Add a common physdev UID to all ports in the Qualcomm SoC, so that they
are all bound together to the same modem object.
This change assumes that no external plugged device ever exports a
QRTR node in the QRTR bus (e.g. that no PCI or USB device would ever
do that). This assumption is likely wrong, but given that we're
currently limiting the QRTR support to the Qualcomm SoC plugin, the
assumption is valid for us for now.
ModemManager[6954]: <warn> [1618260091.319919] [br0] invalid sysfs path read for net/br0
**
ERROR:kerneldevice/mm-kernel-device-generic.c:235:ptr_array_add_sysfs_attribute_link_basename: assertion failed: (array && sysfs_path && attribute)
Bail out! ERROR:kerneldevice/mm-kernel-device-generic.c:235:ptr_array_add_sysfs_attribute_link_basename: assertion failed: (array && sysfs_path && attribute)
Thread 1 "ModemManager" received signal SIGABRT, Aborted.
(gdb) bt
#0 0x00007ffff7438ef5 in raise () at /usr/lib/libc.so.6
#1 0x00007ffff7422862 in abort () at /usr/lib/libc.so.6
#2 0x00007ffff76c2084 in () at /usr/lib/libglib-2.0.so.0
#3 0x00007ffff771d85d in g_assertion_message_expr () at /usr/lib/libglib-2.0.so.0
#4 0x00005555556a5337 in ptr_array_add_sysfs_attribute_link_basename (array=0x5555557de520, sysfs_path=0x0, attribute=0x555555703bd8 "driver", out_value=0x0)
at kerneldevice/mm-kernel-device-generic.c:235
#5 0x00005555556a542b in preload_contents_other (self=0x5555557a6c60) at kerneldevice/mm-kernel-device-generic.c:255
#6 0x00005555556a6485 in preload_contents (self=0x5555557a6c60) at kerneldevice/mm-kernel-device-generic.c:517
#7 0x00005555556a7fb6 in check_preload (self=0x5555557a6c60) at kerneldevice/mm-kernel-device-generic.c:980
#8 0x00005555556a855f in initable_init (initable=0x5555557a6c60, cancellable=0x0, error=0x7fffffffe5c8) at kerneldevice/mm-kernel-device-generic.c:1127
#9 0x00007ffff78a2f41 in g_initable_new_valist () at /usr/lib/libgio-2.0.so.0
#10 0x00007ffff78a2ffa in g_initable_new () at /usr/lib/libgio-2.0.so.0
#11 0x00005555556a8205 in mm_kernel_device_generic_new_with_rules (props=0x5555557a5fa0, rules=0x7fffec001e30, error=0x7fffffffe5c8) at kerneldevice/mm-kernel-device-generic.c:1042
#12 0x00005555556a827d in mm_kernel_device_generic_new (props=0x5555557a5fa0, error=0x7fffffffe5c8) at kerneldevice/mm-kernel-device-generic.c:1063
#13 0x00005555555aba9c in handle_kernel_event (self=0x555555793220, properties=0x5555557a5fa0, error=0x7fffffffe5c8) at mm-base-manager.c:414
We can now setup a 'lower device' object that will be the one
providing all physdev info. This is useful when using virtual ports
that refer to a 'lower' real physical device, in order to bind
together all those ports on the same modem object.
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.
We already have methods to query for interface specific attributes
like class/subclass/protocol, so add a new one for the interface
number, and make sure we use ATTRS{bInterfaceNumber} to load it
always, instead of assuming the ID_USB_INTERFACE_NUM property is set.
In addition to loading port and device properties, we now also allow
loading sysfs properties that are assumed to be static (i.e. their
values won't change since loaded the first time).
Instead of the custom simple implementation which only supported the
'*' modifier in the pattern. This allows us to support e.g. attribute
value matches like e.g. 'DATA[0-9]*_CNTL'.
Until now we did not support ATTRS{} matches against attributes we
don't support in the core codebase. Implement now a simple lookup
mechanism which traverses the tree of sysfs path from the port sysfs
path to the physical device sysfs path, looking up the attribute
requested.
This is not completely equivalent to what udev does, because the udev
rules matching ATTRS would usually also include an additional previous
matching e.g. SUBSYSTEMS and such, so that the ATTRS is looked up
exactly in the device that also matches the additional previous
rules. In our case, we just simplify the logic and return the first
one found.
We were applying the DRIVERS check looking only at the single port
driver. Instead, we now lookup and cache all drivers found in the
device tree, and apply the loose DRIVERS check properly looking at all
of them.
We were applying the SUBSYSTEMS and SUBSYSTEM check looking at the
sysfs path and just hoping the subsystem we're looking for is included
in the path itself. Instead, we now lookup and cache all susystems
found in the device tree, and apply the loose SUBSYSTEMS check
properly looking at all of them.
E.g. we can now match SUBSYSTEMS="mhi_q" in the following device tree,
without needing it to be found in the sysfs path:
looking at device '/devices/pci0000:00/0000:00:1b.0/0000:01:00.0/1001_00.01.00_MBIM/mhi_uci_q/mhi_MBIM':
SUBSYSTEM=="mhi_uci_q"
looking at parent device '/devices/pci0000:00/0000:00:1b.0/0000:01:00.0/1001_00.01.00_MBIM':
SUBSYSTEMS=="mhi_q"
looking at parent device '/devices/pci0000:00/0000:00:1b.0/0000:01:00.0':
SUBSYSTEMS=="pci"
The generic backend implementation was really based on detecting USB
devices, not so much devices in other subsystems. This patch puts the
generic backend at the same level as the udev backend to support
non-USB modems.
USB, PCI, PCMCIA... all these different system buses expose modems in
different ways. Instead of having single methods to attempt to load
different things for all these device types, detect first which is the
system bus in use, and then perform a bus-specific operation to
preload the different things.
We can just subclass the methods to check whether a given property
exists and to get it as a string, and then implement in the generic
class the actual boolean/int/hex type getters common for all.
If multiple kernel device types inherit from MMKernerDevice, and those
are compared against each other, the current logic returns TRUE if the
G_OBJECT_TYPE of one of them is smaller than the other. This function is
checking for equality, so returning FALSE is enough.
This allows MMKernelDevice::cmp to compare two kernel devices
with different object types, which means that subclasses can
continue to only handle comparisons with their own type. The
order may not be stable across builds.
E.g. it may be one more layer up:
looking at device '/devices/pci0000:00/0000:00:14.0/usb1/1-8/1-8.4/1-8.4:1.1/ttyUSB1/tty/ttyUSB1':
KERNEL=="ttyUSB1"
SUBSYSTEM=="tty"
DRIVER==""
looking at parent device '/devices/pci0000:00/0000:00:14.0/usb1/1-8/1-8.4/1-8.4:1.1/ttyUSB1':
KERNELS=="ttyUSB1"
SUBSYSTEMS=="usb-serial"
DRIVERS=="option1"
ATTRS{port_number}=="0"
looking at parent device '/devices/pci0000:00/0000:00:14.0/usb1/1-8/1-8.4/1-8.4:1.1':
KERNELS=="1-8.4:1.1"
SUBSYSTEMS=="usb"
DRIVERS=="option"
ATTRS{bNumEndpoints}=="02"
ATTRS{supports_autosuspend}=="1"
ATTRS{bInterfaceNumber}=="01"
ATTRS{bAlternateSetting}==" 0"
ATTRS{bInterfaceSubClass}=="06"
ATTRS{bInterfaceProtocol}=="13"
ATTRS{interface}=="Huawei Mobile Connect - Application"
ATTRS{bInterfaceClass}=="ff"
ATTRS{authorized}=="1"
kerneldevice/mm-kernel-device-udev.c: In function ‘mm_kernel_device_udev_new_from_properties’:
kerneldevice/mm-kernel-device-udev.c:770:70: error: declaration of ‘properties’ shadows a global declaration [-Werror=shadow]
770 | mm_kernel_device_udev_new_from_properties (MMKernelEventProperties *properties,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~
kerneldevice/mm-kernel-device-udev.c:38:20: note: shadowed declaration is here
38 | static GParamSpec *properties[PROP_LAST];
| ^~~~~~~~~~
kerneldevice/mm-kernel-device-generic.c: In function ‘mm_kernel_device_generic_new_with_rules’:
kerneldevice/mm-kernel-device-generic.c:954:68: error: declaration of ‘properties’ shadows a global declaration [-Werror=shadow]
954 | mm_kernel_device_generic_new_with_rules (MMKernelEventProperties *properties,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~
kerneldevice/mm-kernel-device-generic.c:46:20: note: shadowed declaration is here
46 | static GParamSpec *properties[PROP_LAST];
| ^~~~~~~~~~
kerneldevice/mm-kernel-device-generic.c: In function ‘mm_kernel_device_generic_new’:
kerneldevice/mm-kernel-device-generic.c:971:57: error: declaration of ‘properties’ shadows a global declaration [-Werror=shadow]
971 | mm_kernel_device_generic_new (MMKernelEventProperties *properties,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~
kerneldevice/mm-kernel-device-generic.c:46:20: note: shadowed declaration is here
46 | static GParamSpec *properties[PROP_LAST];
| ^~~~~~~~~~
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
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.
This patch implicitly enables in the generic device backend the
manual-only greylist (ID_MM_DEVICE_MANUAL_SCAN_ONLY tag) and the
platform TTY whitelist (ID_MM_PLATFORM_DRIVER_PROBE), which were not
being applied.
A USB device may report its manufacturer through the Manufacturer String
Descriptor. That information can sometimes be useful, e.g. MBIM doesn't
define a standard command to query the manufacturer. This patch adds a
method to MMKernelDevice to retrieve the manufacturer information from
the underlying physical device, which can later be used by
MMBroadbandModemMbim to report the manufacturer of a MBIM modem.
==5075== 198 bytes in 18 blocks are definitely lost in loss record 4,486 of 4,761
==5075== at 0x4C2CE5F: malloc (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==5075== by 0x66E3028: g_malloc (in /usr/lib/libglib-2.0.so.0.5200.3)
==5075== by 0x66CAD42: g_path_get_basename (in /usr/lib/libglib-2.0.so.0.5200.3)
==5075== by 0x20BB7B: preload_interface_sysfs_path (mm-kernel-device-generic.c:184)
==5075== by 0x20C885: preload_contents (mm-kernel-device-generic.c:383)
==5075== by 0x20E2E1: check_preload (mm-kernel-device-generic.c:830)
==5075== by 0x20EA6F: set_property (mm-kernel-device-generic.c:953)
==5075== by 0x6455359: ??? (in /usr/lib/libgobject-2.0.so.0.5200.3)
==5075== by 0x6456DBC: g_object_new_valist (in /usr/lib/libgobject-2.0.so.0.5200.3)
==5075== by 0x6105E95: g_initable_new_valist (in /usr/lib/libgio-2.0.so.0.5200.3)
==5075== by 0x6105F58: g_initable_new (in /usr/lib/libgio-2.0.so.0.5200.3)
==5075== by 0x20E848: mm_kernel_device_generic_new_with_rules (mm-kernel-device-generic.c:901)
==5075== 333 bytes in 37 blocks are definitely lost in loss record 4,612 of 4,761
==5075== at 0x4C2CE5F: malloc (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==5075== by 0x66E3028: g_malloc (in /usr/lib/libglib-2.0.so.0.5200.3)
==5075== by 0x66CAD42: g_path_get_basename (in /usr/lib/libglib-2.0.so.0.5200.3)
==5075== by 0x20BE20: preload_driver (mm-kernel-device-generic.c:242)
==5075== by 0x20C8E5: preload_contents (mm-kernel-device-generic.c:391)
==5075== by 0x20E2E1: check_preload (mm-kernel-device-generic.c:830)
==5075== by 0x20EA6F: set_property (mm-kernel-device-generic.c:953)
==5075== by 0x6455359: ??? (in /usr/lib/libgobject-2.0.so.0.5200.3)
==5075== by 0x6456DBC: g_object_new_valist (in /usr/lib/libgobject-2.0.so.0.5200.3)
==5075== by 0x6105E95: g_initable_new_valist (in /usr/lib/libgio-2.0.so.0.5200.3)
==5075== by 0x6105F58: g_initable_new (in /usr/lib/libgio-2.0.so.0.5200.3)
==5075== by 0x20E848: mm_kernel_device_generic_new_with_rules (mm-kernel-device-generic.c:901)
get_device_ids() in mm-kernel-device-udev.c accepts a NULL 'vendor' or
'product' argument, but the current implementation could result in a
potential NULL dereferences of the 'vendor' argument. Given that
get_device_ids() is a local helper and its only caller provides a
non-NULL 'vendor' and 'product' argument, this patch removes the NULL
checks (i.e. get_device_ids() expects non-NULL 'vendor' and 'product').
This patch also rearranges the code such that the 'vendor' argument is
updated only when the function returns TRUE, just like how the 'product'
argument is handled.
g_free and g_object_unref are in form of `void (*)(gpointer)`, which
matches the GDestroyNotify signature. An explicit GDestroyNotify cast on
g_free and g_object_unref is thus not needed.
There are 2 main types of udev properties: device-specific and
port-specific.
The port-specific properties are set independently per port (e.g. port
type hints set per interface number for a given vid:pid).
The device-specific properties apply to all ports in the device. Some
of these properties are currently expected in the physical device
(e.g. ID_MM_PLATFORM_DRIVER_PROBE) while some others are expected in
each port (e.g. the plugin udev tag filters).
This patch tries to simplify the logic and just assume that the device
specific tags may be given in either the physical device or the port
device, by providing separate APIs to retrieve port-specific or
device-specific (global) properties. If the same tag is given in both
the device and the port, the one in the device takes preference.
For the generic backend, these new APIs are really useless, as all
device-specific and port-specific properties are always stored in the
port object themselves (there is no 'tree' of devices in the generic
backend, no 'physdev' device).
For the udev backend, though, there really is a difference, as the
tags may be set in port or device.
https://bugs.freedesktop.org/show_bug.cgi?id=100156
find_device_by_physdev_uid() expects a non-NULL UID of the physical
device. However, mm_kernel_device_get_physdev_uid() could potentially
return NULL on MMKernelDeviceUdev if find_physical_gudevdevice() in
mm-kernel-device-udev.c fails to find the physical device. When a NULL
physical device UID is passed to find_device_by_physdev_uid() and used
in g_hash_table_lookup(), it leads to a crash in g_str_hash(), which is
a bit obscure.
This patch updates kernel_device_get_physdev_uid() in
mm-kernel-device-udev.c to fall back to use the device sysfs if the
physical device sysfs isn't available, which is similar to how
kernel_device_get_physdev_uid() mm-kernel-device-generic.c is
implemented.
Aleksander Morgado
Andrew Lassalle
Eric Caruso
Baruch Siach
Ben Chan