Received-SPF: pass (google.com: domain of linux-bluetooth-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Content-Type: text/plain;
        charset=utf-8
Mime-Version: 1.0 (Mac OS X Mail 14.0 \(3654.120.0.1.13\))
Subject: Re: [PATCH] Bluetooth: Fix memory leak of hci device
From:   Marcel Holtmann <marcel@holtmann.org>
In-Reply-To: <CABBYNZL=wSZaLsLAm7Btg6Tmv=K7z903kQxfvRRVPPH0nAcsOA@mail.gmail.com>
Date:   Wed, 13 Oct 2021 08:48:25 +0200
Cc:     Wei Yongjun <weiyongjun1@huawei.com>,
        linux-bluetooth <linux-bluetooth@vger.kernel.org>,
        Johan Hedberg <johan.hedberg@gmail.com>
Content-Transfer-Encoding: 8BIT
Message-Id: <6BD02F26-122B-42FB-8632-DB673A6F336A@holtmann.org>
References: <20211012075634.8041-1-weiyongjun1@huawei.com>
 <3EC78E44-6E60-4B99-B880-2F5CC468C424@holtmann.org>
 <CABBYNZL=wSZaLsLAm7Btg6Tmv=K7z903kQxfvRRVPPH0nAcsOA@mail.gmail.com>
To:     Luiz Augusto von Dentz <luiz.dentz@gmail.com>
Precedence: bulk

Hi Luiz,

>>> Fault injection test reported memory leak of hci device as follows:
>>> 
>>> unreferenced object 0xffff88800b858000 (size 8192):
>>> comm "kworker/0:2", pid 167, jiffies 4294955747 (age 557.148s)
>>> hex dump (first 32 bytes):
>>>   00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>>>   00 00 00 00 00 00 00 00 00 00 00 00 ad 4e ad de  .............N..
>>> backtrace:
>>>   [<0000000070eb1059>] kmem_cache_alloc_trace mm/slub.c:3208
>>>   [<00000000015eb521>] hci_alloc_dev_priv include/linux/slab.h:591
>>>   [<00000000dcfc1e21>] bpa10x_probe include/net/bluetooth/hci_core.h:1240
>>>   [<000000005d3028c7>] usb_probe_interface drivers/usb/core/driver.c:397
>>>   [<00000000cbac9243>] really_probe drivers/base/dd.c:517
>>>   [<0000000024cab3f0>] __driver_probe_device drivers/base/dd.c:751
>>>   [<00000000202135cb>] driver_probe_device drivers/base/dd.c:782
>>>   [<000000000761f2bc>] __device_attach_driver drivers/base/dd.c:899
>>>   [<00000000f7d63134>] bus_for_each_drv drivers/base/bus.c:427
>>>   [<00000000c9551f0b>] __device_attach drivers/base/dd.c:971
>>>   [<000000007f79bd16>] bus_probe_device drivers/base/bus.c:487
>>>   [<000000007bb8b95a>] device_add drivers/base/core.c:3364
>>>   [<000000009564d9ea>] usb_set_configuration drivers/usb/core/message.c:2171
>>>   [<00000000e4657087>] usb_generic_driver_probe drivers/usb/core/generic.c:239
>>>   [<0000000071ede518>] usb_probe_device drivers/usb/core/driver.c:294
>>>   [<00000000cbac9243>] really_probe drivers/base/dd.c:517
>>> 
>>> hci_alloc_dev() do not init the device's flag. And hci_free_dev()
>>> using put_device() to free the memory allocated for this device,
>>> but it calls just kfree(dev) only in case of HCI_UNREGISTER flag
>>> is set. So any error handing before hci_register_dev() success
>>> will cause memory leak.
>>> 
>>> To avoid this behaviour we need to set hdev HCI_UNREGISTER flag
>>> in hci_alloc_dev_priv().
>>> 
>>> Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
>>> 
>>> diff --git a/net/bluetooth/hci_core.c b/net/bluetooth/hci_core.c
>>> index 8a47a3017d61..42410f568e90 100644
>>> --- a/net/bluetooth/hci_core.c
>>> +++ b/net/bluetooth/hci_core.c
>>> @@ -3876,6 +3876,11 @@ struct hci_dev *hci_alloc_dev_priv(int sizeof_priv)
>>>      INIT_DELAYED_WORK(&hdev->cmd_timer, hci_cmd_timeout);
>>>      INIT_DELAYED_WORK(&hdev->ncmd_timer, hci_ncmd_timeout);
>>> 
>>> +     /* We need to set HCI_UNREGISTER flag to correctly release
>>> +      * the device in hci_free_dev()
>>> +      */
>>> +     hci_dev_set_flag(hdev, HCI_UNREGISTER);
>>> +
>> 
>> I can see the point in the bug report, but I don’t see that the fix is correct. Can you prove that this fix is correct when hci_register_dev is actually called.
> 
> I also wonder where is the actual check for HCI_UNREGISTER that the
> commit description says prevents the kfree? hci_free_dev itself just
> calls put_device, so perhaps it is actually talking about the check in
> bt_host_release, anyway in for this to work the HCI_UNREGISTER would
> have to be cleared by hci_register_dev otherwise the likes of
> hci_dev_do_open don't work as it checks if HCI_UNREGISTER had been
> called.
> 
> We also would need to check if it is safe to call hci_release_dev if
> the workqueues, etc, had not been initialized yet, or perhaps don't
> really use HCI_UNREGISTER and just do something like this:
> 
> diff --git a/net/bluetooth/hci_sysfs.c b/net/bluetooth/hci_sysfs.c
> index 7827639ecf5c..81c50b47183f 100644
> --- a/net/bluetooth/hci_sysfs.c
> +++ b/net/bluetooth/hci_sysfs.c
> @@ -86,6 +86,9 @@ static void bt_host_release(struct device *dev)
> 
>        if (hci_dev_test_flag(hdev, HCI_UNREGISTER))
>                hci_release_dev(hdev);
> +       else
> +               kfree(hdev);
> +
>        module_put(THIS_MODULE);
> 
> If this doesn't fix all the leaks that probably means part of the
> hci_release_dev still needs to be executed which can probably be done
> by having the check for HCI_UNREGISTER around the code that does
> actually depend on hci_register_dev.

what we really need to do is move towards devm integration so that all allocation are tied to the driver model. However this might need a cleaner struct device integration.

Regards

Marcel