Hi, This series was mostly driven by a report on how the unfair handling of task_work can negatively impact performance as a whole. - Improve how we recycle/put ring provided kbufs. Not related to the issue at hand, but it does improve that side. - Improve how we deal with non-local task_work. Don't have unbounded looping, and handle them in issue order, like we do for local task_work. - Decouple SQPOLL task_work from actual task_work. -- Jens Axboe