Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <MN2PR01MB535838492432C910F2381F929F6F9@MN2PR01MB5358.prod.exchangelabs.com>
In-Reply-To: <MN2PR01MB535838492432C910F2381F929F6F9@MN2PR01MB5358.prod.exchangelabs.com>
From:   Arnd Bergmann <arnd@arndb.de>
Date:   Thu, 25 Aug 2022 14:29:21 +0200
Message-ID: <CAK8P3a3rteX8N2A-K-8bmvWaUhLJjXLV_8JNLCY0ezBzkQODUg@mail.gmail.com>
Subject: Re: [PATCH] i2c: pasemi: Add IRQ support for Apple Silicon
To:     Arminder Singh <arminders208@outlook.com>
Cc:     linux-kernel@vger.kernel.org, Sven Peter <sven@svenpeter.dev>,
        Alyssa Rosenzweig <alyssa@rosenzweig.io>,
        Hector Martin <marcan@marcan.st>,
        Michael Ellerman <mpe@ellerman.id.au>,
        Benjamin Herrenschmidt <benh@kernel.crashing.org>,
        Paul Mackerras <paulus@samba.org>,
        linux-arm-kernel@lists.infradead.org,
        linuxppc-dev@lists.ozlabs.org, linux-i2c@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Sat, Aug 20, 2022 at 9:45 PM Arminder Singh <arminders208@outlook.com> wrote:
>
> I also fixed a quick checkpatch warning on line 303. "i ++" is now "i++".

In general, anything that is mentioned in the changelog as "also done this"
is worth splitting into a separate patch, or dropping from the patch, to
make reviewing easier.

In this case, I would just not do the trivial change. Alternatively you
can consider doing a larger patch for coding style cleanup on the
patch, as there are possibly other issues as well. Usually it's not worth
changing things unless they hurt readability.

> @@ -80,14 +81,21 @@ static int pasemi_smb_waitready(struct pasemi_smbus *smbus)
>  {
>         int timeout = 10;
>         unsigned int status;
> +       unsigned int bitmask = SMSTA_XEN | SMSTA_MTN;
>
> +       if (smbus->use_irq) {
> +               reinit_completion(&smbus->irq_completion);
> +               reg_write(smbus, REG_IMASK, bitmask);
> +               wait_for_completion_timeout(&smbus->irq_completion, msecs_to_jiffies(10));
>                 status = reg_read(smbus, REG_SMSTA);
>         }
>
> +
>         /* Got NACK? */
>         if (status & SMSTA_MTN)
>                 return -ENXIO;
...
> @@ -356,3 +366,12 @@ int pasemi_i2c_common_probe(struct pasemi_smbus *smbus)
>
>         return 0;
>  }
> +
> +irqreturn_t pasemi_irq_handler(int irq, void *dev_id)
> +{
> +       struct pasemi_smbus *smbus = (struct pasemi_smbus *)dev_id;
> +
> +       reg_write(smbus, REG_IMASK, 0);
> +       complete(&smbus->irq_completion);
> +       return IRQ_HANDLED;
> +}

I think the completion structure gets out of sync if you run into a
timeout here,
so a subsequent wait_for_completion will never complete after we missed one
interrupt.

Since this already causes a bus reset, I think you can just do
reinit_completion() at the end of pasemi_reset() to avoid this.

        Arnd