This is a small optimization to the RISC-V trap handler. By splitting
the call to sched_run from the trap_handle call, loading the previous
thread ptr can be delayed until after it is determined that a schedule
run is required. This shaves of a few cycles during regular interrupts
that do not trigger the scheduler.
This commit reworks the trap entry to only save the callee-saved
registers when a context switch is required. the caller-saved registers
are always stored and restored to adhere to the RISC-V ABI. This saves
considerable cycles on interrupts.
The `_zep_params_setup()` function will modify the `argv` string passed to it.
This is a problem because that string is re-used on reboot.
The modified string is then later processed in `socket_zep_setup()`, so we have to keep
the memory around.
The `strdup()` fulfills all this and the memory is freed by `execve()` on reboot.
A proper solution would be to parse the strings in `_zep_params_setup()`.
While the hard fault handler prints the offending program counter, it
does not print information about the context triggering the hard fault.
This commit adds a line printing the thread ID and name that triggered
the hard fault. If the hard fault is triggered during an ISR, it only
prints that the hard fault happened during ISR context, not which ISR
triggered it.
The RISC-V timer should only be touched by periph/timer and must not be
initialized and enabled by the IRQ code. The current code can cause an
unhandled interrupt when the timer is not used and the mtime register
hits UINT64_MAX.
We can achieve greater accuracy for the relative timer_set()
if we don't use the generic implementation.
Use the same approach as used by atmega_common to trigger interrupts
for too small offsets.
tests/periph_timer_short_relative_set should now succeed for all intervals.
In most places, picolibc and newlib are the same, so use
the existing newlib code when compiling with picolibc.
Signed-off-by: Keith Packard <keithp@keithp.com>
Allocate and initialize a thread-local block for each thread at the
top of the stack.
Set the tls base when switching to a new thread.
Add tdata/tbss linker instructions to cortex_m and risc-v scripts.
Signed-off-by: Keith Packard <keithp@keithp.com>
---
v2:
Squash fixes
v3:
Replace tabs with spaces
v4:
Add tbss to fe310 linker script
Disable the newlib-nano stubs code when picolibc is in use
Signed-off-by: Keith Packard <keithp@keithp.com>
---
v2:
Squash fixes in
v3:
call stdio_init in _PICOLIBC_ mode to initialize uart
v3:
Remove call to stdio_init from nanostubs_init, always
call from cpu_init.
Picolibc makes atexit state per-thread instead of global, so we can't
register destructors with atexit in a non-thread context as we won't
have any TLS space initialized.
Signed-off-by: Keith Packard <keithp@keithp.com>
Support for picolibc as alternative libc implementation is added with
this commit. For now only cortex-m CPU's are supported.
Enable via PICOLIBC=1
---
v2:
squash fixes in
v3:
Remove picolibc integer printf/scanf stuff from sys/Makefile.include,
it gets set in makefiles/libc/picolibc.mk
fixup for dependency
The MPU on the cortex-m23 has some differences with the MPU on the older
cortex-m devices. It is not implemented in the cortex-m MPU driver. This
removes the available feature as it gives a false sense of security by
advertising the feature, but implementing it with noop's
This adds a placeholder define for when the DMA peripheral available on
the MCU doesn't support channel/trigger filtering. This is the case on
the stm32f1 and stm32f3 family.
The stm32_eth driver was build on top of the internal API periph_eth, which
was unused anywhere. (Additionally, with two obscure exceptions, no functions
where declared in headers, making them pretty hard to use anyway.)
The separation of the driver into two layers incurs overhead, but does not
result in cleaner structure or reuse of code. Thus, this artificial separation
was dropped.
The Ethernet DMA is capable of collecting a frame from multiple chunks, just
like the send function of the netdev interface passes. The send function was
rewritten to just set up the Ethernet DMA up to collect the outgoing frame
while sending. As a result, the send function blocks until the frame is
sent to keep control over the buffers.
This frees 6 KiB of RAM previously used for TX buffers.
