The critical section port already disables interrupts and reads the core id for
the whole section, but spinlock_acquire()/spinlock_release() then disabled
interrupts again to the same level and re-read the core id register.
Add spinlock_acquire_impl()/spinlock_release_impl(), which take a caller-supplied
owner id and skip interrupt management, and reuse them from
spinlock_acquire()/spinlock_release() to avoid duplicated code. The Xtensa and
RISC-V ports now read the core id once and call the impl variants, removing one
core id read and one interrupt mask/restore per critical section enter and exit.
Closes https://github.com/espressif/esp-idf/issues/18908
Tick hook registration lives in esp_system, so its scheduler-suspension
coverage belongs in the esp_system unity test app rather than the
freertos IDF additions suite.
Remove ~50 duplicate local definitions of ALIGN_UP/ALIGN_DOWN/ALIGN_UP_BY/
ALIGN_DOWN_BY across the codebase and replace them with canonical
ESP_ALIGN_UP/ESP_ALIGN_DOWN from esp_macros.h.
On the Linux port, tasks are pthreads that cannot be forcibly paused;
the scheduler "unschedules" a task only by selecting a different next
task, and the outgoing task keeps running until it next blocks itself
in event_wait (e.g. via vPortYield). If the scheduler later switches
back into such a task (event_signal) before it ever parked in
event_wait, that wake is never consumed and event_triggered stays
latched. The task's next voluntary block then returns immediately
instead of blocking (e.g. vTaskDelay(100ms) returning in 0ms), which
desynchronizes kernel and port scheduling state.
Clear the outgoing task's pending event when switching to a different
task, under s_port_mutex and atomically with the scheduling decision.
Any latched-but-unconsumed wake is stale at that point; a legitimate
wake can only be delivered later, once the scheduler selects the task
again as next_thread.
Add event_clear() to the wait_for_event helper to support this.
Verified: minimal repros (busy-wait with another ready task, and
pthread-blocked task) now measure 92-101 ms for vTaskDelay(100ms)
instead of 0 ms; linux_freertos kernel test app 18/18 pass; the
previously failing esp_linenoise host test suite passes 34/34.
Rework the syscall interposition architecture so FreeRTOS works
standalone (without VFS) and VFS optionally overrides with strong
symbols. All kernel FDs are registered with VFS via
esp_vfs_register_fd_with_local_fd, so the application only sees
VFS-allocated FD numbers, preventing numerical collisions between
kernel FDs and VFS-internal FD slots.
FreeRTOS side:
- Create linux_port_coop_internal.h with LINUX_COOP_IO_LOOP,
LINUX_COOP_RESOLVE, linux_coop_yield and FD state table functions
- Export cooperative I/O primitives (freertos_linux_coop_read/write/
open/close/fcntl/select/pread/pwrite/readv/writev/recv/send/
recvfrom/sendto/recvmsg/sendmsg/connect/accept/pselect/poll/
socket/socketpair/pipe/pipe2/dup/dup2/syscalls_init)
- Define weak POSIX symbols calling through to the cooperative
primitives; keep nanosleep/sleep/usleep as strong (not FD-related)
- Refactor freertos_linux_coop_syscalls.h into a pure public API header
VFS side:
- Register the Linux host FS with esp_vfs_register_fs_with_id() as a
proper VFS driver; register stdin/stdout/stderr at init (priority 99)
- Rewrite vfs_linux.c with strong POSIX symbols dispatching through
esp_vfs_*
- Add strong overrides for FD-creating syscalls (open, pipe, pipe2,
socket, socketpair, dup, dup2, accept) that register returned FDs
with VFS
- Add strong overrides for FD-translating syscalls (readv, writev,
recv, send, recvfrom, sendto, recvmsg, sendmsg, connect, pselect,
poll) that translate VFS FD to kernel FD
- Delete vfs_coop_syscalls.c (absorbed into FreeRTOS weak + VFS strong)
- Update CMakeLists.txt: remove vfs_coop_syscalls.c, add linker hook
Moved the linux spinlock.h stub into esp_hw_support/include/linux.
esp_hw_support's INCLUDE_DIRS already places include/linux before
include, so the correct include files are picked up for the linux
target.
- Drop the FreeRTOS-Kernel-SMP linux spinlock.h duplicate.
- rv_decode_test: use PRIV_REQUIRES esp_hw_support instead of
hardcoding component include paths.
Replace the eTaskGetState() polling guard in prvTaskDeleteWithCaps()
with a per-core xTaskGetCurrentTaskHandleForCore() check, so the
predicate matches vTaskDelete()'s immediate-vs-deferred decision. This
closes the taskTASK_SCHEDULED_TO_YIELD race against the IDLE-deferred
TCB cleanup on the SMP preview kernel; behaviour is unchanged on the
default IDF kernel and on single-core targets.
With CONFIG_ESP_MAIN_TASK_AFFINITY_NO_AFFINITY=y the main task could
switch core between registering the idle hook and the while loop.
This would cause a deadlock were the current task was waiting for the
idle hook to run on the same core it's busy waiting.
Merge https://github.com/espressif/esp-idf/pull/16149
ESP32-S31 scheduling and spinlock benchmarks match the relaxed P4 thresholds
better than the generic limits, so use those limits to keep the
FreeRTOS perf cases stable.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This commit adds additional delays to the stress test where we create
and self-delete tasks on PSRAM. iSuch tasks rely on the kernel creating
a special cleanup task to delete them. Hence, the delays in the main
task allow the cleanup task to bescheduled and the deletion to go
through.
The macro expanded with a trailing PRIVILEGED_FUNCTION; token, making it
unusable in expression contexts (e.g. inside assert() or as a function
argument), causing a compile error: "expected ')' before ';' token".
Closes https://github.com/espressif/esp-idf/issues/17948
Made-with: Cursor
The Xtensa load/store handler did not properly handle 8/16-bit
memory access to IRAM regions configured with MALLOC_CAP_IRAM_8BIT
(and CONFIG_ESP32_IRAM_AS_8BIT_ACCESSIBLE_MEMORY=y) from a loop
(LBEG/LEND/LCOUNT) context. This caused the loop to exit after
the first access, instead of continuing to iterate as intended.
Closes https://github.com/espressif/esp-idf/issues/14127
