[PATCH] RPC: expose API for serializing access to RPC transports
authorChuck Lever <cel@netapp.com>
Thu, 25 Aug 2005 23:25:51 +0000 (16:25 -0700)
committerTrond Myklebust <Trond.Myklebust@netapp.com>
Fri, 23 Sep 2005 16:38:40 +0000 (12:38 -0400)
 The next method we abstract is the one that releases a transport,
 allowing another task to have access to the transport.

 Again, one generic version of this is provided for transports that
 don't need the RPC client to perform congestion control, and one
 version is for transports that can use the original Van Jacobson
 implementation in xprt.c.

 Test-plan:
 Use WAN simulation to cause sporadic bursty packet loss.  Look for
 significant regression in performance or client stability.

Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
include/linux/sunrpc/xprt.h
net/sunrpc/xprt.c
net/sunrpc/xprtsock.c

index eee1c68778518c3861d68ab50f6b4608287026de..86833b725bb561a84e0681a74a62be986eefe197 100644 (file)
@@ -133,6 +133,7 @@ struct rpc_xprt;
 struct rpc_xprt_ops {
        void            (*set_buffer_size)(struct rpc_xprt *xprt);
        int             (*reserve_xprt)(struct rpc_task *task);
+       void            (*release_xprt)(struct rpc_xprt *xprt, struct rpc_task *task);
        void            (*connect)(struct rpc_task *task);
        int             (*send_request)(struct rpc_task *task);
        void            (*set_retrans_timeout)(struct rpc_task *task);
@@ -238,6 +239,8 @@ int                 xprt_reserve_xprt_cong(struct rpc_task *task);
 int                    xprt_prepare_transmit(struct rpc_task *task);
 void                   xprt_transmit(struct rpc_task *task);
 int                    xprt_adjust_timeout(struct rpc_rqst *req);
+void                   xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task);
+void                   xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task);
 void                   xprt_release(struct rpc_task *task);
 int                    xprt_destroy(struct rpc_xprt *xprt);
 
index 2d1e8b83dd685203bd51d7944db51df45fcb25b6..e92ea99dd3181f77add4b1c9b8fcbb82a0f119e7 100644 (file)
@@ -153,14 +153,42 @@ static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
        return retval;
 }
 
-
 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
+{
+       struct rpc_task *task;
+       struct rpc_rqst *req;
+
+       if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
+               return;
+
+       task = rpc_wake_up_next(&xprt->resend);
+       if (!task) {
+               task = rpc_wake_up_next(&xprt->sending);
+               if (!task)
+                       goto out_unlock;
+       }
+
+       req = task->tk_rqstp;
+       xprt->snd_task = task;
+       if (req) {
+               req->rq_bytes_sent = 0;
+               req->rq_ntrans++;
+       }
+       return;
+
+out_unlock:
+       smp_mb__before_clear_bit();
+       clear_bit(XPRT_LOCKED, &xprt->state);
+       smp_mb__after_clear_bit();
+}
+
+static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 {
        struct rpc_task *task;
 
        if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
                return;
-       if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
+       if (RPCXPRT_CONGESTED(xprt))
                goto out_unlock;
        task = rpc_wake_up_next(&xprt->resend);
        if (!task) {
@@ -168,7 +196,7 @@ static void __xprt_lock_write_next(struct rpc_xprt *xprt)
                if (!task)
                        goto out_unlock;
        }
-       if (xprt->nocong || __xprt_get_cong(xprt, task)) {
+       if (__xprt_get_cong(xprt, task)) {
                struct rpc_rqst *req = task->tk_rqstp;
                xprt->snd_task = task;
                if (req) {
@@ -183,11 +211,14 @@ out_unlock:
        smp_mb__after_clear_bit();
 }
 
-/*
- * Releases the transport for use by other requests.
+/**
+ * xprt_release_xprt - allow other requests to use a transport
+ * @xprt: transport with other tasks potentially waiting
+ * @task: task that is releasing access to the transport
+ *
+ * Note that "task" can be NULL.  No congestion control is provided.
  */
-static void
-__xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
+void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 {
        if (xprt->snd_task == task) {
                xprt->snd_task = NULL;
@@ -198,11 +229,29 @@ __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
        }
 }
 
-static inline void
-xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
+/**
+ * xprt_release_xprt_cong - allow other requests to use a transport
+ * @xprt: transport with other tasks potentially waiting
+ * @task: task that is releasing access to the transport
+ *
+ * Note that "task" can be NULL.  Another task is awoken to use the
+ * transport if the transport's congestion window allows it.
+ */
+void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
+{
+       if (xprt->snd_task == task) {
+               xprt->snd_task = NULL;
+               smp_mb__before_clear_bit();
+               clear_bit(XPRT_LOCKED, &xprt->state);
+               smp_mb__after_clear_bit();
+               __xprt_lock_write_next_cong(xprt);
+       }
+}
+
+static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 {
        spin_lock_bh(&xprt->transport_lock);
-       __xprt_release_write(xprt, task);
+       xprt->ops->release_xprt(xprt, task);
        spin_unlock_bh(&xprt->transport_lock);
 }
 
@@ -237,7 +286,7 @@ __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
                return;
        req->rq_cong = 0;
        xprt->cong -= RPC_CWNDSCALE;
-       __xprt_lock_write_next(xprt);
+       __xprt_lock_write_next_cong(xprt);
 }
 
 /*
@@ -256,7 +305,7 @@ xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
                cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
                if (cwnd > RPC_MAXCWND(xprt))
                        cwnd = RPC_MAXCWND(xprt);
-               __xprt_lock_write_next(xprt);
+               __xprt_lock_write_next_cong(xprt);
        } else if (result == -ETIMEDOUT) {
                cwnd >>= 1;
                if (cwnd < RPC_CWNDSCALE)
@@ -693,7 +742,7 @@ void xprt_transmit(struct rpc_task *task)
                        task->tk_status = -ENOTCONN;
                else if (!req->rq_received)
                        rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
-               __xprt_release_write(xprt, task);
+               xprt->ops->release_xprt(xprt, task);
                spin_unlock_bh(&xprt->transport_lock);
                return;
        }
@@ -792,7 +841,7 @@ void xprt_release(struct rpc_task *task)
        if (!(req = task->tk_rqstp))
                return;
        spin_lock_bh(&xprt->transport_lock);
-       __xprt_release_write(xprt, task);
+       xprt->ops->release_xprt(xprt, task);
        __xprt_put_cong(xprt, req);
        if (!list_empty(&req->rq_list))
                list_del(&req->rq_list);
index fc4fbe8ea346d2554cbf5bd87b079ddd84d3deed..8589c1ad55e3942a053442d65ee8d952184c926a 100644 (file)
@@ -1046,6 +1046,7 @@ static void xs_connect(struct rpc_task *task)
 static struct rpc_xprt_ops xs_udp_ops = {
        .set_buffer_size        = xs_udp_set_buffer_size,
        .reserve_xprt           = xprt_reserve_xprt_cong,
+       .release_xprt           = xprt_release_xprt_cong,
        .connect                = xs_connect,
        .send_request           = xs_udp_send_request,
        .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
@@ -1056,6 +1057,7 @@ static struct rpc_xprt_ops xs_udp_ops = {
 static struct rpc_xprt_ops xs_tcp_ops = {
        .set_buffer_size        = xs_tcp_set_buffer_size,
        .reserve_xprt           = xprt_reserve_xprt,
+       .release_xprt           = xprt_release_xprt,
        .connect                = xs_connect,
        .send_request           = xs_tcp_send_request,
        .set_retrans_timeout    = xprt_set_retrans_timeout_def,