On 04/17/2013 07:38 AM, Kirill A. Shutemov wrote:
>> > Ugh.  This is essentially a copy-n-paste of code in __do_fault(),
>> > including the comments.  Is there no way to consolidate the code so that
>> > there's less duplication here?
> I've looked into it once again and it seems there's not much space for
> consolidation. Code structure looks very similar, but there are many
> special cases for thp: fallback path, pte vs. pmd, etc. I don't see how we
> can consolidate them in them in sane way.
> I think copy is more maintainable :(

I took the two copies, put them each in a file, changed some of the
_very_ trivial stuff to match (foo=1 vs foo=true) and diffed them.
They're very similar lengths (in lines):

 185  __do_fault
 197  do_huge_linear_fault

If you diff them:

 1 file changed, 68 insertions(+), 56 deletions(-)

That means that of 185 lines in __do_fault(), 129 (70%) of them were
copied *VERBATIM*.  Not similar in structure or appearance.  Bit-for-bit
the same.

I took a stab at consolidating them.  I think we could add a
VM_FAULT_FALLBACK flag to explicitly indicate that we need to do a
huge->small fallback, as well as a FAULT_FLAG_TRANSHUGE to indicate that
a given fault has not attempted to be handled by a huge page.  If we
call __do_fault() with FAULT_FLAG_TRANSHUGE and we get back
VM_FAULT_FALLBACK or VM_FAULT_OOM, then we clear FAULT_FLAG_TRANSHUGE
and retry in handle_mm_fault().

I only went about 1/4 of the way in to __do_fault().  If went and spent
another hour or two, I'm pretty convinced I could push this even further.

Are you still sure you can't do _any_ better than a verbatim copy of 129
lines?