Re: [PATCH net-next v1 00/15] psp: introduce socket-less associations
From: Jamal Hadi Salim
Date: Fri Nov 13 2026 - 17:40:18 EST
On Fri, Nov 13, 2026 at 10:45 AM Jamal Hadi Salim <jhs@xxxxxxxxxxxx> wrote:
>
> On Tue, Nov 10, 2026 at 7:56 PM Todd Freed <toddfreed@xxxxxxxx> wrote:
> >
> > psp associations today are bound to a tcp socket when created. traffic
> > on that socket is encapsulated in transport-mode psp, and encrypted.
> >
> > this series adds associations which aren't bound to a socket, which
> > allows psp encapsulation and encryption to be applied to forwarded
> > traffic. you can use this to transparently apply psp tunnels to flows,
> > with help from userspace to manage the tc rules and orchestrate key
> > exchange.
> >
> > the lifecycle of socket-less associations is managed over the psp
> > netlink family. socket-less associations are stored in a small registry,
> > and are addressable by id. the registry does not contain socket
> > associations.
> >
> > socket-less associations are on the device's association lists, like
> > socket associations. rx ones are therefore marked stale by key rotation
> > in the same way. in this series, nothing acts on socket-less rx
> > associations on key rotation. instead, userspace should replace decap
> > rules for a SPI before its device key has been rotated twice. a
> > follow-up will bring enforcement.
> >
> > series structure,
> >
> > 1, 2/ prepares psp to work with non-tcp packets
> >
> > 3, 4/ adds the netlink commands for managing socket-less associations
> >
> > 5/ adds a METADATA_PSP dst which carries the socket-less psp association
> > on an skb for tx, and the transmit checks for it
> >
> > 6, 7/ keep mirred and bpf from sending decrypted skbs from egress to
> > ingress, since skb->decrypted depends on direction
> >
> > 8/ limit time-wait psp socket validator to sockets actually in
> > time-wait. the new psp actions check for presence of the validator
> >
> > 9/ adds the act_psp tc actions, psp encap and psp decap. in software,
> > encap is egress only.
> >
> > 10/ adds flower matches on psp header fields, for rx matching
> >
> > 11, 12/ add drv-net tests (netdevsim and hardware)
> >
> > 13/ makes tc-testing run dependsOn probes outside the test namespace,
> > where they currently always fail
> >
> > 14/ adds tc-testing tests
> >
> > 15/ Documentation for socket-less psp associations
> >
> > a following series adds hardware offload of the psp actions and flower
> > matches, psp tunnel mode (which is offload only and therefore not
> > included here), encap on ingress (also offload only), and the mlx5
> > implementation for switchdev mode. It's available in the tree below.
> >
> > relation to the psp fixes in net:
> >
> > this series is based on net-next, and overlaps with "psp, tls: don't
> > send offloaded contents in cleartext" in net:
> >
>
> I looked at the net/sched bits also looks like the two sashikos
> already caught at least two issues i would have brought up. I have
> started a test run on our test system to see if anything pops up; i
> dont suspect anything will show up, but if it does, I will followup to
> this posting.
> My suggestion is you address the issues brought up by the sashikos and
> spin v2. I will review when you post a v2 and possibly add my
> reviewed-by.
> In v2:
> If you have, or can capture, the decapped-GSO packet shape that leaves
> the L4 header non-linear (gro_cells/vnet_hdr decap bridged to the
> act_psp
> egress), attaching it to the v2 would be useful evidence for the GSO fix.
>
My bad - only one of the sashikos (gemini) responded. I suggest you
wait for the second one. I think it watches this gemini sashiko and
tries to come up with more clever findings.
Here's my view after trying to see if we can actually reproduce the claims:
The only legit thing is "act_psp, [Low] missing .get_fill_size", you
need to add that.
act_psp, [High] "GSO skb with non-linear L4 header -> OOB read/write"
- this sounds like BS/false positive, but the burden of proof is on
you ;-> If that code is broken, then a lot of kernel GSO code is
broken from the grep-ing. GSO skb's transport header is linear by tc
egress time.
act_psp, secondary claim that the transport_header write "defeats"
psp_dev_encap_ok() - that also looks made up.
Double check.
cheers,
jamal
> > https://lore.kernel.org/netdev/20261111001831.3710881-1-toddfreed@xxxxxxxx
> >
> > when net is merged into net-next:
> >
> > - net/psp/psp_main.c conflicts around psp_dev_encap_ok(), which both
> > add: net's only allows tcp, this series' (patch 2) takes any l4
> > protocol. keep this series' version of the function, and its
> > comments above it and above psp_dev_encapsulate().
> > - tools/testing/selftests/drivers/net/psp.py and
> > tools/testing/selftests/drivers/net/config conflict: keep both sides
> > (imports, constants, the list of test cases; the NET_ACT_* options).
> >
> > relation to the netfilter fixes in nf-next:
> >
> > packets which a psp device decrypted don't have to end in a local psp
> > socket. they can already be forwarded or bridged today, but with
> > socket-less associations that is the intended use, e.g. a router that
> > decaps and forwards. netfilter handles two cases itself, outside the ip
> > stack paths which the net series fixes, and would send such a packet's
> > decrypted contents in clear text:
> >
> > - the bridge refragments packets which bridge conntrack reassembled
> > (br_ip6_fragment(), nf_br_ip_fragment()), without the check of "net:
> > don't fragment decrypted skbs".
> > - the reject code builds its own icmp errors and tcp resets
> > (nf_reject_skb_v4_unreach() and friends), without icmp_send() and
> > icmpv6_send(), which the net series' icmp patches cover.
> >
> > a separate series to nf-next applies the net series' rules to both:
> >
> > "netfilter: bridge: don't refragment decrypted skbs"
> > "netfilter: reject: don't reply to decrypted skbs"
> >
> > https://lore.kernel.org/netfilter-devel/20261111003111.3826942-1-toddfreed@xxxxxxxx
> >
> > it doesn't depend on this series, and the two don't conflict.
> >
> > trees:
> >
> > the full tree, with the net fixes, the nf-next series, this series and
> > the mlx5 offload series, and the conflicts above resolved, is at:
> >
> > https://github.com/toddfreed/psptc branch psptc
> >
> > iproute2 tree, with support for the psp action and the psp flower
> > matches, based on iproute2-next. the selftests need it:
> >
> > https://github.com/toddfreed/iproute2-psptc branch tc-psp
> >
> > testing:
> >
> > - netdevsim: drivers/net psp.py passes (57 tests), and tc-testing
> > psp.json (17 tests).
> > - mlx5 (ConnectX-7, FW 28.48), two hosts: drivers/net psp.py passes:
> > 46 pass, 8 skip (GMAC versions not supported by the device),
> > 3 xfail (netdevsim only).
> > - each patch builds without warnings with gcc and clang, allmodconfig,
> > W=1.
> > Todd Freed (15):
> > psp: get netns from net_device instead of sock
> > psp: take the psp nexthdr from the packet on encap
> > psp: declare psp_assoc_set_tx() in psp.h
> > psp: add support for socket-less psp associations
> > net: add METADATA_PSP dst to carry a psp association on an skb
> > net/sched: act_mirred: don't send decrypted skbs from egress to
> > ingress
> > bpf: don't redirect decrypted skbs from egress to ingress
> > psp: only give time-wait sockets of psp connections a validator
> > net/sched: add act_psp for psp transport mode
> > net/sched: flower: add psp_spi, psp_ver, and psp_dev matches
> > selftests: drv-net: psp: test socket-less associations and the tc psp
> > action
> > selftests: drv-net: psp: test that software encap on ingress drops
> > selftests: tc-testing: run dependsOn probes outside the test namespace
> > selftests: tc-testing: add tests for the psp action
> > Documentation: networking: psp: document socket-less associations
> >
> > Documentation/netlink/specs/psp.yaml | 98 ++-
> > Documentation/netlink/specs/tc.yaml | 14 +
> > Documentation/networking/psp.rst | 139 +++-
> > MAINTAINERS | 3 +
> > .../mellanox/mlx5/core/en_accel/psp_rxtx.c | 2 +-
> > .../net/ethernet/mellanox/mlx5/core/en_main.c | 33 +
> > drivers/net/netdevsim/psp.c | 2 +-
> > include/linux/skbuff.h | 4 +
> > include/net/dst_metadata.h | 19 +
> > include/net/flow_dissector.h | 18 +
> > include/net/psp/functions.h | 91 ++-
> > include/net/psp/types.h | 1 +
> > include/net/tc_act/tc_psp.h | 23 +
> > include/uapi/linux/pkt_cls.h | 6 +
> > include/uapi/linux/psp.h | 6 +
> > include/uapi/linux/tc_act/tc_psp.h | 28 +
> > net/core/dev.c | 48 +-
> > net/core/dst.c | 5 +
> > net/core/filter.c | 9 +
> > net/core/flow_dissector.c | 27 +
> > net/psp/Makefile | 2 +-
> > net/psp/psp-nl-gen.c | 60 ++
> > net/psp/psp-nl-gen.h | 8 +
> > net/psp/psp.h | 7 +
> > net/psp/psp_assoc.c | 120 +++
> > net/psp/psp_main.c | 54 +-
> > net/psp/psp_nl.c | 261 ++++++
> > net/psp/psp_sock.c | 32 +-
> > net/sched/Kconfig | 13 +
> > net/sched/Makefile | 1 +
> > net/sched/act_mirred.c | 6 +
> > net/sched/act_psp.c | 439 ++++++++++
> > net/sched/cls_flower.c | 64 ++
> > tools/testing/selftests/drivers/net/config | 5 +
> > tools/testing/selftests/drivers/net/psp.py | 750 +++++++++++++++++-
> > tools/testing/selftests/tc-testing/config | 2 +
> > .../tc-testing/tc-tests/actions/psp.json | 459 +++++++++++
> > tools/testing/selftests/tc-testing/tdc.py | 2 +-
> > tools/testing/selftests/tc-testing/tdc.sh | 1 +
> > 39 files changed, 2812 insertions(+), 50 deletions(-)
> > create mode 100644 include/net/tc_act/tc_psp.h
> > create mode 100644 include/uapi/linux/tc_act/tc_psp.h
> > create mode 100644 net/psp/psp_assoc.c
> > create mode 100644 net/sched/act_psp.c
> > create mode 100644 tools/testing/selftests/tc-testing/tc-tests/actions/psp.json
> >
> >
> > base-commit: ed04b202027249e2769420893220f54a56e0c59a
> > --
> > 2.53.0-Meta
> >