[PATCH makedumpfile] Avoid false-positive mem_section validation with vmlinux

[Pingfan Liu]

On Thu, Apr 21, 2022 at 7:58 AM HAGIO KAZUHITO(萩尾　一仁)
<k-hagio-ab at nec.com> wrote:
>
> Currently get_mem_section() validates if SYMBOL(mem_section) is the address
> of the mem_section array first.  But there was a report that the first
> validation wrongly returned TRUE with -x vmlinux and SPARSEMEM_EXTREME
> (4.15+) on s390x.  This leads to crash failing statup with the following
> seek error:
>
>   crash: seek error: kernel virtual address: 67fffc2800  type: "memory section root table"
>
> Skip the first validation when satisfying the conditions.
>
> Reported-by: Dave Wysochanski <dwysocha at redhat.com>
> Signed-off-by: Kazuhito Hagio <k-hagio-ab at nec.com>
> ---
>  makedumpfile.c | 31 +++++++++++++++++++++++++++++++
>  1 file changed, 31 insertions(+)
>
> diff --git a/makedumpfile.c b/makedumpfile.c
> index a2f45c84cee3..65d1c7c2f02c 100644
> --- a/makedumpfile.c
> +++ b/makedumpfile.c
> @@ -3698,6 +3698,22 @@ validate_mem_section(unsigned long *mem_sec,
>         return ret;
>  }
>
> +/*
> + * SYMBOL(mem_section) varies with the combination of memory model and
> + * its source:
> + *
> + * SPARSEMEM
> + *   vmcoreinfo: address of mem_section root array
> + *   -x vmlinux: address of mem_section root array
> + *
> + * SPARSEMEM_EXTREME v1
> + *   vmcoreinfo: address of mem_section root array
> + *   -x vmlinux: address of mem_section root array
> + *
> + * SPARSEMEM_EXTREME v2 (with 83e3c48729d9 and a0b1280368d1) 4.15+
> + *   vmcoreinfo: address of mem_section root array
> + *   -x vmlinux: address of pointer to mem_section root array
> + */
>  static int
>  get_mem_section(unsigned int mem_section_size, unsigned long *mem_maps,
>                 unsigned int num_section)
> @@ -3710,12 +3726,27 @@ get_mem_section(unsigned int mem_section_size, unsigned long *mem_maps,
>                     strerror(errno));
>                 return FALSE;
>         }
> +
> +       /*
> +        * There was a report that the first validation wrongly returned TRUE
> +        * with -x vmlinux and SPARSEMEM_EXTREME v2 on s390x, so skip it.
> +        * Howerver, leave the fallback validation as it is for the -i option.
> +        */
> +       if (is_sparsemem_extreme() && info->name_vmlinux) {
> +               unsigned long flag = 0;
> +               if (get_symbol_type_name("mem_section", DWARF_INFO_GET_SYMBOL_TYPE,
> +                                       NULL, &flag)
> +                   && !(flag & TYPE_ARRAY))
> +                       goto skip_1st_validation;
> +       }
> +
>         ret = validate_mem_section(mem_sec, SYMBOL(mem_section),
>                                    mem_section_size, mem_maps, num_section);
>
>         if (!ret && is_sparsemem_extreme()) {
>                 unsigned long mem_section_ptr;
>
> +skip_1st_validation:
>                 if (!readmem(VADDR, SYMBOL(mem_section), &mem_section_ptr,
>                              sizeof(mem_section_ptr)))
>                         goto out;
> --
> 2.27.0
>
Discussed with Kazu off-list, and with his nice help, I got clear why
he drops V1.

Hence,
Reviewed-by: Pingfan Liu <piliu at redhat.com>