linux-kernel - Re: [PATCH 2/2] lkdtm: Provide timing tests for atomic_t vs refcount

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Tue, 25 Jul 2017 12:25:57 -0700
From:   Kees Cook <keescook@...omium.org>
To:     LKML <linux-kernel@...r.kernel.org>
Cc:     Kees Cook <keescook@...omium.org>,
        Ard Biesheuvel <ard.biesheuvel@...aro.org>,
        Li Kun <hw.likun@...wei.com>, Arnd Bergmann <arnd@...db.de>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>
Subject: Re: [PATCH 2/2] lkdtm: Provide timing tests for atomic_t vs refcount_t

On Tue, Jul 25, 2017 at 11:33 AM, Kees Cook <keescook@...omium.org> 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 <keescook@...omium.org>
> ---
>  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