SimplePerf C++

承接上文肴敛，本文主要記錄simpleperf C++部分的代碼的閱讀筆記毕箍。

Main

• main main.cpp
  • RunSimpleperfCmd commond.cpp
    • CreateCommandInstance()
      • command->Run(args)

RecordCommand

RecordCommand : Commond

• Run()
  • CheckPerfEventLimit() // 系統(tǒng)默認(rèn)會(huì)限制用戶開(kāi)啟perf權(quán)限
  • setrlimit(RLIMIT_NOFILE) // 設(shè)置打開(kāi)文件句柄的限制為最大值
  • RunInAppContext // 使用run-as在被監(jiān)控的app權(quán)限下重新運(yùn)行simpleperf
    • run-as packagename
    • cp simpleperf to /data/data/packagename
    • shell ./simpleperf
    • pkill simpleperf
    • cp output_file to output_path
  • PrepareRecording // 準(zhǔn)備錄制
    • PrepareVdsoFile // vdso是一個(gè)特殊的so文件, 由內(nèi)核映射到所有的進(jìn)程中
    • event_selection_set_.AddEventType // Add default event type.
    • event_selection_set_.SetSampleSpeed // 設(shè)置采樣率
    • ExcludeKernel
    • TraceOffCpu // 默認(rèn)指統(tǒng)計(jì)在cpu上事件, 如果開(kāi)啟traceOffcpu則記錄off cpu的所有時(shí)間, 包括在內(nèi)核里等待執(zhí)行的時(shí)間
    • SetEventSelectionFlags
      • event_selection_set_.SetBranchSimpling
      • event_selection_set_.EnableFpCallChainSampling() // 開(kāi)啟FP的堆楌妫回溯模式
      • or event_selection_set_.EnableDwarfCallChainSampling() // 或者開(kāi)啟dwarf模式的堆棧回溯
      • event_selection_set_.SetInherit()
      • event_selection_set_.SetClockId() // 設(shè)置指定的CPU
    • offline_unwinder_.reset(new OfflineUnwinder()) // TODO
    • callchain_joiner // TODO
    • event_selection_set_.AddMonitoredProcesses(pids) // Add monitored targets. 設(shè)置監(jiān)控的進(jìn)程pid
    • event_selection_set_.OpenEventFiles(cpus) // Open perf event files 打開(kāi)perf的event files
    • event_selection_set_.MMapEventFiles() // create mapped buffers. 映射內(nèi)核的循環(huán)隊(duì)列緩存(內(nèi)核會(huì)往這個(gè)buffer里寫event, smipleperf會(huì)去讀event)
    • event_selection_set_.PrepareToReadMmapEventData(callback) // 設(shè)置讀取event的回調(diào)函數(shù)為 RecordCommand::ProcessRecord
    • CreateAndInitRecordFile() // 創(chuàng)建 perf.data. 文件
    • event_selection_set_.GetIOEventLoop()->AddSignalEvents( SIGCHLD, SIGINT, SIGTERM ) // Ctrl+C退出模式, 監(jiān)控退出的信號(hào)量
    • or event_selection_set_.GetIOEventLoop()->AddPeriodicEvent(duration_in_sec_) // duration自動(dòng)退出模式, 時(shí)長(zhǎng)到了自動(dòng)退出
    • jit_debug_reader_ // TODO
  • ProcessRecord // 錄制過(guò)程中數(shù)據(jù)的回調(diào)函數(shù)
    • UpdateRecord
      • UpdateMmapRecord if record type == PERF_RECORD_MMAP // TODO
      • UpdateMmapRecord if record type == PERF_RECORD_MMAP2
      • record->SetCommandName if record type == PERF_RECORD_COMM
    • ShouldOmitRecord() // TODO
    • jit_debug_reader_->UpdateRecord() // TODO
    • SaveRecord
      • SaveRecordForPostUnwinding()
        • record_file_writer_->WriteRecord()
      • or SaveRecordAfterUnwinding()
        • record.AdjustCallChainGeneratedByKernel()
        • UnwindRecord()
          • offline_unwinder_->UnwindCallChain() // 回溯堆棧
          • r.ReplaceRegAndStackWithcallChain(ips) // TODO
        • ExcludeKernelCallChain()
        • thread_tree_.Update() // TODO
        • record_file_writer_->WriteRecord()
      • or SaveRecordWithoutUnwinding()
        • record.AdjustCallChainGeneratedByKernel()
        • ExcludeKernelCallChain()
        • record_file_writer_->WriteRecord()
• DoRecording // 開(kāi)始錄制
  • write "STARTED" to start_profiling_fd_ file // 開(kāi)始前往參數(shù)里的start_profiling文件寫入一個(gè)字符串 "STARTED"并立刻關(guān)閉該文件, 標(biāo)記已經(jīng)開(kāi)啟錄制
  • event_selection_set_.GetIOEventLoop()->RunLoop() // 開(kāi)始IO Loop
  • event_selection_set_.FinishReadMMapEventData() // Loop結(jié)束后, 完成停止讀取buffer
• PostPrecessRecording // 結(jié)束錄制
  • PostUnwindRecords() // Post unwind dwarf callchain 默認(rèn)是在錄制的過(guò)程中回溯堆棧, 會(huì)有時(shí)間上的消耗, 在post-unwind模式下是在錄制完成后再去回溯堆棧.
  • JoinCallChains() // Optionally join Callchains.
  • DumpAdditionalFeatures(args) // Dump additional features, and close record file.
  • record_file_writer_->Close() // 關(guān)閉報(bào)告文件
  • print result // Show brief record result 打印簡(jiǎn)要的報(bào)告結(jié)果

EventSelectionSet

AddEventType

SetSampleFreq

OpenEventFiles

ReadCounters

PrepareToReadMmapEventData

FinishedReadMmapEventData

OfflineUnwinder

UnwindCallChain()

使用 libunwindstack 庫(kù)來(lái)回溯堆棧: https://android.googlesource.com/platform/system/core/+/master/libunwindstack/tools/
保存堆棧的sp和pc

perf

perf_record_type

• PERF_RECORD_MMAP 映射so事件, 用于關(guān)聯(lián)IP到代碼, 返回so映射內(nèi)存的地址, 長(zhǎng)度, 名稱等信息
• PERF_RECORD_MMAP2 mmap調(diào)用事件, executable mappings
• PERF_RECORD_COMM 進(jìn)程名改變事件, 包括新的進(jìn)程名稱等信息
• PERF_RECORD_EXIT 進(jìn)程退出事件
• PERF_RECORD_SAMPLE sample事件, 包括sample數(shù)據(jù)

參考資料: http://www.man7.org/linux/man-pages/man2/perf_event_open.2.html

sample事件:

struct {
  struct perf_event_header header;
  u64    sample_id;   /* if PERF_SAMPLE_IDENTIFIER */
  u64    ip;          /* if PERF_SAMPLE_IP */
  u32    pid, tid;    /* if PERF_SAMPLE_TID */
  u64    time;        /* if PERF_SAMPLE_TIME */
  u64    addr;        /* if PERF_SAMPLE_ADDR */
  u64    id;          /* if PERF_SAMPLE_ID */
  u64    stream_id;   /* if PERF_SAMPLE_STREAM_ID */
  u32    cpu, res;    /* if PERF_SAMPLE_CPU */
  u64    period;      /* if PERF_SAMPLE_PERIOD */
  struct read_format v; /* if PERF_SAMPLE_READ */
  u64    nr;          /* if PERF_SAMPLE_CALLCHAIN */
  u64    ips[nr];     /* if PERF_SAMPLE_CALLCHAIN */
  u32    size;        /* if PERF_SAMPLE_RAW */
  char  data[size];   /* if PERF_SAMPLE_RAW */
  u64    bnr;         /* if PERF_SAMPLE_BRANCH_STACK */
  struct perf_branch_entry lbr[bnr]; /* if PERF_SAMPLE_BRANCH_STACK */
  u64    abi;         /* if PERF_SAMPLE_REGS_USER */
  u64    regs[weight(mask)]; /* if PERF_SAMPLE_REGS_USER */
  u64    size;        /* if PERF_SAMPLE_STACK_USER */
  char   data[size];  /* if PERF_SAMPLE_STACK_USER */
  u64    dyn_size;    /* if PERF_SAMPLE_STACK_USER && size != 0 */
  u64    weight;      /* if PERF_SAMPLE_WEIGHT */
  u64    data_src;    /* if PERF_SAMPLE_DATA_SRC */
  u64    transaction; /* if PERF_SAMPLE_TRANSACTION */
  u64    abi;         /* if PERF_SAMPLE_REGS_INTR */
  u64    regs[weight(mask)]; /* if PERF_SAMPLE_REGS_INTR */
};

record

開(kāi)始錄制

simpleperf record

參數(shù):

#if defined(__ANDROID__)
"--app package_name    Profile the process of an Android application.\n"
"                      On non-rooted devices, the app must be debuggable,\n"
"                      because we use run-as to switch to the app's context.\n"
#endif
"-p pid1,pid2,...       Record events on existing processes. Mutually exclusive\n"
"                       with -a.\n"
"-t tid1,tid2,... Record events on existing threads. Mutually exclusive with -a.\n"
"\n"
"Select monitored event types:\n"
"-e event1[:modifier1],event2[:modifier2],...\n"
"             Select a list of events to record. An event can be:\n"
"               1) an event name listed in `simpleperf list`;\n"
"               2) a raw PMU event in rN format. N is a hex number.\n"
"                  For example, r1b selects event number 0x1b.\n"
"             Modifiers can be added to define how the event should be\n"
"             monitored. Possible modifiers are:\n"
"                u - monitor user space events only\n"
"                k - monitor kernel space events only\n"
"--group event1[:modifier],event2[:modifier2],...\n"
"             Similar to -e option. But events specified in the same --group\n"
"             option are monitored as a group, and scheduled in and out at the\n"
"             same time.\n"
"--trace-offcpu   Generate samples when threads are scheduled off cpu.\n"
"                 Similar to \"-c 1 -e sched:sched_switch\".\n"
"\n"
"Select monitoring options:\n"
"-f freq      Set event sample frequency. It means recording at most [freq]\n"
"             samples every second. For non-tracepoint events, the default\n"
"             option is -f 4000. A -f/-c option affects all event types\n"
"             following it until meeting another -f/-c option. For example,\n"
"             for \"-f 1000 cpu-cycles -c 1 -e sched:sched_switch\", cpu-cycles\n"
"             has sample freq 1000, sched:sched_switch event has sample period 1.\n"
"-c count     Set event sample period. It means recording one sample when\n"
"             [count] events happen. For tracepoint events, the default option\n"
"             is -c 1.\n"
"--call-graph fp | dwarf[,<dump_stack_size>]\n"
"             Enable call graph recording. Use frame pointer or dwarf debug\n"
"             frame as the method to parse call graph in stack.\n"
"             Default is dwarf,65528.\n"
"-g           Same as '--call-graph dwarf'.\n"
"--clockid clock_id      Generate timestamps of samples using selected clock.\n"
"                        Possible values are: realtime, monotonic,\n"
"                        monotonic_raw, boottime, perf. Default is perf.\n"
"--cpu cpu_item1,cpu_item2,...\n"
"             Collect samples only on the selected cpus. cpu_item can be cpu\n"
"             number like 1, or cpu range like 0-3.\n"
"--duration time_in_sec  Monitor for time_in_sec seconds instead of running\n"
"                        [command]. Here time_in_sec may be any positive\n"
"                        floating point number.\n"
"-j branch_filter1,branch_filter2,...\n"
"             Enable taken branch stack sampling. Each sample captures a series\n"
"             of consecutive taken branches.\n"
"             The following filters are defined:\n"
"                any: any type of branch\n"
"                any_call: any function call or system call\n"
"                any_ret: any function return or system call return\n"
"                ind_call: any indirect branch\n"
"                u: only when the branch target is at the user level\n"
"                k: only when the branch target is in the kernel\n"
"             This option requires at least one branch type among any, any_call,\n"
"             any_ret, ind_call.\n"
"-b           Enable taken branch stack sampling. Same as '-j any'.\n"
"-m mmap_pages   Set the size of the buffer used to receiving sample data from\n"
"                the kernel. It should be a power of 2. If not set, the max\n"
"                possible value <= 1024 will be used.\n"
"--no-inherit  Don't record created child threads/processes.\n"
"\n"
"Dwarf unwinding options:\n"
"--post-unwind=(yes|no) If `--call-graph dwarf` option is used, then the user's\n"
"                       stack will be recorded in perf.data and unwound while\n"
"                       recording by default. Use --post-unwind=yes to switch\n"
"                       to unwind after recording.\n"
"--no-unwind   If `--call-graph dwarf` option is used, then the user's stack\n"
"              will be unwound by default. Use this option to disable the\n"
"              unwinding of the user's stack.\n"
"--no-callchain-joiner  If `--call-graph dwarf` option is used, then by default\n"
"                       callchain joiner is used to break the 64k stack limit\n"
"                       and build more complete call graphs. However, the built\n"
"                       call graphs may not be correct in all cases.\n"
"--callchain-joiner-min-matching-nodes count\n"
"               When callchain joiner is used, set the matched nodes needed to join\n"
"               callchains. The count should be >= 1. By default it is 1.\n"
"\n"
"Recording file options:\n"
"--no-dump-kernel-symbols  Don't dump kernel symbols in perf.data. By default\n"
"                          kernel symbols will be dumped when needed.\n"
"--no-dump-symbols       Don't dump symbols in perf.data. By default symbols are\n"
"                        dumped in perf.data, to support reporting in another\n"
"                        environment.\n"
"-o record_file_name    Set record file name, default is perf.data.\n"
"--exit-with-parent            Stop recording when the process starting\n"
"                              simpleperf dies.\n"
"--size-limit SIZE[K|M|G]      Stop recording after SIZE bytes of records.\n"
"                              Default is unlimited.\n"
"--start_profiling_fd fd_no    After starting profiling, write \"STARTED\" to\n"
"                              <fd_no>, then close <fd_no>.\n"
"--symfs <dir>    Look for files with symbols relative to this directory.\n"
"                 This option is used to provide files with symbol table and\n"
"                 debug information, which are used for unwinding and dumping symbols.\n"

其中關(guān)鍵參數(shù):

• --call-graph, 是否開(kāi)啟調(diào)用棧的錄制, 兩種堆椆嘶迹回溯的模式: fp(基于寄存器的stack frame pointer), 默認(rèn)模式為dwarf(基于so的.debug_frame)
• -e, 錄制的事件列表, 可以通過(guò) simpleperf list 來(lái)查看當(dāng)前系統(tǒng)支持的event有哪些. u為只監(jiān)控用戶態(tài), k為只監(jiān)控內(nèi)核態(tài)

關(guān)于事件:

• 硬件事件: 由CPU PMU模塊的中斷事件
• 軟件事件: 由內(nèi)核的中斷事件, 例如tick事件
  • task-clock:u 監(jiān)控函數(shù)的CPU占用情況

關(guān)于堆椏畎荆回溯:

• 使用libunwind庫(kù)

關(guān)于dwarf模式:
dwarf是so里面的 .debug 段, 其中 .debug_frame 可用來(lái)做棽澹回溯, 因?yàn)橛行〤PU不會(huì)保存 Stack Frame Pointer (fp) 到寄存器.

關(guān)于fp模式:
但是在Android的某些機(jī)器上(目前看是Android 6.x, 可能小于7.0的機(jī)器上) 會(huì)提示不支持 dwarf 模式, 使用fp模式, 又會(huì)提示 fp mode will not work will

report

將pref.data數(shù)據(jù)轉(zhuǎn)換為txt數(shù)據(jù)

simpleperf report

參數(shù)列表:

• -g, 帶函數(shù)調(diào)用關(guān)系, callee是被調(diào)用關(guān)系, caller是調(diào)用關(guān)系
• -o, output文件名
• --symfs,

            "report", "report sampling information in perf.data",
            "Usage: simpleperf report [options]\n"
            "    -b            Use the branch-to addresses in sampled take branches instead of\n"
            "                  the instruction addresses. Only valid for perf.data recorded with\n"
            "                  -b/-j option.\n"
            "    --children    Print the overhead accumulated by appearing in the callchain.\n"
            "    --comms comm1,comm2,...\n"
            "                  Report only for selected comms.\n"
            "    --dsos dso1,dso2,...\n"
            "                  Report only for selected dsos.\n"
            "    -g [callee|caller]\n"
            "                  Print call graph. If callee mode is used, the graph shows how\n"
            "                  functions are called from others. Otherwise, the graph shows how\n"
            "                  functions call others. Default is callee mode.\n"
            "    -i <file>     Specify path of record file, default is perf.data.\n"
            "    -n            Print the sample count for each item.\n"
            "    --no-demangle        Don't demangle symbol names.\n"
            "    -o report_file_name  Set report file name, default is stdout.\n"
            "    --pid pid1,pid2,...\n"
            "                  Report only for selected pids.\n"
            "    --sort key1,key2,...\n"
            "                  Select the keys to sort and print the report. Possible keys\n"
            "                  include pid, tid, comm, dso, symbol, dso_from, dso_to, symbol_from\n"
            "                  symbol_to. dso_from, dso_to, symbol_from, symbol_to can only be\n"
            "                  used with -b option. Default keys are \"comm,pid,tid,dso,symbol\"\n"
            "    --symfs <dir> Look for files with symbols relative to this directory.\n"
            "    --tids tid1,tid2,...\n"
            "                  Report only for selected tids.\n"
            "    --vmlinux <file>\n"
            "                  Parse kernel symbols from <file>.\n"),

Event FD

打開(kāi) perf event

linux手冊(cè): http://www.man7.org/linux/man-pages/man2/perf_event_open.2.html

static int perf_event_open(const perf_event_attr& attr, pid_t pid, int cpu,
                           int group_fd, unsigned long flags) {  // NOLINT
 **  return syscall(__NR_perf_event_open, &attr, pid, cpu, group_fd, flags);**
}






std::unique_ptr<EventFd> EventFd::OpenEventFile(const perf_event_attr& attr,
                                                pid_t tid, int cpu,
                                                EventFd* group_event_fd,
                                                bool report_error) {
  std::string event_name = GetEventNameByAttr(attr);
  int group_fd = -1;
  if (group_event_fd != nullptr) {
    group_fd = group_event_fd->perf_event_fd_;
  }
  perf_event_attr real_attr = attr;
  if (attr.freq) {
    uint64_t max_sample_freq;
    if (GetMaxSampleFrequency(&max_sample_freq) && max_sample_freq < attr.sample_freq) {
      static bool warned = false;
      if (!warned) {
        warned = true;
        LOG(INFO) << "Adjust sample freq to max allowed sample freq " << max_sample_freq;
      }
      real_attr.sample_freq = max_sample_freq;
    }
  }
  int perf_event_fd = perf_event_open(real_attr, tid, cpu, group_fd, 0);
  if (perf_event_fd == -1) {
    if (report_error) {
      PLOG(ERROR) << "open perf_event_file (event " << event_name << ", tid "
                  << tid << ", cpu " << cpu << ", group_fd " << group_fd
                  << ") failed";
    } else {
      PLOG(DEBUG) << "open perf_event_file (event " << event_name << ", tid "
                  << tid << ", cpu " << cpu << ", group_fd " << group_fd
                  << ") failed";
    }
    return nullptr;
  }
  if (fcntl(perf_event_fd, F_SETFD, FD_CLOEXEC) == -1) {
    if (report_error) {
      PLOG(ERROR) << "fcntl(FD_CLOEXEC) for perf_event_file (event "
                  << event_name << ", tid " << tid << ", cpu " << cpu
                  << ", group_fd " << group_fd << ") failed";
    } else {
      PLOG(DEBUG) << "fcntl(FD_CLOEXEC) for perf_event_file (event "
                  << event_name << ", tid " << tid << ", cpu " << cpu
                  << ", group_fd " << group_fd << ") failed";
    }
    return nullptr;
  }
  return std::unique_ptr<EventFd>(
      new EventFd(real_attr, perf_event_fd, event_name, tid, cpu));
}

參考資料

Android內(nèi)核版本

Android Version    |API Level  |Linux Kernel in AOSP
\----------------------------------------------------
1.5   Cupcake      |3          |2.6.27
1.6   Donut        |4          |2.6.29
2.0/1 Eclair       |5-7        |2.6.29
2.2.x Froyo        |8          |2.6.32
2.3.x Gingerbread  |9, 10      |2.6.35
3.x.x Honeycomb    |11-13      |2.6.36
4.0.x Ice Cream San|14, 15     |3.0.1
4.1.x Jelly Bean   |16         |3.0.31
4.2.x Jelly Bean   |17         |3.4.0
4.3   Jelly Bean   |18         |3.4.39
4.4   Kit Kat      |19, 20     |3.10
5.x   Lollipop     |21, 22     |3.16.1
6.0   Marshmallow  |23         |3.18.10
7.0   Nougat       |24         |4.4.1
7.1   Nougat       |25         |4.4.1 (To be updated)

Why we suggest profiling on Android >= N devices?

1. Running on a device reflects a real running situation, so we suggest
profiling on real devices instead of emulators.
2. To profile Java code, we need ART running in oat mode, which is only
available >= L for rooted devices, and >= N for non-rooted devices.
3. Old Android versions are likely to be shipped with old kernels (< 3.18),
which may not support profiling features like recording dwarf based call graphs.
4. Old Android versions are likely to be shipped with Arm32 chips. In Arm32
mode, recording stack frame based call graphs doesn't work well.

NOTE ATTRIBUTES

Created Date: 2018-08-09 08:45:52
Last Evernote Update Date: 2018-08-14 08:04:49