On Mon, Nov 29, 2021 at 01:53:01PM -0800, Linus Torvalds wrote:
> On Mon, Nov 29, 2021 at 12:56 PM Catalin Marinas
> <catalin.marinas@xxxxxxx> wrote:
> > I think that would be useful, though it doesn't solve the potential
> > livelock with sub-page faults.
>
> I was assuming we'd just do the sub-page faults.
>
> In fact, I was assuming we'd basically just replace all the PAGE_ALIGN
> and PAGE_SIZE with SUBPAGE_{ALIGN,SIZE}, together with something like
>
> if (size > PAGE_SIZE)
> size = PAGE_SIZE;
>
> to limit that size thing (or possibly make that "min size" be a
> parameter, so that people who have things like that "I need at least
> this initial structure to be copied" issue can document their minimum
> size needs).
Ah, so fault_in_writeable() would never fault in the whole range (if too
large). When copy_to_user() goes beyond the faulted in range, it may
fail and we go back to fault in a bit more of the range. A copy loop
would be equivalent to:
fault_addr = ubuf;
end = ubuf + size;
while (1) {
if (fault_in_writeable(fault_addr,
min(PAGE_SIZE, end - fault_addr)))
break;
left = copy_to_user(ubuf, kbuf, size);
if (!left)
break;
fault_addr = end - left;
}
That should work. I'll think about it tomorrow, getting late over here.
(I may still keep the sub-page probing in the arch code, see my earlier
exchanges with Andreas)
--
Catalin
