Re: [PATCH] s390/kfence: fix page fault reporting

[Alexander Potapenko]

On Mon, Feb 13, 2023 at 7:39 PM Heiko Carstens <hca@xxxxxxxxxxxxx> wrote:
>
> Baoquan He reported lots of KFENCE reports when /proc/kcore is read,
> e.g. with crash or even simpler with dd:
>
>  BUG: KFENCE: invalid read in copy_from_kernel_nofault+0x5e/0x120
>  Invalid read at 0x00000000f4f5149f:
>   copy_from_kernel_nofault+0x5e/0x120
>   read_kcore+0x6b2/0x870
>   proc_reg_read+0x9a/0xf0
>   vfs_read+0x94/0x270
>   ksys_read+0x70/0x100
>   __do_syscall+0x1d0/0x200
>   system_call+0x82/0xb0
>
> The reason for this is that read_kcore() simply reads memory that might
> have been unmapped by KFENCE with copy_from_kernel_nofault(). Any fault due
> to pages being unmapped by KFENCE would be handled gracefully by the fault
> handler (exception table fixup).
>
> However the s390 fault handler first reports the fault, and only afterwards
> would perform the exception table fixup. Most architectures have this in
> reversed order, which also avoids the false positive KFENCE reports when an
> unmapped page is accessed.
>
> Therefore change the s390 fault handler so it handles exception table
> fixups before KFENCE page faults are reported.
>
> Reported-by: Baoquan He <bhe@xxxxxxxxxx>
> Signed-off-by: Heiko Carstens <hca@xxxxxxxxxxxxx>
Acked-by: Alexander Potapenko <glider@xxxxxxxxxx>