On Mon 12 Jun 2023 at 22:29, renmingshuai <[email protected]> wrote:
>>On Fri 09 Jun 2023 at 17:46, renmingshuai <[email protected]> wrote:
>>> When a new chain is added by using tc, one soft lockup alarm will be
>>> generated after delete the prio 0 filter of the chain. To reproduce
>>> the problem, perform the following steps:
>>> (1) tc qdisc add dev eth0 root handle 1: htb default 1
>>> (2) tc chain add dev eth0
>>> (3) tc filter del dev eth0 chain 0 parent 1: prio 0
>>> (4) tc filter add dev eth0 chain 0 parent 1:
>>>
>>>
>>> The refcnt of the chain added by step 2 is equal to 1. After step 3,
>>> the flushing flag of the chain is set to true in the tcf_chain_flush()
>>> called by tc_del_tfilter() because the prio is 0. In this case, if
>>> we add a new filter to this chain, it will never succeed and try again
>>> and again because the refresh flash is always true and refcnt is 1.
>>> A soft lock alarm is generated 20 seconds later.
>>
>>Hi Mingshuai,
>>
>>Thanks for investigating and reporting this!
>>
>>> The stack is show as below:
>>>
>>> Kernel panic - not syncing: softlockup: hung tasks
>>> CPU: 2 PID: 3321861 Comm: tc Kdump: loaded Tainted: G
>>> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996)
>>> Call Trace:
>>> <IRQ>
>>> dump_stack+0x57/0x6e
>>> panic+0x196/0x3ec
>>> watchdog_timer_fn.cold+0x16/0x5c
>>> __run_hrtimer+0x5e/0x190
>>> __hrtimer_run_queues+0x8a/0xe0
>>> hrtimer_interrupt+0x110/0x2c0
>>> ? irqtime_account_irq+0x49/0xf0
>>> __sysvec_apic_timer_interrupt+0x5f/0xe0
>>> asm_call_irq_on_stack+0x12/0x20
>>> </IRQ>
>>> sysvec_apic_timer_interrupt+0x72/0x80
>>> asm_sysvec_apic_timer_interrupt+0x12/0x20
>>> RIP: 0010:mutex_lock+0x24/0x70
>>> RSP: 0018:ffffa836004ab9a8 EFLAGS: 00000246
>>> RAX: 0000000000000000 RBX: ffff95bb02d76700 RCX: 0000000000000000
>>> RDX: ffff95bb27462100 RSI: 0000000000000000 RDI: ffff95ba5b527000
>>> RBP: ffff95ba5b527000 R08: 0000000000000001 R09: ffffa836004abbb8
>>> R10: 000000000000000f R11: 0000000000000000 R12: 0000000000000000
>>> R13: ffff95ba5b527000 R14: ffffa836004abbb8 R15: 0000000000000001
>>> __tcf_chain_put+0x27/0x200
>>> tc_new_tfilter+0x5e8/0x810
>>> ? tc_setup_cb_add+0x210/0x210
>>> rtnetlink_rcv_msg+0x2e3/0x380
>>> ? rtnl_calcit.isra.0+0x120/0x120
>>> netlink_rcv_skb+0x50/0x100
>>> netlink_unicast+0x12d/0x1d0
>>> netlink_sendmsg+0x286/0x490
>>> sock_sendmsg+0x62/0x70
>>> ____sys_sendmsg+0x24c/0x2c0
>>> ? import_iovec+0x17/0x20
>>> ? sendmsg_copy_msghdr+0x80/0xa0
>>> ___sys_sendmsg+0x75/0xc0
>>> ? do_fault_around+0x118/0x160
>>> ? do_read_fault+0x68/0xf0
>>> ? __handle_mm_fault+0x3f9/0x6f0
>>> __sys_sendmsg+0x59/0xa0
>>> do_syscall_64+0x33/0x40
>>> entry_SYSCALL_64_after_hwframe+0x61/0xc6
>>> RIP: 0033:0x7f96705b8247
>>> RSP: 002b:00007ffe552e9dc8 EFLAGS: 00000246 ORIG_RAX: 000000000000002e
>>> RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f96705b8247
>>> RDX: 0000000000000000 RSI: 00007ffe552e9e40 RDI: 0000000000000003
>>> RBP: 0000000000000001 R08: 0000000000000001 R09: 0000558113f678b0
>>> R10: 00007f967069ab00 R11: 0000000000000246 R12: 00000000647ea089
>>> R13: 00007ffe552e9f30 R14: 0000000000000001 R15: 0000558113175f00
>>>
>>> To avoid this case, set chain->flushing to be false if the chain->refcnt
>>> is 2 after flushing the chain when prio is 0. I also add one test to tdc.
>>>
>>> Fixes: 726d061286ce ("net: sched: prevent insertion of new classifiers during chain flush")
>>> Signed-off-by: Mingshuai Ren <[email protected]>
>>> Reviewed-by: Aichun Li <[email protected]>
>>> ---
>>> V1 -> V2:
>>> * Correct the judgment on the value chain->refcnt and add one test to tdc
>>> ---
>>> net/sched/cls_api.c | 7 ++++++
>>> .../tc-testing/tc-tests/infra/filter.json | 25 +++++++++++++++++++
>>> 2 files changed, 32 insertions(+)
>>> create mode 100644 tools/testing/selftests/tc-testing/tc-tests/infra/filter.json
>>>
>>> diff --git a/net/sched/cls_api.c b/net/sched/cls_api.c
>>> index 2621550bfddc..3ea054e03fbf 100644
>>> --- a/net/sched/cls_api.c
>>> +++ b/net/sched/cls_api.c
>>> @@ -2442,6 +2442,13 @@ static int tc_del_tfilter(struct sk_buff *skb, struct nlmsghdr *n,
>>> tfilter_notify_chain(net, skb, block, q, parent, n,
>>> chain, RTM_DELTFILTER, extack);
>>> tcf_chain_flush(chain, rtnl_held);
>>> + /* Set the flushing flags to false to prevent an infinite loop
>>> + * when a new filter is added.
>>> + */
>>> + mutex_lock(&chain->filter_chain_lock);
>>> + if (chain->refcnt == 2)
>>> + chain->flushing = false;
>>> + mutex_unlock(&chain->filter_chain_lock);
>>
>>I don't think this check is enough since there can be concurrent filter
>>ops holding the reference to the chain. This just makes the issue harder
>>to reproduce.
>>
>>I'll try to formulate a fix that takes any potential concurrent users
>>into account and verify it with your test.
>
> Thanks for your reply.
> I didn't find any concurrent scenarios that would "escape" this check.
> That's my understanding.
> During the flush operation, after filter_chain_lock is obtained, no new chain reference
> could be added. After unlock, chain->flushing is set to true. The chain->refcnt and
> chain->action_refcnt are always different. Therefore, chain->flushing will be never set to false.
I agree with your analysis.
> Then there seems to be no concurrency problem until flushing is set to true.
> would you mind tell me the concurrent scenario you're talking about?
What if, for example, chain->refcnt==3 after flush because there is
another task inserting filters to the same chain concurrently? In such
case for chains explicitly created with chains API, chain->flushing flag
will never be reset afterwards and user will get the same lockup when
trying to install any following filters. Also, refcnt==2 only means that
there is no concurrent users when flushing explicitly created chains
(because chains API holds a "passive" reference to such chains). For
regular chains implicitly created by filters API refcnt==2 here means
that there are active concurrent users.