Re: [RFC PATCH v3 08/12] lib: vdso: allow arches to provide vdso data pointer
On Thu, Jan 16, 2020 at 2:35 AM Thomas Gleixner wrote:
>
> static __maybe_unused int
> __cvdso_data_clock_gettime(clockid_t clock, struct __kernel_timespec *ts,
>const struct vdso_data *vd)
> {
> .
> }
>
> static __maybe_unused int
> __cvdso_clock_gettime(clockid_t clock, struct __kernel_timespec *ts)
> {
> const struct vdso_data *vd = __arch_get_vdso_data();
>
> return __cvdso_data_clock_gettime(clock, ts, vd);
> }
>
> and then use __cvdso_data_clock_gettime on PPC and let the other archs
> unmodified.
>
>
FWIW, I did some experiments on x86 with gcc 9.2. gcc 9.2 uses
rip-relative accesses if I simplify the config enough and otherwise
materializes the pointer. Presumably it decides that the code size
reduction is worth it if there are a lot of accesses.
I suspect that tglx's suggestion will be fine or at worst will add
negligible overhead on x86_64.
Re: [RFC PATCH v3 08/12] lib: vdso: allow arches to provide vdso data pointer
Christophe Leroy writes:
> Le 15/01/2020 à 07:15, Christophe Leroy a écrit :
> From your point of view, what should I do:
> A/ __arch_get_vdso_data() handled entirely at arch level and arches
> handing over the vdso data pointer to generic C VDSO functions all the
> time (as in my v2 series) ?
No. That's again moving the same code to all architectures.
> B/ Data pointer being handed over all the way up for arches wanting to
> do so, no changes at all for others (as in my v3 series) ?
Too much ifdeffery
> C/ __arch_get_vdso_data() being called at the outermost generic level
> for arches not interested in handling data pointer from the caller (as
> suggested by Thomas) ?
>
> Andy, with A/ you were concerned about arches being able to do PC
> related accesses. Would it be an issue for C/ as well ? If not, I guess
> C/ would be cleaner than B/ allthought not as clean as A which doesn't
> add any #ifdefery at all.
You can avoid ifdeffery with C if you do:
static __maybe_unused int
__cvdso_data_clock_gettime(clockid_t clock, struct __kernel_timespec *ts,
const struct vdso_data *vd)
{
.
}
static __maybe_unused int
__cvdso_clock_gettime(clockid_t clock, struct __kernel_timespec *ts)
{
const struct vdso_data *vd = __arch_get_vdso_data();
return __cvdso_data_clock_gettime(clock, ts, vd);
}
and then use __cvdso_data_clock_gettime on PPC and let the other archs
unmodified.
Thanks,
tglx
Re: [RFC PATCH v3 08/12] lib: vdso: allow arches to provide vdso data pointer
Thomas, Andy,
Le 15/01/2020 à 07:15, Christophe Leroy a écrit :
Le 15/01/2020 à 00:06, Thomas Gleixner a écrit :
Christophe Leroy writes:
static __maybe_unused int
+#ifdef VDSO_GETS_VD_PTR_FROM_ARCH
+__cvdso_clock_gettime_common(const struct vdso_data *vd, clockid_t
clock,
+ struct __kernel_timespec *ts)
+{
+#else
__cvdso_clock_gettime_common(clockid_t clock, struct
__kernel_timespec *ts)
{
const struct vdso_data *vd = __arch_get_vdso_data();
+#endif
u32 msk;
If we do that, then there is no point in propagating this to the inner
functions. It's perfectly fine to have this distinction at the outermost
level.
In v2, I did it at the arch level (see
https://patchwork.ozlabs.org/patch/1214983/). Andy was concerned about
it being suboptimal for arches which (unlike powerpc) have PC related
data addressing mode.
Wouldn't it be the same issue if doing it at the outermost level of
generic VDSO ?
Any opinion on this ?
From your point of view, what should I do:
A/ __arch_get_vdso_data() handled entirely at arch level and arches
handing over the vdso data pointer to generic C VDSO functions all the
time (as in my v2 series) ?
B/ Data pointer being handed over all the way up for arches wanting to
do so, no changes at all for others (as in my v3 series) ?
C/ __arch_get_vdso_data() being called at the outermost generic level
for arches not interested in handling data pointer from the caller (as
suggested by Thomas) ?
Andy, with A/ you were concerned about arches being able to do PC
related accesses. Would it be an issue for C/ as well ? If not, I guess
C/ would be cleaner than B/ allthought not as clean as A which doesn't
add any #ifdefery at all.
Thanks
Christophe
Re: [RFC PATCH v3 08/12] lib: vdso: allow arches to provide vdso data pointer
Le 15/01/2020 à 00:06, Thomas Gleixner a écrit :
Christophe Leroy writes:
static __maybe_unused int
+#ifdef VDSO_GETS_VD_PTR_FROM_ARCH
+__cvdso_clock_gettime_common(const struct vdso_data *vd, clockid_t clock,
+ struct __kernel_timespec *ts)
+{
+#else
__cvdso_clock_gettime_common(clockid_t clock, struct __kernel_timespec *ts)
{
const struct vdso_data *vd = __arch_get_vdso_data();
+#endif
u32 msk;
If we do that, then there is no point in propagating this to the inner
functions. It's perfectly fine to have this distinction at the outermost
level.
In v2, I did it at the arch level (see
https://patchwork.ozlabs.org/patch/1214983/). Andy was concerned about
it being suboptimal for arches which (unlike powerpc) have PC related
data addressing mode.
Wouldn't it be the same issue if doing it at the outermost level of
generic VDSO ?
As a related question, I noticed that you keep all that ASM voodoo in
the PPC specific code which provides the actual entry points. Is that
ASM code really still necessary? All current users of the generic VDSO
just do something like:
int __vdso_clock_gettime(clockid_t clock, struct __kernel_timespec *ts)
{
return __cvdso_clock_gettime(clock, ts);
}
in the architecture code. Is there a reason why this can't work on PPC?
The problem with powerpc is that VDSO functions have to (just like
system calls) set the SO bit in CR register in case of error, or clear
it if no error. There is no way to do that from the C function, because
there is no way to tell GCC to not play up with CR register on function
return.
Refer discussion at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92769
Christophe
Re: [RFC PATCH v3 08/12] lib: vdso: allow arches to provide vdso data pointer
Christophe Leroy writes:
>
> static __maybe_unused int
> +#ifdef VDSO_GETS_VD_PTR_FROM_ARCH
> +__cvdso_clock_gettime_common(const struct vdso_data *vd, clockid_t clock,
> + struct __kernel_timespec *ts)
> +{
> +#else
> __cvdso_clock_gettime_common(clockid_t clock, struct __kernel_timespec *ts)
> {
> const struct vdso_data *vd = __arch_get_vdso_data();
> +#endif
> u32 msk;
If we do that, then there is no point in propagating this to the inner
functions. It's perfectly fine to have this distinction at the outermost
level.
As a related question, I noticed that you keep all that ASM voodoo in
the PPC specific code which provides the actual entry points. Is that
ASM code really still necessary? All current users of the generic VDSO
just do something like:
int __vdso_clock_gettime(clockid_t clock, struct __kernel_timespec *ts)
{
return __cvdso_clock_gettime(clock, ts);
}
in the architecture code. Is there a reason why this can't work on PPC?
Thanks,
tglx
