Skip to contentSkip to conversation

udp: Use READ_ONCE()/WRITE_ONCE() for hslot->count.

PATCH In review Kuniyuki Iwashima 2 participants 5 messages Apr 23

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 udp: Use READ_ONCE()/WRITE_ONCE() for hslot->count.

__udp4_lib_mcast_demux_lookup() and __udp4_lib_mcast_deliver() reads the number of sockets in the 1-tuple hash table chain locklessly. [0]

Let's use READ_ONCE() and WRITE_ONCE() for hslot->count.

[0]:

BUG: KCSAN: data-race in __udp4_lib_mcast_deliver / udp_lib_unhash

read-write to 0xffff88810194a9c8 of 4 bytes by task 15107 on cpu 0:

 udp_lib_unhash+0x1f4/0x3e0 net/ipv4/udp.c:2261
 sk_common_release+0x7a/0x230 net/core/sock.c:3990
 udp_lib_close+0x15/0x20 include/net/udp.h:324
 inet_release+0xcd/0xf0 net/ipv4/af_inet.c:437
 __sock_release net/socket.c:662 [inline]
 sock_close+0x6b/0x150 net/socket.c:1455
 __fput+0x29b/0x650 fs/file_table.c:469
 ____fput+0x1c/0x30 fs/file_table.c:497
 task_work_run+0x130/0x1a0 kernel/task_work.c:233
 resume_user_mode_work include/linux/resume_user_mode.h:50 [inline]
 __exit_to_user_mode_loop kernel/entry/common.c:67 [inline]
 exit_to_user_mode_loop+0x1f4/0x6f0 kernel/entry/common.c:98
 __exit_to_user_mode_prepare include/linux/irq-entry-common.h:226 [inline]
 syscall_exit_to_user_mode_prepare include/linux/irq-entry-common.h:256 [inline]
 syscall_exit_to_user_mode include/linux/entry-common.h:325 [inline]
 do_syscall_64+0x249/0x370 arch/x86/entry/syscall_64.c:100
 entry_SYSCALL_64_after_hwframe+0x77/0x7f

read to 0xffff88810194a9c8 of 4 bytes by task 15111 on cpu 1:

 __udp4_lib_mcast_deliver+0x9a/0x790 net/ipv4/udp.c:2564
 __udp4_lib_rcv+0x972/0x1250 net/ipv4/udp.c:2749
 udp_rcv+0x4f/0x60 net/ipv4/udp.c:2936
 ip_protocol_deliver_rcu+0x3f7/0x790 net/ipv4/ip_input.c:207
 ip_local_deliver_finish+0x1fc/0x2f0 net/ipv4/ip_input.c:241
 NF_HOOK include/linux/netfilter.h:318 [inline]
 ip_local_deliver+0xe8/0x1e0 net/ipv4/ip_input.c:262
 dst_input include/net/dst.h:480 [inline]
 ip_sublist_rcv_finish net/ipv4/ip_input.c:584 [inline]
 ip_list_rcv_finish net/ipv4/ip_input.c:628 [inline]
 ip_sublist_rcv+0x415/0x6a0 net/ipv4/ip_input.c:644
 ip_list_rcv+0x261/0x290 net/ipv4/ip_input.c:678
 __netif_receive_skb_list_ptype net/core/dev.c:6207 [inline]
 __netif_receive_skb_list_core+0x4dc/0x500 net/core/dev.c:6254
 __netif_receive_skb_list net/core/dev.c:6306 [inline]
 netif_receive_skb_list_internal+0x47d/0x5f0 net/core/dev.c:6397
 netif_receive_skb_list+0x35/0x220 net/core/dev.c:6449
 xdp_recv_frames net/bpf/test_run.c:269 [inline]
 xdp_test_run_batch net/bpf/test_run.c:350 [inline]
 bpf_test_run_xdp_live+0x104c/0x1360 net/bpf/test_run.c:379
 bpf_prog_test_run_xdp+0x57b/0xa10 net/bpf/test_run.c:1397
 bpf_prog_test_run+0x204/0x340 kernel/bpf/syscall.c:4722
 __sys_bpf+0x52e/0x7e0 kernel/bpf/syscall.c:6246
 __do_sys_bpf kernel/bpf/syscall.c:6341 [inline]
 __se_sys_bpf kernel/bpf/syscall.c:6339 [inline]
 __x64_sys_bpf+0x41/0x50 kernel/bpf/syscall.c:6339
 x64_sys_call+0x10cb/0x3020 arch/x86/include/generated/asm/syscalls_64.h:322
 do_syscall_x64 arch/x86/entry/syscall_64.c:63 [inline]
 do_syscall_64+0x12c/0x370 arch/x86/entry/syscall_64.c:94
 entry_SYSCALL_64_after_hwframe+0x77/0x7f

value changed: 0x00000001 -> 0x00000000

Reported by Kernel Concurrency Sanitizer on:

CPU: 1 UID: 0 PID: 15111 Comm: syz.6.4060 Not tainted syzkaller #0 PREEMPT(full)
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 02/12/2026
Fixes 2dc41cff7545 ("udp: Use hash2 for long hash1 chains in __udp*_lib_mcast_deliver.")
Fixes 63c6f81cdde5 ("udp: ipv4: do not waste time in __udp4_lib_mcast_demux_lookup")
Reported-by [email protected]
---
 net/ipv4/udp.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
net/ipv4/udp.c +4 -4
index 0ac2bf4f8759..f7e6045f5985 100644
@@ -323,7 +323,7 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum,
 		sock_set_flag(sk, SOCK_RCU_FREE);
 
 		sk_add_node_rcu(sk, &hslot->head);
-		hslot->count++;
+		WRITE_ONCE(hslot->count, hslot->count + 1);
 		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 
 		hslot2 = udp_hashslot2(udptable, udp_sk(sk)->udp_portaddr_hash);
@@ -2200,7 +2200,7 @@ void udp_lib_unhash(struct sock *sk)
 		if (rcu_access_pointer(sk->sk_reuseport_cb))
 			reuseport_detach_sock(sk);
 		if (sk_del_node_init_rcu(sk)) {
-			hslot->count--;
+			WRITE_ONCE(hslot->count, hslot->count - 1);
 			inet_sk(sk)->inet_num = 0;
 			sock_prot_inuse_add(net, sk->sk_prot, -1);
 
@@ -2471,7 +2471,7 @@ static int __udp4_lib_mcast_deliver(struct net *net, struct sk_buff *skb,
 	hash2_any = 0;
 	hash2 = 0;
 	hslot = udp_hashslot(udptable, net, hnum);
-	use_hash2 = hslot->count > 10;
+	use_hash2 = READ_ONCE(hslot->count) > 10;
 	offset = offsetof(typeof(*sk), sk_node);
 
 	if (use_hash2) {
@@ -2703,7 +2703,7 @@ static struct sock *__udp4_lib_mcast_demux_lookup(struct net *net,
 	hslot = &udptable->hash[slot];
 
 	/* Do not bother scanning a too big list */
-	if (hslot->count > 10)
+	if (READ_ONCE(hslot->count) > 10)
 		return NULL;
 
 	result = NULL;
Signature · 1 line
2.54.0.rc2.533.g4f5dca5207-goog
#1 lore raw
Quoted 9 lines · Kuniyuki Iwashima

__udp4_lib_mcast_demux_lookup() and __udp4_lib_mcast_deliver() reads the number of sockets in the 1-tuple hash table chain locklessly. [0]

Let's use READ_ONCE() and WRITE_ONCE() for hslot->count.

[0]:

BUG: KCSAN: data-race in __udp4_lib_mcast_deliver / udp_lib_unhash

...

Quoted 11 lines

Reported by Kernel Concurrency Sanitizer on:

CPU: 1 UID: 0 PID: 15111 Comm: syz.6.4060 Not tainted syzkaller #0 PREEMPT(full)
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 02/12/2026
Fixes 2dc41cff7545 ("udp: Use hash2 for long hash1 chains in __udp*_lib_mcast_deliver.")
Fixes 63c6f81cdde5 ("udp: ipv4: do not waste time in __udp4_lib_mcast_demux_lookup")
Reported-by [email protected]

---

You forgot to change net/ipv6/udp.c

net/ipv6/udp.c +0 -0
index 15e032194eccc3c29b3e5f3a09214cad60623329..1f4dabe4c350118441f61c1cf0398e698581542c

100644

net/ipv6/udp.c +0 -0
@@ -953,7 +953,7 @@ static int __udp6_lib_mcast_deliver(struct net

*net, struct sk_buff *skb,

        hash2_any = 0;
        hash2 = 0;
        hslot = udp_hashslot(udptable, net, hnum);
-       use_hash2 = hslot->count > 10;
+       use_hash2 = READ_ONCE(hslot->count) > 10;
        offset = offsetof(typeof(*sk), sk_node);

        if (use_hash2) {

Also one part is missing in udp_lib_get_port()

@@ -289,7 +289,7 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum,

                        hash2_nulladdr &= udptable->mask;

                        hslot2 = udp_hashslot2(udptable, slot2);
-                       if (hslot->count < hslot2->count)
+                       if (hslot->count < READ_ONCE(hslot2->count))
                                goto scan_primary_hash;

                        exist = udp_lib_lport_inuse2(net, snum, hslot2, sk);
#2 lore raw
Quoted 19 lines · Eric Dumazet
Quoted 9 lines · Kuniyuki Iwashima

__udp4_lib_mcast_demux_lookup() and __udp4_lib_mcast_deliver() reads the number of sockets in the 1-tuple hash table chain locklessly. [0]

Let's use READ_ONCE() and WRITE_ONCE() for hslot->count.

[0]:

BUG: KCSAN: data-race in __udp4_lib_mcast_deliver / udp_lib_unhash

...

Quoted 5 lines

Reported by Kernel Concurrency Sanitizer on:

CPU: 1 UID: 0 PID: 15111 Comm: syz.6.4060 Not tainted syzkaller #0 PREEMPT(full)
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 02/12/2026
Quoted 8 lines
Quoted 6 lines
Fixes 2dc41cff7545 ("udp: Use hash2 for long hash1 chains in __udp*_lib_mcast_deliver.")
Fixes 63c6f81cdde5 ("udp: ipv4: do not waste time in __udp4_lib_mcast_demux_lookup")
Reported-by [email protected]

---

You forgot to change net/ipv6/udp.c

Ah exactly. Will add this in v2.

Quoted 18 lines
net/ipv6/udp.c +0 -0
index 15e032194eccc3c29b3e5f3a09214cad60623329..1f4dabe4c350118441f61c1cf0398e698581542c

100644

net/ipv6/udp.c +1 -1
@@ -953,7 +953,7 @@ static int __udp6_lib_mcast_deliver(struct net
 *net, struct sk_buff *skb,
        hash2_any = 0;
        hash2 = 0;
        hslot = udp_hashslot(udptable, net, hnum);
-       use_hash2 = hslot->count > 10;
+       use_hash2 = READ_ONCE(hslot->count) > 10;
        offset = offsetof(typeof(*sk), sk_node);
 
        if (use_hash2) {
Quoted 1 line

Also one part is missing in udp_lib_get_port()

I excluded this part since it's under spin_lock_bh(&hslot->lock).

Quoted 10 lines
net/ipv6/udp.c +1 -1
@@ -289,7 +289,7 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum,
                        hash2_nulladdr &= udptable->mask;
 
                        hslot2 = udp_hashslot2(udptable, slot2);
-                       if (hslot->count < hslot2->count)
+                       if (hslot->count < READ_ONCE(hslot2->count))
                                goto scan_primary_hash;
 
                        exist = udp_lib_lport_inuse2(net, snum, hslot2, sk);
#3 lore raw
Quoted 47 lines · Kuniyuki Iwashima
Quoted 19 lines · Eric Dumazet
Quoted 9 lines · Kuniyuki Iwashima

__udp4_lib_mcast_demux_lookup() and __udp4_lib_mcast_deliver() reads the number of sockets in the 1-tuple hash table chain locklessly. [0]

Let's use READ_ONCE() and WRITE_ONCE() for hslot->count.

[0]:

BUG: KCSAN: data-race in __udp4_lib_mcast_deliver / udp_lib_unhash

...

Quoted 5 lines

Reported by Kernel Concurrency Sanitizer on:

CPU: 1 UID: 0 PID: 15111 Comm: syz.6.4060 Not tainted syzkaller #0 PREEMPT(full)
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 02/12/2026
Quoted 8 lines
Quoted 6 lines
Fixes 2dc41cff7545 ("udp: Use hash2 for long hash1 chains in __udp*_lib_mcast_deliver.")
Fixes 63c6f81cdde5 ("udp: ipv4: do not waste time in __udp4_lib_mcast_demux_lookup")
Reported-by [email protected]

---

You forgot to change net/ipv6/udp.c

Ah exactly. Will add this in v2.

Quoted 15 lines
net/ipv6/udp.c +0 -0
index 15e032194eccc3c29b3e5f3a09214cad60623329..1f4dabe4c350118441f61c1cf0398e698581542c

100644

net/ipv6/udp.c +1 -1
@@ -953,7 +953,7 @@ static int __udp6_lib_mcast_deliver(struct net
 *net, struct sk_buff *skb,
        hash2_any = 0;
        hash2 = 0;
        hslot = udp_hashslot(udptable, net, hnum);
-       use_hash2 = hslot->count > 10;
+       use_hash2 = READ_ONCE(hslot->count) > 10;
        offset = offsetof(typeof(*sk), sk_node);
 
Quoted 7 lines
Quoted 4 lines
        if (use_hash2) {

Also one part is missing in udp_lib_get_port()

I excluded this part since it's under spin_lock_bh(&hslot->lock).

But you would need a spin_lock(hslot2->lock), wich we do not hold.

Quoted 10 lines
Quoted 10 lines
net/ipv6/udp.c +1 -1
@@ -289,7 +289,7 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum,
                        hash2_nulladdr &= udptable->mask;
 
                        hslot2 = udp_hashslot2(udptable, slot2);
-                       if (hslot->count < hslot2->count)
+                       if (hslot->count < READ_ONCE(hslot2->count))
                                goto scan_primary_hash;
 
                        exist = udp_lib_lport_inuse2(net, snum, hslot2, sk);
#4 lore raw
Quoted 52 lines · Eric Dumazet
Quoted 50 lines · Kuniyuki Iwashima
Quoted 19 lines · Eric Dumazet
Quoted 9 lines · Kuniyuki Iwashima

__udp4_lib_mcast_demux_lookup() and __udp4_lib_mcast_deliver() reads the number of sockets in the 1-tuple hash table chain locklessly. [0]

Let's use READ_ONCE() and WRITE_ONCE() for hslot->count.

[0]:

BUG: KCSAN: data-race in __udp4_lib_mcast_deliver / udp_lib_unhash

...

Quoted 5 lines

Reported by Kernel Concurrency Sanitizer on:

CPU: 1 UID: 0 PID: 15111 Comm: syz.6.4060 Not tainted syzkaller #0 PREEMPT(full)
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 02/12/2026
Quoted 8 lines
Quoted 6 lines
Fixes 2dc41cff7545 ("udp: Use hash2 for long hash1 chains in __udp*_lib_mcast_deliver.")
Fixes 63c6f81cdde5 ("udp: ipv4: do not waste time in __udp4_lib_mcast_demux_lookup")
Reported-by [email protected]

---

You forgot to change net/ipv6/udp.c

Ah exactly. Will add this in v2.

Quoted 18 lines
net/ipv6/udp.c +0 -0
index 15e032194eccc3c29b3e5f3a09214cad60623329..1f4dabe4c350118441f61c1cf0398e698581542c

100644

net/ipv6/udp.c +1 -1
@@ -953,7 +953,7 @@ static int __udp6_lib_mcast_deliver(struct net
 *net, struct sk_buff *skb,
        hash2_any = 0;
        hash2 = 0;
        hslot = udp_hashslot(udptable, net, hnum);
-       use_hash2 = hslot->count > 10;
+       use_hash2 = READ_ONCE(hslot->count) > 10;
        offset = offsetof(typeof(*sk), sk_node);
 
        if (use_hash2) {
Quoted 6 lines
Quoted 4 lines
Quoted 1 line

Also one part is missing in udp_lib_get_port()

I excluded this part since it's under spin_lock_bh(&hslot->lock).

But you would need a spin_lock(hslot2->lock), wich we do not hold.

Ah, it's hslot2->count ! I'll cover that too, thank you !

Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
net/ipv6/udp.c +1 -1
@@ -289,7 +289,7 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum,
                        hash2_nulladdr &= udptable->mask;
 
                        hslot2 = udp_hashslot2(udptable, slot2);
-                       if (hslot->count < hslot2->count)
+                       if (hslot->count < READ_ONCE(hslot2->count))
                                goto scan_primary_hash;
 
                        exist = udp_lib_lport_inuse2(net, snum, hslot2, sk);