> > Joakim Tjernlund wrote: > > >>OK > >> > >>You asked for it ;). > >> > >>The file netvia_tsa.c is the actual QMC + stuff driver. > >>The file netvia_bri.c is an example driver. > >>The header files are the QMC definitions. > >> > >>Linux kernel version is 2.4.20. > >> > >>Don't hesitate to ask for any explanations. > >> > >>Pantelis > >> > >> > > > >Thanks a lot! > > > >I have just browsed the driver and I noticed that there is room for > >performance > >improvements in netvia_tsa.c w.r.t memcpy(). You can avoid the memcpy() call > >for received packets by using invalidate_dcache_range(). I did this for > >the 8xx_io/enet.c driver a few weeks ago(it's not in BK yet). > > > >See http://lists.linuxppc.org/linuxppc-embedded/200210/msg00317.html > >and > >http://lists.linuxppc.org/linuxppc-embedded/200210/msg00272.html > > > > Jocke > > > Yes, I know the trick. I'm looking forward to see your changes > make it in. As it stands I'm not tracking the BK PPC tree, but > the Linus tree. Being behind the company firewall, BK is not > very useful.
The patch should apply to linus tree as well. There is just a small bug in my second version(the invalidate_dcache_range() call must be moved). > > In this driver case since I'm reading the packet in a non-cachable > area, and not directly in a skb, I'm not convinced that > the invalidate_dcache_range will be a great win. It should be a win since the patch removes the need for non-cachable memory and a memcpy()(only add a invalidate_dcache_range call for non cache coherent CPUs like the 8xx). Jocke > > If you do any performance testing I'm very interested in > your results. > > Pantelis > > P.S. Is there anyone in the list interested in this? > Should I clean it up, and submit it? Is QMC supported on other families than 8xx? If so I think it would be a good idea. The 862 even has the QMC built in so there must be a need for this. ** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/
