Hi ASPj,

On Fri, Aug 5, 2011 at 12:53 PM, Pedro Larbig <pedro.lar...@carhs.de> wrote:
> It sure is NO leak. kmemleak just pointed me to the right way, that the space
> was eaten by skbufs.
> As soon as several nodes go down, the remaining ones empty not only their
> queues but also the associated memory. So kmemleak goes back to 0 MB for the
> kmalloc stuff.

Noted.

> The hack proposed by 游波 does indeed fix that problem:
> --- a/net/mac80211/mesh_hwmp.c
> +++ b/net/mac80211/mesh_hwmp.c
> @@ -387,8 +387,7 @@ static u32 hwmp_route_info_get(struct
> ieee80211_sub_if_data *sdata,
>                        spin_lock_bh(&mpath->state_lock);
>                        if (mpath->flags & MESH_PATH_FIXED)
>                                fresh_info = false;
> -                       else if ((mpath->flags & MESH_PATH_ACTIVE) &&
> -                           (mpath->flags & MESH_PATH_SN_VALID)) {
> +                       else if (mpath->flags & MESH_PATH_ACTIVE) {
>                                if (SN_GT(mpath->sn, orig_sn) ||
>                                    (mpath->sn == orig_sn &&
>                                     new_metric >= mpath->metric)) {
>
> I am currently running your git version plus the above patch, and for the last
> hour, it just works FINE! Workqueues are empty, and no huge amounts of memory
> allocated. Every node is pinging every other node, so all possible routes are
> established.

Great.  We'll review the fix and push it upstream.

> I'll give a SMP build a try in the next few days to see if the locking issues
> I had are also gone.

We'll be expecting your results.

Thanks a lot for the testbed photos and your offer to test new code.

Cheers,

Javier

> PS: If you want me to test some new code, just drop a line, I can keep stuff
> running for days on our testbed, I can also perform automated iperf, ping,
> sniffing and info tasks (regularly dumping neighbors, routes, dBm, bitrates,
> etc).
>
> Here are some nice pictures:
> Our network running the open11s git in idle state, values are signal strength
> in dBm: http://aspj.aircrack-ng.org/o11s/idle.png
> The same network after the memory consumption bug ate a few nodes (notice the
> greatly reduced link count!): http://aspj.aircrack-ng.org/o11s/crashed.png
> And finally after the patch, this time showing the routes used (arrow's color
> is the routes destination, arrow's target is next hop): http://aspj.aircrack-
> ng.org/o11s/patched_routes.png
> [mndii = "Mesh Node II", nodes 18, 19 and 20 are NOT outside of the building
> but one floor above the others]
> _______________________________________________
> Devel mailing list
> Devel@lists.open80211s.org
> http://open80211s.com/mailman/listinfo/devel
>



-- 
Javier Cardona
cozybit Inc.
http://www.cozybit.com
_______________________________________________
Devel mailing list
Devel@lists.open80211s.org
http://open80211s.com/mailman/listinfo/devel

Reply via email to