Tue, Jul 19, 2011 at 10:02:18PM CEST, eric.dumazet@gmail.com wrote:

Le mardi 19 juillet 2011 à 15:52 -0400, Neil Horman a écrit :

quoted

This oops was reported recently when running a pktgen script that called for a
transmitted skb to be cloned and sent 100 times using the clone_skb command to a
bridge device with several attached interface:

...

quoted

Turns out the pktgen driver doesn't actually clone skbs, but rather shares them,
increasing the reference count of the skb for each send operation.  This works
for most drivers because most drivers don't store or care about any state in the
skb itself, but several do.  For instance, the above tun/tap driver and other
soft drivers (vlans, bonding/bridging), all requeue frames to a physical device,
meaning the skb next and prev pointers will be set.  Other drivers also care
about skb state.  The virtio_net driver for instance uses the skb->cb space to
store a vnet header and several converged adapters adjust the data pointer of an
skb to prepend a device control header to the skb.  Drivers expect skbs
submitted for i/o to be in their control and unshared with other users, an
assumption which pktgen is violating, the result being multiple skb users
corrupting one antohers state and producing oopses like the one above.  The
solution is to make pktgen clone the skb for each transmit so as to ensure the
drivers assumptions about private exclusive access to the skb is maintained.

Tested successfully by myself
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Reported-by: Jiri Pirko <redacted>
CC: Eric Dumazet <redacted>
CC: Alexey Dobriyan <redacted>
CC: David S. Miller <davem@davemloft.net>
---

This will kill pktgen performance ?

pktgen is only for sysadmins, and very skilled ones :)

You are right, but it may not cause panic, right? In case this patch
would cause significant performance regression, how about to just forbid
pktgen to run on soft-netdevs ?

Jirka

BTW you forgot to CC pktgen author, Robert Olsson



--
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