[tabgab]

On dev 764db6f, one commit. (Issues are disabled in this repository, so the problem is described here.)

Problem

A region tag that a packet carries before a gNB or a UPF tunnels it (CreationTimeTag, ResidenceTimeTag, ElapsedTimeTag, ... added with Packet::addRegionTag) is gone after the other end of the GTP-U tunnel. This happens in an IP and in an Ethernet PDU session, uplink and downlink, and across the UPF's relay from one tunnel into another.

Tags attached to a chunk survive (the chunk is shared), which hides the problem when a source tags the chunk it creates. Consumers lose their measurements: the end-to-end delay from CreationTimeTag, the residence times of TSN bridging.

Cause

src/simu5g/corenetwork/gtp/GtpUser.cc (dev 764db6f) builds a new Packet for the GTP-U message and moves the data over with insertAtBack(peekData()). That copies chunks only; the region tags of a Packet live in Packet::regionTags, not in its chunks.

handleFromTrafficFlowFilter: gtpPacket = new Packet; insertAtFront(header); insertAtBack(datagram->peekData())
handleFromUdp:               originalPacket = new Packet; pkt->popAtFront<GtpUserMsg>(); insertAtBack(pkt->peekData())
tunnelDownlink:              gtpMsg = new Packet; insertAtFront(header); insertAtBack(datagram->peekData())

INET's convention is that the region tags of a packet travel with its data through every encapsulation. INET's own tunnelling inserts the Ipv4/Udp/Ethernet headers into the same Packet, and Packet::insertAt / removeAt shift the region tags. Where INET builds a second Packet from the data of the first, it copies the region tags explicitly with Packet::copyRegionTags (EthernetCutthroughSource.cc:38, CutthroughSource.cc:38).

Reproduction

Tag a frame with addRegionTag<CreationTimeTag>() where it is about to enter the tunnel (for example as upf.ethBridge.ethernet.fcsInserter pushes it out, or as the gNB PDCP rx entity passes it up), and look for the tag where it has left the tunnel: no byte carries it. The two tests of this PR do this.

Fix

GtpUser creates a new Packet for every GTP-U message and copies the data of the original packet into it with insertAtBack(peekData()). That copies the chunks, so the tags of a chunk survive, but the region tags of the Packet were lost at each end of a tunnel. The three places that build a Packet from another one now copy the region tags of the data with Packet::copyRegionTags, as INET does where it builds a second Packet (EthernetCutthroughSource, CutthroughSource):

// handleFromTrafficFlowFilter
gtpPacket->copyRegionTags(*datagram, datagram->getFrontOffset(), header->getChunkLength(), datagram->getDataLength());
// handleFromUdp
originalPacket->copyRegionTags(*pkt, pkt->getFrontOffset(), b(0), pkt->getDataLength());
// tunnelDownlink
gtpMsg->copyRegionTags(*datagram, datagram->getFrontOffset(), header->getChunkLength(), datagram->getDataLength());

Region tag offsets count from the start of the packet content, popped front bytes included. So the source offset is the front offset of the data, and the target offset is where the data starts in the new packet (after the GTP-U header, or 0 in the decapsulated packet). copyRegionTags clips the copy to the data, so tags over the popped GTP-U header or the removed Ethernet FCS do not leak. tunnelUplinkOverN6 inserts its UDP and IP headers into the same Packet and needs no change.

Not changed on purpose: the packet-level tags (Packet::copyTags). GtpUser sets the ones it needs (PacketProtocolTag, QFI, session) itself; a blanket copy would carry stale ones such as the socket and interface tags of the outer packet.

Tests (tests/module)

• GtpRegionTags.test: Ethernet PDU session, both directions. A probe adds a CreationTimeTag over the whole frame where it is about to enter the tunnel (downlink: packetPushedOut of upf.ethBridge.ethernet.fcsInserter; uplink: sentPacketToUpperLayer of the gNB PDCP rx entity). It counts the tagged bytes where the frame has left the tunnel (downlink: receivedPacketFromUpperLayer of the gNB PDCP tx entity; uplink: packetPushedIn of upf.ethBridge.ethernet.fcsChecker).
• GtpRegionTagsIp.test: IP session, two gNBs. ue[0] at gnb sends UDP to ue[1] at gnb2, so the datagram is encapsulated at gnb, decapsulated and encapsulated again at the UPF (tunnelDownlink), and decapsulated at gnb2.

test	direction	tagged at entry	dev	with the fix
GtpRegionTags	downlink	118 of 118 bytes	0 of 115 (FAIL)	114 of 115 (PASS): only the SDAP header is untagged
GtpRegionTags	uplink	116 of 116 bytes	0 of 118 (FAIL)	112 of 118 (PASS): the 4-byte FCS the UPF rebuilds, and 2 bytes the gNB stack already leaves untagged before GtpUser
GtpRegionTagsIp	UE to UE	129 of 129 bytes	0 of 128 (FAIL)	128 of 128 (PASS)

Each of the three lines is needed: without site 1 or 2 both tests fail, and without site 3 GtpRegionTagsIp fails.

Evidence

Simu5G dev 764db6f against INET 4.7.0, OMNeT++ 6.5.0, release, macOS arm64 (./runtest in tests/module; the build here also carries the macOS intval_t change of #1, which this PR does not contain). The other module tests: 13 PASS + 1 EXPECTEDFAIL (TwoSessionsPerUe). EthernetSessionQosRuleFields fails with and without the fix on this INET (Object (inet::EthernetFcs) has no member named 'dest').

Not run: debug build; the fingerprint, unit and physics suites (no region tag enters a fingerprint ingredient, not checked); INET master; Linux.

[avarga]

The reported issue is a duplicate of Unipisa#309, and the fix (a different one from this) is present in branch v1.8.x. Currently it is recommended to use v1.8.x instead of dev for projects.