayton@redhat.com>

Re: [RFC] new client gssd upcall

[Trond Myklebust]

On Thu, 2008-06-19 at 13:27 -0400, Jeff Layton wrote:
> There's also no reason we couldn't use keys for
> idmap upcalls as well. I'm considering them for a similar idmap scheme
> for CIFS.

Ewww.... NACK, NACK, NACK, NACK....

There is a perfectly good reason why you wouldn't ever want to use keys
for idmap upcalls: keys are user/process/thread objects while idmapd
entries are NFSv4-namespace objects.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [RFC] new client gssd upcall

[Jeff Layton]

On Thu, 19 Jun 2008 14:13:19 -0400
Trond Myklebust <Trond.Myklebust@netapp.com> wrote:

> On Thu, 2008-06-19 at 13:27 -0400, Jeff Layton wrote:
> > There's also no reason we couldn't use keys for
> > idmap upcalls as well. I'm considering them for a similar idmap scheme
> > for CIFS.
> 
> Ewww.... NACK, NACK, NACK, NACK....
> 
> There is a perfectly good reason why you wouldn't ever want to use keys
> for idmap upcalls: keys are user/process/thread objects while idmapd
> entries are NFSv4-namespace objects.
> 

My thinking for CIFS is to use keys to do the upcall and copy the
mapping into a cache that we'll manage independently of the key cache.
The CIFS case is a little different though. We're not mapping usernames
to uid's and vice-versa, but Windows RID's to unix uid's. Still, it's a
somewhat similar problem. The amount of data that we're dealing with in
an idmap upcall is pretty small, so copying it and then destroying the
key wouldn't involve a lot of overhead.

That may not be palatable for NFS though.

-- 
Jeff Layton <jlayton@redhat.com>

Re: [RFC] new client gssd upcall

[peter honeyman]

On Tue, 2008-06-17 at 21:59 -0400, Trond Myklebust wrote:
> Kevin has spent quite some time working on the keyring support, but  
> as far as I understand the amount of time he can continue to spend  
> working for CITI has recently been heavily reduced...

kevin is helping out on a med school project this summer, but there  
has been no change to his appointment: he continues to work full-time  
for CITI.

of course, the amount of time available to work on unfunded efforts is  
limited, but that's nothing new.

	peter