8:12 a.m.
The purpose is to track namespaces in use by logged processes from the
perspective of init_*_ns.
1/6 defines a function to generate them and assigns them.
Use a serial number per namespace (unique across one boot of one kernel)
instead of the inode number (which is claimed to have had the right to change
reserved and is not necessarily unique if there is more than one proc fs). It
could be argued that the inode numbers have now become a defacto interface and
can't change now, but I'm proposing this approach to see if this helps address
some of the objections to the earlier patchset.
2/6 adds access functions to get to the serial numbers in a similar way to
inode access for namespace proc operations.
3/6 implements, as suggested by Serge Hallyn, making these serial numbers
available in /proc/self/ns/{ipc,mnt,net,pid,user,uts}_snum. I chose "snum"
instead of "seq" for consistency with inum and there are a number of other uses
of "seq" in the namespace code.
4/6 exposes proc's ns entries structure which lists a number of useful
operations per namespace type for other subsystems to use.
5/6 provides an example of usage for audit_log_task_info() which is used by
syscall audits, among others. audit_log_task() and audit_common_recv_message()
would be other potential use cases.
Proposed output format:
This differs slightly from Aristeu's patch because of the label conflict with
"pid=" due to including it in existing records rather than it being a seperate
record. The serial numbers are printed in hex.
type=SYSCALL msg=audit(1399651071.433:72): arch=c000003e syscall=272 success=yes exit=0
a0=40000000 a1=ffffffffffffffff a2=0 a3=22 items=0 ppid=1 pid=483 auid=4294967295 uid=0
gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 tty=(none) ses=4294967295
comm="(t-daemon)" exe="/usr/lib/systemd/systemd" netns=97 utsns=2
ipcns=1 pidns=4 userns=3 mntns=5 subj=system_u:system_r:init_t:s0 key=(null)
6/6 tracks the creation and deletion of of namespaces, listing the type of
namespace instance, related namespace id if there is one and the newly minted
serial number.
Proposed output format:
type=NS_INIT msg=audit(1400217435.706:94): pid=524 uid=0 auid=4294967295 ses=4294967295
subj=system_u:system_r:mount_t:s0 type=20000 old_snum=0 snum=a1 res=1
type=NS_DEL msg=audit(1400217435.730:95): pid=524 uid=0 auid=4294967295 ses=4294967295
subj=system_u:system_r:mount_t:s0 type=20000 snum=a1 res=1
v2 -> v3:
Use atomic64_t in ns_serial to simplify it.
Avoid funciton duplication in proc, keying on dentry.
Squash down audit patch to avoid rcu sleep issues.
Add tracking for creation and deletion of namespace instances.
v1 -> v2:
Avoid rollover by switching from an int to a long long.
Change rollover behaviour from simply avoiding zero to raising a BUG.
Expose serial numbers in /proc/<pid>/ns/*_snum.
Expose ns_entries and use it in audit.
Notes:
There has been some progress made for audit in net namespaces and pid
namespaces since this previous thread. net namespaces are now served as peers
by one auditd in the init_net namespace with processes in a non-init_net
namespace being able to write records if they are in the init_user_ns and have
CAP_AUDIT_WRITE. Processes in a non-init_pid_ns can now similarly write
records. As for CAP_AUDIT_READ, I just posted a patchset to check capabilities
of userspace processes that try to join netlink broadcast groups.
This set does not try to solve the non-init namespace audit messages and
auditd problem yet. That will come later, likely with additional auditd
instances running in another namespace with a limited ability to influence the
master auditd. I echo Eric B's idea that messages destined for different
namespaces would have to be tailored for that namespace with references that
make sense (such as the right pid number reported to that pid namespace, and
not leaking info about parents or peers).
Bugs:
Patch 6/6 has a timing bug such that mnt and net namespace initial namespaces
never get logged, I suspect because they are initialized before the audit
subsystem. I've tried moving audit from __initcall to subsys_initcall, but
that doesn't help.
Questions:
Is there a way to link serial numbers of namespaces involved in migration of a
container to another kernel? It sounds like what is needed is a part of a
mangement application that is able to pull the audit rcords from constituent
hosts to build an audit trail of a container.
What additional events should list this information?
Does this present any problematic information leaks? Only CAP_AUDIT_CONTROL
(and proposed CAP_AUDIT_READ) in init_user_ns can get to this information in
the init namespace at the moment from audit. *However*, the addition of the
proc/<pid>/ns/*_snum does make it available to other processes now.
Richard Guy Briggs (6):
namespaces: assign each namespace instance a serial number
namespaces: expose namespace instance serial number in proc_ns_operations
namespaces: expose ns instance serial numbers in proc
namespaces: expose ns_entries
audit: log namespace serial numbers
audit: log creation and deletion of namespace instances
fs/mount.h | 1 +
fs/namespace.c | 12 +++++++++
fs/proc/namespaces.c | 35 +++++++++++++++++++-------
include/linux/audit.h | 15 +++++++++++
include/linux/ipc_namespace.h | 1 +
include/linux/nsproxy.h | 8 ++++++
include/linux/pid_namespace.h | 1 +
include/linux/proc_ns.h | 2 +
include/linux/user_namespace.h | 1 +
include/linux/utsname.h | 1 +
include/net/net_namespace.h | 1 +
include/uapi/linux/audit.h | 2 +
init/version.c | 1 +
ipc/msgutil.c | 1 +
ipc/namespace.c | 20 +++++++++++++++
kernel/audit.c | 53 +++++++++++++++++++++++++++++++++++++++-
kernel/nsproxy.c | 17 +++++++++++++
kernel/pid.c | 1 +
kernel/pid_namespace.c | 19 ++++++++++++++
kernel/user.c | 1 +
kernel/user_namespace.c | 18 +++++++++++++
kernel/utsname.c | 20 +++++++++++++++
net/core/net_namespace.c | 20 ++++++++++++++-
23 files changed, 240 insertions(+), 11 deletions(-)