[v6,2/2] tools/nolibc: fix up size inflate regression

  As reported and suggested by Willy, the inline __sysret() helper
introduces three types of conversions and increases the size:

(1) the "unsigned long" argument to __sysret() forces a sign extension
from all sys_* functions that used to return 'int'

(2) the comparison with the error range now has to be performed on a
'unsigned long' instead of an 'int'

(3) the return value from __sysret() is a 'long' (note, a signed long)
which then has to be turned back to an 'int' before being returned by the
caller to satisfy the caller's prototype.

To fix up this, firstly, let's use macro instead of inline function to
preserves the input type and avoids these useless conversions (1), (3).

Secondly, since all of the sys_* functions have been converted to return
integer, now, it is able to remove comparison to a 'unsigned long'
-MAX_ERRNO (2) and restore the simple sign comparison as before.

Suggested-by: Willy Tarreau <w@1wt.eu>
Link: https://lore.kernel.org/lkml/20230806095846.GB10627@1wt.eu/
Signed-off-by: Zhangjin Wu <falcon@tinylab.org>
---
 tools/include/nolibc/sys.h | 27 ++++++++++-----------------
 1 file changed, 10 insertions(+), 17 deletions(-)
  

On Sat, Aug 12, 2023 at 05:51:53AM +0800, Zhangjin Wu wrote:
> As reported and suggested by Willy, the inline __sysret() helper
> introduces three types of conversions and increases the size:
> 
> (1) the "unsigned long" argument to __sysret() forces a sign extension
> from all sys_* functions that used to return 'int'
> 
> (2) the comparison with the error range now has to be performed on a
> 'unsigned long' instead of an 'int'
> 
> (3) the return value from __sysret() is a 'long' (note, a signed long)
> which then has to be turned back to an 'int' before being returned by the
> caller to satisfy the caller's prototype.
> 
> To fix up this, firstly, let's use macro instead of inline function to
> preserves the input type and avoids these useless conversions (1), (3).
> 
> Secondly, since all of the sys_* functions have been converted to return
> integer, now, it is able to remove comparison to a 'unsigned long'
> -MAX_ERRNO (2) and restore the simple sign comparison as before.
> 
(...)
> +/* Syscall return helper, set errno as -ret when ret < 0 */
> +#define __sysret(arg)                        \
> +({                                           \
> +	__typeof__(arg) __ret = (arg);       \
> +	if (__ret < 0) {                     \
> +		SET_ERRNO(-__ret);           \
> +		__ret = -1L;                 \
> +	}                                    \
> +	__ret;                               \
> +})

Except that this now breaks brk(), mmap() and sbrk() by taking any value
with MSB set as an error. Also you've re-introduced the problem you've
faced with const. See my simplification in the other thread by using "?:"
which does avoids any assignment.

Let's just roll brk(), mmap() and sbrk() to their original, working,
definition:

 static __attribute__((unused))
 void *mmap(void *addr, size_t length, int prot, int flags, int fd, off_t offset)
 {
        void *ret = sys_mmap(addr, length, prot, flags, fd, offset);
 
        if ((unsigned long)ret >= -MAX_ERRNO) {
                SET_ERRNO(-(long)ret);
                ret = MAP_FAILED;
        }
        return ret;
 }

And we're done, you can then keep the simplified __sysret() macro for all
other call places.

Willy
  

Hi, Willy

> On Sat, Aug 12, 2023 at 05:51:53AM +0800, Zhangjin Wu wrote:
> > As reported and suggested by Willy, the inline __sysret() helper
> > introduces three types of conversions and increases the size:
> > 
> > (1) the "unsigned long" argument to __sysret() forces a sign extension
> > from all sys_* functions that used to return 'int'
> > 
> > (2) the comparison with the error range now has to be performed on a
> > 'unsigned long' instead of an 'int'
> > 
> > (3) the return value from __sysret() is a 'long' (note, a signed long)
> > which then has to be turned back to an 'int' before being returned by the
> > caller to satisfy the caller's prototype.
> > 
> > To fix up this, firstly, let's use macro instead of inline function to
> > preserves the input type and avoids these useless conversions (1), (3).
> > 
> > Secondly, since all of the sys_* functions have been converted to return
> > integer, now, it is able to remove comparison to a 'unsigned long'
> > -MAX_ERRNO (2) and restore the simple sign comparison as before.
> > 
> (...)
> > +/* Syscall return helper, set errno as -ret when ret < 0 */
> > +#define __sysret(arg)                        \
> > +({                                           \
> > +	__typeof__(arg) __ret = (arg);       \
> > +	if (__ret < 0) {                     \
> > +		SET_ERRNO(-__ret);           \
> > +		__ret = -1L;                 \
> > +	}                                    \
> > +	__ret;                               \
> > +})
> 
> Except that this now breaks brk(), mmap() and sbrk() by taking any value
> with MSB set as an error. Also you've re-introduced the problem you've
> faced with const. See my simplification in the other thread by using "?:"
> which does avoids any assignment.
>

Yeah, thanks for your explanation in this reply [1], the 'const' flag
only triggers build error on the second 'assign' (__ret == -1L), the
first 'assign' is a definition, it is not problematic. so, your "?:"
method is a great idea to simply return without the second 'assign'.

> Let's just roll brk(), mmap() and sbrk() to their original, working,
> definition:
> 
>  static __attribute__((unused))
>  void *mmap(void *addr, size_t length, int prot, int flags, int fd, off_t offset)
>  {
>         void *ret = sys_mmap(addr, length, prot, flags, fd, offset);
>  
>         if ((unsigned long)ret >= -MAX_ERRNO) {
>                 SET_ERRNO(-(long)ret);
>                 ret = MAP_FAILED;
>         }
>         return ret;
>  }
>

Agree, only left a suggestion here [2] about whether we can apply the 2nd patch
instead of rolling them back, let's discuss it in [2] thread.

> And we're done, you can then keep the simplified __sysret() macro for all
> other call places.
> 

Now, this issue is near to the end ;-)

Thanks!
Zhangjin
---

[1]: https://lore.kernel.org/lkml/20230813085140.GD8237@1wt.eu/#R
[2]: https://lore.kernel.org/lkml/20230813132620.19411-1-falcon@tinylab.org/

> Willy
  

On Sun, Aug 13, 2023 at 09:39:44PM +0800, Zhangjin Wu wrote:
> > Let's just roll brk(), mmap() and sbrk() to their original, working,
> > definition:
> > 
> >  static __attribute__((unused))
> >  void *mmap(void *addr, size_t length, int prot, int flags, int fd, off_t offset)
> >  {
> >         void *ret = sys_mmap(addr, length, prot, flags, fd, offset);
> >  
> >         if ((unsigned long)ret >= -MAX_ERRNO) {
> >                 SET_ERRNO(-(long)ret);
> >                 ret = MAP_FAILED;
> >         }
> >         return ret;
> >  }
> >
> 
> Agree, only left a suggestion here [2] about whether we can apply the 2nd patch
> instead of rolling them back, let's discuss it in [2] thread.
(...)
> [2]: https://lore.kernel.org/lkml/20230813132620.19411-1-falcon@tinylab.org/

I'm sorry but I can't find this "suggestion" in this yet-another-super-
long-description-of-another-idea-of-redesign. In addition it's extremely
painful to constantly have to go through web links to follow a single
conversation. Mail works in threads for a reason. When the same discussion
is handled in many parallel threads it becomes impossible to keep it
focused on a specific topic. This is also why you should stop systematically
responding to a message with yet another redesign suggestion, this is super
hard to follow and it literally takes me several hours a week! And at the
end we've not addressed the initial problem but discussed plenty of other
things.

Thanks,
Willy
  

On Sun, Aug 13, 2023 at 09:39:44PM +0800, Zhangjin Wu wrote:
> > And we're done, you can then keep the simplified __sysret() macro for all
> > other call places.
> > 
> 
> Now, this issue is near to the end ;-)

I've now pushed the simplified fix (without changing the SET_ERRNO()
macro, enough last minute breaking changes for now) in branch
20230815-for-6.6-2.

The tests pass and riscv/loongarch are even very slightly smaller than
before (~8 bytes) but again that doesn't count as it depends on how the
compiler decides to arrange if/else branches.

I'll let Shuah know about these late fixes.

Regards,
Willy