Messages in this thread | | | From | "Wang, Haiyue" <> | Subject | RE: [PATCH v1] mm: migration: fix the FOLL_GET failure on following huge page | Date | Sun, 14 Aug 2022 06:20:13 +0000 |
| |
> -----Original Message----- > From: Andrew Morton <akpm@linux-foundation.org> > Sent: Sunday, August 14, 2022 07:29 > To: Wang, Haiyue <haiyue.wang@intel.com> > Cc: linux-mm@kvack.org; linux-kernel@vger.kernel.org; david@redhat.com; linmiaohe@huawei.com; Huang, > Ying <ying.huang@intel.com>; songmuchun@bytedance.com; naoya.horiguchi@linux.dev > Subject: Re: [PATCH v1] mm: migration: fix the FOLL_GET failure on following huge page > > On Fri, 12 Aug 2022 16:49:21 +0800 Haiyue Wang <haiyue.wang@intel.com> wrote: > > > Not all huge page APIs support FOLL_GET option, so the __NR_move_pages > > will fail to get the page node information for huge page. > > Which ones need fixing?
1. 'follow_huge_pud' arch/s390/mm/hugetlbpage.c
2. 'follow_huge_addr' arch/ia64/mm/hugetlbpage.c
3. 'follow_huge_pgd' mm/hugetlb.c
And I found that only 'pud' and 'pmd' need to check 'is_vm_hugetlb_page' like: pud_huge(*pud) && is_vm_hugetlb_page(vma) pmd_huge(pmdval) && is_vm_hugetlb_page(vma)
So I'm not sure whether my patch can cover 2 & 3 for other huge page use cases except by hugetlbfs.
> > What are the user-visible runtime effects of this bug? >
In my test, the '__NR_move_pages' system call will return '-2' for 1GB huge page memory map when dump the page node information. [Test on linux-5.19 stable]
> Is a -stable backport warranted?
Yes.
Since the mainline has introduced the new patch:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=3218f8712d6bb
The backported needs to rebase like for 5.19:
- if (page && !is_zone_device_page(page)) { + if (page) {
> > > This is an temporary solution to mitigate the racing fix. > > > > After supporting follow huge page by FOLL_GET is done, this fix can be > > reverted safely. > >
| |