Message-ID: <CAGCDX1=Y_UWpmo0RiexYh-MJ8v6wwXZ=14h+NS5cmi5Nboytow@mail.gmail.com>
Date:	Fri, 13 Jul 2012 13:29:14 +0800
From:	Jason Lin <kernel.jason@...il.com>
To:	Andy Fleming <afleming@...il.com>
Cc:	Florian Fainelli <florian@...nwrt.org>, netdev@...r.kernel.org
Subject: Re: [RFC] net:phy:phylib: phy shares the same interrupt with mac

Dear:
But I think there have some performance issues.
Because the phy should execute its own ISR only when the interrupt is
issued by phy.
This will cause some unnecessary MDIO access.

By the way, PHY_DUMMY_IRQ is defined in the platform definition file.
One should get PHY_DUMMY_IRQ by platform_get_irq function.
So the driver will look like this:

/*** platform related file ***************/
static struct resource mac_0_resources[] = {
    {
        .start  = MAC_PA_BASE,
        .end    = MAC_PA_BASE + SZ_4K - 1,
        .flags  = IORESOURCE_MEM,
    }, {
        .start  = IRQ_MAC,
        .flags  = IORESOURCE_IRQ,
    }, {
        .start  = IRQ_PHY,
        .flags  = IORESOURCE_IRQ,
    },
};


/*** driver file ***************/
mydriver_probe(struct platform_device *pdev)
{
    int mac_irq, phy_irq, i;

    mac_irq = platform_get_irq(pdev, 0);
    if (mac_irq < 0)
        return mac_irq;

    phy_irq = platform_get_irq(pdev, 1);
    if (phy_irq < 0)
        phy_irq = PHY_POLL;

    for (i = 0; i < PHY_MAX_ADDR; i++)
        priv->mii_bus->irq[i] = phy_irq;
}

mydriver_napi_poll(struct napi_struct *napi, int budget)
{
    struct mydriver_priv *priv = = container_of(napi, struct
mydriver_priv, napi);
    struct phy_device *phydev = priv->phydev;
    unsigned int status = get_int_status();

#ifdef CONFIG_MAC_CONTROL_PHY_INT
    if (status & PHY_INT) {
        tasklet_schedule(&phydev->phy_task);
    }
#endif

}

/**** phy lib modification ***************/
diff --git a/drivers/net/phy/phy.c b/drivers/net/phy/phy.c
index 3cbda08..482f7e5 100644
--- a/drivers/net/phy/phy.c
+++ b/drivers/net/phy/phy.c
@@ -416,7 +416,7 @@ out_unlock:
 }
 EXPORT_SYMBOL(phy_start_aneg);

-
+static void phy_tasklet(unsigned long data);
 static void phy_change(struct work_struct *work);

 /**
@@ -523,10 +523,12 @@ static irqreturn_t phy_interrupt(int irq, void *phy_dat)
     * context, so we need to disable the irq here.  A work
     * queue will write the PHY to disable and clear the
     * interrupt, and then reenable the irq line. */
-   disable_irq_nosync(irq);
+/* disable_irq_nosync(irq);
    atomic_inc(&phydev->irq_disable);

-   schedule_work(&phydev->phy_queue);
+   schedule_work(&phydev->phy_queue);*/
+
+   tasklet_schedule(&phydev->phy_task);

    return IRQ_HANDLED;
 }
@@ -591,6 +593,7 @@ int phy_start_interrupts(struct phy_device *phydev)
 {
    int err = 0;

+   tasklet_init(&phydev->phy_task, phy_tasklet, (unsigned long)phydev);
    INIT_WORK(&phydev->phy_queue, phy_change);

    atomic_set(&phydev->irq_disable, 0);
@@ -633,6 +636,7 @@ int phy_stop_interrupts(struct phy_device *phydev)
     * possibly pending and take care of the matter below.
     */
    cancel_work_sync(&phydev->phy_queue);
+   tasklet_kill(&phydev->phy_task);
    /*
     * If work indeed has been cancelled, disable_irq() will have
     * been left unbalanced from phy_interrupt() and enable_irq()
@@ -645,7 +649,13 @@ int phy_stop_interrupts(struct phy_device *phydev)
 }
 EXPORT_SYMBOL(phy_stop_interrupts);


-
+static void phy_tasklet(unsigned long data)
+{
+   struct phy_device *phydev = (struct phy_device *)data;
+   disable_irq_nosync(phydev->irq);
+   atomic_inc(&phydev->irq_disable);
+   schedule_work(&phydev->phy_queue);
+}
 /**
  * phy_change - Scheduled by the phy_interrupt/timer to handle PHY changes
  * @work: work_struct that describes the work to be done
diff --git a/include/linux/phy.h b/include/linux/phy.h
index c599f7e..04203a7 100644
--- a/include/linux/phy.h
+++ b/include/linux/phy.h
@@ -330,6 +330,7 @@ struct phy_device {
    /* Interrupt and Polling infrastructure */
    struct work_struct phy_queue;
    struct delayed_work state_queue;
+   struct tasklet_struct phy_task;
    atomic_t irq_disable;

    struct mutex lock;

/*** end modification **********************************/

how about this modification?
F.Y.I.
Thanks.

2012/7/13 Jason Lin <kernel.jason@...il.com>:
> Thank for your reply.
> My case is the second one.
> change request_irq with IRQF_SHARED, and assign the same MAC irq number to phy
> (mii_bus->irq[i] = mac_irq)
> This way can fix my issue.
>
> Thanks again.
>
> 2012/7/13 Andy Fleming <afleming@...il.com>:
>> On Tue, Jul 10, 2012 at 8:32 PM, Jason Lin <kernel.jason@...il.com> wrote:
>>> Dear :
>>> I describe my situation again.
>>> In my hardware design, the interrupt (INT) pin of phy is connected to
>>> the INT input pin of MAC.
>>> In other words, one of triggering interrupt condition of MAC is
>>> related to phy's interrupt.
>>> So the phy will share the same "IRQ pin" with MAC.
>>> As described above, the best solution is the INT pin of phy is
>>> connected to architecture independently.
>>> But, the hardware architecture cannot modify easily.
>>> So I think
>>> 1. We can assign dummy IRQ number (which is a empty IRQ number but
>>> large than zero) to phy.
>>> phydev->irq = PHY_DUMMY_IRQ (this value is depend on architecture)
>>> 2. Change to do the soft IRQ in phy's ISR.
>>>     To schedule workqueue in this soft IRQ.
>>> In this way, the MAC can schedule phy's soft IRQ to do phy's work
>>> queue (to do ack phy's interrupt ... etc).
>>>
>>> Dose it make sense?
>>>
>>
>> If PHY_IGNORE_INTERRUPT doesn't currently work to allow a MAC's driver
>> to manage PHY interrupts, then we need to fix PHY Lib to support this
>> correctly. PHY_IGNORE_INTERRUPT was meant for this purpose. We don't
>> want to start defining interrupt numbers which may conflict with real
>> numbers, and have to be constantly re-defined by various systems to
>> avoid that.
>>
>> Your notion of making the phy_enable_interrupts() and
>> phy_disable_interrupts() code available to the MAC drivers sounds like
>> the right approach to me. But you might want to implement your own
>> version. The biggest problem with PHY interrupts is that masking them
>> requires an MDIO transaction, which is *slow*. If you can mask the
>> interrupt by writing a bit in a memory-mapped register, it's far
>> better to do it that way, at interrupt time, and *then* schedule the
>> PHY code to be run from a work queue. However, your driver probably
>> already *does* have a workqueue of some sort. If so, what you should
>> probably do is implement a new version of phy_change(), that doesn't
>> have to deal with the weird PHY interrupt masking issues. Something
>> like this:
>>
>> mydriver_phy_change(struct mydriver_priv *priv)
>> {
>>         int err;
>>         struct phy_device *phydev = priv->phydev;
>>
>>         if (phydev->drv->did_interrupt &&
>>             !phydev->drv->did_interrupt(phydev))
>>                 goto ignore;
>>
>>         err = phy_disable_interrupts(phydev);
>>
>>         if (err)
>>                 goto phy_err;
>>
>>         mutex_lock(&phydev->lock);
>>         if ((PHY_RUNNING == phydev->state) || (PHY_NOLINK == phydev->state))
>>                 phydev->state = PHY_CHANGELINK;
>>         mutex_unlock(&phydev->lock);
>>
>>         /* Reenable interrupts */
>>         if (PHY_HALTED != phydev->state)
>>                 err = phy_config_interrupt(phydev, PHY_INTERRUPT_ENABLED);
>>
>>         if (err)
>>                 goto irq_enable_err;
>>
>>         /* reschedule state queue work to run as soon as possible */
>>         cancel_delayed_work_sync(&phydev->state_queue);
>>         schedule_delayed_work(&phydev->state_queue, 0);
>>
>>         return;
>>
>> ignore:
>>         return;
>>
>> irq_enable_err:
>> phy_err:
>>         phy_error(phydev);
>> }
>>
>>
>> Of course, now I re-read your question, and I'm not sure I've
>> interpreted it correctly. Are you saying your MAC has control of the
>> PHY interrupt (ie - the interrupt sets a bit in some event register in
>> your MAC, and you can MASK/ACK it from your driver), or that the PHY
>> shares the same interrupt pin?
>>
>> If it's the second one, you don't need to do anything, but allow your
>> MAC driver to register its interrupt as a shared interrupt, and allow
>> the PHY to manage its own interrupts, as usual.
>>
>> Andy
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists