Re: [PATCH] tools/nolibc: add support for Alpha

[Willy Tarreau <w@1wt.eu>]

Hi,

On Mon, Jul 14, 2025 at 07:21:38AM +0200, Thomas Weißschuh wrote:
> Hi Richard,
> 
> On 2025-07-13 16:21:58-0600, Richard Henderson wrote:
> > On 7/13/25 14:08, Thomas Weißschuh wrote:
> > > +++ b/tools/testing/selftests/nolibc/nolibc-test.c
> > > @@ -709,6 +709,10 @@ int run_startup(int min, int max)
> > >   	/* checking NULL for argv/argv0, environ and _auxv is not enough, let's compare with sbrk(0) or &end */
> > >   	extern char end;
> > >   	char *brk = sbrk(0) != (void *)-1 ? sbrk(0) : &end;
> > > +#if defined(__alpha__)
> > > +	/* the ordering above does not work on an alpha kernel */
> > > +	brk = NULL;
> > > +#endif
> > 
> > The syscall api is different for brk on alpha.
> > A change to sys_brk or brk in include/nolibc/sys.h is required.
> 
> You are referring to osf_brk, right?
> I think that should work as-is with the current wrappers.

I finally managed to reinstall my DS10 to build and test this and FWIW
the test passes:

  $ ./nolibc-test 
  Running test 'startup'
  0 argc = 1                                                        [OK]
  1 argv_addr = <0x11fc7b428>                                       [OK]
  2 argv_environ = <0x11fc7b428>                                    [OK]
  3 argv_total = 1                                                  [OK]
  4 argv0_addr = <0x11fc7b665>                                      [OK]
  5 argv0_str = <./nolibc-test>                                     [OK]
  6 argv0_len = 13                                                  [OK]
  7 environ_addr = <0x11fc7b438>                                    [OK]
  8 environ_envp = <0x11fc7b438>                                    [OK]
  9 environ_auxv = <0x11fc7b438>                                    [OK]
  10 environ_total = 175                                            [OK]
  11 environ_HOME = <0x11fc7b6f4>                                   [OK]
  12 auxv_addr = <0x11fc7b4e8>                                      [OK]
  13 auxv_AT_UID = 509                                              [OK]
  14 constructor = 3                                                [OK]
  15 linkage_errno = <0x1200200f8>                                  [OK]
  16 linkage_constr = 3                                             [OK]
  Errors during this test: 0
  
  Running test 'syscall'
  0 access = 0                                                      [OK]
  1 access_bad = -1 EPERM                                           [OK]
  2 clock_getres = 0                                                [OK]
  3 clock_gettime = 0                                               [OK]
  4 clock_settime = -1 EINVAL                                       [OK]
  5 getpid = 9201                                                   [OK]
  6 getppid = 419                                                   [OK]
  7 gettid = 9201                                                   [OK]
  8 getpgid_self = 9201                                             [OK]
  9 getpgid_bad = -1 ESRCH                                          [OK]
  10 kill_0 = 0                                                     [OK]
  11 kill_CONT = 0                                                  [OK]
  12 kill_BADPID = -1 ESRCH                                         [OK]
  13 sbrk_0 = <0x120024000>                                         [OK]
  14 sbrk = 0                                                       [OK]
  15 brk = 0                                                        [OK]
  (...)
  Total number of errors: 0
  Exiting with status 0

The result is exactly the same if I comment that line that resets brk,
as brk was apparently already NULL:

  13 sbrk_0 = <0x120024000>                                         [OK]
  14 sbrk = 0                                                       [OK]
  15 brk = 0                                                        [OK]

> On alpha, mm->brk and mm->arg_start are ordered differently from other
> architectures. Personally I think the nolibc tests are a bit bogus here.

I seem to remember that these are among the older minimal consistency
tests and that it could be time to revisit this :-/

Willy