On 01/31/2017 02:17 PM, Greg Kroah-Hartman wrote: > On Tue, Jan 31, 2017 at 01:58:14PM +0100, Christoph Hellwig wrote: >> On Tue, Jan 31, 2017 at 11:21:02AM +0100, Greg Kroah-Hartman wrote: >>> >>> -next isn't Linus's tree, sometimes stuff sits in there for years :) >>> >>> Anyway, if this is a configfs issue, Christoph and Joel can take a look >>> at it. Any reason you didn't cc: Joel as well (the MAINTAINERS file is >>> your friend...) >> >> It's really a mismatched assumption. The configfs binary file >> code just chunks updates up into a buffer, which only gets flushed >> at ->release time. If we'd move that to ->flush the issue Marek reports >> would be fixed. >> >> But I don't think we want that - triggering a filp_open from the update >> of a _binary_ attribute for a start is wrong. And second doing this >> using ->fs of a random calling process is bound to cause problems. >> >> I think he is using the wrong kind of interface for the job. > > Ah, that's why no one has seen this before :) > > So, the DT overlay code needs to be fixed...
Well I ran into the issue when I loaded DTO using 'cat' which bound a driver which required firmware and the firmware loader uses the filp_open() to load the firmware file from the FS. This crashed my kernel because the current->fs was NULL. The example I provided is a stripped down version which checks the current->fs directly to make things simpler. I think the issue is in the firmware loader (or filp_open() itself?) . Shouldn't the filp_open() somehow assure that the current->fs is valid if it is used within instead of triggering a NULL ptr dereference when called from ie. the configfs binary attribute write callback ? -- Best regards, Marek Vasut