Brad Bishop | 6e60e8b | 2018-02-01 10:27:11 -0500 | [diff] [blame] | 1 | # Implements system state sampling. Called by buildstats.bbclass. |
| 2 | # Because it is a real Python module, it can hold persistent state, |
| 3 | # like open log files and the time of the last sampling. |
| 4 | |
| 5 | import time |
| 6 | import re |
| 7 | import bb.event |
| 8 | |
| 9 | class SystemStats: |
| 10 | def __init__(self, d): |
| 11 | bn = d.getVar('BUILDNAME') |
| 12 | bsdir = os.path.join(d.getVar('BUILDSTATS_BASE'), bn) |
| 13 | bb.utils.mkdirhier(bsdir) |
| 14 | |
| 15 | self.proc_files = [] |
| 16 | for filename, handler in ( |
| 17 | ('diskstats', self._reduce_diskstats), |
| 18 | ('meminfo', self._reduce_meminfo), |
| 19 | ('stat', self._reduce_stat), |
| 20 | ): |
| 21 | # The corresponding /proc files might not exist on the host. |
| 22 | # For example, /proc/diskstats is not available in virtualized |
| 23 | # environments like Linux-VServer. Silently skip collecting |
| 24 | # the data. |
| 25 | if os.path.exists(os.path.join('/proc', filename)): |
| 26 | # In practice, this class gets instantiated only once in |
| 27 | # the bitbake cooker process. Therefore 'append' mode is |
| 28 | # not strictly necessary, but using it makes the class |
| 29 | # more robust should two processes ever write |
| 30 | # concurrently. |
| 31 | destfile = os.path.join(bsdir, '%sproc_%s.log' % ('reduced_' if handler else '', filename)) |
| 32 | self.proc_files.append((filename, open(destfile, 'ab'), handler)) |
| 33 | self.monitor_disk = open(os.path.join(bsdir, 'monitor_disk.log'), 'ab') |
| 34 | # Last time that we sampled /proc data resp. recorded disk monitoring data. |
| 35 | self.last_proc = 0 |
| 36 | self.last_disk_monitor = 0 |
| 37 | # Minimum number of seconds between recording a sample. This |
| 38 | # becames relevant when we get called very often while many |
| 39 | # short tasks get started. Sampling during quiet periods |
| 40 | # depends on the heartbeat event, which fires less often. |
| 41 | self.min_seconds = 1 |
| 42 | |
| 43 | self.meminfo_regex = re.compile(b'^(MemTotal|MemFree|Buffers|Cached|SwapTotal|SwapFree):\s*(\d+)') |
| 44 | self.diskstats_regex = re.compile(b'^([hsv]d.|mtdblock\d|mmcblk\d|cciss/c\d+d\d+.*)$') |
| 45 | self.diskstats_ltime = None |
| 46 | self.diskstats_data = None |
| 47 | self.stat_ltimes = None |
| 48 | |
| 49 | def close(self): |
| 50 | self.monitor_disk.close() |
| 51 | for _, output, _ in self.proc_files: |
| 52 | output.close() |
| 53 | |
| 54 | def _reduce_meminfo(self, time, data): |
| 55 | """ |
| 56 | Extracts 'MemTotal', 'MemFree', 'Buffers', 'Cached', 'SwapTotal', 'SwapFree' |
| 57 | and writes their values into a single line, in that order. |
| 58 | """ |
| 59 | values = {} |
| 60 | for line in data.split(b'\n'): |
| 61 | m = self.meminfo_regex.match(line) |
| 62 | if m: |
| 63 | values[m.group(1)] = m.group(2) |
| 64 | if len(values) == 6: |
| 65 | return (time, |
| 66 | b' '.join([values[x] for x in |
| 67 | (b'MemTotal', b'MemFree', b'Buffers', b'Cached', b'SwapTotal', b'SwapFree')]) + b'\n') |
| 68 | |
| 69 | def _diskstats_is_relevant_line(self, linetokens): |
| 70 | if len(linetokens) != 14: |
| 71 | return False |
| 72 | disk = linetokens[2] |
| 73 | return self.diskstats_regex.match(disk) |
| 74 | |
| 75 | def _reduce_diskstats(self, time, data): |
| 76 | relevant_tokens = filter(self._diskstats_is_relevant_line, map(lambda x: x.split(), data.split(b'\n'))) |
| 77 | diskdata = [0] * 3 |
| 78 | reduced = None |
| 79 | for tokens in relevant_tokens: |
| 80 | # rsect |
| 81 | diskdata[0] += int(tokens[5]) |
| 82 | # wsect |
| 83 | diskdata[1] += int(tokens[9]) |
| 84 | # use |
| 85 | diskdata[2] += int(tokens[12]) |
| 86 | if self.diskstats_ltime: |
| 87 | # We need to compute information about the time interval |
| 88 | # since the last sampling and record the result as sample |
| 89 | # for that point in the past. |
| 90 | interval = time - self.diskstats_ltime |
| 91 | if interval > 0: |
| 92 | sums = [ a - b for a, b in zip(diskdata, self.diskstats_data) ] |
| 93 | readTput = sums[0] / 2.0 * 100.0 / interval |
| 94 | writeTput = sums[1] / 2.0 * 100.0 / interval |
| 95 | util = float( sums[2] ) / 10 / interval |
| 96 | util = max(0.0, min(1.0, util)) |
| 97 | reduced = (self.diskstats_ltime, (readTput, writeTput, util)) |
| 98 | |
| 99 | self.diskstats_ltime = time |
| 100 | self.diskstats_data = diskdata |
| 101 | return reduced |
| 102 | |
| 103 | |
| 104 | def _reduce_nop(self, time, data): |
| 105 | return (time, data) |
| 106 | |
| 107 | def _reduce_stat(self, time, data): |
| 108 | if not data: |
| 109 | return None |
| 110 | # CPU times {user, nice, system, idle, io_wait, irq, softirq} from first line |
| 111 | tokens = data.split(b'\n', 1)[0].split() |
| 112 | times = [ int(token) for token in tokens[1:] ] |
| 113 | reduced = None |
| 114 | if self.stat_ltimes: |
| 115 | user = float((times[0] + times[1]) - (self.stat_ltimes[0] + self.stat_ltimes[1])) |
| 116 | system = float((times[2] + times[5] + times[6]) - (self.stat_ltimes[2] + self.stat_ltimes[5] + self.stat_ltimes[6])) |
| 117 | idle = float(times[3] - self.stat_ltimes[3]) |
| 118 | iowait = float(times[4] - self.stat_ltimes[4]) |
| 119 | |
| 120 | aSum = max(user + system + idle + iowait, 1) |
| 121 | reduced = (time, (user/aSum, system/aSum, iowait/aSum)) |
| 122 | |
| 123 | self.stat_ltimes = times |
| 124 | return reduced |
| 125 | |
| 126 | def sample(self, event, force): |
| 127 | now = time.time() |
| 128 | if (now - self.last_proc > self.min_seconds) or force: |
| 129 | for filename, output, handler in self.proc_files: |
| 130 | with open(os.path.join('/proc', filename), 'rb') as input: |
| 131 | data = input.read() |
| 132 | if handler: |
| 133 | reduced = handler(now, data) |
| 134 | else: |
| 135 | reduced = (now, data) |
| 136 | if reduced: |
| 137 | if isinstance(reduced[1], bytes): |
| 138 | # Use as it is. |
| 139 | data = reduced[1] |
| 140 | else: |
| 141 | # Convert to a single line. |
| 142 | data = (' '.join([str(x) for x in reduced[1]]) + '\n').encode('ascii') |
| 143 | # Unbuffered raw write, less overhead and useful |
| 144 | # in case that we end up with concurrent writes. |
| 145 | os.write(output.fileno(), |
| 146 | ('%.0f\n' % reduced[0]).encode('ascii') + |
| 147 | data + |
| 148 | b'\n') |
| 149 | self.last_proc = now |
| 150 | |
| 151 | if isinstance(event, bb.event.MonitorDiskEvent) and \ |
| 152 | ((now - self.last_disk_monitor > self.min_seconds) or force): |
| 153 | os.write(self.monitor_disk.fileno(), |
| 154 | ('%.0f\n' % now).encode('ascii') + |
| 155 | ''.join(['%s: %d\n' % (dev, sample.total_bytes - sample.free_bytes) |
| 156 | for dev, sample in event.disk_usage.items()]).encode('ascii') + |
| 157 | b'\n') |
| 158 | self.last_disk_monitor = now |