On 12/28/2012 02:25 PM, Stephen Hemminger wrote:

On Fri, 28 Dec 2012 13:43:54 +0800
Jason Wang [off-list ref] wrote:

quoted

On 12/28/2012 08:41 AM, Stephen Hemminger wrote:

quoted

On Fri, 21 Dec 2012 12:26:56 +0800
Jason Wang [off-list ref] wrote:

quoted

On 12/21/2012 11:39 AM, Eric Dumazet wrote:

quoted

On Fri, 2012-12-21 at 11:32 +0800, Jason Wang wrote:

quoted

On 12/21/2012 07:50 AM, Stephen Hemminger wrote:

quoted

On Thu, 20 Dec 2012 15:38:17 -0800
Eric Dumazet [off-list ref] wrote:

quoted

On Thu, 2012-12-20 at 18:16 -0500, Paul Moore wrote:

quoted

[CC'ing netdev in case this is a known problem I just missed ...]

Hi Jason,

I started doing some more testing with the multiqueue TUN changes and I ran 
into a problem when running tunctl: running it once w/o arguments works as 
expected, but running it a second time results in failure and a 
kmem_cache_sanity_check() failure.  The problem appears to be very repeatable 
on my test VM and happens independent of the LSM/SELinux fixup patches.

Have you seen this before?

Obviously code in tun_flow_init() is wrong...

static int tun_flow_init(struct tun_struct *tun)
{
        int i;

        tun->flow_cache = kmem_cache_create("tun_flow_cache",
                                            sizeof(struct tun_flow_entry), 0, 0,
                                            NULL);
        if (!tun->flow_cache)
                return -ENOMEM;
...
}


I have no idea why we would need a kmem_cache per tun_struct,
and why we even need a kmem_cache.

Normally flow malloc/free should be good enough.
It might make sense to use private kmem_cache if doing hlist_nulls.


Acked-by: Stephen Hemminger <redacted>

Should be at least a global cache, I thought I can get some speed-up by
using kmem_cache.

Acked-by: Jason Wang <jasowang@redhat.com>

Was it with SLUB or SLAB ?

Using generic kmalloc-64 is better than a dedicated kmem_cache of 48
bytes per object, as we guarantee each object is on a single cache line.

Right, thanks for the explanation.

I wonder if TUN would be better if it used a array to translate
receive hash to receive queue. This is how real hardware works with the
indirection table, and it would allow RFS acceleration. The current flow
cache stuff is prone to DoS attack and scaling problems with lots of
short lived flows.

The problem of indirection table is hash collision which may even happen
when few flows existed.

Hash collision is fine, as long as the the statistical average of
hash across queue's is approximately equal it will be faster. A simple
array indirection is much faster than walking a hash table.

True, but hash collision may cause some negative effects such as losing
the flow affinity and packet re-ordering in guest which does not exist
in a perfect filter. Maybe we can implement them both and let user to
choose.

quoted

For the RFS, we can open a API/ioctl for userspace to add or remove a
flow cache.

RFS acceleration relies on programming the table. It is easier if
TUN looks more like hardware.

quoted

For the DoS/scaling issue, I have an idea of:
- limit the total number of flow entries in tun/tap
- only update the flow entry every N (say 20 like ixgbe) packets or the
the tcp packet has sync flag
- I'm not sure skb_get_rxhash() is lightweight enough, or change to more
lightweight one?

Ideally the hash should be programmable L2 vs L3, but that is splitting
hairs at this point.

Flow tables are scaling problem, especially on highly loaded servers where
they are most needed.

--
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