Non-Reentrant Class Use In Multithreaded Program

Crag_Hack

It just says a specific instance not separate instances though right? From what you said before implies that non-reentrant class objects cannot even be accessed from different threads even if the instances are isolated to their own thread.

kshegunov

@Crag_Hack said in Non-Reentrant Class Use In Multithreaded Program:

It just says a specific instance not separate instances though right?

It's just the turn of phrase. "specific" here has the meaning of "a", "one" or "any".

From what you said before implies that non-reentrant class objects cannot even be accessed from different threads even if the instances are isolated to their own thread.

Yes, this is sort of what a definition of (non)reentrant is. If you want to get the formal one, here it goes:

A reentrant function is such a function whose effect, when called by two or more threads, is guaranteed to be as if the threads each executed the function one after another in an undefined order, even if the actual execution is interleaved

As per the POSIX C standard: ISO/IEC 9945:1-1996, §2.2.2

Or in other words, a function (or method in the context of C++) can be interrupted (e. g. by a context switch) and then re-entered (hence the name) without introducing side effects.

I personally find that the way it's written in Qt's documentation is more understandable to most, but both definitions are pretty much equivalent.

Crag_Hack

Ah I see, just a misunderstanding of the language on my behalf.

Crag_Hack

Quick follow-up question - it's safe to use qSort and foreach loops in my worker thread right? The reason I ask about foreach is I heard iterators in Qt are not safe to use in multiple threads.

kshegunov

Quick follow-up question - it's safe to use qSort and foreach loops in my worker thread right?

Only if the object isn't shared between threads. Usual rules for reentrant data apply - one can have only one thread writing (edit: without concurrent reads) at any one time. If your object is used from only one thread, then yes, it's safe, if not you need to serialize access (think a QMutex).

Crag_Hack

Thanks kshegunov - that applies to both qsort and foreach loops right?

kshegunov

Naturally, yes.

JonB

@Crag_Hack said in Non-Reentrant Class Use In Multithreaded Program:

Thanks kshegunov - that applies to both qsort and foreach loops right?

Just to be clear, it applies to everything you might write, whether that be foreach, qsort or anything else.

JonB

@kshegunov
I have two questions about using the QMutex pattern to synchronise reads with writes, please:

I assume that if the reader wants to read just one item it must still lock the whole object and the writer must lock the whole object, as the reader/iterator could be thrown by the write creating/deleting an item in the middle of the data?
Given the above, what pattern would we use to allow multiple-simultaneous-readers exclusive from single-writer? A reader would still need to grab the QMutex to prevent a writer, but we don't want to stop a second reader which would also want the QMutex?

kshegunov

These are relatively hard questions to answer, because it will depend on the way the data's organized. I'd make an effort to elaborate the best I can, nonetheless.

For 1:

Race conditions are an inevitable consequence of memory access, so if a race condition occurs depends heavily on the memory layout of the object/memory block in question. Say we are talking about a regular reentrant object, then it is safe in theory to call two methods simultaneously (i.e. from different threads) provided the mentioned two methods work on a different data member of that class. For example, if you have the very simple:

struct MyStruct
{
    int a, b;
};

Then you can completely safely read or modify a from one thread and b from from another as long as you don't touch the other member. The same considerations apply to plain C arrays of data, e.g.:

int * myarray = new int[200];

Then you can read or modify simultaneously from different threads different parts of the array; meaning you can write to the first element while reading the second from different threads safely. If you can guarantee there's no overlap between the memory accesses made from the different threads, then there's no race condition, and you're completely fine. This is exploited, for example, if you have the consumer-producer problem and there's an example in Qt's documentation of how it can be achieved with semaphores. To make your regular objects work like this, however, is often too finicky and error prone, so you'd usually opt for a serial access (i.e. through a mutex). Also see answer to question #2 for related information.

Now, if you use a container (for example std::vector or especially QVector due to implicit sharing) it's a bit more complicated. If you're working with std::vector you're fine obtaining a read or write iterator directly, because there's no data sharing behind the scenes. Then the only problem to solve is like with regular arrays - to ensure there are no concurrent memory accesses of the different elements. If you use QVector (or a QList, QHash, etc.) on the other hand then obtaining the write iterator is by itself a write operation, because the data may be shared and data copy might occur at that point. After the possible detach has happened then the same considerations as for std::vector apply.

For 2:
If you want to permit multiple readers and writers to access a specific data block, then a more elaborate scheme for locking is needed. As you correctly observed mutexes are exclusive, hence the name - mutual exclusive lock. They also guard a single resource, for example the whole array. If you need to have a threading primitive that provides guarding of multiple resources, then you need a semaphore (like in the consumer-producer example sourced above). On the other hand if you're satisfied with exclusive access only for write-induced races, which is exactly your question - allowing a single write at one time, but simultaneous reads, you could use Qt's QReadWriteLock which provides a locking scheme for that specific case (internal implementation isn't that important here, so I'll not venture into it).

Here I also would like to note that if you can prepopulate the data, then it could be a very viable approach to avoid needless locking. Because after you have the data in a container (provided it holds reentrant objects and the classes respect the method's const modifier, meaning no const_cast or mutable), you could use const-only access to that container and data in as many threads as you need. I quite often do that in my work with hashes.

JonB

@kshegunov
First, thanks for your comprehensive explanation.

For #1, you're making it way more complex than I had in mind. Say I have one thread writer & one thread reader, and my data is, say, a list/array of integers (forget vectors), no structures or whatever. I wish to use QMutex. All I wanted to verify is: (a) writer can insert/delete/update list/array; (b) reader wants just to read element #10; (c) confirm that the implementation must be QMutex for whole object/list/array, so that writer cannot change elements while reader trying to read one element.

For #2, QReadWriteLock is just the ticket instead of QMutex, thank you.

kshegunov

@JNBarchan said in Non-Reentrant Class Use In Multithreaded Program:

confirm that the implementation must be QMutex for whole object/list/array, so that writer cannot change elements while reader trying to read one element.

Yes, I confirm. While the mutex is locked the whole object/list/array is owned by the locking thread. All other threads requesting access are waiting and one of them (in an unspecified order) will acquire the mutex when the owning thread unlocks it.

VRonin

@kshegunov said in Non-Reentrant Class Use In Multithreaded Program:

and the classes respect the method's const modifier, meaning no const_cast or mutable

Just a small note, the above is not exhaustive. For example:

class BreakingConst{
int* a;
BreakingConst(int b=0) : a(new int(b)){}
~BreakingConst(){delete a;}
int getIncreasingA() const{ return (*a)++;}
};

has no const_cast or mutable but still breaks if used as described

kshegunov

It's probably somewhat of a fringe use case, but yes, you are right. The const modifier doesn't implicitly propagate to the indirectly referenced data. a is of type int * const in that case, which allows you to actually dereference the pointer and modify the underlying data. The same behavior is true for a heap-allocated array too, however with auto-storage:

class BreakingConst
{
    int a[1];

    BreakingConst(int b=0) { a[0] = b; }
    ~BreakingConst()  { delete a; }
    int getIncreasingA() const  { return a[0]++; }
};

The compiler knows what's going on and will prevent it.

Crag_Hack

Last question for this thread - if you go here to the first answer Kuba is complaining about iterators. Any merit to this claim? Anything to worry about? When he says internal pointers that doesn' t mean a pointer to a qt class does it? Only an iterator pointer right?
He says:

Sidebar: If you use iterators or internal pointers to data on non-const instances, you must forcibly detach() the object before constructing the iterators/pointers. The problem with iterators is that they become invalidated when an object's data is detached, and detaching can happen in any thread where the instance is non-const - so at least one thread will end up with invalid iterators. I won't talk any more of this, the takeaway is that implicitly shared data types are tricky to implement and use safely. With C++11, there's no need for implicit sharing anymore: they were a workaround for the lack of move semantics in C++98.

kshegunov

@Crag_Hack said in Non-Reentrant Class Use In Multithreaded Program:

Any merit to this claim?

Yes, some. See this thread on the mailing list.

Anything to worry about?

Don't call write operations, including non-const iterator retrieval, on an object from different threads!

When he says internal pointers that doesn' t mean a pointer to a qt class does it?

No, he means pointers to data that are kept internally by Qt. You don't have access to them, but you need to take into a account they exist. That's also the reason why in the docs it's said to not keep iterators to a container that may get modified (i.e. the container getting out of scope).

Only an iterator pointer right?

The iterator is a pointer to the data or an object keeping a pointer to the data. So in a sense you can say the iterator is a pointer (and it is for contiguous data classes like QVector).

With C++11, there's no need for implicit sharing anymore: they were a workaround for the lack of move semantics in C++98.

This is incorrect though. Qt's signal-slot mechanism could not work effectively without implicit sharing. Move semantics is not a panacea, and certainly solves no problems when an object has to be queued through Qt's event loop.

Crag_Hack

Thanks kshegunov... so nothing to worry about unless you use iterators right?

kshegunov

@kshegunov said in Non-Reentrant Class Use In Multithreaded Program:

Anything to worry about?

Don't call write operations, including non-const iterator retrieval, on an object from different threads!

Crag_Hack

I have two quick followup questions:

There's no possibility of two different non-reentrant classes sharing the same global (program-wise) resources, e. g. static variables? So say QStorageInfo uses the same globals/statics as QPair, you use QStorageInfo in one thread then QPair in another - they're not gonna fight each other and end up with a race condition and everything is safe right? (arbitrary choice of classes)
I'm not confident enough in my understanding of multithreading to feel comfortable with my code right now - so I'll just state it explicity. I am calling job.sortFileList(); in my worker thread. The job object is a custom class for backup jobs. I send it to the worker thread using signals and slots. The sortFileList() function follows as well as the comparison function. fileList is QList<std::pair<QString, QString> > fileList;

void BackupJob::sortFileList()
{
    qSort(fileList.begin(), fileList.end(), fileListCompare);
}
bool fileListCompare(const std::pair<QString, QString> &leftFileStringPair, const std::pair<QString, QString> &rightFileStringPair)
{...}

fileListCompare() is a function declared in the BackupJob.h file but not a member of the BackupJob class (global right?). Is this call to qSort safe? I'm guessing yes since the worker thread uses its own instance of the class (the class contains no non-reentrant objects, functions, or anything) and the iterators are constant (they're constant right?).
*edit - when I hover above the iterators in Qt Creator they appear as const_iterators not sure why though I didn't declare fileList as const
Thanks for the help!

kshegunov

@Crag_Hack said in Non-Reentrant Class Use In Multithreaded Program:

There's no possibility of two different non-reentrant classes sharing the same global (program-wise) resources, e. g. static variables? So say QStorageInfo uses the same globals/statics as QPair, you use QStorageInfo in one thread then QPair in another - they're not gonna fight each other and end up with a race condition and everything is safe right? (arbitrary choice of classes)

Actually this is exactly what makes the classes non-reentrant - having a shared global/static variable. Objects of those classes are safe to use from different threads only if they're explicitly thread-safe, e.g. QSemaphore/QMutex in contrast to QWidget, which is neither reentrant nor thread-safe.

Is this call to qSort safe?

As long as 2 conditions are met:

the fileListCompare doesn't use global data (i.e. static/global variables) - it's reentrant - or if it does, then all accesses to the globals is serialized between different threads (with a mutex) - it's thread-safe.
fileList isn't modified while qSort is executed. This would be true if BackupJob is reentrant, or more specifically - the thread operates on it's own instance of the fileList data member.

when I hover above the iterators in Qt Creator they appear as const_iterators not sure why though I didn't declare fileList as const

It's a deficiency of the IDE you use. qSort requires non-const iterators. As a side note, you should use std::sort instead of the legacy qSort.