Can this code be somewhat unified with reaper kill, instead of copied?

In D39830#906945, @kib wrote:

Can this code be somewhat unified with reaper kill, instead of copied?

Well, I tried that from the beginning, but there are too many differences in the main part (subtree). I can try to unify taskqueue part, is it ok to move it to kern_sig.c?

Note that stopping process could cause spurious EINTRs returned from some syscalls.
Another thing I am curious about, is the overhead of this approach for more common operation killpg(2) then the reap kill. Are there any consumers that would note the difference?

Did you consider other approaches? For instance, we can add a generation counter to the struct pgrp and increment it on the process addition. I am not sure that this could be made into proper fix for all races, but might be worth trying.

In D39830#907696, @kib wrote:

Note that stopping process could cause spurious EINTRs returned from some syscalls.

Thats sad, then a differ approach is needed indeed.

Another thing I am curious about, is the overhead of this approach for more common operation killpg(2) then the reap kill. Are there any consumers that would note the difference?

I think that broadcast case is most important here as it used by jail subsystem which is widely used.

Did you consider other approaches? For instance, we can add a generation counter to the struct pgrp and increment it on the process addition. I am not sure that this could be made into proper fix for all races, but might be worth trying.

Due to broadcast case of killpg1() we need another approach, for now i dont have a good idea. What if we going to mark signals as broadcast in killpg1() and then ‘copy' such signals to the newly created child during dofork()?