Re: [PATCH] arm64, vmcoreinfo : Append 'MAX_USER_VA_BITS' and 'MAX_PHYSMEM_BITS' to vmcoreinfo

[Bhupesh Sharma <bhsharma@redhat.com>]

Hi Kazu,

On 02/22/2019 12:32 AM, Kazuhito Hagio wrote:
> -----Original Message-----
>> ----- Original Message -----
>>> Hi Kazu,
>>>
>>> On 02/20/2019 02:17 AM, Kazuhito Hagio wrote:
>>>> Hi Bhupesh,
>>>>
>>>> -----Original Message-----
>>>>> I am not sure you got a chance to look at the two regression cases I
>>>>> reported here:
>>>>> <http://lists.infradead.org/pipermail/kexec/2019-February/022449.html>
>>>>>
>>>>> Unfortunately the above suggestion doesn't provide any fix for
>>>>> ARMv8.2-LPA regression (see text under heading '
>>>>> (1). Regression Case 1 (ARMv8.2-LPA enabled kernel)')
>>>>
>>>> As for MAX_PHYSMEM_BITS, I realized that ppc64 makedumpfile can detect
>>>> it because there is only one SECTION_SIZE_BITS for ppc64. I think we
>>>> can use the same way as set_ppc64_max_physmem_bits() does also for
>>>> arm64 for now. I'm going to write it for kernels not having
>>>> NUMBER(MAX_PHYSMEM_BITS) in vmcoreinfo.
>>>
>>> I see two drawbacks with the above approach:
>>>
>>> a). This means that other user-space tools like crash-utility would
>>> still be broken and would probably need to find MAX_PHYSMEM_BITS for
>>> arm64 via a similar (hack'ish ?) approach.
>>>
>>> b). I am looking at the makedumpfile code for 'MAX_PHYSMEM_BITS'
>>> determination for two archs as an example:
>>>
>>> ppc
>>> ---
>>>
>>> int
>>> set_ppc64_max_physmem_bits(void)
>>> {
>>>           long array_len = ARRAY_LENGTH(mem_section);
>>>           /*
>>>            * The older ppc64 kernels uses _MAX_PHYSMEM_BITS as 42 and the
>>>            * newer kernels 3.7 onwards uses 46 bits.
>>>            */
>>>
>>>           info->max_physmem_bits  = _MAX_PHYSMEM_BITS_ORIG ;
>>>           if ((array_len == (NR_MEM_SECTIONS() / _SECTIONS_PER_ROOT_EXTREME()))
>>>                   || (array_len == (NR_MEM_SECTIONS() / _SECTIONS_PER_ROOT())))
>>>                   return TRUE;
>>>
>>>           info->max_physmem_bits  = _MAX_PHYSMEM_BITS_3_7;
>>>           if ((array_len == (NR_MEM_SECTIONS() / _SECTIONS_PER_ROOT_EXTREME()))
>>>                   || (array_len == (NR_MEM_SECTIONS() / _SECTIONS_PER_ROOT())))
>>>                   return TRUE;
>>>
>>>           info->max_physmem_bits  = _MAX_PHYSMEM_BITS_4_19;
>>>           if ((array_len == (NR_MEM_SECTIONS() / _SECTIONS_PER_ROOT_EXTREME()))
>>>                   || (array_len == (NR_MEM_SECTIONS() / _SECTIONS_PER_ROOT())))
>>>                   return TRUE;
>>>
>>>           info->max_physmem_bits  = _MAX_PHYSMEM_BITS_4_20;
>>>           if ((array_len == (NR_MEM_SECTIONS() / _SECTIONS_PER_ROOT_EXTREME()))
>>>                   || (array_len == (NR_MEM_SECTIONS() /  _SECTIONS_PER_ROOT())))
>>>                   return TRUE;
>>>
>>>           return FALSE;
>>> }
>>>
>>> x86_64:
>>> ------
>>>
>>> int
>>> get_versiondep_info_x86_64(void)
>>> {
>>>       /*
>>>        * On linux-2.6.26, MAX_PHYSMEM_BITS is changed to 44 from 40.
>>>        */
>>>       if (info->kernel_version < KERNEL_VERSION(2, 6, 26))
>>>           info->max_physmem_bits  = _MAX_PHYSMEM_BITS_ORIG;
>>>       else if (info->kernel_version < KERNEL_VERSION(2, 6, 31))
>>>           info->max_physmem_bits  = _MAX_PHYSMEM_BITS_2_6_26;
>>>       else if(check_5level_paging())
>>>           info->max_physmem_bits  = _MAX_PHYSMEM_BITS_5LEVEL;
>>>       else
>>>           info->max_physmem_bits  = _MAX_PHYSMEM_BITS_2_6_31;
>>>
>>>       ...
>>> }
>>>
>>> Looking at the above, two questions come to my mind:
>>>
>>> - Do we really need all the above complexity in user-space code, to hoop
>>> across various kernel versions and perform allocations for something
>>> that can be so easily exported via vmcoreinfo? Also we need to see how
>>> portable is the above code for a new kernel version - IMO, it will need
>>> another fix patch when we update to a new kernel version in near future.
>>
>> I agree -- not to mention that the "kernel version" way of determining things
>> does not account for distribution-specific backports.
>>
>>>
>>> - Also do we need to replicate the above implementations across
>>> user-space tools when they can also utilize the vmcoreinfo information
>>> to determine the PA_BITS range without any additional arch/kernel
>>> version specific details as the single point of obtaining this
>>> information from the kernel?
>>>
>>> So, in view of the above, I would still advocate that we use a
>>> vmcoreinfo export for 'MAX_PHYSMEM_BITS' as well to have a uniform
>>> interface for the same across all user-land applications.
>>
>> Again, totally agree.
> 
> I also agree that we should do so. Then it will be better to have
> it in kernel core code, not in arch-specific code.
> 
> Although makedumpfile may have to have the kludge for kernels that
> support 52-bit PA and don't have the exported MAX_PHYSMEM_BITS
> sooner or later..

I was waiting for maintainers Cc'ed in this thread to share their 
opinions on the above suggestion.

We especially need comments from x86 maintainers on this (Boris), since 
x86 and ppc user-space utilities are currently not broken (no regression 
reported) due to a lacking export of 'MAX_PHYSMEM_BITS' via vmcore, 
whereas as I shared earlier arm64 user-space utilities are broken right 
now with 52-bit address space changes.

Thanks,
Bhupesh

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel