Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <9225b221-6f21-96b3-4d21-c3ac477df6de@gmail.com>
Date:   Thu, 13 Apr 2023 23:25:01 -0500
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.10.0
Subject: Re: [PATCH for-next 2/3] RDMA/rtrs: Fix rxe_dealloc_pd warning
Content-Language: en-US
To:     Guoqing Jiang <guoqing.jiang@linux.dev>,
        "Zhijian Li (Fujitsu)" <lizhijian@fujitsu.com>,
        Leon Romanovsky <leon@kernel.org>
Cc:     "haris.iqbal@ionos.com" <haris.iqbal@ionos.com>,
        "jinpu.wang@ionos.com" <jinpu.wang@ionos.com>,
        "jgg@ziepe.ca" <jgg@ziepe.ca>,
        "linux-rdma@vger.kernel.org" <linux-rdma@vger.kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
References: <1681108984-2-1-git-send-email-lizhijian@fujitsu.com>
 <1681108984-2-3-git-send-email-lizhijian@fujitsu.com>
 <20230410120809.GN182481@unreal>
 <0d9c57db-bca3-adb4-71fd-7362e4842917@linux.dev>
 <85323eb2-cfc7-d1b8-3a75-3fa63dde29db@fujitsu.com>
 <20230411122651.GV182481@unreal>
 <aa43746f-77f9-8592-5370-2a5042506ee5@fujitsu.com>
 <8fe62e38-e43d-3d6c-624f-1c8ce5859788@linux.dev>
 <f85cdaaa-b1d6-bc89-e963-8b611ae3667d@fujitsu.com>
 <230b310a-26ef-34f1-4c3b-c2360088ce04@linux.dev>
From:   Bob Pearson <rpearsonhpe@gmail.com>
In-Reply-To: <230b310a-26ef-34f1-4c3b-c2360088ce04@linux.dev>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 4/13/23 22:40, Guoqing Jiang wrote:
> 
> 
> On 4/13/23 16:12, Zhijian Li (Fujitsu) wrote:
>> On 13/04/2023 15:35, Guoqing Jiang wrote:
>>> Hi,
>>>
>>> I take a closer look today.
>>>
>>> On 4/12/23 09:15, Zhijian Li (Fujitsu) wrote:
>>>> On 11/04/2023 20:26, Leon Romanovsky wrote:
>>>>> On Tue, Apr 11, 2023 at 02:43:46AM +0000, Zhijian Li (Fujitsu) wrote:
>>>>>> On 10/04/2023 21:10, Guoqing Jiang wrote:
>>>>>>> On 4/10/23 20:08, Leon Romanovsky wrote:
>>>>>>>> On Mon, Apr 10, 2023 at 06:43:03AM +0000, Li Zhijian wrote:
>>>>>>>>> The warning occurs when destroying PD whose reference count is not zero.
>>>>>>>>>
>>>>>>>>> Precodition: clt_path->s.con_num is 2.
>>>>>>>>> So 2 cm connection will be created as below:
>>>>>>>>> CPU0                                              CPU1
>>>>>>>>> init_conns {                              |
>>>>>>>>>       create_cm() // a. con[0] created        |
>>>>>>>>>                                               |  a'. rtrs_clt_rdma_cm_handler() {
>>>>>>>>>                                               |    rtrs_rdma_addr_resolved()
>>>>>>>>>                                               |      create_con_cq_qp(con); << con[0]
>>>>>>>>>                                               |  }
>>>>>>>>>                                               | in this moment, refcnt of PD was increased to 2+
>>> What do you mean "refcnt of PD"? usecnt in struct ib_pd or dev_ref.
>> I mean usecnt in struct ib_pd
>>
>>
>>
>>>>>>>>>                                               |
>>>>>>>>>       create_cm() // b. cid = 1, failed       |
>>>>>>>>>         destroy_con_cq_qp()                   |
>>>>>>>>>           rtrs_ib_dev_put()                   |
>>>>>>>>>             dev_free()                        |
>>>>>>>>>               ib_dealloc_pd(dev->ib_pd) << PD |
>>>>>>>>>                is destroyed, but refcnt is    |
>>>>>>>>>                still greater than 0           |
>>> Assuming you mean "pd->usecnt". We only allocate pd in con[0] by rtrs_ib_dev_find_or_add,
>>> if con[1] failed to create cm, then alloc_path_reqs -> ib_alloc_mr -> atomic_inc(&pd->usecnt)
> 
> The above can't be invoked, right?
> 
>>> can't be triggered. Is there other places could increase the refcnt?
>> Yes, when create a qp, it will also associate to this PD, that also mean refcnt of PD will be increased.
>>
>> When con[0](create_con_cq_qp) succeeded, refcnt of PD will be 2. and then when con[1] failed, since
>> QP didn't create, refcnt of PD is still 2. con[1]'s cleanup will destroy the PD(ib_dealloc_pd) since dev_ref = 1, after that its
>> refcnt is still 1.
> 
> I can see the path increase usecnt to 1.
> 
> rtrs_cq_qp_create -> create_qp
>                     -> rdma_create_qp
>                         -> ib_create_qp
>                             -> create_qp
>                             -> ib_qp_usecnt_inc which increases pd->usecnt
> 
> Where is another place to increase usecnt to 2?
> 
>>> Then what is the appropriate time to call destroy_con_cq_qp for this scenario?
>>> Otherwise there could be memory leak.
>> we must ensure QP in con[0] is closed before destroying the PD.
>> Currently destroy_con_cq_qp() subroutine will close the opened QP first.
> 
> Let me try another way, with below change, rtrs_ib_dev_put can't be called
> from destroy_con_cq_qp, right?
> 
> +    if (!con->has_dev)
> +        return;
>      if (clt_path->s.dev_ref && !--clt_path->s.dev_ref) {
>          rtrs_ib_dev_put(clt_path->s.dev);
>          clt_path->s.dev = NULL;
> 
> Then when will you dealloc pd and free rtrs_ib_dev?
> 
> Thanks,
> Guoqing

I think that wondering into Leon's reference counting is a really bad idea. Currently the assumed rule
is that rdma-core keeps its ref counts and rxe keeps its. rxe defers the return from rxe_dealloc_pd()
until the rxe ref count drops to zero for that pd, sleeping if necessary. (There is a timeout value set
where rxe will return anyway but it will throw a WARN. If the timeout isn't long enough under heavy load
we could extend it.) If it doesn't happen, or it happens too soon, then there is a ref count bug in rxe
that needs to be fixed.

Fixing rxe ref count bugs is hard enough without entangling rdma-core ref counts into the mix.

Bob