Re: BUG_ON triggered in skb_segment

2018-03-16 Thread Yonghong Song 



On 3/16/18 4:03 PM, Eric Dumazet wrote:



On 03/16/2018 03:37 PM, Yonghong Song wrote:


Eric and Daniel,

I have tried to fix this issue but not really successful.
I tried two hacks:
   . if skb_headlen(list_skb) is not 0, we just pull
     skb_headlen(list_skb) from the skb to make skb_headlen(list_skb) = 0, or
   . if skb_headlen(list_skb) is not 0, we go to the beginning of
     the outer loop which will allocate another nskb for this list_skb.

Both approaches removed the BUG and the packet is able to reach the
remote host. Upon receiving the packet, however, the remote host sends a
reset packet back so connection eventually closed. I did not debug
further on this.

Considering it is tricky to change skb_segment, I hacked test_bpf
kernel module to reproduce the issue. The change reflects the gso packet
structure I got from mlx5. Maybe you could take a look and suggest a fix or a 
direction of how to move forward.

Thanks!

= PATCH  ===

-bash-4.2$ git show
commit 41681ab51f85b4a0ea3416a0a62d6bde74f3af4b
Author: Yonghong Song 
Date:   Fri Mar 16 15:10:02 2018 -0700

     [hack] hack test_bpf module to trigger BUG_ON in skb_segment.

     "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:  [#1] SMP PTI
     [   98.179370] Modules linked in:
     [   98.179371]  test_bpf(+)
     ...

     The BUG happens in function skb_segment:
     ...
     3665 while (pos < offset + len) {
     3666 if (i >= nfrags) {
     3667 BUG_ON(skb_headlen(list_skb));


Note that you do not need to pull data.

Chances are high that skb->head is a page fragment, so can be considered as 
such ;)

Look at users of skb_head_is_locked()


Thanks Eric! skb->head_frag... This may well explain why my original 
hack won't work.


Yonghong





     3668
     3669 i = 0;
     3670 nfrags = 
skb_shinfo(list_skb)->nr_frags;
     3671 frag = skb_shinfo(list_skb)->frags;
     3672 frag_skb = list_skb;
     3673
     3674 BUG_ON(!nfrags);
     ...

     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.

     Signed-off-by: Yonghong Song 

diff --git a/lib/test_bpf.c b/lib/test_bpf.c
index 2efb213..d36a991 100644
--- a/lib/test_bpf.c
+++ b/lib/test_bpf.c
@@ -6574,6 +6574,67 @@ static bool exclude_test(int test_id)
     return test_id < test_range[0] || test_id > test_range[1];
  }

+static struct sk_buff *build_test_skb(void) {
+   u32 tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+   u32 headroom = NET_SKB_PAD + NET_IP_ALIGN + ETH_HLEN;
+   struct sk_buff *skb[2];
+   void *data[2], *page;
+   int i, data_size = 8;
+
+   page = (void *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
+   if (!page)
+   return NULL;
+
+   for (i = 0; i < 2; i++) {
+   data[i] = kzalloc(headroom + tailroom + data_size, GFP_KERNEL);
+   if (!data[i])
+   return NULL;
+   skb[i] = build_skb(data[i], 0);
+   if (!skb[i]) {
+   kfree(data[i]);
+   return NULL;
+   }
+   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], skb_shinfo(skb[i])->nr_frags,
+    page, 0, 64, 64);
+   }
+
+   /* 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];
+}
+
+static void test_skb_segment(void) {
+   netdev_features_t features;
+   struct sk_buff *skb;
+
+   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("Failed in test_skb_segment:build_test_skb!");
+
+   if (skb_segment(skb, features))
+   pr_info("Success in test_skb_segment!");
+   else
+   pr_info("Failed in test_skb_segment!");
+}
+
  static __init int

Re: BUG_ON triggered in skb_segment

2018-03-16 Thread Eric Dumazet 


On 03/16/2018 03:37 PM, Yonghong Song wrote:
> 
> Eric and Daniel,
> 
> I have tried to fix this issue but not really successful.
> I tried two hacks:
>   . if skb_headlen(list_skb) is not 0, we just pull
>     skb_headlen(list_skb) from the skb to make skb_headlen(list_skb) = 0, or
>   . if skb_headlen(list_skb) is not 0, we go to the beginning of
>     the outer loop which will allocate another nskb for this list_skb.
> 
> Both approaches removed the BUG and the packet is able to reach the
> remote host. Upon receiving the packet, however, the remote host sends a
> reset packet back so connection eventually closed. I did not debug
> further on this.
> 
> Considering it is tricky to change skb_segment, I hacked test_bpf
> kernel module to reproduce the issue. The change reflects the gso packet
> structure I got from mlx5. Maybe you could take a look and suggest a fix or a 
> direction of how to move forward.
> 
> Thanks!
> 
> = PATCH  ===
> 
> -bash-4.2$ git show
> commit 41681ab51f85b4a0ea3416a0a62d6bde74f3af4b
> Author: Yonghong Song 
> Date:   Fri Mar 16 15:10:02 2018 -0700
> 
>     [hack] hack test_bpf module to trigger BUG_ON in skb_segment.
> 
>     "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:  [#1] SMP PTI
>     [   98.179370] Modules linked in:
>     [   98.179371]  test_bpf(+)
>     ...
> 
>     The BUG happens in function skb_segment:
>     ...
>     3665 while (pos < offset + len) {
>     3666 if (i >= nfrags) {
>     3667 BUG_ON(skb_headlen(list_skb));

Note that you do not need to pull data.

Chances are high that skb->head is a page fragment, so can be considered as 
such ;)

Look at users of skb_head_is_locked()


>     3668
>     3669 i = 0;
>     3670 nfrags = 
> skb_shinfo(list_skb)->nr_frags;
>     3671 frag = skb_shinfo(list_skb)->frags;
>     3672 frag_skb = list_skb;
>     3673
>     3674 BUG_ON(!nfrags);
>     ...
> 
>     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.
> 
>     Signed-off-by: Yonghong Song 
> 
> diff --git a/lib/test_bpf.c b/lib/test_bpf.c
> index 2efb213..d36a991 100644
> --- a/lib/test_bpf.c
> +++ b/lib/test_bpf.c
> @@ -6574,6 +6574,67 @@ static bool exclude_test(int test_id)
>     return test_id < test_range[0] || test_id > test_range[1];
>  }
> 
> +static struct sk_buff *build_test_skb(void) {
> +   u32 tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
> +   u32 headroom = NET_SKB_PAD + NET_IP_ALIGN + ETH_HLEN;
> +   struct sk_buff *skb[2];
> +   void *data[2], *page;
> +   int i, data_size = 8;
> +
> +   page = (void *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
> +   if (!page)
> +   return NULL;
> +
> +   for (i = 0; i < 2; i++) {
> +   data[i] = kzalloc(headroom + tailroom + data_size, 
> GFP_KERNEL);
> +   if (!data[i])
> +   return NULL;
> +   skb[i] = build_skb(data[i], 0);
> +   if (!skb[i]) {
> +   kfree(data[i]);
> +   return NULL;
> +   }
> +   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], skb_shinfo(skb[i])->nr_frags,
> +    page, 0, 64, 64);
> +   }
> +
> +   /* 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];
> +}
> +
> +static void test_skb_segment(void) {
> +   netdev_features_t features;
> +   struct sk_buff *skb;
> +
> +   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("Failed in test_skb_segment:build_test_skb!");
> +
> +   if (skb_segment(skb, features))
> +   pr_info("Success in test_skb_segment!");
> +   else
> +   

Re: BUG_ON triggered in skb_segment

2018-03-16 Thread Yonghong Song 


Eric and Daniel,

I have tried to fix this issue but not really successful.
I tried two hacks:
  . if skb_headlen(list_skb) is not 0, we just pull
skb_headlen(list_skb) from the skb to make skb_headlen(list_skb) = 
0, or

  . if skb_headlen(list_skb) is not 0, we go to the beginning of
the outer loop which will allocate another nskb for this list_skb.

Both approaches removed the BUG and the packet is able to reach the
remote host. Upon receiving the packet, however, the remote host sends a
reset packet back so connection eventually closed. I did not debug
further on this.

Considering it is tricky to change skb_segment, I hacked test_bpf
kernel module to reproduce the issue. The change reflects the gso packet
structure I got from mlx5. Maybe you could take a look and suggest a fix 
or a direction of how to move forward.


Thanks!

= PATCH  ===

-bash-4.2$ git show
commit 41681ab51f85b4a0ea3416a0a62d6bde74f3af4b
Author: Yonghong Song 
Date:   Fri Mar 16 15:10:02 2018 -0700

[hack] hack test_bpf module to trigger BUG_ON in skb_segment.

"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:  [#1] SMP PTI
[   98.179370] Modules linked in:
[   98.179371]  test_bpf(+)
...

The BUG happens in function skb_segment:
...
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;
3673
3674 BUG_ON(!nfrags);
...

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.

Signed-off-by: Yonghong Song 

diff --git a/lib/test_bpf.c b/lib/test_bpf.c
index 2efb213..d36a991 100644
--- a/lib/test_bpf.c
+++ b/lib/test_bpf.c
@@ -6574,6 +6574,67 @@ static bool exclude_test(int test_id)
return test_id < test_range[0] || test_id > test_range[1];
 }

+static struct sk_buff *build_test_skb(void) {
+   u32 tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+   u32 headroom = NET_SKB_PAD + NET_IP_ALIGN + ETH_HLEN;
+   struct sk_buff *skb[2];
+   void *data[2], *page;
+   int i, data_size = 8;
+
+   page = (void *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
+   if (!page)
+   return NULL;
+
+   for (i = 0; i < 2; i++) {
+   data[i] = kzalloc(headroom + tailroom + data_size, 
GFP_KERNEL);

+   if (!data[i])
+   return NULL;
+   skb[i] = build_skb(data[i], 0);
+   if (!skb[i]) {
+   kfree(data[i]);
+   return NULL;
+   }
+   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], skb_shinfo(skb[i])->nr_frags,
+page, 0, 64, 64);
+   }
+
+   /* 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];
+}
+
+static void test_skb_segment(void) {
+   netdev_features_t features;
+   struct sk_buff *skb;
+
+   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("Failed in test_skb_segment:build_test_skb!");
+
+   if (skb_segment(skb, features))
+   pr_info("Success in test_skb_segment!");
+   else
+   pr_info("Failed in test_skb_segment!");
+}
+
 static __init int test_bpf(void)
 {
int i, err_cnt = 0, pass_cnt = 0;
@@ -6631,7 +6692,8 @@ static int __init test_bpf_init(void)
if (ret < 0)
return ret;

-   ret = test_bpf();
+   // ret = test_bpf();
+   test_skb_segment();

destroy_bpf_tests();
return ret;


=  END 


On 3/13/18 6:15 PM, Eric Dumazet wrote:



On

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Eric Dumazet 



On 03/13/2018 05:35 PM, Eric Dumazet wrote:



On 03/13/2018 05:26 PM, Eric Dumazet wrote:



On 03/13/2018 05:04 PM, Alexei Starovoitov wrote:

On 3/13/18 4:27 PM, Eric Dumazet wrote:



On 03/13/2018 04:09 PM, Alexei Starovoitov wrote:


we have bpf_skb_proto_6_to_4() that was used by cilium for long time.
It's not clear why it's not crashing there, but we cannot just
reject changing proto in bpf programs now.
We have to fix whatever needs to be fixed in skb_segment
(if bug is there) or fix whatever necessary on mlx5 side.
In bpf helper we mark it as SKB_GSO_DODGY just like packets coming
through virtio would do, so if skb_segment() needs to do something
special with skb the SKB_GSO_DODGY flag is already there.


'Fixing' skb_segment(), I did that a long time ago and Herbert Xu was
not happy with the fix and provided something else.


any link to your old patches and discussion?

I think since mlx4 can do tso on them and the packets came out
correct on the wire, there is nothing fundamentally wrong with
changing gso_size. Just tricky to teach skb_segment.



The world is not mlx4 only. Some NIC will ask skb_segment() fallback 
segmentation for various reasons (like skb->len above a given limit 
like 16KB)


Maybe https://www.spinics.net/lists/netdev/msg255549.html



Herbert patch :

commit 9d8506cc2d7ea1f911c72c100193a3677f6668c3
Author: Herbert Xu 
Date:   Thu Nov 21 11:10:04 2013 -0800

     gso: handle new frag_list of frags GRO packets



I found my initial patch.

https://www.spinics.net/lists/netdev/msg255452.html

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Eric Dumazet 



On 03/13/2018 05:26 PM, Eric Dumazet wrote:



On 03/13/2018 05:04 PM, Alexei Starovoitov wrote:

On 3/13/18 4:27 PM, Eric Dumazet wrote:



On 03/13/2018 04:09 PM, Alexei Starovoitov wrote:


we have bpf_skb_proto_6_to_4() that was used by cilium for long time.
It's not clear why it's not crashing there, but we cannot just
reject changing proto in bpf programs now.
We have to fix whatever needs to be fixed in skb_segment
(if bug is there) or fix whatever necessary on mlx5 side.
In bpf helper we mark it as SKB_GSO_DODGY just like packets coming
through virtio would do, so if skb_segment() needs to do something
special with skb the SKB_GSO_DODGY flag is already there.


'Fixing' skb_segment(), I did that a long time ago and Herbert Xu was
not happy with the fix and provided something else.


any link to your old patches and discussion?

I think since mlx4 can do tso on them and the packets came out
correct on the wire, there is nothing fundamentally wrong with
changing gso_size. Just tricky to teach skb_segment.



The world is not mlx4 only. Some NIC will ask skb_segment() fallback 
segmentation for various reasons (like skb->len above a given limit like 
16KB)


Maybe https://www.spinics.net/lists/netdev/msg255549.html



Herbert patch :

commit 9d8506cc2d7ea1f911c72c100193a3677f6668c3
Author: Herbert Xu 
Date:   Thu Nov 21 11:10:04 2013 -0800

gso: handle new frag_list of frags GRO packets

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Eric Dumazet 



On 03/13/2018 05:04 PM, Alexei Starovoitov wrote:

On 3/13/18 4:27 PM, Eric Dumazet wrote:



On 03/13/2018 04:09 PM, Alexei Starovoitov wrote:


we have bpf_skb_proto_6_to_4() that was used by cilium for long time.
It's not clear why it's not crashing there, but we cannot just
reject changing proto in bpf programs now.
We have to fix whatever needs to be fixed in skb_segment
(if bug is there) or fix whatever necessary on mlx5 side.
In bpf helper we mark it as SKB_GSO_DODGY just like packets coming
through virtio would do, so if skb_segment() needs to do something
special with skb the SKB_GSO_DODGY flag is already there.


'Fixing' skb_segment(), I did that a long time ago and Herbert Xu was
not happy with the fix and provided something else.


any link to your old patches and discussion?

I think since mlx4 can do tso on them and the packets came out
correct on the wire, there is nothing fundamentally wrong with
changing gso_size. Just tricky to teach skb_segment.



The world is not mlx4 only. Some NIC will ask skb_segment() fallback 
segmentation for various reasons (like skb->len above a given limit like 
16KB)


Maybe https://www.spinics.net/lists/netdev/msg255549.html

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Alexei Starovoitov 

On 3/13/18 4:27 PM, Eric Dumazet wrote:



On 03/13/2018 04:09 PM, Alexei Starovoitov wrote:


we have bpf_skb_proto_6_to_4() that was used by cilium for long time.
It's not clear why it's not crashing there, but we cannot just
reject changing proto in bpf programs now.
We have to fix whatever needs to be fixed in skb_segment
(if bug is there) or fix whatever necessary on mlx5 side.
In bpf helper we mark it as SKB_GSO_DODGY just like packets coming
through virtio would do, so if skb_segment() needs to do something
special with skb the SKB_GSO_DODGY flag is already there.


'Fixing' skb_segment(), I did that a long time ago and Herbert Xu was
not happy with the fix and provided something else.


any link to your old patches and discussion?

I think since mlx4 can do tso on them and the packets came out
correct on the wire, there is nothing fundamentally wrong with
changing gso_size. Just tricky to teach skb_segment.


GSO_DODGY has nothing to do with the problem really.

Changing gso_size is breaking GRO since it ends up changing the number
of segments on the wire. TCP is not going to be happy, so you'll also
have to fix TCP eventually.

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Eric Dumazet 



On 03/13/2018 04:09 PM, Alexei Starovoitov wrote:


we have bpf_skb_proto_6_to_4() that was used by cilium for long time.
It's not clear why it's not crashing there, but we cannot just
reject changing proto in bpf programs now.
We have to fix whatever needs to be fixed in skb_segment
(if bug is there) or fix whatever necessary on mlx5 side.
In bpf helper we mark it as SKB_GSO_DODGY just like packets coming
through virtio would do, so if skb_segment() needs to do something
special with skb the SKB_GSO_DODGY flag is already there.


'Fixing' skb_segment(), I did that a long time ago and Herbert Xu was 
not happy with the fix and provided something else.


GSO_DODGY has nothing to do with the problem really.

Changing gso_size is breaking GRO since it ends up changing the number 
of segments on the wire. TCP is not going to be happy, so you'll also 
have to fix TCP eventually.

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Daniel Borkmann 
On 03/14/2018 12:09 AM, Alexei Starovoitov wrote:
> On 3/13/18 3:47 PM, Eric Dumazet wrote:
>>
>>
>> On 03/13/2018 03:37 PM, Yonghong Song wrote:
>>> Adding additional cc's:
>>>    Saeed Mahameed as this is most likely mlx5 driver related.
>>>    Diptanu Gon Choudhury who initially reported the issue.
>>>
>>>
>>> On 3/13/18 1:44 AM, Steffen Klassert wrote:
 On Mon, Mar 12, 2018 at 11:25:09PM -0700, Eric Dumazet wrote:
>
>
> On 03/12/2018 11:08 PM, Yonghong Song wrote:
>>
>>
>> On 3/12/18 11:04 PM, Eric Dumazet wrote:
>>>
>>>
>>> On 03/12/2018 10:45 PM, Yonghong Song wrote:
 ...
 Setup:
 =

 The test will involve three machines:
     M_ipv6 <-> M_nat <-> M_ipv4

 The M_nat will do ipv4<->ipv6 address translation and then
 forward packet
 to proper destination. The control plane will configure M_nat
 properly
 will understand virtual ipv4 address for machine M_ipv6, and
 virtual ipv6 address for machine M_ipv4.

 M_nat runs a bpf program, which is attached to clsact (ingress)
 qdisc.
 The program uses bpf_skb_change_proto to do protocol conversion.
 bpf_skb_change_proto will adjust skb header_len and len properly
 based on protocol change.
 After the conversion, the program will make proper change on
 ethhdr and ip4/6 header, recalculate checksum, and send the
 packet out
 through bpf_redirect.

 Experiment:
 ===

 MTU: 1500B for all three machines.

 The tso/lro/gro are enabled on the M_nat box.

 ping works on both ways of M_ipv6 <-> M_ipv4.
 It works for transfering a small file (4KB) between M_ipv6 and
 M_ipv4 (both ways).
 Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4,
 failed with the above BUG_ON, really fast.
 Did not really test from M_ipv4 to M_ipv6 with large file.

 The error path likely to be (also from the above call stack):
     nic -> lro/gro -> bpf_program -> gso (BUG_ON)

 Just out of curiosity, are these packets created with LRO or GRO?
 Usually LRO is disabled if forwarding is enabled on a machine,
 because segmented LGO packets are likely corrupt.
>>>
>>> In our experiments, LRO is disabled.
>>> On mlx5, when GRO is on, the BUG_ON will happen, and
>>> when GRO is off, the BUG_ON will not happen.
>>>

 These packets take an alternative redirect path, so not sure what
 happens here.


 In one of experiments, I explicitly printed the skb->len and
 skb->data_len. The values are below:
     skb_segment: len 2856, data_len 2686
 They should be equal to avoid BUG.

 In another experiment, I got:
     skb_segment: len 1428, data_len 1258

 In both cases, the difference is 170 bytes. Not sure whether
 this is just a coincidence or not.

 Workaround:
 ===

 A workaround to avoid BUG_ON is to disable lro/gro. This way,
 kernel will not receive big packets and hence gso is not really
 called.

 I am not familiar with gso code. Does anybody hit this BUG_ON
 before?
 Any suggestion on how to debug this?

>>>
>>> skb_segment() works if incoming GRO packet is not modified in its
>>> geometry.
>>>
>>> In your case it seems you had to adjust gso_size (calling
>>> skb_decrease_gso_size() or skb_increase_gso_size()), and this breaks
>>> skb_segment() badly, because geometry changes, unless you had
>>> specific MTU/MSS restrictions.
>>>
>>> You will have to make skb_segment() more generic if you really
>>> want this.
>>
>> In net/core/filter.c function bpf_skb_change_proto, which is called
>> in the bpf program, does some GSO adjustment. Could you help check
>> whether it satisfies my above use case or not? Thanks!
>
> As I said this  helper ends up modifying gso_size by +/- 20
> (sizeof(ipv6
> header) - sizeof(ipv4 header))
>
> So it wont work if skb_segment() is called after this change.

 Even HW TSO use gso_size to segment the packets. Would'nt this
 result in broken packets too, if gso_size is modified on a
 forwarding path?

>
> Not clear why the GRO packet is not sent as is (as a TSO packet) since
> mlx4/mlx5 NICs certainly support TSO.
>>>
>>> This is a very good observation.
>>> We did the same experiment on mlx4, the same kernel, the same
>>> userspace apps, the same bpf program. The only difference is mlx4 vs.
>>> mlx5.
>>> The mlx4 works fine with LRO=off and GRO=on, while
>>> mlx5 failed with the same LRO/GRO configuration.
>>>
>>> On mlx4 box, we are able to see TSO packets are

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Alexei Starovoitov 

On 3/13/18 3:47 PM, Eric Dumazet wrote:



On 03/13/2018 03:37 PM, Yonghong Song wrote:

Adding additional cc's:
   Saeed Mahameed as this is most likely mlx5 driver related.
   Diptanu Gon Choudhury who initially reported the issue.


On 3/13/18 1:44 AM, Steffen Klassert wrote:

On Mon, Mar 12, 2018 at 11:25:09PM -0700, Eric Dumazet wrote:



On 03/12/2018 11:08 PM, Yonghong Song wrote:



On 3/12/18 11:04 PM, Eric Dumazet wrote:



On 03/12/2018 10:45 PM, Yonghong Song wrote:

...
Setup:
=

The test will involve three machines:
M_ipv6 <-> M_nat <-> M_ipv4

The M_nat will do ipv4<->ipv6 address translation and then
forward packet
to proper destination. The control plane will configure M_nat
properly
will understand virtual ipv4 address for machine M_ipv6, and
virtual ipv6 address for machine M_ipv4.

M_nat runs a bpf program, which is attached to clsact (ingress)
qdisc.
The program uses bpf_skb_change_proto to do protocol conversion.
bpf_skb_change_proto will adjust skb header_len and len properly
based on protocol change.
After the conversion, the program will make proper change on
ethhdr and ip4/6 header, recalculate checksum, and send the
packet out
through bpf_redirect.

Experiment:
===

MTU: 1500B for all three machines.

The tso/lro/gro are enabled on the M_nat box.

ping works on both ways of M_ipv6 <-> M_ipv4.
It works for transfering a small file (4KB) between M_ipv6 and
M_ipv4 (both ways).
Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4,
failed with the above BUG_ON, really fast.
Did not really test from M_ipv4 to M_ipv6 with large file.

The error path likely to be (also from the above call stack):
nic -> lro/gro -> bpf_program -> gso (BUG_ON)


Just out of curiosity, are these packets created with LRO or GRO?
Usually LRO is disabled if forwarding is enabled on a machine,
because segmented LGO packets are likely corrupt.


In our experiments, LRO is disabled.
On mlx5, when GRO is on, the BUG_ON will happen, and
when GRO is off, the BUG_ON will not happen.



These packets take an alternative redirect path, so not sure what
happens here.



In one of experiments, I explicitly printed the skb->len and
skb->data_len. The values are below:
skb_segment: len 2856, data_len 2686
They should be equal to avoid BUG.

In another experiment, I got:
skb_segment: len 1428, data_len 1258

In both cases, the difference is 170 bytes. Not sure whether
this is just a coincidence or not.

Workaround:
===

A workaround to avoid BUG_ON is to disable lro/gro. This way,
kernel will not receive big packets and hence gso is not really
called.

I am not familiar with gso code. Does anybody hit this BUG_ON
before?
Any suggestion on how to debug this?



skb_segment() works if incoming GRO packet is not modified in its
geometry.

In your case it seems you had to adjust gso_size (calling
skb_decrease_gso_size() or skb_increase_gso_size()), and this breaks
skb_segment() badly, because geometry changes, unless you had
specific MTU/MSS restrictions.

You will have to make skb_segment() more generic if you really
want this.


In net/core/filter.c function bpf_skb_change_proto, which is called
in the bpf program, does some GSO adjustment. Could you help check
whether it satisfies my above use case or not? Thanks!


As I said this  helper ends up modifying gso_size by +/- 20
(sizeof(ipv6
header) - sizeof(ipv4 header))

So it wont work if skb_segment() is called after this change.


Even HW TSO use gso_size to segment the packets. Would'nt this
result in broken packets too, if gso_size is modified on a
forwarding path?



Not clear why the GRO packet is not sent as is (as a TSO packet) since
mlx4/mlx5 NICs certainly support TSO.


This is a very good observation.
We did the same experiment on mlx4, the same kernel, the same
userspace apps, the same bpf program. The only difference is mlx4 vs.
mlx5.
The mlx4 works fine with LRO=off and GRO=on, while
mlx5 failed with the same LRO/GRO configuration.

On mlx4 box, we are able to see TSO packets are increasing as the
large file scp is in progress.
# ethtool -S eth0 | grep tso
  tso_packets: 45495
# ethtool -S eth0 | grep tso
  tso_packets: 45865
# ethtool -S eth0 | grep tso
  tso_packets: 46337
# ethtool -S eth0 | grep tso
  tso_packets: 46724

And use bcc tool to track to func call count for skb_segment
and find it is called 0 times. Clearly, mlx4 is able to take
the packet as TSO and hence the packet will not go to
the stack.

# funccount.py -i 3 'skb_segment'
Tracing 1 functions for "skb_segment"... Hit Ctrl-C to end.

FUNCCOUNT

FUNCCOUNT
...

CC Saeed Mahameed who may help debug and provide some insights
what is the problem.


There are many reasons why a GRO packet might need to be segmented by
software (skb_segment())

This is the step where you have crashes, so really mlx4/mlx5 difference
do not matter.

gso_size should not be dynamically

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Eric Dumazet 



On 03/13/2018 03:37 PM, Yonghong Song wrote:

Adding additional cc's:
   Saeed Mahameed as this is most likely mlx5 driver related.
   Diptanu Gon Choudhury who initially reported the issue.


On 3/13/18 1:44 AM, Steffen Klassert wrote:

On Mon, Mar 12, 2018 at 11:25:09PM -0700, Eric Dumazet wrote:



On 03/12/2018 11:08 PM, Yonghong Song wrote:



On 3/12/18 11:04 PM, Eric Dumazet wrote:



On 03/12/2018 10:45 PM, Yonghong Song wrote:

...
Setup:
=

The test will involve three machines:
    M_ipv6 <-> M_nat <-> M_ipv4

The M_nat will do ipv4<->ipv6 address translation and then
forward packet
to proper destination. The control plane will configure M_nat 
properly

will understand virtual ipv4 address for machine M_ipv6, and
virtual ipv6 address for machine M_ipv4.

M_nat runs a bpf program, which is attached to clsact (ingress) 
qdisc.

The program uses bpf_skb_change_proto to do protocol conversion.
bpf_skb_change_proto will adjust skb header_len and len properly
based on protocol change.
After the conversion, the program will make proper change on
ethhdr and ip4/6 header, recalculate checksum, and send the packet 
out

through bpf_redirect.

Experiment:
===

MTU: 1500B for all three machines.

The tso/lro/gro are enabled on the M_nat box.

ping works on both ways of M_ipv6 <-> M_ipv4.
It works for transfering a small file (4KB) between M_ipv6 and
M_ipv4 (both ways).
Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4,
failed with the above BUG_ON, really fast.
Did not really test from M_ipv4 to M_ipv6 with large file.

The error path likely to be (also from the above call stack):
    nic -> lro/gro -> bpf_program -> gso (BUG_ON)


Just out of curiosity, are these packets created with LRO or GRO?
Usually LRO is disabled if forwarding is enabled on a machine,
because segmented LGO packets are likely corrupt.


In our experiments, LRO is disabled.
On mlx5, when GRO is on, the BUG_ON will happen, and
when GRO is off, the BUG_ON will not happen.



These packets take an alternative redirect path, so not sure what
happens here.



In one of experiments, I explicitly printed the skb->len and
skb->data_len. The values are below:
    skb_segment: len 2856, data_len 2686
They should be equal to avoid BUG.

In another experiment, I got:
    skb_segment: len 1428, data_len 1258

In both cases, the difference is 170 bytes. Not sure whether
this is just a coincidence or not.

Workaround:
===

A workaround to avoid BUG_ON is to disable lro/gro. This way,
kernel will not receive big packets and hence gso is not really 
called.


I am not familiar with gso code. Does anybody hit this BUG_ON before?
Any suggestion on how to debug this?



skb_segment() works if incoming GRO packet is not modified in its
geometry.

In your case it seems you had to adjust gso_size (calling
skb_decrease_gso_size() or skb_increase_gso_size()), and this breaks
skb_segment() badly, because geometry changes, unless you had
specific MTU/MSS restrictions.

You will have to make skb_segment() more generic if you really want 
this.


In net/core/filter.c function bpf_skb_change_proto, which is called
in the bpf program, does some GSO adjustment. Could you help check
whether it satisfies my above use case or not? Thanks!


As I said this  helper ends up modifying gso_size by +/- 20 (sizeof(ipv6
header) - sizeof(ipv4 header))

So it wont work if skb_segment() is called after this change.


Even HW TSO use gso_size to segment the packets. Would'nt this
result in broken packets too, if gso_size is modified on a
forwarding path?



Not clear why the GRO packet is not sent as is (as a TSO packet) since
mlx4/mlx5 NICs certainly support TSO.


This is a very good observation.
We did the same experiment on mlx4, the same kernel, the same userspace 
apps, the same bpf program. The only difference is mlx4 vs. mlx5.

The mlx4 works fine with LRO=off and GRO=on, while
mlx5 failed with the same LRO/GRO configuration.

On mlx4 box, we are able to see TSO packets are increasing as the
large file scp is in progress.
# ethtool -S eth0 | grep tso
  tso_packets: 45495
# ethtool -S eth0 | grep tso
  tso_packets: 45865
# ethtool -S eth0 | grep tso
  tso_packets: 46337
# ethtool -S eth0 | grep tso
  tso_packets: 46724

And use bcc tool to track to func call count for skb_segment
and find it is called 0 times. Clearly, mlx4 is able to take
the packet as TSO and hence the packet will not go to
the stack.

# funccount.py -i 3 'skb_segment'
Tracing 1 functions for "skb_segment"... Hit Ctrl-C to end.

FUNC    COUNT

FUNC    COUNT
...

CC Saeed Mahameed who may help debug and provide some insights
what is the problem.


There are many reasons why a GRO packet might need to be segmented by 
software (skb_segment())


This is the step where you have crashes, so really mlx4/mlx5 difference 
do not matter.


gso_size should not be dynamically changed. This needs to be

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Yonghong Song 

Adding additional cc's:
  Saeed Mahameed as this is most likely mlx5 driver related.
  Diptanu Gon Choudhury who initially reported the issue.


On 3/13/18 1:44 AM, Steffen Klassert wrote:

On Mon, Mar 12, 2018 at 11:25:09PM -0700, Eric Dumazet wrote:



On 03/12/2018 11:08 PM, Yonghong Song wrote:



On 3/12/18 11:04 PM, Eric Dumazet wrote:



On 03/12/2018 10:45 PM, Yonghong Song wrote:

...
Setup:
=

The test will involve three machines:
    M_ipv6 <-> M_nat <-> M_ipv4

The M_nat will do ipv4<->ipv6 address translation and then
forward packet
to proper destination. The control plane will configure M_nat properly
will understand virtual ipv4 address for machine M_ipv6, and
virtual ipv6 address for machine M_ipv4.

M_nat runs a bpf program, which is attached to clsact (ingress) qdisc.
The program uses bpf_skb_change_proto to do protocol conversion.
bpf_skb_change_proto will adjust skb header_len and len properly
based on protocol change.
After the conversion, the program will make proper change on
ethhdr and ip4/6 header, recalculate checksum, and send the packet out
through bpf_redirect.

Experiment:
===

MTU: 1500B for all three machines.

The tso/lro/gro are enabled on the M_nat box.

ping works on both ways of M_ipv6 <-> M_ipv4.
It works for transfering a small file (4KB) between M_ipv6 and
M_ipv4 (both ways).
Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4,
failed with the above BUG_ON, really fast.
Did not really test from M_ipv4 to M_ipv6 with large file.

The error path likely to be (also from the above call stack):
    nic -> lro/gro -> bpf_program -> gso (BUG_ON)


Just out of curiosity, are these packets created with LRO or GRO?
Usually LRO is disabled if forwarding is enabled on a machine,
because segmented LGO packets are likely corrupt.


In our experiments, LRO is disabled.
On mlx5, when GRO is on, the BUG_ON will happen, and
when GRO is off, the BUG_ON will not happen.



These packets take an alternative redirect path, so not sure what
happens here.



In one of experiments, I explicitly printed the skb->len and
skb->data_len. The values are below:
    skb_segment: len 2856, data_len 2686
They should be equal to avoid BUG.

In another experiment, I got:
    skb_segment: len 1428, data_len 1258

In both cases, the difference is 170 bytes. Not sure whether
this is just a coincidence or not.

Workaround:
===

A workaround to avoid BUG_ON is to disable lro/gro. This way,
kernel will not receive big packets and hence gso is not really called.

I am not familiar with gso code. Does anybody hit this BUG_ON before?
Any suggestion on how to debug this?



skb_segment() works if incoming GRO packet is not modified in its
geometry.

In your case it seems you had to adjust gso_size (calling
skb_decrease_gso_size() or skb_increase_gso_size()), and this breaks
skb_segment() badly, because geometry changes, unless you had
specific MTU/MSS restrictions.

You will have to make skb_segment() more generic if you really want this.


In net/core/filter.c function bpf_skb_change_proto, which is called
in the bpf program, does some GSO adjustment. Could you help check
whether it satisfies my above use case or not? Thanks!


As I said this  helper ends up modifying gso_size by +/- 20 (sizeof(ipv6
header) - sizeof(ipv4 header))

So it wont work if skb_segment() is called after this change.


Even HW TSO use gso_size to segment the packets. Would'nt this
result in broken packets too, if gso_size is modified on a
forwarding path?



Not clear why the GRO packet is not sent as is (as a TSO packet) since
mlx4/mlx5 NICs certainly support TSO.


This is a very good observation.
We did the same experiment on mlx4, the same kernel, the same userspace 
apps, the same bpf program. The only difference is mlx4 vs. mlx5.

The mlx4 works fine with LRO=off and GRO=on, while
mlx5 failed with the same LRO/GRO configuration.

On mlx4 box, we are able to see TSO packets are increasing as the
large file scp is in progress.
# ethtool -S eth0 | grep tso
 tso_packets: 45495
# ethtool -S eth0 | grep tso
 tso_packets: 45865
# ethtool -S eth0 | grep tso
 tso_packets: 46337
# ethtool -S eth0 | grep tso
 tso_packets: 46724

And use bcc tool to track to func call count for skb_segment
and find it is called 0 times. Clearly, mlx4 is able to take
the packet as TSO and hence the packet will not go to
the stack.

# funccount.py -i 3 'skb_segment'
Tracing 1 functions for "skb_segment"... Hit Ctrl-C to end.

FUNCCOUNT

FUNCCOUNT
...

CC Saeed Mahameed who may help debug and provide some insights
what is the problem.



If the packets are generated with GRO, there could be data chained
at the frag_list pointer. Most NICs can't offload such skbs, so if
skb_segment() can't split at the frag_list pointer, it will just
segment the packets based on gso_size.

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Steffen Klassert 
On Mon, Mar 12, 2018 at 11:25:09PM -0700, Eric Dumazet wrote:
> 
> 
> On 03/12/2018 11:08 PM, Yonghong Song wrote:
> > 
> > 
> > On 3/12/18 11:04 PM, Eric Dumazet wrote:
> > > 
> > > 
> > > On 03/12/2018 10:45 PM, Yonghong Song wrote:
> > > > ...
> > > > Setup:
> > > > =
> > > > 
> > > > The test will involve three machines:
> > > >    M_ipv6 <-> M_nat <-> M_ipv4
> > > > 
> > > > The M_nat will do ipv4<->ipv6 address translation and then
> > > > forward packet
> > > > to proper destination. The control plane will configure M_nat properly
> > > > will understand virtual ipv4 address for machine M_ipv6, and
> > > > virtual ipv6 address for machine M_ipv4.
> > > > 
> > > > M_nat runs a bpf program, which is attached to clsact (ingress) qdisc.
> > > > The program uses bpf_skb_change_proto to do protocol conversion.
> > > > bpf_skb_change_proto will adjust skb header_len and len properly
> > > > based on protocol change.
> > > > After the conversion, the program will make proper change on
> > > > ethhdr and ip4/6 header, recalculate checksum, and send the packet out
> > > > through bpf_redirect.
> > > > 
> > > > Experiment:
> > > > ===
> > > > 
> > > > MTU: 1500B for all three machines.
> > > > 
> > > > The tso/lro/gro are enabled on the M_nat box.
> > > > 
> > > > ping works on both ways of M_ipv6 <-> M_ipv4.
> > > > It works for transfering a small file (4KB) between M_ipv6 and
> > > > M_ipv4 (both ways).
> > > > Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4,
> > > > failed with the above BUG_ON, really fast.
> > > > Did not really test from M_ipv4 to M_ipv6 with large file.
> > > > 
> > > > The error path likely to be (also from the above call stack):
> > > >    nic -> lro/gro -> bpf_program -> gso (BUG_ON)

Just out of curiosity, are these packets created with LRO or GRO?
Usually LRO is disabled if forwarding is enabled on a machine,
because segmented LGO packets are likely corrupt.

These packets take an alternative redirect path, so not sure what
happens here.

> > > > 
> > > > In one of experiments, I explicitly printed the skb->len and
> > > > skb->data_len. The values are below:
> > > >    skb_segment: len 2856, data_len 2686
> > > > They should be equal to avoid BUG.
> > > > 
> > > > In another experiment, I got:
> > > >    skb_segment: len 1428, data_len 1258
> > > > 
> > > > In both cases, the difference is 170 bytes. Not sure whether
> > > > this is just a coincidence or not.
> > > > 
> > > > Workaround:
> > > > ===
> > > > 
> > > > A workaround to avoid BUG_ON is to disable lro/gro. This way,
> > > > kernel will not receive big packets and hence gso is not really called.
> > > > 
> > > > I am not familiar with gso code. Does anybody hit this BUG_ON before?
> > > > Any suggestion on how to debug this?
> > > > 
> > > 
> > > skb_segment() works if incoming GRO packet is not modified in its
> > > geometry.
> > > 
> > > In your case it seems you had to adjust gso_size (calling
> > > skb_decrease_gso_size() or skb_increase_gso_size()), and this breaks
> > > skb_segment() badly, because geometry changes, unless you had
> > > specific MTU/MSS restrictions.
> > > 
> > > You will have to make skb_segment() more generic if you really want this.
> > 
> > In net/core/filter.c function bpf_skb_change_proto, which is called
> > in the bpf program, does some GSO adjustment. Could you help check
> > whether it satisfies my above use case or not? Thanks!
> 
> As I said this  helper ends up modifying gso_size by +/- 20 (sizeof(ipv6
> header) - sizeof(ipv4 header))
> 
> So it wont work if skb_segment() is called after this change.

Even HW TSO use gso_size to segment the packets. Would'nt this
result in broken packets too, if gso_size is modified on a
forwarding path?

> 
> Not clear why the GRO packet is not sent as is (as a TSO packet) since
> mlx4/mlx5 NICs certainly support TSO.

If the packets are generated with GRO, there could be data chained
at the frag_list pointer. Most NICs can't offload such skbs, so if
skb_segment() can't split at the frag_list pointer, it will just
segment the packets based on gso_size.

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Eric Dumazet 



On 03/12/2018 11:08 PM, Yonghong Song wrote:



On 3/12/18 11:04 PM, Eric Dumazet wrote:



On 03/12/2018 10:45 PM, Yonghong Song wrote:

Hi,

One of our in-house projects, bpf-based NAT, hits a kernel BUG_ON at
net-next function skb_segment, line 3667.

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:
...
#0 [883ffef034f8] machine_kexec at 81044c41
  #1 [883ffef03558] __crash_kexec at 8110c525
  #2 [883ffef03620] crash_kexec at 8110d5cc
  #3 [883ffef03640] oops_end at 8101d7e7
  #4 [883ffef03668] die at 8101deb2
  #5 [883ffef03698] do_trap at 8101a700
  #6 [883ffef036e8] do_error_trap at 8101abfe
  #7 [883ffef037a0] do_invalid_op at 8101acd0
  #8 [883ffef037b0] invalid_op at 81a00bab
 [exception RIP: skb_segment+3044]
 RIP: 817e4dd4  RSP: 883ffef03860  RFLAGS: 00010216
 RAX: 2bf6  RBX: 883feb7aaa00  RCX: 0011
 RDX: 883fb87910c0  RSI: 0011  RDI: 883feb7ab500
 RBP: 883ffef03928   R8: 2ce2   R9: 27da
 R10: 01ea  R11: 2d82  R12: 883f90a1ee80
 R13: 883fb8791120  R14: 883feb7abc00  R15: 2ce2
 ORIG_RAX:   CS: 0010  SS: 0018
  #9 [883ffef03930] tcp_gso_segment at 818713e7
#10 [883ffef03990] tcp4_gso_segment at 818717d8
#11 [883ffef039b0] inet_gso_segment at 81882c9b
#12 [883ffef03a10] skb_mac_gso_segment at 817f39b8
#13 [883ffef03a38] __skb_gso_segment at 817f3ac9
#14 [883ffef03a68] validate_xmit_skb at 817f3eed
#15 [883ffef03aa8] validate_xmit_skb_list at 817f40a2
#16 [883ffef03ad8] sch_direct_xmit at 81824efb
#17 [883ffef03b20] __qdisc_run at 818251aa
#18 [883ffef03b90] __dev_queue_xmit at 817f45ed
#19 [883ffef03c08] dev_queue_xmit at 817f4b90
#20 [883ffef03c18] __bpf_redirect at 81812b66
#21 [883ffef03c40] skb_do_redirect at 81813209
#22 [883ffef03c60] __netif_receive_skb_core at 817f310d
#23 [883ffef03cc8] __netif_receive_skb at 817f32e8
#24 [883ffef03ce8] netif_receive_skb_internal at 817f5538
#25 [883ffef03d10] napi_gro_complete at 817f56c0
#26 [883ffef03d28] dev_gro_receive at 817f5ea6
#27 [883ffef03d78] napi_gro_receive at 817f6168
#28 [883ffef03da0] mlx5e_handle_rx_cqe_mpwrq at 817381c2
#29 [883ffef03e30] mlx5e_poll_rx_cq at 817386c2
#30 [883ffef03e80] mlx5e_napi_poll at 8173926e
#31 [883ffef03ed0] net_rx_action at 817f5a6e
#32 [883ffef03f48] __softirqentry_text_start at 81c000c3
#33 [883ffef03fa8] irq_exit at 8108f515
#34 [883ffef03fb8] do_IRQ at 81a01b11
---  ---
bt: cannot transition from IRQ stack to current process stack:
 IRQ stack pointer: 883ffef034f8
 process stack pointer: 81a01ae9
    current stack base: c9000c5c4000
...
Setup:
=

The test will involve three machines:
   M_ipv6 <-> M_nat <-> M_ipv4

The M_nat will do ipv4<->ipv6 address translation and then forward 
packet

to proper destination. The control plane will configure M_nat properly
will understand virtual ipv4 address for machine M_ipv6, and
virtual ipv6 address for machine M_ipv4.

M_nat runs a bpf program, which is attached to clsact (ingress) qdisc.
The program uses bpf_skb_change_proto to do protocol conversion.
bpf_skb_change_proto will adjust skb header_len and len properly
based on protocol change.
After the conversion, the program will make proper change on
ethhdr and ip4/6 header, recalculate checksum, and send the packet out
through bpf_redirect.

Experiment:
===

MTU: 1500B for all three machines.

The tso/lro/gro are enabled on the M_nat box.

ping works on both ways of M_ipv6 <-> M_ipv4.
It works for transfering a small file (4KB) between M_ipv6 and M_ipv4 
(both ways).
Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4, failed 
with the above BUG_ON, really fast.

Did not really test from M_ipv4 to M_ipv6 with large file.

The error path likely to be (also from the above call stack):
   nic -> lro/gro -> bpf_program -> gso (BUG_ON)

In one of experiments, I explicitly

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Yunsheng Lin 
Hi, Song

On 2018/3/13 13:45, Yonghong Song wrote:
> Hi,
> 
> One of our in-house projects, bpf-based NAT, hits a kernel BUG_ON at
> net-next function skb_segment, line 3667.
> 
> 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:
> ...
> #0 [883ffef034f8] machine_kexec at 81044c41
>  #1 [883ffef03558] __crash_kexec at 8110c525
>  #2 [883ffef03620] crash_kexec at 8110d5cc
>  #3 [883ffef03640] oops_end at 8101d7e7
>  #4 [883ffef03668] die at 8101deb2
>  #5 [883ffef03698] do_trap at 8101a700
>  #6 [883ffef036e8] do_error_trap at 8101abfe
>  #7 [883ffef037a0] do_invalid_op at 8101acd0
>  #8 [883ffef037b0] invalid_op at 81a00bab
> [exception RIP: skb_segment+3044]
> RIP: 817e4dd4  RSP: 883ffef03860  RFLAGS: 00010216
> RAX: 2bf6  RBX: 883feb7aaa00  RCX: 0011
> RDX: 883fb87910c0  RSI: 0011  RDI: 883feb7ab500
> RBP: 883ffef03928   R8: 2ce2   R9: 27da
> R10: 01ea  R11: 2d82  R12: 883f90a1ee80
> R13: 883fb8791120  R14: 883feb7abc00  R15: 2ce2
> ORIG_RAX:   CS: 0010  SS: 0018
>  #9 [883ffef03930] tcp_gso_segment at 818713e7
> #10 [883ffef03990] tcp4_gso_segment at 818717d8
> #11 [883ffef039b0] inet_gso_segment at 81882c9b
> #12 [883ffef03a10] skb_mac_gso_segment at 817f39b8
> #13 [883ffef03a38] __skb_gso_segment at 817f3ac9
> #14 [883ffef03a68] validate_xmit_skb at 817f3eed
> #15 [883ffef03aa8] validate_xmit_skb_list at 817f40a2
> #16 [883ffef03ad8] sch_direct_xmit at 81824efb
> #17 [883ffef03b20] __qdisc_run at 818251aa
> #18 [883ffef03b90] __dev_queue_xmit at 817f45ed
> #19 [883ffef03c08] dev_queue_xmit at 817f4b90
> #20 [883ffef03c18] __bpf_redirect at 81812b66
> #21 [883ffef03c40] skb_do_redirect at 81813209
> #22 [883ffef03c60] __netif_receive_skb_core at 817f310d
> #23 [883ffef03cc8] __netif_receive_skb at 817f32e8
> #24 [883ffef03ce8] netif_receive_skb_internal at 817f5538
> #25 [883ffef03d10] napi_gro_complete at 817f56c0
> #26 [883ffef03d28] dev_gro_receive at 817f5ea6
> #27 [883ffef03d78] napi_gro_receive at 817f6168
> #28 [883ffef03da0] mlx5e_handle_rx_cqe_mpwrq at 817381c2
> #29 [883ffef03e30] mlx5e_poll_rx_cq at 817386c2
> #30 [883ffef03e80] mlx5e_napi_poll at 8173926e
> #31 [883ffef03ed0] net_rx_action at 817f5a6e
> #32 [883ffef03f48] __softirqentry_text_start at 81c000c3
> #33 [883ffef03fa8] irq_exit at 8108f515
> #34 [883ffef03fb8] do_IRQ at 81a01b11
> ---  ---
> bt: cannot transition from IRQ stack to current process stack:
> IRQ stack pointer: 883ffef034f8
> process stack pointer: 81a01ae9
>current stack base: c9000c5c4000
> ...
> Setup:
> =
> 
> The test will involve three machines:
>   M_ipv6 <-> M_nat <-> M_ipv4
> 
> The M_nat will do ipv4<->ipv6 address translation and then forward packet
> to proper destination. The control plane will configure M_nat properly
> will understand virtual ipv4 address for machine M_ipv6, and
> virtual ipv6 address for machine M_ipv4.
> 
> M_nat runs a bpf program, which is attached to clsact (ingress) qdisc.
> The program uses bpf_skb_change_proto to do protocol conversion.
> bpf_skb_change_proto will adjust skb header_len and len properly
> based on protocol change.
> After the conversion, the program will make proper change on
> ethhdr and ip4/6 header, recalculate checksum, and send the packet out
> through bpf_redirect.
> 
> Experiment:
> ===
> 
> MTU: 1500B for all three machines.
> 
> The tso/lro/gro are enabled on the M_nat box.
> 
> ping works on both ways of M_ipv6 <-> M_ipv4.
> It works for transfering a small file (4KB) between M_ipv6 and M_ipv4 (both 
> ways).
> Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4, failed with the 
> above BUG_ON, really fast.
> Did not really test from M_ipv4 to M_ipv6 with large file.
> 
> The error path likely to be (also from the above call

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Yonghong Song 



On 3/12/18 11:04 PM, Eric Dumazet wrote:



On 03/12/2018 10:45 PM, Yonghong Song wrote:

Hi,

One of our in-house projects, bpf-based NAT, hits a kernel BUG_ON at
net-next function skb_segment, line 3667.

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:
...
#0 [883ffef034f8] machine_kexec at 81044c41
  #1 [883ffef03558] __crash_kexec at 8110c525
  #2 [883ffef03620] crash_kexec at 8110d5cc
  #3 [883ffef03640] oops_end at 8101d7e7
  #4 [883ffef03668] die at 8101deb2
  #5 [883ffef03698] do_trap at 8101a700
  #6 [883ffef036e8] do_error_trap at 8101abfe
  #7 [883ffef037a0] do_invalid_op at 8101acd0
  #8 [883ffef037b0] invalid_op at 81a00bab
 [exception RIP: skb_segment+3044]
 RIP: 817e4dd4  RSP: 883ffef03860  RFLAGS: 00010216
 RAX: 2bf6  RBX: 883feb7aaa00  RCX: 0011
 RDX: 883fb87910c0  RSI: 0011  RDI: 883feb7ab500
 RBP: 883ffef03928   R8: 2ce2   R9: 27da
 R10: 01ea  R11: 2d82  R12: 883f90a1ee80
 R13: 883fb8791120  R14: 883feb7abc00  R15: 2ce2
 ORIG_RAX:   CS: 0010  SS: 0018
  #9 [883ffef03930] tcp_gso_segment at 818713e7
#10 [883ffef03990] tcp4_gso_segment at 818717d8
#11 [883ffef039b0] inet_gso_segment at 81882c9b
#12 [883ffef03a10] skb_mac_gso_segment at 817f39b8
#13 [883ffef03a38] __skb_gso_segment at 817f3ac9
#14 [883ffef03a68] validate_xmit_skb at 817f3eed
#15 [883ffef03aa8] validate_xmit_skb_list at 817f40a2
#16 [883ffef03ad8] sch_direct_xmit at 81824efb
#17 [883ffef03b20] __qdisc_run at 818251aa
#18 [883ffef03b90] __dev_queue_xmit at 817f45ed
#19 [883ffef03c08] dev_queue_xmit at 817f4b90
#20 [883ffef03c18] __bpf_redirect at 81812b66
#21 [883ffef03c40] skb_do_redirect at 81813209
#22 [883ffef03c60] __netif_receive_skb_core at 817f310d
#23 [883ffef03cc8] __netif_receive_skb at 817f32e8
#24 [883ffef03ce8] netif_receive_skb_internal at 817f5538
#25 [883ffef03d10] napi_gro_complete at 817f56c0
#26 [883ffef03d28] dev_gro_receive at 817f5ea6
#27 [883ffef03d78] napi_gro_receive at 817f6168
#28 [883ffef03da0] mlx5e_handle_rx_cqe_mpwrq at 817381c2
#29 [883ffef03e30] mlx5e_poll_rx_cq at 817386c2
#30 [883ffef03e80] mlx5e_napi_poll at 8173926e
#31 [883ffef03ed0] net_rx_action at 817f5a6e
#32 [883ffef03f48] __softirqentry_text_start at 81c000c3
#33 [883ffef03fa8] irq_exit at 8108f515
#34 [883ffef03fb8] do_IRQ at 81a01b11
---  ---
bt: cannot transition from IRQ stack to current process stack:
 IRQ stack pointer: 883ffef034f8
 process stack pointer: 81a01ae9
    current stack base: c9000c5c4000
...
Setup:
=

The test will involve three machines:
   M_ipv6 <-> M_nat <-> M_ipv4

The M_nat will do ipv4<->ipv6 address translation and then forward packet
to proper destination. The control plane will configure M_nat properly
will understand virtual ipv4 address for machine M_ipv6, and
virtual ipv6 address for machine M_ipv4.

M_nat runs a bpf program, which is attached to clsact (ingress) qdisc.
The program uses bpf_skb_change_proto to do protocol conversion.
bpf_skb_change_proto will adjust skb header_len and len properly
based on protocol change.
After the conversion, the program will make proper change on
ethhdr and ip4/6 header, recalculate checksum, and send the packet out
through bpf_redirect.

Experiment:
===

MTU: 1500B for all three machines.

The tso/lro/gro are enabled on the M_nat box.

ping works on both ways of M_ipv6 <-> M_ipv4.
It works for transfering a small file (4KB) between M_ipv6 and M_ipv4 
(both ways).
Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4, failed 
with the above BUG_ON, really fast.

Did not really test from M_ipv4 to M_ipv6 with large file.

The error path likely to be (also from the above call stack):
   nic -> lro/gro -> bpf_program -> gso (BUG_ON)

In one of experiments, I explicitly printed the skb->len and 
skb->data_len. The values are

Re: BUG_ON triggered in skb_segment

2018-03-13 Thread Eric Dumazet 



On 03/12/2018 10:45 PM, Yonghong Song wrote:

Hi,

One of our in-house projects, bpf-based NAT, hits a kernel BUG_ON at
net-next function skb_segment, line 3667.

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:
...
#0 [883ffef034f8] machine_kexec at 81044c41
  #1 [883ffef03558] __crash_kexec at 8110c525
  #2 [883ffef03620] crash_kexec at 8110d5cc
  #3 [883ffef03640] oops_end at 8101d7e7
  #4 [883ffef03668] die at 8101deb2
  #5 [883ffef03698] do_trap at 8101a700
  #6 [883ffef036e8] do_error_trap at 8101abfe
  #7 [883ffef037a0] do_invalid_op at 8101acd0
  #8 [883ffef037b0] invalid_op at 81a00bab
     [exception RIP: skb_segment+3044]
     RIP: 817e4dd4  RSP: 883ffef03860  RFLAGS: 00010216
     RAX: 2bf6  RBX: 883feb7aaa00  RCX: 0011
     RDX: 883fb87910c0  RSI: 0011  RDI: 883feb7ab500
     RBP: 883ffef03928   R8: 2ce2   R9: 27da
     R10: 01ea  R11: 2d82  R12: 883f90a1ee80
     R13: 883fb8791120  R14: 883feb7abc00  R15: 2ce2
     ORIG_RAX:   CS: 0010  SS: 0018
  #9 [883ffef03930] tcp_gso_segment at 818713e7
#10 [883ffef03990] tcp4_gso_segment at 818717d8
#11 [883ffef039b0] inet_gso_segment at 81882c9b
#12 [883ffef03a10] skb_mac_gso_segment at 817f39b8
#13 [883ffef03a38] __skb_gso_segment at 817f3ac9
#14 [883ffef03a68] validate_xmit_skb at 817f3eed
#15 [883ffef03aa8] validate_xmit_skb_list at 817f40a2
#16 [883ffef03ad8] sch_direct_xmit at 81824efb
#17 [883ffef03b20] __qdisc_run at 818251aa
#18 [883ffef03b90] __dev_queue_xmit at 817f45ed
#19 [883ffef03c08] dev_queue_xmit at 817f4b90
#20 [883ffef03c18] __bpf_redirect at 81812b66
#21 [883ffef03c40] skb_do_redirect at 81813209
#22 [883ffef03c60] __netif_receive_skb_core at 817f310d
#23 [883ffef03cc8] __netif_receive_skb at 817f32e8
#24 [883ffef03ce8] netif_receive_skb_internal at 817f5538
#25 [883ffef03d10] napi_gro_complete at 817f56c0
#26 [883ffef03d28] dev_gro_receive at 817f5ea6
#27 [883ffef03d78] napi_gro_receive at 817f6168
#28 [883ffef03da0] mlx5e_handle_rx_cqe_mpwrq at 817381c2
#29 [883ffef03e30] mlx5e_poll_rx_cq at 817386c2
#30 [883ffef03e80] mlx5e_napi_poll at 8173926e
#31 [883ffef03ed0] net_rx_action at 817f5a6e
#32 [883ffef03f48] __softirqentry_text_start at 81c000c3
#33 [883ffef03fa8] irq_exit at 8108f515
#34 [883ffef03fb8] do_IRQ at 81a01b11
---  ---
bt: cannot transition from IRQ stack to current process stack:
     IRQ stack pointer: 883ffef034f8
     process stack pointer: 81a01ae9
    current stack base: c9000c5c4000
...
Setup:
=

The test will involve three machines:
   M_ipv6 <-> M_nat <-> M_ipv4

The M_nat will do ipv4<->ipv6 address translation and then forward packet
to proper destination. The control plane will configure M_nat properly
will understand virtual ipv4 address for machine M_ipv6, and
virtual ipv6 address for machine M_ipv4.

M_nat runs a bpf program, which is attached to clsact (ingress) qdisc.
The program uses bpf_skb_change_proto to do protocol conversion.
bpf_skb_change_proto will adjust skb header_len and len properly
based on protocol change.
After the conversion, the program will make proper change on
ethhdr and ip4/6 header, recalculate checksum, and send the packet out
through bpf_redirect.

Experiment:
===

MTU: 1500B for all three machines.

The tso/lro/gro are enabled on the M_nat box.

ping works on both ways of M_ipv6 <-> M_ipv4.
It works for transfering a small file (4KB) between M_ipv6 and M_ipv4 
(both ways).
Transfering a large file (e.g., 4MB) from M_ipv6 to M_ipv4, failed with 
the above BUG_ON, really fast.

Did not really test from M_ipv4 to M_ipv6 with large file.

The error path likely to be (also from the above call stack):
   nic -> lro/gro -> bpf_program -> gso (BUG_ON)

In one of experiments, I explicitly printed the skb->len and 
skb->data_len. The values are below:

   skb_segment: len 2856, data_len