#
# Mark Huang <mlhuang@cs.princeton.edu>
# Andy Bavier <acb@cs.princeton.edu>
+# Faiyaz Ahmed <faiyaza@cs.princeton.edu>
# Copyright (C) 2004-2006 The Trustees of Princeton University
#
-# $Id: swapmon.py,v 1.1 2006/04/28 19:26:59 mlhuang Exp $
-#
import syslog
import os
import socket
import time
-# util-vserver/python/vserver.py allows us to control slices directly
-# from Python
-from vserver import VServer
-
# bwlimit exports a few useful functions like run(), get_xid(), and get_slice()
-import bwlimit
+import plnode.bwlimit as bwlimit
# Utility functions
from pl_mom import *
# Defaults
debug = False
verbose = 0
-datafile = "/var/lib/misc/swapmon.dat"
-
+DATAFILE = "/var/lib/misc/swapmon.dat"
+# xxx fixme - this is broken under git
+VERSION = "$Id$"
# Seconds between process analysis
period = 30
change_thresh = 5
# Swap utilization at which the largest consumer of physical memory is reset
-reset_thresh = 85
+reset_thresh = 80
# Swap utilization at which the machine is rebooted
reboot_thresh = 95
-# Minimum physical memory utilization to be considered the largest consumer
-min_thresh = 10
+# Don't email the same message more than once in the same emailtimeout interval
+email_timeout = 1800
+
+# Physical size threshold to be considered a consumer. Rationale is if there are no procs
+# with a size at least as large as this, then there is a slow leaker; better to just reboot.
+rss_min = 150 * 1024
# System slices that should not be reset (regexps)
system_slices = ['root', PLC_SLICE_PREFIX + '_']
%(date)s %(hostname)s reboot
""".lstrip()
-# Message sent after a hog is reset
-reset_subject = "pl_mom reset slice %(slice)s on %(hostname)s"
-reset_body = \
-"""
+# Message sent to system slices that should not be reset
+alarm_subject = "pl_mom alarm slice %(slice)s on %(hostname)s"
+alarm_body = \
+"""
Sometime before %(date)s, swap space was
nearly exhausted on %(hostname)s.
-Slice %(slice)s was reset since it was the largest consumer of
-physical memory at %(rss)s (%(percent)4.1f%%).
-
-Please reply to this message explaining the nature of your experiment,
-and what you are doing to address the problem.
+System slice %(slice)s was the largest consumer of physical memory at
+%(rss)s (%(percent)4.1f%%) (%(sz)s writable). It was not reset,
+but please verify its behavior.
-%(slice)s processes prior to reset:
+%(slice)s processes prior to alarm:
%(table)s
-%(date)s %(hostname)s reset %(slice)s
+%(date)s %(hostname)s alarm %(slice)s
""".lstrip()
-# Message sent to system slices that should not be reset
-alarm_subject = "pl_mom alarm slice %(slice)s on %(hostname)s"
-alarm_body = \
-"""
+# Message sent after a slice has been killed
+kill_subject = "pl_mom killed slice %(slice)s on %(hostname)s"
+kill_body = \
+"""
Sometime before %(date)s, swap space was
nearly exhausted on %(hostname)s.
-System slice %(slice)s was the largest consumer of physical memory at
-%(rss)s (%(percent)4.1f%%). It was not reset, but please verify its
-behavior.
+Slice %(slice)s was killed since it was the largest consumer of
+physical memory at %(rss)s (%(percent)4.1f%%) (%(sz)s writable)
+after repeated restarts.
-%(slice)s processes prior to alarm:
+Please reply to this message explaining the nature of your experiment,
+and what you are doing to address the problem.
+
+%(slice)s processes prior to reset:
%(table)s
-%(date)s %(hostname)s alarm %(slice)s
+%(date)s %(hostname)s reset %(slice)s
""".lstrip()
+def killsliverprocs(xid):
+ bwlimit.run("/usr/sbin/vkill -s 9 -c %s 0" % xid)
+
+
def usage():
print """
Usage: %s [OPTIONS]...
--reboot-thresh=PERCENT Swap utilization at which the machine is rebooted
--min-thresh=PERCENT Minimum physical memory utilization to be considered a hog
--system-slice=SLICE System slice that should not be reset
+ --status Print memory usage statistics and exit
+ --memstatus Print total memory, total swap, and swap used
-h, --help This message
-""".lstrip() % (sys.argv[0], debug, verbose, datafile, format_period(period))
+""".lstrip() % (sys.argv[0], debug, verbose, DATAFILE, format_period(period))
def slicestat(names = None):
"""
Get status of specified slices (if names is None or empty, all
- slices). vsize and rss are in KiB. Returns
-
+ slices). vsize, sz, and rss are in KiB. Returns
+ PID CONTEXT VSZ SZ RSS %MEM CMD
{xid: {'xid': slice_id,
- 'name': slice_name,
- 'procs': [{'pid': pid, 'xid': slice_id, 'user', username, 'cmd': command,
- 'vsize': virtual_kib, 'rss': physical_kib,
+ 'name': slice_name,
+ 'procs': [{'pid': pid, 'xid': slice_id, 'cmd': command,
+ 'vsize': virtual_kib, 'sz': potential_kib, 'rss': physical_kib,
'pcpu': cpu_percent, 'pmem': mem_percent}]
- 'vsize': total_virtual_kib,
- 'rss': total_physical_kib}}
+ 'vsize': total_virtual_kib,
+ 'sz': total_potential_kib,
+ 'rss': total_physical_kib}}
"""
# Mandatory fields. xid is a virtual field inserted by vps. Make
# sure cmd is last so that it does not get truncated
# automatically.
- fields = ['pid', 'xid', 'user', 'vsize', 'rss', 'pcpu', 'pmem', 'cmd']
+ fields = ['pid', 'xid', 'vsname', 'vsize', 'sz', 'rss', 'pmem', 'cmd']
# vps inserts xid after pid in the output, but ps doesn't know
# what the field means.
ps_fields = list(fields)
ps_fields.remove('xid')
+ ps_fields.remove('vsname')
slices = {}
# Eat the header line. vps depends on the header to figure out
# which column is the PID column, so we can't just tell ps not to
# print it.
- for line in bwlimit.run("/usr/sbin/vps -e -o " + ",".join(ps_fields))[1:]:
+ for line in bwlimit.run("/usr/sbin/vps -e -o " + ":16,".join(ps_fields))[1:]:
# Chomp newline
line = line.strip()
- # Replace "0 MAIN" and "1 ALL_PROC" (the special monikers that
- # vps uses to denote the root context and the "all contexts"
- # context) with "0" so that we can just split() on whitespace.
- line = line.replace("0 MAIN", "0").replace("1 ALL_PROC", "0")
-
# Represent process as a dict of fields
values = line.split(None, len(fields) - 1)
if len(values) != len(fields):
+ if "ERR" in line:
+ pass # ignore spurious error message from vps
+ else:
+ print "slicestat: failed to parse line: " + line
continue
proc = dict(zip(fields, values))
except ValueError:
pass
+ # vps sometimes prints ERR or the name of the slice
+ # instead of a context ID if it
+ # cannot identify the context of an orphaned (usually dying)
+ # process. Skip these processes.
+ if (type(proc['xid']) != int) or (type(proc['vsize']) !=int):
+ if "ERR" in line:
+ pass # ignore spurious error message from vps
+ else:
+ print "slicestat: failed to parse line: " + line
+ continue
+
# Assign (pl_)sshd processes to slice instead of root
m = re.search(r"sshd: ([a-zA-Z_]+)", proc['cmd'])
+
if m is not None:
xid = bwlimit.get_xid(m.group(1))
if xid is not None:
proc['rss'] += 12
# Include additional page table overhead
- if proc['vsize'] > 4096:
- proc['rss'] += 4 * ((proc['vsize'] - 1) / 4096)
+ try:
+ if proc['vsize'] > 4096:
+ proc['rss'] += 4 * ((proc['vsize'] - 1) / 4096)
+ except: pass
if slices.has_key(proc['xid']):
slice = slices[proc['xid']]
else:
- slice = {'xid': proc['xid'], 'name': name, 'procs': [], 'vsize': 0, 'rss': 0}
+ slice = {'xid': proc['xid'], 'name': name, 'procs': [], 'vsize': 0, 'sz': 0, 'rss': 0}
slice['procs'].append(proc)
slice['vsize'] += proc['vsize']
+ slice['sz'] += proc['sz']
slice['rss'] += proc['rss']
slices[proc['xid']] = slice
def memtotal():
"""
- Returns total physical memory on the system in KiB.
+ Returns total physical and swap memory on the system in KiB.
"""
-
+ mem = 0
+ swap = 0
meminfo = open("/proc/meminfo", "r")
- line = meminfo.readline()
+ for line in meminfo.readlines():
+ try:
+ (name, value, kb) = line.split()
+ except:
+ continue
+ if name == "MemTotal:":
+ mem = int(value)
+ elif name == "SwapTotal:":
+ swap = int(value)
meminfo.close()
- if line[0:8] == "MemTotal":
- # MemTotal: 255396 kB
- (name, value, kb) = line.split()
- return int(value)
-
- return 0
+ return (mem, swap)
def swap_used():
"""
Returns swap utilization on the system as a whole percentage (0-100).
"""
-
total_swap = 0
total_used = 0
-
try:
swaps = open("/proc/swaps", "r")
# Eat header line
total_used += int(used)
except ValueEror, err:
pass
- except (IOError, KeyError), err:
- pass
+ except (IOError, KeyError), err: pass
+
+ swapused = 100 * total_used / total_swap
+ if debug: print "%s percent swap used" % swapused
+ return swapused
+
+def summary(slices = None, total_mem = None, total_swap = None):
+ """
+ Return a summary of memory usage by slice.
+ """
+ if not slices: slices = slicestat()
+ slicelist = slices.values()
+ slicelist.sort(lambda a, b: b['sz'] - a['sz'])
+ if total_mem is None or total_swap is None:
+ (total_mem, total_swap) = memtotal()
+
+ table = "%-20s%10s%24s%24s\n\n" % ("Slice", "Processes", "Memory Usage", "Potential Usage")
+ for slice in slicelist:
+ table += "%-20s%10d%16s (%4.1f%%)%16s (%4.1f%%)\n" % \
+ (slice['name'], len(slice['procs']),
+ format_bytes(slice['rss'] * 1024, si = False),
+ 100. * slice['rss'] / total_mem,
+ format_bytes(slice['sz'] * 1024, si = False),
+ 100. * slice['sz'] / (total_mem + total_swap))
+ return table
+
+def formtable(slice, percent):
+ '''
+ Makes pretty message to email with human readable ps values.
+ '''
+ table = "%5s %10s %10s %10s %4s %4s %s\n\n" % \
+ ("PID", "VIRT", "SZ", "RES", '%CPU', '%MEM', 'COMMAND')
+ for proc in slice['procs']:
+ table += "%5s %10s %10s %10s %4.1f %s\n" % \
+ (proc['pid'],
+ format_bytes(proc['vsize'] * 1024, si = False),
+ format_bytes(proc['sz'] * 1024, si = False),
+ format_bytes(proc['rss'] * 1024, si = False),
+ proc['pmem'],
+ proc['cmd'])
+
+ prettytable = {'hostname': socket.gethostname(),
+ 'date': time.asctime(time.gmtime()) + " GMT",
+ 'table': table,
+ 'slice': slice['name'],
+ 'rss': format_bytes(slice['rss'] * 1024, si = False),
+ 'sz': format_bytes(slice['sz'] * 1024, si = False),
+ 'percent': percent}
+ return prettytable
+
+def readdat():
+ '''
+ Return dictionary of vps (slicestat) from datfile left behind by OOM
+ before rebooting. If none file, just grab the latest dict (slicestat)
+ and return that. If dat file found, means we rebooted, send an email to
+ pl_mom@pl.
+ '''
+ try:
+ f = open(DATAFILE, "r+")
+ if verbose:
+ print "Loading %s" % DATAFILE
+ (v, slices) = pickle.load(f)
+ f.close()
+ # Check version of data file
+ if v != VERSION:
+ print "Not using old version '%s' data file %s" % (v, DATAFILE)
+ raise Exception
+
+ params = {'hostname': socket.gethostname(),
+ 'date': time.asctime(time.gmtime()) + " GMT",
+ 'table': summary(slices, total_mem, total_swap)}
+ if debug:
+ print rebooted_subject % params
+ print rebooted_body % params
+ else:
+ slicemail(None, rebooted_subject % params, rebooted_body % params)
+
+ # Delete data file
+ os.unlink(DATAFILE)
+ except Exception:
+ slices = slicestat()
+
+ return slices
+
+
+def writedat(slices):
+ """
+ Write (slices) to pickled datfile.
+ """
+ if verbose: print "Saving %s" % DATAFILE
+ f = open(DATAFILE, "w")
+ pickle.dump((VERSION, slices), f)
+ f.close()
- return 100 * total_used / total_swap
def main():
# Defaults
- global debug, verbose, datafile
- global period, change_thresh, reset_thresh, reboot_thresh, min_thresh, system_slices
+ global debug, verbose, DATAFILE, VERSION
+ global period, change_thresh, reset_thresh, reboot_thresh, rss_min, system_slices
# All slices
names = []
+ timer = period
+ last_used = None
+ used = None
+ warned = []
+ emailed = {}
try:
- longopts = ["debug", "verbose", "file=", "slice=", "help"]
+ longopts = ["debug", "verbose", "file=", "slice=", "status", "memstatus", "help"]
longopts += ["period=", "reset-thresh=", "reboot-thresh=", "min-thresh=", "system-slice="]
(opts, argv) = getopt.getopt(sys.argv[1:], "dvf:s:ph", longopts)
except getopt.GetoptError, err:
elif opt == "-v" or opt == "--verbose":
verbose += 1
elif opt == "-f" or opt == "--file":
- datafile = optval
+ DATAFILE = optval
elif opt == "-s" or opt == "--slice":
names.append(optval)
elif opt == "-p" or opt == "--period":
elif opt == "--reboot-thresh":
reboot_thresh = int(optval)
elif opt == "--min-thresh":
- min_thresh = int(optval)
+ rss_min = int(optval)
elif opt == "--system-slice":
system_slices.append(optval)
+ elif opt == "--status":
+ print summary(slicestat(names))
+ sys.exit(0)
+ elif opt == "--memstatus":
+ (mem, swap) = memtotal()
+ swap_pct = swap_used()
+ print "memory total:", mem
+ print "swap total:", swap
+ print "swap used:", swap_pct
+ sys.exit(0)
else:
usage()
sys.exit(0)
syslog.openlog("swapmon")
sys.stdout = sys.stderr = Logger()
- # Get total physical memory
- total_rss = memtotal()
-
- try:
- f = open(datafile, "r+")
- if verbose:
- print "Loading %s" % datafile
- (version, slices) = pickle.load(f)
- f.close()
- # Check version of data file
- if version != "$Id: swapmon.py,v 1.1 2006/04/28 19:26:59 mlhuang Exp $":
- print "Not using old version '%s' data file %s" % (version, datafile)
- raise Exception
-
- # Send notification if we rebooted the node because of swap exhaustion
- slicelist = slices.values()
- slicelist.sort(lambda a, b: b['rss'] - a['rss'])
-
- table = "%-20s%10s%24s\n\n" % ("Slice", "Processes", "Memory Usage")
- for slice in slicelist:
- table += "%-20s%10d%16s (%4.1f%%)\n" % \
- (slice['name'], len(slice['procs']),
- format_bytes(slice['rss'] * 1024, si = False),
- 100. * slice['rss'] / total_rss)
-
- params = {'hostname': socket.gethostname(),
- 'date': time.asctime(time.gmtime()) + " GMT",
- 'table': table}
-
- if debug:
- print rebooted_subject % params
- print rebooted_body % params
- else:
- slicemail(None, rebooted_subject % params, rebooted_body % params)
-
- # Delete data file
- os.unlink(datafile)
- except Exception:
- version = "$Id: swapmon.py,v 1.1 2006/04/28 19:26:59 mlhuang Exp $"
- slices = {}
+ # Get total memory
+ (total_mem, total_swap) = memtotal()
+ slices = readdat()
# Query process table every 30 seconds, or when a large change in
# swap utilization is detected.
- timer = period
- last_used = None
- used = None
-
- # System slices that we have warned but could not reset
- warned = []
while True:
used = swap_used()
- if last_used is None:
- last_used = used
- if verbose:
- print "%d%% swap consumed" % used
-
+ if last_used is None: last_used = used
+
if used >= reboot_thresh:
# Dump slice state before rebooting
- if verbose:
- print "Saving %s" % datafile
- f = open(datafile, "w")
- pickle.dump((version, slices), f)
- f.close()
-
+ writedat(slices)
# Goodbye, cruel world
print "%d%% swap consumed, rebooting" % used
- if not debug:
- bwlimit.run("/bin/sync; /sbin/reboot -f")
-
+ if not debug: bwlimit.run("/bin/sync; /sbin/reboot -f")
elif used >= reset_thresh:
# Try and find a hog
slicelist = slices.values()
+ # Puts largest on top.
slicelist.sort(lambda a, b: b['rss'] - a['rss'])
for slice in slicelist:
- percent = 100. * slice['rss'] / total_rss
- if percent < min_thresh:
- continue
-
+ percent = 100. * slice['rss'] / total_mem
+ if slice['rss'] < rss_min: continue
print "%d%% swap consumed, slice %s is using %s (%d%%) of memory" % \
- (used,
- slice['name'],
- format_bytes(slice['rss'] * 1024, si = False),
- percent)
-
+ (used,
+ slice['name'],
+ format_bytes(slice['rss'] * 1024, si = False),
+ percent)
slice['procs'].sort(lambda a, b: b['rss'] - a['rss'])
-
- table = "%5s %10s %10s %4s %4s %s\n\n" % ("PID", "VIRT", "RES", '%CPU', '%MEM', 'COMMAND')
- for proc in slice['procs']:
- table += "%5s %10s %10s %4.1f %4.1f %s\n" % \
- (proc['pid'],
- format_bytes(proc['vsize'] * 1024, si = False),
- format_bytes(proc['rss'] * 1024, si = False),
- proc['pcpu'], proc['pmem'], proc['cmd'])
-
- params = {'hostname': socket.gethostname(),
- 'date': time.asctime(time.gmtime()) + " GMT",
- 'table': table,
- 'slice': slice['name'],
- 'rss': format_bytes(slice['rss'] * 1024, si = False),
- 'percent': percent}
-
+ # Make a pretty table.
+ params = formtable(slice, percent)
# Match slice name against system slice patterns
- is_system_slice = filter(None, [re.match(pattern, slice['name']) for pattern in system_slices])
-
- if is_system_slice:
- # Do not reset system slices, just warn once
+ is_system_slice = filter(None,
+ [re.match(pattern, slice['name']) for pattern in system_slices])
+
+ # Do not reset system slices, just warn once
+ if is_system_slice:
if slice['name'] not in warned:
warned.append(slice['name'])
+ print "Warning slice " + slice['name']
if debug:
print alarm_subject % params
print alarm_body % params
else:
- print "Warning slice " + slice['name']
- slicemail(slice['name'], alarm_subject % params, alarm_body % params)
+ slicemail(slice['name'], alarm_subject % params,
+ alarm_body % params)
else:
- # Otherwise, reset
- if debug:
- print reset_subject % params
- print reset_body % params
+ # Reset slice
+ if not debug:
+ if emailed.get(slice['name'], (time.time() + email_timeout + 1)) > (time.time() + email_timeout):
+ slicemail(slice['name'], kill_subject % params, kill_body % params)
+ emailed[slice['name']] = time.time()
else:
- try:
- pid = os.fork()
- if pid == 0:
- print "Resetting slice " + slice['name']
- vserver = VServer(slice['name'])
- vserver.stop()
- vserver.start(wait = False)
- os._exit(0)
- else:
- os.waitpid(pid, 0)
- except Exception, err:
- print "Warning: Exception received while resetting slice %s:" % slice['name'], err
- slicemail(slice['name'], reset_subject % params, reset_body % params)
- break
-
- elif timer <= 0 or used >= (last_used + change_thresh):
+ print kill_subject % params
+ print kill_body % params
+ print "Killing procs in %s" % slice['name']
+ killsliverprocs(slice['xid'])
+
+ # wait period before recalculating swap. If in danger, recalc.
+ if timer <= 0 or used >= (last_used + change_thresh):
if used >= (last_used + change_thresh):
print "%d%% swap consumed, %d%% in last %d seconds" % \
- (used, used - last_used, period - timer)
+ (used, used - last_used, period - timer)
# Get slice state
slices = slicestat(names)
# Reset timer
timer = period
# Keep track of large changes in swap utilization
last_used = used
-
timer -= 1
time.sleep(1)