summaryrefslogtreecommitdiffstats
path: root/net/ipv4
diff options
context:
space:
mode:
authorEric Dumazet <edumazet@google.com>2016-09-15 09:33:02 -0700
committerDavid S. Miller <davem@davemloft.net>2016-09-17 10:05:05 -0400
commit3613b3dbd1ade9a6a626dae1f608c57638eb5e8a (patch)
treed93ff35b7198f4f2ea6f9799f28d79a19ee1d814 /net/ipv4
parente812bd905a5cf00fea95da9df4889dad63d4a36a (diff)
downloadtalos-obmc-linux-3613b3dbd1ade9a6a626dae1f608c57638eb5e8a.tar.gz
talos-obmc-linux-3613b3dbd1ade9a6a626dae1f608c57638eb5e8a.zip
tcp: prepare skbs for better sack shifting
With large BDP TCP flows and lossy networks, it is very important to keep a low number of skbs in the write queue. RACK and SACK processing can perform a linear scan of it. We should avoid putting any payload in skb->head, so that SACK shifting can be done if needed. With this patch, we allow to pack ~0.5 MB per skb instead of the 64KB initially cooked at tcp_sendmsg() time. This gives a reduction of number of skbs in write queue by eight. tcp_rack_detect_loss() likes this. We still allow payload in skb->head for first skb put in the queue, to not impact RPC workloads. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Yuchung Cheng <ycheng@google.com> Acked-by: Yuchung Cheng <ycheng@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'net/ipv4')
-rw-r--r--net/ipv4/tcp.c31
1 files changed, 24 insertions, 7 deletions
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index a13fcb369f52..7dae800092e6 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -1020,17 +1020,31 @@ int tcp_sendpage(struct sock *sk, struct page *page, int offset,
}
EXPORT_SYMBOL(tcp_sendpage);
-static inline int select_size(const struct sock *sk, bool sg)
+/* Do not bother using a page frag for very small frames.
+ * But use this heuristic only for the first skb in write queue.
+ *
+ * Having no payload in skb->head allows better SACK shifting
+ * in tcp_shift_skb_data(), reducing sack/rack overhead, because
+ * write queue has less skbs.
+ * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
+ * This also speeds up tso_fragment(), since it wont fallback
+ * to tcp_fragment().
+ */
+static int linear_payload_sz(bool first_skb)
+{
+ if (first_skb)
+ return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
+ return 0;
+}
+
+static int select_size(const struct sock *sk, bool sg, bool first_skb)
{
const struct tcp_sock *tp = tcp_sk(sk);
int tmp = tp->mss_cache;
if (sg) {
if (sk_can_gso(sk)) {
- /* Small frames wont use a full page:
- * Payload will immediately follow tcp header.
- */
- tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
+ tmp = linear_payload_sz(first_skb);
} else {
int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
@@ -1161,6 +1175,8 @@ restart:
}
if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
+ bool first_skb;
+
new_segment:
/* Allocate new segment. If the interface is SG,
* allocate skb fitting to single page.
@@ -1172,10 +1188,11 @@ new_segment:
process_backlog = false;
goto restart;
}
+ first_skb = skb_queue_empty(&sk->sk_write_queue);
skb = sk_stream_alloc_skb(sk,
- select_size(sk, sg),
+ select_size(sk, sg, first_skb),
sk->sk_allocation,
- skb_queue_empty(&sk->sk_write_queue));
+ first_skb);
if (!skb)
goto wait_for_memory;
OpenPOWER on IntegriCloud