Patch for IPv6 on Windows before Vista

[Sob <so...@hisoftware.cz>]

Hello,

while experimenting with httpd and IPv6 on WinXP, I found out that 
it's not working well.
In short, the code expects that IPv4 mapped addresses and IPV6_V6ONLY 
socket option are supported. This is not true for NTs < 6.0. They 
have IPv4 and IPv6 as two separate stacks and therefore all IPv6 
sockets always behave as V6ONLY and no IPv4 mapped addresses are 
possible. Unfortunately the IPV6_V6ONLY option is not understood, so 
if the code tries to set it, if fails with OS error. APR seems to be 
the best place where it can be fixed.

The attached patch modifies apr_socket_opt_set(). When the requested 
option is APR_IPV6_V6ONLY, it checks Windows version and if it's less 
than Vista, it does not call setsockopt(), but instead it calls 
directly apr_set_option() and pretends that the option was set. Then 
it returns APR_SUCCESS or APR_ENOTIMPL depending on if it was 
requested to set or unset the option.
I'm not exactly sure about returning APR_ENOTIMPL, but httpd would 
fail with anything else. At first I just set "on = 1;" and let the 
original apr_set_option() use it, but that wasn't right, because it 
always returned APR_SUCCESS and some application other than httpd can 
depend on the result and in case it requested to unset the option, 
returning APR_SUCCESS would be wrong. Httpd doesn't really care about 
the result and checks the state later using apr_socket_opt_get(). 
That's why always calling apr_set_option(sock, APR_IPV6_V6ONLY, 1) is 
important. If omited, it would be possible (if compiled with IPv4 
mapped addresses support) for apr_socket_opt_get() to return 0 (= 
socket can accept both IPv4 and 6), but that's not possible on these 
older Windows and if some code is depending on it, it may fail (like 
current httpd).

Sob

Re: Patch for IPv6 on Windows before Vista

["William A. Rowe Jr." <wr...@rowe-clan.net>]

On 4/12/2011 5:44 PM, Guenter Knauf wrote:
> Am 12.04.2011 21:01, schrieb Jeff Trawick:
>> On Tue, Apr 12, 2011 at 2:37 PM, Jeff Trawick<tr...@gmail.com>  wrote:
>>> On Thu, Jun 19, 2008 at 9:56 PM, Sob<so...@hisoftware.cz>  wrote:
>>>> Hello,
>>>>
>>>> while experimenting with httpd and IPv6 on WinXP, I found out that it's not
>>>> working well.
>>>> In short, the code expects that IPv4 mapped addresses and IPV6_V6ONLY socket
>> ...
>>
>> I just noticed that this is also tracked at
>> https://issues.apache.org/bugzilla/show_bug.cgi?id=45321
>>
> a C&P from a recent IRC talk with Gregg L. Smith:

As Jeff points out, this introduces binary incompatibility anyways :(

While we might bundle an IPv6 build with 2.3 beta (and I think we should)...

it isn't going to be possible for 2.2.18 if we want to keep loadmodule
compatibility (which gets at what Gregg was saying about two downloads).

Re: Patch for IPv6 on Windows before Vista

[Guenter Knauf <fu...@apache.org>]

Am 12.04.2011 21:01, schrieb Jeff Trawick:
> On Tue, Apr 12, 2011 at 2:37 PM, Jeff Trawick<tr...@gmail.com>  wrote:
>> On Thu, Jun 19, 2008 at 9:56 PM, Sob<so...@hisoftware.cz>  wrote:
>>> Hello,
>>>
>>> while experimenting with httpd and IPv6 on WinXP, I found out that it's not
>>> working well.
>>> In short, the code expects that IPv4 mapped addresses and IPV6_V6ONLY socket
> ...
>
> I just noticed that this is also tracked at
> https://issues.apache.org/bugzilla/show_bug.cgi?id=45321
>
a C&P from a recent IRC talk with Gregg L. Smith:

Gregg: Hola, reading APR list
Gregg: If the turn on IPv6 in apr 1.4.3, therefore 2.2.18, you are going 
to get a hell of a lot of Windows bug reports unless that patch by Sob 
is comitted
Gregg: XP will be no problem till everyone reads (ah, I can do IPv6 now) 
and installs the IPv6 proto, at which point Apache will crash with a 
duplicate socket error
/me: well, why dont you participate the discussion, and bring up your 
experiences?
/me: well, you made your experience - tell us all!!
Gregg: I do not feel like subscribing to apr, but trust me, I tested it 
a hundred times 2 years ago and know, this is the reason we have 
seperate downloads at the haus
Gregg: copy that, paste it in a message and say this is the info you 
received from me via IRC

.... which I just did :-)

Gün.

Re: Patch for IPv6 on Windows before Vista

[Jeff Trawick <tr...@gmail.com>]

On Tue, Apr 12, 2011 at 2:37 PM, Jeff Trawick <tr...@gmail.com> wrote:
> On Thu, Jun 19, 2008 at 9:56 PM, Sob <so...@hisoftware.cz> wrote:
>> Hello,
>>
>> while experimenting with httpd and IPv6 on WinXP, I found out that it's not
>> working well.
>> In short, the code expects that IPv4 mapped addresses and IPV6_V6ONLY socket
...

I just noticed that this is also tracked at
https://issues.apache.org/bugzilla/show_bug.cgi?id=45321

Re: Patch for IPv6 on Windows before Vista

[Jeff Trawick <tr...@gmail.com>]

On Thu, Jun 19, 2008 at 9:56 PM, Sob <so...@hisoftware.cz> wrote:
> Hello,
>
> while experimenting with httpd and IPv6 on WinXP, I found out that it's not
> working well.
> In short, the code expects that IPv4 mapped addresses and IPV6_V6ONLY socket
> option are supported. This is not true for NTs < 6.0. They have IPv4 and
> IPv6 as two separate stacks and therefore all IPv6 sockets always behave as
> V6ONLY and no IPv4 mapped addresses are possible. Unfortunately the
> IPV6_V6ONLY option is not understood, so if the code tries to set it, if
> fails with OS error. APR seems to be the best place where it can be fixed.
>
> The attached patch modifies apr_socket_opt_set(). When the requested option
> is APR_IPV6_V6ONLY, it checks Windows version and if it's less than Vista,
> it does not call setsockopt(), but instead it calls directly
> apr_set_option() and pretends that the option was set. Then it returns
> APR_SUCCESS or APR_ENOTIMPL depending on if it was requested to set or unset
> the option.
> I'm not exactly sure about returning APR_ENOTIMPL, but httpd would fail with
> anything else. At first I just set "on = 1;" and let the original
> apr_set_option() use it, but that wasn't right, because it always returned
> APR_SUCCESS and some application other than httpd can depend on the result
> and in case it requested to unset the option, returning APR_SUCCESS would be
> wrong. Httpd doesn't really care about the result and checks the state later
> using apr_socket_opt_get(). That's why always calling apr_set_option(sock,
> APR_IPV6_V6ONLY, 1) is important. If omited, it would be possible (if
> compiled with IPv4 mapped addresses support) for apr_socket_opt_get() to
> return 0 (= socket can accept both IPv4 and 6), but that's not possible on
> these older Windows and if some code is depending on it, it may fail (like
> current httpd).
>
> Sob

See also http://msdn.microsoft.com/en-us/library/ms738574%28v=vs.85%29.aspx

< Vista: socket option not supported, but behaves as if it is hard-coded to 1
>= Vista: socket option works, and IPV6_V6ONLY=1 is the default state

Suggested changes to the patch:

Move "apr_set_option(sock, APR_IPV6_V6ONLY, 1);" to creation of the
socket, and make it unconditional (old or new Windows).  That allows
apr_socket_opt_get() to work even if apr_socket_opt_set() isn't called
first.*  (This reflects the default setting of 1 on levels of Windows
that support the option, as documented above and confirmed on Windows
7.)

Add the default definition of IPV6_V6ONLY right in sockopt.c.  It
won't be needed anywhere else.

Thoughts?

FWIW, I'm aware that some folks who have tested different IPv6
scenarios on old/new Windows boxes are using about the same code that
Sob has posted.

*This issue of socket_opt_get() before socket_opt_set() seems to be a
wider issue with APR.  There are a few options which must pass through
to the OS to retrieve, but we don't do that.