kill (0, SIGSTOP) is freezing the system (was: /hurd/init and /hurd/proc)

[Justus Winter]

Hi,

[PATCH 1/7] proc: add proc_mark_essential server code
[PATCH 2/7] hurd: add proc_mark_essential
[PATCH 3/7] init: Mark auth, proc and fs servers as essential

This is a refreshed but otherwise unmodified version of Guillems patch
series presented here:
http://lists.gnu.org/archive/html/bug-hurd/2006-02/msg00081.html

[PATCH 4/7] init: Build fixes

This seems to avoid the need to rebuild the libc.

[PATCH 5/7] proc: Fix miscalculation of count

This was the only issue I found with the patch.

[PATCH 6/7] proc: evil hack, mark all pids<100 essential
[PATCH 7/7] patch omitted

But unfortunately the patch series is only a small part of the
solution. To get a better picture of the problem, killall5 does the
following:

* kill (0, SIGSTOP)

* for all processes, read /proc/x/stat and kill it if it is not
  classified as a kernel thread (or is the current process, or init,
  or is in the same session, something like that).

* kill (0, SIGCONT)

The issues are:

* There are some processes that must not be frozen or the whole system
  is affected. Hardcoding all processes below pid 100 to be essential
  makes the system survive killall5, the processes below 100 that are
  not marked as essential by Guillems patch are:

  exec, /sbin/init, term, pflocal, mach-defpager, null, procfs,
  proxy-defpager, tmpfs, storeio

  I'm not sure about how important it is not to freeze anyone of them,
  but at least procfs must not be stopped b/c killall5 wants to
  iterate over /proc. But it would surely be nice not to freeze
  filesystem translators, and probably the pagers and the term
  translators.

* The question of whether killall5 kills a process or not is not at
  all related to the 'kill (0, x) kills essential processes'
  problem. killall5 parses /proc/*/stat and determines whether a
  process is a kernel thread by looking at the startcode and endcode
  fields.

  /proc/*/stat: startcode and endcode fields are 0, making killall5
  think that the process in question is a kernel thread and is
  therefore not to be killed. So currently, killall5 does not kill any
  processes on Hurd.

* As Markus pointed out in his critique of Guillems patch, any user
  can mark any process as essential, this is probably not what we
  want. Restricting this to root seems easy though.

So the main problem that I see is the question how to determine that a
process is essential in the first place. For some processes it is easy
to mark them after starting them, but for some this might be not
feasible. A lot of important translators are started by the root
filesystem for example.

Thoughts?
Justus