Re: [RESEND PATCH v5 2/5] arm64/crash_core: Export TCR_EL1.T1SZ in vmcoreinfo

[Bhupesh Sharma <bhsharma@redhat.com>]

Hi James,

On 01/11/2020 12:30 AM, Dave Anderson wrote:
> 
> ----- Original Message -----
>> Hi Bhupesh,
>>
>> On 25/12/2019 19:01, Bhupesh Sharma wrote:
>>> On 12/12/2019 04:02 PM, James Morse wrote:
>>>> On 29/11/2019 19:59, Bhupesh Sharma wrote:
>>>>> vabits_actual variable on arm64 indicates the actual VA space size,
>>>>> and allows a single binary to support both 48-bit and 52-bit VA
>>>>> spaces.
>>>>>
>>>>> If the ARMv8.2-LVA optional feature is present, and we are running
>>>>> with a 64KB page size; then it is possible to use 52-bits of address
>>>>> space for both userspace and kernel addresses. However, any kernel
>>>>> binary that supports 52-bit must also be able to fall back to 48-bit
>>>>> at early boot time if the hardware feature is not present.
>>>>>
>>>>> Since TCR_EL1.T1SZ indicates the size offset of the memory region
>>>>> addressed by TTBR1_EL1 (and hence can be used for determining the
>>>>> vabits_actual value) it makes more sense to export the same in
>>>>> vmcoreinfo rather than vabits_actual variable, as the name of the
>>>>> variable can change in future kernel versions, but the architectural
>>>>> constructs like TCR_EL1.T1SZ can be used better to indicate intended
>>>>> specific fields to user-space.
>>>>>
>>>>> User-space utilities like makedumpfile and crash-utility, need to
>>>>> read/write this value from/to vmcoreinfo
>>>>
>>>> (write?)
>>>
>>> Yes, also write so that the vmcoreinfo from an (crashing) arm64 system can
>>> be used for
>>> analysis of the root-cause of panic/crash on say an x86_64 host using
>>> utilities like
>>> crash-utility/gdb.
>>
>> I read this as as "User-space [...] needs to write to vmcoreinfo".

That's correct. But for writing to vmcore dump in the kdump kernel, we 
need to read the symbols from the vmcoreinfo in the primary kernel.

>>>>> for determining if a virtual address lies in the linear map range.
>>>>
>>>> I think this is a fragile example. The debugger shouldn't need to know
>>>> this.
>>>
>>> Well that the current user-space utility design, so I am not sure we can
>>> tweak that too much.
>>>
>>>>> The user-space computation for determining whether an address lies in
>>>>> the linear map range is the same as we have in kernel-space:
>>>>>
>>>>>     #define __is_lm_address(addr)    (!(((u64)addr) & BIT(vabits_actual -
>>>>>     1)))
>>>>
>>>> This was changed with 14c127c957c1 ("arm64: mm: Flip kernel VA space"). If
>>>> user-space
>>>> tools rely on 'knowing' the kernel memory layout, they must have to
>>>> constantly be fixed
>>>> and updated. This is a poor argument for adding this to something that
>>>> ends up as ABI.
>>>
>>> See above. The user-space has to rely on some ABI/guaranteed
>>> hardware-symbols which can be
>>> used for 'determining' the kernel memory layout.
>>
>> I disagree. Everything and anything in the kernel will change. The ABI rules apply to
>> stuff exposed via syscalls and kernel filesystems. It does not apply to kernel internals,
>> like the memory layout we used yesterday. 14c127c957c1 is a case in point.
>>
>> A debugger trying to rely on this sort of thing would have to play catchup whenever it
>> changes.
> 
> Exactly.  That's the whole point.
> 
> The crash utility and makedumpfile are not in the same league as other user-space tools.
> They have always had to "play catchup" precisely because they depend upon kernel internals,
> which constantly change.

I agree with you and DaveA here. Software user-space debuggers are 
dependent on kernel internals (which can change from time-to-time) and 
will have to play catch-up (which has been the case since the very start).

Unfortunately we don't have any clear ABI for software debugging tools - 
may be something to look for in future.

A case in point is gdb/kgdb, which still needs to run with KASLR 
turned-off (nokaslr) for debugging, as it confuses gdb which resolve 
kernel symbol address from symbol table of vmlinux. But we can 
work-around the same in makedumpfile/crash by reading the 'kaslr_offset' 
value. And I have several users telling me now they cannot use gdb on 
KASLR enabled kernel to debug panics, but can makedumpfile + crash 
combination to achieve the same.

So, we should be looking to fix these utilities which are broken since 
the 52-bit changes for arm64. Accordingly, I will try to send the v6
soon while incorporating the comments posted on the v5.

Thanks,
Bhupesh