On Fri, Jul 06, 2018 at 09:20:44AM -0700, Eric Dumazet wrote: > > > On 07/06/2018 08:54 AM, Mark Rutland wrote: > > On Fri, Jul 06, 2018 at 08:39:11AM -0700, Eric Dumazet wrote: > >> > >> > >> On 07/06/2018 07:57 AM, Mark Rutland wrote: > >>> On Fri, Jul 06, 2018 at 07:47:04AM -0700, Eric Dumazet wrote: > >>>> > >>>> > >>>> On 07/06/2018 07:28 AM, Mark Rutland wrote: > >>>>> In tnode_free() we iterate over a callback_head list with a while loop. > >>>>> At the start of the loop body we generate the next head pointer, and at > >>>>> the end of the loop body we generate the tn pointer for the next > >>>>> iteration of the loop by using container_of() on the head pointer to > >>>>> find the tnode, and deriving the kv pointer from this. > >>>>> > >>>>> In the final iteration of the loop, this means that we derive a pointer > >>>>> from NULL, which is undefined behaviour, which UBSAN detects: > >>>> > >>>> There is no dereference, your patch title is misleading. > >>>> > >>>> UBSAN might be fooled, not the C compiler. > >>> > >>> I'm happy to change the title to "avoid undefined behaviour". > >>> > >> > >> Are you planning to change this as well ? > >> > >> include/linux/stddef.h:19:#define offsetof(TYPE, MEMBER) > >> ((size_t)&((TYPE *)0)->MEMBER) > > > > No, because __builtin_offsetof() is used these days (since GCC 4), > > avoiding the undefined behaviour. > > Ah... should we remove the line and declare linux must be compiled > with GCC 4 at least ?
Good question -- IIUC x86 already mandates at least GCC 4.6, as this is necessary for jump labels. There was a push to mandate a more recent compiler generally, but I'm not sure of the current state of things. Arnd, were we planning to mandate at least GCC 4.x? > > I do concede that if this is everywhere it's not worth the effort, and > > from the looks of things, the gnaliest cases are where we do things > > like: > > > > get_user(var, &struct->field) > > > > ... where the user could validly pass a NULL pointer if it wished. > > > > So I guess I'll give up. > > There is value to your patch, since it makes UBSAN happy. > > But please change the title and changelog accordingly. Sure thing. Are you happy with the "deriving a pointer from NULL" wording in the explanation? That wasn't intended to mean a dereference. I'll change the title to: ipv4: fib: avoid undefined behaviour Thanks, Mark.
