| Message ID | 20200226130345.209469-17-jolsa@kernel.org |
|---|---|
| State | Changes Requested |
| Delegated to: | BPF Maintainers |
| Headers | show |
| Series | bpf: Add trampoline and dispatcher to /proc/kallsyms | expand |
On Wed, Feb 26, 2020 at 5:06 AM Jiri Olsa <jolsa@kernel.org> wrote: > > Synthesize bpf images (trampolines/dispatchers) on start, > as ksymbol events from /proc/kallsyms. Having this perf > can recognize samples from those images and perf report > and top shows them correctly. > > The rest of the ksymbol handling is already in place from > for the bpf programs monitoring, so only the initial state > was needed. > > perf report output: > > # Overhead Command Shared Object Symbol > > 12.37% test_progs [kernel.vmlinux] [k] entry_SYSCALL_64 > 11.80% test_progs [kernel.vmlinux] [k] syscall_return_via_sysret > 9.63% test_progs bpf_prog_bcf7977d3b93787c_prog2 [k] bpf_prog_bcf7977d3b93787c_prog2 > 6.90% test_progs bpf_trampoline_24456 [k] bpf_trampoline_24456 > 6.36% test_progs [kernel.vmlinux] [k] memcpy_erms > > Signed-off-by: Jiri Olsa <jolsa@kernel.org> Acked-by: Song Liu <songliubraving@fb.com>
Em Wed, Feb 26, 2020 at 02:03:43PM +0100, Jiri Olsa escreveu: > Synthesize bpf images (trampolines/dispatchers) on start, > as ksymbol events from /proc/kallsyms. Having this perf > can recognize samples from those images and perf report > and top shows them correctly. > > The rest of the ksymbol handling is already in place from > for the bpf programs monitoring, so only the initial state > was needed. Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com> But at some point we should try and consolidate all those kallsym__parse() calls we have in tools/perf/ not to do it that many times, see below _before_ this patch: [root@five ~]# perf probe -x ~/bin/perf kallsyms__parse Added new event: probe_perf:kallsyms__parse (on kallsyms__parse in /home/acme/bin/perf) You can now use it in all perf tools, such as: perf record -e probe_perf:kallsyms__parse -aR sleep 1 [root@five ~]# perf trace -e probe_perf:kallsyms__parse/max-stack=8/ -- perf record sleep 1 0.000 perf/6444 probe_perf:kallsyms__parse(__probe_ip: 4904384) kallsyms__parse (/home/acme/bin/perf) machine__get_running_kernel_start (/home/acme/bin/perf) machine__create_kernel_maps (/home/acme/bin/perf) perf_session__new (/home/acme/bin/perf) cmd_record (/home/acme/bin/perf) run_builtin (/home/acme/bin/perf) main (/home/acme/bin/perf) __libc_start_main (/usr/lib64/libc-2.30.so) 0.124 perf/6444 probe_perf:kallsyms__parse(__probe_ip: 4904384) kallsyms__parse (/home/acme/bin/perf) machine__get_running_kernel_start (/home/acme/bin/perf) machine__create_kernel_maps (/home/acme/bin/perf) perf_session__new (/home/acme/bin/perf) cmd_record (/home/acme/bin/perf) run_builtin (/home/acme/bin/perf) main (/home/acme/bin/perf) __libc_start_main (/usr/lib64/libc-2.30.so) 15.489 perf/6444 probe_perf:kallsyms__parse(__probe_ip: 4904384) kallsyms__parse (/home/acme/bin/perf) machine__create_kernel_maps (/home/acme/bin/perf) perf_session__new (/home/acme/bin/perf) cmd_record (/home/acme/bin/perf) run_builtin (/home/acme/bin/perf) main (/home/acme/bin/perf) __libc_start_main (/usr/lib64/libc-2.30.so) [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.019 MB perf.data (7 samples) ] [root@five ~]# - Arnaldo > perf report output: > > # Overhead Command Shared Object Symbol > > 12.37% test_progs [kernel.vmlinux] [k] entry_SYSCALL_64 > 11.80% test_progs [kernel.vmlinux] [k] syscall_return_via_sysret > 9.63% test_progs bpf_prog_bcf7977d3b93787c_prog2 [k] bpf_prog_bcf7977d3b93787c_prog2 > 6.90% test_progs bpf_trampoline_24456 [k] bpf_trampoline_24456 > 6.36% test_progs [kernel.vmlinux] [k] memcpy_erms > > Signed-off-by: Jiri Olsa <jolsa@kernel.org> > --- > tools/perf/util/bpf-event.c | 98 +++++++++++++++++++++++++++++++++++++ > 1 file changed, 98 insertions(+) > > diff --git a/tools/perf/util/bpf-event.c b/tools/perf/util/bpf-event.c > index a3207d900339..120ec547ae75 100644 > --- a/tools/perf/util/bpf-event.c > +++ b/tools/perf/util/bpf-event.c > @@ -6,6 +6,9 @@ > #include <bpf/libbpf.h> > #include <linux/btf.h> > #include <linux/err.h> > +#include <linux/string.h> > +#include <internal/lib.h> > +#include <symbol/kallsyms.h> > #include "bpf-event.h" > #include "debug.h" > #include "dso.h" > @@ -290,11 +293,87 @@ static int perf_event__synthesize_one_bpf_prog(struct perf_session *session, > return err ? -1 : 0; > } > > +struct kallsyms_parse { > + union perf_event *event; > + perf_event__handler_t process; > + struct machine *machine; > + struct perf_tool *tool; > +}; > + > +static int > +process_bpf_image(char *name, u64 addr, struct kallsyms_parse *data) > +{ > + struct machine *machine = data->machine; > + union perf_event *event = data->event; > + struct perf_record_ksymbol *ksymbol; > + u32 size; > + > + ksymbol = &event->ksymbol; > + > + /* > + * The bpf image (trampoline/dispatcher) size is aligned to > + * page, while it starts little bit after the page boundary. > + */ > + size = page_size - (addr - PERF_ALIGN(addr, page_size)); > + > + *ksymbol = (struct perf_record_ksymbol) { > + .header = { > + .type = PERF_RECORD_KSYMBOL, > + .size = offsetof(struct perf_record_ksymbol, name), > + }, > + .addr = addr, > + .len = size, > + .ksym_type = PERF_RECORD_KSYMBOL_TYPE_BPF, > + .flags = 0, > + }; > + > + strncpy(ksymbol->name, name, KSYM_NAME_LEN); > + ksymbol->header.size += PERF_ALIGN(strlen(name) + 1, sizeof(u64)); > + memset((void *) event + event->header.size, 0, machine->id_hdr_size); > + event->header.size += machine->id_hdr_size; > + > + return perf_tool__process_synth_event(data->tool, event, machine, > + data->process); > +} > + > +static int > +kallsyms_process_symbol(void *data, const char *_name, > + char type __maybe_unused, u64 start) > +{ > + char *module, *name; > + unsigned long id; > + int err = 0; > + > + module = strchr(_name, '\t'); > + if (!module) > + return 0; > + > + /* We are going after [bpf] module ... */ > + if (strcmp(module + 1, "[bpf]")) > + return 0; > + > + name = memdup(_name, (module - _name) + 1); > + if (!name) > + return -ENOMEM; > + > + name[module - _name] = 0; > + > + /* .. and only for trampolines and dispatchers */ > + if ((sscanf(name, "bpf_trampoline_%lu", &id) == 1) || > + (sscanf(name, "bpf_dispatcher_%lu", &id) == 1)) > + err = process_bpf_image(name, start, data); > + > + free(name); > + return err; > +} > + > int perf_event__synthesize_bpf_events(struct perf_session *session, > perf_event__handler_t process, > struct machine *machine, > struct record_opts *opts) > { > + const char *kallsyms_filename = "/proc/kallsyms"; > + struct kallsyms_parse arg; > union perf_event *event; > __u32 id = 0; > int err; > @@ -303,6 +382,8 @@ int perf_event__synthesize_bpf_events(struct perf_session *session, > event = malloc(sizeof(event->bpf) + KSYM_NAME_LEN + machine->id_hdr_size); > if (!event) > return -1; > + > + /* Synthesize all the bpf programs in system. */ > while (true) { > err = bpf_prog_get_next_id(id, &id); > if (err) { > @@ -335,6 +416,23 @@ int perf_event__synthesize_bpf_events(struct perf_session *session, > break; > } > } > + > + /* Synthesize all the bpf images - trampolines/dispatchers. */ > + if (symbol_conf.kallsyms_name != NULL) > + kallsyms_filename = symbol_conf.kallsyms_name; > + > + arg = (struct kallsyms_parse) { > + .event = event, > + .process = process, > + .machine = machine, > + .tool = session->tool, > + }; > + > + if (kallsyms__parse(kallsyms_filename, &arg, kallsyms_process_symbol)) { > + pr_err("%s: failed to synthesize bpf images: %s\n", > + __func__, strerror(errno)); > + } > + > free(event); > return err; > } > -- > 2.24.1
diff --git a/tools/perf/util/bpf-event.c b/tools/perf/util/bpf-event.c index a3207d900339..120ec547ae75 100644 --- a/tools/perf/util/bpf-event.c +++ b/tools/perf/util/bpf-event.c @@ -6,6 +6,9 @@ #include <bpf/libbpf.h> #include <linux/btf.h> #include <linux/err.h> +#include <linux/string.h> +#include <internal/lib.h> +#include <symbol/kallsyms.h> #include "bpf-event.h" #include "debug.h" #include "dso.h" @@ -290,11 +293,87 @@ static int perf_event__synthesize_one_bpf_prog(struct perf_session *session, return err ? -1 : 0; } +struct kallsyms_parse { + union perf_event *event; + perf_event__handler_t process; + struct machine *machine; + struct perf_tool *tool; +}; + +static int +process_bpf_image(char *name, u64 addr, struct kallsyms_parse *data) +{ + struct machine *machine = data->machine; + union perf_event *event = data->event; + struct perf_record_ksymbol *ksymbol; + u32 size; + + ksymbol = &event->ksymbol; + + /* + * The bpf image (trampoline/dispatcher) size is aligned to + * page, while it starts little bit after the page boundary. + */ + size = page_size - (addr - PERF_ALIGN(addr, page_size)); + + *ksymbol = (struct perf_record_ksymbol) { + .header = { + .type = PERF_RECORD_KSYMBOL, + .size = offsetof(struct perf_record_ksymbol, name), + }, + .addr = addr, + .len = size, + .ksym_type = PERF_RECORD_KSYMBOL_TYPE_BPF, + .flags = 0, + }; + + strncpy(ksymbol->name, name, KSYM_NAME_LEN); + ksymbol->header.size += PERF_ALIGN(strlen(name) + 1, sizeof(u64)); + memset((void *) event + event->header.size, 0, machine->id_hdr_size); + event->header.size += machine->id_hdr_size; + + return perf_tool__process_synth_event(data->tool, event, machine, + data->process); +} + +static int +kallsyms_process_symbol(void *data, const char *_name, + char type __maybe_unused, u64 start) +{ + char *module, *name; + unsigned long id; + int err = 0; + + module = strchr(_name, '\t'); + if (!module) + return 0; + + /* We are going after [bpf] module ... */ + if (strcmp(module + 1, "[bpf]")) + return 0; + + name = memdup(_name, (module - _name) + 1); + if (!name) + return -ENOMEM; + + name[module - _name] = 0; + + /* .. and only for trampolines and dispatchers */ + if ((sscanf(name, "bpf_trampoline_%lu", &id) == 1) || + (sscanf(name, "bpf_dispatcher_%lu", &id) == 1)) + err = process_bpf_image(name, start, data); + + free(name); + return err; +} + int perf_event__synthesize_bpf_events(struct perf_session *session, perf_event__handler_t process, struct machine *machine, struct record_opts *opts) { + const char *kallsyms_filename = "/proc/kallsyms"; + struct kallsyms_parse arg; union perf_event *event; __u32 id = 0; int err; @@ -303,6 +382,8 @@ int perf_event__synthesize_bpf_events(struct perf_session *session, event = malloc(sizeof(event->bpf) + KSYM_NAME_LEN + machine->id_hdr_size); if (!event) return -1; + + /* Synthesize all the bpf programs in system. */ while (true) { err = bpf_prog_get_next_id(id, &id); if (err) { @@ -335,6 +416,23 @@ int perf_event__synthesize_bpf_events(struct perf_session *session, break; } } + + /* Synthesize all the bpf images - trampolines/dispatchers. */ + if (symbol_conf.kallsyms_name != NULL) + kallsyms_filename = symbol_conf.kallsyms_name; + + arg = (struct kallsyms_parse) { + .event = event, + .process = process, + .machine = machine, + .tool = session->tool, + }; + + if (kallsyms__parse(kallsyms_filename, &arg, kallsyms_process_symbol)) { + pr_err("%s: failed to synthesize bpf images: %s\n", + __func__, strerror(errno)); + } + free(event); return err; }
Synthesize bpf images (trampolines/dispatchers) on start, as ksymbol events from /proc/kallsyms. Having this perf can recognize samples from those images and perf report and top shows them correctly. The rest of the ksymbol handling is already in place from for the bpf programs monitoring, so only the initial state was needed. perf report output: # Overhead Command Shared Object Symbol 12.37% test_progs [kernel.vmlinux] [k] entry_SYSCALL_64 11.80% test_progs [kernel.vmlinux] [k] syscall_return_via_sysret 9.63% test_progs bpf_prog_bcf7977d3b93787c_prog2 [k] bpf_prog_bcf7977d3b93787c_prog2 6.90% test_progs bpf_trampoline_24456 [k] bpf_trampoline_24456 6.36% test_progs [kernel.vmlinux] [k] memcpy_erms Signed-off-by: Jiri Olsa <jolsa@kernel.org> --- tools/perf/util/bpf-event.c | 98 +++++++++++++++++++++++++++++++++++++ 1 file changed, 98 insertions(+)