[dpdk-dev] [PATCH v1] net/tap: explain how to compile eBPF C file

[Thomas Monjalon]

12/06/2018 15:33, Wiles, Keith:
> 
> > On Jun 12, 2018, at 7:58 AM, Thomas Monjalon <thomas at monjalon.net> wrote:
> > 
> > 12/06/2018 14:36, Wiles, Keith:
> >> 
> >>> On Jun 12, 2018, at 7:26 AM, Thomas Monjalon <thomas at monjalon.net> wrote:
> >>> 
> >>> 11/06/2018 18:35, Wiles, Keith:
> >>>> 
> >>>>> On Jun 11, 2018, at 11:06 AM, Ophir Munk <ophirmu at mellanox.com> wrote:
> >>>>> 
> >>>>> This commit explains how to manually compile the C source file
> >>>>> tap_bpf_program.c into an ELF file using the clang compiler.
> >>>>> The code in tap_bpf_program.c requires definitions found in iproute2
> >>>>> source code. This commit suggests cloning the iproute2 git tree and
> >>>>> include its path in the clang command. It also adds inclusion of file
> >>>>> bpf_api.h (required for eBPF definitions) which is located in iproute2
> >>>>> source tree. For more details refer to TAP documentation.
> >>>>> This commit is related to commits [1] and [2].
> >>>> 
> >>>> Normally I would have suggested that eBPF be disable in the TAP driver as it requires external code and programs, but that ship has sailed.
> >>> 
> >>> The external programs are required only to generate new instructions,
> >>> changing the behaviour of the BPF program.
> >>> Currently, the instructions for RSS behaviour are provided.
> >>> 
> >>>> I would like to see building the tap_bpf_program.o as a target in the Makefile, this way the developer can just run the ‘make bpf_program’ target and it would be simpler and less error prone.
> > 
> > As explained in the documentation, for now there is a dependency on iproute2
> > for the compilation of this BPF program.
> > So we cannot make it as simple as a "make command".
> > Probably that we can rework it to change the dependency.
> > I heard there are some good BPF libraries available now?
> 
> Well the dependence of iproute2 is really no different then requiring say libnuma, they just have to pull the code first to type the ‘make bpf_program’ right?

The iproute2 dependency is different because it is not a library.
The .h file is never packaged.
So we need to download the sources and set -I to this directory.


> If that is the case then a make target make sense to me. If iproute2 is not found then an error, right?


> >>> For this to happen, we need to improve the tools.
> >> 
> >> In what way do we need to improve the tools and which tools are we talking about. Building the .o file below appears to be a simple set of command lines. I have a question in my original email about what tool.
> > 
> > The .o file is only the an intermediate file.
> > The next step (numbered as 5 in this patch) is to extract the section
> > of BPF instructions to be uploaded in the kernel.
> > This step must be done by a "tool". Ophir did it by hacking tc,
> > but it is not upstreamed yet.
> > There could be other ways (possibly easier) to achieve the same result.
> 
> Please change the doc to reflect the tool is not upstreamed yet and the developer needs to figure out how to extract the data from the binary.
> 
> I used objdump -j l3_l4 -s tap_bpf_program.o and got a hex dump of the l3_l4 section
> 
> 0000 bf160000 00000000 61681000 00000000 <Ascii characters>
> ...
> 
> Someone schooled in the art of Python coding should be able to convert that output to a ‘C’ data array. :-)
> 
> > 
> >>> It is a work in progress.
> > 
> > Contributions are welcome.
> > 
> >>> This is a very first step to use Linux BPF with DPDK.
> >>> If there are more interests, we should really streamline its usage
> >>> for all parts of DPDK which runs on top of some kernel code.
> >> 
> >> streamlining other parts of DPDK would be nice, but we are now talking about the tap/eBPF patch.
> 
> Regards,
> Keith
> 
>