[PATCH] arm64 : fix makedumpfile failure on 5.4+ kernels

[Bhupesh SHARMA]

Hello Ioanna,

Thanks for the patch. I am partially at blame here (and also for
top-posting here) as this failure is caused due to the flipped VA
address space support we have on arm64 architecture now with newer
kernels (>= 5.4.0) due to the addition of larger VA addressing space
features (52-bit).

I had sent out a v4 series to fix this issue several months back (see
<https://www.spinics.net/lists/kexec/msg23850.html>) and I was
supposed to send an update for the same (v5), but the kernel patches
(for variable export to vmcoreinfo) took a longer time to get accepted
upstream.

I have shared the updated v5 version earlier today (see
<http://lists.infradead.org/pipermail/kexec/2020-September/021333.html>).
Can you please try the same and share your testing inputs.

Thanks,
Bhupesh




On Thu, Sep 10, 2020 at 6:25 AM Ioanna Alifieraki
<ioanna-maria.alifieraki at canonical.com> wrote:
>
> Currently makedumpfile fails on arm64 for 5.4 and newer kernels with
> the following :
>
> [9.595476] kdump-tools[513]: Starting kdump-tools:
> [9.597988] kdump-tools[519]:  * running makedumpfile -c -d 31 /proc/vmcore /var/crash/202009011332/dump-incomplete
> [9.636915] kdump-tools[537]: calculate_plat_config: PAGE SIZE 0x1000 and VA Bits 47 not supported
> [9.639652] kdump-tools[537]: get_machdep_info_arm64: Can't determine platform config values
> [9.642085] kdump-tools[537]: makedumpfile Failed.
> [9.643064] kdump-tools[519]:  * kdump-tools: makedumpfile failed, falling back to 'cp'
>
> The problem starts at get_versiondep_info_arm64 function.
> This functions uses _stext to calculate the va_bits.
> Up to 5.3 kernels the _stext would be ffff000010081000.
> After commit 14c127c957c1c607 (arm64: mm: Flip kernel VA space),
> _stext is ffff800010081000 which ends up the va_bits getting the value
> of 47, even though the kernel configuration is 48 bits.
> Now _stext has contiguous bits ffff8  and matches the 47 bits
> while in the past it had ffff0 that would match the 48 bits.
> The va_bits variable is already exported in vmcoreinfo and therefore
> that could be solved by reading the value  from there
> (va_bits = NUMBER(VA_BITS)) instead of relying on _stext.
> However, if we do so, the page_offset is not calculated properly.
> Currently :
> info->page_offset = (0xffffffffffffffffUL) << (va_bits - 1);
> The page_offset still depends on the _stext value.
> So read the va_bits from vmcoreinfo and keep the old logic of calculating
> va_bits to calculate page_offset (calc_page_offset_variable).
>
> Signed-off-by: Ioanna Alifieraki <ioanna-maria.alifieraki at canonical.com>
> ---
>  arch/arm64.c | 18 +++++++++++-------
>  1 file changed, 11 insertions(+), 7 deletions(-)
>
> diff --git a/arch/arm64.c b/arch/arm64.c
> index 54d60b4..2b0fca3 100644
> --- a/arch/arm64.c
> +++ b/arch/arm64.c
> @@ -290,6 +290,7 @@ int
>  get_versiondep_info_arm64(void)
>  {
>         ulong _stext;
> +       int calc_page_offset;
>
>         _stext = get_stext_symbol();
>         if (!_stext) {
> @@ -297,25 +298,28 @@ get_versiondep_info_arm64(void)
>                 return FALSE;
>         }
>
> +       va_bits = NUMBER(VA_BITS);
> +
>         /* Derive va_bits as per arch/arm64/Kconfig */
>         if ((_stext & PAGE_OFFSET_36) == PAGE_OFFSET_36) {
> -               va_bits = 36;
> +               calc_page_offset = 36;
>         } else if ((_stext & PAGE_OFFSET_39) == PAGE_OFFSET_39) {
> -               va_bits = 39;
> +               calc_page_offset = 39;
>         } else if ((_stext & PAGE_OFFSET_42) == PAGE_OFFSET_42) {
> -               va_bits = 42;
> +               calc_page_offset = 42;
>         } else if ((_stext & PAGE_OFFSET_47) == PAGE_OFFSET_47) {
> -               va_bits = 47;
> +               calc_page_offset = 47;
>         } else if ((_stext & PAGE_OFFSET_48) == PAGE_OFFSET_48) {
> -               va_bits = 48;
> +               calc_page_offset = 48;
>         } else {
> -               ERRMSG("Cannot find a proper _stext for calculating VA_BITS\n");
> +               ERRMSG("Cannot find a proper _stext for calculating page_offset\n");
>                 return FALSE;
>         }
>
> -       info->page_offset = (0xffffffffffffffffUL) << (va_bits - 1);
> +       info->page_offset = (0xffffffffffffffffUL) << (calc_page_offset - 1);
>
>         DEBUG_MSG("va_bits      : %d\n", va_bits);
> +       DEBUG_MSG("calc_page_offset  : %d\n", calc_page_offset);
>         DEBUG_MSG("page_offset  : %lx\n", info->page_offset);
>
>         return TRUE;
> --
> 2.17.1
>
>
> _______________________________________________
> kexec mailing list
> kexec at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec