Hi, This builds on the stuff that Pavel and I have been tossing around: Patch 1: Enable req cache for the last class of requests, those that end up completing from IRQ context. That's regular file/bdev reads and writes. Patch 2: Enable SLAB_ACCOUNT/memcg accounting for requests Patch 3: Use memcg for the ring array accounting as well. That moves it outside of rlimit memlock, though we retain memlock accounting for registered buffers. -- Jens Axboe