On Wed, Sep 10, 2025 at 2:04 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote: > > On Wed, 10 Sep 2025 11:17:52 -0400 Amery Hung wrote: > > > Larger note: I wonder if we should support "shifting the buffer down" > > > if there's insufficient tailroom. XDP has rather copious headroom, > > > but tailroom may be pretty tight, and it may depend on the length of > > > the headers. So if there's not enough tailroom but there's enough > > > headroom -- should we try to memmove the existing headers? > > > > I think it should. If users want to reserve space for metadata, they > > can check the headroom before pulling data. > > > > If the kfunc does not do memmove(), users are still able to do so in > > XDP programs through bpf_xdp_adjust_head() and memmove(), but it feels > > less easy to use IMO. > > Actually, I don't think adjust_head() would even work. The program can > adjust head and memmove() the header, but there's no way to "punch out" > the end of the head buffer. We can only grow and shrink start of packet > and end of packet. After adjust_head + memmove in the prog buffer would > look something like: Ahh. You are right. > > _ _ _ _ __________ _____ _ _ _ _ ________ > hroom | headers | old | troom | frag0 | > - - - - ---------- ----- - - - - -------- > > and the program has no way to "free" the "old" to let pull grab data > from frag0 in its place... > > skb pull helper can allocate a completely fresh buffer, but IDK if > drivers are ready to have the head buffer swapped under their feet. > So I think that best we can do is have the pull() helper aromatically > memmove the headers. Agree. Will make the kfunc memmove headers if more spaces are needed.
