Re: [PATCH v2] RDMA/siw: Clear association under lock if siw_qp_modify fails in siw_accept

[Zhu Yanjun <yanjun.zhu@linux.dev>]

在 2026/9/1 0:32, Leon Romanovsky 写道:
> On Thu, Aug 27, 2026 at 03:28:20PM +0200, Bernard Metzler wrote:
>> On 27.08.2026 14:55, Guoqing Jiang wrote:
>>> We need to clear cep before release state_lock as siw_qp_llp_close and
>>> siw_qp_modify->siw_qp_llp_close did.
>>>
>>> Otherwise if siw_qp_modify() fails in siw_accept(), the QP's state_lock
>>> is released before the error path cleanup. A concurrent ibv_modify_qp()
>>> transitioning the QP to ERROR can race in this window:
>>>
>>>     siw_accept()                       ibv_modify_qp(ERROR)
>>>     ----------------------             ----------------------
>>>     siw_qp_modify() fails
>>>     up_write(&qp->state_lock)
>>>                                        down_write(&qp->state_lock)
>>>                                        nextstate_from_idle():
>>> 				     if (qp->cep)
>>>                                          siw_cep_put(qp->cep) <- frees cep
>>>                                          qp->cep = NULL
>>>     goto error
>>>       cep->qp = NULL                   <- UAF
>>>
>>> Clear qp->cep and drop the association reference taken by siw_cep_get(),
>>> all under the write lock held from the initial down_write(&qp->state_lock).
>>> Thread B therefore sees qp->cep == NULL, skips its own put, and cannot free
>>> the cep before siw_accept() is done with it.
>>>
>>> Reported-by: Shuangpeng Bai <shuangpeng.kernel@gmail.com>
>>> Link: https://lore.kernel.org/linux-rdma/d6fbe475-a5c2-f975-99b0-a0bd6b6d10e8@linux.dev/T/#m5876c1ff2de8686a9a1173b8f1aa0ff5363a785c
>>> Signed-off-by: Guoqing Jiang <guoqing.jiang@linux.dev>
>>> ---
>>> V2: remove redundant code per Bernard's review
>>>
>>>    drivers/infiniband/sw/siw/siw_cm.c | 7 +++++--
>>>    1 file changed, 5 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/infiniband/sw/siw/siw_cm.c b/drivers/infiniband/sw/siw/siw_cm.c
>>> index 0245b25e7271..ed49818793dd 100644
>>> --- a/drivers/infiniband/sw/siw/siw_cm.c
>>> +++ b/drivers/infiniband/sw/siw/siw_cm.c
>>> @@ -1719,9 +1719,12 @@ int siw_accept(struct iw_cm_id *id, struct iw_cm_conn_param *params)
>>>    			   SIW_QP_ATTR_STATE | SIW_QP_ATTR_LLP_HANDLE |
>>>    				   SIW_QP_ATTR_ORD | SIW_QP_ATTR_IRD |
>>>    				   SIW_QP_ATTR_MPA);
>>> +	if (rv) {
>>> +		qp->cep = NULL;
>>> +		siw_cep_put(cep);
>>> +		goto error_unlock;
>>> +	}
>>>    	up_write(&qp->state_lock);
>>> -	if (rv)
>>> -		goto error;
>>>    	siw_dbg_cep(cep, "[QP %u]: send mpa reply, %d byte pdata\n",
>>>    		    qp_id(qp), params->private_data_len);
>>
>> Looks good. Thank you, Guoqing.
>>
>> Acked-by: Bernard Metzler <bernard.metzler@linux.dev>
> 
> Bernard,
> 
> I will take this patch. However, siw_accept() looks wrong to me. It
> mixes locking, qp->cep handling, and siw_cep_put() in a seemingly
> inconsistent manner.
> 
> For example, here you clear qp->cep and call siw_cep_put() while holding
> the lock, whereas in the error unwind path you do exactly the opposite.
> 
> Be aware that AI-generated suggestions for SIW are complete garbage. If
> you do not clean up these code paths, your driver will soon end up in a
> broken and unmaintainable state. The same applies to RXE.

Agree with you. Thanks a lot.

Zhu Yanjun

> 
> Thanks