On Fri, Nov 2, 2018 at 3:28 PM Dave Hansen <dave.han...@intel.com> wrote: > > On 11/2/18 12:50 PM, Waiman Long wrote: > > On 11/02/2018 03:44 PM, Dave Hansen wrote: > >> On 11/2/18 12:40 PM, Waiman Long wrote: > >>> The 64k+ limit check is kind of arbitrary. So the check is now removed > >>> to just let expand_stack() decide if a segmentation fault should happen. > >> With the 64k check removed, what's the next limit that we bump into? Is > >> it just the stack_guard_gap space above the next-lowest VMA? > > I think it is both the stack_guard_gap space above the next lowest VMA > > and the rlimit(RLIMIT_STACK). > > The gap seems to be hundreds of megabytes, typically where RLIMIT_STACK > is 8MB by default, so RLIMIT_STACK is likely to be the practical limit > that will be hit. So, practically, we've taken a ~64k area that we > would on-demand extend the stack into in one go, and turned that into a > the full ~8MB area that you could have expanded into anyway, but all at > once. > > That doesn't seem too insane, especially since we don't physically back > the 8MB or anything. Logically, it also seems like you *should* be able > to touch any bit of the stack within the rlimit. > > But, on the other hand, as our comments say: "Accessing the stack below > %sp is always a bug." Have we been unsuccessful in convincing our gcc > buddies of this?
FWIW, the old code is a bit bogus. Why are we restricting the range of stack expending addresses for user code without restricting the range of kernel uaccess addresses that would do the same thing? So I think I agree with the patch.