lkml.org 
[lkml]   [2014]   [Jul]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH -mm] mm: refactor page index/offset getters
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


\
 
 \ /
  Last update: 2014-07-29 03:21    [W:0.093 / U:0.776 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site