On 2023/7/17 10:53, Matthew Wilcox wrote:
> On Mon, Jul 17, 2023 at 10:33:14AM +0800, Miaohe Lin wrote:
>> On 2023/7/15 11:50, Matthew Wilcox wrote:
>>> On Sat, Jul 15, 2023 at 11:17:26AM +0800, Miaohe Lin wrote:
>>>> Hwpoisoned dirty swap cache page is kept in the swap cache and there's
>>>> simple interception code in do_swap_page() to catch it. But when trying
>>>> to swapoff, unuse_pte() will wrongly install a general sense of "future
>>>> accesses are invalid" swap entry for hwpoisoned swap cache page due to
>>>> unaware of such type of page. The user will receive SIGBUS signal without
>>>> expected BUS_MCEERR_AR payload.
>>>
>>> Have you observed this, or do you just think it's true?
>>>
>>>> +++ b/mm/swapfile.c
>>>> @@ -1767,7 +1767,8 @@ static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
>>>> swp_entry_t swp_entry;
>>>>
>>>> dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
>>>> - if (hwposioned) {
>>>> + /* Hwpoisoned swapcache page is also !PageUptodate. */
>>>> + if (hwposioned || PageHWPoison(page)) {
>>>
>>> This line makes no sense to me. How do we get here with PageHWPoison()
>>> being true and hwposioned being false?
>>
>> hwposioned will be true iff ksm_might_need_to_copy returns -EHWPOISON.
>> And there's PageUptodate check in ksm_might_need_to_copy before we can return -EHWPOISON:
>>
>> ksm_might_need_to_copy
>> if (!PageUptodate(page))
>> return page; /* let do_swap_page report the error */
>> ^^^
>> Will return here because hwpoisoned swapcache page is !PageUptodate(cleared via me_swapcache_dirty()).
>>
>> Or am I miss something?
>
> Ah! So we don't even get to calling copy_mc_to_kernel(). That seems
> like a bug in ksm_might_need_to_copy(), don't you think? Maybe this
I'm sorry but could you please explain what the bug is?
> would be a better fix:
>
> + if (PageHWPoison(page))
> + return ERR_PTR(-EHWPOISON);
Looks reasonable. We can further avoid accessing contents of hwpoisoned swapcache pages which are
PageHWPoison() && PageUptodate() at first place.
Thanks.
> if (!PageUptodate(page))
> return page;
>
> .
>
