Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Message-ID: <ab98ad0a-4c37-0ab7-02e2-a9bb439646f3@linux.ibm.com>
Date:   Tue, 8 Feb 2022 18:22:09 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:91.0) Gecko/20100101
 Thunderbird/91.5.0
Subject: Re: [BUG] net: smc: possible deadlock in smc_lgr_free() and
 smc_link_down_work()
Content-Language: en-US
To:     Jia-Ju Bai <baijiaju1990@gmail.com>
Cc:     linux-s390@vger.kernel.org, netdev@vger.kernel.org,
        linux-kernel <linux-kernel@vger.kernel.org>, davem@davemloft.net,
        kuba@kernel.org
References: <11fe65b8-eda4-121e-ec32-378b918d0909@gmail.com>
 <0936d5f3-aef2-0553-408b-07b3bb47e36b@linux.ibm.com>
 <9a27b497-80d7-ec6f-c8f1-69bee340f2e1@gmail.com>
From:   Karsten Graul <kgraul@linux.ibm.com>
Organization: IBM Deutschland Research & Development GmbH
In-Reply-To: <9a27b497-80d7-ec6f-c8f1-69bee340f2e1@gmail.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 06/02/2022 16:09, Jia-Ju Bai wrote:
> 
> 
> On 2022/2/2 1:06, Karsten Graul wrote:
>> On 01/02/2022 08:51, Jia-Ju Bai wrote:
>>> Hello,
>>>
>>> My static analysis tool reports a possible deadlock in the smc module in Linux 5.16:
>>>
>>> smc_lgr_free()
>>>    mutex_lock(&lgr->llc_conf_mutex); --> Line 1289 (Lock A)
>>>    smcr_link_clear()
>>>      smc_wr_free_link()
>>>        wait_event(lnk->wr_tx_wait, ...); --> Line 648 (Wait X)
>>>
>>> smc_link_down_work()
>>>    mutex_lock(&lgr->llc_conf_mutex); --> Line 1683 (Lock A)
>>>    smcr_link_down()
>>>      smcr_link_clear()
>>>        smc_wr_free_link()
>>>          smc_wr_wakeup_tx_wait()
>>>            wake_up_all(&lnk->wr_tx_wait); --> Line 78 (Wake X)
>>>
>>> When smc_lgr_free() is executed, "Wait X" is performed by holding "Lock A". If smc_link_down_work() is executed at this time, "Wake X" cannot be performed to wake up "Wait X" in smc_lgr_free(), because "Lock A" has been already hold by smc_lgr_free(), causing a possible deadlock.
>>>
>>> I am not quite sure whether this possible problem is real and how to fix it if it is real.
>>> Any feedback would be appreciated, thanks :)
> 
> Hi Karsten,
> 
> Thanks for the reply and explanation :)
> 
>> A deeper analysis showed up that this reported possible deadlock is actually not a problem.
>>
>> The wait on line 648 in smc_wr.c
>>     wait_event(lnk->wr_tx_wait, (!atomic_read(&lnk->wr_tx_refcnt)));
>> waits as long as the refcount wr_tx_refcnt is not zero.
>>
>> Every time when a caller stops using a link wr_tx_refcnt is decreased, and when it reaches
>> zero the wr_tx_wait is woken up in smc_wr_tx_link_put() in smc_wr.h, line 70:
>>         if (atomic_dec_and_test(&link->wr_tx_refcnt))
>>             wake_up_all(&link->wr_tx_wait);
> 
> Okay, you mean that wake_up_all(&link->wr_tx_wait) in smc_wr_tx_link_put() is used to wake up wait_event() in smc_wr_free_link().
> But I wonder whether wake_up_all(&lnk->wr_tx_wait) in smc_wr_wakeup_tx_wait() can wake up this wait_event()?
> If so, my report is in this case.
> 

Nope, due to the link state handling there is no current caller of smc_wr_wakeup_tx_wait() when 
smc_wr_free_link() starts to wait for the link to become free. First the link state is set to DOWN, 
then all waiters are woken up (and no one will start a new wait) and finally smc_wr_free_link()
"re-uses" the wait queue entry to wait for the link to become free.

I think its that reusing of the wait queue entry that confuses the tool.

>> Multiple callers of smc_wr_tx_link_put() do not run under the llc_conf_mutex lock, and those
>> who run under this mutex are saved against the wait_event() in smc_wr_free_link().
> 
> In fact, my tool also reports some other possible deadlocks invovling smc_wr_tx_link_put(), which can be called by holding llc_conf_mutex.
> There are three examples:
> 
> #BUG 1
> smc_lgr_free()
>   mutex_lock(&lgr->llc_conf_mutex); --> Line 1289 (Lock A)
>   smcr_link_clear()
>     smc_wr_free_link()
>       wait_event(lnk->wr_tx_wait, ...); --> Line 648 (Wait X)
> 
> smcr_buf_unuse()
>   mutex_lock(&lgr->llc_conf_mutex); --> Line 1087 (Lock A)
>   smc_llc_do_delete_rkey()
>     smc_llc_send_delete_rkey()
>       smc_wr_tx_link_put()
>         wake_up_all(&link->wr_tx_wait); --> Line 73 (Wake X)
> 
> #BUG 2
> smc_lgr_free()
>   mutex_lock(&lgr->llc_conf_mutex); --> Line 1289 (Lock A)
>   smcr_link_clear()
>     smc_wr_free_link()
>       wait_event(lnk->wr_tx_wait, ...); --> Line 648 (Wait X)
> 
> smc_link_down_work()
>   mutex_lock(&lgr->llc_conf_mutex); --> Line 1683 (Lock A)
>   smcr_link_down()
>     smc_llc_send_delete_link()
>       smc_wr_tx_link_put()
>         wake_up_all(&link->wr_tx_wait); --> Line 73 (Wake X)
> 
> #BUG 3
> smc_llc_process_cli_delete_link()
>   mutex_lock(&lgr->llc_conf_mutex); --> Line 1578 (Lock A)
>   smc_llc_send_message()
>     smc_llc_add_pending_send()
>       smc_wr_tx_get_free_slot()
>         wait_event_interruptible_timeout(link->wr_tx_wait, ...); --> Line 219 (Wake X)
> 
> smc_llc_process_cli_add_link()
>   mutex_lock(&lgr->llc_conf_mutex); --> Line 1198 (Lock A)
>   smc_llc_cli_add_link_invite()
>     smc_llc_send_add_link()
>       smc_wr_tx_link_put()
>         wake_up_all(&link->wr_tx_wait); --> Line 73 (Wake X)
> 
> I am not quite sure whether these possible problems are real.
> Any feedback would be appreciated, thanks :)

Same here, because the wait queue entry is used in two scenarios and some processing separates
those scenarios, the code checker finds problems that 'should' never happen.

I wonder if it would be acceptable to introduce an extra wait queue entry only for the processing in
smc_wr_free_link(), I reused an existing one to save some memory... but a cleaner code also counts.
Not sure what to prefer.

> 
>>
>> Thank you for reporting this finding! Which tool did you use for this analysis?
> 
> Thanks for your interest :)
> I have implemented a static analysis tool based on LLVM, to detect deadlocks caused by locking cycles and improper waiting/waking operations.
> However, this tool still reports some false positives, and thus I am still improving the accuracy of this tool.
> Suggestions on deadlock detection (especially new/infrequent patterns causing deadlocks) or the tool are welcome ;)
> 
> 
> Best wishes,
> Jia-Ju Bai
> 

-- 
Karsten