[PATCH net/next 3/3] net: ethernet: mtk_eth_soc: use genpool allocator for SRAM

[Daniel Golle]

On Sat, Jun 28, 2025 at 10:13:51AM +0200, Andrew Lunn wrote:
> > +static void *mtk_dma_ring_alloc(struct mtk_eth *eth, size_t size,
> > +				dma_addr_t *dma_handle)
> > +{
> > +	void *dma_ring;
> > +
> > +	if (WARN_ON(mtk_use_legacy_sram(eth)))
> > +		return -ENOMEM;
> > +
> > +	if (eth->sram_pool) {
> > +		dma_ring = (void *)gen_pool_alloc(eth->sram_pool, size);
> > +		if (!dma_ring)
> > +			return dma_ring;
> > +		*dma_handle = gen_pool_virt_to_phys(eth->sram_pool, (unsigned long)dma_ring);
> 
> I don't particularly like all the casting backwards and forwards
> between unsigned long and void *. These two APIs are not really
> compatible with each other. So any sort of wrapping is going to be
> messy.
> 
> Maybe define a cookie union:
> 
> struct mtk_dma_cookie {
> 	union {
> 		unsigned long gen_pool;
> 		void *coherent;
> 	}
> }

I've implemented that idea and the diffstat grew quite a lot. Also,
it didn't really make the code more readable (see below why).

> 
> Only dma_handle appears to be used by the rest of the code, so only
> the _alloc and _free need to know about the union.

That's not true. The void* ring->dma is used to access the RX and TX
descriptors, so keeping it void* is useful and using the struct you
are suggesting will make things even more messy than they already are.

See all the places in the code where we assume ring->dma being void*.
Converting all of those to use struct mtk_dma_cookie will not make things
better imho.

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n1337

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n1345

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n1358

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n1804

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n2172

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n2490

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n2638

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n2668

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/mediatek/mtk_eth_soc.c#n2904

I think keeping the two casts in mtk_dma_ring_alloc() and
mtk_dma_ring_free() is the better option.