QThread::idealThreadCount() now returns 1 if the number
of CPU cores could not be detected.
Change-Id: I60b75c46fbfa2891c28e71fed65589e2ce5a5c17
Reviewed-by: Morten Johan Sørvig <morten.sorvig@qt.io>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Allows setting the stack size for the thread pool
worker threads. Implemented using QThread::stackSize.
Task-number: QTBUG-2568
Change-Id: Ic7f3981289290685195bbaee977a23e0c3c49bf0
Reviewed-by: Morten Johan Sørvig <morten.sorvig@qt.io>
(non static data member initializers)
Also, taking the absolute value of the idealThreadCount()
return value is no longer needed since the function now
returns 1 for the "CPU detection failure" case.
Change-Id: I2214fd15ed24413bba796ead38bbf1355dfd37d9
Reviewed-by: Morten Johan Sørvig <morten.sorvig@qt.io>
A deadline timer is more correct for timeouts.
Also starts the timer before trying to acquire the mutex for more
precise timeouts.
Task-number: QTBUG-58745
Change-Id: I230266a3a5d7b7af33981efed4e6882e5727a41c
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
A QFreeList's MaxIndex is clearly meant to be the sum of its Sizes;
but the parentheses in MaxIndex - (16 - 128 - 1024) doesn't give the
same as without the parentheses. Noticed while reviewing a copy of
this code now used by QReadWriteLock (which has this fix).
Change-Id: I9f68171475b3c6edc18896056b5532183fee92b2
Reviewed-by: Sean Harmer <sean.harmer@kdab.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
MinGW with win32 threads might not have _GLIBCXX_HAS_GTHREADS, so it
needs to be checked.
Amends change 597d4ff796.
Change-Id: I5c91d1dd3967c4922ced315b475d653b8b102ad6
Reviewed-by: Giuseppe D'Angelo <giuseppe.dangelo@kdab.com>
Compilation and link times in CPU seconds with GCC 7, using precompiled
headers (not including moc, rcc, uic, etc. steps or headersclean):
Before After
Debug -O0 198,1 180,3
Debug -Og 240,7 229,2
Release -O3 267,1 249,2
Release LTO 239,4 229,8
QtCore required a little manual adjusting because some files are
bootstrapped into moc itself and into qmake.
Change-Id: I84e363d735b443cb9beefffd14b8b57c10e7da36
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
Align ourselves to the Standard Library and call std::terminate if an
exception leaves the thread entry point (that is, run()).
On platforms using pthreads, thread cancellation needs to be taken in
special consideration, since it looks like it was supported before.
On Glibc, and when using C++, pthread_cancel and pthread_exit are
implemented by throwing a special kind of exception that can be caught,
but must always be rethrown. That exception is then used to activate the
cancellation clean-up handlers. (This is non-Standard C++ behavior.)
So: mimic what libstdc++'s std::thread does to support Glibc's pthread
cancellation.
At this time, it looks like libc++ has no support for this, and when
used in combination with Glibc a thread cancellation results in a crash
(also because it does not seem to terminate() when exceptions leave the
thread).
[ChangeLog][QtCore][QThread] An exception escaping from QThread::run()
will now result in immediate and abnormal program termination. The same
applies if an exception leaves a slot connected directly to the
QThread::started() or QThread::finished() signals.
Change-Id: I73cc93cf06c57018e149a578cc9d4cd0d6fc00ef
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
GCC's -Wshadow complained about the re-use of 'args' as the parameter
for the lambda.
Fix by renaming it to largs (lambda-args) instead.
Change-Id: Ia9e65b82b98ee56181d754868354988399496eef
Reviewed-by: Giuseppe D'Angelo <giuseppe.dangelo@kdab.com>
In the spirit of std::thread, which takes a function to call and its
parameters, and runs it in a new thread. Since the user might want to
connect to signals, move QObjects into the new thread, etc., the new
thread is not immediately started.
Although technically all of this _should_ be implementable in pure
C++11, there is nothing in the Standard to help us not reinvent all the
plumbing: packing the decay'd parameters, storing them, invoking the
function over the parameters (honoring INVOKE/std::invoke semantics).
std::function does not do the job, as it's copiable and therefore does
not support move-only functors; std::bind does not have INVOKE
semantics.
I certainly do not want to reimplement all the required facilities
inside of Qt. Therefore, the full blown implementation requires C++17
(std::invoke).
In order to make this useful also in pre-C++17, there are two additional
implementations (C++11 and C++14) that support just a callable, without
any arguments passed to it. The C++11 implementation makes use of a
class to store and call the callable (even move-only ones); basically,
it's what a closure type for a C++14 lambda would look like.
An alternative implementation could've used some of the existing
facilities inside QObject::connect implementation that store a functor
(for the connect() overload connecting to free functions), namely:
the QtPrivate::QFunctorSlotObject class. However:
* QFunctorSlotObject does not support move-only callables (see
QTBUG-60339);
* QFunctorSlotObject itself is not a callable (apparently by design),
and requires to be wrapped in a lambda that calls call() on it;
* the moment QTBUG-60339 is solved, we'd need the same handwritten
closure to keep QFunctorSlotObject working with move-only callabes.
So: just use the handwritten one.
The C++14 implementation is a simplified version of the C++11 one,
actually using a generalized lambda capture (corresponding to the
handwritten C++11 closure type).
All three implementations use std::async (with a deferred launch policy,
a nice use case for it!) under the hood. It's certainly an overkill for
our use case, as we don't need the std::future, but at least std::async
does all the plumbing for us.
[ChangeLog][QtCore][QThread] Added the QThread::create function.
Change-Id: I339d0be6f689df7d56766839baebda0aa2f7e94c
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
That's before the return type or static, inline, constexpr or such
keywords (if any).
Perl Script:
s/^(\s+)(.*) Q_REQUIRED_RESULT(;)?(\s*\/\/.*)?$/\1Q_REQUIRED_RESULT \2\3\4/
Change-Id: I7814054a102a407d876ffffd14b6a16182f159e2
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
ResultStore never actually exists, only ResutStoreBase does. So casting to
ResultStore<T> and calling its member functions is UB. Put the type dependent
function as template member functions within ResultStoreBase and so we don't
need QtPrivate::ResultStore anymore.
Same goes for the iterator.
Change-Id: I739b9d234ba2238977863df77fde3a4471a9abd2
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
The cancel() function added in 5b11e43e for Qt 5.5 suffers from a
number of problems:
First, if runnable->autoDelete() is true, then the function suffers
from the ABA problem (see documentation written for trytake()).
Second, if runnable->autoDelete() is false, due to cancel() throwing
away crucial information instead of returning it, the caller cannot
know whether the runnable was canceled (and thus has to be deleted),
wasn't found or is currently executing (and thus mustn't be deleted),
or has finished executing (and can be used to extract the result).
Deprecate this dangerous API and replace it with the much more useful
Private::stealRunnable(), promoted to public API and renamed to
tryTake() for consistency with the rest of Qt.
Described the various caveats in the function's documentation.
[ChangeLog][QtCore][QThreadPool] The cancel() function suffers from
several subtle issues and has been replaced with a new tryTake()
function.
Change-Id: I93125935614087efa24b3e3969dd6718aeabaa4f
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
An Acquire barrier in QReadWriteLocker::unlock was missing to synchronize
with the testAndSetOrdered on d_ptr in the lock functions.
The race is between the write of d->writerCount in tryLockForWrite,
and the read in unlock. The acquire on d->mutex is not enough because
it is not on the same object. While that race could be fixed by taking
the newly-allocate()ed d's mutex before publishing it with testAndSet,
there's another race, on 'recursive', between a newly-minted Private*
with recursive == false in tryLockForWrite(), and the read of 'recursive'
in unlock().
Task-number: QTBUG-58917
Change-Id: I10ba36573c0e57468d11e9b77d85045711feaea1
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
It is never written after initialization, and accessed concurrently
outside mutex protection, so make sure that it never changes by
marking it as const.
Change-Id: Ib28b2e8b3067d596a95d1699165a6ebfb15fe674
Reviewed-by: David Faure <david.faure@kdab.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Solves a data race found by TSan.
Since thread and threadId are QAtomicPointer, I've removed the explicit
initialization in the QThreadData constructor
Task-number: QTBUG-58855
Change-Id: I4139d5f93dcb4b429ae9fffd14a34082f2683f76
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
$ git grep -we futexFlagSupport
src/corelib/thread/qmutex_linux.cpp:static QBasicAtomicInt futexFlagSupport = Q_BASIC_ATOMIC_INITIALIZER(-1);
The last user of this variable was removed in 9ef59b5.
Change-Id: I818a13a481ad25baa5ff7d389a737b8801adcfcc
Reviewed-by: David Faure <david.faure@kdab.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
This is a simple RAII class that makes semaphore releasing
reliable in the face of exceptions and early returns.
This code originates from KDTools' KDSemaphoreReleaser[1], but
has been extensively reworked to support C++11 move semantics.
[1] https://docs.kdab.com/kdtools/2.3.0/class_k_d_semaphore_releaser.html
[ChangeLog][QtCore][QSemaphore] Added a new RAII class, QSemaphoreReleaser,
to reliably perform release() calls.
Change-Id: I6aff64d37cc0882b17c4419817bde60b542f34d9
Reviewed-by: Mitch Curtis <mitch.curtis@qt.io>
Reviewed-by: David Faure <david.faure@kdab.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
QThreadPool maintains three containers of QThreadPoolThread*:
- allThreads, a QSet
- waitingThreads, a QQueue
- expiredThreads, also a QQueue
None of the operations on allThreads make use of QSets fast lookup.
The only functions called on it are isEmpty(), count(), insert(),
and swap().
Since therefore QSet adds nothing but overhead, causes indeterminism
(e.g. when deleting threads in Private::reset()) and code bloat, use
the same container for allThreads that underlies QQueue: QList.
Port insert() to append(). Add an assert to verify that we're not
running into an ABA problem here (but this should never fire, since
we're never deleting threads except in Private::reset(), where we
do remove them from allThreads), just in case.
Saves ~0.5KiB in text size on optimized Linux AMD64 GCC 7.0 builds.
Change-Id: I53a4d5ef2c204420f7c8852f1e72ab3d6ea43d08
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
qcore_foundation.mm:
- Can't link to 'fromCGPoint()'
- Undocumented parameter 'point' in QPointF::fromCGPoint()
- Can't link to 'fromCGRect()'
- Undocumented parameter 'rect' in QRectF::fromCGRect()
- Can't link to 'fromCGSize()'
- Undocumented parameter 'size' in QSizeF::fromCGSize()
Change-Id: Ie48f04c7b990634f8c5a836100b1be7854848bb4
Reviewed-by: Venugopal Shivashankar <Venugopal.Shivashankar@qt.io>
error: base class ‘class QBasicAtomicPointer<void>’ should be explicitly initialized in the copy constructor [-Werror=extra]
Change-Id: I2bc52f3c7a574209b213fffd149b4b71f3006be5
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
INTEGRITY doesn't support self-extending stack. The default stack
size for a pthread on INTEGRITY is too small so we have to increase
the default size.
Change-Id: I0787d14938cf5e7e96c35df204212c8e83aa8893
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Reviewed-by: Nikola Velinov <nvelinov@ghs.com>
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
qmutex.h is updated to let clangqdoc document the threads case,
because the no-threads case is not interesting, and clang can
handle everything declared in qmutex.h. This change required
that a few minor qdoc errors be corrected in qmutex.cpp as well.
Change-Id: Icb4122f2179d6aad39dc68376498364820143297
Reviewed-by: Martin Smith <martin.smith@qt.io>
Add parentheses after function names, replace
is different -> differs
Change-Id: I6332db1d1650ed8d8320c5f20cd79d0bf1870e27
Reviewed-by: Mitch Curtis <mitch.curtis@qt.io>
By templating on the <chrono> types and unconditionally using
duration_cast to coerce the duration into a milliseconds, we
allowed code such as
mutex.try_lock_for(10us)
to compile, which is misleading, since it's actually a zero-
timeout try_lock().
Feedback from the std-discussions mailing list is that the
wait_for functions should wait for _at least_ the duration
given, because that is the natural direction of variance
(tasks becoming ready to run might not get a CPU immediately,
causing delays), while an interface that documents to wait
_no more_ than the given duration is promising something it
cannot fulfill.
Fix by converting the given duration to the smallest number
of milliseconds not less than the original duration. If that
is not representable in an int, use INT_MAX, emulating the
effect of a spurious wakeup, which are allowed to happen if
the function returns false in that case.
In the above example, the try_lock_for call is now equivalent
to
mutex.tryLock(1);
The tryLock() docs state that the actual waiting time does
not exceed the given milliseconds, but fixing that is a
separate issue.
Change-Id: Id4cbbea0ecc6fd2f94bb5aef28a1658be3728e52
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Introduce helper functions switch_{on,off,from_to} to make
the code more readable, and prepare everything for later
optimizations reducing the sizes of critical sections (by
locking the mutex later, or even never).
This commit, however, is only concerned with shutting up
tsan.
In waitForResult(), simplified the code by removing an
unneeded if guard: the condition is checked in the while
loop immediately following in the then-block, and the
local variable declaration that precedes the loop is not
worth guarding.
Change-Id: I24bfd864ca96f862302536ad8662065e6f366fa8
Reviewed-by: David Faure <david.faure@kdab.com>
Remove most type traits from qtypetraits.h, but keep the custom
implementation of is_signed/is_unsigned. This gets rid of
BSD-3 licensed code from Google in a public header (hugh!).
The custom implementations for is_signed/is_unsigned are kept
because the implementations in gcc's standard headers do not
work as we expect for enums - both is_signed and is_unsigned
always returns false there - see also
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59027
[ChangeLog][QtCore][General] Qt now relies on type traits from
the C++ standard library.
Change-Id: I3f2188b46949f04ca4482a6ac9afd3482103f0e1
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Morten Johan Sørvig
Liang Qi
Allan Sandfeld Jensen
Edward Welbourne
Friedemann Kleint
Thiago Macieira
Giuseppe D'Angelo
Marc Mutz
Olivier Goffart
Nico Vertriest
Alexander Volkov
Kimmo Ollila
Martin Smith
Ulf Hermann
Oliver Wolff
Kai Koehne