On Mon, Oct 07, 2019 at 09:24:17PM -0700, Linus Torvalds wrote: > On Mon, Oct 7, 2019 at 9:09 PM Linus Torvalds > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > > > Try the attached patch, and then count the number of "rorx" > > instructions in the kernel. Hint: not many. On my personal config, > > this triggers 15 times in the whole kernel build (not counting > > modules). > > .. and four of them are in perf_callchain_user(), and are due to those > "__copy_from_user_nmi()" with either 4-byte or 8-byte copies. > > It might as well just use __get_user() instead. > > The point being that the silly code in the header files is just > pointless. We shouldn't do it. FWIW, the one that looks the most potentiall sensitive in that bunch is arch/x86/kvm/paging_tmpl.h:388: if (unlikely(__copy_from_user(&pte, ptep_user, sizeof(pte)))) in the bowels of KVM page fault handling. I would be very surprised if the rest would be detectable... Anyway, another question you way: what do you think of try/catch approaches to __get_user() blocks, like e.g. restore_sigcontext() is doing? Should that be available outside of arch/*? For that matter, would it be a good idea to convert get_user_ex() users in arch/x86 to unsafe_get_user()?
