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;
Date:   Mon, 14 Mar 2022 08:57:05 -0700
From:   Jakub Kicinski <kuba@kernel.org>
To:     Biao Huang <biao.huang@mediatek.com>
Cc:     David Miller <davem@davemloft.net>,
        Rob Herring <robh+dt@kernel.org>,
        Bartosz Golaszewski <brgl@bgdev.pl>,
        Fabien Parent <fparent@baylibre.com>,
        Felix Fietkau <nbd@nbd.name>, John Crispin <john@phrozen.org>,
        Sean Wang <sean.wang@mediatek.com>,
        Mark Lee <Mark-MC.Lee@mediatek.com>,
        "Matthias Brugger" <matthias.bgg@gmail.com>,
        <netdev@vger.kernel.org>, <devicetree@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>,
        <linux-arm-kernel@lists.infradead.org>,
        <linux-mediatek@lists.infradead.org>,
        Yinghua Pan <ot_yinghua.pan@mediatek.com>,
        <srv_heupstream@mediatek.com>,
        Macpaul Lin <macpaul.lin@mediatek.com>
Subject: Re: [PATCH net-next v2 9/9] net: ethernet: mtk-star-emac: separate
 tx/rx handling with two NAPIs
Message-ID: <20220314085705.32033308@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
In-Reply-To: <2d0ab5290e63069f310987a4423ef2a46f02f1b3.camel@mediatek.com>
References: <20220127015857.9868-1-biao.huang@mediatek.com>
        <20220127015857.9868-10-biao.huang@mediatek.com>
        <20220127194338.01722b3c@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
        <2bdb6c9b5ec90b6c606b7db8c13f8acb34910b36.camel@mediatek.com>
        <20220128074454.46d0ca29@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
        <2d0ab5290e63069f310987a4423ef2a46f02f1b3.camel@mediatek.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Mon, 14 Mar 2022 15:01:23 +0800 Biao Huang wrote:
> > Drivers are expected to stop their queues at the end of xmit routine
> > if
> > the ring can't accommodate another frame. It's more efficient to stop
> > the queues early than have to put skbs already dequeued from the
> > qdisc
> > layer back into the qdiscs.  
> Yes, if descriptors ring is full, it's meaningful to stop the queue 
> at the end of xmit; 
> But driver seems hard to know how many descriptors the next skb will
> request, e.g. 3 descriptors are available for next round send, but the
> next skb may need 4 descriptors, in this case, we still need judge
> whether descriptors are enough for skb transmission, then decide stop
> the queue or not, at the beginning of xmit routine.
> 
> Maybe we should judge ring is full or not at the beginning and the end
> of xmit routine(seems a little redundancy).

Assume the worst case scenario. You set the default ring size to 512,
skb can have at most MAX_SKB_FRAGS fragments (usually 17) so the max
number of descriptors should not be very high, hard to pre-compute,
or problematic compared to the total ring size.