Setting tag lxc-userspace-1.0-11
[lxc-userspace.git] / lxcsu
diff --git a/lxcsu b/lxcsu
index dc24ec1..b05342a 100755 (executable)
--- a/lxcsu
+++ b/lxcsu
 #!/usr/bin/python
 
-
-import setns
-import os
 import sys
+import os
+import setns
+import pwd
+
+from argparse import ArgumentParser
+
+# can set to True here, but also use the -d option
+debug = False
+
+def getarch(f):
+    output = os.popen('readelf -h %s 2>&1'%f).readlines()
+    classlines = [x for x in output if ('Class' in x.split(':')[0])]
+    line = classlines[0]
+    c = line.split(':')[1]
+    if ('ELF64' in c):
+        return 'x86_64'
+    elif ('ELF32' in c):
+        return 'i686'
+    else:
+        raise Exception('Could not determine architecture')
+
+def get_cgroup_subdirs_for_pid(pid):
+    cgroup_info_file = '/proc/%s/cgroup'%pid
+    cgroup_lines = open(cgroup_info_file).read().splitlines()
+    
+    subdirs = {}
+    for line in cgroup_lines:
+        try:
+            _, cgroup_name, subdir = line.split(':')
+            subdirs[cgroup_name] = subdir
+        except Exception, e:
+            print "Error reading cgroup info: %s"%str(e)
+            pass
+    
+    return subdirs
+        
+    
+def umount(fs_dir, opts=''):
+    output = os.popen('/bin/umount %s %s 2>&1'%(opts, fs_dir)).read()
+    return ('device is busy' not in output)
+
+def main ():
+    parser = ArgumentParser()
+    parser.add_argument("-n", "--nonet",
+                                       action="store_true", dest="no_netns", default=False,
+                                       help="Don't enter network namespace")
+    parser.add_argument("-m", "--nomnt",
+                                       action="store_true", dest="no_mntns", default=False,
+                                       help="Don't enter mount namespace")
+    parser.add_argument("-p", "--nopid",
+                                       action="store_true", dest="no_pidns", default=False,
+                                       help="Don't enter pid namespace")
+    parser.add_argument("-r", "--root",
+                                       action="store_true", dest="root", default=False,
+                                       help="Enter as root: be careful")
+    parser.add_argument("-i","--internal",
+                                       action="store_true", dest="internal", default=False,
+                                       help="does *not* prepend '-- -c' to arguments - or invoke lxcsu-internal")
+    parser.add_argument("-d","--debug",
+                                       action='store_true', dest='debug', default=False,
+                                       help="debug option")
+    parser.add_argument("-s","--nosliceuid",
+                                       action='store_true', dest="nosliceuid", default=False,
+                                       help="do not change to slice uid inside of slice")
+    parser.add_argument("-o","--noslicehome",
+                                       action='store_true', dest="noslicehome", default=False,
+                                       help="do not change to slice home directory inside of slice")
+
+    if os.path.exists("/etc/lxcsu_default"):
+        defaults = parser.parse_args(file("/etc/lxcsu_default","r").read().split())
+        parser.set_defaults(**defaults.__dict__)
+
+    parser.add_argument ("slice_name")
+    parser.add_argument ("command_to_run",nargs="*")
+
+    args = parser.parse_args()
+    slice_name=args.slice_name
+
+    # support for either setting debug at the top of this file, or on the command-line
+    if args.debug:
+        global debug
+       debug=True
+
+    # somehow some older nodes won't be able to find the login name in /etc/passwd 
+    # when this is done down the road, so compute slice_uid while in a safe env
+    # even though we don't use the slice_uid any more, this is still 
+    # checked later on as a means to ensure existence of the slice account
+    try:
+        slice_uid = pwd.getpwnam(slice_name).pw_uid
+    except Exception, e:
+        if debug:
+            import traceback
+            print 'error while computing slice_uid',e
+            traceback.print_exc()
+        slice_uid=None
+
+    # unless we run the symlink 'lxcsu-internal', or we specify the -i option, prepend '--' '-c'
+    if sys.argv[0].find('internal')>=0: args.internal=True
+
+    if len(args.command_to_run)>0 and (args.command_to_run[0] == "/sbin/service"):
+        # A quick hack to support nodemanager interfaces.py when restarting
+        # networking in a slice.
+        args.nosliceuid = True
+
+    # plain lxcsu
+    if not args.internal:
+       # no command given: enter interactive shell
+       if not args.command_to_run: args.command_to_run=['/bin/sh']
+       args.command_to_run = [ '-c' ] + [" ".join(args.command_to_run)]
 
-#from optparse import OptionParser
-
-drop_capabilities='cap_sys_admin,cap_sys_boot,cap_sys_module'.split(',')
-
-"""
-parser = OptionParser()
-parser.add_option("-n", "--net",
-                  action="store_true", dest="netns", default=False,
-                  help="Enter network namespace")
-parser.add_option("-m", "--mnt",
-                  action="store_true", dest="mntns", default=False,
-                  help="Enter mount namespace")
-parser.add_option("-p", "--pid",
-                  action="store_true", dest="pidns", default=False,
-                  help="Enter pid namespace")
-
-(options, args) = parser.parse_args()
-"""
-args = sys.argv[1:]
-
-try:
-       slice_name = args[0]
-except IndexError:
-       print "You must specify a vm name"
+    try:
+        cmd = '/usr/bin/virsh --connect lxc:/// domid %s'%slice_name
+        # convert to int as a minimal raincheck
+       driver_pid = int(os.popen(cmd).read().strip())
+        # locate the pid for the - expected - single child, that would be the init for that VM
+        #init_pid = int(open("/proc/%s/task/%s/children"%(driver_pid,driver_pid)).read().strip())
+       init_pid = int(os.popen('pgrep -P %s'%driver_pid).readlines()[0].strip())
+        # Thierry: I am changing the code below to use child_pid instead of driver_pid
+        # for the namespace handling features, that I was able to check
+        # I've left the other ones as they were, i.e. using driver_pid, but I suspect
+        # they chould be changed as well
+
+    except:
+        print "Domain %s not found"%slice_name
        exit(1)
 
-try:
-       cmd = 'grep %s /proc/*/cgroup | grep freezer'%slice_name
-       output = os.popen(cmd).readlines()
-except:
-       print "Error finding slice %s"%slice_name
+    if not driver_pid or not init_pid:
+        print "Domain %s not started"%slice_name
        exit(1)
 
-slice_spec = None
-for e in output:
-       try:
-               l = e.rstrip()
-               path = l.split(':')[0]  
-               comp = l.rsplit(':')[-1]
-               slice_name_check = comp.rsplit('/')[-1]
-
-               if (slice_name_check == slice_name):
-                       slice_path = path
-                       pid = slice_path.split('/')[2]
-                       cmdline = open('/proc/%s/cmdline'%pid).read().rstrip('\n\x00')
-                       if (cmdline == '/sbin/init'):
-                               slice_spec = slice_path
-                               break
-       except:
+    if debug: print "Found driver_pid",driver_pid,'and init_pid=',init_pid
+    # xxx probably init_pid here too
+    arch = getarch('/proc/%s/exe'%driver_pid)
+
+    # Set sysctls specific to slice
+    sysctls = []
+    sysctl_dir = '/etc/planetlab/vsys-attributes/%s'%slice_name
+    if (os.access(sysctl_dir,0)):
+        entries = os.listdir(sysctl_dir)
+       for e in entries:
+            prefix = 'vsys_sysctl.'
+           if (e.startswith(prefix)):
+               sysctl_file = '/'.join([sysctl_dir,e])
+               sysctl_name = e[len(prefix):]
+               sysctl_val = open(sysctl_file).read()
+               sysctls.append((sysctl_file, sysctl_name, sysctl_val))
+
+    # xxx probably init_pid here too
+    subdirs = get_cgroup_subdirs_for_pid(driver_pid) 
+    sysfs_root = '/sys/fs/cgroup'
+
+    # If the slice is frozen, then we'll get an EBUSY when trying to write to the task
+    # list for the freezer cgroup. Since the user couldn't do anything anyway, it's best
+    # in this case to error out the shell. (an alternative would be to un-freeze it,
+    # add the task, and re-freeze it)
+    # Enter cgroups
+    current_cgroup = ''
+    for subsystem in ['cpuset','memory','blkio','cpuacct','cpuacct,cpu','freezer']:
+        try:
+            current_cgroup = subsystem
+
+            # There seems to be a bug in the cgroup schema: cpuacct,cpu can become cpu,cpuacct
+            # We need to handle both
+            task_path_alt = None
+            try:
+               subsystem_comps = subsystem.split(',')
+               subsystem_comps.reverse()
+               subsystem_alt = ','.join(subsystem_comps)
+               tasks_path_alt = [sysfs_root, subsystem_alt, subdirs[subsystem], 'tasks']
+            except Exception,e:
+                pass
+               
+            tasks_path = [sysfs_root,subsystem,subdirs[subsystem],'tasks']
+            tasks_path_str = '/'.join(tasks_path)
+     
+            try:
+                f = open(tasks_path_str, 'w')
+            except:
+                tasks_path_alt_str = '/'.join(tasks_path_alt)
+                f = open(tasks_path_alt_str, 'w')
+
+            f.write(str(os.getpid()))
+            if (subsystem=='freezer'):
+                f.close()
+
+        except Exception,e:
+            if (not subdirs.has_key(subsystem)):
+                pass
+            else:
+                if debug: print e 
+                print "Error assigning cgroup %s (pid=%s) for slice %s"%(current_cgroup,driver_pid, slice_name)
+                exit(1)
+
+
+    def chcontext (path):
+        retcod = setns.chcontext (path)
+        if retcod != 0:
+            print 'WARNING - setns(%s)=>%s (ignored)'%(path,retcod)
+        return retcod
+
+    # Use init_pid and not driver_pid to locate reference namespaces
+    ref_ns = "/proc/%s/ns/"%init_pid
+
+    if True:                    chcontext(ref_ns+'uts')
+    if True:                    chcontext(ref_ns+'ipc')
+       
+    if (not args.no_pidns):     chcontext(ref_ns+'pid')
+    if (not args.no_netns):     chcontext(ref_ns+'net')
+    if (not args.no_mntns):     chcontext(ref_ns+'mnt')
+
+    proc_mounted = False
+    if (not os.access('/proc/self',0)):
+       proc_mounted = True
+       setns.proc_mount()
+
+    for (sysctl_file, sysctl_name, sysctl_val) in sysctls:
+       for fn in ["/sbin/sysctl", "/usr/sbin/sysctl", "/bin/sysctl", "/usr/bin/sysctl"]:
+           if os.path.exists(fn):
+               os.system('%s -w %s=%s  >/dev/null 2>&1'%(fn, sysctl_name,sysctl_val))
                break
+           else:
+               print "Error: image does not have a sysctl binary"
 
-if (not slice_spec or not pid):
-    print "Not started: %s"%slice_name
-    exit(1)
-
-# Enter cgroups
-try:
-    for subsystem in ['cpuset','memory','blkio']:
-        open('/sys/fs/cgroup/%s/libvirt/lxc/%s/tasks'%(subsystem,slice_name),'w').write(str(os.getpid()))
-
-except:
-    print "Error assigning resources: %s"%slice_name
-    exit(1)
-
-try:
-    open('/sys/fs/cgroup/cpuacct/system/libvirtd.service/libvirt/lxc/%s/tasks'%slice_name,'w').write(str(os.getpid()))
-except:
-    print "Error assigning cpuacct: %s" % slice_name
-    exit(1)
-
-# If the slice is frozen, then we'll get an EBUSY when trying to write to the task
-# list for the freezer cgroup. Since the user couldn't do anything anyway, it's best
-# in this case to error out the shell. (an alternative would be to un-freeze it,
-# add the task, and re-freeze it)
-try:
-    f=open('/sys/fs/cgroup/freezer/libvirt/lxc/%s/tasks'%(slice_name),'w')
-    f.write(str(os.getpid()))
-    # note: we need to call f.close() explicitly, or we'll get an exception in
-    # the object destructor, which will not be caught
-    f.close()
-except:
-    print "Error adding task to freezer cgroup. Slice is probably frozen: %s" % slice_name
-    exit(1)
-
-r1 = setns.chcontext('/proc/%s/ns/uts'%pid)
-r2 = setns.chcontext('/proc/%s/ns/ipc'%pid)
-r3 = setns.chcontext('/proc/%s/ns/net'%pid)
-
-open('/proc/lxcsu','w').write(pid)
-open('/proc/pidsu','w').write(pid)
-
-pid = os.fork()
-
-cap_args = map(lambda c:'--drop='+c, drop_capabilities)
-    
-if (pid == 0):
-    os.environ['SHELL'] = '/bin/sh'
-    os.execv('/usr/sbin/capsh',cap_args+['--','--login']+args[1:])
-else:
-    os.waitpid(pid,0)
+    # cgroups is not yet LXC-safe, so we need to use the coarse grained access control
+    # strategy of unmounting the filesystem
+
+    umount_result = True
+    for subsystem in ['cpuset','cpu,cpuacct','memory','devices','freezer','net_cls','blkio','perf_event','systemd']:
+       fs_path = '/sys/fs/cgroup/%s'%subsystem
+       if (not umount(fs_path,'-l')):
+            print 'WARNING - umount failed (ignored) with path=',fs_path
+           pass
+           # Leaving these comments for historical reference
+           #print "Error disabling cgroup access"
+           #exit(1) - Don't need this because failure here implies failure in the call to umount /sys/fs/cgroup
+
+    if (not umount('/sys/fs/cgroup')):
+        print "Error disabling cgroup access"
+        exit(1)
+
+    fork_pid = os.fork()
+
+    if (fork_pid == 0):
+        if (not args.root):
+            setns.drop_caps() 
+            if (args.nosliceuid):
+                # we still want to drop capabilities, but don't want to switch UIDs
+                exec_args = [arch,'/bin/sh','--login',]+args.command_to_run
+            else:
+                if not slice_uid:
+                    print "lxcsu could not spot %s in /etc/passwd - exiting"%slice_name
+                    exit(1)
+                exec_args = [arch,'/usr/bin/sudo','-u',slice_name,'/bin/sh','--login',]+args.command_to_run
+# once we can drop f12, it would be nicer to instead go for
+# exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--user=%s'%slice_name,'--login',]+args.command_to_run
+        else:
+            exec_args = [arch,'/bin/sh','--login']+args.command_to_run
+
+        os.environ['SHELL'] = '/bin/sh'
+        if os.path.exists('/etc/planetlab/lib/bind_public.so'):
+            os.environ['LD_PRELOAD'] = '/etc/planetlab/lib/bind_public.so'
+        if not args.noslicehome:
+            os.environ['HOME'] = '/home/%s'%slice_name
+            os.chdir("/home/%s"%(slice_name))
+        if debug: print 'lxcsu:execv:','/usr/bin/setarch',exec_args
+        os.execv('/usr/bin/setarch',exec_args)
+    else:
+        setns.proc_umount()
+        _,status = os.waitpid(fork_pid,0)
+        exit(os.WEXITSTATUS(status))
+
+if __name__ == '__main__':
+       main()