Skip to contentSkip to conversation

Missing piece from as659

DISCUSSION Matthew Wilcox 3 participants 3 messages Feb 24, 2006

lore
Share & tools
lore
#0 lore raw

Alan, you didn't cc the pci mailing list on the original patch.
http://www.ussg.iu.edu/hypermail/linux/kernel/0602.2/2673.html

You only fix pci_get_subsys; pci_get_class has the same bug.

If it is a bug, of course. It's not clear to me whether it's permissible to call pci_dev_put under a spinlock or not. That boils down to whether kobject ->release methods can sleep or not. That isn't documented in Documentation/kobject.txt and I rather think it should be.

#1 lore raw
Quoted 2 lines · Matthew Wilcox

Alan, you didn't cc the pci mailing list on the original patch.
http://www.ussg.iu.edu/hypermail/linux/kernel/0602.2/2673.html

This seems to be a blind spot of mine. Yours is the second complaint in two days about patches I failed to CC to the appropriate maintainer/mailing-list...

Quoted 1 line

You only fix pci_get_subsys; pci_get_class has the same bug.

Please submit a similar bugfix for pci_get_class, then. I just noticed the log messages from pci_get_subsys because that's the routine that happened to run on the machine I was testing.

Quoted 4 lines

If it is a bug, of course. It's not clear to me whether it's permissible to call pci_dev_put under a spinlock or not. That boils down to whether kobject ->release methods can sleep or not. That isn't documented in Documentation/kobject.txt and I rather think it should be.

It is a bug, but it has been ignored up until recently. Within the last month or two, Greg added a might_sleep() call to put_device(). It wouldn't hurt to do the same thing to kobject_put(), or maybe just kobject_release().

Alan Stern

#2 lore raw
Quoted 11 lines · Alan Stern
Quoted 5 lines · Matthew Wilcox

http://www.ussg.iu.edu/hypermail/linux/kernel/0602.2/2673.html
If it is a bug, of course. It's not clear to me whether it's permissible to call pci_dev_put under a spinlock or not. That boils down to whether kobject ->release methods can sleep or not. That isn't documented in Documentation/kobject.txt and I rather think it should be.

It is a bug, but it has been ignored up until recently. Within the last month or two, Greg added a might_sleep() call to put_device(). It wouldn't hurt to do the same thing to kobject_put(), or maybe just kobject_release().

kobject_put() will not necessarily sleep, it's only the device_put that will, due to the locking in the driver core.

thanks,

greg k-h