[PATCH 0/3] Detect user space support for rpcbind v4

[PATCH 0/3] Detect user space support for rpcbind v4

[Chuck Lever]

Hi Bruce-

Here's what I'm considering.

If the kernel detects that user space doesn't support rpcbind protocol
version 4, it falls back to version 2.  If it is trying to register an
IPv6 RPC service, it fails outright.

All of this can be disabled completely in favor of legacy behavior by
disabling CONFIG_SUNRPC_REGISTER_V4.

In addition, I will add CONFIG options to enable/disable lockd and NFSD
listening on IPv6 sockets, to prevent all this noise for anyone using
classic IPv4-only.  These will appear only if CONFIG_SUNRPC_REGISTER_V4
is enabled.

I haven't yet pursued why the kernel doesn't report the program version
mismatch to higher layers.  That will be needed to make these changes
work correctly.

These patches have been compile-tested only.

---

Chuck Lever (3):
      SUNRPC: Clearer error message when user space is running portmap
      SUNRPC: Clearer error message when user space is running portmap
      SUNRPC: Don't use EPROTONOSUPPORT in svc_register()


 net/sunrpc/svc.c |   99 ++++++++++++++++++++++++++++++++++++++++++++++--------
 1 files changed, 84 insertions(+), 15 deletions(-)

-- 
Chuck Lever <chuck.lever@oracle.com>

[PATCH 1/3] SUNRPC: Don't use EPROTONOSUPPORT in svc_register()

[Chuck Lever]

The RPC client returns EPROTONOSUPPORT if there is a protocol version
mismatch (ie the server doesn't support the RPC protocol version sent
by the client).

Since svc_register uses the RPC client, it's confusing when
svc_register's helpers return this errno for a problem not related to
a program number mismatch.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
---

 net/sunrpc/svc.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
index c51fed4..a278a82 100644
--- a/net/sunrpc/svc.c
+++ b/net/sunrpc/svc.c
@@ -750,7 +750,7 @@ static int __svc_rpcb_register4(const u32 program, const u32 version,
 		netid = RPCBIND_NETID_TCP;
 		break;
 	default:
-		return -EPROTONOSUPPORT;
+		return -ENOPROTOOPT;
 	}
 
 	return rpcb_v4_register(program, version,
@@ -786,7 +786,7 @@ static int __svc_rpcb_register6(const u32 program, const u32 version,
 		netid = RPCBIND_NETID_TCP6;
 		break;
 	default:
-		return -EPROTONOSUPPORT;
+		return -ENOPROTOOPT;
 	}
 
 	return rpcb_v4_register(program, version,

[PATCH 2/3] SUNRPC: Clearer error message when user space is running portmap

[Chuck Lever]

Detect the case where we sent an rpcbind v4 UNSET but user space
doesn't support rpcbind protocol version 4.

We down-shift to version 2 and print a warning once in this case.
Olaf says a v2 UNSET should properly clear all entries in the rpcbind
database.

As svc_unregister() is usually done first to clear old portmapper
entries before trying to set up new ones, a v4 UNSET call will almost
always be the first to detect an rpcbind protocol version mismatch
problem.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
---

 net/sunrpc/svc.c |   51 ++++++++++++++++++++++++++++++++++++++++++++-------
 1 files changed, 44 insertions(+), 7 deletions(-)

diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
index a278a82..e93bc55 100644
--- a/net/sunrpc/svc.c
+++ b/net/sunrpc/svc.c
@@ -721,6 +721,21 @@ EXPORT_SYMBOL_GPL(svc_exit_thread);
 
 #ifdef CONFIG_SUNRPC_REGISTER_V4
 
+static u32 __svc_register_version = 4;
+
+static void __svc_rpcb_prog_mismatch_error(void)
+{
+	if (__svc_register_version == 4) {
+		printk(KERN_WARNING
+			"svc: The kernel tried an rpcbind version 4 request, "
+			"but your portmapper\n"
+			"svc: does not support rpcbind version 4.  "
+			"Switching to rpcbind version 2.\n");
+
+		__svc_register_version = 2;
+	}
+}
+
 /*
  * Register an "inet" protocol family netid with the local
  * rpcbind daemon via an rpcbind v4 SET request.
@@ -896,10 +911,20 @@ int svc_register(const struct svc_serv *serv, const unsigned short proto,
 	return error;
 }
 
+static void __svc_v2_unregister(const u32 program, const u32 version,
+				const char *progname)
+{
+	int error;
+
+	error = rpcb_register(program, version, 0, 0);
+	dprintk("svc: %s(%sv%u), error %d\n",
+			__func__, progname, version, error);
+}
+
 #ifdef CONFIG_SUNRPC_REGISTER_V4
 
-static void __svc_unregister(const u32 program, const u32 version,
-			     const char *progname)
+static int __svc_v4_unregister(const u32 program, const u32 version,
+			       const char *progname)
 {
 	struct sockaddr_in6 sin6 = {
 		.sin6_family		= AF_INET6,
@@ -912,6 +937,22 @@ static void __svc_unregister(const u32 program, const u32 version,
 				(struct sockaddr *)&sin6, "");
 	dprintk("svc: %s(%sv%u), error %d\n",
 			__func__, progname, version, error);
+	return error;
+}
+
+static void __svc_unregister(const u32 program, const u32 version,
+			     const char *progname)
+{
+	if (__svc_register_version == 4) {
+		int error = __svc_v4_unregister(program, version, progname);
+		if (error != -EPROTONOSUPPORT)
+			return;
+	}
+
+	__svc_rpcb_prog_mismatch_error();
+
+	/* Olaf says a v2 UNSET _should_ wipe everything */
+	__svc_v2_unregister(program, version, progname);
 }
 
 #else	/* CONFIG_SUNRPC_REGISTER_V4 */
@@ -919,11 +960,7 @@ static void __svc_unregister(const u32 program, const u32 version,
 static void __svc_unregister(const u32 program, const u32 version,
 			     const char *progname)
 {
-	int error;
-
-	error = rpcb_register(program, version, 0, 0);
-	dprintk("svc: %s(%sv%u), error %d\n",
-			__func__, progname, version, error);
+	__svc_v2_unregister(program, version, progname);
 }
 
 #endif	/* CONFIG_SUNRPC_REGISTER_V4 */

[PATCH 3/3] SUNRPC: Clearer error message when user space is running portmap

[Chuck Lever]

Detect the case where we sent an rpcbind v4 SET but user space doesn't
support rpcbind protocol version 4.

When registering an AF_INET service, a warning is sent to the log once,
and we down-shift to version 2.  With AF_INET6, an error is always sent
to the log, and the registration fails.

This should help distributors and testers determine why their NFSD
service won't start, but normal users/admins should really never see
this message unless something very strange has happened.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
---

 net/sunrpc/svc.c |   44 ++++++++++++++++++++++++++++++++++++++------
 1 files changed, 38 insertions(+), 6 deletions(-)

diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
index e93bc55..daed77f 100644
--- a/net/sunrpc/svc.c
+++ b/net/sunrpc/svc.c
@@ -736,6 +736,18 @@ static void __svc_rpcb_prog_mismatch_error(void)
 	}
 }
 
+static int __svc_rpcb_cant_register(const char *progname, const u32 version)
+{
+	printk(KERN_ERR
+		"svc: Could not register an IPv6 listener for %s v%u "
+		"with the local portmapper.\n"
+		"svc: The kernel tried an rpcbind version 4 request, "
+		"but your portmapper\n"
+		"svc: does not support rpcbind version 4.",
+		progname, version);
+	return -EAFNOSUPPORT;
+}
+
 /*
  * Register an "inet" protocol family netid with the local
  * rpcbind daemon via an rpcbind v4 SET request.
@@ -811,10 +823,14 @@ static int __svc_rpcb_register6(const u32 program, const u32 version,
 /*
  * Register a kernel RPC service via rpcbind version 4.
  *
+ * This logic is complicated by the case where user space is not
+ * running a portmapper that can handle a verison 4 request.
+ *
  * Returns zero on success; a negative errno value is returned
  * if any error occurs.
  */
-static int __svc_register(const u32 program, const u32 version,
+static int __svc_register(const char *progname,
+			  const u32 program, const u32 version,
 			  const sa_family_t family,
 			  const unsigned short protocol,
 			  const unsigned short port)
@@ -823,11 +839,25 @@ static int __svc_register(const u32 program, const u32 version,
 
 	switch (family) {
 	case AF_INET:
-		return __svc_rpcb_register4(program, version,
-						protocol, port);
+		if (__svc_register_version == 4) {
+			error = __svc_rpcb_register4(program, version,
+							protocol, port);
+			if (error != -EPROTONOSUPPORT)
+				return error;
+		}
+
+		/* Down-shift. */
+		__svc_rpcb_prog_mismatch_error();
+		return rpcb_register(program, version, protocol, port);
 	case AF_INET6:
+		/* This won't work if we don't have version 4. */
+		if (__svc_register_version != 4)
+			return __svc_rpcb_cant_register(progname, version);
+
 		error = __svc_rpcb_register6(program, version,
 						protocol, port);
+		if (error == -EPROTONOSUPPORT)
+			return __svc_rpcb_cant_register(progname, version);
 		if (error < 0)
 			return error;
 
@@ -854,7 +884,8 @@ static int __svc_register(const u32 program, const u32 version,
  * Returns zero on success; a negative errno value is returned
  * if any error occurs.
  */
-static int __svc_register(const u32 program, const u32 version,
+static int __svc_register(const char *progname,
+			  const u32 program, const u32 version,
 			  sa_family_t family,
 			  const unsigned short protocol,
 			  const unsigned short port)
@@ -901,8 +932,9 @@ int svc_register(const struct svc_serv *serv, const unsigned short proto,
 			if (progp->pg_vers[i]->vs_hidden)
 				continue;
 
-			error = __svc_register(progp->pg_prog, i,
-						serv->sv_family, proto, port);
+			error = __svc_register(progp->pg_name, progp->pg_prog,
+						i, serv->sv_family,
+						proto, port);
 			if (error < 0)
 				break;
 		}

Re: [PATCH 3/3] SUNRPC: Clearer error message when user space is running portmap

[Trond Myklebust]

On Tue, 2009-01-13 at 12:31 -0500, Chuck Lever wrote:
> Detect the case where we sent an rpcbind v4 SET but user space doesn't
> support rpcbind protocol version 4.
> 
> When registering an AF_INET service, a warning is sent to the log once,
> and we down-shift to version 2.  With AF_INET6, an error is always sent
> to the log, and the registration fails.
> 
> This should help distributors and testers determine why their NFSD
> service won't start, but normal users/admins should really never see
> this message unless something very strange has happened.
> 
> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> ---
> 
>  net/sunrpc/svc.c |   44 ++++++++++++++++++++++++++++++++++++++------
>  1 files changed, 38 insertions(+), 6 deletions(-)
> 
> diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
> index e93bc55..daed77f 100644
> --- a/net/sunrpc/svc.c
> +++ b/net/sunrpc/svc.c
> @@ -736,6 +736,18 @@ static void __svc_rpcb_prog_mismatch_error(void)
>  	}
>  }
>  
> +static int __svc_rpcb_cant_register(const char *progname, const u32 version)
> +{
> +	printk(KERN_ERR
> +		"svc: Could not register an IPv6 listener for %s v%u "
> +		"with the local portmapper.\n"
> +		"svc: The kernel tried an rpcbind version 4 request, "
> +		"but your portmapper\n"
> +		"svc: does not support rpcbind version 4.",
> +		progname, version);
> +	return -EAFNOSUPPORT;
> +}

This is ___way___ too intrusive...

We already have stuff like the NFSv4 callback channel that automatically
assumes that it should use an IPv6 registration as long as CONFIG_IPV6
or CONFIG_IPV6_MODULE are defined.

Compiling in IPv6 without having an IPv6-capable portmapper should
certainly not be flagged as a kernel error.

>  /*
>   * Register an "inet" protocol family netid with the local
>   * rpcbind daemon via an rpcbind v4 SET request.
> @@ -811,10 +823,14 @@ static int __svc_rpcb_register6(const u32 program, const u32 version,
>  /*
>   * Register a kernel RPC service via rpcbind version 4.
>   *
> + * This logic is complicated by the case where user space is not
> + * running a portmapper that can handle a verison 4 request.
> + *
>   * Returns zero on success; a negative errno value is returned
>   * if any error occurs.
>   */
> -static int __svc_register(const u32 program, const u32 version,
> +static int __svc_register(const char *progname,
> +			  const u32 program, const u32 version,
>  			  const sa_family_t family,
>  			  const unsigned short protocol,
>  			  const unsigned short port)
> @@ -823,11 +839,25 @@ static int __svc_register(const u32 program, const u32 version,
>  
>  	switch (family) {
>  	case AF_INET:
> -		return __svc_rpcb_register4(program, version,
> -						protocol, port);
> +		if (__svc_register_version == 4) {
> +			error = __svc_rpcb_register4(program, version,
> +							protocol, port);
> +			if (error != -EPROTONOSUPPORT)
> +				return error;
> +		}
> +
> +		/* Down-shift. */
> +		__svc_rpcb_prog_mismatch_error();
> +		return rpcb_register(program, version, protocol, port);
>  	case AF_INET6:
> +		/* This won't work if we don't have version 4. */
> +		if (__svc_register_version != 4)
> +			return __svc_rpcb_cant_register(progname, version);
> +
>  		error = __svc_rpcb_register6(program, version,
>  						protocol, port);
> +		if (error == -EPROTONOSUPPORT)
> +			return __svc_rpcb_cant_register(progname, version);
>  		if (error < 0)
>  			return error;
>  
> @@ -854,7 +884,8 @@ static int __svc_register(const u32 program, const u32 version,
>   * Returns zero on success; a negative errno value is returned
>   * if any error occurs.
>   */
> -static int __svc_register(const u32 program, const u32 version,
> +static int __svc_register(const char *progname,
> +			  const u32 program, const u32 version,
>  			  sa_family_t family,
>  			  const unsigned short protocol,
>  			  const unsigned short port)
> @@ -901,8 +932,9 @@ int svc_register(const struct svc_serv *serv, const unsigned short proto,
>  			if (progp->pg_vers[i]->vs_hidden)
>  				continue;
>  
> -			error = __svc_register(progp->pg_prog, i,
> -						serv->sv_family, proto, port);
> +			error = __svc_register(progp->pg_name, progp->pg_prog,
> +						i, serv->sv_family,
> +						proto, port);
>  			if (error < 0)
>  				break;
>  		}
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 3/3] SUNRPC: Clearer error message when user space is running portmap

[Chuck Lever]

On Jan 13, 2009, at Jan 13, 2009, 12:44 PM, Trond Myklebust wrote:
> On Tue, 2009-01-13 at 12:31 -0500, Chuck Lever wrote:
>> Detect the case where we sent an rpcbind v4 SET but user space  
>> doesn't
>> support rpcbind protocol version 4.
>>
>> When registering an AF_INET service, a warning is sent to the log  
>> once,
>> and we down-shift to version 2.  With AF_INET6, an error is always  
>> sent
>> to the log, and the registration fails.
>>
>> This should help distributors and testers determine why their NFSD
>> service won't start, but normal users/admins should really never see
>> this message unless something very strange has happened.
>>
>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>> ---
>>
>> net/sunrpc/svc.c |   44 ++++++++++++++++++++++++++++++++++++++------
>> 1 files changed, 38 insertions(+), 6 deletions(-)
>>
>> diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
>> index e93bc55..daed77f 100644
>> --- a/net/sunrpc/svc.c
>> +++ b/net/sunrpc/svc.c
>> @@ -736,6 +736,18 @@ static void __svc_rpcb_prog_mismatch_error(void)
>> 	}
>> }
>>
>> +static int __svc_rpcb_cant_register(const char *progname, const  
>> u32 version)
>> +{
>> +	printk(KERN_ERR
>> +		"svc: Could not register an IPv6 listener for %s v%u "
>> +		"with the local portmapper.\n"
>> +		"svc: The kernel tried an rpcbind version 4 request, "
>> +		"but your portmapper\n"
>> +		"svc: does not support rpcbind version 4.",
>> +		progname, version);
>> +	return -EAFNOSUPPORT;
>> +}
>
> This is ___way___ too intrusive...

> We already have stuff like the NFSv4 callback channel that  
> automatically
> assumes that it should use an IPv6 registration as long as CONFIG_IPV6
> or CONFIG_IPV6_MODULE are defined.

I believe the callback channel service does not register itself with  
the local portmapper.  This function wouldn't be invoked in that case.

> Compiling in IPv6 without having an IPv6-capable portmapper should
> certainly not be flagged as a kernel error.

It's an error if the kernel can't register an RPC service that asked  
to be registered, isn't it?

But the point I _think_ you're making is that enabling kernel IPv6  
support (outside of RPC) shouldn't cause problems like this.  If  
CONFIG_SUNRPC_REGISTER_V4 is not set, then IPv6 RPC listeners for  
lockd and NFSD aren't created at all, even if CONFIG_IPV6 is enabled,  
and this problem is avoided.

(I think proper dependencies exist to make this work right today, but  
it's pretty confusing.  An additional CONFIG entry for explicitly  
enabling lockd and NFSD via IPv6 would make this better understood).

The other problem here is that when registering an IPv6 RPC service,  
the kernel uses an IPv6 loopback address to test whether the user  
space portmapper is actually IPv6 enabled.  I expect this is why the  
new kernel running with the old user space never got "program version  
mismatch" -- the portmapper wasn't listening on IPv6 to reply to the  
request.

This was the reason I wanted to use a connected transport for  
registration upcalls -- then rpcb_v4_register could tell immediately  
that user space wasn't prepared for kernel IPv6 RPC services, instead  
of timing out after many seconds.

We can switch to always use IPv4 for the registration upcall to avoid  
this problem and retain backwards compatibility.  However, this means  
that the kernel can't detect whether user space is really IPv6  
enabled.  It can only tell whether rpcbindv4 is supported.  Do we  
think that is sufficient?  (I'll hazard a guess that the answer is  
"yes").

>> /*
>>  * Register an "inet" protocol family netid with the local
>>  * rpcbind daemon via an rpcbind v4 SET request.
>> @@ -811,10 +823,14 @@ static int __svc_rpcb_register6(const u32  
>> program, const u32 version,
>> /*
>>  * Register a kernel RPC service via rpcbind version 4.
>>  *
>> + * This logic is complicated by the case where user space is not
>> + * running a portmapper that can handle a verison 4 request.
>> + *
>>  * Returns zero on success; a negative errno value is returned
>>  * if any error occurs.
>>  */
>> -static int __svc_register(const u32 program, const u32 version,
>> +static int __svc_register(const char *progname,
>> +			  const u32 program, const u32 version,
>> 			  const sa_family_t family,
>> 			  const unsigned short protocol,
>> 			  const unsigned short port)
>> @@ -823,11 +839,25 @@ static int __svc_register(const u32 program,  
>> const u32 version,
>>
>> 	switch (family) {
>> 	case AF_INET:
>> -		return __svc_rpcb_register4(program, version,
>> -						protocol, port);
>> +		if (__svc_register_version == 4) {
>> +			error = __svc_rpcb_register4(program, version,
>> +							protocol, port);
>> +			if (error != -EPROTONOSUPPORT)
>> +				return error;
>> +		}
>> +
>> +		/* Down-shift. */
>> +		__svc_rpcb_prog_mismatch_error();
>> +		return rpcb_register(program, version, protocol, port);
>> 	case AF_INET6:
>> +		/* This won't work if we don't have version 4. */
>> +		if (__svc_register_version != 4)
>> +			return __svc_rpcb_cant_register(progname, version);
>> +
>> 		error = __svc_rpcb_register6(program, version,
>> 						protocol, port);
>> +		if (error == -EPROTONOSUPPORT)
>> +			return __svc_rpcb_cant_register(progname, version);
>> 		if (error < 0)
>> 			return error;
>>
>> @@ -854,7 +884,8 @@ static int __svc_register(const u32 program,  
>> const u32 version,
>>  * Returns zero on success; a negative errno value is returned
>>  * if any error occurs.
>>  */
>> -static int __svc_register(const u32 program, const u32 version,
>> +static int __svc_register(const char *progname,
>> +			  const u32 program, const u32 version,
>> 			  sa_family_t family,
>> 			  const unsigned short protocol,
>> 			  const unsigned short port)
>> @@ -901,8 +932,9 @@ int svc_register(const struct svc_serv *serv,  
>> const unsigned short proto,
>> 			if (progp->pg_vers[i]->vs_hidden)
>> 				continue;
>>
>> -			error = __svc_register(progp->pg_prog, i,
>> -						serv->sv_family, proto, port);
>> +			error = __svc_register(progp->pg_name, progp->pg_prog,
>> +						i, serv->sv_family,
>> +						proto, port);
>> 			if (error < 0)
>> 				break;
>> 		}
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux- 
>> nfs" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>

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

Re: [PATCH 3/3] SUNRPC: Clearer error message when user space is running portmap

[Chuck Lever]

On Jan 13, 2009, at Jan 13, 2009, 1:26 PM, Chuck Lever wrote:
> On Jan 13, 2009, at Jan 13, 2009, 12:44 PM, Trond Myklebust wrote:
>> On Tue, 2009-01-13 at 12:31 -0500, Chuck Lever wrote:
>>> Detect the case where we sent an rpcbind v4 SET but user space  
>>> doesn't
>>> support rpcbind protocol version 4.
>>>
>>> When registering an AF_INET service, a warning is sent to the log  
>>> once,
>>> and we down-shift to version 2.  With AF_INET6, an error is always  
>>> sent
>>> to the log, and the registration fails.
>>>
>>> This should help distributors and testers determine why their NFSD
>>> service won't start, but normal users/admins should really never see
>>> this message unless something very strange has happened.
>>>
>>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>>> ---
>>>
>>> net/sunrpc/svc.c |   44 ++++++++++++++++++++++++++++++++++++++------
>>> 1 files changed, 38 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
>>> index e93bc55..daed77f 100644
>>> --- a/net/sunrpc/svc.c
>>> +++ b/net/sunrpc/svc.c
>>> @@ -736,6 +736,18 @@ static void  
>>> __svc_rpcb_prog_mismatch_error(void)
>>> 	}
>>> }
>>>
>>> +static int __svc_rpcb_cant_register(const char *progname, const  
>>> u32 version)
>>> +{
>>> +	printk(KERN_ERR
>>> +		"svc: Could not register an IPv6 listener for %s v%u "
>>> +		"with the local portmapper.\n"
>>> +		"svc: The kernel tried an rpcbind version 4 request, "
>>> +		"but your portmapper\n"
>>> +		"svc: does not support rpcbind version 4.",
>>> +		progname, version);
>>> +	return -EAFNOSUPPORT;
>>> +}
>>
>> This is ___way___ too intrusive...
>
>> We already have stuff like the NFSv4 callback channel that  
>> automatically
>> assumes that it should use an IPv6 registration as long as  
>> CONFIG_IPV6
>> or CONFIG_IPV6_MODULE are defined.
>
> I believe the callback channel service does not register itself with  
> the local portmapper.  This function wouldn't be invoked in that case.
>
>> Compiling in IPv6 without having an IPv6-capable portmapper should
>> certainly not be flagged as a kernel error.
>
> It's an error if the kernel can't register an RPC service that asked  
> to be registered, isn't it?
>
> But the point I _think_ you're making is that enabling kernel IPv6  
> support (outside of RPC) shouldn't cause problems like this.  If  
> CONFIG_SUNRPC_REGISTER_V4 is not set, then IPv6 RPC listeners for  
> lockd and NFSD aren't created at all, even if CONFIG_IPV6 is  
> enabled, and this problem is avoided.
>
> (I think proper dependencies exist to make this work right today,  
> but it's pretty confusing.  An additional CONFIG entry for  
> explicitly enabling lockd and NFSD via IPv6 would make this better  
> understood).

The problem now becomes "what to do if user space doesn't support  
registering an IPv6 kernel RPC service?"

The kernel uses a single listener for each RPC service; either IPv6 or  
IPv4.  Since the svc_serv is created first, then after that the  
transport is created and registered, we would have to tear down the  
transport and svc_serv and then try to create these again with only  
AF_INET listeners.

Alternately, we could just have svc_register() fall back and register  
just IPv4, even though it's an AF_INET6 listener.  The code is  
designed so that AF_INET6 listeners handle all the IPv4 traffic too.

One benefit is that with this mechanism, replacing portmapper with  
rpcbind suddenly allows the kernel's RPC services to handle AF_INET6  
traffic (if they support it already).

Does this sound OK?

> The other problem here is that when registering an IPv6 RPC service,  
> the kernel uses an IPv6 loopback address to test whether the user  
> space portmapper is actually IPv6 enabled.  I expect this is why the  
> new kernel running with the old user space never got "program  
> version mismatch" -- the portmapper wasn't listening on IPv6 to  
> reply to the request.

Yes, now tested; this is indeed the case.  Switching to IPv4 loopback  
fixed the hang-and-timeout, and registering an AF_INET6 service fails  
immediately now on my Fedora Core 6 system.

> This was the reason I wanted to use a connected transport for  
> registration upcalls -- then rpcb_v4_register could tell immediately  
> that user space wasn't prepared for kernel IPv6 RPC services,  
> instead of timing out after many seconds.
>
> We can switch to always use IPv4 for the registration upcall to  
> avoid this problem and retain backwards compatibility.  However,  
> this means that the kernel can't detect whether user space is really  
> IPv6 enabled.  It can only tell whether rpcbindv4 is supported.  Do  
> we think that is sufficient?  (I'll hazard a guess that the answer  
> is "yes").
>
>>> /*
>>> * Register an "inet" protocol family netid with the local
>>> * rpcbind daemon via an rpcbind v4 SET request.
>>> @@ -811,10 +823,14 @@ static int __svc_rpcb_register6(const u32  
>>> program, const u32 version,
>>> /*
>>> * Register a kernel RPC service via rpcbind version 4.
>>> *
>>> + * This logic is complicated by the case where user space is not
>>> + * running a portmapper that can handle a verison 4 request.
>>> + *
>>> * Returns zero on success; a negative errno value is returned
>>> * if any error occurs.
>>> */
>>> -static int __svc_register(const u32 program, const u32 version,
>>> +static int __svc_register(const char *progname,
>>> +			  const u32 program, const u32 version,
>>> 			  const sa_family_t family,
>>> 			  const unsigned short protocol,
>>> 			  const unsigned short port)
>>> @@ -823,11 +839,25 @@ static int __svc_register(const u32 program,  
>>> const u32 version,
>>>
>>> 	switch (family) {
>>> 	case AF_INET:
>>> -		return __svc_rpcb_register4(program, version,
>>> -						protocol, port);
>>> +		if (__svc_register_version == 4) {
>>> +			error = __svc_rpcb_register4(program, version,
>>> +							protocol, port);
>>> +			if (error != -EPROTONOSUPPORT)
>>> +				return error;
>>> +		}
>>> +
>>> +		/* Down-shift. */
>>> +		__svc_rpcb_prog_mismatch_error();
>>> +		return rpcb_register(program, version, protocol, port);
>>> 	case AF_INET6:
>>> +		/* This won't work if we don't have version 4. */
>>> +		if (__svc_register_version != 4)
>>> +			return __svc_rpcb_cant_register(progname, version);
>>> +
>>> 		error = __svc_rpcb_register6(program, version,
>>> 						protocol, port);
>>> +		if (error == -EPROTONOSUPPORT)
>>> +			return __svc_rpcb_cant_register(progname, version);
>>> 		if (error < 0)
>>> 			return error;
>>>
>>> @@ -854,7 +884,8 @@ static int __svc_register(const u32 program,  
>>> const u32 version,
>>> * Returns zero on success; a negative errno value is returned
>>> * if any error occurs.
>>> */
>>> -static int __svc_register(const u32 program, const u32 version,
>>> +static int __svc_register(const char *progname,
>>> +			  const u32 program, const u32 version,
>>> 			  sa_family_t family,
>>> 			  const unsigned short protocol,
>>> 			  const unsigned short port)
>>> @@ -901,8 +932,9 @@ int svc_register(const struct svc_serv *serv,  
>>> const unsigned short proto,
>>> 			if (progp->pg_vers[i]->vs_hidden)
>>> 				continue;
>>>
>>> -			error = __svc_register(progp->pg_prog, i,
>>> -						serv->sv_family, proto, port);
>>> +			error = __svc_register(progp->pg_name, progp->pg_prog,
>>> +						i, serv->sv_family,
>>> +						proto, port);
>>> 			if (error < 0)
>>> 				break;
>>> 		}
>>>
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe linux- 
>>> nfs" in
>>> the body of a message to majordomo@vger.kernel.org
>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>
>
> -- 
> Chuck Lever
> chuck[dot]lever[at]oracle[dot]com
>
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs"  
> in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

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