Re: [PATCH v1 2/6] fs: use on-stack-bio if backing device has BDI_CAP_SYNC capability

[Dan Williams]

On Wed, Aug 9, 2017 at 8:04 PM, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> On Wed, Aug 09, 2017 at 11:41:50AM +0900, Minchan Kim wrote:
>> On Tue, Aug 08, 2017 at 07:31:22PM -0700, Matthew Wilcox wrote:
>> > On Wed, Aug 09, 2017 at 10:51:13AM +0900, Minchan Kim wrote:
>> > > On Tue, Aug 08, 2017 at 06:29:04AM -0700, Matthew Wilcox wrote:
>> > > > On Tue, Aug 08, 2017 at 05:49:59AM -0700, Matthew Wilcox wrote:
>> > > > > +     struct bio sbio;
>> > > > > +     struct bio_vec sbvec;
>> > > >
>> > > > ... this needs to be sbvec[nr_pages], of course.
>> > > >
>> > > > > -             bio = mpage_alloc(bdev, blocks[0] << (blkbits - 9),
>> > > > > +             if (bdi_cap_synchronous_io(inode_to_bdi(inode))) {
>> > > > > +                     bio = &sbio;
>> > > > > +                     bio_init(bio, &sbvec, nr_pages);
>> > > >
>> > > > ... and this needs to be 'sbvec', not '&sbvec'.
>> > >
>> > > I don't get it why we need sbvec[nr_pages].
>> > > On-stack-bio works with per-page.
>> > > May I miss something?
>> >
>> > The way I redid it, it will work with an arbitrary number of pages.
>>
>> IIUC, it would be good things with dynamic bio alloction with passing
>> allocated bio back and forth but on-stack bio cannot work like that.
>> It should be done in per-page so it is worth?
>
> I'm not passing the bio back and forth between do_mpage_readpage() and
> its callers.  The version I sent allows for multiple pages in a single
> on-stack bio (when called from mpage_readpages()).

I like it, but do you think we should switch to sbvec[<constant>] to
preclude pathological cases where nr_pages is large?