RE: [PATCH] x86/entry/64: randomize kernel stack offset upon syscall

From: Reshetova, Elena
Date: Fri Apr 12 2019 - 01:36:26 EST


> On Wed, Apr 10, 2019 at 3:24 AM Reshetova, Elena
> <elena.reshetova@xxxxxxxxx> wrote:
> >
> >
> > > > > On Mon, Apr 08, 2019 at 09:13:58AM +0300, Elena Reshetova wrote:
> > > > > > diff --git a/arch/x86/entry/common.c b/arch/x86/entry/common.c
> > > > > > index 7bc105f47d21..38ddc213a5e9 100644
> > > > > > --- a/arch/x86/entry/common.c
> > > > > > +++ b/arch/x86/entry/common.c
> > > > > > @@ -35,6 +35,12 @@
> > > > > > #define CREATE_TRACE_POINTS
> > > > > > #include <trace/events/syscalls.h>
> > > > > >
> > > > > > +#ifdef CONFIG_RANDOMIZE_KSTACK_OFFSET
> > > > > > +#include <linux/random.h>
> > > > > > +
> > > > > > +void *alloca(size_t size);
> > > > > > +#endif
> > > > > > +
> > > > > > #ifdef CONFIG_CONTEXT_TRACKING
> > > > > > /* Called on entry from user mode with IRQs off. */
> > > > > > __visible inline void enter_from_user_mode(void)
> > > > > > @@ -273,6 +279,13 @@ __visible void do_syscall_64(unsigned long nr,
> struct
> > > > pt_regs *regs)
> > > > > > {
> > > > > > struct thread_info *ti;
> > > > > >
> > > > > > +#ifdef CONFIG_RANDOMIZE_KSTACK_OFFSET
> > > > > > + size_t offset = ((size_t)prandom_u32()) % 256;
> > > > > > + char *ptr = alloca(offset);
> > > > > > +
> > > > > > + asm volatile("":"=m"(*ptr));
> > > > > > +#endif
> > > > > > +
> > > > > > enter_from_user_mode();
> > > > > > local_irq_enable();
> > > > > > ti = current_thread_info();
> > > > >
> > > > > Would it make sense to also do this for the compat syscalls
> > > > > (do_fast_syscall_32, do_int80_syscall_32)?
> > > >
> > > > Could someone please include the full patch, with justification and
> > > > performance impact analysis etc.? Can only find the code part of the
> > > > thread on lkml, which leaves out this context.
> > > >
> > >
> > > Sorry, this is very weird, I cannot find it either from lkml, but it was sent there
> > > to begin with (and as visible from reply-to headers).
> > >
> > > Do you want me to resent original version or with "do_fast_syscall_32,
> > > do_int80_syscall_32" additions (I am finishing testing them now).
> >
> > I will resend the original x86_64 now since this is the one I tested and
> > measured properly. The 32 bit changes seem to work fine inside my 32 bit VM,
> > but since I don't have any real 32 bit HW, I am hesitant to send them out without
> > real HW testing and measuring.
> >
> > This is the asm code for 32 bits (note it requires __builtin_alloca definition and not
> just alloca,
> > so I will change the 64 bit version to use it also):
> >
> > #ifdef CONFIG_RANDOMIZE_KSTACK_OFFSET
> > size_t offset = ((size_t)prandom_u32()) % 256;
> > 0xc10025b6 call 0xc146f7d0 <prandom_u32>
> > 0xc10025bb movzbl %al,%eax
> > char *ptr = __builtin_alloca(offset);
> > 0xc10025be add $0x12,%eax
> > 0xc10025c1 and $0x1fc,%eax
> > 0xc10025c6 sub %eax,%esp
> > 0xc10025c8 lea 0x27(%esp),%eax
> > 0xc10025cc and $0xfffffff0,%eax
> >
> > Also, the result is 47 different random offsets produced,
> > which is slightly better than 33 offsets for x86_64.
> >
>
> I would suggest that you macro-ify this thing:
>
> #ifdef WHATEVER
> #define add_random_stack_offset() do { void *addr = ... } while (0)
> #else
> #define add_random_stack_offset() do {} while (0)
> #endif
>
> since you'll end up with more than one call site.

Sure, will do. So, you are ok for this to be also called from do_fast_syscall_32
and do_int80_syscall_32? I can send the resulting patch, just cannot test on any
real 32 bit HW, only VM.

Best Regards,
Elena.