On Tue, Apr 19, 2016 at 11:06:45AM -0600, Jason Gunthorpe wrote: > On Tue, Apr 19, 2016 at 06:36:22AM -0400, Stefan Berger wrote: > > On 04/19/2016 06:12 AM, Jarkko Sakkinen wrote: > > >On Mon, Apr 18, 2016 at 01:26:13PM -0400, Stefan Berger wrote: > > >>From: Jason Gunthorpe <jguntho...@obsidianresearch.com> > > >> > > >>The final thing preventing this was the way the sysfs files were > > >>attached to the pdev. Follow the approach developed for ppi and move > > >>the sysfs files to the chip->dev with symlinks from the pdev > > >>for compatibility. Everything in the core now sanely uses container_of > > >>to get the chip. > > >Can you give me a quick recap why this patch was mandatory to make the > > >patch set work? Which regression in the earlier versions of the patch > > >set this fixes? > > > > The below patch removes usage of dev_get_drvdata() for retrieving the chip. > > With Christophe's series dropping the priv field I now can use the drvdata > > to store proxy_dev rather than re-introducing the priv field in the chip > > structure. Besides that it doesn't seem necessary to use the drvdata field > > to get from the chip to the device if a simple container_of can do it. > > More specifically, since the vtpm patches use a NULL parent, the > approach of putting the sysfs files on the parent is no longer > workable. > > The early vtpm patches simply moved the sysfs files to the tpm_chip > when a parent is NULL, which is inconsistent for userspace. This also > created a problem where drvdata on the chip now had to point back to > the chip, meaning it became unusable for its new intended purpose. > > The fix is to make everything uniform and put the sysfs files in the > correct place for all drivers (under the chip) and use symlinks for > compat.
OK, thanks for explaining this, make perfect sense now. I'll do a more detailed review (and testing) later on. > Jason /Jarkko
