23.2. Background¶
23.2.1. Overview¶
The fatal error manager is called upon detection of an irrecoverable error condition by either RTEMS or the application software. Fatal errors are also used in case it is difficult or impossible to return an error condition by other means, e.g. a return value of a directive call. Fatal errors can be detected from various sources, for example
the executive (RTEMS),
support libraries,
user system code,
user application code, and
processor interrupts and exceptions (data abort, instruction prefetch errors, ECC errors, spurious interrupts, etc.).
RTEMS automatically invokes the fatal error manager upon detection of an error it considers to be fatal. Similarly, the user should invoke the fatal error manager upon detection of a fatal error.
Each user extensions set may include a fatal error handler. The fatal error handler in the initial extension sets can be used to provide access to debuggers and monitors which may be present on the target hardware. If any user-supplied fatal error handlers are installed, the fatal error manager will invoke them. Usually, the board support package provides a fatal error extension which resets the board. If no user handlers are configured or if all the user handler return control to the fatal error manager, then the CPU port provided idle loop executes.
23.2.2. System Termination Procedure¶
The _Terminate()
handler is invoked to terminate the system. It is
called by all services which determine that a system termination is required.
For example, it is called by all higher level directives which announce a fatal
error, see Announcing a Fatal Error.
The first action of the system termination handler is to disable maskable interrupts. This ensures that interrupts on this processor do not interfere with the system termination procedure. This reduces the likelihood to end up in a recursive system termination procedure.
The second action of the system termination handler is to call the fatal extensions of the user extensions.
The fatal extensions are called with three parameters:
the fatal source,
a legacy parameter which is always set to
false
, andan error code with a fatal source dependent content.
The fatal extensions of the initial extension sets are invoked first. For them, the following execution environment is required
a valid stack pointer and enough stack space,
a valid code memory, and
valid read-only data.
In uniprocessor configurations, the read-write data (including .bss
segment) is not required. In SMP configurations, however, the read-write data
must have been initialized to determine the state of the other processors and
request them to shut-down if necessary. The board support package (BSP) may
install an initial extension that performs a system reset. See the BSP
documentation in the RTEMS User Manual for more information how the system
reset is done. The BSP provided fatal extension can be disabled by the
CONFIGURE_DISABLE_BSP_SETTINGS application configuration option. It is
recommended to provide an application-specific fatal extension using the
CONFIGURE_INITIAL_EXTENSIONS application configuration option.
In certain error conditions, it may be unreliable to carry out the following steps of the termination procedure since the read-write data may be corrupt. One of the fatal extensions of the initial extension set should reset the system to stop the system termination procedure.
After invoking the fatal extensions of the initial extension sets, the fatal extensions of the dynamic extension sets are invoked. For this procedure valid read-write data is required.
The last action of the system termination handler is to execute the CPU port provided idle loop with maskable interrupts disabled. Please note, that properly configured applications should not reach this point.
23.2.3. Fatal Sources¶
The following fatal sources are defined for RTEMS via the
rtems_fatal_source
enumeration. Each symbolic name has the
corresponding numeric fatal source in parenthesis.
- INTERNAL_ERROR_CORE (0)
Errors of the core operating system. See Internal Error Codes.
- INTERNAL_ERROR_RTEMS_API (1)
Errors of the Classic API.
- INTERNAL_ERROR_POSIX_API (2)
Errors of the POSIX API.
- RTEMS_FATAL_SOURCE_BDBUF (3)
Fatal source for the block device cache. See
rtems_bdbuf_fatal_code
.- RTEMS_FATAL_SOURCE_APPLICATION (4)
Fatal source for application-specific errors. The fatal code is application-specific.
- RTEMS_FATAL_SOURCE_EXIT (5)
Fatal source of
exit()
. The fatal code is theexit()
status code.- RTEMS_FATAL_SOURCE_BSP (6)
Fatal source for BSP errors. The fatal codes are defined in
<bsp/fatal.h>
. Examples are interrupt and exception initialization. Seebsp_fatal_code
andbsp_fatal()
.- RTEMS_FATAL_SOURCE_ASSERT (7)
Fatal source of
assert()
. The fatal code is the pointer value of the assert context. Seertems_assert_context
.- RTEMS_FATAL_SOURCE_STACK_CHECKER (8)
Fatal source of the stack checker. The fatal code is the object name of the executing task.
- RTEMS_FATAL_SOURCE_EXCEPTION (9)
Fatal source of the exceptions. The fatal code is the pointer value of the exception frame pointer. See
rtems_exception_frame
and rtems_exception_frame_print.- RTEMS_FATAL_SOURCE_SMP (10)
Fatal source of SMP domain. See
SMP_Fatal_code
.- RTEMS_FATAL_SOURCE_PANIC (11)
Fatal source of
rtems_panic()
, see rtems_panic.- RTEMS_FATAL_SOURCE_INVALID_HEAP_FREE (12)
Fatal source for invalid C program heap frees via
free()
. The fatal code is the bad pointer.- RTEMS_FATAL_SOURCE_HEAP (13)
Fatal source for heap errors. The fatal code is the address to a heap error context. See
Heap_Error_context
.
23.2.4. Internal Error Codes¶
The following error codes are defined for the INTERNAL_ERROR_CORE
fatal source. Each symbolic name has the corresponding numeric error code in
parenthesis.
- INTERNAL_ERROR_TOO_LITTLE_WORKSPACE (2)
There is not enough memory for the workspace. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_THREAD_EXITTED (5)
A non-POSIX thread entry function returned. This is an API usage error.
An example code to provoke this fatal error is:
rtems_task task( rtems_task_argument arg ) { /* Classic API tasks must not return */ } void create_bad_task( void ) { rtems_status_code sc; rtems_id task_id; sc = rtems_task_create( rtems_build_name('T', 'A', 'S', 'K'), 1, RTEMS_MINIMUM_STACK_SIZE, RTEMS_DEFAULT_MODES, RTEMS_DEFAULT_ATTRIBUTES, &task_id ); assert( sc == RTEMS_SUCCESSFUL ); sc = rtems_task_start( task_id, task, 0 ); assert( sc == RTEMS_SUCCESSFUL ); }
- INTERNAL_ERROR_INCONSISTENT_MP_INFORMATION (6)
This fatal error can only occur on MPCI configurations. The MPCI nodes or global objects configuration is inconsistent. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_INVALID_NODE (7)
This fatal error can only occur on MPCI configurations. The own MPCI node number is invalid. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_NO_MPCI (8)
This fatal error can only occur on MPCI configurations. There is no MPCI configuration table. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_BAD_PACKET (9)
This fatal error can only occur on MPCI configurations. The MPCI server thread received a bad packet.
- INTERNAL_ERROR_OUT_OF_PACKETS (10)
This fatal error can only occur on MPCI configurations. The MPCI packet pool is empty. It is an application configuration error.
- INTERNAL_ERROR_OUT_OF_GLOBAL_OBJECTS (11)
This fatal error can only occur on MPCI configurations. The MPCI global objects pool is empty. It is an application configuration error.
- INTERNAL_ERROR_OUT_OF_PROXIES (12)
This fatal error can only occur on MPCI configurations. The MPCI thread proxy pool is empty. It is an application configuration error.
- INTERNAL_ERROR_INVALID_GLOBAL_ID (13)
This fatal error can only occur on MPCI configurations. The system cannot find the global object for a specific object identifier. In case this happens, then this is probably an operating system bug.
- INTERNAL_ERROR_NO_MEMORY_FOR_HEAP (23)
There is not enough memory for the C program heap. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_CPU_ISR_INSTALL_VECTOR (24)
The use of
_CPU_ISR_install_vector()
is illegal on this system.- INTERNAL_ERROR_RESOURCE_IN_USE (25)
This fatal error can only occur on debug configurations. It happens in case a thread which owns mutexes is deleted. Mutexes owned by a deleted thread are in an inconsistent state.
- INTERNAL_ERROR_RTEMS_INIT_TASK_ENTRY_IS_NULL (26)
An RTEMS initialization task entry function is NULL. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_THREAD_QUEUE_DEADLOCK (28)
A deadlock was detected during a thread queue enqueue operation.
- INTERNAL_ERROR_THREAD_QUEUE_ENQUEUE_STICKY_FROM_BAD_STATE (29)
This fatal error can only happen in SMP configurations. It is not allowed to obtain MrsP semaphores in a context with thread dispatching disabled, for example interrupt context.
An example code to provoke this fatal error is:
rtems_timer_service_routine bad( rtems_id timer_id, void *arg ) { rtems_id *sem_id; sem_id = arg; rtems_semaphore_obtain( *sem_id, RTEMS_WAIT, RTEMS_NO_TIMEOUT ); assert( 0 ); } rtems_task fire_bad_timer( rtems_task_argument arg ) { rtems_status_code sc; rtems_id sem_id; rtems_id timer_id; sc = rtems_semaphore_create( rtems_build_name('M', 'R', 'S', 'P'), 1, RTEMS_MULTIPROCESSOR_RESOURCE_SHARING | RTEMS_BINARY_SEMAPHORE, 1, &sem_id ); assert( sc == RTEMS_SUCCESSFUL ); sc = rtems_timer_create( rtems_build_name( 'E', 'V', 'I', 'L' ), &timer_id ); assert( sc == RTEMS_SUCCESSFUL ); sc = rtems_semaphore_obtain( sem_id, RTEMS_WAIT, RTEMS_NO_TIMEOUT ); assert( sc == RTEMS_SUCCESSFUL ); sc = rtems_timer_fire_after( timer_id, 1, bad, &sem_id ); assert( sc == RTEMS_SUCCESSFUL ); rtems_task_wake_after( 2 ); assert( 0 ); }
- INTERNAL_ERROR_BAD_THREAD_DISPATCH_DISABLE_LEVEL (30)
It is illegal to call blocking operating system services with thread dispatching disabled, for example in interrupt context.
An example code to provoke this fatal error is:
void bad( rtems_id id, void *arg ) { rtems_task_wake_after( RTEMS_YIELD_PROCESSOR ); assert( 0 ); } void fire_bad_timer( void ) { rtems_status_code sc; rtems_id id; sc = rtems_timer_create( rtems_build_name( 'E', 'V', 'I', 'L' ), &id ); assert( sc == RTEMS_SUCCESSFUL ); sc = rtems_timer_fire_after( id, 1, bad, NULL ); assert( sc == RTEMS_SUCCESSFUL ); rtems_task_wake_after( 2 ); assert( 0 ); }
- INTERNAL_ERROR_BAD_THREAD_DISPATCH_ENVIRONMENT (31)
In SMP configurations, it is a fatal error to call blocking operating system with interrupts disabled, since this prevents delivery of inter-processor interrupts. This could lead to executing threads which are not allowed to execute resulting in undefined system behaviour.
Some CPU ports, for example the ARM Cortex-M port, have a similar problem, since the interrupt state is not a part of the thread context.
This fatal error is detected in the operating system core function
_Thread_Do_dispatch()
responsible to carry out a thread dispatch.An example code to provoke this fatal error is:
void bad( void ) { rtems_interrupt_level level; rtems_interrupt_local_disable( level ); rtems_task_suspend( RTEMS_SELF ); rtems_interrupt_local_enable( level ); }
- INTERNAL_ERROR_RTEMS_INIT_TASK_CREATE_FAILED (32)
The creation of the RTEMS initialization task failed. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_POSIX_INIT_THREAD_CREATE_FAILED (33)
The creation of the POSIX initialization thread failed. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_LIBIO_STDOUT_FD_OPEN_FAILED (36)
Open of the standard output file descriptor failed or resulted in an unexpected file descriptor number. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_LIBIO_STDERR_FD_OPEN_FAILED (37)
Open of the standard error file descriptor failed or resulted in an unexpected file descriptor number. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_ILLEGAL_USE_OF_FLOATING_POINT_UNIT (38)
The floating point unit was used illegally, for example in interrupt context on some architectures.
- INTERNAL_ERROR_ARC4RANDOM_GETENTROPY_FAIL (39)
A
getentropy()
system call failed in one of the ARC4RANDOM(3) functions. This fatal error can only be fixed with a different implementation ofgetentropy()
.- INTERNAL_ERROR_NO_MEMORY_FOR_PER_CPU_DATA (40)
This fatal error may happen during workspace initialization. There is not enough memory available to populate the per-CPU data areas, see <rtems/score/percpudata.h>.
- INTERNAL_ERROR_TOO_LARGE_TLS_SIZE (41)
This fatal error may happen during system initialization. The actual thread-local storage (TLS) size of the application exceeds the configured maximum, see CONFIGURE_MAXIMUM_THREAD_LOCAL_STORAGE_SIZE. You can get the thread-local storage size of an application using the RTEMS tool
rtems-execinfo
.- INTERNAL_ERROR_RTEMS_INIT_TASK_CONSTRUCT_FAILED (42)
The construction of the RTEMS initialization task failed. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_IDLE_THREAD_CREATE_FAILED (43)
The creation of an IDLE task failed. This fatal error may occur during system initialization. It happens if a task create extension fails for an IDLE task.
- INTERNAL_ERROR_NO_MEMORY_FOR_IDLE_TASK_STORAGE (44)
There was not enough memory available to allocate an IDLE task stack. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_IDLE_THREAD_STACK_TOO_SMALL (45)
The task stack size of an IDLE task would have been less than the configured stack size for IDLE tasks, see CONFIGURE_IDLE_TASK_STACK_SIZE. This fatal error may occur during system initialization. It is an application configuration error.
- INTERNAL_ERROR_CANNOT_DISABLE_DATA_CACHE (46)
This fatal error may be caused by rtems_cache_disable_data() if the data cache cannot be disabled for a particular target or configuration. The data cache may be necessary to provide atomic operations. In SMP configurations, the data cache may be required to ensure data coherency. See the BSP documentation in the RTEMS User Manual for more information.