[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <57322951.3040109@gmx.de>
Date: Tue, 10 May 2016 20:32:49 +0200
From: Lino Sanfilippo <LinoSanfilippo@....de>
To: Elad Kanfi <eladkan@...lanox.com>,
Lino Sanfilippo <lsanfil@...vell.com>,
David Miller <davem@...emloft.net>
Cc: Noam Camus <noamca@...lanox.com>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
"abrodkin@...opsys.com" <abrodkin@...opsys.com>,
Tal Zilcer <talz@...lanox.com>,
"netdev@...r.kernel.org" <netdev@...r.kernel.org>
Subject: Re: [PATCH v2 1/2] net: nps_enet: Sync access to packet sent flag
Hi Elad,
On 08.05.2016 15:44, Elad Kanfi wrote:
>
> After reviewing the code and your suggestion, it seems that we can do without the flag tx_packet_sent and therefor the first issue becomes irrelevant.
> The indication that a packet was sent is (tx_skb != NULL) , and the sequence will be:
>
> CPU A:
> 1. tx_skb = skb
> 2. wmb() /* make sure tx_skb reaches the RAM before the HW is informed and the IRQ is fired */
> 3. nps_enet_reg_set(priv, NPS_ENET_REG_TX_CTL, tx_ctrl.value); /* send frame */
>
> CPU B:
> 1. read tx_skb
> 2. if( tx_skb != NULL ) handle tx_skb
> 3. tx_skb = NULL
>
>
Ok, without the tx_packet_sent flag the code becomes simpler. But it
does not mean that we can toss the smp_rmb in the irq handler
completely. We still have to use a read barrier there to ensure that we
see the most recent value of tx_skb. E.g like this:
if (priv->tx_skb != NULL ) {
smp_rmb()
/ * handle tx_skb */
}
With both barriers in place the code should work as expected.
Regards,
Lino
Powered by blists - more mailing lists