Re: [PATCH v2 1/2] PCI: Convert to device_create_managed_software_node()

[Heikki Krogerus]

On Thu, Sep 30, 2021 at 10:04:02AM -0500, Bjorn Helgaas wrote:
> On Thu, Sep 30, 2021 at 03:12:45PM +0300, Heikki Krogerus wrote:
> > In quirk_huawei_pcie_sva(), use device_create_managed_software_node()
> > instead of device_add_properties() to set the "dma-can-stall"
> > property.
> > 
> > Reviewed-by: Andy Shevchenko <andy.shevchenko@xxxxxxxxx>
> > Acked-by: Zhangfei Gao <zhangfei.gao@xxxxxxxxxx>
> > Signed-off-by: Heikki Krogerus <heikki.krogerus@xxxxxxxxxxxxxxx>
> > ---
> > Hi,
> > 
> > The commit message now says what Bjorn requested, except I left out
> > the claim that the patch fixes a lifetime issue.
> 
> Thanks.
> 
> The commit log should help reviewers determine whether the change is
> safe and necessary.  So far it doesn't have any hints along that line.
> 
> Comparing device_add_properties() [1] and
> device_create_managed_software_node() [2], the only difference in this
> case is that the latter sets "swnode->managed = true".  The function
> comment says "managed" means the lifetime of the swnode is tied to the
> lifetime of dev, hence my question about a lifetime issue.
> 
> I can see that one reason for this change is to remove the last caller
> of device_add_properties(), so device_add_properties() itself can be
> removed.  That's a good reason for wanting to do it, and the commit
> log could mention it.

Fair enough. I need to explain the why as well as the what.

I'll improve the commit message, but just to be clear, the goal is
actually not to get rid of device_add_properties(). It is removed in
the second patch together with device_remove_properties() because
there are simply no more users for that API.

> But it doesn't help me figure out whether it's safe.  For that,
> I need to know the effect of setting "managed = true".  Obviously
> it means *something*, but I don't know what.  It looks like the only
> test is in software_node_notify():
> 
>   device_del
>     device_platform_notify_remove
>       software_node_notify_remove
>         sysfs_remove_link(dev_name)
>         sysfs_remove_link("software_node")
>         if (swnode->managed)                 <--
>           set_secondary_fwnode(dev, NULL)
>           kobject_put(&swnode->kobj)
>     device_remove_properties
>       if (is_software_node())
>         fwnode_remove_software_node
>           kobject_put(&swnode->kobj)
>         set_secondary_fwnode(dev, NULL)
> 
> I'm not sure what's going on here; it looks like some redundancy with
> multiple calls of kobject_put() and set_secondary_fwnode().  Maybe you
> are in the process of removing device_remove_properties() as well as
> device_add_properties()?

It'll get removed, but that's not the goal. The goal is to get rid of
the call to it in device_del(), so not the function itself. That call
is the problem here as explained in commit 151f6ff78cdf ("software
node: Provide replacement for device_add_properties()").

I'll split the second patch, and first only remove that
device_remove_properties() call from device_del(), and only after
that remove the functions themselves.

thanks,

-- 
heikki