We already had "${DEVICE}" which uses the interface name.
In times of predictable interface naming, that works well.
It allows the user to generate IDs per device which don't
change when the hardware is replaced.
"${MAC}" is similar, except that is uses the permanent MAC
address of the device. The substitution results in the empty
word, if the device has no permanent MAC address (like software
devices).
The per-device substitutions "${DEVICE}" and "${MAC}" are especially
interesting with "connection.multi-connect=multiple".
- use nm_auto to return early when something goes wrong
- don't modify NMDhcpClient's state until the end, when it looks
like we are (almost) started successfully.
- for IPv4, only attempt to load the lease if we actually are
interested in the address. Also, reduce the scope of the lease
variable, to the one place where we need it.
The client-id is something that we want to determine top-down.
Meaning, if the user specifies it via ipv4.dhcp-client-id, then it
should be used. If the user leaves it unspecified, we choose a
default stable client-id. For the internal DHCP plugin, this is
a node specific client-id based on
- the predictable interface name
- and /etc/machine-id
It's not clear, why we should allow specifying the client-id in
the lease file as a third source of configuration. It really pushes
the configuration first down (when we do DHCP without lease file),
to store an additional bit of configuration for future DHCP attempts.
If the machine-id or the interface-name changes, then so does the
default client-id. In this case, also "ipv4.dhcp-client-id=stable"
changes. It's fair to require that the user keeps the machine-id
stable, if the machine identity doesn't change.
Also, the lease files are stored in /var/lib/NetworkManager, which
is more volatile than /etc/machine-id. So, if we think that machine-id
and interface-name is not stable, why would we assume that we have
a suitable lease file?
Also, if you do:
nmcli connection add con-name "$PROFILE" ... ipv4.dhcp-client-id ''
nmcli connection up $PROFILE
nmcli connection modify "$PROFILE" ipv4.dhcp-client-id mac
nmcli connection up $PROFILE
nmcli connection modify "$PROFILE" ipv4.dhcp-client-id ''
nmcli connection up $PROFILE
wouldn't you expect that the original (default) client-id is used again?
Also, this works badly with global connection defaults in
NetworkManager.conf. If you configure a connection default, previously
already this would always force the client-id and overrule the lease.
That is reasonable, but in which case would you ever want to use
the client-id from the lease?
- if we leave the client-id of sd_dhcp_client unset, it will
anyway generate a node-specific client-id (and may fail if
"/etc/machine-id" is invalid).
Anticipate that, and don't let the client-id unset. In case
we have no client-id from configuration or lease, just generate
the id ourself (using the same algorithm). The advantage is,
that we know it upfront and can store the client-id in the
NMDhcpClient instance. We no longer need to peel it out from
the lease later.
- to generate the IPv4 client-id, we need a valid MAC address. Also,
sd_dhcp_client needs a MAC address for dhcp_network_bind_raw_socket()
as well. Just require that a MAC address is always needed. Likewise,
we need a valid ifindex and ifname set.
- likewise for IPv6 and IPv4, cleanup detecting the arptype and
checking MAC address length. sd_dhcp_client_set_mac() is overly
strict at asserting input arguments, so we must validate them anyway.
- also, now that we always initialize the client-id when starting
the DHCP client, there is no need to retroactively extract it
again when we receive the first lease.
A possible issue is that client_start() has about 136 arguments.
It doesn't get simpler by saving lines of code and writing them
all in the same line.
Wrap the lines.
While at it, use "FALSE" for "enforce_duid" argument, instead of "0".
It's a boolean.
We don't do that for ip4_start() either. The duid/client-id
is stored inside the NMDhcpClient instance, and the function can
access it from there.
Maybe, it is often preferable to have stateless objects and not
relying on ip4_start() to obtain the client ID from the client's
state. However, the purpose of the NMDhcpClient object is to
hold state about DHCP. To simplify the complexity of objects that
inherrently have state, we should be careful about mutating the state.
It adds little additional complexity of only reading the state when
needed anyway. In fact, it adds complexity, because previously
it wasn't enough to check all callers of nm_dhcp_client_get_client_id()
to see where the client-id is used. Instead, one would also need to
follow the @duid argument several layers of the call stack.
There should be lower layers that are concerned with writing
and reading dhclient configuration files. It's wrong to
have a nm_dhcp_dhclient_save_duid() function which requires
the caller to pre-escape the string to write. The caller shouldn't
be concerned with the file format, that's why the function
is used in the first place.
We only used "client_id" for IPv4 and "duid" for IPv6. Merge them.
Another advantage is, that we can share the logging functionality
of _set_client_id().
Drop unused function.
Aside from that, dhclient configuration files support a very complex
syntax. The parser was very naive and insufficient in parsing such
files. It's good we can just drop it.
Why would we do this? The configuration file we are reading back was
written by NetworkManager in the first place.
Maybe when assuming a connection after restart, this information could
be interesting. It however is not actually relevant.
Note how nm_dhcp_client_get_client_id() has only very few callers.
- nm_device_spawn_iface_helper() in 'nm-device.c'. In this case,
we either should use the client-id which we used when starting
DHCP, or none at all.
- ip4_start() in 'nm-dhcp-dhclient.c', but this is before starting
DHCP client and before it was re-read from configuration file.
- in "src/dhcp/nm-dhcp-systemd.c", but this has no effect for
the dhclient plugin.
Our internal DHCP client (from systemd) defaults to a particular client ID.
It is currently exposed as nm_sd_utils_generate_default_dhcp_client_id()
and is based on the systemd implementation.
One problem with that is, that it internally looks up the interface name
with if_indextoname() and reads /etc/machine-id. Both makes it harder
for testing.
Another problem is, that this way of generating the client-id is
currently limited to internal client. Why? If you use dhclient plugin,
you may still want to use the same algorithm. Also, there is no explict
"ipv4.dhcp-client-id" mode to select this client-id (so that it could
be used in combination with "dhclient" plugin).
As such, this code will be useful also aside systemd DHCP plugin.
Hence, the function should not be obviously tied to systemd code.
The implementation is simple enough, and since we already have a
unit-test, refactor the code to our own implementation.
Internal DHCP client generates a default client ID. For one,
we should ensure that this algorithm does not change without
us noticing, for example, when upgrading systemd code. Add
a test, that the generation algorithm works as we expect.
Also note, that the generation algorithm uses siphash24().
That means, siphash24() implementation also must not change
in the future, to ensure the client ID doesn't change. As we
patch systemd sources to use shared/c-siphash, this is not
obviously the case. Luckily c-siphash and systemd's siphash24 do
agree, so all is good. The test is here to ensure that.
Also, previously the generation algorithm is not exposed as a
function, sd_dhcp_client will just generate a client-id when
it needs it. However, later we want to know (and set) the client
id before starting DHCP and not leave it unspecified to an
implementation detail.
This patch only adds a unit-test for the existing DHCP client
ID generation to have something for comparison. In the next
commit this will change further.
- use NMUuid type where appropriate.
- no error handling for generate_duid_from_machine_id().
It cannot fail anymore.
- add thread-safety to generate_duid_from_machine_id() with
double-checked locking.
- use unions for converting the sha256 digest to the target
type.
For testing purpose, it's bad to let nm_utils_stable_id_parse()
directly access nm_utils_get_boot_id_str(). Instead, the function
should have no side-effects.
Since the boot-id is anyway cached, accessing it is cheap. Even
if it likely won't be needed.
Tests might access the secret-key.
For CI builds we may very well build NM as root and also run
unit tests. In such a situation it's bad to persist the secret
key. For example, the SELinux label may be wrong, and subsequently
starting NetworkManager may cause errors. Avoid persisting the secret
key for tests.
Previously, whenever we needed /etc/machine-id we would re-load it
from file. The are 3 downsides of that:
- the smallest downside is the runtime overhead of repeatedly
reading the file and parse it.
- as we read it multiple times, it may change anytime. Most
code in NetworkManager does not expect or handle a change of
the machine-id.
Generally, the admin should make sure that the machine-id is properly
initialized before NetworkManager starts, and not change it. As such,
a change of the machine-id should never happen in practice.
But if it would change, we would get odd behaviors. Note for example
how generate_duid_from_machine_id() already cached the generated DUID
and only read it once.
It's better to pick the machine-id once, and rely to use the same
one for the remainder of the program.
If the admin wants to change the machine-id, NetworkManager must be
restarted as well (in case the admin cares).
Also, as we now only load it once, it makes sense to log an error
(once) when we fail to read the machine-id.
- previously, loading the machine-id could fail each time. And we
have to somehow handle that error. It seems, the best thing what we
anyway can do, is to log an error once and continue with a fake
machine-id. Here we add a fake machine-id based on the secret-key
or the boot-id. Now obtaining a machine-id can no longer fail
and error handling is no longer necessary.
Also, ensure that a machine-id of all zeros is not valid.
Technically, a machine-id is not an RFC 4122 UUID. But it's
the same size, so we also use NMUuid data structure for it.
While at it, also refactor caching of the boot-id and the secret
key. In particular, fix the thread-safety of the double-checked
locking implementations.
- add a commnt about thread-safety.
- minor refactoring initializing the value in nm_utils_get_testing().
Instead of returning the flags we just set, go back to the begin
and re-read the value (which must be initialized by now). No big
difference, but feels a bit nicer to me.
The GChecksum API is cumbersome to use.
For example, g_checksum_get_digest() requires a length input/output
argument. At the same time, GChecksum does not allow you to query its
checksum-type nor the desired digest-length. When you have a GChecksum
at hand, you must always know the digest-length you are going to use.
So, the length parameter is only good for asserting.
Add a macro to make that more convenient.
Benefits: it's less lines of code, and we always do all the asserts
that are due.
- fix thread-safety by adding a memory barrier (g_atomic_pointer_get())
to the double-checked locking pattern when initializing the hash key.
- generate the random data outside the lock. Calling nm_utils_random_bytes()
within the lock is ugly, because we don't want to assume that the function
has no side effects which are prone to dead-lock. There is no problem attempting
to generate the random data without lock, and only use it when the race is won.
Manually disconnecting a profile of course blocks autoconnect of the
same profile. Otherwise, the profile would likely re-activate right
away, which is clearly against the users intention. If the users just
want to re-activate the profile, they should issue `nmcli connection up`
instead, with does a full down and up cycle.
This is more interesting for profiles that have 'connection.multi-connect'
set to 'multiple'. Would you expect that manually deactivating such a
profile blocks autoconnect of the profile on all devices? Maybe
yes, maybe not. Currently that is indeed the case and autoconnect gets
blocked regardless of multi-connect.
In the past, the headers "linux/if.h" and "net/if.h" were incompatible.
That means, we can either include one or the other, but not both.
This is fixed in the meantime, however the issue still exists when
building against older kernel/glibc.
That means, including one of these headers from a header file
is problematic. In particular if it's a header like "nm-platform.h",
which itself is dragged in by many other headers.
Avoid that by not including these headers from "platform.h", but instead
from the source files where needed (or possibly from less popular header
files).
Currently there is no problem. However, this allows an unknowing user to
include <net/if.h> at the same time with "nm-platform.h", which is easy
to get wrong.
Previously we would compile source files from shared/nm-utils
multiple times. That not only slows down compilation, but it makes it
confusing which project require exactly what.
Most of the files in shared/nm-utils are a mixed bag of utility
functions. Just build one libnm-utils-base library. Since the linker
will throw away unused parts, there is no problem that not every user
of libnm-utils-base needs everything.
Also add libnm-utils-udev, which cannot be part of libnm-utils-base as
it has an additional dependency on libudev.
For one, it's not unreasonable that we want to run the same
tests both for gitlab and travis.
Move the actual tests into a script, which is called by both
CI environments.
We still can do something different, based on the environment.
The advantage here is, that the common part will be shared, and
the places where we differ can easily be spot.
https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/44
The ebpf syscall doesn't work on RHEL even if the linux/bpf.h header
is available: let's explicitly disable it.
On Fedora explicitly enable eBPF instead of autodetecting it.
This allows us to somehow work with no PCI (or USB for that matter), such
as some virtual machines. Old ModemManager doesn't like this though, but
there doesn't seem to be a way around it.
https://github.com/NetworkManager/NetworkManager/pull/246
