Re: [PATCH 2/2] CRED: Fix __task_cred()'s lockdep check and banner comment
From: Linus Torvalds
Date: Thu Aug 05 2010 - 12:23:32 EST
On Thu, Aug 5, 2010 at 12:19 AM, Eric W. Biederman
> No. When we send a signal to multiple processes it needs to be an
> atomic operation so that kill -KILL -pgrp won't let processes escape.
> It is what posix specifies, it is what real programs expect, and it
> is the useful semantic in userspace.
Ok. However, in that case, it's not really about the whole list
traversal, it's a totally separate thing, and it's really sad that we
end up using the (rather hot) tasklist_lock for something like that.
With the dcache/inode locks basically going away, I think
tasklist_lock ends up being one of the few hot locks left.
Wouldn't it be much nicer to:
- make it clear that all the "real" signal locking can rely on RCU
- use a separate per-pgrp lock that ends up being the one that gives
the signal _semantic_ meaning?
That would automatically document why we get the lock too, which
certainly isn't clear from the code as-is.
The per-pgrp lock might be something as simple as a silly hash that
just spreads out the process groups over some random number of simple
> With the tasklist_lock the rule in these functions is that the caller
> will take the lock, so we probably make the rule the caller should
> take the lock in the same scenarios for the rcu_read_lock(). Aka just
> everywhere, that today we say just:
I agree that we probably should have done that originally, in order to
not have these bugs show up later. However, I don't think it makes
sense any more, especially not if tasklist_lock isn't even a "real"
lock from a kernel internal consistency standpoint, but has a totally
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/