On Sat, Feb 13, 2016 at 11:50:08PM -0700, Jason Gunthorpe wrote: > On Sun, Feb 14, 2016 at 06:55:12AM +0200, Jarkko Sakkinen wrote: > > On Fri, Feb 12, 2016 at 05:04:29PM -0700, Jason Gunthorpe wrote: > > > This was missed during the struct device conversion, we > > > need to hold a kref on the chip to make sure it isn't freed. > > > > > > Signed-off-by: Jason Gunthorpe <jguntho...@obsidianresearch.com> > > > > I'm bit confused about this patch. What is the regression if this > > needs > > The patch is simply totally broken, the placement of the get_device is > wrong: > > > > @@ -53,6 +53,8 @@ struct tpm_chip *tpm_chip_find_get(int chip_num) > > > chip = pos; > > > break; > > > } > > > + > > > + get_device(&chip->dev); > > It needs to be moved up two lines before the break, into the if > statement.
Right. > As for the urgency - today the tpm core relies on module locking to > try and prevent tpm_chip_unregister from racing with stuff like the > above. That is totally broken in modern kernels, but it is what the > core tries to do. Within that framework the get/put are not needed > because of the module locking. Right, because that gives the guarantee that device has refcount of at least one. > The only time these additional get/put do anything is when we are > racing with tpm_unregister, but if we are racing with unregister then > there are much bigger problems and things will crash anyhow. > > So, this patch is just a tiny step. > > The revised version of this patch with the rw_sem attempts to address > the complete race. Got it. Yeah, I'll drop this from my next pull request. Thanks for the explanation. > Jason /Jarkko