Updates to cpusets
Signed-off-by: Martin J. Bligh <[email protected]>
git-svn-id: http://test.kernel.org/svn/autotest/trunk@475 592f7852-d20e-0410-864c-8624ca9c26a4
diff --git a/client/bin/autotest_utils.py b/client/bin/autotest_utils.py
index 1deee1c..b5dab84 100755
--- a/client/bin/autotest_utils.py
+++ b/client/bin/autotest_utils.py
@@ -1,7 +1,7 @@
"""Convenience functions for use by tests or whomever.
"""
-import os,os.path,shutil,urllib,sys,signal,commands,pickle
+import os,os.path,shutil,urllib,sys,signal,commands,pickle,glob
from error import *
import re,string
@@ -96,7 +96,7 @@
return 0
-def get_file(src, dest):
+def get_file(src, dest, permissions = None):
"""get a file, either from url or local"""
if (src == dest): # no-op here allows clean overrides in tests
return
@@ -108,8 +108,10 @@
except IOError:
sys.stderr.write("Unable to retrieve %s (to %s)\n" % (src, dest))
sys.exit(1)
- return dest
- shutil.copyfile(src, dest)
+ else:
+ shutil.copyfile(src, dest)
+ if permissions:
+ os.chmod(permissions)
return dest
@@ -446,3 +448,14 @@
gig = meg / 1024.0
return "%.2fG" % gig
+
+def numa_nodes():
+ node_paths = glob.glob('/sys/devices/system/node/node*')
+ nodes = [int(re.sub(r'.*node(\d+)', r'\1', x)) for x in node_paths]
+ return (sorted(nodes))
+
+
+def node_size():
+ nodes = max(len(numa_nodes()), 1)
+ return ((memtotal() * 1024) / nodes)
+
diff --git a/client/bin/cpuset.py b/client/bin/cpuset.py
index 16aee97..bc1fc86 100644
--- a/client/bin/cpuset.py
+++ b/client/bin/cpuset.py
@@ -1,15 +1,55 @@
-import os, sys, re
+__author__ = """Copyright Google, Peter Dahl, Martin J. Bligh 2007"""
+
+import os, sys, re, glob
from autotest_utils import *
class cpuset:
- def available_cpus():
- available = set(range(
+ # Convert '1-3,7,9-12' to [1,2,3,7,9,10,11,12]
+ def rangelist_to_list(rangelist):
+ result = []
+ for x in rangelist.split(','):
+ if re.match(r'^(\d+)$', x):
+ result.append(int(x))
+ continue
+ m = re.match(r'^(\d+)-(\d+)$', x)
+ if m:
+ start = int(m.group(1))
+ end = int(m.group(2))
+ result += range(start, end+1)
+ continue
+ raise 'Cannot understand data input %s' % x
+ return result
- def __init__(self, job_name, job_size, job_pid, cpus):
+
+ def available_nems(all_mems):
+ available = set(all_mems)
+ for mems in glob.glob('/dev/cpuset/*/mems'):
+ available -= set(rangelist_to_list(read_one_line(mems)))
+ return available
+
+
+ def print_one_cpuset(name):
+ dir = os.path.join('/dev/cpuset', name)
+ print "%s:" % name
+ print "\tcpus: %s" % read_one_line(dir + '/cpus')
+ mems = read_one_line(dir + '/mems')
+ print "\tmems: %s" % mems
+ memtotal = node_size() * len(rangelist_to_list(mems))
+ print "\tmemtotal: %s" % human_format(memtotal)
+ tasks = [x.rstrip() for x in open(dir + '/tasks').readlines()])
+ print "\ttasks: %s" % ','.join(tasks)
+
+
+ def print_all_cpusets():
+ for cpuset in glob.glob('/dev/cpuset/*'):
+ print_one_cpuset(re.sub(r'.*/', '', cpuset)
+
+
+ def __init__(self, name, job_size, job_pid, cpus):
# Create a cpuset container and move job_pid into it
# Allocate the list "cpus" of cpus to that container
- # job name = arbitrary string tag
+ # name = arbitrary string tag
# job size = reqested memory for job in bytes
# job pid = pid of job we're putting into the container
@@ -19,28 +59,34 @@
system('mount -t cpuset none /dev/cpuset')
cmdline = read_one_line('/proc/cmdline')
- fake_numa_nodes = re.search('numa=fake=(\d+)', cmdline).group(1)
+ all_nodes = numa_nodes()
# Bind the specificed cpus to the cpuset
- self.cpudir = "/dev/cpuset/%s" % job_name
+ self.cpudir = "/dev/cpuset/%s" % name
os.mkdir(self.cpudir)
cpu_spec = ','.join(['%d' % x for x in cpus])
- write_one_line(cpu_spec, '/dev/cpuset/%s/cpus' % job_name)
+ write_one_line(cpu_spec, '/dev/cpuset/%s/cpus' % name)
- node_size = memtotal() * 1024 / fake_numa_nodes
- num_nodes = int(math.ceil(job_size / node_size))
+ # Find some free nodes to use to create this cpuset
+ node_size = memtotal() * 1024 / len(all_nodes)
+ nodes_needed = int(math.ceil(job_size / node_size))
+ mems = available_mems(all_nodes)[-nodes_needed:]
+ alloc_size = human_format(len(mems) * node_size)
+ if len(mems) < nodes_needed:
+ raise "Insufficient memory available"
- if cpu_number == 0: # First cpuset
- m_start = 3
- m_end = m_start + num_nodes
- else:
- m_end = num_fake_numa_nodes - 1
- m_start = m_end - num_nodes
- alloc_size = (m_end - m_start + 1) * node_size
- mems = os.path.join(self.cpudir, 'mems')
- write_one_line("%d-%d" % (m_start, m_end), self.cpudir + '/mems')
- write_one_line("%d" % job_pid, self.cpudir + 'tasks')
- print "Cpuset: pid %d, cpu %d, memory %s --> nodes %d-%d --> %s" % (job_pid, cpu_number, int2KMG(job_size), m_start, m_end, int2KMG(alloc_size))
+ # Set up the cpuset
+ mems_spec = ','.join(['%d' % x for x in mems])
+ write_one_line(mems_spec, os.path.join(self.cpudir, 'mems'))
+ write_one_line('%d'%job_pid, os.path.join(self.cpudir, 'tasks'))
+
+ print "Created cpuset for pid %d, size %s" % \
+ (job_pid, human_format(job_size))
+ self.print()
+
+
+ def print(self):
+ print_one_cpuset(self.name)
def release(self):
