could not create IPv6 socket (AI_ADDRCONFIG)

Hello, I have often seen inquiries about an log message from
PostgreSQL server.

> LOG: could not create IPv6 socket: Address family not supported by protocol

This is emitted on ipv6-disabled environment which was available
for me by the following steps on CentOS 6.5,

- Add 'NETWORKING_IPV6=no' into /etc/sysconfig/network

- Create /etc/modprobe.d/disable-ipv6.conf with the following
content.

> options net-pf-10 off
> install ipv6 /bin/true

- Comment out the entry for ::1 in /etc/hosts

- Reboot.

- [Confirm that "ip a" and "ifconfig -a" don't show ipv6 addrs]

The cause is that the server collects available listen addresses
by getaddrinfo with hint.ai_flags not having AI_ADDRCONFIG set.

pqcomm.c:299 in function StreamServerPort
> hint.ai_family = family;
!> hint.ai_flags = AI_PASSIVE;
> hint.ai_socktype = SOCK_STREAM;

The man page of getaddrinfo says that AI_ADDRCONFIG would help
and I saw it actually did.

Well, I excavated some discussions about this option from ancient
pgsql-hackers, a decade ago.

http://www.postgresql.org/message-id/20030703204355.GA12774@ping.be
> This looks a little broken behaviour to me. My guess is that
> it returns an AF_INET6 socket but doesn't support it in the
> kernel. In that case with the AI_ADDRCONFIG option it
> shouldn't have returned that address. The question is wether
> your getaddrinfo() supports that option, and wether it's
> working or not.

I suppose AI_ADDRCONFIG is a quite obscure option at the time,
but the time seems to have come when we can use this option.

man getaddrinfo
>NOTES
> getaddrinfo() supports the address%scope-id notation for
> specifying the IPv6 scope-ID.
>
> AI_ADDRCONFIG, AI_ALL, and AI_V4MAPPED are available
> since glibc 2.3.3. AI_NUMERICSERV is available since
> glibc 2.3.4.

RHEL/CentOS 4.9 and after seems satisfies the condition. I don't
know about other platforms but it is enough to define it as 0 if
not defined. In addition, it would not be a problem even if the
option doesn't work as expected because the error handling code
added at the time the above conversation took place would act as
the last resort. There would be no way it can work destructively.

Attached patch does this.

Any suggestions ?

regards,

--
Kyotaro Horiguchi
NTT Open Source Software Center

Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> writes:
> Hello, I have often seen inquiries about an log message from
> PostgreSQL server.
>> LOG: could not create IPv6 socket: Address family not supported by protocol

That's merely a harmless log message.

> - hints.ai_flags = AI_PASSIVE;
> + hints.ai_flags = AI_PASSIVE | AI_ADDRCONFIG;

This, on the other hand, might actively break things. It did when we
had it in before (cf the thread you link to and commit df63503dc).
I don't have any faith that systems on which it is broken have vanished
from the face of the earth.

One good reason not to trust this too much is that getaddrinfo() is
fundamentally a userspace DNS access function, and as such it has
no very good way to know if there's currently an IPv4 or IPv6
interface configured on the local system. At minimum there are
obvious race conditions in that.

If we're concerned about users worrying about log messages from
this, I'd rather see us downgrade those log messages to DEBUG level
than risk breaking the code with behaviors that were proven to be
a bad idea a decade ago. But TBH I see no strong need to do anything
here.

regards, tom lane

Tom Lane wrote:
> Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> writes:
> > Hello, I have often seen inquiries about an log message from
> > PostgreSQL server.
> >> LOG: could not create IPv6 socket: Address family not supported by protocol
>
> That's merely a harmless log message.

> If we're concerned about users worrying about log messages from
> this, I'd rather see us downgrade those log messages to DEBUG level
> than risk breaking the code with behaviors that were proven to be
> a bad idea a decade ago. But TBH I see no strong need to do anything
> here.

How about just adding a HINT?

--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services

Alvaro Herrera <alvherre(at)2ndquadrant(dot)com> writes:
> Tom Lane wrote:
>> Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> writes:
>>> Hello, I have often seen inquiries about an log message from
>>> PostgreSQL server.
>>> LOG: could not create IPv6 socket: Address family not supported by protocol

>> That's merely a harmless log message.

> How about just adding a HINT?

Hmm ... maybe, but how would you phrase the hint exactly?

regards, tom lane

Hello,

At Tue, 04 Feb 2014 02:07:08 -0500, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote in <3176(dot)1391497628(at)sss(dot)pgh(dot)pa(dot)us>
> One good reason not to trust this too much is that getaddrinfo() is
> fundamentally a userspace DNS access function, and as such it has
> no very good way to know if there's currently an IPv4 or IPv6
> interface configured on the local system. At minimum there are
> obvious race conditions in that.

A case which would be more common is "::1" in /etc/hosts. I had
following error with this patch for such a case.

| LOG: could not bind IPv4 socket: Address already in use
| HINT: Is another postmaster already running on port 5432? If not, wait a few seconds and retry.

getaddrinfo returned two same entries having the same address
AF_INET "127.0.0.1:14357". One of them is for "::1" in
hosts. This is worse than current behavior X-(

At Tue, 04 Feb 2014 10:31:03 -0500, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote in <12552(dot)1391527863(at)sss(dot)pgh(dot)pa(dot)us>
> Alvaro Herrera <alvherre(at)2ndquadrant(dot)com> writes:
> > Tom Lane wrote:
> >> Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> writes:
> >>> LOG: could not create IPv6 socket: Address family not supported by protocol
>
> >> That's merely a harmless log message.
>
> > How about just adding a HINT?
>
> Hmm ... maybe, but how would you phrase the hint exactly?

Putting the 'exactly' aside, is it something means 'You will get
this message when the feature to handle the address family is
disabled', only for EAFNOSUPPORT ?

Though I don't know whether such a hint is helpful for those who
tend to mind that kind of message.

regards,

--
Kyotaro Horiguchi
NTT Open Source Software Center

Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> writes:
> getaddrinfo returned two same entries having the same address
> AF_INET "127.0.0.1:14357". One of them is for "::1" in
> hosts. This is worse than current behavior X-(

Yeah, the fundamental issue is that getaddrinfo tends to return bogus
info.

>>> How about just adding a HINT?

>> Hmm ... maybe, but how would you phrase the hint exactly?

> Putting the 'exactly' aside, is it something means 'You will get
> this message when the feature to handle the address family is
> disabled', only for EAFNOSUPPORT ?

> Though I don't know whether such a hint is helpful for those who
> tend to mind that kind of message.

I still think the best thing might be to reduce the individual messages
to DEBUG-something, and only produce a LOG entry if we are unable to
bind to *any* of the addresses returned by getaddrinfo.

regards, tom lane