Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Date:   Wed, 17 Oct 2018 14:00:41 +0200
From:   Stefan Agner <stefan@agner.ch>
To:     p.zabel@pengutronix.de
Cc:     airlied@linux.ie, rmk+kernel@armlinux.org.uk,
        l.stach@pengutronix.de, dri-devel@lists.freedesktop.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH] Revert "drm/imx: don't destroy mode objects manually on
 driver unbind"
In-Reply-To: <d55ebfae98b0ccf30895700be0881264@agner.ch>
References: <20181016160923.2042-1-stefan@agner.ch>
 <d55ebfae98b0ccf30895700be0881264@agner.ch>
Message-ID: <66518fd4ba1fe77344d849a0b3949342@agner.ch>
User-Agent: Roundcube Webmail/1.3.7
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 17.10.2018 13:25, Stefan Agner wrote:
> On 16.10.2018 18:09, Stefan Agner wrote:
>> This reverts commit 8e3b16e2117409625b89807de3912ff773aea354.
>>
>> Using the component framework requires all components to undo in
>> ->unbind what ->bind does. Unfortunately that particular commit
>> broke this rule. In particular, this is an issue if a single
>> component during probe fails. In that case, component_bind_all()
>> calls unbind on already succussfully bound components, and then
>> frees memory allocated using devm. If one of those components
>> registered e.g. an encoder with the framework then this leads to
>> use after free situations.
>>
>> Revert the commit to ensure that all components properly undo
>> what ->bind does.
> 
> After Lucas comment mentioning HDMI unbind is not proper I looked
> through all the unbind again. The other unbind functions need some
> fixing too. I did not bother checking whether those were always broken
> or just because things changed (the commit this is reverting was in
> 2016).... Here is what I found:
> 
>>
>> Link:
>> https://www.mail-archive.com/dri-devel@lists.freedesktop.org/msg233327.html
>> Suggested-by: Russell King <rmk+kernel@armlinux.org.uk>
>> Signed-off-by: Stefan Agner <stefan@agner.ch>
>> ---
>>  drivers/gpu/drm/imx/imx-drm-core.c     | 4 ++--
>>  drivers/gpu/drm/imx/imx-ldb.c          | 6 ++++++
>>  drivers/gpu/drm/imx/imx-tve.c          | 3 +++
>>  drivers/gpu/drm/imx/parallel-display.c | 3 +++
>>  4 files changed, 14 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/imx/imx-drm-core.c
>> b/drivers/gpu/drm/imx/imx-drm-core.c
>> index 5ea0c82f9957..caa6061a98ba 100644
>> --- a/drivers/gpu/drm/imx/imx-drm-core.c
>> +++ b/drivers/gpu/drm/imx/imx-drm-core.c
>> @@ -305,11 +305,11 @@ static void imx_drm_unbind(struct device *dev)
>>
>>  	drm_fb_cma_fbdev_fini(drm);
>>
>> -	drm_mode_config_cleanup(drm);
>> -
>>  	component_unbind_all(drm->dev, drm);
>>  	dev_set_drvdata(dev, NULL);
>>
>> +	drm_mode_config_cleanup(drm);
>> +
>>  	drm_dev_put(drm);
>>  }
>>
>> diff --git a/drivers/gpu/drm/imx/imx-ldb.c b/drivers/gpu/drm/imx/imx-ldb.c
>> index 3bd0f8a18e74..592aabc4a262 100644
>> --- a/drivers/gpu/drm/imx/imx-ldb.c
>> +++ b/drivers/gpu/drm/imx/imx-ldb.c
>> @@ -723,6 +723,12 @@ static void imx_ldb_unbind(struct device *dev,
>> struct device *master,
>>  		if (channel->panel)
>>  			drm_panel_detach(channel->panel);
>>
>> +		if (!channel->connector.funcs)
>> +			continue;
>> +
>> +		channel->connector.funcs->destroy(&channel->connector);
>> +		channel->encoder.funcs->destroy(&channel->encoder);
> 
> There can be an encoder and bridge, or an encoder and connector. All of
> them should be properly cleaned up.
> 
> So I guess this should look like this:
> 
>   		if (channel->panel)
>   			drm_panel_detach(channel->panel);
> 
>   		if (channel->bridge)
>   			drm_bridge_detach(channel->bridge);

Actually, when a bridge is attached to an encoder, which is the case
when unbind is getting called, then encoder cleanup takes care of
drm_bridge_detach.

> 
> 		if (channel->connector.funcs)
> 			channel->connector.funcs->destroy(&channel->connector);
> 
> 		if (channel->encoder.funcs)
> 			channel->encoder.funcs->destroy(&channel->encoder);
> 
> 		kfree(channel->edid);
> 		i2c_put_adapter(channel->ddc);
> 
> The last two functions following are only strictly necessary when
> connector is initialized. But its safe to call them with null, so I
> would just call them always.
> 
> 
>> +
>>  		kfree(channel->edid);
>>  		i2c_put_adapter(channel->ddc);
>>  	}
>> diff --git a/drivers/gpu/drm/imx/imx-tve.c b/drivers/gpu/drm/imx/imx-tve.c
>> index cffd3310240e..8d6e89ce1edb 100644
>> --- a/drivers/gpu/drm/imx/imx-tve.c
>> +++ b/drivers/gpu/drm/imx/imx-tve.c
>> @@ -673,6 +673,9 @@ static void imx_tve_unbind(struct device *dev,
>> struct device *master,
>>  {
>>  	struct imx_tve *tve = dev_get_drvdata(dev);
>>
>> +	tve->connector.funcs->destroy(&tve->connector);
>> +	tve->encoder.funcs->destroy(&tve->encoder);
>> +
> 
> Cleanup of tve->ddc missing.
> 
>>  	if (!IS_ERR(tve->dac_reg))
>>  		regulator_disable(tve->dac_reg);
>>  }
>> diff --git a/drivers/gpu/drm/imx/parallel-display.c
>> b/drivers/gpu/drm/imx/parallel-display.c
>> index aefd04e18f93..6f11bffcde37 100644
>> --- a/drivers/gpu/drm/imx/parallel-display.c
>> +++ b/drivers/gpu/drm/imx/parallel-display.c
>> @@ -258,6 +258,9 @@ static void imx_pd_unbind(struct device *dev,
>> struct device *master,
>>  	if (imxpd->panel)
>>  		drm_panel_detach(imxpd->panel);
>>
> 
> And in this case a bridge detach is missing.

Same here.

> 
> Will send a v2 with that addressed.

Still valid.

--
Stefan

> 
> --
> Stefan
> 
>> +	imxpd->encoder.funcs->destroy(&imxpd->encoder);
>> +	imxpd->connector.funcs->destroy(&imxpd->connector);
>> +
>>  	kfree(imxpd->edid);
>>  }