Re: [PATCH 5/7] libata: move and reduce locking to the pio data xfer functions

[Tejun Heo <htejun@gmail.com>]

Alan Cox wrote:
>>> I'd say this is a non-starter. It solves nothing and means PIO in libata
>>> is still basically unusable.
>> It doesn't solve the problem completely but still helps, FWIW.  I was
> 
> Most transfers for PIO are a single 512 byte transfer per command.

Not for disks but, yeah, who uses PIO for disks.

>> hoping we could lock only for the last transfer (word).  Would it
>> complicate ->data_xfer() too much?
> 
> I don't think so. We need to complicate it far more to add 32bit
> transfers. We also only have a few ->data_xfer implementations so its
> easy to propogate the chance.
> 
> Why do we need the lock on the last word transferred anyway. I'm missing
> a detail here ?

The problem is that controllers queue IRQ till the end of transfer and
raise it right after the last transfer completes.  If
WQ-active-ignore-IRQ flag is set at that point && we're not holding the
lock, the IRQ handler will ignore the IRQ without clearing it, so we get
nobody-cared right after the last transfer.  So, the last transfer and
clearing of WQ-active-ignore-IRQ flag should be atomic w.r.t. the IRQ
handler.

-- 
tejun