Re: [PATCH v2 4/4] toshiba_acpi: Add blacklist for models with hotkey problems

[Seth Forshee <seth.forshee@canonical.com>]

On Thu, Jan 05, 2012 at 06:26:27PM +0000, Matthew Garrett wrote:
> On Tue, Jan 03, 2012 at 01:02:38PM -0600, Seth Forshee wrote:
> > Several Satellite models have a buggy implementation of the INFO method
> > that causes ACPI exceptions when executed:
> > 
> >  ACPI Error: Result stack is empty! State=ffff88012d70f800 (20110413/dswstate-98)
> >  ACPI Exception: AE_AML_NO_RETURN_VALUE, Missing or null operand (20110413/dsutils-646)
> >  ACPI Exception: AE_AML_NO_RETURN_VALUE, While creating Arg 0 (20110413/dsutils-763)
> >  ACPI Error: Method parse/execution failed [\_SB_.VALZ.GETE] (Node ffff880131175eb0), AE_AML_NO_RETURN_VALUE (20110413/psparse-536)
> >  ACPI Error: Method parse/execution failed [\_SB_.VALZ.INFO] (Node ffff880131175ed8), AE_AML_NO_RETURN_VALUE (20110413/psparse-536)
> >  toshiba_acpi: ACPI INFO method execution failed
> >  toshiba_acpi: Failed to query hotkey event
> 
> Ugh, in several ways. The hotkeys on these machines are presumably 
> supposed to work - do we have any idea what we should be doing?

Here's a run-down of why this happens. First, the relevant sections of
the DSDT:

    Name (EVCD, Package (0x64) {})
    Name (EVCT, 0x00)

    Method (PUTE, 1, Serialized)
    {
        <snip>

        If (LLess (EVCT, 0x64))
        {
            Store (Arg0, Index (EVCD, EVCT))
            Increment (EVCT)
        }

	<snip>

        Notify (VALZ, 0x80)
        Sleep (0x64)
        Return (0x00)
    }

    Method (GETE, 0, Serialized)
    {
        <snip>

        If (LEqual (EVCT, 0x00))
        {
            Release (MUEV)
            Return (0x00)
        }

        Store (DerefOf (Index (EVCD, 0x00)), Local0)
        Store (0x00, Local1)
        While (LLess (Local1, Subtract (0x64, 0x01)))
        {
            Store (DerefOf (Index (EVCD, Add (Local1, 0x01))), Index (
                EVCD, Local1))
            Increment (Local1)
        }

        Decrement (EVCT)
        Release (MUEV)
        If (LNotEqual (EVCT, 0x00))
        {
            And (HKEV, 0x02, Local0)
            If (LEqual (Local0, 0x02))
            {
                Return (0x00)
            }

            Notify (VALZ, 0x80)
            Sleep (0x64)
        }

        Return (Local0)
    }

    Method (INFO, 0, Serialized)
    {
        Store (GETE (), Local0)
        Return (Local0)
    }

So EVCD is a queue of events, and EVCT is the number of events in the
queue. NTFY calls PUTE, which puts an event in the queue. INFO calls
GETE to read an event from the queue. Reading an event consists of
copying out the first object out of EVCD, then copying each subsequent
object to the previous element in EVCD. Inefficient, but it should work.

Except for that fact that EVCD will initially contain uninitialized
objects. So until/unless PUTE writes every element in EVCD that copy
loop is going to dereference unintialized objects, which causes the
exceptions.

GETE/PUTE are the only methods which write to EVCD, so there's not some
initialization method or something like that which we're forgetting to
call. We could do something like call PUTE 64 times during
initialization, but that's relying on a BIOS implementation detail which
just seems like a bad idea.

My only guess is that Windows is more permissive about this sort of
thing than ACPICA is, but by my reading of the spec throwing a fatal
error is what's supposed to happen.

Also a side note of interest: The machines that suffer from this all
also have a WMI interface with code to support hotkeys, and that code
shouldn't suffer from this problem. Unfortunately these models are the
only ones I've seen with this WMI interface, and the code to generate
the hotkey events isn't executed if the OS reports itself as Vista or
newer.