On 2024/4/27 4:33, Jane Chu wrote:
> My apology for the gobbled message earlier.
>
> On 4/26/2024 12:52 PM, Jane Chu wrote:
>> On 4/26/2024 12:05 PM, Matthew Wilcox wrote:
>> [..]
>>> That would be unsafe, the safe way would be if we moved page_folio() after
>>>> the call to __get_hw_poison() in get_any_page() and there would still be one
>>>> remaining user of shake_page() that we can't convert. A safe version of this
>>>> patch would result in a removal of one use of PageHuge() and two uses of
>>>> put_page(), would that be worth submitting?
>>>>
>>>> get_any_page()
>>>> if(__get_hwpoison_page())
>>>> folio = page_folio() /* folio_try_get() returned 1, safe */
>>> I think we should convert __get_hwpoison_page() to return either the folio
>>> or an ERR_PTR or NULL. Also, I think we should delete the "cannot catch
>>> tail" part and just loop in __get_hwpoison_page() until we do catch it.
>>> See try_get_folio() in mm/gup.c for inspiration (although you can't use
>>> it exactly because that code knows that the page is mapped into a page
>>> table, so has a refcount).
>>>
>>> But that's just an immediate assessment; you might find a reason that
>>> doesn't work.
>>
> Besides, in a possible hugetlb demote scenario, it seems to me that we should retry
> get_hwpoison_hugetlb_folio() instead of falling thru to folio_try_get().
>
> static int __get_hwpoison_page(struct page *page, unsigned long flags)
> {
> struct folio *folio = page_folio(page);
> int ret = 0;
> bool hugetlb = false;
>
> ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, false);
> if (hugetlb) {
> /* Make sure hugetlb demotion did not happen from under us. */
> if (folio == page_folio(page))
> return ret;
> if (ret > 0) { <===== still hugetlb, don't fall thru, retry
I tend to agree we should retry get_hwpoison_hugetlb_folio() because folio is still hugetlb in this case.
Below folio_try_get() won't do the right things. This is on my TODO list but as you mentioned this, please
feel free to submit the corresponding patch.
Thanks.
.
