[dm-devel] [PATCH] vmalloc: introduce vmap_pfn for persistent memory

[Mikulas Patocka]

On Wed, 8 Nov 2017, Christoph Hellwig wrote:

> On Tue, Nov 07, 2017 at 05:03:11PM -0500, Mikulas Patocka wrote:
> > Hi
> > 
> > I am developing a driver that uses persistent memory for caching. A 
> > persistent memory device can be mapped in several discontiguous ranges.
> > 
> > The kernel has a function vmap that takes an array of pointers to pages 
> > and maps these pages to contiguous linear address space. However, it can't 
> > be used on persistent memory because persistent memory may not be backed 
> > by page structures.
> > 
> > This patch introduces a new function vmap_pfn, it works like vmap, but 
> > takes an array of pfn_t - so it can be used on persistent memory.
> 
> How is cache flushing going to work for this interface assuming
> that your write to/from the virtual address and expect it to be
> persisted on pmem?

We could use the function clwb() (or arch-independent wrapper dax_flush()) 
- that uses the clflushopt instruction on Broadwell or clwb on Skylake - 
but it is very slow, write performance on Broadwell is only 350MB/s.

So in practice I use the movnti instruction that bypasses cache. The 
write-combining buffer is flushed with sfence.

Mikulas