On Thu, May 23, 2024 at 05:08:29AM +0200, Oscar Salvador wrote:
> On Wed, May 22, 2024 at 05:46:09PM -0400, Peter Xu wrote:
> > > Now, ProcessB still has the page mapped, so upon re-accessing it,
> > > it will trigger a new MCE event. memory-failure code will see that this
> >
> > The question is why accessing that hwpoison entry from ProcB triggers an
> > MCE. Logically that's a swap entry and it should generate a page fault
> > rather than MCE. Then in the pgfault hanlder we don't need that encoded
> > pfn as we have vmf->address.
>
> It would be a swap entry if we reach try_to_umap_one() without trouble.
> Then we have the code that converts it:
>
> ...
> if (PageHWPoison(p))
> pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
> set_{huge_}pte_at
> ...
>
> But maybe we could only do that for ProcA, while ProcB failed to do that,
> which means that for ProcA that is a hwpoisoned-swap-entry, but ProcB still
> has this page mapped as usual, so if ProcB re-access it, that will not
> trigger a fault (because the page is still mapped in its pagetables).
But in that case "whether encode pfn in hwpoison swap entry" doesn't matter
either.. as it's not yet converted to a swap entry, so the pfn is there.
Thanks,
--
Peter Xu
