Re: [PATCH v5 3/3] power: amd power monitor support

[Tyler Retzlaff <roretzla@linux.microsoft.com>]

On Wed, Aug 23, 2023 at 10:19:39AM +0100, Ferruh Yigit wrote:
> On 8/22/2023 11:30 PM, Konstantin Ananyev wrote:
> > 18/08/2023 14:48, Bruce Richardson пишет:
> >> On Fri, Aug 18, 2023 at 02:25:14PM +0100, Ferruh Yigit wrote:
> >>> On 8/17/2023 3:18 PM, Konstantin Ananyev wrote:
> >>>>
> >>>>>> Caution: This message originated from an External Source. Use
> >>>>>> proper caution
> >>>>>> when opening attachments, clicking links, or responding.
> >>>>>>
> >>>>>>
> >>>>>> On Wed, Aug 16, 2023 at 11:59:59AM -0700, Sivaprasad Tummala wrote:
> >>>>>>> mwaitx allows EPYC processors to enter a implementation dependent
> >>>>>>> power/performance optimized state (C1 state) for a specific
> >>>>>>> period or
> >>>>>>> until a store to the monitored address range.
> >>>>>>>
> >>>>>>> Signed-off-by: Sivaprasad Tummala <sivaprasad.tummala@amd.com>
> >>>>>>> Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
> >>>>>>> ---
> >>>>>>>   lib/eal/x86/rte_power_intrinsics.c | 77
> >>>>>>> +++++++++++++++++++++++++-----
> >>>>>>>   1 file changed, 66 insertions(+), 11 deletions(-)
> >>>>>>>
> >>>>>>> diff --git a/lib/eal/x86/rte_power_intrinsics.c
> >>>>>>> b/lib/eal/x86/rte_power_intrinsics.c
> >>>>>>> index 6eb9e50807..b4754e17da 100644
> >>>>>>> --- a/lib/eal/x86/rte_power_intrinsics.c
> >>>>>>> +++ b/lib/eal/x86/rte_power_intrinsics.c
> >>>>>>> @@ -17,6 +17,60 @@ static struct power_wait_status {
> >>>>>>>        volatile void *monitor_addr; /**< NULL if not currently
> >>>>>>> sleeping
> >>>>>>> */  } __rte_cache_aligned wait_status[RTE_MAX_LCORE];
> >>>>>>>
> >>>>>>> +/**
> >>>>>>> + * These functions uses UMONITOR/UMWAIT instructions and will
> >>>>>>> enter C0.2
> >>>>>> state.
> >>>>>>> + * For more information about usage of these instructions, please
> >>>>>>> +refer to
> >>>>>>> + * Intel(R) 64 and IA-32 Architectures Software Developer's Manual.
> >>>>>>> + */
> >>>>>>> +static void intel_umonitor(volatile void *addr) {
> >>>>>>> +     /* UMONITOR */
> >>>>>>> +     asm volatile(".byte 0xf3, 0x0f, 0xae, 0xf7;"
> >>>>>>> +                     :
> >>>>>>> +                     : "D"(addr));
> >>>>>>> +}
> >>>>>>> +
> >>>>>>> +static void intel_umwait(const uint64_t timeout) {
> >>>>>>> +     const uint32_t tsc_l = (uint32_t)timeout;
> >>>>>>> +     const uint32_t tsc_h = (uint32_t)(timeout >> 32);
> >>>>>>> +     /* UMWAIT */
> >>>>>>> +     asm volatile(".byte 0xf2, 0x0f, 0xae, 0xf7;"
> >>>>>>> +                     : /* ignore rflags */
> >>>>>>> +                     : "D"(0), /* enter C0.2 */
> >>>>>>> +                     "a"(tsc_l), "d"(tsc_h)); }
> >>>>>>
> >>>>>> question and perhaps Anatoly Burakov can chime in with expertise.
> >>>>>>
> >>>>>> gcc/clang have built-in intrinsics for umonitor and umwait i
> >>>>>> believe as per our other
> >>>>>> thread of discussion is there a benefit to also providing inline
> >>>>>> assembly over just
> >>>>>> using the intrinsics? I understand that the intrinsics may not
> >>>>>> exist for the monitorx
> >>>>>> and mwaitx below so it is probably necessary for amd.
> >>>>>>
> >>>>>> so the suggestion here is when they are available just use the
> >>>>>> intrinsics.
> >>>>>>
> >>>>>> thanks
> >>>>>>
> >>>>> The gcc built-in functions
> >>>>> __builtin_ia32_monitorx()/__builtin_ia32_mwaitx are available only
> >>>>> when -mmwaitx
> >>>>> is used specific for AMD platforms. On generic builds, these
> >>>>> built-ins are not available and hence inline
> >>>>> assembly is required here.
> >>>>
> >>>> Ok... but we can probably put them into a separate .c file that will
> >>>> be compiled with that specific flag?
> >>>> Same thing can be probably done for Intel specific instructions.
> >>>> In general, I think it is much more preferable to use built-ins vs
> >>>> inline assembly
> >>>> (if possible off-course).
> >>>>
> >>>
> >>> We don't compile different set of files for AMD and Intel, but there are
> >>> runtime checks, so putting into separate file is not much different.
> > 
> > Well, we probably don't compile .c files for particular vendor, but we
> > definitely do compile some .c files for particular ISA extensions.
> > Let say there are files in lib/acl that requires various '-mavx512*'
> > flags, same for other libs and PMDs.
> > So still not clear to me why same approach can't be applied to
> > power_instrincts.c?
> > 
> >>>
> >>> It may be an option to always enable compiler flag (-mmwaitx), I think
> >>> it won't hurt other platforms but I am not sure about implications of
> >>> this to other platforms (what was the motivation for the compiler guys
> >>> to enable these build-ins with specific flag?).
> >>>
> >>> Also this requires detecting compiler that supports 'mmwaitx' or not,
> >>> etc..
> >>>
> >> This is the biggest reason why we have in the past added support for
> >> these
> >> instructions via asm bytes rather than intrinsics. It takes a long
> >> time for
> >> end-user compilers, especially those in LTS releases, to get the
> >> necessary
> >> intrinsics. 
> > 
> > Yep, understand.
> > But why then we can't have both implementations?
> > Let say if WAITPKG is defined we can use builtins for
> > umonitor/umwait/tpause, otherwise we fallback to inline asm implementation.
> > Same story for MWAITX/monitorx.
> > 
> 
> Yes this can be done,
> it can be done either as different .c files per implementation, or as
> #ifdef in same file.
> 
> But eventually asm implementation is required, as fallback, and if we
> will rely on asm implementation anyway, does it worth to have the
> additional checks to be able to use built-in intrinsic?
>
> Does it helps to comment name of the built-in function to inline
> assembly code, to document intention and another possible implementation?

the main value of preferring intrinsics is that when they are available
they also work with msvc/windows. the msvc toolchain does not support
inline asm. so some of the targets have to use intrinsics because that's all
there is.

> 
> >> Consider a user running e.g. RHEL 8, who wants to take
> >> advantages of the latest DPDK features; they should not be required to
> >> upgrade their compiler - and possibly binutils/assembler - to do so.
> >>
> >> /Bruce
> >