Date: Tue, 10 Sep 2002 18:40:02 +0200
From: Vojtech Pavlik <vojtech@suse.cz>
To: "Adam J. Richter" <adam@yggdrasil.com>
Cc: Vojtech Pavlik <vojtech@suse.cz>, linux-kernel@vger.kernel.org
Subject: Re: Patch?: linux-2.5.33/drivers/input/keyboard/atkbd.c allow SETLEDS to fail
Message-ID: <20020910184002.A9919@ucw.cz>
References: <200209091504.IAA01726@baldur.yggdrasil.com> <20020909172004.A2631@ucw.cz> <20020910090520.A731@adam.yggdrasil.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
User-Agent: Mutt/1.2.5i
In-Reply-To: <20020910090520.A731@adam.yggdrasil.com>; from adam@yggdrasil.com on Tue, Sep 10, 2002 at 09:05:20AM -0700
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 11992
Lines: 268

On Tue, Sep 10, 2002 at 09:05:20AM -0700, Adam J. Richter wrote:
> On Mon, Sep 09, 2002 at 05:20:04PM +0200, Vojtech Pavlik wrote:
> > There are not. If SETLEDS fails, then there is the GETID command, which
> > is also allowed to fail (fails on AT (non-PS/2) keyboards, which are
> > still common), and it most likely also fails on your keyboard. That's
> > all.
> > 
> > If you kill the SETLEDS command, then even without a keyboard, the
> > keyboard will be detected, and sent commands, which will have to time
> > out every time.
> 
> 	100 milliseconds, which normally only happens when the user
> hits CapsLock, NumLock or ScrollLock, which generally implies that a
> keyboard is present.

100 milliseconds is a long time. When a keyboard is present, it replies
much faster than that, usually around 1-4 ms.

> I suspect that you could change it to use mdelay
> instead of udelay, so that other processes could do useful work during
> the wait,

Actually, mdelay just calls udelay several times, processes are stopped.
I could do a sleep, but that'd imply quite complex code involving a
kernel thread most likely - keypresses come from interrupt context.

> which might be relevant even with a keyboard plugged in that
> just happens to be slow if one is running a screen saver that cycles
> the keyboard LED's.  If it's important, you could also set a flag
> when SETLEDS fails and not attempt further SETLEDS until another
> sign of life from the keyboard is received (which would also address
> the case of someone unplugging the keyboard after initialization).

We do this for USB keyboards, it might make sense for AT keyboards as
well, but only in case the LED thread is implemented.

> 	I also noticed a problem with plugging in a PS/2 keyboard
> after initialization where it would not work, at least when I also had
> the USB-to-PS/2 emulation enabled (if I booted with PS/2 keyboard
> plugged in, both the USB and the PS/2 keyboards would work).  I have
> not checked the purely non-USB case for whether your code works for
> loading the module without the keyboard present and then plugging the
> keyboard in later.

Autoloading the modules doesn't work yet, but if the modules are loaded
(or better compiled in, as keyboard support used to be), then a keyboard
is autodetected after plugging it in, even if you connect it to the
mouse port.

> 	Without the USB-to-PS/2 emulation on my VD133 motherboard, I
> can't type into lilo. 

Then leave it enabled, it's not a problem.

> Perhaps newer BIOSes have USB keyboard input
> without need for hardware emulation, but I imagine that there are
> quite a few computers with this problem.

Actually the USB code in the Linux kernel does a thing called "USB
emulation takeover", where it takes the USB controller over from BIOS
and disables the PS/2 emulation. This way you can have your keyboard
both in LILO and in Linux using a proper USB driver.

> That would be OK if there
> were some way for Linux to turn off USB-to-PS/2 keyboard emulation
> once the Linux USB code is loaded. 

This is what happens.

> Currently, I get an infinite loop
> of carriage returns when I do "modprobe hid" if I am using USB-to-PS/2
> keyboard emulation.

This is interesting. I never had a report like that. Care to supply more
detail? Also this would mean you have the irqs assigned correctly now?

> 	I agree that USB-to-PS/2 seems to be an ugly kludge, but I
> think that the small cost due to orphaning certain oddball hardware
> combinations is greater than the even smaller benefit of faster
> SETLEDs (which I think could be achieved in other ways).

I'll be adding a switch to not communicate with the keyboard and only
accept keystrokes. This will solve your problem. But I still think it's
better to fix the other problems you have and have the hardware operate
in the intended way.

> 	The rest of this email is just about the PCI interrupt routing.
> 
> > > >I assume you have an USB keyboard, and the BIOS emulates a PS/2 keyboard
> > > >for non-USB capable OSes.
> > > 
> > > >Well, fix Linux irq router code, instead of damaging the keyboard code.
> > > 
> > > 	How is the Linux IRQ routing code broken?  As far as I know,
> > > it only reads the information given to it and cannot reprogram the
> > > hardware.
> > 
> > It can.
> 
> 	Thanks for the pointers to arch/i386/pci/irq.c and
> drivers/pci/quirks.c.  This code programs interrupt routing in a few
> special cases and arch/i386/pci/irq.c has somewhat general underlying
> facilities that could be put to more general use, but I do not see any
> existing programming interface by which a user level program could
> somehow get the kernel to try, for example, to reprogram pci device
> 00:07.2 to IRQ 10.

It's done at boot time, so it's not designed to be programmed by
userspace. However, the usb driver does pci_enable_device(), where the
PCI code will assign an IRQ to it if the BIOS didn't.

> > > 	My understanding is that the BIOS is responsible for setting
> > > up the IRQ routes of each device and writing them into each device's
> > > PCI configuration register dword 15 byte 0 (INTERRUPT_LINE), which is
> > > normally just 8 bits of RAM.
> > 
> > Yes, but because the BIOS often writes nonsense to these config
> > registers, Linux knows several southbridges and can fix the routing by
> > hand.
> > 
> > > The mechanism for *making* those routes
> > > is vendor specific and handled by the BIOS at boot time.  In the
> > > particular case of my VD133, it apparently uses a Via VT82C598 PCI
> > > bridge ("Apollo MVP3 AGP"), which I think is the component that would
> > > have the secret bits for setting up routing from the main PCI bus to
> > > the interrupt controllers.
> > 
> > Most likely the interrupt routing is controlled by the southbridge (the
> > ISA bridge device). See arch/i386/pci/irq.c. Maybe it's as easy as an
> > entry for your southbridge is missing.
> > 
> > > And those bits do appear to be secret.
> > 
> > Not really. Most of them are actually documented. Namely for built-in
> > devices like USB controllers, because those cannot be wired differently
> > by different board manufacturers.
> 
> 	There appears to be more black magic involved than is covered
> by drivers/pci/quirks.c and arch/i386/pci/irq.c.  Thanks for the
> pointers.  From the files that you mentioned and a lot of
> experimentation of setpci and a minor kernel change, I have determined
> that when I have my natsemi ethernet card plugged in, then, and *only
> then*, I can get the USB controller to use the same IRQ as the
> ethernet card, and, with the attached modification to
> arch/i386/pci/irq.c, I can get the kernel to use that IRQ even though
> there is no pirq entry for it.  By the way, this isn't just the USB
> controller being polled whenever an ethernet packet arrives.  I've
> verified that it works with the ethernet cable unplugged, as long as
> the card is present.  However, without the ethernet card installed,
> the only way I have gotten the USB keyboard to work is with the
> USB-to-PS/2 emulation.  Right now, I have the ethernet card installed
> and am composing this email on that USB keyboard via the Linux USB
> drivers.  Thanks again particularly for the pointer to
> arch/i386/pci/irq.c.

Glad to hear you got it to work. More comments below.

> 	Here are the details of what I found, as the information may
> be useful for other purposes in the future.
> 
> 	pirq_via_{get,set} in arch/i386/pci/irq.c show a general
> mechanism for assigning IRQ's though the Via south bridge chip.
> dev_perq_info->irq[pin].link is an index into a table of 4 bit
> nibbles starting at PCI configuration register 0x55 of the south
> bridge.  You write the IRQ that you want to route to into the
> appropriate nibble.  Here is an example from my system, with some
> comments in brackets:
> 
> 	setpci -v -s 00:07.0 55 56 57 58
> 	00:07.0:55 = c0					[ pirq#1 --> IRQ 12 ]
> 	00:07.0:56 = 0b					[ pirq#2 --> IRQ 11 ]
> 	00:07.0:57 = 00
> 	00:07.0:58 = 00
> 
> 	I can change the ethernet's IRQ routing from 11 to, say, 9,
> and watch the ethernet immediately stop, by doing something like
> "setpci -s 00:07.0 56=09".
> 
> 	However, the built-in USB on the VT82C586B does not seem to use
> this mechanism.  There is no PIRQ entry for the pin that the USB controller
> uses, and I cannot, for example, get the USB interrupts to show up on IRQ
> ten by filling all of the southbridge configuration registers from 0x55
> through 0x58 with 0xaa.
> 
> 	It seems that the built-in USB on the VT82C586B has the same
> feature as is described for the Via 686A/B in quirk_via_irqpic in
> drivers/pci/quirks.c:
> 
>  * Via 686A/B:  The PCI_INTERRUPT_LINE register for the on-chip
>  * devices, USB0/1, AC97, MC97, and ACPI, has an unusual feature:
>  * when written, it makes an internal connection to the PIC.

Indeed, the 586b specification says that writing ofset 0x3c of the USB
controller with the irq value causes USB irq to be routed to the irq
written.

> 	This seems to work on my VD133 motherboard also, but it seems
> only to work if I route it to the interrupt being used by the ethernet
> card.

It most likely will work when shared with any PCI interrupt.

> It seems that there is some additional initialization needed

Couldn't find anything like that in the docs. Maybe the other irqs are
used for something else, non-shareable. Maybe you can set them as ISA in
BIOS setup and then the BIOS won't assign anything to them.

> (and yes, I've checked that my choice of IRQ conforms that various IRQ
> masks in arch/i386/pci/irq.c, and I've tried setting the south
> bridge's configuration registers at 0x55 in conjunction with this.
> So, it appears that there is some further interrupt routing
> preparation that the kernel does not know about.

The 586b register description doesn't mention anything else.

> > Try the southbridge instead of the northbridge. I have several of the
> > documents, and can send it to you if you tell me the right chip number.
> 
> 	Great!  Here is output of lspci on my system.  I would be
> interested in documents for any of the relevant chips.  Although I've
> kludged around the problem enough for my own use, I'd prefer to
> develop a more general and correct fix for other VD133 users at least
> and perhaps other boards if they have a similar BIOS problem.
> 
> 00:00.0 Host bridge: VIA Technologies, Inc. VT82C691 [Apollo PRO] (rev 44)
> 00:01.0 PCI bridge: VIA Technologies, Inc. VT82C598 [Apollo MVP3 AGP]
> 00:07.0 ISA bridge: VIA Technologies, Inc. VT82C596 ISA [Apollo PRO] (rev 12)
> 00:07.1 IDE interface: VIA Technologies, Inc. VT82C586 IDE [Apollo] (rev 06)
> 00:07.2 USB Controller: VIA Technologies, Inc. VT82C586B USB (rev 08)
> 00:07.3 Host bridge: VIA Technologies, Inc.: Unknown device 3050 (rev 20)
> 00:10.0 Ethernet controller: National Semiconductor Corporation DP83815 10/100Mbps Ethernet with Wake on LAN
> 01:00.0 VGA compatible controller: ATI Technologies Inc 3D Rage Pro AGP AGP 1X/2X (rev 5c)
> 
> 
> 	Thanks for your help.

586b. Quite old. I have that document. You can get it from
http://twilight.ucw.cz/vt82c586b.pdf

> --- linux-2.5.33/arch/i386/pci/irq.c	2002-08-31 15:05:37.000000000 -0700
> +++ linux/arch/i386/pci/irq.c	2002-09-10 05:29:09.000000000 -0700
> @@ -591,8 +591,13 @@
>  	pirq = info->irq[pin].link;
>  	mask = info->irq[pin].bitmap;
>  	if (!pirq) {
> -		DBG(" -> not routed\n");
> -		return 0;
> +		pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &dev->irq);
> +		if (dev->irq)
> +			return 1;
> +		else {
> +			DBG(" -> not routed\n");
> +			return 0;
> +		}
>  	}
>  	DBG(" -> PIRQ %02x, mask %04x, excl %04x", pirq, mask, pirq_table->exclusive_irqs);
>  	mask &= pcibios_irq_mask;


-- 
Vojtech Pavlik
SuSE Labs
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/