Verify the thread priorities are within the bounds when starting a new
thread and when changing the priority of a thread.
Change-Id: I007b3b249e4b80235b6439cbee44cad2f31973bb
Signed-off-by: Benjamin Walsh <benjamin.walsh@windriver.com>
A race condition would happend if a FIRQ interrupted a
return-from-interrupt from a RIRQ at the wrong moment: if a decision was
already taken which thread to context switch in and the FIRQ woke up
another thread of higher priority, the ready queue would be corrupted.
The solution is to lock interrupts at the moment the interrupt return
code starts looking at the kernel queues. Interrupts do not need to be
unlocked before exiting: the return-from-interrupt (rtie) instruction
will restore the correct interrupt locking state for the thread being
context switched in.
Change-Id: I777665c2faeca7b1f2a77ddd9ee2a520080bae88
Signed-off-by: Benjamin Walsh <benjamin.walsh@windriver.com>
An implementation to flush multiple d-cache lines has been added
per the top-level cache.h API. ZEP-1153 was opened to express
the need for MORE i-cache and d-cache APIs. For example, the current
cache.h API doesn't provide a means to invalidate d-cache lines
and has nothing for i-cache.
I've also modified some of the i-cache related aux registers to have
better names so that they won't be confused with d-cache.
These changes are for
ZEP-1176.
Change-Id: If4c5410451cc40dcd5618fc871093c8febf7e061
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
Adds standard prefix to symbolic option that flags a thread
as essential to system operation.
Change-Id: Ia904a81ce343fdd1cd44caaaeae641d822777f9b
Signed-off-by: Allan Stephens <allan.stephens@windriver.com>
We can derive NUM_REGULAR_IRQ_PRIO_LEVELS by subtracting 1
from CONFIG_IRQ_PRIO_LEVELS if FIRQ is present (which is currently
always the case). If FIRQ is not present, the value will be equal
to CONFIG_IRQ_PRIO_LEVELS since all interrupts will be regular.
Change-Id: Ibefc939e3771bf0adf712127db0d36cb49bf732b
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
Fix the error in thread rescheduling:
Fix Fast IRQ exit routine error when it reschedules threads if
(prio >= 0) || (sched_locked == 0) || (next_thread == _current),
while the correct condition for thread rescheduling is:
(prio >= 0) && (sched_locked == 0) && (next_thread != _current),
Fix regular IRQ error when the regular IRQ exit routine rescheduled
threads when (next_thread == _current) instead of
(next_thread != current).
Increased IDLE_STACK_SIZE for ARC architecture, to hold saved
registers.
Change-Id: I1d87a968e231e13822844b7564567e6ca310cde2
Signed-off-by: Dmitriy Korovkin <dmitriy.korovkin@windriver.com>
Gets rid of unnecessary THREAD_MONITOR_INIT() macro, to be
consistent with the approach taken by _thread_monitor_exit().
Aligns x86 code with the approach used on other architectures.
Revises the associated comments and removes unnecessary
doxygen tags.
Change-Id: Ied1aebcd476afb82f61862b77264efb8a7dc66c9
Signed-off-by: Allan Stephens <allan.stephens@windriver.com>
- the interrupt (both regular and fast) now does not do rescheduling
if the current thread is a coop thread or if the scheduler is not locked
- the _nanokernel.flags cache of _current.flags is not used anymore
(could be a source of bugs) and is not needed in the scheduling algo
- there is no 'task' field in the _nanokernel anymore: scheduling routines
call _get_next_ready_thread instead
- the _nanokernel.fiber field is replaced by a more sophisticated
ready_q, based on the microkernel's priority-bitmap-based one
- thread initialization initializes new fields in the tcs, and does not
initialize obsolete ones
- nano_private includes nano_internal.h from the unified directory
- The FIBER, TASK and PREEMPTIBLE flags do not exist anymore: the thread
priority drives the behaviour
- the tcs uses a dlist for queuing in both ready and wait queues instead
of a custom singly-linked list
- other new fields in the tcs include a schedule-lock count, a
back-pointer to init data (when the task is static) and a pointer to
swap data, needed when a thread pending on _Swap() must be passed more
then just one value (e.g. k_stack_pop() needs an error code and data)
- the 'fiber' and 'task' fields of _nanokernel are replaced with an O(1)
ready queue (taken from the microkernel)
- fiberRtnValueSet() is aliased to _set_thread_return_value since it
also operates on preempt threads now
- _set_thread_return_value_with_data() sets the swap_data field in
addition to a return value from _Swap()
- convenience aliases are created for shorter names:
- _current is defined as _nanokernel.current
- _ready_q is defined as _nanokernel.ready_q
- _Swap() sets the threads's return code to -EAGAIN before swapping out
to prevent timeouts to have to set it (solves hard issues in some
kernel objects).
Change-Id: Ib9690173cbc36c36a9ec67e65590b40d758673de
Signed-off-by: Dmitriy Korovkin <dmitriy.korovkin@windriver.com>
For the EM Starterkit, one SOC I will soon be adding is EM7D.
This SOC has FIRQ, but only has one register bank.
Thus the interrupt handling for FIRQ needs to be different
when CONFIG_RGF_NUM_BANKS==1. The handler must instead push
registers onto the stack in the same stack frame layout that RIRQ uses.
This allows for context switch to be easily done since its compatible.
The common interrupt entry point _isr_enter must save r0 before using
it, because in the FIRQ 1-bank case, it would be destroyed otherwise.
So a global variable named saved_r0 has been added for this reason.
The stack cannot be used to save r0, because it first has to determine
whether its FIRQ or RIRQ here. This change has been tested on the
EM Starterkit with EM7D SOC changes -- coming soon. To make the review
easier, these 3 files are submitted first.
Also, exceptions will no longer use the _firq_stack.
This stack is not needed in the 1-bank case, but an exception stack
is needed. I've added a new stack called _exception_stack,
and made it be 512B, which should be enough for one exception.
See ZEP-966
Change-Id: I6f228b840da7c4db440dd1cfef4ae25336c87f0d
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
The code density registers are NOT saved on the stack.
This is controlled by bit 13 in AUX_IRQ_CTRL, which didn't even have a symbol
defined for it. I've also added _ARC_V2_AUX_IRQ_CTRL_LP for bit 13.
Change-Id: Ie80853b72bed4e60a5cf1cf0a8c905a3d86180d9
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
commit e57b21c78c ("irq: Use lowest priority not a
hard-coded priority 2") introduce a wrong whitespace,
not complying with coding style. Remove it.
Change-Id: Ie7e48843e5da6cb3417773ef8a57cf9a166c70d6
Signed-off-by: Qiu Peiyang <peiyangx.qiu@intel.com>
In this file was found an assumption about how many
priorities are being used. This is configurable with
CONFIG_NUM_IRQ_PRIO_LEVELS, however, so it should be using that
instead. This line of code changes:
or r3,r3,(1<<(CONFIG_NUM_IRQ_PRIO_LEVELS-1))
so as to use the correct bit to OR.
Change-Id: I8c6297e98b5163aa27460a68b203e8a27d1e2506
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
A previous re-work of IRQ priorities was led astray by an incorrect
comment. Priority level 1 is not a non-maskable interrupt priority.
In addition, zero latency IRQs are not implemented on ARC.
Timer driver now doesn't specify IRQ_ZERO_LATENCY (as that wouldn't be
correct) and its IRQ priority is now tunable in Kconfig. The default is 0.
IPM driver on both ARC and x86 side were being configured with hard-coded
priority of 2, which wasn't valid for ARC and caused an assertion failure.
The priority level is now tunable with Kconfig and defaults to 1 for ARC.
Issue: ZEP-693
Change-Id: If76dbfee214be7630d787be0bce4549a1ecbcb5b
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
We have already done this on x86 and ARM. The policy is as follows:
* IRQ priority levels starting at 0 all have the same semantics and
do not have special properties. The priority level is either ignored
on arches which do not support programmable priority levels, or lower
priority levels take precedence over higher ones.
* Special-case priorty levels are specified via flags, in which case
the supplied priority level is ignored.
Issue: ZEP-60
Change-Id: Ic603f49299ee1426fb9350ca29d0b8ef96a1d53a
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
Arches now select whether they want to use the GCC built-ins,
their own assembly implementation, or the generic C code.
At the moment, the SDK compilers only support builtins for ARM
and X86. ZEP-557 opened to investigate further.
Change-Id: I53e411b4967d87f737338379bd482bd653f19422
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
Used by ARC, ARM, Nios II. x86 has alternate code done in assembly.
Linker scripts had some alarming comments about data/BSS overlap,
but the beginning of BSS is aligned so this can't happen even if
the end of data isn't.
The common code doesn't use fake pointer values for the number of
words in these sections, don't compute or export them.
Change-Id: I4291c2a6d0222d0a3e95c140deae7539ebab3cc3
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
There are a few problems with the code being repaired here.
1. A seti was used to re-enable all interrupts, even though the
thread being switched to may have had certain interrupt priorities masked.
2. saved status32 already has SC bit if thats wanted, so its ok to just
restore status32 as-is w/o needing to and off anything.
3. the code is difficult to write using kflag and seti because as you
restore registers, there aren't any to use. But we can exploit a
trick where we pretend an interrupt has occured by setting a bit in
AUX_IRQ_ACT, and then use RTIE instruction to restore status32
atomically with branching to return address. Something about the way
this code was written was causing stack corruptings and crashes in an
application that uses a high rate of both FIRQ and Regular interrupts.
Change-Id: Ia7166d51f0e750c07832ab115b7151ce37ee0278
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
Since firq utilizes a 2nd register bank, and since all of those
many GPRs can be used, the strategy here is to save extra registers,
such as lp_count, lp_start, lp_end into callee saved registers.
These registers are safe to use because the C-ABI followed by the
compiler will cause these to be spilled to the stack if a C function
wants to use them. By selecting upper GPRs, r23-r25, it is very unlikely
the compiler will spill them. This improvement, therefore, can avoid a
d-cache miss since we are avoding memory altogether when saving these.
The struct firq_regs is no longer needed.
Change-Id: I7c0d061908a90376da7a0101b62e804647a20443
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
It was found that the test latency_measure, when compiled
for microkernel, would fail on the ARC. This because the
trap handler, used by irq_offload, wasn't supporting thread switching.
This submission adds the code to do that, and the code size is
bigger only when CONFIG_MICROKERNEL is defined.
To keep code a bit smaller, there is a trick exploited here where
the AE bit is cleared in the STATUS32 register and in AUX_IRQ_ACT,
bit 1 is set, to make it appear as if the machine has interrupted
at priority 1 level. It then can jump into some common interrupt
exit code for regular interrupts and perform an RTIE instruction
to switch into the new thread.
test/latency_measure/microkernel now passes.
Change-Id: I1872a80bb09a259814540567f51721203201679a
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
The .nd on a branch is WRONG if its an unconditional branch. Not needed.
On conditional branches its a compiler feature that is not yet functional
with ARC targets. Typical code for this compiler can use .d to put
something in the delay slot of an instruction, but using .nd is probably
never wanted.
Change-Id: If1017c468e6e7af269ea73daeb4bc223dcc0059f
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
The ARC CPU comes up from reset with i-cache enabled.
It can have garbage in it from a previous run.
The fix is to check the build register for the i-cache, and if its
present, invalidate it fully, and then disable it.
_icache_setup() is called later to turn it on.
Change-Id: I26fae915153841c61e9530d5af2ddb9d0553275b
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
In fault_s.S, changing the word "save" to "safe".
Change-Id: Ia997082a62bf287f09a72b7f0a00d506bd982770
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
Some ARC CPUs can be built with separate instruction bus
and data bus (i.e. Harvard Architecture). Such systems
have only ICCM and DCCM memories. When CONFIG_HARVARD
is defined, the initial stack pointer is set to the
TOP of the DCCM memory. Currently there is no SOC that
existing in Zephyr tree that sets CONFIG_HARVARD, but
this will be coming soon.
Change-Id: I2016d1f472fbdad683a964aa0b65c5263ecfb6cf
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
Some ARC targets can have a data-cache. Although there is no special
instruction to clear exceptions during early init, it is necessary to
invalidate the d-cache BEFORE any data is fetched. The ARC on arduino 101
doesn't have d-cache, and will thus skip this d-cache invalidate.
Also, it is important to set the vector table base register to point to
the interrupt vector table EARLY, so that if an exception is encountered,
the correct vector table is found. Set this base only if it is found to be
different from the one compiled in to the code.
These initialization steps assure that proper exception handling
is in place during early init.
Change-Id: Ie8b5928e5813e104680a6d6510c85d32dc8ed8f3
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
The lr and sr instructions cause a pipeline bubble. There is an efficiency
to be gained if pairs of lr or sr instructions are done right next to
each other. This can avoid some stall cycles.
Also, r14 and r15 can be used with isa-16 instructions.
Change-Id: I4165365b49da910db31e0699a1a6e47114962942
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
By using isa-16 instructions, a bit of code-size can be saved,
and code can be a little faster.
Change-Id: I0567d8274372748f579610e2bd4236ce52c5d6c8
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
Code size can be reduced by replacing ld and st
with ld_s and st_s (if target registers are r0-r3).
Change-Id: Ia70f0aff07fe41a0cfeff2d59dcdadf7c88e1ae8
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
code-size optimization to use small-variant loads/stores with %r13w
Change-Id: Ic9b2b7744f7d465bccb1e59f64e621985ae7d04d
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
There is a BUG here in that the alias for __start was
aliased to the start of the vector table. Yet, on ARC CPUs,
the vector table CANNOT be the entry point, because there
is no code in a vector table. Only addresses appear in each vector.
Thus, the reset vector, at offset 0 in this table, is a raw address.
The top Makefile in zephyr sets the lable __start to be the entry point
like this: -e __start. Debuggers, for example, use this entry point
to know where the first line of code is.
Also, in KConfig, there were duplicate NSIM blocks. One has been
removed.
Change-Id: I480be7d338a8b45b8ea6ef3f55ac2e6c43829452
Signed-off-by: Chuck Jordan <cjordan@synopsys.com>
Fibers initialize this back pointer to NULL as they are (by definition)
not microkernel tasks. Microkernel tasks initialize it to their
corresponding 'ktask_t'.
However for nanokernel systems, the back pointer is always NULL. This
is because there is only one task in a nanokernel system (the background
task) and it can not pend on a nanokernel object--it must poll.
Change-Id: I9840fecc44224bef63d09d587d703720cf33ad57
Signed-off-by: Peter Mitsis <peter.mitsis@windriver.com>
We really should have more faith in the compiler, it generates
code to implement this exactly like the arch-specific assembly
versions, and on ARM is actually 4 bytes shorter.
FUNC_NO_FP used to disable the usual C preamble to update the
frame/stack pointers, which is how the sizes are still the same
or less. It's debatable how useful the occasional use of
FUNC_NO_FP is in practice since it hinders debugging and in a
production build frame pointers should be globally disabled, but
we can address that later.
Change-Id: I6c4b64ab3e3a9b6f91d52fa8c92e6e79a986fc77
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
Of the 3 related functions;
_thread_essential_set()
_thread_essential_clear()
_is_thread_essential()
The first two are parameter-less and always operate on
"_nanokernel.current". The last one takes a 'thread' parameter but will
operate on _nanokernel.current if the parameter is NULL. All calls to
_is_thread_essential() pass NULL!
This change makes the 3 functions consistent by removing the parameter
to the 3rd function. This should also be marginally more efficient,
though consistency was the motivation. This change corrects the doc
preamble to all 3 functions.
(These functions would probably be better as inlines. Also, the choice
of when to use wrappers seems a bit arbitrary. E.g. there's nothing
for setting/testing the "FIBER" flag.)
Change-Id: Ie3589f8a28b227c6d7a3a31b664d3b3e6e9c6d17
Signed-off-by: Geoff Thorpe <geoff.thorpe@nxp.com>
These C variants of atomic operations can work on any arch,
have platforms select them if they don't have ASM equivalents.
Change-Id: I38eb03bb58beff865681ee56ef7bc0fcded1e906
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
Remove hardcoding and make the values configurable. Also make the
Kconfig variables consistent with other architectures.
Change-Id: I69334002303d4d8abaf7363d9134fd5f46ce4eeb
Signed-off-by: Anas Nashif <anas.nashif@intel.com>
lp_count register can be store directly to memory, it is actualy done,
and the instruction that precedes it has no effect.
Change-Id: I8b8fee6abd6f08eea38dd1ab5bbe61c25a2a1f7d
Signed-off-by: Alexandre d'Alton <alexandre.dalton@intel.com>
ARC CPU has stack checking feature that allows to trigger an exception
whenever the stack is incorrectly accessed.
This patch implements the stack_top and stack_base register updates on
context switches, and activates the Stack Checking bit of STATUS32
register when the CPU is in the context of a fiber or task.
As GCC accesses the non-yet allocated stack with frame pointer enabled,
this patch also add the omit-frame-pointer gcc flag in order to work
properly.
Change-Id: Ia9e224085a03bd29d682fb8f51f8e712f2ccb556
Signed-off-by: Alexandre d'Alton <alexandre.dalton@intel.com>
Changed names of Kconfig flags, variables, functions, files and
return codes consistent with names used in the RFC. Updated
relevant comments to match the changes.
Origin: Original
Change-Id: Ie7941032d7ad7af61fc02928f74538745e7966e8
Signed-off-by: Ramesh Thomas <ramesh.thomas@intel.com>
Signed-off-by: Anas Nashif <anas.nashif@intel.com>
The thread monitor allows to iterate over the thread context
structures for each existing thread (fiber/task) in the system.
Thread context structures do not expose thread entry information
directly. Although all the information can be scavenged from memory
stacks. Besides, accessing the information depends on the stack
implementation for each architecture.
By extending the tcs we allow a direct access to the thread
entry point and its parameters, only when thread monitor is
enabled.
It also allows a task to access its kernel task structure
through the first parameter of the thread.
This allows a debugger application to access the information directly
from the thread context structures list.
Change-Id: I0a435942b80eddffdf405016ac4056eb7aa1239c
Signed-off-by: Juan Manuel Cruz <juan.m.cruz.alcaraz@intel.com>
It is semantically identical to CONFIG_SW_ISR_TABLE.
Change-Id: Iff0c47166ee6fb1fd8a0991a67bc863d45c32559
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
The app-facing interface for configuring interrupts was never
formally defined, instead it was defined separately for each arch
in their respective arch-specific header files. Occasionally these
would go out of sync.
Now there is a single irq.h header which defines this interface.
To avoid runtime overhead, these map to _arch_* implementations of
each that must be defined in headers pulled in by arch/cpu.h.
Change-Id: I69afbeff31fd07f981b5b291f3c427296b00a4ef
Signed-off-by: Andrew Boie <andrew.p.boie@intel.com>
