When CONFIG_TIMEOUT_64BIT is y, positive values are relative/delta
timeouts and negative values are absolute timeouts, except
for two special values. -1 is K_WAIT_FOREVER and 0 is K_NO_WAIT.
The reserved value of -1 means INT64_MAX is not a valid argument
to K_TIMEOUT_ABS_TICKS(), but there was no check. If a literal
was passed, a preprocessor/compiler warning would be generated
for overflow, but if a variable was passed as the argument,
then the code would compile but not work correctly since the
absolute timeout would be changed to a relative one. One
example of this is task_wdt_init() if no channels are enabled.
Rather than just fixing task_wdt, and trying to find other cases
in an adhoc way, this CL changes K_TIMEOUT_ABS_TICKS() to
limit the larges value to (INT64_MAX-1). It does so silently,
but given the range of int64_t, there should be no practical
difference.
Also, change the implementation for Z_IS_TIMEOUT_RELATIVE() to
fix the case where INT64_MAX relative timeout was being
improperly reported as being not a relative timeout. This was
again due to the -1 reserved value.
Add some tests for these changes to the timer_api test.
Signed-off-by: Mike J. Chen <mjchen@google.com>
When CONFIG_ZTEST_ASSERT_VERBOSE is 0, the array round_s is not
being used by zassert_true(). So mark it __maybe_unused to
avoid compiler warnings.
Signed-off-by: Daniel Leung <daniel.leung@intel.com>
Fix issues reported by string validation which was added to strings
used in zassert macros.
Signed-off-by: Krzysztof Chruściński <krzysztof.chruscinski@nordicsemi.no>
The new Nordic platforms use GRTC instead of RTC
as the system timer.
Also the nrf54h20 CPUPPR target does not have enough memory
to execute the `timer_behavior` test.
Signed-off-by: Adam Kondraciuk <adam.kondraciuk@nordicsemi.no>
Validate that `K_TIMEOUT_ABS_SEC` works the same way as the other
absolute timeout construction macros.
Signed-off-by: Jordan Yates <jordan@embeint.com>
Test test_timer_remaining is comparing time captured by k_busy_wait and
k_timer_remaining_ticks. Test was accepting 1 tick difference between
those two. If system clock frequency is low (e.g. 100 Hz) 1 tick is a
long time but if system clock is high then 1 tick may not cover for
processing latency. Instead of using fixed 1 tick test is now converting
100 us to ticks (ceiling) to cover for cases where system clock is high.
100 us processing latency time is an arbitrary value.
Signed-off-by: Krzysztof Chruściński <krzysztof.chruscinski@nordicsemi.no>
32-bit multiplication overflows for the 130000 ppm value used currently
on Nordic SoCs and the duration that is configured to 100000 us.
Signed-off-by: Andrzej Głąbek <andrzej.glabek@nordicsemi.no>
On Nordic SoCs, the clock that drives the system timer and the one that
is used in busy-waiting may be significantly skewed, so the test cases
that compare durations derived from those two clocks need to take into
account a proper threshold. After the `z_timeout_expires` function was
corrected in 3d29c9fe54, it turned out
that the threshold was missing in one check and the related test case
started to fail on nRF platforms. This patch adds that threshold there.
Signed-off-by: Andrzej Głąbek <andrzej.glabek@nordicsemi.no>
Let's also explicitly test that absolute timeouts trigger
at the correct time, even if the kernel has not seen
system clock timer announcements for a while.
Signed-off-by: Alberto Escolar Piedras <alberto.escolar.piedras@nordicsemi.no>
rand32.h does not make much sense, since the random subsystem
provides more APIs than just getting a random 32 bits value.
Rename it to random.h and get consistently with other
subsystems.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
When the timer frequency is known at compile time, make sure we can use any
time conversion macro as a global initializer.
Signed-off-by: Keith Packard <keithp@keithp.com>
All of the time_units conversion routines are now macros which means the
test cannot reference them as functions. Instead, create local static
functions which call each one of them and use those instead.
Signed-off-by: Keith Packard <keithp@keithp.com>
SYS_CLOCK_TICKS_PER_SEC of it8xxx2 is 4096 (244us).
Running test_sleep_abs item on it8xxx2 and we get
k_us_to_ticks_ceil32(250) = 2 and late = 2, so it failed.
After we enable the CONFIG_PM, it needs more time to resume
from low power mode, so I modify the logic to <= for passing
the test.
fixes#49605
Signed-off-by: Ruibin Chang <Ruibin.Chang@ite.com.tw>
Add a bunch of missing "zephyr/" prefixes to #include statements in
various test and test framework files.
Signed-off-by: Fabio Baltieri <fabiobaltieri@google.com>
Subtracting with a uint64_t operand yields a uint64_t result, for which
the absolute value is not terribly interesting. Cast the operand to
int64_t.
Use llabs instead of abs as abs takes an int parameter and not an
int64_t. This appears to work even with the minimal C library.
Signed-off-by: Keith Packard <keithp@keithp.com>
In order to bring consistency in-tree, migrate all tests to the new
prefix <zephyr/...>. Note that the conversion has been scripted, refer
to #45388 for more details.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
This patch is testing the test_sleep_abs with a longer
real time slot value. The reason is that for platforms
like stm32wb55rg with PM, the real time slot must be adjusted
because of the LPTIM ticker.
Signed-off-by: Francois Ramu <francois.ramu@st.com>
Extended test to validate that timer API is working as expected
when CONFIG_MULTITHREADING=n.
Signed-off-by: Krzysztof Chruscinski <krzysztof.chruscinski@nordicsemi.no>
The test_timeout_abs case had baked in similar mistakes to the
off-by-one in the absolute timer implementation. FOR THE RECORD:
If you have an absolute timeout expiration set for a tick value "N",
and the current time returned by k_uptime_ticks() is "T", then the
time returned (at the same moment) by any of the *_remaining_ticks()
APIs must ALWAYS AND FOREVER BE EXACTLY "N - T" (also: "N - T > 0"
always, until the moment the kernel ISR hands off control to the first
timeout handler expiring at that tick).
The tick math is exact. No slop is needed on any systems, no matter
whether their clocks divide by milliseconds or not.
The only gotcha is that we need to be sure that the calls don't
interleave with a real time tick advance, which we do here with a
simple retry loop.
But, about slop... This patch also includes a related fix for the
test_sleep_abs(). On an intel_adsp (which has 50 kHz ticks, a
comparatively slow idle resume and interrupt entry, and even has two
CPUs to mess with latency measurements) I would occasionally see the
k_sleep() take more than a tick to wake up from the interrupt handler
until the return to application code. Add some real time slop there
(just 100us) to handle systems like this.
Fixes#32572
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
Absolute timeouts were covered, but nothing was testing their actual
expiration time and there was an off-by-one.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
Correct a bunch of precision/analysis errors in this test:
* Test items weren't consistent about tick alignment and resetting of
the timestamp, so put these steps into init_timer_data() and call
that immediately before k_timer_start().
* Many items would calculate the initial timestamp AFTER
k_timer_start(), leading to an extra (third!) point where the timer
computation could alias by an extra tick. Always do this
consistently before the timer is started (via init_timer-data()).
* Tickless systems with high tick rates can easily advance the system
uptime while the timer ISR is running, so the system can't expect
perfect accuracy even there (this test was originally written for
ticked systmes where the ISR was by definition happening "at the
same time").
(Unfortunately our most popular high tick rate tickless system,
nRF5, also has a clock that doesn't divide milliseconds exactly, so
it had a special path through all these precision comparisons and
avoided the bugs. We finally found it on a x86 HPET system with 10
kHz ticks.)
* The interval validation was placing a minimum bound on the interval
time but not a maximum (this mistake was what had hidden the failure
to reset the timestamp mentioned above).
Longer term, the millisecond precision math in these tests is at this
point an out of control complexity explosion. We should look at
reworking the core OS tests of k_timer to use tick precision (which is
by definition exact) pervasively and leave the millisecond stuff to a
separate layer testing the alternative/legacy APIs.
Fixes#31964 (probably -- that was reported against up_squared, on
which I had trouble reproducing, but it was a common failure on
ehl_crb).
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
Add some error condition or testing cases to verify whether the
robustness of API. Such as give a NULL to some API and check
the response if get result that we were expacted.
Signed-off-by: Jian Kang <jianx.kang@intel.com>
Add a k_usleep() in test_timer_duration_period test to align ticks
before starting the timer. This fixes some rare off-by-1 failures.
Signed-off-by: Daniel Leung <daniel.leung@intel.com>
Move init_timer_data() out of k_usleep() tick alignment.
Compute rem_ticks just after busy_wait_ms() to avoid slew
due to 'now' and 'rem_ms' computations.
With slow CPU 32MHz: -2 Ticks.
Signed-off-by: Alexandre Bourdiol <alexandre.bourdiol@st.com>
Insert k_usleep(1) just before k_timer_start()
to guaranty tick alignment for step "test_timer_k_define"
Signed-off-by: Alexandre Bourdiol <alexandre.bourdiol@st.com>
Add new test cases for timer to improve testing infrastructure.
Add different waiting time in existing cases. For new test cases,
restart timer and check for status of timer.
Signed-off-by: Jian Kang <jianx.kang@intel.com>
Several of the values passed to the conversion failure diagnostic are
unsigned and/or 32-bit values, while all format specifiers are for
signed 64-bit integers. Make the specifiers consistent with the
argument.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
Unit tests were failing to build because random header was included by
kernel_includes.h. The problem is that rand32.h includes a generated
file that is either not generated or not included when building unit
tests. Also, it is better to limit the scope of this file to where it is
used.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
The converted target value for remaining ticks was increased by one to
match original code, which used a one-sided test. The current test is
two-sided, so that increment is already present in the allowed 1 tick
error for boards with no slew, and incorporating it into the absolute
error can cause the test to fail.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
When millisecond/tick conversion is not exact tick delta's are
dependent on the initial tick value. In those cases exact comparisons
need to also allow an adjacent value.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
When HFCLK has a slew making it faster than LFCLK the busy wait can
expire before the timer fires.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
Reduce the duration of the timer test so that it will fire before the
busywait elapses even in the worst case of slew between the tick and
busy-wait clocks.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
This test sets a timer using one clock, waits using a second clock,
then sees whether the remaining time is the expected value. When the
two clocks are skewed the comparison requires a threshold. Provide a
means to estimate the maximum expected error.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
Adjusting the input value to allow round to nearest can cause an
overflow which invalidates the expectation that the 32-bit result is
the low 32 bits of the 64-bit result. If the adjustment overflows do
the full-precision conversion and truncate in the caller.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
When the timer frequency is not a multiple of 1000 converting between
ticks and milliseconds introduces error. Avoid propagating the error
by converting divided values rather than dividing converted values.
Also compensate for observed rate differences between the busywait
clock and the timeout clock.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
A fast timer clock can advance before or after the remaining time
until an event is captured. Verify the expected relationship between
current and remaining time holds for at least one captured current
time.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
When one millisecond is not an integral number of ticks measuring
durations between tick events will sometimes be less than expected to
correct for error that was accumulated between other events. Allow
for that in the duration and period comparisons.
Signed-off-by: Peter Bigot <peter.bigot@nordicsemi.no>
This commit fixes the assertion in test_timer_remaining() that checks
whether the remaining ticks on a timer is less than or equal to half of
the timer duration after a busy wait of that time. If the timer
duration corresponds to an odd number of ticks, 1 should be added to
the upper bound given k_timer_remaining_ticks() returns
<ticks til next deadline> - <elapsed ticks>,
and <elapsed ticks> is truncated to closest integer tick count.
For example, if
dur_ticks = 3277
<ticks til next deadline> = 3277
<elapsed ticks> = 1638.5 rounded to 1638
rem_ticks would be 1639, which is 1 greater than dur_ticks/2.
Fixes#25331
Signed-off-by: Vincent Wan <vincent.wan@linaro.org>
The test of the absolute timeout feature was a simple whitebox test
that inspected the generated ticks field of different constructors for
identity. But it wasn't simple enough, because it was doing a
ticks->ms->ticks conversion (at compile time, sigh) on the input data,
which is obviously lossy on platforms where ticks are shorter than
milliseconds by non-integral factors.
Fix to do the conversion in just one direction.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
This test sets a timer, busy waits for half the duration, and then
checks the remaining time is correct. And it correctly does all its
math in tick precision and aligns to a timer interrupt to eliminate
aliasing due to the tick stride.
But it's waiting using k_busy_wait(), not a timer: "half the duration"
in MICROSECONDS (for k_busy_wait()) is not necessarily representable
as an integer number of TICKS on all platforms. Because k_busy_wait()
always rounds up, we need one extra tick of buffer on those platforms.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
Add support for "absolute" timeouts, which are expressed relative to
system uptime instead of deltas from current time. These allow for
more race-resistant code to be written by allowing application code to
do a single timeout computation, once, and then reuse the timeout
value even if the thread wakes up and needs to suspend again later.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
Kernel timeouts have always been a 32 bit integer despite the
existence of generation macros, and existing code has been
inconsistent about using them. Upcoming commits are going to make the
timeout arguments opaque, so fix things up to be rigorously correct.
Changes include:
+ Adding a K_TIMEOUT_EQ() macro for code that needs to compare timeout
values for equality (e.g. with K_FOREVER or K_NO_WAIT).
+ Adding a k_msleep() synonym for k_sleep() which can continue to take
integral arguments as k_sleep() moves away to timeout arguments.
+ Pervasively using the K_MSEC(), K_SECONDS(), et. al. macros to
generate timeout arguments.
+ Removing the usage of K_NO_WAIT as the final argument to
K_THREAD_DEFINE(). This is just a count of milliseconds and we need
to use a zero.
This patch include no logic changes and should not affect generated
code at all.
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
