Re: [PATCH -next v1 3/3] kernel/events/uprobes: uprobe_write_opcode() rewrite

[David Hildenbrand]

On 10.03.25 18:03, Oleg Nesterov wrote:
> On 03/04, David Hildenbrand wrote:
> > uprobe_write_opcode() does some pretty low-level things that really, it
> > shouldn't be doing:
>
> Agreed. Thanks again for doing this.
>
> David, as I said, I can't review. I don't understand this mm/folio magic
> with or without your changes.

No worries! Thanks for taking a look!

> However. With your changes the code looks "better" and more understandable
> to me. So I'd vote for your patches even if I can't ack them.
>
> But I'd like to ask some stupid (no, really) questions.
> __uprobe_write_opcode() does:
>
> 	/* We're done if we don't find an anonymous folio when unregistering. */
> 	if (!folio_test_anon(folio))
> 		return is_register ? -EFAULT : 0;
>
> Yes, but we do not expect !folio_test_anon() if register == true, right?
> See also below.>
> 	/* Verify that the page content is still as expected. */
> 	if (verify_opcode(fw->page, opcode_vaddr, &opcode) <= 0) {
> 		set_pte_at(vma->vm_mm, vaddr, fw->ptep, fw->pte);
> 		return -EAGAIN;
> 	}
>
> The caller, uprobe_write_opcode(), has already called verify_opcode(),
> why do we need to re-check?

Regarding both questions, the code is fairly racy. Nothing would stop 
user space from (a) modifying that memory (b) zapping the anon page 
using MADV_DONTNEED (if we don't hold the mmap lock in write mode).

Regarding the latter, uprobe_write_opcode() is documented to: "Called 
with mm->mmap_lock held for read or write.".

Note that both checks are fairly cheap.

> But whatever reason we have. Can we change uprobe_write_opcode() to
> "delay" put_page() and instead of

I was debating with myself whether we should do that and went back and 
forth a couple of times.

> 	/* Walk the page tables again, to perform the actual update. */
> 	folio = folio_walk_start(&fw, vma, vaddr, 0);
> 	if (folio) {
> 		ret = __uprobe_write_opcode(vma, &fw, folio, opcode_vaddr,
> 					    opcode);
> 		folio_walk_end(&fw, vma);
> 	} else {
> 		ret = -EAGAIN;
> 	}
>
> do something like
>
> 	/* Walk the page tables again, to perform the actual update. */
> 	ret = -EAGAIN;
> 	folio = folio_walk_start(&fw, vma, vaddr, 0);
> 	if (folio) {
> 		if (fw.page == page) {
> 			WARN_ON(is_register && !folio_test_anon(folio));

Yes, that would work (we could leave the WARN_ON in 
__uprobe_write_opcode), but I am not sure if the end result is better 
better. No strong opinion on the details though.

> 			ret = __uprobe_write_opcode(vma, &fw, folio, opcode_vaddr,
> 					            opcode);
> 		}
> 		folio_walk_end(&fw, vma);
> 	}
>
> ?
>
> Once again, I am not trying to review. I am trying to understand the
> basics of your code.

Any feedback is welcome, thanks!

--
Cheers,

David / dhildenb