Re: [dpdk-dev] [PATCH v7 05/12] net/ether: mark ethernet addresses as being 2-byte aligned

["Richardson, Bruce" <bruce.richardson@intel.com>]

> -----Original Message-----
> From: Olivier Matz [mailto:olivier.matz@6wind.com]
> Sent: Friday, July 5, 2019 3:34 PM
> To: Stephen Hemminger <stephen@networkplumber.org>
> Cc: dev@dpdk.org; Richardson, Bruce <bruce.richardson@intel.com>; Andrew
> Rybchenko <arybchenko@solarflare.com>
> Subject: Re: [dpdk-dev] [PATCH v7 05/12] net/ether: mark ethernet
> addresses as being 2-byte aligned
> 
> On Tue, Jul 02, 2019 at 03:12:40PM -0700, Stephen Hemminger wrote:
> > From: Bruce Richardson <bruce.richardson@intel.com>
> >
> > When including the rte_ether.h header in applications with warnings
> > enabled, a warning was given because of the assumption of 2-byte
> > alignment of ethernet addresses when processing them.
> >
> > .../include/rte_ether.h:149:2: warning: converting a packed ‘const
> >   struct ether_addr’ pointer (alignment 1) to a ‘unaligned_uint16_t’
> >   {aka ‘const short unsigned int’} pointer (alignment 2) may result in
> >   an unaligned pointer value [-Waddress-of-packed-member]
> > 149 |  const unaligned_uint16_t *ea_words = (const unaligned_uint16_t
> *)ea;
> >     |  ^~~~~
> >
> > Since ethernet addresses should always be aligned on a two-byte
> > boundary, we can just inform the compiler of this assumption to remove
> > the warnings and allow us to always access the addresses using 16-bit
> operations.
> >
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
> > Reviewed-by: Andrew Rybchenko <arybchenko@solarflare.com>
> > ---
> >  lib/librte_net/rte_ether.h | 11 ++++++-----
> >  1 file changed, 6 insertions(+), 5 deletions(-)
> >
> > diff --git a/lib/librte_net/rte_ether.h b/lib/librte_net/rte_ether.h
> > index feb35a33c94b..d7b76ddf63eb 100644
> > --- a/lib/librte_net/rte_ether.h
> > +++ b/lib/librte_net/rte_ether.h
> > @@ -58,7 +58,8 @@ extern "C" {
> >   * See http://standards.ieee.org/regauth/groupmac/tutorial.html
> >   */
> >  struct rte_ether_addr {
> > -	uint8_t addr_bytes[RTE_ETHER_ADDR_LEN]; /**< Addr bytes in tx order
> */
> > +	uint8_t addr_bytes[RTE_ETHER_ADDR_LEN] __rte_aligned(2);
> > +	/**< Addr bytes in tx order */
> >  } __attribute__((__packed__));
> >
> >  #define RTE_ETHER_LOCAL_ADMIN_ADDR 0x02 /**< Locally assigned Eth.
> > address. */ @@ -81,8 +82,8 @@ struct rte_ether_addr {  static inline
> > int rte_is_same_ether_addr(const struct rte_ether_addr *ea1,
> >  				     const struct rte_ether_addr *ea2)  {
> > -	const unaligned_uint16_t *w1 = (const uint16_t *)ea1;
> > -	const unaligned_uint16_t *w2 = (const uint16_t *)ea2;
> > +	const uint16_t *w1 = (const uint16_t *)ea1;
> > +	const uint16_t *w2 = (const uint16_t *)ea2;
> >
> >  	return ((w1[0] ^ w2[0]) | (w1[1] ^ w2[1]) | (w1[2] ^ w2[2])) == 0;
> > } @@ -99,7 +100,7 @@ static inline int rte_is_same_ether_addr(const
> > struct rte_ether_addr *ea1,
> >   */
> >  static inline int rte_is_zero_ether_addr(const struct rte_ether_addr
> > *ea)  {
> > -	const unaligned_uint16_t *w = (const uint16_t *)ea;
> > +	const uint16_t *w = (const uint16_t *)ea;
> >
> >  	return (w[0] | w[1] | w[2]) == 0;
> >  }
> > @@ -146,7 +147,7 @@ static inline int rte_is_multicast_ether_addr(const
> struct rte_ether_addr *ea)
> >   */
> >  static inline int rte_is_broadcast_ether_addr(const struct
> > rte_ether_addr *ea)  {
> > -	const unaligned_uint16_t *ea_words = (const unaligned_uint16_t *)ea;
> > +	const uint16_t *ea_words = (const uint16_t *)ea;
> >
> >  	return (ea_words[0] == 0xFFFF && ea_words[1] == 0xFFFF &&
> >  		ea_words[2] == 0xFFFF);
> > --
> > 2.20.1
> >
> 
> Following this discussion:
> https://mails.dpdk.org/archives/dev/2019-July/136590.html
> 
> I still think that changing the ABI without deprecation notice
> is not a good idea.
> 
I'm ok with that. Let's put in a deprecation notice and take this
patch in 19.11.

> The warning issued by the compiler makes me think that the definition of
> unaligned_uint16_t is wrong on intel arch. I made a quick test, and it
> seems that in this particular case, the generated code is the same with
> or without __attribute__((aligned(1))). See: https://godbolt.org/z/NjBNQk
> 
> But changing the definition of unaligned_uint16_t without a deprecation
> notice is not an option either.
> 
> What do you think about using a specific typedef similar to
> unaligned_uint16_t in rte_ether, that has the __attribute__((aligned(1)))
> ?
> It would avoid to change the alignment of struct rte_ether_addr.
> 

I'd like the alignment changed. Since the existing warnings about alignment don’t
seem to be causing anyone any real problems, I suggest we just leave them for
now and fix them by changing the alignment setting for 19.11.

> In parallel, we can talk about changing unaligned_uint16_t for intel
> in another patchset.
> 

Yes, let's fix the broken unaligned definition as a separate issue.

/Bruce