A few IO micro-optimizations for IO polling and NVMe. I'm really working to close the performance gap with userspace drivers, and this gets me halfway there on latency. The fastest hardware I could get measured roundtrip read latency at 5usec with this series that was previously measuring 5.7usec. Note with NVMe, you really need to crank up the interrupt coalescing to see the completion polling benefit. Test pre-setup: echo performance | tee /sys/devices/system/cpu/cpufreq/policy*/scaling_governor echo 0 > /sys/block/nvme0n1/queue/iostats echo -1 > /sys/block/nvme0n1/queue/io_poll_delay nvme set-feature /dev/nvme0 -f 8 -v 0x4ff fio profile: [global] ioengine=pvsync2 rw=randread norandommap direct=1 bs=4k hipri [hi-pri] filename=/dev/nvme0n1 cpus_allowed=2 Keith Busch (3): nvme/pci: Start request after doorbell ring nvme/pci: Remove cq_vector check in IO path block: Polling completion performance optimization drivers/nvme/host/pci.c | 14 +++----------- fs/block_dev.c | 5 ++++- 2 files changed, 7 insertions(+), 12 deletions(-) -- 2.13.6
