Re: [PATCH v4 08/15] ARC: Linux Syscall Interface

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

+CC a few folks working on 64-bit time changes

On 3/26/20 4:52 PM, Joseph Myers wrote:
> On Thu, 12 Mar 2020, Vineet Gupta via Libc-alpha wrote:
> 
>> diff --git a/sysdeps/unix/sysv/linux/arc/bits/socket-constants.h b/sysdeps/unix/sysv/linux/arc/bits/socket-constants.h
>> new file mode 100644
>> index 000000000000..74b0c06edb36
>> --- /dev/null
>> +++ b/sysdeps/unix/sysv/linux/arc/bits/socket-constants.h
> 
> As far as I can see, the only reason for having this header is to use 
> appropriate values of SO_RCVTIMEO and SO_SNDTIMEO for 64-bit time.
> 
> It's best not to need a separate version of this header for all future 
> 32-bit architectures.  So I'd suggest putting the conditionals on the size 
> of time_t in the default sysdeps/unix/sysv/linux/bits/socket-constants.h 
> instead.

OK this was loosely based on RV32 port. I can fix as you say, as a
standalone/separate patch I suppose.

But in grep'ing I see a weird thing: SO_RCVTIMEO in user exposed socket.h has a
totally different value from socket-constants.h - how is that supposed to work.

$git grep SO_RCVTIME

bits/socket.h:348:    SO_RCVTIMEO = 0x1006,     /* Receive timeout.  */
bits/socket.h:349:#define SO_RCVTIMEO SO_RCVTIMEO
sysdeps/unix/sysv/linux/arc/bits/socket-constants.h:33:#define SO_RCVTIMEO 66
sysdeps/unix/sysv/linux/bits/socket-constants.h:33:#define SO_RCVTIMEO 20
sysdeps/unix/sysv/linux/riscv/bits/socket-constants.h:34:# define SO_RCVTIMEO 66
sysdeps/unix/sysv/linux/riscv/bits/socket-constants.h:36:# define SO_RCVTIMEO 20


> Should the choice of values of these macros, for normal user programs, be 
> based on the time_t for the current compilation (as affected by _TIME_BITS 
> in future) or on the time_t for the default ABI in glibc? 

If that's a question for me, I'm afraid Its above my paygrade :-)

> If the latter, 
> it would be a __TIMESIZE conditional in the header (that is, a conditional 
> based on __TIMESIZE == __SYSCALL_WORDSIZE, or using __WORDSIZE when 
> __SYSCALL_WORDSIZE is not defined, since the actual kernel conditional is 
> about whether using the same time_t size as kernel "long"). 


> If the 
> former, it might use __TIMESIZE right now, but that would need to change 
> when _TIME_BITS is supported to a different macro meaning the time_t size 
> for the current compilation.

This approach seems similar to what Alistair has done for generic/bits/typesizes.h
shall we pursue it ?

> 
>> diff --git a/sysdeps/unix/sysv/linux/arc/jmp_buf-macros.h b/sysdeps/unix/sysv/linux/arc/jmp_buf-macros.h
>> new file mode 100644
>> index 000000000000..6c129398483a
>> --- /dev/null
>> +++ b/sysdeps/unix/sysv/linux/arc/jmp_buf-macros.h
>> @@ -0,0 +1,6 @@
>> +#define JMP_BUF_SIZE		(32 + 1 + 64/(8 * sizeof (unsigned long int))) * sizeof (unsigned long int)
>> +#define SIGJMP_BUF_SIZE		(32 + 1 + 64/(8 * sizeof (unsigned long int))) * sizeof (unsigned long int)
> 
> Missing spaces around '/'.

Fixed.

>> +/* The minimum supported kernel version for ARC is 3.9,
>> +   guaranteeing many kernel features.  */
> 
> But actually it's 5.1 at present.

Right, fixed that.

>> +/* All syscall handler come here to avoid generated code bloat due to
>> +   GOT reference  to errno_location or it's equivalent.  */
>> +int
>> +__syscall_error(int err_no)
> 
> Missing space before '('

Oops, fixed now.

> 
>> +{
>> +  __set_errno(-err_no);
> 
> Likewise.

Fixed.
_______________________________________________
linux-snps-arc mailing list
linux-snps-arc@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-snps-arc