For anyone interested, the two cases I mention below are indeed the problems. We've tested two fixes and both work. The first is just allocating extra space. The second is actually implementing ofpbuf_prealloc_headroom() instead of just having it fail which is what it does now.
I like the second approach, but at least my quick implementation has the problem of potentially changing the alignment of the data and I haven't looked through the code enough to know if this might cause trouble (e.g., if there is an expectation somewhere that the IP pointers are aligned). I'll probably push one or both of these to feature branches of the OpenFlow Reference fork on the noxrepo github in case anyone ever wants them. -- Murphy On Jul 7, 2014, at 2:43 AM, Murphy McCauley <murphy.mccau...@gmail.com> wrote: > You should specify what switch you're talking about. As I remember it, this > is a problem with the OpenFlow 1.0 reference implementation's user mode > datapath. There have been other switches based on this code, though. If > you're using one of them, you should specify which -- it may have been fixed > in them, and/or different people may be interested in fixing it. > > If you are indeed talking about the 1.0 reference in user mode... this isn't > widely used anymore, and I don't think anyone ever fixed it. I think I > figured out what the problem was, but I never actually submitted a patch > since even when the bug came to my attention, I had long since moved to Open > vSwitch. Normally, when the switch allocates a buffer, it allocates some > extra headroom so that it can, e.g., add a VLAN tag. But when sending a > packet_in, the switch stores a clone of that buffer, and the clone is exactly > large enough for the current data -- no headroom. So if the controller sends > back a flow_mod or packet_out which refers to the buffer by ID, there isn't > enough room in the clone to actually add a VLAN tag. I think there's a > similar bug for packet_outs where you send the raw data (instead of a buffer > ID). In both cases, it's just a matter of allocating a few extra bytes. > > -- Murphy > > On Jul 6, 2014, at 11:28 PM, scbz...@gmail.com wrote: > >> Is the problem >> ofdatapath: lib/ofpbuf.c:168: ofpbuf_prealloc_headroom: Assertion `size >> <= ofpbuf_headroom(b)' failed. >> solved? >> >> Best Regards >> 杨海祥 >> >> scbz...@gmail.com >> _______________________________________________ >> openflow-discuss mailing list >> openflow-discuss@lists.stanford.edu >> https://mailman.stanford.edu/mailman/listinfo/openflow-discuss >
_______________________________________________ openflow-discuss mailing list openflow-discuss@lists.stanford.edu https://mailman.stanford.edu/mailman/listinfo/openflow-discuss
