Re: [PATCH] fat: Avoid oops when bdi->io_pages==0

[OGAWA Hirofumi]

Jens Axboe <axboe@xxxxxxxxx> writes:

> On 8/31/20 10:56 AM, Matthew Wilcox wrote:
>> On Mon, Aug 31, 2020 at 10:39:26AM -0600, Jens Axboe wrote:
>>> We really should ensure that ->io_pages is always set, imho, instead of
>>> having to work-around it in other spots.
>> 
>> Interestingly, there are only three places in the entire kernel which
>> _use_ bdi->io_pages.  FAT, Verity and the pagecache readahead code.
>> 
>> Verity:
>>                         unsigned long num_ra_pages =
>>                                 min_t(unsigned long, num_blocks_to_hash - i,
>>                                       inode->i_sb->s_bdi->io_pages);
>> 
>> FAT:
>>         if (ra_pages > sb->s_bdi->io_pages)
>>                 ra_pages = rounddown(ra_pages, sb->s_bdi->io_pages);
>> 
>> Pagecache:
>>         max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
>> and
>>         if (req_size > max_pages && bdi->io_pages > max_pages)
>>                 max_pages = min(req_size, bdi->io_pages);
>> 
>> The funny thing is that all three are using it differently.  Verity is
>> taking io_pages to be the maximum amount to readahead.  FAT is using
>> it as the unit of readahead (round down to the previous multiple) and
>> the pagecache uses it to limit reads that exceed the current per-file
>> readahead limit (but allows per-file readahead to exceed io_pages,
>> in which case it has no effect).
>> 
>> So how should it be used?  My inclination is to say that the pagecache
>> is right, by virtue of being the most-used.
>
> When I added ->io_pages, it was for the page cache use case. The others
> grew after that...

FAT and pagecache usage would be similar or same purpose. The both is
using io_pages as optimal IO size.

In pagecache case, it uses io_pages if one request size is exceeding
io_pages. In FAT case, there is perfect knowledge about future/total
request size. So FAT divides request by io_pages, and adjust ra_pages
with knowledge.

I don't know about verity.

Thanks.
-- 
OGAWA Hirofumi <hirofumi@xxxxxxxxxxxxxxxxxx>