linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: "Daniel Ritz" <daniel.ritz@gmx.ch>
To: "Javier Achirica" <achirica@telefonica.net>
Cc: "Jeff Garzik" <jgarzik@pobox.com>,
	"linux-kernel" <linux-kernel@vger.kernel.org>,
	"linux-net" <linux-net@vger.kernel.org>,
	"Jean Tourrilhes" <jt@bougret.hpl.hp.com>,
	"Mike Kershaw" <dragorn@melchior.nerv-un.net>
Subject: Re: [PATCH 2.5] fixes for airo.c
Date: Wed, 23 Jul 2003 11:36:31 +0200	[thread overview]
Message-ID: <000d01c350fd$e625c9d0$a61fc682@alcatel.ch> (raw)
In-Reply-To: Pine.SOL.4.30.0307221014290.3338-100000@tudela.mad.ttd.net

ok, now the braindamaged thing called sourceforge showed the changes, but:
- i don't think the race is fixed. just remove the whole down_trylock()
  crap in the xmit altogether and replace it with a single down(). faster,
  simpler, not racy...and with the schedule_work you win nothing, you lose
speed
- please don't commit bugfixes and new features in the same changeset...
- the loop-forever fix in transmit_allocate: you should have copied the
comment
  changes from my patch too, so the spin-forever-comment goes away...

i look closer when i'm home, having a real operating system to work on, not
this
winblows box at work now..

-daniel


Javier Achirica wrote:
>
> Today I updated the CVS and Sourceforge (airo-linux.sf.net) with the
> latest version (1.53) that (I hope) fixes the race problem. If everything
> is fine, I'll commit the changes to the kernel tree.
>
> Javier Achirica
>
> On Mon, 21 Jul 2003, Daniel Ritz wrote:
>
> > On Mon July 21 2003 21:44, Javier Achirica wrote:
> > >
> > > On Mon, 21 Jul 2003, Daniel Ritz wrote:
> > >
> > > > On Mon July 21 2003 13:00, Javier Achirica wrote:
> > > > >
> > > > > Daniel,
> > > > >
> > > > > Thank you for your patch. Some comments about it:
> > > > >
> > > > > - I'd rather fix whatever is broken in the current code than going
back to
> > > > > spinlocks, as they increase latency and reduce concurrency. In any
case,
> > > > > please check your code. I've seen a spinlock in the interrupt
handler that
> > > > > may lock the system.
> > > >
> > > > but we need to protect from interrupts while accessing the card and
waiting for
> > > > completion. semaphores don't protect you from that.
spin_lock_irqsave does. the
> > > > spin_lock in the interrupt handler is there to protect from
interrupts from
> > > > other processors in a SMP system (see Documentation/spinlocks.txt)
and is btw.
> > > > a no-op on UP. and semaphores are quite heavy....
> > >
> > > Not really. You can still read the received packets from the card (as
> > > you're not issuing any command and are using the other BAP) while a
> > > command is in progress. There are some specific cases in which you
need
> > > to have protection, and that cases are avoided with the down_trylock.
> > >
> >
> > ok, i think i have to look closer...if the card can handle that then we
don't need
> > to irq-protect all the areas i did protect...but i do think that those
down_trylock and
> > then the schedule_work should be replaced by a simple
spinlock_irq_save...
> >
> > i look closer at it tomorrow.
> > you happen to have the tech spec lying aroung?
> >
> > > AFAIK, interrupt serialization is assured by the interrupt handler, so
you
> > > don't need to do that.
> > >
> > > > > - The fix for the transmit code you mention, is about fixing the
returned
> > > > > value in case of error? If not, please explain it to me as I don't
see any
> > > > > other changes.
> > > >
> > > > fixes:
> > > > - return values
> > > > - when to free the skb, when not
> > > > - disabling the queues
> > > > - netif_wake_queue called from the interrupt handler only (and on
the right
> > > >   net_device)
> > > > - i think the priv->xmit stuff and then the schedule_work is evil:
> > > >   if you return 0 from the dev->hard_start_xmit then the network
layer assumes
> > > >   that the packet was kfree_skb()'ed (which does only frees the
packet when the
> > > >   refcount drops to zero.) this is the cause for the keventd
killing, for sure!
> > > >
> > > >   if you return 0 you already kfree_skb()'ed the packet. and that's
it.
> > >
> > > This is where I have the biggest problems. As I've read in
> > > Documentation/networking/driver.txt, looks like the packet needs to be
> > > freed "soon", but doesn't require to be before returning 0 in
> > > hard_start_xmit. Did I get it wrong?
> > >
> >
> > no, i got it wrong. but still...it's the xmit where the oops comes
from....
> >
> > wait. isn't there a race in airo_do_xmit? at high xfer rates (when it
oopses) the
> > queue can wake right after it is stopped in the down_trylock section. so
you can
> > happen to loose an skb 'cos the write to priv->xmit is not protected at
all and
> > there should be a check so that only one skb can be queue there. no?
> > (and then the irq-handler can wake the queue too)
> >
> > ok, i think i got it now. i'll do a new patch tomorrow or so that tries:
> > - to fix the transmit not to oops
> > - to avoid disabling the irq's whenever possible
> > - using spinlocks instead of the heavier semaphores ('cos i think if
it's done cleaner
> >   than i did it now, it's faster than the semas, and to make hch happy
:)
> >
> >
> > > Thanks for your help,
> > > Javier Achirica
> > >
> >
> > rgds
> > -daniel
> >
> >
> >
>


  reply	other threads:[~2003-07-23  9:22 UTC|newest]

Thread overview: 22+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2003-07-17 22:15 [PATCH 2.4] fixes for airo.c Daniel Ritz
2003-07-21 11:00 ` [PATCH 2.5] " Javier Achirica
2003-07-21 12:37   ` Christoph Hellwig
2003-07-21 13:46     ` Javier Achirica
2003-07-21 15:08       ` Mike Kershaw
2003-07-21 18:56         ` Javier Achirica
2003-07-21 17:49   ` Daniel Ritz
2003-07-21 19:44     ` Javier Achirica
2003-07-21 21:01       ` Daniel Ritz
2003-07-21 21:24         ` Javier Achirica
2003-07-22  8:15         ` Javier Achirica
2003-07-23  9:36           ` Daniel Ritz [this message]
2003-07-23 10:26             ` Javier Achirica
2003-07-23 17:56               ` Daniel Ritz
2003-07-23 18:03                 ` Alan Cox
2003-07-23 18:20                   ` Javier Achirica
2003-07-23 18:10                 ` Javier Achirica
2003-07-23 18:20                   ` Alan Cox
2003-07-23 18:52                   ` Daniel Ritz
2003-07-23 20:43                 ` Jeff Garzik
2003-07-23 21:19                   ` Daniel Ritz
2003-07-24 17:07                     ` Jeff Garzik

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='000d01c350fd$e625c9d0$a61fc682@alcatel.ch' \
    --to=daniel.ritz@gmx.ch \
    --cc=achirica@telefonica.net \
    --cc=dragorn@melchior.nerv-un.net \
    --cc=jgarzik@pobox.com \
    --cc=jt@bougret.hpl.hp.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-net@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).