Re: [PATCH 3/4] usb: allocated usb releated dma buffer with kmalloc_node

["Yinghai Lu" <yhlu.kernel@gmail.com>]

On 7/2/07, Greg KH <greg@kroah.com> wrote:
> On Mon, Jul 02, 2007 at 03:36:37PM -0700, Yinghai Lu wrote:
> > [PATCH 3/4] usb: allocated usb releated dma buffer with kmalloc_node
> >
> > For amd64 based two way system. USB always on node0. but dma buffer for urb
> > allocated via kmalloc always get ram on node1. So change to kmalloc_node to
> > get dma_buffer on corresponding node
>
> Are all of these changes really necessary?  You are doing this for some
> allocations that take a _long_ time when sending to the device due to
> the speed of the device.
>
> I could possibly see this making a difference on some drivers, but for
> the core, and for the basic USB structures, I can't imagine it is really
> worth it.
>
> Or do you have numbers showing the differences here?
>
> Patch included fully below for the benifit of the usb list, which you
> should have cc:ed...

dma buffer could be allocated via alloc_pages_coherent. or
kmalloc/dma_map_single.
alloc_pages_coherent get the dma_buffer on corresponding node.
but kmalloc/dma_map_single always get dma_buffer on last node. or say
device is on HT chain node0,  it will get dma buffer on node 7 of 8
socket system.
also on two way system with 4G+4G RAM conf. device on node 0 will get
dma_buffer above 4G, and if the dma_mask is less 32bit, will need
extra iommu mapping.
In my mcp55+io55 system, it show dma_map_single is keepping called by
usb input: keyboard/mouse (8/0x40 bytes), and forcedeth. (0x670bytes)

YH