Re: [patch] networking ipv4: remove total socket usage count from /proc/net/sockstat

[Andy Gospodarek <andy@greyhouse.net>]

On 1/16/06, Jesper Juhl <jesper.juhl@gmail.com> wrote:
> >
> Maybe if you described "your current problem" someone could suggest a
> solution...
>

Sure, I'd be glad to.  If I add up all the entries from the procfiles
(in /proc/net) on my system

packet = 1
netlink = 6
raw = 0
raw6 = 0
tcp = 5
tcp6 = 3
udp = 9
udp6 = 1
unix = 29

I find there are a total of 54 sockets open on my system.

Now this seems to differ from the value in /proc/net/sockstat:
# cat sockstat
sockets: used 59
TCP: inuse 5 orphan 0 tw 0 alloc 8 mem 1
UDP: inuse 9
RAW: inuse 0
FRAG: inuse 0 memory 0

So we are off by 5.  I added some code around the stat collection used
in sockstat to get more detailed info about those sockets and the
output is here.  The values are family[protocol family][socket
family].

family[1][1] = 17        (UNIX/LOCAL,STREAM)
family[1][2] = 12        (UNIX/LOCAL,DGRAM)
family[2][1] = 5         (INET,STREAM)
family[2][2] = 9         (INET,DGRAM)
family[2][3] = 2         (INET,RAW)
family[10][1] = 3        (INET6,STREAM)
family[10][2] = 1        (INET6,DGRAM)
family[10][3] = 3        (INET6,RAW)
family[16][2] = 6        (NETLINK/ROUTE,DGRAM)
family[17][10] = 1       (PACKET,PACKET)
Total = 59

All of these numbers match up with what we saw above, except the
INET/RAW and INET6/RAW sockets.  It seems they aren't being counted
correctly -- which accounts for the 5 missing sockets.  The
decrementing of these values is in sock_release() and seems to get
done correctly other times RAW sockets are created, but not for the 5
sockets in question.

Since the total socket usage seems out of place in that file -- and
quite possibly wrong, it seemed like a nice idea to get rid of it
(prior to understanding the reasoning behind keeping it).  Now it
seems the goal will be to fix the discrepancy between these files.

-andy