From mboxrd@z Thu Jan 1 00:00:00 1970 From: Chuck Lever Subject: Re: [PATCH] mount.nfs: set the default family for lookups based on Defaultproto= setting Date: Fri, 05 Feb 2010 12:13:08 -0500 Message-ID: <4B6C51A4.3070005@oracle.com> References: <1265380055-5650-1-git-send-email-jlayton@redhat.com> <4B6C39D0.3030704@oracle.com> <20100205113052.333f0c05@tlielax.poochiereds.net> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII; format=flowed Cc: steved@redhat.com, linux-nfs@vger.kernel.org To: Jeff Layton Return-path: Received: from rcsinet12.oracle.com ([148.87.113.124]:39508 "EHLO rcsinet12.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752478Ab0BEROi (ORCPT ); Fri, 5 Feb 2010 12:14:38 -0500 In-Reply-To: <20100205113052.333f0c05-9yPaYZwiELC+kQycOl6kW4xkIHaj4LzF@public.gmane.org> Sender: linux-nfs-owner@vger.kernel.org List-ID: Hi Jeff- On 02/05/2010 11:30 AM, Jeff Layton wrote: > On Fri, 05 Feb 2010 10:31:28 -0500 > Chuck Lever wrote: > >> On 02/05/2010 09:27 AM, Jeff Layton wrote: >>> When IPv6 is enabled, the Proto= config file option is treated as a >>> netid, and the address family for lookups is selected based on that >>> setting. The Defaultproto= option however still only affects the >>> protocol setting for the sockets (IPPROTO_*) and not the address family. >>> >>> This patch makes it so that if someone sets the "Defaultproto=" option >>> in the nfsmount.conf, it's used to determine the default address family >>> for lookups as well as the protocol type. >>> >>> This gives users a way to force a particular address family to be used >>> universally for mounts and brings the behavior of the Defaultproto= >>> option in line with the Proto= option. >>> >>> Signed-off-by: Jeff Layton >>> --- >>> utils/mount/configfile.c | 8 ++++++++ >>> utils/mount/network.c | 18 ++++++++---------- >>> 2 files changed, 16 insertions(+), 10 deletions(-) >>> >>> diff --git a/utils/mount/configfile.c b/utils/mount/configfile.c >>> index 6843098..71d3627 100644 >>> --- a/utils/mount/configfile.c >>> +++ b/utils/mount/configfile.c >>> @@ -221,6 +221,8 @@ int inline check_vers(char *mopt, char *field) >>> >>> unsigned long config_default_vers; >>> unsigned long config_default_proto; >>> +extern sa_family_t config_default_family; >>> + >>> /* >>> * Check to see if a default value is being set. >>> * If so, set the appropriate global value which will >>> @@ -242,6 +244,12 @@ int inline default_value(char *mopt) >>> xlog_warn("Unable to set default protocol : %s", >>> strerror(errno)); >>> } >>> +#ifdef IPV6_SUPPORTED >>> + if (!nfs_nfs_proto_family(options,&config_default_family)) { >>> + xlog_warn("Unable to set default family : %s", >>> + strerror(errno)); >>> + } >>> +#endif >> >> Maybe you don't need the #ifdef here? >> >> Aside from making the code more readable, removing the #ifdef would also >> make it so the same code path is followed whether IPv6 is supported or >> not, which makes testing easier. >> >> Just a thought. >> > > They probably aren't essential. It all comes down to what we want the > behavior to be in the following situation: OK... I had thought the behavior would be the same with and without the #ifdef's. Since it clearly doesn't work the same way in both cases, that's a sure sign that we will have testing and documentation problems here if the #ifdef is left in. > If someone had a TIRPC enabled nfs-utils but IPv6 support is disabled, > and then set Defaultproto=tcp6. > > With the ifdef's in place, then the address family part of it would > just be ignored. Without those in place, IPv6 will be forced for the > lookup but we won't actually be able to make use of the addresses > returned. As an admin (who understands how netids work), I would be more surprised if setting the default protocol (netid, really) to tcp6, and then disabling IPv6 on the client, or at build time in nfs-utils, or if the IPv6-related netids were missing from /etc/netconfig, would work at all. It would be more consistent behaviour IMO to make this case fail. I agree with having things "just work" in reasonable cases, but it seems like we are bending the semantics of netids and proto= to make that happen. "tcp6" does not mean "use IPv6 if it's available." It means "always use IPv6." And, "tcp" does not mean "use TCP on either IPv4 or IPv6," it now means "always use TCP on IPv4." This is what we buy with netids and TI-RPC. That's why we use HAVE_LIBTIRPC rather than IPV6_SUPPORTED in nfs_nfs_proto_family and nfs_mount_proto_family. > Granted, it's a bit of a pathological setup, but I don't think the > #ifdef's here really take away anything and may help prevent confusion > with the union of all these different options. > >>> } else { >>> xlog_warn("Unable to alloc memory for default protocol"); >>> } >>> diff --git a/utils/mount/network.c b/utils/mount/network.c >>> index 92bba2d..0ab3bb1 100644 >>> --- a/utils/mount/network.c >>> +++ b/utils/mount/network.c >>> @@ -1331,6 +1331,12 @@ nfs_nfs_port(struct mount_options *options, unsigned long *port) >>> return 1; >>> } >>> >>> +#ifdef IPV6_SUPPORTED By the above argument, this should be HAVE_LIBTIRPC. >>> +sa_family_t config_default_family = AF_UNSPEC; >>> +#else >>> +sa_family_t config_default_family = AF_INET; >>> +#endif >>> + >>> /* >>> * Returns TRUE and fills in @family if a valid NFS protocol option >>> * is found, or FALSE if the option was specified with an invalid value. >>> @@ -1341,11 +1347,7 @@ int nfs_nfs_proto_family(struct mount_options *options, >>> unsigned long protocol; >>> char *option; >>> >>> -#ifdef IPV6_SUPPORTED >>> - *family = AF_UNSPEC; >>> -#else >>> - *family = AF_INET; >>> -#endif >>> + *family = config_default_family; >>> >>> switch (po_rightmost(options, nfs_transport_opttbl)) { >>> case 0: /* udp */ >>> @@ -1488,11 +1490,7 @@ int nfs_mount_proto_family(struct mount_options *options, >>> unsigned long protocol; >>> char *option; >>> >>> -#ifdef HAVE_LIBTIRPC >>> - *family = AF_UNSPEC; >>> -#else >>> - *family = AF_INET; >>> -#endif >>> + *family = config_default_family; >>> >>> option = po_get(options, "mountproto"); >>> if (option != NULL) >> >> > > -- chuck[dot]lever[at]oracle[dot]com