Re: [PATCH v15 8/9] target/loongarch: Adjust functions and structure to support user-mode

[Richard Henderson]

On 6/9/22 23:53, gaosong wrote:
> Hi Richard,
>
> On 2022/6/10 上午2:42, Richard Henderson wrote:
> > >   void helper_asrtle_d(CPULoongArchState *env, target_ulong rj, target_ulong rk)
> > >   {
> > >       if (rj > rk) {
> > > +#ifdef CONFIG_USER_ONLY
> > > +        cpu_loop_exit_sigsegv(env_cpu(env), GETPC(),
> > > +                              MMU_DATA_LOAD, true, GETPC());
> > > +#else
> > >           do_raise_exception(env, EXCCODE_ADEM, GETPC());
> > > +#endif
> >
> > This change is wrong.  First, the kernel's do_ade raises SIGBUS. Second, GETPC() is a 
> > host address, not a guest address.  Third, this highlights the fact that the existing 
> > system code is wrong, and should be setting badvaddr.
> >
> > You need to
> > (1) set badvaddr here, and then
> > (2) handle EXCCODE_ADEM in linux-user/loongarch/cpu_loop.c to 
> > force_fix_fault(TARGET_SIGBUS, TARGET_BUS_ADRERR, env->badvaddr). 
>
> badvaddr is env->pc or base->pc_next?

I don't know.  The documentation for the ASRT{LE,GD}.D instructions is 
incomplete.

However, from the kernel code,

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/loongarch/kernel/traps.c#n360

I can see that the kernel expects *some* value to be set there.  Given that this is the 
same trap used by the bound check memory accesses, I presume that badvaddr is related to 
the memory access not the PC.  So I would expect badvaddr to be RJ.

But that is a guess.  Please check with your hardware engineers.


r~