From: Fred Li <[email protected]>
The crashed kernel version is 5.16.20, and I have not test this patch
because I dont find a way to reproduce it, and the mailine may be
has the same problem.
When using bpf based NAT, hits a kernel BUG_ON at function skb_segment(),
BUG_ON(skb_headlen(list_skb) > len). The bpf calls the bpf_skb_adjust_room
to decrease the gso_size, and then call bpf_redirect send packet out.
call stack:
..
[exception RIP: skb_segment+3016]
RIP: ffffffffb97df2a8 RSP: ffffa3f2cce08728 RFLAGS: 00010293
RAX: 000000000000007d RBX: 00000000fffff7b3 RCX: 0000000000000011
RDX: 0000000000000000 RSI: ffff895ea32c76c0 RDI: 00000000000008c1
RBP: ffffa3f2cce087f8 R8: 000000000000088f R9: 0000000000000011
R10: 000000000000090c R11: ffff895e47e68000 R12: ffff895eb2022f00
R13: 000000000000004b R14: ffff895ecdaf2000 R15: ffff895eb2023f00
ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018
#9 [ffffa3f2cce08720] skb_segment at ffffffffb97ded63
#10 [ffffa3f2cce08800] tcp_gso_segment at ffffffffb98d0320
#11 [ffffa3f2cce08860] tcp4_gso_segment at ffffffffb98d07a3
#12 [ffffa3f2cce08880] inet_gso_segment at ffffffffb98e6de0
#13 [ffffa3f2cce088e0] skb_mac_gso_segment at ffffffffb97f3741
#14 [ffffa3f2cce08918] skb_udp_tunnel_segment at ffffffffb98daa59
#15 [ffffa3f2cce08980] udp4_ufo_fragment at ffffffffb98db471
#16 [ffffa3f2cce089b0] inet_gso_segment at ffffffffb98e6de0
#17 [ffffa3f2cce08a10] skb_mac_gso_segment at ffffffffb97f3741
#18 [ffffa3f2cce08a48] __skb_gso_segment at ffffffffb97f388e
#19 [ffffa3f2cce08a78] validate_xmit_skb at ffffffffb97f3d6e
#20 [ffffa3f2cce08ab8] __dev_queue_xmit at ffffffffb97f4614
#21 [ffffa3f2cce08b50] dev_queue_xmit at ffffffffb97f5030
#22 [ffffa3f2cce08b60] __bpf_redirect at ffffffffb98199a8
#23 [ffffa3f2cce08b88] skb_do_redirect at ffffffffb98205cd
..
The skb has the following properties:
doffset = 66
list_skb = skb_shinfo(skb)->frag_list
list_skb->head_frag = true
skb->len = 2441 && skb->data_len = 2250
skb_shinfo(skb)->nr_frags = 17
skb_shinfo(skb)->gso_size = 75
skb_shinfo(skb)->frags[0...16].bv_len = 125
list_skb->len = 125
list_skb->data_len = 0
3962 struct sk_buff *skb_segment(struct sk_buff *head_skb,
3963 netdev_features_t features)
3964 {
3965 struct sk_buff *segs = NULL;
3966 struct sk_buff *tail = NULL;
..
4181 while (pos < offset + len) {
4182 if (i >= nfrags) {
4183 i = 0;
4184 nfrags = skb_shinfo(list_skb)->nr_frags;
4185 frag = skb_shinfo(list_skb)->frags;
4186 frag_skb = list_skb;
After segment the head_skb's last frag, the (pos == offset+len), so break the
while at line 4181, run into this BUG_ON(), not segment the head_frag frag_list
skb.
Since commit 13acc94eff122(net: permit skb_segment on head_frag frag_list skb),
it is allowed to segment the head_frag frag_list skb.
In commit 3dcbdb134f32 (net: gso: Fix skb_segment splat when splitting gso_size
mangled skb having linear-headed frag_list), it is cleared the NETIF_F_SG if it
has non head_frag skb. It is not cleared the NETIF_F_SG only with one head_frag
frag_list skb.
Signed-off-by: Fred Li <[email protected]>
---
net/core/skbuff.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 001152c..d805a47 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -4070,7 +4070,7 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb,
hsize = skb_headlen(head_skb) - offset;
- if (hsize <= 0 && i >= nfrags && skb_headlen(list_skb) &&
+ if (hsize <= 0 && i >= nfrags && !list_skb->head_frag && skb_headlen(list_skb) &&
(skb_headlen(list_skb) == len || sg)) {
BUG_ON(skb_headlen(list_skb) > len);
--
1.8.3.1
dracoding wrote:
> From: Fred Li <[email protected]>
>
> The crashed kernel version is 5.16.20, and I have not test this patch
> because I dont find a way to reproduce it, and the mailine may be
> has the same problem.
That is a pretty old kernel.
There has been work in this space in the meantime. Such as commit
3dcbdb134f32 ("net: gso: Fix skb_segment splat when splitting gso_size
mangled skb having linear-headed frag_list") or commit 9e4b7a99a03a
("net: gso: fix panic on frag_list with mixed head alloc types").
The patch was based on kernel 6.6.8, the skb properties as
mentioned in [1]. This test will cause system crash without
the patch described in [1].
[1] https://lore.kernel.org/netdev/[email protected]/
Signed-off-by: Fred Li <[email protected]>
---
lib/test_bpf.c | 64 ++++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 64 insertions(+)
diff --git a/lib/test_bpf.c b/lib/test_bpf.c
index ecde42162..a38d2d09c 100644
--- a/lib/test_bpf.c
+++ b/lib/test_bpf.c
@@ -14706,6 +14706,63 @@ static __init struct sk_buff *build_test_skb_linear_no_head_frag(void)
return NULL;
}
+static __init struct sk_buff *build_test_skb_head_frag(void)
+{
+ u32 headroom = 192, doffset = 66, alloc_size = 1536;
+ struct sk_buff *skb[2];
+ struct page *page[17];
+ int i, data_size = 125;
+ int j;
+
+ skb[0] = dev_alloc_skb(headroom + alloc_size);
+ if (!skb[0])
+ return NULL;
+
+ skb_reserve(skb[0], headroom + doffset);
+ skb_put(skb[0], data_size);
+ skb[0]->mac_header = 192;
+
+ skb[0]->protocol = htons(ETH_P_IP);
+ skb[0]->network_header = 206;
+
+ for (i = 0; i < 17; i++) {
+ page[i] = alloc_page(GFP_KERNEL);
+ if (!page[i])
+ goto err_page;
+
+ skb_add_rx_frag(skb[0], i, page[i], 0, data_size, data_size);
+ }
+
+ skb[1] = dev_alloc_skb(headroom + alloc_size);
+ if (!skb[1])
+ goto err_page;
+
+ skb_reserve(skb[1], headroom + doffset);
+ skb_put(skb[1], data_size);
+
+ /* setup shinfo */
+ skb_shinfo(skb[0])->gso_size = 75;
+ skb_shinfo(skb[0])->gso_type = SKB_GSO_TCPV4;
+ skb_shinfo(skb[0])->gso_type |= SKB_GSO_UDP_TUNNEL|SKB_GSO_TCP_FIXEDID|SKB_GSO_DODGY;
+ skb_shinfo(skb[0])->gso_segs = 0;
+ skb_shinfo(skb[0])->frag_list = skb[1];
+ skb_shinfo(skb[0])->hwtstamps.hwtstamp = 1000;
+
+ /* adjust skb[0]'s len */
+ skb[0]->len += skb[1]->len;
+ skb[0]->data_len += skb[1]->len;
+ skb[0]->truesize += skb[1]->truesize;
+
+ return skb[0];
+
+err_page:
+ kfree_skb(skb[0]);
+ for (j = 0; j < i; j++)
+ __free_page(page[j]);
+
+ return NULL;
+}
+
struct skb_segment_test {
const char *descr;
struct sk_buff *(*build_skb)(void);
@@ -14727,6 +14784,13 @@ static struct skb_segment_test skb_segment_tests[] __initconst = {
NETIF_F_LLTX | NETIF_F_GRO |
NETIF_F_IPV6_CSUM | NETIF_F_RXCSUM |
NETIF_F_HW_VLAN_STAG_TX
+ },
+ {
+ .descr = "gso_with_head_frag",
+ .build_skb = build_test_skb_head_frag,
+ .features = NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_GSO_SHIFT |
+ NETIF_F_TSO_ECN | NETIF_F_TSO_MANGLEID | NETIF_F_TSO6 |
+ NETIF_F_GSO_SCTP | NETIF_F_GSO_UDP_L4 | NETIF_F_GSO_FRAGLIST
}
};
--
2.33.0
For kernel 6.6.8, when sg is true and skb_headlen(list_skb) != len, it also has
chance run into this BUG_ON() line 4548.
'''
4544 hsize = skb_headlen(head_skb) - offset;
4545
4546 if (hsize <= 0 && i >= nfrags && skb_headlen(list_skb) &&
4547 (skb_headlen(list_skb) == len || sg)) {
4548 BUG_ON(skb_headlen(list_skb) > len);
4549
4550 nskb = skb_clone(list_skb, GFP_ATOMIC);
'''
As commit 9e4b7a99a03a("net: gso: fix panic on frag_list with mixed head alloc types")
said. It walk the frag_list in skb_segment and clear NETIF_F_SG when there is non head_frag
skb.
But for frag_list only with one head_frag, NETIF_F_SG was not cleared, if skb_headlen(list_skb) != len,
in this case, maybe we can fix it with run into segment as commit 13acc94eff122(net: permit skb_segment on
head_frag frag_list skb).
Any suggestions for resolving this issue.
Thanks
Fred Li