From 0cbe6a8f089e5912a577537c97833546d558c357 Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Mon, 14 Sep 2020 03:20:27 -0700 Subject: [PATCH] tcp: remove SOCK_QUEUE_SHRUNK SOCK_QUEUE_SHRUNK is currently used by TCP as a temporary state that remembers if some room has been made in the rtx queue by an incoming ACK packet. This is later used from tcp_check_space() before considering to send EPOLLOUT. Problem is: If we receive SACK packets, and no packet is removed from RTX queue, we can send fresh packets, thus moving them from write queue to rtx queue and eventually empty the write queue. This stall can happen if TCP_NOTSENT_LOWAT is used. With this fix, we no longer risk stalling sends while holes are repaired, and we can fully use socket sndbuf. This also removes a cache line dirtying for typical RPC workloads. Fixes: c9bee3b7fdec ("tcp: TCP_NOTSENT_LOWAT socket option") Signed-off-by: Eric Dumazet Cc: Soheil Hassas Yeganeh Acked-by: Neal Cardwell Acked-by: Soheil Hassas Yeganeh Signed-off-by: David S. Miller --- include/net/sock.h | 2 -- net/ipv4/tcp_input.c | 23 +++++++---------------- net/ipv4/tcp_output.c | 1 - 3 files changed, 7 insertions(+), 19 deletions(-) diff --git a/include/net/sock.h b/include/net/sock.h index 7dd3051551fb..eaa5cac5e836 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -845,7 +845,6 @@ enum sock_flags { SOCK_RCVTSTAMP, /* %SO_TIMESTAMP setting */ SOCK_RCVTSTAMPNS, /* %SO_TIMESTAMPNS setting */ SOCK_LOCALROUTE, /* route locally only, %SO_DONTROUTE setting */ - SOCK_QUEUE_SHRUNK, /* write queue has been shrunk recently */ SOCK_MEMALLOC, /* VM depends on this socket for swapping */ SOCK_TIMESTAMPING_RX_SOFTWARE, /* %SOF_TIMESTAMPING_RX_SOFTWARE */ SOCK_FASYNC, /* fasync() active */ @@ -1526,7 +1525,6 @@ static inline void sk_mem_uncharge(struct sock *sk, int size) DECLARE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key); static inline void sk_wmem_free_skb(struct sock *sk, struct sk_buff *skb) { - sock_set_flag(sk, SOCK_QUEUE_SHRUNK); sk_wmem_queued_add(sk, -skb->truesize); sk_mem_uncharge(sk, skb->truesize); if (static_branch_unlikely(&tcp_tx_skb_cache_key) && diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 3658ad84f0c6..50834e7f958e 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -5332,12 +5332,6 @@ static bool tcp_should_expand_sndbuf(const struct sock *sk) return true; } -/* When incoming ACK allowed to free some skb from write_queue, - * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket - * on the exit from tcp input handler. - * - * PROBLEM: sndbuf expansion does not work well with largesend. - */ static void tcp_new_space(struct sock *sk) { struct tcp_sock *tp = tcp_sk(sk); @@ -5352,16 +5346,13 @@ static void tcp_new_space(struct sock *sk) static void tcp_check_space(struct sock *sk) { - if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { - sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); - /* pairs with tcp_poll() */ - smp_mb(); - if (sk->sk_socket && - test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { - tcp_new_space(sk); - if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) - tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); - } + /* pairs with tcp_poll() */ + smp_mb(); + if (sk->sk_socket && + test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { + tcp_new_space(sk); + if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) + tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); } } diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index ab79d36ed07f..386978dcd318 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -1682,7 +1682,6 @@ int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) skb->truesize -= delta_truesize; sk_wmem_queued_add(sk, -delta_truesize); sk_mem_uncharge(sk, delta_truesize); - sock_set_flag(sk, SOCK_QUEUE_SHRUNK); } /* Any change of skb->len requires recalculation of tso factor. */