[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 10 Sep 2020 11:01:31 +0100
From: Vincenzo Frascino <vincenzo.frascino@....com>
To: Arnd Bergmann <arnd@...db.de>, Palmer Dabbelt <palmerdabbelt@...gle.com>
Cc: Rich Felker <dalias@...c.org>, musl@...ts.openwall.com
Subject: Re: riscv32 v2
On 9/10/20 8:36 AM, Arnd Bergmann wrote:
> On Thu, Sep 10, 2020 at 1:08 AM Palmer Dabbelt <palmerdabbelt@...gle.com> wrote:
>> On Wed, 09 Sep 2020 14:36:44 PDT (-0700), dalias@...c.org wrote:
>>> On Wed, Sep 09, 2020 at 02:28:55PM -0700, Palmer Dabbelt wrote:
>>>> On Wed, 09 Sep 2020 13:28:27 PDT (-0700), dalias@...c.org wrote:
>>> Possible addition of vdso clock_gettime isn't a blocker for moving
>>> forward with the musl port, but syscall_arch.h should accurately
>>> describe what's available and should not attempt to use vdso before
>>> it's a public kernel interface (e.g. resolving the question of what
>>> the function name will be). So I think it should be removed for now.
>>
>> Sorry if that was confusing, but I definitely agree.
>>
>> I guess my point was that the lack of VDSO clock functions on rv32 was probably
>> an oversight, but one that shouldn't block the port. We definitely can't just
>> make up a kernel interface, particularly as the reason we don't have these on
>> rv32 is because the generic versions of the functions we're using don't appear
>> to run on 32-bit targets.
>>
>> That probably means there's some more subtle issue, though TBH I don't know
>> enough about the 64-bit-ification of time_t for it to just jump out at me. I
>> don't want to derail the thread too much, but I tried the obvious thing
>
> When the vdso for rv64 was added, there was no time64 support in the
> vdso code in general, as this only came with the "generic vdso" infrastructure
> that was added later on, with commit ad5d1122b82f ("riscv: use vDSO
> common flow to reduce the latency of the time-related functions") in v5.8.
>
> At that point it probably should have been added as well.
>
>> --- a/arch/riscv/kernel/vdso/Makefile
>> +++ b/arch/riscv/kernel/vdso/Makefile
>> @@ -7,9 +7,7 @@ ARCH_REL_TYPE_ABS := R_RISCV_32|R_RISCV_64|R_RISCV_JUMP_SLOT
>> include $(srctree)/lib/vdso/Makefile
>> # Symbols present in the vdso
>> vdso-syms = rt_sigreturn
>> -ifdef CONFIG_64BIT
>> vdso-syms += vgettimeofday
>> -endif
>> vdso-syms += getcpu
>> vdso-syms += flush_icache
>>
>> and it doesn't build. I've added Arnd, who might have a better idea of what's
>> going on. Whatever happens, I think the best bet is to just drop the clock
>> functions (specifically __vdso_{clock_gettime,gettimeofday,clock_getres}) from
>> the rv32 port right now.
>
> For rv32 you need clock_gettime64, not clock_gettime, which in the Linux
> ABI refers to the interface with the old timespec. There was some debate
> over whether clock_getres_time64 and gettimeofday_time64 would make
> sense to be added here, but I have so far leaned to the position that these
> are not as performance critical and not worth the effort.
>
> Vincenzo has argued that we might want to extend the generic vdso code
> to include a number of additional syscall implementations, which would
> then include gettimeofday_time64 and clock_getres_time64.
>
I agree with Arnd, clock_getres_time64 and gettimeofday_time64 were not added in
the original port because not considered as performance critical as
clock_gettime64. We might reconsider if there is a strong use case for those.
> Arnd
>
--
Regards,
Vincenzo
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
