> > > Solution: Prevent chip detach while sysctl table is busy, means > > reference counting both the bus and the chip modules. This needs an > > API change on the user side: Move ... sensors/somechip-0-2d to > > sensors/bus-0/somechip-2d. That way bus module is referenced on the > > path. > > Do you seriously think we should do that? This sounds like a very big > change for a near-end-of-life kernel tree. OK, I think I understand this > would be the only way to have references handled correctly. But for now > we don't handle them at all. What about handling chip drivers counts > correctly and forget about bus drivers counts? That's the way things > were done with the old method (up to i2c 2.7.0) so I would feel less > uncomfortable sending such a patch to Marcelo, since there would be no > regression anymore, and still the race condition fixed. Let's start by fixing the chip count. > Ky?sti, what would it take to restore correct module reference counting > on chip drivers? Does it require to update only i2c-core and i2c-proc, > or would we have to update all drivers? Please help me with this. This > is required in order to have any serious patch sent to Marcelo, and > you're the one of us all understanding all these things at best, it > seems. If we only get .owner in adapter and driver structure, I can keep the changes within i2c-core and i2c-proc. Did you have a collection of video4linux drivers, with i2c dependency, easily available ? Seems i2c drivers in kernel tree 2.4.23 drivers/media/video handle reference counting internally; They get reference as soon as the client is attached to adapter. One has to rmmod bus before chip. > > Does not seem _correct_ either. A call to i2c_fill_inode increments > > only i2c-proc module reference. It was already counted as soon as you > > entered /proc/sys/dev/sensors/. > > Is it something that should be fixed, or that already was, or neither? Kernel 2.2 stuff. I was mistaken, there is i2c_fill_inode and i2c_dir_fill_inode and the latter does handle chip reference, but not the adapter. Ironically, for kernels 2.2 it would be easy to fix. I will prepare a set of incremental patches from the CVS with minimal changes, with compatible API. Binary compatibility not maintained. -- Ky?sti M?lkki <kyosti.malkki at welho.com>
