On Tue, Mar 12, 2024 at 6:53 PM 梦龙董 <dongmenglong.8@xxxxxxxxxxxxx> wrote: > > On Wed, Mar 13, 2024 at 12:42 AM Alexei Starovoitov > <alexei.starovoitov@xxxxxxxxx> wrote: > > > > On Mon, Mar 11, 2024 at 7:42 PM 梦龙董 <dongmenglong.8@xxxxxxxxxxxxx> wrote: > > > > [......] > > > > I see. > > I thought you're sharing the trampoline across attachments. > > (since bpf prog is the same). > > That seems to be a good idea, which I hadn't thought before. > > > But above approach cannot possibly work with a shared trampoline. > > You need to create individual trampoline for all attachment > > and point them to single bpf prog. > > > > tbh I'm less excited about this feature now, since sharing > > the prog across different attachments is nice, but it won't scale > > to thousands of attachments. > > I assumed that there will be a single trampoline with max(argno) > > across attachments and attach/detach will scale to thousands. > > > > With individual trampoline this will work for up to a hundred > > attachments max. > > What does "a hundred attachments max" means? Can't I > trace thousands of kernel functions with a bpf program of > tracing multi-link? I mean what time does it take to attach one program to 100 fentry-s ? What is the time for 1k and for 10k ? The kprobe multi test attaches to pretty much all funcs in /sys/kernel/tracing/available_filter_functions and it's fast enough to run in test_progs on every commit in bpf CI. See get_syms() in prog_tests/kprobe_multi_test.c Can this new multi fentry do that? and at what speed? The answer will decide how applicable this api is going to be. Generating different trampolines for every attach point is an approach as well. Pls benchmark it too. > > > > Let's step back. > > What is the exact use case you're trying to solve? > > Not an artificial one as selftest in patch 9, but the real use case? > > I have a tool, which is used to diagnose network problems, > and its name is "nettrace". It will trace many kernel functions, whose > function args contain "skb", like this: > > ./nettrace -p icmp > begin trace... > ***************** ffff889be8fbd500,ffff889be8fbcd00 *************** > [1272349.614564] [dev_gro_receive ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614579] [__netif_receive_skb_core] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614585] [ip_rcv ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614592] [ip_rcv_core ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614599] [skb_clone ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614616] [nf_hook_slow ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614629] [nft_do_chain ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614635] [ip_rcv_finish ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614643] [ip_route_input_slow ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614647] [fib_validate_source ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614652] [ip_local_deliver ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614658] [nf_hook_slow ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614663] [ip_local_deliver_finish] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614666] [icmp_rcv ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614671] [icmp_echo ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614675] [icmp_reply ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614715] [consume_skb ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614722] [packet_rcv ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > [1272349.614725] [consume_skb ] ICMP: 169.254.128.15 -> > 172.27.0.6 ping request, seq: 48220 > > For now, I have to create a bpf program for every kernel > function that I want to trace, which is up to 200. > > With this multi-link, I only need to create 5 bpf program, > like this: > > int BPF_PROG(trace_skb_1, struct *skb); > int BPF_PROG(trace_skb_2, u64 arg0, struct *skb); > int BPF_PROG(trace_skb_3, u64 arg0, u64 arg1, struct *skb); > int BPF_PROG(trace_skb_4, u64 arg0, u64 arg1, u64 arg2, struct *skb); > int BPF_PROG(trace_skb_5, u64 arg0, u64 arg1, u64 arg2, u64 arg3, struct *skb); > > Then, I can attach trace_skb_1 to all the kernel functions that > I want to trace and whose first arg is skb; attach trace_skb_2 to kernel > functions whose 2nd arg is skb, etc. > > Or, I can create only one bpf program and store the index > of skb to the attachment cookie, and attach this program to all > the kernel functions that I want to trace. > > This is my use case. With the multi-link, now I only have > 1 bpf program, 1 bpf link, 200 trampolines, instead of 200 > bpf programs, 200 bpf link and 200 trampolines. I see. The use case makes sense to me. Andrii's retsnoop is used to do similar thing before kprobe multi was introduced. > The shared trampoline you mentioned seems to be a > wonderful idea, which can make the 200 trampolines > to one. Let me have a look, we create a trampoline and > record the max args count of all the target functions, let's > mark it as arg_count. > > During generating the trampoline, we assume that the > function args count is arg_count. During attaching, we > check the consistency of all the target functions, just like > what we do now. For one trampoline to handle all attach points we might need some arch support, but we can start simple. Make btf_func_model with MAX_BPF_FUNC_REG_ARGS by calling btf_distill_func_proto() with func==NULL. And use that to build a trampoline. The challenge is how to use minimal number of trampolines when bpf_progA is attached for func1, func2, func3 and bpf_progB is attached to func3, func4, func5. We'd still need 3 trampolines: for func[12] to call bpf_progA, for func3 to call bpf_progA and bpf_progB, for func[45] to call bpf_progB. Jiri was trying to solve it in the past. His slides from LPC: https://lpc.events/event/16/contributions/1350/attachments/1033/1983/plumbers.pdf Pls study them and his prior patchsets to avoid stepping on the same rakes.
