On Sun, Oct 27, 2013 at 3:24 PM, Greg Kroah-Hartman
On Sun, Oct 27, 2013 at 09:13:29PM +0000, Linus Torvalds wrote:.. and one more case of freeing a delayed work object (likely a kobject again):
This time it looks like it's in the PCI layer, freeing the msi irq information.
It looks like that code simply does
and the problem is that the "entry->kobj" may have *other* references
to it, thanks to people accessing it through /sys, so despite doing a
kojbect_del/kobject_put(), it's not at all ok to then do a "kfree()"
on it. The embedded kobj might still be in use.
Afaik, that code should do the kfree() on the kobject in the _release_
method, not synchronously like that.
We already have a msi_kobj_release(), I'm wondering why that doesn't
do the kfree().
Bjorn? Yinghai? Greg, comments about that msi kobj usage?
Ick, it really should be doing a kfree() in the release only. Bjorn has
had a bunch of changes in this area recently, perhaps they are in
linux-next waiting for 3.13, and I've talked to him about getting rid of
all of the kobjects for msi files, as I don't think it's needed at all.
IIRC, you said you might take a look at converting this to attributes
on the train back home, so I haven't looked into it myself :)
Bjorn, don't you have a fix for this problem already done somewhere?
Yep, this is clearly wrong, and Veaceslav has a patch that moves the
kfree() to the release function. I've been waiting for a consolidated
repost of all his MSI-related fixes, but maybe he's been waiting for