Skip to contentSkip to conversation

bpf: reuseport: add cond_resched_rcu() in reuseport_array_free()

PATCH In review Zijing Yin 2 participants 2 messages Apr 10 BPF

lore
Share & tools
lore

Zijing Yin

author
#0 lore raw

PATCH 1/1 bpf: reuseport: add cond_resched_rcu() in reuseport_array_free()

reuseport_array_free() iterates over all map entries inside rcu_read_lock() to detach sockets from the array. When max_entries is very large (e.g., hundreds of millions), this loop runs for an extended period without yielding the CPU, triggering RCU stall warnings in the kworker thread that executes bpf_map_free_deferred().

The observed stall occurs because the loop has no scheduling point:

  rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
  Workqueue: events_unbound bpf_map_free_deferred
  Call Trace:
   reuseport_array_free+0x1ec/0x470 kernel/bpf/reuseport_array.c:127
   bpf_map_free_deferred+0x34a/0x7e0 kernel/bpf/syscall.c:893
   process_one_work+0x952/0x1a80
   worker_thread+0x87b/0x11f0

Add cond_resched_rcu() in the loop body to allow the scheduler to run and RCU grace periods to complete. This is safe because each iteration processes a single entry independently, sk->sk_callback_lock is not held across the yield point, and the map is fully detached from userspace so no concurrent insertions can occur.

This follows an established pattern for long-running kernel loops that must run under rcu_read_lock(). The closest precedent is in another BPF map free function:

  kernel/bpf/hashtab.c:1600
    htab_free_malloced_internal_structs()
      rcu_read_lock();
      for (i = 0; i < htab->n_buckets; i++) {
          ... walk bucket ...
          cond_resched_rcu();
      }
      rcu_read_unlock();
Fixes 5dc4c4b7d4e8 ("bpf: Introduce BPF_MAP_TYPE_REUSEPORT_SOCKARRAY")

---
Base: bpf-next.git master branch

      (tip a0c584fc18056709c8e047a82a6045d6c209f4ce
       "bpf: Fix use-after-free in offloaded map/prog info fill"
       as of 2026-04-09).

Tested with CONFIG_PREEMPT_RCU=y, CONFIG_KASAN=y (inline), CONFIG_SMP=n (single vCPU QEMU VM), gcc 13.3.0.

To reproduce: create a BPF_MAP_TYPE_REUSEPORT_SOCKARRAY with max_entries >= 100M, set rcu_cpu_stall_timeout low, pin the CPU with a SCHED_FIFO thread so the kworker stays in rcu_read_lock() long enough to trip the stall timeout, then close the fd. Without the fix the reuseport_array_free() kworker stalls RCU reliably; with the fix, cond_resched_rcu() yields periodically and no stall is observed. Reproducer (C source): repro_reuseport.c (https://pastebin.com/YjdwqdX1)

 kernel/bpf/reuseport_array.c | 2 ++
 1 file changed, 2 insertions(+)
kernel/bpf/reuseport_array.c +2 -0
index 49b8e5a0c6b4f..e3c789b80e2b8 100644
@@ -5,6 +5,7 @@
 #include <linux/bpf.h>
 #include <linux/err.h>
 #include <linux/sock_diag.h>
+#include <linux/rcupdate_wait.h>
 #include <net/sock_reuseport.h>
 #include <linux/btf_ids.h>
 
@@ -136,6 +137,7 @@ static void reuseport_array_free(struct bpf_map *map)
 			write_unlock_bh(&sk->sk_callback_lock);
 			RCU_INIT_POINTER(array->ptrs[i], NULL);
 		}
+		cond_resched_rcu();
 	}
 	rcu_read_unlock();
 
Signature · 1 line
2.43.0
#1 lore raw
Quoted 47 lines · Zijing Yin

reuseport_array_free() iterates over all map entries inside rcu_read_lock() to detach sockets from the array. When max_entries is very large (e.g., hundreds of millions), this loop runs for an extended period without yielding the CPU, triggering RCU stall warnings in the kworker thread that executes bpf_map_free_deferred().

The observed stall occurs because the loop has no scheduling point:

  rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
  Workqueue: events_unbound bpf_map_free_deferred
  Call Trace:
   reuseport_array_free+0x1ec/0x470 kernel/bpf/reuseport_array.c:127
   bpf_map_free_deferred+0x34a/0x7e0 kernel/bpf/syscall.c:893
   process_one_work+0x952/0x1a80
   worker_thread+0x87b/0x11f0

Add cond_resched_rcu() in the loop body to allow the scheduler to run and RCU grace periods to complete. This is safe because each iteration processes a single entry independently, sk->sk_callback_lock is not held across the yield point, and the map is fully detached from userspace so no concurrent insertions can occur.

This follows an established pattern for long-running kernel loops that must run under rcu_read_lock(). The closest precedent is in another BPF map free function:

  kernel/bpf/hashtab.c:1600
    htab_free_malloced_internal_structs()
      rcu_read_lock();
      for (i = 0; i < htab->n_buckets; i++) {
          ... walk bucket ...
          cond_resched_rcu();
      }
      rcu_read_unlock();
Fixes 5dc4c4b7d4e8 ("bpf: Introduce BPF_MAP_TYPE_REUSEPORT_SOCKARRAY")

---
Base: bpf-next.git master branch

      (tip a0c584fc18056709c8e047a82a6045d6c209f4ce
       "bpf: Fix use-after-free in offloaded map/prog info fill"
       as of 2026-04-09).

Tested with CONFIG_PREEMPT_RCU=y, CONFIG_KASAN=y (inline), CONFIG_SMP=n (single vCPU QEMU VM), gcc 13.3.0.

Quoted 7 lines

To reproduce: create a BPF_MAP_TYPE_REUSEPORT_SOCKARRAY with max_entries >= 100M, set rcu_cpu_stall_timeout low, pin the CPU with a SCHED_FIFO thread so the kworker stays in rcu_read_lock() long enough to trip the stall timeout, then close the fd. Without the fix the reuseport_array_free() kworker stalls RCU reliably; with the fix, cond_resched_rcu() yields periodically and no stall is observed. Reproducer (C source): repro_reuseport.c (https://pastebin.com/YjdwqdX1)

This is not a realistic scenario that is worth fixing.

pw-bot: cr