Now that the normal passthrough end_io path doesn't need the request anymore, we can kill the explicit blk_mq_free_request() and just pass back RQ_END_IO_FREE instead. This enables the batched completion from freeing batches of requests at the time. This brings passthrough IO performance at least on par with bdev based O_DIRECT with io_uring. With this and batche allocations, peak performance goes from 110M IOPS to 122M IOPS. For IRQ based, passthrough is now also about 10% faster than previously, going from ~61M to ~67M IOPS. Co-developed-by: Stefan Roesch <shr@xxxxxx> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx> --- drivers/nvme/host/ioctl.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/drivers/nvme/host/ioctl.c b/drivers/nvme/host/ioctl.c index 1ccc9dd6d434..25f2f6df1602 100644 --- a/drivers/nvme/host/ioctl.c +++ b/drivers/nvme/host/ioctl.c @@ -430,8 +430,7 @@ static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req, else io_uring_cmd_complete_in_task(ioucmd, nvme_uring_task_cb); - blk_mq_free_request(req); - return RQ_END_IO_NONE; + return RQ_END_IO_FREE; } static enum rq_end_io_ret nvme_uring_cmd_end_io_meta(struct request *req, -- 2.35.1
