Re: [PATCH v7 5/6] serial: Support for RS-485 multipoint addresses

[Ilpo Järvinen]

On Thu, 16 Jun 2022, Jiri Slaby wrote:

> On 16. 06. 22, 7:04, Ilpo Järvinen wrote:
> > On Wed, 15 Jun 2022, Andy Shevchenko wrote:
> > 
> > > On Wed, Jun 15, 2022 at 03:48:28PM +0300, Ilpo Järvinen wrote:
> > > > Add support for RS-485 multipoint addressing using 9th bit [*]. The
> > > > addressing mode is configured through .rs485_config().
> > > > 
> > > > ADDRB in termios indicates 9th bit addressing mode is enabled. In this
> > > > mode, 9th bit is used to indicate an address (byte) within the
> > > > communication line. ADDRB can only be enabled/disabled through
> > > > .rs485_config() that is also responsible for setting the destination and
> > > > receiver (filter) addresses.
> > > > 
> > > > [*] Technically, RS485 is just an electronic spec and does not itself
> > > > specify the 9th bit addressing mode but 9th bit seems at least
> > > > "semi-standard" way to do addressing with RS485.
> > > > 
> > > > Cc: Arnd Bergmann <arnd@xxxxxxxx>
> > > > Cc: Jonathan Corbet <corbet@xxxxxxx>
> > > > Cc: linux-api@xxxxxxxxxxxxxxx
> > > > Cc: linux-doc@xxxxxxxxxxxxxxx
> > > > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > > > Cc: linux-arch@xxxxxxxxxxxxxxx
> > > 
> > > Hmm... In order to reduce commit messages you can move these Cc:s after
> > > the
> > > cutter line ('---').
> > 
> > Ok, although the toolchain I use didn't support preserving --- content
> > so I had to create hack to preserve them, hopefully nothing backfires due
> > to the hack. :-)
> > 
> > > > -	__u32	padding[5];		/* Memory is cheap, new structs
> > > > -					   are a royal PITA .. */
> > > > +	__u8	addr_recv;
> > > > +	__u8	addr_dest;
> > > > +	__u8	padding[2 + 4 * sizeof(__u32)];		/* Memory is cheap,
> > > > new structs
> > > > +							 * are a royal PITA ..
> > > > */
> > > 
> > > I'm not sure it's an equivalent. I would leave u32 members  untouched, so
> > > something like
> > > 
> > > 	__u8	addr_recv;
> > > 	__u8	addr_dest;
> > > 	__u8	padding0[2];		/* Memory is cheap, new structs
> > > 	__u32	padding1[4];		 * are a royal PITA .. */
> > > 
> > > And repeating about `pahole` tool which may be useful here to check for
> > > ABI
> > > potential changes.
> > 
> > I cannot take __u32 padding[] away like that, this is an uapi header.
> 
> Yeah, but it's padding after all. I would personally break it for example as
> Andy suggests (if pahole shows no differences in size on both 32/64 bit) and
> wait if something breaks. To be honest, I'd not expect anyone to touch it. And
> if someone does, we would fix it somehow and they should too...

I realized there are plenty of anonymous unions already in include/uapi/ 
so I think I can keep padding[5] too:

        union {
                /* v1 */
                __u32   padding[5];             /* Memory is cheap, new structs are a pain */

                /* v2 (adds addressing mode fields) */
                struct {
                        __u8    addr_recv;
                        __u8    addr_dest;
                        __u8    padding0[2];
                        __u32   padding1[4];
                };
        };

I'll just skip manual pahole step and add a few BUILD_BUG_ON()s and use 
our build bot to do a quick check over all archs it builds for, that gives 
much better confidence on it being ok:

	BUILD_BUG_ON(((&rs485.delay_rts_after_send) + 1) != &rs485.padding[0]);
	BUILD_BUG_ON(&rs485.padding[1] != &rs485.padding1[0]);
	BUILD_BUG_ON(sizeof(rs485) != ((u8 *)(&rs485.padding[4]) - ((u8 *)&rs485.flags) + sizeof(__u32)));


-- 
 i.