On Wed, Aug 11, 2021 at 12:13:23PM +0100, Christoph Hellwig wrote: > On Wed, Aug 11, 2021 at 09:26:33AM +0100, Christoph Hellwig wrote: > > I really don't like all the layering violations in here. What is the > > problem with a simple (optional) percpu cache in the bio_set? Something > > like the completely untested patch below: > > A slightly updated version that actually compiles and survives minimal > testing is here: > > http://git.infradead.org/users/hch/block.git/shortlog/refs/heads/bio-cache Just wondering if the percpu bio cache may perform better than slab which should be very similar with percpu allocation. And here the cached bio is just one fixed length bio with inline bvecs. BTW, in the patch of 'io_uring: ask for bio caching', you have to not set IOCB_ALLOC_CACHE just like what Jens did, otherwise it will break when using io_uring over queue without POLL capability. Thanks, Ming
