tools/perf/scripts/python/arm-cs-trace-disasm.py

   1 # SPDX-License-Identifier: GPL-2.0
   2 # arm-cs-trace-disasm.py: ARM CoreSight Trace Dump With Disassember
   3 #
   4 # Author: Tor Jeremiassen <tor@ti.com>
   5 #         Mathieu Poirier <mathieu.poirier@linaro.org>
   6 #         Leo Yan <leo.yan@linaro.org>
   7 #         Al Grant <Al.Grant@arm.com>
   8
   9 from __future__ import print_function
  10 import os
  11 from os import path
  12 import sys
  13 import re
  14 from subprocess import *
  15 from optparse import OptionParser, make_option
  16
  17 from perf_trace_context import perf_set_itrace_options, \
  18         perf_sample_insn, perf_sample_srccode
  19
  20 # Below are some example commands for using this script.
  21 #
  22 # Output disassembly with objdump:
  23 #  perf script -s scripts/python/arm-cs-trace-disasm.py \
  24 #               -- -d objdump -k path/to/vmlinux
  25 # Output disassembly with llvm-objdump:
  26 #  perf script -s scripts/python/arm-cs-trace-disasm.py \
  27 #               -- -d llvm-objdump-11 -k path/to/vmlinux
  28 # Output only source line and symbols:
  29 #  perf script -s scripts/python/arm-cs-trace-disasm.py
  30
  31 # Command line parsing.
  32 option_list = [
  33         # formatting options for the bottom entry of the stack
  34         make_option("-k", "--vmlinux", dest="vmlinux_name",
  35                     help="Set path to vmlinux file"),
  36         make_option("-d", "--objdump", dest="objdump_name",
  37                     help="Set path to objdump executable file"),
  38         make_option("-v", "--verbose", dest="verbose",
  39                     action="store_true", default=False,
  40                     help="Enable debugging log")
  41 ]
  42
  43 parser = OptionParser(option_list=option_list)
  44 (options, args) = parser.parse_args()
  45
  46 # Initialize global dicts and regular expression
  47 disasm_cache = dict()
  48 cpu_data = dict()
  49 disasm_re = re.compile("^\s*([0-9a-fA-F]+):")
  50 disasm_func_re = re.compile("^\s*([0-9a-fA-F]+)\s.*:")
  51 cache_size = 64*1024
  52
  53 glb_source_file_name    = None
  54 glb_line_number         = None
  55 glb_dso                 = None
  56
  57 def get_optional(perf_dict, field):
  58        if field in perf_dict:
  59                return perf_dict[field]
  60        return "[unknown]"
  61
  62 def get_offset(perf_dict, field):
  63         if field in perf_dict:
  64                 return "+%#x" % perf_dict[field]
  65         return ""
  66
  67 def get_dso_file_path(dso_name, dso_build_id):
  68         if (dso_name == "[kernel.kallsyms]" or dso_name == "vmlinux"):
  69                 if (options.vmlinux_name):
  70                         return options.vmlinux_name;
  71                 else:
  72                         return dso_name
  73
  74         if (dso_name == "[vdso]") :
  75                 append = "/vdso"
  76         else:
  77                 append = "/elf"
  78
  79         dso_path = os.environ['PERF_BUILDID_DIR'] + "/" + dso_name + "/" + dso_build_id + append;
  80         # Replace duplicate slash chars to single slash char
  81         dso_path = dso_path.replace('//', '/', 1)
  82         return dso_path
  83
  84 def read_disam(dso_fname, dso_start, start_addr, stop_addr):
  85         addr_range = str(start_addr) + ":" + str(stop_addr) + ":" + dso_fname
  86
  87         # Don't let the cache get too big, clear it when it hits max size
  88         if (len(disasm_cache) > cache_size):
  89                 disasm_cache.clear();
  90
  91         if addr_range in disasm_cache:
  92                 disasm_output = disasm_cache[addr_range];
  93         else:
  94                 start_addr = start_addr - dso_start;
  95                 stop_addr = stop_addr - dso_start;
  96                 disasm = [ options.objdump_name, "-d", "-z",
  97                            "--start-address="+format(start_addr,"#x"),
  98                            "--stop-address="+format(stop_addr,"#x") ]
  99                 disasm += [ dso_fname ]
 100                 disasm_output = check_output(disasm).decode('utf-8').split('\n')
 101                 disasm_cache[addr_range] = disasm_output
 102
 103         return disasm_output
 104
 105 def print_disam(dso_fname, dso_start, start_addr, stop_addr):
 106         for line in read_disam(dso_fname, dso_start, start_addr, stop_addr):
 107                 m = disasm_func_re.search(line)
 108                 if m is None:
 109                         m = disasm_re.search(line)
 110                         if m is None:
 111                                 continue
 112                 print("\t" + line)
 113
 114 def print_sample(sample):
 115         print("Sample = { cpu: %04d addr: 0x%016x phys_addr: 0x%016x ip: 0x%016x " \
 116               "pid: %d tid: %d period: %d time: %d }" % \
 117               (sample['cpu'], sample['addr'], sample['phys_addr'], \
 118                sample['ip'], sample['pid'], sample['tid'], \
 119                sample['period'], sample['time']))
 120
 121 def trace_begin():
 122         print('ARM CoreSight Trace Data Assembler Dump')
 123
 124 def trace_end():
 125         print('End')
 126
 127 def trace_unhandled(event_name, context, event_fields_dict):
 128         print(' '.join(['%s=%s'%(k,str(v))for k,v in sorted(event_fields_dict.items())]))
 129
 130 def common_start_str(comm, sample):
 131         sec = int(sample["time"] / 1000000000)
 132         ns = sample["time"] % 1000000000
 133         cpu = sample["cpu"]
 134         pid = sample["pid"]
 135         tid = sample["tid"]
 136         return "%16s %5u/%-5u [%04u] %9u.%09u  " % (comm, pid, tid, cpu, sec, ns)
 137
 138 # This code is copied from intel-pt-events.py for printing source code
 139 # line and symbols.
 140 def print_srccode(comm, param_dict, sample, symbol, dso):
 141         ip = sample["ip"]
 142         if symbol == "[unknown]":
 143                 start_str = common_start_str(comm, sample) + ("%x" % ip).rjust(16).ljust(40)
 144         else:
 145                 offs = get_offset(param_dict, "symoff")
 146                 start_str = common_start_str(comm, sample) + (symbol + offs).ljust(40)
 147
 148         global glb_source_file_name
 149         global glb_line_number
 150         global glb_dso
 151
 152         source_file_name, line_number, source_line = perf_sample_srccode(perf_script_context)
 153         if source_file_name:
 154                 if glb_line_number == line_number and glb_source_file_name == source_file_name:
 155                         src_str = ""
 156                 else:
 157                         if len(source_file_name) > 40:
 158                                 src_file = ("..." + source_file_name[-37:]) + " "
 159                         else:
 160                                 src_file = source_file_name.ljust(41)
 161
 162                         if source_line is None:
 163                                 src_str = src_file + str(line_number).rjust(4) + " <source not found>"
 164                         else:
 165                                 src_str = src_file + str(line_number).rjust(4) + " " + source_line
 166                 glb_dso = None
 167         elif dso == glb_dso:
 168                 src_str = ""
 169         else:
 170                 src_str = dso
 171                 glb_dso = dso
 172
 173         glb_line_number = line_number
 174         glb_source_file_name = source_file_name
 175
 176         print(start_str, src_str)
 177
 178 def process_event(param_dict):
 179         global cache_size
 180         global options
 181
 182         sample = param_dict["sample"]
 183         comm = param_dict["comm"]
 184
 185         name = param_dict["ev_name"]
 186         dso = get_optional(param_dict, "dso")
 187         dso_bid = get_optional(param_dict, "dso_bid")
 188         dso_start = get_optional(param_dict, "dso_map_start")
 189         dso_end = get_optional(param_dict, "dso_map_end")
 190         symbol = get_optional(param_dict, "symbol")
 191
 192         if (options.verbose == True):
 193                 print("Event type: %s" % name)
 194                 print_sample(sample)
 195
 196         # If cannot find dso so cannot dump assembler, bail out
 197         if (dso == '[unknown]'):
 198                 return
 199
 200         # Validate dso start and end addresses
 201         if ((dso_start == '[unknown]') or (dso_end == '[unknown]')):
 202                 print("Failed to find valid dso map for dso %s" % dso)
 203                 return
 204
 205         if (name[0:12] == "instructions"):
 206                 print_srccode(comm, param_dict, sample, symbol, dso)
 207                 return
 208
 209         # Don't proceed if this event is not a branch sample, .
 210         if (name[0:8] != "branches"):
 211                 return
 212
 213         cpu = sample["cpu"]
 214         ip = sample["ip"]
 215         addr = sample["addr"]
 216
 217         # Initialize CPU data if it's empty, and directly return back
 218         # if this is the first tracing event for this CPU.
 219         if (cpu_data.get(str(cpu) + 'addr') == None):
 220                 cpu_data[str(cpu) + 'addr'] = addr
 221                 return
 222
 223         # The format for packet is:
 224         #
 225         #                 +------------+------------+------------+
 226         #  sample_prev:   |    addr    |    ip      |    cpu     |
 227         #                 +------------+------------+------------+
 228         #  sample_next:   |    addr    |    ip      |    cpu     |
 229         #                 +------------+------------+------------+
 230         #
 231         # We need to combine the two continuous packets to get the instruction
 232         # range for sample_prev::cpu:
 233         #
 234         #     [ sample_prev::addr .. sample_next::ip ]
 235         #
 236         # For this purose, sample_prev::addr is stored into cpu_data structure
 237         # and read back for 'start_addr' when the new packet comes, and we need
 238         # to use sample_next::ip to calculate 'stop_addr', plusing extra 4 for
 239         # 'stop_addr' is for the sake of objdump so the final assembler dump can
 240         # include last instruction for sample_next::ip.
 241         start_addr = cpu_data[str(cpu) + 'addr']
 242         stop_addr  = ip + 4
 243
 244         # Record for previous sample packet
 245         cpu_data[str(cpu) + 'addr'] = addr
 246
 247         # Handle CS_ETM_TRACE_ON packet if start_addr=0 and stop_addr=4
 248         if (start_addr == 0 and stop_addr == 4):
 249                 print("CPU%d: CS_ETM_TRACE_ON packet is inserted" % cpu)
 250                 return
 251
 252         if (start_addr < int(dso_start) or start_addr > int(dso_end)):
 253                 print("Start address 0x%x is out of range [ 0x%x .. 0x%x ] for dso %s" % (start_addr, int(dso_start), int(dso_end), dso))
 254                 return
 255
 256         if (stop_addr < int(dso_start) or stop_addr > int(dso_end)):
 257                 print("Stop address 0x%x is out of range [ 0x%x .. 0x%x ] for dso %s" % (stop_addr, int(dso_start), int(dso_end), dso))
 258                 return
 259
 260         if (options.objdump_name != None):
 261                 # It doesn't need to decrease virtual memory offset for disassembly
 262                 # for kernel dso, so in this case we set vm_start to zero.
 263                 if (dso == "[kernel.kallsyms]"):
 264                         dso_vm_start = 0
 265                 else:
 266                         dso_vm_start = int(dso_start)
 267
 268                 dso_fname = get_dso_file_path(dso, dso_bid)
 269                 if path.exists(dso_fname):
 270                         print_disam(dso_fname, dso_vm_start, start_addr, stop_addr)
 271                 else:
 272                         print("Failed to find dso %s for address range [ 0x%x .. 0x%x ]" % (dso, start_addr, stop_addr))
 273
 274         print_srccode(comm, param_dict, sample, symbol, dso)