On Tue, Jul 25, 2017 at 11:33 AM, Kees Cook <[email protected]> wrote:
> While not a crash test, this does provide two tight atomic_t and
> refcount_t loops for performance comparisons:
>
> cd /sys/kernel/debug/provoke-crash
> perf stat -B -- cat <(echo ATOMIC_TIMING) > DIRECT
> perf stat -B -- cat <(echo REFCOUNT_TIMING) > DIRECT
>
> Looking a CPU cycles is the best way to example the fast-path (rather
> than instruction counts, since conditional jumps will be executed but
> will be negligible due to branch-prediction).
>
> Signed-off-by: Kees Cook <[email protected]>
> ---
> drivers/misc/lkdtm.h | 2 ++
> drivers/misc/lkdtm_core.c | 2 ++
> drivers/misc/lkdtm_refcount.c | 44
> +++++++++++++++++++++++++++++++++++++++++++
> 3 files changed, 48 insertions(+)
>
> diff --git a/drivers/misc/lkdtm.h b/drivers/misc/lkdtm.h
> index 241d6f3a26dd..0b86947b31f5 100644
> --- a/drivers/misc/lkdtm.h
> +++ b/drivers/misc/lkdtm.h
> @@ -57,6 +57,8 @@ void lkdtm_REFCOUNT_DEC_SATURATED(void);
> void lkdtm_REFCOUNT_ADD_SATURATED(void);
> void lkdtm_REFCOUNT_DEC_AND_TEST_SATURATED(void);
> void lkdtm_REFCOUNT_SUB_AND_TEST_SATURATED(void);
> +void lkdtm_REFCOUNT_TIMING(void);
> +void lkdtm_ATOMIC_TIMING(void);
>
> /* lkdtm_rodata.c */
> void lkdtm_rodata_do_nothing(void);
> diff --git a/drivers/misc/lkdtm_core.c b/drivers/misc/lkdtm_core.c
> index c3b045aec10e..5f949a05b307 100644
> --- a/drivers/misc/lkdtm_core.c
> +++ b/drivers/misc/lkdtm_core.c
> @@ -234,6 +234,8 @@ struct crashtype crashtypes[] = {
> CRASHTYPE(REFCOUNT_ADD_SATURATED),
> CRASHTYPE(REFCOUNT_DEC_AND_TEST_SATURATED),
> CRASHTYPE(REFCOUNT_SUB_AND_TEST_SATURATED),
> + CRASHTYPE(REFCOUNT_TIMING),
> + CRASHTYPE(ATOMIC_TIMING),
> CRASHTYPE(USERCOPY_HEAP_SIZE_TO),
> CRASHTYPE(USERCOPY_HEAP_SIZE_FROM),
> CRASHTYPE(USERCOPY_HEAP_FLAG_TO),
> diff --git a/drivers/misc/lkdtm_refcount.c b/drivers/misc/lkdtm_refcount.c
> index 234d39a07a99..8bad7baf0119 100644
> --- a/drivers/misc/lkdtm_refcount.c
> +++ b/drivers/misc/lkdtm_refcount.c
> @@ -306,3 +306,47 @@ void lkdtm_REFCOUNT_SUB_AND_TEST_SATURATED(void)
>
> check_saturated(&sat);
> }
> +
> +/* Used to time the existing atomic_t when used for reference counting */
> +void lkdtm_ATOMIC_TIMING(void)
> +{
> + unsigned int i;
> + atomic_t count = ATOMIC_INIT(1);
> +
> + for (i = 0; i < INT_MAX; i++)
> + atomic_inc(&count);
> +
> + for (i = INT_MAX; i >= 0; i--)
> + if (atomic_dec_and_test(&count))
> + break;
> +
> + if (i != 0)
> + pr_err("atomic timing: out of sync up/down cycle: %u\n", i);
> + else
> + pr_info("atomic timing: done\n");
> +}
> +
> +/*
> + * This can be compared to ATOMIC_TIMING when implementing fast refcount
> + * protections. Looking at the number of CPU cycles tells the real story
> + * about performance. For example:
> + * cd /sys/kernel/debug/provoke-crash
> + * perf stat -B -- cat <(echo REFCOUNT_TIMING) > DIRECT
> + */
> +void lkdtm_REFCOUNT_TIMING(void)
> +{
> + unsigned int i;
> + refcount_t count = REFCOUNT_INIT(1);
> +
> + for (i = 0; i < INT_MAX; i++)
Oops, bug snuck in at the last moment: this INT_MAX above and below
(also in atomic_t above) should be INT_MAX - 1 since we start at 1,
not zero.
> + refcount_inc(&count);
> +
> + for (i = INT_MAX; i >= 0; i--)
> + if (refcount_dec_and_test(&count))
> + break;
> +
> + if (i != 0)
> + pr_err("refcount: out of sync up/down cycle: %u\n", i);
> + else
> + pr_info("refcount timing: done\n");
> +}
> --
> 2.7.4
>
--
Kees Cook
Pixel Security
