Date:   Sat, 23 May 2020 00:25:47 +0000
From:   Saeed Mahameed <saeedm@...lanox.com>
To:     Huy Nguyen <huyn@...lanox.com>,
        "davem@...emloft.net" <davem@...emloft.net>
CC:     Boris Pismenny <borisp@...lanox.com>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "steffen.klassert@...unet.com" <steffen.klassert@...unet.com>,
        Raed Salem <raeds@...lanox.com>,
        "huyn@...dia.com" <huyn@...dia.com>
Subject: Re: [PATCH] xfrm: Fix double ESP trailer insertion in IPsec crypto
 offload

On Thu, 2020-05-21 at 16:49 -0500, Huy Nguyen wrote:
> During IPsec performance testing, we see bad ICMP checksum. The issue
> is that
> the error packet that has duplicated ESP trailer. For example, this
> below ping reply skb is
> collected at mlx5e_xmit. This ping reply skb length is 154 because it
> has
> extra duplicate 20 bytes of ESP trailer. The correct length is 134.
>   skb len=154 headroom=2 headlen=154 tailroom=36
>   mac=(2,14) net=(16,20) trans=36
>   shinfo(txflags=0 nr_frags=0 gso(size=0 type=0 segs=0))
>   csum(0xd21a62ff ip_summed=0 complete_sw=0 valid=0 level=0)
>   hash(0x0 sw=0 l4=0) proto=0x0800 pkttype=0 iif=0
>   dev name=enp4s0f0np0 feat=0x0x001ca1829fd14ba9
>   sk family=2 type=3 proto=1
>   skb headroom: 00000000: 00 00
>   skb linear:   00000000: b8 59 9f da d6 6a b8 59 9f da d5 52 08 00
> 45 00
>   skb linear:   00000010: 00 8c 76 0f 00 00 40 32 80 5f c0 a8 01 41
> c0 a8
>   skb linear:   00000020: 01 40 8e 20 a1 20 00 39 03 28 c0 a8 01 41
> c0 a8
>   skb linear:   00000030: 01 40 00 00 12 ec cf ba 03 24 97 cf a9 5e
> 00 00
>   skb linear:   00000040: 00 00 13 34 07 00 00 00 00 00 10 11 12 13
> 14 15
>   skb linear:   00000050: 16 17 18 19 1a 1b 1c 1d 1e 1f 20 21 22 23
> 24 25
>   skb linear:   00000060: 26 27 28 29 2a 2b 2c 2d 2e 2f 30 31 32 33
> 34 35
>   skb linear:   00000070: 36 37 01 02 02 01 00 00 00 00 00 00 00 00
> 00 00
>   skb linear:   00000080: 00 00 00 00 00 00 01 02 02 01 00 00 00 00
> 00 00
>   skb linear:   00000090: 00 00 00 00 00 00 00 00 00 00
>   skb tailroom: 00000000: 00 00 00 00 00 00 00 00 00 00 00 00 a8 50
> 69 d7
>   skb tailroom: 00000010: 96 9f ff ff a8 50 69 d7 96 9f ff ff c0 01
> 58 d0
>   skb tailroom: 00000020: 96 9f ff ff
> 

You don't need to attach the whole debug dumps you have pulled to find
out what the root cause is, we do believe you ;-).

The above dump is just a random and cluttered information that doesn't
really help the cause of this patch's commit message, it is perfectly
fine to just say:

Duplicated ESP trailer can occur due to double validation of xfrm xmit
handler in case of packet xmit re-queue after 1st validation due to the
reason you listed below.

> We figure out that the packet goes through two sch_direct_xmit from
> qdsic.
> The first one is from ip_output and the later one is from NET_TX
> softirq. Below are the two stack traces on the same packet. The first
> one
> fails to send the packet because netif_xmit_frozen_or_stopped is true
> and
> the packet gets dev_requeue_skb. However at this stage, the packet
> already has the ESP trailer. Fix by marking the skb with XFRM_XMIT
> bit after
> the packet is handled by validate_xmit_xfrm to avoid duplicate ESP
> trailer insertion.
> 
> 1st one via ip_output
>   dump_stack+0x66/0x90
>   esp_output_head+0x21a/0x520 [esp4]
>   esp_xmit+0x12e/0x270 [esp4_offload]
>   ? ktime_get+0x36/0xa0
>   validate_xmit_xfrm+0x247/0x2f0
>   ? validate_xmit_skb+0x1d/0x270
>   validate_xmit_skb_list+0x46/0x70
>   sch_direct_xmit+0x18a/0x320
>   __qdisc_run+0x144/0x530
>   __dev_queue_xmit+0x3bb/0x8a0
>   ip_finish_output2+0x3ee/0x5b0
>   ip_output+0x6d/0xe0
> 
> 2nd one via NET_TX softirq
>   dump_stack+0x66/0x90
>   esp_output_head.cold.29+0x22/0x27 [esp4]
>   esp_xmit+0x12e/0x270 [esp4_offload]
>   validate_xmit_xfrm+0x247/0x2f0
>   ? validate_xmit_skb+0x1d/0x270
>   validate_xmit_skb_list+0x46/0x70
>   sch_direct_xmit+0x18a/0x320
>   __qdisc_run+0x144/0x530
>   net_tx_action+0x15d/0x240
>   __do_softirq+0xdf/0x2e5
>   irq_exit+0xdb/0xe0
>   smp_apic_timer_interrupt+0x74/0x130
>   apic_timer_interrupt+0xf/0x20
> 

Same, this comes from your own debug code.. doesn't help the cause of
the commit message. you can just describe the flows that might
retrigger double validation on the skb.

So please improve commit message and avoid clutter.

> issue: 2143007
> Fixes: f6e27114a60a ("net: Add a xfrm validate function to
> validate_xmit_skb")
> Change-Id: I2bc1a189b8160cd90b66b44212b4d44bbdebcaea

Please remove "issue:" and "Change-Id:" clutter.

> Signed-off-by: Huy Nguyen <huyn@...lanox.com>
> Reviewed-by: Boris Pismenny <borisp@...lanox.com>
> Reviewed-by: Raed Salem <raeds@...lanox.com>
> ---
>  include/net/xfrm.h     | 1 +
>  net/xfrm/xfrm_device.c | 4 +++-
>  2 files changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/include/net/xfrm.h b/include/net/xfrm.h
> index 8f71c11..0302470 100644
> --- a/include/net/xfrm.h
> +++ b/include/net/xfrm.h
> @@ -1013,6 +1013,7 @@ struct xfrm_offload {
>  #define	XFRM_GRO		32
>  #define	XFRM_ESP_NO_TRAILER	64
>  #define	XFRM_DEV_RESUME		128
> +#define	XFRM_XMIT		256
>  
>  	__u32			status;
>  #define CRYPTO_SUCCESS				1
> diff --git a/net/xfrm/xfrm_device.c b/net/xfrm/xfrm_device.c
> index 6cc7f7f..c122e3e 100644
> --- a/net/xfrm/xfrm_device.c
> +++ b/net/xfrm/xfrm_device.c
> @@ -110,7 +110,7 @@ struct sk_buff *validate_xmit_xfrm(struct sk_buff
> *skb, netdev_features_t featur
>  	struct xfrm_offload *xo = xfrm_offload(skb);
>  	struct sec_path *sp;
>  
> -	if (!xo)
> +	if (!xo || (xo->flags & XFRM_XMIT))
>  		return skb;
>  
>  	if (!(features & NETIF_F_HW_ESP))
> @@ -131,6 +131,8 @@ struct sk_buff *validate_xmit_xfrm(struct sk_buff
> *skb, netdev_features_t featur
>  		return skb;
>  	}
>  
> +	xo->flags |= XFRM_XMIT;
> +

XFRM_XMIT sounds like a poor name, as you explained the packet is not
actually transmitted, but re-scheduled for later even after it was
already validated/handled by xfrm, i would pick a different name

perhaps XFRM_XMIT_VALID.

>  	if (skb_is_gso(skb)) {
>  		struct net_device *dev = skb->dev;
>  

Powered by blists - more mailing lists