Re: [PATCH v2 06/22] slab: Reimplement page_slab()

[Vlastimil Babka <vbabka@suse.cz>]

On 9/11/25 00:46, Christoph Lameter (Ampere) wrote:
> On Wed, 10 Sep 2025, Matthew Wilcox (Oracle) wrote:
> 
>> In order to separate slabs from folios, we need to convert from any page
>> in a slab to the slab directly without going through a page to folio
>> conversion first.  page_slab() is a little different from other memdesc
>> converters we have in that it will return NULL if the page is not part
>> of a slab.  This will be the normal style for memdesc converters in
>> the future.
> 
> We are looking forward to another level of indirection in order
> to get to struct slab?
> 
>> -#define page_slab(p)		(_Generic((p),				\
>> -	const struct page *:	(const struct slab *)(p),		\
>> -	struct page *:		(struct slab *)(p)))
>> +static inline struct slab *page_slab(const struct page *page)
>> +{
>> +	unsigned long head;
>> +
>> +	head = READ_ONCE(page->compound_head);
>> +	if (head & 1)
>> +		page = (struct page *)(head - 1);
>> +	if (data_race(page->page_type >> 24) != PGTY_slab)
>> +		page = NULL;
>> +
>> +	return (struct slab *)page;
>> +}
> 
> compound_head() is open coded in there.

But since this is replacing going through page_folio() (in the following
patches) that does compound_head() itself, we're not really adding a new
indirection, no? Except to kfence here which knows the page pointer is head
already and now we're adding an unnecessary compound_head(). But it affects
only kfence_init_pool() and not any fastpaths, AFAICS.

I guess later with memdescs we can have a scheme where all head and tail
memdescs could contain the struct slab pointer so we don't have go through
the head one first? Which would replace one indirection with another and not
add a new one either.