Hi Ophir, <Snip>
> Hi Vipin, > I missed your point: > You claim that TAP should work regardless of any pi.proto values. > Can you confirm that for ALL kernels versions (past and future)? I have tested with 3.13.0 , 4.4.0 with patch fix. > > > -----Original Message----- > > From: Ophir Munk > > Sent: Saturday, April 21, 2018 12:49 AM > > To: Varghese, Vipin <vipin.vargh...@intel.com>; dev@dpdk.org; > > pascal.ma...@6wind.com; Yigit, Ferruh <ferruh.yi...@intel.com>; Thomas > > Monjalon <tho...@monjalon.net>; Olga Shern <ol...@mellanox.com>; > > Shahaf Shuler <shah...@mellanox.com> > > Subject: RE: [dpdk-dev] [PATCH 1/2] net/tap: add tun support > > > > Hi Vipin, > > > > Please find comments inline. > > > > > -----Original Message----- > > > From: Varghese, Vipin [mailto:vipin.vargh...@intel.com] > > > Sent: Friday, April 13, 2018 6:18 AM > > > To: Ophir Munk <ophi...@mellanox.com>; dev@dpdk.org; > > > pascal.ma...@6wind.com; Yigit, Ferruh <ferruh.yi...@intel.com>; > > > Thomas Monjalon <tho...@monjalon.net>; Olga Shern > > > <ol...@mellanox.com>; Shahaf Shuler <shah...@mellanox.com> > > > Subject: RE: [dpdk-dev] [PATCH 1/2] net/tap: add tun support > > > <Snip> > > > > 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@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. > > Thanks for sharing detailed example overview. But as you mentioned this will break ' 4025C2' and ' 64D4DA', This will not solve for the correction patch ' https://dpdk.org/dev/patchwork/patch/37986/'. Only choice left is separate tx_burst for TAP and TUN PMD, as we do not want to check PMD type on each call. Questions: 1) Is this ok to split tx_burst and have redundant code? 2) Does applications transparently send packets coming from Physical NIC to TAP interface? Does not the application Modifies the DEST MAC addr to TAP interface? > > > > > > > > 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; > > > """
