Re: [PATCH] Consolidate Linux readahead() implementations

[Adhemerval Zanella <adhemerval dot zanella at linaro dot org>]

On 23/09/2016 11:11, Yury Norov wrote:
> On Fri, Sep 23, 2016 at 10:32:35AM -0300, Adhemerval Zanella wrote:
>>
>> Indeed, unfortunately tile seems to get its own readahead definition.
>> However I think it should not prevent us to use my previous strategy,
>> we can follow the SH example for pread (where it adds a dummy argument
>> before offset), and do something as:
>>
>> sysdeps/unix/sysv/linux/tile/readahead.c
>>
>> #include <sysdep.h>
>>
>> #ifndef _LP64
>> /* Although tile 32-bit ABI passed 64-bit arguments in even registers,
>>    readahead interface does not follow this convention.  */
>> # undef __ALIGNMENT_ARG
>> #endif
>>
>> #include <sysdeps/unix/sysv/linux/readhead.c>
> 
> Currently it looks like this to me (see below). If you think that separated file
> is better than new option - I'm OK with it, but I think it's strange because in
> other patches of series you introduce options (if I'm not mistake).

I prefer to not add any more __ASSUME macro until it is more general
and for this specific case (tile seems the only supported ABI that
implements readhead different than usual ABI).

> 
> We also have 2 another implementations - in linux/wordsize-64/syscalls.list
> and linux/mips/mips64/n32/syscalls.list.
> 
> I think wordsize-64 is safe to generalize, but I'm worry about mips64. If we'll
> choose adding new options and so having a single file, it seems, we'll have to
> add another option for mips64/n32, like this:

My understanding is mipc64n32 adds it on syscall list to avoid current
default Linux implementation where it splits the off64_t (since mips64n32
passes off64_t in only one register afaik).  I think it is safe here,
since SYSCALL_LL64 for mips64n32 will correctly pass only one argument
instead of splitting it.

> 
> #if __ASSUME_READAHEAD_NO_PAIRS
> # define SYSCALL_LL64(val) (val)
> #endif
> 
> If we choose 3 implementations, we can introduce no option, but have
> generic, tile and mips separated versions.
> 
> Yury.
> 
> --
> diff --git a/sysdeps/unix/sysv/linux/kernel-features.h b/sysdeps/unix/sysv/linux/kernel-features.h
> index 71ce57a..ba7d745 100644
> --- a/sysdeps/unix/sysv/linux/kernel-features.h
> +++ b/sysdeps/unix/sysv/linux/kernel-features.h
> @@ -50,6 +50,11 @@
>  #define __ASSUME_ST_INO_64_BIT		1
>  #endif
>  
> +#ifndef __ASSUME_READAHEAD_ALIGN
> +/* readahead() adds padding to registers if this control is enabled.  */
> +# define __ASSUME_READAHEAD_ALIGN	1
> +#endif
> +
>  /* The statfs64 syscalls are available in 2.5.74 (but not for alpha).  */
>  #define __ASSUME_STATFS64	1
>  
> diff --git a/sysdeps/unix/sysv/linux/readahead.c b/sysdeps/unix/sysv/linux/readahead.c
> index 92e5428..eea3142 100644
> --- a/sysdeps/unix/sysv/linux/readahead.c
> +++ b/sysdeps/unix/sysv/linux/readahead.c
> @@ -23,16 +23,20 @@
>  #include <sysdep.h>
>  #include <sys/syscall.h>
>  
> +#include <kernel-features.h>
>  
>  #ifdef __NR_readahead
>  
> +#if !__ASSUME_READAHEAD_ALIGN
> +# undef __ALIGNMENT_ARG
> +# define __ALIGNMENT_ARG
> +#endif
> +
>  ssize_t
>  __readahead (int fd, off64_t offset, size_t count)
>  {
> -  return INLINE_SYSCALL (readahead, 4, fd,
> -			 __LONG_LONG_PAIR ((off_t) (offset >> 32),
> -					   (off_t) (offset & 0xffffffff)),
> -			 count);
> +  return INLINE_SYSCALL_CALL (readahead, fd, __ALIGNMENT_ARG
> +			      SYSCALL_LL64 (offset));
>  }
>  #else
>  ssize_t
> diff --git a/sysdeps/unix/sysv/linux/tile/kernel-features.h b/sysdeps/unix/sysv/linux/tile/kernel-features.h
> index ded0e43..15ad2d3 100644
> --- a/sysdeps/unix/sysv/linux/tile/kernel-features.h
> +++ b/sysdeps/unix/sysv/linux/tile/kernel-features.h
> @@ -24,4 +24,5 @@
>  #ifndef _LP64
>  # define __ASSUME_ALIGNED_REGISTER_PAIRS	1
>  # define __ASSUME_FADVISE64_64_NO_ALIGN		1
> +# define __ASSUME_READAHEAD_ALIGN		0
>  #endif
>