> I've did some measurement of the TX path on an embedded system (2.6.23 kernel) > When bridging packets from an ethernet device to wireless there is > loss of 12% in the CPU utilization and equivalent throughput reduction > in data packets that are checked and expanded in > ieee80211_subif_start_xmit function. Yeah, I figured. > The expansion is just due to bigger size of the 80211 header size, yet > the whole packet is reallocated and copied. As we are reaching 11n > rates 200bps and up this starts to be visible. > When header was reserved to the proper side already in the ethernet > driver the overhead was gone. Yet this doesn't seems to be a correct > solution for bridging or forwarding. > For example Iwlwifi HW supports scattered packets this would allow > just reallocate the header, providing data portion is aligned. Right. Not all hardware supports this though, but even when it does I don't see what we can do unless we want to do all this inside mac80211 which I'd rather not. SKBs don't have a way to say "I need N bytes writable headroom but I can do s/g operation for the rest" which is what we'd need. If that was available, we could even work with clones, then we could pull away the ethernet header and build the 802.11/device headers in separate buffers, demoting the old 'skb head' buffer to a data buffer... johannes
Attachment:
signature.asc
Description: This is a digitally signed message part
