On Tue, Oct 18, 2022 at 6:39 AM Bob Pearson <rpearsonhpe@xxxxxxxxx> wrote: > > This patch series implements work queues as an alternative for > the main tasklets in the rdma_rxe driver. The first few patches > perform some cleanups of the current tasklet code followed by a > patch that makes the internal API for task execution pluggable and > implements an inline and a tasklet based set of functions. > The remaining patches cleanup the qp reset and error code in the > three tasklets and modify the locking logic to prevent making > multiple calls to the tasklet scheduling routine. Finally after > this preparation the work queue equivalent set of functions is > added and module parameters are implemented to allow tuning the > task types. > > The advantages of the work queue version of deferred task execution > is mainly that the work queue variant has much better scalability > and overall performance than the tasklet variant. The tasklet > performance saturates with one connected queue pair and stays constant. > The work queue performance is slightly better for one queue pair but > scales up with the number of connected queue pairs. The perftest > microbenchmarks in local loopback mode (not a very realistic test > case) can reach approximately 100Gb/sec with work queues compared to > about 16Gb/sec for tasklets. > > This patch series is derived from an earlier patch set developed by > Ian Ziemba at HPE which is used in some Lustre storage clients attached > to Lustre servers with hard RoCE v2 NICs. > > Bob Pearson (16): > RDMA/rxe: Remove init of task locks from rxe_qp.c > RDMA/rxe: Removed unused name from rxe_task struct > RDMA/rxe: Split rxe_run_task() into two subroutines > RDMA/rxe: Make rxe_do_task static > RDMA/rxe: Rename task->state_lock to task->lock > RDMA/rxe: Make task interface pluggable > RDMA/rxe: Simplify reset state handling in rxe_resp.c > RDMA/rxe: Split rxe_drain_resp_pkts() > RDMA/rxe: Handle qp error in rxe_resp.c > RDMA/rxe: Cleanup comp tasks in rxe_qp.c > RDMA/rxe: Remove __rxe_do_task() > RDMA/rxe: Make tasks schedule each other > RDMA/rxe: Implement disable/enable_task() > RDMA/rxe: Replace TASK_STATE_START by TASK_STATE_IDLE > RDMA/rxe: Add workqueue support for tasks > RDMA/rxe: Add parameters to control task type > > drivers/infiniband/sw/rxe/rxe.c | 9 +- > drivers/infiniband/sw/rxe/rxe_comp.c | 35 ++- > drivers/infiniband/sw/rxe/rxe_net.c | 4 +- > drivers/infiniband/sw/rxe/rxe_qp.c | 87 +++---- > drivers/infiniband/sw/rxe/rxe_req.c | 10 +- > drivers/infiniband/sw/rxe/rxe_resp.c | 75 ++++-- > drivers/infiniband/sw/rxe/rxe_task.c | 354 ++++++++++++++++++++------ > drivers/infiniband/sw/rxe/rxe_task.h | 76 +++--- > drivers/infiniband/sw/rxe/rxe_verbs.c | 8 +- > 9 files changed, 451 insertions(+), 207 deletions(-) > > > base-commit: 9abf2313adc1ca1b6180c508c25f22f9395cc780 The patch series is not applying cleanly over the mentioned commit for me. Patch 'PATCH for-next 05/16] RDMA/rxe: Rename task->state_lock to task->lock.' fails at "drivers/infiniband/sw/rxe/rxe_task.c:103". I corrected that manually, then it fails in the next commit. Didn't check after that. Is it the same for others or is it just me? > -- > 2.34.1 >
