Add add-standard-as-possible version of C++11 <mutex>.
A lot of the stuff in there is generic, but the actual mutex classes and
call_once need to interface with the OS.
For those, they're not available in ARMC5 or IAR; retargetting would be
necessary for ARMC6 and GCC, and I've yet to investigate how whether
that's possible. So for now I'm using local implementations.
Although `Mutex` in principle could support `timed_mutex` and
`recursive_timed_mutex`, we don't have `chrono` for the time parameters,
so hold off for now.
For the generic stuff like mstd::unique_lock, they are aliased to
std::unique_lock where possible.
* Adjust definition to make the default constructor `constexpr`.
This permits use in classes that want lazy initialization and their
own `constexpr` constructor, such as `mstd::mutex`.
* Add `get_no_init()` method to allow an explicit optimisation for
paths that know they won be the first call (such as
`mstd::mutex::unlock`).
* Add `destroy()` method to permit destruction of the contained object.
(`SingletonPtr`'s destructor does not call its destructor - a cheat
to omit destructors of static objects). Needed if using in a class
that needs proper destruction.
Regularise the C++ support glue, adding `<mstd_type_traits>` etc.
These include the base toolchain file, backfill `namespace std` as much
as possible for ARM C 5, and then arrange to create unified
`namespace mstd`, which can incorporate toolchain bypasses
(eg `mstd::swap` for ARM C 5, or `mstd::atomic`), and include local
implementations of post-ARM C++14 stuff.
All APIs in `namespace mstd` are intended to function as their
`namespace std` equivalent, and their presence there indicates they
are functional on all toolchains, and should be safe to use in
an Mbed OS build (including not unreasonable memory footprint).
See README.md for more info.
For all compilers add post-C++14 stuff to namespace mbed:
* invoke, invoke_result, is_invocable, is_invocable_r,
is_nothrow_invocable, is_nothrow_invocable_r, unwrap_reference,
unwrap_ref_decay
For ARM C 5, add C++11 bits based on above to namespace std:
* result_of, reference_wrapper, ref, cref, mem_fn
Revert back to older behaviour where we hold deep sleep lock only while
timing a sleep. Previous version was a speed optimisation, but broke
some tests.
Add a C++ `Atomic<T>` template to make atomics even easier. Basically
compatible with C++11 `std::atomic<T>`, but using the underlying
`core_util_atomic_xxx` functions from mbed_atomic.h, so appropriate for
synchronising with interrupts and optimised for uniprocessor.
One extra piece of functionality beyond the `core_util_atomic_xxx`
functions is the ability to have an arbitrary atomic type - eg a small
structure with 2 `uint16_t`s can be stored in a `uint32_t` container.
Avoid template ambiguities using type_identity_t.
Previously the compiler would be unable to figure out whether
uint8_t x;
core_util_atomic_store(&x, 0);
should invoke core_util_atomic_store<uint8_t>, matching the pointer
type, or core_util_atomic_store<int>, matching the value, leading to
an ambiguity error.
Templates now select only on the type of the atomic pointer parameter.
Even though C/C++11 don't offer pre-op forms (that do the operation then
return the new value) of their freestanding functions, we will be making
them visible via pre-op operators on `Atomic<T>` (++, --, +=, -=, &=,
|=, ^=).
It is easier to do a pre-op than a post-op, as we can use one
fewer register in the assembler, so it's worth optimising for what will
be quite common cases. Make these forms accessible for `Atomic<T>`, but
don't document them for standalone use at this stage.
When the function `error` is called without ending in a newline
greentea will not display the error. This is because greentea is
performing line buffering.
This patch ensures that all calls to `error` end with a newline. This
is done by adding an additional newline to end end of the string.
Example code exhibiting this problem:
error("This function is not supported.");
Greentea output before this change:
++ MbedOS Error Info ++
Error Status: 0x80FF0100 Code: 256 Module: 255
Error Message: Fatal Run-time error
Location: 0x1001766D
Error Value: 0x0
Current Thread: main Id: 0x8003FD8 Entry: 0x1000E64F StackSize: 0x1000 StackMem: 0x8004AE8 SP: 0x8005898
For more info, visit: https://mbed.com/s/error?error=0x80FF0100&tgt=target
-- MbedOS Error Info --
test suite run finished after 240.50 sec...
Greentea output after this change:
++ MbedOS Error Info ++
Error Status: 0x80FF0100 Code: 256 Module: 255
Error Message: Fatal Run-time error
Location: 0x1001766D
Error Value: 0x0
Current Thread: main Id: 0x8003FD8 Entry: 0x1000E64F StackSize: 0x1000 StackMem: 0x8004AE8 SP: 0x8005898
For more info, visit: https://mbed.com/s/error?error=0x80FF0100&tgt=target
-- MbedOS Error Info --
This function is not supported.
test suite run finished after 240.50 sec...
Static keyword causes problems when trying to use force-inlined
functions from normal inlined functions. This is not legal:
static inline void forced() { }
inline void normal() { forced(); }
You cannot reference internal-linkage things from external-linkage
inline functions.
Removal of the static implies that in C there would need to be a
non-inline definition in case anyone calls it non-inlined, but if the
force attribute is doing its job, that should not happen.
Only significant in-tree user of the MBED_FORCEINLINE macro is
the atomic operations - making this change permits atomic operations
from non-static inline functions.
As we start trying to use new facilities, we're likely to need some more
helpers.
In particular, ARM C 5 has no C++11 support in its library at all, so
to avoid totally breaking it we need some backup.
For the other toolchains, we can add a few C++17/C++20/TS extensions
into namespace mbed to make life a little easier.
* For ARM C 5: C++14 type_traits subset, std::move, std::forward,
std::array, std::initializer_list, std::begin, std::end,
std::align, std::maxalign_t, std::aligned_storage,
alignof + alignas macro replacements.
* For ARM C 5: MBED_CONSTEXPR_FN_14 and MBED_CONSTEXPR_OBJ_14 to
mark things that can only be constexpr in C++14 or later.
* For other compilers: mbed::void_t, mbed::type_identity,
mbed::conjunction, mbed::disjunction, mbed::negation,
mbed::experimental::nonesuch, mbed::experimental::is_detected family,
mbed::remove_cvref, mbed::as_const.
The bulk of Callback.h was auto-generated as 6 specialisations, handling
zero to five arguments.
This can now be handled without specialisation using C++11 variadic
templates, reducing the file from 4,900 lines to 900 lines.
This should reduce compilation time, and offset potential increases from
use of `<type_traits>` or a local `mbed_cxxsupport.h` equivalent.
Several other improvents to `Callback` are possible and/or desirable
with C++11, such as the ability to store lambdas, but this commit is
purely the variadic simplification.
Newer language standards have standard forms for `MBED_NORETURN` and
`MBED_ALIGN` attributes. Use them when available.
C++14 also adds `[[deprecated]]`, but as it needs to go in the middle of
structure definitions as `class [[deprecated]] MyClass`, it's not a
total drop-in-replacemend for `MBED_DEPRECATED`, so that is not
attempted here.
Using standard forms increases the chances that code analysis tools such
Coverity will recognise them - particularly important for "no return".
The `int : 0` bitfield this produced could force integer alignment onto
the structure it was placed in, making a structure that should be 1 byte
be 4 bytes.
Change `int` to `bool` to minimise alignment impact - should be to
nothing.
Alignment/size problem was revealed in a `sizeof` check in an
`Atomic<uint8_t>` test.
M23 doesn't implement Main Extension. So like M0/M0+, these registers HFSR/
MMFSR/BFSR/UFSR/DFSR are not present on M23. Remove access to them in mbed
fault handler for M23 targets.
Use `+ 1` to set the "Thumb" indicator on the `delay_loop` routine
address, rather than `| 1`. That makes it something that can be done
by the linker, rather than needing to be done at run-time.
Saves one instruction and one cycle.
As the timer code became more generic, coping with initialization on
demand, and variable width and speed us_ticker_api implementations,
wait_us has gradually gotten slower and slower.
Some platforms have reportedly seen overhead of wait_us() increase from
10µs to 30µs. These changes should fully reverse that drop, and even
make it better than ever.
Add fast paths for platforms that provide compile-time information about
us_ticker. Speed and code size is improved further if:
* Timer has >= 2^32 microsecond range, or better still is 32-bit 1MHz.
* Platform implements us_ticker_read() as a macro
* Timer is initialised at boot, rather than first use
The latter initialisation option is the default for STM, as this has
always been the case.
Minor tweaks to fix ARM C 5 compatibility.
Pushing "ns_list.h" include to first makes sure "ns_types.h" is included
first, meaning it gets to define `__STDC_LIMIT_MACROS` before the first
include of <stdint.h>, which ensures that UINT8_MAX etc are defined.
- Changed the process into static method
- used the singletonptr for creating the low power ticker instance
- Added the mbed stub into cmake build for cellularnonipsocket,loramacrypto
-Added the mock class function to mock mbed_assert_internal
-Added the unit test case to test start,kick,stop
-Modified the interface api name from is_alive to process
-added the unit test cases for process
-SW watchdog has interface name start(),stop(),kick() Sw watchdog internally has static list and shared across multiple instance of SW watchdog
- Sw watchdog initialize timeout value,unique string via constructor whenever threads created sw watchdog object
-Threads make sure pass proper timeout value,Unique string while creating the instance.
-start() called by components(BLE,WIFI etc.,),it adds the entry into static list with few details current count ,etc.,
-kick() called by registered components(BLE,WIFI etc.) to reset current count to zero.
-is_alive - interface API to mbed_watchdog_manager
-implementation optimization
- mbed_watchdog_mgr has interface name mbed_wdog_manager_start(),mbed_wdog_manager_stop(),mbed_wdog_manager_kick()
- HwWatchdog is going to attach with LowPowerTIcker for periodic callback functionality
- mbed_wdog_manager_start() will either get start either by BL/RTOS Aps,it reads the timeout value specified via macro and macro gets defined in target.json file.
- mbed_wdog_manager_start() internally configure below HAL hw watchdog with timeout specified via target.json
- mbed_wdog_manager_start() internally divide the timeout(specified in target.json) by the 2 and attach LowPowerTicker with periodic callback of hw_kick()
- mbed_wdog_manager_start() internally create one instance of sw watchdog class,to access the static list data structure of sw watchdog class
- mbed_wdog_manager_kick() function periodically get called and refresh the hw watchdog to avoid watchdog reset
- converted C++ code into C based APIs
- added boolean to control watchdog start and stop
- Added detach from ticker on stop API
In case the CThunk object is deleted without having called the
entry() function (and thus _entry is NULL), cthunk_free_real()
will fail with an assertion.
Correct C++14 operation of the alloc wrappers requires us to define
custom sized delete operators.
Their presence won't cause any problem for people compiling as C++03 or
C++11.
Add unified syntax directives to make the atomic assembler work when GCC
is building for M23.
GCC actually uses unified syntax when compiling C code, but puts
`.syntax divided` before each piece of inline assembly when targetting
Thumb-1 type devices like M0 and M23 for backwards compatibility. We can
overcome this with our own `.syntax unified`. The command-line option
`-masm-syntax-unified` intended to override this globally has been
broken from GCC 6 to 8.0.
Some targets may not provide zero-wait-state flash performance. Export this function
to be overridable for targets to provide more accurate implementation like locating
'delay_loop_code' in SRAM.
Imports working McuBoot for reset.
Updates microsec ticker driver.
Default baudrate is set to 115200 to see TF-M boot messages.
Stack top is set to scatter file dependent and not hard-coded.
Volatile makes no real difference when we're using assembler, or locked
functions, but leaving it off could be more efficient for the basic
loads and stores. So add non-volatile overloads in C++ for them.
Reimplement atomic code in inline assembly. This can improve
optimisation, and avoids potential architectural problems with using
LDREX/STREX intrinsics.
API further extended:
* Bitwise operations (fetch_and/fetch_or/fetch_xor)
* fetch_add and fetch_sub (like incr/decr, but returning old value -
aligning with C++11)
* compare_exchange_weak
* Explicit memory order specification
* Basic freestanding template overloads for C++
This gives our existing C implementation essentially all the functionality
needed by C++11.
An actual Atomic<T> template based upon these C functions could follow.
Once a fatal error is in progress, it's not useful to trap RTX errors
or mutex problems, so short-circuit the checks.
This makes it more likely that we may be able to get the console
initialised if it is being written to for the first time by `mbed_error`
in a difficult context - such as an RTX error callback from inside an
SVCall.
For example, the one-line program
osMutexAcquire(NULL, 0);
will generate an RTX error trap, then `mbed_error` will try to call
`write(STDERR_FILENO)` to print the error, which will prompt mbed_retarget to
construct a singleton `UARTSerial`. This would trap in the mutex
for the singleton or the construction of the UARTSerial itself, if
we didn't allow this leniency. If we clear the mutex checks, then
`UARTSerial::write_unbuffered` will work.
User uses of `MBED_MAKE_ERROR` assemble a new error that gets its bottom
16 bits from an existing negative 32-bit error code. This lead to
undefined behaviour when `<<` was used on it - even though it was a
shift by zero!
Avoid the undefined behaviour warning from Clang by masking before
shifting.
In rtos-less code, heap is defined by assuming one-region. Through weak-reference to
ARM_LIB_HEAP, heap definition is fixed if ARM_LIB_HEAP is defined.
Please note the heap address of the both the banks must not be contigious else
GCC considers it to be single memory bank and does allocation across the banks,
which might result into hard-fault
Assert failure took a critical section before calling `mbed_error`.
There's no need to take a critical section on assert failure -
mbed_error does not do this, and is designed to operate from normal
contexts.
Avoiding the critical section will improve the chances of console
initialisation due to assert failure working nicely.
New `target.console-uart` option added to indicate whether a target has
a console UART on STDIO_UART_TX/RX/RTS/CTS pins. (The existing option
`target.console-uart-flow-control` indicates whether RTS and or CTS is
available in addition to TX and RX).
The option defaults to true, and is currently true on all platforms. It
only applies if DEVICE_SERIAL is true, so no need to go through and mark
it false for non-SERIAL platforms.
An application can turn off target.console-uart to save ROM/power/etc if
they don't want to use the serial console. If this is turned off, the
console won't be activated for stdin/stdout, but the application is
still free to open `UARTSerial(STDIO_UART_TX, STDIO_UART_RX)`
themselves.
This provides the ability to generate really small delays - it's often
the case that wait_us() takes multiple microseconds to set up, so
having an alternative suitable for <10us delays is useful.
There have been a few local implementations - it makes sense to
centralise them as they need retuning for each new ARM core.
Based on the local implementation inside the Atmel 802.15.4 driver.
Add a necessary helper to allow FileHandle objects to be obtained
from POSIX file descriptors.
Primary envisaged use case is to act on STDIN_FILENO etc, eg to
set it non-blocking or use sigio, or to use the enable API.
Add API to dynamically enable/disable input and output on FileHandles.
Aim is to allow power saving by indicating that we permanently or
temporarily do not require input, so that for example serial Rx
interrupts can be released, and deep sleep permitted.
Memory model for RTOS and No RTOS was initially single stack and heap,
only few targets implemented 2-region RAM model.
2-region RAM model is applied for all toolchains and targets.
GCC: __wrap__sbrk was implemented for 2-region ram model, with switch to 2-region
for all targets, we do not need target specific implementation of this API
Also _sbrk is WEAK function, hence can be over written in target folder for
special cases
The following commits: #8039, #9092 added Boot/ISR stack definition to all scatter files (ARM_LIB_STACK).
This has changed memory model for RTOS-less builds to 2-region memory model and caused failure in case of rtos less builds.
This PR defines valid heap/stack regions for rtos-less builds.
Fix the following build warning
Compile [ 83.4%]: mbed_error.c
[Warning] mbed_error.c@71,21: 'compute_crc32' defined but not used [-Wunused-function]
compute_crc32 usage is guraded with #define, but not the definition.
Use the same #define around the definition of compute_crc32()
SharedPtr::reset did not actually set the stored pointer value.
Correct this, with other minor tidies:
* Ensure counter is set to NULL if pointer is reset to NULL
* Be consistent about not clearing pointers in decrement_counter().
->move() operator was not touching unused data fields, therefore
leaving uninitialised data and failing the comparison.
Fixed by initialising all fields to zero before moving.
mbed_die was calling wait_ms in a critical section - this is deprecated
behaviour, and caused a fatal error in debug builds, potentially leading
to an infinite reboot loop if this was caused due to a reboot limit
halt.
Switch to using wait_us - this is safe in a critical section. This does
trigger a call to mbed_warning, due to the large parameter, but that is
harmless - it doesn't output anything to the console, and won't
overwrite the error context if it already contains something.
The pre-main check of error_reboot_count was applied repeatedly on
every boot, meaning that once the reboot limit was hit, every subsequent
reset would halt before main, until something managed to clear
or corrupt the error context.
Set the is_error_processed flag before halting, so that when an external
agent resets us while we're halted, we do not report the error and halt
again.
A couple of places in mbed_retarget.cpp were testing for either ARMC5 or
ARMC6 in a long-winded fashion. Testing for __ARMCC_VERSION being
defined is sufficient.
Similar to SingletonPtr, use atomic accesses when loading the guard word
outside the lock, and when storing, to ensure no races for threads that
don't take the lock.
Lack of atomics unlikely to be a problem in current builds, but code
could conceivably be subject to reordering if link-time optimisation was
enabled.
SingletonPtr's implementation wasn't totally safe - see "C++ and the
Perils of Double-Checked Locking"by Meyers and Alexandrescu. No problems
observed in practice, but it was potentially susceptible to compiler
optimisation (or maybe even SMP issues).
Now that we have atomic loads and stores, the function can be made safe,
avoiding any potential races for threads that don't take the lock:
ensure that the unlocked load is atomic, and that the pointer store is
atomic.
See https://preshing.com/20130930/double-checked-locking-is-fixed-in-cpp11/
for more discussion.
As barriers were added in #9247, review comments pointed out that atomic
stores needed a second barrier, and they got one. But atomic_flag_clear
was missed.
Add atomic load and store functions, and add barriers to the existing atomic
functions.
File currently has no explicit barriers - we don't support SMP, so don't
need CPU barriers.
But we do need to worry about compiler barriers - particularly if link time
optimisation is activated so that the compiler can see inside these
functions. The assembler or intrinsics that access PRIMASK for
enter/exit critical act as barriers, but LDREX, STREX and simple
volatile pointer loads and stores do not.
Critical section count/state variables are synchronised by IRQ disabling and
critical section calls themselves, so do not need to be volatile.
This eliminates a couple of unnecessary reads of the counter variable.
The DEVICE_FOO macros are always defined (either 0 or 1).
This patch replaces any instances of a define check on a DEVICE_FOO
macro with value test instead.
Signed-off-by: Alastair D'Silva <alastair@d-silva.org>
As part of work to improve the debugging of exceptions, have
Mbed OS make an effort to make exceptions more precise in debug builds
at the cost of performance.
Related pyOCD work:
https://github.com/mbedmicro/pyOCD/pull/430
Make the option positively named, and as it is a platform config
option make sure it only affects platform code.
HAL functions still remain available even if platform is told not
to use them.
MPU lock under- or overflow is not a likely random runtime failure - it
will be a mismatched lock/unlock programming error, so should be
detected during development. Make the checks asserts so they can be
left out from release builds. MBED_ASSERT is also a bit smaller than
MBED_ERROR in develop builds.
Kevin Bracey
Martin Kojtal
Arto Kinnunen
Russ Butler
Tymoteusz Bloch
Przemyslaw Stekiel
Chun-Chieh Li
Anna Bridge
Rajkumar Kanagaraj
Teijo Kinnunen
Steven Cartmell
Jaakko Korhonen
Gabor Kertesz
Marcus Chang
M. Rahimi
Deepika
Cruz Monrreal
Shrikant Tudavekar
d-kato
David Saada
Mahesh Mahadevan
Michael Schwarcz
Senthil Ramakrishnan
Jimmy Brisson
Sarah Marsh
Naveen Kaje
Veijo Pesonen
Seppo Takalo
paul-szczepanek-arm
kegilbert
bcostm
Alastair D'Silva
Russ Butler
Filip Jagodzinski