[PATCH net-next v3 0/9] tunnels: add core and gre drop reasons

[Anton Danilov]

Only vxlan reports drop reasons among the tunnel drivers today.
Everything else, on both the receive and the transmit side, ends in a
plain kfree_skb(), so a packet that a tunnel throws away is invisible
to dropwatch, drop_monitor and perf trace -e skb:kfree_skb. The device
counters group the failures coarsely: rx_errors and tx_errors each
cover half a dozen unrelated conditions.

This series covers the generic paths shared by ipip, sit, gre and their
IPv6 counterparts, plus the GRE specific parsing, on both directions.
A later series will do the same for geneve, bareudp, fou and the
remaining IP in IP drivers.

Patches 1-2 convert the generic receive paths, ip_tunnel_rcv() and
__ip6_tnl_rcv(). Two reasons are added:

  TNL_OPT_MISMATCH  the options a packet carries do not match the
                    tunnel configuration
  TNL_OLD_SEQ       the sequence number is older than the one the
                    tunnel expects, next to the existing
                    TCP_OLD_SEQUENCE

The second one has a failure mode worth naming: when a peer reboots,
its outgoing sequence number restarts at zero and the receiver drops
everything until its own counter catches up. That is indistinguishable
from a misconfiguration by the counters alone.

Patches 3-5 do the GRE specific receive path. gre_parse_header()
returns -EINVAL for six different reasons, and the only detail its
callers could get was a csum_err flag that none of them read: both
ip_gre and ip6_gre declared it, passed it in and ignored it. It is
replaced by a drop reason. Three reasons are added, mirroring vxlan:
GRE_INVALID_HDR, GRE_CSUM and GRE_TUNNEL_NOT_FOUND.

Patches 6-9 do the transmit side, some eighty failure paths across
ip_tunnel, ip_gre, ip6_tunnel and ip6_gre. One reason is added,
TNL_ENCAP, for a failure to build the encapsulation header. Patch 8 is
a small preparation: prepare_ip6gre_xmit_other() cannot fail, so it is
made void rather than given a drop reason for a branch that never runs.

The transmit side has its own case worth naming: tnl_update_pmtu()
returns -E2BIG after it has already sent the ICMP error back, which is
path MTU discovery working exactly as intended, yet the drop lands in
tx_errors next to genuine failures. An MTU black hole cannot be told
from a broken route by looking at the counters.

Drop reasons on transmit are not new: vxlan already reports several
from its xmit path, and ip_tunnel_core.c reports RECURSION_LIMIT.

Changes since v2:
- ipxip6_tnl_xmit() no longer overwrites the reason ip6_tnl_xmit()
  reported: the PKT_TOO_BIG assignment ran for every error, not just
  -EMSGSIZE, and ip6_tnl_xmit() already sets that one itself, so every
  ip4ip6, ip6ip6 and mplsip6 transmit failure was reported as an MTU
  problem
- the NBMA branch of ip6_tnl_xmit() now reports NO_TX_TARGET instead of
  IP_OUTNOROUTES for an skb that carries no destination, matching the
  sibling branch a few lines above and ip_tunnel_xmit()
- the __iptunnel_pull_header() failures now report NOMEM instead of
  HDR_TRUNC: the helper returns -ENOMEM for every failure mode and also
  fails when skb_unclone() cannot allocate, the way vxlan_rcv() labels
  it. The length checks keep HDR_TRUNC
- fixed the call graph in patch 2: the exported ip6_tnl_rcv() is only
  reached from ip6_gre, while ip4ip6, ip6ip6 and mplsip6 come through
  ipxip6_rcv()
- reworded what a NULL reason means for gre_parse_header(): a checksum
  failure is not reported, but the checksum is still computed
- new patch 8: prepare_ip6gre_xmit_other() cannot fail, its only return
  is "return 0", so it is made void and its caller stops checking it.
  v2 attached IPV6_BAD_EXTHDR to that branch, which never runs
- the source address selection of a collect_md tunnel in ip6_tnl_xmit()
  now reports IP_OUTNOROUTES instead of NO_TX_TARGET: the destination
  is known and the route was found, and the IPv6 stack itself treats a
  failed source selection as a failed lookup, the way vxlan reports it
- the commit message of the ip6_tunnel transmit patch now lists every
  reason the patch uses: TNL_ENCAP and UNHANDLED_PROTO were missing and
  RECURSION_LIMIT also covers the address conflict check
- cover letter and commit messages made precise: the transmit side
  labels some eighty failure paths, not forty; vti does not use the
  converted paths; tnl_update_pmtu() returns -E2BIG for IPv6 payloads
  too, not only for IPv4 ones with DF set
- the three selftest patches stay dropped
- v2: https://lore.kernel.org/netdev/20260913034937.875068-1-littlesmilingcloud@xxxxxxxxx/
- v1: https://lore.kernel.org/netdev/20260831215137.549324-1-littlesmilingcloud@xxxxxxxxx/

Anton Danilov (9):
  ip_tunnel: add drop reasons to the generic RX path
  ip6_tunnel: add drop reasons to the generic RX path
  gre: make gre_parse_header() report a drop reason
  ip_gre: add drop reasons to the RX path
  ip6_gre: add drop reasons to the RX path
  ip_tunnel: add drop reasons to the transmit path
  ip_gre: add drop reasons to the transmit path
  ip6_gre: make prepare_ip6gre_xmit_other() void
  ip6_tunnel: add drop reasons to the transmit path

 include/net/dropreason-core.h |  38 ++++++++
 include/net/gre.h             |   2 +-
 include/net/ip6_tunnel.h      |   3 +-
 net/ipv4/gre_demux.c          |  52 +++++++---
 net/ipv4/ip_gre.c             | 144 +++++++++++++++++++---------
 net/ipv4/ip_tunnel.c          |  60 +++++++++---
 net/ipv6/ip6_gre.c            | 175 +++++++++++++++++++++++-----------
 net/ipv6/ip6_tunnel.c         |  95 +++++++++++++-----
 8 files changed, 419 insertions(+), 150 deletions(-)

-- 
2.47.3