Re: [Ocfs2-devel] [PATCH 1/3] ext3/ext4: Factor out disk addressability check

[Eric Sandeen <sandeen@redhat.com>]

Jan Kara wrote:
> On Thu 12-08-10 15:29:49, Joel Becker wrote:
>> diff --git a/fs/libfs.c b/fs/libfs.c
>> index dcaf972..f099566 100644
>> --- a/fs/libfs.c
>> +++ b/fs/libfs.c
>> @@ -955,6 +955,35 @@ int generic_file_fsync(struct file *file, int datasync)
>>  }
>>  EXPORT_SYMBOL(generic_file_fsync);
>>  
>> +/**
>> + * generic_check_addressable - Check addressability of file system
>> + * @blocksize_bits:	log of file system block size
>> + * @num_blocks:		number of blocks in file system
>> + *
>> + * Determine whether a file system with @num_blocks blocks (and a
>> + * block size of 2**@blocksize_bits) is addressable by the sector_t
>> + * and page cache of the system.  Return 0 if so and -EFBIG otherwise.
>> + */
>> +int generic_check_addressable(unsigned blocksize_bits, u64 num_blocks)
>> +{
>> +	u64 last_fs_block = num_blocks - 1;
>> +
>> +	if (unlikely(num_blocks == 0))
>> +		return 0;
>> +
>> +	if ((blocksize_bits < 9) || (blocksize_bits > PAGE_CACHE_SHIFT))
>> +		return -EINVAL;
>> +
>> +	if ((last_fs_block >
>> +	     (sector_t)(~0ULL) >> (blocksize_bits - 9)) ||
>> +	    (last_fs_block >
>> +	     (pgoff_t)(~0ULL) >> (PAGE_CACHE_SHIFT - blocksize_bits))) {
>             ^^^ I don't get the pgoff_t check. Shouldn't it rather be
> (u64)(pgoff_t)(~0ULL) << (PAGE_CACHE_SHIFT - blocksize_bits)?

Argh that was my fault...  Thankfully not too many 1k-blocksize-formatted
16T devices out there, I guess.

I went through the math again and also came up with:

total fs pages is blocks / (blocks per page)
total pages is blocks / (1 << PAGE_CACHE_SHIFT / 1 << blocksize_bits)
total pages is blocks / (1 << (PAGE_CACHE_SHIFT - blocksize_bits))
total pages is blocks * (1 >> (PAGE_CACHE_SHIFT - blocksize_bits))
total pages is blocks >> (PAGE_CACHE_SHIFT - blocksize_bits)

too big if total pages is > (pgoff_t)(~0ULL)
too big if blocks >> (PAGE_CACHE_SHIFT - blocksize_bits) > (pgoff_t)(~0ULL)
too big if blocks > (pgoff_t)(~0ULL) << (PAGE_CACHE_SHIFT - blocksize_bits)
and to not overflow:
too big if blocks > (u64)(pgoff_t)(~0ULL) << (PAGE_CACHE_SHIFT - blocksize_bits)

so seems like the test is:

last_fs_block > (u64)(pgoff_t)(~0ULL) << (PAGE_CACHE_SHIFT - blocksize_bits)

Given the density of the logic in the helper it seems like maybe breaking it
up and adding specific comments might be helpful to the reader:

	/* can IO layers fit total fs sectors in a sector_t? */
	if (last_fs_block >
	    (sector_t)(~0ULL) >> (blocksize_bits - 9))
		return -EFBIG;

	/* can page cache fit total fs pages in a pgoff_t? */
	if (last_fs_block >
	    (u64)(pgoff_t)(~0ULL) << (PAGE_CACHE_SHIFT - blocksize_bits)
		return -EFBIG;

Or something like that.

Sorry for chiming in late...

-Eric

> Because on 32-bit arch we are able to address 16TB device, which is for 1KB
> blocksize 1<<34 blocks. But your math gives 1<<30 blocks...
> 
> 									Honza
>