[Linux-cachefs] [RFC PATCH 1/7] cifs: Transition from ->readpages() to ->readahead()

[David Howells]

Matthew Wilcox <willy at infradead.org> wrote:

> On Tue, Jan 25, 2022 at 01:57:44PM +0000, David Howells wrote:
> > +	while (readahead_count(ractl) - ractl->_batch_count) {
> 
> You do understand that prefixing a structure member with an '_' means
> "Don't use this", right?  If I could get the compiler to prevent you, I
> would.

Yes, I know.  However, as previously discussed, I think that your
implementation of readahead batching doesn't work right - hence the need to
apply compensation to the values returned by the accessor functions.

Btw, I end up doing this:

		for (i = 0; i < nr_pages; i++)
			if (!readahead_folio(ractl))
				BUG();

in patch 5.  I want to create a batch, but I don't want to be given the array
of addresses of the folios as I'm going to use an xarray-class iterator.
Further, _batch_count at this point is some value related to just the last
folio and not the batch as a whole:-/

(Also, the above won't work if any folios retrieved are larger than a page)

Note that cifs_readahead() is removed in patch 7 and readahead functionality
is offloaded to netfslib, so I'm not sure it's worth spending much time on
fixing.

[I should mention that netfs_readahead() also does:

	while (readahead_folio(ractl))
		;
which could probably be replaced with something better that doesn't keep
taking and dropping the RCU readlock.]

Would you object if I added a function like:

	static inline
	unsigned int readahead_commit_batch(struct readahead_control *rac)
	{
		BUG_ON(rac->_batch_count > rac->_nr_pages);
		rac->_nr_pages -= rac->_batch_count;
		rac->_index += rac->_batch_count;
		rac->_batch_count = 0;
	}

It could then be called from both __readahead_folio() and __readahead_batch().
For __readahead_folio(), the duplicate setting of _batch_count should be
optimised away on the path where a folio is returned.  I could then call this
from the loop in cifs before going round again.

I'd also like to consider adding something like:

	static inline
	void readahead_set_batch(struct readahead_control *rac)
	{
		unsigned int i = 0;
		struct page *page;
		XA_STATE(xas, &rac->mapping->i_pages, 0);

		BUG_ON(rac->_batch_count > rac->_nr_pages);
		rac->_nr_pages -= rac->_batch_count;
		rac->_index += rac->_batch_count;
		rac->_batch_count = 0;

		xas_set(&xas, rac->_index);
		rcu_read_lock();
		xas_for_each(&xas, page, rac->_index + rac->_nr_pages - 1) {
			if (xas_retry(&xas, page))
				continue;
			VM_BUG_ON_PAGE(!PageLocked(page), page);
			VM_BUG_ON_PAGE(PageTail(page), page);
			rac->_batch_count += thp_nr_pages(page);
		}
		rcu_read_unlock();
	}

so that netfslib can use it to load all the pages it is given into a batch
without retrieving the page pointers.

David