Re: [PATCH?] uprobes: change uprobe_write_opcode() to modify the page directly

[Oleg Nesterov <oleg@redhat.com>]

On 12/03, H. Peter Anvin wrote:
>
> On 12/03/2013 02:01 PM, Linus Torvalds wrote:
> > On Tue, Dec 3, 2013 at 12:54 PM, Oleg Nesterov <oleg@redhat.com> wrote:
> >>
> >> So do you think the patch I sent is wrong? Why?
> >
> > I think the TLB shootdown should guarantee that it's ok on other
> > CPU's, since that's basically what we do on mmap.
> >
>
> I think that is true for other CPUs; however, there are definitely CPUs
> out there (which Linux supports) for which you have to synchronize the I
> and D sides "manually" after writing code through memory, at least
> through the CPU.  That is at least one reason why MIPS has a
> cacheflush() system call, for example.

OK, probably (with or without the patch I sent) uprobe_write_opcode() needs
flush_icache_page(). Altough it is nop on x86 and powerpc (architectures
we currently support).

But I still can't understand your "There is no architecture-independent
way to make code globally visible". If this is true, then how, say,
do_swap_page() can work?

So I still think the patch should work (I'll add flush_icache_page).

> > But looking closer at this, I think I see why the old code did what it
> > did. I think it's breaking shared mmap pages on purpose rather than
> > dirtying them. Which is probably the right thing to do.
>
> In other words, treating them as MAP_PRIVATE?  Wouldn't it be better to
> throw an error if we can't honor the semantics of the mapping that we
> are using?

Yes, uprobes never writes to MAP_SHARED vmas.

Oleg.