Re: [dpdk-dev] [EXT] [PATCH 5/6] build: add option for armv8 crypto extension

[Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>]

> >>> On Apr 15, 2019, at 1:13 PM, Honnappa Nagarahalli
> >>> <Honnappa.Nagarahalli@arm.com> wrote:
> >>>
> >>>>>>> Subject: [EXT] [PATCH 5/6] build: add option for armv8 crypto
> >>>>>>> extension
> >>>>>>>
> >>>>>>> CONFIG_RTE_MACHINE="armv8a"
> >>>>>>> +CONFIG_RTE_ENABLE_ARMV8_CRYPTO=y
> >>>>>>
> >>>>>> This approach is not scalable. Even, it is not good for BlueField
> >>>>>> as you you need to maintain two images.
> >>>>>>
> >>>>>> Unlike other CPU flags, arm64's crypto cpu flag is really _optional_.
> >>>>>> Access to crypto instructions is always at under runtime check.
> >>>>>> See the following in rte_armv8_pmd.c
> >>>>>>
> >>>>>>
> >>>>>>   /* Check CPU for support for AES instruction set */
> >>>>>>   if (!rte_cpu_get_flag_enabled(RTE_CPUFLAG_AES)) {
> >>>>>>       ARMV8_CRYPTO_LOG_ERR(
> >>>>>>           "AES instructions not supported by CPU");
> >>>>>>       return -EFAULT;
> >>>>>>   }
> >>>>>>
> >>>>>>   /* Check CPU for support for SHA instruction set */
> >>>>>>   if (!rte_cpu_get_flag_enabled(RTE_CPUFLAG_SHA1) ||
> >>>>>>       !rte_cpu_get_flag_enabled(RTE_CPUFLAG_SHA2)) {
> >>>>>>       ARMV8_CRYPTO_LOG_ERR(
> >>>>>>           "SHA1/SHA2 instructions not supported by CPU");
> >>>>>>       return -EFAULT;
> >>>>>>   }
> >>>>>>
> >>>>>> So In order to avoid one more config flags specific to armv8 in
> >>>>>> meson and makefile build infra And avoid the need for 6/6 patch.
> >>>>>> IMO, # Introduce optional CPU flag scheme in eal. Treat armv8
> >>>>>> crypto as optional flag # Skip the eal init check for optional flag.
> >>>>>>
> >>>>>> Do you see any issues with that approach?
> >>>>>
> >>>>> I also thought about that approach and that was my number 1 priority.
> >>>>> But, I had one question came to my mind. Maybe, arm people can
> >>>>> confirm it. Is it 100% guaranteed that compiler never makes use of
> >>>>> any of crypto instructions even if there's no specific
> >>>>> asm/intrinsic code?  The crypto extension has aes, pmull,
> >>>>> sha1 and sha2. In case of rte_memcpy() for x86, for example,
> >>>>> compiler may optimize code using avx512f instructions even though
> >>>>> it is written specifically with avx2 intrinsics (__mm256_*) unless
> >>>>> avx512f is
> >>> disabled.
> >>>>>
> >>>>> If a complier expert in arm (or anyone else) confirm it is
> >>>>> completely **optional**, then I'd love to take that approach for sure.
> >>>>>
> >>>>> Copied dpdk-on-arm ML.
> >>>>>
> >>>> I do not know the answer, will have to check with the compiler team.
> >>>> I will get
> >>> back on this.
> >>>
> >>> Any update yet?
> >> Currently, enabling 'crypto' flag will generate the crypto
> >> instructions only when crypto intrinsics are used. However, when
> >> 'sha3' (part of 8.2 crypto) flag is
> >
> > The default image is 8.1 spec and except octeontx2 every other SoC is
I am not following this. I think the default image is 8.0.

> > 8.1 and For octeotx2 crypto is supported. If so, Should we worry this case?
I assume we all are talking about the distro/binary portable build. IMO, we should not just look at the existing SoCs.
The CPU specific builds have the freedom to compile as per their corresponding support.

> 
> Right, it sounds to me that we can disable the option without having the new
> config flag until such instructions get needed. According to gcc-8 release note
> [1], currently '+crypto' implies '+aes' and '+sha2' while '+sha3' and '+sm4' are
> newly introduced. Given that armv8 crypto PMD uses external binary of
> Marvell. I don't see any reason to enable '+crypto'. How about simply disable
> it from armv8 build configs?
I think it should be fine. But, this alone is not enough. The run time detection of the crypto feature and hooking up the correct pointers needs to be added.

> 
> diff --git a/config/arm/meson.build b/config/arm/meson.build index
> 7fa6ed3105..abc8cf346c 100644
> --- a/config/arm/meson.build
> +++ b/config/arm/meson.build
> @@ -74,7 +74,7 @@ flags_octeontx2_extra = [
>         ['RTE_USE_C11_MEM_MODEL', true]]
> 
>  machine_args_generic = [
> -       ['default', ['-march=armv8-a+crc+crypto']],
> +       ['default', ['-march=armv8-a+crc']],
>         ['native', ['-march=native']],
>         ['0xd03', ['-mcpu=cortex-a53']],
>         ['0xd04', ['-mcpu=cortex-a35']], diff --git
> a/mk/machine/armv8a/rte.vars.mk b/mk/machine/armv8a/rte.vars.mk index
> 8252efbb7b..5e3ffc3adf 100644
> --- a/mk/machine/armv8a/rte.vars.mk
> +++ b/mk/machine/armv8a/rte.vars.mk
> @@ -28,4 +28,4 @@
>  # CPU_LDFLAGS =
>  # CPU_ASFLAGS =
> 
> -MACHINE_CFLAGS += -march=armv8-a+crc+crypto
> +MACHINE_CFLAGS += -march=armv8-a+crc
> 
> 
> [1] https://gcc.gnu.org/gcc-8/changes.html
> 
> Thanks,
> Yongseok
> 
> >> enabled, compiler can generate 3-way exclusive OR instructions beyond
> >> the intrinsics.
> >
> > The very same problem will be applicable for Linux kernel too for
> distribution binary case.
> > If the above statement is true about 8.2 crypto and crypto generation
> > without Intrinsics then we need to see how linux kernel handling that
> > and align our solution based on that.
Yes, the compiler team cited Linux kernel example, I have not verified it myself.

> >
> >> Compiler team cannot provide a guarantee that other crypto
> >> instructions will not be used beyond the intrinsics.
> >>
> >> The current suggestion is to use GNU indirect function [1] or
> >> similar. I am not
> >
> > Not sure how it helps? If we know the compiler is generating a
> > specific function With crypto instruction then we can generate
> > _alternative_ function for the same With hwcap?.How do we know which
> function compiler using compiler instructions?
This feature is similar to using function pointers and choosing which function pointer to use at run time. If this feature is used, the function pointer to use is decided during dynamic linking stage.
Either ways, we need to have 2 sets of crypto PMD drivers. One that implements the actual functionality using crypto intrinsics/assembly. Only, this code needs to be compiled with '+crypto'. Second driver that implements just stubs and returns error. This code will be compiled without '+crypto'. At run time, depending on the HWCAP, the correct driver/function pointers need to be hooked up.

> >
> >
> >> sure on GNU indirect function portability.
> >
> > We are using HWCAP scheme, So we may not need the very exact GNU
> > indirect scheme to fix the issue.
Agree, using indirect functions is not a must.

> >
> >>
> >> [1]
> >> https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwil
> >> lnewton.name%2F2013%2F07%2F02%2Fusing-gnu-indirect-
> functions%2F&amp;d
> >>
> ata=02%7C01%7Cyskoh%40mellanox.com%7Cda8fb7ed03e7406ded8908d6c
> ee6d759
> >> %7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C63692388818
> 9316743&amp;
> >>
> sdata=x5XNd5WZ3EtiprPMiFzaskvigX8K0AoXA2w%2BKiN156c%3D&amp;res
> erved=0
> >>
> >>>
> >>> Thanks
> >>> Yongseok