Re: CVS commit: src/external/gpl3/gcc

[Rin Okuyama <rokuyama.rk%gmail.com@localhost>]

Hi again,

On 2020/09/10 21:53, Kamil Rytarowski wrote:
> Module Name:	src
> Committed By:	kamil
> Date:		Thu Sep 10 12:53:06 UTC 2020
>
> Modified Files:
> 	src/external/gpl3/gcc/dist/libsanitizer/sanitizer_common:
> 	    sanitizer_linux_libcdep.cc
> 	src/external/gpl3/gcc/lib: Makefile.sanitizer
>
> Log Message:
> Avoid using internal RTLD/libpthread/libc symbol in sanitizers
...
> Index: src/external/gpl3/gcc/dist/libsanitizer/sanitizer_common/sanitizer_linux_libcdep.cc
> diff -u src/external/gpl3/gcc/dist/libsanitizer/sanitizer_common/sanitizer_linux_libcdep.cc:1.15 src/external/gpl3/gcc/dist/libsanitizer/sanitizer_common/sanitizer_linux_libcdep.cc:1.16
> --- src/external/gpl3/gcc/dist/libsanitizer/sanitizer_common/sanitizer_linux_libcdep.cc:1.15	Mon Sep  7 07:10:43 2020
> +++ src/external/gpl3/gcc/dist/libsanitizer/sanitizer_common/sanitizer_linux_libcdep.cc	Thu Sep 10 12:53:05 2020
> @@ -47,6 +47,7 @@
>   #if SANITIZER_NETBSD
>   #include <sys/sysctl.h>
>   #include <sys/tls.h>
> +#include <lwp.h>
>   #endif
>   
>   #if SANITIZER_SOLARIS
> @@ -417,13 +418,7 @@ uptr ThreadSelf() {
>   
>   #if SANITIZER_NETBSD
>   static struct tls_tcb * ThreadSelfTlsTcb() {
> -  struct tls_tcb * tcb = NULL;
> -# ifdef __HAVE___LWP_GETTCB_FAST
> -  tcb = (struct tls_tcb *)__lwp_gettcb_fast();
> -# elif defined(__HAVE___LWP_GETPRIVATE_FAST)
> -  tcb = (struct tls_tcb *)__lwp_getprivate_fast();
> -# endif
> -  return tcb;
> +  return (struct tls_tcb *)_lwp_getprivate();
>   }
>   
>   uptr ThreadSelf() {

This change breaks at least mips and powerpc, in which the return value of
__lwp_getprivate(2), i.e., curlwp->l_private is not tcb address itself, but
biased one. On the other hand, the return value of __lwp_gettcb_fast() is
unbiased address; see sys/arch/{mips,powerpc}/include/mcontext.h.

For powerpc, I recently attempted to change l_private to store tcb address
itself:

http://www.nerv.org/netbsd/?q=id:20200621T004000Z.95c1a18070b53713ce2c763df7f40743bf74172c

But I reverted it soon as requested by joerg:

http://www.nerv.org/netbsd/?q=id:20200622T053457Z.05db3be87b5ad499f5d1adba755bc573fd241c87

His reasoning was that kernel must not know the ABI details in userland.
I fully agree with this. See above links for more details.

Thanks,
rin