Re: [PATCH net-next 6/7] net: phy: smsc: Cache interrupt mask

From: Lukas Wunner <lukas@wunner.de>
To: Andrew Lunn <andrew@lunn.ch>
Cc: Steve Glendinning <steve.glendinning@shawell.net>,
	UNGLinuxDriver@microchip.com, Oliver Neukum <oneukum@suse.com>,
	"David S. Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>, Paolo Abeni <pabeni@redhat.com>,
	netdev@vger.kernel.org, linux-usb@vger.kernel.org,
	Andre Edich <andre.edich@microchip.com>,
	Oleksij Rempel <o.rempel@pengutronix.de>,
	Martyn Welch <martyn.welch@collabora.com>,
	Gabriel Hojda <ghojda@yo2urs.ro>,
	Christoph Fritz <chf.fritz@googlemail.com>,
	Lino Sanfilippo <LinoSanfilippo@gmx.de>,
	Philipp Rosenberger <p.rosenberger@kunbus.com>,
	Heiner Kallweit <hkallweit1@gmail.com>,
	Russell King <linux@armlinux.org.uk>
Subject: Re: [PATCH net-next 6/7] net: phy: smsc: Cache interrupt mask
Date: Wed, 27 Apr 2022 14:51:10 +0200	[thread overview]
Message-ID: <20220427125110.GA7941@wunner.de> (raw)
In-Reply-To: <Ymkzno2IbyNbFrEL@lunn.ch>

On Wed, Apr 27, 2022 at 02:14:22PM +0200, Andrew Lunn wrote:
> On Wed, Apr 27, 2022 at 07:48:06AM +0200, Lukas Wunner wrote:
> > Cache the interrupt mask to avoid re-reading it from the PHY upon every
> > interrupt.  The PHY may be located on a USB device, so the additional
> > read may unnecessarily increase interrupt overhead and latency.
> 
> I don't think your justification is valid. The MDIO bus is clocked at
> 2.5MHz. So even if you are using USB 1.1 at 12MHz, the USB overheads
> are not particularly large. At 480Mbps they are pretty insignificant.
> 
> In general, we consider PHYs as slow devices, they take over 1 second
> to negotiate a link and declare it up. So we don't do this sort of
> micro optimization.
> 
> What i think is relevant here is that you could have an interrupt
> storm going on because you don't mask interrupts? It is not a true
> storm, due to the way USB works, more of a light shower. Do you have
> any statistics to show this code actually reduces the amount of rain
> in a significant way?

TBH the primary motivation for this change is that it simplifies the
succeeding commit ("Cope with hot-removal in interrupt handler").

Additionally it seemed silly to me to re-read the interrupt mask
every time for no reason at all.  To test and debug this series
I logged every MDIO read/write and these nonsensical transactions
are very visible and very annoying in the log output.

So yeah, maybe the latency argument isn't very strong, but there
are other arguments which I didn't deem necessary mentioning in
the commit message as they seemed somewhat egotistical. :)

Thanks,

Lukas