[dpdk-dev] [PATCH 1/2] net/tap: add tun support

[Ophir Munk]

Hi Vipin,

Please find comments inline.

> -----Original Message-----
> From: Varghese, Vipin [mailto:vipin.varghese at intel.com]
> Sent: Friday, April 13, 2018 6:18 AM
> To: Ophir Munk <ophirmu at mellanox.com>; dev at dpdk.org;
> pascal.mazon at 6wind.com; Yigit, Ferruh <ferruh.yigit at intel.com>; Thomas
> Monjalon <thomas at monjalon.net>; Olga Shern <olgas at mellanox.com>;
> Shahaf Shuler <shahafs at mellanox.com>
> Subject: RE: [dpdk-dev] [PATCH 1/2] net/tap: add tun support
> 
> Hi Ophir,
> 
> Please find my answers inline to the queries.
> 
> > -----Original Message-----
> > From: Ophir Munk [mailto:ophirmu at mellanox.com]
> > Sent: Thursday, April 12, 2018 5:19 PM
> > To: Varghese, Vipin <vipin.varghese at intel.com>; dev at dpdk.org;
> > pascal.mazon at 6wind.com; Yigit, Ferruh <ferruh.yigit at intel.com>; Thomas
> > Monjalon <thomas at monjalon.net>; Olga Shern <olgas at mellanox.com>;
> > Shahaf Shuler <shahafs at mellanox.com>
> > Subject: RE: [dpdk-dev] [PATCH 1/2] net/tap: add tun support
> >
> > Hi Vipin,
> > This patch (adding TUN to TAP) has been Acked and accepted in next-net
> > branch.
> > I have some questions regarding the implementation (please find below).
> >
> > > -----Original Message-----
> > > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Vipin Varghese
> > > Sent: Tuesday, April 03, 2018 12:38 AM
> > > To: dev at dpdk.org; pascal.mazon at 6wind.com; ferruh.yigit at intel.com
> > > Cc: Vipin Varghese <vipin.varghese at intel.com>
> > > Subject: [dpdk-dev] [PATCH 1/2] net/tap: add tun support
> > >
> > > The change adds functional TUN PMD logic to the existing TAP PMD.
> > > TUN PMD can be initialized with 'net_tunX' where 'X' represents unique
> id.
> > > PMD supports argument interface, while MAC address and remote are
> > > not supported.
> > >
> >
> > [...]
> >
> > >
> > > +		/*
> > > +		 * TUN and TAP are created with IFF_NO_PI disabled.
> > > +		 * For TUN PMD this mandatory as fields are used by
> > > +		 * Kernel tun.c to determine whether its IP or non IP
> > > +		 * packets.
> > > +		 *
> > > +		 * The logic fetches the first byte of data from mbuf.
> > > +		 * compares whether its v4 or v6. If none matches default
> > > +		 * value 0x00 is taken for protocol field.
> > > +		 */
> > > +		char *buff_data = rte_pktmbuf_mtod(seg, void *);
> > > +		j = (*buff_data & 0xf0);
> > > +		if (j & (0x40 | 0x60))
> > > +			pi.proto = (j == 0x40) ? 0x0008 : 0xdd86;
> > > +
> >
> > 1. Accessing the first byte here assumes it is the first IP header
> > byte (layer 3) which is correct for TUN.
> > For TAP however the first byte belongs to Ethernet destination address
> > (layer 2).
> > Please explain how this logic will work for TAP.
> 
> Based on linux code base '/driver/net/tap.c' and '/driver/net/tun.c' from
> 3.13. to  4.16,
> 
> Please find my observation below
> 1. File: tun.c, function: tun_get_user, check for 'tun->flags &
> TUN_TYPE_MASK' is done and if non ip is taken counter 'rx_dropped' is
> updated.
> 2. File: tap.c, there are no checks for 'tap->flags' for IFF_NO_PI in rx data
> path. Counter 'rx_dropped' is updated in 'tap_handle_frame'.
> 

I understand that in kernel implementation there is no check for tap->flags in file tap.c, however I think there is a bug in dpdk rte_eth_tap.c file.
Please find below an example which demonstrates this claim.

> Please find my reasoning below
> 1. First approach was to have separate function for tap and tun TX and RX.
> But this will introduce code duplication, hence reworked the code as above.

I agree. Avoiding code duplication is a good approach. 

> 2. During my internal testing assigning dummy value for protocol field in TAP
> packets, did not show a difference in behaviour. May be there are some
> specific cases this failing.
> 
> If there difference in behaviour, can please share the same?
> 

Please consider the following example:
I am running testpmd with a TAP device, --forward-mode=csum. 
I am injecting a TCP packet, which is forwarded back (mac addresses swapped) to the sender. 
Using gdb I set a breakpoint at pmd_tx_burst() in file rte_eth_tap.c

Looking at the following code inside pmd_tx_burst():

527                 char *buff_data = rte_pktmbuf_mtod(seg, void *);
528                 j = (*buff_data & 0xf0);
529                 pi.proto = (j == 0x40) ? 0x0008 :
530                                 (j == 0x60) ? 0xdd86 : 0x00;

I am printing the first 20 bytes of buff_data in line 527:

(gdb) p/x *(unsigned char *)buff_data at 20
$3 = {0x0, 0x25, 0x88, 0x10, 0x66, 0x2, 0xf4, 0x52, 0x14, 0x7a, 0x59, 0x81, 0x8, 0x0, 0x45, 0x0, 0x4, 0xdf, 0x0, 0x1}

The gdb printout refers to:
6 bytes of destination MAC address: 0x0, 0x25, 0x88, 0x10, 0x66, 0x2
6 bytes of source MAC address: 0xf4, 0x52, 0x14, 0x7a, 0x59, 0x81
2 bytes of Ethernet type: 0x8, 0x0 - (IPv4)
IP header starting with 0x45, ... which is the byte (0x45) that "j" should have looked at

In the case of TAP - buff_data starts with the destination MAC address of the sender (0x0, ...). 
The code in line 528 expects that buff_data would start with an IP header protocol (e.g. 0x45), but it is not the case for TAP.
In my case j=0x0 (line 528) which is harmless (as it ends up with setting pi.proto=0x00, which is correct for TAP).
However, if the sender had an Intel NIC - the destination MAC address could have started with:
$3 = {0x40, 0x25, 0xC2, ...
Or-
$3 = {0x64, 0xD4, 0xDA, ...

as 4025C2 and 64D4DA are reserved prefixes for Intel Ethernet MAC addresses, see: http://www.coffer.com/mac_find/?string=intel

In this case pi.proto could end up with 0x0008 or 0xdd86 instead of 0x0 as expected for TAP.

I hope that this example clarifies the bug I am referring to. 

> >
> > 2. If the first TUN byte contains 0x2X (which is neither IPv4 nor
> > IPv6) it will end up by setting ip.proto as 0xdd86.
> > Please explain how this logic will work for non-IP packets in TUN
> 
> I see your point. You are correct about this. Thanks for pointing out, may I
> send correction for this as
> 
> """
> -		if (j & (0x40 | 0x60))
> -			pi.proto = (j == 0x40) ? 0x0008 : 0xdd86;
> +		pi.proto = (j == 0x40) ? 0x0008 :
> +					(j == 0x60) ? 0xdd86 :
> +					0x00;
> """