- Fixed configuration loading, when the Process Monitor Method
selected failed to work (e.g.: when ebpf modules are missing,
auditd is not configured, etc).
In this scenario, on the first run, the daemon was not loading the
configuration from disk, and was instead using the default values, which
led to some glitches (unexpected reconnections to the GUI, checksums
option ignored, etc).
- added more tests.
Discard events from the queue if we can't send events to the channel.
It can happen if the case of a deadlock such as in ac3053af327e38acad4ff4024ed688ed68d6ae30
In this situation, reading from the ring buffer is blocked, which
eventually causes it to fill up.
When the ring buffer is full, bpf_ringbuf_output() fails with error -11.
Under very high load of kernel events, combined with a
non-zero value for the ebpf.QueueEventsSize config option,
the cache could end up blocking the events workers if the option to
compute process checksums was enabled.
When receiving the paths of a binary from kernel, the path can be a
symlink to the real binary on disk, and we were not obtaining the real
path to the binary.
Now we'll try to obtain it. If we fail, we'll use the path received from
the kernel. And if it's empty, we'll fallback to read it from proc.
Closes#1357
netlink.AddrSubscribeWithOptions() was causing systemd-resolved not to
obtain DNS servers, and therefore the system was unable to resolve
domain names.
This ssue started happening on Fedora 42, with systemd-resolved 257 and
kernels 6.14.x.
Using netlink.AddrSubscribe() solves this issue for now.
Closes#1343
Changed Error to Warning log level when failing to load hooks.
Except for udp_tunnel6_xmit_skb, which depends on the module
ip6_udp_tunnel to be loaded, which by default is not the case.
In order not to flood the logs with this error, log level changed to
Debug.
In order to intercept UDP tunnels traffic, we use iptunnel_xmit for IPv4
connections. The equivalent of IPv6 would be ip6tunnel_xmit, but
unfortunately it's not public.
So instead of ip6tunnel_xmit, we'll use udp_tunnel6_xmit_skb.
This function receives two parameters besides the struct sock*, in6_addr
*saddr and in6_addr *daddr. If these addresses are not set, usually the
connection is a localhost connection.
On the other hand, this function is defined in the module
ip6_udp_tunnel. If this module is not loaded when the daemon starts,
we'll fail to intercept the connections.
Closes: #1250
- Preprend the chroot of a process to the path:
Even if a process is executed from a chroot (or a different mount
namespace), the kernel reports the path as it were from the host.
For chrooted processes, now we prepend the chroot to the path:
Instead of /usr/bin/curl -> /var/cache/pbuilder/usr/bin/curl
Note: if the process is emulated (with qemu for example), the
checksum will be of the emulator.
- For proceses launched from /proc, like /proc/self/exe or
/proc/<pid>/fd/<number>, we were replacing in the cmdline "/proc/*"
with the absolute path of the binary:
Path: /usr/bin/curl
Cmdline reported: /proc/<pid>/fd/<number> -L 1.1.1.1
Cmdline "fixed": /usr/bin/curl -L 1.1.1.1
Now we only do that for /proc/self/exe, and leave
/proc/<pid>/fd/<number> as is, since Path is already the absolute
path to the binary.
Path: /usr/bin/spotify
Cmdline reported: /proc/self/exe
Cmdline "fixed": /usr/bin/spotify
Path: /usr/bin/curl
Cmdline reported: /proc/<pid>/fd/<number> -L 1.1.1.1
Cmdline final: /proc/<pid>/fd/<number> -L 1.1.1.1
- better process interception on aarch64 and armhf.
- on armhf getting proc arguments is ignored due to an error, but we
were not marking the event as INCOMPLETE_ARGS, so we were not reading
the compete cmdline.
Up until now we've been using https://github.com/iovisor/gobpf/ for
intercepting events from the kernel, and it has worked very well.
Unfortunately it's deprecated since 2021 (iovisor/gobpf#304), which means
that some of the known issues and limitations won't be fixed.
After evaluating some options, I've migrated the code from gobpf to
github.com/cilium/ebpf.
More info: #1222, #1312
- Moved Info message to Trace.
- When applying a default action on a connection (because the daemon is
not connected to the GUI), log the connection details. It eases the
task of creating rules without a GUI.
- Minimize the risk of race conditions when we're prompting the user to
allow/deny a connection, while we're still reading proc's environ file.
(this was actually a leak).
- Preallocate the Env map with the expected environ vars.
Sometimes we may receive a connection event after the exit of a
process:
[exec] /bin/xxx, pid 1234
[exit] /bin/xxx, pid 1234
[new conn] pid 1234 -> process unknown (on exec event and no /proc entry)
In these scenarios, we delay the deletion from cache a little
bit, to keep the PID available for a longer time.
When the Process Connector is used to intercept exec events, get and
build the process tree of a process.
PROCESS CONNECTOR feature was added here: 7a9bb17829
- Calculate the ram usage of a process in the daemon, using the page
size of the system.
- Added new functions to read some details of a process, so we can use
them in other parts of the code.
Allow to customize:
- EventsWorkers: number of goroutines to handle kernel events.
Default 8.
- QueueEventsSize: max number of events in the queue.
By default 0, meaning that it'll relay on the available goroutines to
process the events. If it's > 0, and the daemon can't process the
events fast enough, they'll be queued. Once the queue is full, it'll
behave as it was of size 0.
If there're lost events, a message will be logged: "Lost ebpf events..."
We build the parent process tree of a process when it's executed
for the first time.
Now we also build the tree when an already running process opens a new
outbound connection by the first time.
if an invalid opensnitch-procs.o module was loaded, we were flooding
the log with errors.
In these cases stop processing events after 20 errors (random, we should
have no errors).
This may occur if the module is malformed (valid .o ebpf module but
different structs, etc), or when loading modules from other versions.
Closes: #1099#1082
There's a long running task that monitors established connections every
~2s.
When a connection is not found via ebpf or proc, sometimes it's found
there so we can use the inode to search for the process.
However on some systems the netlink call to dump the sockets may fail
continuously, wasting resources. It'll also fail if you block connections
to port 0 (common case for ICMP packets).
So if there're too many errors dumpng the sockets, stop this task for
these cases.
- When discovering the hierarchy of a process, reuse components of
the tree if they're already on cache, to improve speed and reduce
mem allocs.
- When building the tree of a proces, rebuild the tree if the first
component doesn't have pid 1. Otherwise reuse the tree.
Simplify the cache of connections by storing only the PID of a process,
instead of the Process object.
We can obtain the Process object from the cache of processes by PID.
On this test we assumed that there would always be reading stats for our
own process /proc/self, but on restricted environments that might not
alwys be the case. Anyway, a value of 0 is not an error in itself.
Closes#1075
Now it's possible to configure eBPF modules path from the
default-config.json file:
"Ebpf": {
"ModulesPath": "..."
}
If the option is not provided, or if it's empty, we'll keep loading from
the default directories:
- /usr/local/lib/opensnitchd/ebpf
- /usr/lib/opensnitchd/ebpf
- /etc/opensnitchd/ebpf (deprecated, will be removed in the future).
Closes#928
- Fixed several leaks.
- Cache of events reorganized and improved.
* items are added faster.
* proc details are rebuilt if needed (checksums, proc tree, etc)
* proc's tree is reused if we've got the parent in cache.
rel: #413
Sometimes we receive /proc/self/exe as the path of the process (electron
apps).
Since a couple of systemd versions ago, some processes spawned by
systemd are reported as /proc/self/fd/<number>.
In these cases reading the symbolic link /proc/<pid>/exe points to the
file on disk.