Re: [PATCH v4 08/12] PCI: Introduce /sys/bus/pci/devices/.../remove

[Alex Chiang <achiang@hp.com>]

* Andrew Morton <akpm@linux-foundation.org>:
> On Wed, 18 Mar 2009 16:40:06 -0600 Alex Chiang <achiang@hp.com> wrote:
> 
> > This patch adds an attribute named "remove" to a PCI device's sysfs
> > directory.  Writing a non-zero value to this attribute will remove the PCI
> > device and any children of it.
> > 
> > Trent Piepho wrote the original implementation and documentation.
> > 
> > Thanks to Vegard Nossum for testing under kmemcheck and finding locking
> > issues with the sysfs interface.
> > 
> > ...
> >
> > --- a/drivers/pci/pci-sysfs.c
> > +++ b/drivers/pci/pci-sysfs.c
> > @@ -246,6 +246,47 @@ struct bus_attribute pci_bus_attrs[] = {
> >  	__ATTR(rescan, S_IWUSR, NULL, bus_rescan_store),
> >  	__ATTR_NULL
> >  };
> > +
> > +static void remove_callback(struct device *dev)
> > +{
> > +	int bridge = 0;
> > +	struct pci_dev *pdev = to_pci_dev(dev);
> > +
> > +	mutex_lock(&pci_remove_rescan_mutex);
> > +
> > +	if (pdev->subordinate)
> > +		bridge = 1;
> > +
> > +	pci_remove_bus_device(pdev);
> > +	if (bridge && list_empty(&pdev->bus->devices))
> > +		pci_remove_bus(pdev->bus);
> > +
> > +	mutex_unlock(&pci_remove_rescan_mutex);
> > +}
> > +
> > +static ssize_t
> > +remove_store(struct device *dev, struct device_attribute *dummy,
> > +	     const char *buf, size_t count)
> > +{
> > +	int ret = 0;
> > +	unsigned long val;
> > +	struct pci_dev *pdev = to_pci_dev(dev);
> > +
> > +	if (strict_strtoul(buf, 0, &val) < 0)
> > +		return -EINVAL;
> > +
> > +	if (!capable(CAP_SYS_ADMIN))
> > +		return -EPERM;
> > +
> > +	if (pdev->subordinate && pci_is_root_bus(pdev->bus))
> > +		return -EBUSY;
> > +
> > +	if (val)
> > +		ret = device_schedule_callback(dev, remove_callback);
> > +	if (ret)
> > +		count = ret;
> > +	return count;
> > +}
> >  #endif
> 
> It is very hard for the reader (this one at least) to work out why
> device_schedule_callback() is used here, instead of simply doing the work
> directly.
> 
> The way to solve that problem is to add a code comment.

Hm, I thought it was well-known that a sysfs attribute cannot
remove itself without deadlocking. Thus, we need to use this
callback mechanism.

This thread has the most recent discussion:

	http://thread.gmane.org/gmane.linux.kernel/805033

I will add a comment to the code.

> Given that we're in a sysfs write() handler where no relevant locks at all
> are held, it seems rather weird that we cannot perform this operation
> synchronously, but no doubt the comment will explain all of this.
> 
> Do we need the CAP_SYS_ADMIN check if the sysfs file permissions are
> correct?  (I keep on asking this then forgetting the answer).

I will remove this (as per Greg's advice in a later mail).

> The device_schedule_callback() thing exposes us to (I assume) a pile of
> races, the most obvious of which is "what locking or refcounting keeps
> *dev alive?".  It would be nice to see an analysis/description of the
> lifetime issues here.  Perhaps in the changelog, preferably in code
> comments.

No races; device_schedule_callback() takes a ref on dev, pinning
it until the callback handler returns, after which it releases
the ref. You can see this in sysfs_schedule_callback.

Thanks for the review.

/ac