[PATCH net v3 0/2] kcm: reject the combination of KCM and TLS

KCM's data_ready callback calls the lower socket's read_sock() inline,
from softirq. With kTLS RX that is tls_sw_read_sock(), which can sleep.
v2 deferred KCM parsing to a workqueue whenever read_sock() is not
tcp_read_sock(), and reworked the strparser workqueue to go with it. As
suggested in review [1], v3 instead rejects the combination, as was
done for sockmap in commit 460e6486617c ("tls: reject the combination
of TLS and sockmap"). It has never worked with kTLS RX: before commit
662fbcec32f4 ("net/tls: implement ->read_sock()") KCM parsed TLS
ciphertext as messages, and since then it sleeps in atomic context.

Patch 1 makes kcm_attach() refuse sockets that have any ULP, and
tls_init() refuse sockets attached to KCM. Patch 2 adds tls selftests
for both orders.

Patch 1 changes user-visible behaviour and is tagged for stable. Its
message calls out the errno change and the one refused setup that
previously passed data: TLS TX-only, in either order.

On the new sock flag: patch 1 adds SOCK_KCM to enum sock_flags, which
leaves one free sk_flags bit on 32-bit. I looked at the alternatives
first:

 - sk_user_data alone cannot be the marker. sunrpc and nvme-tcp set it
   on sockets that tlshd later upgrades to TLS, so it is already set
   when tls_init() runs and tls_init() cannot reject on it.

 - A new SK_USER_DATA_KCM tag, alongside SK_USER_DATA_PSOCK, is not
   safe. The tags live in the low bits of the stored pointer, and bit 3
   is not free: reuseport_array stores &array->ptrs[index] | NOCOPY |
   BPF, where ptrs[] holds 8-byte pointers, so every odd index has bit 3
   set and widening SK_USER_DATA_PTRMASK would corrupt it. It would also
   need masking at each of KCM's raw sk_user_data reads.

 - There is precedent for a flag owned by one subsystem: SOCK_XDP
   (commit e4a2a3048ed9) is set only by tun.

If you'd rather not spend an sk_flags bit on this, the alternative I
would implement is a bool in struct tcp_sock under
IS_ENABLED(CONFIG_AF_KCM), like is_mptcp and syn_smc. KCM only attaches
TCP sockets and the TLS ULP is TCP-only, so both checks can use
tcp_sk(). Happy to respin that way.

Changes in v3:
 - Reject KCM and ULPs in both orders instead of deferring KCM parsing
   to a workqueue.
 - Drop the strparser workqueue patch; it is no longer needed for this
   fix and may be posted to net-next separately.
 - Add selftests.

v1: https://lore.kernel.org/netdev/20261013033920.2069285-1-rjethwani@purestorage.com/ (local)
v2: https://lore.kernel.org/netdev/20261016025059.289202-1-rjethwani@purestorage.com/ (local)

[1] https://lore.kernel.org/netdev/3108e934-ec00-458b-8720-450554b35b55@linux.dev/ (local)

Rishikesh Jethwani (2):
  kcm: reject attaching sockets with a ULP, and TLS on KCM sockets
  selftests: tls: test that TLS and KCM are mutually exclusive

 include/net/sock.h                 |   1 +
 net/kcm/kcmsock.c                  |  12 +++
 net/tls/tls_main.c                 |   6 ++
 tools/testing/selftests/net/config |   1 +
 tools/testing/selftests/net/tls.c  | 144 +++++++++++++++++++++++++++++
 5 files changed, 164 insertions(+)

-- 
2.50.1