Re: Possibly SATA related freeze killed networking and RAID

[Tejun Heo <htejun@gmail.com>]

Phillip Susi wrote:
> Tejun Heo wrote:
>> Because SFF ATA controller don't have IRQ pending bit.  You don't know
>> whether IRQ is raised or not.  Plus, accessing the status register which
>> clears pending IRQ can be very slow on PATA machines.  It has to go
>> through the PCI and ATA bus and come back.  So, unconditionally trying
>> to clear IRQ by accessing Status can incur noticeable overhead if the
>> IRQ is shared with devices which raise a lot of IRQs.
> 
> There HAS to be a way to determine if that device generated the
> interrupt, or the interrupt can not be shared.  Since the kernel said
> nobody cared about the interrupt, that indicates that the sata driver
> checked the status register and realized the sata chip didn't generate
> the interrupt, and returned to the kernel letting it know that the
> interrupt was not for it.

Surprise, surprise.  There's no way to tell whether the controller
raised interrupt or not if command is not in progress.  As I said
before, there's no IRQ pending bit.  While processing commands, you can
tell by looking at other status registers but when there's nothing in
flight and the controller determines it's a good time to raise a
spurious interrupt, there's no way you can tell.  That dang SFF
interface is like 15+ years old.

But we can still make things pretty robust.  We're working on it.

Thanks.

-- 
tejun