Skip to contentSkip to conversation

net: reduce rtnetlink_rcv_msg() stack usage

PATCH Applied Eric Dumazet 3 participants 3 messages Jul 10, 2024 – Jul 12, 2024

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 net: reduce rtnetlink_rcv_msg() stack usage

IFLA_MAX is increasing slowly but surely.

Some compilers use more than 512 bytes of stack in rtnetlink_rcv_msg() because it calls rtnl_calcit() for RTM_GETLINK message.

Use noinline_for_stack attribute to not inline rtnl_calcit(), and directly use nla_for_each_attr_type() (Jakub suggestion) because we only care about IFLA_EXT_MASK at this stage.

---
 net/core/rtnetlink.c | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)
net/core/rtnetlink.c +12 -6
index eabfc8290f5e29f2ef3e5c1481715ae9056ea689..87e67194f24046a8420bbb51c19fb0a686b9b06b 100644
@@ -3969,22 +3969,28 @@ static int rtnl_dellinkprop(struct sk_buff *skb, struct nlmsghdr *nlh,
 	return rtnl_linkprop(RTM_DELLINKPROP, skb, nlh, extack);
 }
 
-static u32 rtnl_calcit(struct sk_buff *skb, struct nlmsghdr *nlh)
+static noinline_for_stack u32 rtnl_calcit(struct sk_buff *skb,
+					  struct nlmsghdr *nlh)
 {
 	struct net *net = sock_net(skb->sk);
 	size_t min_ifinfo_dump_size = 0;
-	struct nlattr *tb[IFLA_MAX+1];
 	u32 ext_filter_mask = 0;
 	struct net_device *dev;
-	int hdrlen;
+	struct nlattr *nla;
+	int hdrlen, rem;
 
 	/* Same kernel<->userspace interface hack as in rtnl_dump_ifinfo. */
 	hdrlen = nlmsg_len(nlh) < sizeof(struct ifinfomsg) ?
 		 sizeof(struct rtgenmsg) : sizeof(struct ifinfomsg);
 
-	if (nlmsg_parse_deprecated(nlh, hdrlen, tb, IFLA_MAX, ifla_policy, NULL) >= 0) {
-		if (tb[IFLA_EXT_MASK])
-			ext_filter_mask = nla_get_u32(tb[IFLA_EXT_MASK]);
+	if (nlh->nlmsg_len < nlmsg_msg_size(hdrlen))
+		return NLMSG_GOODSIZE;
+
+	nla_for_each_attr_type(nla, IFLA_EXT_MASK,
+			       nlmsg_attrdata(nlh, hdrlen),
+			       nlmsg_attrlen(nlh, hdrlen), rem) {
+		if (nla_len(nla) == sizeof(u32))
+			ext_filter_mask = nla_get_u32(nla);
 	}
 
 	if (!ext_filter_mask)
Signature · 1 line
2.45.2.993.g49e7a77208-goog
reviewer Reviewed-by on v2 1/1
#1 lore raw
Quoted 10 lines · Eric Dumazet

IFLA_MAX is increasing slowly but surely.

Some compilers use more than 512 bytes of stack in rtnetlink_rcv_msg() because it calls rtnl_calcit() for RTM_GETLINK message.

Use noinline_for_stack attribute to not inline rtnl_calcit(), and directly use nla_for_each_attr_type() (Jakub suggestion) because we only care about IFLA_EXT_MASK at this stage.

Quoted 30 lines
---
 net/core/rtnetlink.c | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)
net/core/rtnetlink.c +4 -4
index eabfc8290f5e29f2ef3e5c1481715ae9056ea689..87e67194f24046a8420bbb51c19fb0a686b9b06b 100644
@@ -3969,22 +3969,28 @@ static int rtnl_dellinkprop(struct sk_buff *skb, struct nlmsghdr *nlh,
 	return rtnl_linkprop(RTM_DELLINKPROP, skb, nlh, extack);
 }
 
-static u32 rtnl_calcit(struct sk_buff *skb, struct nlmsghdr *nlh)
+static noinline_for_stack u32 rtnl_calcit(struct sk_buff *skb,
+					  struct nlmsghdr *nlh)
 {
 	struct net *net = sock_net(skb->sk);
 	size_t min_ifinfo_dump_size = 0;
-	struct nlattr *tb[IFLA_MAX+1];
 	u32 ext_filter_mask = 0;
 	struct net_device *dev;
-	int hdrlen;
+	struct nlattr *nla;
+	int hdrlen, rem;
 
 	/* Same kernel<->userspace interface hack as in rtnl_dump_ifinfo. */
 	hdrlen = nlmsg_len(nlh) < sizeof(struct ifinfomsg) ?
 		 sizeof(struct rtgenmsg) : sizeof(struct ifinfomsg);
 
-	if (nlmsg_parse_deprecated(nlh, hdrlen, tb, IFLA_MAX, ifla_policy, NULL) >= 0) {

net/core/rtnetlink.c static int rtnl_dellinkprop(struct sk_buff *skb, struct nlmsghdr *nlh, · line 3993

Quoted 10 lines
net/core/rtnetlink.c +8 -2
-		if (tb[IFLA_EXT_MASK])
-			ext_filter_mask = nla_get_u32(tb[IFLA_EXT_MASK]);
+	if (nlh->nlmsg_len < nlmsg_msg_size(hdrlen))
+		return NLMSG_GOODSIZE;
+
+	nla_for_each_attr_type(nla, IFLA_EXT_MASK,
+			       nlmsg_attrdata(nlh, hdrlen),
+			       nlmsg_attrlen(nlh, hdrlen), rem) {
+		if (nla_len(nla) == sizeof(u32))
+			ext_filter_mask = nla_get_u32(nla);

I guess that in practice we can break here.
But perhaps there is some case where there is more than one IFLA_EXT_MASK attribute and it changes behaviour.

Quoted 7 lines
net/core/rtnetlink.c +0 -0
 	}
 
 	if (!ext_filter_mask)
Signature · 1 line
2.45.2.993.g49e7a77208-goog
#2 lore raw
In reply to Eric Dumazet
This patch was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Hello:

This patch was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Quoted 10 lines · On Wed, 10 Jul 2024 15:16:53 +0000 you wrote:

IFLA_MAX is increasing slowly but surely.

Some compilers use more than 512 bytes of stack in rtnetlink_rcv_msg() because it calls rtnl_calcit() for RTM_GETLINK message.

Use noinline_for_stack attribute to not inline rtnl_calcit(), and directly use nla_for_each_attr_type() (Jakub suggestion) because we only care about IFLA_EXT_MASK at this stage.

[...]

Here is the summary with links:
- [v2,net-next] net: reduce rtnetlink_rcv_msg() stack usage

    https://git.kernel.org/netdev/net-next/c/cef4902b0fad

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html