Re: [PATCH 1/6] block: Introduce bio_for_each_page()

[Kent Overstreet <kmo@daterainc.com>]

On Wed, Sep 25, 2013 at 02:17:02PM -0700, Zach Brown wrote:
> >  void zero_fill_bio(struct bio *bio)
> >  {
> > -	unsigned long flags;
> >  	struct bio_vec bv;
> >  	struct bvec_iter iter;
> >  
> > -	bio_for_each_segment(bv, bio, iter) {
> > +#if defined(CONFIG_HIGHMEM) || defined(ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE)
> > +	bio_for_each_page(bv, bio, iter) {
> > +		unsigned long flags;
> >  		char *data = bvec_kmap_irq(&bv, &flags);
> >  		memset(data, 0, bv.bv_len);
> >  		flush_dcache_page(bv.bv_page);
> >  		bvec_kunmap_irq(data, &flags);
> >  	}
> > +#else
> > +	bio_for_each_segment(bv, bio, iter)
> > +		memset(page_address(bv.bv_page) + bv.bv_offset,
> > +		       0, bv.bv_len);
> > +#endif
> 
> This looks pretty sketchy.  I'd expect this to be doable with one loop
> and that bvec_kmap_irq() and flush_dcache_page() would fall back to
> page_address() and nops when they're not needed.
> 
> Where did this come from?

It's just that if we need the kmap or the flush_dcache_page we have to
process the bio one 4k page at a time - if not, we can process 64k (or
whatever) bvecs all at once. That doesn't just save us memcpy calls, we
can also avoid all the machinery in bio_for_each_page() for chunking up
large bvecs into single page bvecs.

I can definitely think of better ways to do this, but I figured I'd wait
and see if other code ends up wanting to switch between
bio_for_each_page() and bio_for_each_segment() and why.