CVE-2026-23450

*** Pendiente de traducción *** In the Linux kernel, the following vulnerability has been resolved:<br /> <br /> net/smc: fix NULL dereference and UAF in smc_tcp_syn_recv_sock()<br /> <br /> Syzkaller reported a panic in smc_tcp_syn_recv_sock() [1].<br /> <br /> smc_tcp_syn_recv_sock() is called in the TCP receive path<br /> (softirq) via icsk_af_ops-&gt;syn_recv_sock on the clcsock (TCP<br /> listening socket). It reads sk_user_data to get the smc_sock<br /> pointer. However, when the SMC listen socket is being closed<br /> concurrently, smc_close_active() sets clcsock-&gt;sk_user_data<br /> to NULL under sk_callback_lock, and then the smc_sock itself<br /> can be freed via sock_put() in smc_release().<br /> <br /> This leads to two issues:<br /> <br /> 1) NULL pointer dereference: sk_user_data is NULL when<br /> accessed.<br /> 2) Use-after-free: sk_user_data is read as non-NULL, but the<br /> smc_sock is freed before its fields (e.g., queued_smc_hs,<br /> ori_af_ops) are accessed.<br /> <br /> The race window looks like this (the syzkaller crash [1]<br /> triggers via the SYN cookie path: tcp_get_cookie_sock() -&gt;<br /> smc_tcp_syn_recv_sock(), but the normal tcp_check_req() path<br /> has the same race):<br /> <br /> CPU A (softirq) CPU B (process ctx)<br /> <br /> tcp_v4_rcv()<br /> TCP_NEW_SYN_RECV:<br /> sk = req-&gt;rsk_listener<br /> sock_hold(sk)<br /> /* No lock on listener */<br /> smc_close_active():<br /> write_lock_bh(cb_lock)<br /> sk_user_data = NULL<br /> write_unlock_bh(cb_lock)<br /> ...<br /> smc_clcsock_release()<br /> sock_put(smc-&gt;sk) x2<br /> -&gt; smc_sock freed!<br /> tcp_check_req()<br /> smc_tcp_syn_recv_sock():<br /> smc = user_data(sk)<br /> -&gt; NULL or dangling<br /> smc-&gt;queued_smc_hs<br /> -&gt; crash!<br /> <br /> Note that the clcsock and smc_sock are two independent objects<br /> with separate refcounts. TCP stack holds a reference on the<br /> clcsock, which keeps it alive, but this does NOT prevent the<br /> smc_sock from being freed.<br /> <br /> Fix this by using RCU and refcount_inc_not_zero() to safely<br /> access smc_sock. Since smc_tcp_syn_recv_sock() is called in<br /> the TCP three-way handshake path, taking read_lock_bh on<br /> sk_callback_lock is too heavy and would not survive a SYN<br /> flood attack. Using rcu_read_lock() is much more lightweight.<br /> <br /> - Set SOCK_RCU_FREE on the SMC listen socket so that<br /> smc_sock freeing is deferred until after the RCU grace<br /> period. This guarantees the memory is still valid when<br /> accessed inside rcu_read_lock().<br /> - Use rcu_read_lock() to protect reading sk_user_data.<br /> - Use refcount_inc_not_zero(&amp;smc-&gt;sk.sk_refcnt) to pin the<br /> smc_sock. If the refcount has already reached zero (close<br /> path completed), it returns false and we bail out safely.<br /> <br /> Note: smc_hs_congested() has a similar lockless read of<br /> sk_user_data without rcu_read_lock(), but it only checks for<br /> NULL and accesses the global smc_hs_wq, never dereferencing<br /> any smc_sock field, so it is not affected.<br /> <br /> Reproducer was verified with mdelay injection and smc_run,<br /> the issue no longer occurs with this patch applied.<br /> <br /> [1] https://syzkaller.appspot.com/bug?extid=827ae2bfb3a3529333e9