NET_DMA: free skbs periodically
authorSteven J. Magnani <steve@digidescorp.com>
Tue, 16 Mar 2010 05:22:44 +0000 (05:22 +0000)
committerDavid S. Miller <davem@davemloft.net>
Sat, 20 Mar 2010 21:29:02 +0000 (14:29 -0700)
Under NET_DMA, data transfer can grind to a halt when userland issues a
large read on a socket with a high RCVLOWAT (i.e., 512 KB for both).
This appears to be because the NET_DMA design queues up lots of memcpy
operations, but doesn't issue or wait for them (and thus free the
associated skbs) until it is time for tcp_recvmesg() to return.
The socket hangs when its TCP window goes to zero before enough data is
available to satisfy the read.

Periodically issue asynchronous memcpy operations, and free skbs for ones
that have completed, to prevent sockets from going into zero-window mode.

Signed-off-by: Steven J. Magnani <steve@digidescorp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp.c

index ae16f809e716fbf4c46526c27bd8fb24266c6a1c..6afb6d8662b2cebb984bec82657ab4221e95e464 100644 (file)
@@ -1254,6 +1254,39 @@ static void tcp_prequeue_process(struct sock *sk)
        tp->ucopy.memory = 0;
 }
 
+#ifdef CONFIG_NET_DMA
+static void tcp_service_net_dma(struct sock *sk, bool wait)
+{
+       dma_cookie_t done, used;
+       dma_cookie_t last_issued;
+       struct tcp_sock *tp = tcp_sk(sk);
+
+       if (!tp->ucopy.dma_chan)
+               return;
+
+       last_issued = tp->ucopy.dma_cookie;
+       dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
+
+       do {
+               if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
+                                             last_issued, &done,
+                                             &used) == DMA_SUCCESS) {
+                       /* Safe to free early-copied skbs now */
+                       __skb_queue_purge(&sk->sk_async_wait_queue);
+                       break;
+               } else {
+                       struct sk_buff *skb;
+                       while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
+                              (dma_async_is_complete(skb->dma_cookie, done,
+                                                     used) == DMA_SUCCESS)) {
+                               __skb_dequeue(&sk->sk_async_wait_queue);
+                               kfree_skb(skb);
+                       }
+               }
+       } while (wait);
+}
+#endif
+
 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
 {
        struct sk_buff *skb;
@@ -1546,6 +1579,10 @@ int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
                        /* __ Set realtime policy in scheduler __ */
                }
 
+#ifdef CONFIG_NET_DMA
+               if (tp->ucopy.dma_chan)
+                       dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
+#endif
                if (copied >= target) {
                        /* Do not sleep, just process backlog. */
                        release_sock(sk);
@@ -1554,6 +1591,7 @@ int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
                        sk_wait_data(sk, &timeo);
 
 #ifdef CONFIG_NET_DMA
+               tcp_service_net_dma(sk, false);  /* Don't block */
                tp->ucopy.wakeup = 0;
 #endif
 
@@ -1633,6 +1671,9 @@ do_prequeue:
                                                copied = -EFAULT;
                                        break;
                                }
+
+                               dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
+
                                if ((offset + used) == skb->len)
                                        copied_early = 1;
 
@@ -1702,27 +1743,9 @@ skip_copy:
        }
 
 #ifdef CONFIG_NET_DMA
-       if (tp->ucopy.dma_chan) {
-               dma_cookie_t done, used;
-
-               dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
-
-               while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
-                                                tp->ucopy.dma_cookie, &done,
-                                                &used) == DMA_IN_PROGRESS) {
-                       /* do partial cleanup of sk_async_wait_queue */
-                       while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
-                              (dma_async_is_complete(skb->dma_cookie, done,
-                                                     used) == DMA_SUCCESS)) {
-                               __skb_dequeue(&sk->sk_async_wait_queue);
-                               kfree_skb(skb);
-                       }
-               }
+       tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
+       tp->ucopy.dma_chan = NULL;
 
-               /* Safe to free early-copied skbs now */
-               __skb_queue_purge(&sk->sk_async_wait_queue);
-               tp->ucopy.dma_chan = NULL;
-       }
        if (tp->ucopy.pinned_list) {
                dma_unpin_iovec_pages(tp->ucopy.pinned_list);
                tp->ucopy.pinned_list = NULL;