Re: libata git tree, mbox queue status and contents

[Tejun Heo <htejun@gmail.com>]

Alan Cox wrote:
>> * Albert: irq_on/off.  Really need to give this some thought.  Not sure 
>> I like where this model is going.  Polling and twiddling irq on/off 
>> should be kept to a minimum, because it's sorta an admission that the 
>> host state machine has broken down, and we need to bandaid.  Its a 
>> bandaid not a root-cause solution.
> 
> I think of it more as an admission that the IDE design is lacking in a
> few areas. No suprise as its an emulation of a 15 year old interface that
> was normally used polled.

I second Alan's opinion.  The root cause of the problem is that the
controller lacks reliable IRQ masking and pending register.  If you add
a slightly malfunctioning controller and/or device into the mix, it's
guaranteed to break.

And there's another addition to the mess from SATA land which I
discovered a few days ago.  Some SATA controllers with SFF interface
don't send non-command H2D Reg FIS when ATA_NIEN is toggled, so the
updated ATA_NIEN is sent together with the next command && some SATA
devices don't honor ATA_NIEN on command H2D Reg FIS.  Fun!

I think we should go combined IRQ + polling mode Jeff once mentioned so
that we automatically fallback to polling if IRQ doesn't work for some
reason.  To do that, we need reliable IRQ masking && the only way to
achieve proper IRQ masking on SFF controllers seems plugging it from the
host side.

Thanks.

-- 
tejun