Take the 2-minute tour ×
Unix & Linux Stack Exchange is a question and answer site for users of Linux, FreeBSD and other Un*x-like operating systems.. It's 100% free, no registration required.

I want to determine which process has the other end of a UNIX socket.

Specifically, I'm asking about one that was created with socketpair(), though the problem is the same for any UNIX socket.

I have a program parent which creates a socketpair(AF_UNIX, SOCK_STREAM, 0, fds), and fork()s. The parent process closes fds[1] and keeps fds[0] to communicate. The child does the opposite, close(fds[0]); s=fds[1]. Then the child exec()s another program, child1. The two can communicate back and forth via this socketpair.

Now, let's say I know who parent is, but I want to figure out who child1 is. How do I do this?

There are several tools at my disposal, but none can tell me which process is on the other end of the socket. I have tried:

  • lsof -c progname
  • lsof -c parent -c child1
  • ls -l /proc/$(pidof server)/fd
  • cat /proc/net/unix

Basically, I can see the two sockets, and everything about them, but cannot tell that they are connected. I am trying to determine which FD in the parent is communicating with which child process.

share|improve this question

5 Answers 5

up vote 16 down vote accepted

Since kernel 3.3, it is possible using ss — see Stéphane Chazelas's answer.

In older versions, according to the author of lsof, it was impossible to find this out: the Linux kernel does not expose this information. Source: 2003 thread on comp.unix.admin.

The number shown in /proc/$pid/fd/$fd is the socket's inode number in the virtual socket filesystem. When you create a pipe or socket pair, each end successively receives an inode number. The numbers are attributed sequentially, so there is a high probability that the numbers differ by 1, but this is not guaranteed (either because the first socket was N and N+1 was already in use due to wrapping, or because some other thread was scheduled between the two inode allocations and that thread created some inodes too).

I checked the definition of socketpair in kernel 2.6.39, and the two ends of the socket are not correlated except by the type-specific socketpair method. For unix sockets, that's unix_socketpair in net/unix/af_unix.c.

share|improve this answer
2  
Thanks @Gillles. I do recall reading something about that a while back, but was unable to find it again. I may just have to go writing a patch for /proc/net/unix. –  Jonathon Reinhart Jul 9 '11 at 22:09
    
And yes, I'd made that observation with the increasing inode numbers, and currently that's what I'm working with. However, as you noted, it is not guaranteed. The process I'm looking at has at least 40 open unix sockets, and I saw one instance where the N+1 did not hold true. Bummer. –  Jonathon Reinhart Jul 9 '11 at 22:11
1  
@JonathonReinhart I checked the definition of socketpair, and the two ends of the socket are not correlated except by the type-specific socketpair method. For unix sockets, that's unix_socketpair in `net/unix/af_unix.c. It would be nice to have this information for pipes, too. –  Gilles Jul 9 '11 at 22:35
3  
As @mikeserv found out, newer Linux kernels (since December 2011, 3.3 and above) expose that information via a netlink API which you can query with the ss command. So it looks like this answer is now obsolete. –  Stéphane Chazelas Mar 17 at 14:58

Erkki Seppala actually has a tool that retrieves this information from the Linux kernel with gdb.. It's available here.

share|improve this answer
2  
Very useful information! Even if the tool didn't work out of the box for me (it caused a kernel Oops), the idea helped me to identify the other end. I described my solution on Stack Overflow. –  MvG Aug 15 '12 at 20:33

Note: I now maintain a lsof wrapper that combines both approaches described here and also adds information for peers of loopback TCP connections at https://github.com/stephane-chazelas/misc-scripts/blob/master/lsofc

Linux-3.3 and above.

On Linux, since kernel version 3.3 (and provided the UNIX_DIAG feature is built in the kernel), the peer of a given unix domain socket (includes socketpairs) can be obtained using a new netlink based API. The ss utility makes use of that API to retrieve and display information on the list of unix domain sockets on the system including peer information.

The sockets are identified by their inode number. Note that it's not related to the filesystem inode of the socket file.

For instance in:

$ ss -x
[...]
u_str  ESTAB    0    0   @/tmp/.X11-unix/X0 3435997     * 3435996

it says that socket 3435997 (that was bound to the ABSTRACT socket /tmp/.X11-unix/X0) is connected with socket 3435996. The -p option can tell you which process(es) have that socket open. It does that by doing some readlinks on /proc/$pid/fd/*, so it can only do that on processes you own (unless you're root). For instance here:

$ sudo ss -xp
[...]
u_str  ESTAB  0  0  @/tmp/.X11-unix/X0 3435997 * 3435996 users:(("Xorg",pid=3080,fd=83))
[...]
$ sudo ls -l /proc/3080/fd/23
lrwx------ 1 root root 64 Mar 12 16:34 /proc/3080/fd/83 -> socket:[3435997]

To find out what process(es) has 3435996, you can look up its own entry in the output of ss -xp:

$ ss -xp | awk '$6 == 3435996'
u_str  ESTAB  0  0  * 3435996  * 3435997 users:(("xterm",pid=29215,fd=3))

You could also use this script as a wrapper around lsof to easily show the relevant information there:

#! /usr/bin/perl
# lsof wrapper to add peer information for unix domain socket.
# Needs Linux 3.3 or above and CONFIG_UNIX_DIAG enabled.

# retrieve peer and direction information from ss
my (%peer, %dir);
open SS, '-|', 'ss', '-nexa';
while (<SS>) {
  if (/\s(\d+)\s+\*\s+(\d+) ([<-]-[->])$/) {
    $peer{$1} = $2;
    $dir{$1} = $3;
  }
}
close SS;

# Now get info about processes tied to sockets using lsof
my (%fields, %proc);
open LSOF, '-|', 'lsof', '-nPUFpcfin';
while (<LSOF>) {
  if (/(.)(.*)/) {
    $fields{$1} = $2;
    if ($1 eq 'n') {
      $proc{$fields{i}}->{"$fields{c},$fields{p}" .
      ($fields{n} =~ m{^([@/].*?)( type=\w+)?$} ? ",$1" : "")} = "";
    }
  }
}
close LSOF;

# and finally process the lsof output
open LSOF, '-|', 'lsof', @ARGV;
while (<LSOF>) {
  chomp;
  if (/\sunix\s+\S+\s+\S+\s+(\d+)\s/) {
    my $peer = $peer{$1};
    if (defined($peer)) {
      $_ .= $peer ?
            " ${dir{$1}} $peer\[" . (join("|", keys%{$proc{$peer}})||"?") . "]" :
            "[LISTENING]";
    }
  }
  print "$_\n";
}
close LSOF or exit(1);

For example:

$ sudo that-lsof-wrapper -ad3 -p 29215
COMMAND   PID     USER   FD   TYPE             DEVICE SIZE/OFF    NODE NAME
xterm   29215 stephane    3u  unix 0xffff8800a07da4c0      0t0 3435996 type=STREAM <-> 3435997[Xorg,3080,@/tmp/.X11-unix/X0]

Before linux-3.3

The old Linux API to retrieve unix socket information is via the /proc/net/unix text file. It lists all the Unix domain sockets (including socketpairs). The first field in there (if not hidden to non-superusers with the kernel.kptr_restrict sysctl parameter) as already explained by @Totor contains the kernel address of a unix_sock structure that contains a peer field pointing to the corresponding peer unix_sock. It's also what lsof outputs for the DEVICE column on a Unix socket.

Now getting the value of that peer field means being able to read kernel memory and know the offset of that peer field with regards to the unix_sock address.

Several gdb-based and systemtap-based solutions have already been given but they require gdb/systemtap and Linux kernel debug symbols for the running kernel being installed which is generally not the case on production systems.

Hardcoding the offset is not really an option as that varies with kernel version.

Now we can use a heuristic approach at determining the offset: have our tool create a dummy socketpair (then we know the address of both peers), and search for the address of the peer around the memory at the other end to determine the offset.

Here is a proof-of-concept script that does just that using perl (successfully tested with kernel 2.4.27 and 2.6.32 on i386 and 3.13 and 3.16 on amd64). Like above, it works as a wrapper around lsof:

For example:

$ that-lsof-wrapper -aUc nm-applet
COMMAND    PID     USER   FD   TYPE             DEVICE SIZE/OFF  NODE NAME
nm-applet 4183 stephane    4u  unix 0xffff8800a055eb40      0t0 36888 type=STREAM -> 0xffff8800a055e7c0[dbus-daemon,4190,@/tmp/dbus-AiBCXOnuP6]
nm-applet 4183 stephane    7u  unix 0xffff8800a055e440      0t0 36890 type=STREAM -> 0xffff8800a055e0c0[Xorg,3080,@/tmp/.X11-unix/X0]
nm-applet 4183 stephane    8u  unix 0xffff8800a05c1040      0t0 36201 type=STREAM -> 0xffff8800a05c13c0[dbus-daemon,4118,@/tmp/dbus-yxxNr1NkYC]
nm-applet 4183 stephane   11u  unix 0xffff8800a055d080      0t0 36219 type=STREAM -> 0xffff8800a055d400[dbus-daemon,4118,@/tmp/dbus-yxxNr1NkYC]
nm-applet 4183 stephane   12u  unix 0xffff88022e0dfb80      0t0 36221 type=STREAM -> 0xffff88022e0df800[dbus-daemon,2268,/var/run/dbus/system_bus_socket]
nm-applet 4183 stephane   13u  unix 0xffff88022e0f80c0      0t0 37025 type=STREAM -> 0xffff88022e29ec00[dbus-daemon,2268,/var/run/dbus/system_bus_socket]

Here's the script:

#! /usr/bin/perl
# wrapper around lsof to add peer information for Unix
# domain sockets. needs lsof, and superuser privileges.
# Copyright Stephane Chazelas 2015, public domain.
# example: sudo this-lsof-wrapper -aUc Xorg
use Socket;

open K, "<", "/proc/kcore" or die "open kcore: $!";
read K, $h, 8192 # should be more than enough
 or die "read kcore: $!";

# parse ELF header
my ($t,$o,$n) = unpack("x4Cx[C19L!]L!x[L!C8]S", $h);
$t = $t == 1 ? "L3x4Lx12" : "Lx4QQx8Qx16"; # program header ELF32 or ELF64
my @headers = unpack("x$o($t)$n",$h);

# read data from kcore at given address (obtaining file offset from ELF
# @headers)
sub readaddr {
  my @h = @headers;
  my ($addr, $length) = @_;
  my $offset;
  while (my ($t, $o, $v, $s) = splice @h, 0, 4) {
    if ($addr >= $v && $addr < $v + $s) {
      $offset = $o + $addr - $v;
      if ($addr + $length - $v > $s) {
        $length = $s - ($addr - $v);
      }
      last;
    }
  }
  return undef unless defined($offset);
  seek K, $offset, 0 or die "seek kcore: $!";
  my $ret;
  read K, $ret, $length or die "read($length) kcore \@$offset: $!";
  return $ret;
}

# create a dummy socketpair to try find the offset in the
# kernel structure
socketpair(Rdr, Wtr, AF_UNIX, SOCK_STREAM, PF_UNSPEC)
 or die "socketpair: $!";
$r = readlink("/proc/self/fd/" . fileno(Rdr)) or die "readlink Rdr: $!";
$r =~ /\[(\d+)/; $r = $1;
$w = readlink("/proc/self/fd/" . fileno(Wtr)) or die "readlink Wtr: $!";
$w =~ /\[(\d+)/; $w = $1;
# now $r and $w contain the socket inodes of both ends of the socketpair
die "Can't determine peer offset" unless $r && $w;

# get the inode->address mapping
open U, "<", "/proc/net/unix" or die "open unix: $!";
while (<U>) {
  if (/^([0-9a-f]+):(?:\s+\S+){5}\s+(\d+)/) {
    $addr{$2} = hex $1;
  }
}
close U;

die "Can't determine peer offset" unless $addr{$r} && $addr{$w};

# read 2048 bytes starting at the address of Rdr and hope to find
# the address of Wtr referenced somewhere in there.
$around = readaddr $addr{$r}, 2048;
my $offset = 0;
my $ptr_size = length(pack("L!",0));
my $found;
for (unpack("L!*", $around)) {
  if ($_ == $addr{$w}) {
    $found = 1;
    last;
  }
  $offset += $ptr_size;
}
die "Can't determine peer offset" unless $found;

my %peer;
# now retrieve peer for each socket
for my $inode (keys %addr) {
  $peer{$addr{$inode}} = unpack("L!", readaddr($addr{$inode}+$offset,$ptr_size));
}
close K;

# Now get info about processes tied to sockets using lsof
my (%fields, %proc);
open LSOF, '-|', 'lsof', '-nPUFpcfdn';
while (<LSOF>) {
  if (/(.)(.*)/) {
    $fields{$1} = $2;
    if ($1 eq 'n') {
      $proc{hex($fields{d})}->{"$fields{c},$fields{p}" .
      ($fields{n} =~ m{^([@/].*?)( type=\w+)?$} ? ",$1" : "")} = "";
    }
  }
}
close LSOF;

# and finally process the lsof output
open LSOF, '-|', 'lsof', @ARGV;
while (<LSOF>) {
  chomp;
  for my $addr (/0x[0-9a-f]+/g) {
    $addr = hex $addr;
    my $peer = $peer{$addr};
    if (defined($peer)) {
      $_ .= $peer ?
            sprintf(" -> 0x%x[", $peer) . join("|", keys%{$proc{$peer}}) . "]" :
            "[LISTENING]";
      last;
    }
  }
  print "$_\n";
}
close LSOF or exit(1);
share|improve this answer
1  
@mikeserv, that's a follow-up on that comment. Not being able to find the other end of unix sockets is something that has always annoyed me (often when trying to find X clients and there was a recent question about that). I'll try and see if a similar approach can be used for pseudo-terminals and suggest those to the lsof author. –  Stéphane Chazelas Mar 17 at 9:26
1  
I still can't believe this isn't provided by the kernel itself! I should really submit a patch, if for nothing else but to discover why it doesn't already exist. –  Jonathon Reinhart Mar 17 at 10:10
1  
does ss not do this? It's kind of over my head, but ss -px lists a lot of unix sockets with peer information like: users: ("nacl_helper",pid=18992,fd=6),("chrome",pid=18987,fd=6),("chrome",pid=18975,fd=‌​5)) u_str ESTAB\t0\t0\t/run/dbus/system_bus_socket 8760\t\t* 15068 and the column headings are... State\tRecv-Q\tSend-Q\tLocal Address:Port\tPeer Address:Port –  mikeserv Mar 17 at 12:31
1  
Also, if I do lsof -c terminology I can see terminolo 12731\tmikeserv\t12u\tunix\t0xffff880600e82680\t0t0\t1312426\ttype=STREAM but if I do ss -px | grep terminology I get: u_str\tESTAB\t0\t0\t* 1312426\t*1315046\tusers:(("terminology",pid=12731,fd=12)) –  mikeserv Mar 17 at 12:37
1  
@mikeserv, it looks like it does indeed! It seems I've been wasting a lot of time lately... –  Stéphane Chazelas Mar 17 at 13:00

Since kernel 3.3

You can now get this information with ss:

# ss -xp

Now you can see in the Peer column an ID (inode number) which corresponds to another ID in the Local column. Matching IDs are the two ends of a socket.

Note: The UNIX_DIAG option must be enabled in your kernel.

Before kernel 3.3

Linux didn't expose this information to userland.

However, by looking into kernel memory, we can access this information.

Note: This answer does so by using gdb, however, please see @StéphaneChazelas' answer which is more elaborated in this regard.

# lsof | grep whatever
mysqld 14450 (...) unix 0xffff8801011e8280 (...) /var/run/mysqld/mysqld.sock
mysqld 14450 (...) unix 0xffff8801011e9600 (...) /var/run/mysqld/mysqld.sock

There is 2 different sockets, 1 listening and 1 established. The hexa number is the address to the corresponding kernel unix_sock structure, having a peer attribute being the address of the other end of the socket (also a unix_sock structure instance).

Now we can use gdb to find the peer within kernel memory:

# gdb /usr/lib/debug/boot/vmlinux-3.2.0-4-amd64 /proc/kcore
(gdb) print ((struct unix_sock*)0xffff8801011e9600)->peer
$1 = (struct sock *) 0xffff880171f078c0

# lsof | grep 0xffff880171f078c0
mysql 14815 (...) unix 0xffff880171f078c0 (...) socket

Here you go, the other end of the socket is held by mysql, PID 14815.

Your kernel must be compiled with KCORE_ELF to use /proc/kcore. Also, you need a version of your kernel image with debugging symbols. On Debian 7, apt-get install linux-image-3.2.0-4-amd64-dbg will provide this file.

No need for the debuggable kernel image...

If you don't have (or don't want to keep) the debugging kernel image on the system, you can give gdb the memory offset to "manually" access the peer value. This offset value usually differ with kernel version or architecture.

On my kernel, I know the offset is 680 bytes, that is 85 times 64 bits. So I can do:

# gdb /boot/vmlinux-3.2.0-4-amd64 /proc/kcore
(gdb) print ((void**)0xffff8801011e9600)[85]
$1 = (void *) 0xffff880171f078c0

Voilà, same result as above.

If you have the same kernel running on several machine, it is easier to use this variant because you don't need the debug image, only the offset value.

To (easily) discover this offset value at first, you do need the debug image:

$ pahole -C unix_sock /usr/lib/debug/boot/vmlinux-3.2.0-4-amd64
struct unix_sock {
  (...)
  struct sock *              peer;                 /*   680     8 */
  (...)
}

Here you go, 680 bytes, this is 85 x 64 bits, or 170 x 32 bits.

Most of the credit for this answer goes to MvG.

share|improve this answer
2  
Another approach to retrieve the offset could be to create a socketpair, identify the corresponding entries in /proc/net/unix based on inode numbers from readlinks on /proc/pif/fd/*, and scan memory around the address of one socket for the address of the other. That could make for a reasonably portable (accross Linux versions and architectures) that could be implemented by lsof itself. I'll try to come up with a PoC. –  Stéphane Chazelas Mar 13 at 19:37
2  
I've now added such a PoC which seems to work well on the systems I've tested. –  Stéphane Chazelas Mar 17 at 0:19

This solution, though working, is of limited interest since if you have a recent-enough systemtap, chances are you'll have a recent-enough kernel where you can use ss based approaches, and if you're on an older kernel, that other solution, though more hacky is more likely to work and doesn't require addition software.

Still useful as a demonstration of how to use systemtap for this kind of task.

If on a recent Linux system with a working systemtap (1.8 or newer), you could use the script below to post-process the output of lsof:

For example:

$ lsof -aUc nm-applet | sudo that-script
COMMAND    PID     USER   FD   TYPE             DEVICE SIZE/OFF  NODE NAME
nm-applet 4183 stephane    4u  unix 0xffff8800a055eb40      0t0 36888 type=STREAM -> 0xffff8800a055e7c0[dbus-daemon,4190,@/tmp/dbus-AiBCXOnuP6]
nm-applet 4183 stephane    7u  unix 0xffff8800a055e440      0t0 36890 type=STREAM -> 0xffff8800a055e0c0[Xorg,3080,@/tmp/.X11-unix/X0]
nm-applet 4183 stephane    8u  unix 0xffff8800a05c1040      0t0 36201 type=STREAM -> 0xffff8800a05c13c0[dbus-daemon,4118,@/tmp/dbus-yxxNr1NkYC]
nm-applet 4183 stephane   11u  unix 0xffff8800a055d080      0t0 36219 type=STREAM -> 0xffff8800a055d400[dbus-daemon,4118,@/tmp/dbus-yxxNr1NkYC]
nm-applet 4183 stephane   12u  unix 0xffff88022e0dfb80      0t0 36221 type=STREAM -> 0xffff88022e0df800[dbus-daemon,2268,/var/run/dbus/system_bus_socket]
nm-applet 4183 stephane   13u  unix 0xffff88022e0f80c0      0t0 37025 type=STREAM -> 0xffff88022e29ec00[dbus-daemon,2268,/var/run/dbus/system_bus_socket]

(if you see 0x0000000000000000 above instead of 0xffff..., it's because the kernel.kptr_restrict sysctl parameter is set on your system which causes kernel pointers to be hidden from non-privileged processes, in which case you'll need to run lsof as root to get a meaningful result).

This script doesn't make any attempt to cope with socket file names with newline characters, but nor does lsof (nor does lsof cope with blanks or colons).

systemtap here is used to dump the address and peer address of all the unix_sock structures in the unix_socket_table hash in the kernel.

Only tested on Linux 3.16 amd64 with systemtap 2.6, and 3.13 with 2.3.

#! /usr/bin/perl
# meant to process lsof output to try and find the peer of a given
# unix domain socket. Needs a working systemtap, lsof, and superuser
# privileges. Copyright Stephane Chazelas 2015, public domain.
# Example: lsof -aUc X | sudo this-script
open STAP, '-|', 'stap', '-e', q{
  probe begin {
    offset = &@cast(0, "struct sock")->__sk_common->skc_node;
    for (i = 0; i < 512; i++) 
      for (p = @var("unix_socket_table@net/unix/af_unix.c")[i]->first;
           p;
           p=@cast(p, "struct hlist_node")->next
          ) {
        sock = p - offset;
        printf("%p %p\n", sock, @cast(sock, "struct unix_sock")->peer);
    }
    exit()
  }
};  
my %peer;
while (<STAP>) {
  chomp;
  my ($a, $b) = split;
  $peer{$a} = $b;
}
close STAP;

my %f, %addr;
open LSOF, '-|', 'lsof', '-nPUFpcfdn';
while (<LSOF>) {
  if (/(.)(.*)/) {
    $f{$1} = $2;
    if ($1 eq 'n') {
      $addr{$f{d}}->{"$f{c},$f{p}" . ($f{n} =~ m{^([@/].*?)( type=\w+)?$} ? ",$1" : "")} = "";
    }
  }
}
close LSOF;

while (<>) {
  chomp;
  for my $addr (/0x[0-9a-f]+/g) {
    my $peer = $peer{$addr};
    if (defined($peer)) {
      $_ .= $peer eq '0x0' ?
            "[LISTENING]" :
            " -> $peer\[" . join("|", keys%{$addr{$peer}}) . "]";
      last;
    }
  }
  print "$_\n";
}
share|improve this answer
    
parse error: unknown statistic operator @var: am I missing something? –  Totor Mar 13 at 21:09
    
@Totor, @var was added in systemtap 1.8, 2012-06-17 (latest is 2.7) –  Stéphane Chazelas Mar 13 at 21:18

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.