Re: [PATCH] lib/vdso: Use __arch_use_vsyscall() to indicate fallback

[Andy Lutomirski]

On Thu, Oct 17, 2019 at 7:57 PM Huacai Chen <chenhc@xxxxxxxxxx> wrote:
>
> In do_hres(), we currently use whether the return value of __arch_get_
> hw_counter() is negtive to indicate fallback, but this is not a good
> idea. Because:
>
> 1, ARM64 returns ULL_MAX but MIPS returns 0 when clock_mode is invalid;
> 2, For a 64bit counter, a "negtive" value of counter is actually valid.

s/negtive/negative

What's the actual bug?  Is it that MIPS is returning 0 but the check
is < 0?  Sounds like MIPS should get fixed.

>
> To solve this problem, we use U64_MAX as the only "invalid" return
> value -- this is still not fully correct, but has no problem in most
> cases.

I'm sort of okay with that, but...

> Moreover, all vdso time-related functions should rely on the
> return value of __arch_use_vsyscall(), because update_vdso_data() and
> update_vsyscall_tz() also rely on it. So, in the core functions of
> __cvdso_gettimeofday(), __cvdso_clock_gettime() and __cvdso_clock_
> getres(), if __arch_use_vsyscall() returns false, we use the fallback
> functions directly.

__arch_use_vsyscall() is not currently intended for use in the vDSO at all.

>
> Fixes: 00b26474c2f1613d7ab894c5 ("lib/vdso: Provide generic VDSO implementation")
> Cc: stable@xxxxxxxxxxxxxxx
> Cc: Arnd Bergmann <arnd@xxxxxxxx>
> Cc: Paul Burton <paul.burton@xxxxxxxx>
> Cc: linux-mips@xxxxxxxxxxxxxxx
> Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> Signed-off-by: Huacai Chen <chenhc@xxxxxxxxxx>
> ---
>  arch/arm64/include/asm/vdso/vsyscall.h |  2 +-
>  arch/mips/include/asm/vdso/vsyscall.h  |  2 +-
>  include/asm-generic/vdso/vsyscall.h    |  2 +-
>  lib/vdso/gettimeofday.c                | 12 +++++++++++-
>  4 files changed, 14 insertions(+), 4 deletions(-)
>
> diff --git a/arch/arm64/include/asm/vdso/vsyscall.h b/arch/arm64/include/asm/vdso/vsyscall.h
> index 0c731bf..406e6de 100644
> --- a/arch/arm64/include/asm/vdso/vsyscall.h
> +++ b/arch/arm64/include/asm/vdso/vsyscall.h
> @@ -31,7 +31,7 @@ int __arm64_get_clock_mode(struct timekeeper *tk)
>  #define __arch_get_clock_mode __arm64_get_clock_mode
>
>  static __always_inline
> -int __arm64_use_vsyscall(struct vdso_data *vdata)
> +int __arm64_use_vsyscall(const struct vdso_data *vdata)
>  {
>         return !vdata[CS_HRES_COARSE].clock_mode;
>  }
> diff --git a/arch/mips/include/asm/vdso/vsyscall.h b/arch/mips/include/asm/vdso/vsyscall.h
> index 1953147..8b10dd7 100644
> --- a/arch/mips/include/asm/vdso/vsyscall.h
> +++ b/arch/mips/include/asm/vdso/vsyscall.h
> @@ -29,7 +29,7 @@ int __mips_get_clock_mode(struct timekeeper *tk)
>  #define __arch_get_clock_mode __mips_get_clock_mode
>
>  static __always_inline
> -int __mips_use_vsyscall(struct vdso_data *vdata)
> +int __mips_use_vsyscall(const struct vdso_data *vdata)
>  {
>         return (vdata[CS_HRES_COARSE].clock_mode != VDSO_CLOCK_NONE);
>  }
> diff --git a/include/asm-generic/vdso/vsyscall.h b/include/asm-generic/vdso/vsyscall.h
> index e94b1978..ac05a625 100644
> --- a/include/asm-generic/vdso/vsyscall.h
> +++ b/include/asm-generic/vdso/vsyscall.h
> @@ -26,7 +26,7 @@ static __always_inline int __arch_get_clock_mode(struct timekeeper *tk)
>  #endif /* __arch_get_clock_mode */
>
>  #ifndef __arch_use_vsyscall
> -static __always_inline int __arch_use_vsyscall(struct vdso_data *vdata)
> +static __always_inline int __arch_use_vsyscall(const struct vdso_data *vdata)
>  {
>         return 1;
>  }
> diff --git a/lib/vdso/gettimeofday.c b/lib/vdso/gettimeofday.c
> index e630e7f..4ad062e 100644
> --- a/lib/vdso/gettimeofday.c
> +++ b/lib/vdso/gettimeofday.c
> @@ -9,6 +9,7 @@
>  #include <linux/hrtimer_defs.h>
>  #include <vdso/datapage.h>
>  #include <vdso/helpers.h>
> +#include <vdso/vsyscall.h>
>
>  /*
>   * The generic vDSO implementation requires that gettimeofday.h
> @@ -50,7 +51,7 @@ static int do_hres(const struct vdso_data *vd, clockid_t clk,
>                 cycles = __arch_get_hw_counter(vd->clock_mode);
>                 ns = vdso_ts->nsec;
>                 last = vd->cycle_last;
> -               if (unlikely((s64)cycles < 0))
> +               if (unlikely(cycles == U64_MAX))
>                         return -1;

I would actually prefer:

if (unlikely(cycles < last))

or perhaps:

if (unlikely((s64)(cycles-last) < 0))

which would have the nice side effect of getting rid of the annoying
x86 special case in vdso_calc_delta().  The former version is
compatible with U64_MAX, whereas the latter version would need the
error case to return last-1 or similar.  The benefit of the latter
version is that it can survive wrap-around.

>
>                 ns += vdso_calc_delta(cycles, last, vd->mask, vd->mult);
> @@ -91,6 +92,9 @@ __cvdso_clock_gettime_common(clockid_t clock, struct __kernel_timespec *ts)
>         if (unlikely((u32) clock >= MAX_CLOCKS))
>                 return -1;
>
> +       if (!__arch_use_vsyscall(vd))
> +               return -1;
> +

NAK.  I don't think this is helpful or correct.  It doesn't appear to
do anything valid, and it's racy.

>         /*
>          * Convert the clockid to a bitmask and use it to check which
>          * clocks are handled in the VDSO directly.
> @@ -145,6 +149,9 @@ __cvdso_gettimeofday(struct __kernel_old_timeval *tv, struct timezone *tz)
>  {
>         const struct vdso_data *vd = __arch_get_vdso_data();
>
> +       if (!__arch_use_vsyscall(vd))
> +               return gettimeofday_fallback(tv, tz);
> +

Ditto.

>         if (likely(tv != NULL)) {
>                 struct __kernel_timespec ts;
>
> @@ -189,6 +196,9 @@ int __cvdso_clock_getres_common(clockid_t clock, struct __kernel_timespec *res)
>         if (unlikely((u32) clock >= MAX_CLOCKS))
>                 return -1;
>
> +       if (!__arch_use_vsyscall(vd))
> +               return -1;
> +

Ditto.