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

[Oleg Nesterov]

On 04/23, Peter Zijlstra wrote:
>
> On Mon, 2012-04-23 at 12:54 +0530, Srikar Dronamraju wrote:
> > * Peter Zijlstra <peterz@xxxxxxxxxxxxx> [2012-04-23 09:14:00]:
> >
> > > On Fri, 2012-04-20 at 20:37 +0200, Oleg Nesterov wrote:
> > > > Say, a user wants to probe /sbin/init only. What if init forks?
> > > > We should remove breakpoints from child->mm somehow.
> > >
> > > How is that hard? dup_mmap() only copies the VMAs, this doesn't actually
> > > copy the breakpoint. So the child doesn't have a breakpoint to be
> > > removed.
> > >
> >
> > Because the pages are COWED, the breakpoint gets copied over to the
> > child. If we dont want the breakpoints to be not visible to the child,
> > then we would have to remove them explicitly based on the filter (i.e if
> > and if we had inserted breakpoints conditionally based on filter).
>
> I thought we didn't COW shared maps since the fault handler will fill in
> the pages right and only anon stuff gets copied.

Confused...

Do you mean the "Don't copy ptes where a page fault will fill them correctly"
check in copy_page_range() ? Yes, but this vma should have ->anon_vma != NULL
if it has the breakpoint installed by uprobes.

Yes, we do not COW this page during dup_mmap(), but the new child's pte
should point to the same page with bp.

OK, I guess I misunderstood.

> > Once we add the conditional breakpoint insertion (which is tricky),
>
> How so?

I agree with Srikar this doesn't look simple to me. First of all,
currently it is not easy to find the tasks which use this ->mm.
OK, we can simply do for_each_process() under tasklist, but this is
not very nice.

But again, to me this is not the main problem.

> > Conditional removal
> > of breakpoints in fork path would just be an extension of the
> > conditional breakpoint insertion.
>
> Right, I don't think that removal is particularly hard if needed.

I agree that remove_breakpoint() itself is not that hard, probably.

But the whole idea of filtering is not clear to me. I mean, when/how
we should call the filter, and what should be the argument.
task_struct? Probably, but I am not sure.

And btw fork()->dup_mmap() should call the filter too. Suppose that
uprobe_consumer wants to trace the task T and its children, this looks
very natural.

And we need to rework uprobe_register(). It can't simply return if
this (inode, offset) already has the consumer.

So far I think this needs more thinking. And imho we should merge the
working code Srikar already has, then try to add this (agreed, very
important) optimization.

Oleg.

--
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/