On Wed, Jan 8, 2020 at 4:35 PM Martin KaFai Lau <kafai@xxxxxx> wrote:
>
> This series introduces BPF STRUCT_OPS. It is an infra to allow
> implementing some specific kernel's function pointers in BPF.
> The first use case included in this series is to implement
> TCP congestion control algorithm in BPF (i.e. implement
> struct tcp_congestion_ops in BPF).
>
> There has been attempt to move the TCP CC to the user space
> (e.g. CCP in TCP). The common arguments are faster turn around,
> get away from long-tail kernel versions in production...etc,
> which are legit points.
>
> BPF has been the continuous effort to join both kernel and
> userspace upsides together (e.g. XDP to gain the performance
> advantage without bypassing the kernel). The recent BPF
> advancements (in particular BTF-aware verifier, BPF trampoline,
> BPF CO-RE...) made implementing kernel struct ops (e.g. tcp cc)
> possible in BPF.
>
> The idea is to allow implementing tcp_congestion_ops in bpf.
> It allows a faster turnaround for testing algorithm in the
> production while leveraging the existing (and continue growing) BPF
> feature/framework instead of building one specifically for
> userspace TCP CC.
>
> Please see individual patch for details.
>
> The bpftool support will be posted in follow-up patches.
>
> v4:
> - Expose tcp_ca_find() to tcp.h in patch 7.
> It is used to check the same bpf-tcp-cc
> does not exist to guarantee the register()
> will succeed.
> - set_memory_ro() and then set_memory_x() only after all
> trampolines are written to the image in patch 6. (Daniel)
> spinlock is replaced by mutex because set_memory_*
> requires sleepable context.
Applied. Thanks
Please address any future follow up
and please remember to provide 'why' details in commit log
no matter how obvious the patch looks as Arnaldo pointed out.
Re: 'bpftool module' command.
I think it's too early to call anything 'a module',
since in this context people will immediately assume 'a kernel module'
It's a loaded phrase with a lot of consequences.
bpf-tcp-cc cubic and dctcp do look like kernel modules, but they are
not kernel modules at all. imo making them look like kernel modules
has plenty of downsides.
So as an immediate followup for bpftool I'd recommend to stick
with 'bpftool struct_ops' command or whatever other name.
Just not 'bpftool module'.
I think there is a makefile issue with selftests.
make clean;make -j50
kept failing for me three times in a row with:
progs/bpf_dctcp.c:138:4: warning: implicit declaration of function
'bpf_tcp_send_ack' is invalid in C99 [-Wimplicit-function-declaration]
bpf_tcp_send_ack(sk, *prior_rcv_nxt);
but single threaded 'make clean;make' succeeded.
Andrii, you have a chance to take a look and reproduce would be great.
