Last time, we saw that object destruction is a hidden callout that can cause you to violate the rule against calling out to external code while holding an internal lock. To recap, here’s the class we were using as our example:
class ThingManager
{
private:
std::mutex things_lock_;
std::vector<std::shared_ptr<Thing>> things_;
public:
void AddThing(std::shared_ptr<Thing> thing)
{
std::lock_guard guard(things_lock_);
things_.push_back(std::move(thing));
}
void RemoveThingById(int32_t id)
{
std::lock_guard guard(things_lock_);
auto it = std::find_if(things_.begin(), things_.end(),
[&](auto&& thing)
{
return thing->id() == id;
});
if (it != things_.end()) {
things_.erase(it);
}
}
};
We saw that there was a problem at the call to things_.erase() because the erasure will result in the destruction of the shared_ptr, which could in turn trigger the destruction of an object you don’t control. And that uncontrolled code could try to call back into the ThingManager, resulting in a mess.
Solving this problem can be tricky in general, although things are a bit easier for us in this case because we know some properties of shared_ptr, most importantly that it is noexcept movable.
What we want to do is to remove the element from the vector without destructing it immediately. We want to extend the object’s lifetime until after the lock has been released.
Here’s one way of doing it that works specifically for shared_ptr, taking advantage of its inexpensive default constructor.
void RemoveThingById(int32_t id)
{
std::shared_ptr<Thing> removed;
std::lock_guard guard(things_lock_);
auto it = std::find_if(things_.begin(), things_.end(),
[&](auto&& thing)
{
return thing->id() == id;
});
if (it != things_.end()) {
removed = std::move(*it);
things_.erase(it);
}
}
Before entering the lock, we create an empty shared_ptr and use that to hold the element we intend to remove. C++ destructs objects in reverse order of construction, so declaring it before the guard means that the removed variable destructs after the guard is destructed; in other words, the removed variable destructs outside the lock.
If the container is a std::map or std::unordered_map, you can accomplish this delayed destruction in general by using the extract method to extract the entire node, and then destruct the node outside the lock. For std::list and std::forward_list, you can splice the element out of the container into a temporary initially-empty list. But for containers like vectors, you’ll need to move the object out of the container, in which case you’re counting on the move operation not doing anything scary.
Mind you, you’re already assuming that the move operation isn’t doing anything scary, because your vector mutation operations are going to move objects around, too.
Fortunately, for things like shared_ptr, unique_ptr, and (in Windows) COM wrappers, the move operations are indeed not scary. They just shuffle pointers around. (I’m assuming you’re moving into an empty object. Obviously, if the destination of the move is not empty, then you’re going to run scary code when the previous contents of the destination are released.)
Here are some ways of extracting elements from a container for later destruction:
// map, multimap, set, multiset, // and unordered versions of same auto removed = source.extract(it); // vector, dequeue auto removed = std::move(*it); source.erase(*it); // list, forward_list std::list<T> removed; // or std::forward_list removed.splice(removed.begin(), source, it); return removed;
If you need to delay-destruct multiple items, you could extract them into a temporary vector.
template<typename FwdIt, typename OutIt, typename Pr>
FwdIt extract_if(FwdIt first, const FwdIt last, OutIt out, Pr pred)
{
first = std::find_if_not(first, last, std::ref(pred));
auto next = first;
for (; first != last; ++first) {
if (pred(*first)) {
*out = std::move(*first);
++out;
} else {
*next = std::move(*first);
++next;
}
}
return next;
}
std::vector<T> removed;
v.erase(extract_if(v.begin(), v.end(), std::back_inserter(removed), filter), v.end());
Keep the removed object alive beyond the release of the lock, so that the objects are destructed outside the lock.
Bonus chatter: It would be convenient if vector, list, and forward_list also had extract methods which returned the removed object, similar to the node extractors for the map and set types. Until then, we’ll have to write our own:
template<typename T>
T extract(
std::vector<T>& v,
typename std::vector<T>::iterator it)
{
auto cleanup = wil::scope_exit([&] { v.erase(it); });
return std::move(*it);
}
template<typename T>
std::list<T> extract(
std::list<T>& source,
typename std::list<T>::const_iterator it)
{
std::list<T> removed;
removed.splice(removed.begin(), source, it);
return removed;
}
template<typename T>
std::forward_list<T> extract_after(
std::forward_list<T>& source,
typename std::forward_list<T>::const_iterator it)
{
std::forward_list<T> removed;
removed.splice_after(removed.before_begin(), source, it);
return removed;
}
Exercise: Why did I use a scope_exit in the vector extractor? Why not
template<typename T>
T extract(std::vector<T>& v,
std::vector<T>::iterator it)
{
auto result = std::move(*it);
v.erase(it);
return result;
}
Update: Wrote extract_if because the previous version used remove_if incorrectly.
Is it because of result gets optimized away, and the move doesn’t actually happen until *after* the erase? The scoped exit ensures it runs after the move, as the stack is unwound?
For the removing multiple items example, remove_if suffers from the same issue, it may overwrite some items in the front of the vector so they may get destructed under the lock. Seems like this case requires a home-made algorithm for move-and-remove or filter-out (naming is hard), which does the same as remove_if, but moves all the non-matching items out (including those that the standard remove_if would overwrite)
Oops, thanks. Fixed.
I’m guessing that the answer to the exercise is “vector<bool> is weird”.
T might not have a move constructor, but an expensive copy constructor which then gets called twice.
If T doesn’t have a move constructor, then everything falls apart, because the original will be destructed under the lock by the
erase.Right – the bigger picture. I didn’t take that into account 🙁
A related reentrancy problem I've run into is having a ThingManager::clear() function that empties out the list of things under a mutex, where the Thing destructor can call back into ThingManager to access the list of things (IIRC, it was a map instead of a vector, and that accessor was doing some kind of lookup into the map).
The solution there was to swap the things list with an empty collection under the mutex, then destruct...
Well the non-scope_exit version will make more typename T move constructor and more destructor calls in unoptomised builds if you say “auto x = extract(vec, it)”.
I don’t think there’s a problem with the vector possibly realloc’ing on erase, because won’t properly coded move constructors, etc handle that? The code doesn’t seem to use invalidated iterators due to erase.
This reminds me of the explosives handling that the stdlib maintainers must do to ensure that exception guarantees are met for each of those containers. Carefully isolating all the risky business and then letting it explode at a time they can recover (if allowed). Moreover this code isn't immune will throw if the mutex has any issues... which it can for example if a is used in code that needs a instead.
Long...