[BUG] [PATCH] Infinite loop in arch/i386/kernel/io_apic.c

["Sam Varshavchik" <mrsam@courier-mta.com>]

[-- Attachment #1: Type: text/plain, Size: 680 bytes --]

Good morning/afternoon/evening/night, 

I see that there was some debug code added to io_apic.c in kernel 2.4.6, 
that's compiled by default. 

It looks to me like there's a typo in the APIC_LOCKUP_DEBUG code that 
results in certain hardware setups potentially triggering an infinite loop, 
while holding the ioapic spinlock: namely when the irq_2_pin list contains 
more than one list member: the loop pointer gets reinitialized each time at 
the top of the loop - kablooey. 

After upgrading from 2.4.3 to 2.4.7, my 440GX motherboard (with a couple of 
PCI devices) keels over fast enough that even the NMI watchdog doesn't help, 
and the following patch fixes it. 

-- 
Sam 


[-- Attachment #2: linux-2.4.6-ioapicdebugfix.patch --]
[-- Type: text/plain, Size: 906 bytes --]

*** linux/arch/i386/kernel/io_apic.c.orig	Tue Oct  9 21:11:10 2001
--- linux/arch/i386/kernel/io_apic.c	Tue Oct  9 21:13:03 2001
***************
*** 1248,1261 ****
  	ack_APIC_irq();
  
  	if (!(v & (1 << (i & 0x1f)))) {
  #ifdef APIC_MISMATCH_DEBUG
  		atomic_inc(&irq_mis_count);
  #endif
  		spin_lock(&ioapic_lock);
  		__mask_and_edge_IO_APIC_irq(irq);
  #ifdef APIC_LOCKUP_DEBUG
! 		for (;;) {
! 			struct irq_pin_list *entry = irq_2_pin + irq;
  			unsigned int reg;
  
  			if (entry->pin == -1)
--- 1248,1264 ----
  	ack_APIC_irq();
  
  	if (!(v & (1 << (i & 0x1f)))) {
+ #ifdef APIC_LOCKUP_DEBUG
+ 		struct irq_pin_list *entry;
+ #endif
+ 
  #ifdef APIC_MISMATCH_DEBUG
  		atomic_inc(&irq_mis_count);
  #endif
  		spin_lock(&ioapic_lock);
  		__mask_and_edge_IO_APIC_irq(irq);
  #ifdef APIC_LOCKUP_DEBUG
! 		for (entry = irq_2_pin + irq;;) {
  			unsigned int reg;
  
  			if (entry->pin == -1)