Now incorporate is fixed to copy across the properties we need in the
superclass, we don't need to do the subclass ->probe().
Note, we still need to do the subclassed ->probe() when using
FU_UDEV_DEVICE_FLAG_VENDOR_FROM_PARENT or when looking at properties
on the parent device.
This also removes the spurious 'already set GType to FuVliUsbhubDevice,
ignoring FuVliUsbhubDevice' messages when running the daemon.
This allows us to print better warning strings, and in the future
would allow us to profile each operation in a meaningful way.
Also, add context to some of the progress steps as required.
tristate features will automatically disable if dependencies marked
as required are missing.
Packagers can manually override using `auto_features`.
Link: https://mesonbuild.com/Build-options.html#features
We were calling g_module_symbol() 2703 times, which is actually more
expensive than you'd think.
It also means the plugins are actually what we tell people they are:
A set of vfuncs that get run. The reality before that they were dlsym'd
functions that get called at pretty random times.
We only had to pile everything into the src/fuzzing/firmware directory
because honggfuzz could not cope with more than one input path.
This way each plugin is self contained and easy to copy.
Also, install the fuzzing builder objects as this fixes the installed
tests when srcdir does not exist.
Based on a patch by Jan Tojnar <jtojnar@gmail.com>, many thanks.
In reality these are not super interesting as they only happen on
`->write()` and not `->parse()`.
In other news, the fuzzer now appreciates how critical the alignment
is, which is probably a good thing generally.
Fixes https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=40088
It's actually quite hard to build a front-end for fwupd at the moment
as you're never sure when the progress bar is going to zip back to 0%
and start all over again. Some plugins go 0..100% for write, others
go 0..100% for erase, then again for write, then *again* for verify.
By creating a helper object we can easily split up the progress of the
specific task, e.g. write_firmware().
We can encode at the plugin level "the erase takes 50% of the time, the
write takes 40% and the read takes 10%". This means we can have a
progressbar which goes up just once at a consistent speed.
There are now multiple plugins using drm_dp_aux_dev interface which
may potentially be combined with an amdgpu. Prevent exercising this
interface with any plugin using DP aux unless a new enough kernel is
installed.
There is a lot of code in fwupd that just assigns a shared object type to
a FuPlugin, and then for each device on that plugin assigns that same shared
object to each FuDevice.
Rather than proxy several kinds of information stores over two different levels
of abstraction create a 'context' which contains the shared *system* state
between the daemon, the plugins and the daemon.
This will allow us to hold other per-machine state in the future, for instance
the system battery level or AC state.
This makes a lot more sense; we can parse a firmware and export the same XML
we would use in a .builder.xml file. This allows us to two two things:
* Check we can round trip from XML -> binary -> XML
* Using a .builder.xml file we can check ->write() is endian safe
This allows us to 'nest' firmware formats, and removes a ton of duplication.
The aim here is to deprecate FuFirmwareImage -- it's almost always acting
as a 'child' FuFirmware instance, and even copies most of the vfuncs to allow
custom types. If I'm struggling to work out what should be a FuFirmware and
what should be a FuFirmwareImage then a plugin author has no hope.
For simple payloads we were adding bytes into an image and then the image into
a firmware. This gets really messy when most plugins are treating the FuFirmware
*as* the binary firmware file.
The GBytes saved in the FuFirmware would be considered the payload with the
aim of not using FuFirmwareImage in the single-image case.
The best way of not getting something wrong is to not require it in the first
place...
All plugins now use DeviceInstanceId-style quirk matches and we can just drop
the prefix in all files. We were treating HwId=, Guid= and DeviceInstanceId= in
exactly the same way -- they're just converted to GUIDs when building the silo!
Devices may want to support more than one protocol, and for some devices
(e.g. Unifying peripherals stuck in bootloader mode) you might not even be able
to query for the correct protocol anyway.
It is far too easy to forget to set FWUPD_DEVICE_FLAG_NO_GUID_MATCHING for new
plugins, and without it it all works really well *until* a user has two devices
of the same type installed at the same time and then one 'disappears' for hard
to explain reasons. Typically we only need it for replug anyway!
Explicitly opt-in to this rarely-required behaviour, with the default to just
use the physical and logical IDs. Also document the update behavior for each
plugin to explain why the flag is being used.
This allows you to have two identical Unifying plugged in without one of them
being hidden from the user, at the same time allowing a HIDRAW<->USB transition
when going to and from bootloader and runtime modes.
This removes the workaround added in 99eb3f06b6.
Fixes https://github.com/fwupd/fwupd/issues/2915
This allows a device subclass to call the parent method after doing an initial
action, or even deliberately not call the *generic* parent method at all.
It also simplifies the plugins; you no longer have to remember what the plugin
is deriving from and accidentally clobber the wrong superclass method.
This allows us to check that the plugin is writing into a mutable buffer. Also
fix up the plugins that are currently 'wrong' and use the new function for the
plugins doing the right thing.
At the moment FuChunks are sometimes mutable, and sometimes immutable, and it's
all a bit too low level for comfort.
Before we can do any kind of optimisation or verification we need plugins to
stop reading directly from the C structure. The aim here is to make FuChunk
optionally mutable without making assumptions about the memory model, and also
to be able to introspect it for the docs.
When this is done, include:
* Including the hash
* Including anything that is not ABI stable in plugins yet
Suggested-by: Simon McVittie <smcv@debian.org>
The end year is legally and functionally redundant, and more importantly causes
cherry-pick conflicts when trying to maintain old branches. Use git for history.
This is a safer version of g_bytes_new_from_bytes() which returns a GError
if the offsets are invalid rather than emitting a critical warning.
This prevents a critical warning and potential crash when parsing invalid
bcm57xx firmware.
