Re: [PATCH v2] mm/swapfile: unuse_pte can map random data if swap read fails

[Miaohe Lin <linmiaohe@huawei.com>]

On 2022/4/19 20:12, David Hildenbrand wrote:
> On 19.04.22 14:00, Miaohe Lin wrote:
>> On 2022/4/19 19:46, David Hildenbrand wrote:
>> ...
>>>> Do you mean that we should set the pfn to 0 for the hwpoison marker so that we can
>>>> distinguish swapin error case from real hwpoison case?
>>>
>>> I am not sure if we really have to distinguish. However, "0" seems to
>>> make sense to indicate "this is not an actual problematic PFN, the
>>> information is simply no longer around due to a hardware issue.
>>>
>>
>> IMHO, we have to distinguish. For example, we might need to return VM_FAULT_SIGBUS
>> instead of VM_FAULT_HWPOISON when user accesses the error page. Or should we simply
>> return VM_FAULT_HWPOISON to simplify the handling?
> 
> Hm, you're right. In e.g., x86 do_sigbus() we would send an BUS_MCEERR_AR.
> 
> So yes, if we reuse is_hwpoison_entry() we'd have to convert to either
> VM_FAULT_HWPOISON or VM_FAULT_SIGBUS.
> 
> Something like "is_error_entry()" that can further be refined to
> hwpoison or swapin could make sense. But what you have here is straight
> forward to me as well. Whatever you/others prefer.

IMHO, I prefer to use the separated swapin error maker because reusing hwpoison entry
might make things more complicated and looks somewhat obscure.

> 
> Acked-by: David Hildenbrand <david@redhat.com>

Many thanks for your Acked-by tag!

> 
> 
> NIT: I'd make the terminology make_swapin_error_entry() consistent with
> SWAP_READ_ERROR and especially existing SWP_.
> 
> For example, calling the latter SWP_SWAPIN_ERROR

This looks better. Thanks! Will change to rename the relevant terminology in next
version if no one has an objection. :)

>