[PATCH net 0/3] tcp: take care of empty skbs in write queue

[PATCH net 0/3] tcp: take care of empty skbs in write queue

[Eric Dumazet]

We understood recently that TCP sockets could have an empty
skb at the tail of the write queue, leading to various problems.

This patch series :

1) Make sure we do not send an empty packet since this
   was unintended and causing crashes in old kernels.

2) Change tcp_write_queue_empty() to not be fooled by
   the presence of an empty skb.

3) Fix a bug that could trigger suboptimal epoll()
   application behavior under memory pressure.

Eric Dumazet (3):
  tcp: do not send empty skb from tcp_write_xmit()
  tcp: refine tcp_write_queue_empty() implementation
  tcp: refine rule to allow EPOLLOUT generation under mem pressure

 include/net/tcp.h     | 11 ++++++++++-
 net/ipv4/tcp.c        |  6 ++----
 net/ipv4/tcp_output.c | 13 +++++++++++--
 3 files changed, 23 insertions(+), 7 deletions(-)

-- 
2.24.1.735.g03f4e72817-goog

[PATCH net 1/3] tcp: do not send empty skb from tcp_write_xmit()

[Eric Dumazet]

Backport of commit fdfc5c8594c2 ("tcp: remove empty skb from
write queue in error cases") in linux-4.14 stable triggered
various bugs. One of them has been fixed in commit ba2ddb43f270
("tcp: Don't dequeue SYN/FIN-segments from write-queue"), but
we still have crashes in some occasions.

Root-cause is that when tcp_sendmsg() has allocated a fresh
skb and could not append a fragment before being blocked
in sk_stream_wait_memory(), tcp_write_xmit() might be called
and decide to send this fresh and empty skb.

Sending an empty packet is not only silly, it might have caused
many issues we had in the past with tp->packets_out being
out of sync.

Fixes: c65f7f00c587 ("[TCP]: Simplify SKB data portion allocation with NETIF_F_SG.")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Christoph Paasch <cpaasch@apple.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Cc: Jason Baron <jbaron@akamai.com>
---
 net/ipv4/tcp_output.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index b184f03d743715ef4b2d166ceae651529be77953..57f434a8e41ffd6bc584cb4d9e87703491a378c1 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -2438,6 +2438,14 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 		if (tcp_small_queue_check(sk, skb, 0))
 			break;
 
+		/* Argh, we hit an empty skb(), presumably a thread
+		 * is sleeping in sendmsg()/sk_stream_wait_memory().
+		 * We do not want to send a pure-ack packet and have
+		 * a strange looking rtx queue with empty packet(s).
+		 */
+		if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
+			break;
+
 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
 			break;
 
-- 
2.24.1.735.g03f4e72817-goog

[PATCH net 2/3] tcp: refine tcp_write_queue_empty() implementation

[Eric Dumazet]

Due to how tcp_sendmsg() is implemented, we can have an empty
skb at the tail of the write queue.

Most [1] tcp_write_queue_empty() callers want to know if there is
anything to send (payload and/or FIN)

Instead of checking if the sk_write_queue is empty, we need
to test if tp->write_seq == tp->snd_nxt

[1] tcp_send_fin() was the only caller that expected to
 see if an skb was in the write queue, I have changed the code
 to reuse the tcp_write_queue_tail() result.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
---
 include/net/tcp.h     | 11 ++++++++++-
 net/ipv4/tcp_output.c |  5 +++--
 2 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/include/net/tcp.h b/include/net/tcp.h
index 86b9a8766648824c0f122f6c01f55d59bd0d7d72..e460ea7f767ba627972a63a974cae80357808366 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -1766,9 +1766,18 @@ static inline bool tcp_skb_is_last(const struct sock *sk,
 	return skb_queue_is_last(&sk->sk_write_queue, skb);
 }
 
+/**
+ * tcp_write_queue_empty - test if any payload (or FIN) is available in write queue
+ * @sk: socket
+ *
+ * Since the write queue can have a temporary empty skb in it,
+ * we must not use "return skb_queue_empty(&sk->sk_write_queue)"
+ */
 static inline bool tcp_write_queue_empty(const struct sock *sk)
 {
-	return skb_queue_empty(&sk->sk_write_queue);
+	const struct tcp_sock *tp = tcp_sk(sk);
+
+	return tp->write_seq == tp->snd_nxt;
 }
 
 static inline bool tcp_rtx_queue_empty(const struct sock *sk)
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 57f434a8e41ffd6bc584cb4d9e87703491a378c1..36902d08473ec7e45a654234b407217ee6c65fb1 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -3129,7 +3129,7 @@ void sk_forced_mem_schedule(struct sock *sk, int size)
  */
 void tcp_send_fin(struct sock *sk)
 {
-	struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
+	struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
 	struct tcp_sock *tp = tcp_sk(sk);
 
 	/* Optimization, tack on the FIN if we have one skb in write queue and
@@ -3137,6 +3137,7 @@ void tcp_send_fin(struct sock *sk)
 	 * Note: in the latter case, FIN packet will be sent after a timeout,
 	 * as TCP stack thinks it has already been transmitted.
 	 */
+	tskb = tail;
 	if (!tskb && tcp_under_memory_pressure(sk))
 		tskb = skb_rb_last(&sk->tcp_rtx_queue);
 
@@ -3144,7 +3145,7 @@ void tcp_send_fin(struct sock *sk)
 		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
 		TCP_SKB_CB(tskb)->end_seq++;
 		tp->write_seq++;
-		if (tcp_write_queue_empty(sk)) {
+		if (!tail) {
 			/* This means tskb was already sent.
 			 * Pretend we included the FIN on previous transmit.
 			 * We need to set tp->snd_nxt to the value it would have
-- 
2.24.1.735.g03f4e72817-goog

[PATCH net 3/3] tcp: refine rule to allow EPOLLOUT generation under mem pressure

[Eric Dumazet]

At the time commit ce5ec440994b ("tcp: ensure epoll edge trigger
wakeup when write queue is empty") was added to the kernel,
we still had a single write queue, combining rtx and write queues.

Once we moved the rtx queue into a separate rb-tree, testing
if sk_write_queue is empty has been suboptimal.

Indeed, if we have packets in the rtx queue, we probably want
to delay the EPOLLOUT generation at the time incoming packets
will free them, making room, but more importantly avoiding
flooding application with EPOLLOUT events.

Solution is to use tcp_rtx_and_write_queues_empty() helper.

Fixes: 75c119afe14f ("tcp: implement rb-tree based retransmit queue")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Neal Cardwell <ncardwell@google.com>
---
 net/ipv4/tcp.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 8a39ee79489192c02385aaadc8d1ae969fb55d23..716938313a32534f312c6b90f66ff7870177b4a5 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -1087,8 +1087,7 @@ ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 		goto out;
 out_err:
 	/* make sure we wake any epoll edge trigger waiter */
-	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
-		     err == -EAGAIN)) {
+	if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
 		sk->sk_write_space(sk);
 		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
 	}
@@ -1419,8 +1418,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 	sock_zerocopy_put_abort(uarg, true);
 	err = sk_stream_error(sk, flags, err);
 	/* make sure we wake any epoll edge trigger waiter */
-	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
-		     err == -EAGAIN)) {
+	if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
 		sk->sk_write_space(sk);
 		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
 	}
-- 
2.24.1.735.g03f4e72817-goog

Re: [PATCH net 1/3] tcp: do not send empty skb from tcp_write_xmit()

[Soheil Hassas Yeganeh]

On Thu, Dec 12, 2019 at 3:55 PM Eric Dumazet <edumazet@google.com> wrote:
>
> Backport of commit fdfc5c8594c2 ("tcp: remove empty skb from
> write queue in error cases") in linux-4.14 stable triggered
> various bugs. One of them has been fixed in commit ba2ddb43f270
> ("tcp: Don't dequeue SYN/FIN-segments from write-queue"), but
> we still have crashes in some occasions.
>
> Root-cause is that when tcp_sendmsg() has allocated a fresh
> skb and could not append a fragment before being blocked
> in sk_stream_wait_memory(), tcp_write_xmit() might be called
> and decide to send this fresh and empty skb.
>
> Sending an empty packet is not only silly, it might have caused
> many issues we had in the past with tp->packets_out being
> out of sync.
>
> Fixes: c65f7f00c587 ("[TCP]: Simplify SKB data portion allocation with NETIF_F_SG.")
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Christoph Paasch <cpaasch@apple.com>
> Acked-by: Neal Cardwell <ncardwell@google.com>
> Cc: Jason Baron <jbaron@akamai.com>

Acked-by: Soheil Hassas Yeganeh <soheil@google.com>

> ---
>  net/ipv4/tcp_output.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
>
> diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
> index b184f03d743715ef4b2d166ceae651529be77953..57f434a8e41ffd6bc584cb4d9e87703491a378c1 100644
> --- a/net/ipv4/tcp_output.c
> +++ b/net/ipv4/tcp_output.c
> @@ -2438,6 +2438,14 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
>                 if (tcp_small_queue_check(sk, skb, 0))
>                         break;
>
> +               /* Argh, we hit an empty skb(), presumably a thread
> +                * is sleeping in sendmsg()/sk_stream_wait_memory().
> +                * We do not want to send a pure-ack packet and have
> +                * a strange looking rtx queue with empty packet(s).
> +                */
> +               if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
> +                       break;
> +
>                 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
>                         break;
>
> --
> 2.24.1.735.g03f4e72817-goog
>

Re: [PATCH net 2/3] tcp: refine tcp_write_queue_empty() implementation

[Soheil Hassas Yeganeh]

On Thu, Dec 12, 2019 at 3:55 PM Eric Dumazet <edumazet@google.com> wrote:
>
> Due to how tcp_sendmsg() is implemented, we can have an empty
> skb at the tail of the write queue.
>
> Most [1] tcp_write_queue_empty() callers want to know if there is
> anything to send (payload and/or FIN)
>
> Instead of checking if the sk_write_queue is empty, we need
> to test if tp->write_seq == tp->snd_nxt
>
> [1] tcp_send_fin() was the only caller that expected to
>  see if an skb was in the write queue, I have changed the code
>  to reuse the tcp_write_queue_tail() result.
>
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Neal Cardwell <ncardwell@google.com>

Acked-by: Soheil Hassas Yeganeh <soheil@google.com>

Nice catch!
> ---
>  include/net/tcp.h     | 11 ++++++++++-
>  net/ipv4/tcp_output.c |  5 +++--
>  2 files changed, 13 insertions(+), 3 deletions(-)
>
> diff --git a/include/net/tcp.h b/include/net/tcp.h
> index 86b9a8766648824c0f122f6c01f55d59bd0d7d72..e460ea7f767ba627972a63a974cae80357808366 100644
> --- a/include/net/tcp.h
> +++ b/include/net/tcp.h
> @@ -1766,9 +1766,18 @@ static inline bool tcp_skb_is_last(const struct sock *sk,
>         return skb_queue_is_last(&sk->sk_write_queue, skb);
>  }
>
> +/**
> + * tcp_write_queue_empty - test if any payload (or FIN) is available in write queue
> + * @sk: socket
> + *
> + * Since the write queue can have a temporary empty skb in it,
> + * we must not use "return skb_queue_empty(&sk->sk_write_queue)"
> + */
>  static inline bool tcp_write_queue_empty(const struct sock *sk)
>  {
> -       return skb_queue_empty(&sk->sk_write_queue);
> +       const struct tcp_sock *tp = tcp_sk(sk);
> +
> +       return tp->write_seq == tp->snd_nxt;
>  }
>
>  static inline bool tcp_rtx_queue_empty(const struct sock *sk)
> diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
> index 57f434a8e41ffd6bc584cb4d9e87703491a378c1..36902d08473ec7e45a654234b407217ee6c65fb1 100644
> --- a/net/ipv4/tcp_output.c
> +++ b/net/ipv4/tcp_output.c
> @@ -3129,7 +3129,7 @@ void sk_forced_mem_schedule(struct sock *sk, int size)
>   */
>  void tcp_send_fin(struct sock *sk)
>  {
> -       struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
> +       struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
>         struct tcp_sock *tp = tcp_sk(sk);
>
>         /* Optimization, tack on the FIN if we have one skb in write queue and
> @@ -3137,6 +3137,7 @@ void tcp_send_fin(struct sock *sk)
>          * Note: in the latter case, FIN packet will be sent after a timeout,
>          * as TCP stack thinks it has already been transmitted.
>          */
> +       tskb = tail;
>         if (!tskb && tcp_under_memory_pressure(sk))
>                 tskb = skb_rb_last(&sk->tcp_rtx_queue);
>
> @@ -3144,7 +3145,7 @@ void tcp_send_fin(struct sock *sk)
>                 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
>                 TCP_SKB_CB(tskb)->end_seq++;
>                 tp->write_seq++;
> -               if (tcp_write_queue_empty(sk)) {
> +               if (!tail) {
>                         /* This means tskb was already sent.
>                          * Pretend we included the FIN on previous transmit.
>                          * We need to set tp->snd_nxt to the value it would have
> --
> 2.24.1.735.g03f4e72817-goog
>

Re: [PATCH net 3/3] tcp: refine rule to allow EPOLLOUT generation under mem pressure

[Soheil Hassas Yeganeh]

On Thu, Dec 12, 2019 at 3:55 PM Eric Dumazet <edumazet@google.com> wrote:
>
> At the time commit ce5ec440994b ("tcp: ensure epoll edge trigger
> wakeup when write queue is empty") was added to the kernel,
> we still had a single write queue, combining rtx and write queues.
>
> Once we moved the rtx queue into a separate rb-tree, testing
> if sk_write_queue is empty has been suboptimal.
>
> Indeed, if we have packets in the rtx queue, we probably want
> to delay the EPOLLOUT generation at the time incoming packets
> will free them, making room, but more importantly avoiding
> flooding application with EPOLLOUT events.
>
> Solution is to use tcp_rtx_and_write_queues_empty() helper.
>
> Fixes: 75c119afe14f ("tcp: implement rb-tree based retransmit queue")
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Cc: Jason Baron <jbaron@akamai.com>
> Cc: Neal Cardwell <ncardwell@google.com>

Acked-by: Soheil Hassas Yeganeh <soheil@google.com>

> ---
>  net/ipv4/tcp.c | 6 ++----
>  1 file changed, 2 insertions(+), 4 deletions(-)
>
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index 8a39ee79489192c02385aaadc8d1ae969fb55d23..716938313a32534f312c6b90f66ff7870177b4a5 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -1087,8 +1087,7 @@ ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
>                 goto out;
>  out_err:
>         /* make sure we wake any epoll edge trigger waiter */
> -       if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
> -                    err == -EAGAIN)) {
> +       if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
>                 sk->sk_write_space(sk);
>                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
>         }
> @@ -1419,8 +1418,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
>         sock_zerocopy_put_abort(uarg, true);
>         err = sk_stream_error(sk, flags, err);
>         /* make sure we wake any epoll edge trigger waiter */
> -       if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
> -                    err == -EAGAIN)) {
> +       if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
>                 sk->sk_write_space(sk);
>                 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
>         }
> --
> 2.24.1.735.g03f4e72817-goog
>

Re: [PATCH net 0/3] tcp: take care of empty skbs in write queue

[Jakub Kicinski]

On Thu, 12 Dec 2019 12:55:28 -0800, Eric Dumazet wrote:
> We understood recently that TCP sockets could have an empty
> skb at the tail of the write queue, leading to various problems.
> 
> This patch series :
> 
> 1) Make sure we do not send an empty packet since this
>    was unintended and causing crashes in old kernels.
> 
> 2) Change tcp_write_queue_empty() to not be fooled by
>    the presence of an empty skb.
> 
> 3) Fix a bug that could trigger suboptimal epoll()
>    application behavior under memory pressure.

Applied, thank you!

I've only queued patch 1 for stable/all, please let me know if I should
also queue 2 and/or 3.