On Wed, Dec 20, 2017 at 01:43:13PM +0100, Martin Pieuchot wrote: > > rtm_ifchg() calls route_input() which needs the net lock. > > Does it? What needs the lock in route_input()? I thought routing > sockets are still protected by the KERNEL_LOCK().
My fault. I have misread the code. Your change is fine. > > May we access the ifp->if_flags without net lock? As we want to > > compare them to their value before we changed them, would it be > > even better to put the "oif_flags = ifp->if_flags;" under the lock? > > Flags are modified by the driver, not by the stack. Some of them are > event immutable and might be better moved to a different field. You have marked if_flags and if_xflags with an [N]. Reading happens everywhere in the stack. Carp changes them often. I think we have the kernel lock there. Note that if_flags is short so access is not atomic. Perhaps we should split the flags like I did with the sb_flagsintr a long time ago. Changing them to int would be an easy first step. > The goal is to stop grabbing the lock for per-driver ioctl(2). This > would fix the problem that sleeping in if_ioctl() might block network > traffic for a long time. That means the limit between driver and stack > fields have to be clear. You just pointed out that `if_flags' has some > issues. Thanks for the explanation. The goal sounds reasonable. But that means we have to check all the drivers and especially the pseudo-drivers' ioctl functions before unlocking (*ifp->if_ioctl)(). bluhm
