On Wed, Mar 8, 2023 at 7:13 PM Bruce Elrick <bruce.elrick(a)canonical.com&gt; wrote: > Hello all, > > I'm not sure if this list is appropriate for questions so please let > me know and otherwise ignore if this message is not appropriate. > > I'm trying to help someone who is finally migrating from iptables to > nftables on the back-end and needs to therefore migrate their audit > capability. > > Currently they have a single simple audit rule to detect when there is > a iptable change from any audit user apart from their service user > using a rule like the accepted answer given in this[0] StackExchange > question, although with added filters on the auid (I have to admit I > don't know the origin of auid=-1 events): > > auditctl -a exit,always -F arch=b64 -F a2=64 -F auid!=-1 -F > auid!=${serviceuser_uid} -S setsockopt -k iptablesChange > > They are migrating from Ubuntu bionic to jammy and still using the > iptables front-end but since the back-end changes from default > iptables to default nftables they need to change their audit rules > > They did strace testing and noted the syscall changing from > > setsockopt(4, SOL_IP, IPT_SO_SET_REPLACE, > "filter\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., > 80952) = 0 > > to > > sendto(3, [{nlmsg_len=20, > nlmsg_type=NFNL_SUBSYS_NFTABLES<<8|NFT_MSG_GETGEN, > nlmsg_flags=NLM_F_REQUEST, nlmsg_seq=0, nlmsg_pid=0}, > {nfgen_family=AF_UNSPEC, version=NFNETLINK_V0, res_id=htons(0)}], 20, > 0, {sa_family=AF_NETLINK, nl_pid=0, nl_groups=00000000}, 12) = 20 > > between the two versions. > > In my own testing, I decided to approach from the audit tools > perspective so I created a broad rule to capture all system call > related to a test user: > > auditctl -a always,exit -S all -F auid=1001 # 1001 is uid of testuser > > Then I tried various operations using my testuser such as > iptables-restore of either a default-accept rule set with no rules or > with one or two simple drop rules. I also tested adding just a single > iptables rule. I then used ausearch to discover what the audit system > captured: > > # ausearch -i -m NETFILTER_CFG > ... > ---- > type=PROCTITLE msg=audit(03/07/2023 17:18:55.152:143044) : > proctitle=iptables-restore > type=SYSCALL msg=audit(03/07/2023 17:18:55.152:143044) : > arch=x86_64 syscall=sendmsg success=yes exit=764 a0=0x3 > a1=0x7ffdb0e98db0 a2=0x0 a3=0x7ffdb0e98d9c items=0 ppid=5673 pid=5676 > auid=testuser uid=root gid=root euid=root suid=root fsuid=root > egid=root sgid=root fsgid=root tty=pts2 ses=108 comm=iptables-restor > exe=/usr/sbin/xtables-nft-multi subj=unconfined key=(null) > type=NETFILTER_CFG msg=audit(03/07/2023 17:18:55.152:143044) : > table=filter:30 family=ipv4 entries=12 op=nft_unregister_table > pid=5676 subj=unconfined comm=iptables-restor > type=NETFILTER_CFG msg=audit(03/07/2023 17:18:55.152:143044) : > table=filter:30 family=ipv4 entries=7 op=nft_register_chain pid=5676 > subj=unconfined comm=iptables-restor > ---- > type=PROCTITLE msg=audit(03/07/2023 17:23:04.390:144459) : > proctitle=sudo /usr/sbin/iptables -A OUTPUT -d 10.100.249.64 -j DROP > type=SOCKADDR msg=audit(03/07/2023 17:23:04.390:144459) : saddr={ > saddr_fam=netlink nlnk-fam=16 nlnk-pid=0 } > type=SYSCALL msg=audit(03/07/2023 17:23:04.390:144459) : > arch=x86_64 syscall=sendmsg success=yes exit=304 a0=0x3 > a1=0x7ffc80659110 a2=0x0 a3=0x7ffc806590fc items=0 ppid=5703 pid=5704 > auid=testuser uid=root gid=root euid=root suid=root fsuid=root > egid=root sgid=root fsgid=root tty=pts2 ses=108 comm=iptables > exe=/usr/sbin/xtables-nft-multi subj=unconfined key=(null) > type=NETFILTER_CFG msg=audit(03/07/2023 17:23:04.390:144459) : > table=filter:31 family=ipv4 entries=1 op=nft_register_rule pid=5704 > subj=unconfined comm=iptables > > The event sequences seem to make sense with the sockaddr function > selecting the netlink family which agrees with the strace output. > > With the change in the back-end to nftables, I can see in either case > that the setsockopt system call with a nice, crisp, single argument > (a2=64/IPT_SO_SET_REPLACE) option with either a sendto or sendmsg > system call but with a pointer to a message structure. I read that > audit rules cannot filter using data inside struct arguments. > > My naive interpretation of this is that I'd need to have a rule that > captures all sendmsg syscalls with (auid!=-1 and > auid!=${serviceuser_uid} but I don't know enough about socket syscall > usage to know whether this is too much. I see that write(2) to a > socket is the same as send(2) without the flags so I might assume that > most socket syscalls that are sending data use write(2) and not > send/sendto/sendmsg(2) but I worry this would be too much audit data. > > Anyone care to comment or point me in the correct direction? The problem I think you're going to have, and I believe you've already suspected it, is that auditing socket writes is going to result in a firehose of records. However, unless you have an exclude filter for NETFILTER_CFG records I believe they will be generated without an explicit filter rule triggering their generation. Or am I misunderstanding your question? -- paul-moore.com

On Thu, Mar 9, 2023 at 11:33 AM Bruce Elrick <bruce.elrick(a)canonical.com&gt; wrote: > > I think I need to clarify where I'm confused ;-) > > With iptables you could write a rule that would only catch system > calls that were for iptables changes. That is, you didn't need to > capture *all* setsockopt calls (not that there would be lots of > *those*) but rather you could add the a2=64 to only get the > op=IPT_SO_SET_REPLACE ones. > > With netfilter, however, since the control interface is netlink and > netlink requires a message to a socket and messages are structs, there > is no way to have a similarly narrow audit rule as in the case of > iptables. > > That's the first thing I want to confirm: whether my understanding > above is correct? Yes, you are correct. > I'm confused because your answer implies I'm correct > but you didn't explicitly confirm that my interpretation of how it > works was correct. > > You talk about having an exclude filter on NETFILTER_CFG (or rather > exclude everything except NETFILTER_CFG??) but my understanding is > that you can only do that filtering after the fact using ausearch or > writing some sort of correlation code using the auparse library. The kernel implements an exclude filter which is described in the auditctl(8) manpage: "Add a rule to the event type exclusion filter list. This list is used to filter events that you do not want to see. For example, if you do not want to see any avc messages, you would using this list to record that. Events can be excluded by process ID, user ID, group ID, login user ID, message type, subject context, or executable name. The action is ignored and uses its default of "never". Taken from https://man7.org/linux/man-pages/man8/auditctl.8.html However, in my last reply I wasn't advocating for this use of the exclude filter, I was simply trying to explain that unless you are explicitly excluding the creation of NETFILTER_CFG records via the exclude filter you should be seeing NETFILTER_CFG in your audit stream with basic auditing enabled.

> It just seemed surprising that there is a non-trivial loss of audit > functionality but that I could not find any obvious discussion about > that. By obvious discussion I mean as explicitly as what I'm trying to > say here. Unfortunately it is a fairly common practice for kernel features to be added, and removed, without consulting with the various Linux Kernel security developers, e.g. audit, SELinux, LSM, etc. Sometimes we are successful in retrofitting the necessary security and/or auditing hooks, sometimes we are limited due to design choices. > The other thing I'm trying to understand is how heavy an audit load > would it be to have an audit rule that captures *all* sendmsg calls > (well, all except where auid=-1 or auid=${serviceuser_uid}). I don't > have a good enough understanding of systems programming to know where > and how often the sendmsg is called. Of course I know this is highly > dependent on workload, but my knowledge is limited enough that I I can > convince myself both that the audit load would be not trivial but > still manageable in most cases but also I can convince myself that no > same sysadmin would consider running such an audit rule. With file IO > it's easy to distinguish that file opens are worth auditing but file > reads and writes would be insane to audit. It's not so clear for me > for sockets. This is going to be dependent on both the workloads and applications used on the system, there is no one "right" answer here. -- paul-moore.com -- Linux-audit mailing list Linux-audit(a)redhat.com https://listman.redhat.com/mailman/listinfo/linux-audit

Anyway, I think I need to spend some time playing until that "aha!" moment comes. It's feels a lot closer thanks to both of your responses and I really apprecaite the time you've taken to read my emails and respond to them.

On Fri, Mar 10, 2023 at 9:36 AM Steve Grubb <sgrubb(a)redhat.com&gt; wrote: > On Thursday, March 9, 2023 5:52:28 PM EST Bruce Elrick wrote: > > Anyway, I think I need to spend some time playing until that "aha!" > > moment comes. It's feels a lot closer thanks to both of your responses > > and I really apprecaite the time you've taken to read my emails and > > respond to them. > > There are simple events which are one line and compound events which are > multiple lines - called records. The simple events tend to be hardwired and > not optional. For example, logins are hardwired; kernel config changes are > hardwired; authentication is hardwired. Reading Steve's response I'm not sure we use the same terminology, or perhaps we explain it a bit differently. Regardless, here is a quick definition that I stick to when discussing audit: "audit record": An audit record is a single line in the audit log that consists of a timestamp, record type (type=XXX), and a series of fields which are dependent on the record type. Here is an example of a SYSCALL record: type=SYSCALL msg=audit(03/10/2023 10:59:00.797:563) : arch=x86_64 syscall=bpf success=yes exit=12 a0=BPF_PROG_LOAD a1=0x7ffde0efc650 a2=0x80 a3=0x13 items=0 ppid=1 pid=2683 auid=root uid=root gid=root euid=root suid=root fsuid=root egid=root sgid=root fsgid=root tty=(none) ses=10 comm=systemd exe=/usr/lib/systemd/systemd subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 key=(null) "audit event": An audit event consists of multiple audit records grouped together by a single timestamp. Single record audit events are allowed and do exist. There is no upper bound on the number of records allowed in an audit event. Here is an example of an audit event consisting of PROCTITLE, SYSCALL, and BPF audit records: type=PROCTITLE msg=audit(03/10/2023 10:59:00.797:563) : proctitle=(systemd) type=SYSCALL msg=audit(03/10/2023 10:59:00.797:563) : arch=x86_64 syscall=bpf success=yes exit=12 a0=BPF_PROG_LOAD a1=0x7ffde0efc650 a2=0x80 a3=0x13 items=0 ppid=1 pid=2683 auid=root uid=root gid=root euid=root suid=root fsuid=root egid=root sgid=root fsgid=root tty=(none) ses=10 comm=systemd exe=/usr/lib/systemd/systemd subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 key=(null) type=BPF msg=audit(03/10/2023 10:59:00.797:563) : prog-id=172 op=LOAD

I hope that helps. -- paul-moore.com