On Fri, 6 Jul 2018 17:53:18 -0700, Alexei Starovoitov wrote: > On Fri, Jul 06, 2018 at 05:08:47PM -0700, Jakub Kicinski wrote: > > On Fri, 6 Jul 2018 16:42:51 -0700, Alexei Starovoitov wrote: > > > On Fri, Jul 06, 2018 at 02:33:58PM -0700, Jakub Kicinski wrote: > > > > On Fri, 6 Jul 2018 09:30:42 -0700, Alexei Starovoitov wrote: > > > > > On Thu, Jul 05, 2018 at 10:18:23AM -0700, Jakub Kicinski wrote: > > > > > > > > > > > > I'm also not 100% on board with the argument that "future" FW can > > > > > > reshuffle things whatever way it wants to. Is the assumption that > > > > > > future ASICs/FW will be designed to always use the "blessed" BTF > > > > > > format? Or will it be reconfigurable at runtime? > > > > > > > > > > let's table configuration of metadata aside for a second. > > > > > > > > > > Describing metedata layout in BTF allows NICs to disclose everything > > > > > NIC has to users in a standard and generic way. > > > > > Whether firmware is reconfigurable on the fly or has to reflashed > > > > > and hw powercycled to have new md layout (and corresponding BTF > > > > > description) > > > > > is a separate discussion. > > > > > Saeed's proposal introduces the concept of 'offset' inside 'struct > > > > > xdp_md_info' > > > > > to reach 'hash' value in metadata. > > > > > Essentially it's a run-time way to access 'hash' instead of > > > > > build-time. > > > > > So bpf program would need two loads to read csum or hash field > > > > > instead of one. > > > > > With BTF the layout of metadata is known to the program at build-time. > > > > > > > > > > To reiterate the proposal: > > > > > - driver+firmware keep layout of the metadata in BTF format (either > > > > > in the driver > > > > > or driver can read it from firmware) > > > > > - 'bpftool read-metadata-desc eth0 > md_desc.h' command will query > > > > > the driver and > > > > > generate normal C header file based on BTF in the given NIC > > > > > - user does #include "md_desc.h" and bpf program can access md->csum > > > > > or md->hash > > > > > with direct single load out of metadata area in front of the packet > > > > > - llvm compiles bpf program and records how program is doing this > > > > > md->csum accesses > > > > > in BTF format as well (the compiler will be keeping such records > > > > > for __sk_buff and all other structs too, but that's separate > > > > > discussion) > > > > > - during sys_bpf(prog_load) the kernel checks (via supplied BTF) that > > > > > the way the program > > > > > accesses metadata (and other structs) matches BTF from the driver, > > > > > so no surprises if driver+firmware got updated, but program is not > > > > > recompiled > > > > > - every NIC can have their own layout of metadata and its own meaning > > > > > of the fields, > > > > > but would be good to standardize at least a few common fields like > > > > > hash > > > > > > > > Can I expose HW descriptors this way, though, or is the proposal to > > > > copy this data into the packet buffer? > > > > > > That crossed my mind too. We can use BTF to describe HW descriptors too, > > > but I don't think it would buy us anything. AF_XDP approach is better. > > > > > > > > Once this is working we can do more cool things with BTF. > > > > > Like doing offset rewriting at program load time similar to what we > > > > > plan > > > > > to do for tracing. Tracing programs will be doing 'task->pid' access > > > > > and the kernel will adjust offsetof(struct task_struct, pid) during > > > > > program load > > > > > depending on BTF for the kernel. > > > > > The same trick we can do for networking metadata. > > > > > The program will contain load instruction md->hash that will get > > > > > automatically > > > > > adjusted to proper offset depending on BTF of 'hash' field in the NIC. > > > > > For now I'm proposing _not_ to go that far with offset rewriting and > > > > > start > > > > > with simple steps described above. > > > > > > > > Why? :( Could we please go with the rewrite/driver helpers instead of > > > > impacting fast paths of the drivers yet again? This rewrite should be > > > > easier than task->pid, because we have the synthetic user space struct > > > > xdp_md definition. > > > > > > I don't understand 'impact fast path yet again' concern. > > > If NIC has certain metadata today, just derscribe what it has in BTF > > > and supply the actual per-packet md to xdp program as-is. > > > No changes for fast path at all. > > > Future rewritting will be done by the bpf/xdp core with zero > > > changes to the driver. All driver has to do is to provide BTF. > > > > I'm confused. AFAIK most *existing* NICs have the metadata in the > > "descriptor", i.e. not in the packet buffer. So if the NIC just > > describes what it has, and there is no data shuffling/copying > > (performance) then we have to expose the descriptor, no? > > which piece of sw put that data into desciptor ? > I bet it's firmware. It could have stored it into pre-packet data, no? > I'd like to avoid _all_ copies. > Right now xdp program can only see a pointer to packet and pre-packet. > If we need another pointer to a piece of the packet descriptor, > that's also fine. Both pre-packet metadata and pieces of descriptor > can be described in BTF.
Okay, if we expose another pointer then it's possible to avoid copies. But please keep in mind that descriptors are very compact, there is a lot of interdependencies between fields and the fields can shift depending on the type of packet. HW/FW guys always quote the 64B packet performance as a reason why things can't be simple. We can't consume 50% of PCIe bandwidth to DMA the metadata alone. > I'd like to push back on firmware folks that should be listening > to feedback from driver folks and kernel stack instead of saying > 'here is hw spec that firmware provides'. Firmware is software. > It can change and should be open to change by the community > with proper maintainership. I think the way to influence FW/HW is to provide a strong and well justified standard, and if we just expose raw HW data we are doing just the opposite. We can claim BTF driver provides is not uABI, but I will personally no longer feel comfortable with modifying descriptor formats (and I've done that for the NFP in the past - grep for chained_metadata_format).
