Re: [PATCH v9 0/8] kdump: reduce vmcore size and capture time

[Wandun <chenwandun1@gmail.com>]

On 11/10/26 16:47, Mike Rapoport wrote:
> Hi Wandun,
> 
> On Mon, Nov 09, 2026 at 08:23:16PM +0800, Wandun wrote:
>>
>> On 10/21/26 16:30, Wandun Chen wrote:
>>> From: Wandun Chen <chenwandun@lixiang.com>
>>>
>>> On SoCs that carve out large firmware-owned reserved memory (GPU,
>>> camera ISP, ...), kdump currently dumps those carveouts as part of
>>> system RAM even though their contents are firmware state that is not
>>> useful for kernel crash analysis.
>>>
>>> ACPI systems already filter reserved memory out of the vmcore through
>>> their existing path: firmware reports those carveouts as unusable in
>>> the EFI memory map, they get MEMBLOCK_NOMAP set in memblock.memory
>>> (is_usable_memory(), efi-init.c), and the ELF header builder has
>>> always skipped nomap regions; only DT-based systems currently fail
>>> to filter these regions, which is what this series addresses.
>>>
>>> This series introduces a MEMBLOCK_NODUMP flag (suggested by Rob) in
>>> memblock to filter reserved memory on DT-based architectures. A new
>>> flag is needed because MEMBLOCK_NOMAP can only express regions without
>>> a linear mapping, while MEMBLOCK_NODUMP can filter regions that are
>>> linearly mapped. Reserved regions are marked MEMBLOCK_NODUMP by default
>>> so kdump omits them; reusable CMA regions are different because their
>>> pages are handed back to the buddy allocator and may carry crash-relevant
>>> data.
>>>
>>> Since the reserved memory regions are filtered out, the vmcore is
>>> smaller in size and faster to produce.
>>>
>>> The series is based on linux-next and is organized as follows:
>>>
>>> Patches 1-3: Preparation and bugfixes: switch riscv crash_mem to
>>>              memblock, fold the duplicated per-arch memblock walks
>>>              into the weak defaults, and serialize crash header
>>>              preparation against memory hotplug.
>>> Patches 4-7: NODUMP infrastructure: switch crash_core to
>>>              for_each_mem_region(), introduce the MEMBLOCK_NODUMP
>>>              flag, add a dumpable flag to struct reserved_mem, and
>>>              mark /reserved-memory entries with MEMBLOCK_NODUMP.
>>> Patch 8:     Skip MEMBLOCK_NODUMP regions when building the vmcore
>>>              ELF header.
>>
>> Hi all,
>>
>> A gentle ping on this series.
>>
>> I'd appreciate any further feedback, particularly on the
>> MEMBLOCK_NODUMP discussion and the issues raised in the Sashiko
>> review.
> 
> I'm fine with MEMBLOCK_NODUMP.
> The issue Sashiko raised about bound check against cmem->max_nr_ranges
> seems valid. 

Hi Mike,

Thanks a lot for the review.

Some analysis and plan about "bound check against cmem->max_nr_ranges":

After patch 3, the remaining race is x86-specific. The weak defaults are
safe: every runtime modification of memblock.memory, hot-add and hot-remove
alike, happens under device_hotplug_lock, so with the lock held the count
and populate passes walk a stable snapshot.

x86 instead walks iomem_resource, and Xen balloon inserts its new System RAM
resource before taking device_hotplug_lock() (reserve_additional_memory),
which is the one remaining hole.

The issue is also pre-existing: before patch 3 there was no locking at all,
every hotplug path could race the two passes.

The proper fix is to switch x86 over to the memblock-based weak defaults,
as patch 2 already did for arm64/loongarch/riscv. Would you prefer to have
it in this series, or as a separate follow-up? Both work fine for me.

> 
> I didn't check the reserved mem one.

The two Sashiko findings on the reserved-mem (patch 7) will also be addressed
in next version.

>  
>> Please let me know if there are any other concerns or suggestions
>> before I prepare the next revision.
> 
> I believe that should be rebased to account for reserved memory changes.

Sure, the next verison will be rebased on latest linux-next.

> The riscv change from resources to memblock needs an Ack from riscv
> maintainers.

Hi Paul, Palmer, Albert,

Patch 1 builds the crash_mem ranges for the vmcore ELF header from
memblock instead of the resource tree, so the later MEMBLOCK_NODUMP
patches can filter reserved memory from the vmcore on the memblock
side. With this, the vmcore size reduction the series brings also
applies to riscv.

Could you take a look when you have a chance?


Best regards,
Wandun

> 
> And with all that I can take the next version via kexec tree.
>  
>> Thanks,
>> Wandun
>>
>>>
>>> v8 --> v9:
>>> 1. Patch 7: Avoid a potential boot panic on SH when NODUMP marking requires
>>>    memblock resizing before it is allowed, found by Sashiko in [1]. SH supports
>>>    neither CONFIG_KEXEC_FILE nor CONFIG_CRASH_HOTPLUG, so skip marking when
>>>    both are disabled.
>>>
>>> v6 --> v8 (v7 mainly tried to fix the initrd issue on release in v6,
>>> that approach is superseded here):
>>> 1. Stop marking /memreserve/ regions with NODUMP. v6 marked
>>>    /memreserve/ regions with NODUMP. But a bootloader may specify
>>>    a /memreserve/ entry for memory that the kernel later frees
>>>    (e.g. the initrd), and those pages would then be wrongly
>>>    filtered out of the vmcore after being reused as normal memory.
>>>    The v7 fix (clearing the flag in free_reserved_area()) caused
>>>    a lockless runtime mutation of memblock.memory [2]. Better solutions
>>>    exist but need extra preparation, so temporarily dropped the patch
>>>    that introduced the problem to keep this series simple in v8.
>>> 2. Fix a build failure in patch 2 with both ARCH_KEEP_MEMBLOCK and
>>>    MEMORY_HOTPLUG are unset. There the memblock arrays live in
>>>    .init.data, and the weak defaults referencing them create section
>>>    mismatches, which turn into build errors when SECTION_MISMATCH_WARN_ONLY
>>>    is also unset.
>>> 3. RISCV recently gained crash hotplug support, so adapt its handler
>>>    like x86 in patch 3: use __crash_prepare_headers() instead of
>>>    crash_prepare_headers() to avoid self-deadlock.
>>> 4. Update the cover letter per review feedback: describe how ACPI
>>>    systems filter reserved memory and why a new flag is needed
>>>    instead of reusing NOMAP.
>>>
>>> v6 --> v7:
>>> 1. Fix a stale MEMBLOCK_NODUMP flag issue [3]. Memory reserved and
>>>    marked as NODUMP may later be freed (the initrd, for example),
>>>    so the NODUMP flag must be cleared accordingly. Since the fix is
>>>    part of the NODUMP infrastructure introduced by patch 5, it has
>>>    been folded into that patch.
>>> 2. Drop the "generic" related description in the cover letter.
>>> 3. Collect the Acked-by and Reviewed-by tags received during the v6
>>>    review (thanks to Baoquan, Mike, Rob and Marek). Patch 5 is the
>>>    only one without a tag, as its Reviewed-by was dropped when the
>>>    fix was folded in.
>>> 4. Drop patch 1 of v6, which has been merged separately.
>>>
>>>
>>> v5 --> v6:
>>> 1. Serialize crash header preparation against memory hotplug to avoid
>>>    out-of-bounds or use-after-free issues.
>>> 2. MEMBLOCK_NODUMP marking is now done after memblock allows resizing,
>>>    avoiding a panic from too few regions before resize is permitted.
>>> 3. Reordered the patches, put pre-existing bugfixes earlier in the series.
>>>
>>> [1] https://sashiko.dev/#/patchset/20261016061549.4181390-1-chenwandun1@gmail.com?part=7
>>> [2] https://lore.kernel.org/lkml/20260922090252.B01541F000FF@smtp.kernel.org/
>>> [3] https://sashiko.dev/#/message/20260917220750.GA4013814-robh%40kernel.org
>>>
>>> v5: https://lore.kernel.org/lkml/20260806093519.2907771-1-chenwandun1@gmail.com/
>>> v4: https://lore.kernel.org/lkml/20260630074715.4126796-1-chenwandun1@gmail.com/
>>>
>>> Wandun Chen (8):
>>>   riscv: build crash_mem ranges from memblock instead of resource tree
>>>   crash_core: fold duplicated memblock arch hooks into the weak default
>>>   crash_core: serialize crash header preparation against hotplug
>>>   crash_core: replace for_each_mem_range() with for_each_mem_region()
>>>   memblock: introduce MEMBLOCK_NODUMP flag
>>>   of: reserved_mem: add dumpable flag to opt-in vmcore
>>>   of: reserved_mem: mark /reserved-memory entries with MEMBLOCK_NODUMP
>>>   crash_core: skip MEMBLOCK_NODUMP regions when building vmcore ELF
>>>     header
>>>
>>>  arch/arm64/kernel/machine_kexec_file.c     | 29 ------------
>>>  arch/loongarch/kernel/machine_kexec_file.c | 27 -----------
>>>  arch/riscv/Kconfig                         |  2 +-
>>>  arch/riscv/kernel/crash.c                  |  2 +-
>>>  arch/riscv/kernel/machine_kexec_file.c     | 33 --------------
>>>  arch/x86/kernel/crash.c                    |  2 +-
>>>  drivers/of/fdt.c                           |  1 +
>>>  drivers/of/of_private.h                    |  1 +
>>>  drivers/of/of_reserved_mem.c               | 18 ++++++++
>>>  include/linux/crash_core.h                 |  2 +
>>>  include/linux/memblock.h                   |  9 ++++
>>>  include/linux/of_reserved_mem.h            |  1 +
>>>  kernel/crash_core.c                        | 53 +++++++++++++++++++++-
>>>  kernel/dma/contiguous.c                    |  1 +
>>>  mm/memblock.c                              | 16 +++++++
>>>  15 files changed, 103 insertions(+), 94 deletions(-)
>>>
>>
>