From: Olaf Kirch <okir@lst.de>
To: Jarek Poplawski <jarkao2@o2.pl>
Cc: David Miller <davem@davemloft.net>, netdev@vger.kernel.org
Subject: Re: Races in net_rx_action vs netpoll?
Date: Thu, 12 Jul 2007 15:54:32 +0200 [thread overview]
Message-ID: <200707121554.33915.okir@lst.de> (raw)
In-Reply-To: <20070712125949.GC1708@ff.dom.local>
Hi Jarek,
On Thursday 12 July 2007 14:59, Jarek Poplawski wrote:
> > +#ifdef CONFIG_NETPOLL
> > + /* Prevent race with netpoll - yes, this is a kludge.
> > + * But at least it doesn't penalize the non-netpoll
> > + * code path. */
>
> Alas, this can penalize those who have it enabled (e.g. by distro),
> but don't use.
Well, the test_bit is actually cheap; it's not atomic, and has no memory
ordering requirements by all I know. The costly thing is set_bit/clear_bit
in poll_napi; and you only ever get there when you *use* netpoll.
> And it looks like _netif_rx_complete should be a better place,
> at least considering such cards as: 8139too, skge, sungem and
> maybe more (according to 2.6.22).
Why?
> > + set_bit(__LINK_STATE_POLL_LIST_FROZEN, &np->dev->state);
> > npinfo->rx_flags |= NETPOLL_RX_DROP;
>
> I wonder, why this flag cannot be used for this check?
I tried, but it made the patch rather icky. netpoll_info is defined
in netpoll.h, which includes netdevice.h. So you cannot inline the
check, and have to use an out-of-line function instead, along the
lines of
extern int am_i_being_called_by_poll_napi(struct net_device *);
netif_rx_complete(struct net_device *dev)
{
#ifdef CONFIG_NETPOLL
if (unlikely(dev->npinfo && am_i_being_called_by_poll_napi(dev))
return;
#endif
...
}
If you don't mind that, yes - this flag (or better, a newly introduced
NETPOLL_RX_NAPI) may work as well.
One thing I was a little worried about was whether dev->npinfo can
go away all of a sudden. It's really just protected by an rcu_readlock...
> BTW, I'd be very glad if somebody could hint me what is the main
> reason for such troublesome function as poll_napi: if it's about
> performance isn't this enough to increase budget for netpoll in
> net_rx_action?
I think one reason is that you want to get the kernel oops out even
when the machine is so hosed that it doesn't even service softirqs
anymore.
Olaf
--
Olaf Kirch | --- o --- Nous sommes du soleil we love when we play
okir@lst.de | / | \ sol.dhoop.naytheet.ah kin.ir.samse.qurax
next prev parent reply other threads:[~2007-07-12 13:54 UTC|newest]
Thread overview: 14+ messages / expand[flat|nested] mbox.gz Atom feed top
2007-07-04 12:16 Races in net_rx_action vs netpoll? Olaf Kirch
2007-07-09 22:27 ` David Miller
2007-07-10 10:44 ` Olaf Kirch
2007-07-11 5:44 ` David Miller
2007-07-11 7:41 ` Olaf Kirch
2007-07-12 2:33 ` David Miller
2007-07-19 15:19 ` Olaf Kirch
2007-07-19 16:27 ` Stephen Hemminger
2007-07-22 7:05 ` David Miller
2007-07-24 10:26 ` Jarek Poplawski
2007-07-12 12:59 ` Jarek Poplawski
2007-07-12 13:54 ` Olaf Kirch [this message]
2007-07-13 8:55 ` Jarek Poplawski
2007-07-16 8:06 ` Jarek Poplawski
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=200707121554.33915.okir@lst.de \
--to=okir@lst.de \
--cc=davem@davemloft.net \
--cc=jarkao2@o2.pl \
--cc=netdev@vger.kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).