page_mapping vs page_file_mapping for swap cache pages

page_mapping vs page_file_mapping for swap cache pages

[Matthew Wilcox]

If we call page_mapping(page) and PageSwapCache is true, we do this:

        if (unlikely(PageSwapCache(page))) {
                swp_entry_t entry;

                entry.val = page_private(page);
                return swap_address_space(entry);
        }

#define swap_address_space(entry)                           \
        (&swapper_spaces[swp_type(entry)][swp_offset(entry) \
                >> SWAP_ADDRESS_SPACE_SHIFT])

(i think we could make that more readable by adding
#define swp_as(entry)	(swp_offset(entry) >> SWAP_ADDRESS_SPACE_SHIFT)
but i digress)

If, instead, we call page_file_mapping(page) and PageSwapCache is true,
we do this:

        return page_swap_info(page)->swap_file->f_mapping;

struct swap_info_struct *page_swap_info(struct page *page)
{
        swp_entry_t entry = { .val = page_private(page) };
        return swp_swap_info(entry);
}

struct swap_info_struct *swp_swap_info(swp_entry_t entry)
{
        return swap_type_to_swap_info(swp_type(entry));
}

static struct swap_info_struct *swap_type_to_swap_info(int type)
{
        if (type >= READ_ONCE(nr_swapfiles))
                return NULL;

        smp_rmb();      /* Pairs with smp_wmb in alloc_swap_info. */
        return READ_ONCE(swap_info[type]);
}

So ... are these different address spaces from each other?  If not,
why do we have such a complicated way of finding the address space
for page_file_mapping()?

Re: page_mapping vs page_file_mapping for swap cache pages

[Huang, Ying]

Matthew Wilcox <willy@infradead.org> writes:

> If we call page_mapping(page) and PageSwapCache is true, we do this:
>
>         if (unlikely(PageSwapCache(page))) {
>                 swp_entry_t entry;
>
>                 entry.val = page_private(page);
>                 return swap_address_space(entry);
>         }
>
> #define swap_address_space(entry)                           \
>         (&swapper_spaces[swp_type(entry)][swp_offset(entry) \
>                 >> SWAP_ADDRESS_SPACE_SHIFT])
>
> (i think we could make that more readable by adding
> #define swp_as(entry)	(swp_offset(entry) >> SWAP_ADDRESS_SPACE_SHIFT)
> but i digress)

Yes.  This sounds good.

> If, instead, we call page_file_mapping(page) and PageSwapCache is true,
> we do this:
>
>         return page_swap_info(page)->swap_file->f_mapping;
>
> struct swap_info_struct *page_swap_info(struct page *page)
> {
>         swp_entry_t entry = { .val = page_private(page) };
>         return swp_swap_info(entry);
> }
>
> struct swap_info_struct *swp_swap_info(swp_entry_t entry)
> {
>         return swap_type_to_swap_info(swp_type(entry));
> }
>
> static struct swap_info_struct *swap_type_to_swap_info(int type)
> {
>         if (type >= READ_ONCE(nr_swapfiles))
>                 return NULL;
>
>         smp_rmb();      /* Pairs with smp_wmb in alloc_swap_info. */
>         return READ_ONCE(swap_info[type]);
> }
>
> So ... are these different address spaces from each other?

Yes.  They are different.  The address space from page_mapping(page) is
used to hold pages in swap cache, they are not associated with some
inodes in fact.  While the address space from page_file_mapping(page) is
used to read/write swap pages sometimes (please check
__swap_writepage()).  But page_file_mapping() isn't called there.

Best Regards,
Huang, Ying

> If not, why do we have such a complicated way of finding the address
> space for page_file_mapping()?