Re: Bloat caused by unnecessary calls to compound_head()?

From: Kirill A. Shutemov
Date: Sun Mar 27 2016 - 15:47:17 EST

On Sat, Mar 26, 2016 at 01:50:49PM -0500, Eric Biggers wrote:
> Hi,
> I noticed that after the recent "page-flags" patchset, there are an excessive
> number of calls to compound_head() in certain places.
> For example, the frequently executed mark_page_accessed() function already
> starts out by calling compound_head(), but then each time it tests a page flag
> afterwards, there is an extra, seemingly unnecessary, call to compound_head().
> This causes a series of instructions like the following to appear no fewer than
> 10 times throughout the function:
> ffffffff81119db4: 48 8b 53 20 mov 0x20(%rbx),%rdx
> ffffffff81119db8: 48 8d 42 ff lea -0x1(%rdx),%rax
> ffffffff81119dbc: 83 e2 01 and $0x1,%edx
> ffffffff81119dbf: 48 0f 44 c3 cmove %rbx,%rax
> ffffffff81119dc3: 48 8b 00 mov (%rax),%rax
> Part of the problem, I suppose, is that the compiler doesn't know that the pages
> can't be linked more than one level deep.
> Is this a known tradeoff, and have any possible solutions been considered?

<I'm sick, so my judgment may be off>

Yes, it's known problem. And I've tried to approach it few times without
satisfying results.

Your mail made me try again.

The idea is to introduce new type to indicate head page --
'struct head_page' -- it's compatible with struct page on memory layout,
but distinct from C point of view. compound_head() should return pointer
of that type. For the proof-of-concept I've introduced new helper --

Then we can make page-flag helpers to accept both types, by converting
them to macros and use __builtin_types_compatible_p().

When a page-flag helper sees pointer to 'struct head_page' as an argument,
it can safely assume that it deals with head or non-compound page and therefore
can bypass all policy restrictions and get rid of compound_head() calls.

I'll send proof-of-concept patches in reply to this message. The code is
not pretty. I myself consider the idea rather ugly.

Any comments are welcome.

Kirill A. Shutemov