netfilter: nf_tables: don't queue packet path object notifications

Summary

In the Linux kernel, the following vulnerability has been resolved: netfilter: nf_tables: don't queue packet path object notifications All file:line references below are against v7.2-rc4 (ac5b0e5651b1). The trace was captured on 7.2.0-rc6-kasan72rc6 (075b74841bd0), where the same lines apply. nft_obj_notify() is exported and reached from the packet path. Its only in-tree caller is nft_quota_obj_eval() (net/netfilter/nft_quota.c:68), which notifies with GFP_ATOMIC while evaluating a rule for a transiting packet, holding no mutex. Since commit 67cc570edaa0 ("netfilter: nf_tables: coalesce multiple notifications into one skbuff") that notification is no longer sent immediately. __nft_obj_notify() queues it onto nft_net->notify_list via nft_notify_enqueue() (net/netfilter/nf_tables_api.c:1211), which is a bare list_add_tail(). notify_list has no lock of its own (include/net/netfilter/nf_tables.h:1951), it is serialised by commit_mutex: the six other enqueue sites all run inside a netlink transaction, and the drain in nft_commit_notify() (net/netfilter/nf_tables_api.c:10746) does list_del() + kfree_skb() from nf_tables_commit() with commit_mutex held. Sending packets through a chain that references a depleted quota object therefore races an unlocked list_add_tail() against list_del() + kfree_skb() on another CPU. The WRITE_ONCE(prev->next, new) in __list_add() then stores through an sk_buff that has already been freed: BUG: KASAN: slab-use-after-free in __nft_obj_notify+0x2c5/0x2d0 Write of size 8 at addr ff110001047183c0 by task poc/76 CPU: 0 UID: 1000 PID: 76 Comm: poc Tainted: G W 7.2.0-rc6-kasan72rc6 #4 Call Trace: <IRQ> __nft_obj_notify (include/linux/list.h:164 include/linux/list.h:191 net/netfilter/nf_tables_api.c:1211 net/netfilter/nf_tables_api.c:8743) nft_quota_obj_eval (net/netfilter/nft_quota.c:68) nft_do_chain_inet nf_hook_slow __ip_local_out ip_push_pending_frames udp_send_skb udp_sendmsg __x64_sys_sendto Allocated by task 77: __alloc_skb (net/core/skbuff.c:704) __nft_obj_notify (include/net/netlink.h:1055 net/netfilter/nf_tables_api.c:8731) nft_quota_obj_eval (net/netfilter/nft_quota.c:68) nft_do_chain Freed by task 79: nf_tables_commit (include/linux/skbuff.h:1332 net/netfilter/nf_tables_api.c:10759 net/netfilter/nf_tables_api.c:11185) nfnetlink_rcv_batch (net/netfilter/nfnetlink.c:574) netlink_unicast netlink_sendmsg The buggy address belongs to the cache skbuff_head_cache of size 232 Queueing from the packet path is wrong even leaving the race aside: notify_list is only drained by nft_commit_notify() from nf_tables_commit() (:11185), so a notification enqueued outside a transaction is not sent until some later netlink batch commits, if one ever does. The gfp argument that nft_obj_notify() still takes is a leftover of the pre-67cc570edaa0 behaviour, where this path called nfnetlink_send() directly. Restore that: split the message construction out into nft_obj_notify_alloc() and let each caller decide what to do with the skb. nft_obj_notify(), the exported one reached from the packet path, sends it straight away; nf_tables_obj_notify(), which runs under commit_mutex, keeps queueing it, so transaction notifications are still coalesced.