From 3163613c5bc805dadac8ea157648eefd46747cae Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Fri, 1 Mar 2019 16:09:31 -0300 Subject: [PATCH] perf bpf: Automatically add BTF ELF markers MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The libbpf loader expects that some __btf_map_ structs be in place with the keys and values types of maps so that one can store the struct definitions and have them sent to the kernel via sys_bpf(fd, cmd = BTF_LOAD) and then later be retrievable via sys_bpf(fd, cmd = BPF_OBJ_GET_INFO_BY_FD) for use by tools such as 'bpftool map dump id MAP_ID'. Since we already have this for defining maps in 'perf trace' BPF events: bpf_map(name, _type, type_key, type_val, _max_entries) As used in the tools/perf/examples/bpf/augmented_raw_syscalls.c: --- 8< --- struct syscall { bool enabled; }; bpf_map(syscalls, ARRAY, int, struct syscall, 512); --- 8< --- All we need is to get all that already available info, piggyback on the 'bpf_map' define in tools/perf/include/bpf/bpf.h, that is included by 'perf trace' BPF programs and do that without requiring changes to the BPF programs already defining maps using 'bpf_map()'. So this is what we have before this patch: 1) With this in ~/.perfconfig to dump .c events as .o, aka save a copy so that we can use the .o later as a pre-compiled BPF bytecode: # grep '\[llvm\]' -A2 ~/.perfconfig [llvm] dump-obj = true clang-opt = -g # # clang --version clang version 9.0.0 (https://git.llvm.org/git/clang.git/ 7906282d3afec5dfdc2b27943fd6c0309086c507) (https://git.llvm.org/git/llvm.git/ a1b5de1ff8ae8bc79dc8e86e1f82565229bd0500) Target: x86_64-unknown-linux-gnu Thread model: posix InstalledDir: /opt/llvm/bin 2) Note the -g there so that we get clang to generate debuginfo, and since the target is 'bpf' it will generate the BTF info in this clang version (9.0). 3) Run a simple 'perf record' specifiying as an event the augmented_raw_syscalls.c source code: # perf record -e /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.c sleep 1 LLVM: dumping /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.025 MB perf.data ] # file /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o: ELF 64-bit LSB relocatable, eBPF, version 1 (SYSV), with debug_info, not stripped 4) Look at the BTF structs encoded in it: # pahole -F btf --sizes /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o syscall_enter_args 64 0 augmented_filename 264 0 syscall 1 0 syscall_exit_args 24 0 bpf_map 28 0 # # pahole -F btf -C syscalls /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o # pahole -F btf -C syscall /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o struct syscall { bool enabled; /* 0 1 */ /* size: 1, cachelines: 1, members: 1 */ /* last cacheline: 1 bytes */ }; # 5) Ok, with just this we don't have the markers expected by the libbpf loader and when we run with this BPF bytecode, because we have: # grep '\[trace\]' -A1 ~/.perfconfig [trace] add_events = /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o # 6) Lets do a 'perf trace' system wide session using this BPF program: # perf trace -e *mmsg,open* Cache2 I/O/6885 openat(AT_FDCWD, "/home/acme/.cache/mozilla/firefox/ina67tev.default/cache2/entries/BA220AB2914006A7AE96D27BE6EA13DD77519FCA", O_RDWR|O_CREAT|O_TRUNC, S_IRUSR|S_IWUSR) = 106 Cache2 I/O/6885 openat(AT_FDCWD, "/proc/self/mountinfo", O_RDONLY) = 121 Cache2 I/O/6885 openat(AT_FDCWD, "/proc/self/mountinfo", O_RDONLY) = 121 Cache2 I/O/6885 openat(AT_FDCWD, "/proc/self/mountinfo", O_RDONLY) = 121 Cache2 I/O/6885 openat(AT_FDCWD, "/proc/self/mountinfo", O_RDONLY) = 121 DNS Res~ver #3/23340 openat(AT_FDCWD, "/etc/hosts", O_RDONLY|O_CLOEXEC) = 106 DNS Res~ver #3/23340 sendmmsg(106, 0x7f252f1fcaf0, 2, MSG_NOSIGNAL) = 2 Cache2 I/O/6885 openat(AT_FDCWD, "/home/acme/.cache/mozilla/firefox/ina67tev.default/cache2/entries/BA220AB2914006A7AE96D27BE6EA13DD77519FCA", O_RDWR) = 106 lighttpd/18915 openat(AT_FDCWD, "/proc/loadavg", O_RDONLY) = 12 7) While it runs lets see the maps that 'perf trace' + libbpf's BPF loader loaded into the kernel via sys_bpf(fd, BPF_BTF_LOAD, ...): # bpftool map list | tail -6 149: perf_event_array name __augmented_sys flags 0x0 key 4B value 4B max_entries 8 memlock 4096B 150: array name syscalls flags 0x0 key 4B value 1B max_entries 512 memlock 8192B 151: hash name pids_filtered flags 0x0 key 4B value 1B max_entries 64 memlock 8192B # 8) Dump the "pids_filtered", map, that will have one entry per PID that 'perf trace' wants filtered, which includes its own, to avoid a tracing feedback loop (perf trace shows the syscalls it does which generates more syscalls that it has to show that...), it also auto-filters the 'gnome-terminal' and 'sshd' parent PIDs, for the same reason: # bpftool map dump id 151 key: a5 0c 00 00 value: 01 key: 14 63 00 00 value: 01 Found 2 elements # 9) Since there is no BTF info available, it does a generic hex dump :-\ 10) Now, with this patch applied, we'll do steps 3 to 6 again and look with pahole if there are extra structs encoded in BTF: # pahole -F btf --sizes /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o syscall_enter_args 64 0 augmented_filename 264 0 syscall 1 0 syscall_exit_args 24 0 bpf_map 28 0 ____btf_map___augmented_syscalls__ 8 0 ____btf_map_syscalls 8 0 ____btf_map_pids_filtered 8 0 # 11) Yes, those __btf_map_ + the map names, lets see how they look like: # pahole -F btf -C ____btf_map_syscalls /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o struct ____btf_map_syscalls { int key; /* 0 4 */ struct syscall value; /* 4 1 */ /* size: 8, cachelines: 1, members: 2 */ /* padding: 3 */ /* last cacheline: 8 bytes */ }; # 12) Lets repeat step 7 to get the new map ids: # bpftool map list | tail -6 155: perf_event_array name __augmented_sys flags 0x0 key 4B value 4B max_entries 8 memlock 4096B 156: array name syscalls flags 0x0 key 4B value 1B max_entries 512 memlock 8192B 157: hash name pids_filtered flags 0x0 key 4B value 1B max_entries 64 memlock 8192B # 13) And finally lets dump the 'pids_filtered': # bpftool map dump id 157 [{ "key": 3237, "value": true },{ "key": 26435, "value": true } ] # Looks much better! BTF info was used to interpret the key as an integer and the value as a struct with just one boolean member, so to make it more compact, show just the 'true' value where we saw '01'. Now to make 'perf trace --dump-map' to use BTF! Cc: Adrian Hunter Cc: Alexei Starovoitov Cc: Andrii Nakryiko Cc: Daniel Borkmann Cc: Jiri Olsa Cc: Luis Cláudio Gonçalves Cc: Martin KaFai Lau Cc: Namhyung Kim Cc: Song Liu Cc: Wang Nan Cc: Yonghong Song Link: https://lkml.kernel.org/n/tip-ybuf9wpkm30xk28iq7jbwb40@git.kernel.org Signed-off-by: Arnaldo Carvalho de Melo --- tools/perf/include/bpf/bpf.h | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/tools/perf/include/bpf/bpf.h b/tools/perf/include/bpf/bpf.h index 5df7ed9d9020..2eac6d804b2d 100644 --- a/tools/perf/include/bpf/bpf.h +++ b/tools/perf/include/bpf/bpf.h @@ -24,7 +24,13 @@ struct bpf_map SEC("maps") name = { \ .key_size = sizeof(type_key), \ .value_size = sizeof(type_val), \ .max_entries = _max_entries, \ -} +}; \ +struct ____btf_map_##name { \ + type_key key; \ + type_val value; \ +}; \ +struct ____btf_map_##name __attribute__((section(".maps." #name), used)) \ + ____btf_map_##name = { } /* * FIXME: this should receive .max_entries as a parameter, as careful