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:   Thu, 17 Feb 2022 20:36:04 -0800
From:   Jakub Kicinski <kuba@kernel.org>
To:     Vincent Whitchurch <vincent.whitchurch@axis.com>
Cc:     Giuseppe Cavallaro <peppe.cavallaro@st.com>,
        Alexandre Torgue <alexandre.torgue@foss.st.com>,
        Jose Abreu <joabreu@synopsys.com>,
        "David S. Miller" <davem@davemloft.net>,
        Maxime Coquelin <mcoquelin.stm32@gmail.com>, <kernel@axis.com>,
        Lars Persson <larper@axis.com>,
        Srinivas Kandagatla <srinivas.kandagatla@st.com>,
        <netdev@vger.kernel.org>,
        <linux-stm32@st-md-mailman.stormreply.com>,
        <linux-arm-kernel@lists.infradead.org>,
        <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] net: stmmac: Enable NAPI before interrupts go live
Message-ID: <20220217203604.39e318d0@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
In-Reply-To: <20220217145527.2696444-1-vincent.whitchurch@axis.com>
References: <20220217145527.2696444-1-vincent.whitchurch@axis.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Thu, 17 Feb 2022 15:55:26 +0100 Vincent Whitchurch wrote:
> From: Lars Persson <larper@axis.com>
> 
> The stmmac_open function has a race window between enabling the RX
> path and its interrupt to the point where napi_enabled is called.
> 
> A chatty network with plenty of broadcast/multicast traffic has the
> potential to completely fill the RX ring before the interrupt handler
> is installed. In this scenario the single interrupt taken will find
> napi disabled and the RX ring will not be processed. No further RX
> interrupt will be delivered because the ring is full.
> 
> The RX stall could eventually clear because the TX path will trigger a
> DMA interrupt once the tx_coal_frames threshold is reached and then
> NAPI becomes scheduled.

LGTM, although now the ndo_open and ndo_stop paths are not symmetrical.
Is there no way to mask the IRQs so that they don't fire immediately?
More common flow (IMO) would be:
 - request irq
 - mask irq
 - populate rings
 - start dma
 - enable napi
 - unmask irq
Other than the difference in flow between open/stop there may also be
some unpleasantness around restarting tx queues twice with the patch
as is.