* Eric Dumazet | 2010-07-03 08:23:25 [+0200]:

quoted text
>Sebastian
Eric,

quoted text
>Most modern NICS are multiqueue, so that each cpu can use a queue on its
>own without slowing down other cpus.
True. gianfar uses multiqueue howver in its napi poll one CPU grabs a
locks and inspects every queue one by one.

quoted text
>Yet rx recycling has one queue per device, defeating part of the
>multiqueue goal.
Well, we could add rx recycling per queue if you prefer it that way.

quoted text
>Patch 6/8 even touches dev->refcnt on each emerg packet
>Patch 6/8 adds 8 bytes (emerg_dev) to skb. Oh well...
>
>Adding cache layers, especially dumb ones like this one, is probably the
>sign something more fundamental is broken somewhere.

The memory for the skbs is comming from slab. slab's latency is getting
quite high on memory pressure. Therefore I try to avoid talking to slab
while I'm allocating a new skb. So I created a pool of pre-allocated.

quoted text
>I do believe for example that netdev_alloc_skb() should not try to use
>the node affinity of the device, but use current cpu node for sk_buff at
>least, and possibly for data part too.
>
>One other problem of skb are the two memory blocs involved, and fact
>that first one (skb) is already very big and fat, and filled/dirtied
>many cycles before its use in RX path.
>
>Maybe its time to provide new API, so that a driver can build an skb at
>the time RX interrupt is handled, not at the time the rx ring buffer is
>renewed. RX ring should only provide the data part to NIC, and skb
>should be built when NIC delivers the frame, so that we provide to IP
>stack a real hot skb.

And where are you allocating the memory from? Remember, SLAB could be
slow :)

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