| Message ID | 20240311093526.1010158-1-dongmenglong.8@bytedance.com |
|---|---|
| Headers | show |
| Series | bpf: make tracing program support multi-link | expand |
On Tue, Mar 12, 2024 at 9:55 AM Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote: > > On Mon, Mar 11, 2024 at 2:35 AM Menglong Dong > <dongmenglong.8@bytedance.com> wrote: > > > > By default, the kernel btf that we load during loading program will be > > freed after the programs are loaded in bpf_object_load(). However, we > > still need to use these btf for tracing of multi-link during attaching. > > Therefore, we don't free the btfs until the bpf object is closed if any > > bpf programs of the type multi-link tracing exist. > > > > Meanwhile, introduce the new api bpf_object__free_btf() to manually free > > the btfs after attaching. > > > > Signed-off-by: Menglong Dong <dongmenglong.8@bytedance.com> > > --- > > tools/lib/bpf/libbpf.c | 47 ++++++++++++++++++++++++++++++---------- > > tools/lib/bpf/libbpf.h | 2 ++ > > tools/lib/bpf/libbpf.map | 1 + > > 3 files changed, 38 insertions(+), 12 deletions(-) > > > > diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c > > index 567ad367e7aa..fd5428494a7e 100644 > > --- a/tools/lib/bpf/libbpf.c > > +++ b/tools/lib/bpf/libbpf.c > > @@ -8267,6 +8267,39 @@ static int bpf_object_prepare_struct_ops(struct bpf_object *obj) > > return 0; > > } > > > > +void bpf_object__free_btfs(struct bpf_object *obj) > > +{ > > + int i; > > + > > + /* clean up module BTFs */ > > + for (i = 0; i < obj->btf_module_cnt; i++) { > > + close(obj->btf_modules[i].fd); > > + btf__free(obj->btf_modules[i].btf); > > + free(obj->btf_modules[i].name); > > + } > > + free(obj->btf_modules); > > + obj->btf_modules = NULL; > > + obj->btf_module_cnt = 0; > > + > > + /* clean up vmlinux BTF */ > > + btf__free(obj->btf_vmlinux); > > + obj->btf_vmlinux = NULL; > > +} > > + > > +static void bpf_object_early_free_btf(struct bpf_object *obj) > > +{ > > + struct bpf_program *prog; > > + > > + bpf_object__for_each_program(prog, obj) { > > + if (prog->expected_attach_type == BPF_TRACE_FENTRY_MULTI || > > + prog->expected_attach_type == BPF_TRACE_FEXIT_MULTI || > > + prog->expected_attach_type == BPF_MODIFY_RETURN_MULTI) > > + return; > > + } > > + > > + bpf_object__free_btfs(obj); > > +} > > + > > static int bpf_object_load(struct bpf_object *obj, int extra_log_level, const char *target_btf_path) > > { > > int err, i; > > @@ -8307,18 +8340,7 @@ static int bpf_object_load(struct bpf_object *obj, int extra_log_level, const ch > > /* clean up fd_array */ > > zfree(&obj->fd_array); > > > > - /* clean up module BTFs */ > > - for (i = 0; i < obj->btf_module_cnt; i++) { > > - close(obj->btf_modules[i].fd); > > - btf__free(obj->btf_modules[i].btf); > > - free(obj->btf_modules[i].name); > > - } > > - free(obj->btf_modules); > > - > > - /* clean up vmlinux BTF */ > > - btf__free(obj->btf_vmlinux); > > - obj->btf_vmlinux = NULL; > > - > > + bpf_object_early_free_btf(obj); > > obj->loaded = true; /* doesn't matter if successfully or not */ > > > > if (err) > > @@ -8791,6 +8813,7 @@ void bpf_object__close(struct bpf_object *obj) > > usdt_manager_free(obj->usdt_man); > > obj->usdt_man = NULL; > > > > + bpf_object__free_btfs(obj); > > bpf_gen__free(obj->gen_loader); > > bpf_object__elf_finish(obj); > > bpf_object_unload(obj); > > diff --git a/tools/lib/bpf/libbpf.h b/tools/lib/bpf/libbpf.h > > index 5723cbbfcc41..c41a909ea4c1 100644 > > --- a/tools/lib/bpf/libbpf.h > > +++ b/tools/lib/bpf/libbpf.h > > @@ -299,6 +299,8 @@ LIBBPF_API struct bpf_program * > > bpf_object__find_program_by_name(const struct bpf_object *obj, > > const char *name); > > > > +LIBBPF_API void bpf_object__free_btfs(struct bpf_object *obj); > > + > > It shouldn't be exported. > libbpf should clean it up when bpf_object is freed. Yes, libbpf will clean up the btfs when bpf_object is freed in this commit. And I'm trying to offer a way to early free the btfs by the users manual to reduce the memory usage. Or, the btfs that we opened will keep existing until we close the bpf_object. This is optional, I can remove it if you prefer. Thanks! Menglong Dong
On Mon, Mar 11, 2024 at 7:05 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > > > > +LIBBPF_API void bpf_object__free_btfs(struct bpf_object *obj); > > > + > > > > It shouldn't be exported. > > libbpf should clean it up when bpf_object is freed. > > Yes, libbpf will clean up the btfs when bpf_object is freed in > this commit. And I'm trying to offer a way to early free the btfs > by the users manual to reduce the memory usage. Or, the > btfs that we opened will keep existing until we close the > bpf_object. > > This is optional, I can remove it if you prefer. Let's not extend libbpf api unless we really need to. bpf_program__attach_trace_multi_opts() and *skel*__attach() can probably free them. I don't see a use case where you'd want to keep them afterwards.
On Tue, Mar 12, 2024 at 10:13 AM Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote: > > On Mon, Mar 11, 2024 at 7:05 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > > > > > > > +LIBBPF_API void bpf_object__free_btfs(struct bpf_object *obj); > > > > + > > > > > > It shouldn't be exported. > > > libbpf should clean it up when bpf_object is freed. > > > > Yes, libbpf will clean up the btfs when bpf_object is freed in > > this commit. And I'm trying to offer a way to early free the btfs > > by the users manual to reduce the memory usage. Or, the > > btfs that we opened will keep existing until we close the > > bpf_object. > > > > This is optional, I can remove it if you prefer. > > Let's not extend libbpf api unless we really need to. > bpf_program__attach_trace_multi_opts() and > *skel*__attach() can probably free them. That's a good idea! Should we add a "bool free_btf" field to struct bpf_trace_multi_opts? bpf_program__attach_trace_multi_opts() can be called multi times for a bpf_object, which has multi bpf program of type tracing multi-link. Or, can we free the btfs automatically if we found all tracing multi-link programs are attached? Thanks! Menglong Dong > I don't see a use case where you'd want to keep them afterwards.
On Tue, Mar 12, 2024 at 9:51 AM Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote: > > On Mon, Mar 11, 2024 at 2:35 AM Menglong Dong > <dongmenglong.8@bytedance.com> wrote: > > > > Add target btf to the function args of bpf_check_attach_target(), then > > the caller can specify the btf to check. > > > > Signed-off-by: Menglong Dong <dongmenglong.8@bytedance.com> > > --- > > include/linux/bpf_verifier.h | 1 + > > kernel/bpf/syscall.c | 6 ++++-- > > kernel/bpf/trampoline.c | 1 + > > kernel/bpf/verifier.c | 8 +++++--- > > 4 files changed, 11 insertions(+), 5 deletions(-) > > > > diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h > > index 4b0f6600e499..6cb20efcfac3 100644 > > --- a/include/linux/bpf_verifier.h > > +++ b/include/linux/bpf_verifier.h > > @@ -811,6 +811,7 @@ static inline void bpf_trampoline_unpack_key(u64 key, u32 *obj_id, u32 *btf_id) > > int bpf_check_attach_target(struct bpf_verifier_log *log, > > const struct bpf_prog *prog, > > const struct bpf_prog *tgt_prog, > > + struct btf *btf, > > u32 btf_id, > > struct bpf_attach_target_info *tgt_info); > > void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab); > > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c > > index d1cd645ef9ac..6128c3131141 100644 > > --- a/kernel/bpf/syscall.c > > +++ b/kernel/bpf/syscall.c > > @@ -3401,9 +3401,11 @@ static int bpf_tracing_prog_attach(struct bpf_prog *prog, > > * need a new trampoline and a check for compatibility > > */ > > struct bpf_attach_target_info tgt_info = {}; > > + struct btf *btf; > > > > - err = bpf_check_attach_target(NULL, prog, tgt_prog, btf_id, > > - &tgt_info); > > + btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf; > > I think it's better to keep this bit inside bpf_check_attach_target(), > since a lot of other code in there is working with if (tgt_prog) ... > so if the caller messes up passing tgt_prog->aux->btf with tgt_prog > the bug will be difficult to debug. In the previous version, I pass the attach_btf with the following way: + origin_btf = prog->aux->attach_btf; + /* use the new attach_btf to check the target */ + prog->aux->attach_btf = attach_btf; err = bpf_check_attach_target(NULL, prog, tgt_prog, btf_id, &tgt_info); + prog->aux->attach_btf = origin_btf; And Jiri suggested to add the attach_btf to the function args of bpf_check_attach_target(). Ennn....Should I convert to the old way? Thanks! Menglong Dong > > > + err = bpf_check_attach_target(NULL, prog, tgt_prog, btf, > > + btf_id, &tgt_info);
On Mon, Mar 11, 2024 at 7:42 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > On Tue, Mar 12, 2024 at 10:09 AM Alexei Starovoitov > <alexei.starovoitov@gmail.com> wrote: > > > > On Mon, Mar 11, 2024 at 7:01 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > > > On Tue, Mar 12, 2024 at 9:46 AM Alexei Starovoitov > > > <alexei.starovoitov@gmail.com> wrote: > > > > > > > > On Mon, Mar 11, 2024 at 2:34 AM Menglong Dong > > > > <dongmenglong.8@bytedance.com> wrote: > > > > > > > > > > In this commit, we add the 'accessed_args' field to struct bpf_prog_aux, > > > > > which is used to record the accessed index of the function args in > > > > > btf_ctx_access(). > > > > > > > > > > Meanwhile, we add the function btf_check_func_part_match() to compare the > > > > > accessed function args of two function prototype. This function will be > > > > > used in the following commit. > > > > > > > > > > Signed-off-by: Menglong Dong <dongmenglong.8@bytedance.com> > > > > > --- > > > > > include/linux/bpf.h | 4 ++ > > > > > kernel/bpf/btf.c | 108 +++++++++++++++++++++++++++++++++++++++++++- > > > > > 2 files changed, 110 insertions(+), 2 deletions(-) > > > > > > > > > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h > > > > > index 95e07673cdc1..0f677fdcfcc7 100644 > > > > > --- a/include/linux/bpf.h > > > > > +++ b/include/linux/bpf.h > > > > > @@ -1461,6 +1461,7 @@ struct bpf_prog_aux { > > > > > const struct btf_type *attach_func_proto; > > > > > /* function name for valid attach_btf_id */ > > > > > const char *attach_func_name; > > > > > + u64 accessed_args; > > > > > struct bpf_prog **func; > > > > > void *jit_data; /* JIT specific data. arch dependent */ > > > > > struct bpf_jit_poke_descriptor *poke_tab; > > > > > @@ -2565,6 +2566,9 @@ struct bpf_reg_state; > > > > > int btf_prepare_func_args(struct bpf_verifier_env *env, int subprog); > > > > > int btf_check_type_match(struct bpf_verifier_log *log, const struct bpf_prog *prog, > > > > > struct btf *btf, const struct btf_type *t); > > > > > +int btf_check_func_part_match(struct btf *btf1, const struct btf_type *t1, > > > > > + struct btf *btf2, const struct btf_type *t2, > > > > > + u64 func_args); > > > > > const char *btf_find_decl_tag_value(const struct btf *btf, const struct btf_type *pt, > > > > > int comp_idx, const char *tag_key); > > > > > int btf_find_next_decl_tag(const struct btf *btf, const struct btf_type *pt, > > > > > diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c > > > > > index 170d017e8e4a..c2a0299d4358 100644 > > > > > --- a/kernel/bpf/btf.c > > > > > +++ b/kernel/bpf/btf.c > > > > > @@ -6125,19 +6125,24 @@ static bool is_int_ptr(struct btf *btf, const struct btf_type *t) > > > > > } > > > > > > > > > > static u32 get_ctx_arg_idx(struct btf *btf, const struct btf_type *func_proto, > > > > > - int off) > > > > > + int off, int *aligned_idx) > > > > > { > > > > > const struct btf_param *args; > > > > > const struct btf_type *t; > > > > > u32 offset = 0, nr_args; > > > > > int i; > > > > > > > > > > + if (aligned_idx) > > > > > + *aligned_idx = -ENOENT; > > > > > + > > > > > if (!func_proto) > > > > > return off / 8; > > > > > > > > > > nr_args = btf_type_vlen(func_proto); > > > > > args = (const struct btf_param *)(func_proto + 1); > > > > > for (i = 0; i < nr_args; i++) { > > > > > + if (aligned_idx && offset == off) > > > > > + *aligned_idx = i; > > > > > t = btf_type_skip_modifiers(btf, args[i].type, NULL); > > > > > offset += btf_type_is_ptr(t) ? 8 : roundup(t->size, 8); > > > > > if (off < offset) > > > > > @@ -6207,7 +6212,7 @@ bool btf_ctx_access(int off, int size, enum bpf_access_type type, > > > > > tname, off); > > > > > return false; > > > > > } > > > > > - arg = get_ctx_arg_idx(btf, t, off); > > > > > + arg = get_ctx_arg_idx(btf, t, off, NULL); > > > > > args = (const struct btf_param *)(t + 1); > > > > > /* if (t == NULL) Fall back to default BPF prog with > > > > > * MAX_BPF_FUNC_REG_ARGS u64 arguments. > > > > > @@ -6217,6 +6222,9 @@ bool btf_ctx_access(int off, int size, enum bpf_access_type type, > > > > > /* skip first 'void *__data' argument in btf_trace_##name typedef */ > > > > > args++; > > > > > nr_args--; > > > > > + prog->aux->accessed_args |= (1 << (arg + 1)); > > > > > + } else { > > > > > + prog->aux->accessed_args |= (1 << arg); > > > > > > > > What do you need this aligned_idx for ? > > > > I'd expect that above "accessed_args |= (1 << arg);" is enough. > > > > > > > > > > Which aligned_idx? No aligned_idx in the btf_ctx_access(), and > > > aligned_idx is only used in the btf_check_func_part_match(). > > > > > > In the btf_check_func_part_match(), I need to compare the > > > t1->args[i] and t2->args[j], which have the same offset. And > > > the aligned_idx is to find the "j" according to the offset of > > > t1->args[i]. > > > > And that's my question. > > Why you don't do the max of accessed_args across all attach > > points and do btf_check_func_type_match() to that argno > > instead of nargs1. > > This 'offset += btf_type_is_ptr(t1) ? 8 : roundup... > > is odd. > > Hi, I'm trying to make the bpf flexible enough. Let's take an example, > now we have the bpf program: > > int test1_result = 0; > int BPF_PROG(test1, int a, long b, char c) > { > test1_result = a + c; > return 0; > } > > In this program, only the 1st and 3rd arg is accessed. So all kernel > functions whose 1st arg is int and 3rd arg is char can be attached > by this bpf program, even if their 2nd arg is different. > > And let's take another example for struct. This is our bpf program: > > int test1_result = 0; > int BPF_PROG(test1, long a, long b, char c) > { > test1_result = c; > return 0; > } > > Only the 3rd arg is accessed. And we have following kernel function: > > int kernel_function1(long a, long b, char c) > { > xxx > } > > struct test1 { > long a; > long b; > }; > int kernel_function2(struct test1 a, char b) > { > xxx > } > > The kernel_function1 and kernel_function2 should be compatible, > as the bpf program only accessed the ctx[2], whose offset is 16. > And the arg in kernel_function1() with offset 16 is "char c", the > arg in kernel_function2() with offset 16 is "char b", which is > compatible. I see. I thought you're sharing the trampoline across attachments. (since bpf prog is the same). 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. 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?
On Wed, Mar 13, 2024 at 05:25:35PM -0700, Alexei Starovoitov wrote: > On Tue, Mar 12, 2024 at 6:53 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > On Wed, Mar 13, 2024 at 12:42 AM Alexei Starovoitov > > <alexei.starovoitov@gmail.com> wrote: > > > > > > On Mon, Mar 11, 2024 at 7:42 PM 梦龙董 <dongmenglong.8@bytedance.com> 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. yep, I refrained from commenting not to take you down the same path I did, but if you insist.. ;-) I managed to forgot almost all of it, but the IIRC the main pain point was that at some point I had to split existing trampoline which caused the whole trampolines management and error paths to become a mess I tried to explain things in [1] changelog and the latest patchset is in [0] feel free to use/take anything, but I advice strongly against it ;-) please let me know if I can help jirka [0] https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git/log/?h=bpf/batch [1] https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git/commit/?h=bpf/batch&id=52a1d4acdf55df41e99ca2cea51865e6821036ce
On Thu, Mar 14, 2024 at 8:27 AM Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote: > > On Tue, Mar 12, 2024 at 6:53 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: [......] > > 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. I see. Creating plenty of trampolines does take a lot of time, and I'll do testing on it. > > > > > > > Let's step back. [......] > > 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.
On Thu, Mar 14, 2024 at 2:29 PM Jiri Olsa <olsajiri@gmail.com> wrote: > > On Wed, Mar 13, 2024 at 05:25:35PM -0700, Alexei Starovoitov wrote: > > On Tue, Mar 12, 2024 at 6:53 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > > > On Wed, Mar 13, 2024 at 12:42 AM Alexei Starovoitov > > > <alexei.starovoitov@gmail.com> wrote: > > > > > > > > On Mon, Mar 11, 2024 at 7:42 PM 梦龙董 <dongmenglong.8@bytedance.com> 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. > > yep, I refrained from commenting not to take you down the same path > I did, but if you insist.. ;-) > > I managed to forgot almost all of it, but the IIRC the main pain point > was that at some point I had to split existing trampoline which caused > the whole trampolines management and error paths to become a mess > > I tried to explain things in [1] changelog and the latest patchset is in [0] > > feel free to use/take anything, but I advice strongly against it ;-) > please let me know if I can help I have to say that I have not gone far enough to encounter this problem, and I didn't dig enough to be aware of the complexity. I suspect that I can't overcome this challenge. The only thing that I thought when I hear about the "shared trampoline" is to fallback and not use the shared trampoline for the kernel functions who already have a trampoline. Anyway, let's have a try on it, based on your research. Thanks! Menglong Dong > > jirka > > > [0] https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git/log/?h=bpf/batch > [1] https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git/commit/?h=bpf/batch&id=52a1d4acdf55df41e99ca2cea51865e6821036ce
On Fri, Mar 15, 2024 at 4:00 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > On Thu, Mar 14, 2024 at 8:27 AM Alexei Starovoitov > <alexei.starovoitov@gmail.com> wrote: > > > > On Tue, Mar 12, 2024 at 6:53 PM 梦龙董 <dongmenglong.8@bytedance.com> wrote: > [......] > > > 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. > > I see. Creating plenty of trampolines does take a lot of time, > and I'll do testing on it. > I have done a simple benchmark on creating 1000 trampolines. It is slow, quite slow, which consume up to 60s. We can't do it this way. Now, I have a bad idea. How about we introduce a "dynamic trampoline"? The basic logic of it can be: """ save regs bpfs = trampoline_lookup_ip(ip) fentry = bpfs->fentries while fentry: fentry(ctx) fentry = fentry->next call origin save return value fexit = bpfs->fexits while fexit: fexit(ctx) fexit = fexit->next xxxxxx """ And we lookup the "bpfs" by the function ip in a hash map in trampoline_lookup_ip. The type of "bpfs" is: struct bpf_array { struct bpf_prog *fentries; struct bpf_prog *fexits; struct bpf_prog *modify_returns; } When we need to attach the bpf progA to function A/B/C, we only need to create the bpf_arrayA, bpf_arrayB, bpf_arrayC and add the progA to them, and insert them to the hash map "direct_call_bpfs", and attach the "dynamic trampoline" to A/B/C. If bpf_arrayA exist, just add progA to the tail of bpf_arrayA->fentries. When we need to attach progB to B/C, just add progB to bpf_arrayB->fentries and bpf_arrayB->fentries. Compared to the trampoline, extra overhead is introduced by the hash lookuping. I have not begun to code yet, and I am not sure the overhead is acceptable. Considering that we also need to do hash lookup by the function in kprobe_multi, maybe the overhead is acceptable? Thanks! Menglong Dong > > > > > > > > > > Let's step back. > [......] > > > > 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.
On Thu, 28 Mar 2024 22:43:46 +0800 梦龙董 <dongmenglong.8@bytedance.com> wrote: > I have done a simple benchmark on creating 1000 > trampolines. It is slow, quite slow, which consume up to > 60s. We can't do it this way. > > Now, I have a bad idea. How about we introduce > a "dynamic trampoline"? The basic logic of it can be: > > """ > save regs > bpfs = trampoline_lookup_ip(ip) > fentry = bpfs->fentries > while fentry: > fentry(ctx) > fentry = fentry->next > > call origin > save return value > > fexit = bpfs->fexits > while fexit: > fexit(ctx) > fexit = fexit->next > > xxxxxx > """ > > And we lookup the "bpfs" by the function ip in a hash map > in trampoline_lookup_ip. The type of "bpfs" is: > > struct bpf_array { > struct bpf_prog *fentries; > struct bpf_prog *fexits; > struct bpf_prog *modify_returns; > } > > When we need to attach the bpf progA to function A/B/C, > we only need to create the bpf_arrayA, bpf_arrayB, bpf_arrayC > and add the progA to them, and insert them to the hash map > "direct_call_bpfs", and attach the "dynamic trampoline" to > A/B/C. If bpf_arrayA exist, just add progA to the tail of > bpf_arrayA->fentries. When we need to attach progB to > B/C, just add progB to bpf_arrayB->fentries and > bpf_arrayB->fentries. > > Compared to the trampoline, extra overhead is introduced > by the hash lookuping. > > I have not begun to code yet, and I am not sure the overhead is > acceptable. Considering that we also need to do hash lookup > by the function in kprobe_multi, maybe the overhead is > acceptable? Sounds like you are just recreating the function management that ftrace has. It also can add thousands of trampolines very quickly, because it does it in batches. It takes special synchronization steps to attach to fentry. ftrace (and I believe multi-kprobes) updates all the attachments for each step, so the synchronization needed is only done once. If you really want to have thousands of functions, why not just register it with ftrace itself. It will give you the arguments via the ftrace_regs structure. Can't you just register a program as the callback? It will probably make your accounting much easier, and just let ftrace handle the fentry logic. That's what it was made to do. -- Steve
On Thu, Mar 28, 2024 at 8:10 AM Steven Rostedt <rostedt@goodmis.org> wrote: > > On Thu, 28 Mar 2024 22:43:46 +0800 > 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > I have done a simple benchmark on creating 1000 > > trampolines. It is slow, quite slow, which consume up to > > 60s. We can't do it this way. > > > > Now, I have a bad idea. How about we introduce > > a "dynamic trampoline"? The basic logic of it can be: > > > > """ > > save regs > > bpfs = trampoline_lookup_ip(ip) > > fentry = bpfs->fentries > > while fentry: > > fentry(ctx) > > fentry = fentry->next > > > > call origin > > save return value > > > > fexit = bpfs->fexits > > while fexit: > > fexit(ctx) > > fexit = fexit->next > > > > xxxxxx > > """ > > > > And we lookup the "bpfs" by the function ip in a hash map > > in trampoline_lookup_ip. The type of "bpfs" is: > > > > struct bpf_array { > > struct bpf_prog *fentries; > > struct bpf_prog *fexits; > > struct bpf_prog *modify_returns; > > } > > > > When we need to attach the bpf progA to function A/B/C, > > we only need to create the bpf_arrayA, bpf_arrayB, bpf_arrayC > > and add the progA to them, and insert them to the hash map > > "direct_call_bpfs", and attach the "dynamic trampoline" to > > A/B/C. If bpf_arrayA exist, just add progA to the tail of > > bpf_arrayA->fentries. When we need to attach progB to > > B/C, just add progB to bpf_arrayB->fentries and > > bpf_arrayB->fentries. > > > > Compared to the trampoline, extra overhead is introduced > > by the hash lookuping. > > > > I have not begun to code yet, and I am not sure the overhead is > > acceptable. Considering that we also need to do hash lookup > > by the function in kprobe_multi, maybe the overhead is > > acceptable? > > Sounds like you are just recreating the function management that ftrace > has. It also can add thousands of trampolines very quickly, because it does > it in batches. It takes special synchronization steps to attach to fentry. > ftrace (and I believe multi-kprobes) updates all the attachments for each > step, so the synchronization needed is only done once. > > If you really want to have thousands of functions, why not just register it > with ftrace itself. It will give you the arguments via the ftrace_regs > structure. Can't you just register a program as the callback? > > It will probably make your accounting much easier, and just let ftrace > handle the fentry logic. That's what it was made to do. Absolutely agree. There is no point re-inventing this logic. Menlong, before you hook up into ftrace check whether it's going to be any different from kprobe-multi, since it's the same ftrace underneath. I suspect it will look exactly the same. So it sounds like multi-fentry idea will be shelved once again.
On Thu, Mar 28, 2024 at 8:10 AM Steven Rostedt <rostedt@goodmis.org> wrote: > > On Thu, 28 Mar 2024 22:43:46 +0800 > 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > I have done a simple benchmark on creating 1000 > > trampolines. It is slow, quite slow, which consume up to > > 60s. We can't do it this way. > > > > Now, I have a bad idea. How about we introduce > > a "dynamic trampoline"? The basic logic of it can be: > > > > """ > > save regs > > bpfs = trampoline_lookup_ip(ip) > > fentry = bpfs->fentries > > while fentry: > > fentry(ctx) > > fentry = fentry->next > > > > call origin > > save return value > > > > fexit = bpfs->fexits > > while fexit: > > fexit(ctx) > > fexit = fexit->next > > > > xxxxxx > > """ > > > > And we lookup the "bpfs" by the function ip in a hash map > > in trampoline_lookup_ip. The type of "bpfs" is: > > > > struct bpf_array { > > struct bpf_prog *fentries; > > struct bpf_prog *fexits; > > struct bpf_prog *modify_returns; > > } > > > > When we need to attach the bpf progA to function A/B/C, > > we only need to create the bpf_arrayA, bpf_arrayB, bpf_arrayC > > and add the progA to them, and insert them to the hash map > > "direct_call_bpfs", and attach the "dynamic trampoline" to > > A/B/C. If bpf_arrayA exist, just add progA to the tail of > > bpf_arrayA->fentries. When we need to attach progB to > > B/C, just add progB to bpf_arrayB->fentries and > > bpf_arrayB->fentries. > > > > Compared to the trampoline, extra overhead is introduced > > by the hash lookuping. > > > > I have not begun to code yet, and I am not sure the overhead is > > acceptable. Considering that we also need to do hash lookup > > by the function in kprobe_multi, maybe the overhead is > > acceptable? > > Sounds like you are just recreating the function management that ftrace > has. It also can add thousands of trampolines very quickly, because it does > it in batches. It takes special synchronization steps to attach to fentry. > ftrace (and I believe multi-kprobes) updates all the attachments for each > step, so the synchronization needed is only done once. > > If you really want to have thousands of functions, why not just register it > with ftrace itself. It will give you the arguments via the ftrace_regs > structure. Can't you just register a program as the callback? > > It will probably make your accounting much easier, and just let ftrace > handle the fentry logic. That's what it was made to do. > I thought I'll just ask instead of digging through code, sorry for being lazy :) Is there any way to pass pt_regs/ftrace_regs captured before function execution to a return probe (fexit/kretprobe)? I.e., how hard is it to pass input function arguments to a kretprobe? That's the biggest advantage of fexit over kretprobe, and if we can make these original pt_regs/ftrace_regs available to kretprobe, then multi-kretprobe will effectively be this multi-fexit. > -- Steve
On Thu, Mar 28, 2024 at 11:11 PM Steven Rostedt <rostedt@goodmis.org> wrote: > > On Thu, 28 Mar 2024 22:43:46 +0800 > 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > I have done a simple benchmark on creating 1000 > > trampolines. It is slow, quite slow, which consume up to > > 60s. We can't do it this way. > > > > Now, I have a bad idea. How about we introduce > > a "dynamic trampoline"? The basic logic of it can be: > > > > """ > > save regs > > bpfs = trampoline_lookup_ip(ip) > > fentry = bpfs->fentries > > while fentry: > > fentry(ctx) > > fentry = fentry->next > > > > call origin > > save return value > > > > fexit = bpfs->fexits > > while fexit: > > fexit(ctx) > > fexit = fexit->next > > > > xxxxxx > > """ > > > > And we lookup the "bpfs" by the function ip in a hash map > > in trampoline_lookup_ip. The type of "bpfs" is: > > > > struct bpf_array { > > struct bpf_prog *fentries; > > struct bpf_prog *fexits; > > struct bpf_prog *modify_returns; > > } > > > > When we need to attach the bpf progA to function A/B/C, > > we only need to create the bpf_arrayA, bpf_arrayB, bpf_arrayC > > and add the progA to them, and insert them to the hash map > > "direct_call_bpfs", and attach the "dynamic trampoline" to > > A/B/C. If bpf_arrayA exist, just add progA to the tail of > > bpf_arrayA->fentries. When we need to attach progB to > > B/C, just add progB to bpf_arrayB->fentries and > > bpf_arrayB->fentries. > > > > Compared to the trampoline, extra overhead is introduced > > by the hash lookuping. > > > > I have not begun to code yet, and I am not sure the overhead is > > acceptable. Considering that we also need to do hash lookup > > by the function in kprobe_multi, maybe the overhead is > > acceptable? > > Sounds like you are just recreating the function management that ftrace > has. It also can add thousands of trampolines very quickly, because it does > it in batches. It takes special synchronization steps to attach to fentry. > ftrace (and I believe multi-kprobes) updates all the attachments for each > step, so the synchronization needed is only done once. > Yes, it is fast to register a trampoline for a kernel function in the managed ftrace in register_fentry->register_ftrace_direct->ftrace_add_rec_direct. And it will add the trampoline to the hash table "direct_functions". And the trampoline will be called in the following step (I'm not sure if I understand it correctly): ftrace_regs_caller | __ftrace_ops_list_func -> call_direct_funcs -> save trampoline to pt_regs->origin_ax | call pt_regs->origin_ax if not NULL The logic above means that we can only call a trampoline once, and a kernel function can only have one trampoline. The original idea of mine is to register all the shared trampoline to the managed ftrace. For example, if we have the shared trampoline1 for function A/B/C, and shared trampoline2 for function B/C/D, then I register trampoline1 and trampoline2 for function B/C. However, it can't work, as we can't call 2 trampolines for a function. Then, I thought that we could create a "dynamic trampoline". The logic for the non-ftrace-managed case is simple, we only need to replace the "nop" of all the target functions to "call dynamic_trampoline". And for the ftrace-managed case, the logic is the same too, except that the trampoline that we add to the "direct_functions" hash is the dynamic-trampoline: ftrace_regs_caller | __ftrace_ops_list_func -> call_direct_funcs -> save dynamic-trampoline to pt_regs->origin_ax | call pt_regs->origin_ax(dynamic-trampoline) if not NULL And in the dynamic-trampoline, we can call prog1 for A, call prog1 and prog2 for B/C, call prog2 for D. And the register is fast enough. > If you really want to have thousands of functions, why not just register it > with ftrace itself. It will give you the arguments via the ftrace_regs > structure. Can't you just register a program as the callback? > Ennn...I don't understand. The main purpose for me to use TRACING is: 1. we can directly access the memory, which is more efficient. 2. we can obtain the function args in FEXIT, which kretprobe can't do it. And this is the main reason. Thanks! Menglong Dong > It will probably make your accounting much easier, and just let ftrace > handle the fentry logic. That's what it was made to do. > > -- Steve
On Fri, Mar 29, 2024 at 7:17 AM Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote: > > On Thu, Mar 28, 2024 at 8:10 AM Steven Rostedt <rostedt@goodmis.org> wrote: > > > > On Thu, 28 Mar 2024 22:43:46 +0800 > > 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > > I have done a simple benchmark on creating 1000 > > > trampolines. It is slow, quite slow, which consume up to > > > 60s. We can't do it this way. > > > > > > Now, I have a bad idea. How about we introduce > > > a "dynamic trampoline"? The basic logic of it can be: > > > > > > """ > > > save regs > > > bpfs = trampoline_lookup_ip(ip) > > > fentry = bpfs->fentries > > > while fentry: > > > fentry(ctx) > > > fentry = fentry->next > > > > > > call origin > > > save return value > > > > > > fexit = bpfs->fexits > > > while fexit: > > > fexit(ctx) > > > fexit = fexit->next > > > > > > xxxxxx > > > """ > > > > > > And we lookup the "bpfs" by the function ip in a hash map > > > in trampoline_lookup_ip. The type of "bpfs" is: > > > > > > struct bpf_array { > > > struct bpf_prog *fentries; > > > struct bpf_prog *fexits; > > > struct bpf_prog *modify_returns; > > > } > > > > > > When we need to attach the bpf progA to function A/B/C, > > > we only need to create the bpf_arrayA, bpf_arrayB, bpf_arrayC > > > and add the progA to them, and insert them to the hash map > > > "direct_call_bpfs", and attach the "dynamic trampoline" to > > > A/B/C. If bpf_arrayA exist, just add progA to the tail of > > > bpf_arrayA->fentries. When we need to attach progB to > > > B/C, just add progB to bpf_arrayB->fentries and > > > bpf_arrayB->fentries. > > > > > > Compared to the trampoline, extra overhead is introduced > > > by the hash lookuping. > > > > > > I have not begun to code yet, and I am not sure the overhead is > > > acceptable. Considering that we also need to do hash lookup > > > by the function in kprobe_multi, maybe the overhead is > > > acceptable? > > > > Sounds like you are just recreating the function management that ftrace > > has. It also can add thousands of trampolines very quickly, because it does > > it in batches. It takes special synchronization steps to attach to fentry. > > ftrace (and I believe multi-kprobes) updates all the attachments for each > > step, so the synchronization needed is only done once. > > > > If you really want to have thousands of functions, why not just register it > > with ftrace itself. It will give you the arguments via the ftrace_regs > > structure. Can't you just register a program as the callback? > > > > It will probably make your accounting much easier, and just let ftrace > > handle the fentry logic. That's what it was made to do. > > Absolutely agree. > There is no point re-inventing this logic. > > Menlong, > before you hook up into ftrace check whether > it's going to be any different from kprobe-multi, > since it's the same ftrace underneath. > I suspect it will look exactly the same. Yeah, I dig it a little. I think it is different. For multi-kprobe, it registers a ftrace_ops to ftrace_ops_list for every bpf program. This means that we can register 2 or more multi-kprobe in the same function. The bpf is called in the following step: ftrace_regs_caller | __ftrace_ops_list_func -> fprobe_handler -> kprobe_multi_link_handler -> run BPF And for trampoline, it needs to be called directly, so it can't be registered as a callback to ftrace_ops_list. It need to be called in the following step: ftrace_regs_caller | __ftrace_ops_list_func -> call_direct_funcs -> save trampoline to pt_regs->origin_ax | call pt_regs->origin_ax if not NULL > So it sounds like multi-fentry idea will be shelved once again. Enn...this is the best solution that I can think of. If it doesn't work, I suspect it will be shelved again. Thanks! Menglong Dong
On Sat, Mar 30, 2024 at 7:28 AM Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote: > > On Thu, Mar 28, 2024 at 8:10 AM Steven Rostedt <rostedt@goodmis.org> wrote: > > > > On Thu, 28 Mar 2024 22:43:46 +0800 > > 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > > I have done a simple benchmark on creating 1000 > > > trampolines. It is slow, quite slow, which consume up to > > > 60s. We can't do it this way. > > > > > > Now, I have a bad idea. How about we introduce > > > a "dynamic trampoline"? The basic logic of it can be: > > > > > > """ > > > save regs > > > bpfs = trampoline_lookup_ip(ip) > > > fentry = bpfs->fentries > > > while fentry: > > > fentry(ctx) > > > fentry = fentry->next > > > > > > call origin > > > save return value > > > > > > fexit = bpfs->fexits > > > while fexit: > > > fexit(ctx) > > > fexit = fexit->next > > > > > > xxxxxx > > > """ > > > > > > And we lookup the "bpfs" by the function ip in a hash map > > > in trampoline_lookup_ip. The type of "bpfs" is: > > > > > > struct bpf_array { > > > struct bpf_prog *fentries; > > > struct bpf_prog *fexits; > > > struct bpf_prog *modify_returns; > > > } > > > > > > When we need to attach the bpf progA to function A/B/C, > > > we only need to create the bpf_arrayA, bpf_arrayB, bpf_arrayC > > > and add the progA to them, and insert them to the hash map > > > "direct_call_bpfs", and attach the "dynamic trampoline" to > > > A/B/C. If bpf_arrayA exist, just add progA to the tail of > > > bpf_arrayA->fentries. When we need to attach progB to > > > B/C, just add progB to bpf_arrayB->fentries and > > > bpf_arrayB->fentries. > > > > > > Compared to the trampoline, extra overhead is introduced > > > by the hash lookuping. > > > > > > I have not begun to code yet, and I am not sure the overhead is > > > acceptable. Considering that we also need to do hash lookup > > > by the function in kprobe_multi, maybe the overhead is > > > acceptable? > > > > Sounds like you are just recreating the function management that ftrace > > has. It also can add thousands of trampolines very quickly, because it does > > it in batches. It takes special synchronization steps to attach to fentry. > > ftrace (and I believe multi-kprobes) updates all the attachments for each > > step, so the synchronization needed is only done once. > > > > If you really want to have thousands of functions, why not just register it > > with ftrace itself. It will give you the arguments via the ftrace_regs > > structure. Can't you just register a program as the callback? > > > > It will probably make your accounting much easier, and just let ftrace > > handle the fentry logic. That's what it was made to do. > > > > I thought I'll just ask instead of digging through code, sorry for > being lazy :) Is there any way to pass pt_regs/ftrace_regs captured > before function execution to a return probe (fexit/kretprobe)? I.e., > how hard is it to pass input function arguments to a kretprobe? That's > the biggest advantage of fexit over kretprobe, and if we can make > these original pt_regs/ftrace_regs available to kretprobe, then > multi-kretprobe will effectively be this multi-fexit. Yes, we can use multi-kretprobe instead of multi-fexit if we can obtain the function args in kretprobe. I think it's hard. The reason that we can obtain the function args is that we have a trampoline, and it call the origin function for FEXIT. If we do the same for multi-kretprobe, we need to modify ftrace_regs_caller to: ftrace_regs_caller | __ftrace_ops_list_func | call all multi-kprobe callbacks | call orgin | call all multi-kretprobe callbacks | call bpf trampoline(for TRACING) However, this logic conflicts with bpf trampoline, as it can also call the origin function. What's more, the FENTRY should be called before the "call origin" above. I'm sure if I understand correctly, as I have not figured out how multi-kretprobe works in fprobe. Thanks! Menglong Dong > > > -- Steve
On Fri, 29 Mar 2024 16:28:33 -0700 Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote: > I thought I'll just ask instead of digging through code, sorry for > being lazy :) Is there any way to pass pt_regs/ftrace_regs captured > before function execution to a return probe (fexit/kretprobe)? I.e., > how hard is it to pass input function arguments to a kretprobe? That's > the biggest advantage of fexit over kretprobe, and if we can make > these original pt_regs/ftrace_regs available to kretprobe, then > multi-kretprobe will effectively be this multi-fexit. This should be possible with the updates that Masami is doing with the fgraph code. -- Steve
On Sat, Mar 30, 2024 at 08:27:55AM -0400, Steven Rostedt wrote: > On Fri, 29 Mar 2024 16:28:33 -0700 > Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote: > > > I thought I'll just ask instead of digging through code, sorry for > > being lazy :) Is there any way to pass pt_regs/ftrace_regs captured > > before function execution to a return probe (fexit/kretprobe)? I.e., > > how hard is it to pass input function arguments to a kretprobe? That's > > the biggest advantage of fexit over kretprobe, and if we can make > > these original pt_regs/ftrace_regs available to kretprobe, then > > multi-kretprobe will effectively be this multi-fexit. > > This should be possible with the updates that Masami is doing with the > fgraph code. yes, I have bpf kprobe-multi link support for that [0] (it's on top of Masami's fprobe-over-fgraph changes) we discussed that in [1] jirka [0] https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git/log/?h=bpf/session_data [1] https://lore.kernel.org/bpf/20240228090242.4040210-1-jolsa@kernel.org/
On Sat, 30 Mar 2024 11:18:29 +0800 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > If you really want to have thousands of functions, why not just register it > > with ftrace itself. It will give you the arguments via the ftrace_regs > > structure. Can't you just register a program as the callback? > > > > Ennn...I don't understand. The main purpose for > me to use TRACING is: > > 1. we can directly access the memory, which is more > efficient. I'm not sure what you mean by the above. Access what memory? > 2. we can obtain the function args in FEXIT, which > kretprobe can't do it. And this is the main reason. I didn't mention kretprobe. If you need access to the exit of the function, you can use Masami's fgraph update. fentry -> ftrace_trampoline -> your_code For fgraph: fentry -> ftrace_trampoline -> fgraph [sets up return call] -> your_entry_code function ret -> fgraph_ret_handler -> your_exit_code And you will be able to pass data from the entry to the exit code, including parameters. -- Steve
On Sat, Mar 30, 2024 at 10:52 AM Jiri Olsa <olsajiri@gmail.com> wrote: > > On Sat, Mar 30, 2024 at 08:27:55AM -0400, Steven Rostedt wrote: > > On Fri, 29 Mar 2024 16:28:33 -0700 > > Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote: > > > > > I thought I'll just ask instead of digging through code, sorry for > > > being lazy :) Is there any way to pass pt_regs/ftrace_regs captured > > > before function execution to a return probe (fexit/kretprobe)? I.e., > > > how hard is it to pass input function arguments to a kretprobe? That's > > > the biggest advantage of fexit over kretprobe, and if we can make > > > these original pt_regs/ftrace_regs available to kretprobe, then > > > multi-kretprobe will effectively be this multi-fexit. > > > > This should be possible with the updates that Masami is doing with the > > fgraph code. > > yes, I have bpf kprobe-multi link support for that [0] (it's on top of > Masami's fprobe-over-fgraph changes) we discussed that in [1] Sorry, I forgot the regs/args part, mostly remembering we discussed session cookie ideas. Thanks for reminder! > > jirka > > [0] https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git/log/?h=bpf/session_data > [1] https://lore.kernel.org/bpf/20240228090242.4040210-1-jolsa@kernel.org/
On Sun, Mar 31, 2024 at 3:34 AM Steven Rostedt <rostedt@goodmis.org> wrote: > > On Sat, 30 Mar 2024 11:18:29 +0800 > 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > If you really want to have thousands of functions, why not just register it > > > with ftrace itself. It will give you the arguments via the ftrace_regs > > > structure. Can't you just register a program as the callback? > > > > > > > Ennn...I don't understand. The main purpose for > > me to use TRACING is: > > > > 1. we can directly access the memory, which is more > > efficient. > > I'm not sure what you mean by the above. Access what memory? > We need to use the helper of bpf_probe_read_kernel when we read "skb->sk" in kprobe, and the "skb" is the 1st arg in ip_rcv(). And we can directly read "skb->sk" in tracing, which is more efficient. Isn't it? > > 2. we can obtain the function args in FEXIT, which > > kretprobe can't do it. And this is the main reason. > > I didn't mention kretprobe. If you need access to the exit of the function, > you can use Masami's fgraph update. > > fentry -> ftrace_trampoline -> your_code > > For fgraph: > > fentry -> ftrace_trampoline -> fgraph [sets up return call] -> your_entry_code > > function ret -> fgraph_ret_handler -> your_exit_code > > And you will be able to pass data from the entry to the exit code, > including parameters. Yeah, the fgraph sounds like a nice solution to my problem. I'll have a try on it. Thanks! Menglong Dong > > -- Steve > >
On Mon, 1 Apr 2024 10:28:17 +0800 梦龙董 <dongmenglong.8@bytedance.com> wrote: > On Sun, Mar 31, 2024 at 3:34 AM Steven Rostedt <rostedt@goodmis.org> wrote: > > > > On Sat, 30 Mar 2024 11:18:29 +0800 > > 梦龙董 <dongmenglong.8@bytedance.com> wrote: > > > > > > If you really want to have thousands of functions, why not just register it > > > > with ftrace itself. It will give you the arguments via the ftrace_regs > > > > structure. Can't you just register a program as the callback? > > > > > > > > > > Ennn...I don't understand. The main purpose for > > > me to use TRACING is: > > > > > > 1. we can directly access the memory, which is more > > > efficient. > > > > I'm not sure what you mean by the above. Access what memory? > > > > We need to use the helper of bpf_probe_read_kernel > when we read "skb->sk" in kprobe, and the "skb" is the > 1st arg in ip_rcv(). And we can directly read "skb->sk" > in tracing, which is more efficient. Isn't it? If you add a ftrace_ops function handler that calls a BPF program, I don't see why you can't just give it the parameters it needs instead of using bpf helpers. It's no different than using a trampoline to do the same thing. -- Steve
For now, the BPF program of type BPF_PROG_TYPE_TRACING is not allowed to be attached to multiple hooks, and we have to create a BPF program for each kernel function, for which we want to trace, even through all the program have the same (or similar) logic. This can consume extra memory, and make the program loading slow if we have plenty of kernel function to trace. In this series, we add the support to allow attaching a tracing BPF program to multi hooks, which is similar to BPF_TRACE_KPROBE_MULTI. In the 1st patch, we add the support to record index of the accessed function args of the target for tracing program. Meanwhile, we add the function btf_check_func_part_match() to compare the accessed function args of two function prototype. This function will be used in the next commit. In the 2nd patch, we refactor the struct modules_array to ptr_array, as we need similar function to hold the target btf, target program and kernel modules that we reference to in the following commit. In the 3rd patch, we introduce the struct bpf_tramp_link_conn to be the bridge between bpf_link and trampoline, as the releation between bpf_link and trampoline is not one-to-one anymore. In the 4th patch, we add the struct bpf_tramp_multi_link and bpf_trampoline_multi_{link,unlink}_prog for multi-link of trampoline. In the 5th patch, we add target btf to the function args of bpf_check_attach_target(), then the caller can specify the btf to check. The 6th patch is the main part to add multi-link supporting for tracing. For now, only the following attach type is supported: BPF_TRACE_FENTRY_MULTI BPF_TRACE_FEXIT_MULTI BPF_MODIFY_RETURN_MULTI The attach type of BPF_TRACE_RAW_TP has different link type, so we skip this part in this series for now. In the 7th and 8th patches, we add multi-link supporting of tracing to libbpf. Note that we don't free btfs that we load after the bpf programs are loaded into the kernel now if any programs of type tracing multi-link existing, as we need to lookup the btf types during attaching. In the 9th patch, we add the testcases for this series. Changes since v1: - According to the advice of Alexei, introduce multi-link for tracing instead of attaching a tracing program to multiple trampolines with creating multi instance of bpf_link. Menglong Dong (9): bpf: tracing: add support to record and check the accessed args bpf: refactor the modules_array to ptr_array bpf: trampoline: introduce struct bpf_tramp_link_conn bpf: trampoline: introduce bpf_tramp_multi_link bpf: verifier: add btf to the function args of bpf_check_attach_target bpf: tracing: add multi-link support libbpf: don't free btf if program of multi-link tracing existing libbpf: add support for the multi-link of tracing selftests/bpf: add testcases for multi-link of tracing arch/arm64/net/bpf_jit_comp.c | 4 +- arch/riscv/net/bpf_jit_comp64.c | 4 +- arch/s390/net/bpf_jit_comp.c | 4 +- arch/x86/net/bpf_jit_comp.c | 4 +- include/linux/bpf.h | 51 ++- include/linux/bpf_verifier.h | 1 + include/uapi/linux/bpf.h | 10 + kernel/bpf/bpf_struct_ops.c | 2 +- kernel/bpf/btf.c | 113 ++++- kernel/bpf/syscall.c | 425 +++++++++++++++++- kernel/bpf/trampoline.c | 97 +++- kernel/bpf/verifier.c | 24 +- kernel/trace/bpf_trace.c | 48 +- net/bpf/test_run.c | 3 + net/core/bpf_sk_storage.c | 2 + tools/bpf/bpftool/common.c | 3 + tools/include/uapi/linux/bpf.h | 10 + tools/lib/bpf/bpf.c | 10 + tools/lib/bpf/bpf.h | 6 + tools/lib/bpf/libbpf.c | 215 ++++++++- tools/lib/bpf/libbpf.h | 16 + tools/lib/bpf/libbpf.map | 2 + .../selftests/bpf/bpf_testmod/bpf_testmod.c | 49 ++ .../bpf/prog_tests/tracing_multi_link.c | 153 +++++++ .../selftests/bpf/progs/tracing_multi_test.c | 209 +++++++++ 25 files changed, 1366 insertions(+), 99 deletions(-) create mode 100644 tools/testing/selftests/bpf/prog_tests/tracing_multi_link.c create mode 100644 tools/testing/selftests/bpf/progs/tracing_multi_test.c