Question about SIOCGIFCONF

Question about SIOCGIFCONF

[Jeffrey Merkey]

Why is SIOGICONF only instrumented to return a single interface lo for
example.  I noticed that ifconfig always uses /proc/net/dev but the
older SIOCGIFCONF ioctl seems to be busted.  Anyone have an
explanation or is this just how the shit is these days or is the
fucking thing broken (seems to be). ?

Jeff

Re: Question about SIOCGIFCONF

[Eric Dumazet]

Le jeudi 27 mai 2010 à 21:02 -0600, Jeffrey Merkey a écrit :
> Why is SIOGICONF only instrumented to return a single interface lo for
> example.  I noticed that ifconfig always uses /proc/net/dev but the
> older SIOCGIFCONF ioctl seems to be busted.  Anyone have an
> explanation or is this just how the shit is these days or is the
> fucking thing broken (seems to be). ?

Shit comes from you eyes maybe ?

Correction : Shit comes from your eyes, definitely.

Proof :

# strace -o /tmp/STRACE ifconfig -a
# grep SIOCGIFCONF /tmp/STRACE
ioctl(4, SIOCGIFCONF, {120, {{"lo", {AF_INET, inet_addr("127.0.0.1")}},
{"wlan0", {AF_INET, inet_addr("192.168.1.21")}}, {"ppp0", {AF_INET,
inet_addr("10.150.51.210")}}}}) = 0


Part of ifconfig :

    ifc.ifc_buf = NULL;
    for (;;) {
        ifc.ifc_len = sizeof(struct ifreq) * numreqs;
        ifc.ifc_buf = xrealloc(ifc.ifc_buf, ifc.ifc_len);

        if (ioctl(skfd, SIOCGIFCONF, &ifc) < 0) {
            perror("SIOCGIFCONF");
            goto out;
        }
        if (ifc.ifc_len == sizeof(struct ifreq) * numreqs) {
            /* assume it overflowed and try again */
            numreqs += 10;
            continue;
        }
        break;
    }

maybe numreqs should be firt initialized to 64, then doubled each
round...

Re: Question about SIOCGIFCONF

[Steven Whitehouse]

Hi,

On Fri, 2010-05-28 at 06:57 +0200, Eric Dumazet wrote:
> Le jeudi 27 mai 2010 à 21:02 -0600, Jeffrey Merkey a écrit :
> > Why is SIOGICONF only instrumented to return a single interface lo for
> > example.  I noticed that ifconfig always uses /proc/net/dev but the
> > older SIOCGIFCONF ioctl seems to be busted.  Anyone have an
> > explanation or is this just how the shit is these days or is the
> > fucking thing broken (seems to be). ?
> 
> Shit comes from you eyes maybe ?
> 
> Correction : Shit comes from your eyes, definitely.
> 
> Proof :
> 
> # strace -o /tmp/STRACE ifconfig -a
> # grep SIOCGIFCONF /tmp/STRACE
> ioctl(4, SIOCGIFCONF, {120, {{"lo", {AF_INET, inet_addr("127.0.0.1")}},
> {"wlan0", {AF_INET, inet_addr("192.168.1.21")}}, {"ppp0", {AF_INET,
> inet_addr("10.150.51.210")}}}}) = 0
> 
> 
> Part of ifconfig :
> 
>     ifc.ifc_buf = NULL;
>     for (;;) {
>         ifc.ifc_len = sizeof(struct ifreq) * numreqs;
>         ifc.ifc_buf = xrealloc(ifc.ifc_buf, ifc.ifc_len);
> 
>         if (ioctl(skfd, SIOCGIFCONF, &ifc) < 0) {
>             perror("SIOCGIFCONF");
>             goto out;
>         }
>         if (ifc.ifc_len == sizeof(struct ifreq) * numreqs) {
>             /* assume it overflowed and try again */
>             numreqs += 10;
>             continue;
>         }
>         break;
>     }
> 
> maybe numreqs should be firt initialized to 64, then doubled each
> round...
> 

Yes, except that this will fail if you add a protocol that is not known
about by ifconfig and it has a struct sockaddr which is larger than the
16 bytes allowed for in the ifreq. There used to be a DECnet
implementation of this ioctl() but we had to get rid of it because it
broke every single SIOCGFICONF using program (including ifconfig) on the
system for that reason.

This is partly historical. Some systems have a struct sockaddr which
includes a length parameter which means that you can always parse the
returned values, even if they are larger than 16 bytes and the protocol
is unknown. Linux does not have this length field, so you have to know
in advance a full list of all protocols which have sockaddr's greater
than 16 bytes in length in order to parse the results.

The upshot of all that is that it is much better to use rtnetlink, as
per iproute2 instead of this ioctl which is obsolete for most purposes.
A quick grep through the code suggests that ipv4 is the only protocol
which implements this ioctl currently,

Steve.

Re: Question about SIOCGIFCONF

[Jeffrey Merkey]

Actually,  I verified last night it only returns interfaces which have
been bound to an IP address.  It does not return any interfaces which
are active but for which an IP address has not been bound.  So what I
said is accurate.  it's fucking busted.

/proc/net/dev returns ALL interfaces.  this ioctl does not.

Jeff

On Thu, May 27, 2010 at 10:57 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> Le jeudi 27 mai 2010 à 21:02 -0600, Jeffrey Merkey a écrit :
>> Why is SIOGICONF only instrumented to return a single interface lo for
>> example.  I noticed that ifconfig always uses /proc/net/dev but the
>> older SIOCGIFCONF ioctl seems to be busted.  Anyone have an
>> explanation or is this just how the shit is these days or is the
>> fucking thing broken (seems to be). ?
>
> Shit comes from you eyes maybe ?
>
> Correction : Shit comes from your eyes, definitely.
>
> Proof :
>
> # strace -o /tmp/STRACE ifconfig -a
> # grep SIOCGIFCONF /tmp/STRACE
> ioctl(4, SIOCGIFCONF, {120, {{"lo", {AF_INET, inet_addr("127.0.0.1")}},
> {"wlan0", {AF_INET, inet_addr("192.168.1.21")}}, {"ppp0", {AF_INET,
> inet_addr("10.150.51.210")}}}}) = 0
>
>
> Part of ifconfig :
>
>    ifc.ifc_buf = NULL;
>    for (;;) {
>        ifc.ifc_len = sizeof(struct ifreq) * numreqs;
>        ifc.ifc_buf = xrealloc(ifc.ifc_buf, ifc.ifc_len);
>
>        if (ioctl(skfd, SIOCGIFCONF, &ifc) < 0) {
>            perror("SIOCGIFCONF");
>            goto out;
>        }
>        if (ifc.ifc_len == sizeof(struct ifreq) * numreqs) {
>            /* assume it overflowed and try again */
>            numreqs += 10;
>            continue;
>        }
>        break;
>    }
>
> maybe numreqs should be firt initialized to 64, then doubled each
> round...
>
>
>

Re: Question about SIOCGIFCONF

[Jeffrey Merkey]

Review of the net-tools source code for IFCONFIG indicates that when
/proc is not loaded and/or /proc/net/dev is not available, IFCONFIG
will attempt to use this ioctl to determine which interfaces are
present in the system.  Since the ioctl will not report unbound
interfaces which are active, IFCONFIG will not properly report or
detect network adapters which are unbound.  This seems to be a hole,
although most of the time I assume /proc will always be mounted.
Someone should review this and make a decision as to whether or not
this could be a problem.  At any rate, it does not work as advertised.

Jeff

On Fri, May 28, 2010 at 12:31 PM, Jeffrey Merkey <jeffmerkey@gmail.com> wrote:
> Actually,  I verified last night it only returns interfaces which have
> been bound to an IP address.  It does not return any interfaces which
> are active but for which an IP address has not been bound.  So what I
> said is accurate.  it's fucking busted.
>
> /proc/net/dev returns ALL interfaces.  this ioctl does not.
>
> Jeff
>
> On Thu, May 27, 2010 at 10:57 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>> Le jeudi 27 mai 2010 à 21:02 -0600, Jeffrey Merkey a écrit :
>>> Why is SIOGICONF only instrumented to return a single interface lo for
>>> example.  I noticed that ifconfig always uses /proc/net/dev but the
>>> older SIOCGIFCONF ioctl seems to be busted.  Anyone have an
>>> explanation or is this just how the shit is these days or is the
>>> fucking thing broken (seems to be). ?
>>
>> Shit comes from you eyes maybe ?
>>
>> Correction : Shit comes from your eyes, definitely.
>>
>> Proof :
>>
>> # strace -o /tmp/STRACE ifconfig -a
>> # grep SIOCGIFCONF /tmp/STRACE
>> ioctl(4, SIOCGIFCONF, {120, {{"lo", {AF_INET, inet_addr("127.0.0.1")}},
>> {"wlan0", {AF_INET, inet_addr("192.168.1.21")}}, {"ppp0", {AF_INET,
>> inet_addr("10.150.51.210")}}}}) = 0
>>
>>
>> Part of ifconfig :
>>
>>    ifc.ifc_buf = NULL;
>>    for (;;) {
>>        ifc.ifc_len = sizeof(struct ifreq) * numreqs;
>>        ifc.ifc_buf = xrealloc(ifc.ifc_buf, ifc.ifc_len);
>>
>>        if (ioctl(skfd, SIOCGIFCONF, &ifc) < 0) {
>>            perror("SIOCGIFCONF");
>>            goto out;
>>        }
>>        if (ifc.ifc_len == sizeof(struct ifreq) * numreqs) {
>>            /* assume it overflowed and try again */
>>            numreqs += 10;
>>            continue;
>>        }
>>        break;
>>    }
>>
>> maybe numreqs should be firt initialized to 64, then doubled each
>> round...
>>
>>
>>
>

Re: Question about SIOCGIFCONF

[Jeffrey Merkey]

The code in question is net-tools/lib/interface.c function
if_readproc() and associated routines.  Looks like a hole.

Jeff

On Fri, May 28, 2010 at 1:02 PM, Jeffrey Merkey <jeffmerkey@gmail.com> wrote:
> Review of the net-tools source code for IFCONFIG indicates that when
> /proc is not loaded and/or /proc/net/dev is not available, IFCONFIG
> will attempt to use this ioctl to determine which interfaces are
> present in the system.  Since the ioctl will not report unbound
> interfaces which are active, IFCONFIG will not properly report or
> detect network adapters which are unbound.  This seems to be a hole,
> although most of the time I assume /proc will always be mounted.
> Someone should review this and make a decision as to whether or not
> this could be a problem.  At any rate, it does not work as advertised.
>
> Jeff
>
> On Fri, May 28, 2010 at 12:31 PM, Jeffrey Merkey <jeffmerkey@gmail.com> wrote:
>> Actually,  I verified last night it only returns interfaces which have
>> been bound to an IP address.  It does not return any interfaces which
>> are active but for which an IP address has not been bound.  So what I
>> said is accurate.  it's fucking busted.
>>
>> /proc/net/dev returns ALL interfaces.  this ioctl does not.
>>
>> Jeff
>>
>> On Thu, May 27, 2010 at 10:57 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>>> Le jeudi 27 mai 2010 à 21:02 -0600, Jeffrey Merkey a écrit :
>>>> Why is SIOGICONF only instrumented to return a single interface lo for
>>>> example.  I noticed that ifconfig always uses /proc/net/dev but the
>>>> older SIOCGIFCONF ioctl seems to be busted.  Anyone have an
>>>> explanation or is this just how the shit is these days or is the
>>>> fucking thing broken (seems to be). ?
>>>
>>> Shit comes from you eyes maybe ?
>>>
>>> Correction : Shit comes from your eyes, definitely.
>>>
>>> Proof :
>>>
>>> # strace -o /tmp/STRACE ifconfig -a
>>> # grep SIOCGIFCONF /tmp/STRACE
>>> ioctl(4, SIOCGIFCONF, {120, {{"lo", {AF_INET, inet_addr("127.0.0.1")}},
>>> {"wlan0", {AF_INET, inet_addr("192.168.1.21")}}, {"ppp0", {AF_INET,
>>> inet_addr("10.150.51.210")}}}}) = 0
>>>
>>>
>>> Part of ifconfig :
>>>
>>>    ifc.ifc_buf = NULL;
>>>    for (;;) {
>>>        ifc.ifc_len = sizeof(struct ifreq) * numreqs;
>>>        ifc.ifc_buf = xrealloc(ifc.ifc_buf, ifc.ifc_len);
>>>
>>>        if (ioctl(skfd, SIOCGIFCONF, &ifc) < 0) {
>>>            perror("SIOCGIFCONF");
>>>            goto out;
>>>        }
>>>        if (ifc.ifc_len == sizeof(struct ifreq) * numreqs) {
>>>            /* assume it overflowed and try again */
>>>            numreqs += 10;
>>>            continue;
>>>        }
>>>        break;
>>>    }
>>>
>>> maybe numreqs should be firt initialized to 64, then doubled each
>>> round...
>>>
>>>
>>>
>>
>

Re: Question about SIOCGIFCONF

[Eric Dumazet]

Le vendredi 28 mai 2010 à 13:05 -0600, Jeffrey Merkey a écrit :
> The code in question is net-tools/lib/interface.c function
> if_readproc() and associated routines.  Looks like a hole.
> 
> Jeff
> 
> On Fri, May 28, 2010 at 1:02 PM, Jeffrey Merkey <jeffmerkey@gmail.com> wrote:
> > Review of the net-tools source code for IFCONFIG indicates that when
> > /proc is not loaded and/or /proc/net/dev is not available, IFCONFIG
> > will attempt to use this ioctl to determine which interfaces are
> > present in the system.  Since the ioctl will not report unbound
> > interfaces which are active, IFCONFIG will not properly report or
> > detect network adapters which are unbound.  This seems to be a hole,
> > although most of the time I assume /proc will always be mounted.
> > Someone should review this and make a decision as to whether or not
> > this could be a problem.  At any rate, it does not work as advertised.
> >

Maybe you could forget about a 20 years old legacy program and use the
real thing : ip

ip link
ip addr
...

Alternatively, you could rewrite ifconfig to use modern API.

(Not depending on /proc , at all)

Fwd: Question about SIOCGIFCONF

[Jeffrey Merkey]

---------- Forwarded message ----------
From: Jeffrey Merkey <jeffmerkey@gmail.com>
Date: Fri, May 28, 2010 at 3:57 PM
Subject: Re: Question about SIOCGIFCONF
To: Eric Dumazet <eric.dumazet@gmail.com>


Not a bad suggestion, however, redhat and all the major distros still
use ifconfig.   Do you think this should be fixed or is it a
non-issue?

Jeff

On Fri, May 28, 2010 at 2:06 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> Le vendredi 28 mai 2010 à 13:05 -0600, Jeffrey Merkey a écrit :
>> The code in question is net-tools/lib/interface.c function
>> if_readproc() and associated routines.  Looks like a hole.
>>
>> Jeff
>>
>> On Fri, May 28, 2010 at 1:02 PM, Jeffrey Merkey <jeffmerkey@gmail.com> wrote:
>> > Review of the net-tools source code for IFCONFIG indicates that when
>> > /proc is not loaded and/or /proc/net/dev is not available, IFCONFIG
>> > will attempt to use this ioctl to determine which interfaces are
>> > present in the system.  Since the ioctl will not report unbound
>> > interfaces which are active, IFCONFIG will not properly report or
>> > detect network adapters which are unbound.  This seems to be a hole,
>> > although most of the time I assume /proc will always be mounted.
>> > Someone should review this and make a decision as to whether or not
>> > this could be a problem.  At any rate, it does not work as advertised.
>> >
>
> Maybe you could forget about a 20 years old legacy program and use the
> real thing : ip
>
> ip link
> ip addr
> ...
>
> Alternatively, you could rewrite ifconfig to use modern API.
>
> (Not depending on /proc , at all)
>
>
>
>