Re: [PATCH -mm] mm: refactor page index/offset getters

[Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>]

On Mon, Jul 28, 2014 at 04:29:52PM -0400, Johannes Weiner wrote:
> On Tue, Jul 15, 2014 at 12:41:12PM -0400, Naoya Horiguchi wrote:
> > @@ -399,28 +399,24 @@ static inline struct page *read_mapping_page(struct address_space *mapping,
> >  }
> >  
> >  /*
> > - * Get the offset in PAGE_SIZE.
> > + * Return the 4kB page offset of the given page.
> >   * (TODO: hugepage should have ->index in PAGE_SIZE)
> >   */
> > -static inline pgoff_t page_to_pgoff(struct page *page)
> > +static inline pgoff_t page_pgoff(struct page *page)
> >  {
> > -	if (unlikely(PageHeadHuge(page)))
> > -		return page->index << compound_order(page);
> > -	else
> > -		return page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
> > +	if (unlikely(PageHuge(page))) {
> > +		VM_BUG_ON_PAGE(PageTail(page), page);
> > +		return page_index(page) << compound_order(page);
> > +	} else
> > +		return page_index(page) << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
> >  }
> 
> I just bisected the VM refusing to swap and triggering OOM kills to
> this patch, which is likely the same bug you reported a couple days
> back when you had this patch in your private tree.

Right, thanks.
And sorry for taking your time for my poor testing.

> Changing page->index to page_index() makes this function return the
> swap offset rather than the virtual PFN, but rmap uses this to index
> Into virtual address space.  Thus, swapcache pages can no longer be
> found from try_to_unmap() and reclaim fails.

I missed the fact that swap code needs both swap offset (stored in
page->private) and virtual PFN (in page->index), so unifying offset
getters to a single helper is completely wrong.

> We can't simply change it back to page->index, however, because the
> swapout path, which requires the swap offset, also uses this function
> through page_offset().  Virtual address space functions and page cache
> address space functions can't use the same helpers, and the helpers
> should likely be named distinctly so that they are not confused and
> it's clear what is being asked.

OK.

>  Plus, the patch forced every fs using
> page_offset() to suddenly check PageHuge(), which is a function call.

OK, PageHuge() check should be done only in vm code.

> How about
> 
> o page_offset() for use by filesystems, based on page->index

And many drivers code and network code use this, so I shouldn't have
touched this :(

> o page_virt_pgoff() for use on virtual memory math, based on
>   page->index and respecting PageHuge()

This is what current code does with page_to_pgoff().

> 
> o page_mapping_pgoff() for use by swapping and when working on
>   mappings that could be swapper_space.

page_file_offset() does this.

So it seems to me OK to just rename them with enough comments.

Thanks,
Naoya Horiguchi

> o page_mapping_offset() likewise, just in bytes

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>