summaryrefslogtreecommitdiffstats
path: root/tests/auto/corelib
diff options
context:
space:
mode:
authorGiuseppe D'Angelo <giuseppe.dangelo@kdab.com>2021-11-08 22:24:51 +0100
committerGiuseppe D'Angelo <giuseppe.dangelo@kdab.com>2021-11-26 02:14:52 +0100
commit8c9875893bbd7b1355e36d9501d8cb1f84cb6b4d (patch)
treec399cd37dbccd0abd9e781b7c410e263f0d8d622 /tests/auto/corelib
parent7039fb1e425016ed60383e3752695edbc28a4201 (diff)
Q(Multi)Map: prevent dangling key/value after detach()
Q(Multi)Map mutating functions that take reference to a key and/or a value (e.g. insert(), take(), etc.) must make sure that those references are still valid -- that is, that the referred objects are still alive -- after the detach() call done inside those functions. In fact, if the key/value are references into *this, one must take extra steps in order to preserve them across the detach(). Consider the scenario where one has two shallow copies of QMap, each accessed by a different thread, and each thread calls a mutating function on its copy, using a reference into the map (e.g. map.take(map.firstKey())). Let's call the shared payload of this QMap SP, with its refcount of 2; it's important to note that the argument (call it A) passed to the mutating function belongs to SP. Each thread may then find the reference count to be different than 1 and therefore do a detach() from inside the mutating function. Then this could happen: Thread 1: Thread 2: detach() detach() SP refcount != 1 => true SP refcount != 1 => true deep copy from SP deep copy from SP ref() the new copy ref() the new copy SP.deref() => 1 => don't dealloc SP set the new copy as payload SP.deref() => 0 => dealloc SP set the new copy as payload use A to access the new copy use A to access the new copy The order of ref()/deref() SP and the new copy in each thread doesn't really matter here. What really matters is that SP has been destroyed and that means A is a danging reference. Fix this by keeping SP alive in the mutating functions before doing a detach(). This can simply be realized by taking a local copy of the map from within such functions. remove() doesn't suffer from this because its implementation doesn't do a bare detach() but something slightly smarter. Change-Id: Iad974a1ad1bd5ee5d1e9378ae90947bef737b6bb Pick-to: 6.2 Reviewed-by: Marc Mutz <marc.mutz@qt.io> Reviewed-by: MÃ¥rten Nordheim <marten.nordheim@qt.io>
Diffstat (limited to 'tests/auto/corelib')
-rw-r--r--tests/auto/corelib/tools/collections/tst_collections.cpp68
1 files changed, 68 insertions, 0 deletions
diff --git a/tests/auto/corelib/tools/collections/tst_collections.cpp b/tests/auto/corelib/tools/collections/tst_collections.cpp
index 6d21de9ae7..fd857aa875 100644
--- a/tests/auto/corelib/tools/collections/tst_collections.cpp
+++ b/tests/auto/corelib/tools/collections/tst_collections.cpp
@@ -63,6 +63,9 @@ void foo()
#include <QTest>
#include <QVector>
+#include <QScopedPointer>
+#include <QThread>
+#include <QSemaphore>
#include <algorithm>
@@ -123,6 +126,15 @@ private slots:
void foreach_2();
void insert_remove_loop();
+
+ void detachAssociativeContainerQMap() { detachAssociativeContainerImpl<QMap>(); }
+ void detachAssociativeContainerQMultiMap() { detachAssociativeContainerImpl<QMultiMap>(); }
+ void detachAssociativeContainerQHash() { detachAssociativeContainerImpl<QHash>(); }
+ void detachAssociativeContainerQMultiHash() { detachAssociativeContainerImpl<QMultiHash>(); }
+
+private:
+ template <template<typename, typename> typename Container>
+ void detachAssociativeContainerImpl();
};
struct LargeStatic {
@@ -3545,7 +3557,63 @@ void tst_Collections::insert_remove_loop()
insert_remove_loop_impl<QVarLengthArray<std::string, 15>>();
}
+template <template<typename, typename> typename Container>
+void tst_Collections::detachAssociativeContainerImpl()
+{
+ constexpr int RUNS = 50;
+
+ for (int run = 0; run < RUNS; ++run) {
+ Container<int, int> container;
+
+ for (int i = 0; i < 1'000; ++i) {
+ container.insert(i, i);
+ container.insert(i, i); // for multi-keyed containers
+ }
+
+ const auto it = container.constBegin();
+ const auto &key = it.key();
+ const auto &value = it.value();
+ const auto keyCopy = key;
+ const auto valueCopy = value;
+
+ QSemaphore sem1, sem2;
+ auto detachInAnotherThread = [&sem1, &sem2, copy = container]() mutable {
+ sem1.release();
+ sem2.acquire();
+ copy.clear(); // <==
+ };
+ QScopedPointer thread(QThread::create(std::move(detachInAnotherThread)));
+ thread->start();
+
+ sem2.release();
+ sem1.acquire();
+
+ // The following call may detach (because the container is
+ // shared), and then use key/value to search+insert.
+ //
+ // This means that key/value, as references, have to be valid
+ // throughout the insertion procedure. Note that they are
+ // references into the container *itself*; and that the
+ // insertion procedure is working on a new (detached) copy of
+ // the container's payload.
+ //
+ // There is now a possible scenario in which the clear() above
+ // finds the copy's refcount at 1, hence not perform a detach,
+ // and destroy its payload. But key/value were references into
+ // *that* payload (it's the payload that `container` itself
+ // used to share). If inside insert() we don't take extra
+ // measures to keep the payload alive, now they're dangling and
+ // the insertion will malfunction.
+
+ container.insert(key, value);
+
+ QVERIFY(container.contains(keyCopy));
+ QCOMPARE(container.value(keyCopy), valueCopy);
+
+ thread->wait();
+ }
+}
QTEST_APPLESS_MAIN(tst_Collections)
#include "tst_collections.moc"