On 2024/9/6 10:42, Kefeng Wang wrote:
> Like commit a873dfe1032a ("mm, hwpoison: try to recover from copy-on
> write faults"), there is another path which could crash because it does
> not have recovery code where poison is consumed by the kernel in
> do_cow_fault(), a crash calltrace shown below on old kernel, but it
> could be happened in the lastest mainline code,
>
> CPU: 7 PID: 3248 Comm: mpi Kdump: loaded Tainted: G OE 5.10.0 #1
> pc : copy_page+0xc/0xbc
> lr : copy_user_highpage+0x50/0x9c
> Call trace:
> copy_page+0xc/0xbc
> do_cow_fault+0x118/0x2bc
> do_fault+0x40/0x1a4
> handle_pte_fault+0x154/0x230
> __handle_mm_fault+0x1a8/0x38c
> handle_mm_fault+0xf0/0x250
> do_page_fault+0x184/0x454
> do_translation_fault+0xac/0xd4
> do_mem_abort+0x44/0xbc
>
> Fix it by using copy_mc_user_highpage() to handle this case and return
> VM_FAULT_HWPOISON for cow fault.
>
> Signed-off-by: Kefeng Wang <wangkefeng.wang@xxxxxxxxxx>
> ---
> mm/memory.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/mm/memory.c b/mm/memory.c
> index 42674c0748cb..d310c073a1b3 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -5089,7 +5089,10 @@ static vm_fault_t do_cow_fault(struct vm_fault *vmf)
> if (ret & VM_FAULT_DONE_COW)
> return ret;
>
> - copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
> + if (copy_mc_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma)) {
> + ret = VM_FAULT_HWPOISON;
> + goto uncharge_out;
> + }
When copy_mc_user_highpage fails, we should have vmf->page locked and hold the extra refcnt
of vmf->page. So we should call unlock_page(vmf->page) and put_page(vmf->page) before goto
uncharge_out?
Thanks.
.
