NET_DMA: free skbs periodically
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>
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index ae16f80..6afb6d8 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -1254,6 +1254,39 @@
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 @@
/* __ 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 @@
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 @@
copied = -EFAULT;
break;
}
+
+ dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
+
if ((offset + used) == skb->len)
copied_early = 1;
@@ -1702,27 +1743,9 @@
}
#ifdef CONFIG_NET_DMA
- if (tp->ucopy.dma_chan) {
- dma_cookie_t done, used;
+ tcp_service_net_dma(sk, true); /* Wait for queue to drain */
+ tp->ucopy.dma_chan = NULL;
- 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);
- }
- }
-
- /* 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;