On Tue, 7 Apr 2020, Andy Lutomirski wrote:

> 
> > On Apr 7, 2020, at 8:01 AM, Mikulas Patocka <mpatocka@redhat.com> wrote:
> > 
> > [ resending this to x86 maintainers ]
> > 
> > Hi
> > 
> > I tested performance of various methods how to write to optane-based
> > persistent memory, and found out that non-temporal stores achieve 
> > throughput 1.3 GB/s. 8 cached stores immediatelly followed by clflushopt 
> > or clwb achieve throughput 1.6 GB/s.
> > 
> > memcpy_flushcache uses non-temporal stores, I modified it to use cached 
> > stores + clflushopt and it improved performance of the dm-writecache 
> > target significantly:
> > 
> > dm-writecache throughput:
> > (dd if=/dev/zero of=/dev/mapper/wc bs=64k oflag=direct)
> > writecache block size   512             1024            2048            4096
> > movnti                  496 MB/s        642 MB/s        725 MB/s        744 MB/s
> > clflushopt              373 MB/s        688 MB/s        1.1 GB/s        1.2 GB/s
> > 
> > For block size 512, movnti works better, for larger block sizes, 
> > clflushopt is better.
> > 
> > I was also testing the novafs filesystem, it is not upstream, but it 
> > benefitted from similar change in __memcpy_flushcache and 
> > __copy_user_nocache:
> > write throughput on big files - movnti: 662 MB/s, clwb: 1323 MB/s
> > write throughput on small files - movnti: 621 MB/s, clwb: 1013 MB/s
> > 
> > 
> > I submit this patch for __memcpy_flushcache that improves dm-writecache 
> > performance.
> > 
> > Other ideas - should we introduce memcpy_to_pmem instead of modifying 
> > memcpy_flushcache and move this logic there? Or should I modify the 
> > dm-writecache target directly to use clflushopt with no change to the 
> > architecture-specific code?
> > 
> > Mikulas
> > 
> > 
> > 
> > 
> > From: Mikulas Patocka <mpatocka@redhat.com>
> > 
> > I tested dm-writecache performance on a machine with Optane nvdimm and it
> > turned out that for larger writes, cached stores + cache flushing perform
> > better than non-temporal stores. This is the throughput of dm-writecache
> > measured with this command:
> > dd if=/dev/zero of=/dev/mapper/wc bs=64 oflag=direct
> > 
> > block size    512        1024        2048        4096
> > movnti        496 MB/s    642 MB/s    725 MB/s    744 MB/s
> > clflushopt    373 MB/s    688 MB/s    1.1 GB/s    1.2 GB/s
> > 
> > We can see that for smaller block, movnti performs better, but for larger
> > blocks, clflushopt has better performance.
> > 
> > This patch changes the function __memcpy_flushcache accordingly, so that
> > with size >= 768 it performs cached stores and cache flushing. Note that
> > we must not use the new branch if the CPU doesn't have clflushopt - in
> > that case, the kernel would use inefficient "clflush" instruction that has
> > very bad performance.
> > 
> > Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
> > 
> > ---
> > arch/x86/lib/usercopy_64.c |   36 ++++++++++++++++++++++++++++++++++++
> > 1 file changed, 36 insertions(+)
> > 
> > Index: linux-2.6/arch/x86/lib/usercopy_64.c
> > ===================================================================
> > --- linux-2.6.orig/arch/x86/lib/usercopy_64.c    2020-03-24 15:15:36.644945091 -0400
> > +++ linux-2.6/arch/x86/lib/usercopy_64.c    2020-03-30 07:17:51.450290007 -0400
> > @@ -152,6 +152,42 @@ void __memcpy_flushcache(void *_dst, con
> >            return;
> >    }
> > 
> > +    if (static_cpu_has(X86_FEATURE_CLFLUSHOPT) && size >= 768 && likely(boot_cpu_data.x86_clflush_size == 64)) {
> > +        while (!IS_ALIGNED(dest, 64)) {
> > +            asm("movq    (%0), %%r8\n"
> > +                "movnti  %%r8,   (%1)\n"
> > +                :: "r" (source), "r" (dest)
> > +                : "memory", "r8");
> > +            dest += 8;
> > +            source += 8;
> > +            size -= 8;
> > +        }
> > +        do {
> > +            asm("movq    (%0), %%r8\n"
> > +                "movq   8(%0), %%r9\n"
> > +                "movq  16(%0), %%r10\n"
> > +                "movq  24(%0), %%r11\n"
> > +                "movq    %%r8,   (%1)\n"
> > +                "movq    %%r9,  8(%1)\n"
> > +                "movq   %%r10, 16(%1)\n"
> > +                "movq   %%r11, 24(%1)\n"
> > +                "movq  32(%0), %%r8\n"
> > +                "movq  40(%0), %%r9\n"
> > +                "movq  48(%0), %%r10\n"
> > +                "movq  56(%0), %%r11\n"
> > +                "movq    %%r8, 32(%1)\n"
> > +                "movq    %%r9, 40(%1)\n"
> > +                "movq   %%r10, 48(%1)\n"
> > +                "movq   %%r11, 56(%1)\n"
> > +                :: "r" (source), "r" (dest)
> > +                : "memory", "r8", "r9", "r10", "r11");
> 
> Does this actually work better than the corresponding C code?
> 
> Also, that memory clobber probably isn’t doing your code generation any 
> favors.  Experimentally, you have the constraints wrong. An “r” 

The existing "movnti" loop uses exactly the same constraints (and the 
"memory" clobber).

> constraint doesn’t tell GCC that you are dereferencing the pointer.  
> You need to use “m” with a correctly-sized type.

But you would use
	"=m"(*(char *)dest),"=m"(*((char *)dest + 8)),"=m"((char *)dest + 16))...
and so on, until you run out of argument numbers.

> But I bet plain C is at least as good.

I tried to replace it with
	memcpy((void *)dest, (void *)src, 64);

The compiler inlined the memcpy function into 8 loads and 8 stores. 
However, the whole function __memcpy_flushcache consumed one more saved 
register and the machine code was a few bytes longer.

Mikulas