The eapol state machine parameters are now built inside device.c when
the network connection is attempted. The reason is that the device
object knows about network settings, wiphy constraints and should
contain the main 'management' logic.
netdev now manages the actual low-level process of building association
messages, detecting authentication events, etc.
Keep an updated sorted list of networks in addition to the "networks"
hashmap. The list can be queried through the GetOrderedNetworks dbus
method.
We also take advantage of that list to get rid of a single
l_hashmap_foreach in new_scan_results.
Since the rankmod value only makes sense for autoconnectable networks,
change network_rankmod to return an indication of whether the rankmod is
valid as a boolean instead of as a double, as discussed before.
Do nothing in device_disassociated if device->connected_network
indicates we are not associated. This may happen if the device was
connected since before iwd was started, this should possibly be fixed
separately by querying device state when device is detected.
for network_seen and network_connected
Only accept a struct network pointer instead of separately the ssid and
security type. This is needed so we can do some more simplification in
the next patch by having access to the network struct.
Turn netdev watches into device watches. The intent is to refactor out
netdev specific details into its own class and move device specific
logic into device.c away from wiphy.c
The lost beacon event can be received when iwd thinks netdev is
diconnected if it was connected before iwd started, and then
netdev_disassociated will segfault.
Instead of passing the user_data parameter in every __eapol_rx_packet
call to be used by EAPOL in all tx_packet calls, add
eapol_sm_set_tx_user_data function that sets the value of user_data for
all subsequent tx_packet calls. This way tx_packet can be called from
places that are not necessarily inside an __eapol_rx_packet call.
==2469== 24 bytes in 1 blocks are still reachable in loss record 1 of 1
==2469== at 0x4C2B970: malloc (vg_replace_malloc.c:296)
==2469== by 0x40E6DD: l_malloc (util.c:62)
==2469== by 0x40F1CD: l_queue_new (queue.c:63)
==2469== by 0x40D534: scan_init (scan.c:796)
==2469== by 0x403AC3: nl80211_appeared (wiphy.c:2121)
==2469== by 0x415FF3: get_family_callback (genl.c:987)
==2469== by 0x415A4F: process_request (genl.c:381)
==2469== by 0x415A4F: received_data (genl.c:492)
==2469== by 0x413184: io_callback (io.c:120)
==2469== by 0x4127C2: l_main_run (main.c:346)
==2469== by 0x40253E: main (main.c:171)
Refactoring the entire scan code, and this part seems to not be
supported by the target kernels. Revisit / redo this functionality once
things become a bit clearer.
In the D-bus .Connect call return an error immediately if we
find that there's no common cipher supported between iwd, the
network adapter and the AP. This is to avoid asking the agent
for the passkey if we know the connection will fail.
An alternative would be to only show networks that we can connect
to in the scan results on D-bus but I suspect that would cause
more pain to users debugging their wifi setups on average.
For now, if a passphrase is needed we check once before querying
for passphrase and recheck afterwards when we're about to
associate.
Instead of passing in the RSN/WPA elements, simply pass in the
configured cipher. This will make the implementation of the install_gtk
callback much simpler.
When handling repeated 4-Way Handshakes from the AP there will be no
.Connect() call pending so we need to check that netdev->connect_pending
is non-NULL. It may be a good idea to check this even during initial
handshake.
When disconnect is triggered locally, we do not clean up properly.
==4336== at 0x4C2B970: malloc (vg_replace_malloc.c:296)
==4336== by 0x40CEED: l_malloc (util.c:62)
==4336== by 0x40F46A: l_settings_new (settings.c:82)
==4336== by 0x40CE2E: storage_network_open (storage.c:180)
==4336== by 0x40498E: network_connect_psk (wiphy.c:307)
==4336== by 0x40498E: network_connect (wiphy.c:359)
==4336== by 0x41D7EE: _dbus_object_tree_dispatch (dbus-service.c:845)
==4336== by 0x416A16: message_read_handler (dbus.c:297)
==4336== by 0x411984: io_callback (io.c:120)
==4336== by 0x410FC2: l_main_run (main.c:346)
==4336== by 0x40253E: main (main.c:171)
This happens when connecting / disconnecting successfully multiple
times.
==4336== 64 bytes in 2 blocks are definitely lost in loss record 9 of 11
==4336== at 0x4C2B970: malloc (vg_replace_malloc.c:296)
==4336== by 0x40CEED: l_malloc (util.c:62)
==4336== by 0x40D6D9: l_util_from_hexstring (util.c:493)
==4336== by 0x4049C6: network_connect_psk (wiphy.c:315)
==4336== by 0x4049C6: network_connect (wiphy.c:359)
==4336== by 0x41D7EE: _dbus_object_tree_dispatch (dbus-service.c:845)
==4336== by 0x416A16: message_read_handler (dbus.c:297)
==4336== by 0x411984: io_callback (io.c:120)
==4336== by 0x410FC2: l_main_run (main.c:346)
==4336== by 0x40253E: main (main.c:171)
Instead of storing the network pointer for each BSS, store it on the
netdev object. This saves space inside struct bss and makes longer term
refactoring simpler.
==4249== 231 (32 direct, 199 indirect) bytes in 1 blocks are definitely
lost in loss record 10 of 10
==4249== at 0x4C2B970: malloc (vg_replace_malloc.c:296)
==4249== by 0x40CF5D: l_malloc (util.c:62)
==4249== by 0x40F4DA: l_settings_new (settings.c:82)
==4249== by 0x40CE9E: storage_network_open (storage.c:180)
==4249== by 0x40499E: network_connect_psk (wiphy.c:307)
==4249== by 0x40499E: network_connect (wiphy.c:359)
==4249== by 0x41D85E: _dbus_object_tree_dispatch (dbus-service.c:845)
==4249== by 0x416A86: message_read_handler (dbus.c:297)
==4249== by 0x4119F4: io_callback (io.c:120)
==4249== by 0x411032: l_main_run (main.c:346)
==4249== by 0x40253E: main (main.c:171)
This patch saves off the PSK generated based on the passphrase provided
by the agent/user. The PSK is saved only if the connection is
successful.
Subsequent connection attempts to the known AP use the PSK saved on the
filesystem (default /var/lib/iwd/<ssid>.psk). If the connection fails,
the agent will again be asked for the passphrase on the next attempt.
CMD_DEAUTHENTICATE seems to carry only the management frame pdu
information. CMD_DISCONNECT is carrying the information that is
actually needed by us:
> Event: Disconnect (0x30) len 28 1140.118545
Wiphy: 0 (0x00000000)
Interface Index: 3 (0x00000003)
Reason Code: 2 (0x0002)
Disconnect by AP: true
We will ignore non-UTF8 based SSIDs. Support for non-UTF8 SSIDs seems
to be of dubious value in the real world as the vast majority of
consumer devices would not even allow such SSIDs to be configured or
used.
There also seems to be no compelling argument to support such SSIDs, so
until that argument arrives, non-UTF8 SSIDs will be filtered out. This
makes the D-Bus API and implementation much easier.
We need to set the linkmode and operstate after successful
authentication.
Initial value for linkmode is 1 (user space controlled) and
IF_OPER_DORMANT for opermode. After successful authentication,
the operstate is set to IF_OPER_UP.
More specific details can be seen in kernel sources at
https://www.kernel.org/doc/Documentation/networking/operstates.txt
CC src/wiphy.o
src/wiphy.c: In function ‘eapol_read’:
src/wiphy.c:172:24: error: argument to ‘sizeof’ in ‘memset’ call is the same expression as the destination; did you mean to remove the addressof? [-Werror=sizeof-pointer-memaccess]
memset(&sll, 0, sizeof(&sll));
^
We can give reply to connect DBus call in associating event only
when we are connecting to Open network. For PSK AP, the reply can
only be sent after we have finished 4-way handshaking.
The success or not of a scan command is found from the message directly.
There's no need to look for any attribute from the scan netlink answer.
The message is an error message or not, and that tells if the scan has
been started or not.
Modifying a bit how networks are stored inside the hashtable:
1 - instead of the network id, the network's object path is used
2 - network holds the pointer of the object path
3 - the hashtable does not free the key (network_free() will)
This permits to optimize on:
1 - one memory allocation used for 2 distinct things
2 - remove the need to re-compute the object path (and the id) when it's
needed, it can use dircetly the one stored in the network structure.
Request a passphrase via Agent if none is set at the time network is
being connected. When freeing a network, cancel any outstanding Agent
requests and free allocated memory.
l_genl_family_send only returns request id. If request
failed at low level, current implementation does not handle that.
In case of request failure clear pending dbus messages.
Open networks do not contain a RSN element, so storing a 256 byte buffer
was too expensive.
This patch also has the side-effect of fixing detection of Open
Networks. Prior to this, if the scan results did not contain an RSN IE,
the 'rsne' variable would be set to all zeros. scan_get_ssid_security
would then be called, but instead of a NULL struct ie_rsn_info, a
non-null, but zerod out ie_rsn_info would be passed in. This caused the
code to work, but for the wrong reasons.
Instead of mallocing the ssid buffer, use a static array. This removes an
extra couple of malloc/free operations and should result in less memory
utilization on average.
Authenticate event on wiphy mlme notification does not provide
enough information on which network/bss authentication command
was sent. BSS and network information is required to send associate
command to AP. So cache bss pointer in netdev struct and retrieve
on wiphy mlme notifications.
Instead of storing multiple copies of the same BSS (one hanging off the
netdev object and one hanging off the network object), we instead store
the BSS list only on the netdev object.
The network object gets a pointer to the BSS structure on the netdev
list. As a side effect, the BSS list is always sorted properly.
Instead of always mallocing space for the ssid array, and then freeing
it in most circumstances, do the opposite. Only allocate the array once
it is actually needed. This has the side effect of removing an unneeded
variable and making the code simpler.
The idea here is that network object will contain a list of BSS
that have the same SSID and security setting. From user point of view,
we will connect to a certain SSID instead of connecting directly to
a BSS. We pick the best BSS automatically from the BSS list when
connecting to a SSID.
MLME notify function prints error if wiphy or netdev is missing.
The error text in this case talks about scan notification instead
of more proper MLME notification.
Make sure to print some errors if attributes cannot be appended
to a message. It is dangerous to ignore the return code from
l_genl_msg_append_attr() because the kernel might act weirdly
if some attribute is missing.
Add rudimentary support for mac80211 scheduled scan feature.
This is done so that kernel support for task called "Bind
NL80211_CMD_START_SCHED_SCAN to netlink socket" from TODO
file can be tested. The current scan interval is set to 60
seconds which is probably too fast for the final version.
Because none of the attributes are assigned until after the DeviceAdded
signal is emitted, the signal appears with invalid properties. For now,
move the netdev structure fill-out into the if statement.
If the netdev attributes can change, then these need to be handled
separately and appropriate signals to be sent.
