[LU-8875] Change to new RDMA done callback mechanism

Blocker — Thu, 1 Oct 2020 22:20:33 +0000

The new done callback is implemented in three ways:

1- Direct (no polling)
2- softirq (our callback gets called by the IRQ_POLL_SOFTIRQ mechanism)
3- WorkQueues

It is very tempting to replace the kiblnd_scheduler() and its use of wait queues with the WorkQueue approach. However, this has two major problems:

1- There is no way to bind the WorkQueue to a specific CPT without submitting a change to the RDMA code base. I'm not interested in doing this.
2- It is unclear how the kernel threads for WorkQueues are created/destroyed. If not done efficiently, this will cause a performance degradation to LNet.

So, my recommendation is to bind our current kiblnd_cq_completion() to the softirq callback (with necessary semantic changes). The main loop for the scheduler, kiblnd_scheduler(), will need to be updated to not do any polling of the cq as that will be done for us by the new callback mechanism. All of o2iblnd needs to be scanned for any cq polling and that needs to be turned off.

Whamcloud Community JIRA

[LU-8875] Change to new RDMA done callback mechanism