serial: 8250: Move alpha-specific quirk out of the core

struct uart_8250_port contains mcr_mask and mcr_force members whose
sole purpose is to work around an alpha-specific quirk.  This code
doesn't belong in the core where it is executed by everyone else,
so move it to a proper ->set_mctrl callback which is used on alpha only.

The alpha-specific quirk was introduced in January 1995:
https://git.kernel.org/pub/scm/linux/kernel/git/history/history.git/diff/drivers/char/serial.c?h=1.1.83

The members in struct uart_8250_port were added in 2002:
https://git.kernel.org/history/history/c/4524aad27854

The quirk applies to non-PCI alphas and arch/alpha/Kconfig specifies
"select FORCE_PCI if !ALPHA_JENSEN".  So apparently the only affected
machine is the EISA-based Jensen that Linus was working on back then:
https://lore.kernel.org/all/CAHk-=wj1JWZ3sCrGz16nxEj7=0O+srMg6Ah3iPTDXSPKEws_SA@mail.gmail.com/

Up until now the quirk is not applied unless CONFIG_PCI is disabled.
If users forget to do that, the serial ports aren't usable on Jensen
and the machine may not boot in the first place.  Avoid by confining
the quirk to CONFIG_ALPHA_JENSEN instead.

Signed-off-by: Lukas Wunner <lukas@wunner.de>
Cc: Russell King <rmk+kernel@armlinux.org.uk>
Cc: Ulrich Teichert <krypton@ulrich-teichert.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
---
Compile-tested only.

 drivers/tty/serial/8250/8250.h       | 11 +----------
 drivers/tty/serial/8250/8250_alpha.c | 15 +++++++++++++++
 drivers/tty/serial/8250/8250_core.c  |  8 +++-----
 drivers/tty/serial/8250/8250_port.c  |  2 +-
 drivers/tty/serial/8250/Makefile     |  1 +
 include/linux/serial_8250.h          |  2 --
 6 files changed, 21 insertions(+), 18 deletions(-)
 create mode 100644 drivers/tty/serial/8250/8250_alpha.c

On Thu, Dec 16, 2021 at 09:36:20PM +0100, Ulrich Teichert wrote:
> > struct uart_8250_port contains mcr_mask and mcr_force members whose
> > sole purpose is to work around an alpha-specific quirk.  This code
> > doesn't belong in the core where it is executed by everyone else,
> > so move it to a proper ->set_mctrl callback which is used on alpha only.
[...]
> > The quirk applies to non-PCI alphas and arch/alpha/Kconfig specifies
> > "select FORCE_PCI if !ALPHA_JENSEN".  So apparently the only affected
> > machine is the EISA-based Jensen that Linus was working on back then:
[...]
> > Up until now the quirk is not applied unless CONFIG_PCI is disabled.
> > If users forget to do that, the serial ports aren't usable on Jensen
> > and the machine may not boot in the first place.  Avoid by confining
> > the quirk to CONFIG_ALPHA_JENSEN instead.
> 
> Wouldn't that mean that you can't use a generic Alpha kernel on the Jensen
> anymore? CONFIG_ALPHA_JENSEN is only set if you specifically select the
> Jensen as target, not when you build a generic kernel. That would be a step
> back in my opinion, as the Debian generic kernel from debian-ports did
> as least boot up on real hardware and the serial console worked just fine

The generic Alpha kernel has CONFIG_PCI=y, so the quirk is not applied,
both with and without the present patch.

You should be able to trigger the lockup that the quirk seeks to avoid
by closing the tty of either of the serial ports.  E.g., if you're using
the serial console, "cat" something to the other serial port's tty.
That will clear TIOCM_OUT2 in serial8250_do_shutdown() and should thus
provoke the lockup.  Alternatively, compile and run the little program
below on the Jensen.  (Pass a serial port tty as argument.)

Should you not be able to reproduce the lockup, then the quirk wouldn't
be necessary anymore and could be removed.

Thanks!

Lukas

-- >8 --

#include <unistd.h>
#include <termios.h>
#include <fcntl.h>
#include <sys/ioctl.h>

#define TIOCM_OUT1	0x2000
#define TIOCM_OUT2	0x4000

int main(int argc, char* argv[]) {
	int fd, ret, flags;

	if (argc < 2)
		return 1;

	fd = open(argv[1], O_RDWR);
	if (fd < 0)
		return 2;

	ret = ioctl(fd, TIOCMGET, &flags);
	if (ret < 0)
		goto close;

	flags &= ~(TIOCM_OUT1 | TIOCM_OUT2);
	ret = ioctl(fd, TIOCMSET, &flags);

close:
	close(fd);

	return ret;
}

On Sun, Dec 19, 2021 at 06:37:56PM +0100, Ulrich Teichert wrote:
> Thanks for the explanation and the reproducer program. I tried yesterday
> to install a recent Linux-Alpha distribution on any of my Alphas to
> be able to build natively and check my self-build kernels before burning
> a CD-ROM, but I failed. Neither gentoo nor Debian nor T2 was installable on
> my Miatas (the Personal Workstations) or my Avanti (AlphaStation 400),
> either failing to boot or recognizing the disks or the CD-ROM, or switching
> to a framebuffer which did not work or just dying on the hardware detection.
> That was a bit frustrating, but I have to follow plan B now and do cross-builds
> from amd_x64 to Alpha - that's easy for the kernel (infact, that's what
> I've been doing from the start), but generating something bootable that
> way is a challenge and needs more time than I usually have during the week,
> but I will tackle it over my xmas vacation,

No worries, whenever you get to spend more time with the Jensen and come
across issues with the serial ports, we can look into it.

I've just sent out an updated version of the patch which also applies
the Jensen quirk if a generic kernel is used.  It auto-detects at
runtime whether the kernel is running on a Jensen.  If on the other hand
the kernel is compiled specifically for the Jensen, the quirk is applied
unconditionally.

I've also realized that a code comment in sunsu.c refers to the Alpha
quirk, so I'm updating it to point to the new location.  It looks like
a JavaStation thin client of the same era had a similar hardware erratum
as the Jensen.

Thanks,

Lukas