On 2024/9/6 10:42, Kefeng Wang wrote:
> Similar to other poison recovery, use copy_mc_user_highpage() to
> avoid potentially kernel panic during copy page in copy_present_page()
> from fork, once copy failed due to hwpoison in source page, we need
> to break out of copy in copy_pte_range() and release prealloc folio,
> so copy_mc_user_highpage() is moved ahead before set *prealloc to NULL.
>
> Signed-off-by: Kefeng Wang <wangkefeng.wang@xxxxxxxxxx>
> ---
> mm/memory.c | 10 +++++++---
> 1 file changed, 7 insertions(+), 3 deletions(-)
>
> diff --git a/mm/memory.c b/mm/memory.c
> index d310c073a1b3..6e7b78e49d1a 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -926,8 +926,11 @@ copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma
> * We have a prealloc page, all good! Take it
> * over and copy the page & arm it.
> */
> +
> + if (copy_mc_user_highpage(&new_folio->page, page, addr, src_vma))
> + return -EHWPOISON;
> +
> *prealloc = NULL;
> - copy_user_highpage(&new_folio->page, page, addr, src_vma);
> __folio_mark_uptodate(new_folio);
> folio_add_new_anon_rmap(new_folio, dst_vma, addr, RMAP_EXCLUSIVE);
> folio_add_lru_vma(new_folio, dst_vma);
> @@ -1166,8 +1169,9 @@ copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
> /*
> * If we need a pre-allocated page for this pte, drop the
> * locks, allocate, and try again.
> + * If copy failed due to hwpoison in source page, break out.
> */
> - if (unlikely(ret == -EAGAIN))
> + if (unlikely(ret == -EAGAIN || ret == -EHWPOISON))
Will it be better to put checking ret against -EHWPOISON in a new line? -EAGAIN case will enter the
loop again but -EHWPOISON case never does.
> break;
> if (unlikely(prealloc)) {
> /*
> @@ -1197,7 +1201,7 @@ copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
> goto out;
> }
> entry.val = 0;
> - } else if (ret == -EBUSY) {
> + } else if (ret == -EBUSY || unlikely(ret == -EHWPOISON)) {
The caller of copy_pte_range() always set errno to -ENOMEM. So fork will failed with ENOMEM even if the real
cause is failed to copy due to hwpoison in source page. It's a pity.
Thanks.
.
