Re: [PATCH] Add a text_poke syscall v2

[Oleg Nesterov <oleg@redhat.com>]

On 11/29, H. Peter Anvin wrote:
>
> On 11/29/2013 12:05 PM, Oleg Nesterov wrote:
> >
> > Can't we invalidate pte (so that any user will stuck in page fault),
> > update the page(s), restore the pte and drop the locks?
> >
> > This way sys_text_poke() won't be x86-specific, and it will be per-mm.
> >
>
> Hmmm... if we hold mmap_sem() this pretty much will be the net result,
> no?

Yes, down_write(mmap_sem) is enough to block the page faults.

But we need pte-lock anyway, to avoid the races with, say, try_to_unmap().
(and of course, we need to retry if page_check_address() fails).

This actually means that if we want to update a single page we could
use down_read(). But in general we need to update 2 pages. Or even more
if we generalize sys_text_poke(), perhaps it should be renamed in this
case, but this is off-topic.

> That would mean no additional tests needed on the page fault path.

Not sure I understand... but of course we should not change the fault
paths in any case.

>  What I'm not sure of is whether or not it is actually safe to hold
> mmap_sem across all the code we need

Let me repeat, I do not understand vm enough to answer authoritatively.

But I think this should be safe. I do not see why it should not, but
see above.

Oleg.