Re: qemu-arm64: CONFIG_ARM64_64K_PAGES=y kernel crash on qemu-arm64 with Linux next-20241210 and above

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

在 2024/12/20 01:40, Dan Carpenter 写道:
> On Thu, Dec 19, 2024 at 10:44:12AM +1030, Qu Wenruo wrote:
>>
>>
>> 在 2024/12/19 06:37, Qu Wenruo 写道:
>>>
>>>
>>> 在 2024/12/19 02:22, Naresh Kamboju 写道:
>>>> On Wed, 18 Dec 2024 at 17:33, Naresh Kamboju
>>>> <naresh.kamboju@linaro.org> wrote:
>>>>>
>>>>> The following kernel crash noticed on qemu-arm64 while running the
>>>>> Linux next-20241210 tag (to next-20241218) kernel built with
>>>>>    - CONFIG_ARM64_64K_PAGES=y
>>>>>    - CONFIG_ARM64_16K_PAGES=y
>>>>> and running LTP smoke tests.
>>>>>
>>>>> First seen on Linux next-20241210.
>>>>>     Good: next-20241209
>>>>>     Bad:  next-20241210 and next-20241218
>>>>>
>>>>> qemu-arm64: 9.1.2
>>>>>
>>>>> Anyone noticed this ?
>>>>>
>>>>
>>>> Anders bisected this reported regression and found,
>>>> # first bad commit:
>>>>     [9c1d66793b6faa00106ae4c866359578bfc012d2]
>>>>     btrfs: validate system chunk array at btrfs_validate_super()
>>>
>>> Weird, I run daily fstests with 64K page sized aarch64 VM.
>>>
>>> But never hit a crash on this.
>>>
>>> And the original crash call trace only points back to ext4, not btrfs.
>>>
>
> Yeah.  But it's in the memory allocator so it looks like memory
> corruption.  After the ext4 crash then random other stuff starts
> crashing as well when it allocates memory.
>
>>> Mind to test it with KASAN enabled?
>>
>
> Anders is going to try that later and report back.
>
>> Another thing is, how do you enable both 16K and 64K page size at the
>> same time?
>>
>> The Kconfig should only select one page size IIRC.
>
> Right.  We tested 4k, 16k and 64k.  4k pages worked.
>
>>
>> And for the bisection, does it focus on the test failure or the crash?
>>
>
> The crash.

For the failure part, I got the reason, it's indeed the patch, where we
call btrfs_check_chunk_valid() but fs_info->sectorsize is still in the
default value (4096), not the real one from the superblock.

Thus it will always report false alerts if the on-disk super block is
not using 4K sectorsize.

I'll fix it soon.

But sorry I didn't see why the false alert is related to the crash, the
only new memory allocation done in that patch is for a dummy extent
buffer, which should always be freed.

Anyway in the next version I'll get rid of the memory allocation completely.

Thanks,
Qu
>
> regards,
> dan carpenter
>
>