Re: [PATCH v5 3/5] bnx2x: Eliminate duplicate barriers on weakly-ordered archs

From: Sinan Kaya
Date: Fri Mar 23 2018 - 12:31:22 EST


On 3/23/2018 12:20 PM, David Miller wrote:
> From: Sinan Kaya <okaya@xxxxxxxxxxxxxx>
> Date: Thu, 22 Mar 2018 13:10:00 -0400
>
>> Code includes wmb() followed by writel(). writel() already has a
>> barrier on some architectures like arm64.
> ...
>> @@ -4155,7 +4155,7 @@ netdev_tx_t bnx2x_start_xmit(struct sk_buff *skb, struct net_device *dev)
>> txdata->tx_db.data.prod += nbd;
>> barrier();
>>
>> - DOORBELL(bp, txdata->cid, txdata->tx_db.raw);
>> + DOORBELL_RELAXED(bp, txdata->cid, txdata->tx_db.raw);
>>
>> mmiowb();
> ...
>> @@ -2592,7 +2592,7 @@ static int bnx2x_run_loopback(struct bnx2x *bp, int loopback_mode)
>>
>> txdata->tx_db.data.prod += 2;
>> barrier();
>> - DOORBELL(bp, txdata->cid, txdata->tx_db.raw);
>> + DOORBELL_RELAXED(bp, txdata->cid, txdata->tx_db.raw);
>
> These are compiler barriers being used here, not wmb().
>
> Look, if I can't see a clear:
>
> wmb()
> writel()
>
> sequence in the patch hunks, I am going to keep pushing back on
> these changes.

Sorry, you got me confused now.

If you look at the code closer, you'll see this.

wmb();

txdata->tx_db.data.prod += nbd;
barrier();

DOORBELL(bp, txdata->cid, txdata->tx_db.raw);

and you also asked me to rename DOORBELL to DOORBELL_RELAXED() to make
it obvious that we have a relaxed operator inside the macro.

Did I miss something?

of course, treating barrier() universally as a write barrier is wrong.

>
> Thank you.
>


--
Sinan Kaya
Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project.