On Tue, Aug 18, 2020 at 04:34:18AM +0200, Jann Horn wrote: > I was looking at some usercopy stuff and noticed that > arch_within_stack_frames() (the helper used by the usercopy > bounds-checking logic to detect copies that cross stack frames) seems > to be a no-op on almost all modern kernel configurations. Yeah, this was unfortunate. > It is only defined for x86 - no implementation for e.g. arm64 exists > at all. The x86 version requires CONFIG_FRAME_POINTER, which is only > selected by CONFIG_UNWINDER_FRAME_POINTER (whereas the more modern > choice, and default, for x86-64 is CONFIG_UNWINDER_ORC). usercopy hardening landed in upstream _just_ before the ORC unwinder. > Personally, I don't feel very attached to that check; but if people > are interested in keeping it, it should probably be reworked to use > the proper x86 unwinder API: unwind_start(), unwind_next_frame(), > unwind_get_return_address_ptr() and unwind_done() together would > probably help with this. Otherwise, it should probably be removed, > since in that case it's pretty much going to just be bitrot? Right -- this was discussed at the time, and it seemed like it might be a relatively expensive change. The coverage provided by arch_within_stack_frames() just made sure that a given vector didn't cross frames (and didn't reside entirely between frames). I'm not too attached to the check, though I might be curious to see just _how_ expensive it would be to implement it with the unwinder API. -- Kees Cook
