Android platform assigns the same socket id if the socket identifier in
conf->ctrl_interface and global->params.ctrl_interface (parameter for
android_get_control_socket) point to the same Android specific control
socket. This ends up having two eloop socket handlers registered for the
same file descriptor and thus, two attempt to receive and process each
command. This can result in unexpected failure, e.g., the prefix IFNAME=
for any command is valid for global socket handler, but results in
UNKNOWN COMMAND response from the per-interface ctrl socket handler).
Since it might be possible to end up with this type of invalid
configuration in OTA upgrade, compare the socket identifiers and do not
open the ctrl socket on the respective interface if both point to same.
This allows the Wi-Fi framework to use the global control interface.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
These prints were at DEBUG level (-d), but they can be very frequent, so
drop them to MSGDUMP (-dd). This allows the prints to be suppressed in
common debugging cases while still leaving them easily enablable to
debug control interface issues without having to enable excessive
debugging.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
This makes it easier to track the output queue state on a control
interface socket and determine what could be causing issues with running
out of space in the buffer.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
This commit implements hostapd global control interface notifications
infrastructure. hostapd global control interface clients issue
ATTACH/DETACH commands to register and deregister with hostapd
correspondingly - the same way as for any other hostapd/wpa_supplicant
control interface.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The control interface commands may include passwords or other private
key material, so clear it explicitly from memory as soon as the
temporary buffer is not needed anymore.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Instead of an int variable with magic values 0, 1, 2, use an enum that
gives clearer meaning to the values now that the original boolean type
global argument is not really a boolean anymore.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
It is easier to debug issues related to the wpa_supplicant control
interfaces being left behind in attached state when the debug log file
can be used to determine whether a specific monitor socket was a global
or per-interface one.
Signed-off-by: Jouni Malinen <j@w1.fi>
On Android, the control socket being used may be the socket that is
created when wpa_supplicant is started as a /init.*.rc service. Such a
socket is maintained as a key-value pair in Android's environment.
Closing this control socket would leave wpa_supplicant in a bad state.
When wpa_supplicant re-opens the ctrl_iface socket, it will query the
Android's environment, and will be returned with the same socket
descriptor that has already been closed.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
If this chmod() call fails, the global control interface is allowed to
be used since there was no change to its group. Anyway, it can be
helpful to note the error case in debug log instead of silently ignoring
it.
Signed-off-by: Jouni Malinen <j@w1.fi>
The group name is not used on these paths, so just remove it from the
directory name without updating gid_str to point to the unused group
name.
Signed-off-by: Jouni Malinen <j@w1.fi>
Since the global ctrl_iface can be used with IFNAME= prefix to send
commands to be processed by per-interface code, it should have the same
(well, close to same since the prefix takes some space) limits on
command length as the per-interface ctrl_iface. Increase the buffer from
256 to 4096 to achieve this.
Signed-hostap: Jouni Malinen <jouni@qca.qualcomm.com>
This can be used to override previously set ctrl_interface value in a
way that clears the variable to NULL instead of empty string. The only
real use case for this is to disable per-interface ctrl_interface from
the additional control file (-I<file>) in case ctrl_interface was set in
the main configuration file. It should be noted that zero-length
ctrl_interface parameter can be used to initiate some control interface
backends, so simpler designs were not available for this.
The format of the new parameter is not exactly cleanest due to
configuration file parsing assumptions. For example:
ctrl_interface=....
no_ctrl_interface=
would end up with ctrl_interface=NULL.
Signed-hostap: Jouni Malinen <jouni@qca.qualcomm.com>
Previously, wpa_msg_register_cb() was called only from successful
completion of per-interface control interface initialization. This would
leave the callback unregistered in case only the global control
interface is used which would result in not delivering control interface
events on the global interface. Fix this by registering the callback
handler also from successful initialization of the global control
interface.
Signed-hostap: Jouni Malinen <jouni@qca.qualcomm.com>
If wpa_supplicant ctrl_iface clients are misbehaving and refusing to
read replies or event messages from wpa_supplicant, the single socket
used in wpa_supplicant to send messages can reach the maximum send
buffer limit. When that happens, no more responses to any client can be
sent. Work around this by closed and reopening the socket in case such a
failure state is detected. This is obviously not desirable since it
breaks existing connected sockets, but is needed to avoid leaving
wpa_supplicant completely unable to respond to any client. Cleaner fix
for this may require more considerable changes in the ctrl_iface design
to move to connection oriented design to allow each client to be handled
separately and unreachability to be detected more reliably.
Signed-hostap: Jouni Malinen <jouni@qca.qualcomm.com>
Prepare reply buffer first for all cases and then use a single sendto()
call instead of three calls depending on reply type. This allows simpler
error handling for control interface send operations.
Signed-hostap: Jouni Malinen <jouni@qca.qualcomm.com>
This replaces number of perror() calls with wpa_printf() to get the
error messages embedded within rest of the debug messages in the same
stream instead of pushing these to stderr which may get directed to
another location.
Signed-hostap: Jouni Malinen <j@w1.fi>
Reduce the wait time for the monitor control interfaces to get
messages on wpa_supplicant de-init etc., as this significantly delays
the shutdown of the wpa_supplicant.
Signed-hostap: Ilan Peer <ilan.peer@intel.com>
In case a control interface socket is detached because of sendmsg()
failing for the socket, function call to detach the socket uses a
pointer to the socket information in the structure to be freed. Reorder
code to print socket info before freeing the data to avoid use of freed
memory in case debug prints are enabled.
Signed-hostap: Jouni Malinen <j@w1.fi>
This keeps wpa_supplicant from hanging forever if the other end of the
socket dies. This is similar to the earlier commit
4fdc8def88 to make the global control
interface befave in the same way.
Signed-hostap: Jouni Malinen <j@w1.fi>
It looks like some of the global control interface cases ended up
blocking in sendmsg() when trying to send an event. Since this can block
all wpa_supplicant processing for multiple seconds, this is very
undesirable. Avoid this by requesting sendmsg() to return an error
rather than waiting for the message to be sent.
Signed-hostap: Jouni Malinen <jouni@qca.qualcomm.com>
These events are sent as a special case to both the group interface and
"parent interface" (i.e., the interface that was used for managing P2P
negotiation). The latter is not really correct event, so get rid of it
with the new global control interface design where there is no need to
support legacy upper layer implementations.
Signed-hostap: Jouni Malinen <j@w1.fi>
This function can be used instead of wpa_msg() and wpa_msg_ctrl() to
indicate that an event is not specific to a network interface.
Signed-hostap: Jouni Malinen <j@w1.fi>
The ATTACH/DETACH mechanism to request event messages from
wpa_supplicant can now be used through the global control interface,
too. This results in events from all interfaces being delivered through
a single monitor socket. "IFNAME=<ifname> " prefix is used on events
that are specific to an interface.
Signed-hostap: Jouni Malinen <j@w1.fi>
The wpa_supplicant global control interface parameter can now be used to
explicitly specify an abstract UNIX domain socket (Linux specific
extension) with "@abstract:" prefix and an Android control socket with
"@android:" prefix.
Signed-hostap: Jouni Malinen <j@w1.fi>
The optional -G<group> command line argument can be used to specify the
group that can access the global control interface.
Signed-hostap: Jouni Malinen <j@w1.fi>
wpa_supplicant is started from /init.*.rc on Android and that seems
to be using umask 0077 which would leave the control interface
directory without group access. This breaks things since Wi-Fi
framework assumes that this directory can be accessed by other
applications in the wifi group. Fix this by adding group access even
if umask value would prevent this.
In most cases, this issue was not hit since the control interface
directory is normally created by that same init.*.rc file with suitable
mode and wpa_supplicant is killed in the way that does not allow it to
remove the file. However, if wpa_supplicant is allowed stop cleanly, it
will remove the directory and the next start could result with the Wi-Fi
framework not being able to use Wi-Fi (and GUI not showing Wi-Fi getting
enabled).
Signed-hostap: Jouni Malinen <jouni@qca.qualcomm.com>
These bursts can result in control interface monitors being detached
even if the external program is still working properly. Use much larger
error threshold for ENOBUFS to avoid this.
This patch allows wpa_supplicant to compile on Debian's kfreebsd
architectures.
Patch by Stefan Lippers-Hollmann based on work done by Petr Salinger
and Emmanuel Bouthenot for 0.6.X (http://bugs.debian.org/480572).
When using umask 0077, the control interface directory was left without
group read/execute permissions even if the configuration file explicitly
asked for the group to be allowed to access the control interface. Fix
this by adding read/execute permissions for group if a specific group is
defined in the configuration. [Bug 199]
Removing just sun_family is not portable since some systems (e.g.,
FreeBSD) include an additional sun_len field. Using offsetof should be
portable. In addition, set sun_ken for FreeBSD.
