[dpdk-dev] [PATCH] Fix KNI compiling issue on IBM Power

[Thomas Monjalon]

2014-12-04 10:32, Neil Horman:
> On Thu, Dec 04, 2014 at 02:47:03PM +0100, Thomas Monjalon wrote:
> > 2014-12-04 08:29, Neil Horman:
> > > On Thu, Dec 04, 2014 at 12:59:31PM +0100, Thomas Monjalon wrote:
> > > > > Because of different cache line size, the alignment of struct
> > > > > rte_kni_mbuf in rte_kni_common.h doesn't work on IBM Power. This patch
> > > > > changed from 64 to RTE_CACHE_LINE_SIZE micro to do the alignment.
> > > > > 
> > > > > Signed-off-by: Chao Zhu <chaozhu at linux.vnet.ibm.com>
> > > > 
> > > > Acked-by: Thomas Monjalon <thomas.monjalon at 6wind.com>
> > > > 
> > > > Applied
> > > > 
> > > Woah!  Slow down here, I'm not sure if this makes sense to fix his way.  The
> > > exact same ifndef/define/endif construct is used for this macro in rte_memory.h.
> > > Currently their defined to the same vaule, but if that ever changes, this macro
> > > will return different values based on the order in which header files are
> > > included.  That doesn't seem appropriate at all.
> > 
> > I agree (was my comment) but the patch was applied as a hot fix.
> > A better fix has to be found for DPDK 2.0.
> > Do you agree this fix is enough for DPDK 1.8 release?
> > 
> I really don't like the idea of hacks like this being used.

It's not really a hack to replace a hardcoded value by a constant.
I think you should agree it's better (but not perfect).

> Truthfully, I would rather the KNI just not be built on power for now,
> it is after all a new feature for which not everything works yet (e.g. the
> acl library and the ixgbe rxtx vec code).
> With this in place, KNI will build now, but it means that anything
> changes cache line sizes until it gets fixed properly runs the risk of
> introducing wierd behavioral issues at compile time.

It was also the case before: 64 was hardcoded for KNI.

> I'm also concerned about the fact that, since we have no bug tracker for DPDK,
> indicating that there will be an improved fix in 2.0 isn't really a guarantee,
> in that it requires that someone remember to do it.

Please be confident that I keep it noted and I'll do what I can to have it
properly fixed.
By the way, submitting a fix now would store the need in patchwork.

> > > > I wonder if we could try to guess the cache line size instead of
> > > > configuring it in many places.
> > > > Maybe we could use something like sysconf(_SC_LEVEL1_DCACHE_LINESIZE)?
> > > > 
> > > This is a good idea, but I think its a bit broken for a few reasons:
> > > 
> > > 1) _SC_LEVEL1_DCACHE_LINESIZE I don't think is POSIX mandated, so there is every
> > > possibility that the above won't work on BSD
> > > 
> > > 2) While getting the cache line size dynamically is a great idea, dpdk has
> > > several locations that size structures based on processor cache line size, which
> > > implicitly requires a static cache line definition.
> > 
> > It can be guessed dynamically in the first build step (kind of configure).
> > 
> That would work, though that seems like cause to really start redesigning the
> build system to use autoconf/automake so we can run utilities to do that sort of
> thing more easily (not opposed to that mind you, just illustrating that its more
> work)

I'm convinced we need to work on the build system but it's another discussion
for next weeks. Speaking about that, the AF_PACKET PMD cannot be enabled because
dependencies are not checked before building it.

> > > It seems the right thing to do, in my mind is to define RTE_CACHE_LINE_SIZE per
> > > arch (perhaps in common/include/arch/<arch>/rte_<something>.h), then just let
> > > the build break if a given arch doesn't define it (i.e. make definig that value
> > > an arch reqirement).
> > 
> > It's the other option. For IBM Power, it's currently overwritten in the Makefile:
> > 	http://dpdk.org/browse/dpdk/tree/mk/arch/ppc_64/rte.vars.mk
> > 
> Thats a sensible solution in my mind, though it is limited by the assumption
> that any given arch has only a single cache line size (I dno't think thats a
> problem, but it might be).  If it is, the dynamic solution above is superior.

I think we won't solve the hypothetical problem of heterogeneous CPUs in
first step. I'd like to start with your proposal of a arch variable.

-- 
Thomas