On 2022/2/25 20:38, Jan Kara wrote:
> On Fri 25-02-22 18:28:37, Zhang Yi wrote:
>> The same to commit 1c2d14212b15 ("ext2: Fix underflow in ext2_max_size()")
>> in ext2 filesystem, ext4 driver has the same issue with 64K block size
>> and ^huge_file, fix this issue the same as ext2. This patch also revert
>> commit 75ca6ad408f4 ("ext4: fix loff_t overflow in ext4_max_bitmap_size()")
>> because it's no longer needed.
>>
>> Signed-off-by: Zhang Yi <yi.zhang@xxxxxxxxxx>
>
> Thanks for the patch. I would not refer to ext2 patch in the changelog - it
> is better to have it self-contained. AFAIU the problem is that (meta_blocks
>> upper_limit) for 64k blocksize and ^huge_file and so upper_limit would
> underflow during the computations, am I right?
Thanks for the review. Yes, I will rewrite the change log.
>
> Also two comments below:
>
>> diff --git a/fs/ext4/super.c b/fs/ext4/super.c
>> index c5021ca0a28a..95608c2127e7 100644
>> --- a/fs/ext4/super.c
>> +++ b/fs/ext4/super.c
>> @@ -3468,8 +3468,9 @@ static loff_t ext4_max_size(int blkbits, int has_huge_files)
>> */
>> static loff_t ext4_max_bitmap_size(int bits, int has_huge_files)
>> {
>> - unsigned long long upper_limit, res = EXT4_NDIR_BLOCKS;
>> + loff_t upper_limit, res = EXT4_NDIR_BLOCKS;
>> int meta_blocks;
>> + unsigned int ppb = 1 << (bits - 2);
>>
>> /*
>> * This is calculated to be the largest file size for a dense, block
>> @@ -3501,27 +3502,42 @@ static loff_t ext4_max_bitmap_size(int bits, int has_huge_files)
>>
>> }
>>
>> - /* indirect blocks */
>> - meta_blocks = 1;
>> - /* double indirect blocks */
>> - meta_blocks += 1 + (1LL << (bits-2));
>> - /* tripple indirect blocks */
>> - meta_blocks += 1 + (1LL << (bits-2)) + (1LL << (2*(bits-2)));
>> -
>> - upper_limit -= meta_blocks;
>> - upper_limit <<= bits;
>> -
>> + /* Compute how many blocks we can address by block tree */
>> res += 1LL << (bits-2);
>> res += 1LL << (2*(bits-2));
>> res += 1LL << (3*(bits-2));
>
> When you have the 'ppb' convenience variable, perhaps you can update this
> math to:
>
> res = EXT4_NDIR_BLOCKS + ppb + ppb*ppb + ((long long)ppb)*ppb*ppb;
>
> It is easier to understand and matches how you compute meta_blocks as well.
>
>> + /* Compute how many metadata blocks are needed */
>> + meta_blocks = 1;
>> + meta_blocks += 1 + ppb;
>> + meta_blocks += 1 + ppb + ppb * ppb;
>> + /* Does block tree limit file size? */
>> + if (res + meta_blocks <= upper_limit)
>> + goto check_lfs;
>> +
>> + res = upper_limit;
>> + /* How many metadata blocks are needed for addressing upper_limit? */
>> + upper_limit -= EXT4_NDIR_BLOCKS;
>> + /* indirect blocks */
>> + meta_blocks = 1;
>> + upper_limit -= ppb;
>> + /* double indirect blocks */
>> + if (upper_limit < ppb * ppb) {
>> + meta_blocks += 1 + DIV_ROUND_UP_ULL(upper_limit, ppb);
>> + res -= meta_blocks;
>> + goto check_lfs;
>> + }
>> + meta_blocks += 1 + ppb;
>> + upper_limit -= ppb * ppb;
>> + /* tripple indirect blocks for the rest */
>> + meta_blocks += 1 + DIV_ROUND_UP_ULL(upper_limit, ppb) +
>> + DIV_ROUND_UP_ULL(upper_limit, ppb*ppb);
>> + res -= meta_blocks;
>> +check_lfs:
>> res <<= bits;
>
> Cannot this overflow loff_t again? I mean if upper_limit == (1 << 48) - 1
> and we have 64k blocksize, 'res' will be larger than (1 << 47) and thus
> res << 16 will be greater than 1 << 63 => negative... Am I missing
> something?
>
If upper_limit==(1 << 48) - 1, we could address the whole data blocks, the 'res'
is equal to EXT4_NDIR_BLOCKS + ppb + ppb*ppb + ((long long)ppb)*ppb*ppb, it's
smaller than (1 << 43) - 1, so res << 16 is still smaller 1 << 59, so it cannot
overflow loff_t again.
Thanks,
Yi.
