The original design for the 64-bit futex had the token count replicated
in the high part. But the constructor wasn't setting it. This is one of
the issues I had noticed when investigating QTBUG-66875, but didn't need
to address since the the fix I ended up applying in commit
081c001deb made that unnecessary: the high
part only had the number of waiters.
Unfortunately, when commit f502270c0f06daba75c1da381bd1658d81aa7bba
brought back the token count in the high part, I failed to correct that
problem. As a consequence, every QSemaphore that was initialized in the
constructor to a non-zero value would eventually deadlock.
This commit fixes that oversight.
Task-number: QTBUG-67481
Change-Id: I662b8f168c74440ab1a8fffd1522be6b85adb4d0
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
Reviewed-by: Svenn-Arne Dragly <svenn-arne.dragly@qt.io>
This time, the issue was that we could race a wait and a wake. A multi-
token waiter would not notice that the number of tokens changed because
it only performed a fetch-and-OR, then waited on the high part which did
not change. That means the futex_wait() would succeed, when we wanted it
to have failed.
So we have to bring back a portion of what commit
081c001deb removed: we need to keep both
the token count and the waiter count in the high word.
Task-number: QTBUG-67214
Change-Id: I04a43ee94975482f9e32fffd151e467a9e0030b3
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The future does not actually have a result() member function that takes an int.
The correct function is resultAt(). But that is also available directly in the
QFutureWatcher, so refer to that instead of advising to get to the future.
Change-Id: I53d267b4b48b1171bf611e11130b9dacabc059a4
Reviewed-by: Martin Smith <martin.smith@qt.io>
Since we won't use the high bit of the low 32-bit word at all, we don't
need the AND with 0x7fffffff either. Just cast.
Change-Id: I72f5230ad59948f784eafffd151aa5a7dee995db
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
When the thread that got woken up by release() is supposed to release()
to wake up another thread, we were deadlocking. This happened because we
cleared the bit indicating that there was contention when the first
release(). Instead of storing a single bit, we now store the number of
threads waiting.
Task-number: QTBUG-66875
Change-Id: I72f5230ad59948f784eafffd15193873502ecba4
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
No longer needed. The comment about missing constexpr support is
incorrect: MSVC 2015 does have constexpr issues, but they don't affect
our use of std::atomic.
Change-Id: Ie9d9215342d449c48a11fffd151d11208137f00d
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
MSVC atomics still use the generic version, instead of qatomic_cxx11.h.
The implementation of fetchAndSub is implemented on top of fetchAndAdd,
but produced a warning with unsigned types.
Change-Id: I72f5230ad59948f784eafffd151aa53435b75298
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
all of these iterator classes already have a member operator+, which allows iter+int.
This commits addes non-member operator+, which allows int+iter, and forwards to the member
QList and QArrayData iterators now satisfy RandomAccessIterator concept
Change-Id: I25c1dd8cea299e735d5a5e288dbe23dc1d7a1933
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Since pthread_cond_timedwait takes absolute time instead of relative
time like most POSIX API, there's a small gain in performance here: we
avoid an extra system call to get the current time.
Task-number: QTBUG-64266
Change-Id: I25d85d86649448d5b2b3fffd1451138568091f50
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
It was added by Sam back in 2007, and we've removed other instances of
the same pattern since then. It doesn't make any sense today.
Change-Id: I0f3cb299e312648fd9dc96c639dab4c77fcb48c7
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Timur Pocheptsov <timur.pocheptsov@qt.io>
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@qt.io>
A step towards having the application do its event dispatching though the
thread data's dispatcher, like QEventLoop, instead of keeping two references
to the same dispatcher, one in QCoreApplicationPrivate and one in QThreadData.
Change-Id: I7b215e7e99869d25638ec67f0666f632a508cc0f
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Leaving the logic of starting up the event dispatcher to the call site,
unified both the case of a custom event dispatcher and the default
event dispatcher.
The data argument is left in due to the static nature of the function.
Change-Id: Ia2020e39ccc67cd5a583d4e614dd978b2ec44dba
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Passing on the thread ID is confusing, as it's not really what the
function does. The QNX code path can resolve the thread ID by itself.
Change-Id: I5f0d54621058576cdcf3707d36a11762fe2383c8
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The fixes included adding missing '!' characters to qdoc comment
markers, correct misspelled words, adding documentation for an
anonymous enum type, and replacing Q_QDOC with Q_CLANG_QDOC.
There remain 12 qdoc link warnings in QtBase.
Change-Id: I00447722e6e029f5aed273b3cd571cef33c119b4
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
This update corrects several qdoc warnings about undocumented
parameters, which are actually caused when the return type is not
included in the \fn command, or when the return type is mis-specified
(includes static, or lacks needed template parameters).
Change-Id: Ic49139b88424e93609fbd01bc0836436d13a8f9a
Reviewed-by: Topi Reiniö <topi.reinio@qt.io>
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
After including windows.h, interface is a define that expands to
"struct" (unless WIN32_LEAN_AND_MEAN is defined). This name is
used as a normal identifier in multiple places within Qt.
This has already been worked around in a number of places
(in e.g. 3ba61d9baa and
786d23bb49).
After qrandom.h was included in <QtCore/QtCore>, this header
implicitly includes <random>. In libc++ on windows, this header
then transitively includes windows.h, exposing the clash with
the name "interface" in even more locations than before.
For cases within qtbase internals, it could also alternatively be
handled by defining WIN32_LEAN_AND_MEAN while building QtDbus, but
for occurrences in public headers, the undef trick needs to be
used.
Change-Id: I89754f38f55ae7f2145255a2c8a71b23492be6a1
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
clang required adding template clauses to a few \fn commands.
|| defined(Q_CLANG_QDOC) was also added in qmutex.h.
Change-Id: I7e61f460a8f8f15032094fb35c02f73721a5eb8a
Reviewed-by: Topi Reiniö <topi.reinio@qt.io>
... and make sure we can compile without it. In particular,
Qt Concurrent depends on QFuture, so we specify it as a condition,
and QtConcurrentException should not depend on future but on
concurrent.
Change-Id: I65b158021cecb19f227554cc8b5df7a139fbfe78
Reviewed-by: Martin Smith <martin.smith@qt.io>
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
This update corrects several uses of #ifdef macros
that needed updating because qdoc now uses clang to
parse header files.
Change-Id: I285efa4629a1a5d5bcbfaf701eeafbd0e9f1e43e
Reviewed-by: Topi Reiniö <topi.reinio@qt.io>
This update corrects about 200 qdoc warnings caused by incomplete
\fn commands for member functions of template classes. It can be
used as an example of how to fix \fn commands that suddenly cause
qdoc warnings now that qdoc uses clang to parse the \fn commands.
For example, with the old qdoc, we had this \fn command, which the
old qdoc handled correctly:
\fn QAtomicInteger::operator T() const
For the new clang-based qdoc, this \fn command must be written this way:
\fn template <typename T> QAtomicInteger<T>::operator T() const
However, the documentation generated by the clang-based qdoc looks the
same as it did in the old qdoc.
Change-Id: I7803b3b7ec7c6b8b3cc1be789bc36921438f527e
Reviewed-by: Topi Reiniö <topi.reinio@qt.io>
This update corrects many qdoc warnings in the documentation
for the QFuture classes caused by incomplete \fn commands.
Template text and parameters was added.
Change-Id: I360c9db191230b19a9b174a43468d3de1eb24549
Reviewed-by: Topi Reiniö <topi.reinio@qt.io>
... to make user code buildable with gcc [-Werror=zero-as-null-pointer-constant].
Change-Id: I309953acd7154511660302aa9826410276cfe41b
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
To signal a thread to cancel, nothing more than a std::atomic_flag is
needed, but the implementation actually used mutexes, and weird
run-state introspection, so we can't just swap it out for a
std::atomic_flag.
Instead, we retain the principal logic, however weird it is, and just
optimize the common case where isInterruptionRequested() is called
from the secondary thread, repeatedly. We add a fast-path that just
checks that d->interruptionRequested is not set. That requires nothing
more than a relaxed atomic load, because there's no new value read
that could be used as a signal to the secondary thread that some
condition changed.
"What signal?", you may ask. Well, one can think of users doing this:
void cancel() {
m_why = tr("&Canceled");
requestIterruption();
}
void run() override {
while (!isInterruptionRequested()) {
doWork();
}
emit progress(100, 100, m_why);
}
We need to keep this code working, at least until Qt 6.
But the code can already now only rely on synchronization if
isInterruptionRequested() returns true. If it returns false, then
requestInterruption() has not been called, yet, and any modifications
done prior to the requestInterruption() call are not visible in the
secondary thead.
So we still lock the mutex, and in general don't change the semantics
of the functions, except that we don't lock the mutex in the case
where the flag wasn't set in the first place.
This makes calling isInterruptionRequested() as cheap as it can get,
assuming a lock-free implementation, of course.
I opted to use a std::atomic<bool> instead of QAtomicInt, as the
latter does not have loadRelaxed()/storeRelaxed(), and because it
future-proofs the code.
Change-Id: I67faf36b8de73d2723f9cdd66c416010d0873d98
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
The standard calculateUnixPriority provides values that are almost
invariably inappropriate with even LowestPriority mapping to
something higher than the priority of any other thread on the
system.
[ChangeLog][QtCore][QThread] Changed how Qt thread priorities are
mapped to QNX system thread priorities.
Task-number: QTBUG-53357
Change-Id: I205035c4ca7dcafabda7a9a9b06cc52c67c6d2b2
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
We do not touch anything mutex-protected in the path towards the qWarning(), so
the mutex lock is not needed. It may actually be harmful, since a message handler
may check isInterruptionRequested(), which would then deadlock.
Otherwise, we're just decreasing the size of the critical section — always a
worthwhile goal.
Change-Id: I26aa7e3dc087ff7efaccff1d4dc788ba00ba183f
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
The issue was introduced by eaee1209f0, so
it affected only 5.9.2.
[ChangeLog][QtCore][QSemaphore] Fixed a regression that would make
tryAcquire() not to wait forever if the timeout was a negative
value. Note: new code is advised to only use -1 to indicate "forever",
as some other functions taking timeout periods do not accept other
values.
Task-number: QTBUG-64413
Change-Id: I57a1bd6e0c194530b732fffd14f58fce60d5dfc9
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Commit 02dc39fa8e added the constructor
for the bootstrapped mode. For the regular mode, we hadn't needed, since
the {} syntax guaranteed initialization for us.
Turns out there's at least one compiler that doesn't think it was enough
(GCC for QNX 7).
Task-number: QTBUG-64451
Change-Id: Ic632b4163d784b83951cfffd14f6766b4cb4eb64
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Reviewed-by: Liang Qi <liang.qi@qt.io>
ultrix and reliant have not seen a release since 1995. dgux not since
2001. bsdi not since 2003. irix not since 2006. osf not since 2010.
dynix... unclear, but no later than 2002. symbian needs no mention.
All considered obsolete, all gone.
sco and unixware are effectively obsolete. Remove them until someone
expresses a real need.
Change-Id: Ia3d9d370016adce9213ae5ad0ef965ef8de2a3ff
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
We don't actually need it in that case, but as the code that uses it is
disabled by a constant expression we cannot disable the variable itself
by a macro. The static_cast makes sure the compiler does not complain
about implicitly casting a 64bit value to a 32bit one.
thread/qsemaphore.cpp:156:59: error: large integer implicitly truncated to unsigned type [-Werror=overflow]
Task-number: QTBUG-64261
Change-Id: I96f53e28b290e57033737b4f994f8af5b5666587
Reviewed-by: Liang Qi <liang.qi@qt.io>
It's perfectly benign, but I spent a lot of time debugging this and
trying to figure out how to solve something that didn't need solving. So
document for posterity.
For an adopted thread, the TLS destructors or the adopted thread watcher
on Windows will call QThreadData::deref():
- QThreadData::deref(), count drops to zero
-> delete this;
- ~QThreadData() deletes the QAdoptedThread
-> delete t;
- ~QThreadPrivate() calls deref() again
-> data->deref();
- QThreadData::deref(), count drops to -1, no action taken
- ~QObjectPrivate() calls deref() yet again
-> threadData->deref()
- QThreadData::deref(), count drops to -2, no action taken
Change-Id: Icaa86fc7b54d4b368c0efffd14ee448e0796e8d7
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@qt.io>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
By judiciously positioning of the bits, we can optimize for the case of
threads trying to acquire a single token, which is what QSemaphore
should be mostly used for, as it matches the POSIX Semaphore API
(sem_wait, sem_timedwait and sem_trywait). If there are only waiters
waiting for a single token, we know that adding n tokens means n threads
can wake up.
This optimizes for multi-token waiters too. For example, if we have 50
single-token waiters and 50 multi-token waiters, a sem.release(5) will
wake up 55 threads instead of 100.
Change-Id: I209fcd5dbc2b4e5381cffffd14de5550c75d2600
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
QBasicMutex and QMutex are the same in bootstrap mode.
Change-Id: Icaa86fc7b54d4b368c0efffd14eed63343ddb51b
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
