Fixes test timeouts like this one:
https://gitlab.gnome.org/GNOME/glib/-/jobs/4827270
The race will continue to be reproduced when running the tests not under
valgrind.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
Some tweakings of the time spend during warm up. That mostly matters if
you set very short "--seconds", which can make sense for quickly
checking something. Then the warmup should not take more thatn a certain
percentage of the requested runs.
When we have a constant factor, we still want not to run for more than
10% of the overall test time ... except, we still want to run at least
ESTIMATE_ROUND_TIME_N_RUNS (because we skip the estimation step below).
Also, adjust WARM_UP_ALWAYS_SEC to be only 20% of the test time, for
short test runs.
Also, don't print the messages about "Estimating round time" with a
fixed "--factor".
One test round aims to run for 8msec (TARGET_ROUND_TIME).
As the "--seconds" parameter previously took whole integer numbers, that
meant that we would run at least 125 rounds.
For a quick run, we should also support even faster runs, e.g. to select
only 0.5 seconds.
Historically, there was a verbose mode and a non-verbose mode.
In non-verbose mode (the default), we would still print two lines:
Running test property-set
Property set per second: 39329344
Later, this was changed to include the test name in the second line, so
we would print:
Running test property-set
property-set: Property set per second: 39329344
But this first line is really just noise, making parsing and reading the
results harder. Hence a "--quiet" mode was added, that only printed one
line per test while keeping the previous default behavior. And all was
good.
Except, unless you want verbose mode, this "Running test" line is still
not very useful and mainly clutters the output.
Supporess it now also in normal mode. It is now only printed in verbose
mode.
This also makes the "--quiet" option do nothing. The option is still
there, maybe we find a future use and we should not break the command
line API by dropping an argument.
g_object_set() optimizes the case where there are no signals connected.
Add a test that sets the property with signals. Obviously, this one is
much slower, since we will freeze and thaw the notifications.
For the test, we actually care to find the fastest test run (and take
"min_elapsed"). That is useful, because that is the run where we
possibly have the least interference from external factors, it was the
run where the CPU solved the problem as fast as it could.
As such, we should not reject the first 5% as additional warm up. If the
first 5% are slower (and part of "warmup"), then they are anyway not
considered. If there is a the fastest run in the first 5 percent, then
we want to take that.
Also note, that the calculation of "avg_elapsed" was wrong, since it
divided by the full "num_rounds" while only summing 95% of the runs.
This is fixed too by now considering all runs.
Fixes: 282d536fd229 ('tests/performance: ensure to always warm up for 2 seconds')
- fix and improve usage output for "performance-run.sh" script.
- add a sleep after compilation. It seems to me, that the first run
usually performs better, which might be because the temperature of the
CPU raises and the CPU gets throttled. Unclear whether that is really
the case, but add a sleep so that all runs start under similar
conditions where the CPU was idle for a moment.
When running the test (without parameters), it estimates a factor for
the run size for each test. That is useful for running a reasonable size
of the test, on different machines.
However, when comparing two runs, it seems important that both runs
share a common factor. Otherwise, the factor is determined differently,
and the test is less comparable. For that there is the "--factor" option
or the GLIB_PERFORMANCE_FACTOR environment variable.
However, the factor option can only set the factors for all tests at the
same time. Optimally, one factor is roughly suitable for all tests, but
it is not, as currently the detected factors on my machine are widely
different
$ ./build/gobject/tests/performance/performance -v > p
$ cat p | sed -n -e 's/^Running test //p' -e 's/.*correction factor //p' | sed 'N;s/\n/ /'
simple-construction 34.78
simple-construction1 145.45
complex-construction 11.08
complex-construction1 20.46
complex-construction2 23.74
finalization 4.74
type-check 37.74
emit-unhandled 5.63
emit-unhandled-empty 49.69
emit-unhandled-generic 7.17
emit-unhandled-generic-empty 50.63
emit-unhandled-args 5.20
emit-handled 3.86
emit-handled-empty 4.01
emit-handled-generic 3.96
emit-handled-generic-empty 7.04
emit-handled-args 3.78
notify-unhandled 52.63
notify-by-pspec-unhandled 156.86
notify-handled 2.55
notify-by-pspec-handled 2.66
property-set 34.63
property-get 32.92
refcount 0.83
refcount-1 2.30
refcount-toggle 1.33
Adjust the base factors with these measurements.
PERFORMANCE_FILE="./gobject/tests/performance/performance.c"
IFS=$'\n'
for LINE in $(cat p | sed -n -e 's/^Running test //p' -e 's/.*correction factor //p' | sed 'N;s/\n/ /') ; do
(
IFS=' '
set -- $LINE
TESTNAME="$1"
FACTOR="$2"
LINENUMBER="$(grep -n "^ \"$TESTNAME\",$" "$PERFORMANCE_FILE" | cut -d: -f1)"
LINENUMBER=$((LINENUMBER + 2))
OLD_FACTOR="$(sed -n "$LINENUMBER s/^ \([0-9]\+\),$/\1/p" "$PERFORMANCE_FILE")"
NEW_FACTOR="$(awk -v factor="$FACTOR" -v old_factor="$OLD_FACTOR" 'BEGIN {print int(factor * old_factor + 0.5)}')"
sed -i "$LINENUMBER s/^ \([0-9]\+\),$/ $NEW_FACTOR,/" "$PERFORMANCE_FILE"
)
done
Afterwards, we get comparable factors:
$ ./build/gobject/tests/performance/performance -v > p2
$ cat p2 | sed -n -e 's/^Running test //p' -e 's/.*correction factor //p' | sed 'N;s/\n/ /'
simple-construction 0.98
simple-construction1 0.75
complex-construction 0.99
complex-construction1 0.96
complex-construction2 1.02
finalization 1.05
type-check 0.98
emit-unhandled 1.01
emit-unhandled-empty 1.10
emit-unhandled-generic 1.03
emit-unhandled-generic-empty 1.07
...
Of course, this measurement was taken in my setup. But I think it
brings the base factors into a comparable range for most users.
Also, the commit message shows an ugly script how you can re-generate
this for your own purposes.
Move the factor inside the PerformanceTest structure, so it can be
programatically accessed.
More importantly, the number is now expressed directly beside the test
setup (the PerformanceTest structure), all at one place.
Also, each test now gets a separate factor.
This change will be useful in the next commit. So far there is no
notable change in behavior.
Despite assigning the function to a variable, gcc can still detect that
the function never changes and most of the test code is optimized out.
Initialize it somewhere, where the compiler cannot prove that this
function pointer is always set to the same value.
We could also make the pointer volatile, but this approach seems
preferable to me.
During "as-installed" testing, we should search the GIR_DIR for GIR XML,
instead of hard-coding that it is `${prefix}/share/gir-1.0`. This is
not the case on at least Debian, in order to make it possible to
install more than one architecture's flavour of `GLib-2.0.gir`,
which contains some architecture-specific `#define`s.
Also search GOBJECT_INTROSPECTION_DATADIR/GIR_SUFFIX (in practice
something like `/usr/share/gir-1.0` in all cases) to accommodate
distributions like Debian that move the architecture-independent
majority of GIR XML into /usr/share to avoid duplication, leaving
only the architecture-specific minority of files like `GLib-2.0.gir`
in the GIR_DIR.
Signed-off-by: Simon McVittie <smcv@collabora.com>
We were reusing the same logic everywhere, while we can just reuse an
unique class to base our tests on that avoids having to copy-and-paste
code for no good reason
Add some basic support for having glib-tests-only python libraries that
can be shared across the various projects, so that we don't have to
maintain multiple copies of them.
It’s a false positive, but points to a slightly unnecessary use of a
global variable to store something which could be computed per-test.
scan-build thought that arrays of size `n_handlers` could have
uninitialised values accessed in them if `n_handlers` were to change
value part-way through a test (which it didn’t).
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
It’s deprecated, but I was modifying it anyway and it didn’t have any
coverage, so let’s add a simple test (as suggested by Michael
Catanzaro).
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
If we enable `-Wfloat-conversion`, these warn about a possible loss of
precision due to an implicit conversion from `double` to some other
numeric type.
The warning is correct: there is a possible loss of precision here. In
these instances, we don’t care, as the floating point arithmetic is
being done to do some imprecise scaling or imprecise timing. A loss of
precision is not a problem.
So, add an explicit cast to squash the warning.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
Helps: #3405
The tests were using a lot of signed `int`s when actually the values
being handled were always non-negative. Use `unsigned int` consistently
throughout.
Take the opportunity to move declarations of loop iterator variables
into the loops.
This introduces no functional changes.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
Helps: #3405
There were various places where (signed or unsigned) integer assertions
were being used to compare `double` or `float` values, resulting in an
implicit integer conversion.
This causes a warning when building with `-Wfloat-conversion`.
Improve the specificity of the tests by using the most-specific numeric
assertions through all `param` tests.
For the conversion tests, this means using the assertion function
associated with the target type, not the source type.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
Helps: #3405
As spotted by `-Wfloat-conversion`. Doubles which could not be
accurately represented as floats may have erroneously failed bounds
validation.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
Helps: #3405
These consistently fail on scheduled CI runs, which is not helping our
ability to catch Hurd regressions.
For example, https://gitlab.gnome.org/GNOME/glib/-/jobs/3709402
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
See: #3148
Bumping the reference count from 1 to 2 (and back) is more expensive,
due to the check for toggle notifications.
We have a performance test already that hits that code path. Avoid that
for he "property-{get,set}" tests, so we avoid the known overhead and
test more relevant parts.
Despite all the efforts, there still seems to be a lot of noise in the
performance measurement. Especially, the first iterations seem to run
faster. Maybe that is because the kernel didn't yet determine that the
process is CPU bound and is less likely to schedule it out Or maybe it's
because burning the cycles heats up the CPU and it gets throttled after
a while. It's unclear why, and it's even unclear whether this really
happens. But from my observations, it seems to do.
Hence, more warm up.
- the first time we enter the test, ensure that we keep the CPU busy for
at 2 seconds. This additional warm up (WARM_UP_ALWAYS_SEC) is
global, and not per test.
- for each test, ignore the first 5% of the runs. It seems those tend to
run faster, thus skewing the results.
- if the user specifies a "--factor", the warm up operations are the
same and independent from external factors (such as time
measurements).
Note that this matters the most, when you want to run the executable
twice in a row and compare the results.
By default, the test estimates a run factor for each test. This means,
if you run performance under `perf`, the results are not comparable,
as the run time depends on the estimated factor.
Add an option, to set a fixed factor.
Of course, there is only one factor argument for all tests. Quite
possibly, you would want to run each test individually with a factor
appropriate for the test. On the other hand, all tests should be tuned
so that the same factor gives a similar test duration. So this may not
be a concern, or the tests should be adjusted. In any case, the option
is most useful when running only one test explicitly.
You can get a suitable factor by running the test once with "--verbose".
Another use case is if you run the benchmark under valgrind. Valgrind
slows down the run so much, that the estimated factor would be quite
off. As a result, the chosen code paths are different from the real run.
By setting the factor, the timing measurements don't affect the executed
code.
The default output is annoyingly verbose. You see
Running test simple-construction
simple-construction: Millions of constructed objects per second: 33.498
Running test simple-construction1
simple-construction1: Millions of constructed objects per second: 142.493
Running test complex-construction
complex-construction: Millions of constructed objects per second: 14.304
Running test complex-construction1
...
where the "Running test" lines just clutter the output. In fact so much
so, that my terminal fills up and I don't see the output of all tests in
one page. The "Running test" line is not so useful, because I mostly
care about the test result, and that line already contains the test
name.
Add an option to silence this.
Previously, the result lines are not unique, for example
Running test simple-construction
Millions of constructed objects per second: 27.629
Running test simple-construction1
Millions of constructed objects per second: 151.879
...
That is undesirable, because we might want to parse the test results
with a script, and that's easier when the line is unique.
Change to:
Running test simple-construction
simple-construction: Millions of constructed objects per second: 27.629
Running test simple-construction1
simple-construction1: Millions of constructed objects per second: 151.879
...
On a fast laptop, this test currently takes about 7s to run, which is a
significant portion of the overall test suite time.
On a slower CI machine, especially running the test under valgrind, the
test can time out.
There’s no need to always run so many iterations: we run the tests under
CI so often that it’s likely a failure will eventually be hit (if there
is a bug) even with fewer iterations. We also now run the tests once a
week with `-m slow`, so the original iteration count will also still be
used then.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
It’s more helpful to always register the test, even if it’s normally
skipped, since then the skip is recorded in the test logs so people can
see what’s ‘missing’ from them.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
GSList doesn't seem the best choice here. It's benefits are that it's
relatively convenient to use (albeit not very efficient) and that an
empty list requires only the pointer to the list's head.
But for non-empty list, we need to allocate GSList elements. We can do
better, by writing more code.
I think it's worth optimizing GObject, at the expense of a bit(?) more
complicated code. The complicated code is still entirely self-contained,
so unless you review WeakRefData usage, it doesn't need to bother you.
Note that this can be easily measure to be a bit faster. But I think the
more important part is to safe some allocations. Often objects are
long-lived, and the GWeakRef will be tracked for a long time. It is
interesting, to optimize the memory usage of that.
- if the list only contains one weak reference, it's interned/embedded in
WeakRefData.list.one. Otherwise, an array is allocated and tracked
at WeakRefData.list.many.
- when the buffer grows, we double the size. When the buffer shrinks,
we reallocate to 50% when 75% are empty. When the buffer shrinks to
length 1, we free it (so that "list.one" is always used with a length
of 1).
That means, at worst case we waste 75% of the allocated buffer,
which is a choice in the hope that future weak references will be
registered, and that this is a suitable strategy.
- on architectures like x86_68, does this not increase the size of
WeakRefData.
Also, the number of weak-refs is now limited to 65535, and now an
assertion fails when you try to register more than that. But note that
the internal tracking just uses a linear search, so you really don't
want to register thousands of weak references on an object. If you do
that, the current implementation is not suitable anyway and you must
rethink your approach. Nor does it make sense to optimize the
implementation for such a use case. Instead, the implementation is
optimized for a few (one!) weak reference per object.
The implementation of GWeakRef tracks weak references in a way, that
requires linear search. That is probably best, for an expected low
number of entries (e.g. compared to the overhead of having a hash
table). However, it means, if you create thousands of weak references,
performance start to degrade.
Add a test that creates 64k weak references. Just to see how it goes.
g_object_weak_ref() documentation refers to GWeakRef as thread-safe
replacement. However, it's not clear to me, how GWeakRef is a
replacement for a callback. I think, it means, that you combine
g_object_weak_ref() with GWeakRef, to both hold a (thread-safe) weak
reference and get a notification on destruction.
Add a test, that GWeakRef is already cleared inside the GWeakNotify
callback.
This isn’t normally hit because it’s in a test which is disabled unless
run with `-m thorough`.
The data is owned by `g_test_add_data_func_full()` until the end of the
process.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
This isn’t normally hit because it’s in a test which is disabled unless
run with `-m thorough`.
The `GParamSpec` is initially floating, but its floating ref is sunk by
`g_object_interface_install_property()` (regardless of whether that call
succeeds or aborts). The behaviour of
`g_object_interface_install_property()` in this respect may have changed
more recently than the test was written.
Signed-off-by: Philip Withnall <pwithnall@gnome.org>
Usually, after g_pointer_bit_lock() we want to read the pointer that we
have. In many cases, when we g_pointer_bit_lock() a pointer, we can
access it afterwards without atomic, as nobody is going to modify the
pointer then.
However, gdataset also supports g_datalist_set_flags(), so the pointer
may change at any time and we must always use atomics to read it. For
that reason, g_datalist_lock_and_get() does an atomic read right after
g_pointer_bit_lock().
g_pointer_bit_lock() can easily access the value that it just set. Add
g_pointer_bit_lock_and_get() which can return the value that gets set
afterwards.
Aside from saving the second atomic-get in certain scenarios, the
returned value is also atomically the one that we just set.
The existing g_pointer_bit_lock() and g_pointer_bit_unlock() API
requires the user to understand/reimplement how bits of the pointer get
mangled. Add helper functions for that.
The useful thing to do with g_pointer_bit_lock() API is to get/set
pointers while having it locked. For example, to set the pointer a user
can do:
g_pointer_bit_lock (&lockptr, lock_bit);
ptr2 = set_bit_pointer_as_if_locked(ptr, lock_bit);
g_atomic_pointer_set (&lockptr, ptr2);
g_pointer_bit_unlock (&lockptr, lock_bit);
That has several problems:
- it requires one extra atomic operations (3 instead of 2, in the
non-contended case).
- the first g_atomic_pointer_set() already wakes blocked threads,
which find themselves still being locked and needs to go back to
sleep.
- the user needs to re-implement how bit-locking mangles the pointer so
that it looks as if it were locked.
- while the user tries to re-implement what glib does to mangle the
pointer for bitlocking, there is no immediate guarantee that they get
it right.
Now we can do instead:
g_pointer_bit_lock(&lockptr, lock_bit);
g_pointer_bit_unlock_and_set(&lockptr, lock_bit, ptr, 0);
This will also emit a critical if @ptr has the locked bit set.
g_pointer_bit_lock() really only works with pointers that have a certain
alignment, and the lowest bits unset. Otherwise, there is no space to
encode both the locking and all pointer values. The new assertion helps
to catch such bugs.
Also, g_pointer_bit_lock_mask_ptr() is here, so we can do:
g_pointer_bit_lock(&lockptr, lock_bit);
/* set a pointer separately, when g_pointer_bit_unlock_and_set() is unsuitable. */
g_atomic_pointer_set(&lockptr, g_pointer_bit_lock_mask_ptr(ptr, lock_bit, TRUE, 0, NULL));
...
g_pointer_bit_unlock(&lockptr, lock_bit);
and:
g_pointer_bit_lock(&lockptr, lock_bit);
/* read the real pointer after getting the lock. */
ptr = g_pointer_bit_lock_mask_ptr(lockptr, lock_bit, FALSE, 0, NULL));
...
g_pointer_bit_unlock(&lockptr, lock_bit);
In general, we must not call out to external, unknown code while holding
a lock. That is prone to dead lock.
g_object_ref() can emit a toggle notification. In g_weak_ref_set(), we
must not do that while holding the lock.
GWeakRef calls g_object_ref() while holding a read lock.
g_object_ref() can emit a toggle notification.
If you then inside the toggle notification setup a GWeakRef
(to the same or another object), the code will try to get
a write lock. Deadlock will happen.
Add a test to "show" that (the breaking part is commented out).
Will be fixed next.
In g_object_unref(), we call _object_unref_clear_weak_locations() before
and after dispose() already. At those places it is necessary to do.
Calling it a third time during g_object_real_dispose() seems not useful,
has unnecessary overhead and actually undesirable.
In particular, because g_object_real_dispose() is the implementation for
the virtual function GObject.dispose(). For subclasses that override dispose(),
it's not well defined at which point they should chain up the parent
implementation (for dispose(), I'd argue that usually they chain up at
the end of their own code). If they chain up at the end, this has no
effect.
This only really matters if you try to register GWeakRef during dipose
and/or resurrect the object.
static void dispose(GObject *object)
{
g_weak_ref_set(&global_weak_ref, object);
global_ref = g_object_ref(object);
G_OBJECT_CLASS (parent_class)->dispose (object);
}
the object was resurrected, but g_object_real_dispose() would clear the
weak ref. That is not desirable, nor does it make sense.
Instead, the virtual function dispose() is called from two places, from
g_object_unref() and g_object_run_dispose(). In both cases, it is
ensured that weak locations are cleared *after* dispatching the virtual
function. Don't do it somewhere in the middle from
g_object_real_dispose().
The previous g_object_unref() was racy. There were three places where we
decremented the ref count, but still accessed the object afterwards
(while assuming that somebody else might still hold a reference). For
example:
if (!g_atomic_int_compare_and_exchange_full ((int *) &object->ref_count,
old_ref, old_ref - 1,
&old_ref))
continue;
TRACE (GOBJECT_OBJECT_UNREF (object, G_TYPE_FROM_INSTANCE (object), old_ref));
/* if we went from 2->1 we need to notify toggle refs if any */
if (old_ref == 2 && OBJECT_HAS_TOGGLE_REF (object))
{
/* The last ref being held in this case is owned by the toggle_ref */
toggle_refs_notify (object, TRUE);
}
After we decrement the reference count (and gave up our reference), we
are only allowed to access object if we know we have the only possible
reference to it. In particular, if old_ref is larger than 1, then
somebody else holds references and races against destroying object.
The object might be a dangling pointer already.
This is slightly complicated due to toggle references and clearing of
weak-locations.
For toggle references, we must take a lock on the mutex. Luckily, that
is only necessary, when the current reference count is exactly 2.
Note that we emit the TRACE() after the ref count was already decreased.
If another thread unrefs the object, inside the TRACE() we might have a
dangling pointer. That would only be fixable, by emitting the TRACE()
before the actual unref (which has its own problems). This problem
already existed previously.
The change to the test is necessary and correct. Before this patch,
g_object_unref() would call dispose() and decrement the reference count
right after.
In the test case at gobject/tests/reference.c:1108, the reference count
after dispose and decrement is 1. Then it thaws the queue notification,
which emits a property changed signal. The test then proceeds to
reference the object again and notifying the toggle reference.
Previously, the toggle reference was notified 3 times.
After this change, the property changed signal is emitted before
decreasing the reference count. Taking a reference then does not cause
an additional toggle on+off, so in total only one toggle happens.
That accounts for the change in the test. The new behavior is
correct.
Aside from checking that we're accessing the global GParamSpecPool
without necessarily initializing GObjectClass, we should also verify
that we're doing so safely without the class init lock.
