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;
Subject: Re: [BUG] bus: mhi: possible ABBA deadlock in mhi_pm_m0_transition()
 and mhi_send_cmd()
To:     Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>,
        Bhaumik Bhatt <bbhatt@codeaurora.org>
Cc:     hemantk@codeaurora.org, linux-arm-msm@vger.kernel.org,
        linux-kernel <linux-kernel@vger.kernel.org>
References: <0a213c92-32d5-efc8-079b-dd20d5ecfe20@gmail.com>
 <2fbeeb8dc647dc88e69ec6624c2078d1@codeaurora.org>
 <20210716064512.GE3323@workstation>
From:   Jia-Ju Bai <baijiaju1990@gmail.com>
Message-ID: <34ea09b0-8622-90f5-ee70-68c27081de39@gmail.com>
Date:   Mon, 19 Jul 2021 10:08:55 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:68.0) Gecko/20100101
 Thunderbird/68.4.1
MIME-Version: 1.0
In-Reply-To: <20210716064512.GE3323@workstation>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 8bit
Content-Language: en-US
Precedence: bulk


On 2021/7/16 14:45, Manivannan Sadhasivam wrote:
> On Thu, Jul 15, 2021 at 09:57:22AM -0700, Bhaumik Bhatt wrote:
>> On 2021-07-15 02:45 AM, Jia-Ju Bai wrote:
>>> Hello,
>>>
>>> I find there is a possible ABBA deadlock in the MHI driver in Linux
>>> 5.10:
>>>
>>> In mhi_pm_m0_transition():
>>> 262:     read_lock_bh(&mhi_cntrl->pm_lock);
>>> 281:     spin_lock_irq(&mhi_cmd->lock);
>>>
>>> In mhi_send_cmd():
>>> 1181:   spin_lock_bh(&mhi_cmd->lock);
>>> 1207:   read_lock_bh(&mhi_cntrl->pm_lock);
>>>
>>> When mhi_pm_m0_transition() and mhi_send_cmd() are concurrently
>>> executed, the deadlock can occur.
>>>
>>> I check the code and find a possible case of such concurrent execution:
>>>
>>> #CPU1:
>>> mhi_poll (mhi_event->process_event(...))
>>>    mhi_process_ctrl_ev_ring
>>>      mhi_pm_m0_transition
>>>
>>> #CPU2:
>>> mhi_prepare_for_transfer
>>>    mhi_prepare_channel
>>>      mhi_send_cmd
>>>
>>> Note that mhi_poll() and mhi_prepare_for_transfer() are both exported
>>> by EXPORT_SYMBOL.
>>> Thus, I guess these two functions could be concurrently called by a MHI
>>> driver.
>>>
>>> I am not quite sure whether this possible deadlock is real and how to
>>> fix it if it is real.
>>> Any feedback would be appreciated, thanks :)
>>>
>>>
>>> Best wishes,
>>> Jia-Ju Bai
>> Few pointers from your example:
>>
>> 1. mhi_poll() is currently not used by any client upstream yet.
> Then this shouldn't be added in first place... :/
>
>> 2. Polling is not to be used for single event ring (shared control + data)
>> cases
>> since it is meant to be for client drivers with dedicated data packets only.
>> 3. mhi_send_cmd() will always be called after an mhi_pm_m0_transition() has
>> completed by design since we wait for the device to be held in M0 prior to
>> it.
>>
> But client can be unloaded during M0 event!
>
> Anyway, I don't think the deadlock scenario is valid because of the usage
> of "read_lock_bh()". So if "mhi_send_cmd()" has acquired
> "spin_lock_bh(&mhi_cmd->lock)", it can always acquire
> "read_lock_bh(&mhi_cntrl->pm_lock)" as multiple readers can acquire the
> read lock.
>
> Deadlock would only occur if one of the functions take write lock.
>
> Thanks for auditing.

Thanks for the feedback and explanation :)


Best wishes,
Jia-Ju Bai