- Save the FPU registers into the tcb so they don't get lost if the stack
frame for xcp.regs moves (as it does)
- Handle interger and FPU register save/load separately
- Integer registers are saved/loaded always, like before
- FPU registers are only saved during a context switch:
- Save ONLY if FPU is dirty
- Restore always if FPU has been used (not in FSTATE_OFF, FSTATE_INIT)
- Remove all lazy-FPU related logic from the macros, it is not needed
Why? The tcb can contain info that is needed by the context switch
routine. One example is lazy-FPU handling; the integer registers can
be stored into the stack, because they are always stored & restored.
Lazy-FPU however needs a non-volatile location to store the FPU registers
as the save feature will skip saving a clean FPU, but the restore must
always restore the FPU registers if the thread uses FPU.
This adds support for the CoreMMC v3.1 FPGA driver as described
in Microchip Handbook HB0510. The driver doesn't support DMA.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
USBD has many limitation that make it hard to work properly:
- only one EasyDMA channel which must be shared for all EPs
- only one DMA transfer can take place at a time
- some registers are unavailable during DMA transfer
- in case of any problems, the peripheral silently blocks,
or lose the transmitted bytes without information for the user
This commit is trying to fix these problem and makes the USBS stack more reliable.
Tested with high-speed CDCACM data transfers and that's the best I've been able to get in terms of stability.
if config_walltime_signal is enabled, NuttX_sim will receive a lot of
signals, the socket api will break and errno will be EINTR, masking irq
before calling the host socket api will avoid this problem.
Signed-off-by: zhanghongyu <zhanghongyu@xiaomi.com>
In SMP mode, the fpu owner may switch from core0 to core1,
so it is necessary to force saving the FPU context when a
context switch occurs.
This PR fixed the crash issue mentioned in #8799.
Signed-off-by: zhangyuan21 <zhangyuan21@xiaomi.com>
If a kernel stack exists, use that whenever the user process is in
privileged mode, i.e. running an exception or in system call. Previously
the exception context was stored into the user's stack, which is not ideal.
Why?
1. Because the exception entry status (REG_INT_CTX) is needed by the
kernel, and this is now in user memory which requires that the correct
user mappings are active when it is accessed.
2. The user must currently account for the exception stack frame (which
is BIG) in its own stack allocation. Moving the exception context save
to the kernel stack offloads this responsibility from the user to the
kernel, which is IMO the correct behavior.
3. The kernel access to user memory is currently allowed without condition,
however this is not ideal either. The privileged mode status CSR allows
blocking access to user memory via the STATUS_SUM-bit, which should be
disabled by default and only enabled when access to user space is really
needed. This patch allows implementing such features.
This is preparation to use kernel stack for everything when the user
process enters the kernel. Now the user stack is in use when the user
process runs a system call, which might not be the safest option.
This is a minimalistic SBI implementation for NuttX.
Provides a single service for now:
- Access to machine timer
Provides a start trampoline to start NuttX in S-mode:
- Exceptions / faults are delegated to S-mode.
- External interrupts are delegated to S-mode.
Machine mode timer is used as follows:
- The timer compare match register reload happens in M-mode, via
call gate "riscv_sbi_set_timer"
- The compare match event is dispatched to S-mode ISR, which will
notify the kernel to advance time
- Clearing the STIP interrupt does not work from S-mode,
so the call gate does this from M-mode
The only supported (tested) target for now is MPFS.
CortexR52 can have a optional FPU.
- VFPv3 with FP16
- Option 1: 16 x double-prevision registers - -mfpu=vfpv3-d16-fp16
- Option 1: 32 x double-prevision registers - -mfpu=vfpv3-fp16
Summary:
- I noticed that ./tools/configure.sh fvp-armv8r:nsh_smp shows
warning: (ARCH_CHIP_FVP_R52 && ARCH_CHIP_FVP_R82) selects ARMV8R_HAVE_GICv3 which has unmet direct dependencies (ARCH_ARM && ARCH_ARMV8R)
- I think ARMV8R_HAVE_GICv3 is only used for aarch32.
- This commit fixes this issue.
Impact:
- None
Testing:
- Tested with nsh_smp on FVP
Signed-off-by: Masayuki Ishikawa <Masayuki.Ishikawa@jp.sony.com>
This adds functionality to map pages dynamically into kernel virtual
memory. This allows implementing I/O remap for example, which is a useful
(future) feature.
Now, the first target is to support mapping user pages for the kernel.
Why? There are some userspace structures that might be needed when the
userspace process is not running. Semaphores are one such example. Signals
and the WDT timeout both need access to the user semaphore to work
properly. Even though for this only obtaining the kernel addressable
page pool virtual address is needed, for completeness a procedure is
provided to map several pages.
During initialization it isn't possible to use up_puts once it's
protected against concurrent access through a mutex lock. Instead,
using up_putc makes it similar to ESP32S2 and ESP32S3 and perfectly
fits for showprogress usage.
Basic work required for uniprocessor CortexR52 (ARMv8R AARCH32) using
GICv3 and CP15 mapped arch timer.
Tested on ARM FVP 11.20.
Port is based on ARMv8R AARCH64 and ARMv7R code. Excuse possible copy-paste leftovers.
Remove unnecessary reading of the status register when loading / unloading
the FIFOs. Reading from the IP block is slow due to BUS synchronization and
this basically makes the SPI busy loop for no reason at all, destroying the
CPU usage.
The overall benefit of these changes is approx. 25%-points, which is a
MASSIVE improvement.
Adds a driver for an FPGA fabric / CoreSPI implementation.
Supports multiple instances, assuming they reside in some base address,
offsettable by a constant value.
Add an interface that validate if EasyDMA transfer is possible.
EasyDMA cannot access flash memory which can cause hard to detect silent bugs.
This feature is enabled if CONFIG_DEBUG_FEATURES=y and CONFIG_DEBUG_ASSERTIONS=y.
External function to query vbus status. Reading from the block requires
the clock, but if no devices are open -> vbus detect does not work.
This creates a chicken / egg problem, if vbus detect is used to start
the usb device.
bmp388 works poorly as the system fires STOPs even in
a beginning of a transaction. Don't let unrelated STOPs
to distort the data flow.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
This adds 2 more FPGA I2Cs. Also rework the indexing
so that it matches the earlier work without major changes.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
This fixes the following issue:
- After sending the address, the driver writes an extra zero
Without this patch, the extra write causes an extra ACK that would
terminate the sequence prematurely. This is observed as data read
corruption.
With this fix, the condition is detected precisely. That being the
case, the sequence is continued with a repeated start, after which
the read continues normally.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
This incorporates an fpga i2c driver into the existing i2c driver.
This fpga i2c works almost 100% as the MSS i2c, but the difference
is that the fpga driver terminates all transactions with a stop
sent -interrupt. That needs to be handled.
Fpga clock source is also different, act accordingly.
SEC2TICK(10) is an overkill to any app, use just one second instead.
modifyreg32s are simplified as well, no need to clear and set
as set is enough.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
I2C status register reset value (0xf8) was not handled properly causing unnecessary bus resets.
Added critical section to mpfs_i2c_reset() and removed unnecessary interrupt disabling elsewhere.
Confirmation of the IN request must be done immediately after all data has been transferred,
otherwise sending data when more than one request has been added to the queue will
not work properly.
- Update TrustedFirmare-M instructions to latest version of STM32CubeL5
- Increase idle thread stack size to not overflow during system init
- Select ARCH_HAVE_TRUSTZONE for STM32L5
- Set CONFIG_ARCH_TRUSTZONE_NONSECURE for stm32l562e-dk:nsh, since NuttX
is running in the Non-secure world.
See https://github.com/apache/nuttx/issues/9316
Signed-off-by: Michael Jung <michael.jung@secore.ly>
