Re: [PATCH] mm/dmapool: localize page allocations

[Matthew Wilcox]

On Thu, May 17, 2018 at 03:37:21PM -0400, Sinan Kaya wrote:
> On 5/17/2018 2:18 PM, Matthew Wilcox wrote:
> > On Thu, May 17, 2018 at 01:36:19PM -0400, Sinan Kaya wrote:
> >> Try to keep the pool closer to the device's NUMA node by changing kmalloc()
> >> to kmalloc_node() and devres_alloc() to devres_alloc_node().
> > Have you measured any performance gains by doing this?  The thing is that
> > these allocations are for the metadata about the page, and the page is
> > going to be used by CPUs in every node.  So it's not clear to me that
> > allocating it on the node nearest to the device is going to be any sort
> > of a win.
> > 
> 
> It is true that this is metadata but it is one of the things that is most
> frequently used in spite of its small size.
> 
> I don't think it makes any sense to cross a chip boundary for accessing a
> pointer location on every single pool allocation. 
> 
> Remember that the CPU core that is running this driver is most probably on
> the same NUMA node as the device itself.

Umm ... says who?  If my process is running on NUMA node 5 and I submit
an I/O, it should be allocating from a pool on node 5, not from a pool
on whichever node the device is attached to.

If it actually makes a performance difference, then NVMe should allocate
one pool per queue, rather than one pool per device like it currently
does.

> Also, if it was a one time init kind of thing, I'd say "yeah, leave it alone". 
> DMA pool is used by a wide range of drivers and it is used to allocate
> fixed size buffers at runtime. 

 * DMA Pool allocator
 *
 * Copyright 2001 David Brownell
 * Copyright 2007 Intel Corporation
 *   Author: Matthew Wilcox <willy@xxxxxxxxxxxxxxx>

I know what it's used for.