Re: [PATCH iwl-next 01/12] libeth: add cacheline / struct alignment helpers

[Przemek Kitszel]

On 5/30/24 03:34, Jakub Kicinski wrote:
> On Tue, 28 May 2024 15:48:35 +0200 Alexander Lobakin wrote:
> > diff --git a/scripts/kernel-doc b/scripts/kernel-doc
> > index 95a59ac78f82..d0cf9a2d82de 100755
> > --- a/scripts/kernel-doc
> > +++ b/scripts/kernel-doc
> > @@ -1155,6 +1155,7 @@ sub dump_struct($$) {
> >           $members =~ s/\bstruct_group_attr\s*\(([^,]*,){2}/STRUCT_GROUP(/gos;
> >           $members =~ s/\bstruct_group_tagged\s*\(([^,]*),([^,]*),/struct $1 $2; STRUCT_GROUP(/gos;
> >           $members =~ s/\b__struct_group\s*\(([^,]*,){3}/STRUCT_GROUP(/gos;
> > +        $members =~ s/\blibeth_cacheline_group\s*\(([^,]*,)/struct { } $1; STRUCT_GROUP(/gos;
> >           $members =~ s/\bSTRUCT_GROUP(\(((?:(?>[^)(]+)|(?1))*)\))[^;]*;/$2/gos;
> >   
> >           my $args = qr{([^,)]+)};
>
> Having per-driver grouping defines is a no-go.

[1]

> Do you need the defines in the first place?

this patch was a tough one for me too, but I see the idea promising

> Are you sure the assert you're adding are not going to explode
> on some weird arch? Honestly, patch 5 feels like a little too
> much for a driver..

definitively some of the patch 5 should be added here as doc/example,
but it would be even better to simplify a bit

--

I think that "mark this struct as explicit split into cachelines" is
a hard hard C problem in general, especially in the kernel context, 
*but* I think that this could be simplified for your use case - split
into exactly 3 (possibly empty) sections: mostly-Read, RW, COLD?

Given that it will be a generic solution (would fix the [1] above),
and be also easier to use, like:

 CACHELINE_STRUCT_GROUP(idpf_q_vector,
	CACHELINE_STRUCT_GROUP_RD(/* read mostly */
		struct idpf_vport *vport;
		u16 num_rxq;
		u16 num_txq;
		u16 num_bufq;
		u16 num_complq;
		struct idpf_rx_queue **rx;
		struct idpf_tx_queue **tx;
		struct idpf_buf_queue **bufq;
		struct idpf_compl_queue **complq;
		struct idpf_intr_reg intr_reg;
	),
	CACHELINE_STRUCT_GROUP_RW(
		struct napi_struct napi;
		u16 total_events;
		struct dim tx_dim;
		u16 tx_itr_value;
		bool tx_intr_mode;
		u32 tx_itr_idx;
		struct dim rx_dim;
		u16 rx_itr_value;
		bool rx_intr_mode;
		u32 rx_itr_idx;
	),
	CACHELINE_STRUCT_GROUP_COLD(
		u16 v_idx;
		cpumask_var_t affinity_mask;
	)
);

Note that those three inner macros have distinct meaningful names not to
have this working, but to aid human reader, then checkpatch/check-kdoc.
Technically could be all the same CACHELINE_GROUP().

I'm not sure if (at most) 3 cacheline groups are fine for the general
case, but it would be best to have just one variant of the
CACHELINE_STRUCT_GROUP(), perhaps as a vararg.