net/sched: sch_frag: add generic packet fragment support.
authorwenxu <wenxu@ucloud.cn>
Wed, 25 Nov 2020 04:01:23 +0000 (12:01 +0800)
committerJakub Kicinski <kuba@kernel.org>
Fri, 27 Nov 2020 22:36:02 +0000 (14:36 -0800)
Currently kernel tc subsystem can do conntrack in cat_ct. But when several
fragment packets go through the act_ct, function tcf_ct_handle_fragments
will defrag the packets to a big one. But the last action will redirect
mirred to a device which maybe lead the reassembly big packet over the mtu
of target device.

This patch add support for a xmit hook to mirred, that gets executed before
xmiting the packet. Then, when act_ct gets loaded, it configs that hook.
The frag xmit hook maybe reused by other modules.

Signed-off-by: wenxu <wenxu@ucloud.cn>
Acked-by: Cong Wang <cong.wang@bytedance.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
include/net/act_api.h
include/net/sch_generic.h
net/sched/Makefile
net/sched/act_api.c
net/sched/act_ct.c
net/sched/act_mirred.c
net/sched/sch_frag.c [new file with mode: 0644]

index 8721492..55dab60 100644 (file)
@@ -239,6 +239,12 @@ int tcf_action_check_ctrlact(int action, struct tcf_proto *tp,
                             struct netlink_ext_ack *newchain);
 struct tcf_chain *tcf_action_set_ctrlact(struct tc_action *a, int action,
                                         struct tcf_chain *newchain);
+
+#ifdef CONFIG_INET
+DECLARE_STATIC_KEY_FALSE(tcf_frag_xmit_count);
+#endif
+
+int tcf_dev_queue_xmit(struct sk_buff *skb, int (*xmit)(struct sk_buff *skb));
 #endif /* CONFIG_NET_CLS_ACT */
 
 static inline void tcf_action_stats_update(struct tc_action *a, u64 bytes,
index dd74f06..162ed62 100644 (file)
@@ -1281,4 +1281,6 @@ void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
 void mini_qdisc_pair_block_init(struct mini_Qdisc_pair *miniqp,
                                struct tcf_block *block);
 
+int sch_frag_xmit_hook(struct sk_buff *skb, int (*xmit)(struct sk_buff *skb));
+
 #endif
index 66bbf9a..dd14ef4 100644 (file)
@@ -5,6 +5,7 @@
 
 obj-y  := sch_generic.o sch_mq.o
 
+obj-$(CONFIG_INET)             += sch_frag.o
 obj-$(CONFIG_NET_SCHED)                += sch_api.o sch_blackhole.o
 obj-$(CONFIG_NET_CLS)          += cls_api.o
 obj-$(CONFIG_NET_CLS_ACT)      += act_api.o
index 99db1c7..2e85b63 100644 (file)
 #include <net/act_api.h>
 #include <net/netlink.h>
 
+#ifdef CONFIG_INET
+DEFINE_STATIC_KEY_FALSE(tcf_frag_xmit_count);
+EXPORT_SYMBOL_GPL(tcf_frag_xmit_count);
+#endif
+
+int tcf_dev_queue_xmit(struct sk_buff *skb, int (*xmit)(struct sk_buff *skb))
+{
+#ifdef CONFIG_INET
+       if (static_branch_unlikely(&tcf_frag_xmit_count))
+               return sch_frag_xmit_hook(skb, xmit);
+#endif
+
+       return xmit(skb);
+}
+EXPORT_SYMBOL_GPL(tcf_dev_queue_xmit);
+
 static void tcf_action_goto_chain_exec(const struct tc_action *a,
                                       struct tcf_result *res)
 {
index aba3cd8..61092c5 100644 (file)
@@ -1541,6 +1541,8 @@ static int __init ct_init_module(void)
        if (err)
                goto err_register;
 
+       static_branch_inc(&tcf_frag_xmit_count);
+
        return 0;
 
 err_register:
@@ -1552,6 +1554,7 @@ err_tbl_init:
 
 static void __exit ct_cleanup_module(void)
 {
+       static_branch_dec(&tcf_frag_xmit_count);
        tcf_unregister_action(&act_ct_ops, &ct_net_ops);
        tcf_ct_flow_tables_uninit();
        destroy_workqueue(act_ct_wq);
index 17d0095..7153c67 100644 (file)
@@ -210,7 +210,7 @@ static int tcf_mirred_forward(bool want_ingress, struct sk_buff *skb)
        int err;
 
        if (!want_ingress)
-               err = dev_queue_xmit(skb);
+               err = tcf_dev_queue_xmit(skb, dev_queue_xmit);
        else
                err = netif_receive_skb(skb);
 
diff --git a/net/sched/sch_frag.c b/net/sched/sch_frag.c
new file mode 100644 (file)
index 0000000..e1e77d3
--- /dev/null
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
+#include <net/netlink.h>
+#include <net/sch_generic.h>
+#include <net/dst.h>
+#include <net/ip.h>
+#include <net/ip6_fib.h>
+
+struct sch_frag_data {
+       unsigned long dst;
+       struct qdisc_skb_cb cb;
+       __be16 inner_protocol;
+       u16 vlan_tci;
+       __be16 vlan_proto;
+       unsigned int l2_len;
+       u8 l2_data[VLAN_ETH_HLEN];
+       int (*xmit)(struct sk_buff *skb);
+};
+
+static DEFINE_PER_CPU(struct sch_frag_data, sch_frag_data_storage);
+
+static int sch_frag_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
+{
+       struct sch_frag_data *data = this_cpu_ptr(&sch_frag_data_storage);
+
+       if (skb_cow_head(skb, data->l2_len) < 0) {
+               kfree_skb(skb);
+               return -ENOMEM;
+       }
+
+       __skb_dst_copy(skb, data->dst);
+       *qdisc_skb_cb(skb) = data->cb;
+       skb->inner_protocol = data->inner_protocol;
+       if (data->vlan_tci & VLAN_CFI_MASK)
+               __vlan_hwaccel_put_tag(skb, data->vlan_proto,
+                                      data->vlan_tci & ~VLAN_CFI_MASK);
+       else
+               __vlan_hwaccel_clear_tag(skb);
+
+       /* Reconstruct the MAC header.  */
+       skb_push(skb, data->l2_len);
+       memcpy(skb->data, &data->l2_data, data->l2_len);
+       skb_postpush_rcsum(skb, skb->data, data->l2_len);
+       skb_reset_mac_header(skb);
+
+       return data->xmit(skb);
+}
+
+static void sch_frag_prepare_frag(struct sk_buff *skb,
+                                 int (*xmit)(struct sk_buff *skb))
+{
+       unsigned int hlen = skb_network_offset(skb);
+       struct sch_frag_data *data;
+
+       data = this_cpu_ptr(&sch_frag_data_storage);
+       data->dst = skb->_skb_refdst;
+       data->cb = *qdisc_skb_cb(skb);
+       data->xmit = xmit;
+       data->inner_protocol = skb->inner_protocol;
+       if (skb_vlan_tag_present(skb))
+               data->vlan_tci = skb_vlan_tag_get(skb) | VLAN_CFI_MASK;
+       else
+               data->vlan_tci = 0;
+       data->vlan_proto = skb->vlan_proto;
+       data->l2_len = hlen;
+       memcpy(&data->l2_data, skb->data, hlen);
+
+       memset(IPCB(skb), 0, sizeof(struct inet_skb_parm));
+       skb_pull(skb, hlen);
+}
+
+static unsigned int
+sch_frag_dst_get_mtu(const struct dst_entry *dst)
+{
+       return dst->dev->mtu;
+}
+
+static struct dst_ops sch_frag_dst_ops = {
+       .family = AF_UNSPEC,
+       .mtu = sch_frag_dst_get_mtu,
+};
+
+static int sch_fragment(struct net *net, struct sk_buff *skb,
+                       u16 mru, int (*xmit)(struct sk_buff *skb))
+{
+       int ret = -1;
+
+       if (skb_network_offset(skb) > VLAN_ETH_HLEN) {
+               net_warn_ratelimited("L2 header too long to fragment\n");
+               goto err;
+       }
+
+       if (skb_protocol(skb, true) == htons(ETH_P_IP)) {
+               struct dst_entry sch_frag_dst;
+               unsigned long orig_dst;
+
+               sch_frag_prepare_frag(skb, xmit);
+               dst_init(&sch_frag_dst, &sch_frag_dst_ops, NULL, 1,
+                        DST_OBSOLETE_NONE, DST_NOCOUNT);
+               sch_frag_dst.dev = skb->dev;
+
+               orig_dst = skb->_skb_refdst;
+               skb_dst_set_noref(skb, &sch_frag_dst);
+               IPCB(skb)->frag_max_size = mru;
+
+               ret = ip_do_fragment(net, skb->sk, skb, sch_frag_xmit);
+               refdst_drop(orig_dst);
+       } else if (skb_protocol(skb, true) == htons(ETH_P_IPV6)) {
+               unsigned long orig_dst;
+               struct rt6_info sch_frag_rt;
+
+               sch_frag_prepare_frag(skb, xmit);
+               memset(&sch_frag_rt, 0, sizeof(sch_frag_rt));
+               dst_init(&sch_frag_rt.dst, &sch_frag_dst_ops, NULL, 1,
+                        DST_OBSOLETE_NONE, DST_NOCOUNT);
+               sch_frag_rt.dst.dev = skb->dev;
+
+               orig_dst = skb->_skb_refdst;
+               skb_dst_set_noref(skb, &sch_frag_rt.dst);
+               IP6CB(skb)->frag_max_size = mru;
+
+               ret = ipv6_stub->ipv6_fragment(net, skb->sk, skb,
+                                              sch_frag_xmit);
+               refdst_drop(orig_dst);
+       } else {
+               net_warn_ratelimited("Fail frag %s: eth=%x, MRU=%d, MTU=%d\n",
+                                    netdev_name(skb->dev),
+                                    ntohs(skb_protocol(skb, true)), mru,
+                                    skb->dev->mtu);
+               goto err;
+       }
+
+       return ret;
+err:
+       kfree_skb(skb);
+       return ret;
+}
+
+int sch_frag_xmit_hook(struct sk_buff *skb, int (*xmit)(struct sk_buff *skb))
+{
+       u16 mru = qdisc_skb_cb(skb)->mru;
+       int err;
+
+       if (mru && skb->len > mru + skb->dev->hard_header_len)
+               err = sch_fragment(dev_net(skb->dev), skb, mru, xmit);
+       else
+               err = xmit(skb);
+
+       return err;
+}
+EXPORT_SYMBOL_GPL(sch_frag_xmit_hook);