[Devel] [RFC][PATCH] Add child reaper to struct pspace

[Kirill Korotaev]

Eric W. Biederman wrote:
> Kirill Korotaev <dev at sw.ru> writes:
> 
> 
> 
>>I guess there will be a need of list of tasks... not of pids only...
>>many of loops like do_each_thread()/while_each_thread() has nothing to do with
>>pids
>>and should be narrowed down to loop through the container.
>>
>>Does this logic belong to pid_ns? if yes, then it definetely should be called
>>task_ns.
> 
> 
> Just skimming through I see one or two instances.  Where the existing
> loop uses do_each_thread()/while_each_thread() that we need to change.
> 
> kernel/capabilities.c cap_set_all() is an example.
> 
> However what we are trying to achieve there is to iterate through
> the same list that kill(-1, ) uses.   So we need to replace 
> do_each_thread()/while_each_thread() with something that will
> iterate through everything in the pid namespace.
> 
> Most instances of do_each_thread()/while_each_thread() the kernel
> really does need a global view, and need to be left unchanged.
> 
> Basically the current kernel is short the concept of a process
> group of all processes, and uses the concept of a list of all processes
> instead.  
> 
> Since the two concepts of a list of all tasks, and a list of all processes
> I can see diverge when we have multiple pid namespaces we need to add
> an additional concept, in the kernel.
> 
> Do you know an example in that we need to change to implement a pid
> namespace that goes beyond iterating through the list of processes
> that kill(-1,) uses?

from OVZ patches:

do_each_thread_ve()
  elf_core_dump() (need pid namespace list?)
  zap_threads (need pid namespace list?)
  chroot_fs_refs
  cap_set_all (need pid namespace list?)
  cpt functions (need to freeze VE processes, pid namespace list?)
  sys_setpriority (needs task list for user namespace!)
  sys_getpriority (the same)
  sys_ioprio_set (the same)
  sys_ioprio_get (the same)
  selinux_setprocattr (should be changed with the check for thread_group_empty()???)

for_each_process_ve()
  asids_proc_info (need pid namespace list? in host should print all?)
  kill_something_info (I suppose you changed it already?)

some of these are optimizations which are natural for containers and are good
for scalability (as zap_threads, elf_core_dump etc.).

Thanks,
Kirill
P.S. Sorry for not always replying quickly...