[PATCH] software node: recursively unregister child swnodes

[PATCH] software node: recursively unregister child swnodes

[jorhand]

From: Jordan Hand <jorhand@linux.microsoft.com>

If a child swnode is unregistered after it's parent, it can lead to
undefined behavior.

When a swnode is unregistered, recursively free it's children to avoid
this condition.

Signed-off-by: Jordan Hand <jorhand@linux.microsoft.com>
---
 drivers/base/swnode.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/drivers/base/swnode.c b/drivers/base/swnode.c
index e5eb27375416..e63093b1542b 100644
--- a/drivers/base/swnode.c
+++ b/drivers/base/swnode.c
@@ -619,6 +619,8 @@ static void software_node_release(struct kobject *kobj)
 		property_entries_free(swnode->node->properties);
 		kfree(swnode->node);
 	}
+
+	list_del(&kobj->entry);
 	ida_destroy(&swnode->child_ids);
 	kfree(swnode);
 }
@@ -712,11 +714,6 @@ EXPORT_SYMBOL_GPL(software_node_register_nodes);
  * @nodes: Zero terminated array of software nodes to be unregistered
  *
  * Unregister multiple software nodes at once.
- *
- * NOTE: Be careful using this call if the nodes had parent pointers set up in
- * them before registering.  If so, it is wiser to remove the nodes
- * individually, in the correct order (child before parent) instead of relying
- * on the sequential order of the list of nodes in the array.
  */
 void software_node_unregister_nodes(const struct software_node *nodes)
 {
@@ -839,10 +836,16 @@ EXPORT_SYMBOL_GPL(fwnode_create_software_node);
 void fwnode_remove_software_node(struct fwnode_handle *fwnode)
 {
 	struct swnode *swnode = to_swnode(fwnode);
+	struct swnode *child = NULL;
 
 	if (!swnode)
 		return;
 
+	while (!list_empty(&swnode->children)) {
+		child = list_first_entry_or_null(&swnode->children, struct swnode, entry);
+		fwnode_remove_software_node(&child->fwnode);
+	}
+
 	kobject_put(&swnode->kobj);
 }
 EXPORT_SYMBOL_GPL(fwnode_remove_software_node);
-- 
2.17.1

Re: [PATCH] software node: recursively unregister child swnodes

[Greg Kroah-Hartman]

On Thu, Jun 04, 2020 at 12:36:23PM -0700, jorhand@linux.microsoft.com wrote:
> From: Jordan Hand <jorhand@linux.microsoft.com>
> 
> If a child swnode is unregistered after it's parent, it can lead to
> undefined behavior.

Crashing the system is not really "undefined" :)

> When a swnode is unregistered, recursively free it's children to avoid
> this condition.

Are you sure?  Why would you be unregistering a child after it's parent?
Why not just do not do that?

> 
> Signed-off-by: Jordan Hand <jorhand@linux.microsoft.com>
> ---
>  drivers/base/swnode.c | 13 ++++++++-----
>  1 file changed, 8 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/base/swnode.c b/drivers/base/swnode.c
> index e5eb27375416..e63093b1542b 100644
> --- a/drivers/base/swnode.c
> +++ b/drivers/base/swnode.c
> @@ -619,6 +619,8 @@ static void software_node_release(struct kobject *kobj)
>  		property_entries_free(swnode->node->properties);
>  		kfree(swnode->node);
>  	}
> +
> +	list_del(&kobj->entry);
>  	ida_destroy(&swnode->child_ids);
>  	kfree(swnode);
>  }
> @@ -712,11 +714,6 @@ EXPORT_SYMBOL_GPL(software_node_register_nodes);
>   * @nodes: Zero terminated array of software nodes to be unregistered
>   *
>   * Unregister multiple software nodes at once.
> - *
> - * NOTE: Be careful using this call if the nodes had parent pointers set up in
> - * them before registering.  If so, it is wiser to remove the nodes
> - * individually, in the correct order (child before parent) instead of relying
> - * on the sequential order of the list of nodes in the array.
>   */
>  void software_node_unregister_nodes(const struct software_node *nodes)
>  {
> @@ -839,10 +836,16 @@ EXPORT_SYMBOL_GPL(fwnode_create_software_node);
>  void fwnode_remove_software_node(struct fwnode_handle *fwnode)
>  {
>  	struct swnode *swnode = to_swnode(fwnode);
> +	struct swnode *child = NULL;
>  
>  	if (!swnode)
>  		return;
>  
> +	while (!list_empty(&swnode->children)) {
> +		child = list_first_entry_or_null(&swnode->children, struct swnode, entry);
> +		fwnode_remove_software_node(&child->fwnode);

You should document that you just changed the behaivor here, as you are
now really doing fwnode_remove_software_node_and_all_children().

but again, why?  Who wants to unregister a child before a parent?

thanks,

greg k-h

Re: [PATCH] software node: recursively unregister child swnodes

[Jordan Hand]

On 6/4/20 1:15 PM, Greg Kroah-Hartman wrote:
> On Thu, Jun 04, 2020 at 12:36:23PM -0700, jorhand@linux.microsoft.com wrote:
>> From: Jordan Hand <jorhand@linux.microsoft.com>
>>
>> If a child swnode is unregistered after it's parent, it can lead to
>> undefined behavior.
> 
> Crashing the system is not really "undefined" :)

Fair point :)

> 
>> When a swnode is unregistered, recursively free it's children to avoid
>> this condition.
> 
> Are you sure?  Why would you be unregistering a child after it's parent?
> Why not just do not do that?
> 

The main motivation for doing this was to support 
`software_node_unregister_nodes` so that the passed list of nodes does 
not need to be ordered in any particular way.

I suppose another way to do this would be to add a new function 
`fwnode_remove_software_node_recursive` and just call that from 
`software_node_unregister_nodes`.

That said, I suppose just ordering the nodes so that children come 
before parents would also be fine. My thinking was just that accepting 
any node ordering is simpler.

Thanks,
Jordan

Re: [PATCH] software node: recursively unregister child swnodes

[Jordan Hand]

On 6/4/20 1:57 PM, Jordan Hand wrote:
> On 6/4/20 1:15 PM, Greg Kroah-Hartman wrote:
>> On Thu, Jun 04, 2020 at 12:36:23PM -0700, jorhand@linux.microsoft.com 
> 
> That said, I suppose just ordering the nodes so that children come 
> before parents would also be fine. My thinking was just that accepting 
> any node ordering is simpler.
> 

Oh, actually I just tried that out and software_node_register_nodes 
doesn't allow this (parents must be added before children).

So I still think software_node_register_nodes and 
software_node_unregister_nodes are made more useful if children are 
removed recursively. I'll make some changes for v2 to make the change 
less far-reaching + better documentation.

Thanks,
Jordan

Re: [PATCH] software node: recursively unregister child swnodes

[Greg Kroah-Hartman]

On Thu, Jun 04, 2020 at 01:57:01PM -0700, Jordan Hand wrote:
> On 6/4/20 1:15 PM, Greg Kroah-Hartman wrote:
> > On Thu, Jun 04, 2020 at 12:36:23PM -0700, jorhand@linux.microsoft.com wrote:
> > > From: Jordan Hand <jorhand@linux.microsoft.com>
> > > 
> > > If a child swnode is unregistered after it's parent, it can lead to
> > > undefined behavior.
> > 
> > Crashing the system is not really "undefined" :)
> 
> Fair point :)
> 
> > 
> > > When a swnode is unregistered, recursively free it's children to avoid
> > > this condition.
> > 
> > Are you sure?  Why would you be unregistering a child after it's parent?
> > Why not just do not do that?
> > 
> 
> The main motivation for doing this was to support
> `software_node_unregister_nodes` so that the passed list of nodes does not
> need to be ordered in any particular way.
> 
> I suppose another way to do this would be to add a new function
> `fwnode_remove_software_node_recursive` and just call that from
> `software_node_unregister_nodes`.
> 
> That said, I suppose just ordering the nodes so that children come before
> parents would also be fine. My thinking was just that accepting any node
> ordering is simpler.

Right now, the way the driver model and sysfs/kobjects work is that all
objects must be removed in child-first order.  The problem of your
change where you want to try to remove the devices in parent-first order
is that you do not really know if you still have a reference to a child
device somewhere else, which would prevent this all from happening
correctly, right?

So if you "know" it is safe to drop a child, that's great, and expected.
Don't work to make  this one tiny user of the kobjects (which I'm still
not quite sure why they are kobjects and not devices), do things in a
different way from the rest of the kernel without a strong reason to do
so.

thanks,

greg k-h

Re: [PATCH] software node: recursively unregister child swnodes

[Jordan Hand]

On 6/5/20 12:54 AM, Greg Kroah-Hartman wrote:
> Right now, the way the driver model and sysfs/kobjects work is that all
> objects must be removed in child-first order.  The problem of your
> change where you want to try to remove the devices in parent-first order
> is that you do not really know if you still have a reference to a child
> device somewhere else, which would prevent this all from happening
> correctly, right?
> 
> So if you "know" it is safe to drop a child, that's great, and expected.
> Don't work to make  this one tiny user of the kobjects (which I'm still
> not quite sure why they are kobjects and not devices), do things in a
> different way from the rest of the kernel without a strong reason to do
> so.
> 
> thanks,
> 
> greg k-h
> 

I see, thanks for taking the time to explain, the reason for the 
existing behavior is more clear to me now. I agree it is better to have 
the caller remove the nodes in the correct order rather than having the 
swnode infrastructure try to have some special behavior.

Thanks,
Jordan