On 11/15, Mina Almasry wrote:
tcp_clone_payload() takes references on source skb fragments when
building an MTU probe skb, but never attaches the source skb's zerocopy
ubuf_info (uarg) to the probe skb. When tcp_mtu_probe() subsequently
frees fully consumed source skbs via tcp_eat_one_skb(), skb_release_data()
drops their uarg references and can deliver the zerocopy completion to
userspace while the unacknowledged probe skb still references the user
buffers.
Additionally, tcp_can_coalesce_send_queue_head() only checks
skb_pure_zcopy_same() (which compares SKBFL_PURE_ZEROCOPY), allowing
adjacent zerocopy skbs with different uargs (or a mixed-zerocopy skb and
a non-zerocopy skb) to be coalesced into a single probe skb that can
only hold one destructor_arg.
Reject coalescing across different skb_zcopy() uargs in
tcp_can_coalesce_send_queue_head() and propagate the uarg to the probe
skb via skb_zcopy_set() in tcp_clone_payload().
This is a pre-existing issue found by manual code inspection and LLM
review; it was not triggered in production. Verified with the net_core
KUnit test suite; not tested on real hardware.
Fixes: 736013292e3c ("tcp: let tcp_mtu_probe() build headless packets")
Signed-off-by: Mina Almasry <redacted>
Acked-by: Stanislav Fomichev <sdf@fomichev.me>