On Mon, Apr 19, 2021 at 12:36:19PM +0300, Mike Rapoport wrote:

Well, most if the -4.2% of the performance regression kbuild reported were
due to repeated compount_head(page) in page_mapping(). So the whole point
of this patch is to avoid calling page_mapping().

It's quite ludicrous how many times we call compound_head() in
page_mapping() today:

 page = compound_head(page);
 if (__builtin_expect(!!(PageSlab(page)), 0))
 if (__builtin_expect(!!(PageSwapCache(page)), 0)) {

TESTPAGEFLAG(Slab, slab, PF_NO_TAIL) expands to:

static __always_inline int PageSlab(struct page *page)
{
	PF_POISONED_CHECK(compound_head(page));
	return test_bit(PG_slab, &compound_head(page));
}

static __always_inline int PageSwapCache(struct page *page)
{
        page = compound_head(page);
        return PageSwapBacked(page) && test_bit(PG_swapcache, &page->flags);
}

but then!

TESTPAGEFLAG(SwapBacked, swapbacked, PF_NO_TAIL) also expands like Slab does.

So that's six calls to compound_head(), depending what Kconfig options
you have enabled.

And folio_mapping() is one of the functions I add in the first batch of
patches, so review, etc will be helpful.