[PATCH v9 4/5] arm64: kexec: Add support for crashkernel CMA reservation

[Jinjie Ruan]

On 2026/3/23 18:20, Breno Leitao wrote:
> On Mon, Mar 23, 2026 at 03:27:44PM +0800, Jinjie Ruan wrote:
>> Commit 35c18f2933c5 ("Add a new optional ",cma" suffix to the
>> crashkernel= command line option") and commit ab475510e042 ("kdump:
>> implement reserve_crashkernel_cma") added CMA support for kdump
>> crashkernel reservation.
>>
>> Crash kernel memory reservation wastes production resources if too
>> large, risks kdump failure if too small, and faces allocation difficulties
>> on fragmented systems due to contiguous block constraints. The new
>> CMA-based crashkernel reservation scheme splits the "large fixed
>> reservation" into a "small fixed region + large CMA dynamic region": the
>> CMA memory is available to userspace during normal operation to avoid
>> waste, and is reclaimed for kdump upon crash—saving memory while
>> improving reliability.
>>
>> So extend crashkernel CMA reservation support to arm64. The following
>> changes are made to enable CMA reservation:
>>
>> - Parse and obtain the CMA reservation size along with other crashkernel
>>   parameters.
>> - Call reserve_crashkernel_cma() to allocate the CMA region for kdump.
>> - Include the CMA-reserved ranges for kdump kernel to use.
>> - Exclude the CMA-reserved ranges from the crash kernel memory to
>>   prevent them from being exported through /proc/vmcore, which is already
>>   done in the crash core.
>>
>> Update kernel-parameters.txt to document CMA support for crashkernel on
>> arm64 architecture.
>>
>> Acked-by: Rob Herring (Arm) <robh at kernel.org>
>> Acked-by: Baoquan He <bhe at redhat.com>
>> Acked-by: Mike Rapoport (Microsoft) <rppt at kernel.org>
>> Acked-by: Ard Biesheuvel <ardb at kernel.org>
>> Signed-off-by: Jinjie Ruan <ruanjinjie at huawei.com>
>> ---
>> v7:
>> - Correct the inclusion of CMA-reserved ranges for kdump
>>   kernel in of/kexec.
>> v3:
>> - Add Acked-by.
>> v2:
>> - Free cmem in prepare_elf_headers()
>> - Add the mtivation.
>> ---
>>  Documentation/admin-guide/kernel-parameters.txt | 2 +-
>>  arch/arm64/kernel/machine_kexec_file.c          | 2 +-
>>  arch/arm64/mm/init.c                            | 5 +++--
>>  drivers/of/fdt.c                                | 9 +++++----
>>  drivers/of/kexec.c                              | 9 +++++++++
>>  5 files changed, 19 insertions(+), 8 deletions(-)
>>
>> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
>> index cb850e5290c2..afb3112510f7 100644
>> --- a/Documentation/admin-guide/kernel-parameters.txt
>> +++ b/Documentation/admin-guide/kernel-parameters.txt
>> @@ -1121,7 +1121,7 @@ Kernel parameters
>>  			It will be ignored when crashkernel=X,high is not used
>>  			or memory reserved is below 4G.
>>  	crashkernel=size[KMG],cma
>> -			[KNL, X86, ppc] Reserve additional crash kernel memory from
>> +			[KNL, X86, ARM64, PPC] Reserve additional crash kernel memory from
>>  			CMA. This reservation is usable by the first system's
>>  			userspace memory and kernel movable allocations (memory
>>  			balloon, zswap). Pages allocated from this memory range
>> diff --git a/arch/arm64/kernel/machine_kexec_file.c b/arch/arm64/kernel/machine_kexec_file.c
>> index c338506a580b..cc577d77df00 100644
>> --- a/arch/arm64/kernel/machine_kexec_file.c
>> +++ b/arch/arm64/kernel/machine_kexec_file.c
>> @@ -42,7 +42,7 @@ int arch_kimage_file_post_load_cleanup(struct kimage *image)
>>  #ifdef CONFIG_CRASH_DUMP
>>  unsigned int arch_get_system_nr_ranges(void)
>>  {
>> -	unsigned int nr_ranges = 2; /* for exclusion of crashkernel region */
>> +	unsigned int nr_ranges = 2 + crashk_cma_cnt; /* for exclusion of crashkernel region */
> 
> You update arch_get_system_nr_ranges() to account for CMA ranges, but
> prepare_elf_headers() in the same file (line 51) still has the
> hardcoded:
> 
>         nr_ranges = 2; /* for exclusion of crashkernel region */

I don't see any logic related to prepare_elf_headers() or hardcoded
nr_ranges = 2 in the arm64 implementation.

Did I miss something here?

> 
> and does not exclude CMA ranges from cmem. If the generic crash core
> handles CMA exclusion from vmcore, then shouldn't
> arch_get_system_nr_ranges() also not need this change?
>