Re: [PATCH 157/190] Revert "Input: ad7879 - add check for read errors in interrupt"

[Greg Kroah-Hartman]

On Tue, Apr 27, 2021 at 12:22:12PM -0700, Dmitry Torokhov wrote:
> On Tue, Apr 27, 2021 at 06:55:10PM +0200, Greg Kroah-Hartman wrote:
> > On Wed, Apr 21, 2021 at 10:03:33AM -0700, Dmitry Torokhov wrote:
> > > Hi Greg,
> > > 
> > > On Wed, Apr 21, 2021 at 03:00:32PM +0200, Greg Kroah-Hartman wrote:
> > > > This reverts commit e85bb0beb6498c0dffe18a2f1f16d575bc175c32.
> > > > 
> > > > Commits from @umn.edu addresses have been found to be submitted in "bad
> > > > faith" to try to test the kernel community's ability to review "known
> > > > malicious" changes.  The result of these submissions can be found in a
> > > > paper published at the 42nd IEEE Symposium on Security and Privacy
> > > > entitled, "Open Source Insecurity: Stealthily Introducing
> > > > Vulnerabilities via Hypocrite Commits" written by Qiushi Wu (University
> > > > of Minnesota) and Kangjie Lu (University of Minnesota).
> > > > 
> > > > Because of this, all submissions from this group must be reverted from
> > > > the kernel tree and will need to be re-reviewed again to determine if
> > > > they actually are a valid fix.  Until that work is complete, remove this
> > > > change to ensure that no problems are being introduced into the
> > > > codebase.
> > > 
> > > This one looks really OK to me and does not have to be reverted (unless
> > > Aditya will come clean and show the error introduced?).
> > 
> > I'll drop this revert, but it isn't usually good to be calling printk()
> > from an irq.
> 
> How else do you suggest we tell that something is wrong when
> communicating with the device? For these types of devices the
> communication is essentially unsolicited so we can't pass failure to a
> caller to deal with it (i.e. unlike USB there is no URB posted that we
> could fail and use as a mechanism to signal error to some other layer)
> and while we could invent "something went wrong" input event so far
> there was no interest in having anything like that.
> 
> I'd suggest sending KOBJ_ERROR uevent when a device driver detects
> anomaly in the device it controls, but I wonder how systemd would react
> given past experiences with KOBJ_BIND/KOBJ_UNBIND.
> 
> The message is ratelimited so it will not overfill the logs...

Sending uevents from an irq is not a good idea, as you say :)

I don't know what the best way to "fail" this is, a ratelimited printk()
seems to be about all you can do.  Luckily hardware doesn't fail that
often in this manner.

thanks,

greg k-h