On Fri, Apr 28, 2017 at 11:21:27AM +0200, Thomas Monjalon wrote:
> 28/04/2017 11:03, Bruce Richardson:
> > On Fri, Apr 28, 2017 at 10:56:56AM +0200, Thomas Monjalon wrote:
> > > 28/04/2017 10:15, Bruce Richardson:
> > > > On i686 builds, the uint64_t type is 64-bits in size but is aligned to
> > > > 32-bits only. This causes mbuf fields for rearm_data to not be 16-byte
> > > > aligned on 32-bit builds, which causes errors with some vector PMDs
> > > > which
> > > > expect the rearm data to be aligned as on 64-bit.
> > > >
> > > > Given that we cannot use the extra space in the data structures anyway,
> > > > as
> > > > it's already used on 64-bit builds, we can just force alignment of
> > > > physical
> > > > address structure members to 8-bytes in all cases. This has no effect on
> > > > 64-bit systems, but fixes the updated PMDs on 32-bit.
> > >
> > > I agree to align on 64-bit in mbuf.
> > >
> > > > Fixes: f4356d7ca168 ("net/i40e: eliminate mbuf write on rearm")
> > > > Fixes: f160666a1073 ("net/ixgbe: eliminate mbuf write on rearm")
> > > [...]
> > > > --- a/lib/librte_eal/common/include/rte_memory.h
> > > > +++ b/lib/librte_eal/common/include/rte_memory.h
> > > > -typedef uint64_t phys_addr_t; /**< Physical address definition. */
> > > > +/** Physical address definition. */
> > > > +typedef uint64_t phys_addr_t __rte_aligned(sizeof(uint64_t));
> > >
> > > Why setting this constraint for everyone?
> > >
> > Well, it only has an effect on 32-bit builds, and unless there is a
> > problem, I don't see why not always align them to the extra 8 bytes. If
> > this does cause an issue, I'm happy enough to use #ifdefs, but in the
> > absense of a confirmed problem, I'd rather keep the code clean.
>
> Is it expected for everyone to have every physical addresses aligned on 64?
> I think it can be weird for some applications.
> Why do you think it is cleaner than adding the alignment to the mbuf fields?
>
I'm ok to redo the patch to only make the change to the mbuf value.
However, when researching this, I discovered that gcc apparently already
aligns all non-structure-member uint64_t values on an 8-byte boundary on
32-bit x86 anyway*. [Don't know if this also applies e.g. to 32-bit arm,
but I wouldn't be surprised if it did.] That means the scope of this
only applies to structures with phys_addr values, so it's not a huge
scope.
*Ref: https://gcc.gnu.org/ml/gcc/2009-06/msg00333.html
> PS: It is yet another macro which is not rte_ prefixed.
>
Yes. Not going to fix that in this patch though!
So, do you want a V2 to limit the alignment change to the phys_addr in
the mbuf, rather than generally to physical addresses? I prefer the way
I have it here, but I'm ok to change.
Regards,
/Bruce
