* [PATCH] NET_DMA: free skbs periodically
@ 2010-03-16 15:22 Steven J. Magnani
2010-03-19 3:36 ` David Miller
2010-03-20 21:29 ` David Miller
0 siblings, 2 replies; 3+ messages in thread
From: Steven J. Magnani @ 2010-03-16 15:22 UTC (permalink / raw
To: netdev
Cc: David S. Miller, Alexey Kuznetsov, Pekka Savola (ipv6),
James Morris, Hideaki YOSHIFUJI, Patrick McHardy, linux-kernel,
Steven J. Magnani
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>
---
diff -uprN a/net/ipv4/tcp.c b/net/ipv4/tcp.c
--- a/net/ipv4/tcp.c 2010-03-16 09:15:31.000000000 -0500
+++ b/net/ipv4/tcp.c 2010-03-16 09:59:06.000000000 -0500
@@ -1254,6 +1254,39 @@ static void tcp_prequeue_process(struct
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, stru
/* __ 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, stru
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;
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] NET_DMA: free skbs periodically
2010-03-16 15:22 [PATCH] NET_DMA: free skbs periodically Steven J. Magnani
@ 2010-03-19 3:36 ` David Miller
2010-03-20 21:29 ` David Miller
1 sibling, 0 replies; 3+ messages in thread
From: David Miller @ 2010-03-19 3:36 UTC (permalink / raw
To: steve; +Cc: netdev, kuznet, pekkas, jmorris, yoshfuji, kaber, linux-kernel
From: "Steven J. Magnani" <steve@digidescorp.com>
Date: Tue, 16 Mar 2010 10:22:44 -0500
> 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>
Can one of the NET DMA folks review this.
It's a pretty fundamental problem with how this stuff works
it seems.
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] NET_DMA: free skbs periodically
2010-03-16 15:22 [PATCH] NET_DMA: free skbs periodically Steven J. Magnani
2010-03-19 3:36 ` David Miller
@ 2010-03-20 21:29 ` David Miller
1 sibling, 0 replies; 3+ messages in thread
From: David Miller @ 2010-03-20 21:29 UTC (permalink / raw
To: steve; +Cc: netdev, kuznet, pekkas, jmorris, yoshfuji, kaber, linux-kernel
From: "Steven J. Magnani" <steve@digidescorp.com>
Date: Tue, 16 Mar 2010 10:22:44 -0500
> 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>
Applied, thanks for fixing this bug Steven.
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2010-03-20 21:28 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2010-03-16 15:22 [PATCH] NET_DMA: free skbs periodically Steven J. Magnani
2010-03-19 3:36 ` David Miller
2010-03-20 21:29 ` David Miller
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).