Skip to contentSkip to conversation

2.6 probe.c "pcibus_class" Device Class, release function

DISCUSSION John Rose 3 participants 6 messages Feb 3, 2004 – Feb 4, 2004 PCI

lore
Share & tools
lore

John Rose

author
#0 lore raw

The function release_pcibus_dev() in probe.c defines the release procedure for device class pcibus_class. I want to suggest that this function be scrapped :)

This release function is called in the code path of class_device_unregister(). The pcibus_class devices aren't currently unregistered anywhere, from what I can tell, so this release function is currently unused. The runtime removal of PCI buses from logical partitions on PPC64 requires the unregistration of these class devices. The natural place to do this IMHO is in pci_remove_bus_device() in remove.c.

The problem is that this calls pci_destroy_dev(), which calls put() on the same "bridge" device that the release function does. This should only be done once in the course of removing a pci_bus, and I doubt that we want to change
pci_destroy_dev(). The kfree() of the pci_bus struct is also done in both
pci_remove_bus_device() and release_pcibus_dev().

So the only two operations in the release function are redundantly performed in the place where it makes sense to unregister. For these reasons, I think we should scrap the release function altogether and set that pointer in the struct class to NULL.

Thoughts?
John

#1 lore raw
Quoted 9 lines · John Rose

The function release_pcibus_dev() in probe.c defines the release procedure for device class pcibus_class. I want to suggest that this function be scrapped :)

This release function is called in the code path of class_device_unregister(). The pcibus_class devices aren't currently unregistered anywhere, from what I can tell, so this release function is currently unused. The runtime removal of PCI buses from logical partitions on PPC64 requires the unregistration of these class devices. The natural place to do this IMHO is in pci_remove_bus_device() in remove.c.

You're right that the class device isn't currently unregistered, and that was an oversight in the patch I originally sent. Attatched is a patch that remedies that situation. pci_remove_bus_device() *is* the natural place to unregister the class_dev, and that's just what the patch does.

Quoted 5 lines

The problem is that this calls pci_destroy_dev(), which calls put() on the same "bridge" device that the release function does. This should only be done once in the course of removing a pci_bus, and I doubt that we want to change
pci_destroy_dev(). The kfree() of the pci_bus struct is also done in both
pci_remove_bus_device() and release_pcibus_dev().

Yep. The patch pulls the kfree() out of pci_remove_bus_device() and calls class_device_unregister() in it's place. This will put() the bridge device and free the pci_bus as needed. put() does need to be called twice because the bridge device is get()'d twice: once when the device is registered and once when it's bus device grabs a reference to it.

Quoted 7 lines

So the only two operations in the release function are redundantly performed in the place where it makes sense to unregister. For these reasons, I think we should scrap the release function altogether and set that pointer in the struct class to NULL.

Thoughts?
John

Disagree, for the reasons above. ;) Patch attatched.

Cheers!

-Matt

drivers/pci/remove.c +1 -1
diff -Nurp --exclude-from=/home/mcd/.dontdiff linux-2.6.2-rc3/drivers/pci/remove.c linux-2.6.2-rc3+pcibus_memleak-fix/drivers/pci/remove.c
@@ -83,7 +83,7 @@ void pci_remove_bus_device(struct pci_de
 		list_del(&b->node);
 		spin_unlock(&pci_bus_lock);
 
-		kfree(b);
+		class_device_unregister(&b->class_dev);
 		dev->subordinate = NULL;
 	}
 
author
#2 lore raw
Quoted 3 lines

put() does need to be
called twice because the bridge device is get()'d twice: once when the device is registered and once when it's bus device grabs a reference to it.

Looks great, thanks Matt.

John

author
#3 lore raw

Oops - quick question. Does unregistering blow away the cpuaffinity attr and the bridge symlink?

Thanks-
John

Quoted 7 lines · John Rose
Quoted 3 lines

put() does need to be
called twice because the bridge device is get()'d twice: once when the device is registered and once when it's bus device grabs a reference to it.

Looks great, thanks Matt.

John

author
#4 lore raw

After testing, it would appear that it does. Perfect! Thanks.

John

Quoted 14 lines · John Rose

Oops - quick question. Does unregistering blow away the cpuaffinity attr and the bridge symlink?

Thanks-
John

Quoted 7 lines · John Rose
Quoted 3 lines

put() does need to be
called twice because the bridge device is get()'d twice: once when the device is registered and once when it's bus device grabs a reference to it.

Looks great, thanks Matt.

John

#5 lore raw
In reply to Matthew Dobson
Quoted 18 lines · Matthew Dobson
Quoted 5 lines · John Rose

The function release_pcibus_dev() in probe.c defines the release procedure for
device class pcibus_class. I want to suggest that this function be scrapped :)

Quoted 11 lines

This release function is called in the code path of class_device_unregister().
The pcibus_class devices aren't currently unregistered anywhere, from what I
can tell, so this release function is currently unused. The runtime removal of
PCI buses from logical partitions on PPC64 requires the unregistration of these
class devices. The natural place to do this IMHO is in pci_remove_bus_device()
in remove.c.

Quoted 5 lines

You're right that the class device isn't currently unregistered, and that was an oversight in the patch I originally sent. Attatched is a patch that remedies that situation. pci_remove_bus_device() *is* the natural place to unregister the class_dev, and that's just what the patch does.

Thanks, I've applied this patch to my trees and will send off to Linus in the next round of PCI patches.

greg k-h