net: permit skb_segment on head_frag frag_list skb

One of our in-house projects, bpf-based NAT, hits a kernel BUG_ON at function skb_segment(), line 3667. The bpf program attaches to clsact ingress, calls bpf_skb_change_proto to change protocol from ipv4 to ipv6 or from ipv6 to ipv4, and then calls bpf_redirect to send the changed packet out.
...

    3665                 while (pos < offset + len) {
    3666                         if (i >= nfrags) {
    3667                                 BUG_ON(skb_headlen(list_skb));

...

The triggering input skb has the following properties:

    list_skb = skb->frag_list;
    skb->nfrags != NULL && skb_headlen(list_skb) != 0

and skb_segment() is not able to handle a frag_list skb if its headlen (list_skb->len - list_skb->data_len) is not 0.

Patch #1 provides a simple solution to avoid BUG_ON. If list_skb->head_frag is true, its page-backed frag will be processed before the list_skb->frags.
Patch #2 provides a test case in test_bpf module which constructs a skb and calls skb_segment() directly. The test case is able to trigger the BUG_ON without Patch #1.

The patch has been tested in the following setup:
ipv6_host <-> nat_server <-> ipv4_host
where nat_server has a bpf program doing ipv4<->ipv6 translation and forwarding through clsact hook bpf_skb_change_proto.

Changelog:
v3 -> v4:
. Remove dynamic memory allocation and use rewinding
for both index and frag to remove one branch in fast path, from Alexander.
. Fix a bunch of issues in test_bpf skb_segment() test,
including proper way to allocate skb, proper function argument for skb_add_rx_frag and not freeint skb, etc., from Eric.
v2 -> v3:
. Use starting frag index -1 (instead of 0) to
special process head_frag before other frags in the skb, from Alexander Duyck.
v1 -> v2:
. Removed never-hit BUG_ON, spotted by Linyu Yuan.

Yonghong Song (2):
net: permit skb_segment on head_frag frag_list skb
net: bpf: add a test for skb_segment in test_bpf module

 lib/test_bpf.c    | 91 +++++++++++++++++++++++++++++++++++++++++++++++++++++--
 net/core/skbuff.c | 36 +++++++++++++++-------
 2 files changed, 114 insertions(+), 13 deletions(-)

2.9.5

One of our in-house projects, bpf-based NAT, hits a kernel BUG_ON at function skb_segment(), line 3667. The bpf program attaches to clsact ingress, calls bpf_skb_change_proto to change protocol from ipv4 to ipv6 or from ipv6 to ipv4, and then calls bpf_redirect to send the changed packet out.

3472 struct sk_buff *skb_segment(struct sk_buff *head_skb,
3473 netdev_features_t features)
3474 {
3475 struct sk_buff *segs = NULL;
3476 struct sk_buff *tail = NULL;
...
3665 while (pos < offset + len) {
3666 if (i >= nfrags) {
3667 BUG_ON(skb_headlen(list_skb));
3668

3669                                 i = 0;
3670                                 nfrags = skb_shinfo(list_skb)->nr_frags;
3671                                 frag = skb_shinfo(list_skb)->frags;
3672                                 frag_skb = list_skb;

...

call stack:
...
#1 [ffff883ffef03558] __crash_kexec at ffffffff8110c525 #2 [ffff883ffef03620] crash_kexec at ffffffff8110d5cc #3 [ffff883ffef03640] oops_end at ffffffff8101d7e7 #4 [ffff883ffef03668] die at ffffffff8101deb2 #5 [ffff883ffef03698] do_trap at ffffffff8101a700 #6 [ffff883ffef036e8] do_error_trap at ffffffff8101abfe #7 [ffff883ffef037a0] do_invalid_op at ffffffff8101acd0 #8 [ffff883ffef037b0] invalid_op at ffffffff81a00bab

    [exception RIP: skb_segment+3044]
    RIP: ffffffff817e4dd4  RSP: ffff883ffef03860  RFLAGS: 00010216
    RAX: 0000000000002bf6  RBX: ffff883feb7aaa00  RCX: 0000000000000011
    RDX: ffff883fb87910c0  RSI: 0000000000000011  RDI: ffff883feb7ab500
    RBP: ffff883ffef03928   R8: 0000000000002ce2   R9: 00000000000027da
    R10: 000001ea00000000  R11: 0000000000002d82  R12: ffff883f90a1ee80
    R13: ffff883fb8791120  R14: ffff883feb7abc00  R15: 0000000000002ce2
    ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0018

#9 [ffff883ffef03930] tcp_gso_segment at ffffffff818713e7
--- <IRQ stack> ---
...

The triggering input skb has the following properties:

    list_skb = skb->frag_list;
    skb->nfrags != NULL && skb_headlen(list_skb) != 0

and skb_segment() is not able to handle a frag_list skb if its headlen (list_skb->len - list_skb->data_len) is not 0.

This patch addressed the issue by handling skb_headlen(list_skb) != 0 case properly if list_skb->head_frag is true, which is expected in most cases. The head frag is processed before list_skb->frags are processed.

---
 net/core/skbuff.c | 36 +++++++++++++++++++++++++-----------
 1 file changed, 25 insertions(+), 11 deletions(-)

index 715c134..09f4c24 100644
@@ -3475,7 +3475,7 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb,
 	struct sk_buff *segs = NULL;
 	struct sk_buff *tail = NULL;
 	struct sk_buff *list_skb = skb_shinfo(head_skb)->frag_list;
-	skb_frag_t *frag = skb_shinfo(head_skb)->frags;
+	skb_frag_t *frag = skb_shinfo(head_skb)->frags, head_frag;
 	unsigned int mss = skb_shinfo(head_skb)->gso_size;
 	unsigned int doffset = head_skb->data - skb_mac_header(head_skb);
 	struct sk_buff *frag_skb = head_skb;

@@ -3664,19 +3664,30 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb,
 
 		while (pos < offset + len) {
 			if (i >= nfrags) {
-				BUG_ON(skb_headlen(list_skb));
-
 				i = 0;
 				nfrags = skb_shinfo(list_skb)->nr_frags;
 				frag = skb_shinfo(list_skb)->frags;
-				frag_skb = list_skb;
-
-				BUG_ON(!nfrags);
+				if (skb_headlen(list_skb)) {
+					struct page *page;
+
+					BUG_ON(!list_skb->head_frag);
+
+					page = virt_to_head_page(list_skb->head);
+					head_frag.page.p = page;
+					head_frag.page_offset = list_skb->data -
+						(unsigned char *)page_address(page);
+					head_frag.size = skb_headlen(list_skb);
+					/* to make room for head_frag. */
+					i--; frag--;
+				}
+				if (nfrags) {
+					frag_skb = list_skb;
 
-				if (skb_orphan_frags(frag_skb, GFP_ATOMIC) ||
-				    skb_zerocopy_clone(nskb, frag_skb,
-						       GFP_ATOMIC))
-					goto err;
+					if (skb_orphan_frags(frag_skb, GFP_ATOMIC) ||
+					    skb_zerocopy_clone(nskb, frag_skb,
+							       GFP_ATOMIC))
+						goto err;
+				}
 
 				list_skb = list_skb->next;
 			}

@@ -3689,7 +3700,10 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb,
 				goto err;
 			}
 
-			*nskb_frag = *frag;
+			/* head_frag could be defined in previous outer do/while
+			 * loop iterations.
+			 */
+			*nskb_frag = (i < 0) ? head_frag : *frag;
 			__skb_frag_ref(nskb_frag);
 			size = skb_frag_size(nskb_frag);
 

2.9.5

Without the previous commit,
"modprobe test_bpf" will have the following errors:
...

[   98.149165] ------------[ cut here ]------------
[   98.159362] kernel BUG at net/core/skbuff.c:3667!
[   98.169756] invalid opcode: 0000 [#1] SMP PTI
[   98.179370] Modules linked in:
[   98.179371]  test_bpf(+)

...
which triggers the bug the previous commit intends to fix.

The skbs are constructed to mimic what mlx5 may generate. The packet size/header may not mimic real cases in production. But the processing flow is similar.

---
 lib/test_bpf.c | 91 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 89 insertions(+), 2 deletions(-)

index 2efb213..086a231 100644
@@ -6574,6 +6574,91 @@ static bool exclude_test(int test_id)
 	return test_id < test_range[0] || test_id > test_range[1];
 }
 
+static __init struct sk_buff *build_test_skb(void)
+{
+	u32 headroom = NET_SKB_PAD + NET_IP_ALIGN + ETH_HLEN;
+	struct sk_buff *skb[2];
+	struct page *page[2];
+	int i, data_size = 8;
+
+	for (i = 0; i < 2; i++) {
+		page[i] = alloc_page(GFP_KERNEL);
+		if (!page[i]) {
+			if (i == 0)
+				goto err_page0;
+			else
+				goto err_page1;
+		}
+
+		/* this will set skb[i]->head_frag */
+		skb[i] = dev_alloc_skb(headroom + data_size);
+		if (!skb[i]) {
+			if (i == 0)
+				goto err_skb0;
+			else
+				goto err_skb1;
+		}
+
+		skb_reserve(skb[i], headroom);
+		skb_put(skb[i], data_size);
+		skb[i]->protocol = htons(ETH_P_IP);
+		skb_reset_network_header(skb[i]);
+		skb_set_mac_header(skb[i], -ETH_HLEN);
+
+		skb_add_rx_frag(skb[i], 0, page[i], 0, 64, 64);
+		// skb_headlen(skb[i]): 8, skb[i]->head_frag = 1
+	}
+
+	/* setup shinfo */
+	skb_shinfo(skb[0])->gso_size = 1448;
+	skb_shinfo(skb[0])->gso_type = SKB_GSO_TCPV4;
+	skb_shinfo(skb[0])->gso_type |= SKB_GSO_DODGY;
+	skb_shinfo(skb[0])->gso_segs = 0;
+	skb_shinfo(skb[0])->frag_list = skb[1];
+
+	/* adjust skb[0]'s len */
+	skb[0]->len += skb[1]->len;
+	skb[0]->data_len += skb[1]->data_len;
+	skb[0]->truesize += skb[1]->truesize;
+
+	return skb[0];
+
+err_skb1:
+	__free_page(page[1]);
+err_page1:
+	kfree_skb(skb[0]);
+err_skb0:
+	__free_page(page[0]);
+err_page0:
+	return NULL;
+}
+
+static __init int test_skb_segment(void)
+{
+	netdev_features_t features;
+	struct sk_buff *skb;
+	int ret = -1;
+
+	features = NETIF_F_SG | NETIF_F_GSO_PARTIAL | NETIF_F_IP_CSUM |
+		   NETIF_F_IPV6_CSUM;
+	features |= NETIF_F_RXCSUM;
+	skb = build_test_skb();
+	if (!skb) {
+		pr_info("%s: failed to build_test_skb", __func__);
+		goto done;
+	}
+
+	if (skb_segment(skb, features)) {
+		ret = 0;
+		pr_info("%s: success in skb_segment!", __func__);
+	} else {
+		pr_info("%s: failed in skb_segment!", __func__);
+	}
+	kfree_skb(skb);
+done:
+	return ret;
+}
+
 static __init int test_bpf(void)
 {
 	int i, err_cnt = 0, pass_cnt = 0;

@@ -6632,9 +6717,11 @@ static int __init test_bpf_init(void)
 		return ret;
 
 	ret = test_bpf();
-
 	destroy_bpf_tests();
-	return ret;
+	if (ret)
+		return ret;
+
+	return test_skb_segment();
 }
 
 static void __exit test_bpf_exit(void)

2.9.5

On 03/20/2018 11:47 PM, Yonghong Song wrote:
...

+

I tried with KASAN. It does not panic.
Looking at the code in net/core/dev.c: validate_xmit_skb:

static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
...

        if (netif_needs_gso(skb, features)) {
                struct sk_buff *segs;

                segs = skb_gso_segment(skb, features);
                if (IS_ERR(segs)) {
                        goto out_kfree_skb;
                } else if (segs) {
                        consume_skb(skb);
                        skb = segs;
                }
...
out_kfree_skb:
        kfree_skb(skb);

which also indicates kfree_skb/consume_skb probably is the right way to free skb after skb_gso_segment/skb_segment.

This probably explains why my above kfree_skb(skb) does not crash.

Totally agree. Will make the change. Thanks!