Hi, This abstracts out the alloc cache we have for apoll entries, and extends it to be usable for recv/sendmsg as well. During that abstraction, impose an upper limit for cached entries as well. This yields a 4-5% performance increase running netbench using sendmsg/recvmsg rather than plan send and recv. Post 5.20, I suspect we can get rid of using io_async_msghdr for single vector sendmsg/recvmsg, which will make this less relevant. But as this work isn't done yet, and the support for eg ITER_UBUF isn't upstream yet either, this can help fill the gap. V2: - Abstract cache node out as well, so we can have common helpers for everything. -- Jens Axboe
