linux-stable.git/include/net, branch v6.18.39

Bluetooth: hci_conn: Fix null ptr deref in hci_abort_conn()

2026-07-18T14:53:17+00:00

commit 12917f591cea1af36087dba5b9ec888652f0b42a upstream.

hci_abort_conn() read hci_skb_event(hdev->sent_cmd) when a connection
was pending, but hdev->sent_cmd can be NULL while req_status is still
HCI_REQ_PEND, leading to a NULL pointer dereference and a general
protection fault from the hci_rx_work() receive path.

Instead of inspecting hdev->sent_cmd, track the in-flight create
connection command with a new per-connection HCI_CONN_CREATE flag and
route all cancellation through hci_cancel_connect_sync(), which
dispatches to a dedicated per-type cancel function. The create command
is in exactly one of two states: still queued, or in flight. The cancel
function holds cmd_sync_work_lock across the whole decision: the worker
takes this lock to dequeue every entry, so while it is held a queued
command cannot start running and an in-flight command cannot complete
and let the next command become pending. This keeps the flag test and
hci_cmd_sync_cancel() atomic with respect to the worker, so a queued
command is simply dequeued, and an in-flight command owned by this
connection is cancelled without the risk of cancelling an unrelated
command that became pending in the meantime. CIS uses the same flag
mechanism via HCI_CONN_CREATE_CIS but cannot be dequeued per-connection.

hci_acl_create_conn_sync() and hci_le_create_conn_sync() clear
HCI_CONN_CREATE after the create command completes, but the command
status handler can free conn via hci_conn_del() (for example when the
controller rejects the connection) while the worker is still blocked on
the connection complete event. Hold a reference on conn across the
create command so the flag can be cleared without a use-after-free.

Fixes: a13f316e90fd ("Bluetooth: hci_conn: Consolidate code for aborting connections")
Cc: stable@vger.kernel.org
Suggested-by: XIAO WU 
Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Siwei Zhang 
Signed-off-by: Luiz Augusto von Dentz 
Signed-off-by: Greg Kroah-Hartman

Bluetooth: L2CAP: Fix UAF in channel timeout by holding conn ref

2026-07-18T14:53:15+00:00

commit b66774b48dd98f07254951f74ea6f513efe7ff8b upstream.

l2cap_chan_timeout() runs asynchronously and accesses chan->conn. If
the connection is torn down while the timer is running or pending,
chan->conn can be freed, leading to a use-after-free when the timer
worker attempts to lock conn->lock:

| BUG: KASAN: slab-use-after-free in instrument_atomic_read_write include/linux/instrumented.h:112 [inline]
| BUG: KASAN: slab-use-after-free in atomic_long_try_cmpxchg_acquire include/linux/atomic/atomic-instrumented.h:4456 [inline]
| BUG: KASAN: slab-use-after-free in __mutex_trylock_fast kernel/locking/mutex.c:161 [inline]
| BUG: KASAN: slab-use-after-free in mutex_lock+0x4f/0xa0 kernel/locking/mutex.c:318
| Write of size 8 at addr ffff8881298d9550 by task kworker/2:1/83
|
| CPU: 2 UID: 0 PID: 83 Comm: kworker/2:1 Not tainted 7.1.0-rc6-next-20260601-dirty #6 PREEMPT(full)
| Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-debian-1.17.0-1 04/01/2014
| Workqueue: events l2cap_chan_timeout
| Call Trace:
|  
|  instrument_atomic_read_write include/linux/instrumented.h:112 [inline]
|  atomic_long_try_cmpxchg_acquire include/linux/atomic/atomic-instrumented.h:4456 [inline]
|  __mutex_trylock_fast kernel/locking/mutex.c:161 [inline]
|  mutex_lock+0x4f/0xa0 kernel/locking/mutex.c:318
|  l2cap_chan_timeout+0x5d/0x1b0 net/bluetooth/l2cap_core.c:422
|  process_one_work kernel/workqueue.c:3326 [inline]
|  process_scheduled_works+0x7c8/0xfb0 kernel/workqueue.c:3409
|  worker_thread+0x8a9/0xcf0 kernel/workqueue.c:3490
|  kthread+0x346/0x430 kernel/kthread.c:436
|  ret_from_fork+0x1a3/0x470 arch/x86/kernel/process.c:158
|  ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
|  
|
| Allocated by task 320:
|  l2cap_conn_add+0xa7/0x820 net/bluetooth/l2cap_core.c:7075
|  l2cap_connect_cfm+0xdb/0xd70 net/bluetooth/l2cap_core.c:7452
|  hci_connect_cfm include/net/bluetooth/hci_core.h:2139 [inline]
|  hci_remote_features_evt+0x52f/0x9f0 net/bluetooth/hci_event.c:3760
|  hci_event_func net/bluetooth/hci_event.c:7796 [inline]
|  hci_event_packet+0x561/0xa70 net/bluetooth/hci_event.c:7847
|  hci_rx_work+0x370/0x890 net/bluetooth/hci_core.c:4040
|  process_one_work kernel/workqueue.c:3326 [inline]
|  process_scheduled_works+0x7c8/0xfb0 kernel/workqueue.c:3409
|  worker_thread+0x8a9/0xcf0 kernel/workqueue.c:3490
|  kthread+0x346/0x430 kernel/kthread.c:436
|  ret_from_fork+0x1a3/0x470 arch/x86/kernel/process.c:158
|  ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
|
| Freed by task 322:
|  hci_disconn_cfm include/net/bluetooth/hci_core.h:2154 [inline]
|  hci_conn_hash_flush+0x101/0x1f0 net/bluetooth/hci_conn.c:2736
|  hci_dev_close_sync+0x889/0xde0 net/bluetooth/hci_sync.c:5405
|  hci_dev_do_close net/bluetooth/hci_core.c:502 [inline]
|  hci_unregister_dev+0x1f7/0x370 net/bluetooth/hci_core.c:2679
|  vhci_release+0x12a/0x180 drivers/bluetooth/hci_vhci.c:690
|  __fput+0x369/0x890 fs/file_table.c:510
|  task_work_run+0x160/0x1d0 kernel/task_work.c:233
|  get_signal+0xf5b/0x1120 kernel/signal.c:2810
|  arch_do_signal_or_restart+0x4d/0x600 arch/x86/kernel/signal.c:337
|  __exit_to_user_mode_loop kernel/entry/common.c:64 [inline]
|  exit_to_user_mode_loop+0x85/0x510 kernel/entry/common.c:98
|  do_syscall_64+0x263/0x3d0 arch/x86/entry/syscall_64.c:100
|  entry_SYSCALL_64_after_hwframe+0x77/0x7f
|
| The buggy address belongs to the object at ffff8881298d9400
|  which belongs to the cache kmalloc-512 of size 512
| The buggy address is located 336 bytes inside of
|  freed 512-byte region [ffff8881298d9400, ffff8881298d9600)

Fix it by having chan->conn hold a reference to l2cap_conn (via
l2cap_conn_get) when the channel is added to the connection, and
releasing it in the channel destructor. This ensures the l2cap_conn
remains alive as long as the channel exists.

A new FLAG_DEL channel flag is introduced to indicate that the channel
has been deleted from its connection. l2cap_chan_del() atomically sets
this flag using test_and_set_bit() instead of setting chan->conn to
NULL. All asynchronous workers (l2cap_chan_timeout, l2cap_ack_timeout,
l2cap_monitor_timeout, l2cap_retrans_timeout) and l2cap_chan_send()
check FLAG_DEL to determine whether the channel has been torn down,
rather than testing chan->conn for NULL.

Fixes: 8c8e620467a7 ("Bluetooth: L2CAP: use chan timer to close channels in cleanup_listen()")
Cc: 
Cc: Siwei Zhang 
Cc: Luiz Augusto von Dentz 
Assisted-by: Gemini:gemini-3.1-pro-preview
Reported-by: https://sashiko.dev/#/patchset/20260521021249.3258069-1-oss%40fourdim.xyz
Signed-off-by: Marco Elver 
Signed-off-by: Luiz Augusto von Dentz 
Signed-off-by: Greg Kroah-Hartman

tcp: restore RCU grace period in tcp_ao_destroy_sock

2026-07-18T14:53:13+00:00

commit 8bc4d43bccbd60efe85d0a44d5bf41762f2f0c30 upstream.

Commit 51e547e8c89c ("tcp: Free TCP-AO/TCP-MD5 info/keys without RCU")
removed the call_rcu() callback from tcp_ao_destroy_sock(), arguing that
"the destruction of info/keys is delayed until the socket destructor"
and therefore "no one can discover it anymore".

That argument does not hold for the call site in tcp_connect()
(net/ipv4/tcp_output.c:4327-4332). At that point the socket is in
TCP_SYN_SENT, has already been inserted into the inet ehash by
inet_hash_connect() in tcp_v4_connect(), and is therefore very much
discoverable: any softirq running tcp_v4_rcv() on another CPU can take
the socket out of the ehash, walk into tcp_inbound_hash(), and load
tp->ao_info via implicit RCU before bh_lock_sock_nested() is taken on
the destroying CPU.

The reader path then enters __tcp_ao_do_lookup() (net/ipv4/tcp_ao.c:208)
which re-loads tp->ao_info via rcu_dereference_check(); the re-load can
still observe the (about-to-be-freed) pointer because there is no
synchronize_rcu() between rcu_assign_pointer(tp->ao_info, NULL) and
tcp_ao_info_free() in tcp_ao_destroy_sock(). The captured pointer is
then walked at line 223:

	hlist_for_each_entry_rcu(key, &ao->head, node, ...)

The writer's synchronous kfree() is free to complete between the line
218 re-fetch and the line 223 hlist iteration. The slab is reused
(or simply LIST_POISON1-stamped if not yet reused) and the iteration
walks attacker-controlled or poison memory in softirq context.

Reproducer (no debug shim, stock x86_64 v7.1-rc2 SMP+KASAN, QEMU+KVM):
an unprivileged uid=1000 process inside CLONE_NEWUSER|CLONE_NEWNET
installs TCP_MD5SIG + TCP_AO_ADD_KEY on a TCP socket, sprays forged
TCP-AO segments toward its eventual 4-tuple via raw sockets, then
calls connect(). The md5-wins reconciliation in tcp_connect() fires
tcp_ao_destroy_sock(); the softirq backlog reader on the loopback
NAPI path crashes on the freed ao->head.first walk:

  Oops: general protection fault, probably for non-canonical
    address 0xfbd59c000000002f
  KASAN: maybe wild-memory-access in range
    [0xdead000000000178-0xdead00000000017f]
  CPU: 0 UID: 1000 PID: 100 Comm: repro_userns
  RIP: 0010:__tcp_ao_do_lookup+0x107/0x1c0
  Call Trace: 
    __tcp_ao_do_lookup+0x107/0x1c0
    tcp_ao_inbound_lookup.constprop.0+0x12a/0x200
    tcp_inbound_ao_hash+0x5ea/0x1520
    tcp_inbound_hash+0x7ce/0x1240
    tcp_v4_rcv+0x1e7a/0x3e10
    ...

Restore the RCU grace period: re-add struct rcu_head to tcp_ao_info
and replace the synchronous tcp_ao_info_free() with a call_rcu()
callback. Readers that captured tp->ao_info before rcu_assign_pointer
NULLed it now see the object remain valid until rcu_read_unlock().
With the patch applied the reproducer runs cleanly for 2000 iterations
on the same kernel build.

Fixes: 51e547e8c89c ("tcp: Free TCP-AO/TCP-MD5 info/keys without RCU")
Cc: stable@vger.kernel.org # v6.18+
Reviewed-by: Dmitry Safonov 
Signed-off-by: Michael Bommarito 
Reviewed-by: Eric Dumazet 
Signed-off-by: Dmitry Safonov <0x7f454c46@gmail.com>
Link: https://patch.msgid.link/20260625-tcp-md5-connect-v3-1-1fd313d6c1e0@gmail.com
Signed-off-by: Jakub Kicinski 
Signed-off-by: Greg Kroah-Hartman

net: ip_gre: require CAP_NET_ADMIN in the device netns for changelink

2026-07-04T11:44:17+00:00

commit 8165f7ff57d9667d2bb477ef6af83ede7fed4ad7 upstream.

A tunnel changelink() operates on at most two netns, dev_net(dev) and
the tunnel link netns t->net. They differ once the device is created in
or moved to a netns other than the one the request runs in. The rtnl
changelink path checks CAP_NET_ADMIN only against dev_net(dev), so a
caller privileged there but not in t->net can rewrite a tunnel that
lives in t->net.

Add rtnl_dev_link_net_capable() next to rtnl_get_net_ns_capable() in
net/core/rtnetlink.c. It requires CAP_NET_ADMIN in the link netns and is
skipped when the link netns is dev_net(dev), where the rtnl path already
checked it. The other patches in this series use the same helper.

Gate ipgre_changelink() and erspan_changelink() with it, at the top of
the op before any attribute is parsed, because the parsers update live
tunnel fields first. ipgre_netlink_parms() sets t->collect_md before
ip_tunnel_changelink() runs.

Commit 8b484efd5cb4 ("ip6: vti: Use ip6_tnl.net in
vti6_siocdevprivate().") added the same check on the ioctl path. This
adds it on RTM_NEWLINK.

Reported-by: Xiao Liang 
Closes: https://lore.kernel.org/netdev/CABAhCOSzP1vaThGV35_VnsRCb=87_CPjPVsTHbq905k8A+BuUg@mail.gmail.com/
Fixes: b57708add314 ("gre: add x-netns support")
Cc: stable@vger.kernel.org
Signed-off-by: Maoyi Xie 
Reviewed-by: Kuniyuki Iwashima 
Link: https://patch.msgid.link/20260612085941.3158249-2-maoyixie.tju@gmail.com
Signed-off-by: Jakub Kicinski 
Signed-off-by: Greg Kroah-Hartman

rose: cancel neighbour timers in rose_neigh_put() before freeing

2026-06-27T10:06:48+00:00

commit 9b222cb1d23ff210975e9df5ebab7b011acb6fad upstream.

rose_neigh_put() kfree()s the neighbour but never cancels its ftimer and
t0timer. Until now every caller that dropped the final reference first
called rose_remove_neigh(), which deletes those timers. The socket
heartbeat reaping path drops the last reference directly, so a neighbour
could be freed with t0timer still armed -- it re-arms itself in
rose_t0timer_expiry() -- leading to a use-after-free write in
enqueue_timer().

Cancel both timers with timer_delete_sync() (the synchronous variant, to
wait out a concurrently running, self-rearming handler) in the
refcount-zero branch of rose_neigh_put().

Signed-off-by: Bernard Pidoux 
Signed-off-by: Greg Kroah-Hartman

Bluetooth: L2CAP: reject BR/EDR signaling packets over MTUsig

2026-06-19T11:44:02+00:00

commit dd214733544427587a95f66dbf3adff072568990 upstream.

net/bluetooth/l2cap_core.c:l2cap_sig_channel() accepts BR/EDR
signaling packets up to the channel MTU and dispatches each command
without enforcing the signaling MTU (MTUsig). A Bluetooth BR/EDR peer
within radio range can send a fixed-channel CID 0x0001 packet that is
larger than MTUsig and contains many L2CAP_ECHO_REQ commands before
pairing. In a real-radio stock-kernel run, one 681-byte signaling
packet containing 168 zero-length ECHO_REQ commands made the target
transmit 168 ECHO_RSP frames over about 220 ms.

Impact: a Bluetooth BR/EDR peer within radio range, before pairing, can
force 168 ECHO_RSP frames from one 681-byte fixed-channel signaling
packet containing packed ECHO_REQ commands.

Define Linux's BR/EDR signaling MTU as the spec minimum of 48 bytes and
reject any larger signaling packet with one L2CAP_COMMAND_REJECT_RSP
carrying L2CAP_REJ_MTU_EXCEEDED before any command is dispatched.

The Bluetooth Core spec wording for MTUExceeded says the reject
identifier shall match the first request command in the packet, and
that packets containing only responses shall be silently discarded.
Linux intentionally deviates from that prescription: silently
discarding desynchronizes the peer because the remote stack never
learns its responses were dropped, and locating the first request
command requires walking command headers past MTUsig, i.e. processing
bytes from a packet we have already decided is too large to process.
We therefore always emit one reject and use the identifier from the
first command header, a single fixed-offset byte read.

The unrestricted BR/EDR signaling parser and ECHO_REQ response path both
trace to the initial git import; no later introducing commit is
available for a Fixes tag.

Cc: stable@vger.kernel.org
Suggested-by: Luiz Augusto von Dentz 
Link: https://lore.kernel.org/r/20260518002800.1361430-1-michael.bommarito@gmail.com
Link: https://lore.kernel.org/r/20260520135034.1060859-1-michael.bommarito@gmail.com
Link: https://lore.kernel.org/r/20260521000555.3712030-1-michael.bommarito@gmail.com
Assisted-by: Claude:claude-opus-4-7
Assisted-by: Codex:gpt-5-5-xhigh
Signed-off-by: Michael Bommarito 
Signed-off-by: Luiz Augusto von Dentz 
Signed-off-by: Greg Kroah-Hartman

netfilter: nf_conntrack: destroy stale expectfn expectations on unregister

2026-06-19T11:44:00+00:00

[ Upstream commit c3009418f9fa1dcb3eb86f4d8c92583537b5faa3 ]

NAT helpers such as nf_nat_h323 store a raw pointer to module text in
exp->expectfn (e.g. ip_nat_q931_expect). nf_ct_helper_expectfn_unregister()
only unlinks the callback descriptor and never walks the expectation table,
so an expectation pending at module removal survives with a dangling
exp->expectfn into freed module text.

When the expected connection arrives, init_conntrack() invokes
exp->expectfn(), now a stale pointer into the unloaded module. Reproduced
on a KASAN build by loading the H.323 helpers, creating a Q.931
expectation, unloading nf_nat_h323, then connecting to the expected port:

 Oops: int3: 0000 [#1] SMP KASAN NOPTI
 RIP: 0010:0xffffffffa06102d1
  init_conntrack.isra.0 (net/netfilter/nf_conntrack_core.c:1862)
  nf_conntrack_in (net/netfilter/nf_conntrack_core.c:2049)
  ipv4_conntrack_local (net/netfilter/nf_conntrack_proto.c:223)
  nf_hook_slow (net/netfilter/core.c:619)
  __ip_local_out (net/ipv4/ip_output.c:120)
  __tcp_transmit_skb (net/ipv4/tcp_output.c:1715)
  tcp_connect (net/ipv4/tcp_output.c:4374)
  tcp_v4_connect (net/ipv4/tcp_ipv4.c:345)
  __sys_connect (net/socket.c:2167)
 Modules linked in: nf_conntrack_h323 [last unloaded: nf_nat_h323]

Reaching the dangling state requires CAP_SYS_MODULE in the initial user
namespace to remove a NAT helper that still has live expectations, so this
is a robustness fix; leaving an expectation pointing at freed text is wrong
regardless.

Add nf_ct_helper_expectfn_destroy(), which walks the expectation table and
drops every expectation whose ->expectfn matches the descriptor being torn
down. Call it from each NAT helper's exit path after the existing RCU grace
period, so no expectation outlives the code it points at and no extra
synchronize_rcu() is introduced. With the fix, the same reproducer runs to
completion without the Oops.

Fixes: f587de0e2feb ("[NETFILTER]: nf_conntrack/nf_nat: add H.323 helper port")
Reported-by: Xiang Mei 
Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Weiming Shi 
Signed-off-by: Pablo Neira Ayuso 
Signed-off-by: Sasha Levin

net: guard timestamp cmsgs to real error queue skbs

2026-06-19T11:43:59+00:00

[ Upstream commit 1ee90b77b727df903033db873c75caac5c27ec98 ]

skb_is_err_queue() treats PACKET_OUTGOING as the sole marker for an skb
from sk_error_queue. That assumption is not true for AF_PACKET sockets:
outgoing packet taps are also delivered to packet sockets with
skb->pkt_type == PACKET_OUTGOING, but their skb->cb is owned by AF_PACKET
instead of struct sock_exterr_skb.

If such an skb is received with timestamping enabled, the generic
timestamp cmsg path can read AF_PACKET control-buffer state as
sock_exterr_skb::opt_stats. With SO_RXQ_OVFL enabled, the packet drop
counter overlaps opt_stats. An odd drop count makes the path emit
SCM_TIMESTAMPING_OPT_STATS with skb->len and skb->data. For non-linear
skbs this copies past the linear head and can trigger hardened usercopy or
disclose adjacent heap contents.

Keep skb_is_err_queue() local to net/socket.c, but make it verify that
the PACKET_OUTGOING marker is paired with the sock_rmem_free destructor
installed by sock_queue_err_skb(). AF_PACKET receive skbs use normal
receive ownership and no longer pass as error-queue skbs, while legitimate
sk_error_queue entries keep the PACKET_OUTGOING marker and sock_rmem_free
ownership.

Fixes: 8605330aac5a ("tcp: fix SCM_TIMESTAMPING_OPT_STATS for normal skbs")
Signed-off-by: Kyle Zeng 
Reviewed-by: Kuniyuki Iwashima 
Reviewed-by: Willem de Bruijn 
Link: https://patch.msgid.link/20260607021819.49698-1-kylebot@openai.com
Signed-off-by: Jakub Kicinski 
Signed-off-by: Sasha Levin

net/sched: fix pedit partial COW leading to page cache corruption

2026-06-19T11:43:55+00:00

[ Upstream commit 899ee91156e57784090c5565e4f31bd7dbffbc5a ]

tcf_pedit_act() computes the COW range for skb_ensure_writable()
once before the key loop using tcfp_off_max_hint, but the hint does
not account for the runtime header offset added by typed keys. This
can leave part of the write region un-COW'd.

Fix by moving skb_ensure_writable() inside the per-key loop where
the actual write offset is known, and add overflow checking on the
offset arithmetic. For negative offsets (e.g. Ethernet header edits
at ingress), use skb_cow() to COW the headroom instead. Guard
offset_valid() against INT_MIN, where negation is undefined.

Fixes: 8b796475fd78 ("net/sched: act_pedit: really ensure the skb is writable")
Reported-by: Yiming Qian 
Reported-by: Keenan Dong 
Reported-by: Han Guidong <2045gemini@gmail.com>
Reported-by: Zhang Cen 
Reviewed-by: Han Guidong <2045gemini@gmail.com>
Tested-by: Han Guidong <2045gemini@gmail.com>
Reviewed-by: Davide Caratti 
Tested-by: Davide Caratti 
Reviewed-by: Toke Høiland-Jørgensen 
Tested-by: Toke Høiland-Jørgensen 
Reviewed-by: Victor Nogueira 
Tested-by: Victor Nogueira 
Acked-by: Jamal Hadi Salim 
Signed-off-by: Rajat Gupta 
Link: https://patch.msgid.link/20260531123221.48732-1-jhs@mojatatu.com
Signed-off-by: Jakub Kicinski 
Signed-off-by: Sasha Levin

net/sched: act_api: use RCU with deferred freeing for action lifecycle

2026-06-19T11:43:52+00:00

[ Upstream commit 5057e1aca011e51ef51498c940ef96f3d3e8a305 ]

When NEWTFILTER and DELFILTER are run concurrently it is possible to create a
race with an associated action.

Let's illustrate with CPU0 running NEWTFILTER and CPU1 running DELFILTER:

 0: mutex_lock() <-- holds the idr lock
 0: rcu_read_lock()
 0: p = idr_find(idr, index) <-- action p is valid (RCU protects IDR)
 0: mutex_unlock() <-- releases the idr lock
 1: refcount_dec_and_mutex_lock() <-- refcnt 1->0, mutex held
 1: idr_remove(idr, index) <-- Action removed from IDR
 1: mutex_unlock() <-- mutex released allowing us to delete the action
 1: tcf_action_cleanup(p); kfree(p) <-- Kfrees p immediately, no deferral
 0: refcount_inc_not_zero(&p->tcfa_refcnt) <-- ouch, UAF p points to freed memory

This patch fixes the race condition between NEWTFILTER and DELFILTER by
adding struct rcu_head to tc_action used in the deferral and introducing a
call_rcu() in the delete path to defer the final kfree().

Note: this is a revert of commit d7fb60b9cafb ("net_sched: get rid of tcfa_rcu")
but also modernization/simplification to directly use kfree_rcu().

Let's illustrate the new restored code path:

 0: rcu_read_lock()
 1: refcount_dec_and_mutex_lock() <-- refcnt 1->0, mutex held
 1: idr_remove(idr, index)
 1: mutex_unlock()
 1: call_rcu(&p->tcfa_rcu, tcf_action_rcu_free) <-- defer kfree after grace period
 0: p = idr_find(idr, index)
 0: refcount_inc_not_zero(&p->tcfa_refcnt) <-- fails, refcnt already 0
 1: rcu_read_unlock() <-- release so freeing can run after grace period

After CPU1 calls idr_remove(), the object is no longer reachable through the IDR.
CPU0's subsequent idr_find() will return NULL, and even if it still held a
stale pointer, the immediate kfree() is now deferred until after the RCU grace
period, so no UAF can occur.

Fixes: d7fb60b9cafb ("net_sched: get rid of tcfa_rcu")
Suggested-by: Jakub Kicinski 
Reported-by: Kyle Zeng 
Tested-by: Victor Nogueira 
Tested-by: syzbot@syzkaller.appspotmail.com
Signed-off-by: Jamal Hadi Salim 
Tested-by: Kyle Zeng 
Reviewed-by: Pedro Tammela 
Reviewed-by: Eric Dumazet 
Reviewed-by: Victor Nogueira 
Link: https://patch.msgid.link/20260531160812.68020-1-jhs@mojatatu.com
Signed-off-by: Jakub Kicinski 
Signed-off-by: Sasha Levin