Since pthread_once() is both the initializer and executor of
pthread_once_t, it can have maximally two states. Since the
implementation in Zephyr previously aimed to maximize libc
compatibility, we opted to use the definition of pthread_once_t
from newlib, which is a structure with two ints.
It does not make sense to use 64 bits to manage 2 possible
states. The control for that should effectively be a bool.
We maintain compatibility with newlib by asserting (at build
time), that newlib's pthread_once_t is larger than Zephyr's
new struct pthread_once (which just contains a bool).
This allows us to delete the non-standard pthread_key.h
header file (finally).
Reuse the pthread_pool_lock in order to synchronize the related
init function (so that it is only called maximally once from any
thread). The spinlock is only used to test the state and the
init function is not called with the spinlock held.
The alternative was to use an atomic inside of
struct pthread_once. But again, that could be up to 64-bits with
Zephyr's atomics implementation.
Ideally we would use C11 generics or something to support atomics
on 8, 16, 32, and 64-bit primitives.
Lastly, also update declarations for C11 threads as they mostly
mirror our pthread implementation.
This needed to be done as a single commit in order to ensure
continuity of build.
Signed-off-by: Christopher Friedt <cfriedt@meta.com>
This moves the k_* memory management functions from sys/ into
kernel/ includes, as there are kernel public APIs. The z_*
functions are further separated into the kernel internal
header directory.
Also made a quick change to doxygen to group sys_mem_* into
the OS Memory Management group so they will appear in doc.
Signed-off-by: Daniel Leung <daniel.leung@intel.com>
This change capitalizes on newly added support for dynamic
thread stacks and the existing pthread support to provide
an implementation of the ISO C11 `<threads.h>` API.
Signed-off-by: Christopher Friedt <cfriedt@meta.com>
This prevents the compiler from optimizing calloc into an
infinite recursive call.
For example a call to malloc + memset zero at GCC -O2 will be
replaced by a call to calloc. This causes infinite recursion
if the function being implemented *is* calloc.
fno-builtin-malloc forces the compiler to avoid this optimization.
Signed-off-by: Grant Ramsay <gramsay@enphaseenergy.com>
Leave the malloc partition so that it only contains the heap itself; this
lets the initialization code adjust the address range when configuring the
arena at startup.
Signed-off-by: Keith Packard <keithp@keithp.com>
When multithreading is disabled, the whole mutex infrastructure isn't
available. The common malloc code wasn't checking for this case which
caused build failures.
Signed-off-by: Keith Packard <keithp@keithp.com>
Tested-by: Detlev Zundel dzu@member.fsf.org
Memalign is another name for the posix aligned_alloc function, although it
has weaker restrictions on the relationship between the alignment and size.
memalign() is used internally by the libstdc++ when built for 'newlib'
targets (which includes picolibc) instead of aligned_alloc() due to a bug
in gcc, so we need to provide an implementation of this when using that
library, even though it's not part of the Zephyr C library API.
When a fix for the libstdc++ is merged upstream and can be consider a
reasonable dependency for Zephyr, this work-around can be removed.
Closes: #57899
Signed-off-by: Keith Packard <keithp@keithp.com>
Provide a sensible default for the POSIX architecture,
as now it is possible to build with it.
Signed-off-by: Alberto Escolar Piedras <alberto.escolar.piedras@nordicsemi.no>
This matches the size that would be used with the newlib_nano configuration
and allows several tests to complete which would otherwise fail due to
insufficient heap space.
Signed-off-by: Keith Packard <keithp@keithp.com>
arm, arc and riscv all have special cases for malloc arena alignment that
might be smaller than the minimum required for a C allocator. In
particular, the riscv value might actually be zero, which turns out to be
an invalid alignment value.
Make sure all of these have alignment that meets the C language
requirements for allocation alignment.
Signed-off-by: Keith Packard <keithp@keithp.com>
The malloc arena needs to be aligned to a suitable size on Risc-V, and the
Z_RISCV_STACK_GUARD_SIZE variable is a helpful proxy for what size that
is. However, that is not always a power of two, so round it up to make the
linker capable of performing the necessary allocation.
Signed-off-by: Keith Packard <keithp@keithp.com>
This function may be used by system libraries (like libstdc++), so instead
of gating compile on target language, always build this function and gate
API visibility in header files instead.
Signed-off-by: Keith Packard <keithp@keithp.com>
Not all xtensa targets define the top of usable RAM via the _heap_sentry
address; it looks like the list is limited to esp32, esp32s2, esp32s3 and
intel parts. The first three all define HAS_ESPRESSIF_HAL, so key the test
using that or SOC_FAMILY_INTEL_ADSP.
Signed-off-by: Keith Packard <keithp@keithp.com>
Instead of explicitly initializing the mutex at runtime, use
SYS_MUTEX_DEFINE to initialize it at build time instead.
Signed-off-by: Keith Packard <keithp@keithp.com>
When using the common malloc implementation on systems not using the
minimal C library, allocate a reasonable default malloc heap according to
the following rules (adopted from the Picolibc heap size defaults):
* When an MMU is available, allocate 16kB.
* When USERSPACE is enabled for a device with an MPU require PoT alignment,
allocate 1024 bytes
* Otherwise, use all available memory.
Signed-off-by: Keith Packard <keithp@keithp.com>
When CONFIG_COMMON_LIBC_MALLOC_ARENA_SIZE is set to -1, make the malloc
arena use all remaining RAM. When an MMU exists, allocate the arena at
startup using k_mem_map. Otherwise, compute the available memory
automatically and use that.
When an MPU is being used to manage the malloc arena, make sure the heap
respects any MPU alignment requirements. Otherwise, align the heap to
sizeof(double).
Signed-off-by: Keith Packard <keithp@keithp.com>
Move the abort implementation into common so its shared among the
libc. As part of this start using the common abort on newlib.
Signed-off-by: Kumar Gala <kumar.gala@intel.com>
Move the strnlen implementation into common so its available to any
libc that may not implement strnlen.
Signed-off-by: Kumar Gala <kumar.gala@intel.com>
This will (eventually) permit use of a common malloc implementation with
other C libraries, reducing the amount of Zephyr-specific code required
to support each C library.
Signed-off-by: Keith Packard <keithp@keithp.com>
This allows the C library common library to be empty if no pieces are
needed, eliminating cmake warnings.
Signed-off-by: Keith Packard <keithp@keithp.com>
Introduce a place to share implementations of libc functions that
are needed by different libc versions. Place time() in this common
location so it can be shared when building for either minimal libc or
armclang libc.
Signed-off-by: Kumar Gala <kumar.gala@intel.com>
Christopher Friedt