xprtrdma: Invoke rpcrdma_reply_handler directly from RECV completion
authorChuck Lever <chuck.lever@oracle.com>
Mon, 16 Oct 2017 19:01:30 +0000 (15:01 -0400)
committerAnna Schumaker <Anna.Schumaker@Netapp.com>
Fri, 17 Nov 2017 18:47:54 +0000 (13:47 -0500)
I noticed that the soft IRQ thread looked pretty busy under heavy
I/O workloads. perf suggested one area that was expensive was the
queue_work() call in rpcrdma_wc_receive. That gave me some ideas.

Instead of scheduling a separate worker to process RPC Replies,
promote the Receive completion handler to IB_POLL_WORKQUEUE, and
invoke rpcrdma_reply_handler directly.

Note that the poll workqueue is single-threaded. In order to keep
memory invalidation from serializing all RPC Replies, handle any
necessary invalidation tasks in a separate multi-threaded workqueue.

This provides a two-tier scheme, similar to OS I/O interrupt
handlers: A fast interrupt handler that schedules the slow handler
and re-enables the interrupt, and a slower handler that is invoked
for any needed heavy lifting.

Benefits include:
- One less context switch for RPCs that don't register memory
- Receive completion handling is moved out of soft IRQ context to
  make room for other users of soft IRQ
- The same CPU core now DMA syncs and XDR decodes the Receive buffer

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
net/sunrpc/xprtrdma/rpc_rdma.c
net/sunrpc/xprtrdma/verbs.c
net/sunrpc/xprtrdma/xprt_rdma.h

index 418bcc6b3e1d29b69d2854e2d7e1dbff13310372..430f8b5a8c43152c702b8c9ad79e2544eadc32d6 100644 (file)
@@ -1265,16 +1265,36 @@ out_badheader:
        goto out;
 }
 
+/* Reply handling runs in the poll worker thread. Anything that
+ * might wait is deferred to a separate workqueue.
+ */
+void rpcrdma_deferred_completion(struct work_struct *work)
+{
+       struct rpcrdma_rep *rep =
+                       container_of(work, struct rpcrdma_rep, rr_work);
+       struct rpcrdma_req *req = rpcr_to_rdmar(rep->rr_rqst);
+       struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
+
+       /* Invalidate and unmap the data payloads before waking
+        * the waiting application. This guarantees the memory
+        * regions are properly fenced from the server before the
+        * application accesses the data. It also ensures proper
+        * send flow control: waking the next RPC waits until this
+        * RPC has relinquished all its Send Queue entries.
+        */
+       rpcrdma_mark_remote_invalidation(&req->rl_registered, rep);
+       r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt, &req->rl_registered);
+
+       rpcrdma_complete_rqst(rep);
+}
+
 /* Process received RPC/RDMA messages.
  *
  * Errors must result in the RPC task either being awakened, or
  * allowed to timeout, to discover the errors at that time.
  */
-void
-rpcrdma_reply_handler(struct work_struct *work)
+void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
 {
-       struct rpcrdma_rep *rep =
-                       container_of(work, struct rpcrdma_rep, rr_work);
        struct rpcrdma_xprt *r_xprt = rep->rr_rxprt;
        struct rpc_xprt *xprt = &r_xprt->rx_xprt;
        struct rpcrdma_req *req;
@@ -1320,20 +1340,10 @@ rpcrdma_reply_handler(struct work_struct *work)
        dprintk("RPC:       %s: reply %p completes request %p (xid 0x%08x)\n",
                __func__, rep, req, be32_to_cpu(rep->rr_xid));
 
-       /* Invalidate and unmap the data payloads before waking the
-        * waiting application. This guarantees the memory regions
-        * are properly fenced from the server before the application
-        * accesses the data. It also ensures proper send flow control:
-        * waking the next RPC waits until this RPC has relinquished
-        * all its Send Queue entries.
-        */
-       if (!list_empty(&req->rl_registered)) {
-               rpcrdma_mark_remote_invalidation(&req->rl_registered, rep);
-               r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt,
-                                                   &req->rl_registered);
-       }
-
-       rpcrdma_complete_rqst(rep);
+       if (list_empty(&req->rl_registered))
+               rpcrdma_complete_rqst(rep);
+       else
+               queue_work(rpcrdma_receive_wq, &rep->rr_work);
        return;
 
 out_badstatus:
index 11a1fbf7e59e08e9172f4562652dc55d53240075..d45695408df3b4d7ab70bc6bc90f0dc738cf03e6 100644 (file)
@@ -73,7 +73,7 @@ static void rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt);
 static void rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf);
 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
 
-static struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
+struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
 
 int
 rpcrdma_alloc_wq(void)
@@ -185,7 +185,7 @@ rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
                rpcrdma_update_granted_credits(rep);
 
 out_schedule:
-       queue_work(rpcrdma_receive_wq, &rep->rr_work);
+       rpcrdma_reply_handler(rep);
        return;
 
 out_fail:
@@ -583,7 +583,7 @@ rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
 
        recvcq = ib_alloc_cq(ia->ri_device, NULL,
                             ep->rep_attr.cap.max_recv_wr + 1,
-                            0, IB_POLL_SOFTIRQ);
+                            0, IB_POLL_WORKQUEUE);
        if (IS_ERR(recvcq)) {
                rc = PTR_ERR(recvcq);
                dprintk("RPC:       %s: failed to create recv CQ: %i\n",
@@ -974,7 +974,7 @@ rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
 
        rep->rr_cqe.done = rpcrdma_wc_receive;
        rep->rr_rxprt = r_xprt;
-       INIT_WORK(&rep->rr_work, rpcrdma_reply_handler);
+       INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
        rep->rr_recv_wr.next = NULL;
        rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
        rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
index d68a1351d95e1207dd7e065590c6450db36ae826..a85bcd19b37aeee839d078f07526d5326d06ad3c 100644 (file)
@@ -533,6 +533,8 @@ void rpcrdma_ia_close(struct rpcrdma_ia *);
 bool frwr_is_supported(struct rpcrdma_ia *);
 bool fmr_is_supported(struct rpcrdma_ia *);
 
+extern struct workqueue_struct *rpcrdma_receive_wq;
+
 /*
  * Endpoint calls - xprtrdma/verbs.c
  */
@@ -617,7 +619,8 @@ void rpcrdma_unmap_sges(struct rpcrdma_ia *, struct rpcrdma_req *);
 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
-void rpcrdma_reply_handler(struct work_struct *work);
+void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
+void rpcrdma_deferred_completion(struct work_struct *work);
 
 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
 {