On Mon, Feb 12, 2018 at 03:45:23PM -0800, Florian Fainelli wrote:
> On many platforms, including, but not limited to Brahma-B53 platforms,
> the L1 cache line size is 64bytes. Increasing the value to 128bytes
> appears to be creating performance problems for workloads involving
> network drivers and lots of data movement. In order to keep what was
> introduced with 97303480753e ("arm64: Increase the max granular size"),
> a kernel built for ARCH_THUNDER or ARCH_THUNDER2 will get a 128bytes
> cache line size definition.This approach has been raised before ([1] as an example but you can probably find other threads) and NAK'ed. I really don't want this macro to be configurable as we aim for a single kernel Image. My proposal was to move L1_CACHE_SHIFT back to 6 and ARCH_DMA_MIN_ALIGN to 128 as this is the largest known CWG. The networking code is wrong in assuming SKB_DATA_ALIGN only needs SMP_CACHE_BYTES for DMA alignment but we can add some safety checks (i.e. WARN_ON) in the arch dma ops code if the device is non-coherent. I'll send a patch to the list (hopefully later today). Catalin [1] https://patchwork.kernel.org/patch/8634481/
