skbuff: increase verbosity when dumping skb data
authorWillem de Bruijn <willemb@google.com>
Sun, 7 Jul 2019 09:51:55 +0000 (05:51 -0400)
committerDavid S. Miller <davem@davemloft.net>
Tue, 9 Jul 2019 02:38:46 +0000 (19:38 -0700)
skb_warn_bad_offload and netdev_rx_csum_fault trigger on hard to debug
issues. Dump more state and the header.

Optionally dump the entire packet and linear segment. This is required
to debug checksum bugs that may include bytes past skb_tail_pointer().

Both call sites call this function inside a net_ratelimit() block.
Limit full packet log further to a hard limit of can_dump_full (5).

Based on an earlier patch by Cong Wang, see link below.

Changes v1 -> v2
  - dump frag_list only on full_pkt

Link: https://patchwork.ozlabs.org/patch/1000841/
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/skbuff.h
net/core/dev.c
net/core/skbuff.c

index 7ece49d5f8ef72be44142ed09944cc50f7833efb..1fdfdbb34e8e8d98cb803615ab820227c969d74d 100644 (file)
@@ -1024,6 +1024,7 @@ static inline bool skb_unref(struct sk_buff *skb)
 void skb_release_head_state(struct sk_buff *skb);
 void kfree_skb(struct sk_buff *skb);
 void kfree_skb_list(struct sk_buff *segs);
+void skb_dump(const char *level, const struct sk_buff *skb, bool full_pkt);
 void skb_tx_error(struct sk_buff *skb);
 void consume_skb(struct sk_buff *skb);
 void __consume_stateless_skb(struct sk_buff *skb);
index 58529318b3a947b079153c0aec65d8740563295d..fc676b2610e3c1e7e236b62d4984647bc0e71916 100644 (file)
@@ -2900,12 +2900,10 @@ static void skb_warn_bad_offload(const struct sk_buff *skb)
                else
                        name = netdev_name(dev);
        }
-       WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
-            "gso_type=%d ip_summed=%d\n",
+       skb_dump(KERN_WARNING, skb, false);
+       WARN(1, "%s: caps=(%pNF, %pNF)\n",
             name, dev ? &dev->features : &null_features,
-            skb->sk ? &skb->sk->sk_route_caps : &null_features,
-            skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
-            skb_shinfo(skb)->gso_type, skb->ip_summed);
+            skb->sk ? &skb->sk->sk_route_caps : &null_features);
 }
 
 /*
@@ -3124,13 +3122,7 @@ void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
 {
        if (net_ratelimit()) {
                pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
-               if (dev)
-                       pr_err("dev features: %pNF\n", &dev->features);
-               pr_err("skb len=%u data_len=%u pkt_type=%u gso_size=%u gso_type=%u nr_frags=%u ip_summed=%u csum=%x csum_complete_sw=%d csum_valid=%d csum_level=%u\n",
-                      skb->len, skb->data_len, skb->pkt_type,
-                      skb_shinfo(skb)->gso_size, skb_shinfo(skb)->gso_type,
-                      skb_shinfo(skb)->nr_frags, skb->ip_summed, skb->csum,
-                      skb->csum_complete_sw, skb->csum_valid, skb->csum_level);
+               skb_dump(KERN_ERR, skb, true);
                dump_stack();
        }
 }
index 5323441a12ccf67726c0250587fb92cf98c64998..cdb0ccdaac0bf3c27b3f7b2ab2a4d299b205926b 100644 (file)
@@ -707,6 +707,105 @@ void kfree_skb_list(struct sk_buff *segs)
 }
 EXPORT_SYMBOL(kfree_skb_list);
 
+/* Dump skb information and contents.
+ *
+ * Must only be called from net_ratelimit()-ed paths.
+ *
+ * Dumps up to can_dump_full whole packets if full_pkt, headers otherwise.
+ */
+void skb_dump(const char *level, const struct sk_buff *skb, bool full_pkt)
+{
+       static atomic_t can_dump_full = ATOMIC_INIT(5);
+       struct skb_shared_info *sh = skb_shinfo(skb);
+       struct net_device *dev = skb->dev;
+       struct sock *sk = skb->sk;
+       struct sk_buff *list_skb;
+       bool has_mac, has_trans;
+       int headroom, tailroom;
+       int i, len, seg_len;
+
+       if (full_pkt)
+               full_pkt = atomic_dec_if_positive(&can_dump_full) >= 0;
+
+       if (full_pkt)
+               len = skb->len;
+       else
+               len = min_t(int, skb->len, MAX_HEADER + 128);
+
+       headroom = skb_headroom(skb);
+       tailroom = skb_tailroom(skb);
+
+       has_mac = skb_mac_header_was_set(skb);
+       has_trans = skb_transport_header_was_set(skb);
+
+       printk("%sskb len=%u headroom=%u headlen=%u tailroom=%u\n"
+              "mac=(%d,%d) net=(%d,%d) trans=%d\n"
+              "shinfo(txflags=%u nr_frags=%u gso(size=%hu type=%u segs=%hu))\n"
+              "csum(0x%x ip_summed=%u complete_sw=%u valid=%u level=%u)\n"
+              "hash(0x%x sw=%u l4=%u) proto=0x%04x pkttype=%u iif=%d\n",
+              level, skb->len, headroom, skb_headlen(skb), tailroom,
+              has_mac ? skb->mac_header : -1,
+              has_mac ? skb_mac_header_len(skb) : -1,
+              skb->network_header,
+              has_trans ? skb_network_header_len(skb) : -1,
+              has_trans ? skb->transport_header : -1,
+              sh->tx_flags, sh->nr_frags,
+              sh->gso_size, sh->gso_type, sh->gso_segs,
+              skb->csum, skb->ip_summed, skb->csum_complete_sw,
+              skb->csum_valid, skb->csum_level,
+              skb->hash, skb->sw_hash, skb->l4_hash,
+              ntohs(skb->protocol), skb->pkt_type, skb->skb_iif);
+
+       if (dev)
+               printk("%sdev name=%s feat=0x%pNF\n",
+                      level, dev->name, &dev->features);
+       if (sk)
+               printk("%ssk family=%hu type=%hu proto=%hu\n",
+                      level, sk->sk_family, sk->sk_type, sk->sk_protocol);
+
+       if (full_pkt && headroom)
+               print_hex_dump(level, "skb headroom: ", DUMP_PREFIX_OFFSET,
+                              16, 1, skb->head, headroom, false);
+
+       seg_len = min_t(int, skb_headlen(skb), len);
+       if (seg_len)
+               print_hex_dump(level, "skb linear:   ", DUMP_PREFIX_OFFSET,
+                              16, 1, skb->data, seg_len, false);
+       len -= seg_len;
+
+       if (full_pkt && tailroom)
+               print_hex_dump(level, "skb tailroom: ", DUMP_PREFIX_OFFSET,
+                              16, 1, skb_tail_pointer(skb), tailroom, false);
+
+       for (i = 0; len && i < skb_shinfo(skb)->nr_frags; i++) {
+               skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
+               u32 p_off, p_len, copied;
+               struct page *p;
+               u8 *vaddr;
+
+               skb_frag_foreach_page(frag, frag->page_offset,
+                                     skb_frag_size(frag), p, p_off, p_len,
+                                     copied) {
+                       seg_len = min_t(int, p_len, len);
+                       vaddr = kmap_atomic(p);
+                       print_hex_dump(level, "skb frag:     ",
+                                      DUMP_PREFIX_OFFSET,
+                                      16, 1, vaddr + p_off, seg_len, false);
+                       kunmap_atomic(vaddr);
+                       len -= seg_len;
+                       if (!len)
+                               break;
+               }
+       }
+
+       if (full_pkt && skb_has_frag_list(skb)) {
+               printk("skb fraglist:\n");
+               skb_walk_frags(skb, list_skb)
+                       skb_dump(level, list_skb, true);
+       }
+}
+EXPORT_SYMBOL(skb_dump);
+
 /**
  *     skb_tx_error - report an sk_buff xmit error
  *     @skb: buffer that triggered an error