Re: PROBLEM: APIC on a Pentium Classic SMP, 2.4.21-pre2 and 2.4.21-pre3 ksymoops

[Mikael Pettersson <mikpe@csd.uu.se>]

Maciej W. Rozycki writes:
 > > Fixing the oops is easy (see below), but the real problem is
 > > that 2.4.21-pre2 apparently broke MP table parsing on your HW.
 > > I suggest you sprinkle tracing printk()s in setup/smpboot/mpparse
 > > and compare 2.4.20 (good) and later (bad) to see where things
 > > start to diverge.
 > 
 >  There is no need to -- the problem is already known.  Mikael, if you need
 > additional details on how default MP configurations work in our code, feel

I think I nailed it.

First I found one very strange thing in Mathieu's boot log:

--- mpbug-2.4.20	Wed Sep 10 17:19:05 2003
+++ mpbug-2.4.23-pre3	Wed Sep 10 17:18:44 2003
...
+DMI not present.
 Intel MultiProcessor Specification v1.1
 Virtual Wire compatibility mode.
 Default MP configuration #6

This means construct_default_ISA_mptable() still gets called.
Ok so far.

...
 ENABLING IO-APIC IRQs
 Setting 2 in the phys_id_present_map
 ...changing IO-APIC physical APIC ID to 2 ... ok.

smp_found_config is true, we're now in setup_IO_APIC()
and have completed setup_ioapic_ids_from_mpc(). Ok so far.

-init IO_APIC IRQs
-IO-APIC (apicid-pin) 2-0 not connected.

THIS IS BAD. setup_IO_APIC() calls setup_IO_APIC_IRQs(),
which starts by printk()ing the first line above.
This line is missing from the 2.4.23-pre3 dmesg log, which
seems like an impossibility.

At this point I was thinking "memory corruption",
and the following struck me:

What used to be arrays (mp_irqs[] etc) are now pointers to
memory which is sized and allocated by smp_read_mpc().
In the case when construct_default_ISA_mptable() is called,
smp_read_mpc() is _not_ called, the pointers never get initialised,
and reads and writes of these arrays end up in la-la land.

The fix would be to add allocation and initialisation of
these pointers at the start of construct_default_ISA_mptable().

I'll prepare a patch doing this sometime tomorrow.

/Mikael