nfs4: set source address when callback is generated

when callback is generated in NFSv4 server, it doesn't set the source address. When an alias IP is utilized on NFSv4 server and suppose the client is accessing via that alias IP (e.g. eth0:0), the client invokes the callback to the IP address that is set on the original device (e.g. eth0). This behavior results in timeout of xprt.
The patch sets the IP address that the client should invoke callback to.

Signed-off-by Takuma Umeya <[email protected]>

fs/nfsd/nfs4callback.c +1 -0

index 143da2e..bb5de7f 100644
@@ -483,6 +483,7 @@ int setup_callback_client(struct nfs4_client *clp, struct nfs4_cb_conn *conn)
                .net            = &init_net,
                .address        = (struct sockaddr *) &conn->cb_addr,
                .addrsize       = conn->cb_addrlen,
+               .saddress       = (struct sockaddr *) &conn->cb_saddr,
                .timeout        = &timeparms,
                .program        = &cb_program,
                .version        = 0,

fs/nfsd/nfs4state.c +6 -1

index 116cab9..c7d3171 100644
@@ -1177,9 +1177,11 @@ find_unconfirmed_client_by_str(const char *dname, unsigned int hashval,
 }
 
 static void
-gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
+gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, 
+                                       u32 scopeid, struct svc_rqst *rqstp)
 {
        struct nfs4_cb_conn *conn = &clp->cl_cb_conn;
+       struct sockaddr s_sa;
        unsigned short expected_family;
 
        /* Currently, we only support tcp and tcp6 for the callback channel */

@@ -1205,6 +1207,19 @@ gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
 
        conn->cb_prog = se->se_callback_prog;
        conn->cb_ident = se->se_callback_ident;
+       switch(expected_family){
+               case AF_INET:
+                       ((struct sockaddr_in *)&s_sa)->sin_family = AF_INET;

fs/nfsd/nfs4state.c +1 -1

        return;
 out_err:
        conn->cb_addr.ss_family = AF_UNSPEC;

@@ -1876,7 +1891,7 @@ nfsd4_setclientid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
         * for consistent minorversion use throughout:
         */
        new->cl_minorversion = 0;
-       gen_callback(new, setclid, rpc_get_scope_id(sa));
+       gen_callback(new, setclid, rpc_get_scope_id(sa), rqstp);
        add_to_unconfirmed(new, strhashval);
        setclid->se_clientid.cl_boot = new->cl_clientid.cl_boot;
        setclid->se_clientid.cl_id = new->cl_clientid.cl_id;

fs/nfsd/state.h +1 -1

index 39adc27..0cee03d 100644
@@ -95,6 +95,7 @@ struct nfs4_delegation {
 struct nfs4_cb_conn {
        /* SETCLIENTID info */
        struct sockaddr_storage cb_addr;
+       struct sockaddr_storage cb_saddr;
        size_t                  cb_addrlen;
        u32                     cb_prog; /* used only in 4.0 case;
                                            per-session otherwise */
 
--

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Quoted 51 lines · Takuma Umeya

when callback is generated in NFSv4 server, it doesn't set the source address. When an alias IP is utilized on NFSv4 server and suppose the client is accessing via that alias IP (e.g. eth0:0), the client invokes the callback to the IP address that is set on the original device (e.g. eth0). This behavior results in timeout of xprt.
The patch sets the IP address that the client should invoke callback to.

Signed-off-by Takuma Umeya <[email protected]>

fs/nfsd/nfs4callback.c +1 -0

index 143da2e..bb5de7f 100644
@@ -483,6 +483,7 @@ int setup_callback_client(struct nfs4_client *clp, struct nfs4_cb_conn *conn)
               .net            = &init_net,
               .address        = (struct sockaddr *) &conn->cb_addr,
               .addrsize       = conn->cb_addrlen,
+               .saddress       = (struct sockaddr *) &conn->cb_saddr,
               .timeout        = &timeparms,
               .program        = &cb_program,
               .version        = 0,

fs/nfsd/nfs4state.c +0 -0

index 116cab9..c7d3171 100644
@@ -1177,9 +1177,11 @@ find_unconfirmed_client_by_str(const char *dname, unsigned int hashval,
 }
 

static void

fs/nfsd/nfs4state.c +2 -1

-gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
+gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, 
+                                       u32 scopeid, struct svc_rqst *rqstp)

{
       struct nfs4_cb_conn *conn = &clp->cl_cb_conn;

fs/nfsd/nfs4state.c +9 -0

+       struct sockaddr s_sa;
       unsigned short expected_family;
 
       /* Currently, we only support tcp and tcp6 for the callback channel */

@@ -1205,6 +1207,19 @@ gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
 
       conn->cb_prog = se->se_callback_prog;
       conn->cb_ident = se->se_callback_ident;
+       switch(expected_family){
+               case AF_INET:
+                       ((struct sockaddr_in *)&s_sa)->sin_family = AF_INET;
+                       ((struct sockaddr_in *)&s_sa)->sin_addr = rqstp->
+                                                               rq_daddr.addr;
+                        break;
+               case AF_INET6:
+                       ((struct sockaddr_in6 *)&s_sa)->sin6_family = AF_INET6;

fs/nfsd/nfs4state.c gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid) · line 1222

Quoted 5 lines

fs/nfsd/nfs4state.c +5 -0

+                       ((struct sockaddr_in6 *)&s_sa)->sin6_addr = rqstp->
+                                                               rq_daddr.addr6;
+                       break;
+       }
+       rpc_copy_addr((struct sockaddr *)&conn->cb_saddr, &s_sa);

Chuck, is there any reason we use this special svc_addr_u union and not a struct sockaddr_storage? This seems more verbose than necessary....

--b.

Quoted 30 lines

fs/nfsd/nfs4state.c +0 -0

       return;

out_err:

       conn->cb_addr.ss_family = AF_UNSPEC;

fs/nfsd/nfs4state.c +1 -1

@@ -1876,7 +1891,7 @@ nfsd4_setclientid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
        * for consistent minorversion use throughout:
        */
       new->cl_minorversion = 0;
-       gen_callback(new, setclid, rpc_get_scope_id(sa));
+       gen_callback(new, setclid, rpc_get_scope_id(sa), rqstp);
       add_to_unconfirmed(new, strhashval);
       setclid->se_clientid.cl_boot = new->cl_clientid.cl_boot;
       setclid->se_clientid.cl_id = new->cl_clientid.cl_id;

fs/nfsd/state.h +1 -1

index 39adc27..0cee03d 100644
@@ -95,6 +95,7 @@ struct nfs4_delegation {
 struct nfs4_cb_conn {
       /* SETCLIENTID info */
       struct sockaddr_storage cb_addr;
+       struct sockaddr_storage cb_saddr;
       size_t                  cb_addrlen;
       u32                     cb_prog; /* used only in 4.0 case;
                                           per-session otherwise */
 
--

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

I don't recall creating svc_addr_u, but I'll take a stab at a guess.

It looks like someone thought that we should retain the idea of storing just the address part of the socket address, and not the other stuff (like the family and port, since this code doesn't appear to need that additional information). It greatly reduces the size of the field. A full sockaddr_storage is more than 128 bytes, since it has to be able to store an AF_UNIX pathname.

Doing this, there is a lot less data to keep around, but an IPv6 socket address has other items outside of in6_addr that can be used to form a full address. We decided at some point we could copy this information from the other address storage field in the rqstp.

But the result of this space savings means we must construct a full socket address when needed, using logic such as the above.

Quoted 7 lines · Chuck Lever

I don't recall creating svc_addr_u, but I'll take a stab at a guess.

It looks like someone thought that we should retain the idea of storing just the address part of the socket address, and not the other stuff (like the family and port, since this code doesn't appear to need that additional information). It greatly reduces the size of the field. A full sockaddr_storage is more than 128 bytes, since it has to be able to store an AF_UNIX pathname.

Doing this, there is a lot less data to keep around, but an IPv6 socket address has other items outside of in6_addr that can be used to form a full address. We decided at some point we could copy this information from the other address storage field in the rqstp.

But the result of this space savings means we must construct a full socket address when needed, using logic such as the above.

Seems to me we should either just waste the extra 100 bytes or define something that would be useful elsewhere as well....

But if we do it this way we can at least simplify a little.

--b.

commit 6f3d772fb8a039de8f21d725f5e38c252b4c0efd

Author: Takuma Umeya <[email protected]>
Date:   Wed Dec 15 14:09:01 2010 +0900

nfs4: set source address when callback is generated

when callback is generated in NFSv4 server, it doesn't set the source address. When an alias IP is utilized on NFSv4 server and suppose the client is accessing via that alias IP (e.g. eth0:0), the client invokes the callback to the IP address that is set on the original device (e.g. eth0). This behavior results in timeout of xprt.
The patch sets the IP address that the client should invoke callback to.

Signed-off-by Takuma Umeya <[email protected]>

[[email protected]: Simplify gen_callback arguments, use helper function]

Signed-off-by J. Bruce Fields <[email protected]>

fs/nfsd/nfs4callback.c +1 -0

index a085805..dd183af 100644
@@ -484,6 +484,7 @@ static int setup_callback_client(struct nfs4_client *clp,
 		.net		= &init_net,
 		.address	= (struct sockaddr *) &conn->cb_addr,
 		.addrsize	= conn->cb_addrlen,
+		.saddress	= (struct sockaddr *) &conn->cb_saddr,
 		.timeout	= &timeparms,
 		.program	= &cb_program,
 		.version	= 0,

fs/nfsd/nfs4state.c +14 -0

index 87d4c48..b583e4e 100644
@@ -1163,10 +1163,26 @@ find_unconfirmed_client_by_str(const char *dname, unsigned int hashval)
 	return NULL;
 }
 
+static void rpc_svcaddr2sockaddr(struct sockaddr *sa, unsigned short family, union svc_addr_u *svcaddr)
+{
+	switch (family) {
+	case AF_INET:
+		((struct sockaddr_in *)sa)->sin_family = AF_INET;
+		((struct sockaddr_in *)sa)->sin_addr = svcaddr->addr;
+		return;
+	case AF_INET6:
+		((struct sockaddr_in6 *)sa)->sin6_family = AF_INET6;
+		((struct sockaddr_in6 *)sa)->sin6_addr = svcaddr->addr6;
+		return;
+	}
+}
+

static void

fs/nfsd/nfs4state.c +1 -1

-gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
+gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, struct svc_rqst *rqstp)

{
	struct nfs4_cb_conn *conn = &clp->cl_cb_conn;

fs/nfsd/nfs4state.c +2 -0

+	struct sockaddr	*sa = svc_addr(rqstp);
+	u32 scopeid = rpc_get_scope_id(sa);

	unsigned short expected_family;

	/* Currently, we only support tcp and tcp6 for the callback channel */

fs/nfsd/nfs4state.c +1 -0

@@ -1192,6 +1208,7 @@ gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
 
 	conn->cb_prog = se->se_callback_prog;
 	conn->cb_ident = se->se_callback_ident;
+	rpc_svcaddr2sockaddr((struct sockaddr *)&conn->cb_saddr, expected_family, &rqstp->rq_daddr);
 	return;

out_err:

	conn->cb_addr.ss_family = AF_UNSPEC;

fs/nfsd/nfs4state.c +1 -2

@@ -1768,7 +1785,6 @@ __be32
 nfsd4_setclientid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 		  struct nfsd4_setclientid *setclid)
 {
-	struct sockaddr		*sa = svc_addr(rqstp);
 	struct xdr_netobj 	clname = { 
 		.len = setclid->se_namelen,
 		.data = setclid->se_name,

@@ -1871,7 +1887,7 @@ nfsd4_setclientid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 	 * for consistent minorversion use throughout:
 	 */
 	new->cl_minorversion = 0;
-	gen_callback(new, setclid, rpc_get_scope_id(sa));
+	gen_callback(new, setclid, rqstp);
 	add_to_unconfirmed(new, strhashval);
 	setclid->se_clientid.cl_boot = new->cl_clientid.cl_boot;
 	setclid->se_clientid.cl_id = new->cl_clientid.cl_id;

fs/nfsd/state.h +1 -0

index 84b2302..cf6dc83 100644
@@ -96,6 +96,7 @@ struct nfs4_delegation {
 struct nfs4_cb_conn {
 	/* SETCLIENTID info */
 	struct sockaddr_storage	cb_addr;
+	struct sockaddr_storage	cb_saddr;
 	size_t			cb_addrlen;
 	u32                     cb_prog; /* used only in 4.0 case;
 					    per-session otherwise */

Quoted 11 lines · J. Bruce Fields

Quoted 7 lines · Chuck Lever

I don't recall creating svc_addr_u, but I'll take a stab at a guess.

It looks like someone thought that we should retain the idea of storing just the address part of the socket address, and not the other stuff (like the family and port, since this code doesn't appear to need that additional information). It greatly reduces the size of the field. A full sockaddr_storage is more than 128 bytes, since it has to be able to store an AF_UNIX pathname.

Doing this, there is a lot less data to keep around, but an IPv6 socket address has other items outside of in6_addr that can be used to form a full address. We decided at some point we could copy this information from the other address storage field in the rqstp.

But the result of this space savings means we must construct a full socket address when needed, using logic such as the above.

Seems to me we should either just waste the extra 100 bytes or define something that would be useful elsewhere as well....

In nfs-utils, we define:

	union nfs_sockaddr {
		struct sockaddr_in	s4;
		struct sockaddr_in6	s6;
		struct sockaddr		sa;
	};

A variable of this type is large enough to hold a full IPv6 sockaddr, but is significantly smaller than a sockaddr_storage.

The addition of the "struct sockaddr" element is to enable access to such variables via a "struct sockaddr *" without type punning. This seems to be preferred by gcc over type casting in order to handle optimizations involving address aliasing. It also allows more precise type checking.

A full conversion to use such a construct in kernel RPC and NFS components is, I fear, too late for 2.6.38, but might be considered for a future release if there is consensus on this approach.

Quoted 103 lines

But if we do it this way we can at least simplify a little.

--b.

commit 6f3d772fb8a039de8f21d725f5e38c252b4c0efd

Author: Takuma Umeya <[email protected]>
Date:   Wed Dec 15 14:09:01 2010 +0900

nfs4: set source address when callback is generated

when callback is generated in NFSv4 server, it doesn't set the source address. When an alias IP is utilized on NFSv4 server and suppose the client is accessing via that alias IP (e.g. eth0:0), the client invokes the callback to the IP address that is set on the original device (e.g. eth0). This behavior results in timeout of xprt.
The patch sets the IP address that the client should invoke callback to.

Signed-off-by Takuma Umeya <[email protected]>

[[email protected]: Simplify gen_callback arguments, use helper function]

Signed-off-by J. Bruce Fields <[email protected]>

fs/nfsd/nfs4callback.c +1 -0

index a085805..dd183af 100644
@@ -484,6 +484,7 @@ static int setup_callback_client(struct nfs4_client *clp,
 		.net		= &init_net,
 		.address	= (struct sockaddr *) &conn->cb_addr,
 		.addrsize	= conn->cb_addrlen,
+		.saddress	= (struct sockaddr *) &conn->cb_saddr,
 		.timeout	= &timeparms,
 		.program	= &cb_program,
 		.version	= 0,

fs/nfsd/nfs4state.c +14 -0

index 87d4c48..b583e4e 100644
@@ -1163,10 +1163,26 @@ find_unconfirmed_client_by_str(const char *dname, unsigned int hashval)
 	return NULL;
 }
 
+static void rpc_svcaddr2sockaddr(struct sockaddr *sa, unsigned short family, union svc_addr_u *svcaddr)
+{
+	switch (family) {
+	case AF_INET:
+		((struct sockaddr_in *)sa)->sin_family = AF_INET;
+		((struct sockaddr_in *)sa)->sin_addr = svcaddr->addr;
+		return;
+	case AF_INET6:
+		((struct sockaddr_in6 *)sa)->sin6_family = AF_INET6;
+		((struct sockaddr_in6 *)sa)->sin6_addr = svcaddr->addr6;
+		return;
+	}
+}
+

static void

fs/nfsd/nfs4state.c +1 -1

-gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
+gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, struct svc_rqst *rqstp)

{
	struct nfs4_cb_conn *conn = &clp->cl_cb_conn;

fs/nfsd/nfs4state.c +2 -0

+	struct sockaddr	*sa = svc_addr(rqstp);
+	u32 scopeid = rpc_get_scope_id(sa);

	unsigned short expected_family;

	/* Currently, we only support tcp and tcp6 for the callback channel */

fs/nfsd/nfs4state.c +1 -0

@@ -1192,6 +1208,7 @@ gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
 
 	conn->cb_prog = se->se_callback_prog;
 	conn->cb_ident = se->se_callback_ident;
+	rpc_svcaddr2sockaddr((struct sockaddr *)&conn->cb_saddr, expected_family, &rqstp->rq_daddr);
 	return;

out_err:

	conn->cb_addr.ss_family = AF_UNSPEC;

fs/nfsd/nfs4state.c +1 -2

@@ -1768,7 +1785,6 @@ __be32
 nfsd4_setclientid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 		  struct nfsd4_setclientid *setclid)
 {
-	struct sockaddr		*sa = svc_addr(rqstp);
 	struct xdr_netobj 	clname = { 
 		.len = setclid->se_namelen,
 		.data = setclid->se_name,

@@ -1871,7 +1887,7 @@ nfsd4_setclientid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 	 * for consistent minorversion use throughout:
 	 */
 	new->cl_minorversion = 0;
-	gen_callback(new, setclid, rpc_get_scope_id(sa));
+	gen_callback(new, setclid, rqstp);
 	add_to_unconfirmed(new, strhashval);
 	setclid->se_clientid.cl_boot = new->cl_clientid.cl_boot;
 	setclid->se_clientid.cl_id = new->cl_clientid.cl_id;

fs/nfsd/state.h +1 -0

index 84b2302..cf6dc83 100644
@@ -96,6 +96,7 @@ struct nfs4_delegation {
 struct nfs4_cb_conn {
 	/* SETCLIENTID info */
 	struct sockaddr_storage	cb_addr;
+	struct sockaddr_storage	cb_saddr;
 	size_t			cb_addrlen;
 	u32                     cb_prog; /* used only in 4.0 case;
 					    per-session otherwise */

Signature · 2 lines

Chuck Lever
chuck[dot]lever[at]oracle[dot]com

Quoted 23 lines · Chuck Lever

Quoted 11 lines · J. Bruce Fields

Quoted 7 lines · Chuck Lever

I don't recall creating svc_addr_u, but I'll take a stab at a guess.

It looks like someone thought that we should retain the idea of storing just the address part of the socket address, and not the other stuff (like the family and port, since this code doesn't appear to need that additional information). It greatly reduces the size of the field. A full sockaddr_storage is more than 128 bytes, since it has to be able to store an AF_UNIX pathname.

Doing this, there is a lot less data to keep around, but an IPv6 socket address has other items outside of in6_addr that can be used to form a full address. We decided at some point we could copy this information from the other address storage field in the rqstp.

But the result of this space savings means we must construct a full socket address when needed, using logic such as the above.

Seems to me we should either just waste the extra 100 bytes or define something that would be useful elsewhere as well....

In nfs-utils, we define:

	union nfs_sockaddr {
		struct sockaddr_in	s4;
		struct sockaddr_in6	s6;
		struct sockaddr		sa;
	};

Quoted 3 lines

A variable of this type is large enough to hold a full IPv6 sockaddr, but is significantly smaller than a sockaddr_storage.

The addition of the "struct sockaddr" element is to enable access to such variables via a "struct sockaddr *" without type punning. This seems to be preferred by gcc over type casting in order to handle optimizations involving address aliasing. It also allows more precise type checking.

Sounds reasonable to me.

Quoted 2 lines

A full conversion to use such a construct in kernel RPC and NFS components is, I fear, too late for 2.6.38, but might be considered for a future release if there is consensus on this approach.

OK; I suppose for now I'll apply my revision of Takuma Umeya's patch below (if I didn't screw it up).

--b.

Quoted 111 lines

Quoted 103 lines

But if we do it this way we can at least simplify a little.

--b.

commit 6f3d772fb8a039de8f21d725f5e38c252b4c0efd

Author: Takuma Umeya <[email protected]>
Date:   Wed Dec 15 14:09:01 2010 +0900

nfs4: set source address when callback is generated

when callback is generated in NFSv4 server, it doesn't set the source address. When an alias IP is utilized on NFSv4 server and suppose the client is accessing via that alias IP (e.g. eth0:0), the client invokes the callback to the IP address that is set on the original device (e.g. eth0). This behavior results in timeout of xprt.
The patch sets the IP address that the client should invoke callback to.

Signed-off-by Takuma Umeya <[email protected]>

[[email protected]: Simplify gen_callback arguments, use helper function]

Signed-off-by J. Bruce Fields <[email protected]>

fs/nfsd/nfs4callback.c +1 -0

index a085805..dd183af 100644
@@ -484,6 +484,7 @@ static int setup_callback_client(struct nfs4_client *clp,
 		.net		= &init_net,
 		.address	= (struct sockaddr *) &conn->cb_addr,
 		.addrsize	= conn->cb_addrlen,
+		.saddress	= (struct sockaddr *) &conn->cb_saddr,
 		.timeout	= &timeparms,
 		.program	= &cb_program,
 		.version	= 0,

fs/nfsd/nfs4state.c +14 -0

index 87d4c48..b583e4e 100644
@@ -1163,10 +1163,26 @@ find_unconfirmed_client_by_str(const char *dname, unsigned int hashval)
 	return NULL;
 }
 
+static void rpc_svcaddr2sockaddr(struct sockaddr *sa, unsigned short family, union svc_addr_u *svcaddr)
+{
+	switch (family) {
+	case AF_INET:
+		((struct sockaddr_in *)sa)->sin_family = AF_INET;
+		((struct sockaddr_in *)sa)->sin_addr = svcaddr->addr;
+		return;
+	case AF_INET6:
+		((struct sockaddr_in6 *)sa)->sin6_family = AF_INET6;
+		((struct sockaddr_in6 *)sa)->sin6_addr = svcaddr->addr6;
+		return;
+	}
+}
+

static void

fs/nfsd/nfs4state.c +1 -1

-gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
+gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, struct svc_rqst *rqstp)

{
	struct nfs4_cb_conn *conn = &clp->cl_cb_conn;

fs/nfsd/nfs4state.c +2 -0

+	struct sockaddr	*sa = svc_addr(rqstp);
+	u32 scopeid = rpc_get_scope_id(sa);

	unsigned short expected_family;

	/* Currently, we only support tcp and tcp6 for the callback channel */

fs/nfsd/nfs4state.c +1 -0

@@ -1192,6 +1208,7 @@ gen_callback(struct nfs4_client *clp, struct nfsd4_setclientid *se, u32 scopeid)
 
 	conn->cb_prog = se->se_callback_prog;
 	conn->cb_ident = se->se_callback_ident;
+	rpc_svcaddr2sockaddr((struct sockaddr *)&conn->cb_saddr, expected_family, &rqstp->rq_daddr);
 	return;

out_err:

	conn->cb_addr.ss_family = AF_UNSPEC;

fs/nfsd/nfs4state.c +1 -2

@@ -1768,7 +1785,6 @@ __be32
 nfsd4_setclientid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 		  struct nfsd4_setclientid *setclid)
 {
-	struct sockaddr		*sa = svc_addr(rqstp);
 	struct xdr_netobj 	clname = { 
 		.len = setclid->se_namelen,
 		.data = setclid->se_name,

@@ -1871,7 +1887,7 @@ nfsd4_setclientid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 	 * for consistent minorversion use throughout:
 	 */
 	new->cl_minorversion = 0;
-	gen_callback(new, setclid, rpc_get_scope_id(sa));
+	gen_callback(new, setclid, rqstp);
 	add_to_unconfirmed(new, strhashval);
 	setclid->se_clientid.cl_boot = new->cl_clientid.cl_boot;
 	setclid->se_clientid.cl_id = new->cl_clientid.cl_id;

fs/nfsd/state.h +1 -0

index 84b2302..cf6dc83 100644
@@ -96,6 +96,7 @@ struct nfs4_delegation {
 struct nfs4_cb_conn {
 	/* SETCLIENTID info */
 	struct sockaddr_storage	cb_addr;
+	struct sockaddr_storage	cb_saddr;
 	size_t			cb_addrlen;
 	u32                     cb_prog; /* used only in 4.0 case;
 					    per-session otherwise */

Signature · 2 lines

Chuck Lever
chuck[dot]lever[at]oracle[dot]com

Quoted 48 lines · J. Bruce Fields

Quoted 31 lines · Chuck Lever

Quoted 17 lines · J. Bruce Fields

Quoted 16 lines · Chuck Lever

I don't recall creating svc_addr_u, but I'll take a stab at a guess.

It looks like someone thought that we should retain the idea of storing just the address part of the socket address, and not the other stuff (like the family and port, since this code doesn't appear to need that additional information). It greatly reduces the size of the field. A full sockaddr_storage is more than 128 bytes, since it has to be able to store an AF_UNIX pathname.

Doing this, there is a lot less data to keep around, but an IPv6 socket address has other items outside of in6_addr that can be used to form a full address. We decided at some point we could copy this information from the other address storage field in the rqstp.

Quoted 5 lines

Quoted 2 lines

But the result of this space savings means we must construct a full socket address when needed, using logic such as the above.

Seems to me we should either just waste the extra 100 bytes or define something that would be useful elsewhere as well....

In nfs-utils, we define:

	union nfs_sockaddr { struct sockaddr_in	s4; struct sockaddr_in6
	s6; struct sockaddr		sa; };

Quoted 8 lines

A variable of this type is large enough to hold a full IPv6 sockaddr, but is significantly smaller than a sockaddr_storage.

The addition of the "struct sockaddr" element is to enable access to such variables via a "struct sockaddr *" without type punning. This seems to be preferred by gcc over type casting in order to handle optimizations involving address aliasing. It also allows more precise type checking.

Sounds reasonable to me.

Quoted 4 lines

A full conversion to use such a construct in kernel RPC and NFS components is, I fear, too late for 2.6.38, but might be considered for a future release if there is consensus on this approach.

Quoted 2 lines

OK; I suppose for now I'll apply my revision of Takuma Umeya's patch below (if I didn't screw it up).

My thinking cap is still on vacation. I don't see anything immediately wrong with this as a temporary fix. Has anyone tested this with a multi-homed IPv6 server? How about link-local IPv6 addresses?