Re: 2.6.21-rc7-mm1 BUG at kernel/sched-clock.c:175 init_sched_clock()

[Tejun Heo <htejun@gmail.com>]

Alan Cox wrote:
>> diff --git a/drivers/ata/ahci.c b/drivers/ata/ahci.c
>> index 72c286e..a128717 100644
>> --- a/drivers/ata/ahci.c
>> +++ b/drivers/ata/ahci.c
>> @@ -1369,10 +1369,14 @@ static unsigned int ahci_qc_issue(struct ata_queued_cmd *qc)
>>  
>>  static void ahci_freeze(struct ata_port *ap)
>>  {
>> +	void __iomem *mmio = ap->host->iomap[AHCI_PCI_BAR];
>>  	void __iomem *port_mmio = ahci_port_base(ap);
>>  
>>  	/* turn IRQ off */
>>  	writel(0, port_mmio + PORT_IRQ_MASK);
>> +
>> +	/* clear IRQ pending bit */
>> +	writel(1 << ap->port_no, mmio + HOST_IRQ_STAT);
>>  }
>>  
> 
> Actually now I think about it the problem is even worse. On older systems
> IRQ delivery for SMP machines is asynchronous. So the following can occur
> 
> 
> 		device raises IRQ line
> 		IRQ request hits APIC bus
> 
> 		We write IRQ off
> 		We clear IRQ pending
> 
> 		We read back to ensure the write took
> 	
> 
> 		IRQ arrives on CPU APIC
> 
> (and yes this happens on dual P3 boxes - NE2K has some fantastic bugs at
> one point due to this)

I think the second patch should work regardless of IRQ posting.  No?

-- 
tejun