Re: [PATCH v2 15/16] slab: Allocate frozen pages

[Mel Gorman <mgorman@techsingularity.net>]

On Wed, Aug 10, 2022 at 02:31:11PM +0200, Vlastimil Babka wrote:
> On 8/9/22 19:18, Matthew Wilcox (Oracle) wrote:
> > Since slab does not use the page refcount, it can allocate and
> > free frozen pages, saving one atomic operation per free.
> > 
> > Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
> > Reviewed-by: William Kucharski <william.kucharski@oracle.com>
> 
> AFAICS the problem of has_unmovable_pages() is not addressed:
> https://lore.kernel.org/all/40d658da-6220-e05e-ba0b-d95c82f6bfb3@redhat.com/
> 
> But I don't think it's sustainable approach to enhance the checks there with
> PageSlab() and then with whatever other user will adopt allocating frozen
> pages in the future. I guess it would be better to just be able to detect
> pages on pcplist without false positives. A new page type? Maybe the
> overhead of managing it would be negligible as we set page->index anyway for
> migratetype?
> 

I think page type would be usable to identify a PCP page same as how
it's used to identify a buddy page. Most likely, this could be done in
check_pcp_refill, check_new_pcp (watch DEBUG_VM) and free_pcppages_bulk.
There would be a race between the last refcount being dropped and becoming
a PCP page but I doubt that matters to page isolation as I expect it retries.

The __Clear and __Set operations would add some overhead but it's almost
certainly cheaper than the put_page_testzero in __free_pages().

-- 
Mel Gorman
SUSE Labs