Re: [PATCH] x86/mce: Set PG_hwpoison page flag to avoid the capture kernel panic

[Zhiquan Li]

On 2023/1/31 12:50, HORIGUCHI NAOYA(堀口 直也) wrote:
> On Mon, Jan 30, 2023 at 07:21:15PM +0000, Luck, Tony wrote:
>>> Although I'm not sure that SGX memory or PMEM pages are expected to be
>>> included in kdump, but simply setting PageHWPoison does not work for them?
>>> (Maybe that depends on how kdump handles these types of memory.)
>>
>> SGX/TDX pages can't be dumped. They are encrypted with no way for kdump to
>> get the key.
>>
>> PMEM seems pointless (but I don't know what kdump does here).
>>
>>> As for HugeTLB, kdump utility should parse the struct page and be aware of
>>> HugeTLB pages, so maybe setting PageHWPoison on the head page could work.
>>
>> Or maybe kdump can take not of the PageHWPoison flag on the sub-page of the
>> huge page? It depends on whether there is any benefit to the dump to include the
>> not-poisoned parts of a huge page.
> 
> I think that many kdump users filter out HugeTLB pages (setting dump_level
> to filter "User pages") to reduce the size of kdump.  User pages are not
> much helpful to investigate kernel problems, so filtering all sub-pages in
> hwpoisoned hugepage seems to me not so harmful.
> 
> I don't say that saving healthy subpages has no benefit, but I don't know
> much about usecases where user pages in kdump file help.
> 

Many thanks, Naoya and Tony.

The "TODO" comment was initially added by me. I referenced
memory_failure() that gets rid of the three cases specifically, so I'd
like the cases are fully discussed here to make sure it is a strong fix.

- SGX, as Tony said those pages can't be dumped.
- PMEM, although the memory region is figured out by kexec, kdump
utility doesn't have corresponding implementation for it.  So we can
ignore it.
- HugeTLB
  As we known there are complicated processes for HugeTLB pages at
memory_failure(). However, it doesn't make sense for a routine going to
call panic().  Kernel marks the page hwpoisoned would be enough.

The information is sufficient for kdump to make decision on how to deal
with a hwpoisoned huge pages, even though setting the PageHWPoison flag
on the sub-page of a huge page.  It's kdump's responsibility for the
not-poisoned parts of a huge page.  Currently the implementation is to
exclude the whole huge page simply:

-
https://github.com/makedumpfile/makedumpfile/commit/e8b4f93b3260defe86f5e13ca7536c07f2e32914
- https://bugzilla.redhat.com/show_bug.cgi?id=1181649

Best Regards,
Zhiquan