Re: [RFC PATCH] ARM: Support allocating crashkernel above 4G for LPAE

[Jinjie Ruan <ruanjinjie@huawei.com>]

On 2024/8/5 21:18, Catalin Marinas wrote:
> On Fri, Aug 02, 2024 at 12:01:43PM +0100, Russell King wrote:
>> On Fri, Aug 02, 2024 at 05:25:10PM +0800, Jinjie Ruan wrote:
>>> As ARM LPAE feature support accessing memory beyond the 4G limit, define
>>> HAVE_ARCH_CRASHKERNEL_RESERVATION_HIGH macro to support reserving crash
> 
> At least in 6.11-rc1, there's no trace of such macro anywhere. So not
> sure this patch has any effect (I haven't checked linux-next though).

Sorry, this macro is introduced in linux-next, the -next subject has
been missed.

> 
>>> memory above 4G for ARM32 LPAE.
>>>
>>> No test because there is no LPAE ARM32 hardware.
>>
>> Why are you submitting patches for features you can't test?
>>
>> I'm not going to apply this without it being properly tested, because I
>> don't believe that this will work in the generic case.
>>
>> If the crash kernel is located in memory outside of the lower 4GiB of
>> address space, and there is no alias within physical address space
>> for that memory, then there is *no* *way* for such a kernel to boot.
>>
>> So, right now I believe this patch to be *fundamentally* wrong.
> 
> Indeed. Even on arm64, we keep some crashkernel reservations in the
> lower parts of the memory for ZONE_DMA allocations.

Indeed, it is.

> 
> On arch/arm with LPAE, we could do something similar like forcing some
> lowmem reservation and allowing explicit allocation in the higher ranges
> with crashkernel=<size>,high. We should, of course, force the kdump

In linux-next, with the HAVE_ARCH_CRASHKERNEL_RESERVATION_HIGH macro
defined, it is ok.

> image placement in the lower memory. The user kexec tools must be taught
> to interpret this information and provide a DT accordingly to the crash
> kernel.
>