Re: [PATCHv3, RFC 31/34] thp: initial implementation of do_huge_linear_fault()

["Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>]

Dave Hansen wrote:
> On 04/18/2013 09:09 AM, Kirill A. Shutemov wrote:
> > Dave Hansen wrote:
> >> On 04/17/2013 07:38 AM, Kirill A. Shutemov wrote:
> >> Are you still sure you can't do _any_ better than a verbatim copy of 129
> >> lines?
> > 
> > It seems I was too lazy. Shame on me. :(
> > Here's consolidated version. Only build tested. Does it look better?
> 
> Yeah, it's definitely a step in the right direction.  There rae
> definitely some bugs in there like:
> 
> +	unsigned long haddr = address & PAGE_MASK;

It's not bug. It's bad name for the variable.
See, first 'if (try_huge_pages)'. I update it there for huge page case.

addr_aligned better?

> 
> I do think some of this refactoring stuff
> 
> > -				unlock_page(page);
> > -				vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
> > -				tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
> > -				if (unlikely(tmp &
> > -					  (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
> > -					ret = tmp;
> > +			unlock_page(page);
> > +			vmf.flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE;
> > +			tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
> > +			if (unlikely(tmp &
> > +					(VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
> > +				ret = tmp;
> > +				goto unwritable_page;
> > +			}
> 
> could probably be a separate patch and would make what's going on more
> clear, but it's passable the way it is.  When it is done this way it's
> hard sometimes reading the diff to realize if you are adding code or
> just moving it around.

Will do.

> 
> This stuff:
> 
> >  		if (set_page_dirty(dirty_page))
> > -			dirtied = 1;
> > +			dirtied = true;
> 
> needs to go in another patch for sure.

Ditto.

> One thing I *REALLY* like about doing patches this way is that things
> like this start to pop out:
> 
> > -	ret = vma->vm_ops->fault(vma, &vmf);
> > +	if (try_huge_pages) {
> > +		pgtable = pte_alloc_one(mm, haddr);
> > +		if (unlikely(!pgtable)) {
> > +			ret = VM_FAULT_OOM;
> > +			goto uncharge_out;
> > +		}
> > +		ret = vma->vm_ops->huge_fault(vma, &vmf);
> > +	} else
> > +		ret = vma->vm_ops->fault(vma, &vmf);
> 
> The ->fault is (or can be) essentially per filesystem, and we're going
> to be adding support per-filesystem.  any reason we can't just handle
> this inside the ->fault code and avoid adding huge_fault altogether?

will check. it's on my todo list already.

-- 
 Kirill A. Shutemov