> > @@ -324,6 +325,13 @@ void force_page_cache_ra(struct readahead_control *ractl,
> > * be up to the optimal hardware IO size
> > */
> > index = readahead_index(ractl);
> > + if (!IS_ALIGNED(index, min_nrpages)) {
> > + unsigned long old_index = index;
> > +
> > + index = round_down(index, min_nrpages);
> > + nr_to_read += (old_index - index);
> > + }
>
> new_index = mapping_align_start_index(mapping, index);
> if (new_index != index) {
> nr_to_read += index - new_index;
> index = new_index
Looks good.
> }
>
> > +
> > max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
> > nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
>
> This needs to have a size of at least the minimum folio order size
> so readahead can fill entire folios, not get neutered to the maximum
> IO size the underlying storage supports.
So something like:
> > max_pages = max_t(unsigned long, bdi->io_pages, ra->ra_pages);
> > nr_to_read = min_t(unsigned long, nr_to_read, max_pages);
nr_to_read = max(nr_to_read, min_order);
>
> > + * For higher order address space requirements we ensure no initial reads
> > + * are ever less than the min number of pages required.
> > + *
> > + * We *always* cap the max io size allowed by the device.
> > */
> > -static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
> > +static unsigned long get_init_ra_size(unsigned long size,
> > + unsigned int min_nrpages,
> > + unsigned long max)
> > {
> > unsigned long newsize = roundup_pow_of_two(size);
> >
> > + newsize = max_t(unsigned long, newsize, min_nrpages);
>
> This really doesn't need to care about min_nrpages. That rounding
> can be done in the caller when the new size is returned.
Sounds good.
>
> > if (newsize <= max / 32)
> > newsize = newsize * 4;
>
> >
> >
> > @@ -561,7 +583,11 @@ static void ondemand_readahead(struct readahead_control *ractl,
> > unsigned long add_pages;
> > pgoff_t index = readahead_index(ractl);
> > pgoff_t expected, prev_index;
> > - unsigned int order = folio ? folio_order(folio) : 0;
> > + unsigned int min_order = mapping_min_folio_order(ractl->mapping);
> > + unsigned int min_nrpages = mapping_min_folio_nrpages(ractl->mapping);
> > + unsigned int order = folio ? folio_order(folio) : min_order;
>
> Huh? If we have a folio, then the order is whatever that folio is,
> otherwise we use min_order. What if the folio is larger than
> min_order? Doesn't that mean that this:
>
> > @@ -583,8 +609,8 @@ static void ondemand_readahead(struct readahead_control *ractl,
> > expected = round_down(ra->start + ra->size - ra->async_size,
> > 1UL << order);
> > if (index == expected || index == (ra->start + ra->size)) {
> > - ra->start += ra->size;
> > - ra->size = get_next_ra_size(ra, max_pages);
> > + ra->start += round_down(ra->size, min_nrpages);
> > + ra->size = get_next_ra_size(ra, min_nrpages, max_pages);
>
> may set up the incorrect readahead range because the folio order is
> larger than min_nrpages?
Hmm... So I think we should just increment ra->start by ra->size, and
make sure to round the new size we get from get_next_ra_size() to
min_nrpages. Then we will not disturb the readahead range and always
increase the range in multiples of min_nrpages:
ra->start += ra->size;
ra->size = round_up(get_next_ra_size(ra, max_pages), min_nrpages);
>
> > ra->async_size = ra->size;
> > goto readit;
> > }
> > @@ -603,13 +629,18 @@ static void ondemand_readahead(struct readahead_control *ractl,
> > max_pages);
> > rcu_read_unlock();
> >
> > + start = round_down(start, min_nrpages);
>
> start = mapping_align_start_index(mapping, start);
> > +
> > + VM_BUG_ON(folio->index & (folio_nr_pages(folio) - 1));
> > +
> > if (!start || start - index > max_pages)
> > return;
> >
> > ra->start = start;
> > ra->size = start - index; /* old async_size */
> > +
> > ra->size += req_size;
> > - ra->size = get_next_ra_size(ra, max_pages);
> > + ra->size = get_next_ra_size(ra, min_nrpages, max_pages);
>
> ra->size = max(min_nrpages, get_next_ra_size(ra, max_pages));
If this is a round_up of size instead of max operation, we can
always ensure the ra->start from index aligned to min_nrpages. See my
reasoning in the previous comment.
--
Pankaj
