Re: [PATCH v5 1/4] mfd: add support for Diolan DLN-2 devices

From: Johan Hovold
Date: Wed Sep 24 2014 - 11:10:31 EST


On Wed, Sep 24, 2014 at 05:54:15PM +0300, Octavian Purdila wrote:
> On Wed, Sep 24, 2014 at 4:54 PM, Johan Hovold <johan@xxxxxxxxxx> wrote:
> > On Wed, Sep 24, 2014 at 04:36:22PM +0300, Octavian Purdila wrote:
> >> On Wed, Sep 24, 2014 at 1:48 PM, Johan Hovold <johan@xxxxxxxxxx> wrote:
> >> > On Fri, Sep 19, 2014 at 11:22:42PM +0300, Octavian Purdila wrote:
> >>
> >> <snip>
> >>
> >> >> + * dln2_dev.mod_rx_slots and then the echo header field to index the
> >> >> + * slots field and find the receive context for a particular
> >> >> + * request.
> >> >> + */
> >> >> +struct dln2_mod_rx_slots {
> >> >> + /* RX slots bitmap */
> >> >> + unsigned long bmap;
> >> >> +
> >> >> + /* used to wait for a free RX slot */
> >> >> + wait_queue_head_t wq;
> >> >> +
> >> >> + /* used to wait for an RX operation to complete */
> >> >> + struct dln2_rx_context slots[DLN2_MAX_RX_SLOTS];
> >> >> +
> >> >> + /* device has been disconnected */
> >> >> + bool disconnected;
> >> >
> >> > This belongs in the dln2_dev struct.
> >> >
> >> > I think you're overcomplicating the disconnect handling by intertwining
> >> > it with your slots.
> >> >
> >> > Add a lock, an active-transfer counter, a disconnected flag, and a wait
> >> > queue to struct dln2_dev.
> >> >
> >>
> >> I agree that disconnected is better suited in dln2_dev.
> >>
> >> However, I don't think that we need the active-transfer counter and a
> >> new wait queue. We can simply use the existing waiting queues and the
> >> implicit alloc_rx_slot()/free_rx_slot() calls to see if we are still
> >> waiting for I/O.
> >
> > Just because you can reuse them doesn't mean it's a good idea. By
> > separating a generic disconnect solution from your custom slot
> > implementation we get something that is way easier to verify for
> > correctness and that could also be reused in other drivers.
>
> Maybe I miss-understood what you are proposing, let me try to
> summarize it to see if I got it right.
>
> You are suggesting to add a counter, increment it in alloc_rx_slot(),
> decrement it in free_rx_slot().

No increment it at the start of _dln2_transfer, and decrement it before
returning from that function.

> Then add a new waitqueue in dln2_dev
> and in free_rx_slot() wake it up while in disconnect do a wait_event()
> on it and check for the counter.

Where you also wake the disconnect (or wait-until-sent) wait queue.

> Also, alloc_rx_slot() should fail if
> the disconnect flag is set.

That is not required, but you can bail out early after alloc_rx_slot if
the disconnect flag is set (no locking).

> In this case we are still coupled to the slots implementation, in the
> sense that you would need to understand the slots implementation to
> understand how the disconnect works. We are also doing two wake-up
> operations which I find redundant and which does not add much value in
> clarity (since we still need to wake-up all completions for each
> handle).
>
> I do agree that using a counter instead of checking the bitmaps is
> cleaner though.

You only need to the wake up if disconnected is set when returning from
_dln2_transfer.

Sure, the optimisation bit -- to abort any ongoing transfer -- still
requires some insight into the slot implementation.

But this way everything disconnect related (correctness-wise) is
isolated to _dln2_transfer and dln2_disconnect.

Johan
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/