Hi Marcin,
On Fri, Jan 13, 2017 at 08:44:46AM +0100, Marcin Nowakowski wrote:
> With certain EVA configurations it is possible for the kernel address
> space to overlap user address space, which allows the user to set
> watchpoints on kernel addresses via ptrace.
>
> If a watchpoint is set in the watch exception handling code (after
> exception level has been cleared) then the system will hang in an
> infinite loop when hitting a watchpoint while trying to process it.
>
> To prevent that place all watch exception entry/exit code in a single
> named section and disallow placing watchpoints in that area.
An interesting approach. If I understand correctly, any watch placed on
watch exception entry code up to when watchpoints are disabled in
do_watch and interrupts re-enabled is disallowed, as well as the code
after watchpoints are restored.
Should mips_install_watch_registers() also move into that section?
Should do_watch be made notrace too, so we don't get calls into
unprotected ftrace code?
>
> Signed-off-by: Marcin Nowakowski <marcin.nowakowski@xxxxxxxxxx>
> ---
> arch/mips/kernel/entry.S | 2 +-
> arch/mips/kernel/genex.S | 2 ++
> arch/mips/kernel/ptrace.c | 14 ++++++++++++++
> arch/mips/kernel/traps.c | 2 ++
> arch/mips/kernel/vmlinux.lds.S | 8 ++++++++
> 5 files changed, 27 insertions(+), 1 deletion(-)
>
> diff --git a/arch/mips/kernel/entry.S b/arch/mips/kernel/entry.S
> index 7791840..ef69a64 100644
> --- a/arch/mips/kernel/entry.S
> +++ b/arch/mips/kernel/entry.S
> @@ -24,7 +24,7 @@
> #define __ret_from_irq ret_from_exception
> #endif
>
> - .text
> + .section .text..no_watch
> .align 5
> #ifndef CONFIG_PREEMPT
> FEXPORT(ret_from_exception)
> diff --git a/arch/mips/kernel/genex.S b/arch/mips/kernel/genex.S
> index dc0b296..102a9e8 100644
> --- a/arch/mips/kernel/genex.S
> +++ b/arch/mips/kernel/genex.S
> @@ -433,6 +433,7 @@ NESTED(nmi_handler, PT_SIZE, sp)
> BUILD_HANDLER ftlb ftlb none silent /* #16 */
> BUILD_HANDLER msa msa sti silent /* #21 */
> BUILD_HANDLER mdmx mdmx sti silent /* #22 */
> +.section .text..no_watch
> #ifdef CONFIG_HARDWARE_WATCHPOINTS
> /*
> * For watch, interrupts will be enabled after the watch
> @@ -442,6 +443,7 @@ NESTED(nmi_handler, PT_SIZE, sp)
> #else
> BUILD_HANDLER watch watch sti verbose /* #23 */
> #endif
> +.previous
> BUILD_HANDLER mcheck mcheck cli verbose /* #24 */
> BUILD_HANDLER mt mt sti silent /* #25 */
> BUILD_HANDLER dsp dsp sti silent /* #26 */
> diff --git a/arch/mips/kernel/ptrace.c b/arch/mips/kernel/ptrace.c
> index c8ba260..1c8d75c 100644
> --- a/arch/mips/kernel/ptrace.c
> +++ b/arch/mips/kernel/ptrace.c
> @@ -235,6 +235,17 @@ int ptrace_get_watch_regs(struct task_struct *child,
> return 0;
> }
>
> +static int ptrace_watch_in_watch_code_region(unsigned long addr)
> +{
> + extern unsigned long __nowatch_text_start, __nowatch_text_end;
> + unsigned long start, end;
> +
> + start = (((unsigned long)&__nowatch_text_start) & ~MIPS_WATCHLO_IRW);
> + end = (((unsigned long)&__nowatch_text_end) & ~MIPS_WATCHLO_IRW);
> +
> + return addr >= start && addr < end;
> +}
> +
> int ptrace_set_watch_regs(struct task_struct *child,
> struct pt_watch_regs __user *addr)
> {
> @@ -262,6 +273,9 @@ int ptrace_set_watch_regs(struct task_struct *child,
> return -EINVAL;
> }
> #endif
> + if (ptrace_watch_in_watch_code_region(lt[i]))
> + return -EINVAL;
This would apparently permit userland to probe for the address of the
kernel on EVA kernels, regardless of KASLR (assuming that works with
EVA). Perhaps in that case we should be more restrictive and disallow
any watchpoints in the same segment as the kernel.
Actually for stable kernel purposes, it might be better to do that
first, i.e. disallow any to the same segments as kernel (tagged for
stable), then relax it as you've done here.
> +
> __get_user(ht[i], &addr->WATCH_STYLE.watchhi[i]);
> if (ht[i] & ~MIPS_WATCHHI_MASK)
> return -EINVAL;
> diff --git a/arch/mips/kernel/traps.c b/arch/mips/kernel/traps.c
> index 6c7f9d7..b86ce85 100644
> --- a/arch/mips/kernel/traps.c
> +++ b/arch/mips/kernel/traps.c
> @@ -1509,6 +1509,8 @@ asmlinkage void do_mdmx(struct pt_regs *regs)
> * Called with interrupts disabled.
> */
> asmlinkage void do_watch(struct pt_regs *regs)
> + __attribute__((section(".text..no_watch")));
maybe worth a macro, especially if any other functions do need this.
is a separate declaration really needed? could it go after void in the
definition like notrace does?
Cheers
James
> +asmlinkage void do_watch(struct pt_regs *regs)
> {
> siginfo_t info = { .si_signo = SIGTRAP, .si_code = TRAP_HWBKPT };
> enum ctx_state prev_state;
> diff --git a/arch/mips/kernel/vmlinux.lds.S b/arch/mips/kernel/vmlinux.lds.S
> index d5de675..f76f481 100644
> --- a/arch/mips/kernel/vmlinux.lds.S
> +++ b/arch/mips/kernel/vmlinux.lds.S
> @@ -32,6 +32,13 @@ PHDRS {
> jiffies = jiffies_64;
> #endif
>
> +#define NOWATCH_TEXT \
> + ALIGN_FUNCTION(); \
> + VMLINUX_SYMBOL(__nowatch_text_start) = .; \
> + *(.text..no_watch) \
> + VMLINUX_SYMBOL(__nowatch_text_end) = .;
> +
> +
> SECTIONS
> {
> #ifdef CONFIG_BOOT_ELF64
> @@ -54,6 +61,7 @@ SECTIONS
> _text = .; /* Text and read-only data */
> .text : {
> TEXT_TEXT
> + NOWATCH_TEXT
> SCHED_TEXT
> CPUIDLE_TEXT
> LOCK_TEXT
> --
> 2.7.4
>
Attachment:
signature.asc
Description: Digital signature
