netdev - Re: [PATCH] xen network backend driver

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:	Wed, 19 Jan 2011 18:05:04 +0000
From:	Ben Hutchings <bhutchings@...arflare.com>
To:	Ian Campbell <Ian.Campbell@...citrix.com>
Cc:	"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
	xen-devel <xen-devel@...ts.xensource.com>,
	Jeremy Fitzhardinge <jeremy@...p.org>,
	Konrad Rzeszutek Wilk <konrad.wilk@...cle.com>
Subject: Re: [PATCH] xen network backend driver

On Wed, 2011-01-19 at 17:48 +0000, Ian Campbell wrote:
> Hi Ben,
> 
> Thanks for the very speedy review!
> 
> I don't have many comments other than "yes, you are right".
> 
> There are a couple of things inline below.
> 
> On Wed, 2011-01-19 at 16:40 +0000, Ben Hutchings wrote:
> > On Wed, 2011-01-19 at 15:01 +0000, Ian Campbell wrote:
> > [...]
> > > +	/*
> > > +	 * Initialise a dummy MAC address. We choose the numerically
> > > +	 * largest non-broadcast address to prevent the address getting
> > > +	 * stolen by an Ethernet bridge for STP purposes.
> > > +	 * (FE:FF:FF:FF:FF:FF)
> > > +	 */
> > > +	memset(dev->dev_addr, 0xFF, ETH_ALEN);
> > > +	dev->dev_addr[0] &= ~0x01;
> > 
> > I'm a bit dubious about this.
> 
> Which reminds me that I need to add the hook so that the Xen userspace
> stuff can actually do the right thing and set the MAC address to
> FE:FF:FF:FF:FF:FF itself as it puts the device on the bridge.
> 
> The toolstack has only recently been fixed to even try that though.
> 
> In use these devices aren't typically endpoints which generate or
> receive any actual traffic so letting it pick up a random MAC address by
> default isn't terribly useful. The actual useful MAC address is the one
> which is configured in the frontend.

Right, I understand that.

> > [...]
> > > +static int MODPARM_netback_kthread;
> > > +module_param_named(netback_kthread, MODPARM_netback_kthread, bool, 0);
> > > +MODULE_PARM_DESC(netback_kthread, "Use kernel thread to replace tasklet");
> > > +
> > > +/*
> > > + * Netback bottom half handler.
> > > + * dir indicates the data direction.
> > > + * rx: 1, tx: 0.
> > > + */
> > > +static inline void xen_netbk_bh_handler(struct xen_netbk *netbk, int dir)
> > > +{
> > > +	if (MODPARM_netback_kthread)
> > > +		wake_up(&netbk->kthread.netbk_action_wq);
> > > +	else if (dir)
> > > +		tasklet_schedule(&netbk->tasklet.net_rx_tasklet);
> > > +	else
> > > +		tasklet_schedule(&netbk->tasklet.net_tx_tasklet);
> > > +}
> > 
> > Ugh, please just use NAPI.
> 
> Although I only have a vague concept of what NAPI actually entails in
> practice I think it most likely makes sense.
> 
> Am I right that NAPI only covers the RX case?

All completions should be processed via NAPI, if possible.  The poll
function is given a work budget and each RX completion is assigned a
cost of 1.  TX completions are cheap enough that they aren't budgetted
individually, but they must be limited somehow.  The standard practice
is to consider the budget exhausted after processing an entire TX ring
once.

> Does NAPI processing happen in softirq context?

Yes.

> The reason for the
> existing option to use a kthread was that the tasklets would completely
> swamp the domain 0 CPU under load and prevent anything else from running
> (including e.g. ssh or the toolstack allowing you to fix the
> problem...).

I can see that that could be a problem if dom0's processing power is low
compared to the other domains.

> I guess this is just a case of setting the NAPI weight
> correctly (i.e. appropriately high in this case)?

Sorry, I have not looked at adjusting NAPI weights before.

> Last question before I go an actually investigate NAPI properly: Does
> NAPI also scale out across CPUs? Currently the threads/tasklets are per
> CPU and this is a significant scalability win.
[...]

Not in itself.  NAPI polling will run on the same CPU which scheduled it
(so wherever the IRQ was initially handled).  If the protocol used
between netfront and netback doesn't support RSS then RPS
<http://lwn.net/Articles/362339/> can be used to spread the RX work
across CPUs.

Ben.

-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html