RE: [PATCH v3 net-next] net: mana: Add a driver for Microsoft Azure Network Adapter (MANA)

[Haiyang Zhang]

> -----Original Message-----
> From: David Miller <davem@xxxxxxxxxxxxx>
> Sent: Thursday, April 8, 2021 8:41 PM
> To: Dexuan Cui <decui@xxxxxxxxxxxxx>
> Cc: kuba@xxxxxxxxxx; KY Srinivasan <kys@xxxxxxxxxxxxx>; Haiyang Zhang
> <haiyangz@xxxxxxxxxxxxx>; Stephen Hemminger
> <sthemmin@xxxxxxxxxxxxx>; wei.liu@xxxxxxxxxx; Wei Liu
> <liuwe@xxxxxxxxxxxxx>; netdev@xxxxxxxxxxxxxxx; leon@xxxxxxxxxx;
> andrew@xxxxxxx; bernd@xxxxxxxxxxxxxxxxxxx; rdunlap@xxxxxxxxxxxxx; linux-
> kernel@xxxxxxxxxxxxxxx; linux-hyperv@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH v3 net-next] net: mana: Add a driver for Microsoft Azure
> Network Adapter (MANA)
> 
> From: Dexuan Cui <decui@xxxxxxxxxxxxx>
> Date: Fri, 9 Apr 2021 00:24:51 +0000
> 
> >> From: David Miller <davem@xxxxxxxxxxxxx>
> >> Sent: Thursday, April 8, 2021 4:46 PM
> >> ...
> >> > +struct gdma_msg_hdr {
> >> > +	u32 hdr_type;
> >> > +	u32 msg_type;
> >> > +	u16 msg_version;
> >> > +	u16 hwc_msg_id;
> >> > +	u32 msg_size;
> >> > +} __packed;
> >> > +
> >> > +struct gdma_dev_id {
> >> > +	union {
> >> > +		struct {
> >> > +			u16 type;
> >> > +			u16 instance;
> >> > +		};
> >> > +
> >> > +		u32 as_uint32;
> >> > +	};
> >> > +} __packed;
> >>
> >> Please don't  use __packed unless absolutely necessary.  It generates
> >> suboptimal code (byte at a time
> >> accesses etc.) and for many of these you don't even need it.
> >
> > In the driver code, all the structs/unions marked by __packed are used to
> > talk with the hardware, so I think __packed is necessary here?
> 
> It actually isan't in many cases, check with and without the __packed
> directive
> and see if anything chasnges.
> 
> > Do you think if it's better if we remove all the __packed, and add
> > static_assert(sizeof(struct XXX) == YYY) instead? e.g.
> >
> > @@ -105,7 +105,8 @@ struct gdma_msg_hdr {
> >         u16 msg_version;
> >         u16 hwc_msg_id;
> >         u32 msg_size;
> > -} __packed;
> > +};
> > +static_assert(sizeof(struct gdma_msg_hdr) == 16);
> 
> This won't make sure the structure member offsets are what you expect.
> 
> I think you'll have to go through the structures one-by-one by hand to
> figure out which ones really require the __packed attribute and which do not.

For the structs containing variables with the same sizes, or already size aligned 
variables, we knew the __packed has no effect. And for these structs, it doesn't 
cause performance impact either, correct? 

But in the future, if different sized variables are added, the __packed may 
become necessary again. To prevent anyone accidently forget to add __packed 
when adding new variables to these structs, can we keep the __packed for all 
messages going through the "wire"?

Thanks,
- Haiyang