Re: [RFC PATCH] x86: entry: flush the cache if syscall error

[Kristen C Accardi <kristen@linux.intel.com>]

On Mon, 2018-09-10 at 22:32 +0200, Jann Horn wrote:
> On Mon, Sep 10, 2018 at 9:14 PM Kristen Carlson Accardi
> <kristen@linux.intel.com> wrote:
> > This patch aims to make it harder to perform cache timing attacks
> > on data
> > left behind by system calls. If we have an error returned from a
> > syscall,
> > flush the L1 cache.
> 
> What kind of performance impact does this have on a process that e.g.
> attempts to access a large number of paths to which it has no access,
> and what is the impact on the hyperthread?

This will undoubtedly have a performance impact - I can certainly run
some benchmarks to find out just how much in this scenario.

> 
> (You may want to also CC the X86 maintainers (especially Andy
> Lutomirski) and LKML on this series relatively early - people on the
> kernel-hardening list tend to have a different focus compared to the
> relevant maintainers, so there's some feedback that you'll probably
> only get once you submit this to places other than the
> kernel-hardening@ list.)

Thank you yes, I admit I decided to try to float this idea past a
friendlier audience first :).

> 
> > Signed-off-by: Kristen Carlson Accardi <kristen@linux.intel.com>
> > ---
> >  arch/x86/Kconfig        |  8 ++++++++
> >  arch/x86/entry/common.c | 20 ++++++++++++++++++++
> >  2 files changed, 28 insertions(+)
> > 
> > diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> > index c5ff296bc5d1..8a67642ff9fe 100644
> > --- a/arch/x86/Kconfig
> > +++ b/arch/x86/Kconfig
> > @@ -445,6 +445,14 @@ config RETPOLINE
> >           code are eliminated. Since this includes the syscall
> > entry path,
> >           it is not entirely pointless.
> > 
> > +config SYSCALL_FLUSH
> > +       bool "Clear L1 Cache on syscall errors"
> > +       default y
> > +       help
> > +         Select to allow the L1 cache to be cleared upon return of
> > +         an error code from a syscall. This will reduce the
> > likelyhood of
> > +         speculative execution style attacks on syscalls.
> 
> s/L1/L1D/ ?

I can change this - the documentation for this msr mentioned that on
some processors the icache might be impacted as well - think that's
worth mentioning?

> 
> >  config INTEL_RDT
> >         bool "Intel Resource Director Technology support"
> >         default n
> > diff --git a/arch/x86/entry/common.c b/arch/x86/entry/common.c
> > index 3b2490b81918..77beff541013 100644
> > --- a/arch/x86/entry/common.c
> > +++ b/arch/x86/entry/common.c
> > @@ -268,6 +268,22 @@ __visible inline void
> > syscall_return_slowpath(struct pt_regs *regs)
> >         prepare_exit_to_usermode(regs);
> >  }
> > 
> > +__visible inline void l1_cache_flush(struct pt_regs *regs)
> > +{
> > +       if (IS_ENABLED(CONFIG_SYSCALL_FLUSH)) {
> > +               if (regs->ax == 0 || regs->ax == -EAGAIN ||
> > +                   regs->ax == -EEXIST || regs->ax == -ENOENT ||
> > +                   regs->ax == -EXDEV || regs->ax == -ETIMEDOUT ||
> > +                   regs->ax == -ENOTCONN || regs->ax ==
> > -EINPROGRESS)
> > +                       return;
> > +
> > +               if (static_cpu_has(X86_FEATURE_FLUSH_L1D)) {
> 
> Style nit: Maybe merge this condition into the first if() ?
> 
> > +                       wrmsrl(MSR_IA32_FLUSH_CMD, L1D_FLUSH);
> > +                       return;
> > +               }
> > +       }
> > +}
> > +
> >  #ifdef CONFIG_X86_64
> >  __visible void do_syscall_64(unsigned long nr, struct pt_regs
> > *regs)
> >  {
> > @@ -290,6 +306,8 @@ __visible void do_syscall_64(unsigned long nr,
> > struct pt_regs *regs)
> >                 regs->ax = sys_call_table[nr](regs);
> >         }
> > 
> > +       l1_cache_flush(regs);
> > +
> >         syscall_return_slowpath(regs);
> >  }
> >  #endif
> > @@ -338,6 +356,8 @@ static __always_inline void
> > do_syscall_32_irqs_on(struct pt_regs *regs)
> >  #endif /* CONFIG_IA32_EMULATION */
> >         }
> > 
> > +       l1_cache_flush(regs);
> > +
> >         syscall_return_slowpath(regs);
> >  }
> > 
> > --
> > 2.14.4
> >