Re: [RFC 0/6] uprobes: kill uprobes_srcu/uprobe_srcu_id

[Peter Zijlstra <peterz@infradead.org>]

On Fri, 2012-04-06 at 00:20 +0200, Oleg Nesterov wrote:
> Hello.
> 
> Not for inclusion yet, only for the early review.
> 
> I didn't even try to test these changes, and I am not expert
> in this area. And even _if_ this code is correct, I need to
> re-split these changes anyway, update the changelogs, etc.
> 
> Questions:
> 
> 	- does it make sense?

Maybe, upside is reclaiming that int from task_struct, downside is that
down_write :/ It would be very good not to have to do that. Nor do I
really see how that works.

> 	- can it work or I missed something "in general" ?

So we insert in the rb-tree before we take mmap_sem, this means we can
hit a non-uprobe int3 and still find a uprobe there, no?

> Why:
> 
> 	- It would be nice to remove a member from task_struct.
> 
> 	- Afaics, the usage of uprobes_srcu does not look right,
> 	  at least in theory, see 6/6.
> 
> 	  The comment above delete_uprobe() says:
> 
> 	  	The current unregistering thread waits till all
> 	  	other threads have hit a breakpoint, to acquire
> 	  	the uprobes_treelock before the uprobe is removed
> 	  	from the rbtree.
> 
> 	  but synchronize_srcu() can only help if a thread which
> 	  have hit the breakpoint has already called srcu_read_lock().
> 	  It can't synchronize with read_lock "in future", and there
> 	  is a small window.
> 
> 	  We could probably add another synchronize_sched() before
> 	  synchronize_srcu(), but this doesn't look very nice and

Right, I think that all was written with the assumption that sync_srcu
implied a sync_rcu, which of course we've recently wrecked.