[PATCH v3 net] net: core: dev: add reprocess depth limit for another_round in __netif_receive_skb_core

[PATCH v3 net] net: core: dev: add reprocess depth limit for another_round in __netif_receive_skb_core

[Yizhou Zhao]

In __netif_receive_skb_core(), the another_round label can be reached 
via a TC ingress redirect (bpf_redirect_peer returning -EAGAIN).

Across network namespaces, two BPF programs on peer devices can redirect
packets back and forth indefinitely, creating an unbounded loop that 
monopolizes a CPU core in softirq context. This leads to RCU stalls, 
soft lockups, and system-wide denial of service.

We reproduced it by creating a pair of TC BPF programs across two 
network namespaces that redirect packets to each other, and the RCU 
subsystem detects a stall:

```
[   24.835219] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
[   24.835837] rcu: 	(detected by 0, t=21002 jiffies, g=-627, q=2 ncpus=1)
[   24.835959] rcu: All QSes seen, last rcu_preempt kthread activity 21002 (4294691810-4294670808), jiffies_till_next_fqs=3, root ->qsmask 0x0
[   24.836239] rcu: rcu_preempt kthread starved for 21002 jiffies! g-627 f0x2 RCU_GP_WAIT_FQS(5) ->state=0x0 ->cpu=0
[   24.836362] rcu: 	Unless rcu_preempt kthread gets sufficient CPU time, OOM is now expected behavior.
[   24.836460] rcu: RCU grace-period kthread stack dump:
[   24.836601] task:rcu_preempt     state:R  running task     stack:15448 pid:15    tgid:15    ppid:2      task_flags:0x208040 flags:0x00080000
[   24.837139] Call Trace:
[   24.837568]  <TASK>
[   24.838008]  __schedule+0x4ed/0xea0
[   24.838934]  schedule+0x22/0xd0
[   24.839023]  schedule_timeout+0x81/0x100
[   24.839095]  ? __pfx_process_timeout+0x10/0x10
[   24.839165]  rcu_gp_fqs_loop+0x11b/0x650
[   24.839226]  ? __pfx_rcu_gp_kthread+0x10/0x10
[   24.839282]  rcu_gp_kthread+0x17e/0x210
[   24.839333]  ? __pfx_rcu_gp_kthread+0x10/0x10
[   24.839383]  kthread+0xdd/0x110
[   24.839433]  ? __pfx_kthread+0x10/0x10
[   24.839481]  ret_from_fork+0x1aa/0x260
[   24.839538]  ? __pfx_kthread+0x10/0x10
[   24.839585]  ret_from_fork_asm+0x1a/0x30
[   24.839686]  </TASK>
......
```

Fix this by adding a depth counter when it is about to go to another_round
label. When the counter exceeds XMIT_RECURSION_LIMIT (8), the packet is 
dropped. This follows the same pattern as dev_xmit_recursion() which 
protects the TX redirect path with the same limit.

Reuse SKB_DROP_REASON_TC_RECLASSIFY_LOOP for observability.

Fixes: 9aa1206e8f48 ("bpf: Add redirect_peer helper")
Cc: stable@vger.kernel.org
Reported-by: Yizhou Zhao <zhaoyz24@mails.tsinghua.edu.cn>
Reported-by: Yuxiang Yang <yangyx22@mails.tsinghua.edu.cn>
Reported-by: Xuewei Feng <fengxw06@126.com>
Reported-by: Qi Li <qli01@tsinghua.edu.cn>
Reported-by: Ke Xu <xuke@tsinghua.edu.cn>
Assisted-by: GLM:GLM-5.1
Signed-off-by: Yizhou Zhao <zhaoyz24@mails.tsinghua.edu.cn>
---
Changes in v3:
- Guard redirect_depth declaration with #ifdef CONFIG_NET_INGRESS to
  avoid unused variable warning when CONFIG_NET_INGRESS is not set
- Reorder variable declarations to follow reverse christmas tree style
- Link to v2: https://lore.kernel.org/netdev/20260512022127.7818-1-zhaoyz24@mails.tsinghua.edu.cn/
Changes in v2:
- Move the check just after `another` is set to true to avoid affecting the fast path
- Reuse SKB_DROP_REASON_TC_RECLASSIFY_LOOP to avoid adding new drop reason
- Link to v1: https://lore.kernel.org/netdev/20260511063005.38134-1-zhaoyz24@mails.tsinghua.edu.cn/
---
 net/core/dev.c | 12 ++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 831129f2a..c8e4a1d3f 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -5958,6 +5958,9 @@ static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
    struct net_device *orig_dev;
    bool deliver_exact = false;
+#ifdef CONFIG_NET_INGRESS
+   int redirect_depth = 0;
+#endif
    int ret = NET_RX_DROP;
    __be16 type;

@@ -6031,8 +6034,16 @@ static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
        nf_skip_egress(skb, true);
        skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev,
                     &another);
-       if (another)
+       if (another) {
+           if (unlikely(++redirect_depth > XMIT_RECURSION_LIMIT)) {
+               net_warn_ratelimited(
+                   "%s: redirect loop limit reached, dropping (dev=%s)\n",
+                   __func__, skb->dev->name);
+               drop_reason = SKB_DROP_REASON_TC_RECLASSIFY_LOOP;
+               goto drop;
+           }
            goto another_round;
+       }
        if (!skb)
            goto out;

--
2.43.0

Re: [PATCH v3 net] net: core: dev: add reprocess depth limit for another_round in __netif_receive_skb_core

[Jakub Kicinski]

On Thu, 14 May 2026 20:24:41 +0800 Yizhou Zhao wrote:
> In __netif_receive_skb_core(), the another_round label can be reached 
> via a TC ingress redirect (bpf_redirect_peer returning -EAGAIN).

Does not apply to netdev/net, please rebase+repost.
-- 
pw-bot: cr