Re: [PATCH 6/7] btrfs: Replace open-coded maths with DIV_ROUND_UP

[Nikolay Borisov <nborisov@suse.com>]

On 3.01.19 г. 16:44 ч., David Sterba wrote:
> On Thu, Jan 03, 2019 at 10:50:04AM +0200, Nikolay Borisov wrote:
>> In a couple of places it's required to calculate the number of pages
>> given a start and end offsets. Currently this is opencoded, unify the
>> code base by replacing all such sites with the DIV_ROUND_UP macro. Also,
>> current open-coded sites were buggy in that they were adding
>> 'PAGE_SIZE', rather than 'PAGE_SIZE - 1'.
> 
> Didn't you find it strange that it's so consistently wrong? After a
> closer inspection, you'd find that the end of the range is inclusive. So
> the math is correct and your patch introduces a bug.

But since we are talking about number of pages, why does the range need
to be inclusive. Indeed, let's take delalloc_to_write in
writepage_delalloc. Say we have a 1mb range, 0-1m - that's 256 pages
right so with DIV_ROUND_UP we'll have:

(1048576 + 4096 - 1) / 4096 = 256

with the old version we'll have:

1048576 + 4096 / 4096 = 257

Since delalloc_to_write is used in a context where we care about the
number of pages written I think using DIV_ROUND_UP is correct and it
fixes an off-by-one bug, no ?

Let's take extent_write_locked_range, it's called only from
submit_compressed_extents with the range for the async (compressed) extent:

Say we have an extent which is 2k in ram size and starts at 1m offset.
We'll have:
start = 1048576
end = 1048576 + 2048 - 1 = 1050623

nr_pages in this case should be 1, with DIV_ROUND_UP:

nr_pages = 2047 + 4096 -1 / 4096 = 1 (1,4995 actually but due to integer
math we don't care about floating point part)

with old formula:

nr_pages = (2047 + 4096) / 4096 = 1 (1,4998 but ditto as above).


> 
> end - start + PAGE_SIZE = end + 1 - start + PAGE_SIZE - 1
> 
> Check eg. writepage_delalloc how it sets up page_end.
> 
> The correct use in DIV_ROUND_UP needs +1 adjustment.
>