On Tuesday, November 09, 2010 9:59 PM, Marc Kleine-Budde wrote :
>
> On 11/09/2010 01:26 PM, Tomoya MORINAGA wrote:
> >>>> Can you please explain me your locking sheme? If I understand the
> >>>> documenation correctly the two message interfaces can be used mutual.
> >>>> And you use one for rx the other one for tx.
> >>>
> >>> I show our locking scheme.
> >>> When CPU accesses MessageRAM via IF1, CPU protect until read-modify-write
> >>> so that IF2 access not occurred, vice versa.
> >>
> >> Why is that needed?
> >
> > For MessageRAM data consistency.
>
> As far as I understand the datasheet the access to IF1 and IF2 is
> completely independent. Why do you lock here?
You are right.
Each IF1 and IF2 is completely independent.
All message objects can be accessed via both IF1 and IF2.
There is a possibility that a message object is accessed by IF1 and IF2 simultaneously.
But this driver separates message object by Tx/Rx completely.(Rx:1~26, Tx:27~32)
Thus, these locks are unnecessary.
I will delete these.
>
> [...]
>
> >>>> Please use just "debug" level not warning here. Consider to use
> >>>> netdev_dbg() instead. IMHO the __func__ can be dropped and the
> >>>> "official" name for the error is "Error Warning".
> >>>
> >>> I want to know the reason.
> >>> Why is it not dev_warn but netdev_dbg ?
> >>
> >> If you use warning level it would end up on the console or and in the
> >> syslog. It's quite complicated (for programs) to get information from
> >> there. This is why we send CAN error frames. They hold the same
> >> information but int a binary form, thus it's easier to process.
> >
> > I understand the reason.
> > BTW, Why do you say not dev_dbg but netdev_dbg ?
>
> Sorry - netdev_dbg() is easier to use, its first argument is the
> netdevice, while dev_dbg needs a device and that's deeply hidden in the
> netdevice.
>
I understand.
> [...]
>
> >>>>> +static netdev_tx_t pch_xmit(struct sk_buff *skb, struct net_device *ndev)
> >>>>> +{
> >>>>> + unsigned long flags;
> >>>>> + struct pch_can_priv *priv = netdev_priv(ndev);
> >>>>> + struct can_frame *cf = (struct can_frame *)skb->data;
> >>>>> + int tx_buffer_avail = 0;
> >>>>
> >>>> What I'm totally missing is the TX flow controll. Your driver has to
> >>>> ensure that the package leave the controller in the order that come
> >>>> into the xmit function. Further you have to stop your xmit queue if
> >>>> you're out of tx objects and reenable if you have a object free.
> >>>>
> >>>> Use netif_stop_queue() and netif_wake_queue() for this.
> >>>
> >>> In this code, I think "out of tx objects" cannot be occurred.
> >>
> >> It's not a matter of code it's the hardware. You cannot put more than a
> >> certain number of CAN frames into the hardware. If you have a CAN bus at
> >> a certain speed, you can only send a certain number of CAN frames in a
> >> second. So you cannot push more than this amount of frames/s into the
> >> hardware.
> >>
> >>> Nevertheless, are netif_stop_queue() and netif_wake_queue() is necessary ?
> >>
> >> Yes.
> >
> > I can' understand your issue.
> > Please can you hear my opinion?
> >
> > Please see the head of pch_xmit.
> >
> >>> + if (priv->tx_obj == (PCH_OBJ_NUM + 1)) { /* Point tail Obj + 1 */
> >>> + while (ioread32(&priv->regs->treq2) & 0xfc00)
> >>> + udelay(1);
> >
> > When points tail of Tx message object,
> > this driver waits until completion of all tx messaeg objects.
>
> Looping busy it not an option here.
>
> > Thus, application/driver ought not to be able to put Tx object exceed the number of tx message object.
> > Thus I think these code(netif_stop_queue/netif_wake_queue) are completely redundant.
>
> Nope - please remove the waiting completely and convert your flow
> control to netif_stop_queue/netif_wake_queue.
>
I see.
I will remove like above.
BTW, Let me know the reason.
Could you explain the reason why you deny looping busy loop ?
Thanks,
Tomoya MORINAGA
OKI SEMICONDUCTOR CO., LTD.
On 11/11/2010 10:56 AM, Tomoya MORINAGA wrote:
>>>>> + if (priv->tx_obj == (PCH_OBJ_NUM + 1)) { /* Point tail Obj + 1 */
>>>>> + while (ioread32(&priv->regs->treq2) & 0xfc00)
>>>>> + udelay(1);
>>>
>>> When points tail of Tx message object,
>>> this driver waits until completion of all tx messaeg objects.
>>
>> Looping busy it not an option here.
>>
>>> Thus, application/driver ought not to be able to put Tx object exceed the number of tx message object.
>>> Thus I think these code(netif_stop_queue/netif_wake_queue) are completely redundant.
>>
>> Nope - please remove the waiting completely and convert your flow
>> control to netif_stop_queue/netif_wake_queue.
>>
>
> I see.
> I will remove like above.
>
> BTW, Let me know the reason.
> Could you explain the reason why you deny looping busy loop ?
That would kill performance.
cheers, Marc
--
Pengutronix e.K. | Marc Kleine-Budde |
Industrial Linux Solutions | Phone: +49-231-2826-924 |
Vertretung West/Dortmund | Fax: +49-5121-206917-5555 |
Amtsgericht Hildesheim, HRA 2686 | http://www.pengutronix.de |