Skip to contentSkip to conversation

tcp: return EPOLLOUT from tcp_poll only when notsent_bytes is half the limit

PATCH Applied Soheil Hassas Yeganeh 2 participants 4 messages Sep 14, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 tcp: return EPOLLOUT from tcp_poll only when notsent_bytes is half the limit

From: Soheil Hassas Yeganeh <[email protected]>

If there was any event available on the TCP socket, tcp_poll() will be called to retrieve all the events. In tcp_poll(), we call sk_stream_is_writeable() which returns true as long as we are at least one byte below notsent_lowat. This will result in quite a few spurious EPLLOUT and frequent tiny sendmsg() calls as a result.

Similar to sk_stream_write_space(), use __sk_stream_is_writeable with a wake value of 1, so that we set EPOLLOUT only if half the space is available for write.

---
 net/ipv4/tcp.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
net/ipv4/tcp.c +2 -2
index d3781b6087cb..48c351804efc 100644
@@ -564,7 +564,7 @@ __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 			mask |= EPOLLIN | EPOLLRDNORM;
 
 		if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
-			if (sk_stream_is_writeable(sk)) {
+			if (__sk_stream_is_writeable(sk, 1)) {
 				mask |= EPOLLOUT | EPOLLWRNORM;
 			} else {  /* send SIGIO later */
 				sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
@@ -576,7 +576,7 @@ __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 				 * pairs with the input side.
 				 */
 				smp_mb__after_atomic();
-				if (sk_stream_is_writeable(sk))
+				if (__sk_stream_is_writeable(sk, 1))
 					mask |= EPOLLOUT | EPOLLWRNORM;
 			}
 		} else
Signature · 1 line
2.28.0.618.gf4bc123cb7-goog
#1 lore raw

PATCH 2/2 tcp: schedule EPOLLOUT after a partial sendmsg

From: Soheil Hassas Yeganeh <[email protected]>

For EPOLLET, applications must call sendmsg until they get EAGAIN. Otherwise, there is no guarantee that EPOLLOUT is sent if there was a failure upon memory allocation.

As a result on high-speed NICs, userspace observes multiple small sendmsgs after a partial sendmsg until EAGAIN, since TCP can send 1-2 TSOs in between two sendmsg syscalls:

// One large partial send due to memory allocation failure.
sendmsg(20MB) = 2MB
// Many small sends until EAGAIN.
sendmsg(18MB) = 64KB
sendmsg(17.9MB) = 128KB
sendmsg(17.8MB) = 64KB
...
sendmsg(...) = EAGAIN
// At this point, userspace can assume an EPOLLOUT.

To fix this, set the SOCK_NOSPACE on all partial sendmsg scenarios to guarantee that we send EPOLLOUT after partial sendmsg.

After this commit userspace can assume that it will receive an EPOLLOUT after the first partial sendmsg. This EPOLLOUT will benefit from sk_stream_write_space() logic delaying the EPOLLOUT until significant space is available in write queue.

---
 net/ipv4/tcp.c | 20 +++++++++-----------
 1 file changed, 9 insertions(+), 11 deletions(-)
net/ipv4/tcp.c +9 -11
index 48c351804efc..65057744fac8 100644
@@ -1004,12 +1004,12 @@ ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 		    !tcp_skb_can_collapse_to(skb)) {
 new_segment:
 			if (!sk_stream_memory_free(sk))
-				goto wait_for_sndbuf;
+				goto wait_for_space;
 
 			skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 					tcp_rtx_and_write_queues_empty(sk));
 			if (!skb)
-				goto wait_for_memory;
+				goto wait_for_space;
 
 #ifdef CONFIG_TLS_DEVICE
 			skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
@@ -1028,7 +1028,7 @@ ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 			goto new_segment;
 		}
 		if (!sk_wmem_schedule(sk, copy))
-			goto wait_for_memory;
+			goto wait_for_space;
 
 		if (can_coalesce) {
 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
@@ -1069,9 +1069,8 @@ ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 			tcp_push_one(sk, mss_now);
 		continue;
 
-wait_for_sndbuf:
+wait_for_space:
 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
-wait_for_memory:
 		tcp_push(sk, flags & ~MSG_MORE, mss_now,
 			 TCP_NAGLE_PUSH, size_goal);
 
@@ -1282,7 +1281,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 
 new_segment:
 			if (!sk_stream_memory_free(sk))
-				goto wait_for_sndbuf;
+				goto wait_for_space;
 
 			if (unlikely(process_backlog >= 16)) {
 				process_backlog = 0;
@@ -1293,7 +1292,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 			skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
 						  first_skb);
 			if (!skb)
-				goto wait_for_memory;
+				goto wait_for_space;
 
 			process_backlog++;
 			skb->ip_summed = CHECKSUM_PARTIAL;
@@ -1326,7 +1325,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 			struct page_frag *pfrag = sk_page_frag(sk);
 
 			if (!sk_page_frag_refill(sk, pfrag))
-				goto wait_for_memory;
+				goto wait_for_space;
 
 			if (!skb_can_coalesce(skb, i, pfrag->page,
 					      pfrag->offset)) {
@@ -1340,7 +1339,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
 
 			if (!sk_wmem_schedule(sk, copy))
-				goto wait_for_memory;
+				goto wait_for_space;
 
 			err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
 						       pfrag->page,
@@ -1393,9 +1392,8 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 			tcp_push_one(sk, mss_now);
 		continue;
 
-wait_for_sndbuf:
+wait_for_space:
 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
-wait_for_memory:
 		if (copied)
 			tcp_push(sk, flags & ~MSG_MORE, mss_now,
 				 TCP_NAGLE_PUSH, size_goal);
Signature · 1 line
2.28.0.618.gf4bc123cb7-goog
maintainer Applied
#2 lore raw

Re: [PATCH net-next 2/2] tcp: schedule EPOLLOUT after a partial sendmsg

From: Soheil Hassas Yeganeh <[email protected]>
Date: Mon, 14 Sep 2020 17:52:10 -0400

Quoted 23 lines

From: Soheil Hassas Yeganeh <[email protected]>

For EPOLLET, applications must call sendmsg until they get EAGAIN. Otherwise, there is no guarantee that EPOLLOUT is sent if there was a failure upon memory allocation.

As a result on high-speed NICs, userspace observes multiple small sendmsgs after a partial sendmsg until EAGAIN, since TCP can send 1-2 TSOs in between two sendmsg syscalls:

// One large partial send due to memory allocation failure.
sendmsg(20MB) = 2MB
// Many small sends until EAGAIN.
sendmsg(18MB) = 64KB
sendmsg(17.9MB) = 128KB
sendmsg(17.8MB) = 64KB
...
sendmsg(...) = EAGAIN
// At this point, userspace can assume an EPOLLOUT.

To fix this, set the SOCK_NOSPACE on all partial sendmsg scenarios to guarantee that we send EPOLLOUT after partial sendmsg.

Quoted 7 lines

After this commit userspace can assume that it will receive an EPOLLOUT after the first partial sendmsg. This EPOLLOUT will benefit from sk_stream_write_space() logic delaying the EPOLLOUT until significant space is available in write queue.

Applied.

maintainer Applied
#3 lore raw
In reply to Soheil Hassas Yeganeh

From: Soheil Hassas Yeganeh <[email protected]>
Date: Mon, 14 Sep 2020 17:52:09 -0400

Quoted 8 lines

From: Soheil Hassas Yeganeh <[email protected]>

If there was any event available on the TCP socket, tcp_poll() will be called to retrieve all the events. In tcp_poll(), we call sk_stream_is_writeable() which returns true as long as we are at least one byte below notsent_lowat. This will result in quite a few spurious EPLLOUT and frequent tiny sendmsg() calls as a result.

Quoted 6 lines

Similar to sk_stream_write_space(), use __sk_stream_is_writeable with a wake value of 1, so that we set EPOLLOUT only if half the space is available for write.

Applied.