On Thu, Jan 11, 2018 at 5:11 PM, Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > This patch doesn't affect arch/x86/lib/getuser.S, which I find surprising. > > Of all the user access functions, I actually think that get_user() is > the one most likely to have the result then used speculatively as an > index (the required second dependent read to actually leak data). > > I do *not* see people doing "copy_from_user()" and then somehow using > the thing as an index to another array. I mean, it can happen (copy a > structure, use a member in that structure), but it doesn't seem to be > the most likely thing. > > The most likely thing would seem to be some random ioctl() do a > "get_user()" to get an index, and then using that index. That would > seem to be one of the easier ways to perhaps get that kind of kernel > spectre attack. > > Adding the ASM_IFENCE to __get_user_X() in arch/x86/lib/getuser.S > would seem to go naturally together with the copy_user_64.S changes in > this patch. > > Is there some reason __get_user_X() was overlooked? Those are _the_ > most common user accessor functions that do the address limit > checking. Oversight, I was focused on the uaccess_begin conversions. Yes, let me go add ASM_IFENCE after the ASM_STAC in those paths.
