+static bool nvme_rdma_poller_queue(struct nvme_rdma_queue *queue)
Can we please make this poll_queue? or at least polled_queue?
poller sounds odd..
Changed to nvme_rdma_poll_queue..
- set->nr_maps = 2 /* default + read */;
+ set->nr_maps = HCTX_MAX_TYPES;
}
ret = blk_mq_alloc_tag_set(set);
@@ -864,6 +881,10 @@ static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
ret = PTR_ERR(ctrl->ctrl.connect_q);
goto out_free_tag_set;
}
+
+ if (ctrl->ctrl.opts->nr_poll_queues)
+ blk_queue_flag_set(QUEUE_FLAG_POLL,
+ ctrl->ctrl.connect_q);
The block core is supposed to detect we can poll based on nr_maps > 2,
and then set QUEUE_FLAG_POLL automatically. Although I got the details
wrong for PCI as well, but I just sent a fix..
I'll lose that, but I didn't understand what you got wrong for pci?
(didn't understand the fix either)
+static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
+{
+ struct nvme_rdma_queue *queue = hctx->driver_data;
+ struct ib_cq *cq = queue->ib_cq;
+
+ return ib_process_cq_direct(cq, -1);
I think we can skip the cq local variable here.
Lost..
Otherwise this looks really nice and simple, thanks for looking into it!
Do you have any performance number, especially with Jens' ringbuffer
code?
Well, I can get from my local vm on my laptop on top of soft-roce 7K
iops :) (compared tp 5.5K without polling, but not something to
conclude from)
I don't have any numbers to show for right now...
As I said in the cover letter, we want a way to tell ib_poll_cq_direct
to not count send completions (where we end our sqe), right now we are
probably screwing up poll_success stat...
