On Tue, Jul 23, 2013 at 2:10 PM, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote: > On Mon, 2013-07-22 at 20:47 +0100, Ben Hutchings wrote: >> On Mon, 2013-07-22 at 11:47 -0700, Eric Dumazet wrote: >> > On Mon, 2013-07-22 at 19:38 +0100, Ben Hutchings wrote: >> > > On Mon, 2013-07-22 at 11:29 -0700, Grant Grundler wrote: >> > > > On Mon, Jul 22, 2013 at 10:07 AM, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote: >> > > > ... >> > > > > I guess that if a driver does not advertise NETIF_F_SG, this >> > > > > skb_linearize() call is not needed : All frames reaching your xmit >> > > > > function should already be linear >> > > > >> > > > As Ben Hutchings pointed out, hw_features is still setting this...but >> > > > I'm not sure how that matters. >> > > > >> > > > ax88179_set_features() doesn't allow setting SG or TSO features. But >> > > > I expect it would be "not too difficult" to add such that ethtool >> > > > could set those features after boot. >> > > [...] >> > > >> > > It already can. That's what putting feature flags in hw_features does. >> > >> > My original concern, that inspired this patch, was to remove SG support, >> > as this driver does not have SG support at all. >> > >> > Linearize a full TSO packet needs order-5 allocations, thats likely to >> > fail and lead to very slow TCP performance, because it will only rely on >> > retransmits. >> >> The driver could set gso_max_size to reduce that problem. But I rather >> doubt that TSO followed by skb_linearize() significantly improves >> throughput or CPU-efficiency. (If the device has a 1G link but is >> connected to the host through a USB 2.0 port, then USB is the bottleneck >> and TSO could improve throughput a few percent. But that's a silly >> configuration.) >> >> The real solution would be for someone to add SG support to the usbnet >> core. Trying to support 1GbE with only linear skbs is not a great >> idea... and it can only be a matter of time before there is USB ultra >> speed (or whatever comes after 'super') with 10GbE devices... >> > > This sounds a good idea. > > Is anybody working on adding SG to usbnet ? It depends if size of sg buffer(except for last one) in the sg list can be divided by usb endpoint's max packet size(512 or 1024), at least there is the constraint: http://git.kernel.org/cgit/linux/kernel/git/gregkh/usb.git/commit/?h=usb-next&id=10e232c597ac757e7f8600649f7e872e86de190f I am wondering if network stack can meet that. If not, it might be a bit difficult because lots of USB host controller don't support that, and driver may have to support SG and non-SG at the same time for working well on all HCs. Thanks, -- Ming Lei -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
