Re: [RFC][PATCH] ps command race fix take2 [1/4] list token

[ebiederm@xmission.com (Eric W. Biederman)]

KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> writes:

> This is ps command race fix take2. Unfortunately, against 2.6.18-rc4.
> I'll rebase this to appropriate kernel if O.K. (I think this is RFC)
>
> This patch implements Paul Jackson's idea, 'inserting false link in task list'.

Currently the tasklist_lock is one of the more highly contended locks in
the kernel.  Adding an extra place it is taken is undesirable.
If could see a better algorithm for sending a signal to all processes
in a process groups we could remove the tasklist_lock entirely.

In addition you only solves half the readdir problems.  You don't solve
the seek problem which is returning to an offset you had been to
before.  A relatively rare case but...

> Good point of this approach is cost of searching task is O(N) (N=num of tgids).
> Bad point is lock and kmalloc/kfree.
> I didin't modified thread_list and cpuset's proc list, maybe future work.
>
> If searching pid bitmap is better, please take Erics.

My patch at least needs a good changelog but I believe it will work
better and can be further improved with a better pid data structure
if there is actually a problem there.  Given that I don't take
any locks it should be much friendlier at scale, and the code
was simpler.

However I will miss a few newly forked processes and I don't think your
technique will miss any.  Still neither will miss a process that
existed the entire time.

If nothing else I think it was worth posting so we could contrast the two.

Eric