Re: [RFC PATCH v2 0/3] Prefer working VT console over SPCR and device-tree chosen stdout-path
From: Alper Nebi Yasak
Date: Fri May 15 2020 - 15:27:15 EST
On 13/05/2020 17:37, Petr Mladek wrote:
> On Thu 2020-04-30 19:14:34, Alper Nebi Yasak wrote:
>> | "console=tty0" | (no console arg) |
>> ------------------+-----------------------+-----------------------+
>> QEMU VM | tty0 -WU (EC p ) | ttyAMA0 -W- (EC a) |
>> (w/ SPCR) | ttyAMA0 -W- (E a) |
>> |
>
> The SPCR handling is inconsistent over architectures, see
> https://lkml.kernel.org/r/20180830123849.26163-1-prarit@xxxxxxxxxx
>
> IMHO, arm developers decided that consoles defined by SPCR are always
> enabled when existing.
I'm OK with those being enabled. Though, I hope "not registering tty0"
wasn't an explicit decision, but maybe an oversight/trade-off due to
assuming SPCR code will only run on servers without displays (where tty0
wouldn't matter). (I understand it might be too late to change that.)
So I'd want the 2nd column to be: tty0(EC) ttyAMA0(E) at best, and
ttyAMA0(EC) tty0(E) at worst.
> In 1st column: tty0 is the preferred console because it is defined
> on the commandline.
>
> In 2nd column: tty0 is not enabled at all because another console was
> defined by SPCR. Note that ttySX and ttyX consoles are registered only
> as a fallback when there is no other console defined.
>
> The following code is responsible for the fallback, see register_console()
>
> /*
> * See if we want to use this console driver. If we
> * didn't select a console we take the first one
> * that registers here.
> */
> if (!has_preferred) {
> if (newcon->index < 0)
> newcon->index = 0;
> if (newcon->setup == NULL ||
> newcon->setup(newcon, NULL) == 0) {
> newcon->flags |= CON_ENABLED;
> if (newcon->device) {
> newcon->flags |= CON_CONSDEV;
> has_preferred = true;
> }
> }
> }
>
>
>> ------------------+-----------------------+-----------------------+
>> Chromebook Plus | tty0 -WU (EC p ) | ttyS2 -W- (EC p a) |
>> (w/ stdout-path) | | tty0 -WU (E ) |
>
> Hmm, of_console_check() explicitly ignores the console defined by
> stdout-path when there is a console on the commandline. This explains
> 1st column.
>
> I am not sure about 2nd column. My guess is that ttyX consoles are
> tried first. tty0 is registered as a fallback because there is no
> other console at the moment. ttyS2 is tried later and it is
> registered because it is in stdout-patch and there is no console
> in the command line. It is somehow consistent with CONFIG_VT_CONSOLE
> description.
>
> Sadly, it is different logic than with SPCR :-(
I like the fact that this one has tty0. For example, Debian's installer
iterates over /proc/consoles and launches itself on all the consoles it
finds there, so it wouldn't launch on my chromebook's screen if tty0
wasn't included (just like it doesn't launch on a QEMU aarch64 VM's
framebuffer).
>> ------------------+-----------------------+-----------------------+
>> Chromebook Plus | tty0 -WU (EC p ) | tty0 -WU (EC p ) |
>> (w/o either) | | |
>> ------------------+-----------------------+-----------------------+
>
> This variant is easy and everyone would probably expect this.
I think things run roughly in the following order (from what I can
decipher from kernel messages) and I think it matches your explanations:
| ACPI SPCR | dt chosen stdout-path |
+=================================+=================================+
| acpi_parse_spcr() | |
| -> add_preferred_console(uart0) | |
| (if not on x86) | |
+---------------------------------+---------------------------------+
| console_setup() |
| -> add_preferred_console(tty0) |
| (if console=tty0) |
+---------------------------------+---------------------------------+
| register_console(vt) |
+---------------------------------+---------------------------------+
| | of_console_check() |
| | -> add_preferred_console(uart2) |
| | (if no console arg) |
+---------------------------------+---------------------------------+
| register_console(serial) |
+---------------------------------+---------------------------------+
> Regarding the description of CONFIG_VT_CONSOLE option. I am afraid
> that it was created and true only before SPCR and device tree support
> was introduced.
OK. Assuming these changes won't go any further, maybe I'll try
documenting the current behavior in relevant places.
> Now, it is really sad that SPCR and device tree have different
> behavior even across architectures. But I am afraid that we could
> not change it without breaking many setups.
>
> The only common rules are:
>
> + The last console on the command line should always be the
> preferred one when defined.
>
> + Consoles defined by the device (SPCR, device tree) are used
> when there is no commandline.
>
> + ttyX or ttySX are used as a fallback when nothing else is defined.
>
>
> My suggestion is:
>
> + Fix SPCR setting or device tree of your device when the defaults
> are not as expected.
Maybe I can get QEMU's SPCR use conditional on the existence a
framebuffer, and get distributions to remove stdout-path from certain
device-trees; but that would disable the serial console completely
(instead of having it enabled where tty0 is still preferred).
> + Use command line to force your value when the defaults are not
> as expected and you could not change them.
This works; but I'd have to know the machine's serial configuration in
advance to put it in the cmdline as "console=<serial> console=tty0", or
lose the serial console as in the above. (A "console=dt" like that
"console=spcr" patch you linked to would be useful here if it existed.)
Both seem imperfect in that sense, but tolerable.
> I am afraid that we could not fix your problem on the kernel side. It
> would broke other setups that depend on the existing behavior.
>
> Best Regards,
> Petr
Thanks for the detailed reply.