Re: [PATCH v7.1 07/13] ARC: Linux Syscall Interface

[Vineet Gupta <Vineet.Gupta1@synopsys.com>]

On 7/7/20 12:24 PM, Adhemerval Zanella via Libc-alpha wrote:
> 
> 
> On 06/07/2020 22:25, Vineet Gupta wrote:
>> On 7/6/20 6:20 AM, Adhemerval Zanella via Libc-alpha wrote:
>>>>>> diff --git a/sysdeps/unix/sysv/linux/arc/clone.S b/sysdeps/unix/sysv/linux/arc/clone.S
>>>>
>>>>>> diff --git a/sysdeps/unix/sysv/linux/arc/fixup-asm-unistd.h b/sysdeps/unix/sysv/linux/arc/fixup-asm-unistd.h
>>>>
>>>>>> +
>>>>>> +/* Adjustments to ARC asm-generic syscall ABI (3.9 kernel) for 64-bit time_t
>>>>>> +   support.  */
>>>>>> +
>>>>>> +/* fstat64 and fstatat64 need to be replaced with statx.  */
>>>>>> +
>>>>>> +#undef __NR_fstat64
>>>>>> +#undef __NR_fstatat64
>>>>
>>>> This is certainly needed as they are present in ARC arch-syscall.h but we need to
>>>> use statx.
>>>>
>>>>>> +/* Replace all other 32-bit time syscalls with 64-bit variants.  */
>>>>>> +
>>>>>> +# undef __NR_clock_adjtime
>>>>>> +# undef __NR_clock_getres
>>>>>> +# undef __NR_futex
>>>>>> +# undef __NR_mq_timedreceive
>>>>>> +# undef __NR_mq_timedsend
>>>>>> +# undef __NR_ppoll
>>>>>> +# undef __NR_pselect6
>>>>>> +# undef __NR_recvmmsg
>>>>>> +# undef __NR_rt_sigtimedwait
>>>>>> +# undef __NR_sched_rr_get_interval
>>>>>> +# undef __NR_semtimedop
>>>>>> +# undef __NR_timerfd_settime
>>>>>> +# undef __NR_timerfd_gettime
>>>>>> +# undef __NR_utimensat
>>>>>
>>>>> I am trying to understand why these are required since arc does not define 
>>>>> them in arch-syscall.h.
>>>>
>>>> arch-syscall.h doesn't define them precisely due to these being here. When
>>>> update-syscalls is run, the 32-bit syscalls are generated for ARC (since kernel
>>>> ABI provides these because that was v3.9 circa 2013). Adding them
>>>> fixup-asm-unistd.h removes them (perhaps I need to add this in changelog to
>>>> clarify - atleast for myself).
>>>>
>>>>> And the generic implementation should handle the time64 variant.  If they
>>>>> are not this is something we need to handle it.
>>>>
>>>> At the time we we doing this, arch-syscall.h generation was not yet in place,
>>>> however I tried to undef in generic/sysdep.h for TIMESIZE==64. However I was asked
>>>> me to add this to ARC specific fixup-asm-unistd.h
>>>> https://sourceware.org/pipermail/libc-alpha/2020-March/112395.html
>>>> https://sourceware.org/pipermail/libc-alpha/2020-April/112909.html
>>>
>>> My confusion here, I forgot that this header is only used glibcsyscalls.py
>>> to actually generate arch-syscall.h.
>>>
>>> You changes does look correct.
>>
>> Actually we can add a few more entries here which have 64-bit variants.
>>
>> +# undef __NR_clock_gettime
>> +# undef __NR_clock_nanosleep
>> +# undef __NR_clock_settime
>> +# undef __NR_timer_gettime
>> +# undef __NR_timer_settime
> 
> It should not intefere since ARC also defines __ASSUME_TIME64_SYSCALLS
> and the 32-bit fallback syscalls won't be used in this case.

In an ideal world yes, but as we see from hacks in sysdep.h, there's code which
build/works inadvertently. If nothing else, this documents that those syscalls are
not to be used.


> As a side note, now that arch-syscall.h is based on latest kernel version
> and all the 32-bit ABIs with old 32-bit time_t have upstream support for 
> 64-bit time_t we can simplify a bit some implementation by assuming 
> the 64-bit time_t is always defined and adding a fallback only for
> !define __ASSUME_TIME64_SYSCALLS.

OK.

>>>>>> diff --git a/sysdeps/unix/sysv/linux/arc/sysdep.h b/sysdeps/unix/sysv/linux/arc/sysdep.h
>>>>
>>>> ...
>>>>
>>>>>> +/* 32-bit time syscalls are not available, but the redefines allow generic
>>>>>> +   wrappers to work.  */
>>>>>> +#define __NR_clock_adjtime	__NR_clock_adjtime64
>>>>>> +#define __NR_clock_getres	__NR_clock_getres_time64
>>>>>> +#define __NR_futex		__NR_futex_time64
>>>>>> +#define __NR_mq_timedreceive	__NR_mq_timedreceive_time64
>>>>>> +#define __NR_mq_timedsend	__NR_mq_timedsend_time64
>>>>>> +#define __NR_ppoll		__NR_ppoll_time64
>>>>>> +#define __NR_pselect6		__NR_pselect6_time64
>>>>>> +#define __NR_recvmmsg		__NR_recvmmsg_time64
>>>>>> +#define __NR_rt_sigtimedwait	__NR_rt_sigtimedwait_time64
>>>>>> +#define __NR_sched_rr_get_interval	__NR_sched_rr_get_interval_time64
>>>>>> +#define __NR_semtimedop		__NR_semtimedop_time64
>>>>>> +#define __NR_timerfd_gettime	__NR_timerfd_gettime64
>>>>>> +#define __NR_timerfd_settime	__NR_timerfd_settime64
>>>>>> +#define __NR_utimensat		__NR_utimensat_time64
>>>>>
>>>>> As for the fixup-asm-unistd.h, the generic implementation should handle it
>>>>> without the requirement of the ABI to add such tricks.
>>>>
>>>> fixup-asm-unistd.h is different, but this could be avoided. I know for sure that
>>>> ll code literally expects __NR_futex (atleast used to). But I can remove this and
>>>> see what comes out.
>>>>
>>>>>
>>>>> However it seems that we are still missing support for pselect 
>>>>> (__NR_pselect6_time64), recvmmsg (__NR_recvmmsg_time64), sigtimedwait 
>>>>> (__NR_rt_sigtimedwait_time64), and semtimeop (__NR_semtimedop_time64).
>>>>>
>>>>> I think we can add the redefine hack only the aforementioned symbols for
>>>>> now and removed them once we implement the y2038 support on such symbols
>>>>> (since the expected ABI won't change for ARC, only for old ABIs with
>>>>> 32 time_t support).
>>>>
>>>> Sorry /me horribly confused here.
>>>
>>> Sorry for the confusion, I meant that some of these re-defines are superfluous 
>>> and I would like to have the minimum required re-define to enable the ARC 
>>> support, 
>>
>> Right. The generic code needs a bit more work to eliminate the redefines altogether.
>>
>> 1. Following is not needed
>>
>> -#define __NR_clock_adjtime     __NR_clock_adjtime64
>> -#define __NR_sched_rr_get_interval     __NR_sched_rr_get_interval_time64
>> -#define __NR_mq_timedreceive   __NR_mq_timedreceive_time64
>> -#define __NR_mq_timedsend      __NR_mq_timedsend_time64
>> -#define __NR_timerfd_gettime   __NR_timerfd_gettime64
>> -#define __NR_timerfd_settime   __NR_timerfd_settime64
> 
> Ok.
> 
>>
>> 2. The minimum list needed for ARC (with annotations as to which generic file
>> needs fixing).
>>
>> /* Fix sysdeps/unix/sysv/linux/clock_getcpuclockid.c.  */
>> #define __NR_clock_getres      __NR_clock_getres_time64
> 
> It should be simple since there is no need to provide an extra symbol
> for old ABIs, it would be to just use __NR_clock_getres_time64 if it
> is defined.
> 
>> /* Fix sysdeps/nptl/lowlevellock-futex.h.  */
>> #define __NR_futex             __NR_futex_time64
> 
> It seems Lukasz Majewski has sent a patchset to address it (I haven't
> checked the detail yet).
> 
>> /* Fix sysdeps/unix/sysv/linux/pause.c.  */
>> #define __NR_ppoll             __NR_ppoll_time64
> 
> This is another one similar to clock_getcpuclockid.
> 
>> /* Fix sysdeps/unix/sysv/linux/select.c.  */
>> #define __NR_pselect6          __NR_pselect6_time64
> 
> I have a patch for select [1], it a bit more complex because we need to 
> handle both the timeout and the microblaze lacking support of
> pselect6.
> 
>> /* Fix sysdeps/unix/sysv/linux/recvmmsg.c.  */
>> #define __NR_recvmmsg          __NR_recvmmsg_time64
> 
> I also have a patch for this [1], although it still does not have the
> ancillary data from struct msghdr which might return SCM_TIMESTAMP
> information (which returns struct timespec).  The recvmsg also has
> the same issue regarding ancillary data.
> 
>> /* Fix sysdeps/unix/sysv/linux/sigtimedwait.c.  */
>> #define __NR_rt_sigtimedwait   __NR_rt_sigtimedwait_time64
> 
> The patch for is more straighfoward [1].
> 
>> /* Fix sysdeps/unix/sysv/linux/semtimedop.c.  */
>> #define __NR_semtimedop                __NR_semtimedop_time64
> 
> The patch for is also straighfoward [1].
> 
> 
>> /* Hack sysdeps/unix/sysv/linux/generic/utimes.c (need linux/utimes.c).  */
>> #define __NR_utimensat         __NR_utimensat_time64
> 
> Lukasz Majewski also has a patchset for this, I will check this out.

Thanks for tacking this.

>>> so we can cleanup these later once we enable time64_t support on 
>>> old ABIs as well.
>>
>> IMO the cleanup applies to new ABIs too as generic code should handle those cases
>> w/o these workarounds. But that would delay things further for new ports so I
>> suggest we keep the workarounds and clean things up going fwd.
> 
> I agree, this can be worked in parallel.

Thx for being pragmatic. Given ABI freeze we will merge ARC with workarounds and
clean them up in 2.33 ?

>> BTW, if one were to actually go about fixing those, whats the best approach.
>> Consider the simplest case pause(). For !__NR_pause do we replicate the code for
>> ppoll/ppoll64 handling or simply just call ppoll(). Later has a function call
>> overhead) ? Or there is a paradigm to use __syscallxxx_helper() although that
>> still has a function call overhead.
>>
>> Actually the pause case is really simple as there are no args, so just redefine
>> __NR_xxx trick should suffice w/o going into all the explicit
>> interworking/conversion etc.
>>
>> __libc_pause (void)
>> {
>> #ifdef __NR_pause
>>   return SYSCALL_CANCEL (pause);
>> #else
>>
>>   return SYSCALL_CANCEL (ppoll, NULL, 0, NULL, NULL);
>> #endif
>>
> 
> Each implementation has it ows requirements so I can't really say if
> a helper function does make sense for all of them.  For pause
> specifically we can even simplify to since all architectures have
> either ppoll or ppoll_time64:
> 
>   int
>   __libc_pause (void)
>   {
>   #ifdef __NR_ppoll_time64
>     return SYSCALL_CANCEL (ppoll_time64, NULL, 0, NULL, NULL);
>   #else
>     return SYSCALL_CANCEL (ppoll, NULL, 0, NULL, NULL);
>   #endif
>   }

But how is this compatible with older kernels (and perhaps this is a general
question). I mean one/more ABIs minimum kernel would not have the ppoll or ppoll64
so how will new glibc work with such a kernel ? Is it not required to ?
_______________________________________________
linux-snps-arc mailing list
linux-snps-arc@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-snps-arc