bug#61156: ‘guix container exec’ does not actually change PID namespaces

[Ludovic Courtès]

Currently, when a Guix program runs in separate user, mount, and PID
namespaces (for example via (guix least-authority)), ‘guix container
exec’ fails badly:

  guix container exec 10652 
/gnu/store/720rj90bch716isd8z7lcwrnvz28ap4y-bash-static-5.1.8/bin/sh
  guix container: error: process terminated with signal 11

or, similarly:

  nsenter --preserve-credentials -U -m  -t 10652 -m -U -p -F 
/gnu/store/720rj90bch716isd8z7lcwrnvz28ap4y-bash-static-5.1.8/bin/sh
  Segmentation fault

Stracing reveals that the child process segfaults immediately after
attempting to read /proc/self/exe:

  14111 readlink("/proc/self/exe", 0x7ffccefa29c0, 4096) = -1 ENOENT (No such 
file or directory)
  14111 --- SIGSEGV {si_signo=SIGSEGV, si_code=SEGV_MAPERR, 
si_addr=0xffffffffffffffff} ---

The segfault is due to <https://issues.guix.gnu.org/52671>.

But why isn’t /proc visible in the first place?  It *is* definitely
mounted within that process’s namespace, as confirmed here:

  $ ls -ld /proc/10652/root/proc
  dr-xr-xr-x 326 root root 0 Jan 29 21:55 /proc/10652/root/proc/

The reason is that calling setns(2) on a PID namespace “changes only the
PID namespace that subsequently created child processes of the caller
will be placed in; it does not change the PID namespace of the caller
itself.”

This is why removing ‘-F’ in the ‘nsenter’ command line above solves the
problem.

Conclusion: ’container-excursion’ should fork so that the PID namespace
change takes effect.

Ludo’.