Consolidate Linux readahead() implementations

  On Fri, Sep 23, 2016 at 10:32:35AM -0300, Adhemerval Zanella wrote:
> 
> 
> On 23/09/2016 09:44, Yury Norov wrote:
> > On Fri, Sep 23, 2016 at 08:08:19AM +0200, Florian Weimer wrote:
> >> * Yury Norov:
> >>
> >>> On Thu, Sep 22, 2016 at 06:36:22PM -0300, Adhemerval Zanella wrote:
> >>>> Hi Yury, some comments below.
> >>>>
> >>>> On 22/09/2016 17:44, Yury Norov wrote:
> >>>> I think we can use SYSCALL_LL64 plus __ALIGNMENT_ARG here.  The tricky is to pass
> >>>> the correct argument size, since it used by the macro to select the correct 
> >>>> arch-dependent one.  This is exactly why I proposed the patch to add
> >>>> {INTERNAL,INLINE}_SYSCALL_CALL [1], readahead will be simplified to just:
> >>>>
> >>>> ssize_t
> >>>> __readahead (int fd, off64_t offset, size_t count)
> >>>> {
> >>>>   return INLINE_SYSCALL_CALL (readahead, fd,
> >>>> 			      __ALIGNMENT_ARG SYSCALL_LL64 (offset));
> >>>> }
> >>>>
> >>>> I suggest you to wait the {INTERNAL,INLINE}_SYSCALL_CALL patch to land and
> >>>> based this one on it.  I think I addressed most of Florian comments, I just
> >>>> need to check if assembly generated using the new macros is the same as
> >>>> before (which I expect to).
> >>>>
> >>>> [1] https://sourceware.org/ml/libc-alpha/2016-09/msg00452.html
> >>>
> >>> __ALIGNMENT_ARG is controlled by __ASSUME_ALIGNED_REGISTER_PAIRS,
> >>> as well as __ALIGNMENT_COUNT(). Currently we have 4 ports that
> >>> enable it: mips, arm, powerpc and tile. Mips and arm pass 5 arguments,
> >>> so __ASSUME_ALIGNED_REGISTER_PAIRS is what they need. Powerpc uses
> >>> syscalls.list (thanks for pointing it), and so is not affected by this
> >>> change. But tile is still in case. Is my understanding correct that it
> >>> uses linux/readahead.c as implementation? If so, this patch will break
> >>> tile ABI. That's why I introduced new option.
> >>>
> >>> So, as for me we cannot use __ASSUME_ALIGNED_REGISTER_PAIRS because of
> >>> tile. Is it correct?
> >>
> >> Does readahead even work on tile today?  Maybe it's already broken and
> >> this change fixes it. :)
> >>
> >> Chris?
> > 
> > In kernel 32-bit tile expects 4 arguments:
> > 
> > arch/tile/kernel/sys.c:
> > 61 #if !defined(__tilegx__) || defined(CONFIG_COMPAT)
> > 62 
> > 63 #ifdef __BIG_ENDIAN
> > 64 #define SYSCALL_PAIR(name) u32 name ## _hi, u32 name ## _lo
> > 65 #else
> > 66 #define SYSCALL_PAIR(name) u32 name ## _lo, u32 name ## _hi
> > 67 #endif
> > 68 
> > 69 ssize_t sys32_readahead(int fd, SYSCALL_PAIR(offset), u32 count)
> > 70 {
> > 71         return sys_readahead(fd, ((loff_t)offset_hi << 32) | offset_lo, count);
> > 72 }
> > 73 
> > 74 int sys32_fadvise64_64(int fd, SYSCALL_PAIR(offset),
> > 75                        SYSCALL_PAIR(len), int advice)
> > 76 {
> > 77         return sys_fadvise64_64(fd, ((loff_t)offset_hi << 32) | offset_lo,
> > 78                                 ((loff_t)len_hi << 32) | len_lo, advice);
> > 79 }
> > 80 
> > 81 #endif /* 32-bit syscall wrappers */
> >  
> > So it seems we have no chance to consolidate getdents() completely w/o
> > new option. If no objections, I'll send v2 with removing getdents from
> > powerpc syscalls.list, and rework new option in more suitable way.
> > Objections?
> > 
> > Yury.
> > 
> 
> 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).

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:

#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.

--
  

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
>
  

On Fri, Sep 23, 2016 at 11:24:32AM -0300, Adhemerval Zanella wrote:
> 
> 
> 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.

OK. So for v2:
 - remove arm, mips, wordsize-64, powerpc implementations
 - generalize linux/readahead.c like below
 - add tile exception
 - remove all new __ASSUME 

If it's correct, I'll send new patch this evening or tomorrow.

Yury.

> >  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));
  

On 23/09/2016 12:44, Yury Norov wrote:
> On Fri, Sep 23, 2016 at 11:24:32AM -0300, Adhemerval Zanella wrote:
>>
>>
>> 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.
> 
> OK. So for v2:
>  - remove arm, mips, wordsize-64, powerpc implementations
>  - generalize linux/readahead.c like below
>  - add tile exception
>  - remove all new __ASSUME 
> 
> If it's correct, I'll send new patch this evening or tomorrow.

That would be my recommendation.