On 07/26/2013 08:38 AM, Felix Fietkau wrote:
On 2013-07-26 5:22 PM, Ben Greear wrote:
On 07/26/2013 02:56 AM, Felix Fietkau wrote:
On 2013-07-26 10:53 AM, Johannes Berg wrote:
On Thu, 2013-07-11 at 08:29 -0700, Ben Greear wrote:
I also don't like maintaining two separate hash tables and all that.
I'd reconsider if you actually remove the hash entirely, but that'll be
tricky to walk the station list and will quite possibly make the RX path
there more expensive?
Remove local->sta_hash ?
To be honest, I'm undecided. Yes, I was thinking that, but I also think
having a huge hashtable like that for each virtual interface is way
overkill, in particular for station interfaces that usually have one
peer (the AP) and maybe a few TDLS peers. Or P2P-Device interfaces that
have no peers at all ...
I don't see a good way to improve the hash either, since we don't always
(e.g. in RX path) have the interface address.
How about mixing in the interface address into the hash. Theoretically
you should always have that available, even in the rx path. Multicast
data packets contain the BSSID, so you can get the address from there.
You just need to be careful about checking the DS bits to figure out
which address to use ;)
I think this is a much better solution than duplicating the hash, or
moving it into sdata entirely.
I think I could probably get rid of the big global per wiphy hash and
use the per-wiphy sdata-hash and per-sdata station hashes.
To me, that is cleanest because it gives a nice ownership relationship
between wiphy, sdata, and stations.
For what it's worth, my hashing scheme has been working well on highly
loaded APs and Station machines.
The global hash (with added vif-addr mixing) not only completely fixes
the many-STA-vif case, also has some other advantages compared to the
per-sdata hash:
- Lookup is easier in setups with multiple AP VLANs
- Better cache footprint (especially important for small embedded devices).
- You don't need a separate sdata lookup before the sta lookup.
I'm not convinced that keeping separate hashes is cleaner. Especially in
the AP_VLAN case, ownership is not clear in any way, since there's some
overlap between multiple sdata entities (belonging to the same BSS).
If someone wants to post such a patch, we can run it through our test
rigs, but I have little time or interest for re-doing the
hashing code again at this time. If your approach does fix the performance
issues we saw, then I'll be more than happy to drop my patch and use
your method.
Thanks,
Ben
--
Ben Greear <greearb@xxxxxxxxxxxxxxx>
Candela Technologies Inc http://www.candelatech.com
--
To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
