]> www.infradead.org Git - users/dwmw2/linux.git/commitdiff
xsk: Change the tx writeable condition
authorXuan Zhuo <xuanzhuo@linux.alibaba.com>
Tue, 1 Dec 2020 13:56:58 +0000 (21:56 +0800)
committerDaniel Borkmann <daniel@iogearbox.net>
Thu, 3 Dec 2020 00:14:38 +0000 (01:14 +0100)
Modify the tx writeable condition from the queue is not full to the
number of present tx queues is less than the half of the total number
of queues. Because the tx queue not full is a very short time, this will
cause a large number of EPOLLOUT events, and cause a large number of
process wake up.

Fixes: 35fcde7f8deb ("xsk: support for Tx")
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/508fef55188d4e1160747ead64c6dcda36735880.1606555939.git.xuanzhuo@linux.alibaba.com
net/xdp/xsk.c
net/xdp/xsk_queue.h

index 9bbfd8adbb73f844c111947a7bad7d91aef33f35..62504471fd207cd5254163e4ff892bc209d63d78 100644 (file)
@@ -211,6 +211,14 @@ static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len,
        return 0;
 }
 
+static bool xsk_tx_writeable(struct xdp_sock *xs)
+{
+       if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
+               return false;
+
+       return true;
+}
+
 static bool xsk_is_bound(struct xdp_sock *xs)
 {
        if (READ_ONCE(xs->state) == XSK_BOUND) {
@@ -296,7 +304,8 @@ void xsk_tx_release(struct xsk_buff_pool *pool)
        rcu_read_lock();
        list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
                __xskq_cons_release(xs->tx);
-               xs->sk.sk_write_space(&xs->sk);
+               if (xsk_tx_writeable(xs))
+                       xs->sk.sk_write_space(&xs->sk);
        }
        rcu_read_unlock();
 }
@@ -436,7 +445,8 @@ static int xsk_generic_xmit(struct sock *sk)
 
 out:
        if (sent_frame)
-               sk->sk_write_space(sk);
+               if (xsk_tx_writeable(xs))
+                       sk->sk_write_space(sk);
 
        mutex_unlock(&xs->mutex);
        return err;
@@ -493,7 +503,7 @@ static __poll_t xsk_poll(struct file *file, struct socket *sock,
 
        if (xs->rx && !xskq_prod_is_empty(xs->rx))
                mask |= EPOLLIN | EPOLLRDNORM;
-       if (xs->tx && !xskq_cons_is_full(xs->tx))
+       if (xs->tx && xsk_tx_writeable(xs))
                mask |= EPOLLOUT | EPOLLWRNORM;
 
        return mask;
index cdb9cf3cd1361f9d15c36d081abe0408cf93ab55..9e71b9f27679b91f25d031586f7f99b466ff7502 100644 (file)
@@ -264,6 +264,12 @@ static inline bool xskq_cons_is_full(struct xsk_queue *q)
                q->nentries;
 }
 
+static inline u32 xskq_cons_present_entries(struct xsk_queue *q)
+{
+       /* No barriers needed since data is not accessed */
+       return READ_ONCE(q->ring->producer) - READ_ONCE(q->ring->consumer);
+}
+
 /* Functions for producers */
 
 static inline bool xskq_prod_is_full(struct xsk_queue *q)