linux-stable.git/net/ipv4, branch v7.1.12

inet: frags: strip GSO state from fragments before reassembly

2026-08-28T06:24:19+00:00

commit d5dc1e69fd7258ea605c9952e5d5947539159ae3 upstream.

A virtio_net_hdr (tun/tap, or AF_PACKET with PACKET_VNET_HDR) can mark
an IPv4 or IPv6 fragment as GSO; nothing relates gso_type to frag_off.
inet_frag_reasm_prepare()/inet_frag_reasm_finish() keep the first
fragment's skb as the head of the reassembled datagram, including its
shinfo->gso_size/gso_type/gso_segs, and chain the remaining fragments
on frag_list with whatever linear/paged layout they arrived with.

After ip_defrag() (ip_local_deliver(), nf_defrag_ipv4, ...) the
reassembled skb therefore still claims to be GSO (SKB_GSO_DODGY), and
the next software segmentation point - udp_rcv_segment() on local
delivery, validate_xmit_skb(), or the ip_finish_output_gso() slow
path - hands it to skb_segment(). skb_segment()'s frag_list walk
assumes GRO-shaped input and hits one of its BUG_ON()s. Two writes to
a tap by an unprivileged user in its own userns are enough:

  kernel BUG at net/core/skbuff.c:4899!
  Oops: invalid opcode: 0000 [#1] SMP KASAN NOPTI
  CPU: 0 UID: 1000 PID: 82 Comm: poc Not tainted 7.2.0-pentest+ #2
  RIP: 0010:skb_segment+0x20ca/0x48b0
  Call Trace:
   
   __udp_gso_segment+0x29a/0x27d0
   udp4_ufo_fragment+0x458/0x6c0
   inet_gso_segment+0x429/0x1340
   skb_mac_gso_segment+0x233/0x4f0
   __skb_gso_segment+0x308/0x660
   udp_queue_rcv_skb+0x440/0xad0
   udp_unicast_rcv_skb+0xc7/0x2c0
   udp_rcv+0x16ce/0x2260
   ip_protocol_deliver_rcu+0x197/0x2d0
   ip_local_deliver+0x430/0x690
   ip_rcv+0x16f/0x1f0
   __netif_receive_skb_one_core+0x15e/0x1c0
   __netif_receive_skb+0x1e/0x110
   netif_receive_skb+0xf6/0x5c0
   tun_rx_batched.isra.0+0x3ab/0x790
   tun_get_user+0x17c3/0x3550
   tun_chr_write_iter+0xba/0x1b0
   vfs_write+0x646/0x1130
   
  Kernel panic - not syncing: Fatal exception in interrupt

This runs with BH disabled, so it is a panic rather than an oops. The
same is reachable with CAP_NET_RAW in a netns where a defrag point
precedes a GSO point, and from a guest whose VMM forwards
virtio_net_hdr to a tap. The SKB_GSO_DODGY frag_list checks added by
commit 3dcbdb134f32 ("net: gso: Fix skb_segment splat when splitting
gso_size mangled skb having linear-headed frag_list") and by
commit 9e4b7a99a03a ("net: gso: fix panic on frag_list with mixed head
alloc types") do not cover it: page-backed heads skip them, and kmalloc
heads skip them when gso_size == skb_headlen(head), which the sender
controls.

An skb entering a frag queue is an IP fragment by definition and
cannot legitimately carry GSO state: GRO does not merge fragments and
the stack segments before it fragments, so only untrusted sources are
affected. This has been reachable since
commit f43798c27684 ("tun: Allow GSO using virtio_net_hdr"), the first
path that let userspace attach GSO metadata to an IP fragment. Reset
the GSO fields of every fragment as it is queued, in
inet_frag_queue_insert(), which IPv4, IPv6, nf_conntrack_reasm and
6lowpan reassembly share; then neither the head nor the frag_list
members of the reassembled skb carry them (the members matter too:
the ip_do_fragment()/ip6_fragment() fast paths send them out as they
are). The head may remain CHECKSUM_PARTIAL; that is already accepted
on receive and resolved by skb_checksum_help() in
ip_do_fragment()/ip6_fragment() on forward.

Tested on top of net.git (dc4b95b8fee9), x86_64: the tap reproducer
above, two further IPv4 frag_list geometries that reach
BUG_ON(i >= nfrags) and BUG_ON(!list_skb->head_frag), and an IPv6
fragment-header variant (udp6_ufo_fragment()) each panic the unpatched
kernel; with this patch all four datagrams are delivered intact and
nothing is logged.

Fixes: f43798c27684 ("tun: Allow GSO using virtio_net_hdr")
Cc: stable@kernel.org
Suggested-by: Eric Dumazet 
Signed-off-by: Xinyang Ge 
Signed-off-by: Paolo Abeni 
Reviewed-by: Eric Dumazet 
Link: https://patch.msgid.link/937926e509f2acd8e0e66520dc2b30fd6b4d1687.1787839506.git.pabeni@redhat.com
Signed-off-by: Jakub Kicinski 
Signed-off-by: Greg Kroah-Hartman

ipv4: reject undersized MTUs in ip_do_fragment()

2026-08-27T12:34:37+00:00

commit c0726f0caf8c6b3208552949e17d23634a2f3129 upstream.

ip_do_fragment() subtracts the IPv4 header length from the effective
MTU and passes the resulting payload MTU to ip_frag_next().

If the effective MTU is smaller than hlen + 8, ip_frag_next() rounds
the fragment payload length down to zero. The fragmentation state then
never makes forward progress: state->left, state->ptr and state->offset
stay unchanged while ip_do_fragment() keeps allocating and transmitting
header-only fragments until the softlockup detector fires.

This is reproducible with a route installed using "mtu lock 20", but it
is also reproducible without route MTU lock, for example by forwarding a
packet to a device whose MTU is 20.

Fix it in ip_do_fragment() by rejecting mtu < hlen + 8 with -EMSGSIZE,
matching the existing IPv6 fragmentation check.

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Cc: stable@vger.kernel.org
Reported-by: Vega 
Signed-off-by: Yong Wang 
Signed-off-by: Ren Wei 
Reviewed-by: Ido Schimmel 
Link: https://patch.msgid.link/8809ef6314b98913681b0b370a05a85c2b6cd579.1786599079.git.edragain@163.com
Signed-off-by: Jakub Kicinski 
Signed-off-by: Greg Kroah-Hartman

tcp: fix icsk_ack.ato bitfield overflow

2026-08-23T12:29:32+00:00

[ Upstream commit 60837e4b840a9c3f7ec826e3584df0bc6542a2c2 ]

On cross-region connections we observed delayed ACKs suddenly turning
into immediate ACKs plus a TCP_MAX_QUICKACKS burst, as if the
connection had just received its first data segment.

Commit 95b9a87c6a6b ("tcp: record last received ipv6 flowlabel")
squeezed icsk_ack.ato into 8 bits, sized for TCP_DELACK_MAX. But both
writers still bound ato by icsk_rto, which can be well above 255
jiffies, so the bitfield assignment silently wraps mod 256: repeated
delack timer misses double ato up to icsk_rto, storing 320 as 64 and
256 as 0, and ato == 0 is the "first data packet" sentinel in
tcp_event_data_recv().

Clamp both writers to TCP_DELACK_MAX, which the static_assert already
guarantees to fit and tcp_send_delayed_ack() effectively caps ato at
anyway.

Fixes: 95b9a87c6a6b ("tcp: record last received ipv6 flowlabel")
Signed-off-by: Jiayuan Chen 
Reviewed-by: Neal Cardwell 
Link: https://patch.msgid.link/20260807014437.36687-1-jiayuan.chen@linux.dev
Signed-off-by: Jakub Kicinski 
Signed-off-by: Sasha Levin

bpf, sockmap: Fix sk_redir use-after-free in send verdict

2026-08-19T16:20:31+00:00

commit a76624733730e541e4955fdecf506af2f6b20558 upstream.

sk_psock_msg_verdict() takes a socket reference for psock->sk_redir.
tcp_bpf_send_verdict() copies that pointer while holding the source socket
lock, but does not take a reference for the local copy before dropping the
lock around tcp_bpf_sendmsg_redir().

When apply_bytes keeps the cached verdict active, another sendmsg() on the
same source socket can consume the remaining bytes and release the cached
reference while the first thread still holds only the raw local pointer:

  CPU 0                                  CPU 1
  sk_redir = psock->sk_redir
  apply_bytes remains nonzero
  release_sock(sk)
                                         lock_sock(sk)
                                         apply_bytes reaches zero
                                         psock->sk_redir = NULL
                                         release_sock(sk)
                                         tcp_bpf_sendmsg_redir(sk_redir)
                                         sock_put(sk_redir)
  tcp_bpf_sendmsg_redir(sk_redir)

The final sock_put() can free sk_redir before CPU 0 dereferences it.

KASAN reported:

  BUG: KASAN: slab-use-after-free in tcp_bpf_sendmsg_redir+0xf39/0x1020
  Read of size 8 at addr ffff888108537090 by task poc/87
  Call Trace:
   tcp_bpf_sendmsg_redir+0xf39/0x1020
   tcp_bpf_sendmsg+0x977/0x1a50
   __sys_sendto+0x32c/0x3a0
   __x64_sys_sendto+0xdb/0x1b0
  Allocated by task 85:
   sk_prot_alloc+0x56/0x210
   sk_clone+0x6f/0x14b0
   inet_csk_clone_lock+0x24/0x740
   tcp_create_openreq_child+0x25/0x2710
   tcp_v4_syn_recv_sock+0x10a/0xe00
  Freed by task 0:
   __kasan_slab_free+0x43/0x70
   slab_free_after_rcu_debug+0xa6/0x1e0
   rcu_core+0x50a/0x1850
  Last potentially related work creation:
   __sk_destruct+0x3da/0x540
   sk_psock_destroy+0x81e/0xab0
   process_one_work+0x63a/0x1070

Take a temporary socket reference while the source socket lock still
protects psock->sk_redir, and drop it after tcp_bpf_sendmsg_redir()
returns.  This keeps each unlocked use independent of cached-verdict
ownership.

Fixes: 604326b41a6f ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: Chengfeng Ye 
Reviewed-by: John Fastabend 
Reviewed-by: Emil Tsalapatis 
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/bpf/20260719152207.2892156-1-nicoyip.dev@gmail.com
Signed-off-by: Kumar Kartikeya Dwivedi 
Signed-off-by: Greg Kroah-Hartman

ipv4: fix use-after-free in fib_nhc_update_mtu()

2026-08-19T16:20:25+00:00

commit bc5bde9ce3cc36502839dfe98e068f7303a50982 upstream.

fib_nhc_update_mtu() walks the nexthop exception table under RTNL, but
RTNL does not serialize this walk with PMTU exception updates. The walk
uses rcu_dereference_protected() with a constant true condition without
holding fnhe_lock.

The following interleaving can therefore occur:

  CPU 0                              CPU 1
  fib_nhc_update_mtu()               update_or_create_fnhe()
    load fnhe                          spin_lock_bh(&fnhe_lock)
                                       fnhe_remove_oldest()
                                         unlink fnhe
                                         kfree_rcu(fnhe, rcu)
    
    access fnhe after grace period

KASAN reported:

  BUG: KASAN: slab-use-after-free in fib_nhc_update_mtu+0x3df/0x410
  Read of size 8 at addr ffff888107d49000 by task poc/90
  Call Trace:
   fib_nhc_update_mtu+0x3df/0x410
   fib_sync_mtu+0x7a/0xd0
   fib_netdev_event+0x229/0x3f0
   netif_set_mtu_ext+0x33a/0x570
   dev_set_mtu+0x88/0x120

The same walk updates fnhe_pmtu and fnhe_mtu_locked. These fields form a
pair and other writers serialize them with fnhe_lock. RCU alone prevents
reclamation, but would still allow concurrent writers to leave a mixed
pair.

Walk the table under RCU and acquire fnhe_lock only while updating each
exception. RCU keeps the current entry alive while the short critical
section serializes its paired PMTU fields. This avoids holding the global
lock while scanning all 2048 buckets for every nexthop.

Fixes: af7d6cce5369 ("net: ipv4: update fnhe_pmtu when first hop's MTU changes")
Cc: stable@vger.kernel.org
Suggested-by: Ido Schimmel 
Signed-off-by: Chengfeng Ye 
Reviewed-by: Ido Schimmel 
Link: https://patch.msgid.link/20260807181710.1178747-1-nicoyip.dev@gmail.com
Signed-off-by: Paolo Abeni 
Signed-off-by: Greg Kroah-Hartman

ipv4: Fix fib_nlmsg_size() for RTA_VIA nexthops

2026-08-19T16:20:24+00:00

commit 4ff9548d84945d2cbf9e4c207288063a200ea397 upstream.

fib_nlmsg_size() still estimates nexthop space as if every gateway is
encoded as an IPv4 RTA_GATEWAY attribute. IPv4 routes can also carry an
IPv6 gateway, which fib_nexthop_info() dumps as RTA_VIA.

As a result, route notifications can allocate an skb that is too small.
fib_dump_info() then fails with -EMSGSIZE and rtmsg_fib() hits the
WARN_ON() that marks such failures as a fib_nlmsg_size() bug. With
panic_on_warn set, this becomes a kernel panic.

Mirror the actual nexthop dump layout in fib_nlmsg_size(): account for
IPv6 nexthop gateways dumped as RTA_VIA, for the no-header rtnexthop
layout used inside RTA_MULTIPATH, and for RTA_FLOW only when it is
actually present.

Fixes: d15662682db2 ("ipv4: Allow ipv6 gateway with ipv4 routes")
Cc: stable@vger.kernel.org
Reported-by: Vega 
Signed-off-by: Zihan Xi 
Reviewed-by: Ido Schimmel 
Link: https://patch.msgid.link/6f53fa797fcaeb26966432ed7ae9bb87c4961f37.1785411220.git.zihanx@nebusec.ai
Signed-off-by: Jakub Kicinski 
Signed-off-by: Greg Kroah-Hartman

inet: frags: publish queues before arming timer

2026-08-19T16:20:24+00:00

commit 653d7ddf6cba867777a3d14c4f83ace008c5ad13 upstream.

inet_frag_create() arms the fragment queue timer before inserting the
queue into the fqdir rhashtable. If the namespace fragment timeout is
zero or negative, the timer can run before the queue is published.

The timer callback then marks the queue complete, tries to remove a node
that is not in the hash table yet, and drops the anticipated hash
reference. Creation can subsequently publish the completed queue without
restoring that reference, leaving a stale hash node after the caller drops
the remaining reference.

Publish the queue first and arm the timer while holding the queue lock.
This makes timer expiry wait until the queue is visible in the hash table,
so inet_frag_kill() can remove the node and balance the hash reference.

Fixes: 648700f76b03 ("inet: frags: use rhashtables for reassembly units")
Cc: stable@vger.kernel.org
Reported-by: Vega 
Signed-off-by: Zhiling Zou 
Signed-off-by: Ren Wei 
Link: https://patch.msgid.link/bf66785e7c0c139d7a1900e2f01faeeab344b960.1784948849.git.zhilinz@nebusec.ai
Signed-off-by: Jakub Kicinski 
Signed-off-by: Greg Kroah-Hartman

tcp: fix TFO max_qlen accounting across reuseport migration

2026-08-19T16:20:19+00:00

[ Upstream commit a0ab2ba83e35159d81cec830a92e885ecf8139be ]

A listener's TCP_FASTOPEN max_qlen stops being accurate and lets through
far more pending Fast Open requests than it was configured for.

This only shows up with SO_REUSEPORT listener migration, where closing a
listener hands its still-pending TFO children over to a surviving one.

fastopenq.qlen is charged in tcp_fastopen_create_child() when the child
is created and uncharged in reqsk_fastopen_remove() when the handshake
completes.  The uncharge follows rsk_listener of the request the child
points at, and inet_reqsk_clone() has repointed the child at a new
request owned by the new listener, so the ++ and the -- land on two
different sockets.  The new listener's qlen drifts negative and its
limit no longer binds.

Charge the new listener during migration, like reqsk_queue_migrated()
already does for queue->young and queue->qlen.

Fixes: 54b92e841937 ("tcp: Migrate TCP_ESTABLISHED/TCP_SYN_RECV sockets in accept queues.")
Signed-off-by: Jiayuan Chen 
Reviewed-by: Kuniyuki Iwashima 
Reviewed-by: Eric Dumazet 
Link: https://patch.msgid.link/20260803061739.134737-1-jiayuan.chen@linux.dev
Signed-off-by: Jakub Kicinski 
Signed-off-by: Sasha Levin

udp: fix potential use-after-free in tunnel segmentation

2026-08-19T16:20:18+00:00

[ Upstream commit d0f86fb36eb260abd10007b62c9dcc1028e03e61 ]

__skb_udp_tunnel_segment() gets the UDP header before ensuring the
tunnel header is in the skb head. If the pull reallocates skb->head,
the saved UDP header pointer is no longer valid.

Get the UDP header after the pull to avoid a potential use-after-free.

Fixes: dbef491ebe7f ("udp: Use uh->len instead of skb->len to compute checksum in segmentation")
Signed-off-by: Xuanqiang Luo 
Reviewed-by: Antoine Tenart 
Link: https://patch.msgid.link/20260730093554.68127-1-xuanqiang.luo@linux.dev
Signed-off-by: Jakub Kicinski 
Signed-off-by: Sasha Levin

tcp: do not change rcv_ssthresh in tcp_measure_rcv_mss()

2026-08-19T16:20:17+00:00

[ Upstream commit 0e125ecfe20c077625cf0be8d750d5c3abc0dce9 ]

Commit f5da7c45188e ("tcp: adjust rcvq_space after updating scaling
ratio") replaced the direct window_clamp update in tcp_measure_rcv_mss()
with a call to tcp_set_window_clamp(), a helper that implements the
TCP_WINDOW_CLAMP setsockopt. As a side effect, the helper also shrinks
rcv_ssthresh via __tcp_adjust_rcv_ssthresh().

As a result, each scaling_ratio decrease detected by
tcp_measure_rcv_mss() also cuts rcv_ssthresh. Elsewhere in TCP,
rcv_ssthresh is usually cut under memory pressure and grows via
tcp_grow_window().

Flows whose segment sizes vary keep scaling_ratio oscillating, which
leads to an unstable rcv_ssthresh: a dip of rcv_ssthresh only recovers
via tcp_grow_window(), keeping the advertised window at a relatively
low level even after the ratio itself has recovered, and can even stall
the sender.

Observed on a customer's proxy gateway after upgrading from kernel 6.1
to 6.12: in the worst case, rcv_ssthresh was cut in half by a
scaling_ratio dip. P99 latency jumped from <10ms on 6.1 to ~100ms on
6.12, and almost returned to the 6.1 level with this patch applied.

Restore the plain WRITE_ONCE() update of window_clamp, as introduced
in commit a2cbb1603943 ("tcp: Update window clamping condition"), and
keep the rcvq_space.space adjustment. Now rcv_ssthresh is decoupled from
scaling_ratio changes in tcp_measure_rcv_mss().

Fixes: f5da7c45188e ("tcp: adjust rcvq_space after updating scaling ratio")
Signed-off-by: Nathan Gao 
Link: https://patch.msgid.link/20260725030806.28135-1-zcgao@amazon.com
Signed-off-by: Jakub Kicinski 
Signed-off-by: Sasha Levin