On Tue, May 30, 2023 at 08:39:32AM +1000, Dave Chinner wrote:
> > Option 5: If the folio is both dirty and !uptodate, just refuse to split
> > it, like if somebody else had a reference on it. A less extreme version
> > of #4.
>
> Also seems like a reasonable first step.
>
> > I may have missed some other option. Option 5 seems like the least
> > amount of work.
>
> *nod*
>
> Overall, I think the best way to approach it is to do the simplest,
> most obviously correct thing first. If/when we observe performance
> problems from the simple approach, then we can decide how to solve
> that via one of the more complex approaches...
So the good news is that Option 5 seems to have no regressions and
the functional part of the patch looks like ...
+++ b/fs/iomap/buffered-io.c
@@ -510,11 +510,6 @@ void iomap_invalidate_folio(struct folio *folio, size_t off
set, size_t len)
WARN_ON_ONCE(folio_test_writeback(folio));
folio_cancel_dirty(folio);
iomap_page_release(folio);
- } else if (folio_test_large(folio)) {
- /* Must release the iop so the page can be split */
- VM_WARN_ON_ONCE_FOLIO(!folio_test_uptodate(folio) &&
- folio_test_dirty(folio), folio);
- iomap_page_release(folio);
}
}
EXPORT_SYMBOL_GPL(iomap_invalidate_folio);
I need to modify a few comments and document exactly why this works,
but it seems like a good next step.
