Re: [PATCH v1 0/3] Improve hugetlbfs read on HWPOISON hugepages

[Mike Kravetz <mike.kravetz@oracle.com>]

On 05/18/23 09:10, Jiaqi Yan wrote:
> On Wed, May 17, 2023 at 4:30 PM Mike Kravetz <mike.kravetz@oracle.com> wrote:
> >
> > On 05/17/23 16:09, Jiaqi Yan wrote:
> > > Today when hardware memory is corrupted in a hugetlb hugepage,
> > > kernel leaves the hugepage in pagecache [1]; otherwise future mmap or
> > > read will suject to silent data corruption. This is implemented by
> > > returning -EIO from hugetlb_read_iter immediately if the hugepage has
> > > HWPOISON flag set.
> > >
> > > Since memory_failure already tracks the raw HWPOISON subpages in a
> > > hugepage, a natural improvement is possible: if userspace only asks for
> > > healthy subpages in the pagecache, kernel can return these data.
> >
> > Thanks for putting this together.
> >
> > I recall discussing this some time back, and deciding to wait and see
> > how HGM would progress.  Since it may be some time before HGM goes
> > upstream, it would be reasonable to consider this again.
> 
> This improvement actually does NOT depend on HGM at all. No page table
> related stuff involved here. The other RFC [2] I sent earlier DOES
> require HGM. This improvement was brought up by James when we were
> working on [2]. In "Future Work" section of the cover letter, I
> thought HGM was needed but soon when I code it up, I found I was
> wrong.

Right, this has no HGM dependencies and is actually the only way I can think
of for users to extract some information from a poisoned hugetlb page.

> >
> > One quick question.
> > Do you have an actual use case for this?  It certainly is an improvement
> > over existing functionality.  However, I am not aware of too many (?any?)
> > users actually doing read() calls on hugetlb files.
> 
> I don't have any use case. I did search on Github for around half a
> hour and all the hugetlb usages are done via mmap.
> 

Ok, I was mostly curious as mmap seems to be the most common way of
accessing hugetlb pages.

Even though there is not a known use case today, I think this could be
useful for the reason above: extracting data from a poisoned hugetlb page.
Without HGM this is the only way to extract such data.

Unfortunately, read() is not an option for sysV shared memory or private
mappings.  HGM would help there.
-- 
Mike Kravetz