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;
Feedback-ID: i787e41f1:Fastmail
Date:   Mon, 19 Dec 2022 08:50:00 +0100
From:   Greg KH <greg@kroah.com>
To:     Leesoo Ahn <lsahn@ooseel.net>
Cc:     Oliver Neukum <oneukum@suse.com>,
        "David S. Miller" <davem@davemloft.net>,
        Eric Dumazet <edumazet@google.com>,
        Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>, netdev@vger.kernel.org,
        linux-usb@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] usbnet: jump to rx_cleanup case instead of calling
 skb_queue_tail
Message-ID: <Y6AXqOlCUy7mahgj@kroah.com>
References: <20221217161851.829497-1-lsahn@ooseel.net>
 <Y57VkLKetDsbUUjC@kroah.com>
 <a2e0e98a-1044-908a-15bc-b165ff8b23ea@ooseel.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <a2e0e98a-1044-908a-15bc-b165ff8b23ea@ooseel.net>
Precedence: bulk

On Mon, Dec 19, 2022 at 04:41:16PM +0900, Leesoo Ahn wrote:
> 
> On 22. 12. 18. 17:55, Greg KH wrote:
> > On Sun, Dec 18, 2022 at 01:18:51AM +0900, Leesoo Ahn wrote:
> > > The current source pushes skb into dev->done queue by calling
> > > skb_queue_tail() and then, call skb_dequeue() to pop for rx_cleanup state
> > > to free urb and skb next in usbnet_bh().
> > > It wastes CPU resource with extra instructions. Instead, use return values
> > > jumping to rx_cleanup case directly to free them. Therefore calling
> > > skb_queue_tail() and skb_dequeue() is not necessary.
> > > 
> > > The follows are just showing difference between calling skb_queue_tail()
> > > and using return values jumping to rx_cleanup state directly in usbnet_bh()
> > > in Arm64 instructions with perf tool.
> > > 
> > > ----------- calling skb_queue_tail() -----------
> > >         │     if (!(dev->driver_info->flags & FLAG_RX_ASSEMBLE))
> > >    7.58 │248:   ldr     x0, [x20, #16]
> > >    2.46 │24c:   ldr     w0, [x0, #8]
> > >    1.64 │250: ↑ tbnz    w0, #14, 16c
> > >         │     dev->net->stats.rx_errors++;
> > >    0.57 │254:   ldr     x1, [x20, #184]
> > >    1.64 │258:   ldr     x0, [x1, #336]
> > >    2.65 │25c:   add     x0, x0, #0x1
> > >         │260:   str     x0, [x1, #336]
> > >         │     skb_queue_tail(&dev->done, skb);
> > >    0.38 │264:   mov     x1, x19
> > >         │268:   mov     x0, x21
> > >    2.27 │26c: → bl      skb_queue_tail
> > >    0.57 │270: ↑ b       44    // branch to call skb_dequeue()
> > > 
> > > ----------- jumping to rx_cleanup state -----------
> > >         │     if (!(dev->driver_info->flags & FLAG_RX_ASSEMBLE))
> > >    1.69 │25c:   ldr     x0, [x21, #16]
> > >    4.78 │260:   ldr     w0, [x0, #8]
> > >    3.28 │264: ↑ tbnz    w0, #14, e4    // jump to 'rx_cleanup' state
> > >         │     dev->net->stats.rx_errors++;
> > >    0.09 │268:   ldr     x1, [x21, #184]
> > >    2.72 │26c:   ldr     x0, [x1, #336]
> > >    3.37 │270:   add     x0, x0, #0x1
> > >    0.09 │274:   str     x0, [x1, #336]
> > >    0.66 │278: ↑ b       e4    // branch to 'rx_cleanup' state
> > Interesting, but does this even really matter given the slow speed of
> > the USB hardware?
> 
> It doesn't if USB hardware has slow speed but in software view, it's still
> worth avoiding calling skb_queue_tail() and skb_dequeue() which work with
> spinlock, if possible.

But can you actually measure that in either CPU load or in increased
transfer speeds?

thanks,

greg k-h