On Thu, Jun 19, 2014 at 02:50:18PM +0300, Kirill A. Shutemov wrote:
> > + if (err) {
> > untrack_pfn(vma, pfn, PAGE_ALIGN(size));
> > + if (err != -EBUSY)
> > + zap_page_range_single(vma, addr, size, NULL);
>
> Hm. If I read it correctly, you zap whole range, not only what you've
> set up. Looks wrong.
Yes. I didn't fancy threading the last touched pte back, but that should
be easier if moving to a struct.
> And for after zap, you probably whant to return -EBUSY to caller of
> remap_pfn_range(), not -EINVAL.
No, it has to be EINVAL for my purpose. If we return EBUSY, the caller
will just report VM_NOPAGE back to the fault handler, and the fault will
be retriggered - but the overlapping object will still be present. So the
EINVAL is there to report that the range conflicts with another and that
the caller should abort. It's a nasty semantic that works only when the
concurrent pagefaults are serialised around the call to remap_pfn_range().
The alternative would be to always report EINVAL and clean up, and
export pte_exists() so that the caller can detect when the PTEs have
already been populated by the concurrent fault.
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>