Re: [PATCHv3 5/5] io_uring: cache nodes and mapped buffers

[Keith Busch <kbusch@kernel.org>]

On Tue, Feb 18, 2025 at 12:42:19PM -0800, Caleb Sander Mateos wrote:
> Right, that's a good point that there's a tradeoff. I think always
> allocating space for IO_CACHED_BVECS_SEGS bvecs is reasonable. Maybe
> IO_CACHED_BVECS_SEGS should be slightly smaller so the allocation fits
> nicely in a power of 2? Currently it looks to take up 560 bytes:
> >>> 48 + 16 * 32
> 560
> 
> Using IO_CACHED_BVECS_SEGS = 29 instead would make it 512 bytes:
> >>> 48 + 16 * 29
> 512

Right, and it's even smaller on 32-bit architectures.

I don't think it's worth optimizing the cached object size like this.
It's probably better we optimize for a particular transfer size. If your
bvec is physically discontiguous, 32 bvecs gets you to 128k transfers on
a 4k page platform. That seems like a common transfer size for
benchmarking throughput. It is arbitrary at the end of the day, so I'm
not set on that if there's an argument for something different.