Re: [PATCH net-next 08/13] net: macb: introduce DMA descriptor helpers (is 64bit? is PTP?)

[Théo Lebrun]

Hello Maxime,

On Mon Mar 24, 2025 at 9:55 AM CET, Maxime Chevallier wrote:
> On Fri, 21 Mar 2025 20:09:39 +0100
> Théo Lebrun <theo.lebrun@xxxxxxxxxxx> wrote:
>
>> Introduce macb_dma_is_64b() and macb_dma_is_ptp() helper functions.
>> Many codepaths are made simpler by dropping conditional compilation.
>> 
>> This implies three changes:
>>  - Always compile related structure definitions inside <macb.h>.
>>  - Make the field hw_dma_cap in struct macb always present.
>>  - MACB_EXT_DESC can be dropped as it is useless now.
>> 
>> The common case is:
>> 
>> 	#ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
>> 		struct macb_dma_desc_64 *desc_64;
>> 		if (bp->hw_dma_cap & HW_DMA_CAP_64B) {
>> 			desc_64 = macb_64b_desc(bp, desc);
>> 			// ...
>> 		}
>> 	#endif
>> 
>> And replaced by:
>> 
>> 	struct macb_dma_desc_64 *desc_64;
>> 	if (macb_dma_is_64b(bp)) {
>> 		desc_64 = macb_64b_desc(bp, desc);
>> 		// ...
>> 	}
>
> Just a thought, but this is adding some more branches in the hotpath on
> 32 bits DMA setups. Did you measure any performance changes on
> these platforms (if you have access to one :) )
>
> As the caps can't be changed dynamically, maybe these helpers could be
> replaced more efficiently with some static_key ? This would benefit
> both 32 and 64 bits systems as the following would be more efficient
>
> 	if (bp->hw_dma_cap & HW_DMA_CAP_64B) {
> 		//  ...
> 	}
>
> Just a thought of course, maybe this patch doesn't really hurt perfs :)

Good question! I asked myself the same thing before posting.

We go from:

	void bar(struct macb *bp) {
	#ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
		if (bp->hw_dma_cap & HW_DMA_CAP_64B) {
			foo();
		}
	#endif
	}

To:

	static bool macb_dma_is_64b(struct macb *bp)
	{
		return IS_ENABLED(CONFIG_ARCH_DMA_ADDR_T_64BIT) &&
		       bp->hw_dma_cap & HW_DMA_CAP_64B;
	}

	void bar(struct macb *bp) {
		if (macb_dma_is_64b(bp)) {
			foo();
		}
	}

In the first case, we use explicit preprocessor directives to remove
code if CONFIG_ARCH_DMA_ADDR_T_64BIT isn't defined.

In the second case, our compiler optimises away the IS_ENABLED() call.
 - If false, then the branch doesn't appear.
 - If true, then only the capability check is inlined in bar().
I checked the assembly on arm/arm64/MIPS.

Conclusion: the hotpath doesn't change.

Thanks,

--
Théo Lebrun, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com