Re: [PATCH] mm: Speed up mremap on large regions

From: Joel Fernandes
Date: Tue Oct 09 2018 - 19:04:53 EST


On Wed, Oct 10, 2018 at 01:02:22AM +0300, Kirill A. Shutemov wrote:
> On Tue, Oct 09, 2018 at 01:14:00PM -0700, Joel Fernandes (Google) wrote:
> > Android needs to mremap large regions of memory during memory management
> > related operations. The mremap system call can be really slow if THP is
> > not enabled. The bottleneck is move_page_tables, which is copying each
> > pte at a time, and can be really slow across a large map. Turning on THP
> > may not be a viable option, and is not for us. This patch speeds up the
> > performance for non-THP system by copying at the PMD level when possible.
> >
> > The speed up is three orders of magnitude. On a 1GB mremap, the mremap
> > completion times drops from 160-250 millesconds to 380-400 microseconds.
> >
> > Before:
> > Total mremap time for 1GB data: 242321014 nanoseconds.
> > Total mremap time for 1GB data: 196842467 nanoseconds.
> > Total mremap time for 1GB data: 167051162 nanoseconds.
> >
> > After:
> > Total mremap time for 1GB data: 385781 nanoseconds.
> > Total mremap time for 1GB data: 388959 nanoseconds.
> > Total mremap time for 1GB data: 402813 nanoseconds.
> >
> > Incase THP is enabled, the optimization is skipped. I also flush the
> > tlb every time we do this optimization since I couldn't find a way to
> > determine if the low-level PTEs are dirty. It is seen that the cost of
> > doing so is not much compared the improvement, on both x86-64 and arm64.
>
> Okay. That's interesting.
>
> It makes me wounder why do we pass virtual address to pte_alloc() (and
> pte_alloc_one() inside).
>
> If an arch has real requirement to tight a page table to a virtual address
> than the optimization cannot be used as it is. Per-arch should be fine
> for this case, I guess.
>
> If nobody uses the address we should just drop the argument as a
> preparation to the patch.

I couldn't find any use of the address. But I am wondering why you feel
passing the address is something that can't be done with the optimization.
The pte_alloc only happens if the optimization is not triggered.

Also the clean up of the argument that you're proposing is a bit out of scope
of this patch but yeah we could clean it up in a separate patch if needed. I
don't feel too strongly about that. It seems cosmetic and in the future if
the address that's passed in is needed, then the architecture can use it.

> Anyway, I think the optimization requires some groundwork before it can be
> accepted. At least some explanation why it is safe to move page table from
> one spot in virtual address space to another.

So I did go through several scenarios and its fine to my eyes. I tested
it too and couldn't find any issue. Could you describe your concern a bit
more? The mm->mmap_sem lock is held through out the mremap. Further we are
acquiring needed rmap locks if needed and the ptl locks of the old new
page-table pages. And this same path is already copying pmds for hugepages.

thanks,

- Joel