Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 1 | #!/usr/bin/env python |
| 2 | # |
| 3 | # trace Trace a function and print a trace message based on its |
| 4 | # parameters, with an optional filter. |
| 5 | # |
| 6 | # USAGE: trace [-h] [-p PID] [-v] [-Z STRING_SIZE] [-S] [-M MAX_EVENTS] [-o] |
| 7 | # probe [probe ...] |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 8 | # |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 9 | # Licensed under the Apache License, Version 2.0 (the "License") |
| 10 | # Copyright (C) 2016 Sasha Goldshtein. |
| 11 | |
| 12 | from bcc import BPF |
| 13 | from time import sleep, strftime |
| 14 | import argparse |
| 15 | import re |
| 16 | import ctypes as ct |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 17 | import multiprocessing |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 18 | import os |
| 19 | import traceback |
| 20 | import sys |
| 21 | |
| 22 | class Time(object): |
| 23 | # BPF timestamps come from the monotonic clock. To be able to filter |
| 24 | # and compare them from Python, we need to invoke clock_gettime. |
| 25 | # Adapted from http://stackoverflow.com/a/1205762 |
| 26 | CLOCK_MONOTONIC_RAW = 4 # see <linux/time.h> |
| 27 | |
| 28 | class timespec(ct.Structure): |
| 29 | _fields_ = [ |
| 30 | ('tv_sec', ct.c_long), |
| 31 | ('tv_nsec', ct.c_long) |
| 32 | ] |
| 33 | |
| 34 | librt = ct.CDLL('librt.so.1', use_errno=True) |
| 35 | clock_gettime = librt.clock_gettime |
| 36 | clock_gettime.argtypes = [ct.c_int, ct.POINTER(timespec)] |
| 37 | |
| 38 | @staticmethod |
| 39 | def monotonic_time(): |
| 40 | t = Time.timespec() |
| 41 | if Time.clock_gettime( |
| 42 | Time.CLOCK_MONOTONIC_RAW, ct.pointer(t)) != 0: |
| 43 | errno_ = ct.get_errno() |
| 44 | raise OSError(errno_, os.strerror(errno_)) |
| 45 | return t.tv_sec * 1e9 + t.tv_nsec |
| 46 | |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 47 | class Perf(object): |
| 48 | class perf_event_attr(ct.Structure): |
| 49 | _fields_ = [ |
| 50 | ('type', ct.c_uint), |
| 51 | ('size', ct.c_uint), |
| 52 | ('config', ct.c_ulong), |
| 53 | ('sample_period', ct.c_ulong), |
| 54 | ('sample_type', ct.c_ulong), |
| 55 | ('IGNORE1', ct.c_ulong), |
| 56 | ('IGNORE2', ct.c_ulong), |
| 57 | ('wakeup_events', ct.c_uint), |
| 58 | ('IGNORE3', ct.c_uint), |
| 59 | ('IGNORE4', ct.c_ulong), |
| 60 | ('IGNORE5', ct.c_ulong), |
| 61 | ('IGNORE6', ct.c_ulong), |
| 62 | ('IGNORE7', ct.c_uint), |
| 63 | ('IGNORE8', ct.c_int), |
| 64 | ('IGNORE9', ct.c_ulong), |
| 65 | ('IGNORE10', ct.c_uint), |
| 66 | ('IGNORE11', ct.c_uint) |
| 67 | ] |
| 68 | |
| 69 | NR_PERF_EVENT_OPEN = 298 |
| 70 | PERF_TYPE_TRACEPOINT = 2 |
| 71 | PERF_SAMPLE_RAW = 1024 |
| 72 | PERF_FLAG_FD_CLOEXEC = 8 |
| 73 | PERF_EVENT_IOC_SET_FILTER = 1074275334 |
| 74 | PERF_EVENT_IOC_ENABLE = 9216 |
| 75 | |
| 76 | libc = ct.CDLL('libc.so.6', use_errno=True) |
| 77 | syscall = libc.syscall # not declaring vararg types |
| 78 | ioctl = libc.ioctl # not declaring vararg types |
| 79 | |
| 80 | @staticmethod |
| 81 | def _open_for_cpu(cpu, attr): |
| 82 | pfd = Perf.syscall(Perf.NR_PERF_EVENT_OPEN, ct.byref(attr), |
| 83 | -1, cpu, -1, Perf.PERF_FLAG_FD_CLOEXEC) |
| 84 | if pfd < 0: |
| 85 | errno_ = ct.get_errno() |
| 86 | raise OSError(errno_, os.strerror(errno_)) |
| 87 | if Perf.ioctl(pfd, Perf.PERF_EVENT_IOC_SET_FILTER, |
| 88 | "common_pid == -17") < 0: |
| 89 | errno_ = ct.get_errno() |
| 90 | raise OSError(errno_, os.strerror(errno_)) |
| 91 | if Perf.ioctl(pfd, Perf.PERF_EVENT_IOC_ENABLE, 0) < 0: |
| 92 | errno_ = ct.get_errno() |
| 93 | raise OSError(errno_, os.strerror(errno_)) |
| 94 | |
| 95 | @staticmethod |
| 96 | def perf_event_open(tpoint_id): |
| 97 | attr = Perf.perf_event_attr() |
| 98 | attr.config = tpoint_id |
| 99 | attr.type = Perf.PERF_TYPE_TRACEPOINT |
| 100 | attr.sample_type = Perf.PERF_SAMPLE_RAW |
| 101 | attr.sample_period = 1 |
| 102 | attr.wakeup_events = 1 |
| 103 | for cpu in range(0, multiprocessing.cpu_count()): |
| 104 | Perf._open_for_cpu(cpu, attr) |
| 105 | |
| 106 | class Tracepoint(object): |
| 107 | tracepoints_enabled = 0 |
| 108 | trace_root = "/sys/kernel/debug/tracing" |
| 109 | event_root = os.path.join(trace_root, "events") |
| 110 | |
| 111 | @staticmethod |
| 112 | def generate_decl(): |
| 113 | if Tracepoint.tracepoints_enabled == 0: |
| 114 | return "" |
| 115 | return "\nBPF_HASH(__trace_di, u64, u64);\n" |
| 116 | |
| 117 | @staticmethod |
| 118 | def generate_entry_probe(): |
| 119 | if Tracepoint.tracepoints_enabled == 0: |
| 120 | return "" |
| 121 | return """ |
| 122 | int __trace_entry_update(struct pt_regs *ctx) |
| 123 | { |
| 124 | u64 tid = bpf_get_current_pid_tgid(); |
| 125 | u64 val = ctx->di; |
| 126 | __trace_di.update(&tid, &val); |
| 127 | return 0; |
| 128 | } |
| 129 | """ |
| 130 | |
| 131 | @staticmethod |
| 132 | def enable_tracepoint(category, event): |
| 133 | tp_id = Tracepoint.get_tpoint_id(category, event) |
| 134 | if tp_id == -1: |
| 135 | raise ValueError("no such tracepoint found: %s:%s" % |
| 136 | (category, event)) |
| 137 | Perf.perf_event_open(tp_id) |
| 138 | Tracepoint.tracepoints_enabled += 1 |
| 139 | |
| 140 | @staticmethod |
| 141 | def get_tpoint_id(category, event): |
| 142 | evt_dir = os.path.join(Tracepoint.event_root, category, event) |
| 143 | try: |
| 144 | return int( |
| 145 | open(os.path.join(evt_dir, "id")).read().strip()) |
| 146 | except: |
| 147 | return -1 |
| 148 | |
| 149 | @staticmethod |
| 150 | def get_tpoint_format(category, event): |
| 151 | evt_dir = os.path.join(Tracepoint.event_root, category, event) |
| 152 | try: |
| 153 | return open(os.path.join(evt_dir, "format")).readlines() |
| 154 | except: |
| 155 | return "" |
| 156 | |
| 157 | @staticmethod |
| 158 | def attach(bpf): |
| 159 | if Tracepoint.tracepoints_enabled > 0: |
| 160 | bpf.attach_kprobe(event="tracing_generic_entry_update", |
| 161 | fn_name="__trace_entry_update") |
| 162 | |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 163 | class Probe(object): |
| 164 | probe_count = 0 |
| 165 | max_events = None |
| 166 | event_count = 0 |
| 167 | first_ts = 0 |
| 168 | use_localtime = True |
| 169 | |
| 170 | @classmethod |
| 171 | def configure(cls, args): |
| 172 | cls.max_events = args.max_events |
| 173 | cls.use_localtime = not args.offset |
| 174 | cls.first_ts = Time.monotonic_time() |
| 175 | |
| 176 | def __init__(self, probe, string_size): |
| 177 | self.raw_probe = probe |
| 178 | self.string_size = string_size |
| 179 | Probe.probe_count += 1 |
| 180 | self._parse_probe() |
| 181 | self.probe_num = Probe.probe_count |
| 182 | self.probe_name = "probe_%s_%d" % \ |
| 183 | (self.function, self.probe_num) |
| 184 | |
| 185 | def __str__(self): |
| 186 | return "%s:%s`%s FLT=%s ACT=%s/%s" % (self.probe_type, |
| 187 | self.library, self.function, self.filter, |
| 188 | self.types, self.values) |
| 189 | |
| 190 | def is_default_action(self): |
| 191 | return self.python_format == "" |
| 192 | |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 193 | def _generate_tpoint_entry_struct_fields(self): |
| 194 | format_lines = Tracepoint.get_tpoint_format(self.tp_category, |
| 195 | self.tp_event) |
| 196 | text = "" |
| 197 | for line in format_lines: |
| 198 | match = re.search(r'field:([^;]*);.*size:(\d+);', line) |
| 199 | if match is None: |
| 200 | continue |
| 201 | parts = match.group(1).split() |
| 202 | field_name = parts[-1:][0] |
| 203 | field_type = " ".join(parts[:-1]) |
| 204 | field_size = int(match.group(2)) |
| 205 | if "__data_loc" in field_type: |
| 206 | continue |
| 207 | if field_name.startswith("common_"): |
| 208 | continue |
| 209 | text += " %s %s;\n" % (field_type, field_name) |
| 210 | return text |
| 211 | |
| 212 | def _generate_tpoint_entry_struct(self): |
| 213 | text = """ |
| 214 | struct %s { |
| 215 | u64 __do_not_use__; |
| 216 | %s |
| 217 | }; |
| 218 | """ |
| 219 | self.tp_entry_struct_name = self.probe_name + \ |
| 220 | "_trace_entry" |
| 221 | fields = self._generate_tpoint_entry_struct_fields() |
| 222 | return text % (self.tp_entry_struct_name, fields) |
| 223 | |
| 224 | def _generate_tpoint_entry_prefix(self): |
| 225 | text = """ |
| 226 | u64 tid = bpf_get_current_pid_tgid(); |
| 227 | u64 *di = __trace_di.lookup(&tid); |
| 228 | if (di == 0) { return 0; } |
| 229 | struct %s tp = {}; |
| 230 | bpf_probe_read(&tp, sizeof(tp), (void *)*di); |
| 231 | """ % self.tp_entry_struct_name |
| 232 | return text |
| 233 | |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 234 | def _bail(self, error): |
| 235 | raise ValueError("error parsing probe '%s': %s" % |
| 236 | (self.raw_probe, error)) |
| 237 | |
| 238 | def _parse_probe(self): |
| 239 | text = self.raw_probe |
| 240 | |
| 241 | # Everything until the first space is the probe specifier |
| 242 | first_space = text.find(' ') |
| 243 | spec = text[:first_space] if first_space >= 0 else text |
| 244 | self._parse_spec(spec) |
| 245 | if first_space >= 0: |
| 246 | text = text[first_space:].lstrip() |
| 247 | else: |
| 248 | text = "" |
| 249 | |
| 250 | # If we now have a (, wait for the balanced closing ) and that |
| 251 | # will be the predicate |
| 252 | self.filter = None |
| 253 | if len(text) > 0 and text[0] == "(": |
| 254 | balance = 1 |
| 255 | for i in range(1, len(text)): |
| 256 | if text[i] == "(": |
| 257 | balance += 1 |
| 258 | if text[i] == ")": |
| 259 | balance -= 1 |
| 260 | if balance == 0: |
| 261 | self._parse_filter(text[:i+1]) |
| 262 | text = text[i+1:] |
| 263 | break |
| 264 | if self.filter is None: |
| 265 | self._bail("unmatched end of predicate") |
| 266 | |
| 267 | if self.filter is None: |
| 268 | self.filter = "1" |
| 269 | |
| 270 | # The remainder of the text is the printf action |
| 271 | self._parse_action(text.lstrip()) |
| 272 | |
| 273 | def _parse_spec(self, spec): |
| 274 | parts = spec.split(":") |
| 275 | # Two special cases: 'func' means 'p::func', 'lib:func' means |
| 276 | # 'p:lib:func'. Other combinations need to provide an empty |
| 277 | # value between delimiters, e.g. 'r::func' for a kretprobe on |
| 278 | # the function func. |
| 279 | if len(parts) == 1: |
| 280 | parts = ["p", "", parts[0]] |
| 281 | elif len(parts) == 2: |
| 282 | parts = ["p", parts[0], parts[1]] |
| 283 | if len(parts[0]) == 0: |
| 284 | self.probe_type = "p" |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 285 | elif parts[0] in ["p", "r", "t"]: |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 286 | self.probe_type = parts[0] |
| 287 | else: |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 288 | self._bail("expected '', 'p', 't', or 'r', got '%s'" % |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 289 | parts[0]) |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 290 | if self.probe_type == "t": |
| 291 | self.tp_category = parts[1] |
| 292 | self.tp_event = parts[2] |
| 293 | Tracepoint.enable_tracepoint(self.tp_category, |
| 294 | self.tp_event) |
| 295 | self.library = "" # kernel |
| 296 | self.function = "perf_trace_%s" % self.tp_event |
| 297 | else: |
| 298 | self.library = parts[1] |
| 299 | self.function = parts[2] |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 300 | |
| 301 | def _parse_filter(self, filt): |
| 302 | self.filter = self._replace_args(filt) |
| 303 | |
| 304 | def _parse_types(self, fmt): |
| 305 | for match in re.finditer( |
| 306 | r'[^%]%(s|u|d|llu|lld|hu|hd|x|llx|c)', fmt): |
| 307 | self.types.append(match.group(1)) |
| 308 | fmt = re.sub(r'([^%]%)(u|d|llu|lld|hu|hd)', r'\1d', fmt) |
| 309 | fmt = re.sub(r'([^%]%)(x|llx)', r'\1x', fmt) |
| 310 | self.python_format = fmt.strip('"') |
| 311 | |
| 312 | def _parse_action(self, action): |
| 313 | self.values = [] |
| 314 | self.types = [] |
| 315 | self.python_format = "" |
| 316 | if len(action) == 0: |
| 317 | return |
| 318 | |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 319 | action = action.strip() |
| 320 | match = re.search(r'(\".*\"),?(.*)', action) |
| 321 | if match is None: |
| 322 | self._bail("expected format string in \"s") |
| 323 | |
| 324 | self.raw_format = match.group(1) |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 325 | self._parse_types(self.raw_format) |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 326 | for part in match.group(2).split(','): |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 327 | part = self._replace_args(part) |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 328 | if len(part) > 0: |
| 329 | self.values.append(part) |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 330 | |
| 331 | aliases = { |
| 332 | "retval": "ctx->ax", |
| 333 | "arg1": "ctx->di", |
| 334 | "arg2": "ctx->si", |
| 335 | "arg3": "ctx->dx", |
| 336 | "arg4": "ctx->cx", |
| 337 | "arg5": "ctx->r8", |
| 338 | "arg6": "ctx->r9", |
| 339 | "$uid": "(unsigned)(bpf_get_current_uid_gid() & 0xffffffff)", |
| 340 | "$gid": "(unsigned)(bpf_get_current_uid_gid() >> 32)", |
| 341 | "$pid": "(unsigned)(bpf_get_current_pid_tgid() & 0xffffffff)", |
| 342 | "$tgid": "(unsigned)(bpf_get_current_pid_tgid() >> 32)", |
| 343 | "$cpu": "bpf_get_smp_processor_id()" |
| 344 | } |
| 345 | |
| 346 | def _replace_args(self, expr): |
| 347 | for alias, replacement in Probe.aliases.items(): |
| 348 | expr = expr.replace(alias, replacement) |
| 349 | return expr |
| 350 | |
| 351 | p_type = { "u": ct.c_uint, "d": ct.c_int, |
| 352 | "llu": ct.c_ulonglong, "lld": ct.c_longlong, |
| 353 | "hu": ct.c_ushort, "hd": ct.c_short, |
| 354 | "x": ct.c_uint, "llx": ct.c_ulonglong, |
| 355 | "c": ct.c_ubyte } |
| 356 | |
| 357 | def _generate_python_field_decl(self, idx, fields): |
| 358 | field_type = self.types[idx] |
| 359 | if field_type == "s": |
| 360 | ptype = ct.c_char * self.string_size |
| 361 | else: |
| 362 | ptype = Probe.p_type[field_type] |
| 363 | fields.append(("v%d" % idx, ptype)) |
| 364 | |
| 365 | def _generate_python_data_decl(self): |
| 366 | self.python_struct_name = "%s_%d_Data" % \ |
| 367 | (self.function, self.probe_num) |
| 368 | fields = [ |
| 369 | ("timestamp_ns", ct.c_ulonglong), |
| 370 | ("pid", ct.c_uint), |
| 371 | ("comm", ct.c_char * 16) # TASK_COMM_LEN |
| 372 | ] |
| 373 | for i in range(0, len(self.types)): |
| 374 | self._generate_python_field_decl(i, fields) |
| 375 | return type(self.python_struct_name, (ct.Structure,), |
| 376 | dict(_fields_=fields)) |
| 377 | |
| 378 | c_type = { "u": "unsigned int", "d": "int", |
| 379 | "llu": "unsigned long long", "lld": "long long", |
| 380 | "hu": "unsigned short", "hd": "short", |
| 381 | "x": "unsigned int", "llx": "unsigned long long", |
| 382 | "c": "char" } |
| 383 | fmt_types = c_type.keys() |
| 384 | |
| 385 | def _generate_field_decl(self, idx): |
| 386 | field_type = self.types[idx] |
| 387 | if field_type == "s": |
| 388 | return "char v%d[%d];\n" % (idx, self.string_size) |
| 389 | if field_type in Probe.fmt_types: |
| 390 | return "%s v%d;\n" % (Probe.c_type[field_type], idx) |
| 391 | self._bail("unrecognized format specifier %s" % field_type) |
| 392 | |
| 393 | def _generate_data_decl(self): |
| 394 | # The BPF program will populate values into the struct |
| 395 | # according to the format string, and the Python program will |
| 396 | # construct the final display string. |
| 397 | self.events_name = "%s_events" % self.probe_name |
| 398 | self.struct_name = "%s_data_t" % self.probe_name |
| 399 | |
| 400 | data_fields = "" |
| 401 | for i, field_type in enumerate(self.types): |
| 402 | data_fields += " " + \ |
| 403 | self._generate_field_decl(i) |
| 404 | |
| 405 | text = """ |
| 406 | struct %s |
| 407 | { |
| 408 | u64 timestamp_ns; |
| 409 | u32 pid; |
| 410 | char comm[TASK_COMM_LEN]; |
| 411 | %s |
| 412 | }; |
| 413 | |
| 414 | BPF_PERF_OUTPUT(%s); |
| 415 | """ |
| 416 | return text % (self.struct_name, data_fields, self.events_name) |
| 417 | |
| 418 | def _generate_field_assign(self, idx): |
| 419 | field_type = self.types[idx] |
| 420 | expr = self.values[idx] |
| 421 | if field_type == "s": |
| 422 | return """ |
| 423 | if (%s != 0) { |
| 424 | bpf_probe_read(&__data.v%d, sizeof(__data.v%d), (void *)%s); |
| 425 | } |
| 426 | """ % (expr, idx, idx, expr) |
| 427 | # return ("bpf_probe_read(&__data.v%d, " + \ |
| 428 | # "sizeof(__data.v%d), (char*)%s);\n") % (idx, idx, expr) |
| 429 | # return ("__builtin_memcpy(&__data.v%d, (void *)%s, " + \ |
| 430 | # "sizeof(__data.v%d));\n") % (idx, expr, idx) |
| 431 | if field_type in Probe.fmt_types: |
| 432 | return " __data.v%d = (%s)%s;\n" % \ |
| 433 | (idx, Probe.c_type[field_type], expr) |
| 434 | self._bail("unrecognized field type %s" % field_type) |
| 435 | |
| 436 | def generate_program(self, pid, include_self): |
| 437 | data_decl = self._generate_data_decl() |
| 438 | self.pid = pid |
| 439 | # kprobes don't have built-in pid filters, so we have to add |
| 440 | # it to the function body: |
| 441 | if len(self.library) == 0 and pid != -1: |
| 442 | pid_filter = """ |
| 443 | u32 __pid = bpf_get_current_pid_tgid(); |
| 444 | if (__pid != %d) { return 0; } |
| 445 | """ % pid |
| 446 | elif not include_self: |
| 447 | pid_filter = """ |
| 448 | u32 __pid = bpf_get_current_pid_tgid(); |
| 449 | if (__pid == %d) { return 0; } |
| 450 | """ % os.getpid() |
| 451 | else: |
| 452 | pid_filter = "" |
| 453 | |
| 454 | data_fields = "" |
| 455 | for i, expr in enumerate(self.values): |
| 456 | data_fields += self._generate_field_assign(i) |
| 457 | |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 458 | prefix = "" |
| 459 | if self.probe_type == "t": |
| 460 | data_decl += self._generate_tpoint_entry_struct() |
| 461 | prefix = self._generate_tpoint_entry_prefix() |
| 462 | |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 463 | text = """ |
| 464 | int %s(struct pt_regs *ctx) |
| 465 | { |
| 466 | %s |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 467 | %s |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 468 | if (!(%s)) return 0; |
| 469 | |
| 470 | struct %s __data = {0}; |
| 471 | __data.timestamp_ns = bpf_ktime_get_ns(); |
| 472 | __data.pid = bpf_get_current_pid_tgid(); |
| 473 | bpf_get_current_comm(&__data.comm, sizeof(__data.comm)); |
| 474 | %s |
| 475 | %s.perf_submit(ctx, &__data, sizeof(__data)); |
| 476 | return 0; |
| 477 | } |
| 478 | """ |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 479 | text = text % (self.probe_name, pid_filter, prefix, |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 480 | self.filter, self.struct_name, |
| 481 | data_fields, self.events_name) |
| 482 | |
| 483 | return data_decl + "\n" + text |
| 484 | |
| 485 | @classmethod |
| 486 | def _time_off_str(cls, timestamp_ns): |
| 487 | return "%.6f" % (1e-9 * (timestamp_ns - cls.first_ts)) |
| 488 | |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 489 | auto_includes = { |
| 490 | "linux/time.h" : ["time"], |
| 491 | "linux/fs.h" : ["fs", "file"], |
| 492 | "linux/blkdev.h" : ["bio", "request"], |
| 493 | "linux/slab.h" : ["alloc"], |
| 494 | "linux/netdevice.h" : ["sk_buff"] |
| 495 | } |
| 496 | |
| 497 | @classmethod |
| 498 | def generate_auto_includes(cls, probes): |
| 499 | headers = "" |
| 500 | for header, keywords in cls.auto_includes.items(): |
| 501 | for keyword in keywords: |
| 502 | for probe in probes: |
| 503 | if keyword in probe: |
| 504 | headers += "#include <%s>\n" \ |
| 505 | % header |
| 506 | return headers |
| 507 | |
| 508 | def _display_function(self): |
| 509 | if self.probe_type != 't': |
| 510 | return self.function |
| 511 | else: |
| 512 | return self.function.replace("perf_trace_", "") |
| 513 | |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 514 | def print_event(self, cpu, data, size): |
| 515 | # Cast as the generated structure type and display |
| 516 | # according to the format string in the probe. |
| 517 | event = ct.cast(data, ct.POINTER(self.python_struct)).contents |
| 518 | values = map(lambda i: getattr(event, "v%d" % i), |
| 519 | range(0, len(self.values))) |
| 520 | msg = self.python_format % tuple(values) |
| 521 | time = strftime("%H:%M:%S") if Probe.use_localtime else \ |
| 522 | Probe._time_off_str(event.timestamp_ns) |
| 523 | print("%-8s %-6d %-12s %-16s %s" % \ |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 524 | (time[:8], event.pid, event.comm[:12], |
| 525 | self._display_function(), msg)) |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 526 | |
| 527 | Probe.event_count += 1 |
| 528 | if Probe.max_events is not None and \ |
| 529 | Probe.event_count >= Probe.max_events: |
| 530 | exit() |
| 531 | |
| 532 | def attach(self, bpf, verbose): |
| 533 | if len(self.library) == 0: |
| 534 | self._attach_k(bpf) |
| 535 | else: |
| 536 | self._attach_u(bpf) |
| 537 | self.python_struct = self._generate_python_data_decl() |
| 538 | bpf[self.events_name].open_perf_buffer(self.print_event) |
| 539 | |
| 540 | def _attach_k(self, bpf): |
| 541 | if self.probe_type == "r": |
| 542 | bpf.attach_kretprobe(event=self.function, |
| 543 | fn_name=self.probe_name) |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 544 | elif self.probe_type == "p" or self.probe_type == "t": |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 545 | bpf.attach_kprobe(event=self.function, |
| 546 | fn_name=self.probe_name) |
| 547 | |
| 548 | def _attach_u(self, bpf): |
| 549 | libpath = BPF.find_library(self.library) |
| 550 | if libpath is None: |
| 551 | # This might be an executable (e.g. 'bash') |
| 552 | with os.popen("/usr/bin/which %s 2>/dev/null" % |
| 553 | self.library) as w: |
| 554 | libpath = w.read().strip() |
| 555 | if libpath is None or len(libpath) == 0: |
| 556 | self._bail("unable to find library %s" % self.library) |
| 557 | |
| 558 | if self.probe_type == "r": |
| 559 | bpf.attach_uretprobe(name=libpath, |
| 560 | sym=self.function, |
| 561 | fn_name=self.probe_name, |
| 562 | pid=self.pid) |
| 563 | else: |
| 564 | bpf.attach_uprobe(name=libpath, |
| 565 | sym=self.function, |
| 566 | fn_name=self.probe_name, |
| 567 | pid=self.pid) |
| 568 | |
| 569 | class Tool(object): |
| 570 | examples = """ |
| 571 | EXAMPLES: |
| 572 | |
Sasha Goldshtein | 8acd015 | 2016-02-22 02:25:03 -0800 | [diff] [blame] | 573 | trace do_sys_open |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 574 | Trace the open syscall and print a default trace message when entered |
Sasha Goldshtein | 8acd015 | 2016-02-22 02:25:03 -0800 | [diff] [blame] | 575 | trace 'do_sys_open "%s", arg2' |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 576 | Trace the open syscall and print the filename being opened |
Sasha Goldshtein | 8acd015 | 2016-02-22 02:25:03 -0800 | [diff] [blame] | 577 | trace 'sys_read (arg3 > 20000) "read %d bytes", arg3' |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 578 | Trace the read syscall and print a message for reads >20000 bytes |
| 579 | trace 'r::do_sys_return "%llx", retval' |
| 580 | Trace the return from the open syscall and print the return value |
Sasha Goldshtein | 8acd015 | 2016-02-22 02:25:03 -0800 | [diff] [blame] | 581 | trace 'c:open (arg2 == 42) "%s %d", arg1, arg2' |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 582 | Trace the open() call from libc only if the flags (arg2) argument is 42 |
Sasha Goldshtein | 8acd015 | 2016-02-22 02:25:03 -0800 | [diff] [blame] | 583 | trace 'c:malloc "size = %d", arg1' |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 584 | Trace malloc calls and print the size being allocated |
Sasha Goldshtein | 8acd015 | 2016-02-22 02:25:03 -0800 | [diff] [blame] | 585 | trace 'p:c:write (arg1 == 1) "writing %d bytes to STDOUT", arg3' |
| 586 | Trace the write() call from libc to monitor writes to STDOUT |
| 587 | trace 'r::__kmalloc (retval == 0) "kmalloc failed!" |
| 588 | Trace returns from __kmalloc which returned a null pointer |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 589 | trace 'r:c:malloc (retval) "allocated = %p", retval |
| 590 | Trace returns from malloc and print non-NULL allocated buffers |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 591 | trace 't:block:block_rq_complete "sectors=%d", tp.nr_sector' |
| 592 | Trace the block_rq_complete kernel tracepoint and print # of tx sectors |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 593 | """ |
| 594 | |
| 595 | def __init__(self): |
| 596 | parser = argparse.ArgumentParser(description= |
| 597 | "Attach to functions and print trace messages.", |
| 598 | formatter_class=argparse.RawDescriptionHelpFormatter, |
| 599 | epilog=Tool.examples) |
| 600 | parser.add_argument("-p", "--pid", type=int, |
| 601 | help="id of the process to trace (optional)") |
| 602 | parser.add_argument("-v", "--verbose", action="store_true", |
| 603 | help="print resulting BPF program code before executing") |
| 604 | parser.add_argument("-Z", "--string-size", type=int, |
| 605 | default=80, help="maximum size to read from strings") |
| 606 | parser.add_argument("-S", "--include-self", action="store_true", |
| 607 | help="do not filter trace's own pid from the trace") |
| 608 | parser.add_argument("-M", "--max-events", type=int, |
| 609 | help="number of events to print before quitting") |
| 610 | parser.add_argument("-o", "--offset", action="store_true", |
| 611 | help="use relative time from first traced message") |
| 612 | parser.add_argument(metavar="probe", dest="probes", nargs="+", |
| 613 | help="probe specifier (see examples)") |
| 614 | self.args = parser.parse_args() |
| 615 | |
| 616 | def _create_probes(self): |
| 617 | Probe.configure(self.args) |
| 618 | self.probes = [] |
| 619 | for probe_spec in self.args.probes: |
| 620 | self.probes.append(Probe( |
| 621 | probe_spec, self.args.string_size)) |
| 622 | |
| 623 | def _generate_program(self): |
| 624 | self.program = """ |
| 625 | #include <linux/ptrace.h> |
| 626 | #include <linux/sched.h> /* For TASK_COMM_LEN */ |
| 627 | |
| 628 | """ |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 629 | self.program += Probe.generate_auto_includes( |
| 630 | map(lambda p: p.raw_probe, self.probes)) |
| 631 | self.program += Tracepoint.generate_decl() |
| 632 | self.program += Tracepoint.generate_entry_probe() |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 633 | for probe in self.probes: |
| 634 | self.program += probe.generate_program( |
| 635 | self.args.pid or -1, self.args.include_self) |
| 636 | |
| 637 | if self.args.verbose: |
| 638 | print(self.program) |
| 639 | |
| 640 | def _attach_probes(self): |
| 641 | self.bpf = BPF(text=self.program) |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 642 | Tracepoint.attach(self.bpf) |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 643 | for probe in self.probes: |
| 644 | if self.args.verbose: |
| 645 | print(probe) |
| 646 | probe.attach(self.bpf, self.args.verbose) |
| 647 | |
| 648 | def _main_loop(self): |
| 649 | all_probes_trivial = all(map(Probe.is_default_action, |
| 650 | self.probes)) |
| 651 | |
| 652 | # Print header |
| 653 | print("%-8s %-6s %-12s %-16s %s" % \ |
| 654 | ("TIME", "PID", "COMM", "FUNC", |
| 655 | "-" if not all_probes_trivial else "")) |
| 656 | |
| 657 | while True: |
| 658 | self.bpf.kprobe_poll() |
| 659 | |
| 660 | def run(self): |
| 661 | try: |
| 662 | self._create_probes() |
| 663 | self._generate_program() |
| 664 | self._attach_probes() |
| 665 | self._main_loop() |
| 666 | except: |
| 667 | if self.args.verbose: |
| 668 | traceback.print_exc() |
Sasha Goldshtein | fd60d55 | 2016-03-01 12:15:34 -0800 | [diff] [blame^] | 669 | elif sys.exc_type is not SystemExit: |
Sasha Goldshtein | 38847f0 | 2016-02-22 02:19:24 -0800 | [diff] [blame] | 670 | print(sys.exc_value) |
| 671 | |
| 672 | if __name__ == "__main__": |
| 673 | Tool().run() |