We want to ensure that all errors reported via DBus operations are
normalized to MM-specific errors.
We don't want to return QMI or MBIM specific errors, as those are
protocol specific and we don't want DBus clients to need to rely on
knowing which is the protocol in use by the device.
[1/2] Compiling C object src/ModemManager.p/mm-base-manager.c.o
../src/mm-base-manager.c: In function ‘remove_device_inhibition’:
../src/mm-base-manager.c:1127:20: warning: declaration of ‘l’ shadows a previous local [-Wshadow]
1127 | GList *l;
| ^
../src/mm-base-manager.c:1116:16: note: shadowed declaration is here
1116 | GList *l;
| ^
If the device is not fully removed from the system during inhibition
and port additions happen, we reach a state where the MMDevice object
is still tracked by ModemManager and we also have new port additions
tracked that will require explicit port probings before a new modem
object can be created.
Solve this mixup by faking the removal of all existing device ports,
which will end up completely removing hte MMDevice, so that new port
additions reported afterwards also involve the full device probing
process triggered by the plugin manager.
The issue could be reproduced easily on a MBIM device that also
exposed TTYs, as follows:
* mmcli -m a --inhibit, leave it running
* rmmod cdc_mbim && sleep 5 && modprobe cdc_mbim (so that the
cdc-wdm and net ports go away from the system but NOT the TTYs.)
* Cltr+C to stop the inhibit in the mmcli call.
ModemManager would assert as follows:
0x000079918bcf2a3f (libc.so.6 + 0x00087a3f) pthread_key_delete
0x000079918bca7c6c (libc.so.6 + 0x0003cc6c) gsignal
0x000079918bc93462 (libc.so.6 + 0x00028462) abort
0x000079918bf26f03 (libglib-2.0.so.0 - gtestutils.c: 3253) g_assertion_message
0x000079918bf26f66 (libglib-2.0.so.0 - gtestutils.c: 3279) g_assertion_message_expr
0x0000594ff1093d0c (ModemManager - mm-base-manager.c: 1110) remove_device_inhibition
0x0000594ff1093968 (ModemManager - mm-base-manager.c: 1247) inhibit_device_auth_ready
0x000079918c0536a8 (libgio-2.0.so.0 - gtask.c: 1230) g_task_return_now
0x000079918c0536db (libgio-2.0.so.0 - gtask.c: 1244) complete_in_idle_cb
0x000079918bf053fc (libglib-2.0.so.0 - gmain.c: 3417) g_main_context_dispatch
0x000079918bf05704 (libglib-2.0.so.0 - gmain.c: 4211) g_main_context_iterate
0x000079918bf05978 (libglib-2.0.so.0 - gmain.c: 4411) g_main_loop_run
0x0000594ff108de66 (ModemManager - main.c: 217) main
0x000079918bc936c5 (libc.so.6 + 0x000286c5) __libc_init_first
0x000079918bc93781 (libc.so.6 + 0x00028781) __libc_start_main
0x0000594ff108db80 (ModemManager + 0x00061b80) _start
Instead of creating libmm-plugin* and libmm-shared* libraries that are
dlopen()-ed on runtime, allow incorporating all plugins into the
daemon binary itself.
This makes the startup of the daemon much faster and also avoids
issues with builds that require linker namespace isolation.
Fixes https://gitlab.freedesktop.org/mobile-broadband/ModemManager/-/issues/674
We're bumping the current "INFO" level messages to the new "MSG"
level, also making the new level the default.
The old "INFO" level will be used to setup an intermediate level of
logging which is not as verbose as "DEBUG" but still provides some
capabilities to analyze the behavior of a modem.
The mm_base_modem_sync() method is an asynchronous method that
receives a callback and user data, and therefore we MUST always
complete the async method calling that callback. Set that up with a
GTask as usual.
Also, the mm_base_modem_sync_finish() method should be implemented
along with mm_base_modem_sync(), not in the source file of the
caller of the async method. The finish() always depends on how the
async method was implemented, in our case using a GTask.
Under certain rare conditions (e.g. race between querying devices of a
given subsystem and the kernel tearing those devices down), the
subsystem reported for a GUdevDevice seems to be NULL.
So, ensure both subsystem and name are set on the GUdevDevice before we
process them.
The issue has been observed on GUdevDevices listed by
g_udev_client_query_by_subsystem(), not on the ones asynchronously
reported via uevents, but we add the validity check on both places for
consistency.
Fixes https://gitlab.freedesktop.org/mobile-broadband/ModemManager/-/issues/343
This allows us to know which of the subsystem or name for a
removed device is triggering the assertion from just a stack
trace that contains line information.
Instead of creating new clients internally whenever we need them, just
make sure each MMKernelDeviceUdev object keeps a full reference to the
GUdevClient that generated all GUdevDevices.
There is no point in creating a new kernel device object just to
process a remove event; instead, do any matching with existing
kernel device objects by subsystem and name, which is what the generic
backend already did anyway.
This avoids unnecessary lookup of information in sysfs during removal
events, because the port is anyway already gone when we try to look
those up.
Even if udev support is really built and available.
This is extremely useful to test the udev-less setup without fully
recompiling the whole daemon.
E.g.: the daemon can be run like this:
$ sudo /usr/sbin/ModemManager --debug --test-no-udev
And then, the kernel events may be reported using mmcli like this:
$ sudo mmcli --report-kernel-event-auto-scan
Back in Linux < 3.6 days, the cdc-wdm ports exposed by the QMI driver
were flagged as owned by the 'usb' subsystem. That changed in 3.6 when
the subsystem was renamed to 'usbmisc':
https://mail.gnome.org/archives/networkmanager-list/2012-June/msg00125.html
This patch removes all monitoring of the 'usb' subsystem completely,
which is anyway a valid subsystem but for which we shouldn't need any
special handling. Right now, with newer kernels, we were using that
monitoring exclusively to get notified of full USB device remove
events, which is really not required as we already process the port
removals one by one.
We simplify the logic everywhere that attempted to match either the
'usb' or 'usbmisc' subsystems, and we no longer require the explicit
checks for the port name being named 'cdc-wdm[0-9]*' in the code, as
that is already taken care of by the ID_MM_CANDIDATE udev tag rule.
The auth provider is now a singleton (since 20ab6550), one single
object that lives throughout the whole program execution, and so we
don't need to keep our own full references around. This fix makes
sure we don't attempt to unref a full auth provider reference we
don't own.
Fixes https://gitlab.freedesktop.org/mobile-broadband/ModemManager/issues/179
The auth provider setup is a bit over-engineered. Simplify it by
making a single MMAuthProvider object that may or may not use polkit,
depending on configure options. This object is also setup as a
singleton object using the helper MM_DEFINE_SINGLETON_GETTER().
Instead of having the reference to the object manager server only
while the modem is exported, just keep a reference for as long as the
device object exists. This will make it easier to handle reprobing
logic.
This new method allows users of the ModemManager API to take full
control of a given device.
Unlike other operations in the API, the inhibition is maintained as
long as the caller exists in the bus, or until the same caller
uninhibits the device.
https://gitlab.freedesktop.org/mobile-broadband/ModemManager/issues/98
This prevents errors due to nasty typos in the strings.
We define all symbols in a single header file that is NOT considered
part of the API, as there is no need for MM clients to know about
these tags code-wise. These tags are only meaningful when associated
to devices in udev.
Information of each tag is included in the general API documentation.
https://gitlab.freedesktop.org/mobile-broadband/ModemManager/issues/88
Releasing the port on the device looks benign but because it emits
a signal, it could call device_context_port_released and unref the
MMDevice in port_context_unref. This means the MMDevice might be
disposed before we get to the g_object_ref and the subsequent call
to g_hash_table_remove will try to hash a null string, which makes
MM crash.
The hashtable is keyed on the UID of the MMDevice, and its hash
function is g_str_hash. We shouldn't be passing a GObject into
g_hash_table_remove because calling g_str_hash on an MMDevice is
wrong.
All the previous filter rules were applicable per-port independently.
But, we also want to apply rules on a port based on the existence of
other ports on the same device (e.g. allow TTY if the device also has
a NET port). In this case, we need to wait for all ports to appear and
then apply the additional rules.
We re-use the "min wait time" timeout in the plugin-manager for this
same purpose. This timeout is setup to wait for ports to appear before
starting the probing process (e.g. so that plugin filters like the
forbidden-drivers one work). The very same timeout can therefore be
used to check whether we start the probing or not based on additional
filter rules.
The 'default' filter policy was based on blacklisting as much as
possible and otherwise allow.
The new 'strict' filter policy will be based on whitelisting as much
as much as possible, using custom defined rules, and otherwise forbid
the ports.
The new 'paranoid' filter policy is equivalent to the 'strict' filter
after having applied the blacklist rules from the 'default' filter.
Added a new '--filter-policy=[POLICY]' option in the daemon, which
allows selecting between the supported filter policies. For now, only
two policies are defined:
* default: the default policy used by ModemManager, where it tries
to probe and detect as many modem ports as possible.
* whitelist-only: only devices explicitly tagged via udev (with the
ID_MM_DEVICE_PROCESS tag) will be probed and used.
E.g. forcing a MBIM modem to run in AT-only mode:
# MM_FILTER_RULE_NET=0 \
MM_FILTER_RULE_CDC_WDM=0 \
/usr/sbin/ModemManager --debug
This is just for quick testing for now.
This new object allows configuring the filter rules applied to the
device ports. By default, for now, it implements the same rules as the
MMKernelDevice is_candidate() method, which is obsoleted.
==28888== 622 (280 direct, 342 indirect) bytes in 7 blocks are definitely lost in loss record 2,515 of 2,548
==28888== at 0x6474014: g_type_create_instance (in /usr/lib/libgobject-2.0.so.0.5200.3)
==28888== by 0x6455027: ??? (in /usr/lib/libgobject-2.0.so.0.5200.3)
==28888== by 0x6456A54: g_object_newv (in /usr/lib/libgobject-2.0.so.0.5200.3)
==28888== by 0x6457213: g_object_new (in /usr/lib/libgobject-2.0.so.0.5200.3)
==28888== by 0x4E9DD7D: mm_kernel_event_properties_new (mm-kernel-event-properties.c:422)
==28888== by 0x4E9D8BC: mm_kernel_event_properties_new_from_string (mm-kernel-event-properties.c:283)
==28888== by 0x1465D7: process_initial_kernel_events (mm-base-manager.c:555)
==28888== by 0x14679E: mm_base_manager_start (mm-base-manager.c:581)
==28888== by 0x143CF3: name_acquired_cb (main.c:110)
==28888== by 0x616B805: ??? (in /usr/lib/libgio-2.0.so.0.5200.3)
==28888== by 0x616BA47: ??? (in /usr/lib/libgio-2.0.so.0.5200.3)
==28888== by 0x612FD52: ??? (in /usr/lib/libgio-2.0.so.0.5200.3)
When a USB modem is switching its USB configuration, udev may deliver
the remove events of USB interfaces associated with the old USB
configuration and the add events of USB interfaces associated with the new USB
configuration in an interleaved fashion. An interleaved remove event of USB
interface could trigger the special case handling code in
mm-base-manager.c:device_removed() and incorrectly remove a MMDevice under
probing.
See https://lists.freedesktop.org/archives/modemmanager-devel/2017-August/005626.html
for more details.
This patch adds a check to ensure that only remove events of USB
device (i.e. not interface) can trigger the special handling code.
Aleksander Morgado
Lukas Voegl
Aleksander Morgado
Rukun Mao
Dylan Van Assche
Andrew Lassalle
Eric Caruso
Ben Chan