[1/1] mtd:nand:fix memory leak

Message ID 1522811151-18853-1-git-send-email-wangxidong_97@163.com
State Superseded
Headers show
Series
  • [1/1] mtd:nand:fix memory leak
Related show

Commit Message

Xidong Wang April 4, 2018, 3:05 a.m.
In function tango_nand_probe(), the memory allocated by
clk_get() is not released on the normal path and
the error path that IS_ERR(nfc->chan) returns true.
This will result in a memory leak bug.

Signed-off-by: Xidong Wang <wangxidong_97@163.com>
---
 drivers/mtd/nand/tango_nand.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

Comments

Miquel Raynal April 4, 2018, 6:28 a.m. | #1
Hi Xidong,

As part of a reorganization in the NAND subsystem, you should now
prefix your commit title this way:

        mtd: rawnand: tango: fix memory leak

Not sure if this patch is candidate to cc:stable?

On Wed,  4 Apr 2018 11:05:51 +0800, Xidong Wang
<wangxidong_97@163.com> wrote:

> In function tango_nand_probe(), the memory allocated by
> clk_get() is not released on the normal path and
> the error path that IS_ERR(nfc->chan) returns true.

The fact that the error path returns true looks out of topic, can you
remove it? Just saying that you fix a memory leak is enough I guess.

> This will result in a memory leak bug.
> 
> Signed-off-by: Xidong Wang <wangxidong_97@163.com>
> ---
>  drivers/mtd/nand/tango_nand.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
> index c5bee00b..8083459 100644
> --- a/drivers/mtd/nand/tango_nand.c
> +++ b/drivers/mtd/nand/tango_nand.c
> @@ -648,12 +648,15 @@ static int tango_nand_probe(struct platform_device *pdev)
>  		return PTR_ERR(clk);
>  
>  	nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
> -	if (IS_ERR(nfc->chan))
> +	if (IS_ERR(nfc->chan)) {
> +		clk_put(clk);
>  		return PTR_ERR(nfc->chan);
> +	}
>  
>  	platform_set_drvdata(pdev, nfc);
>  	nand_hw_control_init(&nfc->hw);
>  	nfc->freq_kHz = clk_get_rate(clk) / 1000;
> +	clk_put(clk);

If the clock is used only here, better do the frequency derivation
right after the clock_get(), and follow with a clk_put()? This way you
don't have to change the error path and 'related' actions remain
grouped.

Thanks for fixing this,
Miquèl
Boris Brezillon April 4, 2018, 7:07 a.m. | #2
On Wed, 4 Apr 2018 08:28:07 +0200
Miquel Raynal <miquel.raynal@bootlin.com> wrote:

> Hi Xidong,
> 
> As part of a reorganization in the NAND subsystem, you should now
> prefix your commit title this way:
> 
>         mtd: rawnand: tango: fix memory leak
> 
> Not sure if this patch is candidate to cc:stable?
> 
> On Wed,  4 Apr 2018 11:05:51 +0800, Xidong Wang
> <wangxidong_97@163.com> wrote:
> 
> > In function tango_nand_probe(), the memory allocated by
> > clk_get() is not released on the normal path and
> > the error path that IS_ERR(nfc->chan) returns true.  
> 
> The fact that the error path returns true looks out of topic, can you
> remove it? Just saying that you fix a memory leak is enough I guess.
> 
> > This will result in a memory leak bug.
> > 
> > Signed-off-by: Xidong Wang <wangxidong_97@163.com>
> > ---
> >  drivers/mtd/nand/tango_nand.c | 5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
> > index c5bee00b..8083459 100644
> > --- a/drivers/mtd/nand/tango_nand.c
> > +++ b/drivers/mtd/nand/tango_nand.c
> > @@ -648,12 +648,15 @@ static int tango_nand_probe(struct platform_device *pdev)
> >  		return PTR_ERR(clk);
> >  
> >  	nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
> > -	if (IS_ERR(nfc->chan))
> > +	if (IS_ERR(nfc->chan)) {
> > +		clk_put(clk);
> >  		return PTR_ERR(nfc->chan);
> > +	}
> >  
> >  	platform_set_drvdata(pdev, nfc);
> >  	nand_hw_control_init(&nfc->hw);
> >  	nfc->freq_kHz = clk_get_rate(clk) / 1000;
> > +	clk_put(clk);  
> 
> If the clock is used only here, better do the frequency derivation
> right after the clock_get(), and follow with a clk_put()? This way you
> don't have to change the error path and 'related' actions remain
> grouped.

Hm, definitely not a good idea to release the reference you have on the
clk if the driver depends on it. I recommend using devm_clk_get() to
solve this leak.
Boris Brezillon April 4, 2018, 7:08 a.m. | #3
On Wed, 4 Apr 2018 09:07:10 +0200
Boris Brezillon <boris.brezillon@bootlin.com> wrote:

> On Wed, 4 Apr 2018 08:28:07 +0200
> Miquel Raynal <miquel.raynal@bootlin.com> wrote:
> 
> > Hi Xidong,
> > 
> > As part of a reorganization in the NAND subsystem, you should now
> > prefix your commit title this way:
> > 
> >         mtd: rawnand: tango: fix memory leak
> > 
> > Not sure if this patch is candidate to cc:stable?
> > 
> > On Wed,  4 Apr 2018 11:05:51 +0800, Xidong Wang
> > <wangxidong_97@163.com> wrote:
> >   
> > > In function tango_nand_probe(), the memory allocated by
> > > clk_get() is not released on the normal path and
> > > the error path that IS_ERR(nfc->chan) returns true.    
> > 
> > The fact that the error path returns true looks out of topic, can you
> > remove it? Just saying that you fix a memory leak is enough I guess.
> >   
> > > This will result in a memory leak bug.
> > > 
> > > Signed-off-by: Xidong Wang <wangxidong_97@163.com>
> > > ---
> > >  drivers/mtd/nand/tango_nand.c | 5 ++++-
> > >  1 file changed, 4 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
> > > index c5bee00b..8083459 100644
> > > --- a/drivers/mtd/nand/tango_nand.c
> > > +++ b/drivers/mtd/nand/tango_nand.c
> > > @@ -648,12 +648,15 @@ static int tango_nand_probe(struct platform_device *pdev)
> > >  		return PTR_ERR(clk);
> > >  
> > >  	nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
> > > -	if (IS_ERR(nfc->chan))
> > > +	if (IS_ERR(nfc->chan)) {
> > > +		clk_put(clk);
> > >  		return PTR_ERR(nfc->chan);
> > > +	}
> > >  
> > >  	platform_set_drvdata(pdev, nfc);
> > >  	nand_hw_control_init(&nfc->hw);
> > >  	nfc->freq_kHz = clk_get_rate(clk) / 1000;
> > > +	clk_put(clk);    
> > 
> > If the clock is used only here, better do the frequency derivation
> > right after the clock_get(), and follow with a clk_put()? This way you
> > don't have to change the error path and 'related' actions remain
> > grouped.  
> 
> Hm, definitely not a good idea to release the reference you have on the
> clk if the driver depends on it. I recommend using devm_clk_get() to
> solve this leak.
> 

BTW, it's also weird that the driver does not prepare_enable the clk.
Mark, any comments?
Marc Gonzalez April 5, 2018, 9:12 a.m. | #4
On 04/04/2018 09:08, Boris Brezillon wrote:

> On Wed, 4 Apr 2018 09:07:10 +0200
> Boris Brezillon <boris.brezillon@bootlin.com> wrote:
> 
>> On Wed, 4 Apr 2018 08:28:07 +0200
>> Miquel Raynal <miquel.raynal@bootlin.com> wrote:
>>
>>> Hi Xidong,
>>>
>>> As part of a reorganization in the NAND subsystem, you should now
>>> prefix your commit title this way:
>>>
>>>         mtd: rawnand: tango: fix memory leak
>>>
>>> Not sure if this patch is candidate to cc:stable?
>>>
>>> On Wed,  4 Apr 2018 11:05:51 +0800, Xidong Wang
>>> <wangxidong_97@163.com> wrote:
>>>   
>>>> In function tango_nand_probe(), the memory allocated by
>>>> clk_get() is not released on the normal path and
>>>> the error path that IS_ERR(nfc->chan) returns true.    
>>>
>>> The fact that the error path returns true looks out of topic, can you
>>> remove it? Just saying that you fix a memory leak is enough I guess.
>>>   
>>>> This will result in a memory leak bug.
>>>>
>>>> Signed-off-by: Xidong Wang <wangxidong_97@163.com>
>>>> ---
>>>>  drivers/mtd/nand/tango_nand.c | 5 ++++-
>>>>  1 file changed, 4 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
>>>> index c5bee00b..8083459 100644
>>>> --- a/drivers/mtd/nand/tango_nand.c
>>>> +++ b/drivers/mtd/nand/tango_nand.c
>>>> @@ -648,12 +648,15 @@ static int tango_nand_probe(struct platform_device *pdev)
>>>>  		return PTR_ERR(clk);
>>>>  
>>>>  	nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
>>>> -	if (IS_ERR(nfc->chan))
>>>> +	if (IS_ERR(nfc->chan)) {
>>>> +		clk_put(clk);
>>>>  		return PTR_ERR(nfc->chan);
>>>> +	}
>>>>  
>>>>  	platform_set_drvdata(pdev, nfc);
>>>>  	nand_hw_control_init(&nfc->hw);
>>>>  	nfc->freq_kHz = clk_get_rate(clk) / 1000;
>>>> +	clk_put(clk);    
>>>
>>> If the clock is used only here, better do the frequency derivation
>>> right after the clock_get(), and follow with a clk_put()? This way you
>>> don't have to change the error path and 'related' actions remain
>>> grouped.  
>>
>> Hm, definitely not a good idea to release the reference you have on the
>> clk if the driver depends on it. I recommend using devm_clk_get() to
>> solve this leak.
> 
> BTW, it's also weird that the driver does not prepare_enable the clk.
> Marc, any comments?

I was not aware that clk_get() allocated memory, and required clk_put()
for cleanup. IIRC, I looked at Documentation/clk.txt

On tango, clocks are configured by the boot loader. The existing clk driver
provides only read access to various clocks -- except the CPU clock, which
can be changed by tweaking a post-divider. Tweaking the PLLs requires much
more complex code. The boot loader enables every clock, and Linux has no
way to gate any of them.

In the nfc driver, all I needed was the system frequency, since the NFC is
driven by the system clock (which can never be disabled).

Thus, I wrote the naive (and apparently incorrect)

  clk = clk_get(&pdev->dev, NULL);
  nfc->freq_kHz = clk_get_rate(clk) / 1000;


I suppose the following patch would fix the memory leak, and
matches what Miquèl suggested.

Regards.


diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
index c5bee00b7f5e..fba162af333f 100644
--- a/drivers/mtd/nand/tango_nand.c
+++ b/drivers/mtd/nand/tango_nand.c
@@ -646,6 +646,8 @@ static int tango_nand_probe(struct platform_device *pdev)
        clk = clk_get(&pdev->dev, NULL);
        if (IS_ERR(clk))
                return PTR_ERR(clk);
+       nfc->freq_kHz = clk_get_rate(clk) / 1000;
+       clk_put(clk);
 
        nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
        if (IS_ERR(nfc->chan))
@@ -653,7 +655,6 @@ static int tango_nand_probe(struct platform_device *pdev)
 
        platform_set_drvdata(pdev, nfc);
        nand_hw_control_init(&nfc->hw);
-       nfc->freq_kHz = clk_get_rate(clk) / 1000;
 
        for_each_child_of_node(pdev->dev.of_node, np) {
                err = chip_init(&pdev->dev, np);
Miquel Raynal April 5, 2018, 9:44 a.m. | #5
Hi Marc,

On Thu, 5 Apr 2018 11:12:11 +0200, Marc Gonzalez
<marc.w.gonzalez@free.fr> wrote:

> On 04/04/2018 09:08, Boris Brezillon wrote:
> 
> > On Wed, 4 Apr 2018 09:07:10 +0200
> > Boris Brezillon <boris.brezillon@bootlin.com> wrote:
> >   
> >> On Wed, 4 Apr 2018 08:28:07 +0200
> >> Miquel Raynal <miquel.raynal@bootlin.com> wrote:
> >>  
> >>> Hi Xidong,
> >>>
> >>> As part of a reorganization in the NAND subsystem, you should now
> >>> prefix your commit title this way:
> >>>
> >>>         mtd: rawnand: tango: fix memory leak
> >>>
> >>> Not sure if this patch is candidate to cc:stable?
> >>>
> >>> On Wed,  4 Apr 2018 11:05:51 +0800, Xidong Wang
> >>> <wangxidong_97@163.com> wrote:
> >>>     
> >>>> In function tango_nand_probe(), the memory allocated by
> >>>> clk_get() is not released on the normal path and
> >>>> the error path that IS_ERR(nfc->chan) returns true.      
> >>>
> >>> The fact that the error path returns true looks out of topic, can you
> >>> remove it? Just saying that you fix a memory leak is enough I guess.
> >>>     
> >>>> This will result in a memory leak bug.
> >>>>
> >>>> Signed-off-by: Xidong Wang <wangxidong_97@163.com>
> >>>> ---
> >>>>  drivers/mtd/nand/tango_nand.c | 5 ++++-
> >>>>  1 file changed, 4 insertions(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
> >>>> index c5bee00b..8083459 100644
> >>>> --- a/drivers/mtd/nand/tango_nand.c
> >>>> +++ b/drivers/mtd/nand/tango_nand.c
> >>>> @@ -648,12 +648,15 @@ static int tango_nand_probe(struct platform_device *pdev)
> >>>>  		return PTR_ERR(clk);
> >>>>  
> >>>>  	nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
> >>>> -	if (IS_ERR(nfc->chan))
> >>>> +	if (IS_ERR(nfc->chan)) {
> >>>> +		clk_put(clk);
> >>>>  		return PTR_ERR(nfc->chan);
> >>>> +	}
> >>>>  
> >>>>  	platform_set_drvdata(pdev, nfc);
> >>>>  	nand_hw_control_init(&nfc->hw);
> >>>>  	nfc->freq_kHz = clk_get_rate(clk) / 1000;
> >>>> +	clk_put(clk);      
> >>>
> >>> If the clock is used only here, better do the frequency derivation
> >>> right after the clock_get(), and follow with a clk_put()? This way you
> >>> don't have to change the error path and 'related' actions remain
> >>> grouped.    
> >>
> >> Hm, definitely not a good idea to release the reference you have on the
> >> clk if the driver depends on it. I recommend using devm_clk_get() to
> >> solve this leak.  
> > 
> > BTW, it's also weird that the driver does not prepare_enable the clk.
> > Marc, any comments?  
> 
> I was not aware that clk_get() allocated memory, and required clk_put()
> for cleanup. IIRC, I looked at Documentation/clk.txt

I ignored there was an actual leak too, but the 'struct clk' seems to
be allocated here [1] (cascaded calls from clk_get()) and freed here
[2].

[1] https://elixir.bootlin.com/linux/latest/source/drivers/clk/clk.c#L3044
[2] https://elixir.bootlin.com/linux/latest/source/drivers/clk/clk.c#L3472

> 
> On tango, clocks are configured by the boot loader. The existing clk driver
> provides only read access to various clocks -- except the CPU clock, which
> can be changed by tweaking a post-divider. Tweaking the PLLs requires much
> more complex code. The boot loader enables every clock, and Linux has no
> way to gate any of them.
> 
> In the nfc driver, all I needed was the system frequency, since the NFC is
> driven by the system clock (which can never be disabled).
> 
> Thus, I wrote the naive (and apparently incorrect)
> 
>   clk = clk_get(&pdev->dev, NULL);
>   nfc->freq_kHz = clk_get_rate(clk) / 1000;
> 
> 
> I suppose the following patch would fix the memory leak, and
> matches what Miquèl suggested.

Boris can you confirm:
1/ there is no need to enable the clock from this driver (from the API
   point of view) before the clk_get_rate()?
2/ there is no risk to do the clkd_put() right after instead of keeping
   it until a potential __exit?

Thanks,
Miquèl
Boris Brezillon April 5, 2018, 9:54 a.m. | #6
On Thu, 5 Apr 2018 11:12:11 +0200
Marc Gonzalez <marc.w.gonzalez@free.fr> wrote:

> On 04/04/2018 09:08, Boris Brezillon wrote:
> 
> > On Wed, 4 Apr 2018 09:07:10 +0200
> > Boris Brezillon <boris.brezillon@bootlin.com> wrote:
> >   
> >> On Wed, 4 Apr 2018 08:28:07 +0200
> >> Miquel Raynal <miquel.raynal@bootlin.com> wrote:
> >>  
> >>> Hi Xidong,
> >>>
> >>> As part of a reorganization in the NAND subsystem, you should now
> >>> prefix your commit title this way:
> >>>
> >>>         mtd: rawnand: tango: fix memory leak
> >>>
> >>> Not sure if this patch is candidate to cc:stable?
> >>>
> >>> On Wed,  4 Apr 2018 11:05:51 +0800, Xidong Wang
> >>> <wangxidong_97@163.com> wrote:
> >>>     
> >>>> In function tango_nand_probe(), the memory allocated by
> >>>> clk_get() is not released on the normal path and
> >>>> the error path that IS_ERR(nfc->chan) returns true.      
> >>>
> >>> The fact that the error path returns true looks out of topic, can you
> >>> remove it? Just saying that you fix a memory leak is enough I guess.
> >>>     
> >>>> This will result in a memory leak bug.
> >>>>
> >>>> Signed-off-by: Xidong Wang <wangxidong_97@163.com>
> >>>> ---
> >>>>  drivers/mtd/nand/tango_nand.c | 5 ++++-
> >>>>  1 file changed, 4 insertions(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
> >>>> index c5bee00b..8083459 100644
> >>>> --- a/drivers/mtd/nand/tango_nand.c
> >>>> +++ b/drivers/mtd/nand/tango_nand.c
> >>>> @@ -648,12 +648,15 @@ static int tango_nand_probe(struct platform_device *pdev)
> >>>>  		return PTR_ERR(clk);
> >>>>  
> >>>>  	nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
> >>>> -	if (IS_ERR(nfc->chan))
> >>>> +	if (IS_ERR(nfc->chan)) {
> >>>> +		clk_put(clk);
> >>>>  		return PTR_ERR(nfc->chan);
> >>>> +	}
> >>>>  
> >>>>  	platform_set_drvdata(pdev, nfc);
> >>>>  	nand_hw_control_init(&nfc->hw);
> >>>>  	nfc->freq_kHz = clk_get_rate(clk) / 1000;
> >>>> +	clk_put(clk);      
> >>>
> >>> If the clock is used only here, better do the frequency derivation
> >>> right after the clock_get(), and follow with a clk_put()? This way you
> >>> don't have to change the error path and 'related' actions remain
> >>> grouped.    
> >>
> >> Hm, definitely not a good idea to release the reference you have on the
> >> clk if the driver depends on it. I recommend using devm_clk_get() to
> >> solve this leak.  
> > 
> > BTW, it's also weird that the driver does not prepare_enable the clk.
> > Marc, any comments?  
> 
> I was not aware that clk_get() allocated memory, and required clk_put()
> for cleanup. IIRC, I looked at Documentation/clk.txt
> 
> On tango, clocks are configured by the boot loader. The existing clk driver
> provides only read access to various clocks -- except the CPU clock, which
> can be changed by tweaking a post-divider. Tweaking the PLLs requires much
> more complex code. The boot loader enables every clock, and Linux has no
> way to gate any of them.

Well, even if that's not supported today, it's always a good practice
to retain reference and prepare/enable clks your HW depends on. This
change should be harmless and when/if you someday decide to provide a
way to gate clks, it will work out of the box.

> 
> In the nfc driver, all I needed was the system frequency, since the NFC is
> driven by the system clock (which can never be disabled).
> 
> Thus, I wrote the naive (and apparently incorrect)
> 
>   clk = clk_get(&pdev->dev, NULL);
>   nfc->freq_kHz = clk_get_rate(clk) / 1000;
> 
> 
> I suppose the following patch would fix the memory leak, and
> matches what Miquèl suggested.
> 
> Regards.
> 
> 
> diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
> index c5bee00b7f5e..fba162af333f 100644
> --- a/drivers/mtd/nand/tango_nand.c
> +++ b/drivers/mtd/nand/tango_nand.c
> @@ -646,6 +646,8 @@ static int tango_nand_probe(struct platform_device *pdev)
>         clk = clk_get(&pdev->dev, NULL);

Why not using devm_clk_get() and be done with it?

>         if (IS_ERR(clk))
>                 return PTR_ERR(clk);
> +       nfc->freq_kHz = clk_get_rate(clk) / 1000;
> +       clk_put(clk);

And that's where I disagree. Clearly, you're not following one of the
clk consumer's rule: "when you need a clk, keep a reference to it and
enable it before you start using the HW".

>  
>         nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
>         if (IS_ERR(nfc->chan))
> @@ -653,7 +655,6 @@ static int tango_nand_probe(struct platform_device *pdev)
>  
>         platform_set_drvdata(pdev, nfc);
>         nand_hw_control_init(&nfc->hw);
> -       nfc->freq_kHz = clk_get_rate(clk) / 1000;
>  
>         for_each_child_of_node(pdev->dev.of_node, np) {
>                 err = chip_init(&pdev->dev, np);
Boris Brezillon April 5, 2018, 11:04 a.m. | #7
On Thu, 5 Apr 2018 11:44:10 +0200
Miquel Raynal <miquel.raynal@bootlin.com> wrote:

> Hi Marc,
> 
> On Thu, 5 Apr 2018 11:12:11 +0200, Marc Gonzalez
> <marc.w.gonzalez@free.fr> wrote:
> 
> > On 04/04/2018 09:08, Boris Brezillon wrote:
> >   
> > > On Wed, 4 Apr 2018 09:07:10 +0200
> > > Boris Brezillon <boris.brezillon@bootlin.com> wrote:
> > >     
> > >> On Wed, 4 Apr 2018 08:28:07 +0200
> > >> Miquel Raynal <miquel.raynal@bootlin.com> wrote:
> > >>    
> > >>> Hi Xidong,
> > >>>
> > >>> As part of a reorganization in the NAND subsystem, you should now
> > >>> prefix your commit title this way:
> > >>>
> > >>>         mtd: rawnand: tango: fix memory leak
> > >>>
> > >>> Not sure if this patch is candidate to cc:stable?
> > >>>
> > >>> On Wed,  4 Apr 2018 11:05:51 +0800, Xidong Wang
> > >>> <wangxidong_97@163.com> wrote:
> > >>>       
> > >>>> In function tango_nand_probe(), the memory allocated by
> > >>>> clk_get() is not released on the normal path and
> > >>>> the error path that IS_ERR(nfc->chan) returns true.        
> > >>>
> > >>> The fact that the error path returns true looks out of topic, can you
> > >>> remove it? Just saying that you fix a memory leak is enough I guess.
> > >>>       
> > >>>> This will result in a memory leak bug.
> > >>>>
> > >>>> Signed-off-by: Xidong Wang <wangxidong_97@163.com>
> > >>>> ---
> > >>>>  drivers/mtd/nand/tango_nand.c | 5 ++++-
> > >>>>  1 file changed, 4 insertions(+), 1 deletion(-)
> > >>>>
> > >>>> diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
> > >>>> index c5bee00b..8083459 100644
> > >>>> --- a/drivers/mtd/nand/tango_nand.c
> > >>>> +++ b/drivers/mtd/nand/tango_nand.c
> > >>>> @@ -648,12 +648,15 @@ static int tango_nand_probe(struct platform_device *pdev)
> > >>>>  		return PTR_ERR(clk);
> > >>>>  
> > >>>>  	nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
> > >>>> -	if (IS_ERR(nfc->chan))
> > >>>> +	if (IS_ERR(nfc->chan)) {
> > >>>> +		clk_put(clk);
> > >>>>  		return PTR_ERR(nfc->chan);
> > >>>> +	}
> > >>>>  
> > >>>>  	platform_set_drvdata(pdev, nfc);
> > >>>>  	nand_hw_control_init(&nfc->hw);
> > >>>>  	nfc->freq_kHz = clk_get_rate(clk) / 1000;
> > >>>> +	clk_put(clk);        
> > >>>
> > >>> If the clock is used only here, better do the frequency derivation
> > >>> right after the clock_get(), and follow with a clk_put()? This way you
> > >>> don't have to change the error path and 'related' actions remain
> > >>> grouped.      
> > >>
> > >> Hm, definitely not a good idea to release the reference you have on the
> > >> clk if the driver depends on it. I recommend using devm_clk_get() to
> > >> solve this leak.    
> > > 
> > > BTW, it's also weird that the driver does not prepare_enable the clk.
> > > Marc, any comments?    
> > 
> > I was not aware that clk_get() allocated memory, and required clk_put()
> > for cleanup. IIRC, I looked at Documentation/clk.txt  
> 
> I ignored there was an actual leak too, but the 'struct clk' seems to
> be allocated here [1] (cascaded calls from clk_get()) and freed here
> [2].
> 
> [1] https://elixir.bootlin.com/linux/latest/source/drivers/clk/clk.c#L3044
> [2] https://elixir.bootlin.com/linux/latest/source/drivers/clk/clk.c#L3472
> 
> > 
> > On tango, clocks are configured by the boot loader. The existing clk driver
> > provides only read access to various clocks -- except the CPU clock, which
> > can be changed by tweaking a post-divider. Tweaking the PLLs requires much
> > more complex code. The boot loader enables every clock, and Linux has no
> > way to gate any of them.
> > 
> > In the nfc driver, all I needed was the system frequency, since the NFC is
> > driven by the system clock (which can never be disabled).
> > 
> > Thus, I wrote the naive (and apparently incorrect)
> > 
> >   clk = clk_get(&pdev->dev, NULL);
> >   nfc->freq_kHz = clk_get_rate(clk) / 1000;
> > 
> > 
> > I suppose the following patch would fix the memory leak, and
> > matches what Miquèl suggested.  
> 
> Boris can you confirm:
> 1/ there is no need to enable the clock from this driver (from the API
>    point of view) before the clk_get_rate()?

It's not strictly required, but I'd recommend doing it. Not necessarily
before enabling the clk though.

> 2/ there is no risk to do the clkd_put() right after instead of keeping
>    it until a potential __exit?

It's not a good idea to do that, especially since devm_clk_get() can
release the clk for you when the device is destroyed.
Marc Gonzalez April 5, 2018, 11:26 a.m. | #8
On 05/04/2018 11:54, Boris Brezillon wrote:

> On Thu, 5 Apr 2018 11:12:11 +0200, Marc Gonzalez wrote:
> 
>> I was not aware that clk_get() allocated memory, and required clk_put()
>> for cleanup. IIRC, I looked at Documentation/clk.txt
>>
>> On tango, clocks are configured by the boot loader. The existing clk driver
>> provides only read access to various clocks -- except the CPU clock, which
>> can be changed by tweaking a post-divider. Tweaking the PLLs requires much
>> more complex code. The boot loader enables every clock, and Linux has no
>> way to gate any of them.
> 
> Well, even if that's not supported today, it's always a good practice
> to retain reference and prepare/enable clks your HW depends on. This
> change should be harmless and when/if you someday decide to provide a
> way to gate clks, it will work out of the box.

IIUC, you're saying:
1) use devm_clk_get() instead of clk_get() to solve the memory leak
2) call clk_prepare_enable() before clk_get_rate() even if the former is a no-op today

The following patch implements these suggestions.
(Only compile-tested)

diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
index c5bee00b7f5e..39d190b7521f 100644
--- a/drivers/mtd/nand/tango_nand.c
+++ b/drivers/mtd/nand/tango_nand.c
@@ -643,7 +643,7 @@ static int tango_nand_probe(struct platform_device *pdev)
 
        writel_relaxed(MODE_RAW, nfc->pbus_base + PBUS_PAD_MODE);
 
-       clk = clk_get(&pdev->dev, NULL);
+       clk = devm_clk_get(&pdev->dev, NULL);
        if (IS_ERR(clk))
                return PTR_ERR(clk);
 
@@ -651,6 +651,10 @@ static int tango_nand_probe(struct platform_device *pdev)
        if (IS_ERR(nfc->chan))
                return PTR_ERR(nfc->chan);
 
+       err = clk_prepare_enable(clk);
+       if (err)
+               return err;
+
        platform_set_drvdata(pdev, nfc);
        nand_hw_control_init(&nfc->hw);
        nfc->freq_kHz = clk_get_rate(clk) / 1000;
Boris Brezillon April 5, 2018, 11:47 a.m. | #9
On Thu, 5 Apr 2018 13:26:31 +0200
Marc Gonzalez <marc.w.gonzalez@free.fr> wrote:

> On 05/04/2018 11:54, Boris Brezillon wrote:
> 
> > On Thu, 5 Apr 2018 11:12:11 +0200, Marc Gonzalez wrote:
> >   
> >> I was not aware that clk_get() allocated memory, and required clk_put()
> >> for cleanup. IIRC, I looked at Documentation/clk.txt
> >>
> >> On tango, clocks are configured by the boot loader. The existing clk driver
> >> provides only read access to various clocks -- except the CPU clock, which
> >> can be changed by tweaking a post-divider. Tweaking the PLLs requires much
> >> more complex code. The boot loader enables every clock, and Linux has no
> >> way to gate any of them.  
> > 
> > Well, even if that's not supported today, it's always a good practice
> > to retain reference and prepare/enable clks your HW depends on. This
> > change should be harmless and when/if you someday decide to provide a
> > way to gate clks, it will work out of the box.  
> 
> IIUC, you're saying:
> 1) use devm_clk_get() instead of clk_get() to solve the memory leak
> 2) call clk_prepare_enable() before clk_get_rate() even if the former is a no-op today

+ a disable_unprepare() in the remove path and a another one in the
error path (in case something fails after prepare_enable() has been
called).

> 
> The following patch implements these suggestions.
> (Only compile-tested)
> 
> diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
> index c5bee00b7f5e..39d190b7521f 100644
> --- a/drivers/mtd/nand/tango_nand.c
> +++ b/drivers/mtd/nand/tango_nand.c
> @@ -643,7 +643,7 @@ static int tango_nand_probe(struct platform_device *pdev)
>  
>         writel_relaxed(MODE_RAW, nfc->pbus_base + PBUS_PAD_MODE);
>  
> -       clk = clk_get(&pdev->dev, NULL);
> +       clk = devm_clk_get(&pdev->dev, NULL);
>         if (IS_ERR(clk))
>                 return PTR_ERR(clk);
>  
> @@ -651,6 +651,10 @@ static int tango_nand_probe(struct platform_device *pdev)
>         if (IS_ERR(nfc->chan))
>                 return PTR_ERR(nfc->chan);
>  
> +       err = clk_prepare_enable(clk);
> +       if (err)
> +               return err;
> +
>         platform_set_drvdata(pdev, nfc);
>         nand_hw_control_init(&nfc->hw);
>         nfc->freq_kHz = clk_get_rate(clk) / 1000;
Marc Gonzalez April 5, 2018, noon | #10
On 05/04/2018 13:47, Boris Brezillon wrote:

> On Thu, 5 Apr 2018 13:26:31 +0200, Marc Gonzalez wrote:
> 
>> IIUC, you're saying:
>> 1) use devm_clk_get() instead of clk_get() to solve the memory leak
>> 2) call clk_prepare_enable() before clk_get_rate() even if the former is a no-op today
> 
> + a disable_unprepare() in the remove path and a another one in the
> error path (in case something fails after prepare_enable() has been
> called).

I hate C error-handling with all my heart...

Why is it taking so long for Linux APIs to be turned into devm
managed functions? It makes life so much easier...

Patch

diff --git a/drivers/mtd/nand/tango_nand.c b/drivers/mtd/nand/tango_nand.c
index c5bee00b..8083459 100644
--- a/drivers/mtd/nand/tango_nand.c
+++ b/drivers/mtd/nand/tango_nand.c
@@ -648,12 +648,15 @@  static int tango_nand_probe(struct platform_device *pdev)
 		return PTR_ERR(clk);
 
 	nfc->chan = dma_request_chan(&pdev->dev, "rxtx");
-	if (IS_ERR(nfc->chan))
+	if (IS_ERR(nfc->chan)) {
+		clk_put(clk);
 		return PTR_ERR(nfc->chan);
+	}
 
 	platform_set_drvdata(pdev, nfc);
 	nand_hw_control_init(&nfc->hw);
 	nfc->freq_kHz = clk_get_rate(clk) / 1000;
+	clk_put(clk);
 
 	for_each_child_of_node(pdev->dev.of_node, np) {
 		err = chip_init(&pdev->dev, np);