Re: [PATCH v3 0/10] Uprobes v3

[Srikar Dronamraju <srikar@linux.vnet.ibm.com>]

> 
> Hrmm, you could do something like the above and use write_protect_page()
> along with that lock_page() you do to sync against concurrent faults. It
> won't be the fastest code around but it should work I guess.

Okay, thanks for the code.

> 
> Something like try_to_merge_one_page() in the KSM code:
> 
> write_opcode(struct mm_struct *mm, unsigned long address, void *insn,
> int len)
> {
> 	down_read(mm->mmap_sem);

[ snipped]
> 
> 	kunmap_atomic(vaddr_new);
> 	kunmap_atomic(vaddr_old);
> 
> 	/* XXX: frob mlock state */

I think you are referring to mlock_vma_page() and munlock_vma_page that
we have in try_to_merge_one_page(). However mlock_vma_page and
munlock_vma_page are not exposed outside mm as they are declared inside
mm/internal.h. So do you suggest moving the declarations out of
mm/internal.h or have a wrapper function defined in mm/ that could be
used by both KSM and uprobes.

> 
> 	/* flip pages, do_wp_page() will fail pte_same() and bail */
> 	err = replace_page(vma, new_page, old_page, orig_pte)
> 	if (err) goto fail_more;
> 
> 	/* we're done, let 'em rip */
> 	unlock_page(old_page);
> 	put_page(old_page);
> 
> 	return;
> 
> fail:
> 	/* XXX: fixup stuff */
> }
> 
> Note that this can fail for fun reasons like O_DIRECT IO, but that
> should be very rare indeed on text pages.
>