Skip to contentSkip to conversation

net: ip: avoid OOM kills with large UDP sends over loopback

PATCH Superseded Jakub Kicinski 4 participants 7 messages Jun 23, 2021 – Jun 24, 2021

Subsystems (3)

v3 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Jakub Kicinski

author maintainer
#0 lore raw

PATCH v3 1/1 net: ip: avoid OOM kills with large UDP sends over loopback

Dave observed number of machines hitting OOM on the UDP send path. The workload seems to be sending large UDP packets over loopback. Since loopback has MTU of 64k kernel will try to allocate an skb with up to 64k of head space. This has a good chance of failing under memory pressure. What's worse if the message length is <32k the allocation may trigger an OOM killer.

This is entirely avoidable, we can use an skb with page frags.

af_unix solves a similar problem by limiting the head length to SKB_MAX_ALLOC. This seems like a good and simple approach. It means that UDP messages > 16kB will now use fragments if underlying device supports SG, if extra allocator pressure causes regressions in real workloads we can switch to trying the large allocation first and falling back.

---
 net/ipv4/ip_output.c  | 4 +++-
 net/ipv6/ip6_output.c | 4 +++-
 2 files changed, 6 insertions(+), 2 deletions(-)
net/ipv4/ip_output.c +3 -1
index c3efc7d658f6..790dd28fd198 100644
@@ -1077,7 +1077,9 @@ static int __ip_append_data(struct sock *sk,
 			if ((flags & MSG_MORE) &&
 			    !(rt->dst.dev->features&NETIF_F_SG))
 				alloclen = mtu;
-			else if (!paged)
+			else if (!paged &&
+				 (fraglen + hh_len + 15 < SKB_MAX_ALLOC ||
+				  !(rt->dst.dev->features & NETIF_F_SG)))
 				alloclen = fraglen;
 			else {
 				alloclen = min_t(int, fraglen, MAX_HEADER);
net/ipv6/ip6_output.c +3 -1
index ff4f9ebcf7f6..ae8dbd6cdab1 100644
@@ -1585,7 +1585,9 @@ static int __ip6_append_data(struct sock *sk,
 			if ((flags & MSG_MORE) &&
 			    !(rt->dst.dev->features&NETIF_F_SG))
 				alloclen = mtu;
-			else if (!paged)
+			else if (!paged &&
+				 (fraglen + hh_len < SKB_MAX_ALLOC ||
+				  !(rt->dst.dev->features & NETIF_F_SG)))
 				alloclen = fraglen;
 			else {
 				alloclen = min_t(int, fraglen, MAX_HEADER);
Signature · 1 line
2.31.1
Reviewed-by on v3 1/1
#1 lore raw
Quoted 18 lines · Jakub Kicinski

Dave observed number of machines hitting OOM on the UDP send path. The workload seems to be sending large UDP packets over loopback. Since loopback has MTU of 64k kernel will try to allocate an skb with up to 64k of head space. This has a good chance of failing under memory pressure. What's worse if the message length is <32k the allocation may trigger an OOM killer.

This is entirely avoidable, we can use an skb with page frags.

af_unix solves a similar problem by limiting the head length to SKB_MAX_ALLOC. This seems like a good and simple approach. It means that UDP messages > 16kB will now use fragments if underlying device supports SG, if extra allocator pressure causes regressions in real workloads we can switch to trying the large allocation first and falling back.

Quoted 3 lines

---

Thanks !

#2 lore raw
Quoted 21 lines · Eric Dumazet
Quoted 18 lines · Jakub Kicinski

Dave observed number of machines hitting OOM on the UDP send path. The workload seems to be sending large UDP packets over loopback. Since loopback has MTU of 64k kernel will try to allocate an skb with up to 64k of head space. This has a good chance of failing under memory pressure. What's worse if the message length is <32k the allocation may trigger an OOM killer.

This is entirely avoidable, we can use an skb with page frags.

af_unix solves a similar problem by limiting the head length to SKB_MAX_ALLOC. This seems like a good and simple approach. It means that UDP messages > 16kB will now use fragments if underlying device supports SG, if extra allocator pressure causes regressions in real workloads we can switch to trying the large allocation first and falling back.

Quoted 8 lines
Quoted 3 lines

---

Thanks !

I am taking this back.

IPv6 side also needs to account for sizeof(struct frag_hdr) ?

#3 lore raw
In reply to Jakub Kicinski
Quoted 30 lines · On Wed, 23 Jun 2021 09:23:28 -0700 Jakub Kicinski <[email protected]> wrote:

Dave observed number of machines hitting OOM on the UDP send path. The workload seems to be sending large UDP packets over loopback. Since loopback has MTU of 64k kernel will try to allocate an skb with up to 64k of head space. This has a good chance of failing under memory pressure. What's worse if the message length is <32k the allocation may trigger an OOM killer.

This is entirely avoidable, we can use an skb with page frags.

af_unix solves a similar problem by limiting the head length to SKB_MAX_ALLOC. This seems like a good and simple approach. It means that UDP messages > 16kB will now use fragments if underlying device supports SG, if extra allocator pressure causes regressions in real workloads we can switch to trying the large allocation first and falling back.

---
 net/ipv4/ip_output.c  | 4 +++-
 net/ipv6/ip6_output.c | 4 +++-
 2 files changed, 6 insertions(+), 2 deletions(-)
net/ipv4/ip_output.c +0 -0
index c3efc7d658f6..790dd28fd198 100644

net/ipv4/ip_output.c static int __ip_append_data(struct sock *sk, · line 1081

Quoted 6 lines
net/ipv4/ip_output.c +2 -1
@@ -1077,7 +1077,9 @@ static int __ip_append_data(struct sock *sk,
 			if ((flags & MSG_MORE) &&
 			    !(rt->dst.dev->features&NETIF_F_SG))
 				alloclen = mtu;
-			else if (!paged)
+			else if (!paged &&
+				 (fraglen + hh_len + 15 < SKB_MAX_ALLOC ||

What does the number 15 represent here?

Quoted 9 lines
net/ipv4/ip_output.c +1 -0
+				  !(rt->dst.dev->features & NETIF_F_SG)))
 				alloclen = fraglen;
 			else {
 				alloclen = min_t(int, fraglen, MAX_HEADER);
net/ipv6/ip6_output.c +0 -0
index ff4f9ebcf7f6..ae8dbd6cdab1 100644

net/ipv6/ip6_output.c static int __ip6_append_data(struct sock *sk, · line 1589

Quoted 6 lines
net/ipv6/ip6_output.c +2 -1
@@ -1585,7 +1585,9 @@ static int __ip6_append_data(struct sock *sk,
 			if ((flags & MSG_MORE) &&
 			    !(rt->dst.dev->features&NETIF_F_SG))
 				alloclen = mtu;
-			else if (!paged)
+			else if (!paged &&
+				 (fraglen + hh_len < SKB_MAX_ALLOC ||

The number 15 is not use here.

Quoted 4 lines
net/ipv6/ip6_output.c +1 -0
+				  !(rt->dst.dev->features & NETIF_F_SG)))
 				alloclen = fraglen;
 			else {
 				alloclen = min_t(int, fraglen, MAX_HEADER);
Signature · 4 lines
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer
author maintainer
#4 lore raw
Quoted 5 lines · On Wed, 23 Jun 2021 21:45:55 +0200 Jesper Dangaard Brouer wrote:
Quoted 5 lines
net/ipv4/ip_output.c +0 -0
index c3efc7d658f6..790dd28fd198 100644
@@ -1077,7 +1077,9 @@ static int __ip_append_data(struct sock *sk,
Quoted 8 lines

net/ipv4/ip_output.c static int __ip_append_data(struct sock *sk, · line 1081

Quoted 6 lines
net/ipv4/ip_output.c +2 -1
 			if ((flags & MSG_MORE) &&
 			    !(rt->dst.dev->features&NETIF_F_SG))
 				alloclen = mtu;
-			else if (!paged)
+			else if (!paged &&
+				 (fraglen + hh_len + 15 < SKB_MAX_ALLOC ||  

What does the number 15 represent here?

No idea, it's there on the allocation line, so I need to include it on the size check.

Looking at super old code (2.4.x) it looks like it may have gotten copy & pasted mistakenly? The hard headers are rounded up to 16B, and there is code which does things like:

	skb_alloc(size + dev->hard_header_len + 15);
	skb_reserve(skb, (dev->hard_header_len + 15) & ~15);

in other spots. So if I was to guess I'd say someone decided to add the 15B "to be safe" even though hh_len already includes the round up here.

But that's just my guess. I can't get this simple patch right, so take that with a grain of salt :/

Quoted 22 lines
Quoted 9 lines
net/ipv4/ip_output.c +1 -0
+				  !(rt->dst.dev->features & NETIF_F_SG)))
 				alloclen = fraglen;
 			else {
 				alloclen = min_t(int, fraglen, MAX_HEADER);
net/ipv6/ip6_output.c +0 -0
index ff4f9ebcf7f6..ae8dbd6cdab1 100644

net/ipv6/ip6_output.c static int __ip6_append_data(struct sock *sk, · line 1589

Quoted 6 lines
net/ipv6/ip6_output.c +2 -1
@@ -1585,7 +1585,9 @@ static int __ip6_append_data(struct sock *sk,
 			if ((flags & MSG_MORE) &&
 			    !(rt->dst.dev->features&NETIF_F_SG))
 				alloclen = mtu;
-			else if (!paged)
+			else if (!paged &&
+				 (fraglen + hh_len < SKB_MAX_ALLOC ||  

The number 15 is not use here.

Quoted 4 lines
net/ipv6/ip6_output.c +1 -0
+				  !(rt->dst.dev->features & NETIF_F_SG)))
 				alloclen = fraglen;
 			else {
 				alloclen = min_t(int, fraglen, MAX_HEADER);  
#6 lore raw
In reply to Jesper Dangaard Brouer
Quoted 28 lines · Jesper Dangaard Brouer
Quoted 25 lines · On Wed, 23 Jun 2021 09:23:28 -0700 Jakub Kicinski <[email protected]> wrote:

Dave observed number of machines hitting OOM on the UDP send path. The workload seems to be sending large UDP packets over loopback. Since loopback has MTU of 64k kernel will try to allocate an skb with up to 64k of head space. This has a good chance of failing under memory pressure. What's worse if the message length is <32k the allocation may trigger an OOM killer.

This is entirely avoidable, we can use an skb with page frags.

af_unix solves a similar problem by limiting the head length to SKB_MAX_ALLOC. This seems like a good and simple approach. It means that UDP messages > 16kB will now use fragments if underlying device supports SG, if extra allocator pressure causes regressions in real workloads we can switch to trying the large allocation first and falling back.

---
 net/ipv4/ip_output.c  | 4 +++-
 net/ipv6/ip6_output.c | 4 +++-
 2 files changed, 6 insertions(+), 2 deletions(-)
Quoted 13 lines

net/ipv4/ip_output.c static int __ip_append_data(struct sock *sk, · line 1081

Quoted 11 lines
net/ipv4/ip_output.c +2 -1
index c3efc7d658f6..790dd28fd198 100644
@@ -1077,7 +1077,9 @@ static int __ip_append_data(struct sock *sk,
 			if ((flags & MSG_MORE) &&
 			    !(rt->dst.dev->features&NETIF_F_SG))
 				alloclen = mtu;
-			else if (!paged)
+			else if (!paged &&
+				 (fraglen + hh_len + 15 < SKB_MAX_ALLOC ||

What does the number 15 represent here?

Just look at the existing code, few lines below

skb = alloc_skb(alloclen + hh_len + 15, ...

Quoted 10 lines
Quoted 10 lines
net/ipv4/ip_output.c +1 -0
+				  !(rt->dst.dev->features & NETIF_F_SG)))
 				alloclen = fraglen;
 			else {
 				alloclen = min_t(int, fraglen, MAX_HEADER);
net/ipv6/ip6_output.c +0 -0
index ff4f9ebcf7f6..ae8dbd6cdab1 100644
@@ -1585,7 +1585,9 @@ static int __ip6_append_data(struct sock *sk,
Quoted 8 lines

net/ipv6/ip6_output.c static int __ip6_append_data(struct sock *sk, · line 1589

Quoted 6 lines
net/ipv6/ip6_output.c +2 -1
 			if ((flags & MSG_MORE) &&
 			    !(rt->dst.dev->features&NETIF_F_SG))
 				alloclen = mtu;
-			else if (!paged)
+			else if (!paged &&
+				 (fraglen + hh_len < SKB_MAX_ALLOC ||

The number 15 is not use here.

Because the alloc_skb() done later does not use + 15

Quoted 8 lines
Quoted 5 lines
net/ipv6/ip6_output.c +1 -0
+				  !(rt->dst.dev->features & NETIF_F_SG)))
 				alloclen = fraglen;
 			else {
 				alloclen = min_t(int, fraglen, MAX_HEADER);
#5 lore raw
In reply to Jakub Kicinski
Quoted 23 lines · Jakub Kicinski
Quoted 5 lines · On Wed, 23 Jun 2021 21:45:55 +0200 Jesper Dangaard Brouer wrote:
Quoted 5 lines
net/ipv4/ip_output.c +0 -0
index c3efc7d658f6..790dd28fd198 100644
@@ -1077,7 +1077,9 @@ static int __ip_append_data(struct sock *sk,
Quoted 8 lines

net/ipv4/ip_output.c static int __ip_append_data(struct sock *sk, · line 1081

Quoted 6 lines
net/ipv4/ip_output.c +2 -1
                    if ((flags & MSG_MORE) &&
                        !(rt->dst.dev->features&NETIF_F_SG))
                            alloclen = mtu;
-                   else if (!paged)
+                   else if (!paged &&
+                            (fraglen + hh_len + 15 < SKB_MAX_ALLOC ||

What does the number 15 represent here?

No idea, it's there on the allocation line, so I need to include it on the size check.

Looking at super old code (2.4.x) it looks like it may have gotten copy & pasted mistakenly? The hard headers are rounded up to 16B, and there is code which does things like:

Quoted 5 lines
        skb_alloc(size + dev->hard_header_len + 15);
        skb_reserve(skb, (dev->hard_header_len + 15) & ~15);

in other spots. So if I was to guess I'd say someone decided to add the 15B "to be safe" even though hh_len already includes the round up here.

The 15 seems to come from alignment indeed. Not sure when it was introduced, but until 56951b54e87a there is also this

/*
* Get the memory we require with some space left
for alignment.

                 */

                skb = sock_alloc_send_skb(sk, fraglen+hh_len+15, 0,

flags&MSG_DONTWAIT, &err);