The way the ready thread cache was implemented caused it to not always
be "hot", i.e. there could be some misses, which happened when the
cached thread was taken out of the ready queue. When that happened, it
was not replaced immediately, since doing so could mean that the
replacement might not run because the flow could be interrupted and
another thread could take its place. This was the more conservative
approach that insured that moving a thread to the cache would never be
wasted.
However, this caused two problems:
1. The cache could not be refilled until another thread context-switched
in, since there was no thread in the cache to compare priorities
against.
2. Interrupt exit code would always have to call into C to find what
thread to run when the current thread was not coop and did not have the
scheduler locked. Furthermore, it was possible for this code path to
encounter a cold cache and then it had to find out what thread to run
the long way.
To fix this, filling the cache is now more aggressive, i.e. the next
thread to put in the cache is found even in the case the current cached
thread is context-switched out. This ensures the interrupt exit code is
much faster on the slow path. In addition, since finding the next thread
to run is now always "get it from the cache", which is a simple fetch
from memory (_kernel.ready_q.cache), there is no need to call the more
complex C code.
On the ARM FRDM K64F board, this improvement is seen:
Before:
1- Measure time to switch from ISR back to interrupted task
switching time is 215 tcs = 1791 nsec
2- Measure time from ISR to executing a different task (rescheduled)
switch time is 315 tcs = 2625 nsec
After:
1- Measure time to switch from ISR back to interrupted task
switching time is 130 tcs = 1083 nsec
2- Measure time from ISR to executing a different task (rescheduled)
switch time is 225 tcs = 1875 nsec
These are the most dramatic improvements, but most of the numbers
generated by the latency_measure test are improved.
Fixes ZEP-1401.
Change-Id: I2eaac147048b1ec71a93bd0a285e743a39533973
Signed-off-by: Benjamin Walsh <benjamin.walsh@windriver.com>
The ARM Cortex-M early boot was using a custom stack at the end of the
SRAM instead of the interrupt stack. This works as long as no static
data that needs a known initial value occupies that stack space. This
has probably not been an issue because the .noinit section is at the
very end of the image, but it was still wrong to use that region of
memory for that initial stack.
To be able to use the interrupt stack during early boot, the stack has
to be released before an interrupt can happen. Since ARM Cortex-M uses
PendSV as a very low priority exception for context switching, if a
device driver installs and enables an interrupt during the PRE_KERNEL
initialization points, an interrupt could take precedence over PendSV
while the initial dummy thread has not yet been context switched of and
thus released the interrupt stack. To address this, rather than using
_Swap() and thus triggering PendSV, the initialization logic switches to
the main stack and branches to _main() directly instead.
Fixes ZEP-1309
Change-Id: If0b62cc66470b45b601e63826b5b3306e6a25ae9
Signed-off-by: Benjamin Walsh <benjamin.walsh@windriver.com>
There was a possible race condition when setting the return value of a
thread that is pending, from an ISR.
A kernel function causes a thread to pend, with the following series of
steps:
- disable interrupts
- move current thread to wait_q
- call _Swap
Depending if running on M3/4 or M0+, _Swap will either issue a svc #0,
or pend PendSV directly. The same problem exists in both cases.
M3/4:
__svc will:
- enable interrupts
- trigger __pendsv
M0+:
_Swap() will enable interrupts.
__pendsv will:
- save register context including PSP into the thread struct
If an interrupt occurs between interrupts being enabled them and
__pendsv saving PSP, and the ISR sets the pending thread's return value,
this will happen:
- sees the thread in a wait_q
- removes it
- makes it ready
- calls _set_thread_return_value
- _set_thread_return_value looks at the thread's saved PSP to poke
the value
In this scenario, PSP hasn't yet been updated by __pendsv so it's a
stale value from the previous context switch, resulting in unpredictable
word on the stack getting set to the return value.
There is no way to fix this issue and still have the return value being
delivered directly in the pending thread's exception stack frame, in the
M0+ case. There will always be a window between the unlocking of
interrupts and PendSV being handled. On M3/4, it could be possible with
the mix of SVC and PendSV, since the exception stack frame is created in
the __svc handler. However, because we want to keep the two
implementations as close as possible, and there were talks of moving
M3/4 to using PendSV only, to save an exception, the approach taken
solves both cases.
The approach taken is similar to the ARC and Nios2 ports, where
there is a field in the thread structure that holds the return value.
_Swap() then loads r0/a1 with that value just before returning.
Fixes ZEP-1289.
Change-Id: Iee7e06fe3f8ded84aff918fd43408c7f589344d9
Signed-off-by: Benjamin Walsh <benjamin.walsh@windriver.com>
There was a lot of duplication between architectures for the definition
of threads and the "nanokernel" guts. These have been consolidated.
Now, a common file kernel/unified/include/kernel_structs.h holds the
common definitions. Architectures provide two files to complement it:
kernel_arch_data.h and kernel_arch_func.h. The first one contains at
least the struct _thread_arch and struct _kernel_arch data structures,
as well as the struct _callee_saved and struct _caller_saved register
layouts. The second file contains anything that needs what is provided
by the common stuff in kernel_structs.h. Those two files are only meant
to be included in kernel_structs.h in very specific locations.
The thread data structure has been separated into three major parts:
common struct _thread_base and struct k_thread, and arch-specific struct
_thread_arch. The first and third ones are included in the second.
The struct s_NANO data structure has been split into two: common struct
_kernel and arch-specific struct _kernel_arch. The latter is included in
the former.
Offsets files have also changed: nano_offsets.h has been renamed
kernel_offsets.h and is still included by the arch-specific offsets.c.
Also, since the thread and kernel data structures are now made of
sub-structures, offsets have to be added to make up the full offset.
Some of these additions have been consolidated in shorter symbols,
available from kernel/unified/include/offsets_short.h, which includes an
arch-specific offsets_arch_short.h. Most of the code include
offsets_short.h now instead of offsets.h.
Change-Id: I084645cb7e6db8db69aeaaf162963fe157045d5a
Signed-off-by: Benjamin Walsh <benjamin.walsh@windriver.com>
Benjamin Walsh