The Statistic Value of Dropped Packets of Network Interface Increases Constantly in UEK2/UEK3/UEK4 Kernel

Asset ID:	1-72-1515650.1
Update Date:	2017-12-13
Keywords:

Solution Type Problem Resolution Sure

Solution 1515650.1 : The Statistic Value of Dropped Packets of Network Interface Increases Constantly in UEK2/UEK3/UEK4 Kernel

Applies to:

Exadata X4-2 Full Rack - Version All Versions and later
Linux OS - Version Oracle Linux 6.5 with Unbreakable Enterprise Kernel [3.8.13] to Oracle Linux 7.3 with Unbreakable Enterprise Kerne [3.8.13] [Release OL6U5 to OL7U3]
Linux OS - Version Oracle Linux 5.0 to Oracle Linux 6.9 with Unbreakable Enterprise Kernel [2.6.39] [Release OL5 to OL6U9]
Linux x86-64
Linux x86

Symptoms

The network interface card(NIC) drops packets continuously when run Oracle UEK2 kernel (2.6.39) or UEK3 kernel (3.8.13) or UEK4 and the packets dropping issue does not exist under RHCK kernel (RedHat Compatible Kernel) like 2.6.32-279.el6.

Example below:

# ifconfig
xxx Link encap:Ethernet HWaddr 00:26:B9:37:A5:11
  inet addr:xx.xx.xx.xx Bcast:xx.xx.255.255 Mask:255.254.0.0
  inet6 addr: fe80::226:b9ff:fe37:a581/64 Scope:Link
  UP BROADCAST RUNNING MASTER MULTICAST MTU:1500 Metric:1
  RX packets:39381 errors:0 dropped:131523 overruns:0 frame:0
  TX packets:24450 errors:0 dropped:0 overruns:0 carrier:0
  collisions:0 txqueuelen:0
  RX bytes:3018775 (2.8 MiB) TX bytes:13532143 (12.9 MiB)

This applies to all types of interface: native, bonded, vlan and virtual.

Cause

The packets are always dropped regardless of kernel version. However, kernels prior to 2.6.37 do not count and report these dropped packets at all, just silently discard them,it's only in 2.6.37 and later that the packet drops are counted and reported.

Solution

This is not a bug. All kernels prior to 2.6.37 do not record packets that are received on an interface for which there is no handler. Recording all dropped packets is by design and is intentional from 2.6.37.In the case of active-backup bonds, this includes all packets sent to passive interface. Other packets that will be dropped include broadcast packets that are of no interest to the host.

Note that the "dropped" in ifconfig count situations where a packet was successfully received, but the kernel *chose* to drop the packet because it had no handler for the data in the packet. So, the drop counters are entirely normal (if strange-looking); however, they can point to traffic being on the network that isn't intended to be there (such as misconfigured/rogue hosts).

Relatively low levels of dropped frames, a few percent, is normal and expected; High levels of dropped frames usually indicate misbehaving or rogue hosts elsewhere on the LAN.

Most layer-2 protocols which are conditionally used (e.g., VLANs) use negotiation with switches to avoid forwarding of traffic that is useless to the local host. So when the number of drops is relatively high, there really is traffic coming to the local machine's network port which should not be there -- probably from a host elsewhere on the same network doing improper things or sending a lot of broadcast packets.

Note that problem situations on receive are reported in the "errors", "overruns", and "frame" counters. These count received-but-malformed packets, hardware FIFO overflows, and wire-level packet corruptions, respectively.

References

<BUG:14790368> - THE SECOND SLAVE NIC DROPS PACKETS UNDER MODE 6 IN LINUX 6.3
<BUG:18468859> - PACKET DROPS DUE TO MISSING LLT PROTOCOL HANDLER
<BUG:17960125> - MELLANOX DRIVER SETS PORT TO ETH BUT NO TRAFFIC ACROSS PORT

Attachments

This solution has no attachment