[PATCH v2 2/2] mm: speed up mremap by 500x on large regions

[Joel Fernandes]

On Fri, Oct 12, 2018 at 04:19:46PM +0300, Kirill A. Shutemov wrote:
> On Fri, Oct 12, 2018 at 05:50:46AM -0700, Joel Fernandes wrote:
> > On Fri, Oct 12, 2018 at 02:30:56PM +0300, Kirill A. Shutemov wrote:
> > > On Thu, Oct 11, 2018 at 06:37:56PM -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.
> > > 
> > > I looked into the code more and noticed move_pte() helper called from
> > > move_ptes(). It changes PTE entry to suite new address.
> > > 
> > > It is only defined in non-trivial way on Sparc. I don't know much about
> > > Sparc and it's hard for me to say if the optimization will break anything
> > > there.
> > 
> > Sparc's move_pte seems to be flushing the D-cache to prevent aliasing. It is
> > not modifying the PTE itself AFAICS:
> > 
> > #ifdef DCACHE_ALIASING_POSSIBLE
> > #define __HAVE_ARCH_MOVE_PTE
> > #define move_pte(pte, prot, old_addr, new_addr)                         \
> > ({                                                                      \
> >         pte_t newpte = (pte);                                           \
> >         if (tlb_type != hypervisor && pte_present(pte)) {               \
> >                 unsigned long this_pfn = pte_pfn(pte);                  \
> >                                                                         \
> >                 if (pfn_valid(this_pfn) &&                              \
> >                     (((old_addr) ^ (new_addr)) & (1 << 13)))            \
> >                         flush_dcache_page_all(current->mm,              \
> >                                               pfn_to_page(this_pfn));   \
> >         }                                                               \
> >         newpte;                                                         \
> > })
> > #endif
> > 
> > If its an issue, then how do transparent huge pages work on Sparc?  I don't
> > see the huge page code (move_huge_pages) during mremap doing anything special
> > for Sparc architecture when moving PMDs..
> 
> My *guess* is that it will work fine on Sparc as it apprarently it only
> cares about change in bit 13 of virtual address. It will never happen for
> huge pages or when PTE page tables move.
> 
> But I just realized that the problem is bigger: since we pass new_addr to
> the set_pte_at() we would need to audit all implementations that they are
> safe with just moving PTE page table.
> 
> I would rather go with per-architecture enabling. It's much safer.

I'm Ok with the per-arch enabling, I agree its safer. So I should be adding a
a new __HAVE_ARCH_MOVE_PMD right, or did you have a better name for that?

Also, do you feel we should still need to remove the address argument from
set_pte_alloc? Or should we leave that alone if we do per-arch?
I figure I spent a bunch of time on that already anyway, and its a clean up
anyway, so may as well do it. But perhaps that "pte_alloc cleanup" can then
be a separate patch independent of this series?

> > Also, do we not flush the caches from any path when we munmap address space?
> > We do call do_munmap on the old mapping from mremap after moving to the new one.
> 
> Are you sure about that? It can be hided deeper in architecture-specific
> code.

I am sure we do call do_munmap, I was asking if we flush the caches as well.
If we're enabling this per architecture, then I guess it does not matter for
the purposes of this patch.

thanks,

 - Joel