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?
I have an old one, but I don't think that I'm allowed (by Cisco) to pass
it around.
> > 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 :)
Yes! This is the race that may be causing problems. I'll try to fix it and
we may compare semaphore vs. spinlock implementation.
Javier Achirica
-
: send the line "unsubscribe linux-net" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
