Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com>
Subject: Re: [PATCH i2c-next v6] i2c: aspeed: Handle master/slave combined irq
 events properly
To:     Guenter Roeck <linux@roeck-us.net>
Cc:     Joel Stanley <joel@jms.id.au>, linux-aspeed@lists.ozlabs.org,
        Vernon Mauery <vernon.mauery@linux.intel.com>,
        OpenBMC Maillist <openbmc@lists.ozlabs.org>,
        Brendan Higgins <brendanhiggins@google.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        linux-i2c@vger.kernel.org, jarkko.nikula@linux.intel.com,
        =?UTF-8?Q?C=c3=a9dric_Le_Goater?= <clg@kaod.org>,
        Linux ARM <linux-arm-kernel@lists.infradead.org>,
        James Feist <james.feist@linux.intel.com>
References: <20180823225731.19063-1-jae.hyun.yoo@linux.intel.com>
 <20180911183734.GA21976@roeck-us.net>
 <1f34fe8c-69ef-5f2d-25dc-d5f6037cc558@linux.intel.com>
 <20180911204107.GA26017@roeck-us.net>
 <e5135af4-9105-214f-5de2-65cc03804222@linux.intel.com>
 <CACPK8Xdv3aP_zOHAL0WMgwvkpBHAUesU7OdhW==jgvXjRmt-Zw@mail.gmail.com>
 <20180911233302.GA18799@roeck-us.net>
 <5698ca34-14c9-8d05-c4e6-5acf85ff9d14@linux.intel.com>
 <20180912013449.GA12612@roeck-us.net>
Message-ID: <7fd98646-fb5a-be4d-ce37-84b74e0fa8b3@linux.intel.com>
Date:   Wed, 12 Sep 2018 09:54:51 -0700
User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:52.0) Gecko/20100101
 Thunderbird/52.9.1
MIME-Version: 1.0
In-Reply-To: <20180912013449.GA12612@roeck-us.net>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 9/11/2018 6:34 PM, Guenter Roeck wrote:
> On Tue, Sep 11, 2018 at 04:58:44PM -0700, Jae Hyun Yoo wrote:
>> On 9/11/2018 4:33 PM, Guenter Roeck wrote:
>>> Looking into the patch, clearing the interrupt status at the end of an
>>> interrupt handler is always suspicious and tends to result in race
>>> conditions (because additional interrupts may have arrived while handling
>>> the existing interrupts, or because interrupt handling itself may trigger
>>> another interrupt). With that in mind, the following patch fixes the
>>> problem for me.
>>>
>>> Guenter
>>>
>>> ---
>>>
>>> diff --git a/drivers/i2c/busses/i2c-aspeed.c b/drivers/i2c/busses/i2c-aspeed.c
>>> index c258c4d9a4c0..c488e6950b7c 100644
>>> --- a/drivers/i2c/busses/i2c-aspeed.c
>>> +++ b/drivers/i2c/busses/i2c-aspeed.c
>>> @@ -552,6 +552,8 @@ static irqreturn_t aspeed_i2c_bus_irq(int irq, void *dev_id)
>>>   	spin_lock(&bus->lock);
>>>   	irq_received = readl(bus->base + ASPEED_I2C_INTR_STS_REG);
>>> +	/* Ack all interrupt bits. */
>>> +	writel(irq_received, bus->base + ASPEED_I2C_INTR_STS_REG);
>>>   	irq_remaining = irq_received;
>>>   #if IS_ENABLED(CONFIG_I2C_SLAVE)
>>> @@ -584,8 +586,6 @@ static irqreturn_t aspeed_i2c_bus_irq(int irq, void *dev_id)
>>>   			"irq handled != irq. expected 0x%08x, but was 0x%08x\n",
>>>   			irq_received, irq_handled);
>>> -	/* Ack all interrupt bits. */
>>> -	writel(irq_received, bus->base + ASPEED_I2C_INTR_STS_REG);
>>>   	spin_unlock(&bus->lock);
>>>   	return irq_remaining ? IRQ_NONE : IRQ_HANDLED;
>>>   }
>>>
>>
>> My intention of putting the code at the end of interrupt handler was,
>> to reduce possibility of combined irq calls which is explained in this
>> patch. But YES, I agree with you. It could make a potential race
> 
> Hmm, yes, but that doesn't explain why it would make sense to acknowledge
> the interrupt late. The interrupt ack only means "I am going to handle these
> interrupts". If additional interrupts arrive while the interrupt handler
> is active, those will have to be acknowledged separately.
> 
> Sure, there is a risk that an interrupt arrives while the handler is
> running, and that it is handled but not acknowledged. That can happen
> with pretty much all interrupt handlers, and there are mitigations to
> limit the impact (for example, read the interrupt status register in
> a loop until no more interrupts are pending). But acknowledging
> an interrupt that was possibly not handled is always bad idea.

Well, that's generally right but not always. Sometimes that depends on
hardware and Aspeed I2C is the case.

This is a description from Aspeed AST2500 datasheet:
   I2CD10 Interrupt Status Register
   bit 2 Receive Done Interrupt status
         S/W needs to clear this status bit to allow next data receiving.

It means, driver should hold this bit to prevent transition of hardware
state machine until the driver handles received data, so the bit should
be cleared at the end of interrupt handler.

Let me share my test result. Your code change works on 100KHz bus speed
but doesn't work well on 1MHz bus speed. Investigated that interrupt
handling is fast enough in 100KHz test but in 1MHz, most of data is
corrupted because the bit is cleared at the beginning of interrupt
handler so it allows receiving of the next data but the interrupt
handler isn't fast enough to read the data buffer on time. I checked
this problem on BMC-ME channel which ME sends lots of IPMB packets to
BMC at 1MHz speed. You could simply check the data corruption problem on
the BMC-ME channel.

My thought is, the current code is right for real Aspeed I2C hardware.
It seems that QEMU 3.0 model for witherspoon-bmc doesn't simulate the
actual Aspeed I2C hardware correctly.

Thanks,
Jae