[RFC PATCH 2/2] arm64: Implement vmalloc based thread_info allocator

[sergey.senozhatsky.work@gmail.com (Sergey Senozhatsky)]

On (05/27/15 13:10), Minchan Kim wrote:
> On Tue, May 26, 2015 at 08:29:59PM +0900, Jungseok Lee wrote:
> > On May 25, 2015, at 11:40 PM, Minchan Kim wrote:
> > > Hello Jungseok,
> > 
> > Hi, Minchan,
> > 
> > > On Mon, May 25, 2015 at 01:02:20AM +0900, Jungseok Lee wrote:
> > >> Fork-routine sometimes fails to get a physically contiguous region for
> > >> thread_info on 4KB page system although free memory is enough. That is,
> > >> a physically contiguous region, which is currently 16KB, is not available
> > >> since system memory is fragmented.
> > > 
> > > Order less than PAGE_ALLOC_COSTLY_ORDER should not fail in current
> > > mm implementation. If you saw the order-2,3 high-order allocation fail
> > > maybe your application received SIGKILL by someone. LMK?
> > 
> > Exactly right. The allocation is failed via the following path.
> > 
> > if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
> > 	goto nopage;
> > 
> > IMHO, a reclaim operation would be not needed in this context if memory is
> > allocated from vmalloc space. It means there is no need to traverse shrinker list. 
> 
> For making fork successful with using vmalloc, it's bandaid.
> 
> > 
> > >> This patch tries to solve the problem as allocating thread_info memory
> > >> from vmalloc space, not 1:1 mapping one. The downside is one additional
> > >> page allocation in case of vmalloc. However, vmalloc space is large enough,
> > > 
> > > The size you want to allocate is 16KB in here but additional 4K?
> > > It increases 25% memory footprint, which is huge downside.
> > 
> > I agree with the point, and most people who try to use vmalloc might know the number.
> > However, an interoperation on the number depends on a point of view.
> > 
> > Vmalloc is large enough and not fully utilized in case of ARM64.
> > With the considerations, there is a room to do math as follows.
> > 
> > 4KB / 240GB = 1.5e-8 (4KB page + 3 level combo)
> > 
> > It would be not a huge downside if fork-routine is not damaged due to fragmentation.
> 
> Okay, address size point of view, it wouldn't be significant problem.
> Then, let's see it performance as point of view.
> 
> If we use vmalloc, it needs additional data structure for vmalloc
> management, several additional allocation request, page table hanlding
> and TLB flush.

plus a guard page. I don't see VM_NO_GUARD being passed.

	-ss

> 
> Normally, forking is very frequent operation so we shouldn't do
> make it slow and memory consumption bigger if there isn't big reason.
> 
> > 
> > However, this is one of reasons to add "RFC" prefix in the patch set. How is the
> > additional 4KB interpreted and considered?
> > 
> > Best Regards
> > Jungseok Lee
> 
> -- 
> Kind regards,
> Minchan Kim
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo at kvack.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@kvack.org"> email at kvack.org </a>
>