Android CPU使用率：top和dump cpuinfo的不同

CPU是系统非常重要的资源，在Android中，查看CPU使用情况，可以使用top命令和dump cpuinfo。我记得很久以前，就发现这两者存在不同，初步猜测应该是算法上存在差异。最近需要采集应用CPU的使用率，看了一下两种CPU的计算方法。

1、top

top是比较经典的CPU计算方法，top的代码在androidm/system/core/toolbox/top.c下面，输出process的cpu使用率在print_procs里面：

static void print_procs(void) {
...for (i = 0; i < num_new_procs; i++) {if (new_procs[i]) {old_proc = find_old_proc(new_procs[i]->pid, new_procs[i]->tid);if (old_proc) {new_procs[i]->delta_utime = new_procs[i]->utime - old_proc->utime;new_procs[i]->delta_stime = new_procs[i]->stime - old_proc->stime;} else {new_procs[i]->delta_utime = 0;new_procs[i]->delta_stime = 0;}new_procs[i]->delta_time = new_procs[i]->delta_utime + new_procs[i]->delta_stime;}}total_delta_time = (new_cpu.utime + new_cpu.ntime + new_cpu.stime + new_cpu.itime+ new_cpu.iowtime + new_cpu.irqtime + new_cpu.sirqtime)- (old_cpu.utime + old_cpu.ntime + old_cpu.stime + old_cpu.itime+ old_cpu.iowtime + old_cpu.irqtime + old_cpu.sirqtime);...if (!threads) {printf("%5d %2d %3" PRIu64 "%% %c %5d %6" PRIu64 "K %6" PRIu64 "K %3s %-8.8s %s\n",proc->pid, proc->prs, proc->delta_time * 100 / total_delta_time, proc->state, proc->num_threads,proc->vss / 1024, proc->rss * getpagesize() / 1024, proc->policy, user_str, proc->name[0] != 0 ? proc->name : proc->tname);} else {printf("%5d %5d %2d %3" PRIu64 "%% %c %6" PRIu64 "K %6" PRIu64 "K %3s %-8.8s %-15s %s\n",proc->pid, proc->tid, proc->prs, proc->delta_time * 100 / total_delta_time, proc->state,proc->vss / 1024, proc->rss * getpagesize() / 1024, proc->policy, user_str, proc->tname, proc->name);}
...
}

CPU的计算是proc->delta_time * 100 / total_delta_time。

先看total_delta_time由：

    total_delta_time = (new_cpu.utime + new_cpu.ntime + new_cpu.stime + new_cpu.itime+ new_cpu.iowtime + new_cpu.irqtime + new_cpu.sirqtime)- (old_cpu.utime + old_cpu.ntime + old_cpu.stime + old_cpu.itime+ old_cpu.iowtime + old_cpu.irqtime + old_cpu.sirqtime);

而这些变量的值，是在read_procs通过读取/proc/stat的jiffies得到：

static void read_procs(void) {
...proc_dir = opendir("/proc");if (!proc_dir) die("Could not open /proc.\n");new_procs = calloc(INIT_PROCS * (threads ? THREAD_MULT : 1), sizeof(struct proc_info *));num_new_procs = INIT_PROCS * (threads ? THREAD_MULT : 1);file = fopen("/proc/stat", "r");if (!file) die("Could not open /proc/stat.\n");fscanf(file, "cpu  %lu %lu %lu %lu %lu %lu %lu", &new_cpu.utime, &new_cpu.ntime, &new_cpu.stime,&new_cpu.itime, &new_cpu.iowtime, &new_cpu.irqtime, &new_cpu.sirqtime);fclose(file);

而proc->delta_time是两次读取/proc/pid/stat相减得到：

static int read_stat(char *filename, struct proc_info *proc) {
.../* Scan rest of string. */sscanf(close_paren + 1," %c " "%*d %*d %*d %*d %*d %*d %*d %*d %*d %*d ""%" SCNu64"%" SCNu64 "%*d %*d %*d %*d %*d %*d %*d ""%" SCNu64"%" SCNu64 "%*d %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d ""%d",&proc->state,&proc->utime,&proc->stime,&proc->vss,&proc->rss,&proc->prs);return 0;
}

可见，top是一段时间内，计算process的cpu jiffies与总的cpu jiffies差值得到。

2、dump cpuinfo

dump cpuinfo是Android特有的命令（我一直都android的各种dump、trace非常感兴趣，快玩物丧志了。。。）。dump cpuinfo命令的实现在androidm/frameworks/base/core/java/com/android/internal/os/ProcessCpuTracker.java类里面，方法是printCurrentState:

final public String printCurrentState(long now) {
...int N = mWorkingProcs.size();for (int i=0; i<N; i++) {Stats st = mWorkingProcs.get(i);printProcessCPU(pw, st.added ? " +" : (st.removed ? " -": "  "),st.pid, st.name, (int)st.rel_uptime,st.rel_utime, st.rel_stime, 0, 0, 0, st.rel_minfaults, st.rel_majfaults);if (!st.removed && st.workingThreads != null) {int M = st.workingThreads.size();for (int j=0; j<M; j++) {Stats tst = st.workingThreads.get(j);printProcessCPU(pw,tst.added ? "   +" : (tst.removed ? "   -": "    "),tst.pid, tst.name, (int)st.rel_uptime,tst.rel_utime, tst.rel_stime, 0, 0, 0, 0, 0);}}}...
}

而printProcessCPU输出process CPU的使用情况：

private void printProcessCPU(PrintWriter pw, String prefix, int pid, String label,int totalTime, int user, int system, int iowait, int irq, int softIrq,int minFaults, int majFaults) {pw.print(prefix);if (totalTime == 0) totalTime = 1;printRatio(pw, user+system+iowait+irq+softIrq, totalTime);
...
}

user+system+iowait+irq+softIrq 比totalTime。
看下st个变量的赋值，在collectStats里面：

private int[] collectStats(String statsFile, int parentPid, boolean first,int[] curPids, ArrayList<Stats> allProcs) {int[] pids = Process.getPids(statsFile, curPids);
...final long uptime = SystemClock.uptimeMillis();final long[] procStats = mProcessStatsData;if (!Process.readProcFile(st.statFile.toString(),PROCESS_STATS_FORMAT, null, procStats, null)) {continue;}...if (DEBUG) Slog.v("Load", "Stats changed " + st.name + " pid=" + st.pid+ " utime=" + utime + "-" + st.base_utime+ " stime=" + stime + "-" + st.base_stime+ " minfaults=" + minfaults + "-" + st.base_minfaults+ " majfaults=" + majfaults + "-" + st.base_majfaults);st.rel_uptime = uptime - st.base_uptime;st.base_uptime = uptime;st.rel_utime = (int)(utime - st.base_utime);st.rel_stime = (int)(stime - st.base_stime);st.base_utime = utime;st.base_stime = stime;st.rel_minfaults = (int)(minfaults - st.base_minfaults);st.rel_majfaults = (int)(majfaults - st.base_majfaults);st.base_minfaults = minfaults;st.base_majfaults = majfaults;st.working = true;
...
}

st.rel_utime 和 st.rel_stime还是通过读/proc/pid/stat相减得到，而st.rel_uptime却是通过 SystemClock.uptimeMillis()差值，并不是跟top一样，通过得到总CPU jiffies。

看到这，也就能明白，top跟dump cpuinfo的区别在于：top分母有的是总测CPU jiffies，而dump cpuinfo是uptime，是时间，而并非jiffies，也能解释为什么top出来的cpu，大部分时间会比dump cpuinfo的原因。