X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lxcsu;h=3ee60357ef57ee5f67e8adb9fae7957d0e0a71a4;hb=4d356cca5a99de1f38c7a955966c3f4580651eeb;hp=a9ff3d67c8981a14d3799885a9020edbb20912d5;hpb=a6adbc76585a1a2a46f9ab3011cdb91d6aa42507;p=lxc-userspace.git diff --git a/lxcsu b/lxcsu index a9ff3d6..3ee6035 100755 --- a/lxcsu +++ b/lxcsu @@ -1,19 +1,18 @@ -#!/usr/bin/python +#!/usr/bin/python3 import sys import os import setns -import pdb +import pwd from argparse import ArgumentParser -drop_capabilities='cap_sys_admin,cap_sys_boot,cap_sys_module' - # 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() +def getarch(pid): + exe_filename='/proc/%s/exe'%pid + output = os.popen('readelf -h %s 2>&1'%exe_filename).readlines() classlines = [x for x in output if ('Class' in x.split(':')[0])] line = classlines[0] c = line.split(':')[1] @@ -22,24 +21,24 @@ def getarch(f): elif ('ELF32' in c): return 'i686' else: - raise Exception('Could not determine architecture') + raise Exception('Could not determine architecture for pid %s'%pid) 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) + except Exception as 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) @@ -47,13 +46,13 @@ def umount(fs_dir, opts=''): def main (): parser = ArgumentParser() parser.add_argument("-n", "--nonet", - action="store_true", dest="netns", default=False, + action="store_true", dest="no_netns", default=False, help="Don't enter network namespace") parser.add_argument("-m", "--nomnt", - action="store_true", dest="mntns", default=False, + action="store_true", dest="no_mntns", default=False, help="Don't enter mount namespace") parser.add_argument("-p", "--nopid", - action="store_true", dest="pidns", default=False, + 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, @@ -81,6 +80,24 @@ def main (): 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 as 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 @@ -95,29 +112,29 @@ def main (): if not args.command_to_run: args.command_to_run=['/bin/sh'] args.command_to_run = [ '-c' ] + [" ".join(args.command_to_run)] - # support for either setting debug at the top of this file, or on the command-line - if args.debug: - global debug - debug=True - try: cmd = '/usr/bin/virsh --connect lxc:/// domid %s'%slice_name - pidnum = int(os.popen(cmd).read().rstrip()) + # 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 init_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 + # some should be changed as well + except: - print "Domain %s not found"%slice_name + print("Domain %s not found"%slice_name) exit(1) - pid = '%s'%pidnum - if debug: print "Found pidnum",pidnum - cmdline = open('/proc/%s/cmdline'%pidnum).read().rstrip('\n\x00') - arch = getarch('/proc/%s/exe'%pid) - - if (not pid): - print "Domain %s not started"%slice_name + if not driver_pid or not init_pid: + print("Domain %s not started"%slice_name) exit(1) - if arch is None: - arch = 'x86_64' + if debug: print("Found driver_pid",driver_pid,'and init_pid=',init_pid) + # driver_pid is always x86_64, we need to look at the VM's init process here + arch = getarch(init_pid) # Set sysctls specific to slice sysctls = [] @@ -132,7 +149,8 @@ def main (): sysctl_val = open(sysctl_file).read() sysctls.append((sysctl_file, sysctl_name, sysctl_val)) - subdirs = get_cgroup_subdirs_for_pid(pid) + # 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 @@ -153,12 +171,12 @@ def main (): subsystem_comps.reverse() subsystem_alt = ','.join(subsystem_comps) tasks_path_alt = [sysfs_root, subsystem_alt, subdirs[subsystem], 'tasks'] - except Exception,e: + except Exception as e: pass - + tasks_path = [sysfs_root,subsystem,subdirs[subsystem],'tasks'] tasks_path_str = '/'.join(tasks_path) - + try: f = open(tasks_path_str, 'w') except: @@ -169,32 +187,30 @@ def main (): if (subsystem=='freezer'): f.close() - except Exception,e: - if (not subdirs.has_key(subsystem)): + except Exception as e: + if (subsystem not in subdirs): pass else: - if debug: print e - print "Error assigning cgroup %s (%s) for slice %s"%(current_cgroup,pid, slice_name) + 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) + print('WARNING - setns(%s)=>%s (ignored)'%(path,retcod)) return retcod - chcontext('/proc/%s/ns/uts'%pid) - chcontext('/proc/%s/ns/ipc'%pid) - - if (not args.pidns): - chcontext('/proc/%s/ns/pid'%pid) + # Use init_pid and not driver_pid to locate reference namespaces + ref_ns = "/proc/%s/ns/"%init_pid - if (not args.netns): - chcontext('/proc/%s/ns/net'%pid) + if True: chcontext(ref_ns+'uts') + if True: chcontext(ref_ns+'ipc') - if (not args.mntns): - chcontext('/proc/%s/ns/mnt'%pid) + 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)): @@ -207,7 +223,7 @@ def main (): 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" + print("Error: image does not have a sysctl binary") # cgroups is not yet LXC-safe, so we need to use the coarse grained access control # strategy of unmounting the filesystem @@ -216,55 +232,33 @@ def main (): 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 + 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" + print("Error disabling cgroup access") exit(1) - pid = os.fork() - - # capsh has a --user option starting with f14 - # so if only for f12 we need to fake this one - # - # capsh.c does essentially the following when invoked with --user: - # pwd = getpwnam(user); ... - # ngroups = MAX_GROUPS; - # status = getgrouplist(user, pwd->pw_gid, groups, &ngroups); ... - # status = setgroups(ngroups, groups); ... - # status = setgid(pwd->pw_gid); ... - # status = setuid(pwd->pw_uid); ... - # however we cannot simulate that ourselves because if we did in this process then - # capsh could not be allowed to mess with caps any more - - def getuid (slicename): - import pwd - try: - return pwd.getpwnam(slicename).pw_uid - except: - return - - if (pid == 0): - cap_arg = '--drop='+drop_capabilities + 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,'/usr/sbin/capsh',cap_arg,'--','--login',]+args.command_to_run + exec_args = [arch,'/bin/sh','--login',]+args.command_to_run else: - uid = getuid (slice_name) - if not uid: - print "lxcsu could not spot %s in /etc/passwd - exiting"%slice_name + if not slice_uid: + print("lxcsu could not spot %s in /etc/passwd - exiting"%slice_name) exit(1) - exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--uid=%s'%uid,'--','--login',]+args.command_to_run + 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 +# exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--user=%s'%slice_name,'--login',]+args.command_to_run else: - exec_args = [arch,'/usr/sbin/capsh','--','--login']+args.command_to_run + 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'): @@ -272,11 +266,11 @@ def main (): 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 + if debug: print('lxcsu:execv:','/usr/bin/setarch',exec_args) os.execv('/usr/bin/setarch',exec_args) else: setns.proc_umount() - _,status = os.waitpid(pid,0) + _,status = os.waitpid(fork_pid,0) exit(os.WEXITSTATUS(status)) if __name__ == '__main__':