Re: [PATCH v4] arch/riscv: add Zihintpause support

From: Conor.Dooley
Date: Fri Aug 12 2022 - 03:21:50 EST


On 12/08/2022 07:57, Conor Dooley - M52691 wrote:
> On 11/08/2022 16:17, Palmer Dabbelt wrote:
>> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
>>
>> On Mon, 20 Jun 2022 13:15:25 PDT (-0700), daolu@xxxxxxxxxxxx wrote:
>>> Implement support for the ZiHintPause extension.
>>>
>>> The PAUSE instruction is a HINT that indicates the current hart’s rate
>>> of instruction retirement should be temporarily reduced or paused.
>>>
>>> Reviewed-by: Heiko Stuebner <heiko@xxxxxxxxx>
>>> Tested-by: Heiko Stuebner <heiko@xxxxxxxxx>
>>> Signed-off-by: Dao Lu <daolu@xxxxxxxxxxxx>
>>> ---
>>>
>>> v1 -> v2:
>>>  Remove the usage of static branch, use PAUSE if toolchain supports it
>>> v2 -> v3:
>>>  Added the static branch back, cpu_relax() behavior is kept the same for
>>> systems that do not support ZiHintPause
>>> v3 -> v4:
>>>  Adopted the newly added unified static keys for extensions
>>> ---
>>>  arch/riscv/Makefile                     |  4 ++++
>>>  arch/riscv/include/asm/hwcap.h          |  5 +++++
>>>  arch/riscv/include/asm/vdso/processor.h | 21 ++++++++++++++++++---
>>>  arch/riscv/kernel/cpu.c                 |  1 +
>>>  arch/riscv/kernel/cpufeature.c          |  1 +
>>>  5 files changed, 29 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
>>> index 34cf8a598617..6ddacc6f44b9 100644
>>> --- a/arch/riscv/Makefile
>>> +++ b/arch/riscv/Makefile
>>> @@ -56,6 +56,10 @@ riscv-march-$(CONFIG_RISCV_ISA_C)  := $(riscv-march-y)c
>>>  toolchain-need-zicsr-zifencei := $(call cc-option-yn, -march=$(riscv-march-y)_zicsr_zifencei)
>>>  riscv-march-$(toolchain-need-zicsr-zifencei) := $(riscv-march-y)_zicsr_zifencei
>>>
>>> +# Check if the toolchain supports Zihintpause extension
>>> +toolchain-supports-zihintpause := $(call cc-option-yn, -march=$(riscv-march-y)_zihintpause)
>>> +riscv-march-$(toolchain-supports-zihintpause) := $(riscv-march-y)_zihintpause
>>> +
>>>  KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
>>>  KBUILD_AFLAGS += -march=$(riscv-march-y)
>>>
>>> diff --git a/arch/riscv/include/asm/hwcap.h b/arch/riscv/include/asm/hwcap.h
>>> index e48eebdd2631..dc47019a0b38 100644
>>> --- a/arch/riscv/include/asm/hwcap.h
>>> +++ b/arch/riscv/include/asm/hwcap.h
>>> @@ -8,6 +8,7 @@
>>>  #ifndef _ASM_RISCV_HWCAP_H
>>>  #define _ASM_RISCV_HWCAP_H
>>>
>>> +#include <asm/errno.h>
>>>  #include <linux/bits.h>
>>>  #include <uapi/asm/hwcap.h>
>>>
>>> @@ -54,6 +55,7 @@ extern unsigned long elf_hwcap;
>>>  enum riscv_isa_ext_id {
>>>       RISCV_ISA_EXT_SSCOFPMF = RISCV_ISA_EXT_BASE,
>>>       RISCV_ISA_EXT_SVPBMT,
>>> +     RISCV_ISA_EXT_ZIHINTPAUSE,
>>>       RISCV_ISA_EXT_ID_MAX = RISCV_ISA_EXT_MAX,
>>>  };
>>>
>>> @@ -64,6 +66,7 @@ enum riscv_isa_ext_id {
>>>   */
>>>  enum riscv_isa_ext_key {
>>>       RISCV_ISA_EXT_KEY_FPU,          /* For 'F' and 'D' */
>>> +     RISCV_ISA_EXT_KEY_ZIHINTPAUSE,
>>>       RISCV_ISA_EXT_KEY_MAX,
>>>  };
>>>
>>> @@ -83,6 +86,8 @@ static __always_inline int riscv_isa_ext2key(int num)
>>>               return RISCV_ISA_EXT_KEY_FPU;
>>>       case RISCV_ISA_EXT_d:
>>>               return RISCV_ISA_EXT_KEY_FPU;
>>> +     case RISCV_ISA_EXT_ZIHINTPAUSE:
>>> +             return RISCV_ISA_EXT_KEY_ZIHINTPAUSE;
>>>       default:
>>>               return -EINVAL;
>>>       }
>>> diff --git a/arch/riscv/include/asm/vdso/processor.h b/arch/riscv/include/asm/vdso/processor.h
>>> index 134388cbaaa1..1e4f8b4aef79 100644
>>> --- a/arch/riscv/include/asm/vdso/processor.h
>>> +++ b/arch/riscv/include/asm/vdso/processor.h
>>> @@ -4,15 +4,30 @@
>>>
>>>  #ifndef __ASSEMBLY__
>>>
>>> +#include <linux/jump_label.h>
>>>  #include <asm/barrier.h>
>>> +#include <asm/hwcap.h>
>>>
>>>  static inline void cpu_relax(void)
>>>  {
>>> +     if (!static_branch_likely(&riscv_isa_ext_keys[RISCV_ISA_EXT_KEY_ZIHINTPAUSE])) {
>>>  #ifdef __riscv_muldiv
>>> -     int dummy;
>>> -     /* In lieu of a halt instruction, induce a long-latency stall. */
>>> -     __asm__ __volatile__ ("div %0, %0, zero" : "=r" (dummy));
>>> +             int dummy;
>>> +             /* In lieu of a halt instruction, induce a long-latency stall. */
>>> +             __asm__ __volatile__ ("div %0, %0, zero" : "=r" (dummy));
>>>  #endif
>>> +     } else {
>>> +             /*
>>> +              * Reduce instruction retirement.
>>> +              * This assumes the PC changes.
>>> +              */
>>> +#ifdef __riscv_zihintpause
>>> +             __asm__ __volatile__ ("pause");
>>> +#else
>>> +             /* Encoding of the pause instruction */
>>> +             __asm__ __volatile__ (".4byte 0x100000F");
>>> +#endif
>>> +     }
>>>       barrier();
>>>  }
>>>
>>> diff --git a/arch/riscv/kernel/cpu.c b/arch/riscv/kernel/cpu.c
>>> index fba9e9f46a8c..a123e92b14dd 100644
>>> --- a/arch/riscv/kernel/cpu.c
>>> +++ b/arch/riscv/kernel/cpu.c
>>> @@ -89,6 +89,7 @@ int riscv_of_parent_hartid(struct device_node *node)
>>>  static struct riscv_isa_ext_data isa_ext_arr[] = {
>>>       __RISCV_ISA_EXT_DATA(sscofpmf, RISCV_ISA_EXT_SSCOFPMF),
>>>       __RISCV_ISA_EXT_DATA(svpbmt, RISCV_ISA_EXT_SVPBMT),
>>> +     __RISCV_ISA_EXT_DATA(zihintpause, RISCV_ISA_EXT_ZIHINTPAUSE),
>>>       __RISCV_ISA_EXT_DATA("", RISCV_ISA_EXT_MAX),
>>>  };
>>>
>>> diff --git a/arch/riscv/kernel/cpufeature.c b/arch/riscv/kernel/cpufeature.c
>>> index 1b3ec44e25f5..708df2c0bc34 100644
>>> --- a/arch/riscv/kernel/cpufeature.c
>>> +++ b/arch/riscv/kernel/cpufeature.c
>>> @@ -198,6 +198,7 @@ void __init riscv_fill_hwcap(void)
>>>                       } else {
>>>                               SET_ISA_EXT_MAP("sscofpmf", RISCV_ISA_EXT_SSCOFPMF);
>>>                               SET_ISA_EXT_MAP("svpbmt", RISCV_ISA_EXT_SVPBMT);
>>> +                             SET_ISA_EXT_MAP("zihintpause", RISCV_ISA_EXT_ZIHINTPAUSE);
>>>                       }
>>>  #undef SET_ISA_EXT_MAP
>>>               }
>>
>> Thanks, this is on for-next.  It needs a sparse patch, which I put in as a link.
>
> This breaks the C=1 build for all toolchains, not just new ones as your sparse
> patch suggests. I amn't 100% what my CI is running, but I replicated this on
> my own machine with:

Argh, I went poking around and my toolchain's binutils etc is newer than I thought.
Good for people searching on lore I suppose...
Sorry!
Conor.

>
> sparse --version
> 0.6.4 (Ubuntu: 0.6.4-2)
>
> ---8<---
>   YACC    scripts/dtc/dtc-parser.tab.[ch]
>   HOSTCC  scripts/dtc/libfdt/fdt.o
>   HOSTCC  scripts/dtc/libfdt/fdt_ro.o
>   HOSTCC  scripts/dtc/libfdt/fdt_wip.o
>   UPD     include/generated/uapi/linux/version.h
>   HOSTCC  scripts/dtc/libfdt/fdt_sw.o
>   UPD     include/config/kernel.release
>   HOSTCC  scripts/dtc/libfdt/fdt_rw.o
>   HOSTCC  scripts/dtc/libfdt/fdt_strerror.o
>   HOSTCC  scripts/dtc/libfdt/fdt_empty_tree.o
>   HOSTCC  scripts/dtc/libfdt/fdt_addresses.o
>   HOSTCC  scripts/dtc/libfdt/fdt_overlay.o
>   HOSTCC  scripts/dtc/fdtoverlay.o
>   HOSTCC  scripts/dtc/dtc-lexer.lex.o
>   HOSTCC  scripts/dtc/dtc-parser.tab.o
>   UPD     include/generated/utsrelease.h
>   HOSTLD  scripts/dtc/fdtoverlay
>   HOSTLD  scripts/dtc/dtc
>   HOSTCC  scripts/kallsyms
>   HOSTCC  scripts/sorttable
>   HOSTCC  scripts/asn1_compiler
>   HOSTCC  scripts/selinux/genheaders/genheaders
>   HOSTCC  scripts/selinux/mdp/mdp
>   CC      scripts/mod/empty.o
>   HOSTCC  scripts/mod/mk_elfconfig
>   CC      scripts/mod/devicetable-offsets.s
>   CHECK   ../scripts/mod/empty.c
> invalid argument to '-march': '_zihintpause'
>
> make[2]: *** [../scripts/Makefile.build:250: scripts/mod/empty.o] Error 1
> make[2]: *** Deleting file 'scripts/mod/empty.o'
> make[2]: *** Waiting for unfinished jobs....
> make[1]: *** [/mnt/automation/corp/workspace/ux-test_upstream-next-develop-cd@2/linux/Makefile:1287: prepare0] Error 2
> make[1]: Leaving directory '/mnt/automation/corp/workspace/ux-test_upstream-next-develop-cd@2/linux/builddir'
> make: *** [Makefile:231: __sub-make] Error 2