[PATCHv2] nvme-tcp: align I/O cpu with blk-mq mapping

[Sagi Grimberg]

>> I see how you address multiple controllers falling into the same 
>> mappings case in your patch.
>> You could have selected a different mq_map entry for each controller 
>> (out of the entries that map to the qid).
>>
> Looked at it, but hadn't any idea how to figure out the load.
> The load is actually per-cpu, but we only have per controller structures.
> So we would need to introduce a per-cpu counter, detailing out the
> number of queues scheduled on that CPU.
> But that won't help with the CPU oversubscription issue; we still 
> might have substantially higher number of overall queues than we have 
> CPUs...

I think that it would still be better than what you have right now:

IIUC Right now you will have for all controllers (based on your example):
queue 1: using cpu 6
queue 2: using cpu 9
queue 3: using cpu 18

But selecting a different mq_map entry can give:
ctrl1:
queue 1: using cpu 6
queue 2: using cpu 9
queue 3: using cpu 18

ctrl2:
queue 1: using cpu 7
queue 2: using cpu 10
queue 3: using cpu 19

ctrl3:
queue 1: using cpu 8
queue 2: using cpu 11
queue 3: using cpu 20

ctrl4:
queue 1: using cpu 54
queue 2: using cpu 57
queue 3: using cpu 66

and so on...





>
>>>
>>> Not sure how wq_unbound helps in this case; in theory the workqueue
>>> items can be pushed on arbitrary CPUs, but that only leads to even 
>>> worse
>>> thread bouncing.
>>>
>>> However, topic for ALPSS. We really should have some sore of 
>>> backpressure here.
>>
>> I have a patch that was sitting for some time now, to make the RX 
>> path run directly
>> from softirq, which should make RX execute from the cpu core mapped 
>> to the RSS hash.
>> Perhaps you or your customer can give it a go.
>>
> No s**t. That is pretty much what I wanted to do.
> I'm sure to give it a go.
> Thanks for that!

You will need another prep patch for it:
--
diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
index 3649987c0a2d..b6ea7e337eb8 100644
--- a/drivers/nvme/host/tcp.c
+++ b/drivers/nvme/host/tcp.c
@@ -955,6 +955,18 @@ static int nvme_tcp_recv_skb(read_descriptor_t 
*desc, struct sk_buff *skb,
         return consumed;
  }

+static int nvme_tcp_try_recv_locked(struct nvme_tcp_queue *queue)
+{
+       struct socket *sock = queue->sock;
+       struct sock *sk = sock->sk;
+       read_descriptor_t rd_desc;
+
+       rd_desc.arg.data = queue;
+       rd_desc.count = 1;
+       queue->nr_cqe = 0;
+       return sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
+}
+
  static void nvme_tcp_data_ready(struct sock *sk)
  {
         struct nvme_tcp_queue *queue;
@@ -1251,16 +1263,11 @@ static int nvme_tcp_try_send(struct 
nvme_tcp_queue *queue)

  static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
  {
-       struct socket *sock = queue->sock;
-       struct sock *sk = sock->sk;
-       read_descriptor_t rd_desc;
+       struct sock *sk = queue->sock->sk;
         int consumed;

-       rd_desc.arg.data = queue;
-       rd_desc.count = 1;
         lock_sock(sk);
-       queue->nr_cqe = 0;
-       consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
+       consumed = nvme_tcp_try_recv_locked(queue);
         release_sock(sk);
         return consumed;
  }
--