90 driver main functions repeated the same three strcmp() checks and
calls. I2CSPIDriver<T>::module_main() does it once; drivers with extra
verbs keep their own dispatch.
Assisted-by: Claude:claude-opus-5-5
Add the PX4 arch layer for the i.MX9 family: ADC, HRT on LPTMR with
LPIT deadlines, TPM io timers, pin interrupts, LPSPI with chip selects
for all eight instances, critmon on the DWT cycle counter, a tone alarm
on a TPM channel, version reporting with the 128-bit ECID read from the
Fuse Status Block as the UUID, the board reset entry point, plus the
i.MX95 chip include set, the chip mapping in px4_impl_os.cmake and the
PX4_SOC_ARCH_ID_IMX95 id. DShot uses the shared NXP FlexIO driver with
the i.MX9 FlexIO base, offsets, IRQ, clock root and pad hooks.
Not implemented yet: reset of the M7 through the System Manager, LED
PWM and PWM input capture.
i.MX9 timer configs name their timers Timer::TPMn and mark DShot
capable ones with initIOTimerDshot(), so teach
output_groups_from_timer_config.py both forms.
Co-authored-by: Andre Heinemans <andre.heinemans@nxp.com>
About 1700 call sites loaded the log level as the first argument of
px4_log_modulename(). Per-level functions drop that load and leave one
more register for the format arguments.
Assisted-by: Claude:claude-opus-5-5
CONSTRAINED_FLASH_NO_HELP only emptied the PRINT_MODULE_USAGE_* bodies,
so every call site and its strings stayed in flash. Make the helpers
no-op inlines there, as PRINT_MODULE_DESCRIPTION already is on NuttX;
only the usage line with the docs link is still printed.
Assisted-by: Claude:claude-opus-5-5
The constructor loop that assigns and subscribes every instance was
expanded into each of the ~55 SubscriptionMultiArray members in the
firmware. Move it into a non-template function in Subscription.cpp,
called with the array, its size and the topic.
px4_fmu-v6x_default: 1918712 -> 1917160 bytes of text, -1552 B.
Assisted-by: Claude:claude-fable-5-1
Co-authored-by: Balduin <balduin@auterion.com>
The constructor was defined in the header, so every one of the ~150
module and library classes deriving from ModuleParams expanded it into
its own constructor: the ListNode and child-list initialisation plus
the setParent() call, repeated per class.
Moving the body next to setParent() in module_params.cpp leaves a single
call at each site. px4_fmu-v6x_default: 1934288 -> 1932520 bytes of
text, -1768 B.
Assisted-by: Claude:claude-fable-5-1
Co-authored-by: Balduin <balduin@auterion.com>
The constructor was defined in the header, so each of the ~75 I2C/SPI
drivers expanded the ScheduledWorkItem and I2CSPIInstance initialisation
into its own constructor.
Moving the body to i2c_spi_buses.cpp leaves one call per driver.
px4_fmu-v6x_default: 1932520 -> 1930616 bytes of text, -1904 B.
Assisted-by: Claude:claude-fable-5-1
Co-authored-by: Balduin <balduin@auterion.com>
The base implementation, which walks the child list and then calls
updateParamsImpl(), was defined in the class body. Every module that
calls updateParams() on a parameter update therefore inlined the list
walk. Define it in module_params.cpp next to setParent().
px4_fmu-v6x_default: 1917160 -> 1915056 bytes of text, -2104 B.
Assisted-by: Claude:claude-fable-5-1
Co-authored-by: Balduin <balduin@auterion.com>
* fix(vectornav): limit init based on retries, set max thread prio
Signed-off-by: dirksavage88 <dirksavage88@gmail.com>
* fix(vnav): set lower priority than max
Signed-off-by: dirksavage88 <dirksavage88@gmail.com>
* chore(vnav): remove unused variable
Signed-off-by: dirksavage88 <dirksavage88@gmail.com>
* fix(vnav): set scheduler priority definition in tasks header, base off of PX4_WQ_HP_BASE
Signed-off-by: dirksavage88 <dirksavage88@gmail.com>
---------
Signed-off-by: dirksavage88 <dirksavage88@gmail.com>
A publisher calls call() under the node lock until the subscription is
unregistered. SubscriptionBlocking, SubscriptionCallbackWorkItem and
the MCP CallbackHandler left that to ~SubscriptionCallback(), which runs
after the subclass is gone, so a publication during destruction aborted
with "pure virtual method called". SubscriptionBlocking could also signal
its already destroyed condition variable. SITL mag calibration in SIH hit
this on almost every side.
The new functional test aborts on both subscription types within about
a second without the fix.
Assisted-by: Claude:claude-opus-5-5
While a topic has no device node, every Subscription::update(), copy(),
updated() and advertised() repeats the device node lookup and fails
again. Polling modules touch many optional topics on every loop: in a
SITL flight about 8% of all instructions, and about a quarter of each
MAVLink thread, were such failed lookups.
Device nodes are never removed, so a lookup that failed cannot succeed
before a node is added. Count the nodes added in DeviceMaster and only
retry the lookup once that count has changed. Protected builds, where
user space has no cheap access to it, keep retrying every time.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Benjamin Cabé <benjamin@zephyrproject.org>
This reverts commit d92a0c6b3. In protected builds px4_work_queue is also
linked into the kernel image, where px4_layer's user-side objects collide
with the kernel ones (px4_fmu-v5_protected). The Erbium target no longer
needs the link dependency, since its modules pull in px4_task_spawn_cmd()
and px4_task_join() before px4_work_queue is linked.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Afonso Oliveira <afonso.oliveira707@gmail.com>
Erbium is a RISC-V RV64IMFC chip whose NuttX port is in PX4/NuttX. Add
the PX4 platform layer for it: HRT, board reset, critmon time base and
board identity, the riscv-none-elf toolchain file and the rv64imfc
compiler flags, which include -mno-fdiv because the ET-Minion core has
no hardware for FDIV.S and FSQRT.S.
The HRT reads the machine timer and runs its callouts in a task, since
the NuttX RISC-V trap path does not save the FP registers that callouts
may use.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Afonso Oliveira <afonso.oliveira707@gmail.com>
px4_at24c_initialize() takes a struct i2c_master_s pointer. Without a
declaration in scope the struct is local to the prototype, which is an
error with -Werror on a target that does not include the NuttX I2C
headers before px4_mtd.h.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Afonso Oliveira <afonso.oliveira707@gmail.com>
WorkQueueManager starts and joins its thread with px4_task_spawn_cmd()
and px4_task_join() from px4_layer, but px4_work_queue did not declare
that dependency. Targets with many modules pull those objects in before
px4_work_queue is linked; a target with only a few system commands
fails to link.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Afonso Oliveira <afonso.oliveira707@gmail.com>
events::send(id, log_levels, message) was an inline function, so every
module using it got either a private clone or the body inlined at each
call site. The message is never part of the sent event: it only feeds
the build-time event generator and the optional console print.
Keep a thin inline wrapper so the compiler still drops the unused
message string, and build the event in one out-of-line function that
takes the two log levels as scalars to keep the call sites small. This
saves 1760 bytes of FLASH on px4_fmu-v6x and 1184 bytes on px4_fmu-v2.
Assisted-by: Claude:claude-fable-5-1
Signed-off-by: Michael Rogov Papernov <michael@membrowse.com>
Modules that stop through ModuleBase::exit_and_cleanup() delete themselves
inside Run(), and since 9b62a1701d WorkQueue::Run() then clears
_run_in_progress in the freed object. ASAN aborts SITL on
`battery_simulator stop` and the stop of most other work queue modules.
Track the in-flight item in the WorkQueue instead, which lives until its
own Run() exits on the worker thread.
Assisted-by: Claude:claude-opus-5-5
* fix(dataman_client): keep request topic advertised on client teardown
* fix(uorb): keep shared topics advertised on publisher teardown
Single-instance publications share one advertised flag, so destroying one
publisher must not hide updates from the others.
Keep automatic unadvertising in PublicationMultiBase and remove the
DatamanClient workaround. Cover surviving publishers and instance reuse,
and reset retained input samples between test cases.
Tested: make tests (215/215), make format.
Signed-off-by: jonas <jonas.perolini@rigi.tech>
Assisted-by: Codex
* docs(uorb): correct unadvertise comments and document teardown
The unadvertise comment claimed it has no effect on single-instance
topics, which is how one publisher's teardown came to hide the others.
State what clearing the shared flag does, and document that only
PublicationMulti releases its instance on destruction.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: jonas <jonas.perolini@rigi.tech>
* test(uorb): keep test cases independent of retained samples
A destroyed single-instance publication now leaves its last sample
visible, so ManualControlTest, test_prec_takeoff and mixer_module_tests
passed only in declaration order and failed under --gtest_shuffle.
Reset with an inert sample where one exists, otherwise unadvertise the
test's own publications, and drop module outputs left over from a
previous case.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: jonas <jonas.perolini@rigi.tech>
* fix(boards): drop the PublicationBase destructor from the ITCM lists
The destructor is now trivial and no longer emitted, so itcm_check.py
fails on the missing section.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
---------
Signed-off-by: jonas <jonas.perolini@rigi.tech>
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
Co-authored-by: jonas <jonas.perolini@rigi.tech>
Co-authored-by: Jacob Dahl <dahl.jakejacob@gmail.com>
* fix(platforms/common): sleep with nanosleep() in px4_usleep()
Outside of the SITL lockstep build, px4_usleep() was usleep(). POSIX only requires usleep() to accept less than 1000000 microseconds and lets it fail with EINVAL from there; POSIX.1-2001 marked it obsolescent and POSIX.1-2008 removed it in favour of nanosleep(). glibc, NuttX and macOS sleep for longer anyway, but other implementations return EINVAL at once, and as no caller checks the result, the sleep would silently not happen. More than twenty call sites ask for a second or more, such as uorb top, mavlink, the ESC calibration or voxl2_io, and some durations are computed at run time, such as the one-second idle period of the POSIX HRT thread.
Outside of the lockstep build, px4_usleep() is now a px4_platform function that converts the microseconds to the timespec of nanosleep(), so the call sites keep passing microseconds of any duration. It is not inline: expanding the conversion at every call site grew px4_fmu-v6xrt_default by 4 kB of flash. QURT has no nanosleep(), which PX4 only stubs there, but its usleep() accepts a second or more, so px4_usleep() remains usleep() there. system_usleep() remains usleep(), for the wall-clock sleeps that bypass the lockstep scheduler.
A signal still ends the sleep early, with -1 and EINTR, rather than restarting it for the remaining time: on POSIX, the HRT and work queue threads sleep in px4_usleep(), and hrt_work_queue() and work_queue() wake them with SIGCONT, so a restarted sleep would delay hrt_call_*() callouts by up to a second. On NuttX, no PX4 code signals its tasks, and px4_task_delete() does not depend on the sleep returning: without cancellation points, which no board enables, task_delete() terminates the task wherever it is.
Failures are now -1 with errno, where the usleep() of NuttX returned the error number. No call site checks the result, so none changes. The sleep test now also sleeps for 1.2 s.
Assisted-by: Claude:opus-5.5
* perf(boards): keep the sleep path in ITCM on i.MX RT boards
The ITCM lists of these boards hold the sleep path of NuttX: usleep(), clock_nanosleep() and nxsig_nanosleep(). px4_usleep() now enters it through its own function and nanosleep(), which take the place of usleep(), so the path stays in ITCM. usleep() remains linked for the code that still calls it directly, such as the start-up code of the boards and NSH.
Assisted-by: Claude:opus-5.5
* fix(drivers): sleep for a second or more with px4_usleep()
bmi088_i2c, microstrain, frsky_telemetry and the bit rate detection of the kinetis, stm32 and stm32h7 UAVCAN drivers asked usleep() for a second or more, which POSIX doesn't require it to accept. The usleep() of NuttX accepts it, but px4_usleep() now does on every platform.
Assisted-by: Claude:opus-5.5
Boards that override their UUID, such as SITL and the Linux boards, only had the functions that format the GUID, while DroneCAN and Cyphal send the GUID itself to the other nodes, so these drivers couldn't be linked for them. board_get_mfguid() and board_get_px4_guid() now give the bytes that the formatting functions print.
Assisted-by: Claude:opus-5.5
Signed-off-by: Benjamin Cabé <benjamin@zephyrproject.org>
The fmu-v6xrt boards can be fitted with either of two octal flash parts, and
nothing told the user which one is on the board. The identity is already
known: the bootloader stage reads the JEDEC ID to pick a config, and the app
stage recognises the config from the installed read sequence, so no extra
flash access is needed to report it.
Keep the part name behind a weak board_get_flash_type_name() and append it to
the chip tag when the board provides one, so `ver mcu` and the logged chip
name read e.g. "i.MX RT1170 rA0 IS25WX512M". Boards that do not implement it
are unaffected and their tag is unchanged.
Signed-off-by: Peter van der Perk <peter.vanderperk@nxp.com>
Each default driver instantiation emits the same allocation and initialization failure handling. Share that code outside the template while keeping init() resolved on the concrete driver. Allocation failures are now logged under SPI_I2C.
Extracted from Balduin (mbjd) in PX4/PX4-Autopilot#27816, commit 8a2391730f.
Assisted-by: Codex
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
Co-authored-by: Balduin <balduin@auterion.com>
On NuttX these are two things: / is the filesystem root, and /fs/microsd is the
SD card mounted under it. MAVLink FTP serves the root and confines writes to the
SD card, so the read-only ROMFS at /etc is visible but cannot be written.
On POSIX they were the same directory. PX4_ROOTFSDIR and PX4_STORAGEDIR both
resolved to CONFIG_BOARD_ROOT_PATH, which on SITL is ".", so the working
directory was simultaneously the FTP root and the only writable area. The
consequences were that the ROMFS symlink sat inside the FTP root, and that
_validatePathIsWritable() had nothing meaningful to check against and so was
compiled out on POSIX entirely, leaving no write restriction at all.
Give POSIX the same split. CONFIG_BOARD_ROOT_PATH keeps its meaning as the
storage directory, and a new CONFIG_BOARD_FS_ROOT_PATH names the root FTP
serves, defaulting to the storage path so every existing board is unchanged.
SITL sets the root to "." and storage to "./fs", which mirrors NuttX: logs,
parameters, dataman and eeprom move under ./fs, and etc/ stays in the root as
read-only data.
With storage distinct from the root, the write restriction now applies on every
platform rather than NuttX only, and no longer compares against a hardcoded
prefix length that was wrong for any board not using /fs/microsd.
Assisted-by: Claude:claude-opus-5[1m]
Signed-off-by: Julian Oes <julian@oes.ch>
* fix(uavcan): apply UAVCAN_BITRATE on FlexCAN again
58dfff0de7 turned the SIOCSCANBITRATE path off for i.MX RT, S32K and Kinetis,
which is every board in the tree that builds SocketCAN, so UAVCAN_BITRATE
stopped doing anything at all. It was written before db6f7ec304, and the crash
it works around is that commit's bug: the driver reports bit/s on NuttX 12 and
the caller still divided by 1000, so a 1 Mbit/s board asked its controller for
1000 bit/s on every boot.
Benched on an ARK FMU-v6XRT with the gate removed: 1 Mbit/s -> 500 kbit/s ->
1 Mbit/s round-trips with both DroneCAN nodes coming back OPERAT, buses
error-active, no fault. Requesting a rate the controller cannot reach is also
survivable now -- it is reported and the configured rate is kept, where before
a negative return from here made CanDriver::init() give up and DroneCAN never
started.
The %u for a uint32_t is fixed as well; nothing compiled this branch while the
gate was in place.
* fix(work_queue): size the uavcan and TTY work queues for NuttX 12
NuttX 12.12 costs a few hundred bytes of stack on the paths that go through
the file layer, and that is more than these two defaults left spare. Measured
against the same PX4 tree on 10.3.0:
wq:uavcan 2860/3728 -> 3184/3736 (ARK FMU-v6X, two DroneCAN GNSS nodes)
wq:ttyS4 1088/1704 -> 1432/1712 (ARK FMU-v6XRT, crsf_rc)
load_mon reports wq:ttyS4 low on stack with 280 bytes left, and wq:uavcan sits
at 85 % of its own. 4096 and 2048 restore the margin the 10.3.0 figures had.
* feat(sensors): decouple sensor internal/external classification from the bus
A sensor's internal/external classification was derived from its bus
(px4_i2c_bus_external / px4_spi_bus_external), but a bus is a wire: one bus
routinely serves both chips soldered on the FC and a pinned-out connector, and
any onboard sensor on such a shared bus was classified external. That hands an
onboard mag an operator-settable rotation instead of the board rotation, feeds
an onboard baro's self-heated die temperature into air density as if it were
ambient, and inverts the 75/50 default priority. FMU-v6C and AirBrainH743
worked around it with hand-rolled device_id whitelists behind
BOARD_OVERRIDE_I2C_DEVICE_EXTERNAL; FMU-v6XRT's second onboard baro was simply
misclassified.
Classify the device instead of the bus: drivers publish is_external in
sensor_accel/gyro/mag/baro, derived from the device id by default and
overridden by the new -O start flag ("onboard") for onboard sensors that share
a bus with an external connector. -I/-X stay pure bus probe filters.
The question "is this sensor external" had four answers. It now has one, with
a single override point:
-O -> I2CSPIDriverConfig::external -> Device::set_external() and the
PX4* wrappers -> is_external in the sensor topic
px4_i2c_device_external() was px4_i2c_bus_external() with a device id decode in
front, and calibration::DeviceExternal() forwarded to device_is_external();
both are gone. Device::external() stops being a bus query: it is non-virtual,
defaults to device_is_external(), and takes the declared value through
set_external(), so the five I2C/SPI overrides that used to answer it from the
bus are deleted and cannot diverge again. device::I2C and device::SPI set it
from the config, so every driver built on the bus framework follows -O without
doing anything. px4_i2c_bus_external() and px4_spi_bus_external() survive as
what they honestly are - bus predicates - reachable only through the fallback.
-O is opt-in per driver (BusCLIArguments::support_onboard), the same way -k is:
a flag that every driver advertised but only a handful honoured would be a
silent no-op on the rest. This also drops the special case for the mcp23009 and
mcp23017 GPIO expanders, which use -O for their output state and simply do not
opt in.
sensor_gyro_fifo carries is_external too. VehicleAngularVelocity prefers the
FIFO topic for any IMU that publishes one, so without it the rate controller
would take its rotation from the bus while VehicleIMU took it from the topic -
the same chip, two classifications.
The BOARD_OVERRIDE_I2C_DEVICE_EXTERNAL hook is removed along with both board
implementations, replaced by -O on the affected rc.board_sensors lines.
FMU-v5x, MR-CANHUBK3, KakuteF7, NXT-Dual and MicoAir H743-Lite each start an
onboard barometer on an external bus and get the flag as well.
Assisted-by: Claude:claude-fable-5, Claude:claude-opus-5[1m]
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
* feat(sensors): treat shared buses as first-class topology
A boolean internal/external flag cannot describe a bus that carries both
hard-mounted chips and a connector. -O was an opt-in override for that
case, so most drivers silently ignored it and classification still
followed the bus.
Declare Internal / External / Shared on the bus, probe External and
Shared, and classify each sensor from -I/-s vs -X/-S.
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
* fix(ci): drop stale ITCM symbols and clang-tidy errors
ITCM lists still named calibration::DeviceExternal and
px4_spi_bus_external after both were removed. The host test stubs
forwarded varargs in a way the analyzer rejected, and stripped I2C
headers kept extra trailing newlines.
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
* fix(sensors): inline device_is_external into the header
Reaching it through drivers__device pulled the kernel-only library into
the userspace image of a protected build; giving it a library of its own
put an archive referencing px4_i2c_buses/px4_spi_buses after the board
library that defines them. Inlining sidesteps both.
* fix(sensors): compile tcbp001ta and probe canhubk3 GPS mag
tcbp001ta is not an I2CSPIDriver, so config.external does not exist; it
only ever starts on internal SPI. -X on canhubk3 I2C2 never ran because
that bus is Internal.
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
---------
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
hrt_elapsed_time() was a static inline expanding to ~15 bytes at each
of its ~450 call sites (call to hrt_absolute_time plus 64-bit compare
and subtract). Move the definition into px4_platform so each call site
is a single branch. px4iofirmware does not link px4_platform, so it
compiles the new source directly.
Saves 3592 bytes of flash on px4_fmu-v6x_default.
Co-authored-by: Balduin <balduin@auterion.com>
* refactor(px4_work_queue): out-line ScheduledWorkItem constructor to save flash
The header-inline constructor zero-initialises the hrt_call member at
every derived work-item constructor across ~150 classes. Move the
definition (verbatim) into ScheduledWorkItem.cpp next to the
destructor; the derived constructors already pay a call into the base
constructor chain, so this only removes the duplicated inline stores.
Saves 1376 B of .text on px4_fmu-v6x_default.
Signed-off-by: Balduin <balduin@auterion.com>
* refactor(px4_work_queue): out-line WorkItem::ScheduleNow to save flash
The body was inlined at ~240 call sites; a plain call is smaller at
every one of them.
Saves 672 B FLASH on px4_fmu-v6x_default.
Signed-off-by: Balduin <balduin@auterion.com>
---------
Signed-off-by: Balduin <balduin@auterion.com>
Co-authored-by: Balduin <balduin@auterion.com>
The constructor is inlined into every I2C/SPI driver's command-line
entry point, and with it the ~15 default member initializers and the
32-byte _options zero-fill, costing ~50-70 B per driver. Move the
definition (verbatim) into i2c_spi_buses.cpp, which already holds the
rest of the CLI parsing code. Purely cold-path (driver start/CLI).
Saves 2456 B of .text on px4_fmu-v6x_default.
Signed-off-by: Balduin <balduin@auterion.com>
Co-authored-by: Balduin <balduin@auterion.com>
Every Param<> constructor inlines param_set_used(handle()) followed by
param_get(handle(), &_val), materialising the handle constant twice at
each of the ~2000 instantiation sites. Fold both into a single
out-of-line param_get_mark_used() so each constructor emits one call.
The failsafe_web build stubs the param API itself, so it gets a matching
stub. The smaller constructors let the compiler fully inline the
defaulted FlightTaskDescend constructor on the ITCM boards, so its entry
is dropped from their linker scripts.
Saves 7232 B of .text on px4_fmu-v6x_default.
Assisted-by: Claude:claude-fable-5
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
Co-authored-by: Balduin <balduin@auterion.com>
Both were defined inline in the header, so the List add/remove bodies
were duplicated into the constructor and destructor of every one of the
150+ ModuleParams-derived classes. Construction and destruction are cold
paths; a call is smaller at each site.
Two consumers reach ModuleParams without px4_platform on the link line
and now need the definitions explicitly: health_and_arming_checks (the
functional-ModeManagement test only pulls it in transitively through
modules__commander, after px4_platform) and the failsafe_web emscripten
build, which compiles module_params.cpp directly.
The smaller constructors let the compiler fully inline the defaulted
FlightTaskDescend constructor on the ITCM boards, so its entry is
dropped from their linker scripts.
Saves 3968 B of .text on px4_fmu-v6x_default.
Assisted-by: Claude:claude-fable-5
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
Co-authored-by: Balduin <balduin@auterion.com>
board_query_manifest walked the list for mft[ndx].id == id then
returned &mft[id]. On a sparse list that is the wrong row, and for
PX4_MFT_T100_ETH (id 7) it is out of bounds.
Signed-off-by: Jacob Dahl <dahl.jakejacob@gmail.com>
Make px4::atomic's value constructor and ModuleBase::Descriptor
constexpr so the ~40 per-module 'desc' statics are constant-initialised
instead of each emitting static-init code, and apply the same to the
mixer_module FunctionProvider table.
Signed-off-by: Balduin <balduin@auterion.com>
The ~PublicationBase() destructor was header-inline, so its null-check
plus the orb_get_queue_size()/unadvertise() branch was emitted at every
publication destruction site (PX4 links with bfd ld, no ICF, no LTO).
Move it into Publication.cpp alongside the already out-lined
advertise()/publish(), mirroring #27581.
Saves 2.4 kB of flash on px4_fmu-v6x_default.
publish(), advertise() and get_instance() were instantiated per topic
type and the advertise-check in publish() was additionally inlined at
every call site. Move them into a type-independent PublicationMultiBase
compiled once.
Saves 5.5 kB of flash on px4_fmu-v6x_default.
process_add_subscription() (CONFIG_ORB_COMMUNICATOR only) read _data_valid and
the topic buffer contents while passing a pointer into _data straight to
send_message(), with no synchronization against a concurrent publisher. A
publisher's write() memcpy's into _data and sets _data_valid under ATOMIC_ENTER,
so the remote-init send can read a torn/partially-updated sample - most clearly
for o_queue == 1, where the read slot ((gen - 1) % 1) is exactly the slot being
written.
This is a pre-existing latent race (the previous px4::atomic<uint8_t*> _data only
guarded the pointer, never the bytes) that neither ASan nor TSan can catch because
CONFIG_ORB_COMMUNICATOR is not built in those configs. On VOXL2 it runs on the
muorb RX thread concurrently with publishers.
Snapshot the most recent sample into a temporary buffer under ATOMIC_ENTER, then
send it outside the critical section - mirroring copy(). The buffer is allocated
before the critical section (it cannot be allocated under ATOMIC_ENTER on NuttX)
and send_message() must run outside it anyway, since it can be slow and may call
back into DeviceNode. This only affects the rare subscription-add path, not the
publish hot path.
Signed-off-by: Julian Oes <julian@oes.ch>
_data (the lazily-allocated topic buffer) was a px4::atomic only to make the
double-checked-locking publish safe against concurrent readers. It is allocated
under the node lock() and written exactly once, so reading it under ATOMIC_ENTER
- which is that same lock() on POSIX - serializes the subscriber side against the
allocation without needing an atomic. On NuttX it is a set-once, word-aligned
pointer. The copy() guard is moved inside ATOMIC_ENTER accordingly. (Suggested in
review.)
DeviceNode::open() took the cdev lock only to call mark_as_advertised(), but
_advertised is a px4::atomic, so the single store needs no lock. (Suggested in
review.)
_interval_us and _last_update are read in SubscriptionCallbackWorkItem::call()
(and SubscriptionBlocking::call()) on the publishing thread, while the subscriber
sets them - e.g. from mixer_module's setMaxTopicUpdateRate() on a reconfigure.
That is a genuine cross-thread access that ThreadSanitizer flags.
Instead of making these fields atomic in the base SubscriptionInterval - which is
used single-threaded by 100+ subscriptions - parameterize the class on whether the
interval / last-update storage is atomic and expose two aliases:
SubscriptionInterval plain, single-threaded (the common case)
SubscriptionIntervalAtomic atomic, for callback subscriptions
SubscriptionCallback - and hence all callback subscriptions
(SubscriptionCallbackWorkItem, SubscriptionBlocking, CallbackHandler) - uses the
atomic variant; every other subscription stays plain with no synchronization
overhead.
The updated()/update()/copy() bodies stay out-of-line with explicit instantiation
of both variants, so they are not inlined at every call site (flash size).
The publisher fires callbacks via SubscriptionCallbackWorkItem::call(), which read
the subscriber's _last_generation (and called updated(), which reads it again) to
decide whether to schedule the work item. That made the subscriber's read cursor a
cross-thread field - publisher reads it while the subscriber writes it in copy() -
which TSan flagged and which we had wrapped in an atomic.
Instead, hand the publishing node's generation into call() and let the callback keep
its own _last_scheduled_generation cursor for the count gate (_required_updates).
call() runs under the node lock, so that cursor is only ever touched there and is
serialized - no atomic needed, and using the real generation (not a plain counter)
means a coalesced/missed call() does not drift the batch. The redundant updated()
generation check is dropped (call() only runs right after a publish, so it is always
true); the interval gate (_last_update / _interval_us) is kept unchanged.
With the publisher no longer reading it, _last_generation reverts to a plain unsigned
(subscriber-thread only). Behaviour is unchanged: the count and interval throttles
fire at the same rates (verified via work_queue status - gyro_fft 62.5 Hz,
flight_mode_manager 50 Hz) and TSan stays clean.