[PATCH V2] powerpc/infiniband: Use cache inhibitted and guarded mapping on powerpc

Aneesh Kumar K.V aneesh.kumar at linux.vnet.ibm.com
Mon Apr 25 03:33:10 AEST 2016


Benjamin Herrenschmidt <benh at kernel.crashing.org> writes:

> On Fri, 2016-04-22 at 10:09 -0600, Jason Gunthorpe wrote:
>> On Fri, Apr 22, 2016 at 08:47:56AM +1000, Benjamin Herrenschmidt
>> wrote:
>> > 
>> > On Wed, 2016-04-20 at 03:58 -0400, Aneesh Kumar K.V wrote:
>> > > 
>> > > The driver was requesting for a writethrough mapping. But with
>> > > thoses
>> > > flags we will end up with a SAO mapping because we now have
>> > > memory
>> > > conherence always enabled. ie, the existing mapping will end up
>> > > with
>> > > a WIMG value 0b1110 which is Strong Access Order.
>> > > 
>> > > Update this to use cache inhibitted guarded mapping
>> > Why guarded ? If it's performance sensitive (and the driver has
>> > appropriate barriers where needed), you will get write combining
>> > without guarded, you won't with it.
>> This driver uses uncached write combining on x86
>
> Right so if you want something similar on power, use the _wc variant,
> and make sure you have the appropriate barriers when you need to
> break combining.
>
> Cheers,
> Ben.

The code after change looks like below

#if defined(__powerpc__)
	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
#endif

	/*
	 * don't allow them to later change to readable with mprotect (for when
	 * not initially mapped readable, as is normally the case)
	 */
	vma->vm_flags &= ~VM_MAYREAD;
	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;

	/* We used PAT if wc_cookie == 0 */
	if (!dd->wc_cookie)
		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);

So it uses pgprot_noncached_wc if (!dd->wc_cookie). Hence I was thinking
it needs more stricter mapping otherwise.

-aneesh



More information about the Linuxppc-dev mailing list