RE: [PATCH v3] RISC-V: Enable static-pie.

["Wang, Yanzhang" <yanzhang.wang@intel.com>]

Hi Andreas,

Sorry for the late response. Thanks your comments! And I don't quite
follow your comments. Could you please give some further information?

> (gdb) ni
> _start () at ../sysdeps/riscv/start.S:74
> 74              lla   gp, __global_pointer$
> 1: x/i $pc
> => 0x3ff7f59ec6 <load_gp>:      auipc   gp,0xa0
> (gdb)
> 0x0000003ff7f59eca      74              lla   gp, __global_pointer$
> 1: x/i $pc
> => 0x3ff7f59eca <_start+38>:    add     gp,gp,706
> (gdb)
> load_gp () at ../sysdeps/riscv/start.S:76
> 76              ret
> 1: x/i $pc
> => 0x3ff7f59ece <load_gp+8>:    ret
> (gdb)
> _start () at ../sysdeps/riscv/start.S:51
> 51              mv    a5, a0  /* rtld_fini.  */
> 1: x/i $pc
> => 0x3ff7f59ea8 <_start+4>:     mv      a5,a0
> (gdb)
> 53              la   a0, main
> 1: x/i $pc
> => 0x3ff7f59eaa <_start+6>:     auipc   a0,0x9e
> (gdb)
> 0x0000003ff7f59eae      53              la   a0, main
> 1: x/i $pc
> => 0x3ff7f59eae <_start+10>:    ld      a0,-74(a0)
> (gdb)
> 54              REG_L a1, 0(sp)      /* argc.  */
> 1: x/i $pc
> => 0x3ff7f59eb2 <_start+14>:    ld      a1,0(sp)
> (gdb) i reg a0
> a0             0x0      0

The first, it seems you have constructed a test case which should fail
this patch. But based on your gdb comment, I'm not sure which step is
not correct. Should the a0 register be non-zero or the gp register be
wrong? 

---

The code in the if will try to lookup the __global_pointer$ and move to the
gp register. It will be failed because l->l_scope is NULL when static-pie.
So I added this checking to avoid this failure.

> > +  if (l->l_type == lt_executable && l->l_scope != NULL)


> This is not the right way to test for static PIE, it tries to access a
> non-relocated non-zero pointer.  The mold linker, when producing relative
> relocations, puts the addend both in the relocation addend and the
> relocated field, which is perfectly valid to do, so this is non-zero here.

The second, do you mean there's a case that l->l_scope is not NULL when built
with static-pie? Or we should not access the members in link_map ?

Thanks,
Yanzhang

> -----Original Message-----
> From: Andreas Schwab <schwab@suse.de>
> Sent: Monday, January 29, 2024 8:46 PM
> To: Wang, Yanzhang <yanzhang.wang@intel.com>
> Cc: libc-alpha@sourceware.org; adhemerval.zanella@linaro.org
> Subject: Re: [PATCH v3] RISC-V: Enable static-pie.
> 
> On Jan 02 2024, yanzhang.wang@intel.com wrote:
> 
> > diff --git a/sysdeps/riscv/dl-machine.h b/sysdeps/riscv/dl-machine.h
> > index ffb8c4aaa0..0cbb476c05 100644
> > --- a/sysdeps/riscv/dl-machine.h
> > +++ b/sysdeps/riscv/dl-machine.h
> > @@ -348,7 +348,7 @@ elf_machine_runtime_setup (struct link_map *l,
> struct r_scope_elem *scope[],
> >        gotplt[1] = (ElfW(Addr)) l;
> >      }
> >
> > -  if (l->l_type == lt_executable)
> > +  if (l->l_type == lt_executable && l->l_scope != NULL)
> 
> This is not the right way to test for static PIE, it tries to access a
> non-relocated non-zero pointer.  The mold linker, when producing relative
> relocations, puts the addend both in the relocation addend and the
> relocated field, which is perfectly valid to do, so this is non-zero here.
> 
> --
> Andreas Schwab, SUSE Labs, schwab@suse.de GPG Key fingerprint = 0196 BAD8
> 1CE9 1970 F4BE  1748 E4D4 88E3 0EEA B9D7 "And now for something
> completely different."