[dpdk-dev] [PATCH v2] ring: use aligned memzone allocation

[Olivier Matz]

On Thu, 8 Jun 2017 14:20:52 +0100, Bruce Richardson <bruce.richardson at intel.com> wrote:
> On Thu, Jun 08, 2017 at 02:45:40PM +0200, Olivier Matz wrote:
> > On Tue, 6 Jun 2017 15:56:28 +0100, Bruce Richardson <bruce.richardson at intel.com> wrote:  
> > > On Tue, Jun 06, 2017 at 02:19:21PM +0100, Ananyev, Konstantin wrote:  
> > > > 
> > > >     
> > > > > -----Original Message-----
> > > > > From: Richardson, Bruce
> > > > > Sent: Tuesday, June 6, 2017 1:42 PM
> > > > > To: Ananyev, Konstantin <konstantin.ananyev at intel.com>
> > > > > Cc: Verkamp, Daniel <daniel.verkamp at intel.com>; dev at dpdk.org
> > > > > Subject: Re: [dpdk-dev] [PATCH v2] ring: use aligned memzone allocation
> > > > > 
> > > > > On Tue, Jun 06, 2017 at 10:59:59AM +0100, Ananyev, Konstantin wrote:    
> > > > > >    
> > > > > > > >
> > > > > > > >
> > > > > > > >    
> > > > > > > > >
> > > > > > > > > The PROD/CONS_ALIGN values on x86-64 are set to 2 cache lines, so members    
> > > > > > > > of struct rte_ring are 128 byte aligned,    
> > > > > > > > >and therefore the whole struct needs 128-byte alignment according to the ABI    
> > > > > > > > so that the 128-byte alignment of the fields can be guaranteed.
> > > > > > > >
> > > > > > > > Ah ok, missed the fact that rte_ring is 128B aligned these days.
> > > > > > > > BTW, I probably missed the initial discussion, but what was the reason for that?
> > > > > > > > Konstantin    
> > > > > > >
> > > > > > > I don't know why PROD_ALIGN/CONS_ALIGN use 128 byte alignment; it seems unnecessary if the cache line is only 64 bytes.  An    
> > > > > alternate    
> > > > > > > fix would be to just use cache line alignment for these fields (since memzones are already cache line aligned).    
> > > > > >
> > > > > > Yes, had the same thought.
> > > > > >    
> > > > > > > Maybe there is some deeper  reason for the >= 128-byte alignment logic in rte_ring.h?    
> > > > > >
> > > > > > Might be, would be good to hear opinion the author of that change.    
> > > > > 
> > > > > It gives improved performance for core-2-core transfer.    
> > > > 
> > > > You mean empty cache-line(s) after prod/cons, correct?
> > > > That's ok but why we can't keep them and whole rte_ring aligned on cache-line boundaries?
> > > > Something like that:
> > > > struct rte_ring {
> > > >    ...
> > > >    struct rte_ring_headtail prod __rte_cache_aligned;
> > > >    EMPTY_CACHE_LINE   __rte_cache_aligned;
> > > >    struct rte_ring_headtail cons __rte_cache_aligned;
> > > >    EMPTY_CACHE_LINE   __rte_cache_aligned;
> > > > };
> > > > 
> > > > Konstantin    
> > > 
> > > Sure. That should probably work too. 
> > > 
> > > /Bruce  
> > 
> > I also agree with Konstantin's proposal. One question though: since it
> > changes the alignment constraint of the rte_ring structure, I think it is
> > an ABI breakage: a structure including the rte_ring structure inherits
> > from this constraint.
> > 
> > How could we handle that, knowing this is probably a rare case?
> > 
> >  
> Is it an ABI break so long as we keep the resulting size and field
> placement of the structures the same? The alignment being reduced should
> not be a problem, as 128byte alignment is also valid as 64byte
> alignment, after all.

I'd say yes. Consider the following example:

---8<---
#include <stdio.h>
#include <stdlib.h>

#define ALIGN 64
/* #define ALIGN 128 */

/* dummy rte_ring struct */
struct rte_ring {
	char x[128];
} __attribute__((aligned(ALIGN)));

struct foo {
	struct rte_ring r;
	unsigned bar;
};

int main(void)
{
	struct foo array[2];

	printf("sizeof(ring)=%zu diff=%u\n",
		sizeof(struct rte_ring),
		(unsigned int)((char *)&array[1].r - (char *)array));

	return 0;
}
---8<---

The size of rte_ring is always 128.
diff is 192 or 256, depending on the value of ALIGN.



Olivier