Linux – How to find out the file offset of an opened file

If you attach strace to the process just when it's hung (you can get the pid and queue the command up in advance, in a spare terminal), it'll show the file descriptor of the blocking write.

Trivial example:

$ mkfifo tmp
$ cat /dev/urandom > tmp &
[1] 636226
  # this will block on open until someone opens for reading

$ exec 4<tmp
  # now it should be blocked trying to write

$ strace -p 636226
Process 636226 attached - interrupt to quit
write(1, "L!\f\335\330\27\374\360\212\244c\326\0\356j\374`\310C\30Z\362W\307\365Rv\244?o\225N"..., 4096 <unfinished ...>
^C
Process 636226 detached

I'll try and answer both this and your earlier question as they are related.

The doors to namespaces are files in /proc/*/ns/* and /proc/*/task/*/ns/*.

A namespace is created by a process unsharing its namespace. A namespace can then be made permanent by bind-mounting the ns file to some other place.

That's what ip netns does for instance for net namespaces. It unshares its net namespace and bind-mounts /proc/self/ns/net to /run/netns/netns-name.

In a /proc mounted in the root pid namespace, you can list all the namespaces that have a process in them by doing:

# readlink /proc/*/task/*/ns/* | sort -u
ipc:[4026531839]
mnt:[4026531840]
mnt:[4026531856]
mnt:[4026532469]
net:[4026531956]
net:[4026532375]
pid:[4026531836]
pid:[4026532373]
uts:[4026531838]

The number in square brackets is the inode number.

To get that for a given process:

# ls -Li /proc/1/ns/pid
4026531836 /proc/1/ns/pid

Now, there may be permanent namespaces that don't have any process in them. Finding them out can be a lot trickier AFAICT.

First, you have to bear in mind that there can be several mount namespaces.

# awk '$9 == "proc" {print FILENAME,$0}' /proc/*/task/*/mountinfo | sort -k2 -u
/proc/1070/task/1070/mountinfo 15 19 0:3 / /proc rw,nosuid,nodev,noexec,relatime - proc proc rw
/proc/19877/task/19877/mountinfo 50 49 0:3 / /run/netns/a rw,nosuid,nodev,noexec,relatime shared:2 - proc proc rw
/proc/19877/task/19877/mountinfo 57 40 0:3 / /proc rw,nosuid,nodev,noexec,relatime - proc proc rw
/proc/1070/task/1070/mountinfo 66 39 0:3 / /run/netns/a rw,nosuid,nodev,noexec,relatime shared:2 - proc proc rw
/proc/19877/task/19877/mountinfo 68 67 0:3 / /mnt/1/a rw,nosuid,nodev,noexec,relatime unbindable - proc proc rw

Those /mnt/1/a, /run/netns/a may be namespace files.

We can get an inode number:

# nsenter --mount=/proc/19877/task/19877/ns/mnt -- ls -Li /mnt/1/a
4026532471 /mnt/1/a

But that doesn't tell us much other than it's not in the list computed above.

We can try and enter it as any of the different types:

# nsenter --mount=/proc/19877/task/19877/ns/mnt -- nsenter --pid=/mnt/1/a true
nsenter: reassociate to namespace 'ns/pid' failed: Invalid argument
# nsenter --mount=/proc/19877/task/19877/ns/mnt -- nsenter --mount=/mnt/1/a true
nsenter: reassociate to namespace 'ns/mnt' failed: Invalid argument
# nsenter --mount=/proc/19877/task/19877/ns/mnt -- nsenter --net=/mnt/1/a true
#

OK, that was a net namespace file.

So it would seem we have a method to list the name spaces: list the ns directories of all the tasks, then find all the proc mountpoints in all the /proc/*/task/*/mountinfo and figure out their type by trying to enter them.