[PATCH v4] erofs: support superblock checksum

[Chao Yu]

On 2019/10/23 16:45, Gao Xiang wrote:
> Hi Chao,
> 
> On Wed, Oct 23, 2019 at 04:15:29PM +0800, Chao Yu wrote:
>> Hi, Xiang, Pratik,
>>
>> On 2019/10/23 12:05, Gao Xiang wrote:
> 
> <snip>
> 
>>>  }
>>>  
>>> +static int erofs_superblock_csum_verify(struct super_block *sb, void *sbdata)
>>> +{
>>> +	struct erofs_super_block *dsb;
>>> +	u32 expected_crc, nblocks, crc;
>>> +	void *kaddr;
>>> +	struct page *page;
>>> +	int i;
>>> +
>>> +	dsb = kmemdup(sbdata + EROFS_SUPER_OFFSET,
>>> +		      EROFS_BLKSIZ - EROFS_SUPER_OFFSET, GFP_KERNEL);
>>> +	if (!dsb)
>>> +		return -ENOMEM;
>>> +
>>> +	expected_crc = le32_to_cpu(dsb->checksum);
>>> +	nblocks = le32_to_cpu(dsb->chksum_blocks);
>>
>> Now, we try to use nblocks's value before checking its validation, I guess fuzz
>> test can easily make the value extreme larger, result in checking latter blocks
>> unnecessarily.
>>
>> IMO, we'd better
>> 1. check validation of superblock to make sure all fields in sb are valid
>> 2. use .nblocks to count and check payload blocks following sb
> 
> That is quite a good point. :-)
> 
> My first thought is to check the following payloads of sb (e.g, some per-fs
> metadata should be checked at mount time together. or for small images, check
> the whole image at the mount time) as well since if we introduce a new feature
> to some kernel version, forward compatibility needs to be considered. So it's
> better to make proper scalability, for this case, we have some choices:
>  1) limit `chksum_blocks' upbound at runtime (e.g. refuse >= 65536 blocks,
>     totally 256M.)
>  2) just get rid of the whole `chksum_blocks' mess and checksum the first 4k
>     at all, don't consider any latter scalability.

Xiang, sorry for later reply...

I prefer method 2), let's enable chksum feature only on superblock first,
chksum_blocks feature can be added later.

Thanks,

> 
> Some perferred idea about this? I plan to release erofs-utils v1.0 tomorrow
> and hold up this feature for the next erofs-utils release, but I think we can
> get it ready for v5.5 since it is not quite complex feature...
> 
> Thanks,
> Gao Xiang
> 
> .
>