Re: [PATCH v5 2/2] target/xtensa: Use semihosting/syscalls.h

["Alex Bennée" <alex.bennee@linaro.org>]

Richard Henderson <richard.henderson@linaro.org> writes:

> On 6/28/22 19:08, Max Filippov wrote:
>> On Tue, Jun 28, 2022 at 4:43 AM Richard Henderson
>> <richard.henderson@linaro.org> wrote:
>>>
>>> This separates guest file descriptors from host file descriptors,
>>> and utilizes shared infrastructure for integration with gdbstub.
>>> Remove the xtensa custom console handing and rely on the
>>> generic -semihosting-config handling of chardevs.
>>>
>>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>>> ---
>>>   target/xtensa/cpu.h         |   1 -
>>>   hw/xtensa/sim.c             |   3 -
>>>   target/xtensa/xtensa-semi.c | 226 ++++++++----------------------------
>>>   3 files changed, 50 insertions(+), 180 deletions(-)
>>>
>>> diff --git a/target/xtensa/cpu.h b/target/xtensa/cpu.h
>>> index ea66895e7f..99ac3efd71 100644
>>> --- a/target/xtensa/cpu.h
>>> +++ b/target/xtensa/cpu.h
>>> @@ -612,7 +612,6 @@ void xtensa_translate_init(void);
>>>   void **xtensa_get_regfile_by_name(const char *name, int entries, int bits);
>>>   void xtensa_breakpoint_handler(CPUState *cs);
>>>   void xtensa_register_core(XtensaConfigList *node);
>>> -void xtensa_sim_open_console(Chardev *chr);
>>>   void check_interrupts(CPUXtensaState *s);
>>>   void xtensa_irq_init(CPUXtensaState *env);
>>>   qemu_irq *xtensa_get_extints(CPUXtensaState *env);
>>> diff --git a/hw/xtensa/sim.c b/hw/xtensa/sim.c
>>> index 946c71cb5b..5cca6a170e 100644
>>> --- a/hw/xtensa/sim.c
>>> +++ b/hw/xtensa/sim.c
>>> @@ -87,9 +87,6 @@ XtensaCPU *xtensa_sim_common_init(MachineState *machine)
>>>           xtensa_create_memory_regions(&sysram, "xtensa.sysram",
>>>                                        get_system_memory());
>>>       }
>>> -    if (serial_hd(0)) {
>>> -        xtensa_sim_open_console(serial_hd(0));
>>> -    }
>> I've noticed that with this change '-serial stdio' and its variants
>> are still
>> accepted in the command line, but now they do nothing.
>
> Pardon?  They certainly will do something, via writes to the serial hardware.
>
>
>> This quiet
>> change of behavior is unfortunate. I wonder if it would be acceptable
>> to map the '-serial stdio' option in the presence of '-semihosting' to
>> something like '-chardev stdio,id=id1 -semihosting-config chardev=id1'?
>
> I dunno.  I'm wary of having xtensa be unique here.  Alex, thoughts?

Is semihosting *the* serial hardware for xtensa-sim or is it overriding
another serial interface? I'm wary of adding more magical behaviour for
-serial as it can be confusing enough already what actually gets routed
to it if not doing everything explicitly.

>
>>> +                if (get_user_u32(tv_sec, regs[5]) ||
>>> +                    get_user_u32(tv_usec, regs[5])) {
>> get_user_u32(tv_usec, regs[5] + 4)?
>
> Oops, yes.
>
>>> -                regs[2] = select(fd + 1,
>>> -                                 rq == SELECT_ONE_READ   ? &fdset : NULL,
>>> -                                 rq == SELECT_ONE_WRITE  ? &fdset : NULL,
>>> -                                 rq == SELECT_ONE_EXCEPT ? &fdset : NULL,
>>> -                                 target_tv ? &tv : NULL);
>>> -                regs[3] = errno_h2g(errno);
>>> +                /* Poll timeout is in milliseconds; overflow to infinity. */
>>> +                msec = tv_sec * 1000ull + DIV_ROUND_UP(tv_usec, 1000ull);
>>> +                timeout = msec <= INT32_MAX ? msec : -1;
>>> +            } else {
>>> +                timeout = -1;
>>>               }
>>> +
>>> +            switch (regs[4]) {
>>> +            case SELECT_ONE_READ:
>>> +                events = G_IO_IN;
>>> +                break;
>>> +            case SELECT_ONE_WRITE:
>>> +                events = G_IO_OUT;
>>> +                break;
>>> +            case SELECT_ONE_EXCEPT:
>>> +                events = G_IO_PRI;
>>> +                break;
>>> +            default:
>>> +                xtensa_cb(cs, -1, EINVAL);
>> This doesn't match what there used to be: it was possible to call
>> select_one with rq other than SELECT_ONE_* and that would've
>> passed NULL for all fd sets in the select invocation turning it into
>> a sleep. It would return 0 after the timeout.
>
> Hmm.  Is there any documentation of what it was *supposed* to do?
> Passing rq == 0xdeadbeef and expecting a specific behaviour seems odd.
>
>
> r~


-- 
Alex Bennée