In the methods' check_settings do a more complete early check for
possible certificate / private key misconfiguration, including check
that the certificate and the private key are always present or absent
together and that they actually match each other. Do this by encrypting
and decrypting a small buffer because we have no better API for that.
A method's .check_settings method checks for inconsistent setting files
and prints readable errors so there's no need to do that again in
.load_settings, although at some point after removing the duplicate
error messages from the load_settings methods we agreed to keep minimum
checks that could cause a crash e.g. in a corner case like when the
setting file got modified between the check_settings and the
load_settings call. Some error messages have been re-added to
load_settings after that (e.g. in
bb4e1ebd4f) but they're incomplete and not
useful so remove them.
When this test was written only group 19 was supported. The 'bad_group'
test used, at the time, unsupported group 20. Now group 20 is supported
so this test was expecting a failure. This updates the test to use group
0xff, which is not a valid ECC group and should always fail.
Previously, the storage dir has only been created after a successful
network connection, causing removal of Known Network interface from
Dbus and failure to register dir watcher until daemon is restarted.
A length check was still assuming the 256 bit ECC group. This
was updated to scale with the group. The commit buffer was also
not properly sized. This was changed to allow for the largest
ECC group supported.
SAE was hardcoded to work only with group 19. This change fixes up the
hard coded lengths to allow it to work with group 20 since ELL supports
it. There was also good amount of logic added to support negotiating
groups. Before, since we only supported group 19, we would just reject
the connection to an AP unless it only supported group 19.
This did lead to a discovery of a potential bug in hostapd, which was
worked around in SAE in order to properly support group negotiation.
If an AP receives a commit request with a group it does not support it
should reject the authentication with code 77. According to the spec
it should also include the group number which it is rejecting. This is
not the case with hostapd. To fix this we needed to special case a
length check where we would otherwise fail the connection.
SAE has a clogging test which requires 4 radios to all simultaneously
connect. All the other tests are only using one of these radios, so
in these tests we explicitly disconnect these devices preventing them
from autoconnecting.
Since the EAP-PWD fragmentation test uses group 19 there is test
coverage there for that group. This changes connection_test to use
group 20 instead of 19.
Most of this work was already done after moving ECC into ELL, but
there were still a few places where the 256-bit group was assumed.
This allows the 384-bit group to be used, and theoretically any
other group added to ELL in the future.
If we have a BSS list where all BSS's have been blacklisted we still
need a way to force a connection to that network, instead of having
to wait for the blacklist entry to expire. network_bss_select now
takes a boolean 'fallback_to_blacklist' which causes the selection
to still return a connectable BSS even if the entire list was
blacklisted.
In most cases this is set to true, as these cases are initiated by
DBus calls. The only case where this is not true is inside
station_try_next_bss, where we do want to honor the blacklist.
This both prevents an explicit connect call (where all BSS's are
blacklisted) from trying all the blacklisted BSS's, as well as the
autoconnect case where we simply should not try to connect if all
the BSS's are blacklisted.
There are is some implied behavior here that may not be obvious:
On an explicit DBus connect call IWD will attempt to connect to
any non-blacklisted BSS found under the network. If unsuccessful,
the current BSS will be blacklisted and IWD will try the next
in the list. This will repeat until all BSS's are blacklisted,
and in this case the connect call will fail.
If a connect is tried again when all BSS's are blacklisted IWD
will attempt to connect to the first connectable blacklisted
BSS, and if this fails the connect call will fail. No more
connection attempts will happen until the next DBus call.
If IWD fails to connect to a BSS we can attempt to connect to a different
BSS under the same network and blacklist the first BSS. In the case of an
incorrect PSK (MMPDU code 2 or 23) we will still fail the connection.
station_connect_cb was refactored to better handle the dbus case. Now the
netdev result switch statement is handled before deciding whether to send
a dbus reply. This allows for both cases where we are trying to connect
to the next BSS in autoconnect, as well as in the dbus case.
This makes __station_connect_network even less intelligent by JUST
making it connect to a network, without any state changes. This makes
the rekey logic much cleaner.
We were also changing dbus properties when setting the state to
CONNECTING, so those dbus property change calls were moved into
station_enter_state.
A new driver extended feature bit was added signifying if the driver
supports PTK replacement/rekeying. During a connect, netdev checks
for the driver feature and sets the handshakes 'no_rekey' flag
accordingly.
At some point the AP will decide to rekey which is handled inside
eapol. If no_rekey is unset we rekey as normal and the connection
remains open. If we have set no_rekey eapol will emit
HANDSHAKE_EVENT_REKEY_FAILED, which is now caught inside station. If
this happens our only choice is to fully disconnect and reconnect.
If we receive handshake message 1/4 after we are already connected
the AP is attempting to rekey. This may not be allowed and if not
we do not process the rekey and emit HANDSHAKE_EVENT_REKEY_FAILED
so any listeners can handle accordingly.
The AP structure was getting cleaned up twice. When the DBus stop method came
in we do AP_STOP on nl80211. In this callback the AP was getting freed in
ap_reset. Also when the DBus interface was cleaned up it triggered ap_reset.
Since ap->started gets set to false in ap_reset, we now check this and bail
out if the AP is already stopped.
Fixes:
++++++++ backtrace ++++++++
0 0x7f099c11ef20 in /lib/x86_64-linux-gnu/libc.so.6
1 0x43fed0 in l_queue_foreach() at ell/queue.c:441 (discriminator 3)
2 0x423a6c in ap_reset() at src/ap.c:140
3 0x423b69 in ap_free() at src/ap.c:162
4 0x44ee86 in interface_instance_free() at ell/dbus-service.c:513
5 0x451730 in _dbus_object_tree_remove_interface() at ell/dbus-service.c:1650
6 0x405c07 in netdev_newlink_notify() at src/netdev.c:4449 (discriminator 9)
7 0x440775 in l_hashmap_foreach() at ell/hashmap.c:534
8 0x4455d3 in process_broadcast() at ell/netlink.c:158
9 0x4439b3 in io_callback() at ell/io.c:126
10 0x442c4e in l_main_iterate() at ell/main.c:473
11 0x442d1c in l_main_run() at ell/main.c:516
12 0x442f2b in l_main_run_with_signal() at ell/main.c:644
13 0x403ab3 in main() at src/main.c:504
14 0x7f099c101b97 in /lib/x86_64-linux-gnu/libc.so.6
+++++++++++++++++++++++++++
This will allow for blacklisting a BSS if the connection fails. The
actual blacklist module is simple and must be driven by station. All
it does is add BSS addresses, a timestamp, and a timeout to a queue.
Entries can also be removed, or checked if they exist. The blacklist
timeout is configuratble in main.conf, as well as the blacklist
timeout multiplier and maximum timeout. The multiplier is used after
a blacklisted BSS timeout expires but we still fail to connect on the
next connection attempt. We multiply the current timeout by the
multiplier so the BSS remains in the blacklist for a larger growing
amount of time until it reaches the maximum (24 hours by default).
Soon BSS blacklisting will be added, and in order to properly decide if
a BSS should be blacklisted we need the status code on a failed
connection. This change stores the status code when there is a failure
in netdev and adds a getter to retrieve later. In many cases we have
the actual status code from the AP, but in some corner cases its not
obtainable (e.g. an error sending an NL80211 command) in which case we
just default to MMPDU_REASON_CODE_UNSPECIFIED.
Rather than continue with the pattern of setting netdev->result and
now netdev->last_status_code, the netdev_connect_failed function was
redefined so its no longer used as both a NL80211 callback and called
directly. Instead a new function was added, netdev_disconnect_cb which
just calls netdev_connect_failed. netdev_disconnect_cb should not be
used for all the NL80211 disconnect commands. Now netdev_connect_failed
takes both a result and status code which it sets in the netdev object.
In the case where we were using netdev_connect_failed as a callback we
still need to set the result and last_status_code but at least this is
better than having to set those in all cases.
Remove an unneeded buffer and its memcpy, remove the now unneeded use of
l_checksum_digest_length and use l_checksum_reset instead of creating a
new l_checksum for each chunk.
When using --valgrind, you must also use --verbose iwd, and, depending
on the tests you may also need to include pytests in the verbose flag.
Since anyone using --valgrind definitely wants to see valgrind info
printed they should not need to enable verbose printing. Also, manually
parsing valgrind prints with IWD prints mixed throughout is a nightmare
even for a single test.
This patch uses valgrind's --log-file flag, which is directed to
/tmp/valgrind.log. After the tests runs we can print out this file.
ELL ECC supports group 20 (P384) so OWE can also support it. This also
adds group negotiation, where OWE can choose a different group than the
default if the AP requests it.
A check needed to be added in netdev in order for the negotiation to work.
The RFC says that if a group is not supported association should be rejected
with code 77 (unsupported finite cyclic group) and association should be
started again. This rejection was causing a connect event to be emitted by
the kernel (in addition to an associate event) which would result in netdev
terminating the connection, which we didn't want. Since OWE receives the
rejected associate event it can intelligently decide whether it really wants
to terminate (out of supported groups) or try the next available group.
This also utilizes the new MIC/KEK/KCK length changes, since OWE dictates
the lengths of those keys.
Rather than hard coding to SHA256, we can pass in l_checksum_type
and use that SHA. This will allow for OWE/SAE/PWD to support more
curves that use different SHA algorithms for hashing.
OWE defines KEK/KCK lengths depending on group. This change adds a
case into handshake_get_key_sizes. With OWE we can determine the
key lengths based on the PMK length in the handshake.
In preparation for OWE supporting multiple groups eapol needed some
additional cases to handle the OWE AKM since OWE dictates the KEK,
KCK and MIC key lengths (depending on group).
Right now the PMK is hard coded to 32 bytes, which works for the vast
majority of cases. The only outlier is OWE which can generate a PMK
of 32, 48 or 64 bytes depending on the ECC group used. The PMK length
is already stored in the handshake, so now we can just pass that to
crypto_derive_pairwise_ptk
The crypto_ptk was hard coded for 16 byte KCK/KEK. Depending on the
AKM these can be up to 32 bytes. This changes completely removes the
crypto_ptk struct and adds getters to the handshake object for the
kck and kek. Like before the PTK is derived into a continuous buffer,
and the kck/kek getters take care of returning the proper key offset
depending on AKM.
To allow for larger than 16 byte keys aes_unwrap needed to be
modified to take the kek length.
Unfortunately there is no way to determine the MIC length just from the
eapol frame. 802.11 defined AKMs define the MIC length, but non 802.11
AKMs (e.g. OWE) can define their own MIC length. For this reason it seem
infeasable to track these special AKM's data flow to determine the MIC
length.
To work around this we can just try different MIC lengths (since there
are only 3 after all). This allows us to get key data length and see if
the total packet size equals the frame length + key data length. If the
sizes don't match we can try the next MIC length.
