Re: [PATCH 2/3] zram: try vmalloc() after kmalloc()

[Minchan Kim <minchan@kernel.org>]

On Tue, Nov 24, 2015 at 03:12:58PM +0900, Sergey Senozhatsky wrote:
> [..]
> >  static void *zcomp_lz4_create(void)
> >  {
> > -       return kzalloc(LZ4_MEM_COMPRESS, GFP_KERNEL);
> > +       void *ret;
> > +
> > +       /*
> > +        * This function could be called in swapout/fs write path
> > +        * so we couldn't use GFP_FS|IO. And it assumes we already
> > +        * have at least one stream in zram initialization so we
> > +        * don't do best effort to allocate more stream in here.
> > +        * A default stream will work well without further multiple
> > +        * stream. That's why we use  __GFP_NORETRY|NOWARN|NOMEMALLOC.
> > +        */
> > +       ret = kzalloc(LZ4_MEM_COMPRESS,
> > +                       __GFP_NORETRY|__GFP_NOWARN|__GFP_NOMEMALLOC);
> > +       if (!ret)
> > +               ret = __vmalloc(LZ4_MEM_COMPRESS,
> > +                                       __GFP_NORETRY|__GFP_NOWARN|
> > +                                       __GFP_NOMEMALLOC|__GFP_ZERO,
> > +                                       PAGE_KERNEL);
> > +       return ret;
> >  }
> [..]
> 
> so this change is still questionable. is there a real value in having
> a vmalloc() fallback in the middle of allocations sequence:
> 
> 	zstrm = kmalloc(sizeof(*zstrm), GFP_NOIO);
> 		^^^ ok, can fail here
> 
> 	zstrm->zstrm->private = comp->backend->create();
> 				^^^ kzalloc() and vmalloc() fallback ??
> 
> 	zstrm->buffer = (void *)__get_free_pages(GFP_NOIO | __GFP_ZERO, 1);
> 		^^^ can fail here again.
> 
> can you please comment on this?

Good question.

Actually, failure of allocation came from backend->create as Kyeongdon's
comment because it requires order-3 allocation which is very fragile
in embedded system recenlty(Android, webOS. That's why Joonsoo are solving
the trouble by fixing compaction part). Otherwise, other kmalloc/vmalloc
stuff in our example would be almost no trouble in real practice(Of course,
if you says it might, you're absolutely right. It could fail but I think
it's *really* rare in real practice).

More concern is order-1 allocation rather than kmalloc/vmalloc.
I've got lots of allocation failure reports from product team until now
and frankly speaking, I don't get such order-1 fail report until now.
I guess the reason is that system is almost trobule due to heavy fragmentation
before the notice such failure.

So, I think if we solve order-3 allocation in backend->create,
above problem will be almost solved.

> 
> 
> and I'd prefer it to be a bit different -- use likely path first and
> avoid an assignment in unlikely path.

Personally, I like one return case unless other is really better for
performance. I have trained it for Andrew, I belive. :)
But if you don't like this by performance reason, I will add unlikely
for vmalloc path. If you hate it just by personal preferenece, please
I want to stick my code.


> ... and add GFP_NOIO to both kzalloc() and __vmalloc().

I can add it. The harmness is really ignorable but as I mentioned
at reply of Andrew, what's the benefit with GFP_NOIO?
We couldn't make forward progress with __GFP_RECLAIM in reclaim
context.

> 
> and there is no __GFP_HIGHMEM in __vmalloc() call?

Good to have. Thanks for the hint!

Thanks.

> 
> something like this:
> 
> ---
> 
> 
> 	ret = kzalloc(LZ4_MEM_COMPRESS, GFP_NOIO | __GFP_NORETRY |
> 					__GFP_NOWARN | __GFP_NOMEMALLOC);
> 	if (ret)
> 		return ret;
> 
> 	return __vmalloc(LZ4_MEM_COMPRESS,
> 			GFP_NOIO | __GFP_NOWARN | __GFP_HIGHMEM | __GFP_ZERO,
> 			PAGE_KERNEL);
> 
> 
> 	-ss

-- 
Kind regards,
Minchan Kim