On 9/8/26 7:34 AM, Kuniyuki Iwashima wrote:
On Sun, Sep 6, 2026 at 1:23 AM <[email protected]> wrote:
tcp: Skip cond_resched() in inet_csk_listen_stop() under BPF context
bpf_sock_destroy() runs from the tcp iterator, under rcu_read_lock(). If
the sock is a listener that still has children in its accept queue,
tcp_abort() ends up in inet_csk_listen_stop() and the cond_resched()
there trips the debug check:
BUG: sleeping function called from invalid context at
net/ipv4/inet_connection_sock.c:1523
in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 628, name: test_progs
preempt_count: 0, expected: 0
RCU nest depth: 1, expected: 0
locks held by test_progs/628: 3, last CPU#3:
#0: ffff8881158cee18 (&p->lock){+.+.}-{4:4}, at: bpf_seq_read+0x56/0x1210
#1: ffff8881106bb858 (sk_lock-AF_INET6){+.+.}-{0:0}, at:
bpf_iter_tcp_seq_show+0x32b/0x4b0
#2: ffffffffb435af20 (rcu_read_lock){....}-{1:3}, at:
bpf_iter_run_prog+0x46b/0xde0
CPU: 3 UID: 0 PID: 628 Comm: test_progs Tainted: G W 7.2.0+
#65 PREEMPT
Tainted: [W]=WARN
Call Trace:
<TASK>
dump_stack_lvl+0xc1/0xf0
dump_stack+0x10/0x20
__might_resched+0x3d2/0x610
inet_csk_listen_stop+0x7b/0xbf0
tcp_abort+0x23b/0x3b0
bpf_sock_destroy+0xfc/0x140
bpf_prog_448133d24601754f_iter_tcp6_server+0x81/0x8a
bpf_iter_run_prog+0x538/0xde0
bpf_iter_tcp_seq_show+0x26b/0x4b0
bpf_seq_read+0x424/0x1210
vfs_read+0x197/0xe40
ksys_read+0x119/0x240
__x64_sys_read+0x72/0xc0
x64_sys_call+0x647/0x27e0
do_syscall_64+0xe5/0x610
entry_SYSCALL_64_after_hwframe+0x76/0x7e
RIP: 0033:0x7fad39b28aca
RSP: 002b:00007ffc381c61c0 EFLAGS: 00000246 ORIG_RAX: 0000000000000000
RAX: ffffffffffffffda RBX: 00007ffc381c6a88 RCX: 00007fad39b28aca
RDX: 0000000000000032 RSI: 00007ffc381c6250 RDI: 0000000000000014
RBP: 00007ffc381c61e0 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000003
R13: 0000000000000000 R14: 000055f077c1bbb0 R15: 00007fad3a0f3000
</TASK>
The commit that added the kfunc already guards lock_sock() in tcp_abort()
and udp_abort() with has_current_bpf_ctx(), but missed the listener path.
Do the same for the cond_resched(), it can't reschedule there anyway.
Fixes: 4ddbcb886268 ("bpf: Add bpf_sock_destroy kfunc")
Signed-off-by: Jiayuan Chen <[email protected]>
Is the justification "it can't reschedule there anyway" accurate?
With CONFIG_PREEMPT_DYNAMIC=y booted with preempt=none or preempt=voluntary,
cond_resched() expands to __cond_resched() which can actually reschedule.
The splat in the commit message confirms preempt_count is 0 while RCU nest
depth is 1. With preempt_count==0, should_resched(0) can be true and
__cond_resched() will call preempt_schedule_common() for a real reschedule.
Additionally, in configurations with CONFIG_PREEMPT_RCU=n where
rcu_read_lock() is preempt_disable(), __cond_resched() falls through to
rcu_all_qs() which calls rcu_qs() to report a quiescent state from inside
an RCU read-side critical section. That's a correctness problem beyond just
the debug check.
So the call can either reschedule (PREEMPT_DYNAMIC none/voluntary) or report
a bogus quiescent state (non-preemptible RCU). Could the justification be
reworded to explain that the loop runs inside the iterator's RCU read-side
critical section and must not reschedule or report a quiescent state there?
The code change itself is correct and matches the existing pattern in
tcp_abort() and udp_abort().
or maybe simply remove cond_resched(), hoping 7dadeaa6e851 would
resolve the scheduling issue.
Good suggestion. cond_resched has become old practice under
CONFIG_PREEMPT_LAZY
