[PATCH -next 1/3] PCI: fix double put_device() in error case in pci_create_root_bus()

[PATCH -next 1/3] PCI: fix double put_device() in error case in pci_create_root_bus()

[Yang Yingliang]

If device_add() fails in pci_register_host_bridge(), the brigde device will
be put once, and it will be put again in error path of pci_create_root_bus().
Move the put_device() from pci_create_root_bus() to pci_register_host_bridge()
to fix this problem. And use device_unregister() instead of del_device() and
put_device().

Fixes: 9885440b16b8 ("PCI: Fix pci_host_bridge struct device release/free handling")
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
---
 drivers/pci/probe.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)

diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index c5286b027f00..e500eb9d6468 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -1027,7 +1027,7 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
 
 unregister:
 	put_device(&bridge->dev);
-	device_del(&bridge->dev);
+	device_unregister(&bridge->dev);
 
 free:
 	kfree(bus);
@@ -3037,13 +3037,9 @@ struct pci_bus *pci_create_root_bus(struct device *parent, int bus,
 
 	error = pci_register_host_bridge(bridge);
 	if (error < 0)
-		goto err_out;
+		return NULL;
 
 	return bridge->bus;
-
-err_out:
-	put_device(&bridge->dev);
-	return NULL;
 }
 EXPORT_SYMBOL_GPL(pci_create_root_bus);
 
-- 
2.25.1

[PATCH -next 2/3] PCI: fix possible memory leak in error case in pci_register_host_bridge()

[Yang Yingliang]

If device_register() fails in pci_register_host_bridge(), the refcount
of bus device is leaked, so device name that set by dev_set_name() can
not be freed. Fix this by calling put_device() when device_register()
fails, so the device name will be freed in kobject_cleanup().

Fixes: 37d6a0a6f470 ("PCI: Add pci_register_host_bridge() interface")
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
---
 drivers/pci/probe.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index e500eb9d6468..292d9da146ce 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -948,8 +948,17 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
 	name = dev_name(&bus->dev);
 
 	err = device_register(&bus->dev);
-	if (err)
-		goto unregister;
+	if (err) {
+		/*
+		 * release_pcibus_dev() will decrease the refcount of bridge
+		 * device and free the memory of bus.
+		 * The memory of bus device name will be freed when the refcount
+		 * get to zero.
+		 */
+		put_device(&bus->dev);
+		device_unregister(&bridge->dev);
+		return err;
+	}
 
 	pcibios_add_bus(bus);
 
@@ -1025,10 +1034,6 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
 
 	return 0;
 
-unregister:
-	put_device(&bridge->dev);
-	device_unregister(&bridge->dev);
-
 free:
 	kfree(bus);
 	return err;
-- 
2.25.1

[PATCH -next 3/3] PCI: fix handle error case in pci_alloc_child_bus()

[Yang Yingliang]

Return NULL pointer if device_register() fails, and call put_device()
to free the memory of pci bus and device name.

Fixes: 4f535093cf8f ("PCI: Put pci_dev in device tree as early as possible")
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
---
 drivers/pci/probe.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index 292d9da146ce..c924f4e1ed38 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -1144,7 +1144,10 @@ static struct pci_bus *pci_alloc_child_bus(struct pci_bus *parent,
 add_dev:
 	pci_set_bus_msi_domain(child);
 	ret = device_register(&child->dev);
-	WARN_ON(ret < 0);
+	if (WARN_ON(ret < 0)) {
+		put_device(&child->dev);
+		return NULL;
+	}
 
 	pcibios_add_bus(child);
 
-- 
2.25.1

Re: [PATCH -next 1/3] PCI: fix double put_device() in error case in pci_create_root_bus()

[Bjorn Helgaas]

[+cc Rob]

On Thu, Aug 25, 2022 at 08:27:51PM +0800, Yang Yingliang wrote:
> If device_add() fails in pci_register_host_bridge(), the brigde device will
> be put once, and it will be put again in error path of pci_create_root_bus().
> Move the put_device() from pci_create_root_bus() to pci_register_host_bridge()
> to fix this problem. And use device_unregister() instead of del_device() and
> put_device().

s/brigde/bridge/

> Fixes: 9885440b16b8 ("PCI: Fix pci_host_bridge struct device release/free handling")

If you're fixing a commit from somebody else, please always cc: the
author because the author can help review the change.

> Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
> ---
>  drivers/pci/probe.c | 8 ++------
>  1 file changed, 2 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index c5286b027f00..e500eb9d6468 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -1027,7 +1027,7 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
>  
>  unregister:
>  	put_device(&bridge->dev);
> -	device_del(&bridge->dev);
> +	device_unregister(&bridge->dev);

I don't understand this part.  device_unregister() looks like this:

  void device_unregister(struct device *dev)
  {
    device_del(dev);
    put_device(dev);
  }

So this calls put_device(&bridge->dev) twice, doesn't it?

The "unregister" label looks poorly named.  We only get there if
device_register() *failed*.  We shouldn't need to unregister anything
in that case.

>  free:
>  	kfree(bus);
> @@ -3037,13 +3037,9 @@ struct pci_bus *pci_create_root_bus(struct device *parent, int bus,
>  
>  	error = pci_register_host_bridge(bridge);
>  	if (error < 0)
> -		goto err_out;
> +		return NULL;
>  
>  	return bridge->bus;
> -
> -err_out:
> -	put_device(&bridge->dev);
> -	return NULL;

This part looks right to me.  The get_device() is in
pci_register_host_bridge(), and if pci_register_host_bridge() returns
failure, I think it should first do the corresponding put_device().

>  }
>  EXPORT_SYMBOL_GPL(pci_create_root_bus);
>  
> -- 
> 2.25.1
> 

Re: [PATCH -next 2/3] PCI: fix possible memory leak in error case in pci_register_host_bridge()

[Bjorn Helgaas]

[+cc Arnd, Rob]

On Thu, Aug 25, 2022 at 08:27:52PM +0800, Yang Yingliang wrote:
> If device_register() fails in pci_register_host_bridge(), the refcount
> of bus device is leaked, so device name that set by dev_set_name() can
> not be freed. Fix this by calling put_device() when device_register()
> fails, so the device name will be freed in kobject_cleanup().
>
> Fixes: 37d6a0a6f470 ("PCI: Add pci_register_host_bridge() interface")
> Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
> ---
>  drivers/pci/probe.c | 17 +++++++++++------
>  1 file changed, 11 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index e500eb9d6468..292d9da146ce 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -948,8 +948,17 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
>  	name = dev_name(&bus->dev);
>  
>  	err = device_register(&bus->dev);
> -	if (err)
> -		goto unregister;
> +	if (err) {
> +		/*
> +		 * release_pcibus_dev() will decrease the refcount of bridge
> +		 * device and free the memory of bus.
> +		 * The memory of bus device name will be freed when the refcount
> +		 * get to zero.
> +		 */
> +		put_device(&bus->dev);
> +		device_unregister(&bridge->dev);
> +		return err;
> +	}

Calling put_device(X) after device_register(X) returns failure doesn't
need explanation because that's the standard pattern.  I think that
was just missing before.

In this error case, we previously did called put_device() for the
*bridge* instead of the bus.  That was likely a typo and seems like
the important thing here.

>  	pcibios_add_bus(bus);
>  
> @@ -1025,10 +1034,6 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
>  
>  	return 0;
>  
> -unregister:
> -	put_device(&bridge->dev);
> -	device_unregister(&bridge->dev);
> -
>  free:
>  	kfree(bus);
>  	return err;
> -- 
> 2.25.1
> 

Re: [PATCH -next 1/3] PCI: fix double put_device() in error case in pci_create_root_bus()

[Yang Yingliang]

Hi,

On 2022/8/27 5:14, Bjorn Helgaas wrote:
> [+cc Rob]
>
> On Thu, Aug 25, 2022 at 08:27:51PM +0800, Yang Yingliang wrote:
>> If device_add() fails in pci_register_host_bridge(), the brigde device will
>> be put once, and it will be put again in error path of pci_create_root_bus().
>> Move the put_device() from pci_create_root_bus() to pci_register_host_bridge()
>> to fix this problem. And use device_unregister() instead of del_device() and
>> put_device().
> s/brigde/bridge/
>
>> Fixes: 9885440b16b8 ("PCI: Fix pci_host_bridge struct device release/free handling")
> If you're fixing a commit from somebody else, please always cc: the
> author because the author can help review the change.
OK.
>
>> Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
>> ---
>>   drivers/pci/probe.c | 8 ++------
>>   1 file changed, 2 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
>> index c5286b027f00..e500eb9d6468 100644
>> --- a/drivers/pci/probe.c
>> +++ b/drivers/pci/probe.c
>> @@ -1027,7 +1027,7 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
>>   
>>   unregister:
>>   	put_device(&bridge->dev);
>> -	device_del(&bridge->dev);
>> +	device_unregister(&bridge->dev);
> I don't understand this part.  device_unregister() looks like this:
>
>    void device_unregister(struct device *dev)
>    {
>      device_del(dev);
>      put_device(dev);
>    }
>
> So this calls put_device(&bridge->dev) twice, doesn't it?
>
> The "unregister" label looks poorly named.  We only get there if
> device_register() *failed*.  We shouldn't need to unregister anything
> in that case.
If it goes to the 'unregister' label, the bridge->dev has been register 
sucessfully (device_initialize() called from pci_alloc_host_bridge()
and device_add() called from pci_register_host_bridge()), so it need be 
unregister, and another put_device() is for decreasing
refcount of 'bus->bridge'.

Thanks,
Yang
>
>>   free:
>>   	kfree(bus);
>> @@ -3037,13 +3037,9 @@ struct pci_bus *pci_create_root_bus(struct device *parent, int bus,
>>   
>>   	error = pci_register_host_bridge(bridge);
>>   	if (error < 0)
>> -		goto err_out;
>> +		return NULL;
>>   
>>   	return bridge->bus;
>> -
>> -err_out:
>> -	put_device(&bridge->dev);
>> -	return NULL;
> This part looks right to me.  The get_device() is in
> pci_register_host_bridge(), and if pci_register_host_bridge() returns
> failure, I think it should first do the corresponding put_device().
>
>>   }
>>   EXPORT_SYMBOL_GPL(pci_create_root_bus);
>>   
>> -- 
>> 2.25.1
>>
> .

Re: [PATCH -next 2/3] PCI: fix possible memory leak in error case in pci_register_host_bridge()

[Yang Yingliang]

On 2022/8/27 6:38, Bjorn Helgaas wrote:
> [+cc Arnd, Rob]
>
> On Thu, Aug 25, 2022 at 08:27:52PM +0800, Yang Yingliang wrote:
>> If device_register() fails in pci_register_host_bridge(), the refcount
>> of bus device is leaked, so device name that set by dev_set_name() can
>> not be freed. Fix this by calling put_device() when device_register()
>> fails, so the device name will be freed in kobject_cleanup().
>>
>> Fixes: 37d6a0a6f470 ("PCI: Add pci_register_host_bridge() interface")
>> Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
>> ---
>>   drivers/pci/probe.c | 17 +++++++++++------
>>   1 file changed, 11 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
>> index e500eb9d6468..292d9da146ce 100644
>> --- a/drivers/pci/probe.c
>> +++ b/drivers/pci/probe.c
>> @@ -948,8 +948,17 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
>>   	name = dev_name(&bus->dev);
>>   
>>   	err = device_register(&bus->dev);
>> -	if (err)
>> -		goto unregister;
>> +	if (err) {
>> +		/*
>> +		 * release_pcibus_dev() will decrease the refcount of bridge
>> +		 * device and free the memory of bus.
>> +		 * The memory of bus device name will be freed when the refcount
>> +		 * get to zero.
>> +		 */
>> +		put_device(&bus->dev);
>> +		device_unregister(&bridge->dev);
>> +		return err;
>> +	}
> Calling put_device(X) after device_register(X) returns failure doesn't
> need explanation because that's the standard pattern.  I think that
> was just missing before.
>
> In this error case, we previously did called put_device() for the
> *bridge* instead of the bus.  That was likely a typo and seems like
> the important thing here.
put_device() for the bridge will be called in the callback of put for 
the bus.
So it doesn't call put bridge device here.

Thanks,
Yang
>>   	pcibios_add_bus(bus);
>>   
>> @@ -1025,10 +1034,6 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
>>   
>>   	return 0;
>>   
>> -unregister:
>> -	put_device(&bridge->dev);
>> -	device_unregister(&bridge->dev);
>> -
>>   free:
>>   	kfree(bus);
>>   	return err;
>> -- 
>> 2.25.1
>>
> .