[PATCH 0/3] makedumpfile: hugepage filtering for vmcore dump

[HATAYAMA Daisuke]

(2013/12/03 17:05), Atsushi Kumagai wrote:
> On 2013/11/29 13:57:21, kexec <kexec-bounces at lists.infradead.org> wrote:
>> (2013/11/29 13:23), Atsushi Kumagai wrote:
>>> On 2013/11/29 12:24:45, kexec <kexec-bounces at lists.infradead.org> wrote:
>>>> (2013/11/29 12:02), Atsushi Kumagai wrote:
>>>>> On 2013/11/28 16:50:21, kexec <kexec-bounces at lists.infradead.org> wrote:
>>>>>>>> ping, in case you overlooked this...
>>>>>>>
>>>>>>> Sorry for the delayed response, I prioritize the release of v1.5.5 now.
>>>>>>>
>>>>>>> Thanks for your advice, check_cyclic_buffer_overrun() should be fixed
>>>>>>> as you said. In addition, I'm considering other way to address such case,
>>>>>>> that is to bring the number of "overflowed pages" to the next cycle and
>>>>>>> exclude them at the top of __exclude_unnecessary_pages() like below:
>>>>>>>
>>>>>>>                    /*
>>>>>>>                     * The pages which should be excluded still remain.
>>>>>>>                     */
>>>>>>>                    if (remainder >= 1) {
>>>>>>>                            int i;
>>>>>>>                            unsigned long tmp;
>>>>>>>                            for (i = 0; i < remainder; ++i) {
>>>>>>>                                    if (clear_bit_on_2nd_bitmap_for_kernel(pfn + i)) {
>>>>>>>                                            pfn_user++;
>>>>>>>                                            tmp++;
>>>>>>>                                    }
>>>>>>>                            }
>>>>>>>                            pfn += tmp;
>>>>>>>                            remainder -= tmp;
>>>>>>>                            mem_map += (tmp - 1) * SIZE(page);
>>>>>>>                            continue;
>>>>>>>                    }
>>>>>>>
>>>>>>> If this way works well, then aligning info->buf_size_cyclic will be
>>>>>>> unnecessary.
>>>>>>>
>>>>>>
>>>>>> I selected the current implementation of changing cyclic buffer size becuase
>>>>>> I thought it was simpler than carrying over remaining filtered pages to next cycle
>>>>>> in that there was no need to add extra code in filtering processing.
>>>>>>
>>>>>> I guess the reason why you think this is better now is how to detect maximum order of
>>>>>> huge page is hard in some way, right?
>>>>>
>>>>> The maximum order will be gotten from HUGETLB_PAGE_ORDER or HPAGE_PMD_ORDER,
>>>>> so I don't say it's hard. However, the carrying over method doesn't depend on
>>>>> such kernel symbols, so I think it's robuster.
>>>>>
>>>>
>>>> Then, it's better to remove check_cyclic_buffer_overrun() and rewrite part of free page
>>>> filtering in __exclude_unnecessary_pages(). Could you do that too?
>>>
>>> Sure, I'll modify it too.
>>>
>>
>> This is a suggestion from different point of view...
>>
>> In general, data on crash dump can be corrupted. Thus, order contained in a page
>> descriptor can also be corrupted. For example, if the corrupted value were a huge
>> number, wide range of pages after buddy page would be filtered falsely.
>>
>> So, actually we should sanity check data in crash dump before using them for application
>> level feature. I've picked up order contained in page descriptor, so there would be other
>> data used in makedumpfile that are not checked.
> 
> What you said is reasonable, but how will you do such sanity check ?
> Certain standard values are necessary for sanity check, how will
> you prepare such values ?
> (Get them from kernel source and hard-code them in makedumpfile ?)
> 
>> Unlike diskdump, we no longer need to care about kernel/hardware level data integrity
>> outside of user-land, but we still care about data its own integrity.
>>
>> On the other hand, if we do it, we might face some difficulty, for example, hardness of
>> maintenance or performance bottleneck; it might be the reason why we don't see sanity
>> check in makedumpfile now.
> 
> There are many values which should be checked, e.g. page.flags, page._count,
> page.mapping, list_head.next and so on.
> If we introduce sanity check for them, the issues you mentioned will be appear
> distinctly.
> 
> So I think makedumpfile has to trust crash dump in practice.
> 

Yes, I don't mean such very drastic checking; I understand hardness because I often
handle/write this kind of code; I don't want to fight tremendously many dependencies...

So we need to concentrate on things that can affect makedumpfile's behavior significantly,
e.g. infinite loop caused by broken linked list objects, buffer overrun cauesd by large values
from broken data, etc. We should be able to deal with them by carefully handling
dump data against makedumpfile's runtime data structure, e.g., buffer size.

Is it OK to consider this is a policy of makedumpfile for data corruption?

-- 
Thanks.
HATAYAMA, Daisuke